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 = xmm0[1,1,3,3]
44 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[1,1,3,3]
47 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[0,2,2,3]
58 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
60 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
62 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm4[1,1,3,3]
71 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm2[0,0,2,2]
73 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
75 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[1,1,3,3]
78 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[0,2,2,3]
91 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[0,0,2,2]
101 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[0,0,2,2]
105 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[0,2,2,3]
176 ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
178 ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
180 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm4[1,1,3,3]
189 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm2[0,0,2,2]
191 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
193 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[1,1,3,3]
196 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[0,2,2,3]
208 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
210 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
212 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm4[1,1,3,3]
221 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm2[0,0,2,2]
223 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
225 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[1,1,3,3]
228 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[0,2,2,3]
242 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[0,0,2,2]
252 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[0,0,2,2]
256 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[0,2,2,3]
331 ; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
333 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[1,1,3,3]
335 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm6[1,1,3,3]
344 ; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
346 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
348 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
357 ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm1[0,0,2,2]
359 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
361 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
368 ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
370 ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[1,1,3,3]
372 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[1,1,3,3]
387 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
389 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[1,1,3,3]
391 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm6[1,1,3,3]
400 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
402 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
404 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
413 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm1[0,0,2,2]
415 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
417 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
424 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
426 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[1,1,3,3]
428 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[1,1,3,3]
445 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm7[0,0,2,2]
455 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[0,0,2,2]
465 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[0,0,2,2]
475 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[0,0,2,2]
593 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm5[0,0,2,2]
595 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm2[1,1,3,3]
597 ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm5[1,1,3,3]
606 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2]
608 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
610 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
619 ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
621 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
623 ; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm4[1,1,3,3]
632 ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
634 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
636 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
645 ; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm5[0,0,2,2]
647 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
649 ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3]
656 ; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
658 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
660 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm4[1,1,3,3]
668 ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
670 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
672 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
679 ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm3[0,0,2,2]
681 ; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm0[1,1,3,3]
683 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
702 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm5[0,0,2,2]
704 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm2[1,1,3,3]
706 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm5[1,1,3,3]
715 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2]
717 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
719 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
728 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
730 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
732 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm4[1,1,3,3]
741 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
743 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
745 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
754 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm5[0,0,2,2]
756 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
758 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3]
765 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
767 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
769 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm4[1,1,3,3]
777 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
779 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
781 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
788 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm3[0,0,2,2]
790 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm0[1,1,3,3]
792 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
813 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[0,0,2,2]
824 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[0,0,2,2]
835 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[0,0,2,2]
845 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[0,0,2,2]
855 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[0,0,2,2]
865 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[0,0,2,2]
876 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[0,0,2,2]
886 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm5[0,0,2,2]
1000 ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
1002 ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
1004 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm4[1,1,3,3]
1013 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm2[0,0,2,2]
1015 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
1017 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[1,1,3,3]
1020 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[0,2,2,3]
1032 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
1034 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
1036 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm4[1,1,3,3]
1045 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm2[0,0,2,2]
1047 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
1049 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[1,1,3,3]
1052 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[0,2,2,3]
1066 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[0,0,2,2]
1076 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[0,0,2,2]
1080 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[0,2,2,3]
1174 ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
1176 ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
1178 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm4[1,1,3,3]
1187 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm2[0,0,2,2]
1189 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
1191 ; 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 = xmm0[1,1,3,3]
1224 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[1,1,3,3]
1227 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[0,2,2,3]
1242 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[0,0,2,2]
1252 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[0,0,2,2]
1256 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[0,2,2,3]
1358 ; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
1360 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[1,1,3,3]
1362 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm6[1,1,3,3]
1371 ; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
