llvm/llvm/test/CodeGen/X86/vector-interleaved-load-i8-stride-8.ll

; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
; RUN: llc < %s -mtriple=x86_64-- -mattr=+sse2 | FileCheck %s --check-prefixes=SSE
; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx  | FileCheck %s --check-prefixes=AVX
; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx2 | FileCheck %s --check-prefixes=AVX2
; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx2,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX2-FP
; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx2,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX2-FCP
; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl | FileCheck %s --check-prefixes=AVX512
; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX512-FCP
; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+avx512dq | FileCheck %s --check-prefixes=AVX512DQ
; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+avx512dq,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX512DQ-FCP
; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+avx512bw | FileCheck %s --check-prefixes=AVX512BW
; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+avx512bw,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX512BW-FCP
; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+avx512dq,+avx512bw | FileCheck %s --check-prefixes=AVX512DQ-BW
; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+avx512dq,+avx512bw,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX512DQ-BW-FCP

; These patterns are produced by LoopVectorizer for interleaved loads.

define void @load_i8_stride8_vf2(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr %out.vec2, ptr %out.vec3, ptr %out.vec4, ptr %out.vec5, ptr %out.vec6, ptr %out.vec7) nounwind {
; SSE-LABEL: load_i8_stride8_vf2:
; SSE:       # %bb.0:
; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; SSE-NEXT:    movdqa (%rdi), %xmm1
; SSE-NEXT:    movdqa {{.*#+}} xmm0 = [255,0,0,0,255,0,0,0]
; SSE-NEXT:    pand %xmm1, %xmm0
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    pxor %xmm2, %xmm2
; SSE-NEXT:    movdqa %xmm1, %xmm3
; SSE-NEXT:    punpckhbw {{.*#+}} xmm3 = xmm3[8],xmm2[8],xmm3[9],xmm2[9],xmm3[10],xmm2[10],xmm3[11],xmm2[11],xmm3[12],xmm2[12],xmm3[13],xmm2[13],xmm3[14],xmm2[14],xmm3[15],xmm2[15]
; SSE-NEXT:    movdqa %xmm1, %xmm4
; SSE-NEXT:    punpcklbw {{.*#+}} xmm4 = xmm4[0],xmm2[0],xmm4[1],xmm2[1],xmm4[2],xmm2[2],xmm4[3],xmm2[3],xmm4[4],xmm2[4],xmm4[5],xmm2[5],xmm4[6],xmm2[6],xmm4[7],xmm2[7]
; SSE-NEXT:    movdqa %xmm4, %xmm2
; SSE-NEXT:    punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm2[1,1,1,1]
; SSE-NEXT:    packuswb %xmm5, %xmm5
; SSE-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
; SSE-NEXT:    pshufd {{.*#+}} xmm6 = xmm1[0,2,2,3]
; SSE-NEXT:    pshuflw {{.*#+}} xmm6 = xmm6[1,3,2,3,4,5,6,7]
; SSE-NEXT:    packuswb %xmm6, %xmm6
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[3,3,3,3]
; SSE-NEXT:    packuswb %xmm2, %xmm2
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[3,1,2,3]
; SSE-NEXT:    pshuflw {{.*#+}} xmm7 = xmm1[2,0,2,3,4,5,6,7]
; SSE-NEXT:    packuswb %xmm7, %xmm7
; SSE-NEXT:    punpckhwd {{.*#+}} xmm4 = xmm4[4],xmm3[4],xmm4[5],xmm3[5],xmm4[6],xmm3[6],xmm4[7],xmm3[7]
; SSE-NEXT:    pshufd {{.*#+}} xmm3 = xmm4[1,1,1,1]
; SSE-NEXT:    packuswb %xmm3, %xmm3
; SSE-NEXT:    pshuflw {{.*#+}} xmm1 = xmm1[3,1,2,3,4,5,6,7]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[3,3,3,3]
; SSE-NEXT:    packuswb %xmm4, %xmm4
; SSE-NEXT:    movd %xmm0, %edi
; SSE-NEXT:    movw %di, (%rsi)
; SSE-NEXT:    movd %xmm5, %esi
; SSE-NEXT:    movw %si, (%rdx)
; SSE-NEXT:    movd %xmm6, %edx
; SSE-NEXT:    movw %dx, (%rcx)
; SSE-NEXT:    movd %xmm2, %ecx
; SSE-NEXT:    movw %cx, (%r8)
; SSE-NEXT:    movd %xmm7, %ecx
; SSE-NEXT:    movw %cx, (%r9)
; SSE-NEXT:    movd %xmm3, %ecx
; SSE-NEXT:    movw %cx, (%r11)
; SSE-NEXT:    movd %xmm1, %ecx
; SSE-NEXT:    movw %cx, (%r10)
; SSE-NEXT:    movd %xmm4, %ecx
; SSE-NEXT:    movw %cx, (%rax)
; SSE-NEXT:    retq
;
; AVX-LABEL: load_i8_stride8_vf2:
; AVX:       # %bb.0:
; AVX-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX-NEXT:    vmovdqa (%rdi), %xmm0
; AVX-NEXT:    vpshufb {{.*#+}} xmm1 = xmm0[0,8,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX-NEXT:    vpshufb {{.*#+}} xmm2 = xmm0[1,9,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX-NEXT:    vpshufb {{.*#+}} xmm3 = xmm0[2,10,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX-NEXT:    vpshufb {{.*#+}} xmm4 = xmm0[3,11,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX-NEXT:    vpshufb {{.*#+}} xmm5 = xmm0[4,12,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX-NEXT:    vpshufb {{.*#+}} xmm6 = xmm0[5,13,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX-NEXT:    vpshufb {{.*#+}} xmm7 = xmm0[6,14,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX-NEXT:    vpshufb {{.*#+}} xmm0 = xmm0[7,15,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX-NEXT:    vpextrw $0, %xmm1, (%rsi)
; AVX-NEXT:    vpextrw $0, %xmm2, (%rdx)
; AVX-NEXT:    vpextrw $0, %xmm3, (%rcx)
; AVX-NEXT:    vpextrw $0, %xmm4, (%r8)
; AVX-NEXT:    vpextrw $0, %xmm5, (%r9)
; AVX-NEXT:    vpextrw $0, %xmm6, (%r11)
; AVX-NEXT:    vpextrw $0, %xmm7, (%r10)
; AVX-NEXT:    vpextrw $0, %xmm0, (%rax)
; AVX-NEXT:    retq
;
; AVX2-LABEL: load_i8_stride8_vf2:
; AVX2:       # %bb.0:
; AVX2-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX2-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX2-NEXT:    vmovdqa (%rdi), %xmm0
; AVX2-NEXT:    vpshufb {{.*#+}} xmm1 = xmm0[0,8,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX2-NEXT:    vpshufb {{.*#+}} xmm2 = xmm0[1,9,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX2-NEXT:    vpshufb {{.*#+}} xmm3 = xmm0[2,10,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX2-NEXT:    vpshufb {{.*#+}} xmm4 = xmm0[3,11,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX2-NEXT:    vpshufb {{.*#+}} xmm5 = xmm0[4,12,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX2-NEXT:    vpshufb {{.*#+}} xmm6 = xmm0[5,13,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX2-NEXT:    vpshufb {{.*#+}} xmm7 = xmm0[6,14,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX2-NEXT:    vpshufb {{.*#+}} xmm0 = xmm0[7,15,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX2-NEXT:    vpextrw $0, %xmm1, (%rsi)
; AVX2-NEXT:    vpextrw $0, %xmm2, (%rdx)
; AVX2-NEXT:    vpextrw $0, %xmm3, (%rcx)
; AVX2-NEXT:    vpextrw $0, %xmm4, (%r8)
; AVX2-NEXT:    vpextrw $0, %xmm5, (%r9)
; AVX2-NEXT:    vpextrw $0, %xmm6, (%r11)
; AVX2-NEXT:    vpextrw $0, %xmm7, (%r10)
; AVX2-NEXT:    vpextrw $0, %xmm0, (%rax)
; AVX2-NEXT:    retq
;
; AVX2-FP-LABEL: load_i8_stride8_vf2:
; AVX2-FP:       # %bb.0:
; AVX2-FP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FP-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX2-FP-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX2-FP-NEXT:    vmovdqa (%rdi), %xmm0
; AVX2-FP-NEXT:    vpshufb {{.*#+}} xmm1 = xmm0[0,8,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX2-FP-NEXT:    vpshufb {{.*#+}} xmm2 = xmm0[1,9,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX2-FP-NEXT:    vpshufb {{.*#+}} xmm3 = xmm0[2,10,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX2-FP-NEXT:    vpshufb {{.*#+}} xmm4 = xmm0[3,11,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX2-FP-NEXT:    vpshufb {{.*#+}} xmm5 = xmm0[4,12,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX2-FP-NEXT:    vpshufb {{.*#+}} xmm6 = xmm0[5,13,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX2-FP-NEXT:    vpshufb {{.*#+}} xmm7 = xmm0[6,14,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX2-FP-NEXT:    vpshufb {{.*#+}} xmm0 = xmm0[7,15,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX2-FP-NEXT:    vpextrw $0, %xmm1, (%rsi)
; AVX2-FP-NEXT:    vpextrw $0, %xmm2, (%rdx)
; AVX2-FP-NEXT:    vpextrw $0, %xmm3, (%rcx)
; AVX2-FP-NEXT:    vpextrw $0, %xmm4, (%r8)
; AVX2-FP-NEXT:    vpextrw $0, %xmm5, (%r9)
; AVX2-FP-NEXT:    vpextrw $0, %xmm6, (%r11)
; AVX2-FP-NEXT:    vpextrw $0, %xmm7, (%r10)
; AVX2-FP-NEXT:    vpextrw $0, %xmm0, (%rax)
; AVX2-FP-NEXT:    retq
;
; AVX2-FCP-LABEL: load_i8_stride8_vf2:
; AVX2-FCP:       # %bb.0:
; AVX2-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX2-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX2-FCP-NEXT:    vmovdqa (%rdi), %xmm0
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} xmm1 = xmm0[0,8,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} xmm2 = xmm0[1,9,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} xmm3 = xmm0[2,10,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} xmm4 = xmm0[3,11,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} xmm5 = xmm0[4,12,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} xmm6 = xmm0[5,13,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} xmm7 = xmm0[6,14,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} xmm0 = xmm0[7,15,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX2-FCP-NEXT:    vpextrw $0, %xmm1, (%rsi)
; AVX2-FCP-NEXT:    vpextrw $0, %xmm2, (%rdx)
; AVX2-FCP-NEXT:    vpextrw $0, %xmm3, (%rcx)
; AVX2-FCP-NEXT:    vpextrw $0, %xmm4, (%r8)
; AVX2-FCP-NEXT:    vpextrw $0, %xmm5, (%r9)
; AVX2-FCP-NEXT:    vpextrw $0, %xmm6, (%r11)
; AVX2-FCP-NEXT:    vpextrw $0, %xmm7, (%r10)
; AVX2-FCP-NEXT:    vpextrw $0, %xmm0, (%rax)
; AVX2-FCP-NEXT:    retq
;
; AVX512-LABEL: load_i8_stride8_vf2:
; AVX512:       # %bb.0:
; AVX512-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX512-NEXT:    vmovdqa (%rdi), %xmm0
; AVX512-NEXT:    vpshufb {{.*#+}} xmm1 = xmm0[1,9,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512-NEXT:    vpshufb {{.*#+}} xmm2 = xmm0[2,10,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512-NEXT:    vpshufb {{.*#+}} xmm3 = xmm0[3,11,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512-NEXT:    vpshufb {{.*#+}} xmm4 = xmm0[4,12,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512-NEXT:    vpshufb {{.*#+}} xmm5 = xmm0[5,13,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512-NEXT:    vpshufb {{.*#+}} xmm6 = xmm0[6,14,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512-NEXT:    vpshufb {{.*#+}} xmm7 = xmm0[7,15,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512-NEXT:    vpmovqb %xmm0, (%rsi)
; AVX512-NEXT:    vpextrw $0, %xmm1, (%rdx)
; AVX512-NEXT:    vpextrw $0, %xmm2, (%rcx)
; AVX512-NEXT:    vpextrw $0, %xmm3, (%r8)
; AVX512-NEXT:    vpextrw $0, %xmm4, (%r9)
; AVX512-NEXT:    vpextrw $0, %xmm5, (%r11)
; AVX512-NEXT:    vpextrw $0, %xmm6, (%r10)
; AVX512-NEXT:    vpextrw $0, %xmm7, (%rax)
; AVX512-NEXT:    retq
;
; AVX512-FCP-LABEL: load_i8_stride8_vf2:
; AVX512-FCP:       # %bb.0:
; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX512-FCP-NEXT:    vmovdqa (%rdi), %xmm0
; AVX512-FCP-NEXT:    vpshufb {{.*#+}} xmm1 = xmm0[1,9,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512-FCP-NEXT:    vpshufb {{.*#+}} xmm2 = xmm0[2,10,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512-FCP-NEXT:    vpshufb {{.*#+}} xmm3 = xmm0[3,11,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512-FCP-NEXT:    vpshufb {{.*#+}} xmm4 = xmm0[4,12,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512-FCP-NEXT:    vpshufb {{.*#+}} xmm5 = xmm0[5,13,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512-FCP-NEXT:    vpshufb {{.*#+}} xmm6 = xmm0[6,14,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512-FCP-NEXT:    vpshufb {{.*#+}} xmm7 = xmm0[7,15,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512-FCP-NEXT:    vpmovqb %xmm0, (%rsi)
; AVX512-FCP-NEXT:    vpextrw $0, %xmm1, (%rdx)
; AVX512-FCP-NEXT:    vpextrw $0, %xmm2, (%rcx)
; AVX512-FCP-NEXT:    vpextrw $0, %xmm3, (%r8)
; AVX512-FCP-NEXT:    vpextrw $0, %xmm4, (%r9)
; AVX512-FCP-NEXT:    vpextrw $0, %xmm5, (%r11)
; AVX512-FCP-NEXT:    vpextrw $0, %xmm6, (%r10)
; AVX512-FCP-NEXT:    vpextrw $0, %xmm7, (%rax)
; AVX512-FCP-NEXT:    retq
;
; AVX512DQ-LABEL: load_i8_stride8_vf2:
; AVX512DQ:       # %bb.0:
; AVX512DQ-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512DQ-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX512DQ-NEXT:    vmovdqa (%rdi), %xmm0
; AVX512DQ-NEXT:    vpshufb {{.*#+}} xmm1 = xmm0[1,9,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512DQ-NEXT:    vpshufb {{.*#+}} xmm2 = xmm0[2,10,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512DQ-NEXT:    vpshufb {{.*#+}} xmm3 = xmm0[3,11,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512DQ-NEXT:    vpshufb {{.*#+}} xmm4 = xmm0[4,12,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512DQ-NEXT:    vpshufb {{.*#+}} xmm5 = xmm0[5,13,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512DQ-NEXT:    vpshufb {{.*#+}} xmm6 = xmm0[6,14,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512DQ-NEXT:    vpshufb {{.*#+}} xmm7 = xmm0[7,15,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512DQ-NEXT:    vpmovqb %xmm0, (%rsi)
; AVX512DQ-NEXT:    vpextrw $0, %xmm1, (%rdx)
; AVX512DQ-NEXT:    vpextrw $0, %xmm2, (%rcx)
; AVX512DQ-NEXT:    vpextrw $0, %xmm3, (%r8)
; AVX512DQ-NEXT:    vpextrw $0, %xmm4, (%r9)
; AVX512DQ-NEXT:    vpextrw $0, %xmm5, (%r11)
; AVX512DQ-NEXT:    vpextrw $0, %xmm6, (%r10)
; AVX512DQ-NEXT:    vpextrw $0, %xmm7, (%rax)
; AVX512DQ-NEXT:    retq
;
; AVX512DQ-FCP-LABEL: load_i8_stride8_vf2:
; AVX512DQ-FCP:       # %bb.0:
; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX512DQ-FCP-NEXT:    vmovdqa (%rdi), %xmm0
; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} xmm1 = xmm0[1,9,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} xmm2 = xmm0[2,10,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} xmm3 = xmm0[3,11,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} xmm4 = xmm0[4,12,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} xmm5 = xmm0[5,13,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} xmm6 = xmm0[6,14,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} xmm7 = xmm0[7,15,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512DQ-FCP-NEXT:    vpmovqb %xmm0, (%rsi)
; AVX512DQ-FCP-NEXT:    vpextrw $0, %xmm1, (%rdx)
; AVX512DQ-FCP-NEXT:    vpextrw $0, %xmm2, (%rcx)
; AVX512DQ-FCP-NEXT:    vpextrw $0, %xmm3, (%r8)
; AVX512DQ-FCP-NEXT:    vpextrw $0, %xmm4, (%r9)
; AVX512DQ-FCP-NEXT:    vpextrw $0, %xmm5, (%r11)
; AVX512DQ-FCP-NEXT:    vpextrw $0, %xmm6, (%r10)
