• Home
  • Raw
  • Download

Lines Matching refs:xmm0

58 ; SSE-NEXT:    pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,7,6,5,4]
63 ; AVX-NEXT: vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,7,6,5,4]
77 ; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[3,2,1,0,4,5,6,7]
82 ; AVX-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[3,2,1,0,4,5,6,7]
96 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,0,0,0]
101 ; AVX1-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,0,0,0]
106 ; AVX2-NEXT: vbroadcastss %xmm0, %xmm0
138 ; SSE-NEXT: pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,7,6,5,4]
143 ; AVX-NEXT: vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,7,6,5,4]
155 ; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[3,2,1,0,4,5,6,7]
160 ; AVX-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[3,2,1,0,4,5,6,7]
172 ; SSE-NEXT: pand %xmm1, %xmm0
173 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,2,1,3]
178 ; AVX-NEXT: vpand %xmm1, %xmm0, %xmm0
179 ; AVX-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,2,1,3]
190 ; SSE-NEXT: por %xmm1, %xmm0
191 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,2,1,3]
196 ; AVX-NEXT: vpor %xmm1, %xmm0, %xmm0
197 ; AVX-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,2,1,3]
208 ; SSE-NEXT: pxor %xmm1, %xmm0
209 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,2,1,3]
214 ; AVX-NEXT: vpxor %xmm1, %xmm0, %xmm0
215 ; AVX-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,2,1,3]
226 ; SSE-NEXT: pand %xmm1, %xmm0
227 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,2,1,3]
232 ; AVX-NEXT: vpand %xmm1, %xmm0, %xmm0
233 ; AVX-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,2,1,3]
244 ; SSE-NEXT: por %xmm1, %xmm0
245 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,2,1,3]
250 ; AVX-NEXT: vpor %xmm1, %xmm0, %xmm0
251 ; AVX-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,2,1,3]
262 ; SSE-NEXT: pxor %xmm1, %xmm0
263 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,2,1,3]
268 ; AVX-NEXT: vpxor %xmm1, %xmm0, %xmm0
269 ; AVX-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,2,1,3]
284 ; SSE2-NEXT: pand %xmm1, %xmm0
285 ; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
287 ; SSE2-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
292 ; SSSE3-NEXT: pand %xmm1, %xmm0
293 ; SSSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
295 ; SSSE3-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
300 ; SSE41-NEXT: pand %xmm1, %xmm0
301 ; SSE41-NEXT: pblendw {{.*#+}} xmm0 = xmm0[0,1],xmm2[2,3],xmm0[4,5],xmm2[6,7]
306 ; AVX1-NEXT: vpand %xmm1, %xmm0, %xmm0
307 ; AVX1-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0,1],xmm2[2,3],xmm0[4,5],xmm2[6,7]
312 ; AVX2-NEXT: vpand %xmm1, %xmm0, %xmm0
313 ; AVX2-NEXT: vpblendd {{.*#+}} xmm0 = xmm0[0],xmm2[1],xmm0[2],xmm2[3]
324 ; SSE2-NEXT: por %xmm1, %xmm0
325 ; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
327 ; SSE2-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
332 ; SSSE3-NEXT: por %xmm1, %xmm0
333 ; SSSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
335 ; SSSE3-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
340 ; SSE41-NEXT: por %xmm1, %xmm0
341 ; SSE41-NEXT: pblendw {{.*#+}} xmm0 = xmm0[0,1],xmm2[2,3],xmm0[4,5],xmm2[6,7]
346 ; AVX1-NEXT: vpor %xmm1, %xmm0, %xmm0
347 ; AVX1-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0,1],xmm2[2,3],xmm0[4,5],xmm2[6,7]
352 ; AVX2-NEXT: vpor %xmm1, %xmm0, %xmm0
353 ; AVX2-NEXT: vpblendd {{.*#+}} xmm0 = xmm0[0],xmm2[1],xmm0[2],xmm2[3]
364 ; SSE2-NEXT: xorps %xmm1, %xmm0
365 ; SSE2-NEXT: andps {{.*}}(%rip), %xmm0
370 ; SSSE3-NEXT: xorps %xmm1, %xmm0
371 ; SSSE3-NEXT: andps {{.*}}(%rip), %xmm0
376 ; SSE41-NEXT: pxor %xmm1, %xmm0
378 ; SSE41-NEXT: pblendw {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3],xmm0[4,5],xmm1[6,7]
383 ; AVX1-NEXT: vpxor %xmm1, %xmm0, %xmm0
385 ; AVX1-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3],xmm0[4,5],xmm1[6,7]
390 ; AVX2-NEXT: vpxor %xmm1, %xmm0, %xmm0
392 ; AVX2-NEXT: vpblendd {{.*#+}} xmm0 = xmm0[0],xmm1[1],xmm0[2],xmm1[3]
403 ; SSE2-NEXT: pand %xmm1, %xmm0
404 ; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,3,2,3]
405 ; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm2[0,2,2,3]
406 ; SSE2-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
411 ; SSSE3-NEXT: pand %xmm1, %xmm0
412 ; SSSE3-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,3,2,3]
413 ; SSSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm2[0,2,2,3]
414 ; SSSE3-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
419 ; SSE41-NEXT: pand %xmm1, %xmm0
420 ; SSE41-NEXT: pblendw {{.*#+}} xmm0 = xmm2[0,1],xmm0[2,3],xmm2[4,5],xmm0[6,7]
425 ; AVX1-NEXT: vpand %xmm1, %xmm0, %xmm0
426 ; AVX1-NEXT: vpblendw {{.*#+}} xmm0 = xmm2[0,1],xmm0[2,3],xmm2[4,5],xmm0[6,7]
431 ; AVX2-NEXT: vpand %xmm1, %xmm0, %xmm0
432 ; AVX2-NEXT: vpblendd {{.*#+}} xmm0 = xmm2[0],xmm0[1],xmm2[2],xmm0[3]
443 ; SSE2-NEXT: por %xmm1, %xmm0
444 ; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,3,2,3]
