| blob | 4e30d706c0e20f23b473cd0044d96755fb9db6d6 |
1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+sse2 | FileCheck %s --check-prefixes=SSE,SSE2
3 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+sse4.1 | FileCheck %s --check-prefixes=SSE,SSE41
4 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx | FileCheck %s --check-prefixes=AVX,AVX1
5 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx2 | FileCheck %s --check-prefixes=AVX,AVX2
6 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f | FileCheck %s --check-prefix=AVX512F
7 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f,+avx512vl | FileCheck %s --check-prefix=AVX512VL
8 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f,+avx512bw | FileCheck %s --check-prefix=AVX512BW
9 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f,+avx512bw,+avx512vl | FileCheck %s --check-prefix=AVX512VLBW
10 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f,+avx512vbmi,+avx512vbmi2 | FileCheck %s --check-prefix=AVX512VBMI2
11 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f,+avx512vbmi,+avx512vbmi2,+avx512vl | FileCheck %s --check-prefix=AVX512VLVBMI2
12 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+xop,+avx | FileCheck %s --check-prefixes=XOP,XOPAVX1
13 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+xop,+avx2 | FileCheck %s --check-prefixes=XOP,XOPAVX2
15 ; Just one 32-bit run to make sure we do reasonable things for i64 cases.
16 ; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=+sse2 | FileCheck %s --check-prefix=X86-SSE2
18 declare <2 x i64> @llvm.fshr.v2i64(<2 x i64>, <2 x i64>, <2 x i64>)
19 declare <4 x i32> @llvm.fshr.v4i32(<4 x i32>, <4 x i32>, <4 x i32>)
20 declare <8 x i16> @llvm.fshr.v8i16(<8 x i16>, <8 x i16>, <8 x i16>)
21 declare <16 x i8> @llvm.fshr.v16i8(<16 x i8>, <16 x i8>, <16 x i8>)
23 ;
24 ; Variable Shifts
25 ;
27 define <2 x i64> @var_funnnel_v2i64(<2 x i64> %x, <2 x i64> %amt) nounwind {
28 ; SSE2-LABEL: var_funnnel_v2i64:
29 ; SSE2: # %bb.0:
30 ; SSE2-NEXT: movdqa {{.*#+}} xmm2 = [63,63]
31 ; SSE2-NEXT: pxor %xmm3, %xmm3
32 ; SSE2-NEXT: psubq %xmm1, %xmm3
33 ; SSE2-NEXT: pand %xmm2, %xmm1
34 ; SSE2-NEXT: movdqa %xmm0, %xmm4
35 ; SSE2-NEXT: psrlq %xmm1, %xmm4
36 ; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm1[2,3,2,3]
37 ; SSE2-NEXT: movdqa %xmm0, %xmm5
38 ; SSE2-NEXT: psrlq %xmm1, %xmm5
39 ; SSE2-NEXT: movsd {{.*#+}} xmm5 = xmm4[0],xmm5[1]
40 ; SSE2-NEXT: pand %xmm2, %xmm3
41 ; SSE2-NEXT: movdqa %xmm0, %xmm1
42 ; SSE2-NEXT: psllq %xmm3, %xmm1
43 ; SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm3[2,3,2,3]
44 ; SSE2-NEXT: psllq %xmm2, %xmm0
45 ; SSE2-NEXT: movsd {{.*#+}} xmm0 = xmm1[0],xmm0[1]
46 ; SSE2-NEXT: orpd %xmm5, %xmm0
47 ; SSE2-NEXT: retq
48 ;
49 ; SSE41-LABEL: var_funnnel_v2i64:
50 ; SSE41: # %bb.0:
51 ; SSE41-NEXT: pmovsxbq {{.*#+}} xmm2 = [63,63]
52 ; SSE41-NEXT: pxor %xmm3, %xmm3
53 ; SSE41-NEXT: psubq %xmm1, %xmm3
54 ; SSE41-NEXT: pand %xmm2, %xmm1
55 ; SSE41-NEXT: movdqa %xmm0, %xmm4
56 ; SSE41-NEXT: psrlq %xmm1, %xmm4
57 ; SSE41-NEXT: pshufd {{.*#+}} xmm1 = xmm1[2,3,2,3]
58 ; SSE41-NEXT: movdqa %xmm0, %xmm5
59 ; SSE41-NEXT: psrlq %xmm1, %xmm5
60 ; SSE41-NEXT: pblendw {{.*#+}} xmm5 = xmm4[0,1,2,3],xmm5[4,5,6,7]
61 ; SSE41-NEXT: pand %xmm2, %xmm3
62 ; SSE41-NEXT: movdqa %xmm0, %xmm1
63 ; SSE41-NEXT: psllq %xmm3, %xmm1
64 ; SSE41-NEXT: pshufd {{.*#+}} xmm2 = xmm3[2,3,2,3]
65 ; SSE41-NEXT: psllq %xmm2, %xmm0
66 ; SSE41-NEXT: pblendw {{.*#+}} xmm0 = xmm1[0,1,2,3],xmm0[4,5,6,7]
67 ; SSE41-NEXT: por %xmm5, %xmm0
68 ; SSE41-NEXT: retq
69 ;
70 ; AVX1-LABEL: var_funnnel_v2i64:
71 ; AVX1: # %bb.0:
72 ; AVX1-NEXT: vpmovsxbq {{.*#+}} xmm2 = [63,63]
73 ; AVX1-NEXT: vpand %xmm2, %xmm1, %xmm3
74 ; AVX1-NEXT: vpsrlq %xmm3, %xmm0, %xmm4
75 ; AVX1-NEXT: vpshufd {{.*#+}} xmm3 = xmm3[2,3,2,3]
76 ; AVX1-NEXT: vpsrlq %xmm3, %xmm0, %xmm3
77 ; AVX1-NEXT: vpblendw {{.*#+}} xmm3 = xmm4[0,1,2,3],xmm3[4,5,6,7]
78 ; AVX1-NEXT: vpxor %xmm4, %xmm4, %xmm4
79 ; AVX1-NEXT: vpsubq %xmm1, %xmm4, %xmm1
80 ; AVX1-NEXT: vpand %xmm2, %xmm1, %xmm1
81 ; AVX1-NEXT: vpsllq %xmm1, %xmm0, %xmm2
82 ; AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[2,3,2,3]
83 ; AVX1-NEXT: vpsllq %xmm1, %xmm0, %xmm0
84 ; AVX1-NEXT: vpblendw {{.*#+}} xmm0 = xmm2[0,1,2,3],xmm0[4,5,6,7]
85 ; AVX1-NEXT: vpor %xmm0, %xmm3, %xmm0
86 ; AVX1-NEXT: retq
87 ;
88 ; AVX2-LABEL: var_funnnel_v2i64:
89 ; AVX2: # %bb.0:
90 ; AVX2-NEXT: vpmovsxbq {{.*#+}} xmm2 = [63,63]
91 ; AVX2-NEXT: vpand %xmm2, %xmm1, %xmm3
92 ; AVX2-NEXT: vpsrlvq %xmm3, %xmm0, %xmm3
93 ; AVX2-NEXT: vpxor %xmm4, %xmm4, %xmm4
94 ; AVX2-NEXT: vpsubq %xmm1, %xmm4, %xmm1
95 ; AVX2-NEXT: vpand %xmm2, %xmm1, %xmm1
96 ; AVX2-NEXT: vpsllvq %xmm1, %xmm0, %xmm0
97 ; AVX2-NEXT: vpor %xmm0, %xmm3, %xmm0
98 ; AVX2-NEXT: retq
99 ;
100 ; AVX512F-LABEL: var_funnnel_v2i64:
101 ; AVX512F: # %bb.0:
102 ; AVX512F-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
103 ; AVX512F-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
104 ; AVX512F-NEXT: vprorvq %zmm1, %zmm0, %zmm0
105 ; AVX512F-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
106 ; AVX512F-NEXT: vzeroupper
107 ; AVX512F-NEXT: retq
108 ;
109 ; AVX512VL-LABEL: var_funnnel_v2i64:
110 ; AVX512VL: # %bb.0:
111 ; AVX512VL-NEXT: vprorvq %xmm1, %xmm0, %xmm0
112 ; AVX512VL-NEXT: retq
113 ;
114 ; AVX512BW-LABEL: var_funnnel_v2i64:
115 ; AVX512BW: # %bb.0:
116 ; AVX512BW-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
117 ; AVX512BW-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
118 ; AVX512BW-NEXT: vprorvq %zmm1, %zmm0, %zmm0
119 ; AVX512BW-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
120 ; AVX512BW-NEXT: vzeroupper
121 ; AVX512BW-NEXT: retq
122 ;
123 ; AVX512VLBW-LABEL: var_funnnel_v2i64:
124 ; AVX512VLBW: # %bb.0:
125 ; AVX512VLBW-NEXT: vprorvq %xmm1, %xmm0, %xmm0
126 ; AVX512VLBW-NEXT: retq
127 ;
128 ; AVX512VBMI2-LABEL: var_funnnel_v2i64:
129 ; AVX512VBMI2: # %bb.0:
130 ; AVX512VBMI2-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
131 ; AVX512VBMI2-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
132 ; AVX512VBMI2-NEXT: vprorvq %zmm1, %zmm0, %zmm0
133 ; AVX512VBMI2-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
134 ; AVX512VBMI2-NEXT: vzeroupper
135 ; AVX512VBMI2-NEXT: retq
136 ;
137 ; AVX512VLVBMI2-LABEL: var_funnnel_v2i64:
138 ; AVX512VLVBMI2: # %bb.0:
139 ; AVX512VLVBMI2-NEXT: vprorvq %xmm1, %xmm0, %xmm0
140 ; AVX512VLVBMI2-NEXT: retq
141 ;
142 ; XOP-LABEL: var_funnnel_v2i64:
143 ; XOP: # %bb.0:
144 ; XOP-NEXT: vpxor %xmm2, %xmm2, %xmm2
145 ; XOP-NEXT: vpsubq %xmm1, %xmm2, %xmm1
146 ; XOP-NEXT: vprotq %xmm1, %xmm0, %xmm0
147 ; XOP-NEXT: retq
148 ;
149 ; X86-SSE2-LABEL: var_funnnel_v2i64:
150 ; X86-SSE2: # %bb.0:
151 ; X86-SSE2-NEXT: movdqa {{.*#+}} xmm2 = [63,0,63,0]
152 ; X86-SSE2-NEXT: pxor %xmm3, %xmm3
153 ; X86-SSE2-NEXT: psubq %xmm1, %xmm3
154 ; X86-SSE2-NEXT: pand %xmm2, %xmm1
155 ; X86-SSE2-NEXT: movdqa %xmm0, %xmm4
156 ; X86-SSE2-NEXT: psrlq %xmm1, %xmm4
157 ; X86-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm1[2,3,2,3]
158 ; X86-SSE2-NEXT: movdqa %xmm0, %xmm5
159 ; X86-SSE2-NEXT: psrlq %xmm1, %xmm5
160 ; X86-SSE2-NEXT: movsd {{.*#+}} xmm5 = xmm4[0],xmm5[1]
161 ; X86-SSE2-NEXT: pand %xmm2, %xmm3
162 ; X86-SSE2-NEXT: movdqa %xmm0, %xmm1
163 ; X86-SSE2-NEXT: psllq %xmm3, %xmm1
164 ; X86-SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm3[2,3,2,3]
165 ; X86-SSE2-NEXT: psllq %xmm2, %xmm0
166 ; X86-SSE2-NEXT: movsd {{.*#+}} xmm0 = xmm1[0],xmm0[1]
167 ; X86-SSE2-NEXT: orpd %xmm5, %xmm0
168 ; X86-SSE2-NEXT: retl
169 %res = call <2 x i64> @llvm.fshr.v2i64(<2 x i64> %x, <2 x i64> %x, <2 x i64> %amt)
170 ret <2 x i64> %res
171 }
173 define <4 x i32> @var_funnnel_v4i32(<4 x i32> %x, <4 x i32> %amt) nounwind {
174 ; SSE2-LABEL: var_funnnel_v4i32:
175 ; SSE2: # %bb.0:
176 ; SSE2-NEXT: pxor %xmm2, %xmm2
177 ; SSE2-NEXT: psubd %xmm1, %xmm2
178 ; SSE2-NEXT: pslld $23, %xmm2
179 ; SSE2-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm2
180 ; SSE2-NEXT: paddd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm2
181 ; SSE2-NEXT: cvttps2dq %xmm2, %xmm1
182 ; SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm0[1,1,3,3]
183 ; SSE2-NEXT: pmuludq %xmm1, %xmm0
184 ; SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm0[1,3,2,3]
185 ; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
186 ; SSE2-NEXT: pmuludq %xmm2, %xmm1
187 ; SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm1[1,3,2,3]
188 ; SSE2-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]
189 ; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
190 ; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,2,2,3]
191 ; SSE2-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
192 ; SSE2-NEXT: por %xmm3, %xmm0
193 ; SSE2-NEXT: retq
194 ;
195 ; SSE41-LABEL: var_funnnel_v4i32:
196 ; SSE41: # %bb.0:
197 ; SSE41-NEXT: pxor %xmm2, %xmm2
198 ; SSE41-NEXT: psubd %xmm1, %xmm2
199 ; SSE41-NEXT: pslld $23, %xmm2
200 ; SSE41-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm2
201 ; SSE41-NEXT: paddd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm2
202 ; SSE41-NEXT: cvttps2dq %xmm2, %xmm1
203 ; SSE41-NEXT: pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
204 ; SSE41-NEXT: pshufd {{.*#+}} xmm3 = xmm0[1,1,3,3]
205 ; SSE41-NEXT: pmuludq %xmm2, %xmm3
206 ; SSE41-NEXT: pmuludq %xmm1, %xmm0
207 ; SSE41-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,3,3]
208 ; SSE41-NEXT: pblendw {{.*#+}} xmm1 = xmm1[0,1],xmm3[2,3],xmm1[4,5],xmm3[6,7]
209 ; SSE41-NEXT: pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
210 ; SSE41-NEXT: pblendw {{.*#+}} xmm0 = xmm0[0,1],xmm2[2,3],xmm0[4,5],xmm2[6,7]
211 ; SSE41-NEXT: por %xmm1, %xmm0
212 ; SSE41-NEXT: retq
213 ;
214 ; AVX1-LABEL: var_funnnel_v4i32:
215 ; AVX1: # %bb.0:
216 ; AVX1-NEXT: vpxor %xmm2, %xmm2, %xmm2
217 ; AVX1-NEXT: vpsubd %xmm1, %xmm2, %xmm1
218 ; AVX1-NEXT: vpslld $23, %xmm1, %xmm1
219 ; AVX1-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1, %xmm1
220 ; AVX1-NEXT: vpaddd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1, %xmm1
221 ; AVX1-NEXT: vcvttps2dq %xmm1, %xmm1
222 ; AVX1-NEXT: vpshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
223 ; AVX1-NEXT: vpshufd {{.*#+}} xmm3 = xmm0[1,1,3,3]
224 ; AVX1-NEXT: vpmuludq %xmm2, %xmm3, %xmm2
225 ; AVX1-NEXT: vpmuludq %xmm1, %xmm0, %xmm0
226 ; AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,3,3]
227 ; AVX1-NEXT: vpblendw {{.*#+}} xmm1 = xmm1[0,1],xmm2[2,3],xmm1[4,5],xmm2[6,7]
228 ; AVX1-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[0,0,2,2]
229 ; AVX1-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0,1],xmm2[2,3],xmm0[4,5],xmm2[6,7]
230 ; AVX1-NEXT: vpor %xmm1, %xmm0, %xmm0
231 ; AVX1-NEXT: retq
232 ;
233 ; AVX2-LABEL: var_funnnel_v4i32:
234 ; AVX2: # %bb.0:
235 ; AVX2-NEXT: vpbroadcastd {{.*#+}} xmm2 = [31,31,31,31]
236 ; AVX2-NEXT: vpand %xmm2, %xmm1, %xmm1
237 ; AVX2-NEXT: vpsrlvd %xmm1, %xmm0, %xmm2
238 ; AVX2-NEXT: vpbroadcastd {{.*#+}} xmm3 = [32,32,32,32]
239 ; AVX2-NEXT: vpsubd %xmm1, %xmm3, %xmm1
240 ; AVX2-NEXT: vpsllvd %xmm1, %xmm0, %xmm0
241 ; AVX2-NEXT: vpor %xmm0, %xmm2, %xmm0
242 ; AVX2-NEXT: retq
243 ;
244 ; AVX512F-LABEL: var_funnnel_v4i32:
245 ; AVX512F: # %bb.0:
246 ; AVX512F-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
247 ; AVX512F-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