; AVX512DQ-FCP-NEXT:    vpextrw $0, %xmm7, (%rax)
; AVX512DQ-FCP-NEXT:    retq
;
; AVX512BW-LABEL: load_i8_stride8_vf2:
; AVX512BW:       # %bb.0:
; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX512BW-NEXT:    vmovdqa (%rdi), %xmm0
; AVX512BW-NEXT:    vpshufb {{.*#+}} xmm1 = xmm0[1,9,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512BW-NEXT:    vpshufb {{.*#+}} xmm2 = xmm0[2,10,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512BW-NEXT:    vpshufb {{.*#+}} xmm3 = xmm0[3,11,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512BW-NEXT:    vpshufb {{.*#+}} xmm4 = xmm0[4,12,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512BW-NEXT:    vpshufb {{.*#+}} xmm5 = xmm0[5,13,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512BW-NEXT:    vpshufb {{.*#+}} xmm6 = xmm0[6,14,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512BW-NEXT:    vpshufb {{.*#+}} xmm7 = xmm0[7,15,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512BW-NEXT:    vpmovqb %xmm0, (%rsi)
; AVX512BW-NEXT:    vpextrw $0, %xmm1, (%rdx)
; AVX512BW-NEXT:    vpextrw $0, %xmm2, (%rcx)
; AVX512BW-NEXT:    vpextrw $0, %xmm3, (%r8)
; AVX512BW-NEXT:    vpextrw $0, %xmm4, (%r9)
; AVX512BW-NEXT:    vpextrw $0, %xmm5, (%r11)
; AVX512BW-NEXT:    vpextrw $0, %xmm6, (%r10)
; AVX512BW-NEXT:    vpextrw $0, %xmm7, (%rax)
; AVX512BW-NEXT:    retq
;
; AVX512BW-FCP-LABEL: load_i8_stride8_vf2:
; AVX512BW-FCP:       # %bb.0:
; AVX512BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX512BW-FCP-NEXT:    vmovdqa (%rdi), %xmm0
; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} xmm1 = xmm0[1,9,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} xmm2 = xmm0[2,10,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} xmm3 = xmm0[3,11,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} xmm4 = xmm0[4,12,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} xmm5 = xmm0[5,13,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} xmm6 = xmm0[6,14,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512BW-FCP-NEXT:    vpshufb {{.*#+}} xmm7 = xmm0[7,15,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512BW-FCP-NEXT:    vpmovqb %xmm0, (%rsi)
; AVX512BW-FCP-NEXT:    vpextrw $0, %xmm1, (%rdx)
; AVX512BW-FCP-NEXT:    vpextrw $0, %xmm2, (%rcx)
; AVX512BW-FCP-NEXT:    vpextrw $0, %xmm3, (%r8)
; AVX512BW-FCP-NEXT:    vpextrw $0, %xmm4, (%r9)
; AVX512BW-FCP-NEXT:    vpextrw $0, %xmm5, (%r11)
; AVX512BW-FCP-NEXT:    vpextrw $0, %xmm6, (%r10)
; AVX512BW-FCP-NEXT:    vpextrw $0, %xmm7, (%rax)
; AVX512BW-FCP-NEXT:    retq
;
; AVX512DQ-BW-LABEL: load_i8_stride8_vf2:
; AVX512DQ-BW:       # %bb.0:
; AVX512DQ-BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-BW-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512DQ-BW-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX512DQ-BW-NEXT:    vmovdqa (%rdi), %xmm0
; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} xmm1 = xmm0[1,9,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} xmm2 = xmm0[2,10,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} xmm3 = xmm0[3,11,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} xmm4 = xmm0[4,12,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} xmm5 = xmm0[5,13,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} xmm6 = xmm0[6,14,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512DQ-BW-NEXT:    vpshufb {{.*#+}} xmm7 = xmm0[7,15,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512DQ-BW-NEXT:    vpmovqb %xmm0, (%rsi)
; AVX512DQ-BW-NEXT:    vpextrw $0, %xmm1, (%rdx)
; AVX512DQ-BW-NEXT:    vpextrw $0, %xmm2, (%rcx)
; AVX512DQ-BW-NEXT:    vpextrw $0, %xmm3, (%r8)
; AVX512DQ-BW-NEXT:    vpextrw $0, %xmm4, (%r9)
; AVX512DQ-BW-NEXT:    vpextrw $0, %xmm5, (%r11)
; AVX512DQ-BW-NEXT:    vpextrw $0, %xmm6, (%r10)
; AVX512DQ-BW-NEXT:    vpextrw $0, %xmm7, (%rax)
; AVX512DQ-BW-NEXT:    retq
;
; AVX512DQ-BW-FCP-LABEL: load_i8_stride8_vf2:
; AVX512DQ-BW-FCP:       # %bb.0:
; AVX512DQ-BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512DQ-BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%rdi), %xmm0
; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} xmm1 = xmm0[1,9,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} xmm2 = xmm0[2,10,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} xmm3 = xmm0[3,11,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} xmm4 = xmm0[4,12,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} xmm5 = xmm0[5,13,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} xmm6 = xmm0[6,14,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512DQ-BW-FCP-NEXT:    vpshufb {{.*#+}} xmm7 = xmm0[7,15,u,u,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512DQ-BW-FCP-NEXT:    vpmovqb %xmm0, (%rsi)
; AVX512DQ-BW-FCP-NEXT:    vpextrw $0, %xmm1, (%rdx)
; AVX512DQ-BW-FCP-NEXT:    vpextrw $0, %xmm2, (%rcx)
; AVX512DQ-BW-FCP-NEXT:    vpextrw $0, %xmm3, (%r8)
; AVX512DQ-BW-FCP-NEXT:    vpextrw $0, %xmm4, (%r9)
; AVX512DQ-BW-FCP-NEXT:    vpextrw $0, %xmm5, (%r11)
; AVX512DQ-BW-FCP-NEXT:    vpextrw $0, %xmm6, (%r10)
; AVX512DQ-BW-FCP-NEXT:    vpextrw $0, %xmm7, (%rax)
; AVX512DQ-BW-FCP-NEXT:    retq
  %wide.vec = load <16 x i8>, ptr %in.vec, align 64
  %strided.vec0 = shufflevector <16 x i8> %wide.vec, <16 x i8> poison, <2 x i32> <i32 0, i32 8>
  %strided.vec1 = shufflevector <16 x i8> %wide.vec, <16 x i8> poison, <2 x i32> <i32 1, i32 9>
  %strided.vec2 = shufflevector <16 x i8> %wide.vec, <16 x i8> poison, <2 x i32> <i32 2, i32 10>
  %strided.vec3 = shufflevector <16 x i8> %wide.vec, <16 x i8> poison, <2 x i32> <i32 3, i32 11>
  %strided.vec4 = shufflevector <16 x i8> %wide.vec, <16 x i8> poison, <2 x i32> <i32 4, i32 12>
  %strided.vec5 = shufflevector <16 x i8> %wide.vec, <16 x i8> poison, <2 x i32> <i32 5, i32 13>
  %strided.vec6 = shufflevector <16 x i8> %wide.vec, <16 x i8> poison, <2 x i32> <i32 6, i32 14>
  %strided.vec7 = shufflevector <16 x i8> %wide.vec, <16 x i8> poison, <2 x i32> <i32 7, i32 15>
  store <2 x i8> %strided.vec0, ptr %out.vec0, align 64
  store <2 x i8> %strided.vec1, ptr %out.vec1, align 64
  store <2 x i8> %strided.vec2, ptr %out.vec2, align 64
  store <2 x i8> %strided.vec3, ptr %out.vec3, align 64
  store <2 x i8> %strided.vec4, ptr %out.vec4, align 64
  store <2 x i8> %strided.vec5, ptr %out.vec5, align 64
  store <2 x i8> %strided.vec6, ptr %out.vec6, align 64
  store <2 x i8> %strided.vec7, ptr %out.vec7, align 64
  ret void
}

define void @load_i8_stride8_vf4(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr %out.vec2, ptr %out.vec3, ptr %out.vec4, ptr %out.vec5, ptr %out.vec6, ptr %out.vec7) nounwind {
; SSE-LABEL: load_i8_stride8_vf4:
; SSE:       # %bb.0:
; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; SSE-NEXT:    movdqa (%rdi), %xmm5
; SSE-NEXT:    movdqa 16(%rdi), %xmm8
; SSE-NEXT:    movdqa {{.*#+}} xmm0 = [255,0,0,0,255,0,0,0]
; SSE-NEXT:    movdqa %xmm8, %xmm1
; SSE-NEXT:    pand %xmm0, %xmm1
; SSE-NEXT:    pand %xmm5, %xmm0
; SSE-NEXT:    packuswb %xmm1, %xmm0
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    pxor %xmm3, %xmm3
; SSE-NEXT:    movdqa %xmm8, %xmm7
; SSE-NEXT:    punpckhbw {{.*#+}} xmm7 = xmm7[8],xmm3[8],xmm7[9],xmm3[9],xmm7[10],xmm3[10],xmm7[11],xmm3[11],xmm7[12],xmm3[12],xmm7[13],xmm3[13],xmm7[14],xmm3[14],xmm7[15],xmm3[15]
; SSE-NEXT:    movdqa %xmm8, %xmm2
; SSE-NEXT:    punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3],xmm2[4],xmm3[4],xmm2[5],xmm3[5],xmm2[6],xmm3[6],xmm2[7],xmm3[7]
; SSE-NEXT:    movdqa %xmm2, %xmm10
; SSE-NEXT:    punpcklwd {{.*#+}} xmm10 = xmm10[0],xmm7[0],xmm10[1],xmm7[1],xmm10[2],xmm7[2],xmm10[3],xmm7[3]
; SSE-NEXT:    packuswb %xmm10, %xmm10
; SSE-NEXT:    movdqa {{.*#+}} xmm1 = [65535,0,65535,65535,65535,65535,65535,65535]
; SSE-NEXT:    movdqa %xmm1, %xmm6
; SSE-NEXT:    pandn %xmm10, %xmm6
; SSE-NEXT:    movdqa %xmm5, %xmm9
; SSE-NEXT:    punpckhbw {{.*#+}} xmm9 = xmm9[8],xmm3[8],xmm9[9],xmm3[9],xmm9[10],xmm3[10],xmm9[11],xmm3[11],xmm9[12],xmm3[12],xmm9[13],xmm3[13],xmm9[14],xmm3[14],xmm9[15],xmm3[15]
; SSE-NEXT:    movdqa %xmm5, %xmm4
; SSE-NEXT:    punpcklbw {{.*#+}} xmm4 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3],xmm4[4],xmm3[4],xmm4[5],xmm3[5],xmm4[6],xmm3[6],xmm4[7],xmm3[7]
; SSE-NEXT:    movdqa %xmm4, %xmm11
; SSE-NEXT:    punpcklwd {{.*#+}} xmm11 = xmm11[0],xmm9[0],xmm11[1],xmm9[1],xmm11[2],xmm9[2],xmm11[3],xmm9[3]
; SSE-NEXT:    pshufd {{.*#+}} xmm3 = xmm11[1,1,1,1]
; SSE-NEXT:    packuswb %xmm3, %xmm3
; SSE-NEXT:    pand %xmm1, %xmm3
; SSE-NEXT:    por %xmm6, %xmm3
; SSE-NEXT:    movdqa {{.*#+}} xmm6 = [255,255,255,255,255,255,255,255]
; SSE-NEXT:    pand %xmm6, %xmm8
; SSE-NEXT:    pshufd {{.*#+}} xmm12 = xmm8[0,2,2,3]
; SSE-NEXT:    pshuflw {{.*#+}} xmm12 = xmm12[0,1,1,3,4,5,6,7]
; SSE-NEXT:    packuswb %xmm12, %xmm12
; SSE-NEXT:    movdqa %xmm1, %xmm13
; SSE-NEXT:    pandn %xmm12, %xmm13
; SSE-NEXT:    pand %xmm6, %xmm5
; SSE-NEXT:    pshufd {{.*#+}} xmm6 = xmm5[0,2,2,3]
; SSE-NEXT:    pshuflw {{.*#+}} xmm6 = xmm6[1,3,2,3,4,5,6,7]
; SSE-NEXT:    packuswb %xmm6, %xmm6
; SSE-NEXT:    pand %xmm1, %xmm6
; SSE-NEXT:    por %xmm13, %xmm6
; SSE-NEXT:    pshufd {{.*#+}} xmm10 = xmm10[1,1,2,3]
; SSE-NEXT:    movdqa %xmm1, %xmm12
; SSE-NEXT:    pandn %xmm10, %xmm12
; SSE-NEXT:    pshufd {{.*#+}} xmm10 = xmm11[3,3,3,3]
; SSE-NEXT:    packuswb %xmm10, %xmm10
; SSE-NEXT:    pand %xmm1, %xmm10
; SSE-NEXT:    por %xmm12, %xmm10
; SSE-NEXT:    pshufd {{.*#+}} xmm8 = xmm8[3,1,2,3]
; SSE-NEXT:    pshuflw {{.*#+}} xmm11 = xmm8[0,1,2,0,4,5,6,7]
; SSE-NEXT:    packuswb %xmm11, %xmm11
; SSE-NEXT:    movdqa %xmm1, %xmm12
; SSE-NEXT:    pandn %xmm11, %xmm12
; SSE-NEXT:    pshufd {{.*#+}} xmm11 = xmm5[3,1,2,3]
; SSE-NEXT:    pshuflw {{.*#+}} xmm5 = xmm11[2,0,2,3,4,5,6,7]
; SSE-NEXT:    packuswb %xmm5, %xmm5
; SSE-NEXT:    pand %xmm1, %xmm5
; SSE-NEXT:    por %xmm12, %xmm5
; SSE-NEXT:    punpckhwd {{.*#+}} xmm2 = xmm2[4],xmm7[4],xmm2[5],xmm7[5],xmm2[6],xmm7[6],xmm2[7],xmm7[7]
; SSE-NEXT:    packuswb %xmm2, %xmm2
; SSE-NEXT:    movdqa %xmm1, %xmm12
; SSE-NEXT:    pandn %xmm2, %xmm12
; SSE-NEXT:    punpckhwd {{.*#+}} xmm4 = xmm4[4],xmm9[4],xmm4[5],xmm9[5],xmm4[6],xmm9[6],xmm4[7],xmm9[7]
; SSE-NEXT:    pshufd {{.*#+}} xmm7 = xmm4[1,1,1,1]
; SSE-NEXT:    packuswb %xmm7, %xmm7
; SSE-NEXT:    pand %xmm1, %xmm7
; SSE-NEXT:    por %xmm12, %xmm7
; SSE-NEXT:    pshuflw {{.*#+}} xmm8 = xmm8[0,1,3,1,4,5,6,7]
; SSE-NEXT:    packuswb %xmm8, %xmm8
; SSE-NEXT:    movdqa %xmm1, %xmm9
; SSE-NEXT:    pandn %xmm8, %xmm9
; SSE-NEXT:    pshuflw {{.*#+}} xmm8 = xmm11[3,1,2,3,4,5,6,7]
; SSE-NEXT:    packuswb %xmm8, %xmm8
; SSE-NEXT:    pand %xmm1, %xmm8
; SSE-NEXT:    por %xmm9, %xmm8
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,2,3]
; SSE-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[3,3,3,3]
; SSE-NEXT:    packuswb %xmm4, %xmm4
; SSE-NEXT:    pand %xmm1, %xmm4
; SSE-NEXT:    pandn %xmm2, %xmm1
; SSE-NEXT:    por %xmm4, %xmm1
; SSE-NEXT:    movd %xmm0, (%rsi)
; SSE-NEXT:    movd %xmm3, (%rdx)
; SSE-NEXT:    movd %xmm6, (%rcx)
; SSE-NEXT:    movd %xmm10, (%r8)
; SSE-NEXT:    movd %xmm5, (%r9)
; SSE-NEXT:    movd %xmm7, (%r11)
; SSE-NEXT:    movd %xmm8, (%r10)
; SSE-NEXT:    movd %xmm1, (%rax)
; SSE-NEXT:    retq
;
; AVX-LABEL: load_i8_stride8_vf4:
; AVX:       # %bb.0:
; AVX-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX-NEXT:    vmovd {{.*#+}} xmm0 = [0,8,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vmovdqa (%rdi), %xmm1
; AVX-NEXT:    vmovdqa 16(%rdi), %xmm2
; AVX-NEXT:    vpshufb %xmm0, %xmm2, %xmm3
; AVX-NEXT:    vpshufb %xmm0, %xmm1, %xmm0
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3]
; AVX-NEXT:    vmovd {{.*#+}} xmm3 = [1,9,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vpshufb %xmm3, %xmm2, %xmm4
; AVX-NEXT:    vpshufb %xmm3, %xmm1, %xmm3
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
; AVX-NEXT:    vmovd {{.*#+}} xmm4 = [2,10,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vpshufb %xmm4, %xmm2, %xmm5
; AVX-NEXT:    vpshufb %xmm4, %xmm1, %xmm4
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3]
; AVX-NEXT:    vmovd {{.*#+}} xmm5 = [3,11,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vpshufb %xmm5, %xmm2, %xmm6
; AVX-NEXT:    vpshufb %xmm5, %xmm1, %xmm5
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3]
; AVX-NEXT:    vmovd {{.*#+}} xmm6 = [4,12,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vpshufb %xmm6, %xmm2, %xmm7
; AVX-NEXT:    vpshufb %xmm6, %xmm1, %xmm6
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm6[0],xmm7[0],xmm6[1],xmm7[1],xmm6[2],xmm7[2],xmm6[3],xmm7[3]
; AVX-NEXT:    vmovd {{.*#+}} xmm7 = [5,13,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vpshufb %xmm7, %xmm2, %xmm8
; AVX-NEXT:    vpshufb %xmm7, %xmm1, %xmm7
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm7 = xmm7[0],xmm8[0],xmm7[1],xmm8[1],xmm7[2],xmm8[2],xmm7[3],xmm8[3]
; AVX-NEXT:    vmovd {{.*#+}} xmm8 = [6,14,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vpshufb %xmm8, %xmm2, %xmm9
; AVX-NEXT:    vpshufb %xmm8, %xmm1, %xmm8
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm9[0],xmm8[1],xmm9[1],xmm8[2],xmm9[2],xmm8[3],xmm9[3]
; AVX-NEXT:    vmovd {{.*#+}} xmm9 = [7,15,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vpshufb %xmm9, %xmm2, %xmm2
; AVX-NEXT:    vpshufb %xmm9, %xmm1, %xmm1
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
; AVX-NEXT:    vmovd %xmm0, (%rsi)
; AVX-NEXT:    vmovd %xmm3, (%rdx)
; AVX-NEXT:    vmovd %xmm4, (%rcx)
; AVX-NEXT:    vmovd %xmm5, (%r8)
; AVX-NEXT:    vmovd %xmm6, (%r9)
; AVX-NEXT:    vmovd %xmm7, (%r11)
; AVX-NEXT:    vmovd %xmm8, (%r10)
; AVX-NEXT:    vmovd %xmm1, (%rax)
; AVX-NEXT:    retq
;
; AVX2-LABEL: load_i8_stride8_vf4:
; AVX2:       # %bb.0:
; AVX2-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX2-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm0 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX2-NEXT:    vmovdqa (%rdi), %xmm1
; AVX2-NEXT:    vmovdqa 16(%rdi), %xmm2
; AVX2-NEXT:    vpshufb %xmm0, %xmm2, %xmm3
; AVX2-NEXT:    vpshufb %xmm0, %xmm1, %xmm0
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm3 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX2-NEXT:    vpshufb %xmm3, %xmm2, %xmm4
; AVX2-NEXT:    vpshufb %xmm3, %xmm1, %xmm3
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm4 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX2-NEXT:    vpshufb %xmm4, %xmm2, %xmm5
