Lines Matching refs:X32

2 …i386-unknown-unknown -mattr=+avx512f,+avx512vl | FileCheck %s --check-prefix=ALL --check-prefix=X32
8 ; X32-LABEL: test_mm_broadcastd_epi32:
9 ; X32:       # BB#0:
10 ; X32-NEXT:    vpbroadcastd %xmm0, %xmm0
11 ; X32-NEXT:    retl
24 ; X32-LABEL: test_mm_mask_broadcastd_epi32:
25 ; X32:       # BB#0:
26 ; X32-NEXT:    pushl %eax
27 ; X32-NEXT:  .Ltmp0:
28 ; X32-NEXT:    .cfi_def_cfa_offset 8
29 ; X32-NEXT:    movb {{[0-9]+}}(%esp), %al
30 ; X32-NEXT:    andb $15, %al
31 ; X32-NEXT:    movb %al, (%esp)
32 ; X32-NEXT:    movzbl (%esp), %eax
33 ; X32-NEXT:    kmovw %eax, %k1
34 ; X32-NEXT:    vpbroadcastd %xmm1, %xmm0 {%k1}
35 ; X32-NEXT:    popl %eax
36 ; X32-NEXT:    retl
57 ; X32-LABEL: test_mm_maskz_broadcastd_epi32:
58 ; X32:       # BB#0:
59 ; X32-NEXT:    pushl %eax
60 ; X32-NEXT:  .Ltmp1:
61 ; X32-NEXT:    .cfi_def_cfa_offset 8
62 ; X32-NEXT:    movb {{[0-9]+}}(%esp), %al
63 ; X32-NEXT:    andb $15, %al
64 ; X32-NEXT:    movb %al, (%esp)
65 ; X32-NEXT:    movzbl (%esp), %eax
66 ; X32-NEXT:    kmovw %eax, %k1
67 ; X32-NEXT:    vpbroadcastd %xmm0, %xmm0 {%k1} {z}
68 ; X32-NEXT:    popl %eax
69 ; X32-NEXT:    retl
89 ; X32-LABEL: test_mm256_broadcastd_epi32:
90 ; X32:       # BB#0:
91 ; X32-NEXT:    vpbroadcastd %xmm0, %ymm0
92 ; X32-NEXT:    retl
105 ; X32-LABEL: test_mm256_mask_broadcastd_epi32:
106 ; X32:       # BB#0:
107 ; X32-NEXT:    movb {{[0-9]+}}(%esp), %al
108 ; X32-NEXT:    kmovw %eax, %k1
109 ; X32-NEXT:    vpbroadcastd %xmm1, %ymm0 {%k1}
110 ; X32-NEXT:    retl
127 ; X32-LABEL: test_mm256_maskz_broadcastd_epi32:
128 ; X32:       # BB#0:
129 ; X32-NEXT:    movb {{[0-9]+}}(%esp), %al
130 ; X32-NEXT:    kmovw %eax, %k1
131 ; X32-NEXT:    vpbroadcastd %xmm0, %ymm0 {%k1} {z}
132 ; X32-NEXT:    retl
148 ; X32-LABEL: test_mm_broadcastq_epi64:
149 ; X32:       # BB#0:
150 ; X32-NEXT:    vpbroadcastq %xmm0, %xmm0
151 ; X32-NEXT:    retl
162 ; X32-LABEL: test_mm_mask_broadcastq_epi64:
163 ; X32:       # BB#0:
164 ; X32-NEXT:    pushl %eax
165 ; X32-NEXT:  .Ltmp2:
166 ; X32-NEXT:    .cfi_def_cfa_offset 8
167 ; X32-NEXT:    movb {{[0-9]+}}(%esp), %al
168 ; X32-NEXT:    andb $3, %al
169 ; X32-NEXT:    movb %al, {{[0-9]+}}(%esp)
170 ; X32-NEXT:    movzbl {{[0-9]+}}(%esp), %eax
171 ; X32-NEXT:    kmovw %eax, %k1
172 ; X32-NEXT:    vpbroadcastq %xmm1, %xmm0 {%k1}
173 ; X32-NEXT:    popl %eax
174 ; X32-NEXT:    retl
192 ; X32-LABEL: test_mm_maskz_broadcastq_epi64:
193 ; X32:       # BB#0:
194 ; X32-NEXT:    pushl %eax
