Lines Matching +full:0 +full:x92

7 ; X86:       # %bb.0:
8 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
9 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
10 ; X86-NEXT:    vcompresspd %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0xfd,0x09,0x8a,0xc1]
11 ; X86-NEXT:    vmovdqa %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6f,0xc1]
12 ; X86-NEXT:    retl # encoding: [0xc3]
15 ; X64:       # %bb.0:
16 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
17 ; X64-NEXT:    vcompresspd %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0xfd,0x09,0x8a,0xc1]
18 ; X64-NEXT:    vmovdqa %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6f,0xc1]
19 ; X64-NEXT:    retq # encoding: [0xc3]
21   %extract = shufflevector <8 x i1> %1, <8 x i1> %1, <2 x i32> <i32 0, i32 1>
28 ; X86:       # %bb.0:
29 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
30 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
31 ; X86-NEXT:    vcompresspd %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0xfd,0x89,0x8a,0xc0]
32 ; X86-NEXT:    retl # encoding: [0xc3]
35 ; X64:       # %bb.0:
36 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
37 ; X64-NEXT:    vcompresspd %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0xfd,0x89,0x8a,0xc0]
38 ; X64-NEXT:    retq # encoding: [0xc3]
40   %extract = shufflevector <8 x i1> %1, <8 x i1> %1, <2 x i32> <i32 0, i32 1>
47 ; CHECK:       # %bb.0:
48 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
55 ; X86:       # %bb.0:
56 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
57 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
58 ; X86-NEXT:    vcompressps %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7d,0x09,0x8a,0xc1]
59 ; X86-NEXT:    vmovdqa %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6f,0xc1]
60 ; X86-NEXT:    retl # encoding: [0xc3]
63 ; X64:       # %bb.0:
64 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
65 ; X64-NEXT:    vcompressps %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7d,0x09,0x8a,0xc1]
66 ; X64-NEXT:    vmovdqa %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6f,0xc1]
67 ; X64-NEXT:    retq # encoding: [0xc3]
69   %extract = shufflevector <8 x i1> %1, <8 x i1> %1, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
76 ; X86:       # %bb.0:
77 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
78 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
79 ; X86-NEXT:    vcompressps %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7d,0x89,0x8a,0xc0]
80 ; X86-NEXT:    retl # encoding: [0xc3]
83 ; X64:       # %bb.0:
84 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
85 ; X64-NEXT:    vcompressps %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7d,0x89,0x8a,0xc0]
86 ; X64-NEXT:    retq # encoding: [0xc3]
88   %extract = shufflevector <8 x i1> %1, <8 x i1> %1, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
95 ; CHECK:       # %bb.0:
96 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
103 ; X86:       # %bb.0:
104 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
105 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
106 ; X86-NEXT:    vpcompressq %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0xfd,0x09,0x8b,0xc1]
107 ; X86-NEXT:    vmovdqa %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6f,0xc1]
108 ; X86-NEXT:    retl # encoding: [0xc3]
111 ; X64:       # %bb.0:
112 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
113 ; X64-NEXT:    vpcompressq %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0xfd,0x09,0x8b,0xc1]
114 ; X64-NEXT:    vmovdqa %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6f,0xc1]
115 ; X64-NEXT:    retq # encoding: [0xc3]
117   %extract = shufflevector <8 x i1> %1, <8 x i1> %1, <2 x i32> <i32 0, i32 1>
124 ; X86:       # %bb.0:
125 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
126 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
127 ; X86-NEXT:    vpcompressq %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0xfd,0x89,0x8b,0xc0]
128 ; X86-NEXT:    retl # encoding: [0xc3]
131 ; X64:       # %bb.0:
132 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
133 ; X64-NEXT:    vpcompressq %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0xfd,0x89,0x8b,0xc0]
134 ; X64-NEXT:    retq # encoding: [0xc3]
136   %extract = shufflevector <8 x i1> %1, <8 x i1> %1, <2 x i32> <i32 0, i32 1>
143 ; CHECK:       # %bb.0:
144 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
151 ; X86:       # %bb.0:
152 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
153 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
154 ; X86-NEXT:    vpcompressd %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7d,0x09,0x8b,0xc1]
155 ; X86-NEXT:    vmovdqa %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6f,0xc1]
156 ; X86-NEXT:    retl # encoding: [0xc3]
159 ; X64:       # %bb.0:
160 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
161 ; X64-NEXT:    vpcompressd %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7d,0x09,0x8b,0xc1]
162 ; X64-NEXT:    vmovdqa %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6f,0xc1]
163 ; X64-NEXT:    retq # encoding: [0xc3]
165   %extract = shufflevector <8 x i1> %1, <8 x i1> %1, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
172 ; X86:       # %bb.0:
173 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
174 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
175 ; X86-NEXT:    vpcompressd %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7d,0x89,0x8b,0xc0]
176 ; X86-NEXT:    retl # encoding: [0xc3]
179 ; X64:       # %bb.0:
180 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
181 ; X64-NEXT:    vpcompressd %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7d,0x89,0x8b,0xc0]
182 ; X64-NEXT:    retq # encoding: [0xc3]
184   %extract = shufflevector <8 x i1> %1, <8 x i1> %1, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
191 ; CHECK:       # %bb.0:
192 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
199 ; CHECK:       # %bb.0:
200 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
207 ; X86:       # %bb.0:
208 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
209 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
210 ; X86-NEXT:    vexpandpd %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0xfd,0x09,0x88,0xc8]
211 ; X86-NEXT:    vmovdqa %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6f,0xc1]
212 ; X86-NEXT:    retl # encoding: [0xc3]
215 ; X64:       # %bb.0:
216 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
217 ; X64-NEXT:    vexpandpd %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0xfd,0x09,0x88,0xc8]
218 ; X64-NEXT:    vmovdqa %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6f,0xc1]
219 ; X64-NEXT:    retq # encoding: [0xc3]
221   %extract = shufflevector <8 x i1> %1, <8 x i1> %1, <2 x i32> <i32 0, i32 1>
228 ; X86:       # %bb.0:
229 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
230 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
231 ; X86-NEXT:    vexpandpd %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0xfd,0x89,0x88,0xc0]
232 ; X86-NEXT:    retl # encoding: [0xc3]
235 ; X64:       # %bb.0:
236 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
237 ; X64-NEXT:    vexpandpd %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0xfd,0x89,0x88,0xc0]
238 ; X64-NEXT:    retq # encoding: [0xc3]
240   %extract = shufflevector <8 x i1> %1, <8 x i1> %1, <2 x i32> <i32 0, i32 1>
247 ; CHECK:       # %bb.0:
248 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
255 ; X86:       # %bb.0:
256 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
257 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
258 ; X86-NEXT:    vexpandps %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7d,0x09,0x88,0xc8]
259 ; X86-NEXT:    vmovdqa %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6f,0xc1]
260 ; X86-NEXT:    retl # encoding: [0xc3]
263 ; X64:       # %bb.0:
264 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
265 ; X64-NEXT:    vexpandps %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7d,0x09,0x88,0xc8]
266 ; X64-NEXT:    vmovdqa %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6f,0xc1]
267 ; X64-NEXT:    retq # encoding: [0xc3]
269   %extract = shufflevector <8 x i1> %1, <8 x i1> %1, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
276 ; X86:       # %bb.0:
277 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
278 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
279 ; X86-NEXT:    vexpandps %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7d,0x89,0x88,0xc0]
280 ; X86-NEXT:    retl # encoding: [0xc3]
283 ; X64:       # %bb.0:
284 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
285 ; X64-NEXT:    vexpandps %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7d,0x89,0x88,0xc0]
286 ; X64-NEXT:    retq # encoding: [0xc3]
288   %extract = shufflevector <8 x i1> %1, <8 x i1> %1, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
295 ; CHECK:       # %bb.0:
296 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
303 ; X86:       # %bb.0:
304 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
305 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
306 ; X86-NEXT:    vpexpandq %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0xfd,0x09,0x89,0xc8]
307 ; X86-NEXT:    vmovdqa %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6f,0xc1]
308 ; X86-NEXT:    retl # encoding: [0xc3]
311 ; X64:       # %bb.0:
312 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
313 ; X64-NEXT:    vpexpandq %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0xfd,0x09,0x89,0xc8]
314 ; X64-NEXT:    vmovdqa %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6f,0xc1]
315 ; X64-NEXT:    retq # encoding: [0xc3]
317   %extract = shufflevector <8 x i1> %1, <8 x i1> %1, <2 x i32> <i32 0, i32 1>
324 ; X86:       # %bb.0:
325 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
326 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
327 ; X86-NEXT:    vpexpandq %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0xfd,0x89,0x89,0xc0]
328 ; X86-NEXT:    retl # encoding: [0xc3]
331 ; X64:       # %bb.0:
332 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
333 ; X64-NEXT:    vpexpandq %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0xfd,0x89,0x89,0xc0]
334 ; X64-NEXT:    retq # encoding: [0xc3]
336   %extract = shufflevector <8 x i1> %1, <8 x i1> %1, <2 x i32> <i32 0, i32 1>
343 ; CHECK:       # %bb.0:
344 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
351 ; X86:       # %bb.0:
352 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
353 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
354 ; X86-NEXT:    vpexpandd %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7d,0x09,0x89,0xc8]
355 ; X86-NEXT:    vmovdqa %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6f,0xc1]
356 ; X86-NEXT:    retl # encoding: [0xc3]
359 ; X64:       # %bb.0:
360 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
361 ; X64-NEXT:    vpexpandd %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7d,0x09,0x89,0xc8]
362 ; X64-NEXT:    vmovdqa %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6f,0xc1]
363 ; X64-NEXT:    retq # encoding: [0xc3]
365   %extract = shufflevector <8 x i1> %1, <8 x i1> %1, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
372 ; X86:       # %bb.0:
373 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
374 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
375 ; X86-NEXT:    vpexpandd %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7d,0x89,0x89,0xc0]
376 ; X86-NEXT:    retl # encoding: [0xc3]
379 ; X64:       # %bb.0:
380 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
381 ; X64-NEXT:    vpexpandd %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7d,0x89,0x89,0xc0]
382 ; X64-NEXT:    retq # encoding: [0xc3]
384   %extract = shufflevector <8 x i1> %1, <8 x i1> %1, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
391 ; X86:       # %bb.0:
392 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
393 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
394 ; X86-NEXT:    vcompresspd %ymm0, %ymm1 {%k1} # encoding: [0x62,0xf2,0xfd,0x29,0x8a,0xc1]
395 ; X86-NEXT:    vmovdqa %ymm1, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfd,0x6f,0xc1]
396 ; X86-NEXT:    retl # encoding: [0xc3]
399 ; X64:       # %bb.0:
400 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
401 ; X64-NEXT:    vcompresspd %ymm0, %ymm1 {%k1} # encoding: [0x62,0xf2,0xfd,0x29,0x8a,0xc1]
402 ; X64-NEXT:    vmovdqa %ymm1, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfd,0x6f,0xc1]
403 ; X64-NEXT:    retq # encoding: [0xc3]
405   %extract = shufflevector <8 x i1> %1, <8 x i1> %1, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
412 ; X86:       # %bb.0:
413 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
414 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
415 ; X86-NEXT:    vcompresspd %ymm0, %ymm0 {%k1} {z} # encoding: [0x62,0xf2,0xfd,0xa9,0x8a,0xc0]
416 ; X86-NEXT:    retl # encoding: [0xc3]
419 ; X64:       # %bb.0:
420 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
421 ; X64-NEXT:    vcompresspd %ymm0, %ymm0 {%k1} {z} # encoding: [0x62,0xf2,0xfd,0xa9,0x8a,0xc0]
422 ; X64-NEXT:    retq # encoding: [0xc3]
424   %extract = shufflevector <8 x i1> %1, <8 x i1> %1, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
431 ; CHECK:       # %bb.0:
432 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
439 ; X86:       # %bb.0:
440 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
441 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
442 ; X86-NEXT:    vcompressps %ymm0, %ymm1 {%k1} # encoding: [0x62,0xf2,0x7d,0x29,0x8a,0xc1]
443 ; X86-NEXT:    vmovdqa %ymm1, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfd,0x6f,0xc1]
444 ; X86-NEXT:    retl # encoding: [0xc3]
447 ; X64:       # %bb.0:
448 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
449 ; X64-NEXT:    vcompressps %ymm0, %ymm1 {%k1} # encoding: [0x62,0xf2,0x7d,0x29,0x8a,0xc1]
450 ; X64-NEXT:    vmovdqa %ymm1, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfd,0x6f,0xc1]
451 ; X64-NEXT:    retq # encoding: [0xc3]
459 ; X86:       # %bb.0:
460 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
461 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
462 ; X86-NEXT:    vcompressps %ymm0, %ymm0 {%k1} {z} # encoding: [0x62,0xf2,0x7d,0xa9,0x8a,0xc0]
463 ; X86-NEXT:    retl # encoding: [0xc3]
466 ; X64:       # %bb.0:
467 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
468 ; X64-NEXT:    vcompressps %ymm0, %ymm0 {%k1} {z} # encoding: [0x62,0xf2,0x7d,0xa9,0x8a,0xc0]
469 ; X64-NEXT:    retq # encoding: [0xc3]
477 ; CHECK:       # %bb.0:
478 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
485 ; X86:       # %bb.0:
486 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
487 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
488 ; X86-NEXT:    vpcompressq %ymm0, %ymm1 {%k1} # encoding: [0x62,0xf2,0xfd,0x29,0x8b,0xc1]
489 ; X86-NEXT:    vmovdqa %ymm1, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfd,0x6f,0xc1]
490 ; X86-NEXT:    retl # encoding: [0xc3]
493 ; X64:       # %bb.0:
494 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
495 ; X64-NEXT:    vpcompressq %ymm0, %ymm1 {%k1} # encoding: [0x62,0xf2,0xfd,0x29,0x8b,0xc1]
496 ; X64-NEXT:    vmovdqa %ymm1, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfd,0x6f,0xc1]
497 ; X64-NEXT:    retq # encoding: [0xc3]
499   %extract = shufflevector <8 x i1> %1, <8 x i1> %1, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
506 ; X86:       # %bb.0:
507 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
508 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
509 ; X86-NEXT:    vpcompressq %ymm0, %ymm0 {%k1} {z} # encoding: [0x62,0xf2,0xfd,0xa9,0x8b,0xc0]
510 ; X86-NEXT:    retl # encoding: [0xc3]
513 ; X64:       # %bb.0:
514 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
515 ; X64-NEXT:    vpcompressq %ymm0, %ymm0 {%k1} {z} # encoding: [0x62,0xf2,0xfd,0xa9,0x8b,0xc0]
516 ; X64-NEXT:    retq # encoding: [0xc3]
518   %extract = shufflevector <8 x i1> %1, <8 x i1> %1, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
525 ; CHECK:       # %bb.0:
526 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
533 ; X86:       # %bb.0:
534 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
535 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
536 ; X86-NEXT:    vpcompressd %ymm0, %ymm1 {%k1} # encoding: [0x62,0xf2,0x7d,0x29,0x8b,0xc1]
537 ; X86-NEXT:    vmovdqa %ymm1, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfd,0x6f,0xc1]
538 ; X86-NEXT:    retl # encoding: [0xc3]
541 ; X64:       # %bb.0:
542 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
543 ; X64-NEXT:    vpcompressd %ymm0, %ymm1 {%k1} # encoding: [0x62,0xf2,0x7d,0x29,0x8b,0xc1]
544 ; X64-NEXT:    vmovdqa %ymm1, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfd,0x6f,0xc1]
545 ; X64-NEXT:    retq # encoding: [0xc3]
553 ; X86:       # %bb.0:
554 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
555 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
556 ; X86-NEXT:    vpcompressd %ymm0, %ymm0 {%k1} {z} # encoding: [0x62,0xf2,0x7d,0xa9,0x8b,0xc0]
557 ; X86-NEXT:    retl # encoding: [0xc3]
560 ; X64:       # %bb.0:
561 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
562 ; X64-NEXT:    vpcompressd %ymm0, %ymm0 {%k1} {z} # encoding: [0x62,0xf2,0x7d,0xa9,0x8b,0xc0]
563 ; X64-NEXT:    retq # encoding: [0xc3]
571 ; CHECK:       # %bb.0:
572 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
579 ; CHECK:       # %bb.0:
580 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
587 ; X86:       # %bb.0:
588 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
589 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
590 ; X86-NEXT:    vexpandpd %ymm0, %ymm1 {%k1} # encoding: [0x62,0xf2,0xfd,0x29,0x88,0xc8]
591 ; X86-NEXT:    vmovdqa %ymm1, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfd,0x6f,0xc1]
592 ; X86-NEXT:    retl # encoding: [0xc3]
595 ; X64:       # %bb.0:
596 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
597 ; X64-NEXT:    vexpandpd %ymm0, %ymm1 {%k1} # encoding: [0x62,0xf2,0xfd,0x29,0x88,0xc8]
598 ; X64-NEXT:    vmovdqa %ymm1, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfd,0x6f,0xc1]
599 ; X64-NEXT:    retq # encoding: [0xc3]
601   %extract = shufflevector <8 x i1> %1, <8 x i1> %1, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
608 ; X86:       # %bb.0:
609 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
610 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
611 ; X86-NEXT:    vexpandpd %ymm0, %ymm0 {%k1} {z} # encoding: [0x62,0xf2,0xfd,0xa9,0x88,0xc0]
612 ; X86-NEXT:    retl # encoding: [0xc3]
615 ; X64:       # %bb.0:
616 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
617 ; X64-NEXT:    vexpandpd %ymm0, %ymm0 {%k1} {z} # encoding: [0x62,0xf2,0xfd,0xa9,0x88,0xc0]
618 ; X64-NEXT:    retq # encoding: [0xc3]
620   %extract = shufflevector <8 x i1> %1, <8 x i1> %1, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
627 ; CHECK:       # %bb.0:
628 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
635 ; X86:       # %bb.0:
636 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
637 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
638 ; X86-NEXT:    vexpandps %ymm0, %ymm1 {%k1} # encoding: [0x62,0xf2,0x7d,0x29,0x88,0xc8]
639 ; X86-NEXT:    vmovdqa %ymm1, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfd,0x6f,0xc1]
640 ; X86-NEXT:    retl # encoding: [0xc3]
643 ; X64:       # %bb.0:
644 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
645 ; X64-NEXT:    vexpandps %ymm0, %ymm1 {%k1} # encoding: [0x62,0xf2,0x7d,0x29,0x88,0xc8]
646 ; X64-NEXT:    vmovdqa %ymm1, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfd,0x6f,0xc1]
647 ; X64-NEXT:    retq # encoding: [0xc3]
655 ; X86:       # %bb.0:
656 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
657 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
658 ; X86-NEXT:    vexpandps %ymm0, %ymm0 {%k1} {z} # encoding: [0x62,0xf2,0x7d,0xa9,0x88,0xc0]
659 ; X86-NEXT:    retl # encoding: [0xc3]
662 ; X64:       # %bb.0:
663 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
664 ; X64-NEXT:    vexpandps %ymm0, %ymm0 {%k1} {z} # encoding: [0x62,0xf2,0x7d,0xa9,0x88,0xc0]
665 ; X64-NEXT:    retq # encoding: [0xc3]
673 ; CHECK:       # %bb.0:
674 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
681 ; X86:       # %bb.0:
682 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
683 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
684 ; X86-NEXT:    vpexpandq %ymm0, %ymm1 {%k1} # encoding: [0x62,0xf2,0xfd,0x29,0x89,0xc8]
685 ; X86-NEXT:    vmovdqa %ymm1, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfd,0x6f,0xc1]
686 ; X86-NEXT:    retl # encoding: [0xc3]
689 ; X64:       # %bb.0:
690 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
691 ; X64-NEXT:    vpexpandq %ymm0, %ymm1 {%k1} # encoding: [0x62,0xf2,0xfd,0x29,0x89,0xc8]
692 ; X64-NEXT:    vmovdqa %ymm1, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfd,0x6f,0xc1]
693 ; X64-NEXT:    retq # encoding: [0xc3]
695   %extract = shufflevector <8 x i1> %1, <8 x i1> %1, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
702 ; X86:       # %bb.0:
703 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
704 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
705 ; X86-NEXT:    vpexpandq %ymm0, %ymm0 {%k1} {z} # encoding: [0x62,0xf2,0xfd,0xa9,0x89,0xc0]
706 ; X86-NEXT:    retl # encoding: [0xc3]
709 ; X64:       # %bb.0:
710 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
711 ; X64-NEXT:    vpexpandq %ymm0, %ymm0 {%k1} {z} # encoding: [0x62,0xf2,0xfd,0xa9,0x89,0xc0]
712 ; X64-NEXT:    retq # encoding: [0xc3]
714   %extract = shufflevector <8 x i1> %1, <8 x i1> %1, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
721 ; CHECK:       # %bb.0:
722 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
729 ; X86:       # %bb.0:
730 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
731 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
732 ; X86-NEXT:    vpexpandd %ymm0, %ymm1 {%k1} # encoding: [0x62,0xf2,0x7d,0x29,0x89,0xc8]
733 ; X86-NEXT:    vmovdqa %ymm1, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfd,0x6f,0xc1]
734 ; X86-NEXT:    retl # encoding: [0xc3]
737 ; X64:       # %bb.0:
738 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
739 ; X64-NEXT:    vpexpandd %ymm0, %ymm1 {%k1} # encoding: [0x62,0xf2,0x7d,0x29,0x89,0xc8]
740 ; X64-NEXT:    vmovdqa %ymm1, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfd,0x6f,0xc1]
741 ; X64-NEXT:    retq # encoding: [0xc3]
749 ; X86:       # %bb.0:
750 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
751 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
752 ; X86-NEXT:    vpexpandd %ymm0, %ymm0 {%k1} {z} # encoding: [0x62,0xf2,0x7d,0xa9,0x89,0xc0]
753 ; X86-NEXT:    retl # encoding: [0xc3]
756 ; X64:       # %bb.0:
757 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
758 ; X64-NEXT:    vpexpandd %ymm0, %ymm0 {%k1} {z} # encoding: [0x62,0xf2,0x7d,0xa9,0x89,0xc0]
759 ; X64-NEXT:    retq # encoding: [0xc3]
767 ; CHECK:       # %bb.0:
768 ; CHECK-NEXT:    vcmpleps %ymm1, %ymm0, %k0 # encoding: [0x62,0xf1,0x7c,0x28,0xc2,0xc1,0x02]
769 ; CHECK-NEXT:    kmovw %k0, %eax # encoding: [0xc5,0xf8,0x93,0xc0]
771 ; CHECK-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
772 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
781 ; CHECK:       # %bb.0:
782 ; CHECK-NEXT:    vcmpleps %xmm1, %xmm0, %k0 # encoding: [0x62,0xf1,0x7c,0x08,0xc2,0xc1,0x02]
783 ; CHECK-NEXT:    kmovw %k0, %eax # encoding: [0xc5,0xf8,0x93,0xc0]
785 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
787   %1 = shufflevector <4 x i1> %res, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
795 ; CHECK:       # %bb.0:
796 ; CHECK-NEXT:    vcmplepd %ymm1, %ymm0, %k0 # encoding: [0x62,0xf1,0xfd,0x28,0xc2,0xc1,0x02]
797 ; CHECK-NEXT:    kmovw %k0, %eax # encoding: [0xc5,0xf8,0x93,0xc0]
799 ; CHECK-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
800 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
802   %1 = shufflevector <4 x i1> %res, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
810 ; CHECK:       # %bb.0:
811 ; CHECK-NEXT:    vcmplepd %xmm1, %xmm0, %k0 # encoding: [0x62,0xf1,0xfd,0x08,0xc2,0xc1,0x02]
812 ; CHECK-NEXT:    kmovw %k0, %eax # encoding: [0xc5,0xf8,0x93,0xc0]
814 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
816   %1 = shufflevector <2 x i1> %res, <2 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
824 ; X86:       # %bb.0:
825 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
826 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
827 ; X86-NEXT:    vmaxps %ymm1, %ymm0, %ymm0 {%k1} {z} # encoding: [0x62,0xf1,0x7c,0xa9,0x5f,0xc1]
828 ; X86-NEXT:    retl # encoding: [0xc3]
831 ; X64:       # %bb.0:
832 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
833 ; X64-NEXT:    vmaxps %ymm1, %ymm0, %ymm0 {%k1} {z} # encoding: [0x62,0xf1,0x7c,0xa9,0x5f,0xc1]
834 ; X64-NEXT:    retq # encoding: [0xc3]
843 ; X86:       # %bb.0:
844 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
845 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
846 ; X86-NEXT:    vmaxps %ymm1, %ymm0, %ymm2 {%k1} # encoding: [0x62,0xf1,0x7c,0x29,0x5f,0xd1]
847 ; X86-NEXT:    vmovaps %ymm2, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfc,0x28,0xc2]
848 ; X86-NEXT:    retl # encoding: [0xc3]
851 ; X64:       # %bb.0:
852 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
853 ; X64-NEXT:    vmaxps %ymm1, %ymm0, %ymm2 {%k1} # encoding: [0x62,0xf1,0x7c,0x29,0x5f,0xd1]
854 ; X64-NEXT:    vmovaps %ymm2, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfc,0x28,0xc2]
855 ; X64-NEXT:    retq # encoding: [0xc3]
864 ; CHECK:       # %bb.0:
865 ; CHECK-NEXT:    vmaxps %ymm1, %ymm0, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfc,0x5f,0xc1]
866 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
874 ; X86:       # %bb.0:
875 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
876 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
877 ; X86-NEXT:    vmaxps %xmm1, %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf1,0x7c,0x89,0x5f,0xc1]
878 ; X86-NEXT:    retl # encoding: [0xc3]
881 ; X64:       # %bb.0:
882 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
883 ; X64-NEXT:    vmaxps %xmm1, %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf1,0x7c,0x89,0x5f,0xc1]
884 ; X64-NEXT:    retq # encoding: [0xc3]
887   %extract = shufflevector <8 x i1> %2, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
894 ; X86:       # %bb.0:
895 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
896 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
897 ; X86-NEXT:    vmaxps %xmm1, %xmm0, %xmm2 {%k1} # encoding: [0x62,0xf1,0x7c,0x09,0x5f,0xd1]
898 ; X86-NEXT:    vmovaps %xmm2, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x28,0xc2]
899 ; X86-NEXT:    retl # encoding: [0xc3]
902 ; X64:       # %bb.0:
903 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
904 ; X64-NEXT:    vmaxps %xmm1, %xmm0, %xmm2 {%k1} # encoding: [0x62,0xf1,0x7c,0x09,0x5f,0xd1]
905 ; X64-NEXT:    vmovaps %xmm2, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x28,0xc2]
906 ; X64-NEXT:    retq # encoding: [0xc3]
909   %extract = shufflevector <8 x i1> %2, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
916 ; CHECK:       # %bb.0:
917 ; CHECK-NEXT:    vmaxps %xmm1, %xmm0, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x5f,0xc1]
918 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
926 ; X86:       # %bb.0:
927 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
928 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
929 ; X86-NEXT:    vminps %ymm1, %ymm0, %ymm0 {%k1} {z} # encoding: [0x62,0xf1,0x7c,0xa9,0x5d,0xc1]
930 ; X86-NEXT:    retl # encoding: [0xc3]
933 ; X64:       # %bb.0:
934 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
935 ; X64-NEXT:    vminps %ymm1, %ymm0, %ymm0 {%k1} {z} # encoding: [0x62,0xf1,0x7c,0xa9,0x5d,0xc1]
936 ; X64-NEXT:    retq # encoding: [0xc3]
945 ; X86:       # %bb.0:
946 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
947 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
948 ; X86-NEXT:    vminps %ymm1, %ymm0, %ymm2 {%k1} # encoding: [0x62,0xf1,0x7c,0x29,0x5d,0xd1]
949 ; X86-NEXT:    vmovaps %ymm2, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfc,0x28,0xc2]
950 ; X86-NEXT:    retl # encoding: [0xc3]
953 ; X64:       # %bb.0:
954 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
955 ; X64-NEXT:    vminps %ymm1, %ymm0, %ymm2 {%k1} # encoding: [0x62,0xf1,0x7c,0x29,0x5d,0xd1]
956 ; X64-NEXT:    vmovaps %ymm2, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfc,0x28,0xc2]
957 ; X64-NEXT:    retq # encoding: [0xc3]
966 ; CHECK:       # %bb.0:
967 ; CHECK-NEXT:    vminps %ymm1, %ymm0, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfc,0x5d,0xc1]
968 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
976 ; X86:       # %bb.0:
977 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