; AVX2-NEXT:    vpshufb %xmm4, %xmm1, %xmm4
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm5 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX2-NEXT:    vpshufb %xmm5, %xmm2, %xmm6
; AVX2-NEXT:    vpshufb %xmm5, %xmm1, %xmm5
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm6 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX2-NEXT:    vpshufb %xmm6, %xmm2, %xmm7
; AVX2-NEXT:    vpshufb %xmm6, %xmm1, %xmm6
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm6[0],xmm7[0],xmm6[1],xmm7[1],xmm6[2],xmm7[2],xmm6[3],xmm7[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm7 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX2-NEXT:    vpshufb %xmm7, %xmm2, %xmm8
; AVX2-NEXT:    vpshufb %xmm7, %xmm1, %xmm7
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm7 = xmm7[0],xmm8[0],xmm7[1],xmm8[1],xmm7[2],xmm8[2],xmm7[3],xmm8[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm8 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX2-NEXT:    vpshufb %xmm8, %xmm2, %xmm9
; AVX2-NEXT:    vpshufb %xmm8, %xmm1, %xmm8
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm9[0],xmm8[1],xmm9[1],xmm8[2],xmm9[2],xmm8[3],xmm9[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm9 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX2-NEXT:    vpshufb %xmm9, %xmm2, %xmm2
; AVX2-NEXT:    vpshufb %xmm9, %xmm1, %xmm1
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
; AVX2-NEXT:    vmovd %xmm0, (%rsi)
; AVX2-NEXT:    vmovd %xmm3, (%rdx)
; AVX2-NEXT:    vmovd %xmm4, (%rcx)
; AVX2-NEXT:    vmovd %xmm5, (%r8)
; AVX2-NEXT:    vmovd %xmm6, (%r9)
; AVX2-NEXT:    vmovd %xmm7, (%r11)
; AVX2-NEXT:    vmovd %xmm8, (%r10)
; AVX2-NEXT:    vmovd %xmm1, (%rax)
; AVX2-NEXT:    retq
;
; AVX2-FP-LABEL: load_i8_stride8_vf4:
; AVX2-FP:       # %bb.0:
; AVX2-FP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FP-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX2-FP-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm0 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX2-FP-NEXT:    vmovdqa (%rdi), %xmm1
; AVX2-FP-NEXT:    vmovdqa 16(%rdi), %xmm2
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm2, %xmm3
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm1, %xmm0
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm3 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm2, %xmm4
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm1, %xmm3
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm4 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX2-FP-NEXT:    vpshufb %xmm4, %xmm2, %xmm5
; AVX2-FP-NEXT:    vpshufb %xmm4, %xmm1, %xmm4
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm5 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX2-FP-NEXT:    vpshufb %xmm5, %xmm2, %xmm6
; AVX2-FP-NEXT:    vpshufb %xmm5, %xmm1, %xmm5
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm6 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX2-FP-NEXT:    vpshufb %xmm6, %xmm2, %xmm7
; AVX2-FP-NEXT:    vpshufb %xmm6, %xmm1, %xmm6
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm6[0],xmm7[0],xmm6[1],xmm7[1],xmm6[2],xmm7[2],xmm6[3],xmm7[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm7 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX2-FP-NEXT:    vpshufb %xmm7, %xmm2, %xmm8
; AVX2-FP-NEXT:    vpshufb %xmm7, %xmm1, %xmm7
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm7 = xmm7[0],xmm8[0],xmm7[1],xmm8[1],xmm7[2],xmm8[2],xmm7[3],xmm8[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm8 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX2-FP-NEXT:    vpshufb %xmm8, %xmm2, %xmm9
; AVX2-FP-NEXT:    vpshufb %xmm8, %xmm1, %xmm8
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm9[0],xmm8[1],xmm9[1],xmm8[2],xmm9[2],xmm8[3],xmm9[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm9 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX2-FP-NEXT:    vpshufb %xmm9, %xmm2, %xmm2
; AVX2-FP-NEXT:    vpshufb %xmm9, %xmm1, %xmm1
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
; AVX2-FP-NEXT:    vmovd %xmm0, (%rsi)
; AVX2-FP-NEXT:    vmovd %xmm3, (%rdx)
; AVX2-FP-NEXT:    vmovd %xmm4, (%rcx)
; AVX2-FP-NEXT:    vmovd %xmm5, (%r8)
; AVX2-FP-NEXT:    vmovd %xmm6, (%r9)
; AVX2-FP-NEXT:    vmovd %xmm7, (%r11)
; AVX2-FP-NEXT:    vmovd %xmm8, (%r10)
; AVX2-FP-NEXT:    vmovd %xmm1, (%rax)
; AVX2-FP-NEXT:    retq
;
; AVX2-FCP-LABEL: load_i8_stride8_vf4:
; AVX2-FCP:       # %bb.0:
; AVX2-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX2-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm0 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX2-FCP-NEXT:    vmovdqa (%rdi), %xmm1
; AVX2-FCP-NEXT:    vmovdqa 16(%rdi), %xmm2
; AVX2-FCP-NEXT:    vpshufb %xmm0, %xmm2, %xmm3
; AVX2-FCP-NEXT:    vpshufb %xmm0, %xmm1, %xmm0
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm3 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX2-FCP-NEXT:    vpshufb %xmm3, %xmm2, %xmm4
; AVX2-FCP-NEXT:    vpshufb %xmm3, %xmm1, %xmm3
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm4 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX2-FCP-NEXT:    vpshufb %xmm4, %xmm2, %xmm5
; AVX2-FCP-NEXT:    vpshufb %xmm4, %xmm1, %xmm4
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm5 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX2-FCP-NEXT:    vpshufb %xmm5, %xmm2, %xmm6
; AVX2-FCP-NEXT:    vpshufb %xmm5, %xmm1, %xmm5
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm6 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX2-FCP-NEXT:    vpshufb %xmm6, %xmm2, %xmm7
; AVX2-FCP-NEXT:    vpshufb %xmm6, %xmm1, %xmm6
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm6[0],xmm7[0],xmm6[1],xmm7[1],xmm6[2],xmm7[2],xmm6[3],xmm7[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm7 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX2-FCP-NEXT:    vpshufb %xmm7, %xmm2, %xmm8
; AVX2-FCP-NEXT:    vpshufb %xmm7, %xmm1, %xmm7
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm7 = xmm7[0],xmm8[0],xmm7[1],xmm8[1],xmm7[2],xmm8[2],xmm7[3],xmm8[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm8 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX2-FCP-NEXT:    vpshufb %xmm8, %xmm2, %xmm9
; AVX2-FCP-NEXT:    vpshufb %xmm8, %xmm1, %xmm8
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm9[0],xmm8[1],xmm9[1],xmm8[2],xmm9[2],xmm8[3],xmm9[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm9 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX2-FCP-NEXT:    vpshufb %xmm9, %xmm2, %xmm2
; AVX2-FCP-NEXT:    vpshufb %xmm9, %xmm1, %xmm1
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
; AVX2-FCP-NEXT:    vmovd %xmm0, (%rsi)
; AVX2-FCP-NEXT:    vmovd %xmm3, (%rdx)
; AVX2-FCP-NEXT:    vmovd %xmm4, (%rcx)
; AVX2-FCP-NEXT:    vmovd %xmm5, (%r8)
; AVX2-FCP-NEXT:    vmovd %xmm6, (%r9)
; AVX2-FCP-NEXT:    vmovd %xmm7, (%r11)
; AVX2-FCP-NEXT:    vmovd %xmm8, (%r10)
; AVX2-FCP-NEXT:    vmovd %xmm1, (%rax)
; AVX2-FCP-NEXT:    retq
;
; AVX512-LABEL: load_i8_stride8_vf4:
; AVX512:       # %bb.0:
; AVX512-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX512-NEXT:    vmovdqa (%rdi), %ymm0
; AVX512-NEXT:    vpsrlq $8, %ymm0, %ymm1
; AVX512-NEXT:    vpsrlq $16, %ymm0, %ymm2
; AVX512-NEXT:    vpsrlq $24, %ymm0, %ymm3
; AVX512-NEXT:    vpsrlq $32, %ymm0, %ymm4
; AVX512-NEXT:    vpsrlq $40, %ymm0, %ymm5
; AVX512-NEXT:    vpsrlq $48, %ymm0, %ymm6
; AVX512-NEXT:    vpsrlq $56, %ymm0, %ymm7
; AVX512-NEXT:    vpmovqb %ymm0, (%rsi)
; AVX512-NEXT:    vpmovqb %ymm1, (%rdx)
; AVX512-NEXT:    vpmovqb %ymm2, (%rcx)
; AVX512-NEXT:    vpmovqb %ymm3, (%r8)
; AVX512-NEXT:    vpmovqb %ymm4, (%r9)
; AVX512-NEXT:    vpmovqb %ymm5, (%r11)
; AVX512-NEXT:    vpmovqb %ymm6, (%r10)
; AVX512-NEXT:    vpmovqb %ymm7, (%rax)
; AVX512-NEXT:    vzeroupper
; AVX512-NEXT:    retq
;
; AVX512-FCP-LABEL: load_i8_stride8_vf4:
; AVX512-FCP:       # %bb.0:
; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX512-FCP-NEXT:    vmovdqa (%rdi), %ymm0
; AVX512-FCP-NEXT:    vpsrlq $8, %ymm0, %ymm1
; AVX512-FCP-NEXT:    vpsrlq $16, %ymm0, %ymm2
; AVX512-FCP-NEXT:    vpsrlq $24, %ymm0, %ymm3
; AVX512-FCP-NEXT:    vpsrlq $32, %ymm0, %ymm4
; AVX512-FCP-NEXT:    vpsrlq $40, %ymm0, %ymm5
; AVX512-FCP-NEXT:    vpsrlq $48, %ymm0, %ymm6
; AVX512-FCP-NEXT:    vpsrlq $56, %ymm0, %ymm7
; AVX512-FCP-NEXT:    vpmovqb %ymm0, (%rsi)
; AVX512-FCP-NEXT:    vpmovqb %ymm1, (%rdx)
; AVX512-FCP-NEXT:    vpmovqb %ymm2, (%rcx)
; AVX512-FCP-NEXT:    vpmovqb %ymm3, (%r8)
; AVX512-FCP-NEXT:    vpmovqb %ymm4, (%r9)
; AVX512-FCP-NEXT:    vpmovqb %ymm5, (%r11)
; AVX512-FCP-NEXT:    vpmovqb %ymm6, (%r10)
; AVX512-FCP-NEXT:    vpmovqb %ymm7, (%rax)
; AVX512-FCP-NEXT:    vzeroupper
; AVX512-FCP-NEXT:    retq
;
; AVX512DQ-LABEL: load_i8_stride8_vf4:
; AVX512DQ:       # %bb.0:
; AVX512DQ-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512DQ-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX512DQ-NEXT:    vmovdqa (%rdi), %ymm0
; AVX512DQ-NEXT:    vpsrlq $8, %ymm0, %ymm1
; AVX512DQ-NEXT:    vpsrlq $16, %ymm0, %ymm2
; AVX512DQ-NEXT:    vpsrlq $24, %ymm0, %ymm3
; AVX512DQ-NEXT:    vpsrlq $32, %ymm0, %ymm4
; AVX512DQ-NEXT:    vpsrlq $40, %ymm0, %ymm5
; AVX512DQ-NEXT:    vpsrlq $48, %ymm0, %ymm6
; AVX512DQ-NEXT:    vpsrlq $56, %ymm0, %ymm7
; AVX512DQ-NEXT:    vpmovqb %ymm0, (%rsi)
; AVX512DQ-NEXT:    vpmovqb %ymm1, (%rdx)
; AVX512DQ-NEXT:    vpmovqb %ymm2, (%rcx)
; AVX512DQ-NEXT:    vpmovqb %ymm3, (%r8)
; AVX512DQ-NEXT:    vpmovqb %ymm4, (%r9)
; AVX512DQ-NEXT:    vpmovqb %ymm5, (%r11)
; AVX512DQ-NEXT:    vpmovqb %ymm6, (%r10)
; AVX512DQ-NEXT:    vpmovqb %ymm7, (%rax)
; AVX512DQ-NEXT:    vzeroupper
; AVX512DQ-NEXT:    retq
;
; AVX512DQ-FCP-LABEL: load_i8_stride8_vf4:
; AVX512DQ-FCP:       # %bb.0:
; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX512DQ-FCP-NEXT:    vmovdqa (%rdi), %ymm0
; AVX512DQ-FCP-NEXT:    vpsrlq $8, %ymm0, %ymm1
; AVX512DQ-FCP-NEXT:    vpsrlq $16, %ymm0, %ymm2
; AVX512DQ-FCP-NEXT:    vpsrlq $24, %ymm0, %ymm3
; AVX512DQ-FCP-NEXT:    vpsrlq $32, %ymm0, %ymm4
; AVX512DQ-FCP-NEXT:    vpsrlq $40, %ymm0, %ymm5
; AVX512DQ-FCP-NEXT:    vpsrlq $48, %ymm0, %ymm6
; AVX512DQ-FCP-NEXT:    vpsrlq $56, %ymm0, %ymm7
; AVX512DQ-FCP-NEXT:    vpmovqb %ymm0, (%rsi)
; AVX512DQ-FCP-NEXT:    vpmovqb %ymm1, (%rdx)
; AVX512DQ-FCP-NEXT:    vpmovqb %ymm2, (%rcx)
; AVX512DQ-FCP-NEXT:    vpmovqb %ymm3, (%r8)
; AVX512DQ-FCP-NEXT:    vpmovqb %ymm4, (%r9)
; AVX512DQ-FCP-NEXT:    vpmovqb %ymm5, (%r11)
; AVX512DQ-FCP-NEXT:    vpmovqb %ymm6, (%r10)
; AVX512DQ-FCP-NEXT:    vpmovqb %ymm7, (%rax)
; AVX512DQ-FCP-NEXT:    vzeroupper
; AVX512DQ-FCP-NEXT:    retq
;
; AVX512BW-LABEL: load_i8_stride8_vf4:
; AVX512BW:       # %bb.0:
; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX512BW-NEXT:    vmovdqa (%rdi), %ymm0
; AVX512BW-NEXT:    vpsrlq $8, %ymm0, %ymm1
; AVX512BW-NEXT:    vpsrlq $16, %ymm0, %ymm2
; AVX512BW-NEXT:    vpsrlq $24, %ymm0, %ymm3
; AVX512BW-NEXT:    vpsrlq $32, %ymm0, %ymm4
; AVX512BW-NEXT:    vpsrlq $40, %ymm0, %ymm5
; AVX512BW-NEXT:    vpsrlq $48, %ymm0, %ymm6
; AVX512BW-NEXT:    vpsrlq $56, %ymm0, %ymm7
; AVX512BW-NEXT:    vpmovqb %ymm0, (%rsi)
; AVX512BW-NEXT:    vpmovqb %ymm1, (%rdx)
; AVX512BW-NEXT:    vpmovqb %ymm2, (%rcx)
; AVX512BW-NEXT:    vpmovqb %ymm3, (%r8)
; AVX512BW-NEXT:    vpmovqb %ymm4, (%r9)
; AVX512BW-NEXT:    vpmovqb %ymm5, (%r11)
; AVX512BW-NEXT:    vpmovqb %ymm6, (%r10)
; AVX512BW-NEXT:    vpmovqb %ymm7, (%rax)
; AVX512BW-NEXT:    vzeroupper
; AVX512BW-NEXT:    retq
;
; AVX512BW-FCP-LABEL: load_i8_stride8_vf4:
; AVX512BW-FCP:       # %bb.0:
; AVX512BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX512BW-FCP-NEXT:    vmovdqa (%rdi), %ymm0
; AVX512BW-FCP-NEXT:    vpsrlq $8, %ymm0, %ymm1
; AVX512BW-FCP-NEXT:    vpsrlq $16, %ymm0, %ymm2
; AVX512BW-FCP-NEXT:    vpsrlq $24, %ymm0, %ymm3
; AVX512BW-FCP-NEXT:    vpsrlq $32, %ymm0, %ymm4
; AVX512BW-FCP-NEXT:    vpsrlq $40, %ymm0, %ymm5
; AVX512BW-FCP-NEXT:    vpsrlq $48, %ymm0, %ymm6
; AVX512BW-FCP-NEXT:    vpsrlq $56, %ymm0, %ymm7
; AVX512BW-FCP-NEXT:    vpmovqb %ymm0, (%rsi)
; AVX512BW-FCP-NEXT:    vpmovqb %ymm1, (%rdx)
; AVX512BW-FCP-NEXT:    vpmovqb %ymm2, (%rcx)
; AVX512BW-FCP-NEXT:    vpmovqb %ymm3, (%r8)
; AVX512BW-FCP-NEXT:    vpmovqb %ymm4, (%r9)
; AVX512BW-FCP-NEXT:    vpmovqb %ymm5, (%r11)
; AVX512BW-FCP-NEXT:    vpmovqb %ymm6, (%r10)
; AVX512BW-FCP-NEXT:    vpmovqb %ymm7, (%rax)
; AVX512BW-FCP-NEXT:    vzeroupper
; AVX512BW-FCP-NEXT:    retq
;
; AVX512DQ-BW-LABEL: load_i8_stride8_vf4:
; AVX512DQ-BW:       # %bb.0:
; AVX512DQ-BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-BW-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512DQ-BW-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX512DQ-BW-NEXT:    vmovdqa (%rdi), %ymm0
; AVX512DQ-BW-NEXT:    vpsrlq $8, %ymm0, %ymm1
; AVX512DQ-BW-NEXT:    vpsrlq $16, %ymm0, %ymm2
; AVX512DQ-BW-NEXT:    vpsrlq $24, %ymm0, %ymm3
; AVX512DQ-BW-NEXT:    vpsrlq $32, %ymm0, %ymm4
; AVX512DQ-BW-NEXT:    vpsrlq $40, %ymm0, %ymm5
; AVX512DQ-BW-NEXT:    vpsrlq $48, %ymm0, %ymm6
; AVX512DQ-BW-NEXT:    vpsrlq $56, %ymm0, %ymm7
; AVX512DQ-BW-NEXT:    vpmovqb %ymm0, (%rsi)
; AVX512DQ-BW-NEXT:    vpmovqb %ymm1, (%rdx)
; AVX512DQ-BW-NEXT:    vpmovqb %ymm2, (%rcx)
; AVX512DQ-BW-NEXT:    vpmovqb %ymm3, (%r8)
; AVX512DQ-BW-NEXT:    vpmovqb %ymm4, (%r9)
; AVX512DQ-BW-NEXT:    vpmovqb %ymm5, (%r11)
; AVX512DQ-BW-NEXT:    vpmovqb %ymm6, (%r10)
; AVX512DQ-BW-NEXT:    vpmovqb %ymm7, (%rax)
; AVX512DQ-BW-NEXT:    vzeroupper
; AVX512DQ-BW-NEXT:    retq
;
; AVX512DQ-BW-FCP-LABEL: load_i8_stride8_vf4:
; AVX512DQ-BW-FCP:       # %bb.0:
; AVX512DQ-BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512DQ-BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%rdi), %ymm0
; AVX512DQ-BW-FCP-NEXT:    vpsrlq $8, %ymm0, %ymm1
; AVX512DQ-BW-FCP-NEXT:    vpsrlq $16, %ymm0, %ymm2
; AVX512DQ-BW-FCP-NEXT:    vpsrlq $24, %ymm0, %ymm3
; AVX512DQ-BW-FCP-NEXT:    vpsrlq $32, %ymm0, %ymm4
; AVX512DQ-BW-FCP-NEXT:    vpsrlq $40, %ymm0, %ymm5
; AVX512DQ-BW-FCP-NEXT:    vpsrlq $48, %ymm0, %ymm6
; AVX512DQ-BW-FCP-NEXT:    vpsrlq $56, %ymm0, %ymm7
; AVX512DQ-BW-FCP-NEXT:    vpmovqb %ymm0, (%rsi)
; AVX512DQ-BW-FCP-NEXT:    vpmovqb %ymm1, (%rdx)
; AVX512DQ-BW-FCP-NEXT:    vpmovqb %ymm2, (%rcx)
; AVX512DQ-BW-FCP-NEXT:    vpmovqb %ymm3, (%r8)
; AVX512DQ-BW-FCP-NEXT:    vpmovqb %ymm4, (%r9)
; AVX512DQ-BW-FCP-NEXT:    vpmovqb %ymm5, (%r11)
; AVX512DQ-BW-FCP-NEXT:    vpmovqb %ymm6, (%r10)
