| blob | e837f14d367b250699f4427167720c1610edbcaa |
1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2 ; RUN: llc < %s -mtriple=x86_64-- -mattr=+sse2 | FileCheck %s --check-prefixes=SSE
3 ; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx | FileCheck %s --check-prefixes=AVX
4 ; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx2 | FileCheck %s --check-prefixes=AVX2
5 ; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx2,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX2-FP
6 ; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx2,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX2-FCP
7 ; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl | FileCheck %s --check-prefixes=AVX512
8 ; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX512-FCP
9 ; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+avx512dq | FileCheck %s --check-prefixes=AVX512DQ
10 ; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+avx512dq,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX512DQ-FCP
11 ; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+avx512bw | FileCheck %s --check-prefixes=AVX512BW
12 ; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+avx512bw,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX512BW-FCP
13 ; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+avx512dq,+avx512bw | FileCheck %s --check-prefixes=AVX512DQ-BW
14 ; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+avx512dq,+avx512bw,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX512DQ-BW-FCP
16 ; These patterns are produced by LoopVectorizer for interleaved stores.
18 define void @store_i64_stride8_vf2(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecptr2, ptr %in.vecptr3, ptr %in.vecptr4, ptr %in.vecptr5, ptr %in.vecptr6, ptr %in.vecptr7, ptr %out.vec) nounwind {
19 ; SSE-LABEL: store_i64_stride8_vf2:
20 ; SSE: # %bb.0:
21 ; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax
22 ; SSE-NEXT: movq {{[0-9]+}}(%rsp), %r10
23 ; SSE-NEXT: movq {{[0-9]+}}(%rsp), %r11
24 ; SSE-NEXT: movaps (%rdi), %xmm0
25 ; SSE-NEXT: movaps (%rsi), %xmm1
26 ; SSE-NEXT: movaps (%rdx), %xmm2
27 ; SSE-NEXT: movaps (%rcx), %xmm3
28 ; SSE-NEXT: movaps (%r8), %xmm4
29 ; SSE-NEXT: movaps (%r9), %xmm5
30 ; SSE-NEXT: movaps (%r11), %xmm6
31 ; SSE-NEXT: movaps (%r10), %xmm7
32 ; SSE-NEXT: movaps %xmm0, %xmm8
33 ; SSE-NEXT: movlhps {{.*#+}} xmm8 = xmm8[0],xmm1[0]
34 ; SSE-NEXT: movaps %xmm2, %xmm9
35 ; SSE-NEXT: movlhps {{.*#+}} xmm9 = xmm9[0],xmm3[0]
36 ; SSE-NEXT: movaps %xmm6, %xmm10
37 ; SSE-NEXT: movlhps {{.*#+}} xmm10 = xmm10[0],xmm7[0]
38 ; SSE-NEXT: movaps %xmm4, %xmm11
39 ; SSE-NEXT: movlhps {{.*#+}} xmm11 = xmm11[0],xmm5[0]
40 ; SSE-NEXT: unpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
41 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm3[1]
42 ; SSE-NEXT: unpckhpd {{.*#+}} xmm6 = xmm6[1],xmm7[1]
43 ; SSE-NEXT: unpckhpd {{.*#+}} xmm4 = xmm4[1],xmm5[1]
44 ; SSE-NEXT: movaps %xmm4, 96(%rax)
45 ; SSE-NEXT: movaps %xmm6, 112(%rax)
46 ; SSE-NEXT: movaps %xmm2, 80(%rax)
47 ; SSE-NEXT: movaps %xmm0, 64(%rax)
48 ; SSE-NEXT: movaps %xmm11, 32(%rax)
49 ; SSE-NEXT: movaps %xmm10, 48(%rax)
50 ; SSE-NEXT: movaps %xmm9, 16(%rax)
51 ; SSE-NEXT: movaps %xmm8, (%rax)
52 ; SSE-NEXT: retq
53 ;
54 ; AVX-LABEL: store_i64_stride8_vf2:
55 ; AVX: # %bb.0:
56 ; AVX-NEXT: movq {{[0-9]+}}(%rsp), %rax
57 ; AVX-NEXT: movq {{[0-9]+}}(%rsp), %r10
58 ; AVX-NEXT: movq {{[0-9]+}}(%rsp), %r11
59 ; AVX-NEXT: vmovaps (%rdi), %xmm0
60 ; AVX-NEXT: vmovaps (%rsi), %xmm1
61 ; AVX-NEXT: vmovaps (%r8), %xmm2
62 ; AVX-NEXT: vmovaps (%r9), %xmm3
63 ; AVX-NEXT: vinsertf128 $1, (%rcx), %ymm1, %ymm1
64 ; AVX-NEXT: vinsertf128 $1, (%rdx), %ymm0, %ymm0
65 ; AVX-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
66 ; AVX-NEXT: vinsertf128 $1, (%r11), %ymm3, %ymm3
67 ; AVX-NEXT: vinsertf128 $1, (%r10), %ymm2, %ymm2
68 ; AVX-NEXT: vunpcklpd {{.*#+}} ymm5 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
69 ; AVX-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
70 ; AVX-NEXT: vunpckhpd {{.*#+}} ymm1 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
71 ; AVX-NEXT: vmovaps %ymm1, 96(%rax)
72 ; AVX-NEXT: vmovaps %ymm0, 64(%rax)
73 ; AVX-NEXT: vmovaps %ymm5, 32(%rax)
74 ; AVX-NEXT: vmovaps %ymm4, (%rax)
75 ; AVX-NEXT: vzeroupper
76 ; AVX-NEXT: retq
77 ;
78 ; AVX2-LABEL: store_i64_stride8_vf2:
79 ; AVX2: # %bb.0:
80 ; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %rax
81 ; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %r10
82 ; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %r11
83 ; AVX2-NEXT: vmovaps (%rdi), %xmm0
84 ; AVX2-NEXT: vmovaps (%rdx), %xmm1
85 ; AVX2-NEXT: vmovaps (%r8), %xmm2
86 ; AVX2-NEXT: vmovaps (%r11), %xmm3
87 ; AVX2-NEXT: vinsertf128 $1, (%rsi), %ymm0, %ymm0
88 ; AVX2-NEXT: vinsertf128 $1, (%rcx), %ymm1, %ymm1
89 ; AVX2-NEXT: vinsertf128 $1, (%r9), %ymm2, %ymm2
90 ; AVX2-NEXT: vinsertf128 $1, (%r10), %ymm3, %ymm3
91 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
92 ; AVX2-NEXT: vpermpd {{.*#+}} ymm4 = ymm4[0,2,1,3]
93 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm5 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
94 ; AVX2-NEXT: vpermpd {{.*#+}} ymm5 = ymm5[0,2,1,3]
95 ; AVX2-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
96 ; AVX2-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,2,1,3]
97 ; AVX2-NEXT: vunpckhpd {{.*#+}} ymm1 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
98 ; AVX2-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,2,1,3]
99 ; AVX2-NEXT: vmovaps %ymm1, 96(%rax)
100 ; AVX2-NEXT: vmovaps %ymm0, 64(%rax)
101 ; AVX2-NEXT: vmovaps %ymm5, 32(%rax)
102 ; AVX2-NEXT: vmovaps %ymm4, (%rax)
103 ; AVX2-NEXT: vzeroupper
104 ; AVX2-NEXT: retq
105 ;
106 ; AVX2-FP-LABEL: store_i64_stride8_vf2:
107 ; AVX2-FP: # %bb.0:
108 ; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %rax
109 ; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %r10
110 ; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %r11
111 ; AVX2-FP-NEXT: vmovaps (%rdi), %xmm0
112 ; AVX2-FP-NEXT: vmovaps (%rdx), %xmm1
113 ; AVX2-FP-NEXT: vmovaps (%r8), %xmm2
114 ; AVX2-FP-NEXT: vmovaps (%r11), %xmm3
115 ; AVX2-FP-NEXT: vinsertf128 $1, (%rsi), %ymm0, %ymm0
116 ; AVX2-FP-NEXT: vinsertf128 $1, (%rcx), %ymm1, %ymm1
117 ; AVX2-FP-NEXT: vinsertf128 $1, (%r9), %ymm2, %ymm2
118 ; AVX2-FP-NEXT: vinsertf128 $1, (%r10), %ymm3, %ymm3
119 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
120 ; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm4 = ymm4[0,2,1,3]
121 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm5 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
122 ; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm5 = ymm5[0,2,1,3]
123 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
124 ; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,2,1,3]
125 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm1 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
126 ; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,2,1,3]
127 ; AVX2-FP-NEXT: vmovaps %ymm1, 96(%rax)
128 ; AVX2-FP-NEXT: vmovaps %ymm0, 64(%rax)
129 ; AVX2-FP-NEXT: vmovaps %ymm5, 32(%rax)
130 ; AVX2-FP-NEXT: vmovaps %ymm4, (%rax)
131 ; AVX2-FP-NEXT: vzeroupper
132 ; AVX2-FP-NEXT: retq
133 ;
134 ; AVX2-FCP-LABEL: store_i64_stride8_vf2:
135 ; AVX2-FCP: # %bb.0:
136 ; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
137 ; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10
138 ; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r11
139 ; AVX2-FCP-NEXT: vmovaps (%rdi), %xmm0
140 ; AVX2-FCP-NEXT: vmovaps (%rdx), %xmm1
141 ; AVX2-FCP-NEXT: vmovaps (%r8), %xmm2
142 ; AVX2-FCP-NEXT: vmovaps (%r11), %xmm3
143 ; AVX2-FCP-NEXT: vinsertf128 $1, (%rsi), %ymm0, %ymm0
144 ; AVX2-FCP-NEXT: vinsertf128 $1, (%rcx), %ymm1, %ymm1
145 ; AVX2-FCP-NEXT: vinsertf128 $1, (%r9), %ymm2, %ymm2
146 ; AVX2-FCP-NEXT: vinsertf128 $1, (%r10), %ymm3, %ymm3
147 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
148 ; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm4 = ymm4[0,2,1,3]
149 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm5 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
150 ; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm5 = ymm5[0,2,1,3]
151 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
152 ; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,2,1,3]
153 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm1 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
154 ; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,2,1,3]
155 ; AVX2-FCP-NEXT: vmovaps %ymm1, 96(%rax)
156 ; AVX2-FCP-NEXT: vmovaps %ymm0, 64(%rax)
157 ; AVX2-FCP-NEXT: vmovaps %ymm5, 32(%rax)
158 ; AVX2-FCP-NEXT: vmovaps %ymm4, (%rax)
159 ; AVX2-FCP-NEXT: vzeroupper
160 ; AVX2-FCP-NEXT: retq
161 ;
162 ; AVX512-LABEL: store_i64_stride8_vf2:
163 ; AVX512: # %bb.0:
164 ; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %rax
165 ; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %r10
166 ; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %r11
167 ; AVX512-NEXT: vmovdqa (%rdi), %xmm0
168 ; AVX512-NEXT: vmovdqa (%rdx), %xmm1
169 ; AVX512-NEXT: vmovdqa (%r8), %xmm2
170 ; AVX512-NEXT: vmovdqa (%r11), %xmm3
171 ; AVX512-NEXT: vinserti128 $1, (%rcx), %ymm1, %ymm1
172 ; AVX512-NEXT: vinserti128 $1, (%rsi), %ymm0, %ymm0
173 ; AVX512-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
174 ; AVX512-NEXT: vinserti128 $1, (%r10), %ymm3, %ymm1
175 ; AVX512-NEXT: vinserti128 $1, (%r9), %ymm2, %ymm2
176 ; AVX512-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm1
177 ; AVX512-NEXT: vpmovsxbq {{.*#+}} zmm2 = [0,2,4,6,8,10,12,14]
178 ; AVX512-NEXT: vpermi2q %zmm1, %zmm0, %zmm2
179 ; AVX512-NEXT: vpmovsxbq {{.*#+}} zmm3 = [1,3,5,7,9,11,13,15]
180 ; AVX512-NEXT: vpermi2q %zmm1, %zmm0, %zmm3
181 ; AVX512-NEXT: vmovdqa64 %zmm3, 64(%rax)
182 ; AVX512-NEXT: vmovdqa64 %zmm2, (%rax)
183 ; AVX512-NEXT: vzeroupper
184 ; AVX512-NEXT: retq
185 ;
186 ; AVX512-FCP-LABEL: store_i64_stride8_vf2:
187 ; AVX512-FCP: # %bb.0:
188 ; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
189 ; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10
190 ; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r11
191 ; AVX512-FCP-NEXT: vmovdqa (%rdi), %xmm0
192 ; AVX512-FCP-NEXT: vmovdqa (%rdx), %xmm1
193 ; AVX512-FCP-NEXT: vmovdqa (%r8), %xmm2
194 ; AVX512-FCP-NEXT: vmovdqa (%r11), %xmm3
195 ; AVX512-FCP-NEXT: vinserti128 $1, (%rcx), %ymm1, %ymm1
196 ; AVX512-FCP-NEXT: vinserti128 $1, (%rsi), %ymm0, %ymm0
197 ; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
198 ; AVX512-FCP-NEXT: vinserti128 $1, (%r10), %ymm3, %ymm1
199 ; AVX512-FCP-NEXT: vinserti128 $1, (%r9), %ymm2, %ymm2
200 ; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm1
201 ; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm2 = [0,2,4,6,8,10,12,14]
202 ; AVX512-FCP-NEXT: vpermi2q %zmm1, %zmm0, %zmm2
203 ; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm3 = [1,3,5,7,9,11,13,15]
204 ; AVX512-FCP-NEXT: vpermi2q %zmm1, %zmm0, %zmm3
205 ; AVX512-FCP-NEXT: vmovdqa64 %zmm3, 64(%rax)
206 ; AVX512-FCP-NEXT: vmovdqa64 %zmm2, (%rax)
207 ; AVX512-FCP-NEXT: vzeroupper
208 ; AVX512-FCP-NEXT: retq
209 ;
210 ; AVX512DQ-LABEL: store_i64_stride8_vf2:
211 ; AVX512DQ: # %bb.0:
212 ; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %rax
213 ; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %r10
214 ; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %r11
215 ; AVX512DQ-NEXT: vmovdqa (%rdi), %xmm0
216 ; AVX512DQ-NEXT: vmovdqa (%rdx), %xmm1
217 ; AVX512DQ-NEXT: vmovdqa (%r8), %xmm2
218 ; AVX512DQ-NEXT: vmovdqa (%r11), %xmm3
219 ; AVX512DQ-NEXT: vinserti128 $1, (%rcx), %ymm1, %ymm1
220 ; AVX512DQ-NEXT: vinserti128 $1, (%rsi), %ymm0, %ymm0
221 ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
222 ; AVX512DQ-NEXT: vinserti128 $1, (%r10), %ymm3, %ymm1
223 ; AVX512DQ-NEXT: vinserti128 $1, (%r9), %ymm2, %ymm2
224 ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm1
225 ; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} zmm2 = [0,2,4,6,8,10,12,14]
226 ; AVX512DQ-NEXT: vpermi2q %zmm1, %zmm0, %zmm2
227 ; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} zmm3 = [1,3,5,7,9,11,13,15]
228 ; AVX512DQ-NEXT: vpermi2q %zmm1, %zmm0, %zmm3
229 ; AVX512DQ-NEXT: vmovdqa64 %zmm3, 64(%rax)
230 ; AVX512DQ-NEXT: vmovdqa64 %zmm2, (%rax)
231 ; AVX512DQ-NEXT: vzeroupper
232 ; AVX512DQ-NEXT: retq
233 ;
234 ; AVX512DQ-FCP-LABEL: store_i64_stride8_vf2:
235 ; AVX512DQ-FCP: # %bb.0:
236 ; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
237 ; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10
238 ; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r11
239 ; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %xmm0
240 ; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %xmm1
241 ; AVX512DQ-FCP-NEXT: vmovdqa (%r8), %xmm2
242 ; AVX512DQ-FCP-NEXT: vmovdqa (%r11), %xmm3
243 ; AVX512DQ-FCP-NEXT: vinserti128 $1, (%rcx), %ymm1, %ymm1
244 ; AVX512DQ-FCP-NEXT: vinserti128 $1, (%rsi), %ymm0, %ymm0
245 ; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
246 ; AVX512DQ-FCP-NEXT: vinserti128 $1, (%r10), %ymm3, %ymm1
247 ; AVX512DQ-FCP-NEXT: vinserti128 $1, (%r9), %ymm2, %ymm2
248 ; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm1
249 ; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm2 = [0,2,4,6,8,10,12,14]
250 ; AVX512DQ-FCP-NEXT: vpermi2q %zmm1, %zmm0, %zmm2
251 ; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm3 = [1,3,5,7,9,11,13,15]
252 ; AVX512DQ-FCP-NEXT: vpermi2q %zmm1, %zmm0, %zmm3
253 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, 64(%rax)
254 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, (%rax)
255 ; AVX512DQ-FCP-NEXT: vzeroupper
256 ; AVX512DQ-FCP-NEXT: retq
257 ;
258 ; AVX512BW-LABEL: store_i64_stride8_vf2:
259 ; AVX512BW: # %bb.0:
260 ; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
261 ; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %r10
262 ; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %r11
263 ; AVX512BW-NEXT: vmovdqa (%rdi), %xmm0
264 ; AVX512BW-NEXT: vmovdqa (%rdx), %xmm1
265 ; AVX512BW-NEXT: vmovdqa (%r8), %xmm2
266 ; AVX512BW-NEXT: vmovdqa (%r11), %xmm3
267 ; AVX512BW-NEXT: vinserti128 $1, (%rcx), %ymm1, %ymm1
268 ; AVX512BW-NEXT: vinserti128 $1, (%rsi), %ymm0, %ymm0
269 ; AVX512BW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
270 ; AVX512BW-NEXT: vinserti128 $1, (%r10), %ymm3, %ymm1
271 ; AVX512BW-NEXT: vinserti128 $1, (%r9), %ymm2, %ymm2
272 ; AVX512BW-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm1
273 ; AVX512BW-NEXT: vpmovsxbq {{.*#+}} zmm2 = [0,2,4,6,8,10,12,14]
274 ; AVX512BW-NEXT: vpermi2q %zmm1, %zmm0, %zmm2
275 ; AVX512BW-NEXT: vpmovsxbq {{.*#+}} zmm3 = [1,3,5,7,9,11,13,15]
276 ; AVX512BW-NEXT: vpermi2q %zmm1, %zmm0, %zmm3
277 ; AVX512BW-NEXT: vmovdqa64 %zmm3, 64(%rax)
278 ; AVX512BW-NEXT: vmovdqa64 %zmm2, (%rax)
279 ; AVX512BW-NEXT: vzeroupper
280 ; AVX512BW-NEXT: retq
281 ;
282 ; AVX512BW-FCP-LABEL: store_i64_stride8_vf2:
283 ; AVX512BW-FCP: # %bb.0:
284 ; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
285 ; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10
286 ; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r11
287 ; AVX512BW-FCP-NEXT: vmovdqa (%rdi), %xmm0
288 ; AVX512BW-FCP-NEXT: vmovdqa (%rdx), %xmm1
289 ; AVX512BW-FCP-NEXT: vmovdqa (%r8), %xmm2
290 ; AVX512BW-FCP-NEXT: vmovdqa (%r11), %xmm3
291 ; AVX512BW-FCP-NEXT: vinserti128 $1, (%rcx), %ymm1, %ymm1
292 ; AVX512BW-FCP-NEXT: vinserti128 $1, (%rsi), %ymm0, %ymm0
293 ; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
294 ; AVX512BW-FCP-NEXT: vinserti128 $1, (%r10), %ymm3, %ymm1
295 ; AVX512BW-FCP-NEXT: vinserti128 $1, (%r9), %ymm2, %ymm2
296 ; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm1
297 ; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm2 = [0,2,4,6,8,10,12,14]
298 ; AVX512BW-FCP-NEXT: vpermi2q %zmm1, %zmm0, %zmm2
299 ; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm3 = [1,3,5,7,9,11,13,15]
300 ; AVX512BW-FCP-NEXT: vpermi2q %zmm1, %zmm0, %zmm3
301 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, 64(%rax)
302 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, (%rax)
303 ; AVX512BW-FCP-NEXT: vzeroupper
304 ; AVX512BW-FCP-NEXT: retq
305 ;
306 ; AVX512DQ-BW-LABEL: store_i64_stride8_vf2:
307 ; AVX512DQ-BW: # %bb.0:
308 ; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
309 ; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %r10
310 ; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %r11
311 ; AVX512DQ-BW-NEXT: vmovdqa (%rdi), %xmm0
312 ; AVX512DQ-BW-NEXT: vmovdqa (%rdx), %xmm1
313 ; AVX512DQ-BW-NEXT: vmovdqa (%r8), %xmm2
314 ; AVX512DQ-BW-NEXT: vmovdqa (%r11), %xmm3
315 ; AVX512DQ-BW-NEXT: vinserti128 $1, (%rcx), %ymm1, %ymm1
316 ; AVX512DQ-BW-NEXT: vinserti128 $1, (%rsi), %ymm0, %ymm0
317 ; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
318 ; AVX512DQ-BW-NEXT: vinserti128 $1, (%r10), %ymm3, %ymm1
319 ; AVX512DQ-BW-NEXT: vinserti128 $1, (%r9), %ymm2, %ymm2
320 ; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm1
321 ; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} zmm2 = [0,2,4,6,8,10,12,14]
322 ; AVX512DQ-BW-NEXT: vpermi2q %zmm1, %zmm0, %zmm2
323 ; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} zmm3 = [1,3,5,7,9,11,13,15]
324 ; AVX512DQ-BW-NEXT: vpermi2q %zmm1, %zmm0, %zmm3
325 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, 64(%rax)
326 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, (%rax)
327 ; AVX512DQ-BW-NEXT: vzeroupper
328 ; AVX512DQ-BW-NEXT: retq
329 ;
330 ; AVX512DQ-BW-FCP-LABEL: store_i64_stride8_vf2:
331 ; AVX512DQ-BW-FCP: # %bb.0:
332 ; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
333 ; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10
334 ; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r11
335 ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdi), %xmm0
336 ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdx), %xmm1
337 ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%r8), %xmm2
338 ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%r11), %xmm3
339 ; AVX512DQ-BW-FCP-NEXT: vinserti128 $1, (%rcx), %ymm1, %ymm1
340 ; AVX512DQ-BW-FCP-NEXT: vinserti128 $1, (%rsi), %ymm0, %ymm0
341 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
342 ; AVX512DQ-BW-FCP-NEXT: vinserti128 $1, (%r10), %ymm3, %ymm1
343 ; AVX512DQ-BW-FCP-NEXT: vinserti128 $1, (%r9), %ymm2, %ymm2
344 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm1
345 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm2 = [0,2,4,6,8,10,12,14]
346 ; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm1, %zmm0, %zmm2
347 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm3 = [1,3,5,7,9,11,13,15]
348 ; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm1, %zmm0, %zmm3
349 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, 64(%rax)
350 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, (%rax)
351 ; AVX512DQ-BW-FCP-NEXT: vzeroupper
352 ; AVX512DQ-BW-FCP-NEXT: retq
353 %in.vec0 = load <2 x i64>, ptr %in.vecptr0, align 64
354 %in.vec1 = load <2 x i64>, ptr %in.vecptr1, align 64
355 %in.vec2 = load <2 x i64>, ptr %in.vecptr2, align 64
356 %in.vec3 = load <2 x i64>, ptr %in.vecptr3, align 64
357 %in.vec4 = load <2 x i64>, ptr %in.vecptr4, align 64
358 %in.vec5 = load <2 x i64>, ptr %in.vecptr5, align 64
359 %in.vec6 = load <2 x i64>, ptr %in.vecptr6, align 64
360 %in.vec7 = load <2 x i64>, ptr %in.vecptr7, align 64
361 %1 = shufflevector <2 x i64> %in.vec0, <2 x i64> %in.vec1, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
362 %2 = shufflevector <2 x i64> %in.vec2, <2 x i64> %in.vec3, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
363 %3 = shufflevector <2 x i64> %in.vec4, <2 x i64> %in.vec5, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
364 %4 = shufflevector <2 x i64> %in.vec6, <2 x i64> %in.vec7, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
365 %5 = shufflevector <4 x i64> %1, <4 x i64> %2, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
366 %6 = shufflevector <4 x i64> %3, <4 x i64> %4, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
367 %7 = shufflevector <8 x i64> %5, <8 x i64> %6, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
368 %interleaved.vec = shufflevector <16 x i64> %7, <16 x i64> poison, <16 x i32> <i32 0, i32 2, i32 4, i32 6, i32 8, i32 10, i32 12, i32 14, i32 1, i32 3, i32 5, i32 7, i32 9, i32 11, i32 13, i32 15>
369 store <16 x i64> %interleaved.vec, ptr %out.vec, align 64
370 ret void
371 }
373 define void @store_i64_stride8_vf4(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecptr2, ptr %in.vecptr3, ptr %in.vecptr4, ptr %in.vecptr5, ptr %in.vecptr6, ptr %in.vecptr7, ptr %out.vec) nounwind {
374 ; SSE-LABEL: store_i64_stride8_vf4:
375 ; SSE: # %bb.0:
376 ; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax
377 ; SSE-NEXT: movq {{[0-9]+}}(%rsp), %r10
378 ; SSE-NEXT: movaps (%rdi), %xmm4
379 ; SSE-NEXT: movaps 16(%rdi), %xmm2
380 ; SSE-NEXT: movaps (%rsi), %xmm10
381 ; SSE-NEXT: movaps 16(%rsi), %xmm14
382 ; SSE-NEXT: movaps (%rdx), %xmm1
383 ; SSE-NEXT: movaps 16(%rdx), %xmm3
384 ; SSE-NEXT: movaps (%rcx), %xmm7
385 ; SSE-NEXT: movaps 16(%rcx), %xmm12
386 ; SSE-NEXT: movaps (%r8), %xmm5
387 ; SSE-NEXT: movaps 16(%r8), %xmm0
388 ; SSE-NEXT: movaps (%r9), %xmm13
389 ; SSE-NEXT: movaps (%r10), %xmm6
390 ; SSE-NEXT: movaps 16(%r10), %xmm9
391 ; SSE-NEXT: movaps (%rax), %xmm15
392 ; SSE-NEXT: movaps 16(%rax), %xmm11
393 ; SSE-NEXT: movaps %xmm1, %xmm8
394 ; SSE-NEXT: movlhps {{.*#+}} xmm8 = xmm8[0],xmm7[0]
395 ; SSE-NEXT: movaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
396 ; SSE-NEXT: unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm7[1]
397 ; SSE-NEXT: movaps %xmm4, %xmm7
398 ; SSE-NEXT: movlhps {{.*#+}} xmm7 = xmm7[0],xmm10[0]
399 ; SSE-NEXT: unpckhpd {{.*#+}} xmm4 = xmm4[1],xmm10[1]
400 ; SSE-NEXT: movaps %xmm3, %xmm10
401 ; SSE-NEXT: movlhps {{.*#+}} xmm10 = xmm10[0],xmm12[0]
402 ; SSE-NEXT: unpckhpd {{.*#+}} xmm3 = xmm3[1],xmm12[1]
403 ; SSE-NEXT: movaps %xmm2, %xmm12
404 ; SSE-NEXT: movlhps {{.*#+}} xmm12 = xmm12[0],xmm14[0]
405 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm14[1]
406 ; SSE-NEXT: movaps %xmm6, %xmm14
407 ; SSE-NEXT: movlhps {{.*#+}} xmm14 = xmm14[0],xmm15[0]
408 ; SSE-NEXT: unpckhpd {{.*#+}} xmm6 = xmm6[1],xmm15[1]
409 ; SSE-NEXT: movaps %xmm5, %xmm15
410 ; SSE-NEXT: movlhps {{.*#+}} xmm15 = xmm15[0],xmm13[0]
411 ; SSE-NEXT: unpckhpd {{.*#+}} xmm5 = xmm5[1],xmm13[1]
412 ; SSE-NEXT: movaps %xmm9, %xmm13
413 ; SSE-NEXT: movlhps {{.*#+}} xmm13 = xmm13[0],xmm11[0]
414 ; SSE-NEXT: unpckhpd {{.*#+}} xmm9 = xmm9[1],xmm11[1]
415 ; SSE-NEXT: movaps 16(%r9), %xmm11
416 ; SSE-NEXT: movaps %xmm0, %xmm8
417 ; SSE-NEXT: movlhps {{.*#+}} xmm8 = xmm8[0],xmm11[0]
418 ; SSE-NEXT: unpckhpd {{.*#+}} xmm0 = xmm0[1],xmm11[1]
419 ; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax
420 ; SSE-NEXT: movaps %xmm0, 224(%rax)
421 ; SSE-NEXT: movaps %xmm9, 240(%rax)
422 ; SSE-NEXT: movaps %xmm8, 160(%rax)
423 ; SSE-NEXT: movaps %xmm13, 176(%rax)
424 ; SSE-NEXT: movaps %xmm5, 96(%rax)
425 ; SSE-NEXT: movaps %xmm6, 112(%rax)
426 ; SSE-NEXT: movaps %xmm15, 32(%rax)
427 ; SSE-NEXT: movaps %xmm14, 48(%rax)
428 ; SSE-NEXT: movaps %xmm2, 192(%rax)
429 ; SSE-NEXT: movaps %xmm3, 208(%rax)
430 ; SSE-NEXT: movaps %xmm12, 128(%rax)
431 ; SSE-NEXT: movaps %xmm10, 144(%rax)
432 ; SSE-NEXT: movaps %xmm4, 64(%rax)
433 ; SSE-NEXT: movaps %xmm1, 80(%rax)
434 ; SSE-NEXT: movaps %xmm7, (%rax)
435 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
436 ; SSE-NEXT: movaps %xmm0, 16(%rax)
437 ; SSE-NEXT: retq
438 ;
439 ; AVX-LABEL: store_i64_stride8_vf4:
440 ; AVX: # %bb.0:
441 ; AVX-NEXT: movq {{[0-9]+}}(%rsp), %rax
442 ; AVX-NEXT: movq {{[0-9]+}}(%rsp), %r10
443 ; AVX-NEXT: movq {{[0-9]+}}(%rsp), %r11
444 ; AVX-NEXT: vmovaps (%rdx), %ymm2
445 ; AVX-NEXT: vmovaps (%rcx), %ymm3
446 ; AVX-NEXT: vmovaps (%r11), %ymm1
447 ; AVX-NEXT: vmovaps (%r10), %ymm4
448 ; AVX-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm4[0],ymm1[2],ymm4[2]
449 ; AVX-NEXT: vmovaps 16(%r9), %xmm5
450 ; AVX-NEXT: vmovaps 16(%r8), %xmm6
451 ; AVX-NEXT: vmovlhps {{.*#+}} xmm7 = xmm6[0],xmm5[0]
452 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm7[0,1,2,3],ymm0[4,5,6,7]
453 ; AVX-NEXT: vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm4[1],ymm1[3],ymm4[3]
454 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm4 = xmm6[1],xmm5[1]
455 ; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm4[0,1,2,3],ymm1[4,5,6,7]
456 ; AVX-NEXT: vunpckhpd {{.*#+}} ymm4 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
457 ; AVX-NEXT: vmovaps 16(%rsi), %xmm5
458 ; AVX-NEXT: vmovaps 16(%rdi), %xmm6
459 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm7 = xmm6[1],xmm5[1]
460 ; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm7[0,1,2,3],ymm4[4,5,6,7]
461 ; AVX-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
462 ; AVX-NEXT: vmovlhps {{.*#+}} xmm3 = xmm6[0],xmm5[0]
463 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm3[0,1,2,3],ymm2[4,5,6,7]
464 ; AVX-NEXT: vmovaps (%r9), %xmm3
465 ; AVX-NEXT: vmovaps (%r8), %xmm5
466 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm6 = xmm5[1],xmm3[1]
467 ; AVX-NEXT: vmovaps (%r10), %xmm7
468 ; AVX-NEXT: vmovaps (%r11), %xmm8
469 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm9 = xmm8[1],xmm7[1]
470 ; AVX-NEXT: vmovlhps {{.*#+}} xmm3 = xmm5[0],xmm3[0]
471 ; AVX-NEXT: vmovlhps {{.*#+}} xmm5 = xmm8[0],xmm7[0]
472 ; AVX-NEXT: vmovaps (%rsi), %xmm7
473 ; AVX-NEXT: vmovaps (%rdi), %xmm8
474 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm10 = xmm8[1],xmm7[1]
475 ; AVX-NEXT: vmovaps (%rcx), %xmm11
476 ; AVX-NEXT: vmovaps (%rdx), %xmm12
477 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm13 = xmm12[1],xmm11[1]
478 ; AVX-NEXT: vmovlhps {{.*#+}} xmm7 = xmm8[0],xmm7[0]
479 ; AVX-NEXT: vmovlhps {{.*#+}} xmm8 = xmm12[0],xmm11[0]
480 ; AVX-NEXT: vmovaps %xmm8, 16(%rax)
481 ; AVX-NEXT: vmovaps %xmm7, (%rax)
482 ; AVX-NEXT: vmovaps %xmm13, 80(%rax)
483 ; AVX-NEXT: vmovaps %xmm10, 64(%rax)
484 ; AVX-NEXT: vmovaps %xmm5, 48(%rax)
485 ; AVX-NEXT: vmovaps %xmm3, 32(%rax)
486 ; AVX-NEXT: vmovaps %xmm9, 112(%rax)
487 ; AVX-NEXT: vmovaps %xmm6, 96(%rax)
488 ; AVX-NEXT: vmovaps %ymm2, 128(%rax)
489 ; AVX-NEXT: vmovaps %ymm4, 192(%rax)
490 ; AVX-NEXT: vmovaps %ymm1, 224(%rax)
491 ; AVX-NEXT: vmovaps %ymm0, 160(%rax)
492 ; AVX-NEXT: vzeroupper
493 ; AVX-NEXT: retq
494 ;
495 ; AVX2-LABEL: store_i64_stride8_vf4:
496 ; AVX2: # %bb.0:
497 ; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %rax
498 ; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %r10
499 ; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %r11
500 ; AVX2-NEXT: vmovaps (%rdi), %ymm0
501 ; AVX2-NEXT: vmovaps (%rsi), %ymm1
502 ; AVX2-NEXT: vmovaps (%rdx), %ymm2
503 ; AVX2-NEXT: vmovaps (%rcx), %ymm3
504 ; AVX2-NEXT: vmovaps (%r8), %ymm4
505 ; AVX2-NEXT: vmovaps (%r9), %ymm5
506 ; AVX2-NEXT: vmovaps (%r11), %ymm6
507 ; AVX2-NEXT: vmovaps (%r10), %ymm7
508 ; AVX2-NEXT: vmovaps (%r9), %xmm8
509 ; AVX2-NEXT: vinsertf128 $1, (%r10), %ymm8, %ymm8
510 ; AVX2-NEXT: vmovaps (%r8), %xmm9
511 ; AVX2-NEXT: vinsertf128 $1, (%r11), %ymm9, %ymm9
512 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm10 = ymm9[0],ymm8[0],ymm9[2],ymm8[2]
513 ; AVX2-NEXT: vunpckhpd {{.*#+}} ymm8 = ymm9[1],ymm8[1],ymm9[3],ymm8[3]
514 ; AVX2-NEXT: vmovaps (%rsi), %xmm9
515 ; AVX2-NEXT: vinsertf128 $1, (%rcx), %ymm9, %ymm9
516 ; AVX2-NEXT: vmovaps (%rdi), %xmm11
517 ; AVX2-NEXT: vinsertf128 $1, (%rdx), %ymm11, %ymm11
518 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm12 = ymm11[0],ymm9[0],ymm11[2],ymm9[2]
519 ; AVX2-NEXT: vunpckhpd {{.*#+}} ymm9 = ymm11[1],ymm9[1],ymm11[3],ymm9[3]
520 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm11 = ymm6[0],ymm7[0],ymm6[2],ymm7[2]
521 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm13 = ymm4[0],ymm5[0],ymm4[2],ymm5[2]
522 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm11 = ymm13[2,3],ymm11[2,3]
523 ; AVX2-NEXT: vunpckhpd {{.*#+}} ymm6 = ymm6[1],ymm7[1],ymm6[3],ymm7[3]
524 ; AVX2-NEXT: vunpckhpd {{.*#+}} ymm4 = ymm4[1],ymm5[1],ymm4[3],ymm5[3]
525 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],ymm6[2,3]
526 ; AVX2-NEXT: vunpckhpd {{.*#+}} ymm5 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
527 ; AVX2-NEXT: vunpckhpd {{.*#+}} ymm6 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
528 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm5 = ymm6[2,3],ymm5[2,3]
529 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
530 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
531 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm2[2,3]
532 ; AVX2-NEXT: vmovaps %ymm0, 128(%rax)
533 ; AVX2-NEXT: vmovaps %ymm5, 192(%rax)
534 ; AVX2-NEXT: vmovaps %ymm4, 224(%rax)
535 ; AVX2-NEXT: vmovaps %ymm11, 160(%rax)
536 ; AVX2-NEXT: vmovaps %ymm9, 64(%rax)
537 ; AVX2-NEXT: vmovaps %ymm12, (%rax)
538 ; AVX2-NEXT: vmovaps %ymm8, 96(%rax)
539 ; AVX2-NEXT: vmovaps %ymm10, 32(%rax)
540 ; AVX2-NEXT: vzeroupper
541 ; AVX2-NEXT: retq
542 ;
543 ; AVX2-FP-LABEL: store_i64_stride8_vf4:
544 ; AVX2-FP: # %bb.0:
545 ; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %rax
546 ; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %r10
547 ; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %r11
548 ; AVX2-FP-NEXT: vmovaps (%rdi), %ymm0
549 ; AVX2-FP-NEXT: vmovaps (%rsi), %ymm1
550 ; AVX2-FP-NEXT: vmovaps (%rdx), %ymm2
551 ; AVX2-FP-NEXT: vmovaps (%rcx), %ymm3
552 ; AVX2-FP-NEXT: vmovaps (%r8), %ymm4
553 ; AVX2-FP-NEXT: vmovaps (%r9), %ymm5
554 ; AVX2-FP-NEXT: vmovaps (%r11), %ymm6
555 ; AVX2-FP-NEXT: vmovaps (%r10), %ymm7
556 ; AVX2-FP-NEXT: vmovaps (%r9), %xmm8
557 ; AVX2-FP-NEXT: vinsertf128 $1, (%r10), %ymm8, %ymm8
558 ; AVX2-FP-NEXT: vmovaps (%r8), %xmm9
559 ; AVX2-FP-NEXT: vinsertf128 $1, (%r11), %ymm9, %ymm9
560 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm10 = ymm9[0],ymm8[0],ymm9[2],ymm8[2]
561 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm8 = ymm9[1],ymm8[1],ymm9[3],ymm8[3]
562 ; AVX2-FP-NEXT: vmovaps (%rsi), %xmm9
563 ; AVX2-FP-NEXT: vinsertf128 $1, (%rcx), %ymm9, %ymm9
564 ; AVX2-FP-NEXT: vmovaps (%rdi), %xmm11
565 ; AVX2-FP-NEXT: vinsertf128 $1, (%rdx), %ymm11, %ymm11
566 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm12 = ymm11[0],ymm9[0],ymm11[2],ymm9[2]
567 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm9 = ymm11[1],ymm9[1],ymm11[3],ymm9[3]
568 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm11 = ymm6[0],ymm7[0],ymm6[2],ymm7[2]
569 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm13 = ymm4[0],ymm5[0],ymm4[2],ymm5[2]
570 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm11 = ymm13[2,3],ymm11[2,3]
571 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm6 = ymm6[1],ymm7[1],ymm6[3],ymm7[3]
572 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm4 = ymm4[1],ymm5[1],ymm4[3],ymm5[3]
573 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],ymm6[2,3]
574 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm5 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
575 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm6 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
576 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm5 = ymm6[2,3],ymm5[2,3]
577 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
578 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
579 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm2[2,3]
580 ; AVX2-FP-NEXT: vmovaps %ymm0, 128(%rax)
581 ; AVX2-FP-NEXT: vmovaps %ymm5, 192(%rax)
582 ; AVX2-FP-NEXT: vmovaps %ymm4, 224(%rax)
583 ; AVX2-FP-NEXT: vmovaps %ymm11, 160(%rax)
584 ; AVX2-FP-NEXT: vmovaps %ymm9, 64(%rax)
585 ; AVX2-FP-NEXT: vmovaps %ymm12, (%rax)
586 ; AVX2-FP-NEXT: vmovaps %ymm8, 96(%rax)
587 ; AVX2-FP-NEXT: vmovaps %ymm10, 32(%rax)
588 ; AVX2-FP-NEXT: vzeroupper
589 ; AVX2-FP-NEXT: retq
590 ;
591 ; AVX2-FCP-LABEL: store_i64_stride8_vf4:
592 ; AVX2-FCP: # %bb.0:
593 ; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
594 ; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10
595 ; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r11
596 ; AVX2-FCP-NEXT: vmovaps (%rdi), %ymm0
597 ; AVX2-FCP-NEXT: vmovaps (%rsi), %ymm1
598 ; AVX2-FCP-NEXT: vmovaps (%rdx), %ymm2
599 ; AVX2-FCP-NEXT: vmovaps (%rcx), %ymm3
600 ; AVX2-FCP-NEXT: vmovaps (%r8), %ymm4
601 ; AVX2-FCP-NEXT: vmovaps (%r9), %ymm5
602 ; AVX2-FCP-NEXT: vmovaps (%r11), %ymm6
603 ; AVX2-FCP-NEXT: vmovaps (%r10), %ymm7
604 ; AVX2-FCP-NEXT: vmovaps (%r9), %xmm8
605 ; AVX2-FCP-NEXT: vinsertf128 $1, (%r10), %ymm8, %ymm8
606 ; AVX2-FCP-NEXT: vmovaps (%r8), %xmm9
607 ; AVX2-FCP-NEXT: vinsertf128 $1, (%r11), %ymm9, %ymm9
608 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm10 = ymm9[0],ymm8[0],ymm9[2],ymm8[2]
609 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm8 = ymm9[1],ymm8[1],ymm9[3],ymm8[3]
610 ; AVX2-FCP-NEXT: vmovaps (%rsi), %xmm9
611 ; AVX2-FCP-NEXT: vinsertf128 $1, (%rcx), %ymm9, %ymm9
612 ; AVX2-FCP-NEXT: vmovaps (%rdi), %xmm11
613 ; AVX2-FCP-NEXT: vinsertf128 $1, (%rdx), %ymm11, %ymm11
614 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm12 = ymm11[0],ymm9[0],ymm11[2],ymm9[2]
615 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm9 = ymm11[1],ymm9[1],ymm11[3],ymm9[3]
616 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm11 = ymm6[0],ymm7[0],ymm6[2],ymm7[2]
617 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm13 = ymm4[0],ymm5[0],ymm4[2],ymm5[2]
618 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm11 = ymm13[2,3],ymm11[2,3]
619 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm6 = ymm6[1],ymm7[1],ymm6[3],ymm7[3]
620 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm4 = ymm4[1],ymm5[1],ymm4[3],ymm5[3]
621 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],ymm6[2,3]
622 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm5 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
623 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm6 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
624 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm5 = ymm6[2,3],ymm5[2,3]
625 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
626 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
627 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm2[2,3]
628 ; AVX2-FCP-NEXT: vmovaps %ymm0, 128(%rax)
629 ; AVX2-FCP-NEXT: vmovaps %ymm5, 192(%rax)
630 ; AVX2-FCP-NEXT: vmovaps %ymm4, 224(%rax)
631 ; AVX2-FCP-NEXT: vmovaps %ymm11, 160(%rax)
632 ; AVX2-FCP-NEXT: vmovaps %ymm9, 64(%rax)
633 ; AVX2-FCP-NEXT: vmovaps %ymm12, (%rax)
634 ; AVX2-FCP-NEXT: vmovaps %ymm8, 96(%rax)
635 ; AVX2-FCP-NEXT: vmovaps %ymm10, 32(%rax)
636 ; AVX2-FCP-NEXT: vzeroupper
637 ; AVX2-FCP-NEXT: retq
638 ;
639 ; AVX512-LABEL: store_i64_stride8_vf4:
640 ; AVX512: # %bb.0:
641 ; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %rax
642 ; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %r10
643 ; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %r11
644 ; AVX512-NEXT: vmovdqa (%rdi), %ymm0
645 ; AVX512-NEXT: vmovdqa (%rdx), %ymm1
646 ; AVX512-NEXT: vmovdqa (%r8), %ymm2
647 ; AVX512-NEXT: vmovdqa (%r11), %ymm3
648 ; AVX512-NEXT: vinserti64x4 $1, (%rsi), %zmm0, %zmm0
649 ; AVX512-NEXT: vinserti64x4 $1, (%rcx), %zmm1, %zmm1
650 ; AVX512-NEXT: vinserti64x4 $1, (%r9), %zmm2, %zmm2
651 ; AVX512-NEXT: vinserti64x4 $1, (%r10), %zmm3, %zmm3
652 ; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm4 = [0,4,8,12,0,4,8,12]
653 ; AVX512-NEXT: # zmm4 = mem[0,1,2,3,0,1,2,3]
654 ; AVX512-NEXT: vmovdqa64 %zmm2, %zmm5
655 ; AVX512-NEXT: vpermt2q %zmm3, %zmm4, %zmm5
656 ; AVX512-NEXT: vpermi2q %zmm1, %zmm0, %zmm4
657 ; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm4 = zmm4[0,1,2,3],zmm5[4,5,6,7]
658 ; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm5 = [1,5,9,13,1,5,9,13]
659 ; AVX512-NEXT: # zmm5 = mem[0,1,2,3,0,1,2,3]
660 ; AVX512-NEXT: vmovdqa64 %zmm2, %zmm6
661 ; AVX512-NEXT: vpermt2q %zmm3, %zmm5, %zmm6
662 ; AVX512-NEXT: vpermi2q %zmm1, %zmm0, %zmm5
663 ; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm5 = zmm5[0,1,2,3],zmm6[4,5,6,7]
664 ; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = [2,6,10,14,2,6,10,14]
665 ; AVX512-NEXT: # zmm6 = mem[0,1,2,3,0,1,2,3]
666 ; AVX512-NEXT: vmovdqa64 %zmm2, %zmm7
667 ; AVX512-NEXT: vpermt2q %zmm3, %zmm6, %zmm7
668 ; AVX512-NEXT: vpermi2q %zmm1, %zmm0, %zmm6
669 ; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm6 = zmm6[0,1,2,3],zmm7[4,5,6,7]
670 ; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [3,7,11,15,3,7,11,15]
671 ; AVX512-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]
672 ; AVX512-NEXT: vpermt2q %zmm3, %zmm7, %zmm2
673 ; AVX512-NEXT: vpermt2q %zmm1, %zmm7, %zmm0
674 ; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm2[4,5,6,7]
675 ; AVX512-NEXT: vmovdqa64 %zmm0, 192(%rax)
676 ; AVX512-NEXT: vmovdqa64 %zmm6, 128(%rax)
677 ; AVX512-NEXT: vmovdqa64 %zmm5, 64(%rax)
678 ; AVX512-NEXT: vmovdqa64 %zmm4, (%rax)
679 ; AVX512-NEXT: vzeroupper
680 ; AVX512-NEXT: retq
681 ;
682 ; AVX512-FCP-LABEL: store_i64_stride8_vf4:
683 ; AVX512-FCP: # %bb.0:
684 ; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
685 ; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10
686 ; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r11
687 ; AVX512-FCP-NEXT: vmovdqa (%rdi), %ymm0
688 ; AVX512-FCP-NEXT: vmovdqa (%rdx), %ymm1
689 ; AVX512-FCP-NEXT: vmovdqa (%r8), %ymm2
690 ; AVX512-FCP-NEXT: vmovdqa (%r11), %ymm3
691 ; AVX512-FCP-NEXT: vinserti64x4 $1, (%rsi), %zmm0, %zmm0
692 ; AVX512-FCP-NEXT: vinserti64x4 $1, (%rcx), %zmm1, %zmm1
693 ; AVX512-FCP-NEXT: vinserti64x4 $1, (%r9), %zmm2, %zmm2
694 ; AVX512-FCP-NEXT: vinserti64x4 $1, (%r10), %zmm3, %zmm3
695 ; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm4 = [0,4,8,12,0,4,8,12]
696 ; AVX512-FCP-NEXT: # zmm4 = mem[0,1,2,3,0,1,2,3]
697 ; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm5
698 ; AVX512-FCP-NEXT: vpermt2q %zmm3, %zmm4, %zmm5
699 ; AVX512-FCP-NEXT: vpermi2q %zmm1, %zmm0, %zmm4
700 ; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm4 = zmm4[0,1,2,3],zmm5[4,5,6,7]
701 ; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm5 = [1,5,9,13,1,5,9,13]
702 ; AVX512-FCP-NEXT: # zmm5 = mem[0,1,2,3,0,1,2,3]
703 ; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm6
704 ; AVX512-FCP-NEXT: vpermt2q %zmm3, %zmm5, %zmm6
705 ; AVX512-FCP-NEXT: vpermi2q %zmm1, %zmm0, %zmm5
706 ; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm5 = zmm5[0,1,2,3],zmm6[4,5,6,7]
707 ; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = [2,6,10,14,2,6,10,14]
708 ; AVX512-FCP-NEXT: # zmm6 = mem[0,1,2,3,0,1,2,3]
709 ; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm7
710 ; AVX512-FCP-NEXT: vpermt2q %zmm3, %zmm6, %zmm7
711 ; AVX512-FCP-NEXT: vpermi2q %zmm1, %zmm0, %zmm6
712 ; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm6 = zmm6[0,1,2,3],zmm7[4,5,6,7]
713 ; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [3,7,11,15,3,7,11,15]
714 ; AVX512-FCP-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]
715 ; AVX512-FCP-NEXT: vpermt2q %zmm3, %zmm7, %zmm2
716 ; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm7, %zmm0
717 ; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm2[4,5,6,7]
718 ; AVX512-FCP-NEXT: vmovdqa64 %zmm0, 192(%rax)
719 ; AVX512-FCP-NEXT: vmovdqa64 %zmm6, 128(%rax)
720 ; AVX512-FCP-NEXT: vmovdqa64 %zmm5, 64(%rax)
721 ; AVX512-FCP-NEXT: vmovdqa64 %zmm4, (%rax)
722 ; AVX512-FCP-NEXT: vzeroupper
723 ; AVX512-FCP-NEXT: retq
724 ;
725 ; AVX512DQ-LABEL: store_i64_stride8_vf4:
726 ; AVX512DQ: # %bb.0:
727 ; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %rax
728 ; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %r10
729 ; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %r11
730 ; AVX512DQ-NEXT: vmovdqa (%rdi), %ymm0
731 ; AVX512DQ-NEXT: vmovdqa (%rdx), %ymm1
732 ; AVX512DQ-NEXT: vmovdqa (%r8), %ymm2
733 ; AVX512DQ-NEXT: vmovdqa (%r11), %ymm3
734 ; AVX512DQ-NEXT: vinserti64x4 $1, (%rsi), %zmm0, %zmm0
735 ; AVX512DQ-NEXT: vinserti64x4 $1, (%rcx), %zmm1, %zmm1
736 ; AVX512DQ-NEXT: vinserti64x4 $1, (%r9), %zmm2, %zmm2
737 ; AVX512DQ-NEXT: vinserti64x4 $1, (%r10), %zmm3, %zmm3
738 ; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm4 = [0,4,8,12,0,4,8,12]
739 ; AVX512DQ-NEXT: # zmm4 = mem[0,1,2,3,0,1,2,3]
740 ; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm5
741 ; AVX512DQ-NEXT: vpermt2q %zmm3, %zmm4, %zmm5
742 ; AVX512DQ-NEXT: vpermi2q %zmm1, %zmm0, %zmm4
743 ; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm4 = zmm4[0,1,2,3],zmm5[4,5,6,7]
744 ; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm5 = [1,5,9,13,1,5,9,13]
745 ; AVX512DQ-NEXT: # zmm5 = mem[0,1,2,3,0,1,2,3]
746 ; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm6
747 ; AVX512DQ-NEXT: vpermt2q %zmm3, %zmm5, %zmm6
748 ; AVX512DQ-NEXT: vpermi2q %zmm1, %zmm0, %zmm5
749 ; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm5 = zmm5[0,1,2,3],zmm6[4,5,6,7]
750 ; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = [2,6,10,14,2,6,10,14]
751 ; AVX512DQ-NEXT: # zmm6 = mem[0,1,2,3,0,1,2,3]
752 ; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm7
753 ; AVX512DQ-NEXT: vpermt2q %zmm3, %zmm6, %zmm7
754 ; AVX512DQ-NEXT: vpermi2q %zmm1, %zmm0, %zmm6
755 ; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm6 = zmm6[0,1,2,3],zmm7[4,5,6,7]
756 ; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [3,7,11,15,3,7,11,15]
757 ; AVX512DQ-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]
758 ; AVX512DQ-NEXT: vpermt2q %zmm3, %zmm7, %zmm2
759 ; AVX512DQ-NEXT: vpermt2q %zmm1, %zmm7, %zmm0
760 ; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm2[4,5,6,7]
761 ; AVX512DQ-NEXT: vmovdqa64 %zmm0, 192(%rax)
762 ; AVX512DQ-NEXT: vmovdqa64 %zmm6, 128(%rax)
763 ; AVX512DQ-NEXT: vmovdqa64 %zmm5, 64(%rax)
764 ; AVX512DQ-NEXT: vmovdqa64 %zmm4, (%rax)
765 ; AVX512DQ-NEXT: vzeroupper
766 ; AVX512DQ-NEXT: retq
767 ;
768 ; AVX512DQ-FCP-LABEL: store_i64_stride8_vf4:
769 ; AVX512DQ-FCP: # %bb.0:
770 ; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
771 ; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10
772 ; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r11
773 ; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %ymm0
774 ; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %ymm1
775 ; AVX512DQ-FCP-NEXT: vmovdqa (%r8), %ymm2
776 ; AVX512DQ-FCP-NEXT: vmovdqa (%r11), %ymm3
777 ; AVX512DQ-FCP-NEXT: vinserti64x4 $1, (%rsi), %zmm0, %zmm0
778 ; AVX512DQ-FCP-NEXT: vinserti64x4 $1, (%rcx), %zmm1, %zmm1
779 ; AVX512DQ-FCP-NEXT: vinserti64x4 $1, (%r9), %zmm2, %zmm2
780 ; AVX512DQ-FCP-NEXT: vinserti64x4 $1, (%r10), %zmm3, %zmm3
781 ; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm4 = [0,4,8,12,0,4,8,12]
782 ; AVX512DQ-FCP-NEXT: # zmm4 = mem[0,1,2,3,0,1,2,3]
783 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm5
784 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm3, %zmm4, %zmm5
785 ; AVX512DQ-FCP-NEXT: vpermi2q %zmm1, %zmm0, %zmm4
786 ; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm4 = zmm4[0,1,2,3],zmm5[4,5,6,7]
787 ; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm5 = [1,5,9,13,1,5,9,13]
788 ; AVX512DQ-FCP-NEXT: # zmm5 = mem[0,1,2,3,0,1,2,3]
789 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm6
790 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm3, %zmm5, %zmm6
791 ; AVX512DQ-FCP-NEXT: vpermi2q %zmm1, %zmm0, %zmm5
792 ; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm5 = zmm5[0,1,2,3],zmm6[4,5,6,7]
793 ; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = [2,6,10,14,2,6,10,14]
794 ; AVX512DQ-FCP-NEXT: # zmm6 = mem[0,1,2,3,0,1,2,3]
795 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm7
796 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm3, %zmm6, %zmm7
797 ; AVX512DQ-FCP-NEXT: vpermi2q %zmm1, %zmm0, %zmm6
798 ; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm6 = zmm6[0,1,2,3],zmm7[4,5,6,7]
799 ; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [3,7,11,15,3,7,11,15]
800 ; AVX512DQ-FCP-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]
801 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm3, %zmm7, %zmm2
802 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm7, %zmm0
803 ; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm2[4,5,6,7]
804 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, 192(%rax)
805 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, 128(%rax)
806 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm5, 64(%rax)
807 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm4, (%rax)
808 ; AVX512DQ-FCP-NEXT: vzeroupper
809 ; AVX512DQ-FCP-NEXT: retq
810 ;
811 ; AVX512BW-LABEL: store_i64_stride8_vf4:
812 ; AVX512BW: # %bb.0:
813 ; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
814 ; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %r10
815 ; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %r11
816 ; AVX512BW-NEXT: vmovdqa (%rdi), %ymm0
817 ; AVX512BW-NEXT: vmovdqa (%rdx), %ymm1
818 ; AVX512BW-NEXT: vmovdqa (%r8), %ymm2
819 ; AVX512BW-NEXT: vmovdqa (%r11), %ymm3
820 ; AVX512BW-NEXT: vinserti64x4 $1, (%rsi), %zmm0, %zmm0
821 ; AVX512BW-NEXT: vinserti64x4 $1, (%rcx), %zmm1, %zmm1
822 ; AVX512BW-NEXT: vinserti64x4 $1, (%r9), %zmm2, %zmm2
823 ; AVX512BW-NEXT: vinserti64x4 $1, (%r10), %zmm3, %zmm3
824 ; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm4 = [0,4,8,12,0,4,8,12]
825 ; AVX512BW-NEXT: # zmm4 = mem[0,1,2,3,0,1,2,3]
826 ; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm5
827 ; AVX512BW-NEXT: vpermt2q %zmm3, %zmm4, %zmm5
828 ; AVX512BW-NEXT: vpermi2q %zmm1, %zmm0, %zmm4
829 ; AVX512BW-NEXT: vshufi64x2 {{.*#+}} zmm4 = zmm4[0,1,2,3],zmm5[4,5,6,7]
830 ; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm5 = [1,5,9,13,1,5,9,13]
831 ; AVX512BW-NEXT: # zmm5 = mem[0,1,2,3,0,1,2,3]
832 ; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm6
833 ; AVX512BW-NEXT: vpermt2q %zmm3, %zmm5, %zmm6
834 ; AVX512BW-NEXT: vpermi2q %zmm1, %zmm0, %zmm5
835 ; AVX512BW-NEXT: vshufi64x2 {{.*#+}} zmm5 = zmm5[0,1,2,3],zmm6[4,5,6,7]
836 ; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = [2,6,10,14,2,6,10,14]
837 ; AVX512BW-NEXT: # zmm6 = mem[0,1,2,3,0,1,2,3]
838 ; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm7
839 ; AVX512BW-NEXT: vpermt2q %zmm3, %zmm6, %zmm7
840 ; AVX512BW-NEXT: vpermi2q %zmm1, %zmm0, %zmm6
841 ; AVX512BW-NEXT: vshufi64x2 {{.*#+}} zmm6 = zmm6[0,1,2,3],zmm7[4,5,6,7]
842 ; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [3,7,11,15,3,7,11,15]
843 ; AVX512BW-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]
844 ; AVX512BW-NEXT: vpermt2q %zmm3, %zmm7, %zmm2
845 ; AVX512BW-NEXT: vpermt2q %zmm1, %zmm7, %zmm0
846 ; AVX512BW-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm2[4,5,6,7]
847 ; AVX512BW-NEXT: vmovdqa64 %zmm0, 192(%rax)
848 ; AVX512BW-NEXT: vmovdqa64 %zmm6, 128(%rax)
849 ; AVX512BW-NEXT: vmovdqa64 %zmm5, 64(%rax)
850 ; AVX512BW-NEXT: vmovdqa64 %zmm4, (%rax)
851 ; AVX512BW-NEXT: vzeroupper
852 ; AVX512BW-NEXT: retq
853 ;
854 ; AVX512BW-FCP-LABEL: store_i64_stride8_vf4:
855 ; AVX512BW-FCP: # %bb.0:
856 ; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
857 ; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10
858 ; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r11
859 ; AVX512BW-FCP-NEXT: vmovdqa (%rdi), %ymm0
860 ; AVX512BW-FCP-NEXT: vmovdqa (%rdx), %ymm1
861 ; AVX512BW-FCP-NEXT: vmovdqa (%r8), %ymm2
862 ; AVX512BW-FCP-NEXT: vmovdqa (%r11), %ymm3
863 ; AVX512BW-FCP-NEXT: vinserti64x4 $1, (%rsi), %zmm0, %zmm0
864 ; AVX512BW-FCP-NEXT: vinserti64x4 $1, (%rcx), %zmm1, %zmm1
865 ; AVX512BW-FCP-NEXT: vinserti64x4 $1, (%r9), %zmm2, %zmm2
866 ; AVX512BW-FCP-NEXT: vinserti64x4 $1, (%r10), %zmm3, %zmm3
867 ; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm4 = [0,4,8,12,0,4,8,12]
868 ; AVX512BW-FCP-NEXT: # zmm4 = mem[0,1,2,3,0,1,2,3]
869 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm5
870 ; AVX512BW-FCP-NEXT: vpermt2q %zmm3, %zmm4, %zmm5
871 ; AVX512BW-FCP-NEXT: vpermi2q %zmm1, %zmm0, %zmm4
872 ; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm4 = zmm4[0,1,2,3],zmm5[4,5,6,7]
873 ; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm5 = [1,5,9,13,1,5,9,13]
874 ; AVX512BW-FCP-NEXT: # zmm5 = mem[0,1,2,3,0,1,2,3]
875 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm6
876 ; AVX512BW-FCP-NEXT: vpermt2q %zmm3, %zmm5, %zmm6
877 ; AVX512BW-FCP-NEXT: vpermi2q %zmm1, %zmm0, %zmm5
878 ; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm5 = zmm5[0,1,2,3],zmm6[4,5,6,7]
879 ; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = [2,6,10,14,2,6,10,14]
880 ; AVX512BW-FCP-NEXT: # zmm6 = mem[0,1,2,3,0,1,2,3]
881 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm7
882 ; AVX512BW-FCP-NEXT: vpermt2q %zmm3, %zmm6, %zmm7
883 ; AVX512BW-FCP-NEXT: vpermi2q %zmm1, %zmm0, %zmm6
884 ; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm6 = zmm6[0,1,2,3],zmm7[4,5,6,7]
885 ; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [3,7,11,15,3,7,11,15]
886 ; AVX512BW-FCP-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]
887 ; AVX512BW-FCP-NEXT: vpermt2q %zmm3, %zmm7, %zmm2
888 ; AVX512BW-FCP-NEXT: vpermt2q %zmm1, %zmm7, %zmm0
889 ; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm2[4,5,6,7]
890 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, 192(%rax)
891 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm6, 128(%rax)
892 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm5, 64(%rax)
893 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm4, (%rax)
894 ; AVX512BW-FCP-NEXT: vzeroupper
895 ; AVX512BW-FCP-NEXT: retq
896 ;
897 ; AVX512DQ-BW-LABEL: store_i64_stride8_vf4:
898 ; AVX512DQ-BW: # %bb.0:
899 ; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
900 ; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %r10
901 ; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %r11
902 ; AVX512DQ-BW-NEXT: vmovdqa (%rdi), %ymm0
903 ; AVX512DQ-BW-NEXT: vmovdqa (%rdx), %ymm1
904 ; AVX512DQ-BW-NEXT: vmovdqa (%r8), %ymm2
905 ; AVX512DQ-BW-NEXT: vmovdqa (%r11), %ymm3
906 ; AVX512DQ-BW-NEXT: vinserti64x4 $1, (%rsi), %zmm0, %zmm0
907 ; AVX512DQ-BW-NEXT: vinserti64x4 $1, (%rcx), %zmm1, %zmm1
908 ; AVX512DQ-BW-NEXT: vinserti64x4 $1, (%r9), %zmm2, %zmm2
909 ; AVX512DQ-BW-NEXT: vinserti64x4 $1, (%r10), %zmm3, %zmm3
910 ; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm4 = [0,4,8,12,0,4,8,12]
911 ; AVX512DQ-BW-NEXT: # zmm4 = mem[0,1,2,3,0,1,2,3]
912 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm5
913 ; AVX512DQ-BW-NEXT: vpermt2q %zmm3, %zmm4, %zmm5
914 ; AVX512DQ-BW-NEXT: vpermi2q %zmm1, %zmm0, %zmm4
915 ; AVX512DQ-BW-NEXT: vshufi64x2 {{.*#+}} zmm4 = zmm4[0,1,2,3],zmm5[4,5,6,7]
916 ; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm5 = [1,5,9,13,1,5,9,13]
917 ; AVX512DQ-BW-NEXT: # zmm5 = mem[0,1,2,3,0,1,2,3]
918 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm6
919 ; AVX512DQ-BW-NEXT: vpermt2q %zmm3, %zmm5, %zmm6
920 ; AVX512DQ-BW-NEXT: vpermi2q %zmm1, %zmm0, %zmm5
921 ; AVX512DQ-BW-NEXT: vshufi64x2 {{.*#+}} zmm5 = zmm5[0,1,2,3],zmm6[4,5,6,7]
922 ; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = [2,6,10,14,2,6,10,14]
923 ; AVX512DQ-BW-NEXT: # zmm6 = mem[0,1,2,3,0,1,2,3]
924 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm7
925 ; AVX512DQ-BW-NEXT: vpermt2q %zmm3, %zmm6, %zmm7
926 ; AVX512DQ-BW-NEXT: vpermi2q %zmm1, %zmm0, %zmm6
927 ; AVX512DQ-BW-NEXT: vshufi64x2 {{.*#+}} zmm6 = zmm6[0,1,2,3],zmm7[4,5,6,7]
928 ; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [3,7,11,15,3,7,11,15]
929 ; AVX512DQ-BW-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]
930 ; AVX512DQ-BW-NEXT: vpermt2q %zmm3, %zmm7, %zmm2
931 ; AVX512DQ-BW-NEXT: vpermt2q %zmm1, %zmm7, %zmm0
932 ; AVX512DQ-BW-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm2[4,5,6,7]
933 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, 192(%rax)
934 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm6, 128(%rax)
935 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm5, 64(%rax)
936 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm4, (%rax)
937 ; AVX512DQ-BW-NEXT: vzeroupper
938 ; AVX512DQ-BW-NEXT: retq
939 ;
940 ; AVX512DQ-BW-FCP-LABEL: store_i64_stride8_vf4:
941 ; AVX512DQ-BW-FCP: # %bb.0:
942 ; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
943 ; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10
944 ; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r11
945 ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdi), %ymm0
946 ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdx), %ymm1
947 ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%r8), %ymm2
948 ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%r11), %ymm3
949 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, (%rsi), %zmm0, %zmm0
950 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, (%rcx), %zmm1, %zmm1
951 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, (%r9), %zmm2, %zmm2
952 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, (%r10), %zmm3, %zmm3
953 ; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm4 = [0,4,8,12,0,4,8,12]
954 ; AVX512DQ-BW-FCP-NEXT: # zmm4 = mem[0,1,2,3,0,1,2,3]
955 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm5
956 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm3, %zmm4, %zmm5
957 ; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm1, %zmm0, %zmm4
958 ; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm4 = zmm4[0,1,2,3],zmm5[4,5,6,7]
959 ; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm5 = [1,5,9,13,1,5,9,13]
960 ; AVX512DQ-BW-FCP-NEXT: # zmm5 = mem[0,1,2,3,0,1,2,3]
961 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm6
962 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm3, %zmm5, %zmm6
963 ; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm1, %zmm0, %zmm5
964 ; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm5 = zmm5[0,1,2,3],zmm6[4,5,6,7]
965 ; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = [2,6,10,14,2,6,10,14]
966 ; AVX512DQ-BW-FCP-NEXT: # zmm6 = mem[0,1,2,3,0,1,2,3]
967 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm7
968 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm3, %zmm6, %zmm7
969 ; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm1, %zmm0, %zmm6
970 ; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm6 = zmm6[0,1,2,3],zmm7[4,5,6,7]
971 ; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [3,7,11,15,3,7,11,15]
972 ; AVX512DQ-BW-FCP-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]
973 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm3, %zmm7, %zmm2
974 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm1, %zmm7, %zmm0
975 ; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm2[4,5,6,7]
976 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, 192(%rax)
977 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm6, 128(%rax)
978 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm5, 64(%rax)
979 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm4, (%rax)
980 ; AVX512DQ-BW-FCP-NEXT: vzeroupper
981 ; AVX512DQ-BW-FCP-NEXT: retq
982 %in.vec0 = load <4 x i64>, ptr %in.vecptr0, align 64
983 %in.vec1 = load <4 x i64>, ptr %in.vecptr1, align 64
984 %in.vec2 = load <4 x i64>, ptr %in.vecptr2, align 64
985 %in.vec3 = load <4 x i64>, ptr %in.vecptr3, align 64
986 %in.vec4 = load <4 x i64>, ptr %in.vecptr4, align 64
987 %in.vec5 = load <4 x i64>, ptr %in.vecptr5, align 64
988 %in.vec6 = load <4 x i64>, ptr %in.vecptr6, align 64
989 %in.vec7 = load <4 x i64>, ptr %in.vecptr7, align 64
990 %1 = shufflevector <4 x i64> %in.vec0, <4 x i64> %in.vec1, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
991 %2 = shufflevector <4 x i64> %in.vec2, <4 x i64> %in.vec3, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
992 %3 = shufflevector <4 x i64> %in.vec4, <4 x i64> %in.vec5, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
993 %4 = shufflevector <4 x i64> %in.vec6, <4 x i64> %in.vec7, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
994 %5 = shufflevector <8 x i64> %1, <8 x i64> %2, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
995 %6 = shufflevector <8 x i64> %3, <8 x i64> %4, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
996 %7 = shufflevector <16 x i64> %5, <16 x i64> %6, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
997 %interleaved.vec = shufflevector <32 x i64> %7, <32 x i64> poison, <32 x i32> <i32 0, i32 4, i32 8, i32 12, i32 16, i32 20, i32 24, i32 28, i32 1, i32 5, i32 9, i32 13, i32 17, i32 21, i32 25, i32 29, i32 2, i32 6, i32 10, i32 14, i32 18, i32 22, i32 26, i32 30, i32 3, i32 7, i32 11, i32 15, i32 19, i32 23, i32 27, i32 31>
998 store <32 x i64> %interleaved.vec, ptr %out.vec, align 64
999 ret void
1000 }
1002 define void @store_i64_stride8_vf8(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecptr2, ptr %in.vecptr3, ptr %in.vecptr4, ptr %in.vecptr5, ptr %in.vecptr6, ptr %in.vecptr7, ptr %out.vec) nounwind {
1003 ; SSE-LABEL: store_i64_stride8_vf8:
1004 ; SSE: # %bb.0:
1005 ; SSE-NEXT: subq $152, %rsp
1006 ; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax
1007 ; SSE-NEXT: movq {{[0-9]+}}(%rsp), %r10
1008 ; SSE-NEXT: movaps (%rdi), %xmm7
1009 ; SSE-NEXT: movaps 16(%rdi), %xmm9
1010 ; SSE-NEXT: movaps (%rsi), %xmm3
1011 ; SSE-NEXT: movaps 16(%rsi), %xmm0
1012 ; SSE-NEXT: movaps (%rdx), %xmm8
1013 ; SSE-NEXT: movaps 16(%rdx), %xmm11
1014 ; SSE-NEXT: movaps (%rcx), %xmm4
1015 ; SSE-NEXT: movaps 16(%rcx), %xmm1
1016 ; SSE-NEXT: movaps (%r8), %xmm10
1017 ; SSE-NEXT: movaps 16(%r8), %xmm13
1018 ; SSE-NEXT: movaps (%r9), %xmm5
1019 ; SSE-NEXT: movaps 16(%r9), %xmm2
1020 ; SSE-NEXT: movaps (%r10), %xmm12
1021 ; SSE-NEXT: movaps 16(%r10), %xmm15
1022 ; SSE-NEXT: movaps (%rax), %xmm6
1023 ; SSE-NEXT: movaps %xmm7, %xmm14
1024 ; SSE-NEXT: movlhps {{.*#+}} xmm14 = xmm14[0],xmm3[0]
1025 ; SSE-NEXT: movaps %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
1026 ; SSE-NEXT: unpckhpd {{.*#+}} xmm7 = xmm7[1],xmm3[1]
1027 ; SSE-NEXT: movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
1028 ; SSE-NEXT: movaps %xmm8, %xmm3
1029 ; SSE-NEXT: movlhps {{.*#+}} xmm3 = xmm3[0],xmm4[0]
1030 ; SSE-NEXT: movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
1031 ; SSE-NEXT: unpckhpd {{.*#+}} xmm8 = xmm8[1],xmm4[1]
1032 ; SSE-NEXT: movaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
1033 ; SSE-NEXT: movaps %xmm10, %xmm4
1034 ; SSE-NEXT: movlhps {{.*#+}} xmm4 = xmm4[0],xmm5[0]
1035 ; SSE-NEXT: movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
1036 ; SSE-NEXT: unpckhpd {{.*#+}} xmm10 = xmm10[1],xmm5[1]
1037 ; SSE-NEXT: movaps %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
1038 ; SSE-NEXT: movaps %xmm12, %xmm4
1039 ; SSE-NEXT: movlhps {{.*#+}} xmm4 = xmm4[0],xmm6[0]
1040 ; SSE-NEXT: movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
1041 ; SSE-NEXT: unpckhpd {{.*#+}} xmm12 = xmm12[1],xmm6[1]
1042 ; SSE-NEXT: movaps %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
1043 ; SSE-NEXT: movaps %xmm9, %xmm3
1044 ; SSE-NEXT: movlhps {{.*#+}} xmm3 = xmm3[0],xmm0[0]
1045 ; SSE-NEXT: movaps %xmm3, (%rsp) # 16-byte Spill
1046 ; SSE-NEXT: unpckhpd {{.*#+}} xmm9 = xmm9[1],xmm0[1]
1047 ; SSE-NEXT: movaps %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
1048 ; SSE-NEXT: movaps %xmm11, %xmm0
1049 ; SSE-NEXT: movlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0]
1050 ; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
1051 ; SSE-NEXT: unpckhpd {{.*#+}} xmm11 = xmm11[1],xmm1[1]
1052 ; SSE-NEXT: movaps %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
1053 ; SSE-NEXT: movaps %xmm13, %xmm0
1054 ; SSE-NEXT: movlhps {{.*#+}} xmm0 = xmm0[0],xmm2[0]
1055 ; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
1056 ; SSE-NEXT: unpckhpd {{.*#+}} xmm13 = xmm13[1],xmm2[1]
1057 ; SSE-NEXT: movaps %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
1058 ; SSE-NEXT: movaps 16(%rax), %xmm0
1059 ; SSE-NEXT: movaps %xmm15, %xmm1
1060 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
1061 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
1062 ; SSE-NEXT: unpckhpd {{.*#+}} xmm15 = xmm15[1],xmm0[1]
1063 ; SSE-NEXT: movaps %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
1064 ; SSE-NEXT: movaps 32(%rdi), %xmm13
1065 ; SSE-NEXT: movaps 32(%rsi), %xmm0
1066 ; SSE-NEXT: movaps %xmm13, %xmm1
1067 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
1068 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
1069 ; SSE-NEXT: unpckhpd {{.*#+}} xmm13 = xmm13[1],xmm0[1]
1070 ; SSE-NEXT: movaps 32(%rdx), %xmm11
1071 ; SSE-NEXT: movaps 32(%rcx), %xmm0
1072 ; SSE-NEXT: movaps %xmm11, %xmm15
1073 ; SSE-NEXT: movlhps {{.*#+}} xmm15 = xmm15[0],xmm0[0]
1074 ; SSE-NEXT: unpckhpd {{.*#+}} xmm11 = xmm11[1],xmm0[1]
1075 ; SSE-NEXT: movaps 32(%r8), %xmm10
1076 ; SSE-NEXT: movaps 32(%r9), %xmm0
1077 ; SSE-NEXT: movaps %xmm10, %xmm14
1078 ; SSE-NEXT: movlhps {{.*#+}} xmm14 = xmm14[0],xmm0[0]
1079 ; SSE-NEXT: unpckhpd {{.*#+}} xmm10 = xmm10[1],xmm0[1]
1080 ; SSE-NEXT: movaps 32(%r10), %xmm8
1081 ; SSE-NEXT: movaps 32(%rax), %xmm1
1082 ; SSE-NEXT: movaps %xmm8, %xmm12
1083 ; SSE-NEXT: movlhps {{.*#+}} xmm12 = xmm12[0],xmm1[0]
1084 ; SSE-NEXT: unpckhpd {{.*#+}} xmm8 = xmm8[1],xmm1[1]
1085 ; SSE-NEXT: movaps 48(%rdi), %xmm6
1086 ; SSE-NEXT: movaps 48(%rsi), %xmm0
1087 ; SSE-NEXT: movaps %xmm6, %xmm9
1088 ; SSE-NEXT: movlhps {{.*#+}} xmm9 = xmm9[0],xmm0[0]
1089 ; SSE-NEXT: unpckhpd {{.*#+}} xmm6 = xmm6[1],xmm0[1]
1090 ; SSE-NEXT: movaps 48(%rdx), %xmm5
1091 ; SSE-NEXT: movaps 48(%rcx), %xmm1
1092 ; SSE-NEXT: movaps %xmm5, %xmm7
1093 ; SSE-NEXT: movlhps {{.*#+}} xmm7 = xmm7[0],xmm1[0]
1094 ; SSE-NEXT: unpckhpd {{.*#+}} xmm5 = xmm5[1],xmm1[1]
1095 ; SSE-NEXT: movaps 48(%r8), %xmm1
1096 ; SSE-NEXT: movaps 48(%r9), %xmm2
1097 ; SSE-NEXT: movaps %xmm1, %xmm4
1098 ; SSE-NEXT: movlhps {{.*#+}} xmm4 = xmm4[0],xmm2[0]
1099 ; SSE-NEXT: unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm2[1]
1100 ; SSE-NEXT: movaps 48(%r10), %xmm2
1101 ; SSE-NEXT: movaps 48(%rax), %xmm3
1102 ; SSE-NEXT: movaps %xmm2, %xmm0
1103 ; SSE-NEXT: movlhps {{.*#+}} xmm0 = xmm0[0],xmm3[0]
1104 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm3[1]
1105 ; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax
1106 ; SSE-NEXT: movaps %xmm2, 496(%rax)
1107 ; SSE-NEXT: movaps %xmm1, 480(%rax)
1108 ; SSE-NEXT: movaps %xmm5, 464(%rax)
1109 ; SSE-NEXT: movaps %xmm6, 448(%rax)
1110 ; SSE-NEXT: movaps %xmm0, 432(%rax)
1111 ; SSE-NEXT: movaps %xmm4, 416(%rax)
1112 ; SSE-NEXT: movaps %xmm7, 400(%rax)
1113 ; SSE-NEXT: movaps %xmm9, 384(%rax)
1114 ; SSE-NEXT: movaps %xmm8, 368(%rax)
1115 ; SSE-NEXT: movaps %xmm10, 352(%rax)
1116 ; SSE-NEXT: movaps %xmm11, 336(%rax)
1117 ; SSE-NEXT: movaps %xmm13, 320(%rax)
1118 ; SSE-NEXT: movaps %xmm12, 304(%rax)
1119 ; SSE-NEXT: movaps %xmm14, 288(%rax)
1120 ; SSE-NEXT: movaps %xmm15, 272(%rax)
1121 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
1122 ; SSE-NEXT: movaps %xmm0, 256(%rax)
1123 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
1124 ; SSE-NEXT: movaps %xmm0, 240(%rax)
1125 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
1126 ; SSE-NEXT: movaps %xmm0, 224(%rax)
1127 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
1128 ; SSE-NEXT: movaps %xmm0, 208(%rax)
1129 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
1130 ; SSE-NEXT: movaps %xmm0, 192(%rax)
1131 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
1132 ; SSE-NEXT: movaps %xmm0, 176(%rax)
1133 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
1134 ; SSE-NEXT: movaps %xmm0, 160(%rax)
1135 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
1136 ; SSE-NEXT: movaps %xmm0, 144(%rax)
1137 ; SSE-NEXT: movaps (%rsp), %xmm0 # 16-byte Reload
1138 ; SSE-NEXT: movaps %xmm0, 128(%rax)
1139 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
1140 ; SSE-NEXT: movaps %xmm0, 112(%rax)
1141 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
1142 ; SSE-NEXT: movaps %xmm0, 96(%rax)
1143 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
1144 ; SSE-NEXT: movaps %xmm0, 80(%rax)
1145 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
1146 ; SSE-NEXT: movaps %xmm0, 64(%rax)
1147 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
1148 ; SSE-NEXT: movaps %xmm0, 48(%rax)
1149 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
1150 ; SSE-NEXT: movaps %xmm0, 32(%rax)
1151 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
1152 ; SSE-NEXT: movaps %xmm0, 16(%rax)
1153 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
1154 ; SSE-NEXT: movaps %xmm0, (%rax)
1155 ; SSE-NEXT: addq $152, %rsp
1156 ; SSE-NEXT: retq
1157 ;
1158 ; AVX-LABEL: store_i64_stride8_vf8:
1159 ; AVX: # %bb.0:
1160 ; AVX-NEXT: pushq %rax
1161 ; AVX-NEXT: movq {{[0-9]+}}(%rsp), %r10
1162 ; AVX-NEXT: movq {{[0-9]+}}(%rsp), %rax
1163 ; AVX-NEXT: vmovaps 32(%r9), %xmm1
1164 ; AVX-NEXT: vmovaps 32(%r8), %xmm2
1165 ; AVX-NEXT: vmovlhps {{.*#+}} xmm0 = xmm2[0],xmm1[0]
1166 ; AVX-NEXT: vmovaps (%rax), %xmm5
1167 ; AVX-NEXT: vmovaps 32(%rax), %xmm3
1168 ; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm4
1169 ; AVX-NEXT: vinsertf128 $1, 32(%r10), %ymm0, %ymm0
1170 ; AVX-NEXT: vshufpd {{.*#+}} ymm0 = ymm0[0],ymm4[1],ymm0[2],ymm4[2]
1171 ; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
1172 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm2[1],xmm1[1]
1173 ; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm1, %ymm2
1174 ; AVX-NEXT: vbroadcastsd 40(%r10), %ymm3
1175 ; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm3[4,5,6,7]
1176 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm2[2,3],ymm1[4,5],ymm2[6,7]
1177 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
1178 ; AVX-NEXT: vmovaps 32(%rsi), %xmm3
1179 ; AVX-NEXT: vmovaps 32(%rdi), %xmm4
1180 ; AVX-NEXT: vmovlhps {{.*#+}} xmm2 = xmm4[0],xmm3[0]
1181 ; AVX-NEXT: vmovaps (%rcx), %xmm7
1182 ; AVX-NEXT: vmovaps 32(%rcx), %xmm6
1183 ; AVX-NEXT: vinsertf128 $1, %xmm6, %ymm2, %ymm8
1184 ; AVX-NEXT: vinsertf128 $1, 32(%rdx), %ymm2, %ymm2
1185 ; AVX-NEXT: vshufpd {{.*#+}} ymm0 = ymm2[0],ymm8[1],ymm2[2],ymm8[2]
1186 ; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
1187 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm3 = xmm4[1],xmm3[1]
1188 ; AVX-NEXT: vinsertf128 $1, %xmm6, %ymm3, %ymm4
1189 ; AVX-NEXT: vbroadcastsd 40(%rdx), %ymm6
1190 ; AVX-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm6[4,5,6,7]
1191 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm3[0,1],ymm4[2,3],ymm3[4,5],ymm4[6,7]
1192 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
1193 ; AVX-NEXT: vmovaps (%r9), %xmm6
1194 ; AVX-NEXT: vmovaps (%r8), %xmm8
1195 ; AVX-NEXT: vmovlhps {{.*#+}} xmm4 = xmm8[0],xmm6[0]
1196 ; AVX-NEXT: vinsertf128 $1, %xmm5, %ymm4, %ymm9
1197 ; AVX-NEXT: vinsertf128 $1, (%r10), %ymm4, %ymm4
1198 ; AVX-NEXT: vshufpd {{.*#+}} ymm4 = ymm4[0],ymm9[1],ymm4[2],ymm9[2]
1199 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm6 = xmm8[1],xmm6[1]
1200 ; AVX-NEXT: vinsertf128 $1, %xmm5, %ymm6, %ymm5
1201 ; AVX-NEXT: vbroadcastsd 8(%r10), %ymm8
1202 ; AVX-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm8[4,5,6,7]
1203 ; AVX-NEXT: vblendps {{.*#+}} ymm5 = ymm6[0,1],ymm5[2,3],ymm6[4,5],ymm5[6,7]
1204 ; AVX-NEXT: vmovaps (%rsi), %xmm8
1205 ; AVX-NEXT: vmovaps (%rdi), %xmm9
1206 ; AVX-NEXT: vmovlhps {{.*#+}} xmm6 = xmm9[0],xmm8[0]
1207 ; AVX-NEXT: vinsertf128 $1, %xmm7, %ymm6, %ymm10
1208 ; AVX-NEXT: vinsertf128 $1, (%rdx), %ymm6, %ymm6
1209 ; AVX-NEXT: vshufpd {{.*#+}} ymm6 = ymm6[0],ymm10[1],ymm6[2],ymm10[2]
1210 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm8 = xmm9[1],xmm8[1]
1211 ; AVX-NEXT: vinsertf128 $1, %xmm7, %ymm8, %ymm7
1212 ; AVX-NEXT: vbroadcastsd 8(%rdx), %ymm9
1213 ; AVX-NEXT: vblendps {{.*#+}} ymm8 = ymm8[0,1,2,3],ymm9[4,5,6,7]
1214 ; AVX-NEXT: vblendps {{.*#+}} ymm8 = ymm8[0,1],ymm7[2,3],ymm8[4,5],ymm7[6,7]
1215 ; AVX-NEXT: vmovaps 16(%r9), %xmm10
1216 ; AVX-NEXT: vmovaps 16(%r8), %xmm11
1217 ; AVX-NEXT: vmovlhps {{.*#+}} xmm7 = xmm11[0],xmm10[0]
1218 ; AVX-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0,1,2,3],mem[4,5,6,7]
1219 ; AVX-NEXT: vbroadcastsd 16(%rax), %ymm9
1220 ; AVX-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0,1,2,3,4,5],ymm9[6,7]
1221 ; AVX-NEXT: vmovaps 16(%rsi), %xmm12
1222 ; AVX-NEXT: vmovaps 16(%rdi), %xmm13
1223 ; AVX-NEXT: vmovlhps {{.*#+}} xmm9 = xmm13[0],xmm12[0]
1224 ; AVX-NEXT: vblendps {{.*#+}} ymm9 = ymm9[0,1,2,3],mem[4,5,6,7]
1225 ; AVX-NEXT: vbroadcastsd 16(%rcx), %ymm14
1226 ; AVX-NEXT: vblendps {{.*#+}} ymm9 = ymm9[0,1,2,3,4,5],ymm14[6,7]
1227 ; AVX-NEXT: vmovaps 48(%r9), %xmm14
1228 ; AVX-NEXT: vmovaps 48(%r8), %xmm15
1229 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm15[1],xmm14[1]
1230 ; AVX-NEXT: vbroadcastsd 56(%r10), %ymm1
1231 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm0[0,1,2,3],ymm1[4,5,6,7]
1232 ; AVX-NEXT: vmovlhps {{.*#+}} xmm1 = xmm15[0],xmm14[0]
1233 ; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],mem[4,5,6,7]
1234 ; AVX-NEXT: vbroadcastsd 48(%rax), %ymm14
1235 ; AVX-NEXT: vblendps {{.*#+}} ymm14 = ymm1[0,1,2,3,4,5],ymm14[6,7]
1236 ; AVX-NEXT: vmovaps 48(%rsi), %xmm1
1237 ; AVX-NEXT: vmovaps 48(%rdi), %xmm15
1238 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm15[1],xmm1[1]
1239 ; AVX-NEXT: vbroadcastsd 56(%rdx), %ymm3
1240 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm3[4,5,6,7]
1241 ; AVX-NEXT: vmovlhps {{.*#+}} xmm1 = xmm15[0],xmm1[0]
1242 ; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],mem[4,5,6,7]
1243 ; AVX-NEXT: vbroadcastsd 48(%rcx), %ymm3
1244 ; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5],ymm3[6,7]
1245 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm3 = xmm11[1],xmm10[1]
1246 ; AVX-NEXT: vbroadcastsd 24(%r10), %ymm10
1247 ; AVX-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm10[4,5,6,7]
1248 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm10 = xmm13[1],xmm12[1]
1249 ; AVX-NEXT: vbroadcastsd 24(%rdx), %ymm11
1250 ; AVX-NEXT: vblendps {{.*#+}} ymm10 = ymm10[0,1,2,3],ymm11[4,5,6,7]
1251 ; AVX-NEXT: movq {{[0-9]+}}(%rsp), %rdx
1252 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],mem[6,7]
1253 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],mem[6,7]
1254 ; AVX-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],mem[6,7]
1255 ; AVX-NEXT: vblendps {{.*#+}} ymm10 = ymm10[0,1,2,3,4,5],mem[6,7]
1256 ; AVX-NEXT: vmovaps %ymm10, 192(%rdx)
1257 ; AVX-NEXT: vmovaps %ymm3, 224(%rdx)
1258 ; AVX-NEXT: vmovaps %ymm8, 64(%rdx)
1259 ; AVX-NEXT: vmovapd %ymm6, (%rdx)
1260 ; AVX-NEXT: vmovaps %ymm1, 384(%rdx)
1261 ; AVX-NEXT: vmovaps %ymm0, 448(%rdx)
1262 ; AVX-NEXT: vmovaps %ymm5, 96(%rdx)
1263 ; AVX-NEXT: vmovapd %ymm4, 32(%rdx)
1264 ; AVX-NEXT: vmovaps %ymm14, 416(%rdx)
1265 ; AVX-NEXT: vmovaps %ymm2, 480(%rdx)
1266 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
1267 ; AVX-NEXT: vmovaps %ymm0, 320(%rdx)
1268 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
1269 ; AVX-NEXT: vmovaps %ymm0, 256(%rdx)
1270 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
1271 ; AVX-NEXT: vmovaps %ymm0, 352(%rdx)
1272 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
1273 ; AVX-NEXT: vmovaps %ymm0, 288(%rdx)
1274 ; AVX-NEXT: vmovaps %ymm9, 128(%rdx)
1275 ; AVX-NEXT: vmovaps %ymm7, 160(%rdx)
1276 ; AVX-NEXT: popq %rax
1277 ; AVX-NEXT: vzeroupper
1278 ; AVX-NEXT: retq
1279 ;
1280 ; AVX2-LABEL: store_i64_stride8_vf8:
1281 ; AVX2: # %bb.0:
1282 ; AVX2-NEXT: pushq %rax
1283 ; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %r10
1284 ; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %rax
1285 ; AVX2-NEXT: vmovaps (%rax), %xmm6
1286 ; AVX2-NEXT: vmovaps 32(%rax), %xmm3
1287 ; AVX2-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm0
1288 ; AVX2-NEXT: vmovaps (%r9), %xmm9
1289 ; AVX2-NEXT: vmovaps 32(%r9), %xmm7
1290 ; AVX2-NEXT: vmovaps (%r8), %xmm10
1291 ; AVX2-NEXT: vmovaps 32(%r8), %xmm8
1292 ; AVX2-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm8[1],xmm7[1]
1293 ; AVX2-NEXT: vbroadcastsd 40(%r10), %ymm2
1294 ; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
1295 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
1296 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
1297 ; AVX2-NEXT: vmovaps 32(%rcx), %xmm5
1298 ; AVX2-NEXT: vinsertf128 $1, %xmm5, %ymm0, %ymm1
1299 ; AVX2-NEXT: vmovaps (%rsi), %xmm11
1300 ; AVX2-NEXT: vmovaps 32(%rsi), %xmm13
1301 ; AVX2-NEXT: vmovaps (%rdi), %xmm12
1302 ; AVX2-NEXT: vmovaps 32(%rdi), %xmm14
1303 ; AVX2-NEXT: vunpckhpd {{.*#+}} xmm2 = xmm14[1],xmm13[1]
1304 ; AVX2-NEXT: vbroadcastsd 40(%rdx), %ymm4
1305 ; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm4[4,5,6,7]
1306 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3,4,5],ymm1[6,7]
1307 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
1308 ; AVX2-NEXT: vinsertf128 $1, %xmm6, %ymm0, %ymm2
1309 ; AVX2-NEXT: vunpckhpd {{.*#+}} xmm4 = xmm10[1],xmm9[1]
1310 ; AVX2-NEXT: vbroadcastsd 8(%r10), %ymm15
1311 ; AVX2-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm15[4,5,6,7]
1312 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm4[0,1,2,3,4,5],ymm2[6,7]
1313 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
1314 ; AVX2-NEXT: vunpckhpd {{.*#+}} xmm4 = xmm12[1],xmm11[1]
1315 ; AVX2-NEXT: vbroadcastsd 8(%rdx), %ymm15
1316 ; AVX2-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm15[4,5,6,7]
1317 ; AVX2-NEXT: vmovaps (%rcx), %xmm15
1318 ; AVX2-NEXT: vinsertf128 $1, %xmm15, %ymm0, %ymm0
1319 ; AVX2-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm0[6,7]
1320 ; AVX2-NEXT: vmovlhps {{.*#+}} xmm0 = xmm8[0],xmm7[0]
1321 ; AVX2-NEXT: vmovaps (%rdi), %ymm7
1322 ; AVX2-NEXT: vinsertf128 $1, 32(%r10), %ymm0, %ymm0
1323 ; AVX2-NEXT: vbroadcastsd %xmm3, %ymm3
1324 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm3[6,7]
1325 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
1326 ; AVX2-NEXT: vmovaps (%rsi), %ymm8
1327 ; AVX2-NEXT: vmovlhps {{.*#+}} xmm0 = xmm14[0],xmm13[0]
1328 ; AVX2-NEXT: vmovaps 32(%r8), %ymm1
1329 ; AVX2-NEXT: vinsertf128 $1, 32(%rdx), %ymm0, %ymm0
1330 ; AVX2-NEXT: vbroadcastsd %xmm5, %ymm5
1331 ; AVX2-NEXT: vblendps {{.*#+}} ymm5 = ymm0[0,1,2,3,4,5],ymm5[6,7]
1332 ; AVX2-NEXT: vmovaps (%r8), %ymm13
1333 ; AVX2-NEXT: vmovlhps {{.*#+}} xmm0 = xmm10[0],xmm9[0]
1334 ; AVX2-NEXT: vmovaps (%r9), %ymm14
1335 ; AVX2-NEXT: vinsertf128 $1, (%r10), %ymm0, %ymm0
1336 ; AVX2-NEXT: vbroadcastsd %xmm6, %ymm6
1337 ; AVX2-NEXT: vblendps {{.*#+}} ymm6 = ymm0[0,1,2,3,4,5],ymm6[6,7]
1338 ; AVX2-NEXT: vmovlhps {{.*#+}} xmm0 = xmm12[0],xmm11[0]
1339 ; AVX2-NEXT: vinsertf128 $1, (%rdx), %ymm0, %ymm0
1340 ; AVX2-NEXT: vbroadcastsd %xmm15, %ymm9
1341 ; AVX2-NEXT: vblendps {{.*#+}} ymm9 = ymm0[0,1,2,3,4,5],ymm9[6,7]
1342 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm13[0],ymm14[0],ymm13[2],ymm14[2]
1343 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],mem[2,3]
1344 ; AVX2-NEXT: vbroadcastsd 16(%rax), %ymm10
1345 ; AVX2-NEXT: vblendps {{.*#+}} ymm10 = ymm0[0,1,2,3,4,5],ymm10[6,7]
1346 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm7[0],ymm8[0],ymm7[2],ymm8[2]
1347 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],mem[2,3]
1348 ; AVX2-NEXT: vbroadcastsd 16(%rcx), %ymm11
1349 ; AVX2-NEXT: vblendps {{.*#+}} ymm11 = ymm0[0,1,2,3,4,5],ymm11[6,7]
1350 ; AVX2-NEXT: vmovaps 32(%r9), %ymm0
1351 ; AVX2-NEXT: vunpckhpd {{.*#+}} ymm12 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
1352 ; AVX2-NEXT: vbroadcastsd 56(%r10), %ymm15
1353 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm12 = ymm12[2,3],ymm15[2,3]
1354 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
1355 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],mem[2,3]
1356 ; AVX2-NEXT: vbroadcastsd 48(%rax), %ymm1
1357 ; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm0[0,1,2,3,4,5],ymm1[6,7]
1358 ; AVX2-NEXT: vmovaps 32(%rdi), %ymm1
1359 ; AVX2-NEXT: vmovaps 32(%rsi), %ymm15
1360 ; AVX2-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm15[1],ymm1[3],ymm15[3]
1361 ; AVX2-NEXT: vbroadcastsd 56(%rdx), %ymm3
1362 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm3[2,3]
1363 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm15[0],ymm1[2],ymm15[2]
1364 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm1[2,3],mem[2,3]
1365 ; AVX2-NEXT: vbroadcastsd 48(%rcx), %ymm3
1366 ; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5],ymm3[6,7]
1367 ; AVX2-NEXT: vunpckhpd {{.*#+}} ymm3 = ymm13[1],ymm14[1],ymm13[3],ymm14[3]
1368 ; AVX2-NEXT: vbroadcastsd 24(%r10), %ymm13
1369 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm3 = ymm3[2,3],ymm13[2,3]
1370 ; AVX2-NEXT: vunpckhpd {{.*#+}} ymm7 = ymm7[1],ymm8[1],ymm7[3],ymm8[3]
1371 ; AVX2-NEXT: vbroadcastsd 24(%rdx), %ymm8
1372 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm7 = ymm7[2,3],ymm8[2,3]
1373 ; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %rdx
1374 ; AVX2-NEXT: vblendps {{.*#+}} ymm8 = ymm12[0,1,2,3,4,5],mem[6,7]
1375 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],mem[6,7]
1376 ; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],mem[6,7]
1377 ; AVX2-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0,1,2,3,4,5],mem[6,7]
1378 ; AVX2-NEXT: vmovaps %ymm7, 192(%rdx)
1379 ; AVX2-NEXT: vmovaps %ymm3, 224(%rdx)
1380 ; AVX2-NEXT: vmovaps %ymm1, 384(%rdx)
1381 ; AVX2-NEXT: vmovaps %ymm0, 448(%rdx)
1382 ; AVX2-NEXT: vmovaps %ymm2, 416(%rdx)
1383 ; AVX2-NEXT: vmovaps %ymm8, 480(%rdx)
1384 ; AVX2-NEXT: vmovaps %ymm11, 128(%rdx)
1385 ; AVX2-NEXT: vmovaps %ymm10, 160(%rdx)
1386 ; AVX2-NEXT: vmovaps %ymm4, 64(%rdx)
1387 ; AVX2-NEXT: vmovaps %ymm9, (%rdx)
1388 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
1389 ; AVX2-NEXT: vmovaps %ymm0, 96(%rdx)
1390 ; AVX2-NEXT: vmovaps %ymm6, 32(%rdx)
1391 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
1392 ; AVX2-NEXT: vmovaps %ymm0, 320(%rdx)
1393 ; AVX2-NEXT: vmovaps %ymm5, 256(%rdx)
1394 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
1395 ; AVX2-NEXT: vmovaps %ymm0, 352(%rdx)
1396 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
1397 ; AVX2-NEXT: vmovaps %ymm0, 288(%rdx)
1398 ; AVX2-NEXT: popq %rax
1399 ; AVX2-NEXT: vzeroupper
1400 ; AVX2-NEXT: retq
1401 ;
1402 ; AVX2-FP-LABEL: store_i64_stride8_vf8:
1403 ; AVX2-FP: # %bb.0:
1404 ; AVX2-FP-NEXT: pushq %rax
1405 ; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %r10
1406 ; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %rax
1407 ; AVX2-FP-NEXT: vmovaps (%rax), %xmm6
1408 ; AVX2-FP-NEXT: vmovaps 32(%rax), %xmm3
1409 ; AVX2-FP-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm0
1410 ; AVX2-FP-NEXT: vmovaps (%r9), %xmm9
1411 ; AVX2-FP-NEXT: vmovaps 32(%r9), %xmm7
1412 ; AVX2-FP-NEXT: vmovaps (%r8), %xmm10
1413 ; AVX2-FP-NEXT: vmovaps 32(%r8), %xmm8
1414 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm8[1],xmm7[1]
1415 ; AVX2-FP-NEXT: vbroadcastsd 40(%r10), %ymm2
1416 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
1417 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
1418 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
1419 ; AVX2-FP-NEXT: vmovaps 32(%rcx), %xmm5
1420 ; AVX2-FP-NEXT: vinsertf128 $1, %xmm5, %ymm0, %ymm1
1421 ; AVX2-FP-NEXT: vmovaps (%rsi), %xmm11
1422 ; AVX2-FP-NEXT: vmovaps 32(%rsi), %xmm13
1423 ; AVX2-FP-NEXT: vmovaps (%rdi), %xmm12
1424 ; AVX2-FP-NEXT: vmovaps 32(%rdi), %xmm14
1425 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm2 = xmm14[1],xmm13[1]
1426 ; AVX2-FP-NEXT: vbroadcastsd 40(%rdx), %ymm4
1427 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm4[4,5,6,7]
1428 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3,4,5],ymm1[6,7]
1429 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
1430 ; AVX2-FP-NEXT: vinsertf128 $1, %xmm6, %ymm0, %ymm2
1431 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm4 = xmm10[1],xmm9[1]
1432 ; AVX2-FP-NEXT: vbroadcastsd 8(%r10), %ymm15
1433 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm15[4,5,6,7]
1434 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm4[0,1,2,3,4,5],ymm2[6,7]
1435 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
1436 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm4 = xmm12[1],xmm11[1]
1437 ; AVX2-FP-NEXT: vbroadcastsd 8(%rdx), %ymm15
1438 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm15[4,5,6,7]
1439 ; AVX2-FP-NEXT: vmovaps (%rcx), %xmm15
1440 ; AVX2-FP-NEXT: vinsertf128 $1, %xmm15, %ymm0, %ymm0
1441 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm0[6,7]
1442 ; AVX2-FP-NEXT: vmovlhps {{.*#+}} xmm0 = xmm8[0],xmm7[0]
1443 ; AVX2-FP-NEXT: vmovaps (%rdi), %ymm7
1444 ; AVX2-FP-NEXT: vinsertf128 $1, 32(%r10), %ymm0, %ymm0
1445 ; AVX2-FP-NEXT: vbroadcastsd %xmm3, %ymm3
1446 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm3[6,7]
1447 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
1448 ; AVX2-FP-NEXT: vmovaps (%rsi), %ymm8
1449 ; AVX2-FP-NEXT: vmovlhps {{.*#+}} xmm0 = xmm14[0],xmm13[0]
1450 ; AVX2-FP-NEXT: vmovaps 32(%r8), %ymm1
1451 ; AVX2-FP-NEXT: vinsertf128 $1, 32(%rdx), %ymm0, %ymm0
1452 ; AVX2-FP-NEXT: vbroadcastsd %xmm5, %ymm5
1453 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm5 = ymm0[0,1,2,3,4,5],ymm5[6,7]
1454 ; AVX2-FP-NEXT: vmovaps (%r8), %ymm13
1455 ; AVX2-FP-NEXT: vmovlhps {{.*#+}} xmm0 = xmm10[0],xmm9[0]
1456 ; AVX2-FP-NEXT: vmovaps (%r9), %ymm14
1457 ; AVX2-FP-NEXT: vinsertf128 $1, (%r10), %ymm0, %ymm0
1458 ; AVX2-FP-NEXT: vbroadcastsd %xmm6, %ymm6
1459 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm6 = ymm0[0,1,2,3,4,5],ymm6[6,7]
1460 ; AVX2-FP-NEXT: vmovlhps {{.*#+}} xmm0 = xmm12[0],xmm11[0]
1461 ; AVX2-FP-NEXT: vinsertf128 $1, (%rdx), %ymm0, %ymm0
1462 ; AVX2-FP-NEXT: vbroadcastsd %xmm15, %ymm9
1463 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm9 = ymm0[0,1,2,3,4,5],ymm9[6,7]
1464 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm13[0],ymm14[0],ymm13[2],ymm14[2]
1465 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],mem[2,3]
1466 ; AVX2-FP-NEXT: vbroadcastsd 16(%rax), %ymm10
1467 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm10 = ymm0[0,1,2,3,4,5],ymm10[6,7]
1468 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm7[0],ymm8[0],ymm7[2],ymm8[2]
1469 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],mem[2,3]
1470 ; AVX2-FP-NEXT: vbroadcastsd 16(%rcx), %ymm11
1471 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm11 = ymm0[0,1,2,3,4,5],ymm11[6,7]
1472 ; AVX2-FP-NEXT: vmovaps 32(%r9), %ymm0
1473 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm12 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
1474 ; AVX2-FP-NEXT: vbroadcastsd 56(%r10), %ymm15
1475 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm12 = ymm12[2,3],ymm15[2,3]
1476 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
1477 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],mem[2,3]
1478 ; AVX2-FP-NEXT: vbroadcastsd 48(%rax), %ymm1
1479 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm0[0,1,2,3,4,5],ymm1[6,7]
1480 ; AVX2-FP-NEXT: vmovaps 32(%rdi), %ymm1
1481 ; AVX2-FP-NEXT: vmovaps 32(%rsi), %ymm15
1482 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm15[1],ymm1[3],ymm15[3]
1483 ; AVX2-FP-NEXT: vbroadcastsd 56(%rdx), %ymm3
1484 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm3[2,3]
1485 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm15[0],ymm1[2],ymm15[2]
1486 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm1[2,3],mem[2,3]
1487 ; AVX2-FP-NEXT: vbroadcastsd 48(%rcx), %ymm3
1488 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5],ymm3[6,7]
1489 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm3 = ymm13[1],ymm14[1],ymm13[3],ymm14[3]
1490 ; AVX2-FP-NEXT: vbroadcastsd 24(%r10), %ymm13
1491 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm3 = ymm3[2,3],ymm13[2,3]
1492 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm7 = ymm7[1],ymm8[1],ymm7[3],ymm8[3]
1493 ; AVX2-FP-NEXT: vbroadcastsd 24(%rdx), %ymm8
1494 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm7 = ymm7[2,3],ymm8[2,3]
1495 ; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %rdx
1496 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm8 = ymm12[0,1,2,3,4,5],mem[6,7]
1497 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],mem[6,7]
1498 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],mem[6,7]
1499 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0,1,2,3,4,5],mem[6,7]
1500 ; AVX2-FP-NEXT: vmovaps %ymm7, 192(%rdx)
1501 ; AVX2-FP-NEXT: vmovaps %ymm3, 224(%rdx)
1502 ; AVX2-FP-NEXT: vmovaps %ymm1, 384(%rdx)
1503 ; AVX2-FP-NEXT: vmovaps %ymm0, 448(%rdx)
1504 ; AVX2-FP-NEXT: vmovaps %ymm2, 416(%rdx)
1505 ; AVX2-FP-NEXT: vmovaps %ymm8, 480(%rdx)
1506 ; AVX2-FP-NEXT: vmovaps %ymm11, 128(%rdx)
1507 ; AVX2-FP-NEXT: vmovaps %ymm10, 160(%rdx)
1508 ; AVX2-FP-NEXT: vmovaps %ymm4, 64(%rdx)
1509 ; AVX2-FP-NEXT: vmovaps %ymm9, (%rdx)
1510 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
1511 ; AVX2-FP-NEXT: vmovaps %ymm0, 96(%rdx)
1512 ; AVX2-FP-NEXT: vmovaps %ymm6, 32(%rdx)
1513 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
1514 ; AVX2-FP-NEXT: vmovaps %ymm0, 320(%rdx)
1515 ; AVX2-FP-NEXT: vmovaps %ymm5, 256(%rdx)
1516 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
1517 ; AVX2-FP-NEXT: vmovaps %ymm0, 352(%rdx)
1518 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
1519 ; AVX2-FP-NEXT: vmovaps %ymm0, 288(%rdx)
1520 ; AVX2-FP-NEXT: popq %rax
1521 ; AVX2-FP-NEXT: vzeroupper
1522 ; AVX2-FP-NEXT: retq
1523 ;
1524 ; AVX2-FCP-LABEL: store_i64_stride8_vf8:
1525 ; AVX2-FCP: # %bb.0:
1526 ; AVX2-FCP-NEXT: pushq %rax
1527 ; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10
1528 ; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
1529 ; AVX2-FCP-NEXT: vmovaps (%rax), %xmm6
1530 ; AVX2-FCP-NEXT: vmovaps 32(%rax), %xmm3
1531 ; AVX2-FCP-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm0
1532 ; AVX2-FCP-NEXT: vmovaps (%r9), %xmm9
1533 ; AVX2-FCP-NEXT: vmovaps 32(%r9), %xmm7
1534 ; AVX2-FCP-NEXT: vmovaps (%r8), %xmm10
1535 ; AVX2-FCP-NEXT: vmovaps 32(%r8), %xmm8
1536 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm8[1],xmm7[1]
1537 ; AVX2-FCP-NEXT: vbroadcastsd 40(%r10), %ymm2
1538 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
1539 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
1540 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
1541 ; AVX2-FCP-NEXT: vmovaps 32(%rcx), %xmm5
1542 ; AVX2-FCP-NEXT: vinsertf128 $1, %xmm5, %ymm0, %ymm1
1543 ; AVX2-FCP-NEXT: vmovaps (%rsi), %xmm11
1544 ; AVX2-FCP-NEXT: vmovaps 32(%rsi), %xmm13
1545 ; AVX2-FCP-NEXT: vmovaps (%rdi), %xmm12
1546 ; AVX2-FCP-NEXT: vmovaps 32(%rdi), %xmm14
1547 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm2 = xmm14[1],xmm13[1]
1548 ; AVX2-FCP-NEXT: vbroadcastsd 40(%rdx), %ymm4
1549 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm4[4,5,6,7]
1550 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3,4,5],ymm1[6,7]
1551 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
1552 ; AVX2-FCP-NEXT: vinsertf128 $1, %xmm6, %ymm0, %ymm2
1553 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm4 = xmm10[1],xmm9[1]
1554 ; AVX2-FCP-NEXT: vbroadcastsd 8(%r10), %ymm15
1555 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm15[4,5,6,7]
1556 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm4[0,1,2,3,4,5],ymm2[6,7]
1557 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
1558 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm4 = xmm12[1],xmm11[1]
1559 ; AVX2-FCP-NEXT: vbroadcastsd 8(%rdx), %ymm15
1560 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm15[4,5,6,7]
1561 ; AVX2-FCP-NEXT: vmovaps (%rcx), %xmm15
1562 ; AVX2-FCP-NEXT: vinsertf128 $1, %xmm15, %ymm0, %ymm0
1563 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm0[6,7]
1564 ; AVX2-FCP-NEXT: vmovlhps {{.*#+}} xmm0 = xmm8[0],xmm7[0]
1565 ; AVX2-FCP-NEXT: vmovaps (%rdi), %ymm7
1566 ; AVX2-FCP-NEXT: vinsertf128 $1, 32(%r10), %ymm0, %ymm0
1567 ; AVX2-FCP-NEXT: vbroadcastsd %xmm3, %ymm3
1568 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm3[6,7]
1569 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
1570 ; AVX2-FCP-NEXT: vmovaps (%rsi), %ymm8
1571 ; AVX2-FCP-NEXT: vmovlhps {{.*#+}} xmm0 = xmm14[0],xmm13[0]
1572 ; AVX2-FCP-NEXT: vmovaps 32(%r8), %ymm1
1573 ; AVX2-FCP-NEXT: vinsertf128 $1, 32(%rdx), %ymm0, %ymm0
1574 ; AVX2-FCP-NEXT: vbroadcastsd %xmm5, %ymm5
1575 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm5 = ymm0[0,1,2,3,4,5],ymm5[6,7]
1576 ; AVX2-FCP-NEXT: vmovaps (%r8), %ymm13
1577 ; AVX2-FCP-NEXT: vmovlhps {{.*#+}} xmm0 = xmm10[0],xmm9[0]
1578 ; AVX2-FCP-NEXT: vmovaps (%r9), %ymm14
1579 ; AVX2-FCP-NEXT: vinsertf128 $1, (%r10), %ymm0, %ymm0
1580 ; AVX2-FCP-NEXT: vbroadcastsd %xmm6, %ymm6
1581 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm6 = ymm0[0,1,2,3,4,5],ymm6[6,7]
1582 ; AVX2-FCP-NEXT: vmovlhps {{.*#+}} xmm0 = xmm12[0],xmm11[0]
1583 ; AVX2-FCP-NEXT: vinsertf128 $1, (%rdx), %ymm0, %ymm0
1584 ; AVX2-FCP-NEXT: vbroadcastsd %xmm15, %ymm9
1585 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm9 = ymm0[0,1,2,3,4,5],ymm9[6,7]
1586 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm13[0],ymm14[0],ymm13[2],ymm14[2]
1587 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],mem[2,3]
1588 ; AVX2-FCP-NEXT: vbroadcastsd 16(%rax), %ymm10
1589 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm10 = ymm0[0,1,2,3,4,5],ymm10[6,7]
1590 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm7[0],ymm8[0],ymm7[2],ymm8[2]
1591 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],mem[2,3]
1592 ; AVX2-FCP-NEXT: vbroadcastsd 16(%rcx), %ymm11
1593 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm11 = ymm0[0,1,2,3,4,5],ymm11[6,7]
1594 ; AVX2-FCP-NEXT: vmovaps 32(%r9), %ymm0
1595 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm12 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
1596 ; AVX2-FCP-NEXT: vbroadcastsd 56(%r10), %ymm15
1597 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm12 = ymm12[2,3],ymm15[2,3]
1598 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
1599 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],mem[2,3]
1600 ; AVX2-FCP-NEXT: vbroadcastsd 48(%rax), %ymm1
1601 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm0[0,1,2,3,4,5],ymm1[6,7]
1602 ; AVX2-FCP-NEXT: vmovaps 32(%rdi), %ymm1
1603 ; AVX2-FCP-NEXT: vmovaps 32(%rsi), %ymm15
1604 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm15[1],ymm1[3],ymm15[3]
1605 ; AVX2-FCP-NEXT: vbroadcastsd 56(%rdx), %ymm3
1606 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm3[2,3]
1607 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm1[0],ymm15[0],ymm1[2],ymm15[2]
1608 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm1[2,3],mem[2,3]
1609 ; AVX2-FCP-NEXT: vbroadcastsd 48(%rcx), %ymm3
1610 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5],ymm3[6,7]
1611 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm3 = ymm13[1],ymm14[1],ymm13[3],ymm14[3]
1612 ; AVX2-FCP-NEXT: vbroadcastsd 24(%r10), %ymm13
1613 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm3 = ymm3[2,3],ymm13[2,3]
1614 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm7 = ymm7[1],ymm8[1],ymm7[3],ymm8[3]
1615 ; AVX2-FCP-NEXT: vbroadcastsd 24(%rdx), %ymm8
1616 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm7 = ymm7[2,3],ymm8[2,3]
1617 ; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rdx
1618 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm8 = ymm12[0,1,2,3,4,5],mem[6,7]
1619 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],mem[6,7]
1620 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],mem[6,7]
1621 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0,1,2,3,4,5],mem[6,7]
1622 ; AVX2-FCP-NEXT: vmovaps %ymm7, 192(%rdx)
1623 ; AVX2-FCP-NEXT: vmovaps %ymm3, 224(%rdx)
1624 ; AVX2-FCP-NEXT: vmovaps %ymm1, 384(%rdx)
1625 ; AVX2-FCP-NEXT: vmovaps %ymm0, 448(%rdx)
1626 ; AVX2-FCP-NEXT: vmovaps %ymm2, 416(%rdx)
1627 ; AVX2-FCP-NEXT: vmovaps %ymm8, 480(%rdx)
1628 ; AVX2-FCP-NEXT: vmovaps %ymm11, 128(%rdx)
1629 ; AVX2-FCP-NEXT: vmovaps %ymm10, 160(%rdx)
1630 ; AVX2-FCP-NEXT: vmovaps %ymm4, 64(%rdx)
1631 ; AVX2-FCP-NEXT: vmovaps %ymm9, (%rdx)
1632 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
1633 ; AVX2-FCP-NEXT: vmovaps %ymm0, 96(%rdx)
1634 ; AVX2-FCP-NEXT: vmovaps %ymm6, 32(%rdx)
1635 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
1636 ; AVX2-FCP-NEXT: vmovaps %ymm0, 320(%rdx)
1637 ; AVX2-FCP-NEXT: vmovaps %ymm5, 256(%rdx)
1638 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
1639 ; AVX2-FCP-NEXT: vmovaps %ymm0, 352(%rdx)
1640 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
1641 ; AVX2-FCP-NEXT: vmovaps %ymm0, 288(%rdx)
1642 ; AVX2-FCP-NEXT: popq %rax
1643 ; AVX2-FCP-NEXT: vzeroupper
1644 ; AVX2-FCP-NEXT: retq
1645 ;
1646 ; AVX512-LABEL: store_i64_stride8_vf8:
1647 ; AVX512: # %bb.0:
1648 ; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %rax
1649 ; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %r10
1650 ; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %r11
1651 ; AVX512-NEXT: vmovdqa64 (%rdi), %zmm6
1652 ; AVX512-NEXT: vmovdqa64 (%rsi), %zmm8
1653 ; AVX512-NEXT: vmovdqa64 (%rdx), %zmm7
1654 ; AVX512-NEXT: vmovdqa64 (%rcx), %zmm11
1655 ; AVX512-NEXT: vmovdqa64 (%r8), %zmm0
1656 ; AVX512-NEXT: vmovdqa64 (%r9), %zmm2
1657 ; AVX512-NEXT: vmovdqa64 (%r11), %zmm1
1658 ; AVX512-NEXT: vmovdqa64 (%r10), %zmm3
1659 ; AVX512-NEXT: vbroadcasti32x4 {{.*#+}} zmm4 = [1,9,1,9,1,9,1,9]
1660 ; AVX512-NEXT: # zmm4 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
1661 ; AVX512-NEXT: vmovdqa64 %zmm1, %zmm5
1662 ; AVX512-NEXT: vpermt2q %zmm3, %zmm4, %zmm5
1663 ; AVX512-NEXT: vpermi2q %zmm2, %zmm0, %zmm4
1664 ; AVX512-NEXT: movb $-64, %r8b
1665 ; AVX512-NEXT: kmovw %r8d, %k1
1666 ; AVX512-NEXT: vmovdqa64 %zmm5, %zmm4 {%k1}
1667 ; AVX512-NEXT: vmovdqa (%rsi), %xmm5
1668 ; AVX512-NEXT: vinserti128 $1, (%rcx), %ymm5, %ymm10
1669 ; AVX512-NEXT: vmovdqa (%rdi), %xmm5
1670 ; AVX512-NEXT: vinserti128 $1, (%rdx), %ymm5, %ymm12
1671 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} ymm5 = ymm12[1],ymm10[1],ymm12[3],ymm10[3]
1672 ; AVX512-NEXT: vinserti64x4 $0, %ymm5, %zmm4, %zmm16
1673 ; AVX512-NEXT: vbroadcasti32x4 {{.*#+}} zmm5 = [6,14,6,14,6,14,6,14]
1674 ; AVX512-NEXT: # zmm5 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
1675 ; AVX512-NEXT: vmovdqa64 %zmm7, %zmm9
1676 ; AVX512-NEXT: vpermt2q %zmm11, %zmm5, %zmm9
1677 ; AVX512-NEXT: vmovdqa64 %zmm6, %zmm13
1678 ; AVX512-NEXT: vpermt2q %zmm8, %zmm5, %zmm13
1679 ; AVX512-NEXT: vpblendd {{.*#+}} ymm9 = ymm13[0,1,2,3],ymm9[4,5,6,7]
1680 ; AVX512-NEXT: vpermi2q %zmm2, %zmm0, %zmm5
1681 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} zmm5 {%k1} = zmm1[0],zmm3[0],zmm1[2],zmm3[2],zmm1[4],zmm3[4],zmm1[6],zmm3[6]
1682 ; AVX512-NEXT: vinserti64x4 $0, %ymm9, %zmm5, %zmm5
1683 ; AVX512-NEXT: vbroadcasti32x4 {{.*#+}} zmm9 = [7,15,7,15,7,15,7,15]
1684 ; AVX512-NEXT: # zmm9 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
1685 ; AVX512-NEXT: vmovdqa64 %zmm7, %zmm13
1686 ; AVX512-NEXT: vpermt2q %zmm11, %zmm9, %zmm13
1687 ; AVX512-NEXT: vmovdqa64 %zmm6, %zmm14
1688 ; AVX512-NEXT: vpermt2q %zmm8, %zmm9, %zmm14
1689 ; AVX512-NEXT: vpblendd {{.*#+}} ymm13 = ymm14[0,1,2,3],ymm13[4,5,6,7]
1690 ; AVX512-NEXT: vpermi2q %zmm2, %zmm0, %zmm9
1691 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} zmm9 {%k1} = zmm1[1],zmm3[1],zmm1[3],zmm3[3],zmm1[5],zmm3[5],zmm1[7],zmm3[7]
1692 ; AVX512-NEXT: vinserti64x4 $0, %ymm13, %zmm9, %zmm9
1693 ; AVX512-NEXT: vbroadcasti32x4 {{.*#+}} zmm13 = [4,12,4,12,4,12,4,12]
1694 ; AVX512-NEXT: # zmm13 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
1695 ; AVX512-NEXT: vmovdqa64 %zmm7, %zmm14
1696 ; AVX512-NEXT: vpermt2q %zmm11, %zmm13, %zmm14
1697 ; AVX512-NEXT: vmovdqa64 %zmm6, %zmm15
1698 ; AVX512-NEXT: vpermt2q %zmm8, %zmm13, %zmm15
1699 ; AVX512-NEXT: vpblendd {{.*#+}} ymm14 = ymm15[0,1,2,3],ymm14[4,5,6,7]
1700 ; AVX512-NEXT: vpermi2q %zmm3, %zmm1, %zmm13
1701 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} zmm15 = zmm0[0],zmm2[0],zmm0[2],zmm2[2],zmm0[4],zmm2[4],zmm0[6],zmm2[6]
1702 ; AVX512-NEXT: vmovdqa64 %zmm13, %zmm15 {%k1}
1703 ; AVX512-NEXT: vinserti64x4 $0, %ymm14, %zmm15, %zmm13
1704 ; AVX512-NEXT: vbroadcasti32x4 {{.*#+}} zmm14 = [5,13,5,13,5,13,5,13]
1705 ; AVX512-NEXT: # zmm14 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
1706 ; AVX512-NEXT: vpermt2q %zmm11, %zmm14, %zmm7
1707 ; AVX512-NEXT: vpermt2q %zmm8, %zmm14, %zmm6
1708 ; AVX512-NEXT: vpblendd {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7]
1709 ; AVX512-NEXT: vpermi2q %zmm3, %zmm1, %zmm14
1710 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} zmm7 = zmm0[1],zmm2[1],zmm0[3],zmm2[3],zmm0[5],zmm2[5],zmm0[7],zmm2[7]
1711 ; AVX512-NEXT: vmovdqa64 %zmm14, %zmm7 {%k1}
1712 ; AVX512-NEXT: vinserti64x4 $0, %ymm6, %zmm7, %zmm6
1713 ; AVX512-NEXT: vbroadcasti32x4 {{.*#+}} zmm7 = [0,8,0,8,0,8,0,8]
1714 ; AVX512-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
1715 ; AVX512-NEXT: vmovdqa64 %zmm1, %zmm8
1716 ; AVX512-NEXT: vpermt2q %zmm3, %zmm7, %zmm8
1717 ; AVX512-NEXT: vpermi2q %zmm2, %zmm0, %zmm7
1718 ; AVX512-NEXT: vmovdqa64 %zmm8, %zmm7 {%k1}
1719 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} ymm8 = ymm12[0],ymm10[0],ymm12[2],ymm10[2]
1720 ; AVX512-NEXT: vinserti64x4 $0, %ymm8, %zmm7, %zmm7
1721 ; AVX512-NEXT: vbroadcasti32x4 {{.*#+}} zmm8 = [2,10,2,10,2,10,2,10]
1722 ; AVX512-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
1723 ; AVX512-NEXT: vmovdqa64 %zmm1, %zmm10
1724 ; AVX512-NEXT: vpermt2q %zmm3, %zmm8, %zmm10
1725 ; AVX512-NEXT: vpermi2q %zmm2, %zmm0, %zmm8
1726 ; AVX512-NEXT: vmovdqa64 %zmm10, %zmm8 {%k1}
1727 ; AVX512-NEXT: vmovdqa (%rcx), %ymm10
1728 ; AVX512-NEXT: vmovdqa (%rdx), %ymm11
1729 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} ymm12 = ymm11[0],ymm10[0],ymm11[2],ymm10[2]
1730 ; AVX512-NEXT: vmovdqa (%rsi), %ymm14
1731 ; AVX512-NEXT: vmovdqa (%rdi), %ymm15
1732 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} ymm4 = ymm15[0],ymm14[0],ymm15[2],ymm14[2]
1733 ; AVX512-NEXT: vperm2i128 {{.*#+}} ymm4 = ymm4[2,3],ymm12[2,3]
1734 ; AVX512-NEXT: vinserti64x4 $0, %ymm4, %zmm8, %zmm4
1735 ; AVX512-NEXT: vbroadcasti32x4 {{.*#+}} zmm8 = [3,11,3,11,3,11,3,11]
1736 ; AVX512-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
1737 ; AVX512-NEXT: vpermt2q %zmm3, %zmm8, %zmm1
1738 ; AVX512-NEXT: vpermt2q %zmm2, %zmm8, %zmm0
1739 ; AVX512-NEXT: vmovdqa64 %zmm1, %zmm0 {%k1}
1740 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm11[1],ymm10[1],ymm11[3],ymm10[3]
1741 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} ymm2 = ymm15[1],ymm14[1],ymm15[3],ymm14[3]
1742 ; AVX512-NEXT: vperm2i128 {{.*#+}} ymm1 = ymm2[2,3],ymm1[2,3]
1743 ; AVX512-NEXT: vinserti64x4 $0, %ymm1, %zmm0, %zmm0
1744 ; AVX512-NEXT: vmovdqa64 %zmm0, 192(%rax)
1745 ; AVX512-NEXT: vmovdqa64 %zmm4, 128(%rax)
1746 ; AVX512-NEXT: vmovdqa64 %zmm7, (%rax)
1747 ; AVX512-NEXT: vmovdqa64 %zmm6, 320(%rax)
1748 ; AVX512-NEXT: vmovdqa64 %zmm13, 256(%rax)
1749 ; AVX512-NEXT: vmovdqa64 %zmm9, 448(%rax)
1750 ; AVX512-NEXT: vmovdqa64 %zmm5, 384(%rax)
1751 ; AVX512-NEXT: vmovdqa64 %zmm16, 64(%rax)
1752 ; AVX512-NEXT: vzeroupper
1753 ; AVX512-NEXT: retq
1754 ;
1755 ; AVX512-FCP-LABEL: store_i64_stride8_vf8:
1756 ; AVX512-FCP: # %bb.0:
1757 ; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
1758 ; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10
1759 ; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r11
1760 ; AVX512-FCP-NEXT: vmovdqa64 (%rdi), %zmm6
1761 ; AVX512-FCP-NEXT: vmovdqa64 (%rsi), %zmm8
1762 ; AVX512-FCP-NEXT: vmovdqa64 (%rdx), %zmm7
1763 ; AVX512-FCP-NEXT: vmovdqa64 (%rcx), %zmm11
1764 ; AVX512-FCP-NEXT: vmovdqa64 (%r8), %zmm0
1765 ; AVX512-FCP-NEXT: vmovdqa64 (%r9), %zmm2
1766 ; AVX512-FCP-NEXT: vmovdqa64 (%r11), %zmm1
1767 ; AVX512-FCP-NEXT: vmovdqa64 (%r10), %zmm3
1768 ; AVX512-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm4 = [1,9,1,9,1,9,1,9]
1769 ; AVX512-FCP-NEXT: # zmm4 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
1770 ; AVX512-FCP-NEXT: vmovdqa64 %zmm1, %zmm5
1771 ; AVX512-FCP-NEXT: vpermt2q %zmm3, %zmm4, %zmm5
1772 ; AVX512-FCP-NEXT: vpermi2q %zmm2, %zmm0, %zmm4
1773 ; AVX512-FCP-NEXT: movb $-64, %r8b
1774 ; AVX512-FCP-NEXT: kmovw %r8d, %k1
1775 ; AVX512-FCP-NEXT: vmovdqa64 %zmm5, %zmm4 {%k1}
1776 ; AVX512-FCP-NEXT: vmovdqa (%rsi), %xmm5
1777 ; AVX512-FCP-NEXT: vinserti128 $1, (%rcx), %ymm5, %ymm10
1778 ; AVX512-FCP-NEXT: vmovdqa (%rdi), %xmm5
1779 ; AVX512-FCP-NEXT: vinserti128 $1, (%rdx), %ymm5, %ymm12
1780 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm5 = ymm12[1],ymm10[1],ymm12[3],ymm10[3]
1781 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm5, %zmm4, %zmm16
1782 ; AVX512-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm5 = [6,14,6,14,6,14,6,14]
1783 ; AVX512-FCP-NEXT: # zmm5 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
1784 ; AVX512-FCP-NEXT: vmovdqa64 %zmm7, %zmm9
1785 ; AVX512-FCP-NEXT: vpermt2q %zmm11, %zmm5, %zmm9
1786 ; AVX512-FCP-NEXT: vmovdqa64 %zmm6, %zmm13
1787 ; AVX512-FCP-NEXT: vpermt2q %zmm8, %zmm5, %zmm13
1788 ; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm9 = ymm13[0,1,2,3],ymm9[4,5,6,7]
1789 ; AVX512-FCP-NEXT: vpermi2q %zmm2, %zmm0, %zmm5
1790 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm5 {%k1} = zmm1[0],zmm3[0],zmm1[2],zmm3[2],zmm1[4],zmm3[4],zmm1[6],zmm3[6]
1791 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm9, %zmm5, %zmm5
1792 ; AVX512-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm9 = [7,15,7,15,7,15,7,15]
1793 ; AVX512-FCP-NEXT: # zmm9 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
1794 ; AVX512-FCP-NEXT: vmovdqa64 %zmm7, %zmm13
1795 ; AVX512-FCP-NEXT: vpermt2q %zmm11, %zmm9, %zmm13
1796 ; AVX512-FCP-NEXT: vmovdqa64 %zmm6, %zmm14
1797 ; AVX512-FCP-NEXT: vpermt2q %zmm8, %zmm9, %zmm14
1798 ; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm13 = ymm14[0,1,2,3],ymm13[4,5,6,7]
1799 ; AVX512-FCP-NEXT: vpermi2q %zmm2, %zmm0, %zmm9
1800 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm9 {%k1} = zmm1[1],zmm3[1],zmm1[3],zmm3[3],zmm1[5],zmm3[5],zmm1[7],zmm3[7]
1801 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm13, %zmm9, %zmm9
1802 ; AVX512-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm13 = [4,12,4,12,4,12,4,12]
1803 ; AVX512-FCP-NEXT: # zmm13 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
1804 ; AVX512-FCP-NEXT: vmovdqa64 %zmm7, %zmm14
1805 ; AVX512-FCP-NEXT: vpermt2q %zmm11, %zmm13, %zmm14
1806 ; AVX512-FCP-NEXT: vmovdqa64 %zmm6, %zmm15
1807 ; AVX512-FCP-NEXT: vpermt2q %zmm8, %zmm13, %zmm15
1808 ; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm14 = ymm15[0,1,2,3],ymm14[4,5,6,7]
1809 ; AVX512-FCP-NEXT: vpermi2q %zmm3, %zmm1, %zmm13
1810 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm15 = zmm0[0],zmm2[0],zmm0[2],zmm2[2],zmm0[4],zmm2[4],zmm0[6],zmm2[6]
1811 ; AVX512-FCP-NEXT: vmovdqa64 %zmm13, %zmm15 {%k1}
1812 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm14, %zmm15, %zmm13
1813 ; AVX512-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm14 = [5,13,5,13,5,13,5,13]
1814 ; AVX512-FCP-NEXT: # zmm14 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
1815 ; AVX512-FCP-NEXT: vpermt2q %zmm11, %zmm14, %zmm7
1816 ; AVX512-FCP-NEXT: vpermt2q %zmm8, %zmm14, %zmm6
1817 ; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7]
1818 ; AVX512-FCP-NEXT: vpermi2q %zmm3, %zmm1, %zmm14
1819 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm7 = zmm0[1],zmm2[1],zmm0[3],zmm2[3],zmm0[5],zmm2[5],zmm0[7],zmm2[7]
1820 ; AVX512-FCP-NEXT: vmovdqa64 %zmm14, %zmm7 {%k1}
1821 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm7, %zmm6
1822 ; AVX512-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm7 = [0,8,0,8,0,8,0,8]
1823 ; AVX512-FCP-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
1824 ; AVX512-FCP-NEXT: vmovdqa64 %zmm1, %zmm8
1825 ; AVX512-FCP-NEXT: vpermt2q %zmm3, %zmm7, %zmm8
1826 ; AVX512-FCP-NEXT: vpermi2q %zmm2, %zmm0, %zmm7
1827 ; AVX512-FCP-NEXT: vmovdqa64 %zmm8, %zmm7 {%k1}
1828 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm8 = ymm12[0],ymm10[0],ymm12[2],ymm10[2]
1829 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm8, %zmm7, %zmm7
1830 ; AVX512-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm8 = [2,10,2,10,2,10,2,10]
1831 ; AVX512-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
1832 ; AVX512-FCP-NEXT: vmovdqa64 %zmm1, %zmm10
1833 ; AVX512-FCP-NEXT: vpermt2q %zmm3, %zmm8, %zmm10
1834 ; AVX512-FCP-NEXT: vpermi2q %zmm2, %zmm0, %zmm8
1835 ; AVX512-FCP-NEXT: vmovdqa64 %zmm10, %zmm8 {%k1}
1836 ; AVX512-FCP-NEXT: vmovdqa (%rcx), %ymm10
1837 ; AVX512-FCP-NEXT: vmovdqa (%rdx), %ymm11
1838 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm12 = ymm11[0],ymm10[0],ymm11[2],ymm10[2]
1839 ; AVX512-FCP-NEXT: vmovdqa (%rsi), %ymm14
1840 ; AVX512-FCP-NEXT: vmovdqa (%rdi), %ymm15
1841 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm4 = ymm15[0],ymm14[0],ymm15[2],ymm14[2]
1842 ; AVX512-FCP-NEXT: vperm2i128 {{.*#+}} ymm4 = ymm4[2,3],ymm12[2,3]
1843 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm4, %zmm8, %zmm4
1844 ; AVX512-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm8 = [3,11,3,11,3,11,3,11]
1845 ; AVX512-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
1846 ; AVX512-FCP-NEXT: vpermt2q %zmm3, %zmm8, %zmm1
1847 ; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm8, %zmm0
1848 ; AVX512-FCP-NEXT: vmovdqa64 %zmm1, %zmm0 {%k1}
1849 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm11[1],ymm10[1],ymm11[3],ymm10[3]
1850 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm2 = ymm15[1],ymm14[1],ymm15[3],ymm14[3]
1851 ; AVX512-FCP-NEXT: vperm2i128 {{.*#+}} ymm1 = ymm2[2,3],ymm1[2,3]
1852 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm1, %zmm0, %zmm0
1853 ; AVX512-FCP-NEXT: vmovdqa64 %zmm0, 192(%rax)
1854 ; AVX512-FCP-NEXT: vmovdqa64 %zmm4, 128(%rax)
1855 ; AVX512-FCP-NEXT: vmovdqa64 %zmm7, (%rax)
1856 ; AVX512-FCP-NEXT: vmovdqa64 %zmm6, 320(%rax)
1857 ; AVX512-FCP-NEXT: vmovdqa64 %zmm13, 256(%rax)
1858 ; AVX512-FCP-NEXT: vmovdqa64 %zmm9, 448(%rax)
1859 ; AVX512-FCP-NEXT: vmovdqa64 %zmm5, 384(%rax)
1860 ; AVX512-FCP-NEXT: vmovdqa64 %zmm16, 64(%rax)
1861 ; AVX512-FCP-NEXT: vzeroupper
1862 ; AVX512-FCP-NEXT: retq
1863 ;
1864 ; AVX512DQ-LABEL: store_i64_stride8_vf8:
1865 ; AVX512DQ: # %bb.0:
1866 ; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %rax
1867 ; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %r10
1868 ; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %r11
1869 ; AVX512DQ-NEXT: vmovdqa64 (%rdi), %zmm6
1870 ; AVX512DQ-NEXT: vmovdqa64 (%rsi), %zmm8
1871 ; AVX512DQ-NEXT: vmovdqa64 (%rdx), %zmm7
1872 ; AVX512DQ-NEXT: vmovdqa64 (%rcx), %zmm11
1873 ; AVX512DQ-NEXT: vmovdqa64 (%r8), %zmm0
1874 ; AVX512DQ-NEXT: vmovdqa64 (%r9), %zmm2
1875 ; AVX512DQ-NEXT: vmovdqa64 (%r11), %zmm1
1876 ; AVX512DQ-NEXT: vmovdqa64 (%r10), %zmm3
1877 ; AVX512DQ-NEXT: vbroadcasti32x4 {{.*#+}} zmm4 = [1,9,1,9,1,9,1,9]
1878 ; AVX512DQ-NEXT: # zmm4 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
1879 ; AVX512DQ-NEXT: vmovdqa64 %zmm1, %zmm5
1880 ; AVX512DQ-NEXT: vpermt2q %zmm3, %zmm4, %zmm5
1881 ; AVX512DQ-NEXT: vpermi2q %zmm2, %zmm0, %zmm4
1882 ; AVX512DQ-NEXT: movb $-64, %r8b
1883 ; AVX512DQ-NEXT: kmovw %r8d, %k1
1884 ; AVX512DQ-NEXT: vmovdqa64 %zmm5, %zmm4 {%k1}
1885 ; AVX512DQ-NEXT: vmovdqa (%rsi), %xmm5
1886 ; AVX512DQ-NEXT: vinserti128 $1, (%rcx), %ymm5, %ymm10
1887 ; AVX512DQ-NEXT: vmovdqa (%rdi), %xmm5
1888 ; AVX512DQ-NEXT: vinserti128 $1, (%rdx), %ymm5, %ymm12
1889 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} ymm5 = ymm12[1],ymm10[1],ymm12[3],ymm10[3]
1890 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm5, %zmm4, %zmm16
1891 ; AVX512DQ-NEXT: vbroadcasti32x4 {{.*#+}} zmm5 = [6,14,6,14,6,14,6,14]
1892 ; AVX512DQ-NEXT: # zmm5 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
1893 ; AVX512DQ-NEXT: vmovdqa64 %zmm7, %zmm9
1894 ; AVX512DQ-NEXT: vpermt2q %zmm11, %zmm5, %zmm9
1895 ; AVX512DQ-NEXT: vmovdqa64 %zmm6, %zmm13
1896 ; AVX512DQ-NEXT: vpermt2q %zmm8, %zmm5, %zmm13
1897 ; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm9 = ymm13[0,1,2,3],ymm9[4,5,6,7]
1898 ; AVX512DQ-NEXT: vpermi2q %zmm2, %zmm0, %zmm5
1899 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} zmm5 {%k1} = zmm1[0],zmm3[0],zmm1[2],zmm3[2],zmm1[4],zmm3[4],zmm1[6],zmm3[6]
1900 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm9, %zmm5, %zmm5
1901 ; AVX512DQ-NEXT: vbroadcasti32x4 {{.*#+}} zmm9 = [7,15,7,15,7,15,7,15]
1902 ; AVX512DQ-NEXT: # zmm9 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
1903 ; AVX512DQ-NEXT: vmovdqa64 %zmm7, %zmm13
1904 ; AVX512DQ-NEXT: vpermt2q %zmm11, %zmm9, %zmm13
1905 ; AVX512DQ-NEXT: vmovdqa64 %zmm6, %zmm14
1906 ; AVX512DQ-NEXT: vpermt2q %zmm8, %zmm9, %zmm14
1907 ; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm13 = ymm14[0,1,2,3],ymm13[4,5,6,7]
1908 ; AVX512DQ-NEXT: vpermi2q %zmm2, %zmm0, %zmm9
1909 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} zmm9 {%k1} = zmm1[1],zmm3[1],zmm1[3],zmm3[3],zmm1[5],zmm3[5],zmm1[7],zmm3[7]
1910 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm13, %zmm9, %zmm9
1911 ; AVX512DQ-NEXT: vbroadcasti32x4 {{.*#+}} zmm13 = [4,12,4,12,4,12,4,12]
1912 ; AVX512DQ-NEXT: # zmm13 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
1913 ; AVX512DQ-NEXT: vmovdqa64 %zmm7, %zmm14
1914 ; AVX512DQ-NEXT: vpermt2q %zmm11, %zmm13, %zmm14
1915 ; AVX512DQ-NEXT: vmovdqa64 %zmm6, %zmm15
1916 ; AVX512DQ-NEXT: vpermt2q %zmm8, %zmm13, %zmm15
1917 ; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm14 = ymm15[0,1,2,3],ymm14[4,5,6,7]
1918 ; AVX512DQ-NEXT: vpermi2q %zmm3, %zmm1, %zmm13
1919 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} zmm15 = zmm0[0],zmm2[0],zmm0[2],zmm2[2],zmm0[4],zmm2[4],zmm0[6],zmm2[6]
1920 ; AVX512DQ-NEXT: vmovdqa64 %zmm13, %zmm15 {%k1}
1921 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm14, %zmm15, %zmm13
1922 ; AVX512DQ-NEXT: vbroadcasti32x4 {{.*#+}} zmm14 = [5,13,5,13,5,13,5,13]
1923 ; AVX512DQ-NEXT: # zmm14 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
1924 ; AVX512DQ-NEXT: vpermt2q %zmm11, %zmm14, %zmm7
1925 ; AVX512DQ-NEXT: vpermt2q %zmm8, %zmm14, %zmm6
1926 ; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7]
1927 ; AVX512DQ-NEXT: vpermi2q %zmm3, %zmm1, %zmm14
1928 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} zmm7 = zmm0[1],zmm2[1],zmm0[3],zmm2[3],zmm0[5],zmm2[5],zmm0[7],zmm2[7]
1929 ; AVX512DQ-NEXT: vmovdqa64 %zmm14, %zmm7 {%k1}
1930 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm6, %zmm7, %zmm6
1931 ; AVX512DQ-NEXT: vbroadcasti32x4 {{.*#+}} zmm7 = [0,8,0,8,0,8,0,8]
1932 ; AVX512DQ-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
1933 ; AVX512DQ-NEXT: vmovdqa64 %zmm1, %zmm8
1934 ; AVX512DQ-NEXT: vpermt2q %zmm3, %zmm7, %zmm8
1935 ; AVX512DQ-NEXT: vpermi2q %zmm2, %zmm0, %zmm7
1936 ; AVX512DQ-NEXT: vmovdqa64 %zmm8, %zmm7 {%k1}
1937 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} ymm8 = ymm12[0],ymm10[0],ymm12[2],ymm10[2]
1938 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm8, %zmm7, %zmm7
1939 ; AVX512DQ-NEXT: vbroadcasti32x4 {{.*#+}} zmm8 = [2,10,2,10,2,10,2,10]
1940 ; AVX512DQ-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
1941 ; AVX512DQ-NEXT: vmovdqa64 %zmm1, %zmm10
1942 ; AVX512DQ-NEXT: vpermt2q %zmm3, %zmm8, %zmm10
1943 ; AVX512DQ-NEXT: vpermi2q %zmm2, %zmm0, %zmm8
1944 ; AVX512DQ-NEXT: vmovdqa64 %zmm10, %zmm8 {%k1}
1945 ; AVX512DQ-NEXT: vmovdqa (%rcx), %ymm10
1946 ; AVX512DQ-NEXT: vmovdqa (%rdx), %ymm11
1947 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} ymm12 = ymm11[0],ymm10[0],ymm11[2],ymm10[2]
1948 ; AVX512DQ-NEXT: vmovdqa (%rsi), %ymm14
1949 ; AVX512DQ-NEXT: vmovdqa (%rdi), %ymm15
1950 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} ymm4 = ymm15[0],ymm14[0],ymm15[2],ymm14[2]
1951 ; AVX512DQ-NEXT: vperm2i128 {{.*#+}} ymm4 = ymm4[2,3],ymm12[2,3]
1952 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm4, %zmm8, %zmm4
1953 ; AVX512DQ-NEXT: vbroadcasti32x4 {{.*#+}} zmm8 = [3,11,3,11,3,11,3,11]
1954 ; AVX512DQ-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
1955 ; AVX512DQ-NEXT: vpermt2q %zmm3, %zmm8, %zmm1
1956 ; AVX512DQ-NEXT: vpermt2q %zmm2, %zmm8, %zmm0
1957 ; AVX512DQ-NEXT: vmovdqa64 %zmm1, %zmm0 {%k1}
1958 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm11[1],ymm10[1],ymm11[3],ymm10[3]
1959 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} ymm2 = ymm15[1],ymm14[1],ymm15[3],ymm14[3]
1960 ; AVX512DQ-NEXT: vperm2i128 {{.*#+}} ymm1 = ymm2[2,3],ymm1[2,3]
1961 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm1, %zmm0, %zmm0
1962 ; AVX512DQ-NEXT: vmovdqa64 %zmm0, 192(%rax)
1963 ; AVX512DQ-NEXT: vmovdqa64 %zmm4, 128(%rax)
1964 ; AVX512DQ-NEXT: vmovdqa64 %zmm7, (%rax)
1965 ; AVX512DQ-NEXT: vmovdqa64 %zmm6, 320(%rax)
1966 ; AVX512DQ-NEXT: vmovdqa64 %zmm13, 256(%rax)
1967 ; AVX512DQ-NEXT: vmovdqa64 %zmm9, 448(%rax)
1968 ; AVX512DQ-NEXT: vmovdqa64 %zmm5, 384(%rax)
1969 ; AVX512DQ-NEXT: vmovdqa64 %zmm16, 64(%rax)
1970 ; AVX512DQ-NEXT: vzeroupper
1971 ; AVX512DQ-NEXT: retq
1972 ;
1973 ; AVX512DQ-FCP-LABEL: store_i64_stride8_vf8:
1974 ; AVX512DQ-FCP: # %bb.0:
1975 ; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
1976 ; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10
1977 ; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r11
1978 ; AVX512DQ-FCP-NEXT: vmovdqa64 (%rdi), %zmm6
1979 ; AVX512DQ-FCP-NEXT: vmovdqa64 (%rsi), %zmm8
1980 ; AVX512DQ-FCP-NEXT: vmovdqa64 (%rdx), %zmm7
1981 ; AVX512DQ-FCP-NEXT: vmovdqa64 (%rcx), %zmm11
1982 ; AVX512DQ-FCP-NEXT: vmovdqa64 (%r8), %zmm0
1983 ; AVX512DQ-FCP-NEXT: vmovdqa64 (%r9), %zmm2
1984 ; AVX512DQ-FCP-NEXT: vmovdqa64 (%r11), %zmm1
1985 ; AVX512DQ-FCP-NEXT: vmovdqa64 (%r10), %zmm3
1986 ; AVX512DQ-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm4 = [1,9,1,9,1,9,1,9]
1987 ; AVX512DQ-FCP-NEXT: # zmm4 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
1988 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, %zmm5
1989 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm3, %zmm4, %zmm5
1990 ; AVX512DQ-FCP-NEXT: vpermi2q %zmm2, %zmm0, %zmm4
1991 ; AVX512DQ-FCP-NEXT: movb $-64, %r8b
1992 ; AVX512DQ-FCP-NEXT: kmovw %r8d, %k1
1993 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm5, %zmm4 {%k1}
1994 ; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %xmm5
1995 ; AVX512DQ-FCP-NEXT: vinserti128 $1, (%rcx), %ymm5, %ymm10
1996 ; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %xmm5
1997 ; AVX512DQ-FCP-NEXT: vinserti128 $1, (%rdx), %ymm5, %ymm12
1998 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm5 = ymm12[1],ymm10[1],ymm12[3],ymm10[3]
1999 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm5, %zmm4, %zmm16
2000 ; AVX512DQ-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm5 = [6,14,6,14,6,14,6,14]
2001 ; AVX512DQ-FCP-NEXT: # zmm5 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
2002 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm7, %zmm9
2003 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm11, %zmm5, %zmm9
2004 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, %zmm13
2005 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm8, %zmm5, %zmm13
2006 ; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm9 = ymm13[0,1,2,3],ymm9[4,5,6,7]
2007 ; AVX512DQ-FCP-NEXT: vpermi2q %zmm2, %zmm0, %zmm5
2008 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm5 {%k1} = zmm1[0],zmm3[0],zmm1[2],zmm3[2],zmm1[4],zmm3[4],zmm1[6],zmm3[6]
2009 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm9, %zmm5, %zmm5
2010 ; AVX512DQ-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm9 = [7,15,7,15,7,15,7,15]
2011 ; AVX512DQ-FCP-NEXT: # zmm9 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
2012 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm7, %zmm13
2013 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm11, %zmm9, %zmm13
2014 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, %zmm14
2015 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm8, %zmm9, %zmm14
2016 ; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm13 = ymm14[0,1,2,3],ymm13[4,5,6,7]
2017 ; AVX512DQ-FCP-NEXT: vpermi2q %zmm2, %zmm0, %zmm9
2018 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm9 {%k1} = zmm1[1],zmm3[1],zmm1[3],zmm3[3],zmm1[5],zmm3[5],zmm1[7],zmm3[7]
2019 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm13, %zmm9, %zmm9
2020 ; AVX512DQ-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm13 = [4,12,4,12,4,12,4,12]
2021 ; AVX512DQ-FCP-NEXT: # zmm13 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
2022 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm7, %zmm14
2023 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm11, %zmm13, %zmm14
2024 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, %zmm15
2025 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm8, %zmm13, %zmm15
2026 ; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm14 = ymm15[0,1,2,3],ymm14[4,5,6,7]
2027 ; AVX512DQ-FCP-NEXT: vpermi2q %zmm3, %zmm1, %zmm13
2028 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm15 = zmm0[0],zmm2[0],zmm0[2],zmm2[2],zmm0[4],zmm2[4],zmm0[6],zmm2[6]
2029 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm13, %zmm15 {%k1}
2030 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm14, %zmm15, %zmm13
2031 ; AVX512DQ-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm14 = [5,13,5,13,5,13,5,13]
2032 ; AVX512DQ-FCP-NEXT: # zmm14 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
2033 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm11, %zmm14, %zmm7
2034 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm8, %zmm14, %zmm6
2035 ; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7]
2036 ; AVX512DQ-FCP-NEXT: vpermi2q %zmm3, %zmm1, %zmm14
2037 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm7 = zmm0[1],zmm2[1],zmm0[3],zmm2[3],zmm0[5],zmm2[5],zmm0[7],zmm2[7]
2038 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm14, %zmm7 {%k1}
2039 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm7, %zmm6
2040 ; AVX512DQ-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm7 = [0,8,0,8,0,8,0,8]
2041 ; AVX512DQ-FCP-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
2042 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, %zmm8
2043 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm3, %zmm7, %zmm8
2044 ; AVX512DQ-FCP-NEXT: vpermi2q %zmm2, %zmm0, %zmm7
2045 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm8, %zmm7 {%k1}
2046 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm8 = ymm12[0],ymm10[0],ymm12[2],ymm10[2]
2047 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm8, %zmm7, %zmm7
2048 ; AVX512DQ-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm8 = [2,10,2,10,2,10,2,10]
2049 ; AVX512DQ-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
2050 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, %zmm10
2051 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm3, %zmm8, %zmm10
2052 ; AVX512DQ-FCP-NEXT: vpermi2q %zmm2, %zmm0, %zmm8
2053 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm10, %zmm8 {%k1}
2054 ; AVX512DQ-FCP-NEXT: vmovdqa (%rcx), %ymm10
2055 ; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %ymm11
2056 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm12 = ymm11[0],ymm10[0],ymm11[2],ymm10[2]
2057 ; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %ymm14
2058 ; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %ymm15
2059 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm4 = ymm15[0],ymm14[0],ymm15[2],ymm14[2]
2060 ; AVX512DQ-FCP-NEXT: vperm2i128 {{.*#+}} ymm4 = ymm4[2,3],ymm12[2,3]
2061 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm4, %zmm8, %zmm4
2062 ; AVX512DQ-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm8 = [3,11,3,11,3,11,3,11]
2063 ; AVX512DQ-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
2064 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm3, %zmm8, %zmm1
2065 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm8, %zmm0
2066 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, %zmm0 {%k1}
2067 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm11[1],ymm10[1],ymm11[3],ymm10[3]
2068 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm2 = ymm15[1],ymm14[1],ymm15[3],ymm14[3]
2069 ; AVX512DQ-FCP-NEXT: vperm2i128 {{.*#+}} ymm1 = ymm2[2,3],ymm1[2,3]
2070 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm1, %zmm0, %zmm0
2071 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, 192(%rax)
2072 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm4, 128(%rax)
2073 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm7, (%rax)
2074 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, 320(%rax)
2075 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm13, 256(%rax)
2076 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm9, 448(%rax)
2077 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm5, 384(%rax)
2078 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm16, 64(%rax)
2079 ; AVX512DQ-FCP-NEXT: vzeroupper
2080 ; AVX512DQ-FCP-NEXT: retq
2081 ;
2082 ; AVX512BW-LABEL: store_i64_stride8_vf8:
2083 ; AVX512BW: # %bb.0:
2084 ; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
2085 ; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %r10
2086 ; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %r11
2087 ; AVX512BW-NEXT: vmovdqa64 (%rdi), %zmm6
2088 ; AVX512BW-NEXT: vmovdqa64 (%rsi), %zmm8
2089 ; AVX512BW-NEXT: vmovdqa64 (%rdx), %zmm7
2090 ; AVX512BW-NEXT: vmovdqa64 (%rcx), %zmm11
2091 ; AVX512BW-NEXT: vmovdqa64 (%r8), %zmm0
2092 ; AVX512BW-NEXT: vmovdqa64 (%r9), %zmm2
2093 ; AVX512BW-NEXT: vmovdqa64 (%r11), %zmm1
2094 ; AVX512BW-NEXT: vmovdqa64 (%r10), %zmm3
2095 ; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm4 = [1,9,1,9,1,9,1,9]
2096 ; AVX512BW-NEXT: # zmm4 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
2097 ; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm5
2098 ; AVX512BW-NEXT: vpermt2q %zmm3, %zmm4, %zmm5
2099 ; AVX512BW-NEXT: vpermi2q %zmm2, %zmm0, %zmm4
2100 ; AVX512BW-NEXT: movb $-64, %r8b
2101 ; AVX512BW-NEXT: kmovd %r8d, %k1
2102 ; AVX512BW-NEXT: vmovdqa64 %zmm5, %zmm4 {%k1}
2103 ; AVX512BW-NEXT: vmovdqa (%rsi), %xmm5
2104 ; AVX512BW-NEXT: vinserti128 $1, (%rcx), %ymm5, %ymm10
2105 ; AVX512BW-NEXT: vmovdqa (%rdi), %xmm5
2106 ; AVX512BW-NEXT: vinserti128 $1, (%rdx), %ymm5, %ymm12
2107 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} ymm5 = ymm12[1],ymm10[1],ymm12[3],ymm10[3]
2108 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm5, %zmm4, %zmm16
2109 ; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm5 = [6,14,6,14,6,14,6,14]
2110 ; AVX512BW-NEXT: # zmm5 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
2111 ; AVX512BW-NEXT: vmovdqa64 %zmm7, %zmm9
2112 ; AVX512BW-NEXT: vpermt2q %zmm11, %zmm5, %zmm9
2113 ; AVX512BW-NEXT: vmovdqa64 %zmm6, %zmm13
2114 ; AVX512BW-NEXT: vpermt2q %zmm8, %zmm5, %zmm13
2115 ; AVX512BW-NEXT: vpblendd {{.*#+}} ymm9 = ymm13[0,1,2,3],ymm9[4,5,6,7]
2116 ; AVX512BW-NEXT: vpermi2q %zmm2, %zmm0, %zmm5
2117 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} zmm5 {%k1} = zmm1[0],zmm3[0],zmm1[2],zmm3[2],zmm1[4],zmm3[4],zmm1[6],zmm3[6]
2118 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm9, %zmm5, %zmm5
2119 ; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm9 = [7,15,7,15,7,15,7,15]
2120 ; AVX512BW-NEXT: # zmm9 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
2121 ; AVX512BW-NEXT: vmovdqa64 %zmm7, %zmm13
2122 ; AVX512BW-NEXT: vpermt2q %zmm11, %zmm9, %zmm13
2123 ; AVX512BW-NEXT: vmovdqa64 %zmm6, %zmm14
2124 ; AVX512BW-NEXT: vpermt2q %zmm8, %zmm9, %zmm14
2125 ; AVX512BW-NEXT: vpblendd {{.*#+}} ymm13 = ymm14[0,1,2,3],ymm13[4,5,6,7]
2126 ; AVX512BW-NEXT: vpermi2q %zmm2, %zmm0, %zmm9
2127 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} zmm9 {%k1} = zmm1[1],zmm3[1],zmm1[3],zmm3[3],zmm1[5],zmm3[5],zmm1[7],zmm3[7]
2128 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm13, %zmm9, %zmm9
2129 ; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm13 = [4,12,4,12,4,12,4,12]
2130 ; AVX512BW-NEXT: # zmm13 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
2131 ; AVX512BW-NEXT: vmovdqa64 %zmm7, %zmm14
2132 ; AVX512BW-NEXT: vpermt2q %zmm11, %zmm13, %zmm14
2133 ; AVX512BW-NEXT: vmovdqa64 %zmm6, %zmm15
2134 ; AVX512BW-NEXT: vpermt2q %zmm8, %zmm13, %zmm15
2135 ; AVX512BW-NEXT: vpblendd {{.*#+}} ymm14 = ymm15[0,1,2,3],ymm14[4,5,6,7]
2136 ; AVX512BW-NEXT: vpermi2q %zmm3, %zmm1, %zmm13
2137 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} zmm15 = zmm0[0],zmm2[0],zmm0[2],zmm2[2],zmm0[4],zmm2[4],zmm0[6],zmm2[6]
2138 ; AVX512BW-NEXT: vmovdqa64 %zmm13, %zmm15 {%k1}
2139 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm14, %zmm15, %zmm13
2140 ; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm14 = [5,13,5,13,5,13,5,13]
2141 ; AVX512BW-NEXT: # zmm14 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
2142 ; AVX512BW-NEXT: vpermt2q %zmm11, %zmm14, %zmm7
2143 ; AVX512BW-NEXT: vpermt2q %zmm8, %zmm14, %zmm6
2144 ; AVX512BW-NEXT: vpblendd {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7]
2145 ; AVX512BW-NEXT: vpermi2q %zmm3, %zmm1, %zmm14
2146 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} zmm7 = zmm0[1],zmm2[1],zmm0[3],zmm2[3],zmm0[5],zmm2[5],zmm0[7],zmm2[7]
2147 ; AVX512BW-NEXT: vmovdqa64 %zmm14, %zmm7 {%k1}
2148 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm6, %zmm7, %zmm6
2149 ; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm7 = [0,8,0,8,0,8,0,8]
2150 ; AVX512BW-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
2151 ; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm8
2152 ; AVX512BW-NEXT: vpermt2q %zmm3, %zmm7, %zmm8
2153 ; AVX512BW-NEXT: vpermi2q %zmm2, %zmm0, %zmm7
2154 ; AVX512BW-NEXT: vmovdqa64 %zmm8, %zmm7 {%k1}
2155 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} ymm8 = ymm12[0],ymm10[0],ymm12[2],ymm10[2]
2156 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm8, %zmm7, %zmm7
2157 ; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm8 = [2,10,2,10,2,10,2,10]
2158 ; AVX512BW-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
2159 ; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm10
2160 ; AVX512BW-NEXT: vpermt2q %zmm3, %zmm8, %zmm10
2161 ; AVX512BW-NEXT: vpermi2q %zmm2, %zmm0, %zmm8
2162 ; AVX512BW-NEXT: vmovdqa64 %zmm10, %zmm8 {%k1}
2163 ; AVX512BW-NEXT: vmovdqa (%rcx), %ymm10
2164 ; AVX512BW-NEXT: vmovdqa (%rdx), %ymm11
2165 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} ymm12 = ymm11[0],ymm10[0],ymm11[2],ymm10[2]
2166 ; AVX512BW-NEXT: vmovdqa (%rsi), %ymm14
2167 ; AVX512BW-NEXT: vmovdqa (%rdi), %ymm15
2168 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} ymm4 = ymm15[0],ymm14[0],ymm15[2],ymm14[2]
2169 ; AVX512BW-NEXT: vperm2i128 {{.*#+}} ymm4 = ymm4[2,3],ymm12[2,3]
2170 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm4, %zmm8, %zmm4
2171 ; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm8 = [3,11,3,11,3,11,3,11]
2172 ; AVX512BW-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
2173 ; AVX512BW-NEXT: vpermt2q %zmm3, %zmm8, %zmm1
2174 ; AVX512BW-NEXT: vpermt2q %zmm2, %zmm8, %zmm0
2175 ; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm0 {%k1}
2176 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm11[1],ymm10[1],ymm11[3],ymm10[3]
2177 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} ymm2 = ymm15[1],ymm14[1],ymm15[3],ymm14[3]
2178 ; AVX512BW-NEXT: vperm2i128 {{.*#+}} ymm1 = ymm2[2,3],ymm1[2,3]
2179 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm1, %zmm0, %zmm0
2180 ; AVX512BW-NEXT: vmovdqa64 %zmm0, 192(%rax)
2181 ; AVX512BW-NEXT: vmovdqa64 %zmm4, 128(%rax)
2182 ; AVX512BW-NEXT: vmovdqa64 %zmm7, (%rax)
2183 ; AVX512BW-NEXT: vmovdqa64 %zmm6, 320(%rax)
2184 ; AVX512BW-NEXT: vmovdqa64 %zmm13, 256(%rax)
2185 ; AVX512BW-NEXT: vmovdqa64 %zmm9, 448(%rax)
2186 ; AVX512BW-NEXT: vmovdqa64 %zmm5, 384(%rax)
2187 ; AVX512BW-NEXT: vmovdqa64 %zmm16, 64(%rax)
2188 ; AVX512BW-NEXT: vzeroupper
2189 ; AVX512BW-NEXT: retq
2190 ;
2191 ; AVX512BW-FCP-LABEL: store_i64_stride8_vf8:
2192 ; AVX512BW-FCP: # %bb.0:
2193 ; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
2194 ; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10
2195 ; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r11
2196 ; AVX512BW-FCP-NEXT: vmovdqa64 (%rdi), %zmm6
2197 ; AVX512BW-FCP-NEXT: vmovdqa64 (%rsi), %zmm8
2198 ; AVX512BW-FCP-NEXT: vmovdqa64 (%rdx), %zmm7
2199 ; AVX512BW-FCP-NEXT: vmovdqa64 (%rcx), %zmm11
2200 ; AVX512BW-FCP-NEXT: vmovdqa64 (%r8), %zmm0
2201 ; AVX512BW-FCP-NEXT: vmovdqa64 (%r9), %zmm2
2202 ; AVX512BW-FCP-NEXT: vmovdqa64 (%r11), %zmm1
2203 ; AVX512BW-FCP-NEXT: vmovdqa64 (%r10), %zmm3
2204 ; AVX512BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm4 = [1,9,1,9,1,9,1,9]
2205 ; AVX512BW-FCP-NEXT: # zmm4 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
2206 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm5
2207 ; AVX512BW-FCP-NEXT: vpermt2q %zmm3, %zmm4, %zmm5
2208 ; AVX512BW-FCP-NEXT: vpermi2q %zmm2, %zmm0, %zmm4
2209 ; AVX512BW-FCP-NEXT: movb $-64, %r8b
2210 ; AVX512BW-FCP-NEXT: kmovd %r8d, %k1
2211 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm5, %zmm4 {%k1}
2212 ; AVX512BW-FCP-NEXT: vmovdqa (%rsi), %xmm5
2213 ; AVX512BW-FCP-NEXT: vinserti128 $1, (%rcx), %ymm5, %ymm10
2214 ; AVX512BW-FCP-NEXT: vmovdqa (%rdi), %xmm5
2215 ; AVX512BW-FCP-NEXT: vinserti128 $1, (%rdx), %ymm5, %ymm12
2216 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm5 = ymm12[1],ymm10[1],ymm12[3],ymm10[3]
2217 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm5, %zmm4, %zmm16
2218 ; AVX512BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm5 = [6,14,6,14,6,14,6,14]
2219 ; AVX512BW-FCP-NEXT: # zmm5 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
2220 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm7, %zmm9
2221 ; AVX512BW-FCP-NEXT: vpermt2q %zmm11, %zmm5, %zmm9
2222 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm13
2223 ; AVX512BW-FCP-NEXT: vpermt2q %zmm8, %zmm5, %zmm13
2224 ; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm9 = ymm13[0,1,2,3],ymm9[4,5,6,7]
2225 ; AVX512BW-FCP-NEXT: vpermi2q %zmm2, %zmm0, %zmm5
2226 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm5 {%k1} = zmm1[0],zmm3[0],zmm1[2],zmm3[2],zmm1[4],zmm3[4],zmm1[6],zmm3[6]
2227 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm9, %zmm5, %zmm5
2228 ; AVX512BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm9 = [7,15,7,15,7,15,7,15]
2229 ; AVX512BW-FCP-NEXT: # zmm9 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
2230 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm7, %zmm13
2231 ; AVX512BW-FCP-NEXT: vpermt2q %zmm11, %zmm9, %zmm13
2232 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm14
2233 ; AVX512BW-FCP-NEXT: vpermt2q %zmm8, %zmm9, %zmm14
2234 ; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm13 = ymm14[0,1,2,3],ymm13[4,5,6,7]
2235 ; AVX512BW-FCP-NEXT: vpermi2q %zmm2, %zmm0, %zmm9
2236 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm9 {%k1} = zmm1[1],zmm3[1],zmm1[3],zmm3[3],zmm1[5],zmm3[5],zmm1[7],zmm3[7]
2237 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm13, %zmm9, %zmm9
2238 ; AVX512BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm13 = [4,12,4,12,4,12,4,12]
2239 ; AVX512BW-FCP-NEXT: # zmm13 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
2240 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm7, %zmm14
2241 ; AVX512BW-FCP-NEXT: vpermt2q %zmm11, %zmm13, %zmm14
2242 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm15
2243 ; AVX512BW-FCP-NEXT: vpermt2q %zmm8, %zmm13, %zmm15
2244 ; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm14 = ymm15[0,1,2,3],ymm14[4,5,6,7]
2245 ; AVX512BW-FCP-NEXT: vpermi2q %zmm3, %zmm1, %zmm13
2246 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm15 = zmm0[0],zmm2[0],zmm0[2],zmm2[2],zmm0[4],zmm2[4],zmm0[6],zmm2[6]
2247 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm13, %zmm15 {%k1}
2248 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm14, %zmm15, %zmm13
2249 ; AVX512BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm14 = [5,13,5,13,5,13,5,13]
2250 ; AVX512BW-FCP-NEXT: # zmm14 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
2251 ; AVX512BW-FCP-NEXT: vpermt2q %zmm11, %zmm14, %zmm7
2252 ; AVX512BW-FCP-NEXT: vpermt2q %zmm8, %zmm14, %zmm6
2253 ; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7]
2254 ; AVX512BW-FCP-NEXT: vpermi2q %zmm3, %zmm1, %zmm14
2255 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm7 = zmm0[1],zmm2[1],zmm0[3],zmm2[3],zmm0[5],zmm2[5],zmm0[7],zmm2[7]
2256 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm14, %zmm7 {%k1}
2257 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm7, %zmm6
2258 ; AVX512BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm7 = [0,8,0,8,0,8,0,8]
2259 ; AVX512BW-FCP-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
2260 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm8
2261 ; AVX512BW-FCP-NEXT: vpermt2q %zmm3, %zmm7, %zmm8
2262 ; AVX512BW-FCP-NEXT: vpermi2q %zmm2, %zmm0, %zmm7
2263 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm8, %zmm7 {%k1}
2264 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm8 = ymm12[0],ymm10[0],ymm12[2],ymm10[2]
2265 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm8, %zmm7, %zmm7
2266 ; AVX512BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm8 = [2,10,2,10,2,10,2,10]
2267 ; AVX512BW-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
2268 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm10
2269 ; AVX512BW-FCP-NEXT: vpermt2q %zmm3, %zmm8, %zmm10
2270 ; AVX512BW-FCP-NEXT: vpermi2q %zmm2, %zmm0, %zmm8
2271 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm10, %zmm8 {%k1}
2272 ; AVX512BW-FCP-NEXT: vmovdqa (%rcx), %ymm10
2273 ; AVX512BW-FCP-NEXT: vmovdqa (%rdx), %ymm11
2274 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm12 = ymm11[0],ymm10[0],ymm11[2],ymm10[2]
2275 ; AVX512BW-FCP-NEXT: vmovdqa (%rsi), %ymm14
2276 ; AVX512BW-FCP-NEXT: vmovdqa (%rdi), %ymm15
2277 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm4 = ymm15[0],ymm14[0],ymm15[2],ymm14[2]
2278 ; AVX512BW-FCP-NEXT: vperm2i128 {{.*#+}} ymm4 = ymm4[2,3],ymm12[2,3]
2279 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm4, %zmm8, %zmm4
2280 ; AVX512BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm8 = [3,11,3,11,3,11,3,11]
2281 ; AVX512BW-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
2282 ; AVX512BW-FCP-NEXT: vpermt2q %zmm3, %zmm8, %zmm1
2283 ; AVX512BW-FCP-NEXT: vpermt2q %zmm2, %zmm8, %zmm0
2284 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm0 {%k1}
2285 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm11[1],ymm10[1],ymm11[3],ymm10[3]
2286 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm2 = ymm15[1],ymm14[1],ymm15[3],ymm14[3]
2287 ; AVX512BW-FCP-NEXT: vperm2i128 {{.*#+}} ymm1 = ymm2[2,3],ymm1[2,3]
2288 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm1, %zmm0, %zmm0
2289 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, 192(%rax)
2290 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm4, 128(%rax)
2291 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm7, (%rax)
2292 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm6, 320(%rax)
2293 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm13, 256(%rax)
2294 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm9, 448(%rax)
2295 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm5, 384(%rax)
2296 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm16, 64(%rax)
2297 ; AVX512BW-FCP-NEXT: vzeroupper
2298 ; AVX512BW-FCP-NEXT: retq
2299 ;
2300 ; AVX512DQ-BW-LABEL: store_i64_stride8_vf8:
2301 ; AVX512DQ-BW: # %bb.0:
2302 ; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
2303 ; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %r10
2304 ; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %r11
2305 ; AVX512DQ-BW-NEXT: vmovdqa64 (%rdi), %zmm6
2306 ; AVX512DQ-BW-NEXT: vmovdqa64 (%rsi), %zmm8
2307 ; AVX512DQ-BW-NEXT: vmovdqa64 (%rdx), %zmm7
2308 ; AVX512DQ-BW-NEXT: vmovdqa64 (%rcx), %zmm11
2309 ; AVX512DQ-BW-NEXT: vmovdqa64 (%r8), %zmm0
2310 ; AVX512DQ-BW-NEXT: vmovdqa64 (%r9), %zmm2
2311 ; AVX512DQ-BW-NEXT: vmovdqa64 (%r11), %zmm1
2312 ; AVX512DQ-BW-NEXT: vmovdqa64 (%r10), %zmm3
2313 ; AVX512DQ-BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm4 = [1,9,1,9,1,9,1,9]
2314 ; AVX512DQ-BW-NEXT: # zmm4 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
2315 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm1, %zmm5
2316 ; AVX512DQ-BW-NEXT: vpermt2q %zmm3, %zmm4, %zmm5
2317 ; AVX512DQ-BW-NEXT: vpermi2q %zmm2, %zmm0, %zmm4
2318 ; AVX512DQ-BW-NEXT: movb $-64, %r8b
2319 ; AVX512DQ-BW-NEXT: kmovd %r8d, %k1
2320 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm5, %zmm4 {%k1}
2321 ; AVX512DQ-BW-NEXT: vmovdqa (%rsi), %xmm5
2322 ; AVX512DQ-BW-NEXT: vinserti128 $1, (%rcx), %ymm5, %ymm10
2323 ; AVX512DQ-BW-NEXT: vmovdqa (%rdi), %xmm5
2324 ; AVX512DQ-BW-NEXT: vinserti128 $1, (%rdx), %ymm5, %ymm12
2325 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} ymm5 = ymm12[1],ymm10[1],ymm12[3],ymm10[3]
2326 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm5, %zmm4, %zmm16
2327 ; AVX512DQ-BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm5 = [6,14,6,14,6,14,6,14]
2328 ; AVX512DQ-BW-NEXT: # zmm5 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
2329 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm7, %zmm9
2330 ; AVX512DQ-BW-NEXT: vpermt2q %zmm11, %zmm5, %zmm9
2331 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm6, %zmm13
2332 ; AVX512DQ-BW-NEXT: vpermt2q %zmm8, %zmm5, %zmm13
2333 ; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm9 = ymm13[0,1,2,3],ymm9[4,5,6,7]
2334 ; AVX512DQ-BW-NEXT: vpermi2q %zmm2, %zmm0, %zmm5
2335 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} zmm5 {%k1} = zmm1[0],zmm3[0],zmm1[2],zmm3[2],zmm1[4],zmm3[4],zmm1[6],zmm3[6]
2336 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm9, %zmm5, %zmm5
2337 ; AVX512DQ-BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm9 = [7,15,7,15,7,15,7,15]
2338 ; AVX512DQ-BW-NEXT: # zmm9 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
2339 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm7, %zmm13
2340 ; AVX512DQ-BW-NEXT: vpermt2q %zmm11, %zmm9, %zmm13
2341 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm6, %zmm14
2342 ; AVX512DQ-BW-NEXT: vpermt2q %zmm8, %zmm9, %zmm14
2343 ; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm13 = ymm14[0,1,2,3],ymm13[4,5,6,7]
2344 ; AVX512DQ-BW-NEXT: vpermi2q %zmm2, %zmm0, %zmm9
2345 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} zmm9 {%k1} = zmm1[1],zmm3[1],zmm1[3],zmm3[3],zmm1[5],zmm3[5],zmm1[7],zmm3[7]
2346 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm13, %zmm9, %zmm9
2347 ; AVX512DQ-BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm13 = [4,12,4,12,4,12,4,12]
2348 ; AVX512DQ-BW-NEXT: # zmm13 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
2349 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm7, %zmm14
2350 ; AVX512DQ-BW-NEXT: vpermt2q %zmm11, %zmm13, %zmm14
2351 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm6, %zmm15
2352 ; AVX512DQ-BW-NEXT: vpermt2q %zmm8, %zmm13, %zmm15
2353 ; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm14 = ymm15[0,1,2,3],ymm14[4,5,6,7]
2354 ; AVX512DQ-BW-NEXT: vpermi2q %zmm3, %zmm1, %zmm13
2355 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} zmm15 = zmm0[0],zmm2[0],zmm0[2],zmm2[2],zmm0[4],zmm2[4],zmm0[6],zmm2[6]
2356 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm13, %zmm15 {%k1}
2357 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm14, %zmm15, %zmm13
2358 ; AVX512DQ-BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm14 = [5,13,5,13,5,13,5,13]
2359 ; AVX512DQ-BW-NEXT: # zmm14 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
2360 ; AVX512DQ-BW-NEXT: vpermt2q %zmm11, %zmm14, %zmm7
2361 ; AVX512DQ-BW-NEXT: vpermt2q %zmm8, %zmm14, %zmm6
2362 ; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7]
2363 ; AVX512DQ-BW-NEXT: vpermi2q %zmm3, %zmm1, %zmm14
2364 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} zmm7 = zmm0[1],zmm2[1],zmm0[3],zmm2[3],zmm0[5],zmm2[5],zmm0[7],zmm2[7]
2365 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm14, %zmm7 {%k1}
2366 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm6, %zmm7, %zmm6
2367 ; AVX512DQ-BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm7 = [0,8,0,8,0,8,0,8]
2368 ; AVX512DQ-BW-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
2369 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm1, %zmm8
2370 ; AVX512DQ-BW-NEXT: vpermt2q %zmm3, %zmm7, %zmm8
2371 ; AVX512DQ-BW-NEXT: vpermi2q %zmm2, %zmm0, %zmm7
2372 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm8, %zmm7 {%k1}
2373 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} ymm8 = ymm12[0],ymm10[0],ymm12[2],ymm10[2]
2374 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm8, %zmm7, %zmm7
2375 ; AVX512DQ-BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm8 = [2,10,2,10,2,10,2,10]
2376 ; AVX512DQ-BW-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
2377 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm1, %zmm10
2378 ; AVX512DQ-BW-NEXT: vpermt2q %zmm3, %zmm8, %zmm10
2379 ; AVX512DQ-BW-NEXT: vpermi2q %zmm2, %zmm0, %zmm8
2380 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm10, %zmm8 {%k1}
2381 ; AVX512DQ-BW-NEXT: vmovdqa (%rcx), %ymm10
2382 ; AVX512DQ-BW-NEXT: vmovdqa (%rdx), %ymm11
2383 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} ymm12 = ymm11[0],ymm10[0],ymm11[2],ymm10[2]
2384 ; AVX512DQ-BW-NEXT: vmovdqa (%rsi), %ymm14
2385 ; AVX512DQ-BW-NEXT: vmovdqa (%rdi), %ymm15
2386 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} ymm4 = ymm15[0],ymm14[0],ymm15[2],ymm14[2]
2387 ; AVX512DQ-BW-NEXT: vperm2i128 {{.*#+}} ymm4 = ymm4[2,3],ymm12[2,3]
2388 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm4, %zmm8, %zmm4
2389 ; AVX512DQ-BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm8 = [3,11,3,11,3,11,3,11]
2390 ; AVX512DQ-BW-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
2391 ; AVX512DQ-BW-NEXT: vpermt2q %zmm3, %zmm8, %zmm1
2392 ; AVX512DQ-BW-NEXT: vpermt2q %zmm2, %zmm8, %zmm0
2393 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm1, %zmm0 {%k1}
2394 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm11[1],ymm10[1],ymm11[3],ymm10[3]
2395 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} ymm2 = ymm15[1],ymm14[1],ymm15[3],ymm14[3]
2396 ; AVX512DQ-BW-NEXT: vperm2i128 {{.*#+}} ymm1 = ymm2[2,3],ymm1[2,3]
2397 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm1, %zmm0, %zmm0
2398 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, 192(%rax)
2399 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm4, 128(%rax)
2400 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm7, (%rax)
2401 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm6, 320(%rax)
2402 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm13, 256(%rax)
2403 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm9, 448(%rax)
2404 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm5, 384(%rax)
2405 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm16, 64(%rax)
2406 ; AVX512DQ-BW-NEXT: vzeroupper
2407 ; AVX512DQ-BW-NEXT: retq
2408 ;
2409 ; AVX512DQ-BW-FCP-LABEL: store_i64_stride8_vf8:
2410 ; AVX512DQ-BW-FCP: # %bb.0:
2411 ; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
2412 ; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10
2413 ; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r11
2414 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rdi), %zmm6
2415 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rsi), %zmm8
2416 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rdx), %zmm7
2417 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rcx), %zmm11
2418 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%r8), %zmm0
2419 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%r9), %zmm2
2420 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%r11), %zmm1
2421 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%r10), %zmm3
2422 ; AVX512DQ-BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm4 = [1,9,1,9,1,9,1,9]
2423 ; AVX512DQ-BW-FCP-NEXT: # zmm4 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
2424 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm5
2425 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm3, %zmm4, %zmm5
2426 ; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm2, %zmm0, %zmm4
2427 ; AVX512DQ-BW-FCP-NEXT: movb $-64, %r8b
2428 ; AVX512DQ-BW-FCP-NEXT: kmovd %r8d, %k1
2429 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm5, %zmm4 {%k1}
2430 ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rsi), %xmm5
2431 ; AVX512DQ-BW-FCP-NEXT: vinserti128 $1, (%rcx), %ymm5, %ymm10
2432 ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdi), %xmm5
2433 ; AVX512DQ-BW-FCP-NEXT: vinserti128 $1, (%rdx), %ymm5, %ymm12
2434 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm5 = ymm12[1],ymm10[1],ymm12[3],ymm10[3]
2435 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm5, %zmm4, %zmm16
2436 ; AVX512DQ-BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm5 = [6,14,6,14,6,14,6,14]
2437 ; AVX512DQ-BW-FCP-NEXT: # zmm5 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
2438 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm7, %zmm9
2439 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm11, %zmm5, %zmm9
2440 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm13
2441 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm8, %zmm5, %zmm13
2442 ; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm9 = ymm13[0,1,2,3],ymm9[4,5,6,7]
2443 ; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm2, %zmm0, %zmm5
2444 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm5 {%k1} = zmm1[0],zmm3[0],zmm1[2],zmm3[2],zmm1[4],zmm3[4],zmm1[6],zmm3[6]
2445 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm9, %zmm5, %zmm5
2446 ; AVX512DQ-BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm9 = [7,15,7,15,7,15,7,15]
2447 ; AVX512DQ-BW-FCP-NEXT: # zmm9 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
2448 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm7, %zmm13
2449 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm11, %zmm9, %zmm13
2450 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm14
2451 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm8, %zmm9, %zmm14
2452 ; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm13 = ymm14[0,1,2,3],ymm13[4,5,6,7]
2453 ; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm2, %zmm0, %zmm9
2454 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm9 {%k1} = zmm1[1],zmm3[1],zmm1[3],zmm3[3],zmm1[5],zmm3[5],zmm1[7],zmm3[7]
2455 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm13, %zmm9, %zmm9
2456 ; AVX512DQ-BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm13 = [4,12,4,12,4,12,4,12]
2457 ; AVX512DQ-BW-FCP-NEXT: # zmm13 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
2458 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm7, %zmm14
2459 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm11, %zmm13, %zmm14
2460 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm15
2461 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm8, %zmm13, %zmm15
2462 ; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm14 = ymm15[0,1,2,3],ymm14[4,5,6,7]
2463 ; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm3, %zmm1, %zmm13
2464 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm15 = zmm0[0],zmm2[0],zmm0[2],zmm2[2],zmm0[4],zmm2[4],zmm0[6],zmm2[6]
2465 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm13, %zmm15 {%k1}
2466 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm14, %zmm15, %zmm13
2467 ; AVX512DQ-BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm14 = [5,13,5,13,5,13,5,13]
2468 ; AVX512DQ-BW-FCP-NEXT: # zmm14 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
2469 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm11, %zmm14, %zmm7
2470 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm8, %zmm14, %zmm6
2471 ; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7]
2472 ; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm3, %zmm1, %zmm14
2473 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm7 = zmm0[1],zmm2[1],zmm0[3],zmm2[3],zmm0[5],zmm2[5],zmm0[7],zmm2[7]
2474 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm14, %zmm7 {%k1}
2475 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm7, %zmm6
2476 ; AVX512DQ-BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm7 = [0,8,0,8,0,8,0,8]
2477 ; AVX512DQ-BW-FCP-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
2478 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm8
2479 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm3, %zmm7, %zmm8
2480 ; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm2, %zmm0, %zmm7
2481 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm8, %zmm7 {%k1}
2482 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm8 = ymm12[0],ymm10[0],ymm12[2],ymm10[2]
2483 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm8, %zmm7, %zmm7
2484 ; AVX512DQ-BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm8 = [2,10,2,10,2,10,2,10]
2485 ; AVX512DQ-BW-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
2486 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm10
2487 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm3, %zmm8, %zmm10
2488 ; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm2, %zmm0, %zmm8
2489 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm10, %zmm8 {%k1}
2490 ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rcx), %ymm10
2491 ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdx), %ymm11
2492 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm12 = ymm11[0],ymm10[0],ymm11[2],ymm10[2]
2493 ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rsi), %ymm14
2494 ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdi), %ymm15
2495 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm4 = ymm15[0],ymm14[0],ymm15[2],ymm14[2]
2496 ; AVX512DQ-BW-FCP-NEXT: vperm2i128 {{.*#+}} ymm4 = ymm4[2,3],ymm12[2,3]
2497 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm4, %zmm8, %zmm4
2498 ; AVX512DQ-BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm8 = [3,11,3,11,3,11,3,11]
2499 ; AVX512DQ-BW-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
2500 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm3, %zmm8, %zmm1
2501 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm2, %zmm8, %zmm0
2502 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm0 {%k1}
2503 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm11[1],ymm10[1],ymm11[3],ymm10[3]
2504 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm2 = ymm15[1],ymm14[1],ymm15[3],ymm14[3]
2505 ; AVX512DQ-BW-FCP-NEXT: vperm2i128 {{.*#+}} ymm1 = ymm2[2,3],ymm1[2,3]
2506 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm1, %zmm0, %zmm0
2507 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, 192(%rax)
2508 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm4, 128(%rax)
2509 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm7, (%rax)
2510 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm6, 320(%rax)
2511 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm13, 256(%rax)
2512 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm9, 448(%rax)
2513 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm5, 384(%rax)
2514 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm16, 64(%rax)
2515 ; AVX512DQ-BW-FCP-NEXT: vzeroupper
2516 ; AVX512DQ-BW-FCP-NEXT: retq
2517 %in.vec0 = load <8 x i64>, ptr %in.vecptr0, align 64
2518 %in.vec1 = load <8 x i64>, ptr %in.vecptr1, align 64
2519 %in.vec2 = load <8 x i64>, ptr %in.vecptr2, align 64
2520 %in.vec3 = load <8 x i64>, ptr %in.vecptr3, align 64
2521 %in.vec4 = load <8 x i64>, ptr %in.vecptr4, align 64
2522 %in.vec5 = load <8 x i64>, ptr %in.vecptr5, align 64
2523 %in.vec6 = load <8 x i64>, ptr %in.vecptr6, align 64
2524 %in.vec7 = load <8 x i64>, ptr %in.vecptr7, align 64
2525 %1 = shufflevector <8 x i64> %in.vec0, <8 x i64> %in.vec1, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
2526 %2 = shufflevector <8 x i64> %in.vec2, <8 x i64> %in.vec3, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
2527 %3 = shufflevector <8 x i64> %in.vec4, <8 x i64> %in.vec5, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
2528 %4 = shufflevector <8 x i64> %in.vec6, <8 x i64> %in.vec7, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
2529 %5 = shufflevector <16 x i64> %1, <16 x i64> %2, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
2530 %6 = shufflevector <16 x i64> %3, <16 x i64> %4, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
2531 %7 = shufflevector <32 x i64> %5, <32 x i64> %6, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
2532 %interleaved.vec = shufflevector <64 x i64> %7, <64 x i64> poison, <64 x i32> <i32 0, i32 8, i32 16, i32 24, i32 32, i32 40, i32 48, i32 56, i32 1, i32 9, i32 17, i32 25, i32 33, i32 41, i32 49, i32 57, i32 2, i32 10, i32 18, i32 26, i32 34, i32 42, i32 50, i32 58, i32 3, i32 11, i32 19, i32 27, i32 35, i32 43, i32 51, i32 59, i32 4, i32 12, i32 20, i32 28, i32 36, i32 44, i32 52, i32 60, i32 5, i32 13, i32 21, i32 29, i32 37, i32 45, i32 53, i32 61, i32 6, i32 14, i32 22, i32 30, i32 38, i32 46, i32 54, i32 62, i32 7, i32 15, i32 23, i32 31, i32 39, i32 47, i32 55, i32 63>
2533 store <64 x i64> %interleaved.vec, ptr %out.vec, align 64
2534 ret void
2535 }
2537 define void @store_i64_stride8_vf16(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecptr2, ptr %in.vecptr3, ptr %in.vecptr4, ptr %in.vecptr5, ptr %in.vecptr6, ptr %in.vecptr7, ptr %out.vec) nounwind {
2538 ; SSE-LABEL: store_i64_stride8_vf16:
2539 ; SSE: # %bb.0:
2540 ; SSE-NEXT: subq $664, %rsp # imm = 0x298
2541 ; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax
2542 ; SSE-NEXT: movq {{[0-9]+}}(%rsp), %r10
2543 ; SSE-NEXT: movaps (%rdi), %xmm7
2544 ; SSE-NEXT: movaps 16(%rdi), %xmm8
2545 ; SSE-NEXT: movaps (%rsi), %xmm2
2546 ; SSE-NEXT: movaps 16(%rsi), %xmm0
2547 ; SSE-NEXT: movaps (%rdx), %xmm9
2548 ; SSE-NEXT: movaps 16(%rdx), %xmm10
2549 ; SSE-NEXT: movaps (%rcx), %xmm4
2550 ; SSE-NEXT: movaps 16(%rcx), %xmm1
2551 ; SSE-NEXT: movaps (%r8), %xmm11
2552 ; SSE-NEXT: movaps 16(%r8), %xmm12
2553 ; SSE-NEXT: movaps (%r9), %xmm5
2554 ; SSE-NEXT: movaps 16(%r9), %xmm3
2555 ; SSE-NEXT: movaps (%r10), %xmm13
2556 ; SSE-NEXT: movaps 16(%r10), %xmm15
2557 ; SSE-NEXT: movaps (%rax), %xmm6
2558 ; SSE-NEXT: movaps %xmm7, %xmm14
2559 ; SSE-NEXT: movlhps {{.*#+}} xmm14 = xmm14[0],xmm2[0]
2560 ; SSE-NEXT: movaps %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2561 ; SSE-NEXT: unpckhpd {{.*#+}} xmm7 = xmm7[1],xmm2[1]
2562 ; SSE-NEXT: movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2563 ; SSE-NEXT: movaps %xmm9, %xmm2
2564 ; SSE-NEXT: movlhps {{.*#+}} xmm2 = xmm2[0],xmm4[0]
2565 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2566 ; SSE-NEXT: unpckhpd {{.*#+}} xmm9 = xmm9[1],xmm4[1]
2567 ; SSE-NEXT: movaps %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2568 ; SSE-NEXT: movaps %xmm11, %xmm2
2569 ; SSE-NEXT: movlhps {{.*#+}} xmm2 = xmm2[0],xmm5[0]
2570 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2571 ; SSE-NEXT: unpckhpd {{.*#+}} xmm11 = xmm11[1],xmm5[1]
2572 ; SSE-NEXT: movaps %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2573 ; SSE-NEXT: movaps %xmm13, %xmm4
2574 ; SSE-NEXT: movlhps {{.*#+}} xmm4 = xmm4[0],xmm6[0]
2575 ; SSE-NEXT: movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2576 ; SSE-NEXT: unpckhpd {{.*#+}} xmm13 = xmm13[1],xmm6[1]
2577 ; SSE-NEXT: movaps %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2578 ; SSE-NEXT: movaps %xmm8, %xmm2
2579 ; SSE-NEXT: movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]
2580 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2581 ; SSE-NEXT: unpckhpd {{.*#+}} xmm8 = xmm8[1],xmm0[1]
2582 ; SSE-NEXT: movaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2583 ; SSE-NEXT: movaps %xmm10, %xmm0
2584 ; SSE-NEXT: movlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0]
2585 ; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2586 ; SSE-NEXT: unpckhpd {{.*#+}} xmm10 = xmm10[1],xmm1[1]
2587 ; SSE-NEXT: movaps %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2588 ; SSE-NEXT: movaps %xmm12, %xmm0
2589 ; SSE-NEXT: movlhps {{.*#+}} xmm0 = xmm0[0],xmm3[0]
2590 ; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2591 ; SSE-NEXT: unpckhpd {{.*#+}} xmm12 = xmm12[1],xmm3[1]
2592 ; SSE-NEXT: movaps %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2593 ; SSE-NEXT: movaps 16(%rax), %xmm0
2594 ; SSE-NEXT: movaps %xmm15, %xmm1
2595 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
2596 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2597 ; SSE-NEXT: unpckhpd {{.*#+}} xmm15 = xmm15[1],xmm0[1]
2598 ; SSE-NEXT: movaps %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2599 ; SSE-NEXT: movaps 32(%rdi), %xmm2
2600 ; SSE-NEXT: movaps 32(%rsi), %xmm0
2601 ; SSE-NEXT: movaps %xmm2, %xmm1
2602 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
2603 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2604 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
2605 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2606 ; SSE-NEXT: movaps 32(%rdx), %xmm2
2607 ; SSE-NEXT: movaps 32(%rcx), %xmm0
2608 ; SSE-NEXT: movaps %xmm2, %xmm1
2609 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
2610 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2611 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
2612 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2613 ; SSE-NEXT: movaps 32(%r8), %xmm2
2614 ; SSE-NEXT: movaps 32(%r9), %xmm0
2615 ; SSE-NEXT: movaps %xmm2, %xmm1
2616 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
2617 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2618 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
2619 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2620 ; SSE-NEXT: movaps 32(%r10), %xmm2
2621 ; SSE-NEXT: movaps 32(%rax), %xmm0
2622 ; SSE-NEXT: movaps %xmm2, %xmm1
2623 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
2624 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2625 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
2626 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2627 ; SSE-NEXT: movaps 48(%rdi), %xmm2
2628 ; SSE-NEXT: movaps 48(%rsi), %xmm0
2629 ; SSE-NEXT: movaps %xmm2, %xmm1
2630 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
2631 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2632 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
2633 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2634 ; SSE-NEXT: movaps 48(%rdx), %xmm2
2635 ; SSE-NEXT: movaps 48(%rcx), %xmm0
2636 ; SSE-NEXT: movaps %xmm2, %xmm1
2637 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
2638 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2639 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
2640 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2641 ; SSE-NEXT: movaps 48(%r8), %xmm2
2642 ; SSE-NEXT: movaps 48(%r9), %xmm0
2643 ; SSE-NEXT: movaps %xmm2, %xmm1
2644 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
2645 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2646 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
2647 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2648 ; SSE-NEXT: movaps 48(%r10), %xmm2
2649 ; SSE-NEXT: movaps 48(%rax), %xmm0
2650 ; SSE-NEXT: movaps %xmm2, %xmm1
2651 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
2652 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2653 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
2654 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2655 ; SSE-NEXT: movaps 64(%rdi), %xmm2
2656 ; SSE-NEXT: movaps 64(%rsi), %xmm0
2657 ; SSE-NEXT: movaps %xmm2, %xmm1
2658 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
2659 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2660 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
2661 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2662 ; SSE-NEXT: movaps 64(%rdx), %xmm2
2663 ; SSE-NEXT: movaps 64(%rcx), %xmm0
2664 ; SSE-NEXT: movaps %xmm2, %xmm1
2665 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
2666 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2667 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
2668 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2669 ; SSE-NEXT: movaps 64(%r8), %xmm2
2670 ; SSE-NEXT: movaps 64(%r9), %xmm0
2671 ; SSE-NEXT: movaps %xmm2, %xmm1
2672 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
2673 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2674 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
2675 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2676 ; SSE-NEXT: movaps 64(%r10), %xmm2
2677 ; SSE-NEXT: movaps 64(%rax), %xmm0
2678 ; SSE-NEXT: movaps %xmm2, %xmm1
2679 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
2680 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2681 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
2682 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2683 ; SSE-NEXT: movaps 80(%rdi), %xmm2
2684 ; SSE-NEXT: movaps 80(%rsi), %xmm0
2685 ; SSE-NEXT: movaps %xmm2, %xmm1
2686 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
2687 ; SSE-NEXT: movaps %xmm1, (%rsp) # 16-byte Spill
2688 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
2689 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2690 ; SSE-NEXT: movaps 80(%rdx), %xmm2
2691 ; SSE-NEXT: movaps 80(%rcx), %xmm0
2692 ; SSE-NEXT: movaps %xmm2, %xmm1
2693 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
2694 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2695 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
2696 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2697 ; SSE-NEXT: movaps 80(%r8), %xmm2
2698 ; SSE-NEXT: movaps 80(%r9), %xmm0
2699 ; SSE-NEXT: movaps %xmm2, %xmm1
2700 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
2701 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2702 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
2703 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2704 ; SSE-NEXT: movaps 80(%r10), %xmm2
2705 ; SSE-NEXT: movaps 80(%rax), %xmm0
2706 ; SSE-NEXT: movaps %xmm2, %xmm1
2707 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
2708 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2709 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
2710 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2711 ; SSE-NEXT: movaps 96(%rdi), %xmm13
2712 ; SSE-NEXT: movaps 96(%rsi), %xmm0
2713 ; SSE-NEXT: movaps %xmm13, %xmm1
2714 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
2715 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
2716 ; SSE-NEXT: unpckhpd {{.*#+}} xmm13 = xmm13[1],xmm0[1]
2717 ; SSE-NEXT: movaps 96(%rdx), %xmm10
2718 ; SSE-NEXT: movaps 96(%rcx), %xmm0
2719 ; SSE-NEXT: movaps %xmm10, %xmm15
2720 ; SSE-NEXT: movlhps {{.*#+}} xmm15 = xmm15[0],xmm0[0]
2721 ; SSE-NEXT: unpckhpd {{.*#+}} xmm10 = xmm10[1],xmm0[1]
2722 ; SSE-NEXT: movaps 96(%r8), %xmm11
2723 ; SSE-NEXT: movaps 96(%r9), %xmm0
2724 ; SSE-NEXT: movaps %xmm11, %xmm14
2725 ; SSE-NEXT: movlhps {{.*#+}} xmm14 = xmm14[0],xmm0[0]
2726 ; SSE-NEXT: unpckhpd {{.*#+}} xmm11 = xmm11[1],xmm0[1]
2727 ; SSE-NEXT: movaps 96(%r10), %xmm9
2728 ; SSE-NEXT: movaps 96(%rax), %xmm0
2729 ; SSE-NEXT: movaps %xmm9, %xmm12
2730 ; SSE-NEXT: movlhps {{.*#+}} xmm12 = xmm12[0],xmm0[0]
2731 ; SSE-NEXT: unpckhpd {{.*#+}} xmm9 = xmm9[1],xmm0[1]
2732 ; SSE-NEXT: movaps 112(%rdi), %xmm7
2733 ; SSE-NEXT: movaps 112(%rsi), %xmm0
2734 ; SSE-NEXT: movaps %xmm7, %xmm8
2735 ; SSE-NEXT: movlhps {{.*#+}} xmm8 = xmm8[0],xmm0[0]
2736 ; SSE-NEXT: unpckhpd {{.*#+}} xmm7 = xmm7[1],xmm0[1]
2737 ; SSE-NEXT: movaps 112(%rdx), %xmm5
2738 ; SSE-NEXT: movaps 112(%rcx), %xmm1
2739 ; SSE-NEXT: movaps %xmm5, %xmm6
2740 ; SSE-NEXT: movlhps {{.*#+}} xmm6 = xmm6[0],xmm1[0]
2741 ; SSE-NEXT: unpckhpd {{.*#+}} xmm5 = xmm5[1],xmm1[1]
2742 ; SSE-NEXT: movaps 112(%r8), %xmm1
2743 ; SSE-NEXT: movaps 112(%r9), %xmm2
2744 ; SSE-NEXT: movaps %xmm1, %xmm4
2745 ; SSE-NEXT: movlhps {{.*#+}} xmm4 = xmm4[0],xmm2[0]
2746 ; SSE-NEXT: unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm2[1]
2747 ; SSE-NEXT: movaps 112(%r10), %xmm2
2748 ; SSE-NEXT: movaps 112(%rax), %xmm3
2749 ; SSE-NEXT: movaps %xmm2, %xmm0
2750 ; SSE-NEXT: movlhps {{.*#+}} xmm0 = xmm0[0],xmm3[0]
2751 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm3[1]
2752 ; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax
2753 ; SSE-NEXT: movaps %xmm2, 1008(%rax)
2754 ; SSE-NEXT: movaps %xmm1, 992(%rax)
2755 ; SSE-NEXT: movaps %xmm5, 976(%rax)
2756 ; SSE-NEXT: movaps %xmm7, 960(%rax)
2757 ; SSE-NEXT: movaps %xmm0, 944(%rax)
2758 ; SSE-NEXT: movaps %xmm4, 928(%rax)
2759 ; SSE-NEXT: movaps %xmm6, 912(%rax)
2760 ; SSE-NEXT: movaps %xmm8, 896(%rax)
2761 ; SSE-NEXT: movaps %xmm9, 880(%rax)
2762 ; SSE-NEXT: movaps %xmm11, 864(%rax)
2763 ; SSE-NEXT: movaps %xmm10, 848(%rax)
2764 ; SSE-NEXT: movaps %xmm13, 832(%rax)
2765 ; SSE-NEXT: movaps %xmm12, 816(%rax)
2766 ; SSE-NEXT: movaps %xmm14, 800(%rax)
2767 ; SSE-NEXT: movaps %xmm15, 784(%rax)
2768 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2769 ; SSE-NEXT: movaps %xmm0, 768(%rax)
2770 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2771 ; SSE-NEXT: movaps %xmm0, 752(%rax)
2772 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2773 ; SSE-NEXT: movaps %xmm0, 736(%rax)
2774 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2775 ; SSE-NEXT: movaps %xmm0, 720(%rax)
2776 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2777 ; SSE-NEXT: movaps %xmm0, 704(%rax)
2778 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2779 ; SSE-NEXT: movaps %xmm0, 688(%rax)
2780 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2781 ; SSE-NEXT: movaps %xmm0, 672(%rax)
2782 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2783 ; SSE-NEXT: movaps %xmm0, 656(%rax)
2784 ; SSE-NEXT: movaps (%rsp), %xmm0 # 16-byte Reload
2785 ; SSE-NEXT: movaps %xmm0, 640(%rax)
2786 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2787 ; SSE-NEXT: movaps %xmm0, 624(%rax)
2788 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2789 ; SSE-NEXT: movaps %xmm0, 608(%rax)
2790 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2791 ; SSE-NEXT: movaps %xmm0, 592(%rax)
2792 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2793 ; SSE-NEXT: movaps %xmm0, 576(%rax)
2794 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2795 ; SSE-NEXT: movaps %xmm0, 560(%rax)
2796 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2797 ; SSE-NEXT: movaps %xmm0, 544(%rax)
2798 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2799 ; SSE-NEXT: movaps %xmm0, 528(%rax)
2800 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2801 ; SSE-NEXT: movaps %xmm0, 512(%rax)
2802 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2803 ; SSE-NEXT: movaps %xmm0, 496(%rax)
2804 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2805 ; SSE-NEXT: movaps %xmm0, 480(%rax)
2806 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2807 ; SSE-NEXT: movaps %xmm0, 464(%rax)
2808 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2809 ; SSE-NEXT: movaps %xmm0, 448(%rax)
2810 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2811 ; SSE-NEXT: movaps %xmm0, 432(%rax)
2812 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2813 ; SSE-NEXT: movaps %xmm0, 416(%rax)
2814 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2815 ; SSE-NEXT: movaps %xmm0, 400(%rax)
2816 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2817 ; SSE-NEXT: movaps %xmm0, 384(%rax)
2818 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2819 ; SSE-NEXT: movaps %xmm0, 368(%rax)
2820 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2821 ; SSE-NEXT: movaps %xmm0, 352(%rax)
2822 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2823 ; SSE-NEXT: movaps %xmm0, 336(%rax)
2824 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2825 ; SSE-NEXT: movaps %xmm0, 320(%rax)
2826 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2827 ; SSE-NEXT: movaps %xmm0, 304(%rax)
2828 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2829 ; SSE-NEXT: movaps %xmm0, 288(%rax)
2830 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2831 ; SSE-NEXT: movaps %xmm0, 272(%rax)
2832 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2833 ; SSE-NEXT: movaps %xmm0, 256(%rax)
2834 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2835 ; SSE-NEXT: movaps %xmm0, 240(%rax)
2836 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2837 ; SSE-NEXT: movaps %xmm0, 224(%rax)
2838 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2839 ; SSE-NEXT: movaps %xmm0, 208(%rax)
2840 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2841 ; SSE-NEXT: movaps %xmm0, 192(%rax)
2842 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2843 ; SSE-NEXT: movaps %xmm0, 176(%rax)
2844 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2845 ; SSE-NEXT: movaps %xmm0, 160(%rax)
2846 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2847 ; SSE-NEXT: movaps %xmm0, 144(%rax)
2848 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2849 ; SSE-NEXT: movaps %xmm0, 128(%rax)
2850 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2851 ; SSE-NEXT: movaps %xmm0, 112(%rax)
2852 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2853 ; SSE-NEXT: movaps %xmm0, 96(%rax)
2854 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2855 ; SSE-NEXT: movaps %xmm0, 80(%rax)
2856 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2857 ; SSE-NEXT: movaps %xmm0, 64(%rax)
2858 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2859 ; SSE-NEXT: movaps %xmm0, 48(%rax)
2860 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2861 ; SSE-NEXT: movaps %xmm0, 32(%rax)
2862 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2863 ; SSE-NEXT: movaps %xmm0, 16(%rax)
2864 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
2865 ; SSE-NEXT: movaps %xmm0, (%rax)
2866 ; SSE-NEXT: addq $664, %rsp # imm = 0x298
2867 ; SSE-NEXT: retq
2868 ;
2869 ; AVX-LABEL: store_i64_stride8_vf16:
2870 ; AVX: # %bb.0:
2871 ; AVX-NEXT: subq $488, %rsp # imm = 0x1E8
2872 ; AVX-NEXT: movq {{[0-9]+}}(%rsp), %r10
2873 ; AVX-NEXT: movq {{[0-9]+}}(%rsp), %rax
2874 ; AVX-NEXT: vmovaps (%rsi), %xmm4
2875 ; AVX-NEXT: vmovaps (%rdi), %xmm5
2876 ; AVX-NEXT: vmovlhps {{.*#+}} xmm1 = xmm5[0],xmm4[0]
2877 ; AVX-NEXT: vmovaps (%rcx), %xmm6
2878 ; AVX-NEXT: vmovaps 32(%rcx), %xmm2
2879 ; AVX-NEXT: vmovaps 64(%rcx), %xmm0
2880 ; AVX-NEXT: vinsertf128 $1, %xmm6, %ymm1, %ymm3
2881 ; AVX-NEXT: vinsertf128 $1, (%rdx), %ymm1, %ymm1
2882 ; AVX-NEXT: vshufpd {{.*#+}} ymm1 = ymm1[0],ymm3[1],ymm1[2],ymm3[2]
2883 ; AVX-NEXT: vmovupd %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2884 ; AVX-NEXT: vmovaps (%r9), %xmm7
2885 ; AVX-NEXT: vmovaps (%r8), %xmm8
2886 ; AVX-NEXT: vmovlhps {{.*#+}} xmm9 = xmm8[0],xmm7[0]
2887 ; AVX-NEXT: vmovaps (%rax), %xmm10
2888 ; AVX-NEXT: vmovaps 32(%rax), %xmm3
2889 ; AVX-NEXT: vmovaps 64(%rax), %xmm1
2890 ; AVX-NEXT: vinsertf128 $1, %xmm10, %ymm9, %ymm11
2891 ; AVX-NEXT: vinsertf128 $1, (%r10), %ymm9, %ymm9
2892 ; AVX-NEXT: vshufpd {{.*#+}} ymm9 = ymm9[0],ymm11[1],ymm9[2],ymm11[2]
2893 ; AVX-NEXT: vmovupd %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2894 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm4 = xmm5[1],xmm4[1]
2895 ; AVX-NEXT: vinsertf128 $1, %xmm6, %ymm4, %ymm5
2896 ; AVX-NEXT: vbroadcastsd 8(%rdx), %ymm6
2897 ; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm6[4,5,6,7]
2898 ; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1],ymm5[2,3],ymm4[4,5],ymm5[6,7]
2899 ; AVX-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2900 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm4 = xmm8[1],xmm7[1]
2901 ; AVX-NEXT: vinsertf128 $1, %xmm10, %ymm4, %ymm5
2902 ; AVX-NEXT: vbroadcastsd 8(%r10), %ymm6
2903 ; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm6[4,5,6,7]
2904 ; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1],ymm5[2,3],ymm4[4,5],ymm5[6,7]
2905 ; AVX-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2906 ; AVX-NEXT: vmovaps 32(%rsi), %xmm4
2907 ; AVX-NEXT: vmovaps 32(%rdi), %xmm5
2908 ; AVX-NEXT: vmovlhps {{.*#+}} xmm6 = xmm5[0],xmm4[0]
2909 ; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm6, %ymm7
2910 ; AVX-NEXT: vinsertf128 $1, 32(%rdx), %ymm6, %ymm6
2911 ; AVX-NEXT: vshufpd {{.*#+}} ymm6 = ymm6[0],ymm7[1],ymm6[2],ymm7[2]
2912 ; AVX-NEXT: vmovupd %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2913 ; AVX-NEXT: vmovaps 32(%r9), %xmm6
2914 ; AVX-NEXT: vmovaps 32(%r8), %xmm7
2915 ; AVX-NEXT: vmovlhps {{.*#+}} xmm8 = xmm7[0],xmm6[0]
2916 ; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm8, %ymm9
2917 ; AVX-NEXT: vinsertf128 $1, 32(%r10), %ymm8, %ymm8
2918 ; AVX-NEXT: vshufpd {{.*#+}} ymm8 = ymm8[0],ymm9[1],ymm8[2],ymm9[2]
2919 ; AVX-NEXT: vmovupd %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2920 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm4 = xmm5[1],xmm4[1]
2921 ; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm4, %ymm2
2922 ; AVX-NEXT: vbroadcastsd 40(%rdx), %ymm5
2923 ; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm5[4,5,6,7]
2924 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm4[0,1],ymm2[2,3],ymm4[4,5],ymm2[6,7]
2925 ; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2926 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm2 = xmm7[1],xmm6[1]
2927 ; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm2, %ymm3
2928 ; AVX-NEXT: vbroadcastsd 40(%r10), %ymm4
2929 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm4[4,5,6,7]
2930 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1],ymm3[2,3],ymm2[4,5],ymm3[6,7]
2931 ; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2932 ; AVX-NEXT: vmovaps 64(%rsi), %xmm2
2933 ; AVX-NEXT: vmovaps 64(%rdi), %xmm3
2934 ; AVX-NEXT: vmovlhps {{.*#+}} xmm4 = xmm3[0],xmm2[0]
2935 ; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm4, %ymm5
2936 ; AVX-NEXT: vinsertf128 $1, 64(%rdx), %ymm4, %ymm4
2937 ; AVX-NEXT: vshufpd {{.*#+}} ymm4 = ymm4[0],ymm5[1],ymm4[2],ymm5[2]
2938 ; AVX-NEXT: vmovupd %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2939 ; AVX-NEXT: vmovaps 64(%r9), %xmm4
2940 ; AVX-NEXT: vmovaps 64(%r8), %xmm5
2941 ; AVX-NEXT: vmovlhps {{.*#+}} xmm6 = xmm5[0],xmm4[0]
2942 ; AVX-NEXT: vinsertf128 $1, %xmm1, %ymm6, %ymm7
2943 ; AVX-NEXT: vinsertf128 $1, 64(%r10), %ymm6, %ymm6
2944 ; AVX-NEXT: vshufpd {{.*#+}} ymm6 = ymm6[0],ymm7[1],ymm6[2],ymm7[2]
2945 ; AVX-NEXT: vmovupd %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2946 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm2 = xmm3[1],xmm2[1]
2947 ; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm2, %ymm0
2948 ; AVX-NEXT: vbroadcastsd 72(%rdx), %ymm3
2949 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]
2950 ; AVX-NEXT: vmovaps 96(%rcx), %xmm3
2951 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1],ymm0[2,3],ymm2[4,5],ymm0[6,7]
2952 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2953 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm5[1],xmm4[1]
2954 ; AVX-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
2955 ; AVX-NEXT: vbroadcastsd 72(%r10), %ymm2
2956 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
2957 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7]
2958 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2959 ; AVX-NEXT: vmovaps 96(%rsi), %xmm0
2960 ; AVX-NEXT: vmovaps 96(%rdi), %xmm1
2961 ; AVX-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
2962 ; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm2, %ymm4
2963 ; AVX-NEXT: vinsertf128 $1, 96(%rdx), %ymm2, %ymm2
2964 ; AVX-NEXT: vshufpd {{.*#+}} ymm2 = ymm2[0],ymm4[1],ymm2[2],ymm4[2]
2965 ; AVX-NEXT: vmovupd %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2966 ; AVX-NEXT: vmovaps 96(%rax), %xmm2
2967 ; AVX-NEXT: vmovaps 96(%r9), %xmm4
2968 ; AVX-NEXT: vmovaps 96(%r8), %xmm5
2969 ; AVX-NEXT: vmovlhps {{.*#+}} xmm6 = xmm5[0],xmm4[0]
2970 ; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm6, %ymm7
2971 ; AVX-NEXT: vinsertf128 $1, 96(%r10), %ymm6, %ymm6
2972 ; AVX-NEXT: vshufpd {{.*#+}} ymm6 = ymm6[0],ymm7[1],ymm6[2],ymm7[2]
2973 ; AVX-NEXT: vmovupd %ymm6, (%rsp) # 32-byte Spill
2974 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
2975 ; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm1
2976 ; AVX-NEXT: vbroadcastsd 104(%rdx), %ymm3
2977 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm3[4,5,6,7]
2978 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7]
2979 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2980 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm5[1],xmm4[1]
2981 ; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm1
2982 ; AVX-NEXT: vbroadcastsd 104(%r10), %ymm2
2983 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
2984 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7]
2985 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2986 ; AVX-NEXT: vmovaps 16(%rsi), %xmm0
2987 ; AVX-NEXT: vmovaps 16(%rdi), %xmm1
2988 ; AVX-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
2989 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],mem[4,5,6,7]
2990 ; AVX-NEXT: vbroadcastsd 16(%rcx), %ymm3
2991 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
2992 ; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
2993 ; AVX-NEXT: vmovaps 16(%r9), %xmm2
2994 ; AVX-NEXT: vmovaps 16(%r8), %xmm3
2995 ; AVX-NEXT: vmovlhps {{.*#+}} xmm4 = xmm3[0],xmm2[0]
2996 ; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],mem[4,5,6,7]
2997 ; AVX-NEXT: vbroadcastsd 16(%rax), %ymm5
2998 ; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
2999 ; AVX-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
3000 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
3001 ; AVX-NEXT: vbroadcastsd 24(%rdx), %ymm1
3002 ; AVX-NEXT: vblendps {{.*#+}} ymm6 = ymm0[0,1,2,3],ymm1[4,5,6,7]
3003 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm3[1],xmm2[1]
3004 ; AVX-NEXT: vbroadcastsd 24(%r10), %ymm1
3005 ; AVX-NEXT: vblendps {{.*#+}} ymm10 = ymm0[0,1,2,3],ymm1[4,5,6,7]
3006 ; AVX-NEXT: vmovaps 48(%rsi), %xmm0
3007 ; AVX-NEXT: vmovaps 48(%rdi), %xmm1
3008 ; AVX-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
3009 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],mem[4,5,6,7]
3010 ; AVX-NEXT: vbroadcastsd 48(%rcx), %ymm3
3011 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
3012 ; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
3013 ; AVX-NEXT: vmovaps 48(%r9), %xmm2
3014 ; AVX-NEXT: vmovaps 48(%r8), %xmm4
3015 ; AVX-NEXT: vmovlhps {{.*#+}} xmm3 = xmm4[0],xmm2[0]
3016 ; AVX-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3],mem[4,5,6,7]
3017 ; AVX-NEXT: vbroadcastsd 48(%rax), %ymm7
3018 ; AVX-NEXT: vblendps {{.*#+}} ymm11 = ymm3[0,1,2,3,4,5],ymm7[6,7]
3019 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
3020 ; AVX-NEXT: vbroadcastsd 56(%rdx), %ymm1
3021 ; AVX-NEXT: vblendps {{.*#+}} ymm8 = ymm0[0,1,2,3],ymm1[4,5,6,7]
3022 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm4[1],xmm2[1]
3023 ; AVX-NEXT: vbroadcastsd 56(%r10), %ymm1
3024 ; AVX-NEXT: vblendps {{.*#+}} ymm7 = ymm0[0,1,2,3],ymm1[4,5,6,7]
3025 ; AVX-NEXT: vmovaps 80(%rsi), %xmm5
3026 ; AVX-NEXT: vmovaps 80(%rdi), %xmm2
3027 ; AVX-NEXT: vmovlhps {{.*#+}} xmm3 = xmm2[0],xmm5[0]
3028 ; AVX-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3],mem[4,5,6,7]
3029 ; AVX-NEXT: vbroadcastsd 80(%rcx), %ymm4
3030 ; AVX-NEXT: vblendps {{.*#+}} ymm9 = ymm3[0,1,2,3,4,5],ymm4[6,7]
3031 ; AVX-NEXT: vmovaps 80(%r9), %xmm1
3032 ; AVX-NEXT: vmovaps 80(%r8), %xmm0
3033 ; AVX-NEXT: vmovlhps {{.*#+}} xmm3 = xmm0[0],xmm1[0]
3034 ; AVX-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3],mem[4,5,6,7]
3035 ; AVX-NEXT: vbroadcastsd 80(%rax), %ymm14
3036 ; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm3[0,1,2,3,4,5],ymm14[6,7]
3037 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm2 = xmm2[1],xmm5[1]
3038 ; AVX-NEXT: vbroadcastsd 88(%rdx), %ymm5
3039 ; AVX-NEXT: vblendps {{.*#+}} ymm3 = ymm2[0,1,2,3],ymm5[4,5,6,7]
3040 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
3041 ; AVX-NEXT: vbroadcastsd 88(%r10), %ymm1
3042 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm0[0,1,2,3],ymm1[4,5,6,7]
3043 ; AVX-NEXT: vmovaps 112(%rsi), %xmm1
3044 ; AVX-NEXT: vmovaps 112(%rdi), %xmm5
3045 ; AVX-NEXT: vmovlhps {{.*#+}} xmm14 = xmm5[0],xmm1[0]
3046 ; AVX-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3],mem[4,5,6,7]
3047 ; AVX-NEXT: vbroadcastsd 112(%rcx), %ymm15
3048 ; AVX-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
3049 ; AVX-NEXT: vmovaps 112(%r9), %xmm15
3050 ; AVX-NEXT: vmovaps 112(%r8), %xmm0
3051 ; AVX-NEXT: vmovlhps {{.*#+}} xmm13 = xmm0[0],xmm15[0]
3052 ; AVX-NEXT: vblendps {{.*#+}} ymm13 = ymm13[0,1,2,3],mem[4,5,6,7]
3053 ; AVX-NEXT: vbroadcastsd 112(%rax), %ymm12
3054 ; AVX-NEXT: vblendps {{.*#+}} ymm12 = ymm13[0,1,2,3,4,5],ymm12[6,7]
3055 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm5[1],xmm1[1]
3056 ; AVX-NEXT: vbroadcastsd 120(%rdx), %ymm5
3057 ; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm5[4,5,6,7]
3058 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm15[1]
3059 ; AVX-NEXT: vbroadcastsd 120(%r10), %ymm5
3060 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm5[4,5,6,7]
3061 ; AVX-NEXT: movq {{[0-9]+}}(%rsp), %rdx
3062 ; AVX-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3,4,5],mem[6,7]
3063 ; AVX-NEXT: vblendps {{.*#+}} ymm5 = ymm10[0,1,2,3,4,5],mem[6,7]
3064 ; AVX-NEXT: vblendps {{.*#+}} ymm8 = ymm8[0,1,2,3,4,5],mem[6,7]
3065 ; AVX-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0,1,2,3,4,5],mem[6,7]
3066 ; AVX-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],mem[6,7]
3067 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],mem[6,7]
3068 ; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5],mem[6,7]
3069 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],mem[6,7]
3070 ; AVX-NEXT: vmovaps %ymm0, 992(%rdx)
3071 ; AVX-NEXT: vmovaps %ymm1, 960(%rdx)
3072 ; AVX-NEXT: vmovaps %ymm12, 928(%rdx)
3073 ; AVX-NEXT: vmovaps %ymm14, 896(%rdx)
3074 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3075 ; AVX-NEXT: vmovaps %ymm0, 864(%rdx)
3076 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3077 ; AVX-NEXT: vmovaps %ymm0, 832(%rdx)
3078 ; AVX-NEXT: vmovups (%rsp), %ymm0 # 32-byte Reload
3079 ; AVX-NEXT: vmovaps %ymm0, 800(%rdx)
3080 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3081 ; AVX-NEXT: vmovaps %ymm0, 768(%rdx)
3082 ; AVX-NEXT: vmovaps %ymm2, 736(%rdx)
3083 ; AVX-NEXT: vmovaps %ymm3, 704(%rdx)
3084 ; AVX-NEXT: vmovaps %ymm4, 672(%rdx)
3085 ; AVX-NEXT: vmovaps %ymm9, 640(%rdx)
3086 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3087 ; AVX-NEXT: vmovaps %ymm0, 608(%rdx)
3088 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3089 ; AVX-NEXT: vmovaps %ymm0, 576(%rdx)
3090 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3091 ; AVX-NEXT: vmovaps %ymm0, 544(%rdx)
3092 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3093 ; AVX-NEXT: vmovaps %ymm0, 512(%rdx)
3094 ; AVX-NEXT: vmovaps %ymm7, 480(%rdx)
3095 ; AVX-NEXT: vmovaps %ymm8, 448(%rdx)
3096 ; AVX-NEXT: vmovaps %ymm11, 416(%rdx)
3097 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3098 ; AVX-NEXT: vmovaps %ymm0, 384(%rdx)
3099 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3100 ; AVX-NEXT: vmovaps %ymm0, 352(%rdx)
3101 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3102 ; AVX-NEXT: vmovaps %ymm0, 320(%rdx)
3103 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3104 ; AVX-NEXT: vmovaps %ymm0, 288(%rdx)
3105 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3106 ; AVX-NEXT: vmovaps %ymm0, 256(%rdx)
3107 ; AVX-NEXT: vmovaps %ymm5, 224(%rdx)
3108 ; AVX-NEXT: vmovaps %ymm6, 192(%rdx)
3109 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3110 ; AVX-NEXT: vmovaps %ymm0, 160(%rdx)
3111 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3112 ; AVX-NEXT: vmovaps %ymm0, 128(%rdx)
3113 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3114 ; AVX-NEXT: vmovaps %ymm0, 96(%rdx)
3115 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3116 ; AVX-NEXT: vmovaps %ymm0, 64(%rdx)
3117 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3118 ; AVX-NEXT: vmovaps %ymm0, 32(%rdx)
3119 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3120 ; AVX-NEXT: vmovaps %ymm0, (%rdx)
3121 ; AVX-NEXT: addq $488, %rsp # imm = 0x1E8
3122 ; AVX-NEXT: vzeroupper
3123 ; AVX-NEXT: retq
3124 ;
3125 ; AVX2-LABEL: store_i64_stride8_vf16:
3126 ; AVX2: # %bb.0:
3127 ; AVX2-NEXT: subq $488, %rsp # imm = 0x1E8
3128 ; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %r10
3129 ; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %rax
3130 ; AVX2-NEXT: vmovaps (%rcx), %xmm0
3131 ; AVX2-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
3132 ; AVX2-NEXT: vmovaps 32(%rcx), %xmm3
3133 ; AVX2-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
3134 ; AVX2-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0
3135 ; AVX2-NEXT: vmovaps (%rsi), %xmm2
3136 ; AVX2-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
3137 ; AVX2-NEXT: vmovaps 32(%rsi), %xmm4
3138 ; AVX2-NEXT: vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
3139 ; AVX2-NEXT: vmovaps (%rdi), %xmm1
3140 ; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
3141 ; AVX2-NEXT: vmovaps 32(%rdi), %xmm5
3142 ; AVX2-NEXT: vmovaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
3143 ; AVX2-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm2[1]
3144 ; AVX2-NEXT: vbroadcastsd 8(%rdx), %ymm2
3145 ; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
3146 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
3147 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
3148 ; AVX2-NEXT: vmovaps (%rax), %xmm0
3149 ; AVX2-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
3150 ; AVX2-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0
3151 ; AVX2-NEXT: vmovaps (%r9), %xmm1
3152 ; AVX2-NEXT: vmovaps %xmm1, (%rsp) # 16-byte Spill
3153 ; AVX2-NEXT: vmovaps 32(%r9), %xmm6
3154 ; AVX2-NEXT: vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
3155 ; AVX2-NEXT: vmovaps (%r8), %xmm15
3156 ; AVX2-NEXT: vmovaps 32(%r8), %xmm7
3157 ; AVX2-NEXT: vmovaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
3158 ; AVX2-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm15[1],xmm1[1]
3159 ; AVX2-NEXT: vbroadcastsd 8(%r10), %ymm2
3160 ; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
3161 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
3162 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
3163 ; AVX2-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm0
3164 ; AVX2-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm5[1],xmm4[1]
3165 ; AVX2-NEXT: vbroadcastsd 40(%rdx), %ymm2
3166 ; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
3167 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
3168 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
3169 ; AVX2-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm7[1],xmm6[1]
3170 ; AVX2-NEXT: vbroadcastsd 40(%r10), %ymm1
3171 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
3172 ; AVX2-NEXT: vmovaps 32(%rax), %xmm12
3173 ; AVX2-NEXT: vinsertf128 $1, %xmm12, %ymm0, %ymm1
3174 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
3175 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
3176 ; AVX2-NEXT: vmovaps 64(%rsi), %xmm11
3177 ; AVX2-NEXT: vmovaps 64(%rdi), %xmm10
3178 ; AVX2-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm10[1],xmm11[1]
3179 ; AVX2-NEXT: vbroadcastsd 72(%rdx), %ymm1
3180 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
3181 ; AVX2-NEXT: vmovaps 64(%rcx), %xmm9
3182 ; AVX2-NEXT: vinsertf128 $1, %xmm9, %ymm0, %ymm1
3183 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
3184 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
3185 ; AVX2-NEXT: vmovaps 64(%r9), %xmm8
3186 ; AVX2-NEXT: vmovaps 64(%r8), %xmm7
3187 ; AVX2-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm7[1],xmm8[1]
3188 ; AVX2-NEXT: vbroadcastsd 72(%r10), %ymm1
3189 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
3190 ; AVX2-NEXT: vmovaps 64(%rax), %xmm6
3191 ; AVX2-NEXT: vinsertf128 $1, %xmm6, %ymm0, %ymm1
3192 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
3193 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
3194 ; AVX2-NEXT: vmovaps 96(%rsi), %xmm5
3195 ; AVX2-NEXT: vmovaps 96(%rdi), %xmm4
3196 ; AVX2-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm4[1],xmm5[1]
3197 ; AVX2-NEXT: vbroadcastsd 104(%rdx), %ymm1
3198 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
3199 ; AVX2-NEXT: vmovaps 96(%rcx), %xmm3
3200 ; AVX2-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm1
3201 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
3202 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
3203 ; AVX2-NEXT: vmovaps 96(%r9), %xmm2
3204 ; AVX2-NEXT: vmovaps 96(%r8), %xmm1
3205 ; AVX2-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm2[1]
3206 ; AVX2-NEXT: vbroadcastsd 104(%r10), %ymm14
3207 ; AVX2-NEXT: vblendps {{.*#+}} ymm14 = ymm0[0,1,2,3],ymm14[4,5,6,7]
3208 ; AVX2-NEXT: vmovaps 96(%rax), %xmm0
3209 ; AVX2-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm13
3210 ; AVX2-NEXT: vblendps {{.*#+}} ymm13 = ymm14[0,1,2,3,4,5],ymm13[6,7]
3211 ; AVX2-NEXT: vmovups %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
3212 ; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
3213 ; AVX2-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm13, %xmm13 # 16-byte Folded Reload
3214 ; AVX2-NEXT: # xmm13 = xmm13[0],mem[0]
3215 ; AVX2-NEXT: vinsertf128 $1, (%rdx), %ymm13, %ymm13
3216 ; AVX2-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 16-byte Folded Reload
3217 ; AVX2-NEXT: vblendps {{.*#+}} ymm13 = ymm13[0,1,2,3,4,5],ymm14[6,7]
3218 ; AVX2-NEXT: vmovups %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
3219 ; AVX2-NEXT: vunpcklpd (%rsp), %xmm15, %xmm13 # 16-byte Folded Reload
3220 ; AVX2-NEXT: # xmm13 = xmm15[0],mem[0]
3221 ; AVX2-NEXT: vinsertf128 $1, (%r10), %ymm13, %ymm13
3222 ; AVX2-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 16-byte Folded Reload
3223 ; AVX2-NEXT: vblendps {{.*#+}} ymm13 = ymm13[0,1,2,3,4,5],ymm14[6,7]
3224 ; AVX2-NEXT: vmovups %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
3225 ; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
3226 ; AVX2-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm13, %xmm13 # 16-byte Folded Reload
3227 ; AVX2-NEXT: # xmm13 = xmm13[0],mem[0]
3228 ; AVX2-NEXT: vinsertf128 $1, 32(%rdx), %ymm13, %ymm13
3229 ; AVX2-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 16-byte Folded Reload
3230 ; AVX2-NEXT: vblendps {{.*#+}} ymm13 = ymm13[0,1,2,3,4,5],ymm14[6,7]
3231 ; AVX2-NEXT: vmovups %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
3232 ; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
3233 ; AVX2-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm13, %xmm13 # 16-byte Folded Reload
3234 ; AVX2-NEXT: # xmm13 = xmm13[0],mem[0]
3235 ; AVX2-NEXT: vinsertf128 $1, 32(%r10), %ymm13, %ymm13
3236 ; AVX2-NEXT: vbroadcastsd %xmm12, %ymm12
3237 ; AVX2-NEXT: vblendps {{.*#+}} ymm12 = ymm13[0,1,2,3,4,5],ymm12[6,7]
3238 ; AVX2-NEXT: vmovups %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
3239 ; AVX2-NEXT: vmovlhps {{.*#+}} xmm10 = xmm10[0],xmm11[0]
3240 ; AVX2-NEXT: vinsertf128 $1, 64(%rdx), %ymm10, %ymm10
3241 ; AVX2-NEXT: vbroadcastsd %xmm9, %ymm9
3242 ; AVX2-NEXT: vblendps {{.*#+}} ymm9 = ymm10[0,1,2,3,4,5],ymm9[6,7]
3243 ; AVX2-NEXT: vmovups %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
3244 ; AVX2-NEXT: vmovlhps {{.*#+}} xmm7 = xmm7[0],xmm8[0]
3245 ; AVX2-NEXT: vinsertf128 $1, 64(%r10), %ymm7, %ymm7
3246 ; AVX2-NEXT: vbroadcastsd %xmm6, %ymm6
3247 ; AVX2-NEXT: vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3,4,5],ymm6[6,7]
3248 ; AVX2-NEXT: vmovups %ymm6, (%rsp) # 32-byte Spill
3249 ; AVX2-NEXT: vmovlhps {{.*#+}} xmm4 = xmm4[0],xmm5[0]
3250 ; AVX2-NEXT: vinsertf128 $1, 96(%rdx), %ymm4, %ymm4
3251 ; AVX2-NEXT: vbroadcastsd %xmm3, %ymm3
3252 ; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm4[0,1,2,3,4,5],ymm3[6,7]
3253 ; AVX2-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
3254 ; AVX2-NEXT: vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
3255 ; AVX2-NEXT: vinsertf128 $1, 96(%r10), %ymm1, %ymm1
3256 ; AVX2-NEXT: vbroadcastsd %xmm0, %ymm0
3257 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
3258 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
3259 ; AVX2-NEXT: vmovaps (%rdi), %ymm0
3260 ; AVX2-NEXT: vmovaps (%rsi), %ymm1
3261 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
3262 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
3263 ; AVX2-NEXT: vbroadcastsd 16(%rcx), %ymm3
3264 ; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
3265 ; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
3266 ; AVX2-NEXT: vmovaps (%r8), %ymm2
3267 ; AVX2-NEXT: vmovaps (%r9), %ymm3
3268 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
3269 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],mem[2,3]
3270 ; AVX2-NEXT: vbroadcastsd 16(%rax), %ymm6
3271 ; AVX2-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm6[6,7]
3272 ; AVX2-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
3273 ; AVX2-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
3274 ; AVX2-NEXT: vbroadcastsd 24(%rdx), %ymm1
3275 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm5 = ymm0[2,3],ymm1[2,3]
3276 ; AVX2-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
3277 ; AVX2-NEXT: vbroadcastsd 24(%r10), %ymm2
3278 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm0[2,3],ymm2[2,3]
3279 ; AVX2-NEXT: vmovaps 32(%rdi), %ymm6
3280 ; AVX2-NEXT: vmovaps 32(%rsi), %ymm7
3281 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm6[0],ymm7[0],ymm6[2],ymm7[2]
3282 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
3283 ; AVX2-NEXT: vbroadcastsd 48(%rcx), %ymm3
3284 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3,4,5],ymm3[6,7]
3285 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
3286 ; AVX2-NEXT: vmovaps 32(%r8), %ymm8
3287 ; AVX2-NEXT: vmovaps 32(%r9), %ymm9
3288 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm8[0],ymm9[0],ymm8[2],ymm9[2]
3289 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
3290 ; AVX2-NEXT: vbroadcastsd 48(%rax), %ymm10
3291 ; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm2[0,1,2,3,4,5],ymm10[6,7]
3292 ; AVX2-NEXT: vunpckhpd {{.*#+}} ymm6 = ymm6[1],ymm7[1],ymm6[3],ymm7[3]
3293 ; AVX2-NEXT: vbroadcastsd 56(%rdx), %ymm7
3294 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm7 = ymm6[2,3],ymm7[2,3]
3295 ; AVX2-NEXT: vunpckhpd {{.*#+}} ymm6 = ymm8[1],ymm9[1],ymm8[3],ymm9[3]
3296 ; AVX2-NEXT: vbroadcastsd 56(%r10), %ymm8
3297 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm8 = ymm6[2,3],ymm8[2,3]
3298 ; AVX2-NEXT: vmovaps 64(%rdi), %ymm10
3299 ; AVX2-NEXT: vmovaps 64(%rsi), %ymm11
3300 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm6 = ymm10[0],ymm11[0],ymm10[2],ymm11[2]
3301 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm6 = ymm6[2,3],mem[2,3]
3302 ; AVX2-NEXT: vbroadcastsd 80(%rcx), %ymm9
3303 ; AVX2-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3,4,5],ymm9[6,7]
3304 ; AVX2-NEXT: vmovaps 64(%r8), %ymm1
3305 ; AVX2-NEXT: vmovaps 64(%r9), %ymm0
3306 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm9 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
3307 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm9 = ymm9[2,3],mem[2,3]
3308 ; AVX2-NEXT: vbroadcastsd 80(%rax), %ymm14
3309 ; AVX2-NEXT: vblendps {{.*#+}} ymm9 = ymm9[0,1,2,3,4,5],ymm14[6,7]
3310 ; AVX2-NEXT: vunpckhpd {{.*#+}} ymm10 = ymm10[1],ymm11[1],ymm10[3],ymm11[3]
3311 ; AVX2-NEXT: vbroadcastsd 88(%rdx), %ymm11
3312 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm10 = ymm10[2,3],ymm11[2,3]
3313 ; AVX2-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
3314 ; AVX2-NEXT: vbroadcastsd 88(%r10), %ymm1
3315 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm0[2,3],ymm1[2,3]
3316 ; AVX2-NEXT: vmovaps 96(%rdi), %ymm1
3317 ; AVX2-NEXT: vmovaps 96(%rsi), %ymm11
3318 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm14 = ymm1[0],ymm11[0],ymm1[2],ymm11[2]
3319 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm14 = ymm14[2,3],mem[2,3]
3320 ; AVX2-NEXT: vbroadcastsd 112(%rcx), %ymm15
3321 ; AVX2-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
3322 ; AVX2-NEXT: vmovaps 96(%r8), %ymm15
3323 ; AVX2-NEXT: vmovaps 96(%r9), %ymm0
3324 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm13 = ymm15[0],ymm0[0],ymm15[2],ymm0[2]
3325 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm13 = ymm13[2,3],mem[2,3]
3326 ; AVX2-NEXT: vbroadcastsd 112(%rax), %ymm12
3327 ; AVX2-NEXT: vblendps {{.*#+}} ymm12 = ymm13[0,1,2,3,4,5],ymm12[6,7]
3328 ; AVX2-NEXT: vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm11[1],ymm1[3],ymm11[3]
3329 ; AVX2-NEXT: vbroadcastsd 120(%rdx), %ymm11
3330 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm1[2,3],ymm11[2,3]
3331 ; AVX2-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm15[1],ymm0[1],ymm15[3],ymm0[3]
3332 ; AVX2-NEXT: vbroadcastsd 120(%r10), %ymm11
3333 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm11[2,3]
3334 ; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %rdx
3335 ; AVX2-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3,4,5],mem[6,7]
3336 ; AVX2-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],mem[6,7]
3337 ; AVX2-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0,1,2,3,4,5],mem[6,7]
3338 ; AVX2-NEXT: vblendps {{.*#+}} ymm8 = ymm8[0,1,2,3,4,5],mem[6,7]
3339 ; AVX2-NEXT: vblendps {{.*#+}} ymm10 = ymm10[0,1,2,3,4,5],mem[6,7]
3340 ; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],mem[6,7]
3341 ; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5],mem[6,7]
3342 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],mem[6,7]
3343 ; AVX2-NEXT: vmovaps %ymm0, 992(%rdx)
3344 ; AVX2-NEXT: vmovaps %ymm1, 960(%rdx)
3345 ; AVX2-NEXT: vmovaps %ymm12, 928(%rdx)
3346 ; AVX2-NEXT: vmovaps %ymm14, 896(%rdx)
3347 ; AVX2-NEXT: vmovaps %ymm2, 736(%rdx)
3348 ; AVX2-NEXT: vmovaps %ymm10, 704(%rdx)
3349 ; AVX2-NEXT: vmovaps %ymm9, 672(%rdx)
3350 ; AVX2-NEXT: vmovaps %ymm6, 640(%rdx)
3351 ; AVX2-NEXT: vmovaps %ymm8, 480(%rdx)
3352 ; AVX2-NEXT: vmovaps %ymm7, 448(%rdx)
3353 ; AVX2-NEXT: vmovaps %ymm3, 416(%rdx)
3354 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3355 ; AVX2-NEXT: vmovaps %ymm0, 384(%rdx)
3356 ; AVX2-NEXT: vmovaps %ymm4, 224(%rdx)
3357 ; AVX2-NEXT: vmovaps %ymm5, 192(%rdx)
3358 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3359 ; AVX2-NEXT: vmovaps %ymm0, 160(%rdx)
3360 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3361 ; AVX2-NEXT: vmovaps %ymm0, 128(%rdx)
3362 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3363 ; AVX2-NEXT: vmovaps %ymm0, 864(%rdx)
3364 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3365 ; AVX2-NEXT: vmovaps %ymm0, 832(%rdx)
3366 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3367 ; AVX2-NEXT: vmovaps %ymm0, 800(%rdx)
3368 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3369 ; AVX2-NEXT: vmovaps %ymm0, 768(%rdx)
3370 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3371 ; AVX2-NEXT: vmovaps %ymm0, 608(%rdx)
3372 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3373 ; AVX2-NEXT: vmovaps %ymm0, 576(%rdx)
3374 ; AVX2-NEXT: vmovups (%rsp), %ymm0 # 32-byte Reload
3375 ; AVX2-NEXT: vmovaps %ymm0, 544(%rdx)
3376 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3377 ; AVX2-NEXT: vmovaps %ymm0, 512(%rdx)
3378 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3379 ; AVX2-NEXT: vmovaps %ymm0, 352(%rdx)
3380 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3381 ; AVX2-NEXT: vmovaps %ymm0, 320(%rdx)
3382 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3383 ; AVX2-NEXT: vmovaps %ymm0, 288(%rdx)
3384 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3385 ; AVX2-NEXT: vmovaps %ymm0, 256(%rdx)
3386 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3387 ; AVX2-NEXT: vmovaps %ymm0, 96(%rdx)
3388 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3389 ; AVX2-NEXT: vmovaps %ymm0, 64(%rdx)
3390 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3391 ; AVX2-NEXT: vmovaps %ymm0, 32(%rdx)
3392 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3393 ; AVX2-NEXT: vmovaps %ymm0, (%rdx)
3394 ; AVX2-NEXT: addq $488, %rsp # imm = 0x1E8
3395 ; AVX2-NEXT: vzeroupper
3396 ; AVX2-NEXT: retq
3397 ;
3398 ; AVX2-FP-LABEL: store_i64_stride8_vf16:
3399 ; AVX2-FP: # %bb.0:
3400 ; AVX2-FP-NEXT: subq $488, %rsp # imm = 0x1E8
3401 ; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %r10
3402 ; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %rax
3403 ; AVX2-FP-NEXT: vmovaps (%rcx), %xmm0
3404 ; AVX2-FP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
3405 ; AVX2-FP-NEXT: vmovaps 32(%rcx), %xmm3
3406 ; AVX2-FP-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
3407 ; AVX2-FP-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0
3408 ; AVX2-FP-NEXT: vmovaps (%rsi), %xmm2
3409 ; AVX2-FP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
3410 ; AVX2-FP-NEXT: vmovaps 32(%rsi), %xmm4
3411 ; AVX2-FP-NEXT: vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
3412 ; AVX2-FP-NEXT: vmovaps (%rdi), %xmm1
3413 ; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
3414 ; AVX2-FP-NEXT: vmovaps 32(%rdi), %xmm5
3415 ; AVX2-FP-NEXT: vmovaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
3416 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm2[1]
3417 ; AVX2-FP-NEXT: vbroadcastsd 8(%rdx), %ymm2
3418 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
3419 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
3420 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
3421 ; AVX2-FP-NEXT: vmovaps (%rax), %xmm0
3422 ; AVX2-FP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
3423 ; AVX2-FP-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0
3424 ; AVX2-FP-NEXT: vmovaps (%r9), %xmm1
3425 ; AVX2-FP-NEXT: vmovaps %xmm1, (%rsp) # 16-byte Spill
3426 ; AVX2-FP-NEXT: vmovaps 32(%r9), %xmm6
3427 ; AVX2-FP-NEXT: vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
3428 ; AVX2-FP-NEXT: vmovaps (%r8), %xmm15
3429 ; AVX2-FP-NEXT: vmovaps 32(%r8), %xmm7
3430 ; AVX2-FP-NEXT: vmovaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
3431 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm15[1],xmm1[1]
3432 ; AVX2-FP-NEXT: vbroadcastsd 8(%r10), %ymm2
3433 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
3434 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
3435 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
3436 ; AVX2-FP-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm0
3437 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm5[1],xmm4[1]
3438 ; AVX2-FP-NEXT: vbroadcastsd 40(%rdx), %ymm2
3439 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
3440 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
3441 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
3442 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm7[1],xmm6[1]
3443 ; AVX2-FP-NEXT: vbroadcastsd 40(%r10), %ymm1
3444 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
3445 ; AVX2-FP-NEXT: vmovaps 32(%rax), %xmm12
3446 ; AVX2-FP-NEXT: vinsertf128 $1, %xmm12, %ymm0, %ymm1
3447 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
3448 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
3449 ; AVX2-FP-NEXT: vmovaps 64(%rsi), %xmm11
3450 ; AVX2-FP-NEXT: vmovaps 64(%rdi), %xmm10
3451 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm10[1],xmm11[1]
3452 ; AVX2-FP-NEXT: vbroadcastsd 72(%rdx), %ymm1
3453 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
3454 ; AVX2-FP-NEXT: vmovaps 64(%rcx), %xmm9
3455 ; AVX2-FP-NEXT: vinsertf128 $1, %xmm9, %ymm0, %ymm1
3456 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
3457 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
3458 ; AVX2-FP-NEXT: vmovaps 64(%r9), %xmm8
3459 ; AVX2-FP-NEXT: vmovaps 64(%r8), %xmm7
3460 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm7[1],xmm8[1]
3461 ; AVX2-FP-NEXT: vbroadcastsd 72(%r10), %ymm1
3462 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
3463 ; AVX2-FP-NEXT: vmovaps 64(%rax), %xmm6
3464 ; AVX2-FP-NEXT: vinsertf128 $1, %xmm6, %ymm0, %ymm1
3465 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
3466 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
3467 ; AVX2-FP-NEXT: vmovaps 96(%rsi), %xmm5
3468 ; AVX2-FP-NEXT: vmovaps 96(%rdi), %xmm4
3469 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm4[1],xmm5[1]
3470 ; AVX2-FP-NEXT: vbroadcastsd 104(%rdx), %ymm1
3471 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
3472 ; AVX2-FP-NEXT: vmovaps 96(%rcx), %xmm3
3473 ; AVX2-FP-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm1
3474 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
3475 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
3476 ; AVX2-FP-NEXT: vmovaps 96(%r9), %xmm2
3477 ; AVX2-FP-NEXT: vmovaps 96(%r8), %xmm1
3478 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm2[1]
3479 ; AVX2-FP-NEXT: vbroadcastsd 104(%r10), %ymm14
3480 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm14 = ymm0[0,1,2,3],ymm14[4,5,6,7]
3481 ; AVX2-FP-NEXT: vmovaps 96(%rax), %xmm0