445 ; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm2[0,2,2,3]
446 ; SSE2-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
451 ; SSSE3-NEXT: por %xmm1, %xmm0
452 ; SSSE3-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,3,2,3]
453 ; SSSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm2[0,2,2,3]
454 ; SSSE3-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
459 ; SSE41-NEXT: por %xmm1, %xmm0
460 ; SSE41-NEXT: pblendw {{.*#+}} xmm0 = xmm2[0,1],xmm0[2,3],xmm2[4,5],xmm0[6,7]
465 ; AVX1-NEXT: vpor %xmm1, %xmm0, %xmm0
466 ; AVX1-NEXT: vpblendw {{.*#+}} xmm0 = xmm2[0,1],xmm0[2,3],xmm2[4,5],xmm0[6,7]
471 ; AVX2-NEXT: vpor %xmm1, %xmm0, %xmm0
472 ; AVX2-NEXT: vpblendd {{.*#+}} xmm0 = xmm2[0],xmm0[1],xmm2[2],xmm0[3]
483 ; SSE2-NEXT: xorps %xmm1, %xmm0
484 ; SSE2-NEXT: andps {{.*}}(%rip), %xmm0
489 ; SSSE3-NEXT: xorps %xmm1, %xmm0
490 ; SSSE3-NEXT: andps {{.*}}(%rip), %xmm0
495 ; SSE41-NEXT: pxor %xmm1, %xmm0
497 ; SSE41-NEXT: pblendw {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3],xmm1[4,5],xmm0[6,7]
502 ; AVX1-NEXT: vpxor %xmm1, %xmm0, %xmm0
504 ; AVX1-NEXT: vpblendw {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3],xmm1[4,5],xmm0[6,7]
509 ; AVX2-NEXT: vpxor %xmm1, %xmm0, %xmm0
511 ; AVX2-NEXT: vpblendd {{.*#+}} xmm0 = xmm1[0],xmm0[1],xmm1[2],xmm0[3]
522 ; SSE2-NEXT: pand %xmm1, %xmm0
523 ; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
525 ; SSE2-NEXT: punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
530 ; SSSE3-NEXT: pand %xmm1, %xmm0
531 ; SSSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
533 ; SSSE3-NEXT: punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
538 ; SSE41-NEXT: pand %xmm1, %xmm0
539 ; SSE41-NEXT: pblendw {{.*#+}} xmm0 = xmm0[0,1],xmm2[2,3],xmm0[4,5],xmm2[6,7]
540 ; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,2,1,3]
545 ; AVX1-NEXT: vpand %xmm1, %xmm0, %xmm0
546 ; AVX1-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0,1],xmm2[2,3],xmm0[4,5],xmm2[6,7]
547 ; AVX1-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,2,1,3]
552 ; AVX2-NEXT: vpand %xmm1, %xmm0, %xmm0
553 ; AVX2-NEXT: vpblendd {{.*#+}} xmm0 = xmm0[0],xmm2[1],xmm0[2],xmm2[3]
554 ; AVX2-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,2,1,3]
565 ; SSE2-NEXT: por %xmm1, %xmm0
566 ; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
568 ; SSE2-NEXT: punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
573 ; SSSE3-NEXT: por %xmm1, %xmm0
574 ; SSSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
576 ; SSSE3-NEXT: punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
581 ; SSE41-NEXT: por %xmm1, %xmm0
582 ; SSE41-NEXT: pblendw {{.*#+}} xmm0 = xmm0[0,1],xmm2[2,3],xmm0[4,5],xmm2[6,7]
583 ; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,2,1,3]
588 ; AVX1-NEXT: vpor %xmm1, %xmm0, %xmm0
589 ; AVX1-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0,1],xmm2[2,3],xmm0[4,5],xmm2[6,7]
590 ; AVX1-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,2,1,3]
595 ; AVX2-NEXT: vpor %xmm1, %xmm0, %xmm0
596 ; AVX2-NEXT: vpblendd {{.*#+}} xmm0 = xmm0[0],xmm2[1],xmm0[2],xmm2[3]
597 ; AVX2-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,2,1,3]
608 ; SSE2-NEXT: pxor %xmm1, %xmm0
609 ; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,1,0,2]
611 ; SSE2-NEXT: punpckhqdq {{.*#+}} xmm0 = xmm0[1],xmm1[1]
616 ; SSSE3-NEXT: pxor %xmm1, %xmm0
617 ; SSSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,1,0,2]
619 ; SSSE3-NEXT: punpckhqdq {{.*#+}} xmm0 = xmm0[1],xmm1[1]
624 ; SSE41-NEXT: pxor %xmm1, %xmm0
625 ; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
626 ; SSE41-NEXT: movq {{.*#+}} xmm0 = xmm0[0],zero
631 ; AVX-NEXT: vpxor %xmm1, %xmm0, %xmm0
632 ; AVX-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
633 ; AVX-NEXT: vmovq {{.*#+}} xmm0 = xmm0[0],zero
644 ; SSE2-NEXT: pand %xmm1, %xmm0
645 ; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,3,2,3]
646 ; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm2[0,2,2,3]
647 ; SSE2-NEXT: punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
652 ; SSSE3-NEXT: pand %xmm1, %xmm0
653 ; SSSE3-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,3,2,3]
654 ; SSSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm2[0,2,2,3]
655 ; SSSE3-NEXT: punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
660 ; SSE41-NEXT: pand %xmm1, %xmm0
661 ; SSE41-NEXT: pblendw {{.*#+}} xmm0 = xmm2[0,1],xmm0[2,3],xmm2[4,5],xmm0[6,7]
662 ; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,2,1,3]
667 ; AVX1-NEXT: vpand %xmm1, %xmm0, %xmm0
668 ; AVX1-NEXT: vpblendw {{.*#+}} xmm0 = xmm2[0,1],xmm0[2,3],xmm2[4,5],xmm0[6,7]
669 ; AVX1-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,2,1,3]
674 ; AVX2-NEXT: vpand %xmm1, %xmm0, %xmm0
675 ; AVX2-NEXT: vpblendd {{.*#+}} xmm0 = xmm2[0],xmm0[1],xmm2[2],xmm0[3]
676 ; AVX2-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,2,1,3]
687 ; SSE2-NEXT: por %xmm1, %xmm0
688 ; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,3,2,3]
689 ; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm2[0,2,2,3]
690 ; SSE2-NEXT: punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
