Lines Matching refs:xmm0
23 … CHECK32-NEXT: pmovzxbd {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],z…
24 ; CHECK32-NEXT: pmaddwd {{\.LCPI.*}}, %xmm0
29 … CHECK64-NEXT: pmovzxbd {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],z…
30 ; CHECK64-NEXT: pmaddwd {{.*}}(%rip), %xmm0
35 … SSE4-32-NEXT: pmovzxbd {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],z…
36 ; SSE4-32-NEXT: pmaddwd {{\.LCPI.*}}, %xmm0
41 … SSE4-64-NEXT: pmovzxbd {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],z…
42 ; SSE4-64-NEXT: pmaddwd {{.*}}(%rip), %xmm0
47 …AVX2-32-NEXT: vpmovzxbd {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],z…
48 ; AVX2-32-NEXT: vpmaddwd {{\.LCPI.*}}, %xmm0, %xmm0
53 …AVX2-64-NEXT: vpmovzxbd {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],z…
54 ; AVX2-64-NEXT: vpmaddwd {{.*}}(%rip), %xmm0, %xmm0
59 …12DQ-32-NEXT: vpmovzxbd {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],z…
60 ; AVX512DQ-32-NEXT: vpmaddwd {{\.LCPI.*}}, %xmm0, %xmm0
65 …12DQ-64-NEXT: vpmovzxbd {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],z…
66 ; AVX512DQ-64-NEXT: vpmaddwd {{.*}}(%rip), %xmm0, %xmm0
71 …12BW-32-NEXT: vpmovzxbd {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],z…
72 ; AVX512BW-32-NEXT: vpmaddwd {{\.LCPI.*}}, %xmm0, %xmm0
77 …12BW-64-NEXT: vpmovzxbd {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],z…
78 ; AVX512BW-64-NEXT: vpmaddwd {{.*}}(%rip), %xmm0, %xmm0
83 … KNL-32-NEXT: vpmovzxbd {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],z…
85 ; KNL-32-NEXT: vpmulld %xmm1, %xmm0, %xmm0
90 … KNL-64-NEXT: vpmovzxbd {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],z…
92 ; KNL-64-NEXT: vpmulld %xmm1, %xmm0, %xmm0
102 … pmovzxbw {{.*#+}} xmm1 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0…
103 ; SLM32-NEXT: movdqa {{.*#+}} xmm0 = [18778,18778,18778,18778,18778,18778,18778,18778]
105 ; SLM32-NEXT: pmullw %xmm0, %xmm1
106 ; SLM32-NEXT: pmulhw %xmm0, %xmm2
107 ; SLM32-NEXT: movdqa %xmm1, %xmm0
109 ; SLM32-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],…
114 … pmovzxbw {{.*#+}} xmm1 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0…
115 ; SLM64-NEXT: movdqa {{.*#+}} xmm0 = [18778,18778,18778,18778,18778,18778,18778,18778]
117 ; SLM64-NEXT: pmullw %xmm0, %xmm1
118 ; SLM64-NEXT: pmulhw %xmm0, %xmm2
119 ; SLM64-NEXT: movdqa %xmm1, %xmm0
121 ; SLM64-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],…
126 … pmovzxbw {{.*#+}} xmm1 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0…
127 ; SLOW32-NEXT: movdqa {{.*#+}} xmm0 = [18778,18778,18778,18778,18778,18778,18778,18778]
129 ; SLOW32-NEXT: pmulhw %xmm0, %xmm2
130 ; SLOW32-NEXT: pmullw %xmm0, %xmm1
131 ; SLOW32-NEXT: movdqa %xmm1, %xmm0
132 ; SLOW32-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3]…
138 … pmovzxbw {{.*#+}} xmm1 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0…
139 ; SLOW64-NEXT: movdqa {{.*#+}} xmm0 = [18778,18778,18778,18778,18778,18778,18778,18778]
141 ; SLOW64-NEXT: pmulhw %xmm0, %xmm2
142 ; SLOW64-NEXT: pmullw %xmm0, %xmm1
143 ; SLOW64-NEXT: movdqa %xmm1, %xmm0
144 ; SLOW64-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3]…
150 ; SSE4-32-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
152 … SSE4-32-NEXT: pmovzxbd {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],z…
154 ; SSE4-32-NEXT: pmaddwd %xmm2, %xmm0
160 ; SSE4-64-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