3482 ; AVX2-FP-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm13
3483 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm13 = ymm14[0,1,2,3,4,5],ymm13[6,7]
3484 ; AVX2-FP-NEXT: vmovups %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
3485 ; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
3486 ; AVX2-FP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm13, %xmm13 # 16-byte Folded Reload
3487 ; AVX2-FP-NEXT: # xmm13 = xmm13[0],mem[0]
3488 ; AVX2-FP-NEXT: vinsertf128 $1, (%rdx), %ymm13, %ymm13
3489 ; AVX2-FP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 16-byte Folded Reload
3490 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm13 = ymm13[0,1,2,3,4,5],ymm14[6,7]
3491 ; AVX2-FP-NEXT: vmovups %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
3492 ; AVX2-FP-NEXT: vunpcklpd (%rsp), %xmm15, %xmm13 # 16-byte Folded Reload
3493 ; AVX2-FP-NEXT: # xmm13 = xmm15[0],mem[0]
3494 ; AVX2-FP-NEXT: vinsertf128 $1, (%r10), %ymm13, %ymm13
3495 ; AVX2-FP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 16-byte Folded Reload
3496 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm13 = ymm13[0,1,2,3,4,5],ymm14[6,7]
3497 ; AVX2-FP-NEXT: vmovups %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
3498 ; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
3499 ; AVX2-FP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm13, %xmm13 # 16-byte Folded Reload
3500 ; AVX2-FP-NEXT: # xmm13 = xmm13[0],mem[0]
3501 ; AVX2-FP-NEXT: vinsertf128 $1, 32(%rdx), %ymm13, %ymm13
3502 ; AVX2-FP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 16-byte Folded Reload
3503 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm13 = ymm13[0,1,2,3,4,5],ymm14[6,7]
3504 ; AVX2-FP-NEXT: vmovups %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
3505 ; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
3506 ; AVX2-FP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm13, %xmm13 # 16-byte Folded Reload
3507 ; AVX2-FP-NEXT: # xmm13 = xmm13[0],mem[0]
3508 ; AVX2-FP-NEXT: vinsertf128 $1, 32(%r10), %ymm13, %ymm13
3509 ; AVX2-FP-NEXT: vbroadcastsd %xmm12, %ymm12
3510 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm12 = ymm13[0,1,2,3,4,5],ymm12[6,7]
3511 ; AVX2-FP-NEXT: vmovups %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
3512 ; AVX2-FP-NEXT: vmovlhps {{.*#+}} xmm10 = xmm10[0],xmm11[0]
3513 ; AVX2-FP-NEXT: vinsertf128 $1, 64(%rdx), %ymm10, %ymm10
3514 ; AVX2-FP-NEXT: vbroadcastsd %xmm9, %ymm9
3515 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm9 = ymm10[0,1,2,3,4,5],ymm9[6,7]
3516 ; AVX2-FP-NEXT: vmovups %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
3517 ; AVX2-FP-NEXT: vmovlhps {{.*#+}} xmm7 = xmm7[0],xmm8[0]
3518 ; AVX2-FP-NEXT: vinsertf128 $1, 64(%r10), %ymm7, %ymm7
3519 ; AVX2-FP-NEXT: vbroadcastsd %xmm6, %ymm6
3520 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3,4,5],ymm6[6,7]
3521 ; AVX2-FP-NEXT: vmovups %ymm6, (%rsp) # 32-byte Spill
3522 ; AVX2-FP-NEXT: vmovlhps {{.*#+}} xmm4 = xmm4[0],xmm5[0]
3523 ; AVX2-FP-NEXT: vinsertf128 $1, 96(%rdx), %ymm4, %ymm4
3524 ; AVX2-FP-NEXT: vbroadcastsd %xmm3, %ymm3
3525 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm4[0,1,2,3,4,5],ymm3[6,7]
3526 ; AVX2-FP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
3527 ; AVX2-FP-NEXT: vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
3528 ; AVX2-FP-NEXT: vinsertf128 $1, 96(%r10), %ymm1, %ymm1
3529 ; AVX2-FP-NEXT: vbroadcastsd %xmm0, %ymm0
3530 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
3531 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
3532 ; AVX2-FP-NEXT: vmovaps (%rdi), %ymm0
3533 ; AVX2-FP-NEXT: vmovaps (%rsi), %ymm1
3534 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
3535 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
3536 ; AVX2-FP-NEXT: vbroadcastsd 16(%rcx), %ymm3
3537 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
3538 ; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
3539 ; AVX2-FP-NEXT: vmovaps (%r8), %ymm2
3540 ; AVX2-FP-NEXT: vmovaps (%r9), %ymm3
3541 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
3542 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],mem[2,3]
3543 ; AVX2-FP-NEXT: vbroadcastsd 16(%rax), %ymm6
3544 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm6[6,7]
3545 ; AVX2-FP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
3546 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
3547 ; AVX2-FP-NEXT: vbroadcastsd 24(%rdx), %ymm1
3548 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm5 = ymm0[2,3],ymm1[2,3]
3549 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
3550 ; AVX2-FP-NEXT: vbroadcastsd 24(%r10), %ymm2
3551 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm0[2,3],ymm2[2,3]
3552 ; AVX2-FP-NEXT: vmovaps 32(%rdi), %ymm6
3553 ; AVX2-FP-NEXT: vmovaps 32(%rsi), %ymm7
3554 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm6[0],ymm7[0],ymm6[2],ymm7[2]
3555 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
3556 ; AVX2-FP-NEXT: vbroadcastsd 48(%rcx), %ymm3
3557 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3,4,5],ymm3[6,7]
3558 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
3559 ; AVX2-FP-NEXT: vmovaps 32(%r8), %ymm8
3560 ; AVX2-FP-NEXT: vmovaps 32(%r9), %ymm9
3561 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm8[0],ymm9[0],ymm8[2],ymm9[2]
3562 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
3563 ; AVX2-FP-NEXT: vbroadcastsd 48(%rax), %ymm10
3564 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm2[0,1,2,3,4,5],ymm10[6,7]
3565 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm6 = ymm6[1],ymm7[1],ymm6[3],ymm7[3]
3566 ; AVX2-FP-NEXT: vbroadcastsd 56(%rdx), %ymm7
3567 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm7 = ymm6[2,3],ymm7[2,3]
3568 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm6 = ymm8[1],ymm9[1],ymm8[3],ymm9[3]
3569 ; AVX2-FP-NEXT: vbroadcastsd 56(%r10), %ymm8
3570 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm8 = ymm6[2,3],ymm8[2,3]
3571 ; AVX2-FP-NEXT: vmovaps 64(%rdi), %ymm10
3572 ; AVX2-FP-NEXT: vmovaps 64(%rsi), %ymm11
3573 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm6 = ymm10[0],ymm11[0],ymm10[2],ymm11[2]
3574 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm6 = ymm6[2,3],mem[2,3]
3575 ; AVX2-FP-NEXT: vbroadcastsd 80(%rcx), %ymm9
3576 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3,4,5],ymm9[6,7]
3577 ; AVX2-FP-NEXT: vmovaps 64(%r8), %ymm1
3578 ; AVX2-FP-NEXT: vmovaps 64(%r9), %ymm0
3579 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm9 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
3580 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm9 = ymm9[2,3],mem[2,3]
3581 ; AVX2-FP-NEXT: vbroadcastsd 80(%rax), %ymm14
3582 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm9 = ymm9[0,1,2,3,4,5],ymm14[6,7]
3583 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm10 = ymm10[1],ymm11[1],ymm10[3],ymm11[3]
3584 ; AVX2-FP-NEXT: vbroadcastsd 88(%rdx), %ymm11
3585 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm10 = ymm10[2,3],ymm11[2,3]
3586 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
3587 ; AVX2-FP-NEXT: vbroadcastsd 88(%r10), %ymm1
3588 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm0[2,3],ymm1[2,3]
3589 ; AVX2-FP-NEXT: vmovaps 96(%rdi), %ymm1
3590 ; AVX2-FP-NEXT: vmovaps 96(%rsi), %ymm11
3591 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm14 = ymm1[0],ymm11[0],ymm1[2],ymm11[2]
3592 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm14 = ymm14[2,3],mem[2,3]
3593 ; AVX2-FP-NEXT: vbroadcastsd 112(%rcx), %ymm15
3594 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
3595 ; AVX2-FP-NEXT: vmovaps 96(%r8), %ymm15
3596 ; AVX2-FP-NEXT: vmovaps 96(%r9), %ymm0
3597 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm13 = ymm15[0],ymm0[0],ymm15[2],ymm0[2]
3598 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm13 = ymm13[2,3],mem[2,3]
3599 ; AVX2-FP-NEXT: vbroadcastsd 112(%rax), %ymm12
3600 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm12 = ymm13[0,1,2,3,4,5],ymm12[6,7]
3601 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm11[1],ymm1[3],ymm11[3]
3602 ; AVX2-FP-NEXT: vbroadcastsd 120(%rdx), %ymm11
3603 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm1[2,3],ymm11[2,3]
3604 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm15[1],ymm0[1],ymm15[3],ymm0[3]
3605 ; AVX2-FP-NEXT: vbroadcastsd 120(%r10), %ymm11
3606 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm11[2,3]
3607 ; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %rdx
3608 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3,4,5],mem[6,7]
3609 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],mem[6,7]
3610 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0,1,2,3,4,5],mem[6,7]
3611 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm8 = ymm8[0,1,2,3,4,5],mem[6,7]
3612 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm10 = ymm10[0,1,2,3,4,5],mem[6,7]
3613 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],mem[6,7]
3614 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5],mem[6,7]
3615 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],mem[6,7]
3616 ; AVX2-FP-NEXT: vmovaps %ymm0, 992(%rdx)
3617 ; AVX2-FP-NEXT: vmovaps %ymm1, 960(%rdx)
3618 ; AVX2-FP-NEXT: vmovaps %ymm12, 928(%rdx)
3619 ; AVX2-FP-NEXT: vmovaps %ymm14, 896(%rdx)
3620 ; AVX2-FP-NEXT: vmovaps %ymm2, 736(%rdx)
3621 ; AVX2-FP-NEXT: vmovaps %ymm10, 704(%rdx)
3622 ; AVX2-FP-NEXT: vmovaps %ymm9, 672(%rdx)
3623 ; AVX2-FP-NEXT: vmovaps %ymm6, 640(%rdx)
3624 ; AVX2-FP-NEXT: vmovaps %ymm8, 480(%rdx)
3625 ; AVX2-FP-NEXT: vmovaps %ymm7, 448(%rdx)
3626 ; AVX2-FP-NEXT: vmovaps %ymm3, 416(%rdx)
3627 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3628 ; AVX2-FP-NEXT: vmovaps %ymm0, 384(%rdx)
3629 ; AVX2-FP-NEXT: vmovaps %ymm4, 224(%rdx)
3630 ; AVX2-FP-NEXT: vmovaps %ymm5, 192(%rdx)
3631 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3632 ; AVX2-FP-NEXT: vmovaps %ymm0, 160(%rdx)
3633 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3634 ; AVX2-FP-NEXT: vmovaps %ymm0, 128(%rdx)
3635 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3636 ; AVX2-FP-NEXT: vmovaps %ymm0, 864(%rdx)
3637 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3638 ; AVX2-FP-NEXT: vmovaps %ymm0, 832(%rdx)
3639 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3640 ; AVX2-FP-NEXT: vmovaps %ymm0, 800(%rdx)
3641 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3642 ; AVX2-FP-NEXT: vmovaps %ymm0, 768(%rdx)
3643 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3644 ; AVX2-FP-NEXT: vmovaps %ymm0, 608(%rdx)
3645 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3646 ; AVX2-FP-NEXT: vmovaps %ymm0, 576(%rdx)
3647 ; AVX2-FP-NEXT: vmovups (%rsp), %ymm0 # 32-byte Reload
3648 ; AVX2-FP-NEXT: vmovaps %ymm0, 544(%rdx)
3649 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3650 ; AVX2-FP-NEXT: vmovaps %ymm0, 512(%rdx)
3651 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3652 ; AVX2-FP-NEXT: vmovaps %ymm0, 352(%rdx)
3653 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3654 ; AVX2-FP-NEXT: vmovaps %ymm0, 320(%rdx)
3655 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3656 ; AVX2-FP-NEXT: vmovaps %ymm0, 288(%rdx)
3657 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3658 ; AVX2-FP-NEXT: vmovaps %ymm0, 256(%rdx)
3659 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3660 ; AVX2-FP-NEXT: vmovaps %ymm0, 96(%rdx)
3661 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3662 ; AVX2-FP-NEXT: vmovaps %ymm0, 64(%rdx)
3663 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3664 ; AVX2-FP-NEXT: vmovaps %ymm0, 32(%rdx)
3665 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3666 ; AVX2-FP-NEXT: vmovaps %ymm0, (%rdx)
3667 ; AVX2-FP-NEXT: addq $488, %rsp # imm = 0x1E8
3668 ; AVX2-FP-NEXT: vzeroupper
3669 ; AVX2-FP-NEXT: retq
3670 ;
3671 ; AVX2-FCP-LABEL: store_i64_stride8_vf16:
3672 ; AVX2-FCP: # %bb.0:
3673 ; AVX2-FCP-NEXT: subq $488, %rsp # imm = 0x1E8
3674 ; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10
3675 ; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
3676 ; AVX2-FCP-NEXT: vmovaps (%rcx), %xmm0
3677 ; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
3678 ; AVX2-FCP-NEXT: vmovaps 32(%rcx), %xmm3
3679 ; AVX2-FCP-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
3680 ; AVX2-FCP-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0
3681 ; AVX2-FCP-NEXT: vmovaps (%rsi), %xmm2
3682 ; AVX2-FCP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
3683 ; AVX2-FCP-NEXT: vmovaps 32(%rsi), %xmm4
3684 ; AVX2-FCP-NEXT: vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
3685 ; AVX2-FCP-NEXT: vmovaps (%rdi), %xmm1
3686 ; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
3687 ; AVX2-FCP-NEXT: vmovaps 32(%rdi), %xmm5
3688 ; AVX2-FCP-NEXT: vmovaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
3689 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm2[1]
3690 ; AVX2-FCP-NEXT: vbroadcastsd 8(%rdx), %ymm2
3691 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
3692 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
3693 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
3694 ; AVX2-FCP-NEXT: vmovaps (%rax), %xmm0
3695 ; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
3696 ; AVX2-FCP-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0
3697 ; AVX2-FCP-NEXT: vmovaps (%r9), %xmm1
3698 ; AVX2-FCP-NEXT: vmovaps %xmm1, (%rsp) # 16-byte Spill
3699 ; AVX2-FCP-NEXT: vmovaps 32(%r9), %xmm6
3700 ; AVX2-FCP-NEXT: vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
3701 ; AVX2-FCP-NEXT: vmovaps (%r8), %xmm15
3702 ; AVX2-FCP-NEXT: vmovaps 32(%r8), %xmm7
3703 ; AVX2-FCP-NEXT: vmovaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
3704 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm15[1],xmm1[1]
3705 ; AVX2-FCP-NEXT: vbroadcastsd 8(%r10), %ymm2
3706 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
3707 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
3708 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
3709 ; AVX2-FCP-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm0
3710 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm5[1],xmm4[1]
3711 ; AVX2-FCP-NEXT: vbroadcastsd 40(%rdx), %ymm2
3712 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
3713 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
3714 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
3715 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm7[1],xmm6[1]
3716 ; AVX2-FCP-NEXT: vbroadcastsd 40(%r10), %ymm1
3717 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
3718 ; AVX2-FCP-NEXT: vmovaps 32(%rax), %xmm12
3719 ; AVX2-FCP-NEXT: vinsertf128 $1, %xmm12, %ymm0, %ymm1
3720 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
3721 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
3722 ; AVX2-FCP-NEXT: vmovaps 64(%rsi), %xmm11
3723 ; AVX2-FCP-NEXT: vmovaps 64(%rdi), %xmm10
3724 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm10[1],xmm11[1]
3725 ; AVX2-FCP-NEXT: vbroadcastsd 72(%rdx), %ymm1
3726 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
3727 ; AVX2-FCP-NEXT: vmovaps 64(%rcx), %xmm9
3728 ; AVX2-FCP-NEXT: vinsertf128 $1, %xmm9, %ymm0, %ymm1
3729 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
3730 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
3731 ; AVX2-FCP-NEXT: vmovaps 64(%r9), %xmm8
3732 ; AVX2-FCP-NEXT: vmovaps 64(%r8), %xmm7
3733 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm7[1],xmm8[1]
3734 ; AVX2-FCP-NEXT: vbroadcastsd 72(%r10), %ymm1
3735 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
3736 ; AVX2-FCP-NEXT: vmovaps 64(%rax), %xmm6
3737 ; AVX2-FCP-NEXT: vinsertf128 $1, %xmm6, %ymm0, %ymm1
3738 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
3739 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
3740 ; AVX2-FCP-NEXT: vmovaps 96(%rsi), %xmm5
3741 ; AVX2-FCP-NEXT: vmovaps 96(%rdi), %xmm4
3742 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm4[1],xmm5[1]
3743 ; AVX2-FCP-NEXT: vbroadcastsd 104(%rdx), %ymm1
3744 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
3745 ; AVX2-FCP-NEXT: vmovaps 96(%rcx), %xmm3
3746 ; AVX2-FCP-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm1
3747 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
3748 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
3749 ; AVX2-FCP-NEXT: vmovaps 96(%r9), %xmm2
3750 ; AVX2-FCP-NEXT: vmovaps 96(%r8), %xmm1
3751 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm2[1]
3752 ; AVX2-FCP-NEXT: vbroadcastsd 104(%r10), %ymm14
3753 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm14 = ymm0[0,1,2,3],ymm14[4,5,6,7]
3754 ; AVX2-FCP-NEXT: vmovaps 96(%rax), %xmm0
3755 ; AVX2-FCP-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm13
3756 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm13 = ymm14[0,1,2,3,4,5],ymm13[6,7]
3757 ; AVX2-FCP-NEXT: vmovups %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
3758 ; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
3759 ; AVX2-FCP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm13, %xmm13 # 16-byte Folded Reload
3760 ; AVX2-FCP-NEXT: # xmm13 = xmm13[0],mem[0]
3761 ; AVX2-FCP-NEXT: vinsertf128 $1, (%rdx), %ymm13, %ymm13
3762 ; AVX2-FCP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 16-byte Folded Reload
3763 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm13 = ymm13[0,1,2,3,4,5],ymm14[6,7]
3764 ; AVX2-FCP-NEXT: vmovups %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
3765 ; AVX2-FCP-NEXT: vunpcklpd (%rsp), %xmm15, %xmm13 # 16-byte Folded Reload
3766 ; AVX2-FCP-NEXT: # xmm13 = xmm15[0],mem[0]
3767 ; AVX2-FCP-NEXT: vinsertf128 $1, (%r10), %ymm13, %ymm13
3768 ; AVX2-FCP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 16-byte Folded Reload
3769 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm13 = ymm13[0,1,2,3,4,5],ymm14[6,7]
3770 ; AVX2-FCP-NEXT: vmovups %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
3771 ; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
3772 ; AVX2-FCP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm13, %xmm13 # 16-byte Folded Reload
3773 ; AVX2-FCP-NEXT: # xmm13 = xmm13[0],mem[0]
3774 ; AVX2-FCP-NEXT: vinsertf128 $1, 32(%rdx), %ymm13, %ymm13
3775 ; AVX2-FCP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 16-byte Folded Reload
3776 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm13 = ymm13[0,1,2,3,4,5],ymm14[6,7]
3777 ; AVX2-FCP-NEXT: vmovups %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
3778 ; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
3779 ; AVX2-FCP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm13, %xmm13 # 16-byte Folded Reload
3780 ; AVX2-FCP-NEXT: # xmm13 = xmm13[0],mem[0]
3781 ; AVX2-FCP-NEXT: vinsertf128 $1, 32(%r10), %ymm13, %ymm13
3782 ; AVX2-FCP-NEXT: vbroadcastsd %xmm12, %ymm12
3783 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm12 = ymm13[0,1,2,3,4,5],ymm12[6,7]
3784 ; AVX2-FCP-NEXT: vmovups %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
3785 ; AVX2-FCP-NEXT: vmovlhps {{.*#+}} xmm10 = xmm10[0],xmm11[0]
3786 ; AVX2-FCP-NEXT: vinsertf128 $1, 64(%rdx), %ymm10, %ymm10
3787 ; AVX2-FCP-NEXT: vbroadcastsd %xmm9, %ymm9
3788 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm9 = ymm10[0,1,2,3,4,5],ymm9[6,7]
3789 ; AVX2-FCP-NEXT: vmovups %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
3790 ; AVX2-FCP-NEXT: vmovlhps {{.*#+}} xmm7 = xmm7[0],xmm8[0]
3791 ; AVX2-FCP-NEXT: vinsertf128 $1, 64(%r10), %ymm7, %ymm7
3792 ; AVX2-FCP-NEXT: vbroadcastsd %xmm6, %ymm6
3793 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3,4,5],ymm6[6,7]
3794 ; AVX2-FCP-NEXT: vmovups %ymm6, (%rsp) # 32-byte Spill
3795 ; AVX2-FCP-NEXT: vmovlhps {{.*#+}} xmm4 = xmm4[0],xmm5[0]
3796 ; AVX2-FCP-NEXT: vinsertf128 $1, 96(%rdx), %ymm4, %ymm4
3797 ; AVX2-FCP-NEXT: vbroadcastsd %xmm3, %ymm3
3798 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm3 = ymm4[0,1,2,3,4,5],ymm3[6,7]
3799 ; AVX2-FCP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
3800 ; AVX2-FCP-NEXT: vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
3801 ; AVX2-FCP-NEXT: vinsertf128 $1, 96(%r10), %ymm1, %ymm1
3802 ; AVX2-FCP-NEXT: vbroadcastsd %xmm0, %ymm0
3803 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
3804 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
3805 ; AVX2-FCP-NEXT: vmovaps (%rdi), %ymm0
3806 ; AVX2-FCP-NEXT: vmovaps (%rsi), %ymm1
3807 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
3808 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
3809 ; AVX2-FCP-NEXT: vbroadcastsd 16(%rcx), %ymm3
3810 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
3811 ; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
3812 ; AVX2-FCP-NEXT: vmovaps (%r8), %ymm2
3813 ; AVX2-FCP-NEXT: vmovaps (%r9), %ymm3
3814 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
3815 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],mem[2,3]
3816 ; AVX2-FCP-NEXT: vbroadcastsd 16(%rax), %ymm6
3817 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm6[6,7]
3818 ; AVX2-FCP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
3819 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
3820 ; AVX2-FCP-NEXT: vbroadcastsd 24(%rdx), %ymm1
3821 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm5 = ymm0[2,3],ymm1[2,3]
3822 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
3823 ; AVX2-FCP-NEXT: vbroadcastsd 24(%r10), %ymm2
3824 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm0[2,3],ymm2[2,3]
3825 ; AVX2-FCP-NEXT: vmovaps 32(%rdi), %ymm6
3826 ; AVX2-FCP-NEXT: vmovaps 32(%rsi), %ymm7
3827 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm6[0],ymm7[0],ymm6[2],ymm7[2]
3828 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
3829 ; AVX2-FCP-NEXT: vbroadcastsd 48(%rcx), %ymm3
3830 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3,4,5],ymm3[6,7]
3831 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
3832 ; AVX2-FCP-NEXT: vmovaps 32(%r8), %ymm8
3833 ; AVX2-FCP-NEXT: vmovaps 32(%r9), %ymm9
3834 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm8[0],ymm9[0],ymm8[2],ymm9[2]
3835 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
3836 ; AVX2-FCP-NEXT: vbroadcastsd 48(%rax), %ymm10
3837 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm3 = ymm2[0,1,2,3,4,5],ymm10[6,7]
3838 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm6 = ymm6[1],ymm7[1],ymm6[3],ymm7[3]
3839 ; AVX2-FCP-NEXT: vbroadcastsd 56(%rdx), %ymm7
3840 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm7 = ymm6[2,3],ymm7[2,3]
3841 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm6 = ymm8[1],ymm9[1],ymm8[3],ymm9[3]
3842 ; AVX2-FCP-NEXT: vbroadcastsd 56(%r10), %ymm8
3843 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm8 = ymm6[2,3],ymm8[2,3]
3844 ; AVX2-FCP-NEXT: vmovaps 64(%rdi), %ymm10
3845 ; AVX2-FCP-NEXT: vmovaps 64(%rsi), %ymm11
3846 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm6 = ymm10[0],ymm11[0],ymm10[2],ymm11[2]
3847 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm6 = ymm6[2,3],mem[2,3]
3848 ; AVX2-FCP-NEXT: vbroadcastsd 80(%rcx), %ymm9
3849 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3,4,5],ymm9[6,7]
3850 ; AVX2-FCP-NEXT: vmovaps 64(%r8), %ymm1
3851 ; AVX2-FCP-NEXT: vmovaps 64(%r9), %ymm0
3852 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm9 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
3853 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm9 = ymm9[2,3],mem[2,3]
3854 ; AVX2-FCP-NEXT: vbroadcastsd 80(%rax), %ymm14
3855 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm9 = ymm9[0,1,2,3,4,5],ymm14[6,7]
3856 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm10 = ymm10[1],ymm11[1],ymm10[3],ymm11[3]
3857 ; AVX2-FCP-NEXT: vbroadcastsd 88(%rdx), %ymm11
3858 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm10 = ymm10[2,3],ymm11[2,3]
3859 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
3860 ; AVX2-FCP-NEXT: vbroadcastsd 88(%r10), %ymm1
3861 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm0[2,3],ymm1[2,3]
3862 ; AVX2-FCP-NEXT: vmovaps 96(%rdi), %ymm1
3863 ; AVX2-FCP-NEXT: vmovaps 96(%rsi), %ymm11
3864 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm14 = ymm1[0],ymm11[0],ymm1[2],ymm11[2]
3865 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm14 = ymm14[2,3],mem[2,3]
3866 ; AVX2-FCP-NEXT: vbroadcastsd 112(%rcx), %ymm15
3867 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
3868 ; AVX2-FCP-NEXT: vmovaps 96(%r8), %ymm15
3869 ; AVX2-FCP-NEXT: vmovaps 96(%r9), %ymm0
3870 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm13 = ymm15[0],ymm0[0],ymm15[2],ymm0[2]
3871 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm13 = ymm13[2,3],mem[2,3]
3872 ; AVX2-FCP-NEXT: vbroadcastsd 112(%rax), %ymm12
3873 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm12 = ymm13[0,1,2,3,4,5],ymm12[6,7]
3874 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm1 = ymm1[1],ymm11[1],ymm1[3],ymm11[3]
3875 ; AVX2-FCP-NEXT: vbroadcastsd 120(%rdx), %ymm11
3876 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm1[2,3],ymm11[2,3]
3877 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm15[1],ymm0[1],ymm15[3],ymm0[3]
3878 ; AVX2-FCP-NEXT: vbroadcastsd 120(%r10), %ymm11
3879 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm11[2,3]
3880 ; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rdx
3881 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3,4,5],mem[6,7]
3882 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],mem[6,7]
3883 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0,1,2,3,4,5],mem[6,7]
3884 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm8 = ymm8[0,1,2,3,4,5],mem[6,7]
3885 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm10 = ymm10[0,1,2,3,4,5],mem[6,7]
3886 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],mem[6,7]
3887 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5],mem[6,7]
3888 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],mem[6,7]
3889 ; AVX2-FCP-NEXT: vmovaps %ymm0, 992(%rdx)
3890 ; AVX2-FCP-NEXT: vmovaps %ymm1, 960(%rdx)
3891 ; AVX2-FCP-NEXT: vmovaps %ymm12, 928(%rdx)
3892 ; AVX2-FCP-NEXT: vmovaps %ymm14, 896(%rdx)
3893 ; AVX2-FCP-NEXT: vmovaps %ymm2, 736(%rdx)
3894 ; AVX2-FCP-NEXT: vmovaps %ymm10, 704(%rdx)
3895 ; AVX2-FCP-NEXT: vmovaps %ymm9, 672(%rdx)
3896 ; AVX2-FCP-NEXT: vmovaps %ymm6, 640(%rdx)
3897 ; AVX2-FCP-NEXT: vmovaps %ymm8, 480(%rdx)
3898 ; AVX2-FCP-NEXT: vmovaps %ymm7, 448(%rdx)
3899 ; AVX2-FCP-NEXT: vmovaps %ymm3, 416(%rdx)
3900 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3901 ; AVX2-FCP-NEXT: vmovaps %ymm0, 384(%rdx)
3902 ; AVX2-FCP-NEXT: vmovaps %ymm4, 224(%rdx)
3903 ; AVX2-FCP-NEXT: vmovaps %ymm5, 192(%rdx)
3904 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3905 ; AVX2-FCP-NEXT: vmovaps %ymm0, 160(%rdx)
3906 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3907 ; AVX2-FCP-NEXT: vmovaps %ymm0, 128(%rdx)
3908 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3909 ; AVX2-FCP-NEXT: vmovaps %ymm0, 864(%rdx)
3910 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3911 ; AVX2-FCP-NEXT: vmovaps %ymm0, 832(%rdx)
3912 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3913 ; AVX2-FCP-NEXT: vmovaps %ymm0, 800(%rdx)
3914 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3915 ; AVX2-FCP-NEXT: vmovaps %ymm0, 768(%rdx)
3916 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3917 ; AVX2-FCP-NEXT: vmovaps %ymm0, 608(%rdx)
3918 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3919 ; AVX2-FCP-NEXT: vmovaps %ymm0, 576(%rdx)
3920 ; AVX2-FCP-NEXT: vmovups (%rsp), %ymm0 # 32-byte Reload
3921 ; AVX2-FCP-NEXT: vmovaps %ymm0, 544(%rdx)
3922 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3923 ; AVX2-FCP-NEXT: vmovaps %ymm0, 512(%rdx)
3924 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3925 ; AVX2-FCP-NEXT: vmovaps %ymm0, 352(%rdx)
3926 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3927 ; AVX2-FCP-NEXT: vmovaps %ymm0, 320(%rdx)
3928 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3929 ; AVX2-FCP-NEXT: vmovaps %ymm0, 288(%rdx)
3930 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3931 ; AVX2-FCP-NEXT: vmovaps %ymm0, 256(%rdx)
3932 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3933 ; AVX2-FCP-NEXT: vmovaps %ymm0, 96(%rdx)
3934 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3935 ; AVX2-FCP-NEXT: vmovaps %ymm0, 64(%rdx)
3936 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3937 ; AVX2-FCP-NEXT: vmovaps %ymm0, 32(%rdx)
3938 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
3939 ; AVX2-FCP-NEXT: vmovaps %ymm0, (%rdx)
3940 ; AVX2-FCP-NEXT: addq $488, %rsp # imm = 0x1E8
3941 ; AVX2-FCP-NEXT: vzeroupper
3942 ; AVX2-FCP-NEXT: retq
3943 ;
3944 ; AVX512-LABEL: store_i64_stride8_vf16:
3945 ; AVX512: # %bb.0:
3946 ; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %rax
3947 ; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %r10
3948 ; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %r11
3949 ; AVX512-NEXT: vmovdqa64 64(%rdi), %zmm11
3950 ; AVX512-NEXT: vmovdqa64 (%rdi), %zmm15
3951 ; AVX512-NEXT: vmovdqa64 64(%rsi), %zmm27
3952 ; AVX512-NEXT: vmovdqa64 (%rsi), %zmm17
3953 ; AVX512-NEXT: vmovdqa64 64(%rdx), %zmm14
3954 ; AVX512-NEXT: vmovdqa64 (%rdx), %zmm5
3955 ; AVX512-NEXT: vmovdqa64 64(%rcx), %zmm28
3956 ; AVX512-NEXT: vmovdqa64 (%rcx), %zmm26
3957 ; AVX512-NEXT: vmovdqa64 (%r8), %zmm6
3958 ; AVX512-NEXT: vmovdqa64 64(%r8), %zmm0
3959 ; AVX512-NEXT: vmovdqa64 (%r9), %zmm7
3960 ; AVX512-NEXT: vmovdqa64 64(%r9), %zmm1
3961 ; AVX512-NEXT: vmovdqa64 (%r11), %zmm8
3962 ; AVX512-NEXT: vmovdqa64 64(%r11), %zmm3
3963 ; AVX512-NEXT: vmovdqa64 (%r10), %zmm9
3964 ; AVX512-NEXT: vmovdqa64 64(%r10), %zmm4
3965 ; AVX512-NEXT: vbroadcasti32x4 {{.*#+}} zmm18 = [1,9,1,9,1,9,1,9]
3966 ; AVX512-NEXT: # zmm18 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
3967 ; AVX512-NEXT: vmovdqa64 %zmm8, %zmm10
3968 ; AVX512-NEXT: vpermt2q %zmm9, %zmm18, %zmm10
3969 ; AVX512-NEXT: vmovdqa64 %zmm6, %zmm12
3970 ; AVX512-NEXT: vpermt2q %zmm7, %zmm18, %zmm12
3971 ; AVX512-NEXT: movb $-64, %r8b
3972 ; AVX512-NEXT: kmovw %r8d, %k1
3973 ; AVX512-NEXT: vmovdqa64 %zmm10, %zmm12 {%k1}
3974 ; AVX512-NEXT: vmovdqa (%rsi), %xmm10
3975 ; AVX512-NEXT: vinserti128 $1, (%rcx), %ymm10, %ymm10
3976 ; AVX512-NEXT: vmovdqa64 (%rdi), %xmm16
3977 ; AVX512-NEXT: vinserti32x4 $1, (%rdx), %ymm16, %ymm16
3978 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} ymm19 = ymm16[1],ymm10[1],ymm16[3],ymm10[3]
3979 ; AVX512-NEXT: vinserti64x4 $0, %ymm19, %zmm12, %zmm21
3980 ; AVX512-NEXT: vbroadcasti32x4 {{.*#+}} zmm19 = [0,8,0,8,0,8,0,8]
3981 ; AVX512-NEXT: # zmm19 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
3982 ; AVX512-NEXT: vmovdqa64 %zmm8, %zmm12
3983 ; AVX512-NEXT: vpermt2q %zmm9, %zmm19, %zmm12
3984 ; AVX512-NEXT: vmovdqa64 %zmm6, %zmm20
3985 ; AVX512-NEXT: vpermt2q %zmm7, %zmm19, %zmm20
3986 ; AVX512-NEXT: vmovdqa64 %zmm12, %zmm20 {%k1}
3987 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} ymm10 = ymm16[0],ymm10[0],ymm16[2],ymm10[2]
3988 ; AVX512-NEXT: vinserti64x4 $0, %ymm10, %zmm20, %zmm22
3989 ; AVX512-NEXT: vbroadcasti32x4 {{.*#+}} zmm20 = [5,13,5,13,5,13,5,13]
3990 ; AVX512-NEXT: # zmm20 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
3991 ; AVX512-NEXT: vmovdqa64 %zmm5, %zmm10
3992 ; AVX512-NEXT: vpermt2q %zmm26, %zmm20, %zmm10
3993 ; AVX512-NEXT: vmovdqa64 %zmm15, %zmm12
3994 ; AVX512-NEXT: vpermt2q %zmm17, %zmm20, %zmm12
3995 ; AVX512-NEXT: vpblendd {{.*#+}} ymm10 = ymm12[0,1,2,3],ymm10[4,5,6,7]
3996 ; AVX512-NEXT: vmovdqa64 %zmm8, %zmm12
3997 ; AVX512-NEXT: vpermt2q %zmm9, %zmm20, %zmm12
3998 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} zmm16 = zmm6[1],zmm7[1],zmm6[3],zmm7[3],zmm6[5],zmm7[5],zmm6[7],zmm7[7]
3999 ; AVX512-NEXT: vmovdqa64 %zmm12, %zmm16 {%k1}
4000 ; AVX512-NEXT: vinserti64x4 $0, %ymm10, %zmm16, %zmm29
4001 ; AVX512-NEXT: vbroadcasti32x4 {{.*#+}} zmm23 = [4,12,4,12,4,12,4,12]
4002 ; AVX512-NEXT: # zmm23 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
4003 ; AVX512-NEXT: vmovdqa64 %zmm5, %zmm12
4004 ; AVX512-NEXT: vpermt2q %zmm26, %zmm23, %zmm12
4005 ; AVX512-NEXT: vmovdqa64 %zmm15, %zmm2
4006 ; AVX512-NEXT: vpermt2q %zmm17, %zmm23, %zmm2
4007 ; AVX512-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm12[4,5,6,7]
4008 ; AVX512-NEXT: vmovdqa64 %zmm8, %zmm12
4009 ; AVX512-NEXT: vpermt2q %zmm9, %zmm23, %zmm12
4010 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} zmm16 = zmm6[0],zmm7[0],zmm6[2],zmm7[2],zmm6[4],zmm7[4],zmm6[6],zmm7[6]
4011 ; AVX512-NEXT: vmovdqa64 %zmm12, %zmm16 {%k1}
4012 ; AVX512-NEXT: vinserti64x4 $0, %ymm2, %zmm16, %zmm31
4013 ; AVX512-NEXT: vbroadcasti32x4 {{.*#+}} zmm24 = [7,15,7,15,7,15,7,15]
4014 ; AVX512-NEXT: # zmm24 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
4015 ; AVX512-NEXT: vmovdqa64 %zmm5, %zmm2
4016 ; AVX512-NEXT: vpermt2q %zmm26, %zmm24, %zmm2
4017 ; AVX512-NEXT: vmovdqa64 %zmm15, %zmm13
4018 ; AVX512-NEXT: vpermt2q %zmm17, %zmm24, %zmm13
4019 ; AVX512-NEXT: vpblendd {{.*#+}} ymm2 = ymm13[0,1,2,3],ymm2[4,5,6,7]
4020 ; AVX512-NEXT: vmovdqa64 %zmm6, %zmm13
4021 ; AVX512-NEXT: vpermt2q %zmm7, %zmm24, %zmm13
4022 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} zmm13 {%k1} = zmm8[1],zmm9[1],zmm8[3],zmm9[3],zmm8[5],zmm9[5],zmm8[7],zmm9[7]
4023 ; AVX512-NEXT: vinserti64x4 $0, %ymm2, %zmm13, %zmm16
4024 ; AVX512-NEXT: vbroadcasti32x4 {{.*#+}} zmm25 = [6,14,6,14,6,14,6,14]
4025 ; AVX512-NEXT: # zmm25 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
4026 ; AVX512-NEXT: vpermt2q %zmm26, %zmm25, %zmm5
4027 ; AVX512-NEXT: vpermt2q %zmm17, %zmm25, %zmm15
4028 ; AVX512-NEXT: vpblendd {{.*#+}} ymm2 = ymm15[0,1,2,3],ymm5[4,5,6,7]
4029 ; AVX512-NEXT: vmovdqa64 %zmm6, %zmm5
4030 ; AVX512-NEXT: vpermt2q %zmm7, %zmm25, %zmm5
4031 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} zmm5 {%k1} = zmm8[0],zmm9[0],zmm8[2],zmm9[2],zmm8[4],zmm9[4],zmm8[6],zmm9[6]
4032 ; AVX512-NEXT: vinserti64x4 $0, %ymm2, %zmm5, %zmm17
4033 ; AVX512-NEXT: vmovdqa64 %zmm3, %zmm2
4034 ; AVX512-NEXT: vpermt2q %zmm4, %zmm18, %zmm2
4035 ; AVX512-NEXT: vpermi2q %zmm1, %zmm0, %zmm18
4036 ; AVX512-NEXT: vmovdqa64 %zmm2, %zmm18 {%k1}
4037 ; AVX512-NEXT: vmovdqa 64(%rsi), %xmm2
4038 ; AVX512-NEXT: vinserti128 $1, 64(%rcx), %ymm2, %ymm2
4039 ; AVX512-NEXT: vmovdqa 64(%rdi), %xmm5
4040 ; AVX512-NEXT: vinserti128 $1, 64(%rdx), %ymm5, %ymm5
4041 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} ymm13 = ymm5[1],ymm2[1],ymm5[3],ymm2[3]
4042 ; AVX512-NEXT: vinserti64x4 $0, %ymm13, %zmm18, %zmm18
4043 ; AVX512-NEXT: vmovdqa64 %zmm3, %zmm13
4044 ; AVX512-NEXT: vpermt2q %zmm4, %zmm19, %zmm13
4045 ; AVX512-NEXT: vpermi2q %zmm1, %zmm0, %zmm19
4046 ; AVX512-NEXT: vmovdqa64 %zmm13, %zmm19 {%k1}
4047 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} ymm2 = ymm5[0],ymm2[0],ymm5[2],ymm2[2]
4048 ; AVX512-NEXT: vinserti64x4 $0, %ymm2, %zmm19, %zmm19
4049 ; AVX512-NEXT: vmovdqa64 %zmm14, %zmm2
4050 ; AVX512-NEXT: vpermt2q %zmm28, %zmm20, %zmm2
4051 ; AVX512-NEXT: vmovdqa64 %zmm11, %zmm5
4052 ; AVX512-NEXT: vpermt2q %zmm27, %zmm20, %zmm5
4053 ; AVX512-NEXT: vpblendd {{.*#+}} ymm2 = ymm5[0,1,2,3],ymm2[4,5,6,7]
4054 ; AVX512-NEXT: vpermi2q %zmm4, %zmm3, %zmm20
4055 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} zmm5 = zmm0[1],zmm1[1],zmm0[3],zmm1[3],zmm0[5],zmm1[5],zmm0[7],zmm1[7]
4056 ; AVX512-NEXT: vmovdqa64 %zmm20, %zmm5 {%k1}
4057 ; AVX512-NEXT: vinserti64x4 $0, %ymm2, %zmm5, %zmm20
4058 ; AVX512-NEXT: vmovdqa64 %zmm14, %zmm2
4059 ; AVX512-NEXT: vpermt2q %zmm28, %zmm23, %zmm2
4060 ; AVX512-NEXT: vmovdqa64 %zmm11, %zmm5
4061 ; AVX512-NEXT: vpermt2q %zmm27, %zmm23, %zmm5
4062 ; AVX512-NEXT: vpblendd {{.*#+}} ymm2 = ymm5[0,1,2,3],ymm2[4,5,6,7]
4063 ; AVX512-NEXT: vpermi2q %zmm4, %zmm3, %zmm23
4064 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} zmm5 = zmm0[0],zmm1[0],zmm0[2],zmm1[2],zmm0[4],zmm1[4],zmm0[6],zmm1[6]
4065 ; AVX512-NEXT: vmovdqa64 %zmm23, %zmm5 {%k1}
4066 ; AVX512-NEXT: vinserti64x4 $0, %ymm2, %zmm5, %zmm2
4067 ; AVX512-NEXT: vmovdqa64 %zmm14, %zmm5
4068 ; AVX512-NEXT: vpermt2q %zmm28, %zmm24, %zmm5
4069 ; AVX512-NEXT: vmovdqa64 %zmm11, %zmm13
4070 ; AVX512-NEXT: vpermt2q %zmm27, %zmm24, %zmm13
4071 ; AVX512-NEXT: vpblendd {{.*#+}} ymm5 = ymm13[0,1,2,3],ymm5[4,5,6,7]
4072 ; AVX512-NEXT: vpermi2q %zmm1, %zmm0, %zmm24
4073 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} zmm24 {%k1} = zmm3[1],zmm4[1],zmm3[3],zmm4[3],zmm3[5],zmm4[5],zmm3[7],zmm4[7]
4074 ; AVX512-NEXT: vinserti64x4 $0, %ymm5, %zmm24, %zmm5
4075 ; AVX512-NEXT: vpermt2q %zmm28, %zmm25, %zmm14
4076 ; AVX512-NEXT: vpermt2q %zmm27, %zmm25, %zmm11
4077 ; AVX512-NEXT: vpblendd {{.*#+}} ymm11 = ymm11[0,1,2,3],ymm14[4,5,6,7]
4078 ; AVX512-NEXT: vpermi2q %zmm1, %zmm0, %zmm25
4079 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} zmm25 {%k1} = zmm3[0],zmm4[0],zmm3[2],zmm4[2],zmm3[4],zmm4[4],zmm3[6],zmm4[6]
4080 ; AVX512-NEXT: vinserti64x4 $0, %ymm11, %zmm25, %zmm11
4081 ; AVX512-NEXT: vbroadcasti32x4 {{.*#+}} zmm13 = [3,11,3,11,3,11,3,11]
4082 ; AVX512-NEXT: # zmm13 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
4083 ; AVX512-NEXT: vmovdqa64 %zmm8, %zmm14
4084 ; AVX512-NEXT: vpermt2q %zmm9, %zmm13, %zmm14
4085 ; AVX512-NEXT: vmovdqa64 %zmm6, %zmm15
4086 ; AVX512-NEXT: vpermt2q %zmm7, %zmm13, %zmm15
4087 ; AVX512-NEXT: vmovdqa64 %zmm14, %zmm15 {%k1}
4088 ; AVX512-NEXT: vmovdqa (%rcx), %ymm14
4089 ; AVX512-NEXT: vmovdqa64 64(%rcx), %ymm23
4090 ; AVX512-NEXT: vmovdqa64 (%rdx), %ymm24
4091 ; AVX512-NEXT: vmovdqa64 64(%rdx), %ymm25
4092 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} ymm10 = ymm24[1],ymm14[1],ymm24[3],ymm14[3]
4093 ; AVX512-NEXT: vmovdqa64 (%rsi), %ymm26
4094 ; AVX512-NEXT: vmovdqa64 64(%rsi), %ymm27
4095 ; AVX512-NEXT: vmovdqa64 (%rdi), %ymm28
4096 ; AVX512-NEXT: vmovdqa64 64(%rdi), %ymm30
4097 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} ymm12 = ymm28[1],ymm26[1],ymm28[3],ymm26[3]
4098 ; AVX512-NEXT: vperm2i128 {{.*#+}} ymm10 = ymm12[2,3],ymm10[2,3]
4099 ; AVX512-NEXT: vinserti64x4 $0, %ymm10, %zmm15, %zmm10
4100 ; AVX512-NEXT: vbroadcasti32x4 {{.*#+}} zmm12 = [2,10,2,10,2,10,2,10]
4101 ; AVX512-NEXT: # zmm12 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
4102 ; AVX512-NEXT: vpermt2q %zmm9, %zmm12, %zmm8
4103 ; AVX512-NEXT: vpermt2q %zmm7, %zmm12, %zmm6
4104 ; AVX512-NEXT: vmovdqa64 %zmm8, %zmm6 {%k1}
4105 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm24[0],ymm14[0],ymm24[2],ymm14[2]
4106 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} ymm8 = ymm28[0],ymm26[0],ymm28[2],ymm26[2]
4107 ; AVX512-NEXT: vperm2i128 {{.*#+}} ymm7 = ymm8[2,3],ymm7[2,3]
4108 ; AVX512-NEXT: vinserti64x4 $0, %ymm7, %zmm6, %zmm6
4109 ; AVX512-NEXT: vmovdqa64 %zmm3, %zmm7
4110 ; AVX512-NEXT: vpermt2q %zmm4, %zmm13, %zmm7
4111 ; AVX512-NEXT: vpermi2q %zmm1, %zmm0, %zmm13
4112 ; AVX512-NEXT: vmovdqa64 %zmm7, %zmm13 {%k1}
4113 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} ymm7 = ymm25[1],ymm23[1],ymm25[3],ymm23[3]
4114 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} ymm8 = ymm30[1],ymm27[1],ymm30[3],ymm27[3]
4115 ; AVX512-NEXT: vperm2i128 {{.*#+}} ymm7 = ymm8[2,3],ymm7[2,3]
4116 ; AVX512-NEXT: vinserti64x4 $0, %ymm7, %zmm13, %zmm7
4117 ; AVX512-NEXT: vpermt2q %zmm4, %zmm12, %zmm3
4118 ; AVX512-NEXT: vpermt2q %zmm1, %zmm12, %zmm0
4119 ; AVX512-NEXT: vmovdqa64 %zmm3, %zmm0 {%k1}
4120 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} ymm1 = ymm25[0],ymm23[0],ymm25[2],ymm23[2]
4121 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm30[0],ymm27[0],ymm30[2],ymm27[2]
4122 ; AVX512-NEXT: vperm2i128 {{.*#+}} ymm1 = ymm3[2,3],ymm1[2,3]
4123 ; AVX512-NEXT: vinserti64x4 $0, %ymm1, %zmm0, %zmm0
4124 ; AVX512-NEXT: vmovdqa64 %zmm0, 640(%rax)
4125 ; AVX512-NEXT: vmovdqa64 %zmm7, 704(%rax)
4126 ; AVX512-NEXT: vmovdqa64 %zmm6, 128(%rax)
4127 ; AVX512-NEXT: vmovdqa64 %zmm10, 192(%rax)
4128 ; AVX512-NEXT: vmovdqa64 %zmm11, 896(%rax)
4129 ; AVX512-NEXT: vmovdqa64 %zmm5, 960(%rax)
4130 ; AVX512-NEXT: vmovdqa64 %zmm2, 768(%rax)
4131 ; AVX512-NEXT: vmovdqa64 %zmm20, 832(%rax)
4132 ; AVX512-NEXT: vmovdqa64 %zmm19, 512(%rax)
4133 ; AVX512-NEXT: vmovdqa64 %zmm18, 576(%rax)
4134 ; AVX512-NEXT: vmovdqa64 %zmm17, 384(%rax)
4135 ; AVX512-NEXT: vmovdqa64 %zmm16, 448(%rax)
4136 ; AVX512-NEXT: vmovdqa64 %zmm31, 256(%rax)
4137 ; AVX512-NEXT: vmovdqa64 %zmm29, 320(%rax)
4138 ; AVX512-NEXT: vmovdqa64 %zmm22, (%rax)
4139 ; AVX512-NEXT: vmovdqa64 %zmm21, 64(%rax)
4140 ; AVX512-NEXT: vzeroupper
4141 ; AVX512-NEXT: retq
4142 ;
4143 ; AVX512-FCP-LABEL: store_i64_stride8_vf16:
4144 ; AVX512-FCP: # %bb.0:
4145 ; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
4146 ; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10
4147 ; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r11
4148 ; AVX512-FCP-NEXT: vmovdqa64 64(%rdi), %zmm11
4149 ; AVX512-FCP-NEXT: vmovdqa64 (%rdi), %zmm15
4150 ; AVX512-FCP-NEXT: vmovdqa64 64(%rsi), %zmm27
4151 ; AVX512-FCP-NEXT: vmovdqa64 (%rsi), %zmm17
4152 ; AVX512-FCP-NEXT: vmovdqa64 64(%rdx), %zmm14
4153 ; AVX512-FCP-NEXT: vmovdqa64 (%rdx), %zmm5
4154 ; AVX512-FCP-NEXT: vmovdqa64 64(%rcx), %zmm28
4155 ; AVX512-FCP-NEXT: vmovdqa64 (%rcx), %zmm26
4156 ; AVX512-FCP-NEXT: vmovdqa64 (%r8), %zmm6
4157 ; AVX512-FCP-NEXT: vmovdqa64 64(%r8), %zmm0
4158 ; AVX512-FCP-NEXT: vmovdqa64 (%r9), %zmm7
4159 ; AVX512-FCP-NEXT: vmovdqa64 64(%r9), %zmm1
4160 ; AVX512-FCP-NEXT: vmovdqa64 (%r11), %zmm8
4161 ; AVX512-FCP-NEXT: vmovdqa64 64(%r11), %zmm3
4162 ; AVX512-FCP-NEXT: vmovdqa64 (%r10), %zmm9
4163 ; AVX512-FCP-NEXT: vmovdqa64 64(%r10), %zmm4
4164 ; AVX512-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm18 = [1,9,1,9,1,9,1,9]
4165 ; AVX512-FCP-NEXT: # zmm18 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
4166 ; AVX512-FCP-NEXT: vmovdqa64 %zmm8, %zmm10
4167 ; AVX512-FCP-NEXT: vpermt2q %zmm9, %zmm18, %zmm10
4168 ; AVX512-FCP-NEXT: vmovdqa64 %zmm6, %zmm12
4169 ; AVX512-FCP-NEXT: vpermt2q %zmm7, %zmm18, %zmm12
4170 ; AVX512-FCP-NEXT: movb $-64, %r8b
4171 ; AVX512-FCP-NEXT: kmovw %r8d, %k1
4172 ; AVX512-FCP-NEXT: vmovdqa64 %zmm10, %zmm12 {%k1}
4173 ; AVX512-FCP-NEXT: vmovdqa (%rsi), %xmm10
4174 ; AVX512-FCP-NEXT: vinserti128 $1, (%rcx), %ymm10, %ymm10
4175 ; AVX512-FCP-NEXT: vmovdqa64 (%rdi), %xmm16
4176 ; AVX512-FCP-NEXT: vinserti32x4 $1, (%rdx), %ymm16, %ymm16
4177 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm19 = ymm16[1],ymm10[1],ymm16[3],ymm10[3]
4178 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm19, %zmm12, %zmm21
4179 ; AVX512-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm19 = [0,8,0,8,0,8,0,8]
4180 ; AVX512-FCP-NEXT: # zmm19 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
4181 ; AVX512-FCP-NEXT: vmovdqa64 %zmm8, %zmm12
4182 ; AVX512-FCP-NEXT: vpermt2q %zmm9, %zmm19, %zmm12
4183 ; AVX512-FCP-NEXT: vmovdqa64 %zmm6, %zmm20
4184 ; AVX512-FCP-NEXT: vpermt2q %zmm7, %zmm19, %zmm20
4185 ; AVX512-FCP-NEXT: vmovdqa64 %zmm12, %zmm20 {%k1}
4186 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm10 = ymm16[0],ymm10[0],ymm16[2],ymm10[2]
4187 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm10, %zmm20, %zmm22
4188 ; AVX512-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm20 = [5,13,5,13,5,13,5,13]
4189 ; AVX512-FCP-NEXT: # zmm20 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
4190 ; AVX512-FCP-NEXT: vmovdqa64 %zmm5, %zmm10
4191 ; AVX512-FCP-NEXT: vpermt2q %zmm26, %zmm20, %zmm10
4192 ; AVX512-FCP-NEXT: vmovdqa64 %zmm15, %zmm12
4193 ; AVX512-FCP-NEXT: vpermt2q %zmm17, %zmm20, %zmm12
4194 ; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm10 = ymm12[0,1,2,3],ymm10[4,5,6,7]
4195 ; AVX512-FCP-NEXT: vmovdqa64 %zmm8, %zmm12
4196 ; AVX512-FCP-NEXT: vpermt2q %zmm9, %zmm20, %zmm12
4197 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm16 = zmm6[1],zmm7[1],zmm6[3],zmm7[3],zmm6[5],zmm7[5],zmm6[7],zmm7[7]
4198 ; AVX512-FCP-NEXT: vmovdqa64 %zmm12, %zmm16 {%k1}
4199 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm10, %zmm16, %zmm29
4200 ; AVX512-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm23 = [4,12,4,12,4,12,4,12]
4201 ; AVX512-FCP-NEXT: # zmm23 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
4202 ; AVX512-FCP-NEXT: vmovdqa64 %zmm5, %zmm12
4203 ; AVX512-FCP-NEXT: vpermt2q %zmm26, %zmm23, %zmm12
4204 ; AVX512-FCP-NEXT: vmovdqa64 %zmm15, %zmm2
4205 ; AVX512-FCP-NEXT: vpermt2q %zmm17, %zmm23, %zmm2
4206 ; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm12[4,5,6,7]
4207 ; AVX512-FCP-NEXT: vmovdqa64 %zmm8, %zmm12
4208 ; AVX512-FCP-NEXT: vpermt2q %zmm9, %zmm23, %zmm12
4209 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm16 = zmm6[0],zmm7[0],zmm6[2],zmm7[2],zmm6[4],zmm7[4],zmm6[6],zmm7[6]
4210 ; AVX512-FCP-NEXT: vmovdqa64 %zmm12, %zmm16 {%k1}
4211 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm16, %zmm31
4212 ; AVX512-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm24 = [7,15,7,15,7,15,7,15]
4213 ; AVX512-FCP-NEXT: # zmm24 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
4214 ; AVX512-FCP-NEXT: vmovdqa64 %zmm5, %zmm2
4215 ; AVX512-FCP-NEXT: vpermt2q %zmm26, %zmm24, %zmm2
4216 ; AVX512-FCP-NEXT: vmovdqa64 %zmm15, %zmm13
4217 ; AVX512-FCP-NEXT: vpermt2q %zmm17, %zmm24, %zmm13
4218 ; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm13[0,1,2,3],ymm2[4,5,6,7]
4219 ; AVX512-FCP-NEXT: vmovdqa64 %zmm6, %zmm13
4220 ; AVX512-FCP-NEXT: vpermt2q %zmm7, %zmm24, %zmm13
4221 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm13 {%k1} = zmm8[1],zmm9[1],zmm8[3],zmm9[3],zmm8[5],zmm9[5],zmm8[7],zmm9[7]
4222 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm13, %zmm16
4223 ; AVX512-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm25 = [6,14,6,14,6,14,6,14]
4224 ; AVX512-FCP-NEXT: # zmm25 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
4225 ; AVX512-FCP-NEXT: vpermt2q %zmm26, %zmm25, %zmm5
4226 ; AVX512-FCP-NEXT: vpermt2q %zmm17, %zmm25, %zmm15
4227 ; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm15[0,1,2,3],ymm5[4,5,6,7]
4228 ; AVX512-FCP-NEXT: vmovdqa64 %zmm6, %zmm5
4229 ; AVX512-FCP-NEXT: vpermt2q %zmm7, %zmm25, %zmm5
4230 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm5 {%k1} = zmm8[0],zmm9[0],zmm8[2],zmm9[2],zmm8[4],zmm9[4],zmm8[6],zmm9[6]
4231 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm5, %zmm17
4232 ; AVX512-FCP-NEXT: vmovdqa64 %zmm3, %zmm2
4233 ; AVX512-FCP-NEXT: vpermt2q %zmm4, %zmm18, %zmm2
4234 ; AVX512-FCP-NEXT: vpermi2q %zmm1, %zmm0, %zmm18
4235 ; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm18 {%k1}
4236 ; AVX512-FCP-NEXT: vmovdqa 64(%rsi), %xmm2
4237 ; AVX512-FCP-NEXT: vinserti128 $1, 64(%rcx), %ymm2, %ymm2
4238 ; AVX512-FCP-NEXT: vmovdqa 64(%rdi), %xmm5
4239 ; AVX512-FCP-NEXT: vinserti128 $1, 64(%rdx), %ymm5, %ymm5
4240 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm13 = ymm5[1],ymm2[1],ymm5[3],ymm2[3]
4241 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm13, %zmm18, %zmm18
4242 ; AVX512-FCP-NEXT: vmovdqa64 %zmm3, %zmm13
4243 ; AVX512-FCP-NEXT: vpermt2q %zmm4, %zmm19, %zmm13
4244 ; AVX512-FCP-NEXT: vpermi2q %zmm1, %zmm0, %zmm19
4245 ; AVX512-FCP-NEXT: vmovdqa64 %zmm13, %zmm19 {%k1}
4246 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm2 = ymm5[0],ymm2[0],ymm5[2],ymm2[2]
4247 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm19, %zmm19
4248 ; AVX512-FCP-NEXT: vmovdqa64 %zmm14, %zmm2
4249 ; AVX512-FCP-NEXT: vpermt2q %zmm28, %zmm20, %zmm2
4250 ; AVX512-FCP-NEXT: vmovdqa64 %zmm11, %zmm5
4251 ; AVX512-FCP-NEXT: vpermt2q %zmm27, %zmm20, %zmm5
4252 ; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm5[0,1,2,3],ymm2[4,5,6,7]
4253 ; AVX512-FCP-NEXT: vpermi2q %zmm4, %zmm3, %zmm20
4254 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm5 = zmm0[1],zmm1[1],zmm0[3],zmm1[3],zmm0[5],zmm1[5],zmm0[7],zmm1[7]
4255 ; AVX512-FCP-NEXT: vmovdqa64 %zmm20, %zmm5 {%k1}
4256 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm5, %zmm20
4257 ; AVX512-FCP-NEXT: vmovdqa64 %zmm14, %zmm2
4258 ; AVX512-FCP-NEXT: vpermt2q %zmm28, %zmm23, %zmm2
4259 ; AVX512-FCP-NEXT: vmovdqa64 %zmm11, %zmm5
4260 ; AVX512-FCP-NEXT: vpermt2q %zmm27, %zmm23, %zmm5
4261 ; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm5[0,1,2,3],ymm2[4,5,6,7]
4262 ; AVX512-FCP-NEXT: vpermi2q %zmm4, %zmm3, %zmm23
4263 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm5 = zmm0[0],zmm1[0],zmm0[2],zmm1[2],zmm0[4],zmm1[4],zmm0[6],zmm1[6]
4264 ; AVX512-FCP-NEXT: vmovdqa64 %zmm23, %zmm5 {%k1}
4265 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm5, %zmm2
4266 ; AVX512-FCP-NEXT: vmovdqa64 %zmm14, %zmm5
4267 ; AVX512-FCP-NEXT: vpermt2q %zmm28, %zmm24, %zmm5
4268 ; AVX512-FCP-NEXT: vmovdqa64 %zmm11, %zmm13
4269 ; AVX512-FCP-NEXT: vpermt2q %zmm27, %zmm24, %zmm13
4270 ; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm5 = ymm13[0,1,2,3],ymm5[4,5,6,7]
4271 ; AVX512-FCP-NEXT: vpermi2q %zmm1, %zmm0, %zmm24
4272 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm24 {%k1} = zmm3[1],zmm4[1],zmm3[3],zmm4[3],zmm3[5],zmm4[5],zmm3[7],zmm4[7]
4273 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm5, %zmm24, %zmm5
4274 ; AVX512-FCP-NEXT: vpermt2q %zmm28, %zmm25, %zmm14
4275 ; AVX512-FCP-NEXT: vpermt2q %zmm27, %zmm25, %zmm11
4276 ; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm11 = ymm11[0,1,2,3],ymm14[4,5,6,7]
4277 ; AVX512-FCP-NEXT: vpermi2q %zmm1, %zmm0, %zmm25
4278 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm25 {%k1} = zmm3[0],zmm4[0],zmm3[2],zmm4[2],zmm3[4],zmm4[4],zmm3[6],zmm4[6]
4279 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm11, %zmm25, %zmm11
4280 ; AVX512-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm13 = [3,11,3,11,3,11,3,11]
4281 ; AVX512-FCP-NEXT: # zmm13 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
4282 ; AVX512-FCP-NEXT: vmovdqa64 %zmm8, %zmm14
4283 ; AVX512-FCP-NEXT: vpermt2q %zmm9, %zmm13, %zmm14
4284 ; AVX512-FCP-NEXT: vmovdqa64 %zmm6, %zmm15
4285 ; AVX512-FCP-NEXT: vpermt2q %zmm7, %zmm13, %zmm15
4286 ; AVX512-FCP-NEXT: vmovdqa64 %zmm14, %zmm15 {%k1}
4287 ; AVX512-FCP-NEXT: vmovdqa (%rcx), %ymm14
4288 ; AVX512-FCP-NEXT: vmovdqa64 64(%rcx), %ymm23
4289 ; AVX512-FCP-NEXT: vmovdqa64 (%rdx), %ymm24
4290 ; AVX512-FCP-NEXT: vmovdqa64 64(%rdx), %ymm25
4291 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm10 = ymm24[1],ymm14[1],ymm24[3],ymm14[3]
4292 ; AVX512-FCP-NEXT: vmovdqa64 (%rsi), %ymm26
4293 ; AVX512-FCP-NEXT: vmovdqa64 64(%rsi), %ymm27
4294 ; AVX512-FCP-NEXT: vmovdqa64 (%rdi), %ymm28
4295 ; AVX512-FCP-NEXT: vmovdqa64 64(%rdi), %ymm30
4296 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm12 = ymm28[1],ymm26[1],ymm28[3],ymm26[3]
4297 ; AVX512-FCP-NEXT: vperm2i128 {{.*#+}} ymm10 = ymm12[2,3],ymm10[2,3]
4298 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm10, %zmm15, %zmm10
4299 ; AVX512-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm12 = [2,10,2,10,2,10,2,10]
4300 ; AVX512-FCP-NEXT: # zmm12 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
4301 ; AVX512-FCP-NEXT: vpermt2q %zmm9, %zmm12, %zmm8
4302 ; AVX512-FCP-NEXT: vpermt2q %zmm7, %zmm12, %zmm6
4303 ; AVX512-FCP-NEXT: vmovdqa64 %zmm8, %zmm6 {%k1}
4304 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm24[0],ymm14[0],ymm24[2],ymm14[2]
4305 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm8 = ymm28[0],ymm26[0],ymm28[2],ymm26[2]
4306 ; AVX512-FCP-NEXT: vperm2i128 {{.*#+}} ymm7 = ymm8[2,3],ymm7[2,3]
4307 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm7, %zmm6, %zmm6
4308 ; AVX512-FCP-NEXT: vmovdqa64 %zmm3, %zmm7
4309 ; AVX512-FCP-NEXT: vpermt2q %zmm4, %zmm13, %zmm7
4310 ; AVX512-FCP-NEXT: vpermi2q %zmm1, %zmm0, %zmm13
4311 ; AVX512-FCP-NEXT: vmovdqa64 %zmm7, %zmm13 {%k1}
4312 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm7 = ymm25[1],ymm23[1],ymm25[3],ymm23[3]
4313 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm8 = ymm30[1],ymm27[1],ymm30[3],ymm27[3]
4314 ; AVX512-FCP-NEXT: vperm2i128 {{.*#+}} ymm7 = ymm8[2,3],ymm7[2,3]
4315 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm7, %zmm13, %zmm7
4316 ; AVX512-FCP-NEXT: vpermt2q %zmm4, %zmm12, %zmm3
4317 ; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm12, %zmm0
4318 ; AVX512-FCP-NEXT: vmovdqa64 %zmm3, %zmm0 {%k1}
4319 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm1 = ymm25[0],ymm23[0],ymm25[2],ymm23[2]
4320 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm30[0],ymm27[0],ymm30[2],ymm27[2]
4321 ; AVX512-FCP-NEXT: vperm2i128 {{.*#+}} ymm1 = ymm3[2,3],ymm1[2,3]
4322 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm1, %zmm0, %zmm0
4323 ; AVX512-FCP-NEXT: vmovdqa64 %zmm0, 640(%rax)
4324 ; AVX512-FCP-NEXT: vmovdqa64 %zmm7, 704(%rax)
4325 ; AVX512-FCP-NEXT: vmovdqa64 %zmm6, 128(%rax)
4326 ; AVX512-FCP-NEXT: vmovdqa64 %zmm10, 192(%rax)
4327 ; AVX512-FCP-NEXT: vmovdqa64 %zmm11, 896(%rax)
4328 ; AVX512-FCP-NEXT: vmovdqa64 %zmm5, 960(%rax)
4329 ; AVX512-FCP-NEXT: vmovdqa64 %zmm2, 768(%rax)
4330 ; AVX512-FCP-NEXT: vmovdqa64 %zmm20, 832(%rax)
4331 ; AVX512-FCP-NEXT: vmovdqa64 %zmm19, 512(%rax)
4332 ; AVX512-FCP-NEXT: vmovdqa64 %zmm18, 576(%rax)
4333 ; AVX512-FCP-NEXT: vmovdqa64 %zmm17, 384(%rax)
4334 ; AVX512-FCP-NEXT: vmovdqa64 %zmm16, 448(%rax)
4335 ; AVX512-FCP-NEXT: vmovdqa64 %zmm31, 256(%rax)
4336 ; AVX512-FCP-NEXT: vmovdqa64 %zmm29, 320(%rax)
4337 ; AVX512-FCP-NEXT: vmovdqa64 %zmm22, (%rax)
4338 ; AVX512-FCP-NEXT: vmovdqa64 %zmm21, 64(%rax)
4339 ; AVX512-FCP-NEXT: vzeroupper
4340 ; AVX512-FCP-NEXT: retq
4341 ;
4342 ; AVX512DQ-LABEL: store_i64_stride8_vf16:
4343 ; AVX512DQ: # %bb.0:
4344 ; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %rax
4345 ; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %r10
4346 ; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %r11
4347 ; AVX512DQ-NEXT: vmovdqa64 64(%rdi), %zmm11
4348 ; AVX512DQ-NEXT: vmovdqa64 (%rdi), %zmm15
4349 ; AVX512DQ-NEXT: vmovdqa64 64(%rsi), %zmm27
4350 ; AVX512DQ-NEXT: vmovdqa64 (%rsi), %zmm17
4351 ; AVX512DQ-NEXT: vmovdqa64 64(%rdx), %zmm14
4352 ; AVX512DQ-NEXT: vmovdqa64 (%rdx), %zmm5
4353 ; AVX512DQ-NEXT: vmovdqa64 64(%rcx), %zmm28
4354 ; AVX512DQ-NEXT: vmovdqa64 (%rcx), %zmm26
4355 ; AVX512DQ-NEXT: vmovdqa64 (%r8), %zmm6
4356 ; AVX512DQ-NEXT: vmovdqa64 64(%r8), %zmm0
4357 ; AVX512DQ-NEXT: vmovdqa64 (%r9), %zmm7
4358 ; AVX512DQ-NEXT: vmovdqa64 64(%r9), %zmm1
4359 ; AVX512DQ-NEXT: vmovdqa64 (%r11), %zmm8
4360 ; AVX512DQ-NEXT: vmovdqa64 64(%r11), %zmm3
4361 ; AVX512DQ-NEXT: vmovdqa64 (%r10), %zmm9
4362 ; AVX512DQ-NEXT: vmovdqa64 64(%r10), %zmm4
4363 ; AVX512DQ-NEXT: vbroadcasti32x4 {{.*#+}} zmm18 = [1,9,1,9,1,9,1,9]
4364 ; AVX512DQ-NEXT: # zmm18 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
4365 ; AVX512DQ-NEXT: vmovdqa64 %zmm8, %zmm10
4366 ; AVX512DQ-NEXT: vpermt2q %zmm9, %zmm18, %zmm10
4367 ; AVX512DQ-NEXT: vmovdqa64 %zmm6, %zmm12
4368 ; AVX512DQ-NEXT: vpermt2q %zmm7, %zmm18, %zmm12
4369 ; AVX512DQ-NEXT: movb $-64, %r8b
4370 ; AVX512DQ-NEXT: kmovw %r8d, %k1
4371 ; AVX512DQ-NEXT: vmovdqa64 %zmm10, %zmm12 {%k1}
4372 ; AVX512DQ-NEXT: vmovdqa (%rsi), %xmm10
4373 ; AVX512DQ-NEXT: vinserti128 $1, (%rcx), %ymm10, %ymm10
4374 ; AVX512DQ-NEXT: vmovdqa64 (%rdi), %xmm16
4375 ; AVX512DQ-NEXT: vinserti32x4 $1, (%rdx), %ymm16, %ymm16
4376 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} ymm19 = ymm16[1],ymm10[1],ymm16[3],ymm10[3]
4377 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm19, %zmm12, %zmm21
4378 ; AVX512DQ-NEXT: vbroadcasti32x4 {{.*#+}} zmm19 = [0,8,0,8,0,8,0,8]
4379 ; AVX512DQ-NEXT: # zmm19 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
4380 ; AVX512DQ-NEXT: vmovdqa64 %zmm8, %zmm12
4381 ; AVX512DQ-NEXT: vpermt2q %zmm9, %zmm19, %zmm12
4382 ; AVX512DQ-NEXT: vmovdqa64 %zmm6, %zmm20
4383 ; AVX512DQ-NEXT: vpermt2q %zmm7, %zmm19, %zmm20
4384 ; AVX512DQ-NEXT: vmovdqa64 %zmm12, %zmm20 {%k1}
4385 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} ymm10 = ymm16[0],ymm10[0],ymm16[2],ymm10[2]
4386 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm10, %zmm20, %zmm22
4387 ; AVX512DQ-NEXT: vbroadcasti32x4 {{.*#+}} zmm20 = [5,13,5,13,5,13,5,13]
4388 ; AVX512DQ-NEXT: # zmm20 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
4389 ; AVX512DQ-NEXT: vmovdqa64 %zmm5, %zmm10
4390 ; AVX512DQ-NEXT: vpermt2q %zmm26, %zmm20, %zmm10
4391 ; AVX512DQ-NEXT: vmovdqa64 %zmm15, %zmm12
4392 ; AVX512DQ-NEXT: vpermt2q %zmm17, %zmm20, %zmm12
4393 ; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm10 = ymm12[0,1,2,3],ymm10[4,5,6,7]
4394 ; AVX512DQ-NEXT: vmovdqa64 %zmm8, %zmm12
4395 ; AVX512DQ-NEXT: vpermt2q %zmm9, %zmm20, %zmm12
4396 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} zmm16 = zmm6[1],zmm7[1],zmm6[3],zmm7[3],zmm6[5],zmm7[5],zmm6[7],zmm7[7]
4397 ; AVX512DQ-NEXT: vmovdqa64 %zmm12, %zmm16 {%k1}
4398 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm10, %zmm16, %zmm29
4399 ; AVX512DQ-NEXT: vbroadcasti32x4 {{.*#+}} zmm23 = [4,12,4,12,4,12,4,12]
4400 ; AVX512DQ-NEXT: # zmm23 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
4401 ; AVX512DQ-NEXT: vmovdqa64 %zmm5, %zmm12
4402 ; AVX512DQ-NEXT: vpermt2q %zmm26, %zmm23, %zmm12
4403 ; AVX512DQ-NEXT: vmovdqa64 %zmm15, %zmm2
4404 ; AVX512DQ-NEXT: vpermt2q %zmm17, %zmm23, %zmm2
4405 ; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm12[4,5,6,7]
4406 ; AVX512DQ-NEXT: vmovdqa64 %zmm8, %zmm12
4407 ; AVX512DQ-NEXT: vpermt2q %zmm9, %zmm23, %zmm12
4408 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} zmm16 = zmm6[0],zmm7[0],zmm6[2],zmm7[2],zmm6[4],zmm7[4],zmm6[6],zmm7[6]
4409 ; AVX512DQ-NEXT: vmovdqa64 %zmm12, %zmm16 {%k1}
4410 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm2, %zmm16, %zmm31
4411 ; AVX512DQ-NEXT: vbroadcasti32x4 {{.*#+}} zmm24 = [7,15,7,15,7,15,7,15]
4412 ; AVX512DQ-NEXT: # zmm24 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
4413 ; AVX512DQ-NEXT: vmovdqa64 %zmm5, %zmm2
4414 ; AVX512DQ-NEXT: vpermt2q %zmm26, %zmm24, %zmm2
4415 ; AVX512DQ-NEXT: vmovdqa64 %zmm15, %zmm13
4416 ; AVX512DQ-NEXT: vpermt2q %zmm17, %zmm24, %zmm13
4417 ; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm2 = ymm13[0,1,2,3],ymm2[4,5,6,7]
4418 ; AVX512DQ-NEXT: vmovdqa64 %zmm6, %zmm13
4419 ; AVX512DQ-NEXT: vpermt2q %zmm7, %zmm24, %zmm13
4420 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} zmm13 {%k1} = zmm8[1],zmm9[1],zmm8[3],zmm9[3],zmm8[5],zmm9[5],zmm8[7],zmm9[7]
4421 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm2, %zmm13, %zmm16
4422 ; AVX512DQ-NEXT: vbroadcasti32x4 {{.*#+}} zmm25 = [6,14,6,14,6,14,6,14]
4423 ; AVX512DQ-NEXT: # zmm25 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
4424 ; AVX512DQ-NEXT: vpermt2q %zmm26, %zmm25, %zmm5
4425 ; AVX512DQ-NEXT: vpermt2q %zmm17, %zmm25, %zmm15
4426 ; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm2 = ymm15[0,1,2,3],ymm5[4,5,6,7]
4427 ; AVX512DQ-NEXT: vmovdqa64 %zmm6, %zmm5
4428 ; AVX512DQ-NEXT: vpermt2q %zmm7, %zmm25, %zmm5
4429 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} zmm5 {%k1} = zmm8[0],zmm9[0],zmm8[2],zmm9[2],zmm8[4],zmm9[4],zmm8[6],zmm9[6]
4430 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm2, %zmm5, %zmm17
4431 ; AVX512DQ-NEXT: vmovdqa64 %zmm3, %zmm2
4432 ; AVX512DQ-NEXT: vpermt2q %zmm4, %zmm18, %zmm2
4433 ; AVX512DQ-NEXT: vpermi2q %zmm1, %zmm0, %zmm18
4434 ; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm18 {%k1}
4435 ; AVX512DQ-NEXT: vmovdqa 64(%rsi), %xmm2
4436 ; AVX512DQ-NEXT: vinserti128 $1, 64(%rcx), %ymm2, %ymm2
4437 ; AVX512DQ-NEXT: vmovdqa 64(%rdi), %xmm5
4438 ; AVX512DQ-NEXT: vinserti128 $1, 64(%rdx), %ymm5, %ymm5
4439 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} ymm13 = ymm5[1],ymm2[1],ymm5[3],ymm2[3]
4440 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm13, %zmm18, %zmm18
4441 ; AVX512DQ-NEXT: vmovdqa64 %zmm3, %zmm13
4442 ; AVX512DQ-NEXT: vpermt2q %zmm4, %zmm19, %zmm13
4443 ; AVX512DQ-NEXT: vpermi2q %zmm1, %zmm0, %zmm19
4444 ; AVX512DQ-NEXT: vmovdqa64 %zmm13, %zmm19 {%k1}
4445 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} ymm2 = ymm5[0],ymm2[0],ymm5[2],ymm2[2]
4446 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm2, %zmm19, %zmm19
4447 ; AVX512DQ-NEXT: vmovdqa64 %zmm14, %zmm2
4448 ; AVX512DQ-NEXT: vpermt2q %zmm28, %zmm20, %zmm2
4449 ; AVX512DQ-NEXT: vmovdqa64 %zmm11, %zmm5
4450 ; AVX512DQ-NEXT: vpermt2q %zmm27, %zmm20, %zmm5
4451 ; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm2 = ymm5[0,1,2,3],ymm2[4,5,6,7]
4452 ; AVX512DQ-NEXT: vpermi2q %zmm4, %zmm3, %zmm20
4453 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} zmm5 = zmm0[1],zmm1[1],zmm0[3],zmm1[3],zmm0[5],zmm1[5],zmm0[7],zmm1[7]
4454 ; AVX512DQ-NEXT: vmovdqa64 %zmm20, %zmm5 {%k1}
4455 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm2, %zmm5, %zmm20
4456 ; AVX512DQ-NEXT: vmovdqa64 %zmm14, %zmm2
4457 ; AVX512DQ-NEXT: vpermt2q %zmm28, %zmm23, %zmm2
4458 ; AVX512DQ-NEXT: vmovdqa64 %zmm11, %zmm5
4459 ; AVX512DQ-NEXT: vpermt2q %zmm27, %zmm23, %zmm5
4460 ; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm2 = ymm5[0,1,2,3],ymm2[4,5,6,7]
4461 ; AVX512DQ-NEXT: vpermi2q %zmm4, %zmm3, %zmm23
4462 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} zmm5 = zmm0[0],zmm1[0],zmm0[2],zmm1[2],zmm0[4],zmm1[4],zmm0[6],zmm1[6]
4463 ; AVX512DQ-NEXT: vmovdqa64 %zmm23, %zmm5 {%k1}
4464 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm2, %zmm5, %zmm2
4465 ; AVX512DQ-NEXT: vmovdqa64 %zmm14, %zmm5
4466 ; AVX512DQ-NEXT: vpermt2q %zmm28, %zmm24, %zmm5
4467 ; AVX512DQ-NEXT: vmovdqa64 %zmm11, %zmm13
4468 ; AVX512DQ-NEXT: vpermt2q %zmm27, %zmm24, %zmm13
4469 ; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm5 = ymm13[0,1,2,3],ymm5[4,5,6,7]
4470 ; AVX512DQ-NEXT: vpermi2q %zmm1, %zmm0, %zmm24
4471 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} zmm24 {%k1} = zmm3[1],zmm4[1],zmm3[3],zmm4[3],zmm3[5],zmm4[5],zmm3[7],zmm4[7]
4472 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm5, %zmm24, %zmm5
4473 ; AVX512DQ-NEXT: vpermt2q %zmm28, %zmm25, %zmm14
4474 ; AVX512DQ-NEXT: vpermt2q %zmm27, %zmm25, %zmm11
4475 ; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm11 = ymm11[0,1,2,3],ymm14[4,5,6,7]
4476 ; AVX512DQ-NEXT: vpermi2q %zmm1, %zmm0, %zmm25
4477 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} zmm25 {%k1} = zmm3[0],zmm4[0],zmm3[2],zmm4[2],zmm3[4],zmm4[4],zmm3[6],zmm4[6]
4478 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm11, %zmm25, %zmm11
4479 ; AVX512DQ-NEXT: vbroadcasti32x4 {{.*#+}} zmm13 = [3,11,3,11,3,11,3,11]
4480 ; AVX512DQ-NEXT: # zmm13 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
4481 ; AVX512DQ-NEXT: vmovdqa64 %zmm8, %zmm14
4482 ; AVX512DQ-NEXT: vpermt2q %zmm9, %zmm13, %zmm14
4483 ; AVX512DQ-NEXT: vmovdqa64 %zmm6, %zmm15
4484 ; AVX512DQ-NEXT: vpermt2q %zmm7, %zmm13, %zmm15
4485 ; AVX512DQ-NEXT: vmovdqa64 %zmm14, %zmm15 {%k1}
4486 ; AVX512DQ-NEXT: vmovdqa (%rcx), %ymm14
4487 ; AVX512DQ-NEXT: vmovdqa64 64(%rcx), %ymm23
4488 ; AVX512DQ-NEXT: vmovdqa64 (%rdx), %ymm24
4489 ; AVX512DQ-NEXT: vmovdqa64 64(%rdx), %ymm25
4490 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} ymm10 = ymm24[1],ymm14[1],ymm24[3],ymm14[3]
4491 ; AVX512DQ-NEXT: vmovdqa64 (%rsi), %ymm26
4492 ; AVX512DQ-NEXT: vmovdqa64 64(%rsi), %ymm27
4493 ; AVX512DQ-NEXT: vmovdqa64 (%rdi), %ymm28
4494 ; AVX512DQ-NEXT: vmovdqa64 64(%rdi), %ymm30
4495 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} ymm12 = ymm28[1],ymm26[1],ymm28[3],ymm26[3]
4496 ; AVX512DQ-NEXT: vperm2i128 {{.*#+}} ymm10 = ymm12[2,3],ymm10[2,3]
4497 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm10, %zmm15, %zmm10
4498 ; AVX512DQ-NEXT: vbroadcasti32x4 {{.*#+}} zmm12 = [2,10,2,10,2,10,2,10]
4499 ; AVX512DQ-NEXT: # zmm12 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
4500 ; AVX512DQ-NEXT: vpermt2q %zmm9, %zmm12, %zmm8
4501 ; AVX512DQ-NEXT: vpermt2q %zmm7, %zmm12, %zmm6
4502 ; AVX512DQ-NEXT: vmovdqa64 %zmm8, %zmm6 {%k1}
4503 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm24[0],ymm14[0],ymm24[2],ymm14[2]
4504 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} ymm8 = ymm28[0],ymm26[0],ymm28[2],ymm26[2]
4505 ; AVX512DQ-NEXT: vperm2i128 {{.*#+}} ymm7 = ymm8[2,3],ymm7[2,3]
4506 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm7, %zmm6, %zmm6
4507 ; AVX512DQ-NEXT: vmovdqa64 %zmm3, %zmm7
4508 ; AVX512DQ-NEXT: vpermt2q %zmm4, %zmm13, %zmm7
4509 ; AVX512DQ-NEXT: vpermi2q %zmm1, %zmm0, %zmm13
4510 ; AVX512DQ-NEXT: vmovdqa64 %zmm7, %zmm13 {%k1}
4511 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} ymm7 = ymm25[1],ymm23[1],ymm25[3],ymm23[3]
4512 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} ymm8 = ymm30[1],ymm27[1],ymm30[3],ymm27[3]
4513 ; AVX512DQ-NEXT: vperm2i128 {{.*#+}} ymm7 = ymm8[2,3],ymm7[2,3]
4514 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm7, %zmm13, %zmm7
4515 ; AVX512DQ-NEXT: vpermt2q %zmm4, %zmm12, %zmm3
4516 ; AVX512DQ-NEXT: vpermt2q %zmm1, %zmm12, %zmm0
4517 ; AVX512DQ-NEXT: vmovdqa64 %zmm3, %zmm0 {%k1}
4518 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} ymm1 = ymm25[0],ymm23[0],ymm25[2],ymm23[2]
4519 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm30[0],ymm27[0],ymm30[2],ymm27[2]
4520 ; AVX512DQ-NEXT: vperm2i128 {{.*#+}} ymm1 = ymm3[2,3],ymm1[2,3]
4521 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm1, %zmm0, %zmm0
4522 ; AVX512DQ-NEXT: vmovdqa64 %zmm0, 640(%rax)
4523 ; AVX512DQ-NEXT: vmovdqa64 %zmm7, 704(%rax)
4524 ; AVX512DQ-NEXT: vmovdqa64 %zmm6, 128(%rax)
4525 ; AVX512DQ-NEXT: vmovdqa64 %zmm10, 192(%rax)
4526 ; AVX512DQ-NEXT: vmovdqa64 %zmm11, 896(%rax)
4527 ; AVX512DQ-NEXT: vmovdqa64 %zmm5, 960(%rax)
4528 ; AVX512DQ-NEXT: vmovdqa64 %zmm2, 768(%rax)
4529 ; AVX512DQ-NEXT: vmovdqa64 %zmm20, 832(%rax)
4530 ; AVX512DQ-NEXT: vmovdqa64 %zmm19, 512(%rax)
4531 ; AVX512DQ-NEXT: vmovdqa64 %zmm18, 576(%rax)
4532 ; AVX512DQ-NEXT: vmovdqa64 %zmm17, 384(%rax)
4533 ; AVX512DQ-NEXT: vmovdqa64 %zmm16, 448(%rax)
4534 ; AVX512DQ-NEXT: vmovdqa64 %zmm31, 256(%rax)
4535 ; AVX512DQ-NEXT: vmovdqa64 %zmm29, 320(%rax)
4536 ; AVX512DQ-NEXT: vmovdqa64 %zmm22, (%rax)
4537 ; AVX512DQ-NEXT: vmovdqa64 %zmm21, 64(%rax)
4538 ; AVX512DQ-NEXT: vzeroupper
4539 ; AVX512DQ-NEXT: retq
4540 ;
4541 ; AVX512DQ-FCP-LABEL: store_i64_stride8_vf16:
4542 ; AVX512DQ-FCP: # %bb.0:
4543 ; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
4544 ; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10
4545 ; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r11
4546 ; AVX512DQ-FCP-NEXT: vmovdqa64 64(%rdi), %zmm11
4547 ; AVX512DQ-FCP-NEXT: vmovdqa64 (%rdi), %zmm15
4548 ; AVX512DQ-FCP-NEXT: vmovdqa64 64(%rsi), %zmm27
4549 ; AVX512DQ-FCP-NEXT: vmovdqa64 (%rsi), %zmm17
4550 ; AVX512DQ-FCP-NEXT: vmovdqa64 64(%rdx), %zmm14
4551 ; AVX512DQ-FCP-NEXT: vmovdqa64 (%rdx), %zmm5
4552 ; AVX512DQ-FCP-NEXT: vmovdqa64 64(%rcx), %zmm28
4553 ; AVX512DQ-FCP-NEXT: vmovdqa64 (%rcx), %zmm26
4554 ; AVX512DQ-FCP-NEXT: vmovdqa64 (%r8), %zmm6
4555 ; AVX512DQ-FCP-NEXT: vmovdqa64 64(%r8), %zmm0
4556 ; AVX512DQ-FCP-NEXT: vmovdqa64 (%r9), %zmm7
4557 ; AVX512DQ-FCP-NEXT: vmovdqa64 64(%r9), %zmm1
4558 ; AVX512DQ-FCP-NEXT: vmovdqa64 (%r11), %zmm8
4559 ; AVX512DQ-FCP-NEXT: vmovdqa64 64(%r11), %zmm3
4560 ; AVX512DQ-FCP-NEXT: vmovdqa64 (%r10), %zmm9
4561 ; AVX512DQ-FCP-NEXT: vmovdqa64 64(%r10), %zmm4
4562 ; AVX512DQ-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm18 = [1,9,1,9,1,9,1,9]
4563 ; AVX512DQ-FCP-NEXT: # zmm18 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
4564 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm8, %zmm10
4565 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm9, %zmm18, %zmm10
4566 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, %zmm12
4567 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm7, %zmm18, %zmm12
4568 ; AVX512DQ-FCP-NEXT: movb $-64, %r8b
4569 ; AVX512DQ-FCP-NEXT: kmovw %r8d, %k1
4570 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm10, %zmm12 {%k1}
4571 ; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %xmm10
4572 ; AVX512DQ-FCP-NEXT: vinserti128 $1, (%rcx), %ymm10, %ymm10
4573 ; AVX512DQ-FCP-NEXT: vmovdqa64 (%rdi), %xmm16
4574 ; AVX512DQ-FCP-NEXT: vinserti32x4 $1, (%rdx), %ymm16, %ymm16
4575 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm19 = ymm16[1],ymm10[1],ymm16[3],ymm10[3]
4576 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm19, %zmm12, %zmm21
4577 ; AVX512DQ-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm19 = [0,8,0,8,0,8,0,8]
4578 ; AVX512DQ-FCP-NEXT: # zmm19 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
4579 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm8, %zmm12
4580 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm9, %zmm19, %zmm12
4581 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, %zmm20
4582 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm7, %zmm19, %zmm20
4583 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm12, %zmm20 {%k1}
4584 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm10 = ymm16[0],ymm10[0],ymm16[2],ymm10[2]
4585 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm10, %zmm20, %zmm22
4586 ; AVX512DQ-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm20 = [5,13,5,13,5,13,5,13]
4587 ; AVX512DQ-FCP-NEXT: # zmm20 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
4588 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm5, %zmm10
4589 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm26, %zmm20, %zmm10
4590 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm15, %zmm12
4591 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm17, %zmm20, %zmm12
4592 ; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm10 = ymm12[0,1,2,3],ymm10[4,5,6,7]
4593 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm8, %zmm12
4594 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm9, %zmm20, %zmm12
4595 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm16 = zmm6[1],zmm7[1],zmm6[3],zmm7[3],zmm6[5],zmm7[5],zmm6[7],zmm7[7]
4596 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm12, %zmm16 {%k1}
4597 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm10, %zmm16, %zmm29
4598 ; AVX512DQ-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm23 = [4,12,4,12,4,12,4,12]
4599 ; AVX512DQ-FCP-NEXT: # zmm23 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
4600 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm5, %zmm12
4601 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm26, %zmm23, %zmm12
4602 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm15, %zmm2
4603 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm17, %zmm23, %zmm2
4604 ; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm12[4,5,6,7]
4605 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm8, %zmm12
4606 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm9, %zmm23, %zmm12
4607 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm16 = zmm6[0],zmm7[0],zmm6[2],zmm7[2],zmm6[4],zmm7[4],zmm6[6],zmm7[6]
4608 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm12, %zmm16 {%k1}
4609 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm16, %zmm31
4610 ; AVX512DQ-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm24 = [7,15,7,15,7,15,7,15]
4611 ; AVX512DQ-FCP-NEXT: # zmm24 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
4612 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm5, %zmm2
4613 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm26, %zmm24, %zmm2
4614 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm15, %zmm13
4615 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm17, %zmm24, %zmm13
4616 ; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm13[0,1,2,3],ymm2[4,5,6,7]
4617 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, %zmm13
4618 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm7, %zmm24, %zmm13
4619 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm13 {%k1} = zmm8[1],zmm9[1],zmm8[3],zmm9[3],zmm8[5],zmm9[5],zmm8[7],zmm9[7]
4620 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm13, %zmm16
4621 ; AVX512DQ-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm25 = [6,14,6,14,6,14,6,14]
4622 ; AVX512DQ-FCP-NEXT: # zmm25 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
4623 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm26, %zmm25, %zmm5
4624 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm17, %zmm25, %zmm15
4625 ; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm15[0,1,2,3],ymm5[4,5,6,7]
4626 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, %zmm5
4627 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm7, %zmm25, %zmm5
4628 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm5 {%k1} = zmm8[0],zmm9[0],zmm8[2],zmm9[2],zmm8[4],zmm9[4],zmm8[6],zmm9[6]
4629 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm5, %zmm17
4630 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, %zmm2
4631 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm4, %zmm18, %zmm2
4632 ; AVX512DQ-FCP-NEXT: vpermi2q %zmm1, %zmm0, %zmm18
4633 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm18 {%k1}
4634 ; AVX512DQ-FCP-NEXT: vmovdqa 64(%rsi), %xmm2
4635 ; AVX512DQ-FCP-NEXT: vinserti128 $1, 64(%rcx), %ymm2, %ymm2
4636 ; AVX512DQ-FCP-NEXT: vmovdqa 64(%rdi), %xmm5
4637 ; AVX512DQ-FCP-NEXT: vinserti128 $1, 64(%rdx), %ymm5, %ymm5
4638 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm13 = ymm5[1],ymm2[1],ymm5[3],ymm2[3]
4639 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm13, %zmm18, %zmm18
4640 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, %zmm13
4641 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm4, %zmm19, %zmm13
4642 ; AVX512DQ-FCP-NEXT: vpermi2q %zmm1, %zmm0, %zmm19
4643 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm13, %zmm19 {%k1}
4644 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm2 = ymm5[0],ymm2[0],ymm5[2],ymm2[2]
4645 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm19, %zmm19
4646 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm14, %zmm2
4647 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm28, %zmm20, %zmm2
4648 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm11, %zmm5
4649 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm27, %zmm20, %zmm5
4650 ; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm5[0,1,2,3],ymm2[4,5,6,7]
4651 ; AVX512DQ-FCP-NEXT: vpermi2q %zmm4, %zmm3, %zmm20
4652 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm5 = zmm0[1],zmm1[1],zmm0[3],zmm1[3],zmm0[5],zmm1[5],zmm0[7],zmm1[7]
4653 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm20, %zmm5 {%k1}
4654 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm5, %zmm20
4655 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm14, %zmm2
4656 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm28, %zmm23, %zmm2
4657 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm11, %zmm5
4658 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm27, %zmm23, %zmm5
4659 ; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm5[0,1,2,3],ymm2[4,5,6,7]
4660 ; AVX512DQ-FCP-NEXT: vpermi2q %zmm4, %zmm3, %zmm23
4661 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm5 = zmm0[0],zmm1[0],zmm0[2],zmm1[2],zmm0[4],zmm1[4],zmm0[6],zmm1[6]
4662 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm23, %zmm5 {%k1}
4663 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm5, %zmm2
4664 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm14, %zmm5
4665 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm28, %zmm24, %zmm5
4666 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm11, %zmm13
4667 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm27, %zmm24, %zmm13
4668 ; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm5 = ymm13[0,1,2,3],ymm5[4,5,6,7]
4669 ; AVX512DQ-FCP-NEXT: vpermi2q %zmm1, %zmm0, %zmm24
4670 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm24 {%k1} = zmm3[1],zmm4[1],zmm3[3],zmm4[3],zmm3[5],zmm4[5],zmm3[7],zmm4[7]
4671 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm5, %zmm24, %zmm5
4672 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm28, %zmm25, %zmm14
4673 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm27, %zmm25, %zmm11
4674 ; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm11 = ymm11[0,1,2,3],ymm14[4,5,6,7]
4675 ; AVX512DQ-FCP-NEXT: vpermi2q %zmm1, %zmm0, %zmm25
4676 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm25 {%k1} = zmm3[0],zmm4[0],zmm3[2],zmm4[2],zmm3[4],zmm4[4],zmm3[6],zmm4[6]
4677 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm11, %zmm25, %zmm11
4678 ; AVX512DQ-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm13 = [3,11,3,11,3,11,3,11]
4679 ; AVX512DQ-FCP-NEXT: # zmm13 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
4680 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm8, %zmm14
4681 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm9, %zmm13, %zmm14
4682 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, %zmm15
4683 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm7, %zmm13, %zmm15
4684 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm14, %zmm15 {%k1}
4685 ; AVX512DQ-FCP-NEXT: vmovdqa (%rcx), %ymm14
4686 ; AVX512DQ-FCP-NEXT: vmovdqa64 64(%rcx), %ymm23
4687 ; AVX512DQ-FCP-NEXT: vmovdqa64 (%rdx), %ymm24
4688 ; AVX512DQ-FCP-NEXT: vmovdqa64 64(%rdx), %ymm25
4689 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm10 = ymm24[1],ymm14[1],ymm24[3],ymm14[3]
4690 ; AVX512DQ-FCP-NEXT: vmovdqa64 (%rsi), %ymm26
4691 ; AVX512DQ-FCP-NEXT: vmovdqa64 64(%rsi), %ymm27
4692 ; AVX512DQ-FCP-NEXT: vmovdqa64 (%rdi), %ymm28
4693 ; AVX512DQ-FCP-NEXT: vmovdqa64 64(%rdi), %ymm30
4694 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm12 = ymm28[1],ymm26[1],ymm28[3],ymm26[3]
4695 ; AVX512DQ-FCP-NEXT: vperm2i128 {{.*#+}} ymm10 = ymm12[2,3],ymm10[2,3]
4696 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm10, %zmm15, %zmm10
4697 ; AVX512DQ-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm12 = [2,10,2,10,2,10,2,10]
4698 ; AVX512DQ-FCP-NEXT: # zmm12 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
4699 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm9, %zmm12, %zmm8
4700 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm7, %zmm12, %zmm6
4701 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm8, %zmm6 {%k1}
4702 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm24[0],ymm14[0],ymm24[2],ymm14[2]
4703 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm8 = ymm28[0],ymm26[0],ymm28[2],ymm26[2]
4704 ; AVX512DQ-FCP-NEXT: vperm2i128 {{.*#+}} ymm7 = ymm8[2,3],ymm7[2,3]
4705 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm7, %zmm6, %zmm6
4706 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, %zmm7
4707 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm4, %zmm13, %zmm7
4708 ; AVX512DQ-FCP-NEXT: vpermi2q %zmm1, %zmm0, %zmm13
4709 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm7, %zmm13 {%k1}
4710 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm7 = ymm25[1],ymm23[1],ymm25[3],ymm23[3]
4711 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm8 = ymm30[1],ymm27[1],ymm30[3],ymm27[3]
4712 ; AVX512DQ-FCP-NEXT: vperm2i128 {{.*#+}} ymm7 = ymm8[2,3],ymm7[2,3]
4713 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm7, %zmm13, %zmm7
4714 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm4, %zmm12, %zmm3
4715 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm12, %zmm0
4716 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, %zmm0 {%k1}
4717 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm1 = ymm25[0],ymm23[0],ymm25[2],ymm23[2]
4718 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm30[0],ymm27[0],ymm30[2],ymm27[2]
4719 ; AVX512DQ-FCP-NEXT: vperm2i128 {{.*#+}} ymm1 = ymm3[2,3],ymm1[2,3]
4720 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm1, %zmm0, %zmm0
4721 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, 640(%rax)
4722 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm7, 704(%rax)
4723 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, 128(%rax)
4724 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm10, 192(%rax)
4725 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm11, 896(%rax)
4726 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm5, 960(%rax)
4727 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, 768(%rax)
4728 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm20, 832(%rax)
4729 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm19, 512(%rax)
4730 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm18, 576(%rax)
4731 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm17, 384(%rax)
4732 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm16, 448(%rax)
4733 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm31, 256(%rax)
4734 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm29, 320(%rax)
4735 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm22, (%rax)
4736 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm21, 64(%rax)
4737 ; AVX512DQ-FCP-NEXT: vzeroupper
4738 ; AVX512DQ-FCP-NEXT: retq
4739 ;
4740 ; AVX512BW-LABEL: store_i64_stride8_vf16:
4741 ; AVX512BW: # %bb.0:
4742 ; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
4743 ; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %r10
4744 ; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %r11
4745 ; AVX512BW-NEXT: vmovdqa64 64(%rdi), %zmm11
4746 ; AVX512BW-NEXT: vmovdqa64 (%rdi), %zmm15
4747 ; AVX512BW-NEXT: vmovdqa64 64(%rsi), %zmm27
4748 ; AVX512BW-NEXT: vmovdqa64 (%rsi), %zmm17
4749 ; AVX512BW-NEXT: vmovdqa64 64(%rdx), %zmm14
4750 ; AVX512BW-NEXT: vmovdqa64 (%rdx), %zmm5
4751 ; AVX512BW-NEXT: vmovdqa64 64(%rcx), %zmm28
4752 ; AVX512BW-NEXT: vmovdqa64 (%rcx), %zmm26
4753 ; AVX512BW-NEXT: vmovdqa64 (%r8), %zmm6
4754 ; AVX512BW-NEXT: vmovdqa64 64(%r8), %zmm0
4755 ; AVX512BW-NEXT: vmovdqa64 (%r9), %zmm7
4756 ; AVX512BW-NEXT: vmovdqa64 64(%r9), %zmm1
4757 ; AVX512BW-NEXT: vmovdqa64 (%r11), %zmm8
4758 ; AVX512BW-NEXT: vmovdqa64 64(%r11), %zmm3
4759 ; AVX512BW-NEXT: vmovdqa64 (%r10), %zmm9
4760 ; AVX512BW-NEXT: vmovdqa64 64(%r10), %zmm4
4761 ; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm18 = [1,9,1,9,1,9,1,9]
4762 ; AVX512BW-NEXT: # zmm18 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
4763 ; AVX512BW-NEXT: vmovdqa64 %zmm8, %zmm10
4764 ; AVX512BW-NEXT: vpermt2q %zmm9, %zmm18, %zmm10
4765 ; AVX512BW-NEXT: vmovdqa64 %zmm6, %zmm12
4766 ; AVX512BW-NEXT: vpermt2q %zmm7, %zmm18, %zmm12
4767 ; AVX512BW-NEXT: movb $-64, %r8b
4768 ; AVX512BW-NEXT: kmovd %r8d, %k1
4769 ; AVX512BW-NEXT: vmovdqa64 %zmm10, %zmm12 {%k1}
4770 ; AVX512BW-NEXT: vmovdqa (%rsi), %xmm10
4771 ; AVX512BW-NEXT: vinserti128 $1, (%rcx), %ymm10, %ymm10
4772 ; AVX512BW-NEXT: vmovdqa64 (%rdi), %xmm16
4773 ; AVX512BW-NEXT: vinserti32x4 $1, (%rdx), %ymm16, %ymm16
4774 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} ymm19 = ymm16[1],ymm10[1],ymm16[3],ymm10[3]
4775 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm19, %zmm12, %zmm21
4776 ; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm19 = [0,8,0,8,0,8,0,8]
4777 ; AVX512BW-NEXT: # zmm19 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
4778 ; AVX512BW-NEXT: vmovdqa64 %zmm8, %zmm12
4779 ; AVX512BW-NEXT: vpermt2q %zmm9, %zmm19, %zmm12
4780 ; AVX512BW-NEXT: vmovdqa64 %zmm6, %zmm20
4781 ; AVX512BW-NEXT: vpermt2q %zmm7, %zmm19, %zmm20
4782 ; AVX512BW-NEXT: vmovdqa64 %zmm12, %zmm20 {%k1}
4783 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} ymm10 = ymm16[0],ymm10[0],ymm16[2],ymm10[2]
4784 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm10, %zmm20, %zmm22
4785 ; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm20 = [5,13,5,13,5,13,5,13]
4786 ; AVX512BW-NEXT: # zmm20 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
4787 ; AVX512BW-NEXT: vmovdqa64 %zmm5, %zmm10
4788 ; AVX512BW-NEXT: vpermt2q %zmm26, %zmm20, %zmm10
4789 ; AVX512BW-NEXT: vmovdqa64 %zmm15, %zmm12
4790 ; AVX512BW-NEXT: vpermt2q %zmm17, %zmm20, %zmm12
4791 ; AVX512BW-NEXT: vpblendd {{.*#+}} ymm10 = ymm12[0,1,2,3],ymm10[4,5,6,7]
4792 ; AVX512BW-NEXT: vmovdqa64 %zmm8, %zmm12
4793 ; AVX512BW-NEXT: vpermt2q %zmm9, %zmm20, %zmm12
4794 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} zmm16 = zmm6[1],zmm7[1],zmm6[3],zmm7[3],zmm6[5],zmm7[5],zmm6[7],zmm7[7]
4795 ; AVX512BW-NEXT: vmovdqa64 %zmm12, %zmm16 {%k1}
4796 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm10, %zmm16, %zmm29
4797 ; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm23 = [4,12,4,12,4,12,4,12]
4798 ; AVX512BW-NEXT: # zmm23 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
4799 ; AVX512BW-NEXT: vmovdqa64 %zmm5, %zmm12
4800 ; AVX512BW-NEXT: vpermt2q %zmm26, %zmm23, %zmm12
4801 ; AVX512BW-NEXT: vmovdqa64 %zmm15, %zmm2
4802 ; AVX512BW-NEXT: vpermt2q %zmm17, %zmm23, %zmm2
4803 ; AVX512BW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm12[4,5,6,7]
4804 ; AVX512BW-NEXT: vmovdqa64 %zmm8, %zmm12
4805 ; AVX512BW-NEXT: vpermt2q %zmm9, %zmm23, %zmm12
4806 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} zmm16 = zmm6[0],zmm7[0],zmm6[2],zmm7[2],zmm6[4],zmm7[4],zmm6[6],zmm7[6]
4807 ; AVX512BW-NEXT: vmovdqa64 %zmm12, %zmm16 {%k1}
4808 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm2, %zmm16, %zmm31
4809 ; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm24 = [7,15,7,15,7,15,7,15]
4810 ; AVX512BW-NEXT: # zmm24 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
4811 ; AVX512BW-NEXT: vmovdqa64 %zmm5, %zmm2
4812 ; AVX512BW-NEXT: vpermt2q %zmm26, %zmm24, %zmm2
4813 ; AVX512BW-NEXT: vmovdqa64 %zmm15, %zmm13
4814 ; AVX512BW-NEXT: vpermt2q %zmm17, %zmm24, %zmm13
4815 ; AVX512BW-NEXT: vpblendd {{.*#+}} ymm2 = ymm13[0,1,2,3],ymm2[4,5,6,7]
4816 ; AVX512BW-NEXT: vmovdqa64 %zmm6, %zmm13
4817 ; AVX512BW-NEXT: vpermt2q %zmm7, %zmm24, %zmm13
4818 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} zmm13 {%k1} = zmm8[1],zmm9[1],zmm8[3],zmm9[3],zmm8[5],zmm9[5],zmm8[7],zmm9[7]
4819 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm2, %zmm13, %zmm16
4820 ; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm25 = [6,14,6,14,6,14,6,14]
4821 ; AVX512BW-NEXT: # zmm25 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
4822 ; AVX512BW-NEXT: vpermt2q %zmm26, %zmm25, %zmm5
4823 ; AVX512BW-NEXT: vpermt2q %zmm17, %zmm25, %zmm15
4824 ; AVX512BW-NEXT: vpblendd {{.*#+}} ymm2 = ymm15[0,1,2,3],ymm5[4,5,6,7]
4825 ; AVX512BW-NEXT: vmovdqa64 %zmm6, %zmm5
4826 ; AVX512BW-NEXT: vpermt2q %zmm7, %zmm25, %zmm5
4827 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} zmm5 {%k1} = zmm8[0],zmm9[0],zmm8[2],zmm9[2],zmm8[4],zmm9[4],zmm8[6],zmm9[6]
4828 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm2, %zmm5, %zmm17
4829 ; AVX512BW-NEXT: vmovdqa64 %zmm3, %zmm2
4830 ; AVX512BW-NEXT: vpermt2q %zmm4, %zmm18, %zmm2
4831 ; AVX512BW-NEXT: vpermi2q %zmm1, %zmm0, %zmm18
4832 ; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm18 {%k1}
4833 ; AVX512BW-NEXT: vmovdqa 64(%rsi), %xmm2
4834 ; AVX512BW-NEXT: vinserti128 $1, 64(%rcx), %ymm2, %ymm2
4835 ; AVX512BW-NEXT: vmovdqa 64(%rdi), %xmm5
4836 ; AVX512BW-NEXT: vinserti128 $1, 64(%rdx), %ymm5, %ymm5
4837 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} ymm13 = ymm5[1],ymm2[1],ymm5[3],ymm2[3]
4838 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm13, %zmm18, %zmm18
4839 ; AVX512BW-NEXT: vmovdqa64 %zmm3, %zmm13
4840 ; AVX512BW-NEXT: vpermt2q %zmm4, %zmm19, %zmm13
4841 ; AVX512BW-NEXT: vpermi2q %zmm1, %zmm0, %zmm19
4842 ; AVX512BW-NEXT: vmovdqa64 %zmm13, %zmm19 {%k1}
4843 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} ymm2 = ymm5[0],ymm2[0],ymm5[2],ymm2[2]
4844 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm2, %zmm19, %zmm19
4845 ; AVX512BW-NEXT: vmovdqa64 %zmm14, %zmm2
4846 ; AVX512BW-NEXT: vpermt2q %zmm28, %zmm20, %zmm2
4847 ; AVX512BW-NEXT: vmovdqa64 %zmm11, %zmm5
4848 ; AVX512BW-NEXT: vpermt2q %zmm27, %zmm20, %zmm5
4849 ; AVX512BW-NEXT: vpblendd {{.*#+}} ymm2 = ymm5[0,1,2,3],ymm2[4,5,6,7]
4850 ; AVX512BW-NEXT: vpermi2q %zmm4, %zmm3, %zmm20
4851 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} zmm5 = zmm0[1],zmm1[1],zmm0[3],zmm1[3],zmm0[5],zmm1[5],zmm0[7],zmm1[7]
4852 ; AVX512BW-NEXT: vmovdqa64 %zmm20, %zmm5 {%k1}
4853 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm2, %zmm5, %zmm20
4854 ; AVX512BW-NEXT: vmovdqa64 %zmm14, %zmm2
4855 ; AVX512BW-NEXT: vpermt2q %zmm28, %zmm23, %zmm2
4856 ; AVX512BW-NEXT: vmovdqa64 %zmm11, %zmm5
4857 ; AVX512BW-NEXT: vpermt2q %zmm27, %zmm23, %zmm5
4858 ; AVX512BW-NEXT: vpblendd {{.*#+}} ymm2 = ymm5[0,1,2,3],ymm2[4,5,6,7]
4859 ; AVX512BW-NEXT: vpermi2q %zmm4, %zmm3, %zmm23
4860 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} zmm5 = zmm0[0],zmm1[0],zmm0[2],zmm1[2],zmm0[4],zmm1[4],zmm0[6],zmm1[6]
4861 ; AVX512BW-NEXT: vmovdqa64 %zmm23, %zmm5 {%k1}
4862 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm2, %zmm5, %zmm2
4863 ; AVX512BW-NEXT: vmovdqa64 %zmm14, %zmm5
4864 ; AVX512BW-NEXT: vpermt2q %zmm28, %zmm24, %zmm5
4865 ; AVX512BW-NEXT: vmovdqa64 %zmm11, %zmm13
4866 ; AVX512BW-NEXT: vpermt2q %zmm27, %zmm24, %zmm13
4867 ; AVX512BW-NEXT: vpblendd {{.*#+}} ymm5 = ymm13[0,1,2,3],ymm5[4,5,6,7]
4868 ; AVX512BW-NEXT: vpermi2q %zmm1, %zmm0, %zmm24
4869 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} zmm24 {%k1} = zmm3[1],zmm4[1],zmm3[3],zmm4[3],zmm3[5],zmm4[5],zmm3[7],zmm4[7]
4870 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm5, %zmm24, %zmm5
4871 ; AVX512BW-NEXT: vpermt2q %zmm28, %zmm25, %zmm14
4872 ; AVX512BW-NEXT: vpermt2q %zmm27, %zmm25, %zmm11
4873 ; AVX512BW-NEXT: vpblendd {{.*#+}} ymm11 = ymm11[0,1,2,3],ymm14[4,5,6,7]
4874 ; AVX512BW-NEXT: vpermi2q %zmm1, %zmm0, %zmm25
4875 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} zmm25 {%k1} = zmm3[0],zmm4[0],zmm3[2],zmm4[2],zmm3[4],zmm4[4],zmm3[6],zmm4[6]
4876 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm11, %zmm25, %zmm11
4877 ; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm13 = [3,11,3,11,3,11,3,11]
4878 ; AVX512BW-NEXT: # zmm13 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
4879 ; AVX512BW-NEXT: vmovdqa64 %zmm8, %zmm14
4880 ; AVX512BW-NEXT: vpermt2q %zmm9, %zmm13, %zmm14
4881 ; AVX512BW-NEXT: vmovdqa64 %zmm6, %zmm15
4882 ; AVX512BW-NEXT: vpermt2q %zmm7, %zmm13, %zmm15
4883 ; AVX512BW-NEXT: vmovdqa64 %zmm14, %zmm15 {%k1}
4884 ; AVX512BW-NEXT: vmovdqa (%rcx), %ymm14
4885 ; AVX512BW-NEXT: vmovdqa64 64(%rcx), %ymm23
4886 ; AVX512BW-NEXT: vmovdqa64 (%rdx), %ymm24
4887 ; AVX512BW-NEXT: vmovdqa64 64(%rdx), %ymm25
4888 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} ymm10 = ymm24[1],ymm14[1],ymm24[3],ymm14[3]
4889 ; AVX512BW-NEXT: vmovdqa64 (%rsi), %ymm26
4890 ; AVX512BW-NEXT: vmovdqa64 64(%rsi), %ymm27
4891 ; AVX512BW-NEXT: vmovdqa64 (%rdi), %ymm28
4892 ; AVX512BW-NEXT: vmovdqa64 64(%rdi), %ymm30
4893 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} ymm12 = ymm28[1],ymm26[1],ymm28[3],ymm26[3]
4894 ; AVX512BW-NEXT: vperm2i128 {{.*#+}} ymm10 = ymm12[2,3],ymm10[2,3]
4895 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm10, %zmm15, %zmm10
4896 ; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm12 = [2,10,2,10,2,10,2,10]
4897 ; AVX512BW-NEXT: # zmm12 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
4898 ; AVX512BW-NEXT: vpermt2q %zmm9, %zmm12, %zmm8
4899 ; AVX512BW-NEXT: vpermt2q %zmm7, %zmm12, %zmm6
4900 ; AVX512BW-NEXT: vmovdqa64 %zmm8, %zmm6 {%k1}
4901 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm24[0],ymm14[0],ymm24[2],ymm14[2]
4902 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} ymm8 = ymm28[0],ymm26[0],ymm28[2],ymm26[2]
4903 ; AVX512BW-NEXT: vperm2i128 {{.*#+}} ymm7 = ymm8[2,3],ymm7[2,3]
4904 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm7, %zmm6, %zmm6
4905 ; AVX512BW-NEXT: vmovdqa64 %zmm3, %zmm7
4906 ; AVX512BW-NEXT: vpermt2q %zmm4, %zmm13, %zmm7
4907 ; AVX512BW-NEXT: vpermi2q %zmm1, %zmm0, %zmm13
4908 ; AVX512BW-NEXT: vmovdqa64 %zmm7, %zmm13 {%k1}
4909 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} ymm7 = ymm25[1],ymm23[1],ymm25[3],ymm23[3]
4910 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} ymm8 = ymm30[1],ymm27[1],ymm30[3],ymm27[3]
4911 ; AVX512BW-NEXT: vperm2i128 {{.*#+}} ymm7 = ymm8[2,3],ymm7[2,3]
4912 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm7, %zmm13, %zmm7
4913 ; AVX512BW-NEXT: vpermt2q %zmm4, %zmm12, %zmm3
4914 ; AVX512BW-NEXT: vpermt2q %zmm1, %zmm12, %zmm0
4915 ; AVX512BW-NEXT: vmovdqa64 %zmm3, %zmm0 {%k1}
4916 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} ymm1 = ymm25[0],ymm23[0],ymm25[2],ymm23[2]
4917 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm30[0],ymm27[0],ymm30[2],ymm27[2]
4918 ; AVX512BW-NEXT: vperm2i128 {{.*#+}} ymm1 = ymm3[2,3],ymm1[2,3]
4919 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm1, %zmm0, %zmm0
4920 ; AVX512BW-NEXT: vmovdqa64 %zmm0, 640(%rax)
4921 ; AVX512BW-NEXT: vmovdqa64 %zmm7, 704(%rax)
4922 ; AVX512BW-NEXT: vmovdqa64 %zmm6, 128(%rax)
4923 ; AVX512BW-NEXT: vmovdqa64 %zmm10, 192(%rax)
4924 ; AVX512BW-NEXT: vmovdqa64 %zmm11, 896(%rax)
4925 ; AVX512BW-NEXT: vmovdqa64 %zmm5, 960(%rax)
4926 ; AVX512BW-NEXT: vmovdqa64 %zmm2, 768(%rax)
4927 ; AVX512BW-NEXT: vmovdqa64 %zmm20, 832(%rax)
4928 ; AVX512BW-NEXT: vmovdqa64 %zmm19, 512(%rax)
4929 ; AVX512BW-NEXT: vmovdqa64 %zmm18, 576(%rax)
4930 ; AVX512BW-NEXT: vmovdqa64 %zmm17, 384(%rax)
4931 ; AVX512BW-NEXT: vmovdqa64 %zmm16, 448(%rax)
4932 ; AVX512BW-NEXT: vmovdqa64 %zmm31, 256(%rax)
4933 ; AVX512BW-NEXT: vmovdqa64 %zmm29, 320(%rax)
4934 ; AVX512BW-NEXT: vmovdqa64 %zmm22, (%rax)
4935 ; AVX512BW-NEXT: vmovdqa64 %zmm21, 64(%rax)
4936 ; AVX512BW-NEXT: vzeroupper
4937 ; AVX512BW-NEXT: retq
4938 ;
4939 ; AVX512BW-FCP-LABEL: store_i64_stride8_vf16:
4940 ; AVX512BW-FCP: # %bb.0:
4941 ; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
4942 ; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10
4943 ; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r11
4944 ; AVX512BW-FCP-NEXT: vmovdqa64 64(%rdi), %zmm11
4945 ; AVX512BW-FCP-NEXT: vmovdqa64 (%rdi), %zmm15
4946 ; AVX512BW-FCP-NEXT: vmovdqa64 64(%rsi), %zmm27
4947 ; AVX512BW-FCP-NEXT: vmovdqa64 (%rsi), %zmm17
4948 ; AVX512BW-FCP-NEXT: vmovdqa64 64(%rdx), %zmm14
4949 ; AVX512BW-FCP-NEXT: vmovdqa64 (%rdx), %zmm5
4950 ; AVX512BW-FCP-NEXT: vmovdqa64 64(%rcx), %zmm28
4951 ; AVX512BW-FCP-NEXT: vmovdqa64 (%rcx), %zmm26
4952 ; AVX512BW-FCP-NEXT: vmovdqa64 (%r8), %zmm6
4953 ; AVX512BW-FCP-NEXT: vmovdqa64 64(%r8), %zmm0
4954 ; AVX512BW-FCP-NEXT: vmovdqa64 (%r9), %zmm7
4955 ; AVX512BW-FCP-NEXT: vmovdqa64 64(%r9), %zmm1
4956 ; AVX512BW-FCP-NEXT: vmovdqa64 (%r11), %zmm8
4957 ; AVX512BW-FCP-NEXT: vmovdqa64 64(%r11), %zmm3
4958 ; AVX512BW-FCP-NEXT: vmovdqa64 (%r10), %zmm9
4959 ; AVX512BW-FCP-NEXT: vmovdqa64 64(%r10), %zmm4
4960 ; AVX512BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm18 = [1,9,1,9,1,9,1,9]
4961 ; AVX512BW-FCP-NEXT: # zmm18 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
4962 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm8, %zmm10
4963 ; AVX512BW-FCP-NEXT: vpermt2q %zmm9, %zmm18, %zmm10
4964 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm12
4965 ; AVX512BW-FCP-NEXT: vpermt2q %zmm7, %zmm18, %zmm12
4966 ; AVX512BW-FCP-NEXT: movb $-64, %r8b
4967 ; AVX512BW-FCP-NEXT: kmovd %r8d, %k1
4968 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm10, %zmm12 {%k1}
4969 ; AVX512BW-FCP-NEXT: vmovdqa (%rsi), %xmm10
4970 ; AVX512BW-FCP-NEXT: vinserti128 $1, (%rcx), %ymm10, %ymm10
4971 ; AVX512BW-FCP-NEXT: vmovdqa64 (%rdi), %xmm16
4972 ; AVX512BW-FCP-NEXT: vinserti32x4 $1, (%rdx), %ymm16, %ymm16
4973 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm19 = ymm16[1],ymm10[1],ymm16[3],ymm10[3]
4974 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm19, %zmm12, %zmm21
4975 ; AVX512BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm19 = [0,8,0,8,0,8,0,8]
4976 ; AVX512BW-FCP-NEXT: # zmm19 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
4977 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm8, %zmm12
4978 ; AVX512BW-FCP-NEXT: vpermt2q %zmm9, %zmm19, %zmm12
4979 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm20
4980 ; AVX512BW-FCP-NEXT: vpermt2q %zmm7, %zmm19, %zmm20
4981 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm12, %zmm20 {%k1}
4982 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm10 = ymm16[0],ymm10[0],ymm16[2],ymm10[2]
4983 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm10, %zmm20, %zmm22
4984 ; AVX512BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm20 = [5,13,5,13,5,13,5,13]
4985 ; AVX512BW-FCP-NEXT: # zmm20 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
4986 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm5, %zmm10
4987 ; AVX512BW-FCP-NEXT: vpermt2q %zmm26, %zmm20, %zmm10
4988 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm15, %zmm12
4989 ; AVX512BW-FCP-NEXT: vpermt2q %zmm17, %zmm20, %zmm12
4990 ; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm10 = ymm12[0,1,2,3],ymm10[4,5,6,7]
4991 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm8, %zmm12
4992 ; AVX512BW-FCP-NEXT: vpermt2q %zmm9, %zmm20, %zmm12
4993 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm16 = zmm6[1],zmm7[1],zmm6[3],zmm7[3],zmm6[5],zmm7[5],zmm6[7],zmm7[7]
4994 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm12, %zmm16 {%k1}
4995 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm10, %zmm16, %zmm29
4996 ; AVX512BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm23 = [4,12,4,12,4,12,4,12]
4997 ; AVX512BW-FCP-NEXT: # zmm23 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
4998 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm5, %zmm12
4999 ; AVX512BW-FCP-NEXT: vpermt2q %zmm26, %zmm23, %zmm12
5000 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm15, %zmm2
5001 ; AVX512BW-FCP-NEXT: vpermt2q %zmm17, %zmm23, %zmm2
5002 ; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm12[4,5,6,7]
5003 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm8, %zmm12
5004 ; AVX512BW-FCP-NEXT: vpermt2q %zmm9, %zmm23, %zmm12
5005 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm16 = zmm6[0],zmm7[0],zmm6[2],zmm7[2],zmm6[4],zmm7[4],zmm6[6],zmm7[6]
5006 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm12, %zmm16 {%k1}
5007 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm16, %zmm31
5008 ; AVX512BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm24 = [7,15,7,15,7,15,7,15]
5009 ; AVX512BW-FCP-NEXT: # zmm24 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
5010 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm5, %zmm2
5011 ; AVX512BW-FCP-NEXT: vpermt2q %zmm26, %zmm24, %zmm2
5012 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm15, %zmm13
5013 ; AVX512BW-FCP-NEXT: vpermt2q %zmm17, %zmm24, %zmm13
5014 ; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm13[0,1,2,3],ymm2[4,5,6,7]
5015 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm13
5016 ; AVX512BW-FCP-NEXT: vpermt2q %zmm7, %zmm24, %zmm13
5017 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm13 {%k1} = zmm8[1],zmm9[1],zmm8[3],zmm9[3],zmm8[5],zmm9[5],zmm8[7],zmm9[7]
5018 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm13, %zmm16
5019 ; AVX512BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm25 = [6,14,6,14,6,14,6,14]
5020 ; AVX512BW-FCP-NEXT: # zmm25 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
5021 ; AVX512BW-FCP-NEXT: vpermt2q %zmm26, %zmm25, %zmm5
5022 ; AVX512BW-FCP-NEXT: vpermt2q %zmm17, %zmm25, %zmm15
5023 ; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm15[0,1,2,3],ymm5[4,5,6,7]
5024 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm5
5025 ; AVX512BW-FCP-NEXT: vpermt2q %zmm7, %zmm25, %zmm5
5026 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm5 {%k1} = zmm8[0],zmm9[0],zmm8[2],zmm9[2],zmm8[4],zmm9[4],zmm8[6],zmm9[6]
5027 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm5, %zmm17
5028 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm2
5029 ; AVX512BW-FCP-NEXT: vpermt2q %zmm4, %zmm18, %zmm2
5030 ; AVX512BW-FCP-NEXT: vpermi2q %zmm1, %zmm0, %zmm18
5031 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm18 {%k1}
5032 ; AVX512BW-FCP-NEXT: vmovdqa 64(%rsi), %xmm2
5033 ; AVX512BW-FCP-NEXT: vinserti128 $1, 64(%rcx), %ymm2, %ymm2
5034 ; AVX512BW-FCP-NEXT: vmovdqa 64(%rdi), %xmm5
5035 ; AVX512BW-FCP-NEXT: vinserti128 $1, 64(%rdx), %ymm5, %ymm5
5036 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm13 = ymm5[1],ymm2[1],ymm5[3],ymm2[3]
5037 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm13, %zmm18, %zmm18
5038 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm13
5039 ; AVX512BW-FCP-NEXT: vpermt2q %zmm4, %zmm19, %zmm13
5040 ; AVX512BW-FCP-NEXT: vpermi2q %zmm1, %zmm0, %zmm19
5041 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm13, %zmm19 {%k1}
5042 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm2 = ymm5[0],ymm2[0],ymm5[2],ymm2[2]
5043 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm19, %zmm19
5044 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm14, %zmm2
5045 ; AVX512BW-FCP-NEXT: vpermt2q %zmm28, %zmm20, %zmm2
5046 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm11, %zmm5
5047 ; AVX512BW-FCP-NEXT: vpermt2q %zmm27, %zmm20, %zmm5
5048 ; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm5[0,1,2,3],ymm2[4,5,6,7]
5049 ; AVX512BW-FCP-NEXT: vpermi2q %zmm4, %zmm3, %zmm20
5050 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm5 = zmm0[1],zmm1[1],zmm0[3],zmm1[3],zmm0[5],zmm1[5],zmm0[7],zmm1[7]
5051 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm20, %zmm5 {%k1}
5052 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm5, %zmm20
5053 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm14, %zmm2
5054 ; AVX512BW-FCP-NEXT: vpermt2q %zmm28, %zmm23, %zmm2
5055 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm11, %zmm5
5056 ; AVX512BW-FCP-NEXT: vpermt2q %zmm27, %zmm23, %zmm5
5057 ; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm5[0,1,2,3],ymm2[4,5,6,7]
5058 ; AVX512BW-FCP-NEXT: vpermi2q %zmm4, %zmm3, %zmm23
5059 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm5 = zmm0[0],zmm1[0],zmm0[2],zmm1[2],zmm0[4],zmm1[4],zmm0[6],zmm1[6]
5060 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm23, %zmm5 {%k1}
5061 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm5, %zmm2
5062 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm14, %zmm5
5063 ; AVX512BW-FCP-NEXT: vpermt2q %zmm28, %zmm24, %zmm5
5064 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm11, %zmm13
5065 ; AVX512BW-FCP-NEXT: vpermt2q %zmm27, %zmm24, %zmm13
5066 ; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm5 = ymm13[0,1,2,3],ymm5[4,5,6,7]
5067 ; AVX512BW-FCP-NEXT: vpermi2q %zmm1, %zmm0, %zmm24
5068 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm24 {%k1} = zmm3[1],zmm4[1],zmm3[3],zmm4[3],zmm3[5],zmm4[5],zmm3[7],zmm4[7]
5069 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm5, %zmm24, %zmm5
5070 ; AVX512BW-FCP-NEXT: vpermt2q %zmm28, %zmm25, %zmm14
5071 ; AVX512BW-FCP-NEXT: vpermt2q %zmm27, %zmm25, %zmm11
5072 ; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm11 = ymm11[0,1,2,3],ymm14[4,5,6,7]
5073 ; AVX512BW-FCP-NEXT: vpermi2q %zmm1, %zmm0, %zmm25
5074 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm25 {%k1} = zmm3[0],zmm4[0],zmm3[2],zmm4[2],zmm3[4],zmm4[4],zmm3[6],zmm4[6]
5075 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm11, %zmm25, %zmm11
5076 ; AVX512BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm13 = [3,11,3,11,3,11,3,11]
5077 ; AVX512BW-FCP-NEXT: # zmm13 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
5078 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm8, %zmm14
5079 ; AVX512BW-FCP-NEXT: vpermt2q %zmm9, %zmm13, %zmm14
5080 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm15
5081 ; AVX512BW-FCP-NEXT: vpermt2q %zmm7, %zmm13, %zmm15
5082 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm14, %zmm15 {%k1}
5083 ; AVX512BW-FCP-NEXT: vmovdqa (%rcx), %ymm14
5084 ; AVX512BW-FCP-NEXT: vmovdqa64 64(%rcx), %ymm23
5085 ; AVX512BW-FCP-NEXT: vmovdqa64 (%rdx), %ymm24
5086 ; AVX512BW-FCP-NEXT: vmovdqa64 64(%rdx), %ymm25
5087 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm10 = ymm24[1],ymm14[1],ymm24[3],ymm14[3]
5088 ; AVX512BW-FCP-NEXT: vmovdqa64 (%rsi), %ymm26
5089 ; AVX512BW-FCP-NEXT: vmovdqa64 64(%rsi), %ymm27
5090 ; AVX512BW-FCP-NEXT: vmovdqa64 (%rdi), %ymm28
5091 ; AVX512BW-FCP-NEXT: vmovdqa64 64(%rdi), %ymm30
5092 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm12 = ymm28[1],ymm26[1],ymm28[3],ymm26[3]
5093 ; AVX512BW-FCP-NEXT: vperm2i128 {{.*#+}} ymm10 = ymm12[2,3],ymm10[2,3]
5094 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm10, %zmm15, %zmm10
5095 ; AVX512BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm12 = [2,10,2,10,2,10,2,10]
5096 ; AVX512BW-FCP-NEXT: # zmm12 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
5097 ; AVX512BW-FCP-NEXT: vpermt2q %zmm9, %zmm12, %zmm8
5098 ; AVX512BW-FCP-NEXT: vpermt2q %zmm7, %zmm12, %zmm6
5099 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm8, %zmm6 {%k1}
5100 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm24[0],ymm14[0],ymm24[2],ymm14[2]
5101 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm8 = ymm28[0],ymm26[0],ymm28[2],ymm26[2]
5102 ; AVX512BW-FCP-NEXT: vperm2i128 {{.*#+}} ymm7 = ymm8[2,3],ymm7[2,3]
5103 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm7, %zmm6, %zmm6
5104 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm7
5105 ; AVX512BW-FCP-NEXT: vpermt2q %zmm4, %zmm13, %zmm7
5106 ; AVX512BW-FCP-NEXT: vpermi2q %zmm1, %zmm0, %zmm13
5107 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm7, %zmm13 {%k1}
5108 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm7 = ymm25[1],ymm23[1],ymm25[3],ymm23[3]
5109 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm8 = ymm30[1],ymm27[1],ymm30[3],ymm27[3]
5110 ; AVX512BW-FCP-NEXT: vperm2i128 {{.*#+}} ymm7 = ymm8[2,3],ymm7[2,3]
5111 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm7, %zmm13, %zmm7
5112 ; AVX512BW-FCP-NEXT: vpermt2q %zmm4, %zmm12, %zmm3
5113 ; AVX512BW-FCP-NEXT: vpermt2q %zmm1, %zmm12, %zmm0
5114 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm0 {%k1}
5115 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm1 = ymm25[0],ymm23[0],ymm25[2],ymm23[2]
5116 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm30[0],ymm27[0],ymm30[2],ymm27[2]
5117 ; AVX512BW-FCP-NEXT: vperm2i128 {{.*#+}} ymm1 = ymm3[2,3],ymm1[2,3]
5118 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm1, %zmm0, %zmm0
5119 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, 640(%rax)
5120 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm7, 704(%rax)
5121 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm6, 128(%rax)
5122 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm10, 192(%rax)
5123 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm11, 896(%rax)
5124 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm5, 960(%rax)
5125 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, 768(%rax)
5126 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm20, 832(%rax)
5127 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm19, 512(%rax)
5128 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm18, 576(%rax)
5129 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm17, 384(%rax)
5130 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm16, 448(%rax)
5131 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm31, 256(%rax)
5132 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm29, 320(%rax)
5133 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm22, (%rax)
5134 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm21, 64(%rax)
5135 ; AVX512BW-FCP-NEXT: vzeroupper
5136 ; AVX512BW-FCP-NEXT: retq
5137 ;
5138 ; AVX512DQ-BW-LABEL: store_i64_stride8_vf16:
5139 ; AVX512DQ-BW: # %bb.0:
5140 ; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
5141 ; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %r10
5142 ; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %r11
5143 ; AVX512DQ-BW-NEXT: vmovdqa64 64(%rdi), %zmm11
5144 ; AVX512DQ-BW-NEXT: vmovdqa64 (%rdi), %zmm15
5145 ; AVX512DQ-BW-NEXT: vmovdqa64 64(%rsi), %zmm27
5146 ; AVX512DQ-BW-NEXT: vmovdqa64 (%rsi), %zmm17
5147 ; AVX512DQ-BW-NEXT: vmovdqa64 64(%rdx), %zmm14
5148 ; AVX512DQ-BW-NEXT: vmovdqa64 (%rdx), %zmm5
5149 ; AVX512DQ-BW-NEXT: vmovdqa64 64(%rcx), %zmm28
5150 ; AVX512DQ-BW-NEXT: vmovdqa64 (%rcx), %zmm26
5151 ; AVX512DQ-BW-NEXT: vmovdqa64 (%r8), %zmm6
5152 ; AVX512DQ-BW-NEXT: vmovdqa64 64(%r8), %zmm0
5153 ; AVX512DQ-BW-NEXT: vmovdqa64 (%r9), %zmm7
5154 ; AVX512DQ-BW-NEXT: vmovdqa64 64(%r9), %zmm1
5155 ; AVX512DQ-BW-NEXT: vmovdqa64 (%r11), %zmm8
5156 ; AVX512DQ-BW-NEXT: vmovdqa64 64(%r11), %zmm3
5157 ; AVX512DQ-BW-NEXT: vmovdqa64 (%r10), %zmm9
5158 ; AVX512DQ-BW-NEXT: vmovdqa64 64(%r10), %zmm4
5159 ; AVX512DQ-BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm18 = [1,9,1,9,1,9,1,9]
5160 ; AVX512DQ-BW-NEXT: # zmm18 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
5161 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm8, %zmm10
5162 ; AVX512DQ-BW-NEXT: vpermt2q %zmm9, %zmm18, %zmm10
5163 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm6, %zmm12
5164 ; AVX512DQ-BW-NEXT: vpermt2q %zmm7, %zmm18, %zmm12
5165 ; AVX512DQ-BW-NEXT: movb $-64, %r8b
5166 ; AVX512DQ-BW-NEXT: kmovd %r8d, %k1
5167 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm10, %zmm12 {%k1}
5168 ; AVX512DQ-BW-NEXT: vmovdqa (%rsi), %xmm10
5169 ; AVX512DQ-BW-NEXT: vinserti128 $1, (%rcx), %ymm10, %ymm10
5170 ; AVX512DQ-BW-NEXT: vmovdqa64 (%rdi), %xmm16
5171 ; AVX512DQ-BW-NEXT: vinserti32x4 $1, (%rdx), %ymm16, %ymm16
5172 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} ymm19 = ymm16[1],ymm10[1],ymm16[3],ymm10[3]
5173 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm19, %zmm12, %zmm21
5174 ; AVX512DQ-BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm19 = [0,8,0,8,0,8,0,8]
5175 ; AVX512DQ-BW-NEXT: # zmm19 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
5176 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm8, %zmm12
5177 ; AVX512DQ-BW-NEXT: vpermt2q %zmm9, %zmm19, %zmm12
5178 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm6, %zmm20
5179 ; AVX512DQ-BW-NEXT: vpermt2q %zmm7, %zmm19, %zmm20
5180 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm12, %zmm20 {%k1}
5181 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} ymm10 = ymm16[0],ymm10[0],ymm16[2],ymm10[2]
5182 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm10, %zmm20, %zmm22
5183 ; AVX512DQ-BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm20 = [5,13,5,13,5,13,5,13]
5184 ; AVX512DQ-BW-NEXT: # zmm20 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
5185 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm5, %zmm10
5186 ; AVX512DQ-BW-NEXT: vpermt2q %zmm26, %zmm20, %zmm10
5187 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm15, %zmm12
5188 ; AVX512DQ-BW-NEXT: vpermt2q %zmm17, %zmm20, %zmm12
5189 ; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm10 = ymm12[0,1,2,3],ymm10[4,5,6,7]
5190 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm8, %zmm12
5191 ; AVX512DQ-BW-NEXT: vpermt2q %zmm9, %zmm20, %zmm12
5192 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} zmm16 = zmm6[1],zmm7[1],zmm6[3],zmm7[3],zmm6[5],zmm7[5],zmm6[7],zmm7[7]
5193 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm12, %zmm16 {%k1}
5194 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm10, %zmm16, %zmm29
5195 ; AVX512DQ-BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm23 = [4,12,4,12,4,12,4,12]
5196 ; AVX512DQ-BW-NEXT: # zmm23 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
5197 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm5, %zmm12
5198 ; AVX512DQ-BW-NEXT: vpermt2q %zmm26, %zmm23, %zmm12
5199 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm15, %zmm2
5200 ; AVX512DQ-BW-NEXT: vpermt2q %zmm17, %zmm23, %zmm2
5201 ; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm12[4,5,6,7]
5202 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm8, %zmm12
5203 ; AVX512DQ-BW-NEXT: vpermt2q %zmm9, %zmm23, %zmm12
5204 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} zmm16 = zmm6[0],zmm7[0],zmm6[2],zmm7[2],zmm6[4],zmm7[4],zmm6[6],zmm7[6]
5205 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm12, %zmm16 {%k1}
5206 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm2, %zmm16, %zmm31
5207 ; AVX512DQ-BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm24 = [7,15,7,15,7,15,7,15]
5208 ; AVX512DQ-BW-NEXT: # zmm24 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
5209 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm5, %zmm2
5210 ; AVX512DQ-BW-NEXT: vpermt2q %zmm26, %zmm24, %zmm2
5211 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm15, %zmm13
5212 ; AVX512DQ-BW-NEXT: vpermt2q %zmm17, %zmm24, %zmm13
5213 ; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm2 = ymm13[0,1,2,3],ymm2[4,5,6,7]
5214 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm6, %zmm13
5215 ; AVX512DQ-BW-NEXT: vpermt2q %zmm7, %zmm24, %zmm13
5216 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} zmm13 {%k1} = zmm8[1],zmm9[1],zmm8[3],zmm9[3],zmm8[5],zmm9[5],zmm8[7],zmm9[7]
5217 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm2, %zmm13, %zmm16
5218 ; AVX512DQ-BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm25 = [6,14,6,14,6,14,6,14]
5219 ; AVX512DQ-BW-NEXT: # zmm25 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
5220 ; AVX512DQ-BW-NEXT: vpermt2q %zmm26, %zmm25, %zmm5
5221 ; AVX512DQ-BW-NEXT: vpermt2q %zmm17, %zmm25, %zmm15
5222 ; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm2 = ymm15[0,1,2,3],ymm5[4,5,6,7]
5223 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm6, %zmm5
5224 ; AVX512DQ-BW-NEXT: vpermt2q %zmm7, %zmm25, %zmm5
5225 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} zmm5 {%k1} = zmm8[0],zmm9[0],zmm8[2],zmm9[2],zmm8[4],zmm9[4],zmm8[6],zmm9[6]
5226 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm2, %zmm5, %zmm17
5227 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, %zmm2
5228 ; AVX512DQ-BW-NEXT: vpermt2q %zmm4, %zmm18, %zmm2
5229 ; AVX512DQ-BW-NEXT: vpermi2q %zmm1, %zmm0, %zmm18
5230 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm18 {%k1}
5231 ; AVX512DQ-BW-NEXT: vmovdqa 64(%rsi), %xmm2
5232 ; AVX512DQ-BW-NEXT: vinserti128 $1, 64(%rcx), %ymm2, %ymm2
5233 ; AVX512DQ-BW-NEXT: vmovdqa 64(%rdi), %xmm5
5234 ; AVX512DQ-BW-NEXT: vinserti128 $1, 64(%rdx), %ymm5, %ymm5
5235 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} ymm13 = ymm5[1],ymm2[1],ymm5[3],ymm2[3]
5236 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm13, %zmm18, %zmm18
5237 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, %zmm13
5238 ; AVX512DQ-BW-NEXT: vpermt2q %zmm4, %zmm19, %zmm13
5239 ; AVX512DQ-BW-NEXT: vpermi2q %zmm1, %zmm0, %zmm19
5240 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm13, %zmm19 {%k1}
5241 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} ymm2 = ymm5[0],ymm2[0],ymm5[2],ymm2[2]
5242 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm2, %zmm19, %zmm19
5243 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm14, %zmm2
5244 ; AVX512DQ-BW-NEXT: vpermt2q %zmm28, %zmm20, %zmm2
5245 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm11, %zmm5
5246 ; AVX512DQ-BW-NEXT: vpermt2q %zmm27, %zmm20, %zmm5
5247 ; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm2 = ymm5[0,1,2,3],ymm2[4,5,6,7]
5248 ; AVX512DQ-BW-NEXT: vpermi2q %zmm4, %zmm3, %zmm20
5249 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} zmm5 = zmm0[1],zmm1[1],zmm0[3],zmm1[3],zmm0[5],zmm1[5],zmm0[7],zmm1[7]
5250 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm20, %zmm5 {%k1}
5251 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm2, %zmm5, %zmm20
5252 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm14, %zmm2
5253 ; AVX512DQ-BW-NEXT: vpermt2q %zmm28, %zmm23, %zmm2
5254 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm11, %zmm5
5255 ; AVX512DQ-BW-NEXT: vpermt2q %zmm27, %zmm23, %zmm5
5256 ; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm2 = ymm5[0,1,2,3],ymm2[4,5,6,7]
5257 ; AVX512DQ-BW-NEXT: vpermi2q %zmm4, %zmm3, %zmm23
5258 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} zmm5 = zmm0[0],zmm1[0],zmm0[2],zmm1[2],zmm0[4],zmm1[4],zmm0[6],zmm1[6]
5259 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm23, %zmm5 {%k1}
5260 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm2, %zmm5, %zmm2
5261 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm14, %zmm5
5262 ; AVX512DQ-BW-NEXT: vpermt2q %zmm28, %zmm24, %zmm5
5263 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm11, %zmm13
5264 ; AVX512DQ-BW-NEXT: vpermt2q %zmm27, %zmm24, %zmm13
5265 ; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm5 = ymm13[0,1,2,3],ymm5[4,5,6,7]
5266 ; AVX512DQ-BW-NEXT: vpermi2q %zmm1, %zmm0, %zmm24
5267 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} zmm24 {%k1} = zmm3[1],zmm4[1],zmm3[3],zmm4[3],zmm3[5],zmm4[5],zmm3[7],zmm4[7]
5268 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm5, %zmm24, %zmm5
5269 ; AVX512DQ-BW-NEXT: vpermt2q %zmm28, %zmm25, %zmm14
5270 ; AVX512DQ-BW-NEXT: vpermt2q %zmm27, %zmm25, %zmm11
5271 ; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm11 = ymm11[0,1,2,3],ymm14[4,5,6,7]
5272 ; AVX512DQ-BW-NEXT: vpermi2q %zmm1, %zmm0, %zmm25
5273 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} zmm25 {%k1} = zmm3[0],zmm4[0],zmm3[2],zmm4[2],zmm3[4],zmm4[4],zmm3[6],zmm4[6]
5274 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm11, %zmm25, %zmm11
5275 ; AVX512DQ-BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm13 = [3,11,3,11,3,11,3,11]
5276 ; AVX512DQ-BW-NEXT: # zmm13 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
5277 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm8, %zmm14
5278 ; AVX512DQ-BW-NEXT: vpermt2q %zmm9, %zmm13, %zmm14
5279 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm6, %zmm15
5280 ; AVX512DQ-BW-NEXT: vpermt2q %zmm7, %zmm13, %zmm15
5281 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm14, %zmm15 {%k1}
5282 ; AVX512DQ-BW-NEXT: vmovdqa (%rcx), %ymm14
5283 ; AVX512DQ-BW-NEXT: vmovdqa64 64(%rcx), %ymm23
5284 ; AVX512DQ-BW-NEXT: vmovdqa64 (%rdx), %ymm24
5285 ; AVX512DQ-BW-NEXT: vmovdqa64 64(%rdx), %ymm25
5286 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} ymm10 = ymm24[1],ymm14[1],ymm24[3],ymm14[3]
5287 ; AVX512DQ-BW-NEXT: vmovdqa64 (%rsi), %ymm26
5288 ; AVX512DQ-BW-NEXT: vmovdqa64 64(%rsi), %ymm27
5289 ; AVX512DQ-BW-NEXT: vmovdqa64 (%rdi), %ymm28
5290 ; AVX512DQ-BW-NEXT: vmovdqa64 64(%rdi), %ymm30
5291 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} ymm12 = ymm28[1],ymm26[1],ymm28[3],ymm26[3]
5292 ; AVX512DQ-BW-NEXT: vperm2i128 {{.*#+}} ymm10 = ymm12[2,3],ymm10[2,3]
5293 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm10, %zmm15, %zmm10
5294 ; AVX512DQ-BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm12 = [2,10,2,10,2,10,2,10]
5295 ; AVX512DQ-BW-NEXT: # zmm12 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
5296 ; AVX512DQ-BW-NEXT: vpermt2q %zmm9, %zmm12, %zmm8
5297 ; AVX512DQ-BW-NEXT: vpermt2q %zmm7, %zmm12, %zmm6
5298 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm8, %zmm6 {%k1}
5299 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm24[0],ymm14[0],ymm24[2],ymm14[2]
5300 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} ymm8 = ymm28[0],ymm26[0],ymm28[2],ymm26[2]
5301 ; AVX512DQ-BW-NEXT: vperm2i128 {{.*#+}} ymm7 = ymm8[2,3],ymm7[2,3]
5302 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm7, %zmm6, %zmm6
5303 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, %zmm7
5304 ; AVX512DQ-BW-NEXT: vpermt2q %zmm4, %zmm13, %zmm7
5305 ; AVX512DQ-BW-NEXT: vpermi2q %zmm1, %zmm0, %zmm13
5306 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm7, %zmm13 {%k1}
5307 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} ymm7 = ymm25[1],ymm23[1],ymm25[3],ymm23[3]
5308 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} ymm8 = ymm30[1],ymm27[1],ymm30[3],ymm27[3]
5309 ; AVX512DQ-BW-NEXT: vperm2i128 {{.*#+}} ymm7 = ymm8[2,3],ymm7[2,3]
5310 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm7, %zmm13, %zmm7
5311 ; AVX512DQ-BW-NEXT: vpermt2q %zmm4, %zmm12, %zmm3
5312 ; AVX512DQ-BW-NEXT: vpermt2q %zmm1, %zmm12, %zmm0
5313 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, %zmm0 {%k1}
5314 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} ymm1 = ymm25[0],ymm23[0],ymm25[2],ymm23[2]
5315 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm30[0],ymm27[0],ymm30[2],ymm27[2]
5316 ; AVX512DQ-BW-NEXT: vperm2i128 {{.*#+}} ymm1 = ymm3[2,3],ymm1[2,3]
5317 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm1, %zmm0, %zmm0
5318 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, 640(%rax)
5319 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm7, 704(%rax)
5320 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm6, 128(%rax)
5321 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm10, 192(%rax)
5322 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm11, 896(%rax)
5323 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm5, 960(%rax)
5324 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, 768(%rax)
5325 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm20, 832(%rax)
5326 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm19, 512(%rax)
5327 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm18, 576(%rax)
5328 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm17, 384(%rax)
5329 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm16, 448(%rax)
5330 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm31, 256(%rax)
5331 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm29, 320(%rax)
5332 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm22, (%rax)
5333 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm21, 64(%rax)
5334 ; AVX512DQ-BW-NEXT: vzeroupper
5335 ; AVX512DQ-BW-NEXT: retq
5336 ;
5337 ; AVX512DQ-BW-FCP-LABEL: store_i64_stride8_vf16:
5338 ; AVX512DQ-BW-FCP: # %bb.0:
5339 ; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
5340 ; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10
5341 ; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r11
5342 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 64(%rdi), %zmm11
5343 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rdi), %zmm15
5344 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 64(%rsi), %zmm27
5345 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rsi), %zmm17
5346 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 64(%rdx), %zmm14
5347 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rdx), %zmm5
5348 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 64(%rcx), %zmm28
5349 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rcx), %zmm26
5350 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%r8), %zmm6
5351 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 64(%r8), %zmm0
5352 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%r9), %zmm7
5353 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 64(%r9), %zmm1
5354 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%r11), %zmm8
5355 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 64(%r11), %zmm3
5356 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%r10), %zmm9
5357 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 64(%r10), %zmm4
5358 ; AVX512DQ-BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm18 = [1,9,1,9,1,9,1,9]
5359 ; AVX512DQ-BW-FCP-NEXT: # zmm18 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
5360 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm8, %zmm10
5361 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm9, %zmm18, %zmm10
5362 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm12
5363 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm7, %zmm18, %zmm12
5364 ; AVX512DQ-BW-FCP-NEXT: movb $-64, %r8b
5365 ; AVX512DQ-BW-FCP-NEXT: kmovd %r8d, %k1
5366 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm10, %zmm12 {%k1}
5367 ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rsi), %xmm10
5368 ; AVX512DQ-BW-FCP-NEXT: vinserti128 $1, (%rcx), %ymm10, %ymm10
5369 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rdi), %xmm16
5370 ; AVX512DQ-BW-FCP-NEXT: vinserti32x4 $1, (%rdx), %ymm16, %ymm16
5371 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm19 = ymm16[1],ymm10[1],ymm16[3],ymm10[3]
5372 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm19, %zmm12, %zmm21
5373 ; AVX512DQ-BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm19 = [0,8,0,8,0,8,0,8]
5374 ; AVX512DQ-BW-FCP-NEXT: # zmm19 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
5375 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm8, %zmm12
5376 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm9, %zmm19, %zmm12
5377 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm20
5378 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm7, %zmm19, %zmm20
5379 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm12, %zmm20 {%k1}
5380 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm10 = ymm16[0],ymm10[0],ymm16[2],ymm10[2]
5381 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm10, %zmm20, %zmm22
5382 ; AVX512DQ-BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm20 = [5,13,5,13,5,13,5,13]
5383 ; AVX512DQ-BW-FCP-NEXT: # zmm20 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
5384 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm5, %zmm10
5385 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm26, %zmm20, %zmm10
5386 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm15, %zmm12
5387 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm17, %zmm20, %zmm12
5388 ; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm10 = ymm12[0,1,2,3],ymm10[4,5,6,7]
5389 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm8, %zmm12
5390 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm9, %zmm20, %zmm12
5391 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm16 = zmm6[1],zmm7[1],zmm6[3],zmm7[3],zmm6[5],zmm7[5],zmm6[7],zmm7[7]
5392 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm12, %zmm16 {%k1}
5393 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm10, %zmm16, %zmm29
5394 ; AVX512DQ-BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm23 = [4,12,4,12,4,12,4,12]
5395 ; AVX512DQ-BW-FCP-NEXT: # zmm23 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
5396 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm5, %zmm12
5397 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm26, %zmm23, %zmm12
5398 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm15, %zmm2
5399 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm17, %zmm23, %zmm2
5400 ; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm12[4,5,6,7]
5401 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm8, %zmm12
5402 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm9, %zmm23, %zmm12
5403 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm16 = zmm6[0],zmm7[0],zmm6[2],zmm7[2],zmm6[4],zmm7[4],zmm6[6],zmm7[6]
5404 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm12, %zmm16 {%k1}
5405 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm16, %zmm31
5406 ; AVX512DQ-BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm24 = [7,15,7,15,7,15,7,15]
5407 ; AVX512DQ-BW-FCP-NEXT: # zmm24 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
5408 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm5, %zmm2
5409 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm26, %zmm24, %zmm2
5410 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm15, %zmm13
5411 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm17, %zmm24, %zmm13
5412 ; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm13[0,1,2,3],ymm2[4,5,6,7]
5413 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm13
5414 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm7, %zmm24, %zmm13
5415 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm13 {%k1} = zmm8[1],zmm9[1],zmm8[3],zmm9[3],zmm8[5],zmm9[5],zmm8[7],zmm9[7]
5416 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm13, %zmm16
5417 ; AVX512DQ-BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm25 = [6,14,6,14,6,14,6,14]
5418 ; AVX512DQ-BW-FCP-NEXT: # zmm25 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
5419 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm26, %zmm25, %zmm5
5420 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm17, %zmm25, %zmm15
5421 ; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm15[0,1,2,3],ymm5[4,5,6,7]
5422 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm5
5423 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm7, %zmm25, %zmm5
5424 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm5 {%k1} = zmm8[0],zmm9[0],zmm8[2],zmm9[2],zmm8[4],zmm9[4],zmm8[6],zmm9[6]
5425 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm5, %zmm17
5426 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm2
5427 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm4, %zmm18, %zmm2
5428 ; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm1, %zmm0, %zmm18
5429 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm18 {%k1}
5430 ; AVX512DQ-BW-FCP-NEXT: vmovdqa 64(%rsi), %xmm2
5431 ; AVX512DQ-BW-FCP-NEXT: vinserti128 $1, 64(%rcx), %ymm2, %ymm2
5432 ; AVX512DQ-BW-FCP-NEXT: vmovdqa 64(%rdi), %xmm5
5433 ; AVX512DQ-BW-FCP-NEXT: vinserti128 $1, 64(%rdx), %ymm5, %ymm5
5434 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm13 = ymm5[1],ymm2[1],ymm5[3],ymm2[3]
5435 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm13, %zmm18, %zmm18
5436 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm13
5437 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm4, %zmm19, %zmm13
5438 ; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm1, %zmm0, %zmm19
5439 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm13, %zmm19 {%k1}
5440 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm2 = ymm5[0],ymm2[0],ymm5[2],ymm2[2]
5441 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm19, %zmm19
5442 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm14, %zmm2
5443 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm28, %zmm20, %zmm2
5444 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm11, %zmm5
5445 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm27, %zmm20, %zmm5
5446 ; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm5[0,1,2,3],ymm2[4,5,6,7]
5447 ; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm4, %zmm3, %zmm20
5448 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm5 = zmm0[1],zmm1[1],zmm0[3],zmm1[3],zmm0[5],zmm1[5],zmm0[7],zmm1[7]
5449 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm20, %zmm5 {%k1}
5450 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm5, %zmm20
5451 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm14, %zmm2
5452 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm28, %zmm23, %zmm2
5453 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm11, %zmm5
5454 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm27, %zmm23, %zmm5
5455 ; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm5[0,1,2,3],ymm2[4,5,6,7]
5456 ; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm4, %zmm3, %zmm23
5457 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm5 = zmm0[0],zmm1[0],zmm0[2],zmm1[2],zmm0[4],zmm1[4],zmm0[6],zmm1[6]
5458 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm23, %zmm5 {%k1}
5459 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm5, %zmm2
5460 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm14, %zmm5
5461 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm28, %zmm24, %zmm5
5462 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm11, %zmm13
5463 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm27, %zmm24, %zmm13
5464 ; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm5 = ymm13[0,1,2,3],ymm5[4,5,6,7]
5465 ; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm1, %zmm0, %zmm24
5466 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm24 {%k1} = zmm3[1],zmm4[1],zmm3[3],zmm4[3],zmm3[5],zmm4[5],zmm3[7],zmm4[7]
5467 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm5, %zmm24, %zmm5
5468 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm28, %zmm25, %zmm14
5469 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm27, %zmm25, %zmm11
5470 ; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm11 = ymm11[0,1,2,3],ymm14[4,5,6,7]
5471 ; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm1, %zmm0, %zmm25
5472 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm25 {%k1} = zmm3[0],zmm4[0],zmm3[2],zmm4[2],zmm3[4],zmm4[4],zmm3[6],zmm4[6]
5473 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm11, %zmm25, %zmm11
5474 ; AVX512DQ-BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm13 = [3,11,3,11,3,11,3,11]
5475 ; AVX512DQ-BW-FCP-NEXT: # zmm13 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
5476 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm8, %zmm14
5477 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm9, %zmm13, %zmm14
5478 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm15
5479 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm7, %zmm13, %zmm15
5480 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm14, %zmm15 {%k1}
5481 ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rcx), %ymm14
5482 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 64(%rcx), %ymm23
5483 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rdx), %ymm24
5484 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 64(%rdx), %ymm25
5485 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm10 = ymm24[1],ymm14[1],ymm24[3],ymm14[3]
5486 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rsi), %ymm26
5487 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 64(%rsi), %ymm27
5488 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rdi), %ymm28
5489 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 64(%rdi), %ymm30
5490 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm12 = ymm28[1],ymm26[1],ymm28[3],ymm26[3]
5491 ; AVX512DQ-BW-FCP-NEXT: vperm2i128 {{.*#+}} ymm10 = ymm12[2,3],ymm10[2,3]
5492 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm10, %zmm15, %zmm10
5493 ; AVX512DQ-BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm12 = [2,10,2,10,2,10,2,10]
5494 ; AVX512DQ-BW-FCP-NEXT: # zmm12 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
5495 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm9, %zmm12, %zmm8
5496 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm7, %zmm12, %zmm6
5497 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm8, %zmm6 {%k1}
5498 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm24[0],ymm14[0],ymm24[2],ymm14[2]
5499 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm8 = ymm28[0],ymm26[0],ymm28[2],ymm26[2]
5500 ; AVX512DQ-BW-FCP-NEXT: vperm2i128 {{.*#+}} ymm7 = ymm8[2,3],ymm7[2,3]
5501 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm7, %zmm6, %zmm6
5502 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm7
5503 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm4, %zmm13, %zmm7
5504 ; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm1, %zmm0, %zmm13
5505 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm7, %zmm13 {%k1}
5506 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm7 = ymm25[1],ymm23[1],ymm25[3],ymm23[3]
5507 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm8 = ymm30[1],ymm27[1],ymm30[3],ymm27[3]
5508 ; AVX512DQ-BW-FCP-NEXT: vperm2i128 {{.*#+}} ymm7 = ymm8[2,3],ymm7[2,3]
5509 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm7, %zmm13, %zmm7
5510 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm4, %zmm12, %zmm3
5511 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm1, %zmm12, %zmm0
5512 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm0 {%k1}
5513 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm1 = ymm25[0],ymm23[0],ymm25[2],ymm23[2]
5514 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm30[0],ymm27[0],ymm30[2],ymm27[2]
5515 ; AVX512DQ-BW-FCP-NEXT: vperm2i128 {{.*#+}} ymm1 = ymm3[2,3],ymm1[2,3]
5516 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm1, %zmm0, %zmm0
5517 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, 640(%rax)
5518 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm7, 704(%rax)
5519 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm6, 128(%rax)
5520 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm10, 192(%rax)
5521 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm11, 896(%rax)
5522 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm5, 960(%rax)
5523 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, 768(%rax)
5524 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm20, 832(%rax)
5525 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm19, 512(%rax)
5526 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm18, 576(%rax)
5527 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm17, 384(%rax)
5528 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm16, 448(%rax)
5529 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm31, 256(%rax)
5530 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm29, 320(%rax)
5531 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm22, (%rax)
5532 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm21, 64(%rax)
5533 ; AVX512DQ-BW-FCP-NEXT: vzeroupper
5534 ; AVX512DQ-BW-FCP-NEXT: retq
5535 %in.vec0 = load <16 x i64>, ptr %in.vecptr0, align 64
5536 %in.vec1 = load <16 x i64>, ptr %in.vecptr1, align 64
5537 %in.vec2 = load <16 x i64>, ptr %in.vecptr2, align 64
5538 %in.vec3 = load <16 x i64>, ptr %in.vecptr3, align 64
5539 %in.vec4 = load <16 x i64>, ptr %in.vecptr4, align 64
5540 %in.vec5 = load <16 x i64>, ptr %in.vecptr5, align 64
5541 %in.vec6 = load <16 x i64>, ptr %in.vecptr6, align 64
5542 %in.vec7 = load <16 x i64>, ptr %in.vecptr7, align 64
5543 %1 = shufflevector <16 x i64> %in.vec0, <16 x i64> %in.vec1, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
5544 %2 = shufflevector <16 x i64> %in.vec2, <16 x i64> %in.vec3, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
5545 %3 = shufflevector <16 x i64> %in.vec4, <16 x i64> %in.vec5, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
5546 %4 = shufflevector <16 x i64> %in.vec6, <16 x i64> %in.vec7, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
5547 %5 = shufflevector <32 x i64> %1, <32 x i64> %2, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
5548 %6 = shufflevector <32 x i64> %3, <32 x i64> %4, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
5549 %7 = shufflevector <64 x i64> %5, <64 x i64> %6, <128 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 64, i32 65, i32 66, i32 67, i32 68, i32 69, i32 70, i32 71, i32 72, i32 73, i32 74, i32 75, i32 76, i32 77, i32 78, i32 79, i32 80, i32 81, i32 82, i32 83, i32 84, i32 85, i32 86, i32 87, i32 88, i32 89, i32 90, i32 91, i32 92, i32 93, i32 94, i32 95, i32 96, i32 97, i32 98, i32 99, i32 100, i32 101, i32 102, i32 103, i32 104, i32 105, i32 106, i32 107, i32 108, i32 109, i32 110, i32 111, i32 112, i32 113, i32 114, i32 115, i32 116, i32 117, i32 118, i32 119, i32 120, i32 121, i32 122, i32 123, i32 124, i32 125, i32 126, i32 127>
5550 %interleaved.vec = shufflevector <128 x i64> %7, <128 x i64> poison, <128 x i32> <i32 0, i32 16, i32 32, i32 48, i32 64, i32 80, i32 96, i32 112, i32 1, i32 17, i32 33, i32 49, i32 65, i32 81, i32 97, i32 113, i32 2, i32 18, i32 34, i32 50, i32 66, i32 82, i32 98, i32 114, i32 3, i32 19, i32 35, i32 51, i32 67, i32 83, i32 99, i32 115, i32 4, i32 20, i32 36, i32 52, i32 68, i32 84, i32 100, i32 116, i32 5, i32 21, i32 37, i32 53, i32 69, i32 85, i32 101, i32 117, i32 6, i32 22, i32 38, i32 54, i32 70, i32 86, i32 102, i32 118, i32 7, i32 23, i32 39, i32 55, i32 71, i32 87, i32 103, i32 119, i32 8, i32 24, i32 40, i32 56, i32 72, i32 88, i32 104, i32 120, i32 9, i32 25, i32 41, i32 57, i32 73, i32 89, i32 105, i32 121, i32 10, i32 26, i32 42, i32 58, i32 74, i32 90, i32 106, i32 122, i32 11, i32 27, i32 43, i32 59, i32 75, i32 91, i32 107, i32 123, i32 12, i32 28, i32 44, i32 60, i32 76, i32 92, i32 108, i32 124, i32 13, i32 29, i32 45, i32 61, i32 77, i32 93, i32 109, i32 125, i32 14, i32 30, i32 46, i32 62, i32 78, i32 94, i32 110, i32 126, i32 15, i32 31, i32 47, i32 63, i32 79, i32 95, i32 111, i32 127>
5551 store <128 x i64> %interleaved.vec, ptr %out.vec, align 64
5552 ret void
5553 }
5555 define void @store_i64_stride8_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecptr2, ptr %in.vecptr3, ptr %in.vecptr4, ptr %in.vecptr5, ptr %in.vecptr6, ptr %in.vecptr7, ptr %out.vec) nounwind {
5556 ; SSE-LABEL: store_i64_stride8_vf32:
5557 ; SSE: # %bb.0:
5558 ; SSE-NEXT: subq $1688, %rsp # imm = 0x698
5559 ; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax
5560 ; SSE-NEXT: movq {{[0-9]+}}(%rsp), %r10
5561 ; SSE-NEXT: movaps (%rdi), %xmm7
5562 ; SSE-NEXT: movaps 16(%rdi), %xmm8
5563 ; SSE-NEXT: movaps (%rsi), %xmm1
5564 ; SSE-NEXT: movaps 16(%rsi), %xmm0
5565 ; SSE-NEXT: movaps (%rdx), %xmm9
5566 ; SSE-NEXT: movaps 16(%rdx), %xmm10
5567 ; SSE-NEXT: movaps (%rcx), %xmm3
5568 ; SSE-NEXT: movaps 16(%rcx), %xmm2
5569 ; SSE-NEXT: movaps (%r8), %xmm11
5570 ; SSE-NEXT: movaps 16(%r8), %xmm12
5571 ; SSE-NEXT: movaps (%r9), %xmm5
5572 ; SSE-NEXT: movaps 16(%r9), %xmm4
5573 ; SSE-NEXT: movaps (%r10), %xmm13
5574 ; SSE-NEXT: movaps 16(%r10), %xmm14
5575 ; SSE-NEXT: movaps (%rax), %xmm6
5576 ; SSE-NEXT: movaps %xmm7, %xmm15
5577 ; SSE-NEXT: movlhps {{.*#+}} xmm15 = xmm15[0],xmm1[0]
5578 ; SSE-NEXT: movaps %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5579 ; SSE-NEXT: unpckhpd {{.*#+}} xmm7 = xmm7[1],xmm1[1]
5580 ; SSE-NEXT: movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5581 ; SSE-NEXT: movaps %xmm9, %xmm1
5582 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm3[0]
5583 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5584 ; SSE-NEXT: unpckhpd {{.*#+}} xmm9 = xmm9[1],xmm3[1]
5585 ; SSE-NEXT: movaps %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5586 ; SSE-NEXT: movaps %xmm11, %xmm1
5587 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm5[0]
5588 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5589 ; SSE-NEXT: unpckhpd {{.*#+}} xmm11 = xmm11[1],xmm5[1]
5590 ; SSE-NEXT: movaps %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5591 ; SSE-NEXT: movaps %xmm13, %xmm1
5592 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm6[0]
5593 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5594 ; SSE-NEXT: unpckhpd {{.*#+}} xmm13 = xmm13[1],xmm6[1]
5595 ; SSE-NEXT: movaps %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5596 ; SSE-NEXT: movaps %xmm8, %xmm1
5597 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
5598 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5599 ; SSE-NEXT: unpckhpd {{.*#+}} xmm8 = xmm8[1],xmm0[1]
5600 ; SSE-NEXT: movaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5601 ; SSE-NEXT: movaps %xmm10, %xmm0
5602 ; SSE-NEXT: movlhps {{.*#+}} xmm0 = xmm0[0],xmm2[0]
5603 ; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5604 ; SSE-NEXT: unpckhpd {{.*#+}} xmm10 = xmm10[1],xmm2[1]
5605 ; SSE-NEXT: movaps %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5606 ; SSE-NEXT: movaps %xmm12, %xmm0
5607 ; SSE-NEXT: movlhps {{.*#+}} xmm0 = xmm0[0],xmm4[0]
5608 ; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5609 ; SSE-NEXT: unpckhpd {{.*#+}} xmm12 = xmm12[1],xmm4[1]
5610 ; SSE-NEXT: movaps %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5611 ; SSE-NEXT: movaps 16(%rax), %xmm0
5612 ; SSE-NEXT: movaps %xmm14, %xmm1
5613 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
5614 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5615 ; SSE-NEXT: unpckhpd {{.*#+}} xmm14 = xmm14[1],xmm0[1]
5616 ; SSE-NEXT: movaps %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5617 ; SSE-NEXT: movaps 32(%rdi), %xmm2
5618 ; SSE-NEXT: movaps 32(%rsi), %xmm0
5619 ; SSE-NEXT: movaps %xmm2, %xmm1
5620 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
5621 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5622 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
5623 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5624 ; SSE-NEXT: movaps 32(%rdx), %xmm2
5625 ; SSE-NEXT: movaps 32(%rcx), %xmm0
5626 ; SSE-NEXT: movaps %xmm2, %xmm1
5627 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
5628 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5629 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
5630 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5631 ; SSE-NEXT: movaps 32(%r8), %xmm2
5632 ; SSE-NEXT: movaps 32(%r9), %xmm0
5633 ; SSE-NEXT: movaps %xmm2, %xmm1
5634 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
5635 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5636 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
5637 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5638 ; SSE-NEXT: movaps 32(%r10), %xmm2
5639 ; SSE-NEXT: movaps 32(%rax), %xmm0
5640 ; SSE-NEXT: movaps %xmm2, %xmm1
5641 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
5642 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5643 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
5644 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5645 ; SSE-NEXT: movaps 48(%rdi), %xmm2
5646 ; SSE-NEXT: movaps 48(%rsi), %xmm0
5647 ; SSE-NEXT: movaps %xmm2, %xmm1
5648 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
5649 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5650 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
5651 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5652 ; SSE-NEXT: movaps 48(%rdx), %xmm2
5653 ; SSE-NEXT: movaps 48(%rcx), %xmm0
5654 ; SSE-NEXT: movaps %xmm2, %xmm1
5655 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
5656 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5657 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
5658 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5659 ; SSE-NEXT: movaps 48(%r8), %xmm2
5660 ; SSE-NEXT: movaps 48(%r9), %xmm0
5661 ; SSE-NEXT: movaps %xmm2, %xmm1
5662 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
5663 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5664 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
5665 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5666 ; SSE-NEXT: movaps 48(%r10), %xmm2
5667 ; SSE-NEXT: movaps 48(%rax), %xmm0
5668 ; SSE-NEXT: movaps %xmm2, %xmm1
5669 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
5670 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5671 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
5672 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5673 ; SSE-NEXT: movaps 64(%rdi), %xmm2
5674 ; SSE-NEXT: movaps 64(%rsi), %xmm0
5675 ; SSE-NEXT: movaps %xmm2, %xmm1
5676 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
5677 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5678 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
5679 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5680 ; SSE-NEXT: movaps 64(%rdx), %xmm2
5681 ; SSE-NEXT: movaps 64(%rcx), %xmm0
5682 ; SSE-NEXT: movaps %xmm2, %xmm1
5683 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
5684 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5685 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
5686 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5687 ; SSE-NEXT: movaps 64(%r8), %xmm2
5688 ; SSE-NEXT: movaps 64(%r9), %xmm0
5689 ; SSE-NEXT: movaps %xmm2, %xmm1
5690 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
5691 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5692 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
5693 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5694 ; SSE-NEXT: movaps 64(%r10), %xmm2
5695 ; SSE-NEXT: movaps 64(%rax), %xmm0
5696 ; SSE-NEXT: movaps %xmm2, %xmm1
5697 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
5698 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5699 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
5700 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5701 ; SSE-NEXT: movaps 80(%rdi), %xmm2
5702 ; SSE-NEXT: movaps 80(%rsi), %xmm0
5703 ; SSE-NEXT: movaps %xmm2, %xmm1
5704 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
5705 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5706 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
5707 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5708 ; SSE-NEXT: movaps 80(%rdx), %xmm2
5709 ; SSE-NEXT: movaps 80(%rcx), %xmm0
5710 ; SSE-NEXT: movaps %xmm2, %xmm1
5711 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
5712 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5713 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
5714 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5715 ; SSE-NEXT: movaps 80(%r8), %xmm2
5716 ; SSE-NEXT: movaps 80(%r9), %xmm0
5717 ; SSE-NEXT: movaps %xmm2, %xmm1
5718 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
5719 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5720 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
5721 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5722 ; SSE-NEXT: movaps 80(%r10), %xmm2
5723 ; SSE-NEXT: movaps 80(%rax), %xmm0
5724 ; SSE-NEXT: movaps %xmm2, %xmm1
5725 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
5726 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5727 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
5728 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5729 ; SSE-NEXT: movaps 96(%rdi), %xmm2
5730 ; SSE-NEXT: movaps 96(%rsi), %xmm0
5731 ; SSE-NEXT: movaps %xmm2, %xmm1
5732 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
5733 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5734 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
5735 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5736 ; SSE-NEXT: movaps 96(%rdx), %xmm2
5737 ; SSE-NEXT: movaps 96(%rcx), %xmm0
5738 ; SSE-NEXT: movaps %xmm2, %xmm1
5739 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
5740 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5741 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
5742 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5743 ; SSE-NEXT: movaps 96(%r8), %xmm2
5744 ; SSE-NEXT: movaps 96(%r9), %xmm0
5745 ; SSE-NEXT: movaps %xmm2, %xmm1
5746 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
5747 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5748 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
5749 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5750 ; SSE-NEXT: movaps 96(%r10), %xmm2
5751 ; SSE-NEXT: movaps 96(%rax), %xmm0
5752 ; SSE-NEXT: movaps %xmm2, %xmm1
5753 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
5754 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5755 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
5756 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5757 ; SSE-NEXT: movaps 112(%rdi), %xmm2
5758 ; SSE-NEXT: movaps 112(%rsi), %xmm0
5759 ; SSE-NEXT: movaps %xmm2, %xmm1
5760 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
5761 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5762 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
5763 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5764 ; SSE-NEXT: movaps 112(%rdx), %xmm2
5765 ; SSE-NEXT: movaps 112(%rcx), %xmm0
5766 ; SSE-NEXT: movaps %xmm2, %xmm1
5767 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
5768 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5769 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
5770 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5771 ; SSE-NEXT: movaps 112(%r8), %xmm2
5772 ; SSE-NEXT: movaps 112(%r9), %xmm0
5773 ; SSE-NEXT: movaps %xmm2, %xmm1
5774 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
5775 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5776 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
5777 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5778 ; SSE-NEXT: movaps 112(%r10), %xmm2
5779 ; SSE-NEXT: movaps 112(%rax), %xmm0
5780 ; SSE-NEXT: movaps %xmm2, %xmm1
5781 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
5782 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5783 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
5784 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5785 ; SSE-NEXT: movaps 128(%rdi), %xmm2
5786 ; SSE-NEXT: movaps 128(%rsi), %xmm0
5787 ; SSE-NEXT: movaps %xmm2, %xmm1
5788 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
5789 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5790 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
5791 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5792 ; SSE-NEXT: movaps 128(%rdx), %xmm2
5793 ; SSE-NEXT: movaps 128(%rcx), %xmm0
5794 ; SSE-NEXT: movaps %xmm2, %xmm1
5795 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
5796 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5797 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
5798 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5799 ; SSE-NEXT: movaps 128(%r8), %xmm2
5800 ; SSE-NEXT: movaps 128(%r9), %xmm0
5801 ; SSE-NEXT: movaps %xmm2, %xmm1
5802 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
5803 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5804 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
5805 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5806 ; SSE-NEXT: movaps 128(%r10), %xmm2
5807 ; SSE-NEXT: movaps 128(%rax), %xmm0
5808 ; SSE-NEXT: movaps %xmm2, %xmm1
5809 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
5810 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5811 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
5812 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5813 ; SSE-NEXT: movaps 144(%rdi), %xmm2
5814 ; SSE-NEXT: movaps 144(%rsi), %xmm0
5815 ; SSE-NEXT: movaps %xmm2, %xmm1
5816 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
5817 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5818 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
5819 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5820 ; SSE-NEXT: movaps 144(%rdx), %xmm2
5821 ; SSE-NEXT: movaps 144(%rcx), %xmm0
5822 ; SSE-NEXT: movaps %xmm2, %xmm1
5823 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
5824 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5825 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
5826 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5827 ; SSE-NEXT: movaps 144(%r8), %xmm2
5828 ; SSE-NEXT: movaps 144(%r9), %xmm0
5829 ; SSE-NEXT: movaps %xmm2, %xmm1
5830 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
5831 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5832 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
5833 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5834 ; SSE-NEXT: movaps 144(%r10), %xmm2
5835 ; SSE-NEXT: movaps 144(%rax), %xmm0
5836 ; SSE-NEXT: movaps %xmm2, %xmm1
5837 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
5838 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5839 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
5840 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5841 ; SSE-NEXT: movaps 160(%rdi), %xmm2
5842 ; SSE-NEXT: movaps 160(%rsi), %xmm0
5843 ; SSE-NEXT: movaps %xmm2, %xmm1
5844 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
5845 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5846 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
5847 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5848 ; SSE-NEXT: movaps 160(%rdx), %xmm2
5849 ; SSE-NEXT: movaps 160(%rcx), %xmm0
5850 ; SSE-NEXT: movaps %xmm2, %xmm1
5851 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
5852 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5853 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
5854 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5855 ; SSE-NEXT: movaps 160(%r8), %xmm2
5856 ; SSE-NEXT: movaps 160(%r9), %xmm0
5857 ; SSE-NEXT: movaps %xmm2, %xmm1
5858 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
5859 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5860 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
5861 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5862 ; SSE-NEXT: movaps 160(%r10), %xmm2
5863 ; SSE-NEXT: movaps 160(%rax), %xmm0
5864 ; SSE-NEXT: movaps %xmm2, %xmm1
5865 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
5866 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5867 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
5868 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5869 ; SSE-NEXT: movaps 176(%rdi), %xmm2
5870 ; SSE-NEXT: movaps 176(%rsi), %xmm0
5871 ; SSE-NEXT: movaps %xmm2, %xmm1
5872 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
5873 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5874 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
5875 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5876 ; SSE-NEXT: movaps 176(%rdx), %xmm2
5877 ; SSE-NEXT: movaps 176(%rcx), %xmm0
5878 ; SSE-NEXT: movaps %xmm2, %xmm1
5879 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
5880 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5881 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
5882 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5883 ; SSE-NEXT: movaps 176(%r8), %xmm2
5884 ; SSE-NEXT: movaps 176(%r9), %xmm0
5885 ; SSE-NEXT: movaps %xmm2, %xmm1
5886 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
5887 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5888 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
5889 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5890 ; SSE-NEXT: movaps 176(%r10), %xmm2
5891 ; SSE-NEXT: movaps 176(%rax), %xmm0
5892 ; SSE-NEXT: movaps %xmm2, %xmm1
5893 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
5894 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5895 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
5896 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5897 ; SSE-NEXT: movaps 192(%rdi), %xmm2
5898 ; SSE-NEXT: movaps 192(%rsi), %xmm0
5899 ; SSE-NEXT: movaps %xmm2, %xmm1
5900 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
5901 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5902 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
5903 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5904 ; SSE-NEXT: movaps 192(%rdx), %xmm2
5905 ; SSE-NEXT: movaps 192(%rcx), %xmm0
5906 ; SSE-NEXT: movaps %xmm2, %xmm1
5907 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
5908 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5909 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
5910 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5911 ; SSE-NEXT: movaps 192(%r8), %xmm2
5912 ; SSE-NEXT: movaps 192(%r9), %xmm0
5913 ; SSE-NEXT: movaps %xmm2, %xmm1
5914 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
5915 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5916 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
5917 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5918 ; SSE-NEXT: movaps 192(%r10), %xmm2
5919 ; SSE-NEXT: movaps 192(%rax), %xmm0
5920 ; SSE-NEXT: movaps %xmm2, %xmm1
5921 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
5922 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5923 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
5924 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5925 ; SSE-NEXT: movaps 208(%rdi), %xmm2
5926 ; SSE-NEXT: movaps 208(%rsi), %xmm0
5927 ; SSE-NEXT: movaps %xmm2, %xmm1
5928 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
5929 ; SSE-NEXT: movaps %xmm1, (%rsp) # 16-byte Spill
5930 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
5931 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5932 ; SSE-NEXT: movaps 208(%rdx), %xmm2
5933 ; SSE-NEXT: movaps 208(%rcx), %xmm0
5934 ; SSE-NEXT: movaps %xmm2, %xmm1
5935 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
5936 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5937 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
5938 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5939 ; SSE-NEXT: movaps 208(%r8), %xmm2
5940 ; SSE-NEXT: movaps 208(%r9), %xmm0
5941 ; SSE-NEXT: movaps %xmm2, %xmm1
5942 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
5943 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5944 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
5945 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5946 ; SSE-NEXT: movaps 208(%r10), %xmm2
5947 ; SSE-NEXT: movaps 208(%rax), %xmm0
5948 ; SSE-NEXT: movaps %xmm2, %xmm1
5949 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
5950 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5951 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
5952 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5953 ; SSE-NEXT: movaps 224(%rdi), %xmm14
5954 ; SSE-NEXT: movaps 224(%rsi), %xmm0
5955 ; SSE-NEXT: movaps %xmm14, %xmm1
5956 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
5957 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
5958 ; SSE-NEXT: unpckhpd {{.*#+}} xmm14 = xmm14[1],xmm0[1]
5959 ; SSE-NEXT: movaps 224(%rdx), %xmm10
5960 ; SSE-NEXT: movaps 224(%rcx), %xmm0
5961 ; SSE-NEXT: movaps %xmm10, %xmm15
5962 ; SSE-NEXT: movlhps {{.*#+}} xmm15 = xmm15[0],xmm0[0]
5963 ; SSE-NEXT: unpckhpd {{.*#+}} xmm10 = xmm10[1],xmm0[1]
5964 ; SSE-NEXT: movaps 224(%r8), %xmm12
5965 ; SSE-NEXT: movaps 224(%r9), %xmm0
5966 ; SSE-NEXT: movaps %xmm12, %xmm13
5967 ; SSE-NEXT: movlhps {{.*#+}} xmm13 = xmm13[0],xmm0[0]
5968 ; SSE-NEXT: unpckhpd {{.*#+}} xmm12 = xmm12[1],xmm0[1]
5969 ; SSE-NEXT: movaps 224(%r10), %xmm8
5970 ; SSE-NEXT: movaps 224(%rax), %xmm0
5971 ; SSE-NEXT: movaps %xmm8, %xmm11
5972 ; SSE-NEXT: movlhps {{.*#+}} xmm11 = xmm11[0],xmm0[0]
5973 ; SSE-NEXT: unpckhpd {{.*#+}} xmm8 = xmm8[1],xmm0[1]
5974 ; SSE-NEXT: movaps 240(%rdi), %xmm5
5975 ; SSE-NEXT: movaps 240(%rsi), %xmm0
5976 ; SSE-NEXT: movaps %xmm5, %xmm9
5977 ; SSE-NEXT: movlhps {{.*#+}} xmm9 = xmm9[0],xmm0[0]
5978 ; SSE-NEXT: unpckhpd {{.*#+}} xmm5 = xmm5[1],xmm0[1]
5979 ; SSE-NEXT: movaps 240(%rdx), %xmm6
5980 ; SSE-NEXT: movaps 240(%rcx), %xmm1
5981 ; SSE-NEXT: movaps %xmm6, %xmm7
5982 ; SSE-NEXT: movlhps {{.*#+}} xmm7 = xmm7[0],xmm1[0]
5983 ; SSE-NEXT: unpckhpd {{.*#+}} xmm6 = xmm6[1],xmm1[1]
5984 ; SSE-NEXT: movaps 240(%r8), %xmm1
5985 ; SSE-NEXT: movaps 240(%r9), %xmm2
5986 ; SSE-NEXT: movaps %xmm1, %xmm4
5987 ; SSE-NEXT: movlhps {{.*#+}} xmm4 = xmm4[0],xmm2[0]
5988 ; SSE-NEXT: unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm2[1]
5989 ; SSE-NEXT: movaps 240(%r10), %xmm2
5990 ; SSE-NEXT: movaps 240(%rax), %xmm3
5991 ; SSE-NEXT: movaps %xmm2, %xmm0
5992 ; SSE-NEXT: movlhps {{.*#+}} xmm0 = xmm0[0],xmm3[0]
5993 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm3[1]
5994 ; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax
5995 ; SSE-NEXT: movaps %xmm2, 2032(%rax)
5996 ; SSE-NEXT: movaps %xmm1, 2016(%rax)
5997 ; SSE-NEXT: movaps %xmm6, 2000(%rax)
5998 ; SSE-NEXT: movaps %xmm5, 1984(%rax)
5999 ; SSE-NEXT: movaps %xmm0, 1968(%rax)
6000 ; SSE-NEXT: movaps %xmm4, 1952(%rax)
6001 ; SSE-NEXT: movaps %xmm7, 1936(%rax)
6002 ; SSE-NEXT: movaps %xmm9, 1920(%rax)
6003 ; SSE-NEXT: movaps %xmm8, 1904(%rax)
6004 ; SSE-NEXT: movaps %xmm12, 1888(%rax)
6005 ; SSE-NEXT: movaps %xmm10, 1872(%rax)
6006 ; SSE-NEXT: movaps %xmm14, 1856(%rax)
6007 ; SSE-NEXT: movaps %xmm11, 1840(%rax)
6008 ; SSE-NEXT: movaps %xmm13, 1824(%rax)
6009 ; SSE-NEXT: movaps %xmm15, 1808(%rax)
6010 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6011 ; SSE-NEXT: movaps %xmm0, 1792(%rax)
6012 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6013 ; SSE-NEXT: movaps %xmm0, 1776(%rax)
6014 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6015 ; SSE-NEXT: movaps %xmm0, 1760(%rax)
6016 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6017 ; SSE-NEXT: movaps %xmm0, 1744(%rax)
6018 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6019 ; SSE-NEXT: movaps %xmm0, 1728(%rax)
6020 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6021 ; SSE-NEXT: movaps %xmm0, 1712(%rax)
6022 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6023 ; SSE-NEXT: movaps %xmm0, 1696(%rax)
6024 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6025 ; SSE-NEXT: movaps %xmm0, 1680(%rax)
6026 ; SSE-NEXT: movaps (%rsp), %xmm0 # 16-byte Reload
6027 ; SSE-NEXT: movaps %xmm0, 1664(%rax)
6028 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6029 ; SSE-NEXT: movaps %xmm0, 1648(%rax)
6030 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6031 ; SSE-NEXT: movaps %xmm0, 1632(%rax)
6032 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6033 ; SSE-NEXT: movaps %xmm0, 1616(%rax)
6034 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6035 ; SSE-NEXT: movaps %xmm0, 1600(%rax)
6036 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6037 ; SSE-NEXT: movaps %xmm0, 1584(%rax)
6038 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6039 ; SSE-NEXT: movaps %xmm0, 1568(%rax)
6040 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6041 ; SSE-NEXT: movaps %xmm0, 1552(%rax)
6042 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6043 ; SSE-NEXT: movaps %xmm0, 1536(%rax)
6044 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6045 ; SSE-NEXT: movaps %xmm0, 1520(%rax)
6046 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6047 ; SSE-NEXT: movaps %xmm0, 1504(%rax)
6048 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6049 ; SSE-NEXT: movaps %xmm0, 1488(%rax)
6050 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6051 ; SSE-NEXT: movaps %xmm0, 1472(%rax)
6052 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6053 ; SSE-NEXT: movaps %xmm0, 1456(%rax)
6054 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6055 ; SSE-NEXT: movaps %xmm0, 1440(%rax)
6056 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6057 ; SSE-NEXT: movaps %xmm0, 1424(%rax)
6058 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6059 ; SSE-NEXT: movaps %xmm0, 1408(%rax)
6060 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6061 ; SSE-NEXT: movaps %xmm0, 1392(%rax)
6062 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6063 ; SSE-NEXT: movaps %xmm0, 1376(%rax)
6064 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6065 ; SSE-NEXT: movaps %xmm0, 1360(%rax)
6066 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6067 ; SSE-NEXT: movaps %xmm0, 1344(%rax)
6068 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6069 ; SSE-NEXT: movaps %xmm0, 1328(%rax)
6070 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6071 ; SSE-NEXT: movaps %xmm0, 1312(%rax)
6072 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6073 ; SSE-NEXT: movaps %xmm0, 1296(%rax)
6074 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6075 ; SSE-NEXT: movaps %xmm0, 1280(%rax)
6076 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6077 ; SSE-NEXT: movaps %xmm0, 1264(%rax)
6078 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6079 ; SSE-NEXT: movaps %xmm0, 1248(%rax)
6080 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6081 ; SSE-NEXT: movaps %xmm0, 1232(%rax)
6082 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6083 ; SSE-NEXT: movaps %xmm0, 1216(%rax)
6084 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6085 ; SSE-NEXT: movaps %xmm0, 1200(%rax)
6086 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6087 ; SSE-NEXT: movaps %xmm0, 1184(%rax)
6088 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6089 ; SSE-NEXT: movaps %xmm0, 1168(%rax)
6090 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6091 ; SSE-NEXT: movaps %xmm0, 1152(%rax)
6092 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6093 ; SSE-NEXT: movaps %xmm0, 1136(%rax)
6094 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6095 ; SSE-NEXT: movaps %xmm0, 1120(%rax)
6096 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6097 ; SSE-NEXT: movaps %xmm0, 1104(%rax)
6098 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6099 ; SSE-NEXT: movaps %xmm0, 1088(%rax)
6100 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6101 ; SSE-NEXT: movaps %xmm0, 1072(%rax)
6102 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6103 ; SSE-NEXT: movaps %xmm0, 1056(%rax)
6104 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6105 ; SSE-NEXT: movaps %xmm0, 1040(%rax)
6106 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6107 ; SSE-NEXT: movaps %xmm0, 1024(%rax)
6108 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6109 ; SSE-NEXT: movaps %xmm0, 1008(%rax)
6110 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6111 ; SSE-NEXT: movaps %xmm0, 992(%rax)
6112 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6113 ; SSE-NEXT: movaps %xmm0, 976(%rax)
6114 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6115 ; SSE-NEXT: movaps %xmm0, 960(%rax)
6116 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6117 ; SSE-NEXT: movaps %xmm0, 944(%rax)
6118 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6119 ; SSE-NEXT: movaps %xmm0, 928(%rax)
6120 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6121 ; SSE-NEXT: movaps %xmm0, 912(%rax)
6122 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6123 ; SSE-NEXT: movaps %xmm0, 896(%rax)
6124 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6125 ; SSE-NEXT: movaps %xmm0, 880(%rax)
6126 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6127 ; SSE-NEXT: movaps %xmm0, 864(%rax)
6128 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6129 ; SSE-NEXT: movaps %xmm0, 848(%rax)
6130 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6131 ; SSE-NEXT: movaps %xmm0, 832(%rax)
6132 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6133 ; SSE-NEXT: movaps %xmm0, 816(%rax)
6134 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6135 ; SSE-NEXT: movaps %xmm0, 800(%rax)
6136 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6137 ; SSE-NEXT: movaps %xmm0, 784(%rax)
6138 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6139 ; SSE-NEXT: movaps %xmm0, 768(%rax)
6140 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6141 ; SSE-NEXT: movaps %xmm0, 752(%rax)
6142 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6143 ; SSE-NEXT: movaps %xmm0, 736(%rax)
6144 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6145 ; SSE-NEXT: movaps %xmm0, 720(%rax)
6146 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6147 ; SSE-NEXT: movaps %xmm0, 704(%rax)
6148 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6149 ; SSE-NEXT: movaps %xmm0, 688(%rax)
6150 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6151 ; SSE-NEXT: movaps %xmm0, 672(%rax)
6152 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6153 ; SSE-NEXT: movaps %xmm0, 656(%rax)
6154 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6155 ; SSE-NEXT: movaps %xmm0, 640(%rax)
6156 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6157 ; SSE-NEXT: movaps %xmm0, 624(%rax)
6158 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6159 ; SSE-NEXT: movaps %xmm0, 608(%rax)
6160 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6161 ; SSE-NEXT: movaps %xmm0, 592(%rax)
6162 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6163 ; SSE-NEXT: movaps %xmm0, 576(%rax)
6164 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6165 ; SSE-NEXT: movaps %xmm0, 560(%rax)
6166 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6167 ; SSE-NEXT: movaps %xmm0, 544(%rax)
6168 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6169 ; SSE-NEXT: movaps %xmm0, 528(%rax)
6170 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6171 ; SSE-NEXT: movaps %xmm0, 512(%rax)
6172 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6173 ; SSE-NEXT: movaps %xmm0, 496(%rax)
6174 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6175 ; SSE-NEXT: movaps %xmm0, 480(%rax)
6176 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6177 ; SSE-NEXT: movaps %xmm0, 464(%rax)
6178 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6179 ; SSE-NEXT: movaps %xmm0, 448(%rax)
6180 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6181 ; SSE-NEXT: movaps %xmm0, 432(%rax)
6182 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6183 ; SSE-NEXT: movaps %xmm0, 416(%rax)
6184 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6185 ; SSE-NEXT: movaps %xmm0, 400(%rax)
6186 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6187 ; SSE-NEXT: movaps %xmm0, 384(%rax)
6188 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6189 ; SSE-NEXT: movaps %xmm0, 368(%rax)
6190 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6191 ; SSE-NEXT: movaps %xmm0, 352(%rax)
6192 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6193 ; SSE-NEXT: movaps %xmm0, 336(%rax)
6194 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6195 ; SSE-NEXT: movaps %xmm0, 320(%rax)
6196 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6197 ; SSE-NEXT: movaps %xmm0, 304(%rax)
6198 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6199 ; SSE-NEXT: movaps %xmm0, 288(%rax)
6200 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6201 ; SSE-NEXT: movaps %xmm0, 272(%rax)
6202 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6203 ; SSE-NEXT: movaps %xmm0, 256(%rax)
6204 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6205 ; SSE-NEXT: movaps %xmm0, 240(%rax)
6206 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6207 ; SSE-NEXT: movaps %xmm0, 224(%rax)
6208 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6209 ; SSE-NEXT: movaps %xmm0, 208(%rax)
6210 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6211 ; SSE-NEXT: movaps %xmm0, 192(%rax)
6212 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6213 ; SSE-NEXT: movaps %xmm0, 176(%rax)
6214 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6215 ; SSE-NEXT: movaps %xmm0, 160(%rax)
6216 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6217 ; SSE-NEXT: movaps %xmm0, 144(%rax)
6218 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6219 ; SSE-NEXT: movaps %xmm0, 128(%rax)
6220 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6221 ; SSE-NEXT: movaps %xmm0, 112(%rax)
6222 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6223 ; SSE-NEXT: movaps %xmm0, 96(%rax)
6224 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6225 ; SSE-NEXT: movaps %xmm0, 80(%rax)
6226 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6227 ; SSE-NEXT: movaps %xmm0, 64(%rax)
6228 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6229 ; SSE-NEXT: movaps %xmm0, 48(%rax)
6230 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6231 ; SSE-NEXT: movaps %xmm0, 32(%rax)
6232 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6233 ; SSE-NEXT: movaps %xmm0, 16(%rax)
6234 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
6235 ; SSE-NEXT: movaps %xmm0, (%rax)
6236 ; SSE-NEXT: addq $1688, %rsp # imm = 0x698
6237 ; SSE-NEXT: retq
6238 ;
6239 ; AVX-LABEL: store_i64_stride8_vf32:
6240 ; AVX: # %bb.0:
6241 ; AVX-NEXT: subq $1672, %rsp # imm = 0x688
6242 ; AVX-NEXT: movq {{[0-9]+}}(%rsp), %r10
6243 ; AVX-NEXT: movq {{[0-9]+}}(%rsp), %rax
6244 ; AVX-NEXT: vmovaps (%rsi), %xmm2
6245 ; AVX-NEXT: vmovaps (%rdi), %xmm3
6246 ; AVX-NEXT: vmovlhps {{.*#+}} xmm4 = xmm3[0],xmm2[0]
6247 ; AVX-NEXT: vmovaps (%rcx), %xmm5
6248 ; AVX-NEXT: vmovaps 32(%rcx), %xmm1
6249 ; AVX-NEXT: vmovaps 64(%rcx), %xmm0
6250 ; AVX-NEXT: vinsertf128 $1, %xmm5, %ymm4, %ymm6
6251 ; AVX-NEXT: vinsertf128 $1, (%rdx), %ymm4, %ymm4
6252 ; AVX-NEXT: vshufpd {{.*#+}} ymm4 = ymm4[0],ymm6[1],ymm4[2],ymm6[2]
6253 ; AVX-NEXT: vmovupd %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6254 ; AVX-NEXT: vmovaps (%r9), %xmm4
6255 ; AVX-NEXT: vmovaps (%r8), %xmm6
6256 ; AVX-NEXT: vmovlhps {{.*#+}} xmm7 = xmm6[0],xmm4[0]
6257 ; AVX-NEXT: vmovaps (%rax), %xmm8
6258 ; AVX-NEXT: vinsertf128 $1, %xmm8, %ymm7, %ymm9
6259 ; AVX-NEXT: vinsertf128 $1, (%r10), %ymm7, %ymm7
6260 ; AVX-NEXT: vshufpd {{.*#+}} ymm7 = ymm7[0],ymm9[1],ymm7[2],ymm9[2]
6261 ; AVX-NEXT: vmovupd %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6262 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm2 = xmm3[1],xmm2[1]
6263 ; AVX-NEXT: vinsertf128 $1, %xmm5, %ymm2, %ymm3
6264 ; AVX-NEXT: vbroadcastsd 8(%rdx), %ymm5
6265 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm5[4,5,6,7]
6266 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1],ymm3[2,3],ymm2[4,5],ymm3[6,7]
6267 ; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6268 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm2 = xmm6[1],xmm4[1]
6269 ; AVX-NEXT: vinsertf128 $1, %xmm8, %ymm2, %ymm3
6270 ; AVX-NEXT: vbroadcastsd 8(%r10), %ymm4
6271 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm4[4,5,6,7]
6272 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1],ymm3[2,3],ymm2[4,5],ymm3[6,7]
6273 ; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6274 ; AVX-NEXT: vmovaps 32(%rsi), %xmm3
6275 ; AVX-NEXT: vmovaps 32(%rdi), %xmm4
6276 ; AVX-NEXT: vmovlhps {{.*#+}} xmm2 = xmm4[0],xmm3[0]
6277 ; AVX-NEXT: vinsertf128 $1, %xmm1, %ymm2, %ymm5
6278 ; AVX-NEXT: vinsertf128 $1, 32(%rdx), %ymm2, %ymm2
6279 ; AVX-NEXT: vshufpd {{.*#+}} ymm2 = ymm2[0],ymm5[1],ymm2[2],ymm5[2]
6280 ; AVX-NEXT: vmovupd %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6281 ; AVX-NEXT: vmovaps 32(%r9), %xmm5
6282 ; AVX-NEXT: vmovaps 32(%r8), %xmm6
6283 ; AVX-NEXT: vmovlhps {{.*#+}} xmm7 = xmm6[0],xmm5[0]
6284 ; AVX-NEXT: vmovaps 32(%rax), %xmm8
6285 ; AVX-NEXT: vmovaps 64(%rax), %xmm2
6286 ; AVX-NEXT: vinsertf128 $1, %xmm8, %ymm7, %ymm9
6287 ; AVX-NEXT: vinsertf128 $1, 32(%r10), %ymm7, %ymm7
6288 ; AVX-NEXT: vshufpd {{.*#+}} ymm7 = ymm7[0],ymm9[1],ymm7[2],ymm9[2]
6289 ; AVX-NEXT: vmovupd %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6290 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm3 = xmm4[1],xmm3[1]
6291 ; AVX-NEXT: vinsertf128 $1, %xmm1, %ymm3, %ymm1
6292 ; AVX-NEXT: vbroadcastsd 40(%rdx), %ymm4
6293 ; AVX-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm4[4,5,6,7]
6294 ; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm3[0,1],ymm1[2,3],ymm3[4,5],ymm1[6,7]
6295 ; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6296 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm6[1],xmm5[1]
6297 ; AVX-NEXT: vinsertf128 $1, %xmm8, %ymm1, %ymm3
6298 ; AVX-NEXT: vbroadcastsd 40(%r10), %ymm4
6299 ; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm4[4,5,6,7]
6300 ; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm3[2,3],ymm1[4,5],ymm3[6,7]
6301 ; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6302 ; AVX-NEXT: vmovaps 64(%rsi), %xmm1
6303 ; AVX-NEXT: vmovaps 64(%rdi), %xmm3
6304 ; AVX-NEXT: vmovlhps {{.*#+}} xmm4 = xmm3[0],xmm1[0]
6305 ; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm4, %ymm5
6306 ; AVX-NEXT: vinsertf128 $1, 64(%rdx), %ymm4, %ymm4
6307 ; AVX-NEXT: vshufpd {{.*#+}} ymm4 = ymm4[0],ymm5[1],ymm4[2],ymm5[2]
6308 ; AVX-NEXT: vmovupd %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6309 ; AVX-NEXT: vmovaps 64(%r9), %xmm4
6310 ; AVX-NEXT: vmovaps 64(%r8), %xmm5
6311 ; AVX-NEXT: vmovlhps {{.*#+}} xmm6 = xmm5[0],xmm4[0]
6312 ; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm6, %ymm7
6313 ; AVX-NEXT: vinsertf128 $1, 64(%r10), %ymm6, %ymm6
6314 ; AVX-NEXT: vshufpd {{.*#+}} ymm6 = ymm6[0],ymm7[1],ymm6[2],ymm7[2]
6315 ; AVX-NEXT: vmovupd %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6316 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm3[1],xmm1[1]
6317 ; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm1, %ymm0
6318 ; AVX-NEXT: vbroadcastsd 72(%rdx), %ymm3
6319 ; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm3[4,5,6,7]
6320 ; AVX-NEXT: vmovaps 96(%rcx), %xmm3
6321 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3],ymm1[4,5],ymm0[6,7]
6322 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6323 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm5[1],xmm4[1]
6324 ; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm1
6325 ; AVX-NEXT: vbroadcastsd 72(%r10), %ymm2
6326 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
6327 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7]
6328 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6329 ; AVX-NEXT: vmovaps 96(%rsi), %xmm0
6330 ; AVX-NEXT: vmovaps 96(%rdi), %xmm1
6331 ; AVX-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
6332 ; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm2, %ymm4
6333 ; AVX-NEXT: vinsertf128 $1, 96(%rdx), %ymm2, %ymm2
6334 ; AVX-NEXT: vshufpd {{.*#+}} ymm2 = ymm2[0],ymm4[1],ymm2[2],ymm4[2]
6335 ; AVX-NEXT: vmovupd %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6336 ; AVX-NEXT: vmovaps 96(%r9), %xmm2
6337 ; AVX-NEXT: vmovaps 96(%r8), %xmm4
6338 ; AVX-NEXT: vmovlhps {{.*#+}} xmm5 = xmm4[0],xmm2[0]
6339 ; AVX-NEXT: vmovaps 96(%rax), %xmm6
6340 ; AVX-NEXT: vinsertf128 $1, %xmm6, %ymm5, %ymm7
6341 ; AVX-NEXT: vinsertf128 $1, 96(%r10), %ymm5, %ymm5
6342 ; AVX-NEXT: vshufpd {{.*#+}} ymm5 = ymm5[0],ymm7[1],ymm5[2],ymm7[2]
6343 ; AVX-NEXT: vmovupd %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6344 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
6345 ; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm1
6346 ; AVX-NEXT: vbroadcastsd 104(%rdx), %ymm3
6347 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm3[4,5,6,7]
6348 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7]
6349 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6350 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm4[1],xmm2[1]
6351 ; AVX-NEXT: vinsertf128 $1, %xmm6, %ymm0, %ymm1
6352 ; AVX-NEXT: vbroadcastsd 104(%r10), %ymm2
6353 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
6354 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7]
6355 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6356 ; AVX-NEXT: vmovaps 128(%rsi), %xmm0
6357 ; AVX-NEXT: vmovaps 128(%rdi), %xmm1
6358 ; AVX-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
6359 ; AVX-NEXT: vmovaps 128(%rcx), %xmm3
6360 ; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm2, %ymm4
6361 ; AVX-NEXT: vinsertf128 $1, 128(%rdx), %ymm2, %ymm2
6362 ; AVX-NEXT: vshufpd {{.*#+}} ymm2 = ymm2[0],ymm4[1],ymm2[2],ymm4[2]
6363 ; AVX-NEXT: vmovupd %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6364 ; AVX-NEXT: vmovaps 128(%r9), %xmm2
6365 ; AVX-NEXT: vmovaps 128(%r8), %xmm4
6366 ; AVX-NEXT: vmovlhps {{.*#+}} xmm5 = xmm4[0],xmm2[0]
6367 ; AVX-NEXT: vmovaps 128(%rax), %xmm6
6368 ; AVX-NEXT: vinsertf128 $1, %xmm6, %ymm5, %ymm7
6369 ; AVX-NEXT: vinsertf128 $1, 128(%r10), %ymm5, %ymm5
6370 ; AVX-NEXT: vshufpd {{.*#+}} ymm5 = ymm5[0],ymm7[1],ymm5[2],ymm7[2]
6371 ; AVX-NEXT: vmovupd %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6372 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
6373 ; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm1
6374 ; AVX-NEXT: vbroadcastsd 136(%rdx), %ymm3
6375 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm3[4,5,6,7]
6376 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7]
6377 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6378 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm4[1],xmm2[1]
6379 ; AVX-NEXT: vinsertf128 $1, %xmm6, %ymm0, %ymm1
6380 ; AVX-NEXT: vbroadcastsd 136(%r10), %ymm2
6381 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
6382 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7]
6383 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6384 ; AVX-NEXT: vmovaps 160(%rsi), %xmm0
6385 ; AVX-NEXT: vmovaps 160(%rdi), %xmm1
6386 ; AVX-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
6387 ; AVX-NEXT: vmovaps 160(%rcx), %xmm3
6388 ; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm2, %ymm4
6389 ; AVX-NEXT: vinsertf128 $1, 160(%rdx), %ymm2, %ymm2
6390 ; AVX-NEXT: vshufpd {{.*#+}} ymm2 = ymm2[0],ymm4[1],ymm2[2],ymm4[2]
6391 ; AVX-NEXT: vmovupd %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6392 ; AVX-NEXT: vmovaps 160(%rax), %xmm2
6393 ; AVX-NEXT: vmovaps 160(%r9), %xmm4
6394 ; AVX-NEXT: vmovaps 160(%r8), %xmm5
6395 ; AVX-NEXT: vmovlhps {{.*#+}} xmm6 = xmm5[0],xmm4[0]
6396 ; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm6, %ymm7
6397 ; AVX-NEXT: vinsertf128 $1, 160(%r10), %ymm6, %ymm6
6398 ; AVX-NEXT: vshufpd {{.*#+}} ymm6 = ymm6[0],ymm7[1],ymm6[2],ymm7[2]
6399 ; AVX-NEXT: vmovupd %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6400 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
6401 ; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm1
6402 ; AVX-NEXT: vbroadcastsd 168(%rdx), %ymm3
6403 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm3[4,5,6,7]
6404 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7]
6405 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6406 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm5[1],xmm4[1]
6407 ; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm1
6408 ; AVX-NEXT: vbroadcastsd 168(%r10), %ymm2
6409 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
6410 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7]
6411 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6412 ; AVX-NEXT: vmovaps 192(%rsi), %xmm0
6413 ; AVX-NEXT: vmovaps 192(%rdi), %xmm1
6414 ; AVX-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
6415 ; AVX-NEXT: vmovaps 192(%rcx), %xmm3
6416 ; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm2, %ymm4
6417 ; AVX-NEXT: vinsertf128 $1, 192(%rdx), %ymm2, %ymm2
6418 ; AVX-NEXT: vshufpd {{.*#+}} ymm2 = ymm2[0],ymm4[1],ymm2[2],ymm4[2]
6419 ; AVX-NEXT: vmovupd %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6420 ; AVX-NEXT: vmovaps 192(%rax), %xmm2
6421 ; AVX-NEXT: vmovaps 192(%r9), %xmm4
6422 ; AVX-NEXT: vmovaps 192(%r8), %xmm5
6423 ; AVX-NEXT: vmovlhps {{.*#+}} xmm6 = xmm5[0],xmm4[0]
6424 ; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm6, %ymm7
6425 ; AVX-NEXT: vinsertf128 $1, 192(%r10), %ymm6, %ymm6
6426 ; AVX-NEXT: vshufpd {{.*#+}} ymm6 = ymm6[0],ymm7[1],ymm6[2],ymm7[2]
6427 ; AVX-NEXT: vmovupd %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6428 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
6429 ; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm1
6430 ; AVX-NEXT: vbroadcastsd 200(%rdx), %ymm3
6431 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm3[4,5,6,7]
6432 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7]
6433 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6434 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm5[1],xmm4[1]
6435 ; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm1
6436 ; AVX-NEXT: vbroadcastsd 200(%r10), %ymm2
6437 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
6438 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7]
6439 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6440 ; AVX-NEXT: vmovaps 224(%rsi), %xmm0
6441 ; AVX-NEXT: vmovaps 224(%rdi), %xmm1
6442 ; AVX-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
6443 ; AVX-NEXT: vmovaps 224(%rcx), %xmm3
6444 ; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm2, %ymm4
6445 ; AVX-NEXT: vinsertf128 $1, 224(%rdx), %ymm2, %ymm2
6446 ; AVX-NEXT: vshufpd {{.*#+}} ymm2 = ymm2[0],ymm4[1],ymm2[2],ymm4[2]
6447 ; AVX-NEXT: vmovupd %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6448 ; AVX-NEXT: vmovaps 224(%rax), %xmm2
6449 ; AVX-NEXT: vmovaps 224(%r9), %xmm4
6450 ; AVX-NEXT: vmovaps 224(%r8), %xmm5
6451 ; AVX-NEXT: vmovlhps {{.*#+}} xmm6 = xmm5[0],xmm4[0]
6452 ; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm6, %ymm7
6453 ; AVX-NEXT: vinsertf128 $1, 224(%r10), %ymm6, %ymm6
6454 ; AVX-NEXT: vshufpd {{.*#+}} ymm6 = ymm6[0],ymm7[1],ymm6[2],ymm7[2]
6455 ; AVX-NEXT: vmovupd %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6456 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
6457 ; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm1
6458 ; AVX-NEXT: vbroadcastsd 232(%rdx), %ymm3
6459 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm3[4,5,6,7]
6460 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7]
6461 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6462 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm5[1],xmm4[1]
6463 ; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm1
6464 ; AVX-NEXT: vbroadcastsd 232(%r10), %ymm2
6465 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
6466 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7]
6467 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6468 ; AVX-NEXT: vmovaps 16(%rsi), %xmm0
6469 ; AVX-NEXT: vmovaps 16(%rdi), %xmm1
6470 ; AVX-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
6471 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],mem[4,5,6,7]
6472 ; AVX-NEXT: vbroadcastsd 16(%rcx), %ymm3
6473 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
6474 ; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6475 ; AVX-NEXT: vmovaps 16(%r9), %xmm2
6476 ; AVX-NEXT: vmovaps 16(%r8), %xmm3
6477 ; AVX-NEXT: vmovlhps {{.*#+}} xmm4 = xmm3[0],xmm2[0]
6478 ; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],mem[4,5,6,7]
6479 ; AVX-NEXT: vbroadcastsd 16(%rax), %ymm5
6480 ; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
6481 ; AVX-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6482 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
6483 ; AVX-NEXT: vbroadcastsd 24(%rdx), %ymm1
6484 ; AVX-NEXT: vblendps {{.*#+}} ymm9 = ymm0[0,1,2,3],ymm1[4,5,6,7]
6485 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm3[1],xmm2[1]
6486 ; AVX-NEXT: vbroadcastsd 24(%r10), %ymm1
6487 ; AVX-NEXT: vblendps {{.*#+}} ymm13 = ymm0[0,1,2,3],ymm1[4,5,6,7]
6488 ; AVX-NEXT: vmovaps 48(%rsi), %xmm0
6489 ; AVX-NEXT: vmovaps 48(%rdi), %xmm1
6490 ; AVX-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
6491 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],mem[4,5,6,7]
6492 ; AVX-NEXT: vbroadcastsd 48(%rcx), %ymm3
6493 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
6494 ; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6495 ; AVX-NEXT: vmovaps 48(%r9), %xmm2
6496 ; AVX-NEXT: vmovaps 48(%r8), %xmm3
6497 ; AVX-NEXT: vmovlhps {{.*#+}} xmm4 = xmm3[0],xmm2[0]
6498 ; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],mem[4,5,6,7]
6499 ; AVX-NEXT: vbroadcastsd 48(%rax), %ymm5
6500 ; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
6501 ; AVX-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6502 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
6503 ; AVX-NEXT: vbroadcastsd 56(%rdx), %ymm1
6504 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
6505 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6506 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm3[1],xmm2[1]
6507 ; AVX-NEXT: vbroadcastsd 56(%r10), %ymm1
6508 ; AVX-NEXT: vblendps {{.*#+}} ymm12 = ymm0[0,1,2,3],ymm1[4,5,6,7]
6509 ; AVX-NEXT: vmovaps 80(%rsi), %xmm0
6510 ; AVX-NEXT: vmovaps 80(%rdi), %xmm1
6511 ; AVX-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
6512 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],mem[4,5,6,7]
6513 ; AVX-NEXT: vbroadcastsd 80(%rcx), %ymm3
6514 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
6515 ; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6516 ; AVX-NEXT: vmovaps 80(%r9), %xmm2
6517 ; AVX-NEXT: vmovaps 80(%r8), %xmm3
6518 ; AVX-NEXT: vmovlhps {{.*#+}} xmm4 = xmm3[0],xmm2[0]
6519 ; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],mem[4,5,6,7]
6520 ; AVX-NEXT: vbroadcastsd 80(%rax), %ymm5
6521 ; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
6522 ; AVX-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6523 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
6524 ; AVX-NEXT: vbroadcastsd 88(%rdx), %ymm1
6525 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
6526 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6527 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm3[1],xmm2[1]
6528 ; AVX-NEXT: vbroadcastsd 88(%r10), %ymm1
6529 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
6530 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6531 ; AVX-NEXT: vmovaps 112(%rsi), %xmm0
6532 ; AVX-NEXT: vmovaps 112(%rdi), %xmm1
6533 ; AVX-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
6534 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],mem[4,5,6,7]
6535 ; AVX-NEXT: vbroadcastsd 112(%rcx), %ymm3
6536 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
6537 ; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6538 ; AVX-NEXT: vmovaps 112(%r9), %xmm2
6539 ; AVX-NEXT: vmovaps 112(%r8), %xmm3
6540 ; AVX-NEXT: vmovlhps {{.*#+}} xmm4 = xmm3[0],xmm2[0]
6541 ; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],mem[4,5,6,7]
6542 ; AVX-NEXT: vbroadcastsd 112(%rax), %ymm5
6543 ; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
6544 ; AVX-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6545 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
6546 ; AVX-NEXT: vbroadcastsd 120(%rdx), %ymm1
6547 ; AVX-NEXT: vblendps {{.*#+}} ymm11 = ymm0[0,1,2,3],ymm1[4,5,6,7]
6548 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm3[1],xmm2[1]
6549 ; AVX-NEXT: vbroadcastsd 120(%r10), %ymm1
6550 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
6551 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6552 ; AVX-NEXT: vmovaps 144(%rsi), %xmm0
6553 ; AVX-NEXT: vmovaps 144(%rdi), %xmm1
6554 ; AVX-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
6555 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],mem[4,5,6,7]
6556 ; AVX-NEXT: vbroadcastsd 144(%rcx), %ymm3
6557 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
6558 ; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6559 ; AVX-NEXT: vmovaps 144(%r9), %xmm2
6560 ; AVX-NEXT: vmovaps 144(%r8), %xmm3
6561 ; AVX-NEXT: vmovlhps {{.*#+}} xmm4 = xmm3[0],xmm2[0]
6562 ; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],mem[4,5,6,7]
6563 ; AVX-NEXT: vbroadcastsd 144(%rax), %ymm5
6564 ; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
6565 ; AVX-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6566 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
6567 ; AVX-NEXT: vbroadcastsd 152(%rdx), %ymm1
6568 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
6569 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6570 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm3[1],xmm2[1]
6571 ; AVX-NEXT: vbroadcastsd 152(%r10), %ymm1
6572 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
6573 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6574 ; AVX-NEXT: vmovaps 176(%rsi), %xmm0
6575 ; AVX-NEXT: vmovaps 176(%rdi), %xmm1
6576 ; AVX-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
6577 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],mem[4,5,6,7]
6578 ; AVX-NEXT: vbroadcastsd 176(%rcx), %ymm3
6579 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
6580 ; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6581 ; AVX-NEXT: vmovaps 176(%r9), %xmm2
6582 ; AVX-NEXT: vmovaps 176(%r8), %xmm3
6583 ; AVX-NEXT: vmovlhps {{.*#+}} xmm4 = xmm3[0],xmm2[0]
6584 ; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],mem[4,5,6,7]
6585 ; AVX-NEXT: vbroadcastsd 176(%rax), %ymm5
6586 ; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
6587 ; AVX-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6588 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
6589 ; AVX-NEXT: vbroadcastsd 184(%rdx), %ymm1
6590 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
6591 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6592 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm3[1],xmm2[1]
6593 ; AVX-NEXT: vbroadcastsd 184(%r10), %ymm1
6594 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
6595 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6596 ; AVX-NEXT: vmovaps 208(%rsi), %xmm0
6597 ; AVX-NEXT: vmovaps 208(%rdi), %xmm1
6598 ; AVX-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
6599 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],mem[4,5,6,7]
6600 ; AVX-NEXT: vbroadcastsd 208(%rcx), %ymm3
6601 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
6602 ; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6603 ; AVX-NEXT: vmovaps 208(%r9), %xmm2
6604 ; AVX-NEXT: vmovaps 208(%r8), %xmm3
6605 ; AVX-NEXT: vmovlhps {{.*#+}} xmm4 = xmm3[0],xmm2[0]
6606 ; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],mem[4,5,6,7]
6607 ; AVX-NEXT: vbroadcastsd 208(%rax), %ymm5
6608 ; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
6609 ; AVX-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6610 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
6611 ; AVX-NEXT: vbroadcastsd 216(%rdx), %ymm1
6612 ; AVX-NEXT: vblendps {{.*#+}} ymm10 = ymm0[0,1,2,3],ymm1[4,5,6,7]
6613 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm3[1],xmm2[1]
6614 ; AVX-NEXT: vbroadcastsd 216(%r10), %ymm1
6615 ; AVX-NEXT: vblendps {{.*#+}} ymm8 = ymm0[0,1,2,3],ymm1[4,5,6,7]
6616 ; AVX-NEXT: vmovaps 240(%rsi), %xmm0
6617 ; AVX-NEXT: vmovaps 240(%rdi), %xmm3
6618 ; AVX-NEXT: vmovlhps {{.*#+}} xmm1 = xmm3[0],xmm0[0]
6619 ; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],mem[4,5,6,7]
6620 ; AVX-NEXT: vbroadcastsd 240(%rcx), %ymm5
6621 ; AVX-NEXT: vblendps {{.*#+}} ymm7 = ymm1[0,1,2,3,4,5],ymm5[6,7]
6622 ; AVX-NEXT: vmovaps 240(%r9), %xmm5
6623 ; AVX-NEXT: vmovaps 240(%r8), %xmm6
6624 ; AVX-NEXT: vmovlhps {{.*#+}} xmm15 = xmm6[0],xmm5[0]
6625 ; AVX-NEXT: vblendps {{.*#+}} ymm15 = ymm15[0,1,2,3],mem[4,5,6,7]
6626 ; AVX-NEXT: vbroadcastsd 240(%rax), %ymm14
6627 ; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm15[0,1,2,3,4,5],ymm14[6,7]
6628 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm3[1],xmm0[1]
6629 ; AVX-NEXT: vbroadcastsd 248(%rdx), %ymm3
6630 ; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm0[0,1,2,3],ymm3[4,5,6,7]
6631 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm2 = xmm6[1],xmm5[1]
6632 ; AVX-NEXT: vbroadcastsd 248(%r10), %ymm3
6633 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm3[4,5,6,7]
6634 ; AVX-NEXT: movq {{[0-9]+}}(%rsp), %rdx
6635 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm9[0,1,2,3,4,5],mem[6,7]
6636 ; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6637 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm13[0,1,2,3,4,5],mem[6,7]
6638 ; AVX-NEXT: vmovups %ymm2, (%rsp) # 32-byte Spill
6639 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
6640 ; AVX-NEXT: vblendps {{.*#+}} ymm15 = ymm2[0,1,2,3,4,5],mem[6,7]
6641 ; AVX-NEXT: vblendps {{.*#+}} ymm14 = ymm12[0,1,2,3,4,5],mem[6,7]
6642 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
6643 ; AVX-NEXT: vblendps {{.*#+}} ymm13 = ymm2[0,1,2,3,4,5],mem[6,7]
6644 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
6645 ; AVX-NEXT: vblendps {{.*#+}} ymm12 = ymm2[0,1,2,3,4,5],mem[6,7]
6646 ; AVX-NEXT: vblendps {{.*#+}} ymm11 = ymm11[0,1,2,3,4,5],mem[6,7]
6647 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
6648 ; AVX-NEXT: vblendps {{.*#+}} ymm9 = ymm2[0,1,2,3,4,5],mem[6,7]
6649 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
6650 ; AVX-NEXT: vblendps {{.*#+}} ymm6 = ymm2[0,1,2,3,4,5],mem[6,7]
6651 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
6652 ; AVX-NEXT: vblendps {{.*#+}} ymm5 = ymm2[0,1,2,3,4,5],mem[6,7]
6653 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
6654 ; AVX-NEXT: vblendps {{.*#+}} ymm3 = ymm2[0,1,2,3,4,5],mem[6,7]
6655 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
6656 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],mem[6,7]
6657 ; AVX-NEXT: vblendps {{.*#+}} ymm10 = ymm10[0,1,2,3,4,5],mem[6,7]
6658 ; AVX-NEXT: vblendps {{.*#+}} ymm8 = ymm8[0,1,2,3,4,5],mem[6,7]
6659 ; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5],mem[6,7]
6660 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],mem[6,7]
6661 ; AVX-NEXT: vmovaps %ymm0, 2016(%rdx)
6662 ; AVX-NEXT: vmovaps %ymm1, 1984(%rdx)
6663 ; AVX-NEXT: vmovaps %ymm4, 1952(%rdx)
6664 ; AVX-NEXT: vmovaps %ymm7, 1920(%rdx)
6665 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6666 ; AVX-NEXT: vmovaps %ymm0, 1888(%rdx)
6667 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6668 ; AVX-NEXT: vmovaps %ymm0, 1856(%rdx)
6669 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6670 ; AVX-NEXT: vmovaps %ymm0, 1824(%rdx)
6671 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6672 ; AVX-NEXT: vmovaps %ymm0, 1792(%rdx)
6673 ; AVX-NEXT: vmovaps %ymm8, 1760(%rdx)
6674 ; AVX-NEXT: vmovaps %ymm10, 1728(%rdx)
6675 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6676 ; AVX-NEXT: vmovaps %ymm0, 1696(%rdx)
6677 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6678 ; AVX-NEXT: vmovaps %ymm0, 1664(%rdx)
6679 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6680 ; AVX-NEXT: vmovaps %ymm0, 1632(%rdx)
6681 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6682 ; AVX-NEXT: vmovaps %ymm0, 1600(%rdx)
6683 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6684 ; AVX-NEXT: vmovaps %ymm0, 1568(%rdx)
6685 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6686 ; AVX-NEXT: vmovaps %ymm0, 1536(%rdx)
6687 ; AVX-NEXT: vmovaps %ymm2, 1504(%rdx)
6688 ; AVX-NEXT: vmovaps %ymm3, 1472(%rdx)
6689 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6690 ; AVX-NEXT: vmovaps %ymm0, 1440(%rdx)
6691 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6692 ; AVX-NEXT: vmovaps %ymm0, 1408(%rdx)
6693 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6694 ; AVX-NEXT: vmovaps %ymm0, 1376(%rdx)
6695 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6696 ; AVX-NEXT: vmovaps %ymm0, 1344(%rdx)
6697 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6698 ; AVX-NEXT: vmovaps %ymm0, 1312(%rdx)
6699 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6700 ; AVX-NEXT: vmovaps %ymm0, 1280(%rdx)
6701 ; AVX-NEXT: vmovaps %ymm5, 1248(%rdx)
6702 ; AVX-NEXT: vmovaps %ymm6, 1216(%rdx)
6703 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6704 ; AVX-NEXT: vmovaps %ymm0, 1184(%rdx)
6705 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6706 ; AVX-NEXT: vmovaps %ymm0, 1152(%rdx)
6707 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6708 ; AVX-NEXT: vmovaps %ymm0, 1120(%rdx)
6709 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6710 ; AVX-NEXT: vmovaps %ymm0, 1088(%rdx)
6711 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6712 ; AVX-NEXT: vmovaps %ymm0, 1056(%rdx)
6713 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6714 ; AVX-NEXT: vmovaps %ymm0, 1024(%rdx)
6715 ; AVX-NEXT: vmovaps %ymm9, 992(%rdx)
6716 ; AVX-NEXT: vmovaps %ymm11, 960(%rdx)
6717 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6718 ; AVX-NEXT: vmovaps %ymm0, 928(%rdx)
6719 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6720 ; AVX-NEXT: vmovaps %ymm0, 896(%rdx)
6721 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6722 ; AVX-NEXT: vmovaps %ymm0, 864(%rdx)
6723 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6724 ; AVX-NEXT: vmovaps %ymm0, 832(%rdx)
6725 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6726 ; AVX-NEXT: vmovaps %ymm0, 800(%rdx)
6727 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6728 ; AVX-NEXT: vmovaps %ymm0, 768(%rdx)
6729 ; AVX-NEXT: vmovaps %ymm12, 736(%rdx)
6730 ; AVX-NEXT: vmovaps %ymm13, 704(%rdx)
6731 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6732 ; AVX-NEXT: vmovaps %ymm0, 672(%rdx)
6733 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6734 ; AVX-NEXT: vmovaps %ymm0, 640(%rdx)
6735 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6736 ; AVX-NEXT: vmovaps %ymm0, 608(%rdx)
6737 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6738 ; AVX-NEXT: vmovaps %ymm0, 576(%rdx)
6739 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6740 ; AVX-NEXT: vmovaps %ymm0, 544(%rdx)
6741 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6742 ; AVX-NEXT: vmovaps %ymm0, 512(%rdx)
6743 ; AVX-NEXT: vmovaps %ymm14, 480(%rdx)
6744 ; AVX-NEXT: vmovaps %ymm15, 448(%rdx)
6745 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6746 ; AVX-NEXT: vmovaps %ymm0, 416(%rdx)
6747 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6748 ; AVX-NEXT: vmovaps %ymm0, 384(%rdx)
6749 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6750 ; AVX-NEXT: vmovaps %ymm0, 352(%rdx)
6751 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6752 ; AVX-NEXT: vmovaps %ymm0, 320(%rdx)
6753 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6754 ; AVX-NEXT: vmovaps %ymm0, 288(%rdx)
6755 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6756 ; AVX-NEXT: vmovaps %ymm0, 256(%rdx)
6757 ; AVX-NEXT: vmovups (%rsp), %ymm0 # 32-byte Reload
6758 ; AVX-NEXT: vmovaps %ymm0, 224(%rdx)
6759 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6760 ; AVX-NEXT: vmovaps %ymm0, 192(%rdx)
6761 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6762 ; AVX-NEXT: vmovaps %ymm0, 160(%rdx)
6763 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6764 ; AVX-NEXT: vmovaps %ymm0, 128(%rdx)
6765 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6766 ; AVX-NEXT: vmovaps %ymm0, 96(%rdx)
6767 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6768 ; AVX-NEXT: vmovaps %ymm0, 64(%rdx)
6769 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6770 ; AVX-NEXT: vmovaps %ymm0, 32(%rdx)
6771 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
6772 ; AVX-NEXT: vmovaps %ymm0, (%rdx)
6773 ; AVX-NEXT: addq $1672, %rsp # imm = 0x688
6774 ; AVX-NEXT: vzeroupper
6775 ; AVX-NEXT: retq
6776 ;
6777 ; AVX2-LABEL: store_i64_stride8_vf32:
6778 ; AVX2: # %bb.0:
6779 ; AVX2-NEXT: subq $1704, %rsp # imm = 0x6A8
6780 ; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %r10
6781 ; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %rax
6782 ; AVX2-NEXT: vmovaps (%rcx), %xmm0
6783 ; AVX2-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
6784 ; AVX2-NEXT: vmovaps 32(%rcx), %xmm3
6785 ; AVX2-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
6786 ; AVX2-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0
6787 ; AVX2-NEXT: vmovaps (%rsi), %xmm2
6788 ; AVX2-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
6789 ; AVX2-NEXT: vmovaps 32(%rsi), %xmm4
6790 ; AVX2-NEXT: vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
6791 ; AVX2-NEXT: vmovaps (%rdi), %xmm1
6792 ; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
