Lines Matching refs:AVX512

6 …triple=x86_64-unknown-unknown -mattr=+avx512dq,+avx512vl | FileCheck %s --check-prefixes=AVX,AVX512
7 …triple=x86_64-unknown-unknown -mattr=+avx512bw,+avx512vl | FileCheck %s --check-prefixes=AVX,AVX512
398 ; AVX512-LABEL: test_v8f64_align1:
399 ; AVX512:       # %bb.0:
400 ; AVX512-NEXT:    vmovups (%rdi), %zmm0
401 ; AVX512-NEXT:    retq
427 ; AVX512-LABEL: test_v16f32_align1:
428 ; AVX512:       # %bb.0:
429 ; AVX512-NEXT:    vmovups (%rdi), %zmm0
430 ; AVX512-NEXT:    retq
456 ; AVX512-LABEL: test_v8i64_align1:
457 ; AVX512:       # %bb.0:
458 ; AVX512-NEXT:    vmovups (%rdi), %zmm0
459 ; AVX512-NEXT:    retq
485 ; AVX512-LABEL: test_v16i32_align1:
486 ; AVX512:       # %bb.0:
487 ; AVX512-NEXT:    vmovups (%rdi), %zmm0
488 ; AVX512-NEXT:    retq
514 ; AVX512-LABEL: test_v32i16_align1:
515 ; AVX512:       # %bb.0:
516 ; AVX512-NEXT:    vmovups (%rdi), %zmm0
517 ; AVX512-NEXT:    retq
543 ; AVX512-LABEL: test_v64i8_align1:
544 ; AVX512:       # %bb.0:
545 ; AVX512-NEXT:    vmovups (%rdi), %zmm0
546 ; AVX512-NEXT:    retq
608 ; AVX512-LABEL: test_v8f64_align16:
609 ; AVX512:       # %bb.0:
610 ; AVX512-NEXT:    pushq %rbp
611 ; AVX512-NEXT:    movq %rsp, %rbp
612 ; AVX512-NEXT:    andq $-64, %rsp
613 ; AVX512-NEXT:    subq $128, %rsp
614 ; AVX512-NEXT:    vmovntdqa 48(%rdi), %xmm0
615 ; AVX512-NEXT:    vmovdqa %xmm0, {{[0-9]+}}(%rsp)
616 ; AVX512-NEXT:    vmovntdqa 32(%rdi), %xmm0
617 ; AVX512-NEXT:    vmovdqa %xmm0, {{[0-9]+}}(%rsp)
618 ; AVX512-NEXT:    vmovntdqa 16(%rdi), %xmm0
619 ; AVX512-NEXT:    vmovdqa %xmm0, {{[0-9]+}}(%rsp)
620 ; AVX512-NEXT:    vmovntdqa (%rdi), %xmm0
621 ; AVX512-NEXT:    vmovdqa %xmm0, (%rsp)
622 ; AVX512-NEXT:    vmovaps (%rsp), %zmm0
623 ; AVX512-NEXT:    movq %rbp, %rsp
624 ; AVX512-NEXT:    popq %rbp
625 ; AVX512-NEXT:    retq
687 ; AVX512-LABEL: test_v16f32_align16:
688 ; AVX512:       # %bb.0:
689 ; AVX512-NEXT:    pushq %rbp
690 ; AVX512-NEXT:    movq %rsp, %rbp
691 ; AVX512-NEXT:    andq $-64, %rsp
692 ; AVX512-NEXT:    subq $128, %rsp
693 ; AVX512-NEXT:    vmovntdqa 48(%rdi), %xmm0
694 ; AVX512-NEXT:    vmovdqa %xmm0, {{[0-9]+}}(%rsp)
695 ; AVX512-NEXT:    vmovntdqa 32(%rdi), %xmm0
696 ; AVX512-NEXT:    vmovdqa %xmm0, {{[0-9]+}}(%rsp)
