| blob | 71cfd891bb188c8059e9affa479dcab015df6e8e |
1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2 ; RUN: llc < %s -disable-peephole -mtriple=i686-unknown-unknown -mattr=+avx512bw --show-mc-encoding | FileCheck %s --check-prefixes=CHECK,X86
3 ; RUN: llc < %s -disable-peephole -mtriple=x86_64-unknown-unknown -mattr=+avx512bw --show-mc-encoding | FileCheck %s --check-prefixes=CHECK,X64
5 define i32 @test_int_x86_avx512_kadd_d(<32 x i16> %A, <32 x i16> %B) nounwind {
6 ; CHECK-LABEL: test_int_x86_avx512_kadd_d:
7 ; CHECK: # %bb.0: # %entry
8 ; CHECK-NEXT: vptestmw %zmm0, %zmm0, %k0 # encoding: [0x62,0xf2,0xfd,0x48,0x26,0xc0]
9 ; CHECK-NEXT: vptestmw %zmm1, %zmm1, %k1 # encoding: [0x62,0xf2,0xf5,0x48,0x26,0xc9]
10 ; CHECK-NEXT: kaddd %k1, %k0, %k0 # encoding: [0xc4,0xe1,0xfd,0x4a,0xc1]
11 ; CHECK-NEXT: xorl %eax, %eax # encoding: [0x31,0xc0]
12 ; CHECK-NEXT: kortestd %k0, %k0 # encoding: [0xc4,0xe1,0xf9,0x98,0xc0]
13 ; CHECK-NEXT: sete %al # encoding: [0x0f,0x94,0xc0]
14 ; CHECK-NEXT: vzeroupper # encoding: [0xc5,0xf8,0x77]
15 ; CHECK-NEXT: ret{{[l|q]}} # encoding: [0xc3]
16 entry:
17 %0 = icmp ne <32 x i16> %A, zeroinitializer
18 %1 = icmp ne <32 x i16> %B, zeroinitializer
19 %2 = call <32 x i1> @llvm.x86.avx512.kadd.d(<32 x i1> %0, <32 x i1> %1)
20 %3 = bitcast <32 x i1> %2 to i32
21 %4 = icmp eq i32 %3, 0
22 %5 = zext i1 %4 to i32
23 ret i32 %5
24 }
25 declare <32 x i1> @llvm.x86.avx512.kadd.d(<32 x i1>, <32 x i1>)
27 define i32 @test_int_x86_avx512_kadd_q(<64 x i8> %A, <64 x i8> %B) nounwind {
28 ; X86-LABEL: test_int_x86_avx512_kadd_q:
29 ; X86: # %bb.0: # %entry
30 ; X86-NEXT: vptestmb %zmm0, %zmm0, %k0 # encoding: [0x62,0xf2,0x7d,0x48,0x26,0xc0]
31 ; X86-NEXT: vptestmb %zmm1, %zmm1, %k1 # encoding: [0x62,0xf2,0x75,0x48,0x26,0xc9]
32 ; X86-NEXT: kaddq %k1, %k0, %k0 # encoding: [0xc4,0xe1,0xfc,0x4a,0xc1]
33 ; X86-NEXT: kshiftrq $32, %k0, %k1 # encoding: [0xc4,0xe3,0xf9,0x31,0xc8,0x20]
34 ; X86-NEXT: xorl %eax, %eax # encoding: [0x31,0xc0]
35 ; X86-NEXT: kortestd %k1, %k0 # encoding: [0xc4,0xe1,0xf9,0x98,0xc1]
36 ; X86-NEXT: sete %al # encoding: [0x0f,0x94,0xc0]
37 ; X86-NEXT: vzeroupper # encoding: [0xc5,0xf8,0x77]
38 ; X86-NEXT: retl # encoding: [0xc3]
39 ;
40 ; X64-LABEL: test_int_x86_avx512_kadd_q:
41 ; X64: # %bb.0: # %entry
42 ; X64-NEXT: vptestmb %zmm0, %zmm0, %k0 # encoding: [0x62,0xf2,0x7d,0x48,0x26,0xc0]
43 ; X64-NEXT: vptestmb %zmm1, %zmm1, %k1 # encoding: [0x62,0xf2,0x75,0x48,0x26,0xc9]
44 ; X64-NEXT: kaddq %k1, %k0, %k0 # encoding: [0xc4,0xe1,0xfc,0x4a,0xc1]
45 ; X64-NEXT: xorl %eax, %eax # encoding: [0x31,0xc0]
46 ; X64-NEXT: kortestq %k0, %k0 # encoding: [0xc4,0xe1,0xf8,0x98,0xc0]
47 ; X64-NEXT: sete %al # encoding: [0x0f,0x94,0xc0]
48 ; X64-NEXT: vzeroupper # encoding: [0xc5,0xf8,0x77]
49 ; X64-NEXT: retq # encoding: [0xc3]
50 entry:
51 %0 = icmp ne <64 x i8> %A, zeroinitializer
52 %1 = icmp ne <64 x i8> %B, zeroinitializer
53 %2 = call <64 x i1> @llvm.x86.avx512.kadd.q(<64 x i1> %0, <64 x i1> %1)
54 %3 = bitcast <64 x i1> %2 to i64
55 %4 = icmp eq i64 %3, 0
56 %5 = zext i1 %4 to i32
57 ret i32 %5
58 }
59 declare <64 x i1> @llvm.x86.avx512.kadd.q(<64 x i1>, <64 x i1>)
61 define i32 @test_x86_avx512_ktestc_d(<32 x i16> %A, <32 x i16> %B) {
62 ; CHECK-LABEL: test_x86_avx512_ktestc_d:
63 ; CHECK: # %bb.0:
64 ; CHECK-NEXT: vptestmw %zmm0, %zmm0, %k0 # encoding: [0x62,0xf2,0xfd,0x48,0x26,0xc0]
65 ; CHECK-NEXT: vptestmw %zmm1, %zmm1, %k1 # encoding: [0x62,0xf2,0xf5,0x48,0x26,0xc9]
66 ; CHECK-NEXT: xorl %eax, %eax # encoding: [0x31,0xc0]
67 ; CHECK-NEXT: ktestd %k1, %k0 # encoding: [0xc4,0xe1,0xf9,0x99,0xc1]
68 ; CHECK-NEXT: setb %al # encoding: [0x0f,0x92,0xc0]
69 ; CHECK-NEXT: vzeroupper # encoding: [0xc5,0xf8,0x77]
70 ; CHECK-NEXT: ret{{[l|q]}} # encoding: [0xc3]
71 %1 = icmp ne <32 x i16> %A, zeroinitializer
72 %2 = icmp ne <32 x i16> %B, zeroinitializer
73 %res = call i32 @llvm.x86.avx512.ktestc.d(<32 x i1> %1, <32 x i1> %2) ; <i32> [#uses=1]
74 ret i32 %res
75 }
76 declare i32 @llvm.x86.avx512.ktestc.d(<32 x i1>, <32 x i1>) nounwind readnone
78 define i32 @test_x86_avx512_ktestz_d(<32 x i16> %A, <32 x i16> %B) {
79 ; CHECK-LABEL: test_x86_avx512_ktestz_d:
80 ; CHECK: # %bb.0:
81 ; CHECK-NEXT: vptestmw %zmm0, %zmm0, %k0 # encoding: [0x62,0xf2,0xfd,0x48,0x26,0xc0]
82 ; CHECK-NEXT: vptestmw %zmm1, %zmm1, %k1 # encoding: [0x62,0xf2,0xf5,0x48,0x26,0xc9]
83 ; CHECK-NEXT: xorl %eax, %eax # encoding: [0x31,0xc0]
84 ; CHECK-NEXT: ktestd %k1, %k0 # encoding: [0xc4,0xe1,0xf9,0x99,0xc1]
85 ; CHECK-NEXT: sete %al # encoding: [0x0f,0x94,0xc0]
86 ; CHECK-NEXT: vzeroupper # encoding: [0xc5,0xf8,0x77]
87 ; CHECK-NEXT: ret{{[l|q]}} # encoding: [0xc3]
88 %1 = icmp ne <32 x i16> %A, zeroinitializer
89 %2 = icmp ne <32 x i16> %B, zeroinitializer
90 %res = call i32 @llvm.x86.avx512.ktestz.d(<32 x i1> %1, <32 x i1> %2) ; <i32> [#uses=1]
91 ret i32 %res
92 }
93 declare i32 @llvm.x86.avx512.ktestz.d(<32 x i1>, <32 x i1>) nounwind readnone
95 define i32 @test_x86_avx512_ktestc_q(<64 x i8> %A, <64 x i8> %B) {
96 ; CHECK-LABEL: test_x86_avx512_ktestc_q:
97 ; CHECK: # %bb.0:
98 ; CHECK-NEXT: vptestmb %zmm0, %zmm0, %k0 # encoding: [0x62,0xf2,0x7d,0x48,0x26,0xc0]
99 ; CHECK-NEXT: vptestmb %zmm1, %zmm1, %k1 # encoding: [0x62,0xf2,0x75,0x48,0x26,0xc9]
100 ; CHECK-NEXT: xorl %eax, %eax # encoding: [0x31,0xc0]
101 ; CHECK-NEXT: ktestq %k1, %k0 # encoding: [0xc4,0xe1,0xf8,0x99,0xc1]
102 ; CHECK-NEXT: setb %al # encoding: [0x0f,0x92,0xc0]
103 ; CHECK-NEXT: vzeroupper # encoding: [0xc5,0xf8,0x77]
104 ; CHECK-NEXT: ret{{[l|q]}} # encoding: [0xc3]
105 %1 = icmp ne <64 x i8> %A, zeroinitializer
106 %2 = icmp ne <64 x i8> %B, zeroinitializer
107 %res = call i32 @llvm.x86.avx512.ktestc.q(<64 x i1> %1, <64 x i1> %2) ; <i32> [#uses=1]
108 ret i32 %res
109 }
110 declare i32 @llvm.x86.avx512.ktestc.q(<64 x i1>, <64 x i1>) nounwind readnone
112 define i32 @test_x86_avx512_ktestz_q(<64 x i8> %A, <64 x i8> %B) {
113 ; CHECK-LABEL: test_x86_avx512_ktestz_q:
114 ; CHECK: # %bb.0:
115 ; CHECK-NEXT: vptestmb %zmm0, %zmm0, %k0 # encoding: [0x62,0xf2,0x7d,0x48,0x26,0xc0]
116 ; CHECK-NEXT: vptestmb %zmm1, %zmm1, %k1 # encoding: [0x62,0xf2,0x75,0x48,0x26,0xc9]
117 ; CHECK-NEXT: xorl %eax, %eax # encoding: [0x31,0xc0]
118 ; CHECK-NEXT: ktestq %k1, %k0 # encoding: [0xc4,0xe1,0xf8,0x99,0xc1]
119 ; CHECK-NEXT: sete %al # encoding: [0x0f,0x94,0xc0]
120 ; CHECK-NEXT: vzeroupper # encoding: [0xc5,0xf8,0x77]
121 ; CHECK-NEXT: ret{{[l|q]}} # encoding: [0xc3]
122 %1 = icmp ne <64 x i8> %A, zeroinitializer
123 %2 = icmp ne <64 x i8> %B, zeroinitializer
124 %res = call i32 @llvm.x86.avx512.ktestz.q(<64 x i1> %1, <64 x i1> %2) ; <i32> [#uses=1]
125 ret i32 %res
126 }
127 declare i32 @llvm.x86.avx512.ktestz.q(<64 x i1>, <64 x i1>) nounwind readnone
129 define <32 x i16> @test_mask_packs_epi32_rr_512(<16 x i32> %a, <16 x i32> %b) {
130 ; CHECK-LABEL: test_mask_packs_epi32_rr_512:
131 ; CHECK: # %bb.0:
132 ; CHECK-NEXT: vpackssdw %zmm1, %zmm0, %zmm0 # encoding: [0x62,0xf1,0x7d,0x48,0x6b,0xc1]
133 ; CHECK-NEXT: ret{{[l|q]}} # encoding: [0xc3]
134 %1 = call <32 x i16> @llvm.x86.avx512.packssdw.512(<16 x i32> %a, <16 x i32> %b)
135 ret <32 x i16> %1
136 }
138 define <32 x i16> @test_mask_packs_epi32_rrk_512(<16 x i32> %a, <16 x i32> %b, <32 x i16> %passThru, i32 %mask) {
139 ; X86-LABEL: test_mask_packs_epi32_rrk_512:
140 ; X86: # %bb.0:
141 ; X86-NEXT: kmovd {{[0-9]+}}(%esp), %k1 # encoding: [0xc4,0xe1,0xf9,0x90,0x4c,0x24,0x04]
142 ; X86-NEXT: vpackssdw %zmm1, %zmm0, %zmm2 {%k1} # encoding: [0x62,0xf1,0x7d,0x49,0x6b,0xd1]
143 ; X86-NEXT: vmovdqa64 %zmm2, %zmm0 # encoding: [0x62,0xf1,0xfd,0x48,0x6f,0xc2]
144 ; X86-NEXT: retl # encoding: [0xc3]
145 ;
146 ; X64-LABEL: test_mask_packs_epi32_rrk_512:
147 ; X64: # %bb.0:
148 ; X64-NEXT: kmovd %edi, %k1 # encoding: [0xc5,0xfb,0x92,0xcf]
149 ; X64-NEXT: vpackssdw %zmm1, %zmm0, %zmm2 {%k1} # encoding: [0x62,0xf1,0x7d,0x49,0x6b,0xd1]
150 ; X64-NEXT: vmovdqa64 %zmm2, %zmm0 # encoding: [0x62,0xf1,0xfd,0x48,0x6f,0xc2]
151 ; X64-NEXT: retq # encoding: [0xc3]
152 %1 = call <32 x i16> @llvm.x86.avx512.packssdw.512(<16 x i32> %a, <16 x i32> %b)
153 %2 = bitcast i32 %mask to <32 x i1>
154 %3 = select <32 x i1> %2, <32 x i16> %1, <32 x i16> %passThru
155 ret <32 x i16> %3
156 }
158 define <32 x i16> @test_mask_packs_epi32_rrkz_512(<16 x i32> %a, <16 x i32> %b, i32 %mask) {
159 ; X86-LABEL: test_mask_packs_epi32_rrkz_512:
160 ; X86: # %bb.0:
161 ; X86-NEXT: kmovd {{[0-9]+}}(%esp), %k1 # encoding: [0xc4,0xe1,0xf9,0x90,0x4c,0x24,0x04]
162 ; X86-NEXT: vpackssdw %zmm1, %zmm0, %zmm0 {%k1} {z} # encoding: [0x62,0xf1,0x7d,0xc9,0x6b,0xc1]
163 ; X86-NEXT: retl # encoding: [0xc3]
164 ;
165 ; X64-LABEL: test_mask_packs_epi32_rrkz_512:
166 ; X64: # %bb.0:
167 ; X64-NEXT: kmovd %edi, %k1 # encoding: [0xc5,0xfb,0x92,0xcf]
168 ; X64-NEXT: vpackssdw %zmm1, %zmm0, %zmm0 {%k1} {z} # encoding: [0x62,0xf1,0x7d,0xc9,0x6b,0xc1]
169 ; X64-NEXT: retq # encoding: [0xc3]
170 %1 = call <32 x i16> @llvm.x86.avx512.packssdw.512(<16 x i32> %a, <16 x i32> %b)
171 %2 = bitcast i32 %mask to <32 x i1>
172 %3 = select <32 x i1> %2, <32 x i16> %1, <32 x i16> zeroinitializer
173 ret <32 x i16> %3
174 }
176 define <32 x i16> @test_mask_packs_epi32_rm_512(<16 x i32> %a, <16 x i32>* %ptr_b) {
177 ; X86-LABEL: test_mask_packs_epi32_rm_512:
178 ; X86: # %bb.0:
179 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x04]
180 ; X86-NEXT: vpackssdw (%eax), %zmm0, %zmm0 # encoding: [0x62,0xf1,0x7d,0x48,0x6b,0x00]
181 ; X86-NEXT: retl # encoding: [0xc3]
182 ;
183 ; X64-LABEL: test_mask_packs_epi32_rm_512:
184 ; X64: # %bb.0:
185 ; X64-NEXT: vpackssdw (%rdi), %zmm0, %zmm0 # encoding: [0x62,0xf1,0x7d,0x48,0x6b,0x07]
186 ; X64-NEXT: retq # encoding: [0xc3]
187 %b = load <16 x i32>, <16 x i32>* %ptr_b
188 %1 = call <32 x i16> @llvm.x86.avx512.packssdw.512(<16 x i32> %a, <16 x i32> %b)
189 ret <32 x i16> %1
190 }
192 define <32 x i16> @test_mask_packs_epi32_rmk_512(<16 x i32> %a, <16 x i32>* %ptr_b, <32 x i16> %passThru, i32 %mask) {
193 ; X86-LABEL: test_mask_packs_epi32_rmk_512:
194 ; X86: # %bb.0:
195 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x04]
196 ; X86-NEXT: kmovd {{[0-9]+}}(%esp), %k1 # encoding: [0xc4,0xe1,0xf9,0x90,0x4c,0x24,0x08]
197 ; X86-NEXT: vpackssdw (%eax), %zmm0, %zmm1 {%k1} # encoding: [0x62,0xf1,0x7d,0x49,0x6b,0x08]
198 ; X86-NEXT: vmovdqa64 %zmm1, %zmm0 # encoding: [0x62,0xf1,0xfd,0x48,0x6f,0xc1]
199 ; X86-NEXT: retl # encoding: [0xc3]
200 ;
201 ; X64-LABEL: test_mask_packs_epi32_rmk_512:
202 ; X64: # %bb.0:
203 ; X64-NEXT: kmovd %esi, %k1 # encoding: [0xc5,0xfb,0x92,0xce]
204 ; X64-NEXT: vpackssdw (%rdi), %zmm0, %zmm1 {%k1} # encoding: [0x62,0xf1,0x7d,0x49,0x6b,0x0f]
205 ; X64-NEXT: vmovdqa64 %zmm1, %zmm0 # encoding: [0x62,0xf1,0xfd,0x48,0x6f,0xc1]
206 ; X64-NEXT: retq # encoding: [0xc3]
207 %b = load <16 x i32>, <16 x i32>* %ptr_b
208 %1 = call <32 x i16> @llvm.x86.avx512.packssdw.512(<16 x i32> %a, <16 x i32> %b)
209 %2 = bitcast i32 %mask to <32 x i1>
210 %3 = select <32 x i1> %2, <32 x i16> %1, <32 x i16> %passThru
211 ret <32 x i16> %3
212 }
214 define <32 x i16> @test_mask_packs_epi32_rmkz_512(<16 x i32> %a, <16 x i32>* %ptr_b, i32 %mask) {
215 ; X86-LABEL: test_mask_packs_epi32_rmkz_512:
216 ; X86: # %bb.0:
217 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x04]
218 ; X86-NEXT: kmovd {{[0-9]+}}(%esp), %k1 # encoding: [0xc4,0xe1,0xf9,0x90,0x4c,0x24,0x08]
219 ; X86-NEXT: vpackssdw (%eax), %zmm0, %zmm0 {%k1} {z} # encoding: [0x62,0xf1,0x7d,0xc9,0x6b,0x00]
220 ; X86-NEXT: retl # encoding: [0xc3]
221 ;
222 ; X64-LABEL: test_mask_packs_epi32_rmkz_512:
223 ; X64: # %bb.0:
224 ; X64-NEXT: kmovd %esi, %k1 # encoding: [0xc5,0xfb,0x92,0xce]
225 ; X64-NEXT: vpackssdw (%rdi), %zmm0, %zmm0 {%k1} {z} # encoding: [0x62,0xf1,0x7d,0xc9,0x6b,0x07]
226 ; X64-NEXT: retq # encoding: [0xc3]
227 %b = load <16 x i32>, <16 x i32>* %ptr_b
228 %1 = call <32 x i16> @llvm.x86.avx512.packssdw.512(<16 x i32> %a, <16 x i32> %b)
229 %2 = bitcast i32 %mask to <32 x i1>
230 %3 = select <32 x i1> %2, <32 x i16> %1, <32 x i16> zeroinitializer
231 ret <32 x i16> %3
232 }
234 define <32 x i16> @test_mask_packs_epi32_rmb_512(<16 x i32> %a, i32* %ptr_b) {
235 ; X86-LABEL: test_mask_packs_epi32_rmb_512:
236 ; X86: # %bb.0:
237 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x04]
238 ; X86-NEXT: vpackssdw (%eax){1to16}, %zmm0, %zmm0 # encoding: [0x62,0xf1,0x7d,0x58,0x6b,0x00]
239 ; X86-NEXT: retl # encoding: [0xc3]
240 ;
241 ; X64-LABEL: test_mask_packs_epi32_rmb_512:
242 ; X64: # %bb.0:
243 ; X64-NEXT: vpackssdw (%rdi){1to16}, %zmm0, %zmm0 # encoding: [0x62,0xf1,0x7d,0x58,0x6b,0x07]
244 ; X64-NEXT: retq # encoding: [0xc3]
245 %q = load i32, i32* %ptr_b
246 %vecinit.i = insertelement <16 x i32> undef, i32 %q, i32 0
247 %b = shufflevector <16 x i32> %vecinit.i, <16 x i32> undef, <16 x i32> zeroinitializer
248 %1 = call <32 x i16> @llvm.x86.avx512.packssdw.512(<16 x i32> %a, <16 x i32> %b)
249 ret <32 x i16> %1
250 }
252 define <32 x i16> @test_mask_packs_epi32_rmbk_512(<16 x i32> %a, i32* %ptr_b, <32 x i16> %passThru, i32 %mask) {
253 ; X86-LABEL: test_mask_packs_epi32_rmbk_512:
254 ; X86: # %bb.0:
255 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x04]
256 ; X86-NEXT: kmovd {{[0-9]+}}(%esp), %k1 # encoding: [0xc4,0xe1,0xf9,0x90,0x4c,0x24,0x08]
257 ; X86-NEXT: vpackssdw (%eax){1to16}, %zmm0, %zmm1 {%k1} # encoding: [0x62,0xf1,0x7d,0x59,0x6b,0x08]
258 ; X86-NEXT: vmovdqa64 %zmm1, %zmm0 # encoding: [0x62,0xf1,0xfd,0x48,0x6f,0xc1]
259 ; X86-NEXT: retl # encoding: [0xc3]
260 ;
261 ; X64-LABEL: test_mask_packs_epi32_rmbk_512:
262 ; X64: # %bb.0:
263 ; X64-NEXT: kmovd %esi, %k1 # encoding: [0xc5,0xfb,0x92,0xce]
264 ; X64-NEXT: vpackssdw (%rdi){1to16}, %zmm0, %zmm1 {%k1} # encoding: [0x62,0xf1,0x7d,0x59,0x6b,0x0f]
265 ; X64-NEXT: vmovdqa64 %zmm1, %zmm0 # encoding: [0x62,0xf1,0xfd,0x48,0x6f,0xc1]
266 ; X64-NEXT: retq # encoding: [0xc3]
267 %q = load i32, i32* %ptr_b
268 %vecinit.i = insertelement <16 x i32> undef, i32 %q, i32 0
269 %b = shufflevector <16 x i32> %vecinit.i, <16 x i32> undef, <16 x i32> zeroinitializer
270 %1 = call <32 x i16> @llvm.x86.avx512.packssdw.512(<16 x i32> %a, <16 x i32> %b)
271 %2 = bitcast i32 %mask to <32 x i1>
272 %3 = select <32 x i1> %2, <32 x i16> %1, <32 x i16> %passThru
273 ret <32 x i16> %3
274 }
276 define <32 x i16> @test_mask_packs_epi32_rmbkz_512(<16 x i32> %a, i32* %ptr_b, i32 %mask) {
277 ; X86-LABEL: test_mask_packs_epi32_rmbkz_512:
278 ; X86: # %bb.0:
279 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x04]
280 ; X86-NEXT: kmovd {{[0-9]+}}(%esp), %k1 # encoding: [0xc4,0xe1,0xf9,0x90,0x4c,0x24,0x08]
281 ; X86-NEXT: vpackssdw (%eax){1to16}, %zmm0, %zmm0 {%k1} {z} # encoding: [0x62,0xf1,0x7d,0xd9,0x6b,0x00]
282 ; X86-NEXT: retl # encoding: [0xc3]
283 ;
284 ; X64-LABEL: test_mask_packs_epi32_rmbkz_512:
285 ; X64: # %bb.0:
286 ; X64-NEXT: kmovd %esi, %k1 # encoding: [0xc5,0xfb,0x92,0xce]
287 ; X64-NEXT: vpackssdw (%rdi){1to16}, %zmm0, %zmm0 {%k1} {z} # encoding: [0x62,0xf1,0x7d,0xd9,0x6b,0x07]
288 ; X64-NEXT: retq # encoding: [0xc3]
289 %q = load i32, i32* %ptr_b
290 %vecinit.i = insertelement <16 x i32> undef, i32 %q, i32 0
291 %b = shufflevector <16 x i32> %vecinit.i, <16 x i32> undef, <16 x i32> zeroinitializer
292 %1 = call <32 x i16> @llvm.x86.avx512.packssdw.512(<16 x i32> %a, <16 x i32> %b)
293 %2 = bitcast i32 %mask to <32 x i1>
294 %3 = select <32 x i1> %2, <32 x i16> %1, <32 x i16> zeroinitializer
295 ret <32 x i16> %3
296 }
298 declare <32 x i16> @llvm.x86.avx512.packssdw.512(<16 x i32>, <16 x i32>)
300 define <64 x i8> @test_mask_packs_epi16_rr_512(<32 x i16> %a, <32 x i16> %b) {
301 ; CHECK-LABEL: test_mask_packs_epi16_rr_512:
302 ; CHECK: # %bb.0:
303 ; CHECK-NEXT: vpacksswb %zmm1, %zmm0, %zmm0 # encoding: [0x62,0xf1,0x7d,0x48,0x63,0xc1]
304 ; CHECK-NEXT: ret{{[l|q]}} # encoding: [0xc3]
305 %1 = call <64 x i8> @llvm.x86.avx512.packsswb.512(<32 x i16> %a, <32 x i16> %b)
306 ret <64 x i8> %1
307 }
309 define <64 x i8> @test_mask_packs_epi16_rrk_512(<32 x i16> %a, <32 x i16> %b, <64 x i8> %passThru, i64 %mask) {
310 ; X86-LABEL: test_mask_packs_epi16_rrk_512:
311 ; X86: # %bb.0:
312 ; X86-NEXT: kmovq {{[0-9]+}}(%esp), %k1 # encoding: [0xc4,0xe1,0xf8,0x90,0x4c,0x24,0x04]
313 ; X86-NEXT: vpacksswb %zmm1, %zmm0, %zmm2 {%k1} # encoding: [0x62,0xf1,0x7d,0x49,0x63,0xd1]
314 ; X86-NEXT: vmovdqa64 %zmm2, %zmm0 # encoding: [0x62,0xf1,0xfd,0x48,0x6f,0xc2]
315 ; X86-NEXT: retl # encoding: [0xc3]
316 ;
317 ; X64-LABEL: test_mask_packs_epi16_rrk_512:
318 ; X64: # %bb.0:
319 ; X64-NEXT: kmovq %rdi, %k1 # encoding: [0xc4,0xe1,0xfb,0x92,0xcf]
320 ; X64-NEXT: vpacksswb %zmm1, %zmm0, %zmm2 {%k1} # encoding: [0x62,0xf1,0x7d,0x49,0x63,0xd1]
321 ; X64-NEXT: vmovdqa64 %zmm2, %zmm0 # encoding: [0x62,0xf1,0xfd,0x48,0x6f,0xc2]
322 ; X64-NEXT: retq # encoding: [0xc3]
323 %1 = call <64 x i8> @llvm.x86.avx512.packsswb.512(<32 x i16> %a, <32 x i16> %b)
324 %2 = bitcast i64 %mask to <64 x i1>
325 %3 = select <64 x i1> %2, <64 x i8> %1, <64 x i8> %passThru
326 ret <64 x i8> %3
327 }
329 define <64 x i8> @test_mask_packs_epi16_rrkz_512(<32 x i16> %a, <32 x i16> %b, i64 %mask) {
330 ; X86-LABEL: test_mask_packs_epi16_rrkz_512:
331 ; X86: # %bb.0:
332 ; X86-NEXT: kmovq {{[0-9]+}}(%esp), %k1 # encoding: [0xc4,0xe1,0xf8,0x90,0x4c,0x24,0x04]
333 ; X86-NEXT: vpacksswb %zmm1, %zmm0, %zmm0 {%k1} {z} # encoding: [0x62,0xf1,0x7d,0xc9,0x63,0xc1]
334 ; X86-NEXT: retl # encoding: [0xc3]
335 ;
336 ; X64-LABEL: test_mask_packs_epi16_rrkz_512:
337 ; X64: # %bb.0:
338 ; X64-NEXT: kmovq %rdi, %k1 # encoding: [0xc4,0xe1,0xfb,0x92,0xcf]
339 ; X64-NEXT: vpacksswb %zmm1, %zmm0, %zmm0 {%k1} {z} # encoding: [0x62,0xf1,0x7d,0xc9,0x63,0xc1]
340 ; X64-NEXT: retq # encoding: [0xc3]
341 %1 = call <64 x i8> @llvm.x86.avx512.packsswb.512(<32 x i16> %a, <32 x i16> %b)
342 %2 = bitcast i64 %mask to <64 x i1>
343 %3 = select <64 x i1> %2, <64 x i8> %1, <64 x i8> zeroinitializer
344 ret <64 x i8> %3
345 }
347 define <64 x i8> @test_mask_packs_epi16_rm_512(<32 x i16> %a, <32 x i16>* %ptr_b) {
348 ; X86-LABEL: test_mask_packs_epi16_rm_512:
349 ; X86: # %bb.0:
350 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x04]
351 ; X86-NEXT: vpacksswb (%eax), %zmm0, %zmm0 # encoding: [0x62,0xf1,0x7d,0x48,0x63,0x00]
352 ; X86-NEXT: retl # encoding: [0xc3]
353 ;
354 ; X64-LABEL: test_mask_packs_epi16_rm_512:
355 ; X64: # %bb.0:
356 ; X64-NEXT: vpacksswb (%rdi), %zmm0, %zmm0 # encoding: [0x62,0xf1,0x7d,0x48,0x63,0x07]
357 ; X64-NEXT: retq # encoding: [0xc3]
358 %b = load <32 x i16>, <32 x i16>* %ptr_b
359 %1 = call <64 x i8> @llvm.x86.avx512.packsswb.512(<32 x i16> %a, <32 x i16> %b)
360 ret <64 x i8> %1
361 }
363 define <64 x i8> @test_mask_packs_epi16_rmk_512(<32 x i16> %a, <32 x i16>* %ptr_b, <64 x i8> %passThru, i64 %mask) {
364 ; X86-LABEL: test_mask_packs_epi16_rmk_512:
365 ; X86: # %bb.0:
366 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x04]
367 ; X86-NEXT: kmovq {{[0-9]+}}(%esp), %k1 # encoding: [0xc4,0xe1,0xf8,0x90,0x4c,0x24,0x08]
368 ; X86-NEXT: vpacksswb (%eax), %zmm0, %zmm1 {%k1} # encoding: [0x62,0xf1,0x7d,0x49,0x63,0x08]
369 ; X86-NEXT: vmovdqa64 %zmm1, %zmm0 # encoding: [0x62,0xf1,0xfd,0x48,0x6f,0xc1]
370 ; X86-NEXT: retl # encoding: [0xc3]
371 ;
372 ; X64-LABEL: test_mask_packs_epi16_rmk_512:
373 ; X64: # %bb.0:
374 ; X64-NEXT: kmovq %rsi, %k1 # encoding: [0xc4,0xe1,0xfb,0x92,0xce]
375 ; X64-NEXT: vpacksswb (%rdi), %zmm0, %zmm1 {%k1} # encoding: [0x62,0xf1,0x7d,0x49,0x63,0x0f]
376 ; X64-NEXT: vmovdqa64 %zmm1, %zmm0 # encoding: [0x62,0xf1,0xfd,0x48,0x6f,0xc1]
377 ; X64-NEXT: retq # encoding: [0xc3]
378 %b = load <32 x i16>, <32 x i16>* %ptr_b
379 %1 = call <64 x i8> @llvm.x86.avx512.packsswb.512(<32 x i16> %a, <32 x i16> %b)
380 %2 = bitcast i64 %mask to <64 x i1>
381 %3 = select <64 x i1> %2, <64 x i8> %1, <64 x i8> %passThru
382 ret <64 x i8> %3
383 }
385 define <64 x i8> @test_mask_packs_epi16_rmkz_512(<32 x i16> %a, <32 x i16>* %ptr_b, i64 %mask) {
386 ; X86-LABEL: test_mask_packs_epi16_rmkz_512:
387 ; X86: # %bb.0:
388 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x04]
389 ; X86-NEXT: kmovq {{[0-9]+}}(%esp), %k1 # encoding: [0xc4,0xe1,0xf8,0x90,0x4c,0x24,0x08]
390 ; X86-NEXT: vpacksswb (%eax), %zmm0, %zmm0 {%k1} {z} # encoding: [0x62,0xf1,0x7d,0xc9,0x63,0x00]
391 ; X86-NEXT: retl # encoding: [0xc3]
392 ;
393 ; X64-LABEL: test_mask_packs_epi16_rmkz_512:
394 ; X64: # %bb.0:
395 ; X64-NEXT: kmovq %rsi, %k1 # encoding: [0xc4,0xe1,0xfb,0x92,0xce]
396 ; X64-NEXT: vpacksswb (%rdi), %zmm0, %zmm0 {%k1} {z} # encoding: [0x62,0xf1,0x7d,0xc9,0x63,0x07]
397 ; X64-NEXT: retq # encoding: [0xc3]
398 %b = load <32 x i16>, <32 x i16>* %ptr_b
399 %1 = call <64 x i8> @llvm.x86.avx512.packsswb.512(<32 x i16> %a, <32 x i16> %b)
400 %2 = bitcast i64 %mask to <64 x i1>
401 %3 = select <64 x i1> %2, <64 x i8> %1, <64 x i8> zeroinitializer
402 ret <64 x i8> %3
403 }
405 declare <64 x i8> @llvm.x86.avx512.packsswb.512(<32 x i16>, <32 x i16>)
408 define <32 x i16> @test_mask_packus_epi32_rr_512(<16 x i32> %a, <16 x i32> %b) {
409 ; CHECK-LABEL: test_mask_packus_epi32_rr_512:
410 ; CHECK: # %bb.0:
411 ; CHECK-NEXT: vpackusdw %zmm1, %zmm0, %zmm0 # encoding: [0x62,0xf2,0x7d,0x48,0x2b,0xc1]
412 ; CHECK-NEXT: ret{{[l|q]}} # encoding: [0xc3]
413 %1 = call <32 x i16> @llvm.x86.avx512.packusdw.512(<16 x i32> %a, <16 x i32> %b)
414 ret <32 x i16> %1
415 }
417 define <32 x i16> @test_mask_packus_epi32_rrk_512(<16 x i32> %a, <16 x i32> %b, <32 x i16> %passThru, i32 %mask) {
418 ; X86-LABEL: test_mask_packus_epi32_rrk_512:
419 ; X86: # %bb.0:
420 ; X86-NEXT: kmovd {{[0-9]+}}(%esp), %k1 # encoding: [0xc4,0xe1,0xf9,0x90,0x4c,0x24,0x04]
421 ; X86-NEXT: vpackusdw %zmm1, %zmm0, %zmm2 {%k1} # encoding: [0x62,0xf2,0x7d,0x49,0x2b,0xd1]
422 ; X86-NEXT: vmovdqa64 %zmm2, %zmm0 # encoding: [0x62,0xf1,0xfd,0x48,0x6f,0xc2]
423 ; X86-NEXT: retl # encoding: [0xc3]
424 ;
425 ; X64-LABEL: test_mask_packus_epi32_rrk_512:
426 ; X64: # %bb.0:
427 ; X64-NEXT: kmovd %edi, %k1 # encoding: [0xc5,0xfb,0x92,0xcf]
428 ; X64-NEXT: vpackusdw %zmm1, %zmm0, %zmm2 {%k1} # encoding: [0x62,0xf2,0x7d,0x49,0x2b,0xd1]
429 ; X64-NEXT: vmovdqa64 %zmm2, %zmm0 # encoding: [0x62,0xf1,0xfd,0x48,0x6f,0xc2]
430 ; X64-NEXT: retq # encoding: [0xc3]
431 %1 = call <32 x i16> @llvm.x86.avx512.packusdw.512(<16 x i32> %a, <16 x i32> %b)
432 %2 = bitcast i32 %mask to <32 x i1>
433 %3 = select <32 x i1> %2, <32 x i16> %1, <32 x i16> %passThru
434 ret <32 x i16> %3
435 }
437 define <32 x i16> @test_mask_packus_epi32_rrkz_512(<16 x i32> %a, <16 x i32> %b, i32 %mask) {
438 ; X86-LABEL: test_mask_packus_epi32_rrkz_512:
439 ; X86: # %bb.0:
440 ; X86-NEXT: kmovd {{[0-9]+}}(%esp), %k1 # encoding: [0xc4,0xe1,0xf9,0x90,0x4c,0x24,0x04]
441 ; X86-NEXT: vpackusdw %zmm1, %zmm0, %zmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7d,0xc9,0x2b,0xc1]
442 ; X86-NEXT: retl # encoding: [0xc3]
443 ;
444 ; X64-LABEL: test_mask_packus_epi32_rrkz_512:
445 ; X64: # %bb.0:
446 ; X64-NEXT: kmovd %edi, %k1 # encoding: [0xc5,0xfb,0x92,0xcf]
447 ; X64-NEXT: vpackusdw %zmm1, %zmm0, %zmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7d,0xc9,0x2b,0xc1]
448 ; X64-NEXT: retq # encoding: [0xc3]