162 … SSE4-64-NEXT: pmovzxbd {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],z…
164 ; SSE4-64-NEXT: pmaddwd %xmm2, %xmm0
170 …xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero,xmm0[4…
176 …xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero,xmm0[4…
182 …xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero,xmm0[4…
188 …xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero,xmm0[4…
194 …xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero,xmm0[4…
200 …xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero,xmm0[4…
206 …xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero,xmm0[4…
213 …xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero,xmm0[4…
225 … pmovzxbw {{.*#+}} xmm1 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0…
226 ; SLM32-NEXT: movdqa %xmm0, %xmm3
227 ; SLM32-NEXT: movdqa {{.*#+}} xmm0 = [18778,18778,18778,18778,18778,18778,18778,18778]
232 ; SLM32-NEXT: pmullw %xmm0, %xmm1
233 ; SLM32-NEXT: pmulhw %xmm0, %xmm2
234 ; SLM32-NEXT: pmullw %xmm0, %xmm3
235 ; SLM32-NEXT: pmulhw %xmm0, %xmm4
236 ; SLM32-NEXT: movdqa %xmm1, %xmm0
238 ; SLM32-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],…
246 … pmovzxbw {{.*#+}} xmm1 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0…
247 ; SLM64-NEXT: movdqa %xmm0, %xmm3
248 ; SLM64-NEXT: movdqa {{.*#+}} xmm0 = [18778,18778,18778,18778,18778,18778,18778,18778]
253 ; SLM64-NEXT: pmullw %xmm0, %xmm1
254 ; SLM64-NEXT: pmulhw %xmm0, %xmm2
255 ; SLM64-NEXT: pmullw %xmm0, %xmm3
256 ; SLM64-NEXT: pmulhw %xmm0, %xmm4
257 ; SLM64-NEXT: movdqa %xmm1, %xmm0
259 ; SLM64-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],…
267 ; SLOW32-NEXT: movdqa %xmm0, %xmm3
268 … pmovzxbw {{.*#+}} xmm1 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0…
273 ; SLOW32-NEXT: movdqa %xmm1, %xmm0
274 ; SLOW32-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm4[0],xmm0[1],xmm4[1],xmm0[2],xmm4[2],xmm0[3]…
288 ; SLOW64-NEXT: movdqa %xmm0, %xmm3
289 … pmovzxbw {{.*#+}} xmm1 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0…
294 ; SLOW64-NEXT: movdqa %xmm1, %xmm0
295 ; SLOW64-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm4[0],xmm0[1],xmm4[1],xmm0[2],xmm4[2],xmm0[3]…
309 ; SSE4-32-NEXT: pshufd {{.*#+}} xmm1 = xmm0[3,3,3,3]
311 ; SSE4-32-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
313 ; SSE4-32-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
315 … SSE4-32-NEXT: pmovzxbd {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],z…
317 ; SSE4-32-NEXT: pmaddwd %xmm4, %xmm0
325 ; SSE4-64-NEXT: pshufd {{.*#+}} xmm1 = xmm0[3,3,3,3]
327 ; SSE4-64-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
329 ; SSE4-64-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
331 … SSE4-64-NEXT: pmovzxbd {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],z…
333 ; SSE4-64-NEXT: pmaddwd %xmm4, %xmm0
341 ; AVX2-32-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
343 …xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero,xmm0[4…
351 ; AVX2-64-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
353 …xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero,xmm0[4…
361 …xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero,xmm0[4…
367 …xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero,xmm0[4…
373 …xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero,xmm0[4…
379 …xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero,xmm0[4…
385 …xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero,xmm0[4…
