Lines Matching refs:pshufd

27 ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
29 ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
31 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm4[1,1,3,3]
40 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm2[0,0,2,2]
42 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
44 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[1,1,3,3]
49 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[0,2,2,3]
60 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
62 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
64 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm4[1,1,3,3]
73 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm2[0,0,2,2]
75 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
77 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[1,1,3,3]
82 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[0,2,2,3]
95 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[0,0,2,2]
105 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[0,0,2,2]
109 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[0,2,2,3]
170 ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
172 ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
174 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm4[1,1,3,3]
183 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm2[0,0,2,2]
185 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
187 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[1,1,3,3]
192 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[0,2,2,3]
204 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
206 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
208 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm4[1,1,3,3]
217 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm2[0,0,2,2]
219 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
221 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[1,1,3,3]
226 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[0,2,2,3]
240 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[0,0,2,2]
250 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[0,0,2,2]
254 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[0,2,2,3]
319 ; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
321 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[1,1,3,3]
323 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm6[1,1,3,3]
332 ; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
334 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
336 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
347 ; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
349 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
351 ; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3]
359 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm0[0,0,2,2]
361 ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
363 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
380 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
382 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[1,1,3,3]
384 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm6[1,1,3,3]
393 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
395 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
397 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
408 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
410 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
412 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3]
420 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm0[0,0,2,2]
422 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
424 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
443 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm7[0,0,2,2]
453 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[0,0,2,2]
464 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[0,0,2,2]
473 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[0,0,2,2]
581 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm6[0,0,2,2]
583 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm2[1,1,3,3]
585 ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm6[1,1,3,3]
594 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2]
596 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
598 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
607 ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
609 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
611 ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[1,1,3,3]
620 ; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
622 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
624 ; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm4[1,1,3,3]
635 ; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
637 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
639 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
648 ; SSE2-NEXT:    pshufd {{.*#+}} xmm10 = xmm6[0,0,2,2]
650 ; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm1[1,1,3,3]
652 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm6[1,1,3,3]
662 ; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm5[0,0,2,2]
664 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
666 ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3]
674 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2]
676 ; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm0[1,1,3,3]
678 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
699 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm6[0,0,2,2]
701 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm2[1,1,3,3]
703 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm6[1,1,3,3]
712 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2]
714 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
716 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
725 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
727 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
729 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[1,1,3,3]
738 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
740 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
742 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm4[1,1,3,3]
753 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
755 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
757 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
766 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm10 = xmm6[0,0,2,2]
768 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm1[1,1,3,3]
770 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm6[1,1,3,3]
780 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm5[0,0,2,2]
782 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
784 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3]
792 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2]
794 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm0[1,1,3,3]
796 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
819 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[0,0,2,2]
830 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[0,0,2,2]
841 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[0,0,2,2]
851 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[0,0,2,2]
862 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm7[0,0,2,2]
872 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[0,0,2,2]
883 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[0,0,2,2]
892 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[0,0,2,2]
1004 ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
1006 ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
1008 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm4[1,1,3,3]
1017 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm2[0,0,2,2]
1019 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
1021 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[1,1,3,3]
1026 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[0,2,2,3]
1038 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
1040 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
1042 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm4[1,1,3,3]
1051 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm2[0,0,2,2]
1053 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
1055 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[1,1,3,3]
1060 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[0,2,2,3]
1074 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[0,0,2,2]
1084 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[0,0,2,2]
1088 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[0,2,2,3]
1172 ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
1174 ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
1176 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm4[1,1,3,3]
1185 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm2[0,0,2,2]
1187 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
1189 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[1,1,3,3]
1194 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[0,2,2,3]
1207 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
1209 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