978 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
979 ; X86-NEXT:    vminps %xmm1, %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf1,0x7c,0x89,0x5d,0xc1]
980 ; X86-NEXT:    retl # encoding: [0xc3]
983 ; X64:       # %bb.0:
984 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
985 ; X64-NEXT:    vminps %xmm1, %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf1,0x7c,0x89,0x5d,0xc1]
986 ; X64-NEXT:    retq # encoding: [0xc3]
989   %extract = shufflevector <8 x i1> %2, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
996 ; X86:       # %bb.0:
997 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
998 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
999 ; X86-NEXT:    vminps %xmm1, %xmm0, %xmm2 {%k1} # encoding: [0x62,0xf1,0x7c,0x09,0x5d,0xd1]
1000 ; X86-NEXT:    vmovaps %xmm2, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x28,0xc2]
1001 ; X86-NEXT:    retl # encoding: [0xc3]
1004 ; X64:       # %bb.0:
1005 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
1006 ; X64-NEXT:    vminps %xmm1, %xmm0, %xmm2 {%k1} # encoding: [0x62,0xf1,0x7c,0x09,0x5d,0xd1]
1007 ; X64-NEXT:    vmovaps %xmm2, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x28,0xc2]
1008 ; X64-NEXT:    retq # encoding: [0xc3]
1011   %extract = shufflevector <8 x i1> %2, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1018 ; CHECK:       # %bb.0:
1019 ; CHECK-NEXT:    vminps %xmm1, %xmm0, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x5d,0xc1]
1020 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
1028 ; CHECK:       # %bb.0:
1029 ; CHECK-NEXT:    vgetexppd %ymm0, %ymm0 # encoding: [0x62,0xf2,0xfd,0x28,0x42,0xc0]
1030 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
1039 ; CHECK:       # %bb.0:
1040 ; CHECK-NEXT:    vgetexpps %ymm0, %ymm0 # encoding: [0x62,0xf2,0x7d,0x28,0x42,0xc0]
1041 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
1051 ; CHECK:       # %bb.0:
1052 ; CHECK-NEXT:    vpermt2d %xmm2, %xmm1, %xmm0 # encoding: [0x62,0xf2,0x75,0x08,0x7e,0xc2]
1053 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
1060 ; X86:       # %bb.0:
1061 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
1062 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
1063 ; X86-NEXT:    vpermi2d %xmm2, %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7d,0x09,0x76,0xca]
1064 ; X86-NEXT:    vmovdqa %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6f,0xc1]
1065 ; X86-NEXT:    retl # encoding: [0xc3]
1068 ; X64:       # %bb.0:
1069 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
1070 ; X64-NEXT:    vpermi2d %xmm2, %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7d,0x09,0x76,0xca]
1071 ; X64-NEXT:    vmovdqa %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6f,0xc1]
1072 ; X64-NEXT:    retq # encoding: [0xc3]
1075   %extract = shufflevector <8 x i1> %2, <8 x i1> %2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1082 ; CHECK:       # %bb.0:
1083 ; CHECK-NEXT:    vpermi2d %xmm2, %xmm1, %xmm0 # encoding: [0x62,0xf2,0x75,0x08,0x76,0xc2]
1084 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
1091 ; X86:       # %bb.0:
1092 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
1093 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
1094 ; X86-NEXT:    vpermt2d %xmm2, %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7d,0x09,0x7e,0xca]
1095 ; X86-NEXT:    vmovdqa %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6f,0xc1]
1096 ; X86-NEXT:    retl # encoding: [0xc3]
1099 ; X64:       # %bb.0:
1100 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
1101 ; X64-NEXT:    vpermt2d %xmm2, %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7d,0x09,0x7e,0xca]
1102 ; X64-NEXT:    vmovdqa %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6f,0xc1]
1103 ; X64-NEXT:    retq # encoding: [0xc3]
1106   %extract = shufflevector <8 x i1> %2, <8 x i1> %2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1113 ; X86:       # %bb.0:
1114 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
1115 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
1116 ; X86-NEXT:    vpermi2d %xmm2, %xmm1, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x75,0x89,0x76,0xc2]
1117 ; X86-NEXT:    retl # encoding: [0xc3]
1120 ; X64:       # %bb.0:
1121 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
1122 ; X64-NEXT:    vpermi2d %xmm2, %xmm1, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x75,0x89,0x76,0xc2]
1123 ; X64-NEXT:    retq # encoding: [0xc3]
1126   %extract = shufflevector <8 x i1> %2, <8 x i1> %2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1135 ; CHECK:       # %bb.0:
1136 ; CHECK-NEXT:    vpermt2d %ymm2, %ymm1, %ymm0 # encoding: [0x62,0xf2,0x75,0x28,0x7e,0xc2]
1137 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
1144 ; X86:       # %bb.0:
1145 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
1146 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
1147 ; X86-NEXT:    vpermi2d %ymm2, %ymm0, %ymm1 {%k1} # encoding: [0x62,0xf2,0x7d,0x29,0x76,0xca]
1148 ; X86-NEXT:    vmovdqa %ymm1, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfd,0x6f,0xc1]
1149 ; X86-NEXT:    retl # encoding: [0xc3]
1152 ; X64:       # %bb.0:
1153 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
1154 ; X64-NEXT:    vpermi2d %ymm2, %ymm0, %ymm1 {%k1} # encoding: [0x62,0xf2,0x7d,0x29,0x76,0xca]
1155 ; X64-NEXT:    vmovdqa %ymm1, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfd,0x6f,0xc1]
1156 ; X64-NEXT:    retq # encoding: [0xc3]
1165 ; CHECK:       # %bb.0:
1166 ; CHECK-NEXT:    vpermi2d %ymm2, %ymm1, %ymm0 # encoding: [0x62,0xf2,0x75,0x28,0x76,0xc2]
1167 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
1174 ; X86:       # %bb.0:
1175 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
1176 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
1177 ; X86-NEXT:    vpermt2d %ymm2, %ymm0, %ymm1 {%k1} # encoding: [0x62,0xf2,0x7d,0x29,0x7e,0xca]
1178 ; X86-NEXT:    vmovdqa %ymm1, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfd,0x6f,0xc1]
1179 ; X86-NEXT:    retl # encoding: [0xc3]
1182 ; X64:       # %bb.0:
1183 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
1184 ; X64-NEXT:    vpermt2d %ymm2, %ymm0, %ymm1 {%k1} # encoding: [0x62,0xf2,0x7d,0x29,0x7e,0xca]
1185 ; X64-NEXT:    vmovdqa %ymm1, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfd,0x6f,0xc1]
1186 ; X64-NEXT:    retq # encoding: [0xc3]
1195 ; X86:       # %bb.0:
1196 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
1197 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
1198 ; X86-NEXT:    vpermi2d %ymm2, %ymm1, %ymm0 {%k1} {z} # encoding: [0x62,0xf2,0x75,0xa9,0x76,0xc2]
1199 ; X86-NEXT:    retl # encoding: [0xc3]
1202 ; X64:       # %bb.0:
1203 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
1204 ; X64-NEXT:    vpermi2d %ymm2, %ymm1, %ymm0 {%k1} {z} # encoding: [0x62,0xf2,0x75,0xa9,0x76,0xc2]
1205 ; X64-NEXT:    retq # encoding: [0xc3]
1216 ; CHECK:       # %bb.0:
1217 ; CHECK-NEXT:    vpermt2pd %xmm2, %xmm1, %xmm0 # encoding: [0x62,0xf2,0xf5,0x08,0x7f,0xc2]
1218 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
1225 ; X86:       # %bb.0:
1226 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
1227 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
1228 ; X86-NEXT:    vpermi2pd %xmm2, %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0xfd,0x09,0x77,0xca]
1229 ; X86-NEXT:    vmovapd %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf9,0x28,0xc1]
1230 ; X86-NEXT:    retl # encoding: [0xc3]
1233 ; X64:       # %bb.0:
1234 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
1235 ; X64-NEXT:    vpermi2pd %xmm2, %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0xfd,0x09,0x77,0xca]
1236 ; X64-NEXT:    vmovapd %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf9,0x28,0xc1]
1237 ; X64-NEXT:    retq # encoding: [0xc3]
1241   %extract = shufflevector <8 x i1> %3, <8 x i1> %3, <2 x i32> <i32 0, i32 1>
1250 ; CHECK:       # %bb.0:
1251 ; CHECK-NEXT:    vpermt2pd %ymm2, %ymm1, %ymm0 # encoding: [0x62,0xf2,0xf5,0x28,0x7f,0xc2]
1252 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
1259 ; X86:       # %bb.0:
1260 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
1261 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
1262 ; X86-NEXT:    vpermi2pd %ymm2, %ymm0, %ymm1 {%k1} # encoding: [0x62,0xf2,0xfd,0x29,0x77,0xca]
1263 ; X86-NEXT:    vmovapd %ymm1, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfd,0x28,0xc1]
1264 ; X86-NEXT:    retl # encoding: [0xc3]
1267 ; X64:       # %bb.0:
1268 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
1269 ; X64-NEXT:    vpermi2pd %ymm2, %ymm0, %ymm1 {%k1} # encoding: [0x62,0xf2,0xfd,0x29,0x77,0xca]
1270 ; X64-NEXT:    vmovapd %ymm1, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfd,0x28,0xc1]
1271 ; X64-NEXT:    retq # encoding: [0xc3]
1275   %extract = shufflevector <8 x i1> %3, <8 x i1> %3, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1284 ; CHECK:       # %bb.0:
1285 ; CHECK-NEXT:    vpermt2ps %xmm2, %xmm1, %xmm0 # encoding: [0x62,0xf2,0x75,0x08,0x7f,0xc2]
1286 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
1293 ; X86:       # %bb.0:
1294 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
1295 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
1296 ; X86-NEXT:    vpermi2ps %xmm2, %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7d,0x09,0x77,0xca]
1297 ; X86-NEXT:    vmovaps %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x28,0xc1]
1298 ; X86-NEXT:    retl # encoding: [0xc3]
1301 ; X64:       # %bb.0:
1302 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
1303 ; X64-NEXT:    vpermi2ps %xmm2, %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7d,0x09,0x77,0xca]
1304 ; X64-NEXT:    vmovaps %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x28,0xc1]
1305 ; X64-NEXT:    retq # encoding: [0xc3]
1309   %extract = shufflevector <8 x i1> %3, <8 x i1> %3, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1316 ; X86:       # %bb.0:
1317 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
1318 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
1319 ; X86-NEXT:    vpermi2ps %xmm2, %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7d,0x09,0x77,0xca]
1320 ; X86-NEXT:    vmovaps %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x28,0xc1]
1321 ; X86-NEXT:    retl # encoding: [0xc3]
1324 ; X64:       # %bb.0:
1325 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
1326 ; X64-NEXT:    vpermi2ps %xmm2, %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7d,0x09,0x77,0xca]
1327 ; X64-NEXT:    vmovaps %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x28,0xc1]
1328 ; X64-NEXT:    retq # encoding: [0xc3]
1333   %extract = shufflevector <8 x i1> %3, <8 x i1> %3, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1342 ; CHECK:       # %bb.0:
1343 ; CHECK-NEXT:    vpermt2ps %ymm2, %ymm1, %ymm0 # encoding: [0x62,0xf2,0x75,0x28,0x7f,0xc2]
1344 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
1351 ; X86:       # %bb.0:
1352 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
1353 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
1354 ; X86-NEXT:    vpermi2ps %ymm2, %ymm0, %ymm1 {%k1} # encoding: [0x62,0xf2,0x7d,0x29,0x77,0xca]
1355 ; X86-NEXT:    vmovaps %ymm1, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfc,0x28,0xc1]
1356 ; X86-NEXT:    retl # encoding: [0xc3]
1359 ; X64:       # %bb.0:
1360 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
1361 ; X64-NEXT:    vpermi2ps %ymm2, %ymm0, %ymm1 {%k1} # encoding: [0x62,0xf2,0x7d,0x29,0x77,0xca]
1362 ; X64-NEXT:    vmovaps %ymm1, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfc,0x28,0xc1]
1363 ; X64-NEXT:    retq # encoding: [0xc3]
1375 ; CHECK:       # %bb.0:
1376 ; CHECK-NEXT:    vpermt2q %xmm2, %xmm1, %xmm0 # encoding: [0x62,0xf2,0xf5,0x08,0x7e,0xc2]
1377 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
1384 ; X86:       # %bb.0:
1385 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
1386 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
1387 ; X86-NEXT:    vpermi2q %xmm2, %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0xfd,0x09,0x76,0xca]
1388 ; X86-NEXT:    vmovdqa %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6f,0xc1]
1389 ; X86-NEXT:    retl # encoding: [0xc3]
1392 ; X64:       # %bb.0:
1393 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
1394 ; X64-NEXT:    vpermi2q %xmm2, %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0xfd,0x09,0x76,0xca]
1395 ; X64-NEXT:    vmovdqa %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6f,0xc1]
1396 ; X64-NEXT:    retq # encoding: [0xc3]
1399   %extract = shufflevector <8 x i1> %2, <8 x i1> %2, <2 x i32> <i32 0, i32 1>
1406 ; CHECK:       # %bb.0:
1407 ; CHECK-NEXT:    vpermi2q %xmm2, %xmm1, %xmm0 # encoding: [0x62,0xf2,0xf5,0x08,0x76,0xc2]
1408 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
1415 ; X86:       # %bb.0:
1416 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
1417 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
1418 ; X86-NEXT:    vpermt2q %xmm2, %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0xfd,0x09,0x7e,0xca]
1419 ; X86-NEXT:    vmovdqa %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6f,0xc1]
1420 ; X86-NEXT:    retl # encoding: [0xc3]
1423 ; X64:       # %bb.0:
1424 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
1425 ; X64-NEXT:    vpermt2q %xmm2, %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0xfd,0x09,0x7e,0xca]
1426 ; X64-NEXT:    vmovdqa %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6f,0xc1]
1427 ; X64-NEXT:    retq # encoding: [0xc3]
1430   %extract = shufflevector <8 x i1> %2, <8 x i1> %2, <2 x i32> <i32 0, i32 1>
1437 ; X86:       # %bb.0:
1438 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
1439 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
1440 ; X86-NEXT:    vpermi2q %xmm2, %xmm1, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0xf5,0x89,0x76,0xc2]
1441 ; X86-NEXT:    retl # encoding: [0xc3]
1444 ; X64:       # %bb.0:
1445 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
1446 ; X64-NEXT:    vpermi2q %xmm2, %xmm1, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0xf5,0x89,0x76,0xc2]
1447 ; X64-NEXT:    retq # encoding: [0xc3]
1450   %extract = shufflevector <8 x i1> %2, <8 x i1> %2, <2 x i32> <i32 0, i32 1>
1459 ; CHECK:       # %bb.0:
1460 ; CHECK-NEXT:    vpermt2q %ymm2, %ymm1, %ymm0 # encoding: [0x62,0xf2,0xf5,0x28,0x7e,0xc2]
1461 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
1468 ; X86:       # %bb.0:
1469 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
1470 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
1471 ; X86-NEXT:    vpermi2q %ymm2, %ymm0, %ymm1 {%k1} # encoding: [0x62,0xf2,0xfd,0x29,0x76,0xca]
1472 ; X86-NEXT:    vmovdqa %ymm1, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfd,0x6f,0xc1]
1473 ; X86-NEXT:    retl # encoding: [0xc3]
1476 ; X64:       # %bb.0:
1477 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
1478 ; X64-NEXT:    vpermi2q %ymm2, %ymm0, %ymm1 {%k1} # encoding: [0x62,0xf2,0xfd,0x29,0x76,0xca]
1479 ; X64-NEXT:    vmovdqa %ymm1, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfd,0x6f,0xc1]
1480 ; X64-NEXT:    retq # encoding: [0xc3]
1483   %extract = shufflevector <8 x i1> %2, <8 x i1> %2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1490 ; CHECK:       # %bb.0:
1491 ; CHECK-NEXT:    vpermi2q %ymm2, %ymm1, %ymm0 # encoding: [0x62,0xf2,0xf5,0x28,0x76,0xc2]
1492 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
1499 ; X86:       # %bb.0:
1500 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
1501 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
1502 ; X86-NEXT:    vpermt2q %ymm2, %ymm0, %ymm1 {%k1} # encoding: [0x62,0xf2,0xfd,0x29,0x7e,0xca]
1503 ; X86-NEXT:    vmovdqa %ymm1, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfd,0x6f,0xc1]
1504 ; X86-NEXT:    retl # encoding: [0xc3]
1507 ; X64:       # %bb.0:
1508 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
1509 ; X64-NEXT:    vpermt2q %ymm2, %ymm0, %ymm1 {%k1} # encoding: [0x62,0xf2,0xfd,0x29,0x7e,0xca]
1510 ; X64-NEXT:    vmovdqa %ymm1, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfd,0x6f,0xc1]
1511 ; X64-NEXT:    retq # encoding: [0xc3]
1514   %extract = shufflevector <8 x i1> %2, <8 x i1> %2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1521 ; X86:       # %bb.0:
1522 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
1523 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
1524 ; X86-NEXT:    vpermi2q %ymm2, %ymm1, %ymm0 {%k1} {z} # encoding: [0x62,0xf2,0xf5,0xa9,0x76,0xc2]
1525 ; X86-NEXT:    retl # encoding: [0xc3]
1528 ; X64:       # %bb.0:
1529 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
1530 ; X64-NEXT:    vpermi2q %ymm2, %ymm1, %ymm0 {%k1} {z} # encoding: [0x62,0xf2,0xf5,0xa9,0x76,0xc2]
1531 ; X64-NEXT:    retq # encoding: [0xc3]
1534   %extract = shufflevector <8 x i1> %2, <8 x i1> %2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1543 ; CHECK:       # %bb.0:
1544 ; CHECK-NEXT:    vscalefpd %xmm1, %xmm0, %xmm0 # encoding: [0x62,0xf2,0xfd,0x08,0x2c,0xc1]
1545 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
1552 ; X86:       # %bb.0:
1553 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
1554 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
1555 ; X86-NEXT:    vscalefpd %xmm1, %xmm0, %xmm2 {%k1} # encoding: [0x62,0xf2,0xfd,0x09,0x2c,0xd1]
1556 ; X86-NEXT:    vmovapd %xmm2, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf9,0x28,0xc2]
1557 ; X86-NEXT:    retl # encoding: [0xc3]
1560 ; X64:       # %bb.0:
1561 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
1562 ; X64-NEXT:    vscalefpd %xmm1, %xmm0, %xmm2 {%k1} # encoding: [0x62,0xf2,0xfd,0x09,0x2c,0xd1]
1563 ; X64-NEXT:    vmovapd %xmm2, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf9,0x28,0xc2]
1564 ; X64-NEXT:    retq # encoding: [0xc3]
1573 ; CHECK:       # %bb.0:
1574 ; CHECK-NEXT:    vscalefpd %ymm1, %ymm0, %ymm0 # encoding: [0x62,0xf2,0xfd,0x28,0x2c,0xc1]
1575 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
1582 ; X86:       # %bb.0:
1583 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
1584 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
1585 ; X86-NEXT:    vscalefpd %ymm1, %ymm0, %ymm2 {%k1} # encoding: [0x62,0xf2,0xfd,0x29,0x2c,0xd1]
1586 ; X86-NEXT:    vmovapd %ymm2, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfd,0x28,0xc2]
1587 ; X86-NEXT:    retl # encoding: [0xc3]
1590 ; X64:       # %bb.0:
1591 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
1592 ; X64-NEXT:    vscalefpd %ymm1, %ymm0, %ymm2 {%k1} # encoding: [0x62,0xf2,0xfd,0x29,0x2c,0xd1]
1593 ; X64-NEXT:    vmovapd %ymm2, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfd,0x28,0xc2]
1594 ; X64-NEXT:    retq # encoding: [0xc3]
1603 ; CHECK:       # %bb.0:
1604 ; CHECK-NEXT:    vscalefps %xmm1, %xmm0, %xmm0 # encoding: [0x62,0xf2,0x7d,0x08,0x2c,0xc1]
1605 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
1612 ; X86:       # %bb.0:
1613 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
1614 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
1615 ; X86-NEXT:    vscalefps %xmm1, %xmm0, %xmm2 {%k1} # encoding: [0x62,0xf2,0x7d,0x09,0x2c,0xd1]
1616 ; X86-NEXT:    vmovaps %xmm2, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x28,0xc2]
1617 ; X86-NEXT:    retl # encoding: [0xc3]
1620 ; X64:       # %bb.0:
1621 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
1622 ; X64-NEXT:    vscalefps %xmm1, %xmm0, %xmm2 {%k1} # encoding: [0x62,0xf2,0x7d,0x09,0x2c,0xd1]
1623 ; X64-NEXT:    vmovaps %xmm2, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x28,0xc2]
1624 ; X64-NEXT:    retq # encoding: [0xc3]
1633 ; CHECK:       # %bb.0:
1634 ; CHECK-NEXT:    vscalefps %ymm1, %ymm0, %ymm0 # encoding: [0x62,0xf2,0x7d,0x28,0x2c,0xc1]
1635 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
1642 ; X86:       # %bb.0:
1643 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
1644 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
1645 ; X86-NEXT:    vscalefps %ymm1, %ymm0, %ymm2 {%k1} # encoding: [0x62,0xf2,0x7d,0x29,0x2c,0xd1]
1646 ; X86-NEXT:    vmovaps %ymm2, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfc,0x28,0xc2]
1647 ; X86-NEXT:    retl # encoding: [0xc3]
1650 ; X64:       # %bb.0:
1651 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
1652 ; X64-NEXT:    vscalefps %ymm1, %ymm0, %ymm2 {%k1} # encoding: [0x62,0xf2,0x7d,0x29,0x2c,0xd1]
1653 ; X64-NEXT:    vmovaps %ymm2, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfc,0x28,0xc2]
1654 ; X64-NEXT:    retq # encoding: [0xc3]
1663 ; X86:       # %bb.0:
1664 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
1665 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
1666 ; X86-NEXT:    vpmovqb %xmm0, %xmm2 # encoding: [0x62,0xf2,0x7e,0x08,0x32,0xc2]
1667 ; X86-NEXT:    vpmovqb %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7e,0x09,0x32,0xc1]
1668 ; X86-NEXT:    vpaddb %xmm1, %xmm2, %xmm1 # encoding: [0xc5,0xe9,0xfc,0xc9]
1669 ; X86-NEXT:    vpmovqb %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7e,0x89,0x32,0xc0]
1670 ; X86-NEXT:    vpaddb %xmm0, %xmm1, %xmm0 # encoding: [0xc5,0xf1,0xfc,0xc0]
1671 ; X86-NEXT:    retl # encoding: [0xc3]
1674 ; X64:       # %bb.0:
1675 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
1676 ; X64-NEXT:    vpmovqb %xmm0, %xmm2 # encoding: [0x62,0xf2,0x7e,0x08,0x32,0xc2]
1677 ; X64-NEXT:    vpmovqb %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7e,0x09,0x32,0xc1]
1678 ; X64-NEXT:    vpaddb %xmm1, %xmm2, %xmm1 # encoding: [0xc5,0xe9,0xfc,0xc9]
1679 ; X64-NEXT:    vpmovqb %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7e,0x89,0x32,0xc0]
1680 ; X64-NEXT:    vpaddb %xmm0, %xmm1, %xmm0 # encoding: [0xc5,0xf1,0xfc,0xc0]
1681 ; X64-NEXT:    retq # encoding: [0xc3]
1694 ; X86:       # %bb.0:
1695 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x08]
1696 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
1697 ; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x04]
1698 ; X86-NEXT:    vpmovqb %xmm0, (%eax) # encoding: [0x62,0xf2,0x7e,0x08,0x32,0x00]
1699 ; X86-NEXT:    vpmovqb %xmm0, (%eax) {%k1} # encoding: [0x62,0xf2,0x7e,0x09,0x32,0x00]
1700 ; X86-NEXT:    retl # encoding: [0xc3]
1703 ; X64:       # %bb.0:
1704 ; X64-NEXT:    kmovw %esi, %k1 # encoding: [0xc5,0xf8,0x92,0xce]
1705 ; X64-NEXT:    vpmovqb %xmm0, (%rdi) # encoding: [0x62,0xf2,0x7e,0x08,0x32,0x07]
1706 ; X64-NEXT:    vpmovqb %xmm0, (%rdi) {%k1} # encoding: [0x62,0xf2,0x7e,0x09,0x32,0x07]
1707 ; X64-NEXT:    retq # encoding: [0xc3]
1717 ; X86:       # %bb.0:
1718 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
1719 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
1720 ; X86-NEXT:    vpmovsqb %xmm0, %xmm2 # encoding: [0x62,0xf2,0x7e,0x08,0x22,0xc2]
1721 ; X86-NEXT:    vpmovsqb %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7e,0x09,0x22,0xc1]
1722 ; X86-NEXT:    vpaddb %xmm1, %xmm2, %xmm1 # encoding: [0xc5,0xe9,0xfc,0xc9]
1723 ; X86-NEXT:    vpmovsqb %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7e,0x89,0x22,0xc0]
1724 ; X86-NEXT:    vpaddb %xmm0, %xmm1, %xmm0 # encoding: [0xc5,0xf1,0xfc,0xc0]
1725 ; X86-NEXT:    retl # encoding: [0xc3]
1728 ; X64:       # %bb.0:
1729 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
1730 ; X64-NEXT:    vpmovsqb %xmm0, %xmm2 # encoding: [0x62,0xf2,0x7e,0x08,0x22,0xc2]
1731 ; X64-NEXT:    vpmovsqb %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7e,0x09,0x22,0xc1]
1732 ; X64-NEXT:    vpaddb %xmm1, %xmm2, %xmm1 # encoding: [0xc5,0xe9,0xfc,0xc9]
1733 ; X64-NEXT:    vpmovsqb %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7e,0x89,0x22,0xc0]
1734 ; X64-NEXT:    vpaddb %xmm0, %xmm1, %xmm0 # encoding: [0xc5,0xf1,0xfc,0xc0]
1735 ; X64-NEXT:    retq # encoding: [0xc3]
1748 ; X86:       # %bb.0:
1749 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x08]
1750 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
1751 ; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x04]
1752 ; X86-NEXT:    vpmovsqb %xmm0, (%eax) # encoding: [0x62,0xf2,0x7e,0x08,0x22,0x00]
1753 ; X86-NEXT:    vpmovsqb %xmm0, (%eax) {%k1} # encoding: [0x62,0xf2,0x7e,0x09,0x22,0x00]
1754 ; X86-NEXT:    retl # encoding: [0xc3]
1757 ; X64:       # %bb.0:
1758 ; X64-NEXT:    kmovw %esi, %k1 # encoding: [0xc5,0xf8,0x92,0xce]
1759 ; X64-NEXT:    vpmovsqb %xmm0, (%rdi) # encoding: [0x62,0xf2,0x7e,0x08,0x22,0x07]
1760 ; X64-NEXT:    vpmovsqb %xmm0, (%rdi) {%k1} # encoding: [0x62,0xf2,0x7e,0x09,0x22,0x07]
1761 ; X64-NEXT:    retq # encoding: [0xc3]
1771 ; X86:       # %bb.0:
1772 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
1773 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
1774 ; X86-NEXT:    vpmovusqb %xmm0, %xmm2 # encoding: [0x62,0xf2,0x7e,0x08,0x12,0xc2]
1775 ; X86-NEXT:    vpmovusqb %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7e,0x09,0x12,0xc1]
1776 ; X86-NEXT:    vpaddb %xmm1, %xmm2, %xmm1 # encoding: [0xc5,0xe9,0xfc,0xc9]
1777 ; X86-NEXT:    vpmovusqb %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7e,0x89,0x12,0xc0]
1778 ; X86-NEXT:    vpaddb %xmm0, %xmm1, %xmm0 # encoding: [0xc5,0xf1,0xfc,0xc0]
1779 ; X86-NEXT:    retl # encoding: [0xc3]
1782 ; X64:       # %bb.0:
1783 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
1784 ; X64-NEXT:    vpmovusqb %xmm0, %xmm2 # encoding: [0x62,0xf2,0x7e,0x08,0x12,0xc2]
1785 ; X64-NEXT:    vpmovusqb %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7e,0x09,0x12,0xc1]
1786 ; X64-NEXT:    vpaddb %xmm1, %xmm2, %xmm1 # encoding: [0xc5,0xe9,0xfc,0xc9]
1787 ; X64-NEXT:    vpmovusqb %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7e,0x89,0x12,0xc0]
1788 ; X64-NEXT:    vpaddb %xmm0, %xmm1, %xmm0 # encoding: [0xc5,0xf1,0xfc,0xc0]
1789 ; X64-NEXT:    retq # encoding: [0xc3]
1802 ; X86:       # %bb.0:
1803 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x08]
1804 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
1805 ; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x04]
1806 ; X86-NEXT:    vpmovusqb %xmm0, (%eax) # encoding: [0x62,0xf2,0x7e,0x08,0x12,0x00]
1807 ; X86-NEXT:    vpmovusqb %xmm0, (%eax) {%k1} # encoding: [0x62,0xf2,0x7e,0x09,0x12,0x00]
1808 ; X86-NEXT:    retl # encoding: [0xc3]
1811 ; X64:       # %bb.0:
1812 ; X64-NEXT:    kmovw %esi, %k1 # encoding: [0xc5,0xf8,0x92,0xce]
1813 ; X64-NEXT:    vpmovusqb %xmm0, (%rdi) # encoding: [0x62,0xf2,0x7e,0x08,0x12,0x07]
1814 ; X64-NEXT:    vpmovusqb %xmm0, (%rdi) {%k1} # encoding: [0x62,0xf2,0x7e,0x09,0x12,0x07]
1815 ; X64-NEXT:    retq # encoding: [0xc3]
1825 ; X86:       # %bb.0:
1826 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
1827 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
1828 ; X86-NEXT:    vpmovqb %ymm0, %xmm2 # encoding: [0x62,0xf2,0x7e,0x28,0x32,0xc2]
1829 ; X86-NEXT:    vpmovqb %ymm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7e,0x29,0x32,0xc1]
1830 ; X86-NEXT:    vpaddb %xmm1, %xmm2, %xmm1 # encoding: [0xc5,0xe9,0xfc,0xc9]
1831 ; X86-NEXT:    vpmovqb %ymm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7e,0xa9,0x32,0xc0]
1832 ; X86-NEXT:    vpaddb %xmm0, %xmm1, %xmm0 # encoding: [0xc5,0xf1,0xfc,0xc0]
1833 ; X86-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
1834 ; X86-NEXT:    retl # encoding: [0xc3]
1837 ; X64:       # %bb.0:
1838 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
1839 ; X64-NEXT:    vpmovqb %ymm0, %xmm2 # encoding: [0x62,0xf2,0x7e,0x28,0x32,0xc2]
1840 ; X64-NEXT:    vpmovqb %ymm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7e,0x29,0x32,0xc1]
1841 ; X64-NEXT:    vpaddb %xmm1, %xmm2, %xmm1 # encoding: [0xc5,0xe9,0xfc,0xc9]
1842 ; X64-NEXT:    vpmovqb %ymm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7e,0xa9,0x32,0xc0]
1843 ; X64-NEXT:    vpaddb %xmm0, %xmm1, %xmm0 # encoding: [0xc5,0xf1,0xfc,0xc0]
1844 ; X64-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
1845 ; X64-NEXT:    retq # encoding: [0xc3]
1858 ; X86:       # %bb.0:
1859 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x08]
1860 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
1861 ; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x04]
1862 ; X86-NEXT:    vpmovqb %ymm0, (%eax) # encoding: [0x62,0xf2,0x7e,0x28,0x32,0x00]
1863 ; X86-NEXT:    vpmovqb %ymm0, (%eax) {%k1} # encoding: [0x62,0xf2,0x7e,0x29,0x32,0x00]
1864 ; X86-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
1865 ; X86-NEXT:    retl # encoding: [0xc3]
1868 ; X64:       # %bb.0:
1869 ; X64-NEXT:    kmovw %esi, %k1 # encoding: [0xc5,0xf8,0x92,0xce]
1870 ; X64-NEXT:    vpmovqb %ymm0, (%rdi) # encoding: [0x62,0xf2,0x7e,0x28,0x32,0x07]
1871 ; X64-NEXT:    vpmovqb %ymm0, (%rdi) {%k1} # encoding: [0x62,0xf2,0x7e,0x29,0x32,0x07]
1872 ; X64-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
1873 ; X64-NEXT:    retq # encoding: [0xc3]
1883 ; X86:       # %bb.0:
1884 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
1885 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
1886 ; X86-NEXT:    vpmovsqb %ymm0, %xmm2 # encoding: [0x62,0xf2,0x7e,0x28,0x22,0xc2]
1887 ; X86-NEXT:    vpmovsqb %ymm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7e,0x29,0x22,0xc1]
