Lines Matching refs:AVX512

7 … %s -mtriple=x86_64-unknown-unknown -mattr=+avx512dq,+avx512vl | FileCheck %s --check-prefix=AVX512
8 … %s -mtriple=x86_64-unknown-unknown -mattr=+avx512bw,+avx512vl | FileCheck %s --check-prefix=AVX512
29 ; AVX512-LABEL: test_zero_v2f64_align1:
30 ; AVX512:       # %bb.0:
31 ; AVX512-NEXT:    xorl %eax, %eax
32 ; AVX512-NEXT:    movntiq %rax, 8(%rdi)
33 ; AVX512-NEXT:    movntiq %rax, (%rdi)
34 ; AVX512-NEXT:    retq
54 ; AVX512-LABEL: test_zero_v4f32_align1:
55 ; AVX512:       # %bb.0:
56 ; AVX512-NEXT:    xorl %eax, %eax
57 ; AVX512-NEXT:    movntiq %rax, 8(%rdi)
58 ; AVX512-NEXT:    movntiq %rax, (%rdi)
59 ; AVX512-NEXT:    retq
79 ; AVX512-LABEL: test_zero_v2i64_align1:
80 ; AVX512:       # %bb.0:
81 ; AVX512-NEXT:    xorl %eax, %eax
82 ; AVX512-NEXT:    movntiq %rax, 8(%rdi)
83 ; AVX512-NEXT:    movntiq %rax, (%rdi)
84 ; AVX512-NEXT:    retq
104 ; AVX512-LABEL: test_zero_v4i32_align1:
105 ; AVX512:       # %bb.0:
106 ; AVX512-NEXT:    xorl %eax, %eax
107 ; AVX512-NEXT:    movntiq %rax, 8(%rdi)
108 ; AVX512-NEXT:    movntiq %rax, (%rdi)
109 ; AVX512-NEXT:    retq
129 ; AVX512-LABEL: test_zero_v8i16_align1:
130 ; AVX512:       # %bb.0:
131 ; AVX512-NEXT:    xorl %eax, %eax
132 ; AVX512-NEXT:    movntiq %rax, 8(%rdi)
133 ; AVX512-NEXT:    movntiq %rax, (%rdi)
134 ; AVX512-NEXT:    retq
154 ; AVX512-LABEL: test_zero_v16i8_align1:
155 ; AVX512:       # %bb.0:
156 ; AVX512-NEXT:    xorl %eax, %eax
157 ; AVX512-NEXT:    movntiq %rax, 8(%rdi)
158 ; AVX512-NEXT:    movntiq %rax, (%rdi)
159 ; AVX512-NEXT:    retq
185 ; AVX512-LABEL: test_zero_v4f64_align1:
186 ; AVX512:       # %bb.0:
187 ; AVX512-NEXT:    xorl %eax, %eax
188 ; AVX512-NEXT:    movntiq %rax, 8(%rdi)
189 ; AVX512-NEXT:    movntiq %rax, (%rdi)
190 ; AVX512-NEXT:    movntiq %rax, 24(%rdi)
191 ; AVX512-NEXT:    movntiq %rax, 16(%rdi)
192 ; AVX512-NEXT:    retq
235 ; AVX512-LABEL: test_zero_v8f32_align1:
236 ; AVX512:       # %bb.0:
237 ; AVX512-NEXT:    xorl %eax, %eax
238 ; AVX512-NEXT:    movntiq %rax, 8(%rdi)
239 ; AVX512-NEXT:    movntiq %rax, (%rdi)
240 ; AVX512-NEXT:    movntiq %rax, 24(%rdi)
241 ; AVX512-NEXT:    movntiq %rax, 16(%rdi)
242 ; AVX512-NEXT:    retq
284 ; AVX512-LABEL: test_zero_v4i64_align1:
285 ; AVX512:       # %bb.0:
286 ; AVX512-NEXT:    xorl %eax, %eax
