• Home
  • Raw
  • Download

Lines Matching refs:AVX512BW

5 …rwin -mcpu=knl -mattr=-avx512cd,+avx512bw | FileCheck %s --check-prefix=ALL --check-prefix=AVX512BW
52 ; AVX512BW-LABEL: testv8i64:
53 ; AVX512BW: ## BB#0:
54 ; AVX512BW-NEXT: vpxord %zmm1, %zmm1, %zmm1
55 ; AVX512BW-NEXT: vpsubq %zmm0, %zmm1, %zmm2
56 ; AVX512BW-NEXT: vpandq %zmm2, %zmm0, %zmm0
57 ; AVX512BW-NEXT: vpsubq {{.*}}(%rip){1to8}, %zmm0, %zmm0
58 ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm2 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,1…
59 ; AVX512BW-NEXT: vpandq %zmm2, %zmm0, %zmm3
60 ; AVX512BW-NEXT: vmovdqu8 {{.*#+}} zmm4 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2…
61 ; AVX512BW-NEXT: vpshufb %zmm3, %zmm4, %zmm3
62 ; AVX512BW-NEXT: vpsrlw $4, %zmm0, %zmm0
63 ; AVX512BW-NEXT: vpandq %zmm2, %zmm0, %zmm0
64 ; AVX512BW-NEXT: vpshufb %zmm0, %zmm4, %zmm0
65 ; AVX512BW-NEXT: vpaddb %zmm3, %zmm0, %zmm0
66 ; AVX512BW-NEXT: vpsadbw %zmm1, %zmm0, %zmm0
67 ; AVX512BW-NEXT: retq
93 ; AVX512BW-LABEL: testv8i64u:
94 ; AVX512BW: ## BB#0:
95 ; AVX512BW-NEXT: vpxord %zmm1, %zmm1, %zmm1
96 ; AVX512BW-NEXT: vpsubq %zmm0, %zmm1, %zmm2
97 ; AVX512BW-NEXT: vpandq %zmm2, %zmm0, %zmm0
98 ; AVX512BW-NEXT: vpsubq {{.*}}(%rip){1to8}, %zmm0, %zmm0
99 ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm2 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,1…
100 ; AVX512BW-NEXT: vpandq %zmm2, %zmm0, %zmm3
101 ; AVX512BW-NEXT: vmovdqu8 {{.*#+}} zmm4 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2…
102 ; AVX512BW-NEXT: vpshufb %zmm3, %zmm4, %zmm3
103 ; AVX512BW-NEXT: vpsrlw $4, %zmm0, %zmm0
104 ; AVX512BW-NEXT: vpandq %zmm2, %zmm0, %zmm0
105 ; AVX512BW-NEXT: vpshufb %zmm0, %zmm4, %zmm0
106 ; AVX512BW-NEXT: vpaddb %zmm3, %zmm0, %zmm0
107 ; AVX512BW-NEXT: vpsadbw %zmm1, %zmm0, %zmm0
108 ; AVX512BW-NEXT: retq
170 ; AVX512BW-LABEL: testv16i32:
171 ; AVX512BW: ## BB#0:
172 ; AVX512BW-NEXT: vpxord %zmm1, %zmm1, %zmm1
173 ; AVX512BW-NEXT: vpsubd %zmm0, %zmm1, %zmm2
174 ; AVX512BW-NEXT: vpandd %zmm2, %zmm0, %zmm0
175 ; AVX512BW-NEXT: vpsubd {{.*}}(%rip){1to16}, %zmm0, %zmm0
176 ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm2 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,1…
177 ; AVX512BW-NEXT: vpandq %zmm2, %zmm0, %zmm3
178 ; AVX512BW-NEXT: vmovdqu8 {{.*#+}} zmm4 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2…
179 ; AVX512BW-NEXT: vpshufb %zmm3, %zmm4, %zmm3
