Lines Matching refs:pshufd

26 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2]
28 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
30 ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[1,1,3,3]
35 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[0,2,2,3]
45 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2]
47 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
49 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[1,1,3,3]
54 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[0,2,2,3]
67 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[0,0,2,2]
71 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[0,2,2,3]
128 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2]
130 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
132 ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[1,1,3,3]
137 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[0,2,2,3]
148 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2]
150 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
152 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[1,1,3,3]
157 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[0,2,2,3]
171 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[0,0,2,2]
175 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[0,2,2,3]
236 ; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm5[0,0,2,2]
238 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
240 ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3]
245 ; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm3[0,0,2,2]
247 ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
249 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
268 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm5[0,0,2,2]
270 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
272 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3]
277 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm3[0,0,2,2]
279 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
281 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
303 ; SSE41-NEXT:    pshufd {{.*#+}} xmm3 = xmm6[0,0,2,2]
310 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[0,0,2,2]
434 ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm4[0,0,2,2]
436 ; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm7[1,1,3,3]
438 ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm4[1,1,3,3]
447 ; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm4[0,0,2,2]
449 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[1,1,3,3]
451 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm4[1,1,3,3]
461 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2]
463 ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
465 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
473 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm2[0,0,2,2]
475 ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm0[1,1,3,3]
477 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,3,3]
498 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm4[0,0,2,2]
500 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm7[1,1,3,3]
502 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm4[1,1,3,3]
511 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm4[0,0,2,2]
513 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[1,1,3,3]
515 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm4[1,1,3,3]
525 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2]
527 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
529 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
537 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm2[0,0,2,2]
539 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm0[1,1,3,3]
541 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,3,3]
564 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[0,0,2,2]
575 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[0,0,2,2]
587 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[0,0,2,2]
596 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[0,0,2,2]
697 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2]
699 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
701 ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[1,1,3,3]
706 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[0,2,2,3]
717 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2]
719 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
721 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[1,1,3,3]
726 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[0,2,2,3]
740 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[0,0,2,2]
744 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[0,2,2,3]
822 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2]
824 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
826 ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[1,1,3,3]
831 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[0,2,2,3]
843 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2]
845 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
847 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[1,1,3,3]
852 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[0,2,2,3]
867 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[0,0,2,2]
871 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[0,2,2,3]
958 ; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
960 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
962 ; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3]
970 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm0[0,0,2,2]
972 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
974 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
979 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
981 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[0,2,2,3]
995 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
997 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
999 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3]
1007 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm0[0,0,2,2]
1009 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
1011 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
1016 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
1018 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[0,2,2,3]
1035 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm7[0,0,2,2]
1044 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm5[0,0,2,2]
1131 ; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
1133 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
1135 ; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3]
1143 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm0[0,0,2,2]
1145 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
1147 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
1152 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
1154 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm6[0,2,2,3]
1169 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
1171 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
1173 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3]
1181 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm0[0,0,2,2]
1183 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
1185 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
1190 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
1192 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm6[0,2,2,3]
1210 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm7[0,0,2,2]
1219 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm5[0,0,2,2]
1315 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm5[0,0,2,2]
1317 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[1,1,3,3]
1319 ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm5[1,1,3,3]
1328 ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm1[0,0,2,2]
1330 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
1332 ; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm1[1,1,3,3]
1341 ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm1[0,0,2,2]
1343 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
1345 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
1353 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm0[0,0,2,2]
1355 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
1357 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
1362 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
1364 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[0,2,2,3]
1367 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[0,2,2,3]
1369 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[0,2,2,3]
1388 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm5[0,0,2,2]
1390 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[1,1,3,3]
1392 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm5[1,1,3,3]
1401 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm1[0,0,2,2]
1403 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
1405 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm1[1,1,3,3]
1414 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm1[0,0,2,2]
1416 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
1418 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
1426 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm0[0,0,2,2]
1428 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
1430 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
1435 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
1437 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[0,2,2,3]
1440 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[0,2,2,3]
1442 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[0,2,2,3]
1463 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[0,0,2,2]
1474 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[0,0,2,2]
1486 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[0,0,2,2]
1495 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[0,0,2,2]
1582 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
1664 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
