brintos

brintos / llvm-project-archived public Read only

0
0
Text · 89.6 KiB · 8878801 Raw
1848 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --no_x86_scrub_sp --no_x86_scrub_mem_shuffle2; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f | FileCheck %s --check-prefixes=AVX512,AVX512F3; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512bw | FileCheck %s --check-prefixes=AVX512,AVX512BW4; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512bw,+avx512vbmi | FileCheck %s --check-prefixes=AVX512,AVX512VBMI5 6define <8 x i64> @var_shuffle_v8i64(<8 x i64> %v, <8 x i64> %indices) nounwind {7; AVX512-LABEL: var_shuffle_v8i64:8; AVX512:       # %bb.0:9; AVX512-NEXT:    vpermpd %zmm0, %zmm1, %zmm010; AVX512-NEXT:    retq11  %index0 = extractelement <8 x i64> %indices, i32 012  %index1 = extractelement <8 x i64> %indices, i32 113  %index2 = extractelement <8 x i64> %indices, i32 214  %index3 = extractelement <8 x i64> %indices, i32 315  %index4 = extractelement <8 x i64> %indices, i32 416  %index5 = extractelement <8 x i64> %indices, i32 517  %index6 = extractelement <8 x i64> %indices, i32 618  %index7 = extractelement <8 x i64> %indices, i32 719  %v0 = extractelement <8 x i64> %v, i64 %index020  %v1 = extractelement <8 x i64> %v, i64 %index121  %v2 = extractelement <8 x i64> %v, i64 %index222  %v3 = extractelement <8 x i64> %v, i64 %index323  %v4 = extractelement <8 x i64> %v, i64 %index424  %v5 = extractelement <8 x i64> %v, i64 %index525  %v6 = extractelement <8 x i64> %v, i64 %index626  %v7 = extractelement <8 x i64> %v, i64 %index727  %ret0 = insertelement <8 x i64> undef, i64 %v0, i32 028  %ret1 = insertelement <8 x i64> %ret0, i64 %v1, i32 129  %ret2 = insertelement <8 x i64> %ret1, i64 %v2, i32 230  %ret3 = insertelement <8 x i64> %ret2, i64 %v3, i32 331  %ret4 = insertelement <8 x i64> %ret3, i64 %v4, i32 432  %ret5 = insertelement <8 x i64> %ret4, i64 %v5, i32 533  %ret6 = insertelement <8 x i64> %ret5, i64 %v6, i32 634  %ret7 = insertelement <8 x i64> %ret6, i64 %v7, i32 735  ret <8 x i64> %ret736}37 38define <16 x i32> @var_shuffle_v16i32(<16 x i32> %v, <16 x i32> %indices) nounwind {39; AVX512-LABEL: var_shuffle_v16i32:40; AVX512:       # %bb.0:41; AVX512-NEXT:    vpermps %zmm0, %zmm1, %zmm042; AVX512-NEXT:    retq43  %index0 = extractelement <16 x i32> %indices, i32 044  %index1 = extractelement <16 x i32> %indices, i32 145  %index2 = extractelement <16 x i32> %indices, i32 246  %index3 = extractelement <16 x i32> %indices, i32 347  %index4 = extractelement <16 x i32> %indices, i32 448  %index5 = extractelement <16 x i32> %indices, i32 549  %index6 = extractelement <16 x i32> %indices, i32 650  %index7 = extractelement <16 x i32> %indices, i32 751  %index8 = extractelement <16 x i32> %indices, i32 852  %index9 = extractelement <16 x i32> %indices, i32 953  %index10 = extractelement <16 x i32> %indices, i32 1054  %index11 = extractelement <16 x i32> %indices, i32 1155  %index12 = extractelement <16 x i32> %indices, i32 1256  %index13 = extractelement <16 x i32> %indices, i32 1357  %index14 = extractelement <16 x i32> %indices, i32 1458  %index15 = extractelement <16 x i32> %indices, i32 1559  %v0 = extractelement <16 x i32> %v, i32 %index060  %v1 = extractelement <16 x i32> %v, i32 %index161  %v2 = extractelement <16 x i32> %v, i32 %index262  %v3 = extractelement <16 x i32> %v, i32 %index363  %v4 = extractelement <16 x i32> %v, i32 %index464  %v5 = extractelement <16 x i32> %v, i32 %index565  %v6 = extractelement <16 x i32> %v, i32 %index666  %v7 = extractelement <16 x i32> %v, i32 %index767  %v8 = extractelement <16 x i32> %v, i32 %index868  %v9 = extractelement <16 x i32> %v, i32 %index969  %v10 = extractelement <16 x i32> %v, i32 %index1070  %v11 = extractelement <16 x i32> %v, i32 %index1171  %v12 = extractelement <16 x i32> %v, i32 %index1272  %v13 = extractelement <16 x i32> %v, i32 %index1373  %v14 = extractelement <16 x i32> %v, i32 %index1474  %v15 = extractelement <16 x i32> %v, i32 %index1575  %ret0 = insertelement <16 x i32> undef, i32 %v0, i32 076  %ret1 = insertelement <16 x i32> %ret0, i32 %v1, i32 177  %ret2 = insertelement <16 x i32> %ret1, i32 %v2, i32 278  %ret3 = insertelement <16 x i32> %ret2, i32 %v3, i32 379  %ret4 = insertelement <16 x i32> %ret3, i32 %v4, i32 480  %ret5 = insertelement <16 x i32> %ret4, i32 %v5, i32 581  %ret6 = insertelement <16 x i32> %ret5, i32 %v6, i32 682  %ret7 = insertelement <16 x i32> %ret6, i32 %v7, i32 783  %ret8 = insertelement <16 x i32> %ret7, i32 %v8, i32 884  %ret9 = insertelement <16 x i32> %ret8, i32 %v9, i32 985  %ret10 = insertelement <16 x i32> %ret9, i32 %v10, i32 1086  %ret11 = insertelement <16 x i32> %ret10, i32 %v11, i32 1187  %ret12 = insertelement <16 x i32> %ret11, i32 %v12, i32 1288  %ret13 = insertelement <16 x i32> %ret12, i32 %v13, i32 1389  %ret14 = insertelement <16 x i32> %ret13, i32 %v14, i32 1490  %ret15 = insertelement <16 x i32> %ret14, i32 %v15, i32 1591  ret <16 x i32> %ret1592}93 94define <32 x i16> @var_shuffle_v32i16(<32 x i16> %v, <32 x i16> %indices) nounwind {95; AVX512F-LABEL: var_shuffle_v32i16:96; AVX512F:       # %bb.0:97; AVX512F-NEXT:    pushq %rbp98; AVX512F-NEXT:    movq %rsp, %rbp99; AVX512F-NEXT:    andq $-64, %rsp100; AVX512F-NEXT:    subq $128, %rsp101; AVX512F-NEXT:    vextracti128 $1, %ymm1, %xmm2102; AVX512F-NEXT:    vextracti32x4 $2, %zmm1, %xmm3103; AVX512F-NEXT:    vextracti32x4 $3, %zmm1, %xmm4104; AVX512F-NEXT:    vpextrw $0, %xmm4, %eax105; AVX512F-NEXT:    vmovaps %zmm0, (%rsp)106; AVX512F-NEXT:    andl $31, %eax107; AVX512F-NEXT:    movzwl (%rsp,%rax,2), %eax108; AVX512F-NEXT:    vmovd %eax, %xmm0109; AVX512F-NEXT:    vpextrw $1, %xmm4, %eax110; AVX512F-NEXT:    andl $31, %eax111; AVX512F-NEXT:    vpinsrw $1, (%rsp,%rax,2), %xmm0, %xmm0112; AVX512F-NEXT:    vpextrw $2, %xmm4, %eax113; AVX512F-NEXT:    andl $31, %eax114; AVX512F-NEXT:    vpinsrw $2, (%rsp,%rax,2), %xmm0, %xmm0115; AVX512F-NEXT:    vpextrw $3, %xmm4, %eax116; AVX512F-NEXT:    andl $31, %eax117; AVX512F-NEXT:    vpinsrw $3, (%rsp,%rax,2), %xmm0, %xmm0118; AVX512F-NEXT:    vpextrw $4, %xmm4, %eax119; AVX512F-NEXT:    andl $31, %eax120; AVX512F-NEXT:    vpinsrw $4, (%rsp,%rax,2), %xmm0, %xmm0121; AVX512F-NEXT:    vpextrw $5, %xmm4, %eax122; AVX512F-NEXT:    andl $31, %eax123; AVX512F-NEXT:    vpinsrw $5, (%rsp,%rax,2), %xmm0, %xmm0124; AVX512F-NEXT:    vpextrw $6, %xmm4, %eax125; AVX512F-NEXT:    andl $31, %eax126; AVX512F-NEXT:    vpinsrw $6, (%rsp,%rax,2), %xmm0, %xmm0127; AVX512F-NEXT:    vpextrw $7, %xmm4, %eax128; AVX512F-NEXT:    andl $31, %eax129; AVX512F-NEXT:    vpinsrw $7, (%rsp,%rax,2), %xmm0, %xmm0130; AVX512F-NEXT:    vpextrw $0, %xmm3, %eax131; AVX512F-NEXT:    andl $31, %eax132; AVX512F-NEXT:    movzwl (%rsp,%rax,2), %eax133; AVX512F-NEXT:    vmovd %eax, %xmm4134; AVX512F-NEXT:    vpextrw $1, %xmm3, %eax135; AVX512F-NEXT:    andl $31, %eax136; AVX512F-NEXT:    vpinsrw $1, (%rsp,%rax,2), %xmm4, %xmm4137; AVX512F-NEXT:    vpextrw $2, %xmm3, %eax138; AVX512F-NEXT:    andl $31, %eax139; AVX512F-NEXT:    vpinsrw $2, (%rsp,%rax,2), %xmm4, %xmm4140; AVX512F-NEXT:    vpextrw $3, %xmm3, %eax141; AVX512F-NEXT:    andl $31, %eax142; AVX512F-NEXT:    vpinsrw $3, (%rsp,%rax,2), %xmm4, %xmm4143; AVX512F-NEXT:    vpextrw $4, %xmm3, %eax144; AVX512F-NEXT:    andl $31, %eax145; AVX512F-NEXT:    vpinsrw $4, (%rsp,%rax,2), %xmm4, %xmm4146; AVX512F-NEXT:    vpextrw $5, %xmm3, %eax147; AVX512F-NEXT:    andl $31, %eax148; AVX512F-NEXT:    vpinsrw $5, (%rsp,%rax,2), %xmm4, %xmm4149; AVX512F-NEXT:    vpextrw $6, %xmm3, %eax150; AVX512F-NEXT:    andl $31, %eax151; AVX512F-NEXT:    vpinsrw $6, (%rsp,%rax,2), %xmm4, %xmm4152; AVX512F-NEXT:    vpextrw $7, %xmm3, %eax153; AVX512F-NEXT:    andl $31, %eax154; AVX512F-NEXT:    vpinsrw $7, (%rsp,%rax,2), %xmm4, %xmm3155; AVX512F-NEXT:    vpextrw $0, %xmm2, %eax156; AVX512F-NEXT:    andl $31, %eax157; AVX512F-NEXT:    movzwl (%rsp,%rax,2), %eax158; AVX512F-NEXT:    vmovd %eax, %xmm4159; AVX512F-NEXT:    vpextrw $1, %xmm2, %eax160; AVX512F-NEXT:    andl $31, %eax161; AVX512F-NEXT:    vpinsrw $1, (%rsp,%rax,2), %xmm4, %xmm4162; AVX512F-NEXT:    vpextrw $2, %xmm2, %eax163; AVX512F-NEXT:    andl $31, %eax164; AVX512F-NEXT:    vpinsrw $2, (%rsp,%rax,2), %xmm4, %xmm4165; AVX512F-NEXT:    vpextrw $3, %xmm2, %eax166; AVX512F-NEXT:    andl $31, %eax167; AVX512F-NEXT:    vpinsrw $3, (%rsp,%rax,2), %xmm4, %xmm4168; AVX512F-NEXT:    vpextrw $4, %xmm2, %eax169; AVX512F-NEXT:    andl $31, %eax170; AVX512F-NEXT:    vpinsrw $4, (%rsp,%rax,2), %xmm4, %xmm4171; AVX512F-NEXT:    vpextrw $5, %xmm2, %eax172; AVX512F-NEXT:    andl $31, %eax173; AVX512F-NEXT:    movzwl (%rsp,%rax,2), %eax174; AVX512F-NEXT:    vpinsrw $5, %eax, %xmm4, %xmm4175; AVX512F-NEXT:    vpextrw $6, %xmm2, %eax176; AVX512F-NEXT:    andl $31, %eax177; AVX512F-NEXT:    movzwl (%rsp,%rax,2), %eax178; AVX512F-NEXT:    vpinsrw $6, %eax, %xmm4, %xmm4179; AVX512F-NEXT:    vpextrw $7, %xmm2, %eax180; AVX512F-NEXT:    andl $31, %eax181; AVX512F-NEXT:    movzwl (%rsp,%rax,2), %eax182; AVX512F-NEXT:    vpinsrw $7, %eax, %xmm4, %xmm2183; AVX512F-NEXT:    vpextrw $0, %xmm1, %eax184; AVX512F-NEXT:    andl $31, %eax185; AVX512F-NEXT:    movzwl (%rsp,%rax,2), %eax186; AVX512F-NEXT:    vmovd %eax, %xmm4187; AVX512F-NEXT:    vpextrw $1, %xmm1, %eax188; AVX512F-NEXT:    andl $31, %eax189; AVX512F-NEXT:    vpinsrw $1, (%rsp,%rax,2), %xmm4, %xmm4190; AVX512F-NEXT:    vpextrw $2, %xmm1, %eax191; AVX512F-NEXT:    andl $31, %eax192; AVX512F-NEXT:    vpinsrw $2, (%rsp,%rax,2), %xmm4, %xmm4193; AVX512F-NEXT:    vpextrw $3, %xmm1, %eax194; AVX512F-NEXT:    andl $31, %eax195; AVX512F-NEXT:    vpinsrw $3, (%rsp,%rax,2), %xmm4, %xmm4196; AVX512F-NEXT:    vpextrw $4, %xmm1, %eax197; AVX512F-NEXT:    andl $31, %eax198; AVX512F-NEXT:    vpinsrw $4, (%rsp,%rax,2), %xmm4, %xmm4199; AVX512F-NEXT:    vpextrw $5, %xmm1, %eax200; AVX512F-NEXT:    andl $31, %eax201; AVX512F-NEXT:    vpinsrw $5, (%rsp,%rax,2), %xmm4, %xmm4202; AVX512F-NEXT:    vpextrw $6, %xmm1, %eax203; AVX512F-NEXT:    andl $31, %eax204; AVX512F-NEXT:    vpinsrw $6, (%rsp,%rax,2), %xmm4, %xmm4205; AVX512F-NEXT:    vpextrw $7, %xmm1, %eax206; AVX512F-NEXT:    andl $31, %eax207; AVX512F-NEXT:    vpinsrw $7, (%rsp,%rax,2), %xmm4, %xmm1208; AVX512F-NEXT:    vinserti128 $1, %xmm0, %ymm3, %ymm0209; AVX512F-NEXT:    vinserti128 $1, %xmm2, %ymm1, %ymm1210; AVX512F-NEXT:    vinserti64x4 $1, %ymm0, %zmm1, %zmm0211; AVX512F-NEXT:    movq %rbp, %rsp212; AVX512F-NEXT:    popq %rbp213; AVX512F-NEXT:    retq214;215; AVX512BW-LABEL: var_shuffle_v32i16:216; AVX512BW:       # %bb.0:217; AVX512BW-NEXT:    vpermw %zmm0, %zmm1, %zmm0218; AVX512BW-NEXT:    retq219;220; AVX512VBMI-LABEL: var_shuffle_v32i16:221; AVX512VBMI:       # %bb.0:222; AVX512VBMI-NEXT:    vpermw %zmm0, %zmm1, %zmm0223; AVX512VBMI-NEXT:    retq224  %index0 = extractelement <32 x i16> %indices, i32 0225  %index1 = extractelement <32 x i16> %indices, i32 1226  %index2 = extractelement <32 x i16> %indices, i32 2227  %index3 = extractelement <32 x i16> %indices, i32 3228  %index4 = extractelement <32 x i16> %indices, i32 4229  %index5 = extractelement <32 x i16> %indices, i32 5230  %index6 = extractelement <32 x i16> %indices, i32 6231  %index7 = extractelement <32 x i16> %indices, i32 7232  %index8 = extractelement <32 x i16> %indices, i32 8233  %index9 = extractelement <32 x i16> %indices, i32 9234  %index10 = extractelement <32 x i16> %indices, i32 10235  %index11 = extractelement <32 x i16> %indices, i32 11236  %index12 = extractelement <32 x i16> %indices, i32 12237  %index13 = extractelement <32 x i16> %indices, i32 13238  %index14 = extractelement <32 x i16> %indices, i32 14239  %index15 = extractelement <32 x i16> %indices, i32 15240  %index16 = extractelement <32 x i16> %indices, i32 16241  %index17 = extractelement <32 x i16> %indices, i32 17242  %index18 = extractelement <32 x i16> %indices, i32 18243  %index19 = extractelement <32 x i16> %indices, i32 19244  %index20 = extractelement <32 x i16> %indices, i32 20245  %index21 = extractelement <32 x i16> %indices, i32 21246  %index22 = extractelement <32 x i16> %indices, i32 22247  %index23 = extractelement <32 x i16> %indices, i32 23248  %index24 = extractelement <32 x i16> %indices, i32 24249  %index25 = extractelement <32 x i16> %indices, i32 25250  %index26 = extractelement <32 x i16> %indices, i32 26251  %index27 = extractelement <32 x i16> %indices, i32 27252  %index28 = extractelement <32 x i16> %indices, i32 28253  %index29 = extractelement <32 x i16> %indices, i32 29254  %index30 = extractelement <32 x i16> %indices, i32 30255  %index31 = extractelement <32 x i16> %indices, i32 31256  %v0 = extractelement <32 x i16> %v, i16 %index0257  %v1 = extractelement <32 x i16> %v, i16 %index1258  %v2 = extractelement <32 x i16> %v, i16 %index2259  %v3 = extractelement <32 x i16> %v, i16 %index3260  %v4 = extractelement <32 x i16> %v, i16 %index4261  %v5 = extractelement <32 x i16> %v, i16 %index5262  %v6 = extractelement <32 x i16> %v, i16 %index6263  %v7 = extractelement <32 x i16> %v, i16 %index7264  %v8 = extractelement <32 x i16> %v, i16 %index8265  %v9 = extractelement <32 x i16> %v, i16 %index9266  %v10 = extractelement <32 x i16> %v, i16 %index10267  %v11 = extractelement <32 x i16> %v, i16 %index11268  %v12 = extractelement <32 x i16> %v, i16 %index12269  %v13 = extractelement <32 x i16> %v, i16 %index13270  %v14 = extractelement <32 x i16> %v, i16 %index14271  %v15 = extractelement <32 x i16> %v, i16 %index15272  %v16 = extractelement <32 x i16> %v, i16 %index16273  %v17 = extractelement <32 x i16> %v, i16 %index17274  %v18 = extractelement <32 x i16> %v, i16 %index18275  %v19 = extractelement <32 x i16> %v, i16 %index19276  %v20 = extractelement <32 x i16> %v, i16 %index20277  %v21 = extractelement <32 x i16> %v, i16 %index21278  %v22 = extractelement <32 x i16> %v, i16 %index22279  %v23 = extractelement <32 x i16> %v, i16 %index23280  %v24 = extractelement <32 x i16> %v, i16 %index24281  %v25 = extractelement <32 x i16> %v, i16 %index25282  %v26 = extractelement <32 x i16> %v, i16 %index26283  %v27 = extractelement <32 x i16> %v, i16 %index27284  %v28 = extractelement <32 x i16> %v, i16 %index28285  %v29 = extractelement <32 x i16> %v, i16 %index29286  %v30 = extractelement <32 x i16> %v, i16 %index30287  %v31 = extractelement <32 x i16> %v, i16 %index31288  %ret0 = insertelement <32 x i16> undef, i16 %v0, i32 0289  %ret1 = insertelement <32 x i16> %ret0, i16 %v1, i32 1290  %ret2 = insertelement <32 x i16> %ret1, i16 %v2, i32 2291  %ret3 = insertelement <32 x i16> %ret2, i16 %v3, i32 3292  %ret4 = insertelement <32 x i16> %ret3, i16 %v4, i32 4293  %ret5 = insertelement <32 x i16> %ret4, i16 %v5, i32 5294  %ret6 = insertelement <32 x i16> %ret5, i16 %v6, i32 6295  %ret7 = insertelement <32 x i16> %ret6, i16 %v7, i32 7296  %ret8 = insertelement <32 x i16> %ret7, i16 %v8, i32 8297  %ret9 = insertelement <32 x i16> %ret8, i16 %v9, i32 9298  %ret10 = insertelement <32 x i16> %ret9, i16 %v10, i32 10299  %ret11 = insertelement <32 x i16> %ret10, i16 %v11, i32 11300  %ret12 = insertelement <32 x i16> %ret11, i16 %v12, i32 12301  %ret13 = insertelement <32 x i16> %ret12, i16 %v13, i32 13302  %ret14 = insertelement <32 x i16> %ret13, i16 %v14, i32 14303  %ret15 = insertelement <32 x i16> %ret14, i16 %v15, i32 15304  %ret16 = insertelement <32 x i16> %ret15, i16 %v16, i32 16305  %ret17 = insertelement <32 x i16> %ret16, i16 %v17, i32 17306  %ret18 = insertelement <32 x i16> %ret17, i16 %v18, i32 18307  %ret19 = insertelement <32 x i16> %ret18, i16 %v19, i32 19308  %ret20 = insertelement <32 x i16> %ret19, i16 %v20, i32 20309  %ret21 = insertelement <32 x i16> %ret20, i16 %v21, i32 21310  %ret22 = insertelement <32 x i16> %ret21, i16 %v22, i32 22311  %ret23 = insertelement <32 x i16> %ret22, i16 %v23, i32 23312  %ret24 = insertelement <32 x i16> %ret23, i16 %v24, i32 24313  %ret25 = insertelement <32 x i16> %ret24, i16 %v25, i32 25314  %ret26 = insertelement <32 x i16> %ret25, i16 %v26, i32 26315  %ret27 = insertelement <32 x i16> %ret26, i16 %v27, i32 27316  %ret28 = insertelement <32 x i16> %ret27, i16 %v28, i32 28317  %ret29 = insertelement <32 x i16> %ret28, i16 %v29, i32 29318  %ret30 = insertelement <32 x i16> %ret29, i16 %v30, i32 30319  %ret31 = insertelement <32 x i16> %ret30, i16 %v31, i32 31320  ret <32 x i16> %ret31321}322 323define <64 x i8> @var_shuffle_v64i8(<64 x i8> %v, <64 x i8> %indices) nounwind {324; AVX512F-LABEL: var_shuffle_v64i8:325; AVX512F:       # %bb.0:326; AVX512F-NEXT:    pushq %rbp327; AVX512F-NEXT:    movq %rsp, %rbp328; AVX512F-NEXT:    andq $-64, %rsp329; AVX512F-NEXT:    subq $128, %rsp330; AVX512F-NEXT:    vextracti128 $1, %ymm1, %xmm2331; AVX512F-NEXT:    vextracti32x4 $2, %zmm1, %xmm3332; AVX512F-NEXT:    vextracti32x4 $3, %zmm1, %xmm4333; AVX512F-NEXT:    vpextrb $0, %xmm4, %eax334; AVX512F-NEXT:    vmovaps %zmm0, (%rsp)335; AVX512F-NEXT:    andl $63, %eax336; AVX512F-NEXT:    movzbl (%rsp,%rax), %eax337; AVX512F-NEXT:    vmovd %eax, %xmm0338; AVX512F-NEXT:    vpextrb $1, %xmm4, %eax339; AVX512F-NEXT:    andl $63, %eax340; AVX512F-NEXT:    vpinsrb $1, (%rsp,%rax), %xmm0, %xmm0341; AVX512F-NEXT:    vpextrb $2, %xmm4, %eax342; AVX512F-NEXT:    andl $63, %eax343; AVX512F-NEXT:    vpinsrb $2, (%rsp,%rax), %xmm0, %xmm0344; AVX512F-NEXT:    vpextrb $3, %xmm4, %eax345; AVX512F-NEXT:    andl $63, %eax346; AVX512F-NEXT:    vpinsrb $3, (%rsp,%rax), %xmm0, %xmm0347; AVX512F-NEXT:    vpextrb $4, %xmm4, %eax348; AVX512F-NEXT:    andl $63, %eax349; AVX512F-NEXT:    vpinsrb $4, (%rsp,%rax), %xmm0, %xmm0350; AVX512F-NEXT:    vpextrb $5, %xmm4, %eax351; AVX512F-NEXT:    andl $63, %eax352; AVX512F-NEXT:    vpinsrb $5, (%rsp,%rax), %xmm0, %xmm0353; AVX512F-NEXT:    vpextrb $6, %xmm4, %eax354; AVX512F-NEXT:    andl $63, %eax355; AVX512F-NEXT:    vpinsrb $6, (%rsp,%rax), %xmm0, %xmm0356; AVX512F-NEXT:    vpextrb $7, %xmm4, %eax357; AVX512F-NEXT:    andl $63, %eax358; AVX512F-NEXT:    vpinsrb $7, (%rsp,%rax), %xmm0, %xmm0359; AVX512F-NEXT:    vpextrb $8, %xmm4, %eax360; AVX512F-NEXT:    andl $63, %eax361; AVX512F-NEXT:    vpinsrb $8, (%rsp,%rax), %xmm0, %xmm0362; AVX512F-NEXT:    vpextrb $9, %xmm4, %eax363; AVX512F-NEXT:    andl $63, %eax364; AVX512F-NEXT:    vpinsrb $9, (%rsp,%rax), %xmm0, %xmm0365; AVX512F-NEXT:    vpextrb $10, %xmm4, %eax366; AVX512F-NEXT:    andl $63, %eax367; AVX512F-NEXT:    vpinsrb $10, (%rsp,%rax), %xmm0, %xmm0368; AVX512F-NEXT:    vpextrb $11, %xmm4, %eax369; AVX512F-NEXT:    andl $63, %eax370; AVX512F-NEXT:    vpinsrb $11, (%rsp,%rax), %xmm0, %xmm0371; AVX512F-NEXT:    vpextrb $12, %xmm4, %eax372; AVX512F-NEXT:    andl $63, %eax373; AVX512F-NEXT:    vpinsrb $12, (%rsp,%rax), %xmm0, %xmm0374; AVX512F-NEXT:    vpextrb $13, %xmm4, %eax375; AVX512F-NEXT:    andl $63, %eax376; AVX512F-NEXT:    vpinsrb $13, (%rsp,%rax), %xmm0, %xmm0377; AVX512F-NEXT:    vpextrb $14, %xmm4, %eax378; AVX512F-NEXT:    andl $63, %eax379; AVX512F-NEXT:    vpinsrb $14, (%rsp,%rax), %xmm0, %xmm0380; AVX512F-NEXT:    vpextrb $15, %xmm4, %eax381; AVX512F-NEXT:    andl $63, %eax382; AVX512F-NEXT:    vpinsrb $15, (%rsp,%rax), %xmm0, %xmm0383; AVX512F-NEXT:    vpextrb $0, %xmm3, %eax384; AVX512F-NEXT:    andl $63, %eax385; AVX512F-NEXT:    movzbl (%rsp,%rax), %eax386; AVX512F-NEXT:    vmovd %eax, %xmm4387; AVX512F-NEXT:    vpextrb $1, %xmm3, %eax388; AVX512F-NEXT:    andl $63, %eax389; AVX512F-NEXT:    vpinsrb $1, (%rsp,%rax), %xmm4, %xmm4390; AVX512F-NEXT:    vpextrb $2, %xmm3, %eax391; AVX512F-NEXT:    andl $63, %eax392; AVX512F-NEXT:    vpinsrb $2, (%rsp,%rax), %xmm4, %xmm4393; AVX512F-NEXT:    vpextrb $3, %xmm3, %eax394; AVX512F-NEXT:    andl $63, %eax395; AVX512F-NEXT:    vpinsrb $3, (%rsp,%rax), %xmm4, %xmm4396; AVX512F-NEXT:    vpextrb $4, %xmm3, %eax397; AVX512F-NEXT:    andl $63, %eax398; AVX512F-NEXT:    vpinsrb $4, (%rsp,%rax), %xmm4, %xmm4399; AVX512F-NEXT:    vpextrb $5, %xmm3, %eax400; AVX512F-NEXT:    andl $63, %eax401; AVX512F-NEXT:    vpinsrb $5, (%rsp,%rax), %xmm4, %xmm4402; AVX512F-NEXT:    vpextrb $6, %xmm3, %eax403; AVX512F-NEXT:    andl $63, %eax404; AVX512F-NEXT:    vpinsrb $6, (%rsp,%rax), %xmm4, %xmm4405; AVX512F-NEXT:    vpextrb $7, %xmm3, %eax406; AVX512F-NEXT:    andl $63, %eax407; AVX512F-NEXT:    vpinsrb $7, (%rsp,%rax), %xmm4, %xmm4408; AVX512F-NEXT:    vpextrb $8, %xmm3, %eax409; AVX512F-NEXT:    andl $63, %eax410; AVX512F-NEXT:    vpinsrb $8, (%rsp,%rax), %xmm4, %xmm4411; AVX512F-NEXT:    vpextrb $9, %xmm3, %eax412; AVX512F-NEXT:    andl $63, %eax413; AVX512F-NEXT:    vpinsrb $9, (%rsp,%rax), %xmm4, %xmm4414; AVX512F-NEXT:    vpextrb $10, %xmm3, %eax415; AVX512F-NEXT:    andl $63, %eax416; AVX512F-NEXT:    vpinsrb $10, (%rsp,%rax), %xmm4, %xmm4417; AVX512F-NEXT:    vpextrb $11, %xmm3, %eax418; AVX512F-NEXT:    andl $63, %eax419; AVX512F-NEXT:    vpinsrb $11, (%rsp,%rax), %xmm4, %xmm4420; AVX512F-NEXT:    vpextrb $12, %xmm3, %eax421; AVX512F-NEXT:    andl $63, %eax422; AVX512F-NEXT:    vpinsrb $12, (%rsp,%rax), %xmm4, %xmm4423; AVX512F-NEXT:    vpextrb $13, %xmm3, %eax424; AVX512F-NEXT:    andl $63, %eax425; AVX512F-NEXT:    movzbl (%rsp,%rax), %eax426; AVX512F-NEXT:    vpinsrb $13, %eax, %xmm4, %xmm4427; AVX512F-NEXT:    vpextrb $14, %xmm3, %eax428; AVX512F-NEXT:    andl $63, %eax429; AVX512F-NEXT:    movzbl (%rsp,%rax), %eax430; AVX512F-NEXT:    vpinsrb $14, %eax, %xmm4, %xmm4431; AVX512F-NEXT:    vpextrb $15, %xmm3, %eax432; AVX512F-NEXT:    andl $63, %eax433; AVX512F-NEXT:    movzbl (%rsp,%rax), %eax434; AVX512F-NEXT:    vpinsrb $15, %eax, %xmm4, %xmm3435; AVX512F-NEXT:    vpextrb $0, %xmm2, %eax436; AVX512F-NEXT:    andl $63, %eax437; AVX512F-NEXT:    movzbl (%rsp,%rax), %eax438; AVX512F-NEXT:    vmovd %eax, %xmm4439; AVX512F-NEXT:    vpextrb $1, %xmm2, %eax440; AVX512F-NEXT:    andl $63, %eax441; AVX512F-NEXT:    vpinsrb $1, (%rsp,%rax), %xmm4, %xmm4442; AVX512F-NEXT:    vpextrb $2, %xmm2, %eax443; AVX512F-NEXT:    andl $63, %eax444; AVX512F-NEXT:    vpinsrb $2, (%rsp,%rax), %xmm4, %xmm4445; AVX512F-NEXT:    vpextrb $3, %xmm2, %eax446; AVX512F-NEXT:    andl $63, %eax447; AVX512F-NEXT:    vpinsrb $3, (%rsp,%rax), %xmm4, %xmm4448; AVX512F-NEXT:    vpextrb $4, %xmm2, %eax449; AVX512F-NEXT:    andl $63, %eax450; AVX512F-NEXT:    vpinsrb $4, (%rsp,%rax), %xmm4, %xmm4451; AVX512F-NEXT:    vpextrb $5, %xmm2, %eax452; AVX512F-NEXT:    andl $63, %eax453; AVX512F-NEXT:    vpinsrb $5, (%rsp,%rax), %xmm4, %xmm4454; AVX512F-NEXT:    vpextrb $6, %xmm2, %eax455; AVX512F-NEXT:    andl $63, %eax456; AVX512F-NEXT:    vpinsrb $6, (%rsp,%rax), %xmm4, %xmm4457; AVX512F-NEXT:    vpextrb $7, %xmm2, %eax458; AVX512F-NEXT:    andl $63, %eax459; AVX512F-NEXT:    vpinsrb $7, (%rsp,%rax), %xmm4, %xmm4460; AVX512F-NEXT:    vpextrb $8, %xmm2, %eax461; AVX512F-NEXT:    andl $63, %eax462; AVX512F-NEXT:    vpinsrb $8, (%rsp,%rax), %xmm4, %xmm4463; AVX512F-NEXT:    vpextrb $9, %xmm2, %eax464; AVX512F-NEXT:    andl $63, %eax465; AVX512F-NEXT:    vpinsrb $9, (%rsp,%rax), %xmm4, %xmm4466; AVX512F-NEXT:    vpextrb $10, %xmm2, %eax467; AVX512F-NEXT:    andl $63, %eax468; AVX512F-NEXT:    vpinsrb $10, (%rsp,%rax), %xmm4, %xmm4469; AVX512F-NEXT:    vpextrb $11, %xmm2, %eax470; AVX512F-NEXT:    andl $63, %eax471; AVX512F-NEXT:    vpinsrb $11, (%rsp,%rax), %xmm4, %xmm4472; AVX512F-NEXT:    vpextrb $12, %xmm2, %eax473; AVX512F-NEXT:    andl $63, %eax474; AVX512F-NEXT:    movzbl (%rsp,%rax), %eax475; AVX512F-NEXT:    vpinsrb $12, %eax, %xmm4, %xmm4476; AVX512F-NEXT:    vpextrb $13, %xmm2, %eax477; AVX512F-NEXT:    andl $63, %eax478; AVX512F-NEXT:    movzbl (%rsp,%rax), %eax479; AVX512F-NEXT:    vpinsrb $13, %eax, %xmm4, %xmm4480; AVX512F-NEXT:    vpextrb $14, %xmm2, %eax481; AVX512F-NEXT:    andl $63, %eax482; AVX512F-NEXT:    movzbl (%rsp,%rax), %eax483; AVX512F-NEXT:    vpinsrb $14, %eax, %xmm4, %xmm4484; AVX512F-NEXT:    vpextrb $15, %xmm2, %eax485; AVX512F-NEXT:    andl $63, %eax486; AVX512F-NEXT:    movzbl (%rsp,%rax), %eax487; AVX512F-NEXT:    