1211 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm4[1,1,3,3]
1220 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm2[0,0,2,2]
1222 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
1224 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[1,1,3,3]
1229 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[0,2,2,3]
1244 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[0,0,2,2]
1254 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[0,0,2,2]
1258 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[0,2,2,3]
1350 ; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
1352 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[1,1,3,3]
1354 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm6[1,1,3,3]
1363 ; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
1365 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
1367 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
1378 ; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
1380 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
1382 ; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3]
1390 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm0[0,0,2,2]
1392 ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
1394 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
1412 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
1414 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[1,1,3,3]
1416 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm6[1,1,3,3]
1425 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
1427 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
1429 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
1440 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
1442 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
1444 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3]
1452 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm0[0,0,2,2]
1454 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
1456 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
1476 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm7[0,0,2,2]
1486 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[0,0,2,2]
1497 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[0,0,2,2]
1506 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[0,0,2,2]
1596 ; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
1598 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[1,1,3,3]
1600 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm6[1,1,3,3]
1609 ; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
1611 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
1613 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
1624 ; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
1626 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
1628 ; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3]
1636 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm1[0,0,2,2]
1638 ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
1640 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
1659 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
1661 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[1,1,3,3]
1663 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm6[1,1,3,3]
1672 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
1674 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
1676 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
1687 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
1689 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
1691 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3]
1699 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm1[0,0,2,2]
1701 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
1703 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
1724 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm7[0,0,2,2]
1734 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[0,0,2,2]
1745 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[0,0,2,2]
1754 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[0,0,2,2]
1853 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm7[0,0,2,2]
1855 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm2[1,1,3,3]
1857 ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm7[1,1,3,3]
1866 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2]
1868 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
1870 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
1879 ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
1881 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
1883 ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[1,1,3,3]
1892 ; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
1894 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
1896 ; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm4[1,1,3,3]
1907 ; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
1909 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
1911 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
1920 ; SSE2-NEXT:    pshufd {{.*#+}} xmm10 = xmm6[0,0,2,2]
1922 ; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm0[1,1,3,3]
1924 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[1,1,3,3]
1934 ; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm5[0,0,2,2]
1936 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
1938 ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3]
1946 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2]
1948 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
1950 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
1972 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm7[0,0,2,2]
1974 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm2[1,1,3,3]
1976 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm7[1,1,3,3]
1985 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2]
1987 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
1989 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
1998 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
2000 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
2002 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[1,1,3,3]
2011 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
2013 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
2015 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm4[1,1,3,3]
2026 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
2028 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
2030 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
2039 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm10 = xmm6[0,0,2,2]
2041 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm0[1,1,3,3]
2043 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[1,1,3,3]
2053 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm5[0,0,2,2]
2055 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
2057 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3]
2065 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2]
2067 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
2069 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
2093 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[0,0,2,2]
2104 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[0,0,2,2]
2115 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm5[0,0,2,2]
2125 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[0,0,2,2]
2136 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[0,0,2,2]
2146 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[0,0,2,2]
2157 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[0,0,2,2]
2166 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[0,0,2,2]
2444 ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
2446 ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
2448 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm4[1,1,3,3]
2457 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm2[0,0,2,2]
2459 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
2461 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,3,3]
2480 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
2482 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
2484 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm4[1,1,3,3]
2493 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm2[0,0,2,2]
2495 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
2497 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,3,3]
2515 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[0,0,2,2]
2525 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[0,0,2,2]
2589 ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
2591 ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
2593 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm4[1,1,3,3]
2602 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm2[0,0,2,2]
2604 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
2606 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[1,1,3,3]
2627 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
2629 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
2631 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm4[1,1,3,3]