287 ; AVX512-NEXT:    movntiq %rax, 8(%rdi)
288 ; AVX512-NEXT:    movntiq %rax, (%rdi)
289 ; AVX512-NEXT:    movntiq %rax, 24(%rdi)
290 ; AVX512-NEXT:    movntiq %rax, 16(%rdi)
291 ; AVX512-NEXT:    retq
333 ; AVX512-LABEL: test_zero_v8i32_align1:
334 ; AVX512:       # %bb.0:
335 ; AVX512-NEXT:    xorl %eax, %eax
336 ; AVX512-NEXT:    movntiq %rax, 8(%rdi)
337 ; AVX512-NEXT:    movntiq %rax, (%rdi)
338 ; AVX512-NEXT:    movntiq %rax, 24(%rdi)
339 ; AVX512-NEXT:    movntiq %rax, 16(%rdi)
340 ; AVX512-NEXT:    retq
382 ; AVX512-LABEL: test_zero_v16i16_align1:
383 ; AVX512:       # %bb.0:
384 ; AVX512-NEXT:    xorl %eax, %eax
385 ; AVX512-NEXT:    movntiq %rax, 8(%rdi)
386 ; AVX512-NEXT:    movntiq %rax, (%rdi)
387 ; AVX512-NEXT:    movntiq %rax, 24(%rdi)
388 ; AVX512-NEXT:    movntiq %rax, 16(%rdi)
389 ; AVX512-NEXT:    retq
431 ; AVX512-LABEL: test_zero_v32i8_align1:
432 ; AVX512:       # %bb.0:
433 ; AVX512-NEXT:    xorl %eax, %eax
434 ; AVX512-NEXT:    movntiq %rax, 8(%rdi)
435 ; AVX512-NEXT:    movntiq %rax, (%rdi)
436 ; AVX512-NEXT:    movntiq %rax, 24(%rdi)
437 ; AVX512-NEXT:    movntiq %rax, 16(%rdi)
438 ; AVX512-NEXT:    retq
458 ; AVX512-LABEL: test_zero_v4f64_align16:
459 ; AVX512:       # %bb.0:
460 ; AVX512-NEXT:    vxorps %xmm0, %xmm0, %xmm0
461 ; AVX512-NEXT:    vmovntps %xmm0, 16(%rdi)
462 ; AVX512-NEXT:    vmovntps %xmm0, (%rdi)
463 ; AVX512-NEXT:    retq
483 ; AVX512-LABEL: test_zero_v8f32_align16:
484 ; AVX512:       # %bb.0:
485 ; AVX512-NEXT:    vxorps %xmm0, %xmm0, %xmm0
486 ; AVX512-NEXT:    vmovntps %xmm0, 16(%rdi)
487 ; AVX512-NEXT:    vmovntps %xmm0, (%rdi)
488 ; AVX512-NEXT:    retq
508 ; AVX512-LABEL: test_zero_v4i64_align16:
509 ; AVX512:       # %bb.0:
510 ; AVX512-NEXT:    vxorps %xmm0, %xmm0, %xmm0
511 ; AVX512-NEXT:    vmovntps %xmm0, 16(%rdi)
512 ; AVX512-NEXT:    vmovntps %xmm0, (%rdi)
513 ; AVX512-NEXT:    retq
533 ; AVX512-LABEL: test_zero_v8i32_align16:
534 ; AVX512:       # %bb.0:
535 ; AVX512-NEXT:    vxorps %xmm0, %xmm0, %xmm0
536 ; AVX512-NEXT:    vmovntps %xmm0, 16(%rdi)
537 ; AVX512-NEXT:    vmovntps %xmm0, (%rdi)
538 ; AVX512-NEXT:    retq
558 ; AVX512-LABEL: test_zero_v16i16_align16:
559 ; AVX512:       # %bb.0:
560 ; AVX512-NEXT:    vxorps %xmm0, %xmm0, %xmm0
561 ; AVX512-NEXT:    vmovntps %xmm0, 16(%rdi)
562 ; AVX512-NEXT:    vmovntps %xmm0, (%rdi)
563 ; AVX512-NEXT:    retq