180 ; AVX512BW-NEXT: vpsrlw $4, %zmm0, %zmm0
181 ; AVX512BW-NEXT: vpandq %zmm2, %zmm0, %zmm0
182 ; AVX512BW-NEXT: vpshufb %zmm0, %zmm4, %zmm0
183 ; AVX512BW-NEXT: vpaddb %zmm3, %zmm0, %zmm0
184 ; AVX512BW-NEXT: vpunpckhdq {{.*#+}} zmm2 = zmm0[2],zmm1[2],zmm0[3],zmm1[3],zmm0[6],zmm1[6],zmm0…
185 ; AVX512BW-NEXT: vpsadbw %zmm1, %zmm2, %zmm2
186 ; AVX512BW-NEXT: vpunpckldq {{.*#+}} zmm0 = zmm0[0],zmm1[0],zmm0[1],zmm1[1],zmm0[4],zmm1[4],zmm0…
187 ; AVX512BW-NEXT: vpsadbw %zmm1, %zmm0, %zmm0
188 ; AVX512BW-NEXT: vpackuswb %zmm2, %zmm0, %zmm0
189 ; AVX512BW-NEXT: retq
215 ; AVX512BW-LABEL: testv16i32u:
216 ; AVX512BW: ## BB#0:
217 ; AVX512BW-NEXT: vpxord %zmm1, %zmm1, %zmm1
218 ; AVX512BW-NEXT: vpsubd %zmm0, %zmm1, %zmm2
219 ; AVX512BW-NEXT: vpandd %zmm2, %zmm0, %zmm0
220 ; AVX512BW-NEXT: vpsubd {{.*}}(%rip){1to16}, %zmm0, %zmm0
221 ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm2 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,1…
222 ; AVX512BW-NEXT: vpandq %zmm2, %zmm0, %zmm3
223 ; AVX512BW-NEXT: vmovdqu8 {{.*#+}} zmm4 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2…
224 ; AVX512BW-NEXT: vpshufb %zmm3, %zmm4, %zmm3
225 ; AVX512BW-NEXT: vpsrlw $4, %zmm0, %zmm0
226 ; AVX512BW-NEXT: vpandq %zmm2, %zmm0, %zmm0
227 ; AVX512BW-NEXT: vpshufb %zmm0, %zmm4, %zmm0
228 ; AVX512BW-NEXT: vpaddb %zmm3, %zmm0, %zmm0
229 ; AVX512BW-NEXT: vpunpckhdq {{.*#+}} zmm2 = zmm0[2],zmm1[2],zmm0[3],zmm1[3],zmm0[6],zmm1[6],zmm0…
230 ; AVX512BW-NEXT: vpsadbw %zmm1, %zmm2, %zmm2
231 ; AVX512BW-NEXT: vpunpckldq {{.*#+}} zmm0 = zmm0[0],zmm1[0],zmm0[1],zmm1[1],zmm0[4],zmm1[4],zmm0…
232 ; AVX512BW-NEXT: vpsadbw %zmm1, %zmm0, %zmm0
233 ; AVX512BW-NEXT: vpackuswb %zmm2, %zmm0, %zmm0
234 ; AVX512BW-NEXT: retq
291 ; AVX512BW-LABEL: testv32i16:
292 ; AVX512BW: ## BB#0:
293 ; AVX512BW-NEXT: vpxord %zmm1, %zmm1, %zmm1
294 ; AVX512BW-NEXT: vpsubw %zmm0, %zmm1, %zmm1
295 ; AVX512BW-NEXT: vpandq %zmm1, %zmm0, %zmm0
296 ; AVX512BW-NEXT: vpsubw {{.*}}(%rip), %zmm0, %zmm0
297 ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,1…
298 ; AVX512BW-NEXT: vpandq %zmm1, %zmm0, %zmm2
299 ; AVX512BW-NEXT: vmovdqu8 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2…
300 ; AVX512BW-NEXT: vpshufb %zmm2, %zmm3, %zmm2
301 ; AVX512BW-NEXT: vpsrlw $4, %zmm0, %zmm0
302 ; AVX512BW-NEXT: vpandq %zmm1, %zmm0, %zmm0
303 ; AVX512BW-NEXT: vpshufb %zmm0, %zmm3, %zmm0
304 ; AVX512BW-NEXT: vpaddb %zmm2, %zmm0, %zmm0
305 ; AVX512BW-NEXT: vpsllw $8, %zmm0, %zmm1