6793 ; AVX2-NEXT: vmovaps 32(%rdi), %xmm5
6794 ; AVX2-NEXT: vmovaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
6795 ; AVX2-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm2[1]
6796 ; AVX2-NEXT: vbroadcastsd 8(%rdx), %ymm2
6797 ; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
6798 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
6799 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6800 ; AVX2-NEXT: vmovaps (%rax), %xmm0
6801 ; AVX2-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
6802 ; AVX2-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0
6803 ; AVX2-NEXT: vmovaps (%r9), %xmm2
6804 ; AVX2-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
6805 ; AVX2-NEXT: vmovaps 32(%r9), %xmm6
6806 ; AVX2-NEXT: vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
6807 ; AVX2-NEXT: vmovaps (%r8), %xmm1
6808 ; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
6809 ; AVX2-NEXT: vmovaps 32(%r8), %xmm7
6810 ; AVX2-NEXT: vmovaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
6811 ; AVX2-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm2[1]
6812 ; AVX2-NEXT: vbroadcastsd 8(%r10), %ymm2
6813 ; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
6814 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
6815 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6816 ; AVX2-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm0
6817 ; AVX2-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm5[1],xmm4[1]
6818 ; AVX2-NEXT: vbroadcastsd 40(%rdx), %ymm2
6819 ; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
6820 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
6821 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6822 ; AVX2-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm7[1],xmm6[1]
6823 ; AVX2-NEXT: vbroadcastsd 40(%r10), %ymm1
6824 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
6825 ; AVX2-NEXT: vmovaps 32(%rax), %xmm1
6826 ; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
6827 ; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
6828 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
6829 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6830 ; AVX2-NEXT: vmovaps 64(%rsi), %xmm1
6831 ; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
6832 ; AVX2-NEXT: vmovaps 64(%rdi), %xmm0
6833 ; AVX2-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
6834 ; AVX2-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
6835 ; AVX2-NEXT: vbroadcastsd 72(%rdx), %ymm1
6836 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
6837 ; AVX2-NEXT: vmovaps 64(%rcx), %xmm1
6838 ; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
6839 ; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
6840 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
6841 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6842 ; AVX2-NEXT: vmovaps 64(%r9), %xmm1
6843 ; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
6844 ; AVX2-NEXT: vmovaps 64(%r8), %xmm0
6845 ; AVX2-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
6846 ; AVX2-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
6847 ; AVX2-NEXT: vbroadcastsd 72(%r10), %ymm1
6848 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
6849 ; AVX2-NEXT: vmovaps 64(%rax), %xmm1
6850 ; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
6851 ; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
6852 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
6853 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6854 ; AVX2-NEXT: vmovaps 96(%rsi), %xmm1
6855 ; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
6856 ; AVX2-NEXT: vmovaps 96(%rdi), %xmm0
6857 ; AVX2-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
6858 ; AVX2-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
6859 ; AVX2-NEXT: vbroadcastsd 104(%rdx), %ymm1
6860 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
6861 ; AVX2-NEXT: vmovaps 96(%rcx), %xmm1
6862 ; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
6863 ; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
6864 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
6865 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6866 ; AVX2-NEXT: vmovaps 96(%r9), %xmm1
6867 ; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
6868 ; AVX2-NEXT: vmovaps 96(%r8), %xmm0
6869 ; AVX2-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
6870 ; AVX2-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
6871 ; AVX2-NEXT: vbroadcastsd 104(%r10), %ymm1
6872 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
6873 ; AVX2-NEXT: vmovaps 96(%rax), %xmm1
6874 ; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
6875 ; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
6876 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
6877 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6878 ; AVX2-NEXT: vmovaps 128(%rsi), %xmm1
6879 ; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
6880 ; AVX2-NEXT: vmovaps 128(%rdi), %xmm0
6881 ; AVX2-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
6882 ; AVX2-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
6883 ; AVX2-NEXT: vbroadcastsd 136(%rdx), %ymm1
6884 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
6885 ; AVX2-NEXT: vmovaps 128(%rcx), %xmm1
6886 ; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
6887 ; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
6888 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
6889 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6890 ; AVX2-NEXT: vmovaps 128(%r9), %xmm1
6891 ; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
6892 ; AVX2-NEXT: vmovaps 128(%r8), %xmm0
6893 ; AVX2-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
6894 ; AVX2-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
6895 ; AVX2-NEXT: vbroadcastsd 136(%r10), %ymm1
6896 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
6897 ; AVX2-NEXT: vmovaps 128(%rax), %xmm1
6898 ; AVX2-NEXT: vmovaps %xmm1, (%rsp) # 16-byte Spill
6899 ; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
6900 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
6901 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6902 ; AVX2-NEXT: vmovaps 160(%rsi), %xmm1
6903 ; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
6904 ; AVX2-NEXT: vmovaps 160(%rdi), %xmm0
6905 ; AVX2-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
6906 ; AVX2-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
6907 ; AVX2-NEXT: vbroadcastsd 168(%rdx), %ymm1
6908 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
6909 ; AVX2-NEXT: vmovaps 160(%rcx), %xmm1
6910 ; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
6911 ; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
6912 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
6913 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6914 ; AVX2-NEXT: vmovaps 160(%r9), %xmm0
6915 ; AVX2-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
6916 ; AVX2-NEXT: vmovaps 160(%r8), %xmm13
6917 ; AVX2-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm13[1],xmm0[1]
6918 ; AVX2-NEXT: vbroadcastsd 168(%r10), %ymm1
6919 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
6920 ; AVX2-NEXT: vmovaps 160(%rax), %xmm12
6921 ; AVX2-NEXT: vinsertf128 $1, %xmm12, %ymm0, %ymm1
6922 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
6923 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6924 ; AVX2-NEXT: vmovaps 192(%rsi), %xmm11
6925 ; AVX2-NEXT: vmovaps 192(%rdi), %xmm10
6926 ; AVX2-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm10[1],xmm11[1]
6927 ; AVX2-NEXT: vbroadcastsd 200(%rdx), %ymm1
6928 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
6929 ; AVX2-NEXT: vmovaps 192(%rcx), %xmm9
6930 ; AVX2-NEXT: vinsertf128 $1, %xmm9, %ymm0, %ymm1
6931 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
6932 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6933 ; AVX2-NEXT: vmovaps 192(%r9), %xmm8
6934 ; AVX2-NEXT: vmovaps 192(%r8), %xmm7
6935 ; AVX2-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm7[1],xmm8[1]
6936 ; AVX2-NEXT: vbroadcastsd 200(%r10), %ymm1
6937 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
6938 ; AVX2-NEXT: vmovaps 192(%rax), %xmm6
6939 ; AVX2-NEXT: vinsertf128 $1, %xmm6, %ymm0, %ymm1
6940 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
6941 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6942 ; AVX2-NEXT: vmovaps 224(%rsi), %xmm5
6943 ; AVX2-NEXT: vmovaps 224(%rdi), %xmm4
6944 ; AVX2-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm4[1],xmm5[1]
6945 ; AVX2-NEXT: vbroadcastsd 232(%rdx), %ymm1
6946 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
6947 ; AVX2-NEXT: vmovaps 224(%rcx), %xmm3
6948 ; AVX2-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm1
6949 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
6950 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6951 ; AVX2-NEXT: vmovaps 224(%r9), %xmm2
6952 ; AVX2-NEXT: vmovaps 224(%r8), %xmm1
6953 ; AVX2-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm2[1]
6954 ; AVX2-NEXT: vbroadcastsd 232(%r10), %ymm15
6955 ; AVX2-NEXT: vblendps {{.*#+}} ymm15 = ymm0[0,1,2,3],ymm15[4,5,6,7]
6956 ; AVX2-NEXT: vmovaps 224(%rax), %xmm0
6957 ; AVX2-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm14
6958 ; AVX2-NEXT: vblendps {{.*#+}} ymm14 = ymm15[0,1,2,3,4,5],ymm14[6,7]
6959 ; AVX2-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6960 ; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
6961 ; AVX2-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
6962 ; AVX2-NEXT: # xmm14 = xmm14[0],mem[0]
6963 ; AVX2-NEXT: vinsertf128 $1, (%rdx), %ymm14, %ymm14
6964 ; AVX2-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
6965 ; AVX2-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
6966 ; AVX2-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6967 ; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
6968 ; AVX2-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
6969 ; AVX2-NEXT: # xmm14 = xmm14[0],mem[0]
6970 ; AVX2-NEXT: vinsertf128 $1, (%r10), %ymm14, %ymm14
6971 ; AVX2-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
6972 ; AVX2-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
6973 ; AVX2-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6974 ; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
6975 ; AVX2-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
6976 ; AVX2-NEXT: # xmm14 = xmm14[0],mem[0]
6977 ; AVX2-NEXT: vinsertf128 $1, 32(%rdx), %ymm14, %ymm14
6978 ; AVX2-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
6979 ; AVX2-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
6980 ; AVX2-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6981 ; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
6982 ; AVX2-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
6983 ; AVX2-NEXT: # xmm14 = xmm14[0],mem[0]
6984 ; AVX2-NEXT: vinsertf128 $1, 32(%r10), %ymm14, %ymm14
6985 ; AVX2-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
6986 ; AVX2-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
6987 ; AVX2-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6988 ; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
6989 ; AVX2-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
6990 ; AVX2-NEXT: # xmm14 = xmm14[0],mem[0]
6991 ; AVX2-NEXT: vinsertf128 $1, 64(%rdx), %ymm14, %ymm14
6992 ; AVX2-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
6993 ; AVX2-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
6994 ; AVX2-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
6995 ; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
6996 ; AVX2-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
6997 ; AVX2-NEXT: # xmm14 = xmm14[0],mem[0]
6998 ; AVX2-NEXT: vinsertf128 $1, 64(%r10), %ymm14, %ymm14
6999 ; AVX2-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
7000 ; AVX2-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
7001 ; AVX2-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7002 ; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
7003 ; AVX2-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
7004 ; AVX2-NEXT: # xmm14 = xmm14[0],mem[0]
7005 ; AVX2-NEXT: vinsertf128 $1, 96(%rdx), %ymm14, %ymm14
7006 ; AVX2-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
7007 ; AVX2-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
7008 ; AVX2-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7009 ; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
7010 ; AVX2-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
7011 ; AVX2-NEXT: # xmm14 = xmm14[0],mem[0]
7012 ; AVX2-NEXT: vinsertf128 $1, 96(%r10), %ymm14, %ymm14
7013 ; AVX2-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
7014 ; AVX2-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
7015 ; AVX2-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7016 ; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
7017 ; AVX2-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
7018 ; AVX2-NEXT: # xmm14 = xmm14[0],mem[0]
7019 ; AVX2-NEXT: vinsertf128 $1, 128(%rdx), %ymm14, %ymm14
7020 ; AVX2-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
7021 ; AVX2-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
7022 ; AVX2-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7023 ; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
7024 ; AVX2-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
7025 ; AVX2-NEXT: # xmm14 = xmm14[0],mem[0]
7026 ; AVX2-NEXT: vinsertf128 $1, 128(%r10), %ymm14, %ymm14
7027 ; AVX2-NEXT: vbroadcastsd (%rsp), %ymm15 # 16-byte Folded Reload
7028 ; AVX2-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
7029 ; AVX2-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7030 ; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
7031 ; AVX2-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
7032 ; AVX2-NEXT: # xmm14 = xmm14[0],mem[0]
7033 ; AVX2-NEXT: vinsertf128 $1, 160(%rdx), %ymm14, %ymm14
7034 ; AVX2-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
7035 ; AVX2-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
7036 ; AVX2-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7037 ; AVX2-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm13, %xmm13 # 16-byte Folded Reload
7038 ; AVX2-NEXT: # xmm13 = xmm13[0],mem[0]
7039 ; AVX2-NEXT: vinsertf128 $1, 160(%r10), %ymm13, %ymm13
7040 ; AVX2-NEXT: vbroadcastsd %xmm12, %ymm12
7041 ; AVX2-NEXT: vblendps {{.*#+}} ymm12 = ymm13[0,1,2,3,4,5],ymm12[6,7]
7042 ; AVX2-NEXT: vmovups %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7043 ; AVX2-NEXT: vmovlhps {{.*#+}} xmm10 = xmm10[0],xmm11[0]
7044 ; AVX2-NEXT: vinsertf128 $1, 192(%rdx), %ymm10, %ymm10
7045 ; AVX2-NEXT: vbroadcastsd %xmm9, %ymm9
7046 ; AVX2-NEXT: vblendps {{.*#+}} ymm9 = ymm10[0,1,2,3,4,5],ymm9[6,7]
7047 ; AVX2-NEXT: vmovups %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7048 ; AVX2-NEXT: vmovlhps {{.*#+}} xmm7 = xmm7[0],xmm8[0]
7049 ; AVX2-NEXT: vinsertf128 $1, 192(%r10), %ymm7, %ymm7
7050 ; AVX2-NEXT: vbroadcastsd %xmm6, %ymm6
7051 ; AVX2-NEXT: vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3,4,5],ymm6[6,7]
7052 ; AVX2-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7053 ; AVX2-NEXT: vmovlhps {{.*#+}} xmm4 = xmm4[0],xmm5[0]
7054 ; AVX2-NEXT: vinsertf128 $1, 224(%rdx), %ymm4, %ymm4
7055 ; AVX2-NEXT: vbroadcastsd %xmm3, %ymm3
7056 ; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm4[0,1,2,3,4,5],ymm3[6,7]
7057 ; AVX2-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7058 ; AVX2-NEXT: vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
7059 ; AVX2-NEXT: vinsertf128 $1, 224(%r10), %ymm1, %ymm1
7060 ; AVX2-NEXT: vbroadcastsd %xmm0, %ymm0
7061 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
7062 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7063 ; AVX2-NEXT: vmovaps (%rdi), %ymm0
7064 ; AVX2-NEXT: vmovaps (%rsi), %ymm1
7065 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
7066 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
7067 ; AVX2-NEXT: vbroadcastsd 16(%rcx), %ymm3
7068 ; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
7069 ; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7070 ; AVX2-NEXT: vmovaps (%r8), %ymm2
7071 ; AVX2-NEXT: vmovaps (%r9), %ymm3
7072 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
7073 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],mem[2,3]
7074 ; AVX2-NEXT: vbroadcastsd 16(%rax), %ymm5
7075 ; AVX2-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
7076 ; AVX2-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7077 ; AVX2-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
7078 ; AVX2-NEXT: vbroadcastsd 24(%rdx), %ymm1
7079 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm8 = ymm0[2,3],ymm1[2,3]
7080 ; AVX2-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
7081 ; AVX2-NEXT: vbroadcastsd 24(%r10), %ymm1
7082 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm9 = ymm0[2,3],ymm1[2,3]
7083 ; AVX2-NEXT: vmovaps 32(%rdi), %ymm0
7084 ; AVX2-NEXT: vmovaps 32(%rsi), %ymm1
7085 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
7086 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
7087 ; AVX2-NEXT: vbroadcastsd 48(%rcx), %ymm3
7088 ; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
7089 ; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7090 ; AVX2-NEXT: vmovaps 32(%r8), %ymm2
7091 ; AVX2-NEXT: vmovaps 32(%r9), %ymm3
7092 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
7093 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],mem[2,3]
7094 ; AVX2-NEXT: vbroadcastsd 48(%rax), %ymm5
7095 ; AVX2-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
7096 ; AVX2-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7097 ; AVX2-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
7098 ; AVX2-NEXT: vbroadcastsd 56(%rdx), %ymm1
7099 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm12 = ymm0[2,3],ymm1[2,3]
7100 ; AVX2-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
7101 ; AVX2-NEXT: vbroadcastsd 56(%r10), %ymm1
7102 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
7103 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7104 ; AVX2-NEXT: vmovaps 64(%rdi), %ymm0
7105 ; AVX2-NEXT: vmovaps 64(%rsi), %ymm1
7106 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
7107 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
7108 ; AVX2-NEXT: vbroadcastsd 80(%rcx), %ymm3
7109 ; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
7110 ; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7111 ; AVX2-NEXT: vmovaps 64(%r8), %ymm2
7112 ; AVX2-NEXT: vmovaps 64(%r9), %ymm3
7113 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
7114 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],mem[2,3]
7115 ; AVX2-NEXT: vbroadcastsd 80(%rax), %ymm5
7116 ; AVX2-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
7117 ; AVX2-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7118 ; AVX2-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
7119 ; AVX2-NEXT: vbroadcastsd 88(%rdx), %ymm1
7120 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
7121 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7122 ; AVX2-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
7123 ; AVX2-NEXT: vbroadcastsd 88(%r10), %ymm1
7124 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
7125 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7126 ; AVX2-NEXT: vmovaps 96(%rdi), %ymm0
7127 ; AVX2-NEXT: vmovaps 96(%rsi), %ymm1
7128 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
7129 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
7130 ; AVX2-NEXT: vbroadcastsd 112(%rcx), %ymm3
7131 ; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
7132 ; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7133 ; AVX2-NEXT: vmovaps 96(%r8), %ymm2
7134 ; AVX2-NEXT: vmovaps 96(%r9), %ymm3
7135 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
7136 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],mem[2,3]
7137 ; AVX2-NEXT: vbroadcastsd 112(%rax), %ymm5
7138 ; AVX2-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
7139 ; AVX2-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7140 ; AVX2-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
7141 ; AVX2-NEXT: vbroadcastsd 120(%rdx), %ymm1
7142 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
7143 ; AVX2-NEXT: vmovups %ymm0, (%rsp) # 32-byte Spill
7144 ; AVX2-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
7145 ; AVX2-NEXT: vbroadcastsd 120(%r10), %ymm1
7146 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
7147 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7148 ; AVX2-NEXT: vmovaps 128(%rdi), %ymm0
7149 ; AVX2-NEXT: vmovaps 128(%rsi), %ymm1
7150 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
7151 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
7152 ; AVX2-NEXT: vbroadcastsd 144(%rcx), %ymm3
7153 ; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
7154 ; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7155 ; AVX2-NEXT: vmovaps 128(%r8), %ymm2
7156 ; AVX2-NEXT: vmovaps 128(%r9), %ymm3
7157 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
7158 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],mem[2,3]
7159 ; AVX2-NEXT: vbroadcastsd 144(%rax), %ymm5
7160 ; AVX2-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
7161 ; AVX2-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7162 ; AVX2-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
7163 ; AVX2-NEXT: vbroadcastsd 152(%rdx), %ymm1
7164 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
7165 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7166 ; AVX2-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
7167 ; AVX2-NEXT: vbroadcastsd 152(%r10), %ymm1
7168 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
7169 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7170 ; AVX2-NEXT: vmovaps 160(%rdi), %ymm0
7171 ; AVX2-NEXT: vmovaps 160(%rsi), %ymm1
7172 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
7173 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
7174 ; AVX2-NEXT: vbroadcastsd 176(%rcx), %ymm3
7175 ; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
7176 ; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7177 ; AVX2-NEXT: vmovaps 160(%r8), %ymm2
7178 ; AVX2-NEXT: vmovaps 160(%r9), %ymm3
7179 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
7180 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],mem[2,3]
7181 ; AVX2-NEXT: vbroadcastsd 176(%rax), %ymm5
7182 ; AVX2-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
7183 ; AVX2-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7184 ; AVX2-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
7185 ; AVX2-NEXT: vbroadcastsd 184(%rdx), %ymm1
7186 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
7187 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7188 ; AVX2-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
7189 ; AVX2-NEXT: vbroadcastsd 184(%r10), %ymm1
7190 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
7191 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7192 ; AVX2-NEXT: vmovaps 192(%rdi), %ymm0
7193 ; AVX2-NEXT: vmovaps 192(%rsi), %ymm1
7194 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
7195 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
7196 ; AVX2-NEXT: vbroadcastsd 208(%rcx), %ymm3
7197 ; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
7198 ; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7199 ; AVX2-NEXT: vmovaps 192(%r8), %ymm2
7200 ; AVX2-NEXT: vmovaps 192(%r9), %ymm3
7201 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
7202 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],mem[2,3]
7203 ; AVX2-NEXT: vbroadcastsd 208(%rax), %ymm5
7204 ; AVX2-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
7205 ; AVX2-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7206 ; AVX2-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
7207 ; AVX2-NEXT: vbroadcastsd 216(%rdx), %ymm1
7208 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm13 = ymm0[2,3],ymm1[2,3]
7209 ; AVX2-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
7210 ; AVX2-NEXT: vbroadcastsd 216(%r10), %ymm1
7211 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm11 = ymm0[2,3],ymm1[2,3]
7212 ; AVX2-NEXT: vmovaps 224(%rdi), %ymm0
7213 ; AVX2-NEXT: vmovaps 224(%rsi), %ymm3
7214 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm0[0],ymm3[0],ymm0[2],ymm3[2]
7215 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm1[2,3],mem[2,3]
7216 ; AVX2-NEXT: vbroadcastsd 240(%rcx), %ymm5
7217 ; AVX2-NEXT: vblendps {{.*#+}} ymm10 = ymm1[0,1,2,3,4,5],ymm5[6,7]
7218 ; AVX2-NEXT: vmovaps 224(%r8), %ymm5
7219 ; AVX2-NEXT: vmovaps 224(%r9), %ymm6
7220 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm7 = ymm5[0],ymm6[0],ymm5[2],ymm6[2]
7221 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm7 = ymm7[2,3],mem[2,3]
7222 ; AVX2-NEXT: vbroadcastsd 240(%rax), %ymm15
7223 ; AVX2-NEXT: vblendps {{.*#+}} ymm15 = ymm7[0,1,2,3,4,5],ymm15[6,7]
7224 ; AVX2-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm3[1],ymm0[3],ymm3[3]
7225 ; AVX2-NEXT: vbroadcastsd 248(%rdx), %ymm3
7226 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm0[2,3],ymm3[2,3]
7227 ; AVX2-NEXT: vunpckhpd {{.*#+}} ymm1 = ymm5[1],ymm6[1],ymm5[3],ymm6[3]
7228 ; AVX2-NEXT: vbroadcastsd 248(%r10), %ymm3
7229 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm1[2,3],ymm3[2,3]
7230 ; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %rdx
7231 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm8[0,1,2,3,4,5],mem[6,7]
7232 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7233 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm9[0,1,2,3,4,5],mem[6,7]
7234 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7235 ; AVX2-NEXT: vblendps {{.*#+}} ymm14 = ymm12[0,1,2,3,4,5],mem[6,7]
7236 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7237 ; AVX2-NEXT: vblendps {{.*#+}} ymm12 = ymm0[0,1,2,3,4,5],mem[6,7]
7238 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7239 ; AVX2-NEXT: vblendps {{.*#+}} ymm9 = ymm0[0,1,2,3,4,5],mem[6,7]
7240 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7241 ; AVX2-NEXT: vblendps {{.*#+}} ymm8 = ymm0[0,1,2,3,4,5],mem[6,7]
7242 ; AVX2-NEXT: vmovups (%rsp), %ymm0 # 32-byte Reload
7243 ; AVX2-NEXT: vblendps {{.*#+}} ymm7 = ymm0[0,1,2,3,4,5],mem[6,7]
7244 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7245 ; AVX2-NEXT: vblendps {{.*#+}} ymm6 = ymm0[0,1,2,3,4,5],mem[6,7]
7246 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7247 ; AVX2-NEXT: vblendps {{.*#+}} ymm5 = ymm0[0,1,2,3,4,5],mem[6,7]
7248 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7249 ; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm0[0,1,2,3,4,5],mem[6,7]
7250 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7251 ; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm0[0,1,2,3,4,5],mem[6,7]
7252 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7253 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],mem[6,7]
7254 ; AVX2-NEXT: vblendps {{.*#+}} ymm13 = ymm13[0,1,2,3,4,5],mem[6,7]
7255 ; AVX2-NEXT: vblendps {{.*#+}} ymm11 = ymm11[0,1,2,3,4,5],mem[6,7]
7256 ; AVX2-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],mem[6,7]
7257 ; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],mem[6,7]
7258 ; AVX2-NEXT: vmovaps %ymm2, 2016(%rdx)
7259 ; AVX2-NEXT: vmovaps %ymm4, 1984(%rdx)
7260 ; AVX2-NEXT: vmovaps %ymm15, 1952(%rdx)
7261 ; AVX2-NEXT: vmovaps %ymm10, 1920(%rdx)
7262 ; AVX2-NEXT: vmovaps %ymm11, 1760(%rdx)
7263 ; AVX2-NEXT: vmovaps %ymm13, 1728(%rdx)
7264 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
7265 ; AVX2-NEXT: vmovaps %ymm2, 1696(%rdx)
7266 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
7267 ; AVX2-NEXT: vmovaps %ymm2, 1664(%rdx)
7268 ; AVX2-NEXT: vmovaps %ymm0, 1504(%rdx)
7269 ; AVX2-NEXT: vmovaps %ymm1, 1472(%rdx)
7270 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7271 ; AVX2-NEXT: vmovaps %ymm0, 1440(%rdx)
7272 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7273 ; AVX2-NEXT: vmovaps %ymm0, 1408(%rdx)
7274 ; AVX2-NEXT: vmovaps %ymm3, 1248(%rdx)
7275 ; AVX2-NEXT: vmovaps %ymm5, 1216(%rdx)
7276 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7277 ; AVX2-NEXT: vmovaps %ymm0, 1184(%rdx)
7278 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7279 ; AVX2-NEXT: vmovaps %ymm0, 1152(%rdx)
7280 ; AVX2-NEXT: vmovaps %ymm6, 992(%rdx)
7281 ; AVX2-NEXT: vmovaps %ymm7, 960(%rdx)
7282 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7283 ; AVX2-NEXT: vmovaps %ymm0, 928(%rdx)
7284 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7285 ; AVX2-NEXT: vmovaps %ymm0, 896(%rdx)
7286 ; AVX2-NEXT: vmovaps %ymm8, 736(%rdx)
7287 ; AVX2-NEXT: vmovaps %ymm9, 704(%rdx)
7288 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7289 ; AVX2-NEXT: vmovaps %ymm0, 672(%rdx)
7290 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7291 ; AVX2-NEXT: vmovaps %ymm0, 640(%rdx)
7292 ; AVX2-NEXT: vmovaps %ymm12, 480(%rdx)
7293 ; AVX2-NEXT: vmovaps %ymm14, 448(%rdx)
7294 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7295 ; AVX2-NEXT: vmovaps %ymm0, 416(%rdx)
7296 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7297 ; AVX2-NEXT: vmovaps %ymm0, 384(%rdx)
7298 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7299 ; AVX2-NEXT: vmovaps %ymm0, 224(%rdx)
7300 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7301 ; AVX2-NEXT: vmovaps %ymm0, 192(%rdx)
7302 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7303 ; AVX2-NEXT: vmovaps %ymm0, 160(%rdx)
7304 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7305 ; AVX2-NEXT: vmovaps %ymm0, 128(%rdx)
7306 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7307 ; AVX2-NEXT: vmovaps %ymm0, 1888(%rdx)
7308 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7309 ; AVX2-NEXT: vmovaps %ymm0, 1856(%rdx)
7310 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7311 ; AVX2-NEXT: vmovaps %ymm0, 1824(%rdx)
7312 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7313 ; AVX2-NEXT: vmovaps %ymm0, 1792(%rdx)
7314 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7315 ; AVX2-NEXT: vmovaps %ymm0, 1632(%rdx)
7316 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7317 ; AVX2-NEXT: vmovaps %ymm0, 1600(%rdx)
7318 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7319 ; AVX2-NEXT: vmovaps %ymm0, 1568(%rdx)
7320 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7321 ; AVX2-NEXT: vmovaps %ymm0, 1536(%rdx)
7322 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7323 ; AVX2-NEXT: vmovaps %ymm0, 1376(%rdx)
7324 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7325 ; AVX2-NEXT: vmovaps %ymm0, 1344(%rdx)
7326 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7327 ; AVX2-NEXT: vmovaps %ymm0, 1312(%rdx)
7328 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7329 ; AVX2-NEXT: vmovaps %ymm0, 1280(%rdx)
7330 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7331 ; AVX2-NEXT: vmovaps %ymm0, 1120(%rdx)
7332 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7333 ; AVX2-NEXT: vmovaps %ymm0, 1088(%rdx)
7334 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7335 ; AVX2-NEXT: vmovaps %ymm0, 1056(%rdx)
7336 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7337 ; AVX2-NEXT: vmovaps %ymm0, 1024(%rdx)
7338 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7339 ; AVX2-NEXT: vmovaps %ymm0, 864(%rdx)
7340 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7341 ; AVX2-NEXT: vmovaps %ymm0, 832(%rdx)
7342 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7343 ; AVX2-NEXT: vmovaps %ymm0, 800(%rdx)
7344 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7345 ; AVX2-NEXT: vmovaps %ymm0, 768(%rdx)
7346 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7347 ; AVX2-NEXT: vmovaps %ymm0, 608(%rdx)
7348 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7349 ; AVX2-NEXT: vmovaps %ymm0, 576(%rdx)
7350 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7351 ; AVX2-NEXT: vmovaps %ymm0, 544(%rdx)
7352 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7353 ; AVX2-NEXT: vmovaps %ymm0, 512(%rdx)
7354 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7355 ; AVX2-NEXT: vmovaps %ymm0, 352(%rdx)
7356 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7357 ; AVX2-NEXT: vmovaps %ymm0, 320(%rdx)
7358 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7359 ; AVX2-NEXT: vmovaps %ymm0, 288(%rdx)
7360 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7361 ; AVX2-NEXT: vmovaps %ymm0, 256(%rdx)
7362 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7363 ; AVX2-NEXT: vmovaps %ymm0, 96(%rdx)
7364 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7365 ; AVX2-NEXT: vmovaps %ymm0, 64(%rdx)
7366 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7367 ; AVX2-NEXT: vmovaps %ymm0, 32(%rdx)
7368 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7369 ; AVX2-NEXT: vmovaps %ymm0, (%rdx)
7370 ; AVX2-NEXT: addq $1704, %rsp # imm = 0x6A8
7371 ; AVX2-NEXT: vzeroupper
7372 ; AVX2-NEXT: retq
7373 ;
7374 ; AVX2-FP-LABEL: store_i64_stride8_vf32:
7375 ; AVX2-FP: # %bb.0:
7376 ; AVX2-FP-NEXT: subq $1704, %rsp # imm = 0x6A8
7377 ; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %r10
7378 ; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %rax
7379 ; AVX2-FP-NEXT: vmovaps (%rcx), %xmm0
7380 ; AVX2-FP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7381 ; AVX2-FP-NEXT: vmovaps 32(%rcx), %xmm3
7382 ; AVX2-FP-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7383 ; AVX2-FP-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0
7384 ; AVX2-FP-NEXT: vmovaps (%rsi), %xmm2
7385 ; AVX2-FP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7386 ; AVX2-FP-NEXT: vmovaps 32(%rsi), %xmm4
7387 ; AVX2-FP-NEXT: vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7388 ; AVX2-FP-NEXT: vmovaps (%rdi), %xmm1
7389 ; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7390 ; AVX2-FP-NEXT: vmovaps 32(%rdi), %xmm5
7391 ; AVX2-FP-NEXT: vmovaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7392 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm2[1]
7393 ; AVX2-FP-NEXT: vbroadcastsd 8(%rdx), %ymm2
7394 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
7395 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
7396 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7397 ; AVX2-FP-NEXT: vmovaps (%rax), %xmm0
7398 ; AVX2-FP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7399 ; AVX2-FP-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0
7400 ; AVX2-FP-NEXT: vmovaps (%r9), %xmm2
7401 ; AVX2-FP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7402 ; AVX2-FP-NEXT: vmovaps 32(%r9), %xmm6
7403 ; AVX2-FP-NEXT: vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7404 ; AVX2-FP-NEXT: vmovaps (%r8), %xmm1
7405 ; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7406 ; AVX2-FP-NEXT: vmovaps 32(%r8), %xmm7
7407 ; AVX2-FP-NEXT: vmovaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7408 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm2[1]
7409 ; AVX2-FP-NEXT: vbroadcastsd 8(%r10), %ymm2
7410 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
7411 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
7412 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7413 ; AVX2-FP-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm0
7414 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm5[1],xmm4[1]
7415 ; AVX2-FP-NEXT: vbroadcastsd 40(%rdx), %ymm2
7416 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
7417 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
7418 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7419 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm7[1],xmm6[1]
7420 ; AVX2-FP-NEXT: vbroadcastsd 40(%r10), %ymm1
7421 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
7422 ; AVX2-FP-NEXT: vmovaps 32(%rax), %xmm1
7423 ; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7424 ; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
7425 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
7426 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7427 ; AVX2-FP-NEXT: vmovaps 64(%rsi), %xmm1
7428 ; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7429 ; AVX2-FP-NEXT: vmovaps 64(%rdi), %xmm0
7430 ; AVX2-FP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7431 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
7432 ; AVX2-FP-NEXT: vbroadcastsd 72(%rdx), %ymm1
7433 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
7434 ; AVX2-FP-NEXT: vmovaps 64(%rcx), %xmm1
7435 ; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7436 ; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
7437 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
7438 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7439 ; AVX2-FP-NEXT: vmovaps 64(%r9), %xmm1
7440 ; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7441 ; AVX2-FP-NEXT: vmovaps 64(%r8), %xmm0
7442 ; AVX2-FP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7443 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
7444 ; AVX2-FP-NEXT: vbroadcastsd 72(%r10), %ymm1
7445 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
7446 ; AVX2-FP-NEXT: vmovaps 64(%rax), %xmm1
7447 ; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7448 ; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
7449 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
7450 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7451 ; AVX2-FP-NEXT: vmovaps 96(%rsi), %xmm1
7452 ; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7453 ; AVX2-FP-NEXT: vmovaps 96(%rdi), %xmm0
7454 ; AVX2-FP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7455 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
7456 ; AVX2-FP-NEXT: vbroadcastsd 104(%rdx), %ymm1
7457 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
7458 ; AVX2-FP-NEXT: vmovaps 96(%rcx), %xmm1
7459 ; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7460 ; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
7461 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
7462 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7463 ; AVX2-FP-NEXT: vmovaps 96(%r9), %xmm1
7464 ; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7465 ; AVX2-FP-NEXT: vmovaps 96(%r8), %xmm0
7466 ; AVX2-FP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7467 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
7468 ; AVX2-FP-NEXT: vbroadcastsd 104(%r10), %ymm1
7469 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
7470 ; AVX2-FP-NEXT: vmovaps 96(%rax), %xmm1
7471 ; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7472 ; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
7473 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
7474 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7475 ; AVX2-FP-NEXT: vmovaps 128(%rsi), %xmm1
7476 ; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7477 ; AVX2-FP-NEXT: vmovaps 128(%rdi), %xmm0
7478 ; AVX2-FP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7479 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
7480 ; AVX2-FP-NEXT: vbroadcastsd 136(%rdx), %ymm1
7481 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
7482 ; AVX2-FP-NEXT: vmovaps 128(%rcx), %xmm1
7483 ; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7484 ; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
7485 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
7486 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7487 ; AVX2-FP-NEXT: vmovaps 128(%r9), %xmm1
7488 ; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7489 ; AVX2-FP-NEXT: vmovaps 128(%r8), %xmm0
7490 ; AVX2-FP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7491 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
7492 ; AVX2-FP-NEXT: vbroadcastsd 136(%r10), %ymm1
7493 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
7494 ; AVX2-FP-NEXT: vmovaps 128(%rax), %xmm1
7495 ; AVX2-FP-NEXT: vmovaps %xmm1, (%rsp) # 16-byte Spill
7496 ; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
7497 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
7498 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7499 ; AVX2-FP-NEXT: vmovaps 160(%rsi), %xmm1
7500 ; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7501 ; AVX2-FP-NEXT: vmovaps 160(%rdi), %xmm0
7502 ; AVX2-FP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7503 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
7504 ; AVX2-FP-NEXT: vbroadcastsd 168(%rdx), %ymm1
7505 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
7506 ; AVX2-FP-NEXT: vmovaps 160(%rcx), %xmm1
7507 ; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7508 ; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
7509 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
7510 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7511 ; AVX2-FP-NEXT: vmovaps 160(%r9), %xmm0
7512 ; AVX2-FP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7513 ; AVX2-FP-NEXT: vmovaps 160(%r8), %xmm13
7514 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm13[1],xmm0[1]
7515 ; AVX2-FP-NEXT: vbroadcastsd 168(%r10), %ymm1
7516 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
7517 ; AVX2-FP-NEXT: vmovaps 160(%rax), %xmm12
7518 ; AVX2-FP-NEXT: vinsertf128 $1, %xmm12, %ymm0, %ymm1
7519 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
7520 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7521 ; AVX2-FP-NEXT: vmovaps 192(%rsi), %xmm11
7522 ; AVX2-FP-NEXT: vmovaps 192(%rdi), %xmm10
7523 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm10[1],xmm11[1]
7524 ; AVX2-FP-NEXT: vbroadcastsd 200(%rdx), %ymm1
7525 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
7526 ; AVX2-FP-NEXT: vmovaps 192(%rcx), %xmm9
7527 ; AVX2-FP-NEXT: vinsertf128 $1, %xmm9, %ymm0, %ymm1
7528 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
7529 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7530 ; AVX2-FP-NEXT: vmovaps 192(%r9), %xmm8
7531 ; AVX2-FP-NEXT: vmovaps 192(%r8), %xmm7
7532 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm7[1],xmm8[1]
7533 ; AVX2-FP-NEXT: vbroadcastsd 200(%r10), %ymm1
7534 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
7535 ; AVX2-FP-NEXT: vmovaps 192(%rax), %xmm6
7536 ; AVX2-FP-NEXT: vinsertf128 $1, %xmm6, %ymm0, %ymm1
7537 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
7538 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7539 ; AVX2-FP-NEXT: vmovaps 224(%rsi), %xmm5
7540 ; AVX2-FP-NEXT: vmovaps 224(%rdi), %xmm4
7541 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm4[1],xmm5[1]
7542 ; AVX2-FP-NEXT: vbroadcastsd 232(%rdx), %ymm1
7543 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
7544 ; AVX2-FP-NEXT: vmovaps 224(%rcx), %xmm3
7545 ; AVX2-FP-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm1
7546 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
7547 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7548 ; AVX2-FP-NEXT: vmovaps 224(%r9), %xmm2
7549 ; AVX2-FP-NEXT: vmovaps 224(%r8), %xmm1
7550 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm2[1]
7551 ; AVX2-FP-NEXT: vbroadcastsd 232(%r10), %ymm15
7552 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm15 = ymm0[0,1,2,3],ymm15[4,5,6,7]
7553 ; AVX2-FP-NEXT: vmovaps 224(%rax), %xmm0
7554 ; AVX2-FP-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm14
7555 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm14 = ymm15[0,1,2,3,4,5],ymm14[6,7]
7556 ; AVX2-FP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7557 ; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
7558 ; AVX2-FP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
7559 ; AVX2-FP-NEXT: # xmm14 = xmm14[0],mem[0]
7560 ; AVX2-FP-NEXT: vinsertf128 $1, (%rdx), %ymm14, %ymm14
7561 ; AVX2-FP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
7562 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
7563 ; AVX2-FP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7564 ; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
7565 ; AVX2-FP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
7566 ; AVX2-FP-NEXT: # xmm14 = xmm14[0],mem[0]
7567 ; AVX2-FP-NEXT: vinsertf128 $1, (%r10), %ymm14, %ymm14
7568 ; AVX2-FP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
7569 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
7570 ; AVX2-FP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7571 ; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
7572 ; AVX2-FP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
7573 ; AVX2-FP-NEXT: # xmm14 = xmm14[0],mem[0]
7574 ; AVX2-FP-NEXT: vinsertf128 $1, 32(%rdx), %ymm14, %ymm14
7575 ; AVX2-FP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
7576 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
7577 ; AVX2-FP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7578 ; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
7579 ; AVX2-FP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
7580 ; AVX2-FP-NEXT: # xmm14 = xmm14[0],mem[0]
7581 ; AVX2-FP-NEXT: vinsertf128 $1, 32(%r10), %ymm14, %ymm14
7582 ; AVX2-FP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
7583 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
7584 ; AVX2-FP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7585 ; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
7586 ; AVX2-FP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
7587 ; AVX2-FP-NEXT: # xmm14 = xmm14[0],mem[0]
7588 ; AVX2-FP-NEXT: vinsertf128 $1, 64(%rdx), %ymm14, %ymm14
7589 ; AVX2-FP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
7590 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
7591 ; AVX2-FP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7592 ; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
7593 ; AVX2-FP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
7594 ; AVX2-FP-NEXT: # xmm14 = xmm14[0],mem[0]
7595 ; AVX2-FP-NEXT: vinsertf128 $1, 64(%r10), %ymm14, %ymm14
7596 ; AVX2-FP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
7597 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
7598 ; AVX2-FP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7599 ; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
7600 ; AVX2-FP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
7601 ; AVX2-FP-NEXT: # xmm14 = xmm14[0],mem[0]
7602 ; AVX2-FP-NEXT: vinsertf128 $1, 96(%rdx), %ymm14, %ymm14
7603 ; AVX2-FP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
7604 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
7605 ; AVX2-FP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7606 ; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
7607 ; AVX2-FP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
7608 ; AVX2-FP-NEXT: # xmm14 = xmm14[0],mem[0]
7609 ; AVX2-FP-NEXT: vinsertf128 $1, 96(%r10), %ymm14, %ymm14
7610 ; AVX2-FP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
7611 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
7612 ; AVX2-FP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7613 ; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
7614 ; AVX2-FP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
7615 ; AVX2-FP-NEXT: # xmm14 = xmm14[0],mem[0]
7616 ; AVX2-FP-NEXT: vinsertf128 $1, 128(%rdx), %ymm14, %ymm14
7617 ; AVX2-FP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
7618 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
7619 ; AVX2-FP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7620 ; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
7621 ; AVX2-FP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
7622 ; AVX2-FP-NEXT: # xmm14 = xmm14[0],mem[0]
7623 ; AVX2-FP-NEXT: vinsertf128 $1, 128(%r10), %ymm14, %ymm14
7624 ; AVX2-FP-NEXT: vbroadcastsd (%rsp), %ymm15 # 16-byte Folded Reload
7625 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
7626 ; AVX2-FP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7627 ; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
7628 ; AVX2-FP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
7629 ; AVX2-FP-NEXT: # xmm14 = xmm14[0],mem[0]
7630 ; AVX2-FP-NEXT: vinsertf128 $1, 160(%rdx), %ymm14, %ymm14
7631 ; AVX2-FP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
7632 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
7633 ; AVX2-FP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7634 ; AVX2-FP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm13, %xmm13 # 16-byte Folded Reload
7635 ; AVX2-FP-NEXT: # xmm13 = xmm13[0],mem[0]
7636 ; AVX2-FP-NEXT: vinsertf128 $1, 160(%r10), %ymm13, %ymm13
7637 ; AVX2-FP-NEXT: vbroadcastsd %xmm12, %ymm12
7638 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm12 = ymm13[0,1,2,3,4,5],ymm12[6,7]
7639 ; AVX2-FP-NEXT: vmovups %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7640 ; AVX2-FP-NEXT: vmovlhps {{.*#+}} xmm10 = xmm10[0],xmm11[0]
7641 ; AVX2-FP-NEXT: vinsertf128 $1, 192(%rdx), %ymm10, %ymm10
7642 ; AVX2-FP-NEXT: vbroadcastsd %xmm9, %ymm9
7643 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm9 = ymm10[0,1,2,3,4,5],ymm9[6,7]
7644 ; AVX2-FP-NEXT: vmovups %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7645 ; AVX2-FP-NEXT: vmovlhps {{.*#+}} xmm7 = xmm7[0],xmm8[0]
7646 ; AVX2-FP-NEXT: vinsertf128 $1, 192(%r10), %ymm7, %ymm7
7647 ; AVX2-FP-NEXT: vbroadcastsd %xmm6, %ymm6
7648 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3,4,5],ymm6[6,7]
7649 ; AVX2-FP-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7650 ; AVX2-FP-NEXT: vmovlhps {{.*#+}} xmm4 = xmm4[0],xmm5[0]
7651 ; AVX2-FP-NEXT: vinsertf128 $1, 224(%rdx), %ymm4, %ymm4
7652 ; AVX2-FP-NEXT: vbroadcastsd %xmm3, %ymm3
7653 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm4[0,1,2,3,4,5],ymm3[6,7]
7654 ; AVX2-FP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7655 ; AVX2-FP-NEXT: vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
7656 ; AVX2-FP-NEXT: vinsertf128 $1, 224(%r10), %ymm1, %ymm1
7657 ; AVX2-FP-NEXT: vbroadcastsd %xmm0, %ymm0
7658 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
7659 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7660 ; AVX2-FP-NEXT: vmovaps (%rdi), %ymm0
7661 ; AVX2-FP-NEXT: vmovaps (%rsi), %ymm1
7662 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
7663 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
7664 ; AVX2-FP-NEXT: vbroadcastsd 16(%rcx), %ymm3
7665 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
7666 ; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7667 ; AVX2-FP-NEXT: vmovaps (%r8), %ymm2
7668 ; AVX2-FP-NEXT: vmovaps (%r9), %ymm3
7669 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
7670 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],mem[2,3]
7671 ; AVX2-FP-NEXT: vbroadcastsd 16(%rax), %ymm5
7672 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
7673 ; AVX2-FP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7674 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
7675 ; AVX2-FP-NEXT: vbroadcastsd 24(%rdx), %ymm1
7676 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm8 = ymm0[2,3],ymm1[2,3]
7677 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
7678 ; AVX2-FP-NEXT: vbroadcastsd 24(%r10), %ymm1
7679 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm9 = ymm0[2,3],ymm1[2,3]
7680 ; AVX2-FP-NEXT: vmovaps 32(%rdi), %ymm0
7681 ; AVX2-FP-NEXT: vmovaps 32(%rsi), %ymm1
7682 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
7683 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
7684 ; AVX2-FP-NEXT: vbroadcastsd 48(%rcx), %ymm3
7685 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
7686 ; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7687 ; AVX2-FP-NEXT: vmovaps 32(%r8), %ymm2
7688 ; AVX2-FP-NEXT: vmovaps 32(%r9), %ymm3
7689 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
7690 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],mem[2,3]
7691 ; AVX2-FP-NEXT: vbroadcastsd 48(%rax), %ymm5
7692 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
7693 ; AVX2-FP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7694 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
7695 ; AVX2-FP-NEXT: vbroadcastsd 56(%rdx), %ymm1
7696 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm12 = ymm0[2,3],ymm1[2,3]
7697 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
7698 ; AVX2-FP-NEXT: vbroadcastsd 56(%r10), %ymm1
7699 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
7700 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7701 ; AVX2-FP-NEXT: vmovaps 64(%rdi), %ymm0
7702 ; AVX2-FP-NEXT: vmovaps 64(%rsi), %ymm1
7703 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
7704 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
7705 ; AVX2-FP-NEXT: vbroadcastsd 80(%rcx), %ymm3
7706 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
7707 ; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7708 ; AVX2-FP-NEXT: vmovaps 64(%r8), %ymm2
7709 ; AVX2-FP-NEXT: vmovaps 64(%r9), %ymm3
7710 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
7711 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],mem[2,3]
7712 ; AVX2-FP-NEXT: vbroadcastsd 80(%rax), %ymm5
7713 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
7714 ; AVX2-FP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7715 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
7716 ; AVX2-FP-NEXT: vbroadcastsd 88(%rdx), %ymm1
7717 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
7718 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7719 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
7720 ; AVX2-FP-NEXT: vbroadcastsd 88(%r10), %ymm1
7721 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
7722 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7723 ; AVX2-FP-NEXT: vmovaps 96(%rdi), %ymm0
7724 ; AVX2-FP-NEXT: vmovaps 96(%rsi), %ymm1
7725 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
7726 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
7727 ; AVX2-FP-NEXT: vbroadcastsd 112(%rcx), %ymm3
7728 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
7729 ; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7730 ; AVX2-FP-NEXT: vmovaps 96(%r8), %ymm2
7731 ; AVX2-FP-NEXT: vmovaps 96(%r9), %ymm3
7732 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
7733 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],mem[2,3]
7734 ; AVX2-FP-NEXT: vbroadcastsd 112(%rax), %ymm5
7735 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
7736 ; AVX2-FP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7737 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
7738 ; AVX2-FP-NEXT: vbroadcastsd 120(%rdx), %ymm1
7739 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
7740 ; AVX2-FP-NEXT: vmovups %ymm0, (%rsp) # 32-byte Spill
7741 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
7742 ; AVX2-FP-NEXT: vbroadcastsd 120(%r10), %ymm1
7743 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
7744 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7745 ; AVX2-FP-NEXT: vmovaps 128(%rdi), %ymm0
7746 ; AVX2-FP-NEXT: vmovaps 128(%rsi), %ymm1
7747 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
7748 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
7749 ; AVX2-FP-NEXT: vbroadcastsd 144(%rcx), %ymm3
7750 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
7751 ; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7752 ; AVX2-FP-NEXT: vmovaps 128(%r8), %ymm2
7753 ; AVX2-FP-NEXT: vmovaps 128(%r9), %ymm3
7754 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
7755 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],mem[2,3]
7756 ; AVX2-FP-NEXT: vbroadcastsd 144(%rax), %ymm5
7757 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
7758 ; AVX2-FP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7759 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
7760 ; AVX2-FP-NEXT: vbroadcastsd 152(%rdx), %ymm1
7761 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
7762 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7763 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
7764 ; AVX2-FP-NEXT: vbroadcastsd 152(%r10), %ymm1
7765 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
7766 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7767 ; AVX2-FP-NEXT: vmovaps 160(%rdi), %ymm0
7768 ; AVX2-FP-NEXT: vmovaps 160(%rsi), %ymm1
7769 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
7770 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
7771 ; AVX2-FP-NEXT: vbroadcastsd 176(%rcx), %ymm3
7772 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
7773 ; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7774 ; AVX2-FP-NEXT: vmovaps 160(%r8), %ymm2
7775 ; AVX2-FP-NEXT: vmovaps 160(%r9), %ymm3
7776 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
7777 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],mem[2,3]
7778 ; AVX2-FP-NEXT: vbroadcastsd 176(%rax), %ymm5
7779 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
7780 ; AVX2-FP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7781 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
7782 ; AVX2-FP-NEXT: vbroadcastsd 184(%rdx), %ymm1
7783 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
7784 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7785 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
7786 ; AVX2-FP-NEXT: vbroadcastsd 184(%r10), %ymm1
7787 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
7788 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7789 ; AVX2-FP-NEXT: vmovaps 192(%rdi), %ymm0
7790 ; AVX2-FP-NEXT: vmovaps 192(%rsi), %ymm1
7791 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
7792 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
7793 ; AVX2-FP-NEXT: vbroadcastsd 208(%rcx), %ymm3
7794 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
7795 ; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7796 ; AVX2-FP-NEXT: vmovaps 192(%r8), %ymm2
7797 ; AVX2-FP-NEXT: vmovaps 192(%r9), %ymm3
7798 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
7799 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],mem[2,3]
7800 ; AVX2-FP-NEXT: vbroadcastsd 208(%rax), %ymm5
7801 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
7802 ; AVX2-FP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7803 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
7804 ; AVX2-FP-NEXT: vbroadcastsd 216(%rdx), %ymm1
7805 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm13 = ymm0[2,3],ymm1[2,3]
7806 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
7807 ; AVX2-FP-NEXT: vbroadcastsd 216(%r10), %ymm1
7808 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm11 = ymm0[2,3],ymm1[2,3]
7809 ; AVX2-FP-NEXT: vmovaps 224(%rdi), %ymm0
7810 ; AVX2-FP-NEXT: vmovaps 224(%rsi), %ymm3
7811 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm0[0],ymm3[0],ymm0[2],ymm3[2]
7812 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm1[2,3],mem[2,3]
7813 ; AVX2-FP-NEXT: vbroadcastsd 240(%rcx), %ymm5
7814 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm10 = ymm1[0,1,2,3,4,5],ymm5[6,7]
7815 ; AVX2-FP-NEXT: vmovaps 224(%r8), %ymm5
7816 ; AVX2-FP-NEXT: vmovaps 224(%r9), %ymm6
7817 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm7 = ymm5[0],ymm6[0],ymm5[2],ymm6[2]
7818 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm7 = ymm7[2,3],mem[2,3]
7819 ; AVX2-FP-NEXT: vbroadcastsd 240(%rax), %ymm15
7820 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm15 = ymm7[0,1,2,3,4,5],ymm15[6,7]
7821 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm3[1],ymm0[3],ymm3[3]
7822 ; AVX2-FP-NEXT: vbroadcastsd 248(%rdx), %ymm3
7823 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm0[2,3],ymm3[2,3]
7824 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm1 = ymm5[1],ymm6[1],ymm5[3],ymm6[3]
7825 ; AVX2-FP-NEXT: vbroadcastsd 248(%r10), %ymm3
7826 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm1[2,3],ymm3[2,3]
7827 ; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %rdx
7828 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm8[0,1,2,3,4,5],mem[6,7]
7829 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7830 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm9[0,1,2,3,4,5],mem[6,7]
7831 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7832 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm14 = ymm12[0,1,2,3,4,5],mem[6,7]
7833 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7834 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm12 = ymm0[0,1,2,3,4,5],mem[6,7]
7835 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7836 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm9 = ymm0[0,1,2,3,4,5],mem[6,7]
7837 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7838 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm8 = ymm0[0,1,2,3,4,5],mem[6,7]
7839 ; AVX2-FP-NEXT: vmovups (%rsp), %ymm0 # 32-byte Reload
7840 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm7 = ymm0[0,1,2,3,4,5],mem[6,7]
7841 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7842 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm6 = ymm0[0,1,2,3,4,5],mem[6,7]
7843 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7844 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm5 = ymm0[0,1,2,3,4,5],mem[6,7]
7845 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7846 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm0[0,1,2,3,4,5],mem[6,7]
7847 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7848 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm0[0,1,2,3,4,5],mem[6,7]
7849 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7850 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],mem[6,7]
7851 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm13 = ymm13[0,1,2,3,4,5],mem[6,7]
7852 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm11 = ymm11[0,1,2,3,4,5],mem[6,7]
7853 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],mem[6,7]
7854 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],mem[6,7]
7855 ; AVX2-FP-NEXT: vmovaps %ymm2, 2016(%rdx)
7856 ; AVX2-FP-NEXT: vmovaps %ymm4, 1984(%rdx)
7857 ; AVX2-FP-NEXT: vmovaps %ymm15, 1952(%rdx)
7858 ; AVX2-FP-NEXT: vmovaps %ymm10, 1920(%rdx)
7859 ; AVX2-FP-NEXT: vmovaps %ymm11, 1760(%rdx)
7860 ; AVX2-FP-NEXT: vmovaps %ymm13, 1728(%rdx)
7861 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
7862 ; AVX2-FP-NEXT: vmovaps %ymm2, 1696(%rdx)
7863 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
7864 ; AVX2-FP-NEXT: vmovaps %ymm2, 1664(%rdx)
7865 ; AVX2-FP-NEXT: vmovaps %ymm0, 1504(%rdx)
7866 ; AVX2-FP-NEXT: vmovaps %ymm1, 1472(%rdx)
7867 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7868 ; AVX2-FP-NEXT: vmovaps %ymm0, 1440(%rdx)
7869 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7870 ; AVX2-FP-NEXT: vmovaps %ymm0, 1408(%rdx)
7871 ; AVX2-FP-NEXT: vmovaps %ymm3, 1248(%rdx)
7872 ; AVX2-FP-NEXT: vmovaps %ymm5, 1216(%rdx)
7873 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7874 ; AVX2-FP-NEXT: vmovaps %ymm0, 1184(%rdx)
7875 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7876 ; AVX2-FP-NEXT: vmovaps %ymm0, 1152(%rdx)
7877 ; AVX2-FP-NEXT: vmovaps %ymm6, 992(%rdx)
7878 ; AVX2-FP-NEXT: vmovaps %ymm7, 960(%rdx)
7879 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7880 ; AVX2-FP-NEXT: vmovaps %ymm0, 928(%rdx)
7881 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7882 ; AVX2-FP-NEXT: vmovaps %ymm0, 896(%rdx)
7883 ; AVX2-FP-NEXT: vmovaps %ymm8, 736(%rdx)
7884 ; AVX2-FP-NEXT: vmovaps %ymm9, 704(%rdx)
7885 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7886 ; AVX2-FP-NEXT: vmovaps %ymm0, 672(%rdx)
7887 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7888 ; AVX2-FP-NEXT: vmovaps %ymm0, 640(%rdx)
7889 ; AVX2-FP-NEXT: vmovaps %ymm12, 480(%rdx)
7890 ; AVX2-FP-NEXT: vmovaps %ymm14, 448(%rdx)
7891 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7892 ; AVX2-FP-NEXT: vmovaps %ymm0, 416(%rdx)
7893 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7894 ; AVX2-FP-NEXT: vmovaps %ymm0, 384(%rdx)
7895 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7896 ; AVX2-FP-NEXT: vmovaps %ymm0, 224(%rdx)
7897 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7898 ; AVX2-FP-NEXT: vmovaps %ymm0, 192(%rdx)
7899 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7900 ; AVX2-FP-NEXT: vmovaps %ymm0, 160(%rdx)
7901 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7902 ; AVX2-FP-NEXT: vmovaps %ymm0, 128(%rdx)
7903 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7904 ; AVX2-FP-NEXT: vmovaps %ymm0, 1888(%rdx)
7905 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7906 ; AVX2-FP-NEXT: vmovaps %ymm0, 1856(%rdx)
7907 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7908 ; AVX2-FP-NEXT: vmovaps %ymm0, 1824(%rdx)
7909 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7910 ; AVX2-FP-NEXT: vmovaps %ymm0, 1792(%rdx)
7911 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7912 ; AVX2-FP-NEXT: vmovaps %ymm0, 1632(%rdx)
7913 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7914 ; AVX2-FP-NEXT: vmovaps %ymm0, 1600(%rdx)
7915 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7916 ; AVX2-FP-NEXT: vmovaps %ymm0, 1568(%rdx)
7917 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7918 ; AVX2-FP-NEXT: vmovaps %ymm0, 1536(%rdx)
7919 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7920 ; AVX2-FP-NEXT: vmovaps %ymm0, 1376(%rdx)
7921 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7922 ; AVX2-FP-NEXT: vmovaps %ymm0, 1344(%rdx)
7923 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7924 ; AVX2-FP-NEXT: vmovaps %ymm0, 1312(%rdx)
7925 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7926 ; AVX2-FP-NEXT: vmovaps %ymm0, 1280(%rdx)
7927 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7928 ; AVX2-FP-NEXT: vmovaps %ymm0, 1120(%rdx)
7929 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7930 ; AVX2-FP-NEXT: vmovaps %ymm0, 1088(%rdx)
7931 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7932 ; AVX2-FP-NEXT: vmovaps %ymm0, 1056(%rdx)
7933 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7934 ; AVX2-FP-NEXT: vmovaps %ymm0, 1024(%rdx)
7935 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7936 ; AVX2-FP-NEXT: vmovaps %ymm0, 864(%rdx)
7937 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7938 ; AVX2-FP-NEXT: vmovaps %ymm0, 832(%rdx)
7939 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7940 ; AVX2-FP-NEXT: vmovaps %ymm0, 800(%rdx)
7941 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7942 ; AVX2-FP-NEXT: vmovaps %ymm0, 768(%rdx)
7943 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7944 ; AVX2-FP-NEXT: vmovaps %ymm0, 608(%rdx)
7945 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7946 ; AVX2-FP-NEXT: vmovaps %ymm0, 576(%rdx)
7947 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7948 ; AVX2-FP-NEXT: vmovaps %ymm0, 544(%rdx)
7949 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7950 ; AVX2-FP-NEXT: vmovaps %ymm0, 512(%rdx)
7951 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7952 ; AVX2-FP-NEXT: vmovaps %ymm0, 352(%rdx)
7953 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7954 ; AVX2-FP-NEXT: vmovaps %ymm0, 320(%rdx)
7955 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7956 ; AVX2-FP-NEXT: vmovaps %ymm0, 288(%rdx)
7957 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7958 ; AVX2-FP-NEXT: vmovaps %ymm0, 256(%rdx)
7959 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7960 ; AVX2-FP-NEXT: vmovaps %ymm0, 96(%rdx)
7961 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7962 ; AVX2-FP-NEXT: vmovaps %ymm0, 64(%rdx)
7963 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7964 ; AVX2-FP-NEXT: vmovaps %ymm0, 32(%rdx)
7965 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
7966 ; AVX2-FP-NEXT: vmovaps %ymm0, (%rdx)
7967 ; AVX2-FP-NEXT: addq $1704, %rsp # imm = 0x6A8
7968 ; AVX2-FP-NEXT: vzeroupper
7969 ; AVX2-FP-NEXT: retq
7970 ;
7971 ; AVX2-FCP-LABEL: store_i64_stride8_vf32:
7972 ; AVX2-FCP: # %bb.0:
7973 ; AVX2-FCP-NEXT: subq $1704, %rsp # imm = 0x6A8
7974 ; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10
7975 ; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
7976 ; AVX2-FCP-NEXT: vmovaps (%rcx), %xmm0
7977 ; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7978 ; AVX2-FCP-NEXT: vmovaps 32(%rcx), %xmm3
7979 ; AVX2-FCP-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7980 ; AVX2-FCP-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0
7981 ; AVX2-FCP-NEXT: vmovaps (%rsi), %xmm2
7982 ; AVX2-FCP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7983 ; AVX2-FCP-NEXT: vmovaps 32(%rsi), %xmm4
7984 ; AVX2-FCP-NEXT: vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7985 ; AVX2-FCP-NEXT: vmovaps (%rdi), %xmm1
7986 ; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7987 ; AVX2-FCP-NEXT: vmovaps 32(%rdi), %xmm5
7988 ; AVX2-FCP-NEXT: vmovaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7989 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm2[1]
7990 ; AVX2-FCP-NEXT: vbroadcastsd 8(%rdx), %ymm2
7991 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
7992 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
7993 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
7994 ; AVX2-FCP-NEXT: vmovaps (%rax), %xmm0
7995 ; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7996 ; AVX2-FCP-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0
7997 ; AVX2-FCP-NEXT: vmovaps (%r9), %xmm2
7998 ; AVX2-FCP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
7999 ; AVX2-FCP-NEXT: vmovaps 32(%r9), %xmm6
8000 ; AVX2-FCP-NEXT: vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8001 ; AVX2-FCP-NEXT: vmovaps (%r8), %xmm1
8002 ; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8003 ; AVX2-FCP-NEXT: vmovaps 32(%r8), %xmm7
8004 ; AVX2-FCP-NEXT: vmovaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8005 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm2[1]
8006 ; AVX2-FCP-NEXT: vbroadcastsd 8(%r10), %ymm2
8007 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
8008 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
8009 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
8010 ; AVX2-FCP-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm0
8011 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm5[1],xmm4[1]
8012 ; AVX2-FCP-NEXT: vbroadcastsd 40(%rdx), %ymm2
8013 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
8014 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
8015 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
8016 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm7[1],xmm6[1]
8017 ; AVX2-FCP-NEXT: vbroadcastsd 40(%r10), %ymm1
8018 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
8019 ; AVX2-FCP-NEXT: vmovaps 32(%rax), %xmm1
8020 ; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8021 ; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
8022 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
8023 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
8024 ; AVX2-FCP-NEXT: vmovaps 64(%rsi), %xmm1
8025 ; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8026 ; AVX2-FCP-NEXT: vmovaps 64(%rdi), %xmm0
8027 ; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8028 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
8029 ; AVX2-FCP-NEXT: vbroadcastsd 72(%rdx), %ymm1
8030 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
8031 ; AVX2-FCP-NEXT: vmovaps 64(%rcx), %xmm1
8032 ; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8033 ; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
8034 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
8035 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
8036 ; AVX2-FCP-NEXT: vmovaps 64(%r9), %xmm1
8037 ; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8038 ; AVX2-FCP-NEXT: vmovaps 64(%r8), %xmm0
8039 ; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8040 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
8041 ; AVX2-FCP-NEXT: vbroadcastsd 72(%r10), %ymm1
8042 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
8043 ; AVX2-FCP-NEXT: vmovaps 64(%rax), %xmm1
8044 ; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8045 ; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
8046 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
8047 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
8048 ; AVX2-FCP-NEXT: vmovaps 96(%rsi), %xmm1
8049 ; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8050 ; AVX2-FCP-NEXT: vmovaps 96(%rdi), %xmm0
8051 ; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8052 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
8053 ; AVX2-FCP-NEXT: vbroadcastsd 104(%rdx), %ymm1
8054 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
8055 ; AVX2-FCP-NEXT: vmovaps 96(%rcx), %xmm1
8056 ; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8057 ; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
8058 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
8059 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
8060 ; AVX2-FCP-NEXT: vmovaps 96(%r9), %xmm1
8061 ; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8062 ; AVX2-FCP-NEXT: vmovaps 96(%r8), %xmm0
8063 ; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8064 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
8065 ; AVX2-FCP-NEXT: vbroadcastsd 104(%r10), %ymm1
8066 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
8067 ; AVX2-FCP-NEXT: vmovaps 96(%rax), %xmm1
8068 ; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8069 ; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
8070 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
8071 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
8072 ; AVX2-FCP-NEXT: vmovaps 128(%rsi), %xmm1
8073 ; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8074 ; AVX2-FCP-NEXT: vmovaps 128(%rdi), %xmm0
8075 ; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8076 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
8077 ; AVX2-FCP-NEXT: vbroadcastsd 136(%rdx), %ymm1
8078 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
8079 ; AVX2-FCP-NEXT: vmovaps 128(%rcx), %xmm1
8080 ; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8081 ; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
8082 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
8083 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
8084 ; AVX2-FCP-NEXT: vmovaps 128(%r9), %xmm1
8085 ; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8086 ; AVX2-FCP-NEXT: vmovaps 128(%r8), %xmm0
8087 ; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8088 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
8089 ; AVX2-FCP-NEXT: vbroadcastsd 136(%r10), %ymm1
8090 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
8091 ; AVX2-FCP-NEXT: vmovaps 128(%rax), %xmm1
8092 ; AVX2-FCP-NEXT: vmovaps %xmm1, (%rsp) # 16-byte Spill
8093 ; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
8094 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
8095 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
8096 ; AVX2-FCP-NEXT: vmovaps 160(%rsi), %xmm1
8097 ; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8098 ; AVX2-FCP-NEXT: vmovaps 160(%rdi), %xmm0
8099 ; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8100 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
8101 ; AVX2-FCP-NEXT: vbroadcastsd 168(%rdx), %ymm1
8102 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
8103 ; AVX2-FCP-NEXT: vmovaps 160(%rcx), %xmm1
8104 ; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8105 ; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
8106 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
8107 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
8108 ; AVX2-FCP-NEXT: vmovaps 160(%r9), %xmm0
8109 ; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
8110 ; AVX2-FCP-NEXT: vmovaps 160(%r8), %xmm13
8111 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm13[1],xmm0[1]
8112 ; AVX2-FCP-NEXT: vbroadcastsd 168(%r10), %ymm1
8113 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
8114 ; AVX2-FCP-NEXT: vmovaps 160(%rax), %xmm12
8115 ; AVX2-FCP-NEXT: vinsertf128 $1, %xmm12, %ymm0, %ymm1
8116 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
8117 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
8118 ; AVX2-FCP-NEXT: vmovaps 192(%rsi), %xmm11
8119 ; AVX2-FCP-NEXT: vmovaps 192(%rdi), %xmm10
8120 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm10[1],xmm11[1]
8121 ; AVX2-FCP-NEXT: vbroadcastsd 200(%rdx), %ymm1
8122 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
8123 ; AVX2-FCP-NEXT: vmovaps 192(%rcx), %xmm9
8124 ; AVX2-FCP-NEXT: vinsertf128 $1, %xmm9, %ymm0, %ymm1
8125 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
8126 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
8127 ; AVX2-FCP-NEXT: vmovaps 192(%r9), %xmm8
8128 ; AVX2-FCP-NEXT: vmovaps 192(%r8), %xmm7
8129 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm7[1],xmm8[1]
8130 ; AVX2-FCP-NEXT: vbroadcastsd 200(%r10), %ymm1
8131 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
8132 ; AVX2-FCP-NEXT: vmovaps 192(%rax), %xmm6
8133 ; AVX2-FCP-NEXT: vinsertf128 $1, %xmm6, %ymm0, %ymm1
8134 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
8135 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
8136 ; AVX2-FCP-NEXT: vmovaps 224(%rsi), %xmm5
8137 ; AVX2-FCP-NEXT: vmovaps 224(%rdi), %xmm4
8138 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm4[1],xmm5[1]
8139 ; AVX2-FCP-NEXT: vbroadcastsd 232(%rdx), %ymm1
8140 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
8141 ; AVX2-FCP-NEXT: vmovaps 224(%rcx), %xmm3
8142 ; AVX2-FCP-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm1
8143 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
8144 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
8145 ; AVX2-FCP-NEXT: vmovaps 224(%r9), %xmm2
8146 ; AVX2-FCP-NEXT: vmovaps 224(%r8), %xmm1
8147 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm2[1]
8148 ; AVX2-FCP-NEXT: vbroadcastsd 232(%r10), %ymm15
8149 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm15 = ymm0[0,1,2,3],ymm15[4,5,6,7]
8150 ; AVX2-FCP-NEXT: vmovaps 224(%rax), %xmm0
8151 ; AVX2-FCP-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm14
8152 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm14 = ymm15[0,1,2,3,4,5],ymm14[6,7]
8153 ; AVX2-FCP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
8154 ; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
8155 ; AVX2-FCP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
8156 ; AVX2-FCP-NEXT: # xmm14 = xmm14[0],mem[0]
8157 ; AVX2-FCP-NEXT: vinsertf128 $1, (%rdx), %ymm14, %ymm14
8158 ; AVX2-FCP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
8159 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
8160 ; AVX2-FCP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
8161 ; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
8162 ; AVX2-FCP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
8163 ; AVX2-FCP-NEXT: # xmm14 = xmm14[0],mem[0]
8164 ; AVX2-FCP-NEXT: vinsertf128 $1, (%r10), %ymm14, %ymm14
8165 ; AVX2-FCP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
8166 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
8167 ; AVX2-FCP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
8168 ; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
8169 ; AVX2-FCP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
8170 ; AVX2-FCP-NEXT: # xmm14 = xmm14[0],mem[0]
8171 ; AVX2-FCP-NEXT: vinsertf128 $1, 32(%rdx), %ymm14, %ymm14
8172 ; AVX2-FCP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
8173 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
8174 ; AVX2-FCP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
8175 ; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
8176 ; AVX2-FCP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
8177 ; AVX2-FCP-NEXT: # xmm14 = xmm14[0],mem[0]
8178 ; AVX2-FCP-NEXT: vinsertf128 $1, 32(%r10), %ymm14, %ymm14
8179 ; AVX2-FCP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
8180 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
8181 ; AVX2-FCP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
8182 ; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
8183 ; AVX2-FCP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
8184 ; AVX2-FCP-NEXT: # xmm14 = xmm14[0],mem[0]
8185 ; AVX2-FCP-NEXT: vinsertf128 $1, 64(%rdx), %ymm14, %ymm14
8186 ; AVX2-FCP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
8187 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
8188 ; AVX2-FCP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
8189 ; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
8190 ; AVX2-FCP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
8191 ; AVX2-FCP-NEXT: # xmm14 = xmm14[0],mem[0]
8192 ; AVX2-FCP-NEXT: vinsertf128 $1, 64(%r10), %ymm14, %ymm14
8193 ; AVX2-FCP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
8194 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
8195 ; AVX2-FCP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
8196 ; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
8197 ; AVX2-FCP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
8198 ; AVX2-FCP-NEXT: # xmm14 = xmm14[0],mem[0]
8199 ; AVX2-FCP-NEXT: vinsertf128 $1, 96(%rdx), %ymm14, %ymm14
8200 ; AVX2-FCP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
8201 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
8202 ; AVX2-FCP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
8203 ; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
8204 ; AVX2-FCP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
8205 ; AVX2-FCP-NEXT: # xmm14 = xmm14[0],mem[0]
8206 ; AVX2-FCP-NEXT: vinsertf128 $1, 96(%r10), %ymm14, %ymm14
8207 ; AVX2-FCP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
8208 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
8209 ; AVX2-FCP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
8210 ; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
8211 ; AVX2-FCP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
8212 ; AVX2-FCP-NEXT: # xmm14 = xmm14[0],mem[0]
8213 ; AVX2-FCP-NEXT: vinsertf128 $1, 128(%rdx), %ymm14, %ymm14
8214 ; AVX2-FCP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
8215 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
8216 ; AVX2-FCP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
8217 ; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
8218 ; AVX2-FCP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
8219 ; AVX2-FCP-NEXT: # xmm14 = xmm14[0],mem[0]
8220 ; AVX2-FCP-NEXT: vinsertf128 $1, 128(%r10), %ymm14, %ymm14
8221 ; AVX2-FCP-NEXT: vbroadcastsd (%rsp), %ymm15 # 16-byte Folded Reload
8222 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
8223 ; AVX2-FCP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
8224 ; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
8225 ; AVX2-FCP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
8226 ; AVX2-FCP-NEXT: # xmm14 = xmm14[0],mem[0]
8227 ; AVX2-FCP-NEXT: vinsertf128 $1, 160(%rdx), %ymm14, %ymm14
8228 ; AVX2-FCP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
8229 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
8230 ; AVX2-FCP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
8231 ; AVX2-FCP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm13, %xmm13 # 16-byte Folded Reload
8232 ; AVX2-FCP-NEXT: # xmm13 = xmm13[0],mem[0]
8233 ; AVX2-FCP-NEXT: vinsertf128 $1, 160(%r10), %ymm13, %ymm13
8234 ; AVX2-FCP-NEXT: vbroadcastsd %xmm12, %ymm12
8235 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm12 = ymm13[0,1,2,3,4,5],ymm12[6,7]
8236 ; AVX2-FCP-NEXT: vmovups %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
8237 ; AVX2-FCP-NEXT: vmovlhps {{.*#+}} xmm10 = xmm10[0],xmm11[0]
8238 ; AVX2-FCP-NEXT: vinsertf128 $1, 192(%rdx), %ymm10, %ymm10
8239 ; AVX2-FCP-NEXT: vbroadcastsd %xmm9, %ymm9
8240 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm9 = ymm10[0,1,2,3,4,5],ymm9[6,7]
8241 ; AVX2-FCP-NEXT: vmovups %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
8242 ; AVX2-FCP-NEXT: vmovlhps {{.*#+}} xmm7 = xmm7[0],xmm8[0]
8243 ; AVX2-FCP-NEXT: vinsertf128 $1, 192(%r10), %ymm7, %ymm7
8244 ; AVX2-FCP-NEXT: vbroadcastsd %xmm6, %ymm6
8245 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3,4,5],ymm6[6,7]
8246 ; AVX2-FCP-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
8247 ; AVX2-FCP-NEXT: vmovlhps {{.*#+}} xmm4 = xmm4[0],xmm5[0]
8248 ; AVX2-FCP-NEXT: vinsertf128 $1, 224(%rdx), %ymm4, %ymm4
8249 ; AVX2-FCP-NEXT: vbroadcastsd %xmm3, %ymm3
8250 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm3 = ymm4[0,1,2,3,4,5],ymm3[6,7]
8251 ; AVX2-FCP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
8252 ; AVX2-FCP-NEXT: vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
8253 ; AVX2-FCP-NEXT: vinsertf128 $1, 224(%r10), %ymm1, %ymm1
8254 ; AVX2-FCP-NEXT: vbroadcastsd %xmm0, %ymm0
8255 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
8256 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
8257 ; AVX2-FCP-NEXT: vmovaps (%rdi), %ymm0
8258 ; AVX2-FCP-NEXT: vmovaps (%rsi), %ymm1
8259 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
8260 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
8261 ; AVX2-FCP-NEXT: vbroadcastsd 16(%rcx), %ymm3
8262 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
8263 ; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
8264 ; AVX2-FCP-NEXT: vmovaps (%r8), %ymm2
8265 ; AVX2-FCP-NEXT: vmovaps (%r9), %ymm3
8266 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
8267 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],mem[2,3]
8268 ; AVX2-FCP-NEXT: vbroadcastsd 16(%rax), %ymm5
8269 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
8270 ; AVX2-FCP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
8271 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
8272 ; AVX2-FCP-NEXT: vbroadcastsd 24(%rdx), %ymm1
8273 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm8 = ymm0[2,3],ymm1[2,3]
8274 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
8275 ; AVX2-FCP-NEXT: vbroadcastsd 24(%r10), %ymm1
8276 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm9 = ymm0[2,3],ymm1[2,3]
8277 ; AVX2-FCP-NEXT: vmovaps 32(%rdi), %ymm0
8278 ; AVX2-FCP-NEXT: vmovaps 32(%rsi), %ymm1
8279 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
8280 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
8281 ; AVX2-FCP-NEXT: vbroadcastsd 48(%rcx), %ymm3
8282 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
8283 ; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
8284 ; AVX2-FCP-NEXT: vmovaps 32(%r8), %ymm2
8285 ; AVX2-FCP-NEXT: vmovaps 32(%r9), %ymm3
8286 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
8287 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],mem[2,3]
8288 ; AVX2-FCP-NEXT: vbroadcastsd 48(%rax), %ymm5
8289 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
8290 ; AVX2-FCP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
8291 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
8292 ; AVX2-FCP-NEXT: vbroadcastsd 56(%rdx), %ymm1
8293 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm12 = ymm0[2,3],ymm1[2,3]
8294 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
8295 ; AVX2-FCP-NEXT: vbroadcastsd 56(%r10), %ymm1
8296 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
8297 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
8298 ; AVX2-FCP-NEXT: vmovaps 64(%rdi), %ymm0
8299 ; AVX2-FCP-NEXT: vmovaps 64(%rsi), %ymm1
8300 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
8301 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
8302 ; AVX2-FCP-NEXT: vbroadcastsd 80(%rcx), %ymm3
8303 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
8304 ; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
8305 ; AVX2-FCP-NEXT: vmovaps 64(%r8), %ymm2
8306 ; AVX2-FCP-NEXT: vmovaps 64(%r9), %ymm3
8307 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
8308 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],mem[2,3]
8309 ; AVX2-FCP-NEXT: vbroadcastsd 80(%rax), %ymm5
8310 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
8311 ; AVX2-FCP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
8312 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
8313 ; AVX2-FCP-NEXT: vbroadcastsd 88(%rdx), %ymm1
8314 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
8315 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
8316 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
8317 ; AVX2-FCP-NEXT: vbroadcastsd 88(%r10), %ymm1
8318 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
8319 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
8320 ; AVX2-FCP-NEXT: vmovaps 96(%rdi), %ymm0
8321 ; AVX2-FCP-NEXT: vmovaps 96(%rsi), %ymm1
8322 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
8323 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
8324 ; AVX2-FCP-NEXT: vbroadcastsd 112(%rcx), %ymm3
8325 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
8326 ; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
8327 ; AVX2-FCP-NEXT: vmovaps 96(%r8), %ymm2
8328 ; AVX2-FCP-NEXT: vmovaps 96(%r9), %ymm3
8329 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
8330 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],mem[2,3]
8331 ; AVX2-FCP-NEXT: vbroadcastsd 112(%rax), %ymm5
8332 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
8333 ; AVX2-FCP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
8334 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
8335 ; AVX2-FCP-NEXT: vbroadcastsd 120(%rdx), %ymm1
8336 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
8337 ; AVX2-FCP-NEXT: vmovups %ymm0, (%rsp) # 32-byte Spill
8338 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
8339 ; AVX2-FCP-NEXT: vbroadcastsd 120(%r10), %ymm1
8340 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
8341 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
8342 ; AVX2-FCP-NEXT: vmovaps 128(%rdi), %ymm0
8343 ; AVX2-FCP-NEXT: vmovaps 128(%rsi), %ymm1
8344 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
8345 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
8346 ; AVX2-FCP-NEXT: vbroadcastsd 144(%rcx), %ymm3
8347 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
8348 ; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
8349 ; AVX2-FCP-NEXT: vmovaps 128(%r8), %ymm2
8350 ; AVX2-FCP-NEXT: vmovaps 128(%r9), %ymm3
8351 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
8352 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],mem[2,3]
8353 ; AVX2-FCP-NEXT: vbroadcastsd 144(%rax), %ymm5
8354 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
8355 ; AVX2-FCP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
8356 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
8357 ; AVX2-FCP-NEXT: vbroadcastsd 152(%rdx), %ymm1
8358 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
8359 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
8360 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
8361 ; AVX2-FCP-NEXT: vbroadcastsd 152(%r10), %ymm1
8362 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
8363 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
8364 ; AVX2-FCP-NEXT: vmovaps 160(%rdi), %ymm0
8365 ; AVX2-FCP-NEXT: vmovaps 160(%rsi), %ymm1
8366 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
8367 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
8368 ; AVX2-FCP-NEXT: vbroadcastsd 176(%rcx), %ymm3
8369 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
8370 ; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
8371 ; AVX2-FCP-NEXT: vmovaps 160(%r8), %ymm2
8372 ; AVX2-FCP-NEXT: vmovaps 160(%r9), %ymm3
8373 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
8374 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],mem[2,3]
8375 ; AVX2-FCP-NEXT: vbroadcastsd 176(%rax), %ymm5
8376 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
8377 ; AVX2-FCP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
8378 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
8379 ; AVX2-FCP-NEXT: vbroadcastsd 184(%rdx), %ymm1
8380 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
8381 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
8382 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
8383 ; AVX2-FCP-NEXT: vbroadcastsd 184(%r10), %ymm1
8384 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
8385 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
8386 ; AVX2-FCP-NEXT: vmovaps 192(%rdi), %ymm0
8387 ; AVX2-FCP-NEXT: vmovaps 192(%rsi), %ymm1
8388 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
8389 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
8390 ; AVX2-FCP-NEXT: vbroadcastsd 208(%rcx), %ymm3
8391 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
8392 ; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
8393 ; AVX2-FCP-NEXT: vmovaps 192(%r8), %ymm2
8394 ; AVX2-FCP-NEXT: vmovaps 192(%r9), %ymm3
8395 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
8396 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],mem[2,3]
8397 ; AVX2-FCP-NEXT: vbroadcastsd 208(%rax), %ymm5
8398 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
8399 ; AVX2-FCP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
8400 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
8401 ; AVX2-FCP-NEXT: vbroadcastsd 216(%rdx), %ymm1
8402 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm13 = ymm0[2,3],ymm1[2,3]
8403 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
8404 ; AVX2-FCP-NEXT: vbroadcastsd 216(%r10), %ymm1
8405 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm11 = ymm0[2,3],ymm1[2,3]
8406 ; AVX2-FCP-NEXT: vmovaps 224(%rdi), %ymm0
8407 ; AVX2-FCP-NEXT: vmovaps 224(%rsi), %ymm3
8408 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm1 = ymm0[0],ymm3[0],ymm0[2],ymm3[2]
8409 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm1[2,3],mem[2,3]
8410 ; AVX2-FCP-NEXT: vbroadcastsd 240(%rcx), %ymm5
8411 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm10 = ymm1[0,1,2,3,4,5],ymm5[6,7]
8412 ; AVX2-FCP-NEXT: vmovaps 224(%r8), %ymm5
8413 ; AVX2-FCP-NEXT: vmovaps 224(%r9), %ymm6
8414 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm7 = ymm5[0],ymm6[0],ymm5[2],ymm6[2]
8415 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm7 = ymm7[2,3],mem[2,3]
8416 ; AVX2-FCP-NEXT: vbroadcastsd 240(%rax), %ymm15
8417 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm15 = ymm7[0,1,2,3,4,5],ymm15[6,7]
8418 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm3[1],ymm0[3],ymm3[3]
8419 ; AVX2-FCP-NEXT: vbroadcastsd 248(%rdx), %ymm3
8420 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm0[2,3],ymm3[2,3]
8421 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm1 = ymm5[1],ymm6[1],ymm5[3],ymm6[3]
8422 ; AVX2-FCP-NEXT: vbroadcastsd 248(%r10), %ymm3
8423 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm1[2,3],ymm3[2,3]
8424 ; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rdx
8425 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm8[0,1,2,3,4,5],mem[6,7]
8426 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
8427 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm9[0,1,2,3,4,5],mem[6,7]
8428 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
8429 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm14 = ymm12[0,1,2,3,4,5],mem[6,7]
8430 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
8431 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm12 = ymm0[0,1,2,3,4,5],mem[6,7]
8432 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
8433 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm9 = ymm0[0,1,2,3,4,5],mem[6,7]
8434 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
8435 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm8 = ymm0[0,1,2,3,4,5],mem[6,7]
8436 ; AVX2-FCP-NEXT: vmovups (%rsp), %ymm0 # 32-byte Reload
8437 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm7 = ymm0[0,1,2,3,4,5],mem[6,7]
8438 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
8439 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm6 = ymm0[0,1,2,3,4,5],mem[6,7]
8440 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
8441 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm5 = ymm0[0,1,2,3,4,5],mem[6,7]
8442 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
8443 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm3 = ymm0[0,1,2,3,4,5],mem[6,7]
8444 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
8445 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm0[0,1,2,3,4,5],mem[6,7]
8446 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
8447 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],mem[6,7]
8448 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm13 = ymm13[0,1,2,3,4,5],mem[6,7]
8449 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm11 = ymm11[0,1,2,3,4,5],mem[6,7]
8450 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],mem[6,7]
8451 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],mem[6,7]
8452 ; AVX2-FCP-NEXT: vmovaps %ymm2, 2016(%rdx)
8453 ; AVX2-FCP-NEXT: vmovaps %ymm4, 1984(%rdx)
8454 ; AVX2-FCP-NEXT: vmovaps %ymm15, 1952(%rdx)
8455 ; AVX2-FCP-NEXT: vmovaps %ymm10, 1920(%rdx)
8456 ; AVX2-FCP-NEXT: vmovaps %ymm11, 1760(%rdx)
8457 ; AVX2-FCP-NEXT: vmovaps %ymm13, 1728(%rdx)
8458 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
8459 ; AVX2-FCP-NEXT: vmovaps %ymm2, 1696(%rdx)
8460 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
8461 ; AVX2-FCP-NEXT: vmovaps %ymm2, 1664(%rdx)
8462 ; AVX2-FCP-NEXT: vmovaps %ymm0, 1504(%rdx)
8463 ; AVX2-FCP-NEXT: vmovaps %ymm1, 1472(%rdx)
8464 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
8465 ; AVX2-FCP-NEXT: vmovaps %ymm0, 1440(%rdx)
8466 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
8467 ; AVX2-FCP-NEXT: vmovaps %ymm0, 1408(%rdx)
8468 ; AVX2-FCP-NEXT: vmovaps %ymm3, 1248(%rdx)
8469 ; AVX2-FCP-NEXT: vmovaps %ymm5, 1216(%rdx)
8470 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
8471 ; AVX2-FCP-NEXT: vmovaps %ymm0, 1184(%rdx)
8472 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
8473 ; AVX2-FCP-NEXT: vmovaps %ymm0, 1152(%rdx)
8474 ; AVX2-FCP-NEXT: vmovaps %ymm6, 992(%rdx)
8475 ; AVX2-FCP-NEXT: vmovaps %ymm7, 960(%rdx)
8476 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
8477 ; AVX2-FCP-NEXT: vmovaps %ymm0, 928(%rdx)
8478 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
8479 ; AVX2-FCP-NEXT: vmovaps %ymm0, 896(%rdx)
8480 ; AVX2-FCP-NEXT: vmovaps %ymm8, 736(%rdx)
8481 ; AVX2-FCP-NEXT: vmovaps %ymm9, 704(%rdx)
8482 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
8483 ; AVX2-FCP-NEXT: vmovaps %ymm0, 672(%rdx)
8484 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
8485 ; AVX2-FCP-NEXT: vmovaps %ymm0, 640(%rdx)
8486 ; AVX2-FCP-NEXT: vmovaps %ymm12, 480(%rdx)
8487 ; AVX2-FCP-NEXT: vmovaps %ymm14, 448(%rdx)
8488 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
8489 ; AVX2-FCP-NEXT: vmovaps %ymm0, 416(%rdx)
8490 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
8491 ; AVX2-FCP-NEXT: vmovaps %ymm0, 384(%rdx)
8492 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
8493 ; AVX2-FCP-NEXT: vmovaps %ymm0, 224(%rdx)
8494 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
8495 ; AVX2-FCP-NEXT: vmovaps %ymm0, 192(%rdx)
8496 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
8497 ; AVX2-FCP-NEXT: vmovaps %ymm0, 160(%rdx)
8498 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
8499 ; AVX2-FCP-NEXT: vmovaps %ymm0, 128(%rdx)
8500 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
8501 ; AVX2-FCP-NEXT: vmovaps %ymm0, 1888(%rdx)
8502 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
8503 ; AVX2-FCP-NEXT: vmovaps %ymm0, 1856(%rdx)
8504 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
8505 ; AVX2-FCP-NEXT: vmovaps %ymm0, 1824(%rdx)
8506 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
8507 ; AVX2-FCP-NEXT: vmovaps %ymm0, 1792(%rdx)
8508 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
8509 ; AVX2-FCP-NEXT: vmovaps %ymm0, 1632(%rdx)
8510 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
8511 ; AVX2-FCP-NEXT: vmovaps %ymm0, 1600(%rdx)
8512 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
8513 ; AVX2-FCP-NEXT: vmovaps %ymm0, 1568(%rdx)
8514 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
8515 ; AVX2-FCP-NEXT: vmovaps %ymm0, 1536(%rdx)
8516 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
8517 ; AVX2-FCP-NEXT: vmovaps %ymm0, 1376(%rdx)
8518 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
8519 ; AVX2-FCP-NEXT: vmovaps %ymm0, 1344(%rdx)
8520 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
8521 ; AVX2-FCP-NEXT: vmovaps %ymm0, 1312(%rdx)
8522 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
8523 ; AVX2-FCP-NEXT: vmovaps %ymm0, 1280(%rdx)
8524 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
8525 ; AVX2-FCP-NEXT: vmovaps %ymm0, 1120(%rdx)
8526 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
8527 ; AVX2-FCP-NEXT: vmovaps %ymm0, 1088(%rdx)
8528 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
8529 ; AVX2-FCP-NEXT: vmovaps %ymm0, 1056(%rdx)
8530 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
8531 ; AVX2-FCP-NEXT: vmovaps %ymm0, 1024(%rdx)
8532 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
8533 ; AVX2-FCP-NEXT: vmovaps %ymm0, 864(%rdx)
8534 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
8535 ; AVX2-FCP-NEXT: vmovaps %ymm0, 832(%rdx)
8536 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
8537 ; AVX2-FCP-NEXT: vmovaps %ymm0, 800(%rdx)
8538 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
8539 ; AVX2-FCP-NEXT: vmovaps %ymm0, 768(%rdx)
8540 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
8541 ; AVX2-FCP-NEXT: vmovaps %ymm0, 608(%rdx)
8542 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
8543 ; AVX2-FCP-NEXT: vmovaps %ymm0, 576(%rdx)
8544 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
8545 ; AVX2-FCP-NEXT: vmovaps %ymm0, 544(%rdx)
8546 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
8547 ; AVX2-FCP-NEXT: vmovaps %ymm0, 512(%rdx)
8548 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
8549 ; AVX2-FCP-NEXT: vmovaps %ymm0, 352(%rdx)
8550 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
8551 ; AVX2-FCP-NEXT: vmovaps %ymm0, 320(%rdx)
8552 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
8553 ; AVX2-FCP-NEXT: vmovaps %ymm0, 288(%rdx)
8554 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
8555 ; AVX2-FCP-NEXT: vmovaps %ymm0, 256(%rdx)
8556 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
8557 ; AVX2-FCP-NEXT: vmovaps %ymm0, 96(%rdx)
8558 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
8559 ; AVX2-FCP-NEXT: vmovaps %ymm0, 64(%rdx)
8560 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
8561 ; AVX2-FCP-NEXT: vmovaps %ymm0, 32(%rdx)
8562 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
8563 ; AVX2-FCP-NEXT: vmovaps %ymm0, (%rdx)
8564 ; AVX2-FCP-NEXT: addq $1704, %rsp # imm = 0x6A8
8565 ; AVX2-FCP-NEXT: vzeroupper
8566 ; AVX2-FCP-NEXT: retq
8567 ;
8568 ; AVX512-LABEL: store_i64_stride8_vf32:
8569 ; AVX512: # %bb.0:
8570 ; AVX512-NEXT: subq $2312, %rsp # imm = 0x908
8571 ; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %rax
8572 ; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %r10
8573 ; AVX512-NEXT: vmovaps 128(%rdi), %zmm0
8574 ; AVX512-NEXT: vmovups %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
8575 ; AVX512-NEXT: vmovdqa64 64(%rdi), %zmm6
8576 ; AVX512-NEXT: vmovdqa64 (%rdi), %zmm2
8577 ; AVX512-NEXT: vmovdqa64 128(%rsi), %zmm0
8578 ; AVX512-NEXT: vmovdqa64 64(%rsi), %zmm5
8579 ; AVX512-NEXT: vmovdqa64 (%rsi), %zmm3
8580 ; AVX512-NEXT: vmovaps 192(%rdx), %zmm1
8581 ; AVX512-NEXT: vmovups %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
8582 ; AVX512-NEXT: vmovaps 128(%rdx), %zmm1
8583 ; AVX512-NEXT: vmovups %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
8584 ; AVX512-NEXT: vmovdqa64 64(%rdx), %zmm7
8585 ; AVX512-NEXT: vmovdqa64 (%rdx), %zmm4
8586 ; AVX512-NEXT: vmovdqa64 192(%rcx), %zmm23
8587 ; AVX512-NEXT: vmovdqa64 128(%rcx), %zmm1
8588 ; AVX512-NEXT: vmovdqa64 64(%rcx), %zmm9
8589 ; AVX512-NEXT: vmovdqa64 (%rcx), %zmm8
8590 ; AVX512-NEXT: vmovdqa64 (%r8), %zmm18
8591 ; AVX512-NEXT: vmovdqa64 64(%r8), %zmm16
8592 ; AVX512-NEXT: vmovdqa64 (%r9), %zmm12
8593 ; AVX512-NEXT: vmovdqa64 64(%r9), %zmm19
8594 ; AVX512-NEXT: vmovdqa64 (%r10), %zmm25
8595 ; AVX512-NEXT: vmovdqa64 64(%r10), %zmm17
8596 ; AVX512-NEXT: vmovdqa64 (%rax), %zmm15
8597 ; AVX512-NEXT: vmovdqa64 64(%rax), %zmm20
8598 ; AVX512-NEXT: movb $-64, %r11b
8599 ; AVX512-NEXT: kmovw %r11d, %k1
8600 ; AVX512-NEXT: vbroadcasti32x4 {{.*#+}} zmm27 = [4,12,4,12,4,12,4,12]
8601 ; AVX512-NEXT: # zmm27 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
8602 ; AVX512-NEXT: vmovdqa64 %zmm4, %zmm10
8603 ; AVX512-NEXT: vpermt2q %zmm8, %zmm27, %zmm10
8604 ; AVX512-NEXT: vmovdqa64 %zmm2, %zmm11
8605 ; AVX512-NEXT: vpermt2q %zmm3, %zmm27, %zmm11
8606 ; AVX512-NEXT: vpblendd {{.*#+}} ymm10 = ymm11[0,1,2,3],ymm10[4,5,6,7]
8607 ; AVX512-NEXT: vmovdqa64 %zmm25, %zmm11
8608 ; AVX512-NEXT: vpermt2q %zmm15, %zmm27, %zmm11
8609 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} zmm13 = zmm18[0],zmm12[0],zmm18[2],zmm12[2],zmm18[4],zmm12[4],zmm18[6],zmm12[6]
8610 ; AVX512-NEXT: vmovdqa64 %zmm11, %zmm13 {%k1}
8611 ; AVX512-NEXT: vinserti64x4 $0, %ymm10, %zmm13, %zmm10
8612 ; AVX512-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
8613 ; AVX512-NEXT: vbroadcasti32x4 {{.*#+}} zmm24 = [5,13,5,13,5,13,5,13]
8614 ; AVX512-NEXT: # zmm24 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
8615 ; AVX512-NEXT: vmovdqa64 %zmm4, %zmm10
8616 ; AVX512-NEXT: vpermt2q %zmm8, %zmm24, %zmm10
8617 ; AVX512-NEXT: vmovdqa64 %zmm2, %zmm11
8618 ; AVX512-NEXT: vpermt2q %zmm3, %zmm24, %zmm11
8619 ; AVX512-NEXT: vpblendd {{.*#+}} ymm10 = ymm11[0,1,2,3],ymm10[4,5,6,7]
8620 ; AVX512-NEXT: vmovdqa64 %zmm25, %zmm11
8621 ; AVX512-NEXT: vpermt2q %zmm15, %zmm24, %zmm11
8622 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} zmm14 = zmm18[1],zmm12[1],zmm18[3],zmm12[3],zmm18[5],zmm12[5],zmm18[7],zmm12[7]
8623 ; AVX512-NEXT: vmovdqa64 %zmm11, %zmm14 {%k1}
8624 ; AVX512-NEXT: vinserti64x4 $0, %ymm10, %zmm14, %zmm10
8625 ; AVX512-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
8626 ; AVX512-NEXT: vbroadcasti32x4 {{.*#+}} zmm22 = [6,14,6,14,6,14,6,14]
8627 ; AVX512-NEXT: # zmm22 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
8628 ; AVX512-NEXT: vmovdqa64 %zmm4, %zmm10
8629 ; AVX512-NEXT: vpermt2q %zmm8, %zmm22, %zmm10
8630 ; AVX512-NEXT: vmovdqa64 %zmm2, %zmm11
8631 ; AVX512-NEXT: vpermt2q %zmm3, %zmm22, %zmm11
8632 ; AVX512-NEXT: vpblendd {{.*#+}} ymm10 = ymm11[0,1,2,3],ymm10[4,5,6,7]
8633 ; AVX512-NEXT: vmovdqa64 %zmm18, %zmm11
8634 ; AVX512-NEXT: vpermt2q %zmm12, %zmm22, %zmm11
8635 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} zmm11 {%k1} = zmm25[0],zmm15[0],zmm25[2],zmm15[2],zmm25[4],zmm15[4],zmm25[6],zmm15[6]
8636 ; AVX512-NEXT: vinserti64x4 $0, %ymm10, %zmm11, %zmm10
8637 ; AVX512-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
8638 ; AVX512-NEXT: vbroadcasti32x4 {{.*#+}} zmm21 = [7,15,7,15,7,15,7,15]
8639 ; AVX512-NEXT: # zmm21 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
8640 ; AVX512-NEXT: vpermt2q %zmm8, %zmm21, %zmm4
8641 ; AVX512-NEXT: vpermt2q %zmm3, %zmm21, %zmm2
8642 ; AVX512-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm4[4,5,6,7]
8643 ; AVX512-NEXT: vmovdqa64 %zmm18, %zmm3
8644 ; AVX512-NEXT: vpermt2q %zmm12, %zmm21, %zmm3
8645 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} zmm3 {%k1} = zmm25[1],zmm15[1],zmm25[3],zmm15[3],zmm25[5],zmm15[5],zmm25[7],zmm15[7]
8646 ; AVX512-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
8647 ; AVX512-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
8648 ; AVX512-NEXT: vmovdqa64 %zmm7, %zmm2
8649 ; AVX512-NEXT: vpermt2q %zmm9, %zmm27, %zmm2
8650 ; AVX512-NEXT: vmovdqa64 %zmm6, %zmm3
8651 ; AVX512-NEXT: vpermt2q %zmm5, %zmm27, %zmm3
8652 ; AVX512-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3],ymm2[4,5,6,7]
8653 ; AVX512-NEXT: vmovdqa64 %zmm17, %zmm3
8654 ; AVX512-NEXT: vpermt2q %zmm20, %zmm27, %zmm3
8655 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} zmm4 = zmm16[0],zmm19[0],zmm16[2],zmm19[2],zmm16[4],zmm19[4],zmm16[6],zmm19[6]
8656 ; AVX512-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
8657 ; AVX512-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
8658 ; AVX512-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
8659 ; AVX512-NEXT: vmovdqa64 %zmm7, %zmm2
8660 ; AVX512-NEXT: vpermt2q %zmm9, %zmm24, %zmm2
8661 ; AVX512-NEXT: vmovdqa64 %zmm6, %zmm3
8662 ; AVX512-NEXT: vpermt2q %zmm5, %zmm24, %zmm3
8663 ; AVX512-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3],ymm2[4,5,6,7]
8664 ; AVX512-NEXT: vmovdqa64 %zmm17, %zmm3
8665 ; AVX512-NEXT: vpermt2q %zmm20, %zmm24, %zmm3
8666 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} zmm4 = zmm16[1],zmm19[1],zmm16[3],zmm19[3],zmm16[5],zmm19[5],zmm16[7],zmm19[7]
8667 ; AVX512-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
8668 ; AVX512-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
8669 ; AVX512-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
8670 ; AVX512-NEXT: vmovdqa64 %zmm7, %zmm2
8671 ; AVX512-NEXT: vpermt2q %zmm9, %zmm22, %zmm2
8672 ; AVX512-NEXT: vmovdqa64 %zmm6, %zmm3
8673 ; AVX512-NEXT: vpermt2q %zmm5, %zmm22, %zmm3
8674 ; AVX512-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3],ymm2[4,5,6,7]
8675 ; AVX512-NEXT: vmovdqa64 %zmm16, %zmm3
8676 ; AVX512-NEXT: vpermt2q %zmm19, %zmm22, %zmm3
8677 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} zmm3 {%k1} = zmm17[0],zmm20[0],zmm17[2],zmm20[2],zmm17[4],zmm20[4],zmm17[6],zmm20[6]
8678 ; AVX512-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
8679 ; AVX512-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
8680 ; AVX512-NEXT: vpermt2q %zmm9, %zmm21, %zmm7
8681 ; AVX512-NEXT: vpermt2q %zmm5, %zmm21, %zmm6
8682 ; AVX512-NEXT: vpblendd {{.*#+}} ymm2 = ymm6[0,1,2,3],ymm7[4,5,6,7]
8683 ; AVX512-NEXT: vmovdqa64 %zmm16, %zmm3
8684 ; AVX512-NEXT: vpermt2q %zmm19, %zmm21, %zmm3
8685 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} zmm3 {%k1} = zmm17[1],zmm20[1],zmm17[3],zmm20[3],zmm17[5],zmm20[5],zmm17[7],zmm20[7]
8686 ; AVX512-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
8687 ; AVX512-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
8688 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
8689 ; AVX512-NEXT: vmovdqa64 %zmm3, %zmm2
8690 ; AVX512-NEXT: vpermt2q %zmm1, %zmm27, %zmm2
8691 ; AVX512-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
8692 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
8693 ; AVX512-NEXT: vmovdqa64 %zmm2, %zmm4
8694 ; AVX512-NEXT: vpermt2q %zmm0, %zmm27, %zmm4
8695 ; AVX512-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
8696 ; AVX512-NEXT: vmovdqa64 %zmm3, %zmm4
8697 ; AVX512-NEXT: vpermt2q %zmm1, %zmm24, %zmm4
8698 ; AVX512-NEXT: vmovdqu64 %zmm4, (%rsp) # 64-byte Spill
8699 ; AVX512-NEXT: vmovdqa64 %zmm2, %zmm4
8700 ; AVX512-NEXT: vpermt2q %zmm0, %zmm24, %zmm4
8701 ; AVX512-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
8702 ; AVX512-NEXT: vmovdqa64 %zmm3, %zmm4
8703 ; AVX512-NEXT: vpermt2q %zmm1, %zmm22, %zmm4
8704 ; AVX512-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
8705 ; AVX512-NEXT: vmovdqa64 %zmm2, %zmm4
8706 ; AVX512-NEXT: vpermt2q %zmm0, %zmm22, %zmm4
8707 ; AVX512-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
8708 ; AVX512-NEXT: vpermt2q %zmm1, %zmm21, %zmm3
8709 ; AVX512-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
8710 ; AVX512-NEXT: vpermt2q %zmm0, %zmm21, %zmm2
8711 ; AVX512-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
8712 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
8713 ; AVX512-NEXT: vmovdqa64 %zmm0, %zmm1
8714 ; AVX512-NEXT: vpermt2q %zmm23, %zmm27, %zmm1
8715 ; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
8716 ; AVX512-NEXT: vmovdqa64 %zmm0, %zmm1
8717 ; AVX512-NEXT: vpermt2q %zmm23, %zmm24, %zmm1
8718 ; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
8719 ; AVX512-NEXT: vmovdqa64 %zmm0, %zmm1
8720 ; AVX512-NEXT: vpermt2q %zmm23, %zmm22, %zmm1
8721 ; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
8722 ; AVX512-NEXT: vpermt2q %zmm23, %zmm21, %zmm0
8723 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
8724 ; AVX512-NEXT: vmovdqa64 192(%rdi), %zmm1
8725 ; AVX512-NEXT: vmovdqa64 192(%rsi), %zmm0
8726 ; AVX512-NEXT: vmovdqa64 %zmm1, %zmm2
8727 ; AVX512-NEXT: vpermt2q %zmm0, %zmm27, %zmm2
8728 ; AVX512-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
8729 ; AVX512-NEXT: vmovdqa64 %zmm1, %zmm2
8730 ; AVX512-NEXT: vpermt2q %zmm0, %zmm24, %zmm2
8731 ; AVX512-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
8732 ; AVX512-NEXT: vmovdqa64 %zmm1, %zmm2
8733 ; AVX512-NEXT: vpermt2q %zmm0, %zmm22, %zmm2
8734 ; AVX512-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
8735 ; AVX512-NEXT: vpermt2q %zmm0, %zmm21, %zmm1
8736 ; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
8737 ; AVX512-NEXT: vbroadcasti32x4 {{.*#+}} zmm30 = [0,8,0,8,0,8,0,8]
8738 ; AVX512-NEXT: # zmm30 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
8739 ; AVX512-NEXT: vmovdqa64 %zmm25, %zmm0
8740 ; AVX512-NEXT: vpermt2q %zmm15, %zmm30, %zmm0
8741 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
8742 ; AVX512-NEXT: vbroadcasti32x4 {{.*#+}} zmm31 = [1,9,1,9,1,9,1,9]
8743 ; AVX512-NEXT: # zmm31 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
8744 ; AVX512-NEXT: vmovdqa64 %zmm25, %zmm0
8745 ; AVX512-NEXT: vpermt2q %zmm15, %zmm31, %zmm0
8746 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
8747 ; AVX512-NEXT: vbroadcasti32x4 {{.*#+}} zmm23 = [2,10,2,10,2,10,2,10]
8748 ; AVX512-NEXT: # zmm23 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
8749 ; AVX512-NEXT: vmovdqa64 %zmm25, %zmm0
8750 ; AVX512-NEXT: vpermt2q %zmm15, %zmm23, %zmm0
8751 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
8752 ; AVX512-NEXT: vbroadcasti32x4 {{.*#+}} zmm26 = [3,11,3,11,3,11,3,11]
8753 ; AVX512-NEXT: # zmm26 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
8754 ; AVX512-NEXT: vpermt2q %zmm15, %zmm26, %zmm25
8755 ; AVX512-NEXT: vmovdqu64 %zmm25, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
8756 ; AVX512-NEXT: vmovdqa64 %zmm18, %zmm0
8757 ; AVX512-NEXT: vmovdqa64 %zmm18, %zmm29
8758 ; AVX512-NEXT: vpermt2q %zmm12, %zmm30, %zmm29
8759 ; AVX512-NEXT: vpermt2q %zmm12, %zmm31, %zmm18
8760 ; AVX512-NEXT: vmovdqa64 %zmm0, %zmm1
8761 ; AVX512-NEXT: vpermt2q %zmm12, %zmm23, %zmm1
8762 ; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
8763 ; AVX512-NEXT: vpermt2q %zmm12, %zmm26, %zmm0
8764 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
8765 ; AVX512-NEXT: vmovdqa64 %zmm17, %zmm0
8766 ; AVX512-NEXT: vpermt2q %zmm20, %zmm30, %zmm0
8767 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
8768 ; AVX512-NEXT: vmovdqa64 %zmm17, %zmm0
8769 ; AVX512-NEXT: vpermt2q %zmm20, %zmm31, %zmm0
8770 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
8771 ; AVX512-NEXT: vmovdqa64 %zmm17, %zmm0
8772 ; AVX512-NEXT: vpermt2q %zmm20, %zmm23, %zmm0
8773 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
8774 ; AVX512-NEXT: vpermt2q %zmm20, %zmm26, %zmm17
8775 ; AVX512-NEXT: vmovdqu64 %zmm17, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
8776 ; AVX512-NEXT: vmovdqa64 %zmm16, %zmm14
8777 ; AVX512-NEXT: vpermt2q %zmm19, %zmm30, %zmm14
8778 ; AVX512-NEXT: vmovdqa64 %zmm16, %zmm11
8779 ; AVX512-NEXT: vpermt2q %zmm19, %zmm31, %zmm11
8780 ; AVX512-NEXT: vmovdqa64 %zmm16, %zmm25
8781 ; AVX512-NEXT: vpermt2q %zmm19, %zmm23, %zmm25
8782 ; AVX512-NEXT: vpermt2q %zmm19, %zmm26, %zmm16
8783 ; AVX512-NEXT: vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
8784 ; AVX512-NEXT: vmovdqa64 128(%r10), %zmm19
8785 ; AVX512-NEXT: vmovdqa64 128(%rax), %zmm0
8786 ; AVX512-NEXT: vmovdqa64 %zmm19, %zmm28
8787 ; AVX512-NEXT: vpermt2q %zmm0, %zmm30, %zmm28
8788 ; AVX512-NEXT: vmovdqa64 %zmm19, %zmm1
8789 ; AVX512-NEXT: vpermt2q %zmm0, %zmm31, %zmm1
8790 ; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
8791 ; AVX512-NEXT: vmovdqa64 %zmm19, %zmm13
8792 ; AVX512-NEXT: vpermt2q %zmm0, %zmm27, %zmm13
8793 ; AVX512-NEXT: vmovdqa64 %zmm19, %zmm16
8794 ; AVX512-NEXT: vpermt2q %zmm0, %zmm24, %zmm16
8795 ; AVX512-NEXT: vmovdqa64 128(%r8), %zmm20
8796 ; AVX512-NEXT: vmovdqa64 128(%r9), %zmm1
8797 ; AVX512-NEXT: vmovdqa64 %zmm20, %zmm3
8798 ; AVX512-NEXT: vpermt2q %zmm1, %zmm22, %zmm3
8799 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} zmm3 {%k1} = zmm19[0],zmm0[0],zmm19[2],zmm0[2],zmm19[4],zmm0[4],zmm19[6],zmm0[6]
8800 ; AVX512-NEXT: vmovdqa64 %zmm20, %zmm10
8801 ; AVX512-NEXT: vpermt2q %zmm1, %zmm21, %zmm10
8802 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} zmm10 {%k1} = zmm19[1],zmm0[1],zmm19[3],zmm0[3],zmm19[5],zmm0[5],zmm19[7],zmm0[7]
8803 ; AVX512-NEXT: vmovdqa64 %zmm19, %zmm2
8804 ; AVX512-NEXT: vpermt2q %zmm0, %zmm23, %zmm2
8805 ; AVX512-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
8806 ; AVX512-NEXT: vpermt2q %zmm0, %zmm26, %zmm19
8807 ; AVX512-NEXT: vmovdqa64 %zmm20, %zmm4
8808 ; AVX512-NEXT: vpermt2q %zmm1, %zmm30, %zmm4
8809 ; AVX512-NEXT: vmovdqa64 %zmm20, %zmm6
8810 ; AVX512-NEXT: vpermt2q %zmm1, %zmm31, %zmm6
8811 ; AVX512-NEXT: vmovdqa64 %zmm20, %zmm17
8812 ; AVX512-NEXT: vpermt2q %zmm1, %zmm23, %zmm17
8813 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} zmm5 = zmm20[0],zmm1[0],zmm20[2],zmm1[2],zmm20[4],zmm1[4],zmm20[6],zmm1[6]
8814 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} zmm7 = zmm20[1],zmm1[1],zmm20[3],zmm1[3],zmm20[5],zmm1[5],zmm20[7],zmm1[7]
8815 ; AVX512-NEXT: vpermt2q %zmm1, %zmm26, %zmm20
8816 ; AVX512-NEXT: vmovdqa64 192(%r10), %zmm8
8817 ; AVX512-NEXT: vmovdqa64 192(%rax), %zmm1
8818 ; AVX512-NEXT: vmovdqa64 %zmm8, %zmm12
8819 ; AVX512-NEXT: vpermt2q %zmm1, %zmm30, %zmm12
8820 ; AVX512-NEXT: vmovdqa64 %zmm8, %zmm15
8821 ; AVX512-NEXT: vpermt2q %zmm1, %zmm31, %zmm15
8822 ; AVX512-NEXT: vpermi2q %zmm1, %zmm8, %zmm27
8823 ; AVX512-NEXT: vpermi2q %zmm1, %zmm8, %zmm24
8824 ; AVX512-NEXT: vmovdqa64 192(%r8), %zmm2
8825 ; AVX512-NEXT: vmovdqa64 192(%r9), %zmm0
8826 ; AVX512-NEXT: vpermi2q %zmm0, %zmm2, %zmm22
8827 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} zmm22 {%k1} = zmm8[0],zmm1[0],zmm8[2],zmm1[2],zmm8[4],zmm1[4],zmm8[6],zmm1[6]
8828 ; AVX512-NEXT: vpermi2q %zmm0, %zmm2, %zmm21
8829 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} zmm21 {%k1} = zmm8[1],zmm1[1],zmm8[3],zmm1[3],zmm8[5],zmm1[5],zmm8[7],zmm1[7]
8830 ; AVX512-NEXT: vmovdqa64 %zmm8, %zmm9
8831 ; AVX512-NEXT: vpermt2q %zmm1, %zmm23, %zmm9
8832 ; AVX512-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
8833 ; AVX512-NEXT: vpermt2q %zmm1, %zmm26, %zmm8
8834 ; AVX512-NEXT: vpermi2q %zmm0, %zmm2, %zmm30
8835 ; AVX512-NEXT: vpermi2q %zmm0, %zmm2, %zmm31
8836 ; AVX512-NEXT: vpermi2q %zmm0, %zmm2, %zmm23
8837 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} zmm1 = zmm2[0],zmm0[0],zmm2[2],zmm0[2],zmm2[4],zmm0[4],zmm2[6],zmm0[6]
8838 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} zmm9 = zmm2[1],zmm0[1],zmm2[3],zmm0[3],zmm2[5],zmm0[5],zmm2[7],zmm0[7]
8839 ; AVX512-NEXT: vpermt2q %zmm0, %zmm26, %zmm2
8840 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
8841 ; AVX512-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
8842 ; AVX512-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
8843 ; AVX512-NEXT: vmovdqa64 %zmm13, %zmm5 {%k1}
8844 ; AVX512-NEXT: vinserti64x4 $0, %ymm0, %zmm5, %zmm0
8845 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
8846 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
8847 ; AVX512-NEXT: vpblendd $240, (%rsp), %ymm0, %ymm0 # 32-byte Folded Reload
8848 ; AVX512-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
8849 ; AVX512-NEXT: vmovdqa64 %zmm16, %zmm7 {%k1}
8850 ; AVX512-NEXT: vinserti64x4 $0, %ymm0, %zmm7, %zmm0
8851 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
8852 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
8853 ; AVX512-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
8854 ; AVX512-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
8855 ; AVX512-NEXT: vinserti64x4 $0, %ymm0, %zmm3, %zmm0
8856 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
8857 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
8858 ; AVX512-NEXT: vmovdqa64 %zmm0, %zmm29 {%k1}
8859 ; AVX512-NEXT: vmovdqa (%rsi), %xmm0
8860 ; AVX512-NEXT: vinserti128 $1, (%rcx), %ymm0, %ymm3
8861 ; AVX512-NEXT: vmovdqa (%rdi), %xmm0
8862 ; AVX512-NEXT: vinserti128 $1, (%rdx), %ymm0, %ymm5
8863 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm5[0],ymm3[0],ymm5[2],ymm3[2]
8864 ; AVX512-NEXT: vinserti64x4 $0, %ymm0, %zmm29, %zmm0
8865 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
8866 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
8867 ; AVX512-NEXT: vmovdqa64 %zmm0, %zmm18 {%k1}
8868 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} ymm3 = ymm5[1],ymm3[1],ymm5[3],ymm3[3]
8869 ; AVX512-NEXT: vinserti64x4 $0, %ymm3, %zmm18, %zmm0
8870 ; AVX512-NEXT: vmovdqu64 %zmm0, (%rsp) # 64-byte Spill
8871 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
8872 ; AVX512-NEXT: vmovdqa64 %zmm0, %zmm14 {%k1}
8873 ; AVX512-NEXT: vmovdqa 64(%rsi), %xmm3
8874 ; AVX512-NEXT: vinserti128 $1, 64(%rcx), %ymm3, %ymm3
8875 ; AVX512-NEXT: vmovdqa 64(%rdi), %xmm5
8876 ; AVX512-NEXT: vinserti128 $1, 64(%rdx), %ymm5, %ymm5
8877 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm5[0],ymm3[0],ymm5[2],ymm3[2]
8878 ; AVX512-NEXT: vinserti64x4 $0, %ymm7, %zmm14, %zmm29
8879 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
8880 ; AVX512-NEXT: vmovdqa64 %zmm0, %zmm11 {%k1}
8881 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} ymm3 = ymm5[1],ymm3[1],ymm5[3],ymm3[3]
8882 ; AVX512-NEXT: vinserti64x4 $0, %ymm3, %zmm11, %zmm26
8883 ; AVX512-NEXT: vmovdqa64 %zmm28, %zmm4 {%k1}
8884 ; AVX512-NEXT: vmovdqa 128(%rsi), %xmm3
8885 ; AVX512-NEXT: vinserti128 $1, 128(%rcx), %ymm3, %ymm3
8886 ; AVX512-NEXT: vmovdqa 128(%rdi), %xmm5
8887 ; AVX512-NEXT: vinserti128 $1, 128(%rdx), %ymm5, %ymm5
8888 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm5[0],ymm3[0],ymm5[2],ymm3[2]
8889 ; AVX512-NEXT: vinserti64x4 $0, %ymm7, %zmm4, %zmm28
8890 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
8891 ; AVX512-NEXT: vmovdqa64 %zmm0, %zmm6 {%k1}
8892 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} ymm3 = ymm5[1],ymm3[1],ymm5[3],ymm3[3]
8893 ; AVX512-NEXT: vinserti64x4 $0, %ymm3, %zmm6, %zmm18
8894 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
8895 ; AVX512-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm4 # 32-byte Folded Reload
8896 ; AVX512-NEXT: # ymm4 = ymm0[0,1,2,3],mem[4,5,6,7]
8897 ; AVX512-NEXT: vinserti64x4 $0, %ymm4, %zmm10, %zmm16
8898 ; AVX512-NEXT: vmovdqa64 %zmm12, %zmm30 {%k1}
8899 ; AVX512-NEXT: vmovdqa 192(%rsi), %xmm4
8900 ; AVX512-NEXT: vinserti128 $1, 192(%rcx), %ymm4, %ymm5
8901 ; AVX512-NEXT: vmovdqa 192(%rdi), %xmm4
8902 ; AVX512-NEXT: vinserti128 $1, 192(%rdx), %ymm4, %ymm6
8903 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} ymm4 = ymm6[0],ymm5[0],ymm6[2],ymm5[2]
8904 ; AVX512-NEXT: vinserti64x4 $0, %ymm4, %zmm30, %zmm30
8905 ; AVX512-NEXT: vmovdqa64 %zmm15, %zmm31 {%k1}
8906 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} ymm5 = ymm6[1],ymm5[1],ymm6[3],ymm5[3]
8907 ; AVX512-NEXT: vinserti64x4 $0, %ymm5, %zmm31, %zmm31
8908 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
8909 ; AVX512-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm6 # 32-byte Folded Reload
8910 ; AVX512-NEXT: # ymm6 = ymm0[0,1,2,3],mem[4,5,6,7]
8911 ; AVX512-NEXT: vmovdqa64 %zmm27, %zmm1 {%k1}
8912 ; AVX512-NEXT: vinserti64x4 $0, %ymm6, %zmm1, %zmm6
8913 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
8914 ; AVX512-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm1 # 32-byte Folded Reload
8915 ; AVX512-NEXT: # ymm1 = ymm0[0,1,2,3],mem[4,5,6,7]
8916 ; AVX512-NEXT: vmovdqa64 %zmm24, %zmm9 {%k1}
8917 ; AVX512-NEXT: vinserti64x4 $0, %ymm1, %zmm9, %zmm7
8918 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
8919 ; AVX512-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm1 # 32-byte Folded Reload
8920 ; AVX512-NEXT: # ymm1 = ymm0[0,1,2,3],mem[4,5,6,7]
8921 ; AVX512-NEXT: vinserti64x4 $0, %ymm1, %zmm22, %zmm9
8922 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
8923 ; AVX512-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm1 # 32-byte Folded Reload
8924 ; AVX512-NEXT: # ymm1 = ymm0[0,1,2,3],mem[4,5,6,7]
8925 ; AVX512-NEXT: vinserti64x4 $0, %ymm1, %zmm21, %zmm12
8926 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
8927 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
8928 ; AVX512-NEXT: vmovdqa64 %zmm0, %zmm3 {%k1}
8929 ; AVX512-NEXT: vmovdqa (%rcx), %ymm1
8930 ; AVX512-NEXT: vmovdqa (%rdx), %ymm15
8931 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm15[0],ymm1[0],ymm15[2],ymm1[2]
8932 ; AVX512-NEXT: vmovdqa64 (%rsi), %ymm21
8933 ; AVX512-NEXT: vmovdqa64 (%rdi), %ymm22
8934 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} ymm14 = ymm22[0],ymm21[0],ymm22[2],ymm21[2]
8935 ; AVX512-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm14[2,3],ymm0[2,3]
8936 ; AVX512-NEXT: vinserti64x4 $0, %ymm0, %zmm3, %zmm0
8937 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
8938 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
8939 ; AVX512-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
8940 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm15[1],ymm1[1],ymm15[3],ymm1[3]
8941 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} ymm14 = ymm22[1],ymm21[1],ymm22[3],ymm21[3]
8942 ; AVX512-NEXT: vperm2i128 {{.*#+}} ymm1 = ymm14[2,3],ymm1[2,3]
8943 ; AVX512-NEXT: vinserti64x4 $0, %ymm1, %zmm4, %zmm1
8944 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
8945 ; AVX512-NEXT: vmovdqa64 %zmm3, %zmm25 {%k1}
8946 ; AVX512-NEXT: vmovdqa 64(%rcx), %ymm14
8947 ; AVX512-NEXT: vmovdqa 64(%rdx), %ymm15
8948 ; AVX512-NEXT: vmovdqa64 64(%rsi), %ymm21
8949 ; AVX512-NEXT: vmovdqa64 64(%rdi), %ymm22
8950 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} ymm13 = ymm15[0],ymm14[0],ymm15[2],ymm14[2]
8951 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} ymm11 = ymm22[0],ymm21[0],ymm22[2],ymm21[2]
8952 ; AVX512-NEXT: vperm2i128 {{.*#+}} ymm11 = ymm11[2,3],ymm13[2,3]
8953 ; AVX512-NEXT: vinserti64x4 $0, %ymm11, %zmm25, %zmm11
8954 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
8955 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
8956 ; AVX512-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
8957 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} ymm13 = ymm15[1],ymm14[1],ymm15[3],ymm14[3]
8958 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} ymm14 = ymm22[1],ymm21[1],ymm22[3],ymm21[3]
8959 ; AVX512-NEXT: vperm2i128 {{.*#+}} ymm13 = ymm14[2,3],ymm13[2,3]
8960 ; AVX512-NEXT: vinserti64x4 $0, %ymm13, %zmm4, %zmm13
8961 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
8962 ; AVX512-NEXT: vmovdqa64 %zmm3, %zmm17 {%k1}
8963 ; AVX512-NEXT: vmovdqa 128(%rcx), %ymm14
8964 ; AVX512-NEXT: vmovdqa 128(%rdx), %ymm15
8965 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} ymm4 = ymm15[0],ymm14[0],ymm15[2],ymm14[2]
8966 ; AVX512-NEXT: vmovdqa64 128(%rsi), %ymm21
8967 ; AVX512-NEXT: vmovdqa64 128(%rdi), %ymm22
8968 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm22[0],ymm21[0],ymm22[2],ymm21[2]
8969 ; AVX512-NEXT: vperm2i128 {{.*#+}} ymm3 = ymm3[2,3],ymm4[2,3]
8970 ; AVX512-NEXT: vinserti64x4 $0, %ymm3, %zmm17, %zmm3
8971 ; AVX512-NEXT: vmovdqa64 %zmm19, %zmm20 {%k1}
8972 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} ymm4 = ymm15[1],ymm14[1],ymm15[3],ymm14[3]
8973 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} ymm14 = ymm22[1],ymm21[1],ymm22[3],ymm21[3]
8974 ; AVX512-NEXT: vperm2i128 {{.*#+}} ymm4 = ymm14[2,3],ymm4[2,3]
8975 ; AVX512-NEXT: vinserti64x4 $0, %ymm4, %zmm20, %zmm4
8976 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
8977 ; AVX512-NEXT: vmovdqa64 %zmm5, %zmm23 {%k1}
8978 ; AVX512-NEXT: vmovdqa 192(%rcx), %ymm14
8979 ; AVX512-NEXT: vmovdqa 192(%rdx), %ymm15
8980 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} ymm5 = ymm15[0],ymm14[0],ymm15[2],ymm14[2]
8981 ; AVX512-NEXT: vmovdqa64 192(%rsi), %ymm17
8982 ; AVX512-NEXT: vmovdqa64 192(%rdi), %ymm19
8983 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} ymm10 = ymm19[0],ymm17[0],ymm19[2],ymm17[2]
8984 ; AVX512-NEXT: vperm2i128 {{.*#+}} ymm5 = ymm10[2,3],ymm5[2,3]
8985 ; AVX512-NEXT: vinserti64x4 $0, %ymm5, %zmm23, %zmm5
8986 ; AVX512-NEXT: vmovdqa64 %zmm8, %zmm2 {%k1}
8987 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} ymm8 = ymm15[1],ymm14[1],ymm15[3],ymm14[3]
8988 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} ymm10 = ymm19[1],ymm17[1],ymm19[3],ymm17[3]
8989 ; AVX512-NEXT: vperm2i128 {{.*#+}} ymm8 = ymm10[2,3],ymm8[2,3]
8990 ; AVX512-NEXT: vinserti64x4 $0, %ymm8, %zmm2, %zmm2
8991 ; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %rax
8992 ; AVX512-NEXT: vmovdqa64 %zmm2, 1728(%rax)
8993 ; AVX512-NEXT: vmovdqa64 %zmm5, 1664(%rax)
8994 ; AVX512-NEXT: vmovdqa64 %zmm4, 1216(%rax)
8995 ; AVX512-NEXT: vmovdqa64 %zmm3, 1152(%rax)
8996 ; AVX512-NEXT: vmovdqa64 %zmm13, 704(%rax)
8997 ; AVX512-NEXT: vmovdqa64 %zmm11, 640(%rax)
8998 ; AVX512-NEXT: vmovdqa64 %zmm1, 192(%rax)
8999 ; AVX512-NEXT: vmovdqa64 %zmm0, 128(%rax)
9000 ; AVX512-NEXT: vmovdqa64 %zmm12, 1984(%rax)
9001 ; AVX512-NEXT: vmovdqa64 %zmm9, 1920(%rax)
9002 ; AVX512-NEXT: vmovdqa64 %zmm7, 1856(%rax)
9003 ; AVX512-NEXT: vmovdqa64 %zmm6, 1792(%rax)
9004 ; AVX512-NEXT: vmovdqa64 %zmm31, 1600(%rax)
9005 ; AVX512-NEXT: vmovdqa64 %zmm30, 1536(%rax)
9006 ; AVX512-NEXT: vmovdqa64 %zmm16, 1472(%rax)
9007 ; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
9008 ; AVX512-NEXT: vmovaps %zmm0, 1408(%rax)
9009 ; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
9010 ; AVX512-NEXT: vmovaps %zmm0, 1344(%rax)
9011 ; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
9012 ; AVX512-NEXT: vmovaps %zmm0, 1280(%rax)
9013 ; AVX512-NEXT: vmovdqa64 %zmm18, 1088(%rax)
9014 ; AVX512-NEXT: vmovdqa64 %zmm28, 1024(%rax)
9015 ; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
9016 ; AVX512-NEXT: vmovaps %zmm0, 960(%rax)
9017 ; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
9018 ; AVX512-NEXT: vmovaps %zmm0, 896(%rax)
9019 ; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
9020 ; AVX512-NEXT: vmovaps %zmm0, 832(%rax)
9021 ; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
9022 ; AVX512-NEXT: vmovaps %zmm0, 768(%rax)
9023 ; AVX512-NEXT: vmovdqa64 %zmm26, 576(%rax)
9024 ; AVX512-NEXT: vmovdqa64 %zmm29, 512(%rax)
9025 ; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
9026 ; AVX512-NEXT: vmovaps %zmm0, 448(%rax)
9027 ; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
9028 ; AVX512-NEXT: vmovaps %zmm0, 384(%rax)
9029 ; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
9030 ; AVX512-NEXT: vmovaps %zmm0, 320(%rax)
9031 ; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
9032 ; AVX512-NEXT: vmovaps %zmm0, 256(%rax)
9033 ; AVX512-NEXT: vmovups (%rsp), %zmm0 # 64-byte Reload
9034 ; AVX512-NEXT: vmovaps %zmm0, 64(%rax)
9035 ; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
9036 ; AVX512-NEXT: vmovaps %zmm0, (%rax)
9037 ; AVX512-NEXT: addq $2312, %rsp # imm = 0x908
9038 ; AVX512-NEXT: vzeroupper
9039 ; AVX512-NEXT: retq
9040 ;
9041 ; AVX512-FCP-LABEL: store_i64_stride8_vf32:
9042 ; AVX512-FCP: # %bb.0:
9043 ; AVX512-FCP-NEXT: subq $2312, %rsp # imm = 0x908
9044 ; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
9045 ; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10
9046 ; AVX512-FCP-NEXT: vmovaps 128(%rdi), %zmm0
9047 ; AVX512-FCP-NEXT: vmovups %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9048 ; AVX512-FCP-NEXT: vmovdqa64 64(%rdi), %zmm6
9049 ; AVX512-FCP-NEXT: vmovdqa64 (%rdi), %zmm2
9050 ; AVX512-FCP-NEXT: vmovdqa64 128(%rsi), %zmm0
9051 ; AVX512-FCP-NEXT: vmovdqa64 64(%rsi), %zmm5
9052 ; AVX512-FCP-NEXT: vmovdqa64 (%rsi), %zmm3
9053 ; AVX512-FCP-NEXT: vmovaps 192(%rdx), %zmm1
9054 ; AVX512-FCP-NEXT: vmovups %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9055 ; AVX512-FCP-NEXT: vmovaps 128(%rdx), %zmm1
9056 ; AVX512-FCP-NEXT: vmovups %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9057 ; AVX512-FCP-NEXT: vmovdqa64 64(%rdx), %zmm7
9058 ; AVX512-FCP-NEXT: vmovdqa64 (%rdx), %zmm4
9059 ; AVX512-FCP-NEXT: vmovdqa64 192(%rcx), %zmm23
9060 ; AVX512-FCP-NEXT: vmovdqa64 128(%rcx), %zmm1
9061 ; AVX512-FCP-NEXT: vmovdqa64 64(%rcx), %zmm9
9062 ; AVX512-FCP-NEXT: vmovdqa64 (%rcx), %zmm8
9063 ; AVX512-FCP-NEXT: vmovdqa64 (%r8), %zmm18
9064 ; AVX512-FCP-NEXT: vmovdqa64 64(%r8), %zmm16
9065 ; AVX512-FCP-NEXT: vmovdqa64 (%r9), %zmm12
9066 ; AVX512-FCP-NEXT: vmovdqa64 64(%r9), %zmm19
9067 ; AVX512-FCP-NEXT: vmovdqa64 (%r10), %zmm25
9068 ; AVX512-FCP-NEXT: vmovdqa64 64(%r10), %zmm17
9069 ; AVX512-FCP-NEXT: vmovdqa64 (%rax), %zmm15
9070 ; AVX512-FCP-NEXT: vmovdqa64 64(%rax), %zmm20
9071 ; AVX512-FCP-NEXT: movb $-64, %r11b
9072 ; AVX512-FCP-NEXT: kmovw %r11d, %k1
9073 ; AVX512-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm27 = [4,12,4,12,4,12,4,12]
9074 ; AVX512-FCP-NEXT: # zmm27 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
9075 ; AVX512-FCP-NEXT: vmovdqa64 %zmm4, %zmm10
9076 ; AVX512-FCP-NEXT: vpermt2q %zmm8, %zmm27, %zmm10
9077 ; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm11
9078 ; AVX512-FCP-NEXT: vpermt2q %zmm3, %zmm27, %zmm11
9079 ; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm10 = ymm11[0,1,2,3],ymm10[4,5,6,7]
9080 ; AVX512-FCP-NEXT: vmovdqa64 %zmm25, %zmm11
9081 ; AVX512-FCP-NEXT: vpermt2q %zmm15, %zmm27, %zmm11
9082 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm13 = zmm18[0],zmm12[0],zmm18[2],zmm12[2],zmm18[4],zmm12[4],zmm18[6],zmm12[6]
9083 ; AVX512-FCP-NEXT: vmovdqa64 %zmm11, %zmm13 {%k1}
9084 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm10, %zmm13, %zmm10
9085 ; AVX512-FCP-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9086 ; AVX512-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm24 = [5,13,5,13,5,13,5,13]
9087 ; AVX512-FCP-NEXT: # zmm24 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
9088 ; AVX512-FCP-NEXT: vmovdqa64 %zmm4, %zmm10
9089 ; AVX512-FCP-NEXT: vpermt2q %zmm8, %zmm24, %zmm10
9090 ; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm11
9091 ; AVX512-FCP-NEXT: vpermt2q %zmm3, %zmm24, %zmm11
9092 ; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm10 = ymm11[0,1,2,3],ymm10[4,5,6,7]
9093 ; AVX512-FCP-NEXT: vmovdqa64 %zmm25, %zmm11
9094 ; AVX512-FCP-NEXT: vpermt2q %zmm15, %zmm24, %zmm11
9095 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm14 = zmm18[1],zmm12[1],zmm18[3],zmm12[3],zmm18[5],zmm12[5],zmm18[7],zmm12[7]
9096 ; AVX512-FCP-NEXT: vmovdqa64 %zmm11, %zmm14 {%k1}
9097 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm10, %zmm14, %zmm10
9098 ; AVX512-FCP-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9099 ; AVX512-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm22 = [6,14,6,14,6,14,6,14]
9100 ; AVX512-FCP-NEXT: # zmm22 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
9101 ; AVX512-FCP-NEXT: vmovdqa64 %zmm4, %zmm10
9102 ; AVX512-FCP-NEXT: vpermt2q %zmm8, %zmm22, %zmm10
9103 ; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm11
9104 ; AVX512-FCP-NEXT: vpermt2q %zmm3, %zmm22, %zmm11
9105 ; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm10 = ymm11[0,1,2,3],ymm10[4,5,6,7]
9106 ; AVX512-FCP-NEXT: vmovdqa64 %zmm18, %zmm11
9107 ; AVX512-FCP-NEXT: vpermt2q %zmm12, %zmm22, %zmm11
9108 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm11 {%k1} = zmm25[0],zmm15[0],zmm25[2],zmm15[2],zmm25[4],zmm15[4],zmm25[6],zmm15[6]
9109 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm10, %zmm11, %zmm10
9110 ; AVX512-FCP-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9111 ; AVX512-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm21 = [7,15,7,15,7,15,7,15]
9112 ; AVX512-FCP-NEXT: # zmm21 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
9113 ; AVX512-FCP-NEXT: vpermt2q %zmm8, %zmm21, %zmm4
9114 ; AVX512-FCP-NEXT: vpermt2q %zmm3, %zmm21, %zmm2
9115 ; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm4[4,5,6,7]
9116 ; AVX512-FCP-NEXT: vmovdqa64 %zmm18, %zmm3
9117 ; AVX512-FCP-NEXT: vpermt2q %zmm12, %zmm21, %zmm3
9118 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm3 {%k1} = zmm25[1],zmm15[1],zmm25[3],zmm15[3],zmm25[5],zmm15[5],zmm25[7],zmm15[7]
9119 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
9120 ; AVX512-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9121 ; AVX512-FCP-NEXT: vmovdqa64 %zmm7, %zmm2
9122 ; AVX512-FCP-NEXT: vpermt2q %zmm9, %zmm27, %zmm2
9123 ; AVX512-FCP-NEXT: vmovdqa64 %zmm6, %zmm3
9124 ; AVX512-FCP-NEXT: vpermt2q %zmm5, %zmm27, %zmm3
9125 ; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3],ymm2[4,5,6,7]
9126 ; AVX512-FCP-NEXT: vmovdqa64 %zmm17, %zmm3
9127 ; AVX512-FCP-NEXT: vpermt2q %zmm20, %zmm27, %zmm3
9128 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm4 = zmm16[0],zmm19[0],zmm16[2],zmm19[2],zmm16[4],zmm19[4],zmm16[6],zmm19[6]
9129 ; AVX512-FCP-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
9130 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
9131 ; AVX512-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9132 ; AVX512-FCP-NEXT: vmovdqa64 %zmm7, %zmm2
9133 ; AVX512-FCP-NEXT: vpermt2q %zmm9, %zmm24, %zmm2
9134 ; AVX512-FCP-NEXT: vmovdqa64 %zmm6, %zmm3
9135 ; AVX512-FCP-NEXT: vpermt2q %zmm5, %zmm24, %zmm3
9136 ; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3],ymm2[4,5,6,7]
9137 ; AVX512-FCP-NEXT: vmovdqa64 %zmm17, %zmm3
9138 ; AVX512-FCP-NEXT: vpermt2q %zmm20, %zmm24, %zmm3
9139 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm4 = zmm16[1],zmm19[1],zmm16[3],zmm19[3],zmm16[5],zmm19[5],zmm16[7],zmm19[7]
9140 ; AVX512-FCP-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
9141 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
9142 ; AVX512-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9143 ; AVX512-FCP-NEXT: vmovdqa64 %zmm7, %zmm2
9144 ; AVX512-FCP-NEXT: vpermt2q %zmm9, %zmm22, %zmm2
9145 ; AVX512-FCP-NEXT: vmovdqa64 %zmm6, %zmm3
9146 ; AVX512-FCP-NEXT: vpermt2q %zmm5, %zmm22, %zmm3
9147 ; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3],ymm2[4,5,6,7]
9148 ; AVX512-FCP-NEXT: vmovdqa64 %zmm16, %zmm3
9149 ; AVX512-FCP-NEXT: vpermt2q %zmm19, %zmm22, %zmm3
9150 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm3 {%k1} = zmm17[0],zmm20[0],zmm17[2],zmm20[2],zmm17[4],zmm20[4],zmm17[6],zmm20[6]
9151 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
9152 ; AVX512-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9153 ; AVX512-FCP-NEXT: vpermt2q %zmm9, %zmm21, %zmm7
9154 ; AVX512-FCP-NEXT: vpermt2q %zmm5, %zmm21, %zmm6
9155 ; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm6[0,1,2,3],ymm7[4,5,6,7]
9156 ; AVX512-FCP-NEXT: vmovdqa64 %zmm16, %zmm3
9157 ; AVX512-FCP-NEXT: vpermt2q %zmm19, %zmm21, %zmm3
9158 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm3 {%k1} = zmm17[1],zmm20[1],zmm17[3],zmm20[3],zmm17[5],zmm20[5],zmm17[7],zmm20[7]
9159 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
9160 ; AVX512-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9161 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
9162 ; AVX512-FCP-NEXT: vmovdqa64 %zmm3, %zmm2
9163 ; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm27, %zmm2
9164 ; AVX512-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9165 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
9166 ; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm4
9167 ; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm27, %zmm4
9168 ; AVX512-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9169 ; AVX512-FCP-NEXT: vmovdqa64 %zmm3, %zmm4
9170 ; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm24, %zmm4
9171 ; AVX512-FCP-NEXT: vmovdqu64 %zmm4, (%rsp) # 64-byte Spill
9172 ; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm4
9173 ; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm24, %zmm4
9174 ; AVX512-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9175 ; AVX512-FCP-NEXT: vmovdqa64 %zmm3, %zmm4
9176 ; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm22, %zmm4
9177 ; AVX512-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9178 ; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm4
9179 ; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm22, %zmm4
9180 ; AVX512-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9181 ; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm21, %zmm3
9182 ; AVX512-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9183 ; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm21, %zmm2
9184 ; AVX512-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9185 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
9186 ; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm1
9187 ; AVX512-FCP-NEXT: vpermt2q %zmm23, %zmm27, %zmm1
9188 ; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9189 ; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm1
9190 ; AVX512-FCP-NEXT: vpermt2q %zmm23, %zmm24, %zmm1
9191 ; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9192 ; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm1
9193 ; AVX512-FCP-NEXT: vpermt2q %zmm23, %zmm22, %zmm1
9194 ; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9195 ; AVX512-FCP-NEXT: vpermt2q %zmm23, %zmm21, %zmm0
9196 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9197 ; AVX512-FCP-NEXT: vmovdqa64 192(%rdi), %zmm1
9198 ; AVX512-FCP-NEXT: vmovdqa64 192(%rsi), %zmm0
9199 ; AVX512-FCP-NEXT: vmovdqa64 %zmm1, %zmm2
9200 ; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm27, %zmm2
9201 ; AVX512-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9202 ; AVX512-FCP-NEXT: vmovdqa64 %zmm1, %zmm2
9203 ; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm24, %zmm2
9204 ; AVX512-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9205 ; AVX512-FCP-NEXT: vmovdqa64 %zmm1, %zmm2
9206 ; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm22, %zmm2
9207 ; AVX512-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9208 ; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm21, %zmm1
9209 ; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9210 ; AVX512-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm30 = [0,8,0,8,0,8,0,8]
9211 ; AVX512-FCP-NEXT: # zmm30 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
9212 ; AVX512-FCP-NEXT: vmovdqa64 %zmm25, %zmm0
9213 ; AVX512-FCP-NEXT: vpermt2q %zmm15, %zmm30, %zmm0
9214 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9215 ; AVX512-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm31 = [1,9,1,9,1,9,1,9]
9216 ; AVX512-FCP-NEXT: # zmm31 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
9217 ; AVX512-FCP-NEXT: vmovdqa64 %zmm25, %zmm0
9218 ; AVX512-FCP-NEXT: vpermt2q %zmm15, %zmm31, %zmm0
9219 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9220 ; AVX512-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm23 = [2,10,2,10,2,10,2,10]
9221 ; AVX512-FCP-NEXT: # zmm23 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
9222 ; AVX512-FCP-NEXT: vmovdqa64 %zmm25, %zmm0
9223 ; AVX512-FCP-NEXT: vpermt2q %zmm15, %zmm23, %zmm0
9224 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9225 ; AVX512-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm26 = [3,11,3,11,3,11,3,11]
9226 ; AVX512-FCP-NEXT: # zmm26 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
9227 ; AVX512-FCP-NEXT: vpermt2q %zmm15, %zmm26, %zmm25
9228 ; AVX512-FCP-NEXT: vmovdqu64 %zmm25, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9229 ; AVX512-FCP-NEXT: vmovdqa64 %zmm18, %zmm0
9230 ; AVX512-FCP-NEXT: vmovdqa64 %zmm18, %zmm29
9231 ; AVX512-FCP-NEXT: vpermt2q %zmm12, %zmm30, %zmm29
9232 ; AVX512-FCP-NEXT: vpermt2q %zmm12, %zmm31, %zmm18
9233 ; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm1
9234 ; AVX512-FCP-NEXT: vpermt2q %zmm12, %zmm23, %zmm1
9235 ; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9236 ; AVX512-FCP-NEXT: vpermt2q %zmm12, %zmm26, %zmm0
9237 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9238 ; AVX512-FCP-NEXT: vmovdqa64 %zmm17, %zmm0
9239 ; AVX512-FCP-NEXT: vpermt2q %zmm20, %zmm30, %zmm0
9240 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9241 ; AVX512-FCP-NEXT: vmovdqa64 %zmm17, %zmm0
9242 ; AVX512-FCP-NEXT: vpermt2q %zmm20, %zmm31, %zmm0
9243 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9244 ; AVX512-FCP-NEXT: vmovdqa64 %zmm17, %zmm0
9245 ; AVX512-FCP-NEXT: vpermt2q %zmm20, %zmm23, %zmm0
9246 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9247 ; AVX512-FCP-NEXT: vpermt2q %zmm20, %zmm26, %zmm17
9248 ; AVX512-FCP-NEXT: vmovdqu64 %zmm17, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9249 ; AVX512-FCP-NEXT: vmovdqa64 %zmm16, %zmm14
9250 ; AVX512-FCP-NEXT: vpermt2q %zmm19, %zmm30, %zmm14
9251 ; AVX512-FCP-NEXT: vmovdqa64 %zmm16, %zmm11
9252 ; AVX512-FCP-NEXT: vpermt2q %zmm19, %zmm31, %zmm11
9253 ; AVX512-FCP-NEXT: vmovdqa64 %zmm16, %zmm25
9254 ; AVX512-FCP-NEXT: vpermt2q %zmm19, %zmm23, %zmm25
9255 ; AVX512-FCP-NEXT: vpermt2q %zmm19, %zmm26, %zmm16
9256 ; AVX512-FCP-NEXT: vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9257 ; AVX512-FCP-NEXT: vmovdqa64 128(%r10), %zmm19
9258 ; AVX512-FCP-NEXT: vmovdqa64 128(%rax), %zmm0
9259 ; AVX512-FCP-NEXT: vmovdqa64 %zmm19, %zmm28
9260 ; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm30, %zmm28
9261 ; AVX512-FCP-NEXT: vmovdqa64 %zmm19, %zmm1
9262 ; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm31, %zmm1
9263 ; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9264 ; AVX512-FCP-NEXT: vmovdqa64 %zmm19, %zmm13
9265 ; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm27, %zmm13
9266 ; AVX512-FCP-NEXT: vmovdqa64 %zmm19, %zmm16
9267 ; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm24, %zmm16
9268 ; AVX512-FCP-NEXT: vmovdqa64 128(%r8), %zmm20
9269 ; AVX512-FCP-NEXT: vmovdqa64 128(%r9), %zmm1
9270 ; AVX512-FCP-NEXT: vmovdqa64 %zmm20, %zmm3
9271 ; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm22, %zmm3
9272 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm3 {%k1} = zmm19[0],zmm0[0],zmm19[2],zmm0[2],zmm19[4],zmm0[4],zmm19[6],zmm0[6]
9273 ; AVX512-FCP-NEXT: vmovdqa64 %zmm20, %zmm10
9274 ; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm21, %zmm10
9275 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm10 {%k1} = zmm19[1],zmm0[1],zmm19[3],zmm0[3],zmm19[5],zmm0[5],zmm19[7],zmm0[7]
9276 ; AVX512-FCP-NEXT: vmovdqa64 %zmm19, %zmm2
9277 ; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm23, %zmm2
9278 ; AVX512-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9279 ; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm26, %zmm19
9280 ; AVX512-FCP-NEXT: vmovdqa64 %zmm20, %zmm4
9281 ; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm30, %zmm4
9282 ; AVX512-FCP-NEXT: vmovdqa64 %zmm20, %zmm6
9283 ; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm31, %zmm6
9284 ; AVX512-FCP-NEXT: vmovdqa64 %zmm20, %zmm17
9285 ; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm23, %zmm17
9286 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm5 = zmm20[0],zmm1[0],zmm20[2],zmm1[2],zmm20[4],zmm1[4],zmm20[6],zmm1[6]
9287 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm7 = zmm20[1],zmm1[1],zmm20[3],zmm1[3],zmm20[5],zmm1[5],zmm20[7],zmm1[7]
9288 ; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm26, %zmm20
9289 ; AVX512-FCP-NEXT: vmovdqa64 192(%r10), %zmm8
9290 ; AVX512-FCP-NEXT: vmovdqa64 192(%rax), %zmm1
9291 ; AVX512-FCP-NEXT: vmovdqa64 %zmm8, %zmm12
9292 ; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm30, %zmm12
9293 ; AVX512-FCP-NEXT: vmovdqa64 %zmm8, %zmm15
9294 ; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm31, %zmm15
9295 ; AVX512-FCP-NEXT: vpermi2q %zmm1, %zmm8, %zmm27
9296 ; AVX512-FCP-NEXT: vpermi2q %zmm1, %zmm8, %zmm24
9297 ; AVX512-FCP-NEXT: vmovdqa64 192(%r8), %zmm2
9298 ; AVX512-FCP-NEXT: vmovdqa64 192(%r9), %zmm0
9299 ; AVX512-FCP-NEXT: vpermi2q %zmm0, %zmm2, %zmm22
9300 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm22 {%k1} = zmm8[0],zmm1[0],zmm8[2],zmm1[2],zmm8[4],zmm1[4],zmm8[6],zmm1[6]
9301 ; AVX512-FCP-NEXT: vpermi2q %zmm0, %zmm2, %zmm21
9302 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm21 {%k1} = zmm8[1],zmm1[1],zmm8[3],zmm1[3],zmm8[5],zmm1[5],zmm8[7],zmm1[7]
9303 ; AVX512-FCP-NEXT: vmovdqa64 %zmm8, %zmm9
9304 ; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm23, %zmm9
9305 ; AVX512-FCP-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9306 ; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm26, %zmm8
9307 ; AVX512-FCP-NEXT: vpermi2q %zmm0, %zmm2, %zmm30
9308 ; AVX512-FCP-NEXT: vpermi2q %zmm0, %zmm2, %zmm31
9309 ; AVX512-FCP-NEXT: vpermi2q %zmm0, %zmm2, %zmm23
9310 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm1 = zmm2[0],zmm0[0],zmm2[2],zmm0[2],zmm2[4],zmm0[4],zmm2[6],zmm0[6]
9311 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm9 = zmm2[1],zmm0[1],zmm2[3],zmm0[3],zmm2[5],zmm0[5],zmm2[7],zmm0[7]
9312 ; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm26, %zmm2
9313 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
9314 ; AVX512-FCP-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
9315 ; AVX512-FCP-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
9316 ; AVX512-FCP-NEXT: vmovdqa64 %zmm13, %zmm5 {%k1}
9317 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm5, %zmm0
9318 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9319 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
9320 ; AVX512-FCP-NEXT: vpblendd $240, (%rsp), %ymm0, %ymm0 # 32-byte Folded Reload
9321 ; AVX512-FCP-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
9322 ; AVX512-FCP-NEXT: vmovdqa64 %zmm16, %zmm7 {%k1}
9323 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm7, %zmm0
9324 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9325 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
9326 ; AVX512-FCP-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
9327 ; AVX512-FCP-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
9328 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm3, %zmm0
9329 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9330 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
9331 ; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm29 {%k1}
9332 ; AVX512-FCP-NEXT: vmovdqa (%rsi), %xmm0
9333 ; AVX512-FCP-NEXT: vinserti128 $1, (%rcx), %ymm0, %ymm3
9334 ; AVX512-FCP-NEXT: vmovdqa (%rdi), %xmm0
9335 ; AVX512-FCP-NEXT: vinserti128 $1, (%rdx), %ymm0, %ymm5
9336 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm5[0],ymm3[0],ymm5[2],ymm3[2]
9337 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm29, %zmm0
9338 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9339 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
9340 ; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm18 {%k1}
9341 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm3 = ymm5[1],ymm3[1],ymm5[3],ymm3[3]
9342 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm3, %zmm18, %zmm0
9343 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, (%rsp) # 64-byte Spill
9344 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
9345 ; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm14 {%k1}
9346 ; AVX512-FCP-NEXT: vmovdqa 64(%rsi), %xmm3
9347 ; AVX512-FCP-NEXT: vinserti128 $1, 64(%rcx), %ymm3, %ymm3
9348 ; AVX512-FCP-NEXT: vmovdqa 64(%rdi), %xmm5
9349 ; AVX512-FCP-NEXT: vinserti128 $1, 64(%rdx), %ymm5, %ymm5
9350 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm5[0],ymm3[0],ymm5[2],ymm3[2]
9351 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm7, %zmm14, %zmm29
9352 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
9353 ; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm11 {%k1}
9354 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm3 = ymm5[1],ymm3[1],ymm5[3],ymm3[3]
9355 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm3, %zmm11, %zmm26
9356 ; AVX512-FCP-NEXT: vmovdqa64 %zmm28, %zmm4 {%k1}
9357 ; AVX512-FCP-NEXT: vmovdqa 128(%rsi), %xmm3
9358 ; AVX512-FCP-NEXT: vinserti128 $1, 128(%rcx), %ymm3, %ymm3
9359 ; AVX512-FCP-NEXT: vmovdqa 128(%rdi), %xmm5
9360 ; AVX512-FCP-NEXT: vinserti128 $1, 128(%rdx), %ymm5, %ymm5
9361 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm5[0],ymm3[0],ymm5[2],ymm3[2]
9362 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm7, %zmm4, %zmm28
9363 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
9364 ; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm6 {%k1}
9365 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm3 = ymm5[1],ymm3[1],ymm5[3],ymm3[3]
9366 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm3, %zmm6, %zmm18
9367 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
9368 ; AVX512-FCP-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm4 # 32-byte Folded Reload
9369 ; AVX512-FCP-NEXT: # ymm4 = ymm0[0,1,2,3],mem[4,5,6,7]
9370 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm4, %zmm10, %zmm16
9371 ; AVX512-FCP-NEXT: vmovdqa64 %zmm12, %zmm30 {%k1}
9372 ; AVX512-FCP-NEXT: vmovdqa 192(%rsi), %xmm4
9373 ; AVX512-FCP-NEXT: vinserti128 $1, 192(%rcx), %ymm4, %ymm5
9374 ; AVX512-FCP-NEXT: vmovdqa 192(%rdi), %xmm4
9375 ; AVX512-FCP-NEXT: vinserti128 $1, 192(%rdx), %ymm4, %ymm6
9376 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm4 = ymm6[0],ymm5[0],ymm6[2],ymm5[2]
9377 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm4, %zmm30, %zmm30
9378 ; AVX512-FCP-NEXT: vmovdqa64 %zmm15, %zmm31 {%k1}
9379 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm5 = ymm6[1],ymm5[1],ymm6[3],ymm5[3]
9380 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm5, %zmm31, %zmm31
9381 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
9382 ; AVX512-FCP-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm6 # 32-byte Folded Reload
9383 ; AVX512-FCP-NEXT: # ymm6 = ymm0[0,1,2,3],mem[4,5,6,7]
9384 ; AVX512-FCP-NEXT: vmovdqa64 %zmm27, %zmm1 {%k1}
9385 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm1, %zmm6
9386 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
9387 ; AVX512-FCP-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm1 # 32-byte Folded Reload
9388 ; AVX512-FCP-NEXT: # ymm1 = ymm0[0,1,2,3],mem[4,5,6,7]
9389 ; AVX512-FCP-NEXT: vmovdqa64 %zmm24, %zmm9 {%k1}
9390 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm1, %zmm9, %zmm7
9391 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
9392 ; AVX512-FCP-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm1 # 32-byte Folded Reload
9393 ; AVX512-FCP-NEXT: # ymm1 = ymm0[0,1,2,3],mem[4,5,6,7]
9394 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm1, %zmm22, %zmm9
9395 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
9396 ; AVX512-FCP-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm1 # 32-byte Folded Reload
9397 ; AVX512-FCP-NEXT: # ymm1 = ymm0[0,1,2,3],mem[4,5,6,7]
9398 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm1, %zmm21, %zmm12
9399 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
9400 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
9401 ; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm3 {%k1}
9402 ; AVX512-FCP-NEXT: vmovdqa (%rcx), %ymm1
9403 ; AVX512-FCP-NEXT: vmovdqa (%rdx), %ymm15
9404 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm15[0],ymm1[0],ymm15[2],ymm1[2]
9405 ; AVX512-FCP-NEXT: vmovdqa64 (%rsi), %ymm21
9406 ; AVX512-FCP-NEXT: vmovdqa64 (%rdi), %ymm22
9407 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm14 = ymm22[0],ymm21[0],ymm22[2],ymm21[2]
9408 ; AVX512-FCP-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm14[2,3],ymm0[2,3]
9409 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm3, %zmm0
9410 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
9411 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
9412 ; AVX512-FCP-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
9413 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm15[1],ymm1[1],ymm15[3],ymm1[3]
9414 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm14 = ymm22[1],ymm21[1],ymm22[3],ymm21[3]
9415 ; AVX512-FCP-NEXT: vperm2i128 {{.*#+}} ymm1 = ymm14[2,3],ymm1[2,3]
9416 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm1, %zmm4, %zmm1
9417 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
9418 ; AVX512-FCP-NEXT: vmovdqa64 %zmm3, %zmm25 {%k1}
9419 ; AVX512-FCP-NEXT: vmovdqa 64(%rcx), %ymm14
9420 ; AVX512-FCP-NEXT: vmovdqa 64(%rdx), %ymm15
9421 ; AVX512-FCP-NEXT: vmovdqa64 64(%rsi), %ymm21
9422 ; AVX512-FCP-NEXT: vmovdqa64 64(%rdi), %ymm22
9423 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm13 = ymm15[0],ymm14[0],ymm15[2],ymm14[2]
9424 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm11 = ymm22[0],ymm21[0],ymm22[2],ymm21[2]
9425 ; AVX512-FCP-NEXT: vperm2i128 {{.*#+}} ymm11 = ymm11[2,3],ymm13[2,3]
9426 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm11, %zmm25, %zmm11
9427 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
9428 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
9429 ; AVX512-FCP-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
9430 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm13 = ymm15[1],ymm14[1],ymm15[3],ymm14[3]
9431 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm14 = ymm22[1],ymm21[1],ymm22[3],ymm21[3]
9432 ; AVX512-FCP-NEXT: vperm2i128 {{.*#+}} ymm13 = ymm14[2,3],ymm13[2,3]
9433 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm13, %zmm4, %zmm13
9434 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
9435 ; AVX512-FCP-NEXT: vmovdqa64 %zmm3, %zmm17 {%k1}
9436 ; AVX512-FCP-NEXT: vmovdqa 128(%rcx), %ymm14
9437 ; AVX512-FCP-NEXT: vmovdqa 128(%rdx), %ymm15
9438 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm4 = ymm15[0],ymm14[0],ymm15[2],ymm14[2]
9439 ; AVX512-FCP-NEXT: vmovdqa64 128(%rsi), %ymm21
9440 ; AVX512-FCP-NEXT: vmovdqa64 128(%rdi), %ymm22
9441 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm22[0],ymm21[0],ymm22[2],ymm21[2]
9442 ; AVX512-FCP-NEXT: vperm2i128 {{.*#+}} ymm3 = ymm3[2,3],ymm4[2,3]
9443 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm3, %zmm17, %zmm3
9444 ; AVX512-FCP-NEXT: vmovdqa64 %zmm19, %zmm20 {%k1}
9445 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm4 = ymm15[1],ymm14[1],ymm15[3],ymm14[3]
9446 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm14 = ymm22[1],ymm21[1],ymm22[3],ymm21[3]
9447 ; AVX512-FCP-NEXT: vperm2i128 {{.*#+}} ymm4 = ymm14[2,3],ymm4[2,3]
9448 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm4, %zmm20, %zmm4
9449 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
9450 ; AVX512-FCP-NEXT: vmovdqa64 %zmm5, %zmm23 {%k1}
9451 ; AVX512-FCP-NEXT: vmovdqa 192(%rcx), %ymm14
9452 ; AVX512-FCP-NEXT: vmovdqa 192(%rdx), %ymm15
9453 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm5 = ymm15[0],ymm14[0],ymm15[2],ymm14[2]
9454 ; AVX512-FCP-NEXT: vmovdqa64 192(%rsi), %ymm17
9455 ; AVX512-FCP-NEXT: vmovdqa64 192(%rdi), %ymm19
9456 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm10 = ymm19[0],ymm17[0],ymm19[2],ymm17[2]
9457 ; AVX512-FCP-NEXT: vperm2i128 {{.*#+}} ymm5 = ymm10[2,3],ymm5[2,3]
9458 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm5, %zmm23, %zmm5
9459 ; AVX512-FCP-NEXT: vmovdqa64 %zmm8, %zmm2 {%k1}
9460 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm8 = ymm15[1],ymm14[1],ymm15[3],ymm14[3]
9461 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm10 = ymm19[1],ymm17[1],ymm19[3],ymm17[3]
9462 ; AVX512-FCP-NEXT: vperm2i128 {{.*#+}} ymm8 = ymm10[2,3],ymm8[2,3]
9463 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm8, %zmm2, %zmm2
9464 ; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
9465 ; AVX512-FCP-NEXT: vmovdqa64 %zmm2, 1728(%rax)
9466 ; AVX512-FCP-NEXT: vmovdqa64 %zmm5, 1664(%rax)
9467 ; AVX512-FCP-NEXT: vmovdqa64 %zmm4, 1216(%rax)
9468 ; AVX512-FCP-NEXT: vmovdqa64 %zmm3, 1152(%rax)
9469 ; AVX512-FCP-NEXT: vmovdqa64 %zmm13, 704(%rax)
9470 ; AVX512-FCP-NEXT: vmovdqa64 %zmm11, 640(%rax)
9471 ; AVX512-FCP-NEXT: vmovdqa64 %zmm1, 192(%rax)
9472 ; AVX512-FCP-NEXT: vmovdqa64 %zmm0, 128(%rax)
9473 ; AVX512-FCP-NEXT: vmovdqa64 %zmm12, 1984(%rax)
9474 ; AVX512-FCP-NEXT: vmovdqa64 %zmm9, 1920(%rax)
9475 ; AVX512-FCP-NEXT: vmovdqa64 %zmm7, 1856(%rax)
9476 ; AVX512-FCP-NEXT: vmovdqa64 %zmm6, 1792(%rax)
9477 ; AVX512-FCP-NEXT: vmovdqa64 %zmm31, 1600(%rax)
9478 ; AVX512-FCP-NEXT: vmovdqa64 %zmm30, 1536(%rax)
9479 ; AVX512-FCP-NEXT: vmovdqa64 %zmm16, 1472(%rax)
9480 ; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
9481 ; AVX512-FCP-NEXT: vmovaps %zmm0, 1408(%rax)
9482 ; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
9483 ; AVX512-FCP-NEXT: vmovaps %zmm0, 1344(%rax)
9484 ; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
9485 ; AVX512-FCP-NEXT: vmovaps %zmm0, 1280(%rax)
9486 ; AVX512-FCP-NEXT: vmovdqa64 %zmm18, 1088(%rax)
9487 ; AVX512-FCP-NEXT: vmovdqa64 %zmm28, 1024(%rax)
9488 ; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
9489 ; AVX512-FCP-NEXT: vmovaps %zmm0, 960(%rax)
9490 ; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
9491 ; AVX512-FCP-NEXT: vmovaps %zmm0, 896(%rax)
9492 ; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
9493 ; AVX512-FCP-NEXT: vmovaps %zmm0, 832(%rax)
9494 ; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
9495 ; AVX512-FCP-NEXT: vmovaps %zmm0, 768(%rax)
9496 ; AVX512-FCP-NEXT: vmovdqa64 %zmm26, 576(%rax)
9497 ; AVX512-FCP-NEXT: vmovdqa64 %zmm29, 512(%rax)
9498 ; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
9499 ; AVX512-FCP-NEXT: vmovaps %zmm0, 448(%rax)
9500 ; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
9501 ; AVX512-FCP-NEXT: vmovaps %zmm0, 384(%rax)
9502 ; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
9503 ; AVX512-FCP-NEXT: vmovaps %zmm0, 320(%rax)
9504 ; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
9505 ; AVX512-FCP-NEXT: vmovaps %zmm0, 256(%rax)
9506 ; AVX512-FCP-NEXT: vmovups (%rsp), %zmm0 # 64-byte Reload
9507 ; AVX512-FCP-NEXT: vmovaps %zmm0, 64(%rax)
9508 ; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
9509 ; AVX512-FCP-NEXT: vmovaps %zmm0, (%rax)
9510 ; AVX512-FCP-NEXT: addq $2312, %rsp # imm = 0x908
9511 ; AVX512-FCP-NEXT: vzeroupper
9512 ; AVX512-FCP-NEXT: retq
9513 ;
9514 ; AVX512DQ-LABEL: store_i64_stride8_vf32:
9515 ; AVX512DQ: # %bb.0:
9516 ; AVX512DQ-NEXT: subq $2312, %rsp # imm = 0x908
9517 ; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %rax
9518 ; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %r10
9519 ; AVX512DQ-NEXT: vmovaps 128(%rdi), %zmm0
9520 ; AVX512DQ-NEXT: vmovups %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9521 ; AVX512DQ-NEXT: vmovdqa64 64(%rdi), %zmm6
9522 ; AVX512DQ-NEXT: vmovdqa64 (%rdi), %zmm2
9523 ; AVX512DQ-NEXT: vmovdqa64 128(%rsi), %zmm0
9524 ; AVX512DQ-NEXT: vmovdqa64 64(%rsi), %zmm5
9525 ; AVX512DQ-NEXT: vmovdqa64 (%rsi), %zmm3
9526 ; AVX512DQ-NEXT: vmovaps 192(%rdx), %zmm1
9527 ; AVX512DQ-NEXT: vmovups %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9528 ; AVX512DQ-NEXT: vmovaps 128(%rdx), %zmm1
9529 ; AVX512DQ-NEXT: vmovups %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9530 ; AVX512DQ-NEXT: vmovdqa64 64(%rdx), %zmm7
9531 ; AVX512DQ-NEXT: vmovdqa64 (%rdx), %zmm4
9532 ; AVX512DQ-NEXT: vmovdqa64 192(%rcx), %zmm23
9533 ; AVX512DQ-NEXT: vmovdqa64 128(%rcx), %zmm1
9534 ; AVX512DQ-NEXT: vmovdqa64 64(%rcx), %zmm9
9535 ; AVX512DQ-NEXT: vmovdqa64 (%rcx), %zmm8
9536 ; AVX512DQ-NEXT: vmovdqa64 (%r8), %zmm18
9537 ; AVX512DQ-NEXT: vmovdqa64 64(%r8), %zmm16
9538 ; AVX512DQ-NEXT: vmovdqa64 (%r9), %zmm12
9539 ; AVX512DQ-NEXT: vmovdqa64 64(%r9), %zmm19
9540 ; AVX512DQ-NEXT: vmovdqa64 (%r10), %zmm25
9541 ; AVX512DQ-NEXT: vmovdqa64 64(%r10), %zmm17
9542 ; AVX512DQ-NEXT: vmovdqa64 (%rax), %zmm15
9543 ; AVX512DQ-NEXT: vmovdqa64 64(%rax), %zmm20
9544 ; AVX512DQ-NEXT: movb $-64, %r11b
9545 ; AVX512DQ-NEXT: kmovw %r11d, %k1
9546 ; AVX512DQ-NEXT: vbroadcasti32x4 {{.*#+}} zmm27 = [4,12,4,12,4,12,4,12]
9547 ; AVX512DQ-NEXT: # zmm27 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
9548 ; AVX512DQ-NEXT: vmovdqa64 %zmm4, %zmm10
9549 ; AVX512DQ-NEXT: vpermt2q %zmm8, %zmm27, %zmm10
9550 ; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm11
9551 ; AVX512DQ-NEXT: vpermt2q %zmm3, %zmm27, %zmm11
9552 ; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm10 = ymm11[0,1,2,3],ymm10[4,5,6,7]
9553 ; AVX512DQ-NEXT: vmovdqa64 %zmm25, %zmm11
9554 ; AVX512DQ-NEXT: vpermt2q %zmm15, %zmm27, %zmm11
9555 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} zmm13 = zmm18[0],zmm12[0],zmm18[2],zmm12[2],zmm18[4],zmm12[4],zmm18[6],zmm12[6]
9556 ; AVX512DQ-NEXT: vmovdqa64 %zmm11, %zmm13 {%k1}
9557 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm10, %zmm13, %zmm10
9558 ; AVX512DQ-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9559 ; AVX512DQ-NEXT: vbroadcasti32x4 {{.*#+}} zmm24 = [5,13,5,13,5,13,5,13]
9560 ; AVX512DQ-NEXT: # zmm24 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
9561 ; AVX512DQ-NEXT: vmovdqa64 %zmm4, %zmm10
9562 ; AVX512DQ-NEXT: vpermt2q %zmm8, %zmm24, %zmm10
9563 ; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm11
9564 ; AVX512DQ-NEXT: vpermt2q %zmm3, %zmm24, %zmm11
9565 ; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm10 = ymm11[0,1,2,3],ymm10[4,5,6,7]
9566 ; AVX512DQ-NEXT: vmovdqa64 %zmm25, %zmm11
9567 ; AVX512DQ-NEXT: vpermt2q %zmm15, %zmm24, %zmm11
9568 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} zmm14 = zmm18[1],zmm12[1],zmm18[3],zmm12[3],zmm18[5],zmm12[5],zmm18[7],zmm12[7]
9569 ; AVX512DQ-NEXT: vmovdqa64 %zmm11, %zmm14 {%k1}
9570 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm10, %zmm14, %zmm10
9571 ; AVX512DQ-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9572 ; AVX512DQ-NEXT: vbroadcasti32x4 {{.*#+}} zmm22 = [6,14,6,14,6,14,6,14]
9573 ; AVX512DQ-NEXT: # zmm22 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
9574 ; AVX512DQ-NEXT: vmovdqa64 %zmm4, %zmm10
9575 ; AVX512DQ-NEXT: vpermt2q %zmm8, %zmm22, %zmm10
9576 ; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm11
9577 ; AVX512DQ-NEXT: vpermt2q %zmm3, %zmm22, %zmm11
9578 ; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm10 = ymm11[0,1,2,3],ymm10[4,5,6,7]
9579 ; AVX512DQ-NEXT: vmovdqa64 %zmm18, %zmm11
9580 ; AVX512DQ-NEXT: vpermt2q %zmm12, %zmm22, %zmm11
9581 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} zmm11 {%k1} = zmm25[0],zmm15[0],zmm25[2],zmm15[2],zmm25[4],zmm15[4],zmm25[6],zmm15[6]
9582 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm10, %zmm11, %zmm10
9583 ; AVX512DQ-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9584 ; AVX512DQ-NEXT: vbroadcasti32x4 {{.*#+}} zmm21 = [7,15,7,15,7,15,7,15]
9585 ; AVX512DQ-NEXT: # zmm21 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
9586 ; AVX512DQ-NEXT: vpermt2q %zmm8, %zmm21, %zmm4
9587 ; AVX512DQ-NEXT: vpermt2q %zmm3, %zmm21, %zmm2
9588 ; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm4[4,5,6,7]
9589 ; AVX512DQ-NEXT: vmovdqa64 %zmm18, %zmm3
9590 ; AVX512DQ-NEXT: vpermt2q %zmm12, %zmm21, %zmm3
9591 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} zmm3 {%k1} = zmm25[1],zmm15[1],zmm25[3],zmm15[3],zmm25[5],zmm15[5],zmm25[7],zmm15[7]
9592 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
9593 ; AVX512DQ-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9594 ; AVX512DQ-NEXT: vmovdqa64 %zmm7, %zmm2
9595 ; AVX512DQ-NEXT: vpermt2q %zmm9, %zmm27, %zmm2
9596 ; AVX512DQ-NEXT: vmovdqa64 %zmm6, %zmm3
9597 ; AVX512DQ-NEXT: vpermt2q %zmm5, %zmm27, %zmm3
9598 ; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3],ymm2[4,5,6,7]
9599 ; AVX512DQ-NEXT: vmovdqa64 %zmm17, %zmm3
9600 ; AVX512DQ-NEXT: vpermt2q %zmm20, %zmm27, %zmm3
9601 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} zmm4 = zmm16[0],zmm19[0],zmm16[2],zmm19[2],zmm16[4],zmm19[4],zmm16[6],zmm19[6]
9602 ; AVX512DQ-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
9603 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
9604 ; AVX512DQ-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9605 ; AVX512DQ-NEXT: vmovdqa64 %zmm7, %zmm2
9606 ; AVX512DQ-NEXT: vpermt2q %zmm9, %zmm24, %zmm2
9607 ; AVX512DQ-NEXT: vmovdqa64 %zmm6, %zmm3
9608 ; AVX512DQ-NEXT: vpermt2q %zmm5, %zmm24, %zmm3
9609 ; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3],ymm2[4,5,6,7]
9610 ; AVX512DQ-NEXT: vmovdqa64 %zmm17, %zmm3
9611 ; AVX512DQ-NEXT: vpermt2q %zmm20, %zmm24, %zmm3
9612 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} zmm4 = zmm16[1],zmm19[1],zmm16[3],zmm19[3],zmm16[5],zmm19[5],zmm16[7],zmm19[7]
9613 ; AVX512DQ-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
9614 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
9615 ; AVX512DQ-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9616 ; AVX512DQ-NEXT: vmovdqa64 %zmm7, %zmm2
9617 ; AVX512DQ-NEXT: vpermt2q %zmm9, %zmm22, %zmm2
9618 ; AVX512DQ-NEXT: vmovdqa64 %zmm6, %zmm3
9619 ; AVX512DQ-NEXT: vpermt2q %zmm5, %zmm22, %zmm3
9620 ; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3],ymm2[4,5,6,7]
9621 ; AVX512DQ-NEXT: vmovdqa64 %zmm16, %zmm3
9622 ; AVX512DQ-NEXT: vpermt2q %zmm19, %zmm22, %zmm3
9623 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} zmm3 {%k1} = zmm17[0],zmm20[0],zmm17[2],zmm20[2],zmm17[4],zmm20[4],zmm17[6],zmm20[6]
9624 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
9625 ; AVX512DQ-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9626 ; AVX512DQ-NEXT: vpermt2q %zmm9, %zmm21, %zmm7
9627 ; AVX512DQ-NEXT: vpermt2q %zmm5, %zmm21, %zmm6
9628 ; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm2 = ymm6[0,1,2,3],ymm7[4,5,6,7]
9629 ; AVX512DQ-NEXT: vmovdqa64 %zmm16, %zmm3
9630 ; AVX512DQ-NEXT: vpermt2q %zmm19, %zmm21, %zmm3
9631 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} zmm3 {%k1} = zmm17[1],zmm20[1],zmm17[3],zmm20[3],zmm17[5],zmm20[5],zmm17[7],zmm20[7]
9632 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
9633 ; AVX512DQ-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9634 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
9635 ; AVX512DQ-NEXT: vmovdqa64 %zmm3, %zmm2
9636 ; AVX512DQ-NEXT: vpermt2q %zmm1, %zmm27, %zmm2
9637 ; AVX512DQ-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9638 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
9639 ; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm4
9640 ; AVX512DQ-NEXT: vpermt2q %zmm0, %zmm27, %zmm4
9641 ; AVX512DQ-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9642 ; AVX512DQ-NEXT: vmovdqa64 %zmm3, %zmm4
9643 ; AVX512DQ-NEXT: vpermt2q %zmm1, %zmm24, %zmm4
9644 ; AVX512DQ-NEXT: vmovdqu64 %zmm4, (%rsp) # 64-byte Spill
9645 ; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm4
9646 ; AVX512DQ-NEXT: vpermt2q %zmm0, %zmm24, %zmm4
9647 ; AVX512DQ-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9648 ; AVX512DQ-NEXT: vmovdqa64 %zmm3, %zmm4
9649 ; AVX512DQ-NEXT: vpermt2q %zmm1, %zmm22, %zmm4
9650 ; AVX512DQ-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9651 ; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm4
9652 ; AVX512DQ-NEXT: vpermt2q %zmm0, %zmm22, %zmm4
9653 ; AVX512DQ-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9654 ; AVX512DQ-NEXT: vpermt2q %zmm1, %zmm21, %zmm3
9655 ; AVX512DQ-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9656 ; AVX512DQ-NEXT: vpermt2q %zmm0, %zmm21, %zmm2
9657 ; AVX512DQ-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9658 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
9659 ; AVX512DQ-NEXT: vmovdqa64 %zmm0, %zmm1
9660 ; AVX512DQ-NEXT: vpermt2q %zmm23, %zmm27, %zmm1
9661 ; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9662 ; AVX512DQ-NEXT: vmovdqa64 %zmm0, %zmm1
9663 ; AVX512DQ-NEXT: vpermt2q %zmm23, %zmm24, %zmm1
9664 ; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9665 ; AVX512DQ-NEXT: vmovdqa64 %zmm0, %zmm1
9666 ; AVX512DQ-NEXT: vpermt2q %zmm23, %zmm22, %zmm1
9667 ; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9668 ; AVX512DQ-NEXT: vpermt2q %zmm23, %zmm21, %zmm0
9669 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9670 ; AVX512DQ-NEXT: vmovdqa64 192(%rdi), %zmm1
9671 ; AVX512DQ-NEXT: vmovdqa64 192(%rsi), %zmm0
9672 ; AVX512DQ-NEXT: vmovdqa64 %zmm1, %zmm2
9673 ; AVX512DQ-NEXT: vpermt2q %zmm0, %zmm27, %zmm2
9674 ; AVX512DQ-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9675 ; AVX512DQ-NEXT: vmovdqa64 %zmm1, %zmm2
9676 ; AVX512DQ-NEXT: vpermt2q %zmm0, %zmm24, %zmm2
9677 ; AVX512DQ-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9678 ; AVX512DQ-NEXT: vmovdqa64 %zmm1, %zmm2
9679 ; AVX512DQ-NEXT: vpermt2q %zmm0, %zmm22, %zmm2
9680 ; AVX512DQ-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9681 ; AVX512DQ-NEXT: vpermt2q %zmm0, %zmm21, %zmm1
9682 ; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9683 ; AVX512DQ-NEXT: vbroadcasti32x4 {{.*#+}} zmm30 = [0,8,0,8,0,8,0,8]
9684 ; AVX512DQ-NEXT: # zmm30 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
9685 ; AVX512DQ-NEXT: vmovdqa64 %zmm25, %zmm0
9686 ; AVX512DQ-NEXT: vpermt2q %zmm15, %zmm30, %zmm0
9687 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9688 ; AVX512DQ-NEXT: vbroadcasti32x4 {{.*#+}} zmm31 = [1,9,1,9,1,9,1,9]
9689 ; AVX512DQ-NEXT: # zmm31 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
9690 ; AVX512DQ-NEXT: vmovdqa64 %zmm25, %zmm0
9691 ; AVX512DQ-NEXT: vpermt2q %zmm15, %zmm31, %zmm0
9692 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9693 ; AVX512DQ-NEXT: vbroadcasti32x4 {{.*#+}} zmm23 = [2,10,2,10,2,10,2,10]
9694 ; AVX512DQ-NEXT: # zmm23 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
9695 ; AVX512DQ-NEXT: vmovdqa64 %zmm25, %zmm0
9696 ; AVX512DQ-NEXT: vpermt2q %zmm15, %zmm23, %zmm0
9697 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9698 ; AVX512DQ-NEXT: vbroadcasti32x4 {{.*#+}} zmm26 = [3,11,3,11,3,11,3,11]
9699 ; AVX512DQ-NEXT: # zmm26 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
9700 ; AVX512DQ-NEXT: vpermt2q %zmm15, %zmm26, %zmm25
9701 ; AVX512DQ-NEXT: vmovdqu64 %zmm25, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9702 ; AVX512DQ-NEXT: vmovdqa64 %zmm18, %zmm0
9703 ; AVX512DQ-NEXT: vmovdqa64 %zmm18, %zmm29
9704 ; AVX512DQ-NEXT: vpermt2q %zmm12, %zmm30, %zmm29
9705 ; AVX512DQ-NEXT: vpermt2q %zmm12, %zmm31, %zmm18
9706 ; AVX512DQ-NEXT: vmovdqa64 %zmm0, %zmm1
9707 ; AVX512DQ-NEXT: vpermt2q %zmm12, %zmm23, %zmm1
9708 ; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9709 ; AVX512DQ-NEXT: vpermt2q %zmm12, %zmm26, %zmm0
9710 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9711 ; AVX512DQ-NEXT: vmovdqa64 %zmm17, %zmm0
9712 ; AVX512DQ-NEXT: vpermt2q %zmm20, %zmm30, %zmm0
9713 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9714 ; AVX512DQ-NEXT: vmovdqa64 %zmm17, %zmm0
9715 ; AVX512DQ-NEXT: vpermt2q %zmm20, %zmm31, %zmm0
9716 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9717 ; AVX512DQ-NEXT: vmovdqa64 %zmm17, %zmm0
9718 ; AVX512DQ-NEXT: vpermt2q %zmm20, %zmm23, %zmm0
9719 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9720 ; AVX512DQ-NEXT: vpermt2q %zmm20, %zmm26, %zmm17
9721 ; AVX512DQ-NEXT: vmovdqu64 %zmm17, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9722 ; AVX512DQ-NEXT: vmovdqa64 %zmm16, %zmm14
9723 ; AVX512DQ-NEXT: vpermt2q %zmm19, %zmm30, %zmm14
9724 ; AVX512DQ-NEXT: vmovdqa64 %zmm16, %zmm11
9725 ; AVX512DQ-NEXT: vpermt2q %zmm19, %zmm31, %zmm11
9726 ; AVX512DQ-NEXT: vmovdqa64 %zmm16, %zmm25
9727 ; AVX512DQ-NEXT: vpermt2q %zmm19, %zmm23, %zmm25
9728 ; AVX512DQ-NEXT: vpermt2q %zmm19, %zmm26, %zmm16
9729 ; AVX512DQ-NEXT: vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9730 ; AVX512DQ-NEXT: vmovdqa64 128(%r10), %zmm19
9731 ; AVX512DQ-NEXT: vmovdqa64 128(%rax), %zmm0
9732 ; AVX512DQ-NEXT: vmovdqa64 %zmm19, %zmm28
9733 ; AVX512DQ-NEXT: vpermt2q %zmm0, %zmm30, %zmm28
9734 ; AVX512DQ-NEXT: vmovdqa64 %zmm19, %zmm1
9735 ; AVX512DQ-NEXT: vpermt2q %zmm0, %zmm31, %zmm1
9736 ; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9737 ; AVX512DQ-NEXT: vmovdqa64 %zmm19, %zmm13
9738 ; AVX512DQ-NEXT: vpermt2q %zmm0, %zmm27, %zmm13
9739 ; AVX512DQ-NEXT: vmovdqa64 %zmm19, %zmm16
9740 ; AVX512DQ-NEXT: vpermt2q %zmm0, %zmm24, %zmm16
9741 ; AVX512DQ-NEXT: vmovdqa64 128(%r8), %zmm20
9742 ; AVX512DQ-NEXT: vmovdqa64 128(%r9), %zmm1
9743 ; AVX512DQ-NEXT: vmovdqa64 %zmm20, %zmm3
9744 ; AVX512DQ-NEXT: vpermt2q %zmm1, %zmm22, %zmm3
9745 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} zmm3 {%k1} = zmm19[0],zmm0[0],zmm19[2],zmm0[2],zmm19[4],zmm0[4],zmm19[6],zmm0[6]
9746 ; AVX512DQ-NEXT: vmovdqa64 %zmm20, %zmm10
9747 ; AVX512DQ-NEXT: vpermt2q %zmm1, %zmm21, %zmm10
9748 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} zmm10 {%k1} = zmm19[1],zmm0[1],zmm19[3],zmm0[3],zmm19[5],zmm0[5],zmm19[7],zmm0[7]
9749 ; AVX512DQ-NEXT: vmovdqa64 %zmm19, %zmm2
9750 ; AVX512DQ-NEXT: vpermt2q %zmm0, %zmm23, %zmm2
9751 ; AVX512DQ-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9752 ; AVX512DQ-NEXT: vpermt2q %zmm0, %zmm26, %zmm19
9753 ; AVX512DQ-NEXT: vmovdqa64 %zmm20, %zmm4
9754 ; AVX512DQ-NEXT: vpermt2q %zmm1, %zmm30, %zmm4
9755 ; AVX512DQ-NEXT: vmovdqa64 %zmm20, %zmm6
9756 ; AVX512DQ-NEXT: vpermt2q %zmm1, %zmm31, %zmm6
9757 ; AVX512DQ-NEXT: vmovdqa64 %zmm20, %zmm17
9758 ; AVX512DQ-NEXT: vpermt2q %zmm1, %zmm23, %zmm17
9759 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} zmm5 = zmm20[0],zmm1[0],zmm20[2],zmm1[2],zmm20[4],zmm1[4],zmm20[6],zmm1[6]
9760 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} zmm7 = zmm20[1],zmm1[1],zmm20[3],zmm1[3],zmm20[5],zmm1[5],zmm20[7],zmm1[7]
9761 ; AVX512DQ-NEXT: vpermt2q %zmm1, %zmm26, %zmm20
9762 ; AVX512DQ-NEXT: vmovdqa64 192(%r10), %zmm8
9763 ; AVX512DQ-NEXT: vmovdqa64 192(%rax), %zmm1
9764 ; AVX512DQ-NEXT: vmovdqa64 %zmm8, %zmm12
9765 ; AVX512DQ-NEXT: vpermt2q %zmm1, %zmm30, %zmm12
9766 ; AVX512DQ-NEXT: vmovdqa64 %zmm8, %zmm15
9767 ; AVX512DQ-NEXT: vpermt2q %zmm1, %zmm31, %zmm15
9768 ; AVX512DQ-NEXT: vpermi2q %zmm1, %zmm8, %zmm27
9769 ; AVX512DQ-NEXT: vpermi2q %zmm1, %zmm8, %zmm24
9770 ; AVX512DQ-NEXT: vmovdqa64 192(%r8), %zmm2
9771 ; AVX512DQ-NEXT: vmovdqa64 192(%r9), %zmm0
9772 ; AVX512DQ-NEXT: vpermi2q %zmm0, %zmm2, %zmm22
9773 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} zmm22 {%k1} = zmm8[0],zmm1[0],zmm8[2],zmm1[2],zmm8[4],zmm1[4],zmm8[6],zmm1[6]
9774 ; AVX512DQ-NEXT: vpermi2q %zmm0, %zmm2, %zmm21
9775 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} zmm21 {%k1} = zmm8[1],zmm1[1],zmm8[3],zmm1[3],zmm8[5],zmm1[5],zmm8[7],zmm1[7]
9776 ; AVX512DQ-NEXT: vmovdqa64 %zmm8, %zmm9
9777 ; AVX512DQ-NEXT: vpermt2q %zmm1, %zmm23, %zmm9
9778 ; AVX512DQ-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9779 ; AVX512DQ-NEXT: vpermt2q %zmm1, %zmm26, %zmm8
9780 ; AVX512DQ-NEXT: vpermi2q %zmm0, %zmm2, %zmm30
9781 ; AVX512DQ-NEXT: vpermi2q %zmm0, %zmm2, %zmm31
9782 ; AVX512DQ-NEXT: vpermi2q %zmm0, %zmm2, %zmm23
9783 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} zmm1 = zmm2[0],zmm0[0],zmm2[2],zmm0[2],zmm2[4],zmm0[4],zmm2[6],zmm0[6]
9784 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} zmm9 = zmm2[1],zmm0[1],zmm2[3],zmm0[3],zmm2[5],zmm0[5],zmm2[7],zmm0[7]
9785 ; AVX512DQ-NEXT: vpermt2q %zmm0, %zmm26, %zmm2
9786 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
9787 ; AVX512DQ-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
9788 ; AVX512DQ-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
9789 ; AVX512DQ-NEXT: vmovdqa64 %zmm13, %zmm5 {%k1}
9790 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm0, %zmm5, %zmm0
9791 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9792 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
9793 ; AVX512DQ-NEXT: vpblendd $240, (%rsp), %ymm0, %ymm0 # 32-byte Folded Reload
9794 ; AVX512DQ-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
9795 ; AVX512DQ-NEXT: vmovdqa64 %zmm16, %zmm7 {%k1}
9796 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm0, %zmm7, %zmm0
9797 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9798 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
9799 ; AVX512DQ-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
9800 ; AVX512DQ-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
9801 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm0, %zmm3, %zmm0
9802 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9803 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
9804 ; AVX512DQ-NEXT: vmovdqa64 %zmm0, %zmm29 {%k1}
9805 ; AVX512DQ-NEXT: vmovdqa (%rsi), %xmm0
9806 ; AVX512DQ-NEXT: vinserti128 $1, (%rcx), %ymm0, %ymm3
9807 ; AVX512DQ-NEXT: vmovdqa (%rdi), %xmm0
9808 ; AVX512DQ-NEXT: vinserti128 $1, (%rdx), %ymm0, %ymm5
9809 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm5[0],ymm3[0],ymm5[2],ymm3[2]
9810 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm0, %zmm29, %zmm0
9811 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9812 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
9813 ; AVX512DQ-NEXT: vmovdqa64 %zmm0, %zmm18 {%k1}
9814 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} ymm3 = ymm5[1],ymm3[1],ymm5[3],ymm3[3]
9815 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm3, %zmm18, %zmm0
9816 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, (%rsp) # 64-byte Spill
9817 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
9818 ; AVX512DQ-NEXT: vmovdqa64 %zmm0, %zmm14 {%k1}
9819 ; AVX512DQ-NEXT: vmovdqa 64(%rsi), %xmm3
9820 ; AVX512DQ-NEXT: vinserti128 $1, 64(%rcx), %ymm3, %ymm3
9821 ; AVX512DQ-NEXT: vmovdqa 64(%rdi), %xmm5
9822 ; AVX512DQ-NEXT: vinserti128 $1, 64(%rdx), %ymm5, %ymm5
9823 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm5[0],ymm3[0],ymm5[2],ymm3[2]
9824 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm7, %zmm14, %zmm29
9825 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
9826 ; AVX512DQ-NEXT: vmovdqa64 %zmm0, %zmm11 {%k1}
9827 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} ymm3 = ymm5[1],ymm3[1],ymm5[3],ymm3[3]
9828 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm3, %zmm11, %zmm26
9829 ; AVX512DQ-NEXT: vmovdqa64 %zmm28, %zmm4 {%k1}
9830 ; AVX512DQ-NEXT: vmovdqa 128(%rsi), %xmm3
9831 ; AVX512DQ-NEXT: vinserti128 $1, 128(%rcx), %ymm3, %ymm3
9832 ; AVX512DQ-NEXT: vmovdqa 128(%rdi), %xmm5
9833 ; AVX512DQ-NEXT: vinserti128 $1, 128(%rdx), %ymm5, %ymm5
9834 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm5[0],ymm3[0],ymm5[2],ymm3[2]
9835 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm7, %zmm4, %zmm28
9836 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
9837 ; AVX512DQ-NEXT: vmovdqa64 %zmm0, %zmm6 {%k1}
9838 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} ymm3 = ymm5[1],ymm3[1],ymm5[3],ymm3[3]
9839 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm3, %zmm6, %zmm18
9840 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
9841 ; AVX512DQ-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm4 # 32-byte Folded Reload
9842 ; AVX512DQ-NEXT: # ymm4 = ymm0[0,1,2,3],mem[4,5,6,7]
9843 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm4, %zmm10, %zmm16
9844 ; AVX512DQ-NEXT: vmovdqa64 %zmm12, %zmm30 {%k1}
9845 ; AVX512DQ-NEXT: vmovdqa 192(%rsi), %xmm4
9846 ; AVX512DQ-NEXT: vinserti128 $1, 192(%rcx), %ymm4, %ymm5
9847 ; AVX512DQ-NEXT: vmovdqa 192(%rdi), %xmm4
9848 ; AVX512DQ-NEXT: vinserti128 $1, 192(%rdx), %ymm4, %ymm6
9849 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} ymm4 = ymm6[0],ymm5[0],ymm6[2],ymm5[2]
9850 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm4, %zmm30, %zmm30
9851 ; AVX512DQ-NEXT: vmovdqa64 %zmm15, %zmm31 {%k1}
9852 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} ymm5 = ymm6[1],ymm5[1],ymm6[3],ymm5[3]
9853 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm5, %zmm31, %zmm31
9854 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
9855 ; AVX512DQ-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm6 # 32-byte Folded Reload
9856 ; AVX512DQ-NEXT: # ymm6 = ymm0[0,1,2,3],mem[4,5,6,7]
9857 ; AVX512DQ-NEXT: vmovdqa64 %zmm27, %zmm1 {%k1}
9858 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm6, %zmm1, %zmm6
9859 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
9860 ; AVX512DQ-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm1 # 32-byte Folded Reload
9861 ; AVX512DQ-NEXT: # ymm1 = ymm0[0,1,2,3],mem[4,5,6,7]
9862 ; AVX512DQ-NEXT: vmovdqa64 %zmm24, %zmm9 {%k1}
9863 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm1, %zmm9, %zmm7
9864 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
9865 ; AVX512DQ-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm1 # 32-byte Folded Reload
9866 ; AVX512DQ-NEXT: # ymm1 = ymm0[0,1,2,3],mem[4,5,6,7]
9867 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm1, %zmm22, %zmm9
9868 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
9869 ; AVX512DQ-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm1 # 32-byte Folded Reload
9870 ; AVX512DQ-NEXT: # ymm1 = ymm0[0,1,2,3],mem[4,5,6,7]
9871 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm1, %zmm21, %zmm12
9872 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
9873 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
9874 ; AVX512DQ-NEXT: vmovdqa64 %zmm0, %zmm3 {%k1}
9875 ; AVX512DQ-NEXT: vmovdqa (%rcx), %ymm1
9876 ; AVX512DQ-NEXT: vmovdqa (%rdx), %ymm15
9877 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm15[0],ymm1[0],ymm15[2],ymm1[2]
9878 ; AVX512DQ-NEXT: vmovdqa64 (%rsi), %ymm21
9879 ; AVX512DQ-NEXT: vmovdqa64 (%rdi), %ymm22
9880 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} ymm14 = ymm22[0],ymm21[0],ymm22[2],ymm21[2]
9881 ; AVX512DQ-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm14[2,3],ymm0[2,3]
9882 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm0, %zmm3, %zmm0
9883 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
9884 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
9885 ; AVX512DQ-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
9886 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm15[1],ymm1[1],ymm15[3],ymm1[3]
9887 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} ymm14 = ymm22[1],ymm21[1],ymm22[3],ymm21[3]
9888 ; AVX512DQ-NEXT: vperm2i128 {{.*#+}} ymm1 = ymm14[2,3],ymm1[2,3]
9889 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm1, %zmm4, %zmm1
9890 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
9891 ; AVX512DQ-NEXT: vmovdqa64 %zmm3, %zmm25 {%k1}
9892 ; AVX512DQ-NEXT: vmovdqa 64(%rcx), %ymm14
9893 ; AVX512DQ-NEXT: vmovdqa 64(%rdx), %ymm15
9894 ; AVX512DQ-NEXT: vmovdqa64 64(%rsi), %ymm21
9895 ; AVX512DQ-NEXT: vmovdqa64 64(%rdi), %ymm22
9896 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} ymm13 = ymm15[0],ymm14[0],ymm15[2],ymm14[2]
9897 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} ymm11 = ymm22[0],ymm21[0],ymm22[2],ymm21[2]
9898 ; AVX512DQ-NEXT: vperm2i128 {{.*#+}} ymm11 = ymm11[2,3],ymm13[2,3]
9899 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm11, %zmm25, %zmm11
9900 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
9901 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
9902 ; AVX512DQ-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
9903 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} ymm13 = ymm15[1],ymm14[1],ymm15[3],ymm14[3]
9904 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} ymm14 = ymm22[1],ymm21[1],ymm22[3],ymm21[3]
9905 ; AVX512DQ-NEXT: vperm2i128 {{.*#+}} ymm13 = ymm14[2,3],ymm13[2,3]
9906 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm13, %zmm4, %zmm13
9907 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
9908 ; AVX512DQ-NEXT: vmovdqa64 %zmm3, %zmm17 {%k1}
9909 ; AVX512DQ-NEXT: vmovdqa 128(%rcx), %ymm14
9910 ; AVX512DQ-NEXT: vmovdqa 128(%rdx), %ymm15
9911 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} ymm4 = ymm15[0],ymm14[0],ymm15[2],ymm14[2]
9912 ; AVX512DQ-NEXT: vmovdqa64 128(%rsi), %ymm21
9913 ; AVX512DQ-NEXT: vmovdqa64 128(%rdi), %ymm22
9914 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm22[0],ymm21[0],ymm22[2],ymm21[2]
9915 ; AVX512DQ-NEXT: vperm2i128 {{.*#+}} ymm3 = ymm3[2,3],ymm4[2,3]
9916 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm3, %zmm17, %zmm3
9917 ; AVX512DQ-NEXT: vmovdqa64 %zmm19, %zmm20 {%k1}
9918 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} ymm4 = ymm15[1],ymm14[1],ymm15[3],ymm14[3]
9919 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} ymm14 = ymm22[1],ymm21[1],ymm22[3],ymm21[3]
9920 ; AVX512DQ-NEXT: vperm2i128 {{.*#+}} ymm4 = ymm14[2,3],ymm4[2,3]
9921 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm4, %zmm20, %zmm4
9922 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
9923 ; AVX512DQ-NEXT: vmovdqa64 %zmm5, %zmm23 {%k1}
9924 ; AVX512DQ-NEXT: vmovdqa 192(%rcx), %ymm14
9925 ; AVX512DQ-NEXT: vmovdqa 192(%rdx), %ymm15
9926 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} ymm5 = ymm15[0],ymm14[0],ymm15[2],ymm14[2]
9927 ; AVX512DQ-NEXT: vmovdqa64 192(%rsi), %ymm17
9928 ; AVX512DQ-NEXT: vmovdqa64 192(%rdi), %ymm19
9929 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} ymm10 = ymm19[0],ymm17[0],ymm19[2],ymm17[2]
9930 ; AVX512DQ-NEXT: vperm2i128 {{.*#+}} ymm5 = ymm10[2,3],ymm5[2,3]
9931 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm5, %zmm23, %zmm5
9932 ; AVX512DQ-NEXT: vmovdqa64 %zmm8, %zmm2 {%k1}
9933 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} ymm8 = ymm15[1],ymm14[1],ymm15[3],ymm14[3]
9934 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} ymm10 = ymm19[1],ymm17[1],ymm19[3],ymm17[3]
9935 ; AVX512DQ-NEXT: vperm2i128 {{.*#+}} ymm8 = ymm10[2,3],ymm8[2,3]
9936 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm8, %zmm2, %zmm2
9937 ; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %rax
9938 ; AVX512DQ-NEXT: vmovdqa64 %zmm2, 1728(%rax)
9939 ; AVX512DQ-NEXT: vmovdqa64 %zmm5, 1664(%rax)
9940 ; AVX512DQ-NEXT: vmovdqa64 %zmm4, 1216(%rax)
9941 ; AVX512DQ-NEXT: vmovdqa64 %zmm3, 1152(%rax)
9942 ; AVX512DQ-NEXT: vmovdqa64 %zmm13, 704(%rax)
9943 ; AVX512DQ-NEXT: vmovdqa64 %zmm11, 640(%rax)
9944 ; AVX512DQ-NEXT: vmovdqa64 %zmm1, 192(%rax)
9945 ; AVX512DQ-NEXT: vmovdqa64 %zmm0, 128(%rax)
9946 ; AVX512DQ-NEXT: vmovdqa64 %zmm12, 1984(%rax)
9947 ; AVX512DQ-NEXT: vmovdqa64 %zmm9, 1920(%rax)
9948 ; AVX512DQ-NEXT: vmovdqa64 %zmm7, 1856(%rax)
9949 ; AVX512DQ-NEXT: vmovdqa64 %zmm6, 1792(%rax)
9950 ; AVX512DQ-NEXT: vmovdqa64 %zmm31, 1600(%rax)
9951 ; AVX512DQ-NEXT: vmovdqa64 %zmm30, 1536(%rax)
9952 ; AVX512DQ-NEXT: vmovdqa64 %zmm16, 1472(%rax)
9953 ; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
9954 ; AVX512DQ-NEXT: vmovaps %zmm0, 1408(%rax)
9955 ; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
9956 ; AVX512DQ-NEXT: vmovaps %zmm0, 1344(%rax)
9957 ; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
9958 ; AVX512DQ-NEXT: vmovaps %zmm0, 1280(%rax)
9959 ; AVX512DQ-NEXT: vmovdqa64 %zmm18, 1088(%rax)
9960 ; AVX512DQ-NEXT: vmovdqa64 %zmm28, 1024(%rax)
9961 ; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
9962 ; AVX512DQ-NEXT: vmovaps %zmm0, 960(%rax)
9963 ; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
9964 ; AVX512DQ-NEXT: vmovaps %zmm0, 896(%rax)
9965 ; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
9966 ; AVX512DQ-NEXT: vmovaps %zmm0, 832(%rax)
9967 ; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
9968 ; AVX512DQ-NEXT: vmovaps %zmm0, 768(%rax)
9969 ; AVX512DQ-NEXT: vmovdqa64 %zmm26, 576(%rax)
9970 ; AVX512DQ-NEXT: vmovdqa64 %zmm29, 512(%rax)
9971 ; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
9972 ; AVX512DQ-NEXT: vmovaps %zmm0, 448(%rax)
9973 ; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
9974 ; AVX512DQ-NEXT: vmovaps %zmm0, 384(%rax)
9975 ; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
9976 ; AVX512DQ-NEXT: vmovaps %zmm0, 320(%rax)
9977 ; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
9978 ; AVX512DQ-NEXT: vmovaps %zmm0, 256(%rax)
9979 ; AVX512DQ-NEXT: vmovups (%rsp), %zmm0 # 64-byte Reload
9980 ; AVX512DQ-NEXT: vmovaps %zmm0, 64(%rax)
9981 ; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
9982 ; AVX512DQ-NEXT: vmovaps %zmm0, (%rax)
9983 ; AVX512DQ-NEXT: addq $2312, %rsp # imm = 0x908
9984 ; AVX512DQ-NEXT: vzeroupper
9985 ; AVX512DQ-NEXT: retq
9986 ;
9987 ; AVX512DQ-FCP-LABEL: store_i64_stride8_vf32:
9988 ; AVX512DQ-FCP: # %bb.0:
9989 ; AVX512DQ-FCP-NEXT: subq $2312, %rsp # imm = 0x908
9990 ; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
9991 ; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10
9992 ; AVX512DQ-FCP-NEXT: vmovaps 128(%rdi), %zmm0
9993 ; AVX512DQ-FCP-NEXT: vmovups %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
9994 ; AVX512DQ-FCP-NEXT: vmovdqa64 64(%rdi), %zmm6
9995 ; AVX512DQ-FCP-NEXT: vmovdqa64 (%rdi), %zmm2
9996 ; AVX512DQ-FCP-NEXT: vmovdqa64 128(%rsi), %zmm0
9997 ; AVX512DQ-FCP-NEXT: vmovdqa64 64(%rsi), %zmm5
9998 ; AVX512DQ-FCP-NEXT: vmovdqa64 (%rsi), %zmm3
9999 ; AVX512DQ-FCP-NEXT: vmovaps 192(%rdx), %zmm1
10000 ; AVX512DQ-FCP-NEXT: vmovups %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10001 ; AVX512DQ-FCP-NEXT: vmovaps 128(%rdx), %zmm1
10002 ; AVX512DQ-FCP-NEXT: vmovups %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10003 ; AVX512DQ-FCP-NEXT: vmovdqa64 64(%rdx), %zmm7
10004 ; AVX512DQ-FCP-NEXT: vmovdqa64 (%rdx), %zmm4
10005 ; AVX512DQ-FCP-NEXT: vmovdqa64 192(%rcx), %zmm23
10006 ; AVX512DQ-FCP-NEXT: vmovdqa64 128(%rcx), %zmm1
10007 ; AVX512DQ-FCP-NEXT: vmovdqa64 64(%rcx), %zmm9
10008 ; AVX512DQ-FCP-NEXT: vmovdqa64 (%rcx), %zmm8
10009 ; AVX512DQ-FCP-NEXT: vmovdqa64 (%r8), %zmm18
10010 ; AVX512DQ-FCP-NEXT: vmovdqa64 64(%r8), %zmm16
10011 ; AVX512DQ-FCP-NEXT: vmovdqa64 (%r9), %zmm12
10012 ; AVX512DQ-FCP-NEXT: vmovdqa64 64(%r9), %zmm19
10013 ; AVX512DQ-FCP-NEXT: vmovdqa64 (%r10), %zmm25
10014 ; AVX512DQ-FCP-NEXT: vmovdqa64 64(%r10), %zmm17
10015 ; AVX512DQ-FCP-NEXT: vmovdqa64 (%rax), %zmm15
10016 ; AVX512DQ-FCP-NEXT: vmovdqa64 64(%rax), %zmm20
10017 ; AVX512DQ-FCP-NEXT: movb $-64, %r11b
10018 ; AVX512DQ-FCP-NEXT: kmovw %r11d, %k1
10019 ; AVX512DQ-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm27 = [4,12,4,12,4,12,4,12]
10020 ; AVX512DQ-FCP-NEXT: # zmm27 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
10021 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm4, %zmm10
10022 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm8, %zmm27, %zmm10
10023 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm11
10024 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm3, %zmm27, %zmm11
10025 ; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm10 = ymm11[0,1,2,3],ymm10[4,5,6,7]
10026 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm25, %zmm11
10027 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm15, %zmm27, %zmm11
10028 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm13 = zmm18[0],zmm12[0],zmm18[2],zmm12[2],zmm18[4],zmm12[4],zmm18[6],zmm12[6]
10029 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm11, %zmm13 {%k1}
10030 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm10, %zmm13, %zmm10
10031 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10032 ; AVX512DQ-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm24 = [5,13,5,13,5,13,5,13]
10033 ; AVX512DQ-FCP-NEXT: # zmm24 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
10034 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm4, %zmm10
10035 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm8, %zmm24, %zmm10
10036 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm11
10037 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm3, %zmm24, %zmm11
10038 ; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm10 = ymm11[0,1,2,3],ymm10[4,5,6,7]
10039 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm25, %zmm11
10040 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm15, %zmm24, %zmm11
10041 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm14 = zmm18[1],zmm12[1],zmm18[3],zmm12[3],zmm18[5],zmm12[5],zmm18[7],zmm12[7]
10042 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm11, %zmm14 {%k1}
10043 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm10, %zmm14, %zmm10
10044 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10045 ; AVX512DQ-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm22 = [6,14,6,14,6,14,6,14]
10046 ; AVX512DQ-FCP-NEXT: # zmm22 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
10047 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm4, %zmm10
10048 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm8, %zmm22, %zmm10
10049 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm11
10050 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm3, %zmm22, %zmm11
10051 ; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm10 = ymm11[0,1,2,3],ymm10[4,5,6,7]
10052 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm18, %zmm11
10053 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm12, %zmm22, %zmm11
10054 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm11 {%k1} = zmm25[0],zmm15[0],zmm25[2],zmm15[2],zmm25[4],zmm15[4],zmm25[6],zmm15[6]
10055 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm10, %zmm11, %zmm10
10056 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10057 ; AVX512DQ-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm21 = [7,15,7,15,7,15,7,15]
10058 ; AVX512DQ-FCP-NEXT: # zmm21 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
10059 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm8, %zmm21, %zmm4
10060 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm3, %zmm21, %zmm2
10061 ; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm4[4,5,6,7]
10062 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm18, %zmm3
10063 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm12, %zmm21, %zmm3
10064 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm3 {%k1} = zmm25[1],zmm15[1],zmm25[3],zmm15[3],zmm25[5],zmm15[5],zmm25[7],zmm15[7]
10065 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
10066 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10067 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm7, %zmm2
10068 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm9, %zmm27, %zmm2
10069 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, %zmm3
10070 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm5, %zmm27, %zmm3
10071 ; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3],ymm2[4,5,6,7]
10072 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm17, %zmm3
10073 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm20, %zmm27, %zmm3
10074 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm4 = zmm16[0],zmm19[0],zmm16[2],zmm19[2],zmm16[4],zmm19[4],zmm16[6],zmm19[6]
10075 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
10076 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
10077 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10078 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm7, %zmm2
10079 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm9, %zmm24, %zmm2
10080 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, %zmm3
10081 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm5, %zmm24, %zmm3
10082 ; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3],ymm2[4,5,6,7]
10083 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm17, %zmm3
10084 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm20, %zmm24, %zmm3
10085 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm4 = zmm16[1],zmm19[1],zmm16[3],zmm19[3],zmm16[5],zmm19[5],zmm16[7],zmm19[7]
10086 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
10087 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
10088 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10089 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm7, %zmm2
10090 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm9, %zmm22, %zmm2
10091 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, %zmm3
10092 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm5, %zmm22, %zmm3
10093 ; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3],ymm2[4,5,6,7]
10094 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm16, %zmm3
10095 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm19, %zmm22, %zmm3
10096 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm3 {%k1} = zmm17[0],zmm20[0],zmm17[2],zmm20[2],zmm17[4],zmm20[4],zmm17[6],zmm20[6]
10097 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
10098 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10099 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm9, %zmm21, %zmm7
10100 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm5, %zmm21, %zmm6
10101 ; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm6[0,1,2,3],ymm7[4,5,6,7]
10102 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm16, %zmm3
10103 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm19, %zmm21, %zmm3
10104 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm3 {%k1} = zmm17[1],zmm20[1],zmm17[3],zmm20[3],zmm17[5],zmm20[5],zmm17[7],zmm20[7]
10105 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
10106 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10107 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
10108 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, %zmm2
10109 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm27, %zmm2
10110 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10111 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
10112 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm4
10113 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm27, %zmm4
10114 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10115 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, %zmm4
10116 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm24, %zmm4
10117 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm4, (%rsp) # 64-byte Spill
10118 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm4
10119 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm24, %zmm4
10120 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10121 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, %zmm4
10122 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm22, %zmm4
10123 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10124 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm4
10125 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm22, %zmm4
10126 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10127 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm21, %zmm3
10128 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10129 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm21, %zmm2
10130 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10131 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
10132 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm1
10133 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm23, %zmm27, %zmm1
10134 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10135 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm1
10136 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm23, %zmm24, %zmm1
10137 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10138 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm1
10139 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm23, %zmm22, %zmm1
10140 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10141 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm23, %zmm21, %zmm0
10142 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10143 ; AVX512DQ-FCP-NEXT: vmovdqa64 192(%rdi), %zmm1
10144 ; AVX512DQ-FCP-NEXT: vmovdqa64 192(%rsi), %zmm0
10145 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, %zmm2
10146 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm27, %zmm2
10147 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10148 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, %zmm2
10149 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm24, %zmm2
10150 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10151 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, %zmm2
10152 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm22, %zmm2
10153 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10154 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm21, %zmm1
10155 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10156 ; AVX512DQ-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm30 = [0,8,0,8,0,8,0,8]
10157 ; AVX512DQ-FCP-NEXT: # zmm30 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
10158 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm25, %zmm0
10159 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm15, %zmm30, %zmm0
10160 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10161 ; AVX512DQ-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm31 = [1,9,1,9,1,9,1,9]
10162 ; AVX512DQ-FCP-NEXT: # zmm31 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
10163 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm25, %zmm0
10164 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm15, %zmm31, %zmm0
10165 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10166 ; AVX512DQ-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm23 = [2,10,2,10,2,10,2,10]
10167 ; AVX512DQ-FCP-NEXT: # zmm23 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
10168 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm25, %zmm0
10169 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm15, %zmm23, %zmm0
10170 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10171 ; AVX512DQ-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm26 = [3,11,3,11,3,11,3,11]
10172 ; AVX512DQ-FCP-NEXT: # zmm26 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
10173 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm15, %zmm26, %zmm25
10174 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm25, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10175 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm18, %zmm0
10176 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm18, %zmm29
10177 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm12, %zmm30, %zmm29
10178 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm12, %zmm31, %zmm18
10179 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm1
