; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f | FileCheck %s --check-prefixes=AVX512,AVX512F
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512bw | FileCheck %s --check-prefixes=AVX512,AVX512BW
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512bw,+avx512vbmi | FileCheck %s --check-prefixes=AVX512,AVX512VBMI

define <8 x i64> @var_shuffle_v8i64(<8 x i64> %v, <8 x i64> %indices) nounwind {
; AVX512-LABEL: var_shuffle_v8i64:
; AVX512:       # %bb.0:
; AVX512-NEXT:    vpermpd %zmm0, %zmm1, %zmm0
; AVX512-NEXT:    retq
  %index0 = extractelement <8 x i64> %indices, i32 0
  %index1 = extractelement <8 x i64> %indices, i32 1
  %index2 = extractelement <8 x i64> %indices, i32 2
  %index3 = extractelement <8 x i64> %indices, i32 3
  %index4 = extractelement <8 x i64> %indices, i32 4
  %index5 = extractelement <8 x i64> %indices, i32 5
  %index6 = extractelement <8 x i64> %indices, i32 6
  %index7 = extractelement <8 x i64> %indices, i32 7
  %v0 = extractelement <8 x i64> %v, i64 %index0
  %v1 = extractelement <8 x i64> %v, i64 %index1
  %v2 = extractelement <8 x i64> %v, i64 %index2
  %v3 = extractelement <8 x i64> %v, i64 %index3
  %v4 = extractelement <8 x i64> %v, i64 %index4
  %v5 = extractelement <8 x i64> %v, i64 %index5
  %v6 = extractelement <8 x i64> %v, i64 %index6
  %v7 = extractelement <8 x i64> %v, i64 %index7
  %ret0 = insertelement <8 x i64> undef, i64 %v0, i32 0
  %ret1 = insertelement <8 x i64> %ret0, i64 %v1, i32 1
  %ret2 = insertelement <8 x i64> %ret1, i64 %v2, i32 2
  %ret3 = insertelement <8 x i64> %ret2, i64 %v3, i32 3
  %ret4 = insertelement <8 x i64> %ret3, i64 %v4, i32 4
  %ret5 = insertelement <8 x i64> %ret4, i64 %v5, i32 5
  %ret6 = insertelement <8 x i64> %ret5, i64 %v6, i32 6
  %ret7 = insertelement <8 x i64> %ret6, i64 %v7, i32 7
  ret <8 x i64> %ret7
}

define <16 x i32> @var_shuffle_v16i32(<16 x i32> %v, <16 x i32> %indices) nounwind {
; AVX512-LABEL: var_shuffle_v16i32:
; AVX512:       # %bb.0:
; AVX512-NEXT:    vpermps %zmm0, %zmm1, %zmm0
; AVX512-NEXT:    retq
  %index0 = extractelement <16 x i32> %indices, i32 0
  %index1 = extractelement <16 x i32> %indices, i32 1
  %index2 = extractelement <16 x i32> %indices, i32 2
  %index3 = extractelement <16 x i32> %indices, i32 3
  %index4 = extractelement <16 x i32> %indices, i32 4
  %index5 = extractelement <16 x i32> %indices, i32 5
  %index6 = extractelement <16 x i32> %indices, i32 6
  %index7 = extractelement <16 x i32> %indices, i32 7
  %index8 = extractelement <16 x i32> %indices, i32 8
  %index9 = extractelement <16 x i32> %indices, i32 9
  %index10 = extractelement <16 x i32> %indices, i32 10
  %index11 = extractelement <16 x i32> %indices, i32 11
  %index12 = extractelement <16 x i32> %indices, i32 12
  %index13 = extractelement <16 x i32> %indices, i32 13
  %index14 = extractelement <16 x i32> %indices, i32 14
  %index15 = extractelement <16 x i32> %indices, i32 15
  %v0 = extractelement <16 x i32> %v, i32 %index0
  %v1 = extractelement <16 x i32> %v, i32 %index1
  %v2 = extractelement <16 x i32> %v, i32 %index2
  %v3 = extractelement <16 x i32> %v, i32 %index3
  %v4 = extractelement <16 x i32> %v, i32 %index4
  %v5 = extractelement <16 x i32> %v, i32 %index5
  %v6 = extractelement <16 x i32> %v, i32 %index6
  %v7 = extractelement <16 x i32> %v, i32 %index7
  %v8 = extractelement <16 x i32> %v, i32 %index8
  %v9 = extractelement <16 x i32> %v, i32 %index9
  %v10 = extractelement <16 x i32> %v, i32 %index10
  %v11 = extractelement <16 x i32> %v, i32 %index11
  %v12 = extractelement <16 x i32> %v, i32 %index12
  %v13 = extractelement <16 x i32> %v, i32 %index13
  %v14 = extractelement <16 x i32> %v, i32 %index14
  %v15 = extractelement <16 x i32> %v, i32 %index15
  %ret0 = insertelement <16 x i32> undef, i32 %v0, i32 0
  %ret1 = insertelement <16 x i32> %ret0, i32 %v1, i32 1
  %ret2 = insertelement <16 x i32> %ret1, i32 %v2, i32 2
  %ret3 = insertelement <16 x i32> %ret2, i32 %v3, i32 3
  %ret4 = insertelement <16 x i32> %ret3, i32 %v4, i32 4
  %ret5 = insertelement <16 x i32> %ret4, i32 %v5, i32 5
  %ret6 = insertelement <16 x i32> %ret5, i32 %v6, i32 6
  %ret7 = insertelement <16 x i32> %ret6, i32 %v7, i32 7
  %ret8 = insertelement <16 x i32> %ret7, i32 %v8, i32 8
  %ret9 = insertelement <16 x i32> %ret8, i32 %v9, i32 9
  %ret10 = insertelement <16 x i32> %ret9, i32 %v10, i32 10
  %ret11 = insertelement <16 x i32> %ret10, i32 %v11, i32 11
  %ret12 = insertelement <16 x i32> %ret11, i32 %v12, i32 12
  %ret13 = insertelement <16 x i32> %ret12, i32 %v13, i32 13
  %ret14 = insertelement <16 x i32> %ret13, i32 %v14, i32 14
  %ret15 = insertelement <16 x i32> %ret14, i32 %v15, i32 15
  ret <16 x i32> %ret15
}

define <32 x i16> @var_shuffle_v32i16(<32 x i16> %v, <32 x i16> %indices) nounwind {
; AVX512F-LABEL: var_shuffle_v32i16:
; AVX512F:       # %bb.0:
; AVX512F-NEXT:    pushq %rbp
; AVX512F-NEXT:    movq %rsp, %rbp
; AVX512F-NEXT:    andq $-64, %rsp
; AVX512F-NEXT:    subq $2112, %rsp # imm = 0x840
; AVX512F-NEXT:    vextracti64x4 $1, %zmm1, %ymm2
; AVX512F-NEXT:    vextracti128 $1, %ymm1, %xmm3
; AVX512F-NEXT:    vextracti128 $1, %ymm2, %xmm4
; AVX512F-NEXT:    vmovd %xmm4, %eax
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, (%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    andl $31, %eax
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    movzwl 1536(%rsp,%rax,2), %eax
; AVX512F-NEXT:    vmovd %eax, %xmm0
; AVX512F-NEXT:    vpextrw $1, %xmm4, %eax
; AVX512F-NEXT:    andl $31, %eax
; AVX512F-NEXT:    vpinsrw $1, 1600(%rsp,%rax,2), %xmm0, %xmm0
; AVX512F-NEXT:    vpextrw $2, %xmm4, %eax
; AVX512F-NEXT:    andl $31, %eax
; AVX512F-NEXT:    vpinsrw $2, 1664(%rsp,%rax,2), %xmm0, %xmm0
; AVX512F-NEXT:    vpextrw $3, %xmm4, %eax
; AVX512F-NEXT:    andl $31, %eax
; AVX512F-NEXT:    vpinsrw $3, 1728(%rsp,%rax,2), %xmm0, %xmm0
; AVX512F-NEXT:    vpextrw $4, %xmm4, %eax
; AVX512F-NEXT:    andl $31, %eax
; AVX512F-NEXT:    vpinsrw $4, 1792(%rsp,%rax,2), %xmm0, %xmm0
; AVX512F-NEXT:    vpextrw $5, %xmm4, %eax
; AVX512F-NEXT:    andl $31, %eax
; AVX512F-NEXT:    vpinsrw $5, 1856(%rsp,%rax,2), %xmm0, %xmm0
; AVX512F-NEXT:    vpextrw $6, %xmm4, %eax
; AVX512F-NEXT:    andl $31, %eax
; AVX512F-NEXT:    vpinsrw $6, 1920(%rsp,%rax,2), %xmm0, %xmm0
; AVX512F-NEXT:    vpextrw $7, %xmm4, %eax
; AVX512F-NEXT:    andl $31, %eax
; AVX512F-NEXT:    vpinsrw $7, 1984(%rsp,%rax,2), %xmm0, %xmm0
; AVX512F-NEXT:    vmovd %xmm2, %eax
; AVX512F-NEXT:    andl $31, %eax
; AVX512F-NEXT:    movzwl 1024(%rsp,%rax,2), %eax
; AVX512F-NEXT:    vmovd %eax, %xmm4
; AVX512F-NEXT:    vpextrw $1, %xmm2, %eax
; AVX512F-NEXT:    andl $31, %eax
; AVX512F-NEXT:    vpinsrw $1, 1088(%rsp,%rax,2), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrw $2, %xmm2, %eax
; AVX512F-NEXT:    andl $31, %eax
; AVX512F-NEXT:    vpinsrw $2, 1152(%rsp,%rax,2), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrw $3, %xmm2, %eax
; AVX512F-NEXT:    andl $31, %eax
; AVX512F-NEXT:    vpinsrw $3, 1216(%rsp,%rax,2), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrw $4, %xmm2, %eax
; AVX512F-NEXT:    andl $31, %eax
; AVX512F-NEXT:    vpinsrw $4, 1280(%rsp,%rax,2), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrw $5, %xmm2, %eax
; AVX512F-NEXT:    andl $31, %eax
; AVX512F-NEXT:    vpinsrw $5, 1344(%rsp,%rax,2), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrw $6, %xmm2, %eax
; AVX512F-NEXT:    andl $31, %eax
; AVX512F-NEXT:    vpinsrw $6, 1408(%rsp,%rax,2), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrw $7, %xmm2, %eax
; AVX512F-NEXT:    andl $31, %eax
; AVX512F-NEXT:    vpinsrw $7, 1472(%rsp,%rax,2), %xmm4, %xmm2
; AVX512F-NEXT:    vmovd %xmm3, %eax
; AVX512F-NEXT:    andl $31, %eax
; AVX512F-NEXT:    movzwl 512(%rsp,%rax,2), %eax
; AVX512F-NEXT:    vmovd %eax, %xmm4
; AVX512F-NEXT:    vpextrw $1, %xmm3, %eax
; AVX512F-NEXT:    andl $31, %eax
; AVX512F-NEXT:    vpinsrw $1, 576(%rsp,%rax,2), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrw $2, %xmm3, %eax
; AVX512F-NEXT:    andl $31, %eax
; AVX512F-NEXT:    vpinsrw $2, 640(%rsp,%rax,2), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrw $3, %xmm3, %eax
; AVX512F-NEXT:    andl $31, %eax
; AVX512F-NEXT:    vpinsrw $3, 704(%rsp,%rax,2), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrw $4, %xmm3, %eax
; AVX512F-NEXT:    andl $31, %eax
; AVX512F-NEXT:    vpinsrw $4, 768(%rsp,%rax,2), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrw $5, %xmm3, %eax
; AVX512F-NEXT:    andl $31, %eax
; AVX512F-NEXT:    vpinsrw $5, 832(%rsp,%rax,2), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrw $6, %xmm3, %eax
; AVX512F-NEXT:    andl $31, %eax
; AVX512F-NEXT:    vpinsrw $6, 896(%rsp,%rax,2), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrw $7, %xmm3, %eax
; AVX512F-NEXT:    andl $31, %eax
; AVX512F-NEXT:    vpinsrw $7, 960(%rsp,%rax,2), %xmm4, %xmm3
; AVX512F-NEXT:    vmovd %xmm1, %eax
; AVX512F-NEXT:    andl $31, %eax
; AVX512F-NEXT:    movzwl (%rsp,%rax,2), %eax
; AVX512F-NEXT:    vmovd %eax, %xmm4
; AVX512F-NEXT:    vpextrw $1, %xmm1, %eax
; AVX512F-NEXT:    andl $31, %eax
; AVX512F-NEXT:    vpinsrw $1, 64(%rsp,%rax,2), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrw $2, %xmm1, %eax
; AVX512F-NEXT:    andl $31, %eax
; AVX512F-NEXT:    vpinsrw $2, 128(%rsp,%rax,2), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrw $3, %xmm1, %eax
; AVX512F-NEXT:    andl $31, %eax
; AVX512F-NEXT:    vpinsrw $3, 192(%rsp,%rax,2), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrw $4, %xmm1, %eax
; AVX512F-NEXT:    andl $31, %eax
; AVX512F-NEXT:    vpinsrw $4, 256(%rsp,%rax,2), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrw $5, %xmm1, %eax
; AVX512F-NEXT:    andl $31, %eax
; AVX512F-NEXT:    vpinsrw $5, 320(%rsp,%rax,2), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrw $6, %xmm1, %eax
; AVX512F-NEXT:    andl $31, %eax
; AVX512F-NEXT:    vpinsrw $6, 384(%rsp,%rax,2), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrw $7, %xmm1, %eax
; AVX512F-NEXT:    andl $31, %eax
; AVX512F-NEXT:    vpinsrw $7, 448(%rsp,%rax,2), %xmm4, %xmm1
; AVX512F-NEXT:    vinserti128 $1, %xmm0, %ymm2, %ymm0
; AVX512F-NEXT:    vinserti128 $1, %xmm3, %ymm1, %ymm1
; AVX512F-NEXT:    vinserti64x4 $1, %ymm0, %zmm1, %zmm0
; AVX512F-NEXT:    movq %rbp, %rsp
; AVX512F-NEXT:    popq %rbp
; AVX512F-NEXT:    retq
;
; AVX512BW-LABEL: var_shuffle_v32i16:
; AVX512BW:       # %bb.0:
; AVX512BW-NEXT:    vpermw %zmm0, %zmm1, %zmm0
; AVX512BW-NEXT:    retq
;
; AVX512VBMI-LABEL: var_shuffle_v32i16:
; AVX512VBMI:       # %bb.0:
; AVX512VBMI-NEXT:    vpermw %zmm0, %zmm1, %zmm0
; AVX512VBMI-NEXT:    retq
  %index0 = extractelement <32 x i16> %indices, i32 0
  %index1 = extractelement <32 x i16> %indices, i32 1
  %index2 = extractelement <32 x i16> %indices, i32 2
  %index3 = extractelement <32 x i16> %indices, i32 3
  %index4 = extractelement <32 x i16> %indices, i32 4