449 %1 = call <32 x i16> @llvm.x86.avx512.packusdw.512(<16 x i32> %a, <16 x i32> %b)
450 %2 = bitcast i32 %mask to <32 x i1>
451 %3 = select <32 x i1> %2, <32 x i16> %1, <32 x i16> zeroinitializer
452 ret <32 x i16> %3
453 }
455 define <32 x i16> @test_mask_packus_epi32_rm_512(<16 x i32> %a, <16 x i32>* %ptr_b) {
456 ; X86-LABEL: test_mask_packus_epi32_rm_512:
457 ; X86: # %bb.0:
458 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x04]
459 ; X86-NEXT: vpackusdw (%eax), %zmm0, %zmm0 # encoding: [0x62,0xf2,0x7d,0x48,0x2b,0x00]
460 ; X86-NEXT: retl # encoding: [0xc3]
461 ;
462 ; X64-LABEL: test_mask_packus_epi32_rm_512:
463 ; X64: # %bb.0:
464 ; X64-NEXT: vpackusdw (%rdi), %zmm0, %zmm0 # encoding: [0x62,0xf2,0x7d,0x48,0x2b,0x07]
465 ; X64-NEXT: retq # encoding: [0xc3]
466 %b = load <16 x i32>, <16 x i32>* %ptr_b
467 %1 = call <32 x i16> @llvm.x86.avx512.packusdw.512(<16 x i32> %a, <16 x i32> %b)
468 ret <32 x i16> %1
469 }
471 define <32 x i16> @test_mask_packus_epi32_rmk_512(<16 x i32> %a, <16 x i32>* %ptr_b, <32 x i16> %passThru, i32 %mask) {
472 ; X86-LABEL: test_mask_packus_epi32_rmk_512:
473 ; X86: # %bb.0:
474 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x04]
475 ; X86-NEXT: kmovd {{[0-9]+}}(%esp), %k1 # encoding: [0xc4,0xe1,0xf9,0x90,0x4c,0x24,0x08]
476 ; X86-NEXT: vpackusdw (%eax), %zmm0, %zmm1 {%k1} # encoding: [0x62,0xf2,0x7d,0x49,0x2b,0x08]
477 ; X86-NEXT: vmovdqa64 %zmm1, %zmm0 # encoding: [0x62,0xf1,0xfd,0x48,0x6f,0xc1]
478 ; X86-NEXT: retl # encoding: [0xc3]
479 ;
480 ; X64-LABEL: test_mask_packus_epi32_rmk_512:
481 ; X64: # %bb.0:
482 ; X64-NEXT: kmovd %esi, %k1 # encoding: [0xc5,0xfb,0x92,0xce]
483 ; X64-NEXT: vpackusdw (%rdi), %zmm0, %zmm1 {%k1} # encoding: [0x62,0xf2,0x7d,0x49,0x2b,0x0f]
484 ; X64-NEXT: vmovdqa64 %zmm1, %zmm0 # encoding: [0x62,0xf1,0xfd,0x48,0x6f,0xc1]
485 ; X64-NEXT: retq # encoding: [0xc3]
486 %b = load <16 x i32>, <16 x i32>* %ptr_b
487 %1 = call <32 x i16> @llvm.x86.avx512.packusdw.512(<16 x i32> %a, <16 x i32> %b)
488 %2 = bitcast i32 %mask to <32 x i1>
489 %3 = select <32 x i1> %2, <32 x i16> %1, <32 x i16> %passThru
490 ret <32 x i16> %3
491 }
493 define <32 x i16> @test_mask_packus_epi32_rmkz_512(<16 x i32> %a, <16 x i32>* %ptr_b, i32 %mask) {
494 ; X86-LABEL: test_mask_packus_epi32_rmkz_512:
495 ; X86: # %bb.0:
496 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x04]
497 ; X86-NEXT: kmovd {{[0-9]+}}(%esp), %k1 # encoding: [0xc4,0xe1,0xf9,0x90,0x4c,0x24,0x08]
498 ; X86-NEXT: vpackusdw (%eax), %zmm0, %zmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7d,0xc9,0x2b,0x00]
499 ; X86-NEXT: retl # encoding: [0xc3]
500 ;
501 ; X64-LABEL: test_mask_packus_epi32_rmkz_512:
502 ; X64: # %bb.0:
503 ; X64-NEXT: kmovd %esi, %k1 # encoding: [0xc5,0xfb,0x92,0xce]
504 ; X64-NEXT: vpackusdw (%rdi), %zmm0, %zmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7d,0xc9,0x2b,0x07]
505 ; X64-NEXT: retq # encoding: [0xc3]
506 %b = load <16 x i32>, <16 x i32>* %ptr_b
507 %1 = call <32 x i16> @llvm.x86.avx512.packusdw.512(<16 x i32> %a, <16 x i32> %b)
508 %2 = bitcast i32 %mask to <32 x i1>
509 %3 = select <32 x i1> %2, <32 x i16> %1, <32 x i16> zeroinitializer
510 ret <32 x i16> %3
511 }
513 define <32 x i16> @test_mask_packus_epi32_rmb_512(<16 x i32> %a, i32* %ptr_b) {
514 ; X86-LABEL: test_mask_packus_epi32_rmb_512:
515 ; X86: # %bb.0:
516 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x04]
517 ; X86-NEXT: vpackusdw (%eax){1to16}, %zmm0, %zmm0 # encoding: [0x62,0xf2,0x7d,0x58,0x2b,0x00]
518 ; X86-NEXT: retl # encoding: [0xc3]
519 ;
520 ; X64-LABEL: test_mask_packus_epi32_rmb_512:
521 ; X64: # %bb.0:
522 ; X64-NEXT: vpackusdw (%rdi){1to16}, %zmm0, %zmm0 # encoding: [0x62,0xf2,0x7d,0x58,0x2b,0x07]
523 ; X64-NEXT: retq # encoding: [0xc3]
524 %q = load i32, i32* %ptr_b
525 %vecinit.i = insertelement <16 x i32> undef, i32 %q, i32 0
526 %b = shufflevector <16 x i32> %vecinit.i, <16 x i32> undef, <16 x i32> zeroinitializer
527 %1 = call <32 x i16> @llvm.x86.avx512.packusdw.512(<16 x i32> %a, <16 x i32> %b)
528 ret <32 x i16> %1
529 }
531 define <32 x i16> @test_mask_packus_epi32_rmbk_512(<16 x i32> %a, i32* %ptr_b, <32 x i16> %passThru, i32 %mask) {
532 ; X86-LABEL: test_mask_packus_epi32_rmbk_512:
533 ; X86: # %bb.0:
534 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x04]
535 ; X86-NEXT: kmovd {{[0-9]+}}(%esp), %k1 # encoding: [0xc4,0xe1,0xf9,0x90,0x4c,0x24,0x08]
536 ; X86-NEXT: vpackusdw (%eax){1to16}, %zmm0, %zmm1 {%k1} # encoding: [0x62,0xf2,0x7d,0x59,0x2b,0x08]
537 ; X86-NEXT: vmovdqa64 %zmm1, %zmm0 # encoding: [0x62,0xf1,0xfd,0x48,0x6f,0xc1]
538 ; X86-NEXT: retl # encoding: [0xc3]
539 ;
540 ; X64-LABEL: test_mask_packus_epi32_rmbk_512:
541 ; X64: # %bb.0:
542 ; X64-NEXT: kmovd %esi, %k1 # encoding: [0xc5,0xfb,0x92,0xce]
543 ; X64-NEXT: vpackusdw (%rdi){1to16}, %zmm0, %zmm1 {%k1} # encoding: [0x62,0xf2,0x7d,0x59,0x2b,0x0f]
544 ; X64-NEXT: vmovdqa64 %zmm1, %zmm0 # encoding: [0x62,0xf1,0xfd,0x48,0x6f,0xc1]
545 ; X64-NEXT: retq # encoding: [0xc3]
546 %q = load i32, i32* %ptr_b
547 %vecinit.i = insertelement <16 x i32> undef, i32 %q, i32 0
548 %b = shufflevector <16 x i32> %vecinit.i, <16 x i32> undef, <16 x i32> zeroinitializer
549 %1 = call <32 x i16> @llvm.x86.avx512.packusdw.512(<16 x i32> %a, <16 x i32> %b)
550 %2 = bitcast i32 %mask to <32 x i1>
551 %3 = select <32 x i1> %2, <32 x i16> %1, <32 x i16> %passThru
552 ret <32 x i16> %3
553 }
555 define <32 x i16> @test_mask_packus_epi32_rmbkz_512(<16 x i32> %a, i32* %ptr_b, i32 %mask) {
556 ; X86-LABEL: test_mask_packus_epi32_rmbkz_512:
557 ; X86: # %bb.0:
558 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x04]
559 ; X86-NEXT: kmovd {{[0-9]+}}(%esp), %k1 # encoding: [0xc4,0xe1,0xf9,0x90,0x4c,0x24,0x08]
560 ; X86-NEXT: vpackusdw (%eax){1to16}, %zmm0, %zmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7d,0xd9,0x2b,0x00]
561 ; X86-NEXT: retl # encoding: [0xc3]
562 ;
563 ; X64-LABEL: test_mask_packus_epi32_rmbkz_512:
564 ; X64: # %bb.0:
565 ; X64-NEXT: kmovd %esi, %k1 # encoding: [0xc5,0xfb,0x92,0xce]
566 ; X64-NEXT: vpackusdw (%rdi){1to16}, %zmm0, %zmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7d,0xd9,0x2b,0x07]
567 ; X64-NEXT: retq # encoding: [0xc3]
568 %q = load i32, i32* %ptr_b
569 %vecinit.i = insertelement <16 x i32> undef, i32 %q, i32 0
570 %b = shufflevector <16 x i32> %vecinit.i, <16 x i32> undef, <16 x i32> zeroinitializer
571 %1 = call <32 x i16> @llvm.x86.avx512.packusdw.512(<16 x i32> %a, <16 x i32> %b)
572 %2 = bitcast i32 %mask to <32 x i1>
573 %3 = select <32 x i1> %2, <32 x i16> %1, <32 x i16> zeroinitializer
574 ret <32 x i16> %3
575 }
577 declare <32 x i16> @llvm.x86.avx512.packusdw.512(<16 x i32>, <16 x i32>)
579 define <64 x i8> @test_mask_packus_epi16_rr_512(<32 x i16> %a, <32 x i16> %b) {
580 ; CHECK-LABEL: test_mask_packus_epi16_rr_512:
581 ; CHECK: # %bb.0:
582 ; CHECK-NEXT: vpackuswb %zmm1, %zmm0, %zmm0 # encoding: [0x62,0xf1,0x7d,0x48,0x67,0xc1]
583 ; CHECK-NEXT: ret{{[l|q]}} # encoding: [0xc3]
584 %1 = call <64 x i8> @llvm.x86.avx512.packuswb.512(<32 x i16> %a, <32 x i16> %b)
585 ret <64 x i8> %1
586 }
588 define <64 x i8> @test_mask_packus_epi16_rrk_512(<32 x i16> %a, <32 x i16> %b, <64 x i8> %passThru, i64 %mask) {
589 ; X86-LABEL: test_mask_packus_epi16_rrk_512:
590 ; X86: # %bb.0:
591 ; X86-NEXT: kmovq {{[0-9]+}}(%esp), %k1 # encoding: [0xc4,0xe1,0xf8,0x90,0x4c,0x24,0x04]
592 ; X86-NEXT: vpackuswb %zmm1, %zmm0, %zmm2 {%k1} # encoding: [0x62,0xf1,0x7d,0x49,0x67,0xd1]
593 ; X86-NEXT: vmovdqa64 %zmm2, %zmm0 # encoding: [0x62,0xf1,0xfd,0x48,0x6f,0xc2]
594 ; X86-NEXT: retl # encoding: [0xc3]
595 ;
596 ; X64-LABEL: test_mask_packus_epi16_rrk_512:
597 ; X64: # %bb.0:
598 ; X64-NEXT: kmovq %rdi, %k1 # encoding: [0xc4,0xe1,0xfb,0x92,0xcf]
599 ; X64-NEXT: vpackuswb %zmm1, %zmm0, %zmm2 {%k1} # encoding: [0x62,0xf1,0x7d,0x49,0x67,0xd1]
600 ; X64-NEXT: vmovdqa64 %zmm2, %zmm0 # encoding: [0x62,0xf1,0xfd,0x48,0x6f,0xc2]
601 ; X64-NEXT: retq # encoding: [0xc3]
602 %1 = call <64 x i8> @llvm.x86.avx512.packuswb.512(<32 x i16> %a, <32 x i16> %b)
603 %2 = bitcast i64 %mask to <64 x i1>
604 %3 = select <64 x i1> %2, <64 x i8> %1, <64 x i8> %passThru
605 ret <64 x i8> %3
606 }
608 define <64 x i8> @test_mask_packus_epi16_rrkz_512(<32 x i16> %a, <32 x i16> %b, i64 %mask) {
609 ; X86-LABEL: test_mask_packus_epi16_rrkz_512:
610 ; X86: # %bb.0:
611 ; X86-NEXT: kmovq {{[0-9]+}}(%esp), %k1 # encoding: [0xc4,0xe1,0xf8,0x90,0x4c,0x24,0x04]
612 ; X86-NEXT: vpackuswb %zmm1, %zmm0, %zmm0 {%k1} {z} # encoding: [0x62,0xf1,0x7d,0xc9,0x67,0xc1]
613 ; X86-NEXT: retl # encoding: [0xc3]
614 ;
615 ; X64-LABEL: test_mask_packus_epi16_rrkz_512:
616 ; X64: # %bb.0:
617 ; X64-NEXT: kmovq %rdi, %k1 # encoding: [0xc4,0xe1,0xfb,0x92,0xcf]
618 ; X64-NEXT: vpackuswb %zmm1, %zmm0, %zmm0 {%k1} {z} # encoding: [0x62,0xf1,0x7d,0xc9,0x67,0xc1]
619 ; X64-NEXT: retq # encoding: [0xc3]
620 %1 = call <64 x i8> @llvm.x86.avx512.packuswb.512(<32 x i16> %a, <32 x i16> %b)
621 %2 = bitcast i64 %mask to <64 x i1>
622 %3 = select <64 x i1> %2, <64 x i8> %1, <64 x i8> zeroinitializer
623 ret <64 x i8> %3
624 }
626 define <64 x i8> @test_mask_packus_epi16_rm_512(<32 x i16> %a, <32 x i16>* %ptr_b) {
627 ; X86-LABEL: test_mask_packus_epi16_rm_512:
628 ; X86: # %bb.0:
629 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x04]
630 ; X86-NEXT: vpackuswb (%eax), %zmm0, %zmm0 # encoding: [0x62,0xf1,0x7d,0x48,0x67,0x00]
631 ; X86-NEXT: retl # encoding: [0xc3]
632 ;
633 ; X64-LABEL: test_mask_packus_epi16_rm_512:
634 ; X64: # %bb.0:
635 ; X64-NEXT: vpackuswb (%rdi), %zmm0, %zmm0 # encoding: [0x62,0xf1,0x7d,0x48,0x67,0x07]
636 ; X64-NEXT: retq # encoding: [0xc3]
637 %b = load <32 x i16>, <32 x i16>* %ptr_b
638 %1 = call <64 x i8> @llvm.x86.avx512.packuswb.512(<32 x i16> %a, <32 x i16> %b)
639 ret <64 x i8> %1
640 }
642 define <64 x i8> @test_mask_packus_epi16_rmk_512(<32 x i16> %a, <32 x i16>* %ptr_b, <64 x i8> %passThru, i64 %mask) {
643 ; X86-LABEL: test_mask_packus_epi16_rmk_512:
644 ; X86: # %bb.0:
645 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x04]
646 ; X86-NEXT: kmovq {{[0-9]+}}(%esp), %k1 # encoding: [0xc4,0xe1,0xf8,0x90,0x4c,0x24,0x08]
647 ; X86-NEXT: vpackuswb (%eax), %zmm0, %zmm1 {%k1} # encoding: [0x62,0xf1,0x7d,0x49,0x67,0x08]
648 ; X86-NEXT: vmovdqa64 %zmm1, %zmm0 # encoding: [0x62,0xf1,0xfd,0x48,0x6f,0xc1]
649 ; X86-NEXT: retl # encoding: [0xc3]
650 ;
651 ; X64-LABEL: test_mask_packus_epi16_rmk_512:
652 ; X64: # %bb.0:
653 ; X64-NEXT: kmovq %rsi, %k1 # encoding: [0xc4,0xe1,0xfb,0x92,0xce]
654 ; X64-NEXT: vpackuswb (%rdi), %zmm0, %zmm1 {%k1} # encoding: [0x62,0xf1,0x7d,0x49,0x67,0x0f]
655 ; X64-NEXT: vmovdqa64 %zmm1, %zmm0 # encoding: [0x62,0xf1,0xfd,0x48,0x6f,0xc1]
656 ; X64-NEXT: retq # encoding: [0xc3]
657 %b = load <32 x i16>, <32 x i16>* %ptr_b
658 %1 = call <64 x i8> @llvm.x86.avx512.packuswb.512(<32 x i16> %a, <32 x i16> %b)
659 %2 = bitcast i64 %mask to <64 x i1>
660 %3 = select <64 x i1> %2, <64 x i8> %1, <64 x i8> %passThru
661 ret <64 x i8> %3
662 }
664 define <64 x i8> @test_mask_packus_epi16_rmkz_512(<32 x i16> %a, <32 x i16>* %ptr_b, i64 %mask) {
665 ; X86-LABEL: test_mask_packus_epi16_rmkz_512:
666 ; X86: # %bb.0:
667 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x04]
668 ; X86-NEXT: kmovq {{[0-9]+}}(%esp), %k1 # encoding: [0xc4,0xe1,0xf8,0x90,0x4c,0x24,0x08]
669 ; X86-NEXT: vpackuswb (%eax), %zmm0, %zmm0 {%k1} {z} # encoding: [0x62,0xf1,0x7d,0xc9,0x67,0x00]
670 ; X86-NEXT: retl # encoding: [0xc3]
671 ;
672 ; X64-LABEL: test_mask_packus_epi16_rmkz_512:
673 ; X64: # %bb.0:
674 ; X64-NEXT: kmovq %rsi, %k1 # encoding: [0xc4,0xe1,0xfb,0x92,0xce]