697 ; AVX512-NEXT:    vmovntdqa 16(%rdi), %xmm0
698 ; AVX512-NEXT:    vmovdqa %xmm0, {{[0-9]+}}(%rsp)
699 ; AVX512-NEXT:    vmovntdqa (%rdi), %xmm0
700 ; AVX512-NEXT:    vmovdqa %xmm0, (%rsp)
701 ; AVX512-NEXT:    vmovaps (%rsp), %zmm0
702 ; AVX512-NEXT:    movq %rbp, %rsp
703 ; AVX512-NEXT:    popq %rbp
704 ; AVX512-NEXT:    retq
766 ; AVX512-LABEL: test_v8i64_align16:
767 ; AVX512:       # %bb.0:
768 ; AVX512-NEXT:    pushq %rbp
769 ; AVX512-NEXT:    movq %rsp, %rbp
770 ; AVX512-NEXT:    andq $-64, %rsp
771 ; AVX512-NEXT:    subq $128, %rsp
772 ; AVX512-NEXT:    vmovntdqa 48(%rdi), %xmm0
773 ; AVX512-NEXT:    vmovdqa %xmm0, {{[0-9]+}}(%rsp)
774 ; AVX512-NEXT:    vmovntdqa 32(%rdi), %xmm0
775 ; AVX512-NEXT:    vmovdqa %xmm0, {{[0-9]+}}(%rsp)
776 ; AVX512-NEXT:    vmovntdqa 16(%rdi), %xmm0
777 ; AVX512-NEXT:    vmovdqa %xmm0, {{[0-9]+}}(%rsp)
778 ; AVX512-NEXT:    vmovntdqa (%rdi), %xmm0
779 ; AVX512-NEXT:    vmovdqa %xmm0, (%rsp)
780 ; AVX512-NEXT:    vmovaps (%rsp), %zmm0
781 ; AVX512-NEXT:    movq %rbp, %rsp
782 ; AVX512-NEXT:    popq %rbp
783 ; AVX512-NEXT:    retq
845 ; AVX512-LABEL: test_v16i32_align16:
846 ; AVX512:       # %bb.0:
847 ; AVX512-NEXT:    pushq %rbp
848 ; AVX512-NEXT:    movq %rsp, %rbp
849 ; AVX512-NEXT:    andq $-64, %rsp
850 ; AVX512-NEXT:    subq $128, %rsp
851 ; AVX512-NEXT:    vmovntdqa 48(%rdi), %xmm0
852 ; AVX512-NEXT:    vmovdqa %xmm0, {{[0-9]+}}(%rsp)
853 ; AVX512-NEXT:    vmovntdqa 32(%rdi), %xmm0
854 ; AVX512-NEXT:    vmovdqa %xmm0, {{[0-9]+}}(%rsp)
855 ; AVX512-NEXT:    vmovntdqa 16(%rdi), %xmm0
856 ; AVX512-NEXT:    vmovdqa %xmm0, {{[0-9]+}}(%rsp)
857 ; AVX512-NEXT:    vmovntdqa (%rdi), %xmm0
858 ; AVX512-NEXT:    vmovdqa %xmm0, (%rsp)
859 ; AVX512-NEXT:    vmovaps (%rsp), %zmm0
860 ; AVX512-NEXT:    movq %rbp, %rsp
861 ; AVX512-NEXT:    popq %rbp
862 ; AVX512-NEXT:    retq
924 ; AVX512-LABEL: test_v32i16_align16:
925 ; AVX512:       # %bb.0:
926 ; AVX512-NEXT:    pushq %rbp
927 ; AVX512-NEXT:    movq %rsp, %rbp
928 ; AVX512-NEXT:    andq $-64, %rsp
929 ; AVX512-NEXT:    subq $128, %rsp
930 ; AVX512-NEXT:    vmovntdqa 48(%rdi), %xmm0
931 ; AVX512-NEXT:    vmovdqa %xmm0, {{[0-9]+}}(%rsp)
932 ; AVX512-NEXT:    vmovntdqa 32(%rdi), %xmm0
933 ; AVX512-NEXT:    vmovdqa %xmm0, {{[0-9]+}}(%rsp)