391 …xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero,xmm0[4…
403 ; CHECK32-NEXT: movdqa %xmm0, %xmm2
405 ; CHECK32-NEXT: pmullw %xmm1, %xmm0
406 ; CHECK32-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3…
412 ; CHECK64-NEXT: movdqa %xmm0, %xmm2
414 ; CHECK64-NEXT: pmullw %xmm1, %xmm0
415 ; CHECK64-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3…
420 ; SSE4-32-NEXT: pmovzxwd {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero
421 ; SSE4-32-NEXT: pmulld {{\.LCPI.*}}, %xmm0
426 ; SSE4-64-NEXT: pmovzxwd {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero
427 ; SSE4-64-NEXT: pmulld {{.*}}(%rip), %xmm0
432 ; AVX-32-NEXT: vpmovzxwd {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero
434 ; AVX-32-NEXT: vpmulld %xmm1, %xmm0, %xmm0
439 ; AVX-64-NEXT: vpmovzxwd {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero
441 ; AVX-64-NEXT: vpmulld %xmm1, %xmm0, %xmm0
451 ; SLM32-NEXT: movdqa %xmm0, %xmm1
452 ; SLM32-NEXT: movdqa {{.*#+}} xmm0 = [18778,18778,18778,18778,18778,18778,18778,18778]
454 ; SLM32-NEXT: pmullw %xmm0, %xmm1
455 ; SLM32-NEXT: pmulhuw %xmm0, %xmm2
456 ; SLM32-NEXT: movdqa %xmm1, %xmm0
458 ; SLM32-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],…
463 ; SLM64-NEXT: movdqa %xmm0, %xmm1
464 ; SLM64-NEXT: movdqa {{.*#+}} xmm0 = [18778,18778,18778,18778,18778,18778,18778,18778]
466 ; SLM64-NEXT: pmullw %xmm0, %xmm1
467 ; SLM64-NEXT: pmulhuw %xmm0, %xmm2
468 ; SLM64-NEXT: movdqa %xmm1, %xmm0
470 ; SLM64-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],…
475 ; SLOW32-NEXT: movdqa %xmm0, %xmm1
476 ; SLOW32-NEXT: movdqa {{.*#+}} xmm0 = [18778,18778,18778,18778,18778,18778,18778,18778]
478 ; SLOW32-NEXT: pmulhuw %xmm0, %xmm2
479 ; SLOW32-NEXT: pmullw %xmm0, %xmm1
480 ; SLOW32-NEXT: movdqa %xmm1, %xmm0
481 ; SLOW32-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3]…
487 ; SLOW64-NEXT: movdqa %xmm0, %xmm1
488 ; SLOW64-NEXT: movdqa {{.*#+}} xmm0 = [18778,18778,18778,18778,18778,18778,18778,18778]
490 ; SLOW64-NEXT: pmulhuw %xmm0, %xmm2
491 ; SLOW64-NEXT: pmullw %xmm0, %xmm1
492 ; SLOW64-NEXT: movdqa %xmm1, %xmm0
493 ; SLOW64-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3]…
500 ; SSE4-32-NEXT: pmovzxwd {{.*#+}} xmm2 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero
501 ; SSE4-32-NEXT: punpckhwd {{.*#+}} xmm0 = xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7…
504 ; SSE4-32-NEXT: pmulld %xmm0, %xmm1
505 ; SSE4-32-NEXT: movdqa %xmm2, %xmm0
511 ; SSE4-64-NEXT: pmovzxwd {{.*#+}} xmm2 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero
512 ; SSE4-64-NEXT: punpckhwd {{.*#+}} xmm0 = xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7…
515 ; SSE4-64-NEXT: pmulld %xmm0, %xmm1
516 ; SSE4-64-NEXT: movdqa %xmm2, %xmm0
521 … vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0…
528 … vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0…
541 ; SLM32-NEXT: movdqa %xmm0, %xmm1
542 ; SLM32-NEXT: movdqa {{.*#+}} xmm0 = [18778,18778,18778,18778,18778,18778,18778,18778]
545 ; SLM32-NEXT: pmullw %xmm0, %xmm1
546 ; SLM32-NEXT: pmulhuw %xmm0, %xmm2
547 ; SLM32-NEXT: pmullw %xmm0, %xmm3
548 ; SLM32-NEXT: pmulhuw %xmm0, %xmm4
549 ; SLM32-NEXT: movdqa %xmm1, %xmm0
551 ; SLM32-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],…
560 ; SLM64-NEXT: movdqa %xmm0, %xmm1
561 ; SLM64-NEXT: movdqa {{.*#+}} xmm0 = [18778,18778,18778,18778,18778,18778,18778,18778]
564 ; SLM64-NEXT: pmullw %xmm0, %xmm1
565 ; SLM64-NEXT: pmulhuw %xmm0, %xmm2