248 ; AVX512F-NEXT: vprorvd %zmm1, %zmm0, %zmm0
249 ; AVX512F-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
250 ; AVX512F-NEXT: vzeroupper
251 ; AVX512F-NEXT: retq
252 ;
253 ; AVX512VL-LABEL: var_funnnel_v4i32:
254 ; AVX512VL: # %bb.0:
255 ; AVX512VL-NEXT: vprorvd %xmm1, %xmm0, %xmm0
256 ; AVX512VL-NEXT: retq
257 ;
258 ; AVX512BW-LABEL: var_funnnel_v4i32:
259 ; AVX512BW: # %bb.0:
260 ; AVX512BW-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
261 ; AVX512BW-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
262 ; AVX512BW-NEXT: vprorvd %zmm1, %zmm0, %zmm0
263 ; AVX512BW-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
264 ; AVX512BW-NEXT: vzeroupper
265 ; AVX512BW-NEXT: retq
266 ;
267 ; AVX512VLBW-LABEL: var_funnnel_v4i32:
268 ; AVX512VLBW: # %bb.0:
269 ; AVX512VLBW-NEXT: vprorvd %xmm1, %xmm0, %xmm0
270 ; AVX512VLBW-NEXT: retq
271 ;
272 ; AVX512VBMI2-LABEL: var_funnnel_v4i32:
273 ; AVX512VBMI2: # %bb.0:
274 ; AVX512VBMI2-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
275 ; AVX512VBMI2-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
276 ; AVX512VBMI2-NEXT: vprorvd %zmm1, %zmm0, %zmm0
277 ; AVX512VBMI2-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
278 ; AVX512VBMI2-NEXT: vzeroupper
279 ; AVX512VBMI2-NEXT: retq
280 ;
281 ; AVX512VLVBMI2-LABEL: var_funnnel_v4i32:
282 ; AVX512VLVBMI2: # %bb.0:
283 ; AVX512VLVBMI2-NEXT: vprorvd %xmm1, %xmm0, %xmm0
284 ; AVX512VLVBMI2-NEXT: retq
285 ;
286 ; XOP-LABEL: var_funnnel_v4i32:
287 ; XOP: # %bb.0:
288 ; XOP-NEXT: vpxor %xmm2, %xmm2, %xmm2
289 ; XOP-NEXT: vpsubd %xmm1, %xmm2, %xmm1
290 ; XOP-NEXT: vprotd %xmm1, %xmm0, %xmm0
291 ; XOP-NEXT: retq
292 ;
293 ; X86-SSE2-LABEL: var_funnnel_v4i32:
294 ; X86-SSE2: # %bb.0:
295 ; X86-SSE2-NEXT: pxor %xmm2, %xmm2
296 ; X86-SSE2-NEXT: psubd %xmm1, %xmm2
297 ; X86-SSE2-NEXT: pslld $23, %xmm2
298 ; X86-SSE2-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}, %xmm2
299 ; X86-SSE2-NEXT: paddd {{\.?LCPI[0-9]+_[0-9]+}}, %xmm2
300 ; X86-SSE2-NEXT: cvttps2dq %xmm2, %xmm1
301 ; X86-SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm0[1,1,3,3]
302 ; X86-SSE2-NEXT: pmuludq %xmm1, %xmm0
303 ; X86-SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm0[1,3,2,3]
304 ; X86-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
305 ; X86-SSE2-NEXT: pmuludq %xmm2, %xmm1
306 ; X86-SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm1[1,3,2,3]
307 ; X86-SSE2-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]
308 ; X86-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
309 ; X86-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,2,2,3]
310 ; X86-SSE2-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
311 ; X86-SSE2-NEXT: por %xmm3, %xmm0
312 ; X86-SSE2-NEXT: retl
313 %res = call <4 x i32> @llvm.fshr.v4i32(<4 x i32> %x, <4 x i32> %x, <4 x i32> %amt)
314 ret <4 x i32> %res
315 }
317 define <8 x i16> @var_funnnel_v8i16(<8 x i16> %x, <8 x i16> %amt) nounwind {
318 ; SSE2-LABEL: var_funnnel_v8i16:
319 ; SSE2: # %bb.0:
320 ; SSE2-NEXT: pxor %xmm2, %xmm2
321 ; SSE2-NEXT: psubw %xmm1, %xmm2
322 ; SSE2-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm2
323 ; SSE2-NEXT: movdqa %xmm2, %xmm1
324 ; SSE2-NEXT: punpckhwd {{.*#+}} xmm1 = xmm1[4,4,5,5,6,6,7,7]
325 ; SSE2-NEXT: pslld $23, %xmm1
326 ; SSE2-NEXT: movdqa {{.*#+}} xmm3 = [1065353216,1065353216,1065353216,1065353216]
327 ; SSE2-NEXT: paddd %xmm3, %xmm1
328 ; SSE2-NEXT: cvttps2dq %xmm1, %xmm1
329 ; SSE2-NEXT: pslld $16, %xmm1
330 ; SSE2-NEXT: psrad $16, %xmm1
331 ; SSE2-NEXT: punpcklwd {{.*#+}} xmm2 = xmm2[0,0,1,1,2,2,3,3]
332 ; SSE2-NEXT: pslld $23, %xmm2
333 ; SSE2-NEXT: paddd %xmm3, %xmm2
334 ; SSE2-NEXT: cvttps2dq %xmm2, %xmm2
335 ; SSE2-NEXT: pslld $16, %xmm2
336 ; SSE2-NEXT: psrad $16, %xmm2
337 ; SSE2-NEXT: packssdw %xmm1, %xmm2
338 ; SSE2-NEXT: movdqa %xmm0, %xmm1
339 ; SSE2-NEXT: pmulhuw %xmm2, %xmm1
340 ; SSE2-NEXT: pmullw %xmm2, %xmm0
341 ; SSE2-NEXT: por %xmm1, %xmm0
342 ; SSE2-NEXT: retq
343 ;
344 ; SSE41-LABEL: var_funnnel_v8i16:
345 ; SSE41: # %bb.0:
346 ; SSE41-NEXT: pxor %xmm2, %xmm2
347 ; SSE41-NEXT: psubw %xmm1, %xmm2
348 ; SSE41-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm2
349 ; SSE41-NEXT: pmovzxwd {{.*#+}} xmm1 = xmm2[0],zero,xmm2[1],zero,xmm2[2],zero,xmm2[3],zero
350 ; SSE41-NEXT: punpckhwd {{.*#+}} xmm2 = xmm2[4,4,5,5,6,6,7,7]
351 ; SSE41-NEXT: pslld $23, %xmm2
352 ; SSE41-NEXT: movdqa {{.*#+}} xmm3 = [1065353216,1065353216,1065353216,1065353216]
353 ; SSE41-NEXT: paddd %xmm3, %xmm2
354 ; SSE41-NEXT: cvttps2dq %xmm2, %xmm2
355 ; SSE41-NEXT: pslld $23, %xmm1
356 ; SSE41-NEXT: paddd %xmm3, %xmm1
357 ; SSE41-NEXT: cvttps2dq %xmm1, %xmm1
358 ; SSE41-NEXT: packusdw %xmm2, %xmm1
359 ; SSE41-NEXT: movdqa %xmm0, %xmm2
360 ; SSE41-NEXT: pmulhuw %xmm1, %xmm2
361 ; SSE41-NEXT: pmullw %xmm1, %xmm0
362 ; SSE41-NEXT: por %xmm2, %xmm0
363 ; SSE41-NEXT: retq
364 ;
365 ; AVX1-LABEL: var_funnnel_v8i16:
366 ; AVX1: # %bb.0:
367 ; AVX1-NEXT: vpxor %xmm2, %xmm2, %xmm2
368 ; AVX1-NEXT: vpsubw %xmm1, %xmm2, %xmm1
369 ; AVX1-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1, %xmm1
370 ; AVX1-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm1[4,4,5,5,6,6,7,7]
371 ; AVX1-NEXT: vpslld $23, %xmm2, %xmm2
372 ; AVX1-NEXT: vbroadcastss {{.*#+}} xmm3 = [1065353216,1065353216,1065353216,1065353216]
373 ; AVX1-NEXT: vpaddd %xmm3, %xmm2, %xmm2
374 ; AVX1-NEXT: vcvttps2dq %xmm2, %xmm2
375 ; AVX1-NEXT: vpmovzxwd {{.*#+}} xmm1 = xmm1[0],zero,xmm1[1],zero,xmm1[2],zero,xmm1[3],zero
376 ; AVX1-NEXT: vpslld $23, %xmm1, %xmm1
377 ; AVX1-NEXT: vpaddd %xmm3, %xmm1, %xmm1
378 ; AVX1-NEXT: vcvttps2dq %xmm1, %xmm1
379 ; AVX1-NEXT: vpackusdw %xmm2, %xmm1, %xmm1
380 ; AVX1-NEXT: vpmulhuw %xmm1, %xmm0, %xmm2
381 ; AVX1-NEXT: vpmullw %xmm1, %xmm0, %xmm0
382 ; AVX1-NEXT: vpor %xmm2, %xmm0, %xmm0
383 ; AVX1-NEXT: retq
384 ;
385 ; AVX2-LABEL: var_funnnel_v8i16:
386 ; AVX2: # %bb.0:
387 ; AVX2-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1, %xmm1
388 ; AVX2-NEXT: vpxor %xmm2, %xmm2, %xmm2
389 ; AVX2-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm1[4],xmm2[4],xmm1[5],xmm2[5],xmm1[6],xmm2[6],xmm1[7],xmm2[7]
390 ; AVX2-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm0[4,4,5,5,6,6,7,7]
391 ; AVX2-NEXT: vpsrlvd %xmm3, %xmm4, %xmm3
392 ; AVX2-NEXT: vpblendw {{.*#+}} xmm3 = xmm3[0],xmm2[1],xmm3[2],xmm2[3],xmm3[4],xmm2[5],xmm3[6],xmm2[7]
393 ; AVX2-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3]
394 ; AVX2-NEXT: vpmovzxwd {{.*#+}} xmm1 = xmm1[0],zero,xmm1[1],zero,xmm1[2],zero,xmm1[3],zero
395 ; AVX2-NEXT: vpsrlvd %xmm1, %xmm0, %xmm0
396 ; AVX2-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0],xmm2[1],xmm0[2],xmm2[3],xmm0[4],xmm2[5],xmm0[6],xmm2[7]
397 ; AVX2-NEXT: vpackusdw %xmm3, %xmm0, %xmm0
398 ; AVX2-NEXT: retq
399 ;
400 ; AVX512F-LABEL: var_funnnel_v8i16:
401 ; AVX512F: # %bb.0:
402 ; AVX512F-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1, %xmm1
403 ; AVX512F-NEXT: vpxor %xmm2, %xmm2, %xmm2
404 ; AVX512F-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm1[4],xmm2[4],xmm1[5],xmm2[5],xmm1[6],xmm2[6],xmm1[7],xmm2[7]
405 ; AVX512F-NEXT: vpunpckhwd {{.*#+}} xmm4 = xmm0[4,4,5,5,6,6,7,7]
406 ; AVX512F-NEXT: vpsrlvd %xmm3, %xmm4, %xmm3
407 ; AVX512F-NEXT: vpblendw {{.*#+}} xmm3 = xmm3[0],xmm2[1],xmm3[2],xmm2[3],xmm3[4],xmm2[5],xmm3[6],xmm2[7]
408 ; AVX512F-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3]
409 ; AVX512F-NEXT: vpmovzxwd {{.*#+}} xmm1 = xmm1[0],zero,xmm1[1],zero,xmm1[2],zero,xmm1[3],zero
410 ; AVX512F-NEXT: vpsrlvd %xmm1, %xmm0, %xmm0
411 ; AVX512F-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0],xmm2[1],xmm0[2],xmm2[3],xmm0[4],xmm2[5],xmm0[6],xmm2[7]
412 ; AVX512F-NEXT: vpackusdw %xmm3, %xmm0, %xmm0
413 ; AVX512F-NEXT: retq
414 ;
415 ; AVX512VL-LABEL: var_funnnel_v8i16:
416 ; AVX512VL: # %bb.0:
417 ; AVX512VL-NEXT: vpandd {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to4}, %xmm1, %xmm1
418 ; AVX512VL-NEXT: vpxor %xmm2, %xmm2, %xmm2
419 ; AVX512VL-NEXT: vpunpckhwd {{.*#+}} xmm2 = xmm1[4],xmm2[4],xmm1[5],xmm2[5],xmm1[6],xmm2[6],xmm1[7],xmm2[7]
420 ; AVX512VL-NEXT: vpunpckhwd {{.*#+}} xmm3 = xmm0[4,4,5,5,6,6,7,7]
421 ; AVX512VL-NEXT: vpsrlvd %xmm2, %xmm3, %xmm2
422 ; AVX512VL-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3]
423 ; AVX512VL-NEXT: vpmovzxwd {{.*#+}} xmm1 = xmm1[0],zero,xmm1[1],zero,xmm1[2],zero,xmm1[3],zero
424 ; AVX512VL-NEXT: vpsrlvd %xmm1, %xmm0, %xmm0
425 ; AVX512VL-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm0
426 ; AVX512VL-NEXT: vpmovdw %ymm0, %xmm0
427 ; AVX512VL-NEXT: vzeroupper
428 ; AVX512VL-NEXT: retq
429 ;
430 ; AVX512BW-LABEL: var_funnnel_v8i16:
431 ; AVX512BW: # %bb.0:
432 ; AVX512BW-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
433 ; AVX512BW-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1, %xmm1
434 ; AVX512BW-NEXT: vpsrlvw %zmm1, %zmm0, %zmm2
435 ; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm3 = [16,16,16,16,16,16,16,16]
436 ; AVX512BW-NEXT: vpsubw %xmm1, %xmm3, %xmm1
437 ; AVX512BW-NEXT: vpsllvw %zmm1, %zmm0, %zmm0
438 ; AVX512BW-NEXT: vpor %xmm0, %xmm2, %xmm0
439 ; AVX512BW-NEXT: vzeroupper
440 ; AVX512BW-NEXT: retq
441 ;
442 ; AVX512VLBW-LABEL: var_funnnel_v8i16:
443 ; AVX512VLBW: # %bb.0:
444 ; AVX512VLBW-NEXT: vpandd {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to4}, %xmm1, %xmm1
445 ; AVX512VLBW-NEXT: vpsrlvw %xmm1, %xmm0, %xmm2
446 ; AVX512VLBW-NEXT: vpbroadcastw {{.*#+}} xmm3 = [16,16,16,16,16,16,16,16]
447 ; AVX512VLBW-NEXT: vpsubw %xmm1, %xmm3, %xmm1
448 ; AVX512VLBW-NEXT: vpsllvw %xmm1, %xmm0, %xmm0
449 ; AVX512VLBW-NEXT: vpor %xmm0, %xmm2, %xmm0
450 ; AVX512VLBW-NEXT: retq
451 ;
452 ; AVX512VBMI2-LABEL: var_funnnel_v8i16:
453 ; AVX512VBMI2: # %bb.0:
454 ; AVX512VBMI2-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
455 ; AVX512VBMI2-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
456 ; AVX512VBMI2-NEXT: vpshrdvw %zmm1, %zmm0, %zmm0
457 ; AVX512VBMI2-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
458 ; AVX512VBMI2-NEXT: vzeroupper
459 ; AVX512VBMI2-NEXT: retq
460 ;
461 ; AVX512VLVBMI2-LABEL: var_funnnel_v8i16:
462 ; AVX512VLVBMI2: # %bb.0:
463 ; AVX512VLVBMI2-NEXT: vpshrdvw %xmm1, %xmm0, %xmm0
464 ; AVX512VLVBMI2-NEXT: retq
465 ;
466 ; XOP-LABEL: var_funnnel_v8i16:
467 ; XOP: # %bb.0:
468 ; XOP-NEXT: vpxor %xmm2, %xmm2, %xmm2
469 ; XOP-NEXT: vpsubw %xmm1, %xmm2, %xmm1
470 ; XOP-NEXT: vprotw %xmm1, %xmm0, %xmm0
471 ; XOP-NEXT: retq
472 ;
473 ; X86-SSE2-LABEL: var_funnnel_v8i16:
474 ; X86-SSE2: # %bb.0:
475 ; X86-SSE2-NEXT: pxor %xmm2, %xmm2
476 ; X86-SSE2-NEXT: psubw %xmm1, %xmm2
477 ; X86-SSE2-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}, %xmm2
478 ; X86-SSE2-NEXT: movdqa %xmm2, %xmm1
479 ; X86-SSE2-NEXT: punpckhwd {{.*#+}} xmm1 = xmm1[4,4,5,5,6,6,7,7]
480 ; X86-SSE2-NEXT: pslld $23, %xmm1
481 ; X86-SSE2-NEXT: movdqa {{.*#+}} xmm3 = [1065353216,1065353216,1065353216,1065353216]
482 ; X86-SSE2-NEXT: paddd %xmm3, %xmm1
483 ; X86-SSE2-NEXT: cvttps2dq %xmm1, %xmm1
484 ; X86-SSE2-NEXT: pslld $16, %xmm1
485 ; X86-SSE2-NEXT: psrad $16, %xmm1
486 ; X86-SSE2-NEXT: punpcklwd {{.*#+}} xmm2 = xmm2[0,0,1,1,2,2,3,3]
487 ; X86-SSE2-NEXT: pslld $23, %xmm2
488 ; X86-SSE2-NEXT: paddd %xmm3, %xmm2
489 ; X86-SSE2-NEXT: cvttps2dq %xmm2, %xmm2
490 ; X86-SSE2-NEXT: pslld $16, %xmm2
491 ; X86-SSE2-NEXT: psrad $16, %xmm2
492 ; X86-SSE2-NEXT: packssdw %xmm1, %xmm2