10180 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm12, %zmm23, %zmm1
10181 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10182 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm12, %zmm26, %zmm0
10183 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10184 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm17, %zmm0
10185 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm20, %zmm30, %zmm0
10186 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10187 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm17, %zmm0
10188 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm20, %zmm31, %zmm0
10189 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10190 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm17, %zmm0
10191 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm20, %zmm23, %zmm0
10192 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10193 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm20, %zmm26, %zmm17
10194 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm17, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10195 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm16, %zmm14
10196 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm19, %zmm30, %zmm14
10197 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm16, %zmm11
10198 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm19, %zmm31, %zmm11
10199 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm16, %zmm25
10200 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm19, %zmm23, %zmm25
10201 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm19, %zmm26, %zmm16
10202 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10203 ; AVX512DQ-FCP-NEXT: vmovdqa64 128(%r10), %zmm19
10204 ; AVX512DQ-FCP-NEXT: vmovdqa64 128(%rax), %zmm0
10205 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm19, %zmm28
10206 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm30, %zmm28
10207 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm19, %zmm1
10208 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm31, %zmm1
10209 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10210 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm19, %zmm13
10211 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm27, %zmm13
10212 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm19, %zmm16
10213 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm24, %zmm16
10214 ; AVX512DQ-FCP-NEXT: vmovdqa64 128(%r8), %zmm20
10215 ; AVX512DQ-FCP-NEXT: vmovdqa64 128(%r9), %zmm1
10216 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm20, %zmm3
10217 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm22, %zmm3
10218 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm3 {%k1} = zmm19[0],zmm0[0],zmm19[2],zmm0[2],zmm19[4],zmm0[4],zmm19[6],zmm0[6]
10219 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm20, %zmm10
10220 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm21, %zmm10
10221 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm10 {%k1} = zmm19[1],zmm0[1],zmm19[3],zmm0[3],zmm19[5],zmm0[5],zmm19[7],zmm0[7]
10222 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm19, %zmm2
10223 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm23, %zmm2
10224 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10225 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm26, %zmm19
10226 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm20, %zmm4
10227 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm30, %zmm4
10228 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm20, %zmm6
10229 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm31, %zmm6
10230 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm20, %zmm17
10231 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm23, %zmm17
10232 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm5 = zmm20[0],zmm1[0],zmm20[2],zmm1[2],zmm20[4],zmm1[4],zmm20[6],zmm1[6]
10233 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm7 = zmm20[1],zmm1[1],zmm20[3],zmm1[3],zmm20[5],zmm1[5],zmm20[7],zmm1[7]
10234 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm26, %zmm20
10235 ; AVX512DQ-FCP-NEXT: vmovdqa64 192(%r10), %zmm8
10236 ; AVX512DQ-FCP-NEXT: vmovdqa64 192(%rax), %zmm1
10237 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm8, %zmm12
10238 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm30, %zmm12
10239 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm8, %zmm15
10240 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm31, %zmm15
10241 ; AVX512DQ-FCP-NEXT: vpermi2q %zmm1, %zmm8, %zmm27
10242 ; AVX512DQ-FCP-NEXT: vpermi2q %zmm1, %zmm8, %zmm24
10243 ; AVX512DQ-FCP-NEXT: vmovdqa64 192(%r8), %zmm2
10244 ; AVX512DQ-FCP-NEXT: vmovdqa64 192(%r9), %zmm0
10245 ; AVX512DQ-FCP-NEXT: vpermi2q %zmm0, %zmm2, %zmm22
10246 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm22 {%k1} = zmm8[0],zmm1[0],zmm8[2],zmm1[2],zmm8[4],zmm1[4],zmm8[6],zmm1[6]
10247 ; AVX512DQ-FCP-NEXT: vpermi2q %zmm0, %zmm2, %zmm21
10248 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm21 {%k1} = zmm8[1],zmm1[1],zmm8[3],zmm1[3],zmm8[5],zmm1[5],zmm8[7],zmm1[7]
10249 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm8, %zmm9
10250 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm23, %zmm9
10251 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10252 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm26, %zmm8
10253 ; AVX512DQ-FCP-NEXT: vpermi2q %zmm0, %zmm2, %zmm30
10254 ; AVX512DQ-FCP-NEXT: vpermi2q %zmm0, %zmm2, %zmm31
10255 ; AVX512DQ-FCP-NEXT: vpermi2q %zmm0, %zmm2, %zmm23
10256 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm1 = zmm2[0],zmm0[0],zmm2[2],zmm0[2],zmm2[4],zmm0[4],zmm2[6],zmm0[6]
10257 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm9 = zmm2[1],zmm0[1],zmm2[3],zmm0[3],zmm2[5],zmm0[5],zmm2[7],zmm0[7]
10258 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm26, %zmm2
10259 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
10260 ; AVX512DQ-FCP-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
10261 ; AVX512DQ-FCP-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
10262 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm13, %zmm5 {%k1}
10263 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm5, %zmm0
10264 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10265 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
10266 ; AVX512DQ-FCP-NEXT: vpblendd $240, (%rsp), %ymm0, %ymm0 # 32-byte Folded Reload
10267 ; AVX512DQ-FCP-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
10268 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm16, %zmm7 {%k1}
10269 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm7, %zmm0
10270 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10271 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
10272 ; AVX512DQ-FCP-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
10273 ; AVX512DQ-FCP-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
10274 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm3, %zmm0
10275 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10276 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
10277 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm29 {%k1}
10278 ; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %xmm0
10279 ; AVX512DQ-FCP-NEXT: vinserti128 $1, (%rcx), %ymm0, %ymm3
10280 ; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %xmm0
10281 ; AVX512DQ-FCP-NEXT: vinserti128 $1, (%rdx), %ymm0, %ymm5
10282 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm5[0],ymm3[0],ymm5[2],ymm3[2]
10283 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm29, %zmm0
10284 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10285 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
10286 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm18 {%k1}
10287 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm3 = ymm5[1],ymm3[1],ymm5[3],ymm3[3]
10288 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm3, %zmm18, %zmm0
10289 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, (%rsp) # 64-byte Spill
10290 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
10291 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm14 {%k1}
10292 ; AVX512DQ-FCP-NEXT: vmovdqa 64(%rsi), %xmm3
10293 ; AVX512DQ-FCP-NEXT: vinserti128 $1, 64(%rcx), %ymm3, %ymm3
10294 ; AVX512DQ-FCP-NEXT: vmovdqa 64(%rdi), %xmm5
10295 ; AVX512DQ-FCP-NEXT: vinserti128 $1, 64(%rdx), %ymm5, %ymm5
10296 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm5[0],ymm3[0],ymm5[2],ymm3[2]
10297 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm7, %zmm14, %zmm29
10298 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
10299 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm11 {%k1}
10300 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm3 = ymm5[1],ymm3[1],ymm5[3],ymm3[3]
10301 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm3, %zmm11, %zmm26
10302 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm28, %zmm4 {%k1}
10303 ; AVX512DQ-FCP-NEXT: vmovdqa 128(%rsi), %xmm3
10304 ; AVX512DQ-FCP-NEXT: vinserti128 $1, 128(%rcx), %ymm3, %ymm3
10305 ; AVX512DQ-FCP-NEXT: vmovdqa 128(%rdi), %xmm5
10306 ; AVX512DQ-FCP-NEXT: vinserti128 $1, 128(%rdx), %ymm5, %ymm5
10307 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm5[0],ymm3[0],ymm5[2],ymm3[2]
10308 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm7, %zmm4, %zmm28
10309 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
10310 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm6 {%k1}
10311 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm3 = ymm5[1],ymm3[1],ymm5[3],ymm3[3]
10312 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm3, %zmm6, %zmm18
10313 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
10314 ; AVX512DQ-FCP-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm4 # 32-byte Folded Reload
10315 ; AVX512DQ-FCP-NEXT: # ymm4 = ymm0[0,1,2,3],mem[4,5,6,7]
10316 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm4, %zmm10, %zmm16
10317 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm12, %zmm30 {%k1}
10318 ; AVX512DQ-FCP-NEXT: vmovdqa 192(%rsi), %xmm4
10319 ; AVX512DQ-FCP-NEXT: vinserti128 $1, 192(%rcx), %ymm4, %ymm5
10320 ; AVX512DQ-FCP-NEXT: vmovdqa 192(%rdi), %xmm4
10321 ; AVX512DQ-FCP-NEXT: vinserti128 $1, 192(%rdx), %ymm4, %ymm6
10322 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm4 = ymm6[0],ymm5[0],ymm6[2],ymm5[2]
10323 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm4, %zmm30, %zmm30
10324 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm15, %zmm31 {%k1}
10325 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm5 = ymm6[1],ymm5[1],ymm6[3],ymm5[3]
10326 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm5, %zmm31, %zmm31
10327 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
10328 ; AVX512DQ-FCP-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm6 # 32-byte Folded Reload
10329 ; AVX512DQ-FCP-NEXT: # ymm6 = ymm0[0,1,2,3],mem[4,5,6,7]
10330 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm27, %zmm1 {%k1}
10331 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm1, %zmm6
10332 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
10333 ; AVX512DQ-FCP-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm1 # 32-byte Folded Reload
10334 ; AVX512DQ-FCP-NEXT: # ymm1 = ymm0[0,1,2,3],mem[4,5,6,7]
10335 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm24, %zmm9 {%k1}
10336 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm1, %zmm9, %zmm7
10337 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
10338 ; AVX512DQ-FCP-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm1 # 32-byte Folded Reload
10339 ; AVX512DQ-FCP-NEXT: # ymm1 = ymm0[0,1,2,3],mem[4,5,6,7]
10340 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm1, %zmm22, %zmm9
10341 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
10342 ; AVX512DQ-FCP-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm1 # 32-byte Folded Reload
10343 ; AVX512DQ-FCP-NEXT: # ymm1 = ymm0[0,1,2,3],mem[4,5,6,7]
10344 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm1, %zmm21, %zmm12
10345 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
10346 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
10347 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm3 {%k1}
10348 ; AVX512DQ-FCP-NEXT: vmovdqa (%rcx), %ymm1
10349 ; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %ymm15
10350 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm15[0],ymm1[0],ymm15[2],ymm1[2]
10351 ; AVX512DQ-FCP-NEXT: vmovdqa64 (%rsi), %ymm21
10352 ; AVX512DQ-FCP-NEXT: vmovdqa64 (%rdi), %ymm22
10353 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm14 = ymm22[0],ymm21[0],ymm22[2],ymm21[2]
10354 ; AVX512DQ-FCP-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm14[2,3],ymm0[2,3]
10355 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm3, %zmm0
10356 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
10357 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
10358 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
10359 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm15[1],ymm1[1],ymm15[3],ymm1[3]
10360 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm14 = ymm22[1],ymm21[1],ymm22[3],ymm21[3]
10361 ; AVX512DQ-FCP-NEXT: vperm2i128 {{.*#+}} ymm1 = ymm14[2,3],ymm1[2,3]
10362 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm1, %zmm4, %zmm1
10363 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
10364 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, %zmm25 {%k1}
10365 ; AVX512DQ-FCP-NEXT: vmovdqa 64(%rcx), %ymm14
10366 ; AVX512DQ-FCP-NEXT: vmovdqa 64(%rdx), %ymm15
10367 ; AVX512DQ-FCP-NEXT: vmovdqa64 64(%rsi), %ymm21
10368 ; AVX512DQ-FCP-NEXT: vmovdqa64 64(%rdi), %ymm22
10369 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm13 = ymm15[0],ymm14[0],ymm15[2],ymm14[2]
10370 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm11 = ymm22[0],ymm21[0],ymm22[2],ymm21[2]
10371 ; AVX512DQ-FCP-NEXT: vperm2i128 {{.*#+}} ymm11 = ymm11[2,3],ymm13[2,3]
10372 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm11, %zmm25, %zmm11
10373 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
10374 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
10375 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
10376 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm13 = ymm15[1],ymm14[1],ymm15[3],ymm14[3]
10377 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm14 = ymm22[1],ymm21[1],ymm22[3],ymm21[3]
10378 ; AVX512DQ-FCP-NEXT: vperm2i128 {{.*#+}} ymm13 = ymm14[2,3],ymm13[2,3]
10379 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm13, %zmm4, %zmm13
10380 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
10381 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, %zmm17 {%k1}
10382 ; AVX512DQ-FCP-NEXT: vmovdqa 128(%rcx), %ymm14
10383 ; AVX512DQ-FCP-NEXT: vmovdqa 128(%rdx), %ymm15
10384 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm4 = ymm15[0],ymm14[0],ymm15[2],ymm14[2]
10385 ; AVX512DQ-FCP-NEXT: vmovdqa64 128(%rsi), %ymm21
10386 ; AVX512DQ-FCP-NEXT: vmovdqa64 128(%rdi), %ymm22
10387 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm22[0],ymm21[0],ymm22[2],ymm21[2]
10388 ; AVX512DQ-FCP-NEXT: vperm2i128 {{.*#+}} ymm3 = ymm3[2,3],ymm4[2,3]
10389 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm3, %zmm17, %zmm3
10390 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm19, %zmm20 {%k1}
10391 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm4 = ymm15[1],ymm14[1],ymm15[3],ymm14[3]
10392 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm14 = ymm22[1],ymm21[1],ymm22[3],ymm21[3]
10393 ; AVX512DQ-FCP-NEXT: vperm2i128 {{.*#+}} ymm4 = ymm14[2,3],ymm4[2,3]
10394 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm4, %zmm20, %zmm4
10395 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
10396 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm5, %zmm23 {%k1}
10397 ; AVX512DQ-FCP-NEXT: vmovdqa 192(%rcx), %ymm14
10398 ; AVX512DQ-FCP-NEXT: vmovdqa 192(%rdx), %ymm15
10399 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm5 = ymm15[0],ymm14[0],ymm15[2],ymm14[2]
10400 ; AVX512DQ-FCP-NEXT: vmovdqa64 192(%rsi), %ymm17
10401 ; AVX512DQ-FCP-NEXT: vmovdqa64 192(%rdi), %ymm19
10402 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm10 = ymm19[0],ymm17[0],ymm19[2],ymm17[2]
10403 ; AVX512DQ-FCP-NEXT: vperm2i128 {{.*#+}} ymm5 = ymm10[2,3],ymm5[2,3]
10404 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm5, %zmm23, %zmm5
10405 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm8, %zmm2 {%k1}
10406 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm8 = ymm15[1],ymm14[1],ymm15[3],ymm14[3]
10407 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm10 = ymm19[1],ymm17[1],ymm19[3],ymm17[3]
10408 ; AVX512DQ-FCP-NEXT: vperm2i128 {{.*#+}} ymm8 = ymm10[2,3],ymm8[2,3]
10409 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm8, %zmm2, %zmm2
10410 ; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
10411 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, 1728(%rax)
10412 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm5, 1664(%rax)
10413 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm4, 1216(%rax)
10414 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, 1152(%rax)
10415 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm13, 704(%rax)
10416 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm11, 640(%rax)
10417 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, 192(%rax)
10418 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, 128(%rax)
10419 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm12, 1984(%rax)
10420 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm9, 1920(%rax)
10421 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm7, 1856(%rax)
10422 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, 1792(%rax)
10423 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm31, 1600(%rax)
10424 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm30, 1536(%rax)
10425 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm16, 1472(%rax)
10426 ; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
10427 ; AVX512DQ-FCP-NEXT: vmovaps %zmm0, 1408(%rax)
10428 ; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
10429 ; AVX512DQ-FCP-NEXT: vmovaps %zmm0, 1344(%rax)
10430 ; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
10431 ; AVX512DQ-FCP-NEXT: vmovaps %zmm0, 1280(%rax)
10432 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm18, 1088(%rax)
10433 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm28, 1024(%rax)
10434 ; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
10435 ; AVX512DQ-FCP-NEXT: vmovaps %zmm0, 960(%rax)
10436 ; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
10437 ; AVX512DQ-FCP-NEXT: vmovaps %zmm0, 896(%rax)
10438 ; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
10439 ; AVX512DQ-FCP-NEXT: vmovaps %zmm0, 832(%rax)
10440 ; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
10441 ; AVX512DQ-FCP-NEXT: vmovaps %zmm0, 768(%rax)
10442 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm26, 576(%rax)
10443 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm29, 512(%rax)
10444 ; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
10445 ; AVX512DQ-FCP-NEXT: vmovaps %zmm0, 448(%rax)
10446 ; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
10447 ; AVX512DQ-FCP-NEXT: vmovaps %zmm0, 384(%rax)
10448 ; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
10449 ; AVX512DQ-FCP-NEXT: vmovaps %zmm0, 320(%rax)
10450 ; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
10451 ; AVX512DQ-FCP-NEXT: vmovaps %zmm0, 256(%rax)
10452 ; AVX512DQ-FCP-NEXT: vmovups (%rsp), %zmm0 # 64-byte Reload
10453 ; AVX512DQ-FCP-NEXT: vmovaps %zmm0, 64(%rax)
10454 ; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
10455 ; AVX512DQ-FCP-NEXT: vmovaps %zmm0, (%rax)
10456 ; AVX512DQ-FCP-NEXT: addq $2312, %rsp # imm = 0x908
10457 ; AVX512DQ-FCP-NEXT: vzeroupper
10458 ; AVX512DQ-FCP-NEXT: retq
10459 ;
10460 ; AVX512BW-LABEL: store_i64_stride8_vf32:
10461 ; AVX512BW: # %bb.0:
10462 ; AVX512BW-NEXT: subq $2312, %rsp # imm = 0x908
10463 ; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
10464 ; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %r10
10465 ; AVX512BW-NEXT: vmovaps 128(%rdi), %zmm0
10466 ; AVX512BW-NEXT: vmovups %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10467 ; AVX512BW-NEXT: vmovdqa64 64(%rdi), %zmm6
10468 ; AVX512BW-NEXT: vmovdqa64 (%rdi), %zmm2
10469 ; AVX512BW-NEXT: vmovdqa64 128(%rsi), %zmm0
10470 ; AVX512BW-NEXT: vmovdqa64 64(%rsi), %zmm5
10471 ; AVX512BW-NEXT: vmovdqa64 (%rsi), %zmm3
10472 ; AVX512BW-NEXT: vmovaps 192(%rdx), %zmm1
10473 ; AVX512BW-NEXT: vmovups %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10474 ; AVX512BW-NEXT: vmovaps 128(%rdx), %zmm1
10475 ; AVX512BW-NEXT: vmovups %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10476 ; AVX512BW-NEXT: vmovdqa64 64(%rdx), %zmm7
10477 ; AVX512BW-NEXT: vmovdqa64 (%rdx), %zmm4
10478 ; AVX512BW-NEXT: vmovdqa64 192(%rcx), %zmm23
10479 ; AVX512BW-NEXT: vmovdqa64 128(%rcx), %zmm1
10480 ; AVX512BW-NEXT: vmovdqa64 64(%rcx), %zmm9
10481 ; AVX512BW-NEXT: vmovdqa64 (%rcx), %zmm8
10482 ; AVX512BW-NEXT: vmovdqa64 (%r8), %zmm18
10483 ; AVX512BW-NEXT: vmovdqa64 64(%r8), %zmm16
10484 ; AVX512BW-NEXT: vmovdqa64 (%r9), %zmm12
10485 ; AVX512BW-NEXT: vmovdqa64 64(%r9), %zmm19
10486 ; AVX512BW-NEXT: vmovdqa64 (%r10), %zmm25
10487 ; AVX512BW-NEXT: vmovdqa64 64(%r10), %zmm17
10488 ; AVX512BW-NEXT: vmovdqa64 (%rax), %zmm15
10489 ; AVX512BW-NEXT: vmovdqa64 64(%rax), %zmm20
10490 ; AVX512BW-NEXT: movb $-64, %r11b
10491 ; AVX512BW-NEXT: kmovd %r11d, %k1
10492 ; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm27 = [4,12,4,12,4,12,4,12]
10493 ; AVX512BW-NEXT: # zmm27 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
10494 ; AVX512BW-NEXT: vmovdqa64 %zmm4, %zmm10
10495 ; AVX512BW-NEXT: vpermt2q %zmm8, %zmm27, %zmm10
10496 ; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm11
10497 ; AVX512BW-NEXT: vpermt2q %zmm3, %zmm27, %zmm11
10498 ; AVX512BW-NEXT: vpblendd {{.*#+}} ymm10 = ymm11[0,1,2,3],ymm10[4,5,6,7]
10499 ; AVX512BW-NEXT: vmovdqa64 %zmm25, %zmm11
10500 ; AVX512BW-NEXT: vpermt2q %zmm15, %zmm27, %zmm11
10501 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} zmm13 = zmm18[0],zmm12[0],zmm18[2],zmm12[2],zmm18[4],zmm12[4],zmm18[6],zmm12[6]
10502 ; AVX512BW-NEXT: vmovdqa64 %zmm11, %zmm13 {%k1}
10503 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm10, %zmm13, %zmm10
10504 ; AVX512BW-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10505 ; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm24 = [5,13,5,13,5,13,5,13]
10506 ; AVX512BW-NEXT: # zmm24 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
10507 ; AVX512BW-NEXT: vmovdqa64 %zmm4, %zmm10
10508 ; AVX512BW-NEXT: vpermt2q %zmm8, %zmm24, %zmm10
10509 ; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm11
10510 ; AVX512BW-NEXT: vpermt2q %zmm3, %zmm24, %zmm11
10511 ; AVX512BW-NEXT: vpblendd {{.*#+}} ymm10 = ymm11[0,1,2,3],ymm10[4,5,6,7]
10512 ; AVX512BW-NEXT: vmovdqa64 %zmm25, %zmm11
10513 ; AVX512BW-NEXT: vpermt2q %zmm15, %zmm24, %zmm11
10514 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} zmm14 = zmm18[1],zmm12[1],zmm18[3],zmm12[3],zmm18[5],zmm12[5],zmm18[7],zmm12[7]
10515 ; AVX512BW-NEXT: vmovdqa64 %zmm11, %zmm14 {%k1}
10516 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm10, %zmm14, %zmm10
10517 ; AVX512BW-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10518 ; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm22 = [6,14,6,14,6,14,6,14]
10519 ; AVX512BW-NEXT: # zmm22 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
10520 ; AVX512BW-NEXT: vmovdqa64 %zmm4, %zmm10
10521 ; AVX512BW-NEXT: vpermt2q %zmm8, %zmm22, %zmm10
10522 ; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm11
10523 ; AVX512BW-NEXT: vpermt2q %zmm3, %zmm22, %zmm11
10524 ; AVX512BW-NEXT: vpblendd {{.*#+}} ymm10 = ymm11[0,1,2,3],ymm10[4,5,6,7]
10525 ; AVX512BW-NEXT: vmovdqa64 %zmm18, %zmm11
10526 ; AVX512BW-NEXT: vpermt2q %zmm12, %zmm22, %zmm11
10527 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} zmm11 {%k1} = zmm25[0],zmm15[0],zmm25[2],zmm15[2],zmm25[4],zmm15[4],zmm25[6],zmm15[6]
10528 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm10, %zmm11, %zmm10
10529 ; AVX512BW-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10530 ; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm21 = [7,15,7,15,7,15,7,15]
10531 ; AVX512BW-NEXT: # zmm21 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
10532 ; AVX512BW-NEXT: vpermt2q %zmm8, %zmm21, %zmm4
10533 ; AVX512BW-NEXT: vpermt2q %zmm3, %zmm21, %zmm2
10534 ; AVX512BW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm4[4,5,6,7]
10535 ; AVX512BW-NEXT: vmovdqa64 %zmm18, %zmm3
10536 ; AVX512BW-NEXT: vpermt2q %zmm12, %zmm21, %zmm3
10537 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} zmm3 {%k1} = zmm25[1],zmm15[1],zmm25[3],zmm15[3],zmm25[5],zmm15[5],zmm25[7],zmm15[7]
10538 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
10539 ; AVX512BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10540 ; AVX512BW-NEXT: vmovdqa64 %zmm7, %zmm2
10541 ; AVX512BW-NEXT: vpermt2q %zmm9, %zmm27, %zmm2
10542 ; AVX512BW-NEXT: vmovdqa64 %zmm6, %zmm3
10543 ; AVX512BW-NEXT: vpermt2q %zmm5, %zmm27, %zmm3
10544 ; AVX512BW-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3],ymm2[4,5,6,7]
10545 ; AVX512BW-NEXT: vmovdqa64 %zmm17, %zmm3
10546 ; AVX512BW-NEXT: vpermt2q %zmm20, %zmm27, %zmm3
10547 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} zmm4 = zmm16[0],zmm19[0],zmm16[2],zmm19[2],zmm16[4],zmm19[4],zmm16[6],zmm19[6]
10548 ; AVX512BW-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
10549 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
10550 ; AVX512BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10551 ; AVX512BW-NEXT: vmovdqa64 %zmm7, %zmm2
10552 ; AVX512BW-NEXT: vpermt2q %zmm9, %zmm24, %zmm2
10553 ; AVX512BW-NEXT: vmovdqa64 %zmm6, %zmm3
10554 ; AVX512BW-NEXT: vpermt2q %zmm5, %zmm24, %zmm3
10555 ; AVX512BW-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3],ymm2[4,5,6,7]
10556 ; AVX512BW-NEXT: vmovdqa64 %zmm17, %zmm3
10557 ; AVX512BW-NEXT: vpermt2q %zmm20, %zmm24, %zmm3
10558 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} zmm4 = zmm16[1],zmm19[1],zmm16[3],zmm19[3],zmm16[5],zmm19[5],zmm16[7],zmm19[7]
10559 ; AVX512BW-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
10560 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
10561 ; AVX512BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10562 ; AVX512BW-NEXT: vmovdqa64 %zmm7, %zmm2
10563 ; AVX512BW-NEXT: vpermt2q %zmm9, %zmm22, %zmm2
10564 ; AVX512BW-NEXT: vmovdqa64 %zmm6, %zmm3
10565 ; AVX512BW-NEXT: vpermt2q %zmm5, %zmm22, %zmm3
10566 ; AVX512BW-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3],ymm2[4,5,6,7]
10567 ; AVX512BW-NEXT: vmovdqa64 %zmm16, %zmm3
10568 ; AVX512BW-NEXT: vpermt2q %zmm19, %zmm22, %zmm3
10569 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} zmm3 {%k1} = zmm17[0],zmm20[0],zmm17[2],zmm20[2],zmm17[4],zmm20[4],zmm17[6],zmm20[6]
10570 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
10571 ; AVX512BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10572 ; AVX512BW-NEXT: vpermt2q %zmm9, %zmm21, %zmm7
10573 ; AVX512BW-NEXT: vpermt2q %zmm5, %zmm21, %zmm6
10574 ; AVX512BW-NEXT: vpblendd {{.*#+}} ymm2 = ymm6[0,1,2,3],ymm7[4,5,6,7]
10575 ; AVX512BW-NEXT: vmovdqa64 %zmm16, %zmm3
10576 ; AVX512BW-NEXT: vpermt2q %zmm19, %zmm21, %zmm3
10577 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} zmm3 {%k1} = zmm17[1],zmm20[1],zmm17[3],zmm20[3],zmm17[5],zmm20[5],zmm17[7],zmm20[7]
10578 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
10579 ; AVX512BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10580 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
10581 ; AVX512BW-NEXT: vmovdqa64 %zmm3, %zmm2
10582 ; AVX512BW-NEXT: vpermt2q %zmm1, %zmm27, %zmm2
10583 ; AVX512BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10584 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
10585 ; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm4
10586 ; AVX512BW-NEXT: vpermt2q %zmm0, %zmm27, %zmm4
10587 ; AVX512BW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10588 ; AVX512BW-NEXT: vmovdqa64 %zmm3, %zmm4
10589 ; AVX512BW-NEXT: vpermt2q %zmm1, %zmm24, %zmm4
10590 ; AVX512BW-NEXT: vmovdqu64 %zmm4, (%rsp) # 64-byte Spill
10591 ; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm4
10592 ; AVX512BW-NEXT: vpermt2q %zmm0, %zmm24, %zmm4
10593 ; AVX512BW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10594 ; AVX512BW-NEXT: vmovdqa64 %zmm3, %zmm4
10595 ; AVX512BW-NEXT: vpermt2q %zmm1, %zmm22, %zmm4
10596 ; AVX512BW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10597 ; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm4
10598 ; AVX512BW-NEXT: vpermt2q %zmm0, %zmm22, %zmm4
10599 ; AVX512BW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10600 ; AVX512BW-NEXT: vpermt2q %zmm1, %zmm21, %zmm3
10601 ; AVX512BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10602 ; AVX512BW-NEXT: vpermt2q %zmm0, %zmm21, %zmm2
10603 ; AVX512BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10604 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
10605 ; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm1
10606 ; AVX512BW-NEXT: vpermt2q %zmm23, %zmm27, %zmm1
10607 ; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10608 ; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm1
10609 ; AVX512BW-NEXT: vpermt2q %zmm23, %zmm24, %zmm1
10610 ; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10611 ; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm1
10612 ; AVX512BW-NEXT: vpermt2q %zmm23, %zmm22, %zmm1
10613 ; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10614 ; AVX512BW-NEXT: vpermt2q %zmm23, %zmm21, %zmm0
10615 ; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10616 ; AVX512BW-NEXT: vmovdqa64 192(%rdi), %zmm1
10617 ; AVX512BW-NEXT: vmovdqa64 192(%rsi), %zmm0
10618 ; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm2
10619 ; AVX512BW-NEXT: vpermt2q %zmm0, %zmm27, %zmm2
10620 ; AVX512BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10621 ; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm2
10622 ; AVX512BW-NEXT: vpermt2q %zmm0, %zmm24, %zmm2
10623 ; AVX512BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10624 ; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm2
10625 ; AVX512BW-NEXT: vpermt2q %zmm0, %zmm22, %zmm2
10626 ; AVX512BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10627 ; AVX512BW-NEXT: vpermt2q %zmm0, %zmm21, %zmm1
10628 ; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10629 ; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm30 = [0,8,0,8,0,8,0,8]
10630 ; AVX512BW-NEXT: # zmm30 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
10631 ; AVX512BW-NEXT: vmovdqa64 %zmm25, %zmm0
10632 ; AVX512BW-NEXT: vpermt2q %zmm15, %zmm30, %zmm0
10633 ; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10634 ; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm31 = [1,9,1,9,1,9,1,9]
10635 ; AVX512BW-NEXT: # zmm31 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
10636 ; AVX512BW-NEXT: vmovdqa64 %zmm25, %zmm0
10637 ; AVX512BW-NEXT: vpermt2q %zmm15, %zmm31, %zmm0
10638 ; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10639 ; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm23 = [2,10,2,10,2,10,2,10]
10640 ; AVX512BW-NEXT: # zmm23 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
10641 ; AVX512BW-NEXT: vmovdqa64 %zmm25, %zmm0
10642 ; AVX512BW-NEXT: vpermt2q %zmm15, %zmm23, %zmm0
10643 ; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10644 ; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm26 = [3,11,3,11,3,11,3,11]
10645 ; AVX512BW-NEXT: # zmm26 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
10646 ; AVX512BW-NEXT: vpermt2q %zmm15, %zmm26, %zmm25
10647 ; AVX512BW-NEXT: vmovdqu64 %zmm25, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10648 ; AVX512BW-NEXT: vmovdqa64 %zmm18, %zmm0
10649 ; AVX512BW-NEXT: vmovdqa64 %zmm18, %zmm29
10650 ; AVX512BW-NEXT: vpermt2q %zmm12, %zmm30, %zmm29
10651 ; AVX512BW-NEXT: vpermt2q %zmm12, %zmm31, %zmm18
10652 ; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm1
10653 ; AVX512BW-NEXT: vpermt2q %zmm12, %zmm23, %zmm1
10654 ; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10655 ; AVX512BW-NEXT: vpermt2q %zmm12, %zmm26, %zmm0
10656 ; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10657 ; AVX512BW-NEXT: vmovdqa64 %zmm17, %zmm0
10658 ; AVX512BW-NEXT: vpermt2q %zmm20, %zmm30, %zmm0
10659 ; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10660 ; AVX512BW-NEXT: vmovdqa64 %zmm17, %zmm0
10661 ; AVX512BW-NEXT: vpermt2q %zmm20, %zmm31, %zmm0
10662 ; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10663 ; AVX512BW-NEXT: vmovdqa64 %zmm17, %zmm0
10664 ; AVX512BW-NEXT: vpermt2q %zmm20, %zmm23, %zmm0
10665 ; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10666 ; AVX512BW-NEXT: vpermt2q %zmm20, %zmm26, %zmm17
10667 ; AVX512BW-NEXT: vmovdqu64 %zmm17, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10668 ; AVX512BW-NEXT: vmovdqa64 %zmm16, %zmm14
10669 ; AVX512BW-NEXT: vpermt2q %zmm19, %zmm30, %zmm14
10670 ; AVX512BW-NEXT: vmovdqa64 %zmm16, %zmm11
10671 ; AVX512BW-NEXT: vpermt2q %zmm19, %zmm31, %zmm11
10672 ; AVX512BW-NEXT: vmovdqa64 %zmm16, %zmm25
10673 ; AVX512BW-NEXT: vpermt2q %zmm19, %zmm23, %zmm25
10674 ; AVX512BW-NEXT: vpermt2q %zmm19, %zmm26, %zmm16
10675 ; AVX512BW-NEXT: vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10676 ; AVX512BW-NEXT: vmovdqa64 128(%r10), %zmm19
10677 ; AVX512BW-NEXT: vmovdqa64 128(%rax), %zmm0
10678 ; AVX512BW-NEXT: vmovdqa64 %zmm19, %zmm28
10679 ; AVX512BW-NEXT: vpermt2q %zmm0, %zmm30, %zmm28
10680 ; AVX512BW-NEXT: vmovdqa64 %zmm19, %zmm1
10681 ; AVX512BW-NEXT: vpermt2q %zmm0, %zmm31, %zmm1
10682 ; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10683 ; AVX512BW-NEXT: vmovdqa64 %zmm19, %zmm13
10684 ; AVX512BW-NEXT: vpermt2q %zmm0, %zmm27, %zmm13
10685 ; AVX512BW-NEXT: vmovdqa64 %zmm19, %zmm16
10686 ; AVX512BW-NEXT: vpermt2q %zmm0, %zmm24, %zmm16
10687 ; AVX512BW-NEXT: vmovdqa64 128(%r8), %zmm20
10688 ; AVX512BW-NEXT: vmovdqa64 128(%r9), %zmm1
10689 ; AVX512BW-NEXT: vmovdqa64 %zmm20, %zmm3
10690 ; AVX512BW-NEXT: vpermt2q %zmm1, %zmm22, %zmm3
10691 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} zmm3 {%k1} = zmm19[0],zmm0[0],zmm19[2],zmm0[2],zmm19[4],zmm0[4],zmm19[6],zmm0[6]
10692 ; AVX512BW-NEXT: vmovdqa64 %zmm20, %zmm10
10693 ; AVX512BW-NEXT: vpermt2q %zmm1, %zmm21, %zmm10
10694 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} zmm10 {%k1} = zmm19[1],zmm0[1],zmm19[3],zmm0[3],zmm19[5],zmm0[5],zmm19[7],zmm0[7]
10695 ; AVX512BW-NEXT: vmovdqa64 %zmm19, %zmm2
10696 ; AVX512BW-NEXT: vpermt2q %zmm0, %zmm23, %zmm2
10697 ; AVX512BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10698 ; AVX512BW-NEXT: vpermt2q %zmm0, %zmm26, %zmm19
10699 ; AVX512BW-NEXT: vmovdqa64 %zmm20, %zmm4
10700 ; AVX512BW-NEXT: vpermt2q %zmm1, %zmm30, %zmm4
10701 ; AVX512BW-NEXT: vmovdqa64 %zmm20, %zmm6
10702 ; AVX512BW-NEXT: vpermt2q %zmm1, %zmm31, %zmm6
10703 ; AVX512BW-NEXT: vmovdqa64 %zmm20, %zmm17
10704 ; AVX512BW-NEXT: vpermt2q %zmm1, %zmm23, %zmm17
10705 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} zmm5 = zmm20[0],zmm1[0],zmm20[2],zmm1[2],zmm20[4],zmm1[4],zmm20[6],zmm1[6]
10706 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} zmm7 = zmm20[1],zmm1[1],zmm20[3],zmm1[3],zmm20[5],zmm1[5],zmm20[7],zmm1[7]
10707 ; AVX512BW-NEXT: vpermt2q %zmm1, %zmm26, %zmm20
10708 ; AVX512BW-NEXT: vmovdqa64 192(%r10), %zmm8
10709 ; AVX512BW-NEXT: vmovdqa64 192(%rax), %zmm1
10710 ; AVX512BW-NEXT: vmovdqa64 %zmm8, %zmm12
10711 ; AVX512BW-NEXT: vpermt2q %zmm1, %zmm30, %zmm12
10712 ; AVX512BW-NEXT: vmovdqa64 %zmm8, %zmm15
10713 ; AVX512BW-NEXT: vpermt2q %zmm1, %zmm31, %zmm15
10714 ; AVX512BW-NEXT: vpermi2q %zmm1, %zmm8, %zmm27
10715 ; AVX512BW-NEXT: vpermi2q %zmm1, %zmm8, %zmm24
10716 ; AVX512BW-NEXT: vmovdqa64 192(%r8), %zmm2
10717 ; AVX512BW-NEXT: vmovdqa64 192(%r9), %zmm0
10718 ; AVX512BW-NEXT: vpermi2q %zmm0, %zmm2, %zmm22
10719 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} zmm22 {%k1} = zmm8[0],zmm1[0],zmm8[2],zmm1[2],zmm8[4],zmm1[4],zmm8[6],zmm1[6]
10720 ; AVX512BW-NEXT: vpermi2q %zmm0, %zmm2, %zmm21
10721 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} zmm21 {%k1} = zmm8[1],zmm1[1],zmm8[3],zmm1[3],zmm8[5],zmm1[5],zmm8[7],zmm1[7]
10722 ; AVX512BW-NEXT: vmovdqa64 %zmm8, %zmm9
10723 ; AVX512BW-NEXT: vpermt2q %zmm1, %zmm23, %zmm9
10724 ; AVX512BW-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10725 ; AVX512BW-NEXT: vpermt2q %zmm1, %zmm26, %zmm8
10726 ; AVX512BW-NEXT: vpermi2q %zmm0, %zmm2, %zmm30
10727 ; AVX512BW-NEXT: vpermi2q %zmm0, %zmm2, %zmm31
10728 ; AVX512BW-NEXT: vpermi2q %zmm0, %zmm2, %zmm23
10729 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} zmm1 = zmm2[0],zmm0[0],zmm2[2],zmm0[2],zmm2[4],zmm0[4],zmm2[6],zmm0[6]
10730 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} zmm9 = zmm2[1],zmm0[1],zmm2[3],zmm0[3],zmm2[5],zmm0[5],zmm2[7],zmm0[7]
10731 ; AVX512BW-NEXT: vpermt2q %zmm0, %zmm26, %zmm2
10732 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
10733 ; AVX512BW-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
10734 ; AVX512BW-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
10735 ; AVX512BW-NEXT: vmovdqa64 %zmm13, %zmm5 {%k1}
10736 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm0, %zmm5, %zmm0
10737 ; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10738 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
10739 ; AVX512BW-NEXT: vpblendd $240, (%rsp), %ymm0, %ymm0 # 32-byte Folded Reload
10740 ; AVX512BW-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
10741 ; AVX512BW-NEXT: vmovdqa64 %zmm16, %zmm7 {%k1}
10742 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm0, %zmm7, %zmm0
10743 ; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10744 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
10745 ; AVX512BW-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
10746 ; AVX512BW-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
10747 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm0, %zmm3, %zmm0
10748 ; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10749 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
10750 ; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm29 {%k1}
10751 ; AVX512BW-NEXT: vmovdqa (%rsi), %xmm0
10752 ; AVX512BW-NEXT: vinserti128 $1, (%rcx), %ymm0, %ymm3
10753 ; AVX512BW-NEXT: vmovdqa (%rdi), %xmm0
10754 ; AVX512BW-NEXT: vinserti128 $1, (%rdx), %ymm0, %ymm5
10755 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm5[0],ymm3[0],ymm5[2],ymm3[2]
10756 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm0, %zmm29, %zmm0
10757 ; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10758 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
10759 ; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm18 {%k1}
10760 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} ymm3 = ymm5[1],ymm3[1],ymm5[3],ymm3[3]
10761 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm3, %zmm18, %zmm0
10762 ; AVX512BW-NEXT: vmovdqu64 %zmm0, (%rsp) # 64-byte Spill
10763 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
10764 ; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm14 {%k1}
10765 ; AVX512BW-NEXT: vmovdqa 64(%rsi), %xmm3
10766 ; AVX512BW-NEXT: vinserti128 $1, 64(%rcx), %ymm3, %ymm3
10767 ; AVX512BW-NEXT: vmovdqa 64(%rdi), %xmm5
10768 ; AVX512BW-NEXT: vinserti128 $1, 64(%rdx), %ymm5, %ymm5
10769 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm5[0],ymm3[0],ymm5[2],ymm3[2]
10770 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm7, %zmm14, %zmm29
10771 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
10772 ; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm11 {%k1}
10773 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} ymm3 = ymm5[1],ymm3[1],ymm5[3],ymm3[3]
10774 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm3, %zmm11, %zmm26
10775 ; AVX512BW-NEXT: vmovdqa64 %zmm28, %zmm4 {%k1}
10776 ; AVX512BW-NEXT: vmovdqa 128(%rsi), %xmm3
10777 ; AVX512BW-NEXT: vinserti128 $1, 128(%rcx), %ymm3, %ymm3
10778 ; AVX512BW-NEXT: vmovdqa 128(%rdi), %xmm5
10779 ; AVX512BW-NEXT: vinserti128 $1, 128(%rdx), %ymm5, %ymm5
10780 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm5[0],ymm3[0],ymm5[2],ymm3[2]
10781 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm7, %zmm4, %zmm28
10782 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
10783 ; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm6 {%k1}
10784 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} ymm3 = ymm5[1],ymm3[1],ymm5[3],ymm3[3]
10785 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm3, %zmm6, %zmm18
10786 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
10787 ; AVX512BW-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm4 # 32-byte Folded Reload
10788 ; AVX512BW-NEXT: # ymm4 = ymm0[0,1,2,3],mem[4,5,6,7]
10789 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm4, %zmm10, %zmm16
10790 ; AVX512BW-NEXT: vmovdqa64 %zmm12, %zmm30 {%k1}
10791 ; AVX512BW-NEXT: vmovdqa 192(%rsi), %xmm4
10792 ; AVX512BW-NEXT: vinserti128 $1, 192(%rcx), %ymm4, %ymm5
10793 ; AVX512BW-NEXT: vmovdqa 192(%rdi), %xmm4
10794 ; AVX512BW-NEXT: vinserti128 $1, 192(%rdx), %ymm4, %ymm6
10795 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} ymm4 = ymm6[0],ymm5[0],ymm6[2],ymm5[2]
10796 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm4, %zmm30, %zmm30
10797 ; AVX512BW-NEXT: vmovdqa64 %zmm15, %zmm31 {%k1}
10798 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} ymm5 = ymm6[1],ymm5[1],ymm6[3],ymm5[3]
10799 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm5, %zmm31, %zmm31
10800 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
10801 ; AVX512BW-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm6 # 32-byte Folded Reload
10802 ; AVX512BW-NEXT: # ymm6 = ymm0[0,1,2,3],mem[4,5,6,7]
10803 ; AVX512BW-NEXT: vmovdqa64 %zmm27, %zmm1 {%k1}
10804 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm6, %zmm1, %zmm6
10805 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
10806 ; AVX512BW-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm1 # 32-byte Folded Reload
10807 ; AVX512BW-NEXT: # ymm1 = ymm0[0,1,2,3],mem[4,5,6,7]
10808 ; AVX512BW-NEXT: vmovdqa64 %zmm24, %zmm9 {%k1}
10809 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm1, %zmm9, %zmm7
10810 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
10811 ; AVX512BW-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm1 # 32-byte Folded Reload
10812 ; AVX512BW-NEXT: # ymm1 = ymm0[0,1,2,3],mem[4,5,6,7]
10813 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm1, %zmm22, %zmm9
10814 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
10815 ; AVX512BW-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm1 # 32-byte Folded Reload
10816 ; AVX512BW-NEXT: # ymm1 = ymm0[0,1,2,3],mem[4,5,6,7]
10817 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm1, %zmm21, %zmm12
10818 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
10819 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
10820 ; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm3 {%k1}
10821 ; AVX512BW-NEXT: vmovdqa (%rcx), %ymm1
10822 ; AVX512BW-NEXT: vmovdqa (%rdx), %ymm15
10823 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm15[0],ymm1[0],ymm15[2],ymm1[2]
10824 ; AVX512BW-NEXT: vmovdqa64 (%rsi), %ymm21
10825 ; AVX512BW-NEXT: vmovdqa64 (%rdi), %ymm22
10826 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} ymm14 = ymm22[0],ymm21[0],ymm22[2],ymm21[2]
10827 ; AVX512BW-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm14[2,3],ymm0[2,3]
10828 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm0, %zmm3, %zmm0
10829 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
10830 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
10831 ; AVX512BW-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
10832 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm15[1],ymm1[1],ymm15[3],ymm1[3]
10833 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} ymm14 = ymm22[1],ymm21[1],ymm22[3],ymm21[3]
10834 ; AVX512BW-NEXT: vperm2i128 {{.*#+}} ymm1 = ymm14[2,3],ymm1[2,3]
10835 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm1, %zmm4, %zmm1
10836 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
10837 ; AVX512BW-NEXT: vmovdqa64 %zmm3, %zmm25 {%k1}
10838 ; AVX512BW-NEXT: vmovdqa 64(%rcx), %ymm14
10839 ; AVX512BW-NEXT: vmovdqa 64(%rdx), %ymm15
10840 ; AVX512BW-NEXT: vmovdqa64 64(%rsi), %ymm21
10841 ; AVX512BW-NEXT: vmovdqa64 64(%rdi), %ymm22
10842 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} ymm13 = ymm15[0],ymm14[0],ymm15[2],ymm14[2]
10843 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} ymm11 = ymm22[0],ymm21[0],ymm22[2],ymm21[2]
10844 ; AVX512BW-NEXT: vperm2i128 {{.*#+}} ymm11 = ymm11[2,3],ymm13[2,3]
10845 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm11, %zmm25, %zmm11
10846 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
10847 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
10848 ; AVX512BW-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
10849 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} ymm13 = ymm15[1],ymm14[1],ymm15[3],ymm14[3]
10850 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} ymm14 = ymm22[1],ymm21[1],ymm22[3],ymm21[3]
10851 ; AVX512BW-NEXT: vperm2i128 {{.*#+}} ymm13 = ymm14[2,3],ymm13[2,3]
10852 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm13, %zmm4, %zmm13
10853 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
10854 ; AVX512BW-NEXT: vmovdqa64 %zmm3, %zmm17 {%k1}
10855 ; AVX512BW-NEXT: vmovdqa 128(%rcx), %ymm14
10856 ; AVX512BW-NEXT: vmovdqa 128(%rdx), %ymm15
10857 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} ymm4 = ymm15[0],ymm14[0],ymm15[2],ymm14[2]
10858 ; AVX512BW-NEXT: vmovdqa64 128(%rsi), %ymm21
10859 ; AVX512BW-NEXT: vmovdqa64 128(%rdi), %ymm22
10860 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm22[0],ymm21[0],ymm22[2],ymm21[2]
10861 ; AVX512BW-NEXT: vperm2i128 {{.*#+}} ymm3 = ymm3[2,3],ymm4[2,3]
10862 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm3, %zmm17, %zmm3
10863 ; AVX512BW-NEXT: vmovdqa64 %zmm19, %zmm20 {%k1}
10864 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} ymm4 = ymm15[1],ymm14[1],ymm15[3],ymm14[3]
10865 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} ymm14 = ymm22[1],ymm21[1],ymm22[3],ymm21[3]
10866 ; AVX512BW-NEXT: vperm2i128 {{.*#+}} ymm4 = ymm14[2,3],ymm4[2,3]
10867 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm4, %zmm20, %zmm4
10868 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
10869 ; AVX512BW-NEXT: vmovdqa64 %zmm5, %zmm23 {%k1}
10870 ; AVX512BW-NEXT: vmovdqa 192(%rcx), %ymm14
10871 ; AVX512BW-NEXT: vmovdqa 192(%rdx), %ymm15
10872 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} ymm5 = ymm15[0],ymm14[0],ymm15[2],ymm14[2]
10873 ; AVX512BW-NEXT: vmovdqa64 192(%rsi), %ymm17
10874 ; AVX512BW-NEXT: vmovdqa64 192(%rdi), %ymm19
10875 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} ymm10 = ymm19[0],ymm17[0],ymm19[2],ymm17[2]
10876 ; AVX512BW-NEXT: vperm2i128 {{.*#+}} ymm5 = ymm10[2,3],ymm5[2,3]
10877 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm5, %zmm23, %zmm5
10878 ; AVX512BW-NEXT: vmovdqa64 %zmm8, %zmm2 {%k1}
10879 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} ymm8 = ymm15[1],ymm14[1],ymm15[3],ymm14[3]
10880 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} ymm10 = ymm19[1],ymm17[1],ymm19[3],ymm17[3]
10881 ; AVX512BW-NEXT: vperm2i128 {{.*#+}} ymm8 = ymm10[2,3],ymm8[2,3]
10882 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm8, %zmm2, %zmm2
10883 ; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
10884 ; AVX512BW-NEXT: vmovdqa64 %zmm2, 1728(%rax)
10885 ; AVX512BW-NEXT: vmovdqa64 %zmm5, 1664(%rax)
10886 ; AVX512BW-NEXT: vmovdqa64 %zmm4, 1216(%rax)
10887 ; AVX512BW-NEXT: vmovdqa64 %zmm3, 1152(%rax)
10888 ; AVX512BW-NEXT: vmovdqa64 %zmm13, 704(%rax)
10889 ; AVX512BW-NEXT: vmovdqa64 %zmm11, 640(%rax)
10890 ; AVX512BW-NEXT: vmovdqa64 %zmm1, 192(%rax)
10891 ; AVX512BW-NEXT: vmovdqa64 %zmm0, 128(%rax)
10892 ; AVX512BW-NEXT: vmovdqa64 %zmm12, 1984(%rax)
10893 ; AVX512BW-NEXT: vmovdqa64 %zmm9, 1920(%rax)
10894 ; AVX512BW-NEXT: vmovdqa64 %zmm7, 1856(%rax)
10895 ; AVX512BW-NEXT: vmovdqa64 %zmm6, 1792(%rax)
10896 ; AVX512BW-NEXT: vmovdqa64 %zmm31, 1600(%rax)
10897 ; AVX512BW-NEXT: vmovdqa64 %zmm30, 1536(%rax)
10898 ; AVX512BW-NEXT: vmovdqa64 %zmm16, 1472(%rax)
10899 ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
10900 ; AVX512BW-NEXT: vmovaps %zmm0, 1408(%rax)
10901 ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
10902 ; AVX512BW-NEXT: vmovaps %zmm0, 1344(%rax)
10903 ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
10904 ; AVX512BW-NEXT: vmovaps %zmm0, 1280(%rax)
10905 ; AVX512BW-NEXT: vmovdqa64 %zmm18, 1088(%rax)
10906 ; AVX512BW-NEXT: vmovdqa64 %zmm28, 1024(%rax)
10907 ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
10908 ; AVX512BW-NEXT: vmovaps %zmm0, 960(%rax)
10909 ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
10910 ; AVX512BW-NEXT: vmovaps %zmm0, 896(%rax)
10911 ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
10912 ; AVX512BW-NEXT: vmovaps %zmm0, 832(%rax)
10913 ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
10914 ; AVX512BW-NEXT: vmovaps %zmm0, 768(%rax)
10915 ; AVX512BW-NEXT: vmovdqa64 %zmm26, 576(%rax)
10916 ; AVX512BW-NEXT: vmovdqa64 %zmm29, 512(%rax)
10917 ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
10918 ; AVX512BW-NEXT: vmovaps %zmm0, 448(%rax)
10919 ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
10920 ; AVX512BW-NEXT: vmovaps %zmm0, 384(%rax)
10921 ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
10922 ; AVX512BW-NEXT: vmovaps %zmm0, 320(%rax)
10923 ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
10924 ; AVX512BW-NEXT: vmovaps %zmm0, 256(%rax)
10925 ; AVX512BW-NEXT: vmovups (%rsp), %zmm0 # 64-byte Reload
10926 ; AVX512BW-NEXT: vmovaps %zmm0, 64(%rax)
10927 ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
10928 ; AVX512BW-NEXT: vmovaps %zmm0, (%rax)
10929 ; AVX512BW-NEXT: addq $2312, %rsp # imm = 0x908
10930 ; AVX512BW-NEXT: vzeroupper
10931 ; AVX512BW-NEXT: retq
10932 ;
10933 ; AVX512BW-FCP-LABEL: store_i64_stride8_vf32:
10934 ; AVX512BW-FCP: # %bb.0:
10935 ; AVX512BW-FCP-NEXT: subq $2312, %rsp # imm = 0x908
10936 ; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
10937 ; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10
10938 ; AVX512BW-FCP-NEXT: vmovaps 128(%rdi), %zmm0
10939 ; AVX512BW-FCP-NEXT: vmovups %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10940 ; AVX512BW-FCP-NEXT: vmovdqa64 64(%rdi), %zmm6
10941 ; AVX512BW-FCP-NEXT: vmovdqa64 (%rdi), %zmm2
10942 ; AVX512BW-FCP-NEXT: vmovdqa64 128(%rsi), %zmm0
10943 ; AVX512BW-FCP-NEXT: vmovdqa64 64(%rsi), %zmm5
10944 ; AVX512BW-FCP-NEXT: vmovdqa64 (%rsi), %zmm3
10945 ; AVX512BW-FCP-NEXT: vmovaps 192(%rdx), %zmm1
10946 ; AVX512BW-FCP-NEXT: vmovups %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10947 ; AVX512BW-FCP-NEXT: vmovaps 128(%rdx), %zmm1
10948 ; AVX512BW-FCP-NEXT: vmovups %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10949 ; AVX512BW-FCP-NEXT: vmovdqa64 64(%rdx), %zmm7
10950 ; AVX512BW-FCP-NEXT: vmovdqa64 (%rdx), %zmm4
10951 ; AVX512BW-FCP-NEXT: vmovdqa64 192(%rcx), %zmm23
10952 ; AVX512BW-FCP-NEXT: vmovdqa64 128(%rcx), %zmm1
10953 ; AVX512BW-FCP-NEXT: vmovdqa64 64(%rcx), %zmm9
10954 ; AVX512BW-FCP-NEXT: vmovdqa64 (%rcx), %zmm8
10955 ; AVX512BW-FCP-NEXT: vmovdqa64 (%r8), %zmm18
10956 ; AVX512BW-FCP-NEXT: vmovdqa64 64(%r8), %zmm16
10957 ; AVX512BW-FCP-NEXT: vmovdqa64 (%r9), %zmm12
10958 ; AVX512BW-FCP-NEXT: vmovdqa64 64(%r9), %zmm19
10959 ; AVX512BW-FCP-NEXT: vmovdqa64 (%r10), %zmm25
10960 ; AVX512BW-FCP-NEXT: vmovdqa64 64(%r10), %zmm17
10961 ; AVX512BW-FCP-NEXT: vmovdqa64 (%rax), %zmm15
10962 ; AVX512BW-FCP-NEXT: vmovdqa64 64(%rax), %zmm20
10963 ; AVX512BW-FCP-NEXT: movb $-64, %r11b
10964 ; AVX512BW-FCP-NEXT: kmovd %r11d, %k1
10965 ; AVX512BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm27 = [4,12,4,12,4,12,4,12]
10966 ; AVX512BW-FCP-NEXT: # zmm27 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
10967 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm4, %zmm10
10968 ; AVX512BW-FCP-NEXT: vpermt2q %zmm8, %zmm27, %zmm10
10969 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm11
10970 ; AVX512BW-FCP-NEXT: vpermt2q %zmm3, %zmm27, %zmm11
10971 ; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm10 = ymm11[0,1,2,3],ymm10[4,5,6,7]
10972 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm25, %zmm11
10973 ; AVX512BW-FCP-NEXT: vpermt2q %zmm15, %zmm27, %zmm11
10974 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm13 = zmm18[0],zmm12[0],zmm18[2],zmm12[2],zmm18[4],zmm12[4],zmm18[6],zmm12[6]
10975 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm11, %zmm13 {%k1}
10976 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm10, %zmm13, %zmm10
10977 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10978 ; AVX512BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm24 = [5,13,5,13,5,13,5,13]
10979 ; AVX512BW-FCP-NEXT: # zmm24 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
10980 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm4, %zmm10
10981 ; AVX512BW-FCP-NEXT: vpermt2q %zmm8, %zmm24, %zmm10
10982 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm11
10983 ; AVX512BW-FCP-NEXT: vpermt2q %zmm3, %zmm24, %zmm11
10984 ; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm10 = ymm11[0,1,2,3],ymm10[4,5,6,7]
10985 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm25, %zmm11
10986 ; AVX512BW-FCP-NEXT: vpermt2q %zmm15, %zmm24, %zmm11
10987 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm14 = zmm18[1],zmm12[1],zmm18[3],zmm12[3],zmm18[5],zmm12[5],zmm18[7],zmm12[7]
10988 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm11, %zmm14 {%k1}
10989 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm10, %zmm14, %zmm10
10990 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
10991 ; AVX512BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm22 = [6,14,6,14,6,14,6,14]
10992 ; AVX512BW-FCP-NEXT: # zmm22 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
10993 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm4, %zmm10
10994 ; AVX512BW-FCP-NEXT: vpermt2q %zmm8, %zmm22, %zmm10
10995 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm11
10996 ; AVX512BW-FCP-NEXT: vpermt2q %zmm3, %zmm22, %zmm11
10997 ; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm10 = ymm11[0,1,2,3],ymm10[4,5,6,7]
10998 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm18, %zmm11
10999 ; AVX512BW-FCP-NEXT: vpermt2q %zmm12, %zmm22, %zmm11
11000 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm11 {%k1} = zmm25[0],zmm15[0],zmm25[2],zmm15[2],zmm25[4],zmm15[4],zmm25[6],zmm15[6]
11001 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm10, %zmm11, %zmm10
11002 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11003 ; AVX512BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm21 = [7,15,7,15,7,15,7,15]
11004 ; AVX512BW-FCP-NEXT: # zmm21 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
11005 ; AVX512BW-FCP-NEXT: vpermt2q %zmm8, %zmm21, %zmm4
11006 ; AVX512BW-FCP-NEXT: vpermt2q %zmm3, %zmm21, %zmm2
11007 ; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm4[4,5,6,7]
11008 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm18, %zmm3
11009 ; AVX512BW-FCP-NEXT: vpermt2q %zmm12, %zmm21, %zmm3
11010 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm3 {%k1} = zmm25[1],zmm15[1],zmm25[3],zmm15[3],zmm25[5],zmm15[5],zmm25[7],zmm15[7]
11011 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
11012 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11013 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm7, %zmm2
11014 ; AVX512BW-FCP-NEXT: vpermt2q %zmm9, %zmm27, %zmm2
11015 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm3
11016 ; AVX512BW-FCP-NEXT: vpermt2q %zmm5, %zmm27, %zmm3
11017 ; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3],ymm2[4,5,6,7]
11018 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm17, %zmm3
11019 ; AVX512BW-FCP-NEXT: vpermt2q %zmm20, %zmm27, %zmm3
11020 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm4 = zmm16[0],zmm19[0],zmm16[2],zmm19[2],zmm16[4],zmm19[4],zmm16[6],zmm19[6]
11021 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
11022 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
11023 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11024 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm7, %zmm2
11025 ; AVX512BW-FCP-NEXT: vpermt2q %zmm9, %zmm24, %zmm2
11026 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm3
11027 ; AVX512BW-FCP-NEXT: vpermt2q %zmm5, %zmm24, %zmm3
11028 ; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3],ymm2[4,5,6,7]
11029 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm17, %zmm3
11030 ; AVX512BW-FCP-NEXT: vpermt2q %zmm20, %zmm24, %zmm3
11031 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm4 = zmm16[1],zmm19[1],zmm16[3],zmm19[3],zmm16[5],zmm19[5],zmm16[7],zmm19[7]
11032 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
11033 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
11034 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11035 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm7, %zmm2
11036 ; AVX512BW-FCP-NEXT: vpermt2q %zmm9, %zmm22, %zmm2
11037 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm3
11038 ; AVX512BW-FCP-NEXT: vpermt2q %zmm5, %zmm22, %zmm3
11039 ; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3],ymm2[4,5,6,7]
11040 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm16, %zmm3
11041 ; AVX512BW-FCP-NEXT: vpermt2q %zmm19, %zmm22, %zmm3
11042 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm3 {%k1} = zmm17[0],zmm20[0],zmm17[2],zmm20[2],zmm17[4],zmm20[4],zmm17[6],zmm20[6]
11043 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
11044 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11045 ; AVX512BW-FCP-NEXT: vpermt2q %zmm9, %zmm21, %zmm7
11046 ; AVX512BW-FCP-NEXT: vpermt2q %zmm5, %zmm21, %zmm6
11047 ; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm6[0,1,2,3],ymm7[4,5,6,7]
11048 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm16, %zmm3
11049 ; AVX512BW-FCP-NEXT: vpermt2q %zmm19, %zmm21, %zmm3
11050 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm3 {%k1} = zmm17[1],zmm20[1],zmm17[3],zmm20[3],zmm17[5],zmm20[5],zmm17[7],zmm20[7]
11051 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
11052 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11053 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
11054 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm2
11055 ; AVX512BW-FCP-NEXT: vpermt2q %zmm1, %zmm27, %zmm2
11056 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11057 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
11058 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm4
11059 ; AVX512BW-FCP-NEXT: vpermt2q %zmm0, %zmm27, %zmm4
11060 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11061 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm4
11062 ; AVX512BW-FCP-NEXT: vpermt2q %zmm1, %zmm24, %zmm4
11063 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm4, (%rsp) # 64-byte Spill
11064 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm4
11065 ; AVX512BW-FCP-NEXT: vpermt2q %zmm0, %zmm24, %zmm4
11066 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11067 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm4
11068 ; AVX512BW-FCP-NEXT: vpermt2q %zmm1, %zmm22, %zmm4
11069 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11070 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm4
11071 ; AVX512BW-FCP-NEXT: vpermt2q %zmm0, %zmm22, %zmm4
11072 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11073 ; AVX512BW-FCP-NEXT: vpermt2q %zmm1, %zmm21, %zmm3
11074 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11075 ; AVX512BW-FCP-NEXT: vpermt2q %zmm0, %zmm21, %zmm2
11076 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11077 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
11078 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm1
11079 ; AVX512BW-FCP-NEXT: vpermt2q %zmm23, %zmm27, %zmm1
11080 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11081 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm1
11082 ; AVX512BW-FCP-NEXT: vpermt2q %zmm23, %zmm24, %zmm1
11083 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11084 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm1
11085 ; AVX512BW-FCP-NEXT: vpermt2q %zmm23, %zmm22, %zmm1
11086 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11087 ; AVX512BW-FCP-NEXT: vpermt2q %zmm23, %zmm21, %zmm0
11088 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11089 ; AVX512BW-FCP-NEXT: vmovdqa64 192(%rdi), %zmm1
11090 ; AVX512BW-FCP-NEXT: vmovdqa64 192(%rsi), %zmm0
11091 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm2
11092 ; AVX512BW-FCP-NEXT: vpermt2q %zmm0, %zmm27, %zmm2
11093 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11094 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm2
11095 ; AVX512BW-FCP-NEXT: vpermt2q %zmm0, %zmm24, %zmm2
11096 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11097 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm2
11098 ; AVX512BW-FCP-NEXT: vpermt2q %zmm0, %zmm22, %zmm2
11099 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11100 ; AVX512BW-FCP-NEXT: vpermt2q %zmm0, %zmm21, %zmm1
11101 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11102 ; AVX512BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm30 = [0,8,0,8,0,8,0,8]
11103 ; AVX512BW-FCP-NEXT: # zmm30 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
11104 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm25, %zmm0
11105 ; AVX512BW-FCP-NEXT: vpermt2q %zmm15, %zmm30, %zmm0
11106 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11107 ; AVX512BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm31 = [1,9,1,9,1,9,1,9]
11108 ; AVX512BW-FCP-NEXT: # zmm31 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
11109 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm25, %zmm0
11110 ; AVX512BW-FCP-NEXT: vpermt2q %zmm15, %zmm31, %zmm0
11111 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11112 ; AVX512BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm23 = [2,10,2,10,2,10,2,10]
11113 ; AVX512BW-FCP-NEXT: # zmm23 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
11114 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm25, %zmm0
11115 ; AVX512BW-FCP-NEXT: vpermt2q %zmm15, %zmm23, %zmm0
11116 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11117 ; AVX512BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm26 = [3,11,3,11,3,11,3,11]
11118 ; AVX512BW-FCP-NEXT: # zmm26 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
11119 ; AVX512BW-FCP-NEXT: vpermt2q %zmm15, %zmm26, %zmm25
11120 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm25, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11121 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm18, %zmm0
11122 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm18, %zmm29
11123 ; AVX512BW-FCP-NEXT: vpermt2q %zmm12, %zmm30, %zmm29
11124 ; AVX512BW-FCP-NEXT: vpermt2q %zmm12, %zmm31, %zmm18
11125 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm1
11126 ; AVX512BW-FCP-NEXT: vpermt2q %zmm12, %zmm23, %zmm1
11127 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11128 ; AVX512BW-FCP-NEXT: vpermt2q %zmm12, %zmm26, %zmm0
11129 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11130 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm17, %zmm0
11131 ; AVX512BW-FCP-NEXT: vpermt2q %zmm20, %zmm30, %zmm0
11132 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11133 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm17, %zmm0
11134 ; AVX512BW-FCP-NEXT: vpermt2q %zmm20, %zmm31, %zmm0
11135 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11136 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm17, %zmm0
11137 ; AVX512BW-FCP-NEXT: vpermt2q %zmm20, %zmm23, %zmm0
11138 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11139 ; AVX512BW-FCP-NEXT: vpermt2q %zmm20, %zmm26, %zmm17
11140 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm17, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11141 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm16, %zmm14
11142 ; AVX512BW-FCP-NEXT: vpermt2q %zmm19, %zmm30, %zmm14
11143 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm16, %zmm11
11144 ; AVX512BW-FCP-NEXT: vpermt2q %zmm19, %zmm31, %zmm11
11145 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm16, %zmm25
11146 ; AVX512BW-FCP-NEXT: vpermt2q %zmm19, %zmm23, %zmm25
11147 ; AVX512BW-FCP-NEXT: vpermt2q %zmm19, %zmm26, %zmm16
11148 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11149 ; AVX512BW-FCP-NEXT: vmovdqa64 128(%r10), %zmm19
11150 ; AVX512BW-FCP-NEXT: vmovdqa64 128(%rax), %zmm0
11151 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm19, %zmm28
11152 ; AVX512BW-FCP-NEXT: vpermt2q %zmm0, %zmm30, %zmm28
11153 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm19, %zmm1
11154 ; AVX512BW-FCP-NEXT: vpermt2q %zmm0, %zmm31, %zmm1
11155 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11156 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm19, %zmm13
11157 ; AVX512BW-FCP-NEXT: vpermt2q %zmm0, %zmm27, %zmm13
11158 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm19, %zmm16
11159 ; AVX512BW-FCP-NEXT: vpermt2q %zmm0, %zmm24, %zmm16
11160 ; AVX512BW-FCP-NEXT: vmovdqa64 128(%r8), %zmm20
11161 ; AVX512BW-FCP-NEXT: vmovdqa64 128(%r9), %zmm1
11162 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm20, %zmm3
11163 ; AVX512BW-FCP-NEXT: vpermt2q %zmm1, %zmm22, %zmm3
11164 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm3 {%k1} = zmm19[0],zmm0[0],zmm19[2],zmm0[2],zmm19[4],zmm0[4],zmm19[6],zmm0[6]
11165 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm20, %zmm10
11166 ; AVX512BW-FCP-NEXT: vpermt2q %zmm1, %zmm21, %zmm10
11167 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm10 {%k1} = zmm19[1],zmm0[1],zmm19[3],zmm0[3],zmm19[5],zmm0[5],zmm19[7],zmm0[7]
11168 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm19, %zmm2
11169 ; AVX512BW-FCP-NEXT: vpermt2q %zmm0, %zmm23, %zmm2
11170 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11171 ; AVX512BW-FCP-NEXT: vpermt2q %zmm0, %zmm26, %zmm19
11172 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm20, %zmm4
11173 ; AVX512BW-FCP-NEXT: vpermt2q %zmm1, %zmm30, %zmm4
11174 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm20, %zmm6
11175 ; AVX512BW-FCP-NEXT: vpermt2q %zmm1, %zmm31, %zmm6
11176 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm20, %zmm17
11177 ; AVX512BW-FCP-NEXT: vpermt2q %zmm1, %zmm23, %zmm17
11178 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm5 = zmm20[0],zmm1[0],zmm20[2],zmm1[2],zmm20[4],zmm1[4],zmm20[6],zmm1[6]
11179 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm7 = zmm20[1],zmm1[1],zmm20[3],zmm1[3],zmm20[5],zmm1[5],zmm20[7],zmm1[7]
11180 ; AVX512BW-FCP-NEXT: vpermt2q %zmm1, %zmm26, %zmm20
11181 ; AVX512BW-FCP-NEXT: vmovdqa64 192(%r10), %zmm8
11182 ; AVX512BW-FCP-NEXT: vmovdqa64 192(%rax), %zmm1
11183 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm8, %zmm12
11184 ; AVX512BW-FCP-NEXT: vpermt2q %zmm1, %zmm30, %zmm12
11185 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm8, %zmm15
11186 ; AVX512BW-FCP-NEXT: vpermt2q %zmm1, %zmm31, %zmm15
11187 ; AVX512BW-FCP-NEXT: vpermi2q %zmm1, %zmm8, %zmm27
11188 ; AVX512BW-FCP-NEXT: vpermi2q %zmm1, %zmm8, %zmm24
11189 ; AVX512BW-FCP-NEXT: vmovdqa64 192(%r8), %zmm2
11190 ; AVX512BW-FCP-NEXT: vmovdqa64 192(%r9), %zmm0
11191 ; AVX512BW-FCP-NEXT: vpermi2q %zmm0, %zmm2, %zmm22
11192 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm22 {%k1} = zmm8[0],zmm1[0],zmm8[2],zmm1[2],zmm8[4],zmm1[4],zmm8[6],zmm1[6]
11193 ; AVX512BW-FCP-NEXT: vpermi2q %zmm0, %zmm2, %zmm21
11194 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm21 {%k1} = zmm8[1],zmm1[1],zmm8[3],zmm1[3],zmm8[5],zmm1[5],zmm8[7],zmm1[7]
11195 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm8, %zmm9
11196 ; AVX512BW-FCP-NEXT: vpermt2q %zmm1, %zmm23, %zmm9
11197 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11198 ; AVX512BW-FCP-NEXT: vpermt2q %zmm1, %zmm26, %zmm8
11199 ; AVX512BW-FCP-NEXT: vpermi2q %zmm0, %zmm2, %zmm30
11200 ; AVX512BW-FCP-NEXT: vpermi2q %zmm0, %zmm2, %zmm31
11201 ; AVX512BW-FCP-NEXT: vpermi2q %zmm0, %zmm2, %zmm23
11202 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm1 = zmm2[0],zmm0[0],zmm2[2],zmm0[2],zmm2[4],zmm0[4],zmm2[6],zmm0[6]
11203 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm9 = zmm2[1],zmm0[1],zmm2[3],zmm0[3],zmm2[5],zmm0[5],zmm2[7],zmm0[7]
11204 ; AVX512BW-FCP-NEXT: vpermt2q %zmm0, %zmm26, %zmm2
11205 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
11206 ; AVX512BW-FCP-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
11207 ; AVX512BW-FCP-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
11208 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm13, %zmm5 {%k1}
11209 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm5, %zmm0
11210 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11211 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
11212 ; AVX512BW-FCP-NEXT: vpblendd $240, (%rsp), %ymm0, %ymm0 # 32-byte Folded Reload
11213 ; AVX512BW-FCP-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
11214 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm16, %zmm7 {%k1}
11215 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm7, %zmm0
11216 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11217 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
11218 ; AVX512BW-FCP-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
11219 ; AVX512BW-FCP-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
11220 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm3, %zmm0
11221 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11222 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
11223 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm29 {%k1}
11224 ; AVX512BW-FCP-NEXT: vmovdqa (%rsi), %xmm0
11225 ; AVX512BW-FCP-NEXT: vinserti128 $1, (%rcx), %ymm0, %ymm3
11226 ; AVX512BW-FCP-NEXT: vmovdqa (%rdi), %xmm0
11227 ; AVX512BW-FCP-NEXT: vinserti128 $1, (%rdx), %ymm0, %ymm5
11228 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm5[0],ymm3[0],ymm5[2],ymm3[2]
11229 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm29, %zmm0
11230 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11231 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
11232 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm18 {%k1}
11233 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm3 = ymm5[1],ymm3[1],ymm5[3],ymm3[3]
11234 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm3, %zmm18, %zmm0
11235 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, (%rsp) # 64-byte Spill
11236 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
11237 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm14 {%k1}
11238 ; AVX512BW-FCP-NEXT: vmovdqa 64(%rsi), %xmm3
11239 ; AVX512BW-FCP-NEXT: vinserti128 $1, 64(%rcx), %ymm3, %ymm3
11240 ; AVX512BW-FCP-NEXT: vmovdqa 64(%rdi), %xmm5
11241 ; AVX512BW-FCP-NEXT: vinserti128 $1, 64(%rdx), %ymm5, %ymm5
11242 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm5[0],ymm3[0],ymm5[2],ymm3[2]
11243 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm7, %zmm14, %zmm29
11244 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
11245 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm11 {%k1}
11246 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm3 = ymm5[1],ymm3[1],ymm5[3],ymm3[3]
11247 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm3, %zmm11, %zmm26
11248 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm28, %zmm4 {%k1}
11249 ; AVX512BW-FCP-NEXT: vmovdqa 128(%rsi), %xmm3
11250 ; AVX512BW-FCP-NEXT: vinserti128 $1, 128(%rcx), %ymm3, %ymm3
11251 ; AVX512BW-FCP-NEXT: vmovdqa 128(%rdi), %xmm5
11252 ; AVX512BW-FCP-NEXT: vinserti128 $1, 128(%rdx), %ymm5, %ymm5
11253 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm5[0],ymm3[0],ymm5[2],ymm3[2]
11254 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm7, %zmm4, %zmm28
11255 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
11256 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm6 {%k1}
11257 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm3 = ymm5[1],ymm3[1],ymm5[3],ymm3[3]
11258 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm3, %zmm6, %zmm18
11259 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
11260 ; AVX512BW-FCP-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm4 # 32-byte Folded Reload
11261 ; AVX512BW-FCP-NEXT: # ymm4 = ymm0[0,1,2,3],mem[4,5,6,7]
11262 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm4, %zmm10, %zmm16
11263 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm12, %zmm30 {%k1}
11264 ; AVX512BW-FCP-NEXT: vmovdqa 192(%rsi), %xmm4
11265 ; AVX512BW-FCP-NEXT: vinserti128 $1, 192(%rcx), %ymm4, %ymm5
11266 ; AVX512BW-FCP-NEXT: vmovdqa 192(%rdi), %xmm4
11267 ; AVX512BW-FCP-NEXT: vinserti128 $1, 192(%rdx), %ymm4, %ymm6
11268 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm4 = ymm6[0],ymm5[0],ymm6[2],ymm5[2]
11269 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm4, %zmm30, %zmm30
11270 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm15, %zmm31 {%k1}
11271 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm5 = ymm6[1],ymm5[1],ymm6[3],ymm5[3]
11272 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm5, %zmm31, %zmm31
11273 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
11274 ; AVX512BW-FCP-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm6 # 32-byte Folded Reload
11275 ; AVX512BW-FCP-NEXT: # ymm6 = ymm0[0,1,2,3],mem[4,5,6,7]
11276 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm27, %zmm1 {%k1}
11277 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm1, %zmm6
11278 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
11279 ; AVX512BW-FCP-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm1 # 32-byte Folded Reload
11280 ; AVX512BW-FCP-NEXT: # ymm1 = ymm0[0,1,2,3],mem[4,5,6,7]
11281 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm24, %zmm9 {%k1}
11282 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm1, %zmm9, %zmm7
11283 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
11284 ; AVX512BW-FCP-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm1 # 32-byte Folded Reload
11285 ; AVX512BW-FCP-NEXT: # ymm1 = ymm0[0,1,2,3],mem[4,5,6,7]
11286 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm1, %zmm22, %zmm9
11287 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
11288 ; AVX512BW-FCP-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm1 # 32-byte Folded Reload
11289 ; AVX512BW-FCP-NEXT: # ymm1 = ymm0[0,1,2,3],mem[4,5,6,7]
11290 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm1, %zmm21, %zmm12
11291 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
11292 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
11293 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm3 {%k1}
11294 ; AVX512BW-FCP-NEXT: vmovdqa (%rcx), %ymm1
11295 ; AVX512BW-FCP-NEXT: vmovdqa (%rdx), %ymm15
11296 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm15[0],ymm1[0],ymm15[2],ymm1[2]
11297 ; AVX512BW-FCP-NEXT: vmovdqa64 (%rsi), %ymm21
11298 ; AVX512BW-FCP-NEXT: vmovdqa64 (%rdi), %ymm22
11299 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm14 = ymm22[0],ymm21[0],ymm22[2],ymm21[2]
11300 ; AVX512BW-FCP-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm14[2,3],ymm0[2,3]
11301 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm3, %zmm0
11302 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
11303 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
11304 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
11305 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm15[1],ymm1[1],ymm15[3],ymm1[3]
11306 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm14 = ymm22[1],ymm21[1],ymm22[3],ymm21[3]
11307 ; AVX512BW-FCP-NEXT: vperm2i128 {{.*#+}} ymm1 = ymm14[2,3],ymm1[2,3]
11308 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm1, %zmm4, %zmm1
11309 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
11310 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm25 {%k1}
11311 ; AVX512BW-FCP-NEXT: vmovdqa 64(%rcx), %ymm14
11312 ; AVX512BW-FCP-NEXT: vmovdqa 64(%rdx), %ymm15
11313 ; AVX512BW-FCP-NEXT: vmovdqa64 64(%rsi), %ymm21
11314 ; AVX512BW-FCP-NEXT: vmovdqa64 64(%rdi), %ymm22
11315 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm13 = ymm15[0],ymm14[0],ymm15[2],ymm14[2]
11316 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm11 = ymm22[0],ymm21[0],ymm22[2],ymm21[2]
11317 ; AVX512BW-FCP-NEXT: vperm2i128 {{.*#+}} ymm11 = ymm11[2,3],ymm13[2,3]
11318 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm11, %zmm25, %zmm11
11319 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
11320 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
11321 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
11322 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm13 = ymm15[1],ymm14[1],ymm15[3],ymm14[3]
11323 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm14 = ymm22[1],ymm21[1],ymm22[3],ymm21[3]
11324 ; AVX512BW-FCP-NEXT: vperm2i128 {{.*#+}} ymm13 = ymm14[2,3],ymm13[2,3]
11325 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm13, %zmm4, %zmm13
11326 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
11327 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm17 {%k1}
11328 ; AVX512BW-FCP-NEXT: vmovdqa 128(%rcx), %ymm14
11329 ; AVX512BW-FCP-NEXT: vmovdqa 128(%rdx), %ymm15
11330 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm4 = ymm15[0],ymm14[0],ymm15[2],ymm14[2]
11331 ; AVX512BW-FCP-NEXT: vmovdqa64 128(%rsi), %ymm21
11332 ; AVX512BW-FCP-NEXT: vmovdqa64 128(%rdi), %ymm22
11333 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm22[0],ymm21[0],ymm22[2],ymm21[2]
11334 ; AVX512BW-FCP-NEXT: vperm2i128 {{.*#+}} ymm3 = ymm3[2,3],ymm4[2,3]
11335 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm3, %zmm17, %zmm3
11336 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm19, %zmm20 {%k1}
11337 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm4 = ymm15[1],ymm14[1],ymm15[3],ymm14[3]
11338 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm14 = ymm22[1],ymm21[1],ymm22[3],ymm21[3]
11339 ; AVX512BW-FCP-NEXT: vperm2i128 {{.*#+}} ymm4 = ymm14[2,3],ymm4[2,3]
11340 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm4, %zmm20, %zmm4
11341 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
11342 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm5, %zmm23 {%k1}
11343 ; AVX512BW-FCP-NEXT: vmovdqa 192(%rcx), %ymm14
11344 ; AVX512BW-FCP-NEXT: vmovdqa 192(%rdx), %ymm15
11345 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm5 = ymm15[0],ymm14[0],ymm15[2],ymm14[2]
11346 ; AVX512BW-FCP-NEXT: vmovdqa64 192(%rsi), %ymm17
11347 ; AVX512BW-FCP-NEXT: vmovdqa64 192(%rdi), %ymm19
11348 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm10 = ymm19[0],ymm17[0],ymm19[2],ymm17[2]
11349 ; AVX512BW-FCP-NEXT: vperm2i128 {{.*#+}} ymm5 = ymm10[2,3],ymm5[2,3]
11350 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm5, %zmm23, %zmm5
11351 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm8, %zmm2 {%k1}
11352 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm8 = ymm15[1],ymm14[1],ymm15[3],ymm14[3]
11353 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm10 = ymm19[1],ymm17[1],ymm19[3],ymm17[3]
11354 ; AVX512BW-FCP-NEXT: vperm2i128 {{.*#+}} ymm8 = ymm10[2,3],ymm8[2,3]
11355 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm8, %zmm2, %zmm2
11356 ; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
11357 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, 1728(%rax)
11358 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm5, 1664(%rax)
11359 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm4, 1216(%rax)
11360 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, 1152(%rax)
11361 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm13, 704(%rax)
11362 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm11, 640(%rax)
11363 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm1, 192(%rax)
11364 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, 128(%rax)
11365 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm12, 1984(%rax)
11366 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm9, 1920(%rax)
11367 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm7, 1856(%rax)
11368 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm6, 1792(%rax)
11369 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm31, 1600(%rax)
11370 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm30, 1536(%rax)
11371 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm16, 1472(%rax)
11372 ; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
11373 ; AVX512BW-FCP-NEXT: vmovaps %zmm0, 1408(%rax)
11374 ; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
11375 ; AVX512BW-FCP-NEXT: vmovaps %zmm0, 1344(%rax)
11376 ; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
11377 ; AVX512BW-FCP-NEXT: vmovaps %zmm0, 1280(%rax)
11378 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm18, 1088(%rax)
11379 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm28, 1024(%rax)
11380 ; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
11381 ; AVX512BW-FCP-NEXT: vmovaps %zmm0, 960(%rax)
11382 ; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
11383 ; AVX512BW-FCP-NEXT: vmovaps %zmm0, 896(%rax)
11384 ; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
11385 ; AVX512BW-FCP-NEXT: vmovaps %zmm0, 832(%rax)
11386 ; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
11387 ; AVX512BW-FCP-NEXT: vmovaps %zmm0, 768(%rax)
11388 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm26, 576(%rax)
11389 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm29, 512(%rax)
11390 ; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
11391 ; AVX512BW-FCP-NEXT: vmovaps %zmm0, 448(%rax)
11392 ; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
11393 ; AVX512BW-FCP-NEXT: vmovaps %zmm0, 384(%rax)
11394 ; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
11395 ; AVX512BW-FCP-NEXT: vmovaps %zmm0, 320(%rax)
11396 ; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
11397 ; AVX512BW-FCP-NEXT: vmovaps %zmm0, 256(%rax)
11398 ; AVX512BW-FCP-NEXT: vmovups (%rsp), %zmm0 # 64-byte Reload
11399 ; AVX512BW-FCP-NEXT: vmovaps %zmm0, 64(%rax)
11400 ; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
11401 ; AVX512BW-FCP-NEXT: vmovaps %zmm0, (%rax)
11402 ; AVX512BW-FCP-NEXT: addq $2312, %rsp # imm = 0x908
11403 ; AVX512BW-FCP-NEXT: vzeroupper
11404 ; AVX512BW-FCP-NEXT: retq
11405 ;
11406 ; AVX512DQ-BW-LABEL: store_i64_stride8_vf32:
11407 ; AVX512DQ-BW: # %bb.0:
11408 ; AVX512DQ-BW-NEXT: subq $2312, %rsp # imm = 0x908
11409 ; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
11410 ; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %r10
11411 ; AVX512DQ-BW-NEXT: vmovaps 128(%rdi), %zmm0
11412 ; AVX512DQ-BW-NEXT: vmovups %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11413 ; AVX512DQ-BW-NEXT: vmovdqa64 64(%rdi), %zmm6
11414 ; AVX512DQ-BW-NEXT: vmovdqa64 (%rdi), %zmm2
11415 ; AVX512DQ-BW-NEXT: vmovdqa64 128(%rsi), %zmm0
11416 ; AVX512DQ-BW-NEXT: vmovdqa64 64(%rsi), %zmm5
11417 ; AVX512DQ-BW-NEXT: vmovdqa64 (%rsi), %zmm3
11418 ; AVX512DQ-BW-NEXT: vmovaps 192(%rdx), %zmm1
11419 ; AVX512DQ-BW-NEXT: vmovups %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11420 ; AVX512DQ-BW-NEXT: vmovaps 128(%rdx), %zmm1
11421 ; AVX512DQ-BW-NEXT: vmovups %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11422 ; AVX512DQ-BW-NEXT: vmovdqa64 64(%rdx), %zmm7
11423 ; AVX512DQ-BW-NEXT: vmovdqa64 (%rdx), %zmm4
11424 ; AVX512DQ-BW-NEXT: vmovdqa64 192(%rcx), %zmm23
11425 ; AVX512DQ-BW-NEXT: vmovdqa64 128(%rcx), %zmm1
11426 ; AVX512DQ-BW-NEXT: vmovdqa64 64(%rcx), %zmm9
11427 ; AVX512DQ-BW-NEXT: vmovdqa64 (%rcx), %zmm8
11428 ; AVX512DQ-BW-NEXT: vmovdqa64 (%r8), %zmm18
11429 ; AVX512DQ-BW-NEXT: vmovdqa64 64(%r8), %zmm16
11430 ; AVX512DQ-BW-NEXT: vmovdqa64 (%r9), %zmm12
11431 ; AVX512DQ-BW-NEXT: vmovdqa64 64(%r9), %zmm19
11432 ; AVX512DQ-BW-NEXT: vmovdqa64 (%r10), %zmm25
11433 ; AVX512DQ-BW-NEXT: vmovdqa64 64(%r10), %zmm17
11434 ; AVX512DQ-BW-NEXT: vmovdqa64 (%rax), %zmm15
11435 ; AVX512DQ-BW-NEXT: vmovdqa64 64(%rax), %zmm20
11436 ; AVX512DQ-BW-NEXT: movb $-64, %r11b
11437 ; AVX512DQ-BW-NEXT: kmovd %r11d, %k1
11438 ; AVX512DQ-BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm27 = [4,12,4,12,4,12,4,12]
11439 ; AVX512DQ-BW-NEXT: # zmm27 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
11440 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm4, %zmm10
11441 ; AVX512DQ-BW-NEXT: vpermt2q %zmm8, %zmm27, %zmm10
11442 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm11
11443 ; AVX512DQ-BW-NEXT: vpermt2q %zmm3, %zmm27, %zmm11
11444 ; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm10 = ymm11[0,1,2,3],ymm10[4,5,6,7]
11445 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm25, %zmm11
11446 ; AVX512DQ-BW-NEXT: vpermt2q %zmm15, %zmm27, %zmm11
11447 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} zmm13 = zmm18[0],zmm12[0],zmm18[2],zmm12[2],zmm18[4],zmm12[4],zmm18[6],zmm12[6]
11448 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm11, %zmm13 {%k1}
11449 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm10, %zmm13, %zmm10
11450 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11451 ; AVX512DQ-BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm24 = [5,13,5,13,5,13,5,13]
11452 ; AVX512DQ-BW-NEXT: # zmm24 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
11453 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm4, %zmm10
11454 ; AVX512DQ-BW-NEXT: vpermt2q %zmm8, %zmm24, %zmm10
11455 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm11
11456 ; AVX512DQ-BW-NEXT: vpermt2q %zmm3, %zmm24, %zmm11
11457 ; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm10 = ymm11[0,1,2,3],ymm10[4,5,6,7]
11458 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm25, %zmm11
11459 ; AVX512DQ-BW-NEXT: vpermt2q %zmm15, %zmm24, %zmm11
11460 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} zmm14 = zmm18[1],zmm12[1],zmm18[3],zmm12[3],zmm18[5],zmm12[5],zmm18[7],zmm12[7]
11461 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm11, %zmm14 {%k1}
11462 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm10, %zmm14, %zmm10
11463 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11464 ; AVX512DQ-BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm22 = [6,14,6,14,6,14,6,14]
11465 ; AVX512DQ-BW-NEXT: # zmm22 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
11466 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm4, %zmm10
11467 ; AVX512DQ-BW-NEXT: vpermt2q %zmm8, %zmm22, %zmm10
11468 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm11
11469 ; AVX512DQ-BW-NEXT: vpermt2q %zmm3, %zmm22, %zmm11
11470 ; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm10 = ymm11[0,1,2,3],ymm10[4,5,6,7]
11471 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm18, %zmm11
11472 ; AVX512DQ-BW-NEXT: vpermt2q %zmm12, %zmm22, %zmm11
11473 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} zmm11 {%k1} = zmm25[0],zmm15[0],zmm25[2],zmm15[2],zmm25[4],zmm15[4],zmm25[6],zmm15[6]
11474 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm10, %zmm11, %zmm10
11475 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11476 ; AVX512DQ-BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm21 = [7,15,7,15,7,15,7,15]
11477 ; AVX512DQ-BW-NEXT: # zmm21 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
11478 ; AVX512DQ-BW-NEXT: vpermt2q %zmm8, %zmm21, %zmm4
11479 ; AVX512DQ-BW-NEXT: vpermt2q %zmm3, %zmm21, %zmm2
11480 ; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm4[4,5,6,7]
11481 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm18, %zmm3
11482 ; AVX512DQ-BW-NEXT: vpermt2q %zmm12, %zmm21, %zmm3
11483 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} zmm3 {%k1} = zmm25[1],zmm15[1],zmm25[3],zmm15[3],zmm25[5],zmm15[5],zmm25[7],zmm15[7]
11484 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
11485 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11486 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm7, %zmm2
11487 ; AVX512DQ-BW-NEXT: vpermt2q %zmm9, %zmm27, %zmm2
11488 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm6, %zmm3
11489 ; AVX512DQ-BW-NEXT: vpermt2q %zmm5, %zmm27, %zmm3
11490 ; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3],ymm2[4,5,6,7]
11491 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm17, %zmm3
11492 ; AVX512DQ-BW-NEXT: vpermt2q %zmm20, %zmm27, %zmm3
11493 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} zmm4 = zmm16[0],zmm19[0],zmm16[2],zmm19[2],zmm16[4],zmm19[4],zmm16[6],zmm19[6]
11494 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
11495 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
11496 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11497 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm7, %zmm2
11498 ; AVX512DQ-BW-NEXT: vpermt2q %zmm9, %zmm24, %zmm2
11499 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm6, %zmm3
11500 ; AVX512DQ-BW-NEXT: vpermt2q %zmm5, %zmm24, %zmm3
11501 ; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3],ymm2[4,5,6,7]
11502 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm17, %zmm3
11503 ; AVX512DQ-BW-NEXT: vpermt2q %zmm20, %zmm24, %zmm3
11504 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} zmm4 = zmm16[1],zmm19[1],zmm16[3],zmm19[3],zmm16[5],zmm19[5],zmm16[7],zmm19[7]
11505 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
11506 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
11507 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11508 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm7, %zmm2
11509 ; AVX512DQ-BW-NEXT: vpermt2q %zmm9, %zmm22, %zmm2
11510 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm6, %zmm3
11511 ; AVX512DQ-BW-NEXT: vpermt2q %zmm5, %zmm22, %zmm3
11512 ; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3],ymm2[4,5,6,7]
11513 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm16, %zmm3
11514 ; AVX512DQ-BW-NEXT: vpermt2q %zmm19, %zmm22, %zmm3
11515 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} zmm3 {%k1} = zmm17[0],zmm20[0],zmm17[2],zmm20[2],zmm17[4],zmm20[4],zmm17[6],zmm20[6]
11516 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
11517 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11518 ; AVX512DQ-BW-NEXT: vpermt2q %zmm9, %zmm21, %zmm7
11519 ; AVX512DQ-BW-NEXT: vpermt2q %zmm5, %zmm21, %zmm6
11520 ; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm2 = ymm6[0,1,2,3],ymm7[4,5,6,7]
11521 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm16, %zmm3
11522 ; AVX512DQ-BW-NEXT: vpermt2q %zmm19, %zmm21, %zmm3
11523 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} zmm3 {%k1} = zmm17[1],zmm20[1],zmm17[3],zmm20[3],zmm17[5],zmm20[5],zmm17[7],zmm20[7]
11524 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
11525 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11526 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
11527 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, %zmm2
11528 ; AVX512DQ-BW-NEXT: vpermt2q %zmm1, %zmm27, %zmm2
11529 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11530 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
11531 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm4
11532 ; AVX512DQ-BW-NEXT: vpermt2q %zmm0, %zmm27, %zmm4
11533 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11534 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, %zmm4
11535 ; AVX512DQ-BW-NEXT: vpermt2q %zmm1, %zmm24, %zmm4
11536 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm4, (%rsp) # 64-byte Spill
11537 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm4
11538 ; AVX512DQ-BW-NEXT: vpermt2q %zmm0, %zmm24, %zmm4
11539 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11540 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, %zmm4
11541 ; AVX512DQ-BW-NEXT: vpermt2q %zmm1, %zmm22, %zmm4
11542 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11543 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm4
11544 ; AVX512DQ-BW-NEXT: vpermt2q %zmm0, %zmm22, %zmm4
11545 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11546 ; AVX512DQ-BW-NEXT: vpermt2q %zmm1, %zmm21, %zmm3
11547 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11548 ; AVX512DQ-BW-NEXT: vpermt2q %zmm0, %zmm21, %zmm2
11549 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11550 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
11551 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, %zmm1
11552 ; AVX512DQ-BW-NEXT: vpermt2q %zmm23, %zmm27, %zmm1
11553 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11554 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, %zmm1
11555 ; AVX512DQ-BW-NEXT: vpermt2q %zmm23, %zmm24, %zmm1
11556 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11557 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, %zmm1
11558 ; AVX512DQ-BW-NEXT: vpermt2q %zmm23, %zmm22, %zmm1
11559 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11560 ; AVX512DQ-BW-NEXT: vpermt2q %zmm23, %zmm21, %zmm0
11561 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11562 ; AVX512DQ-BW-NEXT: vmovdqa64 192(%rdi), %zmm1
11563 ; AVX512DQ-BW-NEXT: vmovdqa64 192(%rsi), %zmm0
11564 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm1, %zmm2
11565 ; AVX512DQ-BW-NEXT: vpermt2q %zmm0, %zmm27, %zmm2
11566 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11567 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm1, %zmm2
11568 ; AVX512DQ-BW-NEXT: vpermt2q %zmm0, %zmm24, %zmm2
11569 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11570 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm1, %zmm2
11571 ; AVX512DQ-BW-NEXT: vpermt2q %zmm0, %zmm22, %zmm2
11572 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11573 ; AVX512DQ-BW-NEXT: vpermt2q %zmm0, %zmm21, %zmm1
11574 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11575 ; AVX512DQ-BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm30 = [0,8,0,8,0,8,0,8]
11576 ; AVX512DQ-BW-NEXT: # zmm30 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
11577 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm25, %zmm0
11578 ; AVX512DQ-BW-NEXT: vpermt2q %zmm15, %zmm30, %zmm0
11579 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11580 ; AVX512DQ-BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm31 = [1,9,1,9,1,9,1,9]
11581 ; AVX512DQ-BW-NEXT: # zmm31 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
11582 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm25, %zmm0
11583 ; AVX512DQ-BW-NEXT: vpermt2q %zmm15, %zmm31, %zmm0
11584 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11585 ; AVX512DQ-BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm23 = [2,10,2,10,2,10,2,10]
11586 ; AVX512DQ-BW-NEXT: # zmm23 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
11587 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm25, %zmm0
11588 ; AVX512DQ-BW-NEXT: vpermt2q %zmm15, %zmm23, %zmm0
11589 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11590 ; AVX512DQ-BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm26 = [3,11,3,11,3,11,3,11]
11591 ; AVX512DQ-BW-NEXT: # zmm26 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
11592 ; AVX512DQ-BW-NEXT: vpermt2q %zmm15, %zmm26, %zmm25
11593 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm25, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11594 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm18, %zmm0
11595 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm18, %zmm29
11596 ; AVX512DQ-BW-NEXT: vpermt2q %zmm12, %zmm30, %zmm29
11597 ; AVX512DQ-BW-NEXT: vpermt2q %zmm12, %zmm31, %zmm18
11598 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, %zmm1
11599 ; AVX512DQ-BW-NEXT: vpermt2q %zmm12, %zmm23, %zmm1
11600 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11601 ; AVX512DQ-BW-NEXT: vpermt2q %zmm12, %zmm26, %zmm0
11602 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11603 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm17, %zmm0
11604 ; AVX512DQ-BW-NEXT: vpermt2q %zmm20, %zmm30, %zmm0
11605 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11606 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm17, %zmm0
11607 ; AVX512DQ-BW-NEXT: vpermt2q %zmm20, %zmm31, %zmm0
11608 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11609 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm17, %zmm0
11610 ; AVX512DQ-BW-NEXT: vpermt2q %zmm20, %zmm23, %zmm0
11611 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11612 ; AVX512DQ-BW-NEXT: vpermt2q %zmm20, %zmm26, %zmm17
11613 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm17, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11614 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm16, %zmm14
11615 ; AVX512DQ-BW-NEXT: vpermt2q %zmm19, %zmm30, %zmm14
11616 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm16, %zmm11
11617 ; AVX512DQ-BW-NEXT: vpermt2q %zmm19, %zmm31, %zmm11
11618 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm16, %zmm25
11619 ; AVX512DQ-BW-NEXT: vpermt2q %zmm19, %zmm23, %zmm25
11620 ; AVX512DQ-BW-NEXT: vpermt2q %zmm19, %zmm26, %zmm16
11621 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11622 ; AVX512DQ-BW-NEXT: vmovdqa64 128(%r10), %zmm19
11623 ; AVX512DQ-BW-NEXT: vmovdqa64 128(%rax), %zmm0
11624 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm19, %zmm28
11625 ; AVX512DQ-BW-NEXT: vpermt2q %zmm0, %zmm30, %zmm28
11626 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm19, %zmm1
11627 ; AVX512DQ-BW-NEXT: vpermt2q %zmm0, %zmm31, %zmm1
11628 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11629 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm19, %zmm13
11630 ; AVX512DQ-BW-NEXT: vpermt2q %zmm0, %zmm27, %zmm13
11631 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm19, %zmm16
11632 ; AVX512DQ-BW-NEXT: vpermt2q %zmm0, %zmm24, %zmm16
11633 ; AVX512DQ-BW-NEXT: vmovdqa64 128(%r8), %zmm20
11634 ; AVX512DQ-BW-NEXT: vmovdqa64 128(%r9), %zmm1
11635 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm20, %zmm3
11636 ; AVX512DQ-BW-NEXT: vpermt2q %zmm1, %zmm22, %zmm3
11637 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} zmm3 {%k1} = zmm19[0],zmm0[0],zmm19[2],zmm0[2],zmm19[4],zmm0[4],zmm19[6],zmm0[6]
11638 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm20, %zmm10
11639 ; AVX512DQ-BW-NEXT: vpermt2q %zmm1, %zmm21, %zmm10
11640 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} zmm10 {%k1} = zmm19[1],zmm0[1],zmm19[3],zmm0[3],zmm19[5],zmm0[5],zmm19[7],zmm0[7]
11641 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm19, %zmm2
11642 ; AVX512DQ-BW-NEXT: vpermt2q %zmm0, %zmm23, %zmm2
11643 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11644 ; AVX512DQ-BW-NEXT: vpermt2q %zmm0, %zmm26, %zmm19
11645 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm20, %zmm4
11646 ; AVX512DQ-BW-NEXT: vpermt2q %zmm1, %zmm30, %zmm4
11647 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm20, %zmm6
11648 ; AVX512DQ-BW-NEXT: vpermt2q %zmm1, %zmm31, %zmm6
11649 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm20, %zmm17
11650 ; AVX512DQ-BW-NEXT: vpermt2q %zmm1, %zmm23, %zmm17
11651 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} zmm5 = zmm20[0],zmm1[0],zmm20[2],zmm1[2],zmm20[4],zmm1[4],zmm20[6],zmm1[6]
11652 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} zmm7 = zmm20[1],zmm1[1],zmm20[3],zmm1[3],zmm20[5],zmm1[5],zmm20[7],zmm1[7]
11653 ; AVX512DQ-BW-NEXT: vpermt2q %zmm1, %zmm26, %zmm20
11654 ; AVX512DQ-BW-NEXT: vmovdqa64 192(%r10), %zmm8
11655 ; AVX512DQ-BW-NEXT: vmovdqa64 192(%rax), %zmm1
11656 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm8, %zmm12
11657 ; AVX512DQ-BW-NEXT: vpermt2q %zmm1, %zmm30, %zmm12
11658 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm8, %zmm15
11659 ; AVX512DQ-BW-NEXT: vpermt2q %zmm1, %zmm31, %zmm15
11660 ; AVX512DQ-BW-NEXT: vpermi2q %zmm1, %zmm8, %zmm27
11661 ; AVX512DQ-BW-NEXT: vpermi2q %zmm1, %zmm8, %zmm24
11662 ; AVX512DQ-BW-NEXT: vmovdqa64 192(%r8), %zmm2
11663 ; AVX512DQ-BW-NEXT: vmovdqa64 192(%r9), %zmm0
11664 ; AVX512DQ-BW-NEXT: vpermi2q %zmm0, %zmm2, %zmm22
11665 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} zmm22 {%k1} = zmm8[0],zmm1[0],zmm8[2],zmm1[2],zmm8[4],zmm1[4],zmm8[6],zmm1[6]
11666 ; AVX512DQ-BW-NEXT: vpermi2q %zmm0, %zmm2, %zmm21
11667 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} zmm21 {%k1} = zmm8[1],zmm1[1],zmm8[3],zmm1[3],zmm8[5],zmm1[5],zmm8[7],zmm1[7]
11668 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm8, %zmm9
11669 ; AVX512DQ-BW-NEXT: vpermt2q %zmm1, %zmm23, %zmm9
11670 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11671 ; AVX512DQ-BW-NEXT: vpermt2q %zmm1, %zmm26, %zmm8
11672 ; AVX512DQ-BW-NEXT: vpermi2q %zmm0, %zmm2, %zmm30
11673 ; AVX512DQ-BW-NEXT: vpermi2q %zmm0, %zmm2, %zmm31
11674 ; AVX512DQ-BW-NEXT: vpermi2q %zmm0, %zmm2, %zmm23
11675 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} zmm1 = zmm2[0],zmm0[0],zmm2[2],zmm0[2],zmm2[4],zmm0[4],zmm2[6],zmm0[6]
11676 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} zmm9 = zmm2[1],zmm0[1],zmm2[3],zmm0[3],zmm2[5],zmm0[5],zmm2[7],zmm0[7]
11677 ; AVX512DQ-BW-NEXT: vpermt2q %zmm0, %zmm26, %zmm2
11678 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
11679 ; AVX512DQ-BW-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
11680 ; AVX512DQ-BW-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
11681 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm13, %zmm5 {%k1}
11682 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm0, %zmm5, %zmm0
11683 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11684 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
11685 ; AVX512DQ-BW-NEXT: vpblendd $240, (%rsp), %ymm0, %ymm0 # 32-byte Folded Reload
11686 ; AVX512DQ-BW-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
11687 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm16, %zmm7 {%k1}
11688 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm0, %zmm7, %zmm0
11689 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11690 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
11691 ; AVX512DQ-BW-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
11692 ; AVX512DQ-BW-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
11693 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm0, %zmm3, %zmm0
11694 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11695 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
11696 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, %zmm29 {%k1}
11697 ; AVX512DQ-BW-NEXT: vmovdqa (%rsi), %xmm0
11698 ; AVX512DQ-BW-NEXT: vinserti128 $1, (%rcx), %ymm0, %ymm3
11699 ; AVX512DQ-BW-NEXT: vmovdqa (%rdi), %xmm0
11700 ; AVX512DQ-BW-NEXT: vinserti128 $1, (%rdx), %ymm0, %ymm5
11701 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm5[0],ymm3[0],ymm5[2],ymm3[2]
11702 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm0, %zmm29, %zmm0
11703 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11704 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
11705 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, %zmm18 {%k1}
11706 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} ymm3 = ymm5[1],ymm3[1],ymm5[3],ymm3[3]
11707 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm3, %zmm18, %zmm0
11708 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, (%rsp) # 64-byte Spill
11709 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
11710 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, %zmm14 {%k1}
11711 ; AVX512DQ-BW-NEXT: vmovdqa 64(%rsi), %xmm3
11712 ; AVX512DQ-BW-NEXT: vinserti128 $1, 64(%rcx), %ymm3, %ymm3
11713 ; AVX512DQ-BW-NEXT: vmovdqa 64(%rdi), %xmm5
11714 ; AVX512DQ-BW-NEXT: vinserti128 $1, 64(%rdx), %ymm5, %ymm5
11715 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm5[0],ymm3[0],ymm5[2],ymm3[2]
11716 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm7, %zmm14, %zmm29
11717 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
11718 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, %zmm11 {%k1}
11719 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} ymm3 = ymm5[1],ymm3[1],ymm5[3],ymm3[3]
11720 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm3, %zmm11, %zmm26
11721 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm28, %zmm4 {%k1}
11722 ; AVX512DQ-BW-NEXT: vmovdqa 128(%rsi), %xmm3
11723 ; AVX512DQ-BW-NEXT: vinserti128 $1, 128(%rcx), %ymm3, %ymm3
11724 ; AVX512DQ-BW-NEXT: vmovdqa 128(%rdi), %xmm5
11725 ; AVX512DQ-BW-NEXT: vinserti128 $1, 128(%rdx), %ymm5, %ymm5
11726 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm5[0],ymm3[0],ymm5[2],ymm3[2]
11727 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm7, %zmm4, %zmm28
11728 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
11729 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, %zmm6 {%k1}
11730 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} ymm3 = ymm5[1],ymm3[1],ymm5[3],ymm3[3]
11731 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm3, %zmm6, %zmm18
11732 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
11733 ; AVX512DQ-BW-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm4 # 32-byte Folded Reload
11734 ; AVX512DQ-BW-NEXT: # ymm4 = ymm0[0,1,2,3],mem[4,5,6,7]
11735 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm4, %zmm10, %zmm16
11736 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm12, %zmm30 {%k1}
11737 ; AVX512DQ-BW-NEXT: vmovdqa 192(%rsi), %xmm4
11738 ; AVX512DQ-BW-NEXT: vinserti128 $1, 192(%rcx), %ymm4, %ymm5
11739 ; AVX512DQ-BW-NEXT: vmovdqa 192(%rdi), %xmm4
11740 ; AVX512DQ-BW-NEXT: vinserti128 $1, 192(%rdx), %ymm4, %ymm6
11741 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} ymm4 = ymm6[0],ymm5[0],ymm6[2],ymm5[2]
11742 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm4, %zmm30, %zmm30
11743 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm15, %zmm31 {%k1}
11744 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} ymm5 = ymm6[1],ymm5[1],ymm6[3],ymm5[3]
11745 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm5, %zmm31, %zmm31
11746 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
11747 ; AVX512DQ-BW-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm6 # 32-byte Folded Reload
11748 ; AVX512DQ-BW-NEXT: # ymm6 = ymm0[0,1,2,3],mem[4,5,6,7]
11749 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm27, %zmm1 {%k1}
11750 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm6, %zmm1, %zmm6
11751 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
11752 ; AVX512DQ-BW-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm1 # 32-byte Folded Reload
11753 ; AVX512DQ-BW-NEXT: # ymm1 = ymm0[0,1,2,3],mem[4,5,6,7]
11754 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm24, %zmm9 {%k1}
11755 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm1, %zmm9, %zmm7
11756 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
11757 ; AVX512DQ-BW-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm1 # 32-byte Folded Reload
11758 ; AVX512DQ-BW-NEXT: # ymm1 = ymm0[0,1,2,3],mem[4,5,6,7]
11759 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm1, %zmm22, %zmm9
11760 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
11761 ; AVX512DQ-BW-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm1 # 32-byte Folded Reload
11762 ; AVX512DQ-BW-NEXT: # ymm1 = ymm0[0,1,2,3],mem[4,5,6,7]
11763 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm1, %zmm21, %zmm12
11764 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
11765 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
11766 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, %zmm3 {%k1}
11767 ; AVX512DQ-BW-NEXT: vmovdqa (%rcx), %ymm1
11768 ; AVX512DQ-BW-NEXT: vmovdqa (%rdx), %ymm15
11769 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm15[0],ymm1[0],ymm15[2],ymm1[2]
11770 ; AVX512DQ-BW-NEXT: vmovdqa64 (%rsi), %ymm21
11771 ; AVX512DQ-BW-NEXT: vmovdqa64 (%rdi), %ymm22
11772 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} ymm14 = ymm22[0],ymm21[0],ymm22[2],ymm21[2]
11773 ; AVX512DQ-BW-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm14[2,3],ymm0[2,3]
11774 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm0, %zmm3, %zmm0
11775 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
11776 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
11777 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
11778 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm15[1],ymm1[1],ymm15[3],ymm1[3]
11779 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} ymm14 = ymm22[1],ymm21[1],ymm22[3],ymm21[3]
11780 ; AVX512DQ-BW-NEXT: vperm2i128 {{.*#+}} ymm1 = ymm14[2,3],ymm1[2,3]
11781 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm1, %zmm4, %zmm1
11782 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
11783 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, %zmm25 {%k1}
11784 ; AVX512DQ-BW-NEXT: vmovdqa 64(%rcx), %ymm14
11785 ; AVX512DQ-BW-NEXT: vmovdqa 64(%rdx), %ymm15
11786 ; AVX512DQ-BW-NEXT: vmovdqa64 64(%rsi), %ymm21
11787 ; AVX512DQ-BW-NEXT: vmovdqa64 64(%rdi), %ymm22
11788 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} ymm13 = ymm15[0],ymm14[0],ymm15[2],ymm14[2]
11789 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} ymm11 = ymm22[0],ymm21[0],ymm22[2],ymm21[2]
11790 ; AVX512DQ-BW-NEXT: vperm2i128 {{.*#+}} ymm11 = ymm11[2,3],ymm13[2,3]
11791 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm11, %zmm25, %zmm11
11792 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
11793 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
11794 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
11795 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} ymm13 = ymm15[1],ymm14[1],ymm15[3],ymm14[3]
11796 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} ymm14 = ymm22[1],ymm21[1],ymm22[3],ymm21[3]
11797 ; AVX512DQ-BW-NEXT: vperm2i128 {{.*#+}} ymm13 = ymm14[2,3],ymm13[2,3]
11798 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm13, %zmm4, %zmm13
11799 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
11800 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, %zmm17 {%k1}
11801 ; AVX512DQ-BW-NEXT: vmovdqa 128(%rcx), %ymm14
11802 ; AVX512DQ-BW-NEXT: vmovdqa 128(%rdx), %ymm15
11803 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} ymm4 = ymm15[0],ymm14[0],ymm15[2],ymm14[2]
11804 ; AVX512DQ-BW-NEXT: vmovdqa64 128(%rsi), %ymm21
11805 ; AVX512DQ-BW-NEXT: vmovdqa64 128(%rdi), %ymm22
11806 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm22[0],ymm21[0],ymm22[2],ymm21[2]
11807 ; AVX512DQ-BW-NEXT: vperm2i128 {{.*#+}} ymm3 = ymm3[2,3],ymm4[2,3]
11808 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm3, %zmm17, %zmm3
11809 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm19, %zmm20 {%k1}
11810 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} ymm4 = ymm15[1],ymm14[1],ymm15[3],ymm14[3]
11811 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} ymm14 = ymm22[1],ymm21[1],ymm22[3],ymm21[3]
11812 ; AVX512DQ-BW-NEXT: vperm2i128 {{.*#+}} ymm4 = ymm14[2,3],ymm4[2,3]
11813 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm4, %zmm20, %zmm4
11814 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
11815 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm5, %zmm23 {%k1}
11816 ; AVX512DQ-BW-NEXT: vmovdqa 192(%rcx), %ymm14
11817 ; AVX512DQ-BW-NEXT: vmovdqa 192(%rdx), %ymm15
11818 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} ymm5 = ymm15[0],ymm14[0],ymm15[2],ymm14[2]
11819 ; AVX512DQ-BW-NEXT: vmovdqa64 192(%rsi), %ymm17
11820 ; AVX512DQ-BW-NEXT: vmovdqa64 192(%rdi), %ymm19
11821 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} ymm10 = ymm19[0],ymm17[0],ymm19[2],ymm17[2]
11822 ; AVX512DQ-BW-NEXT: vperm2i128 {{.*#+}} ymm5 = ymm10[2,3],ymm5[2,3]
11823 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm5, %zmm23, %zmm5
11824 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm8, %zmm2 {%k1}
11825 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} ymm8 = ymm15[1],ymm14[1],ymm15[3],ymm14[3]
11826 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} ymm10 = ymm19[1],ymm17[1],ymm19[3],ymm17[3]
11827 ; AVX512DQ-BW-NEXT: vperm2i128 {{.*#+}} ymm8 = ymm10[2,3],ymm8[2,3]
11828 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm8, %zmm2, %zmm2
11829 ; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
11830 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, 1728(%rax)
11831 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm5, 1664(%rax)
11832 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm4, 1216(%rax)
11833 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, 1152(%rax)
11834 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm13, 704(%rax)
11835 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm11, 640(%rax)
11836 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm1, 192(%rax)
11837 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, 128(%rax)
11838 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm12, 1984(%rax)
11839 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm9, 1920(%rax)
11840 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm7, 1856(%rax)
11841 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm6, 1792(%rax)
11842 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm31, 1600(%rax)
11843 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm30, 1536(%rax)
11844 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm16, 1472(%rax)
11845 ; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
11846 ; AVX512DQ-BW-NEXT: vmovaps %zmm0, 1408(%rax)
11847 ; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
11848 ; AVX512DQ-BW-NEXT: vmovaps %zmm0, 1344(%rax)
11849 ; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
11850 ; AVX512DQ-BW-NEXT: vmovaps %zmm0, 1280(%rax)
11851 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm18, 1088(%rax)
11852 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm28, 1024(%rax)
11853 ; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
11854 ; AVX512DQ-BW-NEXT: vmovaps %zmm0, 960(%rax)
11855 ; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
11856 ; AVX512DQ-BW-NEXT: vmovaps %zmm0, 896(%rax)
11857 ; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
11858 ; AVX512DQ-BW-NEXT: vmovaps %zmm0, 832(%rax)
11859 ; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
11860 ; AVX512DQ-BW-NEXT: vmovaps %zmm0, 768(%rax)
11861 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm26, 576(%rax)
11862 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm29, 512(%rax)
11863 ; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
11864 ; AVX512DQ-BW-NEXT: vmovaps %zmm0, 448(%rax)
11865 ; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
11866 ; AVX512DQ-BW-NEXT: vmovaps %zmm0, 384(%rax)
11867 ; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
11868 ; AVX512DQ-BW-NEXT: vmovaps %zmm0, 320(%rax)
11869 ; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
11870 ; AVX512DQ-BW-NEXT: vmovaps %zmm0, 256(%rax)
11871 ; AVX512DQ-BW-NEXT: vmovups (%rsp), %zmm0 # 64-byte Reload
11872 ; AVX512DQ-BW-NEXT: vmovaps %zmm0, 64(%rax)
11873 ; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
11874 ; AVX512DQ-BW-NEXT: vmovaps %zmm0, (%rax)
11875 ; AVX512DQ-BW-NEXT: addq $2312, %rsp # imm = 0x908
11876 ; AVX512DQ-BW-NEXT: vzeroupper
11877 ; AVX512DQ-BW-NEXT: retq
11878 ;
11879 ; AVX512DQ-BW-FCP-LABEL: store_i64_stride8_vf32:
11880 ; AVX512DQ-BW-FCP: # %bb.0:
11881 ; AVX512DQ-BW-FCP-NEXT: subq $2312, %rsp # imm = 0x908
11882 ; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
11883 ; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10
11884 ; AVX512DQ-BW-FCP-NEXT: vmovaps 128(%rdi), %zmm0
11885 ; AVX512DQ-BW-FCP-NEXT: vmovups %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11886 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 64(%rdi), %zmm6
11887 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rdi), %zmm2
11888 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 128(%rsi), %zmm0
11889 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 64(%rsi), %zmm5
11890 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rsi), %zmm3
11891 ; AVX512DQ-BW-FCP-NEXT: vmovaps 192(%rdx), %zmm1
11892 ; AVX512DQ-BW-FCP-NEXT: vmovups %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11893 ; AVX512DQ-BW-FCP-NEXT: vmovaps 128(%rdx), %zmm1
11894 ; AVX512DQ-BW-FCP-NEXT: vmovups %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11895 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 64(%rdx), %zmm7
11896 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rdx), %zmm4
11897 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 192(%rcx), %zmm23
11898 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 128(%rcx), %zmm1
11899 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 64(%rcx), %zmm9
11900 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rcx), %zmm8
11901 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%r8), %zmm18
11902 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 64(%r8), %zmm16
11903 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%r9), %zmm12
11904 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 64(%r9), %zmm19
11905 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%r10), %zmm25
11906 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 64(%r10), %zmm17
11907 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rax), %zmm15
11908 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 64(%rax), %zmm20
11909 ; AVX512DQ-BW-FCP-NEXT: movb $-64, %r11b
11910 ; AVX512DQ-BW-FCP-NEXT: kmovd %r11d, %k1
11911 ; AVX512DQ-BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm27 = [4,12,4,12,4,12,4,12]
11912 ; AVX512DQ-BW-FCP-NEXT: # zmm27 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
11913 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm4, %zmm10
11914 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm8, %zmm27, %zmm10
11915 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm11
11916 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm3, %zmm27, %zmm11
11917 ; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm10 = ymm11[0,1,2,3],ymm10[4,5,6,7]
11918 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm25, %zmm11
11919 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm15, %zmm27, %zmm11
11920 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm13 = zmm18[0],zmm12[0],zmm18[2],zmm12[2],zmm18[4],zmm12[4],zmm18[6],zmm12[6]
11921 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm11, %zmm13 {%k1}
11922 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm10, %zmm13, %zmm10
11923 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11924 ; AVX512DQ-BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm24 = [5,13,5,13,5,13,5,13]
11925 ; AVX512DQ-BW-FCP-NEXT: # zmm24 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
11926 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm4, %zmm10
11927 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm8, %zmm24, %zmm10
11928 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm11
11929 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm3, %zmm24, %zmm11
11930 ; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm10 = ymm11[0,1,2,3],ymm10[4,5,6,7]
11931 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm25, %zmm11
11932 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm15, %zmm24, %zmm11
11933 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm14 = zmm18[1],zmm12[1],zmm18[3],zmm12[3],zmm18[5],zmm12[5],zmm18[7],zmm12[7]
11934 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm11, %zmm14 {%k1}
11935 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm10, %zmm14, %zmm10
11936 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11937 ; AVX512DQ-BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm22 = [6,14,6,14,6,14,6,14]
11938 ; AVX512DQ-BW-FCP-NEXT: # zmm22 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
11939 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm4, %zmm10
11940 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm8, %zmm22, %zmm10
11941 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm11
11942 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm3, %zmm22, %zmm11
11943 ; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm10 = ymm11[0,1,2,3],ymm10[4,5,6,7]
11944 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm18, %zmm11
11945 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm12, %zmm22, %zmm11
11946 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm11 {%k1} = zmm25[0],zmm15[0],zmm25[2],zmm15[2],zmm25[4],zmm15[4],zmm25[6],zmm15[6]
11947 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm10, %zmm11, %zmm10
11948 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11949 ; AVX512DQ-BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm21 = [7,15,7,15,7,15,7,15]
11950 ; AVX512DQ-BW-FCP-NEXT: # zmm21 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
11951 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm8, %zmm21, %zmm4
11952 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm3, %zmm21, %zmm2
11953 ; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm4[4,5,6,7]
11954 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm18, %zmm3
11955 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm12, %zmm21, %zmm3
11956 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm3 {%k1} = zmm25[1],zmm15[1],zmm25[3],zmm15[3],zmm25[5],zmm15[5],zmm25[7],zmm15[7]
11957 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
11958 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11959 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm7, %zmm2
11960 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm9, %zmm27, %zmm2
11961 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm3
11962 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm5, %zmm27, %zmm3
11963 ; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3],ymm2[4,5,6,7]
11964 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm17, %zmm3
11965 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm20, %zmm27, %zmm3
11966 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm4 = zmm16[0],zmm19[0],zmm16[2],zmm19[2],zmm16[4],zmm19[4],zmm16[6],zmm19[6]
11967 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
11968 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
11969 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11970 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm7, %zmm2
11971 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm9, %zmm24, %zmm2
11972 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm3
11973 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm5, %zmm24, %zmm3
11974 ; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3],ymm2[4,5,6,7]
11975 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm17, %zmm3
11976 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm20, %zmm24, %zmm3
11977 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm4 = zmm16[1],zmm19[1],zmm16[3],zmm19[3],zmm16[5],zmm19[5],zmm16[7],zmm19[7]
11978 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
11979 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm4, %zmm2
11980 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11981 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm7, %zmm2
11982 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm9, %zmm22, %zmm2
11983 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm3
11984 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm5, %zmm22, %zmm3
11985 ; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3],ymm2[4,5,6,7]
11986 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm16, %zmm3
11987 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm19, %zmm22, %zmm3
11988 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm3 {%k1} = zmm17[0],zmm20[0],zmm17[2],zmm20[2],zmm17[4],zmm20[4],zmm17[6],zmm20[6]
11989 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
11990 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11991 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm9, %zmm21, %zmm7
11992 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm5, %zmm21, %zmm6
11993 ; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm6[0,1,2,3],ymm7[4,5,6,7]
11994 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm16, %zmm3
11995 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm19, %zmm21, %zmm3
11996 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm3 {%k1} = zmm17[1],zmm20[1],zmm17[3],zmm20[3],zmm17[5],zmm20[5],zmm17[7],zmm20[7]
11997 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm3, %zmm2
11998 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
11999 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
12000 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm2
12001 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm1, %zmm27, %zmm2
12002 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
12003 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
12004 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm4
12005 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm0, %zmm27, %zmm4
12006 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
12007 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm4
12008 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm1, %zmm24, %zmm4
12009 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm4, (%rsp) # 64-byte Spill
12010 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm4
12011 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm0, %zmm24, %zmm4
12012 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
12013 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm4
12014 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm1, %zmm22, %zmm4
12015 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
12016 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm4
12017 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm0, %zmm22, %zmm4
12018 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
12019 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm1, %zmm21, %zmm3
12020 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
12021 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm0, %zmm21, %zmm2
12022 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
12023 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
12024 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm1
12025 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm23, %zmm27, %zmm1
12026 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
12027 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm1
12028 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm23, %zmm24, %zmm1
12029 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
12030 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm1
12031 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm23, %zmm22, %zmm1
12032 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
12033 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm23, %zmm21, %zmm0
12034 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
12035 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 192(%rdi), %zmm1
12036 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 192(%rsi), %zmm0
12037 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm2
12038 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm0, %zmm27, %zmm2
12039 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
12040 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm2
12041 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm0, %zmm24, %zmm2
12042 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
12043 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm2
12044 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm0, %zmm22, %zmm2
12045 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
12046 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm0, %zmm21, %zmm1
12047 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
12048 ; AVX512DQ-BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm30 = [0,8,0,8,0,8,0,8]
12049 ; AVX512DQ-BW-FCP-NEXT: # zmm30 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
12050 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm25, %zmm0
12051 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm15, %zmm30, %zmm0
12052 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
12053 ; AVX512DQ-BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm31 = [1,9,1,9,1,9,1,9]
12054 ; AVX512DQ-BW-FCP-NEXT: # zmm31 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
12055 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm25, %zmm0
12056 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm15, %zmm31, %zmm0
12057 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
12058 ; AVX512DQ-BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm23 = [2,10,2,10,2,10,2,10]
12059 ; AVX512DQ-BW-FCP-NEXT: # zmm23 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
12060 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm25, %zmm0
12061 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm15, %zmm23, %zmm0
12062 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
12063 ; AVX512DQ-BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm26 = [3,11,3,11,3,11,3,11]
12064 ; AVX512DQ-BW-FCP-NEXT: # zmm26 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
12065 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm15, %zmm26, %zmm25
12066 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm25, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
12067 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm18, %zmm0
12068 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm18, %zmm29
12069 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm12, %zmm30, %zmm29
12070 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm12, %zmm31, %zmm18
12071 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm1
12072 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm12, %zmm23, %zmm1
12073 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
12074 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm12, %zmm26, %zmm0
12075 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
12076 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm17, %zmm0
12077 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm20, %zmm30, %zmm0
12078 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
12079 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm17, %zmm0
12080 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm20, %zmm31, %zmm0
12081 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
12082 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm17, %zmm0
12083 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm20, %zmm23, %zmm0
12084 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
12085 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm20, %zmm26, %zmm17
12086 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm17, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
12087 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm16, %zmm14
12088 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm19, %zmm30, %zmm14
12089 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm16, %zmm11
12090 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm19, %zmm31, %zmm11
12091 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm16, %zmm25
12092 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm19, %zmm23, %zmm25
12093 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm19, %zmm26, %zmm16
12094 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
12095 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 128(%r10), %zmm19
12096 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 128(%rax), %zmm0
12097 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm19, %zmm28
12098 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm0, %zmm30, %zmm28
12099 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm19, %zmm1
12100 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm0, %zmm31, %zmm1
12101 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
12102 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm19, %zmm13
12103 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm0, %zmm27, %zmm13
12104 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm19, %zmm16
12105 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm0, %zmm24, %zmm16
12106 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 128(%r8), %zmm20
12107 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 128(%r9), %zmm1
12108 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm20, %zmm3
12109 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm1, %zmm22, %zmm3
12110 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm3 {%k1} = zmm19[0],zmm0[0],zmm19[2],zmm0[2],zmm19[4],zmm0[4],zmm19[6],zmm0[6]
12111 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm20, %zmm10
12112 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm1, %zmm21, %zmm10
12113 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm10 {%k1} = zmm19[1],zmm0[1],zmm19[3],zmm0[3],zmm19[5],zmm0[5],zmm19[7],zmm0[7]
12114 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm19, %zmm2
12115 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm0, %zmm23, %zmm2
12116 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
12117 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm0, %zmm26, %zmm19
12118 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm20, %zmm4
12119 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm1, %zmm30, %zmm4
12120 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm20, %zmm6
12121 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm1, %zmm31, %zmm6
12122 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm20, %zmm17
12123 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm1, %zmm23, %zmm17
12124 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm5 = zmm20[0],zmm1[0],zmm20[2],zmm1[2],zmm20[4],zmm1[4],zmm20[6],zmm1[6]
12125 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm7 = zmm20[1],zmm1[1],zmm20[3],zmm1[3],zmm20[5],zmm1[5],zmm20[7],zmm1[7]
12126 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm1, %zmm26, %zmm20
12127 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 192(%r10), %zmm8
12128 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 192(%rax), %zmm1
12129 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm8, %zmm12
12130 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm1, %zmm30, %zmm12
12131 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm8, %zmm15
12132 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm1, %zmm31, %zmm15
12133 ; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm1, %zmm8, %zmm27
12134 ; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm1, %zmm8, %zmm24
12135 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 192(%r8), %zmm2
12136 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 192(%r9), %zmm0
12137 ; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm0, %zmm2, %zmm22
12138 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm22 {%k1} = zmm8[0],zmm1[0],zmm8[2],zmm1[2],zmm8[4],zmm1[4],zmm8[6],zmm1[6]
12139 ; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm0, %zmm2, %zmm21
12140 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm21 {%k1} = zmm8[1],zmm1[1],zmm8[3],zmm1[3],zmm8[5],zmm1[5],zmm8[7],zmm1[7]
12141 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm8, %zmm9
12142 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm1, %zmm23, %zmm9
12143 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
12144 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm1, %zmm26, %zmm8
12145 ; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm0, %zmm2, %zmm30
12146 ; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm0, %zmm2, %zmm31
12147 ; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm0, %zmm2, %zmm23
12148 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm1 = zmm2[0],zmm0[0],zmm2[2],zmm0[2],zmm2[4],zmm0[4],zmm2[6],zmm0[6]
12149 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm9 = zmm2[1],zmm0[1],zmm2[3],zmm0[3],zmm2[5],zmm0[5],zmm2[7],zmm0[7]
12150 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm0, %zmm26, %zmm2
12151 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
12152 ; AVX512DQ-BW-FCP-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
12153 ; AVX512DQ-BW-FCP-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
12154 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm13, %zmm5 {%k1}
12155 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm5, %zmm0
12156 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
12157 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
12158 ; AVX512DQ-BW-FCP-NEXT: vpblendd $240, (%rsp), %ymm0, %ymm0 # 32-byte Folded Reload
12159 ; AVX512DQ-BW-FCP-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
12160 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm16, %zmm7 {%k1}
12161 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm7, %zmm0
12162 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
12163 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
12164 ; AVX512DQ-BW-FCP-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
12165 ; AVX512DQ-BW-FCP-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
12166 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm3, %zmm0
12167 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
12168 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
12169 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm29 {%k1}
12170 ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rsi), %xmm0
12171 ; AVX512DQ-BW-FCP-NEXT: vinserti128 $1, (%rcx), %ymm0, %ymm3
12172 ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdi), %xmm0
12173 ; AVX512DQ-BW-FCP-NEXT: vinserti128 $1, (%rdx), %ymm0, %ymm5
12174 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm5[0],ymm3[0],ymm5[2],ymm3[2]
12175 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm29, %zmm0
12176 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
12177 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
12178 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm18 {%k1}
12179 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm3 = ymm5[1],ymm3[1],ymm5[3],ymm3[3]
12180 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm3, %zmm18, %zmm0
12181 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, (%rsp) # 64-byte Spill
12182 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
12183 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm14 {%k1}
12184 ; AVX512DQ-BW-FCP-NEXT: vmovdqa 64(%rsi), %xmm3
12185 ; AVX512DQ-BW-FCP-NEXT: vinserti128 $1, 64(%rcx), %ymm3, %ymm3
12186 ; AVX512DQ-BW-FCP-NEXT: vmovdqa 64(%rdi), %xmm5
12187 ; AVX512DQ-BW-FCP-NEXT: vinserti128 $1, 64(%rdx), %ymm5, %ymm5
12188 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm5[0],ymm3[0],ymm5[2],ymm3[2]
12189 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm7, %zmm14, %zmm29
12190 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
12191 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm11 {%k1}
12192 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm3 = ymm5[1],ymm3[1],ymm5[3],ymm3[3]
12193 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm3, %zmm11, %zmm26
12194 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm28, %zmm4 {%k1}
12195 ; AVX512DQ-BW-FCP-NEXT: vmovdqa 128(%rsi), %xmm3
12196 ; AVX512DQ-BW-FCP-NEXT: vinserti128 $1, 128(%rcx), %ymm3, %ymm3
12197 ; AVX512DQ-BW-FCP-NEXT: vmovdqa 128(%rdi), %xmm5
12198 ; AVX512DQ-BW-FCP-NEXT: vinserti128 $1, 128(%rdx), %ymm5, %ymm5
12199 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm5[0],ymm3[0],ymm5[2],ymm3[2]
12200 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm7, %zmm4, %zmm28
12201 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
12202 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm6 {%k1}
12203 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm3 = ymm5[1],ymm3[1],ymm5[3],ymm3[3]
12204 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm3, %zmm6, %zmm18
12205 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
12206 ; AVX512DQ-BW-FCP-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm4 # 32-byte Folded Reload
12207 ; AVX512DQ-BW-FCP-NEXT: # ymm4 = ymm0[0,1,2,3],mem[4,5,6,7]
12208 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm4, %zmm10, %zmm16
12209 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm12, %zmm30 {%k1}
12210 ; AVX512DQ-BW-FCP-NEXT: vmovdqa 192(%rsi), %xmm4
12211 ; AVX512DQ-BW-FCP-NEXT: vinserti128 $1, 192(%rcx), %ymm4, %ymm5
12212 ; AVX512DQ-BW-FCP-NEXT: vmovdqa 192(%rdi), %xmm4
12213 ; AVX512DQ-BW-FCP-NEXT: vinserti128 $1, 192(%rdx), %ymm4, %ymm6
12214 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm4 = ymm6[0],ymm5[0],ymm6[2],ymm5[2]
12215 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm4, %zmm30, %zmm30
12216 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm15, %zmm31 {%k1}
12217 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm5 = ymm6[1],ymm5[1],ymm6[3],ymm5[3]
12218 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm5, %zmm31, %zmm31
12219 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
12220 ; AVX512DQ-BW-FCP-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm6 # 32-byte Folded Reload
12221 ; AVX512DQ-BW-FCP-NEXT: # ymm6 = ymm0[0,1,2,3],mem[4,5,6,7]
12222 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm27, %zmm1 {%k1}
12223 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm1, %zmm6
12224 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
12225 ; AVX512DQ-BW-FCP-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm1 # 32-byte Folded Reload
12226 ; AVX512DQ-BW-FCP-NEXT: # ymm1 = ymm0[0,1,2,3],mem[4,5,6,7]
12227 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm24, %zmm9 {%k1}
12228 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm1, %zmm9, %zmm7
12229 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
12230 ; AVX512DQ-BW-FCP-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm1 # 32-byte Folded Reload
12231 ; AVX512DQ-BW-FCP-NEXT: # ymm1 = ymm0[0,1,2,3],mem[4,5,6,7]
12232 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm1, %zmm22, %zmm9
12233 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
12234 ; AVX512DQ-BW-FCP-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm1 # 32-byte Folded Reload
12235 ; AVX512DQ-BW-FCP-NEXT: # ymm1 = ymm0[0,1,2,3],mem[4,5,6,7]
12236 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm1, %zmm21, %zmm12
12237 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
12238 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
12239 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm3 {%k1}
12240 ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rcx), %ymm1
12241 ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdx), %ymm15
12242 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm15[0],ymm1[0],ymm15[2],ymm1[2]
12243 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rsi), %ymm21
12244 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rdi), %ymm22
12245 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm14 = ymm22[0],ymm21[0],ymm22[2],ymm21[2]
12246 ; AVX512DQ-BW-FCP-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm14[2,3],ymm0[2,3]
12247 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm3, %zmm0
12248 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
12249 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
12250 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
12251 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm15[1],ymm1[1],ymm15[3],ymm1[3]
12252 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm14 = ymm22[1],ymm21[1],ymm22[3],ymm21[3]
12253 ; AVX512DQ-BW-FCP-NEXT: vperm2i128 {{.*#+}} ymm1 = ymm14[2,3],ymm1[2,3]
12254 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm1, %zmm4, %zmm1
12255 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
12256 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm25 {%k1}
12257 ; AVX512DQ-BW-FCP-NEXT: vmovdqa 64(%rcx), %ymm14
12258 ; AVX512DQ-BW-FCP-NEXT: vmovdqa 64(%rdx), %ymm15
12259 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 64(%rsi), %ymm21
12260 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 64(%rdi), %ymm22
12261 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm13 = ymm15[0],ymm14[0],ymm15[2],ymm14[2]
12262 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm11 = ymm22[0],ymm21[0],ymm22[2],ymm21[2]
12263 ; AVX512DQ-BW-FCP-NEXT: vperm2i128 {{.*#+}} ymm11 = ymm11[2,3],ymm13[2,3]
12264 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm11, %zmm25, %zmm11
12265 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
12266 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
12267 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
12268 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm13 = ymm15[1],ymm14[1],ymm15[3],ymm14[3]
12269 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm14 = ymm22[1],ymm21[1],ymm22[3],ymm21[3]
12270 ; AVX512DQ-BW-FCP-NEXT: vperm2i128 {{.*#+}} ymm13 = ymm14[2,3],ymm13[2,3]
12271 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm13, %zmm4, %zmm13
12272 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
12273 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm17 {%k1}
12274 ; AVX512DQ-BW-FCP-NEXT: vmovdqa 128(%rcx), %ymm14
12275 ; AVX512DQ-BW-FCP-NEXT: vmovdqa 128(%rdx), %ymm15
12276 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm4 = ymm15[0],ymm14[0],ymm15[2],ymm14[2]
12277 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 128(%rsi), %ymm21
12278 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 128(%rdi), %ymm22
12279 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm22[0],ymm21[0],ymm22[2],ymm21[2]
12280 ; AVX512DQ-BW-FCP-NEXT: vperm2i128 {{.*#+}} ymm3 = ymm3[2,3],ymm4[2,3]
12281 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm3, %zmm17, %zmm3
12282 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm19, %zmm20 {%k1}
12283 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm4 = ymm15[1],ymm14[1],ymm15[3],ymm14[3]
12284 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm14 = ymm22[1],ymm21[1],ymm22[3],ymm21[3]
12285 ; AVX512DQ-BW-FCP-NEXT: vperm2i128 {{.*#+}} ymm4 = ymm14[2,3],ymm4[2,3]
12286 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm4, %zmm20, %zmm4
12287 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
12288 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm5, %zmm23 {%k1}
12289 ; AVX512DQ-BW-FCP-NEXT: vmovdqa 192(%rcx), %ymm14
12290 ; AVX512DQ-BW-FCP-NEXT: vmovdqa 192(%rdx), %ymm15
12291 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm5 = ymm15[0],ymm14[0],ymm15[2],ymm14[2]
12292 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 192(%rsi), %ymm17
12293 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 192(%rdi), %ymm19
12294 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm10 = ymm19[0],ymm17[0],ymm19[2],ymm17[2]
12295 ; AVX512DQ-BW-FCP-NEXT: vperm2i128 {{.*#+}} ymm5 = ymm10[2,3],ymm5[2,3]
12296 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm5, %zmm23, %zmm5
12297 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm8, %zmm2 {%k1}
12298 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm8 = ymm15[1],ymm14[1],ymm15[3],ymm14[3]
12299 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm10 = ymm19[1],ymm17[1],ymm19[3],ymm17[3]
12300 ; AVX512DQ-BW-FCP-NEXT: vperm2i128 {{.*#+}} ymm8 = ymm10[2,3],ymm8[2,3]
12301 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm8, %zmm2, %zmm2
12302 ; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
12303 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, 1728(%rax)
12304 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm5, 1664(%rax)
12305 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm4, 1216(%rax)
12306 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, 1152(%rax)
12307 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm13, 704(%rax)
12308 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm11, 640(%rax)
12309 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm1, 192(%rax)
12310 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, 128(%rax)
12311 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm12, 1984(%rax)
12312 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm9, 1920(%rax)
12313 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm7, 1856(%rax)
12314 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm6, 1792(%rax)
12315 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm31, 1600(%rax)
12316 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm30, 1536(%rax)
12317 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm16, 1472(%rax)
12318 ; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
12319 ; AVX512DQ-BW-FCP-NEXT: vmovaps %zmm0, 1408(%rax)
12320 ; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
12321 ; AVX512DQ-BW-FCP-NEXT: vmovaps %zmm0, 1344(%rax)
12322 ; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
12323 ; AVX512DQ-BW-FCP-NEXT: vmovaps %zmm0, 1280(%rax)
12324 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm18, 1088(%rax)
12325 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm28, 1024(%rax)
12326 ; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
12327 ; AVX512DQ-BW-FCP-NEXT: vmovaps %zmm0, 960(%rax)
12328 ; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
12329 ; AVX512DQ-BW-FCP-NEXT: vmovaps %zmm0, 896(%rax)
12330 ; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
12331 ; AVX512DQ-BW-FCP-NEXT: vmovaps %zmm0, 832(%rax)
12332 ; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
12333 ; AVX512DQ-BW-FCP-NEXT: vmovaps %zmm0, 768(%rax)
12334 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm26, 576(%rax)
12335 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm29, 512(%rax)
12336 ; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
12337 ; AVX512DQ-BW-FCP-NEXT: vmovaps %zmm0, 448(%rax)
12338 ; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
12339 ; AVX512DQ-BW-FCP-NEXT: vmovaps %zmm0, 384(%rax)
12340 ; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
12341 ; AVX512DQ-BW-FCP-NEXT: vmovaps %zmm0, 320(%rax)
12342 ; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
12343 ; AVX512DQ-BW-FCP-NEXT: vmovaps %zmm0, 256(%rax)
12344 ; AVX512DQ-BW-FCP-NEXT: vmovups (%rsp), %zmm0 # 64-byte Reload
12345 ; AVX512DQ-BW-FCP-NEXT: vmovaps %zmm0, 64(%rax)
12346 ; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
12347 ; AVX512DQ-BW-FCP-NEXT: vmovaps %zmm0, (%rax)
12348 ; AVX512DQ-BW-FCP-NEXT: addq $2312, %rsp # imm = 0x908
12349 ; AVX512DQ-BW-FCP-NEXT: vzeroupper
12350 ; AVX512DQ-BW-FCP-NEXT: retq
12351 %in.vec0 = load <32 x i64>, ptr %in.vecptr0, align 64
12352 %in.vec1 = load <32 x i64>, ptr %in.vecptr1, align 64
12353 %in.vec2 = load <32 x i64>, ptr %in.vecptr2, align 64
12354 %in.vec3 = load <32 x i64>, ptr %in.vecptr3, align 64
12355 %in.vec4 = load <32 x i64>, ptr %in.vecptr4, align 64
12356 %in.vec5 = load <32 x i64>, ptr %in.vecptr5, align 64
12357 %in.vec6 = load <32 x i64>, ptr %in.vecptr6, align 64
12358 %in.vec7 = load <32 x i64>, ptr %in.vecptr7, align 64
12359 %1 = shufflevector <32 x i64> %in.vec0, <32 x i64> %in.vec1, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
12360 %2 = shufflevector <32 x i64> %in.vec2, <32 x i64> %in.vec3, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
12361 %3 = shufflevector <32 x i64> %in.vec4, <32 x i64> %in.vec5, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
12362 %4 = shufflevector <32 x i64> %in.vec6, <32 x i64> %in.vec7, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
12363 %5 = shufflevector <64 x i64> %1, <64 x i64> %2, <128 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 64, i32 65, i32 66, i32 67, i32 68, i32 69, i32 70, i32 71, i32 72, i32 73, i32 74, i32 75, i32 76, i32 77, i32 78, i32 79, i32 80, i32 81, i32 82, i32 83, i32 84, i32 85, i32 86, i32 87, i32 88, i32 89, i32 90, i32 91, i32 92, i32 93, i32 94, i32 95, i32 96, i32 97, i32 98, i32 99, i32 100, i32 101, i32 102, i32 103, i32 104, i32 105, i32 106, i32 107, i32 108, i32 109, i32 110, i32 111, i32 112, i32 113, i32 114, i32 115, i32 116, i32 117, i32 118, i32 119, i32 120, i32 121, i32 122, i32 123, i32 124, i32 125, i32 126, i32 127>
12364 %6 = shufflevector <64 x i64> %3, <64 x i64> %4, <128 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 64, i32 65, i32 66, i32 67, i32 68, i32 69, i32 70, i32 71, i32 72, i32 73, i32 74, i32 75, i32 76, i32 77, i32 78, i32 79, i32 80, i32 81, i32 82, i32 83, i32 84, i32 85, i32 86, i32 87, i32 88, i32 89, i32 90, i32 91, i32 92, i32 93, i32 94, i32 95, i32 96, i32 97, i32 98, i32 99, i32 100, i32 101, i32 102, i32 103, i32 104, i32 105, i32 106, i32 107, i32 108, i32 109, i32 110, i32 111, i32 112, i32 113, i32 114, i32 115, i32 116, i32 117, i32 118, i32 119, i32 120, i32 121, i32 122, i32 123, i32 124, i32 125, i32 126, i32 127>
12365 %7 = shufflevector <128 x i64> %5, <128 x i64> %6, <256 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 64, i32 65, i32 66, i32 67, i32 68, i32 69, i32 70, i32 71, i32 72, i32 73, i32 74, i32 75, i32 76, i32 77, i32 78, i32 79, i32 80, i32 81, i32 82, i32 83, i32 84, i32 85, i32 86, i32 87, i32 88, i32 89, i32 90, i32 91, i32 92, i32 93, i32 94, i32 95, i32 96, i32 97, i32 98, i32 99, i32 100, i32 101, i32 102, i32 103, i32 104, i32 105, i32 106, i32 107, i32 108, i32 109, i32 110, i32 111, i32 112, i32 113, i32 114, i32 115, i32 116, i32 117, i32 118, i32 119, i32 120, i32 121, i32 122, i32 123, i32 124, i32 125, i32 126, i32 127, i32 128, i32 129, i32 130, i32 131, i32 132, i32 133, i32 134, i32 135, i32 136, i32 137, i32 138, i32 139, i32 140, i32 141, i32 142, i32 143, i32 144, i32 145, i32 146, i32 147, i32 148, i32 149, i32 150, i32 151, i32 152, i32 153, i32 154, i32 155, i32 156, i32 157, i32 158, i32 159, i32 160, i32 161, i32 162, i32 163, i32 164, i32 165, i32 166, i32 167, i32 168, i32 169, i32 170, i32 171, i32 172, i32 173, i32 174, i32 175, i32 176, i32 177, i32 178, i32 179, i32 180, i32 181, i32 182, i32 183, i32 184, i32 185, i32 186, i32 187, i32 188, i32 189, i32 190, i32 191, i32 192, i32 193, i32 194, i32 195, i32 196, i32 197, i32 198, i32 199, i32 200, i32 201, i32 202, i32 203, i32 204, i32 205, i32 206, i32 207, i32 208, i32 209, i32 210, i32 211, i32 212, i32 213, i32 214, i32 215, i32 216, i32 217, i32 218, i32 219, i32 220, i32 221, i32 222, i32 223, i32 224, i32 225, i32 226, i32 227, i32 228, i32 229, i32 230, i32 231, i32 232, i32 233, i32 234, i32 235, i32 236, i32 237, i32 238, i32 239, i32 240, i32 241, i32 242, i32 243, i32 244, i32 245, i32 246, i32 247, i32 248, i32 249, i32 250, i32 251, i32 252, i32 253, i32 254, i32 255>
12366 %interleaved.vec = shufflevector <256 x i64> %7, <256 x i64> poison, <256 x i32> <i32 0, i32 32, i32 64, i32 96, i32 128, i32 160, i32 192, i32 224, i32 1, i32 33, i32 65, i32 97, i32 129, i32 161, i32 193, i32 225, i32 2, i32 34, i32 66, i32 98, i32 130, i32 162, i32 194, i32 226, i32 3, i32 35, i32 67, i32 99, i32 131, i32 163, i32 195, i32 227, i32 4, i32 36, i32 68, i32 100, i32 132, i32 164, i32 196, i32 228, i32 5, i32 37, i32 69, i32 101, i32 133, i32 165, i32 197, i32 229, i32 6, i32 38, i32 70, i32 102, i32 134, i32 166, i32 198, i32 230, i32 7, i32 39, i32 71, i32 103, i32 135, i32 167, i32 199, i32 231, i32 8, i32 40, i32 72, i32 104, i32 136, i32 168, i32 200, i32 232, i32 9, i32 41, i32 73, i32 105, i32 137, i32 169, i32 201, i32 233, i32 10, i32 42, i32 74, i32 106, i32 138, i32 170, i32 202, i32 234, i32 11, i32 43, i32 75, i32 107, i32 139, i32 171, i32 203, i32 235, i32 12, i32 44, i32 76, i32 108, i32 140, i32 172, i32 204, i32 236, i32 13, i32 45, i32 77, i32 109, i32 141, i32 173, i32 205, i32 237, i32 14, i32 46, i32 78, i32 110, i32 142, i32 174, i32 206, i32 238, i32 15, i32 47, i32 79, i32 111, i32 143, i32 175, i32 207, i32 239, i32 16, i32 48, i32 80, i32 112, i32 144, i32 176, i32 208, i32 240, i32 17, i32 49, i32 81, i32 113, i32 145, i32 177, i32 209, i32 241, i32 18, i32 50, i32 82, i32 114, i32 146, i32 178, i32 210, i32 242, i32 19, i32 51, i32 83, i32 115, i32 147, i32 179, i32 211, i32 243, i32 20, i32 52, i32 84, i32 116, i32 148, i32 180, i32 212, i32 244, i32 21, i32 53, i32 85, i32 117, i32 149, i32 181, i32 213, i32 245, i32 22, i32 54, i32 86, i32 118, i32 150, i32 182, i32 214, i32 246, i32 23, i32 55, i32 87, i32 119, i32 151, i32 183, i32 215, i32 247, i32 24, i32 56, i32 88, i32 120, i32 152, i32 184, i32 216, i32 248, i32 25, i32 57, i32 89, i32 121, i32 153, i32 185, i32 217, i32 249, i32 26, i32 58, i32 90, i32 122, i32 154, i32 186, i32 218, i32 250, i32 27, i32 59, i32 91, i32 123, i32 155, i32 187, i32 219, i32 251, i32 28, i32 60, i32 92, i32 124, i32 156, i32 188, i32 220, i32 252, i32 29, i32 61, i32 93, i32 125, i32 157, i32 189, i32 221, i32 253, i32 30, i32 62, i32 94, i32 126, i32 158, i32 190, i32 222, i32 254, i32 31, i32 63, i32 95, i32 127, i32 159, i32 191, i32 223, i32 255>
12367 store <256 x i64> %interleaved.vec, ptr %out.vec, align 64
12368 ret void
12369 }
12371 define void @store_i64_stride8_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecptr2, ptr %in.vecptr3, ptr %in.vecptr4, ptr %in.vecptr5, ptr %in.vecptr6, ptr %in.vecptr7, ptr %out.vec) nounwind {
12372 ; SSE-LABEL: store_i64_stride8_vf64:
12373 ; SSE: # %bb.0:
12374 ; SSE-NEXT: subq $3736, %rsp # imm = 0xE98
12375 ; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax
12376 ; SSE-NEXT: movq {{[0-9]+}}(%rsp), %r10
12377 ; SSE-NEXT: movaps (%rdi), %xmm7
12378 ; SSE-NEXT: movaps 16(%rdi), %xmm8
12379 ; SSE-NEXT: movaps (%rsi), %xmm1
12380 ; SSE-NEXT: movaps 16(%rsi), %xmm0
12381 ; SSE-NEXT: movaps (%rdx), %xmm9
12382 ; SSE-NEXT: movaps 16(%rdx), %xmm10
12383 ; SSE-NEXT: movaps (%rcx), %xmm3
12384 ; SSE-NEXT: movaps 16(%rcx), %xmm2
12385 ; SSE-NEXT: movaps 16(%r8), %xmm12
12386 ; SSE-NEXT: movaps (%r8), %xmm11
12387 ; SSE-NEXT: movaps 16(%r9), %xmm4
12388 ; SSE-NEXT: movaps (%r9), %xmm5
12389 ; SSE-NEXT: movaps 16(%r10), %xmm14
12390 ; SSE-NEXT: movaps (%r10), %xmm13
12391 ; SSE-NEXT: movaps (%rax), %xmm6
12392 ; SSE-NEXT: movaps %xmm7, %xmm15
12393 ; SSE-NEXT: movlhps {{.*#+}} xmm15 = xmm15[0],xmm1[0]
12394 ; SSE-NEXT: movaps %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12395 ; SSE-NEXT: unpckhpd {{.*#+}} xmm7 = xmm7[1],xmm1[1]
12396 ; SSE-NEXT: movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12397 ; SSE-NEXT: movaps %xmm9, %xmm1
12398 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm3[0]
12399 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12400 ; SSE-NEXT: unpckhpd {{.*#+}} xmm9 = xmm9[1],xmm3[1]
12401 ; SSE-NEXT: movaps %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12402 ; SSE-NEXT: movaps %xmm11, %xmm1
12403 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm5[0]
12404 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12405 ; SSE-NEXT: unpckhpd {{.*#+}} xmm11 = xmm11[1],xmm5[1]
12406 ; SSE-NEXT: movaps %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12407 ; SSE-NEXT: movaps %xmm13, %xmm1
12408 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm6[0]
12409 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12410 ; SSE-NEXT: unpckhpd {{.*#+}} xmm13 = xmm13[1],xmm6[1]
12411 ; SSE-NEXT: movaps %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12412 ; SSE-NEXT: movaps %xmm8, %xmm1
12413 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12414 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12415 ; SSE-NEXT: unpckhpd {{.*#+}} xmm8 = xmm8[1],xmm0[1]
12416 ; SSE-NEXT: movaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12417 ; SSE-NEXT: movaps %xmm10, %xmm0
12418 ; SSE-NEXT: movlhps {{.*#+}} xmm0 = xmm0[0],xmm2[0]
12419 ; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12420 ; SSE-NEXT: unpckhpd {{.*#+}} xmm10 = xmm10[1],xmm2[1]
12421 ; SSE-NEXT: movaps %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12422 ; SSE-NEXT: movaps %xmm12, %xmm0
12423 ; SSE-NEXT: movlhps {{.*#+}} xmm0 = xmm0[0],xmm4[0]
12424 ; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12425 ; SSE-NEXT: unpckhpd {{.*#+}} xmm12 = xmm12[1],xmm4[1]
12426 ; SSE-NEXT: movaps %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12427 ; SSE-NEXT: movaps 16(%rax), %xmm0
12428 ; SSE-NEXT: movaps %xmm14, %xmm1
12429 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12430 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12431 ; SSE-NEXT: unpckhpd {{.*#+}} xmm14 = xmm14[1],xmm0[1]
12432 ; SSE-NEXT: movaps %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12433 ; SSE-NEXT: movaps 32(%rdi), %xmm2
12434 ; SSE-NEXT: movaps 32(%rsi), %xmm0
12435 ; SSE-NEXT: movaps %xmm2, %xmm1
12436 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12437 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12438 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12439 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12440 ; SSE-NEXT: movaps 32(%rdx), %xmm2
12441 ; SSE-NEXT: movaps 32(%rcx), %xmm0
12442 ; SSE-NEXT: movaps %xmm2, %xmm1
12443 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12444 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12445 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12446 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12447 ; SSE-NEXT: movaps 32(%r8), %xmm2
12448 ; SSE-NEXT: movaps 32(%r9), %xmm0
12449 ; SSE-NEXT: movaps %xmm2, %xmm1
12450 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12451 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12452 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12453 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12454 ; SSE-NEXT: movaps 32(%r10), %xmm2
12455 ; SSE-NEXT: movaps 32(%rax), %xmm0
12456 ; SSE-NEXT: movaps %xmm2, %xmm1
12457 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12458 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12459 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12460 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12461 ; SSE-NEXT: movaps 48(%rdi), %xmm2
12462 ; SSE-NEXT: movaps 48(%rsi), %xmm0
12463 ; SSE-NEXT: movaps %xmm2, %xmm1
12464 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12465 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12466 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12467 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12468 ; SSE-NEXT: movaps 48(%rdx), %xmm2
12469 ; SSE-NEXT: movaps 48(%rcx), %xmm0
12470 ; SSE-NEXT: movaps %xmm2, %xmm1
12471 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12472 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12473 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12474 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12475 ; SSE-NEXT: movaps 48(%r8), %xmm2
12476 ; SSE-NEXT: movaps 48(%r9), %xmm0
12477 ; SSE-NEXT: movaps %xmm2, %xmm1
12478 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12479 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12480 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12481 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12482 ; SSE-NEXT: movaps 48(%r10), %xmm2
12483 ; SSE-NEXT: movaps 48(%rax), %xmm0
12484 ; SSE-NEXT: movaps %xmm2, %xmm1
12485 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12486 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12487 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12488 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12489 ; SSE-NEXT: movaps 64(%rdi), %xmm2
12490 ; SSE-NEXT: movaps 64(%rsi), %xmm0
12491 ; SSE-NEXT: movaps %xmm2, %xmm1
12492 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12493 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12494 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12495 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12496 ; SSE-NEXT: movaps 64(%rdx), %xmm2
12497 ; SSE-NEXT: movaps 64(%rcx), %xmm0
12498 ; SSE-NEXT: movaps %xmm2, %xmm1
12499 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12500 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12501 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12502 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12503 ; SSE-NEXT: movaps 64(%r8), %xmm2
12504 ; SSE-NEXT: movaps 64(%r9), %xmm0
12505 ; SSE-NEXT: movaps %xmm2, %xmm1
12506 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12507 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12508 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12509 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12510 ; SSE-NEXT: movaps 64(%r10), %xmm2
12511 ; SSE-NEXT: movaps 64(%rax), %xmm0
12512 ; SSE-NEXT: movaps %xmm2, %xmm1
12513 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12514 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12515 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12516 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12517 ; SSE-NEXT: movaps 80(%rdi), %xmm2
12518 ; SSE-NEXT: movaps 80(%rsi), %xmm0
12519 ; SSE-NEXT: movaps %xmm2, %xmm1
12520 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12521 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12522 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12523 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12524 ; SSE-NEXT: movaps 80(%rdx), %xmm2
12525 ; SSE-NEXT: movaps 80(%rcx), %xmm0
12526 ; SSE-NEXT: movaps %xmm2, %xmm1
12527 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12528 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12529 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12530 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12531 ; SSE-NEXT: movaps 80(%r8), %xmm2
12532 ; SSE-NEXT: movaps 80(%r9), %xmm0
12533 ; SSE-NEXT: movaps %xmm2, %xmm1
12534 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12535 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12536 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12537 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12538 ; SSE-NEXT: movaps 80(%r10), %xmm2
12539 ; SSE-NEXT: movaps 80(%rax), %xmm0
12540 ; SSE-NEXT: movaps %xmm2, %xmm1
12541 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12542 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12543 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12544 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12545 ; SSE-NEXT: movaps 96(%rdi), %xmm2
12546 ; SSE-NEXT: movaps 96(%rsi), %xmm0
12547 ; SSE-NEXT: movaps %xmm2, %xmm1
12548 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12549 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12550 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12551 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12552 ; SSE-NEXT: movaps 96(%rdx), %xmm2
12553 ; SSE-NEXT: movaps 96(%rcx), %xmm0
12554 ; SSE-NEXT: movaps %xmm2, %xmm1
12555 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12556 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12557 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12558 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12559 ; SSE-NEXT: movaps 96(%r8), %xmm2
12560 ; SSE-NEXT: movaps 96(%r9), %xmm0
12561 ; SSE-NEXT: movaps %xmm2, %xmm1
12562 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12563 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12564 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12565 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12566 ; SSE-NEXT: movaps 96(%r10), %xmm2
12567 ; SSE-NEXT: movaps 96(%rax), %xmm0
12568 ; SSE-NEXT: movaps %xmm2, %xmm1
12569 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12570 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12571 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12572 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12573 ; SSE-NEXT: movaps 112(%rdi), %xmm2
12574 ; SSE-NEXT: movaps 112(%rsi), %xmm0
12575 ; SSE-NEXT: movaps %xmm2, %xmm1
12576 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12577 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12578 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12579 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12580 ; SSE-NEXT: movaps 112(%rdx), %xmm2
12581 ; SSE-NEXT: movaps 112(%rcx), %xmm0
12582 ; SSE-NEXT: movaps %xmm2, %xmm1
12583 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12584 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12585 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12586 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12587 ; SSE-NEXT: movaps 112(%r8), %xmm2
12588 ; SSE-NEXT: movaps 112(%r9), %xmm0
12589 ; SSE-NEXT: movaps %xmm2, %xmm1
12590 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12591 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12592 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12593 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12594 ; SSE-NEXT: movaps 112(%r10), %xmm2
12595 ; SSE-NEXT: movaps 112(%rax), %xmm0
12596 ; SSE-NEXT: movaps %xmm2, %xmm1
12597 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12598 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12599 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12600 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12601 ; SSE-NEXT: movaps 128(%rdi), %xmm2
12602 ; SSE-NEXT: movaps 128(%rsi), %xmm0
12603 ; SSE-NEXT: movaps %xmm2, %xmm1
12604 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12605 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12606 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12607 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12608 ; SSE-NEXT: movaps 128(%rdx), %xmm2
12609 ; SSE-NEXT: movaps 128(%rcx), %xmm0
12610 ; SSE-NEXT: movaps %xmm2, %xmm1
12611 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12612 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12613 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12614 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12615 ; SSE-NEXT: movaps 128(%r8), %xmm2
12616 ; SSE-NEXT: movaps 128(%r9), %xmm0
12617 ; SSE-NEXT: movaps %xmm2, %xmm1
12618 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12619 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12620 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12621 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12622 ; SSE-NEXT: movaps 128(%r10), %xmm2
12623 ; SSE-NEXT: movaps 128(%rax), %xmm0
12624 ; SSE-NEXT: movaps %xmm2, %xmm1
12625 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12626 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12627 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12628 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12629 ; SSE-NEXT: movaps 144(%rdi), %xmm2
12630 ; SSE-NEXT: movaps 144(%rsi), %xmm0
12631 ; SSE-NEXT: movaps %xmm2, %xmm1
12632 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12633 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12634 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12635 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12636 ; SSE-NEXT: movaps 144(%rdx), %xmm2
12637 ; SSE-NEXT: movaps 144(%rcx), %xmm0
12638 ; SSE-NEXT: movaps %xmm2, %xmm1
12639 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12640 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12641 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12642 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12643 ; SSE-NEXT: movaps 144(%r8), %xmm2
12644 ; SSE-NEXT: movaps 144(%r9), %xmm0
12645 ; SSE-NEXT: movaps %xmm2, %xmm1
12646 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12647 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12648 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12649 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12650 ; SSE-NEXT: movaps 144(%r10), %xmm2
12651 ; SSE-NEXT: movaps 144(%rax), %xmm0
12652 ; SSE-NEXT: movaps %xmm2, %xmm1
12653 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12654 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12655 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12656 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12657 ; SSE-NEXT: movaps 160(%rdi), %xmm2
12658 ; SSE-NEXT: movaps 160(%rsi), %xmm0
12659 ; SSE-NEXT: movaps %xmm2, %xmm1
12660 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12661 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12662 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12663 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12664 ; SSE-NEXT: movaps 160(%rdx), %xmm2
12665 ; SSE-NEXT: movaps 160(%rcx), %xmm0
12666 ; SSE-NEXT: movaps %xmm2, %xmm1
12667 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12668 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12669 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12670 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12671 ; SSE-NEXT: movaps 160(%r8), %xmm2
12672 ; SSE-NEXT: movaps 160(%r9), %xmm0
12673 ; SSE-NEXT: movaps %xmm2, %xmm1
12674 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12675 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12676 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12677 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12678 ; SSE-NEXT: movaps 160(%r10), %xmm2
12679 ; SSE-NEXT: movaps 160(%rax), %xmm0
12680 ; SSE-NEXT: movaps %xmm2, %xmm1
12681 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12682 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12683 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12684 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12685 ; SSE-NEXT: movaps 176(%rdi), %xmm2
12686 ; SSE-NEXT: movaps 176(%rsi), %xmm0
12687 ; SSE-NEXT: movaps %xmm2, %xmm1
12688 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12689 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12690 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12691 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12692 ; SSE-NEXT: movaps 176(%rdx), %xmm2
12693 ; SSE-NEXT: movaps 176(%rcx), %xmm0
12694 ; SSE-NEXT: movaps %xmm2, %xmm1
12695 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12696 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12697 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12698 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12699 ; SSE-NEXT: movaps 176(%r8), %xmm2
12700 ; SSE-NEXT: movaps 176(%r9), %xmm0
12701 ; SSE-NEXT: movaps %xmm2, %xmm1
12702 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12703 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12704 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12705 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12706 ; SSE-NEXT: movaps 176(%r10), %xmm2
12707 ; SSE-NEXT: movaps 176(%rax), %xmm0
12708 ; SSE-NEXT: movaps %xmm2, %xmm1
12709 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12710 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12711 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12712 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12713 ; SSE-NEXT: movaps 192(%rdi), %xmm2
12714 ; SSE-NEXT: movaps 192(%rsi), %xmm0
12715 ; SSE-NEXT: movaps %xmm2, %xmm1
12716 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12717 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12718 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12719 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12720 ; SSE-NEXT: movaps 192(%rdx), %xmm2
12721 ; SSE-NEXT: movaps 192(%rcx), %xmm0
12722 ; SSE-NEXT: movaps %xmm2, %xmm1
12723 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12724 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12725 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12726 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12727 ; SSE-NEXT: movaps 192(%r8), %xmm2
12728 ; SSE-NEXT: movaps 192(%r9), %xmm0
12729 ; SSE-NEXT: movaps %xmm2, %xmm1
12730 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12731 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12732 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12733 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12734 ; SSE-NEXT: movaps 192(%r10), %xmm2
12735 ; SSE-NEXT: movaps 192(%rax), %xmm0
12736 ; SSE-NEXT: movaps %xmm2, %xmm1
12737 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12738 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12739 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12740 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12741 ; SSE-NEXT: movaps 208(%rdi), %xmm2
12742 ; SSE-NEXT: movaps 208(%rsi), %xmm0
12743 ; SSE-NEXT: movaps %xmm2, %xmm1
12744 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12745 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12746 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12747 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12748 ; SSE-NEXT: movaps 208(%rdx), %xmm2
12749 ; SSE-NEXT: movaps 208(%rcx), %xmm0
12750 ; SSE-NEXT: movaps %xmm2, %xmm1
12751 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12752 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12753 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12754 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12755 ; SSE-NEXT: movaps 208(%r8), %xmm2
12756 ; SSE-NEXT: movaps 208(%r9), %xmm0
12757 ; SSE-NEXT: movaps %xmm2, %xmm1
12758 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12759 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12760 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12761 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12762 ; SSE-NEXT: movaps 208(%r10), %xmm2
12763 ; SSE-NEXT: movaps 208(%rax), %xmm0
12764 ; SSE-NEXT: movaps %xmm2, %xmm1
12765 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12766 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12767 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12768 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12769 ; SSE-NEXT: movaps 224(%rdi), %xmm2
12770 ; SSE-NEXT: movaps 224(%rsi), %xmm0
12771 ; SSE-NEXT: movaps %xmm2, %xmm1
12772 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12773 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12774 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12775 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12776 ; SSE-NEXT: movaps 224(%rdx), %xmm2
12777 ; SSE-NEXT: movaps 224(%rcx), %xmm0
12778 ; SSE-NEXT: movaps %xmm2, %xmm1
12779 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12780 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12781 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12782 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12783 ; SSE-NEXT: movaps 224(%r8), %xmm2
12784 ; SSE-NEXT: movaps 224(%r9), %xmm0
12785 ; SSE-NEXT: movaps %xmm2, %xmm1
12786 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12787 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12788 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12789 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12790 ; SSE-NEXT: movaps 224(%r10), %xmm2
12791 ; SSE-NEXT: movaps 224(%rax), %xmm0
12792 ; SSE-NEXT: movaps %xmm2, %xmm1
12793 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12794 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12795 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12796 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12797 ; SSE-NEXT: movaps 240(%rdi), %xmm2
12798 ; SSE-NEXT: movaps 240(%rsi), %xmm0
12799 ; SSE-NEXT: movaps %xmm2, %xmm1
12800 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12801 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12802 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12803 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12804 ; SSE-NEXT: movaps 240(%rdx), %xmm2
12805 ; SSE-NEXT: movaps 240(%rcx), %xmm0
12806 ; SSE-NEXT: movaps %xmm2, %xmm1
12807 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12808 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12809 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12810 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12811 ; SSE-NEXT: movaps 240(%r8), %xmm2
12812 ; SSE-NEXT: movaps 240(%r9), %xmm0
12813 ; SSE-NEXT: movaps %xmm2, %xmm1
12814 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12815 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12816 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12817 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12818 ; SSE-NEXT: movaps 240(%r10), %xmm2
12819 ; SSE-NEXT: movaps 240(%rax), %xmm0
12820 ; SSE-NEXT: movaps %xmm2, %xmm1
12821 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12822 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12823 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12824 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12825 ; SSE-NEXT: movaps 256(%rdi), %xmm2
12826 ; SSE-NEXT: movaps 256(%rsi), %xmm0
12827 ; SSE-NEXT: movaps %xmm2, %xmm1
12828 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12829 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12830 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12831 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12832 ; SSE-NEXT: movaps 256(%rdx), %xmm2
12833 ; SSE-NEXT: movaps 256(%rcx), %xmm0
12834 ; SSE-NEXT: movaps %xmm2, %xmm1
12835 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12836 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12837 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12838 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12839 ; SSE-NEXT: movaps 256(%r8), %xmm2
12840 ; SSE-NEXT: movaps 256(%r9), %xmm0
12841 ; SSE-NEXT: movaps %xmm2, %xmm1
12842 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12843 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12844 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12845 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12846 ; SSE-NEXT: movaps 256(%r10), %xmm2
12847 ; SSE-NEXT: movaps 256(%rax), %xmm0
12848 ; SSE-NEXT: movaps %xmm2, %xmm1
12849 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12850 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12851 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12852 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12853 ; SSE-NEXT: movaps 272(%rdi), %xmm2
12854 ; SSE-NEXT: movaps 272(%rsi), %xmm0
12855 ; SSE-NEXT: movaps %xmm2, %xmm1
12856 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12857 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12858 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12859 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12860 ; SSE-NEXT: movaps 272(%rdx), %xmm2
12861 ; SSE-NEXT: movaps 272(%rcx), %xmm0
12862 ; SSE-NEXT: movaps %xmm2, %xmm1
12863 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12864 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12865 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12866 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12867 ; SSE-NEXT: movaps 272(%r8), %xmm2
12868 ; SSE-NEXT: movaps 272(%r9), %xmm0
12869 ; SSE-NEXT: movaps %xmm2, %xmm1
12870 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12871 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12872 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12873 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12874 ; SSE-NEXT: movaps 272(%r10), %xmm2
12875 ; SSE-NEXT: movaps 272(%rax), %xmm0
12876 ; SSE-NEXT: movaps %xmm2, %xmm1
12877 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12878 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12879 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12880 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12881 ; SSE-NEXT: movaps 288(%rdi), %xmm2
12882 ; SSE-NEXT: movaps 288(%rsi), %xmm0
12883 ; SSE-NEXT: movaps %xmm2, %xmm1
12884 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12885 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12886 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12887 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12888 ; SSE-NEXT: movaps 288(%rdx), %xmm2
12889 ; SSE-NEXT: movaps 288(%rcx), %xmm0
12890 ; SSE-NEXT: movaps %xmm2, %xmm1
12891 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12892 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12893 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12894 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12895 ; SSE-NEXT: movaps 288(%r8), %xmm2
12896 ; SSE-NEXT: movaps 288(%r9), %xmm0
12897 ; SSE-NEXT: movaps %xmm2, %xmm1
12898 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12899 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12900 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12901 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12902 ; SSE-NEXT: movaps 288(%r10), %xmm2
12903 ; SSE-NEXT: movaps 288(%rax), %xmm0
12904 ; SSE-NEXT: movaps %xmm2, %xmm1
12905 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12906 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12907 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12908 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12909 ; SSE-NEXT: movaps 304(%rdi), %xmm2
12910 ; SSE-NEXT: movaps 304(%rsi), %xmm0
12911 ; SSE-NEXT: movaps %xmm2, %xmm1
12912 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12913 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12914 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12915 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12916 ; SSE-NEXT: movaps 304(%rdx), %xmm2
12917 ; SSE-NEXT: movaps 304(%rcx), %xmm0
12918 ; SSE-NEXT: movaps %xmm2, %xmm1
12919 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12920 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12921 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12922 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12923 ; SSE-NEXT: movaps 304(%r8), %xmm2
12924 ; SSE-NEXT: movaps 304(%r9), %xmm0
12925 ; SSE-NEXT: movaps %xmm2, %xmm1
12926 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12927 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12928 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12929 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12930 ; SSE-NEXT: movaps 304(%r10), %xmm2
12931 ; SSE-NEXT: movaps 304(%rax), %xmm0
12932 ; SSE-NEXT: movaps %xmm2, %xmm1
12933 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12934 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12935 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12936 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12937 ; SSE-NEXT: movaps 320(%rdi), %xmm2
12938 ; SSE-NEXT: movaps 320(%rsi), %xmm0
12939 ; SSE-NEXT: movaps %xmm2, %xmm1
12940 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12941 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12942 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12943 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12944 ; SSE-NEXT: movaps 320(%rdx), %xmm2
12945 ; SSE-NEXT: movaps 320(%rcx), %xmm0
12946 ; SSE-NEXT: movaps %xmm2, %xmm1
12947 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12948 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12949 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12950 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12951 ; SSE-NEXT: movaps 320(%r8), %xmm2
12952 ; SSE-NEXT: movaps 320(%r9), %xmm0
12953 ; SSE-NEXT: movaps %xmm2, %xmm1
12954 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12955 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12956 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12957 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12958 ; SSE-NEXT: movaps 320(%r10), %xmm2
12959 ; SSE-NEXT: movaps 320(%rax), %xmm0
12960 ; SSE-NEXT: movaps %xmm2, %xmm1
12961 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12962 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12963 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12964 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12965 ; SSE-NEXT: movaps 336(%rdi), %xmm2
12966 ; SSE-NEXT: movaps 336(%rsi), %xmm0
12967 ; SSE-NEXT: movaps %xmm2, %xmm1
12968 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12969 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12970 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12971 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12972 ; SSE-NEXT: movaps 336(%rdx), %xmm2
12973 ; SSE-NEXT: movaps 336(%rcx), %xmm0
12974 ; SSE-NEXT: movaps %xmm2, %xmm1
12975 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12976 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12977 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12978 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12979 ; SSE-NEXT: movaps 336(%r8), %xmm2
12980 ; SSE-NEXT: movaps 336(%r9), %xmm0
12981 ; SSE-NEXT: movaps %xmm2, %xmm1
12982 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12983 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12984 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12985 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12986 ; SSE-NEXT: movaps 336(%r10), %xmm2
12987 ; SSE-NEXT: movaps 336(%rax), %xmm0
12988 ; SSE-NEXT: movaps %xmm2, %xmm1
12989 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12990 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12991 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12992 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12993 ; SSE-NEXT: movaps 352(%rdi), %xmm2
12994 ; SSE-NEXT: movaps 352(%rsi), %xmm0
12995 ; SSE-NEXT: movaps %xmm2, %xmm1
12996 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
12997 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
12998 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
12999 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
13000 ; SSE-NEXT: movaps 352(%rdx), %xmm2
13001 ; SSE-NEXT: movaps 352(%rcx), %xmm0
13002 ; SSE-NEXT: movaps %xmm2, %xmm1
13003 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
13004 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
13005 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
13006 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
13007 ; SSE-NEXT: movaps 352(%r8), %xmm2
13008 ; SSE-NEXT: movaps 352(%r9), %xmm0
13009 ; SSE-NEXT: movaps %xmm2, %xmm1
13010 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
13011 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
13012 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
13013 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
13014 ; SSE-NEXT: movaps 352(%r10), %xmm2
13015 ; SSE-NEXT: movaps 352(%rax), %xmm0
13016 ; SSE-NEXT: movaps %xmm2, %xmm1
13017 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
13018 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
13019 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
13020 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
13021 ; SSE-NEXT: movaps 368(%rdi), %xmm2
13022 ; SSE-NEXT: movaps 368(%rsi), %xmm0
13023 ; SSE-NEXT: movaps %xmm2, %xmm1
13024 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
13025 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
13026 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
13027 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
13028 ; SSE-NEXT: movaps 368(%rdx), %xmm2
13029 ; SSE-NEXT: movaps 368(%rcx), %xmm0
13030 ; SSE-NEXT: movaps %xmm2, %xmm1
13031 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
13032 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
13033 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
13034 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
13035 ; SSE-NEXT: movaps 368(%r8), %xmm2
13036 ; SSE-NEXT: movaps 368(%r9), %xmm0
13037 ; SSE-NEXT: movaps %xmm2, %xmm1
13038 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
13039 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
13040 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
13041 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
13042 ; SSE-NEXT: movaps 368(%r10), %xmm2
13043 ; SSE-NEXT: movaps 368(%rax), %xmm0
13044 ; SSE-NEXT: movaps %xmm2, %xmm1
13045 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
13046 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
13047 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
13048 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
13049 ; SSE-NEXT: movaps 384(%rdi), %xmm2
13050 ; SSE-NEXT: movaps 384(%rsi), %xmm0
13051 ; SSE-NEXT: movaps %xmm2, %xmm1
13052 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
13053 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
13054 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
13055 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
13056 ; SSE-NEXT: movaps 384(%rdx), %xmm2
13057 ; SSE-NEXT: movaps 384(%rcx), %xmm0
13058 ; SSE-NEXT: movaps %xmm2, %xmm1
13059 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
13060 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
13061 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
13062 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
13063 ; SSE-NEXT: movaps 384(%r8), %xmm2
13064 ; SSE-NEXT: movaps 384(%r9), %xmm0
13065 ; SSE-NEXT: movaps %xmm2, %xmm1
13066 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
13067 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
13068 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
13069 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
13070 ; SSE-NEXT: movaps 384(%r10), %xmm2
13071 ; SSE-NEXT: movaps 384(%rax), %xmm0
13072 ; SSE-NEXT: movaps %xmm2, %xmm1
13073 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
13074 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
13075 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
13076 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
13077 ; SSE-NEXT: movaps 400(%rdi), %xmm2
13078 ; SSE-NEXT: movaps 400(%rsi), %xmm0
13079 ; SSE-NEXT: movaps %xmm2, %xmm1
13080 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
13081 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
13082 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
13083 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
13084 ; SSE-NEXT: movaps 400(%rdx), %xmm2
13085 ; SSE-NEXT: movaps 400(%rcx), %xmm0
13086 ; SSE-NEXT: movaps %xmm2, %xmm1
13087 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
13088 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
13089 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
13090 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
13091 ; SSE-NEXT: movaps 400(%r8), %xmm2
13092 ; SSE-NEXT: movaps 400(%r9), %xmm0
13093 ; SSE-NEXT: movaps %xmm2, %xmm1
13094 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
13095 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
13096 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
13097 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
13098 ; SSE-NEXT: movaps 400(%r10), %xmm2
13099 ; SSE-NEXT: movaps 400(%rax), %xmm0
13100 ; SSE-NEXT: movaps %xmm2, %xmm1
13101 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
13102 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
13103 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
13104 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
13105 ; SSE-NEXT: movaps 416(%rdi), %xmm2
13106 ; SSE-NEXT: movaps 416(%rsi), %xmm0
13107 ; SSE-NEXT: movaps %xmm2, %xmm1
13108 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
13109 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
13110 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
13111 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
13112 ; SSE-NEXT: movaps 416(%rdx), %xmm2
13113 ; SSE-NEXT: movaps 416(%rcx), %xmm0
13114 ; SSE-NEXT: movaps %xmm2, %xmm1
13115 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
13116 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
13117 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
13118 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
13119 ; SSE-NEXT: movaps 416(%r8), %xmm2
13120 ; SSE-NEXT: movaps 416(%r9), %xmm0
13121 ; SSE-NEXT: movaps %xmm2, %xmm1
13122 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
13123 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
13124 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
13125 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
13126 ; SSE-NEXT: movaps 416(%r10), %xmm2
13127 ; SSE-NEXT: movaps 416(%rax), %xmm0
13128 ; SSE-NEXT: movaps %xmm2, %xmm1
13129 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
13130 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
13131 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
13132 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
13133 ; SSE-NEXT: movaps 432(%rdi), %xmm2
13134 ; SSE-NEXT: movaps 432(%rsi), %xmm0
13135 ; SSE-NEXT: movaps %xmm2, %xmm1
13136 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
13137 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
13138 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
13139 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
13140 ; SSE-NEXT: movaps 432(%rdx), %xmm2
13141 ; SSE-NEXT: movaps 432(%rcx), %xmm0
13142 ; SSE-NEXT: movaps %xmm2, %xmm1
13143 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
13144 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
13145 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
13146 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
13147 ; SSE-NEXT: movaps 432(%r8), %xmm2
13148 ; SSE-NEXT: movaps 432(%r9), %xmm0
13149 ; SSE-NEXT: movaps %xmm2, %xmm1
13150 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
13151 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
13152 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
13153 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
13154 ; SSE-NEXT: movaps 432(%r10), %xmm2
13155 ; SSE-NEXT: movaps 432(%rax), %xmm0
13156 ; SSE-NEXT: movaps %xmm2, %xmm1
13157 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
13158 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
13159 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
13160 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
13161 ; SSE-NEXT: movaps 448(%rdi), %xmm2
13162 ; SSE-NEXT: movaps 448(%rsi), %xmm0
13163 ; SSE-NEXT: movaps %xmm2, %xmm1
13164 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
13165 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
13166 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
13167 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
13168 ; SSE-NEXT: movaps 448(%rdx), %xmm2
13169 ; SSE-NEXT: movaps 448(%rcx), %xmm0
13170 ; SSE-NEXT: movaps %xmm2, %xmm1
13171 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
13172 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
13173 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
13174 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
13175 ; SSE-NEXT: movaps 448(%r8), %xmm2
13176 ; SSE-NEXT: movaps 448(%r9), %xmm0
13177 ; SSE-NEXT: movaps %xmm2, %xmm1
13178 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
13179 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
13180 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
13181 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
13182 ; SSE-NEXT: movaps 448(%r10), %xmm2
13183 ; SSE-NEXT: movaps 448(%rax), %xmm0
13184 ; SSE-NEXT: movaps %xmm2, %xmm1
13185 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
13186 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
13187 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
13188 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
13189 ; SSE-NEXT: movaps 464(%rdi), %xmm2
13190 ; SSE-NEXT: movaps 464(%rsi), %xmm0
13191 ; SSE-NEXT: movaps %xmm2, %xmm1
13192 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
13193 ; SSE-NEXT: movaps %xmm1, (%rsp) # 16-byte Spill
13194 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
13195 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
13196 ; SSE-NEXT: movaps 464(%rdx), %xmm2
13197 ; SSE-NEXT: movaps 464(%rcx), %xmm0
13198 ; SSE-NEXT: movaps %xmm2, %xmm1
13199 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
13200 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
13201 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
13202 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
13203 ; SSE-NEXT: movaps 464(%r8), %xmm2
13204 ; SSE-NEXT: movaps 464(%r9), %xmm0
13205 ; SSE-NEXT: movaps %xmm2, %xmm1
13206 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
13207 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
13208 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
13209 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
13210 ; SSE-NEXT: movaps 464(%r10), %xmm2
13211 ; SSE-NEXT: movaps 464(%rax), %xmm0
13212 ; SSE-NEXT: movaps %xmm2, %xmm1
13213 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
13214 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
13215 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
13216 ; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
13217 ; SSE-NEXT: movaps 480(%rdi), %xmm13
13218 ; SSE-NEXT: movaps 480(%rsi), %xmm0
13219 ; SSE-NEXT: movaps %xmm13, %xmm1
13220 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
13221 ; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
13222 ; SSE-NEXT: unpckhpd {{.*#+}} xmm13 = xmm13[1],xmm0[1]
13223 ; SSE-NEXT: movaps 480(%rdx), %xmm10
13224 ; SSE-NEXT: movaps 480(%rcx), %xmm0
13225 ; SSE-NEXT: movaps %xmm10, %xmm15
13226 ; SSE-NEXT: movlhps {{.*#+}} xmm15 = xmm15[0],xmm0[0]
13227 ; SSE-NEXT: unpckhpd {{.*#+}} xmm10 = xmm10[1],xmm0[1]
13228 ; SSE-NEXT: movaps 480(%r8), %xmm9
13229 ; SSE-NEXT: movaps 480(%r9), %xmm0
13230 ; SSE-NEXT: movaps %xmm9, %xmm14
13231 ; SSE-NEXT: movlhps {{.*#+}} xmm14 = xmm14[0],xmm0[0]
13232 ; SSE-NEXT: unpckhpd {{.*#+}} xmm9 = xmm9[1],xmm0[1]
13233 ; SSE-NEXT: movaps 480(%r10), %xmm11
13234 ; SSE-NEXT: movaps 480(%rax), %xmm1
13235 ; SSE-NEXT: movaps %xmm11, %xmm12
13236 ; SSE-NEXT: movlhps {{.*#+}} xmm12 = xmm12[0],xmm1[0]
13237 ; SSE-NEXT: unpckhpd {{.*#+}} xmm11 = xmm11[1],xmm1[1]
13238 ; SSE-NEXT: movaps 496(%rdi), %xmm7
13239 ; SSE-NEXT: movaps 496(%rsi), %xmm0
13240 ; SSE-NEXT: movaps %xmm7, %xmm8
13241 ; SSE-NEXT: movlhps {{.*#+}} xmm8 = xmm8[0],xmm0[0]
13242 ; SSE-NEXT: unpckhpd {{.*#+}} xmm7 = xmm7[1],xmm0[1]
13243 ; SSE-NEXT: movaps 496(%rdx), %xmm5
13244 ; SSE-NEXT: movaps 496(%rcx), %xmm1
13245 ; SSE-NEXT: movaps %xmm5, %xmm6
13246 ; SSE-NEXT: movlhps {{.*#+}} xmm6 = xmm6[0],xmm1[0]
13247 ; SSE-NEXT: unpckhpd {{.*#+}} xmm5 = xmm5[1],xmm1[1]
13248 ; SSE-NEXT: movaps 496(%r8), %xmm1
13249 ; SSE-NEXT: movaps 496(%r9), %xmm2
13250 ; SSE-NEXT: movaps %xmm1, %xmm4
13251 ; SSE-NEXT: movlhps {{.*#+}} xmm4 = xmm4[0],xmm2[0]
13252 ; SSE-NEXT: unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm2[1]
13253 ; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rcx
13254 ; SSE-NEXT: movaps 496(%r10), %xmm2
13255 ; SSE-NEXT: movaps 496(%rax), %xmm3
13256 ; SSE-NEXT: movaps %xmm2, %xmm0
13257 ; SSE-NEXT: movlhps {{.*#+}} xmm0 = xmm0[0],xmm3[0]
13258 ; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm3[1]
13259 ; SSE-NEXT: movaps %xmm2, 4080(%rcx)
13260 ; SSE-NEXT: movaps %xmm1, 4064(%rcx)
13261 ; SSE-NEXT: movaps %xmm5, 4048(%rcx)
13262 ; SSE-NEXT: movaps %xmm7, 4032(%rcx)
13263 ; SSE-NEXT: movaps %xmm0, 4016(%rcx)
13264 ; SSE-NEXT: movaps %xmm4, 4000(%rcx)
13265 ; SSE-NEXT: movaps %xmm6, 3984(%rcx)
13266 ; SSE-NEXT: movaps %xmm8, 3968(%rcx)
13267 ; SSE-NEXT: movaps %xmm11, 3952(%rcx)
13268 ; SSE-NEXT: movaps %xmm9, 3936(%rcx)
13269 ; SSE-NEXT: movaps %xmm10, 3920(%rcx)
13270 ; SSE-NEXT: movaps %xmm13, 3904(%rcx)
13271 ; SSE-NEXT: movaps %xmm12, 3888(%rcx)
13272 ; SSE-NEXT: movaps %xmm14, 3872(%rcx)
13273 ; SSE-NEXT: movaps %xmm15, 3856(%rcx)
13274 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13275 ; SSE-NEXT: movaps %xmm0, 3840(%rcx)
13276 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13277 ; SSE-NEXT: movaps %xmm0, 3824(%rcx)
13278 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13279 ; SSE-NEXT: movaps %xmm0, 3808(%rcx)
13280 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13281 ; SSE-NEXT: movaps %xmm0, 3792(%rcx)
13282 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13283 ; SSE-NEXT: movaps %xmm0, 3776(%rcx)
13284 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13285 ; SSE-NEXT: movaps %xmm0, 3760(%rcx)
13286 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13287 ; SSE-NEXT: movaps %xmm0, 3744(%rcx)
13288 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13289 ; SSE-NEXT: movaps %xmm0, 3728(%rcx)
13290 ; SSE-NEXT: movaps (%rsp), %xmm0 # 16-byte Reload
13291 ; SSE-NEXT: movaps %xmm0, 3712(%rcx)
13292 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13293 ; SSE-NEXT: movaps %xmm0, 3696(%rcx)
13294 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13295 ; SSE-NEXT: movaps %xmm0, 3680(%rcx)
13296 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13297 ; SSE-NEXT: movaps %xmm0, 3664(%rcx)