  %index5 = extractelement <32 x i16> %indices, i32 5
  %index6 = extractelement <32 x i16> %indices, i32 6
  %index7 = extractelement <32 x i16> %indices, i32 7
  %index8 = extractelement <32 x i16> %indices, i32 8
  %index9 = extractelement <32 x i16> %indices, i32 9
  %index10 = extractelement <32 x i16> %indices, i32 10
  %index11 = extractelement <32 x i16> %indices, i32 11
  %index12 = extractelement <32 x i16> %indices, i32 12
  %index13 = extractelement <32 x i16> %indices, i32 13
  %index14 = extractelement <32 x i16> %indices, i32 14
  %index15 = extractelement <32 x i16> %indices, i32 15
  %index16 = extractelement <32 x i16> %indices, i32 16
  %index17 = extractelement <32 x i16> %indices, i32 17
  %index18 = extractelement <32 x i16> %indices, i32 18
  %index19 = extractelement <32 x i16> %indices, i32 19
  %index20 = extractelement <32 x i16> %indices, i32 20
  %index21 = extractelement <32 x i16> %indices, i32 21
  %index22 = extractelement <32 x i16> %indices, i32 22
  %index23 = extractelement <32 x i16> %indices, i32 23
  %index24 = extractelement <32 x i16> %indices, i32 24
  %index25 = extractelement <32 x i16> %indices, i32 25
  %index26 = extractelement <32 x i16> %indices, i32 26
  %index27 = extractelement <32 x i16> %indices, i32 27
  %index28 = extractelement <32 x i16> %indices, i32 28
  %index29 = extractelement <32 x i16> %indices, i32 29
  %index30 = extractelement <32 x i16> %indices, i32 30
  %index31 = extractelement <32 x i16> %indices, i32 31
  %v0 = extractelement <32 x i16> %v, i16 %index0
  %v1 = extractelement <32 x i16> %v, i16 %index1
  %v2 = extractelement <32 x i16> %v, i16 %index2
  %v3 = extractelement <32 x i16> %v, i16 %index3
  %v4 = extractelement <32 x i16> %v, i16 %index4
  %v5 = extractelement <32 x i16> %v, i16 %index5
  %v6 = extractelement <32 x i16> %v, i16 %index6
  %v7 = extractelement <32 x i16> %v, i16 %index7
  %v8 = extractelement <32 x i16> %v, i16 %index8
  %v9 = extractelement <32 x i16> %v, i16 %index9
  %v10 = extractelement <32 x i16> %v, i16 %index10
  %v11 = extractelement <32 x i16> %v, i16 %index11
  %v12 = extractelement <32 x i16> %v, i16 %index12
  %v13 = extractelement <32 x i16> %v, i16 %index13
  %v14 = extractelement <32 x i16> %v, i16 %index14
  %v15 = extractelement <32 x i16> %v, i16 %index15
  %v16 = extractelement <32 x i16> %v, i16 %index16
  %v17 = extractelement <32 x i16> %v, i16 %index17
  %v18 = extractelement <32 x i16> %v, i16 %index18
  %v19 = extractelement <32 x i16> %v, i16 %index19
  %v20 = extractelement <32 x i16> %v, i16 %index20
  %v21 = extractelement <32 x i16> %v, i16 %index21
  %v22 = extractelement <32 x i16> %v, i16 %index22
  %v23 = extractelement <32 x i16> %v, i16 %index23
  %v24 = extractelement <32 x i16> %v, i16 %index24
  %v25 = extractelement <32 x i16> %v, i16 %index25
  %v26 = extractelement <32 x i16> %v, i16 %index26
  %v27 = extractelement <32 x i16> %v, i16 %index27
  %v28 = extractelement <32 x i16> %v, i16 %index28
  %v29 = extractelement <32 x i16> %v, i16 %index29
  %v30 = extractelement <32 x i16> %v, i16 %index30
  %v31 = extractelement <32 x i16> %v, i16 %index31
  %ret0 = insertelement <32 x i16> undef, i16 %v0, i32 0
  %ret1 = insertelement <32 x i16> %ret0, i16 %v1, i32 1
  %ret2 = insertelement <32 x i16> %ret1, i16 %v2, i32 2
  %ret3 = insertelement <32 x i16> %ret2, i16 %v3, i32 3
  %ret4 = insertelement <32 x i16> %ret3, i16 %v4, i32 4
  %ret5 = insertelement <32 x i16> %ret4, i16 %v5, i32 5
  %ret6 = insertelement <32 x i16> %ret5, i16 %v6, i32 6
  %ret7 = insertelement <32 x i16> %ret6, i16 %v7, i32 7
  %ret8 = insertelement <32 x i16> %ret7, i16 %v8, i32 8
  %ret9 = insertelement <32 x i16> %ret8, i16 %v9, i32 9
  %ret10 = insertelement <32 x i16> %ret9, i16 %v10, i32 10
  %ret11 = insertelement <32 x i16> %ret10, i16 %v11, i32 11
  %ret12 = insertelement <32 x i16> %ret11, i16 %v12, i32 12
  %ret13 = insertelement <32 x i16> %ret12, i16 %v13, i32 13
  %ret14 = insertelement <32 x i16> %ret13, i16 %v14, i32 14
  %ret15 = insertelement <32 x i16> %ret14, i16 %v15, i32 15
  %ret16 = insertelement <32 x i16> %ret15, i16 %v16, i32 16
  %ret17 = insertelement <32 x i16> %ret16, i16 %v17, i32 17
  %ret18 = insertelement <32 x i16> %ret17, i16 %v18, i32 18
  %ret19 = insertelement <32 x i16> %ret18, i16 %v19, i32 19
  %ret20 = insertelement <32 x i16> %ret19, i16 %v20, i32 20
  %ret21 = insertelement <32 x i16> %ret20, i16 %v21, i32 21
  %ret22 = insertelement <32 x i16> %ret21, i16 %v22, i32 22
  %ret23 = insertelement <32 x i16> %ret22, i16 %v23, i32 23
  %ret24 = insertelement <32 x i16> %ret23, i16 %v24, i32 24
  %ret25 = insertelement <32 x i16> %ret24, i16 %v25, i32 25
  %ret26 = insertelement <32 x i16> %ret25, i16 %v26, i32 26
  %ret27 = insertelement <32 x i16> %ret26, i16 %v27, i32 27
  %ret28 = insertelement <32 x i16> %ret27, i16 %v28, i32 28
  %ret29 = insertelement <32 x i16> %ret28, i16 %v29, i32 29
  %ret30 = insertelement <32 x i16> %ret29, i16 %v30, i32 30
  %ret31 = insertelement <32 x i16> %ret30, i16 %v31, i32 31
  ret <32 x i16> %ret31
}

define <64 x i8> @var_shuffle_v64i8(<64 x i8> %v, <64 x i8> %indices) nounwind {
; AVX512F-LABEL: var_shuffle_v64i8:
; AVX512F:       # %bb.0:
; AVX512F-NEXT:    pushq %rbp
; AVX512F-NEXT:    movq %rsp, %rbp
; AVX512F-NEXT:    andq $-64, %rsp
; AVX512F-NEXT:    subq $4160, %rsp # imm = 0x1040
; AVX512F-NEXT:    vextracti64x4 $1, %zmm1, %ymm2
; AVX512F-NEXT:    vextracti128 $1, %ymm1, %xmm3
; AVX512F-NEXT:    vextracti128 $1, %ymm2, %xmm4
; AVX512F-NEXT:    vmovd %xmm4, %eax
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, (%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    movzbl 3072(%rsp,%rax), %eax
; AVX512F-NEXT:    vmovd %eax, %xmm0
; AVX512F-NEXT:    vpextrb $1, %xmm4, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $1, 3136(%rsp,%rax), %xmm0, %xmm0
; AVX512F-NEXT:    vpextrb $2, %xmm4, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $2, 3200(%rsp,%rax), %xmm0, %xmm0
; AVX512F-NEXT:    vpextrb $3, %xmm4, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $3, 3264(%rsp,%rax), %xmm0, %xmm0
; AVX512F-NEXT:    vpextrb $4, %xmm4, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $4, 3328(%rsp,%rax), %xmm0, %xmm0
; AVX512F-NEXT:    vpextrb $5, %xmm4, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $5, 3392(%rsp,%rax), %xmm0, %xmm0
; AVX512F-NEXT:    vpextrb $6, %xmm4, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $6, 3456(%rsp,%rax), %xmm0, %xmm0
; AVX512F-NEXT:    vpextrb $7, %xmm4, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $7, 3520(%rsp,%rax), %xmm0, %xmm0
; AVX512F-NEXT:    vpextrb $8, %xmm4, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $8, 3584(%rsp,%rax), %xmm0, %xmm0
; AVX512F-NEXT:    vpextrb $9, %xmm4, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $9, 3648(%rsp,%rax), %xmm0, %xmm0
; AVX512F-NEXT:    vpextrb $10, %xmm4, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $10, 3712(%rsp,%rax), %xmm0, %xmm0
; AVX512F-NEXT:    vpextrb $11, %xmm4, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $11, 3776(%rsp,%rax), %xmm0, %xmm0
; AVX512F-NEXT:    vpextrb $12, %xmm4, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $12, 3840(%rsp,%rax), %xmm0, %xmm0
; AVX512F-NEXT:    vpextrb $13, %xmm4, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $13, 3904(%rsp,%rax), %xmm0, %xmm0
; AVX512F-NEXT:    vpextrb $14, %xmm4, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $14, 3968(%rsp,%rax), %xmm0, %xmm0
; AVX512F-NEXT:    vpextrb $15, %xmm4, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $15, 4032(%rsp,%rax), %xmm0, %xmm0
; AVX512F-NEXT:    vmovd %xmm2, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    movzbl 2048(%rsp,%rax), %eax
; AVX512F-NEXT:    vmovd %eax, %xmm4
; AVX512F-NEXT:    vpextrb $1, %xmm2, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $1, 2112(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrb $2, %xmm2, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $2, 2176(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrb $3, %xmm2, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $3, 2240(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrb $4, %xmm2, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $4, 2304(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrb $5, %xmm2, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $5, 2368(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrb $6, %xmm2, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $6, 2432(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrb $7, %xmm2, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $7, 2496(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrb $8, %xmm2, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $8, 2560(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrb $9, %xmm2, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $9, 2624(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrb $10, %xmm2, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $10, 2688(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrb $11, %xmm2, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $11, 2752(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrb $12, %xmm2, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $12, 2816(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrb $13, %xmm2, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $13, 2880(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrb $14, %xmm2, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $14, 2944(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrb $15, %xmm2, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $15, 3008(%rsp,%rax), %xmm4, %xmm2
; AVX512F-NEXT:    vmovd %xmm3, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    movzbl 1024(%rsp,%rax), %eax
; AVX512F-NEXT:    vmovd %eax, %xmm4
; AVX512F-NEXT:    vpextrb $1, %xmm3, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $1, 1088(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrb $2, %xmm3, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $2, 1152(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrb $3, %xmm3, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $3, 1216(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrb $4, %xmm3, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $4, 1280(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrb $5, %xmm3, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $5, 1344(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrb $6, %xmm3, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $6, 1408(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrb $7, %xmm3, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $7, 1472(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrb $8, %xmm3, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $8, 1536(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrb $9, %xmm3, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $9, 1600(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrb $10, %xmm3, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $10, 1664(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrb $11, %xmm3, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $11, 1728(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrb $12, %xmm3, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $12, 1792(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrb $13, %xmm3, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $13, 1856(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrb $14, %xmm3, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $14, 1920(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrb $15, %xmm3, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $15, 1984(%rsp,%rax), %xmm4, %xmm3