1373 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
1375 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
1384 ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm1[0,0,2,2]
1386 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
1388 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
1395 ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
1397 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
1399 ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm4[1,1,3,3]
1402 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[0,2,2,3]
1404 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[0,2,2,3]
1418 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
1420 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[1,1,3,3]
1422 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm6[1,1,3,3]
1431 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
1433 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
1435 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
1444 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm1[0,0,2,2]
1446 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
1448 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
1455 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
1457 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
1459 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm4[1,1,3,3]
1462 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[0,2,2,3]
1464 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[0,2,2,3]
1480 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm7[0,0,2,2]
1490 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[0,0,2,2]
1500 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[0,0,2,2]
1510 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[0,0,2,2]
1610 ; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
1612 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[1,1,3,3]
1614 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm6[1,1,3,3]
1623 ; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
1625 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
1627 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
1636 ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm1[0,0,2,2]
1638 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
1640 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
1647 ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
1649 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
1651 ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm4[1,1,3,3]
1654 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[0,2,2,3]
1656 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[0,2,2,3]
1671 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
1673 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[1,1,3,3]
1675 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm6[1,1,3,3]
1684 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
1686 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
1688 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
1697 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm1[0,0,2,2]
1699 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
1701 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
1708 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
1710 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
1712 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm4[1,1,3,3]
1715 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[0,2,2,3]
1717 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[0,2,2,3]
1734 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm7[0,0,2,2]
1744 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[0,0,2,2]
1754 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[0,0,2,2]
1764 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[0,0,2,2]
1873 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm5[0,0,2,2]
1875 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm1[1,1,3,3]
1877 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm5[1,1,3,3]
1886 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2]
1888 ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm2[1,1,3,3]
1890 ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[1,1,3,3]
1899 ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
1901 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
1903 ; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm4[1,1,3,3]
1912 ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
1914 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
1916 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
1925 ; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm5[0,0,2,2]
1927 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
1929 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm5[1,1,3,3]
1936 ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
1938 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
1940 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
1947 ; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm5[0,0,2,2]
1949 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
1951 ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3]
1958 ; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm2[0,0,2,2]
1960 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
1962 ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
1965 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[0,2,2,3]
1967 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm5[0,2,2,3]
1970 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[0,2,2,3]
1972 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[0,2,2,3]
1991 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm5[0,0,2,2]
1993 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm1[1,1,3,3]
1995 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm5[1,1,3,3]
2004 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2]
2006 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm2[1,1,3,3]
2008 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[1,1,3,3]
2017 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
2019 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
2021 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm4[1,1,3,3]
2030 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
2032 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
2034 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
2043 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm5[0,0,2,2]
2045 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
2047 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm5[1,1,3,3]
2054 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
2056 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
2058 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
2065 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm5[0,0,2,2]
2067 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
2069 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3]
2076 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm2[0,0,2,2]
2078 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
2080 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
2083 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[0,2,2,3]
2085 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm5[0,2,2,3]
2088 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[0,2,2,3]
2090 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[0,2,2,3]
2111 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[0,0,2,2]
2122 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[0,0,2,2]
2133 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm5[0,0,2,2]
2143 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[0,0,2,2]
2153 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[0,0,2,2]
2163 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[0,0,2,2]
2174 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[0,0,2,2]
2184 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[0,0,2,2]
2285 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
2345 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
2700 ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
2702 ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
2704 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm4[1,1,3,3]
2713 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm2[0,0,2,2]
2715 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,3,3]
2717 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,3,3]
2734 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
2736 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
2738 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm4[1,1,3,3]
2747 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm2[0,0,2,2]
2749 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,3,3]
2751 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,3,3]
2767 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[0,0,2,2]
2777 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[0,0,2,2]
2851 ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
2853 ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
2855 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm4[1,1,3,3]
2864 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm2[0,0,2,2]
2866 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