195 ; X32-NEXT:  .Ltmp3:
196 ; X32-NEXT:    .cfi_def_cfa_offset 8
197 ; X32-NEXT:    movb {{[0-9]+}}(%esp), %al
198 ; X32-NEXT:    andb $3, %al
199 ; X32-NEXT:    movb %al, {{[0-9]+}}(%esp)
200 ; X32-NEXT:    movzbl {{[0-9]+}}(%esp), %eax
201 ; X32-NEXT:    kmovw %eax, %k1
202 ; X32-NEXT:    vpbroadcastq %xmm0, %xmm0 {%k1} {z}
203 ; X32-NEXT:    popl %eax
204 ; X32-NEXT:    retl
222 ; X32-LABEL: test_mm256_broadcastq_epi64:
223 ; X32:       # BB#0:
224 ; X32-NEXT:    vpbroadcastq %xmm0, %ymm0
225 ; X32-NEXT:    retl
236 ; X32-LABEL: test_mm256_mask_broadcastq_epi64:
237 ; X32:       # BB#0:
238 ; X32-NEXT:    pushl %eax
239 ; X32-NEXT:  .Ltmp4:
240 ; X32-NEXT:    .cfi_def_cfa_offset 8
241 ; X32-NEXT:    movb {{[0-9]+}}(%esp), %al
242 ; X32-NEXT:    andb $15, %al
243 ; X32-NEXT:    movb %al, (%esp)
244 ; X32-NEXT:    movzbl (%esp), %eax
245 ; X32-NEXT:    kmovw %eax, %k1
246 ; X32-NEXT:    vpbroadcastq %xmm1, %ymm0 {%k1}
247 ; X32-NEXT:    popl %eax
248 ; X32-NEXT:    retl
266 ; X32-LABEL: test_mm256_maskz_broadcastq_epi64:
267 ; X32:       # BB#0:
268 ; X32-NEXT:    pushl %eax
269 ; X32-NEXT:  .Ltmp5:
270 ; X32-NEXT:    .cfi_def_cfa_offset 8
271 ; X32-NEXT:    movb {{[0-9]+}}(%esp), %al
272 ; X32-NEXT:    andb $15, %al
273 ; X32-NEXT:    movb %al, (%esp)
274 ; X32-NEXT:    movzbl (%esp), %eax
275 ; X32-NEXT:    kmovw %eax, %k1
276 ; X32-NEXT:    vpbroadcastq %xmm0, %ymm0 {%k1} {z}
277 ; X32-NEXT:    popl %eax
278 ; X32-NEXT:    retl
296 ; X32-LABEL: test_mm_broadcastsd_pd:
297 ; X32:       # BB#0:
298 ; X32-NEXT:    vmovddup {{.*#+}} xmm0 = xmm0[0,0]
299 ; X32-NEXT:    retl
310 ; X32-LABEL: test_mm_mask_broadcastsd_pd:
311 ; X32:       # BB#0:
312 ; X32-NEXT:    pushl %eax
313 ; X32-NEXT:  .Ltmp6:
314 ; X32-NEXT:    .cfi_def_cfa_offset 8
315 ; X32-NEXT:    movb {{[0-9]+}}(%esp), %al
316 ; X32-NEXT:    andb $3, %al
317 ; X32-NEXT:    movb %al, {{[0-9]+}}(%esp)
318 ; X32-NEXT:    movzbl {{[0-9]+}}(%esp), %eax
319 ; X32-NEXT:    kmovw %eax, %k1
320 ; X32-NEXT:    vmovddup {{.*#+}} xmm0 {%k1} = xmm1[0,0]
321 ; X32-NEXT:    popl %eax
322 ; X32-NEXT:    retl
340 ; X32-LABEL: test_mm_maskz_broadcastsd_pd:
341 ; X32:       # BB#0:
342 ; X32-NEXT:    pushl %eax
343 ; X32-NEXT:  .Ltmp7:
344 ; X32-NEXT:    .cfi_def_cfa_offset 8
345 ; X32-NEXT:    movb {{[0-9]+}}(%esp), %al
346 ; X32-NEXT:    andb $3, %al
347 ; X32-NEXT:    movb %al, {{[0-9]+}}(%esp)
348 ; X32-NEXT:    movzbl {{[0-9]+}}(%esp), %eax
349 ; X32-NEXT:    kmovw %eax, %k1
350 ; X32-NEXT:    vmovddup {{.*#+}} xmm0 {%k1} {z} = xmm0[0,0]