; AVX512F-NEXT:    vmovd %xmm1, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    movzbl (%rsp,%rax), %eax
; AVX512F-NEXT:    vmovd %eax, %xmm4
; AVX512F-NEXT:    vpextrb $1, %xmm1, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $1, 64(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrb $2, %xmm1, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $2, 128(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrb $3, %xmm1, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $3, 192(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrb $4, %xmm1, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $4, 256(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrb $5, %xmm1, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $5, 320(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrb $6, %xmm1, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $6, 384(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrb $7, %xmm1, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $7, 448(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrb $8, %xmm1, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $8, 512(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrb $9, %xmm1, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $9, 576(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrb $10, %xmm1, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $10, 640(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrb $11, %xmm1, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $11, 704(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrb $12, %xmm1, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $12, 768(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrb $13, %xmm1, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $13, 832(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrb $14, %xmm1, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $14, 896(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrb $15, %xmm1, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $15, 960(%rsp,%rax), %xmm4, %xmm1
; AVX512F-NEXT:    vinserti128 $1, %xmm0, %ymm2, %ymm0
; AVX512F-NEXT:    vinserti128 $1, %xmm3, %ymm1, %ymm1
; AVX512F-NEXT:    vinserti64x4 $1, %ymm0, %zmm1, %zmm0
; AVX512F-NEXT:    movq %rbp, %rsp
; AVX512F-NEXT:    popq %rbp
; AVX512F-NEXT:    retq
;
; AVX512BW-LABEL: var_shuffle_v64i8:
; AVX512BW:       # %bb.0:
; AVX512BW-NEXT:    pushq %rbp
; AVX512BW-NEXT:    movq %rsp, %rbp
; AVX512BW-NEXT:    andq $-64, %rsp
; AVX512BW-NEXT:    subq $128, %rsp
; AVX512BW-NEXT:    vextracti128 $1, %ymm1, %xmm2
; AVX512BW-NEXT:    vextracti32x4 $2, %zmm1, %xmm3
; AVX512BW-NEXT:    vextracti32x4 $3, %zmm1, %xmm4
; AVX512BW-NEXT:    vmovd %xmm4, %eax
; AVX512BW-NEXT:    vmovaps %zmm0, (%rsp)
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    movzbl (%rsp,%rax), %eax
; AVX512BW-NEXT:    vmovd %eax, %xmm0
; AVX512BW-NEXT:    vpextrb $1, %xmm4, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $1, (%rsp,%rax), %xmm0, %xmm0
; AVX512BW-NEXT:    vpextrb $2, %xmm4, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $2, (%rsp,%rax), %xmm0, %xmm0
; AVX512BW-NEXT:    vpextrb $3, %xmm4, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $3, (%rsp,%rax), %xmm0, %xmm0
; AVX512BW-NEXT:    vpextrb $4, %xmm4, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $4, (%rsp,%rax), %xmm0, %xmm0
; AVX512BW-NEXT:    vpextrb $5, %xmm4, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $5, (%rsp,%rax), %xmm0, %xmm0
; AVX512BW-NEXT:    vpextrb $6, %xmm4, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $6, (%rsp,%rax), %xmm0, %xmm0
; AVX512BW-NEXT:    vpextrb $7, %xmm4, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $7, (%rsp,%rax), %xmm0, %xmm0
; AVX512BW-NEXT:    vpextrb $8, %xmm4, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $8, (%rsp,%rax), %xmm0, %xmm0
; AVX512BW-NEXT:    vpextrb $9, %xmm4, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $9, (%rsp,%rax), %xmm0, %xmm0
; AVX512BW-NEXT:    vpextrb $10, %xmm4, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $10, (%rsp,%rax), %xmm0, %xmm0
; AVX512BW-NEXT:    vpextrb $11, %xmm4, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $11, (%rsp,%rax), %xmm0, %xmm0
; AVX512BW-NEXT:    vpextrb $12, %xmm4, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $12, (%rsp,%rax), %xmm0, %xmm0
; AVX512BW-NEXT:    vpextrb $13, %xmm4, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $13, (%rsp,%rax), %xmm0, %xmm0
; AVX512BW-NEXT:    vpextrb $14, %xmm4, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $14, (%rsp,%rax), %xmm0, %xmm0
; AVX512BW-NEXT:    vpextrb $15, %xmm4, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $15, (%rsp,%rax), %xmm0, %xmm0
; AVX512BW-NEXT:    vmovd %xmm3, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    movzbl (%rsp,%rax), %eax
; AVX512BW-NEXT:    vmovd %eax, %xmm4
; AVX512BW-NEXT:    vpextrb $1, %xmm3, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $1, (%rsp,%rax), %xmm4, %xmm4
; AVX512BW-NEXT:    vpextrb $2, %xmm3, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $2, (%rsp,%rax), %xmm4, %xmm4
; AVX512BW-NEXT:    vpextrb $3, %xmm3, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $3, (%rsp,%rax), %xmm4, %xmm4
; AVX512BW-NEXT:    vpextrb $4, %xmm3, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $4, (%rsp,%rax), %xmm4, %xmm4
; AVX512BW-NEXT:    vpextrb $5, %xmm3, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $5, (%rsp,%rax), %xmm4, %xmm4
; AVX512BW-NEXT:    vpextrb $6, %xmm3, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $6, (%rsp,%rax), %xmm4, %xmm4
; AVX512BW-NEXT:    vpextrb $7, %xmm3, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $7, (%rsp,%rax), %xmm4, %xmm4
; AVX512BW-NEXT:    vpextrb $8, %xmm3, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $8, (%rsp,%rax), %xmm4, %xmm4
; AVX512BW-NEXT:    vpextrb $9, %xmm3, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $9, (%rsp,%rax), %xmm4, %xmm4
; AVX512BW-NEXT:    vpextrb $10, %xmm3, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $10, (%rsp,%rax), %xmm4, %xmm4
; AVX512BW-NEXT:    vpextrb $11, %xmm3, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $11, (%rsp,%rax), %xmm4, %xmm4
; AVX512BW-NEXT:    vpextrb $12, %xmm3, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $12, (%rsp,%rax), %xmm4, %xmm4
; AVX512BW-NEXT:    vpextrb $13, %xmm3, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    movzbl (%rsp,%rax), %eax
; AVX512BW-NEXT:    vpinsrb $13, %eax, %xmm4, %xmm4
; AVX512BW-NEXT:    vpextrb $14, %xmm3, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    movzbl (%rsp,%rax), %eax
; AVX512BW-NEXT:    vpinsrb $14, %eax, %xmm4, %xmm4
; AVX512BW-NEXT:    vpextrb $15, %xmm3, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    movzbl (%rsp,%rax), %eax
; AVX512BW-NEXT:    vpinsrb $15, %eax, %xmm4, %xmm3
; AVX512BW-NEXT:    vmovd %xmm2, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    movzbl (%rsp,%rax), %eax
; AVX512BW-NEXT:    vmovd %eax, %xmm4
; AVX512BW-NEXT:    vpextrb $1, %xmm2, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $1, (%rsp,%rax), %xmm4, %xmm4
; AVX512BW-NEXT:    vpextrb $2, %xmm2, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $2, (%rsp,%rax), %xmm4, %xmm4
; AVX512BW-NEXT:    vpextrb $3, %xmm2, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $3, (%rsp,%rax), %xmm4, %xmm4
; AVX512BW-NEXT:    vpextrb $4, %xmm2, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $4, (%rsp,%rax), %xmm4, %xmm4
; AVX512BW-NEXT:    vpextrb $5, %xmm2, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $5, (%rsp,%rax), %xmm4, %xmm4
; AVX512BW-NEXT:    vpextrb $6, %xmm2, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $6, (%rsp,%rax), %xmm4, %xmm4
; AVX512BW-NEXT:    vpextrb $7, %xmm2, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $7, (%rsp,%rax), %xmm4, %xmm4
; AVX512BW-NEXT:    vpextrb $8, %xmm2, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $8, (%rsp,%rax), %xmm4, %xmm4
; AVX512BW-NEXT:    vpextrb $9, %xmm2, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $9, (%rsp,%rax), %xmm4, %xmm4
; AVX512BW-NEXT:    vpextrb $10, %xmm2, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $10, (%rsp,%rax), %xmm4, %xmm4
; AVX512BW-NEXT:    vpextrb $11, %xmm2, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $11, (%rsp,%rax), %xmm4, %xmm4
; AVX512BW-NEXT:    vpextrb $12, %xmm2, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    movzbl (%rsp,%rax), %eax
; AVX512BW-NEXT:    vpinsrb $12, %eax, %xmm4, %xmm4
; AVX512BW-NEXT:    vpextrb $13, %xmm2, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    movzbl (%rsp,%rax), %eax
; AVX512BW-NEXT:    vpinsrb $13, %eax, %xmm4, %xmm4
; AVX512BW-NEXT:    vpextrb $14, %xmm2, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    movzbl (%rsp,%rax), %eax
; AVX512BW-NEXT:    vpinsrb $14, %eax, %xmm4, %xmm4
; AVX512BW-NEXT:    vpextrb $15, %xmm2, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    movzbl (%rsp,%rax), %eax
; AVX512BW-NEXT:    vpinsrb $15, %eax, %xmm4, %xmm2
; AVX512BW-NEXT:    vmovd %xmm1, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    movzbl (%rsp,%rax), %eax
; AVX512BW-NEXT:    vmovd %eax, %xmm4
; AVX512BW-NEXT:    vpextrb $1, %xmm1, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $1, (%rsp,%rax), %xmm4, %xmm4
; AVX512BW-NEXT:    vpextrb $2, %xmm1, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $2, (%rsp,%rax), %xmm4, %xmm4
; AVX512BW-NEXT:    vpextrb $3, %xmm1, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $3, (%rsp,%rax), %xmm4, %xmm4
; AVX512BW-NEXT:    vpextrb $4, %xmm1, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $4, (%rsp,%rax), %xmm4, %xmm4
; AVX512BW-NEXT:    vpextrb $5, %xmm1, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $5, (%rsp,%rax), %xmm4, %xmm4
; AVX512BW-NEXT:    vpextrb $6, %xmm1, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $6, (%rsp,%rax), %xmm4, %xmm4
; AVX512BW-NEXT:    vpextrb $7, %xmm1, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $7, (%rsp,%rax), %xmm4, %xmm4