934 ; AVX512-NEXT:    vmovntdqa 16(%rdi), %xmm0
935 ; AVX512-NEXT:    vmovdqa %xmm0, {{[0-9]+}}(%rsp)
936 ; AVX512-NEXT:    vmovntdqa (%rdi), %xmm0
937 ; AVX512-NEXT:    vmovdqa %xmm0, (%rsp)
938 ; AVX512-NEXT:    vmovaps (%rsp), %zmm0
939 ; AVX512-NEXT:    movq %rbp, %rsp
940 ; AVX512-NEXT:    popq %rbp
941 ; AVX512-NEXT:    retq
1003 ; AVX512-LABEL: test_v64i8_align16:
1004 ; AVX512:       # %bb.0:
1005 ; AVX512-NEXT:    pushq %rbp
1006 ; AVX512-NEXT:    movq %rsp, %rbp
1007 ; AVX512-NEXT:    andq $-64, %rsp
1008 ; AVX512-NEXT:    subq $128, %rsp
1009 ; AVX512-NEXT:    vmovntdqa 48(%rdi), %xmm0
1010 ; AVX512-NEXT:    vmovdqa %xmm0, {{[0-9]+}}(%rsp)
1011 ; AVX512-NEXT:    vmovntdqa 32(%rdi), %xmm0
1012 ; AVX512-NEXT:    vmovdqa %xmm0, {{[0-9]+}}(%rsp)
1013 ; AVX512-NEXT:    vmovntdqa 16(%rdi), %xmm0
1014 ; AVX512-NEXT:    vmovdqa %xmm0, {{[0-9]+}}(%rsp)
1015 ; AVX512-NEXT:    vmovntdqa (%rdi), %xmm0
1016 ; AVX512-NEXT:    vmovdqa %xmm0, (%rsp)
1017 ; AVX512-NEXT:    vmovaps (%rsp), %zmm0
1018 ; AVX512-NEXT:    movq %rbp, %rsp
1019 ; AVX512-NEXT:    popq %rbp
1020 ; AVX512-NEXT:    retq
1058 ; AVX512-LABEL: test_v8f64_align32:
1059 ; AVX512:       # %bb.0:
1060 ; AVX512-NEXT:    pushq %rbp
1061 ; AVX512-NEXT:    movq %rsp, %rbp
1062 ; AVX512-NEXT:    andq $-64, %rsp
1063 ; AVX512-NEXT:    subq $128, %rsp
1064 ; AVX512-NEXT:    vmovntdqa 32(%rdi), %ymm0
1065 ; AVX512-NEXT:    vmovdqa %ymm0, {{[0-9]+}}(%rsp)
1066 ; AVX512-NEXT:    vmovntdqa (%rdi), %ymm0
1067 ; AVX512-NEXT:    vmovdqa %ymm0, (%rsp)
1068 ; AVX512-NEXT:    vmovaps (%rsp), %zmm0
1069 ; AVX512-NEXT:    movq %rbp, %rsp
1070 ; AVX512-NEXT:    popq %rbp
1071 ; AVX512-NEXT:    retq
1109 ; AVX512-LABEL: test_v16f32_align32:
1110 ; AVX512:       # %bb.0:
1111 ; AVX512-NEXT:    pushq %rbp
1112 ; AVX512-NEXT:    movq %rsp, %rbp
1113 ; AVX512-NEXT:    andq $-64, %rsp
1114 ; AVX512-NEXT:    subq $128, %rsp
1115 ; AVX512-NEXT:    vmovntdqa 32(%rdi), %ymm0
1116 ; AVX512-NEXT:    vmovdqa %ymm0, {{[0-9]+}}(%rsp)
1117 ; AVX512-NEXT:    vmovntdqa (%rdi), %ymm0
1118 ; AVX512-NEXT:    vmovdqa %ymm0, (%rsp)
1119 ; AVX512-NEXT:    vmovaps (%rsp), %zmm0
1120 ; AVX512-NEXT:    movq %rbp, %rsp
1121 ; AVX512-NEXT:    popq %rbp
1122 ; AVX512-NEXT:    retq