493 ; X86-SSE2-NEXT: movdqa %xmm0, %xmm1
494 ; X86-SSE2-NEXT: pmulhuw %xmm2, %xmm1
495 ; X86-SSE2-NEXT: pmullw %xmm2, %xmm0
496 ; X86-SSE2-NEXT: por %xmm1, %xmm0
497 ; X86-SSE2-NEXT: retl
498 %res = call <8 x i16> @llvm.fshr.v8i16(<8 x i16> %x, <8 x i16> %x, <8 x i16> %amt)
499 ret <8 x i16> %res
500 }
502 define <16 x i8> @var_funnnel_v16i8(<16 x i8> %x, <16 x i8> %amt) nounwind {
503 ; SSE2-LABEL: var_funnnel_v16i8:
504 ; SSE2: # %bb.0:
505 ; SSE2-NEXT: movdqa %xmm0, %xmm2
506 ; SSE2-NEXT: pxor %xmm0, %xmm0
507 ; SSE2-NEXT: pxor %xmm3, %xmm3
508 ; SSE2-NEXT: psubb %xmm1, %xmm3
509 ; SSE2-NEXT: psllw $5, %xmm3
510 ; SSE2-NEXT: pxor %xmm1, %xmm1
511 ; SSE2-NEXT: pcmpgtb %xmm3, %xmm1
512 ; SSE2-NEXT: movdqa %xmm2, %xmm4
513 ; SSE2-NEXT: psrlw $4, %xmm4
514 ; SSE2-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm4
515 ; SSE2-NEXT: movdqa %xmm2, %xmm5
516 ; SSE2-NEXT: psllw $4, %xmm5
517 ; SSE2-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm5
518 ; SSE2-NEXT: por %xmm4, %xmm5
519 ; SSE2-NEXT: pand %xmm1, %xmm5
520 ; SSE2-NEXT: pandn %xmm2, %xmm1
521 ; SSE2-NEXT: por %xmm5, %xmm1
522 ; SSE2-NEXT: movdqa %xmm1, %xmm2
523 ; SSE2-NEXT: psrlw $6, %xmm2
524 ; SSE2-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm2
525 ; SSE2-NEXT: movdqa %xmm1, %xmm4
526 ; SSE2-NEXT: psllw $2, %xmm4
527 ; SSE2-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm4
528 ; SSE2-NEXT: por %xmm2, %xmm4
529 ; SSE2-NEXT: paddb %xmm3, %xmm3
530 ; SSE2-NEXT: pxor %xmm2, %xmm2
531 ; SSE2-NEXT: pcmpgtb %xmm3, %xmm2
532 ; SSE2-NEXT: pand %xmm2, %xmm4
533 ; SSE2-NEXT: pandn %xmm1, %xmm2
534 ; SSE2-NEXT: por %xmm4, %xmm2
535 ; SSE2-NEXT: movdqa %xmm2, %xmm1
536 ; SSE2-NEXT: psrlw $7, %xmm1
537 ; SSE2-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
538 ; SSE2-NEXT: movdqa %xmm2, %xmm4
539 ; SSE2-NEXT: paddb %xmm2, %xmm4
540 ; SSE2-NEXT: por %xmm1, %xmm4
541 ; SSE2-NEXT: paddb %xmm3, %xmm3
542 ; SSE2-NEXT: pcmpgtb %xmm3, %xmm0
543 ; SSE2-NEXT: pand %xmm0, %xmm4
544 ; SSE2-NEXT: pandn %xmm2, %xmm0
545 ; SSE2-NEXT: por %xmm4, %xmm0
546 ; SSE2-NEXT: retq
547 ;
548 ; SSE41-LABEL: var_funnnel_v16i8:
549 ; SSE41: # %bb.0:
550 ; SSE41-NEXT: movdqa %xmm0, %xmm2
551 ; SSE41-NEXT: psrlw $4, %xmm0
552 ; SSE41-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
553 ; SSE41-NEXT: movdqa %xmm2, %xmm3
554 ; SSE41-NEXT: psllw $4, %xmm3
555 ; SSE41-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
556 ; SSE41-NEXT: por %xmm0, %xmm3
557 ; SSE41-NEXT: pxor %xmm0, %xmm0
558 ; SSE41-NEXT: psubb %xmm1, %xmm0
559 ; SSE41-NEXT: psllw $5, %xmm0
560 ; SSE41-NEXT: pblendvb %xmm0, %xmm3, %xmm2
561 ; SSE41-NEXT: movdqa %xmm2, %xmm1
562 ; SSE41-NEXT: psrlw $6, %xmm1
563 ; SSE41-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
564 ; SSE41-NEXT: movdqa %xmm2, %xmm3
565 ; SSE41-NEXT: psllw $2, %xmm3
566 ; SSE41-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3
567 ; SSE41-NEXT: por %xmm1, %xmm3
568 ; SSE41-NEXT: paddb %xmm0, %xmm0
569 ; SSE41-NEXT: pblendvb %xmm0, %xmm3, %xmm2
570 ; SSE41-NEXT: movdqa %xmm2, %xmm1
571 ; SSE41-NEXT: psrlw $7, %xmm1
572 ; SSE41-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
573 ; SSE41-NEXT: movdqa %xmm2, %xmm3
574 ; SSE41-NEXT: paddb %xmm2, %xmm3
575 ; SSE41-NEXT: por %xmm1, %xmm3
576 ; SSE41-NEXT: paddb %xmm0, %xmm0
577 ; SSE41-NEXT: pblendvb %xmm0, %xmm3, %xmm2
578 ; SSE41-NEXT: movdqa %xmm2, %xmm0
579 ; SSE41-NEXT: retq
580 ;
581 ; AVX-LABEL: var_funnnel_v16i8:
582 ; AVX: # %bb.0:
583 ; AVX-NEXT: vpsrlw $4, %xmm0, %xmm2
584 ; AVX-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm2, %xmm2
585 ; AVX-NEXT: vpsllw $4, %xmm0, %xmm3
586 ; AVX-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3, %xmm3
587 ; AVX-NEXT: vpor %xmm2, %xmm3, %xmm2
588 ; AVX-NEXT: vpxor %xmm3, %xmm3, %xmm3
589 ; AVX-NEXT: vpsubb %xmm1, %xmm3, %xmm1
590 ; AVX-NEXT: vpsllw $5, %xmm1, %xmm1
591 ; AVX-NEXT: vpblendvb %xmm1, %xmm2, %xmm0, %xmm0
592 ; AVX-NEXT: vpsrlw $6, %xmm0, %xmm2
593 ; AVX-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm2, %xmm2
594 ; AVX-NEXT: vpsllw $2, %xmm0, %xmm3
595 ; AVX-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3, %xmm3
596 ; AVX-NEXT: vpor %xmm2, %xmm3, %xmm2
597 ; AVX-NEXT: vpaddb %xmm1, %xmm1, %xmm1
598 ; AVX-NEXT: vpblendvb %xmm1, %xmm2, %xmm0, %xmm0
599 ; AVX-NEXT: vpsrlw $7, %xmm0, %xmm2
600 ; AVX-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm2, %xmm2
601 ; AVX-NEXT: vpaddb %xmm0, %xmm0, %xmm3
602 ; AVX-NEXT: vpor %xmm2, %xmm3, %xmm2
603 ; AVX-NEXT: vpaddb %xmm1, %xmm1, %xmm1
604 ; AVX-NEXT: vpblendvb %xmm1, %xmm2, %xmm0, %xmm0
605 ; AVX-NEXT: retq
606 ;
607 ; AVX512F-LABEL: var_funnnel_v16i8:
608 ; AVX512F: # %bb.0:
609 ; AVX512F-NEXT: vpmovzxbd {{.*#+}} zmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero,xmm0[4],zero,zero,zero,xmm0[5],zero,zero,zero,xmm0[6],zero,zero,zero,xmm0[7],zero,zero,zero,xmm0[8],zero,zero,zero,xmm0[9],zero,zero,zero,xmm0[10],zero,zero,zero,xmm0[11],zero,zero,zero,xmm0[12],zero,zero,zero,xmm0[13],zero,zero,zero,xmm0[14],zero,zero,zero,xmm0[15],zero,zero,zero
610 ; AVX512F-NEXT: vpslld $8, %zmm0, %zmm2
611 ; AVX512F-NEXT: vpord %zmm2, %zmm0, %zmm0
612 ; AVX512F-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1, %xmm1
613 ; AVX512F-NEXT: vpmovzxbd {{.*#+}} zmm1 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero,xmm1[2],zero,zero,zero,xmm1[3],zero,zero,zero,xmm1[4],zero,zero,zero,xmm1[5],zero,zero,zero,xmm1[6],zero,zero,zero,xmm1[7],zero,zero,zero,xmm1[8],zero,zero,zero,xmm1[9],zero,zero,zero,xmm1[10],zero,zero,zero,xmm1[11],zero,zero,zero,xmm1[12],zero,zero,zero,xmm1[13],zero,zero,zero,xmm1[14],zero,zero,zero,xmm1[15],zero,zero,zero
614 ; AVX512F-NEXT: vpsrlvd %zmm1, %zmm0, %zmm0
615 ; AVX512F-NEXT: vpmovdb %zmm0, %xmm0
616 ; AVX512F-NEXT: vzeroupper
617 ; AVX512F-NEXT: retq
618 ;
619 ; AVX512VL-LABEL: var_funnnel_v16i8:
620 ; AVX512VL: # %bb.0:
621 ; AVX512VL-NEXT: vpmovzxbd {{.*#+}} zmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero,xmm0[4],zero,zero,zero,xmm0[5],zero,zero,zero,xmm0[6],zero,zero,zero,xmm0[7],zero,zero,zero,xmm0[8],zero,zero,zero,xmm0[9],zero,zero,zero,xmm0[10],zero,zero,zero,xmm0[11],zero,zero,zero,xmm0[12],zero,zero,zero,xmm0[13],zero,zero,zero,xmm0[14],zero,zero,zero,xmm0[15],zero,zero,zero
622 ; AVX512VL-NEXT: vpslld $8, %zmm0, %zmm2
623 ; AVX512VL-NEXT: vpord %zmm2, %zmm0, %zmm0
624 ; AVX512VL-NEXT: vpandd {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to4}, %xmm1, %xmm1
625 ; AVX512VL-NEXT: vpmovzxbd {{.*#+}} zmm1 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero,xmm1[2],zero,zero,zero,xmm1[3],zero,zero,zero,xmm1[4],zero,zero,zero,xmm1[5],zero,zero,zero,xmm1[6],zero,zero,zero,xmm1[7],zero,zero,zero,xmm1[8],zero,zero,zero,xmm1[9],zero,zero,zero,xmm1[10],zero,zero,zero,xmm1[11],zero,zero,zero,xmm1[12],zero,zero,zero,xmm1[13],zero,zero,zero,xmm1[14],zero,zero,zero,xmm1[15],zero,zero,zero
626 ; AVX512VL-NEXT: vpsrlvd %zmm1, %zmm0, %zmm0
627 ; AVX512VL-NEXT: vpmovdb %zmm0, %xmm0
628 ; AVX512VL-NEXT: vzeroupper
629 ; AVX512VL-NEXT: retq
630 ;
631 ; AVX512BW-LABEL: var_funnnel_v16i8:
632 ; AVX512BW: # %bb.0:
633 ; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm0[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
634 ; AVX512BW-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1, %xmm1
635 ; AVX512BW-NEXT: vpxor %xmm3, %xmm3, %xmm3
636 ; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm1[8],xmm3[8],xmm1[9],xmm3[9],xmm1[10],xmm3[10],xmm1[11],xmm3[11],xmm1[12],xmm3[12],xmm1[13],xmm3[13],xmm1[14],xmm3[14],xmm1[15],xmm3[15]
637 ; AVX512BW-NEXT: vpsrlvw %zmm3, %zmm2, %zmm2
638 ; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm3 = [255,255,255,255,255,255,255,255]
639 ; AVX512BW-NEXT: vpand %xmm3, %xmm2, %xmm2
640 ; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
641 ; AVX512BW-NEXT: vpmovzxbw {{.*#+}} xmm1 = xmm1[0],zero,xmm1[1],zero,xmm1[2],zero,xmm1[3],zero,xmm1[4],zero,xmm1[5],zero,xmm1[6],zero,xmm1[7],zero
642 ; AVX512BW-NEXT: vpsrlvw %zmm1, %zmm0, %zmm0
643 ; AVX512BW-NEXT: vpand %xmm3, %xmm0, %xmm0
644 ; AVX512BW-NEXT: vpackuswb %xmm2, %xmm0, %xmm0
645 ; AVX512BW-NEXT: vzeroupper
646 ; AVX512BW-NEXT: retq
647 ;
648 ; AVX512VLBW-LABEL: var_funnnel_v16i8:
649 ; AVX512VLBW: # %bb.0:
650 ; AVX512VLBW-NEXT: vpandd {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to4}, %xmm1, %xmm1
651 ; AVX512VLBW-NEXT: vpxor %xmm2, %xmm2, %xmm2
652 ; AVX512VLBW-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm1[8],xmm2[8],xmm1[9],xmm2[9],xmm1[10],xmm2[10],xmm1[11],xmm2[11],xmm1[12],xmm2[12],xmm1[13],xmm2[13],xmm1[14],xmm2[14],xmm1[15],xmm2[15]
653 ; AVX512VLBW-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm0[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
654 ; AVX512VLBW-NEXT: vpsrlvw %xmm2, %xmm3, %xmm2
655 ; AVX512VLBW-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
656 ; AVX512VLBW-NEXT: vpmovzxbw {{.*#+}} xmm1 = xmm1[0],zero,xmm1[1],zero,xmm1[2],zero,xmm1[3],zero,xmm1[4],zero,xmm1[5],zero,xmm1[6],zero,xmm1[7],zero
657 ; AVX512VLBW-NEXT: vpsrlvw %xmm1, %xmm0, %xmm0
658 ; AVX512VLBW-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm0
659 ; AVX512VLBW-NEXT: vpmovwb %ymm0, %xmm0
660 ; AVX512VLBW-NEXT: vzeroupper
661 ; AVX512VLBW-NEXT: retq
662 ;
663 ; AVX512VBMI2-LABEL: var_funnnel_v16i8:
664 ; AVX512VBMI2: # %bb.0:
665 ; AVX512VBMI2-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm0[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
666 ; AVX512VBMI2-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1, %xmm1
667 ; AVX512VBMI2-NEXT: vpxor %xmm3, %xmm3, %xmm3
668 ; AVX512VBMI2-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm1[8],xmm3[8],xmm1[9],xmm3[9],xmm1[10],xmm3[10],xmm1[11],xmm3[11],xmm1[12],xmm3[12],xmm1[13],xmm3[13],xmm1[14],xmm3[14],xmm1[15],xmm3[15]
669 ; AVX512VBMI2-NEXT: vpsrlvw %zmm3, %zmm2, %zmm2
670 ; AVX512VBMI2-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
671 ; AVX512VBMI2-NEXT: vpmovzxbw {{.*#+}} xmm1 = xmm1[0],zero,xmm1[1],zero,xmm1[2],zero,xmm1[3],zero,xmm1[4],zero,xmm1[5],zero,xmm1[6],zero,xmm1[7],zero
672 ; AVX512VBMI2-NEXT: vpsrlvw %zmm1, %zmm0, %zmm1
673 ; AVX512VBMI2-NEXT: vmovdqa {{.*#+}} xmm0 = [0,2,4,6,8,10,12,14,64,66,68,70,72,74,76,78]
674 ; AVX512VBMI2-NEXT: vpermi2b %zmm2, %zmm1, %zmm0
675 ; AVX512VBMI2-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
676 ; AVX512VBMI2-NEXT: vzeroupper
677 ; AVX512VBMI2-NEXT: retq
678 ;
679 ; AVX512VLVBMI2-LABEL: var_funnnel_v16i8:
680 ; AVX512VLVBMI2: # %bb.0:
681 ; AVX512VLVBMI2-NEXT: vpandd {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to4}, %xmm1, %xmm1
682 ; AVX512VLVBMI2-NEXT: vpxor %xmm2, %xmm2, %xmm2
683 ; AVX512VLVBMI2-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm1[8],xmm2[8],xmm1[9],xmm2[9],xmm1[10],xmm2[10],xmm1[11],xmm2[11],xmm1[12],xmm2[12],xmm1[13],xmm2[13],xmm1[14],xmm2[14],xmm1[15],xmm2[15]
684 ; AVX512VLVBMI2-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm0[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
685 ; AVX512VLVBMI2-NEXT: vpsrlvw %xmm2, %xmm3, %xmm2
686 ; AVX512VLVBMI2-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
687 ; AVX512VLVBMI2-NEXT: vpmovzxbw {{.*#+}} xmm1 = xmm1[0],zero,xmm1[1],zero,xmm1[2],zero,xmm1[3],zero,xmm1[4],zero,xmm1[5],zero,xmm1[6],zero,xmm1[7],zero
688 ; AVX512VLVBMI2-NEXT: vpsrlvw %xmm1, %xmm0, %xmm0