695 ; SSSE3-NEXT: por %xmm1, %xmm0
696 ; SSSE3-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,3,2,3]
697 ; SSSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm2[0,2,2,3]
698 ; SSSE3-NEXT: punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
703 ; SSE41-NEXT: por %xmm1, %xmm0
704 ; SSE41-NEXT: pblendw {{.*#+}} xmm0 = xmm2[0,1],xmm0[2,3],xmm2[4,5],xmm0[6,7]
705 ; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,2,1,3]
710 ; AVX1-NEXT: vpor %xmm1, %xmm0, %xmm0
711 ; AVX1-NEXT: vpblendw {{.*#+}} xmm0 = xmm2[0,1],xmm0[2,3],xmm2[4,5],xmm0[6,7]
712 ; AVX1-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,2,1,3]
717 ; AVX2-NEXT: vpor %xmm1, %xmm0, %xmm0
718 ; AVX2-NEXT: vpblendd {{.*#+}} xmm0 = xmm2[0],xmm0[1],xmm2[2],xmm0[3]
719 ; AVX2-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,2,1,3]
730 ; SSE2-NEXT: pxor %xmm1, %xmm0
731 ; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,3,2,3]
732 ; SSE2-NEXT: pxor %xmm0, %xmm0
733 ; SSE2-NEXT: punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
738 ; SSSE3-NEXT: pxor %xmm1, %xmm0
739 ; SSSE3-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,3,2,3]
740 ; SSSE3-NEXT: pxor %xmm0, %xmm0
741 ; SSSE3-NEXT: punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
746 ; SSE41-NEXT: pxor %xmm1, %xmm0
747 ; SSE41-NEXT: pshufd {{.*#+}} xmm1 = xmm0[0,1,1,3]
748 ; SSE41-NEXT: pxor %xmm0, %xmm0
749 ; SSE41-NEXT: pblendw {{.*#+}} xmm0 = xmm0[0,1,2,3],xmm1[4,5,6,7]
754 ; AVX1-NEXT: vpxor %xmm1, %xmm0, %xmm0
755 ; AVX1-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,1,1,3]
757 ; AVX1-NEXT: vpblendw {{.*#+}} xmm0 = xmm1[0,1,2,3],xmm0[4,5,6,7]
762 ; AVX2-NEXT: vpxor %xmm1, %xmm0, %xmm0
763 ; AVX2-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,1,1,3]
765 ; AVX2-NEXT: vpblendd {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3]
776 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[3,1,0,1]
781 ; AVX-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[3,1,0,1]
791 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[2,1,0,3]
796 ; AVX-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,1,0,3]
806 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[2,1,0,3]
811 ; AVX-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,1,0,3]
821 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,1,0,1]
826 ; AVX1-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,1,0,1]
831 ; AVX2-NEXT: vpbroadcastq %xmm0, %xmm0
841 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[2,3,2,3]
846 ; AVX-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,3,2,3]
856 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[2,3,0,1]
861 ; AVX-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,3,0,1]
871 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,2,0,2]
876 ; AVX-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,2,0,2]
886 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
891 ; AVX-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
901 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,3,2,2]
906 ; AVX-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[1,3,2,2]
916 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,1,3]
921 ; AVX-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[1,1,1,3]
931 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,2,1]
936 ; AVX-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[1,1,2,1]
946 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,1,0,1]
951 ; AVX1-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,1,0,1]
956 ; AVX2-NEXT: vpbroadcastq %xmm0, %xmm0
977 ; SSE-NEXT: movaps %xmm1, %xmm0
982 ; AVX-NEXT: vmovaps %xmm1, %xmm0
999 ; SSE2-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,0],xmm0[3,0]
1000 ; SSE2-NEXT: shufps {{.*#+}} xmm1 = xmm1[2,0],xmm0[0,1]
1001 ; SSE2-NEXT: movaps %xmm1, %xmm0
1006 ; SSSE3-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,0],xmm0[3,0]
1007 ; SSSE3-NEXT: shufps {{.*#+}} xmm1 = xmm1[2,0],xmm0[0,1]
1008 ; SSSE3-NEXT: movaps %xmm1, %xmm0
1014 ; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm0[3,1,0,1]
1015 ; SSE41-NEXT: pblendw {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3],xmm0[4,5,6,7]
1021 ; AVX1-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[3,1,0,1]
1022 ; AVX1-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3],xmm0[4,5,6,7]
1028 ; AVX2-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[3,1,0,1]
1029 ; AVX2-NEXT: vpblendd {{.*#+}} xmm0 = xmm0[0],xmm1[1],xmm0[2,3]
1040 ; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[2,0,2,3]
1041 ; SSE2-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
1047 ; SSSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm0[2,0,2,3]
1048 ; SSSE3-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
1053 ; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm0[2,3,0,1]
1054 ; SSE41-NEXT: pblendw {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3],xmm0[4,5],xmm1[6,7]
1059 ; AVX1-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,3,0,1]
1060 ; AVX1-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3],xmm0[4,5],xmm1[6,7]