2640 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm2[0,0,2,2]
2642 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
2644 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[1,1,3,3]
2664 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[0,0,2,2]
2674 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[0,0,2,2]
2743 ; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
2745 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[1,1,3,3]
2747 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm6[1,1,3,3]
2756 ; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
2758 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
2760 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
2771 ; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
2773 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,3,3]
2775 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[1,1,3,3]
2783 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm1[0,0,2,2]
2785 ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
2787 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
2809 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
2811 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[1,1,3,3]
2813 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm6[1,1,3,3]
2822 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
2824 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
2826 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
2837 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
2839 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,3,3]
2841 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[1,1,3,3]
2849 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm1[0,0,2,2]
2851 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
2853 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
2876 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm7[0,0,2,2]
2886 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[0,0,2,2]
2897 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[0,0,2,2]
2906 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[0,0,2,2]
3002 ; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
3004 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[1,1,3,3]
3006 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm6[1,1,3,3]
3015 ; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
3017 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
3019 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
3030 ; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
3032 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,3,3]
3034 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[1,1,3,3]
3042 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm1[0,0,2,2]
3044 ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
3046 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
3069 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
3071 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[1,1,3,3]
3073 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm6[1,1,3,3]
3082 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
3084 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
3086 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
3097 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
3099 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
3101 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm6[1,1,3,3]
3109 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm0[0,0,2,2]
3111 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
3113 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
3137 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm7[0,0,2,2]
3147 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[0,0,2,2]
3158 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[0,0,2,2]
3167 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[0,0,2,2]
3272 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm7[0,0,2,2]
3274 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm2[1,1,3,3]
3276 ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm7[1,1,3,3]
3285 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2]
3287 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
3289 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
3298 ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
3300 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
3302 ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[1,1,3,3]
3311 ; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
3313 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
3315 ; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm4[1,1,3,3]
3326 ; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
3328 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
3330 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
3339 ; SSE2-NEXT:    pshufd {{.*#+}} xmm10 = xmm6[0,0,2,2]
3341 ; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm0[1,1,3,3]
3343 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[1,1,3,3]
3353 ; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm5[0,0,2,2]
3355 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
3357 ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3]
3365 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2]
3367 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
3369 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
3392 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm7[0,0,2,2]
3394 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm2[1,1,3,3]
3396 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm7[1,1,3,3]
3405 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2]
3407 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
3409 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
3418 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
3420 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
3422 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[1,1,3,3]
3431 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
3433 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
3435 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm4[1,1,3,3]
3446 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
3448 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
3450 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
3459 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm10 = xmm6[0,0,2,2]
3461 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm0[1,1,3,3]
3463 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[1,1,3,3]
3473 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm5[0,0,2,2]
3475 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
3477 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3]
3485 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2]
3487 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
3489 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
3514 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[0,0,2,2]
3525 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[0,0,2,2]
3536 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm5[0,0,2,2]
3546 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[0,0,2,2]
3557 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[0,0,2,2]
3567 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[0,0,2,2]
3578 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[0,0,2,2]
3587 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[0,0,2,2]
3690 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm7[0,0,2,2]
3692 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm1[1,1,3,3]
3694 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm7[1,1,3,3]
3703 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2]
3705 ; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm2[1,1,3,3]
3707 ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[1,1,3,3]
3716 ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
3718 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
3720 ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[1,1,3,3]
3729 ; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
3731 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
3733 ; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm4[1,1,3,3]
3744 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2]
3746 ; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm7[1,1,3,3]
3748 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
3757 ; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
3759 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
3761 ; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3]
3771 ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
3773 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
3775 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
3783 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2]
3785 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
3787 ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
3811 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm7[0,0,2,2]
3813 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm1[1,1,3,3]
3815 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm7[1,1,3,3]