675 ; X64-NEXT: vpackuswb (%rdi), %zmm0, %zmm0 {%k1} {z} # encoding: [0x62,0xf1,0x7d,0xc9,0x67,0x07]
676 ; X64-NEXT: retq # encoding: [0xc3]
677 %b = load <32 x i16>, <32 x i16>* %ptr_b
678 %1 = call <64 x i8> @llvm.x86.avx512.packuswb.512(<32 x i16> %a, <32 x i16> %b)
679 %2 = bitcast i64 %mask to <64 x i1>
680 %3 = select <64 x i1> %2, <64 x i8> %1, <64 x i8> zeroinitializer
681 ret <64 x i8> %3
682 }
684 declare <64 x i8> @llvm.x86.avx512.packuswb.512(<32 x i16>, <32 x i16>)
686 define <32 x i16>@test_int_x86_avx512_mask_vpermt2var_hi_512(<32 x i16> %x0, <32 x i16> %x1, <32 x i16> %x2, i32 %x3) {
687 ; X86-LABEL: test_int_x86_avx512_mask_vpermt2var_hi_512:
688 ; X86: # %bb.0:
689 ; X86-NEXT: vmovdqa64 %zmm1, %zmm3 # encoding: [0x62,0xf1,0xfd,0x48,0x6f,0xd9]
690 ; X86-NEXT: vpermt2w %zmm2, %zmm0, %zmm3 # encoding: [0x62,0xf2,0xfd,0x48,0x7d,0xda]
691 ; X86-NEXT: kmovd {{[0-9]+}}(%esp), %k1 # encoding: [0xc4,0xe1,0xf9,0x90,0x4c,0x24,0x04]
692 ; X86-NEXT: vpermt2w %zmm2, %zmm0, %zmm1 {%k1} # encoding: [0x62,0xf2,0xfd,0x49,0x7d,0xca]
693 ; X86-NEXT: vpaddw %zmm3, %zmm1, %zmm0 # encoding: [0x62,0xf1,0x75,0x48,0xfd,0xc3]
694 ; X86-NEXT: retl # encoding: [0xc3]
695 ;
696 ; X64-LABEL: test_int_x86_avx512_mask_vpermt2var_hi_512:
697 ; X64: # %bb.0:
698 ; X64-NEXT: vmovdqa64 %zmm1, %zmm3 # encoding: [0x62,0xf1,0xfd,0x48,0x6f,0xd9]
699 ; X64-NEXT: vpermt2w %zmm2, %zmm0, %zmm3 # encoding: [0x62,0xf2,0xfd,0x48,0x7d,0xda]
700 ; X64-NEXT: kmovd %edi, %k1 # encoding: [0xc5,0xfb,0x92,0xcf]
701 ; X64-NEXT: vpermt2w %zmm2, %zmm0, %zmm1 {%k1} # encoding: [0x62,0xf2,0xfd,0x49,0x7d,0xca]
702 ; X64-NEXT: vpaddw %zmm3, %zmm1, %zmm0 # encoding: [0x62,0xf1,0x75,0x48,0xfd,0xc3]
703 ; X64-NEXT: retq # encoding: [0xc3]
704 %1 = call <32 x i16> @llvm.x86.avx512.vpermi2var.hi.512(<32 x i16> %x1, <32 x i16> %x0, <32 x i16> %x2)
705 %2 = bitcast i32 %x3 to <32 x i1>
706 %3 = select <32 x i1> %2, <32 x i16> %1, <32 x i16> %x1
707 %4 = call <32 x i16> @llvm.x86.avx512.vpermi2var.hi.512(<32 x i16> %x1, <32 x i16> %x0, <32 x i16> %x2)
708 %res2 = add <32 x i16> %3, %4
709 ret <32 x i16> %res2
710 }
712 define <32 x i16>@test_int_x86_avx512_maskz_vpermt2var_hi_512(<32 x i16> %x0, <32 x i16> %x1, <32 x i16> %x2, i32 %x3) {
713 ; X86-LABEL: test_int_x86_avx512_maskz_vpermt2var_hi_512:
714 ; X86: # %bb.0:
715 ; X86-NEXT: vmovdqa64 %zmm1, %zmm3 # encoding: [0x62,0xf1,0xfd,0x48,0x6f,0xd9]
716 ; X86-NEXT: vpermt2w %zmm2, %zmm0, %zmm3 # encoding: [0x62,0xf2,0xfd,0x48,0x7d,0xda]
717 ; X86-NEXT: kmovd {{[0-9]+}}(%esp), %k1 # encoding: [0xc4,0xe1,0xf9,0x90,0x4c,0x24,0x04]
718 ; X86-NEXT: vpermt2w %zmm2, %zmm0, %zmm1 {%k1} {z} # encoding: [0x62,0xf2,0xfd,0xc9,0x7d,0xca]
719 ; X86-NEXT: vpaddw %zmm3, %zmm1, %zmm0 # encoding: [0x62,0xf1,0x75,0x48,0xfd,0xc3]
720 ; X86-NEXT: retl # encoding: [0xc3]
721 ;
722 ; X64-LABEL: test_int_x86_avx512_maskz_vpermt2var_hi_512:
723 ; X64: # %bb.0:
724 ; X64-NEXT: vmovdqa64 %zmm1, %zmm3 # encoding: [0x62,0xf1,0xfd,0x48,0x6f,0xd9]
725 ; X64-NEXT: vpermt2w %zmm2, %zmm0, %zmm3 # encoding: [0x62,0xf2,0xfd,0x48,0x7d,0xda]
726 ; X64-NEXT: kmovd %edi, %k1 # encoding: [0xc5,0xfb,0x92,0xcf]
727 ; X64-NEXT: vpermt2w %zmm2, %zmm0, %zmm1 {%k1} {z} # encoding: [0x62,0xf2,0xfd,0xc9,0x7d,0xca]
728 ; X64-NEXT: vpaddw %zmm3, %zmm1, %zmm0 # encoding: [0x62,0xf1,0x75,0x48,0xfd,0xc3]
729 ; X64-NEXT: retq # encoding: [0xc3]
730 %1 = call <32 x i16> @llvm.x86.avx512.vpermi2var.hi.512(<32 x i16> %x1, <32 x i16> %x0, <32 x i16> %x2)
731 %2 = bitcast i32 %x3 to <32 x i1>
732 %3 = select <32 x i1> %2, <32 x i16> %1, <32 x i16> zeroinitializer
733 %4 = call <32 x i16> @llvm.x86.avx512.vpermi2var.hi.512(<32 x i16> %x1, <32 x i16> %x0, <32 x i16> %x2)
734 %res2 = add <32 x i16> %3, %4
735 ret <32 x i16> %res2
736 }
738 declare <32 x i16> @llvm.x86.avx512.vpermi2var.hi.512(<32 x i16>, <32 x i16>, <32 x i16>)
740 define <32 x i16>@test_int_x86_avx512_mask_vpermi2var_hi_512(<32 x i16> %x0, <32 x i16> %x1, <32 x i16> %x2, i32 %x3) {
741 ; X86-LABEL: test_int_x86_avx512_mask_vpermi2var_hi_512:
742 ; X86: # %bb.0:
743 ; X86-NEXT: vmovdqa64 %zmm0, %zmm3 # encoding: [0x62,0xf1,0xfd,0x48,0x6f,0xd8]
744 ; X86-NEXT: vpermt2w %zmm2, %zmm1, %zmm3 # encoding: [0x62,0xf2,0xf5,0x48,0x7d,0xda]
745 ; X86-NEXT: kmovd {{[0-9]+}}(%esp), %k1 # encoding: [0xc4,0xe1,0xf9,0x90,0x4c,0x24,0x04]
746 ; X86-NEXT: vpermi2w %zmm2, %zmm0, %zmm1 {%k1} # encoding: [0x62,0xf2,0xfd,0x49,0x75,0xca]
747 ; X86-NEXT: vpaddw %zmm3, %zmm1, %zmm0 # encoding: [0x62,0xf1,0x75,0x48,0xfd,0xc3]
748 ; X86-NEXT: retl # encoding: [0xc3]
749 ;
750 ; X64-LABEL: test_int_x86_avx512_mask_vpermi2var_hi_512:
751 ; X64: # %bb.0:
752 ; X64-NEXT: vmovdqa64 %zmm0, %zmm3 # encoding: [0x62,0xf1,0xfd,0x48,0x6f,0xd8]
753 ; X64-NEXT: vpermt2w %zmm2, %zmm1, %zmm3 # encoding: [0x62,0xf2,0xf5,0x48,0x7d,0xda]
754 ; X64-NEXT: kmovd %edi, %k1 # encoding: [0xc5,0xfb,0x92,0xcf]
755 ; X64-NEXT: vpermi2w %zmm2, %zmm0, %zmm1 {%k1} # encoding: [0x62,0xf2,0xfd,0x49,0x75,0xca]
756 ; X64-NEXT: vpaddw %zmm3, %zmm1, %zmm0 # encoding: [0x62,0xf1,0x75,0x48,0xfd,0xc3]
757 ; X64-NEXT: retq # encoding: [0xc3]
758 %1 = call <32 x i16> @llvm.x86.avx512.vpermi2var.hi.512(<32 x i16> %x0, <32 x i16> %x1, <32 x i16> %x2)
759 %2 = bitcast i32 %x3 to <32 x i1>
760 %3 = select <32 x i1> %2, <32 x i16> %1, <32 x i16> %x1
761 %4 = call <32 x i16> @llvm.x86.avx512.vpermi2var.hi.512(<32 x i16> %x0, <32 x i16> %x1, <32 x i16> %x2)
762 %res2 = add <32 x i16> %3, %4
763 ret <32 x i16> %res2
764 }
766 declare <64 x i8> @llvm.x86.avx512.pshuf.b.512(<64 x i8>, <64 x i8>)
768 define <64 x i8>@test_int_x86_avx512_pshuf_b_512(<64 x i8> %x0, <64 x i8> %x1) {
769 ; CHECK-LABEL: test_int_x86_avx512_pshuf_b_512:
770 ; CHECK: # %bb.0:
771 ; CHECK-NEXT: vpshufb %zmm1, %zmm0, %zmm0 # encoding: [0x62,0xf2,0x7d,0x48,0x00,0xc1]
772 ; CHECK-NEXT: ret{{[l|q]}} # encoding: [0xc3]
773 %res = call <64 x i8> @llvm.x86.avx512.pshuf.b.512(<64 x i8> %x0, <64 x i8> %x1)
774 ret <64 x i8> %res
775 }
777 define <64 x i8>@test_int_x86_avx512_pshuf_b_512_mask(<64 x i8> %x0, <64 x i8> %x1, <64 x i8> %x2, i64 %mask) {
778 ; X86-LABEL: test_int_x86_avx512_pshuf_b_512_mask:
779 ; X86: # %bb.0:
780 ; X86-NEXT: kmovq {{[0-9]+}}(%esp), %k1 # encoding: [0xc4,0xe1,0xf8,0x90,0x4c,0x24,0x04]
781 ; X86-NEXT: vpshufb %zmm1, %zmm0, %zmm2 {%k1} # encoding: [0x62,0xf2,0x7d,0x49,0x00,0xd1]
782 ; X86-NEXT: vmovdqa64 %zmm2, %zmm0 # encoding: [0x62,0xf1,0xfd,0x48,0x6f,0xc2]
783 ; X86-NEXT: retl # encoding: [0xc3]
784 ;
785 ; X64-LABEL: test_int_x86_avx512_pshuf_b_512_mask:
786 ; X64: # %bb.0:
787 ; X64-NEXT: kmovq %rdi, %k1 # encoding: [0xc4,0xe1,0xfb,0x92,0xcf]
788 ; X64-NEXT: vpshufb %zmm1, %zmm0, %zmm2 {%k1} # encoding: [0x62,0xf2,0x7d,0x49,0x00,0xd1]
789 ; X64-NEXT: vmovdqa64 %zmm2, %zmm0 # encoding: [0x62,0xf1,0xfd,0x48,0x6f,0xc2]
790 ; X64-NEXT: retq # encoding: [0xc3]
791 %res = call <64 x i8> @llvm.x86.avx512.pshuf.b.512(<64 x i8> %x0, <64 x i8> %x1)
792 %mask.cast = bitcast i64 %mask to <64 x i1>
793 %res2 = select <64 x i1> %mask.cast, <64 x i8> %res, <64 x i8> %x2
794 ret <64 x i8> %res2
795 }
797 define <64 x i8>@test_int_x86_avx512_pshuf_b_512_maskz(<64 x i8> %x0, <64 x i8> %x1, i64 %mask) {
798 ; X86-LABEL: test_int_x86_avx512_pshuf_b_512_maskz:
799 ; X86: # %bb.0:
800 ; X86-NEXT: kmovq {{[0-9]+}}(%esp), %k1 # encoding: [0xc4,0xe1,0xf8,0x90,0x4c,0x24,0x04]
801 ; X86-NEXT: vpshufb %zmm1, %zmm0, %zmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7d,0xc9,0x00,0xc1]
802 ; X86-NEXT: retl # encoding: [0xc3]
803 ;
804 ; X64-LABEL: test_int_x86_avx512_pshuf_b_512_maskz:
805 ; X64: # %bb.0:
806 ; X64-NEXT: kmovq %rdi, %k1 # encoding: [0xc4,0xe1,0xfb,0x92,0xcf]
807 ; X64-NEXT: vpshufb %zmm1, %zmm0, %zmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7d,0xc9,0x00,0xc1]
808 ; X64-NEXT: retq # encoding: [0xc3]
809 %res = call <64 x i8> @llvm.x86.avx512.pshuf.b.512(<64 x i8> %x0, <64 x i8> %x1)
810 %mask.cast = bitcast i64 %mask to <64 x i1>
811 %res2 = select <64 x i1> %mask.cast, <64 x i8> %res, <64 x i8> zeroinitializer
812 ret <64 x i8> %res2
813 }
815 declare <32 x i16> @llvm.x86.avx512.pmulhu.w.512(<32 x i16>, <32 x i16>)
817 define <32 x i16> @test_int_x86_avx512_mask_pmulhu_w_512(<32 x i16> %x0, <32 x i16> %x1, <32 x i16> %x2, i32 %x3) {
818 ; X86-LABEL: test_int_x86_avx512_mask_pmulhu_w_512:
819 ; X86: # %bb.0:
820 ; X86-NEXT: vpmulhuw %zmm1, %zmm0, %zmm3 # encoding: [0x62,0xf1,0x7d,0x48,0xe4,0xd9]
821 ; X86-NEXT: kmovd {{[0-9]+}}(%esp), %k1 # encoding: [0xc4,0xe1,0xf9,0x90,0x4c,0x24,0x04]
822 ; X86-NEXT: vpmulhuw %zmm1, %zmm0, %zmm2 {%k1} # encoding: [0x62,0xf1,0x7d,0x49,0xe4,0xd1]
823 ; X86-NEXT: vpaddw %zmm3, %zmm2, %zmm0 # encoding: [0x62,0xf1,0x6d,0x48,0xfd,0xc3]
824 ; X86-NEXT: retl # encoding: [0xc3]
825 ;
826 ; X64-LABEL: test_int_x86_avx512_mask_pmulhu_w_512:
827 ; X64: # %bb.0:
828 ; X64-NEXT: vpmulhuw %zmm1, %zmm0, %zmm3 # encoding: [0x62,0xf1,0x7d,0x48,0xe4,0xd9]
829 ; X64-NEXT: kmovd %edi, %k1 # encoding: [0xc5,0xfb,0x92,0xcf]
830 ; X64-NEXT: vpmulhuw %zmm1, %zmm0, %zmm2 {%k1} # encoding: [0x62,0xf1,0x7d,0x49,0xe4,0xd1]
831 ; X64-NEXT: vpaddw %zmm3, %zmm2, %zmm0 # encoding: [0x62,0xf1,0x6d,0x48,0xfd,0xc3]
832 ; X64-NEXT: retq # encoding: [0xc3]
833 %1 = call <32 x i16> @llvm.x86.avx512.pmulhu.w.512(<32 x i16> %x0, <32 x i16> %x1)
834 %2 = bitcast i32 %x3 to <32 x i1>
835 %3 = select <32 x i1> %2, <32 x i16> %1, <32 x i16> %x2
836 %4 = call <32 x i16> @llvm.x86.avx512.pmulhu.w.512(<32 x i16> %x0, <32 x i16> %x1)
837 %res2 = add <32 x i16> %3, %4
838 ret <32 x i16> %res2
839 }
841 declare <32 x i16> @llvm.x86.avx512.pmulh.w.512(<32 x i16>, <32 x i16>)
843 define <32 x i16> @test_int_x86_avx512_mask_pmulh_w_512(<32 x i16> %x0, <32 x i16> %x1, <32 x i16> %x2, i32 %x3) {
844 ; X86-LABEL: test_int_x86_avx512_mask_pmulh_w_512:
845 ; X86: # %bb.0:
846 ; X86-NEXT: vpmulhw %zmm1, %zmm0, %zmm3 # encoding: [0x62,0xf1,0x7d,0x48,0xe5,0xd9]
847 ; X86-NEXT: kmovd {{[0-9]+}}(%esp), %k1 # encoding: [0xc4,0xe1,0xf9,0x90,0x4c,0x24,0x04]
848 ; X86-NEXT: vpmulhw %zmm1, %zmm0, %zmm2 {%k1} # encoding: [0x62,0xf1,0x7d,0x49,0xe5,0xd1]
849 ; X86-NEXT: vpaddw %zmm3, %zmm2, %zmm0 # encoding: [0x62,0xf1,0x6d,0x48,0xfd,0xc3]
850 ; X86-NEXT: retl # encoding: [0xc3]
851 ;
852 ; X64-LABEL: test_int_x86_avx512_mask_pmulh_w_512:
853 ; X64: # %bb.0:
854 ; X64-NEXT: vpmulhw %zmm1, %zmm0, %zmm3 # encoding: [0x62,0xf1,0x7d,0x48,0xe5,0xd9]
855 ; X64-NEXT: kmovd %edi, %k1 # encoding: [0xc5,0xfb,0x92,0xcf]
856 ; X64-NEXT: vpmulhw %zmm1, %zmm0, %zmm2 {%k1} # encoding: [0x62,0xf1,0x7d,0x49,0xe5,0xd1]
857 ; X64-NEXT: vpaddw %zmm3, %zmm2, %zmm0 # encoding: [0x62,0xf1,0x6d,0x48,0xfd,0xc3]
858 ; X64-NEXT: retq # encoding: [0xc3]
859 %1 = call <32 x i16> @llvm.x86.avx512.pmulh.w.512(<32 x i16> %x0, <32 x i16> %x1)
860 %2 = bitcast i32 %x3 to <32 x i1>
861 %3 = select <32 x i1> %2, <32 x i16> %1, <32 x i16> %x2
862 %4 = call <32 x i16> @llvm.x86.avx512.pmulh.w.512(<32 x i16> %x0, <32 x i16> %x1)
863 %res2 = add <32 x i16> %3, %4
864 ret <32 x i16> %res2
865 }
867 declare <32 x i16> @llvm.x86.avx512.pmul.hr.sw.512(<32 x i16>, <32 x i16>)
869 define <32 x i16> @test_int_x86_avx512_mask_pmulhr_sw_512(<32 x i16> %x0, <32 x i16> %x1, <32 x i16> %x2, i32 %x3) {
870 ; X86-LABEL: test_int_x86_avx512_mask_pmulhr_sw_512:
871 ; X86: # %bb.0:
872 ; X86-NEXT: vpmulhrsw %zmm1, %zmm0, %zmm3 # encoding: [0x62,0xf2,0x7d,0x48,0x0b,0xd9]
873 ; X86-NEXT: kmovd {{[0-9]+}}(%esp), %k1 # encoding: [0xc4,0xe1,0xf9,0x90,0x4c,0x24,0x04]
874 ; X86-NEXT: vpmulhrsw %zmm1, %zmm0, %zmm2 {%k1} # encoding: [0x62,0xf2,0x7d,0x49,0x0b,0xd1]