; AVX512BW-NEXT:    vpextrb $8, %xmm1, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $8, (%rsp,%rax), %xmm4, %xmm4
; AVX512BW-NEXT:    vpextrb $9, %xmm1, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $9, (%rsp,%rax), %xmm4, %xmm4
; AVX512BW-NEXT:    vpextrb $10, %xmm1, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $10, (%rsp,%rax), %xmm4, %xmm4
; AVX512BW-NEXT:    vpextrb $11, %xmm1, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    movzbl (%rsp,%rax), %eax
; AVX512BW-NEXT:    vpinsrb $11, %eax, %xmm4, %xmm4
; AVX512BW-NEXT:    vpextrb $12, %xmm1, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    movzbl (%rsp,%rax), %eax
; AVX512BW-NEXT:    vpinsrb $12, %eax, %xmm4, %xmm4
; AVX512BW-NEXT:    vpextrb $13, %xmm1, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    movzbl (%rsp,%rax), %eax
; AVX512BW-NEXT:    vpinsrb $13, %eax, %xmm4, %xmm4
; AVX512BW-NEXT:    vpextrb $14, %xmm1, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    movzbl (%rsp,%rax), %eax
; AVX512BW-NEXT:    vpinsrb $14, %eax, %xmm4, %xmm4
; AVX512BW-NEXT:    vpextrb $15, %xmm1, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    movzbl (%rsp,%rax), %eax
; AVX512BW-NEXT:    vpinsrb $15, %eax, %xmm4, %xmm1
; AVX512BW-NEXT:    vinserti128 $1, %xmm0, %ymm3, %ymm0
; AVX512BW-NEXT:    vinserti128 $1, %xmm2, %ymm1, %ymm1
; AVX512BW-NEXT:    vinserti64x4 $1, %ymm0, %zmm1, %zmm0
; AVX512BW-NEXT:    movq %rbp, %rsp
; AVX512BW-NEXT:    popq %rbp
; AVX512BW-NEXT:    retq
;
; AVX512VBMI-LABEL: var_shuffle_v64i8:
; AVX512VBMI:       # %bb.0:
; AVX512VBMI-NEXT:    vpermb %zmm0, %zmm1, %zmm0
; AVX512VBMI-NEXT:    retq
  %index0 = extractelement <64 x i8> %indices, i32 0
  %index1 = extractelement <64 x i8> %indices, i32 1
  %index2 = extractelement <64 x i8> %indices, i32 2
  %index3 = extractelement <64 x i8> %indices, i32 3
  %index4 = extractelement <64 x i8> %indices, i32 4
  %index5 = extractelement <64 x i8> %indices, i32 5
  %index6 = extractelement <64 x i8> %indices, i32 6
  %index7 = extractelement <64 x i8> %indices, i32 7
  %index8 = extractelement <64 x i8> %indices, i32 8
  %index9 = extractelement <64 x i8> %indices, i32 9
  %index10 = extractelement <64 x i8> %indices, i32 10
  %index11 = extractelement <64 x i8> %indices, i32 11
  %index12 = extractelement <64 x i8> %indices, i32 12
  %index13 = extractelement <64 x i8> %indices, i32 13
  %index14 = extractelement <64 x i8> %indices, i32 14
  %index15 = extractelement <64 x i8> %indices, i32 15
  %index16 = extractelement <64 x i8> %indices, i32 16
  %index17 = extractelement <64 x i8> %indices, i32 17
  %index18 = extractelement <64 x i8> %indices, i32 18
  %index19 = extractelement <64 x i8> %indices, i32 19
  %index20 = extractelement <64 x i8> %indices, i32 20
  %index21 = extractelement <64 x i8> %indices, i32 21
  %index22 = extractelement <64 x i8> %indices, i32 22
  %index23 = extractelement <64 x i8> %indices, i32 23
  %index24 = extractelement <64 x i8> %indices, i32 24
  %index25 = extractelement <64 x i8> %indices, i32 25
  %index26 = extractelement <64 x i8> %indices, i32 26
  %index27 = extractelement <64 x i8> %indices, i32 27
  %index28 = extractelement <64 x i8> %indices, i32 28
  %index29 = extractelement <64 x i8> %indices, i32 29
  %index30 = extractelement <64 x i8> %indices, i32 30
  %index31 = extractelement <64 x i8> %indices, i32 31
  %index32 = extractelement <64 x i8> %indices, i32 32
  %index33 = extractelement <64 x i8> %indices, i32 33
  %index34 = extractelement <64 x i8> %indices, i32 34
  %index35 = extractelement <64 x i8> %indices, i32 35
  %index36 = extractelement <64 x i8> %indices, i32 36
  %index37 = extractelement <64 x i8> %indices, i32 37
  %index38 = extractelement <64 x i8> %indices, i32 38
  %index39 = extractelement <64 x i8> %indices, i32 39
  %index40 = extractelement <64 x i8> %indices, i32 40
  %index41 = extractelement <64 x i8> %indices, i32 41
  %index42 = extractelement <64 x i8> %indices, i32 42
  %index43 = extractelement <64 x i8> %indices, i32 43
  %index44 = extractelement <64 x i8> %indices, i32 44
  %index45 = extractelement <64 x i8> %indices, i32 45
  %index46 = extractelement <64 x i8> %indices, i32 46
  %index47 = extractelement <64 x i8> %indices, i32 47
  %index48 = extractelement <64 x i8> %indices, i32 48
  %index49 = extractelement <64 x i8> %indices, i32 49
  %index50 = extractelement <64 x i8> %indices, i32 50
  %index51 = extractelement <64 x i8> %indices, i32 51
  %index52 = extractelement <64 x i8> %indices, i32 52
  %index53 = extractelement <64 x i8> %indices, i32 53
  %index54 = extractelement <64 x i8> %indices, i32 54
  %index55 = extractelement <64 x i8> %indices, i32 55
  %index56 = extractelement <64 x i8> %indices, i32 56
  %index57 = extractelement <64 x i8> %indices, i32 57
  %index58 = extractelement <64 x i8> %indices, i32 58
  %index59 = extractelement <64 x i8> %indices, i32 59
  %index60 = extractelement <64 x i8> %indices, i32 60
  %index61 = extractelement <64 x i8> %indices, i32 61
  %index62 = extractelement <64 x i8> %indices, i32 62
  %index63 = extractelement <64 x i8> %indices, i32 63
  %v0 = extractelement <64 x i8> %v, i8 %index0
  %v1 = extractelement <64 x i8> %v, i8 %index1
  %v2 = extractelement <64 x i8> %v, i8 %index2
  %v3 = extractelement <64 x i8> %v, i8 %index3
  %v4 = extractelement <64 x i8> %v, i8 %index4
  %v5 = extractelement <64 x i8> %v, i8 %index5
  %v6 = extractelement <64 x i8> %v, i8 %index6
  %v7 = extractelement <64 x i8> %v, i8 %index7
  %v8 = extractelement <64 x i8> %v, i8 %index8
  %v9 = extractelement <64 x i8> %v, i8 %index9
  %v10 = extractelement <64 x i8> %v, i8 %index10
  %v11 = extractelement <64 x i8> %v, i8 %index11
  %v12 = extractelement <64 x i8> %v, i8 %index12
  %v13 = extractelement <64 x i8> %v, i8 %index13
  %v14 = extractelement <64 x i8> %v, i8 %index14
  %v15 = extractelement <64 x i8> %v, i8 %index15
  %v16 = extractelement <64 x i8> %v, i8 %index16
  %v17 = extractelement <64 x i8> %v, i8 %index17
  %v18 = extractelement <64 x i8> %v, i8 %index18
  %v19 = extractelement <64 x i8> %v, i8 %index19
  %v20 = extractelement <64 x i8> %v, i8 %index20
  %v21 = extractelement <64 x i8> %v, i8 %index21
  %v22 = extractelement <64 x i8> %v, i8 %index22
  %v23 = extractelement <64 x i8> %v, i8 %index23
  %v24 = extractelement <64 x i8> %v, i8 %index24
  %v25 = extractelement <64 x i8> %v, i8 %index25
  %v26 = extractelement <64 x i8> %v, i8 %index26
  %v27 = extractelement <64 x i8> %v, i8 %index27
  %v28 = extractelement <64 x i8> %v, i8 %index28
  %v29 = extractelement <64 x i8> %v, i8 %index29
  %v30 = extractelement <64 x i8> %v, i8 %index30
  %v31 = extractelement <64 x i8> %v, i8 %index31
  %v32 = extractelement <64 x i8> %v, i8 %index32
  %v33 = extractelement <64 x i8> %v, i8 %index33
  %v34 = extractelement <64 x i8> %v, i8 %index34
  %v35 = extractelement <64 x i8> %v, i8 %index35
  %v36 = extractelement <64 x i8> %v, i8 %index36
  %v37 = extractelement <64 x i8> %v, i8 %index37
  %v38 = extractelement <64 x i8> %v, i8 %index38
  %v39 = extractelement <64 x i8> %v, i8 %index39
  %v40 = extractelement <64 x i8> %v, i8 %index40
  %v41 = extractelement <64 x i8> %v, i8 %index41
  %v42 = extractelement <64 x i8> %v, i8 %index42
  %v43 = extractelement <64 x i8> %v, i8 %index43
  %v44 = extractelement <64 x i8> %v, i8 %index44
  %v45 = extractelement <64 x i8> %v, i8 %index45
  %v46 = extractelement <64 x i8> %v, i8 %index46
  %v47 = extractelement <64 x i8> %v, i8 %index47
  %v48 = extractelement <64 x i8> %v, i8 %index48
  %v49 = extractelement <64 x i8> %v, i8 %index49
  %v50 = extractelement <64 x i8> %v, i8 %index50
  %v51 = extractelement <64 x i8> %v, i8 %index51
  %v52 = extractelement <64 x i8> %v, i8 %index52
  %v53 = extractelement <64 x i8> %v, i8 %index53
  %v54 = extractelement <64 x i8> %v, i8 %index54
  %v55 = extractelement <64 x i8> %v, i8 %index55
  %v56 = extractelement <64 x i8> %v, i8 %index56
  %v57 = extractelement <64 x i8> %v, i8 %index57
  %v58 = extractelement <64 x i8> %v, i8 %index58
  %v59 = extractelement <64 x i8> %v, i8 %index59
  %v60 = extractelement <64 x i8> %v, i8 %index60
  %v61 = extractelement <64 x i8> %v, i8 %index61
  %v62 = extractelement <64 x i8> %v, i8 %index62
  %v63 = extractelement <64 x i8> %v, i8 %index63
  %ret0 = insertelement <64 x i8> undef, i8 %v0, i32 0
  %ret1 = insertelement <64 x i8> %ret0, i8 %v1, i32 1
  %ret2 = insertelement <64 x i8> %ret1, i8 %v2, i32 2
  %ret3 = insertelement <64 x i8> %ret2, i8 %v3, i32 3
  %ret4 = insertelement <64 x i8> %ret3, i8 %v4, i32 4
  %ret5 = insertelement <64 x i8> %ret4, i8 %v5, i32 5
  %ret6 = insertelement <64 x i8> %ret5, i8 %v6, i32 6
  %ret7 = insertelement <64 x i8> %ret6, i8 %v7, i32 7
  %ret8 = insertelement <64 x i8> %ret7, i8 %v8, i32 8
  %ret9 = insertelement <64 x i8> %ret8, i8 %v9, i32 9
  %ret10 = insertelement <64 x i8> %ret9, i8 %v10, i32 10
  %ret11 = insertelement <64 x i8> %ret10, i8 %v11, i32 11
  %ret12 = insertelement <64 x i8> %ret11, i8 %v12, i32 12
  %ret13 = insertelement <64 x i8> %ret12, i8 %v13, i32 13
  %ret14 = insertelement <64 x i8> %ret13, i8 %v14, i32 14
  %ret15 = insertelement <64 x i8> %ret14, i8 %v15, i32 15
  %ret16 = insertelement <64 x i8> %ret15, i8 %v16, i32 16
  %ret17 = insertelement <64 x i8> %ret16, i8 %v17, i32 17
  %ret18 = insertelement <64 x i8> %ret17, i8 %v18, i32 18
  %ret19 = insertelement <64 x i8> %ret18, i8 %v19, i32 19
  %ret20 = insertelement <64 x i8> %ret19, i8 %v20, i32 20
  %ret21 = insertelement <64 x i8> %ret20, i8 %v21, i32 21
  %ret22 = insertelement <64 x i8> %ret21, i8 %v22, i32 22
  %ret23 = insertelement <64 x i8> %ret22, i8 %v23, i32 23
  %ret24 = insertelement <64 x i8> %ret23, i8 %v24, i32 24
  %ret25 = insertelement <64 x i8> %ret24, i8 %v25, i32 25
  %ret26 = insertelement <64 x i8> %ret25, i8 %v26, i32 26
  %ret27 = insertelement <64 x i8> %ret26, i8 %v27, i32 27
  %ret28 = insertelement <64 x i8> %ret27, i8 %v28, i32 28
  %ret29 = insertelement <64 x i8> %ret28, i8 %v29, i32 29
  %ret30 = insertelement <64 x i8> %ret29, i8 %v30, i32 30
  %ret31 = insertelement <64 x i8> %ret30, i8 %v31, i32 31
  %ret32 = insertelement <64 x i8> %ret31, i8 %v32, i32 32
  %ret33 = insertelement <64 x i8> %ret32, i8 %v33, i32 33
  %ret34 = insertelement <64 x i8> %ret33, i8 %v34, i32 34
  %ret35 = insertelement <64 x i8> %ret34, i8 %v35, i32 35
  %ret36 = insertelement <64 x i8> %ret35, i8 %v36, i32 36
  %ret37 = insertelement <64 x i8> %ret36, i8 %v37, i32 37
  %ret38 = insertelement <64 x i8> %ret37, i8 %v38, i32 38
  %ret39 = insertelement <64 x i8> %ret38, i8 %v39, i32 39
  %ret40 = insertelement <64 x i8> %ret39, i8 %v40, i32 40
  %ret41 = insertelement <64 x i8> %ret40, i8 %v41, i32 41
  %ret42 = insertelement <64 x i8> %ret41, i8 %v42, i32 42
  %ret43 = insertelement <64 x i8> %ret42, i8 %v43, i32 43
  %ret44 = insertelement <64 x i8> %ret43, i8 %v44, i32 44
  %ret45 = insertelement <64 x i8> %ret44, i8 %v45, i32 45