689 ; AVX512VLVBMI2-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm0
690 ; AVX512VLVBMI2-NEXT: vpmovwb %ymm0, %xmm0
691 ; AVX512VLVBMI2-NEXT: vzeroupper
692 ; AVX512VLVBMI2-NEXT: retq
693 ;
694 ; XOP-LABEL: var_funnnel_v16i8:
695 ; XOP: # %bb.0:
696 ; XOP-NEXT: vpxor %xmm2, %xmm2, %xmm2
697 ; XOP-NEXT: vpsubb %xmm1, %xmm2, %xmm1
698 ; XOP-NEXT: vprotb %xmm1, %xmm0, %xmm0
699 ; XOP-NEXT: retq
700 ;
701 ; X86-SSE2-LABEL: var_funnnel_v16i8:
702 ; X86-SSE2: # %bb.0:
703 ; X86-SSE2-NEXT: movdqa %xmm0, %xmm2
704 ; X86-SSE2-NEXT: pxor %xmm0, %xmm0
705 ; X86-SSE2-NEXT: pxor %xmm3, %xmm3
706 ; X86-SSE2-NEXT: psubb %xmm1, %xmm3
707 ; X86-SSE2-NEXT: psllw $5, %xmm3
708 ; X86-SSE2-NEXT: pxor %xmm1, %xmm1
709 ; X86-SSE2-NEXT: pcmpgtb %xmm3, %xmm1
710 ; X86-SSE2-NEXT: movdqa %xmm2, %xmm4
711 ; X86-SSE2-NEXT: psrlw $4, %xmm4
712 ; X86-SSE2-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}, %xmm4
713 ; X86-SSE2-NEXT: movdqa %xmm2, %xmm5
714 ; X86-SSE2-NEXT: psllw $4, %xmm5
715 ; X86-SSE2-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}, %xmm5
716 ; X86-SSE2-NEXT: por %xmm4, %xmm5
717 ; X86-SSE2-NEXT: pand %xmm1, %xmm5
718 ; X86-SSE2-NEXT: pandn %xmm2, %xmm1
719 ; X86-SSE2-NEXT: por %xmm5, %xmm1
720 ; X86-SSE2-NEXT: movdqa %xmm1, %xmm2
721 ; X86-SSE2-NEXT: psrlw $6, %xmm2
722 ; X86-SSE2-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}, %xmm2
723 ; X86-SSE2-NEXT: movdqa %xmm1, %xmm4
724 ; X86-SSE2-NEXT: psllw $2, %xmm4
725 ; X86-SSE2-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}, %xmm4
726 ; X86-SSE2-NEXT: por %xmm2, %xmm4
727 ; X86-SSE2-NEXT: paddb %xmm3, %xmm3
728 ; X86-SSE2-NEXT: pxor %xmm2, %xmm2
729 ; X86-SSE2-NEXT: pcmpgtb %xmm3, %xmm2
730 ; X86-SSE2-NEXT: pand %xmm2, %xmm4
731 ; X86-SSE2-NEXT: pandn %xmm1, %xmm2
732 ; X86-SSE2-NEXT: por %xmm4, %xmm2
733 ; X86-SSE2-NEXT: movdqa %xmm2, %xmm1
734 ; X86-SSE2-NEXT: psrlw $7, %xmm1
735 ; X86-SSE2-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}, %xmm1
736 ; X86-SSE2-NEXT: movdqa %xmm2, %xmm4
737 ; X86-SSE2-NEXT: paddb %xmm2, %xmm4
738 ; X86-SSE2-NEXT: por %xmm1, %xmm4
739 ; X86-SSE2-NEXT: paddb %xmm3, %xmm3
740 ; X86-SSE2-NEXT: pcmpgtb %xmm3, %xmm0
741 ; X86-SSE2-NEXT: pand %xmm0, %xmm4
742 ; X86-SSE2-NEXT: pandn %xmm2, %xmm0
743 ; X86-SSE2-NEXT: por %xmm4, %xmm0
744 ; X86-SSE2-NEXT: retl
745 %res = call <16 x i8> @llvm.fshr.v16i8(<16 x i8> %x, <16 x i8> %x, <16 x i8> %amt)
746 ret <16 x i8> %res
747 }
749 ;
750 ; Uniform Variable Shifts
751 ;
753 define <2 x i64> @splatvar_funnnel_v2i64(<2 x i64> %x, <2 x i64> %amt) nounwind {
754 ; SSE2-LABEL: splatvar_funnnel_v2i64:
755 ; SSE2: # %bb.0:
756 ; SSE2-NEXT: movdqa {{.*#+}} xmm2 = [63,63]
757 ; SSE2-NEXT: pxor %xmm3, %xmm3
758 ; SSE2-NEXT: psubq %xmm1, %xmm3
759 ; SSE2-NEXT: pand %xmm2, %xmm1
760 ; SSE2-NEXT: movdqa %xmm0, %xmm4
761 ; SSE2-NEXT: psrlq %xmm1, %xmm4
762 ; SSE2-NEXT: pand %xmm2, %xmm3
763 ; SSE2-NEXT: psllq %xmm3, %xmm0
764 ; SSE2-NEXT: por %xmm4, %xmm0
765 ; SSE2-NEXT: retq
766 ;
767 ; SSE41-LABEL: splatvar_funnnel_v2i64:
768 ; SSE41: # %bb.0:
769 ; SSE41-NEXT: pmovsxbq {{.*#+}} xmm2 = [63,63]
770 ; SSE41-NEXT: pxor %xmm3, %xmm3
771 ; SSE41-NEXT: psubq %xmm1, %xmm3
772 ; SSE41-NEXT: pand %xmm2, %xmm1
773 ; SSE41-NEXT: movdqa %xmm0, %xmm4
774 ; SSE41-NEXT: psrlq %xmm1, %xmm4
775 ; SSE41-NEXT: pand %xmm2, %xmm3
776 ; SSE41-NEXT: psllq %xmm3, %xmm0
777 ; SSE41-NEXT: por %xmm4, %xmm0
778 ; SSE41-NEXT: retq
779 ;
780 ; AVX-LABEL: splatvar_funnnel_v2i64:
781 ; AVX: # %bb.0:
782 ; AVX-NEXT: vpmovsxbq {{.*#+}} xmm2 = [63,63]
783 ; AVX-NEXT: vpand %xmm2, %xmm1, %xmm3
784 ; AVX-NEXT: vpsrlq %xmm3, %xmm0, %xmm3
785 ; AVX-NEXT: vpxor %xmm4, %xmm4, %xmm4
786 ; AVX-NEXT: vpsubq %xmm1, %xmm4, %xmm1
787 ; AVX-NEXT: vpand %xmm2, %xmm1, %xmm1
788 ; AVX-NEXT: vpsllq %xmm1, %xmm0, %xmm0
789 ; AVX-NEXT: vpor %xmm0, %xmm3, %xmm0
790 ; AVX-NEXT: retq
791 ;
792 ; AVX512F-LABEL: splatvar_funnnel_v2i64:
793 ; AVX512F: # %bb.0:
794 ; AVX512F-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
795 ; AVX512F-NEXT: vpbroadcastq %xmm1, %xmm1
796 ; AVX512F-NEXT: vprorvq %zmm1, %zmm0, %zmm0
797 ; AVX512F-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
798 ; AVX512F-NEXT: vzeroupper
799 ; AVX512F-NEXT: retq
800 ;
801 ; AVX512VL-LABEL: splatvar_funnnel_v2i64:
802 ; AVX512VL: # %bb.0:
803 ; AVX512VL-NEXT: vpbroadcastq %xmm1, %xmm1
804 ; AVX512VL-NEXT: vprorvq %xmm1, %xmm0, %xmm0
805 ; AVX512VL-NEXT: retq
806 ;
807 ; AVX512BW-LABEL: splatvar_funnnel_v2i64:
808 ; AVX512BW: # %bb.0:
809 ; AVX512BW-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
810 ; AVX512BW-NEXT: vpbroadcastq %xmm1, %xmm1
811 ; AVX512BW-NEXT: vprorvq %zmm1, %zmm0, %zmm0
812 ; AVX512BW-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
813 ; AVX512BW-NEXT: vzeroupper
814 ; AVX512BW-NEXT: retq
815 ;
816 ; AVX512VLBW-LABEL: splatvar_funnnel_v2i64:
817 ; AVX512VLBW: # %bb.0:
818 ; AVX512VLBW-NEXT: vpbroadcastq %xmm1, %xmm1
819 ; AVX512VLBW-NEXT: vprorvq %xmm1, %xmm0, %xmm0
820 ; AVX512VLBW-NEXT: retq
821 ;
822 ; AVX512VBMI2-LABEL: splatvar_funnnel_v2i64:
823 ; AVX512VBMI2: # %bb.0:
824 ; AVX512VBMI2-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
825 ; AVX512VBMI2-NEXT: vpbroadcastq %xmm1, %xmm1
826 ; AVX512VBMI2-NEXT: vprorvq %zmm1, %zmm0, %zmm0
827 ; AVX512VBMI2-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
828 ; AVX512VBMI2-NEXT: vzeroupper
829 ; AVX512VBMI2-NEXT: retq
830 ;
831 ; AVX512VLVBMI2-LABEL: splatvar_funnnel_v2i64:
832 ; AVX512VLVBMI2: # %bb.0:
833 ; AVX512VLVBMI2-NEXT: vpbroadcastq %xmm1, %xmm1
834 ; AVX512VLVBMI2-NEXT: vprorvq %xmm1, %xmm0, %xmm0
835 ; AVX512VLVBMI2-NEXT: retq
836 ;
837 ; XOPAVX1-LABEL: splatvar_funnnel_v2i64:
838 ; XOPAVX1: # %bb.0:
839 ; XOPAVX1-NEXT: vpxor %xmm2, %xmm2, %xmm2
840 ; XOPAVX1-NEXT: vpsubq %xmm1, %xmm2, %xmm1
841 ; XOPAVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[0,1,0,1]
842 ; XOPAVX1-NEXT: vprotq %xmm1, %xmm0, %xmm0
843 ; XOPAVX1-NEXT: retq
844 ;
845 ; XOPAVX2-LABEL: splatvar_funnnel_v2i64:
846 ; XOPAVX2: # %bb.0:
847 ; XOPAVX2-NEXT: vpbroadcastq %xmm1, %xmm1
848 ; XOPAVX2-NEXT: vpxor %xmm2, %xmm2, %xmm2
849 ; XOPAVX2-NEXT: vpsubq %xmm1, %xmm2, %xmm1
850 ; XOPAVX2-NEXT: vprotq %xmm1, %xmm0, %xmm0
851 ; XOPAVX2-NEXT: retq
852 ;
853 ; X86-SSE2-LABEL: splatvar_funnnel_v2i64:
854 ; X86-SSE2: # %bb.0:
855 ; X86-SSE2-NEXT: movdqa {{.*#+}} xmm2 = [63,0,63,0]
856 ; X86-SSE2-NEXT: pxor %xmm3, %xmm3
857 ; X86-SSE2-NEXT: psubq %xmm1, %xmm3
858 ; X86-SSE2-NEXT: pand %xmm2, %xmm1
859 ; X86-SSE2-NEXT: movdqa %xmm0, %xmm4
860 ; X86-SSE2-NEXT: psrlq %xmm1, %xmm4
861 ; X86-SSE2-NEXT: pand %xmm2, %xmm3
862 ; X86-SSE2-NEXT: psllq %xmm3, %xmm0
863 ; X86-SSE2-NEXT: por %xmm4, %xmm0
864 ; X86-SSE2-NEXT: retl
865 %splat = shufflevector <2 x i64> %amt, <2 x i64> undef, <2 x i32> zeroinitializer
866 %res = call <2 x i64> @llvm.fshr.v2i64(<2 x i64> %x, <2 x i64> %x, <2 x i64> %splat)
867 ret <2 x i64> %res
868 }
870 define <4 x i32> @splatvar_funnnel_v4i32(<4 x i32> %x, <4 x i32> %amt) nounwind {
871 ; SSE-LABEL: splatvar_funnnel_v4i32:
872 ; SSE: # %bb.0:
873 ; SSE-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
874 ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm0[2,2,3,3]
875 ; SSE-NEXT: psrlq %xmm1, %xmm2
876 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,0,1,1]
877 ; SSE-NEXT: psrlq %xmm1, %xmm0
878 ; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[0,2],xmm2[0,2]
879 ; SSE-NEXT: retq
880 ;
881 ; AVX-LABEL: splatvar_funnnel_v4i32:
882 ; AVX: # %bb.0:
883 ; AVX-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1, %xmm1
884 ; AVX-NEXT: vpshufd {{.*#+}} xmm2 = xmm0[2,2,3,3]
885 ; AVX-NEXT: vpsrlq %xmm1, %xmm2, %xmm2
886 ; AVX-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,0,1,1]
887 ; AVX-NEXT: vpsrlq %xmm1, %xmm0, %xmm0
888 ; AVX-NEXT: vshufps {{.*#+}} xmm0 = xmm0[0,2],xmm2[0,2]
889 ; AVX-NEXT: retq
890 ;
891 ; AVX512F-LABEL: splatvar_funnnel_v4i32:
892 ; AVX512F: # %bb.0:
893 ; AVX512F-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
894 ; AVX512F-NEXT: vpbroadcastd %xmm1, %xmm1
895 ; AVX512F-NEXT: vprorvd %zmm1, %zmm0, %zmm0
896 ; AVX512F-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
897 ; AVX512F-NEXT: vzeroupper
898 ; AVX512F-NEXT: retq
899 ;
900 ; AVX512VL-LABEL: splatvar_funnnel_v4i32:
901 ; AVX512VL: # %bb.0:
902 ; AVX512VL-NEXT: vpbroadcastd %xmm1, %xmm1
903 ; AVX512VL-NEXT: vprorvd %xmm1, %xmm0, %xmm0
904 ; AVX512VL-NEXT: retq
905 ;
906 ; AVX512BW-LABEL: splatvar_funnnel_v4i32:
907 ; AVX512BW: # %bb.0:
908 ; AVX512BW-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
909 ; AVX512BW-NEXT: vpbroadcastd %xmm1, %xmm1
910 ; AVX512BW-NEXT: vprorvd %zmm1, %zmm0, %zmm0
911 ; AVX512BW-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
912 ; AVX512BW-NEXT: vzeroupper
913 ; AVX512BW-NEXT: retq
914 ;
915 ; AVX512VLBW-LABEL: splatvar_funnnel_v4i32:
916 ; AVX512VLBW: # %bb.0:
917 ; AVX512VLBW-NEXT: vpbroadcastd %xmm1, %xmm1
918 ; AVX512VLBW-NEXT: vprorvd %xmm1, %xmm0, %xmm0
919 ; AVX512VLBW-NEXT: retq
920 ;
921 ; AVX512VBMI2-LABEL: splatvar_funnnel_v4i32:
922 ; AVX512VBMI2: # %bb.0:
923 ; AVX512VBMI2-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
924 ; AVX512VBMI2-NEXT: vpbroadcastd %xmm1, %xmm1
925 ; AVX512VBMI2-NEXT: vprorvd %zmm1, %zmm0, %zmm0
926 ; AVX512VBMI2-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
927 ; AVX512VBMI2-NEXT: vzeroupper
928 ; AVX512VBMI2-NEXT: retq
929 ;
930 ; AVX512VLVBMI2-LABEL: splatvar_funnnel_v4i32:
931 ; AVX512VLVBMI2: # %bb.0:
932 ; AVX512VLVBMI2-NEXT: vpbroadcastd %xmm1, %xmm1
933 ; AVX512VLVBMI2-NEXT: vprorvd %xmm1, %xmm0, %xmm0
934 ; AVX512VLVBMI2-NEXT: retq
935 ;
936 ; XOPAVX1-LABEL: splatvar_funnnel_v4i32:
937 ; XOPAVX1: # %bb.0:
938 ; XOPAVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[0,0,0,0]
939 ; XOPAVX1-NEXT: vpxor %xmm2, %xmm2, %xmm2
940 ; XOPAVX1-NEXT: vpsubd %xmm1, %xmm2, %xmm1
941 ; XOPAVX1-NEXT: vprotd %xmm1, %xmm0, %xmm0
942 ; XOPAVX1-NEXT: retq
943 ;
944 ; XOPAVX2-LABEL: splatvar_funnnel_v4i32:
945 ; XOPAVX2: # %bb.0:
946 ; XOPAVX2-NEXT: vpbroadcastd %xmm1, %xmm1
947 ; XOPAVX2-NEXT: vpxor %xmm2, %xmm2, %xmm2
948 ; XOPAVX2-NEXT: vpsubd %xmm1, %xmm2, %xmm1
949 ; XOPAVX2-NEXT: vprotd %xmm1, %xmm0, %xmm0
950 ; XOPAVX2-NEXT: retq
951 ;
952 ; X86-SSE2-LABEL: splatvar_funnnel_v4i32:
953 ; X86-SSE2: # %bb.0:
954 ; X86-SSE2-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}, %xmm1
955 ; X86-SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm0[2,2,3,3]
956 ; X86-SSE2-NEXT: psrlq %xmm1, %xmm2
957 ; X86-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,0,1,1]
958 ; X86-SSE2-NEXT: psrlq %xmm1, %xmm0
959 ; X86-SSE2-NEXT: shufps {{.*#+}} xmm0 = xmm0[0,2],xmm2[0,2]
960 ; X86-SSE2-NEXT: retl
961 %splat = shufflevector <4 x i32> %amt, <4 x i32> undef, <4 x i32> zeroinitializer
962 %res = call <4 x i32> @llvm.fshr.v4i32(<4 x i32> %x, <4 x i32> %x, <4 x i32> %splat)
963 ret <4 x i32> %res
964 }
966 define <8 x i16> @splatvar_funnnel_v8i16(<8 x i16> %x, <8 x i16> %amt) nounwind {
967 ; SSE2-LABEL: splatvar_funnnel_v8i16:
968 ; SSE2: # %bb.0:
969 ; SSE2-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
970 ; SSE2-NEXT: movdqa %xmm0, %xmm2
971 ; SSE2-NEXT: punpckhwd {{.*#+}} xmm2 = xmm2[4],xmm0[4],xmm2[5],xmm0[5],xmm2[6],xmm0[6],xmm2[7],xmm0[7]
972 ; SSE2-NEXT: psrld %xmm1, %xmm2
973 ; SSE2-NEXT: pslld $16, %xmm2