1065 ; AVX2-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,3,0,1]
1066 ; AVX2-NEXT: vpblendd {{.*#+}} xmm0 = xmm0[0],xmm1[1],xmm0[2],xmm1[3]
1076 ; SSE2-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,0],xmm0[1,0]
1077 ; SSE2-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,1],xmm1[0,2]
1082 ; SSSE3-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,0],xmm0[1,0]
1083 ; SSSE3-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,1],xmm1[0,2]
1088 ; SSE41-NEXT: pblendw {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3,4,5,6,7]
1089 ; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm0[3,1,0,1]
1094 ; AVX1-NEXT: vpblendw {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3,4,5,6,7]
1095 ; AVX1-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[3,1,0,1]
1100 ; AVX2-NEXT: vpblendd {{.*#+}} xmm0 = xmm1[0],xmm0[1,2,3]
1101 ; AVX2-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[3,1,0,1]
1111 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,0,3]
1116 ; AVX-NEXT: vpshufd {{.*#+}} xmm0 = xmm1[1,1,0,3]
1126 ; SSE2-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
1127 ; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[3,0,0,0]
1132 ; SSSE3-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
1133 ; SSSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm0[3,0,0,0]
1138 ; SSE41-NEXT: pblendw {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3],xmm0[4,5,6,7]
1139 ; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,0,0,0]
1144 ; AVX1-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3],xmm0[4,5,6,7]
1145 ; AVX1-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[1,0,0,0]
1150 ; AVX2-NEXT: vpblendd {{.*#+}} xmm0 = xmm0[0],xmm1[1],xmm0[2,3]
1151 ; AVX2-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[1,0,0,0]
1161 ; SSE2-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,0],xmm0[2,3]
1163 ; SSE2-NEXT: movaps %xmm1, %xmm0
1168 ; SSSE3-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,0],xmm0[2,3]
1170 ; SSSE3-NEXT: movaps %xmm1, %xmm0
1175 ; SSE41-NEXT: pblendw {{.*#+}} xmm0 = xmm1[0,1,2,3],xmm0[4,5,6,7]
1176 ; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,2,3,0]
1181 ; AVX1-NEXT: vpblendw {{.*#+}} xmm0 = xmm1[0,1,2,3],xmm0[4,5,6,7]
1182 ; AVX1-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,2,3,0]
1187 ; AVX2-NEXT: vpblendd {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3]
1188 ; AVX2-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,2,3,0]
1198 ; SSE2-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
1199 ; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[0,3,0,3]
1204 ; SSSE3-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
1205 ; SSSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm1[0,3,0,3]
1210 ; SSE41-NEXT: pblendw {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3],xmm1[4,5,6,7]
1211 ; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,1,0,1]
1216 ; AVX1-NEXT: vpblendw {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3],xmm1[4,5,6,7]
1217 ; AVX1-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,1,0,1]
1222 ; AVX2-NEXT: vpblendd {{.*#+}} xmm0 = xmm1[0],xmm0[1],xmm1[2,3]
1223 ; AVX2-NEXT: vpbroadcastq %xmm0, %xmm0
1237 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,3]
1242 ; AVX-NEXT: vpshufd {{.*#+}} xmm0 = xmm1[1,1,1,3]
1252 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm1[0,1,0,3]
1257 ; AVX-NEXT: vpshufd {{.*#+}} xmm0 = xmm1[0,1,0,3]
1267 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm1[0,3,2,3]
1272 ; AVX-NEXT: vpshufd {{.*#+}} xmm0 = xmm1[0,3,2,3]
1282 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,1,0,1]
1287 ; AVX1-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,1,0,1]
1292 ; AVX2-NEXT: vpbroadcastq %xmm0, %xmm0
1302 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[2,3,2,3]
1307 ; AVX-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,3,2,3]
1317 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,1,0,1]
1322 ; AVX1-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,1,0,1]
1327 ; AVX2-NEXT: vpbroadcastq %xmm0, %xmm0
1337 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,1,1,0]
1342 ; AVX-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,1,1,0]
1352 ; SSE-NEXT: movaps %xmm1, %xmm0
1357 ; AVX-NEXT: vmovaps %xmm1, %xmm0
1367 ; SSE2-NEXT: movss {{.*#+}} xmm1 = xmm0[0],xmm1[1,2,3]
1368 ; SSE2-NEXT: movaps %xmm1, %xmm0
1373 ; SSSE3-NEXT: movss {{.*#+}} xmm1 = xmm0[0],xmm1[1,2,3]
1374 ; SSSE3-NEXT: movaps %xmm1, %xmm0
1379 ; SSE41-NEXT: blendps {{.*#+}} xmm0 = xmm0[0],xmm1[1,2,3]
1384 ; AVX-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0],xmm1[1,2,3]
1394 ; SSE-NEXT: unpcklpd {{.*#+}} xmm0 = xmm0[0],xmm1[0]
1399 ; AVX-NEXT: vunpcklpd {{.*#+}} xmm0 = xmm0[0],xmm1[0]
1409 ; SSE-NEXT: unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
1410 ; SSE-NEXT: movapd %xmm1, %xmm0
1415 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
1425 ; SSE2-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,0],xmm1[0,0]