13298 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13299 ; SSE-NEXT: movaps %xmm0, 3648(%rcx)
13300 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13301 ; SSE-NEXT: movaps %xmm0, 3632(%rcx)
13302 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13303 ; SSE-NEXT: movaps %xmm0, 3616(%rcx)
13304 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13305 ; SSE-NEXT: movaps %xmm0, 3600(%rcx)
13306 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13307 ; SSE-NEXT: movaps %xmm0, 3584(%rcx)
13308 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13309 ; SSE-NEXT: movaps %xmm0, 3568(%rcx)
13310 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13311 ; SSE-NEXT: movaps %xmm0, 3552(%rcx)
13312 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13313 ; SSE-NEXT: movaps %xmm0, 3536(%rcx)
13314 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13315 ; SSE-NEXT: movaps %xmm0, 3520(%rcx)
13316 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13317 ; SSE-NEXT: movaps %xmm0, 3504(%rcx)
13318 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13319 ; SSE-NEXT: movaps %xmm0, 3488(%rcx)
13320 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13321 ; SSE-NEXT: movaps %xmm0, 3472(%rcx)
13322 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13323 ; SSE-NEXT: movaps %xmm0, 3456(%rcx)
13324 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13325 ; SSE-NEXT: movaps %xmm0, 3440(%rcx)
13326 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13327 ; SSE-NEXT: movaps %xmm0, 3424(%rcx)
13328 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13329 ; SSE-NEXT: movaps %xmm0, 3408(%rcx)
13330 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13331 ; SSE-NEXT: movaps %xmm0, 3392(%rcx)
13332 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13333 ; SSE-NEXT: movaps %xmm0, 3376(%rcx)
13334 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13335 ; SSE-NEXT: movaps %xmm0, 3360(%rcx)
13336 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13337 ; SSE-NEXT: movaps %xmm0, 3344(%rcx)
13338 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13339 ; SSE-NEXT: movaps %xmm0, 3328(%rcx)
13340 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13341 ; SSE-NEXT: movaps %xmm0, 3312(%rcx)
13342 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13343 ; SSE-NEXT: movaps %xmm0, 3296(%rcx)
13344 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13345 ; SSE-NEXT: movaps %xmm0, 3280(%rcx)
13346 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13347 ; SSE-NEXT: movaps %xmm0, 3264(%rcx)
13348 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13349 ; SSE-NEXT: movaps %xmm0, 3248(%rcx)
13350 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13351 ; SSE-NEXT: movaps %xmm0, 3232(%rcx)
13352 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13353 ; SSE-NEXT: movaps %xmm0, 3216(%rcx)
13354 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13355 ; SSE-NEXT: movaps %xmm0, 3200(%rcx)
13356 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13357 ; SSE-NEXT: movaps %xmm0, 3184(%rcx)
13358 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13359 ; SSE-NEXT: movaps %xmm0, 3168(%rcx)
13360 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13361 ; SSE-NEXT: movaps %xmm0, 3152(%rcx)
13362 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13363 ; SSE-NEXT: movaps %xmm0, 3136(%rcx)
13364 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13365 ; SSE-NEXT: movaps %xmm0, 3120(%rcx)
13366 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13367 ; SSE-NEXT: movaps %xmm0, 3104(%rcx)
13368 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13369 ; SSE-NEXT: movaps %xmm0, 3088(%rcx)
13370 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13371 ; SSE-NEXT: movaps %xmm0, 3072(%rcx)
13372 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13373 ; SSE-NEXT: movaps %xmm0, 3056(%rcx)
13374 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13375 ; SSE-NEXT: movaps %xmm0, 3040(%rcx)
13376 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13377 ; SSE-NEXT: movaps %xmm0, 3024(%rcx)
13378 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13379 ; SSE-NEXT: movaps %xmm0, 3008(%rcx)
13380 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13381 ; SSE-NEXT: movaps %xmm0, 2992(%rcx)
13382 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13383 ; SSE-NEXT: movaps %xmm0, 2976(%rcx)
13384 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13385 ; SSE-NEXT: movaps %xmm0, 2960(%rcx)
13386 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13387 ; SSE-NEXT: movaps %xmm0, 2944(%rcx)
13388 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13389 ; SSE-NEXT: movaps %xmm0, 2928(%rcx)
13390 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13391 ; SSE-NEXT: movaps %xmm0, 2912(%rcx)
13392 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13393 ; SSE-NEXT: movaps %xmm0, 2896(%rcx)
13394 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13395 ; SSE-NEXT: movaps %xmm0, 2880(%rcx)
13396 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13397 ; SSE-NEXT: movaps %xmm0, 2864(%rcx)
13398 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13399 ; SSE-NEXT: movaps %xmm0, 2848(%rcx)
13400 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13401 ; SSE-NEXT: movaps %xmm0, 2832(%rcx)
13402 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13403 ; SSE-NEXT: movaps %xmm0, 2816(%rcx)
13404 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13405 ; SSE-NEXT: movaps %xmm0, 2800(%rcx)
13406 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13407 ; SSE-NEXT: movaps %xmm0, 2784(%rcx)
13408 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13409 ; SSE-NEXT: movaps %xmm0, 2768(%rcx)
13410 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13411 ; SSE-NEXT: movaps %xmm0, 2752(%rcx)
13412 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13413 ; SSE-NEXT: movaps %xmm0, 2736(%rcx)
13414 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13415 ; SSE-NEXT: movaps %xmm0, 2720(%rcx)
13416 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13417 ; SSE-NEXT: movaps %xmm0, 2704(%rcx)
13418 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13419 ; SSE-NEXT: movaps %xmm0, 2688(%rcx)
13420 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13421 ; SSE-NEXT: movaps %xmm0, 2672(%rcx)
13422 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13423 ; SSE-NEXT: movaps %xmm0, 2656(%rcx)
13424 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13425 ; SSE-NEXT: movaps %xmm0, 2640(%rcx)
13426 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13427 ; SSE-NEXT: movaps %xmm0, 2624(%rcx)
13428 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13429 ; SSE-NEXT: movaps %xmm0, 2608(%rcx)
13430 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13431 ; SSE-NEXT: movaps %xmm0, 2592(%rcx)
13432 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13433 ; SSE-NEXT: movaps %xmm0, 2576(%rcx)
13434 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13435 ; SSE-NEXT: movaps %xmm0, 2560(%rcx)
13436 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13437 ; SSE-NEXT: movaps %xmm0, 2544(%rcx)
13438 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13439 ; SSE-NEXT: movaps %xmm0, 2528(%rcx)
13440 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13441 ; SSE-NEXT: movaps %xmm0, 2512(%rcx)
13442 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13443 ; SSE-NEXT: movaps %xmm0, 2496(%rcx)
13444 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13445 ; SSE-NEXT: movaps %xmm0, 2480(%rcx)
13446 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13447 ; SSE-NEXT: movaps %xmm0, 2464(%rcx)
13448 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13449 ; SSE-NEXT: movaps %xmm0, 2448(%rcx)
13450 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13451 ; SSE-NEXT: movaps %xmm0, 2432(%rcx)
13452 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13453 ; SSE-NEXT: movaps %xmm0, 2416(%rcx)
13454 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13455 ; SSE-NEXT: movaps %xmm0, 2400(%rcx)
13456 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13457 ; SSE-NEXT: movaps %xmm0, 2384(%rcx)
13458 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13459 ; SSE-NEXT: movaps %xmm0, 2368(%rcx)
13460 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13461 ; SSE-NEXT: movaps %xmm0, 2352(%rcx)
13462 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13463 ; SSE-NEXT: movaps %xmm0, 2336(%rcx)
13464 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13465 ; SSE-NEXT: movaps %xmm0, 2320(%rcx)
13466 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13467 ; SSE-NEXT: movaps %xmm0, 2304(%rcx)
13468 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13469 ; SSE-NEXT: movaps %xmm0, 2288(%rcx)
13470 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13471 ; SSE-NEXT: movaps %xmm0, 2272(%rcx)
13472 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13473 ; SSE-NEXT: movaps %xmm0, 2256(%rcx)
13474 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13475 ; SSE-NEXT: movaps %xmm0, 2240(%rcx)
13476 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13477 ; SSE-NEXT: movaps %xmm0, 2224(%rcx)
13478 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13479 ; SSE-NEXT: movaps %xmm0, 2208(%rcx)
13480 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13481 ; SSE-NEXT: movaps %xmm0, 2192(%rcx)
13482 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13483 ; SSE-NEXT: movaps %xmm0, 2176(%rcx)
13484 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13485 ; SSE-NEXT: movaps %xmm0, 2160(%rcx)
13486 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13487 ; SSE-NEXT: movaps %xmm0, 2144(%rcx)
13488 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13489 ; SSE-NEXT: movaps %xmm0, 2128(%rcx)
13490 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13491 ; SSE-NEXT: movaps %xmm0, 2112(%rcx)
13492 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13493 ; SSE-NEXT: movaps %xmm0, 2096(%rcx)
13494 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13495 ; SSE-NEXT: movaps %xmm0, 2080(%rcx)
13496 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13497 ; SSE-NEXT: movaps %xmm0, 2064(%rcx)
13498 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13499 ; SSE-NEXT: movaps %xmm0, 2048(%rcx)
13500 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13501 ; SSE-NEXT: movaps %xmm0, 2032(%rcx)
13502 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13503 ; SSE-NEXT: movaps %xmm0, 2016(%rcx)
13504 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13505 ; SSE-NEXT: movaps %xmm0, 2000(%rcx)
13506 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13507 ; SSE-NEXT: movaps %xmm0, 1984(%rcx)
13508 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13509 ; SSE-NEXT: movaps %xmm0, 1968(%rcx)
13510 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13511 ; SSE-NEXT: movaps %xmm0, 1952(%rcx)
13512 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13513 ; SSE-NEXT: movaps %xmm0, 1936(%rcx)
13514 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13515 ; SSE-NEXT: movaps %xmm0, 1920(%rcx)
13516 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13517 ; SSE-NEXT: movaps %xmm0, 1904(%rcx)
13518 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13519 ; SSE-NEXT: movaps %xmm0, 1888(%rcx)
13520 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13521 ; SSE-NEXT: movaps %xmm0, 1872(%rcx)
13522 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13523 ; SSE-NEXT: movaps %xmm0, 1856(%rcx)
13524 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13525 ; SSE-NEXT: movaps %xmm0, 1840(%rcx)
13526 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13527 ; SSE-NEXT: movaps %xmm0, 1824(%rcx)
13528 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13529 ; SSE-NEXT: movaps %xmm0, 1808(%rcx)
13530 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13531 ; SSE-NEXT: movaps %xmm0, 1792(%rcx)
13532 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13533 ; SSE-NEXT: movaps %xmm0, 1776(%rcx)
13534 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13535 ; SSE-NEXT: movaps %xmm0, 1760(%rcx)
13536 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13537 ; SSE-NEXT: movaps %xmm0, 1744(%rcx)
13538 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13539 ; SSE-NEXT: movaps %xmm0, 1728(%rcx)
13540 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13541 ; SSE-NEXT: movaps %xmm0, 1712(%rcx)
13542 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13543 ; SSE-NEXT: movaps %xmm0, 1696(%rcx)
13544 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13545 ; SSE-NEXT: movaps %xmm0, 1680(%rcx)
13546 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13547 ; SSE-NEXT: movaps %xmm0, 1664(%rcx)
13548 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13549 ; SSE-NEXT: movaps %xmm0, 1648(%rcx)
13550 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13551 ; SSE-NEXT: movaps %xmm0, 1632(%rcx)
13552 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13553 ; SSE-NEXT: movaps %xmm0, 1616(%rcx)
13554 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13555 ; SSE-NEXT: movaps %xmm0, 1600(%rcx)
13556 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13557 ; SSE-NEXT: movaps %xmm0, 1584(%rcx)
13558 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13559 ; SSE-NEXT: movaps %xmm0, 1568(%rcx)
13560 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13561 ; SSE-NEXT: movaps %xmm0, 1552(%rcx)
13562 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13563 ; SSE-NEXT: movaps %xmm0, 1536(%rcx)
13564 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13565 ; SSE-NEXT: movaps %xmm0, 1520(%rcx)
13566 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13567 ; SSE-NEXT: movaps %xmm0, 1504(%rcx)
13568 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13569 ; SSE-NEXT: movaps %xmm0, 1488(%rcx)
13570 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13571 ; SSE-NEXT: movaps %xmm0, 1472(%rcx)
13572 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13573 ; SSE-NEXT: movaps %xmm0, 1456(%rcx)
13574 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13575 ; SSE-NEXT: movaps %xmm0, 1440(%rcx)
13576 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13577 ; SSE-NEXT: movaps %xmm0, 1424(%rcx)
13578 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13579 ; SSE-NEXT: movaps %xmm0, 1408(%rcx)
13580 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13581 ; SSE-NEXT: movaps %xmm0, 1392(%rcx)
13582 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13583 ; SSE-NEXT: movaps %xmm0, 1376(%rcx)
13584 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13585 ; SSE-NEXT: movaps %xmm0, 1360(%rcx)
13586 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13587 ; SSE-NEXT: movaps %xmm0, 1344(%rcx)
13588 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13589 ; SSE-NEXT: movaps %xmm0, 1328(%rcx)
13590 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13591 ; SSE-NEXT: movaps %xmm0, 1312(%rcx)
13592 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13593 ; SSE-NEXT: movaps %xmm0, 1296(%rcx)
13594 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13595 ; SSE-NEXT: movaps %xmm0, 1280(%rcx)
13596 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13597 ; SSE-NEXT: movaps %xmm0, 1264(%rcx)
13598 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13599 ; SSE-NEXT: movaps %xmm0, 1248(%rcx)
13600 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13601 ; SSE-NEXT: movaps %xmm0, 1232(%rcx)
13602 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13603 ; SSE-NEXT: movaps %xmm0, 1216(%rcx)
13604 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13605 ; SSE-NEXT: movaps %xmm0, 1200(%rcx)
13606 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13607 ; SSE-NEXT: movaps %xmm0, 1184(%rcx)
13608 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13609 ; SSE-NEXT: movaps %xmm0, 1168(%rcx)
13610 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13611 ; SSE-NEXT: movaps %xmm0, 1152(%rcx)
13612 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13613 ; SSE-NEXT: movaps %xmm0, 1136(%rcx)
13614 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13615 ; SSE-NEXT: movaps %xmm0, 1120(%rcx)
13616 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13617 ; SSE-NEXT: movaps %xmm0, 1104(%rcx)
13618 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13619 ; SSE-NEXT: movaps %xmm0, 1088(%rcx)
13620 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13621 ; SSE-NEXT: movaps %xmm0, 1072(%rcx)
13622 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13623 ; SSE-NEXT: movaps %xmm0, 1056(%rcx)
13624 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13625 ; SSE-NEXT: movaps %xmm0, 1040(%rcx)
13626 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13627 ; SSE-NEXT: movaps %xmm0, 1024(%rcx)
13628 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13629 ; SSE-NEXT: movaps %xmm0, 1008(%rcx)
13630 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13631 ; SSE-NEXT: movaps %xmm0, 992(%rcx)
13632 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13633 ; SSE-NEXT: movaps %xmm0, 976(%rcx)
13634 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13635 ; SSE-NEXT: movaps %xmm0, 960(%rcx)
13636 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13637 ; SSE-NEXT: movaps %xmm0, 944(%rcx)
13638 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13639 ; SSE-NEXT: movaps %xmm0, 928(%rcx)
13640 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13641 ; SSE-NEXT: movaps %xmm0, 912(%rcx)
13642 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13643 ; SSE-NEXT: movaps %xmm0, 896(%rcx)
13644 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13645 ; SSE-NEXT: movaps %xmm0, 880(%rcx)
13646 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13647 ; SSE-NEXT: movaps %xmm0, 864(%rcx)
13648 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13649 ; SSE-NEXT: movaps %xmm0, 848(%rcx)
13650 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13651 ; SSE-NEXT: movaps %xmm0, 832(%rcx)
13652 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13653 ; SSE-NEXT: movaps %xmm0, 816(%rcx)
13654 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13655 ; SSE-NEXT: movaps %xmm0, 800(%rcx)
13656 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13657 ; SSE-NEXT: movaps %xmm0, 784(%rcx)
13658 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13659 ; SSE-NEXT: movaps %xmm0, 768(%rcx)
13660 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13661 ; SSE-NEXT: movaps %xmm0, 752(%rcx)
13662 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13663 ; SSE-NEXT: movaps %xmm0, 736(%rcx)
13664 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13665 ; SSE-NEXT: movaps %xmm0, 720(%rcx)
13666 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13667 ; SSE-NEXT: movaps %xmm0, 704(%rcx)
13668 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13669 ; SSE-NEXT: movaps %xmm0, 688(%rcx)
13670 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13671 ; SSE-NEXT: movaps %xmm0, 672(%rcx)
13672 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13673 ; SSE-NEXT: movaps %xmm0, 656(%rcx)
13674 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13675 ; SSE-NEXT: movaps %xmm0, 640(%rcx)
13676 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13677 ; SSE-NEXT: movaps %xmm0, 624(%rcx)
13678 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13679 ; SSE-NEXT: movaps %xmm0, 608(%rcx)
13680 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13681 ; SSE-NEXT: movaps %xmm0, 592(%rcx)
13682 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13683 ; SSE-NEXT: movaps %xmm0, 576(%rcx)
13684 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13685 ; SSE-NEXT: movaps %xmm0, 560(%rcx)
13686 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13687 ; SSE-NEXT: movaps %xmm0, 544(%rcx)
13688 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13689 ; SSE-NEXT: movaps %xmm0, 528(%rcx)
13690 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13691 ; SSE-NEXT: movaps %xmm0, 512(%rcx)
13692 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13693 ; SSE-NEXT: movaps %xmm0, 496(%rcx)
13694 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13695 ; SSE-NEXT: movaps %xmm0, 480(%rcx)
13696 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13697 ; SSE-NEXT: movaps %xmm0, 464(%rcx)
13698 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13699 ; SSE-NEXT: movaps %xmm0, 448(%rcx)
13700 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13701 ; SSE-NEXT: movaps %xmm0, 432(%rcx)
13702 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13703 ; SSE-NEXT: movaps %xmm0, 416(%rcx)
13704 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13705 ; SSE-NEXT: movaps %xmm0, 400(%rcx)
13706 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13707 ; SSE-NEXT: movaps %xmm0, 384(%rcx)
13708 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13709 ; SSE-NEXT: movaps %xmm0, 368(%rcx)
13710 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13711 ; SSE-NEXT: movaps %xmm0, 352(%rcx)
13712 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13713 ; SSE-NEXT: movaps %xmm0, 336(%rcx)
13714 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13715 ; SSE-NEXT: movaps %xmm0, 320(%rcx)
13716 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13717 ; SSE-NEXT: movaps %xmm0, 304(%rcx)
13718 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13719 ; SSE-NEXT: movaps %xmm0, 288(%rcx)
13720 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13721 ; SSE-NEXT: movaps %xmm0, 272(%rcx)
13722 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13723 ; SSE-NEXT: movaps %xmm0, 256(%rcx)
13724 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13725 ; SSE-NEXT: movaps %xmm0, 240(%rcx)
13726 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13727 ; SSE-NEXT: movaps %xmm0, 224(%rcx)
13728 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13729 ; SSE-NEXT: movaps %xmm0, 208(%rcx)
13730 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13731 ; SSE-NEXT: movaps %xmm0, 192(%rcx)
13732 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13733 ; SSE-NEXT: movaps %xmm0, 176(%rcx)
13734 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13735 ; SSE-NEXT: movaps %xmm0, 160(%rcx)
13736 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13737 ; SSE-NEXT: movaps %xmm0, 144(%rcx)
13738 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13739 ; SSE-NEXT: movaps %xmm0, 128(%rcx)
13740 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13741 ; SSE-NEXT: movaps %xmm0, 112(%rcx)
13742 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13743 ; SSE-NEXT: movaps %xmm0, 96(%rcx)
13744 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13745 ; SSE-NEXT: movaps %xmm0, 80(%rcx)
13746 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13747 ; SSE-NEXT: movaps %xmm0, 64(%rcx)
13748 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13749 ; SSE-NEXT: movaps %xmm0, 48(%rcx)
13750 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13751 ; SSE-NEXT: movaps %xmm0, 32(%rcx)
13752 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13753 ; SSE-NEXT: movaps %xmm0, 16(%rcx)
13754 ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
13755 ; SSE-NEXT: movaps %xmm0, (%rcx)
13756 ; SSE-NEXT: addq $3736, %rsp # imm = 0xE98
13757 ; SSE-NEXT: retq
13758 ;
13759 ; AVX-LABEL: store_i64_stride8_vf64:
13760 ; AVX: # %bb.0:
13761 ; AVX-NEXT: subq $3784, %rsp # imm = 0xEC8
13762 ; AVX-NEXT: movq {{[0-9]+}}(%rsp), %r10
13763 ; AVX-NEXT: movq {{[0-9]+}}(%rsp), %rax
13764 ; AVX-NEXT: vmovaps (%rsi), %xmm0
13765 ; AVX-NEXT: vmovaps (%rdi), %xmm1
13766 ; AVX-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
13767 ; AVX-NEXT: vmovaps (%rcx), %xmm3
13768 ; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm2, %ymm4
13769 ; AVX-NEXT: vinsertf128 $1, (%rdx), %ymm2, %ymm2
13770 ; AVX-NEXT: vshufpd {{.*#+}} ymm2 = ymm2[0],ymm4[1],ymm2[2],ymm4[2]
13771 ; AVX-NEXT: vmovupd %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
13772 ; AVX-NEXT: vmovaps (%r9), %xmm2
13773 ; AVX-NEXT: vmovaps (%r8), %xmm4
13774 ; AVX-NEXT: vmovlhps {{.*#+}} xmm5 = xmm4[0],xmm2[0]
13775 ; AVX-NEXT: vmovaps (%rax), %xmm6
13776 ; AVX-NEXT: vinsertf128 $1, %xmm6, %ymm5, %ymm7
13777 ; AVX-NEXT: vinsertf128 $1, (%r10), %ymm5, %ymm5
13778 ; AVX-NEXT: vshufpd {{.*#+}} ymm5 = ymm5[0],ymm7[1],ymm5[2],ymm7[2]
13779 ; AVX-NEXT: vmovupd %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
13780 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
13781 ; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm1
13782 ; AVX-NEXT: vbroadcastsd 8(%rdx), %ymm3
13783 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm3[4,5,6,7]
13784 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7]
13785 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
13786 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm4[1],xmm2[1]
13787 ; AVX-NEXT: vinsertf128 $1, %xmm6, %ymm0, %ymm1
13788 ; AVX-NEXT: vbroadcastsd 8(%r10), %ymm2
13789 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
13790 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7]
13791 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
13792 ; AVX-NEXT: vmovaps 32(%rsi), %xmm4
13793 ; AVX-NEXT: vmovaps 32(%rdi), %xmm5
13794 ; AVX-NEXT: vmovlhps {{.*#+}} xmm1 = xmm5[0],xmm4[0]
13795 ; AVX-NEXT: vmovaps 32(%rcx), %xmm6
13796 ; AVX-NEXT: vmovaps 64(%rcx), %xmm0
13797 ; AVX-NEXT: vinsertf128 $1, %xmm6, %ymm1, %ymm2
13798 ; AVX-NEXT: vinsertf128 $1, 32(%rdx), %ymm1, %ymm1
13799 ; AVX-NEXT: vshufpd {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2],ymm2[2]
13800 ; AVX-NEXT: vmovupd %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
13801 ; AVX-NEXT: vmovaps 32(%r9), %xmm7
13802 ; AVX-NEXT: vmovaps 64(%r9), %xmm1
13803 ; AVX-NEXT: vmovaps 32(%r8), %xmm8
13804 ; AVX-NEXT: vmovaps 64(%r8), %xmm2
13805 ; AVX-NEXT: vmovlhps {{.*#+}} xmm3 = xmm8[0],xmm7[0]
13806 ; AVX-NEXT: vmovaps 32(%rax), %xmm9
13807 ; AVX-NEXT: vinsertf128 $1, %xmm9, %ymm3, %ymm10
13808 ; AVX-NEXT: vinsertf128 $1, 32(%r10), %ymm3, %ymm3
13809 ; AVX-NEXT: vshufpd {{.*#+}} ymm3 = ymm3[0],ymm10[1],ymm3[2],ymm10[2]
13810 ; AVX-NEXT: vmovupd %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
13811 ; AVX-NEXT: vmovaps 64(%rax), %xmm3
13812 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm4 = xmm5[1],xmm4[1]
13813 ; AVX-NEXT: vinsertf128 $1, %xmm6, %ymm4, %ymm5
13814 ; AVX-NEXT: vbroadcastsd 40(%rdx), %ymm6
13815 ; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm6[4,5,6,7]
13816 ; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1],ymm5[2,3],ymm4[4,5],ymm5[6,7]
13817 ; AVX-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
13818 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm4 = xmm8[1],xmm7[1]
13819 ; AVX-NEXT: vinsertf128 $1, %xmm9, %ymm4, %ymm5
13820 ; AVX-NEXT: vbroadcastsd 40(%r10), %ymm6
13821 ; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm6[4,5,6,7]
13822 ; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1],ymm5[2,3],ymm4[4,5],ymm5[6,7]
13823 ; AVX-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
13824 ; AVX-NEXT: vmovaps 64(%rsi), %xmm4
13825 ; AVX-NEXT: vmovaps 64(%rdi), %xmm5
13826 ; AVX-NEXT: vmovlhps {{.*#+}} xmm6 = xmm5[0],xmm4[0]
13827 ; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm6, %ymm7
13828 ; AVX-NEXT: vinsertf128 $1, 64(%rdx), %ymm6, %ymm6
13829 ; AVX-NEXT: vshufpd {{.*#+}} ymm6 = ymm6[0],ymm7[1],ymm6[2],ymm7[2]
13830 ; AVX-NEXT: vmovupd %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
13831 ; AVX-NEXT: vmovlhps {{.*#+}} xmm6 = xmm2[0],xmm1[0]
13832 ; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm6, %ymm7
13833 ; AVX-NEXT: vinsertf128 $1, 64(%r10), %ymm6, %ymm6
13834 ; AVX-NEXT: vshufpd {{.*#+}} ymm6 = ymm6[0],ymm7[1],ymm6[2],ymm7[2]
13835 ; AVX-NEXT: vmovupd %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
13836 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm4 = xmm5[1],xmm4[1]
13837 ; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm4, %ymm0
13838 ; AVX-NEXT: vbroadcastsd 72(%rdx), %ymm5
13839 ; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm5[4,5,6,7]
13840 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm4[0,1],ymm0[2,3],ymm4[4,5],ymm0[6,7]
13841 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
13842 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm2[1],xmm1[1]
13843 ; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm1
13844 ; AVX-NEXT: vbroadcastsd 72(%r10), %ymm2
13845 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
13846 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7]
13847 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
13848 ; AVX-NEXT: vmovaps 96(%rsi), %xmm0
13849 ; AVX-NEXT: vmovaps 96(%rdi), %xmm1
13850 ; AVX-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
13851 ; AVX-NEXT: vmovaps 96(%rcx), %xmm3
13852 ; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm2, %ymm4
13853 ; AVX-NEXT: vinsertf128 $1, 96(%rdx), %ymm2, %ymm2
13854 ; AVX-NEXT: vshufpd {{.*#+}} ymm2 = ymm2[0],ymm4[1],ymm2[2],ymm4[2]
13855 ; AVX-NEXT: vmovupd %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
13856 ; AVX-NEXT: vmovaps 96(%r9), %xmm2
13857 ; AVX-NEXT: vmovaps 96(%r8), %xmm4
13858 ; AVX-NEXT: vmovlhps {{.*#+}} xmm5 = xmm4[0],xmm2[0]
13859 ; AVX-NEXT: vmovaps 96(%rax), %xmm6
13860 ; AVX-NEXT: vinsertf128 $1, %xmm6, %ymm5, %ymm7
13861 ; AVX-NEXT: vinsertf128 $1, 96(%r10), %ymm5, %ymm5
13862 ; AVX-NEXT: vshufpd {{.*#+}} ymm5 = ymm5[0],ymm7[1],ymm5[2],ymm7[2]
13863 ; AVX-NEXT: vmovupd %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
13864 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
13865 ; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm1
13866 ; AVX-NEXT: vbroadcastsd 104(%rdx), %ymm3
13867 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm3[4,5,6,7]
13868 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7]
13869 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
13870 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm4[1],xmm2[1]
13871 ; AVX-NEXT: vinsertf128 $1, %xmm6, %ymm0, %ymm1
13872 ; AVX-NEXT: vbroadcastsd 104(%r10), %ymm2
13873 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
13874 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7]
13875 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
13876 ; AVX-NEXT: vmovaps 128(%rsi), %xmm0
13877 ; AVX-NEXT: vmovaps 128(%rdi), %xmm1
13878 ; AVX-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
13879 ; AVX-NEXT: vmovaps 128(%rcx), %xmm3
13880 ; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm2, %ymm4
13881 ; AVX-NEXT: vinsertf128 $1, 128(%rdx), %ymm2, %ymm2
13882 ; AVX-NEXT: vshufpd {{.*#+}} ymm2 = ymm2[0],ymm4[1],ymm2[2],ymm4[2]
13883 ; AVX-NEXT: vmovupd %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
13884 ; AVX-NEXT: vmovaps 128(%r9), %xmm2
13885 ; AVX-NEXT: vmovaps 128(%r8), %xmm4
13886 ; AVX-NEXT: vmovlhps {{.*#+}} xmm5 = xmm4[0],xmm2[0]
13887 ; AVX-NEXT: vmovaps 128(%rax), %xmm6
13888 ; AVX-NEXT: vinsertf128 $1, %xmm6, %ymm5, %ymm7
13889 ; AVX-NEXT: vinsertf128 $1, 128(%r10), %ymm5, %ymm5
13890 ; AVX-NEXT: vshufpd {{.*#+}} ymm5 = ymm5[0],ymm7[1],ymm5[2],ymm7[2]
13891 ; AVX-NEXT: vmovupd %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
13892 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
13893 ; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm1
13894 ; AVX-NEXT: vbroadcastsd 136(%rdx), %ymm3
13895 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm3[4,5,6,7]
13896 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7]
13897 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
13898 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm4[1],xmm2[1]
13899 ; AVX-NEXT: vinsertf128 $1, %xmm6, %ymm0, %ymm1
13900 ; AVX-NEXT: vbroadcastsd 136(%r10), %ymm2
13901 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
13902 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7]
13903 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
13904 ; AVX-NEXT: vmovaps 160(%rcx), %xmm0
13905 ; AVX-NEXT: vmovaps 160(%rsi), %xmm1
13906 ; AVX-NEXT: vmovaps 160(%rdi), %xmm2
13907 ; AVX-NEXT: vmovlhps {{.*#+}} xmm3 = xmm2[0],xmm1[0]
13908 ; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm3, %ymm4
13909 ; AVX-NEXT: vinsertf128 $1, 160(%rdx), %ymm3, %ymm3
13910 ; AVX-NEXT: vshufpd {{.*#+}} ymm3 = ymm3[0],ymm4[1],ymm3[2],ymm4[2]
13911 ; AVX-NEXT: vmovupd %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
13912 ; AVX-NEXT: vmovaps 160(%r9), %xmm3
13913 ; AVX-NEXT: vmovaps 160(%r8), %xmm4
13914 ; AVX-NEXT: vmovaps 160(%rax), %xmm5
13915 ; AVX-NEXT: vmovlhps {{.*#+}} xmm6 = xmm4[0],xmm3[0]
13916 ; AVX-NEXT: vinsertf128 $1, %xmm5, %ymm6, %ymm7
13917 ; AVX-NEXT: vinsertf128 $1, 160(%r10), %ymm6, %ymm6
13918 ; AVX-NEXT: vshufpd {{.*#+}} ymm6 = ymm6[0],ymm7[1],ymm6[2],ymm7[2]
13919 ; AVX-NEXT: vmovupd %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
13920 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm2[1],xmm1[1]
13921 ; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm1, %ymm0
13922 ; AVX-NEXT: vbroadcastsd 168(%rdx), %ymm2
13923 ; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
13924 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3],ymm1[4,5],ymm0[6,7]
13925 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
13926 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm4[1],xmm3[1]
13927 ; AVX-NEXT: vinsertf128 $1, %xmm5, %ymm0, %ymm1
13928 ; AVX-NEXT: vbroadcastsd 168(%r10), %ymm2
13929 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
13930 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7]
13931 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
13932 ; AVX-NEXT: vmovaps 192(%rsi), %xmm0
13933 ; AVX-NEXT: vmovaps 192(%rdi), %xmm1
13934 ; AVX-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
13935 ; AVX-NEXT: vmovaps 192(%rcx), %xmm3
13936 ; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm2, %ymm4
13937 ; AVX-NEXT: vinsertf128 $1, 192(%rdx), %ymm2, %ymm2
13938 ; AVX-NEXT: vshufpd {{.*#+}} ymm2 = ymm2[0],ymm4[1],ymm2[2],ymm4[2]
13939 ; AVX-NEXT: vmovupd %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
13940 ; AVX-NEXT: vmovaps 192(%r9), %xmm2
13941 ; AVX-NEXT: vmovaps 192(%r8), %xmm4
13942 ; AVX-NEXT: vmovlhps {{.*#+}} xmm5 = xmm4[0],xmm2[0]
13943 ; AVX-NEXT: vmovaps 192(%rax), %xmm6
13944 ; AVX-NEXT: vinsertf128 $1, %xmm6, %ymm5, %ymm7
13945 ; AVX-NEXT: vinsertf128 $1, 192(%r10), %ymm5, %ymm5
13946 ; AVX-NEXT: vshufpd {{.*#+}} ymm5 = ymm5[0],ymm7[1],ymm5[2],ymm7[2]
13947 ; AVX-NEXT: vmovupd %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
13948 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
13949 ; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm1
13950 ; AVX-NEXT: vbroadcastsd 200(%rdx), %ymm3
13951 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm3[4,5,6,7]
13952 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7]
13953 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
13954 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm4[1],xmm2[1]
13955 ; AVX-NEXT: vinsertf128 $1, %xmm6, %ymm0, %ymm1
13956 ; AVX-NEXT: vbroadcastsd 200(%r10), %ymm2
13957 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
13958 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7]
13959 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
13960 ; AVX-NEXT: vmovaps 224(%rcx), %xmm0
13961 ; AVX-NEXT: vmovaps 224(%rsi), %xmm1
13962 ; AVX-NEXT: vmovaps 224(%rdi), %xmm2
13963 ; AVX-NEXT: vmovlhps {{.*#+}} xmm3 = xmm2[0],xmm1[0]
13964 ; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm3, %ymm4
13965 ; AVX-NEXT: vinsertf128 $1, 224(%rdx), %ymm3, %ymm3
13966 ; AVX-NEXT: vshufpd {{.*#+}} ymm3 = ymm3[0],ymm4[1],ymm3[2],ymm4[2]
13967 ; AVX-NEXT: vmovupd %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
13968 ; AVX-NEXT: vmovaps 224(%r9), %xmm3
13969 ; AVX-NEXT: vmovaps 224(%r8), %xmm4
13970 ; AVX-NEXT: vmovaps 224(%rax), %xmm5
13971 ; AVX-NEXT: vmovlhps {{.*#+}} xmm6 = xmm4[0],xmm3[0]
13972 ; AVX-NEXT: vinsertf128 $1, %xmm5, %ymm6, %ymm7
13973 ; AVX-NEXT: vinsertf128 $1, 224(%r10), %ymm6, %ymm6
13974 ; AVX-NEXT: vshufpd {{.*#+}} ymm6 = ymm6[0],ymm7[1],ymm6[2],ymm7[2]
13975 ; AVX-NEXT: vmovupd %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
13976 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm2[1],xmm1[1]
13977 ; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm1, %ymm0
13978 ; AVX-NEXT: vbroadcastsd 232(%rdx), %ymm2
13979 ; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
13980 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3],ymm1[4,5],ymm0[6,7]
13981 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
13982 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm4[1],xmm3[1]
13983 ; AVX-NEXT: vinsertf128 $1, %xmm5, %ymm0, %ymm1
13984 ; AVX-NEXT: vbroadcastsd 232(%r10), %ymm2
13985 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
13986 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7]
13987 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
13988 ; AVX-NEXT: vmovaps 256(%rsi), %xmm0
13989 ; AVX-NEXT: vmovaps 256(%rdi), %xmm1
13990 ; AVX-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
13991 ; AVX-NEXT: vmovaps 256(%rcx), %xmm3
13992 ; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm2, %ymm4
13993 ; AVX-NEXT: vinsertf128 $1, 256(%rdx), %ymm2, %ymm2
13994 ; AVX-NEXT: vshufpd {{.*#+}} ymm2 = ymm2[0],ymm4[1],ymm2[2],ymm4[2]
13995 ; AVX-NEXT: vmovupd %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
13996 ; AVX-NEXT: vmovaps 256(%r9), %xmm2
13997 ; AVX-NEXT: vmovaps 256(%r8), %xmm4
13998 ; AVX-NEXT: vmovlhps {{.*#+}} xmm5 = xmm4[0],xmm2[0]
13999 ; AVX-NEXT: vmovaps 256(%rax), %xmm6
14000 ; AVX-NEXT: vinsertf128 $1, %xmm6, %ymm5, %ymm7
14001 ; AVX-NEXT: vinsertf128 $1, 256(%r10), %ymm5, %ymm5
14002 ; AVX-NEXT: vshufpd {{.*#+}} ymm5 = ymm5[0],ymm7[1],ymm5[2],ymm7[2]
14003 ; AVX-NEXT: vmovupd %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14004 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
14005 ; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm1
14006 ; AVX-NEXT: vbroadcastsd 264(%rdx), %ymm3
14007 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm3[4,5,6,7]
14008 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7]
14009 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14010 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm4[1],xmm2[1]
14011 ; AVX-NEXT: vinsertf128 $1, %xmm6, %ymm0, %ymm1
14012 ; AVX-NEXT: vbroadcastsd 264(%r10), %ymm2
14013 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
14014 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7]
14015 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14016 ; AVX-NEXT: vmovaps 288(%rsi), %xmm0
14017 ; AVX-NEXT: vmovaps 288(%rdi), %xmm1
14018 ; AVX-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
14019 ; AVX-NEXT: vmovaps 288(%rcx), %xmm3
14020 ; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm2, %ymm4
14021 ; AVX-NEXT: vinsertf128 $1, 288(%rdx), %ymm2, %ymm2
14022 ; AVX-NEXT: vshufpd {{.*#+}} ymm2 = ymm2[0],ymm4[1],ymm2[2],ymm4[2]
14023 ; AVX-NEXT: vmovupd %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14024 ; AVX-NEXT: vmovaps 288(%r9), %xmm2
14025 ; AVX-NEXT: vmovaps 288(%r8), %xmm4
14026 ; AVX-NEXT: vmovlhps {{.*#+}} xmm5 = xmm4[0],xmm2[0]
14027 ; AVX-NEXT: vmovaps 288(%rax), %xmm6
14028 ; AVX-NEXT: vinsertf128 $1, %xmm6, %ymm5, %ymm7
14029 ; AVX-NEXT: vinsertf128 $1, 288(%r10), %ymm5, %ymm5
14030 ; AVX-NEXT: vshufpd {{.*#+}} ymm5 = ymm5[0],ymm7[1],ymm5[2],ymm7[2]
14031 ; AVX-NEXT: vmovupd %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14032 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
14033 ; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm1
14034 ; AVX-NEXT: vbroadcastsd 296(%rdx), %ymm3
14035 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm3[4,5,6,7]
14036 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7]
14037 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14038 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm4[1],xmm2[1]
14039 ; AVX-NEXT: vinsertf128 $1, %xmm6, %ymm0, %ymm1
14040 ; AVX-NEXT: vbroadcastsd 296(%r10), %ymm2
14041 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
14042 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7]
14043 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14044 ; AVX-NEXT: vmovaps 320(%rsi), %xmm0
14045 ; AVX-NEXT: vmovaps 320(%rdi), %xmm1
14046 ; AVX-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
14047 ; AVX-NEXT: vmovaps 320(%rcx), %xmm3
14048 ; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm2, %ymm4
14049 ; AVX-NEXT: vinsertf128 $1, 320(%rdx), %ymm2, %ymm2
14050 ; AVX-NEXT: vshufpd {{.*#+}} ymm2 = ymm2[0],ymm4[1],ymm2[2],ymm4[2]
14051 ; AVX-NEXT: vmovupd %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14052 ; AVX-NEXT: vmovaps 320(%r9), %xmm2
14053 ; AVX-NEXT: vmovaps 320(%r8), %xmm4
14054 ; AVX-NEXT: vmovlhps {{.*#+}} xmm5 = xmm4[0],xmm2[0]
14055 ; AVX-NEXT: vmovaps 320(%rax), %xmm6
14056 ; AVX-NEXT: vinsertf128 $1, %xmm6, %ymm5, %ymm7
14057 ; AVX-NEXT: vinsertf128 $1, 320(%r10), %ymm5, %ymm5
14058 ; AVX-NEXT: vshufpd {{.*#+}} ymm5 = ymm5[0],ymm7[1],ymm5[2],ymm7[2]
14059 ; AVX-NEXT: vmovupd %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14060 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
14061 ; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm1
14062 ; AVX-NEXT: vbroadcastsd 328(%rdx), %ymm3
14063 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm3[4,5,6,7]
14064 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7]
14065 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14066 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm4[1],xmm2[1]
14067 ; AVX-NEXT: vinsertf128 $1, %xmm6, %ymm0, %ymm1
14068 ; AVX-NEXT: vbroadcastsd 328(%r10), %ymm2
14069 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
14070 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7]
14071 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14072 ; AVX-NEXT: vmovaps 352(%rcx), %xmm0
14073 ; AVX-NEXT: vmovaps 352(%rsi), %xmm1
14074 ; AVX-NEXT: vmovaps 352(%rdi), %xmm2
14075 ; AVX-NEXT: vmovlhps {{.*#+}} xmm3 = xmm2[0],xmm1[0]
14076 ; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm3, %ymm4
14077 ; AVX-NEXT: vinsertf128 $1, 352(%rdx), %ymm3, %ymm3
14078 ; AVX-NEXT: vshufpd {{.*#+}} ymm3 = ymm3[0],ymm4[1],ymm3[2],ymm4[2]
14079 ; AVX-NEXT: vmovupd %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14080 ; AVX-NEXT: vmovaps 352(%r9), %xmm3
14081 ; AVX-NEXT: vmovaps 352(%r8), %xmm4
14082 ; AVX-NEXT: vmovaps 352(%rax), %xmm5
14083 ; AVX-NEXT: vmovlhps {{.*#+}} xmm6 = xmm4[0],xmm3[0]
14084 ; AVX-NEXT: vinsertf128 $1, %xmm5, %ymm6, %ymm7
14085 ; AVX-NEXT: vinsertf128 $1, 352(%r10), %ymm6, %ymm6
14086 ; AVX-NEXT: vshufpd {{.*#+}} ymm6 = ymm6[0],ymm7[1],ymm6[2],ymm7[2]
14087 ; AVX-NEXT: vmovupd %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14088 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm2[1],xmm1[1]
14089 ; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm1, %ymm0
14090 ; AVX-NEXT: vbroadcastsd 360(%rdx), %ymm2
14091 ; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
14092 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3],ymm1[4,5],ymm0[6,7]
14093 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14094 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm4[1],xmm3[1]
14095 ; AVX-NEXT: vinsertf128 $1, %xmm5, %ymm0, %ymm1
14096 ; AVX-NEXT: vbroadcastsd 360(%r10), %ymm2
14097 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
14098 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7]
14099 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14100 ; AVX-NEXT: vmovaps 384(%rcx), %xmm0
14101 ; AVX-NEXT: vmovaps 384(%rsi), %xmm1
14102 ; AVX-NEXT: vmovaps 384(%rdi), %xmm2
14103 ; AVX-NEXT: vmovlhps {{.*#+}} xmm3 = xmm2[0],xmm1[0]
14104 ; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm3, %ymm4
14105 ; AVX-NEXT: vinsertf128 $1, 384(%rdx), %ymm3, %ymm3
14106 ; AVX-NEXT: vshufpd {{.*#+}} ymm3 = ymm3[0],ymm4[1],ymm3[2],ymm4[2]
14107 ; AVX-NEXT: vmovupd %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14108 ; AVX-NEXT: vmovaps 384(%r9), %xmm3
14109 ; AVX-NEXT: vmovaps 384(%r8), %xmm4
14110 ; AVX-NEXT: vmovaps 384(%rax), %xmm5
14111 ; AVX-NEXT: vmovlhps {{.*#+}} xmm6 = xmm4[0],xmm3[0]
14112 ; AVX-NEXT: vinsertf128 $1, %xmm5, %ymm6, %ymm7
14113 ; AVX-NEXT: vinsertf128 $1, 384(%r10), %ymm6, %ymm6
14114 ; AVX-NEXT: vshufpd {{.*#+}} ymm6 = ymm6[0],ymm7[1],ymm6[2],ymm7[2]
14115 ; AVX-NEXT: vmovupd %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14116 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm2[1],xmm1[1]
14117 ; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm1, %ymm0
14118 ; AVX-NEXT: vbroadcastsd 392(%rdx), %ymm2
14119 ; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
14120 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3],ymm1[4,5],ymm0[6,7]
14121 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14122 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm4[1],xmm3[1]
14123 ; AVX-NEXT: vinsertf128 $1, %xmm5, %ymm0, %ymm1
14124 ; AVX-NEXT: vbroadcastsd 392(%r10), %ymm2
14125 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
14126 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7]
14127 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14128 ; AVX-NEXT: vmovaps 416(%rcx), %xmm0
14129 ; AVX-NEXT: vmovaps 416(%rsi), %xmm1
14130 ; AVX-NEXT: vmovaps 416(%rdi), %xmm2
14131 ; AVX-NEXT: vmovlhps {{.*#+}} xmm3 = xmm2[0],xmm1[0]
14132 ; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm3, %ymm4
14133 ; AVX-NEXT: vinsertf128 $1, 416(%rdx), %ymm3, %ymm3
14134 ; AVX-NEXT: vshufpd {{.*#+}} ymm3 = ymm3[0],ymm4[1],ymm3[2],ymm4[2]
14135 ; AVX-NEXT: vmovupd %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14136 ; AVX-NEXT: vmovaps 416(%r9), %xmm3
14137 ; AVX-NEXT: vmovaps 416(%r8), %xmm4
14138 ; AVX-NEXT: vmovaps 416(%rax), %xmm5
14139 ; AVX-NEXT: vmovlhps {{.*#+}} xmm6 = xmm4[0],xmm3[0]
14140 ; AVX-NEXT: vinsertf128 $1, %xmm5, %ymm6, %ymm7
14141 ; AVX-NEXT: vinsertf128 $1, 416(%r10), %ymm6, %ymm6
14142 ; AVX-NEXT: vshufpd {{.*#+}} ymm6 = ymm6[0],ymm7[1],ymm6[2],ymm7[2]
14143 ; AVX-NEXT: vmovupd %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14144 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm2[1],xmm1[1]
14145 ; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm1, %ymm0
14146 ; AVX-NEXT: vbroadcastsd 424(%rdx), %ymm2
14147 ; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
14148 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3],ymm1[4,5],ymm0[6,7]
14149 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14150 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm4[1],xmm3[1]
14151 ; AVX-NEXT: vinsertf128 $1, %xmm5, %ymm0, %ymm1
14152 ; AVX-NEXT: vbroadcastsd 424(%r10), %ymm2
14153 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
14154 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7]
14155 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14156 ; AVX-NEXT: vmovaps 448(%rcx), %xmm0
14157 ; AVX-NEXT: vmovaps 448(%rsi), %xmm1
14158 ; AVX-NEXT: vmovaps 448(%rdi), %xmm2
14159 ; AVX-NEXT: vmovlhps {{.*#+}} xmm3 = xmm2[0],xmm1[0]
14160 ; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm3, %ymm4
14161 ; AVX-NEXT: vinsertf128 $1, 448(%rdx), %ymm3, %ymm3
14162 ; AVX-NEXT: vshufpd {{.*#+}} ymm3 = ymm3[0],ymm4[1],ymm3[2],ymm4[2]
14163 ; AVX-NEXT: vmovupd %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14164 ; AVX-NEXT: vmovaps 448(%r9), %xmm3
14165 ; AVX-NEXT: vmovaps 448(%r8), %xmm4
14166 ; AVX-NEXT: vmovaps 448(%rax), %xmm5
14167 ; AVX-NEXT: vmovlhps {{.*#+}} xmm6 = xmm4[0],xmm3[0]
14168 ; AVX-NEXT: vinsertf128 $1, %xmm5, %ymm6, %ymm7
14169 ; AVX-NEXT: vinsertf128 $1, 448(%r10), %ymm6, %ymm6
14170 ; AVX-NEXT: vshufpd {{.*#+}} ymm6 = ymm6[0],ymm7[1],ymm6[2],ymm7[2]
14171 ; AVX-NEXT: vmovupd %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14172 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm2[1],xmm1[1]
14173 ; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm1, %ymm0
14174 ; AVX-NEXT: vbroadcastsd 456(%rdx), %ymm2
14175 ; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
14176 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3],ymm1[4,5],ymm0[6,7]
14177 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14178 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm4[1],xmm3[1]
14179 ; AVX-NEXT: vinsertf128 $1, %xmm5, %ymm0, %ymm1
14180 ; AVX-NEXT: vbroadcastsd 456(%r10), %ymm2
14181 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
14182 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7]
14183 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14184 ; AVX-NEXT: vmovaps 480(%rcx), %xmm0
14185 ; AVX-NEXT: vmovaps 480(%rsi), %xmm1
14186 ; AVX-NEXT: vmovaps 480(%rdi), %xmm2
14187 ; AVX-NEXT: vmovlhps {{.*#+}} xmm3 = xmm2[0],xmm1[0]
14188 ; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm3, %ymm4
14189 ; AVX-NEXT: vinsertf128 $1, 480(%rdx), %ymm3, %ymm3
14190 ; AVX-NEXT: vshufpd {{.*#+}} ymm3 = ymm3[0],ymm4[1],ymm3[2],ymm4[2]
14191 ; AVX-NEXT: vmovupd %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14192 ; AVX-NEXT: vmovaps 480(%r9), %xmm3
14193 ; AVX-NEXT: vmovaps 480(%r8), %xmm4
14194 ; AVX-NEXT: vmovaps 480(%rax), %xmm5
14195 ; AVX-NEXT: vmovlhps {{.*#+}} xmm6 = xmm4[0],xmm3[0]
14196 ; AVX-NEXT: vinsertf128 $1, %xmm5, %ymm6, %ymm7
14197 ; AVX-NEXT: vinsertf128 $1, 480(%r10), %ymm6, %ymm6
14198 ; AVX-NEXT: vshufpd {{.*#+}} ymm6 = ymm6[0],ymm7[1],ymm6[2],ymm7[2]
14199 ; AVX-NEXT: vmovupd %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14200 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm2[1],xmm1[1]
14201 ; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm1, %ymm0
14202 ; AVX-NEXT: vbroadcastsd 488(%rdx), %ymm2
14203 ; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
14204 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3],ymm1[4,5],ymm0[6,7]
14205 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14206 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm4[1],xmm3[1]
14207 ; AVX-NEXT: vinsertf128 $1, %xmm5, %ymm0, %ymm1
14208 ; AVX-NEXT: vbroadcastsd 488(%r10), %ymm2
14209 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
14210 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7]
14211 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14212 ; AVX-NEXT: vmovaps 16(%rsi), %xmm0
14213 ; AVX-NEXT: vmovaps 16(%rdi), %xmm1
14214 ; AVX-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
14215 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],mem[4,5,6,7]
14216 ; AVX-NEXT: vbroadcastsd 16(%rcx), %ymm3
14217 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
14218 ; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14219 ; AVX-NEXT: vmovaps 16(%r9), %xmm2
14220 ; AVX-NEXT: vmovaps 16(%r8), %xmm3
14221 ; AVX-NEXT: vmovlhps {{.*#+}} xmm4 = xmm3[0],xmm2[0]
14222 ; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],mem[4,5,6,7]
14223 ; AVX-NEXT: vbroadcastsd 16(%rax), %ymm5
14224 ; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
14225 ; AVX-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14226 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
14227 ; AVX-NEXT: vbroadcastsd 24(%rdx), %ymm1
14228 ; AVX-NEXT: vblendps {{.*#+}} ymm6 = ymm0[0,1,2,3],ymm1[4,5,6,7]
14229 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm3[1],xmm2[1]
14230 ; AVX-NEXT: vbroadcastsd 24(%r10), %ymm1
14231 ; AVX-NEXT: vblendps {{.*#+}} ymm7 = ymm0[0,1,2,3],ymm1[4,5,6,7]
14232 ; AVX-NEXT: vmovaps 48(%rsi), %xmm0
14233 ; AVX-NEXT: vmovaps 48(%rdi), %xmm1
14234 ; AVX-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
14235 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],mem[4,5,6,7]
14236 ; AVX-NEXT: vbroadcastsd 48(%rcx), %ymm3
14237 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
14238 ; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14239 ; AVX-NEXT: vmovaps 48(%r9), %xmm2
14240 ; AVX-NEXT: vmovaps 48(%r8), %xmm3
14241 ; AVX-NEXT: vmovlhps {{.*#+}} xmm4 = xmm3[0],xmm2[0]
14242 ; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],mem[4,5,6,7]
14243 ; AVX-NEXT: vbroadcastsd 48(%rax), %ymm5
14244 ; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
14245 ; AVX-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14246 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
14247 ; AVX-NEXT: vbroadcastsd 56(%rdx), %ymm1
14248 ; AVX-NEXT: vblendps {{.*#+}} ymm9 = ymm0[0,1,2,3],ymm1[4,5,6,7]
14249 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm3[1],xmm2[1]
14250 ; AVX-NEXT: vbroadcastsd 56(%r10), %ymm1
14251 ; AVX-NEXT: vblendps {{.*#+}} ymm10 = ymm0[0,1,2,3],ymm1[4,5,6,7]
14252 ; AVX-NEXT: vmovaps 80(%rsi), %xmm0
14253 ; AVX-NEXT: vmovaps 80(%rdi), %xmm1
14254 ; AVX-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
14255 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],mem[4,5,6,7]
14256 ; AVX-NEXT: vbroadcastsd 80(%rcx), %ymm3
14257 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
14258 ; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14259 ; AVX-NEXT: vmovaps 80(%r9), %xmm2
14260 ; AVX-NEXT: vmovaps 80(%r8), %xmm3
14261 ; AVX-NEXT: vmovlhps {{.*#+}} xmm4 = xmm3[0],xmm2[0]
14262 ; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],mem[4,5,6,7]
14263 ; AVX-NEXT: vbroadcastsd 80(%rax), %ymm5
14264 ; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
14265 ; AVX-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14266 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
14267 ; AVX-NEXT: vbroadcastsd 88(%rdx), %ymm1
14268 ; AVX-NEXT: vblendps {{.*#+}} ymm14 = ymm0[0,1,2,3],ymm1[4,5,6,7]
14269 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm3[1],xmm2[1]
14270 ; AVX-NEXT: vbroadcastsd 88(%r10), %ymm1
14271 ; AVX-NEXT: vblendps {{.*#+}} ymm15 = ymm0[0,1,2,3],ymm1[4,5,6,7]
14272 ; AVX-NEXT: vmovaps 112(%rsi), %xmm0
14273 ; AVX-NEXT: vmovaps 112(%rdi), %xmm1
14274 ; AVX-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
14275 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],mem[4,5,6,7]
14276 ; AVX-NEXT: vbroadcastsd 112(%rcx), %ymm3
14277 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
14278 ; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14279 ; AVX-NEXT: vmovaps 112(%r9), %xmm2
14280 ; AVX-NEXT: vmovaps 112(%r8), %xmm3
14281 ; AVX-NEXT: vmovlhps {{.*#+}} xmm4 = xmm3[0],xmm2[0]
14282 ; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],mem[4,5,6,7]
14283 ; AVX-NEXT: vbroadcastsd 112(%rax), %ymm5
14284 ; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
14285 ; AVX-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14286 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
14287 ; AVX-NEXT: vbroadcastsd 120(%rdx), %ymm1
14288 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
14289 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14290 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm3[1],xmm2[1]
14291 ; AVX-NEXT: vbroadcastsd 120(%r10), %ymm1
14292 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
14293 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14294 ; AVX-NEXT: vmovaps 144(%rsi), %xmm0
14295 ; AVX-NEXT: vmovaps 144(%rdi), %xmm1
14296 ; AVX-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
14297 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],mem[4,5,6,7]
14298 ; AVX-NEXT: vbroadcastsd 144(%rcx), %ymm3
14299 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
14300 ; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14301 ; AVX-NEXT: vmovaps 144(%r9), %xmm2
14302 ; AVX-NEXT: vmovaps 144(%r8), %xmm3
14303 ; AVX-NEXT: vmovlhps {{.*#+}} xmm4 = xmm3[0],xmm2[0]
14304 ; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],mem[4,5,6,7]
14305 ; AVX-NEXT: vbroadcastsd 144(%rax), %ymm5
14306 ; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
14307 ; AVX-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14308 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
14309 ; AVX-NEXT: vbroadcastsd 152(%rdx), %ymm1
14310 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
14311 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14312 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm3[1],xmm2[1]
14313 ; AVX-NEXT: vbroadcastsd 152(%r10), %ymm1
14314 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
14315 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14316 ; AVX-NEXT: vmovaps 176(%rsi), %xmm0
14317 ; AVX-NEXT: vmovaps 176(%rdi), %xmm1
14318 ; AVX-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
14319 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],mem[4,5,6,7]
14320 ; AVX-NEXT: vbroadcastsd 176(%rcx), %ymm3
14321 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
14322 ; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14323 ; AVX-NEXT: vmovaps 176(%r9), %xmm2
14324 ; AVX-NEXT: vmovaps 176(%r8), %xmm3
14325 ; AVX-NEXT: vmovlhps {{.*#+}} xmm4 = xmm3[0],xmm2[0]
14326 ; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],mem[4,5,6,7]
14327 ; AVX-NEXT: vbroadcastsd 176(%rax), %ymm5
14328 ; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
14329 ; AVX-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14330 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
14331 ; AVX-NEXT: vbroadcastsd 184(%rdx), %ymm1
14332 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
14333 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14334 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm3[1],xmm2[1]
14335 ; AVX-NEXT: vbroadcastsd 184(%r10), %ymm1
14336 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
14337 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14338 ; AVX-NEXT: vmovaps 208(%rsi), %xmm0
14339 ; AVX-NEXT: vmovaps 208(%rdi), %xmm1
14340 ; AVX-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
14341 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],mem[4,5,6,7]
14342 ; AVX-NEXT: vbroadcastsd 208(%rcx), %ymm3
14343 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
14344 ; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14345 ; AVX-NEXT: vmovaps 208(%r9), %xmm2
14346 ; AVX-NEXT: vmovaps 208(%r8), %xmm3
14347 ; AVX-NEXT: vmovlhps {{.*#+}} xmm4 = xmm3[0],xmm2[0]
14348 ; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],mem[4,5,6,7]
14349 ; AVX-NEXT: vbroadcastsd 208(%rax), %ymm5
14350 ; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
14351 ; AVX-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14352 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
14353 ; AVX-NEXT: vbroadcastsd 216(%rdx), %ymm1
14354 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
14355 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14356 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm3[1],xmm2[1]
14357 ; AVX-NEXT: vbroadcastsd 216(%r10), %ymm1
14358 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
14359 ; AVX-NEXT: vmovups %ymm0, (%rsp) # 32-byte Spill
14360 ; AVX-NEXT: vmovaps 240(%rsi), %xmm0
14361 ; AVX-NEXT: vmovaps 240(%rdi), %xmm1
14362 ; AVX-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
14363 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],mem[4,5,6,7]
14364 ; AVX-NEXT: vbroadcastsd 240(%rcx), %ymm3
14365 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
14366 ; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14367 ; AVX-NEXT: vmovaps 240(%r9), %xmm2
14368 ; AVX-NEXT: vmovaps 240(%r8), %xmm3
14369 ; AVX-NEXT: vmovlhps {{.*#+}} xmm4 = xmm3[0],xmm2[0]
14370 ; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],mem[4,5,6,7]
14371 ; AVX-NEXT: vbroadcastsd 240(%rax), %ymm5
14372 ; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
14373 ; AVX-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14374 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
14375 ; AVX-NEXT: vbroadcastsd 248(%rdx), %ymm1
14376 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
14377 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14378 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm3[1],xmm2[1]
14379 ; AVX-NEXT: vbroadcastsd 248(%r10), %ymm1
14380 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
14381 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14382 ; AVX-NEXT: vmovaps 272(%rsi), %xmm0
14383 ; AVX-NEXT: vmovaps 272(%rdi), %xmm1
14384 ; AVX-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
14385 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],mem[4,5,6,7]
14386 ; AVX-NEXT: vbroadcastsd 272(%rcx), %ymm3
14387 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
14388 ; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14389 ; AVX-NEXT: vmovaps 272(%r9), %xmm2
14390 ; AVX-NEXT: vmovaps 272(%r8), %xmm3
14391 ; AVX-NEXT: vmovlhps {{.*#+}} xmm4 = xmm3[0],xmm2[0]
14392 ; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],mem[4,5,6,7]
14393 ; AVX-NEXT: vbroadcastsd 272(%rax), %ymm5
14394 ; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
14395 ; AVX-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14396 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
14397 ; AVX-NEXT: vbroadcastsd 280(%rdx), %ymm1
14398 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
14399 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14400 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm3[1],xmm2[1]
14401 ; AVX-NEXT: vbroadcastsd 280(%r10), %ymm1
14402 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
14403 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14404 ; AVX-NEXT: vmovaps 304(%rsi), %xmm0
14405 ; AVX-NEXT: vmovaps 304(%rdi), %xmm1
14406 ; AVX-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
14407 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],mem[4,5,6,7]
14408 ; AVX-NEXT: vbroadcastsd 304(%rcx), %ymm3
14409 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
14410 ; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14411 ; AVX-NEXT: vmovaps 304(%r9), %xmm2
14412 ; AVX-NEXT: vmovaps 304(%r8), %xmm3
14413 ; AVX-NEXT: vmovlhps {{.*#+}} xmm4 = xmm3[0],xmm2[0]
14414 ; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],mem[4,5,6,7]
14415 ; AVX-NEXT: vbroadcastsd 304(%rax), %ymm5
14416 ; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
14417 ; AVX-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14418 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
14419 ; AVX-NEXT: vbroadcastsd 312(%rdx), %ymm1
14420 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
14421 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14422 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm3[1],xmm2[1]
14423 ; AVX-NEXT: vbroadcastsd 312(%r10), %ymm1
14424 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
14425 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14426 ; AVX-NEXT: vmovaps 336(%rsi), %xmm0
14427 ; AVX-NEXT: vmovaps 336(%rdi), %xmm1
14428 ; AVX-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
14429 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],mem[4,5,6,7]
14430 ; AVX-NEXT: vbroadcastsd 336(%rcx), %ymm3
14431 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
14432 ; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14433 ; AVX-NEXT: vmovaps 336(%r9), %xmm2
14434 ; AVX-NEXT: vmovaps 336(%r8), %xmm3
14435 ; AVX-NEXT: vmovlhps {{.*#+}} xmm4 = xmm3[0],xmm2[0]
14436 ; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],mem[4,5,6,7]
14437 ; AVX-NEXT: vbroadcastsd 336(%rax), %ymm5
14438 ; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
14439 ; AVX-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14440 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
14441 ; AVX-NEXT: vbroadcastsd 344(%rdx), %ymm1
14442 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
14443 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14444 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm3[1],xmm2[1]
14445 ; AVX-NEXT: vbroadcastsd 344(%r10), %ymm1
14446 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
14447 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14448 ; AVX-NEXT: vmovaps 368(%rsi), %xmm0
14449 ; AVX-NEXT: vmovaps 368(%rdi), %xmm1
14450 ; AVX-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
14451 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],mem[4,5,6,7]
14452 ; AVX-NEXT: vbroadcastsd 368(%rcx), %ymm3
14453 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
14454 ; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14455 ; AVX-NEXT: vmovaps 368(%r9), %xmm2
14456 ; AVX-NEXT: vmovaps 368(%r8), %xmm3
14457 ; AVX-NEXT: vmovlhps {{.*#+}} xmm4 = xmm3[0],xmm2[0]
14458 ; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],mem[4,5,6,7]
14459 ; AVX-NEXT: vbroadcastsd 368(%rax), %ymm5
14460 ; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
14461 ; AVX-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14462 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
14463 ; AVX-NEXT: vbroadcastsd 376(%rdx), %ymm1
14464 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
14465 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14466 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm3[1],xmm2[1]
14467 ; AVX-NEXT: vbroadcastsd 376(%r10), %ymm1
14468 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
14469 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14470 ; AVX-NEXT: vmovaps 400(%rsi), %xmm0
14471 ; AVX-NEXT: vmovaps 400(%rdi), %xmm1
14472 ; AVX-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
14473 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],mem[4,5,6,7]
14474 ; AVX-NEXT: vbroadcastsd 400(%rcx), %ymm3
14475 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
14476 ; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14477 ; AVX-NEXT: vmovaps 400(%r9), %xmm2
14478 ; AVX-NEXT: vmovaps 400(%r8), %xmm3
14479 ; AVX-NEXT: vmovlhps {{.*#+}} xmm4 = xmm3[0],xmm2[0]
14480 ; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],mem[4,5,6,7]
14481 ; AVX-NEXT: vbroadcastsd 400(%rax), %ymm5
14482 ; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
14483 ; AVX-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14484 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
14485 ; AVX-NEXT: vbroadcastsd 408(%rdx), %ymm1
14486 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
14487 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14488 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm3[1],xmm2[1]
14489 ; AVX-NEXT: vbroadcastsd 408(%r10), %ymm1
14490 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
14491 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14492 ; AVX-NEXT: vmovaps 432(%rsi), %xmm0
14493 ; AVX-NEXT: vmovaps 432(%rdi), %xmm1
14494 ; AVX-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
14495 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],mem[4,5,6,7]
14496 ; AVX-NEXT: vbroadcastsd 432(%rcx), %ymm3
14497 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
14498 ; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14499 ; AVX-NEXT: vmovaps 432(%r9), %xmm2
14500 ; AVX-NEXT: vmovaps 432(%r8), %xmm3
14501 ; AVX-NEXT: vmovlhps {{.*#+}} xmm4 = xmm3[0],xmm2[0]
14502 ; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],mem[4,5,6,7]
14503 ; AVX-NEXT: vbroadcastsd 432(%rax), %ymm5
14504 ; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
14505 ; AVX-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14506 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
14507 ; AVX-NEXT: vbroadcastsd 440(%rdx), %ymm1
14508 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
14509 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14510 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm3[1],xmm2[1]
14511 ; AVX-NEXT: vbroadcastsd 440(%r10), %ymm1
14512 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
14513 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14514 ; AVX-NEXT: vmovaps 464(%rsi), %xmm0
14515 ; AVX-NEXT: vmovaps 464(%rdi), %xmm1
14516 ; AVX-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
14517 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],mem[4,5,6,7]
14518 ; AVX-NEXT: vbroadcastsd 464(%rcx), %ymm3
14519 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
14520 ; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14521 ; AVX-NEXT: vmovaps 464(%r9), %xmm2
14522 ; AVX-NEXT: vmovaps 464(%r8), %xmm3
14523 ; AVX-NEXT: vmovlhps {{.*#+}} xmm4 = xmm3[0],xmm2[0]
14524 ; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],mem[4,5,6,7]
14525 ; AVX-NEXT: vbroadcastsd 464(%rax), %ymm5
14526 ; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
14527 ; AVX-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14528 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
14529 ; AVX-NEXT: vbroadcastsd 472(%rdx), %ymm1
14530 ; AVX-NEXT: vblendps {{.*#+}} ymm13 = ymm0[0,1,2,3],ymm1[4,5,6,7]
14531 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm3[1],xmm2[1]
14532 ; AVX-NEXT: vbroadcastsd 472(%r10), %ymm1
14533 ; AVX-NEXT: vblendps {{.*#+}} ymm12 = ymm0[0,1,2,3],ymm1[4,5,6,7]
14534 ; AVX-NEXT: vmovaps 496(%rsi), %xmm0
14535 ; AVX-NEXT: vmovaps 496(%rdi), %xmm1
14536 ; AVX-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]
14537 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],mem[4,5,6,7]
14538 ; AVX-NEXT: vbroadcastsd 496(%rcx), %ymm3
14539 ; AVX-NEXT: vblendps {{.*#+}} ymm11 = ymm2[0,1,2,3,4,5],ymm3[6,7]
14540 ; AVX-NEXT: vmovaps 496(%r9), %xmm3
14541 ; AVX-NEXT: vmovaps 496(%r8), %xmm4
14542 ; AVX-NEXT: vmovlhps {{.*#+}} xmm2 = xmm4[0],xmm3[0]
14543 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],mem[4,5,6,7]
14544 ; AVX-NEXT: vbroadcastsd 496(%rax), %ymm5
14545 ; AVX-NEXT: vblendps {{.*#+}} ymm8 = ymm2[0,1,2,3,4,5],ymm5[6,7]
14546 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
14547 ; AVX-NEXT: vbroadcastsd 504(%rdx), %ymm1
14548 ; AVX-NEXT: vblendps {{.*#+}} ymm5 = ymm0[0,1,2,3],ymm1[4,5,6,7]
14549 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm4[1],xmm3[1]
14550 ; AVX-NEXT: vbroadcastsd 504(%r10), %ymm1
14551 ; AVX-NEXT: vblendps {{.*#+}} ymm3 = ymm0[0,1,2,3],ymm1[4,5,6,7]
14552 ; AVX-NEXT: movq {{[0-9]+}}(%rsp), %rdx
14553 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm6[0,1,2,3,4,5],mem[6,7]
14554 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14555 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm7[0,1,2,3,4,5],mem[6,7]
14556 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14557 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm9[0,1,2,3,4,5],mem[6,7]
14558 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14559 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm10[0,1,2,3,4,5],mem[6,7]
14560 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14561 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm14[0,1,2,3,4,5],mem[6,7]
14562 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14563 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3,4,5],mem[6,7]
14564 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14565 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14566 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],mem[6,7]
14567 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14568 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14569 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],mem[6,7]
14570 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14571 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14572 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],mem[6,7]
14573 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14574 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14575 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],mem[6,7]
14576 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14577 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14578 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],mem[6,7]
14579 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14580 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14581 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],mem[6,7]
14582 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14583 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14584 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],mem[6,7]
14585 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14586 ; AVX-NEXT: vmovups (%rsp), %ymm0 # 32-byte Reload
14587 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],mem[6,7]
14588 ; AVX-NEXT: vmovups %ymm0, (%rsp) # 32-byte Spill
14589 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14590 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],mem[6,7]
14591 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14592 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14593 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],mem[6,7]
14594 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14595 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14596 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],mem[6,7]
14597 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14598 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14599 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],mem[6,7]
14600 ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14601 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14602 ; AVX-NEXT: vblendps {{.*#+}} ymm15 = ymm0[0,1,2,3,4,5],mem[6,7]
14603 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14604 ; AVX-NEXT: vblendps {{.*#+}} ymm14 = ymm0[0,1,2,3,4,5],mem[6,7]
14605 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14606 ; AVX-NEXT: vblendps {{.*#+}} ymm10 = ymm0[0,1,2,3,4,5],mem[6,7]
14607 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14608 ; AVX-NEXT: vblendps {{.*#+}} ymm9 = ymm0[0,1,2,3,4,5],mem[6,7]
14609 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14610 ; AVX-NEXT: vblendps {{.*#+}} ymm7 = ymm0[0,1,2,3,4,5],mem[6,7]
14611 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14612 ; AVX-NEXT: vblendps {{.*#+}} ymm6 = ymm0[0,1,2,3,4,5],mem[6,7]
14613 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14614 ; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm0[0,1,2,3,4,5],mem[6,7]
14615 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14616 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm0[0,1,2,3,4,5],mem[6,7]
14617 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14618 ; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm0[0,1,2,3,4,5],mem[6,7]
14619 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14620 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],mem[6,7]
14621 ; AVX-NEXT: vblendps {{.*#+}} ymm13 = ymm13[0,1,2,3,4,5],mem[6,7]
14622 ; AVX-NEXT: vblendps {{.*#+}} ymm12 = ymm12[0,1,2,3,4,5],mem[6,7]
14623 ; AVX-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3,4,5],mem[6,7]
14624 ; AVX-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],mem[6,7]
14625 ; AVX-NEXT: vmovaps %ymm3, 4064(%rdx)
14626 ; AVX-NEXT: vmovaps %ymm5, 4032(%rdx)
14627 ; AVX-NEXT: vmovaps %ymm8, 4000(%rdx)
14628 ; AVX-NEXT: vmovaps %ymm11, 3968(%rdx)
14629 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
14630 ; AVX-NEXT: vmovaps %ymm3, 3936(%rdx)
14631 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
14632 ; AVX-NEXT: vmovaps %ymm3, 3904(%rdx)
14633 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
14634 ; AVX-NEXT: vmovaps %ymm3, 3872(%rdx)
14635 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
14636 ; AVX-NEXT: vmovaps %ymm3, 3840(%rdx)
14637 ; AVX-NEXT: vmovaps %ymm12, 3808(%rdx)
14638 ; AVX-NEXT: vmovaps %ymm13, 3776(%rdx)
14639 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
14640 ; AVX-NEXT: vmovaps %ymm3, 3744(%rdx)
14641 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
14642 ; AVX-NEXT: vmovaps %ymm3, 3712(%rdx)
14643 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
14644 ; AVX-NEXT: vmovaps %ymm3, 3680(%rdx)
14645 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
14646 ; AVX-NEXT: vmovaps %ymm3, 3648(%rdx)
14647 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
14648 ; AVX-NEXT: vmovaps %ymm3, 3616(%rdx)
14649 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
14650 ; AVX-NEXT: vmovaps %ymm3, 3584(%rdx)
14651 ; AVX-NEXT: vmovaps %ymm0, 3552(%rdx)
14652 ; AVX-NEXT: vmovaps %ymm1, 3520(%rdx)
14653 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14654 ; AVX-NEXT: vmovaps %ymm0, 3488(%rdx)
14655 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14656 ; AVX-NEXT: vmovaps %ymm0, 3456(%rdx)
14657 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14658 ; AVX-NEXT: vmovaps %ymm0, 3424(%rdx)
14659 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14660 ; AVX-NEXT: vmovaps %ymm0, 3392(%rdx)
14661 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14662 ; AVX-NEXT: vmovaps %ymm0, 3360(%rdx)
14663 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14664 ; AVX-NEXT: vmovaps %ymm0, 3328(%rdx)
14665 ; AVX-NEXT: vmovaps %ymm2, 3296(%rdx)
14666 ; AVX-NEXT: vmovaps %ymm4, 3264(%rdx)
14667 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14668 ; AVX-NEXT: vmovaps %ymm0, 3232(%rdx)
14669 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14670 ; AVX-NEXT: vmovaps %ymm0, 3200(%rdx)
14671 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14672 ; AVX-NEXT: vmovaps %ymm0, 3168(%rdx)
14673 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14674 ; AVX-NEXT: vmovaps %ymm0, 3136(%rdx)
14675 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14676 ; AVX-NEXT: vmovaps %ymm0, 3104(%rdx)
14677 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14678 ; AVX-NEXT: vmovaps %ymm0, 3072(%rdx)
14679 ; AVX-NEXT: vmovaps %ymm6, 3040(%rdx)
14680 ; AVX-NEXT: vmovaps %ymm7, 3008(%rdx)
14681 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14682 ; AVX-NEXT: vmovaps %ymm0, 2976(%rdx)
14683 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14684 ; AVX-NEXT: vmovaps %ymm0, 2944(%rdx)
14685 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14686 ; AVX-NEXT: vmovaps %ymm0, 2912(%rdx)
14687 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14688 ; AVX-NEXT: vmovaps %ymm0, 2880(%rdx)
14689 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14690 ; AVX-NEXT: vmovaps %ymm0, 2848(%rdx)
14691 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14692 ; AVX-NEXT: vmovaps %ymm0, 2816(%rdx)
14693 ; AVX-NEXT: vmovaps %ymm9, 2784(%rdx)
14694 ; AVX-NEXT: vmovaps %ymm10, 2752(%rdx)
14695 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14696 ; AVX-NEXT: vmovaps %ymm0, 2720(%rdx)
14697 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14698 ; AVX-NEXT: vmovaps %ymm0, 2688(%rdx)
14699 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14700 ; AVX-NEXT: vmovaps %ymm0, 2656(%rdx)
14701 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14702 ; AVX-NEXT: vmovaps %ymm0, 2624(%rdx)
14703 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14704 ; AVX-NEXT: vmovaps %ymm0, 2592(%rdx)
14705 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14706 ; AVX-NEXT: vmovaps %ymm0, 2560(%rdx)
14707 ; AVX-NEXT: vmovaps %ymm14, 2528(%rdx)
14708 ; AVX-NEXT: vmovaps %ymm15, 2496(%rdx)
14709 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14710 ; AVX-NEXT: vmovaps %ymm0, 2464(%rdx)
14711 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14712 ; AVX-NEXT: vmovaps %ymm0, 2432(%rdx)
14713 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14714 ; AVX-NEXT: vmovaps %ymm0, 2400(%rdx)
14715 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14716 ; AVX-NEXT: vmovaps %ymm0, 2368(%rdx)
14717 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14718 ; AVX-NEXT: vmovaps %ymm0, 2336(%rdx)
14719 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14720 ; AVX-NEXT: vmovaps %ymm0, 2304(%rdx)
14721 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14722 ; AVX-NEXT: vmovaps %ymm0, 2272(%rdx)
14723 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14724 ; AVX-NEXT: vmovaps %ymm0, 2240(%rdx)
14725 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14726 ; AVX-NEXT: vmovaps %ymm0, 2208(%rdx)
14727 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14728 ; AVX-NEXT: vmovaps %ymm0, 2176(%rdx)
14729 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14730 ; AVX-NEXT: vmovaps %ymm0, 2144(%rdx)
14731 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14732 ; AVX-NEXT: vmovaps %ymm0, 2112(%rdx)
14733 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14734 ; AVX-NEXT: vmovaps %ymm0, 2080(%rdx)
14735 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14736 ; AVX-NEXT: vmovaps %ymm0, 2048(%rdx)
14737 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14738 ; AVX-NEXT: vmovaps %ymm0, 2016(%rdx)
14739 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14740 ; AVX-NEXT: vmovaps %ymm0, 1984(%rdx)
14741 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14742 ; AVX-NEXT: vmovaps %ymm0, 1952(%rdx)
14743 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14744 ; AVX-NEXT: vmovaps %ymm0, 1920(%rdx)
14745 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14746 ; AVX-NEXT: vmovaps %ymm0, 1888(%rdx)
14747 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14748 ; AVX-NEXT: vmovaps %ymm0, 1856(%rdx)
14749 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14750 ; AVX-NEXT: vmovaps %ymm0, 1824(%rdx)
14751 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14752 ; AVX-NEXT: vmovaps %ymm0, 1792(%rdx)
14753 ; AVX-NEXT: vmovups (%rsp), %ymm0 # 32-byte Reload
14754 ; AVX-NEXT: vmovaps %ymm0, 1760(%rdx)
14755 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14756 ; AVX-NEXT: vmovaps %ymm0, 1728(%rdx)
14757 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14758 ; AVX-NEXT: vmovaps %ymm0, 1696(%rdx)
14759 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14760 ; AVX-NEXT: vmovaps %ymm0, 1664(%rdx)
14761 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14762 ; AVX-NEXT: vmovaps %ymm0, 1632(%rdx)
14763 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14764 ; AVX-NEXT: vmovaps %ymm0, 1600(%rdx)
14765 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14766 ; AVX-NEXT: vmovaps %ymm0, 1568(%rdx)
14767 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14768 ; AVX-NEXT: vmovaps %ymm0, 1536(%rdx)
14769 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14770 ; AVX-NEXT: vmovaps %ymm0, 1504(%rdx)
14771 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14772 ; AVX-NEXT: vmovaps %ymm0, 1472(%rdx)
14773 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14774 ; AVX-NEXT: vmovaps %ymm0, 1440(%rdx)
14775 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14776 ; AVX-NEXT: vmovaps %ymm0, 1408(%rdx)
14777 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14778 ; AVX-NEXT: vmovaps %ymm0, 1376(%rdx)
14779 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14780 ; AVX-NEXT: vmovaps %ymm0, 1344(%rdx)
14781 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14782 ; AVX-NEXT: vmovaps %ymm0, 1312(%rdx)
14783 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14784 ; AVX-NEXT: vmovaps %ymm0, 1280(%rdx)
14785 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14786 ; AVX-NEXT: vmovaps %ymm0, 1248(%rdx)
14787 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14788 ; AVX-NEXT: vmovaps %ymm0, 1216(%rdx)
14789 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14790 ; AVX-NEXT: vmovaps %ymm0, 1184(%rdx)
14791 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14792 ; AVX-NEXT: vmovaps %ymm0, 1152(%rdx)
14793 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14794 ; AVX-NEXT: vmovaps %ymm0, 1120(%rdx)
14795 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14796 ; AVX-NEXT: vmovaps %ymm0, 1088(%rdx)
14797 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14798 ; AVX-NEXT: vmovaps %ymm0, 1056(%rdx)
14799 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14800 ; AVX-NEXT: vmovaps %ymm0, 1024(%rdx)
14801 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14802 ; AVX-NEXT: vmovaps %ymm0, 992(%rdx)
14803 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14804 ; AVX-NEXT: vmovaps %ymm0, 960(%rdx)
14805 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14806 ; AVX-NEXT: vmovaps %ymm0, 928(%rdx)
14807 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14808 ; AVX-NEXT: vmovaps %ymm0, 896(%rdx)
14809 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14810 ; AVX-NEXT: vmovaps %ymm0, 864(%rdx)
14811 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14812 ; AVX-NEXT: vmovaps %ymm0, 832(%rdx)
14813 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14814 ; AVX-NEXT: vmovaps %ymm0, 800(%rdx)
14815 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14816 ; AVX-NEXT: vmovaps %ymm0, 768(%rdx)
14817 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14818 ; AVX-NEXT: vmovaps %ymm0, 736(%rdx)
14819 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14820 ; AVX-NEXT: vmovaps %ymm0, 704(%rdx)
14821 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14822 ; AVX-NEXT: vmovaps %ymm0, 672(%rdx)
14823 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14824 ; AVX-NEXT: vmovaps %ymm0, 640(%rdx)
14825 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14826 ; AVX-NEXT: vmovaps %ymm0, 608(%rdx)
14827 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14828 ; AVX-NEXT: vmovaps %ymm0, 576(%rdx)
14829 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14830 ; AVX-NEXT: vmovaps %ymm0, 544(%rdx)
14831 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14832 ; AVX-NEXT: vmovaps %ymm0, 512(%rdx)
14833 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14834 ; AVX-NEXT: vmovaps %ymm0, 480(%rdx)
14835 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14836 ; AVX-NEXT: vmovaps %ymm0, 448(%rdx)
14837 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14838 ; AVX-NEXT: vmovaps %ymm0, 416(%rdx)
14839 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14840 ; AVX-NEXT: vmovaps %ymm0, 384(%rdx)
14841 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14842 ; AVX-NEXT: vmovaps %ymm0, 352(%rdx)
14843 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14844 ; AVX-NEXT: vmovaps %ymm0, 320(%rdx)
14845 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14846 ; AVX-NEXT: vmovaps %ymm0, 288(%rdx)
14847 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14848 ; AVX-NEXT: vmovaps %ymm0, 256(%rdx)
14849 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14850 ; AVX-NEXT: vmovaps %ymm0, 224(%rdx)
14851 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14852 ; AVX-NEXT: vmovaps %ymm0, 192(%rdx)
14853 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14854 ; AVX-NEXT: vmovaps %ymm0, 160(%rdx)
14855 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14856 ; AVX-NEXT: vmovaps %ymm0, 128(%rdx)
14857 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14858 ; AVX-NEXT: vmovaps %ymm0, 96(%rdx)
14859 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14860 ; AVX-NEXT: vmovaps %ymm0, 64(%rdx)
14861 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14862 ; AVX-NEXT: vmovaps %ymm0, 32(%rdx)
14863 ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
14864 ; AVX-NEXT: vmovaps %ymm0, (%rdx)
14865 ; AVX-NEXT: addq $3784, %rsp # imm = 0xEC8
14866 ; AVX-NEXT: vzeroupper
14867 ; AVX-NEXT: retq
14868 ;
14869 ; AVX2-LABEL: store_i64_stride8_vf64:
14870 ; AVX2: # %bb.0:
14871 ; AVX2-NEXT: subq $3880, %rsp # imm = 0xF28
14872 ; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %r10
14873 ; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %rax
14874 ; AVX2-NEXT: vmovaps (%rcx), %xmm0
14875 ; AVX2-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
14876 ; AVX2-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0
14877 ; AVX2-NEXT: vmovaps (%rsi), %xmm2
14878 ; AVX2-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
14879 ; AVX2-NEXT: vmovaps 32(%rsi), %xmm3
14880 ; AVX2-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
14881 ; AVX2-NEXT: vmovaps (%rdi), %xmm1
14882 ; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
14883 ; AVX2-NEXT: vmovaps 32(%rdi), %xmm4
14884 ; AVX2-NEXT: vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
14885 ; AVX2-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm2[1]
14886 ; AVX2-NEXT: vbroadcastsd 8(%rdx), %ymm2
14887 ; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
14888 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
14889 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14890 ; AVX2-NEXT: vmovaps (%rax), %xmm0
14891 ; AVX2-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
14892 ; AVX2-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0
14893 ; AVX2-NEXT: vmovaps (%r9), %xmm2
14894 ; AVX2-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
14895 ; AVX2-NEXT: vmovaps (%r8), %xmm1
14896 ; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
14897 ; AVX2-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm2[1]
14898 ; AVX2-NEXT: vbroadcastsd 8(%r10), %ymm2
14899 ; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
14900 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
14901 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14902 ; AVX2-NEXT: vmovaps 32(%rcx), %xmm0
14903 ; AVX2-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
14904 ; AVX2-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0
14905 ; AVX2-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm4[1],xmm3[1]
14906 ; AVX2-NEXT: vbroadcastsd 40(%rdx), %ymm2
14907 ; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
14908 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
14909 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14910 ; AVX2-NEXT: vmovaps 32(%r9), %xmm1
14911 ; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
14912 ; AVX2-NEXT: vmovaps 32(%r8), %xmm0
14913 ; AVX2-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
14914 ; AVX2-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
14915 ; AVX2-NEXT: vbroadcastsd 40(%r10), %ymm1
14916 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
14917 ; AVX2-NEXT: vmovaps 32(%rax), %xmm1
14918 ; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
14919 ; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
14920 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
14921 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14922 ; AVX2-NEXT: vmovaps 64(%rsi), %xmm1
14923 ; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
14924 ; AVX2-NEXT: vmovaps 64(%rdi), %xmm0
14925 ; AVX2-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
14926 ; AVX2-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
14927 ; AVX2-NEXT: vbroadcastsd 72(%rdx), %ymm1
14928 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
14929 ; AVX2-NEXT: vmovaps 64(%rcx), %xmm1
14930 ; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
14931 ; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
14932 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
14933 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14934 ; AVX2-NEXT: vmovaps 64(%r9), %xmm1
14935 ; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
14936 ; AVX2-NEXT: vmovaps 64(%r8), %xmm0
14937 ; AVX2-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
14938 ; AVX2-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
14939 ; AVX2-NEXT: vbroadcastsd 72(%r10), %ymm1
14940 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
14941 ; AVX2-NEXT: vmovaps 64(%rax), %xmm1
14942 ; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
14943 ; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
14944 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
14945 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14946 ; AVX2-NEXT: vmovaps 96(%rsi), %xmm1
14947 ; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
14948 ; AVX2-NEXT: vmovaps 96(%rdi), %xmm0
14949 ; AVX2-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
14950 ; AVX2-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
14951 ; AVX2-NEXT: vbroadcastsd 104(%rdx), %ymm1
14952 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
14953 ; AVX2-NEXT: vmovaps 96(%rcx), %xmm1
14954 ; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
14955 ; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
14956 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
14957 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14958 ; AVX2-NEXT: vmovaps 96(%r9), %xmm1
14959 ; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
14960 ; AVX2-NEXT: vmovaps 96(%r8), %xmm0
14961 ; AVX2-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
14962 ; AVX2-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
14963 ; AVX2-NEXT: vbroadcastsd 104(%r10), %ymm1
14964 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
14965 ; AVX2-NEXT: vmovaps 96(%rax), %xmm1
14966 ; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
14967 ; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
14968 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
14969 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14970 ; AVX2-NEXT: vmovaps 128(%rsi), %xmm1
14971 ; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
14972 ; AVX2-NEXT: vmovaps 128(%rdi), %xmm0
14973 ; AVX2-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
14974 ; AVX2-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
14975 ; AVX2-NEXT: vbroadcastsd 136(%rdx), %ymm1
14976 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
14977 ; AVX2-NEXT: vmovaps 128(%rcx), %xmm1
14978 ; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
14979 ; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
14980 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
14981 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14982 ; AVX2-NEXT: vmovaps 128(%r9), %xmm1
14983 ; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
14984 ; AVX2-NEXT: vmovaps 128(%r8), %xmm0
14985 ; AVX2-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
14986 ; AVX2-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
14987 ; AVX2-NEXT: vbroadcastsd 136(%r10), %ymm1
14988 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
14989 ; AVX2-NEXT: vmovaps 128(%rax), %xmm1
14990 ; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
14991 ; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
14992 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
14993 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
14994 ; AVX2-NEXT: vmovaps 160(%rsi), %xmm1
14995 ; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
14996 ; AVX2-NEXT: vmovaps 160(%rdi), %xmm0
14997 ; AVX2-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
14998 ; AVX2-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
14999 ; AVX2-NEXT: vbroadcastsd 168(%rdx), %ymm1
15000 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
15001 ; AVX2-NEXT: vmovaps 160(%rcx), %xmm1
15002 ; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
15003 ; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
15004 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
15005 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15006 ; AVX2-NEXT: vmovaps 160(%r9), %xmm1
15007 ; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
15008 ; AVX2-NEXT: vmovaps 160(%r8), %xmm0
15009 ; AVX2-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
15010 ; AVX2-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
15011 ; AVX2-NEXT: vbroadcastsd 168(%r10), %ymm1
15012 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
15013 ; AVX2-NEXT: vmovaps 160(%rax), %xmm1
15014 ; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
15015 ; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
15016 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
15017 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15018 ; AVX2-NEXT: vmovaps 192(%rsi), %xmm1
15019 ; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
15020 ; AVX2-NEXT: vmovaps 192(%rdi), %xmm0
15021 ; AVX2-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
15022 ; AVX2-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
15023 ; AVX2-NEXT: vbroadcastsd 200(%rdx), %ymm1
15024 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
15025 ; AVX2-NEXT: vmovaps 192(%rcx), %xmm1
15026 ; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
15027 ; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
15028 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
15029 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15030 ; AVX2-NEXT: vmovaps 192(%r9), %xmm1
15031 ; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
15032 ; AVX2-NEXT: vmovaps 192(%r8), %xmm0
15033 ; AVX2-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
15034 ; AVX2-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
15035 ; AVX2-NEXT: vbroadcastsd 200(%r10), %ymm1
15036 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
15037 ; AVX2-NEXT: vmovaps 192(%rax), %xmm1
15038 ; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
15039 ; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
15040 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
15041 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15042 ; AVX2-NEXT: vmovaps 224(%rsi), %xmm1
15043 ; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
15044 ; AVX2-NEXT: vmovaps 224(%rdi), %xmm0
15045 ; AVX2-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
15046 ; AVX2-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
15047 ; AVX2-NEXT: vbroadcastsd 232(%rdx), %ymm1
15048 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
15049 ; AVX2-NEXT: vmovaps 224(%rcx), %xmm1
15050 ; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
15051 ; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
15052 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
15053 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15054 ; AVX2-NEXT: vmovaps 224(%r9), %xmm1
15055 ; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
15056 ; AVX2-NEXT: vmovaps 224(%r8), %xmm0
15057 ; AVX2-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
15058 ; AVX2-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
15059 ; AVX2-NEXT: vbroadcastsd 232(%r10), %ymm1
15060 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
15061 ; AVX2-NEXT: vmovaps 224(%rax), %xmm1
15062 ; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
15063 ; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
15064 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
15065 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15066 ; AVX2-NEXT: vmovaps 256(%rsi), %xmm1
15067 ; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
15068 ; AVX2-NEXT: vmovaps 256(%rdi), %xmm0
15069 ; AVX2-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
15070 ; AVX2-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
15071 ; AVX2-NEXT: vbroadcastsd 264(%rdx), %ymm1
15072 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
15073 ; AVX2-NEXT: vmovaps 256(%rcx), %xmm1
15074 ; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
15075 ; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
15076 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
15077 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15078 ; AVX2-NEXT: vmovaps 256(%r9), %xmm1
15079 ; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
15080 ; AVX2-NEXT: vmovaps 256(%r8), %xmm0
15081 ; AVX2-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
15082 ; AVX2-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
15083 ; AVX2-NEXT: vbroadcastsd 264(%r10), %ymm1
15084 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
15085 ; AVX2-NEXT: vmovaps 256(%rax), %xmm1
15086 ; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
15087 ; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
15088 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
15089 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15090 ; AVX2-NEXT: vmovaps 288(%rsi), %xmm1
15091 ; AVX2-NEXT: vmovaps %xmm1, (%rsp) # 16-byte Spill
15092 ; AVX2-NEXT: vmovaps 288(%rdi), %xmm0
15093 ; AVX2-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
15094 ; AVX2-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
15095 ; AVX2-NEXT: vbroadcastsd 296(%rdx), %ymm1
15096 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
15097 ; AVX2-NEXT: vmovaps 288(%rcx), %xmm1
15098 ; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
15099 ; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
15100 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
15101 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15102 ; AVX2-NEXT: vmovaps 288(%r9), %xmm1
15103 ; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
15104 ; AVX2-NEXT: vmovaps 288(%r8), %xmm0
15105 ; AVX2-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
15106 ; AVX2-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
15107 ; AVX2-NEXT: vbroadcastsd 296(%r10), %ymm1
15108 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
15109 ; AVX2-NEXT: vmovaps 288(%rax), %xmm1
15110 ; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
15111 ; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
15112 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
15113 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15114 ; AVX2-NEXT: vmovaps 320(%rsi), %xmm1
15115 ; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
15116 ; AVX2-NEXT: vmovaps 320(%rdi), %xmm0
15117 ; AVX2-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
15118 ; AVX2-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
15119 ; AVX2-NEXT: vbroadcastsd 328(%rdx), %ymm1
15120 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
15121 ; AVX2-NEXT: vmovaps 320(%rcx), %xmm1
15122 ; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
15123 ; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
15124 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
15125 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15126 ; AVX2-NEXT: vmovaps 320(%r9), %xmm1
15127 ; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
15128 ; AVX2-NEXT: vmovaps 320(%r8), %xmm0
15129 ; AVX2-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
15130 ; AVX2-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
15131 ; AVX2-NEXT: vbroadcastsd 328(%r10), %ymm1
15132 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
15133 ; AVX2-NEXT: vmovaps 320(%rax), %xmm1
15134 ; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
15135 ; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
15136 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
15137 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15138 ; AVX2-NEXT: vmovaps 352(%rsi), %xmm1
15139 ; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
15140 ; AVX2-NEXT: vmovaps 352(%rdi), %xmm0
15141 ; AVX2-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
15142 ; AVX2-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
15143 ; AVX2-NEXT: vbroadcastsd 360(%rdx), %ymm1
15144 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
15145 ; AVX2-NEXT: vmovaps 352(%rcx), %xmm1
15146 ; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
15147 ; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
15148 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
15149 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15150 ; AVX2-NEXT: vmovaps 352(%r9), %xmm1
15151 ; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
15152 ; AVX2-NEXT: vmovaps 352(%r8), %xmm0
15153 ; AVX2-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
15154 ; AVX2-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
15155 ; AVX2-NEXT: vbroadcastsd 360(%r10), %ymm1
15156 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
15157 ; AVX2-NEXT: vmovaps 352(%rax), %xmm1
15158 ; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
15159 ; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
15160 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
15161 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15162 ; AVX2-NEXT: vmovaps 384(%rsi), %xmm1
15163 ; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
15164 ; AVX2-NEXT: vmovaps 384(%rdi), %xmm0
15165 ; AVX2-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
15166 ; AVX2-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
15167 ; AVX2-NEXT: vbroadcastsd 392(%rdx), %ymm1
15168 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
15169 ; AVX2-NEXT: vmovaps 384(%rcx), %xmm1
15170 ; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
15171 ; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
15172 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
15173 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15174 ; AVX2-NEXT: vmovaps 384(%r9), %xmm1
15175 ; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
15176 ; AVX2-NEXT: vmovaps 384(%r8), %xmm0
15177 ; AVX2-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
15178 ; AVX2-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
15179 ; AVX2-NEXT: vbroadcastsd 392(%r10), %ymm1
15180 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
15181 ; AVX2-NEXT: vmovaps 384(%rax), %xmm1
15182 ; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
15183 ; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
15184 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
15185 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15186 ; AVX2-NEXT: vmovaps 416(%rsi), %xmm1
15187 ; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
15188 ; AVX2-NEXT: vmovaps 416(%rdi), %xmm0
15189 ; AVX2-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
15190 ; AVX2-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
15191 ; AVX2-NEXT: vbroadcastsd 424(%rdx), %ymm1
15192 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
15193 ; AVX2-NEXT: vmovaps 416(%rcx), %xmm1
15194 ; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
15195 ; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
15196 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
15197 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15198 ; AVX2-NEXT: vmovaps 416(%r9), %xmm0
15199 ; AVX2-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
15200 ; AVX2-NEXT: vmovaps 416(%r8), %xmm13
15201 ; AVX2-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm13[1],xmm0[1]
15202 ; AVX2-NEXT: vbroadcastsd 424(%r10), %ymm1
15203 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
15204 ; AVX2-NEXT: vmovaps 416(%rax), %xmm12
15205 ; AVX2-NEXT: vinsertf128 $1, %xmm12, %ymm0, %ymm1
15206 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
15207 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15208 ; AVX2-NEXT: vmovaps 448(%rsi), %xmm11
15209 ; AVX2-NEXT: vmovaps 448(%rdi), %xmm10
15210 ; AVX2-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm10[1],xmm11[1]
15211 ; AVX2-NEXT: vbroadcastsd 456(%rdx), %ymm1
15212 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
15213 ; AVX2-NEXT: vmovaps 448(%rcx), %xmm9
15214 ; AVX2-NEXT: vinsertf128 $1, %xmm9, %ymm0, %ymm1
15215 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
15216 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15217 ; AVX2-NEXT: vmovaps 448(%r9), %xmm8
15218 ; AVX2-NEXT: vmovaps 448(%r8), %xmm7
15219 ; AVX2-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm7[1],xmm8[1]
15220 ; AVX2-NEXT: vbroadcastsd 456(%r10), %ymm1
15221 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
15222 ; AVX2-NEXT: vmovaps 448(%rax), %xmm6
15223 ; AVX2-NEXT: vinsertf128 $1, %xmm6, %ymm0, %ymm1
15224 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
15225 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15226 ; AVX2-NEXT: vmovaps 480(%rsi), %xmm5
15227 ; AVX2-NEXT: vmovaps 480(%rdi), %xmm4
15228 ; AVX2-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm4[1],xmm5[1]
15229 ; AVX2-NEXT: vbroadcastsd 488(%rdx), %ymm1
15230 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
15231 ; AVX2-NEXT: vmovaps 480(%rcx), %xmm3
15232 ; AVX2-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm1
15233 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
15234 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15235 ; AVX2-NEXT: vmovaps 480(%r9), %xmm2
15236 ; AVX2-NEXT: vmovaps 480(%r8), %xmm1
15237 ; AVX2-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm2[1]
15238 ; AVX2-NEXT: vbroadcastsd 488(%r10), %ymm15
15239 ; AVX2-NEXT: vblendps {{.*#+}} ymm15 = ymm0[0,1,2,3],ymm15[4,5,6,7]
15240 ; AVX2-NEXT: vmovaps 480(%rax), %xmm0
15241 ; AVX2-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm14
15242 ; AVX2-NEXT: vblendps {{.*#+}} ymm14 = ymm15[0,1,2,3,4,5],ymm14[6,7]
15243 ; AVX2-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15244 ; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
15245 ; AVX2-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
15246 ; AVX2-NEXT: # xmm14 = xmm14[0],mem[0]
15247 ; AVX2-NEXT: vinsertf128 $1, (%rdx), %ymm14, %ymm14
15248 ; AVX2-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
15249 ; AVX2-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
15250 ; AVX2-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15251 ; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
15252 ; AVX2-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
15253 ; AVX2-NEXT: # xmm14 = xmm14[0],mem[0]
15254 ; AVX2-NEXT: vinsertf128 $1, (%r10), %ymm14, %ymm14
15255 ; AVX2-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
15256 ; AVX2-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
15257 ; AVX2-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15258 ; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
15259 ; AVX2-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
15260 ; AVX2-NEXT: # xmm14 = xmm14[0],mem[0]
15261 ; AVX2-NEXT: vinsertf128 $1, 32(%rdx), %ymm14, %ymm14
15262 ; AVX2-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
15263 ; AVX2-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
15264 ; AVX2-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15265 ; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
15266 ; AVX2-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
15267 ; AVX2-NEXT: # xmm14 = xmm14[0],mem[0]
15268 ; AVX2-NEXT: vinsertf128 $1, 32(%r10), %ymm14, %ymm14
15269 ; AVX2-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
15270 ; AVX2-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
15271 ; AVX2-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15272 ; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
15273 ; AVX2-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
15274 ; AVX2-NEXT: # xmm14 = xmm14[0],mem[0]
15275 ; AVX2-NEXT: vinsertf128 $1, 64(%rdx), %ymm14, %ymm14
15276 ; AVX2-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
15277 ; AVX2-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
15278 ; AVX2-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15279 ; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
15280 ; AVX2-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
15281 ; AVX2-NEXT: # xmm14 = xmm14[0],mem[0]
15282 ; AVX2-NEXT: vinsertf128 $1, 64(%r10), %ymm14, %ymm14
15283 ; AVX2-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
15284 ; AVX2-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
15285 ; AVX2-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15286 ; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
15287 ; AVX2-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
15288 ; AVX2-NEXT: # xmm14 = xmm14[0],mem[0]
15289 ; AVX2-NEXT: vinsertf128 $1, 96(%rdx), %ymm14, %ymm14
15290 ; AVX2-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
15291 ; AVX2-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
15292 ; AVX2-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15293 ; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
15294 ; AVX2-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
15295 ; AVX2-NEXT: # xmm14 = xmm14[0],mem[0]
15296 ; AVX2-NEXT: vinsertf128 $1, 96(%r10), %ymm14, %ymm14
15297 ; AVX2-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
15298 ; AVX2-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
15299 ; AVX2-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15300 ; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
15301 ; AVX2-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
15302 ; AVX2-NEXT: # xmm14 = xmm14[0],mem[0]
15303 ; AVX2-NEXT: vinsertf128 $1, 128(%rdx), %ymm14, %ymm14
15304 ; AVX2-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
15305 ; AVX2-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
15306 ; AVX2-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15307 ; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
15308 ; AVX2-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
15309 ; AVX2-NEXT: # xmm14 = xmm14[0],mem[0]
15310 ; AVX2-NEXT: vinsertf128 $1, 128(%r10), %ymm14, %ymm14
15311 ; AVX2-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
15312 ; AVX2-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
15313 ; AVX2-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15314 ; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
15315 ; AVX2-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
15316 ; AVX2-NEXT: # xmm14 = xmm14[0],mem[0]
15317 ; AVX2-NEXT: vinsertf128 $1, 160(%rdx), %ymm14, %ymm14
15318 ; AVX2-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
15319 ; AVX2-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
15320 ; AVX2-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15321 ; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
15322 ; AVX2-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
15323 ; AVX2-NEXT: # xmm14 = xmm14[0],mem[0]
15324 ; AVX2-NEXT: vinsertf128 $1, 160(%r10), %ymm14, %ymm14
15325 ; AVX2-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
15326 ; AVX2-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
15327 ; AVX2-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15328 ; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
15329 ; AVX2-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
15330 ; AVX2-NEXT: # xmm14 = xmm14[0],mem[0]
15331 ; AVX2-NEXT: vinsertf128 $1, 192(%rdx), %ymm14, %ymm14
15332 ; AVX2-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
15333 ; AVX2-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
15334 ; AVX2-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15335 ; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
15336 ; AVX2-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
15337 ; AVX2-NEXT: # xmm14 = xmm14[0],mem[0]
15338 ; AVX2-NEXT: vinsertf128 $1, 192(%r10), %ymm14, %ymm14
15339 ; AVX2-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
15340 ; AVX2-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
15341 ; AVX2-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15342 ; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
15343 ; AVX2-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
15344 ; AVX2-NEXT: # xmm14 = xmm14[0],mem[0]
15345 ; AVX2-NEXT: vinsertf128 $1, 224(%rdx), %ymm14, %ymm14
15346 ; AVX2-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
15347 ; AVX2-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
15348 ; AVX2-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15349 ; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
15350 ; AVX2-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
15351 ; AVX2-NEXT: # xmm14 = xmm14[0],mem[0]
15352 ; AVX2-NEXT: vinsertf128 $1, 224(%r10), %ymm14, %ymm14
15353 ; AVX2-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
15354 ; AVX2-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
15355 ; AVX2-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15356 ; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
15357 ; AVX2-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
15358 ; AVX2-NEXT: # xmm14 = xmm14[0],mem[0]
15359 ; AVX2-NEXT: vinsertf128 $1, 256(%rdx), %ymm14, %ymm14
15360 ; AVX2-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
15361 ; AVX2-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
15362 ; AVX2-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15363 ; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
15364 ; AVX2-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
15365 ; AVX2-NEXT: # xmm14 = xmm14[0],mem[0]
15366 ; AVX2-NEXT: vinsertf128 $1, 256(%r10), %ymm14, %ymm14
15367 ; AVX2-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
15368 ; AVX2-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
15369 ; AVX2-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15370 ; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
15371 ; AVX2-NEXT: vunpcklpd (%rsp), %xmm14, %xmm14 # 16-byte Folded Reload
15372 ; AVX2-NEXT: # xmm14 = xmm14[0],mem[0]
15373 ; AVX2-NEXT: vinsertf128 $1, 288(%rdx), %ymm14, %ymm14
15374 ; AVX2-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
15375 ; AVX2-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
15376 ; AVX2-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15377 ; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
15378 ; AVX2-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
15379 ; AVX2-NEXT: # xmm14 = xmm14[0],mem[0]
15380 ; AVX2-NEXT: vinsertf128 $1, 288(%r10), %ymm14, %ymm14
15381 ; AVX2-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
15382 ; AVX2-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
15383 ; AVX2-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15384 ; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
15385 ; AVX2-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
15386 ; AVX2-NEXT: # xmm14 = xmm14[0],mem[0]
15387 ; AVX2-NEXT: vinsertf128 $1, 320(%rdx), %ymm14, %ymm14
15388 ; AVX2-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
15389 ; AVX2-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
15390 ; AVX2-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15391 ; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
15392 ; AVX2-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
15393 ; AVX2-NEXT: # xmm14 = xmm14[0],mem[0]
15394 ; AVX2-NEXT: vinsertf128 $1, 320(%r10), %ymm14, %ymm14
15395 ; AVX2-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
15396 ; AVX2-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
15397 ; AVX2-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15398 ; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
15399 ; AVX2-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
15400 ; AVX2-NEXT: # xmm14 = xmm14[0],mem[0]
15401 ; AVX2-NEXT: vinsertf128 $1, 352(%rdx), %ymm14, %ymm14
15402 ; AVX2-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
15403 ; AVX2-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
15404 ; AVX2-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15405 ; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
15406 ; AVX2-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
15407 ; AVX2-NEXT: # xmm14 = xmm14[0],mem[0]
15408 ; AVX2-NEXT: vinsertf128 $1, 352(%r10), %ymm14, %ymm14
15409 ; AVX2-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
15410 ; AVX2-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
15411 ; AVX2-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15412 ; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
15413 ; AVX2-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
15414 ; AVX2-NEXT: # xmm14 = xmm14[0],mem[0]
15415 ; AVX2-NEXT: vinsertf128 $1, 384(%rdx), %ymm14, %ymm14
15416 ; AVX2-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
15417 ; AVX2-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
15418 ; AVX2-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15419 ; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
15420 ; AVX2-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
15421 ; AVX2-NEXT: # xmm14 = xmm14[0],mem[0]
15422 ; AVX2-NEXT: vinsertf128 $1, 384(%r10), %ymm14, %ymm14
15423 ; AVX2-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
15424 ; AVX2-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
15425 ; AVX2-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15426 ; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
15427 ; AVX2-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
15428 ; AVX2-NEXT: # xmm14 = xmm14[0],mem[0]
15429 ; AVX2-NEXT: vinsertf128 $1, 416(%rdx), %ymm14, %ymm14
15430 ; AVX2-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
15431 ; AVX2-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
15432 ; AVX2-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15433 ; AVX2-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm13, %xmm13 # 16-byte Folded Reload
15434 ; AVX2-NEXT: # xmm13 = xmm13[0],mem[0]
15435 ; AVX2-NEXT: vinsertf128 $1, 416(%r10), %ymm13, %ymm13
15436 ; AVX2-NEXT: vbroadcastsd %xmm12, %ymm12
15437 ; AVX2-NEXT: vblendps {{.*#+}} ymm12 = ymm13[0,1,2,3,4,5],ymm12[6,7]
15438 ; AVX2-NEXT: vmovups %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15439 ; AVX2-NEXT: vmovlhps {{.*#+}} xmm10 = xmm10[0],xmm11[0]
15440 ; AVX2-NEXT: vinsertf128 $1, 448(%rdx), %ymm10, %ymm10
15441 ; AVX2-NEXT: vbroadcastsd %xmm9, %ymm9
15442 ; AVX2-NEXT: vblendps {{.*#+}} ymm9 = ymm10[0,1,2,3,4,5],ymm9[6,7]
15443 ; AVX2-NEXT: vmovups %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15444 ; AVX2-NEXT: vmovlhps {{.*#+}} xmm7 = xmm7[0],xmm8[0]
15445 ; AVX2-NEXT: vinsertf128 $1, 448(%r10), %ymm7, %ymm7
15446 ; AVX2-NEXT: vbroadcastsd %xmm6, %ymm6
15447 ; AVX2-NEXT: vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3,4,5],ymm6[6,7]
15448 ; AVX2-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15449 ; AVX2-NEXT: vmovlhps {{.*#+}} xmm4 = xmm4[0],xmm5[0]
15450 ; AVX2-NEXT: vinsertf128 $1, 480(%rdx), %ymm4, %ymm4
15451 ; AVX2-NEXT: vbroadcastsd %xmm3, %ymm3
15452 ; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm4[0,1,2,3,4,5],ymm3[6,7]
15453 ; AVX2-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15454 ; AVX2-NEXT: vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
15455 ; AVX2-NEXT: vinsertf128 $1, 480(%r10), %ymm1, %ymm1
15456 ; AVX2-NEXT: vbroadcastsd %xmm0, %ymm0
15457 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
15458 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15459 ; AVX2-NEXT: vmovaps (%rdi), %ymm0
15460 ; AVX2-NEXT: vmovaps (%rsi), %ymm1
15461 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
15462 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
15463 ; AVX2-NEXT: vbroadcastsd 16(%rcx), %ymm3
15464 ; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
15465 ; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15466 ; AVX2-NEXT: vmovaps (%r8), %ymm2
15467 ; AVX2-NEXT: vmovaps (%r9), %ymm3
15468 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
15469 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],mem[2,3]
15470 ; AVX2-NEXT: vbroadcastsd 16(%rax), %ymm5
15471 ; AVX2-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
15472 ; AVX2-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15473 ; AVX2-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
15474 ; AVX2-NEXT: vbroadcastsd 24(%rdx), %ymm1
15475 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm6 = ymm0[2,3],ymm1[2,3]
15476 ; AVX2-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
15477 ; AVX2-NEXT: vbroadcastsd 24(%r10), %ymm1
15478 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm7 = ymm0[2,3],ymm1[2,3]
15479 ; AVX2-NEXT: vmovaps 32(%rdi), %ymm0
15480 ; AVX2-NEXT: vmovaps 32(%rsi), %ymm1
15481 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
15482 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
15483 ; AVX2-NEXT: vbroadcastsd 48(%rcx), %ymm3
15484 ; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
15485 ; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15486 ; AVX2-NEXT: vmovaps 32(%r8), %ymm2
15487 ; AVX2-NEXT: vmovaps 32(%r9), %ymm3
15488 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
15489 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],mem[2,3]
15490 ; AVX2-NEXT: vbroadcastsd 48(%rax), %ymm5
15491 ; AVX2-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
15492 ; AVX2-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15493 ; AVX2-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
15494 ; AVX2-NEXT: vbroadcastsd 56(%rdx), %ymm1
15495 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm8 = ymm0[2,3],ymm1[2,3]
15496 ; AVX2-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
15497 ; AVX2-NEXT: vbroadcastsd 56(%r10), %ymm1
15498 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm10 = ymm0[2,3],ymm1[2,3]
15499 ; AVX2-NEXT: vmovaps 64(%rdi), %ymm0
15500 ; AVX2-NEXT: vmovaps 64(%rsi), %ymm1
15501 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
15502 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
15503 ; AVX2-NEXT: vbroadcastsd 80(%rcx), %ymm3
15504 ; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
15505 ; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15506 ; AVX2-NEXT: vmovaps 64(%r8), %ymm2
15507 ; AVX2-NEXT: vmovaps 64(%r9), %ymm3
15508 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
15509 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],mem[2,3]
15510 ; AVX2-NEXT: vbroadcastsd 80(%rax), %ymm5
15511 ; AVX2-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
15512 ; AVX2-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15513 ; AVX2-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
15514 ; AVX2-NEXT: vbroadcastsd 88(%rdx), %ymm1
15515 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm12 = ymm0[2,3],ymm1[2,3]
15516 ; AVX2-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
15517 ; AVX2-NEXT: vbroadcastsd 88(%r10), %ymm1
15518 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm13 = ymm0[2,3],ymm1[2,3]
15519 ; AVX2-NEXT: vmovaps 96(%rdi), %ymm0
15520 ; AVX2-NEXT: vmovaps 96(%rsi), %ymm1
15521 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
15522 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
15523 ; AVX2-NEXT: vbroadcastsd 112(%rcx), %ymm3
15524 ; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
15525 ; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15526 ; AVX2-NEXT: vmovaps 96(%r8), %ymm2
15527 ; AVX2-NEXT: vmovaps 96(%r9), %ymm3
15528 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
15529 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],mem[2,3]
15530 ; AVX2-NEXT: vbroadcastsd 112(%rax), %ymm5
15531 ; AVX2-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
15532 ; AVX2-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15533 ; AVX2-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
15534 ; AVX2-NEXT: vbroadcastsd 120(%rdx), %ymm1
15535 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm14 = ymm0[2,3],ymm1[2,3]
15536 ; AVX2-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
15537 ; AVX2-NEXT: vbroadcastsd 120(%r10), %ymm1
15538 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
15539 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15540 ; AVX2-NEXT: vmovaps 128(%rdi), %ymm0
15541 ; AVX2-NEXT: vmovaps 128(%rsi), %ymm1
15542 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
15543 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
15544 ; AVX2-NEXT: vbroadcastsd 144(%rcx), %ymm3
15545 ; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
15546 ; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15547 ; AVX2-NEXT: vmovaps 128(%r8), %ymm2
15548 ; AVX2-NEXT: vmovaps 128(%r9), %ymm3
15549 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
15550 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],mem[2,3]
15551 ; AVX2-NEXT: vbroadcastsd 144(%rax), %ymm5
15552 ; AVX2-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
15553 ; AVX2-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15554 ; AVX2-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
15555 ; AVX2-NEXT: vbroadcastsd 152(%rdx), %ymm1
15556 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
15557 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15558 ; AVX2-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
15559 ; AVX2-NEXT: vbroadcastsd 152(%r10), %ymm1
15560 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
15561 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15562 ; AVX2-NEXT: vmovaps 160(%rdi), %ymm0
15563 ; AVX2-NEXT: vmovaps 160(%rsi), %ymm1
15564 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
15565 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
15566 ; AVX2-NEXT: vbroadcastsd 176(%rcx), %ymm3
15567 ; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
15568 ; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15569 ; AVX2-NEXT: vmovaps 160(%r8), %ymm2
15570 ; AVX2-NEXT: vmovaps 160(%r9), %ymm3
15571 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
15572 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],mem[2,3]
15573 ; AVX2-NEXT: vbroadcastsd 176(%rax), %ymm5
15574 ; AVX2-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
15575 ; AVX2-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15576 ; AVX2-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
15577 ; AVX2-NEXT: vbroadcastsd 184(%rdx), %ymm1
15578 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
15579 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15580 ; AVX2-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
15581 ; AVX2-NEXT: vbroadcastsd 184(%r10), %ymm1
15582 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
15583 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15584 ; AVX2-NEXT: vmovaps 192(%rdi), %ymm0
15585 ; AVX2-NEXT: vmovaps 192(%rsi), %ymm1
15586 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
15587 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
15588 ; AVX2-NEXT: vbroadcastsd 208(%rcx), %ymm3
15589 ; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
15590 ; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15591 ; AVX2-NEXT: vmovaps 192(%r8), %ymm2
15592 ; AVX2-NEXT: vmovaps 192(%r9), %ymm3
15593 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
15594 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],mem[2,3]
15595 ; AVX2-NEXT: vbroadcastsd 208(%rax), %ymm5
15596 ; AVX2-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
15597 ; AVX2-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15598 ; AVX2-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
15599 ; AVX2-NEXT: vbroadcastsd 216(%rdx), %ymm1
15600 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
15601 ; AVX2-NEXT: vmovups %ymm0, (%rsp) # 32-byte Spill
15602 ; AVX2-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
15603 ; AVX2-NEXT: vbroadcastsd 216(%r10), %ymm1
15604 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
15605 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15606 ; AVX2-NEXT: vmovaps 224(%rdi), %ymm0
15607 ; AVX2-NEXT: vmovaps 224(%rsi), %ymm1
15608 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
15609 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
15610 ; AVX2-NEXT: vbroadcastsd 240(%rcx), %ymm3
15611 ; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
15612 ; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15613 ; AVX2-NEXT: vmovaps 224(%r8), %ymm2
15614 ; AVX2-NEXT: vmovaps 224(%r9), %ymm3
15615 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
15616 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],mem[2,3]
15617 ; AVX2-NEXT: vbroadcastsd 240(%rax), %ymm5
15618 ; AVX2-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
15619 ; AVX2-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15620 ; AVX2-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
15621 ; AVX2-NEXT: vbroadcastsd 248(%rdx), %ymm1
15622 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
15623 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15624 ; AVX2-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
15625 ; AVX2-NEXT: vbroadcastsd 248(%r10), %ymm1
15626 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
15627 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15628 ; AVX2-NEXT: vmovaps 256(%rdi), %ymm0
15629 ; AVX2-NEXT: vmovaps 256(%rsi), %ymm1
15630 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
15631 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
15632 ; AVX2-NEXT: vbroadcastsd 272(%rcx), %ymm3
15633 ; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
15634 ; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15635 ; AVX2-NEXT: vmovaps 256(%r8), %ymm2
15636 ; AVX2-NEXT: vmovaps 256(%r9), %ymm3
15637 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
15638 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],mem[2,3]
15639 ; AVX2-NEXT: vbroadcastsd 272(%rax), %ymm5
15640 ; AVX2-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
15641 ; AVX2-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15642 ; AVX2-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
15643 ; AVX2-NEXT: vbroadcastsd 280(%rdx), %ymm1
15644 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
15645 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15646 ; AVX2-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
15647 ; AVX2-NEXT: vbroadcastsd 280(%r10), %ymm1
15648 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
15649 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15650 ; AVX2-NEXT: vmovaps 288(%rdi), %ymm0
15651 ; AVX2-NEXT: vmovaps 288(%rsi), %ymm1
15652 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
15653 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
15654 ; AVX2-NEXT: vbroadcastsd 304(%rcx), %ymm3
15655 ; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
15656 ; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15657 ; AVX2-NEXT: vmovaps 288(%r8), %ymm2
15658 ; AVX2-NEXT: vmovaps 288(%r9), %ymm3
15659 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
15660 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],mem[2,3]
15661 ; AVX2-NEXT: vbroadcastsd 304(%rax), %ymm5
15662 ; AVX2-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
15663 ; AVX2-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15664 ; AVX2-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
15665 ; AVX2-NEXT: vbroadcastsd 312(%rdx), %ymm1
15666 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
15667 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15668 ; AVX2-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
15669 ; AVX2-NEXT: vbroadcastsd 312(%r10), %ymm1
15670 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
15671 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15672 ; AVX2-NEXT: vmovaps 320(%rdi), %ymm0
15673 ; AVX2-NEXT: vmovaps 320(%rsi), %ymm1
15674 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
15675 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
15676 ; AVX2-NEXT: vbroadcastsd 336(%rcx), %ymm3
15677 ; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
15678 ; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15679 ; AVX2-NEXT: vmovaps 320(%r8), %ymm2
15680 ; AVX2-NEXT: vmovaps 320(%r9), %ymm3
15681 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
15682 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],mem[2,3]
15683 ; AVX2-NEXT: vbroadcastsd 336(%rax), %ymm5
15684 ; AVX2-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
15685 ; AVX2-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15686 ; AVX2-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
15687 ; AVX2-NEXT: vbroadcastsd 344(%rdx), %ymm1
15688 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
15689 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15690 ; AVX2-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
15691 ; AVX2-NEXT: vbroadcastsd 344(%r10), %ymm1
15692 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
15693 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15694 ; AVX2-NEXT: vmovaps 352(%rdi), %ymm0
15695 ; AVX2-NEXT: vmovaps 352(%rsi), %ymm1
15696 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
15697 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
15698 ; AVX2-NEXT: vbroadcastsd 368(%rcx), %ymm3
15699 ; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
15700 ; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15701 ; AVX2-NEXT: vmovaps 352(%r8), %ymm2
15702 ; AVX2-NEXT: vmovaps 352(%r9), %ymm3
15703 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
15704 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],mem[2,3]
15705 ; AVX2-NEXT: vbroadcastsd 368(%rax), %ymm5
15706 ; AVX2-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
15707 ; AVX2-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15708 ; AVX2-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
15709 ; AVX2-NEXT: vbroadcastsd 376(%rdx), %ymm1
15710 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
15711 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15712 ; AVX2-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
15713 ; AVX2-NEXT: vbroadcastsd 376(%r10), %ymm1
15714 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
15715 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15716 ; AVX2-NEXT: vmovaps 384(%rdi), %ymm0
15717 ; AVX2-NEXT: vmovaps 384(%rsi), %ymm1
15718 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
15719 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
15720 ; AVX2-NEXT: vbroadcastsd 400(%rcx), %ymm3
15721 ; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
15722 ; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15723 ; AVX2-NEXT: vmovaps 384(%r8), %ymm2
15724 ; AVX2-NEXT: vmovaps 384(%r9), %ymm3
15725 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
15726 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],mem[2,3]
15727 ; AVX2-NEXT: vbroadcastsd 400(%rax), %ymm5
15728 ; AVX2-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
15729 ; AVX2-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15730 ; AVX2-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
15731 ; AVX2-NEXT: vbroadcastsd 408(%rdx), %ymm1
15732 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
15733 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15734 ; AVX2-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
15735 ; AVX2-NEXT: vbroadcastsd 408(%r10), %ymm1
15736 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
15737 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15738 ; AVX2-NEXT: vmovaps 416(%rdi), %ymm0
15739 ; AVX2-NEXT: vmovaps 416(%rsi), %ymm1
15740 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
15741 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
15742 ; AVX2-NEXT: vbroadcastsd 432(%rcx), %ymm3
15743 ; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
15744 ; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15745 ; AVX2-NEXT: vmovaps 416(%r8), %ymm2
15746 ; AVX2-NEXT: vmovaps 416(%r9), %ymm3
15747 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
15748 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],mem[2,3]
15749 ; AVX2-NEXT: vbroadcastsd 432(%rax), %ymm5
15750 ; AVX2-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
15751 ; AVX2-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15752 ; AVX2-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
15753 ; AVX2-NEXT: vbroadcastsd 440(%rdx), %ymm1
15754 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
15755 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15756 ; AVX2-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
15757 ; AVX2-NEXT: vbroadcastsd 440(%r10), %ymm1
15758 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
15759 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15760 ; AVX2-NEXT: vmovaps 448(%rdi), %ymm0
15761 ; AVX2-NEXT: vmovaps 448(%rsi), %ymm1
15762 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
15763 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
15764 ; AVX2-NEXT: vbroadcastsd 464(%rcx), %ymm3
15765 ; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
15766 ; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15767 ; AVX2-NEXT: vmovaps 448(%r8), %ymm2
15768 ; AVX2-NEXT: vmovaps 448(%r9), %ymm3
15769 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
15770 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],mem[2,3]
15771 ; AVX2-NEXT: vbroadcastsd 464(%rax), %ymm5
15772 ; AVX2-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
15773 ; AVX2-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15774 ; AVX2-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
15775 ; AVX2-NEXT: vbroadcastsd 472(%rdx), %ymm1
15776 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
15777 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15778 ; AVX2-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
15779 ; AVX2-NEXT: vbroadcastsd 472(%r10), %ymm1
15780 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
15781 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15782 ; AVX2-NEXT: vmovaps 480(%rdi), %ymm0
15783 ; AVX2-NEXT: vmovaps 480(%rsi), %ymm1
15784 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
15785 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
15786 ; AVX2-NEXT: vbroadcastsd 496(%rcx), %ymm3
15787 ; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
15788 ; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15789 ; AVX2-NEXT: vmovaps 480(%r8), %ymm3
15790 ; AVX2-NEXT: vmovaps 480(%r9), %ymm4
15791 ; AVX2-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm3[0],ymm4[0],ymm3[2],ymm4[2]
15792 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
15793 ; AVX2-NEXT: vbroadcastsd 496(%rax), %ymm5
15794 ; AVX2-NEXT: vblendps {{.*#+}} ymm15 = ymm2[0,1,2,3,4,5],ymm5[6,7]
15795 ; AVX2-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
15796 ; AVX2-NEXT: vbroadcastsd 504(%rdx), %ymm1
15797 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm11 = ymm0[2,3],ymm1[2,3]
15798 ; AVX2-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm3[1],ymm4[1],ymm3[3],ymm4[3]
15799 ; AVX2-NEXT: vbroadcastsd 504(%r10), %ymm1
15800 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm9 = ymm0[2,3],ymm1[2,3]
15801 ; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %rdx
15802 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm6[0,1,2,3,4,5],mem[6,7]
15803 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15804 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm7[0,1,2,3,4,5],mem[6,7]
15805 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15806 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm8[0,1,2,3,4,5],mem[6,7]
15807 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15808 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm10[0,1,2,3,4,5],mem[6,7]
15809 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15810 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm12[0,1,2,3,4,5],mem[6,7]
15811 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15812 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm13[0,1,2,3,4,5],mem[6,7]
15813 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15814 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm14[0,1,2,3,4,5],mem[6,7]
15815 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15816 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15817 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],mem[6,7]
15818 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15819 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15820 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],mem[6,7]
15821 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15822 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15823 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],mem[6,7]
15824 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15825 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15826 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],mem[6,7]
15827 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15828 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15829 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],mem[6,7]
15830 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15831 ; AVX2-NEXT: vmovups (%rsp), %ymm0 # 32-byte Reload
15832 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],mem[6,7]
15833 ; AVX2-NEXT: vmovups %ymm0, (%rsp) # 32-byte Spill
15834 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15835 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],mem[6,7]
15836 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15837 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15838 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],mem[6,7]
15839 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15840 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15841 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],mem[6,7]
15842 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15843 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15844 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],mem[6,7]
15845 ; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
15846 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15847 ; AVX2-NEXT: vblendps {{.*#+}} ymm14 = ymm0[0,1,2,3,4,5],mem[6,7]
15848 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15849 ; AVX2-NEXT: vblendps {{.*#+}} ymm13 = ymm0[0,1,2,3,4,5],mem[6,7]
15850 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15851 ; AVX2-NEXT: vblendps {{.*#+}} ymm12 = ymm0[0,1,2,3,4,5],mem[6,7]
15852 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15853 ; AVX2-NEXT: vblendps {{.*#+}} ymm10 = ymm0[0,1,2,3,4,5],mem[6,7]
15854 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15855 ; AVX2-NEXT: vblendps {{.*#+}} ymm8 = ymm0[0,1,2,3,4,5],mem[6,7]
15856 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15857 ; AVX2-NEXT: vblendps {{.*#+}} ymm7 = ymm0[0,1,2,3,4,5],mem[6,7]
15858 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15859 ; AVX2-NEXT: vblendps {{.*#+}} ymm6 = ymm0[0,1,2,3,4,5],mem[6,7]
15860 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15861 ; AVX2-NEXT: vblendps {{.*#+}} ymm5 = ymm0[0,1,2,3,4,5],mem[6,7]
15862 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15863 ; AVX2-NEXT: vblendps {{.*#+}} ymm4 = ymm0[0,1,2,3,4,5],mem[6,7]
15864 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15865 ; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm0[0,1,2,3,4,5],mem[6,7]
15866 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15867 ; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm0[0,1,2,3,4,5],mem[6,7]
15868 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15869 ; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm0[0,1,2,3,4,5],mem[6,7]
15870 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15871 ; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],mem[6,7]
15872 ; AVX2-NEXT: vblendps {{.*#+}} ymm11 = ymm11[0,1,2,3,4,5],mem[6,7]
15873 ; AVX2-NEXT: vblendps {{.*#+}} ymm9 = ymm9[0,1,2,3,4,5],mem[6,7]
15874 ; AVX2-NEXT: vmovaps %ymm9, 4064(%rdx)
15875 ; AVX2-NEXT: vmovaps %ymm11, 4032(%rdx)
15876 ; AVX2-NEXT: vmovaps %ymm15, 4000(%rdx)
15877 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload
15878 ; AVX2-NEXT: vmovaps %ymm9, 3968(%rdx)
15879 ; AVX2-NEXT: vmovaps %ymm0, 3808(%rdx)
15880 ; AVX2-NEXT: vmovaps %ymm1, 3776(%rdx)
15881 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15882 ; AVX2-NEXT: vmovaps %ymm0, 3744(%rdx)
15883 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15884 ; AVX2-NEXT: vmovaps %ymm0, 3712(%rdx)
15885 ; AVX2-NEXT: vmovaps %ymm2, 3552(%rdx)
15886 ; AVX2-NEXT: vmovaps %ymm3, 3520(%rdx)
15887 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15888 ; AVX2-NEXT: vmovaps %ymm0, 3488(%rdx)
15889 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15890 ; AVX2-NEXT: vmovaps %ymm0, 3456(%rdx)
15891 ; AVX2-NEXT: vmovaps %ymm4, 3296(%rdx)
15892 ; AVX2-NEXT: vmovaps %ymm5, 3264(%rdx)
15893 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15894 ; AVX2-NEXT: vmovaps %ymm0, 3232(%rdx)
15895 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15896 ; AVX2-NEXT: vmovaps %ymm0, 3200(%rdx)
15897 ; AVX2-NEXT: vmovaps %ymm6, 3040(%rdx)
15898 ; AVX2-NEXT: vmovaps %ymm7, 3008(%rdx)
15899 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15900 ; AVX2-NEXT: vmovaps %ymm0, 2976(%rdx)
15901 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15902 ; AVX2-NEXT: vmovaps %ymm0, 2944(%rdx)
15903 ; AVX2-NEXT: vmovaps %ymm8, 2784(%rdx)
15904 ; AVX2-NEXT: vmovaps %ymm10, 2752(%rdx)
15905 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15906 ; AVX2-NEXT: vmovaps %ymm0, 2720(%rdx)
15907 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15908 ; AVX2-NEXT: vmovaps %ymm0, 2688(%rdx)
15909 ; AVX2-NEXT: vmovaps %ymm12, 2528(%rdx)
15910 ; AVX2-NEXT: vmovaps %ymm13, 2496(%rdx)
15911 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15912 ; AVX2-NEXT: vmovaps %ymm0, 2464(%rdx)
15913 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15914 ; AVX2-NEXT: vmovaps %ymm0, 2432(%rdx)
15915 ; AVX2-NEXT: vmovaps %ymm14, 2272(%rdx)
15916 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15917 ; AVX2-NEXT: vmovaps %ymm0, 2240(%rdx)
15918 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15919 ; AVX2-NEXT: vmovaps %ymm0, 2208(%rdx)
15920 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15921 ; AVX2-NEXT: vmovaps %ymm0, 2176(%rdx)
15922 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15923 ; AVX2-NEXT: vmovaps %ymm0, 2016(%rdx)
15924 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15925 ; AVX2-NEXT: vmovaps %ymm0, 1984(%rdx)
15926 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15927 ; AVX2-NEXT: vmovaps %ymm0, 1952(%rdx)
15928 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15929 ; AVX2-NEXT: vmovaps %ymm0, 1920(%rdx)
15930 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15931 ; AVX2-NEXT: vmovaps %ymm0, 1760(%rdx)
15932 ; AVX2-NEXT: vmovups (%rsp), %ymm0 # 32-byte Reload
15933 ; AVX2-NEXT: vmovaps %ymm0, 1728(%rdx)
15934 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15935 ; AVX2-NEXT: vmovaps %ymm0, 1696(%rdx)
15936 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15937 ; AVX2-NEXT: vmovaps %ymm0, 1664(%rdx)
15938 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15939 ; AVX2-NEXT: vmovaps %ymm0, 1504(%rdx)
15940 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15941 ; AVX2-NEXT: vmovaps %ymm0, 1472(%rdx)
15942 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15943 ; AVX2-NEXT: vmovaps %ymm0, 1440(%rdx)
15944 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15945 ; AVX2-NEXT: vmovaps %ymm0, 1408(%rdx)
15946 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15947 ; AVX2-NEXT: vmovaps %ymm0, 1248(%rdx)
15948 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15949 ; AVX2-NEXT: vmovaps %ymm0, 1216(%rdx)
15950 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15951 ; AVX2-NEXT: vmovaps %ymm0, 1184(%rdx)
15952 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15953 ; AVX2-NEXT: vmovaps %ymm0, 1152(%rdx)
15954 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15955 ; AVX2-NEXT: vmovaps %ymm0, 992(%rdx)
15956 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15957 ; AVX2-NEXT: vmovaps %ymm0, 960(%rdx)
15958 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15959 ; AVX2-NEXT: vmovaps %ymm0, 928(%rdx)
15960 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15961 ; AVX2-NEXT: vmovaps %ymm0, 896(%rdx)
15962 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15963 ; AVX2-NEXT: vmovaps %ymm0, 736(%rdx)
15964 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15965 ; AVX2-NEXT: vmovaps %ymm0, 704(%rdx)
15966 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15967 ; AVX2-NEXT: vmovaps %ymm0, 672(%rdx)
15968 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15969 ; AVX2-NEXT: vmovaps %ymm0, 640(%rdx)
15970 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15971 ; AVX2-NEXT: vmovaps %ymm0, 480(%rdx)
15972 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15973 ; AVX2-NEXT: vmovaps %ymm0, 448(%rdx)
15974 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15975 ; AVX2-NEXT: vmovaps %ymm0, 416(%rdx)
15976 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15977 ; AVX2-NEXT: vmovaps %ymm0, 384(%rdx)
15978 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15979 ; AVX2-NEXT: vmovaps %ymm0, 224(%rdx)
15980 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15981 ; AVX2-NEXT: vmovaps %ymm0, 192(%rdx)
15982 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15983 ; AVX2-NEXT: vmovaps %ymm0, 160(%rdx)
15984 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15985 ; AVX2-NEXT: vmovaps %ymm0, 128(%rdx)
15986 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15987 ; AVX2-NEXT: vmovaps %ymm0, 3936(%rdx)
15988 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15989 ; AVX2-NEXT: vmovaps %ymm0, 3904(%rdx)
15990 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15991 ; AVX2-NEXT: vmovaps %ymm0, 3872(%rdx)
15992 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15993 ; AVX2-NEXT: vmovaps %ymm0, 3840(%rdx)
15994 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15995 ; AVX2-NEXT: vmovaps %ymm0, 3680(%rdx)
15996 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15997 ; AVX2-NEXT: vmovaps %ymm0, 3648(%rdx)
15998 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
15999 ; AVX2-NEXT: vmovaps %ymm0, 3616(%rdx)
16000 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
16001 ; AVX2-NEXT: vmovaps %ymm0, 3584(%rdx)
16002 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
16003 ; AVX2-NEXT: vmovaps %ymm0, 3424(%rdx)
16004 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
16005 ; AVX2-NEXT: vmovaps %ymm0, 3392(%rdx)
16006 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
16007 ; AVX2-NEXT: vmovaps %ymm0, 3360(%rdx)
16008 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
16009 ; AVX2-NEXT: vmovaps %ymm0, 3328(%rdx)
16010 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
16011 ; AVX2-NEXT: vmovaps %ymm0, 3168(%rdx)
16012 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
16013 ; AVX2-NEXT: vmovaps %ymm0, 3136(%rdx)
16014 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
16015 ; AVX2-NEXT: vmovaps %ymm0, 3104(%rdx)
16016 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
16017 ; AVX2-NEXT: vmovaps %ymm0, 3072(%rdx)
16018 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
16019 ; AVX2-NEXT: vmovaps %ymm0, 2912(%rdx)
16020 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
16021 ; AVX2-NEXT: vmovaps %ymm0, 2880(%rdx)
16022 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
16023 ; AVX2-NEXT: vmovaps %ymm0, 2848(%rdx)
16024 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
16025 ; AVX2-NEXT: vmovaps %ymm0, 2816(%rdx)
16026 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
16027 ; AVX2-NEXT: vmovaps %ymm0, 2656(%rdx)
16028 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
16029 ; AVX2-NEXT: vmovaps %ymm0, 2624(%rdx)
16030 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
16031 ; AVX2-NEXT: vmovaps %ymm0, 2592(%rdx)
16032 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
16033 ; AVX2-NEXT: vmovaps %ymm0, 2560(%rdx)
16034 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
16035 ; AVX2-NEXT: vmovaps %ymm0, 2400(%rdx)
16036 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
16037 ; AVX2-NEXT: vmovaps %ymm0, 2368(%rdx)
16038 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
16039 ; AVX2-NEXT: vmovaps %ymm0, 2336(%rdx)
16040 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
16041 ; AVX2-NEXT: vmovaps %ymm0, 2304(%rdx)
16042 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
16043 ; AVX2-NEXT: vmovaps %ymm0, 2144(%rdx)
16044 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
16045 ; AVX2-NEXT: vmovaps %ymm0, 2112(%rdx)
16046 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
16047 ; AVX2-NEXT: vmovaps %ymm0, 2080(%rdx)
16048 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
16049 ; AVX2-NEXT: vmovaps %ymm0, 2048(%rdx)
16050 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
16051 ; AVX2-NEXT: vmovaps %ymm0, 1888(%rdx)
16052 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
16053 ; AVX2-NEXT: vmovaps %ymm0, 1856(%rdx)
16054 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
16055 ; AVX2-NEXT: vmovaps %ymm0, 1824(%rdx)
16056 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
16057 ; AVX2-NEXT: vmovaps %ymm0, 1792(%rdx)
16058 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
16059 ; AVX2-NEXT: vmovaps %ymm0, 1632(%rdx)
16060 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
16061 ; AVX2-NEXT: vmovaps %ymm0, 1600(%rdx)
16062 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
16063 ; AVX2-NEXT: vmovaps %ymm0, 1568(%rdx)
16064 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
16065 ; AVX2-NEXT: vmovaps %ymm0, 1536(%rdx)
16066 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
16067 ; AVX2-NEXT: vmovaps %ymm0, 1376(%rdx)
16068 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
16069 ; AVX2-NEXT: vmovaps %ymm0, 1344(%rdx)
16070 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
16071 ; AVX2-NEXT: vmovaps %ymm0, 1312(%rdx)
16072 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
16073 ; AVX2-NEXT: vmovaps %ymm0, 1280(%rdx)
16074 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
16075 ; AVX2-NEXT: vmovaps %ymm0, 1120(%rdx)
16076 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
16077 ; AVX2-NEXT: vmovaps %ymm0, 1088(%rdx)
16078 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
16079 ; AVX2-NEXT: vmovaps %ymm0, 1056(%rdx)
16080 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
16081 ; AVX2-NEXT: vmovaps %ymm0, 1024(%rdx)
16082 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
16083 ; AVX2-NEXT: vmovaps %ymm0, 864(%rdx)
16084 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
16085 ; AVX2-NEXT: vmovaps %ymm0, 832(%rdx)
16086 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
16087 ; AVX2-NEXT: vmovaps %ymm0, 800(%rdx)
16088 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
16089 ; AVX2-NEXT: vmovaps %ymm0, 768(%rdx)
16090 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
16091 ; AVX2-NEXT: vmovaps %ymm0, 608(%rdx)
16092 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
16093 ; AVX2-NEXT: vmovaps %ymm0, 576(%rdx)
16094 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
16095 ; AVX2-NEXT: vmovaps %ymm0, 544(%rdx)
16096 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
16097 ; AVX2-NEXT: vmovaps %ymm0, 512(%rdx)
16098 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
16099 ; AVX2-NEXT: vmovaps %ymm0, 352(%rdx)
16100 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
16101 ; AVX2-NEXT: vmovaps %ymm0, 320(%rdx)
16102 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
16103 ; AVX2-NEXT: vmovaps %ymm0, 288(%rdx)
16104 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
16105 ; AVX2-NEXT: vmovaps %ymm0, 256(%rdx)
16106 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
16107 ; AVX2-NEXT: vmovaps %ymm0, 96(%rdx)
16108 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
16109 ; AVX2-NEXT: vmovaps %ymm0, 64(%rdx)
16110 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
16111 ; AVX2-NEXT: vmovaps %ymm0, 32(%rdx)
16112 ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
16113 ; AVX2-NEXT: vmovaps %ymm0, (%rdx)
16114 ; AVX2-NEXT: addq $3880, %rsp # imm = 0xF28
16115 ; AVX2-NEXT: vzeroupper
16116 ; AVX2-NEXT: retq
16117 ;
16118 ; AVX2-FP-LABEL: store_i64_stride8_vf64:
16119 ; AVX2-FP: # %bb.0:
16120 ; AVX2-FP-NEXT: subq $3880, %rsp # imm = 0xF28
16121 ; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %r10
16122 ; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %rax
16123 ; AVX2-FP-NEXT: vmovaps (%rcx), %xmm0
16124 ; AVX2-FP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16125 ; AVX2-FP-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0
16126 ; AVX2-FP-NEXT: vmovaps (%rsi), %xmm2
16127 ; AVX2-FP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16128 ; AVX2-FP-NEXT: vmovaps 32(%rsi), %xmm3
16129 ; AVX2-FP-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16130 ; AVX2-FP-NEXT: vmovaps (%rdi), %xmm1
16131 ; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16132 ; AVX2-FP-NEXT: vmovaps 32(%rdi), %xmm4
16133 ; AVX2-FP-NEXT: vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16134 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm2[1]
16135 ; AVX2-FP-NEXT: vbroadcastsd 8(%rdx), %ymm2
16136 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
16137 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
16138 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16139 ; AVX2-FP-NEXT: vmovaps (%rax), %xmm0
16140 ; AVX2-FP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16141 ; AVX2-FP-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0
16142 ; AVX2-FP-NEXT: vmovaps (%r9), %xmm2
16143 ; AVX2-FP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16144 ; AVX2-FP-NEXT: vmovaps (%r8), %xmm1
16145 ; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16146 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm2[1]
16147 ; AVX2-FP-NEXT: vbroadcastsd 8(%r10), %ymm2
16148 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
16149 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
16150 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16151 ; AVX2-FP-NEXT: vmovaps 32(%rcx), %xmm0
16152 ; AVX2-FP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16153 ; AVX2-FP-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0
16154 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm4[1],xmm3[1]
16155 ; AVX2-FP-NEXT: vbroadcastsd 40(%rdx), %ymm2
16156 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
16157 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
16158 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16159 ; AVX2-FP-NEXT: vmovaps 32(%r9), %xmm1
16160 ; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16161 ; AVX2-FP-NEXT: vmovaps 32(%r8), %xmm0
16162 ; AVX2-FP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16163 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
16164 ; AVX2-FP-NEXT: vbroadcastsd 40(%r10), %ymm1
16165 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
16166 ; AVX2-FP-NEXT: vmovaps 32(%rax), %xmm1
16167 ; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16168 ; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
16169 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
16170 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16171 ; AVX2-FP-NEXT: vmovaps 64(%rsi), %xmm1
16172 ; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16173 ; AVX2-FP-NEXT: vmovaps 64(%rdi), %xmm0
16174 ; AVX2-FP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16175 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
16176 ; AVX2-FP-NEXT: vbroadcastsd 72(%rdx), %ymm1
16177 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
16178 ; AVX2-FP-NEXT: vmovaps 64(%rcx), %xmm1
16179 ; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16180 ; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
16181 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
16182 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16183 ; AVX2-FP-NEXT: vmovaps 64(%r9), %xmm1
16184 ; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16185 ; AVX2-FP-NEXT: vmovaps 64(%r8), %xmm0
16186 ; AVX2-FP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16187 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
16188 ; AVX2-FP-NEXT: vbroadcastsd 72(%r10), %ymm1
16189 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
16190 ; AVX2-FP-NEXT: vmovaps 64(%rax), %xmm1
16191 ; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16192 ; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
16193 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
16194 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16195 ; AVX2-FP-NEXT: vmovaps 96(%rsi), %xmm1
16196 ; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16197 ; AVX2-FP-NEXT: vmovaps 96(%rdi), %xmm0
16198 ; AVX2-FP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16199 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
16200 ; AVX2-FP-NEXT: vbroadcastsd 104(%rdx), %ymm1
16201 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
16202 ; AVX2-FP-NEXT: vmovaps 96(%rcx), %xmm1
16203 ; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16204 ; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
16205 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
16206 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16207 ; AVX2-FP-NEXT: vmovaps 96(%r9), %xmm1
16208 ; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16209 ; AVX2-FP-NEXT: vmovaps 96(%r8), %xmm0
16210 ; AVX2-FP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16211 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
16212 ; AVX2-FP-NEXT: vbroadcastsd 104(%r10), %ymm1
16213 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
16214 ; AVX2-FP-NEXT: vmovaps 96(%rax), %xmm1
16215 ; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16216 ; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
16217 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
16218 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16219 ; AVX2-FP-NEXT: vmovaps 128(%rsi), %xmm1
16220 ; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16221 ; AVX2-FP-NEXT: vmovaps 128(%rdi), %xmm0
16222 ; AVX2-FP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16223 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
16224 ; AVX2-FP-NEXT: vbroadcastsd 136(%rdx), %ymm1
16225 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
16226 ; AVX2-FP-NEXT: vmovaps 128(%rcx), %xmm1
16227 ; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16228 ; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
16229 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
16230 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16231 ; AVX2-FP-NEXT: vmovaps 128(%r9), %xmm1
16232 ; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16233 ; AVX2-FP-NEXT: vmovaps 128(%r8), %xmm0
16234 ; AVX2-FP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16235 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
16236 ; AVX2-FP-NEXT: vbroadcastsd 136(%r10), %ymm1
16237 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
16238 ; AVX2-FP-NEXT: vmovaps 128(%rax), %xmm1
16239 ; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16240 ; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
16241 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
16242 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16243 ; AVX2-FP-NEXT: vmovaps 160(%rsi), %xmm1
16244 ; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16245 ; AVX2-FP-NEXT: vmovaps 160(%rdi), %xmm0
16246 ; AVX2-FP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16247 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
16248 ; AVX2-FP-NEXT: vbroadcastsd 168(%rdx), %ymm1
16249 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
16250 ; AVX2-FP-NEXT: vmovaps 160(%rcx), %xmm1
16251 ; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16252 ; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
16253 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
16254 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16255 ; AVX2-FP-NEXT: vmovaps 160(%r9), %xmm1
16256 ; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16257 ; AVX2-FP-NEXT: vmovaps 160(%r8), %xmm0
16258 ; AVX2-FP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16259 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
16260 ; AVX2-FP-NEXT: vbroadcastsd 168(%r10), %ymm1
16261 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
16262 ; AVX2-FP-NEXT: vmovaps 160(%rax), %xmm1
16263 ; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16264 ; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
16265 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
16266 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16267 ; AVX2-FP-NEXT: vmovaps 192(%rsi), %xmm1
16268 ; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16269 ; AVX2-FP-NEXT: vmovaps 192(%rdi), %xmm0
16270 ; AVX2-FP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16271 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
16272 ; AVX2-FP-NEXT: vbroadcastsd 200(%rdx), %ymm1
16273 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
16274 ; AVX2-FP-NEXT: vmovaps 192(%rcx), %xmm1
16275 ; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16276 ; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
16277 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
16278 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16279 ; AVX2-FP-NEXT: vmovaps 192(%r9), %xmm1
16280 ; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16281 ; AVX2-FP-NEXT: vmovaps 192(%r8), %xmm0
16282 ; AVX2-FP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16283 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
16284 ; AVX2-FP-NEXT: vbroadcastsd 200(%r10), %ymm1
16285 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
16286 ; AVX2-FP-NEXT: vmovaps 192(%rax), %xmm1
16287 ; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16288 ; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
16289 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
16290 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16291 ; AVX2-FP-NEXT: vmovaps 224(%rsi), %xmm1
16292 ; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16293 ; AVX2-FP-NEXT: vmovaps 224(%rdi), %xmm0
16294 ; AVX2-FP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16295 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
16296 ; AVX2-FP-NEXT: vbroadcastsd 232(%rdx), %ymm1
16297 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
16298 ; AVX2-FP-NEXT: vmovaps 224(%rcx), %xmm1
16299 ; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16300 ; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
16301 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
16302 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16303 ; AVX2-FP-NEXT: vmovaps 224(%r9), %xmm1
16304 ; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16305 ; AVX2-FP-NEXT: vmovaps 224(%r8), %xmm0
16306 ; AVX2-FP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16307 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
16308 ; AVX2-FP-NEXT: vbroadcastsd 232(%r10), %ymm1
16309 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
16310 ; AVX2-FP-NEXT: vmovaps 224(%rax), %xmm1
16311 ; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16312 ; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
16313 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
16314 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16315 ; AVX2-FP-NEXT: vmovaps 256(%rsi), %xmm1
16316 ; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16317 ; AVX2-FP-NEXT: vmovaps 256(%rdi), %xmm0
16318 ; AVX2-FP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16319 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
16320 ; AVX2-FP-NEXT: vbroadcastsd 264(%rdx), %ymm1
16321 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
16322 ; AVX2-FP-NEXT: vmovaps 256(%rcx), %xmm1
16323 ; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16324 ; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
16325 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
16326 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16327 ; AVX2-FP-NEXT: vmovaps 256(%r9), %xmm1
16328 ; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16329 ; AVX2-FP-NEXT: vmovaps 256(%r8), %xmm0
16330 ; AVX2-FP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16331 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
16332 ; AVX2-FP-NEXT: vbroadcastsd 264(%r10), %ymm1
16333 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
16334 ; AVX2-FP-NEXT: vmovaps 256(%rax), %xmm1
16335 ; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16336 ; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
16337 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
16338 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16339 ; AVX2-FP-NEXT: vmovaps 288(%rsi), %xmm1
16340 ; AVX2-FP-NEXT: vmovaps %xmm1, (%rsp) # 16-byte Spill
16341 ; AVX2-FP-NEXT: vmovaps 288(%rdi), %xmm0
16342 ; AVX2-FP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16343 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
16344 ; AVX2-FP-NEXT: vbroadcastsd 296(%rdx), %ymm1
16345 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
16346 ; AVX2-FP-NEXT: vmovaps 288(%rcx), %xmm1
16347 ; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16348 ; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
16349 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
16350 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16351 ; AVX2-FP-NEXT: vmovaps 288(%r9), %xmm1
16352 ; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16353 ; AVX2-FP-NEXT: vmovaps 288(%r8), %xmm0
16354 ; AVX2-FP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16355 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
16356 ; AVX2-FP-NEXT: vbroadcastsd 296(%r10), %ymm1
16357 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
16358 ; AVX2-FP-NEXT: vmovaps 288(%rax), %xmm1
16359 ; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16360 ; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
16361 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
16362 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16363 ; AVX2-FP-NEXT: vmovaps 320(%rsi), %xmm1
16364 ; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16365 ; AVX2-FP-NEXT: vmovaps 320(%rdi), %xmm0
16366 ; AVX2-FP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16367 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
16368 ; AVX2-FP-NEXT: vbroadcastsd 328(%rdx), %ymm1
16369 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
16370 ; AVX2-FP-NEXT: vmovaps 320(%rcx), %xmm1
16371 ; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16372 ; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
16373 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
16374 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16375 ; AVX2-FP-NEXT: vmovaps 320(%r9), %xmm1
16376 ; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16377 ; AVX2-FP-NEXT: vmovaps 320(%r8), %xmm0
16378 ; AVX2-FP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16379 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
16380 ; AVX2-FP-NEXT: vbroadcastsd 328(%r10), %ymm1
16381 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
16382 ; AVX2-FP-NEXT: vmovaps 320(%rax), %xmm1
16383 ; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16384 ; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
16385 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
16386 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16387 ; AVX2-FP-NEXT: vmovaps 352(%rsi), %xmm1
16388 ; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16389 ; AVX2-FP-NEXT: vmovaps 352(%rdi), %xmm0
16390 ; AVX2-FP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16391 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
16392 ; AVX2-FP-NEXT: vbroadcastsd 360(%rdx), %ymm1
16393 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
16394 ; AVX2-FP-NEXT: vmovaps 352(%rcx), %xmm1
16395 ; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16396 ; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
16397 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
16398 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16399 ; AVX2-FP-NEXT: vmovaps 352(%r9), %xmm1
16400 ; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16401 ; AVX2-FP-NEXT: vmovaps 352(%r8), %xmm0
16402 ; AVX2-FP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16403 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
16404 ; AVX2-FP-NEXT: vbroadcastsd 360(%r10), %ymm1
16405 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
16406 ; AVX2-FP-NEXT: vmovaps 352(%rax), %xmm1
16407 ; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16408 ; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
16409 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
16410 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16411 ; AVX2-FP-NEXT: vmovaps 384(%rsi), %xmm1
16412 ; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16413 ; AVX2-FP-NEXT: vmovaps 384(%rdi), %xmm0
16414 ; AVX2-FP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16415 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
16416 ; AVX2-FP-NEXT: vbroadcastsd 392(%rdx), %ymm1
16417 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
16418 ; AVX2-FP-NEXT: vmovaps 384(%rcx), %xmm1
16419 ; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16420 ; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
16421 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
16422 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16423 ; AVX2-FP-NEXT: vmovaps 384(%r9), %xmm1
16424 ; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16425 ; AVX2-FP-NEXT: vmovaps 384(%r8), %xmm0
16426 ; AVX2-FP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16427 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
16428 ; AVX2-FP-NEXT: vbroadcastsd 392(%r10), %ymm1
16429 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
16430 ; AVX2-FP-NEXT: vmovaps 384(%rax), %xmm1
16431 ; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16432 ; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
16433 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
16434 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16435 ; AVX2-FP-NEXT: vmovaps 416(%rsi), %xmm1
16436 ; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16437 ; AVX2-FP-NEXT: vmovaps 416(%rdi), %xmm0
16438 ; AVX2-FP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16439 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
16440 ; AVX2-FP-NEXT: vbroadcastsd 424(%rdx), %ymm1
16441 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
16442 ; AVX2-FP-NEXT: vmovaps 416(%rcx), %xmm1
16443 ; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16444 ; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
16445 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
16446 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16447 ; AVX2-FP-NEXT: vmovaps 416(%r9), %xmm0
16448 ; AVX2-FP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
16449 ; AVX2-FP-NEXT: vmovaps 416(%r8), %xmm13
16450 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm13[1],xmm0[1]
16451 ; AVX2-FP-NEXT: vbroadcastsd 424(%r10), %ymm1
16452 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
16453 ; AVX2-FP-NEXT: vmovaps 416(%rax), %xmm12
16454 ; AVX2-FP-NEXT: vinsertf128 $1, %xmm12, %ymm0, %ymm1
16455 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
16456 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16457 ; AVX2-FP-NEXT: vmovaps 448(%rsi), %xmm11
16458 ; AVX2-FP-NEXT: vmovaps 448(%rdi), %xmm10
16459 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm10[1],xmm11[1]
16460 ; AVX2-FP-NEXT: vbroadcastsd 456(%rdx), %ymm1
16461 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
16462 ; AVX2-FP-NEXT: vmovaps 448(%rcx), %xmm9
16463 ; AVX2-FP-NEXT: vinsertf128 $1, %xmm9, %ymm0, %ymm1
16464 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
16465 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16466 ; AVX2-FP-NEXT: vmovaps 448(%r9), %xmm8
16467 ; AVX2-FP-NEXT: vmovaps 448(%r8), %xmm7
16468 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm7[1],xmm8[1]
16469 ; AVX2-FP-NEXT: vbroadcastsd 456(%r10), %ymm1
16470 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
16471 ; AVX2-FP-NEXT: vmovaps 448(%rax), %xmm6
16472 ; AVX2-FP-NEXT: vinsertf128 $1, %xmm6, %ymm0, %ymm1
16473 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
16474 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16475 ; AVX2-FP-NEXT: vmovaps 480(%rsi), %xmm5
16476 ; AVX2-FP-NEXT: vmovaps 480(%rdi), %xmm4
16477 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm4[1],xmm5[1]
16478 ; AVX2-FP-NEXT: vbroadcastsd 488(%rdx), %ymm1
16479 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
16480 ; AVX2-FP-NEXT: vmovaps 480(%rcx), %xmm3
16481 ; AVX2-FP-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm1
16482 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
16483 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16484 ; AVX2-FP-NEXT: vmovaps 480(%r9), %xmm2
16485 ; AVX2-FP-NEXT: vmovaps 480(%r8), %xmm1
16486 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm2[1]
16487 ; AVX2-FP-NEXT: vbroadcastsd 488(%r10), %ymm15
16488 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm15 = ymm0[0,1,2,3],ymm15[4,5,6,7]
16489 ; AVX2-FP-NEXT: vmovaps 480(%rax), %xmm0
16490 ; AVX2-FP-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm14
16491 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm14 = ymm15[0,1,2,3,4,5],ymm14[6,7]
16492 ; AVX2-FP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16493 ; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
16494 ; AVX2-FP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
16495 ; AVX2-FP-NEXT: # xmm14 = xmm14[0],mem[0]
16496 ; AVX2-FP-NEXT: vinsertf128 $1, (%rdx), %ymm14, %ymm14
16497 ; AVX2-FP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
16498 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
16499 ; AVX2-FP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16500 ; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
16501 ; AVX2-FP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
16502 ; AVX2-FP-NEXT: # xmm14 = xmm14[0],mem[0]
16503 ; AVX2-FP-NEXT: vinsertf128 $1, (%r10), %ymm14, %ymm14
16504 ; AVX2-FP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
16505 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
16506 ; AVX2-FP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16507 ; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
16508 ; AVX2-FP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
16509 ; AVX2-FP-NEXT: # xmm14 = xmm14[0],mem[0]
16510 ; AVX2-FP-NEXT: vinsertf128 $1, 32(%rdx), %ymm14, %ymm14
16511 ; AVX2-FP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
16512 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
16513 ; AVX2-FP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16514 ; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
16515 ; AVX2-FP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
16516 ; AVX2-FP-NEXT: # xmm14 = xmm14[0],mem[0]
16517 ; AVX2-FP-NEXT: vinsertf128 $1, 32(%r10), %ymm14, %ymm14
16518 ; AVX2-FP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
16519 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
16520 ; AVX2-FP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16521 ; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
16522 ; AVX2-FP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
16523 ; AVX2-FP-NEXT: # xmm14 = xmm14[0],mem[0]
16524 ; AVX2-FP-NEXT: vinsertf128 $1, 64(%rdx), %ymm14, %ymm14
16525 ; AVX2-FP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
16526 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
16527 ; AVX2-FP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16528 ; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
16529 ; AVX2-FP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
16530 ; AVX2-FP-NEXT: # xmm14 = xmm14[0],mem[0]
16531 ; AVX2-FP-NEXT: vinsertf128 $1, 64(%r10), %ymm14, %ymm14
16532 ; AVX2-FP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
16533 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
16534 ; AVX2-FP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16535 ; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
16536 ; AVX2-FP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
16537 ; AVX2-FP-NEXT: # xmm14 = xmm14[0],mem[0]
16538 ; AVX2-FP-NEXT: vinsertf128 $1, 96(%rdx), %ymm14, %ymm14
16539 ; AVX2-FP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
16540 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
16541 ; AVX2-FP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16542 ; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
16543 ; AVX2-FP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
16544 ; AVX2-FP-NEXT: # xmm14 = xmm14[0],mem[0]
16545 ; AVX2-FP-NEXT: vinsertf128 $1, 96(%r10), %ymm14, %ymm14
16546 ; AVX2-FP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
16547 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
16548 ; AVX2-FP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16549 ; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
16550 ; AVX2-FP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
16551 ; AVX2-FP-NEXT: # xmm14 = xmm14[0],mem[0]
16552 ; AVX2-FP-NEXT: vinsertf128 $1, 128(%rdx), %ymm14, %ymm14
16553 ; AVX2-FP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
16554 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
16555 ; AVX2-FP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16556 ; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
16557 ; AVX2-FP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
16558 ; AVX2-FP-NEXT: # xmm14 = xmm14[0],mem[0]
16559 ; AVX2-FP-NEXT: vinsertf128 $1, 128(%r10), %ymm14, %ymm14
16560 ; AVX2-FP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
16561 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
16562 ; AVX2-FP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16563 ; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
16564 ; AVX2-FP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
16565 ; AVX2-FP-NEXT: # xmm14 = xmm14[0],mem[0]
16566 ; AVX2-FP-NEXT: vinsertf128 $1, 160(%rdx), %ymm14, %ymm14
16567 ; AVX2-FP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
16568 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
16569 ; AVX2-FP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16570 ; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
16571 ; AVX2-FP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
16572 ; AVX2-FP-NEXT: # xmm14 = xmm14[0],mem[0]
16573 ; AVX2-FP-NEXT: vinsertf128 $1, 160(%r10), %ymm14, %ymm14
16574 ; AVX2-FP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
16575 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
16576 ; AVX2-FP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16577 ; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
16578 ; AVX2-FP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
16579 ; AVX2-FP-NEXT: # xmm14 = xmm14[0],mem[0]
16580 ; AVX2-FP-NEXT: vinsertf128 $1, 192(%rdx), %ymm14, %ymm14
16581 ; AVX2-FP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
16582 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
16583 ; AVX2-FP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16584 ; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
16585 ; AVX2-FP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
16586 ; AVX2-FP-NEXT: # xmm14 = xmm14[0],mem[0]
16587 ; AVX2-FP-NEXT: vinsertf128 $1, 192(%r10), %ymm14, %ymm14
16588 ; AVX2-FP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
16589 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
16590 ; AVX2-FP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16591 ; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
16592 ; AVX2-FP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
16593 ; AVX2-FP-NEXT: # xmm14 = xmm14[0],mem[0]
16594 ; AVX2-FP-NEXT: vinsertf128 $1, 224(%rdx), %ymm14, %ymm14
16595 ; AVX2-FP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
16596 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
16597 ; AVX2-FP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16598 ; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
16599 ; AVX2-FP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
16600 ; AVX2-FP-NEXT: # xmm14 = xmm14[0],mem[0]
16601 ; AVX2-FP-NEXT: vinsertf128 $1, 224(%r10), %ymm14, %ymm14
16602 ; AVX2-FP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
16603 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
16604 ; AVX2-FP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16605 ; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
16606 ; AVX2-FP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
16607 ; AVX2-FP-NEXT: # xmm14 = xmm14[0],mem[0]
16608 ; AVX2-FP-NEXT: vinsertf128 $1, 256(%rdx), %ymm14, %ymm14
16609 ; AVX2-FP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
16610 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
16611 ; AVX2-FP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16612 ; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
16613 ; AVX2-FP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
16614 ; AVX2-FP-NEXT: # xmm14 = xmm14[0],mem[0]
16615 ; AVX2-FP-NEXT: vinsertf128 $1, 256(%r10), %ymm14, %ymm14
16616 ; AVX2-FP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
16617 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
16618 ; AVX2-FP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16619 ; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
16620 ; AVX2-FP-NEXT: vunpcklpd (%rsp), %xmm14, %xmm14 # 16-byte Folded Reload
16621 ; AVX2-FP-NEXT: # xmm14 = xmm14[0],mem[0]
16622 ; AVX2-FP-NEXT: vinsertf128 $1, 288(%rdx), %ymm14, %ymm14
16623 ; AVX2-FP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
16624 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
16625 ; AVX2-FP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16626 ; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
16627 ; AVX2-FP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
16628 ; AVX2-FP-NEXT: # xmm14 = xmm14[0],mem[0]
16629 ; AVX2-FP-NEXT: vinsertf128 $1, 288(%r10), %ymm14, %ymm14
16630 ; AVX2-FP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
16631 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
16632 ; AVX2-FP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16633 ; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
16634 ; AVX2-FP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
16635 ; AVX2-FP-NEXT: # xmm14 = xmm14[0],mem[0]
16636 ; AVX2-FP-NEXT: vinsertf128 $1, 320(%rdx), %ymm14, %ymm14
16637 ; AVX2-FP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
16638 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
16639 ; AVX2-FP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16640 ; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
16641 ; AVX2-FP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
16642 ; AVX2-FP-NEXT: # xmm14 = xmm14[0],mem[0]
16643 ; AVX2-FP-NEXT: vinsertf128 $1, 320(%r10), %ymm14, %ymm14
16644 ; AVX2-FP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
16645 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
16646 ; AVX2-FP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16647 ; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
16648 ; AVX2-FP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
16649 ; AVX2-FP-NEXT: # xmm14 = xmm14[0],mem[0]
16650 ; AVX2-FP-NEXT: vinsertf128 $1, 352(%rdx), %ymm14, %ymm14
16651 ; AVX2-FP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
16652 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
16653 ; AVX2-FP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16654 ; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
16655 ; AVX2-FP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
16656 ; AVX2-FP-NEXT: # xmm14 = xmm14[0],mem[0]
16657 ; AVX2-FP-NEXT: vinsertf128 $1, 352(%r10), %ymm14, %ymm14
16658 ; AVX2-FP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
16659 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
16660 ; AVX2-FP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16661 ; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
16662 ; AVX2-FP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
16663 ; AVX2-FP-NEXT: # xmm14 = xmm14[0],mem[0]
16664 ; AVX2-FP-NEXT: vinsertf128 $1, 384(%rdx), %ymm14, %ymm14
16665 ; AVX2-FP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
16666 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
16667 ; AVX2-FP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16668 ; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
16669 ; AVX2-FP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
16670 ; AVX2-FP-NEXT: # xmm14 = xmm14[0],mem[0]
16671 ; AVX2-FP-NEXT: vinsertf128 $1, 384(%r10), %ymm14, %ymm14
16672 ; AVX2-FP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
16673 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
16674 ; AVX2-FP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16675 ; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
16676 ; AVX2-FP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
16677 ; AVX2-FP-NEXT: # xmm14 = xmm14[0],mem[0]
16678 ; AVX2-FP-NEXT: vinsertf128 $1, 416(%rdx), %ymm14, %ymm14
16679 ; AVX2-FP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
16680 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
16681 ; AVX2-FP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16682 ; AVX2-FP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm13, %xmm13 # 16-byte Folded Reload
16683 ; AVX2-FP-NEXT: # xmm13 = xmm13[0],mem[0]
16684 ; AVX2-FP-NEXT: vinsertf128 $1, 416(%r10), %ymm13, %ymm13
16685 ; AVX2-FP-NEXT: vbroadcastsd %xmm12, %ymm12
16686 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm12 = ymm13[0,1,2,3,4,5],ymm12[6,7]
16687 ; AVX2-FP-NEXT: vmovups %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16688 ; AVX2-FP-NEXT: vmovlhps {{.*#+}} xmm10 = xmm10[0],xmm11[0]
16689 ; AVX2-FP-NEXT: vinsertf128 $1, 448(%rdx), %ymm10, %ymm10
16690 ; AVX2-FP-NEXT: vbroadcastsd %xmm9, %ymm9
16691 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm9 = ymm10[0,1,2,3,4,5],ymm9[6,7]
16692 ; AVX2-FP-NEXT: vmovups %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16693 ; AVX2-FP-NEXT: vmovlhps {{.*#+}} xmm7 = xmm7[0],xmm8[0]
16694 ; AVX2-FP-NEXT: vinsertf128 $1, 448(%r10), %ymm7, %ymm7
16695 ; AVX2-FP-NEXT: vbroadcastsd %xmm6, %ymm6
16696 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3,4,5],ymm6[6,7]
16697 ; AVX2-FP-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16698 ; AVX2-FP-NEXT: vmovlhps {{.*#+}} xmm4 = xmm4[0],xmm5[0]
16699 ; AVX2-FP-NEXT: vinsertf128 $1, 480(%rdx), %ymm4, %ymm4
16700 ; AVX2-FP-NEXT: vbroadcastsd %xmm3, %ymm3
16701 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm4[0,1,2,3,4,5],ymm3[6,7]
16702 ; AVX2-FP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16703 ; AVX2-FP-NEXT: vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
16704 ; AVX2-FP-NEXT: vinsertf128 $1, 480(%r10), %ymm1, %ymm1
16705 ; AVX2-FP-NEXT: vbroadcastsd %xmm0, %ymm0
16706 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
16707 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16708 ; AVX2-FP-NEXT: vmovaps (%rdi), %ymm0
16709 ; AVX2-FP-NEXT: vmovaps (%rsi), %ymm1
16710 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
16711 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
16712 ; AVX2-FP-NEXT: vbroadcastsd 16(%rcx), %ymm3
16713 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
16714 ; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16715 ; AVX2-FP-NEXT: vmovaps (%r8), %ymm2
16716 ; AVX2-FP-NEXT: vmovaps (%r9), %ymm3
16717 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
16718 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],mem[2,3]
16719 ; AVX2-FP-NEXT: vbroadcastsd 16(%rax), %ymm5
16720 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
16721 ; AVX2-FP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16722 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
16723 ; AVX2-FP-NEXT: vbroadcastsd 24(%rdx), %ymm1
16724 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm6 = ymm0[2,3],ymm1[2,3]
16725 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
16726 ; AVX2-FP-NEXT: vbroadcastsd 24(%r10), %ymm1
16727 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm7 = ymm0[2,3],ymm1[2,3]
16728 ; AVX2-FP-NEXT: vmovaps 32(%rdi), %ymm0
16729 ; AVX2-FP-NEXT: vmovaps 32(%rsi), %ymm1
16730 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
16731 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
16732 ; AVX2-FP-NEXT: vbroadcastsd 48(%rcx), %ymm3
16733 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
16734 ; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16735 ; AVX2-FP-NEXT: vmovaps 32(%r8), %ymm2
16736 ; AVX2-FP-NEXT: vmovaps 32(%r9), %ymm3
16737 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
16738 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],mem[2,3]
16739 ; AVX2-FP-NEXT: vbroadcastsd 48(%rax), %ymm5
16740 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
16741 ; AVX2-FP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16742 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
16743 ; AVX2-FP-NEXT: vbroadcastsd 56(%rdx), %ymm1
16744 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm8 = ymm0[2,3],ymm1[2,3]
16745 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
16746 ; AVX2-FP-NEXT: vbroadcastsd 56(%r10), %ymm1
16747 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm10 = ymm0[2,3],ymm1[2,3]
16748 ; AVX2-FP-NEXT: vmovaps 64(%rdi), %ymm0
16749 ; AVX2-FP-NEXT: vmovaps 64(%rsi), %ymm1
16750 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
16751 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
16752 ; AVX2-FP-NEXT: vbroadcastsd 80(%rcx), %ymm3
16753 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
16754 ; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16755 ; AVX2-FP-NEXT: vmovaps 64(%r8), %ymm2
16756 ; AVX2-FP-NEXT: vmovaps 64(%r9), %ymm3
16757 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
16758 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],mem[2,3]
16759 ; AVX2-FP-NEXT: vbroadcastsd 80(%rax), %ymm5
16760 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
16761 ; AVX2-FP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16762 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
16763 ; AVX2-FP-NEXT: vbroadcastsd 88(%rdx), %ymm1
16764 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm12 = ymm0[2,3],ymm1[2,3]
16765 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
16766 ; AVX2-FP-NEXT: vbroadcastsd 88(%r10), %ymm1
16767 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm13 = ymm0[2,3],ymm1[2,3]
16768 ; AVX2-FP-NEXT: vmovaps 96(%rdi), %ymm0
16769 ; AVX2-FP-NEXT: vmovaps 96(%rsi), %ymm1
16770 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
16771 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
16772 ; AVX2-FP-NEXT: vbroadcastsd 112(%rcx), %ymm3
16773 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
16774 ; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16775 ; AVX2-FP-NEXT: vmovaps 96(%r8), %ymm2
16776 ; AVX2-FP-NEXT: vmovaps 96(%r9), %ymm3
16777 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
16778 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],mem[2,3]
16779 ; AVX2-FP-NEXT: vbroadcastsd 112(%rax), %ymm5
16780 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
16781 ; AVX2-FP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16782 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
16783 ; AVX2-FP-NEXT: vbroadcastsd 120(%rdx), %ymm1
16784 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm14 = ymm0[2,3],ymm1[2,3]
16785 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
16786 ; AVX2-FP-NEXT: vbroadcastsd 120(%r10), %ymm1
16787 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
16788 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16789 ; AVX2-FP-NEXT: vmovaps 128(%rdi), %ymm0
16790 ; AVX2-FP-NEXT: vmovaps 128(%rsi), %ymm1
16791 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
16792 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
16793 ; AVX2-FP-NEXT: vbroadcastsd 144(%rcx), %ymm3
16794 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
16795 ; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16796 ; AVX2-FP-NEXT: vmovaps 128(%r8), %ymm2
16797 ; AVX2-FP-NEXT: vmovaps 128(%r9), %ymm3
16798 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
16799 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],mem[2,3]
16800 ; AVX2-FP-NEXT: vbroadcastsd 144(%rax), %ymm5
16801 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
16802 ; AVX2-FP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16803 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
16804 ; AVX2-FP-NEXT: vbroadcastsd 152(%rdx), %ymm1
16805 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
16806 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16807 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
16808 ; AVX2-FP-NEXT: vbroadcastsd 152(%r10), %ymm1
16809 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
16810 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16811 ; AVX2-FP-NEXT: vmovaps 160(%rdi), %ymm0
16812 ; AVX2-FP-NEXT: vmovaps 160(%rsi), %ymm1
16813 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
16814 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
16815 ; AVX2-FP-NEXT: vbroadcastsd 176(%rcx), %ymm3
16816 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
16817 ; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16818 ; AVX2-FP-NEXT: vmovaps 160(%r8), %ymm2
16819 ; AVX2-FP-NEXT: vmovaps 160(%r9), %ymm3
16820 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
16821 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],mem[2,3]
16822 ; AVX2-FP-NEXT: vbroadcastsd 176(%rax), %ymm5
16823 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
16824 ; AVX2-FP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16825 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
16826 ; AVX2-FP-NEXT: vbroadcastsd 184(%rdx), %ymm1
16827 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
16828 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16829 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
16830 ; AVX2-FP-NEXT: vbroadcastsd 184(%r10), %ymm1
16831 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
16832 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16833 ; AVX2-FP-NEXT: vmovaps 192(%rdi), %ymm0
16834 ; AVX2-FP-NEXT: vmovaps 192(%rsi), %ymm1
16835 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
16836 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
16837 ; AVX2-FP-NEXT: vbroadcastsd 208(%rcx), %ymm3
16838 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
16839 ; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16840 ; AVX2-FP-NEXT: vmovaps 192(%r8), %ymm2
16841 ; AVX2-FP-NEXT: vmovaps 192(%r9), %ymm3
16842 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
16843 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],mem[2,3]
16844 ; AVX2-FP-NEXT: vbroadcastsd 208(%rax), %ymm5
16845 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
16846 ; AVX2-FP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16847 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
16848 ; AVX2-FP-NEXT: vbroadcastsd 216(%rdx), %ymm1
16849 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
16850 ; AVX2-FP-NEXT: vmovups %ymm0, (%rsp) # 32-byte Spill
16851 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
16852 ; AVX2-FP-NEXT: vbroadcastsd 216(%r10), %ymm1
16853 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
16854 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16855 ; AVX2-FP-NEXT: vmovaps 224(%rdi), %ymm0
16856 ; AVX2-FP-NEXT: vmovaps 224(%rsi), %ymm1
16857 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
16858 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
16859 ; AVX2-FP-NEXT: vbroadcastsd 240(%rcx), %ymm3
16860 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
16861 ; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16862 ; AVX2-FP-NEXT: vmovaps 224(%r8), %ymm2
16863 ; AVX2-FP-NEXT: vmovaps 224(%r9), %ymm3
16864 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
16865 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],mem[2,3]
16866 ; AVX2-FP-NEXT: vbroadcastsd 240(%rax), %ymm5
16867 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
16868 ; AVX2-FP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16869 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
16870 ; AVX2-FP-NEXT: vbroadcastsd 248(%rdx), %ymm1
16871 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
16872 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16873 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
16874 ; AVX2-FP-NEXT: vbroadcastsd 248(%r10), %ymm1
16875 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
16876 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16877 ; AVX2-FP-NEXT: vmovaps 256(%rdi), %ymm0
16878 ; AVX2-FP-NEXT: vmovaps 256(%rsi), %ymm1
16879 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
16880 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
16881 ; AVX2-FP-NEXT: vbroadcastsd 272(%rcx), %ymm3
16882 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
16883 ; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16884 ; AVX2-FP-NEXT: vmovaps 256(%r8), %ymm2
16885 ; AVX2-FP-NEXT: vmovaps 256(%r9), %ymm3
16886 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
16887 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],mem[2,3]
16888 ; AVX2-FP-NEXT: vbroadcastsd 272(%rax), %ymm5
16889 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
16890 ; AVX2-FP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16891 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
16892 ; AVX2-FP-NEXT: vbroadcastsd 280(%rdx), %ymm1
16893 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
16894 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16895 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
16896 ; AVX2-FP-NEXT: vbroadcastsd 280(%r10), %ymm1
16897 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
16898 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16899 ; AVX2-FP-NEXT: vmovaps 288(%rdi), %ymm0
16900 ; AVX2-FP-NEXT: vmovaps 288(%rsi), %ymm1