974 ; SSE2-NEXT: psrad $16, %xmm2
975 ; SSE2-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3]
976 ; SSE2-NEXT: psrld %xmm1, %xmm0
977 ; SSE2-NEXT: pslld $16, %xmm0
978 ; SSE2-NEXT: psrad $16, %xmm0
979 ; SSE2-NEXT: packssdw %xmm2, %xmm0
980 ; SSE2-NEXT: retq
981 ;
982 ; SSE41-LABEL: splatvar_funnnel_v8i16:
983 ; SSE41: # %bb.0:
984 ; SSE41-NEXT: pmovsxbq {{.*#+}} xmm2 = [15,0]
985 ; SSE41-NEXT: movdqa %xmm1, %xmm3
986 ; SSE41-NEXT: pand %xmm2, %xmm3
987 ; SSE41-NEXT: movdqa %xmm0, %xmm4
988 ; SSE41-NEXT: psrlw %xmm3, %xmm4
989 ; SSE41-NEXT: pandn %xmm2, %xmm1
990 ; SSE41-NEXT: paddw %xmm0, %xmm0
991 ; SSE41-NEXT: psllw %xmm1, %xmm0
992 ; SSE41-NEXT: por %xmm4, %xmm0
993 ; SSE41-NEXT: retq
994 ;
995 ; AVX-LABEL: splatvar_funnnel_v8i16:
996 ; AVX: # %bb.0:
997 ; AVX-NEXT: vpmovsxbq {{.*#+}} xmm2 = [15,0]
998 ; AVX-NEXT: vpand %xmm2, %xmm1, %xmm3
999 ; AVX-NEXT: vpsrlw %xmm3, %xmm0, %xmm3
1000 ; AVX-NEXT: vpandn %xmm2, %xmm1, %xmm1
1001 ; AVX-NEXT: vpaddw %xmm0, %xmm0, %xmm0
1002 ; AVX-NEXT: vpsllw %xmm1, %xmm0, %xmm0
1003 ; AVX-NEXT: vpor %xmm3, %xmm0, %xmm0
1004 ; AVX-NEXT: retq
1005 ;
1006 ; AVX512F-LABEL: splatvar_funnnel_v8i16:
1007 ; AVX512F: # %bb.0:
1008 ; AVX512F-NEXT: vpmovsxbq {{.*#+}} xmm2 = [15,0]
1009 ; AVX512F-NEXT: vpand %xmm2, %xmm1, %xmm3
1010 ; AVX512F-NEXT: vpsrlw %xmm3, %xmm0, %xmm3
1011 ; AVX512F-NEXT: vpandn %xmm2, %xmm1, %xmm1
1012 ; AVX512F-NEXT: vpaddw %xmm0, %xmm0, %xmm0
1013 ; AVX512F-NEXT: vpsllw %xmm1, %xmm0, %xmm0
1014 ; AVX512F-NEXT: vpor %xmm3, %xmm0, %xmm0
1015 ; AVX512F-NEXT: retq
1016 ;
1017 ; AVX512VL-LABEL: splatvar_funnnel_v8i16:
1018 ; AVX512VL: # %bb.0:
1019 ; AVX512VL-NEXT: vpmovsxbq {{.*#+}} xmm2 = [15,0]
1020 ; AVX512VL-NEXT: vpand %xmm2, %xmm1, %xmm3
1021 ; AVX512VL-NEXT: vpsrlw %xmm3, %xmm0, %xmm3
1022 ; AVX512VL-NEXT: vpandn %xmm2, %xmm1, %xmm1
1023 ; AVX512VL-NEXT: vpaddw %xmm0, %xmm0, %xmm0
1024 ; AVX512VL-NEXT: vpsllw %xmm1, %xmm0, %xmm0
1025 ; AVX512VL-NEXT: vpor %xmm3, %xmm0, %xmm0
1026 ; AVX512VL-NEXT: retq
1027 ;
1028 ; AVX512BW-LABEL: splatvar_funnnel_v8i16:
1029 ; AVX512BW: # %bb.0:
1030 ; AVX512BW-NEXT: vpmovsxbq {{.*#+}} xmm2 = [15,0]
1031 ; AVX512BW-NEXT: vpand %xmm2, %xmm1, %xmm3
1032 ; AVX512BW-NEXT: vpsrlw %xmm3, %xmm0, %xmm3
1033 ; AVX512BW-NEXT: vpandn %xmm2, %xmm1, %xmm1
1034 ; AVX512BW-NEXT: vpaddw %xmm0, %xmm0, %xmm0
1035 ; AVX512BW-NEXT: vpsllw %xmm1, %xmm0, %xmm0
1036 ; AVX512BW-NEXT: vpor %xmm3, %xmm0, %xmm0
1037 ; AVX512BW-NEXT: retq
1038 ;
1039 ; AVX512VLBW-LABEL: splatvar_funnnel_v8i16:
1040 ; AVX512VLBW: # %bb.0:
1041 ; AVX512VLBW-NEXT: vpmovsxbq {{.*#+}} xmm2 = [15,0]
1042 ; AVX512VLBW-NEXT: vpand %xmm2, %xmm1, %xmm3
1043 ; AVX512VLBW-NEXT: vpsrlw %xmm3, %xmm0, %xmm3
1044 ; AVX512VLBW-NEXT: vpandn %xmm2, %xmm1, %xmm1
1045 ; AVX512VLBW-NEXT: vpaddw %xmm0, %xmm0, %xmm0
1046 ; AVX512VLBW-NEXT: vpsllw %xmm1, %xmm0, %xmm0
1047 ; AVX512VLBW-NEXT: vpor %xmm3, %xmm0, %xmm0
1048 ; AVX512VLBW-NEXT: retq
1049 ;
1050 ; AVX512VBMI2-LABEL: splatvar_funnnel_v8i16:
1051 ; AVX512VBMI2: # %bb.0:
1052 ; AVX512VBMI2-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
1053 ; AVX512VBMI2-NEXT: vpbroadcastw %xmm1, %xmm1
1054 ; AVX512VBMI2-NEXT: vpshrdvw %zmm1, %zmm0, %zmm0
1055 ; AVX512VBMI2-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
1056 ; AVX512VBMI2-NEXT: vzeroupper
1057 ; AVX512VBMI2-NEXT: retq
1058 ;
1059 ; AVX512VLVBMI2-LABEL: splatvar_funnnel_v8i16:
1060 ; AVX512VLVBMI2: # %bb.0:
1061 ; AVX512VLVBMI2-NEXT: vpbroadcastw %xmm1, %xmm1
1062 ; AVX512VLVBMI2-NEXT: vpshrdvw %xmm1, %xmm0, %xmm0
1063 ; AVX512VLVBMI2-NEXT: retq
1064 ;
1065 ; XOPAVX1-LABEL: splatvar_funnnel_v8i16:
1066 ; XOPAVX1: # %bb.0:
1067 ; XOPAVX1-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[0,0,0,0,4,5,6,7]
1068 ; XOPAVX1-NEXT: vpxor %xmm2, %xmm2, %xmm2
1069 ; XOPAVX1-NEXT: vpsubw %xmm1, %xmm2, %xmm1
1070 ; XOPAVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[0,0,0,0]
1071 ; XOPAVX1-NEXT: vprotw %xmm1, %xmm0, %xmm0
1072 ; XOPAVX1-NEXT: retq
1073 ;
1074 ; XOPAVX2-LABEL: splatvar_funnnel_v8i16:
1075 ; XOPAVX2: # %bb.0:
1076 ; XOPAVX2-NEXT: vpbroadcastw %xmm1, %xmm1
1077 ; XOPAVX2-NEXT: vpxor %xmm2, %xmm2, %xmm2
1078 ; XOPAVX2-NEXT: vpsubw %xmm1, %xmm2, %xmm1
1079 ; XOPAVX2-NEXT: vprotw %xmm1, %xmm0, %xmm0
1080 ; XOPAVX2-NEXT: retq
1081 ;
1082 ; X86-SSE2-LABEL: splatvar_funnnel_v8i16:
1083 ; X86-SSE2: # %bb.0:
1084 ; X86-SSE2-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}, %xmm1
1085 ; X86-SSE2-NEXT: movdqa %xmm0, %xmm2
1086 ; X86-SSE2-NEXT: punpckhwd {{.*#+}} xmm2 = xmm2[4],xmm0[4],xmm2[5],xmm0[5],xmm2[6],xmm0[6],xmm2[7],xmm0[7]
1087 ; X86-SSE2-NEXT: psrld %xmm1, %xmm2
1088 ; X86-SSE2-NEXT: pslld $16, %xmm2
1089 ; X86-SSE2-NEXT: psrad $16, %xmm2
1090 ; X86-SSE2-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3]
1091 ; X86-SSE2-NEXT: psrld %xmm1, %xmm0
1092 ; X86-SSE2-NEXT: pslld $16, %xmm0
1093 ; X86-SSE2-NEXT: psrad $16, %xmm0
1094 ; X86-SSE2-NEXT: packssdw %xmm2, %xmm0
1095 ; X86-SSE2-NEXT: retl
1096 %splat = shufflevector <8 x i16> %amt, <8 x i16> undef, <8 x i32> zeroinitializer
1097 %res = call <8 x i16> @llvm.fshr.v8i16(<8 x i16> %x, <8 x i16> %x, <8 x i16> %splat)
1098 ret <8 x i16> %res
1099 }
1101 define <16 x i8> @splatvar_funnnel_v16i8(<16 x i8> %x, <16 x i8> %amt) nounwind {
1102 ; SSE2-LABEL: splatvar_funnnel_v16i8:
1103 ; SSE2: # %bb.0:
1104 ; SSE2-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
1105 ; SSE2-NEXT: movdqa %xmm0, %xmm2
1106 ; SSE2-NEXT: punpckhbw {{.*#+}} xmm2 = xmm2[8],xmm0[8],xmm2[9],xmm0[9],xmm2[10],xmm0[10],xmm2[11],xmm0[11],xmm2[12],xmm0[12],xmm2[13],xmm0[13],xmm2[14],xmm0[14],xmm2[15],xmm0[15]
1107 ; SSE2-NEXT: psrlw %xmm1, %xmm2
1108 ; SSE2-NEXT: movdqa {{.*#+}} xmm3 = [255,255,255,255,255,255,255,255]
1109 ; SSE2-NEXT: pand %xmm3, %xmm2
1110 ; SSE2-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
1111 ; SSE2-NEXT: psrlw %xmm1, %xmm0
1112 ; SSE2-NEXT: pand %xmm3, %xmm0
1113 ; SSE2-NEXT: packuswb %xmm2, %xmm0
1114 ; SSE2-NEXT: retq
1115 ;
1116 ; SSE41-LABEL: splatvar_funnnel_v16i8:
1117 ; SSE41: # %bb.0:
1118 ; SSE41-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
1119 ; SSE41-NEXT: movdqa %xmm0, %xmm2
1120 ; SSE41-NEXT: punpckhbw {{.*#+}} xmm2 = xmm2[8],xmm0[8],xmm2[9],xmm0[9],xmm2[10],xmm0[10],xmm2[11],xmm0[11],xmm2[12],xmm0[12],xmm2[13],xmm0[13],xmm2[14],xmm0[14],xmm2[15],xmm0[15]
1121 ; SSE41-NEXT: psrlw %xmm1, %xmm2
1122 ; SSE41-NEXT: pmovzxbw {{.*#+}} xmm3 = [255,255,255,255,255,255,255,255]
1123 ; SSE41-NEXT: pand %xmm3, %xmm2
1124 ; SSE41-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
1125 ; SSE41-NEXT: psrlw %xmm1, %xmm0
1126 ; SSE41-NEXT: pand %xmm3, %xmm0
1127 ; SSE41-NEXT: packuswb %xmm2, %xmm0
1128 ; SSE41-NEXT: retq
1129 ;
1130 ; AVX1-LABEL: splatvar_funnnel_v16i8:
1131 ; AVX1: # %bb.0:
1132 ; AVX1-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1, %xmm1
1133 ; AVX1-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm0[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
1134 ; AVX1-NEXT: vpsrlw %xmm1, %xmm2, %xmm2
1135 ; AVX1-NEXT: vbroadcastss {{.*#+}} xmm3 = [255,255,255,255,255,255,255,255]
1136 ; AVX1-NEXT: vpand %xmm3, %xmm2, %xmm2
1137 ; AVX1-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
1138 ; AVX1-NEXT: vpsrlw %xmm1, %xmm0, %xmm0
1139 ; AVX1-NEXT: vpand %xmm3, %xmm0, %xmm0
1140 ; AVX1-NEXT: vpackuswb %xmm2, %xmm0, %xmm0
1141 ; AVX1-NEXT: retq
1142 ;
1143 ; AVX2-LABEL: splatvar_funnnel_v16i8:
1144 ; AVX2: # %bb.0:
1145 ; AVX2-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1, %xmm1
1146 ; AVX2-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm0[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
1147 ; AVX2-NEXT: vpsrlw %xmm1, %xmm2, %xmm2
1148 ; AVX2-NEXT: vpbroadcastw {{.*#+}} xmm3 = [255,255,255,255,255,255,255,255]
1149 ; AVX2-NEXT: vpand %xmm3, %xmm2, %xmm2
1150 ; AVX2-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
1151 ; AVX2-NEXT: vpsrlw %xmm1, %xmm0, %xmm0
1152 ; AVX2-NEXT: vpand %xmm3, %xmm0, %xmm0
1153 ; AVX2-NEXT: vpackuswb %xmm2, %xmm0, %xmm0
1154 ; AVX2-NEXT: retq
1155 ;
1156 ; AVX512F-LABEL: splatvar_funnnel_v16i8:
1157 ; AVX512F: # %bb.0:
1158 ; AVX512F-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1, %xmm1
1159 ; AVX512F-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm0[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
1160 ; AVX512F-NEXT: vpsrlw %xmm1, %xmm2, %xmm2
1161 ; AVX512F-NEXT: vpbroadcastw {{.*#+}} xmm3 = [255,255,255,255,255,255,255,255]
1162 ; AVX512F-NEXT: vpand %xmm3, %xmm2, %xmm2
1163 ; AVX512F-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
1164 ; AVX512F-NEXT: vpsrlw %xmm1, %xmm0, %xmm0
1165 ; AVX512F-NEXT: vpand %xmm3, %xmm0, %xmm0
1166 ; AVX512F-NEXT: vpackuswb %xmm2, %xmm0, %xmm0
1167 ; AVX512F-NEXT: retq
1168 ;
1169 ; AVX512VL-LABEL: splatvar_funnnel_v16i8:
1170 ; AVX512VL: # %bb.0:
1171 ; AVX512VL-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1, %xmm1
1172 ; AVX512VL-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm0[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
1173 ; AVX512VL-NEXT: vpsrlw %xmm1, %xmm2, %xmm2
1174 ; AVX512VL-NEXT: vpbroadcastd {{.*#+}} xmm3 = [255,255,255,255,255,255,255,255]
1175 ; AVX512VL-NEXT: vpand %xmm3, %xmm2, %xmm2
1176 ; AVX512VL-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
1177 ; AVX512VL-NEXT: vpsrlw %xmm1, %xmm0, %xmm0
1178 ; AVX512VL-NEXT: vpand %xmm3, %xmm0, %xmm0
1179 ; AVX512VL-NEXT: vpackuswb %xmm2, %xmm0, %xmm0
1180 ; AVX512VL-NEXT: retq
1181 ;
1182 ; AVX512BW-LABEL: splatvar_funnnel_v16i8:
1183 ; AVX512BW: # %bb.0:
1184 ; AVX512BW-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1, %xmm1
1185 ; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm0[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
1186 ; AVX512BW-NEXT: vpsrlw %xmm1, %xmm2, %xmm2
1187 ; AVX512BW-NEXT: vpbroadcastw {{.*#+}} xmm3 = [255,255,255,255,255,255,255,255]
1188 ; AVX512BW-NEXT: vpand %xmm3, %xmm2, %xmm2
1189 ; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
1190 ; AVX512BW-NEXT: vpsrlw %xmm1, %xmm0, %xmm0
1191 ; AVX512BW-NEXT: vpand %xmm3, %xmm0, %xmm0
1192 ; AVX512BW-NEXT: vpackuswb %xmm2, %xmm0, %xmm0
1193 ; AVX512BW-NEXT: retq
1194 ;
1195 ; AVX512VLBW-LABEL: splatvar_funnnel_v16i8:
1196 ; AVX512VLBW: # %bb.0:
1197 ; AVX512VLBW-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
1198 ; AVX512VLBW-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm0[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
1199 ; AVX512VLBW-NEXT: vinserti128 $1, %xmm0, %ymm2, %ymm0
1200 ; AVX512VLBW-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1, %xmm1
1201 ; AVX512VLBW-NEXT: vpsrlw %xmm1, %ymm0, %ymm0
1202 ; AVX512VLBW-NEXT: vpmovwb %ymm0, %xmm0
1203 ; AVX512VLBW-NEXT: vzeroupper
1204 ; AVX512VLBW-NEXT: retq
1205 ;
1206 ; AVX512VBMI2-LABEL: splatvar_funnnel_v16i8:
1207 ; AVX512VBMI2: # %bb.0:
1208 ; AVX512VBMI2-NEXT: vmovdqa {{.*#+}} xmm2 = [0,2,4,6,8,10,12,14,64,66,68,70,72,74,76,78]
1209 ; AVX512VBMI2-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1, %xmm1
1210 ; AVX512VBMI2-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm0[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
1211 ; AVX512VBMI2-NEXT: vpsrlw %xmm1, %xmm3, %xmm3
1212 ; AVX512VBMI2-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
1213 ; AVX512VBMI2-NEXT: vpsrlw %xmm1, %xmm0, %xmm0
1214 ; AVX512VBMI2-NEXT: vpermt2b %zmm3, %zmm2, %zmm0