vpinsrb $15, %eax, %xmm4, %xmm2488; AVX512F-NEXT:    vpextrb $0, %xmm1, %eax489; AVX512F-NEXT:    andl $63, %eax490; AVX512F-NEXT:    movzbl (%rsp,%rax), %eax491; AVX512F-NEXT:    vmovd %eax, %xmm4492; AVX512F-NEXT:    vpextrb $1, %xmm1, %eax493; AVX512F-NEXT:    andl $63, %eax494; AVX512F-NEXT:    vpinsrb $1, (%rsp,%rax), %xmm4, %xmm4495; AVX512F-NEXT:    vpextrb $2, %xmm1, %eax496; AVX512F-NEXT:    andl $63, %eax497; AVX512F-NEXT:    vpinsrb $2, (%rsp,%rax), %xmm4, %xmm4498; AVX512F-NEXT:    vpextrb $3, %xmm1, %eax499; AVX512F-NEXT:    andl $63, %eax500; AVX512F-NEXT:    vpinsrb $3, (%rsp,%rax), %xmm4, %xmm4501; AVX512F-NEXT:    vpextrb $4, %xmm1, %eax502; AVX512F-NEXT:    andl $63, %eax503; AVX512F-NEXT:    vpinsrb $4, (%rsp,%rax), %xmm4, %xmm4504; AVX512F-NEXT:    vpextrb $5, %xmm1, %eax505; AVX512F-NEXT:    andl $63, %eax506; AVX512F-NEXT:    vpinsrb $5, (%rsp,%rax), %xmm4, %xmm4507; AVX512F-NEXT:    vpextrb $6, %xmm1, %eax508; AVX512F-NEXT:    andl $63, %eax509; AVX512F-NEXT:    vpinsrb $6, (%rsp,%rax), %xmm4, %xmm4510; AVX512F-NEXT:    vpextrb $7, %xmm1, %eax511; AVX512F-NEXT:    andl $63, %eax512; AVX512F-NEXT:    vpinsrb $7, (%rsp,%rax), %xmm4, %xmm4513; AVX512F-NEXT:    vpextrb $8, %xmm1, %eax514; AVX512F-NEXT:    andl $63, %eax515; AVX512F-NEXT:    vpinsrb $8, (%rsp,%rax), %xmm4, %xmm4516; AVX512F-NEXT:    vpextrb $9, %xmm1, %eax517; AVX512F-NEXT:    andl $63, %eax518; AVX512F-NEXT:    vpinsrb $9, (%rsp,%rax), %xmm4, %xmm4519; AVX512F-NEXT:    vpextrb $10, %xmm1, %eax520; AVX512F-NEXT:    andl $63, %eax521; AVX512F-NEXT:    vpinsrb $10, (%rsp,%rax), %xmm4, %xmm4522; AVX512F-NEXT:    vpextrb $11, %xmm1, %eax523; AVX512F-NEXT:    andl $63, %eax524; AVX512F-NEXT:    movzbl (%rsp,%rax), %eax525; AVX512F-NEXT:    vpinsrb $11, %eax, %xmm4, %xmm4526; AVX512F-NEXT:    vpextrb $12, %xmm1, %eax527; AVX512F-NEXT:    andl $63, %eax528; AVX512F-NEXT:    movzbl (%rsp,%rax), %eax529; AVX512F-NEXT:    vpinsrb $12, %eax, %xmm4, %xmm4530; AVX512F-NEXT:    vpextrb $13, %xmm1, %eax531; AVX512F-NEXT:    andl $63, %eax532; AVX512F-NEXT:    movzbl (%rsp,%rax), %eax533; AVX512F-NEXT:    vpinsrb $13, %eax, %xmm4, %xmm4534; AVX512F-NEXT:    vpextrb $14, %xmm1, %eax535; AVX512F-NEXT:    andl $63, %eax536; AVX512F-NEXT:    movzbl (%rsp,%rax), %eax537; AVX512F-NEXT:    vpinsrb $14, %eax, %xmm4, %xmm4538; AVX512F-NEXT:    vpextrb $15, %xmm1, %eax539; AVX512F-NEXT:    andl $63, %eax540; AVX512F-NEXT:    movzbl (%rsp,%rax), %eax541; AVX512F-NEXT:    vpinsrb $15, %eax, %xmm4, %xmm1542; AVX512F-NEXT:    vinserti128 $1, %xmm0, %ymm3, %ymm0543; AVX512F-NEXT:    vinserti128 $1, %xmm2, %ymm1, %ymm1544; AVX512F-NEXT:    vinserti64x4 $1, %ymm0, %zmm1, %zmm0545; AVX512F-NEXT:    movq %rbp, %rsp546; AVX512F-NEXT:    popq %rbp547; AVX512F-NEXT:    retq548;549; AVX512BW-LABEL: var_shuffle_v64i8:550; AVX512BW:       # %bb.0:551; AVX512BW-NEXT:    pushq %rbp552; AVX512BW-NEXT:    movq %rsp, %rbp553; AVX512BW-NEXT:    andq $-64, %rsp554; AVX512BW-NEXT:    subq $128, %rsp555; AVX512BW-NEXT:    vextracti128 $1, %ymm1, %xmm2556; AVX512BW-NEXT:    vextracti32x4 $2, %zmm1, %xmm3557; AVX512BW-NEXT:    vextracti32x4 $3, %zmm1, %xmm4558; AVX512BW-NEXT:    vpextrb $0, %xmm4, %eax559; AVX512BW-NEXT:    vmovaps %zmm0, (%rsp)560; AVX512BW-NEXT:    andl $63, %eax561; AVX512BW-NEXT:    movzbl (%rsp,%rax), %eax562; AVX512BW-NEXT:    vmovd %eax, %xmm0563; AVX512BW-NEXT:    vpextrb $1, %xmm4, %eax564; AVX512BW-NEXT:    andl $63, %eax565; AVX512BW-NEXT:    vpinsrb $1, (%rsp,%rax), %xmm0, %xmm0566; AVX512BW-NEXT:    vpextrb $2, %xmm4, %eax567; AVX512BW-NEXT:    andl $63, %eax568; AVX512BW-NEXT:    vpinsrb $2, (%rsp,%rax), %xmm0, %xmm0569; AVX512BW-NEXT:    vpextrb $3, %xmm4, %eax570; AVX512BW-NEXT:    andl $63, %eax571; AVX512BW-NEXT:    vpinsrb $3, (%rsp,%rax), %xmm0, %xmm0572; AVX512BW-NEXT:    vpextrb $4, %xmm4, %eax573; AVX512BW-NEXT:    andl $63, %eax574; AVX512BW-NEXT:    vpinsrb $4, (%rsp,%rax), %xmm0, %xmm0575; AVX512BW-NEXT:    vpextrb $5, %xmm4, %eax576; AVX512BW-NEXT:    andl $63, %eax577; AVX512BW-NEXT:    vpinsrb $5, (%rsp,%rax), %xmm0, %xmm0578; AVX512BW-NEXT:    vpextrb $6, %xmm4, %eax579; AVX512BW-NEXT:    andl $63, %eax580; AVX512BW-NEXT:    vpinsrb $6, (%rsp,%rax), %xmm0, %xmm0581; AVX512BW-NEXT:    vpextrb $7, %xmm4, %eax582; AVX512BW-NEXT:    andl $63, %eax583; AVX512BW-NEXT:    vpinsrb $7, (%rsp,%rax), %xmm0, %xmm0584; AVX512BW-NEXT:    vpextrb $8, %xmm4, %eax585; AVX512BW-NEXT:    andl $63, %eax586; AVX512BW-NEXT:    vpinsrb $8, (%rsp,%rax), %xmm0, %xmm0587; AVX512BW-NEXT:    vpextrb $9, %xmm4, %eax588; AVX512BW-NEXT:    andl $63, %eax589; AVX512BW-NEXT:    vpinsrb $9, (%rsp,%rax), %xmm0, %xmm0590; AVX512BW-NEXT:    vpextrb $10, %xmm4, %eax591; AVX512BW-NEXT:    andl $63, %eax592; AVX512BW-NEXT:    vpinsrb $10, (%rsp,%rax), %xmm0, %xmm0593; AVX512BW-NEXT:    vpextrb $11, %xmm4, %eax594; AVX512BW-NEXT:    andl $63, %eax595; AVX512BW-NEXT:    vpinsrb $11, (%rsp,%rax), %xmm0, %xmm0596; AVX512BW-NEXT:    vpextrb $12, %xmm4, %eax597; AVX512BW-NEXT:    andl $63, %eax598; AVX512BW-NEXT:    vpinsrb $12, (%rsp,%rax), %xmm0, %xmm0599; AVX512BW-NEXT:    vpextrb $13, %xmm4, %eax600; AVX512BW-NEXT:    andl $63, %eax601; AVX512BW-NEXT:    vpinsrb $13, (%rsp,%rax), %xmm0, %xmm0602; AVX512BW-NEXT:    vpextrb $14, %xmm4, %eax603; AVX512BW-NEXT:    andl $63, %eax604; AVX512BW-NEXT:    vpinsrb $14, (%rsp,%rax), %xmm0, %xmm0605; AVX512BW-NEXT:    vpextrb $15, %xmm4, %eax606; AVX512BW-NEXT:    andl $63, %eax607; AVX512BW-NEXT:    vpinsrb $15, (%rsp,%rax), %xmm0, %xmm0608; AVX512BW-NEXT:    vpextrb $0, %xmm3, %eax609; AVX512BW-NEXT:    andl $63, %eax610; AVX512BW-NEXT:    movzbl (%rsp,%rax), %eax611; AVX512BW-NEXT:    vmovd %eax, %xmm4612; AVX512BW-NEXT:    vpextrb $1, %xmm3, %eax613; AVX512BW-NEXT:    andl $63, %eax614; AVX512BW-NEXT:    vpinsrb $1, (%rsp,%rax), %xmm4, %xmm4615; AVX512BW-NEXT:    vpextrb $2, %xmm3, %eax616; AVX512BW-NEXT:    andl $63, %eax617; AVX512BW-NEXT:    vpinsrb $2, (%rsp,%rax), %xmm4, %xmm4618; AVX512BW-NEXT:    vpextrb $3, %xmm3, %eax619; AVX512BW-NEXT:    andl $63, %eax620; AVX512BW-NEXT:    vpinsrb $3, (%rsp,%rax), %xmm4, %xmm4621; AVX512BW-NEXT:    vpextrb $4, %xmm3, %eax622; AVX512BW-NEXT:    andl $63, %eax623; AVX512BW-NEXT:    vpinsrb $4, (%rsp,%rax), %xmm4, %xmm4624; AVX512BW-NEXT:    vpextrb $5, %xmm3, %eax625; AVX512BW-NEXT:    andl $63, %eax626; AVX512BW-NEXT:    vpinsrb $5, (%rsp,%rax), %xmm4, %xmm4627; AVX512BW-NEXT:    vpextrb $6, %xmm3, %eax628; AVX512BW-NEXT:    andl $63, %eax629; AVX512BW-NEXT:    vpinsrb $6, (%rsp,%rax), %xmm4, %xmm4630; AVX512BW-NEXT:    vpextrb $7, %xmm3, %eax631; AVX512BW-NEXT:    andl $63, %eax632; AVX512BW-NEXT:    vpinsrb $7, (%rsp,%rax), %xmm4, %xmm4633; AVX512BW-NEXT:    vpextrb $8, %xmm3, %eax634; AVX512BW-NEXT:    andl $63, %eax635; AVX512BW-NEXT:    vpinsrb $8, (%rsp,%rax), %xmm4, %xmm4636; AVX512BW-NEXT:    vpextrb $9, %xmm3, %eax637; AVX512BW-NEXT:    andl $63, %eax638; AVX512BW-NEXT:    vpinsrb $9, (%rsp,%rax), %xmm4, %xmm4639; AVX512BW-NEXT:    vpextrb $10, %xmm3, %eax640; AVX512BW-NEXT:    andl $63, %eax641; AVX512BW-NEXT:    vpinsrb $10, (%rsp,%rax), %xmm4, %xmm4642; AVX512BW-NEXT:    vpextrb $11, %xmm3, %eax643; AVX512BW-NEXT:    andl $63, %eax644; AVX512BW-NEXT:    vpinsrb $11, (%rsp,%rax), %xmm4, %xmm4645; AVX512BW-NEXT:    vpextrb $12, %xmm3, %eax646; AVX512BW-NEXT:    andl $63, %eax647; AVX512BW-NEXT:    vpinsrb $12, (%rsp,%rax), %xmm4, %xmm4648; AVX512BW-NEXT:    vpextrb $13, %xmm3, %eax649; AVX512BW-NEXT:    andl $63, %eax650; AVX512BW-NEXT:    movzbl (%rsp,%rax), %eax651; AVX512BW-NEXT:    vpinsrb $13, %eax, %xmm4, %xmm4652; AVX512BW-NEXT:    vpextrb $14, %xmm3, %eax653; AVX512BW-NEXT:    andl $63, %eax654; AVX512BW-NEXT:    movzbl (%rsp,%rax), %eax655; AVX512BW-NEXT:    vpinsrb $14, %eax, %xmm4, %xmm4656; AVX512BW-NEXT:    vpextrb $15, %xmm3, %eax657; AVX512BW-NEXT:    andl $63, %eax658; AVX512BW-NEXT:    movzbl (%rsp,%rax), %eax659; AVX512BW-NEXT:    vpinsrb $15, %eax, %xmm4, %xmm3660; AVX512BW-NEXT:    vpextrb $0, %xmm2, %eax661; AVX512BW-NEXT:    andl $63, %eax662; AVX512BW-NEXT:    movzbl (%rsp,%rax), %eax663; AVX512BW-NEXT:    vmovd %eax, %xmm4664; AVX512BW-NEXT:    vpextrb $1, %xmm2, %eax665; AVX512BW-NEXT:    andl $63, %eax666; AVX512BW-NEXT:    vpinsrb $1, (%rsp,%rax), %xmm4, %xmm4667; AVX512BW-NEXT:    vpextrb $2, %xmm2, %eax668; AVX512BW-NEXT:    andl $63, %eax669; AVX512BW-NEXT:    vpinsrb $2, (%rsp,%rax), %xmm4, %xmm4670; AVX512BW-NEXT:    vpextrb $3, %xmm2, %eax671; AVX512BW-NEXT:    andl $63, %eax672; AVX512BW-NEXT:    vpinsrb $3, (%rsp,%rax), %xmm4, %xmm4673; AVX512BW-NEXT:    vpextrb $4, %xmm2, %eax674; AVX512BW-NEXT:    andl $63, %eax675; AVX512BW-NEXT:    vpinsrb $4, (%rsp,%rax), %xmm4, %xmm4676; AVX512BW-NEXT:    vpextrb $5, %xmm2, %eax677; AVX512BW-NEXT:    andl $63, %eax678; AVX512BW-NEXT:    vpinsrb $5, (%rsp,%rax), %xmm4, %xmm4679; AVX512BW-NEXT:    vpextrb $6, %xmm2, %eax680; AVX512BW-NEXT:    andl $63, %eax681; AVX512BW-NEXT:    vpinsrb $6, (%rsp,%rax), %xmm4, %xmm4682; AVX512BW-NEXT:    vpextrb $7, %xmm2, %eax683; AVX512BW-NEXT:    andl $63, %eax684; AVX512BW-NEXT:    vpinsrb $7, (%rsp,%rax), %xmm4, %xmm4685; AVX512BW-NEXT:    vpextrb $8, %xmm2, %eax686; AVX512BW-NEXT:    andl $63, %eax687; AVX512BW-NEXT:    vpinsrb $8, (%rsp,%rax), %xmm4, %xmm4688; AVX512BW-NEXT:    vpextrb $9, %xmm2, %eax689; AVX512BW-NEXT:    andl $63, %eax690; AVX512BW-NEXT:    vpinsrb $9, (%rsp,%rax), %xmm4, %xmm4691; AVX512BW-NEXT:    vpextrb $10, %xmm2, %eax692; AVX512BW-NEXT:    andl $63, %eax693; AVX512BW-NEXT:    vpinsrb $10, (%rsp,%rax), %xmm4, %xmm4694; AVX512BW-NEXT:    vpextrb $11, %xmm2, %eax695; AVX512BW-NEXT:    andl $63, %eax696; AVX512BW-NEXT:    vpinsrb $11, (%rsp,%rax), %xmm4, %xmm4697; AVX512BW-NEXT:    vpextrb $12, %xmm2, %eax698; AVX512BW-NEXT:    andl $63, %eax699; AVX512BW-NEXT:    movzbl (%rsp,%rax), %eax700; AVX512BW-NEXT:    vpinsrb $12, %eax, %xmm4, %xmm4701; AVX512BW-NEXT:    vpextrb $13, %xmm2, %eax702; AVX512BW-NEXT:    andl $63, %eax703; AVX512BW-NEXT:    movzbl (%rsp,%rax), %eax704; AVX512BW-NEXT:    vpinsrb $13, %eax, %xmm4, %xmm4705; AVX512BW-NEXT:    vpextrb $14, %xmm2, %eax706; AVX512BW-NEXT:    andl $63, %eax707; AVX512BW-NEXT:    movzbl (%rsp,%rax), %eax708; AVX512BW-NEXT:    vpinsrb $14, %eax, %xmm4, %xmm4709; AVX512BW-NEXT:    vpextrb $15, %xmm2, %eax710; AVX512BW-NEXT:    andl $63, %eax711; AVX512BW-NEXT:    movzbl (%rsp,%rax), %eax712; AVX512BW-NEXT:    vpinsrb $15, %eax, %xmm4, %xmm2713; AVX512BW-NEXT:    vpextrb $0, %xmm1, %eax714; AVX512BW-NEXT:    andl $63, %eax715; AVX512BW-NEXT:    movzbl (%rsp,%rax), %eax716; AVX512BW-NEXT:    vmovd %eax, %xmm4717; AVX512BW-NEXT:    vpextrb $1, %xmm1, %eax718; AVX512BW-NEXT:    andl $63, %eax719; AVX512BW-NEXT:    vpinsrb $1, (%rsp,%rax), %xmm4, %xmm4720; AVX512BW-NEXT:    vpextrb $2, %xmm1, %eax721; AVX512BW-NEXT:    andl $63, %eax722; AVX512BW-NEXT:    vpinsrb $2, (%rsp,%rax), %xmm4, %xmm4723; AVX512BW-NEXT:    vpextrb $3, %xmm1, %eax724; AVX512BW-NEXT:    andl $63, %eax725; AVX512BW-NEXT:    vpinsrb $3, (%rsp,%rax), %xmm4, %xmm4726; AVX512BW-NEXT:    vpextrb $4, %xmm1, %eax727; AVX512BW-NEXT:    andl $63, %eax728; AVX512BW-NEXT:    vpinsrb $4, (%rsp,%rax), %xmm4, %xmm4729; AVX512BW-NEXT:    vpextrb $5, %xmm1, %eax730; AVX512BW-NEXT:    andl $63, %eax731; AVX512BW-NEXT:    vpinsrb $5, (%rsp,%rax), %xmm4, %xmm4732; AVX512BW-NEXT:    vpextrb $6, %xmm1, %eax733; AVX512BW-NEXT:    andl $63, %eax734; AVX512BW-NEXT:    vpinsrb $6, (%rsp,%rax), %xmm4, %xmm4735; AVX512BW-NEXT:    vpextrb $7, %xmm1, %eax736; AVX512BW-NEXT:    andl $63, %eax737; AVX512BW-NEXT:    vpinsrb $7, (%rsp,%rax), %xmm4, %xmm4738; AVX512BW-NEXT:    vpextrb $8, %xmm1, %eax739; AVX512BW-NEXT:    andl $63, %eax740; AVX512BW-NEXT:    vpinsrb $8, (%rsp,%rax), %xmm4, %xmm4741; AVX512BW-NEXT:    vpextrb $9, %xmm1, %eax742; AVX512BW-NEXT:    andl $63, %eax743; AVX512BW-NEXT:    vpinsrb $9, (%rsp,%rax), %xmm4, %xmm4744; AVX512BW-NEXT:    vpextrb $10, %xmm1, %eax745; AVX512BW-NEXT:    andl $63, %eax746; AVX512BW-NEXT:    vpinsrb $10, (%rsp,%rax), %xmm4, %xmm4747; AVX512BW-NEXT:    vpextrb $11, %xmm1, %eax748; AVX512BW-NEXT:    andl $63, %eax749; AVX512BW-NEXT:    movzbl (%rsp,%rax), %eax750; AVX512BW-NEXT:    vpinsrb $11, %eax, %xmm4, %xmm4751; AVX512BW-NEXT:    vpextrb $12, %xmm1, %eax752; AVX512BW-NEXT:    andl $63, %eax753; AVX512BW-NEXT:    movzbl (%rsp,%rax), %eax754; AVX512BW-NEXT:    vpinsrb $12, %eax, %xmm4, %xmm4755; AVX512BW-NEXT:    vpextrb $13, %xmm1, %eax756; AVX512BW-NEXT:    andl $63, %eax757; AVX512BW-NEXT:    movzbl (%rsp,%rax), %eax758; AVX512BW-NEXT:    vpinsrb $13, %eax, %xmm4, %xmm4759; AVX512BW-NEXT:    vpextrb $14, %xmm1, %eax760; AVX512BW-NEXT:    andl $63, %eax761; AVX512BW-NEXT:    movzbl (%rsp,%rax), %eax762; AVX512BW-NEXT:    vpinsrb $14, %eax, %xmm4, %xmm4763; AVX512BW-NEXT:    vpextrb $15, %xmm1, %eax764; AVX512BW-NEXT:    andl $63, %eax765; AVX512BW-NEXT:    movzbl (%rsp,%rax), %eax766; AVX512BW-NEXT:    vpinsrb $15, %eax, %xmm4, %xmm1767; AVX512BW-NEXT:    vinserti128 $1, %xmm0, %ymm3, %ymm0768; AVX512BW-NEXT:    vinserti128 $1, %xmm2, %ymm1, %ymm1769; AVX512BW-NEXT:    vinserti64x4 $1, %ymm0, %zmm1, %zmm0770; AVX512BW-NEXT:    movq %rbp, %rsp771; AVX512BW-NEXT:    popq %rbp772; AVX512BW-NEXT:    retq773;774; AVX512VBMI-LABEL: var_shuffle_v64i8:775; AVX512VBMI:       # %bb.0:776; AVX512VBMI-NEXT:    vpermb %zmm0, %zmm1, %zmm0777; AVX512VBMI-NEXT:    retq778  %index0 = extractelement <64 x i8> %indices, i32 0779  %index1 = extractelement <64 x i8> %indices, i32 1780  %index2 = extractelement <64 x i8> %indices, i32 2781  %index3 = extractelement <64 x i8> %indices, i32 3782  %index4 = extractelement <64 x i8> %indices, i32 4783  %index5 = extractelement <64 x i8> %indices, i32 5784  %index6 = extractelement <64 x i8> %indices, i32 6785  %index7 = extractelement <64 x i8> %indices, i32 7786  %index8 = extractelement <64 x i8> %indices, i32 8787  %index9 = extractelement <64 x i8> %indices, i32 9788  %index10 = extractelement <64 x i8> %indices, i32 10789  %index11 = extractelement <64 x i8> %indices, i32 11790  %index12 = extractelement <64 x i8> %indices, i32 12791  %index13 = extractelement <64 x i8> %indices, i32 13792  %index14 = extractelement <64 x i8> %indices, i32 14793  %index15 = extractelement <64 x i8> %indices, i32 15794  %index16 = extractelement <64 x i8> %indices, i32 16795  %index17 = extractelement <64 x i8> %indices, i32 17796  %index18 = extractelement <64 x i8> %indices, i32 18797  %index19 = extractelement <64 x i8> %indices, i32 19798  %index20 = extractelement <64 x i8> %indices, i32 20799  %index21 = extractelement <64 x i8> %indices, i32 21800  %index22 = extractelement <64 x i8> %indices, i32 22801  %index23 = extractelement <64 x i8> %indices, i32 23802  %index24 = extractelement <64 x i8> %indices, i32 24803  %index25 = extractelement <64 x i8> %indices, i32 25804  %index26 = extractelement <64 x i8> %indices, i32 26805  %index27 = extractelement <64 x i8> %indices, i32 27806  %index28 = extractelement <64 x i8> %indices, i32 28807  %index29 = extractelement <64 x i8> %indices, i32 29808  %index30 = extractelement <64 x i8> %indices, i32 30809  %index31 = extractelement <64 x i8> %indices, i32 31810  %index32 = extractelement <64 x i8> %indices, i32 32811  %index33 = extractelement <64 x i8> %indices, i32 33812  %index34 = extractelement <64 x i8> %indices, i32 34813  %index35 = extractelement <64 x i8> %indices, i32 35814  %index36 = extractelement <64 x i8> %indices, i32 36815  %index37 = extractelement <64 x i8> %indices, i32 37816  %index38 = extractelement <64 x i8> %indices, i32 38817  %index39 = extractelement <64 x i8> %indices, i32 39818  %index40 = extractelement <64 x i8> %indices, i32 40819  %index41 = extractelement <64 x i8> %indices, i32 41820  %index42 = extractelement <64 x i8> %indices, i32 42821  %index43 = extractelement <64 x i8> %indices, i32 43822  %index44 = extractelement <64 x i8> %indices, i32 44823  %index45 = extractelement <64 x i8> %indices, i32 45824  %index46 = extractelement <64 x i8> %indices, i32 46825  %index47 = extractelement <64 x i8> %indices, i32 47826  %index48 = extractelement <64 x i8> %indices, i32 48827  %index49 = extractelement <64 x i8> %indices, i32 49828  %index50 = extractelement <64 x i8> %indices, i32 50829  %index51 = extractelement <64 x i8> %indices, i32 51830  %index52 = extractelement <64 x i8> %indices, i32 52831  %index53 = extractelement <64 x i8> %indices, i32 53832  %index54 = extractelement <64 x i8> %indices, i32 54833  %index55 = extractelement <64 x i8> %indices, i32 55834  %index56 = extractelement <64 x i8> %indices, i32 56835  %index57 = extractelement <64 x i8> %indices, i32 57836  %index58 = extractelement <64 x i8> %indices, i32 58837  %index59 = extractelement <64 x i8> %indices, i32 59838  %index60 = extractelement <64 x i8> %indices, i32 60839  %index61 = extractelement <64 x i8> %indices, i32 61840  %index62 = extractelement <64 x i8> %indices, i32 62841  %index63 = extractelement <64 x i8> %indices, i32 63842  %v0 = extractelement <64 x i8> %v, i8 %index0843  %v1 = extractelement <64 x i8> %v, i8 %index1844  %v2 = extractelement <64 x i8> %v, i8 %index2845  %v3 = extractelement <64 x i8> %v, i8 %index3846  %v4 = extractelement <64 x i8> %v, i8 %index4847  %v5 = extractelement <64 x i8> %v, i8 %index5848  %v6 = extractelement <64 x i8> %v, i8 %index6849  %v7 = extractelement <64 x i8> %v, i8 %index7850  %v8 = extractelement <64 x i8> %v, i8 %index8851  %v9 = extractelement <64 x i8> %v, i8 %index9852  %v10 = extractelement <64 x i8> %v, i8 %index10853  %v11 = extractelement <64 x i8> %v, i8 %index11854  %v12 = extractelement <64 x i8> %v, i8 %index12855  %v13 = extractelement <64 x i8> %v, i8 %index13856  %v14 = extractelement <64 x i8> %v, i8 %index14857  %v15 = extractelement <64 x i8> %v, i8 %index15858  %v16 = extractelement <64 x i8> %v, i8 %index16859  %v17 = extractelement <64 x i8> %v, i8 %index17860  %v18 = extractelement <64 x i8> %v, i8 %index18861  %v19 = extractelement <64 x i8> %v, i8 %index19862  %v20 = extractelement <64 x i8> %v, i8 %index20863  %v21 = extractelement <64 x i8> %v, i8 %index21864  %v22 = extractelement <64 x i8> %v, i8 %index22865  %v23 = extractelement <64 x i8> %v, i8 %index23866  %v24 = extractelement <64 x i8> %v, i8 %index24867  %v25 = extractelement <64 x i8> %v, i8 %index25868  %v26 = extractelement <64 x i8> %v, i8 %index26869  %v27 = extractelement <64 x i8> %v, i8 %index27870  %v28 = extractelement <64 x i8> %v, i8 %index28871  %v29 = extractelement <64 x i8> %v, i8 %index29872  %v30 = extractelement <64 x i8> %v, i8 %index30873  %v31 = extractelement <64 x i8> %v, i8 %index31874  %v32 = extractelement <64 x i8> %v, i8 %index32875  %v33 = extractelement <64 x i8> %v, i8 %index33876  %v34 = extractelement <64 x i8> %v, i8 %index34877  %v35 = extractelement <64 x i8> %v, i8 %index35878  %v36 = extractelement <64 x i8> %v, i8 %index36879  %v37 = extractelement <64 x i8> %v, i8 %index37880  %v38 = extractelement <64 x i8> %v, i8 %index38881  %v39 = extractelement <64 x i8> %v, i8 %index39882  %v40 = extractelement <64 x i8> %v, i8 %index40883  %v41 = extractelement <64 x i8> %v, i8 %index41884  %v42 = extractelement <64 x i8> %v, i8 %index42885  %v43 = extractelement <64 x i8> %v, i8 %index43886  %v44 = extractelement <64 x i8> %v, i8 %index44887  %v45 = extractelement <64 x i8> %v, i8 %index45888  %v46 = extractelement <64 x i8> %v, i8 %index46889  %v47 = extractelement <64 x i8> %v, i8 %index47890  %v48 = extractelement <64 x i8> %v, i8 %index48891  %v49 = extractelement <64 x i8> %v, i8 %index49892  %v50 = extractelement <64 x i8> %v, i8 %index50893  %v51 = extractelement <64 x i8> %v, i8 %index51894  %v52 = extractelement <64 x i8> %v, i8 %index52895  %v53 = extractelement <64 x i8> %v, i8 %index53896  %v54 = extractelement <64 x i8> %v, i8 %index54897  %v55 = extractelement <64 x i8> %v, i8 %index55898  %v56 = extractelement <64 x i8> %v, i8 %index56899  %v57 = extractelement <64 x i8> %v, i8 %index57900  %v58 = extractelement <64 x i8> %v, i8 %index58901  %v59 = extractelement <64 x i8> %v, i8 %index59902  %v60 = extractelement <64 x i8> %v, i8 %index60903  %v61 = extractelement <64 x i8> %v, i8 %index61904  %v62 = extractelement <64 x i8> %v, i8 %index62905  %v63 = extractelement <64 x i8> %v, i8 %index63906  %ret0 = insertelement <64 x i8> undef, i8 %v0, i32 0907  %ret1 = insertelement <64 x i8> %ret0, i8 %v1, i32 1908  %ret2 = insertelement <64 x i8> %ret1, i8 %v2, i32 2909  %ret3 = insertelement <64 x i8> %ret2, i8 %v3, i32 3910  %ret4 = insertelement <64 x i8> %ret3, i8 %v4, i32 4911  %ret5 = insertelement <64 x i8> %ret4, i8 %v5, i32 5912  %ret6 = insertelement <64 x i8> %ret5, i8 %v6, i32 6913  %ret7 = insertelement <64 x i8> %ret6, i8 %v7, i32 7914  %ret8 = insertelement <64 x i8> %ret7, i8 %v8, i32 8915  %ret9 = insertelement <64 x i8> %ret8, i8 %v9, i32 9916  %ret10 = insertelement <64 x i8> %ret9, i8 %v10, i32 10917  %ret11 = insertelement <64 x i8> %ret10, i8 %v11, i32 11918  %ret12 = insertelement <64 x i8> %ret11, i8 %v12, i32 12919  %ret13 = insertelement <64 x i8> %ret12, i8 %v13, i32 13920  %ret14 = insertelement <64 x i8> %ret13, i8 %v14, i32 14921  %ret15 = insertelement <64 x i8> %ret14, i8 %v15, i32 15922  %ret16 = insertelement <64 x i8> %ret15, i8 %v16, i32 16923  %ret17 = insertelement <64 x i8> %ret16, i8 %v17, i32 17924  %ret18 = insertelement <64 x i8> %ret17, i8 %v18, i32 18925  %ret19 = insertelement <64 x i8> %ret18, i8 %v19, i32 19926  %ret20 = insertelement <64 x i8> %ret19, i8 %v20, i32 20927  %ret21 = insertelement <64 x i8> %ret20, i8 %v21, i32 21928  %ret22 = insertelement <64 x i8> %ret21, i8 %v22, i32 22929  %ret23 = insertelement <64 x i8> %ret22, i8 %v23, i32 23930  %ret24 = insertelement <64 x i8> %ret23, i8 %v24, i32 24931  %ret25 = insertelement <64 x i8> %ret24, i8 %v25, i32 25932  %ret26 = insertelement <64 x i8> %ret25, i8 %v26, i32 26933  %ret27 = insertelement <64 x i8> %ret26, i8 %v27, i32 27934  %ret28 = insertelement <64 x i8> %ret27, i8 %v28, i32 28935  %ret29 = insertelement <64 x i8> %ret28, i8 %v29, i32 29936  %ret30 = insertelement <64 x i8> %ret29, i8 %v30, i32 30937  %ret31 = insertelement <64 x i8> %ret30, i8 %v31, i32 31938  %ret32 = insertelement <64 x i8> %ret31, i8 %v32, i32 32939  %ret33 = insertelement <64 x i8> %ret32, i8 %v33, i32 33940  %ret34 = insertelement <64 x i8> %ret33, i8 %v34, i32 34941  %ret35 = insertelement <64 x i8> %ret34, i8 %v35, i32 35942  %ret36 = insertelement <64 x i8> %ret35, i8 %v36, i32 36943  %ret37 = insertelement <64 x i8> %ret36, i8 %v37, i32 37944  %ret38 = insertelement <64 x i8> %ret37, i8 %v38, i32 38945  %ret39 = insertelement <64 x i8> %ret38, i8 %v39, i32 39946  %ret40 = insertelement <64 x i8> %ret39, i8 %v40, i32 40947  %ret41 = insertelement <64 x i8> %ret40, i8 %v41, i32 41948  %ret42 = insertelement <64 x i8> %ret41, i8 %v42, i32 42949  %ret43 = insertelement <64 x i8> %ret42, i8 %v43, i32 43950  %ret44 = insertelement <64 x i8> %ret43, i8 %v44, i32 44951  %ret45 = insertelement <64 x i8> %ret44, i8 %v45, i32 45952  %ret46 = insertelement <64 x i8> %ret45, i8 %v46, i32 46953  %ret47 = insertelement <64 x i8> %ret46, i8 %v47, i32 47954  %ret48 = insertelement <64 x i8> %ret47, i8 %v48, i32 48955  %ret49 = insertelement <64 x i8> %ret48, i8 %v49, i32 49956  %ret50 = insertelement <64 x i8> %ret49, i8 %v50, i32 50957  %ret51 = insertelement <64 x i8> %ret50, i8 %v51, i32 51958  %ret52 = insertelement <64 x i8> %ret51, i8 %v52, i32 52959  %ret53 = insertelement <64 x i8> %ret52, i8 %v53, i32 53960  %ret54 = insertelement <64 x i8> %ret53, i8 %v54, i32 54961  %ret55 = insertelement <64 x i8> %ret54, i8 %v55, i32 55962  %ret56 = insertelement <64 x i8> %ret55, i8 %v56, i32 56963  %ret57 = insertelement <64 x i8> %ret56, i8 %v57, i32 57964  %ret58 = insertelement <64 x i8> %ret57, i8 %v58, i32 