16901 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
16902 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
16903 ; AVX2-FP-NEXT: vbroadcastsd 304(%rcx), %ymm3
16904 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
16905 ; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16906 ; AVX2-FP-NEXT: vmovaps 288(%r8), %ymm2
16907 ; AVX2-FP-NEXT: vmovaps 288(%r9), %ymm3
16908 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
16909 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],mem[2,3]
16910 ; AVX2-FP-NEXT: vbroadcastsd 304(%rax), %ymm5
16911 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
16912 ; AVX2-FP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16913 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
16914 ; AVX2-FP-NEXT: vbroadcastsd 312(%rdx), %ymm1
16915 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
16916 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16917 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
16918 ; AVX2-FP-NEXT: vbroadcastsd 312(%r10), %ymm1
16919 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
16920 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16921 ; AVX2-FP-NEXT: vmovaps 320(%rdi), %ymm0
16922 ; AVX2-FP-NEXT: vmovaps 320(%rsi), %ymm1
16923 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
16924 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
16925 ; AVX2-FP-NEXT: vbroadcastsd 336(%rcx), %ymm3
16926 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
16927 ; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16928 ; AVX2-FP-NEXT: vmovaps 320(%r8), %ymm2
16929 ; AVX2-FP-NEXT: vmovaps 320(%r9), %ymm3
16930 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
16931 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],mem[2,3]
16932 ; AVX2-FP-NEXT: vbroadcastsd 336(%rax), %ymm5
16933 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
16934 ; AVX2-FP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16935 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
16936 ; AVX2-FP-NEXT: vbroadcastsd 344(%rdx), %ymm1
16937 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
16938 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16939 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
16940 ; AVX2-FP-NEXT: vbroadcastsd 344(%r10), %ymm1
16941 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
16942 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16943 ; AVX2-FP-NEXT: vmovaps 352(%rdi), %ymm0
16944 ; AVX2-FP-NEXT: vmovaps 352(%rsi), %ymm1
16945 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
16946 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
16947 ; AVX2-FP-NEXT: vbroadcastsd 368(%rcx), %ymm3
16948 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
16949 ; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16950 ; AVX2-FP-NEXT: vmovaps 352(%r8), %ymm2
16951 ; AVX2-FP-NEXT: vmovaps 352(%r9), %ymm3
16952 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
16953 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],mem[2,3]
16954 ; AVX2-FP-NEXT: vbroadcastsd 368(%rax), %ymm5
16955 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
16956 ; AVX2-FP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16957 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
16958 ; AVX2-FP-NEXT: vbroadcastsd 376(%rdx), %ymm1
16959 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
16960 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16961 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
16962 ; AVX2-FP-NEXT: vbroadcastsd 376(%r10), %ymm1
16963 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
16964 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16965 ; AVX2-FP-NEXT: vmovaps 384(%rdi), %ymm0
16966 ; AVX2-FP-NEXT: vmovaps 384(%rsi), %ymm1
16967 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
16968 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
16969 ; AVX2-FP-NEXT: vbroadcastsd 400(%rcx), %ymm3
16970 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
16971 ; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16972 ; AVX2-FP-NEXT: vmovaps 384(%r8), %ymm2
16973 ; AVX2-FP-NEXT: vmovaps 384(%r9), %ymm3
16974 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
16975 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],mem[2,3]
16976 ; AVX2-FP-NEXT: vbroadcastsd 400(%rax), %ymm5
16977 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
16978 ; AVX2-FP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16979 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
16980 ; AVX2-FP-NEXT: vbroadcastsd 408(%rdx), %ymm1
16981 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
16982 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16983 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
16984 ; AVX2-FP-NEXT: vbroadcastsd 408(%r10), %ymm1
16985 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
16986 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16987 ; AVX2-FP-NEXT: vmovaps 416(%rdi), %ymm0
16988 ; AVX2-FP-NEXT: vmovaps 416(%rsi), %ymm1
16989 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
16990 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
16991 ; AVX2-FP-NEXT: vbroadcastsd 432(%rcx), %ymm3
16992 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
16993 ; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
16994 ; AVX2-FP-NEXT: vmovaps 416(%r8), %ymm2
16995 ; AVX2-FP-NEXT: vmovaps 416(%r9), %ymm3
16996 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
16997 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],mem[2,3]
16998 ; AVX2-FP-NEXT: vbroadcastsd 432(%rax), %ymm5
16999 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
17000 ; AVX2-FP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17001 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
17002 ; AVX2-FP-NEXT: vbroadcastsd 440(%rdx), %ymm1
17003 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
17004 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17005 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
17006 ; AVX2-FP-NEXT: vbroadcastsd 440(%r10), %ymm1
17007 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
17008 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17009 ; AVX2-FP-NEXT: vmovaps 448(%rdi), %ymm0
17010 ; AVX2-FP-NEXT: vmovaps 448(%rsi), %ymm1
17011 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
17012 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
17013 ; AVX2-FP-NEXT: vbroadcastsd 464(%rcx), %ymm3
17014 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
17015 ; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17016 ; AVX2-FP-NEXT: vmovaps 448(%r8), %ymm2
17017 ; AVX2-FP-NEXT: vmovaps 448(%r9), %ymm3
17018 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
17019 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],mem[2,3]
17020 ; AVX2-FP-NEXT: vbroadcastsd 464(%rax), %ymm5
17021 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
17022 ; AVX2-FP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17023 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
17024 ; AVX2-FP-NEXT: vbroadcastsd 472(%rdx), %ymm1
17025 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
17026 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17027 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
17028 ; AVX2-FP-NEXT: vbroadcastsd 472(%r10), %ymm1
17029 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
17030 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17031 ; AVX2-FP-NEXT: vmovaps 480(%rdi), %ymm0
17032 ; AVX2-FP-NEXT: vmovaps 480(%rsi), %ymm1
17033 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
17034 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
17035 ; AVX2-FP-NEXT: vbroadcastsd 496(%rcx), %ymm3
17036 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
17037 ; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17038 ; AVX2-FP-NEXT: vmovaps 480(%r8), %ymm3
17039 ; AVX2-FP-NEXT: vmovaps 480(%r9), %ymm4
17040 ; AVX2-FP-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm3[0],ymm4[0],ymm3[2],ymm4[2]
17041 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
17042 ; AVX2-FP-NEXT: vbroadcastsd 496(%rax), %ymm5
17043 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm15 = ymm2[0,1,2,3,4,5],ymm5[6,7]
17044 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
17045 ; AVX2-FP-NEXT: vbroadcastsd 504(%rdx), %ymm1
17046 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm11 = ymm0[2,3],ymm1[2,3]
17047 ; AVX2-FP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm3[1],ymm4[1],ymm3[3],ymm4[3]
17048 ; AVX2-FP-NEXT: vbroadcastsd 504(%r10), %ymm1
17049 ; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm9 = ymm0[2,3],ymm1[2,3]
17050 ; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %rdx
17051 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm6[0,1,2,3,4,5],mem[6,7]
17052 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17053 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm7[0,1,2,3,4,5],mem[6,7]
17054 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17055 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm8[0,1,2,3,4,5],mem[6,7]
17056 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17057 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm10[0,1,2,3,4,5],mem[6,7]
17058 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17059 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm12[0,1,2,3,4,5],mem[6,7]
17060 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17061 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm13[0,1,2,3,4,5],mem[6,7]
17062 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17063 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm14[0,1,2,3,4,5],mem[6,7]
17064 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17065 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17066 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],mem[6,7]
17067 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17068 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17069 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],mem[6,7]
17070 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17071 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17072 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],mem[6,7]
17073 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17074 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17075 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],mem[6,7]
17076 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17077 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17078 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],mem[6,7]
17079 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17080 ; AVX2-FP-NEXT: vmovups (%rsp), %ymm0 # 32-byte Reload
17081 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],mem[6,7]
17082 ; AVX2-FP-NEXT: vmovups %ymm0, (%rsp) # 32-byte Spill
17083 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17084 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],mem[6,7]
17085 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17086 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17087 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],mem[6,7]
17088 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17089 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17090 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],mem[6,7]
17091 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17092 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17093 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],mem[6,7]
17094 ; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17095 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17096 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm14 = ymm0[0,1,2,3,4,5],mem[6,7]
17097 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17098 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm13 = ymm0[0,1,2,3,4,5],mem[6,7]
17099 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17100 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm12 = ymm0[0,1,2,3,4,5],mem[6,7]
17101 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17102 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm10 = ymm0[0,1,2,3,4,5],mem[6,7]
17103 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17104 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm8 = ymm0[0,1,2,3,4,5],mem[6,7]
17105 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17106 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm7 = ymm0[0,1,2,3,4,5],mem[6,7]
17107 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17108 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm6 = ymm0[0,1,2,3,4,5],mem[6,7]
17109 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17110 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm5 = ymm0[0,1,2,3,4,5],mem[6,7]
17111 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17112 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm4 = ymm0[0,1,2,3,4,5],mem[6,7]
17113 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17114 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm0[0,1,2,3,4,5],mem[6,7]
17115 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17116 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm0[0,1,2,3,4,5],mem[6,7]
17117 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17118 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm0[0,1,2,3,4,5],mem[6,7]
17119 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17120 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],mem[6,7]
17121 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm11 = ymm11[0,1,2,3,4,5],mem[6,7]
17122 ; AVX2-FP-NEXT: vblendps {{.*#+}} ymm9 = ymm9[0,1,2,3,4,5],mem[6,7]
17123 ; AVX2-FP-NEXT: vmovaps %ymm9, 4064(%rdx)
17124 ; AVX2-FP-NEXT: vmovaps %ymm11, 4032(%rdx)
17125 ; AVX2-FP-NEXT: vmovaps %ymm15, 4000(%rdx)
17126 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload
17127 ; AVX2-FP-NEXT: vmovaps %ymm9, 3968(%rdx)
17128 ; AVX2-FP-NEXT: vmovaps %ymm0, 3808(%rdx)
17129 ; AVX2-FP-NEXT: vmovaps %ymm1, 3776(%rdx)
17130 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17131 ; AVX2-FP-NEXT: vmovaps %ymm0, 3744(%rdx)
17132 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17133 ; AVX2-FP-NEXT: vmovaps %ymm0, 3712(%rdx)
17134 ; AVX2-FP-NEXT: vmovaps %ymm2, 3552(%rdx)
17135 ; AVX2-FP-NEXT: vmovaps %ymm3, 3520(%rdx)
17136 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17137 ; AVX2-FP-NEXT: vmovaps %ymm0, 3488(%rdx)
17138 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17139 ; AVX2-FP-NEXT: vmovaps %ymm0, 3456(%rdx)
17140 ; AVX2-FP-NEXT: vmovaps %ymm4, 3296(%rdx)
17141 ; AVX2-FP-NEXT: vmovaps %ymm5, 3264(%rdx)
17142 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17143 ; AVX2-FP-NEXT: vmovaps %ymm0, 3232(%rdx)
17144 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17145 ; AVX2-FP-NEXT: vmovaps %ymm0, 3200(%rdx)
17146 ; AVX2-FP-NEXT: vmovaps %ymm6, 3040(%rdx)
17147 ; AVX2-FP-NEXT: vmovaps %ymm7, 3008(%rdx)
17148 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17149 ; AVX2-FP-NEXT: vmovaps %ymm0, 2976(%rdx)
17150 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17151 ; AVX2-FP-NEXT: vmovaps %ymm0, 2944(%rdx)
17152 ; AVX2-FP-NEXT: vmovaps %ymm8, 2784(%rdx)
17153 ; AVX2-FP-NEXT: vmovaps %ymm10, 2752(%rdx)
17154 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17155 ; AVX2-FP-NEXT: vmovaps %ymm0, 2720(%rdx)
17156 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17157 ; AVX2-FP-NEXT: vmovaps %ymm0, 2688(%rdx)
17158 ; AVX2-FP-NEXT: vmovaps %ymm12, 2528(%rdx)
17159 ; AVX2-FP-NEXT: vmovaps %ymm13, 2496(%rdx)
17160 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17161 ; AVX2-FP-NEXT: vmovaps %ymm0, 2464(%rdx)
17162 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17163 ; AVX2-FP-NEXT: vmovaps %ymm0, 2432(%rdx)
17164 ; AVX2-FP-NEXT: vmovaps %ymm14, 2272(%rdx)
17165 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17166 ; AVX2-FP-NEXT: vmovaps %ymm0, 2240(%rdx)
17167 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17168 ; AVX2-FP-NEXT: vmovaps %ymm0, 2208(%rdx)
17169 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17170 ; AVX2-FP-NEXT: vmovaps %ymm0, 2176(%rdx)
17171 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17172 ; AVX2-FP-NEXT: vmovaps %ymm0, 2016(%rdx)
17173 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17174 ; AVX2-FP-NEXT: vmovaps %ymm0, 1984(%rdx)
17175 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17176 ; AVX2-FP-NEXT: vmovaps %ymm0, 1952(%rdx)
17177 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17178 ; AVX2-FP-NEXT: vmovaps %ymm0, 1920(%rdx)
17179 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17180 ; AVX2-FP-NEXT: vmovaps %ymm0, 1760(%rdx)
17181 ; AVX2-FP-NEXT: vmovups (%rsp), %ymm0 # 32-byte Reload
17182 ; AVX2-FP-NEXT: vmovaps %ymm0, 1728(%rdx)
17183 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17184 ; AVX2-FP-NEXT: vmovaps %ymm0, 1696(%rdx)
17185 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17186 ; AVX2-FP-NEXT: vmovaps %ymm0, 1664(%rdx)
17187 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17188 ; AVX2-FP-NEXT: vmovaps %ymm0, 1504(%rdx)
17189 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17190 ; AVX2-FP-NEXT: vmovaps %ymm0, 1472(%rdx)
17191 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17192 ; AVX2-FP-NEXT: vmovaps %ymm0, 1440(%rdx)
17193 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17194 ; AVX2-FP-NEXT: vmovaps %ymm0, 1408(%rdx)
17195 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17196 ; AVX2-FP-NEXT: vmovaps %ymm0, 1248(%rdx)
17197 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17198 ; AVX2-FP-NEXT: vmovaps %ymm0, 1216(%rdx)
17199 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17200 ; AVX2-FP-NEXT: vmovaps %ymm0, 1184(%rdx)
17201 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17202 ; AVX2-FP-NEXT: vmovaps %ymm0, 1152(%rdx)
17203 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17204 ; AVX2-FP-NEXT: vmovaps %ymm0, 992(%rdx)
17205 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17206 ; AVX2-FP-NEXT: vmovaps %ymm0, 960(%rdx)
17207 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17208 ; AVX2-FP-NEXT: vmovaps %ymm0, 928(%rdx)
17209 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17210 ; AVX2-FP-NEXT: vmovaps %ymm0, 896(%rdx)
17211 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17212 ; AVX2-FP-NEXT: vmovaps %ymm0, 736(%rdx)
17213 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17214 ; AVX2-FP-NEXT: vmovaps %ymm0, 704(%rdx)
17215 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17216 ; AVX2-FP-NEXT: vmovaps %ymm0, 672(%rdx)
17217 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17218 ; AVX2-FP-NEXT: vmovaps %ymm0, 640(%rdx)
17219 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17220 ; AVX2-FP-NEXT: vmovaps %ymm0, 480(%rdx)
17221 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17222 ; AVX2-FP-NEXT: vmovaps %ymm0, 448(%rdx)
17223 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17224 ; AVX2-FP-NEXT: vmovaps %ymm0, 416(%rdx)
17225 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17226 ; AVX2-FP-NEXT: vmovaps %ymm0, 384(%rdx)
17227 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17228 ; AVX2-FP-NEXT: vmovaps %ymm0, 224(%rdx)
17229 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17230 ; AVX2-FP-NEXT: vmovaps %ymm0, 192(%rdx)
17231 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17232 ; AVX2-FP-NEXT: vmovaps %ymm0, 160(%rdx)
17233 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17234 ; AVX2-FP-NEXT: vmovaps %ymm0, 128(%rdx)
17235 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17236 ; AVX2-FP-NEXT: vmovaps %ymm0, 3936(%rdx)
17237 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17238 ; AVX2-FP-NEXT: vmovaps %ymm0, 3904(%rdx)
17239 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17240 ; AVX2-FP-NEXT: vmovaps %ymm0, 3872(%rdx)
17241 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17242 ; AVX2-FP-NEXT: vmovaps %ymm0, 3840(%rdx)
17243 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17244 ; AVX2-FP-NEXT: vmovaps %ymm0, 3680(%rdx)
17245 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17246 ; AVX2-FP-NEXT: vmovaps %ymm0, 3648(%rdx)
17247 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17248 ; AVX2-FP-NEXT: vmovaps %ymm0, 3616(%rdx)
17249 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17250 ; AVX2-FP-NEXT: vmovaps %ymm0, 3584(%rdx)
17251 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17252 ; AVX2-FP-NEXT: vmovaps %ymm0, 3424(%rdx)
17253 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17254 ; AVX2-FP-NEXT: vmovaps %ymm0, 3392(%rdx)
17255 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17256 ; AVX2-FP-NEXT: vmovaps %ymm0, 3360(%rdx)
17257 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17258 ; AVX2-FP-NEXT: vmovaps %ymm0, 3328(%rdx)
17259 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17260 ; AVX2-FP-NEXT: vmovaps %ymm0, 3168(%rdx)
17261 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17262 ; AVX2-FP-NEXT: vmovaps %ymm0, 3136(%rdx)
17263 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17264 ; AVX2-FP-NEXT: vmovaps %ymm0, 3104(%rdx)
17265 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17266 ; AVX2-FP-NEXT: vmovaps %ymm0, 3072(%rdx)
17267 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17268 ; AVX2-FP-NEXT: vmovaps %ymm0, 2912(%rdx)
17269 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17270 ; AVX2-FP-NEXT: vmovaps %ymm0, 2880(%rdx)
17271 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17272 ; AVX2-FP-NEXT: vmovaps %ymm0, 2848(%rdx)
17273 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17274 ; AVX2-FP-NEXT: vmovaps %ymm0, 2816(%rdx)
17275 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17276 ; AVX2-FP-NEXT: vmovaps %ymm0, 2656(%rdx)
17277 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17278 ; AVX2-FP-NEXT: vmovaps %ymm0, 2624(%rdx)
17279 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17280 ; AVX2-FP-NEXT: vmovaps %ymm0, 2592(%rdx)
17281 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17282 ; AVX2-FP-NEXT: vmovaps %ymm0, 2560(%rdx)
17283 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17284 ; AVX2-FP-NEXT: vmovaps %ymm0, 2400(%rdx)
17285 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17286 ; AVX2-FP-NEXT: vmovaps %ymm0, 2368(%rdx)
17287 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17288 ; AVX2-FP-NEXT: vmovaps %ymm0, 2336(%rdx)
17289 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17290 ; AVX2-FP-NEXT: vmovaps %ymm0, 2304(%rdx)
17291 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17292 ; AVX2-FP-NEXT: vmovaps %ymm0, 2144(%rdx)
17293 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17294 ; AVX2-FP-NEXT: vmovaps %ymm0, 2112(%rdx)
17295 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17296 ; AVX2-FP-NEXT: vmovaps %ymm0, 2080(%rdx)
17297 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17298 ; AVX2-FP-NEXT: vmovaps %ymm0, 2048(%rdx)
17299 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17300 ; AVX2-FP-NEXT: vmovaps %ymm0, 1888(%rdx)
17301 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17302 ; AVX2-FP-NEXT: vmovaps %ymm0, 1856(%rdx)
17303 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17304 ; AVX2-FP-NEXT: vmovaps %ymm0, 1824(%rdx)
17305 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17306 ; AVX2-FP-NEXT: vmovaps %ymm0, 1792(%rdx)
17307 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17308 ; AVX2-FP-NEXT: vmovaps %ymm0, 1632(%rdx)
17309 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17310 ; AVX2-FP-NEXT: vmovaps %ymm0, 1600(%rdx)
17311 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17312 ; AVX2-FP-NEXT: vmovaps %ymm0, 1568(%rdx)
17313 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17314 ; AVX2-FP-NEXT: vmovaps %ymm0, 1536(%rdx)
17315 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17316 ; AVX2-FP-NEXT: vmovaps %ymm0, 1376(%rdx)
17317 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17318 ; AVX2-FP-NEXT: vmovaps %ymm0, 1344(%rdx)
17319 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17320 ; AVX2-FP-NEXT: vmovaps %ymm0, 1312(%rdx)
17321 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17322 ; AVX2-FP-NEXT: vmovaps %ymm0, 1280(%rdx)
17323 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17324 ; AVX2-FP-NEXT: vmovaps %ymm0, 1120(%rdx)
17325 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17326 ; AVX2-FP-NEXT: vmovaps %ymm0, 1088(%rdx)
17327 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17328 ; AVX2-FP-NEXT: vmovaps %ymm0, 1056(%rdx)
17329 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17330 ; AVX2-FP-NEXT: vmovaps %ymm0, 1024(%rdx)
17331 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17332 ; AVX2-FP-NEXT: vmovaps %ymm0, 864(%rdx)
17333 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17334 ; AVX2-FP-NEXT: vmovaps %ymm0, 832(%rdx)
17335 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17336 ; AVX2-FP-NEXT: vmovaps %ymm0, 800(%rdx)
17337 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17338 ; AVX2-FP-NEXT: vmovaps %ymm0, 768(%rdx)
17339 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17340 ; AVX2-FP-NEXT: vmovaps %ymm0, 608(%rdx)
17341 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17342 ; AVX2-FP-NEXT: vmovaps %ymm0, 576(%rdx)
17343 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17344 ; AVX2-FP-NEXT: vmovaps %ymm0, 544(%rdx)
17345 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17346 ; AVX2-FP-NEXT: vmovaps %ymm0, 512(%rdx)
17347 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17348 ; AVX2-FP-NEXT: vmovaps %ymm0, 352(%rdx)
17349 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17350 ; AVX2-FP-NEXT: vmovaps %ymm0, 320(%rdx)
17351 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17352 ; AVX2-FP-NEXT: vmovaps %ymm0, 288(%rdx)
17353 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17354 ; AVX2-FP-NEXT: vmovaps %ymm0, 256(%rdx)
17355 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17356 ; AVX2-FP-NEXT: vmovaps %ymm0, 96(%rdx)
17357 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17358 ; AVX2-FP-NEXT: vmovaps %ymm0, 64(%rdx)
17359 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17360 ; AVX2-FP-NEXT: vmovaps %ymm0, 32(%rdx)
17361 ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
17362 ; AVX2-FP-NEXT: vmovaps %ymm0, (%rdx)
17363 ; AVX2-FP-NEXT: addq $3880, %rsp # imm = 0xF28
17364 ; AVX2-FP-NEXT: vzeroupper
17365 ; AVX2-FP-NEXT: retq
17366 ;
17367 ; AVX2-FCP-LABEL: store_i64_stride8_vf64:
17368 ; AVX2-FCP: # %bb.0:
17369 ; AVX2-FCP-NEXT: subq $3880, %rsp # imm = 0xF28
17370 ; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10
17371 ; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
17372 ; AVX2-FCP-NEXT: vmovaps (%rcx), %xmm0
17373 ; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17374 ; AVX2-FCP-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0
17375 ; AVX2-FCP-NEXT: vmovaps (%rsi), %xmm2
17376 ; AVX2-FCP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17377 ; AVX2-FCP-NEXT: vmovaps 32(%rsi), %xmm3
17378 ; AVX2-FCP-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17379 ; AVX2-FCP-NEXT: vmovaps (%rdi), %xmm1
17380 ; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17381 ; AVX2-FCP-NEXT: vmovaps 32(%rdi), %xmm4
17382 ; AVX2-FCP-NEXT: vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17383 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm2[1]
17384 ; AVX2-FCP-NEXT: vbroadcastsd 8(%rdx), %ymm2
17385 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
17386 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
17387 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17388 ; AVX2-FCP-NEXT: vmovaps (%rax), %xmm0
17389 ; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17390 ; AVX2-FCP-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0
17391 ; AVX2-FCP-NEXT: vmovaps (%r9), %xmm2
17392 ; AVX2-FCP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17393 ; AVX2-FCP-NEXT: vmovaps (%r8), %xmm1
17394 ; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17395 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm2[1]
17396 ; AVX2-FCP-NEXT: vbroadcastsd 8(%r10), %ymm2
17397 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
17398 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
17399 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17400 ; AVX2-FCP-NEXT: vmovaps 32(%rcx), %xmm0
17401 ; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17402 ; AVX2-FCP-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0
17403 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm1 = xmm4[1],xmm3[1]
17404 ; AVX2-FCP-NEXT: vbroadcastsd 40(%rdx), %ymm2
17405 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
17406 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
17407 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17408 ; AVX2-FCP-NEXT: vmovaps 32(%r9), %xmm1
17409 ; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17410 ; AVX2-FCP-NEXT: vmovaps 32(%r8), %xmm0
17411 ; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17412 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
17413 ; AVX2-FCP-NEXT: vbroadcastsd 40(%r10), %ymm1
17414 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
17415 ; AVX2-FCP-NEXT: vmovaps 32(%rax), %xmm1
17416 ; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17417 ; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
17418 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
17419 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17420 ; AVX2-FCP-NEXT: vmovaps 64(%rsi), %xmm1
17421 ; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17422 ; AVX2-FCP-NEXT: vmovaps 64(%rdi), %xmm0
17423 ; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17424 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
17425 ; AVX2-FCP-NEXT: vbroadcastsd 72(%rdx), %ymm1
17426 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
17427 ; AVX2-FCP-NEXT: vmovaps 64(%rcx), %xmm1
17428 ; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17429 ; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
17430 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
17431 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17432 ; AVX2-FCP-NEXT: vmovaps 64(%r9), %xmm1
17433 ; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17434 ; AVX2-FCP-NEXT: vmovaps 64(%r8), %xmm0
17435 ; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17436 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
17437 ; AVX2-FCP-NEXT: vbroadcastsd 72(%r10), %ymm1
17438 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
17439 ; AVX2-FCP-NEXT: vmovaps 64(%rax), %xmm1
17440 ; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17441 ; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
17442 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
17443 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17444 ; AVX2-FCP-NEXT: vmovaps 96(%rsi), %xmm1
17445 ; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17446 ; AVX2-FCP-NEXT: vmovaps 96(%rdi), %xmm0
17447 ; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17448 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
17449 ; AVX2-FCP-NEXT: vbroadcastsd 104(%rdx), %ymm1
17450 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
17451 ; AVX2-FCP-NEXT: vmovaps 96(%rcx), %xmm1
17452 ; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17453 ; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
17454 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
17455 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17456 ; AVX2-FCP-NEXT: vmovaps 96(%r9), %xmm1
17457 ; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17458 ; AVX2-FCP-NEXT: vmovaps 96(%r8), %xmm0
17459 ; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17460 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
17461 ; AVX2-FCP-NEXT: vbroadcastsd 104(%r10), %ymm1
17462 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
17463 ; AVX2-FCP-NEXT: vmovaps 96(%rax), %xmm1
17464 ; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17465 ; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
17466 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
17467 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17468 ; AVX2-FCP-NEXT: vmovaps 128(%rsi), %xmm1
17469 ; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17470 ; AVX2-FCP-NEXT: vmovaps 128(%rdi), %xmm0
17471 ; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17472 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
17473 ; AVX2-FCP-NEXT: vbroadcastsd 136(%rdx), %ymm1
17474 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
17475 ; AVX2-FCP-NEXT: vmovaps 128(%rcx), %xmm1
17476 ; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17477 ; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
17478 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
17479 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17480 ; AVX2-FCP-NEXT: vmovaps 128(%r9), %xmm1
17481 ; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17482 ; AVX2-FCP-NEXT: vmovaps 128(%r8), %xmm0
17483 ; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17484 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
17485 ; AVX2-FCP-NEXT: vbroadcastsd 136(%r10), %ymm1
17486 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
17487 ; AVX2-FCP-NEXT: vmovaps 128(%rax), %xmm1
17488 ; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17489 ; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
17490 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
17491 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17492 ; AVX2-FCP-NEXT: vmovaps 160(%rsi), %xmm1
17493 ; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17494 ; AVX2-FCP-NEXT: vmovaps 160(%rdi), %xmm0
17495 ; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17496 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
17497 ; AVX2-FCP-NEXT: vbroadcastsd 168(%rdx), %ymm1
17498 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
17499 ; AVX2-FCP-NEXT: vmovaps 160(%rcx), %xmm1
17500 ; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17501 ; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
17502 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
17503 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17504 ; AVX2-FCP-NEXT: vmovaps 160(%r9), %xmm1
17505 ; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17506 ; AVX2-FCP-NEXT: vmovaps 160(%r8), %xmm0
17507 ; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17508 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
17509 ; AVX2-FCP-NEXT: vbroadcastsd 168(%r10), %ymm1
17510 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
17511 ; AVX2-FCP-NEXT: vmovaps 160(%rax), %xmm1
17512 ; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17513 ; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
17514 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
17515 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17516 ; AVX2-FCP-NEXT: vmovaps 192(%rsi), %xmm1
17517 ; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17518 ; AVX2-FCP-NEXT: vmovaps 192(%rdi), %xmm0
17519 ; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17520 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
17521 ; AVX2-FCP-NEXT: vbroadcastsd 200(%rdx), %ymm1
17522 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
17523 ; AVX2-FCP-NEXT: vmovaps 192(%rcx), %xmm1
17524 ; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17525 ; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
17526 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
17527 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17528 ; AVX2-FCP-NEXT: vmovaps 192(%r9), %xmm1
17529 ; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17530 ; AVX2-FCP-NEXT: vmovaps 192(%r8), %xmm0
17531 ; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17532 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
17533 ; AVX2-FCP-NEXT: vbroadcastsd 200(%r10), %ymm1
17534 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
17535 ; AVX2-FCP-NEXT: vmovaps 192(%rax), %xmm1
17536 ; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17537 ; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
17538 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
17539 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17540 ; AVX2-FCP-NEXT: vmovaps 224(%rsi), %xmm1
17541 ; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17542 ; AVX2-FCP-NEXT: vmovaps 224(%rdi), %xmm0
17543 ; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17544 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
17545 ; AVX2-FCP-NEXT: vbroadcastsd 232(%rdx), %ymm1
17546 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
17547 ; AVX2-FCP-NEXT: vmovaps 224(%rcx), %xmm1
17548 ; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17549 ; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
17550 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
17551 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17552 ; AVX2-FCP-NEXT: vmovaps 224(%r9), %xmm1
17553 ; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17554 ; AVX2-FCP-NEXT: vmovaps 224(%r8), %xmm0
17555 ; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17556 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
17557 ; AVX2-FCP-NEXT: vbroadcastsd 232(%r10), %ymm1
17558 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
17559 ; AVX2-FCP-NEXT: vmovaps 224(%rax), %xmm1
17560 ; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17561 ; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
17562 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
17563 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17564 ; AVX2-FCP-NEXT: vmovaps 256(%rsi), %xmm1
17565 ; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17566 ; AVX2-FCP-NEXT: vmovaps 256(%rdi), %xmm0
17567 ; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17568 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
17569 ; AVX2-FCP-NEXT: vbroadcastsd 264(%rdx), %ymm1
17570 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
17571 ; AVX2-FCP-NEXT: vmovaps 256(%rcx), %xmm1
17572 ; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17573 ; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
17574 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
17575 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17576 ; AVX2-FCP-NEXT: vmovaps 256(%r9), %xmm1
17577 ; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17578 ; AVX2-FCP-NEXT: vmovaps 256(%r8), %xmm0
17579 ; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17580 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
17581 ; AVX2-FCP-NEXT: vbroadcastsd 264(%r10), %ymm1
17582 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
17583 ; AVX2-FCP-NEXT: vmovaps 256(%rax), %xmm1
17584 ; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17585 ; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
17586 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
17587 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17588 ; AVX2-FCP-NEXT: vmovaps 288(%rsi), %xmm1
17589 ; AVX2-FCP-NEXT: vmovaps %xmm1, (%rsp) # 16-byte Spill
17590 ; AVX2-FCP-NEXT: vmovaps 288(%rdi), %xmm0
17591 ; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17592 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
17593 ; AVX2-FCP-NEXT: vbroadcastsd 296(%rdx), %ymm1
17594 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
17595 ; AVX2-FCP-NEXT: vmovaps 288(%rcx), %xmm1
17596 ; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17597 ; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
17598 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
17599 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17600 ; AVX2-FCP-NEXT: vmovaps 288(%r9), %xmm1
17601 ; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17602 ; AVX2-FCP-NEXT: vmovaps 288(%r8), %xmm0
17603 ; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17604 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
17605 ; AVX2-FCP-NEXT: vbroadcastsd 296(%r10), %ymm1
17606 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
17607 ; AVX2-FCP-NEXT: vmovaps 288(%rax), %xmm1
17608 ; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17609 ; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
17610 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
17611 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17612 ; AVX2-FCP-NEXT: vmovaps 320(%rsi), %xmm1
17613 ; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17614 ; AVX2-FCP-NEXT: vmovaps 320(%rdi), %xmm0
17615 ; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17616 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
17617 ; AVX2-FCP-NEXT: vbroadcastsd 328(%rdx), %ymm1
17618 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
17619 ; AVX2-FCP-NEXT: vmovaps 320(%rcx), %xmm1
17620 ; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17621 ; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
17622 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
17623 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17624 ; AVX2-FCP-NEXT: vmovaps 320(%r9), %xmm1
17625 ; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17626 ; AVX2-FCP-NEXT: vmovaps 320(%r8), %xmm0
17627 ; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17628 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
17629 ; AVX2-FCP-NEXT: vbroadcastsd 328(%r10), %ymm1
17630 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
17631 ; AVX2-FCP-NEXT: vmovaps 320(%rax), %xmm1
17632 ; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17633 ; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
17634 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
17635 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17636 ; AVX2-FCP-NEXT: vmovaps 352(%rsi), %xmm1
17637 ; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17638 ; AVX2-FCP-NEXT: vmovaps 352(%rdi), %xmm0
17639 ; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17640 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
17641 ; AVX2-FCP-NEXT: vbroadcastsd 360(%rdx), %ymm1
17642 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
17643 ; AVX2-FCP-NEXT: vmovaps 352(%rcx), %xmm1
17644 ; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17645 ; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
17646 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
17647 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17648 ; AVX2-FCP-NEXT: vmovaps 352(%r9), %xmm1
17649 ; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17650 ; AVX2-FCP-NEXT: vmovaps 352(%r8), %xmm0
17651 ; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17652 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
17653 ; AVX2-FCP-NEXT: vbroadcastsd 360(%r10), %ymm1
17654 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
17655 ; AVX2-FCP-NEXT: vmovaps 352(%rax), %xmm1
17656 ; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17657 ; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
17658 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
17659 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17660 ; AVX2-FCP-NEXT: vmovaps 384(%rsi), %xmm1
17661 ; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17662 ; AVX2-FCP-NEXT: vmovaps 384(%rdi), %xmm0
17663 ; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17664 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
17665 ; AVX2-FCP-NEXT: vbroadcastsd 392(%rdx), %ymm1
17666 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
17667 ; AVX2-FCP-NEXT: vmovaps 384(%rcx), %xmm1
17668 ; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17669 ; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
17670 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
17671 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17672 ; AVX2-FCP-NEXT: vmovaps 384(%r9), %xmm1
17673 ; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17674 ; AVX2-FCP-NEXT: vmovaps 384(%r8), %xmm0
17675 ; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17676 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
17677 ; AVX2-FCP-NEXT: vbroadcastsd 392(%r10), %ymm1
17678 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
17679 ; AVX2-FCP-NEXT: vmovaps 384(%rax), %xmm1
17680 ; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17681 ; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
17682 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
17683 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17684 ; AVX2-FCP-NEXT: vmovaps 416(%rsi), %xmm1
17685 ; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17686 ; AVX2-FCP-NEXT: vmovaps 416(%rdi), %xmm0
17687 ; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17688 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
17689 ; AVX2-FCP-NEXT: vbroadcastsd 424(%rdx), %ymm1
17690 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
17691 ; AVX2-FCP-NEXT: vmovaps 416(%rcx), %xmm1
17692 ; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17693 ; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm1
17694 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
17695 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17696 ; AVX2-FCP-NEXT: vmovaps 416(%r9), %xmm0
17697 ; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
17698 ; AVX2-FCP-NEXT: vmovaps 416(%r8), %xmm13
17699 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm13[1],xmm0[1]
17700 ; AVX2-FCP-NEXT: vbroadcastsd 424(%r10), %ymm1
17701 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
17702 ; AVX2-FCP-NEXT: vmovaps 416(%rax), %xmm12
17703 ; AVX2-FCP-NEXT: vinsertf128 $1, %xmm12, %ymm0, %ymm1
17704 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
17705 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17706 ; AVX2-FCP-NEXT: vmovaps 448(%rsi), %xmm11
17707 ; AVX2-FCP-NEXT: vmovaps 448(%rdi), %xmm10
17708 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm10[1],xmm11[1]
17709 ; AVX2-FCP-NEXT: vbroadcastsd 456(%rdx), %ymm1
17710 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
17711 ; AVX2-FCP-NEXT: vmovaps 448(%rcx), %xmm9
17712 ; AVX2-FCP-NEXT: vinsertf128 $1, %xmm9, %ymm0, %ymm1
17713 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
17714 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17715 ; AVX2-FCP-NEXT: vmovaps 448(%r9), %xmm8
17716 ; AVX2-FCP-NEXT: vmovaps 448(%r8), %xmm7
17717 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm7[1],xmm8[1]
17718 ; AVX2-FCP-NEXT: vbroadcastsd 456(%r10), %ymm1
17719 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
17720 ; AVX2-FCP-NEXT: vmovaps 448(%rax), %xmm6
17721 ; AVX2-FCP-NEXT: vinsertf128 $1, %xmm6, %ymm0, %ymm1
17722 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
17723 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17724 ; AVX2-FCP-NEXT: vmovaps 480(%rsi), %xmm5
17725 ; AVX2-FCP-NEXT: vmovaps 480(%rdi), %xmm4
17726 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm4[1],xmm5[1]
17727 ; AVX2-FCP-NEXT: vbroadcastsd 488(%rdx), %ymm1
17728 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
17729 ; AVX2-FCP-NEXT: vmovaps 480(%rcx), %xmm3
17730 ; AVX2-FCP-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm1
17731 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
17732 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17733 ; AVX2-FCP-NEXT: vmovaps 480(%r9), %xmm2
17734 ; AVX2-FCP-NEXT: vmovaps 480(%r8), %xmm1
17735 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm2[1]
17736 ; AVX2-FCP-NEXT: vbroadcastsd 488(%r10), %ymm15
17737 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm15 = ymm0[0,1,2,3],ymm15[4,5,6,7]
17738 ; AVX2-FCP-NEXT: vmovaps 480(%rax), %xmm0
17739 ; AVX2-FCP-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm14
17740 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm14 = ymm15[0,1,2,3,4,5],ymm14[6,7]
17741 ; AVX2-FCP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17742 ; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
17743 ; AVX2-FCP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
17744 ; AVX2-FCP-NEXT: # xmm14 = xmm14[0],mem[0]
17745 ; AVX2-FCP-NEXT: vinsertf128 $1, (%rdx), %ymm14, %ymm14
17746 ; AVX2-FCP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
17747 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
17748 ; AVX2-FCP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17749 ; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
17750 ; AVX2-FCP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
17751 ; AVX2-FCP-NEXT: # xmm14 = xmm14[0],mem[0]
17752 ; AVX2-FCP-NEXT: vinsertf128 $1, (%r10), %ymm14, %ymm14
17753 ; AVX2-FCP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
17754 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
17755 ; AVX2-FCP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17756 ; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
17757 ; AVX2-FCP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
17758 ; AVX2-FCP-NEXT: # xmm14 = xmm14[0],mem[0]
17759 ; AVX2-FCP-NEXT: vinsertf128 $1, 32(%rdx), %ymm14, %ymm14
17760 ; AVX2-FCP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
17761 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
17762 ; AVX2-FCP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17763 ; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
17764 ; AVX2-FCP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
17765 ; AVX2-FCP-NEXT: # xmm14 = xmm14[0],mem[0]
17766 ; AVX2-FCP-NEXT: vinsertf128 $1, 32(%r10), %ymm14, %ymm14
17767 ; AVX2-FCP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
17768 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
17769 ; AVX2-FCP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17770 ; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
17771 ; AVX2-FCP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
17772 ; AVX2-FCP-NEXT: # xmm14 = xmm14[0],mem[0]
17773 ; AVX2-FCP-NEXT: vinsertf128 $1, 64(%rdx), %ymm14, %ymm14
17774 ; AVX2-FCP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
17775 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
17776 ; AVX2-FCP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17777 ; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
17778 ; AVX2-FCP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
17779 ; AVX2-FCP-NEXT: # xmm14 = xmm14[0],mem[0]
17780 ; AVX2-FCP-NEXT: vinsertf128 $1, 64(%r10), %ymm14, %ymm14
17781 ; AVX2-FCP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
17782 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
17783 ; AVX2-FCP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17784 ; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
17785 ; AVX2-FCP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
17786 ; AVX2-FCP-NEXT: # xmm14 = xmm14[0],mem[0]
17787 ; AVX2-FCP-NEXT: vinsertf128 $1, 96(%rdx), %ymm14, %ymm14
17788 ; AVX2-FCP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
17789 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
17790 ; AVX2-FCP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17791 ; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
17792 ; AVX2-FCP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
17793 ; AVX2-FCP-NEXT: # xmm14 = xmm14[0],mem[0]
17794 ; AVX2-FCP-NEXT: vinsertf128 $1, 96(%r10), %ymm14, %ymm14
17795 ; AVX2-FCP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
17796 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
17797 ; AVX2-FCP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17798 ; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
17799 ; AVX2-FCP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
17800 ; AVX2-FCP-NEXT: # xmm14 = xmm14[0],mem[0]
17801 ; AVX2-FCP-NEXT: vinsertf128 $1, 128(%rdx), %ymm14, %ymm14
17802 ; AVX2-FCP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
17803 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
17804 ; AVX2-FCP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17805 ; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
17806 ; AVX2-FCP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
17807 ; AVX2-FCP-NEXT: # xmm14 = xmm14[0],mem[0]
17808 ; AVX2-FCP-NEXT: vinsertf128 $1, 128(%r10), %ymm14, %ymm14
17809 ; AVX2-FCP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
17810 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
17811 ; AVX2-FCP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17812 ; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
17813 ; AVX2-FCP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
17814 ; AVX2-FCP-NEXT: # xmm14 = xmm14[0],mem[0]
17815 ; AVX2-FCP-NEXT: vinsertf128 $1, 160(%rdx), %ymm14, %ymm14
17816 ; AVX2-FCP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
17817 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
17818 ; AVX2-FCP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17819 ; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
17820 ; AVX2-FCP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
17821 ; AVX2-FCP-NEXT: # xmm14 = xmm14[0],mem[0]
17822 ; AVX2-FCP-NEXT: vinsertf128 $1, 160(%r10), %ymm14, %ymm14
17823 ; AVX2-FCP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
17824 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
17825 ; AVX2-FCP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17826 ; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
17827 ; AVX2-FCP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
17828 ; AVX2-FCP-NEXT: # xmm14 = xmm14[0],mem[0]
17829 ; AVX2-FCP-NEXT: vinsertf128 $1, 192(%rdx), %ymm14, %ymm14
17830 ; AVX2-FCP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
17831 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
17832 ; AVX2-FCP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17833 ; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
17834 ; AVX2-FCP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
17835 ; AVX2-FCP-NEXT: # xmm14 = xmm14[0],mem[0]
17836 ; AVX2-FCP-NEXT: vinsertf128 $1, 192(%r10), %ymm14, %ymm14
17837 ; AVX2-FCP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
17838 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
17839 ; AVX2-FCP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17840 ; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
17841 ; AVX2-FCP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
17842 ; AVX2-FCP-NEXT: # xmm14 = xmm14[0],mem[0]
17843 ; AVX2-FCP-NEXT: vinsertf128 $1, 224(%rdx), %ymm14, %ymm14
17844 ; AVX2-FCP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
17845 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
17846 ; AVX2-FCP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17847 ; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
17848 ; AVX2-FCP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
17849 ; AVX2-FCP-NEXT: # xmm14 = xmm14[0],mem[0]
17850 ; AVX2-FCP-NEXT: vinsertf128 $1, 224(%r10), %ymm14, %ymm14
17851 ; AVX2-FCP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
17852 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
17853 ; AVX2-FCP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17854 ; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
17855 ; AVX2-FCP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
17856 ; AVX2-FCP-NEXT: # xmm14 = xmm14[0],mem[0]
17857 ; AVX2-FCP-NEXT: vinsertf128 $1, 256(%rdx), %ymm14, %ymm14
17858 ; AVX2-FCP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
17859 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
17860 ; AVX2-FCP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17861 ; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
17862 ; AVX2-FCP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
17863 ; AVX2-FCP-NEXT: # xmm14 = xmm14[0],mem[0]
17864 ; AVX2-FCP-NEXT: vinsertf128 $1, 256(%r10), %ymm14, %ymm14
17865 ; AVX2-FCP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
17866 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
17867 ; AVX2-FCP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17868 ; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
17869 ; AVX2-FCP-NEXT: vunpcklpd (%rsp), %xmm14, %xmm14 # 16-byte Folded Reload
17870 ; AVX2-FCP-NEXT: # xmm14 = xmm14[0],mem[0]
17871 ; AVX2-FCP-NEXT: vinsertf128 $1, 288(%rdx), %ymm14, %ymm14
17872 ; AVX2-FCP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
17873 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
17874 ; AVX2-FCP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17875 ; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
17876 ; AVX2-FCP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
17877 ; AVX2-FCP-NEXT: # xmm14 = xmm14[0],mem[0]
17878 ; AVX2-FCP-NEXT: vinsertf128 $1, 288(%r10), %ymm14, %ymm14
17879 ; AVX2-FCP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
17880 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
17881 ; AVX2-FCP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17882 ; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
17883 ; AVX2-FCP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
17884 ; AVX2-FCP-NEXT: # xmm14 = xmm14[0],mem[0]
17885 ; AVX2-FCP-NEXT: vinsertf128 $1, 320(%rdx), %ymm14, %ymm14
17886 ; AVX2-FCP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
17887 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
17888 ; AVX2-FCP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17889 ; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
17890 ; AVX2-FCP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
17891 ; AVX2-FCP-NEXT: # xmm14 = xmm14[0],mem[0]
17892 ; AVX2-FCP-NEXT: vinsertf128 $1, 320(%r10), %ymm14, %ymm14
17893 ; AVX2-FCP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
17894 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
17895 ; AVX2-FCP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17896 ; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
17897 ; AVX2-FCP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
17898 ; AVX2-FCP-NEXT: # xmm14 = xmm14[0],mem[0]
17899 ; AVX2-FCP-NEXT: vinsertf128 $1, 352(%rdx), %ymm14, %ymm14
17900 ; AVX2-FCP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
17901 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
17902 ; AVX2-FCP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17903 ; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
17904 ; AVX2-FCP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
17905 ; AVX2-FCP-NEXT: # xmm14 = xmm14[0],mem[0]
17906 ; AVX2-FCP-NEXT: vinsertf128 $1, 352(%r10), %ymm14, %ymm14
17907 ; AVX2-FCP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
17908 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
17909 ; AVX2-FCP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17910 ; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
17911 ; AVX2-FCP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
17912 ; AVX2-FCP-NEXT: # xmm14 = xmm14[0],mem[0]
17913 ; AVX2-FCP-NEXT: vinsertf128 $1, 384(%rdx), %ymm14, %ymm14
17914 ; AVX2-FCP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
17915 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
17916 ; AVX2-FCP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17917 ; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
17918 ; AVX2-FCP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
17919 ; AVX2-FCP-NEXT: # xmm14 = xmm14[0],mem[0]
17920 ; AVX2-FCP-NEXT: vinsertf128 $1, 384(%r10), %ymm14, %ymm14
17921 ; AVX2-FCP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
17922 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
17923 ; AVX2-FCP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17924 ; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
17925 ; AVX2-FCP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm14 # 16-byte Folded Reload
17926 ; AVX2-FCP-NEXT: # xmm14 = xmm14[0],mem[0]
17927 ; AVX2-FCP-NEXT: vinsertf128 $1, 416(%rdx), %ymm14, %ymm14
17928 ; AVX2-FCP-NEXT: vbroadcastsd {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 16-byte Folded Reload
17929 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm15[6,7]
17930 ; AVX2-FCP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17931 ; AVX2-FCP-NEXT: vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm13, %xmm13 # 16-byte Folded Reload
17932 ; AVX2-FCP-NEXT: # xmm13 = xmm13[0],mem[0]
17933 ; AVX2-FCP-NEXT: vinsertf128 $1, 416(%r10), %ymm13, %ymm13
17934 ; AVX2-FCP-NEXT: vbroadcastsd %xmm12, %ymm12
17935 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm12 = ymm13[0,1,2,3,4,5],ymm12[6,7]
17936 ; AVX2-FCP-NEXT: vmovups %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17937 ; AVX2-FCP-NEXT: vmovlhps {{.*#+}} xmm10 = xmm10[0],xmm11[0]
17938 ; AVX2-FCP-NEXT: vinsertf128 $1, 448(%rdx), %ymm10, %ymm10
17939 ; AVX2-FCP-NEXT: vbroadcastsd %xmm9, %ymm9
17940 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm9 = ymm10[0,1,2,3,4,5],ymm9[6,7]
17941 ; AVX2-FCP-NEXT: vmovups %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17942 ; AVX2-FCP-NEXT: vmovlhps {{.*#+}} xmm7 = xmm7[0],xmm8[0]
17943 ; AVX2-FCP-NEXT: vinsertf128 $1, 448(%r10), %ymm7, %ymm7
17944 ; AVX2-FCP-NEXT: vbroadcastsd %xmm6, %ymm6
17945 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3,4,5],ymm6[6,7]
17946 ; AVX2-FCP-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17947 ; AVX2-FCP-NEXT: vmovlhps {{.*#+}} xmm4 = xmm4[0],xmm5[0]
17948 ; AVX2-FCP-NEXT: vinsertf128 $1, 480(%rdx), %ymm4, %ymm4
17949 ; AVX2-FCP-NEXT: vbroadcastsd %xmm3, %ymm3
17950 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm3 = ymm4[0,1,2,3,4,5],ymm3[6,7]
17951 ; AVX2-FCP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17952 ; AVX2-FCP-NEXT: vmovlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
17953 ; AVX2-FCP-NEXT: vinsertf128 $1, 480(%r10), %ymm1, %ymm1
17954 ; AVX2-FCP-NEXT: vbroadcastsd %xmm0, %ymm0
17955 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
17956 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17957 ; AVX2-FCP-NEXT: vmovaps (%rdi), %ymm0
17958 ; AVX2-FCP-NEXT: vmovaps (%rsi), %ymm1
17959 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
17960 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
17961 ; AVX2-FCP-NEXT: vbroadcastsd 16(%rcx), %ymm3
17962 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
17963 ; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17964 ; AVX2-FCP-NEXT: vmovaps (%r8), %ymm2
17965 ; AVX2-FCP-NEXT: vmovaps (%r9), %ymm3
17966 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
17967 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],mem[2,3]
17968 ; AVX2-FCP-NEXT: vbroadcastsd 16(%rax), %ymm5
17969 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
17970 ; AVX2-FCP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17971 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
17972 ; AVX2-FCP-NEXT: vbroadcastsd 24(%rdx), %ymm1
17973 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm6 = ymm0[2,3],ymm1[2,3]
17974 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
17975 ; AVX2-FCP-NEXT: vbroadcastsd 24(%r10), %ymm1
17976 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm7 = ymm0[2,3],ymm1[2,3]
17977 ; AVX2-FCP-NEXT: vmovaps 32(%rdi), %ymm0
17978 ; AVX2-FCP-NEXT: vmovaps 32(%rsi), %ymm1
17979 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
17980 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
17981 ; AVX2-FCP-NEXT: vbroadcastsd 48(%rcx), %ymm3
17982 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
17983 ; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17984 ; AVX2-FCP-NEXT: vmovaps 32(%r8), %ymm2
17985 ; AVX2-FCP-NEXT: vmovaps 32(%r9), %ymm3
17986 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
17987 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],mem[2,3]
17988 ; AVX2-FCP-NEXT: vbroadcastsd 48(%rax), %ymm5
17989 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
17990 ; AVX2-FCP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
17991 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
17992 ; AVX2-FCP-NEXT: vbroadcastsd 56(%rdx), %ymm1
17993 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm8 = ymm0[2,3],ymm1[2,3]
17994 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
17995 ; AVX2-FCP-NEXT: vbroadcastsd 56(%r10), %ymm1
17996 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm10 = ymm0[2,3],ymm1[2,3]
17997 ; AVX2-FCP-NEXT: vmovaps 64(%rdi), %ymm0
17998 ; AVX2-FCP-NEXT: vmovaps 64(%rsi), %ymm1
17999 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
18000 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
18001 ; AVX2-FCP-NEXT: vbroadcastsd 80(%rcx), %ymm3
18002 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
18003 ; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
18004 ; AVX2-FCP-NEXT: vmovaps 64(%r8), %ymm2
18005 ; AVX2-FCP-NEXT: vmovaps 64(%r9), %ymm3
18006 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
18007 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],mem[2,3]
18008 ; AVX2-FCP-NEXT: vbroadcastsd 80(%rax), %ymm5
18009 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
18010 ; AVX2-FCP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
18011 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
18012 ; AVX2-FCP-NEXT: vbroadcastsd 88(%rdx), %ymm1
18013 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm12 = ymm0[2,3],ymm1[2,3]
18014 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
18015 ; AVX2-FCP-NEXT: vbroadcastsd 88(%r10), %ymm1
18016 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm13 = ymm0[2,3],ymm1[2,3]
18017 ; AVX2-FCP-NEXT: vmovaps 96(%rdi), %ymm0
18018 ; AVX2-FCP-NEXT: vmovaps 96(%rsi), %ymm1
18019 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
18020 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
18021 ; AVX2-FCP-NEXT: vbroadcastsd 112(%rcx), %ymm3
18022 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
18023 ; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
18024 ; AVX2-FCP-NEXT: vmovaps 96(%r8), %ymm2
18025 ; AVX2-FCP-NEXT: vmovaps 96(%r9), %ymm3
18026 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
18027 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],mem[2,3]
18028 ; AVX2-FCP-NEXT: vbroadcastsd 112(%rax), %ymm5
18029 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
18030 ; AVX2-FCP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
18031 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
18032 ; AVX2-FCP-NEXT: vbroadcastsd 120(%rdx), %ymm1
18033 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm14 = ymm0[2,3],ymm1[2,3]
18034 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
18035 ; AVX2-FCP-NEXT: vbroadcastsd 120(%r10), %ymm1
18036 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
18037 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
18038 ; AVX2-FCP-NEXT: vmovaps 128(%rdi), %ymm0
18039 ; AVX2-FCP-NEXT: vmovaps 128(%rsi), %ymm1
18040 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
18041 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
18042 ; AVX2-FCP-NEXT: vbroadcastsd 144(%rcx), %ymm3
18043 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
18044 ; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
18045 ; AVX2-FCP-NEXT: vmovaps 128(%r8), %ymm2
18046 ; AVX2-FCP-NEXT: vmovaps 128(%r9), %ymm3
18047 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
18048 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],mem[2,3]
18049 ; AVX2-FCP-NEXT: vbroadcastsd 144(%rax), %ymm5
18050 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
18051 ; AVX2-FCP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
18052 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
18053 ; AVX2-FCP-NEXT: vbroadcastsd 152(%rdx), %ymm1
18054 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
18055 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
18056 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
18057 ; AVX2-FCP-NEXT: vbroadcastsd 152(%r10), %ymm1
18058 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
18059 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
18060 ; AVX2-FCP-NEXT: vmovaps 160(%rdi), %ymm0
18061 ; AVX2-FCP-NEXT: vmovaps 160(%rsi), %ymm1
18062 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
18063 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
18064 ; AVX2-FCP-NEXT: vbroadcastsd 176(%rcx), %ymm3
18065 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
18066 ; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
18067 ; AVX2-FCP-NEXT: vmovaps 160(%r8), %ymm2
18068 ; AVX2-FCP-NEXT: vmovaps 160(%r9), %ymm3
18069 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
18070 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],mem[2,3]
18071 ; AVX2-FCP-NEXT: vbroadcastsd 176(%rax), %ymm5
18072 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
18073 ; AVX2-FCP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
18074 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
18075 ; AVX2-FCP-NEXT: vbroadcastsd 184(%rdx), %ymm1
18076 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
18077 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
18078 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
18079 ; AVX2-FCP-NEXT: vbroadcastsd 184(%r10), %ymm1
18080 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
18081 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
18082 ; AVX2-FCP-NEXT: vmovaps 192(%rdi), %ymm0
18083 ; AVX2-FCP-NEXT: vmovaps 192(%rsi), %ymm1
18084 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
18085 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
18086 ; AVX2-FCP-NEXT: vbroadcastsd 208(%rcx), %ymm3
18087 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
18088 ; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
18089 ; AVX2-FCP-NEXT: vmovaps 192(%r8), %ymm2
18090 ; AVX2-FCP-NEXT: vmovaps 192(%r9), %ymm3
18091 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
18092 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],mem[2,3]
18093 ; AVX2-FCP-NEXT: vbroadcastsd 208(%rax), %ymm5
18094 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
18095 ; AVX2-FCP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
18096 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
18097 ; AVX2-FCP-NEXT: vbroadcastsd 216(%rdx), %ymm1
18098 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
18099 ; AVX2-FCP-NEXT: vmovups %ymm0, (%rsp) # 32-byte Spill
18100 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
18101 ; AVX2-FCP-NEXT: vbroadcastsd 216(%r10), %ymm1
18102 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
18103 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
18104 ; AVX2-FCP-NEXT: vmovaps 224(%rdi), %ymm0
18105 ; AVX2-FCP-NEXT: vmovaps 224(%rsi), %ymm1
18106 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
18107 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
18108 ; AVX2-FCP-NEXT: vbroadcastsd 240(%rcx), %ymm3
18109 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
18110 ; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
18111 ; AVX2-FCP-NEXT: vmovaps 224(%r8), %ymm2
18112 ; AVX2-FCP-NEXT: vmovaps 224(%r9), %ymm3
18113 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
18114 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],mem[2,3]
18115 ; AVX2-FCP-NEXT: vbroadcastsd 240(%rax), %ymm5
18116 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
18117 ; AVX2-FCP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
18118 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
18119 ; AVX2-FCP-NEXT: vbroadcastsd 248(%rdx), %ymm1
18120 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
18121 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
18122 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
18123 ; AVX2-FCP-NEXT: vbroadcastsd 248(%r10), %ymm1
18124 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
18125 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
18126 ; AVX2-FCP-NEXT: vmovaps 256(%rdi), %ymm0
18127 ; AVX2-FCP-NEXT: vmovaps 256(%rsi), %ymm1
18128 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
18129 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
18130 ; AVX2-FCP-NEXT: vbroadcastsd 272(%rcx), %ymm3
18131 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
18132 ; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
18133 ; AVX2-FCP-NEXT: vmovaps 256(%r8), %ymm2
18134 ; AVX2-FCP-NEXT: vmovaps 256(%r9), %ymm3
18135 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
18136 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],mem[2,3]
18137 ; AVX2-FCP-NEXT: vbroadcastsd 272(%rax), %ymm5
18138 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
18139 ; AVX2-FCP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
18140 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
18141 ; AVX2-FCP-NEXT: vbroadcastsd 280(%rdx), %ymm1
18142 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
18143 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
18144 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
18145 ; AVX2-FCP-NEXT: vbroadcastsd 280(%r10), %ymm1
18146 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
18147 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
18148 ; AVX2-FCP-NEXT: vmovaps 288(%rdi), %ymm0
18149 ; AVX2-FCP-NEXT: vmovaps 288(%rsi), %ymm1
18150 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
18151 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
18152 ; AVX2-FCP-NEXT: vbroadcastsd 304(%rcx), %ymm3
18153 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
18154 ; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
18155 ; AVX2-FCP-NEXT: vmovaps 288(%r8), %ymm2
18156 ; AVX2-FCP-NEXT: vmovaps 288(%r9), %ymm3
18157 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
18158 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],mem[2,3]
18159 ; AVX2-FCP-NEXT: vbroadcastsd 304(%rax), %ymm5
18160 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
18161 ; AVX2-FCP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
18162 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
18163 ; AVX2-FCP-NEXT: vbroadcastsd 312(%rdx), %ymm1
18164 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
18165 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
18166 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
18167 ; AVX2-FCP-NEXT: vbroadcastsd 312(%r10), %ymm1
18168 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
18169 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
18170 ; AVX2-FCP-NEXT: vmovaps 320(%rdi), %ymm0
18171 ; AVX2-FCP-NEXT: vmovaps 320(%rsi), %ymm1
18172 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
18173 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
18174 ; AVX2-FCP-NEXT: vbroadcastsd 336(%rcx), %ymm3
18175 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
18176 ; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
18177 ; AVX2-FCP-NEXT: vmovaps 320(%r8), %ymm2
18178 ; AVX2-FCP-NEXT: vmovaps 320(%r9), %ymm3
18179 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
18180 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],mem[2,3]
18181 ; AVX2-FCP-NEXT: vbroadcastsd 336(%rax), %ymm5
18182 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
18183 ; AVX2-FCP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
18184 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
18185 ; AVX2-FCP-NEXT: vbroadcastsd 344(%rdx), %ymm1
18186 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
18187 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
18188 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
18189 ; AVX2-FCP-NEXT: vbroadcastsd 344(%r10), %ymm1
18190 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
18191 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
18192 ; AVX2-FCP-NEXT: vmovaps 352(%rdi), %ymm0
18193 ; AVX2-FCP-NEXT: vmovaps 352(%rsi), %ymm1
18194 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
18195 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
18196 ; AVX2-FCP-NEXT: vbroadcastsd 368(%rcx), %ymm3
18197 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
18198 ; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
18199 ; AVX2-FCP-NEXT: vmovaps 352(%r8), %ymm2
18200 ; AVX2-FCP-NEXT: vmovaps 352(%r9), %ymm3
18201 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
18202 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],mem[2,3]
18203 ; AVX2-FCP-NEXT: vbroadcastsd 368(%rax), %ymm5
18204 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
18205 ; AVX2-FCP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
18206 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
18207 ; AVX2-FCP-NEXT: vbroadcastsd 376(%rdx), %ymm1
18208 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
18209 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
18210 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
18211 ; AVX2-FCP-NEXT: vbroadcastsd 376(%r10), %ymm1
18212 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
18213 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
18214 ; AVX2-FCP-NEXT: vmovaps 384(%rdi), %ymm0
18215 ; AVX2-FCP-NEXT: vmovaps 384(%rsi), %ymm1
18216 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
18217 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
18218 ; AVX2-FCP-NEXT: vbroadcastsd 400(%rcx), %ymm3
18219 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
18220 ; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
18221 ; AVX2-FCP-NEXT: vmovaps 384(%r8), %ymm2
18222 ; AVX2-FCP-NEXT: vmovaps 384(%r9), %ymm3
18223 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
18224 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],mem[2,3]
18225 ; AVX2-FCP-NEXT: vbroadcastsd 400(%rax), %ymm5
18226 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
18227 ; AVX2-FCP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
18228 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
18229 ; AVX2-FCP-NEXT: vbroadcastsd 408(%rdx), %ymm1
18230 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
18231 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
18232 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
18233 ; AVX2-FCP-NEXT: vbroadcastsd 408(%r10), %ymm1
18234 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
18235 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
18236 ; AVX2-FCP-NEXT: vmovaps 416(%rdi), %ymm0
18237 ; AVX2-FCP-NEXT: vmovaps 416(%rsi), %ymm1
18238 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
18239 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
18240 ; AVX2-FCP-NEXT: vbroadcastsd 432(%rcx), %ymm3
18241 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
18242 ; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
18243 ; AVX2-FCP-NEXT: vmovaps 416(%r8), %ymm2
18244 ; AVX2-FCP-NEXT: vmovaps 416(%r9), %ymm3
18245 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
18246 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],mem[2,3]
18247 ; AVX2-FCP-NEXT: vbroadcastsd 432(%rax), %ymm5
18248 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
18249 ; AVX2-FCP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
18250 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
18251 ; AVX2-FCP-NEXT: vbroadcastsd 440(%rdx), %ymm1
18252 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
18253 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
18254 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
18255 ; AVX2-FCP-NEXT: vbroadcastsd 440(%r10), %ymm1
18256 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
18257 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
18258 ; AVX2-FCP-NEXT: vmovaps 448(%rdi), %ymm0
18259 ; AVX2-FCP-NEXT: vmovaps 448(%rsi), %ymm1
18260 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
18261 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
18262 ; AVX2-FCP-NEXT: vbroadcastsd 464(%rcx), %ymm3
18263 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
18264 ; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
18265 ; AVX2-FCP-NEXT: vmovaps 448(%r8), %ymm2
18266 ; AVX2-FCP-NEXT: vmovaps 448(%r9), %ymm3
18267 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]
18268 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm4 = ymm4[2,3],mem[2,3]
18269 ; AVX2-FCP-NEXT: vbroadcastsd 464(%rax), %ymm5
18270 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm5[6,7]
18271 ; AVX2-FCP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
18272 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
18273 ; AVX2-FCP-NEXT: vbroadcastsd 472(%rdx), %ymm1
18274 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
18275 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
18276 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]
18277 ; AVX2-FCP-NEXT: vbroadcastsd 472(%r10), %ymm1
18278 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]
18279 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
18280 ; AVX2-FCP-NEXT: vmovaps 480(%rdi), %ymm0
18281 ; AVX2-FCP-NEXT: vmovaps 480(%rsi), %ymm1
18282 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[2],ymm1[2]
18283 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
18284 ; AVX2-FCP-NEXT: vbroadcastsd 496(%rcx), %ymm3
18285 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
18286 ; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
18287 ; AVX2-FCP-NEXT: vmovaps 480(%r8), %ymm3
18288 ; AVX2-FCP-NEXT: vmovaps 480(%r9), %ymm4
18289 ; AVX2-FCP-NEXT: vunpcklpd {{.*#+}} ymm2 = ymm3[0],ymm4[0],ymm3[2],ymm4[2]
18290 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3],mem[2,3]
18291 ; AVX2-FCP-NEXT: vbroadcastsd 496(%rax), %ymm5
18292 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm15 = ymm2[0,1,2,3,4,5],ymm5[6,7]
18293 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[3],ymm1[3]
18294 ; AVX2-FCP-NEXT: vbroadcastsd 504(%rdx), %ymm1
18295 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm11 = ymm0[2,3],ymm1[2,3]
18296 ; AVX2-FCP-NEXT: vunpckhpd {{.*#+}} ymm0 = ymm3[1],ymm4[1],ymm3[3],ymm4[3]
18297 ; AVX2-FCP-NEXT: vbroadcastsd 504(%r10), %ymm1
18298 ; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm9 = ymm0[2,3],ymm1[2,3]
18299 ; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rdx
18300 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm6[0,1,2,3,4,5],mem[6,7]
18301 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
18302 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm7[0,1,2,3,4,5],mem[6,7]
18303 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
18304 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm8[0,1,2,3,4,5],mem[6,7]
18305 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
18306 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm10[0,1,2,3,4,5],mem[6,7]
18307 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
18308 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm12[0,1,2,3,4,5],mem[6,7]
18309 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
18310 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm13[0,1,2,3,4,5],mem[6,7]
18311 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
18312 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm14[0,1,2,3,4,5],mem[6,7]
18313 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
18314 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18315 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],mem[6,7]
18316 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
18317 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18318 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],mem[6,7]
18319 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
18320 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18321 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],mem[6,7]
18322 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
18323 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18324 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],mem[6,7]
18325 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
18326 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18327 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],mem[6,7]
18328 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
18329 ; AVX2-FCP-NEXT: vmovups (%rsp), %ymm0 # 32-byte Reload
18330 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],mem[6,7]
18331 ; AVX2-FCP-NEXT: vmovups %ymm0, (%rsp) # 32-byte Spill
18332 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18333 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],mem[6,7]
18334 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
18335 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18336 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],mem[6,7]
18337 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
18338 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18339 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],mem[6,7]
18340 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
18341 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18342 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],mem[6,7]
18343 ; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
18344 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18345 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm14 = ymm0[0,1,2,3,4,5],mem[6,7]
18346 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18347 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm13 = ymm0[0,1,2,3,4,5],mem[6,7]
18348 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18349 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm12 = ymm0[0,1,2,3,4,5],mem[6,7]
18350 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18351 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm10 = ymm0[0,1,2,3,4,5],mem[6,7]
18352 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18353 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm8 = ymm0[0,1,2,3,4,5],mem[6,7]
18354 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18355 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm7 = ymm0[0,1,2,3,4,5],mem[6,7]
18356 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18357 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm6 = ymm0[0,1,2,3,4,5],mem[6,7]
18358 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18359 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm5 = ymm0[0,1,2,3,4,5],mem[6,7]
18360 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18361 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm0[0,1,2,3,4,5],mem[6,7]
18362 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18363 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm3 = ymm0[0,1,2,3,4,5],mem[6,7]
18364 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18365 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm0[0,1,2,3,4,5],mem[6,7]
18366 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18367 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm0[0,1,2,3,4,5],mem[6,7]
18368 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18369 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],mem[6,7]
18370 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm11 = ymm11[0,1,2,3,4,5],mem[6,7]
18371 ; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm9 = ymm9[0,1,2,3,4,5],mem[6,7]
18372 ; AVX2-FCP-NEXT: vmovaps %ymm9, 4064(%rdx)
18373 ; AVX2-FCP-NEXT: vmovaps %ymm11, 4032(%rdx)
18374 ; AVX2-FCP-NEXT: vmovaps %ymm15, 4000(%rdx)
18375 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload
18376 ; AVX2-FCP-NEXT: vmovaps %ymm9, 3968(%rdx)
18377 ; AVX2-FCP-NEXT: vmovaps %ymm0, 3808(%rdx)
18378 ; AVX2-FCP-NEXT: vmovaps %ymm1, 3776(%rdx)
18379 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18380 ; AVX2-FCP-NEXT: vmovaps %ymm0, 3744(%rdx)
18381 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18382 ; AVX2-FCP-NEXT: vmovaps %ymm0, 3712(%rdx)
18383 ; AVX2-FCP-NEXT: vmovaps %ymm2, 3552(%rdx)
18384 ; AVX2-FCP-NEXT: vmovaps %ymm3, 3520(%rdx)
18385 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18386 ; AVX2-FCP-NEXT: vmovaps %ymm0, 3488(%rdx)
18387 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18388 ; AVX2-FCP-NEXT: vmovaps %ymm0, 3456(%rdx)
18389 ; AVX2-FCP-NEXT: vmovaps %ymm4, 3296(%rdx)
18390 ; AVX2-FCP-NEXT: vmovaps %ymm5, 3264(%rdx)
18391 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18392 ; AVX2-FCP-NEXT: vmovaps %ymm0, 3232(%rdx)
18393 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18394 ; AVX2-FCP-NEXT: vmovaps %ymm0, 3200(%rdx)
18395 ; AVX2-FCP-NEXT: vmovaps %ymm6, 3040(%rdx)
18396 ; AVX2-FCP-NEXT: vmovaps %ymm7, 3008(%rdx)
18397 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18398 ; AVX2-FCP-NEXT: vmovaps %ymm0, 2976(%rdx)
18399 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18400 ; AVX2-FCP-NEXT: vmovaps %ymm0, 2944(%rdx)
18401 ; AVX2-FCP-NEXT: vmovaps %ymm8, 2784(%rdx)
18402 ; AVX2-FCP-NEXT: vmovaps %ymm10, 2752(%rdx)
18403 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18404 ; AVX2-FCP-NEXT: vmovaps %ymm0, 2720(%rdx)
18405 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18406 ; AVX2-FCP-NEXT: vmovaps %ymm0, 2688(%rdx)
18407 ; AVX2-FCP-NEXT: vmovaps %ymm12, 2528(%rdx)
18408 ; AVX2-FCP-NEXT: vmovaps %ymm13, 2496(%rdx)
18409 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18410 ; AVX2-FCP-NEXT: vmovaps %ymm0, 2464(%rdx)
18411 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18412 ; AVX2-FCP-NEXT: vmovaps %ymm0, 2432(%rdx)
18413 ; AVX2-FCP-NEXT: vmovaps %ymm14, 2272(%rdx)
18414 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18415 ; AVX2-FCP-NEXT: vmovaps %ymm0, 2240(%rdx)
18416 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18417 ; AVX2-FCP-NEXT: vmovaps %ymm0, 2208(%rdx)
18418 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18419 ; AVX2-FCP-NEXT: vmovaps %ymm0, 2176(%rdx)
18420 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18421 ; AVX2-FCP-NEXT: vmovaps %ymm0, 2016(%rdx)
18422 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18423 ; AVX2-FCP-NEXT: vmovaps %ymm0, 1984(%rdx)
18424 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18425 ; AVX2-FCP-NEXT: vmovaps %ymm0, 1952(%rdx)
18426 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18427 ; AVX2-FCP-NEXT: vmovaps %ymm0, 1920(%rdx)
18428 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18429 ; AVX2-FCP-NEXT: vmovaps %ymm0, 1760(%rdx)
18430 ; AVX2-FCP-NEXT: vmovups (%rsp), %ymm0 # 32-byte Reload
18431 ; AVX2-FCP-NEXT: vmovaps %ymm0, 1728(%rdx)
18432 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18433 ; AVX2-FCP-NEXT: vmovaps %ymm0, 1696(%rdx)
18434 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18435 ; AVX2-FCP-NEXT: vmovaps %ymm0, 1664(%rdx)
18436 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18437 ; AVX2-FCP-NEXT: vmovaps %ymm0, 1504(%rdx)
18438 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18439 ; AVX2-FCP-NEXT: vmovaps %ymm0, 1472(%rdx)
18440 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18441 ; AVX2-FCP-NEXT: vmovaps %ymm0, 1440(%rdx)
18442 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18443 ; AVX2-FCP-NEXT: vmovaps %ymm0, 1408(%rdx)
18444 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18445 ; AVX2-FCP-NEXT: vmovaps %ymm0, 1248(%rdx)
18446 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18447 ; AVX2-FCP-NEXT: vmovaps %ymm0, 1216(%rdx)
18448 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18449 ; AVX2-FCP-NEXT: vmovaps %ymm0, 1184(%rdx)
18450 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18451 ; AVX2-FCP-NEXT: vmovaps %ymm0, 1152(%rdx)
18452 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18453 ; AVX2-FCP-NEXT: vmovaps %ymm0, 992(%rdx)
18454 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18455 ; AVX2-FCP-NEXT: vmovaps %ymm0, 960(%rdx)
18456 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18457 ; AVX2-FCP-NEXT: vmovaps %ymm0, 928(%rdx)
18458 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18459 ; AVX2-FCP-NEXT: vmovaps %ymm0, 896(%rdx)
18460 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18461 ; AVX2-FCP-NEXT: vmovaps %ymm0, 736(%rdx)
18462 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18463 ; AVX2-FCP-NEXT: vmovaps %ymm0, 704(%rdx)
18464 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18465 ; AVX2-FCP-NEXT: vmovaps %ymm0, 672(%rdx)
18466 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18467 ; AVX2-FCP-NEXT: vmovaps %ymm0, 640(%rdx)
18468 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18469 ; AVX2-FCP-NEXT: vmovaps %ymm0, 480(%rdx)
18470 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18471 ; AVX2-FCP-NEXT: vmovaps %ymm0, 448(%rdx)
18472 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18473 ; AVX2-FCP-NEXT: vmovaps %ymm0, 416(%rdx)
18474 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18475 ; AVX2-FCP-NEXT: vmovaps %ymm0, 384(%rdx)
18476 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18477 ; AVX2-FCP-NEXT: vmovaps %ymm0, 224(%rdx)
18478 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18479 ; AVX2-FCP-NEXT: vmovaps %ymm0, 192(%rdx)
18480 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18481 ; AVX2-FCP-NEXT: vmovaps %ymm0, 160(%rdx)
18482 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18483 ; AVX2-FCP-NEXT: vmovaps %ymm0, 128(%rdx)
18484 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18485 ; AVX2-FCP-NEXT: vmovaps %ymm0, 3936(%rdx)
18486 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18487 ; AVX2-FCP-NEXT: vmovaps %ymm0, 3904(%rdx)
18488 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18489 ; AVX2-FCP-NEXT: vmovaps %ymm0, 3872(%rdx)
18490 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18491 ; AVX2-FCP-NEXT: vmovaps %ymm0, 3840(%rdx)
18492 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18493 ; AVX2-FCP-NEXT: vmovaps %ymm0, 3680(%rdx)
18494 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18495 ; AVX2-FCP-NEXT: vmovaps %ymm0, 3648(%rdx)
18496 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18497 ; AVX2-FCP-NEXT: vmovaps %ymm0, 3616(%rdx)
18498 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18499 ; AVX2-FCP-NEXT: vmovaps %ymm0, 3584(%rdx)
18500 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18501 ; AVX2-FCP-NEXT: vmovaps %ymm0, 3424(%rdx)
18502 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18503 ; AVX2-FCP-NEXT: vmovaps %ymm0, 3392(%rdx)
18504 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18505 ; AVX2-FCP-NEXT: vmovaps %ymm0, 3360(%rdx)
18506 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18507 ; AVX2-FCP-NEXT: vmovaps %ymm0, 3328(%rdx)
18508 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18509 ; AVX2-FCP-NEXT: vmovaps %ymm0, 3168(%rdx)
18510 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18511 ; AVX2-FCP-NEXT: vmovaps %ymm0, 3136(%rdx)
18512 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18513 ; AVX2-FCP-NEXT: vmovaps %ymm0, 3104(%rdx)
18514 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18515 ; AVX2-FCP-NEXT: vmovaps %ymm0, 3072(%rdx)
18516 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18517 ; AVX2-FCP-NEXT: vmovaps %ymm0, 2912(%rdx)
18518 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18519 ; AVX2-FCP-NEXT: vmovaps %ymm0, 2880(%rdx)
18520 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18521 ; AVX2-FCP-NEXT: vmovaps %ymm0, 2848(%rdx)
18522 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18523 ; AVX2-FCP-NEXT: vmovaps %ymm0, 2816(%rdx)
18524 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18525 ; AVX2-FCP-NEXT: vmovaps %ymm0, 2656(%rdx)
18526 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18527 ; AVX2-FCP-NEXT: vmovaps %ymm0, 2624(%rdx)
18528 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18529 ; AVX2-FCP-NEXT: vmovaps %ymm0, 2592(%rdx)
18530 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18531 ; AVX2-FCP-NEXT: vmovaps %ymm0, 2560(%rdx)
18532 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18533 ; AVX2-FCP-NEXT: vmovaps %ymm0, 2400(%rdx)
18534 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18535 ; AVX2-FCP-NEXT: vmovaps %ymm0, 2368(%rdx)
18536 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18537 ; AVX2-FCP-NEXT: vmovaps %ymm0, 2336(%rdx)
18538 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18539 ; AVX2-FCP-NEXT: vmovaps %ymm0, 2304(%rdx)
18540 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18541 ; AVX2-FCP-NEXT: vmovaps %ymm0, 2144(%rdx)
18542 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18543 ; AVX2-FCP-NEXT: vmovaps %ymm0, 2112(%rdx)
18544 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18545 ; AVX2-FCP-NEXT: vmovaps %ymm0, 2080(%rdx)
18546 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18547 ; AVX2-FCP-NEXT: vmovaps %ymm0, 2048(%rdx)
18548 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18549 ; AVX2-FCP-NEXT: vmovaps %ymm0, 1888(%rdx)
18550 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18551 ; AVX2-FCP-NEXT: vmovaps %ymm0, 1856(%rdx)
18552 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18553 ; AVX2-FCP-NEXT: vmovaps %ymm0, 1824(%rdx)
18554 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18555 ; AVX2-FCP-NEXT: vmovaps %ymm0, 1792(%rdx)
18556 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18557 ; AVX2-FCP-NEXT: vmovaps %ymm0, 1632(%rdx)
18558 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18559 ; AVX2-FCP-NEXT: vmovaps %ymm0, 1600(%rdx)
18560 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18561 ; AVX2-FCP-NEXT: vmovaps %ymm0, 1568(%rdx)
18562 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18563 ; AVX2-FCP-NEXT: vmovaps %ymm0, 1536(%rdx)
18564 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18565 ; AVX2-FCP-NEXT: vmovaps %ymm0, 1376(%rdx)
18566 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18567 ; AVX2-FCP-NEXT: vmovaps %ymm0, 1344(%rdx)
18568 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18569 ; AVX2-FCP-NEXT: vmovaps %ymm0, 1312(%rdx)
18570 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18571 ; AVX2-FCP-NEXT: vmovaps %ymm0, 1280(%rdx)
18572 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18573 ; AVX2-FCP-NEXT: vmovaps %ymm0, 1120(%rdx)
18574 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18575 ; AVX2-FCP-NEXT: vmovaps %ymm0, 1088(%rdx)
18576 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18577 ; AVX2-FCP-NEXT: vmovaps %ymm0, 1056(%rdx)
18578 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18579 ; AVX2-FCP-NEXT: vmovaps %ymm0, 1024(%rdx)
18580 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18581 ; AVX2-FCP-NEXT: vmovaps %ymm0, 864(%rdx)
18582 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18583 ; AVX2-FCP-NEXT: vmovaps %ymm0, 832(%rdx)
18584 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18585 ; AVX2-FCP-NEXT: vmovaps %ymm0, 800(%rdx)
18586 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18587 ; AVX2-FCP-NEXT: vmovaps %ymm0, 768(%rdx)
18588 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18589 ; AVX2-FCP-NEXT: vmovaps %ymm0, 608(%rdx)
18590 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18591 ; AVX2-FCP-NEXT: vmovaps %ymm0, 576(%rdx)
18592 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18593 ; AVX2-FCP-NEXT: vmovaps %ymm0, 544(%rdx)
18594 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18595 ; AVX2-FCP-NEXT: vmovaps %ymm0, 512(%rdx)
18596 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18597 ; AVX2-FCP-NEXT: vmovaps %ymm0, 352(%rdx)
18598 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18599 ; AVX2-FCP-NEXT: vmovaps %ymm0, 320(%rdx)
18600 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18601 ; AVX2-FCP-NEXT: vmovaps %ymm0, 288(%rdx)
18602 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18603 ; AVX2-FCP-NEXT: vmovaps %ymm0, 256(%rdx)
18604 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18605 ; AVX2-FCP-NEXT: vmovaps %ymm0, 96(%rdx)
18606 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18607 ; AVX2-FCP-NEXT: vmovaps %ymm0, 64(%rdx)
18608 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18609 ; AVX2-FCP-NEXT: vmovaps %ymm0, 32(%rdx)
18610 ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
18611 ; AVX2-FCP-NEXT: vmovaps %ymm0, (%rdx)
18612 ; AVX2-FCP-NEXT: addq $3880, %rsp # imm = 0xF28
18613 ; AVX2-FCP-NEXT: vzeroupper
18614 ; AVX2-FCP-NEXT: retq
18615 ;
18616 ; AVX512-LABEL: store_i64_stride8_vf64:
18617 ; AVX512: # %bb.0:
18618 ; AVX512-NEXT: subq $5384, %rsp # imm = 0x1508
18619 ; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %rax
18620 ; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %r10
18621 ; AVX512-NEXT: vmovdqa64 128(%rdi), %zmm0
18622 ; AVX512-NEXT: vmovdqa64 64(%rdi), %zmm4
18623 ; AVX512-NEXT: vmovdqa64 (%rdi), %zmm14
18624 ; AVX512-NEXT: vmovdqa64 128(%rsi), %zmm1
18625 ; AVX512-NEXT: vmovdqa64 64(%rsi), %zmm5
18626 ; AVX512-NEXT: vmovdqa64 (%rsi), %zmm10
18627 ; AVX512-NEXT: vmovdqa64 128(%rdx), %zmm2
18628 ; AVX512-NEXT: vmovdqa64 64(%rdx), %zmm7
18629 ; AVX512-NEXT: vmovdqa64 (%rdx), %zmm12
18630 ; AVX512-NEXT: vmovdqa64 128(%rcx), %zmm3
18631 ; AVX512-NEXT: vmovdqa64 64(%rcx), %zmm9
18632 ; AVX512-NEXT: vmovdqa64 (%rcx), %zmm15
18633 ; AVX512-NEXT: vmovdqa64 (%r8), %zmm30
18634 ; AVX512-NEXT: vmovdqa64 64(%r8), %zmm18
18635 ; AVX512-NEXT: vmovdqa64 128(%r8), %zmm11
18636 ; AVX512-NEXT: vmovdqa64 (%r9), %zmm24
18637 ; AVX512-NEXT: vmovdqa64 64(%r9), %zmm28
18638 ; AVX512-NEXT: vmovdqa64 128(%r9), %zmm22
18639 ; AVX512-NEXT: vmovdqa64 (%r10), %zmm26
18640 ; AVX512-NEXT: vmovdqa64 64(%r10), %zmm31
18641 ; AVX512-NEXT: vmovdqa64 128(%r10), %zmm16
18642 ; AVX512-NEXT: vmovdqa64 (%rax), %zmm17
18643 ; AVX512-NEXT: vmovdqa64 64(%rax), %zmm27
18644 ; AVX512-NEXT: vmovdqa64 128(%rax), %zmm13
18645 ; AVX512-NEXT: movb $-64, %r11b
18646 ; AVX512-NEXT: kmovw %r11d, %k1
18647 ; AVX512-NEXT: vbroadcasti32x4 {{.*#+}} zmm19 = [4,12,4,12,4,12,4,12]
18648 ; AVX512-NEXT: # zmm19 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
18649 ; AVX512-NEXT: vmovdqa64 %zmm12, %zmm8
18650 ; AVX512-NEXT: vpermt2q %zmm15, %zmm19, %zmm8
18651 ; AVX512-NEXT: vmovdqa64 %zmm14, %zmm6
18652 ; AVX512-NEXT: vpermt2q %zmm10, %zmm19, %zmm6
18653 ; AVX512-NEXT: vpblendd {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm8[4,5,6,7]
18654 ; AVX512-NEXT: vmovdqa64 %zmm26, %zmm8
18655 ; AVX512-NEXT: vpermt2q %zmm17, %zmm19, %zmm8
18656 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} zmm20 = zmm30[0],zmm24[0],zmm30[2],zmm24[2],zmm30[4],zmm24[4],zmm30[6],zmm24[6]
18657 ; AVX512-NEXT: vmovdqa64 %zmm8, %zmm20 {%k1}
18658 ; AVX512-NEXT: vinserti64x4 $0, %ymm6, %zmm20, %zmm6
18659 ; AVX512-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18660 ; AVX512-NEXT: vbroadcasti32x4 {{.*#+}} zmm21 = [5,13,5,13,5,13,5,13]
18661 ; AVX512-NEXT: # zmm21 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
18662 ; AVX512-NEXT: vmovdqa64 %zmm12, %zmm6
18663 ; AVX512-NEXT: vpermt2q %zmm15, %zmm21, %zmm6
18664 ; AVX512-NEXT: vmovdqa64 %zmm14, %zmm8
18665 ; AVX512-NEXT: vpermt2q %zmm10, %zmm21, %zmm8
18666 ; AVX512-NEXT: vpblendd {{.*#+}} ymm6 = ymm8[0,1,2,3],ymm6[4,5,6,7]
18667 ; AVX512-NEXT: vmovdqa64 %zmm26, %zmm8
18668 ; AVX512-NEXT: vpermt2q %zmm17, %zmm21, %zmm8
18669 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} zmm20 = zmm30[1],zmm24[1],zmm30[3],zmm24[3],zmm30[5],zmm24[5],zmm30[7],zmm24[7]
18670 ; AVX512-NEXT: vmovdqa64 %zmm8, %zmm20 {%k1}
18671 ; AVX512-NEXT: vinserti64x4 $0, %ymm6, %zmm20, %zmm6
18672 ; AVX512-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18673 ; AVX512-NEXT: vbroadcasti32x4 {{.*#+}} zmm23 = [6,14,6,14,6,14,6,14]
18674 ; AVX512-NEXT: # zmm23 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
18675 ; AVX512-NEXT: vmovdqa64 %zmm12, %zmm6
18676 ; AVX512-NEXT: vpermt2q %zmm15, %zmm23, %zmm6
18677 ; AVX512-NEXT: vmovdqa64 %zmm14, %zmm8
18678 ; AVX512-NEXT: vpermt2q %zmm10, %zmm23, %zmm8
18679 ; AVX512-NEXT: vpblendd {{.*#+}} ymm6 = ymm8[0,1,2,3],ymm6[4,5,6,7]
18680 ; AVX512-NEXT: vmovdqa64 %zmm30, %zmm8
18681 ; AVX512-NEXT: vpermt2q %zmm24, %zmm23, %zmm8
18682 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} zmm8 {%k1} = zmm26[0],zmm17[0],zmm26[2],zmm17[2],zmm26[4],zmm17[4],zmm26[6],zmm17[6]
18683 ; AVX512-NEXT: vinserti64x4 $0, %ymm6, %zmm8, %zmm6
18684 ; AVX512-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18685 ; AVX512-NEXT: vbroadcasti32x4 {{.*#+}} zmm25 = [7,15,7,15,7,15,7,15]
18686 ; AVX512-NEXT: # zmm25 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
18687 ; AVX512-NEXT: vpermt2q %zmm15, %zmm25, %zmm12
18688 ; AVX512-NEXT: vpermt2q %zmm10, %zmm25, %zmm14
18689 ; AVX512-NEXT: vpblendd {{.*#+}} ymm6 = ymm14[0,1,2,3],ymm12[4,5,6,7]
18690 ; AVX512-NEXT: vmovdqa64 %zmm30, %zmm8
18691 ; AVX512-NEXT: vpermt2q %zmm24, %zmm25, %zmm8
18692 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} zmm8 {%k1} = zmm26[1],zmm17[1],zmm26[3],zmm17[3],zmm26[5],zmm17[5],zmm26[7],zmm17[7]
18693 ; AVX512-NEXT: vinserti64x4 $0, %ymm6, %zmm8, %zmm6
18694 ; AVX512-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18695 ; AVX512-NEXT: vmovdqa64 %zmm7, %zmm6
18696 ; AVX512-NEXT: vpermt2q %zmm9, %zmm19, %zmm6
18697 ; AVX512-NEXT: vmovdqa64 %zmm4, %zmm8
18698 ; AVX512-NEXT: vpermt2q %zmm5, %zmm19, %zmm8
18699 ; AVX512-NEXT: vpblendd {{.*#+}} ymm6 = ymm8[0,1,2,3],ymm6[4,5,6,7]
18700 ; AVX512-NEXT: vmovdqa64 %zmm31, %zmm8
18701 ; AVX512-NEXT: vpermt2q %zmm27, %zmm19, %zmm8
18702 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} zmm10 = zmm18[0],zmm28[0],zmm18[2],zmm28[2],zmm18[4],zmm28[4],zmm18[6],zmm28[6]
18703 ; AVX512-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}
18704 ; AVX512-NEXT: vinserti64x4 $0, %ymm6, %zmm10, %zmm6
18705 ; AVX512-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18706 ; AVX512-NEXT: vmovdqa64 %zmm7, %zmm6
18707 ; AVX512-NEXT: vpermt2q %zmm9, %zmm21, %zmm6
18708 ; AVX512-NEXT: vmovdqa64 %zmm4, %zmm8
18709 ; AVX512-NEXT: vpermt2q %zmm5, %zmm21, %zmm8
18710 ; AVX512-NEXT: vpblendd {{.*#+}} ymm6 = ymm8[0,1,2,3],ymm6[4,5,6,7]
18711 ; AVX512-NEXT: vmovdqa64 %zmm31, %zmm8
18712 ; AVX512-NEXT: vpermt2q %zmm27, %zmm21, %zmm8
18713 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} zmm10 = zmm18[1],zmm28[1],zmm18[3],zmm28[3],zmm18[5],zmm28[5],zmm18[7],zmm28[7]
18714 ; AVX512-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}
18715 ; AVX512-NEXT: vinserti64x4 $0, %ymm6, %zmm10, %zmm6
18716 ; AVX512-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18717 ; AVX512-NEXT: vmovdqa64 %zmm7, %zmm6
18718 ; AVX512-NEXT: vpermt2q %zmm9, %zmm23, %zmm6
18719 ; AVX512-NEXT: vmovdqa64 %zmm4, %zmm8
18720 ; AVX512-NEXT: vpermt2q %zmm5, %zmm23, %zmm8
18721 ; AVX512-NEXT: vpblendd {{.*#+}} ymm6 = ymm8[0,1,2,3],ymm6[4,5,6,7]
18722 ; AVX512-NEXT: vmovdqu64 %zmm18, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18723 ; AVX512-NEXT: vmovdqa64 %zmm18, %zmm8
18724 ; AVX512-NEXT: vpermt2q %zmm28, %zmm23, %zmm8
18725 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} zmm8 {%k1} = zmm31[0],zmm27[0],zmm31[2],zmm27[2],zmm31[4],zmm27[4],zmm31[6],zmm27[6]
18726 ; AVX512-NEXT: vinserti64x4 $0, %ymm6, %zmm8, %zmm6
18727 ; AVX512-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18728 ; AVX512-NEXT: vpermt2q %zmm9, %zmm25, %zmm7
18729 ; AVX512-NEXT: vpermt2q %zmm5, %zmm25, %zmm4
18730 ; AVX512-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm7[4,5,6,7]
18731 ; AVX512-NEXT: vmovdqa64 %zmm18, %zmm5
18732 ; AVX512-NEXT: vpermt2q %zmm28, %zmm25, %zmm5
18733 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} zmm5 {%k1} = zmm31[1],zmm27[1],zmm31[3],zmm27[3],zmm31[5],zmm27[5],zmm31[7],zmm27[7]
18734 ; AVX512-NEXT: vinserti64x4 $0, %ymm4, %zmm5, %zmm4
18735 ; AVX512-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18736 ; AVX512-NEXT: vmovdqa64 %zmm2, %zmm4
18737 ; AVX512-NEXT: vpermt2q %zmm3, %zmm19, %zmm4
18738 ; AVX512-NEXT: vmovdqa64 %zmm0, %zmm5
18739 ; AVX512-NEXT: vpermt2q %zmm1, %zmm19, %zmm5
18740 ; AVX512-NEXT: vpblendd {{.*#+}} ymm4 = ymm5[0,1,2,3],ymm4[4,5,6,7]
18741 ; AVX512-NEXT: vmovdqa64 %zmm16, %zmm5
18742 ; AVX512-NEXT: vpermt2q %zmm13, %zmm19, %zmm5
18743 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} zmm6 = zmm11[0],zmm22[0],zmm11[2],zmm22[2],zmm11[4],zmm22[4],zmm11[6],zmm22[6]
18744 ; AVX512-NEXT: vmovdqa64 %zmm5, %zmm6 {%k1}
18745 ; AVX512-NEXT: vinserti64x4 $0, %ymm4, %zmm6, %zmm4
18746 ; AVX512-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18747 ; AVX512-NEXT: vmovdqa64 %zmm2, %zmm4
18748 ; AVX512-NEXT: vpermt2q %zmm3, %zmm21, %zmm4
18749 ; AVX512-NEXT: vmovdqa64 %zmm0, %zmm5
18750 ; AVX512-NEXT: vpermt2q %zmm1, %zmm21, %zmm5
18751 ; AVX512-NEXT: vpblendd {{.*#+}} ymm4 = ymm5[0,1,2,3],ymm4[4,5,6,7]
18752 ; AVX512-NEXT: vmovdqa64 %zmm16, %zmm5
18753 ; AVX512-NEXT: vpermt2q %zmm13, %zmm21, %zmm5
18754 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} zmm6 = zmm11[1],zmm22[1],zmm11[3],zmm22[3],zmm11[5],zmm22[5],zmm11[7],zmm22[7]
18755 ; AVX512-NEXT: vmovdqa64 %zmm5, %zmm6 {%k1}
18756 ; AVX512-NEXT: vinserti64x4 $0, %ymm4, %zmm6, %zmm4
18757 ; AVX512-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18758 ; AVX512-NEXT: vmovdqa64 %zmm2, %zmm4
18759 ; AVX512-NEXT: vpermt2q %zmm3, %zmm23, %zmm4
18760 ; AVX512-NEXT: vmovdqa64 %zmm0, %zmm5
18761 ; AVX512-NEXT: vpermt2q %zmm1, %zmm23, %zmm5
18762 ; AVX512-NEXT: vpblendd {{.*#+}} ymm4 = ymm5[0,1,2,3],ymm4[4,5,6,7]
18763 ; AVX512-NEXT: vmovdqa64 %zmm11, %zmm5
18764 ; AVX512-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18765 ; AVX512-NEXT: vmovdqu64 %zmm22, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18766 ; AVX512-NEXT: vpermt2q %zmm22, %zmm23, %zmm5
18767 ; AVX512-NEXT: vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18768 ; AVX512-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18769 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} zmm5 {%k1} = zmm16[0],zmm13[0],zmm16[2],zmm13[2],zmm16[4],zmm13[4],zmm16[6],zmm13[6]
18770 ; AVX512-NEXT: vinserti64x4 $0, %ymm4, %zmm5, %zmm4
18771 ; AVX512-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18772 ; AVX512-NEXT: vmovdqa64 192(%rdi), %zmm4
18773 ; AVX512-NEXT: vpermt2q %zmm3, %zmm25, %zmm2
18774 ; AVX512-NEXT: vmovdqa64 192(%rdx), %zmm3
18775 ; AVX512-NEXT: vpermt2q %zmm1, %zmm25, %zmm0
18776 ; AVX512-NEXT: vmovdqa64 192(%rcx), %zmm1
18777 ; AVX512-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
18778 ; AVX512-NEXT: vmovdqa64 %zmm11, %zmm2
18779 ; AVX512-NEXT: vpermt2q %zmm22, %zmm25, %zmm2
18780 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} zmm2 {%k1} = zmm16[1],zmm13[1],zmm16[3],zmm13[3],zmm16[5],zmm13[5],zmm16[7],zmm13[7]
18781 ; AVX512-NEXT: vinserti64x4 $0, %ymm0, %zmm2, %zmm0
18782 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18783 ; AVX512-NEXT: vmovdqa64 %zmm3, %zmm0
18784 ; AVX512-NEXT: vpermt2q %zmm1, %zmm19, %zmm0
18785 ; AVX512-NEXT: vmovdqa64 192(%rsi), %zmm2
18786 ; AVX512-NEXT: vmovdqa64 %zmm4, %zmm5
18787 ; AVX512-NEXT: vpermt2q %zmm2, %zmm19, %zmm5
18788 ; AVX512-NEXT: vpblendd {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm0[4,5,6,7]
18789 ; AVX512-NEXT: vmovdqa64 192(%r10), %zmm9
18790 ; AVX512-NEXT: vmovdqa64 192(%rax), %zmm20
18791 ; AVX512-NEXT: vmovdqa64 %zmm9, %zmm6
18792 ; AVX512-NEXT: vpermt2q %zmm20, %zmm19, %zmm6
18793 ; AVX512-NEXT: vmovdqa64 192(%r8), %zmm8
18794 ; AVX512-NEXT: vmovdqa64 192(%r9), %zmm0
18795 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} zmm7 = zmm8[0],zmm0[0],zmm8[2],zmm0[2],zmm8[4],zmm0[4],zmm8[6],zmm0[6]
18796 ; AVX512-NEXT: vmovdqa64 %zmm6, %zmm7 {%k1}
18797 ; AVX512-NEXT: vinserti64x4 $0, %ymm5, %zmm7, %zmm5
18798 ; AVX512-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18799 ; AVX512-NEXT: vmovdqa64 %zmm3, %zmm5
18800 ; AVX512-NEXT: vpermt2q %zmm1, %zmm21, %zmm5
18801 ; AVX512-NEXT: vmovdqa64 %zmm4, %zmm6
18802 ; AVX512-NEXT: vpermt2q %zmm2, %zmm21, %zmm6
18803 ; AVX512-NEXT: vpblendd {{.*#+}} ymm5 = ymm6[0,1,2,3],ymm5[4,5,6,7]
18804 ; AVX512-NEXT: vmovdqa64 %zmm9, %zmm6
18805 ; AVX512-NEXT: vpermt2q %zmm20, %zmm21, %zmm6
18806 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} zmm7 = zmm8[1],zmm0[1],zmm8[3],zmm0[3],zmm8[5],zmm0[5],zmm8[7],zmm0[7]
18807 ; AVX512-NEXT: vmovdqa64 %zmm6, %zmm7 {%k1}
18808 ; AVX512-NEXT: vinserti64x4 $0, %ymm5, %zmm7, %zmm5
18809 ; AVX512-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18810 ; AVX512-NEXT: vmovdqa64 %zmm3, %zmm5
18811 ; AVX512-NEXT: vpermt2q %zmm1, %zmm23, %zmm5
18812 ; AVX512-NEXT: vmovdqa64 %zmm4, %zmm6
18813 ; AVX512-NEXT: vpermt2q %zmm2, %zmm23, %zmm6
18814 ; AVX512-NEXT: vpblendd {{.*#+}} ymm5 = ymm6[0,1,2,3],ymm5[4,5,6,7]
18815 ; AVX512-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18816 ; AVX512-NEXT: vmovdqa64 %zmm8, %zmm6
18817 ; AVX512-NEXT: vpermt2q %zmm0, %zmm23, %zmm6
18818 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} zmm6 {%k1} = zmm9[0],zmm20[0],zmm9[2],zmm20[2],zmm9[4],zmm20[4],zmm9[6],zmm20[6]
18819 ; AVX512-NEXT: vinserti64x4 $0, %ymm5, %zmm6, %zmm5
18820 ; AVX512-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18821 ; AVX512-NEXT: vpermt2q %zmm1, %zmm25, %zmm3
18822 ; AVX512-NEXT: vpermt2q %zmm2, %zmm25, %zmm4
18823 ; AVX512-NEXT: vpblendd {{.*#+}} ymm1 = ymm4[0,1,2,3],ymm3[4,5,6,7]
18824 ; AVX512-NEXT: vmovdqa64 %zmm8, %zmm2
18825 ; AVX512-NEXT: vpermt2q %zmm0, %zmm25, %zmm2
18826 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} zmm2 {%k1} = zmm9[1],zmm20[1],zmm9[3],zmm20[3],zmm9[5],zmm20[5],zmm9[7],zmm20[7]
18827 ; AVX512-NEXT: vinserti64x4 $0, %ymm1, %zmm2, %zmm1
18828 ; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18829 ; AVX512-NEXT: vmovdqa64 256(%rdx), %zmm1
18830 ; AVX512-NEXT: vmovdqa64 256(%rcx), %zmm3
18831 ; AVX512-NEXT: vmovdqa64 %zmm1, %zmm2
18832 ; AVX512-NEXT: vpermt2q %zmm3, %zmm19, %zmm2
18833 ; AVX512-NEXT: vmovdqa64 256(%rdi), %zmm4
18834 ; AVX512-NEXT: vmovdqa64 256(%rsi), %zmm5
18835 ; AVX512-NEXT: vmovdqa64 %zmm4, %zmm6
18836 ; AVX512-NEXT: vpermt2q %zmm5, %zmm19, %zmm6
18837 ; AVX512-NEXT: vpblendd {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm2[4,5,6,7]
18838 ; AVX512-NEXT: vmovdqa64 256(%r10), %zmm13
18839 ; AVX512-NEXT: vmovdqa64 256(%rax), %zmm2
18840 ; AVX512-NEXT: vmovdqa64 %zmm13, %zmm7
18841 ; AVX512-NEXT: vpermt2q %zmm2, %zmm19, %zmm7
18842 ; AVX512-NEXT: vmovdqa64 256(%r8), %zmm18
18843 ; AVX512-NEXT: vmovdqa64 256(%r9), %zmm15
18844 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} zmm8 = zmm18[0],zmm15[0],zmm18[2],zmm15[2],zmm18[4],zmm15[4],zmm18[6],zmm15[6]
18845 ; AVX512-NEXT: vmovdqa64 %zmm7, %zmm8 {%k1}
18846 ; AVX512-NEXT: vinserti64x4 $0, %ymm6, %zmm8, %zmm6
18847 ; AVX512-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18848 ; AVX512-NEXT: vmovdqa64 %zmm1, %zmm6
18849 ; AVX512-NEXT: vpermt2q %zmm3, %zmm21, %zmm6
18850 ; AVX512-NEXT: vmovdqa64 %zmm4, %zmm7
18851 ; AVX512-NEXT: vpermt2q %zmm5, %zmm21, %zmm7
18852 ; AVX512-NEXT: vpblendd {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]
18853 ; AVX512-NEXT: vmovdqa64 %zmm13, %zmm7
18854 ; AVX512-NEXT: vpermt2q %zmm2, %zmm21, %zmm7
18855 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} zmm8 = zmm18[1],zmm15[1],zmm18[3],zmm15[3],zmm18[5],zmm15[5],zmm18[7],zmm15[7]
18856 ; AVX512-NEXT: vmovdqa64 %zmm7, %zmm8 {%k1}
18857 ; AVX512-NEXT: vinserti64x4 $0, %ymm6, %zmm8, %zmm6
18858 ; AVX512-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18859 ; AVX512-NEXT: vmovdqa64 %zmm1, %zmm6
18860 ; AVX512-NEXT: vpermt2q %zmm3, %zmm23, %zmm6
18861 ; AVX512-NEXT: vmovdqa64 %zmm4, %zmm7
18862 ; AVX512-NEXT: vpermt2q %zmm5, %zmm23, %zmm7
18863 ; AVX512-NEXT: vpblendd {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]
18864 ; AVX512-NEXT: vmovdqa64 %zmm18, %zmm7
18865 ; AVX512-NEXT: vpermt2q %zmm15, %zmm23, %zmm7
18866 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} zmm7 {%k1} = zmm13[0],zmm2[0],zmm13[2],zmm2[2],zmm13[4],zmm2[4],zmm13[6],zmm2[6]
18867 ; AVX512-NEXT: vinserti64x4 $0, %ymm6, %zmm7, %zmm6
18868 ; AVX512-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18869 ; AVX512-NEXT: vpermt2q %zmm3, %zmm25, %zmm1
18870 ; AVX512-NEXT: vpermt2q %zmm5, %zmm25, %zmm4
18871 ; AVX512-NEXT: vpblendd {{.*#+}} ymm1 = ymm4[0,1,2,3],ymm1[4,5,6,7]
18872 ; AVX512-NEXT: vmovdqa64 %zmm18, %zmm3
18873 ; AVX512-NEXT: vpermt2q %zmm15, %zmm25, %zmm3
18874 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} zmm3 {%k1} = zmm13[1],zmm2[1],zmm13[3],zmm2[3],zmm13[5],zmm2[5],zmm13[7],zmm2[7]
18875 ; AVX512-NEXT: vinserti64x4 $0, %ymm1, %zmm3, %zmm1
18876 ; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18877 ; AVX512-NEXT: vmovdqa64 320(%rdx), %zmm1
18878 ; AVX512-NEXT: vmovdqa64 320(%rcx), %zmm4
18879 ; AVX512-NEXT: vmovdqa64 %zmm1, %zmm3
18880 ; AVX512-NEXT: vpermt2q %zmm4, %zmm19, %zmm3
18881 ; AVX512-NEXT: vmovdqa64 320(%rdi), %zmm6
18882 ; AVX512-NEXT: vmovdqa64 320(%rsi), %zmm7
18883 ; AVX512-NEXT: vmovdqa64 %zmm6, %zmm5
18884 ; AVX512-NEXT: vpermt2q %zmm7, %zmm19, %zmm5
18885 ; AVX512-NEXT: vpblendd {{.*#+}} ymm8 = ymm5[0,1,2,3],ymm3[4,5,6,7]
18886 ; AVX512-NEXT: vmovdqa64 320(%r10), %zmm16
18887 ; AVX512-NEXT: vmovdqa64 320(%rax), %zmm22
18888 ; AVX512-NEXT: vmovdqa64 %zmm16, %zmm12
18889 ; AVX512-NEXT: vpermt2q %zmm22, %zmm19, %zmm12
18890 ; AVX512-NEXT: vmovdqa64 320(%r8), %zmm29
18891 ; AVX512-NEXT: vmovdqa64 320(%r9), %zmm5
18892 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} zmm14 = zmm29[0],zmm5[0],zmm29[2],zmm5[2],zmm29[4],zmm5[4],zmm29[6],zmm5[6]
18893 ; AVX512-NEXT: vmovdqa64 %zmm12, %zmm14 {%k1}
18894 ; AVX512-NEXT: vinserti64x4 $0, %ymm8, %zmm14, %zmm3
18895 ; AVX512-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18896 ; AVX512-NEXT: vmovdqa64 %zmm1, %zmm8
18897 ; AVX512-NEXT: vpermt2q %zmm4, %zmm21, %zmm8
18898 ; AVX512-NEXT: vmovdqa64 %zmm6, %zmm12
18899 ; AVX512-NEXT: vpermt2q %zmm7, %zmm21, %zmm12
18900 ; AVX512-NEXT: vpblendd {{.*#+}} ymm8 = ymm12[0,1,2,3],ymm8[4,5,6,7]
18901 ; AVX512-NEXT: vmovdqa64 %zmm16, %zmm12
18902 ; AVX512-NEXT: vpermt2q %zmm22, %zmm21, %zmm12
18903 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} zmm14 = zmm29[1],zmm5[1],zmm29[3],zmm5[3],zmm29[5],zmm5[5],zmm29[7],zmm5[7]
18904 ; AVX512-NEXT: vmovdqa64 %zmm12, %zmm14 {%k1}
18905 ; AVX512-NEXT: vinserti64x4 $0, %ymm8, %zmm14, %zmm3
18906 ; AVX512-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18907 ; AVX512-NEXT: vmovdqa64 %zmm1, %zmm8
18908 ; AVX512-NEXT: vpermt2q %zmm4, %zmm23, %zmm8
18909 ; AVX512-NEXT: vmovdqa64 %zmm6, %zmm12
18910 ; AVX512-NEXT: vpermt2q %zmm7, %zmm23, %zmm12
18911 ; AVX512-NEXT: vpblendd {{.*#+}} ymm8 = ymm12[0,1,2,3],ymm8[4,5,6,7]
18912 ; AVX512-NEXT: vmovdqa64 %zmm29, %zmm12
18913 ; AVX512-NEXT: vpermt2q %zmm5, %zmm23, %zmm12
18914 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} zmm12 {%k1} = zmm16[0],zmm22[0],zmm16[2],zmm22[2],zmm16[4],zmm22[4],zmm16[6],zmm22[6]
18915 ; AVX512-NEXT: vinserti64x4 $0, %ymm8, %zmm12, %zmm3
18916 ; AVX512-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18917 ; AVX512-NEXT: vpermt2q %zmm4, %zmm25, %zmm1
18918 ; AVX512-NEXT: vpermt2q %zmm7, %zmm25, %zmm6
18919 ; AVX512-NEXT: vpblendd {{.*#+}} ymm1 = ymm6[0,1,2,3],ymm1[4,5,6,7]
18920 ; AVX512-NEXT: vmovdqa64 %zmm29, %zmm4
18921 ; AVX512-NEXT: vpermt2q %zmm5, %zmm25, %zmm4
18922 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} zmm4 {%k1} = zmm16[1],zmm22[1],zmm16[3],zmm22[3],zmm16[5],zmm22[5],zmm16[7],zmm22[7]
18923 ; AVX512-NEXT: vinserti64x4 $0, %ymm1, %zmm4, %zmm1
18924 ; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18925 ; AVX512-NEXT: vmovdqa64 384(%rdx), %zmm3
18926 ; AVX512-NEXT: vmovdqa64 384(%rcx), %zmm1
18927 ; AVX512-NEXT: vmovdqa64 %zmm3, %zmm4
18928 ; AVX512-NEXT: vpermt2q %zmm1, %zmm19, %zmm4
18929 ; AVX512-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18930 ; AVX512-NEXT: vmovdqa64 %zmm3, %zmm4
18931 ; AVX512-NEXT: vpermt2q %zmm1, %zmm21, %zmm4
18932 ; AVX512-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18933 ; AVX512-NEXT: vmovdqa64 %zmm3, %zmm4
18934 ; AVX512-NEXT: vpermt2q %zmm1, %zmm23, %zmm4
18935 ; AVX512-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18936 ; AVX512-NEXT: vpermt2q %zmm1, %zmm25, %zmm3
18937 ; AVX512-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18938 ; AVX512-NEXT: vmovdqa64 384(%rdi), %zmm3
18939 ; AVX512-NEXT: vmovdqa64 384(%rsi), %zmm1
18940 ; AVX512-NEXT: vmovdqa64 %zmm3, %zmm4
18941 ; AVX512-NEXT: vpermt2q %zmm1, %zmm19, %zmm4
18942 ; AVX512-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18943 ; AVX512-NEXT: vmovdqa64 %zmm3, %zmm4
18944 ; AVX512-NEXT: vpermt2q %zmm1, %zmm21, %zmm4
18945 ; AVX512-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18946 ; AVX512-NEXT: vmovdqa64 %zmm3, %zmm4
18947 ; AVX512-NEXT: vpermt2q %zmm1, %zmm23, %zmm4
18948 ; AVX512-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18949 ; AVX512-NEXT: vpermt2q %zmm1, %zmm25, %zmm3
18950 ; AVX512-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18951 ; AVX512-NEXT: vmovdqa64 448(%rdx), %zmm3
18952 ; AVX512-NEXT: vmovdqa64 448(%rcx), %zmm1
18953 ; AVX512-NEXT: vmovdqa64 %zmm3, %zmm4
18954 ; AVX512-NEXT: vpermt2q %zmm1, %zmm19, %zmm4
18955 ; AVX512-NEXT: vmovdqu64 %zmm4, (%rsp) # 64-byte Spill
18956 ; AVX512-NEXT: vmovdqa64 %zmm3, %zmm4
18957 ; AVX512-NEXT: vpermt2q %zmm1, %zmm21, %zmm4
18958 ; AVX512-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18959 ; AVX512-NEXT: vmovdqa64 %zmm3, %zmm4
18960 ; AVX512-NEXT: vpermt2q %zmm1, %zmm23, %zmm4
18961 ; AVX512-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18962 ; AVX512-NEXT: vpermt2q %zmm1, %zmm25, %zmm3
18963 ; AVX512-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18964 ; AVX512-NEXT: vmovdqa64 448(%rdi), %zmm3
18965 ; AVX512-NEXT: vmovdqa64 448(%rsi), %zmm1
18966 ; AVX512-NEXT: vmovdqa64 %zmm3, %zmm4
18967 ; AVX512-NEXT: vpermt2q %zmm1, %zmm19, %zmm4
18968 ; AVX512-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18969 ; AVX512-NEXT: vmovdqa64 %zmm3, %zmm4
18970 ; AVX512-NEXT: vpermt2q %zmm1, %zmm21, %zmm4
18971 ; AVX512-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18972 ; AVX512-NEXT: vmovdqa64 %zmm3, %zmm4
18973 ; AVX512-NEXT: vpermt2q %zmm1, %zmm23, %zmm4
18974 ; AVX512-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18975 ; AVX512-NEXT: vpermt2q %zmm1, %zmm25, %zmm3
18976 ; AVX512-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18977 ; AVX512-NEXT: vbroadcasti32x4 {{.*#+}} zmm11 = [0,8,0,8,0,8,0,8]
18978 ; AVX512-NEXT: # zmm11 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
18979 ; AVX512-NEXT: vmovdqa64 %zmm26, %zmm1
18980 ; AVX512-NEXT: vpermt2q %zmm17, %zmm11, %zmm1
18981 ; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18982 ; AVX512-NEXT: vbroadcasti32x4 {{.*#+}} zmm12 = [1,9,1,9,1,9,1,9]
18983 ; AVX512-NEXT: # zmm12 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
18984 ; AVX512-NEXT: vmovdqa64 %zmm26, %zmm1
18985 ; AVX512-NEXT: vpermt2q %zmm17, %zmm12, %zmm1
18986 ; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18987 ; AVX512-NEXT: vbroadcasti32x4 {{.*#+}} zmm10 = [2,10,2,10,2,10,2,10]
18988 ; AVX512-NEXT: # zmm10 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
18989 ; AVX512-NEXT: vmovdqa64 %zmm26, %zmm1
18990 ; AVX512-NEXT: vpermt2q %zmm17, %zmm10, %zmm1
18991 ; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18992 ; AVX512-NEXT: vbroadcasti32x4 {{.*#+}} zmm4 = [3,11,3,11,3,11,3,11]
18993 ; AVX512-NEXT: # zmm4 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
18994 ; AVX512-NEXT: vpermt2q %zmm17, %zmm4, %zmm26
18995 ; AVX512-NEXT: vmovdqu64 %zmm26, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18996 ; AVX512-NEXT: vmovdqa64 %zmm30, %zmm1
18997 ; AVX512-NEXT: vpermt2q %zmm24, %zmm11, %zmm1
18998 ; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
18999 ; AVX512-NEXT: vmovdqa64 %zmm30, %zmm1
19000 ; AVX512-NEXT: vpermt2q %zmm24, %zmm12, %zmm1
19001 ; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19002 ; AVX512-NEXT: vmovdqa64 %zmm30, %zmm1
19003 ; AVX512-NEXT: vpermt2q %zmm24, %zmm10, %zmm1
19004 ; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19005 ; AVX512-NEXT: vpermt2q %zmm24, %zmm4, %zmm30
19006 ; AVX512-NEXT: vmovdqu64 %zmm30, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19007 ; AVX512-NEXT: vmovdqa64 %zmm31, %zmm1
19008 ; AVX512-NEXT: vpermt2q %zmm27, %zmm11, %zmm1
19009 ; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19010 ; AVX512-NEXT: vmovdqa64 %zmm31, %zmm1
19011 ; AVX512-NEXT: vpermt2q %zmm27, %zmm12, %zmm1
19012 ; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19013 ; AVX512-NEXT: vmovdqa64 %zmm31, %zmm1
19014 ; AVX512-NEXT: vpermt2q %zmm27, %zmm10, %zmm1
19015 ; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19016 ; AVX512-NEXT: vpermt2q %zmm27, %zmm4, %zmm31
19017 ; AVX512-NEXT: vmovdqu64 %zmm31, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19018 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
19019 ; AVX512-NEXT: vmovdqa64 %zmm3, %zmm1
19020 ; AVX512-NEXT: vpermt2q %zmm28, %zmm11, %zmm1
19021 ; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19022 ; AVX512-NEXT: vmovdqa64 %zmm3, %zmm1
19023 ; AVX512-NEXT: vpermt2q %zmm28, %zmm12, %zmm1
19024 ; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19025 ; AVX512-NEXT: vmovdqa64 %zmm3, %zmm31
19026 ; AVX512-NEXT: vpermt2q %zmm28, %zmm10, %zmm31
19027 ; AVX512-NEXT: vpermt2q %zmm28, %zmm4, %zmm3
19028 ; AVX512-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19029 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
19030 ; AVX512-NEXT: vmovdqa64 %zmm28, %zmm3
19031 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
19032 ; AVX512-NEXT: vpermt2q %zmm1, %zmm11, %zmm3
19033 ; AVX512-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19034 ; AVX512-NEXT: vmovdqa64 %zmm28, %zmm3
19035 ; AVX512-NEXT: vpermt2q %zmm1, %zmm12, %zmm3
19036 ; AVX512-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19037 ; AVX512-NEXT: vmovdqa64 %zmm28, %zmm3
19038 ; AVX512-NEXT: vpermt2q %zmm1, %zmm10, %zmm3
19039 ; AVX512-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19040 ; AVX512-NEXT: vpermt2q %zmm1, %zmm4, %zmm28
19041 ; AVX512-NEXT: vmovdqu64 %zmm28, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19042 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
19043 ; AVX512-NEXT: vmovdqa64 %zmm6, %zmm3
19044 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
19045 ; AVX512-NEXT: vpermt2q %zmm1, %zmm11, %zmm3
19046 ; AVX512-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19047 ; AVX512-NEXT: vmovdqa64 %zmm6, %zmm3
19048 ; AVX512-NEXT: vpermt2q %zmm1, %zmm12, %zmm3
19049 ; AVX512-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19050 ; AVX512-NEXT: vmovdqa64 %zmm6, %zmm27
19051 ; AVX512-NEXT: vpermt2q %zmm1, %zmm10, %zmm27
19052 ; AVX512-NEXT: vpermt2q %zmm1, %zmm4, %zmm6
19053 ; AVX512-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19054 ; AVX512-NEXT: vmovdqa64 %zmm9, %zmm1
19055 ; AVX512-NEXT: vpermt2q %zmm20, %zmm11, %zmm1
19056 ; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19057 ; AVX512-NEXT: vmovdqa64 %zmm9, %zmm1
19058 ; AVX512-NEXT: vpermt2q %zmm20, %zmm12, %zmm1
19059 ; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19060 ; AVX512-NEXT: vmovdqa64 %zmm9, %zmm1
19061 ; AVX512-NEXT: vpermt2q %zmm20, %zmm10, %zmm1
19062 ; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19063 ; AVX512-NEXT: vpermt2q %zmm20, %zmm4, %zmm9
19064 ; AVX512-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19065 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
19066 ; AVX512-NEXT: vmovdqa64 %zmm1, %zmm3
19067 ; AVX512-NEXT: vpermt2q %zmm0, %zmm11, %zmm3
19068 ; AVX512-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19069 ; AVX512-NEXT: vmovdqa64 %zmm1, %zmm3
19070 ; AVX512-NEXT: vpermt2q %zmm0, %zmm12, %zmm3
19071 ; AVX512-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19072 ; AVX512-NEXT: vmovdqa64 %zmm1, %zmm20
19073 ; AVX512-NEXT: vpermt2q %zmm0, %zmm10, %zmm20
19074 ; AVX512-NEXT: vpermt2q %zmm0, %zmm4, %zmm1
19075 ; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19076 ; AVX512-NEXT: vmovdqa64 %zmm13, %zmm0
19077 ; AVX512-NEXT: vpermt2q %zmm2, %zmm11, %zmm0
19078 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19079 ; AVX512-NEXT: vmovdqa64 %zmm13, %zmm0
19080 ; AVX512-NEXT: vpermt2q %zmm2, %zmm12, %zmm0
19081 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19082 ; AVX512-NEXT: vmovdqa64 %zmm13, %zmm0
19083 ; AVX512-NEXT: vpermt2q %zmm2, %zmm10, %zmm0
19084 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19085 ; AVX512-NEXT: vpermt2q %zmm2, %zmm4, %zmm13
19086 ; AVX512-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19087 ; AVX512-NEXT: vmovdqa64 %zmm18, %zmm0
19088 ; AVX512-NEXT: vpermt2q %zmm15, %zmm11, %zmm0
19089 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19090 ; AVX512-NEXT: vmovdqa64 %zmm18, %zmm0
19091 ; AVX512-NEXT: vpermt2q %zmm15, %zmm12, %zmm0
19092 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19093 ; AVX512-NEXT: vmovdqa64 %zmm18, %zmm28
19094 ; AVX512-NEXT: vpermt2q %zmm15, %zmm10, %zmm28
19095 ; AVX512-NEXT: vpermt2q %zmm15, %zmm4, %zmm18
19096 ; AVX512-NEXT: vmovdqu64 %zmm18, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19097 ; AVX512-NEXT: vmovdqa64 %zmm16, %zmm0
19098 ; AVX512-NEXT: vpermt2q %zmm22, %zmm11, %zmm0
19099 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19100 ; AVX512-NEXT: vmovdqa64 %zmm16, %zmm0
19101 ; AVX512-NEXT: vpermt2q %zmm22, %zmm12, %zmm0
19102 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19103 ; AVX512-NEXT: vmovdqa64 %zmm16, %zmm0
19104 ; AVX512-NEXT: vpermt2q %zmm22, %zmm10, %zmm0
19105 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19106 ; AVX512-NEXT: vpermt2q %zmm22, %zmm4, %zmm16
19107 ; AVX512-NEXT: vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19108 ; AVX512-NEXT: vmovdqa64 %zmm29, %zmm0
19109 ; AVX512-NEXT: vpermt2q %zmm5, %zmm11, %zmm0
19110 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19111 ; AVX512-NEXT: vmovdqa64 %zmm29, %zmm0
19112 ; AVX512-NEXT: vpermt2q %zmm5, %zmm12, %zmm0
19113 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19114 ; AVX512-NEXT: vmovdqa64 %zmm29, %zmm22
19115 ; AVX512-NEXT: vpermt2q %zmm5, %zmm10, %zmm22
19116 ; AVX512-NEXT: vpermt2q %zmm5, %zmm4, %zmm29
19117 ; AVX512-NEXT: vmovdqa64 384(%r10), %zmm13
19118 ; AVX512-NEXT: vmovdqa64 384(%rax), %zmm0
19119 ; AVX512-NEXT: vmovdqa64 %zmm13, %zmm1
19120 ; AVX512-NEXT: vpermt2q %zmm0, %zmm11, %zmm1
19121 ; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19122 ; AVX512-NEXT: vmovdqa64 %zmm13, %zmm1
19123 ; AVX512-NEXT: vpermt2q %zmm0, %zmm12, %zmm1
19124 ; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19125 ; AVX512-NEXT: vmovdqa64 %zmm13, %zmm14
19126 ; AVX512-NEXT: vpermt2q %zmm0, %zmm19, %zmm14
19127 ; AVX512-NEXT: vmovdqa64 %zmm13, %zmm15
19128 ; AVX512-NEXT: vpermt2q %zmm0, %zmm21, %zmm15
19129 ; AVX512-NEXT: vmovdqa64 384(%r8), %zmm5
19130 ; AVX512-NEXT: vmovdqa64 384(%r9), %zmm2
19131 ; AVX512-NEXT: vmovdqa64 %zmm5, %zmm8
19132 ; AVX512-NEXT: vpermt2q %zmm2, %zmm23, %zmm8
19133 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} zmm8 {%k1} = zmm13[0],zmm0[0],zmm13[2],zmm0[2],zmm13[4],zmm0[4],zmm13[6],zmm0[6]
19134 ; AVX512-NEXT: vmovdqa64 %zmm5, %zmm6
19135 ; AVX512-NEXT: vpermt2q %zmm2, %zmm25, %zmm6
19136 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} zmm6 {%k1} = zmm13[1],zmm0[1],zmm13[3],zmm0[3],zmm13[5],zmm0[5],zmm13[7],zmm0[7]
19137 ; AVX512-NEXT: vmovdqa64 %zmm13, %zmm30
19138 ; AVX512-NEXT: vpermt2q %zmm0, %zmm10, %zmm30
19139 ; AVX512-NEXT: vpermt2q %zmm0, %zmm4, %zmm13
19140 ; AVX512-NEXT: vmovdqa64 %zmm5, %zmm24
19141 ; AVX512-NEXT: vpermt2q %zmm2, %zmm11, %zmm24
19142 ; AVX512-NEXT: vmovdqa64 %zmm5, %zmm26
19143 ; AVX512-NEXT: vpermt2q %zmm2, %zmm12, %zmm26
19144 ; AVX512-NEXT: vmovdqa64 %zmm5, %zmm17
19145 ; AVX512-NEXT: vpermt2q %zmm2, %zmm10, %zmm17
19146 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} zmm7 = zmm5[0],zmm2[0],zmm5[2],zmm2[2],zmm5[4],zmm2[4],zmm5[6],zmm2[6]
19147 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} zmm9 = zmm5[1],zmm2[1],zmm5[3],zmm2[3],zmm5[5],zmm2[5],zmm5[7],zmm2[7]
19148 ; AVX512-NEXT: vpermt2q %zmm2, %zmm4, %zmm5
19149 ; AVX512-NEXT: vmovdqa64 448(%r10), %zmm16
19150 ; AVX512-NEXT: vmovdqa64 448(%rax), %zmm1
19151 ; AVX512-NEXT: vmovdqa64 %zmm16, %zmm0
19152 ; AVX512-NEXT: vpermt2q %zmm1, %zmm11, %zmm0
19153 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19154 ; AVX512-NEXT: vmovdqa64 %zmm16, %zmm0
19155 ; AVX512-NEXT: vpermt2q %zmm1, %zmm12, %zmm0
19156 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19157 ; AVX512-NEXT: vpermi2q %zmm1, %zmm16, %zmm19
19158 ; AVX512-NEXT: vpermi2q %zmm1, %zmm16, %zmm21
19159 ; AVX512-NEXT: vmovdqa64 448(%r8), %zmm3
19160 ; AVX512-NEXT: vmovdqa64 448(%r9), %zmm0
19161 ; AVX512-NEXT: vpermi2q %zmm0, %zmm3, %zmm23
19162 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} zmm23 {%k1} = zmm16[0],zmm1[0],zmm16[2],zmm1[2],zmm16[4],zmm1[4],zmm16[6],zmm1[6]
19163 ; AVX512-NEXT: vpermi2q %zmm0, %zmm3, %zmm25
19164 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} zmm25 {%k1} = zmm16[1],zmm1[1],zmm16[3],zmm1[3],zmm16[5],zmm1[5],zmm16[7],zmm1[7]
19165 ; AVX512-NEXT: vmovdqa64 %zmm16, %zmm18
19166 ; AVX512-NEXT: vpermt2q %zmm1, %zmm10, %zmm18
19167 ; AVX512-NEXT: vpermt2q %zmm1, %zmm4, %zmm16
19168 ; AVX512-NEXT: vpermi2q %zmm0, %zmm3, %zmm11
19169 ; AVX512-NEXT: vpermi2q %zmm0, %zmm3, %zmm12
19170 ; AVX512-NEXT: vpermi2q %zmm0, %zmm3, %zmm10
19171 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} zmm1 = zmm3[0],zmm0[0],zmm3[2],zmm0[2],zmm3[4],zmm0[4],zmm3[6],zmm0[6]
19172 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} zmm2 = zmm3[1],zmm0[1],zmm3[3],zmm0[3],zmm3[5],zmm0[5],zmm3[7],zmm0[7]
19173 ; AVX512-NEXT: vpermt2q %zmm0, %zmm4, %zmm3
19174 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
19175 ; AVX512-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
19176 ; AVX512-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
19177 ; AVX512-NEXT: vmovdqa64 %zmm14, %zmm7 {%k1}
19178 ; AVX512-NEXT: vinserti64x4 $0, %ymm0, %zmm7, %zmm0
19179 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19180 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
19181 ; AVX512-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
19182 ; AVX512-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
19183 ; AVX512-NEXT: vmovdqa64 %zmm15, %zmm9 {%k1}
19184 ; AVX512-NEXT: vinserti64x4 $0, %ymm0, %zmm9, %zmm0
19185 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19186 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
19187 ; AVX512-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
19188 ; AVX512-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
19189 ; AVX512-NEXT: vinserti64x4 $0, %ymm0, %zmm8, %zmm0
19190 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19191 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
19192 ; AVX512-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
19193 ; AVX512-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
19194 ; AVX512-NEXT: vinserti64x4 $0, %ymm0, %zmm6, %zmm0
19195 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19196 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
19197 ; AVX512-NEXT: vpblendd $240, (%rsp), %ymm0, %ymm0 # 32-byte Folded Reload
19198 ; AVX512-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
19199 ; AVX512-NEXT: vmovdqa64 %zmm19, %zmm1 {%k1}
19200 ; AVX512-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
19201 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19202 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
19203 ; AVX512-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
19204 ; AVX512-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
19205 ; AVX512-NEXT: vmovdqa64 %zmm21, %zmm2 {%k1}
19206 ; AVX512-NEXT: vinserti64x4 $0, %ymm0, %zmm2, %zmm0
19207 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19208 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
19209 ; AVX512-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
19210 ; AVX512-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
19211 ; AVX512-NEXT: vinserti64x4 $0, %ymm0, %zmm23, %zmm0
19212 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19213 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
19214 ; AVX512-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
19215 ; AVX512-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
19216 ; AVX512-NEXT: vinserti64x4 $0, %ymm0, %zmm25, %zmm0
19217 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19218 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
19219 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
19220 ; AVX512-NEXT: vmovdqa64 %zmm0, %zmm8 {%k1}
19221 ; AVX512-NEXT: vmovdqa (%rcx), %ymm0
19222 ; AVX512-NEXT: vmovdqa (%rdx), %ymm1
19223 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
19224 ; AVX512-NEXT: vmovdqa (%rsi), %ymm4
19225 ; AVX512-NEXT: vmovdqa (%rdi), %ymm6
19226 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm6[0],ymm4[0],ymm6[2],ymm4[2]
19227 ; AVX512-NEXT: vperm2i128 {{.*#+}} ymm2 = ymm7[2,3],ymm2[2,3]
19228 ; AVX512-NEXT: vinserti64x4 $0, %ymm2, %zmm8, %zmm2
19229 ; AVX512-NEXT: vmovdqu64 %zmm2, (%rsp) # 64-byte Spill
19230 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
19231 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
19232 ; AVX512-NEXT: vmovdqa64 %zmm2, %zmm7 {%k1}
19233 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
19234 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm6[1],ymm4[1],ymm6[3],ymm4[3]
19235 ; AVX512-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
19236 ; AVX512-NEXT: vinserti64x4 $0, %ymm0, %zmm7, %zmm0
19237 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19238 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
19239 ; AVX512-NEXT: vmovdqa64 %zmm0, %zmm31 {%k1}
19240 ; AVX512-NEXT: vmovdqa 64(%rcx), %ymm0
19241 ; AVX512-NEXT: vmovdqa 64(%rdx), %ymm1
19242 ; AVX512-NEXT: vmovdqa 64(%rsi), %ymm2
19243 ; AVX512-NEXT: vmovdqa 64(%rdi), %ymm4
19244 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} ymm6 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
19245 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm4[0],ymm2[0],ymm4[2],ymm2[2]
19246 ; AVX512-NEXT: vperm2i128 {{.*#+}} ymm6 = ymm7[2,3],ymm6[2,3]
19247 ; AVX512-NEXT: vinserti64x4 $0, %ymm6, %zmm31, %zmm6
19248 ; AVX512-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19249 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
19250 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
19251 ; AVX512-NEXT: vmovdqa64 %zmm6, %zmm7 {%k1}
19252 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
19253 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm4[1],ymm2[1],ymm4[3],ymm2[3]
19254 ; AVX512-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
19255 ; AVX512-NEXT: vinserti64x4 $0, %ymm0, %zmm7, %zmm8
19256 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
19257 ; AVX512-NEXT: vmovdqa64 %zmm0, %zmm27 {%k1}
19258 ; AVX512-NEXT: vmovdqa 128(%rcx), %ymm0
19259 ; AVX512-NEXT: vmovdqa 128(%rdx), %ymm1
19260 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
19261 ; AVX512-NEXT: vmovdqa 128(%rsi), %ymm4
19262 ; AVX512-NEXT: vmovdqa 128(%rdi), %ymm6
19263 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm6[0],ymm4[0],ymm6[2],ymm4[2]
19264 ; AVX512-NEXT: vperm2i128 {{.*#+}} ymm2 = ymm7[2,3],ymm2[2,3]
19265 ; AVX512-NEXT: vinserti64x4 $0, %ymm2, %zmm27, %zmm23
19266 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
19267 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
19268 ; AVX512-NEXT: vmovdqa64 %zmm2, %zmm7 {%k1}
19269 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
19270 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm6[1],ymm4[1],ymm6[3],ymm4[3]
19271 ; AVX512-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
19272 ; AVX512-NEXT: vinserti64x4 $0, %ymm0, %zmm7, %zmm25
19273 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
19274 ; AVX512-NEXT: vmovdqa64 %zmm0, %zmm20 {%k1}
19275 ; AVX512-NEXT: vmovdqa 192(%rcx), %ymm0
19276 ; AVX512-NEXT: vmovdqa 192(%rdx), %ymm1
19277 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
19278 ; AVX512-NEXT: vmovdqa 192(%rsi), %ymm4
19279 ; AVX512-NEXT: vmovdqa 192(%rdi), %ymm6
19280 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm6[0],ymm4[0],ymm6[2],ymm4[2]
19281 ; AVX512-NEXT: vperm2i128 {{.*#+}} ymm2 = ymm7[2,3],ymm2[2,3]
19282 ; AVX512-NEXT: vinserti64x4 $0, %ymm2, %zmm20, %zmm21
19283 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
19284 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
19285 ; AVX512-NEXT: vmovdqa64 %zmm2, %zmm7 {%k1}
19286 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
19287 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm6[1],ymm4[1],ymm6[3],ymm4[3]
19288 ; AVX512-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
19289 ; AVX512-NEXT: vinserti64x4 $0, %ymm0, %zmm7, %zmm19
19290 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
19291 ; AVX512-NEXT: vmovdqa64 %zmm0, %zmm28 {%k1}
19292 ; AVX512-NEXT: vmovdqa 256(%rcx), %ymm0
19293 ; AVX512-NEXT: vmovdqa 256(%rdx), %ymm1
19294 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
19295 ; AVX512-NEXT: vmovdqa 256(%rsi), %ymm4
19296 ; AVX512-NEXT: vmovdqa 256(%rdi), %ymm6
19297 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm6[0],ymm4[0],ymm6[2],ymm4[2]
19298 ; AVX512-NEXT: vperm2i128 {{.*#+}} ymm2 = ymm7[2,3],ymm2[2,3]
19299 ; AVX512-NEXT: vinserti64x4 $0, %ymm2, %zmm28, %zmm15
19300 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
19301 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
19302 ; AVX512-NEXT: vmovdqa64 %zmm2, %zmm7 {%k1}
19303 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
19304 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm6[1],ymm4[1],ymm6[3],ymm4[3]
19305 ; AVX512-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
19306 ; AVX512-NEXT: vinserti64x4 $0, %ymm0, %zmm7, %zmm28
19307 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
19308 ; AVX512-NEXT: vmovdqa64 %zmm0, %zmm22 {%k1}
19309 ; AVX512-NEXT: vmovdqa 320(%rcx), %ymm0
19310 ; AVX512-NEXT: vmovdqa 320(%rdx), %ymm1
19311 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
19312 ; AVX512-NEXT: vmovdqa 320(%rsi), %ymm4
19313 ; AVX512-NEXT: vmovdqa 320(%rdi), %ymm6
19314 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm6[0],ymm4[0],ymm6[2],ymm4[2]
19315 ; AVX512-NEXT: vperm2i128 {{.*#+}} ymm2 = ymm7[2,3],ymm2[2,3]
19316 ; AVX512-NEXT: vinserti64x4 $0, %ymm2, %zmm22, %zmm22
19317 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
19318 ; AVX512-NEXT: vmovdqa64 %zmm2, %zmm29 {%k1}
19319 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
19320 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm6[1],ymm4[1],ymm6[3],ymm4[3]
19321 ; AVX512-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
19322 ; AVX512-NEXT: vinserti64x4 $0, %ymm0, %zmm29, %zmm27
19323 ; AVX512-NEXT: vmovdqa64 %zmm30, %zmm17 {%k1}
19324 ; AVX512-NEXT: vmovdqa 384(%rcx), %ymm0
19325 ; AVX512-NEXT: vmovdqa 384(%rdx), %ymm1
19326 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
19327 ; AVX512-NEXT: vmovdqa 384(%rsi), %ymm4
19328 ; AVX512-NEXT: vmovdqa 384(%rdi), %ymm6
19329 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm6[0],ymm4[0],ymm6[2],ymm4[2]
19330 ; AVX512-NEXT: vperm2i128 {{.*#+}} ymm2 = ymm7[2,3],ymm2[2,3]
19331 ; AVX512-NEXT: vinserti64x4 $0, %ymm2, %zmm17, %zmm14
19332 ; AVX512-NEXT: vmovdqa64 %zmm13, %zmm5 {%k1}
19333 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
19334 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm6[1],ymm4[1],ymm6[3],ymm4[3]
19335 ; AVX512-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
19336 ; AVX512-NEXT: vinserti64x4 $0, %ymm0, %zmm5, %zmm13
19337 ; AVX512-NEXT: vmovdqa64 %zmm18, %zmm10 {%k1}
19338 ; AVX512-NEXT: vmovdqa 448(%rcx), %ymm0
19339 ; AVX512-NEXT: vmovdqa 448(%rdx), %ymm1
19340 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
19341 ; AVX512-NEXT: vmovdqa 448(%rsi), %ymm4
19342 ; AVX512-NEXT: vmovdqa 448(%rdi), %ymm5
19343 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} ymm6 = ymm5[0],ymm4[0],ymm5[2],ymm4[2]
19344 ; AVX512-NEXT: vperm2i128 {{.*#+}} ymm2 = ymm6[2,3],ymm2[2,3]
19345 ; AVX512-NEXT: vinserti64x4 $0, %ymm2, %zmm10, %zmm10
19346 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
19347 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm5[1],ymm4[1],ymm5[3],ymm4[3]
19348 ; AVX512-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
19349 ; AVX512-NEXT: vmovdqa64 %zmm16, %zmm3 {%k1}
19350 ; AVX512-NEXT: vinserti64x4 $0, %ymm0, %zmm3, %zmm9
19351 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
19352 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
19353 ; AVX512-NEXT: vmovdqa64 %zmm0, %zmm3 {%k1}
19354 ; AVX512-NEXT: vmovdqa (%rsi), %xmm0
19355 ; AVX512-NEXT: vinserti128 $1, (%rcx), %ymm0, %ymm1
19356 ; AVX512-NEXT: vmovdqa (%rdi), %xmm0
19357 ; AVX512-NEXT: vinserti128 $1, (%rdx), %ymm0, %ymm2
19358 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm2[0],ymm1[0],ymm2[2],ymm1[2]
19359 ; AVX512-NEXT: vinserti64x4 $0, %ymm0, %zmm3, %zmm0
19360 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
19361 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
19362 ; AVX512-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
19363 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm2[1],ymm1[1],ymm2[3],ymm1[3]
19364 ; AVX512-NEXT: vinserti64x4 $0, %ymm1, %zmm4, %zmm2
19365 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
19366 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
19367 ; AVX512-NEXT: vmovdqa64 %zmm1, %zmm5 {%k1}
19368 ; AVX512-NEXT: vmovdqa 64(%rsi), %xmm1
19369 ; AVX512-NEXT: vinserti128 $1, 64(%rcx), %ymm1, %ymm1
19370 ; AVX512-NEXT: vmovdqa 64(%rdi), %xmm3
19371 ; AVX512-NEXT: vinserti128 $1, 64(%rdx), %ymm3, %ymm4
19372 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm4[0],ymm1[0],ymm4[2],ymm1[2]
19373 ; AVX512-NEXT: vinserti64x4 $0, %ymm3, %zmm5, %zmm3
19374 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
19375 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
19376 ; AVX512-NEXT: vmovdqa64 %zmm5, %zmm6 {%k1}
19377 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm4[1],ymm1[1],ymm4[3],ymm1[3]
19378 ; AVX512-NEXT: vinserti64x4 $0, %ymm1, %zmm6, %zmm5
19379 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
19380 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
19381 ; AVX512-NEXT: vmovdqa64 %zmm1, %zmm7 {%k1}
19382 ; AVX512-NEXT: vmovdqa 128(%rsi), %xmm1
19383 ; AVX512-NEXT: vinserti128 $1, 128(%rcx), %ymm1, %ymm1
19384 ; AVX512-NEXT: vmovdqa 128(%rdi), %xmm4
19385 ; AVX512-NEXT: vinserti128 $1, 128(%rdx), %ymm4, %ymm6
19386 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} ymm4 = ymm6[0],ymm1[0],ymm6[2],ymm1[2]
19387 ; AVX512-NEXT: vinserti64x4 $0, %ymm4, %zmm7, %zmm4
19388 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
19389 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
19390 ; AVX512-NEXT: vmovdqa64 %zmm7, %zmm16 {%k1}
19391 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm6[1],ymm1[1],ymm6[3],ymm1[3]
19392 ; AVX512-NEXT: vinserti64x4 $0, %ymm1, %zmm16, %zmm6
19393 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
19394 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
19395 ; AVX512-NEXT: vmovdqa64 %zmm1, %zmm17 {%k1}
19396 ; AVX512-NEXT: vmovdqa 192(%rsi), %xmm1
19397 ; AVX512-NEXT: vinserti128 $1, 192(%rcx), %ymm1, %ymm1
19398 ; AVX512-NEXT: vmovdqa 192(%rdi), %xmm7
19399 ; AVX512-NEXT: vinserti32x4 $1, 192(%rdx), %ymm7, %ymm16
19400 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm16[0],ymm1[0],ymm16[2],ymm1[2]
19401 ; AVX512-NEXT: vinserti64x4 $0, %ymm7, %zmm17, %zmm7
19402 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
19403 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
19404 ; AVX512-NEXT: vmovdqa64 %zmm17, %zmm18 {%k1}
19405 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm16[1],ymm1[1],ymm16[3],ymm1[3]
19406 ; AVX512-NEXT: vinserti64x4 $0, %ymm1, %zmm18, %zmm1
19407 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
19408 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
19409 ; AVX512-NEXT: vmovdqa64 %zmm16, %zmm20 {%k1}
19410 ; AVX512-NEXT: vmovdqa64 256(%rsi), %xmm16
19411 ; AVX512-NEXT: vinserti32x4 $1, 256(%rcx), %ymm16, %ymm16
19412 ; AVX512-NEXT: vmovdqa64 256(%rdi), %xmm17
19413 ; AVX512-NEXT: vinserti32x4 $1, 256(%rdx), %ymm17, %ymm17
19414 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} ymm18 = ymm17[0],ymm16[0],ymm17[2],ymm16[2]
19415 ; AVX512-NEXT: vinserti64x4 $0, %ymm18, %zmm20, %zmm29
19416 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
19417 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
19418 ; AVX512-NEXT: vmovdqa64 %zmm18, %zmm20 {%k1}
19419 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} ymm16 = ymm17[1],ymm16[1],ymm17[3],ymm16[3]
19420 ; AVX512-NEXT: vinserti64x4 $0, %ymm16, %zmm20, %zmm30
19421 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
19422 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
19423 ; AVX512-NEXT: vmovdqa64 %zmm16, %zmm20 {%k1}
19424 ; AVX512-NEXT: vmovdqa64 320(%rsi), %xmm16
19425 ; AVX512-NEXT: vinserti32x4 $1, 320(%rcx), %ymm16, %ymm16
19426 ; AVX512-NEXT: vmovdqa64 320(%rdi), %xmm17
19427 ; AVX512-NEXT: vinserti32x4 $1, 320(%rdx), %ymm17, %ymm17
19428 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} ymm18 = ymm17[0],ymm16[0],ymm17[2],ymm16[2]
19429 ; AVX512-NEXT: vinserti64x4 $0, %ymm18, %zmm20, %zmm18
19430 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
19431 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
19432 ; AVX512-NEXT: vmovdqa64 %zmm20, %zmm31 {%k1}
19433 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} ymm16 = ymm17[1],ymm16[1],ymm17[3],ymm16[3]
19434 ; AVX512-NEXT: vinserti64x4 $0, %ymm16, %zmm31, %zmm16
19435 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
19436 ; AVX512-NEXT: vmovdqa64 %zmm17, %zmm24 {%k1}
19437 ; AVX512-NEXT: vmovdqa64 384(%rsi), %xmm17
19438 ; AVX512-NEXT: vinserti32x4 $1, 384(%rcx), %ymm17, %ymm17
19439 ; AVX512-NEXT: vmovdqa64 384(%rdi), %xmm20
19440 ; AVX512-NEXT: vinserti32x4 $1, 384(%rdx), %ymm20, %ymm20
19441 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} ymm31 = ymm20[0],ymm17[0],ymm20[2],ymm17[2]
19442 ; AVX512-NEXT: vinserti64x4 $0, %ymm31, %zmm24, %zmm24
19443 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
19444 ; AVX512-NEXT: vmovdqa64 %zmm31, %zmm26 {%k1}
19445 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} ymm17 = ymm20[1],ymm17[1],ymm20[3],ymm17[3]
19446 ; AVX512-NEXT: vinserti64x4 $0, %ymm17, %zmm26, %zmm17
19447 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
19448 ; AVX512-NEXT: vmovdqa64 %zmm20, %zmm11 {%k1}
19449 ; AVX512-NEXT: vmovdqa64 448(%rsi), %xmm20
19450 ; AVX512-NEXT: vinserti32x4 $1, 448(%rcx), %ymm20, %ymm20
19451 ; AVX512-NEXT: vmovdqa64 448(%rdi), %xmm26
19452 ; AVX512-NEXT: vinserti32x4 $1, 448(%rdx), %ymm26, %ymm26
19453 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} ymm31 = ymm26[0],ymm20[0],ymm26[2],ymm20[2]
19454 ; AVX512-NEXT: vinserti64x4 $0, %ymm31, %zmm11, %zmm11
19455 ; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
19456 ; AVX512-NEXT: vmovdqa64 %zmm31, %zmm12 {%k1}
19457 ; AVX512-NEXT: vpunpckhqdq {{.*#+}} ymm20 = ymm26[1],ymm20[1],ymm26[3],ymm20[3]
19458 ; AVX512-NEXT: vinserti64x4 $0, %ymm20, %zmm12, %zmm12
19459 ; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %rax
19460 ; AVX512-NEXT: vmovdqa64 %zmm9, 3776(%rax)
19461 ; AVX512-NEXT: vmovdqa64 %zmm10, 3712(%rax)
19462 ; AVX512-NEXT: vmovdqa64 %zmm13, 3264(%rax)
19463 ; AVX512-NEXT: vmovdqa64 %zmm14, 3200(%rax)
19464 ; AVX512-NEXT: vmovdqa64 %zmm27, 2752(%rax)
19465 ; AVX512-NEXT: vmovdqa64 %zmm22, 2688(%rax)
19466 ; AVX512-NEXT: vmovdqa64 %zmm28, 2240(%rax)
19467 ; AVX512-NEXT: vmovdqa64 %zmm15, 2176(%rax)
19468 ; AVX512-NEXT: vmovdqa64 %zmm19, 1728(%rax)
19469 ; AVX512-NEXT: vmovdqa64 %zmm21, 1664(%rax)
19470 ; AVX512-NEXT: vmovdqa64 %zmm25, 1216(%rax)
19471 ; AVX512-NEXT: vmovdqa64 %zmm23, 1152(%rax)
19472 ; AVX512-NEXT: vmovdqa64 %zmm8, 704(%rax)
19473 ; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
19474 ; AVX512-NEXT: vmovaps %zmm8, 640(%rax)
19475 ; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
19476 ; AVX512-NEXT: vmovaps %zmm8, 192(%rax)
19477 ; AVX512-NEXT: vmovups (%rsp), %zmm8 # 64-byte Reload
19478 ; AVX512-NEXT: vmovaps %zmm8, 128(%rax)
19479 ; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
19480 ; AVX512-NEXT: vmovaps %zmm8, 4032(%rax)
19481 ; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
19482 ; AVX512-NEXT: vmovaps %zmm8, 3968(%rax)
19483 ; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
19484 ; AVX512-NEXT: vmovaps %zmm8, 3904(%rax)
19485 ; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
19486 ; AVX512-NEXT: vmovaps %zmm8, 3840(%rax)
19487 ; AVX512-NEXT: vmovdqa64 %zmm12, 3648(%rax)
19488 ; AVX512-NEXT: vmovdqa64 %zmm11, 3584(%rax)
19489 ; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
19490 ; AVX512-NEXT: vmovaps %zmm8, 3520(%rax)
19491 ; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
19492 ; AVX512-NEXT: vmovaps %zmm8, 3456(%rax)
19493 ; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
19494 ; AVX512-NEXT: vmovaps %zmm8, 3392(%rax)
19495 ; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
19496 ; AVX512-NEXT: vmovaps %zmm8, 3328(%rax)
19497 ; AVX512-NEXT: vmovdqa64 %zmm17, 3136(%rax)
19498 ; AVX512-NEXT: vmovdqa64 %zmm24, 3072(%rax)
19499 ; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
19500 ; AVX512-NEXT: vmovaps %zmm8, 3008(%rax)
19501 ; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
19502 ; AVX512-NEXT: vmovaps %zmm8, 2944(%rax)
19503 ; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
19504 ; AVX512-NEXT: vmovaps %zmm8, 2880(%rax)
19505 ; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
19506 ; AVX512-NEXT: vmovaps %zmm8, 2816(%rax)
19507 ; AVX512-NEXT: vmovdqa64 %zmm16, 2624(%rax)
19508 ; AVX512-NEXT: vmovdqa64 %zmm18, 2560(%rax)
19509 ; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
19510 ; AVX512-NEXT: vmovaps %zmm8, 2496(%rax)
19511 ; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
19512 ; AVX512-NEXT: vmovaps %zmm8, 2432(%rax)
19513 ; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
19514 ; AVX512-NEXT: vmovaps %zmm8, 2368(%rax)
19515 ; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
19516 ; AVX512-NEXT: vmovaps %zmm8, 2304(%rax)
19517 ; AVX512-NEXT: vmovdqa64 %zmm30, 2112(%rax)
19518 ; AVX512-NEXT: vmovdqa64 %zmm29, 2048(%rax)
19519 ; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
19520 ; AVX512-NEXT: vmovaps %zmm8, 1984(%rax)
19521 ; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
19522 ; AVX512-NEXT: vmovaps %zmm8, 1920(%rax)
19523 ; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
19524 ; AVX512-NEXT: vmovaps %zmm8, 1856(%rax)
19525 ; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
19526 ; AVX512-NEXT: vmovaps %zmm8, 1792(%rax)
19527 ; AVX512-NEXT: vmovdqa64 %zmm1, 1600(%rax)
19528 ; AVX512-NEXT: vmovdqa64 %zmm7, 1536(%rax)
19529 ; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
19530 ; AVX512-NEXT: vmovaps %zmm1, 1472(%rax)
19531 ; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
19532 ; AVX512-NEXT: vmovaps %zmm1, 1408(%rax)
19533 ; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
19534 ; AVX512-NEXT: vmovaps %zmm1, 1344(%rax)
19535 ; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
19536 ; AVX512-NEXT: vmovaps %zmm1, 1280(%rax)
19537 ; AVX512-NEXT: vmovdqa64 %zmm6, 1088(%rax)
19538 ; AVX512-NEXT: vmovdqa64 %zmm4, 1024(%rax)
19539 ; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
19540 ; AVX512-NEXT: vmovaps %zmm1, 960(%rax)
19541 ; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
19542 ; AVX512-NEXT: vmovaps %zmm1, 896(%rax)
19543 ; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
19544 ; AVX512-NEXT: vmovaps %zmm1, 832(%rax)
19545 ; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
19546 ; AVX512-NEXT: vmovaps %zmm1, 768(%rax)
19547 ; AVX512-NEXT: vmovdqa64 %zmm5, 576(%rax)
19548 ; AVX512-NEXT: vmovdqa64 %zmm3, 512(%rax)
19549 ; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
19550 ; AVX512-NEXT: vmovaps %zmm1, 448(%rax)
19551 ; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
19552 ; AVX512-NEXT: vmovaps %zmm1, 384(%rax)
19553 ; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
19554 ; AVX512-NEXT: vmovaps %zmm1, 320(%rax)
19555 ; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
19556 ; AVX512-NEXT: vmovaps %zmm1, 256(%rax)
19557 ; AVX512-NEXT: vmovdqa64 %zmm2, 64(%rax)
19558 ; AVX512-NEXT: vmovdqa64 %zmm0, (%rax)
19559 ; AVX512-NEXT: addq $5384, %rsp # imm = 0x1508
19560 ; AVX512-NEXT: vzeroupper
19561 ; AVX512-NEXT: retq
19562 ;
19563 ; AVX512-FCP-LABEL: store_i64_stride8_vf64:
19564 ; AVX512-FCP: # %bb.0:
19565 ; AVX512-FCP-NEXT: subq $5384, %rsp # imm = 0x1508
19566 ; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
19567 ; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10
19568 ; AVX512-FCP-NEXT: vmovdqa64 128(%rdi), %zmm0
19569 ; AVX512-FCP-NEXT: vmovdqa64 64(%rdi), %zmm4
19570 ; AVX512-FCP-NEXT: vmovdqa64 (%rdi), %zmm14
19571 ; AVX512-FCP-NEXT: vmovdqa64 128(%rsi), %zmm1
19572 ; AVX512-FCP-NEXT: vmovdqa64 64(%rsi), %zmm5
19573 ; AVX512-FCP-NEXT: vmovdqa64 (%rsi), %zmm10
19574 ; AVX512-FCP-NEXT: vmovdqa64 128(%rdx), %zmm2
19575 ; AVX512-FCP-NEXT: vmovdqa64 64(%rdx), %zmm7
19576 ; AVX512-FCP-NEXT: vmovdqa64 (%rdx), %zmm12
19577 ; AVX512-FCP-NEXT: vmovdqa64 128(%rcx), %zmm3
19578 ; AVX512-FCP-NEXT: vmovdqa64 64(%rcx), %zmm9
19579 ; AVX512-FCP-NEXT: vmovdqa64 (%rcx), %zmm15
19580 ; AVX512-FCP-NEXT: vmovdqa64 (%r8), %zmm30
19581 ; AVX512-FCP-NEXT: vmovdqa64 64(%r8), %zmm18
19582 ; AVX512-FCP-NEXT: vmovdqa64 128(%r8), %zmm11
19583 ; AVX512-FCP-NEXT: vmovdqa64 (%r9), %zmm24
19584 ; AVX512-FCP-NEXT: vmovdqa64 64(%r9), %zmm28
19585 ; AVX512-FCP-NEXT: vmovdqa64 128(%r9), %zmm22
19586 ; AVX512-FCP-NEXT: vmovdqa64 (%r10), %zmm26
19587 ; AVX512-FCP-NEXT: vmovdqa64 64(%r10), %zmm31
19588 ; AVX512-FCP-NEXT: vmovdqa64 128(%r10), %zmm16
19589 ; AVX512-FCP-NEXT: vmovdqa64 (%rax), %zmm17
19590 ; AVX512-FCP-NEXT: vmovdqa64 64(%rax), %zmm27
19591 ; AVX512-FCP-NEXT: vmovdqa64 128(%rax), %zmm13
19592 ; AVX512-FCP-NEXT: movb $-64, %r11b
19593 ; AVX512-FCP-NEXT: kmovw %r11d, %k1
19594 ; AVX512-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm19 = [4,12,4,12,4,12,4,12]
19595 ; AVX512-FCP-NEXT: # zmm19 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
19596 ; AVX512-FCP-NEXT: vmovdqa64 %zmm12, %zmm8
19597 ; AVX512-FCP-NEXT: vpermt2q %zmm15, %zmm19, %zmm8
19598 ; AVX512-FCP-NEXT: vmovdqa64 %zmm14, %zmm6
19599 ; AVX512-FCP-NEXT: vpermt2q %zmm10, %zmm19, %zmm6
19600 ; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm8[4,5,6,7]
19601 ; AVX512-FCP-NEXT: vmovdqa64 %zmm26, %zmm8
19602 ; AVX512-FCP-NEXT: vpermt2q %zmm17, %zmm19, %zmm8
19603 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm20 = zmm30[0],zmm24[0],zmm30[2],zmm24[2],zmm30[4],zmm24[4],zmm30[6],zmm24[6]
19604 ; AVX512-FCP-NEXT: vmovdqa64 %zmm8, %zmm20 {%k1}
19605 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm20, %zmm6
19606 ; AVX512-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19607 ; AVX512-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm21 = [5,13,5,13,5,13,5,13]
19608 ; AVX512-FCP-NEXT: # zmm21 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
19609 ; AVX512-FCP-NEXT: vmovdqa64 %zmm12, %zmm6
19610 ; AVX512-FCP-NEXT: vpermt2q %zmm15, %zmm21, %zmm6
19611 ; AVX512-FCP-NEXT: vmovdqa64 %zmm14, %zmm8
19612 ; AVX512-FCP-NEXT: vpermt2q %zmm10, %zmm21, %zmm8
19613 ; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm8[0,1,2,3],ymm6[4,5,6,7]
19614 ; AVX512-FCP-NEXT: vmovdqa64 %zmm26, %zmm8
19615 ; AVX512-FCP-NEXT: vpermt2q %zmm17, %zmm21, %zmm8
19616 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm20 = zmm30[1],zmm24[1],zmm30[3],zmm24[3],zmm30[5],zmm24[5],zmm30[7],zmm24[7]
19617 ; AVX512-FCP-NEXT: vmovdqa64 %zmm8, %zmm20 {%k1}
19618 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm20, %zmm6
19619 ; AVX512-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19620 ; AVX512-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm23 = [6,14,6,14,6,14,6,14]
19621 ; AVX512-FCP-NEXT: # zmm23 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
19622 ; AVX512-FCP-NEXT: vmovdqa64 %zmm12, %zmm6
19623 ; AVX512-FCP-NEXT: vpermt2q %zmm15, %zmm23, %zmm6
19624 ; AVX512-FCP-NEXT: vmovdqa64 %zmm14, %zmm8
19625 ; AVX512-FCP-NEXT: vpermt2q %zmm10, %zmm23, %zmm8
19626 ; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm8[0,1,2,3],ymm6[4,5,6,7]
19627 ; AVX512-FCP-NEXT: vmovdqa64 %zmm30, %zmm8
19628 ; AVX512-FCP-NEXT: vpermt2q %zmm24, %zmm23, %zmm8
19629 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm8 {%k1} = zmm26[0],zmm17[0],zmm26[2],zmm17[2],zmm26[4],zmm17[4],zmm26[6],zmm17[6]
19630 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm8, %zmm6
19631 ; AVX512-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19632 ; AVX512-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm25 = [7,15,7,15,7,15,7,15]
19633 ; AVX512-FCP-NEXT: # zmm25 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
19634 ; AVX512-FCP-NEXT: vpermt2q %zmm15, %zmm25, %zmm12
19635 ; AVX512-FCP-NEXT: vpermt2q %zmm10, %zmm25, %zmm14
19636 ; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm14[0,1,2,3],ymm12[4,5,6,7]
19637 ; AVX512-FCP-NEXT: vmovdqa64 %zmm30, %zmm8
19638 ; AVX512-FCP-NEXT: vpermt2q %zmm24, %zmm25, %zmm8
19639 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm8 {%k1} = zmm26[1],zmm17[1],zmm26[3],zmm17[3],zmm26[5],zmm17[5],zmm26[7],zmm17[7]
19640 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm8, %zmm6
19641 ; AVX512-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19642 ; AVX512-FCP-NEXT: vmovdqa64 %zmm7, %zmm6
19643 ; AVX512-FCP-NEXT: vpermt2q %zmm9, %zmm19, %zmm6
19644 ; AVX512-FCP-NEXT: vmovdqa64 %zmm4, %zmm8
19645 ; AVX512-FCP-NEXT: vpermt2q %zmm5, %zmm19, %zmm8
19646 ; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm8[0,1,2,3],ymm6[4,5,6,7]
19647 ; AVX512-FCP-NEXT: vmovdqa64 %zmm31, %zmm8
19648 ; AVX512-FCP-NEXT: vpermt2q %zmm27, %zmm19, %zmm8
19649 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm10 = zmm18[0],zmm28[0],zmm18[2],zmm28[2],zmm18[4],zmm28[4],zmm18[6],zmm28[6]
19650 ; AVX512-FCP-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}
19651 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm10, %zmm6
19652 ; AVX512-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19653 ; AVX512-FCP-NEXT: vmovdqa64 %zmm7, %zmm6
19654 ; AVX512-FCP-NEXT: vpermt2q %zmm9, %zmm21, %zmm6
19655 ; AVX512-FCP-NEXT: vmovdqa64 %zmm4, %zmm8
19656 ; AVX512-FCP-NEXT: vpermt2q %zmm5, %zmm21, %zmm8
19657 ; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm8[0,1,2,3],ymm6[4,5,6,7]
19658 ; AVX512-FCP-NEXT: vmovdqa64 %zmm31, %zmm8
19659 ; AVX512-FCP-NEXT: vpermt2q %zmm27, %zmm21, %zmm8
19660 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm10 = zmm18[1],zmm28[1],zmm18[3],zmm28[3],zmm18[5],zmm28[5],zmm18[7],zmm28[7]
19661 ; AVX512-FCP-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}
19662 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm10, %zmm6
19663 ; AVX512-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19664 ; AVX512-FCP-NEXT: vmovdqa64 %zmm7, %zmm6
19665 ; AVX512-FCP-NEXT: vpermt2q %zmm9, %zmm23, %zmm6
19666 ; AVX512-FCP-NEXT: vmovdqa64 %zmm4, %zmm8
19667 ; AVX512-FCP-NEXT: vpermt2q %zmm5, %zmm23, %zmm8
19668 ; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm8[0,1,2,3],ymm6[4,5,6,7]
19669 ; AVX512-FCP-NEXT: vmovdqu64 %zmm18, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19670 ; AVX512-FCP-NEXT: vmovdqa64 %zmm18, %zmm8
19671 ; AVX512-FCP-NEXT: vpermt2q %zmm28, %zmm23, %zmm8
19672 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm8 {%k1} = zmm31[0],zmm27[0],zmm31[2],zmm27[2],zmm31[4],zmm27[4],zmm31[6],zmm27[6]
19673 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm8, %zmm6
19674 ; AVX512-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19675 ; AVX512-FCP-NEXT: vpermt2q %zmm9, %zmm25, %zmm7
19676 ; AVX512-FCP-NEXT: vpermt2q %zmm5, %zmm25, %zmm4
19677 ; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm7[4,5,6,7]
19678 ; AVX512-FCP-NEXT: vmovdqa64 %zmm18, %zmm5
19679 ; AVX512-FCP-NEXT: vpermt2q %zmm28, %zmm25, %zmm5
19680 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm5 {%k1} = zmm31[1],zmm27[1],zmm31[3],zmm27[3],zmm31[5],zmm27[5],zmm31[7],zmm27[7]
19681 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm4, %zmm5, %zmm4
19682 ; AVX512-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19683 ; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm4
19684 ; AVX512-FCP-NEXT: vpermt2q %zmm3, %zmm19, %zmm4
19685 ; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm5
19686 ; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm19, %zmm5
19687 ; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm5[0,1,2,3],ymm4[4,5,6,7]
19688 ; AVX512-FCP-NEXT: vmovdqa64 %zmm16, %zmm5
19689 ; AVX512-FCP-NEXT: vpermt2q %zmm13, %zmm19, %zmm5
19690 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm6 = zmm11[0],zmm22[0],zmm11[2],zmm22[2],zmm11[4],zmm22[4],zmm11[6],zmm22[6]
19691 ; AVX512-FCP-NEXT: vmovdqa64 %zmm5, %zmm6 {%k1}
19692 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm4, %zmm6, %zmm4
19693 ; AVX512-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19694 ; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm4
19695 ; AVX512-FCP-NEXT: vpermt2q %zmm3, %zmm21, %zmm4
19696 ; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm5
19697 ; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm21, %zmm5
19698 ; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm5[0,1,2,3],ymm4[4,5,6,7]
19699 ; AVX512-FCP-NEXT: vmovdqa64 %zmm16, %zmm5
19700 ; AVX512-FCP-NEXT: vpermt2q %zmm13, %zmm21, %zmm5
19701 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm6 = zmm11[1],zmm22[1],zmm11[3],zmm22[3],zmm11[5],zmm22[5],zmm11[7],zmm22[7]
19702 ; AVX512-FCP-NEXT: vmovdqa64 %zmm5, %zmm6 {%k1}
19703 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm4, %zmm6, %zmm4
19704 ; AVX512-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19705 ; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm4
19706 ; AVX512-FCP-NEXT: vpermt2q %zmm3, %zmm23, %zmm4
19707 ; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm5
19708 ; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm23, %zmm5
19709 ; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm5[0,1,2,3],ymm4[4,5,6,7]
19710 ; AVX512-FCP-NEXT: vmovdqa64 %zmm11, %zmm5
19711 ; AVX512-FCP-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19712 ; AVX512-FCP-NEXT: vmovdqu64 %zmm22, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19713 ; AVX512-FCP-NEXT: vpermt2q %zmm22, %zmm23, %zmm5
19714 ; AVX512-FCP-NEXT: vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19715 ; AVX512-FCP-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19716 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm5 {%k1} = zmm16[0],zmm13[0],zmm16[2],zmm13[2],zmm16[4],zmm13[4],zmm16[6],zmm13[6]
19717 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm4, %zmm5, %zmm4
19718 ; AVX512-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19719 ; AVX512-FCP-NEXT: vmovdqa64 192(%rdi), %zmm4
19720 ; AVX512-FCP-NEXT: vpermt2q %zmm3, %zmm25, %zmm2
19721 ; AVX512-FCP-NEXT: vmovdqa64 192(%rdx), %zmm3
19722 ; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm25, %zmm0
19723 ; AVX512-FCP-NEXT: vmovdqa64 192(%rcx), %zmm1
19724 ; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
19725 ; AVX512-FCP-NEXT: vmovdqa64 %zmm11, %zmm2
19726 ; AVX512-FCP-NEXT: vpermt2q %zmm22, %zmm25, %zmm2
19727 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm2 {%k1} = zmm16[1],zmm13[1],zmm16[3],zmm13[3],zmm16[5],zmm13[5],zmm16[7],zmm13[7]
19728 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm2, %zmm0
19729 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19730 ; AVX512-FCP-NEXT: vmovdqa64 %zmm3, %zmm0
19731 ; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm19, %zmm0
19732 ; AVX512-FCP-NEXT: vmovdqa64 192(%rsi), %zmm2
19733 ; AVX512-FCP-NEXT: vmovdqa64 %zmm4, %zmm5
19734 ; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm19, %zmm5
19735 ; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm0[4,5,6,7]
19736 ; AVX512-FCP-NEXT: vmovdqa64 192(%r10), %zmm9
19737 ; AVX512-FCP-NEXT: vmovdqa64 192(%rax), %zmm20
19738 ; AVX512-FCP-NEXT: vmovdqa64 %zmm9, %zmm6
19739 ; AVX512-FCP-NEXT: vpermt2q %zmm20, %zmm19, %zmm6
19740 ; AVX512-FCP-NEXT: vmovdqa64 192(%r8), %zmm8
19741 ; AVX512-FCP-NEXT: vmovdqa64 192(%r9), %zmm0
19742 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm7 = zmm8[0],zmm0[0],zmm8[2],zmm0[2],zmm8[4],zmm0[4],zmm8[6],zmm0[6]
19743 ; AVX512-FCP-NEXT: vmovdqa64 %zmm6, %zmm7 {%k1}
19744 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm5, %zmm7, %zmm5
19745 ; AVX512-FCP-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19746 ; AVX512-FCP-NEXT: vmovdqa64 %zmm3, %zmm5
19747 ; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm21, %zmm5
19748 ; AVX512-FCP-NEXT: vmovdqa64 %zmm4, %zmm6
19749 ; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm21, %zmm6
19750 ; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm5 = ymm6[0,1,2,3],ymm5[4,5,6,7]
19751 ; AVX512-FCP-NEXT: vmovdqa64 %zmm9, %zmm6
19752 ; AVX512-FCP-NEXT: vpermt2q %zmm20, %zmm21, %zmm6
19753 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm7 = zmm8[1],zmm0[1],zmm8[3],zmm0[3],zmm8[5],zmm0[5],zmm8[7],zmm0[7]
19754 ; AVX512-FCP-NEXT: vmovdqa64 %zmm6, %zmm7 {%k1}
19755 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm5, %zmm7, %zmm5
19756 ; AVX512-FCP-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19757 ; AVX512-FCP-NEXT: vmovdqa64 %zmm3, %zmm5
19758 ; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm23, %zmm5
19759 ; AVX512-FCP-NEXT: vmovdqa64 %zmm4, %zmm6
19760 ; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm23, %zmm6
19761 ; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm5 = ymm6[0,1,2,3],ymm5[4,5,6,7]
19762 ; AVX512-FCP-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19763 ; AVX512-FCP-NEXT: vmovdqa64 %zmm8, %zmm6
19764 ; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm23, %zmm6
19765 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm6 {%k1} = zmm9[0],zmm20[0],zmm9[2],zmm20[2],zmm9[4],zmm20[4],zmm9[6],zmm20[6]
19766 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm5, %zmm6, %zmm5
19767 ; AVX512-FCP-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19768 ; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm25, %zmm3
19769 ; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm25, %zmm4
19770 ; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm1 = ymm4[0,1,2,3],ymm3[4,5,6,7]
19771 ; AVX512-FCP-NEXT: vmovdqa64 %zmm8, %zmm2
19772 ; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm25, %zmm2
19773 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm2 {%k1} = zmm9[1],zmm20[1],zmm9[3],zmm20[3],zmm9[5],zmm20[5],zmm9[7],zmm20[7]
19774 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm1, %zmm2, %zmm1
19775 ; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19776 ; AVX512-FCP-NEXT: vmovdqa64 256(%rdx), %zmm1
19777 ; AVX512-FCP-NEXT: vmovdqa64 256(%rcx), %zmm3
19778 ; AVX512-FCP-NEXT: vmovdqa64 %zmm1, %zmm2
19779 ; AVX512-FCP-NEXT: vpermt2q %zmm3, %zmm19, %zmm2
19780 ; AVX512-FCP-NEXT: vmovdqa64 256(%rdi), %zmm4
19781 ; AVX512-FCP-NEXT: vmovdqa64 256(%rsi), %zmm5
19782 ; AVX512-FCP-NEXT: vmovdqa64 %zmm4, %zmm6
19783 ; AVX512-FCP-NEXT: vpermt2q %zmm5, %zmm19, %zmm6
19784 ; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm2[4,5,6,7]
19785 ; AVX512-FCP-NEXT: vmovdqa64 256(%r10), %zmm13
19786 ; AVX512-FCP-NEXT: vmovdqa64 256(%rax), %zmm2
19787 ; AVX512-FCP-NEXT: vmovdqa64 %zmm13, %zmm7
19788 ; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm19, %zmm7
19789 ; AVX512-FCP-NEXT: vmovdqa64 256(%r8), %zmm18
19790 ; AVX512-FCP-NEXT: vmovdqa64 256(%r9), %zmm15
19791 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm8 = zmm18[0],zmm15[0],zmm18[2],zmm15[2],zmm18[4],zmm15[4],zmm18[6],zmm15[6]
19792 ; AVX512-FCP-NEXT: vmovdqa64 %zmm7, %zmm8 {%k1}
19793 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm8, %zmm6
19794 ; AVX512-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19795 ; AVX512-FCP-NEXT: vmovdqa64 %zmm1, %zmm6
19796 ; AVX512-FCP-NEXT: vpermt2q %zmm3, %zmm21, %zmm6
19797 ; AVX512-FCP-NEXT: vmovdqa64 %zmm4, %zmm7
19798 ; AVX512-FCP-NEXT: vpermt2q %zmm5, %zmm21, %zmm7
19799 ; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]
19800 ; AVX512-FCP-NEXT: vmovdqa64 %zmm13, %zmm7
19801 ; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm21, %zmm7
19802 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm8 = zmm18[1],zmm15[1],zmm18[3],zmm15[3],zmm18[5],zmm15[5],zmm18[7],zmm15[7]
19803 ; AVX512-FCP-NEXT: vmovdqa64 %zmm7, %zmm8 {%k1}
19804 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm8, %zmm6
19805 ; AVX512-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19806 ; AVX512-FCP-NEXT: vmovdqa64 %zmm1, %zmm6
19807 ; AVX512-FCP-NEXT: vpermt2q %zmm3, %zmm23, %zmm6
19808 ; AVX512-FCP-NEXT: vmovdqa64 %zmm4, %zmm7
19809 ; AVX512-FCP-NEXT: vpermt2q %zmm5, %zmm23, %zmm7
19810 ; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]
19811 ; AVX512-FCP-NEXT: vmovdqa64 %zmm18, %zmm7
19812 ; AVX512-FCP-NEXT: vpermt2q %zmm15, %zmm23, %zmm7
19813 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm7 {%k1} = zmm13[0],zmm2[0],zmm13[2],zmm2[2],zmm13[4],zmm2[4],zmm13[6],zmm2[6]
19814 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm7, %zmm6
19815 ; AVX512-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19816 ; AVX512-FCP-NEXT: vpermt2q %zmm3, %zmm25, %zmm1
19817 ; AVX512-FCP-NEXT: vpermt2q %zmm5, %zmm25, %zmm4
19818 ; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm1 = ymm4[0,1,2,3],ymm1[4,5,6,7]
19819 ; AVX512-FCP-NEXT: vmovdqa64 %zmm18, %zmm3
19820 ; AVX512-FCP-NEXT: vpermt2q %zmm15, %zmm25, %zmm3
19821 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm3 {%k1} = zmm13[1],zmm2[1],zmm13[3],zmm2[3],zmm13[5],zmm2[5],zmm13[7],zmm2[7]
19822 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm1, %zmm3, %zmm1
19823 ; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19824 ; AVX512-FCP-NEXT: vmovdqa64 320(%rdx), %zmm1
19825 ; AVX512-FCP-NEXT: vmovdqa64 320(%rcx), %zmm4
19826 ; AVX512-FCP-NEXT: vmovdqa64 %zmm1, %zmm3
19827 ; AVX512-FCP-NEXT: vpermt2q %zmm4, %zmm19, %zmm3
19828 ; AVX512-FCP-NEXT: vmovdqa64 320(%rdi), %zmm6
19829 ; AVX512-FCP-NEXT: vmovdqa64 320(%rsi), %zmm7
19830 ; AVX512-FCP-NEXT: vmovdqa64 %zmm6, %zmm5
19831 ; AVX512-FCP-NEXT: vpermt2q %zmm7, %zmm19, %zmm5
19832 ; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm8 = ymm5[0,1,2,3],ymm3[4,5,6,7]
19833 ; AVX512-FCP-NEXT: vmovdqa64 320(%r10), %zmm16
19834 ; AVX512-FCP-NEXT: vmovdqa64 320(%rax), %zmm22
19835 ; AVX512-FCP-NEXT: vmovdqa64 %zmm16, %zmm12
19836 ; AVX512-FCP-NEXT: vpermt2q %zmm22, %zmm19, %zmm12
19837 ; AVX512-FCP-NEXT: vmovdqa64 320(%r8), %zmm29
19838 ; AVX512-FCP-NEXT: vmovdqa64 320(%r9), %zmm5
19839 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm14 = zmm29[0],zmm5[0],zmm29[2],zmm5[2],zmm29[4],zmm5[4],zmm29[6],zmm5[6]
19840 ; AVX512-FCP-NEXT: vmovdqa64 %zmm12, %zmm14 {%k1}
19841 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm8, %zmm14, %zmm3
19842 ; AVX512-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19843 ; AVX512-FCP-NEXT: vmovdqa64 %zmm1, %zmm8
19844 ; AVX512-FCP-NEXT: vpermt2q %zmm4, %zmm21, %zmm8
19845 ; AVX512-FCP-NEXT: vmovdqa64 %zmm6, %zmm12
19846 ; AVX512-FCP-NEXT: vpermt2q %zmm7, %zmm21, %zmm12
19847 ; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm8 = ymm12[0,1,2,3],ymm8[4,5,6,7]
19848 ; AVX512-FCP-NEXT: vmovdqa64 %zmm16, %zmm12
19849 ; AVX512-FCP-NEXT: vpermt2q %zmm22, %zmm21, %zmm12
19850 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm14 = zmm29[1],zmm5[1],zmm29[3],zmm5[3],zmm29[5],zmm5[5],zmm29[7],zmm5[7]
19851 ; AVX512-FCP-NEXT: vmovdqa64 %zmm12, %zmm14 {%k1}
19852 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm8, %zmm14, %zmm3
19853 ; AVX512-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19854 ; AVX512-FCP-NEXT: vmovdqa64 %zmm1, %zmm8
19855 ; AVX512-FCP-NEXT: vpermt2q %zmm4, %zmm23, %zmm8
19856 ; AVX512-FCP-NEXT: vmovdqa64 %zmm6, %zmm12
19857 ; AVX512-FCP-NEXT: vpermt2q %zmm7, %zmm23, %zmm12
19858 ; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm8 = ymm12[0,1,2,3],ymm8[4,5,6,7]
19859 ; AVX512-FCP-NEXT: vmovdqa64 %zmm29, %zmm12
19860 ; AVX512-FCP-NEXT: vpermt2q %zmm5, %zmm23, %zmm12
19861 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm12 {%k1} = zmm16[0],zmm22[0],zmm16[2],zmm22[2],zmm16[4],zmm22[4],zmm16[6],zmm22[6]
19862 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm8, %zmm12, %zmm3
19863 ; AVX512-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19864 ; AVX512-FCP-NEXT: vpermt2q %zmm4, %zmm25, %zmm1
19865 ; AVX512-FCP-NEXT: vpermt2q %zmm7, %zmm25, %zmm6
19866 ; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm1 = ymm6[0,1,2,3],ymm1[4,5,6,7]
19867 ; AVX512-FCP-NEXT: vmovdqa64 %zmm29, %zmm4
19868 ; AVX512-FCP-NEXT: vpermt2q %zmm5, %zmm25, %zmm4
19869 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm4 {%k1} = zmm16[1],zmm22[1],zmm16[3],zmm22[3],zmm16[5],zmm22[5],zmm16[7],zmm22[7]
19870 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm1, %zmm4, %zmm1
19871 ; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19872 ; AVX512-FCP-NEXT: vmovdqa64 384(%rdx), %zmm3
19873 ; AVX512-FCP-NEXT: vmovdqa64 384(%rcx), %zmm1
19874 ; AVX512-FCP-NEXT: vmovdqa64 %zmm3, %zmm4
19875 ; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm19, %zmm4
19876 ; AVX512-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19877 ; AVX512-FCP-NEXT: vmovdqa64 %zmm3, %zmm4
19878 ; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm21, %zmm4
19879 ; AVX512-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19880 ; AVX512-FCP-NEXT: vmovdqa64 %zmm3, %zmm4
19881 ; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm23, %zmm4
19882 ; AVX512-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19883 ; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm25, %zmm3
19884 ; AVX512-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19885 ; AVX512-FCP-NEXT: vmovdqa64 384(%rdi), %zmm3
19886 ; AVX512-FCP-NEXT: vmovdqa64 384(%rsi), %zmm1
19887 ; AVX512-FCP-NEXT: vmovdqa64 %zmm3, %zmm4
19888 ; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm19, %zmm4
19889 ; AVX512-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19890 ; AVX512-FCP-NEXT: vmovdqa64 %zmm3, %zmm4
19891 ; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm21, %zmm4
19892 ; AVX512-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19893 ; AVX512-FCP-NEXT: vmovdqa64 %zmm3, %zmm4
19894 ; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm23, %zmm4
19895 ; AVX512-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19896 ; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm25, %zmm3
19897 ; AVX512-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19898 ; AVX512-FCP-NEXT: vmovdqa64 448(%rdx), %zmm3
19899 ; AVX512-FCP-NEXT: vmovdqa64 448(%rcx), %zmm1
19900 ; AVX512-FCP-NEXT: vmovdqa64 %zmm3, %zmm4
19901 ; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm19, %zmm4
19902 ; AVX512-FCP-NEXT: vmovdqu64 %zmm4, (%rsp) # 64-byte Spill
19903 ; AVX512-FCP-NEXT: vmovdqa64 %zmm3, %zmm4
19904 ; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm21, %zmm4
19905 ; AVX512-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19906 ; AVX512-FCP-NEXT: vmovdqa64 %zmm3, %zmm4
19907 ; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm23, %zmm4
19908 ; AVX512-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19909 ; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm25, %zmm3
19910 ; AVX512-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19911 ; AVX512-FCP-NEXT: vmovdqa64 448(%rdi), %zmm3
19912 ; AVX512-FCP-NEXT: vmovdqa64 448(%rsi), %zmm1
19913 ; AVX512-FCP-NEXT: vmovdqa64 %zmm3, %zmm4
19914 ; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm19, %zmm4
19915 ; AVX512-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19916 ; AVX512-FCP-NEXT: vmovdqa64 %zmm3, %zmm4
19917 ; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm21, %zmm4
19918 ; AVX512-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19919 ; AVX512-FCP-NEXT: vmovdqa64 %zmm3, %zmm4
19920 ; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm23, %zmm4
19921 ; AVX512-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19922 ; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm25, %zmm3
19923 ; AVX512-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19924 ; AVX512-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm11 = [0,8,0,8,0,8,0,8]
19925 ; AVX512-FCP-NEXT: # zmm11 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
19926 ; AVX512-FCP-NEXT: vmovdqa64 %zmm26, %zmm1
19927 ; AVX512-FCP-NEXT: vpermt2q %zmm17, %zmm11, %zmm1
19928 ; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19929 ; AVX512-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm12 = [1,9,1,9,1,9,1,9]
19930 ; AVX512-FCP-NEXT: # zmm12 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
19931 ; AVX512-FCP-NEXT: vmovdqa64 %zmm26, %zmm1
19932 ; AVX512-FCP-NEXT: vpermt2q %zmm17, %zmm12, %zmm1
19933 ; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19934 ; AVX512-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm10 = [2,10,2,10,2,10,2,10]
19935 ; AVX512-FCP-NEXT: # zmm10 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
19936 ; AVX512-FCP-NEXT: vmovdqa64 %zmm26, %zmm1
19937 ; AVX512-FCP-NEXT: vpermt2q %zmm17, %zmm10, %zmm1
19938 ; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19939 ; AVX512-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm4 = [3,11,3,11,3,11,3,11]
19940 ; AVX512-FCP-NEXT: # zmm4 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
19941 ; AVX512-FCP-NEXT: vpermt2q %zmm17, %zmm4, %zmm26
19942 ; AVX512-FCP-NEXT: vmovdqu64 %zmm26, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19943 ; AVX512-FCP-NEXT: vmovdqa64 %zmm30, %zmm1
19944 ; AVX512-FCP-NEXT: vpermt2q %zmm24, %zmm11, %zmm1
19945 ; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19946 ; AVX512-FCP-NEXT: vmovdqa64 %zmm30, %zmm1
19947 ; AVX512-FCP-NEXT: vpermt2q %zmm24, %zmm12, %zmm1
19948 ; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19949 ; AVX512-FCP-NEXT: vmovdqa64 %zmm30, %zmm1
19950 ; AVX512-FCP-NEXT: vpermt2q %zmm24, %zmm10, %zmm1
19951 ; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19952 ; AVX512-FCP-NEXT: vpermt2q %zmm24, %zmm4, %zmm30
19953 ; AVX512-FCP-NEXT: vmovdqu64 %zmm30, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19954 ; AVX512-FCP-NEXT: vmovdqa64 %zmm31, %zmm1
19955 ; AVX512-FCP-NEXT: vpermt2q %zmm27, %zmm11, %zmm1
19956 ; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19957 ; AVX512-FCP-NEXT: vmovdqa64 %zmm31, %zmm1
19958 ; AVX512-FCP-NEXT: vpermt2q %zmm27, %zmm12, %zmm1
19959 ; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19960 ; AVX512-FCP-NEXT: vmovdqa64 %zmm31, %zmm1
19961 ; AVX512-FCP-NEXT: vpermt2q %zmm27, %zmm10, %zmm1
19962 ; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19963 ; AVX512-FCP-NEXT: vpermt2q %zmm27, %zmm4, %zmm31
19964 ; AVX512-FCP-NEXT: vmovdqu64 %zmm31, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19965 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
19966 ; AVX512-FCP-NEXT: vmovdqa64 %zmm3, %zmm1
19967 ; AVX512-FCP-NEXT: vpermt2q %zmm28, %zmm11, %zmm1
19968 ; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19969 ; AVX512-FCP-NEXT: vmovdqa64 %zmm3, %zmm1
19970 ; AVX512-FCP-NEXT: vpermt2q %zmm28, %zmm12, %zmm1
19971 ; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19972 ; AVX512-FCP-NEXT: vmovdqa64 %zmm3, %zmm31
19973 ; AVX512-FCP-NEXT: vpermt2q %zmm28, %zmm10, %zmm31
19974 ; AVX512-FCP-NEXT: vpermt2q %zmm28, %zmm4, %zmm3
19975 ; AVX512-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19976 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
19977 ; AVX512-FCP-NEXT: vmovdqa64 %zmm28, %zmm3
19978 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
19979 ; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm11, %zmm3
19980 ; AVX512-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19981 ; AVX512-FCP-NEXT: vmovdqa64 %zmm28, %zmm3
19982 ; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm12, %zmm3
19983 ; AVX512-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19984 ; AVX512-FCP-NEXT: vmovdqa64 %zmm28, %zmm3
19985 ; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm10, %zmm3
19986 ; AVX512-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19987 ; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm4, %zmm28
19988 ; AVX512-FCP-NEXT: vmovdqu64 %zmm28, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19989 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
19990 ; AVX512-FCP-NEXT: vmovdqa64 %zmm6, %zmm3
19991 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
19992 ; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm11, %zmm3
19993 ; AVX512-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19994 ; AVX512-FCP-NEXT: vmovdqa64 %zmm6, %zmm3
19995 ; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm12, %zmm3
19996 ; AVX512-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
19997 ; AVX512-FCP-NEXT: vmovdqa64 %zmm6, %zmm27
19998 ; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm10, %zmm27
19999 ; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm4, %zmm6
20000 ; AVX512-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20001 ; AVX512-FCP-NEXT: vmovdqa64 %zmm9, %zmm1
20002 ; AVX512-FCP-NEXT: vpermt2q %zmm20, %zmm11, %zmm1
20003 ; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20004 ; AVX512-FCP-NEXT: vmovdqa64 %zmm9, %zmm1
20005 ; AVX512-FCP-NEXT: vpermt2q %zmm20, %zmm12, %zmm1
20006 ; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20007 ; AVX512-FCP-NEXT: vmovdqa64 %zmm9, %zmm1
20008 ; AVX512-FCP-NEXT: vpermt2q %zmm20, %zmm10, %zmm1
20009 ; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20010 ; AVX512-FCP-NEXT: vpermt2q %zmm20, %zmm4, %zmm9
20011 ; AVX512-FCP-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20012 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
20013 ; AVX512-FCP-NEXT: vmovdqa64 %zmm1, %zmm3
20014 ; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm11, %zmm3
20015 ; AVX512-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20016 ; AVX512-FCP-NEXT: vmovdqa64 %zmm1, %zmm3
20017 ; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm12, %zmm3
20018 ; AVX512-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20019 ; AVX512-FCP-NEXT: vmovdqa64 %zmm1, %zmm20
20020 ; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm10, %zmm20
20021 ; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm4, %zmm1
20022 ; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20023 ; AVX512-FCP-NEXT: vmovdqa64 %zmm13, %zmm0
20024 ; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm11, %zmm0
20025 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20026 ; AVX512-FCP-NEXT: vmovdqa64 %zmm13, %zmm0
20027 ; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm12, %zmm0
20028 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20029 ; AVX512-FCP-NEXT: vmovdqa64 %zmm13, %zmm0
20030 ; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm10, %zmm0
20031 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20032 ; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm4, %zmm13
20033 ; AVX512-FCP-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20034 ; AVX512-FCP-NEXT: vmovdqa64 %zmm18, %zmm0
20035 ; AVX512-FCP-NEXT: vpermt2q %zmm15, %zmm11, %zmm0
20036 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20037 ; AVX512-FCP-NEXT: vmovdqa64 %zmm18, %zmm0
20038 ; AVX512-FCP-NEXT: vpermt2q %zmm15, %zmm12, %zmm0
20039 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20040 ; AVX512-FCP-NEXT: vmovdqa64 %zmm18, %zmm28
20041 ; AVX512-FCP-NEXT: vpermt2q %zmm15, %zmm10, %zmm28
20042 ; AVX512-FCP-NEXT: vpermt2q %zmm15, %zmm4, %zmm18
20043 ; AVX512-FCP-NEXT: vmovdqu64 %zmm18, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20044 ; AVX512-FCP-NEXT: vmovdqa64 %zmm16, %zmm0
20045 ; AVX512-FCP-NEXT: vpermt2q %zmm22, %zmm11, %zmm0
20046 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20047 ; AVX512-FCP-NEXT: vmovdqa64 %zmm16, %zmm0
20048 ; AVX512-FCP-NEXT: vpermt2q %zmm22, %zmm12, %zmm0
20049 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20050 ; AVX512-FCP-NEXT: vmovdqa64 %zmm16, %zmm0
20051 ; AVX512-FCP-NEXT: vpermt2q %zmm22, %zmm10, %zmm0
20052 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20053 ; AVX512-FCP-NEXT: vpermt2q %zmm22, %zmm4, %zmm16
20054 ; AVX512-FCP-NEXT: vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20055 ; AVX512-FCP-NEXT: vmovdqa64 %zmm29, %zmm0
20056 ; AVX512-FCP-NEXT: vpermt2q %zmm5, %zmm11, %zmm0
20057 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20058 ; AVX512-FCP-NEXT: vmovdqa64 %zmm29, %zmm0
20059 ; AVX512-FCP-NEXT: vpermt2q %zmm5, %zmm12, %zmm0
20060 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20061 ; AVX512-FCP-NEXT: vmovdqa64 %zmm29, %zmm22
20062 ; AVX512-FCP-NEXT: vpermt2q %zmm5, %zmm10, %zmm22
20063 ; AVX512-FCP-NEXT: vpermt2q %zmm5, %zmm4, %zmm29
20064 ; AVX512-FCP-NEXT: vmovdqa64 384(%r10), %zmm13
20065 ; AVX512-FCP-NEXT: vmovdqa64 384(%rax), %zmm0
20066 ; AVX512-FCP-NEXT: vmovdqa64 %zmm13, %zmm1
20067 ; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm11, %zmm1
20068 ; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20069 ; AVX512-FCP-NEXT: vmovdqa64 %zmm13, %zmm1
20070 ; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm12, %zmm1
20071 ; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20072 ; AVX512-FCP-NEXT: vmovdqa64 %zmm13, %zmm14
20073 ; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm19, %zmm14
20074 ; AVX512-FCP-NEXT: vmovdqa64 %zmm13, %zmm15
20075 ; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm21, %zmm15
20076 ; AVX512-FCP-NEXT: vmovdqa64 384(%r8), %zmm5
20077 ; AVX512-FCP-NEXT: vmovdqa64 384(%r9), %zmm2
20078 ; AVX512-FCP-NEXT: vmovdqa64 %zmm5, %zmm8
20079 ; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm23, %zmm8
20080 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm8 {%k1} = zmm13[0],zmm0[0],zmm13[2],zmm0[2],zmm13[4],zmm0[4],zmm13[6],zmm0[6]
20081 ; AVX512-FCP-NEXT: vmovdqa64 %zmm5, %zmm6
20082 ; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm25, %zmm6
20083 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm6 {%k1} = zmm13[1],zmm0[1],zmm13[3],zmm0[3],zmm13[5],zmm0[5],zmm13[7],zmm0[7]
20084 ; AVX512-FCP-NEXT: vmovdqa64 %zmm13, %zmm30
20085 ; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm10, %zmm30
20086 ; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm4, %zmm13
20087 ; AVX512-FCP-NEXT: vmovdqa64 %zmm5, %zmm24
20088 ; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm11, %zmm24
20089 ; AVX512-FCP-NEXT: vmovdqa64 %zmm5, %zmm26
20090 ; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm12, %zmm26
20091 ; AVX512-FCP-NEXT: vmovdqa64 %zmm5, %zmm17
20092 ; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm10, %zmm17
20093 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm7 = zmm5[0],zmm2[0],zmm5[2],zmm2[2],zmm5[4],zmm2[4],zmm5[6],zmm2[6]
20094 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm9 = zmm5[1],zmm2[1],zmm5[3],zmm2[3],zmm5[5],zmm2[5],zmm5[7],zmm2[7]
20095 ; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm4, %zmm5
20096 ; AVX512-FCP-NEXT: vmovdqa64 448(%r10), %zmm16
20097 ; AVX512-FCP-NEXT: vmovdqa64 448(%rax), %zmm1
20098 ; AVX512-FCP-NEXT: vmovdqa64 %zmm16, %zmm0
20099 ; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm11, %zmm0
20100 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20101 ; AVX512-FCP-NEXT: vmovdqa64 %zmm16, %zmm0
20102 ; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm12, %zmm0
20103 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20104 ; AVX512-FCP-NEXT: vpermi2q %zmm1, %zmm16, %zmm19
20105 ; AVX512-FCP-NEXT: vpermi2q %zmm1, %zmm16, %zmm21
20106 ; AVX512-FCP-NEXT: vmovdqa64 448(%r8), %zmm3
20107 ; AVX512-FCP-NEXT: vmovdqa64 448(%r9), %zmm0
20108 ; AVX512-FCP-NEXT: vpermi2q %zmm0, %zmm3, %zmm23
20109 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm23 {%k1} = zmm16[0],zmm1[0],zmm16[2],zmm1[2],zmm16[4],zmm1[4],zmm16[6],zmm1[6]
20110 ; AVX512-FCP-NEXT: vpermi2q %zmm0, %zmm3, %zmm25
20111 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm25 {%k1} = zmm16[1],zmm1[1],zmm16[3],zmm1[3],zmm16[5],zmm1[5],zmm16[7],zmm1[7]
20112 ; AVX512-FCP-NEXT: vmovdqa64 %zmm16, %zmm18
20113 ; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm10, %zmm18
20114 ; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm4, %zmm16
20115 ; AVX512-FCP-NEXT: vpermi2q %zmm0, %zmm3, %zmm11
20116 ; AVX512-FCP-NEXT: vpermi2q %zmm0, %zmm3, %zmm12
20117 ; AVX512-FCP-NEXT: vpermi2q %zmm0, %zmm3, %zmm10
20118 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm1 = zmm3[0],zmm0[0],zmm3[2],zmm0[2],zmm3[4],zmm0[4],zmm3[6],zmm0[6]
20119 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm2 = zmm3[1],zmm0[1],zmm3[3],zmm0[3],zmm3[5],zmm0[5],zmm3[7],zmm0[7]
20120 ; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm4, %zmm3
20121 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
20122 ; AVX512-FCP-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
20123 ; AVX512-FCP-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
20124 ; AVX512-FCP-NEXT: vmovdqa64 %zmm14, %zmm7 {%k1}
20125 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm7, %zmm0
20126 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20127 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
20128 ; AVX512-FCP-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
20129 ; AVX512-FCP-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
20130 ; AVX512-FCP-NEXT: vmovdqa64 %zmm15, %zmm9 {%k1}
20131 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm9, %zmm0
20132 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20133 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
20134 ; AVX512-FCP-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
20135 ; AVX512-FCP-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
20136 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm8, %zmm0
20137 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20138 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
20139 ; AVX512-FCP-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
20140 ; AVX512-FCP-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
20141 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm6, %zmm0
20142 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20143 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
20144 ; AVX512-FCP-NEXT: vpblendd $240, (%rsp), %ymm0, %ymm0 # 32-byte Folded Reload
20145 ; AVX512-FCP-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
20146 ; AVX512-FCP-NEXT: vmovdqa64 %zmm19, %zmm1 {%k1}
20147 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
20148 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20149 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
20150 ; AVX512-FCP-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
20151 ; AVX512-FCP-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
20152 ; AVX512-FCP-NEXT: vmovdqa64 %zmm21, %zmm2 {%k1}
20153 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm2, %zmm0
20154 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20155 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
20156 ; AVX512-FCP-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
20157 ; AVX512-FCP-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
20158 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm23, %zmm0
20159 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20160 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
20161 ; AVX512-FCP-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
20162 ; AVX512-FCP-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
20163 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm25, %zmm0
20164 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20165 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
20166 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
20167 ; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm8 {%k1}
20168 ; AVX512-FCP-NEXT: vmovdqa (%rcx), %ymm0
20169 ; AVX512-FCP-NEXT: vmovdqa (%rdx), %ymm1
20170 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
20171 ; AVX512-FCP-NEXT: vmovdqa (%rsi), %ymm4
20172 ; AVX512-FCP-NEXT: vmovdqa (%rdi), %ymm6
20173 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm6[0],ymm4[0],ymm6[2],ymm4[2]
20174 ; AVX512-FCP-NEXT: vperm2i128 {{.*#+}} ymm2 = ymm7[2,3],ymm2[2,3]
20175 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm8, %zmm2
20176 ; AVX512-FCP-NEXT: vmovdqu64 %zmm2, (%rsp) # 64-byte Spill
20177 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
20178 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
20179 ; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm7 {%k1}
20180 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
20181 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm6[1],ymm4[1],ymm6[3],ymm4[3]
20182 ; AVX512-FCP-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
20183 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm7, %zmm0
20184 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20185 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
20186 ; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm31 {%k1}
20187 ; AVX512-FCP-NEXT: vmovdqa 64(%rcx), %ymm0
20188 ; AVX512-FCP-NEXT: vmovdqa 64(%rdx), %ymm1
20189 ; AVX512-FCP-NEXT: vmovdqa 64(%rsi), %ymm2
20190 ; AVX512-FCP-NEXT: vmovdqa 64(%rdi), %ymm4
20191 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm6 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
20192 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm4[0],ymm2[0],ymm4[2],ymm2[2]
20193 ; AVX512-FCP-NEXT: vperm2i128 {{.*#+}} ymm6 = ymm7[2,3],ymm6[2,3]
20194 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm31, %zmm6
20195 ; AVX512-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20196 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
20197 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
20198 ; AVX512-FCP-NEXT: vmovdqa64 %zmm6, %zmm7 {%k1}
20199 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
20200 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm4[1],ymm2[1],ymm4[3],ymm2[3]
20201 ; AVX512-FCP-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
20202 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm7, %zmm8
20203 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
20204 ; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm27 {%k1}
20205 ; AVX512-FCP-NEXT: vmovdqa 128(%rcx), %ymm0
20206 ; AVX512-FCP-NEXT: vmovdqa 128(%rdx), %ymm1
20207 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
20208 ; AVX512-FCP-NEXT: vmovdqa 128(%rsi), %ymm4
20209 ; AVX512-FCP-NEXT: vmovdqa 128(%rdi), %ymm6
20210 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm6[0],ymm4[0],ymm6[2],ymm4[2]
20211 ; AVX512-FCP-NEXT: vperm2i128 {{.*#+}} ymm2 = ymm7[2,3],ymm2[2,3]
20212 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm27, %zmm23
20213 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
20214 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
20215 ; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm7 {%k1}
20216 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
20217 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm6[1],ymm4[1],ymm6[3],ymm4[3]
20218 ; AVX512-FCP-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
20219 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm7, %zmm25
20220 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
20221 ; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm20 {%k1}
20222 ; AVX512-FCP-NEXT: vmovdqa 192(%rcx), %ymm0
20223 ; AVX512-FCP-NEXT: vmovdqa 192(%rdx), %ymm1
20224 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
20225 ; AVX512-FCP-NEXT: vmovdqa 192(%rsi), %ymm4
20226 ; AVX512-FCP-NEXT: vmovdqa 192(%rdi), %ymm6
20227 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm6[0],ymm4[0],ymm6[2],ymm4[2]
20228 ; AVX512-FCP-NEXT: vperm2i128 {{.*#+}} ymm2 = ymm7[2,3],ymm2[2,3]
20229 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm20, %zmm21
20230 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
20231 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
20232 ; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm7 {%k1}
20233 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
20234 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm6[1],ymm4[1],ymm6[3],ymm4[3]
20235 ; AVX512-FCP-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
20236 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm7, %zmm19
20237 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
20238 ; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm28 {%k1}
20239 ; AVX512-FCP-NEXT: vmovdqa 256(%rcx), %ymm0
20240 ; AVX512-FCP-NEXT: vmovdqa 256(%rdx), %ymm1
20241 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
20242 ; AVX512-FCP-NEXT: vmovdqa 256(%rsi), %ymm4
20243 ; AVX512-FCP-NEXT: vmovdqa 256(%rdi), %ymm6
20244 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm6[0],ymm4[0],ymm6[2],ymm4[2]
20245 ; AVX512-FCP-NEXT: vperm2i128 {{.*#+}} ymm2 = ymm7[2,3],ymm2[2,3]
20246 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm28, %zmm15
20247 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
20248 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
20249 ; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm7 {%k1}
20250 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
20251 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm6[1],ymm4[1],ymm6[3],ymm4[3]
20252 ; AVX512-FCP-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
20253 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm7, %zmm28
20254 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
20255 ; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm22 {%k1}
20256 ; AVX512-FCP-NEXT: vmovdqa 320(%rcx), %ymm0
20257 ; AVX512-FCP-NEXT: vmovdqa 320(%rdx), %ymm1
20258 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
20259 ; AVX512-FCP-NEXT: vmovdqa 320(%rsi), %ymm4
20260 ; AVX512-FCP-NEXT: vmovdqa 320(%rdi), %ymm6
20261 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm6[0],ymm4[0],ymm6[2],ymm4[2]
20262 ; AVX512-FCP-NEXT: vperm2i128 {{.*#+}} ymm2 = ymm7[2,3],ymm2[2,3]
20263 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm22, %zmm22
20264 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
20265 ; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm29 {%k1}
20266 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
20267 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm6[1],ymm4[1],ymm6[3],ymm4[3]
20268 ; AVX512-FCP-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
20269 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm29, %zmm27
20270 ; AVX512-FCP-NEXT: vmovdqa64 %zmm30, %zmm17 {%k1}
20271 ; AVX512-FCP-NEXT: vmovdqa 384(%rcx), %ymm0
20272 ; AVX512-FCP-NEXT: vmovdqa 384(%rdx), %ymm1
20273 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
20274 ; AVX512-FCP-NEXT: vmovdqa 384(%rsi), %ymm4
20275 ; AVX512-FCP-NEXT: vmovdqa 384(%rdi), %ymm6
20276 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm6[0],ymm4[0],ymm6[2],ymm4[2]
20277 ; AVX512-FCP-NEXT: vperm2i128 {{.*#+}} ymm2 = ymm7[2,3],ymm2[2,3]
20278 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm17, %zmm14
20279 ; AVX512-FCP-NEXT: vmovdqa64 %zmm13, %zmm5 {%k1}
20280 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
20281 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm6[1],ymm4[1],ymm6[3],ymm4[3]
20282 ; AVX512-FCP-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
20283 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm5, %zmm13
20284 ; AVX512-FCP-NEXT: vmovdqa64 %zmm18, %zmm10 {%k1}
20285 ; AVX512-FCP-NEXT: vmovdqa 448(%rcx), %ymm0
20286 ; AVX512-FCP-NEXT: vmovdqa 448(%rdx), %ymm1
20287 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
20288 ; AVX512-FCP-NEXT: vmovdqa 448(%rsi), %ymm4
20289 ; AVX512-FCP-NEXT: vmovdqa 448(%rdi), %ymm5
20290 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm6 = ymm5[0],ymm4[0],ymm5[2],ymm4[2]
20291 ; AVX512-FCP-NEXT: vperm2i128 {{.*#+}} ymm2 = ymm6[2,3],ymm2[2,3]
20292 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm10, %zmm10
20293 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
20294 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm5[1],ymm4[1],ymm5[3],ymm4[3]
20295 ; AVX512-FCP-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
20296 ; AVX512-FCP-NEXT: vmovdqa64 %zmm16, %zmm3 {%k1}
20297 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm3, %zmm9
20298 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
20299 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
20300 ; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm3 {%k1}
20301 ; AVX512-FCP-NEXT: vmovdqa (%rsi), %xmm0
20302 ; AVX512-FCP-NEXT: vinserti128 $1, (%rcx), %ymm0, %ymm1
20303 ; AVX512-FCP-NEXT: vmovdqa (%rdi), %xmm0
20304 ; AVX512-FCP-NEXT: vinserti128 $1, (%rdx), %ymm0, %ymm2
20305 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm2[0],ymm1[0],ymm2[2],ymm1[2]
20306 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm3, %zmm0
20307 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
20308 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
20309 ; AVX512-FCP-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