351 ; X32-NEXT:    popl %eax
352 ; X32-NEXT:    retl
370 ; X32-LABEL: test_mm256_broadcastsd_pd:
371 ; X32:       # BB#0:
372 ; X32-NEXT:    vbroadcastsd %xmm0, %ymm0
373 ; X32-NEXT:    retl
384 ; X32-LABEL: test_mm256_mask_broadcastsd_pd:
385 ; X32:       # BB#0:
386 ; X32-NEXT:    pushl %eax
387 ; X32-NEXT:  .Ltmp8:
388 ; X32-NEXT:    .cfi_def_cfa_offset 8
389 ; X32-NEXT:    movb {{[0-9]+}}(%esp), %al
390 ; X32-NEXT:    andb $15, %al
391 ; X32-NEXT:    movb %al, (%esp)
392 ; X32-NEXT:    movzbl (%esp), %eax
393 ; X32-NEXT:    kmovw %eax, %k1
394 ; X32-NEXT:    vbroadcastsd %xmm1, %ymm0 {%k1}
395 ; X32-NEXT:    popl %eax
396 ; X32-NEXT:    retl
414 ; X32-LABEL: test_mm256_maskz_broadcastsd_pd:
415 ; X32:       # BB#0:
416 ; X32-NEXT:    pushl %eax
417 ; X32-NEXT:  .Ltmp9:
418 ; X32-NEXT:    .cfi_def_cfa_offset 8
419 ; X32-NEXT:    movb {{[0-9]+}}(%esp), %al
420 ; X32-NEXT:    andb $15, %al
421 ; X32-NEXT:    movb %al, (%esp)
422 ; X32-NEXT:    movzbl (%esp), %eax
423 ; X32-NEXT:    kmovw %eax, %k1
424 ; X32-NEXT:    vbroadcastsd %xmm0, %ymm0 {%k1} {z}
425 ; X32-NEXT:    popl %eax
426 ; X32-NEXT:    retl
444 ; X32-LABEL: test_mm_broadcastss_ps:
445 ; X32:       # BB#0:
446 ; X32-NEXT:    vbroadcastss %xmm0, %xmm0
447 ; X32-NEXT:    retl
458 ; X32-LABEL: test_mm_mask_broadcastss_ps:
459 ; X32:       # BB#0:
460 ; X32-NEXT:    pushl %eax
461 ; X32-NEXT:  .Ltmp10:
462 ; X32-NEXT:    .cfi_def_cfa_offset 8
463 ; X32-NEXT:    movb {{[0-9]+}}(%esp), %al
464 ; X32-NEXT:    andb $15, %al
465 ; X32-NEXT:    movb %al, (%esp)
466 ; X32-NEXT:    movzbl (%esp), %eax
467 ; X32-NEXT:    kmovw %eax, %k1
468 ; X32-NEXT:    vbroadcastss %xmm1, %xmm0 {%k1}
469 ; X32-NEXT:    popl %eax
470 ; X32-NEXT:    retl
488 ; X32-LABEL: test_mm_maskz_broadcastss_ps:
489 ; X32:       # BB#0:
490 ; X32-NEXT:    pushl %eax
491 ; X32-NEXT:  .Ltmp11:
492 ; X32-NEXT:    .cfi_def_cfa_offset 8
493 ; X32-NEXT:    movb {{[0-9]+}}(%esp), %al
494 ; X32-NEXT:    andb $15, %al
495 ; X32-NEXT:    movb %al, (%esp)
496 ; X32-NEXT:    movzbl (%esp), %eax
497 ; X32-NEXT:    kmovw %eax, %k1
498 ; X32-NEXT:    vbroadcastss %xmm0, %xmm0 {%k1} {z}
499 ; X32-NEXT:    popl %eax
500 ; X32-NEXT:    retl
518 ; X32-LABEL: test_mm256_broadcastss_ps:
519 ; X32:       # BB#0:
520 ; X32-NEXT:    vbroadcastss %xmm0, %ymm0
521 ; X32-NEXT:    retl
532 ; X32-LABEL: test_mm256_mask_broadcastss_ps:
533 ; X32:       # BB#0:
534 ; X32-NEXT:    movb {{[0-9]+}}(%esp), %al
535 ; X32-NEXT:    kmovw %eax, %k1