3824 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2]
3826 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm2[1,1,3,3]
3828 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[1,1,3,3]
3837 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
3839 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
3841 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[1,1,3,3]
3850 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
3852 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
3854 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm4[1,1,3,3]
3865 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2]
3867 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm7[1,1,3,3]
3869 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
3878 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
3880 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
3882 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3]
3892 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
3894 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
3896 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
3904 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2]
3906 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
3908 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
3934 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[0,0,2,2]
3945 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[0,0,2,2]
3956 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm5[0,0,2,2]
3966 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[0,0,2,2]
3977 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[0,0,2,2]
3987 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[0,0,2,2]
3998 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[0,0,2,2]
4007 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[0,0,2,2]
4116 ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
4118 ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3]
4120 ; SSE2-NEXT:    pshufd {{.*#+}} xmm11 = xmm0[1,1,3,3]
4129 ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm2[0,0,2,2]
4131 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
4133 ; SSE2-NEXT:    pshufd {{.*#+}} xmm12 = xmm2[1,1,3,3]
4142 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm2[0,0,2,2]
4144 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
4146 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm2[1,1,3,3]
4155 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2]
4157 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
4159 ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm2[1,1,3,3]
4168 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2]
4170 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
4172 ; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm2[1,1,3,3]
4181 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2]
4183 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
4185 ; SSE2-NEXT:    pshufd {{.*#+}} xmm15 = xmm2[1,1,3,3]
4194 ; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm3[0,0,2,2]
4196 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
4198 ; SSE2-NEXT:    pshufd {{.*#+}} xmm13 = xmm3[1,1,3,3]
4207 ; SSE2-NEXT:    pshufd {{.*#+}} xmm10 = xmm7[0,0,2,2]
4209 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
4211 ; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm7[1,1,3,3]
4222 ; SSE2-NEXT:    pshufd {{.*#+}} xmm10 = xmm2[0,0,2,2]
4224 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
4226 ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
4235 ; SSE2-NEXT:    pshufd {{.*#+}} xmm10 = xmm7[0,0,2,2]
4237 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm0[1,1,3,3]
4239 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm7[1,1,3,3]
4249 ; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm3[0,0,2,2]
4251 ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
4253 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
4262 ; SSE2-NEXT:    pshufd {{.*#+}} xmm10 = xmm7[0,0,2,2]
4264 ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
4266 ; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm7[1,1,3,3]
4277 ; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm3[0,0,2,2]
4279 ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
4281 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
4290 ; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm5[0,0,2,2]
4292 ; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm2[1,1,3,3]
4294 ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm5[1,1,3,3]
4304 ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
4306 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
4308 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
4316 ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm3[0,0,2,2]
4318 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
4320 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
4347 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
4349 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3]
4351 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm11 = xmm0[1,1,3,3]
4360 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm2[0,0,2,2]
4362 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
4364 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm12 = xmm2[1,1,3,3]
4373 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm2[0,0,2,2]
4375 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
4377 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm2[1,1,3,3]
4386 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2]
4388 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
4390 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm2[1,1,3,3]
4399 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2]
4401 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
4403 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm2[1,1,3,3]
4412 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2]
4414 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
4416 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm15 = xmm2[1,1,3,3]
4425 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm3[0,0,2,2]
4427 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
4429 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm13 = xmm3[1,1,3,3]
4438 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm10 = xmm7[0,0,2,2]
4440 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
4442 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm7[1,1,3,3]
4453 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm10 = xmm2[0,0,2,2]
4455 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
4457 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
4466 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm10 = xmm7[0,0,2,2]
4468 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm0[1,1,3,3]
4470 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm7[1,1,3,3]
4480 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm3[0,0,2,2]
4482 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
4484 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
4493 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm10 = xmm7[0,0,2,2]
4495 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
4497 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm7[1,1,3,3]
4508 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm3[0,0,2,2]
4510 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
4512 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
4521 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm5[0,0,2,2]
4523 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm2[1,1,3,3]
4525 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm5[1,1,3,3]
4535 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
4537 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
4539 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
4547 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm3[0,0,2,2]
4549 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
4551 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
4580 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[0,0,2,2]
4591 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm5[0,0,2,2]
4602 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm5[0,0,2,2]
4613 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm5[0,0,2,2]
4624 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[0,0,2,2]
4635 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm5[0,0,2,2]
4646 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[0,0,2,2]
4656 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm7[0,0,2,2]
4667 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm5[0,0,2,2]
4677 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[0,0,2,2]
4688 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[0,0,2,2]
4698 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm5[0,0,2,2]
4710 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[0,0,2,2]
4720 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[0,0,2,2]
4731 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[0,0,2,2]
4740 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[0,0,2,2]