; AVX512DQ-BW-FCP-NEXT:    vpmovqb %ymm7, (%rax)
; AVX512DQ-BW-FCP-NEXT:    vzeroupper
; AVX512DQ-BW-FCP-NEXT:    retq
  %wide.vec = load <32 x i8>, ptr %in.vec, align 64
  %strided.vec0 = shufflevector <32 x i8> %wide.vec, <32 x i8> poison, <4 x i32> <i32 0, i32 8, i32 16, i32 24>
  %strided.vec1 = shufflevector <32 x i8> %wide.vec, <32 x i8> poison, <4 x i32> <i32 1, i32 9, i32 17, i32 25>
  %strided.vec2 = shufflevector <32 x i8> %wide.vec, <32 x i8> poison, <4 x i32> <i32 2, i32 10, i32 18, i32 26>
  %strided.vec3 = shufflevector <32 x i8> %wide.vec, <32 x i8> poison, <4 x i32> <i32 3, i32 11, i32 19, i32 27>
  %strided.vec4 = shufflevector <32 x i8> %wide.vec, <32 x i8> poison, <4 x i32> <i32 4, i32 12, i32 20, i32 28>
  %strided.vec5 = shufflevector <32 x i8> %wide.vec, <32 x i8> poison, <4 x i32> <i32 5, i32 13, i32 21, i32 29>
  %strided.vec6 = shufflevector <32 x i8> %wide.vec, <32 x i8> poison, <4 x i32> <i32 6, i32 14, i32 22, i32 30>
  %strided.vec7 = shufflevector <32 x i8> %wide.vec, <32 x i8> poison, <4 x i32> <i32 7, i32 15, i32 23, i32 31>
  store <4 x i8> %strided.vec0, ptr %out.vec0, align 64
  store <4 x i8> %strided.vec1, ptr %out.vec1, align 64
  store <4 x i8> %strided.vec2, ptr %out.vec2, align 64
  store <4 x i8> %strided.vec3, ptr %out.vec3, align 64
  store <4 x i8> %strided.vec4, ptr %out.vec4, align 64
  store <4 x i8> %strided.vec5, ptr %out.vec5, align 64
  store <4 x i8> %strided.vec6, ptr %out.vec6, align 64
  store <4 x i8> %strided.vec7, ptr %out.vec7, align 64
  ret void
}

define void @load_i8_stride8_vf8(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr %out.vec2, ptr %out.vec3, ptr %out.vec4, ptr %out.vec5, ptr %out.vec6, ptr %out.vec7) nounwind {
; SSE-LABEL: load_i8_stride8_vf8:
; SSE:       # %bb.0:
; SSE-NEXT:    pushq %rax
; SSE-NEXT:    movdqa (%rdi), %xmm12
; SSE-NEXT:    movdqa 16(%rdi), %xmm11
; SSE-NEXT:    movdqa 32(%rdi), %xmm9
; SSE-NEXT:    movdqa 48(%rdi), %xmm10
; SSE-NEXT:    movdqa {{.*#+}} xmm0 = [255,0,0,0,255,0,0,0]
; SSE-NEXT:    movdqa %xmm11, %xmm1
; SSE-NEXT:    pand %xmm0, %xmm1
; SSE-NEXT:    movdqa %xmm12, %xmm2
; SSE-NEXT:    pand %xmm0, %xmm2
; SSE-NEXT:    packuswb %xmm1, %xmm2
; SSE-NEXT:    packuswb %xmm2, %xmm2
; SSE-NEXT:    movdqa %xmm10, %xmm1
; SSE-NEXT:    pand %xmm0, %xmm1
; SSE-NEXT:    pand %xmm9, %xmm0
; SSE-NEXT:    packuswb %xmm1, %xmm0
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    packuswb %xmm0, %xmm2
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pxor %xmm7, %xmm7
; SSE-NEXT:    movdqa %xmm11, %xmm0
; SSE-NEXT:    punpckhbw {{.*#+}} xmm0 = xmm0[8],xmm7[8],xmm0[9],xmm7[9],xmm0[10],xmm7[10],xmm0[11],xmm7[11],xmm0[12],xmm7[12],xmm0[13],xmm7[13],xmm0[14],xmm7[14],xmm0[15],xmm7[15]
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm11, %xmm14
; SSE-NEXT:    punpcklbw {{.*#+}} xmm14 = xmm14[0],xmm7[0],xmm14[1],xmm7[1],xmm14[2],xmm7[2],xmm14[3],xmm7[3],xmm14[4],xmm7[4],xmm14[5],xmm7[5],xmm14[6],xmm7[6],xmm14[7],xmm7[7]
; SSE-NEXT:    movdqa %xmm14, %xmm15
; SSE-NEXT:    punpcklwd {{.*#+}} xmm15 = xmm15[0],xmm0[0],xmm15[1],xmm0[1],xmm15[2],xmm0[2],xmm15[3],xmm0[3]
; SSE-NEXT:    packuswb %xmm15, %xmm15
; SSE-NEXT:    movdqa {{.*#+}} xmm3 = [65535,0,65535,65535,65535,65535,65535,65535]
; SSE-NEXT:    movdqa %xmm3, %xmm0
; SSE-NEXT:    pandn %xmm15, %xmm0
; SSE-NEXT:    movdqa %xmm12, %xmm1
; SSE-NEXT:    punpckhbw {{.*#+}} xmm1 = xmm1[8],xmm7[8],xmm1[9],xmm7[9],xmm1[10],xmm7[10],xmm1[11],xmm7[11],xmm1[12],xmm7[12],xmm1[13],xmm7[13],xmm1[14],xmm7[14],xmm1[15],xmm7[15]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm12, %xmm5
; SSE-NEXT:    punpcklbw {{.*#+}} xmm5 = xmm5[0],xmm7[0],xmm5[1],xmm7[1],xmm5[2],xmm7[2],xmm5[3],xmm7[3],xmm5[4],xmm7[4],xmm5[5],xmm7[5],xmm5[6],xmm7[6],xmm5[7],xmm7[7]
; SSE-NEXT:    movdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklwd {{.*#+}} xmm5 = xmm5[0],xmm1[0],xmm5[1],xmm1[1],xmm5[2],xmm1[2],xmm5[3],xmm1[3]
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm5[1,1,1,1]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm3, %xmm1
; SSE-NEXT:    por %xmm0, %xmm1
; SSE-NEXT:    movdqa %xmm10, %xmm0
; SSE-NEXT:    punpckhbw {{.*#+}} xmm0 = xmm0[8],xmm7[8],xmm0[9],xmm7[9],xmm0[10],xmm7[10],xmm0[11],xmm7[11],xmm0[12],xmm7[12],xmm0[13],xmm7[13],xmm0[14],xmm7[14],xmm0[15],xmm7[15]
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm10, %xmm13
; SSE-NEXT:    punpcklbw {{.*#+}} xmm13 = xmm13[0],xmm7[0],xmm13[1],xmm7[1],xmm13[2],xmm7[2],xmm13[3],xmm7[3],xmm13[4],xmm7[4],xmm13[5],xmm7[5],xmm13[6],xmm7[6],xmm13[7],xmm7[7]
; SSE-NEXT:    movdqa %xmm13, %xmm4
; SSE-NEXT:    punpcklwd {{.*#+}} xmm4 = xmm4[0],xmm0[0],xmm4[1],xmm0[1],xmm4[2],xmm0[2],xmm4[3],xmm0[3]
; SSE-NEXT:    packuswb %xmm4, %xmm4
; SSE-NEXT:    pshufd {{.*#+}} xmm8 = xmm4[0,0,2,3]
; SSE-NEXT:    movdqa {{.*#+}} xmm2 = [65535,65535,65535,0,65535,65535,65535,65535]
; SSE-NEXT:    movdqa %xmm2, %xmm6
; SSE-NEXT:    pandn %xmm8, %xmm6
; SSE-NEXT:    movdqa %xmm9, %xmm8
; SSE-NEXT:    punpckhbw {{.*#+}} xmm8 = xmm8[8],xmm7[8],xmm8[9],xmm7[9],xmm8[10],xmm7[10],xmm8[11],xmm7[11],xmm8[12],xmm7[12],xmm8[13],xmm7[13],xmm8[14],xmm7[14],xmm8[15],xmm7[15]
; SSE-NEXT:    movdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm9, %xmm0
; SSE-NEXT:    punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm7[0],xmm0[1],xmm7[1],xmm0[2],xmm7[2],xmm0[3],xmm7[3],xmm0[4],xmm7[4],xmm0[5],xmm7[5],xmm0[6],xmm7[6],xmm0[7],xmm7[7]
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm0, %xmm7
; SSE-NEXT:    punpcklwd {{.*#+}} xmm7 = xmm7[0],xmm8[0],xmm7[1],xmm8[1],xmm7[2],xmm8[2],xmm7[3],xmm8[3]
; SSE-NEXT:    pshufd {{.*#+}} xmm8 = xmm7[0,1,1,3]
; SSE-NEXT:    packuswb %xmm8, %xmm8
; SSE-NEXT:    pand %xmm2, %xmm8
; SSE-NEXT:    por %xmm6, %xmm8
; SSE-NEXT:    pshufd {{.*#+}} xmm6 = xmm8[1,1,1,1]
; SSE-NEXT:    punpckldq {{.*#+}} xmm1 = xmm1[0],xmm6[0],xmm1[1],xmm6[1]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa {{.*#+}} xmm6 = [255,255,255,255,255,255,255,255]
; SSE-NEXT:    pand %xmm6, %xmm11
; SSE-NEXT:    pshufd {{.*#+}} xmm8 = xmm11[0,2,2,3]
; SSE-NEXT:    pshuflw {{.*#+}} xmm8 = xmm8[0,1,1,3,4,5,6,7]
; SSE-NEXT:    packuswb %xmm8, %xmm8
; SSE-NEXT:    movdqa %xmm3, %xmm1
; SSE-NEXT:    pandn %xmm8, %xmm1
; SSE-NEXT:    pand %xmm6, %xmm12
; SSE-NEXT:    pshufd {{.*#+}} xmm8 = xmm12[0,2,2,3]
; SSE-NEXT:    pshuflw {{.*#+}} xmm8 = xmm8[1,3,2,3,4,5,6,7]
; SSE-NEXT:    packuswb %xmm8, %xmm8
; SSE-NEXT:    pand %xmm3, %xmm8
; SSE-NEXT:    por %xmm1, %xmm8
; SSE-NEXT:    pand %xmm6, %xmm10
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm10[0,1,2,0]
; SSE-NEXT:    pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,5,7,5]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    movdqa %xmm2, %xmm0
; SSE-NEXT:    pandn %xmm1, %xmm0
; SSE-NEXT:    pand %xmm6, %xmm9
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm9[0,1,2,0]
; SSE-NEXT:    pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,7,5,6,7]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm2, %xmm1
; SSE-NEXT:    por %xmm0, %xmm1
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
; SSE-NEXT:    punpckldq {{.*#+}} xmm8 = xmm8[0],xmm0[0],xmm8[1],xmm0[1]
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm15[1,1,2,3]
; SSE-NEXT:    movdqa %xmm3, %xmm1
; SSE-NEXT:    pandn %xmm0, %xmm1
; SSE-NEXT:    pshufd {{.*#+}} xmm15 = xmm5[3,3,3,3]
; SSE-NEXT:    packuswb %xmm15, %xmm15
; SSE-NEXT:    pand %xmm3, %xmm15
; SSE-NEXT:    por %xmm1, %xmm15
; SSE-NEXT:    movdqa %xmm2, %xmm0
; SSE-NEXT:    pandn %xmm4, %xmm0
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm7[2,2,3,3]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm2, %xmm1
; SSE-NEXT:    por %xmm0, %xmm1
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
; SSE-NEXT:    punpckldq {{.*#+}} xmm15 = xmm15[0],xmm0[0],xmm15[1],xmm0[1]
; SSE-NEXT:    pshufd {{.*#+}} xmm4 = xmm11[3,1,2,3]
; SSE-NEXT:    pshuflw {{.*#+}} xmm0 = xmm4[0,1,2,0,4,5,6,7]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    movdqa %xmm3, %xmm1
; SSE-NEXT:    pandn %xmm0, %xmm1
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm12[3,1,2,3]
; SSE-NEXT:    pshuflw {{.*#+}} xmm11 = xmm5[2,0,2,3,4,5,6,7]
; SSE-NEXT:    packuswb %xmm11, %xmm11
; SSE-NEXT:    pand %xmm3, %xmm11
; SSE-NEXT:    por %xmm1, %xmm11
; SSE-NEXT:    pshufd {{.*#+}} xmm10 = xmm10[0,1,1,3]
; SSE-NEXT:    pshufhw {{.*#+}} xmm0 = xmm10[0,1,2,3,4,5,4,6]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    movdqa %xmm2, %xmm1
; SSE-NEXT:    pandn %xmm0, %xmm1
; SSE-NEXT:    pshufd {{.*#+}} xmm12 = xmm9[0,1,1,3]
; SSE-NEXT:    pshufhw {{.*#+}} xmm0 = xmm12[0,1,2,3,4,6,6,7]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    pand %xmm2, %xmm0
; SSE-NEXT:    por %xmm1, %xmm0
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,1,1]
; SSE-NEXT:    punpckldq {{.*#+}} xmm11 = xmm11[0],xmm0[0],xmm11[1],xmm0[1]
; SSE-NEXT:    punpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Folded Reload
; SSE-NEXT:    # xmm14 = xmm14[4],mem[4],xmm14[5],mem[5],xmm14[6],mem[6],xmm14[7],mem[7]
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; SSE-NEXT:    punpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Folded Reload
; SSE-NEXT:    # xmm6 = xmm6[4],mem[4],xmm6[5],mem[5],xmm6[6],mem[6],xmm6[7],mem[7]
; SSE-NEXT:    packuswb %xmm14, %xmm14
; SSE-NEXT:    movdqa %xmm3, %xmm0
; SSE-NEXT:    pandn %xmm14, %xmm0
; SSE-NEXT:    pshufd {{.*#+}} xmm9 = xmm6[1,1,1,1]
; SSE-NEXT:    packuswb %xmm9, %xmm9
; SSE-NEXT:    pand %xmm3, %xmm9
; SSE-NEXT:    por %xmm0, %xmm9
; SSE-NEXT:    punpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Folded Reload
; SSE-NEXT:    # xmm13 = xmm13[4],mem[4],xmm13[5],mem[5],xmm13[6],mem[6],xmm13[7],mem[7]
; SSE-NEXT:    packuswb %xmm13, %xmm13
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm13[0,0,2,3]
; SSE-NEXT:    movdqa %xmm2, %xmm1
; SSE-NEXT:    pandn %xmm0, %xmm1
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
; SSE-NEXT:    punpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Folded Reload
; SSE-NEXT:    # xmm7 = xmm7[4],mem[4],xmm7[5],mem[5],xmm7[6],mem[6],xmm7[7],mem[7]
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm7[0,1,1,3]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    pand %xmm2, %xmm0
; SSE-NEXT:    por %xmm1, %xmm0
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,1,1]
; SSE-NEXT:    punpckldq {{.*#+}} xmm9 = xmm9[0],xmm0[0],xmm9[1],xmm0[1]
; SSE-NEXT:    pshuflw {{.*#+}} xmm0 = xmm4[0,1,3,1,4,5,6,7]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    movdqa %xmm3, %xmm1
; SSE-NEXT:    pandn %xmm0, %xmm1
; SSE-NEXT:    pshuflw {{.*#+}} xmm4 = xmm5[3,1,2,3,4,5,6,7]
; SSE-NEXT:    packuswb %xmm4, %xmm4
; SSE-NEXT:    pand %xmm3, %xmm4
; SSE-NEXT:    por %xmm1, %xmm4
; SSE-NEXT:    pshufhw {{.*#+}} xmm0 = xmm10[0,1,2,3,4,5,5,7]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    movdqa %xmm2, %xmm1
; SSE-NEXT:    pandn %xmm0, %xmm1
; SSE-NEXT:    pshufhw {{.*#+}} xmm0 = xmm12[0,1,2,3,5,7,6,7]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    pand %xmm2, %xmm0
; SSE-NEXT:    por %xmm1, %xmm0
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,1,1]
; SSE-NEXT:    punpckldq {{.*#+}} xmm4 = xmm4[0],xmm0[0],xmm4[1],xmm0[1]
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm14[1,1,2,3]
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm6[3,3,3,3]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm3, %xmm1
; SSE-NEXT:    pandn %xmm0, %xmm3
; SSE-NEXT:    por %xmm1, %xmm3
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm7[2,2,3,3]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    pand %xmm2, %xmm0
; SSE-NEXT:    pandn %xmm13, %xmm2
; SSE-NEXT:    por %xmm0, %xmm2
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,1,1]
; SSE-NEXT:    punpckldq {{.*#+}} xmm3 = xmm3[0],xmm0[0],xmm3[1],xmm0[1]
; SSE-NEXT:    pshufd $236, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
; SSE-NEXT:    # xmm0 = mem[0,3,2,3]
; SSE-NEXT:    movq %xmm0, (%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movlps %xmm0, (%rdx)
; SSE-NEXT:    movq %xmm8, (%rcx)
; SSE-NEXT:    movq %xmm15, (%r8)
; SSE-NEXT:    movq %xmm11, (%r9)
; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT:    movq %xmm9, (%rax)
; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT:    movq %xmm4, (%rax)
; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT:    movq %xmm3, (%rax)
; SSE-NEXT:    popq %rax
; SSE-NEXT:    retq
;
; AVX-LABEL: load_i8_stride8_vf8:
; AVX:       # %bb.0:
; AVX-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX-NEXT:    vmovd {{.*#+}} xmm4 = [0,0,0,8,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vmovdqa (%rdi), %xmm0
; AVX-NEXT:    vmovdqa 16(%rdi), %xmm1
; AVX-NEXT:    vmovdqa 32(%rdi), %xmm2
; AVX-NEXT:    vmovdqa 48(%rdi), %xmm3
; AVX-NEXT:    vpshufb %xmm4, %xmm3, %xmm5
; AVX-NEXT:    vpshufb %xmm4, %xmm2, %xmm4
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3]
; AVX-NEXT:    vmovd {{.*#+}} xmm5 = [0,8,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vpshufb %xmm5, %xmm1, %xmm6
; AVX-NEXT:    vpshufb %xmm5, %xmm0, %xmm5
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm4 = xmm5[0,1],xmm4[2,3],xmm5[4,5,6,7]
; AVX-NEXT:    vmovd {{.*#+}} xmm5 = [0,0,1,9,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vpshufb %xmm5, %xmm3, %xmm6
; AVX-NEXT:    vpshufb %xmm5, %xmm2, %xmm5
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3]
; AVX-NEXT:    vmovd {{.*#+}} xmm6 = [1,9,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vpshufb %xmm6, %xmm1, %xmm7
; AVX-NEXT:    vpshufb %xmm6, %xmm0, %xmm6
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm6[0],xmm7[0],xmm6[1],xmm7[1],xmm6[2],xmm7[2],xmm6[3],xmm7[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm5 = xmm6[0,1],xmm5[2,3],xmm6[4,5,6,7]
; AVX-NEXT:    vmovd {{.*#+}} xmm6 = [0,0,2,10,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vpshufb %xmm6, %xmm3, %xmm7
; AVX-NEXT:    vpshufb %xmm6, %xmm2, %xmm6
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm6[0],xmm7[0],xmm6[1],xmm7[1],xmm6[2],xmm7[2],xmm6[3],xmm7[3]
; AVX-NEXT:    vmovd {{.*#+}} xmm7 = [2,10,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vpshufb %xmm7, %xmm1, %xmm8
; AVX-NEXT:    vpshufb %xmm7, %xmm0, %xmm7