875 ; X86-NEXT: vpaddw %zmm3, %zmm2, %zmm0 # encoding: [0x62,0xf1,0x6d,0x48,0xfd,0xc3]
876 ; X86-NEXT: retl # encoding: [0xc3]
877 ;
878 ; X64-LABEL: test_int_x86_avx512_mask_pmulhr_sw_512:
879 ; X64: # %bb.0:
880 ; X64-NEXT: vpmulhrsw %zmm1, %zmm0, %zmm3 # encoding: [0x62,0xf2,0x7d,0x48,0x0b,0xd9]
881 ; X64-NEXT: kmovd %edi, %k1 # encoding: [0xc5,0xfb,0x92,0xcf]
882 ; X64-NEXT: vpmulhrsw %zmm1, %zmm0, %zmm2 {%k1} # encoding: [0x62,0xf2,0x7d,0x49,0x0b,0xd1]
883 ; X64-NEXT: vpaddw %zmm3, %zmm2, %zmm0 # encoding: [0x62,0xf1,0x6d,0x48,0xfd,0xc3]
884 ; X64-NEXT: retq # encoding: [0xc3]
885 %1 = call <32 x i16> @llvm.x86.avx512.pmul.hr.sw.512(<32 x i16> %x0, <32 x i16> %x1)
886 %2 = bitcast i32 %x3 to <32 x i1>
887 %3 = select <32 x i1> %2, <32 x i16> %1, <32 x i16> %x2
888 %4 = call <32 x i16> @llvm.x86.avx512.pmul.hr.sw.512(<32 x i16> %x0, <32 x i16> %x1)
889 %res2 = add <32 x i16> %3, %4
890 ret <32 x i16> %res2
891 }
893 define <32 x i8>@test_int_x86_avx512_mask_pmov_wb_512(<32 x i16> %x0, <32 x i8> %x1, i32 %x2) {
894 ; X86-LABEL: test_int_x86_avx512_mask_pmov_wb_512:
895 ; X86: # %bb.0:
896 ; X86-NEXT: vpmovwb %zmm0, %ymm2 # encoding: [0x62,0xf2,0x7e,0x48,0x30,0xc2]
897 ; X86-NEXT: kmovd {{[0-9]+}}(%esp), %k1 # encoding: [0xc4,0xe1,0xf9,0x90,0x4c,0x24,0x04]
898 ; X86-NEXT: vpmovwb %zmm0, %ymm1 {%k1} # encoding: [0x62,0xf2,0x7e,0x49,0x30,0xc1]
899 ; X86-NEXT: vpmovwb %zmm0, %ymm0 {%k1} {z} # encoding: [0x62,0xf2,0x7e,0xc9,0x30,0xc0]
900 ; X86-NEXT: vpaddb %ymm0, %ymm1, %ymm0 # encoding: [0xc5,0xf5,0xfc,0xc0]
901 ; X86-NEXT: vpaddb %ymm0, %ymm2, %ymm0 # encoding: [0xc5,0xed,0xfc,0xc0]
902 ; X86-NEXT: retl # encoding: [0xc3]
903 ;
904 ; X64-LABEL: test_int_x86_avx512_mask_pmov_wb_512:
905 ; X64: # %bb.0:
906 ; X64-NEXT: vpmovwb %zmm0, %ymm2 # encoding: [0x62,0xf2,0x7e,0x48,0x30,0xc2]
907 ; X64-NEXT: kmovd %edi, %k1 # encoding: [0xc5,0xfb,0x92,0xcf]
908 ; X64-NEXT: vpmovwb %zmm0, %ymm1 {%k1} # encoding: [0x62,0xf2,0x7e,0x49,0x30,0xc1]
909 ; X64-NEXT: vpmovwb %zmm0, %ymm0 {%k1} {z} # encoding: [0x62,0xf2,0x7e,0xc9,0x30,0xc0]
910 ; X64-NEXT: vpaddb %ymm0, %ymm1, %ymm0 # encoding: [0xc5,0xf5,0xfc,0xc0]
911 ; X64-NEXT: vpaddb %ymm0, %ymm2, %ymm0 # encoding: [0xc5,0xed,0xfc,0xc0]
912 ; X64-NEXT: retq # encoding: [0xc3]
913 %1 = trunc <32 x i16> %x0 to <32 x i8>
914 %2 = trunc <32 x i16> %x0 to <32 x i8>
915 %3 = bitcast i32 %x2 to <32 x i1>
916 %4 = select <32 x i1> %3, <32 x i8> %2, <32 x i8> %x1
917 %5 = trunc <32 x i16> %x0 to <32 x i8>
918 %6 = bitcast i32 %x2 to <32 x i1>
919 %7 = select <32 x i1> %6, <32 x i8> %5, <32 x i8> zeroinitializer
920 %res3 = add <32 x i8> %1, %4
921 %res4 = add <32 x i8> %res3, %7
922 ret <32 x i8> %res4
923 }
925 declare void @llvm.x86.avx512.mask.pmov.wb.mem.512(i8* %ptr, <32 x i16>, i32)
927 define void @test_int_x86_avx512_mask_pmov_wb_mem_512(i8* %ptr, <32 x i16> %x1, i32 %x2) {
928 ; X86-LABEL: test_int_x86_avx512_mask_pmov_wb_mem_512:
929 ; X86: # %bb.0:
930 ; X86-NEXT: kmovd {{[0-9]+}}(%esp), %k1 # encoding: [0xc4,0xe1,0xf9,0x90,0x4c,0x24,0x08]
931 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x04]
932 ; X86-NEXT: vpmovwb %zmm0, (%eax) # encoding: [0x62,0xf2,0x7e,0x48,0x30,0x00]
933 ; X86-NEXT: vpmovwb %zmm0, (%eax) {%k1} # encoding: [0x62,0xf2,0x7e,0x49,0x30,0x00]
934 ; X86-NEXT: vzeroupper # encoding: [0xc5,0xf8,0x77]
935 ; X86-NEXT: retl # encoding: [0xc3]
936 ;
937 ; X64-LABEL: test_int_x86_avx512_mask_pmov_wb_mem_512:
938 ; X64: # %bb.0:
939 ; X64-NEXT: kmovd %esi, %k1 # encoding: [0xc5,0xfb,0x92,0xce]
940 ; X64-NEXT: vpmovwb %zmm0, (%rdi) # encoding: [0x62,0xf2,0x7e,0x48,0x30,0x07]
941 ; X64-NEXT: vpmovwb %zmm0, (%rdi) {%k1} # encoding: [0x62,0xf2,0x7e,0x49,0x30,0x07]
942 ; X64-NEXT: vzeroupper # encoding: [0xc5,0xf8,0x77]
943 ; X64-NEXT: retq # encoding: [0xc3]
944 call void @llvm.x86.avx512.mask.pmov.wb.mem.512(i8* %ptr, <32 x i16> %x1, i32 -1)
945 call void @llvm.x86.avx512.mask.pmov.wb.mem.512(i8* %ptr, <32 x i16> %x1, i32 %x2)
946 ret void
947 }
949 declare <32 x i8> @llvm.x86.avx512.mask.pmovs.wb.512(<32 x i16>, <32 x i8>, i32)
951 define <32 x i8>@test_int_x86_avx512_mask_pmovs_wb_512(<32 x i16> %x0, <32 x i8> %x1, i32 %x2) {
952 ; X86-LABEL: test_int_x86_avx512_mask_pmovs_wb_512:
953 ; X86: # %bb.0:
954 ; X86-NEXT: kmovd {{[0-9]+}}(%esp), %k1 # encoding: [0xc4,0xe1,0xf9,0x90,0x4c,0x24,0x04]
955 ; X86-NEXT: vpmovswb %zmm0, %ymm1 {%k1} # encoding: [0x62,0xf2,0x7e,0x49,0x20,0xc1]
956 ; X86-NEXT: vpmovswb %zmm0, %ymm2 {%k1} {z} # encoding: [0x62,0xf2,0x7e,0xc9,0x20,0xc2]
957 ; X86-NEXT: vpaddb %ymm2, %ymm1, %ymm1 # encoding: [0xc5,0xf5,0xfc,0xca]
958 ; X86-NEXT: vpmovswb %zmm0, %ymm0 # encoding: [0x62,0xf2,0x7e,0x48,0x20,0xc0]
959 ; X86-NEXT: vpaddb %ymm1, %ymm0, %ymm0 # encoding: [0xc5,0xfd,0xfc,0xc1]
960 ; X86-NEXT: retl # encoding: [0xc3]
961 ;
962 ; X64-LABEL: test_int_x86_avx512_mask_pmovs_wb_512:
963 ; X64: # %bb.0:
964 ; X64-NEXT: kmovd %edi, %k1 # encoding: [0xc5,0xfb,0x92,0xcf]
965 ; X64-NEXT: vpmovswb %zmm0, %ymm1 {%k1} # encoding: [0x62,0xf2,0x7e,0x49,0x20,0xc1]
966 ; X64-NEXT: vpmovswb %zmm0, %ymm2 {%k1} {z} # encoding: [0x62,0xf2,0x7e,0xc9,0x20,0xc2]
967 ; X64-NEXT: vpaddb %ymm2, %ymm1, %ymm1 # encoding: [0xc5,0xf5,0xfc,0xca]
968 ; X64-NEXT: vpmovswb %zmm0, %ymm0 # encoding: [0x62,0xf2,0x7e,0x48,0x20,0xc0]
969 ; X64-NEXT: vpaddb %ymm1, %ymm0, %ymm0 # encoding: [0xc5,0xfd,0xfc,0xc1]
970 ; X64-NEXT: retq # encoding: [0xc3]
971 %res0 = call <32 x i8> @llvm.x86.avx512.mask.pmovs.wb.512(<32 x i16> %x0, <32 x i8> %x1, i32 -1)
972 %res1 = call <32 x i8> @llvm.x86.avx512.mask.pmovs.wb.512(<32 x i16> %x0, <32 x i8> %x1, i32 %x2)
973 %res2 = call <32 x i8> @llvm.x86.avx512.mask.pmovs.wb.512(<32 x i16> %x0, <32 x i8> zeroinitializer, i32 %x2)
974 %res3 = add <32 x i8> %res0, %res1
975 %res4 = add <32 x i8> %res3, %res2
976 ret <32 x i8> %res4
977 }
979 declare void @llvm.x86.avx512.mask.pmovs.wb.mem.512(i8* %ptr, <32 x i16>, i32)
981 define void @test_int_x86_avx512_mask_pmovs_wb_mem_512(i8* %ptr, <32 x i16> %x1, i32 %x2) {
982 ; X86-LABEL: test_int_x86_avx512_mask_pmovs_wb_mem_512:
983 ; X86: # %bb.0:
984 ; X86-NEXT: kmovd {{[0-9]+}}(%esp), %k1 # encoding: [0xc4,0xe1,0xf9,0x90,0x4c,0x24,0x08]
985 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x04]
986 ; X86-NEXT: vpmovswb %zmm0, (%eax) # encoding: [0x62,0xf2,0x7e,0x48,0x20,0x00]
987 ; X86-NEXT: vpmovswb %zmm0, (%eax) {%k1} # encoding: [0x62,0xf2,0x7e,0x49,0x20,0x00]
988 ; X86-NEXT: vzeroupper # encoding: [0xc5,0xf8,0x77]
989 ; X86-NEXT: retl # encoding: [0xc3]
990 ;
991 ; X64-LABEL: test_int_x86_avx512_mask_pmovs_wb_mem_512:
992 ; X64: # %bb.0:
993 ; X64-NEXT: kmovd %esi, %k1 # encoding: [0xc5,0xfb,0x92,0xce]
994 ; X64-NEXT: vpmovswb %zmm0, (%rdi) # encoding: [0x62,0xf2,0x7e,0x48,0x20,0x07]
995 ; X64-NEXT: vpmovswb %zmm0, (%rdi) {%k1} # encoding: [0x62,0xf2,0x7e,0x49,0x20,0x07]
996 ; X64-NEXT: vzeroupper # encoding: [0xc5,0xf8,0x77]
997 ; X64-NEXT: retq # encoding: [0xc3]
998 call void @llvm.x86.avx512.mask.pmovs.wb.mem.512(i8* %ptr, <32 x i16> %x1, i32 -1)
999 call void @llvm.x86.avx512.mask.pmovs.wb.mem.512(i8* %ptr, <32 x i16> %x1, i32 %x2)
1000 ret void
1001 }
1003 declare <32 x i8> @llvm.x86.avx512.mask.pmovus.wb.512(<32 x i16>, <32 x i8>, i32)
1005 define <32 x i8>@test_int_x86_avx512_mask_pmovus_wb_512(<32 x i16> %x0, <32 x i8> %x1, i32 %x2) {
1006 ; X86-LABEL: test_int_x86_avx512_mask_pmovus_wb_512:
1007 ; X86: # %bb.0:
1008 ; X86-NEXT: kmovd {{[0-9]+}}(%esp), %k1 # encoding: [0xc4,0xe1,0xf9,0x90,0x4c,0x24,0x04]
1009 ; X86-NEXT: vpmovuswb %zmm0, %ymm1 {%k1} # encoding: [0x62,0xf2,0x7e,0x49,0x10,0xc1]
1010 ; X86-NEXT: vpmovuswb %zmm0, %ymm2 {%k1} {z} # encoding: [0x62,0xf2,0x7e,0xc9,0x10,0xc2]
1011 ; X86-NEXT: vpaddb %ymm2, %ymm1, %ymm1 # encoding: [0xc5,0xf5,0xfc,0xca]
1012 ; X86-NEXT: vpmovuswb %zmm0, %ymm0 # encoding: [0x62,0xf2,0x7e,0x48,0x10,0xc0]
1013 ; X86-NEXT: vpaddb %ymm1, %ymm0, %ymm0 # encoding: [0xc5,0xfd,0xfc,0xc1]
1014 ; X86-NEXT: retl # encoding: [0xc3]
1015 ;
1016 ; X64-LABEL: test_int_x86_avx512_mask_pmovus_wb_512:
1017 ; X64: # %bb.0:
1018 ; X64-NEXT: kmovd %edi, %k1 # encoding: [0xc5,0xfb,0x92,0xcf]
1019 ; X64-NEXT: vpmovuswb %zmm0, %ymm1 {%k1} # encoding: [0x62,0xf2,0x7e,0x49,0x10,0xc1]
1020 ; X64-NEXT: vpmovuswb %zmm0, %ymm2 {%k1} {z} # encoding: [0x62,0xf2,0x7e,0xc9,0x10,0xc2]
1021 ; X64-NEXT: vpaddb %ymm2, %ymm1, %ymm1 # encoding: [0xc5,0xf5,0xfc,0xca]
1022 ; X64-NEXT: vpmovuswb %zmm0, %ymm0 # encoding: [0x62,0xf2,0x7e,0x48,0x10,0xc0]
1023 ; X64-NEXT: vpaddb %ymm1, %ymm0, %ymm0 # encoding: [0xc5,0xfd,0xfc,0xc1]
1024 ; X64-NEXT: retq # encoding: [0xc3]
1025 %res0 = call <32 x i8> @llvm.x86.avx512.mask.pmovus.wb.512(<32 x i16> %x0, <32 x i8> %x1, i32 -1)
1026 %res1 = call <32 x i8> @llvm.x86.avx512.mask.pmovus.wb.512(<32 x i16> %x0, <32 x i8> %x1, i32 %x2)
1027 %res2 = call <32 x i8> @llvm.x86.avx512.mask.pmovus.wb.512(<32 x i16> %x0, <32 x i8> zeroinitializer, i32 %x2)
1028 %res3 = add <32 x i8> %res0, %res1
1029 %res4 = add <32 x i8> %res3, %res2
1030 ret <32 x i8> %res4
1031 }
1033 declare void @llvm.x86.avx512.mask.pmovus.wb.mem.512(i8* %ptr, <32 x i16>, i32)
1035 define void @test_int_x86_avx512_mask_pmovus_wb_mem_512(i8* %ptr, <32 x i16> %x1, i32 %x2) {
1036 ; X86-LABEL: test_int_x86_avx512_mask_pmovus_wb_mem_512:
1037 ; X86: # %bb.0:
1038 ; X86-NEXT: kmovd {{[0-9]+}}(%esp), %k1 # encoding: [0xc4,0xe1,0xf9,0x90,0x4c,0x24,0x08]
1039 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x04]
1040 ; X86-NEXT: vpmovuswb %zmm0, (%eax) # encoding: [0x62,0xf2,0x7e,0x48,0x10,0x00]
1041 ; X86-NEXT: vpmovuswb %zmm0, (%eax) {%k1} # encoding: [0x62,0xf2,0x7e,0x49,0x10,0x00]
1042 ; X86-NEXT: vzeroupper # encoding: [0xc5,0xf8,0x77]
1043 ; X86-NEXT: retl # encoding: [0xc3]
1044 ;
1045 ; X64-LABEL: test_int_x86_avx512_mask_pmovus_wb_mem_512:
1046 ; X64: # %bb.0:
1047 ; X64-NEXT: kmovd %esi, %k1 # encoding: [0xc5,0xfb,0x92,0xce]
1048 ; X64-NEXT: vpmovuswb %zmm0, (%rdi) # encoding: [0x62,0xf2,0x7e,0x48,0x10,0x07]
1049 ; X64-NEXT: vpmovuswb %zmm0, (%rdi) {%k1} # encoding: [0x62,0xf2,0x7e,0x49,0x10,0x07]
1050 ; X64-NEXT: vzeroupper # encoding: [0xc5,0xf8,0x77]
1051 ; X64-NEXT: retq # encoding: [0xc3]
1052 call void @llvm.x86.avx512.mask.pmovus.wb.mem.512(i8* %ptr, <32 x i16> %x1, i32 -1)
1053 call void @llvm.x86.avx512.mask.pmovus.wb.mem.512(i8* %ptr, <32 x i16> %x1, i32 %x2)
1054 ret void
1055 }
1057 declare <32 x i16> @llvm.x86.avx512.pmaddubs.w.512(<64 x i8>, <64 x i8>)
1059 define <32 x i16> @test_int_x86_avx512_mask_pmaddubs_w_512(<64 x i8> %x0, <64 x i8> %x1, <32 x i16> %x2, i32 %x3) {
1060 ; X86-LABEL: test_int_x86_avx512_mask_pmaddubs_w_512:
1061 ; X86: # %bb.0:
1062 ; X86-NEXT: vpmaddubsw %zmm1, %zmm0, %zmm3 # encoding: [0x62,0xf2,0x7d,0x48,0x04,0xd9]
1063 ; X86-NEXT: kmovd {{[0-9]+}}(%esp), %k1 # encoding: [0xc4,0xe1,0xf9,0x90,0x4c,0x24,0x04]
1064 ; X86-NEXT: vpmaddubsw %zmm1, %zmm0, %zmm2 {%k1} # encoding: [0x62,0xf2,0x7d,0x49,0x04,0xd1]
1065 ; X86-NEXT: vpaddw %zmm3, %zmm2, %zmm0 # encoding: [0x62,0xf1,0x6d,0x48,0xfd,0xc3]
1066 ; X86-NEXT: retl # encoding: [0xc3]
1067 ;
1068 ; X64-LABEL: test_int_x86_avx512_mask_pmaddubs_w_512:
1069 ; X64: # %bb.0:
1070 ; X64-NEXT: vpmaddubsw %zmm1, %zmm0, %zmm3 # encoding: [0x62,0xf2,0x7d,0x48,0x04,0xd9]
1071 ; X64-NEXT: kmovd %edi, %k1 # encoding: [0xc5,0xfb,0x92,0xcf]
1072 ; X64-NEXT: vpmaddubsw %zmm1, %zmm0, %zmm2 {%k1} # encoding: [0x62,0xf2,0x7d,0x49,0x04,0xd1]
1073 ; X64-NEXT: vpaddw %zmm3, %zmm2, %zmm0 # encoding: [0x62,0xf1,0x6d,0x48,0xfd,0xc3]
1074 ; X64-NEXT: retq # encoding: [0xc3]
1075 %1 = call <32 x i16> @llvm.x86.avx512.pmaddubs.w.512(<64 x i8> %x0, <64 x i8> %x1)