20310 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm2[1],ymm1[1],ymm2[3],ymm1[3]
20311 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm1, %zmm4, %zmm2
20312 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
20313 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
20314 ; AVX512-FCP-NEXT: vmovdqa64 %zmm1, %zmm5 {%k1}
20315 ; AVX512-FCP-NEXT: vmovdqa 64(%rsi), %xmm1
20316 ; AVX512-FCP-NEXT: vinserti128 $1, 64(%rcx), %ymm1, %ymm1
20317 ; AVX512-FCP-NEXT: vmovdqa 64(%rdi), %xmm3
20318 ; AVX512-FCP-NEXT: vinserti128 $1, 64(%rdx), %ymm3, %ymm4
20319 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm4[0],ymm1[0],ymm4[2],ymm1[2]
20320 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm3, %zmm5, %zmm3
20321 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
20322 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
20323 ; AVX512-FCP-NEXT: vmovdqa64 %zmm5, %zmm6 {%k1}
20324 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm4[1],ymm1[1],ymm4[3],ymm1[3]
20325 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm1, %zmm6, %zmm5
20326 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
20327 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
20328 ; AVX512-FCP-NEXT: vmovdqa64 %zmm1, %zmm7 {%k1}
20329 ; AVX512-FCP-NEXT: vmovdqa 128(%rsi), %xmm1
20330 ; AVX512-FCP-NEXT: vinserti128 $1, 128(%rcx), %ymm1, %ymm1
20331 ; AVX512-FCP-NEXT: vmovdqa 128(%rdi), %xmm4
20332 ; AVX512-FCP-NEXT: vinserti128 $1, 128(%rdx), %ymm4, %ymm6
20333 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm4 = ymm6[0],ymm1[0],ymm6[2],ymm1[2]
20334 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm4, %zmm7, %zmm4
20335 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
20336 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
20337 ; AVX512-FCP-NEXT: vmovdqa64 %zmm7, %zmm16 {%k1}
20338 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm6[1],ymm1[1],ymm6[3],ymm1[3]
20339 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm1, %zmm16, %zmm6
20340 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
20341 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
20342 ; AVX512-FCP-NEXT: vmovdqa64 %zmm1, %zmm17 {%k1}
20343 ; AVX512-FCP-NEXT: vmovdqa 192(%rsi), %xmm1
20344 ; AVX512-FCP-NEXT: vinserti128 $1, 192(%rcx), %ymm1, %ymm1
20345 ; AVX512-FCP-NEXT: vmovdqa 192(%rdi), %xmm7
20346 ; AVX512-FCP-NEXT: vinserti32x4 $1, 192(%rdx), %ymm7, %ymm16
20347 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm16[0],ymm1[0],ymm16[2],ymm1[2]
20348 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm7, %zmm17, %zmm7
20349 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
20350 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
20351 ; AVX512-FCP-NEXT: vmovdqa64 %zmm17, %zmm18 {%k1}
20352 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm16[1],ymm1[1],ymm16[3],ymm1[3]
20353 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm1, %zmm18, %zmm1
20354 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
20355 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
20356 ; AVX512-FCP-NEXT: vmovdqa64 %zmm16, %zmm20 {%k1}
20357 ; AVX512-FCP-NEXT: vmovdqa64 256(%rsi), %xmm16
20358 ; AVX512-FCP-NEXT: vinserti32x4 $1, 256(%rcx), %ymm16, %ymm16
20359 ; AVX512-FCP-NEXT: vmovdqa64 256(%rdi), %xmm17
20360 ; AVX512-FCP-NEXT: vinserti32x4 $1, 256(%rdx), %ymm17, %ymm17
20361 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm18 = ymm17[0],ymm16[0],ymm17[2],ymm16[2]
20362 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm18, %zmm20, %zmm29
20363 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
20364 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
20365 ; AVX512-FCP-NEXT: vmovdqa64 %zmm18, %zmm20 {%k1}
20366 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm16 = ymm17[1],ymm16[1],ymm17[3],ymm16[3]
20367 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm16, %zmm20, %zmm30
20368 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
20369 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
20370 ; AVX512-FCP-NEXT: vmovdqa64 %zmm16, %zmm20 {%k1}
20371 ; AVX512-FCP-NEXT: vmovdqa64 320(%rsi), %xmm16
20372 ; AVX512-FCP-NEXT: vinserti32x4 $1, 320(%rcx), %ymm16, %ymm16
20373 ; AVX512-FCP-NEXT: vmovdqa64 320(%rdi), %xmm17
20374 ; AVX512-FCP-NEXT: vinserti32x4 $1, 320(%rdx), %ymm17, %ymm17
20375 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm18 = ymm17[0],ymm16[0],ymm17[2],ymm16[2]
20376 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm18, %zmm20, %zmm18
20377 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
20378 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
20379 ; AVX512-FCP-NEXT: vmovdqa64 %zmm20, %zmm31 {%k1}
20380 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm16 = ymm17[1],ymm16[1],ymm17[3],ymm16[3]
20381 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm16, %zmm31, %zmm16
20382 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
20383 ; AVX512-FCP-NEXT: vmovdqa64 %zmm17, %zmm24 {%k1}
20384 ; AVX512-FCP-NEXT: vmovdqa64 384(%rsi), %xmm17
20385 ; AVX512-FCP-NEXT: vinserti32x4 $1, 384(%rcx), %ymm17, %ymm17
20386 ; AVX512-FCP-NEXT: vmovdqa64 384(%rdi), %xmm20
20387 ; AVX512-FCP-NEXT: vinserti32x4 $1, 384(%rdx), %ymm20, %ymm20
20388 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm31 = ymm20[0],ymm17[0],ymm20[2],ymm17[2]
20389 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm31, %zmm24, %zmm24
20390 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
20391 ; AVX512-FCP-NEXT: vmovdqa64 %zmm31, %zmm26 {%k1}
20392 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm17 = ymm20[1],ymm17[1],ymm20[3],ymm17[3]
20393 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm17, %zmm26, %zmm17
20394 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
20395 ; AVX512-FCP-NEXT: vmovdqa64 %zmm20, %zmm11 {%k1}
20396 ; AVX512-FCP-NEXT: vmovdqa64 448(%rsi), %xmm20
20397 ; AVX512-FCP-NEXT: vinserti32x4 $1, 448(%rcx), %ymm20, %ymm20
20398 ; AVX512-FCP-NEXT: vmovdqa64 448(%rdi), %xmm26
20399 ; AVX512-FCP-NEXT: vinserti32x4 $1, 448(%rdx), %ymm26, %ymm26
20400 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm31 = ymm26[0],ymm20[0],ymm26[2],ymm20[2]
20401 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm31, %zmm11, %zmm11
20402 ; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
20403 ; AVX512-FCP-NEXT: vmovdqa64 %zmm31, %zmm12 {%k1}
20404 ; AVX512-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm20 = ymm26[1],ymm20[1],ymm26[3],ymm20[3]
20405 ; AVX512-FCP-NEXT: vinserti64x4 $0, %ymm20, %zmm12, %zmm12
20406 ; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
20407 ; AVX512-FCP-NEXT: vmovdqa64 %zmm9, 3776(%rax)
20408 ; AVX512-FCP-NEXT: vmovdqa64 %zmm10, 3712(%rax)
20409 ; AVX512-FCP-NEXT: vmovdqa64 %zmm13, 3264(%rax)
20410 ; AVX512-FCP-NEXT: vmovdqa64 %zmm14, 3200(%rax)
20411 ; AVX512-FCP-NEXT: vmovdqa64 %zmm27, 2752(%rax)
20412 ; AVX512-FCP-NEXT: vmovdqa64 %zmm22, 2688(%rax)
20413 ; AVX512-FCP-NEXT: vmovdqa64 %zmm28, 2240(%rax)
20414 ; AVX512-FCP-NEXT: vmovdqa64 %zmm15, 2176(%rax)
20415 ; AVX512-FCP-NEXT: vmovdqa64 %zmm19, 1728(%rax)
20416 ; AVX512-FCP-NEXT: vmovdqa64 %zmm21, 1664(%rax)
20417 ; AVX512-FCP-NEXT: vmovdqa64 %zmm25, 1216(%rax)
20418 ; AVX512-FCP-NEXT: vmovdqa64 %zmm23, 1152(%rax)
20419 ; AVX512-FCP-NEXT: vmovdqa64 %zmm8, 704(%rax)
20420 ; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
20421 ; AVX512-FCP-NEXT: vmovaps %zmm8, 640(%rax)
20422 ; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
20423 ; AVX512-FCP-NEXT: vmovaps %zmm8, 192(%rax)
20424 ; AVX512-FCP-NEXT: vmovups (%rsp), %zmm8 # 64-byte Reload
20425 ; AVX512-FCP-NEXT: vmovaps %zmm8, 128(%rax)
20426 ; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
20427 ; AVX512-FCP-NEXT: vmovaps %zmm8, 4032(%rax)
20428 ; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
20429 ; AVX512-FCP-NEXT: vmovaps %zmm8, 3968(%rax)
20430 ; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
20431 ; AVX512-FCP-NEXT: vmovaps %zmm8, 3904(%rax)
20432 ; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
20433 ; AVX512-FCP-NEXT: vmovaps %zmm8, 3840(%rax)
20434 ; AVX512-FCP-NEXT: vmovdqa64 %zmm12, 3648(%rax)
20435 ; AVX512-FCP-NEXT: vmovdqa64 %zmm11, 3584(%rax)
20436 ; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
20437 ; AVX512-FCP-NEXT: vmovaps %zmm8, 3520(%rax)
20438 ; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
20439 ; AVX512-FCP-NEXT: vmovaps %zmm8, 3456(%rax)
20440 ; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
20441 ; AVX512-FCP-NEXT: vmovaps %zmm8, 3392(%rax)
20442 ; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
20443 ; AVX512-FCP-NEXT: vmovaps %zmm8, 3328(%rax)
20444 ; AVX512-FCP-NEXT: vmovdqa64 %zmm17, 3136(%rax)
20445 ; AVX512-FCP-NEXT: vmovdqa64 %zmm24, 3072(%rax)
20446 ; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
20447 ; AVX512-FCP-NEXT: vmovaps %zmm8, 3008(%rax)
20448 ; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
20449 ; AVX512-FCP-NEXT: vmovaps %zmm8, 2944(%rax)
20450 ; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
20451 ; AVX512-FCP-NEXT: vmovaps %zmm8, 2880(%rax)
20452 ; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
20453 ; AVX512-FCP-NEXT: vmovaps %zmm8, 2816(%rax)
20454 ; AVX512-FCP-NEXT: vmovdqa64 %zmm16, 2624(%rax)
20455 ; AVX512-FCP-NEXT: vmovdqa64 %zmm18, 2560(%rax)
20456 ; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
20457 ; AVX512-FCP-NEXT: vmovaps %zmm8, 2496(%rax)
20458 ; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
20459 ; AVX512-FCP-NEXT: vmovaps %zmm8, 2432(%rax)
20460 ; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
20461 ; AVX512-FCP-NEXT: vmovaps %zmm8, 2368(%rax)
20462 ; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
20463 ; AVX512-FCP-NEXT: vmovaps %zmm8, 2304(%rax)
20464 ; AVX512-FCP-NEXT: vmovdqa64 %zmm30, 2112(%rax)
20465 ; AVX512-FCP-NEXT: vmovdqa64 %zmm29, 2048(%rax)
20466 ; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
20467 ; AVX512-FCP-NEXT: vmovaps %zmm8, 1984(%rax)
20468 ; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
20469 ; AVX512-FCP-NEXT: vmovaps %zmm8, 1920(%rax)
20470 ; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
20471 ; AVX512-FCP-NEXT: vmovaps %zmm8, 1856(%rax)
20472 ; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
20473 ; AVX512-FCP-NEXT: vmovaps %zmm8, 1792(%rax)
20474 ; AVX512-FCP-NEXT: vmovdqa64 %zmm1, 1600(%rax)
20475 ; AVX512-FCP-NEXT: vmovdqa64 %zmm7, 1536(%rax)
20476 ; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
20477 ; AVX512-FCP-NEXT: vmovaps %zmm1, 1472(%rax)
20478 ; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
20479 ; AVX512-FCP-NEXT: vmovaps %zmm1, 1408(%rax)
20480 ; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
20481 ; AVX512-FCP-NEXT: vmovaps %zmm1, 1344(%rax)
20482 ; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
20483 ; AVX512-FCP-NEXT: vmovaps %zmm1, 1280(%rax)
20484 ; AVX512-FCP-NEXT: vmovdqa64 %zmm6, 1088(%rax)
20485 ; AVX512-FCP-NEXT: vmovdqa64 %zmm4, 1024(%rax)
20486 ; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
20487 ; AVX512-FCP-NEXT: vmovaps %zmm1, 960(%rax)
20488 ; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
20489 ; AVX512-FCP-NEXT: vmovaps %zmm1, 896(%rax)
20490 ; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
20491 ; AVX512-FCP-NEXT: vmovaps %zmm1, 832(%rax)
20492 ; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
20493 ; AVX512-FCP-NEXT: vmovaps %zmm1, 768(%rax)
20494 ; AVX512-FCP-NEXT: vmovdqa64 %zmm5, 576(%rax)
20495 ; AVX512-FCP-NEXT: vmovdqa64 %zmm3, 512(%rax)
20496 ; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
20497 ; AVX512-FCP-NEXT: vmovaps %zmm1, 448(%rax)
20498 ; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
20499 ; AVX512-FCP-NEXT: vmovaps %zmm1, 384(%rax)
20500 ; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
20501 ; AVX512-FCP-NEXT: vmovaps %zmm1, 320(%rax)
20502 ; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
20503 ; AVX512-FCP-NEXT: vmovaps %zmm1, 256(%rax)
20504 ; AVX512-FCP-NEXT: vmovdqa64 %zmm2, 64(%rax)
20505 ; AVX512-FCP-NEXT: vmovdqa64 %zmm0, (%rax)
20506 ; AVX512-FCP-NEXT: addq $5384, %rsp # imm = 0x1508
20507 ; AVX512-FCP-NEXT: vzeroupper
20508 ; AVX512-FCP-NEXT: retq
20509 ;
20510 ; AVX512DQ-LABEL: store_i64_stride8_vf64:
20511 ; AVX512DQ: # %bb.0:
20512 ; AVX512DQ-NEXT: subq $5384, %rsp # imm = 0x1508
20513 ; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %rax
20514 ; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %r10
20515 ; AVX512DQ-NEXT: vmovdqa64 128(%rdi), %zmm0
20516 ; AVX512DQ-NEXT: vmovdqa64 64(%rdi), %zmm4
20517 ; AVX512DQ-NEXT: vmovdqa64 (%rdi), %zmm14
20518 ; AVX512DQ-NEXT: vmovdqa64 128(%rsi), %zmm1
20519 ; AVX512DQ-NEXT: vmovdqa64 64(%rsi), %zmm5
20520 ; AVX512DQ-NEXT: vmovdqa64 (%rsi), %zmm10
20521 ; AVX512DQ-NEXT: vmovdqa64 128(%rdx), %zmm2
20522 ; AVX512DQ-NEXT: vmovdqa64 64(%rdx), %zmm7
20523 ; AVX512DQ-NEXT: vmovdqa64 (%rdx), %zmm12
20524 ; AVX512DQ-NEXT: vmovdqa64 128(%rcx), %zmm3
20525 ; AVX512DQ-NEXT: vmovdqa64 64(%rcx), %zmm9
20526 ; AVX512DQ-NEXT: vmovdqa64 (%rcx), %zmm15
20527 ; AVX512DQ-NEXT: vmovdqa64 (%r8), %zmm30
20528 ; AVX512DQ-NEXT: vmovdqa64 64(%r8), %zmm18
20529 ; AVX512DQ-NEXT: vmovdqa64 128(%r8), %zmm11
20530 ; AVX512DQ-NEXT: vmovdqa64 (%r9), %zmm24
20531 ; AVX512DQ-NEXT: vmovdqa64 64(%r9), %zmm28
20532 ; AVX512DQ-NEXT: vmovdqa64 128(%r9), %zmm22
20533 ; AVX512DQ-NEXT: vmovdqa64 (%r10), %zmm26
20534 ; AVX512DQ-NEXT: vmovdqa64 64(%r10), %zmm31
20535 ; AVX512DQ-NEXT: vmovdqa64 128(%r10), %zmm16
20536 ; AVX512DQ-NEXT: vmovdqa64 (%rax), %zmm17
20537 ; AVX512DQ-NEXT: vmovdqa64 64(%rax), %zmm27
20538 ; AVX512DQ-NEXT: vmovdqa64 128(%rax), %zmm13
20539 ; AVX512DQ-NEXT: movb $-64, %r11b
20540 ; AVX512DQ-NEXT: kmovw %r11d, %k1
20541 ; AVX512DQ-NEXT: vbroadcasti32x4 {{.*#+}} zmm19 = [4,12,4,12,4,12,4,12]
20542 ; AVX512DQ-NEXT: # zmm19 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
20543 ; AVX512DQ-NEXT: vmovdqa64 %zmm12, %zmm8
20544 ; AVX512DQ-NEXT: vpermt2q %zmm15, %zmm19, %zmm8
20545 ; AVX512DQ-NEXT: vmovdqa64 %zmm14, %zmm6
20546 ; AVX512DQ-NEXT: vpermt2q %zmm10, %zmm19, %zmm6
20547 ; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm8[4,5,6,7]
20548 ; AVX512DQ-NEXT: vmovdqa64 %zmm26, %zmm8
20549 ; AVX512DQ-NEXT: vpermt2q %zmm17, %zmm19, %zmm8
20550 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} zmm20 = zmm30[0],zmm24[0],zmm30[2],zmm24[2],zmm30[4],zmm24[4],zmm30[6],zmm24[6]
20551 ; AVX512DQ-NEXT: vmovdqa64 %zmm8, %zmm20 {%k1}
20552 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm6, %zmm20, %zmm6
20553 ; AVX512DQ-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20554 ; AVX512DQ-NEXT: vbroadcasti32x4 {{.*#+}} zmm21 = [5,13,5,13,5,13,5,13]
20555 ; AVX512DQ-NEXT: # zmm21 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
20556 ; AVX512DQ-NEXT: vmovdqa64 %zmm12, %zmm6
20557 ; AVX512DQ-NEXT: vpermt2q %zmm15, %zmm21, %zmm6
20558 ; AVX512DQ-NEXT: vmovdqa64 %zmm14, %zmm8
20559 ; AVX512DQ-NEXT: vpermt2q %zmm10, %zmm21, %zmm8
20560 ; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm6 = ymm8[0,1,2,3],ymm6[4,5,6,7]
20561 ; AVX512DQ-NEXT: vmovdqa64 %zmm26, %zmm8
20562 ; AVX512DQ-NEXT: vpermt2q %zmm17, %zmm21, %zmm8
20563 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} zmm20 = zmm30[1],zmm24[1],zmm30[3],zmm24[3],zmm30[5],zmm24[5],zmm30[7],zmm24[7]
20564 ; AVX512DQ-NEXT: vmovdqa64 %zmm8, %zmm20 {%k1}
20565 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm6, %zmm20, %zmm6
20566 ; AVX512DQ-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20567 ; AVX512DQ-NEXT: vbroadcasti32x4 {{.*#+}} zmm23 = [6,14,6,14,6,14,6,14]
20568 ; AVX512DQ-NEXT: # zmm23 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
20569 ; AVX512DQ-NEXT: vmovdqa64 %zmm12, %zmm6
20570 ; AVX512DQ-NEXT: vpermt2q %zmm15, %zmm23, %zmm6
20571 ; AVX512DQ-NEXT: vmovdqa64 %zmm14, %zmm8
20572 ; AVX512DQ-NEXT: vpermt2q %zmm10, %zmm23, %zmm8
20573 ; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm6 = ymm8[0,1,2,3],ymm6[4,5,6,7]
20574 ; AVX512DQ-NEXT: vmovdqa64 %zmm30, %zmm8
20575 ; AVX512DQ-NEXT: vpermt2q %zmm24, %zmm23, %zmm8
20576 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} zmm8 {%k1} = zmm26[0],zmm17[0],zmm26[2],zmm17[2],zmm26[4],zmm17[4],zmm26[6],zmm17[6]
20577 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm6, %zmm8, %zmm6
20578 ; AVX512DQ-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20579 ; AVX512DQ-NEXT: vbroadcasti32x4 {{.*#+}} zmm25 = [7,15,7,15,7,15,7,15]
20580 ; AVX512DQ-NEXT: # zmm25 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
20581 ; AVX512DQ-NEXT: vpermt2q %zmm15, %zmm25, %zmm12
20582 ; AVX512DQ-NEXT: vpermt2q %zmm10, %zmm25, %zmm14
20583 ; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm6 = ymm14[0,1,2,3],ymm12[4,5,6,7]
20584 ; AVX512DQ-NEXT: vmovdqa64 %zmm30, %zmm8
20585 ; AVX512DQ-NEXT: vpermt2q %zmm24, %zmm25, %zmm8
20586 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} zmm8 {%k1} = zmm26[1],zmm17[1],zmm26[3],zmm17[3],zmm26[5],zmm17[5],zmm26[7],zmm17[7]
20587 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm6, %zmm8, %zmm6
20588 ; AVX512DQ-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20589 ; AVX512DQ-NEXT: vmovdqa64 %zmm7, %zmm6
20590 ; AVX512DQ-NEXT: vpermt2q %zmm9, %zmm19, %zmm6
20591 ; AVX512DQ-NEXT: vmovdqa64 %zmm4, %zmm8
20592 ; AVX512DQ-NEXT: vpermt2q %zmm5, %zmm19, %zmm8
20593 ; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm6 = ymm8[0,1,2,3],ymm6[4,5,6,7]
20594 ; AVX512DQ-NEXT: vmovdqa64 %zmm31, %zmm8
20595 ; AVX512DQ-NEXT: vpermt2q %zmm27, %zmm19, %zmm8
20596 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} zmm10 = zmm18[0],zmm28[0],zmm18[2],zmm28[2],zmm18[4],zmm28[4],zmm18[6],zmm28[6]
20597 ; AVX512DQ-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}
20598 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm6, %zmm10, %zmm6
20599 ; AVX512DQ-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20600 ; AVX512DQ-NEXT: vmovdqa64 %zmm7, %zmm6
20601 ; AVX512DQ-NEXT: vpermt2q %zmm9, %zmm21, %zmm6
20602 ; AVX512DQ-NEXT: vmovdqa64 %zmm4, %zmm8
20603 ; AVX512DQ-NEXT: vpermt2q %zmm5, %zmm21, %zmm8
20604 ; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm6 = ymm8[0,1,2,3],ymm6[4,5,6,7]
20605 ; AVX512DQ-NEXT: vmovdqa64 %zmm31, %zmm8
20606 ; AVX512DQ-NEXT: vpermt2q %zmm27, %zmm21, %zmm8
20607 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} zmm10 = zmm18[1],zmm28[1],zmm18[3],zmm28[3],zmm18[5],zmm28[5],zmm18[7],zmm28[7]
20608 ; AVX512DQ-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}
20609 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm6, %zmm10, %zmm6
20610 ; AVX512DQ-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20611 ; AVX512DQ-NEXT: vmovdqa64 %zmm7, %zmm6
20612 ; AVX512DQ-NEXT: vpermt2q %zmm9, %zmm23, %zmm6
20613 ; AVX512DQ-NEXT: vmovdqa64 %zmm4, %zmm8
20614 ; AVX512DQ-NEXT: vpermt2q %zmm5, %zmm23, %zmm8
20615 ; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm6 = ymm8[0,1,2,3],ymm6[4,5,6,7]
20616 ; AVX512DQ-NEXT: vmovdqu64 %zmm18, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20617 ; AVX512DQ-NEXT: vmovdqa64 %zmm18, %zmm8
20618 ; AVX512DQ-NEXT: vpermt2q %zmm28, %zmm23, %zmm8
20619 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} zmm8 {%k1} = zmm31[0],zmm27[0],zmm31[2],zmm27[2],zmm31[4],zmm27[4],zmm31[6],zmm27[6]
20620 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm6, %zmm8, %zmm6
20621 ; AVX512DQ-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20622 ; AVX512DQ-NEXT: vpermt2q %zmm9, %zmm25, %zmm7
20623 ; AVX512DQ-NEXT: vpermt2q %zmm5, %zmm25, %zmm4
20624 ; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm7[4,5,6,7]
20625 ; AVX512DQ-NEXT: vmovdqa64 %zmm18, %zmm5
20626 ; AVX512DQ-NEXT: vpermt2q %zmm28, %zmm25, %zmm5
20627 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} zmm5 {%k1} = zmm31[1],zmm27[1],zmm31[3],zmm27[3],zmm31[5],zmm27[5],zmm31[7],zmm27[7]
20628 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm4, %zmm5, %zmm4
20629 ; AVX512DQ-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20630 ; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm4
20631 ; AVX512DQ-NEXT: vpermt2q %zmm3, %zmm19, %zmm4
20632 ; AVX512DQ-NEXT: vmovdqa64 %zmm0, %zmm5
20633 ; AVX512DQ-NEXT: vpermt2q %zmm1, %zmm19, %zmm5
20634 ; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm4 = ymm5[0,1,2,3],ymm4[4,5,6,7]
20635 ; AVX512DQ-NEXT: vmovdqa64 %zmm16, %zmm5
20636 ; AVX512DQ-NEXT: vpermt2q %zmm13, %zmm19, %zmm5
20637 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} zmm6 = zmm11[0],zmm22[0],zmm11[2],zmm22[2],zmm11[4],zmm22[4],zmm11[6],zmm22[6]
20638 ; AVX512DQ-NEXT: vmovdqa64 %zmm5, %zmm6 {%k1}
20639 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm4, %zmm6, %zmm4
20640 ; AVX512DQ-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20641 ; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm4
20642 ; AVX512DQ-NEXT: vpermt2q %zmm3, %zmm21, %zmm4
20643 ; AVX512DQ-NEXT: vmovdqa64 %zmm0, %zmm5
20644 ; AVX512DQ-NEXT: vpermt2q %zmm1, %zmm21, %zmm5
20645 ; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm4 = ymm5[0,1,2,3],ymm4[4,5,6,7]
20646 ; AVX512DQ-NEXT: vmovdqa64 %zmm16, %zmm5
20647 ; AVX512DQ-NEXT: vpermt2q %zmm13, %zmm21, %zmm5
20648 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} zmm6 = zmm11[1],zmm22[1],zmm11[3],zmm22[3],zmm11[5],zmm22[5],zmm11[7],zmm22[7]
20649 ; AVX512DQ-NEXT: vmovdqa64 %zmm5, %zmm6 {%k1}
20650 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm4, %zmm6, %zmm4
20651 ; AVX512DQ-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20652 ; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm4
20653 ; AVX512DQ-NEXT: vpermt2q %zmm3, %zmm23, %zmm4
20654 ; AVX512DQ-NEXT: vmovdqa64 %zmm0, %zmm5
20655 ; AVX512DQ-NEXT: vpermt2q %zmm1, %zmm23, %zmm5
20656 ; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm4 = ymm5[0,1,2,3],ymm4[4,5,6,7]
20657 ; AVX512DQ-NEXT: vmovdqa64 %zmm11, %zmm5
20658 ; AVX512DQ-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20659 ; AVX512DQ-NEXT: vmovdqu64 %zmm22, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20660 ; AVX512DQ-NEXT: vpermt2q %zmm22, %zmm23, %zmm5
20661 ; AVX512DQ-NEXT: vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20662 ; AVX512DQ-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20663 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} zmm5 {%k1} = zmm16[0],zmm13[0],zmm16[2],zmm13[2],zmm16[4],zmm13[4],zmm16[6],zmm13[6]
20664 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm4, %zmm5, %zmm4
20665 ; AVX512DQ-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20666 ; AVX512DQ-NEXT: vmovdqa64 192(%rdi), %zmm4
20667 ; AVX512DQ-NEXT: vpermt2q %zmm3, %zmm25, %zmm2
20668 ; AVX512DQ-NEXT: vmovdqa64 192(%rdx), %zmm3
20669 ; AVX512DQ-NEXT: vpermt2q %zmm1, %zmm25, %zmm0
20670 ; AVX512DQ-NEXT: vmovdqa64 192(%rcx), %zmm1
20671 ; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
20672 ; AVX512DQ-NEXT: vmovdqa64 %zmm11, %zmm2
20673 ; AVX512DQ-NEXT: vpermt2q %zmm22, %zmm25, %zmm2
20674 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} zmm2 {%k1} = zmm16[1],zmm13[1],zmm16[3],zmm13[3],zmm16[5],zmm13[5],zmm16[7],zmm13[7]
20675 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm0, %zmm2, %zmm0
20676 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20677 ; AVX512DQ-NEXT: vmovdqa64 %zmm3, %zmm0
20678 ; AVX512DQ-NEXT: vpermt2q %zmm1, %zmm19, %zmm0
20679 ; AVX512DQ-NEXT: vmovdqa64 192(%rsi), %zmm2
20680 ; AVX512DQ-NEXT: vmovdqa64 %zmm4, %zmm5
20681 ; AVX512DQ-NEXT: vpermt2q %zmm2, %zmm19, %zmm5
20682 ; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm0[4,5,6,7]
20683 ; AVX512DQ-NEXT: vmovdqa64 192(%r10), %zmm9
20684 ; AVX512DQ-NEXT: vmovdqa64 192(%rax), %zmm20
20685 ; AVX512DQ-NEXT: vmovdqa64 %zmm9, %zmm6
20686 ; AVX512DQ-NEXT: vpermt2q %zmm20, %zmm19, %zmm6
20687 ; AVX512DQ-NEXT: vmovdqa64 192(%r8), %zmm8
20688 ; AVX512DQ-NEXT: vmovdqa64 192(%r9), %zmm0
20689 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} zmm7 = zmm8[0],zmm0[0],zmm8[2],zmm0[2],zmm8[4],zmm0[4],zmm8[6],zmm0[6]
20690 ; AVX512DQ-NEXT: vmovdqa64 %zmm6, %zmm7 {%k1}
20691 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm5, %zmm7, %zmm5
20692 ; AVX512DQ-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20693 ; AVX512DQ-NEXT: vmovdqa64 %zmm3, %zmm5
20694 ; AVX512DQ-NEXT: vpermt2q %zmm1, %zmm21, %zmm5
20695 ; AVX512DQ-NEXT: vmovdqa64 %zmm4, %zmm6
20696 ; AVX512DQ-NEXT: vpermt2q %zmm2, %zmm21, %zmm6
20697 ; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm5 = ymm6[0,1,2,3],ymm5[4,5,6,7]
20698 ; AVX512DQ-NEXT: vmovdqa64 %zmm9, %zmm6
20699 ; AVX512DQ-NEXT: vpermt2q %zmm20, %zmm21, %zmm6
20700 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} zmm7 = zmm8[1],zmm0[1],zmm8[3],zmm0[3],zmm8[5],zmm0[5],zmm8[7],zmm0[7]
20701 ; AVX512DQ-NEXT: vmovdqa64 %zmm6, %zmm7 {%k1}
20702 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm5, %zmm7, %zmm5
20703 ; AVX512DQ-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20704 ; AVX512DQ-NEXT: vmovdqa64 %zmm3, %zmm5
20705 ; AVX512DQ-NEXT: vpermt2q %zmm1, %zmm23, %zmm5
20706 ; AVX512DQ-NEXT: vmovdqa64 %zmm4, %zmm6
20707 ; AVX512DQ-NEXT: vpermt2q %zmm2, %zmm23, %zmm6
20708 ; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm5 = ymm6[0,1,2,3],ymm5[4,5,6,7]
20709 ; AVX512DQ-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20710 ; AVX512DQ-NEXT: vmovdqa64 %zmm8, %zmm6
20711 ; AVX512DQ-NEXT: vpermt2q %zmm0, %zmm23, %zmm6
20712 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} zmm6 {%k1} = zmm9[0],zmm20[0],zmm9[2],zmm20[2],zmm9[4],zmm20[4],zmm9[6],zmm20[6]
20713 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm5, %zmm6, %zmm5
20714 ; AVX512DQ-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20715 ; AVX512DQ-NEXT: vpermt2q %zmm1, %zmm25, %zmm3
20716 ; AVX512DQ-NEXT: vpermt2q %zmm2, %zmm25, %zmm4
20717 ; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm1 = ymm4[0,1,2,3],ymm3[4,5,6,7]
20718 ; AVX512DQ-NEXT: vmovdqa64 %zmm8, %zmm2
20719 ; AVX512DQ-NEXT: vpermt2q %zmm0, %zmm25, %zmm2
20720 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} zmm2 {%k1} = zmm9[1],zmm20[1],zmm9[3],zmm20[3],zmm9[5],zmm20[5],zmm9[7],zmm20[7]
20721 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm1, %zmm2, %zmm1
20722 ; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20723 ; AVX512DQ-NEXT: vmovdqa64 256(%rdx), %zmm1
20724 ; AVX512DQ-NEXT: vmovdqa64 256(%rcx), %zmm3
20725 ; AVX512DQ-NEXT: vmovdqa64 %zmm1, %zmm2
20726 ; AVX512DQ-NEXT: vpermt2q %zmm3, %zmm19, %zmm2
20727 ; AVX512DQ-NEXT: vmovdqa64 256(%rdi), %zmm4
20728 ; AVX512DQ-NEXT: vmovdqa64 256(%rsi), %zmm5
20729 ; AVX512DQ-NEXT: vmovdqa64 %zmm4, %zmm6
20730 ; AVX512DQ-NEXT: vpermt2q %zmm5, %zmm19, %zmm6
20731 ; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm2[4,5,6,7]
20732 ; AVX512DQ-NEXT: vmovdqa64 256(%r10), %zmm13
20733 ; AVX512DQ-NEXT: vmovdqa64 256(%rax), %zmm2
20734 ; AVX512DQ-NEXT: vmovdqa64 %zmm13, %zmm7
20735 ; AVX512DQ-NEXT: vpermt2q %zmm2, %zmm19, %zmm7
20736 ; AVX512DQ-NEXT: vmovdqa64 256(%r8), %zmm18
20737 ; AVX512DQ-NEXT: vmovdqa64 256(%r9), %zmm15
20738 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} zmm8 = zmm18[0],zmm15[0],zmm18[2],zmm15[2],zmm18[4],zmm15[4],zmm18[6],zmm15[6]
20739 ; AVX512DQ-NEXT: vmovdqa64 %zmm7, %zmm8 {%k1}
20740 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm6, %zmm8, %zmm6
20741 ; AVX512DQ-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20742 ; AVX512DQ-NEXT: vmovdqa64 %zmm1, %zmm6
20743 ; AVX512DQ-NEXT: vpermt2q %zmm3, %zmm21, %zmm6
20744 ; AVX512DQ-NEXT: vmovdqa64 %zmm4, %zmm7
20745 ; AVX512DQ-NEXT: vpermt2q %zmm5, %zmm21, %zmm7
20746 ; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]
20747 ; AVX512DQ-NEXT: vmovdqa64 %zmm13, %zmm7
20748 ; AVX512DQ-NEXT: vpermt2q %zmm2, %zmm21, %zmm7
20749 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} zmm8 = zmm18[1],zmm15[1],zmm18[3],zmm15[3],zmm18[5],zmm15[5],zmm18[7],zmm15[7]
20750 ; AVX512DQ-NEXT: vmovdqa64 %zmm7, %zmm8 {%k1}
20751 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm6, %zmm8, %zmm6
20752 ; AVX512DQ-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20753 ; AVX512DQ-NEXT: vmovdqa64 %zmm1, %zmm6
20754 ; AVX512DQ-NEXT: vpermt2q %zmm3, %zmm23, %zmm6
20755 ; AVX512DQ-NEXT: vmovdqa64 %zmm4, %zmm7
20756 ; AVX512DQ-NEXT: vpermt2q %zmm5, %zmm23, %zmm7
20757 ; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]
20758 ; AVX512DQ-NEXT: vmovdqa64 %zmm18, %zmm7
20759 ; AVX512DQ-NEXT: vpermt2q %zmm15, %zmm23, %zmm7
20760 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} zmm7 {%k1} = zmm13[0],zmm2[0],zmm13[2],zmm2[2],zmm13[4],zmm2[4],zmm13[6],zmm2[6]
20761 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm6, %zmm7, %zmm6
20762 ; AVX512DQ-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20763 ; AVX512DQ-NEXT: vpermt2q %zmm3, %zmm25, %zmm1
20764 ; AVX512DQ-NEXT: vpermt2q %zmm5, %zmm25, %zmm4
20765 ; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm1 = ymm4[0,1,2,3],ymm1[4,5,6,7]
20766 ; AVX512DQ-NEXT: vmovdqa64 %zmm18, %zmm3
20767 ; AVX512DQ-NEXT: vpermt2q %zmm15, %zmm25, %zmm3
20768 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} zmm3 {%k1} = zmm13[1],zmm2[1],zmm13[3],zmm2[3],zmm13[5],zmm2[5],zmm13[7],zmm2[7]
20769 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm1, %zmm3, %zmm1
20770 ; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20771 ; AVX512DQ-NEXT: vmovdqa64 320(%rdx), %zmm1
20772 ; AVX512DQ-NEXT: vmovdqa64 320(%rcx), %zmm4
20773 ; AVX512DQ-NEXT: vmovdqa64 %zmm1, %zmm3
20774 ; AVX512DQ-NEXT: vpermt2q %zmm4, %zmm19, %zmm3
20775 ; AVX512DQ-NEXT: vmovdqa64 320(%rdi), %zmm6
20776 ; AVX512DQ-NEXT: vmovdqa64 320(%rsi), %zmm7
20777 ; AVX512DQ-NEXT: vmovdqa64 %zmm6, %zmm5
20778 ; AVX512DQ-NEXT: vpermt2q %zmm7, %zmm19, %zmm5
20779 ; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm8 = ymm5[0,1,2,3],ymm3[4,5,6,7]
20780 ; AVX512DQ-NEXT: vmovdqa64 320(%r10), %zmm16
20781 ; AVX512DQ-NEXT: vmovdqa64 320(%rax), %zmm22
20782 ; AVX512DQ-NEXT: vmovdqa64 %zmm16, %zmm12
20783 ; AVX512DQ-NEXT: vpermt2q %zmm22, %zmm19, %zmm12
20784 ; AVX512DQ-NEXT: vmovdqa64 320(%r8), %zmm29
20785 ; AVX512DQ-NEXT: vmovdqa64 320(%r9), %zmm5
20786 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} zmm14 = zmm29[0],zmm5[0],zmm29[2],zmm5[2],zmm29[4],zmm5[4],zmm29[6],zmm5[6]
20787 ; AVX512DQ-NEXT: vmovdqa64 %zmm12, %zmm14 {%k1}
20788 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm8, %zmm14, %zmm3
20789 ; AVX512DQ-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20790 ; AVX512DQ-NEXT: vmovdqa64 %zmm1, %zmm8
20791 ; AVX512DQ-NEXT: vpermt2q %zmm4, %zmm21, %zmm8
20792 ; AVX512DQ-NEXT: vmovdqa64 %zmm6, %zmm12
20793 ; AVX512DQ-NEXT: vpermt2q %zmm7, %zmm21, %zmm12
20794 ; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm8 = ymm12[0,1,2,3],ymm8[4,5,6,7]
20795 ; AVX512DQ-NEXT: vmovdqa64 %zmm16, %zmm12
20796 ; AVX512DQ-NEXT: vpermt2q %zmm22, %zmm21, %zmm12
20797 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} zmm14 = zmm29[1],zmm5[1],zmm29[3],zmm5[3],zmm29[5],zmm5[5],zmm29[7],zmm5[7]
20798 ; AVX512DQ-NEXT: vmovdqa64 %zmm12, %zmm14 {%k1}
20799 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm8, %zmm14, %zmm3
20800 ; AVX512DQ-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20801 ; AVX512DQ-NEXT: vmovdqa64 %zmm1, %zmm8
20802 ; AVX512DQ-NEXT: vpermt2q %zmm4, %zmm23, %zmm8
20803 ; AVX512DQ-NEXT: vmovdqa64 %zmm6, %zmm12
20804 ; AVX512DQ-NEXT: vpermt2q %zmm7, %zmm23, %zmm12
20805 ; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm8 = ymm12[0,1,2,3],ymm8[4,5,6,7]
20806 ; AVX512DQ-NEXT: vmovdqa64 %zmm29, %zmm12
20807 ; AVX512DQ-NEXT: vpermt2q %zmm5, %zmm23, %zmm12
20808 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} zmm12 {%k1} = zmm16[0],zmm22[0],zmm16[2],zmm22[2],zmm16[4],zmm22[4],zmm16[6],zmm22[6]
20809 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm8, %zmm12, %zmm3
20810 ; AVX512DQ-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20811 ; AVX512DQ-NEXT: vpermt2q %zmm4, %zmm25, %zmm1
20812 ; AVX512DQ-NEXT: vpermt2q %zmm7, %zmm25, %zmm6
20813 ; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm1 = ymm6[0,1,2,3],ymm1[4,5,6,7]
20814 ; AVX512DQ-NEXT: vmovdqa64 %zmm29, %zmm4
20815 ; AVX512DQ-NEXT: vpermt2q %zmm5, %zmm25, %zmm4
20816 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} zmm4 {%k1} = zmm16[1],zmm22[1],zmm16[3],zmm22[3],zmm16[5],zmm22[5],zmm16[7],zmm22[7]
20817 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm1, %zmm4, %zmm1
20818 ; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20819 ; AVX512DQ-NEXT: vmovdqa64 384(%rdx), %zmm3
20820 ; AVX512DQ-NEXT: vmovdqa64 384(%rcx), %zmm1
20821 ; AVX512DQ-NEXT: vmovdqa64 %zmm3, %zmm4
20822 ; AVX512DQ-NEXT: vpermt2q %zmm1, %zmm19, %zmm4
20823 ; AVX512DQ-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20824 ; AVX512DQ-NEXT: vmovdqa64 %zmm3, %zmm4
20825 ; AVX512DQ-NEXT: vpermt2q %zmm1, %zmm21, %zmm4
20826 ; AVX512DQ-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20827 ; AVX512DQ-NEXT: vmovdqa64 %zmm3, %zmm4
20828 ; AVX512DQ-NEXT: vpermt2q %zmm1, %zmm23, %zmm4
20829 ; AVX512DQ-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20830 ; AVX512DQ-NEXT: vpermt2q %zmm1, %zmm25, %zmm3
20831 ; AVX512DQ-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20832 ; AVX512DQ-NEXT: vmovdqa64 384(%rdi), %zmm3
20833 ; AVX512DQ-NEXT: vmovdqa64 384(%rsi), %zmm1
20834 ; AVX512DQ-NEXT: vmovdqa64 %zmm3, %zmm4
20835 ; AVX512DQ-NEXT: vpermt2q %zmm1, %zmm19, %zmm4
20836 ; AVX512DQ-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20837 ; AVX512DQ-NEXT: vmovdqa64 %zmm3, %zmm4
20838 ; AVX512DQ-NEXT: vpermt2q %zmm1, %zmm21, %zmm4
20839 ; AVX512DQ-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20840 ; AVX512DQ-NEXT: vmovdqa64 %zmm3, %zmm4
20841 ; AVX512DQ-NEXT: vpermt2q %zmm1, %zmm23, %zmm4
20842 ; AVX512DQ-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20843 ; AVX512DQ-NEXT: vpermt2q %zmm1, %zmm25, %zmm3
20844 ; AVX512DQ-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20845 ; AVX512DQ-NEXT: vmovdqa64 448(%rdx), %zmm3
20846 ; AVX512DQ-NEXT: vmovdqa64 448(%rcx), %zmm1
20847 ; AVX512DQ-NEXT: vmovdqa64 %zmm3, %zmm4
20848 ; AVX512DQ-NEXT: vpermt2q %zmm1, %zmm19, %zmm4
20849 ; AVX512DQ-NEXT: vmovdqu64 %zmm4, (%rsp) # 64-byte Spill
20850 ; AVX512DQ-NEXT: vmovdqa64 %zmm3, %zmm4
20851 ; AVX512DQ-NEXT: vpermt2q %zmm1, %zmm21, %zmm4
20852 ; AVX512DQ-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20853 ; AVX512DQ-NEXT: vmovdqa64 %zmm3, %zmm4
20854 ; AVX512DQ-NEXT: vpermt2q %zmm1, %zmm23, %zmm4
20855 ; AVX512DQ-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20856 ; AVX512DQ-NEXT: vpermt2q %zmm1, %zmm25, %zmm3
20857 ; AVX512DQ-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20858 ; AVX512DQ-NEXT: vmovdqa64 448(%rdi), %zmm3
20859 ; AVX512DQ-NEXT: vmovdqa64 448(%rsi), %zmm1
20860 ; AVX512DQ-NEXT: vmovdqa64 %zmm3, %zmm4
20861 ; AVX512DQ-NEXT: vpermt2q %zmm1, %zmm19, %zmm4
20862 ; AVX512DQ-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20863 ; AVX512DQ-NEXT: vmovdqa64 %zmm3, %zmm4
20864 ; AVX512DQ-NEXT: vpermt2q %zmm1, %zmm21, %zmm4
20865 ; AVX512DQ-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20866 ; AVX512DQ-NEXT: vmovdqa64 %zmm3, %zmm4
20867 ; AVX512DQ-NEXT: vpermt2q %zmm1, %zmm23, %zmm4
20868 ; AVX512DQ-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20869 ; AVX512DQ-NEXT: vpermt2q %zmm1, %zmm25, %zmm3
20870 ; AVX512DQ-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20871 ; AVX512DQ-NEXT: vbroadcasti32x4 {{.*#+}} zmm11 = [0,8,0,8,0,8,0,8]
20872 ; AVX512DQ-NEXT: # zmm11 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
20873 ; AVX512DQ-NEXT: vmovdqa64 %zmm26, %zmm1
20874 ; AVX512DQ-NEXT: vpermt2q %zmm17, %zmm11, %zmm1
20875 ; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20876 ; AVX512DQ-NEXT: vbroadcasti32x4 {{.*#+}} zmm12 = [1,9,1,9,1,9,1,9]
20877 ; AVX512DQ-NEXT: # zmm12 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
20878 ; AVX512DQ-NEXT: vmovdqa64 %zmm26, %zmm1
20879 ; AVX512DQ-NEXT: vpermt2q %zmm17, %zmm12, %zmm1
20880 ; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20881 ; AVX512DQ-NEXT: vbroadcasti32x4 {{.*#+}} zmm10 = [2,10,2,10,2,10,2,10]
20882 ; AVX512DQ-NEXT: # zmm10 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
20883 ; AVX512DQ-NEXT: vmovdqa64 %zmm26, %zmm1
20884 ; AVX512DQ-NEXT: vpermt2q %zmm17, %zmm10, %zmm1
20885 ; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20886 ; AVX512DQ-NEXT: vbroadcasti32x4 {{.*#+}} zmm4 = [3,11,3,11,3,11,3,11]
20887 ; AVX512DQ-NEXT: # zmm4 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
20888 ; AVX512DQ-NEXT: vpermt2q %zmm17, %zmm4, %zmm26
20889 ; AVX512DQ-NEXT: vmovdqu64 %zmm26, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20890 ; AVX512DQ-NEXT: vmovdqa64 %zmm30, %zmm1
20891 ; AVX512DQ-NEXT: vpermt2q %zmm24, %zmm11, %zmm1
20892 ; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20893 ; AVX512DQ-NEXT: vmovdqa64 %zmm30, %zmm1
20894 ; AVX512DQ-NEXT: vpermt2q %zmm24, %zmm12, %zmm1
20895 ; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20896 ; AVX512DQ-NEXT: vmovdqa64 %zmm30, %zmm1
20897 ; AVX512DQ-NEXT: vpermt2q %zmm24, %zmm10, %zmm1
20898 ; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20899 ; AVX512DQ-NEXT: vpermt2q %zmm24, %zmm4, %zmm30
20900 ; AVX512DQ-NEXT: vmovdqu64 %zmm30, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20901 ; AVX512DQ-NEXT: vmovdqa64 %zmm31, %zmm1
20902 ; AVX512DQ-NEXT: vpermt2q %zmm27, %zmm11, %zmm1
20903 ; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20904 ; AVX512DQ-NEXT: vmovdqa64 %zmm31, %zmm1
20905 ; AVX512DQ-NEXT: vpermt2q %zmm27, %zmm12, %zmm1
20906 ; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20907 ; AVX512DQ-NEXT: vmovdqa64 %zmm31, %zmm1
20908 ; AVX512DQ-NEXT: vpermt2q %zmm27, %zmm10, %zmm1
20909 ; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20910 ; AVX512DQ-NEXT: vpermt2q %zmm27, %zmm4, %zmm31
20911 ; AVX512DQ-NEXT: vmovdqu64 %zmm31, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20912 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
20913 ; AVX512DQ-NEXT: vmovdqa64 %zmm3, %zmm1
20914 ; AVX512DQ-NEXT: vpermt2q %zmm28, %zmm11, %zmm1
20915 ; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20916 ; AVX512DQ-NEXT: vmovdqa64 %zmm3, %zmm1
20917 ; AVX512DQ-NEXT: vpermt2q %zmm28, %zmm12, %zmm1
20918 ; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20919 ; AVX512DQ-NEXT: vmovdqa64 %zmm3, %zmm31
20920 ; AVX512DQ-NEXT: vpermt2q %zmm28, %zmm10, %zmm31
20921 ; AVX512DQ-NEXT: vpermt2q %zmm28, %zmm4, %zmm3
20922 ; AVX512DQ-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20923 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
20924 ; AVX512DQ-NEXT: vmovdqa64 %zmm28, %zmm3
20925 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
20926 ; AVX512DQ-NEXT: vpermt2q %zmm1, %zmm11, %zmm3
20927 ; AVX512DQ-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20928 ; AVX512DQ-NEXT: vmovdqa64 %zmm28, %zmm3
20929 ; AVX512DQ-NEXT: vpermt2q %zmm1, %zmm12, %zmm3
20930 ; AVX512DQ-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20931 ; AVX512DQ-NEXT: vmovdqa64 %zmm28, %zmm3
20932 ; AVX512DQ-NEXT: vpermt2q %zmm1, %zmm10, %zmm3
20933 ; AVX512DQ-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20934 ; AVX512DQ-NEXT: vpermt2q %zmm1, %zmm4, %zmm28
20935 ; AVX512DQ-NEXT: vmovdqu64 %zmm28, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20936 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
20937 ; AVX512DQ-NEXT: vmovdqa64 %zmm6, %zmm3
20938 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
20939 ; AVX512DQ-NEXT: vpermt2q %zmm1, %zmm11, %zmm3
20940 ; AVX512DQ-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20941 ; AVX512DQ-NEXT: vmovdqa64 %zmm6, %zmm3
20942 ; AVX512DQ-NEXT: vpermt2q %zmm1, %zmm12, %zmm3
20943 ; AVX512DQ-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20944 ; AVX512DQ-NEXT: vmovdqa64 %zmm6, %zmm27
20945 ; AVX512DQ-NEXT: vpermt2q %zmm1, %zmm10, %zmm27
20946 ; AVX512DQ-NEXT: vpermt2q %zmm1, %zmm4, %zmm6
20947 ; AVX512DQ-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20948 ; AVX512DQ-NEXT: vmovdqa64 %zmm9, %zmm1
20949 ; AVX512DQ-NEXT: vpermt2q %zmm20, %zmm11, %zmm1
20950 ; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20951 ; AVX512DQ-NEXT: vmovdqa64 %zmm9, %zmm1
20952 ; AVX512DQ-NEXT: vpermt2q %zmm20, %zmm12, %zmm1
20953 ; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20954 ; AVX512DQ-NEXT: vmovdqa64 %zmm9, %zmm1
20955 ; AVX512DQ-NEXT: vpermt2q %zmm20, %zmm10, %zmm1
20956 ; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20957 ; AVX512DQ-NEXT: vpermt2q %zmm20, %zmm4, %zmm9
20958 ; AVX512DQ-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20959 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
20960 ; AVX512DQ-NEXT: vmovdqa64 %zmm1, %zmm3
20961 ; AVX512DQ-NEXT: vpermt2q %zmm0, %zmm11, %zmm3
20962 ; AVX512DQ-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20963 ; AVX512DQ-NEXT: vmovdqa64 %zmm1, %zmm3
20964 ; AVX512DQ-NEXT: vpermt2q %zmm0, %zmm12, %zmm3
20965 ; AVX512DQ-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20966 ; AVX512DQ-NEXT: vmovdqa64 %zmm1, %zmm20
20967 ; AVX512DQ-NEXT: vpermt2q %zmm0, %zmm10, %zmm20
20968 ; AVX512DQ-NEXT: vpermt2q %zmm0, %zmm4, %zmm1
20969 ; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20970 ; AVX512DQ-NEXT: vmovdqa64 %zmm13, %zmm0
20971 ; AVX512DQ-NEXT: vpermt2q %zmm2, %zmm11, %zmm0
20972 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20973 ; AVX512DQ-NEXT: vmovdqa64 %zmm13, %zmm0
20974 ; AVX512DQ-NEXT: vpermt2q %zmm2, %zmm12, %zmm0
20975 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20976 ; AVX512DQ-NEXT: vmovdqa64 %zmm13, %zmm0
20977 ; AVX512DQ-NEXT: vpermt2q %zmm2, %zmm10, %zmm0
20978 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20979 ; AVX512DQ-NEXT: vpermt2q %zmm2, %zmm4, %zmm13
20980 ; AVX512DQ-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20981 ; AVX512DQ-NEXT: vmovdqa64 %zmm18, %zmm0
20982 ; AVX512DQ-NEXT: vpermt2q %zmm15, %zmm11, %zmm0
20983 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20984 ; AVX512DQ-NEXT: vmovdqa64 %zmm18, %zmm0
20985 ; AVX512DQ-NEXT: vpermt2q %zmm15, %zmm12, %zmm0
20986 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20987 ; AVX512DQ-NEXT: vmovdqa64 %zmm18, %zmm28
20988 ; AVX512DQ-NEXT: vpermt2q %zmm15, %zmm10, %zmm28
20989 ; AVX512DQ-NEXT: vpermt2q %zmm15, %zmm4, %zmm18
20990 ; AVX512DQ-NEXT: vmovdqu64 %zmm18, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20991 ; AVX512DQ-NEXT: vmovdqa64 %zmm16, %zmm0
20992 ; AVX512DQ-NEXT: vpermt2q %zmm22, %zmm11, %zmm0
20993 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20994 ; AVX512DQ-NEXT: vmovdqa64 %zmm16, %zmm0
20995 ; AVX512DQ-NEXT: vpermt2q %zmm22, %zmm12, %zmm0
20996 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
20997 ; AVX512DQ-NEXT: vmovdqa64 %zmm16, %zmm0
20998 ; AVX512DQ-NEXT: vpermt2q %zmm22, %zmm10, %zmm0
20999 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21000 ; AVX512DQ-NEXT: vpermt2q %zmm22, %zmm4, %zmm16
21001 ; AVX512DQ-NEXT: vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21002 ; AVX512DQ-NEXT: vmovdqa64 %zmm29, %zmm0
21003 ; AVX512DQ-NEXT: vpermt2q %zmm5, %zmm11, %zmm0
21004 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21005 ; AVX512DQ-NEXT: vmovdqa64 %zmm29, %zmm0
21006 ; AVX512DQ-NEXT: vpermt2q %zmm5, %zmm12, %zmm0
21007 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21008 ; AVX512DQ-NEXT: vmovdqa64 %zmm29, %zmm22
21009 ; AVX512DQ-NEXT: vpermt2q %zmm5, %zmm10, %zmm22
21010 ; AVX512DQ-NEXT: vpermt2q %zmm5, %zmm4, %zmm29
21011 ; AVX512DQ-NEXT: vmovdqa64 384(%r10), %zmm13
21012 ; AVX512DQ-NEXT: vmovdqa64 384(%rax), %zmm0
21013 ; AVX512DQ-NEXT: vmovdqa64 %zmm13, %zmm1
21014 ; AVX512DQ-NEXT: vpermt2q %zmm0, %zmm11, %zmm1
21015 ; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21016 ; AVX512DQ-NEXT: vmovdqa64 %zmm13, %zmm1
21017 ; AVX512DQ-NEXT: vpermt2q %zmm0, %zmm12, %zmm1
21018 ; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21019 ; AVX512DQ-NEXT: vmovdqa64 %zmm13, %zmm14
21020 ; AVX512DQ-NEXT: vpermt2q %zmm0, %zmm19, %zmm14
21021 ; AVX512DQ-NEXT: vmovdqa64 %zmm13, %zmm15
21022 ; AVX512DQ-NEXT: vpermt2q %zmm0, %zmm21, %zmm15
21023 ; AVX512DQ-NEXT: vmovdqa64 384(%r8), %zmm5
21024 ; AVX512DQ-NEXT: vmovdqa64 384(%r9), %zmm2
21025 ; AVX512DQ-NEXT: vmovdqa64 %zmm5, %zmm8
21026 ; AVX512DQ-NEXT: vpermt2q %zmm2, %zmm23, %zmm8
21027 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} zmm8 {%k1} = zmm13[0],zmm0[0],zmm13[2],zmm0[2],zmm13[4],zmm0[4],zmm13[6],zmm0[6]
21028 ; AVX512DQ-NEXT: vmovdqa64 %zmm5, %zmm6
21029 ; AVX512DQ-NEXT: vpermt2q %zmm2, %zmm25, %zmm6
21030 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} zmm6 {%k1} = zmm13[1],zmm0[1],zmm13[3],zmm0[3],zmm13[5],zmm0[5],zmm13[7],zmm0[7]
21031 ; AVX512DQ-NEXT: vmovdqa64 %zmm13, %zmm30
21032 ; AVX512DQ-NEXT: vpermt2q %zmm0, %zmm10, %zmm30
21033 ; AVX512DQ-NEXT: vpermt2q %zmm0, %zmm4, %zmm13
21034 ; AVX512DQ-NEXT: vmovdqa64 %zmm5, %zmm24
21035 ; AVX512DQ-NEXT: vpermt2q %zmm2, %zmm11, %zmm24
21036 ; AVX512DQ-NEXT: vmovdqa64 %zmm5, %zmm26
21037 ; AVX512DQ-NEXT: vpermt2q %zmm2, %zmm12, %zmm26
21038 ; AVX512DQ-NEXT: vmovdqa64 %zmm5, %zmm17
21039 ; AVX512DQ-NEXT: vpermt2q %zmm2, %zmm10, %zmm17
21040 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} zmm7 = zmm5[0],zmm2[0],zmm5[2],zmm2[2],zmm5[4],zmm2[4],zmm5[6],zmm2[6]
21041 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} zmm9 = zmm5[1],zmm2[1],zmm5[3],zmm2[3],zmm5[5],zmm2[5],zmm5[7],zmm2[7]
21042 ; AVX512DQ-NEXT: vpermt2q %zmm2, %zmm4, %zmm5
21043 ; AVX512DQ-NEXT: vmovdqa64 448(%r10), %zmm16
21044 ; AVX512DQ-NEXT: vmovdqa64 448(%rax), %zmm1
21045 ; AVX512DQ-NEXT: vmovdqa64 %zmm16, %zmm0
21046 ; AVX512DQ-NEXT: vpermt2q %zmm1, %zmm11, %zmm0
21047 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21048 ; AVX512DQ-NEXT: vmovdqa64 %zmm16, %zmm0
21049 ; AVX512DQ-NEXT: vpermt2q %zmm1, %zmm12, %zmm0
21050 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21051 ; AVX512DQ-NEXT: vpermi2q %zmm1, %zmm16, %zmm19
21052 ; AVX512DQ-NEXT: vpermi2q %zmm1, %zmm16, %zmm21
21053 ; AVX512DQ-NEXT: vmovdqa64 448(%r8), %zmm3
21054 ; AVX512DQ-NEXT: vmovdqa64 448(%r9), %zmm0
21055 ; AVX512DQ-NEXT: vpermi2q %zmm0, %zmm3, %zmm23
21056 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} zmm23 {%k1} = zmm16[0],zmm1[0],zmm16[2],zmm1[2],zmm16[4],zmm1[4],zmm16[6],zmm1[6]
21057 ; AVX512DQ-NEXT: vpermi2q %zmm0, %zmm3, %zmm25
21058 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} zmm25 {%k1} = zmm16[1],zmm1[1],zmm16[3],zmm1[3],zmm16[5],zmm1[5],zmm16[7],zmm1[7]
21059 ; AVX512DQ-NEXT: vmovdqa64 %zmm16, %zmm18
21060 ; AVX512DQ-NEXT: vpermt2q %zmm1, %zmm10, %zmm18
21061 ; AVX512DQ-NEXT: vpermt2q %zmm1, %zmm4, %zmm16
21062 ; AVX512DQ-NEXT: vpermi2q %zmm0, %zmm3, %zmm11
21063 ; AVX512DQ-NEXT: vpermi2q %zmm0, %zmm3, %zmm12
21064 ; AVX512DQ-NEXT: vpermi2q %zmm0, %zmm3, %zmm10
21065 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} zmm1 = zmm3[0],zmm0[0],zmm3[2],zmm0[2],zmm3[4],zmm0[4],zmm3[6],zmm0[6]
21066 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} zmm2 = zmm3[1],zmm0[1],zmm3[3],zmm0[3],zmm3[5],zmm0[5],zmm3[7],zmm0[7]
21067 ; AVX512DQ-NEXT: vpermt2q %zmm0, %zmm4, %zmm3
21068 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
21069 ; AVX512DQ-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
21070 ; AVX512DQ-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
21071 ; AVX512DQ-NEXT: vmovdqa64 %zmm14, %zmm7 {%k1}
21072 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm0, %zmm7, %zmm0
21073 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21074 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
21075 ; AVX512DQ-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
21076 ; AVX512DQ-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
21077 ; AVX512DQ-NEXT: vmovdqa64 %zmm15, %zmm9 {%k1}
21078 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm0, %zmm9, %zmm0
21079 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21080 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
21081 ; AVX512DQ-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
21082 ; AVX512DQ-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
21083 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm0, %zmm8, %zmm0
21084 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21085 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
21086 ; AVX512DQ-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
21087 ; AVX512DQ-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
21088 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm0, %zmm6, %zmm0
21089 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21090 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
21091 ; AVX512DQ-NEXT: vpblendd $240, (%rsp), %ymm0, %ymm0 # 32-byte Folded Reload
21092 ; AVX512DQ-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
21093 ; AVX512DQ-NEXT: vmovdqa64 %zmm19, %zmm1 {%k1}
21094 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
21095 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21096 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
21097 ; AVX512DQ-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
21098 ; AVX512DQ-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
21099 ; AVX512DQ-NEXT: vmovdqa64 %zmm21, %zmm2 {%k1}
21100 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm0, %zmm2, %zmm0
21101 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21102 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
21103 ; AVX512DQ-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
21104 ; AVX512DQ-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
21105 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm0, %zmm23, %zmm0
21106 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21107 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
21108 ; AVX512DQ-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
21109 ; AVX512DQ-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
21110 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm0, %zmm25, %zmm0
21111 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21112 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
21113 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
21114 ; AVX512DQ-NEXT: vmovdqa64 %zmm0, %zmm8 {%k1}
21115 ; AVX512DQ-NEXT: vmovdqa (%rcx), %ymm0
21116 ; AVX512DQ-NEXT: vmovdqa (%rdx), %ymm1
21117 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
21118 ; AVX512DQ-NEXT: vmovdqa (%rsi), %ymm4
21119 ; AVX512DQ-NEXT: vmovdqa (%rdi), %ymm6
21120 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm6[0],ymm4[0],ymm6[2],ymm4[2]
21121 ; AVX512DQ-NEXT: vperm2i128 {{.*#+}} ymm2 = ymm7[2,3],ymm2[2,3]
21122 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm2, %zmm8, %zmm2
21123 ; AVX512DQ-NEXT: vmovdqu64 %zmm2, (%rsp) # 64-byte Spill
21124 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
21125 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
21126 ; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm7 {%k1}
21127 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
21128 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm6[1],ymm4[1],ymm6[3],ymm4[3]
21129 ; AVX512DQ-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
21130 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm0, %zmm7, %zmm0
21131 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21132 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
21133 ; AVX512DQ-NEXT: vmovdqa64 %zmm0, %zmm31 {%k1}
21134 ; AVX512DQ-NEXT: vmovdqa 64(%rcx), %ymm0
21135 ; AVX512DQ-NEXT: vmovdqa 64(%rdx), %ymm1
21136 ; AVX512DQ-NEXT: vmovdqa 64(%rsi), %ymm2
21137 ; AVX512DQ-NEXT: vmovdqa 64(%rdi), %ymm4
21138 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} ymm6 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
21139 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm4[0],ymm2[0],ymm4[2],ymm2[2]
21140 ; AVX512DQ-NEXT: vperm2i128 {{.*#+}} ymm6 = ymm7[2,3],ymm6[2,3]
21141 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm6, %zmm31, %zmm6
21142 ; AVX512DQ-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21143 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
21144 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
21145 ; AVX512DQ-NEXT: vmovdqa64 %zmm6, %zmm7 {%k1}
21146 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
21147 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm4[1],ymm2[1],ymm4[3],ymm2[3]
21148 ; AVX512DQ-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
21149 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm0, %zmm7, %zmm8
21150 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
21151 ; AVX512DQ-NEXT: vmovdqa64 %zmm0, %zmm27 {%k1}
21152 ; AVX512DQ-NEXT: vmovdqa 128(%rcx), %ymm0
21153 ; AVX512DQ-NEXT: vmovdqa 128(%rdx), %ymm1
21154 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
21155 ; AVX512DQ-NEXT: vmovdqa 128(%rsi), %ymm4
21156 ; AVX512DQ-NEXT: vmovdqa 128(%rdi), %ymm6
21157 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm6[0],ymm4[0],ymm6[2],ymm4[2]
21158 ; AVX512DQ-NEXT: vperm2i128 {{.*#+}} ymm2 = ymm7[2,3],ymm2[2,3]
21159 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm2, %zmm27, %zmm23
21160 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
21161 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
21162 ; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm7 {%k1}
21163 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
21164 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm6[1],ymm4[1],ymm6[3],ymm4[3]
21165 ; AVX512DQ-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
21166 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm0, %zmm7, %zmm25
21167 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
21168 ; AVX512DQ-NEXT: vmovdqa64 %zmm0, %zmm20 {%k1}
21169 ; AVX512DQ-NEXT: vmovdqa 192(%rcx), %ymm0
21170 ; AVX512DQ-NEXT: vmovdqa 192(%rdx), %ymm1
21171 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
21172 ; AVX512DQ-NEXT: vmovdqa 192(%rsi), %ymm4
21173 ; AVX512DQ-NEXT: vmovdqa 192(%rdi), %ymm6
21174 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm6[0],ymm4[0],ymm6[2],ymm4[2]
21175 ; AVX512DQ-NEXT: vperm2i128 {{.*#+}} ymm2 = ymm7[2,3],ymm2[2,3]
21176 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm2, %zmm20, %zmm21
21177 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
21178 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
21179 ; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm7 {%k1}
21180 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
21181 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm6[1],ymm4[1],ymm6[3],ymm4[3]
21182 ; AVX512DQ-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
21183 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm0, %zmm7, %zmm19
21184 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
21185 ; AVX512DQ-NEXT: vmovdqa64 %zmm0, %zmm28 {%k1}
21186 ; AVX512DQ-NEXT: vmovdqa 256(%rcx), %ymm0
21187 ; AVX512DQ-NEXT: vmovdqa 256(%rdx), %ymm1
21188 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
21189 ; AVX512DQ-NEXT: vmovdqa 256(%rsi), %ymm4
21190 ; AVX512DQ-NEXT: vmovdqa 256(%rdi), %ymm6
21191 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm6[0],ymm4[0],ymm6[2],ymm4[2]
21192 ; AVX512DQ-NEXT: vperm2i128 {{.*#+}} ymm2 = ymm7[2,3],ymm2[2,3]
21193 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm2, %zmm28, %zmm15
21194 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
21195 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
21196 ; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm7 {%k1}
21197 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
21198 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm6[1],ymm4[1],ymm6[3],ymm4[3]
21199 ; AVX512DQ-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
21200 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm0, %zmm7, %zmm28
21201 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
21202 ; AVX512DQ-NEXT: vmovdqa64 %zmm0, %zmm22 {%k1}
21203 ; AVX512DQ-NEXT: vmovdqa 320(%rcx), %ymm0
21204 ; AVX512DQ-NEXT: vmovdqa 320(%rdx), %ymm1
21205 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
21206 ; AVX512DQ-NEXT: vmovdqa 320(%rsi), %ymm4
21207 ; AVX512DQ-NEXT: vmovdqa 320(%rdi), %ymm6
21208 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm6[0],ymm4[0],ymm6[2],ymm4[2]
21209 ; AVX512DQ-NEXT: vperm2i128 {{.*#+}} ymm2 = ymm7[2,3],ymm2[2,3]
21210 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm2, %zmm22, %zmm22
21211 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
21212 ; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm29 {%k1}
21213 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
21214 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm6[1],ymm4[1],ymm6[3],ymm4[3]
21215 ; AVX512DQ-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
21216 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm0, %zmm29, %zmm27
21217 ; AVX512DQ-NEXT: vmovdqa64 %zmm30, %zmm17 {%k1}
21218 ; AVX512DQ-NEXT: vmovdqa 384(%rcx), %ymm0
21219 ; AVX512DQ-NEXT: vmovdqa 384(%rdx), %ymm1
21220 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
21221 ; AVX512DQ-NEXT: vmovdqa 384(%rsi), %ymm4
21222 ; AVX512DQ-NEXT: vmovdqa 384(%rdi), %ymm6
21223 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm6[0],ymm4[0],ymm6[2],ymm4[2]
21224 ; AVX512DQ-NEXT: vperm2i128 {{.*#+}} ymm2 = ymm7[2,3],ymm2[2,3]
21225 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm2, %zmm17, %zmm14
21226 ; AVX512DQ-NEXT: vmovdqa64 %zmm13, %zmm5 {%k1}
21227 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
21228 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm6[1],ymm4[1],ymm6[3],ymm4[3]
21229 ; AVX512DQ-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
21230 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm0, %zmm5, %zmm13
21231 ; AVX512DQ-NEXT: vmovdqa64 %zmm18, %zmm10 {%k1}
21232 ; AVX512DQ-NEXT: vmovdqa 448(%rcx), %ymm0
21233 ; AVX512DQ-NEXT: vmovdqa 448(%rdx), %ymm1
21234 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
21235 ; AVX512DQ-NEXT: vmovdqa 448(%rsi), %ymm4
21236 ; AVX512DQ-NEXT: vmovdqa 448(%rdi), %ymm5
21237 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} ymm6 = ymm5[0],ymm4[0],ymm5[2],ymm4[2]
21238 ; AVX512DQ-NEXT: vperm2i128 {{.*#+}} ymm2 = ymm6[2,3],ymm2[2,3]
21239 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm2, %zmm10, %zmm10
21240 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
21241 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm5[1],ymm4[1],ymm5[3],ymm4[3]
21242 ; AVX512DQ-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
21243 ; AVX512DQ-NEXT: vmovdqa64 %zmm16, %zmm3 {%k1}
21244 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm0, %zmm3, %zmm9
21245 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
21246 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
21247 ; AVX512DQ-NEXT: vmovdqa64 %zmm0, %zmm3 {%k1}
21248 ; AVX512DQ-NEXT: vmovdqa (%rsi), %xmm0
21249 ; AVX512DQ-NEXT: vinserti128 $1, (%rcx), %ymm0, %ymm1
21250 ; AVX512DQ-NEXT: vmovdqa (%rdi), %xmm0
21251 ; AVX512DQ-NEXT: vinserti128 $1, (%rdx), %ymm0, %ymm2
21252 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm2[0],ymm1[0],ymm2[2],ymm1[2]
21253 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm0, %zmm3, %zmm0
21254 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
21255 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
21256 ; AVX512DQ-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
21257 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm2[1],ymm1[1],ymm2[3],ymm1[3]
21258 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm1, %zmm4, %zmm2
21259 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
21260 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
21261 ; AVX512DQ-NEXT: vmovdqa64 %zmm1, %zmm5 {%k1}
21262 ; AVX512DQ-NEXT: vmovdqa 64(%rsi), %xmm1
21263 ; AVX512DQ-NEXT: vinserti128 $1, 64(%rcx), %ymm1, %ymm1
21264 ; AVX512DQ-NEXT: vmovdqa 64(%rdi), %xmm3
21265 ; AVX512DQ-NEXT: vinserti128 $1, 64(%rdx), %ymm3, %ymm4
21266 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm4[0],ymm1[0],ymm4[2],ymm1[2]
21267 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm3, %zmm5, %zmm3
21268 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
21269 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
21270 ; AVX512DQ-NEXT: vmovdqa64 %zmm5, %zmm6 {%k1}
21271 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm4[1],ymm1[1],ymm4[3],ymm1[3]
21272 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm1, %zmm6, %zmm5
21273 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
21274 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
21275 ; AVX512DQ-NEXT: vmovdqa64 %zmm1, %zmm7 {%k1}
21276 ; AVX512DQ-NEXT: vmovdqa 128(%rsi), %xmm1
21277 ; AVX512DQ-NEXT: vinserti128 $1, 128(%rcx), %ymm1, %ymm1
21278 ; AVX512DQ-NEXT: vmovdqa 128(%rdi), %xmm4
21279 ; AVX512DQ-NEXT: vinserti128 $1, 128(%rdx), %ymm4, %ymm6
21280 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} ymm4 = ymm6[0],ymm1[0],ymm6[2],ymm1[2]
21281 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm4, %zmm7, %zmm4
21282 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
21283 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
21284 ; AVX512DQ-NEXT: vmovdqa64 %zmm7, %zmm16 {%k1}
21285 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm6[1],ymm1[1],ymm6[3],ymm1[3]
21286 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm1, %zmm16, %zmm6
21287 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
21288 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
21289 ; AVX512DQ-NEXT: vmovdqa64 %zmm1, %zmm17 {%k1}
21290 ; AVX512DQ-NEXT: vmovdqa 192(%rsi), %xmm1
21291 ; AVX512DQ-NEXT: vinserti128 $1, 192(%rcx), %ymm1, %ymm1
21292 ; AVX512DQ-NEXT: vmovdqa 192(%rdi), %xmm7
21293 ; AVX512DQ-NEXT: vinserti32x4 $1, 192(%rdx), %ymm7, %ymm16
21294 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm16[0],ymm1[0],ymm16[2],ymm1[2]
21295 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm7, %zmm17, %zmm7
21296 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
21297 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
21298 ; AVX512DQ-NEXT: vmovdqa64 %zmm17, %zmm18 {%k1}
21299 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm16[1],ymm1[1],ymm16[3],ymm1[3]
21300 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm1, %zmm18, %zmm1
21301 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
21302 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
21303 ; AVX512DQ-NEXT: vmovdqa64 %zmm16, %zmm20 {%k1}
21304 ; AVX512DQ-NEXT: vmovdqa64 256(%rsi), %xmm16
21305 ; AVX512DQ-NEXT: vinserti32x4 $1, 256(%rcx), %ymm16, %ymm16
21306 ; AVX512DQ-NEXT: vmovdqa64 256(%rdi), %xmm17
21307 ; AVX512DQ-NEXT: vinserti32x4 $1, 256(%rdx), %ymm17, %ymm17
21308 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} ymm18 = ymm17[0],ymm16[0],ymm17[2],ymm16[2]
21309 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm18, %zmm20, %zmm29
21310 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
21311 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
21312 ; AVX512DQ-NEXT: vmovdqa64 %zmm18, %zmm20 {%k1}
21313 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} ymm16 = ymm17[1],ymm16[1],ymm17[3],ymm16[3]
21314 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm16, %zmm20, %zmm30
21315 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
21316 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
21317 ; AVX512DQ-NEXT: vmovdqa64 %zmm16, %zmm20 {%k1}
21318 ; AVX512DQ-NEXT: vmovdqa64 320(%rsi), %xmm16
21319 ; AVX512DQ-NEXT: vinserti32x4 $1, 320(%rcx), %ymm16, %ymm16
21320 ; AVX512DQ-NEXT: vmovdqa64 320(%rdi), %xmm17
21321 ; AVX512DQ-NEXT: vinserti32x4 $1, 320(%rdx), %ymm17, %ymm17
21322 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} ymm18 = ymm17[0],ymm16[0],ymm17[2],ymm16[2]
21323 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm18, %zmm20, %zmm18
21324 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
21325 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
21326 ; AVX512DQ-NEXT: vmovdqa64 %zmm20, %zmm31 {%k1}
21327 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} ymm16 = ymm17[1],ymm16[1],ymm17[3],ymm16[3]
21328 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm16, %zmm31, %zmm16
21329 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
21330 ; AVX512DQ-NEXT: vmovdqa64 %zmm17, %zmm24 {%k1}
21331 ; AVX512DQ-NEXT: vmovdqa64 384(%rsi), %xmm17
21332 ; AVX512DQ-NEXT: vinserti32x4 $1, 384(%rcx), %ymm17, %ymm17
21333 ; AVX512DQ-NEXT: vmovdqa64 384(%rdi), %xmm20
21334 ; AVX512DQ-NEXT: vinserti32x4 $1, 384(%rdx), %ymm20, %ymm20
21335 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} ymm31 = ymm20[0],ymm17[0],ymm20[2],ymm17[2]
21336 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm31, %zmm24, %zmm24
21337 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
21338 ; AVX512DQ-NEXT: vmovdqa64 %zmm31, %zmm26 {%k1}
21339 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} ymm17 = ymm20[1],ymm17[1],ymm20[3],ymm17[3]
21340 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm17, %zmm26, %zmm17
21341 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
21342 ; AVX512DQ-NEXT: vmovdqa64 %zmm20, %zmm11 {%k1}
21343 ; AVX512DQ-NEXT: vmovdqa64 448(%rsi), %xmm20
21344 ; AVX512DQ-NEXT: vinserti32x4 $1, 448(%rcx), %ymm20, %ymm20
21345 ; AVX512DQ-NEXT: vmovdqa64 448(%rdi), %xmm26
21346 ; AVX512DQ-NEXT: vinserti32x4 $1, 448(%rdx), %ymm26, %ymm26
21347 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} ymm31 = ymm26[0],ymm20[0],ymm26[2],ymm20[2]
21348 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm31, %zmm11, %zmm11
21349 ; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
21350 ; AVX512DQ-NEXT: vmovdqa64 %zmm31, %zmm12 {%k1}
21351 ; AVX512DQ-NEXT: vpunpckhqdq {{.*#+}} ymm20 = ymm26[1],ymm20[1],ymm26[3],ymm20[3]
21352 ; AVX512DQ-NEXT: vinserti64x4 $0, %ymm20, %zmm12, %zmm12
21353 ; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %rax
21354 ; AVX512DQ-NEXT: vmovdqa64 %zmm9, 3776(%rax)
21355 ; AVX512DQ-NEXT: vmovdqa64 %zmm10, 3712(%rax)
21356 ; AVX512DQ-NEXT: vmovdqa64 %zmm13, 3264(%rax)
21357 ; AVX512DQ-NEXT: vmovdqa64 %zmm14, 3200(%rax)
21358 ; AVX512DQ-NEXT: vmovdqa64 %zmm27, 2752(%rax)
21359 ; AVX512DQ-NEXT: vmovdqa64 %zmm22, 2688(%rax)
21360 ; AVX512DQ-NEXT: vmovdqa64 %zmm28, 2240(%rax)
21361 ; AVX512DQ-NEXT: vmovdqa64 %zmm15, 2176(%rax)
21362 ; AVX512DQ-NEXT: vmovdqa64 %zmm19, 1728(%rax)
21363 ; AVX512DQ-NEXT: vmovdqa64 %zmm21, 1664(%rax)
21364 ; AVX512DQ-NEXT: vmovdqa64 %zmm25, 1216(%rax)
21365 ; AVX512DQ-NEXT: vmovdqa64 %zmm23, 1152(%rax)
21366 ; AVX512DQ-NEXT: vmovdqa64 %zmm8, 704(%rax)
21367 ; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
21368 ; AVX512DQ-NEXT: vmovaps %zmm8, 640(%rax)
21369 ; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
21370 ; AVX512DQ-NEXT: vmovaps %zmm8, 192(%rax)
21371 ; AVX512DQ-NEXT: vmovups (%rsp), %zmm8 # 64-byte Reload
21372 ; AVX512DQ-NEXT: vmovaps %zmm8, 128(%rax)
21373 ; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
21374 ; AVX512DQ-NEXT: vmovaps %zmm8, 4032(%rax)
21375 ; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
21376 ; AVX512DQ-NEXT: vmovaps %zmm8, 3968(%rax)
21377 ; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
21378 ; AVX512DQ-NEXT: vmovaps %zmm8, 3904(%rax)
21379 ; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
21380 ; AVX512DQ-NEXT: vmovaps %zmm8, 3840(%rax)
21381 ; AVX512DQ-NEXT: vmovdqa64 %zmm12, 3648(%rax)
21382 ; AVX512DQ-NEXT: vmovdqa64 %zmm11, 3584(%rax)
21383 ; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
21384 ; AVX512DQ-NEXT: vmovaps %zmm8, 3520(%rax)
21385 ; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
21386 ; AVX512DQ-NEXT: vmovaps %zmm8, 3456(%rax)
21387 ; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
21388 ; AVX512DQ-NEXT: vmovaps %zmm8, 3392(%rax)
21389 ; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
21390 ; AVX512DQ-NEXT: vmovaps %zmm8, 3328(%rax)
21391 ; AVX512DQ-NEXT: vmovdqa64 %zmm17, 3136(%rax)
21392 ; AVX512DQ-NEXT: vmovdqa64 %zmm24, 3072(%rax)
21393 ; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
21394 ; AVX512DQ-NEXT: vmovaps %zmm8, 3008(%rax)
21395 ; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
21396 ; AVX512DQ-NEXT: vmovaps %zmm8, 2944(%rax)
21397 ; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
21398 ; AVX512DQ-NEXT: vmovaps %zmm8, 2880(%rax)
21399 ; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
21400 ; AVX512DQ-NEXT: vmovaps %zmm8, 2816(%rax)
21401 ; AVX512DQ-NEXT: vmovdqa64 %zmm16, 2624(%rax)
21402 ; AVX512DQ-NEXT: vmovdqa64 %zmm18, 2560(%rax)
21403 ; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
21404 ; AVX512DQ-NEXT: vmovaps %zmm8, 2496(%rax)
21405 ; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
21406 ; AVX512DQ-NEXT: vmovaps %zmm8, 2432(%rax)
21407 ; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
21408 ; AVX512DQ-NEXT: vmovaps %zmm8, 2368(%rax)
21409 ; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
21410 ; AVX512DQ-NEXT: vmovaps %zmm8, 2304(%rax)
21411 ; AVX512DQ-NEXT: vmovdqa64 %zmm30, 2112(%rax)
21412 ; AVX512DQ-NEXT: vmovdqa64 %zmm29, 2048(%rax)
21413 ; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
21414 ; AVX512DQ-NEXT: vmovaps %zmm8, 1984(%rax)
21415 ; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
21416 ; AVX512DQ-NEXT: vmovaps %zmm8, 1920(%rax)
21417 ; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
21418 ; AVX512DQ-NEXT: vmovaps %zmm8, 1856(%rax)
21419 ; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
21420 ; AVX512DQ-NEXT: vmovaps %zmm8, 1792(%rax)
21421 ; AVX512DQ-NEXT: vmovdqa64 %zmm1, 1600(%rax)
21422 ; AVX512DQ-NEXT: vmovdqa64 %zmm7, 1536(%rax)
21423 ; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
21424 ; AVX512DQ-NEXT: vmovaps %zmm1, 1472(%rax)
21425 ; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
21426 ; AVX512DQ-NEXT: vmovaps %zmm1, 1408(%rax)
21427 ; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
21428 ; AVX512DQ-NEXT: vmovaps %zmm1, 1344(%rax)
21429 ; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
21430 ; AVX512DQ-NEXT: vmovaps %zmm1, 1280(%rax)
21431 ; AVX512DQ-NEXT: vmovdqa64 %zmm6, 1088(%rax)
21432 ; AVX512DQ-NEXT: vmovdqa64 %zmm4, 1024(%rax)
21433 ; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
21434 ; AVX512DQ-NEXT: vmovaps %zmm1, 960(%rax)
21435 ; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
21436 ; AVX512DQ-NEXT: vmovaps %zmm1, 896(%rax)
21437 ; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
21438 ; AVX512DQ-NEXT: vmovaps %zmm1, 832(%rax)
21439 ; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
21440 ; AVX512DQ-NEXT: vmovaps %zmm1, 768(%rax)
21441 ; AVX512DQ-NEXT: vmovdqa64 %zmm5, 576(%rax)
21442 ; AVX512DQ-NEXT: vmovdqa64 %zmm3, 512(%rax)
21443 ; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
21444 ; AVX512DQ-NEXT: vmovaps %zmm1, 448(%rax)
21445 ; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
21446 ; AVX512DQ-NEXT: vmovaps %zmm1, 384(%rax)
21447 ; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
21448 ; AVX512DQ-NEXT: vmovaps %zmm1, 320(%rax)
21449 ; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
21450 ; AVX512DQ-NEXT: vmovaps %zmm1, 256(%rax)
21451 ; AVX512DQ-NEXT: vmovdqa64 %zmm2, 64(%rax)
21452 ; AVX512DQ-NEXT: vmovdqa64 %zmm0, (%rax)
21453 ; AVX512DQ-NEXT: addq $5384, %rsp # imm = 0x1508
21454 ; AVX512DQ-NEXT: vzeroupper
21455 ; AVX512DQ-NEXT: retq
21456 ;
21457 ; AVX512DQ-FCP-LABEL: store_i64_stride8_vf64:
21458 ; AVX512DQ-FCP: # %bb.0:
21459 ; AVX512DQ-FCP-NEXT: subq $5384, %rsp # imm = 0x1508
21460 ; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
21461 ; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10
21462 ; AVX512DQ-FCP-NEXT: vmovdqa64 128(%rdi), %zmm0
21463 ; AVX512DQ-FCP-NEXT: vmovdqa64 64(%rdi), %zmm4
21464 ; AVX512DQ-FCP-NEXT: vmovdqa64 (%rdi), %zmm14
21465 ; AVX512DQ-FCP-NEXT: vmovdqa64 128(%rsi), %zmm1
21466 ; AVX512DQ-FCP-NEXT: vmovdqa64 64(%rsi), %zmm5
21467 ; AVX512DQ-FCP-NEXT: vmovdqa64 (%rsi), %zmm10
21468 ; AVX512DQ-FCP-NEXT: vmovdqa64 128(%rdx), %zmm2
21469 ; AVX512DQ-FCP-NEXT: vmovdqa64 64(%rdx), %zmm7
21470 ; AVX512DQ-FCP-NEXT: vmovdqa64 (%rdx), %zmm12
21471 ; AVX512DQ-FCP-NEXT: vmovdqa64 128(%rcx), %zmm3
21472 ; AVX512DQ-FCP-NEXT: vmovdqa64 64(%rcx), %zmm9
21473 ; AVX512DQ-FCP-NEXT: vmovdqa64 (%rcx), %zmm15
21474 ; AVX512DQ-FCP-NEXT: vmovdqa64 (%r8), %zmm30
21475 ; AVX512DQ-FCP-NEXT: vmovdqa64 64(%r8), %zmm18
21476 ; AVX512DQ-FCP-NEXT: vmovdqa64 128(%r8), %zmm11
21477 ; AVX512DQ-FCP-NEXT: vmovdqa64 (%r9), %zmm24
21478 ; AVX512DQ-FCP-NEXT: vmovdqa64 64(%r9), %zmm28
21479 ; AVX512DQ-FCP-NEXT: vmovdqa64 128(%r9), %zmm22
21480 ; AVX512DQ-FCP-NEXT: vmovdqa64 (%r10), %zmm26
21481 ; AVX512DQ-FCP-NEXT: vmovdqa64 64(%r10), %zmm31
21482 ; AVX512DQ-FCP-NEXT: vmovdqa64 128(%r10), %zmm16
21483 ; AVX512DQ-FCP-NEXT: vmovdqa64 (%rax), %zmm17
21484 ; AVX512DQ-FCP-NEXT: vmovdqa64 64(%rax), %zmm27
21485 ; AVX512DQ-FCP-NEXT: vmovdqa64 128(%rax), %zmm13
21486 ; AVX512DQ-FCP-NEXT: movb $-64, %r11b
21487 ; AVX512DQ-FCP-NEXT: kmovw %r11d, %k1
21488 ; AVX512DQ-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm19 = [4,12,4,12,4,12,4,12]
21489 ; AVX512DQ-FCP-NEXT: # zmm19 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
21490 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm12, %zmm8
21491 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm15, %zmm19, %zmm8
21492 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm14, %zmm6
21493 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm10, %zmm19, %zmm6
21494 ; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm8[4,5,6,7]
21495 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm26, %zmm8
21496 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm17, %zmm19, %zmm8
21497 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm20 = zmm30[0],zmm24[0],zmm30[2],zmm24[2],zmm30[4],zmm24[4],zmm30[6],zmm24[6]
21498 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm8, %zmm20 {%k1}
21499 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm20, %zmm6
21500 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21501 ; AVX512DQ-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm21 = [5,13,5,13,5,13,5,13]
21502 ; AVX512DQ-FCP-NEXT: # zmm21 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
21503 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm12, %zmm6
21504 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm15, %zmm21, %zmm6
21505 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm14, %zmm8
21506 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm10, %zmm21, %zmm8
21507 ; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm8[0,1,2,3],ymm6[4,5,6,7]
21508 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm26, %zmm8
21509 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm17, %zmm21, %zmm8
21510 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm20 = zmm30[1],zmm24[1],zmm30[3],zmm24[3],zmm30[5],zmm24[5],zmm30[7],zmm24[7]
21511 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm8, %zmm20 {%k1}
21512 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm20, %zmm6
21513 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21514 ; AVX512DQ-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm23 = [6,14,6,14,6,14,6,14]
21515 ; AVX512DQ-FCP-NEXT: # zmm23 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
21516 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm12, %zmm6
21517 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm15, %zmm23, %zmm6
21518 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm14, %zmm8
21519 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm10, %zmm23, %zmm8
21520 ; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm8[0,1,2,3],ymm6[4,5,6,7]
21521 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm30, %zmm8
21522 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm24, %zmm23, %zmm8
21523 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm8 {%k1} = zmm26[0],zmm17[0],zmm26[2],zmm17[2],zmm26[4],zmm17[4],zmm26[6],zmm17[6]
21524 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm8, %zmm6
21525 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21526 ; AVX512DQ-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm25 = [7,15,7,15,7,15,7,15]
21527 ; AVX512DQ-FCP-NEXT: # zmm25 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
21528 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm15, %zmm25, %zmm12
21529 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm10, %zmm25, %zmm14
21530 ; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm14[0,1,2,3],ymm12[4,5,6,7]
21531 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm30, %zmm8
21532 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm24, %zmm25, %zmm8
21533 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm8 {%k1} = zmm26[1],zmm17[1],zmm26[3],zmm17[3],zmm26[5],zmm17[5],zmm26[7],zmm17[7]
21534 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm8, %zmm6
21535 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21536 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm7, %zmm6
21537 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm9, %zmm19, %zmm6
21538 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm4, %zmm8
21539 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm5, %zmm19, %zmm8
21540 ; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm8[0,1,2,3],ymm6[4,5,6,7]
21541 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm31, %zmm8
21542 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm27, %zmm19, %zmm8
21543 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm10 = zmm18[0],zmm28[0],zmm18[2],zmm28[2],zmm18[4],zmm28[4],zmm18[6],zmm28[6]
21544 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}
21545 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm10, %zmm6
21546 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21547 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm7, %zmm6
21548 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm9, %zmm21, %zmm6
21549 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm4, %zmm8
21550 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm5, %zmm21, %zmm8
21551 ; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm8[0,1,2,3],ymm6[4,5,6,7]
21552 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm31, %zmm8
21553 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm27, %zmm21, %zmm8
21554 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm10 = zmm18[1],zmm28[1],zmm18[3],zmm28[3],zmm18[5],zmm28[5],zmm18[7],zmm28[7]
21555 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}
21556 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm10, %zmm6
21557 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21558 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm7, %zmm6
21559 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm9, %zmm23, %zmm6
21560 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm4, %zmm8
21561 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm5, %zmm23, %zmm8
21562 ; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm8[0,1,2,3],ymm6[4,5,6,7]
21563 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm18, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21564 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm18, %zmm8
21565 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm28, %zmm23, %zmm8
21566 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm8 {%k1} = zmm31[0],zmm27[0],zmm31[2],zmm27[2],zmm31[4],zmm27[4],zmm31[6],zmm27[6]
21567 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm8, %zmm6
21568 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21569 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm9, %zmm25, %zmm7
21570 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm5, %zmm25, %zmm4
21571 ; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm7[4,5,6,7]
21572 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm18, %zmm5
21573 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm28, %zmm25, %zmm5
21574 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm5 {%k1} = zmm31[1],zmm27[1],zmm31[3],zmm27[3],zmm31[5],zmm27[5],zmm31[7],zmm27[7]
21575 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm4, %zmm5, %zmm4
21576 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21577 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm4
21578 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm3, %zmm19, %zmm4
21579 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm5
21580 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm19, %zmm5
21581 ; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm5[0,1,2,3],ymm4[4,5,6,7]
21582 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm16, %zmm5
21583 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm13, %zmm19, %zmm5
21584 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm6 = zmm11[0],zmm22[0],zmm11[2],zmm22[2],zmm11[4],zmm22[4],zmm11[6],zmm22[6]
21585 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm5, %zmm6 {%k1}
21586 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm4, %zmm6, %zmm4
21587 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21588 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm4
21589 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm3, %zmm21, %zmm4
21590 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm5
21591 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm21, %zmm5
21592 ; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm5[0,1,2,3],ymm4[4,5,6,7]
21593 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm16, %zmm5
21594 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm13, %zmm21, %zmm5
21595 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm6 = zmm11[1],zmm22[1],zmm11[3],zmm22[3],zmm11[5],zmm22[5],zmm11[7],zmm22[7]
21596 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm5, %zmm6 {%k1}
21597 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm4, %zmm6, %zmm4
21598 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21599 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm4
21600 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm3, %zmm23, %zmm4
21601 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm5
21602 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm23, %zmm5
21603 ; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm5[0,1,2,3],ymm4[4,5,6,7]
21604 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm11, %zmm5
21605 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21606 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm22, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21607 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm22, %zmm23, %zmm5
21608 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21609 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21610 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm5 {%k1} = zmm16[0],zmm13[0],zmm16[2],zmm13[2],zmm16[4],zmm13[4],zmm16[6],zmm13[6]
21611 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm4, %zmm5, %zmm4
21612 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21613 ; AVX512DQ-FCP-NEXT: vmovdqa64 192(%rdi), %zmm4
21614 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm3, %zmm25, %zmm2
21615 ; AVX512DQ-FCP-NEXT: vmovdqa64 192(%rdx), %zmm3
21616 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm25, %zmm0
21617 ; AVX512DQ-FCP-NEXT: vmovdqa64 192(%rcx), %zmm1
21618 ; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
21619 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm11, %zmm2
21620 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm22, %zmm25, %zmm2
21621 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm2 {%k1} = zmm16[1],zmm13[1],zmm16[3],zmm13[3],zmm16[5],zmm13[5],zmm16[7],zmm13[7]
21622 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm2, %zmm0
21623 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21624 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, %zmm0
21625 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm19, %zmm0
21626 ; AVX512DQ-FCP-NEXT: vmovdqa64 192(%rsi), %zmm2
21627 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm4, %zmm5
21628 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm19, %zmm5
21629 ; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm0[4,5,6,7]
21630 ; AVX512DQ-FCP-NEXT: vmovdqa64 192(%r10), %zmm9
21631 ; AVX512DQ-FCP-NEXT: vmovdqa64 192(%rax), %zmm20
21632 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm9, %zmm6
21633 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm20, %zmm19, %zmm6
21634 ; AVX512DQ-FCP-NEXT: vmovdqa64 192(%r8), %zmm8
21635 ; AVX512DQ-FCP-NEXT: vmovdqa64 192(%r9), %zmm0
21636 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm7 = zmm8[0],zmm0[0],zmm8[2],zmm0[2],zmm8[4],zmm0[4],zmm8[6],zmm0[6]
21637 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, %zmm7 {%k1}
21638 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm5, %zmm7, %zmm5
21639 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21640 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, %zmm5
21641 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm21, %zmm5
21642 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm4, %zmm6
21643 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm21, %zmm6
21644 ; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm5 = ymm6[0,1,2,3],ymm5[4,5,6,7]
21645 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm9, %zmm6
21646 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm20, %zmm21, %zmm6
21647 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm7 = zmm8[1],zmm0[1],zmm8[3],zmm0[3],zmm8[5],zmm0[5],zmm8[7],zmm0[7]
21648 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, %zmm7 {%k1}
21649 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm5, %zmm7, %zmm5
21650 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21651 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, %zmm5
21652 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm23, %zmm5
21653 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm4, %zmm6
21654 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm23, %zmm6
21655 ; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm5 = ymm6[0,1,2,3],ymm5[4,5,6,7]
21656 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21657 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm8, %zmm6
21658 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm23, %zmm6
21659 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm6 {%k1} = zmm9[0],zmm20[0],zmm9[2],zmm20[2],zmm9[4],zmm20[4],zmm9[6],zmm20[6]
21660 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm5, %zmm6, %zmm5
21661 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21662 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm25, %zmm3
21663 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm25, %zmm4
21664 ; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm1 = ymm4[0,1,2,3],ymm3[4,5,6,7]
21665 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm8, %zmm2
21666 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm25, %zmm2
21667 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm2 {%k1} = zmm9[1],zmm20[1],zmm9[3],zmm20[3],zmm9[5],zmm20[5],zmm9[7],zmm20[7]
21668 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm1, %zmm2, %zmm1
21669 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21670 ; AVX512DQ-FCP-NEXT: vmovdqa64 256(%rdx), %zmm1
21671 ; AVX512DQ-FCP-NEXT: vmovdqa64 256(%rcx), %zmm3
21672 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, %zmm2
21673 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm3, %zmm19, %zmm2
21674 ; AVX512DQ-FCP-NEXT: vmovdqa64 256(%rdi), %zmm4
21675 ; AVX512DQ-FCP-NEXT: vmovdqa64 256(%rsi), %zmm5
21676 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm4, %zmm6
21677 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm5, %zmm19, %zmm6
21678 ; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm2[4,5,6,7]
21679 ; AVX512DQ-FCP-NEXT: vmovdqa64 256(%r10), %zmm13
21680 ; AVX512DQ-FCP-NEXT: vmovdqa64 256(%rax), %zmm2
21681 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm13, %zmm7
21682 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm19, %zmm7
21683 ; AVX512DQ-FCP-NEXT: vmovdqa64 256(%r8), %zmm18
21684 ; AVX512DQ-FCP-NEXT: vmovdqa64 256(%r9), %zmm15
21685 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm8 = zmm18[0],zmm15[0],zmm18[2],zmm15[2],zmm18[4],zmm15[4],zmm18[6],zmm15[6]
21686 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm7, %zmm8 {%k1}
21687 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm8, %zmm6
21688 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21689 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, %zmm6
21690 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm3, %zmm21, %zmm6
21691 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm4, %zmm7
21692 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm5, %zmm21, %zmm7
21693 ; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]
21694 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm13, %zmm7
21695 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm21, %zmm7
21696 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm8 = zmm18[1],zmm15[1],zmm18[3],zmm15[3],zmm18[5],zmm15[5],zmm18[7],zmm15[7]
21697 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm7, %zmm8 {%k1}
21698 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm8, %zmm6
21699 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21700 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, %zmm6
21701 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm3, %zmm23, %zmm6
21702 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm4, %zmm7
21703 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm5, %zmm23, %zmm7
21704 ; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]
21705 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm18, %zmm7
21706 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm15, %zmm23, %zmm7
21707 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm7 {%k1} = zmm13[0],zmm2[0],zmm13[2],zmm2[2],zmm13[4],zmm2[4],zmm13[6],zmm2[6]
21708 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm7, %zmm6
21709 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21710 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm3, %zmm25, %zmm1
21711 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm5, %zmm25, %zmm4
21712 ; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm1 = ymm4[0,1,2,3],ymm1[4,5,6,7]
21713 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm18, %zmm3
21714 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm15, %zmm25, %zmm3
21715 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm3 {%k1} = zmm13[1],zmm2[1],zmm13[3],zmm2[3],zmm13[5],zmm2[5],zmm13[7],zmm2[7]
21716 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm1, %zmm3, %zmm1
21717 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21718 ; AVX512DQ-FCP-NEXT: vmovdqa64 320(%rdx), %zmm1
21719 ; AVX512DQ-FCP-NEXT: vmovdqa64 320(%rcx), %zmm4
21720 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, %zmm3
21721 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm4, %zmm19, %zmm3
21722 ; AVX512DQ-FCP-NEXT: vmovdqa64 320(%rdi), %zmm6
21723 ; AVX512DQ-FCP-NEXT: vmovdqa64 320(%rsi), %zmm7
21724 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, %zmm5
21725 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm7, %zmm19, %zmm5
21726 ; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm8 = ymm5[0,1,2,3],ymm3[4,5,6,7]
21727 ; AVX512DQ-FCP-NEXT: vmovdqa64 320(%r10), %zmm16
21728 ; AVX512DQ-FCP-NEXT: vmovdqa64 320(%rax), %zmm22
21729 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm16, %zmm12
21730 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm22, %zmm19, %zmm12
21731 ; AVX512DQ-FCP-NEXT: vmovdqa64 320(%r8), %zmm29
21732 ; AVX512DQ-FCP-NEXT: vmovdqa64 320(%r9), %zmm5
21733 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm14 = zmm29[0],zmm5[0],zmm29[2],zmm5[2],zmm29[4],zmm5[4],zmm29[6],zmm5[6]
21734 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm12, %zmm14 {%k1}
21735 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm8, %zmm14, %zmm3
21736 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21737 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, %zmm8
21738 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm4, %zmm21, %zmm8
21739 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, %zmm12
21740 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm7, %zmm21, %zmm12
21741 ; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm8 = ymm12[0,1,2,3],ymm8[4,5,6,7]
21742 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm16, %zmm12
21743 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm22, %zmm21, %zmm12
21744 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm14 = zmm29[1],zmm5[1],zmm29[3],zmm5[3],zmm29[5],zmm5[5],zmm29[7],zmm5[7]
21745 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm12, %zmm14 {%k1}
21746 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm8, %zmm14, %zmm3
21747 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21748 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, %zmm8
21749 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm4, %zmm23, %zmm8
21750 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, %zmm12
21751 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm7, %zmm23, %zmm12
21752 ; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm8 = ymm12[0,1,2,3],ymm8[4,5,6,7]
21753 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm29, %zmm12
21754 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm5, %zmm23, %zmm12
21755 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm12 {%k1} = zmm16[0],zmm22[0],zmm16[2],zmm22[2],zmm16[4],zmm22[4],zmm16[6],zmm22[6]
21756 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm8, %zmm12, %zmm3
21757 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21758 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm4, %zmm25, %zmm1
21759 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm7, %zmm25, %zmm6
21760 ; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm1 = ymm6[0,1,2,3],ymm1[4,5,6,7]
21761 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm29, %zmm4
21762 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm5, %zmm25, %zmm4
21763 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm4 {%k1} = zmm16[1],zmm22[1],zmm16[3],zmm22[3],zmm16[5],zmm22[5],zmm16[7],zmm22[7]
21764 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm1, %zmm4, %zmm1
21765 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21766 ; AVX512DQ-FCP-NEXT: vmovdqa64 384(%rdx), %zmm3
21767 ; AVX512DQ-FCP-NEXT: vmovdqa64 384(%rcx), %zmm1
21768 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, %zmm4
21769 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm19, %zmm4
21770 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21771 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, %zmm4
21772 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm21, %zmm4
21773 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21774 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, %zmm4
21775 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm23, %zmm4
21776 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21777 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm25, %zmm3
21778 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21779 ; AVX512DQ-FCP-NEXT: vmovdqa64 384(%rdi), %zmm3
21780 ; AVX512DQ-FCP-NEXT: vmovdqa64 384(%rsi), %zmm1
21781 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, %zmm4
21782 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm19, %zmm4
21783 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21784 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, %zmm4
21785 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm21, %zmm4
21786 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21787 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, %zmm4
21788 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm23, %zmm4
21789 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21790 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm25, %zmm3
21791 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21792 ; AVX512DQ-FCP-NEXT: vmovdqa64 448(%rdx), %zmm3
21793 ; AVX512DQ-FCP-NEXT: vmovdqa64 448(%rcx), %zmm1
21794 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, %zmm4
21795 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm19, %zmm4
21796 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm4, (%rsp) # 64-byte Spill
21797 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, %zmm4
21798 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm21, %zmm4
21799 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21800 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, %zmm4
21801 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm23, %zmm4
21802 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21803 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm25, %zmm3
21804 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21805 ; AVX512DQ-FCP-NEXT: vmovdqa64 448(%rdi), %zmm3
21806 ; AVX512DQ-FCP-NEXT: vmovdqa64 448(%rsi), %zmm1
21807 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, %zmm4
21808 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm19, %zmm4
21809 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21810 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, %zmm4
21811 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm21, %zmm4
21812 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21813 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, %zmm4
21814 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm23, %zmm4
21815 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21816 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm25, %zmm3
21817 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21818 ; AVX512DQ-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm11 = [0,8,0,8,0,8,0,8]
21819 ; AVX512DQ-FCP-NEXT: # zmm11 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
21820 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm26, %zmm1
21821 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm17, %zmm11, %zmm1
21822 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21823 ; AVX512DQ-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm12 = [1,9,1,9,1,9,1,9]
21824 ; AVX512DQ-FCP-NEXT: # zmm12 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
21825 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm26, %zmm1
21826 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm17, %zmm12, %zmm1
21827 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21828 ; AVX512DQ-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm10 = [2,10,2,10,2,10,2,10]
21829 ; AVX512DQ-FCP-NEXT: # zmm10 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
21830 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm26, %zmm1
21831 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm17, %zmm10, %zmm1
21832 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21833 ; AVX512DQ-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm4 = [3,11,3,11,3,11,3,11]
21834 ; AVX512DQ-FCP-NEXT: # zmm4 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
21835 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm17, %zmm4, %zmm26
21836 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm26, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21837 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm30, %zmm1
21838 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm24, %zmm11, %zmm1
21839 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21840 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm30, %zmm1
21841 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm24, %zmm12, %zmm1
21842 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21843 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm30, %zmm1
21844 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm24, %zmm10, %zmm1
21845 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21846 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm24, %zmm4, %zmm30
21847 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm30, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21848 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm31, %zmm1
21849 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm27, %zmm11, %zmm1
21850 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21851 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm31, %zmm1
21852 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm27, %zmm12, %zmm1
21853 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21854 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm31, %zmm1
21855 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm27, %zmm10, %zmm1
21856 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21857 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm27, %zmm4, %zmm31
21858 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm31, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21859 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
21860 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, %zmm1
21861 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm28, %zmm11, %zmm1
21862 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21863 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, %zmm1
21864 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm28, %zmm12, %zmm1
21865 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21866 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, %zmm31
21867 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm28, %zmm10, %zmm31
21868 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm28, %zmm4, %zmm3
21869 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21870 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
21871 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm28, %zmm3
21872 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
21873 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm11, %zmm3
21874 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21875 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm28, %zmm3
21876 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm12, %zmm3
21877 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21878 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm28, %zmm3
21879 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm10, %zmm3
21880 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21881 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm4, %zmm28
21882 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm28, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21883 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
21884 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, %zmm3
21885 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
21886 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm11, %zmm3
21887 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21888 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, %zmm3
21889 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm12, %zmm3
21890 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21891 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, %zmm27
21892 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm10, %zmm27
21893 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm4, %zmm6
21894 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21895 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm9, %zmm1
21896 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm20, %zmm11, %zmm1
21897 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21898 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm9, %zmm1
21899 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm20, %zmm12, %zmm1
21900 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21901 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm9, %zmm1
21902 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm20, %zmm10, %zmm1
21903 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21904 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm20, %zmm4, %zmm9
21905 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21906 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
21907 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, %zmm3
21908 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm11, %zmm3
21909 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21910 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, %zmm3
21911 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm12, %zmm3
21912 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21913 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, %zmm20
21914 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm10, %zmm20
21915 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm4, %zmm1
21916 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21917 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm13, %zmm0
21918 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm11, %zmm0
21919 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21920 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm13, %zmm0
21921 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm12, %zmm0
21922 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21923 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm13, %zmm0
21924 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm10, %zmm0
21925 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21926 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm4, %zmm13
21927 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21928 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm18, %zmm0
21929 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm15, %zmm11, %zmm0
21930 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21931 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm18, %zmm0
21932 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm15, %zmm12, %zmm0
21933 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21934 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm18, %zmm28
21935 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm15, %zmm10, %zmm28
21936 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm15, %zmm4, %zmm18
21937 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm18, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21938 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm16, %zmm0
21939 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm22, %zmm11, %zmm0
21940 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21941 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm16, %zmm0
21942 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm22, %zmm12, %zmm0
21943 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21944 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm16, %zmm0
21945 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm22, %zmm10, %zmm0
21946 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21947 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm22, %zmm4, %zmm16
21948 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21949 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm29, %zmm0
21950 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm5, %zmm11, %zmm0
21951 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21952 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm29, %zmm0
21953 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm5, %zmm12, %zmm0
21954 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21955 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm29, %zmm22
21956 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm5, %zmm10, %zmm22
21957 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm5, %zmm4, %zmm29
21958 ; AVX512DQ-FCP-NEXT: vmovdqa64 384(%r10), %zmm13
21959 ; AVX512DQ-FCP-NEXT: vmovdqa64 384(%rax), %zmm0
21960 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm13, %zmm1
21961 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm11, %zmm1
21962 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21963 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm13, %zmm1
21964 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm12, %zmm1
21965 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21966 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm13, %zmm14
21967 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm19, %zmm14
21968 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm13, %zmm15
21969 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm21, %zmm15
21970 ; AVX512DQ-FCP-NEXT: vmovdqa64 384(%r8), %zmm5
21971 ; AVX512DQ-FCP-NEXT: vmovdqa64 384(%r9), %zmm2
21972 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm5, %zmm8
21973 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm23, %zmm8
21974 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm8 {%k1} = zmm13[0],zmm0[0],zmm13[2],zmm0[2],zmm13[4],zmm0[4],zmm13[6],zmm0[6]
21975 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm5, %zmm6
21976 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm25, %zmm6
21977 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm6 {%k1} = zmm13[1],zmm0[1],zmm13[3],zmm0[3],zmm13[5],zmm0[5],zmm13[7],zmm0[7]
21978 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm13, %zmm30
21979 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm10, %zmm30
21980 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm4, %zmm13
21981 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm5, %zmm24
21982 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm11, %zmm24
21983 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm5, %zmm26
21984 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm12, %zmm26
21985 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm5, %zmm17
21986 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm10, %zmm17
21987 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm7 = zmm5[0],zmm2[0],zmm5[2],zmm2[2],zmm5[4],zmm2[4],zmm5[6],zmm2[6]
21988 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm9 = zmm5[1],zmm2[1],zmm5[3],zmm2[3],zmm5[5],zmm2[5],zmm5[7],zmm2[7]
21989 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm4, %zmm5
21990 ; AVX512DQ-FCP-NEXT: vmovdqa64 448(%r10), %zmm16
21991 ; AVX512DQ-FCP-NEXT: vmovdqa64 448(%rax), %zmm1
21992 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm16, %zmm0
21993 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm11, %zmm0
21994 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21995 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm16, %zmm0
21996 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm12, %zmm0
21997 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
21998 ; AVX512DQ-FCP-NEXT: vpermi2q %zmm1, %zmm16, %zmm19
21999 ; AVX512DQ-FCP-NEXT: vpermi2q %zmm1, %zmm16, %zmm21
22000 ; AVX512DQ-FCP-NEXT: vmovdqa64 448(%r8), %zmm3
22001 ; AVX512DQ-FCP-NEXT: vmovdqa64 448(%r9), %zmm0
22002 ; AVX512DQ-FCP-NEXT: vpermi2q %zmm0, %zmm3, %zmm23
22003 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm23 {%k1} = zmm16[0],zmm1[0],zmm16[2],zmm1[2],zmm16[4],zmm1[4],zmm16[6],zmm1[6]
22004 ; AVX512DQ-FCP-NEXT: vpermi2q %zmm0, %zmm3, %zmm25
22005 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm25 {%k1} = zmm16[1],zmm1[1],zmm16[3],zmm1[3],zmm16[5],zmm1[5],zmm16[7],zmm1[7]
22006 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm16, %zmm18
22007 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm10, %zmm18
22008 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm4, %zmm16
22009 ; AVX512DQ-FCP-NEXT: vpermi2q %zmm0, %zmm3, %zmm11
22010 ; AVX512DQ-FCP-NEXT: vpermi2q %zmm0, %zmm3, %zmm12
22011 ; AVX512DQ-FCP-NEXT: vpermi2q %zmm0, %zmm3, %zmm10
22012 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm1 = zmm3[0],zmm0[0],zmm3[2],zmm0[2],zmm3[4],zmm0[4],zmm3[6],zmm0[6]
22013 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm2 = zmm3[1],zmm0[1],zmm3[3],zmm0[3],zmm3[5],zmm0[5],zmm3[7],zmm0[7]
22014 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm4, %zmm3
22015 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
22016 ; AVX512DQ-FCP-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
22017 ; AVX512DQ-FCP-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
22018 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm14, %zmm7 {%k1}
22019 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm7, %zmm0
22020 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22021 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
22022 ; AVX512DQ-FCP-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
22023 ; AVX512DQ-FCP-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
22024 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm15, %zmm9 {%k1}
22025 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm9, %zmm0
22026 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22027 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
22028 ; AVX512DQ-FCP-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
22029 ; AVX512DQ-FCP-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
22030 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm8, %zmm0
22031 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22032 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
22033 ; AVX512DQ-FCP-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
22034 ; AVX512DQ-FCP-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
22035 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm6, %zmm0
22036 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22037 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
22038 ; AVX512DQ-FCP-NEXT: vpblendd $240, (%rsp), %ymm0, %ymm0 # 32-byte Folded Reload
22039 ; AVX512DQ-FCP-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
22040 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm19, %zmm1 {%k1}
22041 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
22042 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22043 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
22044 ; AVX512DQ-FCP-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
22045 ; AVX512DQ-FCP-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
22046 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm21, %zmm2 {%k1}
22047 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm2, %zmm0
22048 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22049 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
22050 ; AVX512DQ-FCP-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
22051 ; AVX512DQ-FCP-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
22052 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm23, %zmm0
22053 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22054 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
22055 ; AVX512DQ-FCP-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
22056 ; AVX512DQ-FCP-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
22057 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm25, %zmm0
22058 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22059 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
22060 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
22061 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm8 {%k1}
22062 ; AVX512DQ-FCP-NEXT: vmovdqa (%rcx), %ymm0
22063 ; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %ymm1
22064 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
22065 ; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %ymm4
22066 ; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %ymm6
22067 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm6[0],ymm4[0],ymm6[2],ymm4[2]
22068 ; AVX512DQ-FCP-NEXT: vperm2i128 {{.*#+}} ymm2 = ymm7[2,3],ymm2[2,3]
22069 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm8, %zmm2
22070 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm2, (%rsp) # 64-byte Spill
22071 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
22072 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
22073 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm7 {%k1}
22074 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
22075 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm6[1],ymm4[1],ymm6[3],ymm4[3]
22076 ; AVX512DQ-FCP-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
22077 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm7, %zmm0
22078 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22079 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
22080 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm31 {%k1}
22081 ; AVX512DQ-FCP-NEXT: vmovdqa 64(%rcx), %ymm0
22082 ; AVX512DQ-FCP-NEXT: vmovdqa 64(%rdx), %ymm1
22083 ; AVX512DQ-FCP-NEXT: vmovdqa 64(%rsi), %ymm2
22084 ; AVX512DQ-FCP-NEXT: vmovdqa 64(%rdi), %ymm4
22085 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm6 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
22086 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm4[0],ymm2[0],ymm4[2],ymm2[2]
22087 ; AVX512DQ-FCP-NEXT: vperm2i128 {{.*#+}} ymm6 = ymm7[2,3],ymm6[2,3]
22088 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm31, %zmm6
22089 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22090 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
22091 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
22092 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, %zmm7 {%k1}
22093 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
22094 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm4[1],ymm2[1],ymm4[3],ymm2[3]
22095 ; AVX512DQ-FCP-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
22096 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm7, %zmm8
22097 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
22098 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm27 {%k1}
22099 ; AVX512DQ-FCP-NEXT: vmovdqa 128(%rcx), %ymm0
22100 ; AVX512DQ-FCP-NEXT: vmovdqa 128(%rdx), %ymm1
22101 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
22102 ; AVX512DQ-FCP-NEXT: vmovdqa 128(%rsi), %ymm4
22103 ; AVX512DQ-FCP-NEXT: vmovdqa 128(%rdi), %ymm6
22104 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm6[0],ymm4[0],ymm6[2],ymm4[2]
22105 ; AVX512DQ-FCP-NEXT: vperm2i128 {{.*#+}} ymm2 = ymm7[2,3],ymm2[2,3]
22106 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm27, %zmm23
22107 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
22108 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
22109 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm7 {%k1}
22110 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
22111 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm6[1],ymm4[1],ymm6[3],ymm4[3]
22112 ; AVX512DQ-FCP-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
22113 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm7, %zmm25
22114 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
22115 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm20 {%k1}
22116 ; AVX512DQ-FCP-NEXT: vmovdqa 192(%rcx), %ymm0
22117 ; AVX512DQ-FCP-NEXT: vmovdqa 192(%rdx), %ymm1
22118 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
22119 ; AVX512DQ-FCP-NEXT: vmovdqa 192(%rsi), %ymm4
22120 ; AVX512DQ-FCP-NEXT: vmovdqa 192(%rdi), %ymm6
22121 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm6[0],ymm4[0],ymm6[2],ymm4[2]
22122 ; AVX512DQ-FCP-NEXT: vperm2i128 {{.*#+}} ymm2 = ymm7[2,3],ymm2[2,3]
22123 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm20, %zmm21
22124 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
22125 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
22126 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm7 {%k1}
22127 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
22128 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm6[1],ymm4[1],ymm6[3],ymm4[3]
22129 ; AVX512DQ-FCP-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
22130 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm7, %zmm19
22131 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
22132 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm28 {%k1}
22133 ; AVX512DQ-FCP-NEXT: vmovdqa 256(%rcx), %ymm0
22134 ; AVX512DQ-FCP-NEXT: vmovdqa 256(%rdx), %ymm1
22135 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
22136 ; AVX512DQ-FCP-NEXT: vmovdqa 256(%rsi), %ymm4
22137 ; AVX512DQ-FCP-NEXT: vmovdqa 256(%rdi), %ymm6
22138 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm6[0],ymm4[0],ymm6[2],ymm4[2]
22139 ; AVX512DQ-FCP-NEXT: vperm2i128 {{.*#+}} ymm2 = ymm7[2,3],ymm2[2,3]
22140 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm28, %zmm15
22141 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
22142 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
22143 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm7 {%k1}
22144 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
22145 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm6[1],ymm4[1],ymm6[3],ymm4[3]
22146 ; AVX512DQ-FCP-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
22147 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm7, %zmm28
22148 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
22149 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm22 {%k1}
22150 ; AVX512DQ-FCP-NEXT: vmovdqa 320(%rcx), %ymm0
22151 ; AVX512DQ-FCP-NEXT: vmovdqa 320(%rdx), %ymm1
22152 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
22153 ; AVX512DQ-FCP-NEXT: vmovdqa 320(%rsi), %ymm4
22154 ; AVX512DQ-FCP-NEXT: vmovdqa 320(%rdi), %ymm6
22155 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm6[0],ymm4[0],ymm6[2],ymm4[2]
22156 ; AVX512DQ-FCP-NEXT: vperm2i128 {{.*#+}} ymm2 = ymm7[2,3],ymm2[2,3]
22157 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm22, %zmm22
22158 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
22159 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm29 {%k1}
22160 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
22161 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm6[1],ymm4[1],ymm6[3],ymm4[3]
22162 ; AVX512DQ-FCP-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
22163 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm29, %zmm27
22164 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm30, %zmm17 {%k1}
22165 ; AVX512DQ-FCP-NEXT: vmovdqa 384(%rcx), %ymm0
22166 ; AVX512DQ-FCP-NEXT: vmovdqa 384(%rdx), %ymm1
22167 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
22168 ; AVX512DQ-FCP-NEXT: vmovdqa 384(%rsi), %ymm4
22169 ; AVX512DQ-FCP-NEXT: vmovdqa 384(%rdi), %ymm6
22170 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm6[0],ymm4[0],ymm6[2],ymm4[2]
22171 ; AVX512DQ-FCP-NEXT: vperm2i128 {{.*#+}} ymm2 = ymm7[2,3],ymm2[2,3]
22172 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm17, %zmm14
22173 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm13, %zmm5 {%k1}
22174 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
22175 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm6[1],ymm4[1],ymm6[3],ymm4[3]
22176 ; AVX512DQ-FCP-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
22177 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm5, %zmm13
22178 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm18, %zmm10 {%k1}
22179 ; AVX512DQ-FCP-NEXT: vmovdqa 448(%rcx), %ymm0
22180 ; AVX512DQ-FCP-NEXT: vmovdqa 448(%rdx), %ymm1
22181 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
22182 ; AVX512DQ-FCP-NEXT: vmovdqa 448(%rsi), %ymm4
22183 ; AVX512DQ-FCP-NEXT: vmovdqa 448(%rdi), %ymm5
22184 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm6 = ymm5[0],ymm4[0],ymm5[2],ymm4[2]
22185 ; AVX512DQ-FCP-NEXT: vperm2i128 {{.*#+}} ymm2 = ymm6[2,3],ymm2[2,3]
22186 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm10, %zmm10
22187 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
22188 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm5[1],ymm4[1],ymm5[3],ymm4[3]
22189 ; AVX512DQ-FCP-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
22190 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm16, %zmm3 {%k1}
22191 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm3, %zmm9
22192 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
22193 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
22194 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm3 {%k1}
22195 ; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %xmm0
22196 ; AVX512DQ-FCP-NEXT: vinserti128 $1, (%rcx), %ymm0, %ymm1
22197 ; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %xmm0
22198 ; AVX512DQ-FCP-NEXT: vinserti128 $1, (%rdx), %ymm0, %ymm2
22199 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm2[0],ymm1[0],ymm2[2],ymm1[2]
22200 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm3, %zmm0
22201 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
22202 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
22203 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
22204 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm2[1],ymm1[1],ymm2[3],ymm1[3]
22205 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm1, %zmm4, %zmm2
22206 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
22207 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
22208 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, %zmm5 {%k1}
22209 ; AVX512DQ-FCP-NEXT: vmovdqa 64(%rsi), %xmm1
22210 ; AVX512DQ-FCP-NEXT: vinserti128 $1, 64(%rcx), %ymm1, %ymm1
22211 ; AVX512DQ-FCP-NEXT: vmovdqa 64(%rdi), %xmm3
22212 ; AVX512DQ-FCP-NEXT: vinserti128 $1, 64(%rdx), %ymm3, %ymm4
22213 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm4[0],ymm1[0],ymm4[2],ymm1[2]
22214 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm3, %zmm5, %zmm3
22215 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
22216 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
22217 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm5, %zmm6 {%k1}
22218 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm4[1],ymm1[1],ymm4[3],ymm1[3]
22219 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm1, %zmm6, %zmm5
22220 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
22221 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
22222 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, %zmm7 {%k1}
22223 ; AVX512DQ-FCP-NEXT: vmovdqa 128(%rsi), %xmm1
22224 ; AVX512DQ-FCP-NEXT: vinserti128 $1, 128(%rcx), %ymm1, %ymm1
22225 ; AVX512DQ-FCP-NEXT: vmovdqa 128(%rdi), %xmm4
22226 ; AVX512DQ-FCP-NEXT: vinserti128 $1, 128(%rdx), %ymm4, %ymm6
22227 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm4 = ymm6[0],ymm1[0],ymm6[2],ymm1[2]
22228 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm4, %zmm7, %zmm4
22229 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
22230 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
22231 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm7, %zmm16 {%k1}
22232 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm6[1],ymm1[1],ymm6[3],ymm1[3]
22233 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm1, %zmm16, %zmm6
22234 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
22235 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
22236 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, %zmm17 {%k1}
22237 ; AVX512DQ-FCP-NEXT: vmovdqa 192(%rsi), %xmm1
22238 ; AVX512DQ-FCP-NEXT: vinserti128 $1, 192(%rcx), %ymm1, %ymm1
22239 ; AVX512DQ-FCP-NEXT: vmovdqa 192(%rdi), %xmm7
22240 ; AVX512DQ-FCP-NEXT: vinserti32x4 $1, 192(%rdx), %ymm7, %ymm16
22241 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm16[0],ymm1[0],ymm16[2],ymm1[2]
22242 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm7, %zmm17, %zmm7
22243 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
22244 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
22245 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm17, %zmm18 {%k1}
22246 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm16[1],ymm1[1],ymm16[3],ymm1[3]
22247 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm1, %zmm18, %zmm1
22248 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
22249 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
22250 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm16, %zmm20 {%k1}
22251 ; AVX512DQ-FCP-NEXT: vmovdqa64 256(%rsi), %xmm16
22252 ; AVX512DQ-FCP-NEXT: vinserti32x4 $1, 256(%rcx), %ymm16, %ymm16
22253 ; AVX512DQ-FCP-NEXT: vmovdqa64 256(%rdi), %xmm17
22254 ; AVX512DQ-FCP-NEXT: vinserti32x4 $1, 256(%rdx), %ymm17, %ymm17
22255 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm18 = ymm17[0],ymm16[0],ymm17[2],ymm16[2]
22256 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm18, %zmm20, %zmm29
22257 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
22258 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
22259 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm18, %zmm20 {%k1}
22260 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm16 = ymm17[1],ymm16[1],ymm17[3],ymm16[3]
22261 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm16, %zmm20, %zmm30
22262 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
22263 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
22264 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm16, %zmm20 {%k1}
22265 ; AVX512DQ-FCP-NEXT: vmovdqa64 320(%rsi), %xmm16
22266 ; AVX512DQ-FCP-NEXT: vinserti32x4 $1, 320(%rcx), %ymm16, %ymm16
22267 ; AVX512DQ-FCP-NEXT: vmovdqa64 320(%rdi), %xmm17
22268 ; AVX512DQ-FCP-NEXT: vinserti32x4 $1, 320(%rdx), %ymm17, %ymm17
22269 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm18 = ymm17[0],ymm16[0],ymm17[2],ymm16[2]
22270 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm18, %zmm20, %zmm18
22271 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
22272 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
22273 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm20, %zmm31 {%k1}
22274 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm16 = ymm17[1],ymm16[1],ymm17[3],ymm16[3]
22275 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm16, %zmm31, %zmm16
22276 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
22277 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm17, %zmm24 {%k1}
22278 ; AVX512DQ-FCP-NEXT: vmovdqa64 384(%rsi), %xmm17
22279 ; AVX512DQ-FCP-NEXT: vinserti32x4 $1, 384(%rcx), %ymm17, %ymm17
22280 ; AVX512DQ-FCP-NEXT: vmovdqa64 384(%rdi), %xmm20
22281 ; AVX512DQ-FCP-NEXT: vinserti32x4 $1, 384(%rdx), %ymm20, %ymm20
22282 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm31 = ymm20[0],ymm17[0],ymm20[2],ymm17[2]
22283 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm31, %zmm24, %zmm24
22284 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
22285 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm31, %zmm26 {%k1}
22286 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm17 = ymm20[1],ymm17[1],ymm20[3],ymm17[3]
22287 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm17, %zmm26, %zmm17
22288 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
22289 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm20, %zmm11 {%k1}
22290 ; AVX512DQ-FCP-NEXT: vmovdqa64 448(%rsi), %xmm20
22291 ; AVX512DQ-FCP-NEXT: vinserti32x4 $1, 448(%rcx), %ymm20, %ymm20
22292 ; AVX512DQ-FCP-NEXT: vmovdqa64 448(%rdi), %xmm26
22293 ; AVX512DQ-FCP-NEXT: vinserti32x4 $1, 448(%rdx), %ymm26, %ymm26
22294 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm31 = ymm26[0],ymm20[0],ymm26[2],ymm20[2]
22295 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm31, %zmm11, %zmm11
22296 ; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
22297 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm31, %zmm12 {%k1}
22298 ; AVX512DQ-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm20 = ymm26[1],ymm20[1],ymm26[3],ymm20[3]
22299 ; AVX512DQ-FCP-NEXT: vinserti64x4 $0, %ymm20, %zmm12, %zmm12
22300 ; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
22301 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm9, 3776(%rax)
22302 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm10, 3712(%rax)
22303 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm13, 3264(%rax)
22304 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm14, 3200(%rax)
22305 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm27, 2752(%rax)
22306 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm22, 2688(%rax)
22307 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm28, 2240(%rax)
22308 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm15, 2176(%rax)
22309 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm19, 1728(%rax)
22310 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm21, 1664(%rax)
22311 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm25, 1216(%rax)
22312 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm23, 1152(%rax)
22313 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm8, 704(%rax)
22314 ; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
22315 ; AVX512DQ-FCP-NEXT: vmovaps %zmm8, 640(%rax)
22316 ; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
22317 ; AVX512DQ-FCP-NEXT: vmovaps %zmm8, 192(%rax)
22318 ; AVX512DQ-FCP-NEXT: vmovups (%rsp), %zmm8 # 64-byte Reload
22319 ; AVX512DQ-FCP-NEXT: vmovaps %zmm8, 128(%rax)
22320 ; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
22321 ; AVX512DQ-FCP-NEXT: vmovaps %zmm8, 4032(%rax)
22322 ; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
22323 ; AVX512DQ-FCP-NEXT: vmovaps %zmm8, 3968(%rax)
22324 ; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
22325 ; AVX512DQ-FCP-NEXT: vmovaps %zmm8, 3904(%rax)
22326 ; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
22327 ; AVX512DQ-FCP-NEXT: vmovaps %zmm8, 3840(%rax)
22328 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm12, 3648(%rax)
22329 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm11, 3584(%rax)
22330 ; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
22331 ; AVX512DQ-FCP-NEXT: vmovaps %zmm8, 3520(%rax)
22332 ; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
22333 ; AVX512DQ-FCP-NEXT: vmovaps %zmm8, 3456(%rax)
22334 ; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
22335 ; AVX512DQ-FCP-NEXT: vmovaps %zmm8, 3392(%rax)
22336 ; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
22337 ; AVX512DQ-FCP-NEXT: vmovaps %zmm8, 3328(%rax)
22338 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm17, 3136(%rax)
22339 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm24, 3072(%rax)
22340 ; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
22341 ; AVX512DQ-FCP-NEXT: vmovaps %zmm8, 3008(%rax)
22342 ; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
22343 ; AVX512DQ-FCP-NEXT: vmovaps %zmm8, 2944(%rax)
22344 ; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
22345 ; AVX512DQ-FCP-NEXT: vmovaps %zmm8, 2880(%rax)
22346 ; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
22347 ; AVX512DQ-FCP-NEXT: vmovaps %zmm8, 2816(%rax)
22348 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm16, 2624(%rax)
22349 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm18, 2560(%rax)
22350 ; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
22351 ; AVX512DQ-FCP-NEXT: vmovaps %zmm8, 2496(%rax)
22352 ; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
22353 ; AVX512DQ-FCP-NEXT: vmovaps %zmm8, 2432(%rax)
22354 ; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
22355 ; AVX512DQ-FCP-NEXT: vmovaps %zmm8, 2368(%rax)
22356 ; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
22357 ; AVX512DQ-FCP-NEXT: vmovaps %zmm8, 2304(%rax)
22358 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm30, 2112(%rax)
22359 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm29, 2048(%rax)
22360 ; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
22361 ; AVX512DQ-FCP-NEXT: vmovaps %zmm8, 1984(%rax)
22362 ; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
22363 ; AVX512DQ-FCP-NEXT: vmovaps %zmm8, 1920(%rax)
22364 ; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
22365 ; AVX512DQ-FCP-NEXT: vmovaps %zmm8, 1856(%rax)
22366 ; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
22367 ; AVX512DQ-FCP-NEXT: vmovaps %zmm8, 1792(%rax)
22368 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, 1600(%rax)
22369 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm7, 1536(%rax)
22370 ; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
22371 ; AVX512DQ-FCP-NEXT: vmovaps %zmm1, 1472(%rax)
22372 ; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
22373 ; AVX512DQ-FCP-NEXT: vmovaps %zmm1, 1408(%rax)
22374 ; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
22375 ; AVX512DQ-FCP-NEXT: vmovaps %zmm1, 1344(%rax)
22376 ; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
22377 ; AVX512DQ-FCP-NEXT: vmovaps %zmm1, 1280(%rax)
22378 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, 1088(%rax)
22379 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm4, 1024(%rax)
22380 ; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
22381 ; AVX512DQ-FCP-NEXT: vmovaps %zmm1, 960(%rax)
22382 ; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
22383 ; AVX512DQ-FCP-NEXT: vmovaps %zmm1, 896(%rax)
22384 ; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
22385 ; AVX512DQ-FCP-NEXT: vmovaps %zmm1, 832(%rax)
22386 ; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
22387 ; AVX512DQ-FCP-NEXT: vmovaps %zmm1, 768(%rax)
22388 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm5, 576(%rax)
22389 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, 512(%rax)
22390 ; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
22391 ; AVX512DQ-FCP-NEXT: vmovaps %zmm1, 448(%rax)
22392 ; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
22393 ; AVX512DQ-FCP-NEXT: vmovaps %zmm1, 384(%rax)
22394 ; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
22395 ; AVX512DQ-FCP-NEXT: vmovaps %zmm1, 320(%rax)
22396 ; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
22397 ; AVX512DQ-FCP-NEXT: vmovaps %zmm1, 256(%rax)
22398 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, 64(%rax)
22399 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, (%rax)
22400 ; AVX512DQ-FCP-NEXT: addq $5384, %rsp # imm = 0x1508
22401 ; AVX512DQ-FCP-NEXT: vzeroupper
22402 ; AVX512DQ-FCP-NEXT: retq
22403 ;
22404 ; AVX512BW-LABEL: store_i64_stride8_vf64:
22405 ; AVX512BW: # %bb.0:
22406 ; AVX512BW-NEXT: subq $5384, %rsp # imm = 0x1508
22407 ; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
22408 ; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %r10
22409 ; AVX512BW-NEXT: vmovdqa64 128(%rdi), %zmm0
22410 ; AVX512BW-NEXT: vmovdqa64 64(%rdi), %zmm4
22411 ; AVX512BW-NEXT: vmovdqa64 (%rdi), %zmm14
22412 ; AVX512BW-NEXT: vmovdqa64 128(%rsi), %zmm1
22413 ; AVX512BW-NEXT: vmovdqa64 64(%rsi), %zmm5
22414 ; AVX512BW-NEXT: vmovdqa64 (%rsi), %zmm10
22415 ; AVX512BW-NEXT: vmovdqa64 128(%rdx), %zmm2
22416 ; AVX512BW-NEXT: vmovdqa64 64(%rdx), %zmm7
22417 ; AVX512BW-NEXT: vmovdqa64 (%rdx), %zmm12
22418 ; AVX512BW-NEXT: vmovdqa64 128(%rcx), %zmm3
22419 ; AVX512BW-NEXT: vmovdqa64 64(%rcx), %zmm9
22420 ; AVX512BW-NEXT: vmovdqa64 (%rcx), %zmm15
22421 ; AVX512BW-NEXT: vmovdqa64 (%r8), %zmm30
22422 ; AVX512BW-NEXT: vmovdqa64 64(%r8), %zmm18
22423 ; AVX512BW-NEXT: vmovdqa64 128(%r8), %zmm11
22424 ; AVX512BW-NEXT: vmovdqa64 (%r9), %zmm24
22425 ; AVX512BW-NEXT: vmovdqa64 64(%r9), %zmm28
22426 ; AVX512BW-NEXT: vmovdqa64 128(%r9), %zmm22
22427 ; AVX512BW-NEXT: vmovdqa64 (%r10), %zmm26
22428 ; AVX512BW-NEXT: vmovdqa64 64(%r10), %zmm31
22429 ; AVX512BW-NEXT: vmovdqa64 128(%r10), %zmm16
22430 ; AVX512BW-NEXT: vmovdqa64 (%rax), %zmm17
22431 ; AVX512BW-NEXT: vmovdqa64 64(%rax), %zmm27
22432 ; AVX512BW-NEXT: vmovdqa64 128(%rax), %zmm13
22433 ; AVX512BW-NEXT: movb $-64, %r11b
22434 ; AVX512BW-NEXT: kmovd %r11d, %k1
22435 ; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm19 = [4,12,4,12,4,12,4,12]
22436 ; AVX512BW-NEXT: # zmm19 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
22437 ; AVX512BW-NEXT: vmovdqa64 %zmm12, %zmm8
22438 ; AVX512BW-NEXT: vpermt2q %zmm15, %zmm19, %zmm8
22439 ; AVX512BW-NEXT: vmovdqa64 %zmm14, %zmm6
22440 ; AVX512BW-NEXT: vpermt2q %zmm10, %zmm19, %zmm6
22441 ; AVX512BW-NEXT: vpblendd {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm8[4,5,6,7]
22442 ; AVX512BW-NEXT: vmovdqa64 %zmm26, %zmm8
22443 ; AVX512BW-NEXT: vpermt2q %zmm17, %zmm19, %zmm8
22444 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} zmm20 = zmm30[0],zmm24[0],zmm30[2],zmm24[2],zmm30[4],zmm24[4],zmm30[6],zmm24[6]
22445 ; AVX512BW-NEXT: vmovdqa64 %zmm8, %zmm20 {%k1}
22446 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm6, %zmm20, %zmm6
22447 ; AVX512BW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22448 ; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm21 = [5,13,5,13,5,13,5,13]
22449 ; AVX512BW-NEXT: # zmm21 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
22450 ; AVX512BW-NEXT: vmovdqa64 %zmm12, %zmm6
22451 ; AVX512BW-NEXT: vpermt2q %zmm15, %zmm21, %zmm6
22452 ; AVX512BW-NEXT: vmovdqa64 %zmm14, %zmm8
22453 ; AVX512BW-NEXT: vpermt2q %zmm10, %zmm21, %zmm8
22454 ; AVX512BW-NEXT: vpblendd {{.*#+}} ymm6 = ymm8[0,1,2,3],ymm6[4,5,6,7]
22455 ; AVX512BW-NEXT: vmovdqa64 %zmm26, %zmm8
22456 ; AVX512BW-NEXT: vpermt2q %zmm17, %zmm21, %zmm8
22457 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} zmm20 = zmm30[1],zmm24[1],zmm30[3],zmm24[3],zmm30[5],zmm24[5],zmm30[7],zmm24[7]
22458 ; AVX512BW-NEXT: vmovdqa64 %zmm8, %zmm20 {%k1}
22459 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm6, %zmm20, %zmm6
22460 ; AVX512BW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22461 ; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm23 = [6,14,6,14,6,14,6,14]
22462 ; AVX512BW-NEXT: # zmm23 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
22463 ; AVX512BW-NEXT: vmovdqa64 %zmm12, %zmm6
22464 ; AVX512BW-NEXT: vpermt2q %zmm15, %zmm23, %zmm6
22465 ; AVX512BW-NEXT: vmovdqa64 %zmm14, %zmm8
22466 ; AVX512BW-NEXT: vpermt2q %zmm10, %zmm23, %zmm8
22467 ; AVX512BW-NEXT: vpblendd {{.*#+}} ymm6 = ymm8[0,1,2,3],ymm6[4,5,6,7]
22468 ; AVX512BW-NEXT: vmovdqa64 %zmm30, %zmm8
22469 ; AVX512BW-NEXT: vpermt2q %zmm24, %zmm23, %zmm8
22470 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} zmm8 {%k1} = zmm26[0],zmm17[0],zmm26[2],zmm17[2],zmm26[4],zmm17[4],zmm26[6],zmm17[6]
22471 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm6, %zmm8, %zmm6
22472 ; AVX512BW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22473 ; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm25 = [7,15,7,15,7,15,7,15]
22474 ; AVX512BW-NEXT: # zmm25 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
22475 ; AVX512BW-NEXT: vpermt2q %zmm15, %zmm25, %zmm12
22476 ; AVX512BW-NEXT: vpermt2q %zmm10, %zmm25, %zmm14
22477 ; AVX512BW-NEXT: vpblendd {{.*#+}} ymm6 = ymm14[0,1,2,3],ymm12[4,5,6,7]
22478 ; AVX512BW-NEXT: vmovdqa64 %zmm30, %zmm8
22479 ; AVX512BW-NEXT: vpermt2q %zmm24, %zmm25, %zmm8
22480 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} zmm8 {%k1} = zmm26[1],zmm17[1],zmm26[3],zmm17[3],zmm26[5],zmm17[5],zmm26[7],zmm17[7]
22481 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm6, %zmm8, %zmm6
22482 ; AVX512BW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22483 ; AVX512BW-NEXT: vmovdqa64 %zmm7, %zmm6
22484 ; AVX512BW-NEXT: vpermt2q %zmm9, %zmm19, %zmm6
22485 ; AVX512BW-NEXT: vmovdqa64 %zmm4, %zmm8
22486 ; AVX512BW-NEXT: vpermt2q %zmm5, %zmm19, %zmm8
22487 ; AVX512BW-NEXT: vpblendd {{.*#+}} ymm6 = ymm8[0,1,2,3],ymm6[4,5,6,7]
22488 ; AVX512BW-NEXT: vmovdqa64 %zmm31, %zmm8
22489 ; AVX512BW-NEXT: vpermt2q %zmm27, %zmm19, %zmm8
22490 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} zmm10 = zmm18[0],zmm28[0],zmm18[2],zmm28[2],zmm18[4],zmm28[4],zmm18[6],zmm28[6]
22491 ; AVX512BW-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}
22492 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm6, %zmm10, %zmm6
22493 ; AVX512BW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22494 ; AVX512BW-NEXT: vmovdqa64 %zmm7, %zmm6
22495 ; AVX512BW-NEXT: vpermt2q %zmm9, %zmm21, %zmm6
22496 ; AVX512BW-NEXT: vmovdqa64 %zmm4, %zmm8
22497 ; AVX512BW-NEXT: vpermt2q %zmm5, %zmm21, %zmm8
22498 ; AVX512BW-NEXT: vpblendd {{.*#+}} ymm6 = ymm8[0,1,2,3],ymm6[4,5,6,7]
22499 ; AVX512BW-NEXT: vmovdqa64 %zmm31, %zmm8
22500 ; AVX512BW-NEXT: vpermt2q %zmm27, %zmm21, %zmm8
22501 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} zmm10 = zmm18[1],zmm28[1],zmm18[3],zmm28[3],zmm18[5],zmm28[5],zmm18[7],zmm28[7]
22502 ; AVX512BW-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}
22503 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm6, %zmm10, %zmm6
22504 ; AVX512BW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22505 ; AVX512BW-NEXT: vmovdqa64 %zmm7, %zmm6
22506 ; AVX512BW-NEXT: vpermt2q %zmm9, %zmm23, %zmm6
22507 ; AVX512BW-NEXT: vmovdqa64 %zmm4, %zmm8
22508 ; AVX512BW-NEXT: vpermt2q %zmm5, %zmm23, %zmm8
22509 ; AVX512BW-NEXT: vpblendd {{.*#+}} ymm6 = ymm8[0,1,2,3],ymm6[4,5,6,7]
22510 ; AVX512BW-NEXT: vmovdqu64 %zmm18, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22511 ; AVX512BW-NEXT: vmovdqa64 %zmm18, %zmm8
22512 ; AVX512BW-NEXT: vpermt2q %zmm28, %zmm23, %zmm8
22513 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} zmm8 {%k1} = zmm31[0],zmm27[0],zmm31[2],zmm27[2],zmm31[4],zmm27[4],zmm31[6],zmm27[6]
22514 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm6, %zmm8, %zmm6
22515 ; AVX512BW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22516 ; AVX512BW-NEXT: vpermt2q %zmm9, %zmm25, %zmm7
22517 ; AVX512BW-NEXT: vpermt2q %zmm5, %zmm25, %zmm4
22518 ; AVX512BW-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm7[4,5,6,7]
22519 ; AVX512BW-NEXT: vmovdqa64 %zmm18, %zmm5
22520 ; AVX512BW-NEXT: vpermt2q %zmm28, %zmm25, %zmm5
22521 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} zmm5 {%k1} = zmm31[1],zmm27[1],zmm31[3],zmm27[3],zmm31[5],zmm27[5],zmm31[7],zmm27[7]
22522 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm4, %zmm5, %zmm4
22523 ; AVX512BW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22524 ; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm4
22525 ; AVX512BW-NEXT: vpermt2q %zmm3, %zmm19, %zmm4
22526 ; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm5
22527 ; AVX512BW-NEXT: vpermt2q %zmm1, %zmm19, %zmm5
22528 ; AVX512BW-NEXT: vpblendd {{.*#+}} ymm4 = ymm5[0,1,2,3],ymm4[4,5,6,7]
22529 ; AVX512BW-NEXT: vmovdqa64 %zmm16, %zmm5
22530 ; AVX512BW-NEXT: vpermt2q %zmm13, %zmm19, %zmm5
22531 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} zmm6 = zmm11[0],zmm22[0],zmm11[2],zmm22[2],zmm11[4],zmm22[4],zmm11[6],zmm22[6]
22532 ; AVX512BW-NEXT: vmovdqa64 %zmm5, %zmm6 {%k1}
22533 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm4, %zmm6, %zmm4
22534 ; AVX512BW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22535 ; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm4
22536 ; AVX512BW-NEXT: vpermt2q %zmm3, %zmm21, %zmm4
22537 ; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm5
22538 ; AVX512BW-NEXT: vpermt2q %zmm1, %zmm21, %zmm5
22539 ; AVX512BW-NEXT: vpblendd {{.*#+}} ymm4 = ymm5[0,1,2,3],ymm4[4,5,6,7]
22540 ; AVX512BW-NEXT: vmovdqa64 %zmm16, %zmm5
22541 ; AVX512BW-NEXT: vpermt2q %zmm13, %zmm21, %zmm5
22542 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} zmm6 = zmm11[1],zmm22[1],zmm11[3],zmm22[3],zmm11[5],zmm22[5],zmm11[7],zmm22[7]
22543 ; AVX512BW-NEXT: vmovdqa64 %zmm5, %zmm6 {%k1}
22544 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm4, %zmm6, %zmm4
22545 ; AVX512BW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22546 ; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm4
22547 ; AVX512BW-NEXT: vpermt2q %zmm3, %zmm23, %zmm4
22548 ; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm5
22549 ; AVX512BW-NEXT: vpermt2q %zmm1, %zmm23, %zmm5
22550 ; AVX512BW-NEXT: vpblendd {{.*#+}} ymm4 = ymm5[0,1,2,3],ymm4[4,5,6,7]
22551 ; AVX512BW-NEXT: vmovdqa64 %zmm11, %zmm5
22552 ; AVX512BW-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22553 ; AVX512BW-NEXT: vmovdqu64 %zmm22, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22554 ; AVX512BW-NEXT: vpermt2q %zmm22, %zmm23, %zmm5
22555 ; AVX512BW-NEXT: vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22556 ; AVX512BW-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22557 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} zmm5 {%k1} = zmm16[0],zmm13[0],zmm16[2],zmm13[2],zmm16[4],zmm13[4],zmm16[6],zmm13[6]
22558 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm4, %zmm5, %zmm4
22559 ; AVX512BW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22560 ; AVX512BW-NEXT: vmovdqa64 192(%rdi), %zmm4
22561 ; AVX512BW-NEXT: vpermt2q %zmm3, %zmm25, %zmm2
22562 ; AVX512BW-NEXT: vmovdqa64 192(%rdx), %zmm3
22563 ; AVX512BW-NEXT: vpermt2q %zmm1, %zmm25, %zmm0
22564 ; AVX512BW-NEXT: vmovdqa64 192(%rcx), %zmm1
22565 ; AVX512BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
22566 ; AVX512BW-NEXT: vmovdqa64 %zmm11, %zmm2
22567 ; AVX512BW-NEXT: vpermt2q %zmm22, %zmm25, %zmm2
22568 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} zmm2 {%k1} = zmm16[1],zmm13[1],zmm16[3],zmm13[3],zmm16[5],zmm13[5],zmm16[7],zmm13[7]
22569 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm0, %zmm2, %zmm0
22570 ; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22571 ; AVX512BW-NEXT: vmovdqa64 %zmm3, %zmm0
22572 ; AVX512BW-NEXT: vpermt2q %zmm1, %zmm19, %zmm0
22573 ; AVX512BW-NEXT: vmovdqa64 192(%rsi), %zmm2
22574 ; AVX512BW-NEXT: vmovdqa64 %zmm4, %zmm5
22575 ; AVX512BW-NEXT: vpermt2q %zmm2, %zmm19, %zmm5
22576 ; AVX512BW-NEXT: vpblendd {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm0[4,5,6,7]
22577 ; AVX512BW-NEXT: vmovdqa64 192(%r10), %zmm9
22578 ; AVX512BW-NEXT: vmovdqa64 192(%rax), %zmm20
22579 ; AVX512BW-NEXT: vmovdqa64 %zmm9, %zmm6
22580 ; AVX512BW-NEXT: vpermt2q %zmm20, %zmm19, %zmm6
22581 ; AVX512BW-NEXT: vmovdqa64 192(%r8), %zmm8
22582 ; AVX512BW-NEXT: vmovdqa64 192(%r9), %zmm0
22583 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} zmm7 = zmm8[0],zmm0[0],zmm8[2],zmm0[2],zmm8[4],zmm0[4],zmm8[6],zmm0[6]
22584 ; AVX512BW-NEXT: vmovdqa64 %zmm6, %zmm7 {%k1}
22585 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm5, %zmm7, %zmm5
22586 ; AVX512BW-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22587 ; AVX512BW-NEXT: vmovdqa64 %zmm3, %zmm5
22588 ; AVX512BW-NEXT: vpermt2q %zmm1, %zmm21, %zmm5
22589 ; AVX512BW-NEXT: vmovdqa64 %zmm4, %zmm6
22590 ; AVX512BW-NEXT: vpermt2q %zmm2, %zmm21, %zmm6
22591 ; AVX512BW-NEXT: vpblendd {{.*#+}} ymm5 = ymm6[0,1,2,3],ymm5[4,5,6,7]
22592 ; AVX512BW-NEXT: vmovdqa64 %zmm9, %zmm6
22593 ; AVX512BW-NEXT: vpermt2q %zmm20, %zmm21, %zmm6
22594 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} zmm7 = zmm8[1],zmm0[1],zmm8[3],zmm0[3],zmm8[5],zmm0[5],zmm8[7],zmm0[7]
22595 ; AVX512BW-NEXT: vmovdqa64 %zmm6, %zmm7 {%k1}
22596 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm5, %zmm7, %zmm5
22597 ; AVX512BW-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22598 ; AVX512BW-NEXT: vmovdqa64 %zmm3, %zmm5
22599 ; AVX512BW-NEXT: vpermt2q %zmm1, %zmm23, %zmm5
22600 ; AVX512BW-NEXT: vmovdqa64 %zmm4, %zmm6
22601 ; AVX512BW-NEXT: vpermt2q %zmm2, %zmm23, %zmm6
22602 ; AVX512BW-NEXT: vpblendd {{.*#+}} ymm5 = ymm6[0,1,2,3],ymm5[4,5,6,7]
22603 ; AVX512BW-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22604 ; AVX512BW-NEXT: vmovdqa64 %zmm8, %zmm6
22605 ; AVX512BW-NEXT: vpermt2q %zmm0, %zmm23, %zmm6
22606 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} zmm6 {%k1} = zmm9[0],zmm20[0],zmm9[2],zmm20[2],zmm9[4],zmm20[4],zmm9[6],zmm20[6]
22607 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm5, %zmm6, %zmm5
22608 ; AVX512BW-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22609 ; AVX512BW-NEXT: vpermt2q %zmm1, %zmm25, %zmm3
22610 ; AVX512BW-NEXT: vpermt2q %zmm2, %zmm25, %zmm4
22611 ; AVX512BW-NEXT: vpblendd {{.*#+}} ymm1 = ymm4[0,1,2,3],ymm3[4,5,6,7]
22612 ; AVX512BW-NEXT: vmovdqa64 %zmm8, %zmm2
22613 ; AVX512BW-NEXT: vpermt2q %zmm0, %zmm25, %zmm2
22614 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} zmm2 {%k1} = zmm9[1],zmm20[1],zmm9[3],zmm20[3],zmm9[5],zmm20[5],zmm9[7],zmm20[7]
22615 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm1, %zmm2, %zmm1
22616 ; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22617 ; AVX512BW-NEXT: vmovdqa64 256(%rdx), %zmm1
22618 ; AVX512BW-NEXT: vmovdqa64 256(%rcx), %zmm3
22619 ; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm2
22620 ; AVX512BW-NEXT: vpermt2q %zmm3, %zmm19, %zmm2
22621 ; AVX512BW-NEXT: vmovdqa64 256(%rdi), %zmm4
22622 ; AVX512BW-NEXT: vmovdqa64 256(%rsi), %zmm5
22623 ; AVX512BW-NEXT: vmovdqa64 %zmm4, %zmm6
22624 ; AVX512BW-NEXT: vpermt2q %zmm5, %zmm19, %zmm6
22625 ; AVX512BW-NEXT: vpblendd {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm2[4,5,6,7]
22626 ; AVX512BW-NEXT: vmovdqa64 256(%r10), %zmm13
22627 ; AVX512BW-NEXT: vmovdqa64 256(%rax), %zmm2
22628 ; AVX512BW-NEXT: vmovdqa64 %zmm13, %zmm7
22629 ; AVX512BW-NEXT: vpermt2q %zmm2, %zmm19, %zmm7
22630 ; AVX512BW-NEXT: vmovdqa64 256(%r8), %zmm18
22631 ; AVX512BW-NEXT: vmovdqa64 256(%r9), %zmm15
22632 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} zmm8 = zmm18[0],zmm15[0],zmm18[2],zmm15[2],zmm18[4],zmm15[4],zmm18[6],zmm15[6]
22633 ; AVX512BW-NEXT: vmovdqa64 %zmm7, %zmm8 {%k1}
22634 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm6, %zmm8, %zmm6
22635 ; AVX512BW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22636 ; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm6
22637 ; AVX512BW-NEXT: vpermt2q %zmm3, %zmm21, %zmm6
22638 ; AVX512BW-NEXT: vmovdqa64 %zmm4, %zmm7
22639 ; AVX512BW-NEXT: vpermt2q %zmm5, %zmm21, %zmm7
22640 ; AVX512BW-NEXT: vpblendd {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]
22641 ; AVX512BW-NEXT: vmovdqa64 %zmm13, %zmm7
22642 ; AVX512BW-NEXT: vpermt2q %zmm2, %zmm21, %zmm7
22643 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} zmm8 = zmm18[1],zmm15[1],zmm18[3],zmm15[3],zmm18[5],zmm15[5],zmm18[7],zmm15[7]
22644 ; AVX512BW-NEXT: vmovdqa64 %zmm7, %zmm8 {%k1}
22645 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm6, %zmm8, %zmm6
22646 ; AVX512BW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22647 ; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm6
22648 ; AVX512BW-NEXT: vpermt2q %zmm3, %zmm23, %zmm6
22649 ; AVX512BW-NEXT: vmovdqa64 %zmm4, %zmm7
22650 ; AVX512BW-NEXT: vpermt2q %zmm5, %zmm23, %zmm7
22651 ; AVX512BW-NEXT: vpblendd {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]
22652 ; AVX512BW-NEXT: vmovdqa64 %zmm18, %zmm7
22653 ; AVX512BW-NEXT: vpermt2q %zmm15, %zmm23, %zmm7
22654 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} zmm7 {%k1} = zmm13[0],zmm2[0],zmm13[2],zmm2[2],zmm13[4],zmm2[4],zmm13[6],zmm2[6]
22655 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm6, %zmm7, %zmm6
22656 ; AVX512BW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22657 ; AVX512BW-NEXT: vpermt2q %zmm3, %zmm25, %zmm1
22658 ; AVX512BW-NEXT: vpermt2q %zmm5, %zmm25, %zmm4
22659 ; AVX512BW-NEXT: vpblendd {{.*#+}} ymm1 = ymm4[0,1,2,3],ymm1[4,5,6,7]
22660 ; AVX512BW-NEXT: vmovdqa64 %zmm18, %zmm3
22661 ; AVX512BW-NEXT: vpermt2q %zmm15, %zmm25, %zmm3
22662 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} zmm3 {%k1} = zmm13[1],zmm2[1],zmm13[3],zmm2[3],zmm13[5],zmm2[5],zmm13[7],zmm2[7]
22663 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm1, %zmm3, %zmm1
22664 ; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22665 ; AVX512BW-NEXT: vmovdqa64 320(%rdx), %zmm1
22666 ; AVX512BW-NEXT: vmovdqa64 320(%rcx), %zmm4
22667 ; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm3
22668 ; AVX512BW-NEXT: vpermt2q %zmm4, %zmm19, %zmm3
22669 ; AVX512BW-NEXT: vmovdqa64 320(%rdi), %zmm6
22670 ; AVX512BW-NEXT: vmovdqa64 320(%rsi), %zmm7
22671 ; AVX512BW-NEXT: vmovdqa64 %zmm6, %zmm5
22672 ; AVX512BW-NEXT: vpermt2q %zmm7, %zmm19, %zmm5
22673 ; AVX512BW-NEXT: vpblendd {{.*#+}} ymm8 = ymm5[0,1,2,3],ymm3[4,5,6,7]
22674 ; AVX512BW-NEXT: vmovdqa64 320(%r10), %zmm16
22675 ; AVX512BW-NEXT: vmovdqa64 320(%rax), %zmm22
22676 ; AVX512BW-NEXT: vmovdqa64 %zmm16, %zmm12
22677 ; AVX512BW-NEXT: vpermt2q %zmm22, %zmm19, %zmm12
22678 ; AVX512BW-NEXT: vmovdqa64 320(%r8), %zmm29
22679 ; AVX512BW-NEXT: vmovdqa64 320(%r9), %zmm5
22680 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} zmm14 = zmm29[0],zmm5[0],zmm29[2],zmm5[2],zmm29[4],zmm5[4],zmm29[6],zmm5[6]
22681 ; AVX512BW-NEXT: vmovdqa64 %zmm12, %zmm14 {%k1}
22682 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm8, %zmm14, %zmm3
22683 ; AVX512BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22684 ; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm8
22685 ; AVX512BW-NEXT: vpermt2q %zmm4, %zmm21, %zmm8
22686 ; AVX512BW-NEXT: vmovdqa64 %zmm6, %zmm12
22687 ; AVX512BW-NEXT: vpermt2q %zmm7, %zmm21, %zmm12
22688 ; AVX512BW-NEXT: vpblendd {{.*#+}} ymm8 = ymm12[0,1,2,3],ymm8[4,5,6,7]
22689 ; AVX512BW-NEXT: vmovdqa64 %zmm16, %zmm12
22690 ; AVX512BW-NEXT: vpermt2q %zmm22, %zmm21, %zmm12
22691 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} zmm14 = zmm29[1],zmm5[1],zmm29[3],zmm5[3],zmm29[5],zmm5[5],zmm29[7],zmm5[7]
22692 ; AVX512BW-NEXT: vmovdqa64 %zmm12, %zmm14 {%k1}
22693 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm8, %zmm14, %zmm3
22694 ; AVX512BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22695 ; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm8
22696 ; AVX512BW-NEXT: vpermt2q %zmm4, %zmm23, %zmm8
22697 ; AVX512BW-NEXT: vmovdqa64 %zmm6, %zmm12
22698 ; AVX512BW-NEXT: vpermt2q %zmm7, %zmm23, %zmm12
22699 ; AVX512BW-NEXT: vpblendd {{.*#+}} ymm8 = ymm12[0,1,2,3],ymm8[4,5,6,7]
22700 ; AVX512BW-NEXT: vmovdqa64 %zmm29, %zmm12
22701 ; AVX512BW-NEXT: vpermt2q %zmm5, %zmm23, %zmm12
22702 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} zmm12 {%k1} = zmm16[0],zmm22[0],zmm16[2],zmm22[2],zmm16[4],zmm22[4],zmm16[6],zmm22[6]
22703 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm8, %zmm12, %zmm3
22704 ; AVX512BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22705 ; AVX512BW-NEXT: vpermt2q %zmm4, %zmm25, %zmm1
22706 ; AVX512BW-NEXT: vpermt2q %zmm7, %zmm25, %zmm6
22707 ; AVX512BW-NEXT: vpblendd {{.*#+}} ymm1 = ymm6[0,1,2,3],ymm1[4,5,6,7]
22708 ; AVX512BW-NEXT: vmovdqa64 %zmm29, %zmm4
22709 ; AVX512BW-NEXT: vpermt2q %zmm5, %zmm25, %zmm4
22710 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} zmm4 {%k1} = zmm16[1],zmm22[1],zmm16[3],zmm22[3],zmm16[5],zmm22[5],zmm16[7],zmm22[7]
22711 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm1, %zmm4, %zmm1
22712 ; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22713 ; AVX512BW-NEXT: vmovdqa64 384(%rdx), %zmm3
22714 ; AVX512BW-NEXT: vmovdqa64 384(%rcx), %zmm1
22715 ; AVX512BW-NEXT: vmovdqa64 %zmm3, %zmm4
22716 ; AVX512BW-NEXT: vpermt2q %zmm1, %zmm19, %zmm4
22717 ; AVX512BW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22718 ; AVX512BW-NEXT: vmovdqa64 %zmm3, %zmm4
22719 ; AVX512BW-NEXT: vpermt2q %zmm1, %zmm21, %zmm4
22720 ; AVX512BW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22721 ; AVX512BW-NEXT: vmovdqa64 %zmm3, %zmm4
22722 ; AVX512BW-NEXT: vpermt2q %zmm1, %zmm23, %zmm4
22723 ; AVX512BW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22724 ; AVX512BW-NEXT: vpermt2q %zmm1, %zmm25, %zmm3
22725 ; AVX512BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22726 ; AVX512BW-NEXT: vmovdqa64 384(%rdi), %zmm3
22727 ; AVX512BW-NEXT: vmovdqa64 384(%rsi), %zmm1
22728 ; AVX512BW-NEXT: vmovdqa64 %zmm3, %zmm4
22729 ; AVX512BW-NEXT: vpermt2q %zmm1, %zmm19, %zmm4
22730 ; AVX512BW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22731 ; AVX512BW-NEXT: vmovdqa64 %zmm3, %zmm4
22732 ; AVX512BW-NEXT: vpermt2q %zmm1, %zmm21, %zmm4
22733 ; AVX512BW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22734 ; AVX512BW-NEXT: vmovdqa64 %zmm3, %zmm4
22735 ; AVX512BW-NEXT: vpermt2q %zmm1, %zmm23, %zmm4
22736 ; AVX512BW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22737 ; AVX512BW-NEXT: vpermt2q %zmm1, %zmm25, %zmm3
22738 ; AVX512BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22739 ; AVX512BW-NEXT: vmovdqa64 448(%rdx), %zmm3
22740 ; AVX512BW-NEXT: vmovdqa64 448(%rcx), %zmm1
22741 ; AVX512BW-NEXT: vmovdqa64 %zmm3, %zmm4
22742 ; AVX512BW-NEXT: vpermt2q %zmm1, %zmm19, %zmm4
22743 ; AVX512BW-NEXT: vmovdqu64 %zmm4, (%rsp) # 64-byte Spill
22744 ; AVX512BW-NEXT: vmovdqa64 %zmm3, %zmm4
22745 ; AVX512BW-NEXT: vpermt2q %zmm1, %zmm21, %zmm4
22746 ; AVX512BW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22747 ; AVX512BW-NEXT: vmovdqa64 %zmm3, %zmm4
22748 ; AVX512BW-NEXT: vpermt2q %zmm1, %zmm23, %zmm4
22749 ; AVX512BW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22750 ; AVX512BW-NEXT: vpermt2q %zmm1, %zmm25, %zmm3
22751 ; AVX512BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22752 ; AVX512BW-NEXT: vmovdqa64 448(%rdi), %zmm3
22753 ; AVX512BW-NEXT: vmovdqa64 448(%rsi), %zmm1
22754 ; AVX512BW-NEXT: vmovdqa64 %zmm3, %zmm4
22755 ; AVX512BW-NEXT: vpermt2q %zmm1, %zmm19, %zmm4
22756 ; AVX512BW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22757 ; AVX512BW-NEXT: vmovdqa64 %zmm3, %zmm4
22758 ; AVX512BW-NEXT: vpermt2q %zmm1, %zmm21, %zmm4
22759 ; AVX512BW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22760 ; AVX512BW-NEXT: vmovdqa64 %zmm3, %zmm4
22761 ; AVX512BW-NEXT: vpermt2q %zmm1, %zmm23, %zmm4
22762 ; AVX512BW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22763 ; AVX512BW-NEXT: vpermt2q %zmm1, %zmm25, %zmm3
22764 ; AVX512BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22765 ; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm11 = [0,8,0,8,0,8,0,8]
22766 ; AVX512BW-NEXT: # zmm11 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
22767 ; AVX512BW-NEXT: vmovdqa64 %zmm26, %zmm1
22768 ; AVX512BW-NEXT: vpermt2q %zmm17, %zmm11, %zmm1
22769 ; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22770 ; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm12 = [1,9,1,9,1,9,1,9]
22771 ; AVX512BW-NEXT: # zmm12 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
22772 ; AVX512BW-NEXT: vmovdqa64 %zmm26, %zmm1
22773 ; AVX512BW-NEXT: vpermt2q %zmm17, %zmm12, %zmm1
22774 ; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22775 ; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm10 = [2,10,2,10,2,10,2,10]
22776 ; AVX512BW-NEXT: # zmm10 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
22777 ; AVX512BW-NEXT: vmovdqa64 %zmm26, %zmm1
22778 ; AVX512BW-NEXT: vpermt2q %zmm17, %zmm10, %zmm1
22779 ; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22780 ; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm4 = [3,11,3,11,3,11,3,11]
22781 ; AVX512BW-NEXT: # zmm4 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
22782 ; AVX512BW-NEXT: vpermt2q %zmm17, %zmm4, %zmm26
22783 ; AVX512BW-NEXT: vmovdqu64 %zmm26, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22784 ; AVX512BW-NEXT: vmovdqa64 %zmm30, %zmm1
22785 ; AVX512BW-NEXT: vpermt2q %zmm24, %zmm11, %zmm1
22786 ; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22787 ; AVX512BW-NEXT: vmovdqa64 %zmm30, %zmm1
22788 ; AVX512BW-NEXT: vpermt2q %zmm24, %zmm12, %zmm1
22789 ; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22790 ; AVX512BW-NEXT: vmovdqa64 %zmm30, %zmm1
22791 ; AVX512BW-NEXT: vpermt2q %zmm24, %zmm10, %zmm1
22792 ; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22793 ; AVX512BW-NEXT: vpermt2q %zmm24, %zmm4, %zmm30
22794 ; AVX512BW-NEXT: vmovdqu64 %zmm30, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22795 ; AVX512BW-NEXT: vmovdqa64 %zmm31, %zmm1
22796 ; AVX512BW-NEXT: vpermt2q %zmm27, %zmm11, %zmm1
22797 ; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22798 ; AVX512BW-NEXT: vmovdqa64 %zmm31, %zmm1
22799 ; AVX512BW-NEXT: vpermt2q %zmm27, %zmm12, %zmm1
22800 ; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22801 ; AVX512BW-NEXT: vmovdqa64 %zmm31, %zmm1
22802 ; AVX512BW-NEXT: vpermt2q %zmm27, %zmm10, %zmm1
22803 ; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22804 ; AVX512BW-NEXT: vpermt2q %zmm27, %zmm4, %zmm31
22805 ; AVX512BW-NEXT: vmovdqu64 %zmm31, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22806 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
22807 ; AVX512BW-NEXT: vmovdqa64 %zmm3, %zmm1
22808 ; AVX512BW-NEXT: vpermt2q %zmm28, %zmm11, %zmm1
22809 ; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22810 ; AVX512BW-NEXT: vmovdqa64 %zmm3, %zmm1
22811 ; AVX512BW-NEXT: vpermt2q %zmm28, %zmm12, %zmm1
22812 ; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22813 ; AVX512BW-NEXT: vmovdqa64 %zmm3, %zmm31
22814 ; AVX512BW-NEXT: vpermt2q %zmm28, %zmm10, %zmm31
22815 ; AVX512BW-NEXT: vpermt2q %zmm28, %zmm4, %zmm3
22816 ; AVX512BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22817 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
22818 ; AVX512BW-NEXT: vmovdqa64 %zmm28, %zmm3
22819 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
22820 ; AVX512BW-NEXT: vpermt2q %zmm1, %zmm11, %zmm3
22821 ; AVX512BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22822 ; AVX512BW-NEXT: vmovdqa64 %zmm28, %zmm3
22823 ; AVX512BW-NEXT: vpermt2q %zmm1, %zmm12, %zmm3
22824 ; AVX512BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22825 ; AVX512BW-NEXT: vmovdqa64 %zmm28, %zmm3
22826 ; AVX512BW-NEXT: vpermt2q %zmm1, %zmm10, %zmm3
22827 ; AVX512BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22828 ; AVX512BW-NEXT: vpermt2q %zmm1, %zmm4, %zmm28
22829 ; AVX512BW-NEXT: vmovdqu64 %zmm28, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22830 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
22831 ; AVX512BW-NEXT: vmovdqa64 %zmm6, %zmm3
22832 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
22833 ; AVX512BW-NEXT: vpermt2q %zmm1, %zmm11, %zmm3
22834 ; AVX512BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22835 ; AVX512BW-NEXT: vmovdqa64 %zmm6, %zmm3
22836 ; AVX512BW-NEXT: vpermt2q %zmm1, %zmm12, %zmm3
22837 ; AVX512BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22838 ; AVX512BW-NEXT: vmovdqa64 %zmm6, %zmm27
22839 ; AVX512BW-NEXT: vpermt2q %zmm1, %zmm10, %zmm27
22840 ; AVX512BW-NEXT: vpermt2q %zmm1, %zmm4, %zmm6
22841 ; AVX512BW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22842 ; AVX512BW-NEXT: vmovdqa64 %zmm9, %zmm1
22843 ; AVX512BW-NEXT: vpermt2q %zmm20, %zmm11, %zmm1
22844 ; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22845 ; AVX512BW-NEXT: vmovdqa64 %zmm9, %zmm1
22846 ; AVX512BW-NEXT: vpermt2q %zmm20, %zmm12, %zmm1
22847 ; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22848 ; AVX512BW-NEXT: vmovdqa64 %zmm9, %zmm1
22849 ; AVX512BW-NEXT: vpermt2q %zmm20, %zmm10, %zmm1
22850 ; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22851 ; AVX512BW-NEXT: vpermt2q %zmm20, %zmm4, %zmm9
22852 ; AVX512BW-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22853 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
22854 ; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm3
22855 ; AVX512BW-NEXT: vpermt2q %zmm0, %zmm11, %zmm3
22856 ; AVX512BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22857 ; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm3
22858 ; AVX512BW-NEXT: vpermt2q %zmm0, %zmm12, %zmm3
22859 ; AVX512BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22860 ; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm20
22861 ; AVX512BW-NEXT: vpermt2q %zmm0, %zmm10, %zmm20
22862 ; AVX512BW-NEXT: vpermt2q %zmm0, %zmm4, %zmm1
22863 ; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22864 ; AVX512BW-NEXT: vmovdqa64 %zmm13, %zmm0
22865 ; AVX512BW-NEXT: vpermt2q %zmm2, %zmm11, %zmm0
22866 ; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22867 ; AVX512BW-NEXT: vmovdqa64 %zmm13, %zmm0
22868 ; AVX512BW-NEXT: vpermt2q %zmm2, %zmm12, %zmm0
22869 ; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22870 ; AVX512BW-NEXT: vmovdqa64 %zmm13, %zmm0
22871 ; AVX512BW-NEXT: vpermt2q %zmm2, %zmm10, %zmm0
22872 ; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22873 ; AVX512BW-NEXT: vpermt2q %zmm2, %zmm4, %zmm13
22874 ; AVX512BW-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22875 ; AVX512BW-NEXT: vmovdqa64 %zmm18, %zmm0
22876 ; AVX512BW-NEXT: vpermt2q %zmm15, %zmm11, %zmm0
22877 ; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22878 ; AVX512BW-NEXT: vmovdqa64 %zmm18, %zmm0
22879 ; AVX512BW-NEXT: vpermt2q %zmm15, %zmm12, %zmm0
22880 ; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22881 ; AVX512BW-NEXT: vmovdqa64 %zmm18, %zmm28
22882 ; AVX512BW-NEXT: vpermt2q %zmm15, %zmm10, %zmm28
22883 ; AVX512BW-NEXT: vpermt2q %zmm15, %zmm4, %zmm18
22884 ; AVX512BW-NEXT: vmovdqu64 %zmm18, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22885 ; AVX512BW-NEXT: vmovdqa64 %zmm16, %zmm0
22886 ; AVX512BW-NEXT: vpermt2q %zmm22, %zmm11, %zmm0
22887 ; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22888 ; AVX512BW-NEXT: vmovdqa64 %zmm16, %zmm0
22889 ; AVX512BW-NEXT: vpermt2q %zmm22, %zmm12, %zmm0
22890 ; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22891 ; AVX512BW-NEXT: vmovdqa64 %zmm16, %zmm0
22892 ; AVX512BW-NEXT: vpermt2q %zmm22, %zmm10, %zmm0
22893 ; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22894 ; AVX512BW-NEXT: vpermt2q %zmm22, %zmm4, %zmm16
22895 ; AVX512BW-NEXT: vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22896 ; AVX512BW-NEXT: vmovdqa64 %zmm29, %zmm0
22897 ; AVX512BW-NEXT: vpermt2q %zmm5, %zmm11, %zmm0
22898 ; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22899 ; AVX512BW-NEXT: vmovdqa64 %zmm29, %zmm0
22900 ; AVX512BW-NEXT: vpermt2q %zmm5, %zmm12, %zmm0
22901 ; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22902 ; AVX512BW-NEXT: vmovdqa64 %zmm29, %zmm22
22903 ; AVX512BW-NEXT: vpermt2q %zmm5, %zmm10, %zmm22
22904 ; AVX512BW-NEXT: vpermt2q %zmm5, %zmm4, %zmm29
22905 ; AVX512BW-NEXT: vmovdqa64 384(%r10), %zmm13
22906 ; AVX512BW-NEXT: vmovdqa64 384(%rax), %zmm0
22907 ; AVX512BW-NEXT: vmovdqa64 %zmm13, %zmm1
22908 ; AVX512BW-NEXT: vpermt2q %zmm0, %zmm11, %zmm1
22909 ; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22910 ; AVX512BW-NEXT: vmovdqa64 %zmm13, %zmm1
22911 ; AVX512BW-NEXT: vpermt2q %zmm0, %zmm12, %zmm1
22912 ; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22913 ; AVX512BW-NEXT: vmovdqa64 %zmm13, %zmm14
22914 ; AVX512BW-NEXT: vpermt2q %zmm0, %zmm19, %zmm14
22915 ; AVX512BW-NEXT: vmovdqa64 %zmm13, %zmm15
22916 ; AVX512BW-NEXT: vpermt2q %zmm0, %zmm21, %zmm15
22917 ; AVX512BW-NEXT: vmovdqa64 384(%r8), %zmm5
22918 ; AVX512BW-NEXT: vmovdqa64 384(%r9), %zmm2
22919 ; AVX512BW-NEXT: vmovdqa64 %zmm5, %zmm8
22920 ; AVX512BW-NEXT: vpermt2q %zmm2, %zmm23, %zmm8
22921 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} zmm8 {%k1} = zmm13[0],zmm0[0],zmm13[2],zmm0[2],zmm13[4],zmm0[4],zmm13[6],zmm0[6]
22922 ; AVX512BW-NEXT: vmovdqa64 %zmm5, %zmm6
22923 ; AVX512BW-NEXT: vpermt2q %zmm2, %zmm25, %zmm6
22924 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} zmm6 {%k1} = zmm13[1],zmm0[1],zmm13[3],zmm0[3],zmm13[5],zmm0[5],zmm13[7],zmm0[7]
22925 ; AVX512BW-NEXT: vmovdqa64 %zmm13, %zmm30
22926 ; AVX512BW-NEXT: vpermt2q %zmm0, %zmm10, %zmm30
22927 ; AVX512BW-NEXT: vpermt2q %zmm0, %zmm4, %zmm13
22928 ; AVX512BW-NEXT: vmovdqa64 %zmm5, %zmm24
22929 ; AVX512BW-NEXT: vpermt2q %zmm2, %zmm11, %zmm24
22930 ; AVX512BW-NEXT: vmovdqa64 %zmm5, %zmm26
22931 ; AVX512BW-NEXT: vpermt2q %zmm2, %zmm12, %zmm26
22932 ; AVX512BW-NEXT: vmovdqa64 %zmm5, %zmm17
22933 ; AVX512BW-NEXT: vpermt2q %zmm2, %zmm10, %zmm17
22934 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} zmm7 = zmm5[0],zmm2[0],zmm5[2],zmm2[2],zmm5[4],zmm2[4],zmm5[6],zmm2[6]
22935 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} zmm9 = zmm5[1],zmm2[1],zmm5[3],zmm2[3],zmm5[5],zmm2[5],zmm5[7],zmm2[7]
22936 ; AVX512BW-NEXT: vpermt2q %zmm2, %zmm4, %zmm5
22937 ; AVX512BW-NEXT: vmovdqa64 448(%r10), %zmm16
22938 ; AVX512BW-NEXT: vmovdqa64 448(%rax), %zmm1
22939 ; AVX512BW-NEXT: vmovdqa64 %zmm16, %zmm0
22940 ; AVX512BW-NEXT: vpermt2q %zmm1, %zmm11, %zmm0
22941 ; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22942 ; AVX512BW-NEXT: vmovdqa64 %zmm16, %zmm0
22943 ; AVX512BW-NEXT: vpermt2q %zmm1, %zmm12, %zmm0
22944 ; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22945 ; AVX512BW-NEXT: vpermi2q %zmm1, %zmm16, %zmm19
22946 ; AVX512BW-NEXT: vpermi2q %zmm1, %zmm16, %zmm21
22947 ; AVX512BW-NEXT: vmovdqa64 448(%r8), %zmm3
22948 ; AVX512BW-NEXT: vmovdqa64 448(%r9), %zmm0
22949 ; AVX512BW-NEXT: vpermi2q %zmm0, %zmm3, %zmm23
22950 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} zmm23 {%k1} = zmm16[0],zmm1[0],zmm16[2],zmm1[2],zmm16[4],zmm1[4],zmm16[6],zmm1[6]
22951 ; AVX512BW-NEXT: vpermi2q %zmm0, %zmm3, %zmm25
22952 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} zmm25 {%k1} = zmm16[1],zmm1[1],zmm16[3],zmm1[3],zmm16[5],zmm1[5],zmm16[7],zmm1[7]
22953 ; AVX512BW-NEXT: vmovdqa64 %zmm16, %zmm18
22954 ; AVX512BW-NEXT: vpermt2q %zmm1, %zmm10, %zmm18
22955 ; AVX512BW-NEXT: vpermt2q %zmm1, %zmm4, %zmm16
22956 ; AVX512BW-NEXT: vpermi2q %zmm0, %zmm3, %zmm11
22957 ; AVX512BW-NEXT: vpermi2q %zmm0, %zmm3, %zmm12
22958 ; AVX512BW-NEXT: vpermi2q %zmm0, %zmm3, %zmm10
22959 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} zmm1 = zmm3[0],zmm0[0],zmm3[2],zmm0[2],zmm3[4],zmm0[4],zmm3[6],zmm0[6]
22960 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} zmm2 = zmm3[1],zmm0[1],zmm3[3],zmm0[3],zmm3[5],zmm0[5],zmm3[7],zmm0[7]
22961 ; AVX512BW-NEXT: vpermt2q %zmm0, %zmm4, %zmm3
22962 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
22963 ; AVX512BW-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
22964 ; AVX512BW-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
22965 ; AVX512BW-NEXT: vmovdqa64 %zmm14, %zmm7 {%k1}
22966 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm0, %zmm7, %zmm0
22967 ; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22968 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
22969 ; AVX512BW-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
22970 ; AVX512BW-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
22971 ; AVX512BW-NEXT: vmovdqa64 %zmm15, %zmm9 {%k1}
22972 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm0, %zmm9, %zmm0
22973 ; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22974 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
22975 ; AVX512BW-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
22976 ; AVX512BW-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
22977 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm0, %zmm8, %zmm0
22978 ; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22979 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
22980 ; AVX512BW-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
22981 ; AVX512BW-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
22982 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm0, %zmm6, %zmm0
22983 ; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22984 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
22985 ; AVX512BW-NEXT: vpblendd $240, (%rsp), %ymm0, %ymm0 # 32-byte Folded Reload
22986 ; AVX512BW-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
22987 ; AVX512BW-NEXT: vmovdqa64 %zmm19, %zmm1 {%k1}
22988 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
22989 ; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22990 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
22991 ; AVX512BW-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
22992 ; AVX512BW-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
22993 ; AVX512BW-NEXT: vmovdqa64 %zmm21, %zmm2 {%k1}
22994 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm0, %zmm2, %zmm0
22995 ; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
22996 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
22997 ; AVX512BW-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
22998 ; AVX512BW-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
22999 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm0, %zmm23, %zmm0
23000 ; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23001 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
23002 ; AVX512BW-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
23003 ; AVX512BW-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
23004 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm0, %zmm25, %zmm0
23005 ; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23006 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
23007 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
23008 ; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm8 {%k1}
23009 ; AVX512BW-NEXT: vmovdqa (%rcx), %ymm0
23010 ; AVX512BW-NEXT: vmovdqa (%rdx), %ymm1
23011 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
23012 ; AVX512BW-NEXT: vmovdqa (%rsi), %ymm4
23013 ; AVX512BW-NEXT: vmovdqa (%rdi), %ymm6
23014 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm6[0],ymm4[0],ymm6[2],ymm4[2]
23015 ; AVX512BW-NEXT: vperm2i128 {{.*#+}} ymm2 = ymm7[2,3],ymm2[2,3]
23016 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm2, %zmm8, %zmm2
23017 ; AVX512BW-NEXT: vmovdqu64 %zmm2, (%rsp) # 64-byte Spill
23018 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
23019 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
23020 ; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm7 {%k1}
23021 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
23022 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm6[1],ymm4[1],ymm6[3],ymm4[3]
23023 ; AVX512BW-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
23024 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm0, %zmm7, %zmm0
23025 ; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23026 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
23027 ; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm31 {%k1}
23028 ; AVX512BW-NEXT: vmovdqa 64(%rcx), %ymm0
23029 ; AVX512BW-NEXT: vmovdqa 64(%rdx), %ymm1
23030 ; AVX512BW-NEXT: vmovdqa 64(%rsi), %ymm2
23031 ; AVX512BW-NEXT: vmovdqa 64(%rdi), %ymm4
23032 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} ymm6 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
23033 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm4[0],ymm2[0],ymm4[2],ymm2[2]
23034 ; AVX512BW-NEXT: vperm2i128 {{.*#+}} ymm6 = ymm7[2,3],ymm6[2,3]
23035 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm6, %zmm31, %zmm6
23036 ; AVX512BW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23037 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
23038 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
23039 ; AVX512BW-NEXT: vmovdqa64 %zmm6, %zmm7 {%k1}
23040 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
23041 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm4[1],ymm2[1],ymm4[3],ymm2[3]
23042 ; AVX512BW-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
23043 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm0, %zmm7, %zmm8
23044 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
23045 ; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm27 {%k1}
23046 ; AVX512BW-NEXT: vmovdqa 128(%rcx), %ymm0
23047 ; AVX512BW-NEXT: vmovdqa 128(%rdx), %ymm1
23048 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
23049 ; AVX512BW-NEXT: vmovdqa 128(%rsi), %ymm4
23050 ; AVX512BW-NEXT: vmovdqa 128(%rdi), %ymm6
23051 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm6[0],ymm4[0],ymm6[2],ymm4[2]
23052 ; AVX512BW-NEXT: vperm2i128 {{.*#+}} ymm2 = ymm7[2,3],ymm2[2,3]
23053 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm2, %zmm27, %zmm23
23054 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
23055 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
23056 ; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm7 {%k1}
23057 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
23058 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm6[1],ymm4[1],ymm6[3],ymm4[3]
23059 ; AVX512BW-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
23060 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm0, %zmm7, %zmm25
23061 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
23062 ; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm20 {%k1}
23063 ; AVX512BW-NEXT: vmovdqa 192(%rcx), %ymm0
23064 ; AVX512BW-NEXT: vmovdqa 192(%rdx), %ymm1
23065 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
23066 ; AVX512BW-NEXT: vmovdqa 192(%rsi), %ymm4
23067 ; AVX512BW-NEXT: vmovdqa 192(%rdi), %ymm6
23068 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm6[0],ymm4[0],ymm6[2],ymm4[2]
23069 ; AVX512BW-NEXT: vperm2i128 {{.*#+}} ymm2 = ymm7[2,3],ymm2[2,3]
23070 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm2, %zmm20, %zmm21
23071 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
23072 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
23073 ; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm7 {%k1}
23074 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
23075 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm6[1],ymm4[1],ymm6[3],ymm4[3]
23076 ; AVX512BW-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
23077 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm0, %zmm7, %zmm19
23078 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
23079 ; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm28 {%k1}
23080 ; AVX512BW-NEXT: vmovdqa 256(%rcx), %ymm0
23081 ; AVX512BW-NEXT: vmovdqa 256(%rdx), %ymm1
23082 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
23083 ; AVX512BW-NEXT: vmovdqa 256(%rsi), %ymm4
23084 ; AVX512BW-NEXT: vmovdqa 256(%rdi), %ymm6
23085 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm6[0],ymm4[0],ymm6[2],ymm4[2]
23086 ; AVX512BW-NEXT: vperm2i128 {{.*#+}} ymm2 = ymm7[2,3],ymm2[2,3]
23087 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm2, %zmm28, %zmm15
23088 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
23089 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
23090 ; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm7 {%k1}
23091 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
23092 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm6[1],ymm4[1],ymm6[3],ymm4[3]
23093 ; AVX512BW-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
23094 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm0, %zmm7, %zmm28
23095 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
23096 ; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm22 {%k1}
23097 ; AVX512BW-NEXT: vmovdqa 320(%rcx), %ymm0
23098 ; AVX512BW-NEXT: vmovdqa 320(%rdx), %ymm1
23099 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
23100 ; AVX512BW-NEXT: vmovdqa 320(%rsi), %ymm4
23101 ; AVX512BW-NEXT: vmovdqa 320(%rdi), %ymm6
23102 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm6[0],ymm4[0],ymm6[2],ymm4[2]
23103 ; AVX512BW-NEXT: vperm2i128 {{.*#+}} ymm2 = ymm7[2,3],ymm2[2,3]
23104 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm2, %zmm22, %zmm22
23105 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
23106 ; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm29 {%k1}
23107 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
23108 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm6[1],ymm4[1],ymm6[3],ymm4[3]
23109 ; AVX512BW-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
23110 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm0, %zmm29, %zmm27
23111 ; AVX512BW-NEXT: vmovdqa64 %zmm30, %zmm17 {%k1}
23112 ; AVX512BW-NEXT: vmovdqa 384(%rcx), %ymm0
23113 ; AVX512BW-NEXT: vmovdqa 384(%rdx), %ymm1
23114 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
23115 ; AVX512BW-NEXT: vmovdqa 384(%rsi), %ymm4
23116 ; AVX512BW-NEXT: vmovdqa 384(%rdi), %ymm6
23117 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm6[0],ymm4[0],ymm6[2],ymm4[2]
23118 ; AVX512BW-NEXT: vperm2i128 {{.*#+}} ymm2 = ymm7[2,3],ymm2[2,3]
23119 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm2, %zmm17, %zmm14
23120 ; AVX512BW-NEXT: vmovdqa64 %zmm13, %zmm5 {%k1}
23121 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
23122 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm6[1],ymm4[1],ymm6[3],ymm4[3]
23123 ; AVX512BW-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
23124 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm0, %zmm5, %zmm13
23125 ; AVX512BW-NEXT: vmovdqa64 %zmm18, %zmm10 {%k1}
23126 ; AVX512BW-NEXT: vmovdqa 448(%rcx), %ymm0
23127 ; AVX512BW-NEXT: vmovdqa 448(%rdx), %ymm1
23128 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
23129 ; AVX512BW-NEXT: vmovdqa 448(%rsi), %ymm4
23130 ; AVX512BW-NEXT: vmovdqa 448(%rdi), %ymm5
23131 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} ymm6 = ymm5[0],ymm4[0],ymm5[2],ymm4[2]
23132 ; AVX512BW-NEXT: vperm2i128 {{.*#+}} ymm2 = ymm6[2,3],ymm2[2,3]
23133 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm2, %zmm10, %zmm10
23134 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
23135 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm5[1],ymm4[1],ymm5[3],ymm4[3]
23136 ; AVX512BW-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
23137 ; AVX512BW-NEXT: vmovdqa64 %zmm16, %zmm3 {%k1}
23138 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm0, %zmm3, %zmm9
23139 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
23140 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
23141 ; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm3 {%k1}
23142 ; AVX512BW-NEXT: vmovdqa (%rsi), %xmm0
23143 ; AVX512BW-NEXT: vinserti128 $1, (%rcx), %ymm0, %ymm1
23144 ; AVX512BW-NEXT: vmovdqa (%rdi), %xmm0
23145 ; AVX512BW-NEXT: vinserti128 $1, (%rdx), %ymm0, %ymm2
23146 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm2[0],ymm1[0],ymm2[2],ymm1[2]
23147 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm0, %zmm3, %zmm0
23148 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
23149 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
23150 ; AVX512BW-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
23151 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm2[1],ymm1[1],ymm2[3],ymm1[3]
23152 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm1, %zmm4, %zmm2
23153 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
23154 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
23155 ; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm5 {%k1}
23156 ; AVX512BW-NEXT: vmovdqa 64(%rsi), %xmm1
23157 ; AVX512BW-NEXT: vinserti128 $1, 64(%rcx), %ymm1, %ymm1
23158 ; AVX512BW-NEXT: vmovdqa 64(%rdi), %xmm3
23159 ; AVX512BW-NEXT: vinserti128 $1, 64(%rdx), %ymm3, %ymm4
23160 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm4[0],ymm1[0],ymm4[2],ymm1[2]
23161 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm3, %zmm5, %zmm3
23162 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
23163 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
23164 ; AVX512BW-NEXT: vmovdqa64 %zmm5, %zmm6 {%k1}
23165 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm4[1],ymm1[1],ymm4[3],ymm1[3]
23166 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm1, %zmm6, %zmm5
23167 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
23168 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
23169 ; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm7 {%k1}
23170 ; AVX512BW-NEXT: vmovdqa 128(%rsi), %xmm1
23171 ; AVX512BW-NEXT: vinserti128 $1, 128(%rcx), %ymm1, %ymm1
23172 ; AVX512BW-NEXT: vmovdqa 128(%rdi), %xmm4
23173 ; AVX512BW-NEXT: vinserti128 $1, 128(%rdx), %ymm4, %ymm6
23174 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} ymm4 = ymm6[0],ymm1[0],ymm6[2],ymm1[2]
23175 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm4, %zmm7, %zmm4
23176 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
23177 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
23178 ; AVX512BW-NEXT: vmovdqa64 %zmm7, %zmm16 {%k1}
23179 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm6[1],ymm1[1],ymm6[3],ymm1[3]
23180 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm1, %zmm16, %zmm6
23181 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
23182 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
23183 ; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm17 {%k1}
23184 ; AVX512BW-NEXT: vmovdqa 192(%rsi), %xmm1
23185 ; AVX512BW-NEXT: vinserti128 $1, 192(%rcx), %ymm1, %ymm1
23186 ; AVX512BW-NEXT: vmovdqa 192(%rdi), %xmm7
23187 ; AVX512BW-NEXT: vinserti32x4 $1, 192(%rdx), %ymm7, %ymm16
23188 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm16[0],ymm1[0],ymm16[2],ymm1[2]
23189 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm7, %zmm17, %zmm7
23190 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
23191 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
23192 ; AVX512BW-NEXT: vmovdqa64 %zmm17, %zmm18 {%k1}
23193 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm16[1],ymm1[1],ymm16[3],ymm1[3]
23194 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm1, %zmm18, %zmm1
23195 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
23196 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
23197 ; AVX512BW-NEXT: vmovdqa64 %zmm16, %zmm20 {%k1}
23198 ; AVX512BW-NEXT: vmovdqa64 256(%rsi), %xmm16
23199 ; AVX512BW-NEXT: vinserti32x4 $1, 256(%rcx), %ymm16, %ymm16
23200 ; AVX512BW-NEXT: vmovdqa64 256(%rdi), %xmm17
23201 ; AVX512BW-NEXT: vinserti32x4 $1, 256(%rdx), %ymm17, %ymm17
23202 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} ymm18 = ymm17[0],ymm16[0],ymm17[2],ymm16[2]
23203 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm18, %zmm20, %zmm29
23204 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
23205 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
23206 ; AVX512BW-NEXT: vmovdqa64 %zmm18, %zmm20 {%k1}
23207 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} ymm16 = ymm17[1],ymm16[1],ymm17[3],ymm16[3]
23208 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm16, %zmm20, %zmm30
23209 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
23210 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
23211 ; AVX512BW-NEXT: vmovdqa64 %zmm16, %zmm20 {%k1}
23212 ; AVX512BW-NEXT: vmovdqa64 320(%rsi), %xmm16
23213 ; AVX512BW-NEXT: vinserti32x4 $1, 320(%rcx), %ymm16, %ymm16
23214 ; AVX512BW-NEXT: vmovdqa64 320(%rdi), %xmm17
23215 ; AVX512BW-NEXT: vinserti32x4 $1, 320(%rdx), %ymm17, %ymm17
23216 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} ymm18 = ymm17[0],ymm16[0],ymm17[2],ymm16[2]
23217 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm18, %zmm20, %zmm18
23218 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
23219 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
23220 ; AVX512BW-NEXT: vmovdqa64 %zmm20, %zmm31 {%k1}
23221 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} ymm16 = ymm17[1],ymm16[1],ymm17[3],ymm16[3]
23222 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm16, %zmm31, %zmm16
23223 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
23224 ; AVX512BW-NEXT: vmovdqa64 %zmm17, %zmm24 {%k1}
23225 ; AVX512BW-NEXT: vmovdqa64 384(%rsi), %xmm17
23226 ; AVX512BW-NEXT: vinserti32x4 $1, 384(%rcx), %ymm17, %ymm17
23227 ; AVX512BW-NEXT: vmovdqa64 384(%rdi), %xmm20
23228 ; AVX512BW-NEXT: vinserti32x4 $1, 384(%rdx), %ymm20, %ymm20
23229 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} ymm31 = ymm20[0],ymm17[0],ymm20[2],ymm17[2]
23230 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm31, %zmm24, %zmm24
23231 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
23232 ; AVX512BW-NEXT: vmovdqa64 %zmm31, %zmm26 {%k1}
23233 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} ymm17 = ymm20[1],ymm17[1],ymm20[3],ymm17[3]
23234 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm17, %zmm26, %zmm17
23235 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
23236 ; AVX512BW-NEXT: vmovdqa64 %zmm20, %zmm11 {%k1}
23237 ; AVX512BW-NEXT: vmovdqa64 448(%rsi), %xmm20
23238 ; AVX512BW-NEXT: vinserti32x4 $1, 448(%rcx), %ymm20, %ymm20
23239 ; AVX512BW-NEXT: vmovdqa64 448(%rdi), %xmm26
23240 ; AVX512BW-NEXT: vinserti32x4 $1, 448(%rdx), %ymm26, %ymm26
23241 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} ymm31 = ymm26[0],ymm20[0],ymm26[2],ymm20[2]
23242 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm31, %zmm11, %zmm11
23243 ; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
23244 ; AVX512BW-NEXT: vmovdqa64 %zmm31, %zmm12 {%k1}
23245 ; AVX512BW-NEXT: vpunpckhqdq {{.*#+}} ymm20 = ymm26[1],ymm20[1],ymm26[3],ymm20[3]
23246 ; AVX512BW-NEXT: vinserti64x4 $0, %ymm20, %zmm12, %zmm12
23247 ; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
23248 ; AVX512BW-NEXT: vmovdqa64 %zmm9, 3776(%rax)
23249 ; AVX512BW-NEXT: vmovdqa64 %zmm10, 3712(%rax)
23250 ; AVX512BW-NEXT: vmovdqa64 %zmm13, 3264(%rax)
23251 ; AVX512BW-NEXT: vmovdqa64 %zmm14, 3200(%rax)
23252 ; AVX512BW-NEXT: vmovdqa64 %zmm27, 2752(%rax)
23253 ; AVX512BW-NEXT: vmovdqa64 %zmm22, 2688(%rax)
23254 ; AVX512BW-NEXT: vmovdqa64 %zmm28, 2240(%rax)
23255 ; AVX512BW-NEXT: vmovdqa64 %zmm15, 2176(%rax)
23256 ; AVX512BW-NEXT: vmovdqa64 %zmm19, 1728(%rax)
23257 ; AVX512BW-NEXT: vmovdqa64 %zmm21, 1664(%rax)
23258 ; AVX512BW-NEXT: vmovdqa64 %zmm25, 1216(%rax)
23259 ; AVX512BW-NEXT: vmovdqa64 %zmm23, 1152(%rax)
23260 ; AVX512BW-NEXT: vmovdqa64 %zmm8, 704(%rax)
23261 ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
23262 ; AVX512BW-NEXT: vmovaps %zmm8, 640(%rax)
23263 ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
23264 ; AVX512BW-NEXT: vmovaps %zmm8, 192(%rax)
23265 ; AVX512BW-NEXT: vmovups (%rsp), %zmm8 # 64-byte Reload
23266 ; AVX512BW-NEXT: vmovaps %zmm8, 128(%rax)
23267 ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
23268 ; AVX512BW-NEXT: vmovaps %zmm8, 4032(%rax)
23269 ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
23270 ; AVX512BW-NEXT: vmovaps %zmm8, 3968(%rax)
23271 ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
23272 ; AVX512BW-NEXT: vmovaps %zmm8, 3904(%rax)
23273 ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
23274 ; AVX512BW-NEXT: vmovaps %zmm8, 3840(%rax)
23275 ; AVX512BW-NEXT: vmovdqa64 %zmm12, 3648(%rax)
23276 ; AVX512BW-NEXT: vmovdqa64 %zmm11, 3584(%rax)
23277 ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
23278 ; AVX512BW-NEXT: vmovaps %zmm8, 3520(%rax)
23279 ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
23280 ; AVX512BW-NEXT: vmovaps %zmm8, 3456(%rax)
23281 ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
23282 ; AVX512BW-NEXT: vmovaps %zmm8, 3392(%rax)
23283 ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
23284 ; AVX512BW-NEXT: vmovaps %zmm8, 3328(%rax)
23285 ; AVX512BW-NEXT: vmovdqa64 %zmm17, 3136(%rax)
23286 ; AVX512BW-NEXT: vmovdqa64 %zmm24, 3072(%rax)
23287 ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
23288 ; AVX512BW-NEXT: vmovaps %zmm8, 3008(%rax)
23289 ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
23290 ; AVX512BW-NEXT: vmovaps %zmm8, 2944(%rax)
23291 ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
23292 ; AVX512BW-NEXT: vmovaps %zmm8, 2880(%rax)
23293 ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
23294 ; AVX512BW-NEXT: vmovaps %zmm8, 2816(%rax)
23295 ; AVX512BW-NEXT: vmovdqa64 %zmm16, 2624(%rax)
23296 ; AVX512BW-NEXT: vmovdqa64 %zmm18, 2560(%rax)
23297 ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
23298 ; AVX512BW-NEXT: vmovaps %zmm8, 2496(%rax)
23299 ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
23300 ; AVX512BW-NEXT: vmovaps %zmm8, 2432(%rax)
23301 ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
23302 ; AVX512BW-NEXT: vmovaps %zmm8, 2368(%rax)
23303 ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
23304 ; AVX512BW-NEXT: vmovaps %zmm8, 2304(%rax)
23305 ; AVX512BW-NEXT: vmovdqa64 %zmm30, 2112(%rax)
23306 ; AVX512BW-NEXT: vmovdqa64 %zmm29, 2048(%rax)
23307 ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
23308 ; AVX512BW-NEXT: vmovaps %zmm8, 1984(%rax)
23309 ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
23310 ; AVX512BW-NEXT: vmovaps %zmm8, 1920(%rax)
23311 ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
23312 ; AVX512BW-NEXT: vmovaps %zmm8, 1856(%rax)
23313 ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
23314 ; AVX512BW-NEXT: vmovaps %zmm8, 1792(%rax)
23315 ; AVX512BW-NEXT: vmovdqa64 %zmm1, 1600(%rax)
23316 ; AVX512BW-NEXT: vmovdqa64 %zmm7, 1536(%rax)
23317 ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
23318 ; AVX512BW-NEXT: vmovaps %zmm1, 1472(%rax)
23319 ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
23320 ; AVX512BW-NEXT: vmovaps %zmm1, 1408(%rax)
23321 ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
23322 ; AVX512BW-NEXT: vmovaps %zmm1, 1344(%rax)
23323 ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
23324 ; AVX512BW-NEXT: vmovaps %zmm1, 1280(%rax)
23325 ; AVX512BW-NEXT: vmovdqa64 %zmm6, 1088(%rax)
23326 ; AVX512BW-NEXT: vmovdqa64 %zmm4, 1024(%rax)
23327 ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
23328 ; AVX512BW-NEXT: vmovaps %zmm1, 960(%rax)
23329 ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
23330 ; AVX512BW-NEXT: vmovaps %zmm1, 896(%rax)
23331 ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
23332 ; AVX512BW-NEXT: vmovaps %zmm1, 832(%rax)
23333 ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
23334 ; AVX512BW-NEXT: vmovaps %zmm1, 768(%rax)
23335 ; AVX512BW-NEXT: vmovdqa64 %zmm5, 576(%rax)
23336 ; AVX512BW-NEXT: vmovdqa64 %zmm3, 512(%rax)
23337 ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
23338 ; AVX512BW-NEXT: vmovaps %zmm1, 448(%rax)
23339 ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
23340 ; AVX512BW-NEXT: vmovaps %zmm1, 384(%rax)
23341 ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
23342 ; AVX512BW-NEXT: vmovaps %zmm1, 320(%rax)
23343 ; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
23344 ; AVX512BW-NEXT: vmovaps %zmm1, 256(%rax)
23345 ; AVX512BW-NEXT: vmovdqa64 %zmm2, 64(%rax)
23346 ; AVX512BW-NEXT: vmovdqa64 %zmm0, (%rax)
23347 ; AVX512BW-NEXT: addq $5384, %rsp # imm = 0x1508
23348 ; AVX512BW-NEXT: vzeroupper
23349 ; AVX512BW-NEXT: retq
23350 ;
23351 ; AVX512BW-FCP-LABEL: store_i64_stride8_vf64:
23352 ; AVX512BW-FCP: # %bb.0:
23353 ; AVX512BW-FCP-NEXT: subq $5384, %rsp # imm = 0x1508
23354 ; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
23355 ; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10
23356 ; AVX512BW-FCP-NEXT: vmovdqa64 128(%rdi), %zmm0
23357 ; AVX512BW-FCP-NEXT: vmovdqa64 64(%rdi), %zmm4
23358 ; AVX512BW-FCP-NEXT: vmovdqa64 (%rdi), %zmm14
23359 ; AVX512BW-FCP-NEXT: vmovdqa64 128(%rsi), %zmm1
23360 ; AVX512BW-FCP-NEXT: vmovdqa64 64(%rsi), %zmm5
23361 ; AVX512BW-FCP-NEXT: vmovdqa64 (%rsi), %zmm10
23362 ; AVX512BW-FCP-NEXT: vmovdqa64 128(%rdx), %zmm2
23363 ; AVX512BW-FCP-NEXT: vmovdqa64 64(%rdx), %zmm7
23364 ; AVX512BW-FCP-NEXT: vmovdqa64 (%rdx), %zmm12
23365 ; AVX512BW-FCP-NEXT: vmovdqa64 128(%rcx), %zmm3
23366 ; AVX512BW-FCP-NEXT: vmovdqa64 64(%rcx), %zmm9
23367 ; AVX512BW-FCP-NEXT: vmovdqa64 (%rcx), %zmm15
23368 ; AVX512BW-FCP-NEXT: vmovdqa64 (%r8), %zmm30
23369 ; AVX512BW-FCP-NEXT: vmovdqa64 64(%r8), %zmm18
23370 ; AVX512BW-FCP-NEXT: vmovdqa64 128(%r8), %zmm11
23371 ; AVX512BW-FCP-NEXT: vmovdqa64 (%r9), %zmm24
23372 ; AVX512BW-FCP-NEXT: vmovdqa64 64(%r9), %zmm28
23373 ; AVX512BW-FCP-NEXT: vmovdqa64 128(%r9), %zmm22
23374 ; AVX512BW-FCP-NEXT: vmovdqa64 (%r10), %zmm26
23375 ; AVX512BW-FCP-NEXT: vmovdqa64 64(%r10), %zmm31
23376 ; AVX512BW-FCP-NEXT: vmovdqa64 128(%r10), %zmm16
23377 ; AVX512BW-FCP-NEXT: vmovdqa64 (%rax), %zmm17
23378 ; AVX512BW-FCP-NEXT: vmovdqa64 64(%rax), %zmm27
23379 ; AVX512BW-FCP-NEXT: vmovdqa64 128(%rax), %zmm13
23380 ; AVX512BW-FCP-NEXT: movb $-64, %r11b
23381 ; AVX512BW-FCP-NEXT: kmovd %r11d, %k1
23382 ; AVX512BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm19 = [4,12,4,12,4,12,4,12]
23383 ; AVX512BW-FCP-NEXT: # zmm19 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
23384 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm12, %zmm8
23385 ; AVX512BW-FCP-NEXT: vpermt2q %zmm15, %zmm19, %zmm8
23386 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm14, %zmm6
23387 ; AVX512BW-FCP-NEXT: vpermt2q %zmm10, %zmm19, %zmm6
23388 ; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm8[4,5,6,7]
23389 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm26, %zmm8
23390 ; AVX512BW-FCP-NEXT: vpermt2q %zmm17, %zmm19, %zmm8
23391 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm20 = zmm30[0],zmm24[0],zmm30[2],zmm24[2],zmm30[4],zmm24[4],zmm30[6],zmm24[6]
23392 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm8, %zmm20 {%k1}
23393 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm20, %zmm6
23394 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23395 ; AVX512BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm21 = [5,13,5,13,5,13,5,13]
23396 ; AVX512BW-FCP-NEXT: # zmm21 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
23397 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm12, %zmm6
23398 ; AVX512BW-FCP-NEXT: vpermt2q %zmm15, %zmm21, %zmm6
23399 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm14, %zmm8
23400 ; AVX512BW-FCP-NEXT: vpermt2q %zmm10, %zmm21, %zmm8
23401 ; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm8[0,1,2,3],ymm6[4,5,6,7]
23402 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm26, %zmm8
23403 ; AVX512BW-FCP-NEXT: vpermt2q %zmm17, %zmm21, %zmm8
23404 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm20 = zmm30[1],zmm24[1],zmm30[3],zmm24[3],zmm30[5],zmm24[5],zmm30[7],zmm24[7]
23405 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm8, %zmm20 {%k1}
23406 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm20, %zmm6
23407 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23408 ; AVX512BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm23 = [6,14,6,14,6,14,6,14]
23409 ; AVX512BW-FCP-NEXT: # zmm23 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
23410 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm12, %zmm6
23411 ; AVX512BW-FCP-NEXT: vpermt2q %zmm15, %zmm23, %zmm6
23412 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm14, %zmm8
23413 ; AVX512BW-FCP-NEXT: vpermt2q %zmm10, %zmm23, %zmm8
23414 ; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm8[0,1,2,3],ymm6[4,5,6,7]
23415 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm30, %zmm8
23416 ; AVX512BW-FCP-NEXT: vpermt2q %zmm24, %zmm23, %zmm8
23417 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm8 {%k1} = zmm26[0],zmm17[0],zmm26[2],zmm17[2],zmm26[4],zmm17[4],zmm26[6],zmm17[6]
23418 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm8, %zmm6
23419 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23420 ; AVX512BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm25 = [7,15,7,15,7,15,7,15]
23421 ; AVX512BW-FCP-NEXT: # zmm25 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
23422 ; AVX512BW-FCP-NEXT: vpermt2q %zmm15, %zmm25, %zmm12
23423 ; AVX512BW-FCP-NEXT: vpermt2q %zmm10, %zmm25, %zmm14
23424 ; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm14[0,1,2,3],ymm12[4,5,6,7]
23425 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm30, %zmm8
23426 ; AVX512BW-FCP-NEXT: vpermt2q %zmm24, %zmm25, %zmm8
23427 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm8 {%k1} = zmm26[1],zmm17[1],zmm26[3],zmm17[3],zmm26[5],zmm17[5],zmm26[7],zmm17[7]
23428 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm8, %zmm6
23429 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23430 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm7, %zmm6
23431 ; AVX512BW-FCP-NEXT: vpermt2q %zmm9, %zmm19, %zmm6
23432 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm4, %zmm8
23433 ; AVX512BW-FCP-NEXT: vpermt2q %zmm5, %zmm19, %zmm8
23434 ; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm8[0,1,2,3],ymm6[4,5,6,7]
23435 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm31, %zmm8
23436 ; AVX512BW-FCP-NEXT: vpermt2q %zmm27, %zmm19, %zmm8
23437 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm10 = zmm18[0],zmm28[0],zmm18[2],zmm28[2],zmm18[4],zmm28[4],zmm18[6],zmm28[6]
23438 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}
23439 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm10, %zmm6
23440 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23441 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm7, %zmm6
23442 ; AVX512BW-FCP-NEXT: vpermt2q %zmm9, %zmm21, %zmm6
23443 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm4, %zmm8
23444 ; AVX512BW-FCP-NEXT: vpermt2q %zmm5, %zmm21, %zmm8
23445 ; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm8[0,1,2,3],ymm6[4,5,6,7]
23446 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm31, %zmm8
23447 ; AVX512BW-FCP-NEXT: vpermt2q %zmm27, %zmm21, %zmm8
23448 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm10 = zmm18[1],zmm28[1],zmm18[3],zmm28[3],zmm18[5],zmm28[5],zmm18[7],zmm28[7]
23449 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}
23450 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm10, %zmm6
23451 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23452 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm7, %zmm6
23453 ; AVX512BW-FCP-NEXT: vpermt2q %zmm9, %zmm23, %zmm6
23454 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm4, %zmm8
23455 ; AVX512BW-FCP-NEXT: vpermt2q %zmm5, %zmm23, %zmm8
23456 ; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm8[0,1,2,3],ymm6[4,5,6,7]
23457 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm18, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23458 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm18, %zmm8
23459 ; AVX512BW-FCP-NEXT: vpermt2q %zmm28, %zmm23, %zmm8
23460 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm8 {%k1} = zmm31[0],zmm27[0],zmm31[2],zmm27[2],zmm31[4],zmm27[4],zmm31[6],zmm27[6]
23461 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm8, %zmm6
23462 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23463 ; AVX512BW-FCP-NEXT: vpermt2q %zmm9, %zmm25, %zmm7
23464 ; AVX512BW-FCP-NEXT: vpermt2q %zmm5, %zmm25, %zmm4