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm7 = xmm7[0],xmm8[0],xmm7[1],xmm8[1],xmm7[2],xmm8[2],xmm7[3],xmm8[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm6 = xmm7[0,1],xmm6[2,3],xmm7[4,5,6,7]
; AVX-NEXT:    vmovd {{.*#+}} xmm7 = [0,0,3,11,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vpshufb %xmm7, %xmm3, %xmm8
; AVX-NEXT:    vpshufb %xmm7, %xmm2, %xmm7
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm7 = xmm7[0],xmm8[0],xmm7[1],xmm8[1],xmm7[2],xmm8[2],xmm7[3],xmm8[3]
; AVX-NEXT:    vmovd {{.*#+}} xmm8 = [3,11,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vpshufb %xmm8, %xmm1, %xmm9
; AVX-NEXT:    vpshufb %xmm8, %xmm0, %xmm8
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm9[0],xmm8[1],xmm9[1],xmm8[2],xmm9[2],xmm8[3],xmm9[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm7 = xmm8[0,1],xmm7[2,3],xmm8[4,5,6,7]
; AVX-NEXT:    vmovd {{.*#+}} xmm8 = [0,0,4,12,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vpshufb %xmm8, %xmm3, %xmm9
; AVX-NEXT:    vpshufb %xmm8, %xmm2, %xmm8
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm9[0],xmm8[1],xmm9[1],xmm8[2],xmm9[2],xmm8[3],xmm9[3]
; AVX-NEXT:    vmovd {{.*#+}} xmm9 = [4,12,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vpshufb %xmm9, %xmm1, %xmm10
; AVX-NEXT:    vpshufb %xmm9, %xmm0, %xmm9
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm9[0],xmm10[0],xmm9[1],xmm10[1],xmm9[2],xmm10[2],xmm9[3],xmm10[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm8 = xmm9[0,1],xmm8[2,3],xmm9[4,5,6,7]
; AVX-NEXT:    vmovd {{.*#+}} xmm9 = [0,0,5,13,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vpshufb %xmm9, %xmm3, %xmm10
; AVX-NEXT:    vpshufb %xmm9, %xmm2, %xmm9
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm9[0],xmm10[0],xmm9[1],xmm10[1],xmm9[2],xmm10[2],xmm9[3],xmm10[3]
; AVX-NEXT:    vmovd {{.*#+}} xmm10 = [5,13,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vpshufb %xmm10, %xmm1, %xmm11
; AVX-NEXT:    vpshufb %xmm10, %xmm0, %xmm10
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm10[0],xmm11[0],xmm10[1],xmm11[1],xmm10[2],xmm11[2],xmm10[3],xmm11[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm9 = xmm10[0,1],xmm9[2,3],xmm10[4,5,6,7]
; AVX-NEXT:    vmovd {{.*#+}} xmm10 = [0,0,6,14,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vpshufb %xmm10, %xmm3, %xmm11
; AVX-NEXT:    vpshufb %xmm10, %xmm2, %xmm10
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm10[0],xmm11[0],xmm10[1],xmm11[1],xmm10[2],xmm11[2],xmm10[3],xmm11[3]
; AVX-NEXT:    vmovd {{.*#+}} xmm11 = [6,14,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vpshufb %xmm11, %xmm1, %xmm12
; AVX-NEXT:    vpshufb %xmm11, %xmm0, %xmm11
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm11[0],xmm12[0],xmm11[1],xmm12[1],xmm11[2],xmm12[2],xmm11[3],xmm12[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm10 = xmm11[0,1],xmm10[2,3],xmm11[4,5,6,7]
; AVX-NEXT:    vmovd {{.*#+}} xmm11 = [0,0,7,15,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vpshufb %xmm11, %xmm3, %xmm3
; AVX-NEXT:    vpshufb %xmm11, %xmm2, %xmm2
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
; AVX-NEXT:    vmovd {{.*#+}} xmm3 = [7,15,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vpshufb %xmm3, %xmm1, %xmm1
; AVX-NEXT:    vpshufb %xmm3, %xmm0, %xmm0
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0,1],xmm2[2,3],xmm0[4,5,6,7]
; AVX-NEXT:    vmovq %xmm4, (%rsi)
; AVX-NEXT:    vmovq %xmm5, (%rdx)
; AVX-NEXT:    vmovq %xmm6, (%rcx)
; AVX-NEXT:    vmovq %xmm7, (%r8)
; AVX-NEXT:    vmovq %xmm8, (%r9)
; AVX-NEXT:    vmovq %xmm9, (%r11)
; AVX-NEXT:    vmovq %xmm10, (%r10)
; AVX-NEXT:    vmovq %xmm0, (%rax)
; AVX-NEXT:    retq
;
; AVX2-LABEL: load_i8_stride8_vf8:
; AVX2:       # %bb.0:
; AVX2-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX2-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX2-NEXT:    vmovdqa (%rdi), %xmm0
; AVX2-NEXT:    vmovdqa 16(%rdi), %xmm1
; AVX2-NEXT:    vmovdqa 32(%rdi), %xmm2
; AVX2-NEXT:    vmovdqa 48(%rdi), %xmm3
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm4 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX2-NEXT:    vpshufb %xmm4, %xmm3, %xmm5
; AVX2-NEXT:    vpshufb %xmm4, %xmm2, %xmm4
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm5 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX2-NEXT:    vpshufb %xmm5, %xmm1, %xmm6
; AVX2-NEXT:    vpshufb %xmm5, %xmm0, %xmm5
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm4 = xmm5[0],xmm4[1],xmm5[2,3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm5 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX2-NEXT:    vpshufb %xmm5, %xmm3, %xmm6
; AVX2-NEXT:    vpshufb %xmm5, %xmm2, %xmm5
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm6 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX2-NEXT:    vpshufb %xmm6, %xmm1, %xmm7
; AVX2-NEXT:    vpshufb %xmm6, %xmm0, %xmm6
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm6[0],xmm7[0],xmm6[1],xmm7[1],xmm6[2],xmm7[2],xmm6[3],xmm7[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm5 = xmm6[0],xmm5[1],xmm6[2,3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm6 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX2-NEXT:    vpshufb %xmm6, %xmm3, %xmm7
; AVX2-NEXT:    vpshufb %xmm6, %xmm2, %xmm6
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm6[0],xmm7[0],xmm6[1],xmm7[1],xmm6[2],xmm7[2],xmm6[3],xmm7[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm7 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX2-NEXT:    vpshufb %xmm7, %xmm1, %xmm8
; AVX2-NEXT:    vpshufb %xmm7, %xmm0, %xmm7
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm7 = xmm7[0],xmm8[0],xmm7[1],xmm8[1],xmm7[2],xmm8[2],xmm7[3],xmm8[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm6 = xmm7[0],xmm6[1],xmm7[2,3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm7 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX2-NEXT:    vpshufb %xmm7, %xmm3, %xmm8
; AVX2-NEXT:    vpshufb %xmm7, %xmm2, %xmm7
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm7 = xmm7[0],xmm8[0],xmm7[1],xmm8[1],xmm7[2],xmm8[2],xmm7[3],xmm8[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm8 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX2-NEXT:    vpshufb %xmm8, %xmm1, %xmm9
; AVX2-NEXT:    vpshufb %xmm8, %xmm0, %xmm8
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm9[0],xmm8[1],xmm9[1],xmm8[2],xmm9[2],xmm8[3],xmm9[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm7 = xmm8[0],xmm7[1],xmm8[2,3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm8 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX2-NEXT:    vpshufb %xmm8, %xmm3, %xmm9
; AVX2-NEXT:    vpshufb %xmm8, %xmm2, %xmm8
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm9[0],xmm8[1],xmm9[1],xmm8[2],xmm9[2],xmm8[3],xmm9[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm9 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX2-NEXT:    vpshufb %xmm9, %xmm1, %xmm10
; AVX2-NEXT:    vpshufb %xmm9, %xmm0, %xmm9
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm9[0],xmm10[0],xmm9[1],xmm10[1],xmm9[2],xmm10[2],xmm9[3],xmm10[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm8 = xmm9[0],xmm8[1],xmm9[2,3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm9 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX2-NEXT:    vpshufb %xmm9, %xmm3, %xmm10
; AVX2-NEXT:    vpshufb %xmm9, %xmm2, %xmm9
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm9[0],xmm10[0],xmm9[1],xmm10[1],xmm9[2],xmm10[2],xmm9[3],xmm10[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm10 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX2-NEXT:    vpshufb %xmm10, %xmm1, %xmm11
; AVX2-NEXT:    vpshufb %xmm10, %xmm0, %xmm10
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm10[0],xmm11[0],xmm10[1],xmm11[1],xmm10[2],xmm11[2],xmm10[3],xmm11[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm9 = xmm10[0],xmm9[1],xmm10[2,3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm10 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX2-NEXT:    vpshufb %xmm10, %xmm3, %xmm11
; AVX2-NEXT:    vpshufb %xmm10, %xmm2, %xmm10
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm10[0],xmm11[0],xmm10[1],xmm11[1],xmm10[2],xmm11[2],xmm10[3],xmm11[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm11 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX2-NEXT:    vpshufb %xmm11, %xmm1, %xmm12
; AVX2-NEXT:    vpshufb %xmm11, %xmm0, %xmm11
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm11[0],xmm12[0],xmm11[1],xmm12[1],xmm11[2],xmm12[2],xmm11[3],xmm12[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm10 = xmm11[0],xmm10[1],xmm11[2,3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm11 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX2-NEXT:    vpshufb %xmm11, %xmm3, %xmm3
; AVX2-NEXT:    vpshufb %xmm11, %xmm2, %xmm2
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm3 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX2-NEXT:    vpshufb %xmm3, %xmm1, %xmm1
; AVX2-NEXT:    vpshufb %xmm3, %xmm0, %xmm0
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm0 = xmm0[0],xmm2[1],xmm0[2,3]
; AVX2-NEXT:    vmovq %xmm4, (%rsi)
; AVX2-NEXT:    vmovq %xmm5, (%rdx)
; AVX2-NEXT:    vmovq %xmm6, (%rcx)
; AVX2-NEXT:    vmovq %xmm7, (%r8)
; AVX2-NEXT:    vmovq %xmm8, (%r9)
; AVX2-NEXT:    vmovq %xmm9, (%r11)
; AVX2-NEXT:    vmovq %xmm10, (%r10)
; AVX2-NEXT:    vmovq %xmm0, (%rax)
; AVX2-NEXT:    retq
;
; AVX2-FP-LABEL: load_i8_stride8_vf8:
; AVX2-FP:       # %bb.0:
; AVX2-FP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FP-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX2-FP-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX2-FP-NEXT:    vmovdqa (%rdi), %xmm0
; AVX2-FP-NEXT:    vmovdqa 16(%rdi), %xmm1
; AVX2-FP-NEXT:    vmovdqa 32(%rdi), %xmm2
; AVX2-FP-NEXT:    vmovdqa 48(%rdi), %xmm3
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm4 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX2-FP-NEXT:    vpshufb %xmm4, %xmm3, %xmm5
; AVX2-FP-NEXT:    vpshufb %xmm4, %xmm2, %xmm4
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm5 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX2-FP-NEXT:    vpshufb %xmm5, %xmm1, %xmm6
; AVX2-FP-NEXT:    vpshufb %xmm5, %xmm0, %xmm5
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm4 = xmm5[0],xmm4[1],xmm5[2,3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm5 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX2-FP-NEXT:    vpshufb %xmm5, %xmm3, %xmm6
; AVX2-FP-NEXT:    vpshufb %xmm5, %xmm2, %xmm5
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm6 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX2-FP-NEXT:    vpshufb %xmm6, %xmm1, %xmm7
; AVX2-FP-NEXT:    vpshufb %xmm6, %xmm0, %xmm6
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm6[0],xmm7[0],xmm6[1],xmm7[1],xmm6[2],xmm7[2],xmm6[3],xmm7[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm5 = xmm6[0],xmm5[1],xmm6[2,3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm6 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX2-FP-NEXT:    vpshufb %xmm6, %xmm3, %xmm7
; AVX2-FP-NEXT:    vpshufb %xmm6, %xmm2, %xmm6
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm6[0],xmm7[0],xmm6[1],xmm7[1],xmm6[2],xmm7[2],xmm6[3],xmm7[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm7 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX2-FP-NEXT:    vpshufb %xmm7, %xmm1, %xmm8
; AVX2-FP-NEXT:    vpshufb %xmm7, %xmm0, %xmm7
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm7 = xmm7[0],xmm8[0],xmm7[1],xmm8[1],xmm7[2],xmm8[2],xmm7[3],xmm8[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm6 = xmm7[0],xmm6[1],xmm7[2,3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm7 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX2-FP-NEXT:    vpshufb %xmm7, %xmm3, %xmm8
; AVX2-FP-NEXT:    vpshufb %xmm7, %xmm2, %xmm7
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm7 = xmm7[0],xmm8[0],xmm7[1],xmm8[1],xmm7[2],xmm8[2],xmm7[3],xmm8[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm8 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX2-FP-NEXT:    vpshufb %xmm8, %xmm1, %xmm9
; AVX2-FP-NEXT:    vpshufb %xmm8, %xmm0, %xmm8
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm9[0],xmm8[1],xmm9[1],xmm8[2],xmm9[2],xmm8[3],xmm9[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm7 = xmm8[0],xmm7[1],xmm8[2,3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm8 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX2-FP-NEXT:    vpshufb %xmm8, %xmm3, %xmm9
; AVX2-FP-NEXT:    vpshufb %xmm8, %xmm2, %xmm8
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm9[0],xmm8[1],xmm9[1],xmm8[2],xmm9[2],xmm8[3],xmm9[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm9 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX2-FP-NEXT:    vpshufb %xmm9, %xmm1, %xmm10
; AVX2-FP-NEXT:    vpshufb %xmm9, %xmm0, %xmm9
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm9[0],xmm10[0],xmm9[1],xmm10[1],xmm9[2],xmm10[2],xmm9[3],xmm10[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm8 = xmm9[0],xmm8[1],xmm9[2,3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm9 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX2-FP-NEXT:    vpshufb %xmm9, %xmm3, %xmm10
; AVX2-FP-NEXT:    vpshufb %xmm9, %xmm2, %xmm9
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm9[0],xmm10[0],xmm9[1],xmm10[1],xmm9[2],xmm10[2],xmm9[3],xmm10[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm10 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX2-FP-NEXT:    vpshufb %xmm10, %xmm1, %xmm11
; AVX2-FP-NEXT:    vpshufb %xmm10, %xmm0, %xmm10
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm10[0],xmm11[0],xmm10[1],xmm11[1],xmm10[2],xmm11[2],xmm10[3],xmm11[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm9 = xmm10[0],xmm9[1],xmm10[2,3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm10 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX2-FP-NEXT:    vpshufb %xmm10, %xmm3, %xmm11
; AVX2-FP-NEXT:    vpshufb %xmm10, %xmm2, %xmm10
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm10[0],xmm11[0],xmm10[1],xmm11[1],xmm10[2],xmm11[2],xmm10[3],xmm11[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm11 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX2-FP-NEXT:    vpshufb %xmm11, %xmm1, %xmm12
; AVX2-FP-NEXT:    vpshufb %xmm11, %xmm0, %xmm11