1160 ; AVX512-LABEL: test_v8i64_align32:
1161 ; AVX512:       # %bb.0:
1162 ; AVX512-NEXT:    pushq %rbp
1163 ; AVX512-NEXT:    movq %rsp, %rbp
1164 ; AVX512-NEXT:    andq $-64, %rsp
1165 ; AVX512-NEXT:    subq $128, %rsp
1166 ; AVX512-NEXT:    vmovntdqa 32(%rdi), %ymm0
1167 ; AVX512-NEXT:    vmovdqa %ymm0, {{[0-9]+}}(%rsp)
1168 ; AVX512-NEXT:    vmovntdqa (%rdi), %ymm0
1169 ; AVX512-NEXT:    vmovdqa %ymm0, (%rsp)
1170 ; AVX512-NEXT:    vmovaps (%rsp), %zmm0
1171 ; AVX512-NEXT:    movq %rbp, %rsp
1172 ; AVX512-NEXT:    popq %rbp
1173 ; AVX512-NEXT:    retq
1211 ; AVX512-LABEL: test_v16i32_align32:
1212 ; AVX512:       # %bb.0:
1213 ; AVX512-NEXT:    pushq %rbp
1214 ; AVX512-NEXT:    movq %rsp, %rbp
1215 ; AVX512-NEXT:    andq $-64, %rsp
1216 ; AVX512-NEXT:    subq $128, %rsp
1217 ; AVX512-NEXT:    vmovntdqa 32(%rdi), %ymm0
1218 ; AVX512-NEXT:    vmovdqa %ymm0, {{[0-9]+}}(%rsp)
1219 ; AVX512-NEXT:    vmovntdqa (%rdi), %ymm0
1220 ; AVX512-NEXT:    vmovdqa %ymm0, (%rsp)
1221 ; AVX512-NEXT:    vmovaps (%rsp), %zmm0
1222 ; AVX512-NEXT:    movq %rbp, %rsp
1223 ; AVX512-NEXT:    popq %rbp
1224 ; AVX512-NEXT:    retq
1262 ; AVX512-LABEL: test_v32i16_align32:
1263 ; AVX512:       # %bb.0:
1264 ; AVX512-NEXT:    pushq %rbp
1265 ; AVX512-NEXT:    movq %rsp, %rbp
1266 ; AVX512-NEXT:    andq $-64, %rsp
1267 ; AVX512-NEXT:    subq $128, %rsp
1268 ; AVX512-NEXT:    vmovntdqa 32(%rdi), %ymm0
1269 ; AVX512-NEXT:    vmovdqa %ymm0, {{[0-9]+}}(%rsp)
1270 ; AVX512-NEXT:    vmovntdqa (%rdi), %ymm0
1271 ; AVX512-NEXT:    vmovdqa %ymm0, (%rsp)
1272 ; AVX512-NEXT:    vmovaps (%rsp), %zmm0
1273 ; AVX512-NEXT:    movq %rbp, %rsp
1274 ; AVX512-NEXT:    popq %rbp
1275 ; AVX512-NEXT:    retq
1313 ; AVX512-LABEL: test_v64i8_align32:
1314 ; AVX512:       # %bb.0:
1315 ; AVX512-NEXT:    pushq %rbp
1316 ; AVX512-NEXT:    movq %rsp, %rbp
1317 ; AVX512-NEXT:    andq $-64, %rsp
1318 ; AVX512-NEXT:    subq $128, %rsp
1319 ; AVX512-NEXT:    vmovntdqa 32(%rdi), %ymm0
1320 ; AVX512-NEXT:    vmovdqa %ymm0, {{[0-9]+}}(%rsp)
1321 ; AVX512-NEXT:    vmovntdqa (%rdi), %ymm0
1322 ; AVX512-NEXT:    vmovdqa %ymm0, (%rsp)
1323 ; AVX512-NEXT:    vmovaps (%rsp), %zmm0
1324 ; AVX512-NEXT:    movq %rbp, %rsp
1325 ; AVX512-NEXT:    popq %rbp
1326 ; AVX512-NEXT:    retq