58965  %ret59 = insertelement <64 x i8> %ret58, i8 %v59, i32 59966  %ret60 = insertelement <64 x i8> %ret59, i8 %v60, i32 60967  %ret61 = insertelement <64 x i8> %ret60, i8 %v61, i32 61968  %ret62 = insertelement <64 x i8> %ret61, i8 %v62, i32 62969  %ret63 = insertelement <64 x i8> %ret62, i8 %v63, i32 63970  ret <64 x i8> %ret63971}972 973define <8 x double> @var_shuffle_v8f64(<8 x double> %v, <8 x i64> %indices) nounwind {974; AVX512-LABEL: var_shuffle_v8f64:975; AVX512:       # %bb.0:976; AVX512-NEXT:    vpermpd %zmm0, %zmm1, %zmm0977; AVX512-NEXT:    retq978  %index0 = extractelement <8 x i64> %indices, i32 0979  %index1 = extractelement <8 x i64> %indices, i32 1980  %index2 = extractelement <8 x i64> %indices, i32 2981  %index3 = extractelement <8 x i64> %indices, i32 3982  %index4 = extractelement <8 x i64> %indices, i32 4983  %index5 = extractelement <8 x i64> %indices, i32 5984  %index6 = extractelement <8 x i64> %indices, i32 6985  %index7 = extractelement <8 x i64> %indices, i32 7986  %v0 = extractelement <8 x double> %v, i64 %index0987  %v1 = extractelement <8 x double> %v, i64 %index1988  %v2 = extractelement <8 x double> %v, i64 %index2989  %v3 = extractelement <8 x double> %v, i64 %index3990  %v4 = extractelement <8 x double> %v, i64 %index4991  %v5 = extractelement <8 x double> %v, i64 %index5992  %v6 = extractelement <8 x double> %v, i64 %index6993  %v7 = extractelement <8 x double> %v, i64 %index7994  %ret0 = insertelement <8 x double> undef, double %v0, i32 0995  %ret1 = insertelement <8 x double> %ret0, double %v1, i32 1996  %ret2 = insertelement <8 x double> %ret1, double %v2, i32 2997  %ret3 = insertelement <8 x double> %ret2, double %v3, i32 3998  %ret4 = insertelement <8 x double> %ret3, double %v4, i32 4999  %ret5 = insertelement <8 x double> %ret4, double %v5, i32 51000  %ret6 = insertelement <8 x double> %ret5, double %v6, i32 61001  %ret7 = insertelement <8 x double> %ret6, double %v7, i32 71002  ret <8 x double> %ret71003}1004 1005define <16 x float> @var_shuffle_v16f32(<16 x float> %v, <16 x i32> %indices) nounwind {1006; AVX512-LABEL: var_shuffle_v16f32:1007; AVX512:       # %bb.0:1008; AVX512-NEXT:    vpermps %zmm0, %zmm1, %zmm01009; AVX512-NEXT:    retq1010  %index0 = extractelement <16 x i32> %indices, i32 01011  %index1 = extractelement <16 x i32> %indices, i32 11012  %index2 = extractelement <16 x i32> %indices, i32 21013  %index3 = extractelement <16 x i32> %indices, i32 31014  %index4 = extractelement <16 x i32> %indices, i32 41015  %index5 = extractelement <16 x i32> %indices, i32 51016  %index6 = extractelement <16 x i32> %indices, i32 61017  %index7 = extractelement <16 x i32> %indices, i32 71018  %index8 = extractelement <16 x i32> %indices, i32 81019  %index9 = extractelement <16 x i32> %indices, i32 91020  %index10 = extractelement <16 x i32> %indices, i32 101021  %index11 = extractelement <16 x i32> %indices, i32 111022  %index12 = extractelement <16 x i32> %indices, i32 121023  %index13 = extractelement <16 x i32> %indices, i32 131024  %index14 = extractelement <16 x i32> %indices, i32 141025  %index15 = extractelement <16 x i32> %indices, i32 151026  %v0 = extractelement <16 x float> %v, i32 %index01027  %v1 = extractelement <16 x float> %v, i32 %index11028  %v2 = extractelement <16 x float> %v, i32 %index21029  %v3 = extractelement <16 x float> %v, i32 %index31030  %v4 = extractelement <16 x float> %v, i32 %index41031  %v5 = extractelement <16 x float> %v, i32 %index51032  %v6 = extractelement <16 x float> %v, i32 %index61033  %v7 = extractelement <16 x float> %v, i32 %index71034  %v8 = extractelement <16 x float> %v, i32 %index81035  %v9 = extractelement <16 x float> %v, i32 %index91036  %v10 = extractelement <16 x float> %v, i32 %index101037  %v11 = extractelement <16 x float> %v, i32 %index111038  %v12 = extractelement <16 x float> %v, i32 %index121039  %v13 = extractelement <16 x float> %v, i32 %index131040  %v14 = extractelement <16 x float> %v, i32 %index141041  %v15 = extractelement <16 x float> %v, i32 %index151042  %ret0 = insertelement <16 x float> undef, float %v0, i32 01043  %ret1 = insertelement <16 x float> %ret0, float %v1, i32 11044  %ret2 = insertelement <16 x float> %ret1, float %v2, i32 21045  %ret3 = insertelement <16 x float> %ret2, float %v3, i32 31046  %ret4 = insertelement <16 x float> %ret3, float %v4, i32 41047  %ret5 = insertelement <16 x float> %ret4, float %v5, i32 51048  %ret6 = insertelement <16 x float> %ret5, float %v6, i32 61049  %ret7 = insertelement <16 x float> %ret6, float %v7, i32 71050  %ret8 = insertelement <16 x float> %ret7, float %v8, i32 81051  %ret9 = insertelement <16 x float> %ret8, float %v9, i32 91052  %ret10 = insertelement <16 x float> %ret9, float %v10, i32 101053  %ret11 = insertelement <16 x float> %ret10, float %v11, i32 111054  %ret12 = insertelement <16 x float> %ret11, float %v12, i32 121055  %ret13 = insertelement <16 x float> %ret12, float %v13, i32 131056  %ret14 = insertelement <16 x float> %ret13, float %v14, i32 141057  %ret15 = insertelement <16 x float> %ret14, float %v15, i32 151058  ret <16 x float> %ret151059}1060 1061define void @var_cvt_shuffle_v64f32_v64i8_idx(ptr %dst, <64 x i8> %src, i32 %b) nounwind {1062; AVX512F-LABEL: var_cvt_shuffle_v64f32_v64i8_idx:1063; AVX512F:       # %bb.0:1064; AVX512F-NEXT:    pushq %rbp1065; AVX512F-NEXT:    movq %rsp, %rbp1066; AVX512F-NEXT:    andq $-64, %rsp1067; AVX512F-NEXT:    subq $128, %rsp1068; AVX512F-NEXT:    # kill: def $esi killed $esi def $rsi1069; AVX512F-NEXT:    vpbroadcastd %esi, %zmm21070; AVX512F-NEXT:    vpaddd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm11071; AVX512F-NEXT:    vmovd %xmm1, %eax1072; AVX512F-NEXT:    vmovaps %zmm0, (%rsp)1073; AVX512F-NEXT:    andl $63, %eax1074; AVX512F-NEXT:    movzbl (%rsp,%rax), %eax1075; AVX512F-NEXT:    vmovd %eax, %xmm01076; AVX512F-NEXT:    vpextrd $1, %xmm1, %eax1077; AVX512F-NEXT:    andl $63, %eax1078; AVX512F-NEXT:    vpinsrb $1, (%rsp,%rax), %xmm0, %xmm01079; AVX512F-NEXT:    vpextrd $2, %xmm1, %eax1080; AVX512F-NEXT:    andl $63, %eax1081; AVX512F-NEXT:    vpinsrb $2, (%rsp,%rax), %xmm0, %xmm01082; AVX512F-NEXT:    vpextrd $3, %xmm1, %eax1083; AVX512F-NEXT:    vextracti128 $1, %ymm1, %xmm31084; AVX512F-NEXT:    andl $63, %eax1085; AVX512F-NEXT:    vpinsrb $3, (%rsp,%rax), %xmm0, %xmm01086; AVX512F-NEXT:    vmovd %xmm3, %eax1087; AVX512F-NEXT:    andl $63, %eax1088; AVX512F-NEXT:    vpinsrb $4, (%rsp,%rax), %xmm0, %xmm01089; AVX512F-NEXT:    vpextrd $1, %xmm3, %eax1090; AVX512F-NEXT:    andl $63, %eax1091; AVX512F-NEXT:    vpinsrb $5, (%rsp,%rax), %xmm0, %xmm01092; AVX512F-NEXT:    vpextrd $2, %xmm3, %eax1093; AVX512F-NEXT:    andl $63, %eax1094; AVX512F-NEXT:    vpinsrb $6, (%rsp,%rax), %xmm0, %xmm01095; AVX512F-NEXT:    vpextrd $3, %xmm3, %eax1096; AVX512F-NEXT:    vextracti32x4 $2, %zmm1, %xmm51097; AVX512F-NEXT:    andl $63, %eax1098; AVX512F-NEXT:    vpinsrb $7, (%rsp,%rax), %xmm0, %xmm01099; AVX512F-NEXT:    vmovd %xmm5, %eax1100; AVX512F-NEXT:    andl $63, %eax1101; AVX512F-NEXT:    vpinsrb $8, (%rsp,%rax), %xmm0, %xmm01102; AVX512F-NEXT:    vpextrd $1, %xmm5, %eax1103; AVX512F-NEXT:    andl $63, %eax1104; AVX512F-NEXT:    vpinsrb $9, (%rsp,%rax), %xmm0, %xmm01105; AVX512F-NEXT:    vpaddd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm41106; AVX512F-NEXT:    vpaddd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm31107; AVX512F-NEXT:    vpaddd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm21108; AVX512F-NEXT:    andl $63, %esi1109; AVX512F-NEXT:    vpinsrb $10, (%rsp,%rsi), %xmm0, %xmm01110; AVX512F-NEXT:    vpextrd $3, %xmm5, %eax1111; AVX512F-NEXT:    vextracti32x4 $3, %zmm1, %xmm11112; AVX512F-NEXT:    andl $63, %eax1113; AVX512F-NEXT:    vpinsrb $11, (%rsp,%rax), %xmm0, %xmm01114; AVX512F-NEXT:    vmovd %xmm1, %eax1115; AVX512F-NEXT:    andl $63, %eax1116; AVX512F-NEXT:    vpinsrb $12, (%rsp,%rax), %xmm0, %xmm01117; AVX512F-NEXT:    vpextrd $1, %xmm1, %eax1118; AVX512F-NEXT:    andl $63, %eax1119; AVX512F-NEXT:    vpinsrb $13, (%rsp,%rax), %xmm0, %xmm01120; AVX512F-NEXT:    vpextrd $2, %xmm1, %eax1121; AVX512F-NEXT:    andl $63, %eax1122; AVX512F-NEXT:    vpinsrb $14, (%rsp,%rax), %xmm0, %xmm01123; AVX512F-NEXT:    vpextrd $3, %xmm1, %eax1124; AVX512F-NEXT:    andl $63, %eax1125; AVX512F-NEXT:    vpinsrb $15, (%rsp,%rax), %xmm0, %xmm01126; AVX512F-NEXT:    vmovd %xmm4, %eax1127; AVX512F-NEXT:    andl $63, %eax1128; AVX512F-NEXT:    movzbl (%rsp,%rax), %eax1129; AVX512F-NEXT:    vmovd %eax, %xmm11130; AVX512F-NEXT:    vpextrd $1, %xmm4, %eax1131; AVX512F-NEXT:    andl $63, %eax1132; AVX512F-NEXT:    vpinsrb $1, (%rsp,%rax), %xmm1, %xmm11133; AVX512F-NEXT:    vpextrd $2, %xmm4, %eax1134; AVX512F-NEXT:    andl $63, %eax1135; AVX512F-NEXT:    vpinsrb $2, (%rsp,%rax), %xmm1, %xmm11136; AVX512F-NEXT:    vpextrd $3, %xmm4, %eax1137; AVX512F-NEXT:    vextracti128 $1, %ymm4, %xmm51138; AVX512F-NEXT:    andl $63, %eax1139; AVX512F-NEXT:    vpinsrb $3, (%rsp,%rax), %xmm1, %xmm11140; AVX512F-NEXT:    vmovd %xmm5, %eax1141; AVX512F-NEXT:    andl $63, %eax1142; AVX512F-NEXT:    vpinsrb $4, (%rsp,%rax), %xmm1, %xmm11143; AVX512F-NEXT:    vpextrd $1, %xmm5, %eax1144; AVX512F-NEXT:    andl $63, %eax1145; AVX512F-NEXT:    vpinsrb $5, (%rsp,%rax), %xmm1, %xmm11146; AVX512F-NEXT:    vpextrd $2, %xmm5, %eax1147; AVX512F-NEXT:    andl $63, %eax1148; AVX512F-NEXT:    vpinsrb $6, (%rsp,%rax), %xmm1, %xmm11149; AVX512F-NEXT:    vpextrd $3, %xmm5, %eax1150; AVX512F-NEXT:    vextracti32x4 $2, %zmm4, %xmm51151; AVX512F-NEXT:    andl $63, %eax1152; AVX512F-NEXT:    vpinsrb $7, (%rsp,%rax), %xmm1, %xmm11153; AVX512F-NEXT:    vmovd %xmm5, %eax1154; AVX512F-NEXT:    andl $63, %eax1155; AVX512F-NEXT:    vpinsrb $8, (%rsp,%rax), %xmm1, %xmm11156; AVX512F-NEXT:    vpextrd $1, %xmm5, %eax1157; AVX512F-NEXT:    andl $63, %eax1158; AVX512F-NEXT:    vpinsrb $9, (%rsp,%rax), %xmm1, %xmm11159; AVX512F-NEXT:    vpextrd $2, %xmm5, %eax1160; AVX512F-NEXT:    andl $63, %eax1161; AVX512F-NEXT:    vpinsrb $10, (%rsp,%rax), %xmm1, %xmm61162; AVX512F-NEXT:    vpextrd $3, %xmm5, %eax1163; AVX512F-NEXT:    vextracti32x4 $3, %zmm4, %xmm11164; AVX512F-NEXT:    andl $63, %eax1165; AVX512F-NEXT:    vpinsrb $11, (%rsp,%rax), %xmm6, %xmm41166; AVX512F-NEXT:    vmovd %xmm1, %eax1167; AVX512F-NEXT:    andl $63, %eax1168; AVX512F-NEXT:    vpinsrb $12, (%rsp,%rax), %xmm4, %xmm41169; AVX512F-NEXT:    vmovd %xmm3, %eax1170; AVX512F-NEXT:    andl $63, %eax1171; AVX512F-NEXT:    movzbl (%rsp,%rax), %eax1172; AVX512F-NEXT:    vmovd %eax, %xmm51173; AVX512F-NEXT:    vpextrd $1, %xmm3, %eax1174; AVX512F-NEXT:    andl $63, %eax1175; AVX512F-NEXT:    vpinsrb $1, (%rsp,%rax), %xmm5, %xmm51176; AVX512F-NEXT:    vpextrd $2, %xmm3, %eax1177; AVX512F-NEXT:    andl $63, %eax1178; AVX512F-NEXT:    vpinsrb $2, (%rsp,%rax), %xmm5, %xmm51179; AVX512F-NEXT:    vpextrd $3, %xmm3, %eax1180; AVX512F-NEXT:    vextracti128 $1, %ymm3, %xmm61181; AVX512F-NEXT:    andl $63, %eax1182; AVX512F-NEXT:    vpinsrb $3, (%rsp,%rax), %xmm5, %xmm51183; AVX512F-NEXT:    vmovd %xmm6, %eax1184; AVX512F-NEXT:    andl $63, %eax1185; AVX512F-NEXT:    vpinsrb $4, (%rsp,%rax), %xmm5, %xmm51186; AVX512F-NEXT:    vpextrd $1, %xmm6, %eax1187; AVX512F-NEXT:    andl $63, %eax1188; AVX512F-NEXT:    vpinsrb $5, (%rsp,%rax), %xmm5, %xmm51189; AVX512F-NEXT:    vpextrd $2, %xmm6, %eax1190; AVX512F-NEXT:    andl $63, %eax1191; AVX512F-NEXT:    vpinsrb $6, (%rsp,%rax), %xmm5, %xmm71192; AVX512F-NEXT:    vpextrd $3, %xmm6, %eax1193; AVX512F-NEXT:    vextracti32x4 $2, %zmm3, %xmm51194; AVX512F-NEXT:    andl $63, %eax1195; AVX512F-NEXT:    vpinsrb $7, (%rsp,%rax), %xmm7, %xmm61196; AVX512F-NEXT:    vmovd %xmm5, %eax1197; AVX512F-NEXT:    andl $63, %eax1198; AVX512F-NEXT:    vpinsrb $8, (%rsp,%rax), %xmm6, %xmm61199; AVX512F-NEXT:    vpextrd $1, %xmm5, %eax1200; AVX512F-NEXT:    andl $63, %eax1201; AVX512F-NEXT:    vpinsrb $9, (%rsp,%rax), %xmm6, %xmm61202; AVX512F-NEXT:    vpextrd $2, %xmm5, %eax1203; AVX512F-NEXT:    andl $63, %eax1204; AVX512F-NEXT:    vpinsrb $10, (%rsp,%rax), %xmm6, %xmm61205; AVX512F-NEXT:    vmovd %xmm2, %eax1206; AVX512F-NEXT:    andl $63, %eax1207; AVX512F-NEXT:    movzbl (%rsp,%rax), %eax1208; AVX512F-NEXT:    vmovd %eax, %xmm71209; AVX512F-NEXT:    vpextrd $1, %xmm2, %eax1210; AVX512F-NEXT:    andl $63, %eax1211; AVX512F-NEXT:    vpinsrb $1, (%rsp,%rax), %xmm7, %xmm71212; AVX512F-NEXT:    vpextrd $2, %xmm2, %eax1213; AVX512F-NEXT:    andl $63, %eax1214; AVX512F-NEXT:    vpinsrb $2, (%rsp,%rax), %xmm7, %xmm71215; AVX512F-NEXT:    vpextrd $3, %xmm2, %eax1216; AVX512F-NEXT:    vextracti128 $1, %ymm2, %xmm81217; AVX512F-NEXT:    andl $63, %eax1218; AVX512F-NEXT:    vpinsrb $3, (%rsp,%rax), %xmm7, %xmm71219; AVX512F-NEXT:    vmovd %xmm8, %eax1220; AVX512F-NEXT:    andl $63, %eax1221; AVX512F-NEXT:    vpinsrb $4, (%rsp,%rax), %xmm7, %xmm71222; AVX512F-NEXT:    vpextrd $1, %xmm8, %eax1223; AVX512F-NEXT:    andl $63, %eax1224; AVX512F-NEXT:    vpinsrb $5, (%rsp,%rax), %xmm7, %xmm71225; AVX512F-NEXT:    vpextrd $2, %xmm8, %eax1226; AVX512F-NEXT:    andl $63, %eax1227; AVX512F-NEXT:    vpinsrb $6, (%rsp,%rax), %xmm7, %xmm71228; AVX512F-NEXT:    vpextrd $3, %xmm8, %eax1229; AVX512F-NEXT:    vextracti32x4 $2, %zmm2, %xmm81230; AVX512F-NEXT:    andl $63, %eax1231; AVX512F-NEXT:    vpinsrb $7, (%rsp,%rax), %xmm7, %xmm71232; AVX512F-NEXT:    vmovd %xmm8, %eax1233; AVX512F-NEXT:    andl $63, %eax1234; AVX512F-NEXT:    vpinsrb $8, (%rsp,%rax), %xmm7, %xmm71235; AVX512F-NEXT:    vpextrd $1, %xmm8, %eax1236; AVX512F-NEXT:    andl $63, %eax1237; AVX512F-NEXT:    vpinsrb $9, (%rsp,%rax), %xmm7, %xmm71238; AVX512F-NEXT:    vpextrd $2, %xmm8, %eax1239; AVX512F-NEXT:    andl $63, %eax1240; AVX512F-NEXT:    movzbl (%rsp,%rax), %eax1241; AVX512F-NEXT:    vpinsrb $10, %eax, %xmm7, %xmm71242; AVX512F-NEXT:    vpextrd $3, %xmm8, %eax1243; AVX512F-NEXT:    vextracti32x4 $3, %zmm2, %xmm21244; AVX512F-NEXT:    andl $63, %eax1245; AVX512F-NEXT:    movzbl (%rsp,%rax), %eax1246; AVX512F-NEXT:    vpinsrb $11, %eax, %xmm7, %xmm71247; AVX512F-NEXT:    vmovd %xmm2, %eax1248; AVX512F-NEXT:    andl $63, %eax1249; AVX512F-NEXT:    movzbl (%rsp,%rax), %eax1250; AVX512F-NEXT:    vpinsrb $12, %eax, %xmm7, %xmm71251; AVX512F-NEXT:    vpextrd $1, %xmm2, %eax1252; AVX512F-NEXT:    andl $63, %eax1253; AVX512F-NEXT:    movzbl (%rsp,%rax), %eax1254; AVX512F-NEXT:    vpinsrb $13, %eax, %xmm7, %xmm71255; AVX512F-NEXT:    vpextrd $2, %xmm2, %eax1256; AVX512F-NEXT:    andl $63, %eax1257; AVX512F-NEXT:    movzbl (%rsp,%rax), %eax1258; AVX512F-NEXT:    vpinsrb $14, %eax, %xmm7, %xmm71259; AVX512F-NEXT:    vpextrd $3, %xmm2, %eax1260; AVX512F-NEXT:    andl $63, %eax1261; AVX512F-NEXT:    movzbl (%rsp,%rax), %eax1262; AVX512F-NEXT:    vpinsrb $15, %eax, %xmm7, %xmm21263; AVX512F-NEXT:    vpextrd $3, %xmm5, %eax1264; AVX512F-NEXT:    vextracti32x4 $3, %zmm3, %xmm31265; AVX512F-NEXT:    andl $63, %eax1266; AVX512F-NEXT:    movzbl (%rsp,%rax), %eax1267; AVX512F-NEXT:    vpinsrb $11, %eax, %xmm6, %xmm51268; AVX512F-NEXT:    vmovd %xmm3, %eax1269; AVX512F-NEXT:    andl $63, %eax1270; AVX512F-NEXT:    movzbl (%rsp,%rax), %eax1271; AVX512F-NEXT:    vpinsrb $12, %eax, %xmm5, %xmm51272; AVX512F-NEXT:    vpextrd $1, %xmm3, %eax1273; AVX512F-NEXT:    andl $63, %eax1274; AVX512F-NEXT:    movzbl (%rsp,%rax), %eax1275; AVX512F-NEXT:    vpinsrb $13, %eax, %xmm5, %xmm51276; AVX512F-NEXT:    vpextrd $2, %xmm3, %eax1277; AVX512F-NEXT:    andl $63, %eax1278; AVX512F-NEXT:    movzbl (%rsp,%rax), %eax1279; AVX512F-NEXT:    vpinsrb $14, %eax, %xmm5, %xmm51280; AVX512F-NEXT:    vpextrd $3, %xmm3, %eax1281; AVX512F-NEXT:    andl $63, %eax1282; AVX512F-NEXT:    movzbl (%rsp,%rax), %eax1283; AVX512F-NEXT:    vpinsrb $15, %eax, %xmm5, %xmm31284; AVX512F-NEXT:    vpextrd $1, %xmm1, %eax1285; AVX512F-NEXT:    andl $63, %eax1286; AVX512F-NEXT:    movzbl (%rsp,%rax), %eax1287; AVX512F-NEXT:    vpinsrb $13, %eax, %xmm4, %xmm41288; AVX512F-NEXT:    vpextrd $2, %xmm1, %eax1289; AVX512F-NEXT:    andl $63, %eax1290; AVX512F-NEXT:    movzbl (%rsp,%rax), %eax1291; AVX512F-NEXT:    vpinsrb $14, %eax, %xmm4, %xmm41292; AVX512F-NEXT:    vpextrd $3, %xmm1, %eax1293; AVX512F-NEXT:    andl $63, %eax1294; AVX512F-NEXT:    movzbl (%rsp,%rax), %eax1295; AVX512F-NEXT:    vpinsrb $15, %eax, %xmm4, %xmm11296; AVX512F-NEXT:    vpmovsxbd %xmm2, %zmm21297; AVX512F-NEXT:    vcvtdq2ps %zmm2, %zmm21298; AVX512F-NEXT:    vpmovsxbd %xmm3, %zmm31299; AVX512F-NEXT:    vcvtdq2ps %zmm3, %zmm31300; AVX512F-NEXT:    vpmovsxbd %xmm1, %zmm11301; AVX512F-NEXT:    vcvtdq2ps %zmm1, %zmm11302; AVX512F-NEXT:    vpmovsxbd %xmm0, %zmm01303; AVX512F-NEXT:    vcvtdq2ps %zmm0, %zmm01304; AVX512F-NEXT:    vmovaps %zmm0, 192(%rdi)1305; AVX512F-NEXT:    vmovaps %zmm1, 128(%rdi)1306; AVX512F-NEXT:    vmovaps %zmm3, 64(%rdi)1307; AVX512F-NEXT:    vmovaps %zmm2, (%rdi)1308; AVX512F-NEXT:    movq %rbp, %rsp1309; AVX512F-NEXT:    popq %rbp1310; AVX512F-NEXT:    vzeroupper1311; AVX512F-NEXT:    retq1312;1313; AVX512BW-LABEL: var_cvt_shuffle_v64f32_v64i8_idx:1314; AVX512BW:       # %bb.0:1315; AVX512BW-NEXT:    pushq %rbp1316; AVX512BW-NEXT:    movq %rsp, %rbp1317; AVX512BW-NEXT:    andq $-64, %rsp1318; AVX512BW-NEXT:    subq $128, %rsp1319; AVX512BW-NEXT:    # kill: def $esi killed $esi def $rsi1320; AVX512BW-NEXT:    vpbroadcastd %esi, %zmm21321; AVX512BW-NEXT:    vpaddd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm11322; AVX512BW-NEXT:    vmovd %xmm1, %eax1323; AVX512BW-NEXT:    vmovaps %zmm0, (%rsp)1324; AVX512BW-NEXT:    andl $63, %eax1325; AVX512BW-NEXT:    movzbl (%rsp,%rax), %eax1326; AVX512BW-NEXT:    vmovd %eax, %xmm01327; AVX512BW-NEXT:    vpextrd $1, %xmm1, %eax1328; AVX512BW-NEXT:    andl $63, %eax1329; AVX512BW-NEXT:    vpinsrb $1, (%rsp,%rax), %xmm0, %xmm01330; AVX512BW-NEXT:    vpextrd $2, %xmm1, %eax1331; AVX512BW-NEXT:    andl $63, %eax1332; AVX512BW-NEXT:    vpinsrb $2, (%rsp,%rax), %xmm0, %xmm01333; AVX512BW-NEXT:    vpextrd $3, %xmm1, %eax1334; AVX512BW-NEXT:    vextracti128 $1, %ymm1, %xmm31335; AVX512BW-NEXT:    andl $63, %eax1336; AVX512BW-NEXT:    vpinsrb $3, (%rsp,%rax), %xmm0, %xmm01337; AVX512BW-NEXT:    vmovd %xmm3, %eax1338; AVX512BW-NEXT:    andl $63, %eax1339; AVX512BW-NEXT:    vpinsrb $4, (%rsp,%rax), %xmm0, %xmm01340; AVX512BW-NEXT:    vpextrd $1, %xmm3, %eax1341; AVX512BW-NEXT:    andl $63, %eax1342; AVX512BW-NEXT:    vpinsrb $5, (%rsp,%rax), %xmm0, %xmm01343; AVX512BW-NEXT:    vpextrd $2, %xmm3, %eax1344; AVX512BW-NEXT:    andl $63, %eax1345; AVX512BW-NEXT:    vpinsrb $6, (%rsp,%rax), %xmm0, %xmm01346; AVX512BW-NEXT:    vpextrd $3, %xmm3, %eax1347; AVX512BW-NEXT:    vextracti32x4 $2, %zmm1, %xmm51348; AVX512BW-NEXT:    andl $63, %eax1349; AVX512BW-NEXT:    vpinsrb $7, (%rsp,%rax), %xmm0, %xmm01350; AVX512BW-NEXT:    vmovd %xmm5, %eax1351; AVX512BW-NEXT:    andl $63, %eax1352; AVX512BW-NEXT:    vpinsrb $8, (%rsp,%rax), %xmm0, %xmm01353; AVX512BW-NEXT:    vpextrd $1, %xmm5, %eax1354; AVX512BW-NEXT:    andl $63, %eax1355; AVX512BW-NEXT:    vpinsrb $9, (%rsp,%rax), %xmm0, %xmm01356; AVX512BW-NEXT:    vpaddd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm41357; AVX512BW-NEXT:    vpaddd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm31358; AVX512BW-NEXT:    vpaddd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm21359; AVX512BW-NEXT:    andl $63, %esi1360; AVX512BW-NEXT:    vpinsrb $10, (%rsp,%rsi), %xmm0, %xmm01361; AVX512BW-NEXT:    vpextrd $3, %xmm5, %eax1362; AVX512BW-NEXT:    vextracti32x4 $3, %zmm1, %xmm11363; AVX512BW-NEXT:    andl $63, %eax1364; AVX512BW-NEXT:    vpinsrb $11, (%rsp,%rax), %xmm0, %xmm01365; AVX512BW-NEXT:    vmovd %xmm1, %eax1366; AVX512BW-NEXT:    andl $63, %eax1367; AVX512BW-NEXT:    vpinsrb $12, (%rsp,%rax), %xmm0, %xmm01368; AVX512BW-NEXT:    vpextrd $1, %xmm1, %eax1369; AVX512BW-NEXT:    andl $63, %eax1370; AVX512BW-NEXT:    vpinsrb $13, (%rsp,%rax), %xmm0, %xmm01371; AVX512BW-NEXT:    vpextrd $2, %xmm1, %eax1372; AVX512BW-NEXT:    andl $63, %eax1373; AVX512BW-NEXT:    vpinsrb $14, (%rsp,%rax), %xmm0, %xmm01374; AVX512BW-NEXT:    vpextrd $3, %xmm1, %eax1375; AVX512BW-NEXT:    andl $63, %eax1376; AVX512BW-NEXT:    vpinsrb $15, (%rsp,%rax), %xmm0, %xmm01377; AVX512BW-NEXT:    vmovd %xmm4, %eax1378; AVX512BW-NEXT:    andl $63, %eax1379; AVX512BW-NEXT:    movzbl (%rsp,%rax), %eax1380; AVX512BW-NEXT:    vmovd %eax, %xmm11381; AVX512BW-NEXT:    vpextrd $1, %xmm4, %eax1382; AVX512BW-NEXT:    andl $63, %eax1383; AVX512BW-NEXT:    vpinsrb $1, (%rsp,%rax), %xmm1, %xmm11384; AVX512BW-NEXT:    vpextrd $2, %xmm4, %eax1385; AVX512BW-NEXT:    andl $63, %eax1386; AVX512BW-NEXT:    vpinsrb $2, (%rsp,%rax), %xmm1, %xmm11387; AVX512BW-NEXT:    vpextrd $3, %xmm4, %eax1388; AVX512BW-NEXT:    vextracti128 $1, %ymm4, %xmm51389; AVX512BW-NEXT:    andl $63, %eax1390; AVX512BW-NEXT:    vpinsrb $3, (%rsp,%rax), %xmm1, %xmm11391; AVX512BW-NEXT:    vmovd %xmm5, %eax1392; AVX512BW-NEXT:    andl $63, %eax1393; AVX512BW-NEXT:    vpinsrb $4, (%rsp,%rax), %xmm1, %xmm11394; AVX512BW-NEXT:    vpextrd $1, %xmm5, %eax1395; AVX512BW-NEXT:    andl $63, %eax1396; AVX512BW-NEXT:    vpinsrb $5, (%rsp,%rax), %xmm1, %xmm11397; AVX512BW-NEXT:    vpextrd $2, %xmm5, %eax1398; AVX512BW-NEXT:    andl $63, %eax1399; AVX512BW-NEXT:    vpinsrb $6, (%rsp,%rax), %xmm1, %xmm11400; AVX512BW-NEXT:    vpextrd $3, %xmm5, %eax1401; AVX512BW-NEXT:    vextracti32x4 $2, %zmm4, %xmm51402; AVX512BW-NEXT:    andl $63, %eax1403; AVX512BW-NEXT:    vpinsrb $7, (%rsp,%rax), %xmm1, %xmm11404; AVX512BW-NEXT:    vmovd %xmm5, %eax1405; AVX512BW-NEXT:    andl $63, %eax1406; AVX512BW-NEXT:    vpinsrb $8, (%rsp,%rax), %xmm1, %xmm11407; AVX512BW-NEXT:    vpextrd $1, %xmm5, %eax1408; AVX512BW-NEXT:    andl $63, %eax1409; AVX512BW-NEXT:    vpinsrb $9, (%rsp,%rax), %xmm1, %xmm11410; AVX512BW-NEXT:    vpextrd $2, %xmm5, %eax1411; AVX512BW-NEXT:    andl $63, %eax1412; AVX512BW-NEXT:    vpinsrb $10, (%rsp,%rax), %xmm1, %xmm61413; AVX512BW-NEXT:    vpextrd $3, %xmm5, %eax1414; AVX512BW-NEXT:    vextracti32x4 $3, %zmm4, %xmm11415; AVX512BW-NEXT:    andl $63, %eax1416; AVX512BW-NEXT:    vpinsrb $11, (%rsp,%rax), %xmm6, %xmm41417; AVX512BW-NEXT:    vmovd %xmm1, %eax1418; AVX512BW-NEXT:    andl $63, %eax1419; AVX512BW-NEXT:    vpinsrb $12, (%rsp,%rax), %xmm4, %xmm41420; AVX512BW-NEXT:    vmovd %xmm3, %eax1421; AVX512BW-NEXT:    andl $63, %eax1422; AVX512BW-NEXT:    movzbl (%rsp,%rax), %eax1423; AVX512BW-NEXT:    vmovd %eax, %xmm51424; AVX512BW-NEXT:    vpextrd $1, %xmm3, %eax1425; AVX512BW-NEXT:    andl $63, %eax1426; AVX512BW-NEXT:    vpinsrb $1, (%rsp,%rax), %xmm5, %xmm51427; AVX512BW-NEXT:    vpextrd $2, %xmm3, %eax1428; AVX512BW-NEXT:    andl $63, %eax1429; AVX512BW-NEXT:    vpinsrb $2, (%rsp,%rax), %xmm5, %xmm51430; AVX512BW-NEXT:    vpextrd $3, %xmm3, %eax1431; AVX512BW-NEXT:    vextracti128 $1, %ymm3, %xmm61432; AVX512BW-NEXT:    andl $63, %eax1433; AVX512BW-NEXT:    vpinsrb $3, (%rsp,%rax), %xmm5, %xmm51434; AVX512BW-NEXT:    vmovd %xmm6, %eax1435; AVX512BW-NEXT:    andl $63, %eax1436; AVX512BW-NEXT:    vpinsrb $4, (%rsp,%rax), %xmm5, %xmm51437; AVX512BW-NEXT:    vpextrd $1, %xmm6, %eax1438; AVX512BW-NEXT:    andl $63, %eax1439; AVX512BW-NEXT:    vpinsrb $5, (%rsp,%rax), %xmm5, %xmm51440; AVX512BW-NEXT:    vpextrd $2, %xmm6, %eax1441; AVX512BW-NEXT:    andl $63, %eax1442; AVX512BW-NEXT:    vpinsrb $6, (%rsp,%rax), %xmm5, %xmm71443; AVX512BW-NEXT:    vpextrd $3, %xmm6, %eax1444; AVX512BW-NEXT:    vextracti32x4 $2, %zmm3, %xmm51445; AVX512BW-NEXT:    andl $63, %eax1446; AVX512BW-NEXT:    vpinsrb $7, (%rsp,%rax), %xmm7, %xmm61447; AVX512BW-NEXT:    vmovd %xmm5, %eax1448; AVX512BW-NEXT:    andl $63, %eax1449; AVX512BW-NEXT:    vpinsrb $8, (%rsp,%rax), %xmm6, %xmm61450; AVX512BW-NEXT:    vpextrd $1, %xmm5, %eax1451; AVX512BW-NEXT:    andl $63, %eax1452; AVX512BW-NEXT:    vpinsrb $9, (%rsp,%rax), %xmm6, %xmm61453; AVX512BW-NEXT:    vpextrd $2, %xmm5, %eax1454; AVX512BW-NEXT:    andl $63, %eax1455; AVX512BW-NEXT:    vpinsrb $10, (%rsp,%rax), %xmm6, %xmm61456; AVX512BW-NEXT:    vmovd %xmm2, %eax1457; AVX512BW-NEXT:    andl $63, %eax1458; AVX512BW-NEXT:    movzbl (%rsp,%rax), %eax1459; AVX512BW-NEXT:    vmovd %eax, %xmm71460; AVX512BW-NEXT:    vpextrd $1, %xmm2, %eax1461; AVX512BW-NEXT:    andl $63, %eax1462; AVX512BW-NEXT:    vpinsrb $1, (%rsp,%rax), %xmm7, %xmm71463; AVX512BW-NEXT:    vpextrd $2, %xmm2, %eax1464; AVX512BW-NEXT:    andl $63, %eax1465; AVX512BW-NEXT:    vpinsrb $2, (%rsp,%rax), %xmm7, %xmm71466; AVX512BW-NEXT:    vpextrd $3, %xmm2, %eax1467; AVX512BW-NEXT:    vextracti128 $1, %ymm2, %xmm81468; AVX512BW-NEXT:    andl $63, %eax1469; AVX512BW-NEXT:    vpinsrb $3, (%rsp,%rax), %xmm7, %xmm71470; AVX512BW-NEXT:    vmovd %xmm8, %eax1471; AVX512BW-NEXT:    andl $63, %eax1472; AVX512BW-NEXT:    vpinsrb $4, (%rsp,%rax), %xmm7, %xmm71473; AVX512BW-NEXT:    vpextrd $1, %xmm8, %eax1474; AVX512BW-NEXT:    andl $63, %eax1475; AVX512BW-NEXT:    vpinsrb $5, (%rsp,%rax), %xmm7, %xmm71476; AVX512BW-NEXT:    vpextrd $2, %xmm8, %eax1477; AVX512BW-NEXT:    andl $63, %eax1478; AVX512BW-NEXT:    vpinsrb $6, (%rsp,%rax), %xmm7, %xmm71479; AVX512BW-NEXT:    vpextrd $3, %xmm8, %eax1480; AVX512BW-NEXT:    vextracti32x4 $2, %zmm2, %xmm81481; AVX512BW-NEXT:    andl $63, %eax1482; AVX512BW-NEXT:    vpinsrb $7, (%rsp,%rax), %xmm7, %xmm71483; AVX512BW-NEXT:    vmovd %xmm8, %eax1484; AVX512BW-NEXT:    andl $63, %eax1485; AVX512BW-NEXT:    vpinsrb $8, (%rsp,%rax), %xmm7, %xmm71486; AVX512BW-NEXT:    vpextrd $1, %xmm8, %eax1487; AVX512BW-NEXT:    andl $63, %eax1488; AVX512BW-NEXT:    vpinsrb $9, (%rsp,%rax), %xmm7, %xmm71489; AVX512BW-NEXT:    vpextrd $2, %xmm8, %eax1490; AVX512BW-NEXT:    andl $63, %eax1491; AVX512BW-NEXT:    movzbl (%rsp,%rax), %eax1492; AVX512BW-NEXT:    vpinsrb $10, %eax, %xmm7, %xmm71493; AVX512BW-NEXT:    vpextrd $3, %xmm8, %eax1494; AVX512BW-NEXT:    vextracti32x4 $3, %zmm2, %xmm21495; AVX512BW-NEXT:    andl $63, %eax1496; AVX512BW-NEXT:    movzbl (%rsp,%rax), %eax1497; AVX512BW-NEXT:    vpinsrb $11, %eax, %xmm7, %xmm71498; AVX512BW-NEXT:    vmovd %xmm2, %eax1499; AVX512BW-NEXT:    andl $63, %eax1500; AVX512BW-NEXT:    movzbl (%rsp,%rax), %eax1501; AVX512BW-NEXT:    vpinsrb $12, %eax, %xmm7, %xmm71502; AVX512BW-NEXT:    vpextrd $1, %xmm2, %eax1503; AVX512BW-NEXT:    andl $63, %eax1504; AVX512BW-NEXT:    movzbl (%rsp,%rax), %eax1505; AVX512BW-NEXT:    vpinsrb $13, %eax, %xmm7, %xmm71506; AVX512BW-NEXT:    vpextrd $2, %xmm2, %eax1507; AVX512BW-NEXT:    andl $63, %eax1508; AVX512BW-NEXT:    movzbl (%rsp,%rax), %eax1509; AVX512BW-NEXT:    vpinsrb $14, %eax, %xmm7, %xmm71510; AVX512BW-NEXT:    vpextrd $3, %xmm2, %eax1511; AVX512BW-NEXT:    andl $63, %eax1512; AVX512BW-NEXT:    movzbl (%rsp,%rax), %eax1513; AVX512BW-NEXT:    vpinsrb $15, %eax, %xmm7, %xmm21514; AVX512BW-NEXT:    vpextrd $3, %xmm5, %eax1515; AVX512BW-NEXT:    vextracti32x4 $3, %zmm3, %xmm31516; AVX512BW-NEXT:    andl $63, %eax1517; AVX512BW-NEXT:    movzbl (%rsp,%rax), %eax1518; AVX512BW-NEXT:    vpinsrb $11, %eax, %xmm6, %xmm51519; AVX512BW-NEXT:    vmovd %xmm3, %eax1520; AVX512BW-NEXT:    andl $63, %eax1521; AVX512BW-NEXT:    movzbl (%rsp,%rax), %eax1522; AVX512BW-NEXT:    vpinsrb $12, %eax, %xmm5, %xmm51523; AVX512BW-NEXT:    vpextrd $1, %xmm3, %eax1524; AVX512BW-NEXT:    andl $63, %eax1525; AVX512BW-NEXT:    movzbl (%rsp,%rax), %eax1526; AVX512BW-NEXT:    vpinsrb $13, %eax, %xmm5, %xmm51527; AVX512BW-NEXT:    vpextrd $2, %xmm3, %eax1528; AVX512BW-NEXT:    andl $63, %eax1529; AVX512BW-NEXT:    movzbl (%rsp,%rax), %eax1530; AVX512BW-NEXT:    vpinsrb $14, %eax, %xmm5, %xmm51531; AVX512BW-NEXT:    vpextrd $3, %xmm3, %eax1532; AVX512BW-NEXT:    andl $63, %eax1533; AVX512BW-NEXT:    movzbl (%rsp,%rax), %eax1534; AVX512BW-NEXT:    vpinsrb $15, %eax, %xmm5, %xmm31535; AVX512BW-NEXT:    vpextrd $1, %xmm1, %eax1536; AVX512BW-NEXT:    andl $63, %eax1537; AVX512BW-NEXT:    movzbl (%rsp,%rax), %eax1538; AVX512BW-NEXT:    vpinsrb $13, %eax, %xmm4, %xmm41539; AVX512BW-NEXT:    vpextrd $2, %xmm1, %eax1540; AVX512BW-NEXT:    andl $63, %eax1541; AVX512BW-NEXT:    movzbl (%rsp,%rax), %eax1542; AVX512BW-NEXT:    vpinsrb $14, %eax, %xmm4, %xmm41543; AVX512BW-NEXT:    vpextrd $3, %xmm1, %eax1544; AVX512BW-NEXT:    andl $63, %eax1545; AVX512BW-NEXT:    movzbl (%rsp,%rax), %eax1546; AVX512BW-NEXT:    vpinsrb $15, %eax, %xmm4, %xmm11547; AVX512BW-NEXT:    vpmovsxbd %xmm2, %zmm21548; AVX512BW-NEXT:    vcvtdq2ps %zmm2, %zmm21549; AVX512BW-NEXT:    vpmovsxbd %xmm3, %zmm31550; AVX512BW-NEXT:    vcvtdq2ps %zmm3, %zmm31551; AVX512BW-NEXT:    vpmovsxbd %xmm1, %zmm11552; AVX512BW-NEXT:    vcvtdq2ps %zmm1, %zmm11553; AVX512BW-NEXT:    vpmovsxbd %xmm0, %zmm01554; AVX512BW-NEXT:    vcvtdq2ps %zmm0, %zmm01555; AVX512BW-NEXT:    vmovaps %zmm0, 192(%rdi)1556; AVX512BW-NEXT:    vmovaps %zmm1, 128(%rdi)1557; AVX512BW-NEXT:    vmovaps %zmm3, 64(%rdi)1558; AVX512BW-NEXT:    vmovaps %zmm2, (%rdi)1559; AVX512BW-NEXT:    movq %rbp, %rsp1560; AVX512BW-NEXT:    popq %rbp1561; AVX512BW-NEXT:    vzeroupper1562; AVX512BW-NEXT:    retq1563;1564; AVX512VBMI-LABEL: var_cvt_shuffle_v64f32_v64i8_idx:1565; AVX512VBMI:       # %bb.0:1566; AVX512VBMI-NEXT:    pushq %rbp1567; AVX512VBMI-NEXT:    movq %rsp, %rbp1568; AVX512VBMI-NEXT:    andq $-64, %rsp1569; AVX512VBMI-NEXT:    subq $128, %rsp1570; AVX512VBMI-NEXT:    # kill: def $esi killed $esi def $rsi1571; AVX512VBMI-NEXT:    vpbroadcastd %esi, %zmm11572; AVX512VBMI-NEXT:    vpaddd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm21573; AVX512VBMI-NEXT:    vmovd %xmm2, %eax1574; AVX512VBMI-NEXT:    vmovdqa64 %zmm0, (%rsp)1575; AVX512VBMI-NEXT:    andl $63, %eax1576; AVX512VBMI-NEXT:    movzbl (%rsp,%rax), %eax1577; AVX512VBMI-NEXT:    vmovd %eax, %xmm31578; AVX512VBMI-NEXT:    vpextrd $1, %xmm2, %eax1579; AVX512VBMI-NEXT:    andl $63, %eax1580; AVX512VBMI-NEXT:    vpinsrb $1, (%rsp,%rax), %xmm3, %xmm31581; AVX512VBMI-NEXT:    vpextrd $2, %xmm2, %eax1582; AVX512VBMI-NEXT:    andl $63, %eax1583; AVX512VBMI-NEXT:    vpinsrb $2, (%rsp,%rax), %xmm3, %xmm31584; AVX512VBMI-NEXT:    vpextrd $3, %xmm2, %eax1585; AVX512VBMI-NEXT:    vextracti128 $1, %ymm2, %xmm41586; AVX512VBMI-NEXT:    andl $63, %eax1587; AVX512VBMI-NEXT:    vpinsrb $3, (%rsp,%rax), %xmm3, %xmm31588; AVX512VBMI-NEXT:    vmovd %xmm4, %eax1589; AVX512VBMI-NEXT:    andl $63, %eax1590; AVX512VBMI-NEXT:    vpinsrb $4, (%rsp,%rax), %xmm3, %xmm31591; AVX512VBMI-NEXT:    vpextrd $1, %xmm4, %eax1592; AVX512VBMI-NEXT:    andl $63, %eax1593; AVX512VBMI-NEXT:    vpinsrb $5, (%rsp,%rax), %xmm3, %xmm31594; AVX512VBMI-NEXT:    vpextrd $2, %xmm4, %eax1595; AVX512VBMI-NEXT:    andl $63, %eax1596; AVX512VBMI-NEXT:    vpinsrb $6, (%rsp,%rax), %xmm3, %xmm31597; AVX512VBMI-NEXT:    vpextrd $3, %xmm4, %eax1598; AVX512VBMI-NEXT:    vextracti32x4 $2, %zmm2, %xmm41599; AVX512VBMI-NEXT:    andl $63, %eax1600; AVX512VBMI-NEXT:    vpinsrb $7, (%rsp,%rax), %xmm3, %xmm31601; AVX512VBMI-NEXT:    vmovd %xmm4, %eax1602; AVX512VBMI-NEXT:    andl $63, %eax1603; AVX512VBMI-NEXT:    vpinsrb $8, (%rsp,%rax), %xmm3, %xmm31604; AVX512VBMI-NEXT:    vpextrd $1, %xmm4, %eax1605; AVX512VBMI-NEXT:    andl $63, %eax1606; AVX512VBMI-NEXT:    vpinsrb $9, (%rsp,%rax), %xmm3, %xmm31607; AVX512VBMI-NEXT:    andl $63, %esi1608; AVX512VBMI-NEXT:    vpinsrb $10, (%rsp,%rsi), %xmm3, %xmm31609; AVX512VBMI-NEXT:    vpextrd $3, %xmm4, %eax1610; AVX512VBMI-NEXT:    vextracti32x4 $3, %zmm2, %xmm21611; AVX512VBMI-NEXT:    andl $63, %eax1612; AVX512VBMI-NEXT:    vpinsrb $11, (%rsp,%rax), %xmm3, %xmm31613; AVX512VBMI-NEXT:    vmovd %xmm2, %eax1614; AVX512VBMI-NEXT:    andl $63, %eax1615; AVX512VBMI-NEXT:    vpinsrb $12, (%rsp,%rax), %xmm3, %xmm31616; AVX512VBMI-NEXT:    vpextrd $1, %xmm2, %eax1617; AVX512VBMI-NEXT:    andl $63, %eax1618; AVX512VBMI-NEXT:    vpinsrb $13, (%rsp,%rax), %xmm3, %xmm31619; AVX512VBMI-NEXT:    vpextrd $2, %xmm2, %eax1620; AVX512VBMI-NEXT:    andl $63, %eax1621; AVX512VBMI-NEXT:    vpinsrb $14, (%rsp,%rax), %xmm3, %xmm31622; AVX512VBMI-NEXT:    vpextrd $3, %xmm2, %eax1623; AVX512VBMI-NEXT:    andl $63, %eax1624; AVX512VBMI-NEXT:    vpinsrb $15, (%rsp,%rax), %xmm3, %xmm21625; AVX512VBMI-NEXT:    vpaddd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm31626; AVX512VBMI-NEXT:    vpmovdb %zmm3, %xmm31627; AVX512VBMI-NEXT:    vpaddd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm41628; AVX512VBMI-NEXT:    vpmovdb %zmm4, %xmm41629; AVX512VBMI-NEXT:    vpaddd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %zmm11630; AVX512VBMI-NEXT:    vpmovdb %zmm1, %xmm11631; AVX512VBMI-NEXT:    vpmovsxbd %xmm2, %zmm21632; AVX512VBMI-NEXT:    vcvtdq2ps %zmm2, %zmm21633; AVX512VBMI-NEXT:    vpermb %zmm0, %zmm1, %zmm11634; AVX512VBMI-NEXT:    vpmovsxbd %xmm1, %zmm11635; AVX512VBMI-NEXT:    vcvtdq2ps %zmm1, %zmm11636; AVX512VBMI-NEXT:    vpermb %zmm0, %zmm4, %zmm41637; AVX512VBMI-NEXT:    vpmovsxbd %xmm4, %zmm41638; AVX512VBMI-NEXT:    vcvtdq2ps %zmm4, %zmm41639; AVX512VBMI-NEXT:    vpermb %zmm0, %zmm3, %zmm01640; AVX512VBMI-NEXT:    vpmovsxbd %xmm0, %zmm01641; AVX512VBMI-NEXT:    vcvtdq2ps %zmm0, %zmm01642; AVX512VBMI-NEXT:    vmovaps %zmm0, 128(%rdi)1643; AVX512VBMI-NEXT:    vmovaps %zmm4, 64(%rdi)1644; AVX512VBMI-NEXT:    vmovaps %zmm1, (%rdi)1645; AVX512VBMI-NEXT:    vmovaps %zmm2, 192(%rdi)1646; AVX512VBMI-NEXT:    movq %rbp, %rsp1647; AVX512VBMI-NEXT:    popq %rbp1648; AVX512VBMI-NEXT:    vzeroupper1649; AVX512VBMI-NEXT:    retq1650  %b_broadcast_init = insertelement <64 x i32> undef, i32 %b, i32 01651  %b_broadcast = shufflevector <64 x i32> %b_broadcast_init, <64 x i32> undef, <64 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 undef, i32 0, i32 0, i32 0, i32 0, i32 0>1652  %sub_add__b_broadcast_ = add <64 x i32> %b_broadcast, <i32 58, i32 57, i32 56, i32 55, i32 54, i32 53, i32 52, i32 51, i32 50, i32 49, i32 48, i32 47, i32 46, i32 45, i32 44, i32 43, i32 42, i32 41, i32 40, i32 39, i32 38, i32 37, i32 36, i32 35, i32 34, i32 33, i32 32, i32 31, i32 30, i32 29, i32 28, i32 27, i32 26, i32 25, i32 24, i32 23, i32 22, i32 21, i32 20, i32 19, i32 18, i32 17, i32 16, i32 15, i32 14, i32 13, i32 12, i32 11, i32 10, i32 9, i32 8, i32 7, i32 6, i32 5, i32 4, i32 3, i32 2, i32 1, i32 undef, i32 -1, i32 -2, i32 -3, i32 -4, i32 -5>1653  %index_0.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 01654  %index_1.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 11655  %index_2.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 21656  %index_3.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 31657  %index_4.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 41658  %index_5.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 51659  %index_6.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 61660  %index_7.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 71661  %index_8.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 81662  %index_9.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 91663  %index_10.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 101664  %index_11.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 111665  %index_12.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 121666  %index_13.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 131667  %index_14.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 141668  %index_15.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 151669  %index_16.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 161670  %index_17.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 171671  %index_18.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 181672  %index_19.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 191673  %index_20.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 201674  %index_21.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 211675  %index_22.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 221676  %index_23.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 231677  %index_24.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 241678  %index_25.