  %ret46 = insertelement <64 x i8> %ret45, i8 %v46, i32 46
  %ret47 = insertelement <64 x i8> %ret46, i8 %v47, i32 47
  %ret48 = insertelement <64 x i8> %ret47, i8 %v48, i32 48
  %ret49 = insertelement <64 x i8> %ret48, i8 %v49, i32 49
  %ret50 = insertelement <64 x i8> %ret49, i8 %v50, i32 50
  %ret51 = insertelement <64 x i8> %ret50, i8 %v51, i32 51
  %ret52 = insertelement <64 x i8> %ret51, i8 %v52, i32 52
  %ret53 = insertelement <64 x i8> %ret52, i8 %v53, i32 53
  %ret54 = insertelement <64 x i8> %ret53, i8 %v54, i32 54
  %ret55 = insertelement <64 x i8> %ret54, i8 %v55, i32 55
  %ret56 = insertelement <64 x i8> %ret55, i8 %v56, i32 56
  %ret57 = insertelement <64 x i8> %ret56, i8 %v57, i32 57
  %ret58 = insertelement <64 x i8> %ret57, i8 %v58, i32 58
  %ret59 = insertelement <64 x i8> %ret58, i8 %v59, i32 59
  %ret60 = insertelement <64 x i8> %ret59, i8 %v60, i32 60
  %ret61 = insertelement <64 x i8> %ret60, i8 %v61, i32 61
  %ret62 = insertelement <64 x i8> %ret61, i8 %v62, i32 62
  %ret63 = insertelement <64 x i8> %ret62, i8 %v63, i32 63
  ret <64 x i8> %ret63
}

define <8 x double> @var_shuffle_v8f64(<8 x double> %v, <8 x i64> %indices) nounwind {
; AVX512-LABEL: var_shuffle_v8f64:
; AVX512:       # %bb.0:
; AVX512-NEXT:    vpermpd %zmm0, %zmm1, %zmm0
; AVX512-NEXT:    retq
  %index0 = extractelement <8 x i64> %indices, i32 0
  %index1 = extractelement <8 x i64> %indices, i32 1
  %index2 = extractelement <8 x i64> %indices, i32 2
  %index3 = extractelement <8 x i64> %indices, i32 3
  %index4 = extractelement <8 x i64> %indices, i32 4
  %index5 = extractelement <8 x i64> %indices, i32 5
  %index6 = extractelement <8 x i64> %indices, i32 6
  %index7 = extractelement <8 x i64> %indices, i32 7
  %v0 = extractelement <8 x double> %v, i64 %index0
  %v1 = extractelement <8 x double> %v, i64 %index1
  %v2 = extractelement <8 x double> %v, i64 %index2
  %v3 = extractelement <8 x double> %v, i64 %index3
  %v4 = extractelement <8 x double> %v, i64 %index4
  %v5 = extractelement <8 x double> %v, i64 %index5
  %v6 = extractelement <8 x double> %v, i64 %index6
  %v7 = extractelement <8 x double> %v, i64 %index7
  %ret0 = insertelement <8 x double> undef, double %v0, i32 0
  %ret1 = insertelement <8 x double> %ret0, double %v1, i32 1
  %ret2 = insertelement <8 x double> %ret1, double %v2, i32 2
  %ret3 = insertelement <8 x double> %ret2, double %v3, i32 3
  %ret4 = insertelement <8 x double> %ret3, double %v4, i32 4
  %ret5 = insertelement <8 x double> %ret4, double %v5, i32 5
  %ret6 = insertelement <8 x double> %ret5, double %v6, i32 6
  %ret7 = insertelement <8 x double> %ret6, double %v7, i32 7
  ret <8 x double> %ret7
}

define <16 x float> @var_shuffle_v16f32(<16 x float> %v, <16 x i32> %indices) nounwind {
; AVX512-LABEL: var_shuffle_v16f32:
; AVX512:       # %bb.0:
; AVX512-NEXT:    vpermps %zmm0, %zmm1, %zmm0
; AVX512-NEXT:    retq
  %index0 = extractelement <16 x i32> %indices, i32 0
  %index1 = extractelement <16 x i32> %indices, i32 1
  %index2 = extractelement <16 x i32> %indices, i32 2
  %index3 = extractelement <16 x i32> %indices, i32 3
  %index4 = extractelement <16 x i32> %indices, i32 4
  %index5 = extractelement <16 x i32> %indices, i32 5
  %index6 = extractelement <16 x i32> %indices, i32 6
  %index7 = extractelement <16 x i32> %indices, i32 7
  %index8 = extractelement <16 x i32> %indices, i32 8
  %index9 = extractelement <16 x i32> %indices, i32 9
  %index10 = extractelement <16 x i32> %indices, i32 10
  %index11 = extractelement <16 x i32> %indices, i32 11
  %index12 = extractelement <16 x i32> %indices, i32 12
  %index13 = extractelement <16 x i32> %indices, i32 13
  %index14 = extractelement <16 x i32> %indices, i32 14
  %index15 = extractelement <16 x i32> %indices, i32 15
  %v0 = extractelement <16 x float> %v, i32 %index0
  %v1 = extractelement <16 x float> %v, i32 %index1
  %v2 = extractelement <16 x float> %v, i32 %index2
  %v3 = extractelement <16 x float> %v, i32 %index3
  %v4 = extractelement <16 x float> %v, i32 %index4
  %v5 = extractelement <16 x float> %v, i32 %index5
  %v6 = extractelement <16 x float> %v, i32 %index6
  %v7 = extractelement <16 x float> %v, i32 %index7
  %v8 = extractelement <16 x float> %v, i32 %index8
  %v9 = extractelement <16 x float> %v, i32 %index9
  %v10 = extractelement <16 x float> %v, i32 %index10
  %v11 = extractelement <16 x float> %v, i32 %index11
  %v12 = extractelement <16 x float> %v, i32 %index12
  %v13 = extractelement <16 x float> %v, i32 %index13
  %v14 = extractelement <16 x float> %v, i32 %index14
  %v15 = extractelement <16 x float> %v, i32 %index15
  %ret0 = insertelement <16 x float> undef, float %v0, i32 0
  %ret1 = insertelement <16 x float> %ret0, float %v1, i32 1
  %ret2 = insertelement <16 x float> %ret1, float %v2, i32 2
  %ret3 = insertelement <16 x float> %ret2, float %v3, i32 3
  %ret4 = insertelement <16 x float> %ret3, float %v4, i32 4
  %ret5 = insertelement <16 x float> %ret4, float %v5, i32 5
  %ret6 = insertelement <16 x float> %ret5, float %v6, i32 6
  %ret7 = insertelement <16 x float> %ret6, float %v7, i32 7
  %ret8 = insertelement <16 x float> %ret7, float %v8, i32 8
  %ret9 = insertelement <16 x float> %ret8, float %v9, i32 9
  %ret10 = insertelement <16 x float> %ret9, float %v10, i32 10
  %ret11 = insertelement <16 x float> %ret10, float %v11, i32 11
  %ret12 = insertelement <16 x float> %ret11, float %v12, i32 12
  %ret13 = insertelement <16 x float> %ret12, float %v13, i32 13
  %ret14 = insertelement <16 x float> %ret13, float %v14, i32 14
  %ret15 = insertelement <16 x float> %ret14, float %v15, i32 15
  ret <16 x float> %ret15
}

define void @var_cvt_shuffle_v64f32_v64i8_idx(<64 x float>* %dst, <64 x i8> %src, i32 %b) nounwind {
; AVX512F-LABEL: var_cvt_shuffle_v64f32_v64i8_idx:
; AVX512F:       # %bb.0:
; AVX512F-NEXT:    pushq %rbp
; AVX512F-NEXT:    movq %rsp, %rbp
; AVX512F-NEXT:    andq $-64, %rsp
; AVX512F-NEXT:    subq $4160, %rsp # imm = 0x1040
; AVX512F-NEXT:    # kill: def $esi killed $esi def $rsi
; AVX512F-NEXT:    vpbroadcastd %esi, %zmm4
; AVX512F-NEXT:    vpaddd {{.*}}(%rip), %zmm4, %zmm1
; AVX512F-NEXT:    vpaddd {{.*}}(%rip), %zmm4, %zmm2
; AVX512F-NEXT:    vpaddd {{.*}}(%rip), %zmm4, %zmm3
; AVX512F-NEXT:    vpaddd {{.*}}(%rip), %zmm4, %zmm4
; AVX512F-NEXT:    vmovd %xmm4, %eax
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vextractf64x4 $1, %zmm0, {{[0-9]+}}(%rsp)
; AVX512F-NEXT:    vmovaps %ymm0, (%rsp)
; AVX512F-NEXT:    movzbl 3968(%rsp,%rax), %eax
; AVX512F-NEXT:    vmovd %eax, %xmm0
; AVX512F-NEXT:    vpextrd $1, %xmm4, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $1, 3904(%rsp,%rax), %xmm0, %xmm0
; AVX512F-NEXT:    vpextrd $2, %xmm4, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $2, 3840(%rsp,%rax), %xmm0, %xmm0
; AVX512F-NEXT:    vpextrd $3, %xmm4, %eax
; AVX512F-NEXT:    vextracti128 $1, %ymm4, %xmm5
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $3, 3776(%rsp,%rax), %xmm0, %xmm0
; AVX512F-NEXT:    vmovd %xmm5, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $4, 3712(%rsp,%rax), %xmm0, %xmm0
; AVX512F-NEXT:    vpextrd $1, %xmm5, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $5, 3648(%rsp,%rax), %xmm0, %xmm0
; AVX512F-NEXT:    vpextrd $2, %xmm5, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $6, 3584(%rsp,%rax), %xmm0, %xmm0
; AVX512F-NEXT:    vpextrd $3, %xmm5, %eax
; AVX512F-NEXT:    vextracti32x4 $2, %zmm4, %xmm5
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $7, 3520(%rsp,%rax), %xmm0, %xmm0
; AVX512F-NEXT:    vmovd %xmm5, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $8, 3456(%rsp,%rax), %xmm0, %xmm0
; AVX512F-NEXT:    vpextrd $1, %xmm5, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $9, 3392(%rsp,%rax), %xmm0, %xmm0
; AVX512F-NEXT:    vpextrd $2, %xmm5, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $10, 3328(%rsp,%rax), %xmm0, %xmm0
; AVX512F-NEXT:    vpextrd $3, %xmm5, %eax
; AVX512F-NEXT:    vextracti32x4 $3, %zmm4, %xmm4
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $11, 3264(%rsp,%rax), %xmm0, %xmm0
; AVX512F-NEXT:    vmovd %xmm4, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $12, 3200(%rsp,%rax), %xmm0, %xmm0
; AVX512F-NEXT:    vpextrd $1, %xmm4, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $13, 3136(%rsp,%rax), %xmm0, %xmm0
; AVX512F-NEXT:    vpextrd $2, %xmm4, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $14, 3072(%rsp,%rax), %xmm0, %xmm0
; AVX512F-NEXT:    vpextrd $3, %xmm4, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $15, 3008(%rsp,%rax), %xmm0, %xmm0
; AVX512F-NEXT:    vmovd %xmm3, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    movzbl 2944(%rsp,%rax), %eax
; AVX512F-NEXT:    vmovd %eax, %xmm4
; AVX512F-NEXT:    vpextrd $1, %xmm3, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $1, 2880(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrd $2, %xmm3, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $2, 2816(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrd $3, %xmm3, %eax
; AVX512F-NEXT:    vextracti128 $1, %ymm3, %xmm5
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $3, 2752(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vmovd %xmm5, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $4, 2688(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrd $1, %xmm5, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $5, 2624(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrd $2, %xmm5, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $6, 2560(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrd $3, %xmm5, %eax
; AVX512F-NEXT:    vextracti32x4 $2, %zmm3, %xmm5
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $7, 2496(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vmovd %xmm5, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $8, 2432(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrd $1, %xmm5, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $9, 2368(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrd $2, %xmm5, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $10, 2304(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrd $3, %xmm5, %eax
; AVX512F-NEXT:    vextracti32x4 $3, %zmm3, %xmm3
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $11, 2240(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vmovd %xmm3, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $12, 2176(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrd $1, %xmm3, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $13, 2112(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrd $2, %xmm3, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $14, 2048(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrd $3, %xmm3, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $15, 1984(%rsp,%rax), %xmm4, %xmm3
; AVX512F-NEXT:    vmovd %xmm2, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    movzbl 1920(%rsp,%rax), %eax