536 ; X32-NEXT:    vbroadcastss %xmm1, %ymm0 {%k1}
537 ; X32-NEXT:    retl
551 ; X32-LABEL: test_mm256_maskz_broadcastss_ps:
552 ; X32:       # BB#0:
553 ; X32-NEXT:    movb {{[0-9]+}}(%esp), %al
554 ; X32-NEXT:    kmovw %eax, %k1
555 ; X32-NEXT:    vbroadcastss %xmm0, %ymm0 {%k1} {z}
556 ; X32-NEXT:    retl
570 ; X32-LABEL: test_mm_movddup_pd:
571 ; X32:       # BB#0:
572 ; X32-NEXT:    vmovddup {{.*#+}} xmm0 = xmm0[0,0]
573 ; X32-NEXT:    retl
584 ; X32-LABEL: test_mm_mask_movddup_pd:
585 ; X32:       # BB#0:
586 ; X32-NEXT:    pushl %eax
587 ; X32-NEXT:  .Ltmp12:
588 ; X32-NEXT:    .cfi_def_cfa_offset 8
589 ; X32-NEXT:    movb {{[0-9]+}}(%esp), %al
590 ; X32-NEXT:    andb $3, %al
591 ; X32-NEXT:    movb %al, {{[0-9]+}}(%esp)
592 ; X32-NEXT:    movzbl {{[0-9]+}}(%esp), %eax
593 ; X32-NEXT:    kmovw %eax, %k1
594 ; X32-NEXT:    vmovddup {{.*#+}} xmm0 {%k1} = xmm1[0,0]
595 ; X32-NEXT:    popl %eax
596 ; X32-NEXT:    retl
614 ; X32-LABEL: test_mm_maskz_movddup_pd:
615 ; X32:       # BB#0:
616 ; X32-NEXT:    pushl %eax
617 ; X32-NEXT:  .Ltmp13:
618 ; X32-NEXT:    .cfi_def_cfa_offset 8
619 ; X32-NEXT:    movb {{[0-9]+}}(%esp), %al
620 ; X32-NEXT:    andb $3, %al
621 ; X32-NEXT:    movb %al, {{[0-9]+}}(%esp)
622 ; X32-NEXT:    movzbl {{[0-9]+}}(%esp), %eax
623 ; X32-NEXT:    kmovw %eax, %k1
624 ; X32-NEXT:    vmovddup {{.*#+}} xmm0 {%k1} {z} = xmm0[0,0]
625 ; X32-NEXT:    popl %eax
626 ; X32-NEXT:    retl
644 ; X32-LABEL: test_mm256_movddup_pd:
645 ; X32:       # BB#0:
646 ; X32-NEXT:    vmovddup {{.*#+}} ymm0 = ymm0[0,0,2,2]
647 ; X32-NEXT:    retl
658 ; X32-LABEL: test_mm256_mask_movddup_pd:
659 ; X32:       # BB#0:
660 ; X32-NEXT:    pushl %eax
661 ; X32-NEXT:  .Ltmp14:
662 ; X32-NEXT:    .cfi_def_cfa_offset 8
663 ; X32-NEXT:    movb {{[0-9]+}}(%esp), %al
664 ; X32-NEXT:    andb $15, %al
665 ; X32-NEXT:    movb %al, (%esp)
666 ; X32-NEXT:    movzbl (%esp), %eax
667 ; X32-NEXT:    kmovw %eax, %k1
668 ; X32-NEXT:    vmovddup {{.*#+}} ymm0 {%k1} = ymm1[0,0,2,2]
669 ; X32-NEXT:    popl %eax
670 ; X32-NEXT:    retl
688 ; X32-LABEL: test_mm256_maskz_movddup_pd:
689 ; X32:       # BB#0:
690 ; X32-NEXT:    pushl %eax
691 ; X32-NEXT:  .Ltmp15:
692 ; X32-NEXT:    .cfi_def_cfa_offset 8
693 ; X32-NEXT:    movb {{[0-9]+}}(%esp), %al
694 ; X32-NEXT:    andb $15, %al
695 ; X32-NEXT:    movb %al, (%esp)
696 ; X32-NEXT:    movzbl (%esp), %eax
697 ; X32-NEXT:    kmovw %eax, %k1
698 ; X32-NEXT:    vmovddup {{.*#+}} ymm0 {%k1} {z} = ymm0[0,0,2,2]
699 ; X32-NEXT:    popl %eax