566 ; SLM64-NEXT: pmullw %xmm0, %xmm3
567 ; SLM64-NEXT: pmulhuw %xmm0, %xmm4
568 ; SLM64-NEXT: movdqa %xmm1, %xmm0
570 ; SLM64-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],…
579 ; SLOW32-NEXT: movdqa %xmm0, %xmm1
581 ; SLOW32-NEXT: movdqa %xmm0, %xmm4
584 ; SLOW32-NEXT: movdqa %xmm1, %xmm0
585 ; SLOW32-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm4[0],xmm0[1],xmm4[1],xmm0[2],xmm4[2],xmm0[3]…
598 ; SLOW64-NEXT: movdqa %xmm0, %xmm1
600 ; SLOW64-NEXT: movdqa %xmm0, %xmm4
603 ; SLOW64-NEXT: movdqa %xmm1, %xmm0
604 ; SLOW64-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm4[0],xmm0[1],xmm4[1],xmm0[2],xmm4[2],xmm0[3]…
616 ; SSE4-32-NEXT: movdqa %xmm0, %xmm4
620 ; SSE4-32-NEXT: pmovzxwd {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero
623 ; SSE4-32-NEXT: pmulld %xmm3, %xmm0
632 ; SSE4-64-NEXT: movdqa %xmm0, %xmm4
636 ; SSE4-64-NEXT: pmovzxwd {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero
639 ; SSE4-64-NEXT: pmulld %xmm3, %xmm0
650 … vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0…
660 … vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0…
689 … CHECK32-NEXT: pmovzxbd {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],z…
690 ; CHECK32-NEXT: pmaddwd {{\.LCPI.*}}, %xmm0
695 … CHECK64-NEXT: pmovzxbd {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],z…
696 ; CHECK64-NEXT: pmaddwd {{.*}}(%rip), %xmm0
701 … SSE4-32-NEXT: pmovzxbd {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],z…
702 ; SSE4-32-NEXT: pmaddwd {{\.LCPI.*}}, %xmm0
707 … SSE4-64-NEXT: pmovzxbd {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],z…
708 ; SSE4-64-NEXT: pmaddwd {{.*}}(%rip), %xmm0
713 …AVX2-32-NEXT: vpmovzxbd {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],z…
714 ; AVX2-32-NEXT: vpmaddwd {{\.LCPI.*}}, %xmm0, %xmm0
719 …AVX2-64-NEXT: vpmovzxbd {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],z…
720 ; AVX2-64-NEXT: vpmaddwd {{.*}}(%rip), %xmm0, %xmm0
725 …12DQ-32-NEXT: vpmovzxbd {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],z…
726 ; AVX512DQ-32-NEXT: vpmaddwd {{\.LCPI.*}}, %xmm0, %xmm0
731 …12DQ-64-NEXT: vpmovzxbd {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],z…
732 ; AVX512DQ-64-NEXT: vpmaddwd {{.*}}(%rip), %xmm0, %xmm0
737 …12BW-32-NEXT: vpmovzxbd {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],z…
738 ; AVX512BW-32-NEXT: vpmaddwd {{\.LCPI.*}}, %xmm0, %xmm0
743 …12BW-64-NEXT: vpmovzxbd {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],z…
744 ; AVX512BW-64-NEXT: vpmaddwd {{.*}}(%rip), %xmm0, %xmm0
749 … KNL-32-NEXT: vpmovzxbd {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],z…
751 ; KNL-32-NEXT: vpmulld %xmm1, %xmm0, %xmm0
756 … KNL-64-NEXT: vpmovzxbd {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],z…
758 ; KNL-64-NEXT: vpmulld %xmm1, %xmm0, %xmm0
769 ; SLM32-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
770 ; SLM32-NEXT: pmovzxbd {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zer…
772 ; SLM32-NEXT: pmaddwd %xmm2, %xmm0
779 ; SLM64-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
780 ; SLM64-NEXT: pmovzxbd {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zer…
782 ; SLM64-NEXT: pmaddwd %xmm2, %xmm0
788 ; SLOW32-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
790 ; SLOW32-NEXT: pmovzxbd {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],ze…
792 ; SLOW32-NEXT: pmaddwd %xmm2, %xmm0
798 ; SLOW64-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