1426 ; SSE2-NEXT: shufps {{.*#+}} xmm0 = xmm0[2,0],xmm1[2,3]
1431 ; SSSE3-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,0],xmm1[0,0]
1432 ; SSSE3-NEXT: shufps {{.*#+}} xmm0 = xmm0[2,0],xmm1[2,3]
1437 ; SSE41-NEXT: blendps {{.*#+}} xmm0 = xmm1[0],xmm0[1],xmm1[2,3]
1442 ; AVX-NEXT: vblendps {{.*#+}} xmm0 = xmm1[0],xmm0[1],xmm1[2,3]
1452 ; SSE-NEXT: movaps %xmm1, %xmm0
1457 ; AVX-NEXT: vmovaps %xmm1, %xmm0
1467 ; SSE2-NEXT: movss {{.*#+}} xmm1 = xmm0[0],xmm1[1,2,3]
1468 ; SSE2-NEXT: movaps %xmm1, %xmm0
1473 ; SSSE3-NEXT: movss {{.*#+}} xmm1 = xmm0[0],xmm1[1,2,3]
1474 ; SSSE3-NEXT: movaps %xmm1, %xmm0
1479 ; SSE41-NEXT: pblendw {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3,4,5,6,7]
1484 ; AVX1-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3,4,5,6,7]
1489 ; AVX2-NEXT: vpblendd {{.*#+}} xmm0 = xmm0[0],xmm1[1,2,3]
1499 ; SSE-NEXT: punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
1504 ; AVX-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
1514 ; SSE-NEXT: punpckhqdq {{.*#+}} xmm1 = xmm1[1],xmm0[1]
1515 ; SSE-NEXT: movdqa %xmm1, %xmm0
1520 ; AVX-NEXT: vpunpckhqdq {{.*#+}} xmm0 = xmm1[1],xmm0[1]
1530 ; SSE2-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,0],xmm1[0,0]
1531 ; SSE2-NEXT: shufps {{.*#+}} xmm0 = xmm0[2,0],xmm1[2,3]
1536 ; SSSE3-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,0],xmm1[0,0]
1537 ; SSSE3-NEXT: shufps {{.*#+}} xmm0 = xmm0[2,0],xmm1[2,3]
1542 ; SSE41-NEXT: pblendw {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3],xmm1[4,5,6,7]
1547 ; AVX1-NEXT: vpblendw {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3],xmm1[4,5,6,7]
1552 ; AVX2-NEXT: vpblendd {{.*#+}} xmm0 = xmm1[0],xmm0[1],xmm1[2,3]
1571 ; SSE2-NEXT: movss {{.*#+}} xmm1 = xmm0[0],xmm1[1,2,3]
1572 ; SSE2-NEXT: movaps %xmm1, %xmm0
1577 ; SSSE3-NEXT: movss {{.*#+}} xmm1 = xmm0[0],xmm1[1,2,3]
1578 ; SSSE3-NEXT: movaps %xmm1, %xmm0
1583 ; SSE41-NEXT: blendps {{.*#+}} xmm0 = xmm0[0],xmm1[1,2,3]
1588 ; AVX-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0],xmm1[1,2,3]
1598 ; SSE-NEXT: unpcklpd {{.*#+}} xmm0 = xmm0[0],xmm1[0]
1603 ; AVX-NEXT: vunpcklpd {{.*#+}} xmm0 = xmm0[0],xmm1[0]
1613 ; SSE-NEXT: unpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
1618 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
1628 ; SSE2-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,0],xmm1[0,0]
1629 ; SSE2-NEXT: shufps {{.*#+}} xmm0 = xmm0[2,0],xmm1[2,3]
1634 ; SSSE3-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,0],xmm1[0,0]
1635 ; SSSE3-NEXT: shufps {{.*#+}} xmm0 = xmm0[2,0],xmm1[2,3]
1640 ; SSE41-NEXT: blendps {{.*#+}} xmm0 = xmm1[0],xmm0[1],xmm1[2,3]
1645 ; AVX-NEXT: vblendps {{.*#+}} xmm0 = xmm1[0],xmm0[1],xmm1[2,3]
1664 ; SSE2-NEXT: movss {{.*#+}} xmm1 = xmm0[0],xmm1[1,2,3]
1665 ; SSE2-NEXT: movaps %xmm1, %xmm0
1670 ; SSSE3-NEXT: movss {{.*#+}} xmm1 = xmm0[0],xmm1[1,2,3]
1671 ; SSSE3-NEXT: movaps %xmm1, %xmm0
1676 ; SSE41-NEXT: pblendw {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3,4,5,6,7]
1681 ; AVX1-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3,4,5,6,7]
1686 ; AVX2-NEXT: vpblendd {{.*#+}} xmm0 = xmm0[0],xmm1[1,2,3]
1696 ; SSE-NEXT: punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
1701 ; AVX-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
1711 ; SSE-NEXT: punpckhqdq {{.*#+}} xmm0 = xmm0[1],xmm1[1]
1716 ; AVX-NEXT: vpunpckhqdq {{.*#+}} xmm0 = xmm0[1],xmm1[1]
1726 ; SSE2-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,0],xmm1[0,0]
1727 ; SSE2-NEXT: shufps {{.*#+}} xmm0 = xmm0[2,0],xmm1[2,3]
1732 ; SSSE3-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,0],xmm1[0,0]
1733 ; SSSE3-NEXT: shufps {{.*#+}} xmm0 = xmm0[2,0],xmm1[2,3]
1738 ; SSE41-NEXT: pblendw {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3],xmm1[4,5,6,7]
1743 ; AVX1-NEXT: vpblendw {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3],xmm1[4,5,6,7]
1748 ; AVX2-NEXT: vpblendd {{.*#+}} xmm0 = xmm1[0],xmm0[1],xmm1[2,3]
1758 ; SSE-NEXT: movdqa %xmm0, %xmm2
1760 ; SSE-NEXT: punpckhqdq {{.*#+}} xmm0 = xmm0[1],xmm1[1]
1767 ; AVX1-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm0[0],xmm1[0]
1768 ; AVX1-NEXT: vpunpckhqdq {{.*#+}} xmm0 = xmm0[1],xmm1[1]
1776 ; AVX2-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm0[0],xmm1[0]
1777 ; AVX2-NEXT: vpunpckhqdq {{.*#+}} xmm0 = xmm0[1],xmm1[1]
1790 ; SSE-NEXT: movq {{.*#+}} xmm0 = mem[0],zero
1791 ; SSE-NEXT: movhpd {{.*#+}} xmm0 = xmm0[0],mem[0]
1796 ; AVX-NEXT: vmovq {{.*#+}} xmm0 = mem[0],zero
1797 ; AVX-NEXT: vmovhpd {{.*#+}} xmm0 = xmm0[0],mem[0]
1810 ; SSE-NEXT: movups %xmm0, (%rdi)
1815 ; AVX-NEXT: vpermilpd {{.*#+}} xmm1 = xmm0[1,0]
1816 ; AVX-NEXT: vinsertps {{.*#+}} xmm1 = xmm0[0,1],xmm1[0],xmm0[3]
1817 ; AVX-NEXT: vinsertps {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[3]
1818 ; AVX-NEXT: vmovups %xmm0, (%rdi)
1836 ; SSE-NEXT: movaps %xmm1, %xmm0
1841 ; AVX-NEXT: vpermilps {{.*#+}} xmm0 = xmm1[0,1,2,0]
1852 ; SSE2-NEXT: movaps %xmm1, %xmm0
1857 ; SSSE3-NEXT: movddup {{.*#+}} xmm0 = xmm1[0,0]
1862 ; SSE41-NEXT: movddup {{.*#+}} xmm0 = xmm1[0,0]
1867 ; AVX-NEXT: vmovddup {{.*#+}} xmm0 = xmm1[0,0]