700 ; X32-NEXT:    retl
718 ; X32-LABEL: test_mm_movehdup_ps:
719 ; X32:       # BB#0:
720 ; X32-NEXT:    vmovshdup {{.*#+}} xmm0 = xmm0[1,1,3,3]
721 ; X32-NEXT:    retl
732 ; X32-LABEL: test_mm_mask_movehdup_ps:
733 ; X32:       # BB#0:
734 ; X32-NEXT:    pushl %eax
735 ; X32-NEXT:  .Ltmp16:
736 ; X32-NEXT:    .cfi_def_cfa_offset 8
737 ; X32-NEXT:    movb {{[0-9]+}}(%esp), %al
738 ; X32-NEXT:    andb $15, %al
739 ; X32-NEXT:    movb %al, (%esp)
740 ; X32-NEXT:    movzbl (%esp), %eax
741 ; X32-NEXT:    kmovw %eax, %k1
742 ; X32-NEXT:    vmovshdup {{.*#+}} xmm0 {%k1} = xmm1[1,1,3,3]
743 ; X32-NEXT:    popl %eax
744 ; X32-NEXT:    retl
762 ; X32-LABEL: test_mm_maskz_movehdup_ps:
763 ; X32:       # BB#0:
764 ; X32-NEXT:    pushl %eax
765 ; X32-NEXT:  .Ltmp17:
766 ; X32-NEXT:    .cfi_def_cfa_offset 8
767 ; X32-NEXT:    movb {{[0-9]+}}(%esp), %al
768 ; X32-NEXT:    andb $15, %al
769 ; X32-NEXT:    movb %al, (%esp)
770 ; X32-NEXT:    movzbl (%esp), %eax
771 ; X32-NEXT:    kmovw %eax, %k1
772 ; X32-NEXT:    vmovshdup {{.*#+}} xmm0 {%k1} {z} = xmm0[1,1,3,3]
773 ; X32-NEXT:    popl %eax
774 ; X32-NEXT:    retl
792 ; X32-LABEL: test_mm256_movehdup_ps:
793 ; X32:       # BB#0:
794 ; X32-NEXT:    vmovshdup {{.*#+}} ymm0 = ymm0[1,1,3,3,5,5,7,7]
795 ; X32-NEXT:    retl
806 ; X32-LABEL: test_mm256_mask_movehdup_ps:
807 ; X32:       # BB#0:
808 ; X32-NEXT:    movb {{[0-9]+}}(%esp), %al
809 ; X32-NEXT:    kmovw %eax, %k1
810 ; X32-NEXT:    vmovshdup {{.*#+}} ymm0 {%k1} = ymm1[1,1,3,3,5,5,7,7]
811 ; X32-NEXT:    retl
825 ; X32-LABEL: test_mm256_maskz_movehdup_ps:
826 ; X32:       # BB#0:
827 ; X32-NEXT:    movb {{[0-9]+}}(%esp), %al
828 ; X32-NEXT:    kmovw %eax, %k1
829 ; X32-NEXT:    vmovshdup {{.*#+}} ymm0 {%k1} {z} = ymm0[1,1,3,3,5,5,7,7]
830 ; X32-NEXT:    retl
844 ; X32-LABEL: test_mm_moveldup_ps:
845 ; X32:       # BB#0:
846 ; X32-NEXT:    vmovsldup {{.*#+}} xmm0 = xmm0[0,0,2,2]
847 ; X32-NEXT:    retl
858 ; X32-LABEL: test_mm_mask_moveldup_ps:
859 ; X32:       # BB#0:
860 ; X32-NEXT:    pushl %eax
861 ; X32-NEXT:  .Ltmp18:
862 ; X32-NEXT:    .cfi_def_cfa_offset 8
863 ; X32-NEXT:    movb {{[0-9]+}}(%esp), %al
864 ; X32-NEXT:    andb $15, %al
865 ; X32-NEXT:    movb %al, (%esp)
866 ; X32-NEXT:    movzbl (%esp), %eax
867 ; X32-NEXT:    kmovw %eax, %k1
868 ; X32-NEXT:    vmovsldup {{.*#+}} xmm0 {%k1} = xmm1[0,0,2,2]
869 ; X32-NEXT:    popl %eax
870 ; X32-NEXT:    retl
888 ; X32-LABEL: test_mm_maskz_moveldup_ps:
889 ; X32:       # BB#0:
890 ; X32-NEXT:    pushl %eax
891 ; X32-NEXT:  .Ltmp19:
892 ; X32-NEXT:    .cfi_def_cfa_offset 8
893 ; X32-NEXT:    movb {{[0-9]+}}(%esp), %al
894 ; X32-NEXT:    andb $15, %al
895 ; X32-NEXT:    movb %al, (%esp)
896 ; X32-NEXT:    movzbl (%esp), %eax
897 ; X32-NEXT:    kmovw %eax, %k1
898 ; X32-NEXT:    vmovsldup {{.*#+}} xmm0 {%k1} {z} = xmm0[0,0,2,2]
899 ; X32-NEXT:    popl %eax
900 ; X32-NEXT:    retl
918 ; X32-LABEL: test_mm256_moveldup_ps:
919 ; X32:       # BB#0:
920 ; X32-NEXT:    vmovsldup {{.*#+}} ymm0 = ymm0[0,0,2,2,4,4,6,6]
921 ; X32-NEXT:    retl
932 ; X32-LABEL: test_mm256_mask_moveldup_ps:
933 ; X32:       # BB#0:
934 ; X32-NEXT:    movb {{[0-9]+}}(%esp), %al
935 ; X32-NEXT:    kmovw %eax, %k1
936 ; X32-NEXT:    vmovsldup {{.*#+}} ymm0 {%k1} = ymm1[0,0,2,2,4,4,6,6]
937 ; X32-NEXT:    retl
951 ; X32-LABEL: test_mm256_maskz_moveldup_ps:
952 ; X32:       # BB#0:
953 ; X32-NEXT:    movb {{[0-9]+}}(%esp), %al
954 ; X32-NEXT:    kmovw %eax, %k1
955 ; X32-NEXT:    vmovsldup {{.*#+}} ymm0 {%k1} {z} = ymm0[0,0,2,2,4,4,6,6]
956 ; X32-NEXT:    retl
970 ; X32-LABEL: test_mm256_permutex_epi64:
971 ; X32:       # BB#0:
972 ; X32-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[3,0,0,0]
973 ; X32-NEXT:    retl
984 ; X32-LABEL: test_mm256_mask_permutex_epi64:
985 ; X32:       # BB#0:
986 ; X32-NEXT:    pushl %eax
987 ; X32-NEXT:  .Ltmp20:
988 ; X32-NEXT:    .cfi_def_cfa_offset 8
989 ; X32-NEXT:    movb {{[0-9]+}}(%esp), %al
990 ; X32-NEXT:    andb $15, %al
991 ; X32-NEXT:    movb %al, (%esp)
992 ; X32-NEXT:    movzbl (%esp), %eax
993 ; X32-NEXT:    kmovw %eax, %k1
994 ; X32-NEXT:    vpermq {{.*#+}} ymm0 {%k1} = ymm1[1,0,0,0]
995 ; X32-NEXT:    popl %eax
996 ; X32-NEXT:    retl
1014 ; X32-LABEL: test_mm256_maskz_permutex_epi64:
1015 ; X32:       # BB#0:
1016 ; X32-NEXT:    pushl %eax
1017 ; X32-NEXT:  .Ltmp21:
1018 ; X32-NEXT:    .cfi_def_cfa_offset 8
1019 ; X32-NEXT:    movb {{[0-9]+}}(%esp), %al
1020 ; X32-NEXT:    andb $15, %al
1021 ; X32-NEXT:    movb %al, (%esp)
1022 ; X32-NEXT:    movzbl (%esp), %eax
1023 ; X32-NEXT:    kmovw %eax, %k1
1024 ; X32-NEXT:    vpermq {{.*#+}} ymm0 {%k1} {z} = ymm0[1,0,0,0]
1025 ; X32-NEXT:    popl %eax
1026 ; X32-NEXT:    retl
1044 ; X32-LABEL: test_mm256_permutex_pd:
1045 ; X32:       # BB#0:
1046 ; X32-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[3,0,0,0]
1047 ; X32-NEXT:    retl
1058 ; X32-LABEL: test_mm256_mask_permutex_pd:
1059 ; X32:       # BB#0:
1060 ; X32-NEXT:    pushl %eax
1061 ; X32-NEXT:  .Ltmp22:
1062 ; X32-NEXT:    .cfi_def_cfa_offset 8