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm11[0],xmm12[0],xmm11[1],xmm12[1],xmm11[2],xmm12[2],xmm11[3],xmm12[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm10 = xmm11[0],xmm10[1],xmm11[2,3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm11 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX2-FP-NEXT:    vpshufb %xmm11, %xmm3, %xmm3
; AVX2-FP-NEXT:    vpshufb %xmm11, %xmm2, %xmm2
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm3 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm1, %xmm1
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm0, %xmm0
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm0[0],xmm2[1],xmm0[2,3]
; AVX2-FP-NEXT:    vmovq %xmm4, (%rsi)
; AVX2-FP-NEXT:    vmovq %xmm5, (%rdx)
; AVX2-FP-NEXT:    vmovq %xmm6, (%rcx)
; AVX2-FP-NEXT:    vmovq %xmm7, (%r8)
; AVX2-FP-NEXT:    vmovq %xmm8, (%r9)
; AVX2-FP-NEXT:    vmovq %xmm9, (%r11)
; AVX2-FP-NEXT:    vmovq %xmm10, (%r10)
; AVX2-FP-NEXT:    vmovq %xmm0, (%rax)
; AVX2-FP-NEXT:    retq
;
; AVX2-FCP-LABEL: load_i8_stride8_vf8:
; AVX2-FCP:       # %bb.0:
; AVX2-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX2-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX2-FCP-NEXT:    vmovdqa (%rdi), %xmm0
; AVX2-FCP-NEXT:    vmovdqa 16(%rdi), %xmm1
; AVX2-FCP-NEXT:    vmovdqa 32(%rdi), %xmm2
; AVX2-FCP-NEXT:    vmovdqa 48(%rdi), %xmm3
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm4 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX2-FCP-NEXT:    vpshufb %xmm4, %xmm3, %xmm5
; AVX2-FCP-NEXT:    vpshufb %xmm4, %xmm2, %xmm4
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm5 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX2-FCP-NEXT:    vpshufb %xmm5, %xmm1, %xmm6
; AVX2-FCP-NEXT:    vpshufb %xmm5, %xmm0, %xmm5
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm4 = xmm5[0],xmm4[1],xmm5[2,3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm5 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX2-FCP-NEXT:    vpshufb %xmm5, %xmm3, %xmm6
; AVX2-FCP-NEXT:    vpshufb %xmm5, %xmm2, %xmm5
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm6 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX2-FCP-NEXT:    vpshufb %xmm6, %xmm1, %xmm7
; AVX2-FCP-NEXT:    vpshufb %xmm6, %xmm0, %xmm6
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm6[0],xmm7[0],xmm6[1],xmm7[1],xmm6[2],xmm7[2],xmm6[3],xmm7[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm5 = xmm6[0],xmm5[1],xmm6[2,3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm6 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX2-FCP-NEXT:    vpshufb %xmm6, %xmm3, %xmm7
; AVX2-FCP-NEXT:    vpshufb %xmm6, %xmm2, %xmm6
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm6[0],xmm7[0],xmm6[1],xmm7[1],xmm6[2],xmm7[2],xmm6[3],xmm7[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm7 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX2-FCP-NEXT:    vpshufb %xmm7, %xmm1, %xmm8
; AVX2-FCP-NEXT:    vpshufb %xmm7, %xmm0, %xmm7
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm7 = xmm7[0],xmm8[0],xmm7[1],xmm8[1],xmm7[2],xmm8[2],xmm7[3],xmm8[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm6 = xmm7[0],xmm6[1],xmm7[2,3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm7 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX2-FCP-NEXT:    vpshufb %xmm7, %xmm3, %xmm8
; AVX2-FCP-NEXT:    vpshufb %xmm7, %xmm2, %xmm7
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm7 = xmm7[0],xmm8[0],xmm7[1],xmm8[1],xmm7[2],xmm8[2],xmm7[3],xmm8[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm8 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX2-FCP-NEXT:    vpshufb %xmm8, %xmm1, %xmm9
; AVX2-FCP-NEXT:    vpshufb %xmm8, %xmm0, %xmm8
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm9[0],xmm8[1],xmm9[1],xmm8[2],xmm9[2],xmm8[3],xmm9[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm7 = xmm8[0],xmm7[1],xmm8[2,3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm8 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX2-FCP-NEXT:    vpshufb %xmm8, %xmm3, %xmm9
; AVX2-FCP-NEXT:    vpshufb %xmm8, %xmm2, %xmm8
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm9[0],xmm8[1],xmm9[1],xmm8[2],xmm9[2],xmm8[3],xmm9[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm9 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX2-FCP-NEXT:    vpshufb %xmm9, %xmm1, %xmm10
; AVX2-FCP-NEXT:    vpshufb %xmm9, %xmm0, %xmm9
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm9[0],xmm10[0],xmm9[1],xmm10[1],xmm9[2],xmm10[2],xmm9[3],xmm10[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm8 = xmm9[0],xmm8[1],xmm9[2,3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm9 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX2-FCP-NEXT:    vpshufb %xmm9, %xmm3, %xmm10
; AVX2-FCP-NEXT:    vpshufb %xmm9, %xmm2, %xmm9
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm9[0],xmm10[0],xmm9[1],xmm10[1],xmm9[2],xmm10[2],xmm9[3],xmm10[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm10 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX2-FCP-NEXT:    vpshufb %xmm10, %xmm1, %xmm11
; AVX2-FCP-NEXT:    vpshufb %xmm10, %xmm0, %xmm10
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm10[0],xmm11[0],xmm10[1],xmm11[1],xmm10[2],xmm11[2],xmm10[3],xmm11[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm9 = xmm10[0],xmm9[1],xmm10[2,3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm10 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX2-FCP-NEXT:    vpshufb %xmm10, %xmm3, %xmm11
; AVX2-FCP-NEXT:    vpshufb %xmm10, %xmm2, %xmm10
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm10[0],xmm11[0],xmm10[1],xmm11[1],xmm10[2],xmm11[2],xmm10[3],xmm11[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm11 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX2-FCP-NEXT:    vpshufb %xmm11, %xmm1, %xmm12
; AVX2-FCP-NEXT:    vpshufb %xmm11, %xmm0, %xmm11
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm11[0],xmm12[0],xmm11[1],xmm12[1],xmm11[2],xmm12[2],xmm11[3],xmm12[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm10 = xmm11[0],xmm10[1],xmm11[2,3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm11 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX2-FCP-NEXT:    vpshufb %xmm11, %xmm3, %xmm3
; AVX2-FCP-NEXT:    vpshufb %xmm11, %xmm2, %xmm2
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm3 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX2-FCP-NEXT:    vpshufb %xmm3, %xmm1, %xmm1
; AVX2-FCP-NEXT:    vpshufb %xmm3, %xmm0, %xmm0
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm0[0],xmm2[1],xmm0[2,3]
; AVX2-FCP-NEXT:    vmovq %xmm4, (%rsi)
; AVX2-FCP-NEXT:    vmovq %xmm5, (%rdx)
; AVX2-FCP-NEXT:    vmovq %xmm6, (%rcx)
; AVX2-FCP-NEXT:    vmovq %xmm7, (%r8)
; AVX2-FCP-NEXT:    vmovq %xmm8, (%r9)
; AVX2-FCP-NEXT:    vmovq %xmm9, (%r11)
; AVX2-FCP-NEXT:    vmovq %xmm10, (%r10)
; AVX2-FCP-NEXT:    vmovq %xmm0, (%rax)
; AVX2-FCP-NEXT:    retq
;
; AVX512-LABEL: load_i8_stride8_vf8:
; AVX512:       # %bb.0:
; AVX512-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX512-NEXT:    vmovdqa64 (%rdi), %zmm0
; AVX512-NEXT:    vpsrlq $8, %zmm0, %zmm1
; AVX512-NEXT:    vpsrlq $16, %zmm0, %zmm2
; AVX512-NEXT:    vpsrlq $24, %zmm0, %zmm3
; AVX512-NEXT:    vpsrlq $32, %zmm0, %zmm4
; AVX512-NEXT:    vpsrlq $40, %zmm0, %zmm5
; AVX512-NEXT:    vpsrlq $48, %zmm0, %zmm6
; AVX512-NEXT:    vpsrlq $56, %zmm0, %zmm7
; AVX512-NEXT:    vpmovqb %zmm0, (%rsi)
; AVX512-NEXT:    vpmovqb %zmm1, (%rdx)
; AVX512-NEXT:    vpmovqb %zmm2, (%rcx)
; AVX512-NEXT:    vpmovqb %zmm3, (%r8)
; AVX512-NEXT:    vpmovqb %zmm4, (%r9)
; AVX512-NEXT:    vpmovqb %zmm5, (%r11)
; AVX512-NEXT:    vpmovqb %zmm6, (%r10)
; AVX512-NEXT:    vpmovqb %zmm7, (%rax)
; AVX512-NEXT:    vzeroupper
; AVX512-NEXT:    retq
;
; AVX512-FCP-LABEL: load_i8_stride8_vf8:
; AVX512-FCP:       # %bb.0:
; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX512-FCP-NEXT:    vmovdqa64 (%rdi), %zmm0
; AVX512-FCP-NEXT:    vpsrlq $8, %zmm0, %zmm1
; AVX512-FCP-NEXT:    vpsrlq $16, %zmm0, %zmm2
; AVX512-FCP-NEXT:    vpsrlq $24, %zmm0, %zmm3
; AVX512-FCP-NEXT:    vpsrlq $32, %zmm0, %zmm4
; AVX512-FCP-NEXT:    vpsrlq $40, %zmm0, %zmm5
; AVX512-FCP-NEXT:    vpsrlq $48, %zmm0, %zmm6
; AVX512-FCP-NEXT:    vpsrlq $56, %zmm0, %zmm7
; AVX512-FCP-NEXT:    vpmovqb %zmm0, (%rsi)
; AVX512-FCP-NEXT:    vpmovqb %zmm1, (%rdx)
; AVX512-FCP-NEXT:    vpmovqb %zmm2, (%rcx)
; AVX512-FCP-NEXT:    vpmovqb %zmm3, (%r8)
; AVX512-FCP-NEXT:    vpmovqb %zmm4, (%r9)
; AVX512-FCP-NEXT:    vpmovqb %zmm5, (%r11)
; AVX512-FCP-NEXT:    vpmovqb %zmm6, (%r10)
; AVX512-FCP-NEXT:    vpmovqb %zmm7, (%rax)
; AVX512-FCP-NEXT:    vzeroupper
; AVX512-FCP-NEXT:    retq
;
; AVX512DQ-LABEL: load_i8_stride8_vf8:
; AVX512DQ:       # %bb.0:
; AVX512DQ-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512DQ-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX512DQ-NEXT:    vmovdqa64 (%rdi), %zmm0
; AVX512DQ-NEXT:    vpsrlq $8, %zmm0, %zmm1
; AVX512DQ-NEXT:    vpsrlq $16, %zmm0, %zmm2
; AVX512DQ-NEXT:    vpsrlq $24, %zmm0, %zmm3
; AVX512DQ-NEXT:    vpsrlq $32, %zmm0, %zmm4
; AVX512DQ-NEXT:    vpsrlq $40, %zmm0, %zmm5
; AVX512DQ-NEXT:    vpsrlq $48, %zmm0, %zmm6
; AVX512DQ-NEXT:    vpsrlq $56, %zmm0, %zmm7
; AVX512DQ-NEXT:    vpmovqb %zmm0, (%rsi)
; AVX512DQ-NEXT:    vpmovqb %zmm1, (%rdx)
; AVX512DQ-NEXT:    vpmovqb %zmm2, (%rcx)
; AVX512DQ-NEXT:    vpmovqb %zmm3, (%r8)
; AVX512DQ-NEXT:    vpmovqb %zmm4, (%r9)
; AVX512DQ-NEXT:    vpmovqb %zmm5, (%r11)
; AVX512DQ-NEXT:    vpmovqb %zmm6, (%r10)
; AVX512DQ-NEXT:    vpmovqb %zmm7, (%rax)
; AVX512DQ-NEXT:    vzeroupper
; AVX512DQ-NEXT:    retq
;
; AVX512DQ-FCP-LABEL: load_i8_stride8_vf8:
; AVX512DQ-FCP:       # %bb.0:
; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX512DQ-FCP-NEXT:    vmovdqa64 (%rdi), %zmm0
; AVX512DQ-FCP-NEXT:    vpsrlq $8, %zmm0, %zmm1
; AVX512DQ-FCP-NEXT:    vpsrlq $16, %zmm0, %zmm2
; AVX512DQ-FCP-NEXT:    vpsrlq $24, %zmm0, %zmm3
; AVX512DQ-FCP-NEXT:    vpsrlq $32, %zmm0, %zmm4
; AVX512DQ-FCP-NEXT:    vpsrlq $40, %zmm0, %zmm5
; AVX512DQ-FCP-NEXT:    vpsrlq $48, %zmm0, %zmm6
; AVX512DQ-FCP-NEXT:    vpsrlq $56, %zmm0, %zmm7
; AVX512DQ-FCP-NEXT:    vpmovqb %zmm0, (%rsi)
; AVX512DQ-FCP-NEXT:    vpmovqb %zmm1, (%rdx)
; AVX512DQ-FCP-NEXT:    vpmovqb %zmm2, (%rcx)
; AVX512DQ-FCP-NEXT:    vpmovqb %zmm3, (%r8)
; AVX512DQ-FCP-NEXT:    vpmovqb %zmm4, (%r9)
; AVX512DQ-FCP-NEXT:    vpmovqb %zmm5, (%r11)
; AVX512DQ-FCP-NEXT:    vpmovqb %zmm6, (%r10)
; AVX512DQ-FCP-NEXT:    vpmovqb %zmm7, (%rax)
; AVX512DQ-FCP-NEXT:    vzeroupper
; AVX512DQ-FCP-NEXT:    retq
;
; AVX512BW-LABEL: load_i8_stride8_vf8:
; AVX512BW:       # %bb.0:
; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX512BW-NEXT:    vmovdqa64 (%rdi), %zmm0
; AVX512BW-NEXT:    vpsrlq $8, %zmm0, %zmm1
; AVX512BW-NEXT:    vpsrlq $16, %zmm0, %zmm2
; AVX512BW-NEXT:    vpsrlq $24, %zmm0, %zmm3
; AVX512BW-NEXT:    vpsrlq $32, %zmm0, %zmm4
; AVX512BW-NEXT:    vpsrlq $40, %zmm0, %zmm5
; AVX512BW-NEXT:    vpsrlq $48, %zmm0, %zmm6
; AVX512BW-NEXT:    vpsrlq $56, %zmm0, %zmm7
; AVX512BW-NEXT:    vpmovqb %zmm0, (%rsi)
; AVX512BW-NEXT:    vpmovqb %zmm1, (%rdx)
; AVX512BW-NEXT:    vpmovqb %zmm2, (%rcx)
; AVX512BW-NEXT:    vpmovqb %zmm3, (%r8)
; AVX512BW-NEXT:    vpmovqb %zmm4, (%r9)
; AVX512BW-NEXT:    vpmovqb %zmm5, (%r11)
; AVX512BW-NEXT:    vpmovqb %zmm6, (%r10)
; AVX512BW-NEXT:    vpmovqb %zmm7, (%rax)
; AVX512BW-NEXT:    vzeroupper
; AVX512BW-NEXT:    retq
;
; AVX512BW-FCP-LABEL: load_i8_stride8_vf8:
; AVX512BW-FCP:       # %bb.0:
; AVX512BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX512BW-FCP-NEXT:    vmovdqa64 (%rdi), %zmm0
; AVX512BW-FCP-NEXT:    vpsrlq $8, %zmm0, %zmm1
; AVX512BW-FCP-NEXT:    vpsrlq $16, %zmm0, %zmm2
; AVX512BW-FCP-NEXT:    vpsrlq $24, %zmm0, %zmm3
; AVX512BW-FCP-NEXT:    vpsrlq $32, %zmm0, %zmm4
; AVX512BW-FCP-NEXT:    vpsrlq $40, %zmm0, %zmm5
; AVX512BW-FCP-NEXT:    vpsrlq $48, %zmm0, %zmm6
; AVX512BW-FCP-NEXT:    vpsrlq $56, %zmm0, %zmm7
; AVX512BW-FCP-NEXT:    vpmovqb %zmm0, (%rsi)
; AVX512BW-FCP-NEXT:    vpmovqb %zmm1, (%rdx)
; AVX512BW-FCP-NEXT:    vpmovqb %zmm2, (%rcx)
; AVX512BW-FCP-NEXT:    vpmovqb %zmm3, (%r8)
; AVX512BW-FCP-NEXT:    vpmovqb %zmm4, (%r9)
; AVX512BW-FCP-NEXT:    vpmovqb %zmm5, (%r11)
; AVX512BW-FCP-NEXT:    vpmovqb %zmm6, (%r10)
; AVX512BW-FCP-NEXT:    vpmovqb %zmm7, (%rax)
; AVX512BW-FCP-NEXT:    vzeroupper
; AVX512BW-FCP-NEXT:    retq
;
; AVX512DQ-BW-LABEL: load_i8_stride8_vf8:
; AVX512DQ-BW:       # %bb.0:
; AVX512DQ-BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-BW-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512DQ-BW-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX512DQ-BW-NEXT:    vmovdqa64 (%rdi), %zmm0
; AVX512DQ-BW-NEXT:    vpsrlq $8, %zmm0, %zmm1
; AVX512DQ-BW-NEXT:    vpsrlq $16, %zmm0, %zmm2
; AVX512DQ-BW-NEXT:    vpsrlq $24, %zmm0, %zmm3
; AVX512DQ-BW-NEXT:    vpsrlq $32, %zmm0, %zmm4
; AVX512DQ-BW-NEXT:    vpsrlq $40, %zmm0, %zmm5
; AVX512DQ-BW-NEXT:    vpsrlq $48, %zmm0, %zmm6
; AVX512DQ-BW-NEXT:    vpsrlq $56, %zmm0, %zmm7
; AVX512DQ-BW-NEXT:    vpmovqb %zmm0, (%rsi)
; AVX512DQ-BW-NEXT:    vpmovqb %zmm1, (%rdx)
; AVX512DQ-BW-NEXT:    vpmovqb %zmm2, (%rcx)
; AVX512DQ-BW-NEXT:    vpmovqb %zmm3, (%r8)
; AVX512DQ-BW-NEXT:    vpmovqb %zmm4, (%r9)
; AVX512DQ-BW-NEXT:    vpmovqb %zmm5, (%r11)
; AVX512DQ-BW-NEXT:    vpmovqb %zmm6, (%r10)
; AVX512DQ-BW-NEXT:    vpmovqb %zmm7, (%rax)
; AVX512DQ-BW-NEXT:    vzeroupper
; AVX512DQ-BW-NEXT:    retq
;
; AVX512DQ-BW-FCP-LABEL: load_i8_stride8_vf8:
; AVX512DQ-BW-FCP:       # %bb.0:
; AVX512DQ-BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512DQ-BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 (%rdi), %zmm0
; AVX512DQ-BW-FCP-NEXT:    vpsrlq $8, %zmm0, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vpsrlq $16, %zmm0, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vpsrlq $24, %zmm0, %zmm3
; AVX512DQ-BW-FCP-NEXT:    vpsrlq $32, %zmm0, %zmm4
; AVX512DQ-BW-FCP-NEXT:    vpsrlq $40, %zmm0, %zmm5
; AVX512DQ-BW-FCP-NEXT:    vpsrlq $48, %zmm0, %zmm6
; AVX512DQ-BW-FCP-NEXT:    vpsrlq $56, %zmm0, %zmm7
; AVX512DQ-BW-FCP-NEXT:    vpmovqb %zmm0, (%rsi)
; AVX512DQ-BW-FCP-NEXT:    vpmovqb %zmm1, (%rdx)
; AVX512DQ-BW-FCP-NEXT:    vpmovqb %zmm2, (%rcx)
; AVX512DQ-BW-FCP-NEXT:    vpmovqb %zmm3, (%r8)
; AVX512DQ-BW-FCP-NEXT:    vpmovqb %zmm4, (%r9)
; AVX512DQ-BW-FCP-NEXT:    vpmovqb %zmm5, (%r11)
; AVX512DQ-BW-FCP-NEXT:    vpmovqb %zmm6, (%r10)
; AVX512DQ-BW-FCP-NEXT:    vpmovqb %zmm7, (%rax)
; AVX512DQ-BW-FCP-NEXT:    vzeroupper
; AVX512DQ-BW-FCP-NEXT:    retq