; AVX512F-NEXT:    vmovd %eax, %xmm4
; AVX512F-NEXT:    vpextrd $1, %xmm2, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $1, 1856(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrd $2, %xmm2, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $2, 1792(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrd $3, %xmm2, %eax
; AVX512F-NEXT:    vextracti128 $1, %ymm2, %xmm5
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $3, 1728(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vmovd %xmm5, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $4, 1664(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrd $1, %xmm5, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $5, 1600(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrd $2, %xmm5, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $6, 1536(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrd $3, %xmm5, %eax
; AVX512F-NEXT:    vextracti32x4 $2, %zmm2, %xmm5
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $7, 1472(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vmovd %xmm5, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $8, 1408(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrd $1, %xmm5, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $9, 1344(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrd $2, %xmm5, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $10, 1280(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrd $3, %xmm5, %eax
; AVX512F-NEXT:    vextracti32x4 $3, %zmm2, %xmm2
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $11, 1216(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vmovd %xmm2, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $12, 1152(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrd $1, %xmm2, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $13, 1088(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrd $2, %xmm2, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $14, 1024(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrd $3, %xmm2, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $15, 960(%rsp,%rax), %xmm4, %xmm2
; AVX512F-NEXT:    vmovd %xmm1, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    movzbl 896(%rsp,%rax), %eax
; AVX512F-NEXT:    vmovd %eax, %xmm4
; AVX512F-NEXT:    vpextrd $1, %xmm1, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $1, 832(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrd $2, %xmm1, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $2, 768(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrd $3, %xmm1, %eax
; AVX512F-NEXT:    vextracti128 $1, %ymm1, %xmm5
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $3, 704(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vmovd %xmm5, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $4, 640(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrd $1, %xmm5, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $5, 576(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrd $2, %xmm5, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $6, 512(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrd $3, %xmm5, %eax
; AVX512F-NEXT:    vextracti32x4 $2, %zmm1, %xmm5
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $7, 448(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vmovd %xmm5, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $8, 384(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrd $1, %xmm5, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $9, 320(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    andl $63, %esi
; AVX512F-NEXT:    vpinsrb $10, 4032(%rsp,%rsi), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrd $3, %xmm5, %eax
; AVX512F-NEXT:    vextracti32x4 $3, %zmm1, %xmm1
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $11, 256(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vmovd %xmm1, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $12, 192(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrd $1, %xmm1, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $13, 128(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrd $2, %xmm1, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $14, 64(%rsp,%rax), %xmm4, %xmm4
; AVX512F-NEXT:    vpextrd $3, %xmm1, %eax
; AVX512F-NEXT:    andl $63, %eax
; AVX512F-NEXT:    vpinsrb $15, (%rsp,%rax), %xmm4, %xmm1
; AVX512F-NEXT:    vpmovsxbd %xmm0, %zmm0
; AVX512F-NEXT:    vcvtdq2ps %zmm0, %zmm0
; AVX512F-NEXT:    vpmovsxbd %xmm3, %zmm3
; AVX512F-NEXT:    vcvtdq2ps %zmm3, %zmm3
; AVX512F-NEXT:    vpmovsxbd %xmm2, %zmm2
; AVX512F-NEXT:    vcvtdq2ps %zmm2, %zmm2
; AVX512F-NEXT:    vpmovsxbd %xmm1, %zmm1
; AVX512F-NEXT:    vcvtdq2ps %zmm1, %zmm1
; AVX512F-NEXT:    vmovaps %zmm1, 192(%rdi)
; AVX512F-NEXT:    vmovaps %zmm2, 128(%rdi)
; AVX512F-NEXT:    vmovaps %zmm3, 64(%rdi)
; AVX512F-NEXT:    vmovaps %zmm0, (%rdi)
; AVX512F-NEXT:    movq %rbp, %rsp
; AVX512F-NEXT:    popq %rbp
; AVX512F-NEXT:    vzeroupper
; AVX512F-NEXT:    retq
;
; AVX512BW-LABEL: var_cvt_shuffle_v64f32_v64i8_idx:
; AVX512BW:       # %bb.0:
; AVX512BW-NEXT:    pushq %rbp
; AVX512BW-NEXT:    movq %rsp, %rbp
; AVX512BW-NEXT:    andq $-64, %rsp
; AVX512BW-NEXT:    subq $128, %rsp
; AVX512BW-NEXT:    # kill: def $esi killed $esi def $rsi
; AVX512BW-NEXT:    vpbroadcastd %esi, %zmm2
; AVX512BW-NEXT:    vpaddd {{.*}}(%rip), %zmm2, %zmm1
; AVX512BW-NEXT:    vmovd %xmm1, %eax
; AVX512BW-NEXT:    vmovaps %zmm0, (%rsp)
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    movzbl (%rsp,%rax), %eax
; AVX512BW-NEXT:    vmovd %eax, %xmm0
; AVX512BW-NEXT:    vpextrd $1, %xmm1, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $1, (%rsp,%rax), %xmm0, %xmm0
; AVX512BW-NEXT:    vpextrd $2, %xmm1, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $2, (%rsp,%rax), %xmm0, %xmm0
; AVX512BW-NEXT:    vpextrd $3, %xmm1, %eax
; AVX512BW-NEXT:    vextracti128 $1, %ymm1, %xmm3
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $3, (%rsp,%rax), %xmm0, %xmm0
; AVX512BW-NEXT:    vmovd %xmm3, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $4, (%rsp,%rax), %xmm0, %xmm0
; AVX512BW-NEXT:    vpextrd $1, %xmm3, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $5, (%rsp,%rax), %xmm0, %xmm0
; AVX512BW-NEXT:    vpextrd $2, %xmm3, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $6, (%rsp,%rax), %xmm0, %xmm0
; AVX512BW-NEXT:    vpextrd $3, %xmm3, %eax
; AVX512BW-NEXT:    vextracti32x4 $2, %zmm1, %xmm4
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $7, (%rsp,%rax), %xmm0, %xmm0
; AVX512BW-NEXT:    vmovd %xmm4, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $8, (%rsp,%rax), %xmm0, %xmm0
; AVX512BW-NEXT:    vpextrd $1, %xmm4, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $9, (%rsp,%rax), %xmm0, %xmm0
; AVX512BW-NEXT:    vpaddd {{.*}}(%rip), %zmm2, %zmm5
; AVX512BW-NEXT:    vpaddd {{.*}}(%rip), %zmm2, %zmm3
; AVX512BW-NEXT:    vpaddd {{.*}}(%rip), %zmm2, %zmm2
; AVX512BW-NEXT:    andl $63, %esi
; AVX512BW-NEXT:    vpinsrb $10, (%rsp,%rsi), %xmm0, %xmm0
; AVX512BW-NEXT:    vpextrd $3, %xmm4, %eax
; AVX512BW-NEXT:    vextracti32x4 $3, %zmm1, %xmm1
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $11, (%rsp,%rax), %xmm0, %xmm0
; AVX512BW-NEXT:    vmovd %xmm1, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $12, (%rsp,%rax), %xmm0, %xmm0
; AVX512BW-NEXT:    vpextrd $1, %xmm1, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $13, (%rsp,%rax), %xmm0, %xmm0
; AVX512BW-NEXT:    vpextrd $2, %xmm1, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $14, (%rsp,%rax), %xmm0, %xmm0
; AVX512BW-NEXT:    vpextrd $3, %xmm1, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $15, (%rsp,%rax), %xmm0, %xmm8
; AVX512BW-NEXT:    vmovd %xmm5, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    movzbl (%rsp,%rax), %eax
; AVX512BW-NEXT:    vmovd %eax, %xmm1
; AVX512BW-NEXT:    vpextrd $1, %xmm5, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $1, (%rsp,%rax), %xmm1, %xmm1
; AVX512BW-NEXT:    vpextrd $2, %xmm5, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $2, (%rsp,%rax), %xmm1, %xmm1
; AVX512BW-NEXT:    vpextrd $3, %xmm5, %eax
; AVX512BW-NEXT:    vextracti128 $1, %ymm5, %xmm4
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $3, (%rsp,%rax), %xmm1, %xmm1
; AVX512BW-NEXT:    vmovd %xmm4, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $4, (%rsp,%rax), %xmm1, %xmm1
; AVX512BW-NEXT:    vpextrd $1, %xmm4, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $5, (%rsp,%rax), %xmm1, %xmm1
; AVX512BW-NEXT:    vpextrd $2, %xmm4, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $6, (%rsp,%rax), %xmm1, %xmm1
; AVX512BW-NEXT:    vpextrd $3, %xmm4, %eax
; AVX512BW-NEXT:    vextracti32x4 $2, %zmm5, %xmm4
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $7, (%rsp,%rax), %xmm1, %xmm1
; AVX512BW-NEXT:    vmovd %xmm4, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $8, (%rsp,%rax), %xmm1, %xmm1
; AVX512BW-NEXT:    vpextrd $1, %xmm4, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $9, (%rsp,%rax), %xmm1, %xmm1
; AVX512BW-NEXT:    vpextrd $2, %xmm4, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $10, (%rsp,%rax), %xmm1, %xmm6
; AVX512BW-NEXT:    vpextrd $3, %xmm4, %eax
; AVX512BW-NEXT:    vextracti32x4 $3, %zmm5, %xmm1
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $11, (%rsp,%rax), %xmm6, %xmm4
; AVX512BW-NEXT:    vmovd %xmm1, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $12, (%rsp,%rax), %xmm4, %xmm4
; AVX512BW-NEXT:    vmovd %xmm3, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    movzbl (%rsp,%rax), %eax
; AVX512BW-NEXT:    vmovd %eax, %xmm5
; AVX512BW-NEXT:    vpextrd $1, %xmm3, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $1, (%rsp,%rax), %xmm5, %xmm5
; AVX512BW-NEXT:    vpextrd $2, %xmm3, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $2, (%rsp,%rax), %xmm5, %xmm5
; AVX512BW-NEXT:    vpextrd $3, %xmm3, %eax
; AVX512BW-NEXT:    vextracti128 $1, %ymm3, %xmm6
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $3, (%rsp,%rax), %xmm5, %xmm5
; AVX512BW-NEXT:    vmovd %xmm6, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $4, (%rsp,%rax), %xmm5, %xmm5
; AVX512BW-NEXT:    vpextrd $1, %xmm6, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $5, (%rsp,%rax), %xmm5, %xmm5
; AVX512BW-NEXT:    vpextrd $2, %xmm6, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $6, (%rsp,%rax), %xmm5, %xmm7
; AVX512BW-NEXT:    vpextrd $3, %xmm6, %eax
; AVX512BW-NEXT:    vextracti32x4 $2, %zmm3, %xmm5
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $7, (%rsp,%rax), %xmm7, %xmm6
; AVX512BW-NEXT:    vmovd %xmm5, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $8, (%rsp,%rax), %xmm6, %xmm6
; AVX512BW-NEXT:    vpextrd $1, %xmm5, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $9, (%rsp,%rax), %xmm6, %xmm6
; AVX512BW-NEXT:    vpextrd $2, %xmm5, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $10, (%rsp,%rax), %xmm6, %xmm6