1888 ; X86-NEXT:    vpaddb %xmm1, %xmm2, %xmm1 # encoding: [0xc5,0xe9,0xfc,0xc9]
1889 ; X86-NEXT:    vpmovsqb %ymm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7e,0xa9,0x22,0xc0]
1890 ; X86-NEXT:    vpaddb %xmm0, %xmm1, %xmm0 # encoding: [0xc5,0xf1,0xfc,0xc0]
1891 ; X86-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
1892 ; X86-NEXT:    retl # encoding: [0xc3]
1895 ; X64:       # %bb.0:
1896 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
1897 ; X64-NEXT:    vpmovsqb %ymm0, %xmm2 # encoding: [0x62,0xf2,0x7e,0x28,0x22,0xc2]
1898 ; X64-NEXT:    vpmovsqb %ymm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7e,0x29,0x22,0xc1]
1899 ; X64-NEXT:    vpaddb %xmm1, %xmm2, %xmm1 # encoding: [0xc5,0xe9,0xfc,0xc9]
1900 ; X64-NEXT:    vpmovsqb %ymm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7e,0xa9,0x22,0xc0]
1901 ; X64-NEXT:    vpaddb %xmm0, %xmm1, %xmm0 # encoding: [0xc5,0xf1,0xfc,0xc0]
1902 ; X64-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
1903 ; X64-NEXT:    retq # encoding: [0xc3]
1916 ; X86:       # %bb.0:
1917 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x08]
1918 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
1919 ; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x04]
1920 ; X86-NEXT:    vpmovsqb %ymm0, (%eax) # encoding: [0x62,0xf2,0x7e,0x28,0x22,0x00]
1921 ; X86-NEXT:    vpmovsqb %ymm0, (%eax) {%k1} # encoding: [0x62,0xf2,0x7e,0x29,0x22,0x00]
1922 ; X86-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
1923 ; X86-NEXT:    retl # encoding: [0xc3]
1926 ; X64:       # %bb.0:
1927 ; X64-NEXT:    kmovw %esi, %k1 # encoding: [0xc5,0xf8,0x92,0xce]
1928 ; X64-NEXT:    vpmovsqb %ymm0, (%rdi) # encoding: [0x62,0xf2,0x7e,0x28,0x22,0x07]
1929 ; X64-NEXT:    vpmovsqb %ymm0, (%rdi) {%k1} # encoding: [0x62,0xf2,0x7e,0x29,0x22,0x07]
1930 ; X64-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
1931 ; X64-NEXT:    retq # encoding: [0xc3]
1941 ; X86:       # %bb.0:
1942 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
1943 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
1944 ; X86-NEXT:    vpmovusqb %ymm0, %xmm2 # encoding: [0x62,0xf2,0x7e,0x28,0x12,0xc2]
1945 ; X86-NEXT:    vpmovusqb %ymm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7e,0x29,0x12,0xc1]
1946 ; X86-NEXT:    vpaddb %xmm1, %xmm2, %xmm1 # encoding: [0xc5,0xe9,0xfc,0xc9]
1947 ; X86-NEXT:    vpmovusqb %ymm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7e,0xa9,0x12,0xc0]
1948 ; X86-NEXT:    vpaddb %xmm0, %xmm1, %xmm0 # encoding: [0xc5,0xf1,0xfc,0xc0]
1949 ; X86-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
1950 ; X86-NEXT:    retl # encoding: [0xc3]
1953 ; X64:       # %bb.0:
1954 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
1955 ; X64-NEXT:    vpmovusqb %ymm0, %xmm2 # encoding: [0x62,0xf2,0x7e,0x28,0x12,0xc2]
1956 ; X64-NEXT:    vpmovusqb %ymm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7e,0x29,0x12,0xc1]
1957 ; X64-NEXT:    vpaddb %xmm1, %xmm2, %xmm1 # encoding: [0xc5,0xe9,0xfc,0xc9]
1958 ; X64-NEXT:    vpmovusqb %ymm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7e,0xa9,0x12,0xc0]
1959 ; X64-NEXT:    vpaddb %xmm0, %xmm1, %xmm0 # encoding: [0xc5,0xf1,0xfc,0xc0]
1960 ; X64-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
1961 ; X64-NEXT:    retq # encoding: [0xc3]
1974 ; X86:       # %bb.0:
1975 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x08]
1976 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
1977 ; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x04]
1978 ; X86-NEXT:    vpmovusqb %ymm0, (%eax) # encoding: [0x62,0xf2,0x7e,0x28,0x12,0x00]
1979 ; X86-NEXT:    vpmovusqb %ymm0, (%eax) {%k1} # encoding: [0x62,0xf2,0x7e,0x29,0x12,0x00]
1980 ; X86-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
1981 ; X86-NEXT:    retl # encoding: [0xc3]
1984 ; X64:       # %bb.0:
1985 ; X64-NEXT:    kmovw %esi, %k1 # encoding: [0xc5,0xf8,0x92,0xce]
1986 ; X64-NEXT:    vpmovusqb %ymm0, (%rdi) # encoding: [0x62,0xf2,0x7e,0x28,0x12,0x07]
1987 ; X64-NEXT:    vpmovusqb %ymm0, (%rdi) {%k1} # encoding: [0x62,0xf2,0x7e,0x29,0x12,0x07]
1988 ; X64-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
1989 ; X64-NEXT:    retq # encoding: [0xc3]
1999 ; X86:       # %bb.0:
2000 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
2001 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
2002 ; X86-NEXT:    vpmovqw %xmm0, %xmm2 # encoding: [0x62,0xf2,0x7e,0x08,0x34,0xc2]
2003 ; X86-NEXT:    vpmovqw %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7e,0x09,0x34,0xc1]
2004 ; X86-NEXT:    vpaddw %xmm1, %xmm2, %xmm1 # encoding: [0xc5,0xe9,0xfd,0xc9]
2005 ; X86-NEXT:    vpmovqw %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7e,0x89,0x34,0xc0]
2006 ; X86-NEXT:    vpaddw %xmm0, %xmm1, %xmm0 # encoding: [0xc5,0xf1,0xfd,0xc0]
2007 ; X86-NEXT:    retl # encoding: [0xc3]
2010 ; X64:       # %bb.0:
2011 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
2012 ; X64-NEXT:    vpmovqw %xmm0, %xmm2 # encoding: [0x62,0xf2,0x7e,0x08,0x34,0xc2]
2013 ; X64-NEXT:    vpmovqw %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7e,0x09,0x34,0xc1]
2014 ; X64-NEXT:    vpaddw %xmm1, %xmm2, %xmm1 # encoding: [0xc5,0xe9,0xfd,0xc9]
2015 ; X64-NEXT:    vpmovqw %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7e,0x89,0x34,0xc0]
2016 ; X64-NEXT:    vpaddw %xmm0, %xmm1, %xmm0 # encoding: [0xc5,0xf1,0xfd,0xc0]
2017 ; X64-NEXT:    retq # encoding: [0xc3]
2030 ; X86:       # %bb.0:
2031 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x08]
2032 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
2033 ; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x04]
2034 ; X86-NEXT:    vpmovqw %xmm0, (%eax) # encoding: [0x62,0xf2,0x7e,0x08,0x34,0x00]
2035 ; X86-NEXT:    vpmovqw %xmm0, (%eax) {%k1} # encoding: [0x62,0xf2,0x7e,0x09,0x34,0x00]
2036 ; X86-NEXT:    retl # encoding: [0xc3]
2039 ; X64:       # %bb.0:
2040 ; X64-NEXT:    kmovw %esi, %k1 # encoding: [0xc5,0xf8,0x92,0xce]
2041 ; X64-NEXT:    vpmovqw %xmm0, (%rdi) # encoding: [0x62,0xf2,0x7e,0x08,0x34,0x07]
2042 ; X64-NEXT:    vpmovqw %xmm0, (%rdi) {%k1} # encoding: [0x62,0xf2,0x7e,0x09,0x34,0x07]
2043 ; X64-NEXT:    retq # encoding: [0xc3]
2053 ; X86:       # %bb.0:
2054 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
2055 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
2056 ; X86-NEXT:    vpmovsqw %xmm0, %xmm2 # encoding: [0x62,0xf2,0x7e,0x08,0x24,0xc2]
2057 ; X86-NEXT:    vpmovsqw %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7e,0x09,0x24,0xc1]
2058 ; X86-NEXT:    vpaddw %xmm1, %xmm2, %xmm1 # encoding: [0xc5,0xe9,0xfd,0xc9]
2059 ; X86-NEXT:    vpmovsqw %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7e,0x89,0x24,0xc0]
2060 ; X86-NEXT:    vpaddw %xmm0, %xmm1, %xmm0 # encoding: [0xc5,0xf1,0xfd,0xc0]
2061 ; X86-NEXT:    retl # encoding: [0xc3]
2064 ; X64:       # %bb.0:
2065 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
2066 ; X64-NEXT:    vpmovsqw %xmm0, %xmm2 # encoding: [0x62,0xf2,0x7e,0x08,0x24,0xc2]
2067 ; X64-NEXT:    vpmovsqw %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7e,0x09,0x24,0xc1]
2068 ; X64-NEXT:    vpaddw %xmm1, %xmm2, %xmm1 # encoding: [0xc5,0xe9,0xfd,0xc9]
2069 ; X64-NEXT:    vpmovsqw %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7e,0x89,0x24,0xc0]
2070 ; X64-NEXT:    vpaddw %xmm0, %xmm1, %xmm0 # encoding: [0xc5,0xf1,0xfd,0xc0]
2071 ; X64-NEXT:    retq # encoding: [0xc3]
2084 ; X86:       # %bb.0:
2085 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x08]
2086 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
2087 ; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x04]
2088 ; X86-NEXT:    vpmovsqw %xmm0, (%eax) # encoding: [0x62,0xf2,0x7e,0x08,0x24,0x00]
2089 ; X86-NEXT:    vpmovsqw %xmm0, (%eax) {%k1} # encoding: [0x62,0xf2,0x7e,0x09,0x24,0x00]
2090 ; X86-NEXT:    retl # encoding: [0xc3]
2093 ; X64:       # %bb.0:
2094 ; X64-NEXT:    kmovw %esi, %k1 # encoding: [0xc5,0xf8,0x92,0xce]
2095 ; X64-NEXT:    vpmovsqw %xmm0, (%rdi) # encoding: [0x62,0xf2,0x7e,0x08,0x24,0x07]
2096 ; X64-NEXT:    vpmovsqw %xmm0, (%rdi) {%k1} # encoding: [0x62,0xf2,0x7e,0x09,0x24,0x07]
2097 ; X64-NEXT:    retq # encoding: [0xc3]
2107 ; X86:       # %bb.0:
2108 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
2109 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
2110 ; X86-NEXT:    vpmovusqw %xmm0, %xmm2 # encoding: [0x62,0xf2,0x7e,0x08,0x14,0xc2]
2111 ; X86-NEXT:    vpmovusqw %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7e,0x09,0x14,0xc1]
2112 ; X86-NEXT:    vpaddw %xmm1, %xmm2, %xmm1 # encoding: [0xc5,0xe9,0xfd,0xc9]
2113 ; X86-NEXT:    vpmovusqw %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7e,0x89,0x14,0xc0]
2114 ; X86-NEXT:    vpaddw %xmm0, %xmm1, %xmm0 # encoding: [0xc5,0xf1,0xfd,0xc0]
2115 ; X86-NEXT:    retl # encoding: [0xc3]
2118 ; X64:       # %bb.0:
2119 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
2120 ; X64-NEXT:    vpmovusqw %xmm0, %xmm2 # encoding: [0x62,0xf2,0x7e,0x08,0x14,0xc2]
2121 ; X64-NEXT:    vpmovusqw %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7e,0x09,0x14,0xc1]
2122 ; X64-NEXT:    vpaddw %xmm1, %xmm2, %xmm1 # encoding: [0xc5,0xe9,0xfd,0xc9]
2123 ; X64-NEXT:    vpmovusqw %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7e,0x89,0x14,0xc0]
2124 ; X64-NEXT:    vpaddw %xmm0, %xmm1, %xmm0 # encoding: [0xc5,0xf1,0xfd,0xc0]
2125 ; X64-NEXT:    retq # encoding: [0xc3]
2138 ; X86:       # %bb.0:
2139 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x08]
2140 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
2141 ; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x04]
2142 ; X86-NEXT:    vpmovusqw %xmm0, (%eax) # encoding: [0x62,0xf2,0x7e,0x08,0x14,0x00]
2143 ; X86-NEXT:    vpmovusqw %xmm0, (%eax) {%k1} # encoding: [0x62,0xf2,0x7e,0x09,0x14,0x00]
2144 ; X86-NEXT:    retl # encoding: [0xc3]
2147 ; X64:       # %bb.0:
2148 ; X64-NEXT:    kmovw %esi, %k1 # encoding: [0xc5,0xf8,0x92,0xce]
2149 ; X64-NEXT:    vpmovusqw %xmm0, (%rdi) # encoding: [0x62,0xf2,0x7e,0x08,0x14,0x07]
2150 ; X64-NEXT:    vpmovusqw %xmm0, (%rdi) {%k1} # encoding: [0x62,0xf2,0x7e,0x09,0x14,0x07]
2151 ; X64-NEXT:    retq # encoding: [0xc3]
2161 ; X86:       # %bb.0:
2162 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
2163 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
2164 ; X86-NEXT:    vpmovqw %ymm0, %xmm2 # encoding: [0x62,0xf2,0x7e,0x28,0x34,0xc2]
2165 ; X86-NEXT:    vpmovqw %ymm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7e,0x29,0x34,0xc1]
2166 ; X86-NEXT:    vpaddw %xmm1, %xmm2, %xmm1 # encoding: [0xc5,0xe9,0xfd,0xc9]
2167 ; X86-NEXT:    vpmovqw %ymm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7e,0xa9,0x34,0xc0]
2168 ; X86-NEXT:    vpaddw %xmm0, %xmm1, %xmm0 # encoding: [0xc5,0xf1,0xfd,0xc0]
2169 ; X86-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
2170 ; X86-NEXT:    retl # encoding: [0xc3]
2173 ; X64:       # %bb.0:
2174 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
2175 ; X64-NEXT:    vpmovqw %ymm0, %xmm2 # encoding: [0x62,0xf2,0x7e,0x28,0x34,0xc2]
2176 ; X64-NEXT:    vpmovqw %ymm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7e,0x29,0x34,0xc1]
2177 ; X64-NEXT:    vpaddw %xmm1, %xmm2, %xmm1 # encoding: [0xc5,0xe9,0xfd,0xc9]
2178 ; X64-NEXT:    vpmovqw %ymm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7e,0xa9,0x34,0xc0]
2179 ; X64-NEXT:    vpaddw %xmm0, %xmm1, %xmm0 # encoding: [0xc5,0xf1,0xfd,0xc0]
2180 ; X64-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
2181 ; X64-NEXT:    retq # encoding: [0xc3]
2194 ; X86:       # %bb.0:
2195 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x08]
2196 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
2197 ; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x04]
2198 ; X86-NEXT:    vpmovqw %ymm0, (%eax) # encoding: [0x62,0xf2,0x7e,0x28,0x34,0x00]
2199 ; X86-NEXT:    vpmovqw %ymm0, (%eax) {%k1} # encoding: [0x62,0xf2,0x7e,0x29,0x34,0x00]
2200 ; X86-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
2201 ; X86-NEXT:    retl # encoding: [0xc3]
2204 ; X64:       # %bb.0:
2205 ; X64-NEXT:    kmovw %esi, %k1 # encoding: [0xc5,0xf8,0x92,0xce]
2206 ; X64-NEXT:    vpmovqw %ymm0, (%rdi) # encoding: [0x62,0xf2,0x7e,0x28,0x34,0x07]
2207 ; X64-NEXT:    vpmovqw %ymm0, (%rdi) {%k1} # encoding: [0x62,0xf2,0x7e,0x29,0x34,0x07]
2208 ; X64-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
2209 ; X64-NEXT:    retq # encoding: [0xc3]
2219 ; X86:       # %bb.0:
2220 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
2221 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
2222 ; X86-NEXT:    vpmovsqw %ymm0, %xmm2 # encoding: [0x62,0xf2,0x7e,0x28,0x24,0xc2]
2223 ; X86-NEXT:    vpmovsqw %ymm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7e,0x29,0x24,0xc1]
2224 ; X86-NEXT:    vpaddw %xmm1, %xmm2, %xmm1 # encoding: [0xc5,0xe9,0xfd,0xc9]
2225 ; X86-NEXT:    vpmovsqw %ymm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7e,0xa9,0x24,0xc0]
2226 ; X86-NEXT:    vpaddw %xmm0, %xmm1, %xmm0 # encoding: [0xc5,0xf1,0xfd,0xc0]
2227 ; X86-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
2228 ; X86-NEXT:    retl # encoding: [0xc3]
2231 ; X64:       # %bb.0:
2232 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
2233 ; X64-NEXT:    vpmovsqw %ymm0, %xmm2 # encoding: [0x62,0xf2,0x7e,0x28,0x24,0xc2]
2234 ; X64-NEXT:    vpmovsqw %ymm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7e,0x29,0x24,0xc1]
2235 ; X64-NEXT:    vpaddw %xmm1, %xmm2, %xmm1 # encoding: [0xc5,0xe9,0xfd,0xc9]
2236 ; X64-NEXT:    vpmovsqw %ymm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7e,0xa9,0x24,0xc0]
2237 ; X64-NEXT:    vpaddw %xmm0, %xmm1, %xmm0 # encoding: [0xc5,0xf1,0xfd,0xc0]
2238 ; X64-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
2239 ; X64-NEXT:    retq # encoding: [0xc3]
2252 ; X86:       # %bb.0:
2253 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x08]
2254 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
2255 ; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x04]
2256 ; X86-NEXT:    vpmovsqw %ymm0, (%eax) # encoding: [0x62,0xf2,0x7e,0x28,0x24,0x00]
2257 ; X86-NEXT:    vpmovsqw %ymm0, (%eax) {%k1} # encoding: [0x62,0xf2,0x7e,0x29,0x24,0x00]
2258 ; X86-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
2259 ; X86-NEXT:    retl # encoding: [0xc3]
2262 ; X64:       # %bb.0:
2263 ; X64-NEXT:    kmovw %esi, %k1 # encoding: [0xc5,0xf8,0x92,0xce]
2264 ; X64-NEXT:    vpmovsqw %ymm0, (%rdi) # encoding: [0x62,0xf2,0x7e,0x28,0x24,0x07]
2265 ; X64-NEXT:    vpmovsqw %ymm0, (%rdi) {%k1} # encoding: [0x62,0xf2,0x7e,0x29,0x24,0x07]
2266 ; X64-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
2267 ; X64-NEXT:    retq # encoding: [0xc3]
2277 ; X86:       # %bb.0:
2278 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
2279 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
2280 ; X86-NEXT:    vpmovusqw %ymm0, %xmm2 # encoding: [0x62,0xf2,0x7e,0x28,0x14,0xc2]
2281 ; X86-NEXT:    vpmovusqw %ymm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7e,0x29,0x14,0xc1]
2282 ; X86-NEXT:    vpaddw %xmm1, %xmm2, %xmm1 # encoding: [0xc5,0xe9,0xfd,0xc9]
2283 ; X86-NEXT:    vpmovusqw %ymm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7e,0xa9,0x14,0xc0]
2284 ; X86-NEXT:    vpaddw %xmm0, %xmm1, %xmm0 # encoding: [0xc5,0xf1,0xfd,0xc0]
2285 ; X86-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
2286 ; X86-NEXT:    retl # encoding: [0xc3]
2289 ; X64:       # %bb.0:
2290 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
2291 ; X64-NEXT:    vpmovusqw %ymm0, %xmm2 # encoding: [0x62,0xf2,0x7e,0x28,0x14,0xc2]
2292 ; X64-NEXT:    vpmovusqw %ymm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7e,0x29,0x14,0xc1]
2293 ; X64-NEXT:    vpaddw %xmm1, %xmm2, %xmm1 # encoding: [0xc5,0xe9,0xfd,0xc9]
2294 ; X64-NEXT:    vpmovusqw %ymm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7e,0xa9,0x14,0xc0]
2295 ; X64-NEXT:    vpaddw %xmm0, %xmm1, %xmm0 # encoding: [0xc5,0xf1,0xfd,0xc0]
2296 ; X64-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
2297 ; X64-NEXT:    retq # encoding: [0xc3]
2310 ; X86:       # %bb.0:
2311 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x08]
2312 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
2313 ; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x04]
2314 ; X86-NEXT:    vpmovusqw %ymm0, (%eax) # encoding: [0x62,0xf2,0x7e,0x28,0x14,0x00]
2315 ; X86-NEXT:    vpmovusqw %ymm0, (%eax) {%k1} # encoding: [0x62,0xf2,0x7e,0x29,0x14,0x00]
2316 ; X86-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
2317 ; X86-NEXT:    retl # encoding: [0xc3]
2320 ; X64:       # %bb.0:
2321 ; X64-NEXT:    kmovw %esi, %k1 # encoding: [0xc5,0xf8,0x92,0xce]
2322 ; X64-NEXT:    vpmovusqw %ymm0, (%rdi) # encoding: [0x62,0xf2,0x7e,0x28,0x14,0x07]
2323 ; X64-NEXT:    vpmovusqw %ymm0, (%rdi) {%k1} # encoding: [0x62,0xf2,0x7e,0x29,0x14,0x07]
2324 ; X64-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
2325 ; X64-NEXT:    retq # encoding: [0xc3]
2335 ; X86:       # %bb.0:
2336 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
2337 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
2338 ; X86-NEXT:    vpmovqd %xmm0, %xmm2 # encoding: [0x62,0xf2,0x7e,0x08,0x35,0xc2]
2339 ; X86-NEXT:    vpmovqd %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7e,0x09,0x35,0xc1]
2340 ; X86-NEXT:    vpaddd %xmm1, %xmm2, %xmm1 # EVEX TO VEX Compression encoding: [0xc5,0xe9,0xfe,0xc9]
2341 ; X86-NEXT:    vpmovqd %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7e,0x89,0x35,0xc0]
2342 ; X86-NEXT:    vpaddd %xmm0, %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf1,0xfe,0xc0]
2343 ; X86-NEXT:    retl # encoding: [0xc3]
2346 ; X64:       # %bb.0:
2347 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
2348 ; X64-NEXT:    vpmovqd %xmm0, %xmm2 # encoding: [0x62,0xf2,0x7e,0x08,0x35,0xc2]
2349 ; X64-NEXT:    vpmovqd %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7e,0x09,0x35,0xc1]
2350 ; X64-NEXT:    vpaddd %xmm1, %xmm2, %xmm1 # EVEX TO VEX Compression encoding: [0xc5,0xe9,0xfe,0xc9]
2351 ; X64-NEXT:    vpmovqd %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7e,0x89,0x35,0xc0]
2352 ; X64-NEXT:    vpaddd %xmm0, %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf1,0xfe,0xc0]
2353 ; X64-NEXT:    retq # encoding: [0xc3]
2366 ; X86:       # %bb.0:
2367 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x08]
2368 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
2369 ; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x04]
2370 ; X86-NEXT:    vpmovqd %xmm0, (%eax) # encoding: [0x62,0xf2,0x7e,0x08,0x35,0x00]
2371 ; X86-NEXT:    vpmovqd %xmm0, (%eax) {%k1} # encoding: [0x62,0xf2,0x7e,0x09,0x35,0x00]
2372 ; X86-NEXT:    retl # encoding: [0xc3]
2375 ; X64:       # %bb.0:
2376 ; X64-NEXT:    kmovw %esi, %k1 # encoding: [0xc5,0xf8,0x92,0xce]
2377 ; X64-NEXT:    vpmovqd %xmm0, (%rdi) # encoding: [0x62,0xf2,0x7e,0x08,0x35,0x07]
2378 ; X64-NEXT:    vpmovqd %xmm0, (%rdi) {%k1} # encoding: [0x62,0xf2,0x7e,0x09,0x35,0x07]
2379 ; X64-NEXT:    retq # encoding: [0xc3]
2389 ; X86:       # %bb.0:
2390 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
2391 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
2392 ; X86-NEXT:    vpmovsqd %xmm0, %xmm2 # encoding: [0x62,0xf2,0x7e,0x08,0x25,0xc2]
2393 ; X86-NEXT:    vpmovsqd %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7e,0x09,0x25,0xc1]
2394 ; X86-NEXT:    vpaddd %xmm1, %xmm2, %xmm1 # EVEX TO VEX Compression encoding: [0xc5,0xe9,0xfe,0xc9]
2395 ; X86-NEXT:    vpmovsqd %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7e,0x89,0x25,0xc0]
2396 ; X86-NEXT:    vpaddd %xmm0, %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf1,0xfe,0xc0]
2397 ; X86-NEXT:    retl # encoding: [0xc3]
2400 ; X64:       # %bb.0:
2401 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
2402 ; X64-NEXT:    vpmovsqd %xmm0, %xmm2 # encoding: [0x62,0xf2,0x7e,0x08,0x25,0xc2]
2403 ; X64-NEXT:    vpmovsqd %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7e,0x09,0x25,0xc1]
2404 ; X64-NEXT:    vpaddd %xmm1, %xmm2, %xmm1 # EVEX TO VEX Compression encoding: [0xc5,0xe9,0xfe,0xc9]
2405 ; X64-NEXT:    vpmovsqd %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7e,0x89,0x25,0xc0]
2406 ; X64-NEXT:    vpaddd %xmm0, %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf1,0xfe,0xc0]
2407 ; X64-NEXT:    retq # encoding: [0xc3]
2420 ; X86:       # %bb.0:
2421 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x08]
2422 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
2423 ; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x04]
2424 ; X86-NEXT:    vpmovsqd %xmm0, (%eax) # encoding: [0x62,0xf2,0x7e,0x08,0x25,0x00]
2425 ; X86-NEXT:    vpmovsqd %xmm0, (%eax) {%k1} # encoding: [0x62,0xf2,0x7e,0x09,0x25,0x00]
2426 ; X86-NEXT:    retl # encoding: [0xc3]
2429 ; X64:       # %bb.0:
2430 ; X64-NEXT:    kmovw %esi, %k1 # encoding: [0xc5,0xf8,0x92,0xce]
2431 ; X64-NEXT:    vpmovsqd %xmm0, (%rdi) # encoding: [0x62,0xf2,0x7e,0x08,0x25,0x07]
2432 ; X64-NEXT:    vpmovsqd %xmm0, (%rdi) {%k1} # encoding: [0x62,0xf2,0x7e,0x09,0x25,0x07]
2433 ; X64-NEXT:    retq # encoding: [0xc3]
2443 ; X86:       # %bb.0:
2444 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
2445 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
2446 ; X86-NEXT:    vpmovusqd %xmm0, %xmm2 # encoding: [0x62,0xf2,0x7e,0x08,0x15,0xc2]
2447 ; X86-NEXT:    vpmovusqd %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7e,0x09,0x15,0xc1]
2448 ; X86-NEXT:    vpaddd %xmm1, %xmm2, %xmm1 # EVEX TO VEX Compression encoding: [0xc5,0xe9,0xfe,0xc9]
2449 ; X86-NEXT:    vpmovusqd %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7e,0x89,0x15,0xc0]
2450 ; X86-NEXT:    vpaddd %xmm0, %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf1,0xfe,0xc0]
2451 ; X86-NEXT:    retl # encoding: [0xc3]
2454 ; X64:       # %bb.0:
2455 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
2456 ; X64-NEXT:    vpmovusqd %xmm0, %xmm2 # encoding: [0x62,0xf2,0x7e,0x08,0x15,0xc2]
2457 ; X64-NEXT:    vpmovusqd %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7e,0x09,0x15,0xc1]
2458 ; X64-NEXT:    vpaddd %xmm1, %xmm2, %xmm1 # EVEX TO VEX Compression encoding: [0xc5,0xe9,0xfe,0xc9]
2459 ; X64-NEXT:    vpmovusqd %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7e,0x89,0x15,0xc0]
2460 ; X64-NEXT:    vpaddd %xmm0, %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf1,0xfe,0xc0]
2461 ; X64-NEXT:    retq # encoding: [0xc3]
2474 ; X86:       # %bb.0:
2475 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x08]
2476 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
2477 ; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x04]
2478 ; X86-NEXT:    vpmovusqd %xmm0, (%eax) # encoding: [0x62,0xf2,0x7e,0x08,0x15,0x00]
2479 ; X86-NEXT:    vpmovusqd %xmm0, (%eax) {%k1} # encoding: [0x62,0xf2,0x7e,0x09,0x15,0x00]
2480 ; X86-NEXT:    retl # encoding: [0xc3]
2483 ; X64:       # %bb.0:
2484 ; X64-NEXT:    kmovw %esi, %k1 # encoding: [0xc5,0xf8,0x92,0xce]
2485 ; X64-NEXT:    vpmovusqd %xmm0, (%rdi) # encoding: [0x62,0xf2,0x7e,0x08,0x15,0x07]
2486 ; X64-NEXT:    vpmovusqd %xmm0, (%rdi) {%k1} # encoding: [0x62,0xf2,0x7e,0x09,0x15,0x07]
2487 ; X64-NEXT:    retq # encoding: [0xc3]
2495 ; CHECK:       # %bb.0:
2496 ; CHECK-NEXT:    vpmovqd %ymm0, %xmm0 # encoding: [0x62,0xf2,0x7e,0x28,0x35,0xc0]
2497 ; CHECK-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
2498 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
2505 ; X86:       # %bb.0:
2506 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
2507 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
2508 ; X86-NEXT:    vpmovqd %ymm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7e,0x29,0x35,0xc1]
2509 ; X86-NEXT:    vmovdqa %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6f,0xc1]
2510 ; X86-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
2511 ; X86-NEXT:    retl # encoding: [0xc3]
2514 ; X64:       # %bb.0:
2515 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
2516 ; X64-NEXT:    vpmovqd %ymm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7e,0x29,0x35,0xc1]
2517 ; X64-NEXT:    vmovdqa %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6f,0xc1]
2518 ; X64-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
2519 ; X64-NEXT:    retq # encoding: [0xc3]
2522   %extract1 = shufflevector <8 x i1> %2, <8 x i1> %2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
2529 ; X86:       # %bb.0:
2530 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
2531 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
2532 ; X86-NEXT:    vpmovqd %ymm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7e,0xa9,0x35,0xc0]
2533 ; X86-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
2534 ; X86-NEXT:    retl # encoding: [0xc3]
2537 ; X64:       # %bb.0:
2538 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
2539 ; X64-NEXT:    vpmovqd %ymm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7e,0xa9,0x35,0xc0]
2540 ; X64-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
2541 ; X64-NEXT:    retq # encoding: [0xc3]
2544   %extract = shufflevector <8 x i1> %2, <8 x i1> %2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
2553 ; X86:       # %bb.0:
2554 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x08]
2555 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
2556 ; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x04]
2557 ; X86-NEXT:    vpmovqd %ymm0, (%eax) # encoding: [0x62,0xf2,0x7e,0x28,0x35,0x00]
2558 ; X86-NEXT:    vpmovqd %ymm0, (%eax) {%k1} # encoding: [0x62,0xf2,0x7e,0x29,0x35,0x00]
2559 ; X86-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
2560 ; X86-NEXT:    retl # encoding: [0xc3]
2563 ; X64:       # %bb.0:
2564 ; X64-NEXT:    kmovw %esi, %k1 # encoding: [0xc5,0xf8,0x92,0xce]
2565 ; X64-NEXT:    vpmovqd %ymm0, (%rdi) # encoding: [0x62,0xf2,0x7e,0x28,0x35,0x07]
2566 ; X64-NEXT:    vpmovqd %ymm0, (%rdi) {%k1} # encoding: [0x62,0xf2,0x7e,0x29,0x35,0x07]
2567 ; X64-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
2568 ; X64-NEXT:    retq # encoding: [0xc3]
2578 ; CHECK:       # %bb.0:
2579 ; CHECK-NEXT:    vpmovsqd %ymm0, %xmm0 # encoding: [0x62,0xf2,0x7e,0x28,0x25,0xc0]
2580 ; CHECK-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
2581 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
2588 ; X86:       # %bb.0:
2589 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
2590 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
2591 ; X86-NEXT:    vpmovsqd %ymm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7e,0x29,0x25,0xc1]
2592 ; X86-NEXT:    vmovdqa %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6f,0xc1]
2593 ; X86-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
2594 ; X86-NEXT:    retl # encoding: [0xc3]
2597 ; X64:       # %bb.0:
2598 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
2599 ; X64-NEXT:    vpmovsqd %ymm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7e,0x29,0x25,0xc1]
2600 ; X64-NEXT:    vmovdqa %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6f,0xc1]
2601 ; X64-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
2602 ; X64-NEXT:    retq # encoding: [0xc3]
2609 ; X86:       # %bb.0:
2610 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
2611 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
2612 ; X86-NEXT:    vpmovsqd %ymm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7e,0xa9,0x25,0xc0]
2613 ; X86-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
2614 ; X86-NEXT:    retl # encoding: [0xc3]
2617 ; X64:       # %bb.0:
2618 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
2619 ; X64-NEXT:    vpmovsqd %ymm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7e,0xa9,0x25,0xc0]
2620 ; X64-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
2621 ; X64-NEXT:    retq # encoding: [0xc3]
2630 ; X86:       # %bb.0:
2631 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x08]
2632 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
2633 ; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x04]
2634 ; X86-NEXT:    vpmovsqd %ymm0, (%eax) # encoding: [0x62,0xf2,0x7e,0x28,0x25,0x00]
2635 ; X86-NEXT:    vpmovsqd %ymm0, (%eax) {%k1} # encoding: [0x62,0xf2,0x7e,0x29,0x25,0x00]
2636 ; X86-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
2637 ; X86-NEXT:    retl # encoding: [0xc3]
2640 ; X64:       # %bb.0:
2641 ; X64-NEXT:    kmovw %esi, %k1 # encoding: [0xc5,0xf8,0x92,0xce]
2642 ; X64-NEXT:    vpmovsqd %ymm0, (%rdi) # encoding: [0x62,0xf2,0x7e,0x28,0x25,0x07]
2643 ; X64-NEXT:    vpmovsqd %ymm0, (%rdi) {%k1} # encoding: [0x62,0xf2,0x7e,0x29,0x25,0x07]
2644 ; X64-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
2645 ; X64-NEXT:    retq # encoding: [0xc3]
2655 ; CHECK:       # %bb.0:
2656 ; CHECK-NEXT:    vpmovusqd %ymm0, %xmm0 # encoding: [0x62,0xf2,0x7e,0x28,0x15,0xc0]
2657 ; CHECK-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
2658 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
2665 ; X86:       # %bb.0:
2666 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
2667 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
2668 ; X86-NEXT:    vpmovusqd %ymm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7e,0x29,0x15,0xc1]
2669 ; X86-NEXT:    vmovdqa %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6f,0xc1]