1215 ; AVX512VBMI2-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
1216 ; AVX512VBMI2-NEXT: vzeroupper
1217 ; AVX512VBMI2-NEXT: retq
1218 ;
1219 ; AVX512VLVBMI2-LABEL: splatvar_funnnel_v16i8:
1220 ; AVX512VLVBMI2: # %bb.0:
1221 ; AVX512VLVBMI2-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
1222 ; AVX512VLVBMI2-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm0[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
1223 ; AVX512VLVBMI2-NEXT: vinserti128 $1, %xmm0, %ymm2, %ymm0
1224 ; AVX512VLVBMI2-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1, %xmm1
1225 ; AVX512VLVBMI2-NEXT: vpsrlw %xmm1, %ymm0, %ymm0
1226 ; AVX512VLVBMI2-NEXT: vpmovwb %ymm0, %xmm0
1227 ; AVX512VLVBMI2-NEXT: vzeroupper
1228 ; AVX512VLVBMI2-NEXT: retq
1229 ;
1230 ; XOPAVX1-LABEL: splatvar_funnnel_v16i8:
1231 ; XOPAVX1: # %bb.0:
1232 ; XOPAVX1-NEXT: vpxor %xmm2, %xmm2, %xmm2
1233 ; XOPAVX1-NEXT: vpshufb %xmm2, %xmm1, %xmm1
1234 ; XOPAVX1-NEXT: vpsubb %xmm1, %xmm2, %xmm1
1235 ; XOPAVX1-NEXT: vprotb %xmm1, %xmm0, %xmm0
1236 ; XOPAVX1-NEXT: retq
1237 ;
1238 ; XOPAVX2-LABEL: splatvar_funnnel_v16i8:
1239 ; XOPAVX2: # %bb.0:
1240 ; XOPAVX2-NEXT: vpbroadcastb %xmm1, %xmm1
1241 ; XOPAVX2-NEXT: vpxor %xmm2, %xmm2, %xmm2
1242 ; XOPAVX2-NEXT: vpsubb %xmm1, %xmm2, %xmm1
1243 ; XOPAVX2-NEXT: vprotb %xmm1, %xmm0, %xmm0
1244 ; XOPAVX2-NEXT: retq
1245 ;
1246 ; X86-SSE2-LABEL: splatvar_funnnel_v16i8:
1247 ; X86-SSE2: # %bb.0:
1248 ; X86-SSE2-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}, %xmm1
1249 ; X86-SSE2-NEXT: movdqa %xmm0, %xmm2
1250 ; X86-SSE2-NEXT: punpckhbw {{.*#+}} xmm2 = xmm2[8],xmm0[8],xmm2[9],xmm0[9],xmm2[10],xmm0[10],xmm2[11],xmm0[11],xmm2[12],xmm0[12],xmm2[13],xmm0[13],xmm2[14],xmm0[14],xmm2[15],xmm0[15]
1251 ; X86-SSE2-NEXT: psrlw %xmm1, %xmm2
1252 ; X86-SSE2-NEXT: movdqa {{.*#+}} xmm3 = [255,255,255,255,255,255,255,255]
1253 ; X86-SSE2-NEXT: pand %xmm3, %xmm2
1254 ; X86-SSE2-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
1255 ; X86-SSE2-NEXT: psrlw %xmm1, %xmm0
1256 ; X86-SSE2-NEXT: pand %xmm3, %xmm0
1257 ; X86-SSE2-NEXT: packuswb %xmm2, %xmm0
1258 ; X86-SSE2-NEXT: retl
1259 %splat = shufflevector <16 x i8> %amt, <16 x i8> undef, <16 x i32> zeroinitializer
1260 %res = call <16 x i8> @llvm.fshr.v16i8(<16 x i8> %x, <16 x i8> %x, <16 x i8> %splat)
1261 ret <16 x i8> %res
1262 }
1264 ;
1265 ; Constant Shifts
1266 ;
1268 define <2 x i64> @constant_funnnel_v2i64(<2 x i64> %x) nounwind {
1269 ; SSE2-LABEL: constant_funnnel_v2i64:
1270 ; SSE2: # %bb.0:
1271 ; SSE2-NEXT: movdqa %xmm0, %xmm1
1272 ; SSE2-NEXT: psllq $60, %xmm1
1273 ; SSE2-NEXT: movdqa %xmm0, %xmm2
1274 ; SSE2-NEXT: psllq $50, %xmm2
1275 ; SSE2-NEXT: movsd {{.*#+}} xmm2 = xmm1[0],xmm2[1]
1276 ; SSE2-NEXT: movdqa %xmm0, %xmm1
1277 ; SSE2-NEXT: psrlq $4, %xmm1
1278 ; SSE2-NEXT: psrlq $14, %xmm0
1279 ; SSE2-NEXT: movsd {{.*#+}} xmm0 = xmm1[0],xmm0[1]
1280 ; SSE2-NEXT: orpd %xmm2, %xmm0
1281 ; SSE2-NEXT: retq
1282 ;
1283 ; SSE41-LABEL: constant_funnnel_v2i64:
1284 ; SSE41: # %bb.0:
1285 ; SSE41-NEXT: movdqa %xmm0, %xmm1
1286 ; SSE41-NEXT: psllq $50, %xmm1
1287 ; SSE41-NEXT: movdqa %xmm0, %xmm2
1288 ; SSE41-NEXT: psllq $60, %xmm2
1289 ; SSE41-NEXT: pblendw {{.*#+}} xmm2 = xmm2[0,1,2,3],xmm1[4,5,6,7]
1290 ; SSE41-NEXT: movdqa %xmm0, %xmm1
1291 ; SSE41-NEXT: psrlq $14, %xmm1
1292 ; SSE41-NEXT: psrlq $4, %xmm0
1293 ; SSE41-NEXT: pblendw {{.*#+}} xmm0 = xmm0[0,1,2,3],xmm1[4,5,6,7]
1294 ; SSE41-NEXT: por %xmm2, %xmm0
1295 ; SSE41-NEXT: retq
1296 ;
1297 ; AVX1-LABEL: constant_funnnel_v2i64:
1298 ; AVX1: # %bb.0:
1299 ; AVX1-NEXT: vpsllq $50, %xmm0, %xmm1
1300 ; AVX1-NEXT: vpsllq $60, %xmm0, %xmm2
1301 ; AVX1-NEXT: vpblendw {{.*#+}} xmm1 = xmm2[0,1,2,3],xmm1[4,5,6,7]
1302 ; AVX1-NEXT: vpsrlq $14, %xmm0, %xmm2
1303 ; AVX1-NEXT: vpsrlq $4, %xmm0, %xmm0
1304 ; AVX1-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0,1,2,3],xmm2[4,5,6,7]
1305 ; AVX1-NEXT: vpor %xmm1, %xmm0, %xmm0
1306 ; AVX1-NEXT: retq
1307 ;
1308 ; AVX2-LABEL: constant_funnnel_v2i64:
1309 ; AVX2: # %bb.0:
1310 ; AVX2-NEXT: vpsllvq {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm1
1311 ; AVX2-NEXT: vpsrlvq {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
1312 ; AVX2-NEXT: vpor %xmm1, %xmm0, %xmm0
1313 ; AVX2-NEXT: retq
1314 ;
1315 ; AVX512F-LABEL: constant_funnnel_v2i64:
1316 ; AVX512F: # %bb.0:
1317 ; AVX512F-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
1318 ; AVX512F-NEXT: vpmovsxbq {{.*#+}} xmm1 = [4,14]
1319 ; AVX512F-NEXT: vprorvq %zmm1, %zmm0, %zmm0
1320 ; AVX512F-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
1321 ; AVX512F-NEXT: vzeroupper
1322 ; AVX512F-NEXT: retq
1323 ;
1324 ; AVX512VL-LABEL: constant_funnnel_v2i64:
1325 ; AVX512VL: # %bb.0:
1326 ; AVX512VL-NEXT: vprorvq {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
1327 ; AVX512VL-NEXT: retq
1328 ;
1329 ; AVX512BW-LABEL: constant_funnnel_v2i64:
1330 ; AVX512BW: # %bb.0:
1331 ; AVX512BW-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
1332 ; AVX512BW-NEXT: vpmovsxbq {{.*#+}} xmm1 = [4,14]
1333 ; AVX512BW-NEXT: vprorvq %zmm1, %zmm0, %zmm0
1334 ; AVX512BW-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
1335 ; AVX512BW-NEXT: vzeroupper
1336 ; AVX512BW-NEXT: retq
1337 ;
1338 ; AVX512VLBW-LABEL: constant_funnnel_v2i64:
1339 ; AVX512VLBW: # %bb.0:
1340 ; AVX512VLBW-NEXT: vprorvq {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
1341 ; AVX512VLBW-NEXT: retq
1342 ;
1343 ; AVX512VBMI2-LABEL: constant_funnnel_v2i64:
1344 ; AVX512VBMI2: # %bb.0:
1345 ; AVX512VBMI2-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
1346 ; AVX512VBMI2-NEXT: vpmovsxbq {{.*#+}} xmm1 = [4,14]
1347 ; AVX512VBMI2-NEXT: vprorvq %zmm1, %zmm0, %zmm0
1348 ; AVX512VBMI2-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
1349 ; AVX512VBMI2-NEXT: vzeroupper
1350 ; AVX512VBMI2-NEXT: retq
1351 ;
1352 ; AVX512VLVBMI2-LABEL: constant_funnnel_v2i64:
1353 ; AVX512VLVBMI2: # %bb.0:
1354 ; AVX512VLVBMI2-NEXT: vprorvq {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
1355 ; AVX512VLVBMI2-NEXT: retq
1356 ;
1357 ; XOP-LABEL: constant_funnnel_v2i64:
1358 ; XOP: # %bb.0:
1359 ; XOP-NEXT: vprotq {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
1360 ; XOP-NEXT: retq
1361 ;
1362 ; X86-SSE2-LABEL: constant_funnnel_v2i64:
1363 ; X86-SSE2: # %bb.0:
1364 ; X86-SSE2-NEXT: movdqa %xmm0, %xmm1
1365 ; X86-SSE2-NEXT: psllq $60, %xmm1
1366 ; X86-SSE2-NEXT: movdqa %xmm0, %xmm2
1367 ; X86-SSE2-NEXT: psllq $50, %xmm2
1368 ; X86-SSE2-NEXT: movsd {{.*#+}} xmm2 = xmm1[0],xmm2[1]
1369 ; X86-SSE2-NEXT: movdqa %xmm0, %xmm1
1370 ; X86-SSE2-NEXT: psrlq $4, %xmm1
1371 ; X86-SSE2-NEXT: psrlq $14, %xmm0
1372 ; X86-SSE2-NEXT: movsd {{.*#+}} xmm0 = xmm1[0],xmm0[1]
1373 ; X86-SSE2-NEXT: orpd %xmm2, %xmm0
1374 ; X86-SSE2-NEXT: retl
1375 %res = call <2 x i64> @llvm.fshr.v2i64(<2 x i64> %x, <2 x i64> %x, <2 x i64> <i64 4, i64 14>)
1376 ret <2 x i64> %res
1377 }
1379 define <4 x i32> @constant_funnnel_v4i32(<4 x i32> %x) nounwind {
1380 ; SSE2-LABEL: constant_funnnel_v4i32:
1381 ; SSE2: # %bb.0:
1382 ; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,3,3]
1383 ; SSE2-NEXT: pmuludq {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
1384 ; SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm0[1,3,2,3]
1385 ; SSE2-NEXT: pmuludq {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
1386 ; SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm1[1,3,2,3]
1387 ; SSE2-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1]
1388 ; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
1389 ; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,2,2,3]
1390 ; SSE2-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
1391 ; SSE2-NEXT: por %xmm2, %xmm0
1392 ; SSE2-NEXT: retq
1393 ;
1394 ; SSE41-LABEL: constant_funnnel_v4i32:
1395 ; SSE41: # %bb.0:
1396 ; SSE41-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,3,3]
1397 ; SSE41-NEXT: pmuludq {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
1398 ; SSE41-NEXT: pmuludq {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
1399 ; SSE41-NEXT: pshufd {{.*#+}} xmm2 = xmm0[1,1,3,3]
1400 ; SSE41-NEXT: pblendw {{.*#+}} xmm2 = xmm2[0,1],xmm1[2,3],xmm2[4,5],xmm1[6,7]
1401 ; SSE41-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,0,2,2]
1402 ; SSE41-NEXT: pblendw {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3],xmm0[4,5],xmm1[6,7]
1403 ; SSE41-NEXT: por %xmm2, %xmm0
1404 ; SSE41-NEXT: retq
1405 ;
1406 ; AVX1-LABEL: constant_funnnel_v4i32:
1407 ; AVX1: # %bb.0:
1408 ; AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,3,3]
1409 ; AVX1-NEXT: vpmuludq {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1, %xmm1
1410 ; AVX1-NEXT: vpmuludq {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
1411 ; AVX1-NEXT: vpshufd {{.*#+}} xmm2 = xmm0[1,1,3,3]
1412 ; AVX1-NEXT: vpblendw {{.*#+}} xmm2 = xmm2[0,1],xmm1[2,3],xmm2[4,5],xmm1[6,7]
1413 ; AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[0,0,2,2]
1414 ; AVX1-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3],xmm0[4,5],xmm1[6,7]
1415 ; AVX1-NEXT: vpor %xmm2, %xmm0, %xmm0
1416 ; AVX1-NEXT: retq
1417 ;
1418 ; AVX2-LABEL: constant_funnnel_v4i32:
1419 ; AVX2: # %bb.0:
1420 ; AVX2-NEXT: vpsrlvd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm1
1421 ; AVX2-NEXT: vpsllvd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
1422 ; AVX2-NEXT: vpor %xmm1, %xmm0, %xmm0
1423 ; AVX2-NEXT: retq
1424 ;
1425 ; AVX512F-LABEL: constant_funnnel_v4i32:
1426 ; AVX512F: # %bb.0:
1427 ; AVX512F-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
1428 ; AVX512F-NEXT: vpmovsxbd {{.*#+}} xmm1 = [4,5,6,7]
1429 ; AVX512F-NEXT: vprorvd %zmm1, %zmm0, %zmm0
1430 ; AVX512F-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
1431 ; AVX512F-NEXT: vzeroupper
1432 ; AVX512F-NEXT: retq
1433 ;
1434 ; AVX512VL-LABEL: constant_funnnel_v4i32:
1435 ; AVX512VL: # %bb.0:
1436 ; AVX512VL-NEXT: vprorvd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
1437 ; AVX512VL-NEXT: retq
1438 ;
1439 ; AVX512BW-LABEL: constant_funnnel_v4i32:
1440 ; AVX512BW: # %bb.0:
1441 ; AVX512BW-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
1442 ; AVX512BW-NEXT: vpmovsxbd {{.*#+}} xmm1 = [4,5,6,7]
1443 ; AVX512BW-NEXT: vprorvd %zmm1, %zmm0, %zmm0
1444 ; AVX512BW-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
1445 ; AVX512BW-NEXT: vzeroupper
1446 ; AVX512BW-NEXT: retq
1447 ;
1448 ; AVX512VLBW-LABEL: constant_funnnel_v4i32:
1449 ; AVX512VLBW: # %bb.0:
1450 ; AVX512VLBW-NEXT: vprorvd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
1451 ; AVX512VLBW-NEXT: retq
1452 ;
1453 ; AVX512VBMI2-LABEL: constant_funnnel_v4i32:
1454 ; AVX512VBMI2: # %bb.0:
1455 ; AVX512VBMI2-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
1456 ; AVX512VBMI2-NEXT: vpmovsxbd {{.*#+}} xmm1 = [4,5,6,7]
1457 ; AVX512VBMI2-NEXT: vprorvd %zmm1, %zmm0, %zmm0
1458 ; AVX512VBMI2-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
1459 ; AVX512VBMI2-NEXT: vzeroupper
1460 ; AVX512VBMI2-NEXT: retq
1461 ;
1462 ; AVX512VLVBMI2-LABEL: constant_funnnel_v4i32:
1463 ; AVX512VLVBMI2: # %bb.0:
1464 ; AVX512VLVBMI2-NEXT: vprorvd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
1465 ; AVX512VLVBMI2-NEXT: retq
1466 ;
1467 ; XOP-LABEL: constant_funnnel_v4i32:
1468 ; XOP: # %bb.0:
1469 ; XOP-NEXT: vprotd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