2868 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[1,1,3,3]
2887 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
2889 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
2891 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm4[1,1,3,3]
2900 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm2[0,0,2,2]
2902 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
2904 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[1,1,3,3]
2922 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[0,0,2,2]
2932 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[0,0,2,2]
3011 ; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
3013 ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm4[1,1,3,3]
3015 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm6[1,1,3,3]
3024 ; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm2[0,0,2,2]
3026 ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm1[1,1,3,3]
3028 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[1,1,3,3]
3037 ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm2[0,0,2,2]
3039 ; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm0[1,1,3,3]
3041 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,3,3]
3047 ; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm5[0,0,2,2]
3049 ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
3051 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm5[1,1,3,3]
3071 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
3073 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[1,1,3,3]
3075 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm6[1,1,3,3]
3084 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
3086 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
3088 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
3097 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm1[0,0,2,2]
3099 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm0[1,1,3,3]
3101 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
3108 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
3110 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
3112 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm4[1,1,3,3]
3133 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm7[0,0,2,2]
3143 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[0,0,2,2]
3153 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[0,0,2,2]
3163 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[0,0,2,2]
3269 ; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
3271 ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm4[1,1,3,3]
3273 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm6[1,1,3,3]
3282 ; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm2[0,0,2,2]
3284 ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm1[1,1,3,3]
3286 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[1,1,3,3]
3295 ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm2[0,0,2,2]
3297 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
3299 ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
3305 ; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm5[0,0,2,2]
3307 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
3309 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm5[1,1,3,3]
3330 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
3332 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[1,1,3,3]
3334 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm6[1,1,3,3]
3343 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
3345 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
3347 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
3356 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm1[0,0,2,2]
3358 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
3360 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
3367 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
3369 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
3371 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm4[1,1,3,3]
3393 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm7[0,0,2,2]
3403 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[0,0,2,2]
3413 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[0,0,2,2]
3423 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[0,0,2,2]
3538 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[0,0,2,2]
3540 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm2[1,1,3,3]
3542 ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm6[1,1,3,3]
3551 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2]
3553 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
3555 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
3564 ; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
3566 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
3568 ; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm4[1,1,3,3]
3577 ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
3579 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
3581 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
3590 ; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm5[0,0,2,2]
3592 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
3594 ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3]
3601 ; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
3603 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,3,3]
3605 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[1,1,3,3]
3613 ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
3615 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
3617 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
3624 ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm3[0,0,2,2]
3626 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
3628 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
3649 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[0,0,2,2]
3651 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm2[1,1,3,3]
3653 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm6[1,1,3,3]
3662 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2]
3664 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
3666 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
3675 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
3677 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
3679 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm4[1,1,3,3]
3688 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
3690 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
3692 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
3701 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm5[0,0,2,2]
3703 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
3705 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3]
3712 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
3714 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,3,3]
3716 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[1,1,3,3]
3724 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
3726 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
3728 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
3735 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm3[0,0,2,2]
3737 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
3739 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
3762 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[0,0,2,2]
3773 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[0,0,2,2]
3784 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm5[0,0,2,2]
3794 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[0,0,2,2]
3804 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[0,0,2,2]
3814 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[0,0,2,2]
3825 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[0,0,2,2]
3835 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[0,0,2,2]
3938 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm6[0,0,2,2]
3940 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm1[1,1,3,3]
3942 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm6[1,1,3,3]
3951 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2]
3953 ; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm2[1,1,3,3]
3955 ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[1,1,3,3]
3964 ; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
3966 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
3968 ; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm4[1,1,3,3]
3977 ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
3979 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
3981 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
3990 ; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm5[0,0,2,2]
3992 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
3994 ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3]
4001 ; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
4003 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
4005 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm4[1,1,3,3]
4013 ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
4015 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
4017 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
4024 ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm2[0,0,2,2]
4026 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
4028 ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