2670 ; X86-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
2671 ; X86-NEXT:    retl # encoding: [0xc3]
2674 ; X64:       # %bb.0:
2675 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
2676 ; X64-NEXT:    vpmovusqd %ymm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7e,0x29,0x15,0xc1]
2677 ; X64-NEXT:    vmovdqa %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6f,0xc1]
2678 ; X64-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
2679 ; X64-NEXT:    retq # encoding: [0xc3]
2686 ; X86:       # %bb.0:
2687 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
2688 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
2689 ; X86-NEXT:    vpmovusqd %ymm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7e,0xa9,0x15,0xc0]
2690 ; X86-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
2691 ; X86-NEXT:    retl # encoding: [0xc3]
2694 ; X64:       # %bb.0:
2695 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
2696 ; X64-NEXT:    vpmovusqd %ymm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7e,0xa9,0x15,0xc0]
2697 ; X64-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
2698 ; X64-NEXT:    retq # encoding: [0xc3]
2707 ; X86:       # %bb.0:
2708 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x08]
2709 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
2710 ; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x04]
2711 ; X86-NEXT:    vpmovusqd %ymm0, (%eax) # encoding: [0x62,0xf2,0x7e,0x28,0x15,0x00]
2712 ; X86-NEXT:    vpmovusqd %ymm0, (%eax) {%k1} # encoding: [0x62,0xf2,0x7e,0x29,0x15,0x00]
2713 ; X86-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
2714 ; X86-NEXT:    retl # encoding: [0xc3]
2717 ; X64:       # %bb.0:
2718 ; X64-NEXT:    kmovw %esi, %k1 # encoding: [0xc5,0xf8,0x92,0xce]
2719 ; X64-NEXT:    vpmovusqd %ymm0, (%rdi) # encoding: [0x62,0xf2,0x7e,0x28,0x15,0x07]
2720 ; X64-NEXT:    vpmovusqd %ymm0, (%rdi) {%k1} # encoding: [0x62,0xf2,0x7e,0x29,0x15,0x07]
2721 ; X64-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
2722 ; X64-NEXT:    retq # encoding: [0xc3]
2732 ; X86:       # %bb.0:
2733 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
2734 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
2735 ; X86-NEXT:    vpmovdb %xmm0, %xmm2 # encoding: [0x62,0xf2,0x7e,0x08,0x31,0xc2]
2736 ; X86-NEXT:    vpmovdb %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7e,0x09,0x31,0xc1]
2737 ; X86-NEXT:    vpaddb %xmm1, %xmm2, %xmm1 # encoding: [0xc5,0xe9,0xfc,0xc9]
2738 ; X86-NEXT:    vpmovdb %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7e,0x89,0x31,0xc0]
2739 ; X86-NEXT:    vpaddb %xmm0, %xmm1, %xmm0 # encoding: [0xc5,0xf1,0xfc,0xc0]
2740 ; X86-NEXT:    retl # encoding: [0xc3]
2743 ; X64:       # %bb.0:
2744 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
2745 ; X64-NEXT:    vpmovdb %xmm0, %xmm2 # encoding: [0x62,0xf2,0x7e,0x08,0x31,0xc2]
2746 ; X64-NEXT:    vpmovdb %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7e,0x09,0x31,0xc1]
2747 ; X64-NEXT:    vpaddb %xmm1, %xmm2, %xmm1 # encoding: [0xc5,0xe9,0xfc,0xc9]
2748 ; X64-NEXT:    vpmovdb %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7e,0x89,0x31,0xc0]
2749 ; X64-NEXT:    vpaddb %xmm0, %xmm1, %xmm0 # encoding: [0xc5,0xf1,0xfc,0xc0]
2750 ; X64-NEXT:    retq # encoding: [0xc3]
2763 ; X86:       # %bb.0:
2764 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x08]
2765 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
2766 ; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x04]
2767 ; X86-NEXT:    vpmovdb %xmm0, (%eax) # encoding: [0x62,0xf2,0x7e,0x08,0x31,0x00]
2768 ; X86-NEXT:    vpmovdb %xmm0, (%eax) {%k1} # encoding: [0x62,0xf2,0x7e,0x09,0x31,0x00]
2769 ; X86-NEXT:    retl # encoding: [0xc3]
2772 ; X64:       # %bb.0:
2773 ; X64-NEXT:    kmovw %esi, %k1 # encoding: [0xc5,0xf8,0x92,0xce]
2774 ; X64-NEXT:    vpmovdb %xmm0, (%rdi) # encoding: [0x62,0xf2,0x7e,0x08,0x31,0x07]
2775 ; X64-NEXT:    vpmovdb %xmm0, (%rdi) {%k1} # encoding: [0x62,0xf2,0x7e,0x09,0x31,0x07]
2776 ; X64-NEXT:    retq # encoding: [0xc3]
2786 ; X86:       # %bb.0:
2787 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
2788 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
2789 ; X86-NEXT:    vpmovsdb %xmm0, %xmm2 # encoding: [0x62,0xf2,0x7e,0x08,0x21,0xc2]
2790 ; X86-NEXT:    vpmovsdb %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7e,0x09,0x21,0xc1]
2791 ; X86-NEXT:    vpaddb %xmm1, %xmm2, %xmm1 # encoding: [0xc5,0xe9,0xfc,0xc9]
2792 ; X86-NEXT:    vpmovsdb %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7e,0x89,0x21,0xc0]
2793 ; X86-NEXT:    vpaddb %xmm0, %xmm1, %xmm0 # encoding: [0xc5,0xf1,0xfc,0xc0]
2794 ; X86-NEXT:    retl # encoding: [0xc3]
2797 ; X64:       # %bb.0:
2798 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
2799 ; X64-NEXT:    vpmovsdb %xmm0, %xmm2 # encoding: [0x62,0xf2,0x7e,0x08,0x21,0xc2]
2800 ; X64-NEXT:    vpmovsdb %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7e,0x09,0x21,0xc1]
2801 ; X64-NEXT:    vpaddb %xmm1, %xmm2, %xmm1 # encoding: [0xc5,0xe9,0xfc,0xc9]
2802 ; X64-NEXT:    vpmovsdb %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7e,0x89,0x21,0xc0]
2803 ; X64-NEXT:    vpaddb %xmm0, %xmm1, %xmm0 # encoding: [0xc5,0xf1,0xfc,0xc0]
2804 ; X64-NEXT:    retq # encoding: [0xc3]
2817 ; X86:       # %bb.0:
2818 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x08]
2819 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
2820 ; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x04]
2821 ; X86-NEXT:    vpmovsdb %xmm0, (%eax) # encoding: [0x62,0xf2,0x7e,0x08,0x21,0x00]
2822 ; X86-NEXT:    vpmovsdb %xmm0, (%eax) {%k1} # encoding: [0x62,0xf2,0x7e,0x09,0x21,0x00]
2823 ; X86-NEXT:    retl # encoding: [0xc3]
2826 ; X64:       # %bb.0:
2827 ; X64-NEXT:    kmovw %esi, %k1 # encoding: [0xc5,0xf8,0x92,0xce]
2828 ; X64-NEXT:    vpmovsdb %xmm0, (%rdi) # encoding: [0x62,0xf2,0x7e,0x08,0x21,0x07]
2829 ; X64-NEXT:    vpmovsdb %xmm0, (%rdi) {%k1} # encoding: [0x62,0xf2,0x7e,0x09,0x21,0x07]
2830 ; X64-NEXT:    retq # encoding: [0xc3]
2840 ; X86:       # %bb.0:
2841 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
2842 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
2843 ; X86-NEXT:    vpmovusdb %xmm0, %xmm2 # encoding: [0x62,0xf2,0x7e,0x08,0x11,0xc2]
2844 ; X86-NEXT:    vpmovusdb %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7e,0x09,0x11,0xc1]
2845 ; X86-NEXT:    vpaddb %xmm1, %xmm2, %xmm1 # encoding: [0xc5,0xe9,0xfc,0xc9]
2846 ; X86-NEXT:    vpmovusdb %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7e,0x89,0x11,0xc0]
2847 ; X86-NEXT:    vpaddb %xmm0, %xmm1, %xmm0 # encoding: [0xc5,0xf1,0xfc,0xc0]
2848 ; X86-NEXT:    retl # encoding: [0xc3]
2851 ; X64:       # %bb.0:
2852 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
2853 ; X64-NEXT:    vpmovusdb %xmm0, %xmm2 # encoding: [0x62,0xf2,0x7e,0x08,0x11,0xc2]
2854 ; X64-NEXT:    vpmovusdb %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7e,0x09,0x11,0xc1]
2855 ; X64-NEXT:    vpaddb %xmm1, %xmm2, %xmm1 # encoding: [0xc5,0xe9,0xfc,0xc9]
2856 ; X64-NEXT:    vpmovusdb %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7e,0x89,0x11,0xc0]
2857 ; X64-NEXT:    vpaddb %xmm0, %xmm1, %xmm0 # encoding: [0xc5,0xf1,0xfc,0xc0]
2858 ; X64-NEXT:    retq # encoding: [0xc3]
2871 ; X86:       # %bb.0:
2872 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x08]
2873 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
2874 ; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x04]
2875 ; X86-NEXT:    vpmovusdb %xmm0, (%eax) # encoding: [0x62,0xf2,0x7e,0x08,0x11,0x00]
2876 ; X86-NEXT:    vpmovusdb %xmm0, (%eax) {%k1} # encoding: [0x62,0xf2,0x7e,0x09,0x11,0x00]
2877 ; X86-NEXT:    retl # encoding: [0xc3]
2880 ; X64:       # %bb.0:
2881 ; X64-NEXT:    kmovw %esi, %k1 # encoding: [0xc5,0xf8,0x92,0xce]
2882 ; X64-NEXT:    vpmovusdb %xmm0, (%rdi) # encoding: [0x62,0xf2,0x7e,0x08,0x11,0x07]
2883 ; X64-NEXT:    vpmovusdb %xmm0, (%rdi) {%k1} # encoding: [0x62,0xf2,0x7e,0x09,0x11,0x07]
2884 ; X64-NEXT:    retq # encoding: [0xc3]
2894 ; X86:       # %bb.0:
2895 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
2896 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
2897 ; X86-NEXT:    vpmovdb %ymm0, %xmm2 # encoding: [0x62,0xf2,0x7e,0x28,0x31,0xc2]
2898 ; X86-NEXT:    vpmovdb %ymm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7e,0x29,0x31,0xc1]
2899 ; X86-NEXT:    vpaddb %xmm1, %xmm2, %xmm1 # encoding: [0xc5,0xe9,0xfc,0xc9]
2900 ; X86-NEXT:    vpmovdb %ymm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7e,0xa9,0x31,0xc0]
2901 ; X86-NEXT:    vpaddb %xmm0, %xmm1, %xmm0 # encoding: [0xc5,0xf1,0xfc,0xc0]
2902 ; X86-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
2903 ; X86-NEXT:    retl # encoding: [0xc3]
2906 ; X64:       # %bb.0:
2907 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
2908 ; X64-NEXT:    vpmovdb %ymm0, %xmm2 # encoding: [0x62,0xf2,0x7e,0x28,0x31,0xc2]
2909 ; X64-NEXT:    vpmovdb %ymm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7e,0x29,0x31,0xc1]
2910 ; X64-NEXT:    vpaddb %xmm1, %xmm2, %xmm1 # encoding: [0xc5,0xe9,0xfc,0xc9]
2911 ; X64-NEXT:    vpmovdb %ymm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7e,0xa9,0x31,0xc0]
2912 ; X64-NEXT:    vpaddb %xmm0, %xmm1, %xmm0 # encoding: [0xc5,0xf1,0xfc,0xc0]
2913 ; X64-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
2914 ; X64-NEXT:    retq # encoding: [0xc3]
2927 ; X86:       # %bb.0:
2928 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x08]
2929 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
2930 ; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x04]
2931 ; X86-NEXT:    vpmovdb %ymm0, (%eax) # encoding: [0x62,0xf2,0x7e,0x28,0x31,0x00]
2932 ; X86-NEXT:    vpmovdb %ymm0, (%eax) {%k1} # encoding: [0x62,0xf2,0x7e,0x29,0x31,0x00]
2933 ; X86-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
2934 ; X86-NEXT:    retl # encoding: [0xc3]
2937 ; X64:       # %bb.0:
2938 ; X64-NEXT:    kmovw %esi, %k1 # encoding: [0xc5,0xf8,0x92,0xce]
2939 ; X64-NEXT:    vpmovdb %ymm0, (%rdi) # encoding: [0x62,0xf2,0x7e,0x28,0x31,0x07]
2940 ; X64-NEXT:    vpmovdb %ymm0, (%rdi) {%k1} # encoding: [0x62,0xf2,0x7e,0x29,0x31,0x07]
2941 ; X64-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
2942 ; X64-NEXT:    retq # encoding: [0xc3]
2952 ; X86:       # %bb.0:
2953 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
2954 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
2955 ; X86-NEXT:    vpmovsdb %ymm0, %xmm2 # encoding: [0x62,0xf2,0x7e,0x28,0x21,0xc2]
2956 ; X86-NEXT:    vpmovsdb %ymm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7e,0x29,0x21,0xc1]
2957 ; X86-NEXT:    vpaddb %xmm1, %xmm2, %xmm1 # encoding: [0xc5,0xe9,0xfc,0xc9]
2958 ; X86-NEXT:    vpmovsdb %ymm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7e,0xa9,0x21,0xc0]
2959 ; X86-NEXT:    vpaddb %xmm0, %xmm1, %xmm0 # encoding: [0xc5,0xf1,0xfc,0xc0]
2960 ; X86-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
2961 ; X86-NEXT:    retl # encoding: [0xc3]
2964 ; X64:       # %bb.0:
2965 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
2966 ; X64-NEXT:    vpmovsdb %ymm0, %xmm2 # encoding: [0x62,0xf2,0x7e,0x28,0x21,0xc2]
2967 ; X64-NEXT:    vpmovsdb %ymm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7e,0x29,0x21,0xc1]
2968 ; X64-NEXT:    vpaddb %xmm1, %xmm2, %xmm1 # encoding: [0xc5,0xe9,0xfc,0xc9]
2969 ; X64-NEXT:    vpmovsdb %ymm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7e,0xa9,0x21,0xc0]
2970 ; X64-NEXT:    vpaddb %xmm0, %xmm1, %xmm0 # encoding: [0xc5,0xf1,0xfc,0xc0]
2971 ; X64-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
2972 ; X64-NEXT:    retq # encoding: [0xc3]
2985 ; X86:       # %bb.0:
2986 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x08]
2987 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
2988 ; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x04]
2989 ; X86-NEXT:    vpmovsdb %ymm0, (%eax) # encoding: [0x62,0xf2,0x7e,0x28,0x21,0x00]
2990 ; X86-NEXT:    vpmovsdb %ymm0, (%eax) {%k1} # encoding: [0x62,0xf2,0x7e,0x29,0x21,0x00]
2991 ; X86-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
2992 ; X86-NEXT:    retl # encoding: [0xc3]
2995 ; X64:       # %bb.0:
2996 ; X64-NEXT:    kmovw %esi, %k1 # encoding: [0xc5,0xf8,0x92,0xce]
2997 ; X64-NEXT:    vpmovsdb %ymm0, (%rdi) # encoding: [0x62,0xf2,0x7e,0x28,0x21,0x07]
2998 ; X64-NEXT:    vpmovsdb %ymm0, (%rdi) {%k1} # encoding: [0x62,0xf2,0x7e,0x29,0x21,0x07]
2999 ; X64-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
3000 ; X64-NEXT:    retq # encoding: [0xc3]
3010 ; X86:       # %bb.0:
3011 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
3012 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
3013 ; X86-NEXT:    vpmovusdb %ymm0, %xmm2 # encoding: [0x62,0xf2,0x7e,0x28,0x11,0xc2]
3014 ; X86-NEXT:    vpmovusdb %ymm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7e,0x29,0x11,0xc1]
3015 ; X86-NEXT:    vpaddb %xmm1, %xmm2, %xmm1 # encoding: [0xc5,0xe9,0xfc,0xc9]
3016 ; X86-NEXT:    vpmovusdb %ymm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7e,0xa9,0x11,0xc0]
3017 ; X86-NEXT:    vpaddb %xmm0, %xmm1, %xmm0 # encoding: [0xc5,0xf1,0xfc,0xc0]
3018 ; X86-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
3019 ; X86-NEXT:    retl # encoding: [0xc3]
3022 ; X64:       # %bb.0:
3023 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
3024 ; X64-NEXT:    vpmovusdb %ymm0, %xmm2 # encoding: [0x62,0xf2,0x7e,0x28,0x11,0xc2]
3025 ; X64-NEXT:    vpmovusdb %ymm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7e,0x29,0x11,0xc1]
3026 ; X64-NEXT:    vpaddb %xmm1, %xmm2, %xmm1 # encoding: [0xc5,0xe9,0xfc,0xc9]
3027 ; X64-NEXT:    vpmovusdb %ymm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7e,0xa9,0x11,0xc0]
3028 ; X64-NEXT:    vpaddb %xmm0, %xmm1, %xmm0 # encoding: [0xc5,0xf1,0xfc,0xc0]
3029 ; X64-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
3030 ; X64-NEXT:    retq # encoding: [0xc3]
3043 ; X86:       # %bb.0:
3044 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x08]
3045 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
3046 ; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x04]
3047 ; X86-NEXT:    vpmovusdb %ymm0, (%eax) # encoding: [0x62,0xf2,0x7e,0x28,0x11,0x00]
3048 ; X86-NEXT:    vpmovusdb %ymm0, (%eax) {%k1} # encoding: [0x62,0xf2,0x7e,0x29,0x11,0x00]
3049 ; X86-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
3050 ; X86-NEXT:    retl # encoding: [0xc3]
3053 ; X64:       # %bb.0:
3054 ; X64-NEXT:    kmovw %esi, %k1 # encoding: [0xc5,0xf8,0x92,0xce]
3055 ; X64-NEXT:    vpmovusdb %ymm0, (%rdi) # encoding: [0x62,0xf2,0x7e,0x28,0x11,0x07]
3056 ; X64-NEXT:    vpmovusdb %ymm0, (%rdi) {%k1} # encoding: [0x62,0xf2,0x7e,0x29,0x11,0x07]
3057 ; X64-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
3058 ; X64-NEXT:    retq # encoding: [0xc3]
3068 ; X86:       # %bb.0:
3069 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
3070 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
3071 ; X86-NEXT:    vpmovdw %xmm0, %xmm2 # encoding: [0x62,0xf2,0x7e,0x08,0x33,0xc2]
3072 ; X86-NEXT:    vpmovdw %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7e,0x09,0x33,0xc1]
3073 ; X86-NEXT:    vpaddw %xmm1, %xmm2, %xmm1 # encoding: [0xc5,0xe9,0xfd,0xc9]
3074 ; X86-NEXT:    vpmovdw %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7e,0x89,0x33,0xc0]
3075 ; X86-NEXT:    vpaddw %xmm0, %xmm1, %xmm0 # encoding: [0xc5,0xf1,0xfd,0xc0]
3076 ; X86-NEXT:    retl # encoding: [0xc3]
3079 ; X64:       # %bb.0:
3080 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
3081 ; X64-NEXT:    vpmovdw %xmm0, %xmm2 # encoding: [0x62,0xf2,0x7e,0x08,0x33,0xc2]
3082 ; X64-NEXT:    vpmovdw %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7e,0x09,0x33,0xc1]
3083 ; X64-NEXT:    vpaddw %xmm1, %xmm2, %xmm1 # encoding: [0xc5,0xe9,0xfd,0xc9]
3084 ; X64-NEXT:    vpmovdw %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7e,0x89,0x33,0xc0]
3085 ; X64-NEXT:    vpaddw %xmm0, %xmm1, %xmm0 # encoding: [0xc5,0xf1,0xfd,0xc0]
3086 ; X64-NEXT:    retq # encoding: [0xc3]
3099 ; X86:       # %bb.0:
3100 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x08]
3101 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
3102 ; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x04]
3103 ; X86-NEXT:    vpmovdw %xmm0, (%eax) # encoding: [0x62,0xf2,0x7e,0x08,0x33,0x00]
3104 ; X86-NEXT:    vpmovdw %xmm0, (%eax) {%k1} # encoding: [0x62,0xf2,0x7e,0x09,0x33,0x00]
3105 ; X86-NEXT:    retl # encoding: [0xc3]
3108 ; X64:       # %bb.0:
3109 ; X64-NEXT:    kmovw %esi, %k1 # encoding: [0xc5,0xf8,0x92,0xce]
3110 ; X64-NEXT:    vpmovdw %xmm0, (%rdi) # encoding: [0x62,0xf2,0x7e,0x08,0x33,0x07]
3111 ; X64-NEXT:    vpmovdw %xmm0, (%rdi) {%k1} # encoding: [0x62,0xf2,0x7e,0x09,0x33,0x07]
3112 ; X64-NEXT:    retq # encoding: [0xc3]
3122 ; X86:       # %bb.0:
3123 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
3124 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
3125 ; X86-NEXT:    vpmovsdw %xmm0, %xmm2 # encoding: [0x62,0xf2,0x7e,0x08,0x23,0xc2]
3126 ; X86-NEXT:    vpmovsdw %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7e,0x09,0x23,0xc1]
3127 ; X86-NEXT:    vpaddw %xmm1, %xmm2, %xmm1 # encoding: [0xc5,0xe9,0xfd,0xc9]
3128 ; X86-NEXT:    vpmovsdw %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7e,0x89,0x23,0xc0]
3129 ; X86-NEXT:    vpaddw %xmm0, %xmm1, %xmm0 # encoding: [0xc5,0xf1,0xfd,0xc0]
3130 ; X86-NEXT:    retl # encoding: [0xc3]
3133 ; X64:       # %bb.0:
3134 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
3135 ; X64-NEXT:    vpmovsdw %xmm0, %xmm2 # encoding: [0x62,0xf2,0x7e,0x08,0x23,0xc2]
3136 ; X64-NEXT:    vpmovsdw %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7e,0x09,0x23,0xc1]
3137 ; X64-NEXT:    vpaddw %xmm1, %xmm2, %xmm1 # encoding: [0xc5,0xe9,0xfd,0xc9]
3138 ; X64-NEXT:    vpmovsdw %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7e,0x89,0x23,0xc0]
3139 ; X64-NEXT:    vpaddw %xmm0, %xmm1, %xmm0 # encoding: [0xc5,0xf1,0xfd,0xc0]
3140 ; X64-NEXT:    retq # encoding: [0xc3]
3153 ; X86:       # %bb.0:
3154 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x08]
3155 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
3156 ; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x04]
3157 ; X86-NEXT:    vpmovsdw %xmm0, (%eax) # encoding: [0x62,0xf2,0x7e,0x08,0x23,0x00]
3158 ; X86-NEXT:    vpmovsdw %xmm0, (%eax) {%k1} # encoding: [0x62,0xf2,0x7e,0x09,0x23,0x00]
3159 ; X86-NEXT:    retl # encoding: [0xc3]
3162 ; X64:       # %bb.0:
3163 ; X64-NEXT:    kmovw %esi, %k1 # encoding: [0xc5,0xf8,0x92,0xce]
3164 ; X64-NEXT:    vpmovsdw %xmm0, (%rdi) # encoding: [0x62,0xf2,0x7e,0x08,0x23,0x07]
3165 ; X64-NEXT:    vpmovsdw %xmm0, (%rdi) {%k1} # encoding: [0x62,0xf2,0x7e,0x09,0x23,0x07]
3166 ; X64-NEXT:    retq # encoding: [0xc3]
3176 ; X86:       # %bb.0:
3177 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
3178 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
3179 ; X86-NEXT:    vpmovusdw %xmm0, %xmm2 # encoding: [0x62,0xf2,0x7e,0x08,0x13,0xc2]
3180 ; X86-NEXT:    vpmovusdw %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7e,0x09,0x13,0xc1]
3181 ; X86-NEXT:    vpaddw %xmm1, %xmm2, %xmm1 # encoding: [0xc5,0xe9,0xfd,0xc9]
3182 ; X86-NEXT:    vpmovusdw %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7e,0x89,0x13,0xc0]
3183 ; X86-NEXT:    vpaddw %xmm0, %xmm1, %xmm0 # encoding: [0xc5,0xf1,0xfd,0xc0]
3184 ; X86-NEXT:    retl # encoding: [0xc3]
3187 ; X64:       # %bb.0:
3188 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
3189 ; X64-NEXT:    vpmovusdw %xmm0, %xmm2 # encoding: [0x62,0xf2,0x7e,0x08,0x13,0xc2]
3190 ; X64-NEXT:    vpmovusdw %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7e,0x09,0x13,0xc1]
3191 ; X64-NEXT:    vpaddw %xmm1, %xmm2, %xmm1 # encoding: [0xc5,0xe9,0xfd,0xc9]
3192 ; X64-NEXT:    vpmovusdw %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7e,0x89,0x13,0xc0]
3193 ; X64-NEXT:    vpaddw %xmm0, %xmm1, %xmm0 # encoding: [0xc5,0xf1,0xfd,0xc0]
3194 ; X64-NEXT:    retq # encoding: [0xc3]
3207 ; X86:       # %bb.0:
3208 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x08]
3209 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
3210 ; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x04]
3211 ; X86-NEXT:    vpmovusdw %xmm0, (%eax) # encoding: [0x62,0xf2,0x7e,0x08,0x13,0x00]
3212 ; X86-NEXT:    vpmovusdw %xmm0, (%eax) {%k1} # encoding: [0x62,0xf2,0x7e,0x09,0x13,0x00]
3213 ; X86-NEXT:    retl # encoding: [0xc3]
3216 ; X64:       # %bb.0:
3217 ; X64-NEXT:    kmovw %esi, %k1 # encoding: [0xc5,0xf8,0x92,0xce]
3218 ; X64-NEXT:    vpmovusdw %xmm0, (%rdi) # encoding: [0x62,0xf2,0x7e,0x08,0x13,0x07]
3219 ; X64-NEXT:    vpmovusdw %xmm0, (%rdi) {%k1} # encoding: [0x62,0xf2,0x7e,0x09,0x13,0x07]
3220 ; X64-NEXT:    retq # encoding: [0xc3]
3230 ; X86:       # %bb.0:
3231 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
3232 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
3233 ; X86-NEXT:    vpmovdw %ymm0, %xmm2 # encoding: [0x62,0xf2,0x7e,0x28,0x33,0xc2]
3234 ; X86-NEXT:    vpmovdw %ymm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7e,0x29,0x33,0xc1]
3235 ; X86-NEXT:    vpaddw %xmm1, %xmm2, %xmm1 # encoding: [0xc5,0xe9,0xfd,0xc9]
3236 ; X86-NEXT:    vpmovdw %ymm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7e,0xa9,0x33,0xc0]
3237 ; X86-NEXT:    vpaddw %xmm0, %xmm1, %xmm0 # encoding: [0xc5,0xf1,0xfd,0xc0]
3238 ; X86-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
3239 ; X86-NEXT:    retl # encoding: [0xc3]
3242 ; X64:       # %bb.0:
3243 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
3244 ; X64-NEXT:    vpmovdw %ymm0, %xmm2 # encoding: [0x62,0xf2,0x7e,0x28,0x33,0xc2]
3245 ; X64-NEXT:    vpmovdw %ymm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7e,0x29,0x33,0xc1]
3246 ; X64-NEXT:    vpaddw %xmm1, %xmm2, %xmm1 # encoding: [0xc5,0xe9,0xfd,0xc9]
3247 ; X64-NEXT:    vpmovdw %ymm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7e,0xa9,0x33,0xc0]
3248 ; X64-NEXT:    vpaddw %xmm0, %xmm1, %xmm0 # encoding: [0xc5,0xf1,0xfd,0xc0]
3249 ; X64-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
3250 ; X64-NEXT:    retq # encoding: [0xc3]
3263 ; X86:       # %bb.0:
3264 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x08]
3265 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
3266 ; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x04]
3267 ; X86-NEXT:    vpmovdw %ymm0, (%eax) # encoding: [0x62,0xf2,0x7e,0x28,0x33,0x00]
3268 ; X86-NEXT:    vpmovdw %ymm0, (%eax) {%k1} # encoding: [0x62,0xf2,0x7e,0x29,0x33,0x00]
3269 ; X86-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
3270 ; X86-NEXT:    retl # encoding: [0xc3]
3273 ; X64:       # %bb.0:
3274 ; X64-NEXT:    kmovw %esi, %k1 # encoding: [0xc5,0xf8,0x92,0xce]
3275 ; X64-NEXT:    vpmovdw %ymm0, (%rdi) # encoding: [0x62,0xf2,0x7e,0x28,0x33,0x07]
3276 ; X64-NEXT:    vpmovdw %ymm0, (%rdi) {%k1} # encoding: [0x62,0xf2,0x7e,0x29,0x33,0x07]
3277 ; X64-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
3278 ; X64-NEXT:    retq # encoding: [0xc3]
3288 ; X86:       # %bb.0:
3289 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
3290 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
3291 ; X86-NEXT:    vpmovsdw %ymm0, %xmm2 # encoding: [0x62,0xf2,0x7e,0x28,0x23,0xc2]
3292 ; X86-NEXT:    vpmovsdw %ymm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7e,0x29,0x23,0xc1]
3293 ; X86-NEXT:    vpaddw %xmm1, %xmm2, %xmm1 # encoding: [0xc5,0xe9,0xfd,0xc9]
3294 ; X86-NEXT:    vpmovsdw %ymm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7e,0xa9,0x23,0xc0]
3295 ; X86-NEXT:    vpaddw %xmm0, %xmm1, %xmm0 # encoding: [0xc5,0xf1,0xfd,0xc0]
3296 ; X86-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
3297 ; X86-NEXT:    retl # encoding: [0xc3]
3300 ; X64:       # %bb.0:
3301 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
3302 ; X64-NEXT:    vpmovsdw %ymm0, %xmm2 # encoding: [0x62,0xf2,0x7e,0x28,0x23,0xc2]
3303 ; X64-NEXT:    vpmovsdw %ymm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7e,0x29,0x23,0xc1]
3304 ; X64-NEXT:    vpaddw %xmm1, %xmm2, %xmm1 # encoding: [0xc5,0xe9,0xfd,0xc9]
3305 ; X64-NEXT:    vpmovsdw %ymm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7e,0xa9,0x23,0xc0]
3306 ; X64-NEXT:    vpaddw %xmm0, %xmm1, %xmm0 # encoding: [0xc5,0xf1,0xfd,0xc0]
3307 ; X64-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
3308 ; X64-NEXT:    retq # encoding: [0xc3]
3321 ; X86:       # %bb.0:
3322 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x08]
3323 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
3324 ; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x04]
3325 ; X86-NEXT:    vpmovsdw %ymm0, (%eax) # encoding: [0x62,0xf2,0x7e,0x28,0x23,0x00]
3326 ; X86-NEXT:    vpmovsdw %ymm0, (%eax) {%k1} # encoding: [0x62,0xf2,0x7e,0x29,0x23,0x00]
3327 ; X86-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
3328 ; X86-NEXT:    retl # encoding: [0xc3]
3331 ; X64:       # %bb.0:
3332 ; X64-NEXT:    kmovw %esi, %k1 # encoding: [0xc5,0xf8,0x92,0xce]
3333 ; X64-NEXT:    vpmovsdw %ymm0, (%rdi) # encoding: [0x62,0xf2,0x7e,0x28,0x23,0x07]
3334 ; X64-NEXT:    vpmovsdw %ymm0, (%rdi) {%k1} # encoding: [0x62,0xf2,0x7e,0x29,0x23,0x07]
3335 ; X64-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
3336 ; X64-NEXT:    retq # encoding: [0xc3]
3346 ; X86:       # %bb.0:
3347 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
3348 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
3349 ; X86-NEXT:    vpmovusdw %ymm0, %xmm2 # encoding: [0x62,0xf2,0x7e,0x28,0x13,0xc2]
3350 ; X86-NEXT:    vpmovusdw %ymm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7e,0x29,0x13,0xc1]
3351 ; X86-NEXT:    vpaddw %xmm1, %xmm2, %xmm1 # encoding: [0xc5,0xe9,0xfd,0xc9]
3352 ; X86-NEXT:    vpmovusdw %ymm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7e,0xa9,0x13,0xc0]
3353 ; X86-NEXT:    vpaddw %xmm0, %xmm1, %xmm0 # encoding: [0xc5,0xf1,0xfd,0xc0]
3354 ; X86-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
3355 ; X86-NEXT:    retl # encoding: [0xc3]
3358 ; X64:       # %bb.0:
3359 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
3360 ; X64-NEXT:    vpmovusdw %ymm0, %xmm2 # encoding: [0x62,0xf2,0x7e,0x28,0x13,0xc2]
3361 ; X64-NEXT:    vpmovusdw %ymm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7e,0x29,0x13,0xc1]
3362 ; X64-NEXT:    vpaddw %xmm1, %xmm2, %xmm1 # encoding: [0xc5,0xe9,0xfd,0xc9]
3363 ; X64-NEXT:    vpmovusdw %ymm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7e,0xa9,0x13,0xc0]
3364 ; X64-NEXT:    vpaddw %xmm0, %xmm1, %xmm0 # encoding: [0xc5,0xf1,0xfd,0xc0]
3365 ; X64-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
3366 ; X64-NEXT:    retq # encoding: [0xc3]
3379 ; X86:       # %bb.0:
3380 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x08]
3381 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
3382 ; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x04]
3383 ; X86-NEXT:    vpmovusdw %ymm0, (%eax) # encoding: [0x62,0xf2,0x7e,0x28,0x13,0x00]
3384 ; X86-NEXT:    vpmovusdw %ymm0, (%eax) {%k1} # encoding: [0x62,0xf2,0x7e,0x29,0x13,0x00]
3385 ; X86-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
3386 ; X86-NEXT:    retl # encoding: [0xc3]
3389 ; X64:       # %bb.0:
3390 ; X64-NEXT:    kmovw %esi, %k1 # encoding: [0xc5,0xf8,0x92,0xce]
3391 ; X64-NEXT:    vpmovusdw %ymm0, (%rdi) # encoding: [0x62,0xf2,0x7e,0x28,0x13,0x07]
3392 ; X64-NEXT:    vpmovusdw %ymm0, (%rdi) {%k1} # encoding: [0x62,0xf2,0x7e,0x29,0x13,0x07]
3393 ; X64-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
3394 ; X64-NEXT:    retq # encoding: [0xc3]
3404 ; X86:       # %bb.0:
3405 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
3406 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
3407 ; X86-NEXT:    vcvtpd2dq %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf1,0xff,0x09,0xe6,0xc8]
3408 ; X86-NEXT:    vcvtpd2dq %xmm0, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xfb,0xe6,0xc0]
3409 ; X86-NEXT:    vpaddd %xmm0, %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf1,0xfe,0xc0]
3410 ; X86-NEXT:    retl # encoding: [0xc3]
3413 ; X64:       # %bb.0:
3414 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
3415 ; X64-NEXT:    vcvtpd2dq %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf1,0xff,0x09,0xe6,0xc8]
3416 ; X64-NEXT:    vcvtpd2dq %xmm0, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xfb,0xe6,0xc0]
3417 ; X64-NEXT:    vpaddd %xmm0, %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf1,0xfe,0xc0]
3418 ; X64-NEXT:    retq # encoding: [0xc3]
3427 ; CHECK:       # %bb.0:
3428 ; CHECK-NEXT:    vcvtpd2dq %xmm0, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xfb,0xe6,0xc0]
3429 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
3431   %res3 = shufflevector <4 x i32> %res2, <4 x i32> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 4, i32 5>
3437 ; X86:       # %bb.0:
3438 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
3439 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
3440 ; X86-NEXT:    vcvtpd2dq %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf1,0xff,0x09,0xe6,0xc8]
3441 ; X86-NEXT:    vmovaps %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x28,0xc1]
3442 ; X86-NEXT:    retl # encoding: [0xc3]