1470 ; XOP-NEXT: retq
1471 ;
1472 ; X86-SSE2-LABEL: constant_funnnel_v4i32:
1473 ; X86-SSE2: # %bb.0:
1474 ; X86-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,3,3]
1475 ; X86-SSE2-NEXT: pmuludq {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0
1476 ; X86-SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm0[1,3,2,3]
1477 ; X86-SSE2-NEXT: pmuludq {{\.?LCPI[0-9]+_[0-9]+}}, %xmm1
1478 ; X86-SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm1[1,3,2,3]
1479 ; X86-SSE2-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1]
1480 ; X86-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
1481 ; X86-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,2,2,3]
1482 ; X86-SSE2-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
1483 ; X86-SSE2-NEXT: por %xmm2, %xmm0
1484 ; X86-SSE2-NEXT: retl
1485 %res = call <4 x i32> @llvm.fshr.v4i32(<4 x i32> %x, <4 x i32> %x, <4 x i32> <i32 4, i32 5, i32 6, i32 7>)
1486 ret <4 x i32> %res
1487 }
1489 define <8 x i16> @constant_funnnel_v8i16(<8 x i16> %x) nounwind {
1490 ; SSE-LABEL: constant_funnnel_v8i16:
1491 ; SSE: # %bb.0:
1492 ; SSE-NEXT: movdqa {{.*#+}} xmm1 = [1,32768,16384,8192,4096,2048,1024,512]
1493 ; SSE-NEXT: movdqa %xmm0, %xmm2
1494 ; SSE-NEXT: pmulhuw %xmm1, %xmm2
1495 ; SSE-NEXT: pmullw %xmm1, %xmm0
1496 ; SSE-NEXT: por %xmm2, %xmm0
1497 ; SSE-NEXT: retq
1498 ;
1499 ; AVX-LABEL: constant_funnnel_v8i16:
1500 ; AVX: # %bb.0:
1501 ; AVX-NEXT: vmovdqa {{.*#+}} xmm1 = [1,32768,16384,8192,4096,2048,1024,512]
1502 ; AVX-NEXT: vpmulhuw %xmm1, %xmm0, %xmm2
1503 ; AVX-NEXT: vpmullw %xmm1, %xmm0, %xmm0
1504 ; AVX-NEXT: vpor %xmm2, %xmm0, %xmm0
1505 ; AVX-NEXT: retq
1506 ;
1507 ; AVX512F-LABEL: constant_funnnel_v8i16:
1508 ; AVX512F: # %bb.0:
1509 ; AVX512F-NEXT: vmovdqa {{.*#+}} xmm1 = [1,32768,16384,8192,4096,2048,1024,512]
1510 ; AVX512F-NEXT: vpmulhuw %xmm1, %xmm0, %xmm2
1511 ; AVX512F-NEXT: vpmullw %xmm1, %xmm0, %xmm0
1512 ; AVX512F-NEXT: vpor %xmm2, %xmm0, %xmm0
1513 ; AVX512F-NEXT: retq
1514 ;
1515 ; AVX512VL-LABEL: constant_funnnel_v8i16:
1516 ; AVX512VL: # %bb.0:
1517 ; AVX512VL-NEXT: vmovdqa {{.*#+}} xmm1 = [1,32768,16384,8192,4096,2048,1024,512]
1518 ; AVX512VL-NEXT: vpmulhuw %xmm1, %xmm0, %xmm2
1519 ; AVX512VL-NEXT: vpmullw %xmm1, %xmm0, %xmm0
1520 ; AVX512VL-NEXT: vpor %xmm2, %xmm0, %xmm0
1521 ; AVX512VL-NEXT: retq
1522 ;
1523 ; AVX512BW-LABEL: constant_funnnel_v8i16:
1524 ; AVX512BW: # %bb.0:
1525 ; AVX512BW-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
1526 ; AVX512BW-NEXT: vpmovsxbw {{.*#+}} xmm1 = [16,1,2,3,4,5,6,7]
1527 ; AVX512BW-NEXT: vpsrlvw %zmm1, %zmm0, %zmm1
1528 ; AVX512BW-NEXT: vpmovsxbw {{.*#+}} xmm2 = [0,15,14,13,12,11,10,9]
1529 ; AVX512BW-NEXT: vpsllvw %zmm2, %zmm0, %zmm0
1530 ; AVX512BW-NEXT: vpor %xmm1, %xmm0, %xmm0
1531 ; AVX512BW-NEXT: vzeroupper
1532 ; AVX512BW-NEXT: retq
1533 ;
1534 ; AVX512VLBW-LABEL: constant_funnnel_v8i16:
1535 ; AVX512VLBW: # %bb.0:
1536 ; AVX512VLBW-NEXT: vpsrlvw {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm1
1537 ; AVX512VLBW-NEXT: vpsllvw {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
1538 ; AVX512VLBW-NEXT: vpor %xmm1, %xmm0, %xmm0
1539 ; AVX512VLBW-NEXT: retq
1540 ;
1541 ; AVX512VBMI2-LABEL: constant_funnnel_v8i16:
1542 ; AVX512VBMI2: # %bb.0:
1543 ; AVX512VBMI2-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
1544 ; AVX512VBMI2-NEXT: vpmovsxbw {{.*#+}} xmm1 = [0,1,2,3,4,5,6,7]
1545 ; AVX512VBMI2-NEXT: vpshrdvw %zmm1, %zmm0, %zmm0
1546 ; AVX512VBMI2-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
1547 ; AVX512VBMI2-NEXT: vzeroupper
1548 ; AVX512VBMI2-NEXT: retq
1549 ;
1550 ; AVX512VLVBMI2-LABEL: constant_funnnel_v8i16:
1551 ; AVX512VLVBMI2: # %bb.0:
1552 ; AVX512VLVBMI2-NEXT: vpshrdvw {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
1553 ; AVX512VLVBMI2-NEXT: retq
1554 ;
1555 ; XOP-LABEL: constant_funnnel_v8i16:
1556 ; XOP: # %bb.0:
1557 ; XOP-NEXT: vprotw {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
1558 ; XOP-NEXT: retq
1559 ;
1560 ; X86-SSE2-LABEL: constant_funnnel_v8i16:
1561 ; X86-SSE2: # %bb.0:
1562 ; X86-SSE2-NEXT: movdqa {{.*#+}} xmm1 = [1,32768,16384,8192,4096,2048,1024,512]
1563 ; X86-SSE2-NEXT: movdqa %xmm0, %xmm2
1564 ; X86-SSE2-NEXT: pmulhuw %xmm1, %xmm2
1565 ; X86-SSE2-NEXT: pmullw %xmm1, %xmm0
1566 ; X86-SSE2-NEXT: por %xmm2, %xmm0
1567 ; X86-SSE2-NEXT: retl
1568 %res = call <8 x i16> @llvm.fshr.v8i16(<8 x i16> %x, <8 x i16> %x, <8 x i16> <i16 0, i16 1, i16 2, i16 3, i16 4, i16 5, i16 6, i16 7>)
1569 ret <8 x i16> %res
1570 }
1572 define <16 x i8> @constant_funnnel_v16i8(<16 x i8> %x) nounwind {
1573 ; SSE-LABEL: constant_funnnel_v16i8:
1574 ; SSE: # %bb.0:
1575 ; SSE-NEXT: movdqa %xmm0, %xmm1
1576 ; SSE-NEXT: punpckhbw {{.*#+}} xmm1 = xmm1[8],xmm0[8],xmm1[9],xmm0[9],xmm1[10],xmm0[10],xmm1[11],xmm0[11],xmm1[12],xmm0[12],xmm1[13],xmm0[13],xmm1[14],xmm0[14],xmm1[15],xmm0[15]
1577 ; SSE-NEXT: pmullw {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1 # [1,2,4,8,16,32,64,128]
1578 ; SSE-NEXT: psrlw $8, %xmm1
1579 ; SSE-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
1580 ; SSE-NEXT: pmullw {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0 # [1,128,64,32,16,8,4,2]
1581 ; SSE-NEXT: psrlw $8, %xmm0
1582 ; SSE-NEXT: packuswb %xmm1, %xmm0
1583 ; SSE-NEXT: retq
1584 ;
1585 ; AVX-LABEL: constant_funnnel_v16i8:
1586 ; AVX: # %bb.0:
1587 ; AVX-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm0[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
1588 ; AVX-NEXT: vpmullw {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1, %xmm1 # [1,2,4,8,16,32,64,128]
1589 ; AVX-NEXT: vpsrlw $8, %xmm1, %xmm1
1590 ; AVX-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
1591 ; AVX-NEXT: vpmullw {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0 # [1,128,64,32,16,8,4,2]
1592 ; AVX-NEXT: vpsrlw $8, %xmm0, %xmm0
1593 ; AVX-NEXT: vpackuswb %xmm1, %xmm0, %xmm0
1594 ; AVX-NEXT: retq
1595 ;
1596 ; AVX512F-LABEL: constant_funnnel_v16i8:
1597 ; AVX512F: # %bb.0:
1598 ; AVX512F-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm0[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
1599 ; AVX512F-NEXT: vpmullw {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1, %xmm1 # [1,2,4,8,16,32,64,128]
1600 ; AVX512F-NEXT: vpsrlw $8, %xmm1, %xmm1
1601 ; AVX512F-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
1602 ; AVX512F-NEXT: vpmullw {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0 # [1,128,64,32,16,8,4,2]
1603 ; AVX512F-NEXT: vpsrlw $8, %xmm0, %xmm0
1604 ; AVX512F-NEXT: vpackuswb %xmm1, %xmm0, %xmm0
1605 ; AVX512F-NEXT: retq
1606 ;
1607 ; AVX512VL-LABEL: constant_funnnel_v16i8:
1608 ; AVX512VL: # %bb.0:
1609 ; AVX512VL-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm0[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
1610 ; AVX512VL-NEXT: vpmullw {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1, %xmm1 # [1,2,4,8,16,32,64,128]
1611 ; AVX512VL-NEXT: vpsrlw $8, %xmm1, %xmm1
1612 ; AVX512VL-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
1613 ; AVX512VL-NEXT: vpmullw {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0 # [1,128,64,32,16,8,4,2]
1614 ; AVX512VL-NEXT: vpsrlw $8, %xmm0, %xmm0
1615 ; AVX512VL-NEXT: vpackuswb %xmm1, %xmm0, %xmm0
1616 ; AVX512VL-NEXT: retq
1617 ;
1618 ; AVX512BW-LABEL: constant_funnnel_v16i8:
1619 ; AVX512BW: # %bb.0:
1620 ; AVX512BW-NEXT: vpmovsxbw {{.*#+}} xmm1 = [0,1,2,3,4,5,6,7]
1621 ; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm0[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
1622 ; AVX512BW-NEXT: vpsllvw %zmm1, %zmm2, %zmm1
1623 ; AVX512BW-NEXT: vpsrlw $8, %xmm1, %xmm1
1624 ; AVX512BW-NEXT: vpmovsxbw {{.*#+}} xmm2 = [0,7,6,5,4,3,2,1]
1625 ; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
1626 ; AVX512BW-NEXT: vpsllvw %zmm2, %zmm0, %zmm0
1627 ; AVX512BW-NEXT: vpsrlw $8, %xmm0, %xmm0
1628 ; AVX512BW-NEXT: vpackuswb %xmm1, %xmm0, %xmm0
1629 ; AVX512BW-NEXT: vzeroupper
1630 ; AVX512BW-NEXT: retq
1631 ;
1632 ; AVX512VLBW-LABEL: constant_funnnel_v16i8:
1633 ; AVX512VLBW: # %bb.0:
1634 ; AVX512VLBW-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm0[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
1635 ; AVX512VLBW-NEXT: vpsllvw {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1, %xmm1
1636 ; AVX512VLBW-NEXT: vpsrlw $8, %xmm1, %xmm1
1637 ; AVX512VLBW-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
1638 ; AVX512VLBW-NEXT: vpsllvw {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
1639 ; AVX512VLBW-NEXT: vpsrlw $8, %xmm0, %xmm0
1640 ; AVX512VLBW-NEXT: vpackuswb %xmm1, %xmm0, %xmm0
1641 ; AVX512VLBW-NEXT: retq
1642 ;
1643 ; AVX512VBMI2-LABEL: constant_funnnel_v16i8:
1644 ; AVX512VBMI2: # %bb.0:
1645 ; AVX512VBMI2-NEXT: vpmovsxbw {{.*#+}} xmm1 = [0,1,2,3,4,5,6,7]
1646 ; AVX512VBMI2-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm0[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
1647 ; AVX512VBMI2-NEXT: vpsllvw %zmm1, %zmm2, %zmm1
1648 ; AVX512VBMI2-NEXT: vpsrlw $8, %xmm1, %xmm1
1649 ; AVX512VBMI2-NEXT: vpmovsxbw {{.*#+}} xmm2 = [0,7,6,5,4,3,2,1]
1650 ; AVX512VBMI2-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
1651 ; AVX512VBMI2-NEXT: vpsllvw %zmm2, %zmm0, %zmm0
1652 ; AVX512VBMI2-NEXT: vpsrlw $8, %xmm0, %xmm0
1653 ; AVX512VBMI2-NEXT: vpackuswb %xmm1, %xmm0, %xmm0
1654 ; AVX512VBMI2-NEXT: vzeroupper
1655 ; AVX512VBMI2-NEXT: retq
1656 ;
1657 ; AVX512VLVBMI2-LABEL: constant_funnnel_v16i8:
1658 ; AVX512VLVBMI2: # %bb.0:
1659 ; AVX512VLVBMI2-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm0[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
1660 ; AVX512VLVBMI2-NEXT: vpsllvw {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1, %xmm1
1661 ; AVX512VLVBMI2-NEXT: vpsrlw $8, %xmm1, %xmm1
1662 ; AVX512VLVBMI2-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
1663 ; AVX512VLVBMI2-NEXT: vpsllvw {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
1664 ; AVX512VLVBMI2-NEXT: vpsrlw $8, %xmm0, %xmm0
1665 ; AVX512VLVBMI2-NEXT: vpackuswb %xmm1, %xmm0, %xmm0
1666 ; AVX512VLVBMI2-NEXT: retq
1667 ;
1668 ; XOP-LABEL: constant_funnnel_v16i8:
1669 ; XOP: # %bb.0:
1670 ; XOP-NEXT: vprotb {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
1671 ; XOP-NEXT: retq
1672 ;
1673 ; X86-SSE2-LABEL: constant_funnnel_v16i8:
1674 ; X86-SSE2: # %bb.0:
1675 ; X86-SSE2-NEXT: movdqa %xmm0, %xmm1
1676 ; X86-SSE2-NEXT: punpckhbw {{.*#+}} xmm1 = xmm1[8],xmm0[8],xmm1[9],xmm0[9],xmm1[10],xmm0[10],xmm1[11],xmm0[11],xmm1[12],xmm0[12],xmm1[13],xmm0[13],xmm1[14],xmm0[14],xmm1[15],xmm0[15]
1677 ; X86-SSE2-NEXT: pmullw {{\.?LCPI[0-9]+_[0-9]+}}, %xmm1 # [1,2,4,8,16,32,64,128]
1678 ; X86-SSE2-NEXT: psrlw $8, %xmm1
1679 ; X86-SSE2-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
1680 ; X86-SSE2-NEXT: pmullw {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0 # [1,128,64,32,16,8,4,2]
1681 ; X86-SSE2-NEXT: psrlw $8, %xmm0
1682 ; X86-SSE2-NEXT: packuswb %xmm1, %xmm0
1683 ; X86-SSE2-NEXT: retl
1684 %res = call <16 x i8> @llvm.fshr.v16i8(<16 x i8> %x, <16 x i8> %x, <16 x i8> <i8 0, i8 1, i8 2, i8 3, i8 4, i8 5, i8 6, i8 7, i8 8, i8 7, i8 6, i8 5, i8 4, i8 3, i8 2, i8 1>)
1685 ret <16 x i8> %res
1686 }
1688 ;
1689 ; Uniform Constant Shifts
1690 ;
1692 define <2 x i64> @splatconstant_funnnel_v2i64(<2 x i64> %x) nounwind {
1693 ; SSE-LABEL: splatconstant_funnnel_v2i64:
1694 ; SSE: # %bb.0:
1695 ; SSE-NEXT: movdqa %xmm0, %xmm1
1696 ; SSE-NEXT: psllq $50, %xmm1
1697 ; SSE-NEXT: psrlq $14, %xmm0
1698 ; SSE-NEXT: por %xmm1, %xmm0
1699 ; SSE-NEXT: retq
1700 ;
1701 ; AVX-LABEL: splatconstant_funnnel_v2i64:
1702 ; AVX: # %bb.0:
1703 ; AVX-NEXT: vpsllq $50, %xmm0, %xmm1
1704 ; AVX-NEXT: vpsrlq $14, %xmm0, %xmm0
1705 ; AVX-NEXT: vpor %xmm1, %xmm0, %xmm0
1706 ; AVX-NEXT: retq
1707 ;
1708 ; AVX512F-LABEL: splatconstant_funnnel_v2i64:
1709 ; AVX512F: # %bb.0:
1710 ; AVX512F-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
1711 ; AVX512F-NEXT: vprorq $14, %zmm0, %zmm0
1712 ; AVX512F-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
1713 ; AVX512F-NEXT: vzeroupper
1714 ; AVX512F-NEXT: retq
1715 ;
1716 ; AVX512VL-LABEL: splatconstant_funnnel_v2i64:
1717 ; AVX512VL: # %bb.0:
1718 ; AVX512VL-NEXT: vprorq $14, %xmm0, %xmm0
1719 ; AVX512VL-NEXT: retq
1720 ;
1721 ; AVX512BW-LABEL: splatconstant_funnnel_v2i64:
1722 ; AVX512BW: # %bb.0:
1723 ; AVX512BW-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
1724 ; AVX512BW-NEXT: vprorq $14, %zmm0, %zmm0
1725 ; AVX512BW-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
1726 ; AVX512BW-NEXT: vzeroupper
1727 ; AVX512BW-NEXT: retq
1728 ;
1729 ; AVX512VLBW-LABEL: splatconstant_funnnel_v2i64:
1730 ; AVX512VLBW: # %bb.0:
1731 ; AVX512VLBW-NEXT: vprorq $14, %xmm0, %xmm0
1732 ; AVX512VLBW-NEXT: retq
1733 ;
1734 ; AVX512VBMI2-LABEL: splatconstant_funnnel_v2i64:
1735 ; AVX512VBMI2: # %bb.0:
1736 ; AVX512VBMI2-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
1737 ; AVX512VBMI2-NEXT: vprorq $14, %zmm0, %zmm0
1738 ; AVX512VBMI2-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
1739 ; AVX512VBMI2-NEXT: vzeroupper
1740 ; AVX512VBMI2-NEXT: retq
1741 ;
1742 ; AVX512VLVBMI2-LABEL: splatconstant_funnnel_v2i64:
1743 ; AVX512VLVBMI2: # %bb.0:
1744 ; AVX512VLVBMI2-NEXT: vprorq $14, %xmm0, %xmm0
1745 ; AVX512VLVBMI2-NEXT: retq
1746 ;
1747 ; XOP-LABEL: splatconstant_funnnel_v2i64:
1748 ; XOP: # %bb.0:
1749 ; XOP-NEXT: vprotq $50, %xmm0, %xmm0
1750 ; XOP-NEXT: retq
1751 ;
1752 ; X86-SSE2-LABEL: splatconstant_funnnel_v2i64:
1753 ; X86-SSE2: # %bb.0:
1754 ; X86-SSE2-NEXT: movdqa %xmm0, %xmm1
1755 ; X86-SSE2-NEXT: psllq $50, %xmm1
1756 ; X86-SSE2-NEXT: psrlq $14, %xmm0
1757 ; X86-SSE2-NEXT: por %xmm1, %xmm0
1758 ; X86-SSE2-NEXT: retl
1759 %res = call <2 x i64> @llvm.fshr.v2i64(<2 x i64> %x, <2 x i64> %x, <2 x i64> <i64 14, i64 14>)
1760 ret <2 x i64> %res
1761 }
1763 define <4 x i32> @splatconstant_funnnel_v4i32(<4 x i32> %x) nounwind {
1764 ; SSE-LABEL: splatconstant_funnnel_v4i32:
1765 ; SSE: # %bb.0:
1766 ; SSE-NEXT: movdqa %xmm0, %xmm1
1767 ; SSE-NEXT: psrld $4, %xmm1
1768 ; SSE-NEXT: pslld $28, %xmm0
1769 ; SSE-NEXT: por %xmm1, %xmm0
1770 ; SSE-NEXT: retq
1771 ;
1772 ; AVX-LABEL: splatconstant_funnnel_v4i32:
1773 ; AVX: # %bb.0:
1774 ; AVX-NEXT: vpsrld $4, %xmm0, %xmm1
1775 ; AVX-NEXT: vpslld $28, %xmm0, %xmm0
1776 ; AVX-NEXT: vpor %xmm1, %xmm0, %xmm0
1777 ; AVX-NEXT: retq
1778 ;
1779 ; AVX512F-LABEL: splatconstant_funnnel_v4i32:
1780 ; AVX512F: # %bb.0:
1781 ; AVX512F-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
1782 ; AVX512F-NEXT: vprord $4, %zmm0, %zmm0
1783 ; AVX512F-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
1784 ; AVX512F-NEXT: vzeroupper
1785 ; AVX512F-NEXT: retq
1786 ;
1787 ; AVX512VL-LABEL: splatconstant_funnnel_v4i32:
1788 ; AVX512VL: # %bb.0:
1789 ; AVX512VL-NEXT: vprord $4, %xmm0, %xmm0
1790 ; AVX512VL-NEXT: retq
1791 ;
1792 ; AVX512BW-LABEL: splatconstant_funnnel_v4i32:
1793 ; AVX512BW: # %bb.0:
1794 ; AVX512BW-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
1795 ; AVX512BW-NEXT: vprord $4, %zmm0, %zmm0
1796 ; AVX512BW-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
1797 ; AVX512BW-NEXT: vzeroupper
1798 ; AVX512BW-NEXT: retq
1799 ;
1800 ; AVX512VLBW-LABEL: splatconstant_funnnel_v4i32:
1801 ; AVX512VLBW: # %bb.0:
1802 ; AVX512VLBW-NEXT: vprord $4, %xmm0, %xmm0
1803 ; AVX512VLBW-NEXT: retq
1804 ;
1805 ; AVX512VBMI2-LABEL: splatconstant_funnnel_v4i32:
1806 ; AVX512VBMI2: # %bb.0:
1807 ; AVX512VBMI2-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
1808 ; AVX512VBMI2-NEXT: vprord $4, %zmm0, %zmm0
1809 ; AVX512VBMI2-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
1810 ; AVX512VBMI2-NEXT: vzeroupper
1811 ; AVX512VBMI2-NEXT: retq
1812 ;
1813 ; AVX512VLVBMI2-LABEL: splatconstant_funnnel_v4i32:
1814 ; AVX512VLVBMI2: # %bb.0:
1815 ; AVX512VLVBMI2-NEXT: vprord $4, %xmm0, %xmm0
1816 ; AVX512VLVBMI2-NEXT: retq
1817 ;
1818 ; XOP-LABEL: splatconstant_funnnel_v4i32:
1819 ; XOP: # %bb.0:
1820 ; XOP-NEXT: vprotd $28, %xmm0, %xmm0
1821 ; XOP-NEXT: retq
1822 ;
1823 ; X86-SSE2-LABEL: splatconstant_funnnel_v4i32:
1824 ; X86-SSE2: # %bb.0:
1825 ; X86-SSE2-NEXT: movdqa %xmm0, %xmm1
1826 ; X86-SSE2-NEXT: psrld $4, %xmm1
1827 ; X86-SSE2-NEXT: pslld $28, %xmm0
1828 ; X86-SSE2-NEXT: por %xmm1, %xmm0
1829 ; X86-SSE2-NEXT: retl
1830 %res = call <4 x i32> @llvm.fshr.v4i32(<4 x i32> %x, <4 x i32> %x, <4 x i32> <i32 4, i32 4, i32 4, i32 4>)
1831 ret <4 x i32> %res
1832 }
1834 define <8 x i16> @splatconstant_funnnel_v8i16(<8 x i16> %x) nounwind {
1835 ; SSE-LABEL: splatconstant_funnnel_v8i16:
1836 ; SSE: # %bb.0:
1837 ; SSE-NEXT: movdqa %xmm0, %xmm1
1838 ; SSE-NEXT: psrlw $7, %xmm1
1839 ; SSE-NEXT: psllw $9, %xmm0
1840 ; SSE-NEXT: por %xmm1, %xmm0
1841 ; SSE-NEXT: retq
1842 ;
1843 ; AVX-LABEL: splatconstant_funnnel_v8i16:
1844 ; AVX: # %bb.0:
1845 ; AVX-NEXT: vpsrlw $7, %xmm0, %xmm1
1846 ; AVX-NEXT: vpsllw $9, %xmm0, %xmm0
1847 ; AVX-NEXT: vpor %xmm1, %xmm0, %xmm0
1848 ; AVX-NEXT: retq
1849 ;
1850 ; AVX512F-LABEL: splatconstant_funnnel_v8i16:
1851 ; AVX512F: # %bb.0:
1852 ; AVX512F-NEXT: vpsrlw $7, %xmm0, %xmm1
1853 ; AVX512F-NEXT: vpsllw $9, %xmm0, %xmm0
1854 ; AVX512F-NEXT: vpor %xmm1, %xmm0, %xmm0
1855 ; AVX512F-NEXT: retq
1856 ;
1857 ; AVX512VL-LABEL: splatconstant_funnnel_v8i16:
1858 ; AVX512VL: # %bb.0:
1859 ; AVX512VL-NEXT: vpsrlw $7, %xmm0, %xmm1
1860 ; AVX512VL-NEXT: vpsllw $9, %xmm0, %xmm0
1861 ; AVX512VL-NEXT: vpor %xmm1, %xmm0, %xmm0
1862 ; AVX512VL-NEXT: retq
1863 ;
1864 ; AVX512BW-LABEL: splatconstant_funnnel_v8i16:
1865 ; AVX512BW: # %bb.0:
1866 ; AVX512BW-NEXT: vpsrlw $7, %xmm0, %xmm1
1867 ; AVX512BW-NEXT: vpsllw $9, %xmm0, %xmm0
1868 ; AVX512BW-NEXT: vpor %xmm1, %xmm0, %xmm0
1869 ; AVX512BW-NEXT: retq
1870 ;
1871 ; AVX512VLBW-LABEL: splatconstant_funnnel_v8i16:
1872 ; AVX512VLBW: # %bb.0:
1873 ; AVX512VLBW-NEXT: vpsrlw $7, %xmm0, %xmm1
1874 ; AVX512VLBW-NEXT: vpsllw $9, %xmm0, %xmm0
1875 ; AVX512VLBW-NEXT: vpor %xmm1, %xmm0, %xmm0
1876 ; AVX512VLBW-NEXT: retq
1877 ;
1878 ; AVX512VBMI2-LABEL: splatconstant_funnnel_v8i16:
1879 ; AVX512VBMI2: # %bb.0:
1880 ; AVX512VBMI2-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
1881 ; AVX512VBMI2-NEXT: vpshrdw $7, %zmm0, %zmm0, %zmm0
1882 ; AVX512VBMI2-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
1883 ; AVX512VBMI2-NEXT: vzeroupper
1884 ; AVX512VBMI2-NEXT: retq
1885 ;
1886 ; AVX512VLVBMI2-LABEL: splatconstant_funnnel_v8i16:
1887 ; AVX512VLVBMI2: # %bb.0:
1888 ; AVX512VLVBMI2-NEXT: vpshrdw $7, %xmm0, %xmm0, %xmm0
1889 ; AVX512VLVBMI2-NEXT: retq
1890 ;
1891 ; XOP-LABEL: splatconstant_funnnel_v8i16:
1892 ; XOP: # %bb.0:
1893 ; XOP-NEXT: vprotw $9, %xmm0, %xmm0
1894 ; XOP-NEXT: retq
1895 ;
1896 ; X86-SSE2-LABEL: splatconstant_funnnel_v8i16:
1897 ; X86-SSE2: # %bb.0:
1898 ; X86-SSE2-NEXT: movdqa %xmm0, %xmm1
1899 ; X86-SSE2-NEXT: psrlw $7, %xmm1
1900 ; X86-SSE2-NEXT: psllw $9, %xmm0
1901 ; X86-SSE2-NEXT: por %xmm1, %xmm0
1902 ; X86-SSE2-NEXT: retl
1903 %res = call <8 x i16> @llvm.fshr.v8i16(<8 x i16> %x, <8 x i16> %x, <8 x i16> <i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7>)
1904 ret <8 x i16> %res
1905 }
1907 define <16 x i8> @splatconstant_funnnel_v16i8(<16 x i8> %x) nounwind {
1908 ; SSE-LABEL: splatconstant_funnnel_v16i8:
1909 ; SSE: # %bb.0:
1910 ; SSE-NEXT: movdqa %xmm0, %xmm1
1911 ; SSE-NEXT: psrlw $4, %xmm1
1912 ; SSE-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
1913 ; SSE-NEXT: psllw $4, %xmm0
1914 ; SSE-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
1915 ; SSE-NEXT: por %xmm1, %xmm0
1916 ; SSE-NEXT: retq
1917 ;
1918 ; AVX-LABEL: splatconstant_funnnel_v16i8:
1919 ; AVX: # %bb.0:
1920 ; AVX-NEXT: vpsrlw $4, %xmm0, %xmm1
1921 ; AVX-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1, %xmm1
1922 ; AVX-NEXT: vpsllw $4, %xmm0, %xmm0
1923 ; AVX-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
1924 ; AVX-NEXT: vpor %xmm1, %xmm0, %xmm0
1925 ; AVX-NEXT: retq
1926 ;
1927 ; AVX512F-LABEL: splatconstant_funnnel_v16i8:
1928 ; AVX512F: # %bb.0:
1929 ; AVX512F-NEXT: vpsllw $4, %xmm0, %xmm1
1930 ; AVX512F-NEXT: vpsrlw $4, %xmm0, %xmm0
1931 ; AVX512F-NEXT: vpternlogd {{.*#+}} zmm0 = zmm0 ^ (mem & (zmm0 ^ zmm1))
1932 ; AVX512F-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
1933 ; AVX512F-NEXT: vzeroupper
1934 ; AVX512F-NEXT: retq
1935 ;
1936 ; AVX512VL-LABEL: splatconstant_funnnel_v16i8:
1937 ; AVX512VL: # %bb.0:
1938 ; AVX512VL-NEXT: vpsllw $4, %xmm0, %xmm1
1939 ; AVX512VL-NEXT: vpsrlw $4, %xmm0, %xmm0
1940 ; AVX512VL-NEXT: vpternlogd {{.*#+}} xmm0 = xmm0 ^ (mem & (xmm0 ^ xmm1))
1941 ; AVX512VL-NEXT: retq
1942 ;
1943 ; AVX512BW-LABEL: splatconstant_funnnel_v16i8:
1944 ; AVX512BW: # %bb.0:
1945 ; AVX512BW-NEXT: vpsllw $4, %xmm0, %xmm1
1946 ; AVX512BW-NEXT: vpsrlw $4, %xmm0, %xmm0
1947 ; AVX512BW-NEXT: vpternlogd {{.*#+}} zmm0 = zmm0 ^ (mem & (zmm0 ^ zmm1))
1948 ; AVX512BW-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
1949 ; AVX512BW-NEXT: vzeroupper
1950 ; AVX512BW-NEXT: retq
1951 ;
1952 ; AVX512VLBW-LABEL: splatconstant_funnnel_v16i8:
1953 ; AVX512VLBW: # %bb.0:
1954 ; AVX512VLBW-NEXT: vpsllw $4, %xmm0, %xmm1
1955 ; AVX512VLBW-NEXT: vpsrlw $4, %xmm0, %xmm0
1956 ; AVX512VLBW-NEXT: vpternlogd {{.*#+}} xmm0 = xmm0 ^ (mem & (xmm0 ^ xmm1))
1957 ; AVX512VLBW-NEXT: retq
1958 ;
1959 ; AVX512VBMI2-LABEL: splatconstant_funnnel_v16i8:
1960 ; AVX512VBMI2: # %bb.0:
1961 ; AVX512VBMI2-NEXT: vpsllw $4, %xmm0, %xmm1
1962 ; AVX512VBMI2-NEXT: vpsrlw $4, %xmm0, %xmm0
1963 ; AVX512VBMI2-NEXT: vpternlogd {{.*#+}} zmm0 = zmm0 ^ (mem & (zmm0 ^ zmm1))
1964 ; AVX512VBMI2-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
1965 ; AVX512VBMI2-NEXT: vzeroupper
1966 ; AVX512VBMI2-NEXT: retq
1967 ;
1968 ; AVX512VLVBMI2-LABEL: splatconstant_funnnel_v16i8:
1969 ; AVX512VLVBMI2: # %bb.0:
1970 ; AVX512VLVBMI2-NEXT: vpsllw $4, %xmm0, %xmm1
1971 ; AVX512VLVBMI2-NEXT: vpsrlw $4, %xmm0, %xmm0
1972 ; AVX512VLVBMI2-NEXT: vpternlogd {{.*#+}} xmm0 = xmm0 ^ (mem & (xmm0 ^ xmm1))
1973 ; AVX512VLVBMI2-NEXT: retq
1974 ;
1975 ; XOP-LABEL: splatconstant_funnnel_v16i8:
1976 ; XOP: # %bb.0:
1977 ; XOP-NEXT: vprotb $4, %xmm0, %xmm0
1978 ; XOP-NEXT: retq
1979 ;
1980 ; X86-SSE2-LABEL: splatconstant_funnnel_v16i8:
1981 ; X86-SSE2: # %bb.0:
1982 ; X86-SSE2-NEXT: movdqa %xmm0, %xmm1
1983 ; X86-SSE2-NEXT: psrlw $4, %xmm1
1984 ; X86-SSE2-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}, %xmm1
1985 ; X86-SSE2-NEXT: psllw $4, %xmm0
1986 ; X86-SSE2-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0
1987 ; X86-SSE2-NEXT: por %xmm1, %xmm0
1988 ; X86-SSE2-NEXT: retl
1989 %res = call <16 x i8> @llvm.fshr.v16i8(<16 x i8> %x, <16 x i8> %x, <16 x i8> <i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4>)
1990 ret <16 x i8> %res
1991 }