1076 %2 = bitcast i32 %x3 to <32 x i1>
1077 %3 = select <32 x i1> %2, <32 x i16> %1, <32 x i16> %x2
1078 %4 = call <32 x i16> @llvm.x86.avx512.pmaddubs.w.512(<64 x i8> %x0, <64 x i8> %x1)
1079 %res2 = add <32 x i16> %3, %4
1080 ret <32 x i16> %res2
1081 }
1083 declare <16 x i32> @llvm.x86.avx512.pmaddw.d.512(<32 x i16>, <32 x i16>)
1085 define <16 x i32> @test_int_x86_avx512_mask_pmaddw_d_512(<32 x i16> %x0, <32 x i16> %x1, <16 x i32> %x2, i16 %x3) {
1086 ; X86-LABEL: test_int_x86_avx512_mask_pmaddw_d_512:
1087 ; X86: # %bb.0:
1088 ; X86-NEXT: vpmaddwd %zmm1, %zmm0, %zmm3 # encoding: [0x62,0xf1,0x7d,0x48,0xf5,0xd9]
1089 ; X86-NEXT: kmovw {{[0-9]+}}(%esp), %k1 # encoding: [0xc5,0xf8,0x90,0x4c,0x24,0x04]
1090 ; X86-NEXT: vpmaddwd %zmm1, %zmm0, %zmm2 {%k1} # encoding: [0x62,0xf1,0x7d,0x49,0xf5,0xd1]
1091 ; X86-NEXT: vpaddd %zmm3, %zmm2, %zmm0 # encoding: [0x62,0xf1,0x6d,0x48,0xfe,0xc3]
1092 ; X86-NEXT: retl # encoding: [0xc3]
1093 ;
1094 ; X64-LABEL: test_int_x86_avx512_mask_pmaddw_d_512:
1095 ; X64: # %bb.0:
1096 ; X64-NEXT: vpmaddwd %zmm1, %zmm0, %zmm3 # encoding: [0x62,0xf1,0x7d,0x48,0xf5,0xd9]
1097 ; X64-NEXT: kmovd %edi, %k1 # encoding: [0xc5,0xfb,0x92,0xcf]
1098 ; X64-NEXT: vpmaddwd %zmm1, %zmm0, %zmm2 {%k1} # encoding: [0x62,0xf1,0x7d,0x49,0xf5,0xd1]
1099 ; X64-NEXT: vpaddd %zmm3, %zmm2, %zmm0 # encoding: [0x62,0xf1,0x6d,0x48,0xfe,0xc3]
1100 ; X64-NEXT: retq # encoding: [0xc3]
1101 %1 = call <16 x i32> @llvm.x86.avx512.pmaddw.d.512(<32 x i16> %x0, <32 x i16> %x1)
1102 %2 = bitcast i16 %x3 to <16 x i1>
1103 %3 = select <16 x i1> %2, <16 x i32> %1, <16 x i32> %x2
1104 %4 = call <16 x i32> @llvm.x86.avx512.pmaddw.d.512(<32 x i16> %x0, <32 x i16> %x1)
1105 %res2 = add <16 x i32> %3, %4
1106 ret <16 x i32> %res2
1107 }
1109 declare <32 x i16> @llvm.x86.avx512.dbpsadbw.512(<64 x i8>, <64 x i8>, i32)
1111 define <32 x i16>@test_int_x86_avx512_mask_dbpsadbw_512(<64 x i8> %x0, <64 x i8> %x1, <32 x i16> %x3, i32 %x4) {
1112 ; X86-LABEL: test_int_x86_avx512_mask_dbpsadbw_512:
1113 ; X86: # %bb.0:
1114 ; X86-NEXT: vdbpsadbw $2, %zmm1, %zmm0, %zmm3 # encoding: [0x62,0xf3,0x7d,0x48,0x42,0xd9,0x02]
1115 ; X86-NEXT: kmovd {{[0-9]+}}(%esp), %k1 # encoding: [0xc4,0xe1,0xf9,0x90,0x4c,0x24,0x04]
1116 ; X86-NEXT: vdbpsadbw $2, %zmm1, %zmm0, %zmm2 {%k1} # encoding: [0x62,0xf3,0x7d,0x49,0x42,0xd1,0x02]
1117 ; X86-NEXT: vdbpsadbw $2, %zmm1, %zmm0, %zmm0 {%k1} {z} # encoding: [0x62,0xf3,0x7d,0xc9,0x42,0xc1,0x02]
1118 ; X86-NEXT: vpaddw %zmm3, %zmm0, %zmm0 # encoding: [0x62,0xf1,0x7d,0x48,0xfd,0xc3]
1119 ; X86-NEXT: vpaddw %zmm0, %zmm2, %zmm0 # encoding: [0x62,0xf1,0x6d,0x48,0xfd,0xc0]
1120 ; X86-NEXT: retl # encoding: [0xc3]
1121 ;
1122 ; X64-LABEL: test_int_x86_avx512_mask_dbpsadbw_512:
1123 ; X64: # %bb.0:
1124 ; X64-NEXT: vdbpsadbw $2, %zmm1, %zmm0, %zmm3 # encoding: [0x62,0xf3,0x7d,0x48,0x42,0xd9,0x02]
1125 ; X64-NEXT: kmovd %edi, %k1 # encoding: [0xc5,0xfb,0x92,0xcf]
1126 ; X64-NEXT: vdbpsadbw $2, %zmm1, %zmm0, %zmm2 {%k1} # encoding: [0x62,0xf3,0x7d,0x49,0x42,0xd1,0x02]
1127 ; X64-NEXT: vdbpsadbw $2, %zmm1, %zmm0, %zmm0 {%k1} {z} # encoding: [0x62,0xf3,0x7d,0xc9,0x42,0xc1,0x02]
1128 ; X64-NEXT: vpaddw %zmm3, %zmm0, %zmm0 # encoding: [0x62,0xf1,0x7d,0x48,0xfd,0xc3]
1129 ; X64-NEXT: vpaddw %zmm0, %zmm2, %zmm0 # encoding: [0x62,0xf1,0x6d,0x48,0xfd,0xc0]
1130 ; X64-NEXT: retq # encoding: [0xc3]
1131 %1 = call <32 x i16> @llvm.x86.avx512.dbpsadbw.512(<64 x i8> %x0, <64 x i8> %x1, i32 2)
1132 %2 = bitcast i32 %x4 to <32 x i1>
1133 %3 = select <32 x i1> %2, <32 x i16> %1, <32 x i16> %x3
1134 %4 = call <32 x i16> @llvm.x86.avx512.dbpsadbw.512(<64 x i8> %x0, <64 x i8> %x1, i32 2)
1135 %5 = bitcast i32 %x4 to <32 x i1>
1136 %6 = select <32 x i1> %5, <32 x i16> %4, <32 x i16> zeroinitializer
1137 %7 = call <32 x i16> @llvm.x86.avx512.dbpsadbw.512(<64 x i8> %x0, <64 x i8> %x1, i32 2)
1138 %res3 = add <32 x i16> %3, %6
1139 %res4 = add <32 x i16> %res3, %7
1140 ret <32 x i16> %res4
1141 }
1143 declare <8 x i64> @llvm.x86.avx512.psad.bw.512(<64 x i8>, <64 x i8>)
1145 define <8 x i64>@test_int_x86_avx512_mask_psadb_w_512(<64 x i8> %x0, <64 x i8> %x1, <64 x i8> %x2){
1146 ; CHECK-LABEL: test_int_x86_avx512_mask_psadb_w_512:
1147 ; CHECK: # %bb.0:
1148 ; CHECK-NEXT: vpsadbw %zmm1, %zmm0, %zmm1 # encoding: [0x62,0xf1,0x7d,0x48,0xf6,0xc9]
1149 ; CHECK-NEXT: vpsadbw %zmm2, %zmm0, %zmm0 # encoding: [0x62,0xf1,0x7d,0x48,0xf6,0xc2]
1150 ; CHECK-NEXT: vpaddq %zmm0, %zmm1, %zmm0 # encoding: [0x62,0xf1,0xf5,0x48,0xd4,0xc0]
1151 ; CHECK-NEXT: ret{{[l|q]}} # encoding: [0xc3]
1152 %res = call <8 x i64> @llvm.x86.avx512.psad.bw.512(<64 x i8> %x0, <64 x i8> %x1)
1153 %res1 = call <8 x i64> @llvm.x86.avx512.psad.bw.512(<64 x i8> %x0, <64 x i8> %x2)
1154 %res2 = add <8 x i64> %res, %res1
1155 ret <8 x i64> %res2
1156 }
1158 declare <32 x i16> @llvm.x86.avx512.psrlv.w.512(<32 x i16>, <32 x i16>) nounwind readnone
1160 define <32 x i16> @test_x86_avx512_psrlv_w_512_const() optsize {
1161 ; X86-LABEL: test_x86_avx512_psrlv_w_512_const:
1162 ; X86: # %bb.0:
1163 ; X86-NEXT: vmovdqa64 {{.*#+}} zmm0 = [4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,65535]
1164 ; X86-NEXT: # encoding: [0x62,0xf1,0xfd,0x48,0x6f,0x05,A,A,A,A]
1165 ; X86-NEXT: # fixup A - offset: 6, value: {{\.LCPI.*}}, kind: FK_Data_4
1166 ; X86-NEXT: vpsrlvw {{\.LCPI.*}}, %zmm0, %zmm0 # encoding: [0x62,0xf2,0xfd,0x48,0x10,0x05,A,A,A,A]
1167 ; X86-NEXT: # fixup A - offset: 6, value: {{\.LCPI.*}}, kind: FK_Data_4
1168 ; X86-NEXT: retl # encoding: [0xc3]
1169 ;
1170 ; X64-LABEL: test_x86_avx512_psrlv_w_512_const:
1171 ; X64: # %bb.0:
1172 ; X64-NEXT: vmovdqa64 {{.*#+}} zmm0 = [4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,65535]
1173 ; X64-NEXT: # encoding: [0x62,0xf1,0xfd,0x48,0x6f,0x05,A,A,A,A]
1174 ; X64-NEXT: # fixup A - offset: 6, value: {{\.LCPI.*}}-4, kind: reloc_riprel_4byte
1175 ; X64-NEXT: vpsrlvw {{.*}}(%rip), %zmm0, %zmm0 # encoding: [0x62,0xf2,0xfd,0x48,0x10,0x05,A,A,A,A]
1176 ; X64-NEXT: # fixup A - offset: 6, value: {{\.LCPI.*}}-4, kind: reloc_riprel_4byte
1177 ; X64-NEXT: retq # encoding: [0xc3]
1178 %res1 = call <32 x i16> @llvm.x86.avx512.psrlv.w.512(<32 x i16> <i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 -1>, <32 x i16> <i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 -1>)
1179 ret <32 x i16> %res1
1180 }
1182 define <32 x i16>@test_int_x86_avx512_mask_psrlv32hi(<32 x i16> %x0, <32 x i16> %x1, <32 x i16> %x2, i32 %x3) {
1183 ; X86-LABEL: test_int_x86_avx512_mask_psrlv32hi:
1184 ; X86: # %bb.0:
1185 ; X86-NEXT: vpsrlvw %zmm1, %zmm0, %zmm3 # encoding: [0x62,0xf2,0xfd,0x48,0x10,0xd9]
1186 ; X86-NEXT: kmovd {{[0-9]+}}(%esp), %k1 # encoding: [0xc4,0xe1,0xf9,0x90,0x4c,0x24,0x04]
1187 ; X86-NEXT: vpsrlvw %zmm1, %zmm0, %zmm2 {%k1} # encoding: [0x62,0xf2,0xfd,0x49,0x10,0xd1]
1188 ; X86-NEXT: vpsrlvw %zmm1, %zmm0, %zmm0 {%k1} {z} # encoding: [0x62,0xf2,0xfd,0xc9,0x10,0xc1]
1189 ; X86-NEXT: vpaddw %zmm3, %zmm0, %zmm0 # encoding: [0x62,0xf1,0x7d,0x48,0xfd,0xc3]
1190 ; X86-NEXT: vpaddw %zmm0, %zmm2, %zmm0 # encoding: [0x62,0xf1,0x6d,0x48,0xfd,0xc0]
1191 ; X86-NEXT: retl # encoding: [0xc3]
1192 ;
1193 ; X64-LABEL: test_int_x86_avx512_mask_psrlv32hi:
1194 ; X64: # %bb.0:
1195 ; X64-NEXT: vpsrlvw %zmm1, %zmm0, %zmm3 # encoding: [0x62,0xf2,0xfd,0x48,0x10,0xd9]
1196 ; X64-NEXT: kmovd %edi, %k1 # encoding: [0xc5,0xfb,0x92,0xcf]
1197 ; X64-NEXT: vpsrlvw %zmm1, %zmm0, %zmm2 {%k1} # encoding: [0x62,0xf2,0xfd,0x49,0x10,0xd1]
1198 ; X64-NEXT: vpsrlvw %zmm1, %zmm0, %zmm0 {%k1} {z} # encoding: [0x62,0xf2,0xfd,0xc9,0x10,0xc1]
1199 ; X64-NEXT: vpaddw %zmm3, %zmm0, %zmm0 # encoding: [0x62,0xf1,0x7d,0x48,0xfd,0xc3]
1200 ; X64-NEXT: vpaddw %zmm0, %zmm2, %zmm0 # encoding: [0x62,0xf1,0x6d,0x48,0xfd,0xc0]
1201 ; X64-NEXT: retq # encoding: [0xc3]
1202 %1 = call <32 x i16> @llvm.x86.avx512.psrlv.w.512(<32 x i16> %x0, <32 x i16> %x1)
1203 %2 = bitcast i32 %x3 to <32 x i1>
1204 %3 = select <32 x i1> %2, <32 x i16> %1, <32 x i16> %x2
1205 %4 = call <32 x i16> @llvm.x86.avx512.psrlv.w.512(<32 x i16> %x0, <32 x i16> %x1)
1206 %5 = bitcast i32 %x3 to <32 x i1>
1207 %6 = select <32 x i1> %5, <32 x i16> %4, <32 x i16> zeroinitializer
1208 %7 = call <32 x i16> @llvm.x86.avx512.psrlv.w.512(<32 x i16> %x0, <32 x i16> %x1)
1209 %res3 = add <32 x i16> %3, %6
1210 %res4 = add <32 x i16> %res3, %7
1211 ret <32 x i16> %res4
1212 }
1214 declare <32 x i16> @llvm.x86.avx512.psrav.w.512(<32 x i16>, <32 x i16>)
1216 define <32 x i16>@test_int_x86_avx512_mask_psrav32_hi(<32 x i16> %x0, <32 x i16> %x1, <32 x i16> %x2, i32 %x3) {
1217 ; X86-LABEL: test_int_x86_avx512_mask_psrav32_hi:
1218 ; X86: # %bb.0:
1219 ; X86-NEXT: vpsravw %zmm1, %zmm0, %zmm3 # encoding: [0x62,0xf2,0xfd,0x48,0x11,0xd9]
1220 ; X86-NEXT: kmovd {{[0-9]+}}(%esp), %k1 # encoding: [0xc4,0xe1,0xf9,0x90,0x4c,0x24,0x04]
1221 ; X86-NEXT: vpsravw %zmm1, %zmm0, %zmm2 {%k1} # encoding: [0x62,0xf2,0xfd,0x49,0x11,0xd1]
1222 ; X86-NEXT: vpsravw %zmm1, %zmm0, %zmm0 {%k1} {z} # encoding: [0x62,0xf2,0xfd,0xc9,0x11,0xc1]
1223 ; X86-NEXT: vpaddw %zmm3, %zmm0, %zmm0 # encoding: [0x62,0xf1,0x7d,0x48,0xfd,0xc3]
1224 ; X86-NEXT: vpaddw %zmm0, %zmm2, %zmm0 # encoding: [0x62,0xf1,0x6d,0x48,0xfd,0xc0]
1225 ; X86-NEXT: retl # encoding: [0xc3]
1226 ;
1227 ; X64-LABEL: test_int_x86_avx512_mask_psrav32_hi:
1228 ; X64: # %bb.0:
1229 ; X64-NEXT: vpsravw %zmm1, %zmm0, %zmm3 # encoding: [0x62,0xf2,0xfd,0x48,0x11,0xd9]
1230 ; X64-NEXT: kmovd %edi, %k1 # encoding: [0xc5,0xfb,0x92,0xcf]
1231 ; X64-NEXT: vpsravw %zmm1, %zmm0, %zmm2 {%k1} # encoding: [0x62,0xf2,0xfd,0x49,0x11,0xd1]
1232 ; X64-NEXT: vpsravw %zmm1, %zmm0, %zmm0 {%k1} {z} # encoding: [0x62,0xf2,0xfd,0xc9,0x11,0xc1]
1233 ; X64-NEXT: vpaddw %zmm3, %zmm0, %zmm0 # encoding: [0x62,0xf1,0x7d,0x48,0xfd,0xc3]
1234 ; X64-NEXT: vpaddw %zmm0, %zmm2, %zmm0 # encoding: [0x62,0xf1,0x6d,0x48,0xfd,0xc0]
1235 ; X64-NEXT: retq # encoding: [0xc3]
1236 %1 = call <32 x i16> @llvm.x86.avx512.psrav.w.512(<32 x i16> %x0, <32 x i16> %x1)
1237 %2 = bitcast i32 %x3 to <32 x i1>
1238 %3 = select <32 x i1> %2, <32 x i16> %1, <32 x i16> %x2
1239 %4 = call <32 x i16> @llvm.x86.avx512.psrav.w.512(<32 x i16> %x0, <32 x i16> %x1)
1240 %5 = bitcast i32 %x3 to <32 x i1>
1241 %6 = select <32 x i1> %5, <32 x i16> %4, <32 x i16> zeroinitializer
1242 %7 = call <32 x i16> @llvm.x86.avx512.psrav.w.512(<32 x i16> %x0, <32 x i16> %x1)
1243 %res3 = add <32 x i16> %3, %6
1244 %res4 = add <32 x i16> %res3, %7
1245 ret <32 x i16> %res4
1246 }
1248 define <32 x i16>@test_int_x86_avx512_mask_psrav32_hi_const(<32 x i16> %x0, <32 x i16> %x1, <32 x i16> %x2, i32 %x3) {
1249 ; X86-LABEL: test_int_x86_avx512_mask_psrav32_hi_const:
1250 ; X86: # %bb.0:
1251 ; X86-NEXT: vmovdqa64 {{.*#+}} zmm0 = [2,9,65524,23,65510,37,65496,51,2,9,65524,23,65510,37,65496,51,2,9,65524,23,65510,37,65496,51,2,9,65524,23,65510,37,65496,51]
1252 ; X86-NEXT: # encoding: [0x62,0xf1,0xfd,0x48,0x6f,0x05,A,A,A,A]
1253 ; X86-NEXT: # fixup A - offset: 6, value: {{\.LCPI.*}}, kind: FK_Data_4
1254 ; X86-NEXT: vpsravw {{\.LCPI.*}}, %zmm0, %zmm0 # encoding: [0x62,0xf2,0xfd,0x48,0x11,0x05,A,A,A,A]
1255 ; X86-NEXT: # fixup A - offset: 6, value: {{\.LCPI.*}}, kind: FK_Data_4
1256 ; X86-NEXT: retl # encoding: [0xc3]
1257 ;
1258 ; X64-LABEL: test_int_x86_avx512_mask_psrav32_hi_const:
1259 ; X64: # %bb.0:
1260 ; X64-NEXT: vmovdqa64 {{.*#+}} zmm0 = [2,9,65524,23,65510,37,65496,51,2,9,65524,23,65510,37,65496,51,2,9,65524,23,65510,37,65496,51,2,9,65524,23,65510,37,65496,51]
1261 ; X64-NEXT: # encoding: [0x62,0xf1,0xfd,0x48,0x6f,0x05,A,A,A,A]