583 ; AVX512-LABEL: test_zero_v32i8_align16:
584 ; AVX512:       # %bb.0:
585 ; AVX512-NEXT:    vxorps %xmm0, %xmm0, %xmm0
586 ; AVX512-NEXT:    vmovntps %xmm0, 16(%rdi)
587 ; AVX512-NEXT:    vmovntps %xmm0, (%rdi)
588 ; AVX512-NEXT:    retq
622 ; AVX512-LABEL: test_zero_v8f64_align1:
623 ; AVX512:       # %bb.0:
624 ; AVX512-NEXT:    xorl %eax, %eax
625 ; AVX512-NEXT:    movntiq %rax, 24(%rdi)
626 ; AVX512-NEXT:    movntiq %rax, 16(%rdi)
627 ; AVX512-NEXT:    movntiq %rax, 8(%rdi)
628 ; AVX512-NEXT:    movntiq %rax, (%rdi)
629 ; AVX512-NEXT:    movntiq %rax, 56(%rdi)
630 ; AVX512-NEXT:    movntiq %rax, 48(%rdi)
631 ; AVX512-NEXT:    movntiq %rax, 40(%rdi)
632 ; AVX512-NEXT:    movntiq %rax, 32(%rdi)
633 ; AVX512-NEXT:    retq
692 ; AVX512-LABEL: test_zero_v16f32_align1:
693 ; AVX512:       # %bb.0:
694 ; AVX512-NEXT:    xorl %eax, %eax
695 ; AVX512-NEXT:    movntiq %rax, 24(%rdi)
696 ; AVX512-NEXT:    movntiq %rax, 16(%rdi)
697 ; AVX512-NEXT:    movntiq %rax, 8(%rdi)
698 ; AVX512-NEXT:    movntiq %rax, (%rdi)
699 ; AVX512-NEXT:    movntiq %rax, 56(%rdi)
700 ; AVX512-NEXT:    movntiq %rax, 48(%rdi)
701 ; AVX512-NEXT:    movntiq %rax, 40(%rdi)
702 ; AVX512-NEXT:    movntiq %rax, 32(%rdi)
703 ; AVX512-NEXT:    retq
761 ; AVX512-LABEL: test_zero_v8i64_align1:
762 ; AVX512:       # %bb.0:
763 ; AVX512-NEXT:    xorl %eax, %eax
764 ; AVX512-NEXT:    movntiq %rax, 24(%rdi)
765 ; AVX512-NEXT:    movntiq %rax, 16(%rdi)
766 ; AVX512-NEXT:    movntiq %rax, 8(%rdi)
767 ; AVX512-NEXT:    movntiq %rax, (%rdi)
768 ; AVX512-NEXT:    movntiq %rax, 56(%rdi)
769 ; AVX512-NEXT:    movntiq %rax, 48(%rdi)
770 ; AVX512-NEXT:    movntiq %rax, 40(%rdi)
771 ; AVX512-NEXT:    movntiq %rax, 32(%rdi)
772 ; AVX512-NEXT:    retq
830 ; AVX512-LABEL: test_zero_v16i32_align1:
831 ; AVX512:       # %bb.0:
832 ; AVX512-NEXT:    xorl %eax, %eax
833 ; AVX512-NEXT:    movntiq %rax, 24(%rdi)
834 ; AVX512-NEXT:    movntiq %rax, 16(%rdi)
835 ; AVX512-NEXT:    movntiq %rax, 8(%rdi)
836 ; AVX512-NEXT:    movntiq %rax, (%rdi)
837 ; AVX512-NEXT:    movntiq %rax, 56(%rdi)
838 ; AVX512-NEXT:    movntiq %rax, 48(%rdi)
839 ; AVX512-NEXT:    movntiq %rax, 40(%rdi)
840 ; AVX512-NEXT:    movntiq %rax, 32(%rdi)
841 ; AVX512-NEXT:    retq
899 ; AVX512-LABEL: test_zero_v32i16_align1:
900 ; AVX512:       # %bb.0:
901 ; AVX512-NEXT:    xorl %eax, %eax