306 ; AVX512BW-NEXT: vpaddb %zmm0, %zmm1, %zmm0
307 ; AVX512BW-NEXT: vpsrlw $8, %zmm0, %zmm0
308 ; AVX512BW-NEXT: retq
365 ; AVX512BW-LABEL: testv32i16u:
366 ; AVX512BW: ## BB#0:
367 ; AVX512BW-NEXT: vpxord %zmm1, %zmm1, %zmm1
368 ; AVX512BW-NEXT: vpsubw %zmm0, %zmm1, %zmm1
369 ; AVX512BW-NEXT: vpandq %zmm1, %zmm0, %zmm0
370 ; AVX512BW-NEXT: vpsubw {{.*}}(%rip), %zmm0, %zmm0
371 ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,1…
372 ; AVX512BW-NEXT: vpandq %zmm1, %zmm0, %zmm2
373 ; AVX512BW-NEXT: vmovdqu8 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2…
374 ; AVX512BW-NEXT: vpshufb %zmm2, %zmm3, %zmm2
375 ; AVX512BW-NEXT: vpsrlw $4, %zmm0, %zmm0
376 ; AVX512BW-NEXT: vpandq %zmm1, %zmm0, %zmm0
377 ; AVX512BW-NEXT: vpshufb %zmm0, %zmm3, %zmm0
378 ; AVX512BW-NEXT: vpaddb %zmm2, %zmm0, %zmm0
379 ; AVX512BW-NEXT: vpsllw $8, %zmm0, %zmm1
380 ; AVX512BW-NEXT: vpaddb %zmm0, %zmm1, %zmm0
381 ; AVX512BW-NEXT: vpsrlw $8, %zmm0, %zmm0
382 ; AVX512BW-NEXT: retq
430 ; AVX512BW-LABEL: testv64i8:
431 ; AVX512BW: ## BB#0:
432 ; AVX512BW-NEXT: vpxord %zmm1, %zmm1, %zmm1
433 ; AVX512BW-NEXT: vpsubb %zmm0, %zmm1, %zmm1
434 ; AVX512BW-NEXT: vpandq %zmm1, %zmm0, %zmm0
435 ; AVX512BW-NEXT: vpsubb {{.*}}(%rip), %zmm0, %zmm0
436 ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,1…
437 ; AVX512BW-NEXT: vpandq %zmm1, %zmm0, %zmm2
438 ; AVX512BW-NEXT: vmovdqu8 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2…
439 ; AVX512BW-NEXT: vpshufb %zmm2, %zmm3, %zmm2
440 ; AVX512BW-NEXT: vpsrlw $4, %zmm0, %zmm0
441 ; AVX512BW-NEXT: vpandq %zmm1, %zmm0, %zmm0
442 ; AVX512BW-NEXT: vpshufb %zmm0, %zmm3, %zmm0
443 ; AVX512BW-NEXT: vpaddb %zmm2, %zmm0, %zmm0
444 ; AVX512BW-NEXT: retq
492 ; AVX512BW-LABEL: testv64i8u:
493 ; AVX512BW: ## BB#0:
494 ; AVX512BW-NEXT: vpxord %zmm1, %zmm1, %zmm1
495 ; AVX512BW-NEXT: vpsubb %zmm0, %zmm1, %zmm1
496 ; AVX512BW-NEXT: vpandq %zmm1, %zmm0, %zmm0
497 ; AVX512BW-NEXT: vpsubb {{.*}}(%rip), %zmm0, %zmm0
498 ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,1…
499 ; AVX512BW-NEXT: vpandq %zmm1, %zmm0, %zmm2
500 ; AVX512BW-NEXT: vmovdqu8 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2…
501 ; AVX512BW-NEXT: vpshufb %zmm2, %zmm3, %zmm2
502 ; AVX512BW-NEXT: vpsrlw $4, %zmm0, %zmm0
503 ; AVX512BW-NEXT: vpandq %zmm1, %zmm0, %zmm0
504 ; AVX512BW-NEXT: vpshufb %zmm0, %zmm3, %zmm0
505 ; AVX512BW-NEXT: vpaddb %zmm2, %zmm0, %zmm0
506 ; AVX512BW-NEXT: retq