3445 ; X64:       # %bb.0:
3446 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
3447 ; X64-NEXT:    vcvtpd2dq %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf1,0xff,0x09,0xe6,0xc8]
3448 ; X64-NEXT:    vmovaps %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x28,0xc1]
3449 ; X64-NEXT:    retq # encoding: [0xc3]
3451   %res1 = shufflevector <4 x i32> %res, <4 x i32> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 4, i32 5>
3459 ; CHECK:       # %bb.0:
3460 ; CHECK-NEXT:    vcvtpd2ps %xmm0, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf9,0x5a,0xc0]
3461 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
3468 ; X86:       # %bb.0:
3469 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
3470 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
3471 ; X86-NEXT:    vcvtpd2ps %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf1,0xfd,0x09,0x5a,0xc8]
3472 ; X86-NEXT:    vmovaps %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x28,0xc1]
3473 ; X86-NEXT:    retl # encoding: [0xc3]
3476 ; X64:       # %bb.0:
3477 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
3478 ; X64-NEXT:    vcvtpd2ps %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf1,0xfd,0x09,0x5a,0xc8]
3479 ; X64-NEXT:    vmovaps %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x28,0xc1]
3480 ; X64-NEXT:    retq # encoding: [0xc3]
3487 ; CHECK:       # %bb.0:
3488 ; CHECK-NEXT:    vcvtpd2ps %xmm0, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf9,0x5a,0xc0]
3489 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
3491   %res3 = shufflevector <4 x float> %res2, <4 x float> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 4, i32 5>
3497 ; X86:       # %bb.0:
3498 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
3499 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
3500 ; X86-NEXT:    vcvtpd2ps %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf1,0xfd,0x09,0x5a,0xc8]
3501 ; X86-NEXT:    vmovaps %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x28,0xc1]
3502 ; X86-NEXT:    retl # encoding: [0xc3]
3505 ; X64:       # %bb.0:
3506 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
3507 ; X64-NEXT:    vcvtpd2ps %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf1,0xfd,0x09,0x5a,0xc8]
3508 ; X64-NEXT:    vmovaps %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x28,0xc1]
3509 ; X64-NEXT:    retq # encoding: [0xc3]
3511   %res1 = shufflevector <4 x float> %res, <4 x float> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 4, i32 5>
3519 ; CHECK:       # %bb.0:
3520 ; CHECK-NEXT:    vcvtpd2udq %xmm0, %xmm0 # encoding: [0x62,0xf1,0xfc,0x08,0x79,0xc0]
3521 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
3528 ; X86:       # %bb.0:
3529 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
3530 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
3531 ; X86-NEXT:    vcvtpd2udq %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf1,0xfc,0x09,0x79,0xc8]
3532 ; X86-NEXT:    vmovaps %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x28,0xc1]
3533 ; X86-NEXT:    retl # encoding: [0xc3]
3536 ; X64:       # %bb.0:
3537 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
3538 ; X64-NEXT:    vcvtpd2udq %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf1,0xfc,0x09,0x79,0xc8]
3539 ; X64-NEXT:    vmovaps %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x28,0xc1]
3540 ; X64-NEXT:    retq # encoding: [0xc3]
3547 ; CHECK:       # %bb.0:
3548 ; CHECK-NEXT:    vcvtpd2udq %xmm0, %xmm0 # encoding: [0x62,0xf1,0xfc,0x08,0x79,0xc0]
3549 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
3551   %res3 = shufflevector <4 x i32> %res2, <4 x i32> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 4, i32 5>
3557 ; X86:       # %bb.0:
3558 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
3559 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
3560 ; X86-NEXT:    vcvtpd2udq %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf1,0xfc,0x09,0x79,0xc8]
3561 ; X86-NEXT:    vmovaps %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x28,0xc1]
3562 ; X86-NEXT:    retl # encoding: [0xc3]
3565 ; X64:       # %bb.0:
3566 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
3567 ; X64-NEXT:    vcvtpd2udq %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf1,0xfc,0x09,0x79,0xc8]
3568 ; X64-NEXT:    vmovaps %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x28,0xc1]
3569 ; X64-NEXT:    retq # encoding: [0xc3]
3571   %res1 = shufflevector <4 x i32> %res, <4 x i32> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 4, i32 5>
3579 ; CHECK:       # %bb.0:
3580 ; CHECK-NEXT:    vcvtpd2udq %ymm0, %xmm0 # encoding: [0x62,0xf1,0xfc,0x28,0x79,0xc0]
3581 ; CHECK-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
3582 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
3589 ; X86:       # %bb.0:
3590 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
3591 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
3592 ; X86-NEXT:    vcvtpd2udq %ymm0, %xmm1 {%k1} # encoding: [0x62,0xf1,0xfc,0x29,0x79,0xc8]
3593 ; X86-NEXT:    vmovaps %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x28,0xc1]
3594 ; X86-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
3595 ; X86-NEXT:    retl # encoding: [0xc3]
3598 ; X64:       # %bb.0:
3599 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
3600 ; X64-NEXT:    vcvtpd2udq %ymm0, %xmm1 {%k1} # encoding: [0x62,0xf1,0xfc,0x29,0x79,0xc8]
3601 ; X64-NEXT:    vmovaps %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x28,0xc1]
3602 ; X64-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
3603 ; X64-NEXT:    retq # encoding: [0xc3]
3612 ; CHECK:       # %bb.0:
3613 ; CHECK-NEXT:    vcvtps2dq %xmm0, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf9,0x5b,0xc0]
3614 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
3621 ; X86:       # %bb.0:
3622 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
3623 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
3624 ; X86-NEXT:    vcvtps2dq %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf1,0x7d,0x09,0x5b,0xc8]
3625 ; X86-NEXT:    vmovaps %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x28,0xc1]
3626 ; X86-NEXT:    retl # encoding: [0xc3]
3629 ; X64:       # %bb.0:
3630 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
3631 ; X64-NEXT:    vcvtps2dq %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf1,0x7d,0x09,0x5b,0xc8]
3632 ; X64-NEXT:    vmovaps %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x28,0xc1]
3633 ; X64-NEXT:    retq # encoding: [0xc3]
3642 ; CHECK:       # %bb.0:
3643 ; CHECK-NEXT:    vcvtps2dq %ymm0, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfd,0x5b,0xc0]
3644 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
3651 ; X86:       # %bb.0:
3652 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
3653 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
3654 ; X86-NEXT:    vcvtps2dq %ymm0, %ymm1 {%k1} # encoding: [0x62,0xf1,0x7d,0x29,0x5b,0xc8]
3655 ; X86-NEXT:    vmovaps %ymm1, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfc,0x28,0xc1]
3656 ; X86-NEXT:    retl # encoding: [0xc3]
3659 ; X64:       # %bb.0:
3660 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
3661 ; X64-NEXT:    vcvtps2dq %ymm0, %ymm1 {%k1} # encoding: [0x62,0xf1,0x7d,0x29,0x5b,0xc8]
3662 ; X64-NEXT:    vmovaps %ymm1, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfc,0x28,0xc1]
3663 ; X64-NEXT:    retq # encoding: [0xc3]
3672 ; CHECK:       # %bb.0:
3673 ; CHECK-NEXT:    vcvtps2udq %xmm0, %xmm0 # encoding: [0x62,0xf1,0x7c,0x08,0x79,0xc0]
3674 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
3681 ; X86:       # %bb.0:
3682 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
3683 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
3684 ; X86-NEXT:    vcvtps2udq %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf1,0x7c,0x09,0x79,0xc8]
3685 ; X86-NEXT:    vmovaps %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x28,0xc1]
3686 ; X86-NEXT:    retl # encoding: [0xc3]
3689 ; X64:       # %bb.0:
3690 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
3691 ; X64-NEXT:    vcvtps2udq %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf1,0x7c,0x09,0x79,0xc8]
3692 ; X64-NEXT:    vmovaps %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x28,0xc1]
3693 ; X64-NEXT:    retq # encoding: [0xc3]
3702 ; CHECK:       # %bb.0:
3703 ; CHECK-NEXT:    vcvtps2udq %ymm0, %ymm0 # encoding: [0x62,0xf1,0x7c,0x28,0x79,0xc0]
3704 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
3711 ; X86:       # %bb.0:
3712 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
3713 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
3714 ; X86-NEXT:    vcvtps2udq %ymm0, %ymm1 {%k1} # encoding: [0x62,0xf1,0x7c,0x29,0x79,0xc8]
3715 ; X86-NEXT:    vmovaps %ymm1, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfc,0x28,0xc1]
3716 ; X86-NEXT:    retl # encoding: [0xc3]
3719 ; X64:       # %bb.0:
3720 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
3721 ; X64-NEXT:    vcvtps2udq %ymm0, %ymm1 {%k1} # encoding: [0x62,0xf1,0x7c,0x29,0x79,0xc8]
3722 ; X64-NEXT:    vmovaps %ymm1, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfc,0x28,0xc1]
3723 ; X64-NEXT:    retq # encoding: [0xc3]
3732 ; CHECK:       # %bb.0:
3733 ; CHECK-NEXT:    vcvttpd2dq %xmm0, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf9,0xe6,0xc0]
3734 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
3741 ; X86:       # %bb.0:
3742 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
3743 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
3744 ; X86-NEXT:    vcvttpd2dq %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf1,0xfd,0x09,0xe6,0xc8]
3745 ; X86-NEXT:    vmovaps %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x28,0xc1]
3746 ; X86-NEXT:    retl # encoding: [0xc3]
3749 ; X64:       # %bb.0:
3750 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
3751 ; X64-NEXT:    vcvttpd2dq %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf1,0xfd,0x09,0xe6,0xc8]
3752 ; X64-NEXT:    vmovaps %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x28,0xc1]
3753 ; X64-NEXT:    retq # encoding: [0xc3]
3760 ; CHECK:       # %bb.0:
3761 ; CHECK-NEXT:    vcvttpd2dq %xmm0, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf9,0xe6,0xc0]
3762 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
3764   %res3 = shufflevector <4 x i32> %res2, <4 x i32> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 4, i32 5>
3770 ; X86:       # %bb.0:
3771 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
3772 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
3773 ; X86-NEXT:    vcvttpd2dq %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf1,0xfd,0x09,0xe6,0xc8]
3774 ; X86-NEXT:    vmovaps %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x28,0xc1]
3775 ; X86-NEXT:    retl # encoding: [0xc3]
3778 ; X64:       # %bb.0:
3779 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
3780 ; X64-NEXT:    vcvttpd2dq %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf1,0xfd,0x09,0xe6,0xc8]
3781 ; X64-NEXT:    vmovaps %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x28,0xc1]
3782 ; X64-NEXT:    retq # encoding: [0xc3]
3784   %res1 = shufflevector <4 x i32> %res, <4 x i32> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 4, i32 5>
3792 ; CHECK:       # %bb.0:
3793 ; CHECK-NEXT:    vcvttpd2udq %xmm0, %xmm0 # encoding: [0x62,0xf1,0xfc,0x08,0x78,0xc0]
3794 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
3801 ; X86:       # %bb.0:
3802 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
3803 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
3804 ; X86-NEXT:    vcvttpd2udq %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf1,0xfc,0x09,0x78,0xc8]
3805 ; X86-NEXT:    vmovaps %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x28,0xc1]
3806 ; X86-NEXT:    retl # encoding: [0xc3]
3809 ; X64:       # %bb.0:
3810 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
3811 ; X64-NEXT:    vcvttpd2udq %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf1,0xfc,0x09,0x78,0xc8]
3812 ; X64-NEXT:    vmovaps %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x28,0xc1]
3813 ; X64-NEXT:    retq # encoding: [0xc3]
3820 ; CHECK:       # %bb.0:
3821 ; CHECK-NEXT:    vcvttpd2udq %xmm0, %xmm0 # encoding: [0x62,0xf1,0xfc,0x08,0x78,0xc0]
3822 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
3824   %res3 = shufflevector <4 x i32> %res2, <4 x i32> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 4, i32 5>
3830 ; X86:       # %bb.0:
3831 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
3832 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
3833 ; X86-NEXT:    vcvttpd2udq %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf1,0xfc,0x09,0x78,0xc8]
3834 ; X86-NEXT:    vmovaps %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x28,0xc1]
3835 ; X86-NEXT:    retl # encoding: [0xc3]
3838 ; X64:       # %bb.0:
3839 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
3840 ; X64-NEXT:    vcvttpd2udq %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf1,0xfc,0x09,0x78,0xc8]
3841 ; X64-NEXT:    vmovaps %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x28,0xc1]
3842 ; X64-NEXT:    retq # encoding: [0xc3]
3844   %res1 = shufflevector <4 x i32> %res, <4 x i32> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 4, i32 5>
3852 ; CHECK:       # %bb.0:
3853 ; CHECK-NEXT:    vcvttpd2udq %ymm0, %xmm0 # encoding: [0x62,0xf1,0xfc,0x28,0x78,0xc0]
3854 ; CHECK-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
3855 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
3862 ; X86:       # %bb.0:
3863 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
3864 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
3865 ; X86-NEXT:    vcvttpd2udq %ymm0, %xmm1 {%k1} # encoding: [0x62,0xf1,0xfc,0x29,0x78,0xc8]
3866 ; X86-NEXT:    vmovaps %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x28,0xc1]
3867 ; X86-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
3868 ; X86-NEXT:    retl # encoding: [0xc3]
3871 ; X64:       # %bb.0:
3872 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
3873 ; X64-NEXT:    vcvttpd2udq %ymm0, %xmm1 {%k1} # encoding: [0x62,0xf1,0xfc,0x29,0x78,0xc8]
3874 ; X64-NEXT:    vmovaps %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x28,0xc1]
3875 ; X64-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
3876 ; X64-NEXT:    retq # encoding: [0xc3]
3885 ; CHECK:       # %bb.0:
3886 ; CHECK-NEXT:    vcvttps2udq %xmm0, %xmm0 # encoding: [0x62,0xf1,0x7c,0x08,0x78,0xc0]
3887 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
3894 ; X86:       # %bb.0:
3895 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
3896 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
3897 ; X86-NEXT:    vcvttps2udq %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf1,0x7c,0x09,0x78,0xc8]
3898 ; X86-NEXT:    vmovaps %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x28,0xc1]
3899 ; X86-NEXT:    retl # encoding: [0xc3]
3902 ; X64:       # %bb.0:
3903 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
3904 ; X64-NEXT:    vcvttps2udq %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf1,0x7c,0x09,0x78,0xc8]
3905 ; X64-NEXT:    vmovaps %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x28,0xc1]
3906 ; X64-NEXT:    retq # encoding: [0xc3]
3915 ; CHECK:       # %bb.0:
3916 ; CHECK-NEXT:    vcvttps2udq %ymm0, %ymm0 # encoding: [0x62,0xf1,0x7c,0x28,0x78,0xc0]
3917 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
3924 ; X86:       # %bb.0:
3925 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
3926 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
3927 ; X86-NEXT:    vcvttps2udq %ymm0, %ymm1 {%k1} # encoding: [0x62,0xf1,0x7c,0x29,0x78,0xc8]
3928 ; X86-NEXT:    vmovaps %ymm1, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfc,0x28,0xc1]
3929 ; X86-NEXT:    retl # encoding: [0xc3]
3932 ; X64:       # %bb.0:
3933 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
3934 ; X64-NEXT:    vcvttps2udq %ymm0, %ymm1 {%k1} # encoding: [0x62,0xf1,0x7c,0x29,0x78,0xc8]
3935 ; X64-NEXT:    vmovaps %ymm1, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfc,0x28,0xc1]
3936 ; X64-NEXT:    retq # encoding: [0xc3]
3945 ; X86:       # %bb.0:
3946 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
3947 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
3948 ; X86-NEXT:    vrndscalepd $4, %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf3,0xfd,0x09,0x09,0xc8,0x04]
3949 ; X86-NEXT:    vrndscalepd $88, %xmm0, %xmm0 # encoding: [0x62,0xf3,0xfd,0x08,0x09,0xc0,0x58]
3950 ; X86-NEXT:    vaddpd %xmm0, %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf1,0x58,0xc0]
3951 ; X86-NEXT:    retl # encoding: [0xc3]
3954 ; X64:       # %bb.0:
3955 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
3956 ; X64-NEXT:    vrndscalepd $4, %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf3,0xfd,0x09,0x09,0xc8,0x04]
3957 ; X64-NEXT:    vrndscalepd $88, %xmm0, %xmm0 # encoding: [0x62,0xf3,0xfd,0x08,0x09,0xc0,0x58]
3958 ; X64-NEXT:    vaddpd %xmm0, %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf1,0x58,0xc0]
3959 ; X64-NEXT:    retq # encoding: [0xc3]
3970 ; X86:       # %bb.0:
3971 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
3972 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
3973 ; X86-NEXT:    vrndscalepd $4, %ymm0, %ymm1 {%k1} # encoding: [0x62,0xf3,0xfd,0x29,0x09,0xc8,0x04]
3974 ; X86-NEXT:    vrndscalepd $88, %ymm0, %ymm0 # encoding: [0x62,0xf3,0xfd,0x28,0x09,0xc0,0x58]
3975 ; X86-NEXT:    vaddpd %ymm0, %ymm1, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xf5,0x58,0xc0]
3976 ; X86-NEXT:    retl # encoding: [0xc3]
3979 ; X64:       # %bb.0:
3980 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
3981 ; X64-NEXT:    vrndscalepd $4, %ymm0, %ymm1 {%k1} # encoding: [0x62,0xf3,0xfd,0x29,0x09,0xc8,0x04]
3982 ; X64-NEXT:    vrndscalepd $88, %ymm0, %ymm0 # encoding: [0x62,0xf3,0xfd,0x28,0x09,0xc0,0x58]
3983 ; X64-NEXT:    vaddpd %ymm0, %ymm1, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xf5,0x58,0xc0]
3984 ; X64-NEXT:    retq # encoding: [0xc3]
3995 ; X86:       # %bb.0:
3996 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
3997 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
3998 ; X86-NEXT:    vrndscaleps $88, %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf3,0x7d,0x09,0x08,0xc8,0x58]
3999 ; X86-NEXT:    vroundps $4, %xmm0, %xmm0 # EVEX TO VEX Compression encoding: [0xc4,0xe3,0x79,0x08,0xc0,0x04]
4000 ; X86-NEXT:    vaddps %xmm0, %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf0,0x58,0xc0]
4001 ; X86-NEXT:    retl # encoding: [0xc3]
4004 ; X64:       # %bb.0:
4005 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
4006 ; X64-NEXT:    vrndscaleps $88, %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf3,0x7d,0x09,0x08,0xc8,0x58]
4007 ; X64-NEXT:    vroundps $4, %xmm0, %xmm0 # EVEX TO VEX Compression encoding: [0xc4,0xe3,0x79,0x08,0xc0,0x04]
4008 ; X64-NEXT:    vaddps %xmm0, %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf0,0x58,0xc0]
4009 ; X64-NEXT:    retq # encoding: [0xc3]
4020 ; X86:       # %bb.0:
4021 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
4022 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
4023 ; X86-NEXT:    vrndscaleps $5, %ymm0, %ymm1 {%k1} # encoding: [0x62,0xf3,0x7d,0x29,0x08,0xc8,0x05]
4024 ; X86-NEXT:    vrndscaleps $66, %ymm0, %ymm0 # encoding: [0x62,0xf3,0x7d,0x28,0x08,0xc0,0x42]
4025 ; X86-NEXT:    vaddps %ymm0, %ymm1, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xf4,0x58,0xc0]
4026 ; X86-NEXT:    retl # encoding: [0xc3]
4029 ; X64:       # %bb.0:
4030 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
4031 ; X64-NEXT:    vrndscaleps $5, %ymm0, %ymm1 {%k1} # encoding: [0x62,0xf3,0x7d,0x29,0x08,0xc8,0x05]
4032 ; X64-NEXT:    vrndscaleps $66, %ymm0, %ymm0 # encoding: [0x62,0xf3,0x7d,0x28,0x08,0xc0,0x42]
4033 ; X64-NEXT:    vaddps %ymm0, %ymm1, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xf4,0x58,0xc0]
4034 ; X64-NEXT:    retq # encoding: [0xc3]
4045 ; X86:       # %bb.0:
4046 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
4047 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
4048 ; X86-NEXT:    vgetmantpd $11, %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf3,0xfd,0x09,0x26,0xc8,0x0b]
4049 ; X86-NEXT:    vgetmantpd $12, %xmm0, %xmm2 {%k1} {z} # encoding: [0x62,0xf3,0xfd,0x89,0x26,0xd0,0x0c]
4050 ; X86-NEXT:    vgetmantpd $13, %xmm0, %xmm0 # encoding: [0x62,0xf3,0xfd,0x08,0x26,0xc0,0x0d]
4051 ; X86-NEXT:    vaddpd %xmm0, %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf1,0x58,0xc0]
4052 ; X86-NEXT:    vaddpd %xmm0, %xmm2, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xe9,0x58,0xc0]
4053 ; X86-NEXT:    retl # encoding: [0xc3]
4056 ; X64:       # %bb.0:
4057 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
4058 ; X64-NEXT:    vgetmantpd $11, %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf3,0xfd,0x09,0x26,0xc8,0x0b]
4059 ; X64-NEXT:    vgetmantpd $12, %xmm0, %xmm2 {%k1} {z} # encoding: [0x62,0xf3,0xfd,0x89,0x26,0xd0,0x0c]
4060 ; X64-NEXT:    vgetmantpd $13, %xmm0, %xmm0 # encoding: [0x62,0xf3,0xfd,0x08,0x26,0xc0,0x0d]
4061 ; X64-NEXT:    vaddpd %xmm0, %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf1,0x58,0xc0]
4062 ; X64-NEXT:    vaddpd %xmm0, %xmm2, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xe9,0x58,0xc0]
4063 ; X64-NEXT:    retq # encoding: [0xc3]
4076 ; X86:       # %bb.0:
4077 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
4078 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
4079 ; X86-NEXT:    vgetmantpd $11, %ymm0, %ymm1 {%k1} # encoding: [0x62,0xf3,0xfd,0x29,0x26,0xc8,0x0b]
4080 ; X86-NEXT:    vgetmantpd $12, %ymm0, %ymm0 # encoding: [0x62,0xf3,0xfd,0x28,0x26,0xc0,0x0c]
4081 ; X86-NEXT:    vaddpd %ymm0, %ymm1, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xf5,0x58,0xc0]
4082 ; X86-NEXT:    retl # encoding: [0xc3]
4085 ; X64:       # %bb.0:
4086 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
4087 ; X64-NEXT:    vgetmantpd $11, %ymm0, %ymm1 {%k1} # encoding: [0x62,0xf3,0xfd,0x29,0x26,0xc8,0x0b]
4088 ; X64-NEXT:    vgetmantpd $12, %ymm0, %ymm0 # encoding: [0x62,0xf3,0xfd,0x28,0x26,0xc0,0x0c]
4089 ; X64-NEXT:    vaddpd %ymm0, %ymm1, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xf5,0x58,0xc0]
4090 ; X64-NEXT:    retq # encoding: [0xc3]
4101 ; X86:       # %bb.0:
4102 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
4103 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
4104 ; X86-NEXT:    vgetmantps $11, %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf3,0x7d,0x09,0x26,0xc8,0x0b]
4105 ; X86-NEXT:    vgetmantps $12, %xmm0, %xmm0 # encoding: [0x62,0xf3,0x7d,0x08,0x26,0xc0,0x0c]
4106 ; X86-NEXT:    vaddps %xmm0, %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf0,0x58,0xc0]
4107 ; X86-NEXT:    retl # encoding: [0xc3]
4110 ; X64:       # %bb.0:
4111 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
4112 ; X64-NEXT:    vgetmantps $11, %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf3,0x7d,0x09,0x26,0xc8,0x0b]
4113 ; X64-NEXT:    vgetmantps $12, %xmm0, %xmm0 # encoding: [0x62,0xf3,0x7d,0x08,0x26,0xc0,0x0c]
4114 ; X64-NEXT:    vaddps %xmm0, %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf0,0x58,0xc0]
4115 ; X64-NEXT:    retq # encoding: [0xc3]
4126 ; X86:       # %bb.0:
4127 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
4128 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
4129 ; X86-NEXT:    vgetmantps $11, %ymm0, %ymm1 {%k1} # encoding: [0x62,0xf3,0x7d,0x29,0x26,0xc8,0x0b]
4130 ; X86-NEXT:    vgetmantps $12, %ymm0, %ymm0 # encoding: [0x62,0xf3,0x7d,0x28,0x26,0xc0,0x0c]
4131 ; X86-NEXT:    vaddps %ymm0, %ymm1, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xf4,0x58,0xc0]
4132 ; X86-NEXT:    retl # encoding: [0xc3]
4135 ; X64:       # %bb.0:
4136 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
4137 ; X64-NEXT:    vgetmantps $11, %ymm0, %ymm1 {%k1} # encoding: [0x62,0xf3,0x7d,0x29,0x26,0xc8,0x0b]
4138 ; X64-NEXT:    vgetmantps $12, %ymm0, %ymm0 # encoding: [0x62,0xf3,0x7d,0x28,0x26,0xc0,0x0c]
4139 ; X64-NEXT:    vaddps %ymm0, %ymm1, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xf4,0x58,0xc0]
4140 ; X64-NEXT:    retq # encoding: [0xc3]
4151 ; CHECK:       # %bb.0:
4152 ; CHECK-NEXT:    vpternlogd $33, %xmm2, %xmm1, %xmm0 # encoding: [0x62,0xf3,0x75,0x08,0x25,0xc2,0x21]
4154 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
4161 ; X86:       # %bb.0:
4162 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
4163 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
4164 ; X86-NEXT:    vpternlogd $33, %xmm2, %xmm1, %xmm0 {%k1} # encoding: [0x62,0xf3,0x75,0x09,0x25,0xc2,0x21]
4166 ; X86-NEXT:    retl # encoding: [0xc3]
4169 ; X64:       # %bb.0:
4170 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
4171 ; X64-NEXT:    vpternlogd $33, %xmm2, %xmm1, %xmm0 {%k1} # encoding: [0x62,0xf3,0x75,0x09,0x25,0xc2,0x21]
4173 ; X64-NEXT:    retq # encoding: [0xc3]
4176   %extract = shufflevector <8 x i1> %2, <8 x i1> %2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
4185 ; X86:       # %bb.0:
4186 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
4187 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
4188 ; X86-NEXT:    vpternlogd $33, %xmm2, %xmm1, %xmm0 {%k1} {z} # encoding: [0x62,0xf3,0x75,0x89,0x25,0xc2,0x21]
4190 ; X86-NEXT:    retl # encoding: [0xc3]
4193 ; X64:       # %bb.0:
4194 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
4195 ; X64-NEXT:    vpternlogd $33, %xmm2, %xmm1, %xmm0 {%k1} {z} # encoding: [0x62,0xf3,0x75,0x89,0x25,0xc2,0x21]
4197 ; X64-NEXT:    retq # encoding: [0xc3]
4200   %extract = shufflevector <8 x i1> %2, <8 x i1> %2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
4209 ; CHECK:       # %bb.0:
4210 ; CHECK-NEXT:    vpternlogd $33, %ymm2, %ymm1, %ymm0 # encoding: [0x62,0xf3,0x75,0x28,0x25,0xc2,0x21]
4212 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
4219 ; X86:       # %bb.0:
4220 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
4221 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
4222 ; X86-NEXT:    vpternlogd $33, %ymm2, %ymm1, %ymm0 {%k1} # encoding: [0x62,0xf3,0x75,0x29,0x25,0xc2,0x21]
4224 ; X86-NEXT:    retl # encoding: [0xc3]
4227 ; X64:       # %bb.0:
4228 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
4229 ; X64-NEXT:    vpternlogd $33, %ymm2, %ymm1, %ymm0 {%k1} # encoding: [0x62,0xf3,0x75,0x29,0x25,0xc2,0x21]
4231 ; X64-NEXT:    retq # encoding: [0xc3]
4242 ; X86:       # %bb.0:
4243 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
4244 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
4245 ; X86-NEXT:    vpternlogd $33, %ymm2, %ymm1, %ymm0 {%k1} {z} # encoding: [0x62,0xf3,0x75,0xa9,0x25,0xc2,0x21]
4247 ; X86-NEXT:    retl # encoding: [0xc3]
4250 ; X64:       # %bb.0:
4251 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
4252 ; X64-NEXT:    vpternlogd $33, %ymm2, %ymm1, %ymm0 {%k1} {z} # encoding: [0x62,0xf3,0x75,0xa9,0x25,0xc2,0x21]
4254 ; X64-NEXT:    retq # encoding: [0xc3]
4265 ; CHECK:       # %bb.0:
4266 ; CHECK-NEXT:    vpternlogq $33, %xmm2, %xmm1, %xmm0 # encoding: [0x62,0xf3,0xf5,0x08,0x25,0xc2,0x21]
4268 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
4275 ; X86:       # %bb.0:
4276 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
4277 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
4278 ; X86-NEXT:    vpternlogq $33, %xmm2, %xmm1, %xmm0 {%k1} # encoding: [0x62,0xf3,0xf5,0x09,0x25,0xc2,0x21]
4280 ; X86-NEXT:    retl # encoding: [0xc3]
4283 ; X64:       # %bb.0:
4284 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
4285 ; X64-NEXT:    vpternlogq $33, %xmm2, %xmm1, %xmm0 {%k1} # encoding: [0x62,0xf3,0xf5,0x09,0x25,0xc2,0x21]
4287 ; X64-NEXT:    retq # encoding: [0xc3]
4290   %extract = shufflevector <8 x i1> %2, <8 x i1> %2, <2 x i32> <i32 0, i32 1>
4297 ; X86:       # %bb.0:
4298 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
4299 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
4300 ; X86-NEXT:    vpternlogq $33, %xmm2, %xmm1, %xmm0 {%k1} {z} # encoding: [0x62,0xf3,0xf5,0x89,0x25,0xc2,0x21]
4302 ; X86-NEXT:    retl # encoding: [0xc3]
4305 ; X64:       # %bb.0:
4306 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
4307 ; X64-NEXT:    vpternlogq $33, %xmm2, %xmm1, %xmm0 {%k1} {z} # encoding: [0x62,0xf3,0xf5,0x89,0x25,0xc2,0x21]
4309 ; X64-NEXT:    retq # encoding: [0xc3]
4312   %extract = shufflevector <8 x i1> %2, <8 x i1> %2, <2 x i32> <i32 0, i32 1>
4321 ; CHECK:       # %bb.0:
4322 ; CHECK-NEXT:    vpternlogq $33, %ymm2, %ymm1, %ymm0 # encoding: [0x62,0xf3,0xf5,0x28,0x25,0xc2,0x21]
4324 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
4331 ; X86:       # %bb.0:
4332 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
4333 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
4334 ; X86-NEXT:    vpternlogq $33, %ymm2, %ymm1, %ymm0 {%k1} # encoding: [0x62,0xf3,0xf5,0x29,0x25,0xc2,0x21]
4336 ; X86-NEXT:    retl # encoding: [0xc3]
4339 ; X64:       # %bb.0:
4340 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
4341 ; X64-NEXT:    vpternlogq $33, %ymm2, %ymm1, %ymm0 {%k1} # encoding: [0x62,0xf3,0xf5,0x29,0x25,0xc2,0x21]