800 ; SLOW64-NEXT: pmovzxbd {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],ze…
802 ; SLOW64-NEXT: pmaddwd %xmm2, %xmm0
808 ; SSE4-32-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
810 … SSE4-32-NEXT: pmovzxbd {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],z…
812 ; SSE4-32-NEXT: pmaddwd %xmm2, %xmm0
818 ; SSE4-64-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
820 … SSE4-64-NEXT: pmovzxbd {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],z…
822 ; SSE4-64-NEXT: pmaddwd %xmm2, %xmm0
828 …xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero,xmm0[4…
834 …xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero,xmm0[4…
840 …xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero,xmm0[4…
846 …xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero,xmm0[4…
852 …xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero,xmm0[4…
858 …xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero,xmm0[4…
864 …xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero,xmm0[4…
871 …xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero,xmm0[4…
883 ; SLM32-NEXT: pshufd {{.*#+}} xmm1 = xmm0[3,3,3,3]
885 ; SLM32-NEXT: pshufd {{.*#+}} xmm4 = xmm0[1,1,1,1]
887 ; SLM32-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
888 ; SLM32-NEXT: pmovzxbd {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zer…
891 ; SLM32-NEXT: pmaddwd %xmm5, %xmm0
899 ; SLM64-NEXT: pshufd {{.*#+}} xmm1 = xmm0[3,3,3,3]
901 ; SLM64-NEXT: pshufd {{.*#+}} xmm4 = xmm0[1,1,1,1]
903 ; SLM64-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
904 ; SLM64-NEXT: pmovzxbd {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zer…
907 ; SLM64-NEXT: pmaddwd %xmm5, %xmm0
915 ; SLOW32-NEXT: pshufd {{.*#+}} xmm1 = xmm0[3,3,3,3]
917 ; SLOW32-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
919 ; SLOW32-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
921 ; SLOW32-NEXT: pmovzxbd {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],ze…
923 ; SLOW32-NEXT: pmaddwd %xmm4, %xmm0
931 ; SLOW64-NEXT: pshufd {{.*#+}} xmm1 = xmm0[3,3,3,3]
933 ; SLOW64-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
935 ; SLOW64-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
937 ; SLOW64-NEXT: pmovzxbd {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],ze…
939 ; SLOW64-NEXT: pmaddwd %xmm4, %xmm0
947 ; SSE4-32-NEXT: pshufd {{.*#+}} xmm1 = xmm0[3,3,3,3]
949 ; SSE4-32-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
951 ; SSE4-32-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
953 … SSE4-32-NEXT: pmovzxbd {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],z…
955 ; SSE4-32-NEXT: pmaddwd %xmm4, %xmm0
963 ; SSE4-64-NEXT: pshufd {{.*#+}} xmm1 = xmm0[3,3,3,3]
965 ; SSE4-64-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
967 ; SSE4-64-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
969 … SSE4-64-NEXT: pmovzxbd {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],z…
971 ; SSE4-64-NEXT: pmaddwd %xmm4, %xmm0
979 ; AVX2-32-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
981 …xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero,xmm0[4…
989 ; AVX2-64-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
991 …xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero,xmm0[4…
999 …xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero,xmm0[4…
1005 …xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero,xmm0[4…
1011 …xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero,xmm0[4…