4050 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm6[0,0,2,2]
4052 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm1[1,1,3,3]
4054 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm6[1,1,3,3]
4063 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2]
4065 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm2[1,1,3,3]
4067 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[1,1,3,3]
4076 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
4078 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
4080 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm4[1,1,3,3]
4089 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
4091 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
4093 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
4102 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm5[0,0,2,2]
4104 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
4106 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3]
4113 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
4115 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
4117 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm4[1,1,3,3]
4125 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
4127 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
4129 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
4136 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm2[0,0,2,2]
4138 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
4140 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
4164 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[0,0,2,2]
4175 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[0,0,2,2]
4186 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm5[0,0,2,2]
4196 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[0,0,2,2]
4206 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[0,0,2,2]
4216 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[0,0,2,2]
4227 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[0,0,2,2]
4237 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[0,0,2,2]
4348 ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm6[0,0,2,2]
4350 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
4352 ; SSE2-NEXT:    pshufd {{.*#+}} xmm11 = xmm6[1,1,3,3]
4361 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm2[0,0,2,2]
4363 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
4365 ; SSE2-NEXT:    pshufd {{.*#+}} xmm12 = xmm2[1,1,3,3]
4374 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2]
4376 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
4378 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm2[1,1,3,3]
4387 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2]
4389 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
4391 ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm2[1,1,3,3]
4400 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2]
4402 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
4404 ; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm2[1,1,3,3]
4413 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2]
4415 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
4417 ; SSE2-NEXT:    pshufd {{.*#+}} xmm15 = xmm2[1,1,3,3]
4426 ; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm3[0,0,2,2]
4428 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
4430 ; SSE2-NEXT:    pshufd {{.*#+}} xmm13 = xmm3[1,1,3,3]
4439 ; SSE2-NEXT:    pshufd {{.*#+}} xmm10 = xmm6[0,0,2,2]
4441 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
4443 ; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3]
4452 ; SSE2-NEXT:    pshufd {{.*#+}} xmm8 = xmm2[0,0,2,2]
4454 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
4456 ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
4463 ; SSE2-NEXT:    pshufd {{.*#+}} xmm8 = xmm6[0,0,2,2]
4465 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm0[1,1,3,3]
4467 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[1,1,3,3]
4475 ; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm3[0,0,2,2]
4477 ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
4479 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
4486 ; SSE2-NEXT:    pshufd {{.*#+}} xmm8 = xmm6[0,0,2,2]
4488 ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
4490 ; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3]
4499 ; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm3[0,0,2,2]
4501 ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
4503 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
4510 ; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm5[0,0,2,2]
4512 ; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm2[1,1,3,3]
4514 ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm5[1,1,3,3]
4522 ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
4524 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
4526 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
4533 ; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm5[0,0,2,2]
4535 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
4537 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm5[1,1,3,3]
4562 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm6[0,0,2,2]
4564 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
4566 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm11 = xmm6[1,1,3,3]
4575 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm2[0,0,2,2]
4577 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
4579 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm12 = xmm2[1,1,3,3]
4588 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2]
4590 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
4592 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm2[1,1,3,3]
4601 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2]
4603 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
4605 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm2[1,1,3,3]
4614 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2]
4616 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
4618 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm2[1,1,3,3]
4627 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2]
4629 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
4631 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm15 = xmm2[1,1,3,3]
4640 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm3[0,0,2,2]
4642 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
4644 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm13 = xmm3[1,1,3,3]
4653 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm10 = xmm6[0,0,2,2]
4655 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
4657 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3]
4666 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm8 = xmm2[0,0,2,2]
4668 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
4670 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
4677 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm8 = xmm6[0,0,2,2]
4679 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm0[1,1,3,3]
4681 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[1,1,3,3]
4689 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm3[0,0,2,2]
4691 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
4693 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
4700 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm8 = xmm6[0,0,2,2]
4702 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
4704 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3]
4713 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm3[0,0,2,2]
4715 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
4717 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
4724 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm5[0,0,2,2]
4726 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm2[1,1,3,3]
4728 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm5[1,1,3,3]
4736 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
4738 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
4740 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
4747 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm5[0,0,2,2]
4749 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
4751 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm5[1,1,3,3]
4778 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm5[0,0,2,2]
4789 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[0,0,2,2]
4800 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[0,0,2,2]
4811 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[0,0,2,2]
4822 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[0,0,2,2]
4833 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm5[0,0,2,2]
4844 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[0,0,2,2]
4854 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm7[0,0,2,2]
4864 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[0,0,2,2]
4874 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[0,0,2,2]
4885 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[0,0,2,2]
4895 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[0,0,2,2]
4907 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[0,0,2,2]
4917 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[0,0,2,2]
4928 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[0,0,2,2]
4938 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm5[0,0,2,2]