1262 ; X64-NEXT: # fixup A - offset: 6, value: {{\.LCPI.*}}-4, kind: reloc_riprel_4byte
1263 ; X64-NEXT: vpsravw {{.*}}(%rip), %zmm0, %zmm0 # encoding: [0x62,0xf2,0xfd,0x48,0x11,0x05,A,A,A,A]
1264 ; X64-NEXT: # fixup A - offset: 6, value: {{\.LCPI.*}}-4, kind: reloc_riprel_4byte
1265 ; X64-NEXT: retq # encoding: [0xc3]
1266 %1 = call <32 x i16> @llvm.x86.avx512.psrav.w.512(<32 x i16> <i16 2, i16 9, i16 -12, i16 23, i16 -26, i16 37, i16 -40, i16 51, i16 2, i16 9, i16 -12, i16 23, i16 -26, i16 37, i16 -40, i16 51, i16 2, i16 9, i16 -12, i16 23, i16 -26, i16 37, i16 -40, i16 51, i16 2, i16 9, i16 -12, i16 23, i16 -26, i16 37, i16 -40, i16 51>, <32 x i16> <i16 1, i16 10, i16 35, i16 52, i16 69, i16 9, i16 16, i16 49, i16 1, i16 10, i16 35, i16 52, i16 69, i16 9, i16 16, i16 49, i16 1, i16 10, i16 35, i16 52, i16 69, i16 9, i16 16, i16 49, i16 1, i16 10, i16 35, i16 52, i16 69, i16 9, i16 16, i16 49>)
1267 ret <32 x i16> %1
1268 }
1270 define <32 x i16>@test_int_x86_avx512_mask_psllv32hi(<32 x i16> %x0, <32 x i16> %x1, <32 x i16> %x2, i32 %x3) {
1271 ; X86-LABEL: test_int_x86_avx512_mask_psllv32hi:
1272 ; X86: # %bb.0:
1273 ; X86-NEXT: vpsllvw %zmm1, %zmm0, %zmm3 # encoding: [0x62,0xf2,0xfd,0x48,0x12,0xd9]
1274 ; X86-NEXT: kmovd {{[0-9]+}}(%esp), %k1 # encoding: [0xc4,0xe1,0xf9,0x90,0x4c,0x24,0x04]
1275 ; X86-NEXT: vpsllvw %zmm1, %zmm0, %zmm2 {%k1} # encoding: [0x62,0xf2,0xfd,0x49,0x12,0xd1]
1276 ; X86-NEXT: vpsllvw %zmm1, %zmm0, %zmm0 {%k1} {z} # encoding: [0x62,0xf2,0xfd,0xc9,0x12,0xc1]
1277 ; X86-NEXT: vpaddw %zmm3, %zmm0, %zmm0 # encoding: [0x62,0xf1,0x7d,0x48,0xfd,0xc3]
1278 ; X86-NEXT: vpaddw %zmm0, %zmm2, %zmm0 # encoding: [0x62,0xf1,0x6d,0x48,0xfd,0xc0]
1279 ; X86-NEXT: retl # encoding: [0xc3]
1280 ;
1281 ; X64-LABEL: test_int_x86_avx512_mask_psllv32hi:
1282 ; X64: # %bb.0:
1283 ; X64-NEXT: vpsllvw %zmm1, %zmm0, %zmm3 # encoding: [0x62,0xf2,0xfd,0x48,0x12,0xd9]
1284 ; X64-NEXT: kmovd %edi, %k1 # encoding: [0xc5,0xfb,0x92,0xcf]
1285 ; X64-NEXT: vpsllvw %zmm1, %zmm0, %zmm2 {%k1} # encoding: [0x62,0xf2,0xfd,0x49,0x12,0xd1]
1286 ; X64-NEXT: vpsllvw %zmm1, %zmm0, %zmm0 {%k1} {z} # encoding: [0x62,0xf2,0xfd,0xc9,0x12,0xc1]
1287 ; X64-NEXT: vpaddw %zmm3, %zmm0, %zmm0 # encoding: [0x62,0xf1,0x7d,0x48,0xfd,0xc3]
1288 ; X64-NEXT: vpaddw %zmm0, %zmm2, %zmm0 # encoding: [0x62,0xf1,0x6d,0x48,0xfd,0xc0]
1289 ; X64-NEXT: retq # encoding: [0xc3]
1290 %1 = call <32 x i16> @llvm.x86.avx512.psllv.w.512(<32 x i16> %x0, <32 x i16> %x1)
1291 %2 = bitcast i32 %x3 to <32 x i1>
1292 %3 = select <32 x i1> %2, <32 x i16> %1, <32 x i16> %x2
1293 %4 = call <32 x i16> @llvm.x86.avx512.psllv.w.512(<32 x i16> %x0, <32 x i16> %x1)
1294 %5 = bitcast i32 %x3 to <32 x i1>
1295 %6 = select <32 x i1> %5, <32 x i16> %4, <32 x i16> zeroinitializer
1296 %7 = call <32 x i16> @llvm.x86.avx512.psllv.w.512(<32 x i16> %x0, <32 x i16> %x1)
1297 %res3 = add <32 x i16> %3, %6
1298 %res4 = add <32 x i16> %res3, %7
1299 ret <32 x i16> %res4
1300 }
1302 declare <32 x i16> @llvm.x86.avx512.permvar.hi.512(<32 x i16>, <32 x i16>)
1304 define <32 x i16>@test_int_x86_avx512_mask_permvar_hi_512(<32 x i16> %x0, <32 x i16> %x1, <32 x i16> %x2, i32 %x3) {
1305 ; X86-LABEL: test_int_x86_avx512_mask_permvar_hi_512:
1306 ; X86: # %bb.0:
1307 ; X86-NEXT: vpermw %zmm0, %zmm1, %zmm3 # encoding: [0x62,0xf2,0xf5,0x48,0x8d,0xd8]
1308 ; X86-NEXT: kmovd {{[0-9]+}}(%esp), %k1 # encoding: [0xc4,0xe1,0xf9,0x90,0x4c,0x24,0x04]
1309 ; X86-NEXT: vpermw %zmm0, %zmm1, %zmm2 {%k1} # encoding: [0x62,0xf2,0xf5,0x49,0x8d,0xd0]
1310 ; X86-NEXT: vpermw %zmm0, %zmm1, %zmm0 {%k1} {z} # encoding: [0x62,0xf2,0xf5,0xc9,0x8d,0xc0]
1311 ; X86-NEXT: vpaddw %zmm3, %zmm0, %zmm0 # encoding: [0x62,0xf1,0x7d,0x48,0xfd,0xc3]
1312 ; X86-NEXT: vpaddw %zmm0, %zmm2, %zmm0 # encoding: [0x62,0xf1,0x6d,0x48,0xfd,0xc0]
1313 ; X86-NEXT: retl # encoding: [0xc3]
1314 ;
1315 ; X64-LABEL: test_int_x86_avx512_mask_permvar_hi_512:
1316 ; X64: # %bb.0:
1317 ; X64-NEXT: vpermw %zmm0, %zmm1, %zmm3 # encoding: [0x62,0xf2,0xf5,0x48,0x8d,0xd8]
1318 ; X64-NEXT: kmovd %edi, %k1 # encoding: [0xc5,0xfb,0x92,0xcf]
1319 ; X64-NEXT: vpermw %zmm0, %zmm1, %zmm2 {%k1} # encoding: [0x62,0xf2,0xf5,0x49,0x8d,0xd0]
1320 ; X64-NEXT: vpermw %zmm0, %zmm1, %zmm0 {%k1} {z} # encoding: [0x62,0xf2,0xf5,0xc9,0x8d,0xc0]
1321 ; X64-NEXT: vpaddw %zmm3, %zmm0, %zmm0 # encoding: [0x62,0xf1,0x7d,0x48,0xfd,0xc3]
1322 ; X64-NEXT: vpaddw %zmm0, %zmm2, %zmm0 # encoding: [0x62,0xf1,0x6d,0x48,0xfd,0xc0]
1323 ; X64-NEXT: retq # encoding: [0xc3]
1324 %1 = call <32 x i16> @llvm.x86.avx512.permvar.hi.512(<32 x i16> %x0, <32 x i16> %x1)
1325 %2 = bitcast i32 %x3 to <32 x i1>
1326 %3 = select <32 x i1> %2, <32 x i16> %1, <32 x i16> %x2
1327 %4 = call <32 x i16> @llvm.x86.avx512.permvar.hi.512(<32 x i16> %x0, <32 x i16> %x1)
1328 %5 = bitcast i32 %x3 to <32 x i1>
1329 %6 = select <32 x i1> %5, <32 x i16> %4, <32 x i16> zeroinitializer
1330 %7 = call <32 x i16> @llvm.x86.avx512.permvar.hi.512(<32 x i16> %x0, <32 x i16> %x1)
1331 %res3 = add <32 x i16> %3, %6
1332 %res4 = add <32 x i16> %res3, %7
1333 ret <32 x i16> %res4
1334 }
1336 define <32 x i16> @test_x86_avx512_psll_w_512(<32 x i16> %a0, <8 x i16> %a1) {
1337 ; CHECK-LABEL: test_x86_avx512_psll_w_512:
1338 ; CHECK: # %bb.0:
1339 ; CHECK-NEXT: vpsllw %xmm1, %zmm0, %zmm0 # encoding: [0x62,0xf1,0x7d,0x48,0xf1,0xc1]
1340 ; CHECK-NEXT: ret{{[l|q]}} # encoding: [0xc3]
1341 %res = call <32 x i16> @llvm.x86.avx512.psll.w.512(<32 x i16> %a0, <8 x i16> %a1) ; <<32 x i16>> [#uses=1]
1342 ret <32 x i16> %res
1343 }
1344 define <32 x i16> @test_x86_avx512_mask_psll_w_512(<32 x i16> %a0, <8 x i16> %a1, <32 x i16> %passthru, i32 %mask) {
1345 ; X86-LABEL: test_x86_avx512_mask_psll_w_512:
1346 ; X86: # %bb.0:
1347 ; X86-NEXT: kmovd {{[0-9]+}}(%esp), %k1 # encoding: [0xc4,0xe1,0xf9,0x90,0x4c,0x24,0x04]
1348 ; X86-NEXT: vpsllw %xmm1, %zmm0, %zmm2 {%k1} # encoding: [0x62,0xf1,0x7d,0x49,0xf1,0xd1]
1349 ; X86-NEXT: vmovdqa64 %zmm2, %zmm0 # encoding: [0x62,0xf1,0xfd,0x48,0x6f,0xc2]
1350 ; X86-NEXT: retl # encoding: [0xc3]
1351 ;
1352 ; X64-LABEL: test_x86_avx512_mask_psll_w_512:
1353 ; X64: # %bb.0:
1354 ; X64-NEXT: kmovd %edi, %k1 # encoding: [0xc5,0xfb,0x92,0xcf]
1355 ; X64-NEXT: vpsllw %xmm1, %zmm0, %zmm2 {%k1} # encoding: [0x62,0xf1,0x7d,0x49,0xf1,0xd1]
1356 ; X64-NEXT: vmovdqa64 %zmm2, %zmm0 # encoding: [0x62,0xf1,0xfd,0x48,0x6f,0xc2]
1357 ; X64-NEXT: retq # encoding: [0xc3]
1358 %res = call <32 x i16> @llvm.x86.avx512.psll.w.512(<32 x i16> %a0, <8 x i16> %a1) ; <<32 x i16>> [#uses=1]
1359 %mask.cast = bitcast i32 %mask to <32 x i1>
1360 %res2 = select <32 x i1> %mask.cast, <32 x i16> %res, <32 x i16> %passthru
1361 ret <32 x i16> %res2
1362 }
1363 define <32 x i16> @test_x86_avx512_maskz_psll_w_512(<32 x i16> %a0, <8 x i16> %a1, i32 %mask) {
1364 ; X86-LABEL: test_x86_avx512_maskz_psll_w_512:
1365 ; X86: # %bb.0:
1366 ; X86-NEXT: kmovd {{[0-9]+}}(%esp), %k1 # encoding: [0xc4,0xe1,0xf9,0x90,0x4c,0x24,0x04]
1367 ; X86-NEXT: vpsllw %xmm1, %zmm0, %zmm0 {%k1} {z} # encoding: [0x62,0xf1,0x7d,0xc9,0xf1,0xc1]
1368 ; X86-NEXT: retl # encoding: [0xc3]
1369 ;
1370 ; X64-LABEL: test_x86_avx512_maskz_psll_w_512:
1371 ; X64: # %bb.0:
1372 ; X64-NEXT: kmovd %edi, %k1 # encoding: [0xc5,0xfb,0x92,0xcf]
1373 ; X64-NEXT: vpsllw %xmm1, %zmm0, %zmm0 {%k1} {z} # encoding: [0x62,0xf1,0x7d,0xc9,0xf1,0xc1]
1374 ; X64-NEXT: retq # encoding: [0xc3]
1375 %res = call <32 x i16> @llvm.x86.avx512.psll.w.512(<32 x i16> %a0, <8 x i16> %a1) ; <<32 x i16>> [#uses=1]
1376 %mask.cast = bitcast i32 %mask to <32 x i1>
1377 %res2 = select <32 x i1> %mask.cast, <32 x i16> %res, <32 x i16> zeroinitializer
1378 ret <32 x i16> %res2
1379 }
1380 declare <32 x i16> @llvm.x86.avx512.psll.w.512(<32 x i16>, <8 x i16>) nounwind readnone
1383 define <32 x i16> @test_x86_avx512_psllv_w_512_const() optsize {
1384 ; X86-LABEL: test_x86_avx512_psllv_w_512_const:
1385 ; X86: # %bb.0:
1386 ; X86-NEXT: vmovdqa64 {{.*#+}} zmm0 = [4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,65535]
1387 ; X86-NEXT: # encoding: [0x62,0xf1,0xfd,0x48,0x6f,0x05,A,A,A,A]
1388 ; X86-NEXT: # fixup A - offset: 6, value: {{\.LCPI.*}}, kind: FK_Data_4
1389 ; X86-NEXT: vpsllvw {{\.LCPI.*}}, %zmm0, %zmm0 # encoding: [0x62,0xf2,0xfd,0x48,0x12,0x05,A,A,A,A]
1390 ; X86-NEXT: # fixup A - offset: 6, value: {{\.LCPI.*}}, kind: FK_Data_4
1391 ; X86-NEXT: retl # encoding: [0xc3]
1392 ;
1393 ; X64-LABEL: test_x86_avx512_psllv_w_512_const:
1394 ; X64: # %bb.0:
1395 ; X64-NEXT: vmovdqa64 {{.*#+}} zmm0 = [4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,65535]
1396 ; X64-NEXT: # encoding: [0x62,0xf1,0xfd,0x48,0x6f,0x05,A,A,A,A]
1397 ; X64-NEXT: # fixup A - offset: 6, value: {{\.LCPI.*}}-4, kind: reloc_riprel_4byte
1398 ; X64-NEXT: vpsllvw {{.*}}(%rip), %zmm0, %zmm0 # encoding: [0x62,0xf2,0xfd,0x48,0x12,0x05,A,A,A,A]
1399 ; X64-NEXT: # fixup A - offset: 6, value: {{\.LCPI.*}}-4, kind: reloc_riprel_4byte
1400 ; X64-NEXT: retq # encoding: [0xc3]
1401 %res1 = call <32 x i16> @llvm.x86.avx512.psllv.w.512(<32 x i16> <i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 4, i16 -1>, <32 x i16> <i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 -1>)
1402 ret <32 x i16> %res1
1403 }
1404 declare <32 x i16> @llvm.x86.avx512.psllv.w.512(<32 x i16>, <32 x i16>) nounwind readnone
1406 define <32 x i16> @test_x86_avx512_pslli_w_512(<32 x i16> %a0) {
1407 ; CHECK-LABEL: test_x86_avx512_pslli_w_512:
1408 ; CHECK: # %bb.0:
1409 ; CHECK-NEXT: vpsllw $7, %zmm0, %zmm0 # encoding: [0x62,0xf1,0x7d,0x48,0x71,0xf0,0x07]
1410 ; CHECK-NEXT: ret{{[l|q]}} # encoding: [0xc3]
1411 %res = call <32 x i16> @llvm.x86.avx512.pslli.w.512(<32 x i16> %a0, i32 7) ; <<32 x i16>> [#uses=1]
1412 ret <32 x i16> %res
1413 }
1414 define <32 x i16> @test_x86_avx512_mask_pslli_w_512(<32 x i16> %a0, <32 x i16> %passthru, i32 %mask) {
1415 ; X86-LABEL: test_x86_avx512_mask_pslli_w_512:
1416 ; X86: # %bb.0:
1417 ; X86-NEXT: kmovd {{[0-9]+}}(%esp), %k1 # encoding: [0xc4,0xe1,0xf9,0x90,0x4c,0x24,0x04]
1418 ; X86-NEXT: vpsllw $7, %zmm0, %zmm1 {%k1} # encoding: [0x62,0xf1,0x75,0x49,0x71,0xf0,0x07]
1419 ; X86-NEXT: vmovdqa64 %zmm1, %zmm0 # encoding: [0x62,0xf1,0xfd,0x48,0x6f,0xc1]
1420 ; X86-NEXT: retl # encoding: [0xc3]
1421 ;
1422 ; X64-LABEL: test_x86_avx512_mask_pslli_w_512:
1423 ; X64: # %bb.0:
1424 ; X64-NEXT: kmovd %edi, %k1 # encoding: [0xc5,0xfb,0x92,0xcf]
1425 ; X64-NEXT: vpsllw $7, %zmm0, %zmm1 {%k1} # encoding: [0x62,0xf1,0x75,0x49,0x71,0xf0,0x07]
1426 ; X64-NEXT: vmovdqa64 %zmm1, %zmm0 # encoding: [0x62,0xf1,0xfd,0x48,0x6f,0xc1]
1427 ; X64-NEXT: retq # encoding: [0xc3]
1428 %res = call <32 x i16> @llvm.x86.avx512.pslli.w.512(<32 x i16> %a0, i32 7) ; <<32 x i16>> [#uses=1]
1429 %mask.cast = bitcast i32 %mask to <32 x i1>
1430 %res2 = select <32 x i1> %mask.cast, <32 x i16> %res, <32 x i16> %passthru
1431 ret <32 x i16> %res2
1432 }
1433 define <32 x i16> @test_x86_avx512_maskz_pslli_w_512(<32 x i16> %a0, i32 %mask) {
1434 ; X86-LABEL: test_x86_avx512_maskz_pslli_w_512:
1435 ; X86: # %bb.0:
1436 ; X86-NEXT: kmovd {{[0-9]+}}(%esp), %k1 # encoding: [0xc4,0xe1,0xf9,0x90,0x4c,0x24,0x04]
1437 ; X86-NEXT: vpsllw $7, %zmm0, %zmm0 {%k1} {z} # encoding: [0x62,0xf1,0x7d,0xc9,0x71,0xf0,0x07]
1438 ; X86-NEXT: retl # encoding: [0xc3]
1439 ;
1440 ; X64-LABEL: test_x86_avx512_maskz_pslli_w_512:
1441 ; X64: # %bb.0:
1442 ; X64-NEXT: kmovd %edi, %k1 # encoding: [0xc5,0xfb,0x92,0xcf]
1443 ; X64-NEXT: vpsllw $7, %zmm0, %zmm0 {%k1} {z} # encoding: [0x62,0xf1,0x7d,0xc9,0x71,0xf0,0x07]
1444 ; X64-NEXT: retq # encoding: [0xc3]
1445 %res = call <32 x i16> @llvm.x86.avx512.pslli.w.512(<32 x i16> %a0, i32 7) ; <<32 x i16>> [#uses=1]