1063 ; X32-NEXT:    movb {{[0-9]+}}(%esp), %al
1064 ; X32-NEXT:    andb $15, %al
1065 ; X32-NEXT:    movb %al, (%esp)
1066 ; X32-NEXT:    movzbl (%esp), %eax
1067 ; X32-NEXT:    kmovw %eax, %k1
1068 ; X32-NEXT:    vpermpd {{.*#+}} ymm0 {%k1} = ymm1[1,0,0,0]
1069 ; X32-NEXT:    popl %eax
1070 ; X32-NEXT:    retl
1088 ; X32-LABEL: test_mm256_maskz_permutex_pd:
1089 ; X32:       # BB#0:
1090 ; X32-NEXT:    pushl %eax
1091 ; X32-NEXT:  .Ltmp23:
1092 ; X32-NEXT:    .cfi_def_cfa_offset 8
1093 ; X32-NEXT:    movb {{[0-9]+}}(%esp), %al
1094 ; X32-NEXT:    andb $15, %al
1095 ; X32-NEXT:    movb %al, (%esp)
1096 ; X32-NEXT:    movzbl (%esp), %eax
1097 ; X32-NEXT:    kmovw %eax, %k1
1098 ; X32-NEXT:    vpermpd {{.*#+}} ymm0 {%k1} {z} = ymm0[1,0,0,0]
1099 ; X32-NEXT:    popl %eax
1100 ; X32-NEXT:    retl
1118 ; X32-LABEL: test_mm_shuffle_pd:
1119 ; X32:       # BB#0:
1120 ; X32-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
1121 ; X32-NEXT:    retl
1132 ; X32-LABEL: test_mm_mask_shuffle_pd:
1133 ; X32:       # BB#0:
1134 ; X32-NEXT:    pushl %eax
1135 ; X32-NEXT:  .Ltmp24:
1136 ; X32-NEXT:    .cfi_def_cfa_offset 8
1137 ; X32-NEXT:    movb {{[0-9]+}}(%esp), %al
1138 ; X32-NEXT:    andb $3, %al
1139 ; X32-NEXT:    movb %al, {{[0-9]+}}(%esp)
1140 ; X32-NEXT:    movzbl {{[0-9]+}}(%esp), %eax
1141 ; X32-NEXT:    kmovw %eax, %k1
1142 ; X32-NEXT:    vunpckhpd {{.*#+}} xmm0 {%k1} = xmm1[1],xmm2[1]
1143 ; X32-NEXT:    popl %eax
1144 ; X32-NEXT:    retl
1162 ; X32-LABEL: test_mm_maskz_shuffle_pd:
1163 ; X32:       # BB#0:
1164 ; X32-NEXT:    pushl %eax
1165 ; X32-NEXT:  .Ltmp25:
1166 ; X32-NEXT:    .cfi_def_cfa_offset 8
1167 ; X32-NEXT:    movb {{[0-9]+}}(%esp), %al
1168 ; X32-NEXT:    andb $3, %al
1169 ; X32-NEXT:    movb %al, {{[0-9]+}}(%esp)
1170 ; X32-NEXT:    movzbl {{[0-9]+}}(%esp), %eax
1171 ; X32-NEXT:    kmovw %eax, %k1
1172 ; X32-NEXT:    vunpckhpd {{.*#+}} xmm0 {%k1} {z} = xmm0[1],xmm1[1]
1173 ; X32-NEXT:    popl %eax
1174 ; X32-NEXT:    retl
1192 ; X32-LABEL: test_mm256_shuffle_pd:
1193 ; X32:       # BB#0:
1194 ; X32-NEXT:    vshufpd {{.*#+}} ymm0 = ymm0[1],ymm1[1],ymm0[2],ymm1[2]
1195 ; X32-NEXT:    retl
1206 ; X32-LABEL: test_mm256_mask_shuffle_pd:
1207 ; X32:       # BB#0:
1208 ; X32-NEXT:    pushl %eax
1209 ; X32-NEXT:  .Ltmp26:
1210 ; X32-NEXT:    .cfi_def_cfa_offset 8
1211 ; X32-NEXT:    movb {{[0-9]+}}(%esp), %al
1212 ; X32-NEXT:    andb $15, %al
1213 ; X32-NEXT:    movb %al, (%esp)
1214 ; X32-NEXT:    movzbl (%esp), %eax
1215 ; X32-NEXT:    kmovw %eax, %k1