  %wide.vec = load <64 x i8>, ptr %in.vec, align 64
  %strided.vec0 = shufflevector <64 x i8> %wide.vec, <64 x i8> poison, <8 x i32> <i32 0, i32 8, i32 16, i32 24, i32 32, i32 40, i32 48, i32 56>
  %strided.vec1 = shufflevector <64 x i8> %wide.vec, <64 x i8> poison, <8 x i32> <i32 1, i32 9, i32 17, i32 25, i32 33, i32 41, i32 49, i32 57>
  %strided.vec2 = shufflevector <64 x i8> %wide.vec, <64 x i8> poison, <8 x i32> <i32 2, i32 10, i32 18, i32 26, i32 34, i32 42, i32 50, i32 58>
  %strided.vec3 = shufflevector <64 x i8> %wide.vec, <64 x i8> poison, <8 x i32> <i32 3, i32 11, i32 19, i32 27, i32 35, i32 43, i32 51, i32 59>
  %strided.vec4 = shufflevector <64 x i8> %wide.vec, <64 x i8> poison, <8 x i32> <i32 4, i32 12, i32 20, i32 28, i32 36, i32 44, i32 52, i32 60>
  %strided.vec5 = shufflevector <64 x i8> %wide.vec, <64 x i8> poison, <8 x i32> <i32 5, i32 13, i32 21, i32 29, i32 37, i32 45, i32 53, i32 61>
  %strided.vec6 = shufflevector <64 x i8> %wide.vec, <64 x i8> poison, <8 x i32> <i32 6, i32 14, i32 22, i32 30, i32 38, i32 46, i32 54, i32 62>
  %strided.vec7 = shufflevector <64 x i8> %wide.vec, <64 x i8> poison, <8 x i32> <i32 7, i32 15, i32 23, i32 31, i32 39, i32 47, i32 55, i32 63>
  store <8 x i8> %strided.vec0, ptr %out.vec0, align 64
  store <8 x i8> %strided.vec1, ptr %out.vec1, align 64
  store <8 x i8> %strided.vec2, ptr %out.vec2, align 64
  store <8 x i8> %strided.vec3, ptr %out.vec3, align 64
  store <8 x i8> %strided.vec4, ptr %out.vec4, align 64
  store <8 x i8> %strided.vec5, ptr %out.vec5, align 64
  store <8 x i8> %strided.vec6, ptr %out.vec6, align 64
  store <8 x i8> %strided.vec7, ptr %out.vec7, align 64
  ret void
}

define void @load_i8_stride8_vf16(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr %out.vec2, ptr %out.vec3, ptr %out.vec4, ptr %out.vec5, ptr %out.vec6, ptr %out.vec7) nounwind {
; SSE-LABEL: load_i8_stride8_vf16:
; SSE:       # %bb.0:
; SSE-NEXT:    subq $328, %rsp # imm = 0x148
; SSE-NEXT:    movdqa (%rdi), %xmm6
; SSE-NEXT:    movdqa 16(%rdi), %xmm14
; SSE-NEXT:    movdqa 32(%rdi), %xmm13
; SSE-NEXT:    movdqa 48(%rdi), %xmm8
; SSE-NEXT:    movdqa 64(%rdi), %xmm15
; SSE-NEXT:    movdqa 80(%rdi), %xmm4
; SSE-NEXT:    movdqa 96(%rdi), %xmm12
; SSE-NEXT:    movdqa 112(%rdi), %xmm9
; SSE-NEXT:    movdqa {{.*#+}} xmm3 = [255,0,0,0,255,0,0,0]
; SSE-NEXT:    movdqa %xmm9, %xmm0
; SSE-NEXT:    pand %xmm3, %xmm0
; SSE-NEXT:    movdqa %xmm12, %xmm1
; SSE-NEXT:    pand %xmm3, %xmm1
; SSE-NEXT:    packuswb %xmm0, %xmm1
; SSE-NEXT:    packuswb %xmm1, %xmm0
; SSE-NEXT:    movdqa %xmm4, %xmm1
; SSE-NEXT:    pand %xmm3, %xmm1
; SSE-NEXT:    movdqa %xmm15, %xmm2
; SSE-NEXT:    pand %xmm3, %xmm2
; SSE-NEXT:    packuswb %xmm1, %xmm2
; SSE-NEXT:    packuswb %xmm2, %xmm2
; SSE-NEXT:    packuswb %xmm0, %xmm2
; SSE-NEXT:    movdqa %xmm8, %xmm0
; SSE-NEXT:    pand %xmm3, %xmm0
; SSE-NEXT:    movdqa %xmm13, %xmm1
; SSE-NEXT:    pand %xmm3, %xmm1
; SSE-NEXT:    packuswb %xmm0, %xmm1
; SSE-NEXT:    packuswb %xmm1, %xmm0
; SSE-NEXT:    movdqa %xmm14, %xmm1
; SSE-NEXT:    pand %xmm3, %xmm1
; SSE-NEXT:    pand %xmm6, %xmm3
; SSE-NEXT:    packuswb %xmm1, %xmm3
; SSE-NEXT:    packuswb %xmm3, %xmm3
; SSE-NEXT:    packuswb %xmm0, %xmm3
; SSE-NEXT:    shufps {{.*#+}} xmm3 = xmm3[0,3],xmm2[0,3]
; SSE-NEXT:    movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pxor %xmm7, %xmm7
; SSE-NEXT:    movdqa %xmm14, %xmm0
; SSE-NEXT:    punpckhbw {{.*#+}} xmm0 = xmm0[8],xmm7[8],xmm0[9],xmm7[9],xmm0[10],xmm7[10],xmm0[11],xmm7[11],xmm0[12],xmm7[12],xmm0[13],xmm7[13],xmm0[14],xmm7[14],xmm0[15],xmm7[15]
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm14, %xmm11
; SSE-NEXT:    punpcklbw {{.*#+}} xmm11 = xmm11[0],xmm7[0],xmm11[1],xmm7[1],xmm11[2],xmm7[2],xmm11[3],xmm7[3],xmm11[4],xmm7[4],xmm11[5],xmm7[5],xmm11[6],xmm7[6],xmm11[7],xmm7[7]
; SSE-NEXT:    movdqa %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklwd {{.*#+}} xmm11 = xmm11[0],xmm0[0],xmm11[1],xmm0[1],xmm11[2],xmm0[2],xmm11[3],xmm0[3]
; SSE-NEXT:    packuswb %xmm11, %xmm11
; SSE-NEXT:    movdqa {{.*#+}} xmm10 = [65535,0,65535,65535,65535,65535,65535,65535]
; SSE-NEXT:    movdqa %xmm10, %xmm1
; SSE-NEXT:    pandn %xmm11, %xmm1
; SSE-NEXT:    movdqa %xmm6, %xmm0
; SSE-NEXT:    movdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm6, %xmm2
; SSE-NEXT:    punpckhbw {{.*#+}} xmm2 = xmm2[8],xmm7[8],xmm2[9],xmm7[9],xmm2[10],xmm7[10],xmm2[11],xmm7[11],xmm2[12],xmm7[12],xmm2[13],xmm7[13],xmm2[14],xmm7[14],xmm2[15],xmm7[15]
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm7[0],xmm0[1],xmm7[1],xmm0[2],xmm7[2],xmm0[3],xmm7[3],xmm0[4],xmm7[4],xmm0[5],xmm7[5],xmm0[6],xmm7[6],xmm0[7],xmm7[7]
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],xmm2[3]
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,1,1]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    pand %xmm10, %xmm0
; SSE-NEXT:    por %xmm1, %xmm0
; SSE-NEXT:    movdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm8, %xmm1
; SSE-NEXT:    punpckhbw {{.*#+}} xmm1 = xmm1[8],xmm7[8],xmm1[9],xmm7[9],xmm1[10],xmm7[10],xmm1[11],xmm7[11],xmm1[12],xmm7[12],xmm1[13],xmm7[13],xmm1[14],xmm7[14],xmm1[15],xmm7[15]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklbw {{.*#+}} xmm8 = xmm8[0],xmm7[0],xmm8[1],xmm7[1],xmm8[2],xmm7[2],xmm8[3],xmm7[3],xmm8[4],xmm7[4],xmm8[5],xmm7[5],xmm8[6],xmm7[6],xmm8[7],xmm7[7]
; SSE-NEXT:    movdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklwd {{.*#+}} xmm8 = xmm8[0],xmm1[0],xmm8[1],xmm1[1],xmm8[2],xmm1[2],xmm8[3],xmm1[3]
; SSE-NEXT:    packuswb %xmm8, %xmm8
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm8[0,0,2,3]
; SSE-NEXT:    movdqa {{.*#+}} xmm5 = [65535,65535,65535,0,65535,65535,65535,65535]
; SSE-NEXT:    movdqa %xmm5, %xmm6
; SSE-NEXT:    pandn %xmm1, %xmm6
; SSE-NEXT:    movdqa %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm13, %xmm1
; SSE-NEXT:    punpckhbw {{.*#+}} xmm1 = xmm1[8],xmm7[8],xmm1[9],xmm7[9],xmm1[10],xmm7[10],xmm1[11],xmm7[11],xmm1[12],xmm7[12],xmm1[13],xmm7[13],xmm1[14],xmm7[14],xmm1[15],xmm7[15]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklbw {{.*#+}} xmm13 = xmm13[0],xmm7[0],xmm13[1],xmm7[1],xmm13[2],xmm7[2],xmm13[3],xmm7[3],xmm13[4],xmm7[4],xmm13[5],xmm7[5],xmm13[6],xmm7[6],xmm13[7],xmm7[7]
; SSE-NEXT:    movdqa %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklwd {{.*#+}} xmm13 = xmm13[0],xmm1[0],xmm13[1],xmm1[1],xmm13[2],xmm1[2],xmm13[3],xmm1[3]
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm13[0,1,1,3]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm5, %xmm1
; SSE-NEXT:    por %xmm6, %xmm1
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,1,1]
; SSE-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
; SSE-NEXT:    movdqa %xmm9, %xmm1
; SSE-NEXT:    movdqa %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm9, %xmm2
; SSE-NEXT:    punpckhbw {{.*#+}} xmm2 = xmm2[8],xmm7[8],xmm2[9],xmm7[9],xmm2[10],xmm7[10],xmm2[11],xmm7[11],xmm2[12],xmm7[12],xmm2[13],xmm7[13],xmm2[14],xmm7[14],xmm2[15],xmm7[15]
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm7[0],xmm1[1],xmm7[1],xmm1[2],xmm7[2],xmm1[3],xmm7[3],xmm1[4],xmm7[4],xmm1[5],xmm7[5],xmm1[6],xmm7[6],xmm1[7],xmm7[7]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[0,1,2,2]
; SSE-NEXT:    movdqa {{.*#+}} xmm6 = [65535,65535,65535,65535,65535,65535,65535,0]
; SSE-NEXT:    movdqa %xmm6, %xmm9
; SSE-NEXT:    pandn %xmm1, %xmm9
; SSE-NEXT:    movdqa %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm12, %xmm1
; SSE-NEXT:    punpckhbw {{.*#+}} xmm1 = xmm1[8],xmm7[8],xmm1[9],xmm7[9],xmm1[10],xmm7[10],xmm1[11],xmm7[11],xmm1[12],xmm7[12],xmm1[13],xmm7[13],xmm1[14],xmm7[14],xmm1[15],xmm7[15]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklbw {{.*#+}} xmm12 = xmm12[0],xmm7[0],xmm12[1],xmm7[1],xmm12[2],xmm7[2],xmm12[3],xmm7[3],xmm12[4],xmm7[4],xmm12[5],xmm7[5],xmm12[6],xmm7[6],xmm12[7],xmm7[7]
; SSE-NEXT:    movdqa %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklwd {{.*#+}} xmm12 = xmm12[0],xmm1[0],xmm12[1],xmm1[1],xmm12[2],xmm1[2],xmm12[3],xmm1[3]
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm12[0,1,1,3]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm6, %xmm1
; SSE-NEXT:    por %xmm9, %xmm1
; SSE-NEXT:    movdqa %xmm4, %xmm3
; SSE-NEXT:    movdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm4, %xmm2
; SSE-NEXT:    punpckhbw {{.*#+}} xmm2 = xmm2[8],xmm7[8],xmm2[9],xmm7[9],xmm2[10],xmm7[10],xmm2[11],xmm7[11],xmm2[12],xmm7[12],xmm2[13],xmm7[13],xmm2[14],xmm7[14],xmm2[15],xmm7[15]
; SSE-NEXT:    movdqa %xmm2, %xmm4
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm3, %xmm2
; SSE-NEXT:    punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm7[0],xmm2[1],xmm7[1],xmm2[2],xmm7[2],xmm2[3],xmm7[3],xmm2[4],xmm7[4],xmm2[5],xmm7[5],xmm2[6],xmm7[6],xmm2[7],xmm7[7]
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm15, %xmm3
; SSE-NEXT:    punpckhbw {{.*#+}} xmm3 = xmm3[8],xmm7[8],xmm3[9],xmm7[9],xmm3[10],xmm7[10],xmm3[11],xmm7[11],xmm3[12],xmm7[12],xmm3[13],xmm7[13],xmm3[14],xmm7[14],xmm3[15],xmm7[15]
; SSE-NEXT:    movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklbw {{.*#+}} xmm15 = xmm15[0],xmm7[0],xmm15[1],xmm7[1],xmm15[2],xmm7[2],xmm15[3],xmm7[3],xmm15[4],xmm7[4],xmm15[5],xmm7[5],xmm15[6],xmm7[6],xmm15[7],xmm7[7]
; SSE-NEXT:    movdqa %xmm15, (%rsp) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm2, %xmm7
; SSE-NEXT:    punpcklwd {{.*#+}} xmm7 = xmm7[0],xmm4[0],xmm7[1],xmm4[1],xmm7[2],xmm4[2],xmm7[3],xmm4[3]
; SSE-NEXT:    packuswb %xmm7, %xmm7
; SSE-NEXT:    movdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa {{.*#+}} xmm4 = [65535,65535,65535,65535,65535,0,65535,65535]
; SSE-NEXT:    movdqa %xmm4, %xmm9
; SSE-NEXT:    pandn %xmm7, %xmm9
; SSE-NEXT:    movdqa %xmm15, %xmm7
; SSE-NEXT:    punpcklwd {{.*#+}} xmm7 = xmm7[0],xmm3[0],xmm7[1],xmm3[1],xmm7[2],xmm3[2],xmm7[3],xmm3[3]
; SSE-NEXT:    pshufd {{.*#+}} xmm15 = xmm7[1,1,1,1]
; SSE-NEXT:    packuswb %xmm15, %xmm15
; SSE-NEXT:    pand %xmm4, %xmm15
; SSE-NEXT:    por %xmm9, %xmm15
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm15[2,2,2,2]
; SSE-NEXT:    punpckhdq {{.*#+}} xmm2 = xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; SSE-NEXT:    movsd {{.*#+}} xmm2 = xmm0[0],xmm2[1]
; SSE-NEXT:    movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa {{.*#+}} xmm1 = [255,255,255,255,255,255,255,255]
; SSE-NEXT:    pand %xmm1, %xmm14
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm14[0,2,2,3]
; SSE-NEXT:    pshuflw {{.*#+}} xmm0 = xmm0[0,1,1,3,4,5,6,7]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    movdqa %xmm10, %xmm9
; SSE-NEXT:    pandn %xmm0, %xmm9
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    pand %xmm1, %xmm0
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
; SSE-NEXT:    pshuflw {{.*#+}} xmm0 = xmm0[1,3,2,3,4,5,6,7]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    pand %xmm10, %xmm0
; SSE-NEXT:    por %xmm9, %xmm0
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; SSE-NEXT:    pand %xmm1, %xmm2
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm9 = xmm2[0,1,2,0]
; SSE-NEXT:    pshufhw {{.*#+}} xmm9 = xmm9[0,1,2,3,4,5,7,5]
; SSE-NEXT:    packuswb %xmm9, %xmm9
; SSE-NEXT:    movdqa %xmm5, %xmm15
; SSE-NEXT:    pandn %xmm9, %xmm15
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; SSE-NEXT:    pand %xmm1, %xmm2
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm9 = xmm2[0,1,2,0]
; SSE-NEXT:    pshufhw {{.*#+}} xmm9 = xmm9[0,1,2,3,7,5,6,7]
; SSE-NEXT:    packuswb %xmm9, %xmm9
; SSE-NEXT:    pand %xmm5, %xmm9
; SSE-NEXT:    por %xmm15, %xmm9
; SSE-NEXT:    pshufd {{.*#+}} xmm9 = xmm9[1,1,1,1]
; SSE-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm9[0],xmm0[1],xmm9[1]
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; SSE-NEXT:    pand %xmm1, %xmm2
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm9 = xmm2[0,1,2,0]
; SSE-NEXT:    pshufhw {{.*#+}} xmm9 = xmm9[0,1,2,3,4,5,7,5]
; SSE-NEXT:    packuswb %xmm9, %xmm9
; SSE-NEXT:    movdqa %xmm6, %xmm15
; SSE-NEXT:    pandn %xmm9, %xmm15
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; SSE-NEXT:    pand %xmm1, %xmm2
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm9 = xmm2[0,1,2,0]
; SSE-NEXT:    pshufhw {{.*#+}} xmm9 = xmm9[0,1,2,3,7,5,6,7]
; SSE-NEXT:    packuswb %xmm9, %xmm9
; SSE-NEXT:    pand %xmm6, %xmm9
; SSE-NEXT:    por %xmm15, %xmm9
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; SSE-NEXT:    pand %xmm1, %xmm2
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm15 = xmm2[0,2,2,3]
; SSE-NEXT:    pshuflw {{.*#+}} xmm15 = xmm15[0,1,1,3,4,5,6,7]
; SSE-NEXT:    packuswb %xmm15, %xmm15
; SSE-NEXT:    movdqa %xmm4, %xmm2
; SSE-NEXT:    pandn %xmm15, %xmm2
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; SSE-NEXT:    pand %xmm1, %xmm15
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm15[0,2,2,3]
; SSE-NEXT:    pshuflw {{.*#+}} xmm1 = xmm1[1,3,2,3,4,5,6,7]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm4, %xmm1
; SSE-NEXT:    por %xmm2, %xmm1
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[2,2,2,2]
; SSE-NEXT:    punpckhdq {{.*#+}} xmm1 = xmm1[2],xmm9[2],xmm1[3],xmm9[3]
; SSE-NEXT:    movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]
; SSE-NEXT:    movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm11[1,1,2,3]
; SSE-NEXT:    movdqa %xmm10, %xmm1
; SSE-NEXT:    pandn %xmm0, %xmm1
; SSE-NEXT:    pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
; SSE-NEXT:    # xmm0 = mem[3,3,3,3]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    pand %xmm10, %xmm0
; SSE-NEXT:    por %xmm1, %xmm0
; SSE-NEXT:    movdqa %xmm5, %xmm1
; SSE-NEXT:    pandn %xmm8, %xmm1
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm13[2,2,3,3]
; SSE-NEXT:    packuswb %xmm2, %xmm2
; SSE-NEXT:    pand %xmm5, %xmm2
; SSE-NEXT:    por %xmm1, %xmm2
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[1,1,1,1]
; SSE-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
; SSE-NEXT:    movdqa %xmm6, %xmm1
; SSE-NEXT:    pandn {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm12[2,2,3,3]
; SSE-NEXT:    packuswb %xmm2, %xmm2
; SSE-NEXT:    pand %xmm6, %xmm2
; SSE-NEXT:    por %xmm1, %xmm2
; SSE-NEXT:    pshufd $244, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[0,1,3,3]
; SSE-NEXT:    movdqa %xmm4, %xmm8
; SSE-NEXT:    pandn %xmm1, %xmm8