1446 %mask.cast = bitcast i32 %mask to <32 x i1>
1447 %res2 = select <32 x i1> %mask.cast, <32 x i16> %res, <32 x i16> zeroinitializer
1448 ret <32 x i16> %res2
1449 }
1450 declare <32 x i16> @llvm.x86.avx512.pslli.w.512(<32 x i16>, i32) nounwind readnone
1453 define <32 x i16> @test_x86_avx512_psra_w_512(<32 x i16> %a0, <8 x i16> %a1) {
1454 ; CHECK-LABEL: test_x86_avx512_psra_w_512:
1455 ; CHECK: # %bb.0:
1456 ; CHECK-NEXT: vpsraw %xmm1, %zmm0, %zmm0 # encoding: [0x62,0xf1,0x7d,0x48,0xe1,0xc1]
1457 ; CHECK-NEXT: ret{{[l|q]}} # encoding: [0xc3]
1458 %res = call <32 x i16> @llvm.x86.avx512.psra.w.512(<32 x i16> %a0, <8 x i16> %a1) ; <<32 x i16>> [#uses=1]
1459 ret <32 x i16> %res
1460 }
1461 define <32 x i16> @test_x86_avx512_mask_psra_w_512(<32 x i16> %a0, <8 x i16> %a1, <32 x i16> %passthru, i32 %mask) {
1462 ; X86-LABEL: test_x86_avx512_mask_psra_w_512:
1463 ; X86: # %bb.0:
1464 ; X86-NEXT: kmovd {{[0-9]+}}(%esp), %k1 # encoding: [0xc4,0xe1,0xf9,0x90,0x4c,0x24,0x04]
1465 ; X86-NEXT: vpsraw %xmm1, %zmm0, %zmm2 {%k1} # encoding: [0x62,0xf1,0x7d,0x49,0xe1,0xd1]
1466 ; X86-NEXT: vmovdqa64 %zmm2, %zmm0 # encoding: [0x62,0xf1,0xfd,0x48,0x6f,0xc2]
1467 ; X86-NEXT: retl # encoding: [0xc3]
1468 ;
1469 ; X64-LABEL: test_x86_avx512_mask_psra_w_512:
1470 ; X64: # %bb.0:
1471 ; X64-NEXT: kmovd %edi, %k1 # encoding: [0xc5,0xfb,0x92,0xcf]
1472 ; X64-NEXT: vpsraw %xmm1, %zmm0, %zmm2 {%k1} # encoding: [0x62,0xf1,0x7d,0x49,0xe1,0xd1]
1473 ; X64-NEXT: vmovdqa64 %zmm2, %zmm0 # encoding: [0x62,0xf1,0xfd,0x48,0x6f,0xc2]
1474 ; X64-NEXT: retq # encoding: [0xc3]
1475 %res = call <32 x i16> @llvm.x86.avx512.psra.w.512(<32 x i16> %a0, <8 x i16> %a1) ; <<32 x i16>> [#uses=1]
1476 %mask.cast = bitcast i32 %mask to <32 x i1>
1477 %res2 = select <32 x i1> %mask.cast, <32 x i16> %res, <32 x i16> %passthru
1478 ret <32 x i16> %res2
1479 }
1480 define <32 x i16> @test_x86_avx512_maskz_psra_w_512(<32 x i16> %a0, <8 x i16> %a1, i32 %mask) {
1481 ; X86-LABEL: test_x86_avx512_maskz_psra_w_512:
1482 ; X86: # %bb.0:
1483 ; X86-NEXT: kmovd {{[0-9]+}}(%esp), %k1 # encoding: [0xc4,0xe1,0xf9,0x90,0x4c,0x24,0x04]
1484 ; X86-NEXT: vpsraw %xmm1, %zmm0, %zmm0 {%k1} {z} # encoding: [0x62,0xf1,0x7d,0xc9,0xe1,0xc1]
1485 ; X86-NEXT: retl # encoding: [0xc3]
1486 ;
1487 ; X64-LABEL: test_x86_avx512_maskz_psra_w_512:
1488 ; X64: # %bb.0:
1489 ; X64-NEXT: kmovd %edi, %k1 # encoding: [0xc5,0xfb,0x92,0xcf]
1490 ; X64-NEXT: vpsraw %xmm1, %zmm0, %zmm0 {%k1} {z} # encoding: [0x62,0xf1,0x7d,0xc9,0xe1,0xc1]
1491 ; X64-NEXT: retq # encoding: [0xc3]
1492 %res = call <32 x i16> @llvm.x86.avx512.psra.w.512(<32 x i16> %a0, <8 x i16> %a1) ; <<32 x i16>> [#uses=1]
1493 %mask.cast = bitcast i32 %mask to <32 x i1>
1494 %res2 = select <32 x i1> %mask.cast, <32 x i16> %res, <32 x i16> zeroinitializer
1495 ret <32 x i16> %res2
1496 }
1497 declare <32 x i16> @llvm.x86.avx512.psra.w.512(<32 x i16>, <8 x i16>) nounwind readnone
1500 define <32 x i16> @test_x86_avx512_psrai_w_512(<32 x i16> %a0) {
1501 ; CHECK-LABEL: test_x86_avx512_psrai_w_512:
1502 ; CHECK: # %bb.0:
1503 ; CHECK-NEXT: vpsraw $7, %zmm0, %zmm0 # encoding: [0x62,0xf1,0x7d,0x48,0x71,0xe0,0x07]
1504 ; CHECK-NEXT: ret{{[l|q]}} # encoding: [0xc3]
1505 %res = call <32 x i16> @llvm.x86.avx512.psrai.w.512(<32 x i16> %a0, i32 7) ; <<32 x i16>> [#uses=1]
1506 ret <32 x i16> %res
1507 }
1508 define <32 x i16> @test_x86_avx512_mask_psrai_w_512(<32 x i16> %a0, <32 x i16> %passthru, i32 %mask) {
1509 ; X86-LABEL: test_x86_avx512_mask_psrai_w_512:
1510 ; X86: # %bb.0:
1511 ; X86-NEXT: kmovd {{[0-9]+}}(%esp), %k1 # encoding: [0xc4,0xe1,0xf9,0x90,0x4c,0x24,0x04]
1512 ; X86-NEXT: vpsraw $7, %zmm0, %zmm1 {%k1} # encoding: [0x62,0xf1,0x75,0x49,0x71,0xe0,0x07]
1513 ; X86-NEXT: vmovdqa64 %zmm1, %zmm0 # encoding: [0x62,0xf1,0xfd,0x48,0x6f,0xc1]
1514 ; X86-NEXT: retl # encoding: [0xc3]
1515 ;
1516 ; X64-LABEL: test_x86_avx512_mask_psrai_w_512:
1517 ; X64: # %bb.0:
1518 ; X64-NEXT: kmovd %edi, %k1 # encoding: [0xc5,0xfb,0x92,0xcf]
1519 ; X64-NEXT: vpsraw $7, %zmm0, %zmm1 {%k1} # encoding: [0x62,0xf1,0x75,0x49,0x71,0xe0,0x07]
1520 ; X64-NEXT: vmovdqa64 %zmm1, %zmm0 # encoding: [0x62,0xf1,0xfd,0x48,0x6f,0xc1]
1521 ; X64-NEXT: retq # encoding: [0xc3]
1522 %res = call <32 x i16> @llvm.x86.avx512.psrai.w.512(<32 x i16> %a0, i32 7) ; <<32 x i16>> [#uses=1]
1523 %mask.cast = bitcast i32 %mask to <32 x i1>
1524 %res2 = select <32 x i1> %mask.cast, <32 x i16> %res, <32 x i16> %passthru
1525 ret <32 x i16> %res2
1526 }
1527 define <32 x i16> @test_x86_avx512_maskz_psrai_w_512(<32 x i16> %a0, <32 x i16> %passthru, i32 %mask) {
1528 ; X86-LABEL: test_x86_avx512_maskz_psrai_w_512:
1529 ; X86: # %bb.0:
1530 ; X86-NEXT: kmovd {{[0-9]+}}(%esp), %k1 # encoding: [0xc4,0xe1,0xf9,0x90,0x4c,0x24,0x04]
1531 ; X86-NEXT: vpsraw $7, %zmm0, %zmm0 {%k1} {z} # encoding: [0x62,0xf1,0x7d,0xc9,0x71,0xe0,0x07]
1532 ; X86-NEXT: retl # encoding: [0xc3]
1533 ;
1534 ; X64-LABEL: test_x86_avx512_maskz_psrai_w_512:
1535 ; X64: # %bb.0:
1536 ; X64-NEXT: kmovd %edi, %k1 # encoding: [0xc5,0xfb,0x92,0xcf]
1537 ; X64-NEXT: vpsraw $7, %zmm0, %zmm0 {%k1} {z} # encoding: [0x62,0xf1,0x7d,0xc9,0x71,0xe0,0x07]
1538 ; X64-NEXT: retq # encoding: [0xc3]
1539 %res = call <32 x i16> @llvm.x86.avx512.psrai.w.512(<32 x i16> %a0, i32 7) ; <<32 x i16>> [#uses=1]
1540 %mask.cast = bitcast i32 %mask to <32 x i1>
1541 %res2 = select <32 x i1> %mask.cast, <32 x i16> %res, <32 x i16> zeroinitializer
1542 ret <32 x i16> %res2
1543 }
1544 declare <32 x i16> @llvm.x86.avx512.psrai.w.512(<32 x i16>, i32) nounwind readnone
1547 define <32 x i16> @test_x86_avx512_psrl_w_512(<32 x i16> %a0, <8 x i16> %a1) {
1548 ; CHECK-LABEL: test_x86_avx512_psrl_w_512:
1549 ; CHECK: # %bb.0:
1550 ; CHECK-NEXT: vpsrlw %xmm1, %zmm0, %zmm0 # encoding: [0x62,0xf1,0x7d,0x48,0xd1,0xc1]
1551 ; CHECK-NEXT: ret{{[l|q]}} # encoding: [0xc3]
1552 %res = call <32 x i16> @llvm.x86.avx512.psrl.w.512(<32 x i16> %a0, <8 x i16> %a1) ; <<32 x i16>> [#uses=1]
1553 ret <32 x i16> %res
1554 }
1555 define <32 x i16> @test_x86_avx512_mask_psrl_w_512(<32 x i16> %a0, <8 x i16> %a1, <32 x i16> %passthru, i32 %mask) {
1556 ; X86-LABEL: test_x86_avx512_mask_psrl_w_512:
1557 ; X86: # %bb.0:
1558 ; X86-NEXT: kmovd {{[0-9]+}}(%esp), %k1 # encoding: [0xc4,0xe1,0xf9,0x90,0x4c,0x24,0x04]
1559 ; X86-NEXT: vpsrlw %xmm1, %zmm0, %zmm2 {%k1} # encoding: [0x62,0xf1,0x7d,0x49,0xd1,0xd1]
1560 ; X86-NEXT: vmovdqa64 %zmm2, %zmm0 # encoding: [0x62,0xf1,0xfd,0x48,0x6f,0xc2]
1561 ; X86-NEXT: retl # encoding: [0xc3]
1562 ;
1563 ; X64-LABEL: test_x86_avx512_mask_psrl_w_512:
1564 ; X64: # %bb.0:
1565 ; X64-NEXT: kmovd %edi, %k1 # encoding: [0xc5,0xfb,0x92,0xcf]
1566 ; X64-NEXT: vpsrlw %xmm1, %zmm0, %zmm2 {%k1} # encoding: [0x62,0xf1,0x7d,0x49,0xd1,0xd1]
1567 ; X64-NEXT: vmovdqa64 %zmm2, %zmm0 # encoding: [0x62,0xf1,0xfd,0x48,0x6f,0xc2]
1568 ; X64-NEXT: retq # encoding: [0xc3]
1569 %res = call <32 x i16> @llvm.x86.avx512.psrl.w.512(<32 x i16> %a0, <8 x i16> %a1) ; <<32 x i16>> [#uses=1]
1570 %mask.cast = bitcast i32 %mask to <32 x i1>
1571 %res2 = select <32 x i1> %mask.cast, <32 x i16> %res, <32 x i16> %passthru
1572 ret <32 x i16> %res2
1573 }
1574 define <32 x i16> @test_x86_avx512_maskz_psrl_w_512(<32 x i16> %a0, <8 x i16> %a1, i32 %mask) {
1575 ; X86-LABEL: test_x86_avx512_maskz_psrl_w_512:
1576 ; X86: # %bb.0:
1577 ; X86-NEXT: kmovd {{[0-9]+}}(%esp), %k1 # encoding: [0xc4,0xe1,0xf9,0x90,0x4c,0x24,0x04]
1578 ; X86-NEXT: vpsrlw %xmm1, %zmm0, %zmm0 {%k1} {z} # encoding: [0x62,0xf1,0x7d,0xc9,0xd1,0xc1]
1579 ; X86-NEXT: retl # encoding: [0xc3]
1580 ;
1581 ; X64-LABEL: test_x86_avx512_maskz_psrl_w_512:
1582 ; X64: # %bb.0:
1583 ; X64-NEXT: kmovd %edi, %k1 # encoding: [0xc5,0xfb,0x92,0xcf]
1584 ; X64-NEXT: vpsrlw %xmm1, %zmm0, %zmm0 {%k1} {z} # encoding: [0x62,0xf1,0x7d,0xc9,0xd1,0xc1]
1585 ; X64-NEXT: retq # encoding: [0xc3]
1586 %res = call <32 x i16> @llvm.x86.avx512.psrl.w.512(<32 x i16> %a0, <8 x i16> %a1) ; <<32 x i16>> [#uses=1]
1587 %mask.cast = bitcast i32 %mask to <32 x i1>
1588 %res2 = select <32 x i1> %mask.cast, <32 x i16> %res, <32 x i16> zeroinitializer
1589 ret <32 x i16> %res2
1590 }
1591 declare <32 x i16> @llvm.x86.avx512.psrl.w.512(<32 x i16>, <8 x i16>) nounwind readnone
1593 define <32 x i16> @test_x86_avx512_psrl_w_512_load(<32 x i16> %a0, <8 x i16>* %p) {
1594 ; X86-LABEL: test_x86_avx512_psrl_w_512_load:
1595 ; X86: # %bb.0:
1596 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x04]
1597 ; X86-NEXT: vpsrlw (%eax), %zmm0, %zmm0 # encoding: [0x62,0xf1,0x7d,0x48,0xd1,0x00]
1598 ; X86-NEXT: retl # encoding: [0xc3]
1599 ;
1600 ; X64-LABEL: test_x86_avx512_psrl_w_512_load:
1601 ; X64: # %bb.0:
1602 ; X64-NEXT: vpsrlw (%rdi), %zmm0, %zmm0 # encoding: [0x62,0xf1,0x7d,0x48,0xd1,0x07]
1603 ; X64-NEXT: retq # encoding: [0xc3]
1604 %a1 = load <8 x i16>, <8 x i16>* %p
1605 %res = call <32 x i16> @llvm.x86.avx512.psrl.w.512(<32 x i16> %a0, <8 x i16> %a1) ; <<32 x i16>> [#uses=1]
1606 ret <32 x i16> %res
1607 }
1609 define <32 x i16> @test_x86_avx512_psrli_w_512(<32 x i16> %a0) {
1610 ; CHECK-LABEL: test_x86_avx512_psrli_w_512:
1611 ; CHECK: # %bb.0:
1612 ; CHECK-NEXT: vpsrlw $7, %zmm0, %zmm0 # encoding: [0x62,0xf1,0x7d,0x48,0x71,0xd0,0x07]
1613 ; CHECK-NEXT: ret{{[l|q]}} # encoding: [0xc3]
1614 %res = call <32 x i16> @llvm.x86.avx512.psrli.w.512(<32 x i16> %a0, i32 7) ; <<32 x i16>> [#uses=1]
1615 ret <32 x i16> %res
1616 }
1617 define <32 x i16> @test_x86_avx512_mask_psrli_w_512(<32 x i16> %a0, <32 x i16> %passthru, i32 %mask) {
1618 ; X86-LABEL: test_x86_avx512_mask_psrli_w_512:
1619 ; X86: # %bb.0:
1620 ; X86-NEXT: kmovd {{[0-9]+}}(%esp), %k1 # encoding: [0xc4,0xe1,0xf9,0x90,0x4c,0x24,0x04]
1621 ; X86-NEXT: vpsrlw $7, %zmm0, %zmm1 {%k1} # encoding: [0x62,0xf1,0x75,0x49,0x71,0xd0,0x07]
1622 ; X86-NEXT: vmovdqa64 %zmm1, %zmm0 # encoding: [0x62,0xf1,0xfd,0x48,0x6f,0xc1]
1623 ; X86-NEXT: retl # encoding: [0xc3]
1624 ;
1625 ; X64-LABEL: test_x86_avx512_mask_psrli_w_512:
1626 ; X64: # %bb.0:
1627 ; X64-NEXT: kmovd %edi, %k1 # encoding: [0xc5,0xfb,0x92,0xcf]
1628 ; X64-NEXT: vpsrlw $7, %zmm0, %zmm1 {%k1} # encoding: [0x62,0xf1,0x75,0x49,0x71,0xd0,0x07]
1629 ; X64-NEXT: vmovdqa64 %zmm1, %zmm0 # encoding: [0x62,0xf1,0xfd,0x48,0x6f,0xc1]
1630 ; X64-NEXT: retq # encoding: [0xc3]
1631 %res = call <32 x i16> @llvm.x86.avx512.psrli.w.512(<32 x i16> %a0, i32 7) ; <<32 x i16>> [#uses=1]
1632 %mask.cast = bitcast i32 %mask to <32 x i1>
1633 %res2 = select <32 x i1> %mask.cast, <32 x i16> %res, <32 x i16> %passthru
1634 ret <32 x i16> %res2
1635 }
1636 define <32 x i16> @test_x86_avx512_maskz_psrli_w_512(<32 x i16> %a0, i32 %mask) {
1637 ; X86-LABEL: test_x86_avx512_maskz_psrli_w_512:
1638 ; X86: # %bb.0:
1639 ; X86-NEXT: kmovd {{[0-9]+}}(%esp), %k1 # encoding: [0xc4,0xe1,0xf9,0x90,0x4c,0x24,0x04]
1640 ; X86-NEXT: vpsrlw $7, %zmm0, %zmm0 {%k1} {z} # encoding: [0x62,0xf1,0x7d,0xc9,0x71,0xd0,0x07]
1641 ; X86-NEXT: retl # encoding: [0xc3]
1642 ;
1643 ; X64-LABEL: test_x86_avx512_maskz_psrli_w_512:
1644 ; X64: # %bb.0:
1645 ; X64-NEXT: kmovd %edi, %k1 # encoding: [0xc5,0xfb,0x92,0xcf]
1646 ; X64-NEXT: vpsrlw $7, %zmm0, %zmm0 {%k1} {z} # encoding: [0x62,0xf1,0x7d,0xc9,0x71,0xd0,0x07]
1647 ; X64-NEXT: retq # encoding: [0xc3]
1648 %res = call <32 x i16> @llvm.x86.avx512.psrli.w.512(<32 x i16> %a0, i32 7) ; <<32 x i16>> [#uses=1]
1649 %mask.cast = bitcast i32 %mask to <32 x i1>
1650 %res2 = select <32 x i1> %mask.cast, <32 x i16> %res, <32 x i16> zeroinitializer
1651 ret <32 x i16> %res2
1652 }
1653 declare <32 x i16> @llvm.x86.avx512.psrli.w.512(<32 x i16>, i32) nounwind readnone