1017 …xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero,xmm0[4…
1023 …xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero,xmm0[4…
1029 …xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero,xmm0[4…
1040 ; CHECK32-NEXT: pmovzxwd {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero
1041 ; CHECK32-NEXT: pmulld {{\.LCPI.*}}, %xmm0
1046 ; CHECK64-NEXT: pmovzxwd {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero
1047 ; CHECK64-NEXT: pmulld {{.*}}(%rip), %xmm0
1052 ; SSE4-32-NEXT: pmovzxwd {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero
1053 ; SSE4-32-NEXT: pmulld {{\.LCPI.*}}, %xmm0
1058 ; SSE4-64-NEXT: pmovzxwd {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero
1059 ; SSE4-64-NEXT: pmulld {{.*}}(%rip), %xmm0
1064 ; AVX-32-NEXT: vpmovzxwd {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero
1066 ; AVX-32-NEXT: vpmulld %xmm1, %xmm0, %xmm0
1071 ; AVX-64-NEXT: vpmovzxwd {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero
1073 ; AVX-64-NEXT: vpmulld %xmm1, %xmm0, %xmm0
1084 ; CHECK32-NEXT: pmovzxwd {{.*#+}} xmm2 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero
1085 ; CHECK32-NEXT: punpckhwd {{.*#+}} xmm0 = xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7…
1088 ; CHECK32-NEXT: pmulld %xmm0, %xmm1
1089 ; CHECK32-NEXT: movdqa %xmm2, %xmm0
1095 ; CHECK64-NEXT: pmovzxwd {{.*#+}} xmm2 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero
1096 ; CHECK64-NEXT: punpckhwd {{.*#+}} xmm0 = xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7…
1099 ; CHECK64-NEXT: pmulld %xmm0, %xmm1
1100 ; CHECK64-NEXT: movdqa %xmm2, %xmm0
1106 ; SSE4-32-NEXT: pmovzxwd {{.*#+}} xmm2 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero
1107 ; SSE4-32-NEXT: punpckhwd {{.*#+}} xmm0 = xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7…
1110 ; SSE4-32-NEXT: pmulld %xmm0, %xmm1
1111 ; SSE4-32-NEXT: movdqa %xmm2, %xmm0
1117 ; SSE4-64-NEXT: pmovzxwd {{.*#+}} xmm2 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero
1118 ; SSE4-64-NEXT: punpckhwd {{.*#+}} xmm0 = xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7…
1121 ; SSE4-64-NEXT: pmulld %xmm0, %xmm1
1122 ; SSE4-64-NEXT: movdqa %xmm2, %xmm0
1127 … vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0…
1134 … vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0…
1146 ; SLM32-NEXT: movdqa %xmm0, %xmm4
1149 ; SLM32-NEXT: pmovzxwd {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero
1154 ; SLM32-NEXT: pmulld %xmm3, %xmm0
1162 ; SLM64-NEXT: movdqa %xmm0, %xmm4
1165 ; SLM64-NEXT: pmovzxwd {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero
1170 ; SLM64-NEXT: pmulld %xmm3, %xmm0
1178 ; SLOW32-NEXT: movdqa %xmm0, %xmm4
1182 ; SLOW32-NEXT: pmovzxwd {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero
1185 ; SLOW32-NEXT: pmulld %xmm3, %xmm0
1194 ; SLOW64-NEXT: movdqa %xmm0, %xmm4
1198 ; SLOW64-NEXT: pmovzxwd {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero
1201 ; SLOW64-NEXT: pmulld %xmm3, %xmm0
1210 ; SSE4-32-NEXT: movdqa %xmm0, %xmm4
1214 ; SSE4-32-NEXT: pmovzxwd {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero
1217 ; SSE4-32-NEXT: pmulld %xmm3, %xmm0
1226 ; SSE4-64-NEXT: movdqa %xmm0, %xmm4
1230 ; SSE4-64-NEXT: pmovzxwd {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero
1233 ; SSE4-64-NEXT: pmulld %xmm3, %xmm0
1244 … vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0…
1254 … vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0…