; AVX512BW-NEXT:    vmovd %xmm2, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    movzbl (%rsp,%rax), %eax
; AVX512BW-NEXT:    vmovd %eax, %xmm7
; AVX512BW-NEXT:    vpextrd $1, %xmm2, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $1, (%rsp,%rax), %xmm7, %xmm7
; AVX512BW-NEXT:    vpextrd $2, %xmm2, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $2, (%rsp,%rax), %xmm7, %xmm7
; AVX512BW-NEXT:    vpextrd $3, %xmm2, %eax
; AVX512BW-NEXT:    vextracti128 $1, %ymm2, %xmm0
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $3, (%rsp,%rax), %xmm7, %xmm7
; AVX512BW-NEXT:    vmovd %xmm0, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $4, (%rsp,%rax), %xmm7, %xmm7
; AVX512BW-NEXT:    vpextrd $1, %xmm0, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $5, (%rsp,%rax), %xmm7, %xmm7
; AVX512BW-NEXT:    vpextrd $2, %xmm0, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $6, (%rsp,%rax), %xmm7, %xmm7
; AVX512BW-NEXT:    vpextrd $3, %xmm0, %eax
; AVX512BW-NEXT:    vextracti32x4 $2, %zmm2, %xmm0
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $7, (%rsp,%rax), %xmm7, %xmm7
; AVX512BW-NEXT:    vmovd %xmm0, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $8, (%rsp,%rax), %xmm7, %xmm7
; AVX512BW-NEXT:    vpextrd $1, %xmm0, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    vpinsrb $9, (%rsp,%rax), %xmm7, %xmm7
; AVX512BW-NEXT:    vpextrd $2, %xmm0, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    movzbl (%rsp,%rax), %eax
; AVX512BW-NEXT:    vpinsrb $10, %eax, %xmm7, %xmm7
; AVX512BW-NEXT:    vpextrd $3, %xmm0, %eax
; AVX512BW-NEXT:    vextracti32x4 $3, %zmm2, %xmm0
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    movzbl (%rsp,%rax), %eax
; AVX512BW-NEXT:    vpinsrb $11, %eax, %xmm7, %xmm2
; AVX512BW-NEXT:    vmovd %xmm0, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    movzbl (%rsp,%rax), %eax
; AVX512BW-NEXT:    vpinsrb $12, %eax, %xmm2, %xmm2
; AVX512BW-NEXT:    vpextrd $1, %xmm0, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    movzbl (%rsp,%rax), %eax
; AVX512BW-NEXT:    vpinsrb $13, %eax, %xmm2, %xmm2
; AVX512BW-NEXT:    vpextrd $2, %xmm0, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    movzbl (%rsp,%rax), %eax
; AVX512BW-NEXT:    vpinsrb $14, %eax, %xmm2, %xmm2
; AVX512BW-NEXT:    vpextrd $3, %xmm0, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    movzbl (%rsp,%rax), %eax
; AVX512BW-NEXT:    vpinsrb $15, %eax, %xmm2, %xmm0
; AVX512BW-NEXT:    vpextrd $3, %xmm5, %eax
; AVX512BW-NEXT:    vextracti32x4 $3, %zmm3, %xmm2
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    movzbl (%rsp,%rax), %eax
; AVX512BW-NEXT:    vpinsrb $11, %eax, %xmm6, %xmm3
; AVX512BW-NEXT:    vmovd %xmm2, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    movzbl (%rsp,%rax), %eax
; AVX512BW-NEXT:    vpinsrb $12, %eax, %xmm3, %xmm3
; AVX512BW-NEXT:    vpextrd $1, %xmm2, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    movzbl (%rsp,%rax), %eax
; AVX512BW-NEXT:    vpinsrb $13, %eax, %xmm3, %xmm3
; AVX512BW-NEXT:    vpextrd $2, %xmm2, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    movzbl (%rsp,%rax), %eax
; AVX512BW-NEXT:    vpinsrb $14, %eax, %xmm3, %xmm3
; AVX512BW-NEXT:    vpextrd $3, %xmm2, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    movzbl (%rsp,%rax), %eax
; AVX512BW-NEXT:    vpinsrb $15, %eax, %xmm3, %xmm2
; AVX512BW-NEXT:    vpextrd $1, %xmm1, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    movzbl (%rsp,%rax), %eax
; AVX512BW-NEXT:    vpinsrb $13, %eax, %xmm4, %xmm3
; AVX512BW-NEXT:    vpextrd $2, %xmm1, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    movzbl (%rsp,%rax), %eax
; AVX512BW-NEXT:    vpinsrb $14, %eax, %xmm3, %xmm3
; AVX512BW-NEXT:    vpextrd $3, %xmm1, %eax
; AVX512BW-NEXT:    andl $63, %eax
; AVX512BW-NEXT:    movzbl (%rsp,%rax), %eax
; AVX512BW-NEXT:    vpinsrb $15, %eax, %xmm3, %xmm1
; AVX512BW-NEXT:    vpmovsxbd %xmm0, %zmm0
; AVX512BW-NEXT:    vcvtdq2ps %zmm0, %zmm0
; AVX512BW-NEXT:    vpmovsxbd %xmm2, %zmm2
; AVX512BW-NEXT:    vcvtdq2ps %zmm2, %zmm2
; AVX512BW-NEXT:    vpmovsxbd %xmm1, %zmm1
; AVX512BW-NEXT:    vcvtdq2ps %zmm1, %zmm1
; AVX512BW-NEXT:    vpmovsxbd %xmm8, %zmm3
; AVX512BW-NEXT:    vcvtdq2ps %zmm3, %zmm3
; AVX512BW-NEXT:    vmovaps %zmm3, 192(%rdi)
; AVX512BW-NEXT:    vmovaps %zmm1, 128(%rdi)
; AVX512BW-NEXT:    vmovaps %zmm2, 64(%rdi)
; AVX512BW-NEXT:    vmovaps %zmm0, (%rdi)
; AVX512BW-NEXT:    movq %rbp, %rsp
; AVX512BW-NEXT:    popq %rbp
; AVX512BW-NEXT:    vzeroupper
; AVX512BW-NEXT:    retq
;
; AVX512VBMI-LABEL: var_cvt_shuffle_v64f32_v64i8_idx:
; AVX512VBMI:       # %bb.0:
; AVX512VBMI-NEXT:    pushq %rbp
; AVX512VBMI-NEXT:    movq %rsp, %rbp
; AVX512VBMI-NEXT:    andq $-64, %rsp
; AVX512VBMI-NEXT:    subq $128, %rsp
; AVX512VBMI-NEXT:    # kill: def $esi killed $esi def $rsi
; AVX512VBMI-NEXT:    vpbroadcastd %esi, %zmm1
; AVX512VBMI-NEXT:    vpaddd {{.*}}(%rip), %zmm1, %zmm2
; AVX512VBMI-NEXT:    vmovd %xmm2, %eax
; AVX512VBMI-NEXT:    vmovdqa64 %zmm0, (%rsp)
; AVX512VBMI-NEXT:    andl $63, %eax
; AVX512VBMI-NEXT:    movzbl (%rsp,%rax), %eax
; AVX512VBMI-NEXT:    vmovd %eax, %xmm3
; AVX512VBMI-NEXT:    vpextrd $1, %xmm2, %eax
; AVX512VBMI-NEXT:    andl $63, %eax
; AVX512VBMI-NEXT:    vpinsrb $1, (%rsp,%rax), %xmm3, %xmm3
; AVX512VBMI-NEXT:    vpextrd $2, %xmm2, %eax
; AVX512VBMI-NEXT:    andl $63, %eax
; AVX512VBMI-NEXT:    vpinsrb $2, (%rsp,%rax), %xmm3, %xmm3
; AVX512VBMI-NEXT:    vpextrd $3, %xmm2, %eax
; AVX512VBMI-NEXT:    vextracti128 $1, %ymm2, %xmm4
; AVX512VBMI-NEXT:    andl $63, %eax
; AVX512VBMI-NEXT:    vpinsrb $3, (%rsp,%rax), %xmm3, %xmm3
; AVX512VBMI-NEXT:    vmovd %xmm4, %eax
; AVX512VBMI-NEXT:    andl $63, %eax
; AVX512VBMI-NEXT:    vpinsrb $4, (%rsp,%rax), %xmm3, %xmm3
; AVX512VBMI-NEXT:    vpextrd $1, %xmm4, %eax
; AVX512VBMI-NEXT:    andl $63, %eax
; AVX512VBMI-NEXT:    vpinsrb $5, (%rsp,%rax), %xmm3, %xmm3
; AVX512VBMI-NEXT:    vpextrd $2, %xmm4, %eax
; AVX512VBMI-NEXT:    andl $63, %eax
; AVX512VBMI-NEXT:    vpinsrb $6, (%rsp,%rax), %xmm3, %xmm3
; AVX512VBMI-NEXT:    vpextrd $3, %xmm4, %eax
; AVX512VBMI-NEXT:    vextracti32x4 $2, %zmm2, %xmm4
; AVX512VBMI-NEXT:    andl $63, %eax
; AVX512VBMI-NEXT:    vpinsrb $7, (%rsp,%rax), %xmm3, %xmm3
; AVX512VBMI-NEXT:    vmovd %xmm4, %eax
; AVX512VBMI-NEXT:    andl $63, %eax
; AVX512VBMI-NEXT:    vpinsrb $8, (%rsp,%rax), %xmm3, %xmm3
; AVX512VBMI-NEXT:    vpextrd $1, %xmm4, %eax
; AVX512VBMI-NEXT:    andl $63, %eax
; AVX512VBMI-NEXT:    vpinsrb $9, (%rsp,%rax), %xmm3, %xmm3
; AVX512VBMI-NEXT:    andl $63, %esi
; AVX512VBMI-NEXT:    vpinsrb $10, (%rsp,%rsi), %xmm3, %xmm3
; AVX512VBMI-NEXT:    vpextrd $3, %xmm4, %eax
; AVX512VBMI-NEXT:    vextracti32x4 $3, %zmm2, %xmm2
; AVX512VBMI-NEXT:    andl $63, %eax
; AVX512VBMI-NEXT:    vpinsrb $11, (%rsp,%rax), %xmm3, %xmm3
; AVX512VBMI-NEXT:    vmovd %xmm2, %eax
; AVX512VBMI-NEXT:    andl $63, %eax
; AVX512VBMI-NEXT:    vpinsrb $12, (%rsp,%rax), %xmm3, %xmm3
; AVX512VBMI-NEXT:    vpextrd $1, %xmm2, %eax
; AVX512VBMI-NEXT:    andl $63, %eax
; AVX512VBMI-NEXT:    vpinsrb $13, (%rsp,%rax), %xmm3, %xmm3
; AVX512VBMI-NEXT:    vpextrd $2, %xmm2, %eax
; AVX512VBMI-NEXT:    andl $63, %eax
; AVX512VBMI-NEXT:    vpinsrb $14, (%rsp,%rax), %xmm3, %xmm3
; AVX512VBMI-NEXT:    vpextrd $3, %xmm2, %eax
; AVX512VBMI-NEXT:    andl $63, %eax
; AVX512VBMI-NEXT:    vpinsrb $15, (%rsp,%rax), %xmm3, %xmm2
; AVX512VBMI-NEXT:    vpaddd {{.*}}(%rip), %zmm1, %zmm3
; AVX512VBMI-NEXT:    vpmovdb %zmm3, %xmm3
; AVX512VBMI-NEXT:    vpaddd {{.*}}(%rip), %zmm1, %zmm4
; AVX512VBMI-NEXT:    vpmovdb %zmm4, %xmm4
; AVX512VBMI-NEXT:    vpaddd {{.*}}(%rip), %zmm1, %zmm1
; AVX512VBMI-NEXT:    vpmovdb %zmm1, %xmm1
; AVX512VBMI-NEXT:    vpmovsxbd %xmm2, %zmm2
; AVX512VBMI-NEXT:    vcvtdq2ps %zmm2, %zmm2
; AVX512VBMI-NEXT:    vpermb %zmm0, %zmm1, %zmm1
; AVX512VBMI-NEXT:    vpmovsxbd %xmm1, %zmm1
; AVX512VBMI-NEXT:    vcvtdq2ps %zmm1, %zmm1
; AVX512VBMI-NEXT:    vpermb %zmm0, %zmm4, %zmm4
; AVX512VBMI-NEXT:    vpmovsxbd %xmm4, %zmm4
; AVX512VBMI-NEXT:    vcvtdq2ps %zmm4, %zmm4
; AVX512VBMI-NEXT:    vpermb %zmm0, %zmm3, %zmm0
; AVX512VBMI-NEXT:    vpmovsxbd %xmm0, %zmm0
; AVX512VBMI-NEXT:    vcvtdq2ps %zmm0, %zmm0
; AVX512VBMI-NEXT:    vmovaps %zmm0, 128(%rdi)
; AVX512VBMI-NEXT:    vmovaps %zmm4, 64(%rdi)
; AVX512VBMI-NEXT:    vmovaps %zmm1, (%rdi)
; AVX512VBMI-NEXT:    vmovaps %zmm2, 192(%rdi)
; AVX512VBMI-NEXT:    movq %rbp, %rsp
; AVX512VBMI-NEXT:    popq %rbp
; AVX512VBMI-NEXT:    vzeroupper
; AVX512VBMI-NEXT:    retq
  %b_broadcast_init = insertelement <64 x i32> undef, i32 %b, i32 0
  %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>
  %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>
  %index_0.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 0
  %index_1.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 1
  %index_2.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 2
  %index_3.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 3
  %index_4.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 4
  %index_5.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 5
  %index_6.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 6
  %index_7.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 7
  %index_8.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 8
  %index_9.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 9
  %index_10.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 10
  %index_11.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 11
  %index_12.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 12
  %index_13.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 13
  %index_14.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 14
  %index_15.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 15
  %index_16.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 16
  %index_17.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 17
  %index_18.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 18
  %index_19.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 19
  %index_20.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 20
  %index_21.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 21
  %index_22.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 22
  %index_23.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 23
  %index_24.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 24
  %index_25.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 25
  %index_26.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 26
  %index_27.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 27
  %index_28.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 28
  %index_29.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 29
  %index_30.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 30
  %index_31.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 31
  %index_32.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 32