1877 ; SSE2-NEXT: shufps {{.*#+}} xmm0 = xmm0[0,0],xmm1[3,0]
1878 ; SSE2-NEXT: shufps {{.*#+}} xmm0 = xmm0[0,2],xmm1[2,3]
1883 ; SSSE3-NEXT: shufps {{.*#+}} xmm0 = xmm0[0,0],xmm1[3,0]
1884 ; SSSE3-NEXT: shufps {{.*#+}} xmm0 = xmm0[0,2],xmm1[2,3]
1889 ; SSE41-NEXT: blendpd {{.*#+}} xmm0 = xmm0[0],xmm1[1]
1890 ; SSE41-NEXT: shufps {{.*#+}} xmm0 = xmm0[0,3,2,3]
1895 ; AVX-NEXT: vblendpd {{.*#+}} xmm0 = xmm0[0],xmm1[1]
1896 ; AVX-NEXT: vpermilps {{.*#+}} xmm0 = xmm0[0,3,2,3]
1907 ; SSE-NEXT: movaps %xmm1, %xmm0
1912 ; AVX-NEXT: vpermilps {{.*#+}} xmm0 = xmm1[1,1,2,3]
1926 …}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3],xmm1[4],xmm0[4],xmm1[5],
1927 ; SSE2-NEXT: punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],x…
1928 ; SSE2-NEXT: movd {{.*#+}} xmm0 = mem[0],zero,zero,zero
1929 ; SSE2-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
1930 ; SSE2-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3]
1931 ; SSE2-NEXT: movss {{.*#+}} xmm0 = xmm1[0],xmm0[1,2,3]
1937 …}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3],xmm1[4],xmm0[4],xmm1[5],
1938 … SSSE3-NEXT: punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],
1939 ; SSSE3-NEXT: movd {{.*#+}} xmm0 = mem[0],zero,zero,zero
1940 ; SSSE3-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
1941 ; SSSE3-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3]
1942 ; SSSE3-NEXT: movss {{.*#+}} xmm0 = xmm1[0],xmm0[1,2,3]
1948 ; SSE41-NEXT: pmovzxbd {{.*#+}} xmm0 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,z…
1949 ; SSE41-NEXT: pblendw {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3,4,5,6,7]
1954 ; AVX1-NEXT: vpmovzxbd {{.*#+}} xmm0 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,z…
1956 ; AVX1-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3,4,5,6,7]
1961 ; AVX2-NEXT: vpmovzxbd {{.*#+}} xmm0 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,z…
1963 ; AVX2-NEXT: vpblendd {{.*#+}} xmm0 = xmm0[0],xmm1[1,2,3]
1975 ; SSE2-NEXT: movd {{.*#+}} xmm0 = mem[0],zero,zero,zero
1976 ; SSE2-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
1977 ; SSE2-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3]
1979 …}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3],xmm1[4],xmm0[4],xmm1[5],
1980 ; SSE2-NEXT: punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],x…
1981 ; SSE2-NEXT: punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
1986 ; SSSE3-NEXT: movd {{.*#+}} xmm0 = mem[0],zero,zero,zero
1987 ; SSSE3-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
1988 ; SSSE3-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3]
1990 …}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3],xmm1[4],xmm0[4],xmm1[5],
1991 … SSSE3-NEXT: punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],
1992 ; SSSE3-NEXT: punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
1997 ; SSE41-NEXT: pmovzxbd {{.*#+}} xmm0 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,z…
1999 ; SSE41-NEXT: punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
2004 ; AVX-NEXT: vpmovzxbd {{.*#+}} xmm0 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,ze…
2006 ; AVX-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
2019 …}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3],xmm1[4],xmm0[4],xmm1[5],
2020 ; SSE2-NEXT: punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],x…
2021 ; SSE2-NEXT: movd {{.*#+}} xmm0 = mem[0],zero,zero,zero
2022 ; SSE2-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
2023 ; SSE2-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3]
2024 ; SSE2-NEXT: punpckhqdq {{.*#+}} xmm0 = xmm0[1],xmm1[1]
2030 …}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3],xmm1[4],xmm0[4],xmm1[5],
2031 … SSSE3-NEXT: punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],
2032 ; SSSE3-NEXT: movd {{.*#+}} xmm0 = mem[0],zero,zero,zero
2033 ; SSSE3-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
2034 ; SSSE3-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3]
2035 ; SSSE3-NEXT: punpckhqdq {{.*#+}} xmm0 = xmm0[1],xmm1[1]
2041 ; SSE41-NEXT: pmovzxbd {{.*#+}} xmm0 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,z…
2042 ; SSE41-NEXT: punpckhqdq {{.*#+}} xmm0 = xmm0[1],xmm1[1]
2047 ; AVX-NEXT: vpmovzxbd {{.*#+}} xmm0 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,ze…
2049 ; AVX-NEXT: vpunpckhqdq {{.*#+}} xmm0 = xmm1[1],xmm0[1]
2061 ; SSE2-NEXT: movd {{.*#+}} xmm0 = mem[0],zero,zero,zero
2062 ; SSE2-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
2063 ; SSE2-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3]
2065 …}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3],xmm1[4],xmm0[4],xmm1[5],
2066 ; SSE2-NEXT: punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],x…
2067 ; SSE2-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,0],xmm1[0,0]
2068 ; SSE2-NEXT: shufps {{.*#+}} xmm0 = xmm0[2,0],xmm1[2,3]
2073 ; SSSE3-NEXT: movd {{.*#+}} xmm0 = mem[0],zero,zero,zero
2074 ; SSSE3-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
2075 ; SSSE3-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3]