4343 ; X64-NEXT:    retq # encoding: [0xc3]
4346   %extract = shufflevector <8 x i1> %2, <8 x i1> %2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
4353 ; X86:       # %bb.0:
4354 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
4355 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
4356 ; X86-NEXT:    vpternlogq $33, %ymm2, %ymm1, %ymm0 {%k1} {z} # encoding: [0x62,0xf3,0xf5,0xa9,0x25,0xc2,0x21]
4358 ; X86-NEXT:    retl # encoding: [0xc3]
4361 ; X64:       # %bb.0:
4362 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
4363 ; X64-NEXT:    vpternlogq $33, %ymm2, %ymm1, %ymm0 {%k1} {z} # encoding: [0x62,0xf3,0xf5,0xa9,0x25,0xc2,0x21]
4365 ; X64-NEXT:    retq # encoding: [0xc3]
4368   %extract = shufflevector <8 x i1> %2, <8 x i1> %2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
4375 ; X86:       # %bb.0:
4376 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
4377 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
4378 ; X86-NEXT:    vcvtps2ph $2, %xmm0, %xmm2 # EVEX TO VEX Compression encoding: [0xc4,0xe3,0x79,0x1d,0xc2,0x02]
4379 ; X86-NEXT:    vcvtps2ph $10, %xmm0, %xmm3 {%k1} {z} # encoding: [0x62,0xf3,0x7d,0x89,0x1d,0xc3,0x0a]
4380 ; X86-NEXT:    vpaddw %xmm3, %xmm2, %xmm2 # encoding: [0xc5,0xe9,0xfd,0xd3]
4381 ; X86-NEXT:    vcvtps2ph $11, %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf3,0x7d,0x09,0x1d,0xc1,0x0b]
4382 ; X86-NEXT:    vpaddw %xmm2, %xmm1, %xmm0 # encoding: [0xc5,0xf1,0xfd,0xc2]
4383 ; X86-NEXT:    retl # encoding: [0xc3]
4386 ; X64:       # %bb.0:
4387 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
4388 ; X64-NEXT:    vcvtps2ph $2, %xmm0, %xmm2 # EVEX TO VEX Compression encoding: [0xc4,0xe3,0x79,0x1d,0xc2,0x02]
4389 ; X64-NEXT:    vcvtps2ph $10, %xmm0, %xmm3 {%k1} {z} # encoding: [0x62,0xf3,0x7d,0x89,0x1d,0xc3,0x0a]
4390 ; X64-NEXT:    vpaddw %xmm3, %xmm2, %xmm2 # encoding: [0xc5,0xe9,0xfd,0xd3]
4391 ; X64-NEXT:    vcvtps2ph $11, %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf3,0x7d,0x09,0x1d,0xc1,0x0b]
4392 ; X64-NEXT:    vpaddw %xmm2, %xmm1, %xmm0 # encoding: [0xc5,0xf1,0xfd,0xc2]
4393 ; X64-NEXT:    retq # encoding: [0xc3]
4406 ; X86:       # %bb.0:
4407 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
4408 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
4409 ; X86-NEXT:    vcvtps2ph $2, %ymm0, %xmm2 # EVEX TO VEX Compression encoding: [0xc4,0xe3,0x7d,0x1d,0xc2,0x02]
4410 ; X86-NEXT:    vcvtps2ph $11, %ymm0, %xmm3 {%k1} {z} # encoding: [0x62,0xf3,0x7d,0xa9,0x1d,0xc3,0x0b]
4411 ; X86-NEXT:    vpaddw %xmm3, %xmm2, %xmm2 # encoding: [0xc5,0xe9,0xfd,0xd3]
4412 ; X86-NEXT:    vcvtps2ph $12, %ymm0, %xmm1 {%k1} # encoding: [0x62,0xf3,0x7d,0x29,0x1d,0xc1,0x0c]
4413 ; X86-NEXT:    vpaddw %xmm2, %xmm1, %xmm0 # encoding: [0xc5,0xf1,0xfd,0xc2]
4414 ; X86-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
4415 ; X86-NEXT:    retl # encoding: [0xc3]
4418 ; X64:       # %bb.0:
4419 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
4420 ; X64-NEXT:    vcvtps2ph $2, %ymm0, %xmm2 # EVEX TO VEX Compression encoding: [0xc4,0xe3,0x7d,0x1d,0xc2,0x02]
4421 ; X64-NEXT:    vcvtps2ph $11, %ymm0, %xmm3 {%k1} {z} # encoding: [0x62,0xf3,0x7d,0xa9,0x1d,0xc3,0x0b]
4422 ; X64-NEXT:    vpaddw %xmm3, %xmm2, %xmm2 # encoding: [0xc5,0xe9,0xfd,0xd3]
4423 ; X64-NEXT:    vcvtps2ph $12, %ymm0, %xmm1 {%k1} # encoding: [0x62,0xf3,0x7d,0x29,0x1d,0xc1,0x0c]
4424 ; X64-NEXT:    vpaddw %xmm2, %xmm1, %xmm0 # encoding: [0xc5,0xf1,0xfd,0xc2]
4425 ; X64-NEXT:    vzeroupper # encoding: [0xc5,0xf8,0x77]
4426 ; X64-NEXT:    retq # encoding: [0xc3]
4439 ; CHECK:       # %bb.0:
4440 ; CHECK-NEXT:    vrsqrt14ps %ymm0, %ymm0 # encoding: [0x62,0xf2,0x7d,0x28,0x4e,0xc0]
4441 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
4448 ; X86:       # %bb.0:
4449 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
4450 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
4451 ; X86-NEXT:    vrsqrt14ps %ymm0, %ymm0 {%k1} {z} # encoding: [0x62,0xf2,0x7d,0xa9,0x4e,0xc0]
4452 ; X86-NEXT:    retl # encoding: [0xc3]
4455 ; X64:       # %bb.0:
4456 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
4457 ; X64-NEXT:    vrsqrt14ps %ymm0, %ymm0 {%k1} {z} # encoding: [0x62,0xf2,0x7d,0xa9,0x4e,0xc0]
4458 ; X64-NEXT:    retq # encoding: [0xc3]
4465 ; X86:       # %bb.0:
4466 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
4467 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
4468 ; X86-NEXT:    vrsqrt14ps %ymm0, %ymm1 {%k1} # encoding: [0x62,0xf2,0x7d,0x29,0x4e,0xc8]
4469 ; X86-NEXT:    vmovaps %ymm1, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfc,0x28,0xc1]
4470 ; X86-NEXT:    retl # encoding: [0xc3]
4473 ; X64:       # %bb.0:
4474 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
4475 ; X64-NEXT:    vrsqrt14ps %ymm0, %ymm1 {%k1} # encoding: [0x62,0xf2,0x7d,0x29,0x4e,0xc8]
4476 ; X64-NEXT:    vmovaps %ymm1, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfc,0x28,0xc1]
4477 ; X64-NEXT:    retq # encoding: [0xc3]
4484 ; CHECK:       # %bb.0:
4485 ; CHECK-NEXT:    vrsqrt14ps %xmm0, %xmm0 # encoding: [0x62,0xf2,0x7d,0x08,0x4e,0xc0]
4486 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
4493 ; X86:       # %bb.0:
4494 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
4495 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
4496 ; X86-NEXT:    vrsqrt14ps %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7d,0x89,0x4e,0xc0]
4497 ; X86-NEXT:    retl # encoding: [0xc3]
4500 ; X64:       # %bb.0:
4501 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
4502 ; X64-NEXT:    vrsqrt14ps %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7d,0x89,0x4e,0xc0]
4503 ; X64-NEXT:    retq # encoding: [0xc3]
4510 ; X86:       # %bb.0:
4511 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
4512 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
4513 ; X86-NEXT:    vrsqrt14ps %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7d,0x09,0x4e,0xc8]
4514 ; X86-NEXT:    vmovaps %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x28,0xc1]
4515 ; X86-NEXT:    retl # encoding: [0xc3]
4518 ; X64:       # %bb.0:
4519 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
4520 ; X64-NEXT:    vrsqrt14ps %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7d,0x09,0x4e,0xc8]
4521 ; X64-NEXT:    vmovaps %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x28,0xc1]
4522 ; X64-NEXT:    retq # encoding: [0xc3]
4532 ; CHECK:       # %bb.0:
4533 ; CHECK-NEXT:    vrcp14ps %ymm0, %ymm0 # encoding: [0x62,0xf2,0x7d,0x28,0x4c,0xc0]
4534 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
4541 ; X86:       # %bb.0:
4542 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
4543 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
4544 ; X86-NEXT:    vrcp14ps %ymm0, %ymm0 {%k1} {z} # encoding: [0x62,0xf2,0x7d,0xa9,0x4c,0xc0]
4545 ; X86-NEXT:    retl # encoding: [0xc3]
4548 ; X64:       # %bb.0:
4549 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
4550 ; X64-NEXT:    vrcp14ps %ymm0, %ymm0 {%k1} {z} # encoding: [0x62,0xf2,0x7d,0xa9,0x4c,0xc0]
4551 ; X64-NEXT:    retq # encoding: [0xc3]
4558 ; X86:       # %bb.0:
4559 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
4560 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
4561 ; X86-NEXT:    vrcp14ps %ymm0, %ymm1 {%k1} # encoding: [0x62,0xf2,0x7d,0x29,0x4c,0xc8]
4562 ; X86-NEXT:    vmovaps %ymm1, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfc,0x28,0xc1]
4563 ; X86-NEXT:    retl # encoding: [0xc3]
4566 ; X64:       # %bb.0:
4567 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
4568 ; X64-NEXT:    vrcp14ps %ymm0, %ymm1 {%k1} # encoding: [0x62,0xf2,0x7d,0x29,0x4c,0xc8]
4569 ; X64-NEXT:    vmovaps %ymm1, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfc,0x28,0xc1]
4570 ; X64-NEXT:    retq # encoding: [0xc3]
4577 ; CHECK:       # %bb.0:
4578 ; CHECK-NEXT:    vrcp14ps %xmm0, %xmm0 # encoding: [0x62,0xf2,0x7d,0x08,0x4c,0xc0]
4579 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
4586 ; X86:       # %bb.0:
4587 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
4588 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
4589 ; X86-NEXT:    vrcp14ps %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7d,0x89,0x4c,0xc0]
4590 ; X86-NEXT:    retl # encoding: [0xc3]
4593 ; X64:       # %bb.0:
4594 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
4595 ; X64-NEXT:    vrcp14ps %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7d,0x89,0x4c,0xc0]
4596 ; X64-NEXT:    retq # encoding: [0xc3]
4603 ; X86:       # %bb.0:
4604 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
4605 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
4606 ; X86-NEXT:    vrcp14ps %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7d,0x09,0x4c,0xc8]
4607 ; X86-NEXT:    vmovaps %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x28,0xc1]
4608 ; X86-NEXT:    retl # encoding: [0xc3]
4611 ; X64:       # %bb.0:
4612 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
4613 ; X64-NEXT:    vrcp14ps %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7d,0x09,0x4c,0xc8]
4614 ; X64-NEXT:    vmovaps %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x28,0xc1]
4615 ; X64-NEXT:    retq # encoding: [0xc3]
4625 ; CHECK:       # %bb.0:
4626 ; CHECK-NEXT:    vrsqrt14pd %ymm0, %ymm0 # encoding: [0x62,0xf2,0xfd,0x28,0x4e,0xc0]
4627 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
4634 ; X86:       # %bb.0:
4635 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
4636 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
4637 ; X86-NEXT:    vrsqrt14pd %ymm0, %ymm0 {%k1} {z} # encoding: [0x62,0xf2,0xfd,0xa9,0x4e,0xc0]
4638 ; X86-NEXT:    retl # encoding: [0xc3]
4641 ; X64:       # %bb.0:
4642 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
4643 ; X64-NEXT:    vrsqrt14pd %ymm0, %ymm0 {%k1} {z} # encoding: [0x62,0xf2,0xfd,0xa9,0x4e,0xc0]
4644 ; X64-NEXT:    retq # encoding: [0xc3]
4651 ; X86:       # %bb.0:
4652 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
4653 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
4654 ; X86-NEXT:    vrsqrt14pd %ymm0, %ymm1 {%k1} # encoding: [0x62,0xf2,0xfd,0x29,0x4e,0xc8]
4655 ; X86-NEXT:    vmovapd %ymm1, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfd,0x28,0xc1]
4656 ; X86-NEXT:    retl # encoding: [0xc3]
4659 ; X64:       # %bb.0:
4660 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
4661 ; X64-NEXT:    vrsqrt14pd %ymm0, %ymm1 {%k1} # encoding: [0x62,0xf2,0xfd,0x29,0x4e,0xc8]
4662 ; X64-NEXT:    vmovapd %ymm1, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfd,0x28,0xc1]
4663 ; X64-NEXT:    retq # encoding: [0xc3]
4670 ; CHECK:       # %bb.0:
4671 ; CHECK-NEXT:    vrsqrt14pd %xmm0, %xmm0 # encoding: [0x62,0xf2,0xfd,0x08,0x4e,0xc0]
4672 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
4679 ; X86:       # %bb.0:
4680 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
4681 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
4682 ; X86-NEXT:    vrsqrt14pd %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0xfd,0x89,0x4e,0xc0]
4683 ; X86-NEXT:    retl # encoding: [0xc3]
4686 ; X64:       # %bb.0:
4687 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
4688 ; X64-NEXT:    vrsqrt14pd %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0xfd,0x89,0x4e,0xc0]
4689 ; X64-NEXT:    retq # encoding: [0xc3]
4696 ; X86:       # %bb.0:
4697 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
4698 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
4699 ; X86-NEXT:    vrsqrt14pd %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0xfd,0x09,0x4e,0xc8]
4700 ; X86-NEXT:    vmovapd %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf9,0x28,0xc1]
4701 ; X86-NEXT:    retl # encoding: [0xc3]
4704 ; X64:       # %bb.0:
4705 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
4706 ; X64-NEXT:    vrsqrt14pd %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0xfd,0x09,0x4e,0xc8]
4707 ; X64-NEXT:    vmovapd %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf9,0x28,0xc1]
4708 ; X64-NEXT:    retq # encoding: [0xc3]
4718 ; CHECK:       # %bb.0:
4719 ; CHECK-NEXT:    vrcp14pd %ymm0, %ymm0 # encoding: [0x62,0xf2,0xfd,0x28,0x4c,0xc0]
4720 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
4727 ; X86:       # %bb.0:
4728 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
4729 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
4730 ; X86-NEXT:    vrcp14pd %ymm0, %ymm0 {%k1} {z} # encoding: [0x62,0xf2,0xfd,0xa9,0x4c,0xc0]
4731 ; X86-NEXT:    retl # encoding: [0xc3]
4734 ; X64:       # %bb.0:
4735 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
4736 ; X64-NEXT:    vrcp14pd %ymm0, %ymm0 {%k1} {z} # encoding: [0x62,0xf2,0xfd,0xa9,0x4c,0xc0]
4737 ; X64-NEXT:    retq # encoding: [0xc3]
4744 ; X86:       # %bb.0:
4745 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
4746 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
4747 ; X86-NEXT:    vrcp14pd %ymm0, %ymm1 {%k1} # encoding: [0x62,0xf2,0xfd,0x29,0x4c,0xc8]
4748 ; X86-NEXT:    vmovapd %ymm1, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfd,0x28,0xc1]
4749 ; X86-NEXT:    retl # encoding: [0xc3]
4752 ; X64:       # %bb.0:
4753 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
4754 ; X64-NEXT:    vrcp14pd %ymm0, %ymm1 {%k1} # encoding: [0x62,0xf2,0xfd,0x29,0x4c,0xc8]
4755 ; X64-NEXT:    vmovapd %ymm1, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfd,0x28,0xc1]
4756 ; X64-NEXT:    retq # encoding: [0xc3]
4763 ; CHECK:       # %bb.0:
4764 ; CHECK-NEXT:    vrcp14pd %xmm0, %xmm0 # encoding: [0x62,0xf2,0xfd,0x08,0x4c,0xc0]
4765 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
4772 ; X86:       # %bb.0:
4773 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
4774 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
4775 ; X86-NEXT:    vrcp14pd %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0xfd,0x89,0x4c,0xc0]
4776 ; X86-NEXT:    retl # encoding: [0xc3]
4779 ; X64:       # %bb.0:
4780 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
4781 ; X64-NEXT:    vrcp14pd %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0xfd,0x89,0x4c,0xc0]
4782 ; X64-NEXT:    retq # encoding: [0xc3]
4789 ; X86:       # %bb.0:
4790 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
4791 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
4792 ; X86-NEXT:    vrcp14pd %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0xfd,0x09,0x4c,0xc8]
4793 ; X86-NEXT:    vmovapd %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf9,0x28,0xc1]
4794 ; X86-NEXT:    retl # encoding: [0xc3]
4797 ; X64:       # %bb.0:
4798 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
4799 ; X64-NEXT:    vrcp14pd %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0xfd,0x09,0x4c,0xc8]
4800 ; X64-NEXT:    vmovapd %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf9,0x28,0xc1]
4801 ; X64-NEXT:    retq # encoding: [0xc3]
4813 ; CHECK:       # %bb.0:
4814 ; CHECK-NEXT:    vpermpd %ymm0, %ymm1, %ymm0 # encoding: [0x62,0xf2,0xf5,0x28,0x16,0xc0]
4815 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
4822 ; X86:       # %bb.0:
4823 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
4824 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
4825 ; X86-NEXT:    vpermpd %ymm0, %ymm1, %ymm2 {%k1} # encoding: [0x62,0xf2,0xf5,0x29,0x16,0xd0]
4826 ; X86-NEXT:    vmovapd %ymm2, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfd,0x28,0xc2]
4827 ; X86-NEXT:    retl # encoding: [0xc3]
4830 ; X64:       # %bb.0:
4831 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
4832 ; X64-NEXT:    vpermpd %ymm0, %ymm1, %ymm2 {%k1} # encoding: [0x62,0xf2,0xf5,0x29,0x16,0xd0]
4833 ; X64-NEXT:    vmovapd %ymm2, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfd,0x28,0xc2]
4834 ; X64-NEXT:    retq # encoding: [0xc3]
4837   %extract1 = shufflevector <8 x i1> %2, <8 x i1> %2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
4844 ; X86:       # %bb.0:
4845 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
4846 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
4847 ; X86-NEXT:    vpermpd %ymm0, %ymm1, %ymm0 {%k1} {z} # encoding: [0x62,0xf2,0xf5,0xa9,0x16,0xc0]
4848 ; X86-NEXT:    retl # encoding: [0xc3]
4851 ; X64:       # %bb.0:
4852 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
4853 ; X64-NEXT:    vpermpd %ymm0, %ymm1, %ymm0 {%k1} {z} # encoding: [0x62,0xf2,0xf5,0xa9,0x16,0xc0]
4854 ; X64-NEXT:    retq # encoding: [0xc3]
4857   %extract1 = shufflevector <8 x i1> %2, <8 x i1> %2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
4866 ; CHECK:       # %bb.0:
4867 ; CHECK-NEXT:    vpermpd %ymm0, %ymm1, %ymm0 # encoding: [0x62,0xf2,0xf5,0x28,0x16,0xc0]
4868 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
4875 ; X86:       # %bb.0:
4876 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
4877 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
4878 ; X86-NEXT:    vpermq %ymm0, %ymm1, %ymm2 {%k1} # encoding: [0x62,0xf2,0xf5,0x29,0x36,0xd0]
4879 ; X86-NEXT:    vmovdqa %ymm2, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfd,0x6f,0xc2]
4880 ; X86-NEXT:    retl # encoding: [0xc3]
4883 ; X64:       # %bb.0:
4884 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
4885 ; X64-NEXT:    vpermq %ymm0, %ymm1, %ymm2 {%k1} # encoding: [0x62,0xf2,0xf5,0x29,0x36,0xd0]
4886 ; X64-NEXT:    vmovdqa %ymm2, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfd,0x6f,0xc2]
4887 ; X64-NEXT:    retq # encoding: [0xc3]
4890   %extract1 = shufflevector <8 x i1> %2, <8 x i1> %2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
4897 ; X86:       # %bb.0:
4898 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
4899 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
4900 ; X86-NEXT:    vpermq %ymm0, %ymm1, %ymm0 {%k1} {z} # encoding: [0x62,0xf2,0xf5,0xa9,0x36,0xc0]
4901 ; X86-NEXT:    retl # encoding: [0xc3]
4904 ; X64:       # %bb.0:
4905 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
4906 ; X64-NEXT:    vpermq %ymm0, %ymm1, %ymm0 {%k1} {z} # encoding: [0x62,0xf2,0xf5,0xa9,0x36,0xc0]
4907 ; X64-NEXT:    retq # encoding: [0xc3]
4910   %extract1 = shufflevector <8 x i1> %2, <8 x i1> %2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
4919 ; X86:       # %bb.0:
4920 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
4921 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
4922 ; X86-NEXT:    vmovapd %xmm0, %xmm3 # EVEX TO VEX Compression encoding: [0xc5,0xf9,0x28,0xd8]
4923 ; X86-NEXT:    vfixupimmpd $5, %xmm2, %xmm1, %xmm3 {%k1} # encoding: [0x62,0xf3,0xf5,0x09,0x54,0xda,0x05]
4924 ; X86-NEXT:    vxorpd %xmm4, %xmm4, %xmm4 # encoding: [0xc5,0xd9,0x57,0xe4]
4925 ; X86-NEXT:    vfixupimmpd $4, %xmm2, %xmm1, %xmm4 {%k1} {z} # encoding: [0x62,0xf3,0xf5,0x89,0x54,0xe2,0x04]
4926 ; X86-NEXT:    vaddpd %xmm4, %xmm3, %xmm3 # EVEX TO VEX Compression encoding: [0xc5,0xe1,0x58,0xdc]
4927 ; X86-NEXT:    vfixupimmpd $3, %xmm2, %xmm1, %xmm0 # encoding: [0x62,0xf3,0xf5,0x08,0x54,0xc2,0x03]
4928 ; X86-NEXT:    vaddpd %xmm0, %xmm3, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xe1,0x58,0xc0]
4929 ; X86-NEXT:    retl # encoding: [0xc3]
4932 ; X64:       # %bb.0:
4933 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
4934 ; X64-NEXT:    vmovapd %xmm0, %xmm3 # EVEX TO VEX Compression encoding: [0xc5,0xf9,0x28,0xd8]
4935 ; X64-NEXT:    vfixupimmpd $5, %xmm2, %xmm1, %xmm3 {%k1} # encoding: [0x62,0xf3,0xf5,0x09,0x54,0xda,0x05]
4936 ; X64-NEXT:    vxorpd %xmm4, %xmm4, %xmm4 # encoding: [0xc5,0xd9,0x57,0xe4]
4937 ; X64-NEXT:    vfixupimmpd $4, %xmm2, %xmm1, %xmm4 {%k1} {z} # encoding: [0x62,0xf3,0xf5,0x89,0x54,0xe2,0x04]
4938 ; X64-NEXT:    vaddpd %xmm4, %xmm3, %xmm3 # EVEX TO VEX Compression encoding: [0xc5,0xe1,0x58,0xdc]
4939 ; X64-NEXT:    vfixupimmpd $3, %xmm2, %xmm1, %xmm0 # encoding: [0x62,0xf3,0xf5,0x08,0x54,0xc2,0x03]
4940 ; X64-NEXT:    vaddpd %xmm0, %xmm3, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xe1,0x58,0xc0]
4941 ; X64-NEXT:    retq # encoding: [0xc3]
4954 ; X86:       # %bb.0:
4955 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
4956 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
4957 ; X86-NEXT:    vmovapd %xmm0, %xmm3 # EVEX TO VEX Compression encoding: [0xc5,0xf9,0x28,0xd8]
4958 ; X86-NEXT:    vfixupimmpd $5, %xmm2, %xmm1, %xmm3 {%k1} {z} # encoding: [0x62,0xf3,0xf5,0x89,0x54,0xda,0x05]
4959 ; X86-NEXT:    vxorpd %xmm2, %xmm2, %xmm2 # encoding: [0xc5,0xe9,0x57,0xd2]
4960 ; X86-NEXT:    vfixupimmpd $3, %xmm2, %xmm1, %xmm0 {%k1} {z} # encoding: [0x62,0xf3,0xf5,0x89,0x54,0xc2,0x03]
4961 ; X86-NEXT:    vaddpd %xmm0, %xmm3, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xe1,0x58,0xc0]
4962 ; X86-NEXT:    retl # encoding: [0xc3]
4965 ; X64:       # %bb.0:
4966 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
4967 ; X64-NEXT:    vmovapd %xmm0, %xmm3 # EVEX TO VEX Compression encoding: [0xc5,0xf9,0x28,0xd8]
4968 ; X64-NEXT:    vfixupimmpd $5, %xmm2, %xmm1, %xmm3 {%k1} {z} # encoding: [0x62,0xf3,0xf5,0x89,0x54,0xda,0x05]
4969 ; X64-NEXT:    vxorpd %xmm2, %xmm2, %xmm2 # encoding: [0xc5,0xe9,0x57,0xd2]
4970 ; X64-NEXT:    vfixupimmpd $3, %xmm2, %xmm1, %xmm0 {%k1} {z} # encoding: [0x62,0xf3,0xf5,0x89,0x54,0xc2,0x03]
4971 ; X64-NEXT:    vaddpd %xmm0, %xmm3, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xe1,0x58,0xc0]
4972 ; X64-NEXT:    retq # encoding: [0xc3]
4985 ; X86:       # %bb.0:
4986 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
4987 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
4988 ; X86-NEXT:    vmovapd %ymm0, %ymm3 # EVEX TO VEX Compression encoding: [0xc5,0xfd,0x28,0xd8]
4989 ; X86-NEXT:    vfixupimmpd $4, %ymm2, %ymm1, %ymm3 {%k1} # encoding: [0x62,0xf3,0xf5,0x29,0x54,0xda,0x04]
4990 ; X86-NEXT:    vxorpd %xmm4, %xmm4, %xmm4 # encoding: [0xc5,0xd9,0x57,0xe4]
4991 ; X86-NEXT:    vfixupimmpd $5, %ymm2, %ymm1, %ymm4 {%k1} {z} # encoding: [0x62,0xf3,0xf5,0xa9,0x54,0xe2,0x05]
4992 ; X86-NEXT:    vaddpd %ymm4, %ymm3, %ymm3 # EVEX TO VEX Compression encoding: [0xc5,0xe5,0x58,0xdc]
4993 ; X86-NEXT:    vfixupimmpd $3, %ymm2, %ymm1, %ymm0 # encoding: [0x62,0xf3,0xf5,0x28,0x54,0xc2,0x03]
4994 ; X86-NEXT:    vaddpd %ymm0, %ymm3, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xe5,0x58,0xc0]
4995 ; X86-NEXT:    retl # encoding: [0xc3]
4998 ; X64:       # %bb.0:
4999 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
5000 ; X64-NEXT:    vmovapd %ymm0, %ymm3 # EVEX TO VEX Compression encoding: [0xc5,0xfd,0x28,0xd8]
5001 ; X64-NEXT:    vfixupimmpd $4, %ymm2, %ymm1, %ymm3 {%k1} # encoding: [0x62,0xf3,0xf5,0x29,0x54,0xda,0x04]
5002 ; X64-NEXT:    vxorpd %xmm4, %xmm4, %xmm4 # encoding: [0xc5,0xd9,0x57,0xe4]
5003 ; X64-NEXT:    vfixupimmpd $5, %ymm2, %ymm1, %ymm4 {%k1} {z} # encoding: [0x62,0xf3,0xf5,0xa9,0x54,0xe2,0x05]
5004 ; X64-NEXT:    vaddpd %ymm4, %ymm3, %ymm3 # EVEX TO VEX Compression encoding: [0xc5,0xe5,0x58,0xdc]
5005 ; X64-NEXT:    vfixupimmpd $3, %ymm2, %ymm1, %ymm0 # encoding: [0x62,0xf3,0xf5,0x28,0x54,0xc2,0x03]
5006 ; X64-NEXT:    vaddpd %ymm0, %ymm3, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xe5,0x58,0xc0]
5007 ; X64-NEXT:    retq # encoding: [0xc3]
5020 ; X86:       # %bb.0:
5021 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
5022 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
5023 ; X86-NEXT:    vmovapd %ymm0, %ymm3 # EVEX TO VEX Compression encoding: [0xc5,0xfd,0x28,0xd8]
5024 ; X86-NEXT:    vfixupimmpd $5, %ymm2, %ymm1, %ymm3 {%k1} {z} # encoding: [0x62,0xf3,0xf5,0xa9,0x54,0xda,0x05]
5025 ; X86-NEXT:    vxorpd %xmm4, %xmm4, %xmm4 # encoding: [0xc5,0xd9,0x57,0xe4]
5026 ; X86-NEXT:    vmovapd %ymm0, %ymm5 # EVEX TO VEX Compression encoding: [0xc5,0xfd,0x28,0xe8]
5027 ; X86-NEXT:    vfixupimmpd $4, %ymm4, %ymm1, %ymm5 {%k1} {z} # encoding: [0x62,0xf3,0xf5,0xa9,0x54,0xec,0x04]
5028 ; X86-NEXT:    vaddpd %ymm5, %ymm3, %ymm3 # EVEX TO VEX Compression encoding: [0xc5,0xe5,0x58,0xdd]
5029 ; X86-NEXT:    vfixupimmpd $3, %ymm2, %ymm1, %ymm0 # encoding: [0x62,0xf3,0xf5,0x28,0x54,0xc2,0x03]
5030 ; X86-NEXT:    vaddpd %ymm0, %ymm3, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xe5,0x58,0xc0]
5031 ; X86-NEXT:    retl # encoding: [0xc3]
5034 ; X64:       # %bb.0:
5035 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
5036 ; X64-NEXT:    vmovapd %ymm0, %ymm3 # EVEX TO VEX Compression encoding: [0xc5,0xfd,0x28,0xd8]
5037 ; X64-NEXT:    vfixupimmpd $5, %ymm2, %ymm1, %ymm3 {%k1} {z} # encoding: [0x62,0xf3,0xf5,0xa9,0x54,0xda,0x05]
5038 ; X64-NEXT:    vxorpd %xmm4, %xmm4, %xmm4 # encoding: [0xc5,0xd9,0x57,0xe4]
5039 ; X64-NEXT:    vmovapd %ymm0, %ymm5 # EVEX TO VEX Compression encoding: [0xc5,0xfd,0x28,0xe8]
5040 ; X64-NEXT:    vfixupimmpd $4, %ymm4, %ymm1, %ymm5 {%k1} {z} # encoding: [0x62,0xf3,0xf5,0xa9,0x54,0xec,0x04]
5041 ; X64-NEXT:    vaddpd %ymm5, %ymm3, %ymm3 # EVEX TO VEX Compression encoding: [0xc5,0xe5,0x58,0xdd]
5042 ; X64-NEXT:    vfixupimmpd $3, %ymm2, %ymm1, %ymm0 # encoding: [0x62,0xf3,0xf5,0x28,0x54,0xc2,0x03]
5043 ; X64-NEXT:    vaddpd %ymm0, %ymm3, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xe5,0x58,0xc0]
5044 ; X64-NEXT:    retq # encoding: [0xc3]
5057 ; X86:       # %bb.0:
5058 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
5059 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
5060 ; X86-NEXT:    vmovaps %xmm0, %xmm3 # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x28,0xd8]
5061 ; X86-NEXT:    vfixupimmps $5, %xmm2, %xmm1, %xmm3 {%k1} # encoding: [0x62,0xf3,0x75,0x09,0x54,0xda,0x05]
5062 ; X86-NEXT:    vxorps %xmm4, %xmm4, %xmm4 # encoding: [0xc5,0xd8,0x57,0xe4]
5063 ; X86-NEXT:    vmovaps %xmm0, %xmm5 # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x28,0xe8]
5064 ; X86-NEXT:    vfixupimmps $6, %xmm4, %xmm1, %xmm5 {%k1} # encoding: [0x62,0xf3,0x75,0x09,0x54,0xec,0x06]
5065 ; X86-NEXT:    vaddps %xmm5, %xmm3, %xmm3 # EVEX TO VEX Compression encoding: [0xc5,0xe0,0x58,0xdd]
5066 ; X86-NEXT:    vfixupimmps $7, %xmm2, %xmm1, %xmm0 # encoding: [0x62,0xf3,0x75,0x08,0x54,0xc2,0x07]
5067 ; X86-NEXT:    vaddps %xmm0, %xmm3, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xe0,0x58,0xc0]
5068 ; X86-NEXT:    retl # encoding: [0xc3]
5071 ; X64:       # %bb.0:
5072 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
5073 ; X64-NEXT:    vmovaps %xmm0, %xmm3 # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x28,0xd8]
5074 ; X64-NEXT:    vfixupimmps $5, %xmm2, %xmm1, %xmm3 {%k1} # encoding: [0x62,0xf3,0x75,0x09,0x54,0xda,0x05]
5075 ; X64-NEXT:    vxorps %xmm4, %xmm4, %xmm4 # encoding: [0xc5,0xd8,0x57,0xe4]
5076 ; X64-NEXT:    vmovaps %xmm0, %xmm5 # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x28,0xe8]
5077 ; X64-NEXT:    vfixupimmps $6, %xmm4, %xmm1, %xmm5 {%k1} # encoding: [0x62,0xf3,0x75,0x09,0x54,0xec,0x06]
5078 ; X64-NEXT:    vaddps %xmm5, %xmm3, %xmm3 # EVEX TO VEX Compression encoding: [0xc5,0xe0,0x58,0xdd]
5079 ; X64-NEXT:    vfixupimmps $7, %xmm2, %xmm1, %xmm0 # encoding: [0x62,0xf3,0x75,0x08,0x54,0xc2,0x07]
5080 ; X64-NEXT:    vaddps %xmm0, %xmm3, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xe0,0x58,0xc0]
5081 ; X64-NEXT:    retq # encoding: [0xc3]
5094 ; X86:       # %bb.0:
5095 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
5096 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
5097 ; X86-NEXT:    vmovaps %xmm0, %xmm3 # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x28,0xd8]
5098 ; X86-NEXT:    vfixupimmps $5, %xmm2, %xmm1, %xmm3 {%k1} {z} # encoding: [0x62,0xf3,0x75,0x89,0x54,0xda,0x05]
5099 ; X86-NEXT:    vxorps %xmm4, %xmm4, %xmm4 # encoding: [0xc5,0xd8,0x57,0xe4]
5100 ; X86-NEXT:    vmovaps %xmm0, %xmm5 # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x28,0xe8]
5101 ; X86-NEXT:    vfixupimmps $6, %xmm4, %xmm1, %xmm5 {%k1} {z} # encoding: [0x62,0xf3,0x75,0x89,0x54,0xec,0x06]
5102 ; X86-NEXT:    vaddps %xmm5, %xmm3, %xmm3 # EVEX TO VEX Compression encoding: [0xc5,0xe0,0x58,0xdd]
5103 ; X86-NEXT:    vfixupimmps $7, %xmm2, %xmm1, %xmm0 # encoding: [0x62,0xf3,0x75,0x08,0x54,0xc2,0x07]
5104 ; X86-NEXT:    vaddps %xmm0, %xmm3, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xe0,0x58,0xc0]
5105 ; X86-NEXT:    retl # encoding: [0xc3]
5108 ; X64:       # %bb.0:
5109 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
5110 ; X64-NEXT:    vmovaps %xmm0, %xmm3 # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x28,0xd8]
5111 ; X64-NEXT:    vfixupimmps $5, %xmm2, %xmm1, %xmm3 {%k1} {z} # encoding: [0x62,0xf3,0x75,0x89,0x54,0xda,0x05]
5112 ; X64-NEXT:    vxorps %xmm4, %xmm4, %xmm4 # encoding: [0xc5,0xd8,0x57,0xe4]
5113 ; X64-NEXT:    vmovaps %xmm0, %xmm5 # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x28,0xe8]
5114 ; X64-NEXT:    vfixupimmps $6, %xmm4, %xmm1, %xmm5 {%k1} {z} # encoding: [0x62,0xf3,0x75,0x89,0x54,0xec,0x06]
5115 ; X64-NEXT:    vaddps %xmm5, %xmm3, %xmm3 # EVEX TO VEX Compression encoding: [0xc5,0xe0,0x58,0xdd]
5116 ; X64-NEXT:    vfixupimmps $7, %xmm2, %xmm1, %xmm0 # encoding: [0x62,0xf3,0x75,0x08,0x54,0xc2,0x07]
5117 ; X64-NEXT:    vaddps %xmm0, %xmm3, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xe0,0x58,0xc0]
5118 ; X64-NEXT:    retq # encoding: [0xc3]
5131 ; X86:       # %bb.0:
5132 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
5133 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