  %index_33.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 33
  %index_34.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 34
  %index_35.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 35
  %index_36.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 36
  %index_37.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 37
  %index_38.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 38
  %index_39.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 39
  %index_40.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 40
  %index_41.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 41
  %index_42.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 42
  %index_43.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 43
  %index_44.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 44
  %index_45.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 45
  %index_46.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 46
  %index_47.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 47
  %index_48.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 48
  %index_49.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 49
  %index_50.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 50
  %index_51.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 51
  %index_52.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 52
  %index_53.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 53
  %index_54.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 54
  %index_55.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 55
  %index_56.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 56
  %index_57.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 57
  %index_59.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 59
  %index_60.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 60
  %index_61.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 61
  %index_62.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 62
  %index_63.i.i = extractelement <64 x i32> %sub_add__b_broadcast_, i32 63
  %v_0.i.i = extractelement <64 x i8> %src, i32 %index_0.i.i
  %v_1.i.i = extractelement <64 x i8> %src, i32 %index_1.i.i
  %v_2.i.i = extractelement <64 x i8> %src, i32 %index_2.i.i
  %v_3.i.i = extractelement <64 x i8> %src, i32 %index_3.i.i
  %v_4.i.i = extractelement <64 x i8> %src, i32 %index_4.i.i
  %v_5.i.i = extractelement <64 x i8> %src, i32 %index_5.i.i
  %v_6.i.i = extractelement <64 x i8> %src, i32 %index_6.i.i
  %v_7.i.i = extractelement <64 x i8> %src, i32 %index_7.i.i
  %v_8.i.i = extractelement <64 x i8> %src, i32 %index_8.i.i
  %v_9.i.i = extractelement <64 x i8> %src, i32 %index_9.i.i
  %v_10.i.i = extractelement <64 x i8> %src, i32 %index_10.i.i
  %v_11.i.i = extractelement <64 x i8> %src, i32 %index_11.i.i
  %v_12.i.i = extractelement <64 x i8> %src, i32 %index_12.i.i
  %v_13.i.i = extractelement <64 x i8> %src, i32 %index_13.i.i
  %v_14.i.i = extractelement <64 x i8> %src, i32 %index_14.i.i
  %v_15.i.i = extractelement <64 x i8> %src, i32 %index_15.i.i
  %v_16.i.i = extractelement <64 x i8> %src, i32 %index_16.i.i
  %v_17.i.i = extractelement <64 x i8> %src, i32 %index_17.i.i
  %v_18.i.i = extractelement <64 x i8> %src, i32 %index_18.i.i
  %v_19.i.i = extractelement <64 x i8> %src, i32 %index_19.i.i
  %v_20.i.i = extractelement <64 x i8> %src, i32 %index_20.i.i
  %v_21.i.i = extractelement <64 x i8> %src, i32 %index_21.i.i
  %v_22.i.i = extractelement <64 x i8> %src, i32 %index_22.i.i
  %v_23.i.i = extractelement <64 x i8> %src, i32 %index_23.i.i
  %v_24.i.i = extractelement <64 x i8> %src, i32 %index_24.i.i
  %v_25.i.i = extractelement <64 x i8> %src, i32 %index_25.i.i
  %v_26.i.i = extractelement <64 x i8> %src, i32 %index_26.i.i
  %v_27.i.i = extractelement <64 x i8> %src, i32 %index_27.i.i
  %v_28.i.i = extractelement <64 x i8> %src, i32 %index_28.i.i
  %v_29.i.i = extractelement <64 x i8> %src, i32 %index_29.i.i
  %v_30.i.i = extractelement <64 x i8> %src, i32 %index_30.i.i
  %v_31.i.i = extractelement <64 x i8> %src, i32 %index_31.i.i
  %v_32.i.i = extractelement <64 x i8> %src, i32 %index_32.i.i
  %v_33.i.i = extractelement <64 x i8> %src, i32 %index_33.i.i
  %v_34.i.i = extractelement <64 x i8> %src, i32 %index_34.i.i
  %v_35.i.i = extractelement <64 x i8> %src, i32 %index_35.i.i
  %v_36.i.i = extractelement <64 x i8> %src, i32 %index_36.i.i
  %v_37.i.i = extractelement <64 x i8> %src, i32 %index_37.i.i
  %v_38.i.i = extractelement <64 x i8> %src, i32 %index_38.i.i
  %v_39.i.i = extractelement <64 x i8> %src, i32 %index_39.i.i
  %v_40.i.i = extractelement <64 x i8> %src, i32 %index_40.i.i
  %v_41.i.i = extractelement <64 x i8> %src, i32 %index_41.i.i
  %v_42.i.i = extractelement <64 x i8> %src, i32 %index_42.i.i
  %v_43.i.i = extractelement <64 x i8> %src, i32 %index_43.i.i
  %v_44.i.i = extractelement <64 x i8> %src, i32 %index_44.i.i
  %v_45.i.i = extractelement <64 x i8> %src, i32 %index_45.i.i
  %v_46.i.i = extractelement <64 x i8> %src, i32 %index_46.i.i
  %v_47.i.i = extractelement <64 x i8> %src, i32 %index_47.i.i
  %v_48.i.i = extractelement <64 x i8> %src, i32 %index_48.i.i
  %v_49.i.i = extractelement <64 x i8> %src, i32 %index_49.i.i
  %v_50.i.i = extractelement <64 x i8> %src, i32 %index_50.i.i
  %v_51.i.i = extractelement <64 x i8> %src, i32 %index_51.i.i
  %v_52.i.i = extractelement <64 x i8> %src, i32 %index_52.i.i
  %v_53.i.i = extractelement <64 x i8> %src, i32 %index_53.i.i
  %v_54.i.i = extractelement <64 x i8> %src, i32 %index_54.i.i
  %v_55.i.i = extractelement <64 x i8> %src, i32 %index_55.i.i
  %v_56.i.i = extractelement <64 x i8> %src, i32 %index_56.i.i
  %v_57.i.i = extractelement <64 x i8> %src, i32 %index_57.i.i
  %v_58.i.i = extractelement <64 x i8> %src, i32 %b
  %v_59.i.i = extractelement <64 x i8> %src, i32 %index_59.i.i
  %v_60.i.i = extractelement <64 x i8> %src, i32 %index_60.i.i
  %v_61.i.i = extractelement <64 x i8> %src, i32 %index_61.i.i
  %v_62.i.i = extractelement <64 x i8> %src, i32 %index_62.i.i
  %v_63.i.i = extractelement <64 x i8> %src, i32 %index_63.i.i
  %dst_0.i.i = insertelement <64 x i8> undef, i8 %v_0.i.i, i32 0
  %dst_1.i.i = insertelement <64 x i8> %dst_0.i.i, i8 %v_1.i.i, i32 1
  %dst_2.i.i = insertelement <64 x i8> %dst_1.i.i, i8 %v_2.i.i, i32 2
  %dst_3.i.i = insertelement <64 x i8> %dst_2.i.i, i8 %v_3.i.i, i32 3
  %dst_4.i.i = insertelement <64 x i8> %dst_3.i.i, i8 %v_4.i.i, i32 4
  %dst_5.i.i = insertelement <64 x i8> %dst_4.i.i, i8 %v_5.i.i, i32 5
  %dst_6.i.i = insertelement <64 x i8> %dst_5.i.i, i8 %v_6.i.i, i32 6
  %dst_7.i.i = insertelement <64 x i8> %dst_6.i.i, i8 %v_7.i.i, i32 7
  %dst_8.i.i = insertelement <64 x i8> %dst_7.i.i, i8 %v_8.i.i, i32 8
  %dst_9.i.i = insertelement <64 x i8> %dst_8.i.i, i8 %v_9.i.i, i32 9
  %dst_10.i.i = insertelement <64 x i8> %dst_9.i.i, i8 %v_10.i.i, i32 10
  %dst_11.i.i = insertelement <64 x i8> %dst_10.i.i, i8 %v_11.i.i, i32 11
  %dst_12.i.i = insertelement <64 x i8> %dst_11.i.i, i8 %v_12.i.i, i32 12
  %dst_13.i.i = insertelement <64 x i8> %dst_12.i.i, i8 %v_13.i.i, i32 13
  %dst_14.i.i = insertelement <64 x i8> %dst_13.i.i, i8 %v_14.i.i, i32 14
  %dst_15.i.i = insertelement <64 x i8> %dst_14.i.i, i8 %v_15.i.i, i32 15
  %dst_16.i.i = insertelement <64 x i8> %dst_15.i.i, i8 %v_16.i.i, i32 16
  %dst_17.i.i = insertelement <64 x i8> %dst_16.i.i, i8 %v_17.i.i, i32 17
  %dst_18.i.i = insertelement <64 x i8> %dst_17.i.i, i8 %v_18.i.i, i32 18
  %dst_19.i.i = insertelement <64 x i8> %dst_18.i.i, i8 %v_19.i.i, i32 19
  %dst_20.i.i = insertelement <64 x i8> %dst_19.i.i, i8 %v_20.i.i, i32 20
  %dst_21.i.i = insertelement <64 x i8> %dst_20.i.i, i8 %v_21.i.i, i32 21
  %dst_22.i.i = insertelement <64 x i8> %dst_21.i.i, i8 %v_22.i.i, i32 22
  %dst_23.i.i = insertelement <64 x i8> %dst_22.i.i, i8 %v_23.i.i, i32 23
  %dst_24.i.i = insertelement <64 x i8> %dst_23.i.i, i8 %v_24.i.i, i32 24
  %dst_25.i.i = insertelement <64 x i8> %dst_24.i.i, i8 %v_25.i.i, i32 25
  %dst_26.i.i = insertelement <64 x i8> %dst_25.i.i, i8 %v_26.i.i, i32 26
  %dst_27.i.i = insertelement <64 x i8> %dst_26.i.i, i8 %v_27.i.i, i32 27
  %dst_28.i.i = insertelement <64 x i8> %dst_27.i.i, i8 %v_28.i.i, i32 28
  %dst_29.i.i = insertelement <64 x i8> %dst_28.i.i, i8 %v_29.i.i, i32 29
  %dst_30.i.i = insertelement <64 x i8> %dst_29.i.i, i8 %v_30.i.i, i32 30
  %dst_31.i.i = insertelement <64 x i8> %dst_30.i.i, i8 %v_31.i.i, i32 31
  %dst_32.i.i = insertelement <64 x i8> %dst_31.i.i, i8 %v_32.i.i, i32 32
  %dst_33.i.i = insertelement <64 x i8> %dst_32.i.i, i8 %v_33.i.i, i32 33
  %dst_34.i.i = insertelement <64 x i8> %dst_33.i.i, i8 %v_34.i.i, i32 34
  %dst_35.i.i = insertelement <64 x i8> %dst_34.i.i, i8 %v_35.i.i, i32 35
  %dst_36.i.i = insertelement <64 x i8> %dst_35.i.i, i8 %v_36.i.i, i32 36
  %dst_37.i.i = insertelement <64 x i8> %dst_36.i.i, i8 %v_37.i.i, i32 37
  %dst_38.i.i = insertelement <64 x i8> %dst_37.i.i, i8 %v_38.i.i, i32 38
  %dst_39.i.i = insertelement <64 x i8> %dst_38.i.i, i8 %v_39.i.i, i32 39
  %dst_40.i.i = insertelement <64 x i8> %dst_39.i.i, i8 %v_40.i.i, i32 40
  %dst_41.i.i = insertelement <64 x i8> %dst_40.i.i, i8 %v_41.i.i, i32 41
  %dst_42.i.i = insertelement <64 x i8> %dst_41.i.i, i8 %v_42.i.i, i32 42
  %dst_43.i.i = insertelement <64 x i8> %dst_42.i.i, i8 %v_43.i.i, i32 43
  %dst_44.i.i = insertelement <64 x i8> %dst_43.i.i, i8 %v_44.i.i, i32 44
  %dst_45.i.i = insertelement <64 x i8> %dst_44.i.i, i8 %v_45.i.i, i32 45
  %dst_46.i.i = insertelement <64 x i8> %dst_45.i.i, i8 %v_46.i.i, i32 46
  %dst_47.i.i = insertelement <64 x i8> %dst_46.i.i, i8 %v_47.i.i, i32 47
  %dst_48.i.i = insertelement <64 x i8> %dst_47.i.i, i8 %v_48.i.i, i32 48
  %dst_49.i.i = insertelement <64 x i8> %dst_48.i.i, i8 %v_49.i.i, i32 49
  %dst_50.i.i = insertelement <64 x i8> %dst_49.i.i, i8 %v_50.i.i, i32 50
  %dst_51.i.i = insertelement <64 x i8> %dst_50.i.i, i8 %v_51.i.i, i32 51
  %dst_52.i.i = insertelement <64 x i8> %dst_51.i.i, i8 %v_52.i.i, i32 52
  %dst_53.i.i = insertelement <64 x i8> %dst_52.i.i, i8 %v_53.i.i, i32 53
  %dst_54.i.i = insertelement <64 x i8> %dst_53.i.i, i8 %v_54.i.i, i32 54
  %dst_55.i.i = insertelement <64 x i8> %dst_54.i.i, i8 %v_55.i.i, i32 55
  %dst_56.i.i = insertelement <64 x i8> %dst_55.i.i, i8 %v_56.i.i, i32 56
  %dst_57.i.i = insertelement <64 x i8> %dst_56.i.i, i8 %v_57.i.i, i32 57
  %dst_58.i.i = insertelement <64 x i8> %dst_57.i.i, i8 %v_58.i.i, i32 58
  %dst_59.i.i = insertelement <64 x i8> %dst_58.i.i, i8 %v_59.i.i, i32 59
  %dst_60.i.i = insertelement <64 x i8> %dst_59.i.i, i8 %v_60.i.i, i32 60
  %dst_61.i.i = insertelement <64 x i8> %dst_60.i.i, i8 %v_61.i.i, i32 61
  %dst_62.i.i = insertelement <64 x i8> %dst_61.i.i, i8 %v_62.i.i, i32 62
  %dst_63.i.i = insertelement <64 x i8> %dst_62.i.i, i8 %v_63.i.i, i32 63
  %shuf_load_to_float = sitofp <64 x i8> %dst_63.i.i to <64 x float>
  store <64 x float> %shuf_load_to_float, <64 x float>* %dst
  ret void
}