2077 …}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3],xmm1[4],xmm0[4],xmm1[5],
2078 … SSSE3-NEXT: punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],
2079 ; SSSE3-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,0],xmm1[0,0]
2080 ; SSSE3-NEXT: shufps {{.*#+}} xmm0 = xmm0[2,0],xmm1[2,3]
2086 ; SSE41-NEXT: pmovzxbd {{.*#+}} xmm0 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,z…
2087 ; SSE41-NEXT: pblendw {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3],xmm0[4,5,6,7]
2092 ; AVX1-NEXT: vpmovzxbd {{.*#+}} xmm0 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,z…
2094 ; AVX1-NEXT: vpblendw {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3],xmm1[4,5,6,7]
2099 ; AVX2-NEXT: vpmovzxbd {{.*#+}} xmm0 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,z…
2101 ; AVX2-NEXT: vpblendd {{.*#+}} xmm0 = xmm1[0],xmm0[1],xmm1[2,3]
2143 ; SSE2-NEXT: movsd {{.*#+}} xmm0 = xmm1[0],xmm0[1]
2148 ; SSSE3-NEXT: movsd {{.*#+}} xmm0 = xmm1[0],xmm0[1]
2153 ; SSE41-NEXT: blendpd {{.*#+}} xmm0 = xmm1[0],xmm0[1]
2158 ; AVX-NEXT: vblendpd {{.*#+}} xmm0 = xmm1[0],xmm0[1]
2168 ; SSE2-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,2],xmm0[1,3]
2170 ; SSE2-NEXT: movaps %xmm1, %xmm0
2175 ; SSSE3-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,2],xmm0[1,3]
2177 ; SSSE3-NEXT: movaps %xmm1, %xmm0
2182 ; SSE41-NEXT: blendps {{.*#+}} xmm0 = xmm1[0],xmm0[1],xmm1[2],xmm0[3]
2187 ; AVX-NEXT: vblendps {{.*#+}} xmm0 = xmm1[0],xmm0[1],xmm1[2],xmm0[3]
2197 ; SSE2-NEXT: movss {{.*#+}} xmm1 = xmm0[0],xmm1[1,2,3]
2198 ; SSE2-NEXT: movaps %xmm1, %xmm0
2203 ; SSSE3-NEXT: movss {{.*#+}} xmm1 = xmm0[0],xmm1[1,2,3]
2204 ; SSSE3-NEXT: movaps %xmm1, %xmm0
2209 ; SSE41-NEXT: blendps {{.*#+}} xmm0 = xmm0[0],xmm1[1,2,3]
2214 ; AVX-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0],xmm1[1,2,3]
2225 ; SSE-NEXT: punpckhqdq {{.*#+}} xmm1 = xmm1[1],xmm0[1]
2226 ; SSE-NEXT: movdqa %xmm1, %xmm0
2231 ; AVX-NEXT: vpunpckhqdq {{.*#+}} xmm0 = xmm1[1],xmm0[1]
2241 ; SSE-NEXT: punpckhqdq {{.*#+}} xmm1 = xmm1[1],xmm0[1]
2242 ; SSE-NEXT: movdqa %xmm1, %xmm0
2247 ; AVX-NEXT: vpunpckhqdq {{.*#+}} xmm0 = xmm1[1],xmm0[1]
2257 ; SSE-NEXT: punpckhqdq {{.*#+}} xmm1 = xmm1[1],xmm0[1]
2258 ; SSE-NEXT: movdqa %xmm1, %xmm0
2263 ; AVX-NEXT: vpunpckhqdq {{.*#+}} xmm0 = xmm1[1],xmm0[1]
2277 ; SSE2-NEXT: movsd {{.*#+}} xmm0 = xmm1[0],xmm0[1]
2282 ; SSSE3-NEXT: movsd {{.*#+}} xmm0 = xmm1[0],xmm0[1]
2287 ; SSE41-NEXT: blendpd {{.*#+}} xmm0 = xmm1[0],xmm0[1]
2292 ; AVX-NEXT: vblendpd {{.*#+}} xmm0 = xmm1[0],xmm0[1]
2302 ; SSE-NEXT: unpcklpd {{.*#+}} xmm0 = xmm0[0],xmm1[0]
2307 ; AVX-NEXT: vunpcklpd {{.*#+}} xmm0 = xmm0[0],xmm1[0]
2317 ; SSE-NEXT: unpcklpd {{.*#+}} xmm0 = xmm0[0],xmm1[0]
2322 ; AVX-NEXT: vunpcklpd {{.*#+}} xmm0 = xmm0[0],xmm1[0]
2332 ; SSE-NEXT: unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
2333 ; SSE-NEXT: movapd %xmm1, %xmm0
2338 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
2348 ; SSE2-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]
2349 ; SSE2-NEXT: movapd %xmm1, %xmm0
2354 ; SSSE3-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]
2355 ; SSSE3-NEXT: movapd %xmm1, %xmm0
2360 ; SSE41-NEXT: blendpd {{.*#+}} xmm0 = xmm0[0],xmm1[1]
2365 ; AVX-NEXT: vblendpd {{.*#+}} xmm0 = xmm0[0],xmm1[1]
2388 ; SSE2-NEXT: movlhps {{.*#+}} xmm0 = xmm0[0,0]
2393 ; SSSE3-NEXT: movddup {{.*#+}} xmm0 = xmm0[0,0]
2398 ; SSE41-NEXT: movddup {{.*#+}} xmm0 = xmm0[0,0]
2403 ; AVX-NEXT: vmovddup {{.*#+}} xmm0 = xmm0[0,0]
2413 ; SSE2-NEXT: movlhps {{.*#+}} xmm0 = xmm0[0,0]
2418 ; SSSE3-NEXT: movddup {{.*#+}} xmm0 = xmm0[0,0]
2423 ; SSE41-NEXT: movddup {{.*#+}} xmm0 = xmm0[0,0]
2428 ; AVX-NEXT: vmovddup {{.*#+}} xmm0 = xmm0[0,0]
2438 ; SSE-NEXT: movhlps {{.*#+}} xmm0 = xmm0[1,1]
2443 ; AVX-NEXT: vpermilpd {{.*#+}} xmm0 = xmm0[1,1]
2462 ; SSE2-NEXT: movsd {{.*#+}} xmm0 = xmm1[0],xmm0[1]
2467 ; SSSE3-NEXT: movsd {{.*#+}} xmm0 = xmm1[0],xmm0[1]
2472 ; SSE41-NEXT: blendpd {{.*#+}} xmm0 = xmm1[0],xmm0[1]
2477 ; AVX-NEXT: vblendpd {{.*#+}} xmm0 = xmm1[0],xmm0[1]
2487 ; SSE-NEXT: unpcklpd {{.*#+}} xmm0 = xmm0[0],xmm1[0]
2492 ; AVX-NEXT: vunpcklpd {{.*#+}} xmm0 = xmm0[0],xmm1[0]
2502 ; SSE-NEXT: unpcklpd {{.*#+}} xmm0 = xmm0[0],xmm1[0]
2507 ; AVX-NEXT: vunpcklpd {{.*#+}} xmm0 = xmm0[0],xmm1[0]
2517 ; SSE-NEXT: unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm0[1]
2518 ; SSE-NEXT: movapd %xmm1, %xmm0
2523 ; AVX-NEXT: vunpckhpd {{.*#+}} xmm0 = xmm1[1],xmm0[1]
2533 ; SSE2-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]
2534 ; SSE2-NEXT: movapd %xmm1, %xmm0
2539 ; SSSE3-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]
2540 ; SSSE3-NEXT: movapd %xmm1, %xmm0
2545 ; SSE41-NEXT: blendpd {{.*#+}} xmm0 = xmm0[0],xmm1[1]
2550 ; AVX-NEXT: vblendpd {{.*#+}} xmm0 = xmm0[0],xmm1[1]
2579 ; SSE2-NEXT: movlhps {{.*#+}} xmm0 = xmm0[0,0]
2584 ; SSSE3-NEXT: movddup {{.*#+}} xmm0 = xmm0[0,0]
2589 ; SSE41-NEXT: movddup {{.*#+}} xmm0 = xmm0[0,0]
2594 ; AVX-NEXT: vmovddup {{.*#+}} xmm0 = xmm0[0,0]
2604 ; SSE2-NEXT: movlhps {{.*#+}} xmm0 = xmm0[0,0]
2609 ; SSSE3-NEXT: movddup {{.*#+}} xmm0 = xmm0[0,0]