2014 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2]
2016 ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,3,3]
2018 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
2037 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2]
2039 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
2041 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[1,1,3,3]
2059 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[0,0,2,2]
2119 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2]
2121 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
2123 ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[1,1,3,3]
2143 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[0,0,2,2]
2145 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
2147 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[1,1,3,3]
2167 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[0,0,2,2]
2233 ; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
2235 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm0[1,1,3,3]
2237 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[1,1,3,3]
2245 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2]
2247 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
2249 ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
2270 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
2272 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
2274 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3]
2282 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm4[0,0,2,2]
2284 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
2286 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
2308 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm7[0,0,2,2]
2317 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[0,0,2,2]
2410 ; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
2412 ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[1,1,3,3]
2414 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm6[1,1,3,3]
2422 ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
2424 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
2426 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
2448 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
2450 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[1,1,3,3]
2452 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm6[1,1,3,3]
2460 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]
2462 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
2464 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
2487 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm7[0,0,2,2]
2496 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm5[0,0,2,2]
2598 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm4[0,0,2,2]
2600 ; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm7[1,1,3,3]
2602 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm4[1,1,3,3]
2611 ; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm4[0,0,2,2]
2613 ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[1,1,3,3]
2615 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[1,1,3,3]
2625 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm2[0,0,2,2]
2627 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
2629 ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
2637 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm1[0,0,2,2]
2639 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
2641 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
2664 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm4[0,0,2,2]
2666 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm7[1,1,3,3]
2668 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm4[1,1,3,3]
2677 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm4[0,0,2,2]
2679 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[1,1,3,3]
2681 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[1,1,3,3]
2691 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm2[0,0,2,2]
2693 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
2695 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
2703 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm1[0,0,2,2]
2705 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
2707 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
2732 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[0,0,2,2]
2743 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[0,0,2,2]
2755 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[0,0,2,2]
2764 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[0,0,2,2]
2856 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[0,0,2,2]
2858 ; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm7[1,1,3,3]
2860 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
2869 ; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm5[0,0,2,2]
2871 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
2873 ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3]
2883 ; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm3[0,0,2,2]
2885 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
2887 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
2895 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm0[0,0,2,2]
2897 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[1,1,3,3]
2899 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
2923 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[0,0,2,2]
2925 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm7[1,1,3,3]
2927 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
2936 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm5[0,0,2,2]
2938 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
2940 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3]
2950 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm3[0,0,2,2]
2952 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
2954 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
2962 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm0[0,0,2,2]
2964 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[1,1,3,3]
2966 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
2992 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[0,0,2,2]
3003 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[0,0,2,2]
3015 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm5[0,0,2,2]
3024 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[0,0,2,2]
3123 ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm5[0,0,2,2]
3125 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
3127 ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm5[1,1,3,3]
3136 ; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm5[0,0,2,2]
3138 ; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm0[1,1,3,3]
3140 ; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm5[1,1,3,3]
3150 ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm3[0,0,2,2]
3152 ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
3154 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
3163 ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm2[0,0,2,2]
3165 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
3167 ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
3178 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2]
3180 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
3182 ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
3191 ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm3[0,0,2,2]
3193 ; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm1[1,1,3,3]
3195 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
3205 ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm3[0,0,2,2]
3207 ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
3209 ; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
3217 ; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm2[0,0,2,2]
3219 ; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
3221 ; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
3248 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm5[0,0,2,2]
3250 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
3252 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm5[1,1,3,3]
3261 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm5[0,0,2,2]
3263 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm7 = xmm0[1,1,3,3]
3265 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm5[1,1,3,3]
3275 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm3[0,0,2,2]
3277 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
3279 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
3288 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm2[0,0,2,2]
3290 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
3292 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
3303 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm2[0,0,2,2]
3305 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
3307 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
3316 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm3[0,0,2,2]
3318 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm6 = xmm1[1,1,3,3]
3320 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
3330 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm3[0,0,2,2]
3332 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
3334 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
3342 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm5 = xmm2[0,0,2,2]
3344 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
3346 ; SSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
3375 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm5[0,0,2,2]
3386 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm7[0,0,2,2]
3398 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm5[0,0,2,2]
3409 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm7[0,0,2,2]
3422 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm5[0,0,2,2]
3433 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm7[0,0,2,2]
3445 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm7[0,0,2,2]
3454 ; SSE41-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[0,0,2,2]