1216 ; X32-NEXT:    vshufpd {{.*#+}} ymm0 {%k1} = ymm1[1],ymm2[1],ymm1[2],ymm2[2]
1217 ; X32-NEXT:    popl %eax
1218 ; X32-NEXT:    retl
1236 ; X32-LABEL: test_mm256_maskz_shuffle_pd:
1237 ; X32:       # BB#0:
1238 ; X32-NEXT:    pushl %eax
1239 ; X32-NEXT:  .Ltmp27:
1240 ; X32-NEXT:    .cfi_def_cfa_offset 8
1241 ; X32-NEXT:    movb {{[0-9]+}}(%esp), %al
1242 ; X32-NEXT:    andb $15, %al
1243 ; X32-NEXT:    movb %al, (%esp)
1244 ; X32-NEXT:    movzbl (%esp), %eax
1245 ; X32-NEXT:    kmovw %eax, %k1
1246 ; X32-NEXT:    vshufpd {{.*#+}} ymm0 {%k1} {z} = ymm0[1],ymm1[1],ymm0[2],ymm1[2]
1247 ; X32-NEXT:    popl %eax
1248 ; X32-NEXT:    retl
1266 ; X32-LABEL: test_mm_shuffle_ps:
1267 ; X32:       # BB#0:
1268 ; X32-NEXT:    vshufps {{.*#+}} xmm0 = xmm0[0,1],xmm1[0,0]
1269 ; X32-NEXT:    retl
1280 ; X32-LABEL: test_mm_mask_shuffle_ps:
1281 ; X32:       # BB#0:
1282 ; X32-NEXT:    pushl %eax
1283 ; X32-NEXT:  .Ltmp28:
1284 ; X32-NEXT:    .cfi_def_cfa_offset 8
1285 ; X32-NEXT:    movb {{[0-9]+}}(%esp), %al
1286 ; X32-NEXT:    andb $15, %al
1287 ; X32-NEXT:    movb %al, (%esp)
1288 ; X32-NEXT:    movzbl (%esp), %eax
1289 ; X32-NEXT:    kmovw %eax, %k1
1290 ; X32-NEXT:    vshufps {{.*#+}} xmm0 {%k1} = xmm1[0,1],xmm2[0,0]
1291 ; X32-NEXT:    popl %eax
1292 ; X32-NEXT:    retl
1310 ; X32-LABEL: test_mm_maskz_shuffle_ps:
1311 ; X32:       # BB#0:
1312 ; X32-NEXT:    pushl %eax
1313 ; X32-NEXT:  .Ltmp29:
1314 ; X32-NEXT:    .cfi_def_cfa_offset 8
1315 ; X32-NEXT:    movb {{[0-9]+}}(%esp), %al
1316 ; X32-NEXT:    andb $15, %al
1317 ; X32-NEXT:    movb %al, (%esp)
1318 ; X32-NEXT:    movzbl (%esp), %eax
1319 ; X32-NEXT:    kmovw %eax, %k1
1320 ; X32-NEXT:    vshufps {{.*#+}} xmm0 {%k1} {z} = xmm0[0,1],xmm1[0,0]
1321 ; X32-NEXT:    popl %eax
1322 ; X32-NEXT:    retl
1340 ; X32-LABEL: test_mm256_shuffle_ps:
1341 ; X32:       # BB#0:
1342 ; X32-NEXT:    vshufps {{.*#+}} ymm0 = ymm0[0,1],ymm1[0,0],ymm0[4,5],ymm1[4,4]
1343 ; X32-NEXT:    retl
1354 ; X32-LABEL: test_mm256_mask_shuffle_ps:
1355 ; X32:       # BB#0:
1356 ; X32-NEXT:    movb {{[0-9]+}}(%esp), %al
1357 ; X32-NEXT:    kmovw %eax, %k1
1358 ; X32-NEXT:    vshufps {{.*#+}} ymm0 {%k1} = ymm1[0,1],ymm2[0,0],ymm1[4,5],ymm2[4,4]
1359 ; X32-NEXT:    retl
1373 ; X32-LABEL: test_mm256_maskz_shuffle_ps:
1374 ; X32:       # BB#0:
1375 ; X32-NEXT:    movb {{[0-9]+}}(%esp), %al
1376 ; X32-NEXT:    kmovw %eax, %k1
1377 ; X32-NEXT:    vshufps {{.*#+}} ymm0 {%k1} {z} = ymm0[0,1],ymm1[0,0],ymm0[4,5],ymm1[4,4]
1378 ; X32-NEXT:    retl