902 ; AVX512-NEXT:    movntiq %rax, 24(%rdi)
903 ; AVX512-NEXT:    movntiq %rax, 16(%rdi)
904 ; AVX512-NEXT:    movntiq %rax, 8(%rdi)
905 ; AVX512-NEXT:    movntiq %rax, (%rdi)
906 ; AVX512-NEXT:    movntiq %rax, 56(%rdi)
907 ; AVX512-NEXT:    movntiq %rax, 48(%rdi)
908 ; AVX512-NEXT:    movntiq %rax, 40(%rdi)
909 ; AVX512-NEXT:    movntiq %rax, 32(%rdi)
910 ; AVX512-NEXT:    retq
968 ; AVX512-LABEL: test_zero_v64i8_align1:
969 ; AVX512:       # %bb.0:
970 ; AVX512-NEXT:    xorl %eax, %eax
971 ; AVX512-NEXT:    movntiq %rax, 24(%rdi)
972 ; AVX512-NEXT:    movntiq %rax, 16(%rdi)
973 ; AVX512-NEXT:    movntiq %rax, 8(%rdi)
974 ; AVX512-NEXT:    movntiq %rax, (%rdi)
975 ; AVX512-NEXT:    movntiq %rax, 56(%rdi)
976 ; AVX512-NEXT:    movntiq %rax, 48(%rdi)
977 ; AVX512-NEXT:    movntiq %rax, 40(%rdi)
978 ; AVX512-NEXT:    movntiq %rax, 32(%rdi)
979 ; AVX512-NEXT:    retq
1003 ; AVX512-LABEL: test_zero_v8f64_align16:
1004 ; AVX512:       # %bb.0:
1005 ; AVX512-NEXT:    vxorps %xmm0, %xmm0, %xmm0
1006 ; AVX512-NEXT:    vmovntps %xmm0, 16(%rdi)
1007 ; AVX512-NEXT:    vmovntps %xmm0, (%rdi)
1008 ; AVX512-NEXT:    vmovntps %xmm0, 48(%rdi)
1009 ; AVX512-NEXT:    vmovntps %xmm0, 32(%rdi)
1010 ; AVX512-NEXT:    retq
1034 ; AVX512-LABEL: test_zero_v16f32_align16:
1035 ; AVX512:       # %bb.0:
1036 ; AVX512-NEXT:    vxorps %xmm0, %xmm0, %xmm0
1037 ; AVX512-NEXT:    vmovntps %xmm0, 16(%rdi)
1038 ; AVX512-NEXT:    vmovntps %xmm0, (%rdi)
1039 ; AVX512-NEXT:    vmovntps %xmm0, 48(%rdi)
1040 ; AVX512-NEXT:    vmovntps %xmm0, 32(%rdi)
1041 ; AVX512-NEXT:    retq
1065 ; AVX512-LABEL: test_zero_v8i64_align16:
1066 ; AVX512:       # %bb.0:
1067 ; AVX512-NEXT:    vxorps %xmm0, %xmm0, %xmm0
1068 ; AVX512-NEXT:    vmovntps %xmm0, 16(%rdi)
1069 ; AVX512-NEXT:    vmovntps %xmm0, (%rdi)
1070 ; AVX512-NEXT:    vmovntps %xmm0, 48(%rdi)
1071 ; AVX512-NEXT:    vmovntps %xmm0, 32(%rdi)
1072 ; AVX512-NEXT:    retq
1096 ; AVX512-LABEL: test_zero_v16i32_align16:
1097 ; AVX512:       # %bb.0:
1098 ; AVX512-NEXT:    vxorps %xmm0, %xmm0, %xmm0
1099 ; AVX512-NEXT:    vmovntps %xmm0, 16(%rdi)
1100 ; AVX512-NEXT:    vmovntps %xmm0, (%rdi)
1101 ; AVX512-NEXT:    vmovntps %xmm0, 48(%rdi)
1102 ; AVX512-NEXT:    vmovntps %xmm0, 32(%rdi)
1103 ; AVX512-NEXT:    retq
1127 ; AVX512-LABEL: test_zero_v32i16_align16:
1128 ; AVX512:       # %bb.0:
1129 ; AVX512-NEXT:    vxorps %xmm0, %xmm0, %xmm0
1130 ; AVX512-NEXT:    vmovntps %xmm0, 16(%rdi)
1131 ; AVX512-NEXT:    vmovntps %xmm0, (%rdi)
1132 ; AVX512-NEXT:    vmovntps %xmm0, 48(%rdi)
1133 ; AVX512-NEXT:    vmovntps %xmm0, 32(%rdi)
1134 ; AVX512-NEXT:    retq
1158 ; AVX512-LABEL: test_zero_v64i8_align16:
1159 ; AVX512:       # %bb.0:
1160 ; AVX512-NEXT:    vxorps %xmm0, %xmm0, %xmm0
1161 ; AVX512-NEXT:    vmovntps %xmm0, 16(%rdi)
1162 ; AVX512-NEXT:    vmovntps %xmm0, (%rdi)
1163 ; AVX512-NEXT:    vmovntps %xmm0, 48(%rdi)
1164 ; AVX512-NEXT:    vmovntps %xmm0, 32(%rdi)
1165 ; AVX512-NEXT:    retq
1188 ; AVX512-LABEL: test_zero_v8f64_align32:
1189 ; AVX512:       # %bb.0:
1190 ; AVX512-NEXT:    vxorps %xmm0, %xmm0, %xmm0
1191 ; AVX512-NEXT:    vmovntps %ymm0, 32(%rdi)
1192 ; AVX512-NEXT:    vmovntps %ymm0, (%rdi)
1193 ; AVX512-NEXT:    vzeroupper
1194 ; AVX512-NEXT:    retq
1217 ; AVX512-LABEL: test_zero_v16f32_align32:
1218 ; AVX512:       # %bb.0:
1219 ; AVX512-NEXT:    vxorps %xmm0, %xmm0, %xmm0
1220 ; AVX512-NEXT:    vmovntps %ymm0, 32(%rdi)
1221 ; AVX512-NEXT:    vmovntps %ymm0, (%rdi)
1222 ; AVX512-NEXT:    vzeroupper
1223 ; AVX512-NEXT:    retq
1246 ; AVX512-LABEL: test_zero_v8i64_align32:
1247 ; AVX512:       # %bb.0:
1248 ; AVX512-NEXT:    vxorps %xmm0, %xmm0, %xmm0
1249 ; AVX512-NEXT:    vmovntps %ymm0, 32(%rdi)
1250 ; AVX512-NEXT:    vmovntps %ymm0, (%rdi)
1251 ; AVX512-NEXT:    vzeroupper
1252 ; AVX512-NEXT:    retq
1275 ; AVX512-LABEL: test_zero_v16i32_align32:
1276 ; AVX512:       # %bb.0:
1277 ; AVX512-NEXT:    vxorps %xmm0, %xmm0, %xmm0
1278 ; AVX512-NEXT:    vmovntps %ymm0, 32(%rdi)
1279 ; AVX512-NEXT:    vmovntps %ymm0, (%rdi)
1280 ; AVX512-NEXT:    vzeroupper
1281 ; AVX512-NEXT:    retq
1304 ; AVX512-LABEL: test_zero_v32i16_align32:
1305 ; AVX512:       # %bb.0:
1306 ; AVX512-NEXT:    vxorps %xmm0, %xmm0, %xmm0
1307 ; AVX512-NEXT:    vmovntps %ymm0, 32(%rdi)
1308 ; AVX512-NEXT:    vmovntps %ymm0, (%rdi)
1309 ; AVX512-NEXT:    vzeroupper
1310 ; AVX512-NEXT:    retq
1333 ; AVX512-LABEL: test_zero_v64i8_align32:
1334 ; AVX512:       # %bb.0:
1335 ; AVX512-NEXT:    vxorps %xmm0, %xmm0, %xmm0
1336 ; AVX512-NEXT:    vmovntps %ymm0, 32(%rdi)
1337 ; AVX512-NEXT:    vmovntps %ymm0, (%rdi)
1338 ; AVX512-NEXT:    vzeroupper
1339 ; AVX512-NEXT:    retq