2614 ; SSE41-NEXT: movddup {{.*#+}} xmm0 = xmm0[0,0]
2619 ; AVX-NEXT: vmovddup {{.*#+}} xmm0 = xmm0[0,0]
2629 ; SSE-NEXT: movhlps {{.*#+}} xmm0 = xmm0[1,1]
2634 ; AVX-NEXT: vpermilpd {{.*#+}} xmm0 = xmm0[1,1]
2659 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm1[3,2,1,0]
2660 ; SSE-NEXT: movdqa %xmm0, %xmm1
2665 ; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm0
2666 ; AVX1-NEXT: vpaddd {{.*}}(%rip), %xmm0, %xmm0
2667 ; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0
2687 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm3[3,2,1,0]
2693 ; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm0
2694 ; AVX1-NEXT: vpaddd {{.*}}(%rip), %xmm0, %xmm0
2695 ; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm0
2714 ; SSE2-NEXT: shufps {{.*#+}} xmm1 = xmm1[2,0],xmm0[1,0]
2715 ; SSE2-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,2],xmm0[2,3]
2716 ; SSE2-NEXT: movaps %xmm1, %xmm0
2721 ; SSSE3-NEXT: shufps {{.*#+}} xmm1 = xmm1[2,0],xmm0[1,0]
2722 ; SSSE3-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,2],xmm0[2,3]
2723 ; SSSE3-NEXT: movaps %xmm1, %xmm0
2728 ; SSE41-NEXT: insertps {{.*#+}} xmm0 = xmm1[2],xmm0[1,2,3]
2733 ; AVX-NEXT: vinsertps {{.*#+}} xmm0 = xmm1[2],xmm0[1,2,3]
2744 ; SSE2-NEXT: shufps {{.*#+}} xmm1 = xmm1[2,0],xmm0[0,0]
2745 ; SSE2-NEXT: shufps {{.*#+}} xmm1 = xmm1[2,0],xmm0[2,3]
2746 ; SSE2-NEXT: movaps %xmm1, %xmm0
2751 ; SSSE3-NEXT: shufps {{.*#+}} xmm1 = xmm1[2,0],xmm0[0,0]
2752 ; SSSE3-NEXT: shufps {{.*#+}} xmm1 = xmm1[2,0],xmm0[2,3]
2753 ; SSSE3-NEXT: movaps %xmm1, %xmm0
2758 ; SSE41-NEXT: insertps {{.*#+}} xmm0 = xmm0[0],xmm1[2],xmm0[2,3]
2763 ; AVX-NEXT: vinsertps {{.*#+}} xmm0 = xmm0[0],xmm1[2],xmm0[2,3]
2774 ; SSE2-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,0],xmm0[3,0]
2775 ; SSE2-NEXT: shufps {{.*#+}} xmm0 = xmm0[0,1],xmm1[0,2]
2780 ; SSSE3-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,0],xmm0[3,0]
2781 ; SSSE3-NEXT: shufps {{.*#+}} xmm0 = xmm0[0,1],xmm1[0,2]
2786 ; SSE41-NEXT: insertps {{.*#+}} xmm0 = xmm0[0,1],xmm1[0],xmm0[3]
2791 ; AVX-NEXT: vinsertps {{.*#+}} xmm0 = xmm0[0,1],xmm1[0],xmm0[3]
2802 ; SSE2-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,0],xmm0[2,0]
2803 ; SSE2-NEXT: shufps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,0]
2808 ; SSSE3-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,0],xmm0[2,0]
2809 ; SSSE3-NEXT: shufps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,0]
2814 ; SSE41-NEXT: insertps {{.*#+}} xmm0 = xmm0[0,1,2],xmm1[0]
2819 ; AVX-NEXT: vinsertps {{.*#+}} xmm0 = xmm0[0,1,2],xmm1[0]
2831 ; SSE2-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero
2833 ; SSE2-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,0],xmm0[3,0]
2834 ; SSE2-NEXT: shufps {{.*#+}} xmm0 = xmm0[0,1],xmm1[0,2]
2835 ; SSE2-NEXT: movaps %xmm0, (%rsi)
2840 ; SSSE3-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero
2842 ; SSSE3-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,0],xmm0[3,0]
2843 ; SSSE3-NEXT: shufps {{.*#+}} xmm0 = xmm0[0,1],xmm1[0,2]
2844 ; SSSE3-NEXT: movaps %xmm0, (%rsi)
2849 ; SSE41-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero
2851 ; SSE41-NEXT: blendpd {{.*#+}} xmm1 = xmm0[0],xmm1[1]
2857 ; AVX-NEXT: vmovsd {{.*#+}} xmm0 = mem[0],zero
2859 ; AVX-NEXT: vblendpd {{.*#+}} xmm0 = xmm0[0],xmm1[1]
2860 ; AVX-NEXT: vmovapd %xmm0, (%rsi)
2874 ; SSE-NEXT: movaps %xmm0, %xmm1
2876 ; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[0,2,0,2]
2877 ; SSE-NEXT: addps %xmm0, %xmm1
2878 ; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
2883 ; AVX-NEXT: vpermilps {{.*#+}} xmm1 = xmm0[1,3,1,3]
2884 ; AVX-NEXT: vpermilps {{.*#+}} xmm0 = xmm0[0,2,0,2]
2885 ; AVX-NEXT: vaddps %xmm0, %xmm1, %xmm1
2886 ; AVX-NEXT: vunpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
2899 ; SSE2-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,0,1,2]
2900 ; SSE2-NEXT: movaps %xmm0, %xmm2
2902 ; SSE2-NEXT: addps %xmm0, %xmm2
2903 ; SSE2-NEXT: movaps %xmm2, %xmm0
2908 ; SSSE3-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,0,1,2]
2909 ; SSSE3-NEXT: movaps %xmm0, %xmm2
2911 ; SSSE3-NEXT: addps %xmm0, %xmm2
2912 ; SSSE3-NEXT: movaps %xmm2, %xmm0
2917 ; SSE41-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,0,1,2]
2918 ; SSE41-NEXT: blendps {{.*#+}} xmm1 = xmm1[0],xmm0[1,2,3]
2919 ; SSE41-NEXT: addps %xmm1, %xmm0
2924 ; AVX-NEXT: vpermilps {{.*#+}} xmm0 = xmm0[3,0,1,2]
2925 ; AVX-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0],xmm0[1,2,3]
2926 ; AVX-NEXT: vaddps %xmm1, %xmm0, %xmm0
2938 ; SSE2-NEXT: movsd {{.*#+}} xmm2 = xmm0[0],xmm2[1]
2939 ; SSE2-NEXT: movapd %xmm2, %xmm0
2940 ; SSE2-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,0],xmm3[3,2]
2947 ; SSSE3-NEXT: movsd {{.*#+}} xmm2 = xmm0[0],xmm2[1]
2948 ; SSSE3-NEXT: movapd %xmm2, %xmm0
2949 ; SSSE3-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,0],xmm3[3,2]
2956 ; SSE41-NEXT: blendpd {{.*#+}} xmm0 = xmm0[0],xmm2[1]
2957 ; SSE41-NEXT: movapd %xmm0, %xmm1
2959 ; SSE41-NEXT: shufps {{.*#+}} xmm3 = xmm3[1,0],xmm0[3,2]
2960 ; SSE41-NEXT: movaps %xmm1, %xmm0