23465 ; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm7[4,5,6,7]
23466 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm18, %zmm5
23467 ; AVX512BW-FCP-NEXT: vpermt2q %zmm28, %zmm25, %zmm5
23468 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm5 {%k1} = zmm31[1],zmm27[1],zmm31[3],zmm27[3],zmm31[5],zmm27[5],zmm31[7],zmm27[7]
23469 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm4, %zmm5, %zmm4
23470 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23471 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm4
23472 ; AVX512BW-FCP-NEXT: vpermt2q %zmm3, %zmm19, %zmm4
23473 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm5
23474 ; AVX512BW-FCP-NEXT: vpermt2q %zmm1, %zmm19, %zmm5
23475 ; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm5[0,1,2,3],ymm4[4,5,6,7]
23476 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm16, %zmm5
23477 ; AVX512BW-FCP-NEXT: vpermt2q %zmm13, %zmm19, %zmm5
23478 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm6 = zmm11[0],zmm22[0],zmm11[2],zmm22[2],zmm11[4],zmm22[4],zmm11[6],zmm22[6]
23479 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm5, %zmm6 {%k1}
23480 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm4, %zmm6, %zmm4
23481 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23482 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm4
23483 ; AVX512BW-FCP-NEXT: vpermt2q %zmm3, %zmm21, %zmm4
23484 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm5
23485 ; AVX512BW-FCP-NEXT: vpermt2q %zmm1, %zmm21, %zmm5
23486 ; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm5[0,1,2,3],ymm4[4,5,6,7]
23487 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm16, %zmm5
23488 ; AVX512BW-FCP-NEXT: vpermt2q %zmm13, %zmm21, %zmm5
23489 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm6 = zmm11[1],zmm22[1],zmm11[3],zmm22[3],zmm11[5],zmm22[5],zmm11[7],zmm22[7]
23490 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm5, %zmm6 {%k1}
23491 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm4, %zmm6, %zmm4
23492 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23493 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm4
23494 ; AVX512BW-FCP-NEXT: vpermt2q %zmm3, %zmm23, %zmm4
23495 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm5
23496 ; AVX512BW-FCP-NEXT: vpermt2q %zmm1, %zmm23, %zmm5
23497 ; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm5[0,1,2,3],ymm4[4,5,6,7]
23498 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm11, %zmm5
23499 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23500 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm22, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23501 ; AVX512BW-FCP-NEXT: vpermt2q %zmm22, %zmm23, %zmm5
23502 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23503 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23504 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm5 {%k1} = zmm16[0],zmm13[0],zmm16[2],zmm13[2],zmm16[4],zmm13[4],zmm16[6],zmm13[6]
23505 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm4, %zmm5, %zmm4
23506 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23507 ; AVX512BW-FCP-NEXT: vmovdqa64 192(%rdi), %zmm4
23508 ; AVX512BW-FCP-NEXT: vpermt2q %zmm3, %zmm25, %zmm2
23509 ; AVX512BW-FCP-NEXT: vmovdqa64 192(%rdx), %zmm3
23510 ; AVX512BW-FCP-NEXT: vpermt2q %zmm1, %zmm25, %zmm0
23511 ; AVX512BW-FCP-NEXT: vmovdqa64 192(%rcx), %zmm1
23512 ; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
23513 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm11, %zmm2
23514 ; AVX512BW-FCP-NEXT: vpermt2q %zmm22, %zmm25, %zmm2
23515 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm2 {%k1} = zmm16[1],zmm13[1],zmm16[3],zmm13[3],zmm16[5],zmm13[5],zmm16[7],zmm13[7]
23516 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm2, %zmm0
23517 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23518 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm0
23519 ; AVX512BW-FCP-NEXT: vpermt2q %zmm1, %zmm19, %zmm0
23520 ; AVX512BW-FCP-NEXT: vmovdqa64 192(%rsi), %zmm2
23521 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm4, %zmm5
23522 ; AVX512BW-FCP-NEXT: vpermt2q %zmm2, %zmm19, %zmm5
23523 ; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm0[4,5,6,7]
23524 ; AVX512BW-FCP-NEXT: vmovdqa64 192(%r10), %zmm9
23525 ; AVX512BW-FCP-NEXT: vmovdqa64 192(%rax), %zmm20
23526 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm9, %zmm6
23527 ; AVX512BW-FCP-NEXT: vpermt2q %zmm20, %zmm19, %zmm6
23528 ; AVX512BW-FCP-NEXT: vmovdqa64 192(%r8), %zmm8
23529 ; AVX512BW-FCP-NEXT: vmovdqa64 192(%r9), %zmm0
23530 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm7 = zmm8[0],zmm0[0],zmm8[2],zmm0[2],zmm8[4],zmm0[4],zmm8[6],zmm0[6]
23531 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm7 {%k1}
23532 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm5, %zmm7, %zmm5
23533 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23534 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm5
23535 ; AVX512BW-FCP-NEXT: vpermt2q %zmm1, %zmm21, %zmm5
23536 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm4, %zmm6
23537 ; AVX512BW-FCP-NEXT: vpermt2q %zmm2, %zmm21, %zmm6
23538 ; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm5 = ymm6[0,1,2,3],ymm5[4,5,6,7]
23539 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm9, %zmm6
23540 ; AVX512BW-FCP-NEXT: vpermt2q %zmm20, %zmm21, %zmm6
23541 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm7 = zmm8[1],zmm0[1],zmm8[3],zmm0[3],zmm8[5],zmm0[5],zmm8[7],zmm0[7]
23542 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm7 {%k1}
23543 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm5, %zmm7, %zmm5
23544 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23545 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm5
23546 ; AVX512BW-FCP-NEXT: vpermt2q %zmm1, %zmm23, %zmm5
23547 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm4, %zmm6
23548 ; AVX512BW-FCP-NEXT: vpermt2q %zmm2, %zmm23, %zmm6
23549 ; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm5 = ymm6[0,1,2,3],ymm5[4,5,6,7]
23550 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23551 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm8, %zmm6
23552 ; AVX512BW-FCP-NEXT: vpermt2q %zmm0, %zmm23, %zmm6
23553 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm6 {%k1} = zmm9[0],zmm20[0],zmm9[2],zmm20[2],zmm9[4],zmm20[4],zmm9[6],zmm20[6]
23554 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm5, %zmm6, %zmm5
23555 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23556 ; AVX512BW-FCP-NEXT: vpermt2q %zmm1, %zmm25, %zmm3
23557 ; AVX512BW-FCP-NEXT: vpermt2q %zmm2, %zmm25, %zmm4
23558 ; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm1 = ymm4[0,1,2,3],ymm3[4,5,6,7]
23559 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm8, %zmm2
23560 ; AVX512BW-FCP-NEXT: vpermt2q %zmm0, %zmm25, %zmm2
23561 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm2 {%k1} = zmm9[1],zmm20[1],zmm9[3],zmm20[3],zmm9[5],zmm20[5],zmm9[7],zmm20[7]
23562 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm1, %zmm2, %zmm1
23563 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23564 ; AVX512BW-FCP-NEXT: vmovdqa64 256(%rdx), %zmm1
23565 ; AVX512BW-FCP-NEXT: vmovdqa64 256(%rcx), %zmm3
23566 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm2
23567 ; AVX512BW-FCP-NEXT: vpermt2q %zmm3, %zmm19, %zmm2
23568 ; AVX512BW-FCP-NEXT: vmovdqa64 256(%rdi), %zmm4
23569 ; AVX512BW-FCP-NEXT: vmovdqa64 256(%rsi), %zmm5
23570 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm4, %zmm6
23571 ; AVX512BW-FCP-NEXT: vpermt2q %zmm5, %zmm19, %zmm6
23572 ; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm2[4,5,6,7]
23573 ; AVX512BW-FCP-NEXT: vmovdqa64 256(%r10), %zmm13
23574 ; AVX512BW-FCP-NEXT: vmovdqa64 256(%rax), %zmm2
23575 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm13, %zmm7
23576 ; AVX512BW-FCP-NEXT: vpermt2q %zmm2, %zmm19, %zmm7
23577 ; AVX512BW-FCP-NEXT: vmovdqa64 256(%r8), %zmm18
23578 ; AVX512BW-FCP-NEXT: vmovdqa64 256(%r9), %zmm15
23579 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm8 = zmm18[0],zmm15[0],zmm18[2],zmm15[2],zmm18[4],zmm15[4],zmm18[6],zmm15[6]
23580 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm7, %zmm8 {%k1}
23581 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm8, %zmm6
23582 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23583 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm6
23584 ; AVX512BW-FCP-NEXT: vpermt2q %zmm3, %zmm21, %zmm6
23585 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm4, %zmm7
23586 ; AVX512BW-FCP-NEXT: vpermt2q %zmm5, %zmm21, %zmm7
23587 ; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]
23588 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm13, %zmm7
23589 ; AVX512BW-FCP-NEXT: vpermt2q %zmm2, %zmm21, %zmm7
23590 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm8 = zmm18[1],zmm15[1],zmm18[3],zmm15[3],zmm18[5],zmm15[5],zmm18[7],zmm15[7]
23591 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm7, %zmm8 {%k1}
23592 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm8, %zmm6
23593 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23594 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm6
23595 ; AVX512BW-FCP-NEXT: vpermt2q %zmm3, %zmm23, %zmm6
23596 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm4, %zmm7
23597 ; AVX512BW-FCP-NEXT: vpermt2q %zmm5, %zmm23, %zmm7
23598 ; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]
23599 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm18, %zmm7
23600 ; AVX512BW-FCP-NEXT: vpermt2q %zmm15, %zmm23, %zmm7
23601 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm7 {%k1} = zmm13[0],zmm2[0],zmm13[2],zmm2[2],zmm13[4],zmm2[4],zmm13[6],zmm2[6]
23602 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm7, %zmm6
23603 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23604 ; AVX512BW-FCP-NEXT: vpermt2q %zmm3, %zmm25, %zmm1
23605 ; AVX512BW-FCP-NEXT: vpermt2q %zmm5, %zmm25, %zmm4
23606 ; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm1 = ymm4[0,1,2,3],ymm1[4,5,6,7]
23607 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm18, %zmm3
23608 ; AVX512BW-FCP-NEXT: vpermt2q %zmm15, %zmm25, %zmm3
23609 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm3 {%k1} = zmm13[1],zmm2[1],zmm13[3],zmm2[3],zmm13[5],zmm2[5],zmm13[7],zmm2[7]
23610 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm1, %zmm3, %zmm1
23611 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23612 ; AVX512BW-FCP-NEXT: vmovdqa64 320(%rdx), %zmm1
23613 ; AVX512BW-FCP-NEXT: vmovdqa64 320(%rcx), %zmm4
23614 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm3
23615 ; AVX512BW-FCP-NEXT: vpermt2q %zmm4, %zmm19, %zmm3
23616 ; AVX512BW-FCP-NEXT: vmovdqa64 320(%rdi), %zmm6
23617 ; AVX512BW-FCP-NEXT: vmovdqa64 320(%rsi), %zmm7
23618 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm5
23619 ; AVX512BW-FCP-NEXT: vpermt2q %zmm7, %zmm19, %zmm5
23620 ; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm8 = ymm5[0,1,2,3],ymm3[4,5,6,7]
23621 ; AVX512BW-FCP-NEXT: vmovdqa64 320(%r10), %zmm16
23622 ; AVX512BW-FCP-NEXT: vmovdqa64 320(%rax), %zmm22
23623 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm16, %zmm12
23624 ; AVX512BW-FCP-NEXT: vpermt2q %zmm22, %zmm19, %zmm12
23625 ; AVX512BW-FCP-NEXT: vmovdqa64 320(%r8), %zmm29
23626 ; AVX512BW-FCP-NEXT: vmovdqa64 320(%r9), %zmm5
23627 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm14 = zmm29[0],zmm5[0],zmm29[2],zmm5[2],zmm29[4],zmm5[4],zmm29[6],zmm5[6]
23628 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm12, %zmm14 {%k1}
23629 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm8, %zmm14, %zmm3
23630 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23631 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm8
23632 ; AVX512BW-FCP-NEXT: vpermt2q %zmm4, %zmm21, %zmm8
23633 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm12
23634 ; AVX512BW-FCP-NEXT: vpermt2q %zmm7, %zmm21, %zmm12
23635 ; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm8 = ymm12[0,1,2,3],ymm8[4,5,6,7]
23636 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm16, %zmm12
23637 ; AVX512BW-FCP-NEXT: vpermt2q %zmm22, %zmm21, %zmm12
23638 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm14 = zmm29[1],zmm5[1],zmm29[3],zmm5[3],zmm29[5],zmm5[5],zmm29[7],zmm5[7]
23639 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm12, %zmm14 {%k1}
23640 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm8, %zmm14, %zmm3
23641 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23642 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm8
23643 ; AVX512BW-FCP-NEXT: vpermt2q %zmm4, %zmm23, %zmm8
23644 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm12
23645 ; AVX512BW-FCP-NEXT: vpermt2q %zmm7, %zmm23, %zmm12
23646 ; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm8 = ymm12[0,1,2,3],ymm8[4,5,6,7]
23647 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm29, %zmm12
23648 ; AVX512BW-FCP-NEXT: vpermt2q %zmm5, %zmm23, %zmm12
23649 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm12 {%k1} = zmm16[0],zmm22[0],zmm16[2],zmm22[2],zmm16[4],zmm22[4],zmm16[6],zmm22[6]
23650 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm8, %zmm12, %zmm3
23651 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23652 ; AVX512BW-FCP-NEXT: vpermt2q %zmm4, %zmm25, %zmm1
23653 ; AVX512BW-FCP-NEXT: vpermt2q %zmm7, %zmm25, %zmm6
23654 ; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm1 = ymm6[0,1,2,3],ymm1[4,5,6,7]
23655 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm29, %zmm4
23656 ; AVX512BW-FCP-NEXT: vpermt2q %zmm5, %zmm25, %zmm4
23657 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm4 {%k1} = zmm16[1],zmm22[1],zmm16[3],zmm22[3],zmm16[5],zmm22[5],zmm16[7],zmm22[7]
23658 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm1, %zmm4, %zmm1
23659 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23660 ; AVX512BW-FCP-NEXT: vmovdqa64 384(%rdx), %zmm3
23661 ; AVX512BW-FCP-NEXT: vmovdqa64 384(%rcx), %zmm1
23662 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm4
23663 ; AVX512BW-FCP-NEXT: vpermt2q %zmm1, %zmm19, %zmm4
23664 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23665 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm4
23666 ; AVX512BW-FCP-NEXT: vpermt2q %zmm1, %zmm21, %zmm4
23667 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23668 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm4
23669 ; AVX512BW-FCP-NEXT: vpermt2q %zmm1, %zmm23, %zmm4
23670 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23671 ; AVX512BW-FCP-NEXT: vpermt2q %zmm1, %zmm25, %zmm3
23672 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23673 ; AVX512BW-FCP-NEXT: vmovdqa64 384(%rdi), %zmm3
23674 ; AVX512BW-FCP-NEXT: vmovdqa64 384(%rsi), %zmm1
23675 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm4
23676 ; AVX512BW-FCP-NEXT: vpermt2q %zmm1, %zmm19, %zmm4
23677 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23678 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm4
23679 ; AVX512BW-FCP-NEXT: vpermt2q %zmm1, %zmm21, %zmm4
23680 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23681 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm4
23682 ; AVX512BW-FCP-NEXT: vpermt2q %zmm1, %zmm23, %zmm4
23683 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23684 ; AVX512BW-FCP-NEXT: vpermt2q %zmm1, %zmm25, %zmm3
23685 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23686 ; AVX512BW-FCP-NEXT: vmovdqa64 448(%rdx), %zmm3
23687 ; AVX512BW-FCP-NEXT: vmovdqa64 448(%rcx), %zmm1
23688 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm4
23689 ; AVX512BW-FCP-NEXT: vpermt2q %zmm1, %zmm19, %zmm4
23690 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm4, (%rsp) # 64-byte Spill
23691 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm4
23692 ; AVX512BW-FCP-NEXT: vpermt2q %zmm1, %zmm21, %zmm4
23693 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23694 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm4
23695 ; AVX512BW-FCP-NEXT: vpermt2q %zmm1, %zmm23, %zmm4
23696 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23697 ; AVX512BW-FCP-NEXT: vpermt2q %zmm1, %zmm25, %zmm3
23698 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23699 ; AVX512BW-FCP-NEXT: vmovdqa64 448(%rdi), %zmm3
23700 ; AVX512BW-FCP-NEXT: vmovdqa64 448(%rsi), %zmm1
23701 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm4
23702 ; AVX512BW-FCP-NEXT: vpermt2q %zmm1, %zmm19, %zmm4
23703 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23704 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm4
23705 ; AVX512BW-FCP-NEXT: vpermt2q %zmm1, %zmm21, %zmm4
23706 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23707 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm4
23708 ; AVX512BW-FCP-NEXT: vpermt2q %zmm1, %zmm23, %zmm4
23709 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23710 ; AVX512BW-FCP-NEXT: vpermt2q %zmm1, %zmm25, %zmm3
23711 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23712 ; AVX512BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm11 = [0,8,0,8,0,8,0,8]
23713 ; AVX512BW-FCP-NEXT: # zmm11 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
23714 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm26, %zmm1
23715 ; AVX512BW-FCP-NEXT: vpermt2q %zmm17, %zmm11, %zmm1
23716 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23717 ; AVX512BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm12 = [1,9,1,9,1,9,1,9]
23718 ; AVX512BW-FCP-NEXT: # zmm12 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
23719 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm26, %zmm1
23720 ; AVX512BW-FCP-NEXT: vpermt2q %zmm17, %zmm12, %zmm1
23721 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23722 ; AVX512BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm10 = [2,10,2,10,2,10,2,10]
23723 ; AVX512BW-FCP-NEXT: # zmm10 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
23724 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm26, %zmm1
23725 ; AVX512BW-FCP-NEXT: vpermt2q %zmm17, %zmm10, %zmm1
23726 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23727 ; AVX512BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm4 = [3,11,3,11,3,11,3,11]
23728 ; AVX512BW-FCP-NEXT: # zmm4 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
23729 ; AVX512BW-FCP-NEXT: vpermt2q %zmm17, %zmm4, %zmm26
23730 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm26, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23731 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm30, %zmm1
23732 ; AVX512BW-FCP-NEXT: vpermt2q %zmm24, %zmm11, %zmm1
23733 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23734 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm30, %zmm1
23735 ; AVX512BW-FCP-NEXT: vpermt2q %zmm24, %zmm12, %zmm1
23736 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23737 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm30, %zmm1
23738 ; AVX512BW-FCP-NEXT: vpermt2q %zmm24, %zmm10, %zmm1
23739 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23740 ; AVX512BW-FCP-NEXT: vpermt2q %zmm24, %zmm4, %zmm30
23741 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm30, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23742 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm31, %zmm1
23743 ; AVX512BW-FCP-NEXT: vpermt2q %zmm27, %zmm11, %zmm1
23744 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23745 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm31, %zmm1
23746 ; AVX512BW-FCP-NEXT: vpermt2q %zmm27, %zmm12, %zmm1
23747 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23748 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm31, %zmm1
23749 ; AVX512BW-FCP-NEXT: vpermt2q %zmm27, %zmm10, %zmm1
23750 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23751 ; AVX512BW-FCP-NEXT: vpermt2q %zmm27, %zmm4, %zmm31
23752 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm31, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23753 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
23754 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm1
23755 ; AVX512BW-FCP-NEXT: vpermt2q %zmm28, %zmm11, %zmm1
23756 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23757 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm1
23758 ; AVX512BW-FCP-NEXT: vpermt2q %zmm28, %zmm12, %zmm1
23759 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23760 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm31
23761 ; AVX512BW-FCP-NEXT: vpermt2q %zmm28, %zmm10, %zmm31
23762 ; AVX512BW-FCP-NEXT: vpermt2q %zmm28, %zmm4, %zmm3
23763 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23764 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
23765 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm28, %zmm3
23766 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
23767 ; AVX512BW-FCP-NEXT: vpermt2q %zmm1, %zmm11, %zmm3
23768 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23769 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm28, %zmm3
23770 ; AVX512BW-FCP-NEXT: vpermt2q %zmm1, %zmm12, %zmm3
23771 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23772 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm28, %zmm3
23773 ; AVX512BW-FCP-NEXT: vpermt2q %zmm1, %zmm10, %zmm3
23774 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23775 ; AVX512BW-FCP-NEXT: vpermt2q %zmm1, %zmm4, %zmm28
23776 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm28, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23777 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
23778 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm3
23779 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
23780 ; AVX512BW-FCP-NEXT: vpermt2q %zmm1, %zmm11, %zmm3
23781 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23782 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm3
23783 ; AVX512BW-FCP-NEXT: vpermt2q %zmm1, %zmm12, %zmm3
23784 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23785 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm27
23786 ; AVX512BW-FCP-NEXT: vpermt2q %zmm1, %zmm10, %zmm27
23787 ; AVX512BW-FCP-NEXT: vpermt2q %zmm1, %zmm4, %zmm6
23788 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23789 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm9, %zmm1
23790 ; AVX512BW-FCP-NEXT: vpermt2q %zmm20, %zmm11, %zmm1
23791 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23792 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm9, %zmm1
23793 ; AVX512BW-FCP-NEXT: vpermt2q %zmm20, %zmm12, %zmm1
23794 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23795 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm9, %zmm1
23796 ; AVX512BW-FCP-NEXT: vpermt2q %zmm20, %zmm10, %zmm1
23797 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23798 ; AVX512BW-FCP-NEXT: vpermt2q %zmm20, %zmm4, %zmm9
23799 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23800 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
23801 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm3
23802 ; AVX512BW-FCP-NEXT: vpermt2q %zmm0, %zmm11, %zmm3
23803 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23804 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm3
23805 ; AVX512BW-FCP-NEXT: vpermt2q %zmm0, %zmm12, %zmm3
23806 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23807 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm20
23808 ; AVX512BW-FCP-NEXT: vpermt2q %zmm0, %zmm10, %zmm20
23809 ; AVX512BW-FCP-NEXT: vpermt2q %zmm0, %zmm4, %zmm1
23810 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23811 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm13, %zmm0
23812 ; AVX512BW-FCP-NEXT: vpermt2q %zmm2, %zmm11, %zmm0
23813 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23814 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm13, %zmm0
23815 ; AVX512BW-FCP-NEXT: vpermt2q %zmm2, %zmm12, %zmm0
23816 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23817 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm13, %zmm0
23818 ; AVX512BW-FCP-NEXT: vpermt2q %zmm2, %zmm10, %zmm0
23819 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23820 ; AVX512BW-FCP-NEXT: vpermt2q %zmm2, %zmm4, %zmm13
23821 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23822 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm18, %zmm0
23823 ; AVX512BW-FCP-NEXT: vpermt2q %zmm15, %zmm11, %zmm0
23824 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23825 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm18, %zmm0
23826 ; AVX512BW-FCP-NEXT: vpermt2q %zmm15, %zmm12, %zmm0
23827 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23828 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm18, %zmm28
23829 ; AVX512BW-FCP-NEXT: vpermt2q %zmm15, %zmm10, %zmm28
23830 ; AVX512BW-FCP-NEXT: vpermt2q %zmm15, %zmm4, %zmm18
23831 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm18, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23832 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm16, %zmm0
23833 ; AVX512BW-FCP-NEXT: vpermt2q %zmm22, %zmm11, %zmm0
23834 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23835 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm16, %zmm0
23836 ; AVX512BW-FCP-NEXT: vpermt2q %zmm22, %zmm12, %zmm0
23837 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23838 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm16, %zmm0
23839 ; AVX512BW-FCP-NEXT: vpermt2q %zmm22, %zmm10, %zmm0
23840 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23841 ; AVX512BW-FCP-NEXT: vpermt2q %zmm22, %zmm4, %zmm16
23842 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23843 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm29, %zmm0
23844 ; AVX512BW-FCP-NEXT: vpermt2q %zmm5, %zmm11, %zmm0
23845 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23846 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm29, %zmm0
23847 ; AVX512BW-FCP-NEXT: vpermt2q %zmm5, %zmm12, %zmm0
23848 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23849 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm29, %zmm22
23850 ; AVX512BW-FCP-NEXT: vpermt2q %zmm5, %zmm10, %zmm22
23851 ; AVX512BW-FCP-NEXT: vpermt2q %zmm5, %zmm4, %zmm29
23852 ; AVX512BW-FCP-NEXT: vmovdqa64 384(%r10), %zmm13
23853 ; AVX512BW-FCP-NEXT: vmovdqa64 384(%rax), %zmm0
23854 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm13, %zmm1
23855 ; AVX512BW-FCP-NEXT: vpermt2q %zmm0, %zmm11, %zmm1
23856 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23857 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm13, %zmm1
23858 ; AVX512BW-FCP-NEXT: vpermt2q %zmm0, %zmm12, %zmm1
23859 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23860 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm13, %zmm14
23861 ; AVX512BW-FCP-NEXT: vpermt2q %zmm0, %zmm19, %zmm14
23862 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm13, %zmm15
23863 ; AVX512BW-FCP-NEXT: vpermt2q %zmm0, %zmm21, %zmm15
23864 ; AVX512BW-FCP-NEXT: vmovdqa64 384(%r8), %zmm5
23865 ; AVX512BW-FCP-NEXT: vmovdqa64 384(%r9), %zmm2
23866 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm5, %zmm8
23867 ; AVX512BW-FCP-NEXT: vpermt2q %zmm2, %zmm23, %zmm8
23868 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm8 {%k1} = zmm13[0],zmm0[0],zmm13[2],zmm0[2],zmm13[4],zmm0[4],zmm13[6],zmm0[6]
23869 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm5, %zmm6
23870 ; AVX512BW-FCP-NEXT: vpermt2q %zmm2, %zmm25, %zmm6
23871 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm6 {%k1} = zmm13[1],zmm0[1],zmm13[3],zmm0[3],zmm13[5],zmm0[5],zmm13[7],zmm0[7]
23872 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm13, %zmm30
23873 ; AVX512BW-FCP-NEXT: vpermt2q %zmm0, %zmm10, %zmm30
23874 ; AVX512BW-FCP-NEXT: vpermt2q %zmm0, %zmm4, %zmm13
23875 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm5, %zmm24
23876 ; AVX512BW-FCP-NEXT: vpermt2q %zmm2, %zmm11, %zmm24
23877 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm5, %zmm26
23878 ; AVX512BW-FCP-NEXT: vpermt2q %zmm2, %zmm12, %zmm26
23879 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm5, %zmm17
23880 ; AVX512BW-FCP-NEXT: vpermt2q %zmm2, %zmm10, %zmm17
23881 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm7 = zmm5[0],zmm2[0],zmm5[2],zmm2[2],zmm5[4],zmm2[4],zmm5[6],zmm2[6]
23882 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm9 = zmm5[1],zmm2[1],zmm5[3],zmm2[3],zmm5[5],zmm2[5],zmm5[7],zmm2[7]
23883 ; AVX512BW-FCP-NEXT: vpermt2q %zmm2, %zmm4, %zmm5
23884 ; AVX512BW-FCP-NEXT: vmovdqa64 448(%r10), %zmm16
23885 ; AVX512BW-FCP-NEXT: vmovdqa64 448(%rax), %zmm1
23886 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm16, %zmm0
23887 ; AVX512BW-FCP-NEXT: vpermt2q %zmm1, %zmm11, %zmm0
23888 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23889 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm16, %zmm0
23890 ; AVX512BW-FCP-NEXT: vpermt2q %zmm1, %zmm12, %zmm0
23891 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23892 ; AVX512BW-FCP-NEXT: vpermi2q %zmm1, %zmm16, %zmm19
23893 ; AVX512BW-FCP-NEXT: vpermi2q %zmm1, %zmm16, %zmm21
23894 ; AVX512BW-FCP-NEXT: vmovdqa64 448(%r8), %zmm3
23895 ; AVX512BW-FCP-NEXT: vmovdqa64 448(%r9), %zmm0
23896 ; AVX512BW-FCP-NEXT: vpermi2q %zmm0, %zmm3, %zmm23
23897 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm23 {%k1} = zmm16[0],zmm1[0],zmm16[2],zmm1[2],zmm16[4],zmm1[4],zmm16[6],zmm1[6]
23898 ; AVX512BW-FCP-NEXT: vpermi2q %zmm0, %zmm3, %zmm25
23899 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm25 {%k1} = zmm16[1],zmm1[1],zmm16[3],zmm1[3],zmm16[5],zmm1[5],zmm16[7],zmm1[7]
23900 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm16, %zmm18
23901 ; AVX512BW-FCP-NEXT: vpermt2q %zmm1, %zmm10, %zmm18
23902 ; AVX512BW-FCP-NEXT: vpermt2q %zmm1, %zmm4, %zmm16
23903 ; AVX512BW-FCP-NEXT: vpermi2q %zmm0, %zmm3, %zmm11
23904 ; AVX512BW-FCP-NEXT: vpermi2q %zmm0, %zmm3, %zmm12
23905 ; AVX512BW-FCP-NEXT: vpermi2q %zmm0, %zmm3, %zmm10
23906 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm1 = zmm3[0],zmm0[0],zmm3[2],zmm0[2],zmm3[4],zmm0[4],zmm3[6],zmm0[6]
23907 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm2 = zmm3[1],zmm0[1],zmm3[3],zmm0[3],zmm3[5],zmm0[5],zmm3[7],zmm0[7]
23908 ; AVX512BW-FCP-NEXT: vpermt2q %zmm0, %zmm4, %zmm3
23909 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
23910 ; AVX512BW-FCP-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
23911 ; AVX512BW-FCP-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
23912 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm14, %zmm7 {%k1}
23913 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm7, %zmm0
23914 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23915 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
23916 ; AVX512BW-FCP-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
23917 ; AVX512BW-FCP-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
23918 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm15, %zmm9 {%k1}
23919 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm9, %zmm0
23920 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23921 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
23922 ; AVX512BW-FCP-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
23923 ; AVX512BW-FCP-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
23924 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm8, %zmm0
23925 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23926 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
23927 ; AVX512BW-FCP-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
23928 ; AVX512BW-FCP-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
23929 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm6, %zmm0
23930 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23931 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
23932 ; AVX512BW-FCP-NEXT: vpblendd $240, (%rsp), %ymm0, %ymm0 # 32-byte Folded Reload
23933 ; AVX512BW-FCP-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
23934 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm19, %zmm1 {%k1}
23935 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
23936 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23937 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
23938 ; AVX512BW-FCP-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
23939 ; AVX512BW-FCP-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
23940 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm21, %zmm2 {%k1}
23941 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm2, %zmm0
23942 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23943 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
23944 ; AVX512BW-FCP-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
23945 ; AVX512BW-FCP-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
23946 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm23, %zmm0
23947 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23948 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
23949 ; AVX512BW-FCP-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
23950 ; AVX512BW-FCP-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
23951 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm25, %zmm0
23952 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23953 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
23954 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
23955 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm8 {%k1}
23956 ; AVX512BW-FCP-NEXT: vmovdqa (%rcx), %ymm0
23957 ; AVX512BW-FCP-NEXT: vmovdqa (%rdx), %ymm1
23958 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
23959 ; AVX512BW-FCP-NEXT: vmovdqa (%rsi), %ymm4
23960 ; AVX512BW-FCP-NEXT: vmovdqa (%rdi), %ymm6
23961 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm6[0],ymm4[0],ymm6[2],ymm4[2]
23962 ; AVX512BW-FCP-NEXT: vperm2i128 {{.*#+}} ymm2 = ymm7[2,3],ymm2[2,3]
23963 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm8, %zmm2
23964 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm2, (%rsp) # 64-byte Spill
23965 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
23966 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
23967 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm7 {%k1}
23968 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
23969 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm6[1],ymm4[1],ymm6[3],ymm4[3]
23970 ; AVX512BW-FCP-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
23971 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm7, %zmm0
23972 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23973 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
23974 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm31 {%k1}
23975 ; AVX512BW-FCP-NEXT: vmovdqa 64(%rcx), %ymm0
23976 ; AVX512BW-FCP-NEXT: vmovdqa 64(%rdx), %ymm1
23977 ; AVX512BW-FCP-NEXT: vmovdqa 64(%rsi), %ymm2
23978 ; AVX512BW-FCP-NEXT: vmovdqa 64(%rdi), %ymm4
23979 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm6 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
23980 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm4[0],ymm2[0],ymm4[2],ymm2[2]
23981 ; AVX512BW-FCP-NEXT: vperm2i128 {{.*#+}} ymm6 = ymm7[2,3],ymm6[2,3]
23982 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm31, %zmm6
23983 ; AVX512BW-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
23984 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
23985 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
23986 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm7 {%k1}
23987 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
23988 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm4[1],ymm2[1],ymm4[3],ymm2[3]
23989 ; AVX512BW-FCP-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
23990 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm7, %zmm8
23991 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
23992 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm27 {%k1}
23993 ; AVX512BW-FCP-NEXT: vmovdqa 128(%rcx), %ymm0
23994 ; AVX512BW-FCP-NEXT: vmovdqa 128(%rdx), %ymm1
23995 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
23996 ; AVX512BW-FCP-NEXT: vmovdqa 128(%rsi), %ymm4
23997 ; AVX512BW-FCP-NEXT: vmovdqa 128(%rdi), %ymm6
23998 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm6[0],ymm4[0],ymm6[2],ymm4[2]
23999 ; AVX512BW-FCP-NEXT: vperm2i128 {{.*#+}} ymm2 = ymm7[2,3],ymm2[2,3]
24000 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm27, %zmm23
24001 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
24002 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
24003 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm7 {%k1}
24004 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
24005 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm6[1],ymm4[1],ymm6[3],ymm4[3]
24006 ; AVX512BW-FCP-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
24007 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm7, %zmm25
24008 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
24009 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm20 {%k1}
24010 ; AVX512BW-FCP-NEXT: vmovdqa 192(%rcx), %ymm0
24011 ; AVX512BW-FCP-NEXT: vmovdqa 192(%rdx), %ymm1
24012 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
24013 ; AVX512BW-FCP-NEXT: vmovdqa 192(%rsi), %ymm4
24014 ; AVX512BW-FCP-NEXT: vmovdqa 192(%rdi), %ymm6
24015 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm6[0],ymm4[0],ymm6[2],ymm4[2]
24016 ; AVX512BW-FCP-NEXT: vperm2i128 {{.*#+}} ymm2 = ymm7[2,3],ymm2[2,3]
24017 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm20, %zmm21
24018 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
24019 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
24020 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm7 {%k1}
24021 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
24022 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm6[1],ymm4[1],ymm6[3],ymm4[3]
24023 ; AVX512BW-FCP-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
24024 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm7, %zmm19
24025 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
24026 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm28 {%k1}
24027 ; AVX512BW-FCP-NEXT: vmovdqa 256(%rcx), %ymm0
24028 ; AVX512BW-FCP-NEXT: vmovdqa 256(%rdx), %ymm1
24029 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
24030 ; AVX512BW-FCP-NEXT: vmovdqa 256(%rsi), %ymm4
24031 ; AVX512BW-FCP-NEXT: vmovdqa 256(%rdi), %ymm6
24032 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm6[0],ymm4[0],ymm6[2],ymm4[2]
24033 ; AVX512BW-FCP-NEXT: vperm2i128 {{.*#+}} ymm2 = ymm7[2,3],ymm2[2,3]
24034 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm28, %zmm15
24035 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
24036 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
24037 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm7 {%k1}
24038 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
24039 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm6[1],ymm4[1],ymm6[3],ymm4[3]
24040 ; AVX512BW-FCP-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
24041 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm7, %zmm28
24042 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
24043 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm22 {%k1}
24044 ; AVX512BW-FCP-NEXT: vmovdqa 320(%rcx), %ymm0
24045 ; AVX512BW-FCP-NEXT: vmovdqa 320(%rdx), %ymm1
24046 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
24047 ; AVX512BW-FCP-NEXT: vmovdqa 320(%rsi), %ymm4
24048 ; AVX512BW-FCP-NEXT: vmovdqa 320(%rdi), %ymm6
24049 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm6[0],ymm4[0],ymm6[2],ymm4[2]
24050 ; AVX512BW-FCP-NEXT: vperm2i128 {{.*#+}} ymm2 = ymm7[2,3],ymm2[2,3]
24051 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm22, %zmm22
24052 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
24053 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm29 {%k1}
24054 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
24055 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm6[1],ymm4[1],ymm6[3],ymm4[3]
24056 ; AVX512BW-FCP-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
24057 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm29, %zmm27
24058 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm30, %zmm17 {%k1}
24059 ; AVX512BW-FCP-NEXT: vmovdqa 384(%rcx), %ymm0
24060 ; AVX512BW-FCP-NEXT: vmovdqa 384(%rdx), %ymm1
24061 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
24062 ; AVX512BW-FCP-NEXT: vmovdqa 384(%rsi), %ymm4
24063 ; AVX512BW-FCP-NEXT: vmovdqa 384(%rdi), %ymm6
24064 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm6[0],ymm4[0],ymm6[2],ymm4[2]
24065 ; AVX512BW-FCP-NEXT: vperm2i128 {{.*#+}} ymm2 = ymm7[2,3],ymm2[2,3]
24066 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm17, %zmm14
24067 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm13, %zmm5 {%k1}
24068 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
24069 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm6[1],ymm4[1],ymm6[3],ymm4[3]
24070 ; AVX512BW-FCP-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
24071 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm5, %zmm13
24072 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm18, %zmm10 {%k1}
24073 ; AVX512BW-FCP-NEXT: vmovdqa 448(%rcx), %ymm0
24074 ; AVX512BW-FCP-NEXT: vmovdqa 448(%rdx), %ymm1
24075 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
24076 ; AVX512BW-FCP-NEXT: vmovdqa 448(%rsi), %ymm4
24077 ; AVX512BW-FCP-NEXT: vmovdqa 448(%rdi), %ymm5
24078 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm6 = ymm5[0],ymm4[0],ymm5[2],ymm4[2]
24079 ; AVX512BW-FCP-NEXT: vperm2i128 {{.*#+}} ymm2 = ymm6[2,3],ymm2[2,3]
24080 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm10, %zmm10
24081 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
24082 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm5[1],ymm4[1],ymm5[3],ymm4[3]
24083 ; AVX512BW-FCP-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
24084 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm16, %zmm3 {%k1}
24085 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm3, %zmm9
24086 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
24087 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
24088 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm3 {%k1}
24089 ; AVX512BW-FCP-NEXT: vmovdqa (%rsi), %xmm0
24090 ; AVX512BW-FCP-NEXT: vinserti128 $1, (%rcx), %ymm0, %ymm1
24091 ; AVX512BW-FCP-NEXT: vmovdqa (%rdi), %xmm0
24092 ; AVX512BW-FCP-NEXT: vinserti128 $1, (%rdx), %ymm0, %ymm2
24093 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm2[0],ymm1[0],ymm2[2],ymm1[2]
24094 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm3, %zmm0
24095 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
24096 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
24097 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
24098 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm2[1],ymm1[1],ymm2[3],ymm1[3]
24099 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm1, %zmm4, %zmm2
24100 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
24101 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
24102 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm5 {%k1}
24103 ; AVX512BW-FCP-NEXT: vmovdqa 64(%rsi), %xmm1
24104 ; AVX512BW-FCP-NEXT: vinserti128 $1, 64(%rcx), %ymm1, %ymm1
24105 ; AVX512BW-FCP-NEXT: vmovdqa 64(%rdi), %xmm3
24106 ; AVX512BW-FCP-NEXT: vinserti128 $1, 64(%rdx), %ymm3, %ymm4
24107 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm4[0],ymm1[0],ymm4[2],ymm1[2]
24108 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm3, %zmm5, %zmm3
24109 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
24110 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
24111 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm5, %zmm6 {%k1}
24112 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm4[1],ymm1[1],ymm4[3],ymm1[3]
24113 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm1, %zmm6, %zmm5
24114 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
24115 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
24116 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm7 {%k1}
24117 ; AVX512BW-FCP-NEXT: vmovdqa 128(%rsi), %xmm1
24118 ; AVX512BW-FCP-NEXT: vinserti128 $1, 128(%rcx), %ymm1, %ymm1
24119 ; AVX512BW-FCP-NEXT: vmovdqa 128(%rdi), %xmm4
24120 ; AVX512BW-FCP-NEXT: vinserti128 $1, 128(%rdx), %ymm4, %ymm6
24121 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm4 = ymm6[0],ymm1[0],ymm6[2],ymm1[2]
24122 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm4, %zmm7, %zmm4
24123 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
24124 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
24125 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm7, %zmm16 {%k1}
24126 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm6[1],ymm1[1],ymm6[3],ymm1[3]
24127 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm1, %zmm16, %zmm6
24128 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
24129 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
24130 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm17 {%k1}
24131 ; AVX512BW-FCP-NEXT: vmovdqa 192(%rsi), %xmm1
24132 ; AVX512BW-FCP-NEXT: vinserti128 $1, 192(%rcx), %ymm1, %ymm1
24133 ; AVX512BW-FCP-NEXT: vmovdqa 192(%rdi), %xmm7
24134 ; AVX512BW-FCP-NEXT: vinserti32x4 $1, 192(%rdx), %ymm7, %ymm16
24135 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm16[0],ymm1[0],ymm16[2],ymm1[2]
24136 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm7, %zmm17, %zmm7
24137 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
24138 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
24139 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm17, %zmm18 {%k1}
24140 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm16[1],ymm1[1],ymm16[3],ymm1[3]
24141 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm1, %zmm18, %zmm1
24142 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
24143 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
24144 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm16, %zmm20 {%k1}
24145 ; AVX512BW-FCP-NEXT: vmovdqa64 256(%rsi), %xmm16
24146 ; AVX512BW-FCP-NEXT: vinserti32x4 $1, 256(%rcx), %ymm16, %ymm16
24147 ; AVX512BW-FCP-NEXT: vmovdqa64 256(%rdi), %xmm17
24148 ; AVX512BW-FCP-NEXT: vinserti32x4 $1, 256(%rdx), %ymm17, %ymm17
24149 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm18 = ymm17[0],ymm16[0],ymm17[2],ymm16[2]
24150 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm18, %zmm20, %zmm29
24151 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
24152 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
24153 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm18, %zmm20 {%k1}
24154 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm16 = ymm17[1],ymm16[1],ymm17[3],ymm16[3]
24155 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm16, %zmm20, %zmm30
24156 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
24157 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
24158 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm16, %zmm20 {%k1}
24159 ; AVX512BW-FCP-NEXT: vmovdqa64 320(%rsi), %xmm16
24160 ; AVX512BW-FCP-NEXT: vinserti32x4 $1, 320(%rcx), %ymm16, %ymm16
24161 ; AVX512BW-FCP-NEXT: vmovdqa64 320(%rdi), %xmm17
24162 ; AVX512BW-FCP-NEXT: vinserti32x4 $1, 320(%rdx), %ymm17, %ymm17
24163 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm18 = ymm17[0],ymm16[0],ymm17[2],ymm16[2]
24164 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm18, %zmm20, %zmm18
24165 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
24166 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
24167 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm20, %zmm31 {%k1}
24168 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm16 = ymm17[1],ymm16[1],ymm17[3],ymm16[3]
24169 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm16, %zmm31, %zmm16
24170 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
24171 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm17, %zmm24 {%k1}
24172 ; AVX512BW-FCP-NEXT: vmovdqa64 384(%rsi), %xmm17
24173 ; AVX512BW-FCP-NEXT: vinserti32x4 $1, 384(%rcx), %ymm17, %ymm17
24174 ; AVX512BW-FCP-NEXT: vmovdqa64 384(%rdi), %xmm20
24175 ; AVX512BW-FCP-NEXT: vinserti32x4 $1, 384(%rdx), %ymm20, %ymm20
24176 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm31 = ymm20[0],ymm17[0],ymm20[2],ymm17[2]
24177 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm31, %zmm24, %zmm24
24178 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
24179 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm31, %zmm26 {%k1}
24180 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm17 = ymm20[1],ymm17[1],ymm20[3],ymm17[3]
24181 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm17, %zmm26, %zmm17
24182 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
24183 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm20, %zmm11 {%k1}
24184 ; AVX512BW-FCP-NEXT: vmovdqa64 448(%rsi), %xmm20
24185 ; AVX512BW-FCP-NEXT: vinserti32x4 $1, 448(%rcx), %ymm20, %ymm20
24186 ; AVX512BW-FCP-NEXT: vmovdqa64 448(%rdi), %xmm26
24187 ; AVX512BW-FCP-NEXT: vinserti32x4 $1, 448(%rdx), %ymm26, %ymm26
24188 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm31 = ymm26[0],ymm20[0],ymm26[2],ymm20[2]
24189 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm31, %zmm11, %zmm11
24190 ; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
24191 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm31, %zmm12 {%k1}
24192 ; AVX512BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm20 = ymm26[1],ymm20[1],ymm26[3],ymm20[3]
24193 ; AVX512BW-FCP-NEXT: vinserti64x4 $0, %ymm20, %zmm12, %zmm12
24194 ; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
24195 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm9, 3776(%rax)
24196 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm10, 3712(%rax)
24197 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm13, 3264(%rax)
24198 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm14, 3200(%rax)
24199 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm27, 2752(%rax)
24200 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm22, 2688(%rax)
24201 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm28, 2240(%rax)
24202 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm15, 2176(%rax)
24203 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm19, 1728(%rax)
24204 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm21, 1664(%rax)
24205 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm25, 1216(%rax)
24206 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm23, 1152(%rax)
24207 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm8, 704(%rax)
24208 ; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
24209 ; AVX512BW-FCP-NEXT: vmovaps %zmm8, 640(%rax)
24210 ; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
24211 ; AVX512BW-FCP-NEXT: vmovaps %zmm8, 192(%rax)
24212 ; AVX512BW-FCP-NEXT: vmovups (%rsp), %zmm8 # 64-byte Reload
24213 ; AVX512BW-FCP-NEXT: vmovaps %zmm8, 128(%rax)
24214 ; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
24215 ; AVX512BW-FCP-NEXT: vmovaps %zmm8, 4032(%rax)
24216 ; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
24217 ; AVX512BW-FCP-NEXT: vmovaps %zmm8, 3968(%rax)
24218 ; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
24219 ; AVX512BW-FCP-NEXT: vmovaps %zmm8, 3904(%rax)
24220 ; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
24221 ; AVX512BW-FCP-NEXT: vmovaps %zmm8, 3840(%rax)
24222 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm12, 3648(%rax)
24223 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm11, 3584(%rax)
24224 ; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
24225 ; AVX512BW-FCP-NEXT: vmovaps %zmm8, 3520(%rax)
24226 ; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
24227 ; AVX512BW-FCP-NEXT: vmovaps %zmm8, 3456(%rax)
24228 ; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
24229 ; AVX512BW-FCP-NEXT: vmovaps %zmm8, 3392(%rax)
24230 ; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
24231 ; AVX512BW-FCP-NEXT: vmovaps %zmm8, 3328(%rax)
24232 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm17, 3136(%rax)
24233 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm24, 3072(%rax)
24234 ; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
24235 ; AVX512BW-FCP-NEXT: vmovaps %zmm8, 3008(%rax)
24236 ; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
24237 ; AVX512BW-FCP-NEXT: vmovaps %zmm8, 2944(%rax)
24238 ; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
24239 ; AVX512BW-FCP-NEXT: vmovaps %zmm8, 2880(%rax)
24240 ; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
24241 ; AVX512BW-FCP-NEXT: vmovaps %zmm8, 2816(%rax)
24242 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm16, 2624(%rax)
24243 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm18, 2560(%rax)
24244 ; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
24245 ; AVX512BW-FCP-NEXT: vmovaps %zmm8, 2496(%rax)
24246 ; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
24247 ; AVX512BW-FCP-NEXT: vmovaps %zmm8, 2432(%rax)
24248 ; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
24249 ; AVX512BW-FCP-NEXT: vmovaps %zmm8, 2368(%rax)
24250 ; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
24251 ; AVX512BW-FCP-NEXT: vmovaps %zmm8, 2304(%rax)
24252 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm30, 2112(%rax)
24253 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm29, 2048(%rax)
24254 ; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
24255 ; AVX512BW-FCP-NEXT: vmovaps %zmm8, 1984(%rax)
24256 ; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
24257 ; AVX512BW-FCP-NEXT: vmovaps %zmm8, 1920(%rax)
24258 ; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
24259 ; AVX512BW-FCP-NEXT: vmovaps %zmm8, 1856(%rax)
24260 ; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
24261 ; AVX512BW-FCP-NEXT: vmovaps %zmm8, 1792(%rax)
24262 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm1, 1600(%rax)
24263 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm7, 1536(%rax)
24264 ; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
24265 ; AVX512BW-FCP-NEXT: vmovaps %zmm1, 1472(%rax)
24266 ; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
24267 ; AVX512BW-FCP-NEXT: vmovaps %zmm1, 1408(%rax)
24268 ; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
24269 ; AVX512BW-FCP-NEXT: vmovaps %zmm1, 1344(%rax)
24270 ; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
24271 ; AVX512BW-FCP-NEXT: vmovaps %zmm1, 1280(%rax)
24272 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm6, 1088(%rax)
24273 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm4, 1024(%rax)
24274 ; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
24275 ; AVX512BW-FCP-NEXT: vmovaps %zmm1, 960(%rax)
24276 ; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
24277 ; AVX512BW-FCP-NEXT: vmovaps %zmm1, 896(%rax)
24278 ; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
24279 ; AVX512BW-FCP-NEXT: vmovaps %zmm1, 832(%rax)
24280 ; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
24281 ; AVX512BW-FCP-NEXT: vmovaps %zmm1, 768(%rax)
24282 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm5, 576(%rax)
24283 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, 512(%rax)
24284 ; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
24285 ; AVX512BW-FCP-NEXT: vmovaps %zmm1, 448(%rax)
24286 ; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
24287 ; AVX512BW-FCP-NEXT: vmovaps %zmm1, 384(%rax)
24288 ; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
24289 ; AVX512BW-FCP-NEXT: vmovaps %zmm1, 320(%rax)
24290 ; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
24291 ; AVX512BW-FCP-NEXT: vmovaps %zmm1, 256(%rax)
24292 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, 64(%rax)
24293 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, (%rax)
24294 ; AVX512BW-FCP-NEXT: addq $5384, %rsp # imm = 0x1508
24295 ; AVX512BW-FCP-NEXT: vzeroupper
24296 ; AVX512BW-FCP-NEXT: retq
24297 ;
24298 ; AVX512DQ-BW-LABEL: store_i64_stride8_vf64:
24299 ; AVX512DQ-BW: # %bb.0:
24300 ; AVX512DQ-BW-NEXT: subq $5384, %rsp # imm = 0x1508
24301 ; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
24302 ; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %r10
24303 ; AVX512DQ-BW-NEXT: vmovdqa64 128(%rdi), %zmm0
24304 ; AVX512DQ-BW-NEXT: vmovdqa64 64(%rdi), %zmm4
24305 ; AVX512DQ-BW-NEXT: vmovdqa64 (%rdi), %zmm14
24306 ; AVX512DQ-BW-NEXT: vmovdqa64 128(%rsi), %zmm1
24307 ; AVX512DQ-BW-NEXT: vmovdqa64 64(%rsi), %zmm5
24308 ; AVX512DQ-BW-NEXT: vmovdqa64 (%rsi), %zmm10
24309 ; AVX512DQ-BW-NEXT: vmovdqa64 128(%rdx), %zmm2
24310 ; AVX512DQ-BW-NEXT: vmovdqa64 64(%rdx), %zmm7
24311 ; AVX512DQ-BW-NEXT: vmovdqa64 (%rdx), %zmm12
24312 ; AVX512DQ-BW-NEXT: vmovdqa64 128(%rcx), %zmm3
24313 ; AVX512DQ-BW-NEXT: vmovdqa64 64(%rcx), %zmm9
24314 ; AVX512DQ-BW-NEXT: vmovdqa64 (%rcx), %zmm15
24315 ; AVX512DQ-BW-NEXT: vmovdqa64 (%r8), %zmm30
24316 ; AVX512DQ-BW-NEXT: vmovdqa64 64(%r8), %zmm18
24317 ; AVX512DQ-BW-NEXT: vmovdqa64 128(%r8), %zmm11
24318 ; AVX512DQ-BW-NEXT: vmovdqa64 (%r9), %zmm24
24319 ; AVX512DQ-BW-NEXT: vmovdqa64 64(%r9), %zmm28
24320 ; AVX512DQ-BW-NEXT: vmovdqa64 128(%r9), %zmm22
24321 ; AVX512DQ-BW-NEXT: vmovdqa64 (%r10), %zmm26
24322 ; AVX512DQ-BW-NEXT: vmovdqa64 64(%r10), %zmm31
24323 ; AVX512DQ-BW-NEXT: vmovdqa64 128(%r10), %zmm16
24324 ; AVX512DQ-BW-NEXT: vmovdqa64 (%rax), %zmm17
24325 ; AVX512DQ-BW-NEXT: vmovdqa64 64(%rax), %zmm27
24326 ; AVX512DQ-BW-NEXT: vmovdqa64 128(%rax), %zmm13
24327 ; AVX512DQ-BW-NEXT: movb $-64, %r11b
24328 ; AVX512DQ-BW-NEXT: kmovd %r11d, %k1
24329 ; AVX512DQ-BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm19 = [4,12,4,12,4,12,4,12]
24330 ; AVX512DQ-BW-NEXT: # zmm19 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
24331 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm12, %zmm8
24332 ; AVX512DQ-BW-NEXT: vpermt2q %zmm15, %zmm19, %zmm8
24333 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm14, %zmm6
24334 ; AVX512DQ-BW-NEXT: vpermt2q %zmm10, %zmm19, %zmm6
24335 ; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm8[4,5,6,7]
24336 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm26, %zmm8
24337 ; AVX512DQ-BW-NEXT: vpermt2q %zmm17, %zmm19, %zmm8
24338 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} zmm20 = zmm30[0],zmm24[0],zmm30[2],zmm24[2],zmm30[4],zmm24[4],zmm30[6],zmm24[6]
24339 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm8, %zmm20 {%k1}
24340 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm6, %zmm20, %zmm6
24341 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24342 ; AVX512DQ-BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm21 = [5,13,5,13,5,13,5,13]
24343 ; AVX512DQ-BW-NEXT: # zmm21 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
24344 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm12, %zmm6
24345 ; AVX512DQ-BW-NEXT: vpermt2q %zmm15, %zmm21, %zmm6
24346 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm14, %zmm8
24347 ; AVX512DQ-BW-NEXT: vpermt2q %zmm10, %zmm21, %zmm8
24348 ; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm6 = ymm8[0,1,2,3],ymm6[4,5,6,7]
24349 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm26, %zmm8
24350 ; AVX512DQ-BW-NEXT: vpermt2q %zmm17, %zmm21, %zmm8
24351 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} zmm20 = zmm30[1],zmm24[1],zmm30[3],zmm24[3],zmm30[5],zmm24[5],zmm30[7],zmm24[7]
24352 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm8, %zmm20 {%k1}
24353 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm6, %zmm20, %zmm6
24354 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24355 ; AVX512DQ-BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm23 = [6,14,6,14,6,14,6,14]
24356 ; AVX512DQ-BW-NEXT: # zmm23 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
24357 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm12, %zmm6
24358 ; AVX512DQ-BW-NEXT: vpermt2q %zmm15, %zmm23, %zmm6
24359 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm14, %zmm8
24360 ; AVX512DQ-BW-NEXT: vpermt2q %zmm10, %zmm23, %zmm8
24361 ; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm6 = ymm8[0,1,2,3],ymm6[4,5,6,7]
24362 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm30, %zmm8
24363 ; AVX512DQ-BW-NEXT: vpermt2q %zmm24, %zmm23, %zmm8
24364 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} zmm8 {%k1} = zmm26[0],zmm17[0],zmm26[2],zmm17[2],zmm26[4],zmm17[4],zmm26[6],zmm17[6]
24365 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm6, %zmm8, %zmm6
24366 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24367 ; AVX512DQ-BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm25 = [7,15,7,15,7,15,7,15]
24368 ; AVX512DQ-BW-NEXT: # zmm25 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
24369 ; AVX512DQ-BW-NEXT: vpermt2q %zmm15, %zmm25, %zmm12
24370 ; AVX512DQ-BW-NEXT: vpermt2q %zmm10, %zmm25, %zmm14
24371 ; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm6 = ymm14[0,1,2,3],ymm12[4,5,6,7]
24372 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm30, %zmm8
24373 ; AVX512DQ-BW-NEXT: vpermt2q %zmm24, %zmm25, %zmm8
24374 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} zmm8 {%k1} = zmm26[1],zmm17[1],zmm26[3],zmm17[3],zmm26[5],zmm17[5],zmm26[7],zmm17[7]
24375 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm6, %zmm8, %zmm6
24376 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24377 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm7, %zmm6
24378 ; AVX512DQ-BW-NEXT: vpermt2q %zmm9, %zmm19, %zmm6
24379 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm4, %zmm8
24380 ; AVX512DQ-BW-NEXT: vpermt2q %zmm5, %zmm19, %zmm8
24381 ; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm6 = ymm8[0,1,2,3],ymm6[4,5,6,7]
24382 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm31, %zmm8
24383 ; AVX512DQ-BW-NEXT: vpermt2q %zmm27, %zmm19, %zmm8
24384 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} zmm10 = zmm18[0],zmm28[0],zmm18[2],zmm28[2],zmm18[4],zmm28[4],zmm18[6],zmm28[6]
24385 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}
24386 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm6, %zmm10, %zmm6
24387 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24388 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm7, %zmm6
24389 ; AVX512DQ-BW-NEXT: vpermt2q %zmm9, %zmm21, %zmm6
24390 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm4, %zmm8
24391 ; AVX512DQ-BW-NEXT: vpermt2q %zmm5, %zmm21, %zmm8
24392 ; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm6 = ymm8[0,1,2,3],ymm6[4,5,6,7]
24393 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm31, %zmm8
24394 ; AVX512DQ-BW-NEXT: vpermt2q %zmm27, %zmm21, %zmm8
24395 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} zmm10 = zmm18[1],zmm28[1],zmm18[3],zmm28[3],zmm18[5],zmm28[5],zmm18[7],zmm28[7]
24396 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}
24397 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm6, %zmm10, %zmm6
24398 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24399 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm7, %zmm6
24400 ; AVX512DQ-BW-NEXT: vpermt2q %zmm9, %zmm23, %zmm6
24401 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm4, %zmm8
24402 ; AVX512DQ-BW-NEXT: vpermt2q %zmm5, %zmm23, %zmm8
24403 ; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm6 = ymm8[0,1,2,3],ymm6[4,5,6,7]
24404 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm18, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24405 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm18, %zmm8
24406 ; AVX512DQ-BW-NEXT: vpermt2q %zmm28, %zmm23, %zmm8
24407 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} zmm8 {%k1} = zmm31[0],zmm27[0],zmm31[2],zmm27[2],zmm31[4],zmm27[4],zmm31[6],zmm27[6]
24408 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm6, %zmm8, %zmm6
24409 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24410 ; AVX512DQ-BW-NEXT: vpermt2q %zmm9, %zmm25, %zmm7
24411 ; AVX512DQ-BW-NEXT: vpermt2q %zmm5, %zmm25, %zmm4
24412 ; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm7[4,5,6,7]
24413 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm18, %zmm5
24414 ; AVX512DQ-BW-NEXT: vpermt2q %zmm28, %zmm25, %zmm5
24415 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} zmm5 {%k1} = zmm31[1],zmm27[1],zmm31[3],zmm27[3],zmm31[5],zmm27[5],zmm31[7],zmm27[7]
24416 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm4, %zmm5, %zmm4
24417 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24418 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm4
24419 ; AVX512DQ-BW-NEXT: vpermt2q %zmm3, %zmm19, %zmm4
24420 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, %zmm5
24421 ; AVX512DQ-BW-NEXT: vpermt2q %zmm1, %zmm19, %zmm5
24422 ; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm4 = ymm5[0,1,2,3],ymm4[4,5,6,7]
24423 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm16, %zmm5
24424 ; AVX512DQ-BW-NEXT: vpermt2q %zmm13, %zmm19, %zmm5
24425 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} zmm6 = zmm11[0],zmm22[0],zmm11[2],zmm22[2],zmm11[4],zmm22[4],zmm11[6],zmm22[6]
24426 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm5, %zmm6 {%k1}
24427 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm4, %zmm6, %zmm4
24428 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24429 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm4
24430 ; AVX512DQ-BW-NEXT: vpermt2q %zmm3, %zmm21, %zmm4
24431 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, %zmm5
24432 ; AVX512DQ-BW-NEXT: vpermt2q %zmm1, %zmm21, %zmm5
24433 ; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm4 = ymm5[0,1,2,3],ymm4[4,5,6,7]
24434 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm16, %zmm5
24435 ; AVX512DQ-BW-NEXT: vpermt2q %zmm13, %zmm21, %zmm5
24436 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} zmm6 = zmm11[1],zmm22[1],zmm11[3],zmm22[3],zmm11[5],zmm22[5],zmm11[7],zmm22[7]
24437 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm5, %zmm6 {%k1}
24438 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm4, %zmm6, %zmm4
24439 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24440 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm4
24441 ; AVX512DQ-BW-NEXT: vpermt2q %zmm3, %zmm23, %zmm4
24442 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, %zmm5
24443 ; AVX512DQ-BW-NEXT: vpermt2q %zmm1, %zmm23, %zmm5
24444 ; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm4 = ymm5[0,1,2,3],ymm4[4,5,6,7]
24445 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm11, %zmm5
24446 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24447 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm22, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24448 ; AVX512DQ-BW-NEXT: vpermt2q %zmm22, %zmm23, %zmm5
24449 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24450 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24451 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} zmm5 {%k1} = zmm16[0],zmm13[0],zmm16[2],zmm13[2],zmm16[4],zmm13[4],zmm16[6],zmm13[6]
24452 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm4, %zmm5, %zmm4
24453 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24454 ; AVX512DQ-BW-NEXT: vmovdqa64 192(%rdi), %zmm4
24455 ; AVX512DQ-BW-NEXT: vpermt2q %zmm3, %zmm25, %zmm2
24456 ; AVX512DQ-BW-NEXT: vmovdqa64 192(%rdx), %zmm3
24457 ; AVX512DQ-BW-NEXT: vpermt2q %zmm1, %zmm25, %zmm0
24458 ; AVX512DQ-BW-NEXT: vmovdqa64 192(%rcx), %zmm1
24459 ; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
24460 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm11, %zmm2
24461 ; AVX512DQ-BW-NEXT: vpermt2q %zmm22, %zmm25, %zmm2
24462 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} zmm2 {%k1} = zmm16[1],zmm13[1],zmm16[3],zmm13[3],zmm16[5],zmm13[5],zmm16[7],zmm13[7]
24463 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm0, %zmm2, %zmm0
24464 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24465 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, %zmm0
24466 ; AVX512DQ-BW-NEXT: vpermt2q %zmm1, %zmm19, %zmm0
24467 ; AVX512DQ-BW-NEXT: vmovdqa64 192(%rsi), %zmm2
24468 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm4, %zmm5
24469 ; AVX512DQ-BW-NEXT: vpermt2q %zmm2, %zmm19, %zmm5
24470 ; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm0[4,5,6,7]
24471 ; AVX512DQ-BW-NEXT: vmovdqa64 192(%r10), %zmm9
24472 ; AVX512DQ-BW-NEXT: vmovdqa64 192(%rax), %zmm20
24473 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm9, %zmm6
24474 ; AVX512DQ-BW-NEXT: vpermt2q %zmm20, %zmm19, %zmm6
24475 ; AVX512DQ-BW-NEXT: vmovdqa64 192(%r8), %zmm8
24476 ; AVX512DQ-BW-NEXT: vmovdqa64 192(%r9), %zmm0
24477 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} zmm7 = zmm8[0],zmm0[0],zmm8[2],zmm0[2],zmm8[4],zmm0[4],zmm8[6],zmm0[6]
24478 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm6, %zmm7 {%k1}
24479 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm5, %zmm7, %zmm5
24480 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24481 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, %zmm5
24482 ; AVX512DQ-BW-NEXT: vpermt2q %zmm1, %zmm21, %zmm5
24483 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm4, %zmm6
24484 ; AVX512DQ-BW-NEXT: vpermt2q %zmm2, %zmm21, %zmm6
24485 ; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm5 = ymm6[0,1,2,3],ymm5[4,5,6,7]
24486 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm9, %zmm6
24487 ; AVX512DQ-BW-NEXT: vpermt2q %zmm20, %zmm21, %zmm6
24488 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} zmm7 = zmm8[1],zmm0[1],zmm8[3],zmm0[3],zmm8[5],zmm0[5],zmm8[7],zmm0[7]
24489 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm6, %zmm7 {%k1}
24490 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm5, %zmm7, %zmm5
24491 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24492 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, %zmm5
24493 ; AVX512DQ-BW-NEXT: vpermt2q %zmm1, %zmm23, %zmm5
24494 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm4, %zmm6
24495 ; AVX512DQ-BW-NEXT: vpermt2q %zmm2, %zmm23, %zmm6
24496 ; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm5 = ymm6[0,1,2,3],ymm5[4,5,6,7]
24497 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24498 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm8, %zmm6
24499 ; AVX512DQ-BW-NEXT: vpermt2q %zmm0, %zmm23, %zmm6
24500 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} zmm6 {%k1} = zmm9[0],zmm20[0],zmm9[2],zmm20[2],zmm9[4],zmm20[4],zmm9[6],zmm20[6]
24501 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm5, %zmm6, %zmm5
24502 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24503 ; AVX512DQ-BW-NEXT: vpermt2q %zmm1, %zmm25, %zmm3
24504 ; AVX512DQ-BW-NEXT: vpermt2q %zmm2, %zmm25, %zmm4
24505 ; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm1 = ymm4[0,1,2,3],ymm3[4,5,6,7]
24506 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm8, %zmm2
24507 ; AVX512DQ-BW-NEXT: vpermt2q %zmm0, %zmm25, %zmm2
24508 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} zmm2 {%k1} = zmm9[1],zmm20[1],zmm9[3],zmm20[3],zmm9[5],zmm20[5],zmm9[7],zmm20[7]
24509 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm1, %zmm2, %zmm1
24510 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24511 ; AVX512DQ-BW-NEXT: vmovdqa64 256(%rdx), %zmm1
24512 ; AVX512DQ-BW-NEXT: vmovdqa64 256(%rcx), %zmm3
24513 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm1, %zmm2
24514 ; AVX512DQ-BW-NEXT: vpermt2q %zmm3, %zmm19, %zmm2
24515 ; AVX512DQ-BW-NEXT: vmovdqa64 256(%rdi), %zmm4
24516 ; AVX512DQ-BW-NEXT: vmovdqa64 256(%rsi), %zmm5
24517 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm4, %zmm6
24518 ; AVX512DQ-BW-NEXT: vpermt2q %zmm5, %zmm19, %zmm6
24519 ; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm2[4,5,6,7]
24520 ; AVX512DQ-BW-NEXT: vmovdqa64 256(%r10), %zmm13
24521 ; AVX512DQ-BW-NEXT: vmovdqa64 256(%rax), %zmm2
24522 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm13, %zmm7
24523 ; AVX512DQ-BW-NEXT: vpermt2q %zmm2, %zmm19, %zmm7
24524 ; AVX512DQ-BW-NEXT: vmovdqa64 256(%r8), %zmm18
24525 ; AVX512DQ-BW-NEXT: vmovdqa64 256(%r9), %zmm15
24526 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} zmm8 = zmm18[0],zmm15[0],zmm18[2],zmm15[2],zmm18[4],zmm15[4],zmm18[6],zmm15[6]
24527 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm7, %zmm8 {%k1}
24528 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm6, %zmm8, %zmm6
24529 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24530 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm1, %zmm6
24531 ; AVX512DQ-BW-NEXT: vpermt2q %zmm3, %zmm21, %zmm6
24532 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm4, %zmm7
24533 ; AVX512DQ-BW-NEXT: vpermt2q %zmm5, %zmm21, %zmm7
24534 ; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]
24535 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm13, %zmm7
24536 ; AVX512DQ-BW-NEXT: vpermt2q %zmm2, %zmm21, %zmm7
24537 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} zmm8 = zmm18[1],zmm15[1],zmm18[3],zmm15[3],zmm18[5],zmm15[5],zmm18[7],zmm15[7]
24538 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm7, %zmm8 {%k1}
24539 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm6, %zmm8, %zmm6
24540 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24541 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm1, %zmm6
24542 ; AVX512DQ-BW-NEXT: vpermt2q %zmm3, %zmm23, %zmm6
24543 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm4, %zmm7
24544 ; AVX512DQ-BW-NEXT: vpermt2q %zmm5, %zmm23, %zmm7
24545 ; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]
24546 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm18, %zmm7
24547 ; AVX512DQ-BW-NEXT: vpermt2q %zmm15, %zmm23, %zmm7
24548 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} zmm7 {%k1} = zmm13[0],zmm2[0],zmm13[2],zmm2[2],zmm13[4],zmm2[4],zmm13[6],zmm2[6]
24549 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm6, %zmm7, %zmm6
24550 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24551 ; AVX512DQ-BW-NEXT: vpermt2q %zmm3, %zmm25, %zmm1
24552 ; AVX512DQ-BW-NEXT: vpermt2q %zmm5, %zmm25, %zmm4
24553 ; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm1 = ymm4[0,1,2,3],ymm1[4,5,6,7]
24554 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm18, %zmm3
24555 ; AVX512DQ-BW-NEXT: vpermt2q %zmm15, %zmm25, %zmm3
24556 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} zmm3 {%k1} = zmm13[1],zmm2[1],zmm13[3],zmm2[3],zmm13[5],zmm2[5],zmm13[7],zmm2[7]
24557 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm1, %zmm3, %zmm1
24558 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24559 ; AVX512DQ-BW-NEXT: vmovdqa64 320(%rdx), %zmm1
24560 ; AVX512DQ-BW-NEXT: vmovdqa64 320(%rcx), %zmm4
24561 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm1, %zmm3
24562 ; AVX512DQ-BW-NEXT: vpermt2q %zmm4, %zmm19, %zmm3
24563 ; AVX512DQ-BW-NEXT: vmovdqa64 320(%rdi), %zmm6
24564 ; AVX512DQ-BW-NEXT: vmovdqa64 320(%rsi), %zmm7
24565 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm6, %zmm5
24566 ; AVX512DQ-BW-NEXT: vpermt2q %zmm7, %zmm19, %zmm5
24567 ; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm8 = ymm5[0,1,2,3],ymm3[4,5,6,7]
24568 ; AVX512DQ-BW-NEXT: vmovdqa64 320(%r10), %zmm16
24569 ; AVX512DQ-BW-NEXT: vmovdqa64 320(%rax), %zmm22
24570 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm16, %zmm12
24571 ; AVX512DQ-BW-NEXT: vpermt2q %zmm22, %zmm19, %zmm12
24572 ; AVX512DQ-BW-NEXT: vmovdqa64 320(%r8), %zmm29
24573 ; AVX512DQ-BW-NEXT: vmovdqa64 320(%r9), %zmm5
24574 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} zmm14 = zmm29[0],zmm5[0],zmm29[2],zmm5[2],zmm29[4],zmm5[4],zmm29[6],zmm5[6]
24575 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm12, %zmm14 {%k1}
24576 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm8, %zmm14, %zmm3
24577 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24578 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm1, %zmm8
24579 ; AVX512DQ-BW-NEXT: vpermt2q %zmm4, %zmm21, %zmm8
24580 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm6, %zmm12
24581 ; AVX512DQ-BW-NEXT: vpermt2q %zmm7, %zmm21, %zmm12
24582 ; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm8 = ymm12[0,1,2,3],ymm8[4,5,6,7]
24583 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm16, %zmm12
24584 ; AVX512DQ-BW-NEXT: vpermt2q %zmm22, %zmm21, %zmm12
24585 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} zmm14 = zmm29[1],zmm5[1],zmm29[3],zmm5[3],zmm29[5],zmm5[5],zmm29[7],zmm5[7]
24586 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm12, %zmm14 {%k1}
24587 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm8, %zmm14, %zmm3
24588 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24589 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm1, %zmm8
24590 ; AVX512DQ-BW-NEXT: vpermt2q %zmm4, %zmm23, %zmm8
24591 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm6, %zmm12
24592 ; AVX512DQ-BW-NEXT: vpermt2q %zmm7, %zmm23, %zmm12
24593 ; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm8 = ymm12[0,1,2,3],ymm8[4,5,6,7]
24594 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm29, %zmm12
24595 ; AVX512DQ-BW-NEXT: vpermt2q %zmm5, %zmm23, %zmm12
24596 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} zmm12 {%k1} = zmm16[0],zmm22[0],zmm16[2],zmm22[2],zmm16[4],zmm22[4],zmm16[6],zmm22[6]
24597 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm8, %zmm12, %zmm3
24598 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24599 ; AVX512DQ-BW-NEXT: vpermt2q %zmm4, %zmm25, %zmm1
24600 ; AVX512DQ-BW-NEXT: vpermt2q %zmm7, %zmm25, %zmm6
24601 ; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm1 = ymm6[0,1,2,3],ymm1[4,5,6,7]
24602 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm29, %zmm4
24603 ; AVX512DQ-BW-NEXT: vpermt2q %zmm5, %zmm25, %zmm4
24604 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} zmm4 {%k1} = zmm16[1],zmm22[1],zmm16[3],zmm22[3],zmm16[5],zmm22[5],zmm16[7],zmm22[7]
24605 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm1, %zmm4, %zmm1
24606 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24607 ; AVX512DQ-BW-NEXT: vmovdqa64 384(%rdx), %zmm3
24608 ; AVX512DQ-BW-NEXT: vmovdqa64 384(%rcx), %zmm1
24609 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, %zmm4
24610 ; AVX512DQ-BW-NEXT: vpermt2q %zmm1, %zmm19, %zmm4
24611 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24612 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, %zmm4
24613 ; AVX512DQ-BW-NEXT: vpermt2q %zmm1, %zmm21, %zmm4
24614 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24615 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, %zmm4
24616 ; AVX512DQ-BW-NEXT: vpermt2q %zmm1, %zmm23, %zmm4
24617 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24618 ; AVX512DQ-BW-NEXT: vpermt2q %zmm1, %zmm25, %zmm3
24619 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24620 ; AVX512DQ-BW-NEXT: vmovdqa64 384(%rdi), %zmm3
24621 ; AVX512DQ-BW-NEXT: vmovdqa64 384(%rsi), %zmm1
24622 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, %zmm4
24623 ; AVX512DQ-BW-NEXT: vpermt2q %zmm1, %zmm19, %zmm4
24624 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24625 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, %zmm4
24626 ; AVX512DQ-BW-NEXT: vpermt2q %zmm1, %zmm21, %zmm4
24627 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24628 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, %zmm4
24629 ; AVX512DQ-BW-NEXT: vpermt2q %zmm1, %zmm23, %zmm4
24630 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24631 ; AVX512DQ-BW-NEXT: vpermt2q %zmm1, %zmm25, %zmm3
24632 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24633 ; AVX512DQ-BW-NEXT: vmovdqa64 448(%rdx), %zmm3
24634 ; AVX512DQ-BW-NEXT: vmovdqa64 448(%rcx), %zmm1
24635 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, %zmm4
24636 ; AVX512DQ-BW-NEXT: vpermt2q %zmm1, %zmm19, %zmm4
24637 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm4, (%rsp) # 64-byte Spill
24638 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, %zmm4
24639 ; AVX512DQ-BW-NEXT: vpermt2q %zmm1, %zmm21, %zmm4
24640 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24641 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, %zmm4
24642 ; AVX512DQ-BW-NEXT: vpermt2q %zmm1, %zmm23, %zmm4
24643 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24644 ; AVX512DQ-BW-NEXT: vpermt2q %zmm1, %zmm25, %zmm3
24645 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24646 ; AVX512DQ-BW-NEXT: vmovdqa64 448(%rdi), %zmm3
24647 ; AVX512DQ-BW-NEXT: vmovdqa64 448(%rsi), %zmm1
24648 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, %zmm4
24649 ; AVX512DQ-BW-NEXT: vpermt2q %zmm1, %zmm19, %zmm4
24650 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24651 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, %zmm4
24652 ; AVX512DQ-BW-NEXT: vpermt2q %zmm1, %zmm21, %zmm4
24653 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24654 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, %zmm4
24655 ; AVX512DQ-BW-NEXT: vpermt2q %zmm1, %zmm23, %zmm4
24656 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24657 ; AVX512DQ-BW-NEXT: vpermt2q %zmm1, %zmm25, %zmm3
24658 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24659 ; AVX512DQ-BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm11 = [0,8,0,8,0,8,0,8]
24660 ; AVX512DQ-BW-NEXT: # zmm11 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
24661 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm26, %zmm1
24662 ; AVX512DQ-BW-NEXT: vpermt2q %zmm17, %zmm11, %zmm1
24663 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24664 ; AVX512DQ-BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm12 = [1,9,1,9,1,9,1,9]
24665 ; AVX512DQ-BW-NEXT: # zmm12 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
24666 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm26, %zmm1
24667 ; AVX512DQ-BW-NEXT: vpermt2q %zmm17, %zmm12, %zmm1
24668 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24669 ; AVX512DQ-BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm10 = [2,10,2,10,2,10,2,10]
24670 ; AVX512DQ-BW-NEXT: # zmm10 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
24671 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm26, %zmm1
24672 ; AVX512DQ-BW-NEXT: vpermt2q %zmm17, %zmm10, %zmm1
24673 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24674 ; AVX512DQ-BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm4 = [3,11,3,11,3,11,3,11]
24675 ; AVX512DQ-BW-NEXT: # zmm4 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
24676 ; AVX512DQ-BW-NEXT: vpermt2q %zmm17, %zmm4, %zmm26
24677 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm26, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24678 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm30, %zmm1
24679 ; AVX512DQ-BW-NEXT: vpermt2q %zmm24, %zmm11, %zmm1
24680 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24681 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm30, %zmm1
24682 ; AVX512DQ-BW-NEXT: vpermt2q %zmm24, %zmm12, %zmm1
24683 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24684 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm30, %zmm1
24685 ; AVX512DQ-BW-NEXT: vpermt2q %zmm24, %zmm10, %zmm1
24686 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24687 ; AVX512DQ-BW-NEXT: vpermt2q %zmm24, %zmm4, %zmm30
24688 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm30, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24689 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm31, %zmm1
24690 ; AVX512DQ-BW-NEXT: vpermt2q %zmm27, %zmm11, %zmm1
24691 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24692 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm31, %zmm1
24693 ; AVX512DQ-BW-NEXT: vpermt2q %zmm27, %zmm12, %zmm1
24694 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24695 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm31, %zmm1
24696 ; AVX512DQ-BW-NEXT: vpermt2q %zmm27, %zmm10, %zmm1
24697 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24698 ; AVX512DQ-BW-NEXT: vpermt2q %zmm27, %zmm4, %zmm31
24699 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm31, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24700 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
24701 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, %zmm1
24702 ; AVX512DQ-BW-NEXT: vpermt2q %zmm28, %zmm11, %zmm1
24703 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24704 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, %zmm1
24705 ; AVX512DQ-BW-NEXT: vpermt2q %zmm28, %zmm12, %zmm1
24706 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24707 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, %zmm31
24708 ; AVX512DQ-BW-NEXT: vpermt2q %zmm28, %zmm10, %zmm31
24709 ; AVX512DQ-BW-NEXT: vpermt2q %zmm28, %zmm4, %zmm3
24710 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24711 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
24712 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm28, %zmm3
24713 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
24714 ; AVX512DQ-BW-NEXT: vpermt2q %zmm1, %zmm11, %zmm3
24715 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24716 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm28, %zmm3
24717 ; AVX512DQ-BW-NEXT: vpermt2q %zmm1, %zmm12, %zmm3
24718 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24719 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm28, %zmm3
24720 ; AVX512DQ-BW-NEXT: vpermt2q %zmm1, %zmm10, %zmm3
24721 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24722 ; AVX512DQ-BW-NEXT: vpermt2q %zmm1, %zmm4, %zmm28
24723 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm28, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24724 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
24725 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm6, %zmm3
24726 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
24727 ; AVX512DQ-BW-NEXT: vpermt2q %zmm1, %zmm11, %zmm3
24728 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24729 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm6, %zmm3
24730 ; AVX512DQ-BW-NEXT: vpermt2q %zmm1, %zmm12, %zmm3
24731 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24732 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm6, %zmm27
24733 ; AVX512DQ-BW-NEXT: vpermt2q %zmm1, %zmm10, %zmm27
24734 ; AVX512DQ-BW-NEXT: vpermt2q %zmm1, %zmm4, %zmm6
24735 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24736 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm9, %zmm1
24737 ; AVX512DQ-BW-NEXT: vpermt2q %zmm20, %zmm11, %zmm1
24738 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24739 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm9, %zmm1
24740 ; AVX512DQ-BW-NEXT: vpermt2q %zmm20, %zmm12, %zmm1
24741 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24742 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm9, %zmm1
24743 ; AVX512DQ-BW-NEXT: vpermt2q %zmm20, %zmm10, %zmm1
24744 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24745 ; AVX512DQ-BW-NEXT: vpermt2q %zmm20, %zmm4, %zmm9
24746 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24747 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
24748 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm1, %zmm3
24749 ; AVX512DQ-BW-NEXT: vpermt2q %zmm0, %zmm11, %zmm3
24750 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24751 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm1, %zmm3
24752 ; AVX512DQ-BW-NEXT: vpermt2q %zmm0, %zmm12, %zmm3
24753 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24754 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm1, %zmm20
24755 ; AVX512DQ-BW-NEXT: vpermt2q %zmm0, %zmm10, %zmm20
24756 ; AVX512DQ-BW-NEXT: vpermt2q %zmm0, %zmm4, %zmm1
24757 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24758 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm13, %zmm0
24759 ; AVX512DQ-BW-NEXT: vpermt2q %zmm2, %zmm11, %zmm0
24760 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24761 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm13, %zmm0
24762 ; AVX512DQ-BW-NEXT: vpermt2q %zmm2, %zmm12, %zmm0
24763 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24764 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm13, %zmm0
24765 ; AVX512DQ-BW-NEXT: vpermt2q %zmm2, %zmm10, %zmm0
24766 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24767 ; AVX512DQ-BW-NEXT: vpermt2q %zmm2, %zmm4, %zmm13
24768 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24769 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm18, %zmm0
24770 ; AVX512DQ-BW-NEXT: vpermt2q %zmm15, %zmm11, %zmm0
24771 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24772 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm18, %zmm0
24773 ; AVX512DQ-BW-NEXT: vpermt2q %zmm15, %zmm12, %zmm0
24774 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24775 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm18, %zmm28
24776 ; AVX512DQ-BW-NEXT: vpermt2q %zmm15, %zmm10, %zmm28
24777 ; AVX512DQ-BW-NEXT: vpermt2q %zmm15, %zmm4, %zmm18
24778 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm18, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24779 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm16, %zmm0
24780 ; AVX512DQ-BW-NEXT: vpermt2q %zmm22, %zmm11, %zmm0
24781 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24782 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm16, %zmm0
24783 ; AVX512DQ-BW-NEXT: vpermt2q %zmm22, %zmm12, %zmm0
24784 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24785 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm16, %zmm0
24786 ; AVX512DQ-BW-NEXT: vpermt2q %zmm22, %zmm10, %zmm0
24787 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24788 ; AVX512DQ-BW-NEXT: vpermt2q %zmm22, %zmm4, %zmm16
24789 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24790 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm29, %zmm0
24791 ; AVX512DQ-BW-NEXT: vpermt2q %zmm5, %zmm11, %zmm0
24792 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24793 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm29, %zmm0
24794 ; AVX512DQ-BW-NEXT: vpermt2q %zmm5, %zmm12, %zmm0
24795 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24796 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm29, %zmm22
24797 ; AVX512DQ-BW-NEXT: vpermt2q %zmm5, %zmm10, %zmm22
24798 ; AVX512DQ-BW-NEXT: vpermt2q %zmm5, %zmm4, %zmm29
24799 ; AVX512DQ-BW-NEXT: vmovdqa64 384(%r10), %zmm13
24800 ; AVX512DQ-BW-NEXT: vmovdqa64 384(%rax), %zmm0
24801 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm13, %zmm1
24802 ; AVX512DQ-BW-NEXT: vpermt2q %zmm0, %zmm11, %zmm1
24803 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24804 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm13, %zmm1
24805 ; AVX512DQ-BW-NEXT: vpermt2q %zmm0, %zmm12, %zmm1
24806 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24807 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm13, %zmm14
24808 ; AVX512DQ-BW-NEXT: vpermt2q %zmm0, %zmm19, %zmm14
24809 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm13, %zmm15
24810 ; AVX512DQ-BW-NEXT: vpermt2q %zmm0, %zmm21, %zmm15
24811 ; AVX512DQ-BW-NEXT: vmovdqa64 384(%r8), %zmm5
24812 ; AVX512DQ-BW-NEXT: vmovdqa64 384(%r9), %zmm2
24813 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm5, %zmm8
24814 ; AVX512DQ-BW-NEXT: vpermt2q %zmm2, %zmm23, %zmm8
24815 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} zmm8 {%k1} = zmm13[0],zmm0[0],zmm13[2],zmm0[2],zmm13[4],zmm0[4],zmm13[6],zmm0[6]
24816 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm5, %zmm6
24817 ; AVX512DQ-BW-NEXT: vpermt2q %zmm2, %zmm25, %zmm6
24818 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} zmm6 {%k1} = zmm13[1],zmm0[1],zmm13[3],zmm0[3],zmm13[5],zmm0[5],zmm13[7],zmm0[7]
24819 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm13, %zmm30
24820 ; AVX512DQ-BW-NEXT: vpermt2q %zmm0, %zmm10, %zmm30
24821 ; AVX512DQ-BW-NEXT: vpermt2q %zmm0, %zmm4, %zmm13
24822 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm5, %zmm24
24823 ; AVX512DQ-BW-NEXT: vpermt2q %zmm2, %zmm11, %zmm24
24824 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm5, %zmm26
24825 ; AVX512DQ-BW-NEXT: vpermt2q %zmm2, %zmm12, %zmm26
24826 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm5, %zmm17
24827 ; AVX512DQ-BW-NEXT: vpermt2q %zmm2, %zmm10, %zmm17
24828 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} zmm7 = zmm5[0],zmm2[0],zmm5[2],zmm2[2],zmm5[4],zmm2[4],zmm5[6],zmm2[6]
24829 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} zmm9 = zmm5[1],zmm2[1],zmm5[3],zmm2[3],zmm5[5],zmm2[5],zmm5[7],zmm2[7]
24830 ; AVX512DQ-BW-NEXT: vpermt2q %zmm2, %zmm4, %zmm5
24831 ; AVX512DQ-BW-NEXT: vmovdqa64 448(%r10), %zmm16
24832 ; AVX512DQ-BW-NEXT: vmovdqa64 448(%rax), %zmm1
24833 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm16, %zmm0
24834 ; AVX512DQ-BW-NEXT: vpermt2q %zmm1, %zmm11, %zmm0
24835 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24836 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm16, %zmm0
24837 ; AVX512DQ-BW-NEXT: vpermt2q %zmm1, %zmm12, %zmm0
24838 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24839 ; AVX512DQ-BW-NEXT: vpermi2q %zmm1, %zmm16, %zmm19
24840 ; AVX512DQ-BW-NEXT: vpermi2q %zmm1, %zmm16, %zmm21
24841 ; AVX512DQ-BW-NEXT: vmovdqa64 448(%r8), %zmm3
24842 ; AVX512DQ-BW-NEXT: vmovdqa64 448(%r9), %zmm0
24843 ; AVX512DQ-BW-NEXT: vpermi2q %zmm0, %zmm3, %zmm23
24844 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} zmm23 {%k1} = zmm16[0],zmm1[0],zmm16[2],zmm1[2],zmm16[4],zmm1[4],zmm16[6],zmm1[6]
24845 ; AVX512DQ-BW-NEXT: vpermi2q %zmm0, %zmm3, %zmm25
24846 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} zmm25 {%k1} = zmm16[1],zmm1[1],zmm16[3],zmm1[3],zmm16[5],zmm1[5],zmm16[7],zmm1[7]
24847 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm16, %zmm18
24848 ; AVX512DQ-BW-NEXT: vpermt2q %zmm1, %zmm10, %zmm18
24849 ; AVX512DQ-BW-NEXT: vpermt2q %zmm1, %zmm4, %zmm16
24850 ; AVX512DQ-BW-NEXT: vpermi2q %zmm0, %zmm3, %zmm11
24851 ; AVX512DQ-BW-NEXT: vpermi2q %zmm0, %zmm3, %zmm12
24852 ; AVX512DQ-BW-NEXT: vpermi2q %zmm0, %zmm3, %zmm10
24853 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} zmm1 = zmm3[0],zmm0[0],zmm3[2],zmm0[2],zmm3[4],zmm0[4],zmm3[6],zmm0[6]
24854 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} zmm2 = zmm3[1],zmm0[1],zmm3[3],zmm0[3],zmm3[5],zmm0[5],zmm3[7],zmm0[7]
24855 ; AVX512DQ-BW-NEXT: vpermt2q %zmm0, %zmm4, %zmm3
24856 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
24857 ; AVX512DQ-BW-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
24858 ; AVX512DQ-BW-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
24859 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm14, %zmm7 {%k1}
24860 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm0, %zmm7, %zmm0
24861 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24862 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
24863 ; AVX512DQ-BW-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
24864 ; AVX512DQ-BW-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
24865 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm15, %zmm9 {%k1}
24866 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm0, %zmm9, %zmm0
24867 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24868 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
24869 ; AVX512DQ-BW-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
24870 ; AVX512DQ-BW-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
24871 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm0, %zmm8, %zmm0
24872 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24873 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
24874 ; AVX512DQ-BW-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
24875 ; AVX512DQ-BW-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
24876 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm0, %zmm6, %zmm0
24877 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24878 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
24879 ; AVX512DQ-BW-NEXT: vpblendd $240, (%rsp), %ymm0, %ymm0 # 32-byte Folded Reload
24880 ; AVX512DQ-BW-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
24881 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm19, %zmm1 {%k1}
24882 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
24883 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24884 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
24885 ; AVX512DQ-BW-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
24886 ; AVX512DQ-BW-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
24887 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm21, %zmm2 {%k1}
24888 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm0, %zmm2, %zmm0
24889 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24890 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
24891 ; AVX512DQ-BW-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
24892 ; AVX512DQ-BW-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
24893 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm0, %zmm23, %zmm0
24894 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24895 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
24896 ; AVX512DQ-BW-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
24897 ; AVX512DQ-BW-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
24898 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm0, %zmm25, %zmm0
24899 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24900 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
24901 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
24902 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, %zmm8 {%k1}
24903 ; AVX512DQ-BW-NEXT: vmovdqa (%rcx), %ymm0
24904 ; AVX512DQ-BW-NEXT: vmovdqa (%rdx), %ymm1
24905 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
24906 ; AVX512DQ-BW-NEXT: vmovdqa (%rsi), %ymm4
24907 ; AVX512DQ-BW-NEXT: vmovdqa (%rdi), %ymm6
24908 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm6[0],ymm4[0],ymm6[2],ymm4[2]
24909 ; AVX512DQ-BW-NEXT: vperm2i128 {{.*#+}} ymm2 = ymm7[2,3],ymm2[2,3]
24910 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm2, %zmm8, %zmm2
24911 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm2, (%rsp) # 64-byte Spill
24912 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
24913 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
24914 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm7 {%k1}
24915 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
24916 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm6[1],ymm4[1],ymm6[3],ymm4[3]
24917 ; AVX512DQ-BW-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
24918 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm0, %zmm7, %zmm0
24919 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24920 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
24921 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, %zmm31 {%k1}
24922 ; AVX512DQ-BW-NEXT: vmovdqa 64(%rcx), %ymm0
24923 ; AVX512DQ-BW-NEXT: vmovdqa 64(%rdx), %ymm1
24924 ; AVX512DQ-BW-NEXT: vmovdqa 64(%rsi), %ymm2
24925 ; AVX512DQ-BW-NEXT: vmovdqa 64(%rdi), %ymm4
24926 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} ymm6 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
24927 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm4[0],ymm2[0],ymm4[2],ymm2[2]
24928 ; AVX512DQ-BW-NEXT: vperm2i128 {{.*#+}} ymm6 = ymm7[2,3],ymm6[2,3]
24929 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm6, %zmm31, %zmm6
24930 ; AVX512DQ-BW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
24931 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
24932 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
24933 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm6, %zmm7 {%k1}
24934 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
24935 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm4[1],ymm2[1],ymm4[3],ymm2[3]
24936 ; AVX512DQ-BW-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
24937 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm0, %zmm7, %zmm8
24938 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
24939 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, %zmm27 {%k1}
24940 ; AVX512DQ-BW-NEXT: vmovdqa 128(%rcx), %ymm0
24941 ; AVX512DQ-BW-NEXT: vmovdqa 128(%rdx), %ymm1
24942 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
24943 ; AVX512DQ-BW-NEXT: vmovdqa 128(%rsi), %ymm4
24944 ; AVX512DQ-BW-NEXT: vmovdqa 128(%rdi), %ymm6
24945 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm6[0],ymm4[0],ymm6[2],ymm4[2]
24946 ; AVX512DQ-BW-NEXT: vperm2i128 {{.*#+}} ymm2 = ymm7[2,3],ymm2[2,3]
24947 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm2, %zmm27, %zmm23
24948 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
24949 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
24950 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm7 {%k1}
24951 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
24952 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm6[1],ymm4[1],ymm6[3],ymm4[3]
24953 ; AVX512DQ-BW-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
24954 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm0, %zmm7, %zmm25
24955 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
24956 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, %zmm20 {%k1}
24957 ; AVX512DQ-BW-NEXT: vmovdqa 192(%rcx), %ymm0
24958 ; AVX512DQ-BW-NEXT: vmovdqa 192(%rdx), %ymm1
24959 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
24960 ; AVX512DQ-BW-NEXT: vmovdqa 192(%rsi), %ymm4
24961 ; AVX512DQ-BW-NEXT: vmovdqa 192(%rdi), %ymm6
24962 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm6[0],ymm4[0],ymm6[2],ymm4[2]
24963 ; AVX512DQ-BW-NEXT: vperm2i128 {{.*#+}} ymm2 = ymm7[2,3],ymm2[2,3]
24964 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm2, %zmm20, %zmm21
24965 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
24966 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
24967 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm7 {%k1}
24968 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
24969 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm6[1],ymm4[1],ymm6[3],ymm4[3]
24970 ; AVX512DQ-BW-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
24971 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm0, %zmm7, %zmm19
24972 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
24973 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, %zmm28 {%k1}
24974 ; AVX512DQ-BW-NEXT: vmovdqa 256(%rcx), %ymm0
24975 ; AVX512DQ-BW-NEXT: vmovdqa 256(%rdx), %ymm1
24976 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
24977 ; AVX512DQ-BW-NEXT: vmovdqa 256(%rsi), %ymm4
24978 ; AVX512DQ-BW-NEXT: vmovdqa 256(%rdi), %ymm6
24979 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm6[0],ymm4[0],ymm6[2],ymm4[2]
24980 ; AVX512DQ-BW-NEXT: vperm2i128 {{.*#+}} ymm2 = ymm7[2,3],ymm2[2,3]
24981 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm2, %zmm28, %zmm15
24982 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
24983 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
24984 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm7 {%k1}
24985 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
24986 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm6[1],ymm4[1],ymm6[3],ymm4[3]
24987 ; AVX512DQ-BW-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
24988 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm0, %zmm7, %zmm28
24989 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
24990 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, %zmm22 {%k1}
24991 ; AVX512DQ-BW-NEXT: vmovdqa 320(%rcx), %ymm0
24992 ; AVX512DQ-BW-NEXT: vmovdqa 320(%rdx), %ymm1
24993 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
24994 ; AVX512DQ-BW-NEXT: vmovdqa 320(%rsi), %ymm4
24995 ; AVX512DQ-BW-NEXT: vmovdqa 320(%rdi), %ymm6
24996 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm6[0],ymm4[0],ymm6[2],ymm4[2]
24997 ; AVX512DQ-BW-NEXT: vperm2i128 {{.*#+}} ymm2 = ymm7[2,3],ymm2[2,3]
24998 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm2, %zmm22, %zmm22
24999 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
25000 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm29 {%k1}
25001 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
25002 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm6[1],ymm4[1],ymm6[3],ymm4[3]
25003 ; AVX512DQ-BW-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
25004 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm0, %zmm29, %zmm27
25005 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm30, %zmm17 {%k1}
25006 ; AVX512DQ-BW-NEXT: vmovdqa 384(%rcx), %ymm0
25007 ; AVX512DQ-BW-NEXT: vmovdqa 384(%rdx), %ymm1
25008 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
25009 ; AVX512DQ-BW-NEXT: vmovdqa 384(%rsi), %ymm4
25010 ; AVX512DQ-BW-NEXT: vmovdqa 384(%rdi), %ymm6
25011 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm6[0],ymm4[0],ymm6[2],ymm4[2]
25012 ; AVX512DQ-BW-NEXT: vperm2i128 {{.*#+}} ymm2 = ymm7[2,3],ymm2[2,3]
25013 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm2, %zmm17, %zmm14
25014 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm13, %zmm5 {%k1}
25015 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
25016 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm6[1],ymm4[1],ymm6[3],ymm4[3]
25017 ; AVX512DQ-BW-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
25018 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm0, %zmm5, %zmm13
25019 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm18, %zmm10 {%k1}
25020 ; AVX512DQ-BW-NEXT: vmovdqa 448(%rcx), %ymm0
25021 ; AVX512DQ-BW-NEXT: vmovdqa 448(%rdx), %ymm1
25022 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
25023 ; AVX512DQ-BW-NEXT: vmovdqa 448(%rsi), %ymm4
25024 ; AVX512DQ-BW-NEXT: vmovdqa 448(%rdi), %ymm5
25025 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} ymm6 = ymm5[0],ymm4[0],ymm5[2],ymm4[2]
25026 ; AVX512DQ-BW-NEXT: vperm2i128 {{.*#+}} ymm2 = ymm6[2,3],ymm2[2,3]
25027 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm2, %zmm10, %zmm10
25028 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
25029 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm5[1],ymm4[1],ymm5[3],ymm4[3]
25030 ; AVX512DQ-BW-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
25031 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm16, %zmm3 {%k1}
25032 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm0, %zmm3, %zmm9
25033 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
25034 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
25035 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, %zmm3 {%k1}
25036 ; AVX512DQ-BW-NEXT: vmovdqa (%rsi), %xmm0
25037 ; AVX512DQ-BW-NEXT: vinserti128 $1, (%rcx), %ymm0, %ymm1
25038 ; AVX512DQ-BW-NEXT: vmovdqa (%rdi), %xmm0
25039 ; AVX512DQ-BW-NEXT: vinserti128 $1, (%rdx), %ymm0, %ymm2
25040 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm2[0],ymm1[0],ymm2[2],ymm1[2]
25041 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm0, %zmm3, %zmm0
25042 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
25043 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
25044 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
25045 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm2[1],ymm1[1],ymm2[3],ymm1[3]
25046 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm1, %zmm4, %zmm2
25047 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
25048 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
25049 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm1, %zmm5 {%k1}
25050 ; AVX512DQ-BW-NEXT: vmovdqa 64(%rsi), %xmm1
25051 ; AVX512DQ-BW-NEXT: vinserti128 $1, 64(%rcx), %ymm1, %ymm1
25052 ; AVX512DQ-BW-NEXT: vmovdqa 64(%rdi), %xmm3
25053 ; AVX512DQ-BW-NEXT: vinserti128 $1, 64(%rdx), %ymm3, %ymm4
25054 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm4[0],ymm1[0],ymm4[2],ymm1[2]
25055 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm3, %zmm5, %zmm3
25056 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
25057 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
25058 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm5, %zmm6 {%k1}
25059 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm4[1],ymm1[1],ymm4[3],ymm1[3]
25060 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm1, %zmm6, %zmm5
25061 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
25062 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
25063 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm1, %zmm7 {%k1}
25064 ; AVX512DQ-BW-NEXT: vmovdqa 128(%rsi), %xmm1
25065 ; AVX512DQ-BW-NEXT: vinserti128 $1, 128(%rcx), %ymm1, %ymm1
25066 ; AVX512DQ-BW-NEXT: vmovdqa 128(%rdi), %xmm4
25067 ; AVX512DQ-BW-NEXT: vinserti128 $1, 128(%rdx), %ymm4, %ymm6
25068 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} ymm4 = ymm6[0],ymm1[0],ymm6[2],ymm1[2]
25069 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm4, %zmm7, %zmm4
25070 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
25071 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
25072 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm7, %zmm16 {%k1}
25073 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm6[1],ymm1[1],ymm6[3],ymm1[3]
25074 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm1, %zmm16, %zmm6
25075 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
25076 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
25077 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm1, %zmm17 {%k1}
25078 ; AVX512DQ-BW-NEXT: vmovdqa 192(%rsi), %xmm1
25079 ; AVX512DQ-BW-NEXT: vinserti128 $1, 192(%rcx), %ymm1, %ymm1
25080 ; AVX512DQ-BW-NEXT: vmovdqa 192(%rdi), %xmm7
25081 ; AVX512DQ-BW-NEXT: vinserti32x4 $1, 192(%rdx), %ymm7, %ymm16
25082 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm16[0],ymm1[0],ymm16[2],ymm1[2]
25083 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm7, %zmm17, %zmm7
25084 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
25085 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
25086 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm17, %zmm18 {%k1}
25087 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm16[1],ymm1[1],ymm16[3],ymm1[3]
25088 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm1, %zmm18, %zmm1
25089 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
25090 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
25091 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm16, %zmm20 {%k1}
25092 ; AVX512DQ-BW-NEXT: vmovdqa64 256(%rsi), %xmm16
25093 ; AVX512DQ-BW-NEXT: vinserti32x4 $1, 256(%rcx), %ymm16, %ymm16
25094 ; AVX512DQ-BW-NEXT: vmovdqa64 256(%rdi), %xmm17
25095 ; AVX512DQ-BW-NEXT: vinserti32x4 $1, 256(%rdx), %ymm17, %ymm17
25096 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} ymm18 = ymm17[0],ymm16[0],ymm17[2],ymm16[2]
25097 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm18, %zmm20, %zmm29
25098 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
25099 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
25100 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm18, %zmm20 {%k1}
25101 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} ymm16 = ymm17[1],ymm16[1],ymm17[3],ymm16[3]
25102 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm16, %zmm20, %zmm30
25103 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
25104 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
25105 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm16, %zmm20 {%k1}
25106 ; AVX512DQ-BW-NEXT: vmovdqa64 320(%rsi), %xmm16
25107 ; AVX512DQ-BW-NEXT: vinserti32x4 $1, 320(%rcx), %ymm16, %ymm16
25108 ; AVX512DQ-BW-NEXT: vmovdqa64 320(%rdi), %xmm17
25109 ; AVX512DQ-BW-NEXT: vinserti32x4 $1, 320(%rdx), %ymm17, %ymm17
25110 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} ymm18 = ymm17[0],ymm16[0],ymm17[2],ymm16[2]
25111 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm18, %zmm20, %zmm18
25112 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
25113 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
25114 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm20, %zmm31 {%k1}
25115 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} ymm16 = ymm17[1],ymm16[1],ymm17[3],ymm16[3]
25116 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm16, %zmm31, %zmm16
25117 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
25118 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm17, %zmm24 {%k1}
25119 ; AVX512DQ-BW-NEXT: vmovdqa64 384(%rsi), %xmm17
25120 ; AVX512DQ-BW-NEXT: vinserti32x4 $1, 384(%rcx), %ymm17, %ymm17
25121 ; AVX512DQ-BW-NEXT: vmovdqa64 384(%rdi), %xmm20
25122 ; AVX512DQ-BW-NEXT: vinserti32x4 $1, 384(%rdx), %ymm20, %ymm20
25123 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} ymm31 = ymm20[0],ymm17[0],ymm20[2],ymm17[2]
25124 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm31, %zmm24, %zmm24
25125 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
25126 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm31, %zmm26 {%k1}
25127 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} ymm17 = ymm20[1],ymm17[1],ymm20[3],ymm17[3]
25128 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm17, %zmm26, %zmm17
25129 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
25130 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm20, %zmm11 {%k1}
25131 ; AVX512DQ-BW-NEXT: vmovdqa64 448(%rsi), %xmm20
25132 ; AVX512DQ-BW-NEXT: vinserti32x4 $1, 448(%rcx), %ymm20, %ymm20
25133 ; AVX512DQ-BW-NEXT: vmovdqa64 448(%rdi), %xmm26
25134 ; AVX512DQ-BW-NEXT: vinserti32x4 $1, 448(%rdx), %ymm26, %ymm26
25135 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} ymm31 = ymm26[0],ymm20[0],ymm26[2],ymm20[2]
25136 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm31, %zmm11, %zmm11
25137 ; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
25138 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm31, %zmm12 {%k1}
25139 ; AVX512DQ-BW-NEXT: vpunpckhqdq {{.*#+}} ymm20 = ymm26[1],ymm20[1],ymm26[3],ymm20[3]
25140 ; AVX512DQ-BW-NEXT: vinserti64x4 $0, %ymm20, %zmm12, %zmm12
25141 ; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
25142 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm9, 3776(%rax)
25143 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm10, 3712(%rax)
25144 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm13, 3264(%rax)
25145 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm14, 3200(%rax)
25146 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm27, 2752(%rax)
25147 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm22, 2688(%rax)
25148 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm28, 2240(%rax)
25149 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm15, 2176(%rax)
25150 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm19, 1728(%rax)
25151 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm21, 1664(%rax)
25152 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm25, 1216(%rax)
25153 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm23, 1152(%rax)
25154 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm8, 704(%rax)
25155 ; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
25156 ; AVX512DQ-BW-NEXT: vmovaps %zmm8, 640(%rax)
25157 ; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
25158 ; AVX512DQ-BW-NEXT: vmovaps %zmm8, 192(%rax)
25159 ; AVX512DQ-BW-NEXT: vmovups (%rsp), %zmm8 # 64-byte Reload
25160 ; AVX512DQ-BW-NEXT: vmovaps %zmm8, 128(%rax)
25161 ; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
25162 ; AVX512DQ-BW-NEXT: vmovaps %zmm8, 4032(%rax)
25163 ; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
25164 ; AVX512DQ-BW-NEXT: vmovaps %zmm8, 3968(%rax)
25165 ; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
25166 ; AVX512DQ-BW-NEXT: vmovaps %zmm8, 3904(%rax)
25167 ; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
25168 ; AVX512DQ-BW-NEXT: vmovaps %zmm8, 3840(%rax)
25169 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm12, 3648(%rax)
25170 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm11, 3584(%rax)
25171 ; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
25172 ; AVX512DQ-BW-NEXT: vmovaps %zmm8, 3520(%rax)
25173 ; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
25174 ; AVX512DQ-BW-NEXT: vmovaps %zmm8, 3456(%rax)
25175 ; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
25176 ; AVX512DQ-BW-NEXT: vmovaps %zmm8, 3392(%rax)
25177 ; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
25178 ; AVX512DQ-BW-NEXT: vmovaps %zmm8, 3328(%rax)
25179 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm17, 3136(%rax)
25180 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm24, 3072(%rax)
25181 ; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
25182 ; AVX512DQ-BW-NEXT: vmovaps %zmm8, 3008(%rax)
25183 ; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
25184 ; AVX512DQ-BW-NEXT: vmovaps %zmm8, 2944(%rax)
25185 ; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
25186 ; AVX512DQ-BW-NEXT: vmovaps %zmm8, 2880(%rax)
25187 ; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
25188 ; AVX512DQ-BW-NEXT: vmovaps %zmm8, 2816(%rax)
25189 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm16, 2624(%rax)
25190 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm18, 2560(%rax)
25191 ; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
25192 ; AVX512DQ-BW-NEXT: vmovaps %zmm8, 2496(%rax)
25193 ; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
25194 ; AVX512DQ-BW-NEXT: vmovaps %zmm8, 2432(%rax)
25195 ; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
25196 ; AVX512DQ-BW-NEXT: vmovaps %zmm8, 2368(%rax)
25197 ; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
25198 ; AVX512DQ-BW-NEXT: vmovaps %zmm8, 2304(%rax)
25199 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm30, 2112(%rax)
25200 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm29, 2048(%rax)
25201 ; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
25202 ; AVX512DQ-BW-NEXT: vmovaps %zmm8, 1984(%rax)
25203 ; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
25204 ; AVX512DQ-BW-NEXT: vmovaps %zmm8, 1920(%rax)
25205 ; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
25206 ; AVX512DQ-BW-NEXT: vmovaps %zmm8, 1856(%rax)
25207 ; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
25208 ; AVX512DQ-BW-NEXT: vmovaps %zmm8, 1792(%rax)
25209 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm1, 1600(%rax)
25210 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm7, 1536(%rax)
25211 ; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
25212 ; AVX512DQ-BW-NEXT: vmovaps %zmm1, 1472(%rax)
25213 ; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
25214 ; AVX512DQ-BW-NEXT: vmovaps %zmm1, 1408(%rax)
25215 ; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
25216 ; AVX512DQ-BW-NEXT: vmovaps %zmm1, 1344(%rax)
25217 ; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
25218 ; AVX512DQ-BW-NEXT: vmovaps %zmm1, 1280(%rax)
25219 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm6, 1088(%rax)
25220 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm4, 1024(%rax)
25221 ; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
25222 ; AVX512DQ-BW-NEXT: vmovaps %zmm1, 960(%rax)
25223 ; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
25224 ; AVX512DQ-BW-NEXT: vmovaps %zmm1, 896(%rax)
25225 ; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
25226 ; AVX512DQ-BW-NEXT: vmovaps %zmm1, 832(%rax)
25227 ; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
25228 ; AVX512DQ-BW-NEXT: vmovaps %zmm1, 768(%rax)
25229 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm5, 576(%rax)
25230 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, 512(%rax)
25231 ; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
25232 ; AVX512DQ-BW-NEXT: vmovaps %zmm1, 448(%rax)
25233 ; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
25234 ; AVX512DQ-BW-NEXT: vmovaps %zmm1, 384(%rax)
25235 ; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
25236 ; AVX512DQ-BW-NEXT: vmovaps %zmm1, 320(%rax)
25237 ; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
25238 ; AVX512DQ-BW-NEXT: vmovaps %zmm1, 256(%rax)
25239 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, 64(%rax)
25240 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, (%rax)
25241 ; AVX512DQ-BW-NEXT: addq $5384, %rsp # imm = 0x1508
25242 ; AVX512DQ-BW-NEXT: vzeroupper
25243 ; AVX512DQ-BW-NEXT: retq
25244 ;
25245 ; AVX512DQ-BW-FCP-LABEL: store_i64_stride8_vf64:
25246 ; AVX512DQ-BW-FCP: # %bb.0:
25247 ; AVX512DQ-BW-FCP-NEXT: subq $5384, %rsp # imm = 0x1508
25248 ; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
25249 ; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10
25250 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 128(%rdi), %zmm0
25251 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 64(%rdi), %zmm4
25252 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rdi), %zmm14
25253 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 128(%rsi), %zmm1
25254 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 64(%rsi), %zmm5
25255 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rsi), %zmm10
25256 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 128(%rdx), %zmm2
25257 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 64(%rdx), %zmm7
25258 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rdx), %zmm12
25259 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 128(%rcx), %zmm3
25260 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 64(%rcx), %zmm9
25261 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rcx), %zmm15
25262 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%r8), %zmm30
25263 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 64(%r8), %zmm18
25264 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 128(%r8), %zmm11
25265 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%r9), %zmm24
25266 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 64(%r9), %zmm28
25267 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 128(%r9), %zmm22
25268 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%r10), %zmm26
25269 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 64(%r10), %zmm31
25270 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 128(%r10), %zmm16
25271 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rax), %zmm17
25272 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 64(%rax), %zmm27
25273 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 128(%rax), %zmm13
25274 ; AVX512DQ-BW-FCP-NEXT: movb $-64, %r11b
25275 ; AVX512DQ-BW-FCP-NEXT: kmovd %r11d, %k1
25276 ; AVX512DQ-BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm19 = [4,12,4,12,4,12,4,12]
25277 ; AVX512DQ-BW-FCP-NEXT: # zmm19 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
25278 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm12, %zmm8
25279 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm15, %zmm19, %zmm8
25280 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm14, %zmm6
25281 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm10, %zmm19, %zmm6
25282 ; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm8[4,5,6,7]
25283 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm26, %zmm8
25284 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm17, %zmm19, %zmm8
25285 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm20 = zmm30[0],zmm24[0],zmm30[2],zmm24[2],zmm30[4],zmm24[4],zmm30[6],zmm24[6]
25286 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm8, %zmm20 {%k1}
25287 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm20, %zmm6
25288 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25289 ; AVX512DQ-BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm21 = [5,13,5,13,5,13,5,13]
25290 ; AVX512DQ-BW-FCP-NEXT: # zmm21 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
25291 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm12, %zmm6
25292 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm15, %zmm21, %zmm6
25293 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm14, %zmm8
25294 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm10, %zmm21, %zmm8
25295 ; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm8[0,1,2,3],ymm6[4,5,6,7]
25296 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm26, %zmm8
25297 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm17, %zmm21, %zmm8
25298 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm20 = zmm30[1],zmm24[1],zmm30[3],zmm24[3],zmm30[5],zmm24[5],zmm30[7],zmm24[7]
25299 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm8, %zmm20 {%k1}
25300 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm20, %zmm6
25301 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25302 ; AVX512DQ-BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm23 = [6,14,6,14,6,14,6,14]
25303 ; AVX512DQ-BW-FCP-NEXT: # zmm23 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
25304 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm12, %zmm6
25305 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm15, %zmm23, %zmm6
25306 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm14, %zmm8
25307 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm10, %zmm23, %zmm8
25308 ; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm8[0,1,2,3],ymm6[4,5,6,7]
25309 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm30, %zmm8
25310 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm24, %zmm23, %zmm8
25311 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm8 {%k1} = zmm26[0],zmm17[0],zmm26[2],zmm17[2],zmm26[4],zmm17[4],zmm26[6],zmm17[6]
25312 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm8, %zmm6
25313 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25314 ; AVX512DQ-BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm25 = [7,15,7,15,7,15,7,15]
25315 ; AVX512DQ-BW-FCP-NEXT: # zmm25 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
25316 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm15, %zmm25, %zmm12
25317 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm10, %zmm25, %zmm14
25318 ; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm14[0,1,2,3],ymm12[4,5,6,7]
25319 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm30, %zmm8
25320 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm24, %zmm25, %zmm8
25321 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm8 {%k1} = zmm26[1],zmm17[1],zmm26[3],zmm17[3],zmm26[5],zmm17[5],zmm26[7],zmm17[7]
25322 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm8, %zmm6
25323 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25324 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm7, %zmm6
25325 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm9, %zmm19, %zmm6
25326 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm4, %zmm8
25327 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm5, %zmm19, %zmm8
25328 ; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm8[0,1,2,3],ymm6[4,5,6,7]
25329 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm31, %zmm8
25330 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm27, %zmm19, %zmm8
25331 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm10 = zmm18[0],zmm28[0],zmm18[2],zmm28[2],zmm18[4],zmm28[4],zmm18[6],zmm28[6]
25332 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}
25333 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm10, %zmm6
25334 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25335 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm7, %zmm6
25336 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm9, %zmm21, %zmm6
25337 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm4, %zmm8
25338 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm5, %zmm21, %zmm8
25339 ; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm8[0,1,2,3],ymm6[4,5,6,7]
25340 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm31, %zmm8
25341 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm27, %zmm21, %zmm8
25342 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm10 = zmm18[1],zmm28[1],zmm18[3],zmm28[3],zmm18[5],zmm28[5],zmm18[7],zmm28[7]
25343 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm8, %zmm10 {%k1}
25344 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm10, %zmm6
25345 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25346 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm7, %zmm6
25347 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm9, %zmm23, %zmm6
25348 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm4, %zmm8
25349 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm5, %zmm23, %zmm8
25350 ; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm8[0,1,2,3],ymm6[4,5,6,7]
25351 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm18, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25352 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm18, %zmm8
25353 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm28, %zmm23, %zmm8
25354 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm8 {%k1} = zmm31[0],zmm27[0],zmm31[2],zmm27[2],zmm31[4],zmm27[4],zmm31[6],zmm27[6]
25355 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm8, %zmm6
25356 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25357 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm9, %zmm25, %zmm7
25358 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm5, %zmm25, %zmm4
25359 ; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm7[4,5,6,7]
25360 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm18, %zmm5
25361 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm28, %zmm25, %zmm5
25362 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm5 {%k1} = zmm31[1],zmm27[1],zmm31[3],zmm27[3],zmm31[5],zmm27[5],zmm31[7],zmm27[7]
25363 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm4, %zmm5, %zmm4
25364 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25365 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm4
25366 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm3, %zmm19, %zmm4
25367 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm5
25368 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm1, %zmm19, %zmm5
25369 ; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm5[0,1,2,3],ymm4[4,5,6,7]
25370 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm16, %zmm5
25371 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm13, %zmm19, %zmm5
25372 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm6 = zmm11[0],zmm22[0],zmm11[2],zmm22[2],zmm11[4],zmm22[4],zmm11[6],zmm22[6]
25373 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm5, %zmm6 {%k1}
25374 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm4, %zmm6, %zmm4
25375 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25376 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm4
25377 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm3, %zmm21, %zmm4
25378 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm5
25379 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm1, %zmm21, %zmm5
25380 ; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm5[0,1,2,3],ymm4[4,5,6,7]
25381 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm16, %zmm5
25382 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm13, %zmm21, %zmm5
25383 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm6 = zmm11[1],zmm22[1],zmm11[3],zmm22[3],zmm11[5],zmm22[5],zmm11[7],zmm22[7]
25384 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm5, %zmm6 {%k1}
25385 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm4, %zmm6, %zmm4
25386 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25387 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm4
25388 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm3, %zmm23, %zmm4
25389 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm5
25390 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm1, %zmm23, %zmm5
25391 ; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm5[0,1,2,3],ymm4[4,5,6,7]
25392 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm11, %zmm5
25393 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25394 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm22, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25395 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm22, %zmm23, %zmm5
25396 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25397 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25398 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm5 {%k1} = zmm16[0],zmm13[0],zmm16[2],zmm13[2],zmm16[4],zmm13[4],zmm16[6],zmm13[6]
25399 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm4, %zmm5, %zmm4
25400 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25401 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 192(%rdi), %zmm4
25402 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm3, %zmm25, %zmm2
25403 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 192(%rdx), %zmm3
25404 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm1, %zmm25, %zmm0
25405 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 192(%rcx), %zmm1
25406 ; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
25407 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm11, %zmm2
25408 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm22, %zmm25, %zmm2
25409 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm2 {%k1} = zmm16[1],zmm13[1],zmm16[3],zmm13[3],zmm16[5],zmm13[5],zmm16[7],zmm13[7]
25410 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm2, %zmm0
25411 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25412 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm0
25413 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm1, %zmm19, %zmm0
25414 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 192(%rsi), %zmm2
25415 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm4, %zmm5
25416 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm2, %zmm19, %zmm5
25417 ; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm0[4,5,6,7]
25418 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 192(%r10), %zmm9
25419 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 192(%rax), %zmm20
25420 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm9, %zmm6
25421 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm20, %zmm19, %zmm6
25422 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 192(%r8), %zmm8
25423 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 192(%r9), %zmm0
25424 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm7 = zmm8[0],zmm0[0],zmm8[2],zmm0[2],zmm8[4],zmm0[4],zmm8[6],zmm0[6]
25425 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm7 {%k1}
25426 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm5, %zmm7, %zmm5
25427 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25428 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm5
25429 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm1, %zmm21, %zmm5
25430 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm4, %zmm6
25431 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm2, %zmm21, %zmm6
25432 ; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm5 = ymm6[0,1,2,3],ymm5[4,5,6,7]
25433 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm9, %zmm6
25434 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm20, %zmm21, %zmm6
25435 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm7 = zmm8[1],zmm0[1],zmm8[3],zmm0[3],zmm8[5],zmm0[5],zmm8[7],zmm0[7]
25436 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm7 {%k1}
25437 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm5, %zmm7, %zmm5
25438 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25439 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm5
25440 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm1, %zmm23, %zmm5
25441 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm4, %zmm6
25442 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm2, %zmm23, %zmm6
25443 ; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm5 = ymm6[0,1,2,3],ymm5[4,5,6,7]
25444 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25445 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm8, %zmm6
25446 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm0, %zmm23, %zmm6
25447 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm6 {%k1} = zmm9[0],zmm20[0],zmm9[2],zmm20[2],zmm9[4],zmm20[4],zmm9[6],zmm20[6]
25448 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm5, %zmm6, %zmm5
25449 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25450 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm1, %zmm25, %zmm3
25451 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm2, %zmm25, %zmm4
25452 ; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm1 = ymm4[0,1,2,3],ymm3[4,5,6,7]
25453 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm8, %zmm2
25454 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm0, %zmm25, %zmm2
25455 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm2 {%k1} = zmm9[1],zmm20[1],zmm9[3],zmm20[3],zmm9[5],zmm20[5],zmm9[7],zmm20[7]
25456 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm1, %zmm2, %zmm1
25457 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25458 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 256(%rdx), %zmm1
25459 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 256(%rcx), %zmm3
25460 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm2
25461 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm3, %zmm19, %zmm2
25462 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 256(%rdi), %zmm4
25463 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 256(%rsi), %zmm5
25464 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm4, %zmm6
25465 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm5, %zmm19, %zmm6
25466 ; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm2[4,5,6,7]
25467 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 256(%r10), %zmm13
25468 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 256(%rax), %zmm2
25469 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm13, %zmm7
25470 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm2, %zmm19, %zmm7
25471 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 256(%r8), %zmm18
25472 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 256(%r9), %zmm15
25473 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm8 = zmm18[0],zmm15[0],zmm18[2],zmm15[2],zmm18[4],zmm15[4],zmm18[6],zmm15[6]
25474 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm7, %zmm8 {%k1}
25475 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm8, %zmm6
25476 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25477 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm6
25478 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm3, %zmm21, %zmm6
25479 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm4, %zmm7
25480 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm5, %zmm21, %zmm7
25481 ; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]
25482 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm13, %zmm7
25483 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm2, %zmm21, %zmm7
25484 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm8 = zmm18[1],zmm15[1],zmm18[3],zmm15[3],zmm18[5],zmm15[5],zmm18[7],zmm15[7]
25485 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm7, %zmm8 {%k1}
25486 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm8, %zmm6
25487 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25488 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm6
25489 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm3, %zmm23, %zmm6
25490 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm4, %zmm7
25491 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm5, %zmm23, %zmm7
25492 ; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]
25493 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm18, %zmm7
25494 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm15, %zmm23, %zmm7
25495 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm7 {%k1} = zmm13[0],zmm2[0],zmm13[2],zmm2[2],zmm13[4],zmm2[4],zmm13[6],zmm2[6]
25496 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm7, %zmm6
25497 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25498 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm3, %zmm25, %zmm1
25499 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm5, %zmm25, %zmm4
25500 ; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm1 = ymm4[0,1,2,3],ymm1[4,5,6,7]
25501 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm18, %zmm3
25502 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm15, %zmm25, %zmm3
25503 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm3 {%k1} = zmm13[1],zmm2[1],zmm13[3],zmm2[3],zmm13[5],zmm2[5],zmm13[7],zmm2[7]
25504 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm1, %zmm3, %zmm1
25505 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25506 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 320(%rdx), %zmm1
25507 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 320(%rcx), %zmm4
25508 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm3
25509 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm4, %zmm19, %zmm3
25510 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 320(%rdi), %zmm6
25511 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 320(%rsi), %zmm7
25512 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm5
25513 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm7, %zmm19, %zmm5
25514 ; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm8 = ymm5[0,1,2,3],ymm3[4,5,6,7]
25515 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 320(%r10), %zmm16
25516 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 320(%rax), %zmm22
25517 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm16, %zmm12
25518 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm22, %zmm19, %zmm12
25519 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 320(%r8), %zmm29
25520 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 320(%r9), %zmm5
25521 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm14 = zmm29[0],zmm5[0],zmm29[2],zmm5[2],zmm29[4],zmm5[4],zmm29[6],zmm5[6]
25522 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm12, %zmm14 {%k1}
25523 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm8, %zmm14, %zmm3
25524 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25525 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm8
25526 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm4, %zmm21, %zmm8
25527 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm12
25528 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm7, %zmm21, %zmm12
25529 ; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm8 = ymm12[0,1,2,3],ymm8[4,5,6,7]
25530 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm16, %zmm12
25531 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm22, %zmm21, %zmm12
25532 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm14 = zmm29[1],zmm5[1],zmm29[3],zmm5[3],zmm29[5],zmm5[5],zmm29[7],zmm5[7]
25533 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm12, %zmm14 {%k1}
25534 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm8, %zmm14, %zmm3
25535 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25536 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm8
25537 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm4, %zmm23, %zmm8
25538 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm12
25539 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm7, %zmm23, %zmm12
25540 ; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm8 = ymm12[0,1,2,3],ymm8[4,5,6,7]
25541 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm29, %zmm12
25542 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm5, %zmm23, %zmm12
25543 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm12 {%k1} = zmm16[0],zmm22[0],zmm16[2],zmm22[2],zmm16[4],zmm22[4],zmm16[6],zmm22[6]
25544 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm8, %zmm12, %zmm3
25545 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25546 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm4, %zmm25, %zmm1
25547 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm7, %zmm25, %zmm6
25548 ; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm1 = ymm6[0,1,2,3],ymm1[4,5,6,7]
25549 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm29, %zmm4
25550 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm5, %zmm25, %zmm4
25551 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm4 {%k1} = zmm16[1],zmm22[1],zmm16[3],zmm22[3],zmm16[5],zmm22[5],zmm16[7],zmm22[7]
25552 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm1, %zmm4, %zmm1
25553 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25554 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 384(%rdx), %zmm3
25555 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 384(%rcx), %zmm1
25556 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm4
25557 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm1, %zmm19, %zmm4
25558 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25559 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm4
25560 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm1, %zmm21, %zmm4
25561 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25562 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm4
25563 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm1, %zmm23, %zmm4
25564 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25565 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm1, %zmm25, %zmm3
25566 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25567 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 384(%rdi), %zmm3
25568 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 384(%rsi), %zmm1
25569 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm4
25570 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm1, %zmm19, %zmm4
25571 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25572 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm4
25573 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm1, %zmm21, %zmm4
25574 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25575 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm4
25576 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm1, %zmm23, %zmm4
25577 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25578 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm1, %zmm25, %zmm3
25579 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25580 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 448(%rdx), %zmm3
25581 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 448(%rcx), %zmm1
25582 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm4
25583 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm1, %zmm19, %zmm4
25584 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm4, (%rsp) # 64-byte Spill
25585 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm4
25586 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm1, %zmm21, %zmm4
25587 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25588 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm4
25589 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm1, %zmm23, %zmm4
25590 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25591 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm1, %zmm25, %zmm3
25592 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25593 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 448(%rdi), %zmm3
25594 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 448(%rsi), %zmm1
25595 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm4
25596 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm1, %zmm19, %zmm4
25597 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25598 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm4
25599 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm1, %zmm21, %zmm4
25600 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25601 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm4
25602 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm1, %zmm23, %zmm4
25603 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25604 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm1, %zmm25, %zmm3
25605 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25606 ; AVX512DQ-BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm11 = [0,8,0,8,0,8,0,8]
25607 ; AVX512DQ-BW-FCP-NEXT: # zmm11 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
25608 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm26, %zmm1
25609 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm17, %zmm11, %zmm1
25610 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25611 ; AVX512DQ-BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm12 = [1,9,1,9,1,9,1,9]
25612 ; AVX512DQ-BW-FCP-NEXT: # zmm12 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
25613 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm26, %zmm1
25614 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm17, %zmm12, %zmm1
25615 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25616 ; AVX512DQ-BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm10 = [2,10,2,10,2,10,2,10]
25617 ; AVX512DQ-BW-FCP-NEXT: # zmm10 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
25618 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm26, %zmm1
25619 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm17, %zmm10, %zmm1
25620 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25621 ; AVX512DQ-BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm4 = [3,11,3,11,3,11,3,11]
25622 ; AVX512DQ-BW-FCP-NEXT: # zmm4 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
25623 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm17, %zmm4, %zmm26
25624 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm26, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25625 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm30, %zmm1
25626 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm24, %zmm11, %zmm1
25627 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25628 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm30, %zmm1
25629 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm24, %zmm12, %zmm1
25630 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25631 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm30, %zmm1
25632 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm24, %zmm10, %zmm1
25633 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25634 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm24, %zmm4, %zmm30
25635 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm30, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25636 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm31, %zmm1
25637 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm27, %zmm11, %zmm1
25638 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25639 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm31, %zmm1
25640 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm27, %zmm12, %zmm1
25641 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25642 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm31, %zmm1
25643 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm27, %zmm10, %zmm1
25644 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25645 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm27, %zmm4, %zmm31
25646 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm31, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25647 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
25648 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm1
25649 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm28, %zmm11, %zmm1
25650 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25651 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm1
25652 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm28, %zmm12, %zmm1
25653 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25654 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm31
25655 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm28, %zmm10, %zmm31
25656 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm28, %zmm4, %zmm3
25657 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25658 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm28 # 64-byte Reload
25659 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm28, %zmm3
25660 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
25661 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm1, %zmm11, %zmm3
25662 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25663 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm28, %zmm3
25664 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm1, %zmm12, %zmm3
25665 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25666 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm28, %zmm3
25667 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm1, %zmm10, %zmm3
25668 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25669 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm1, %zmm4, %zmm28
25670 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm28, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25671 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
25672 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm3
25673 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
25674 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm1, %zmm11, %zmm3
25675 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25676 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm3
25677 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm1, %zmm12, %zmm3
25678 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25679 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm27
25680 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm1, %zmm10, %zmm27
25681 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm1, %zmm4, %zmm6
25682 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25683 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm9, %zmm1
25684 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm20, %zmm11, %zmm1
25685 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25686 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm9, %zmm1
25687 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm20, %zmm12, %zmm1
25688 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25689 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm9, %zmm1
25690 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm20, %zmm10, %zmm1
25691 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25692 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm20, %zmm4, %zmm9
25693 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25694 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
25695 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm3
25696 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm0, %zmm11, %zmm3
25697 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25698 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm3
25699 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm0, %zmm12, %zmm3
25700 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25701 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm20
25702 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm0, %zmm10, %zmm20
25703 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm0, %zmm4, %zmm1
25704 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25705 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm13, %zmm0
25706 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm2, %zmm11, %zmm0
25707 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25708 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm13, %zmm0
25709 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm2, %zmm12, %zmm0
25710 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25711 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm13, %zmm0
25712 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm2, %zmm10, %zmm0
25713 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25714 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm2, %zmm4, %zmm13
25715 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25716 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm18, %zmm0
25717 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm15, %zmm11, %zmm0
25718 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25719 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm18, %zmm0
25720 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm15, %zmm12, %zmm0
25721 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25722 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm18, %zmm28
25723 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm15, %zmm10, %zmm28
25724 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm15, %zmm4, %zmm18
25725 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm18, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25726 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm16, %zmm0
25727 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm22, %zmm11, %zmm0
25728 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25729 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm16, %zmm0
25730 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm22, %zmm12, %zmm0
25731 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25732 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm16, %zmm0
25733 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm22, %zmm10, %zmm0
25734 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25735 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm22, %zmm4, %zmm16
25736 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25737 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm29, %zmm0
25738 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm5, %zmm11, %zmm0
25739 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25740 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm29, %zmm0
25741 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm5, %zmm12, %zmm0
25742 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25743 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm29, %zmm22
25744 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm5, %zmm10, %zmm22
25745 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm5, %zmm4, %zmm29
25746 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 384(%r10), %zmm13
25747 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 384(%rax), %zmm0
25748 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm13, %zmm1
25749 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm0, %zmm11, %zmm1
25750 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25751 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm13, %zmm1
25752 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm0, %zmm12, %zmm1
25753 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25754 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm13, %zmm14
25755 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm0, %zmm19, %zmm14
25756 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm13, %zmm15
25757 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm0, %zmm21, %zmm15
25758 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 384(%r8), %zmm5
25759 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 384(%r9), %zmm2
25760 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm5, %zmm8
25761 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm2, %zmm23, %zmm8
25762 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm8 {%k1} = zmm13[0],zmm0[0],zmm13[2],zmm0[2],zmm13[4],zmm0[4],zmm13[6],zmm0[6]
25763 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm5, %zmm6
25764 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm2, %zmm25, %zmm6
25765 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm6 {%k1} = zmm13[1],zmm0[1],zmm13[3],zmm0[3],zmm13[5],zmm0[5],zmm13[7],zmm0[7]
25766 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm13, %zmm30
25767 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm0, %zmm10, %zmm30
25768 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm0, %zmm4, %zmm13
25769 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm5, %zmm24
25770 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm2, %zmm11, %zmm24
25771 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm5, %zmm26
25772 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm2, %zmm12, %zmm26
25773 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm5, %zmm17
25774 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm2, %zmm10, %zmm17
25775 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm7 = zmm5[0],zmm2[0],zmm5[2],zmm2[2],zmm5[4],zmm2[4],zmm5[6],zmm2[6]
25776 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm9 = zmm5[1],zmm2[1],zmm5[3],zmm2[3],zmm5[5],zmm2[5],zmm5[7],zmm2[7]
25777 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm2, %zmm4, %zmm5
25778 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 448(%r10), %zmm16
25779 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 448(%rax), %zmm1
25780 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm16, %zmm0
25781 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm1, %zmm11, %zmm0
25782 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25783 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm16, %zmm0
25784 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm1, %zmm12, %zmm0
25785 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25786 ; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm1, %zmm16, %zmm19
25787 ; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm1, %zmm16, %zmm21
25788 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 448(%r8), %zmm3
25789 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 448(%r9), %zmm0
25790 ; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm0, %zmm3, %zmm23
25791 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm23 {%k1} = zmm16[0],zmm1[0],zmm16[2],zmm1[2],zmm16[4],zmm1[4],zmm16[6],zmm1[6]
25792 ; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm0, %zmm3, %zmm25
25793 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm25 {%k1} = zmm16[1],zmm1[1],zmm16[3],zmm1[3],zmm16[5],zmm1[5],zmm16[7],zmm1[7]
25794 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm16, %zmm18
25795 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm1, %zmm10, %zmm18
25796 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm1, %zmm4, %zmm16
25797 ; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm0, %zmm3, %zmm11
25798 ; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm0, %zmm3, %zmm12
25799 ; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm0, %zmm3, %zmm10
25800 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} zmm1 = zmm3[0],zmm0[0],zmm3[2],zmm0[2],zmm3[4],zmm0[4],zmm3[6],zmm0[6]
25801 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} zmm2 = zmm3[1],zmm0[1],zmm3[3],zmm0[3],zmm3[5],zmm0[5],zmm3[7],zmm0[7]
25802 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm0, %zmm4, %zmm3
25803 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
25804 ; AVX512DQ-BW-FCP-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
25805 ; AVX512DQ-BW-FCP-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
25806 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm14, %zmm7 {%k1}
25807 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm7, %zmm0
25808 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25809 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
25810 ; AVX512DQ-BW-FCP-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
25811 ; AVX512DQ-BW-FCP-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
25812 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm15, %zmm9 {%k1}
25813 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm9, %zmm0
25814 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25815 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
25816 ; AVX512DQ-BW-FCP-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
25817 ; AVX512DQ-BW-FCP-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
25818 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm8, %zmm0
25819 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25820 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
25821 ; AVX512DQ-BW-FCP-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
25822 ; AVX512DQ-BW-FCP-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
25823 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm6, %zmm0
25824 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25825 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
25826 ; AVX512DQ-BW-FCP-NEXT: vpblendd $240, (%rsp), %ymm0, %ymm0 # 32-byte Folded Reload
25827 ; AVX512DQ-BW-FCP-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
25828 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm19, %zmm1 {%k1}
25829 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm1, %zmm0
25830 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25831 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
25832 ; AVX512DQ-BW-FCP-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
25833 ; AVX512DQ-BW-FCP-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
25834 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm21, %zmm2 {%k1}
25835 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm2, %zmm0
25836 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25837 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
25838 ; AVX512DQ-BW-FCP-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
25839 ; AVX512DQ-BW-FCP-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
25840 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm23, %zmm0
25841 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25842 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
25843 ; AVX512DQ-BW-FCP-NEXT: vpblendd $240, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
25844 ; AVX512DQ-BW-FCP-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4,5,6,7]
25845 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm25, %zmm0
25846 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25847 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
25848 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
25849 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm8 {%k1}
25850 ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rcx), %ymm0
25851 ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdx), %ymm1
25852 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
25853 ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rsi), %ymm4
25854 ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdi), %ymm6
25855 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm6[0],ymm4[0],ymm6[2],ymm4[2]
25856 ; AVX512DQ-BW-FCP-NEXT: vperm2i128 {{.*#+}} ymm2 = ymm7[2,3],ymm2[2,3]
25857 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm8, %zmm2
25858 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm2, (%rsp) # 64-byte Spill
25859 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
25860 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
25861 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm7 {%k1}
25862 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
25863 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm6[1],ymm4[1],ymm6[3],ymm4[3]
25864 ; AVX512DQ-BW-FCP-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
25865 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm7, %zmm0
25866 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25867 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
25868 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm31 {%k1}
25869 ; AVX512DQ-BW-FCP-NEXT: vmovdqa 64(%rcx), %ymm0
25870 ; AVX512DQ-BW-FCP-NEXT: vmovdqa 64(%rdx), %ymm1
25871 ; AVX512DQ-BW-FCP-NEXT: vmovdqa 64(%rsi), %ymm2
25872 ; AVX512DQ-BW-FCP-NEXT: vmovdqa 64(%rdi), %ymm4
25873 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm6 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
25874 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm4[0],ymm2[0],ymm4[2],ymm2[2]
25875 ; AVX512DQ-BW-FCP-NEXT: vperm2i128 {{.*#+}} ymm6 = ymm7[2,3],ymm6[2,3]
25876 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm6, %zmm31, %zmm6
25877 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
25878 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
25879 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
25880 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm7 {%k1}
25881 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
25882 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm4[1],ymm2[1],ymm4[3],ymm2[3]
25883 ; AVX512DQ-BW-FCP-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
25884 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm7, %zmm8
25885 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
25886 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm27 {%k1}
25887 ; AVX512DQ-BW-FCP-NEXT: vmovdqa 128(%rcx), %ymm0
25888 ; AVX512DQ-BW-FCP-NEXT: vmovdqa 128(%rdx), %ymm1
25889 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
25890 ; AVX512DQ-BW-FCP-NEXT: vmovdqa 128(%rsi), %ymm4
25891 ; AVX512DQ-BW-FCP-NEXT: vmovdqa 128(%rdi), %ymm6
25892 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm6[0],ymm4[0],ymm6[2],ymm4[2]
25893 ; AVX512DQ-BW-FCP-NEXT: vperm2i128 {{.*#+}} ymm2 = ymm7[2,3],ymm2[2,3]
25894 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm27, %zmm23
25895 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
25896 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
25897 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm7 {%k1}
25898 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
25899 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm6[1],ymm4[1],ymm6[3],ymm4[3]
25900 ; AVX512DQ-BW-FCP-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
25901 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm7, %zmm25
25902 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
25903 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm20 {%k1}
25904 ; AVX512DQ-BW-FCP-NEXT: vmovdqa 192(%rcx), %ymm0
25905 ; AVX512DQ-BW-FCP-NEXT: vmovdqa 192(%rdx), %ymm1
25906 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
25907 ; AVX512DQ-BW-FCP-NEXT: vmovdqa 192(%rsi), %ymm4
25908 ; AVX512DQ-BW-FCP-NEXT: vmovdqa 192(%rdi), %ymm6
25909 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm6[0],ymm4[0],ymm6[2],ymm4[2]
25910 ; AVX512DQ-BW-FCP-NEXT: vperm2i128 {{.*#+}} ymm2 = ymm7[2,3],ymm2[2,3]
25911 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm20, %zmm21
25912 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
25913 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
25914 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm7 {%k1}
25915 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
25916 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm6[1],ymm4[1],ymm6[3],ymm4[3]
25917 ; AVX512DQ-BW-FCP-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
25918 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm7, %zmm19
25919 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
25920 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm28 {%k1}
25921 ; AVX512DQ-BW-FCP-NEXT: vmovdqa 256(%rcx), %ymm0
25922 ; AVX512DQ-BW-FCP-NEXT: vmovdqa 256(%rdx), %ymm1
25923 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
25924 ; AVX512DQ-BW-FCP-NEXT: vmovdqa 256(%rsi), %ymm4
25925 ; AVX512DQ-BW-FCP-NEXT: vmovdqa 256(%rdi), %ymm6
25926 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm6[0],ymm4[0],ymm6[2],ymm4[2]
25927 ; AVX512DQ-BW-FCP-NEXT: vperm2i128 {{.*#+}} ymm2 = ymm7[2,3],ymm2[2,3]
25928 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm28, %zmm15
25929 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
25930 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
25931 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm7 {%k1}
25932 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
25933 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm6[1],ymm4[1],ymm6[3],ymm4[3]
25934 ; AVX512DQ-BW-FCP-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
25935 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm7, %zmm28
25936 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
25937 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm22 {%k1}
25938 ; AVX512DQ-BW-FCP-NEXT: vmovdqa 320(%rcx), %ymm0
25939 ; AVX512DQ-BW-FCP-NEXT: vmovdqa 320(%rdx), %ymm1
25940 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
25941 ; AVX512DQ-BW-FCP-NEXT: vmovdqa 320(%rsi), %ymm4
25942 ; AVX512DQ-BW-FCP-NEXT: vmovdqa 320(%rdi), %ymm6
25943 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm6[0],ymm4[0],ymm6[2],ymm4[2]
25944 ; AVX512DQ-BW-FCP-NEXT: vperm2i128 {{.*#+}} ymm2 = ymm7[2,3],ymm2[2,3]
25945 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm22, %zmm22
25946 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
25947 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm29 {%k1}
25948 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
25949 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm6[1],ymm4[1],ymm6[3],ymm4[3]
25950 ; AVX512DQ-BW-FCP-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
25951 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm29, %zmm27
25952 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm30, %zmm17 {%k1}
25953 ; AVX512DQ-BW-FCP-NEXT: vmovdqa 384(%rcx), %ymm0
25954 ; AVX512DQ-BW-FCP-NEXT: vmovdqa 384(%rdx), %ymm1
25955 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
25956 ; AVX512DQ-BW-FCP-NEXT: vmovdqa 384(%rsi), %ymm4
25957 ; AVX512DQ-BW-FCP-NEXT: vmovdqa 384(%rdi), %ymm6
25958 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm6[0],ymm4[0],ymm6[2],ymm4[2]
25959 ; AVX512DQ-BW-FCP-NEXT: vperm2i128 {{.*#+}} ymm2 = ymm7[2,3],ymm2[2,3]
25960 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm17, %zmm14
25961 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm13, %zmm5 {%k1}
25962 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
25963 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm6[1],ymm4[1],ymm6[3],ymm4[3]
25964 ; AVX512DQ-BW-FCP-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
25965 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm5, %zmm13
25966 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm18, %zmm10 {%k1}
25967 ; AVX512DQ-BW-FCP-NEXT: vmovdqa 448(%rcx), %ymm0
25968 ; AVX512DQ-BW-FCP-NEXT: vmovdqa 448(%rdx), %ymm1
25969 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm2 = ymm1[0],ymm0[0],ymm1[2],ymm0[2]
25970 ; AVX512DQ-BW-FCP-NEXT: vmovdqa 448(%rsi), %ymm4
25971 ; AVX512DQ-BW-FCP-NEXT: vmovdqa 448(%rdi), %ymm5
25972 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm6 = ymm5[0],ymm4[0],ymm5[2],ymm4[2]
25973 ; AVX512DQ-BW-FCP-NEXT: vperm2i128 {{.*#+}} ymm2 = ymm6[2,3],ymm2[2,3]
25974 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm2, %zmm10, %zmm10
25975 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm0 = ymm1[1],ymm0[1],ymm1[3],ymm0[3]
25976 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm5[1],ymm4[1],ymm5[3],ymm4[3]
25977 ; AVX512DQ-BW-FCP-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm1[2,3],ymm0[2,3]
25978 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm16, %zmm3 {%k1}
25979 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm3, %zmm9
25980 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload
25981 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
25982 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm3 {%k1}
25983 ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rsi), %xmm0
25984 ; AVX512DQ-BW-FCP-NEXT: vinserti128 $1, (%rcx), %ymm0, %ymm1
25985 ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdi), %xmm0
25986 ; AVX512DQ-BW-FCP-NEXT: vinserti128 $1, (%rdx), %ymm0, %ymm2
25987 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm2[0],ymm1[0],ymm2[2],ymm1[2]
25988 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm0, %zmm3, %zmm0
25989 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload
25990 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload
25991 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm4 {%k1}
25992 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm2[1],ymm1[1],ymm2[3],ymm1[3]
25993 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm1, %zmm4, %zmm2
25994 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
25995 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
25996 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm5 {%k1}
25997 ; AVX512DQ-BW-FCP-NEXT: vmovdqa 64(%rsi), %xmm1
25998 ; AVX512DQ-BW-FCP-NEXT: vinserti128 $1, 64(%rcx), %ymm1, %ymm1
25999 ; AVX512DQ-BW-FCP-NEXT: vmovdqa 64(%rdi), %xmm3
26000 ; AVX512DQ-BW-FCP-NEXT: vinserti128 $1, 64(%rdx), %ymm3, %ymm4
26001 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm3 = ymm4[0],ymm1[0],ymm4[2],ymm1[2]
26002 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm3, %zmm5, %zmm3
26003 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Reload
26004 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload
26005 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm5, %zmm6 {%k1}
26006 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm4[1],ymm1[1],ymm4[3],ymm1[3]
26007 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm1, %zmm6, %zmm5
26008 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
26009 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
26010 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm7 {%k1}
26011 ; AVX512DQ-BW-FCP-NEXT: vmovdqa 128(%rsi), %xmm1
26012 ; AVX512DQ-BW-FCP-NEXT: vinserti128 $1, 128(%rcx), %ymm1, %ymm1
26013 ; AVX512DQ-BW-FCP-NEXT: vmovdqa 128(%rdi), %xmm4
26014 ; AVX512DQ-BW-FCP-NEXT: vinserti128 $1, 128(%rdx), %ymm4, %ymm6
26015 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm4 = ymm6[0],ymm1[0],ymm6[2],ymm1[2]
26016 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm4, %zmm7, %zmm4
26017 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload
26018 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
26019 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm7, %zmm16 {%k1}
26020 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm6[1],ymm1[1],ymm6[3],ymm1[3]
26021 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm1, %zmm16, %zmm6
26022 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
26023 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
26024 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm17 {%k1}
26025 ; AVX512DQ-BW-FCP-NEXT: vmovdqa 192(%rsi), %xmm1
26026 ; AVX512DQ-BW-FCP-NEXT: vinserti128 $1, 192(%rcx), %ymm1, %ymm1
26027 ; AVX512DQ-BW-FCP-NEXT: vmovdqa 192(%rdi), %xmm7
26028 ; AVX512DQ-BW-FCP-NEXT: vinserti32x4 $1, 192(%rdx), %ymm7, %ymm16
26029 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm7 = ymm16[0],ymm1[0],ymm16[2],ymm1[2]
26030 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm7, %zmm17, %zmm7
26031 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
26032 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
26033 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm17, %zmm18 {%k1}
26034 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm1 = ymm16[1],ymm1[1],ymm16[3],ymm1[3]
26035 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm1, %zmm18, %zmm1
26036 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
26037 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
26038 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm16, %zmm20 {%k1}
26039 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 256(%rsi), %xmm16
26040 ; AVX512DQ-BW-FCP-NEXT: vinserti32x4 $1, 256(%rcx), %ymm16, %ymm16
26041 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 256(%rdi), %xmm17
26042 ; AVX512DQ-BW-FCP-NEXT: vinserti32x4 $1, 256(%rdx), %ymm17, %ymm17
26043 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm18 = ymm17[0],ymm16[0],ymm17[2],ymm16[2]
26044 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm18, %zmm20, %zmm29
26045 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm18 # 64-byte Reload
26046 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
26047 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm18, %zmm20 {%k1}
26048 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm16 = ymm17[1],ymm16[1],ymm17[3],ymm16[3]
26049 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm16, %zmm20, %zmm30
26050 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload
26051 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
26052 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm16, %zmm20 {%k1}
26053 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 320(%rsi), %xmm16
26054 ; AVX512DQ-BW-FCP-NEXT: vinserti32x4 $1, 320(%rcx), %ymm16, %ymm16
26055 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 320(%rdi), %xmm17
26056 ; AVX512DQ-BW-FCP-NEXT: vinserti32x4 $1, 320(%rdx), %ymm17, %ymm17
26057 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm18 = ymm17[0],ymm16[0],ymm17[2],ymm16[2]
26058 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm18, %zmm20, %zmm18
26059 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
26060 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
26061 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm20, %zmm31 {%k1}
26062 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm16 = ymm17[1],ymm16[1],ymm17[3],ymm16[3]
26063 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm16, %zmm31, %zmm16
26064 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm17 # 64-byte Reload
26065 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm17, %zmm24 {%k1}
26066 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 384(%rsi), %xmm17
26067 ; AVX512DQ-BW-FCP-NEXT: vinserti32x4 $1, 384(%rcx), %ymm17, %ymm17
26068 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 384(%rdi), %xmm20
26069 ; AVX512DQ-BW-FCP-NEXT: vinserti32x4 $1, 384(%rdx), %ymm20, %ymm20
26070 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm31 = ymm20[0],ymm17[0],ymm20[2],ymm17[2]
26071 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm31, %zmm24, %zmm24
26072 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
26073 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm31, %zmm26 {%k1}
26074 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm17 = ymm20[1],ymm17[1],ymm20[3],ymm17[3]
26075 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm17, %zmm26, %zmm17
26076 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm20 # 64-byte Reload
26077 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm20, %zmm11 {%k1}
26078 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 448(%rsi), %xmm20
26079 ; AVX512DQ-BW-FCP-NEXT: vinserti32x4 $1, 448(%rcx), %ymm20, %ymm20
26080 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 448(%rdi), %xmm26
26081 ; AVX512DQ-BW-FCP-NEXT: vinserti32x4 $1, 448(%rdx), %ymm26, %ymm26
26082 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm31 = ymm26[0],ymm20[0],ymm26[2],ymm20[2]
26083 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm31, %zmm11, %zmm11
26084 ; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload
26085 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm31, %zmm12 {%k1}
26086 ; AVX512DQ-BW-FCP-NEXT: vpunpckhqdq {{.*#+}} ymm20 = ymm26[1],ymm20[1],ymm26[3],ymm20[3]
26087 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $0, %ymm20, %zmm12, %zmm12
26088 ; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax
26089 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm9, 3776(%rax)
26090 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm10, 3712(%rax)
26091 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm13, 3264(%rax)
26092 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm14, 3200(%rax)
26093 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm27, 2752(%rax)
26094 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm22, 2688(%rax)
26095 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm28, 2240(%rax)
26096 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm15, 2176(%rax)
26097 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm19, 1728(%rax)
26098 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm21, 1664(%rax)
26099 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm25, 1216(%rax)
26100 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm23, 1152(%rax)
26101 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm8, 704(%rax)
26102 ; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
26103 ; AVX512DQ-BW-FCP-NEXT: vmovaps %zmm8, 640(%rax)
26104 ; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
26105 ; AVX512DQ-BW-FCP-NEXT: vmovaps %zmm8, 192(%rax)
26106 ; AVX512DQ-BW-FCP-NEXT: vmovups (%rsp), %zmm8 # 64-byte Reload
26107 ; AVX512DQ-BW-FCP-NEXT: vmovaps %zmm8, 128(%rax)
26108 ; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
26109 ; AVX512DQ-BW-FCP-NEXT: vmovaps %zmm8, 4032(%rax)
26110 ; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
26111 ; AVX512DQ-BW-FCP-NEXT: vmovaps %zmm8, 3968(%rax)
26112 ; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
26113 ; AVX512DQ-BW-FCP-NEXT: vmovaps %zmm8, 3904(%rax)
26114 ; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
26115 ; AVX512DQ-BW-FCP-NEXT: vmovaps %zmm8, 3840(%rax)
26116 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm12, 3648(%rax)
26117 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm11, 3584(%rax)
26118 ; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
26119 ; AVX512DQ-BW-FCP-NEXT: vmovaps %zmm8, 3520(%rax)
26120 ; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
26121 ; AVX512DQ-BW-FCP-NEXT: vmovaps %zmm8, 3456(%rax)
26122 ; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
26123 ; AVX512DQ-BW-FCP-NEXT: vmovaps %zmm8, 3392(%rax)
26124 ; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
26125 ; AVX512DQ-BW-FCP-NEXT: vmovaps %zmm8, 3328(%rax)
26126 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm17, 3136(%rax)
26127 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm24, 3072(%rax)
26128 ; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
26129 ; AVX512DQ-BW-FCP-NEXT: vmovaps %zmm8, 3008(%rax)
26130 ; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
26131 ; AVX512DQ-BW-FCP-NEXT: vmovaps %zmm8, 2944(%rax)
26132 ; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
26133 ; AVX512DQ-BW-FCP-NEXT: vmovaps %zmm8, 2880(%rax)
26134 ; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
26135 ; AVX512DQ-BW-FCP-NEXT: vmovaps %zmm8, 2816(%rax)
26136 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm16, 2624(%rax)
26137 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm18, 2560(%rax)
26138 ; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
26139 ; AVX512DQ-BW-FCP-NEXT: vmovaps %zmm8, 2496(%rax)
26140 ; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
26141 ; AVX512DQ-BW-FCP-NEXT: vmovaps %zmm8, 2432(%rax)
26142 ; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
26143 ; AVX512DQ-BW-FCP-NEXT: vmovaps %zmm8, 2368(%rax)
26144 ; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
26145 ; AVX512DQ-BW-FCP-NEXT: vmovaps %zmm8, 2304(%rax)
26146 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm30, 2112(%rax)
26147 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm29, 2048(%rax)
26148 ; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
26149 ; AVX512DQ-BW-FCP-NEXT: vmovaps %zmm8, 1984(%rax)
26150 ; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
26151 ; AVX512DQ-BW-FCP-NEXT: vmovaps %zmm8, 1920(%rax)
26152 ; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
26153 ; AVX512DQ-BW-FCP-NEXT: vmovaps %zmm8, 1856(%rax)
26154 ; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload
26155 ; AVX512DQ-BW-FCP-NEXT: vmovaps %zmm8, 1792(%rax)
26156 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm1, 1600(%rax)
26157 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm7, 1536(%rax)
26158 ; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
26159 ; AVX512DQ-BW-FCP-NEXT: vmovaps %zmm1, 1472(%rax)
26160 ; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
26161 ; AVX512DQ-BW-FCP-NEXT: vmovaps %zmm1, 1408(%rax)
26162 ; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
26163 ; AVX512DQ-BW-FCP-NEXT: vmovaps %zmm1, 1344(%rax)
26164 ; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
26165 ; AVX512DQ-BW-FCP-NEXT: vmovaps %zmm1, 1280(%rax)
26166 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm6, 1088(%rax)
26167 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm4, 1024(%rax)
26168 ; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
26169 ; AVX512DQ-BW-FCP-NEXT: vmovaps %zmm1, 960(%rax)
26170 ; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
26171 ; AVX512DQ-BW-FCP-NEXT: vmovaps %zmm1, 896(%rax)
26172 ; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
26173 ; AVX512DQ-BW-FCP-NEXT: vmovaps %zmm1, 832(%rax)
26174 ; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
26175 ; AVX512DQ-BW-FCP-NEXT: vmovaps %zmm1, 768(%rax)
26176 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm5, 576(%rax)
26177 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, 512(%rax)
26178 ; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
26179 ; AVX512DQ-BW-FCP-NEXT: vmovaps %zmm1, 448(%rax)
26180 ; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
26181 ; AVX512DQ-BW-FCP-NEXT: vmovaps %zmm1, 384(%rax)
26182 ; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
26183 ; AVX512DQ-BW-FCP-NEXT: vmovaps %zmm1, 320(%rax)
26184 ; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
26185 ; AVX512DQ-BW-FCP-NEXT: vmovaps %zmm1, 256(%rax)
26186 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, 64(%rax)
26187 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, (%rax)
26188 ; AVX512DQ-BW-FCP-NEXT: addq $5384, %rsp # imm = 0x1508
26189 ; AVX512DQ-BW-FCP-NEXT: vzeroupper
26190 ; AVX512DQ-BW-FCP-NEXT: retq
26191 %in.vec0 = load <64 x i64>, ptr %in.vecptr0, align 64
26192 %in.vec1 = load <64 x i64>, ptr %in.vecptr1, align 64
26193 %in.vec2 = load <64 x i64>, ptr %in.vecptr2, align 64
26194 %in.vec3 = load <64 x i64>, ptr %in.vecptr3, align 64
26195 %in.vec4 = load <64 x i64>, ptr %in.vecptr4, align 64
26196 %in.vec5 = load <64 x i64>, ptr %in.vecptr5, align 64
26197 %in.vec6 = load <64 x i64>, ptr %in.vecptr6, align 64
26198 %in.vec7 = load <64 x i64>, ptr %in.vecptr7, align 64
26199 %1 = shufflevector <64 x i64> %in.vec0, <64 x i64> %in.vec1, <128 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 64, i32 65, i32 66, i32 67, i32 68, i32 69, i32 70, i32 71, i32 72, i32 73, i32 74, i32 75, i32 76, i32 77, i32 78, i32 79, i32 80, i32 81, i32 82, i32 83, i32 84, i32 85, i32 86, i32 87, i32 88, i32 89, i32 90, i32 91, i32 92, i32 93, i32 94, i32 95, i32 96, i32 97, i32 98, i32 99, i32 100, i32 101, i32 102, i32 103, i32 104, i32 105, i32 106, i32 107, i32 108, i32 109, i32 110, i32 111, i32 112, i32 113, i32 114, i32 115, i32 116, i32 117, i32 118, i32 119, i32 120, i32 121, i32 122, i32 123, i32 124, i32 125, i32 126, i32 127>
26200 %2 = shufflevector <64 x i64> %in.vec2, <64 x i64> %in.vec3, <128 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 64, i32 65, i32 66, i32 67, i32 68, i32 69, i32 70, i32 71, i32 72, i32 73, i32 74, i32 75, i32 76, i32 77, i32 78, i32 79, i32 80, i32 81, i32 82, i32 83, i32 84, i32 85, i32 86, i32 87, i32 88, i32 89, i32 90, i32 91, i32 92, i32 93, i32 94, i32 95, i32 96, i32 97, i32 98, i32 99, i32 100, i32 101, i32 102, i32 103, i32 104, i32 105, i32 106, i32 107, i32 108, i32 109, i32 110, i32 111, i32 112, i32 113, i32 114, i32 115, i32 116, i32 117, i32 118, i32 119, i32 120, i32 121, i32 122, i32 123, i32 124, i32 125, i32 126, i32 127>
26201 %3 = shufflevector <64 x i64> %in.vec4, <64 x i64> %in.vec5, <128 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 64, i32 65, i32 66, i32 67, i32 68, i32 69, i32 70, i32 71, i32 72, i32 73, i32 74, i32 75, i32 76, i32 77, i32 78, i32 79, i32 80, i32 81, i32 82, i32 83, i32 84, i32 85, i32 86, i32 87, i32 88, i32 89, i32 90, i32 91, i32 92, i32 93, i32 94, i32 95, i32 96, i32 97, i32 98, i32 99, i32 100, i32 101, i32 102, i32 103, i32 104, i32 105, i32 106, i32 107, i32 108, i32 109, i32 110, i32 111, i32 112, i32 113, i32 114, i32 115, i32 116, i32 117, i32 118, i32 119, i32 120, i32 121, i32 122, i32 123, i32 124, i32 125, i32 126, i32 127>
26202 %4 = shufflevector <64 x i64> %in.vec6, <64 x i64> %in.vec7, <128 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 64, i32 65, i32 66, i32 67, i32 68, i32 69, i32 70, i32 71, i32 72, i32 73, i32 74, i32 75, i32 76, i32 77, i32 78, i32 79, i32 80, i32 81, i32 82, i32 83, i32 84, i32 85, i32 86, i32 87, i32 88, i32 89, i32 90, i32 91, i32 92, i32 93, i32 94, i32 95, i32 96, i32 97, i32 98, i32 99, i32 100, i32 101, i32 102, i32 103, i32 104, i32 105, i32 106, i32 107, i32 108, i32 109, i32 110, i32 111, i32 112, i32 113, i32 114, i32 115, i32 116, i32 117, i32 118, i32 119, i32 120, i32 121, i32 122, i32 123, i32 124, i32 125, i32 126, i32 127>
26203 %5 = shufflevector <128 x i64> %1, <128 x i64> %2, <256 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 64, i32 65, i32 66, i32 67, i32 68, i32 69, i32 70, i32 71, i32 72, i32 73, i32 74, i32 75, i32 76, i32 77, i32 78, i32 79, i32 80, i32 81, i32 82, i32 83, i32 84, i32 85, i32 86, i32 87, i32 88, i32 89, i32 90, i32 91, i32 92, i32 93, i32 94, i32 95, i32 96, i32 97, i32 98, i32 99, i32 100, i32 101, i32 102, i32 103, i32 104, i32 105, i32 106, i32 107, i32 108, i32 109, i32 110, i32 111, i32 112, i32 113, i32 114, i32 115, i32 116, i32 117, i32 118, i32 119, i32 120, i32 121, i32 122, i32 123, i32 124, i32 125, i32 126, i32 127, i32 128, i32 129, i32 130, i32 131, i32 132, i32 133, i32 134, i32 135, i32 136, i32 137, i32 138, i32 139, i32 140, i32 141, i32 142, i32 143, i32 144, i32 145, i32 146, i32 147, i32 148, i32 149, i32 150, i32 151, i32 152, i32 153, i32 154, i32 155, i32 156, i32 157, i32 158, i32 159, i32 160, i32 161, i32 162, i32 163, i32 164, i32 165, i32 166, i32 167, i32 168, i32 169, i32 170, i32 171, i32 172, i32 173, i32 174, i32 175, i32 176, i32 177, i32 178, i32 179, i32 180, i32 181, i32 182, i32 183, i32 184, i32 185, i32 186, i32 187, i32 188, i32 189, i32 190, i32 191, i32 192, i32 193, i32 194, i32 195, i32 196, i32 197, i32 198, i32 199, i32 200, i32 201, i32 202, i32 203, i32 204, i32 205, i32 206, i32 207, i32 208, i32 209, i32 210, i32 211, i32 212, i32 213, i32 214, i32 215, i32 216, i32 217, i32 218, i32 219, i32 220, i32 221, i32 222, i32 223, i32 224, i32 225, i32 226, i32 227, i32 228, i32 229, i32 230, i32 231, i32 232, i32 233, i32 234, i32 235, i32 236, i32 237, i32 238, i32 239, i32 240, i32 241, i32 242, i32 243, i32 244, i32 245, i32 246, i32 247, i32 248, i32 249, i32 250, i32 251, i32 252, i32 253, i32 254, i32 255>
26204 %6 = shufflevector <128 x i64> %3, <128 x i64> %4, <256 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 64, i32 65, i32 66, i32 67, i32 68, i32 69, i32 70, i32 71, i32 72, i32 73, i32 74, i32 75, i32 76, i32 77, i32 78, i32 79, i32 80, i32 81, i32 82, i32 83, i32 84, i32 85, i32 86, i32 87, i32 88, i32 89, i32 90, i32 91, i32 92, i32 93, i32 94, i32 95, i32 96, i32 97, i32 98, i32 99, i32 100, i32 101, i32 102, i32 103, i32 104, i32 105, i32 106, i32 107, i32 108, i32 109, i32 110, i32 111, i32 112, i32 113, i32 114, i32 115, i32 116, i32 117, i32 118, i32 119, i32 120, i32 121, i32 122, i32 123, i32 124, i32 125, i32 126, i32 127, i32 128, i32 129, i32 130, i32 131, i32 132, i32 133, i32 134, i32 135, i32 136, i32 137, i32 138, i32 139, i32 140, i32 141, i32 142, i32 143, i32 144, i32 145, i32 146, i32 147, i32 148, i32 149, i32 150, i32 151, i32 152, i32 153, i32 154, i32 155, i32 156, i32 157, i32 158, i32 159, i32 160, i32 161, i32 162, i32 163, i32 164, i32 165, i32 166, i32 167, i32 168, i32 169, i32 170, i32 171, i32 172, i32 173, i32 174, i32 175, i32 176, i32 177, i32 178, i32 179, i32 180, i32 181, i32 182, i32 183, i32 184, i32 185, i32 186, i32 187, i32 188, i32 189, i32 190, i32 191, i32 192, i32 193, i32 194, i32 195, i32 196, i32 197, i32 198, i32 199, i32 200, i32 201, i32 202, i32 203, i32 204, i32 205, i32 206, i32 207, i32 208, i32 209, i32 210, i32 211, i32 212, i32 213, i32 214, i32 215, i32 216, i32 217, i32 218, i32 219, i32 220, i32 221, i32 222, i32 223, i32 224, i32 225, i32 226, i32 227, i32 228, i32 229, i32 230, i32 231, i32 232, i32 233, i32 234, i32 235, i32 236, i32 237, i32 238, i32 239, i32 240, i32 241, i32 242, i32 243, i32 244, i32 245, i32 246, i32 247, i32 248, i32 249, i32 250, i32 251, i32 252, i32 253, i32 254, i32 255>
26205 %7 = shufflevector <256 x i64> %5, <256 x i64> %6, <512 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 64, i32 65, i32 66, i32 67, i32 68, i32 69, i32 70, i32 71, i32 72, i32 73, i32 74, i32 75, i32 76, i32 77, i32 78, i32 79, i32 80, i32 81, i32 82, i32 83, i32 84, i32 85, i32 86, i32 87, i32 88, i32 89, i32 90, i32 91, i32 92, i32 93, i32 94, i32 95, i32 96, i32 97, i32 98, i32 99, i32 100, i32 101, i32 102, i32 103, i32 104, i32 105, i32 106, i32 107, i32 108, i32 109, i32 110, i32 111, i32 112, i32 113, i32 114, i32 115, i32 116, i32 117, i32 118, i32 119, i32 120, i32 121, i32 122, i32 123, i32 124, i32 125, i32 126, i32 127, i32 128, i32 129, i32 130, i32 131, i32 132, i32 133, i32 134, i32 135, i32 136, i32 137, i32 138, i32 139, i32 140, i32 141, i32 142, i32 143, i32 144, i32 145, i32 146, i32 147, i32 148, i32 149, i32 150, i32 151, i32 152, i32 153, i32 154, i32 155, i32 156, i32 157, i32 158, i32 159, i32 160, i32 161, i32 162, i32 163, i32 164, i32 165, i32 166, i32 167, i32 168, i32 169, i32 170, i32 171, i32 172, i32 173, i32 174, i32 175, i32 176, i32 177, i32 178, i32 179, i32 180, i32 181, i32 182, i32 183, i32 184, i32 185, i32 186, i32 187, i32 188, i32 189, i32 190, i32 191, i32 192, i32 193, i32 194, i32 195, i32 196, i32 197, i32 198, i32 199, i32 200, i32 201, i32 202, i32 203, i32 204, i32 205, i32 206, i32 207, i32 208, i32 209, i32 210, i32 211, i32 212, i32 213, i32 214, i32 215, i32 216, i32 217, i32 218, i32 219, i32 220, i32 221, i32 222, i32 223, i32 224, i32 225, i32 226, i32 227, i32 228, i32 229, i32 230, i32 231, i32 232, i32 233, i32 234, i32 235, i32 236, i32 237, i32 238, i32 239, i32 240, i32 241, i32 242, i32 243, i32 244, i32 245, i32 246, i32 247, i32 248, i32 249, i32 250, i32 251, i32 252, i32 253, i32 254, i32 255, i32 256, i32 257, i32 258, i32 259, i32 260, i32 261, i32 262, i32 263, i32 264, i32 265, i32 266, i32 267, i32 268, i32 269, i32 270, i32 271, i32 272, i32 273, i32 274, i32 275, i32 276, i32 277, i32 278, i32 279, i32 280, i32 281, i32 282, i32 283, i32 284, i32 285, i32 286, i32 287, i32 288, i32 289, i32 290, i32 291, i32 292, i32 293, i32 294, i32 295, i32 296, i32 297, i32 298, i32 299, i32 300, i32 301, i32 302, i32 303, i32 304, i32 305, i32 306, i32 307, i32 308, i32 309, i32 310, i32 311, i32 312, i32 313, i32 314, i32 315, i32 316, i32 317, i32 318, i32 319, i32 320, i32 321, i32 322, i32 323, i32 324, i32 325, i32 326, i32 327, i32 328, i32 329, i32 330, i32 331, i32 332, i32 333, i32 334, i32 335, i32 336, i32 337, i32 338, i32 339, i32 340, i32 341, i32 342, i32 343, i32 344, i32 345, i32 346, i32 347, i32 348, i32 349, i32 350, i32 351, i32 352, i32 353, i32 354, i32 355, i32 356, i32 357, i32 358, i32 359, i32 360, i32 361, i32 362, i32 363, i32 364, i32 365, i32 366, i32 367, i32 368, i32 369, i32 370, i32 371, i32 372, i32 373, i32 374, i32 375, i32 376, i32 377, i32 378, i32 379, i32 380, i32 381, i32 382, i32 383, i32 384, i32 385, i32 386, i32 387, i32 388, i32 389, i32 390, i32 391, i32 392, i32 393, i32 394, i32 395, i32 396, i32 397, i32 398, i32 399, i32 400, i32 401, i32 402, i32 403, i32 404, i32 405, i32 406, i32 407, i32 408, i32 409, i32 410, i32 411, i32 412, i32 413, i32 414, i32 415, i32 416, i32 417, i32 418, i32 419, i32 420, i32 421, i32 422, i32 423, i32 424, i32 425, i32 426, i32 427, i32 428, i32 429, i32 430, i32 431, i32 432, i32 433, i32 434, i32 435, i32 436, i32 437, i32 438, i32 439, i32 440, i32 441, i32 442, i32 443, i32 444, i32 445, i32 446, i32 447, i32 448, i32 449, i32 450, i32 451, i32 452, i32 453, i32 454, i32 455, i32 456, i32 457, i32 458, i32 459, i32 460, i32 461, i32 462, i32 463, i32 464, i32 465, i32 466, i32 467, i32 468, i32 469, i32 470, i32 471, i32 472, i32 473, i32 474, i32 475, i32 476, i32 477, i32 478, i32 479, i32 480, i32 481, i32 482, i32 483, i32 484, i32 485, i32 486, i32 487, i32 488, i32 489, i32 490, i32 491, i32 492, i32 493, i32 494, i32 495, i32 496, i32 497, i32 498, i32 499, i32 500, i32 501, i32 502, i32 503, i32 504, i32 505, i32 506, i32 507, i32 508, i32 509, i32 510, i32 511>
26206 %interleaved.vec = shufflevector <512 x i64> %7, <512 x i64> poison, <512 x i32> <i32 0, i32 64, i32 128, i32 192, i32 256, i32 320, i32 384, i32 448, i32 1, i32 65, i32 129, i32 193, i32 257, i32 321, i32 385, i32 449, i32 2, i32 66, i32 130, i32 194, i32 258, i32 322, i32 386, i32 450, i32 3, i32 67, i32 131, i32 195, i32 259, i32 323, i32 387, i32 451, i32 4, i32 68, i32 132, i32 196, i32 260, i32 324, i32 388, i32 452, i32 5, i32 69, i32 133, i32 197, i32 261, i32 325, i32 389, i32 453, i32 6, i32 70, i32 134, i32 198, i32 262, i32 326, i32 390, i32 454, i32 7, i32 71, i32 135, i32 199, i32 263, i32 327, i32 391, i32 455, i32 8, i32 72, i32 136, i32 200, i32 264, i32 328, i32 392, i32 456, i32 9, i32 73, i32 137, i32 201, i32 265, i32 329, i32 393, i32 457, i32 10, i32 74, i32 138, i32 202, i32 266, i32 330, i32 394, i32 458, i32 11, i32 75, i32 139, i32 203, i32 267, i32 331, i32 395, i32 459, i32 12, i32 76, i32 140, i32 204, i32 268, i32 332, i32 396, i32 460, i32 13, i32 77, i32 141, i32 205, i32 269, i32 333, i32 397, i32 461, i32 14, i32 78, i32 142, i32 206, i32 270, i32 334, i32 398, i32 462, i32 15, i32 79, i32 143, i32 207, i32 271, i32 335, i32 399, i32 463, i32 16, i32 80, i32 144, i32 208, i32 272, i32 336, i32 400, i32 464, i32 17, i32 81, i32 145, i32 209, i32 273, i32 337, i32 401, i32 465, i32 18, i32 82, i32 146, i32 210, i32 274, i32 338, i32 402, i32 466, i32 19, i32 83, i32 147, i32 211, i32 275, i32 339, i32 403, i32 467, i32 20, i32 84, i32 148, i32 212, i32 276, i32 340, i32 404, i32 468, i32 21, i32 85, i32 149, i32 213, i32 277, i32 341, i32 405, i32 469, i32 22, i32 86, i32 150, i32 214, i32 278, i32 342, i32 406, i32 470, i32 23, i32 87, i32 151, i32 215, i32 279, i32 343, i32 407, i32 471, i32 24, i32 88, i32 152, i32 216, i32 280, i32 344, i32 408, i32 472, i32 25, i32 89, i32 153, i32 217, i32 281, i32 345, i32 409, i32 473, i32 26, i32 90, i32 154, i32 218, i32 282, i32 346, i32 410, i32 474, i32 27, i32 91, i32 155, i32 219, i32 283, i32 347, i32 411, i32 475, i32 28, i32 92, i32 156, i32 220, i32 284, i32 348, i32 412, i32 476, i32 29, i32 93, i32 157, i32 221, i32 285, i32 349, i32 413, i32 477, i32 30, i32 94, i32 158, i32 222, i32 286, i32 350, i32 414, i32 478, i32 31, i32 95, i32 159, i32 223, i32 287, i32 351, i32 415, i32 479, i32 32, i32 96, i32 160, i32 224, i32 288, i32 352, i32 416, i32 480, i32 33, i32 97, i32 161, i32 225, i32 289, i32 353, i32 417, i32 481, i32 34, i32 98, i32 162, i32 226, i32 290, i32 354, i32 418, i32 482, i32 35, i32 99, i32 163, i32 227, i32 291, i32 355, i32 419, i32 483, i32 36, i32 100, i32 164, i32 228, i32 292, i32 356, i32 420, i32 484, i32 37, i32 101, i32 165, i32 229, i32 293, i32 357, i32 421, i32 485, i32 38, i32 102, i32 166, i32 230, i32 294, i32 358, i32 422, i32 486, i32 39, i32 103, i32 167, i32 231, i32 295, i32 359, i32 423, i32 487, i32 40, i32 104, i32 168, i32 232, i32 296, i32 360, i32 424, i32 488, i32 41, i32 105, i32 169, i32 233, i32 297, i32 361, i32 425, i32 489, i32 42, i32 106, i32 170, i32 234, i32 298, i32 362, i32 426, i32 490, i32 43, i32 107, i32 171, i32 235, i32 299, i32 363, i32 427, i32 491, i32 44, i32 108, i32 172, i32 236, i32 300, i32 364, i32 428, i32 492, i32 45, i32 109, i32 173, i32 237, i32 301, i32 365, i32 429, i32 493, i32 46, i32 110, i32 174, i32 238, i32 302, i32 366, i32 430, i32 494, i32 47, i32 111, i32 175, i32 239, i32 303, i32 367, i32 431, i32 495, i32 48, i32 112, i32 176, i32 240, i32 304, i32 368, i32 432, i32 496, i32 49, i32 113, i32 177, i32 241, i32 305, i32 369, i32 433, i32 497, i32 50, i32 114, i32 178, i32 242, i32 306, i32 370, i32 434, i32 498, i32 51, i32 115, i32 179, i32 243, i32 307, i32 371, i32 435, i32 499, i32 52, i32 116, i32 180, i32 244, i32 308, i32 372, i32 436, i32 500, i32 53, i32 117, i32 181, i32 245, i32 309, i32 373, i32 437, i32 501, i32 54, i32 118, i32 182, i32 246, i32 310, i32 374, i32 438, i32 502, i32 55, i32 119, i32 183, i32 247, i32 311, i32 375, i32 439, i32 503, i32 56, i32 120, i32 184, i32 248, i32 312, i32 376, i32 440, i32 504, i32 57, i32 121, i32 185, i32 249, i32 313, i32 377, i32 441, i32 505, i32 58, i32 122, i32 186, i32 250, i32 314, i32 378, i32 442, i32 506, i32 59, i32 123, i32 187, i32 251, i32 315, i32 379, i32 443, i32 507, i32 60, i32 124, i32 188, i32 252, i32 316, i32 380, i32 444, i32 508, i32 61, i32 125, i32 189, i32 253, i32 317, i32 381, i32 445, i32 509, i32 62, i32 126, i32 190, i32 254, i32 318, i32 382, i32 446, i32 510, i32 63, i32 127, i32 191, i32 255, i32 319, i32 383, i32 447, i32 511>
26207 store <512 x i64> %interleaved.vec, ptr %out.vec, align 64
26208 ret void
26209 }