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 251679  %index_26.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 261680  %index_27.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 271681  %index_28.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 281682  %index_29.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 291683  %index_30.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 301684  %index_31.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 311685  %index_32.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 321686  %index_33.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 331687  %index_34.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 341688  %index_35.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 351689  %index_36.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 361690  %index_37.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 371691  %index_38.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 381692  %index_39.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 391693  %index_40.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 401694  %index_41.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 411695  %index_42.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 421696  %index_43.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 431697  %index_44.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 441698  %index_45.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 451699  %index_46.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 461700  %index_47.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 471701  %index_48.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 481702  %index_49.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 491703  %index_50.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 501704  %index_51.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 511705  %index_52.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 521706  %index_53.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 531707  %index_54.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 541708  %index_55.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 551709  %index_56.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 561710  %index_57.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 571711  %index_59.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 591712  %index_60.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 601713  %index_61.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 611714  %index_62.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 621715  %index_63.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 631716  %v_0.i.i = extractelement <64 x i8> %src, i32 %index_0.i.i1717  %v_1.i.i = extractelement <64 x i8> %src, i32 %index_1.i.i1718  %v_2.i.i = extractelement <64 x i8> %src, i32 %index_2.i.i1719  %v_3.i.i = extractelement <64 x i8> %src, i32 %index_3.i.i1720  %v_4.i.i = extractelement <64 x i8> %src, i32 %index_4.i.i1721  %v_5.i.i = extractelement <64 x i8> %src, i32 %index_5.i.i1722  %v_6.i.i = extractelement <64 x i8> %src, i32 %index_6.i.i1723  %v_7.i.i = extractelement <64 x i8> %src, i32 %index_7.i.i1724  %v_8.i.i = extractelement <64 x i8> %src, i32 %index_8.i.i1725  %v_9.i.i = extractelement <64 x i8> %src, i32 %index_9.i.i1726  %v_10.i.i = extractelement <64 x i8> %src, i32 %index_10.i.i1727  %v_11.i.i = extractelement <64 x i8> %src, i32 %index_11.i.i1728  %v_12.i.i = extractelement <64 x i8> %src, i32 %index_12.i.i1729  %v_13.i.i = extractelement <64 x i8> %src, i32 %index_13.i.i1730  %v_14.i.i = extractelement <64 x i8> %src, i32 %index_14.i.i1731  %v_15.i.i = extractelement <64 x i8> %src, i32 %index_15.i.i1732  %v_16.i.i = extractelement <64 x i8> %src, i32 %index_16.i.i1733  %v_17.i.i = extractelement <64 x i8> %src, i32 %index_17.i.i1734  %v_18.i.i = extractelement <64 x i8> %src, i32 %index_18.i.i1735  %v_19.i.i = extractelement <64 x i8> %src, i32 %index_19.i.i1736  %v_20.i.i = extractelement <64 x i8> %src, i32 %index_20.i.i1737  %v_21.i.i = extractelement <64 x i8> %src, i32 %index_21.i.i1738  %v_22.i.i = extractelement <64 x i8> %src, i32 %index_22.i.i1739  %v_23.i.i = extractelement <64 x i8> %src, i32 %index_23.i.i1740  %v_24.i.i = extractelement <64 x i8> %src, i32 %index_24.i.i1741  %v_25.i.i = extractelement <64 x i8> %src, i32 %index_25.i.i1742  %v_26.i.i = extractelement <64 x i8> %src, i32 %index_26.i.i1743  %v_27.i.i = extractelement <64 x i8> %src, i32 %index_27.i.i1744  %v_28.i.i = extractelement <64 x i8> %src, i32 %index_28.i.i1745  %v_29.i.i = extractelement <64 x i8> %src, i32 %index_29.i.i1746  %v_30.i.i = extractelement <64 x i8> %src, i32 %index_30.i.i1747  %v_31.i.i = extractelement <64 x i8> %src, i32 %index_31.i.i1748  %v_32.i.i = extractelement <64 x i8> %src, i32 %index_32.i.i1749  %v_33.i.i = extractelement <64 x i8> %src, i32 %index_33.i.i1750  %v_34.i.i = extractelement <64 x i8> %src, i32 %index_34.i.i1751  %v_35.i.i = extractelement <64 x i8> %src, i32 %index_35.i.i1752  %v_36.i.i = extractelement <64 x i8> %src, i32 %index_36.i.i1753  %v_37.i.i = extractelement <64 x i8> %src, i32 %index_37.i.i1754  %v_38.i.i = extractelement <64 x i8> %src, i32 %index_38.i.i1755  %v_39.i.i = extractelement <64 x i8> %src, i32 %index_39.i.i1756  %v_40.i.i = extractelement <64 x i8> %src, i32 %index_40.i.i1757  %v_41.i.i = extractelement <64 x i8> %src, i32 %index_41.i.i1758  %v_42.i.i = extractelement <64 x i8> %src, i32 %index_42.i.i1759  %v_43.i.i = extractelement <64 x i8> %src, i32 %index_43.i.i1760  %v_44.i.i = extractelement <64 x i8> %src, i32 %index_44.i.i1761  %v_45.i.i = extractelement <64 x i8> %src, i32 %index_45.i.i1762  %v_46.i.i = extractelement <64 x i8> %src, i32 %index_46.i.i1763  %v_47.i.i = extractelement <64 x i8> %src, i32 %index_47.i.i1764  %v_48.i.i = extractelement <64 x i8> %src, i32 %index_48.i.i1765  %v_49.i.i = extractelement <64 x i8> %src, i32 %index_49.i.i1766  %v_50.i.i = extractelement <64 x i8> %src, i32 %index_50.i.i1767  %v_51.i.i = extractelement <64 x i8> %src, i32 %index_51.i.i1768  %v_52.i.i = extractelement <64 x i8> %src, i32 %index_52.i.i1769  %v_53.i.i = extractelement <64 x i8> %src, i32 %index_53.i.i1770  %v_54.i.i = extractelement <64 x i8> %src, i32 %index_54.i.i1771  %v_55.i.i = extractelement <64 x i8> %src, i32 %index_55.i.i1772  %v_56.i.i = extractelement <64 x i8> %src, i32 %index_56.i.i1773  %v_57.i.i = extractelement <64 x i8> %src, i32 %index_57.i.i1774  %v_58.i.i = extractelement <64 x i8> %src, i32 %b1775  %v_59.i.i = extractelement <64 x i8> %src, i32 %index_59.i.i1776  %v_60.i.i = extractelement <64 x i8> %src, i32 %index_60.i.i1777  %v_61.i.i = extractelement <64 x i8> %src, i32 %index_61.i.i1778  %v_62.i.i = extractelement <64 x i8> %src, i32 %index_62.i.i1779  %v_63.i.i = extractelement <64 x i8> %src, i32 %index_63.i.i1780  %dst_0.i.i = insertelement <64 x i8> undef, i8 %v_0.i.i, i32 01781  %dst_1.i.i = insertelement <64 x i8> %dst_0.i.i, i8 %v_1.i.i, i32 11782  %dst_2.i.i = insertelement <64 x i8> %dst_1.i.i, i8 %v_2.i.i, i32 21783  %dst_3.i.i = insertelement <64 x i8> %dst_2.i.i, i8 %v_3.i.i, i32 31784  %dst_4.i.i = insertelement <64 x i8> %dst_3.i.i, i8 %v_4.i.i, i32 41785  %dst_5.i.i = insertelement <64 x i8> %dst_4.i.i, i8 %v_5.i.i, i32 51786  %dst_6.i.i = insertelement <64 x i8> %dst_5.i.i, i8 %v_6.i.i, i32 61787  %dst_7.i.i = insertelement <64 x i8> %dst_6.i.i, i8 %v_7.i.i, i32 71788  %dst_8.i.i = insertelement <64 x i8> %dst_7.i.i, i8 %v_8.i.i, i32 81789  %dst_9.i.i = insertelement <64 x i8> %dst_8.i.i, i8 %v_9.i.i, i32 91790  %dst_10.i.i = insertelement <64 x i8> %dst_9.i.i, i8 %v_10.i.i, i32 101791  %dst_11.i.i = insertelement <64 x i8> %dst_10.i.i, i8 %v_11.i.i, i32 111792  %dst_12.i.i = insertelement <64 x i8> %dst_11.i.i, i8 %v_12.i.i, i32 121793  %dst_13.i.i = insertelement <64 x i8> %dst_12.i.i, i8 %v_13.i.i, i32 131794  %dst_14.i.i = insertelement <64 x i8> %dst_13.i.i, i8 %v_14.i.i, i32 141795  %dst_15.i.i = insertelement <64 x i8> %dst_14.i.i, i8 %v_15.i.i, i32 151796  %dst_16.i.i = insertelement <64 x i8> %dst_15.i.i, i8 %v_16.i.i, i32 161797  %dst_17.i.i = insertelement <64 x i8> %dst_16.i.i, i8 %v_17.i.i, i32 171798  %dst_18.i.i = insertelement <64 x i8> %dst_17.i.i, i8 %v_18.i.i, i32 181799  %dst_19.i.i = insertelement <64 x i8> %dst_18.i.i, i8 %v_19.i.i, i32 191800  %dst_20.i.i = insertelement <64 x i8> %dst_19.i.i, i8 %v_20.i.i, i32 201801  %dst_21.i.i = insertelement <64 x i8> %dst_20.i.i, i8 %v_21.i.i, i32 211802  %dst_22.i.i = insertelement <64 x i8> %dst_21.i.i, i8 %v_22.i.i, i32 221803  %dst_23.i.i = insertelement <64 x i8> %dst_22.i.i, i8 %v_23.i.i, i32 231804  %dst_24.i.i = insertelement <64 x i8> %dst_23.i.i, i8 %v_24.i.i, i32 241805  %dst_25.i.i = insertelement <64 x i8> %dst_24.i.i, i8 %v_25.i.i, i32 251806  %dst_26.i.i = insertelement <64 x i8> %dst_25.i.i, i8 %v_26.i.i, i32 261807  %dst_27.i.i = insertelement <64 x i8> %dst_26.i.i, i8 %v_27.i.i, i32 271808  %dst_28.i.i = insertelement <64 x i8> %dst_27.i.i, i8 %v_28.i.i, i32 281809  %dst_29.i.i = insertelement <64 x i8> %dst_28.i.i, i8 %v_29.i.i, i32 291810  %dst_30.i.i = insertelement <64 x i8> %dst_29.i.i, i8 %v_30.i.i, i32 301811  %dst_31.i.i = insertelement <64 x i8> %dst_30.i.i, i8 %v_31.i.i, i32 311812  %dst_32.i.i = insertelement <64 x i8> %dst_31.i.i, i8 %v_32.i.i, i32 321813  %dst_33.i.i = insertelement <64 x i8> %dst_32.i.i, i8 %v_33.i.i, i32 331814  %dst_34.i.i = insertelement <64 x i8> %dst_33.i.i, i8 %v_34.i.i, i32 341815  %dst_35.i.i = insertelement <64 x i8> %dst_34.i.i, i8 %v_35.i.i, i32 351816  %dst_36.i.i = insertelement <64 x i8> %dst_35.i.i, i8 %v_36.i.i, i32 361817  %dst_37.i.i = insertelement <64 x i8> %dst_36.i.i, i8 %v_37.i.i, i32 371818  %dst_38.i.i = insertelement <64 x i8> %dst_37.i.i, i8 %v_38.i.i, i32 381819  %dst_39.i.i = insertelement <64 x i8> %dst_38.i.i, i8 %v_39.i.i, i32 391820  %dst_40.i.i = insertelement <64 x i8> %dst_39.i.i, i8 %v_40.i.i, i32 401821  %dst_41.i.i = insertelement <64 x i8> %dst_40.i.i, i8 %v_41.i.i, i32 411822  %dst_42.i.i = insertelement <64 x i8> %dst_41.i.i, i8 %v_42.i.i, i32 421823  %dst_43.i.i = insertelement <64 x i8> %dst_42.i.i, i8 %v_43.i.i, i32 431824  %dst_44.i.i = insertelement <64 x i8> %dst_43.i.i, i8 %v_44.i.i, i32 441825  %dst_45.i.i = insertelement <64 x i8> %dst_44.i.i, i8 %v_45.i.i, i32 451826  %dst_46.i.i = insertelement <64 x i8> %dst_45.i.i, i8 %v_46.i.i, i32 461827  %dst_47.i.i = insertelement <64 x i8> %dst_46.i.i, i8 %v_47.i.i, i32 471828  %dst_48.i.i = insertelement <64 x i8> %dst_47.i.i, i8 %v_48.i.i, i32 481829  %dst_49.i.i = insertelement <64 x i8> %dst_48.i.i, i8 %v_49.i.i, i32 491830  %dst_50.i.i = insertelement <64 x i8> %dst_49.i.i, i8 %v_50.i.i, i32 501831  %dst_51.i.i = insertelement <64 x i8> %dst_50.i.i, i8 %v_51.i.i, i32 511832  %dst_52.i.i = insertelement <64 x i8> %dst_51.i.i, i8 %v_52.i.i, i32 521833  %dst_53.i.i = insertelement <64 x i8> %dst_52.i.i, i8 %v_53.i.i, i32 531834  %dst_54.i.i = insertelement <64 x i8> %dst_53.i.i, i8 %v_54.i.i, i32 541835  %dst_55.i.i = insertelement <64 x i8> %dst_54.i.i, i8 %v_55.i.i, i32 551836  %dst_56.i.i = insertelement <64 x i8> %dst_55.i.i, i8 %v_56.i.i, i32 561837  %dst_57.i.i = insertelement <64 x i8> %dst_56.i.i, i8 %v_57.i.i, i32 571838  %dst_58.i.i = insertelement <64 x i8> %dst_57.i.i, i8 %v_58.i.i, i32 581839  %dst_59.i.i = insertelement <64 x i8> %dst_58.i.i, i8 %v_59.i.i, i32 591840  %dst_60.i.i = insertelement <64 x i8> %dst_59.i.i, i8 %v_60.i.i, i32 601841  %dst_61.i.i = insertelement <64 x i8> %dst_60.i.i, i8 %v_61.i.i, i32 611842  %dst_62.i.i = insertelement <64 x i8> %dst_61.i.i, i8 %v_62.i.i, i32 621843  %dst_63.i.i = insertelement <64 x i8> %dst_62.i.i, i8 %v_63.i.i, i32 631844  %shuf_load_to_float = sitofp <64 x i8> %dst_63.i.i to <64 x float>1845  store <64 x float> %shuf_load_to_float, ptr %dst1846  ret void1847}1848