5134 ; X86-NEXT:    vmovaps %ymm0, %ymm3 # EVEX TO VEX Compression encoding: [0xc5,0xfc,0x28,0xd8]
5135 ; X86-NEXT:    vfixupimmps $5, %ymm2, %ymm1, %ymm3 {%k1} # encoding: [0x62,0xf3,0x75,0x29,0x54,0xda,0x05]
5136 ; X86-NEXT:    vxorps %xmm4, %xmm4, %xmm4 # encoding: [0xc5,0xd8,0x57,0xe4]
5137 ; X86-NEXT:    vmovaps %ymm0, %ymm5 # EVEX TO VEX Compression encoding: [0xc5,0xfc,0x28,0xe8]
5138 ; X86-NEXT:    vfixupimmps $6, %ymm4, %ymm1, %ymm5 {%k1} # encoding: [0x62,0xf3,0x75,0x29,0x54,0xec,0x06]
5139 ; X86-NEXT:    vaddps %ymm5, %ymm3, %ymm3 # EVEX TO VEX Compression encoding: [0xc5,0xe4,0x58,0xdd]
5140 ; X86-NEXT:    vfixupimmps $7, %ymm2, %ymm1, %ymm0 # encoding: [0x62,0xf3,0x75,0x28,0x54,0xc2,0x07]
5141 ; X86-NEXT:    vaddps %ymm0, %ymm3, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xe4,0x58,0xc0]
5142 ; X86-NEXT:    retl # encoding: [0xc3]
5145 ; X64:       # %bb.0:
5146 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
5147 ; X64-NEXT:    vmovaps %ymm0, %ymm3 # EVEX TO VEX Compression encoding: [0xc5,0xfc,0x28,0xd8]
5148 ; X64-NEXT:    vfixupimmps $5, %ymm2, %ymm1, %ymm3 {%k1} # encoding: [0x62,0xf3,0x75,0x29,0x54,0xda,0x05]
5149 ; X64-NEXT:    vxorps %xmm4, %xmm4, %xmm4 # encoding: [0xc5,0xd8,0x57,0xe4]
5150 ; X64-NEXT:    vmovaps %ymm0, %ymm5 # EVEX TO VEX Compression encoding: [0xc5,0xfc,0x28,0xe8]
5151 ; X64-NEXT:    vfixupimmps $6, %ymm4, %ymm1, %ymm5 {%k1} # encoding: [0x62,0xf3,0x75,0x29,0x54,0xec,0x06]
5152 ; X64-NEXT:    vaddps %ymm5, %ymm3, %ymm3 # EVEX TO VEX Compression encoding: [0xc5,0xe4,0x58,0xdd]
5153 ; X64-NEXT:    vfixupimmps $7, %ymm2, %ymm1, %ymm0 # encoding: [0x62,0xf3,0x75,0x28,0x54,0xc2,0x07]
5154 ; X64-NEXT:    vaddps %ymm0, %ymm3, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xe4,0x58,0xc0]
5155 ; X64-NEXT:    retq # encoding: [0xc3]
5168 ; X86:       # %bb.0:
5169 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
5170 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
5171 ; X86-NEXT:    vmovaps %ymm0, %ymm3 # EVEX TO VEX Compression encoding: [0xc5,0xfc,0x28,0xd8]
5172 ; X86-NEXT:    vfixupimmps $5, %ymm2, %ymm1, %ymm3 {%k1} {z} # encoding: [0x62,0xf3,0x75,0xa9,0x54,0xda,0x05]
5173 ; X86-NEXT:    vxorps %xmm4, %xmm4, %xmm4 # encoding: [0xc5,0xd8,0x57,0xe4]
5174 ; X86-NEXT:    vmovaps %ymm0, %ymm5 # EVEX TO VEX Compression encoding: [0xc5,0xfc,0x28,0xe8]
5175 ; X86-NEXT:    vfixupimmps $6, %ymm4, %ymm1, %ymm5 {%k1} {z} # encoding: [0x62,0xf3,0x75,0xa9,0x54,0xec,0x06]
5176 ; X86-NEXT:    vaddps %ymm5, %ymm3, %ymm3 # EVEX TO VEX Compression encoding: [0xc5,0xe4,0x58,0xdd]
5177 ; X86-NEXT:    vfixupimmps $7, %ymm2, %ymm1, %ymm0 # encoding: [0x62,0xf3,0x75,0x28,0x54,0xc2,0x07]
5178 ; X86-NEXT:    vaddps %ymm0, %ymm3, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xe4,0x58,0xc0]
5179 ; X86-NEXT:    retl # encoding: [0xc3]
5182 ; X64:       # %bb.0:
5183 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
5184 ; X64-NEXT:    vmovaps %ymm0, %ymm3 # EVEX TO VEX Compression encoding: [0xc5,0xfc,0x28,0xd8]
5185 ; X64-NEXT:    vfixupimmps $5, %ymm2, %ymm1, %ymm3 {%k1} {z} # encoding: [0x62,0xf3,0x75,0xa9,0x54,0xda,0x05]
5186 ; X64-NEXT:    vxorps %xmm4, %xmm4, %xmm4 # encoding: [0xc5,0xd8,0x57,0xe4]
5187 ; X64-NEXT:    vmovaps %ymm0, %ymm5 # EVEX TO VEX Compression encoding: [0xc5,0xfc,0x28,0xe8]
5188 ; X64-NEXT:    vfixupimmps $6, %ymm4, %ymm1, %ymm5 {%k1} {z} # encoding: [0x62,0xf3,0x75,0xa9,0x54,0xec,0x06]
5189 ; X64-NEXT:    vaddps %ymm5, %ymm3, %ymm3 # EVEX TO VEX Compression encoding: [0xc5,0xe4,0x58,0xdd]
5190 ; X64-NEXT:    vfixupimmps $7, %ymm2, %ymm1, %ymm0 # encoding: [0x62,0xf3,0x75,0x28,0x54,0xc2,0x07]
5191 ; X64-NEXT:    vaddps %ymm0, %ymm3, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xe4,0x58,0xc0]
5192 ; X64-NEXT:    retq # encoding: [0xc3]
5203 ; CHECK:       # %bb.0:
5204 ; CHECK-NEXT:    vpsraq %xmm1, %xmm0, %xmm0 # encoding: [0x62,0xf1,0xfd,0x08,0xe2,0xc1]
5205 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
5211 ; X86:       # %bb.0:
5212 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
5213 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
5214 ; X86-NEXT:    vpsraq %xmm1, %xmm0, %xmm2 {%k1} # encoding: [0x62,0xf1,0xfd,0x09,0xe2,0xd1]
5215 ; X86-NEXT:    vmovdqa %xmm2, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6f,0xc2]
5216 ; X86-NEXT:    retl # encoding: [0xc3]
5219 ; X64:       # %bb.0:
5220 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
5221 ; X64-NEXT:    vpsraq %xmm1, %xmm0, %xmm2 {%k1} # encoding: [0x62,0xf1,0xfd,0x09,0xe2,0xd1]
5222 ; X64-NEXT:    vmovdqa %xmm2, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6f,0xc2]
5223 ; X64-NEXT:    retq # encoding: [0xc3]
5226   %mask.extract = shufflevector <8 x i1> %mask.cast, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
5232 ; X86:       # %bb.0:
5233 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
5234 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
5235 ; X86-NEXT:    vpsraq %xmm1, %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf1,0xfd,0x89,0xe2,0xc1]
5236 ; X86-NEXT:    retl # encoding: [0xc3]
5239 ; X64:       # %bb.0:
5240 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
5241 ; X64-NEXT:    vpsraq %xmm1, %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf1,0xfd,0x89,0xe2,0xc1]
5242 ; X64-NEXT:    retq # encoding: [0xc3]
5245   %mask.extract = shufflevector <8 x i1> %mask.cast, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
5254 ; CHECK:       # %bb.0:
5255 ; CHECK-NEXT:    vpsraq %xmm1, %ymm0, %ymm0 # encoding: [0x62,0xf1,0xfd,0x28,0xe2,0xc1]
5256 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
5262 ; X86:       # %bb.0:
5263 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
5264 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
5265 ; X86-NEXT:    vpsraq %xmm1, %ymm0, %ymm2 {%k1} # encoding: [0x62,0xf1,0xfd,0x29,0xe2,0xd1]
5266 ; X86-NEXT:    vmovdqa %ymm2, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfd,0x6f,0xc2]
5267 ; X86-NEXT:    retl # encoding: [0xc3]
5270 ; X64:       # %bb.0:
5271 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
5272 ; X64-NEXT:    vpsraq %xmm1, %ymm0, %ymm2 {%k1} # encoding: [0x62,0xf1,0xfd,0x29,0xe2,0xd1]
5273 ; X64-NEXT:    vmovdqa %ymm2, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfd,0x6f,0xc2]
5274 ; X64-NEXT:    retq # encoding: [0xc3]
5277   %mask.extract = shufflevector <8 x i1> %mask.cast, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
5283 ; X86:       # %bb.0:
5284 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
5285 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
5286 ; X86-NEXT:    vpsraq %xmm1, %ymm0, %ymm0 {%k1} {z} # encoding: [0x62,0xf1,0xfd,0xa9,0xe2,0xc1]
5287 ; X86-NEXT:    retl # encoding: [0xc3]
5290 ; X64:       # %bb.0:
5291 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
5292 ; X64-NEXT:    vpsraq %xmm1, %ymm0, %ymm0 {%k1} {z} # encoding: [0x62,0xf1,0xfd,0xa9,0xe2,0xc1]
5293 ; X64-NEXT:    retq # encoding: [0xc3]
5296   %mask.extract = shufflevector <8 x i1> %mask.cast, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
5305 ; CHECK:       # %bb.0:
5306 ; CHECK-NEXT:    vpsraq $7, %xmm0, %xmm0 # encoding: [0x62,0xf1,0xfd,0x08,0x72,0xe0,0x07]
5307 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
5313 ; X86:       # %bb.0:
5314 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
5315 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
5316 ; X86-NEXT:    vpsraq $7, %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf1,0xf5,0x09,0x72,0xe0,0x07]
5317 ; X86-NEXT:    vmovdqa %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6f,0xc1]
5318 ; X86-NEXT:    retl # encoding: [0xc3]
5321 ; X64:       # %bb.0:
5322 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
5323 ; X64-NEXT:    vpsraq $7, %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf1,0xf5,0x09,0x72,0xe0,0x07]
5324 ; X64-NEXT:    vmovdqa %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6f,0xc1]
5325 ; X64-NEXT:    retq # encoding: [0xc3]
5328   %mask.extract = shufflevector <8 x i1> %mask.cast, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
5334 ; X86:       # %bb.0:
5335 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
5336 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
5337 ; X86-NEXT:    vpsraq $7, %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf1,0xfd,0x89,0x72,0xe0,0x07]
5338 ; X86-NEXT:    retl # encoding: [0xc3]
5341 ; X64:       # %bb.0:
5342 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
5343 ; X64-NEXT:    vpsraq $7, %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf1,0xfd,0x89,0x72,0xe0,0x07]
5344 ; X64-NEXT:    retq # encoding: [0xc3]
5347   %mask.extract = shufflevector <8 x i1> %mask.cast, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
5356 ; CHECK:       # %bb.0:
5357 ; CHECK-NEXT:    vpsraq $7, %ymm0, %ymm0 # encoding: [0x62,0xf1,0xfd,0x28,0x72,0xe0,0x07]
5358 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
5364 ; X86:       # %bb.0:
5365 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
5366 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
5367 ; X86-NEXT:    vpsraq $7, %ymm0, %ymm1 {%k1} # encoding: [0x62,0xf1,0xf5,0x29,0x72,0xe0,0x07]
5368 ; X86-NEXT:    vmovdqa %ymm1, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfd,0x6f,0xc1]
5369 ; X86-NEXT:    retl # encoding: [0xc3]
5372 ; X64:       # %bb.0:
5373 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
5374 ; X64-NEXT:    vpsraq $7, %ymm0, %ymm1 {%k1} # encoding: [0x62,0xf1,0xf5,0x29,0x72,0xe0,0x07]
5375 ; X64-NEXT:    vmovdqa %ymm1, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfd,0x6f,0xc1]
5376 ; X64-NEXT:    retq # encoding: [0xc3]
5379   %mask.extract = shufflevector <8 x i1> %mask.cast, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
5385 ; X86:       # %bb.0:
5386 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
5387 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
5388 ; X86-NEXT:    vpsraq $7, %ymm0, %ymm0 {%k1} {z} # encoding: [0x62,0xf1,0xfd,0xa9,0x72,0xe0,0x07]
5389 ; X86-NEXT:    retl # encoding: [0xc3]
5392 ; X64:       # %bb.0:
5393 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
5394 ; X64-NEXT:    vpsraq $7, %ymm0, %ymm0 {%k1} {z} # encoding: [0x62,0xf1,0xfd,0xa9,0x72,0xe0,0x07]
5395 ; X64-NEXT:    retq # encoding: [0xc3]
5398   %mask.extract = shufflevector <8 x i1> %mask.cast, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
5406 ; CHECK:       # %bb.0:
5407 ; CHECK-NEXT:    vpsravq %xmm1, %xmm0, %xmm0 # encoding: [0x62,0xf2,0xfd,0x08,0x46,0xc1]
5408 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
5415 ; X86:       # %bb.0:
5416 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
5417 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
5418 ; X86-NEXT:    vpsravq %xmm1, %xmm0, %xmm2 {%k1} # encoding: [0x62,0xf2,0xfd,0x09,0x46,0xd1]
5419 ; X86-NEXT:    vmovdqa %xmm2, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6f,0xc2]
5420 ; X86-NEXT:    retl # encoding: [0xc3]
5423 ; X64:       # %bb.0:
5424 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
5425 ; X64-NEXT:    vpsravq %xmm1, %xmm0, %xmm2 {%k1} # encoding: [0x62,0xf2,0xfd,0x09,0x46,0xd1]
5426 ; X64-NEXT:    vmovdqa %xmm2, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6f,0xc2]
5427 ; X64-NEXT:    retq # encoding: [0xc3]
5430   %mask.extract = shufflevector <8 x i1> %mask.cast, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
5437 ; X86:       # %bb.0:
5438 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
5439 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
5440 ; X86-NEXT:    vpsravq %xmm1, %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0xfd,0x89,0x46,0xc1]
5441 ; X86-NEXT:    retl # encoding: [0xc3]
5444 ; X64:       # %bb.0:
5445 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
5446 ; X64-NEXT:    vpsravq %xmm1, %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0xfd,0x89,0x46,0xc1]
5447 ; X64-NEXT:    retq # encoding: [0xc3]
5450   %mask.extract = shufflevector <8 x i1> %mask.cast, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
5459 ; CHECK:       # %bb.0:
5460 ; CHECK-NEXT:    vpsravq %ymm1, %ymm0, %ymm0 # encoding: [0x62,0xf2,0xfd,0x28,0x46,0xc1]
5461 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
5468 ; X86:       # %bb.0:
5469 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
5470 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
5471 ; X86-NEXT:    vpsravq %ymm1, %ymm0, %ymm2 {%k1} # encoding: [0x62,0xf2,0xfd,0x29,0x46,0xd1]
5472 ; X86-NEXT:    vmovdqa %ymm2, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfd,0x6f,0xc2]
5473 ; X86-NEXT:    retl # encoding: [0xc3]
5476 ; X64:       # %bb.0:
5477 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
5478 ; X64-NEXT:    vpsravq %ymm1, %ymm0, %ymm2 {%k1} # encoding: [0x62,0xf2,0xfd,0x29,0x46,0xd1]
5479 ; X64-NEXT:    vmovdqa %ymm2, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfd,0x6f,0xc2]
5480 ; X64-NEXT:    retq # encoding: [0xc3]
5483   %mask.extract = shufflevector <8 x i1> %mask.cast, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
5490 ; X86:       # %bb.0:
5491 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
5492 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
5493 ; X86-NEXT:    vpsravq %ymm1, %ymm0, %ymm0 {%k1} {z} # encoding: [0x62,0xf2,0xfd,0xa9,0x46,0xc1]
5494 ; X86-NEXT:    retl # encoding: [0xc3]
5497 ; X64:       # %bb.0:
5498 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
5499 ; X64-NEXT:    vpsravq %ymm1, %ymm0, %ymm0 {%k1} {z} # encoding: [0x62,0xf2,0xfd,0xa9,0x46,0xc1]
5500 ; X64-NEXT:    retq # encoding: [0xc3]
5503   %mask.extract = shufflevector <8 x i1> %mask.cast, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
5512 ; CHECK:       # %bb.0:
5513 ; CHECK-NEXT:    vfmadd213ps %ymm2, %ymm1, %ymm0 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x75,0xa8,0xc2]
5515 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
5522 ; X86:       # %bb.0:
5523 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
5524 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
5525 ; X86-NEXT:    vfmadd132ps %ymm1, %ymm2, %ymm0 {%k1} # encoding: [0x62,0xf2,0x6d,0x29,0x98,0xc1]
5527 ; X86-NEXT:    retl # encoding: [0xc3]
5530 ; X64:       # %bb.0:
5531 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
5532 ; X64-NEXT:    vfmadd132ps %ymm1, %ymm2, %ymm0 {%k1} # encoding: [0x62,0xf2,0x6d,0x29,0x98,0xc1]
5534 ; X64-NEXT:    retq # encoding: [0xc3]
5543 ; CHECK:       # %bb.0:
5544 ; CHECK-NEXT:    vfmadd213ps %xmm2, %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x71,0xa8,0xc2]
5546 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
5553 ; X86:       # %bb.0:
5554 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
5555 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
5556 ; X86-NEXT:    vfmadd132ps %xmm1, %xmm2, %xmm0 {%k1} # encoding: [0x62,0xf2,0x6d,0x09,0x98,0xc1]
5558 ; X86-NEXT:    retl # encoding: [0xc3]
5561 ; X64:       # %bb.0:
5562 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
5563 ; X64-NEXT:    vfmadd132ps %xmm1, %xmm2, %xmm0 {%k1} # encoding: [0x62,0xf2,0x6d,0x09,0x98,0xc1]
5565 ; X64-NEXT:    retq # encoding: [0xc3]
5568   %extract = shufflevector <8 x i1> %2, <8 x i1> %2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
5575 ; CHECK:       # %bb.0:
5576 ; CHECK-NEXT:    vfmadd213pd %ymm2, %ymm1, %ymm0 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0xf5,0xa8,0xc2]
5578 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
5585 ; X86:       # %bb.0:
5586 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
5587 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
5588 ; X86-NEXT:    vfmadd132pd %ymm1, %ymm2, %ymm0 {%k1} # encoding: [0x62,0xf2,0xed,0x29,0x98,0xc1]
5590 ; X86-NEXT:    retl # encoding: [0xc3]
5593 ; X64:       # %bb.0:
5594 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
5595 ; X64-NEXT:    vfmadd132pd %ymm1, %ymm2, %ymm0 {%k1} # encoding: [0x62,0xf2,0xed,0x29,0x98,0xc1]
5597 ; X64-NEXT:    retq # encoding: [0xc3]
5600   %extract = shufflevector <8 x i1> %2, <8 x i1> %2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
5607 ; CHECK:       # %bb.0:
5608 ; CHECK-NEXT:    vfmadd213pd %xmm2, %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0xf1,0xa8,0xc2]
5610 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
5617 ; X86:       # %bb.0:
5618 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
5619 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
5620 ; X86-NEXT:    vfmadd132pd %xmm1, %xmm2, %xmm0 {%k1} # encoding: [0x62,0xf2,0xed,0x09,0x98,0xc1]
5622 ; X86-NEXT:    retl # encoding: [0xc3]
5625 ; X64:       # %bb.0:
5626 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
5627 ; X64-NEXT:    vfmadd132pd %xmm1, %xmm2, %xmm0 {%k1} # encoding: [0x62,0xf2,0xed,0x09,0x98,0xc1]
5629 ; X64-NEXT:    retq # encoding: [0xc3]
5632   %extract = shufflevector <8 x i1> %2, <8 x i1> %2, <2 x i32> <i32 0, i32 1>
5639 ; X86:       # %bb.0:
5640 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
5641 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
5642 ; X86-NEXT:    vfmadd231pd %xmm1, %xmm0, %xmm2 {%k1} # encoding: [0x62,0xf2,0xfd,0x09,0xb8,0xd1]
5644 ; X86-NEXT:    vmovapd %xmm2, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf9,0x28,0xc2]
5645 ; X86-NEXT:    retl # encoding: [0xc3]
5648 ; X64:       # %bb.0:
5649 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
5650 ; X64-NEXT:    vfmadd231pd %xmm1, %xmm0, %xmm2 {%k1} # encoding: [0x62,0xf2,0xfd,0x09,0xb8,0xd1]
5652 ; X64-NEXT:    vmovapd %xmm2, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf9,0x28,0xc2]
5653 ; X64-NEXT:    retq # encoding: [0xc3]
5656   %extract = shufflevector <8 x i1> %2, <8 x i1> %2, <2 x i32> <i32 0, i32 1>
5663 ; X86:       # %bb.0:
5664 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
5665 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
5666 ; X86-NEXT:    vfmadd213pd %xmm2, %xmm1, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0xf5,0x89,0xa8,0xc2]
5668 ; X86-NEXT:    retl # encoding: [0xc3]
5671 ; X64:       # %bb.0:
5672 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
5673 ; X64-NEXT:    vfmadd213pd %xmm2, %xmm1, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0xf5,0x89,0xa8,0xc2]
5675 ; X64-NEXT:    retq # encoding: [0xc3]
5678   %extract = shufflevector <8 x i1> %2, <8 x i1> %2, <2 x i32> <i32 0, i32 1>
5685 ; X86:       # %bb.0:
5686 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
5687 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
5688 ; X86-NEXT:    vfmadd231pd %ymm1, %ymm0, %ymm2 {%k1} # encoding: [0x62,0xf2,0xfd,0x29,0xb8,0xd1]
5690 ; X86-NEXT:    vmovapd %ymm2, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfd,0x28,0xc2]
5691 ; X86-NEXT:    retl # encoding: [0xc3]
5694 ; X64:       # %bb.0:
5695 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
5696 ; X64-NEXT:    vfmadd231pd %ymm1, %ymm0, %ymm2 {%k1} # encoding: [0x62,0xf2,0xfd,0x29,0xb8,0xd1]
5698 ; X64-NEXT:    vmovapd %ymm2, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfd,0x28,0xc2]
5699 ; X64-NEXT:    retq # encoding: [0xc3]
5702   %extract = shufflevector <8 x i1> %2, <8 x i1> %2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
5709 ; X86:       # %bb.0:
5710 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
5711 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
5712 ; X86-NEXT:    vfmadd213pd %ymm2, %ymm1, %ymm0 {%k1} {z} # encoding: [0x62,0xf2,0xf5,0xa9,0xa8,0xc2]
5714 ; X86-NEXT:    retl # encoding: [0xc3]
5717 ; X64:       # %bb.0:
5718 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
5719 ; X64-NEXT:    vfmadd213pd %ymm2, %ymm1, %ymm0 {%k1} {z} # encoding: [0x62,0xf2,0xf5,0xa9,0xa8,0xc2]
5721 ; X64-NEXT:    retq # encoding: [0xc3]
5724   %extract = shufflevector <8 x i1> %2, <8 x i1> %2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
5731 ; X86:       # %bb.0:
5732 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
5733 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
5734 ; X86-NEXT:    vfmadd231ps %xmm1, %xmm0, %xmm2 {%k1} # encoding: [0x62,0xf2,0x7d,0x09,0xb8,0xd1]
5736 ; X86-NEXT:    vmovaps %xmm2, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x28,0xc2]
5737 ; X86-NEXT:    retl # encoding: [0xc3]
5740 ; X64:       # %bb.0:
5741 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
5742 ; X64-NEXT:    vfmadd231ps %xmm1, %xmm0, %xmm2 {%k1} # encoding: [0x62,0xf2,0x7d,0x09,0xb8,0xd1]
5744 ; X64-NEXT:    vmovaps %xmm2, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x28,0xc2]
5745 ; X64-NEXT:    retq # encoding: [0xc3]
5748   %extract = shufflevector <8 x i1> %2, <8 x i1> %2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
5755 ; X86:       # %bb.0:
5756 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
5757 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
5758 ; X86-NEXT:    vfmadd213ps %xmm2, %xmm1, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x75,0x89,0xa8,0xc2]
5760 ; X86-NEXT:    retl # encoding: [0xc3]
5763 ; X64:       # %bb.0:
5764 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
5765 ; X64-NEXT:    vfmadd213ps %xmm2, %xmm1, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x75,0x89,0xa8,0xc2]
5767 ; X64-NEXT:    retq # encoding: [0xc3]
5770   %extract = shufflevector <8 x i1> %2, <8 x i1> %2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
5777 ; X86:       # %bb.0:
5778 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
5779 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
5780 ; X86-NEXT:    vfmadd231ps %ymm1, %ymm0, %ymm2 {%k1} # encoding: [0x62,0xf2,0x7d,0x29,0xb8,0xd1]
5782 ; X86-NEXT:    vmovaps %ymm2, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfc,0x28,0xc2]
5783 ; X86-NEXT:    retl # encoding: [0xc3]
5786 ; X64:       # %bb.0:
5787 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
5788 ; X64-NEXT:    vfmadd231ps %ymm1, %ymm0, %ymm2 {%k1} # encoding: [0x62,0xf2,0x7d,0x29,0xb8,0xd1]
5790 ; X64-NEXT:    vmovaps %ymm2, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfc,0x28,0xc2]
5791 ; X64-NEXT:    retq # encoding: [0xc3]
5800 ; X86:       # %bb.0:
5801 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
5802 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
5803 ; X86-NEXT:    vfmadd213ps %ymm2, %ymm1, %ymm0 {%k1} {z} # encoding: [0x62,0xf2,0x75,0xa9,0xa8,0xc2]
5805 ; X86-NEXT:    retl # encoding: [0xc3]
5808 ; X64:       # %bb.0:
5809 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
5810 ; X64-NEXT:    vfmadd213ps %ymm2, %ymm1, %ymm0 {%k1} {z} # encoding: [0x62,0xf2,0x75,0xa9,0xa8,0xc2]
5812 ; X64-NEXT:    retq # encoding: [0xc3]
5821 ; X86:       # %bb.0:
5822 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
5823 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
5824 ; X86-NEXT:    vfmsub231pd %xmm1, %xmm0, %xmm2 {%k1} # encoding: [0x62,0xf2,0xfd,0x09,0xba,0xd1]
5826 ; X86-NEXT:    vmovapd %xmm2, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf9,0x28,0xc2]
5827 ; X86-NEXT:    retl # encoding: [0xc3]
5830 ; X64:       # %bb.0:
5831 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
5832 ; X64-NEXT:    vfmsub231pd %xmm1, %xmm0, %xmm2 {%k1} # encoding: [0x62,0xf2,0xfd,0x09,0xba,0xd1]
5834 ; X64-NEXT:    vmovapd %xmm2, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf9,0x28,0xc2]
5835 ; X64-NEXT:    retq # encoding: [0xc3]
5839   %extract = shufflevector <8 x i1> %3, <8 x i1> %3, <2 x i32> <i32 0, i32 1>
5846 ; X86:       # %bb.0:
5847 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
5848 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
5849 ; X86-NEXT:    vfmsub231pd %ymm1, %ymm0, %ymm2 {%k1} # encoding: [0x62,0xf2,0xfd,0x29,0xba,0xd1]
5851 ; X86-NEXT:    vmovapd %ymm2, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfd,0x28,0xc2]
5852 ; X86-NEXT:    retl # encoding: [0xc3]
5855 ; X64:       # %bb.0:
5856 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
5857 ; X64-NEXT:    vfmsub231pd %ymm1, %ymm0, %ymm2 {%k1} # encoding: [0x62,0xf2,0xfd,0x29,0xba,0xd1]
5859 ; X64-NEXT:    vmovapd %ymm2, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfd,0x28,0xc2]
5860 ; X64-NEXT:    retq # encoding: [0xc3]
5864   %extract = shufflevector <8 x i1> %3, <8 x i1> %3, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
5871 ; X86:       # %bb.0:
5872 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
5873 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
5874 ; X86-NEXT:    vfmsub231ps %xmm1, %xmm0, %xmm2 {%k1} # encoding: [0x62,0xf2,0x7d,0x09,0xba,0xd1]
5876 ; X86-NEXT:    vmovaps %xmm2, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x28,0xc2]
5877 ; X86-NEXT:    retl # encoding: [0xc3]
5880 ; X64:       # %bb.0:
5881 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
5882 ; X64-NEXT:    vfmsub231ps %xmm1, %xmm0, %xmm2 {%k1} # encoding: [0x62,0xf2,0x7d,0x09,0xba,0xd1]
5884 ; X64-NEXT:    vmovaps %xmm2, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x28,0xc2]
5885 ; X64-NEXT:    retq # encoding: [0xc3]
5889   %extract = shufflevector <8 x i1> %3, <8 x i1> %3, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
5896 ; X86:       # %bb.0:
5897 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
5898 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
5899 ; X86-NEXT:    vfmsub231ps %ymm1, %ymm0, %ymm2 {%k1} # encoding: [0x62,0xf2,0x7d,0x29,0xba,0xd1]
5901 ; X86-NEXT:    vmovaps %ymm2, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfc,0x28,0xc2]
5902 ; X86-NEXT:    retl # encoding: [0xc3]
5905 ; X64:       # %bb.0:
5906 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
5907 ; X64-NEXT:    vfmsub231ps %ymm1, %ymm0, %ymm2 {%k1} # encoding: [0x62,0xf2,0x7d,0x29,0xba,0xd1]
5909 ; X64-NEXT:    vmovaps %ymm2, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfc,0x28,0xc2]
5910 ; X64-NEXT:    retq # encoding: [0xc3]
5920 ; CHECK:       # %bb.0:
5921 ; CHECK-NEXT:    vfnmadd213ps %ymm2, %ymm1, %ymm0 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x75,0xac,0xc2]
5923 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
5931 ; X86:       # %bb.0:
5932 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
5933 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
5934 ; X86-NEXT:    vfnmadd132ps %ymm1, %ymm2, %ymm0 {%k1} # encoding: [0x62,0xf2,0x6d,0x29,0x9c,0xc1]
5936 ; X86-NEXT:    retl # encoding: [0xc3]
5939 ; X64:       # %bb.0:
5940 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
5941 ; X64-NEXT:    vfnmadd132ps %ymm1, %ymm2, %ymm0 {%k1} # encoding: [0x62,0xf2,0x6d,0x29,0x9c,0xc1]
5943 ; X64-NEXT:    retq # encoding: [0xc3]
5953 ; CHECK:       # %bb.0:
5954 ; CHECK-NEXT:    vfnmadd213ps %xmm2, %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x71,0xac,0xc2]
5956 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
5964 ; X86:       # %bb.0:
5965 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
5966 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
5967 ; X86-NEXT:    vfnmadd132ps %xmm1, %xmm2, %xmm0 {%k1} # encoding: [0x62,0xf2,0x6d,0x09,0x9c,0xc1]
5969 ; X86-NEXT:    retl # encoding: [0xc3]
5972 ; X64:       # %bb.0:
5973 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
5974 ; X64-NEXT:    vfnmadd132ps %xmm1, %xmm2, %xmm0 {%k1} # encoding: [0x62,0xf2,0x6d,0x09,0x9c,0xc1]
5976 ; X64-NEXT:    retq # encoding: [0xc3]
5980   %extract = shufflevector <8 x i1> %3, <8 x i1> %3, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
5987 ; CHECK:       # %bb.0:
5988 ; CHECK-NEXT:    vfnmadd213pd %ymm2, %ymm1, %ymm0 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0xf5,0xac,0xc2]
5990 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
5998 ; X86:       # %bb.0:
5999 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
6000 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
6001 ; X86-NEXT:    vfnmadd132pd %ymm1, %ymm2, %ymm0 {%k1} # encoding: [0x62,0xf2,0xed,0x29,0x9c,0xc1]
6003 ; X86-NEXT:    retl # encoding: [0xc3]
6006 ; X64:       # %bb.0:
6007 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
6008 ; X64-NEXT:    vfnmadd132pd %ymm1, %ymm2, %ymm0 {%k1} # encoding: [0x62,0xf2,0xed,0x29,0x9c,0xc1]
6010 ; X64-NEXT:    retq # encoding: [0xc3]
6014   %extract = shufflevector <8 x i1> %3, <8 x i1> %3, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
6021 ; CHECK:       # %bb.0:
6022 ; CHECK-NEXT:    vfnmadd213pd %xmm2, %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0xf1,0xac,0xc2]
6024 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
6032 ; X86:       # %bb.0:
6033 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
6034 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
6035 ; X86-NEXT:    vfnmadd132pd %xmm1, %xmm2, %xmm0 {%k1} # encoding: [0x62,0xf2,0xed,0x09,0x9c,0xc1]
6037 ; X86-NEXT:    retl # encoding: [0xc3]
6040 ; X64:       # %bb.0:
6041 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
6042 ; X64-NEXT:    vfnmadd132pd %xmm1, %xmm2, %xmm0 {%k1} # encoding: [0x62,0xf2,0xed,0x09,0x9c,0xc1]
6044 ; X64-NEXT:    retq # encoding: [0xc3]
6048   %extract = shufflevector <8 x i1> %3, <8 x i1> %3, <2 x i32> <i32 0, i32 1>
6055 ; CHECK:       # %bb.0:
6056 ; CHECK-NEXT:    vfnmsub213ps %ymm2, %ymm1, %ymm0 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x75,0xae,0xc2]
6058 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
6067 ; X86:       # %bb.0:
6068 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
6069 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
6070 ; X86-NEXT:    vfnmsub132ps %ymm1, %ymm2, %ymm0 {%k1} # encoding: [0x62,0xf2,0x6d,0x29,0x9e,0xc1]
6072 ; X86-NEXT:    retl # encoding: [0xc3]
6075 ; X64:       # %bb.0:
6076 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
6077 ; X64-NEXT:    vfnmsub132ps %ymm1, %ymm2, %ymm0 {%k1} # encoding: [0x62,0xf2,0x6d,0x29,0x9e,0xc1]
6079 ; X64-NEXT:    retq # encoding: [0xc3]
6090 ; CHECK:       # %bb.0:
6091 ; CHECK-NEXT:    vfnmsub213ps %xmm2, %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x71,0xae,0xc2]
6093 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
6102 ; X86:       # %bb.0:
6103 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
6104 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
6105 ; X86-NEXT:    vfnmsub132ps %xmm1, %xmm2, %xmm0 {%k1} # encoding: [0x62,0xf2,0x6d,0x09,0x9e,0xc1]
6107 ; X86-NEXT:    retl # encoding: [0xc3]
6110 ; X64:       # %bb.0:
6111 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
6112 ; X64-NEXT:    vfnmsub132ps %xmm1, %xmm2, %xmm0 {%k1} # encoding: [0x62,0xf2,0x6d,0x09,0x9e,0xc1]
6114 ; X64-NEXT:    retq # encoding: [0xc3]
6119   %extract = shufflevector <8 x i1> %4, <8 x i1> %4, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
6126 ; CHECK:       # %bb.0:
6127 ; CHECK-NEXT:    vfnmsub213pd %ymm2, %ymm1, %ymm0 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0xf5,0xae,0xc2]
6129 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
6138 ; X86:       # %bb.0:
6139 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
6140 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
6141 ; X86-NEXT:    vfnmsub132pd %ymm1, %ymm2, %ymm0 {%k1} # encoding: [0x62,0xf2,0xed,0x29,0x9e,0xc1]
6143 ; X86-NEXT:    retl # encoding: [0xc3]
6146 ; X64:       # %bb.0:
6147 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
6148 ; X64-NEXT:    vfnmsub132pd %ymm1, %ymm2, %ymm0 {%k1} # encoding: [0x62,0xf2,0xed,0x29,0x9e,0xc1]
6150 ; X64-NEXT:    retq # encoding: [0xc3]
6155   %extract = shufflevector <8 x i1> %4, <8 x i1> %4, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
6162 ; CHECK:       # %bb.0:
6163 ; CHECK-NEXT:    vfnmsub213pd %xmm2, %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0xf1,0xae,0xc2]
6165 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
6174 ; X86:       # %bb.0:
6175 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
6176 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
6177 ; X86-NEXT:    vfnmsub132pd %xmm1, %xmm2, %xmm0 {%k1} # encoding: [0x62,0xf2,0xed,0x09,0x9e,0xc1]
6179 ; X86-NEXT:    retl # encoding: [0xc3]
6182 ; X64:       # %bb.0:
6183 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
6184 ; X64-NEXT:    vfnmsub132pd %xmm1, %xmm2, %xmm0 {%k1} # encoding: [0x62,0xf2,0xed,0x09,0x9e,0xc1]
6186 ; X64-NEXT:    retq # encoding: [0xc3]
6191   %extract = shufflevector <8 x i1> %4, <8 x i1> %4, <2 x i32> <i32 0, i32 1>
6198 ; X86:       # %bb.0:
6199 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
6200 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
6201 ; X86-NEXT:    vfnmsub231pd %xmm1, %xmm0, %xmm2 {%k1} # encoding: [0x62,0xf2,0xfd,0x09,0xbe,0xd1]
6203 ; X86-NEXT:    vmovapd %xmm2, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf9,0x28,0xc2]
6204 ; X86-NEXT:    retl # encoding: [0xc3]
6207 ; X64:       # %bb.0:
6208 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
6209 ; X64-NEXT:    vfnmsub231pd %xmm1, %xmm0, %xmm2 {%k1} # encoding: [0x62,0xf2,0xfd,0x09,0xbe,0xd1]
6211 ; X64-NEXT:    vmovapd %xmm2, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf9,0x28,0xc2]
6212 ; X64-NEXT:    retq # encoding: [0xc3]
6217   %extract = shufflevector <8 x i1> %4, <8 x i1> %4, <2 x i32> <i32 0, i32 1>
6224 ; X86:       # %bb.0:
6225 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
6226 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
6227 ; X86-NEXT:    vfnmsub231pd %ymm1, %ymm0, %ymm2 {%k1} # encoding: [0x62,0xf2,0xfd,0x29,0xbe,0xd1]
6229 ; X86-NEXT:    vmovapd %ymm2, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfd,0x28,0xc2]
6230 ; X86-NEXT:    retl # encoding: [0xc3]
6233 ; X64:       # %bb.0:
6234 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
6235 ; X64-NEXT:    vfnmsub231pd %ymm1, %ymm0, %ymm2 {%k1} # encoding: [0x62,0xf2,0xfd,0x29,0xbe,0xd1]
6237 ; X64-NEXT:    vmovapd %ymm2, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfd,0x28,0xc2]
6238 ; X64-NEXT:    retq # encoding: [0xc3]
6243   %extract = shufflevector <8 x i1> %4, <8 x i1> %4, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
6250 ; X86:       # %bb.0:
6251 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
6252 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
6253 ; X86-NEXT:    vfnmsub231ps %xmm1, %xmm0, %xmm2 {%k1} # encoding: [0x62,0xf2,0x7d,0x09,0xbe,0xd1]
6255 ; X86-NEXT:    vmovaps %xmm2, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x28,0xc2]
6256 ; X86-NEXT:    retl # encoding: [0xc3]
6259 ; X64:       # %bb.0:
6260 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
6261 ; X64-NEXT:    vfnmsub231ps %xmm1, %xmm0, %xmm2 {%k1} # encoding: [0x62,0xf2,0x7d,0x09,0xbe,0xd1]
6263 ; X64-NEXT:    vmovaps %xmm2, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x28,0xc2]
6264 ; X64-NEXT:    retq # encoding: [0xc3]
6269   %extract = shufflevector <8 x i1> %4, <8 x i1> %4, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
6276 ; X86:       # %bb.0:
6277 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
6278 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
6279 ; X86-NEXT:    vfnmsub231ps %ymm1, %ymm0, %ymm2 {%k1} # encoding: [0x62,0xf2,0x7d,0x29,0xbe,0xd1]
6281 ; X86-NEXT:    vmovaps %ymm2, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfc,0x28,0xc2]
6282 ; X86-NEXT:    retl # encoding: [0xc3]
6285 ; X64:       # %bb.0:
6286 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
6287 ; X64-NEXT:    vfnmsub231ps %ymm1, %ymm0, %ymm2 {%k1} # encoding: [0x62,0xf2,0x7d,0x29,0xbe,0xd1]
6289 ; X64-NEXT:    vmovaps %ymm2, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfc,0x28,0xc2]
6290 ; X64-NEXT:    retq # encoding: [0xc3]
6301 ; CHECK:       # %bb.0:
6302 ; CHECK-NEXT:    vfmaddsub213ps %ymm2, %ymm1, %ymm0 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x75,0xa6,0xc2]
6304 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
6308   %4 = shufflevector <8 x float> %3, <8 x float> %1, <8 x i32> <i32 0, i32 9, i32 2, i32 11, i32 4, i32 13, i32 6, i32 15>
6314 ; X86:       # %bb.0:
6315 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
6316 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
6317 ; X86-NEXT:    vfmaddsub132ps %ymm1, %ymm2, %ymm0 {%k1} # encoding: [0x62,0xf2,0x6d,0x29,0x96,0xc1]
6319 ; X86-NEXT:    retl # encoding: [0xc3]
6322 ; X64:       # %bb.0:
6323 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
6324 ; X64-NEXT:    vfmaddsub132ps %ymm1, %ymm2, %ymm0 {%k1} # encoding: [0x62,0xf2,0x6d,0x29,0x96,0xc1]
6326 ; X64-NEXT:    retq # encoding: [0xc3]
6330   %4 = shufflevector <8 x float> %3, <8 x float> %1, <8 x i32> <i32 0, i32 9, i32 2, i32 11, i32 4, i32 13, i32 6, i32 15>
6338 ; CHECK:       # %bb.0:
6339 ; CHECK-NEXT:    vfmaddsub213ps %xmm2, %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x71,0xa6,0xc2]
6341 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
6345   %4 = shufflevector <4 x float> %3, <4 x float> %1, <4 x i32> <i32 0, i32 5, i32 2, i32 7>
6351 ; X86:       # %bb.0:
6352 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
6353 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
6354 ; X86-NEXT:    vfmaddsub132ps %xmm1, %xmm2, %xmm0 {%k1} # encoding: [0x62,0xf2,0x6d,0x09,0x96,0xc1]
6356 ; X86-NEXT:    retl # encoding: [0xc3]
6359 ; X64:       # %bb.0:
6360 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
6361 ; X64-NEXT:    vfmaddsub132ps %xmm1, %xmm2, %xmm0 {%k1} # encoding: [0x62,0xf2,0x6d,0x09,0x96,0xc1]
6363 ; X64-NEXT:    retq # encoding: [0xc3]
6367   %4 = shufflevector <4 x float> %3, <4 x float> %1, <4 x i32> <i32 0, i32 5, i32 2, i32 7>
6369   %extract = shufflevector <8 x i1> %5, <8 x i1> %5, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
6376 ; CHECK:       # %bb.0:
6377 ; CHECK-NEXT:    vfmaddsub213pd %ymm2, %ymm1, %ymm0 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0xf5,0xa6,0xc2]
6379 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
6383   %4 = shufflevector <4 x double> %3, <4 x double> %1, <4 x i32> <i32 0, i32 5, i32 2, i32 7>
6389 ; X86:       # %bb.0:
6390 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
6391 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
6392 ; X86-NEXT:    vfmaddsub132pd %ymm1, %ymm2, %ymm0 {%k1} # encoding: [0x62,0xf2,0xed,0x29,0x96,0xc1]
6394 ; X86-NEXT:    retl # encoding: [0xc3]
6397 ; X64:       # %bb.0:
6398 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
6399 ; X64-NEXT:    vfmaddsub132pd %ymm1, %ymm2, %ymm0 {%k1} # encoding: [0x62,0xf2,0xed,0x29,0x96,0xc1]
6401 ; X64-NEXT:    retq # encoding: [0xc3]
6405   %4 = shufflevector <4 x double> %3, <4 x double> %1, <4 x i32> <i32 0, i32 5, i32 2, i32 7>
6407   %extract = shufflevector <8 x i1> %5, <8 x i1> %5, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
6414 ; CHECK:       # %bb.0:
6415 ; CHECK-NEXT:    vfmaddsub213pd %xmm2, %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0xf1,0xa6,0xc2]
6417 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
6421   %4 = shufflevector <2 x double> %3, <2 x double> %1, <2 x i32> <i32 0, i32 3>
6427 ; X86:       # %bb.0:
6428 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
6429 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
6430 ; X86-NEXT:    vfmaddsub132pd %xmm1, %xmm2, %xmm0 {%k1} # encoding: [0x62,0xf2,0xed,0x09,0x96,0xc1]
6432 ; X86-NEXT:    retl # encoding: [0xc3]
6435 ; X64:       # %bb.0:
6436 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
6437 ; X64-NEXT:    vfmaddsub132pd %xmm1, %xmm2, %xmm0 {%k1} # encoding: [0x62,0xf2,0xed,0x09,0x96,0xc1]
6439 ; X64-NEXT:    retq # encoding: [0xc3]
6443   %4 = shufflevector <2 x double> %3, <2 x double> %1, <2 x i32> <i32 0, i32 3>
6445   %extract = shufflevector <8 x i1> %5, <8 x i1> %5, <2 x i32> <i32 0, i32 1>
6452 ; X86:       # %bb.0:
6453 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
6454 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
6455 ; X86-NEXT:    vfmaddsub231pd %xmm1, %xmm0, %xmm2 {%k1} # encoding: [0x62,0xf2,0xfd,0x09,0xb6,0xd1]
6457 ; X86-NEXT:    vmovapd %xmm2, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf9,0x28,0xc2]
6458 ; X86-NEXT:    retl # encoding: [0xc3]
6461 ; X64:       # %bb.0:
6462 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
6463 ; X64-NEXT:    vfmaddsub231pd %xmm1, %xmm0, %xmm2 {%k1} # encoding: [0x62,0xf2,0xfd,0x09,0xb6,0xd1]
6465 ; X64-NEXT:    vmovapd %xmm2, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf9,0x28,0xc2]
6466 ; X64-NEXT:    retq # encoding: [0xc3]
6470   %4 = shufflevector <2 x double> %3, <2 x double> %1, <2 x i32> <i32 0, i32 3>
6472   %extract = shufflevector <8 x i1> %5, <8 x i1> %5, <2 x i32> <i32 0, i32 1>
6479 ; X86:       # %bb.0:
6480 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
6481 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
6482 ; X86-NEXT:    vfmaddsub213pd %xmm2, %xmm1, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0xf5,0x89,0xa6,0xc2]
6484 ; X86-NEXT:    retl # encoding: [0xc3]
6487 ; X64:       # %bb.0:
6488 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
6489 ; X64-NEXT:    vfmaddsub213pd %xmm2, %xmm1, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0xf5,0x89,0xa6,0xc2]
6491 ; X64-NEXT:    retq # encoding: [0xc3]
6495   %4 = shufflevector <2 x double> %3, <2 x double> %1, <2 x i32> <i32 0, i32 3>
6497   %extract = shufflevector <8 x i1> %5, <8 x i1> %5, <2 x i32> <i32 0, i32 1>
6504 ; X86:       # %bb.0:
6505 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
6506 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
6507 ; X86-NEXT:    vfmaddsub231pd %ymm1, %ymm0, %ymm2 {%k1} # encoding: [0x62,0xf2,0xfd,0x29,0xb6,0xd1]
6509 ; X86-NEXT:    vmovapd %ymm2, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfd,0x28,0xc2]
6510 ; X86-NEXT:    retl # encoding: [0xc3]
6513 ; X64:       # %bb.0:
6514 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
6515 ; X64-NEXT:    vfmaddsub231pd %ymm1, %ymm0, %ymm2 {%k1} # encoding: [0x62,0xf2,0xfd,0x29,0xb6,0xd1]
6517 ; X64-NEXT:    vmovapd %ymm2, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfd,0x28,0xc2]
6518 ; X64-NEXT:    retq # encoding: [0xc3]
6522   %4 = shufflevector <4 x double> %3, <4 x double> %1, <4 x i32> <i32 0, i32 5, i32 2, i32 7>
6524   %extract = shufflevector <8 x i1> %5, <8 x i1> %5, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
6531 ; X86:       # %bb.0:
6532 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
6533 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
6534 ; X86-NEXT:    vfmaddsub213pd %ymm2, %ymm1, %ymm0 {%k1} {z} # encoding: [0x62,0xf2,0xf5,0xa9,0xa6,0xc2]
6536 ; X86-NEXT:    retl # encoding: [0xc3]
6539 ; X64:       # %bb.0:
6540 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
6541 ; X64-NEXT:    vfmaddsub213pd %ymm2, %ymm1, %ymm0 {%k1} {z} # encoding: [0x62,0xf2,0xf5,0xa9,0xa6,0xc2]
6543 ; X64-NEXT:    retq # encoding: [0xc3]
6547   %4 = shufflevector <4 x double> %3, <4 x double> %1, <4 x i32> <i32 0, i32 5, i32 2, i32 7>
6549   %extract = shufflevector <8 x i1> %5, <8 x i1> %5, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
6556 ; X86:       # %bb.0:
6557 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
6558 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
6559 ; X86-NEXT:    vfmaddsub231ps %xmm1, %xmm0, %xmm2 {%k1} # encoding: [0x62,0xf2,0x7d,0x09,0xb6,0xd1]
6561 ; X86-NEXT:    vmovaps %xmm2, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x28,0xc2]
6562 ; X86-NEXT:    retl # encoding: [0xc3]
6565 ; X64:       # %bb.0:
6566 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
6567 ; X64-NEXT:    vfmaddsub231ps %xmm1, %xmm0, %xmm2 {%k1} # encoding: [0x62,0xf2,0x7d,0x09,0xb6,0xd1]
6569 ; X64-NEXT:    vmovaps %xmm2, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x28,0xc2]
6570 ; X64-NEXT:    retq # encoding: [0xc3]
6574   %4 = shufflevector <4 x float> %3, <4 x float> %1, <4 x i32> <i32 0, i32 5, i32 2, i32 7>
6576   %extract = shufflevector <8 x i1> %5, <8 x i1> %5, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
6583 ; X86:       # %bb.0:
6584 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
6585 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
6586 ; X86-NEXT:    vfmaddsub213ps %xmm2, %xmm1, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x75,0x89,0xa6,0xc2]
6588 ; X86-NEXT:    retl # encoding: [0xc3]
6591 ; X64:       # %bb.0:
6592 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
6593 ; X64-NEXT:    vfmaddsub213ps %xmm2, %xmm1, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x75,0x89,0xa6,0xc2]
6595 ; X64-NEXT:    retq # encoding: [0xc3]
6599   %4 = shufflevector <4 x float> %3, <4 x float> %1, <4 x i32> <i32 0, i32 5, i32 2, i32 7>
6601   %extract = shufflevector <8 x i1> %5, <8 x i1> %5, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
6608 ; X86:       # %bb.0:
6609 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
6610 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
6611 ; X86-NEXT:    vfmaddsub231ps %ymm1, %ymm0, %ymm2 {%k1} # encoding: [0x62,0xf2,0x7d,0x29,0xb6,0xd1]
6613 ; X86-NEXT:    vmovaps %ymm2, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfc,0x28,0xc2]
6614 ; X86-NEXT:    retl # encoding: [0xc3]
6617 ; X64:       # %bb.0:
6618 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
6619 ; X64-NEXT:    vfmaddsub231ps %ymm1, %ymm0, %ymm2 {%k1} # encoding: [0x62,0xf2,0x7d,0x29,0xb6,0xd1]
6621 ; X64-NEXT:    vmovaps %ymm2, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfc,0x28,0xc2]
6622 ; X64-NEXT:    retq # encoding: [0xc3]
6626   %4 = shufflevector <8 x float> %3, <8 x float> %1, <8 x i32> <i32 0, i32 9, i32 2, i32 11, i32 4, i32 13, i32 6, i32 15>
6634 ; X86:       # %bb.0:
6635 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
6636 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
6637 ; X86-NEXT:    vfmaddsub213ps %ymm2, %ymm1, %ymm0 {%k1} {z} # encoding: [0x62,0xf2,0x75,0xa9,0xa6,0xc2]
6639 ; X86-NEXT:    retl # encoding: [0xc3]
6642 ; X64:       # %bb.0:
6643 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
6644 ; X64-NEXT:    vfmaddsub213ps %ymm2, %ymm1, %ymm0 {%k1} {z} # encoding: [0x62,0xf2,0x75,0xa9,0xa6,0xc2]
6646 ; X64-NEXT:    retq # encoding: [0xc3]
6650   %4 = shufflevector <8 x float> %3, <8 x float> %1, <8 x i32> <i32 0, i32 9, i32 2, i32 11, i32 4, i32 13, i32 6, i32 15>
6658 ; X86:       # %bb.0:
6659 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
6660 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
6661 ; X86-NEXT:    vfmsubadd231pd %xmm1, %xmm0, %xmm2 {%k1} # encoding: [0x62,0xf2,0xfd,0x09,0xb7,0xd1]
6663 ; X86-NEXT:    vmovapd %xmm2, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf9,0x28,0xc2]
6664 ; X86-NEXT:    retl # encoding: [0xc3]
6667 ; X64:       # %bb.0:
6668 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
6669 ; X64-NEXT:    vfmsubadd231pd %xmm1, %xmm0, %xmm2 {%k1} # encoding: [0x62,0xf2,0xfd,0x09,0xb7,0xd1]
6671 ; X64-NEXT:    vmovapd %xmm2, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf9,0x28,0xc2]
6672 ; X64-NEXT:    retq # encoding: [0xc3]
6676   %4 = shufflevector <2 x double> %1, <2 x double> %3, <2 x i32> <i32 0, i32 3>
6678   %extract = shufflevector <8 x i1> %5, <8 x i1> %5, <2 x i32> <i32 0, i32 1>
6685 ; X86:       # %bb.0:
6686 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
6687 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
6688 ; X86-NEXT:    vfmsubadd231pd %ymm1, %ymm0, %ymm2 {%k1} # encoding: [0x62,0xf2,0xfd,0x29,0xb7,0xd1]
6690 ; X86-NEXT:    vmovapd %ymm2, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfd,0x28,0xc2]
6691 ; X86-NEXT:    retl # encoding: [0xc3]
6694 ; X64:       # %bb.0:
6695 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
6696 ; X64-NEXT:    vfmsubadd231pd %ymm1, %ymm0, %ymm2 {%k1} # encoding: [0x62,0xf2,0xfd,0x29,0xb7,0xd1]
6698 ; X64-NEXT:    vmovapd %ymm2, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfd,0x28,0xc2]
6699 ; X64-NEXT:    retq # encoding: [0xc3]
6703   %4 = shufflevector <4 x double> %1, <4 x double> %3, <4 x i32> <i32 0, i32 5, i32 2, i32 7>
6705   %extract = shufflevector <8 x i1> %5, <8 x i1> %5, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
6712 ; X86:       # %bb.0:
6713 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
6714 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
6715 ; X86-NEXT:    vfmsubadd231ps %xmm1, %xmm0, %xmm2 {%k1} # encoding: [0x62,0xf2,0x7d,0x09,0xb7,0xd1]
6717 ; X86-NEXT:    vmovaps %xmm2, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x28,0xc2]
6718 ; X86-NEXT:    retl # encoding: [0xc3]
6721 ; X64:       # %bb.0:
6722 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
6723 ; X64-NEXT:    vfmsubadd231ps %xmm1, %xmm0, %xmm2 {%k1} # encoding: [0x62,0xf2,0x7d,0x09,0xb7,0xd1]
6725 ; X64-NEXT:    vmovaps %xmm2, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x28,0xc2]
6726 ; X64-NEXT:    retq # encoding: [0xc3]
6730   %4 = shufflevector <4 x float> %1, <4 x float> %3, <4 x i32> <i32 0, i32 5, i32 2, i32 7>
6732   %extract = shufflevector <8 x i1> %5, <8 x i1> %5, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
6739 ; X86:       # %bb.0:
6740 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
6741 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
6742 ; X86-NEXT:    vfmsubadd231ps %ymm1, %ymm0, %ymm2 {%k1} # encoding: [0x62,0xf2,0x7d,0x29,0xb7,0xd1]
6744 ; X86-NEXT:    vmovaps %ymm2, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfc,0x28,0xc2]
6745 ; X86-NEXT:    retl # encoding: [0xc3]
6748 ; X64:       # %bb.0:
6749 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
6750 ; X64-NEXT:    vfmsubadd231ps %ymm1, %ymm0, %ymm2 {%k1} # encoding: [0x62,0xf2,0x7d,0x29,0xb7,0xd1]
6752 ; X64-NEXT:    vmovaps %ymm2, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfc,0x28,0xc2]
6753 ; X64-NEXT:    retq # encoding: [0xc3]
6757   %4 = shufflevector <8 x float> %1, <8 x float> %3, <8 x i32> <i32 0, i32 9, i32 2, i32 11, i32 4, i32 13, i32 6, i32 15>
6765 ; X86:       # %bb.0:
6766 ; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x04]
6767 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %ecx # encoding: [0x0f,0xb6,0x4c,0x24,0x08]
6768 ; X86-NEXT:    kmovw %ecx, %k1 # encoding: [0xc5,0xf8,0x92,0xc9]
6769 ; X86-NEXT:    vfmadd213ps (%eax), %xmm1, %xmm0 {%k1} # encoding: [0x62,0xf2,0x75,0x09,0xa8,0x00]
6771 ; X86-NEXT:    retl # encoding: [0xc3]
6774 ; X64:       # %bb.0:
6775 ; X64-NEXT:    kmovw %esi, %k1 # encoding: [0xc5,0xf8,0x92,0xce]
6776 ; X64-NEXT:    vfmadd213ps (%rdi), %xmm1, %xmm0 {%k1} # encoding: [0x62,0xf2,0x75,0x09,0xa8,0x07]
6778 ; X64-NEXT:    retq # encoding: [0xc3]
6782   %extract = shufflevector <8 x i1> %2, <8 x i1> %2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
6789 ; X86:       # %bb.0:
6790 ; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x04]
6791 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %ecx # encoding: [0x0f,0xb6,0x4c,0x24,0x08]
6792 ; X86-NEXT:    kmovw %ecx, %k1 # encoding: [0xc5,0xf8,0x92,0xc9]
6793 ; X86-NEXT:    vfmadd213ps (%eax), %xmm1, %xmm0 {%k1} # encoding: [0x62,0xf2,0x75,0x09,0xa8,0x00]
6795 ; X86-NEXT:    retl # encoding: [0xc3]
6798 ; X64:       # %bb.0:
6799 ; X64-NEXT:    kmovw %esi, %k1 # encoding: [0xc5,0xf8,0x92,0xce]
6800 ; X64-NEXT:    vfmadd213ps (%rdi), %xmm1, %xmm0 {%k1} # encoding: [0x62,0xf2,0x75,0x09,0xa8,0x07]
6802 ; X64-NEXT:    retq # encoding: [0xc3]
6806   %extract = shufflevector <8 x i1> %2, <8 x i1> %2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
6813 ; X86:       # %bb.0:
6814 ; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x04]
6815 ; X86-NEXT:    vfmadd213ps (%eax), %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x71,0xa8,0x00]
6817 ; X86-NEXT:    retl # encoding: [0xc3]
6820 ; X64:       # %bb.0:
6821 ; X64-NEXT:    vfmadd213ps (%rdi), %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x71,0xa8,0x07]
6823 ; X64-NEXT:    retq # encoding: [0xc3]
6831 ; X86:       # %bb.0:
6832 ; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x04]
6833 ; X86-NEXT:    vfmadd213ps (%eax), %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x71,0xa8,0x00]
6835 ; X86-NEXT:    retl # encoding: [0xc3]
6838 ; X64:       # %bb.0:
6839 ; X64-NEXT:    vfmadd213ps (%rdi), %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x71,0xa8,0x07]
6841 ; X64-NEXT:    retq # encoding: [0xc3]
6849 ; X86:       # %bb.0:
6850 ; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x04]
6851 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %ecx # encoding: [0x0f,0xb6,0x4c,0x24,0x08]
6852 ; X86-NEXT:    kmovw %ecx, %k1 # encoding: [0xc5,0xf8,0x92,0xc9]
6853 ; X86-NEXT:    vfmadd213ps (%eax){1to4}, %xmm1, %xmm0 {%k1} # encoding: [0x62,0xf2,0x75,0x19,0xa8,0x00]
6855 ; X86-NEXT:    retl # encoding: [0xc3]
6858 ; X64:       # %bb.0:
6859 ; X64-NEXT:    kmovw %esi, %k1 # encoding: [0xc5,0xf8,0x92,0xce]
6860 ; X64-NEXT:    vfmadd213ps (%rdi){1to4}, %xmm1, %xmm0 {%k1} # encoding: [0x62,0xf2,0x75,0x19,0xa8,0x07]
6862 ; X64-NEXT:    retq # encoding: [0xc3]
6864   %vecinit.i = insertelement <4 x float> undef, float %q, i32 0
6870   %extract = shufflevector <8 x i1> %2, <8 x i1> %2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
6877 ; X86:       # %bb.0:
6878 ; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x04]
6879 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %ecx # encoding: [0x0f,0xb6,0x4c,0x24,0x08]
6880 ; X86-NEXT:    kmovw %ecx, %k1 # encoding: [0xc5,0xf8,0x92,0xc9]
6881 ; X86-NEXT:    vfmadd213ps (%eax){1to4}, %xmm1, %xmm0 {%k1} # encoding: [0x62,0xf2,0x75,0x19,0xa8,0x00]
6883 ; X86-NEXT:    retl # encoding: [0xc3]
6886 ; X64:       # %bb.0:
6887 ; X64-NEXT:    kmovw %esi, %k1 # encoding: [0xc5,0xf8,0x92,0xce]
6888 ; X64-NEXT:    vfmadd213ps (%rdi){1to4}, %xmm1, %xmm0 {%k1} # encoding: [0x62,0xf2,0x75,0x19,0xa8,0x07]
6890 ; X64-NEXT:    retq # encoding: [0xc3]
6892   %vecinit.i = insertelement <4 x float> undef, float %q, i32 0
6898   %extract = shufflevector <8 x i1> %2, <8 x i1> %2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
6905 ; X86:       # %bb.0:
6906 ; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x04]
6907 ; X86-NEXT:    vfmadd213ps (%eax){1to4}, %xmm1, %xmm0 # encoding: [0x62,0xf2,0x75,0x18,0xa8,0x00]
6909 ; X86-NEXT:    retl # encoding: [0xc3]
6912 ; X64:       # %bb.0:
6913 ; X64-NEXT:    vfmadd213ps (%rdi){1to4}, %xmm1, %xmm0 # encoding: [0x62,0xf2,0x75,0x18,0xa8,0x07]
6915 ; X64-NEXT:    retq # encoding: [0xc3]
6917   %vecinit.i = insertelement <4 x float> undef, float %q, i32 0
6927 ; X86:       # %bb.0:
6928 ; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x04]
6929 ; X86-NEXT:    vfmadd213ps (%eax){1to4}, %xmm1, %xmm0 # encoding: [0x62,0xf2,0x75,0x18,0xa8,0x00]
6931 ; X86-NEXT:    retl # encoding: [0xc3]
6934 ; X64:       # %bb.0:
6935 ; X64-NEXT:    vfmadd213ps (%rdi){1to4}, %xmm1, %xmm0 # encoding: [0x62,0xf2,0x75,0x18,0xa8,0x07]
6937 ; X64-NEXT:    retq # encoding: [0xc3]
6939   %vecinit.i = insertelement <4 x float> undef, float %q, i32 0
6949 ; X86:       # %bb.0:
6950 ; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x04]
6951 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %ecx # encoding: [0x0f,0xb6,0x4c,0x24,0x08]
6952 ; X86-NEXT:    kmovw %ecx, %k1 # encoding: [0xc5,0xf8,0x92,0xc9]
6953 ; X86-NEXT:    vfmadd213pd (%eax), %xmm1, %xmm0 {%k1} # encoding: [0x62,0xf2,0xf5,0x09,0xa8,0x00]
6955 ; X86-NEXT:    retl # encoding: [0xc3]
6958 ; X64:       # %bb.0:
6959 ; X64-NEXT:    kmovw %esi, %k1 # encoding: [0xc5,0xf8,0x92,0xce]
6960 ; X64-NEXT:    vfmadd213pd (%rdi), %xmm1, %xmm0 {%k1} # encoding: [0x62,0xf2,0xf5,0x09,0xa8,0x07]
6962 ; X64-NEXT:    retq # encoding: [0xc3]
6966   %extract = shufflevector <8 x i1> %2, <8 x i1> %2, <2 x i32> <i32 0, i32 1>
6973 ; X86:       # %bb.0:
6974 ; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x04]
6975 ; X86-NEXT:    vfmadd213pd (%eax), %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0xf1,0xa8,0x00]
6977 ; X86-NEXT:    retl # encoding: [0xc3]
6980 ; X64:       # %bb.0:
6981 ; X64-NEXT:    vfmadd213pd (%rdi), %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0xf1,0xa8,0x07]
6983 ; X64-NEXT:    retq # encoding: [0xc3]
6991 ; X86:       # %bb.0:
6992 ; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x04]
6993 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %ecx # encoding: [0x0f,0xb6,0x4c,0x24,0x08]
6994 ; X86-NEXT:    kmovw %ecx, %k1 # encoding: [0xc5,0xf8,0x92,0xc9]
6995 ; X86-NEXT:    vfmadd213pd (%eax), %ymm1, %ymm0 {%k1} # encoding: [0x62,0xf2,0xf5,0x29,0xa8,0x00]
6997 ; X86-NEXT:    retl # encoding: [0xc3]
7000 ; X64:       # %bb.0:
7001 ; X64-NEXT:    kmovw %esi, %k1 # encoding: [0xc5,0xf8,0x92,0xce]
7002 ; X64-NEXT:    vfmadd213pd (%rdi), %ymm1, %ymm0 {%k1} # encoding: [0x62,0xf2,0xf5,0x29,0xa8,0x07]
7004 ; X64-NEXT:    retq # encoding: [0xc3]
7008   %extract = shufflevector <8 x i1> %2, <8 x i1> %2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
7015 ; X86:       # %bb.0:
7016 ; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x04]
7017 ; X86-NEXT:    vfmadd213pd (%eax), %ymm1, %ymm0 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0xf5,0xa8,0x00]
7019 ; X86-NEXT:    retl # encoding: [0xc3]
7022 ; X64:       # %bb.0:
7023 ; X64-NEXT:    vfmadd213pd (%rdi), %ymm1, %ymm0 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0xf5,0xa8,0x07]
7025 ; X64-NEXT:    retq # encoding: [0xc3]
7033 ; X86:       # %bb.0:
7034 ; X86-NEXT:    vpmovsxbd {{.*#+}} ymm1 = [14,13,6,3,5,15,0,1]
7035 ; X86-NEXT:    # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x7d,0x21,0x0d,A,A,A,A]
7036 ; X86-NEXT:    # fixup A - offset: 5, value: {{\.?LCPI[0-9]+_[0-9]+}}, kind: FK_Data_4
7037 ; X86-NEXT:    vpermps %zmm0, %zmm1, %zmm0 # encoding: [0x62,0xf2,0x75,0x48,0x16,0xc0]
7039 ; X86-NEXT:    retl # encoding: [0xc3]
7042 ; X64:       # %bb.0:
7043 ; X64-NEXT:    vpmovsxbd {{.*#+}} ymm1 = [14,13,6,3,5,15,0,1]
7044 ; X64-NEXT:    # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x7d,0x21,0x0d,A,A,A,A]
7045 ; X64-NEXT:    # fixup A - offset: 5, value: {{\.?LCPI[0-9]+_[0-9]+}}-4, kind: reloc_riprel_4byte
7046 ; X64-NEXT:    vpermps %zmm0, %zmm1, %zmm0 # encoding: [0x62,0xf2,0x75,0x48,0x16,0xc0]
7048 ; X64-NEXT:    retq # encoding: [0xc3]
7049   %1 = shufflevector <16 x i32> %a, <16 x i32> poison, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
7051   %3 = tail call <8 x i32> @llvm.x86.avx512.vpermi2var.d.256(<8 x i32> %1, <8 x i32> <i32 14, i32 13, i32 6, i32 3, i32 5, i32 15, i32 0, i32 1>, <8 x i32> %2)