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm7[3,3,3,3]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm4, %xmm1
; SSE-NEXT:    por %xmm8, %xmm1
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[2,2,2,2]
; SSE-NEXT:    punpckhdq {{.*#+}} xmm1 = xmm1[2],xmm2[2],xmm1[3],xmm2[3]
; SSE-NEXT:    movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]
; SSE-NEXT:    movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm3 = xmm14[3,1,2,3]
; SSE-NEXT:    pshuflw {{.*#+}} xmm0 = xmm3[0,1,2,0,4,5,6,7]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    movdqa %xmm10, %xmm1
; SSE-NEXT:    pandn %xmm0, %xmm1
; SSE-NEXT:    pshufd $231, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
; SSE-NEXT:    # xmm0 = mem[3,1,2,3]
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshuflw {{.*#+}} xmm0 = xmm0[2,0,2,3,4,5,6,7]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    pand %xmm10, %xmm0
; SSE-NEXT:    por %xmm1, %xmm0
; SSE-NEXT:    pshufd $212, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[0,1,1,3]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,5,4,6]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    movdqa %xmm5, %xmm8
; SSE-NEXT:    pandn %xmm1, %xmm8
; SSE-NEXT:    pshufd $212, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[0,1,1,3]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,6,6,7]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm5, %xmm1
; SSE-NEXT:    por %xmm8, %xmm1
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,1,1]
; SSE-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
; SSE-NEXT:    pshufd $212, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[0,1,1,3]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,5,4,6]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    movdqa %xmm6, %xmm9
; SSE-NEXT:    pandn %xmm1, %xmm9
; SSE-NEXT:    pshufd $212, {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Folded Reload
; SSE-NEXT:    # xmm8 = mem[0,1,1,3]
; SSE-NEXT:    pshufhw {{.*#+}} xmm1 = xmm8[0,1,2,3,4,6,6,7]
; SSE-NEXT:    packuswb %xmm1, %xmm14
; SSE-NEXT:    pand %xmm6, %xmm14
; SSE-NEXT:    por %xmm9, %xmm14
; SSE-NEXT:    pshufd $231, {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Folded Reload
; SSE-NEXT:    # xmm12 = mem[3,1,2,3]
; SSE-NEXT:    pshuflw {{.*#+}} xmm1 = xmm12[0,1,2,0,4,5,6,7]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    movdqa %xmm4, %xmm9
; SSE-NEXT:    pandn %xmm1, %xmm9
; SSE-NEXT:    pshufd {{.*#+}} xmm13 = xmm15[3,1,2,3]
; SSE-NEXT:    pshuflw {{.*#+}} xmm1 = xmm13[2,0,2,3,4,5,6,7]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm4, %xmm1
; SSE-NEXT:    por %xmm9, %xmm1
; SSE-NEXT:    pshufd {{.*#+}} xmm11 = xmm1[2,2,2,2]
; SSE-NEXT:    punpckhdq {{.*#+}} xmm11 = xmm11[2],xmm14[2],xmm11[3],xmm14[3]
; SSE-NEXT:    movsd {{.*#+}} xmm11 = xmm0[0],xmm11[1]
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    punpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
; SSE-NEXT:    # xmm0 = xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7]
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    punpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = xmm1[4],mem[4],xmm1[5],mem[5],xmm1[6],mem[6],xmm1[7],mem[7]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm10, %xmm9
; SSE-NEXT:    pandn %xmm0, %xmm9
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,1,1]
; SSE-NEXT:    packuswb %xmm2, %xmm2
; SSE-NEXT:    pand %xmm10, %xmm2
; SSE-NEXT:    por %xmm9, %xmm2
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    punpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
; SSE-NEXT:    # xmm0 = xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm9 = xmm0[0,0,2,3]
; SSE-NEXT:    movdqa %xmm5, %xmm14
; SSE-NEXT:    pandn %xmm9, %xmm14
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    punpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
; SSE-NEXT:    # xmm0 = xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7]
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm9 = xmm0[0,1,1,3]
; SSE-NEXT:    packuswb %xmm9, %xmm9
; SSE-NEXT:    pand %xmm5, %xmm9
; SSE-NEXT:    por %xmm14, %xmm9
; SSE-NEXT:    pshufd {{.*#+}} xmm9 = xmm9[1,1,1,1]
; SSE-NEXT:    punpckldq {{.*#+}} xmm2 = xmm2[0],xmm9[0],xmm2[1],xmm9[1]
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    punpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
; SSE-NEXT:    # xmm0 = xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7]
; SSE-NEXT:    packuswb %xmm0, %xmm14
; SSE-NEXT:    pshufd {{.*#+}} xmm9 = xmm14[0,1,2,2]
; SSE-NEXT:    movdqa %xmm6, %xmm15
; SSE-NEXT:    pandn %xmm9, %xmm15
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    punpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
; SSE-NEXT:    # xmm0 = xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7]
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm9 = xmm0[0,1,1,3]
; SSE-NEXT:    packuswb %xmm9, %xmm1
; SSE-NEXT:    pand %xmm6, %xmm1
; SSE-NEXT:    por %xmm15, %xmm1
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
; SSE-NEXT:    punpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Folded Reload
; SSE-NEXT:    # xmm7 = xmm7[4],mem[4],xmm7[5],mem[5],xmm7[6],mem[6],xmm7[7],mem[7]
; SSE-NEXT:    movdqa (%rsp), %xmm0 # 16-byte Reload
; SSE-NEXT:    punpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
; SSE-NEXT:    # xmm0 = xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7]
; SSE-NEXT:    movdqa %xmm0, (%rsp) # 16-byte Spill
; SSE-NEXT:    packuswb %xmm7, %xmm9
; SSE-NEXT:    movdqa %xmm4, %xmm15
; SSE-NEXT:    pandn %xmm9, %xmm15
; SSE-NEXT:    pshufd {{.*#+}} xmm7 = xmm0[1,1,1,1]
; SSE-NEXT:    packuswb %xmm7, %xmm7
; SSE-NEXT:    pand %xmm4, %xmm7
; SSE-NEXT:    por %xmm15, %xmm7
; SSE-NEXT:    pshufd {{.*#+}} xmm15 = xmm7[2,2,2,2]
; SSE-NEXT:    punpckhdq {{.*#+}} xmm15 = xmm15[2],xmm1[2],xmm15[3],xmm1[3]
; SSE-NEXT:    movsd {{.*#+}} xmm15 = xmm2[0],xmm15[1]
; SSE-NEXT:    pshuflw {{.*#+}} xmm0 = xmm3[0,1,3,1,4,5,6,7]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    movdqa %xmm10, %xmm1
; SSE-NEXT:    pandn %xmm0, %xmm1
; SSE-NEXT:    pshuflw $231, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
; SSE-NEXT:    # xmm0 = mem[3,1,2,3,4,5,6,7]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    pand %xmm10, %xmm0
; SSE-NEXT:    por %xmm1, %xmm0
; SSE-NEXT:    pshufhw $212, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[0,1,2,3,4,5,5,7]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    movdqa %xmm5, %xmm7
; SSE-NEXT:    pandn %xmm1, %xmm7
; SSE-NEXT:    pshufhw $237, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[0,1,2,3,5,7,6,7]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm5, %xmm1
; SSE-NEXT:    por %xmm7, %xmm1
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,1,1]
; SSE-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
; SSE-NEXT:    pshufhw $212, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[0,1,2,3,4,5,5,7]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    movdqa %xmm6, %xmm7
; SSE-NEXT:    pandn %xmm1, %xmm7
; SSE-NEXT:    pshufhw {{.*#+}} xmm1 = xmm8[0,1,2,3,5,7,6,7]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm6, %xmm1
; SSE-NEXT:    por %xmm7, %xmm1
; SSE-NEXT:    pshuflw {{.*#+}} xmm7 = xmm12[0,1,3,1,4,5,6,7]
; SSE-NEXT:    packuswb %xmm7, %xmm7
; SSE-NEXT:    movdqa %xmm4, %xmm8
; SSE-NEXT:    pandn %xmm7, %xmm8
; SSE-NEXT:    pshuflw {{.*#+}} xmm7 = xmm13[3,1,2,3,4,5,6,7]
; SSE-NEXT:    packuswb %xmm7, %xmm7
; SSE-NEXT:    pand %xmm4, %xmm7
; SSE-NEXT:    por %xmm8, %xmm7
; SSE-NEXT:    pshufd {{.*#+}} xmm7 = xmm7[2,2,2,2]
; SSE-NEXT:    punpckhdq {{.*#+}} xmm7 = xmm7[2],xmm1[2],xmm7[3],xmm1[3]
; SSE-NEXT:    movsd {{.*#+}} xmm7 = xmm0[0],xmm7[1]
; SSE-NEXT:    pshufd $229, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
; SSE-NEXT:    # xmm0 = mem[1,1,2,3]
; SSE-NEXT:    pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[3,3,3,3]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm10, %xmm1
; SSE-NEXT:    pandn %xmm0, %xmm10
; SSE-NEXT:    por %xmm1, %xmm10
; SSE-NEXT:    pshufd $250, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
; SSE-NEXT:    # xmm0 = mem[2,2,3,3]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    pand %xmm5, %xmm0
; SSE-NEXT:    pandn {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Folded Reload
; SSE-NEXT:    por %xmm0, %xmm5
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm5[1,1,1,1]
; SSE-NEXT:    punpckldq {{.*#+}} xmm10 = xmm10[0],xmm0[0],xmm10[1],xmm0[1]
; SSE-NEXT:    pshufd $250, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
; SSE-NEXT:    # xmm0 = mem[2,2,3,3]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    pand %xmm6, %xmm0
; SSE-NEXT:    pandn %xmm14, %xmm6
; SSE-NEXT:    por %xmm0, %xmm6
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm9[0,1,3,3]
; SSE-NEXT:    pshufd $255, (%rsp), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[3,3,3,3]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm4, %xmm1
; SSE-NEXT:    pandn %xmm0, %xmm4
; SSE-NEXT:    por %xmm1, %xmm4
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[2,2,2,2]
; SSE-NEXT:    punpckhdq {{.*#+}} xmm0 = xmm0[2],xmm6[2],xmm0[3],xmm6[3]
; SSE-NEXT:    movsd {{.*#+}} xmm0 = xmm10[0],xmm0[1]
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, (%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, (%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, (%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, (%r8)
; SSE-NEXT:    movapd %xmm11, (%r9)
; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT:    movapd %xmm15, (%rax)
; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT:    movapd %xmm7, (%rax)
; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT:    movapd %xmm0, (%rax)
; SSE-NEXT:    addq $328, %rsp # imm = 0x148
; SSE-NEXT:    retq
;
; AVX-LABEL: load_i8_stride8_vf16:
; AVX:       # %bb.0:
; AVX-NEXT:    vbroadcastss {{.*#+}} xmm0 = [0,0,0,8,0,0,0,8,0,0,0,8,0,0,0,8]
; AVX-NEXT:    vmovdqa 112(%rdi), %xmm8
; AVX-NEXT:    vpshufb %xmm0, %xmm8, %xmm3
; AVX-NEXT:    vmovdqa 96(%rdi), %xmm2
; AVX-NEXT:    vpshufb %xmm0, %xmm2, %xmm0
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3]
; AVX-NEXT:    vbroadcastss {{.*#+}} xmm3 = [0,8,0,0,0,8,0,0,0,8,0,0,0,8,0,0]
; AVX-NEXT:    vmovdqa 80(%rdi), %xmm4
; AVX-NEXT:    vpshufb %xmm3, %xmm4, %xmm6
; AVX-NEXT:    vmovdqa 64(%rdi), %xmm5
; AVX-NEXT:    vpshufb %xmm3, %xmm5, %xmm3
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm6[0],xmm3[1],xmm6[1],xmm3[2],xmm6[2],xmm3[3],xmm6[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm0 = xmm3[0,1,2,3,4,5],xmm0[6,7]
; AVX-NEXT:    vmovd {{.*#+}} xmm9 = [0,0,0,8,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vmovdqa (%rdi), %xmm1
; AVX-NEXT:    vmovdqa 16(%rdi), %xmm3
; AVX-NEXT:    vmovdqa 32(%rdi), %xmm6
; AVX-NEXT:    vmovdqa 48(%rdi), %xmm7
; AVX-NEXT:    vpshufb %xmm9, %xmm7, %xmm10
; AVX-NEXT:    vpshufb %xmm9, %xmm6, %xmm9
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm9[0],xmm10[0],xmm9[1],xmm10[1],xmm9[2],xmm10[2],xmm9[3],xmm10[3]
; AVX-NEXT:    vmovd {{.*#+}} xmm10 = [0,8,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vpshufb %xmm10, %xmm3, %xmm11
; AVX-NEXT:    vpshufb %xmm10, %xmm1, %xmm10
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm10[0],xmm11[0],xmm10[1],xmm11[1],xmm10[2],xmm11[2],xmm10[3],xmm11[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm9 = xmm10[0,1],xmm9[2,3],xmm10[4,5,6,7]
; AVX-NEXT:    vpblendw {{.*#+}} xmm0 = xmm9[0,1,2,3],xmm0[4,5,6,7]
; AVX-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vbroadcastss {{.*#+}} xmm9 = [0,0,1,9,0,0,1,9,0,0,1,9,0,0,1,9]
; AVX-NEXT:    vpshufb %xmm9, %xmm8, %xmm10
; AVX-NEXT:    vpshufb %xmm9, %xmm2, %xmm9
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm9[0],xmm10[0],xmm9[1],xmm10[1],xmm9[2],xmm10[2],xmm9[3],xmm10[3]
; AVX-NEXT:    vbroadcastss {{.*#+}} xmm10 = [1,9,0,0,1,9,0,0,1,9,0,0,1,9,0,0]
; AVX-NEXT:    vpshufb %xmm10, %xmm4, %xmm11
; AVX-NEXT:    vpshufb %xmm10, %xmm5, %xmm10
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm10[0],xmm11[0],xmm10[1],xmm11[1],xmm10[2],xmm11[2],xmm10[3],xmm11[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm9 = xmm10[0,1,2,3,4,5],xmm9[6,7]
; AVX-NEXT:    vmovd {{.*#+}} xmm10 = [0,0,1,9,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vpshufb %xmm10, %xmm7, %xmm11
; AVX-NEXT:    vpshufb %xmm10, %xmm6, %xmm10
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm10[0],xmm11[0],xmm10[1],xmm11[1],xmm10[2],xmm11[2],xmm10[3],xmm11[3]
; AVX-NEXT:    vmovd {{.*#+}} xmm11 = [1,9,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vpshufb %xmm11, %xmm3, %xmm12
; AVX-NEXT:    vpshufb %xmm11, %xmm1, %xmm11
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm11[0],xmm12[0],xmm11[1],xmm12[1],xmm11[2],xmm12[2],xmm11[3],xmm12[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm10 = xmm11[0,1],xmm10[2,3],xmm11[4,5,6,7]
; AVX-NEXT:    vpblendw {{.*#+}} xmm0 = xmm10[0,1,2,3],xmm9[4,5,6,7]
; AVX-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vbroadcastss {{.*#+}} xmm10 = [0,0,2,10,0,0,2,10,0,0,2,10,0,0,2,10]
; AVX-NEXT:    vpshufb %xmm10, %xmm8, %xmm11
; AVX-NEXT:    vpshufb %xmm10, %xmm2, %xmm10
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm10[0],xmm11[0],xmm10[1],xmm11[1],xmm10[2],xmm11[2],xmm10[3],xmm11[3]
; AVX-NEXT:    vbroadcastss {{.*#+}} xmm11 = [2,10,0,0,2,10,0,0,2,10,0,0,2,10,0,0]
; AVX-NEXT:    vpshufb %xmm11, %xmm4, %xmm12
; AVX-NEXT:    vpshufb %xmm11, %xmm5, %xmm11
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm11[0],xmm12[0],xmm11[1],xmm12[1],xmm11[2],xmm12[2],xmm11[3],xmm12[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm10 = xmm11[0,1,2,3,4,5],xmm10[6,7]
; AVX-NEXT:    vmovd {{.*#+}} xmm11 = [0,0,2,10,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vpshufb %xmm11, %xmm7, %xmm12
; AVX-NEXT:    vpshufb %xmm11, %xmm6, %xmm11
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm11[0],xmm12[0],xmm11[1],xmm12[1],xmm11[2],xmm12[2],xmm11[3],xmm12[3]
; AVX-NEXT:    vmovd {{.*#+}} xmm12 = [2,10,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vpshufb %xmm12, %xmm3, %xmm13
; AVX-NEXT:    vpshufb %xmm12, %xmm1, %xmm12
; AVX-NEXT:    vmovdqa %xmm1, %xmm9
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm12[0],xmm13[0],xmm12[1],xmm13[1],xmm12[2],xmm13[2],xmm12[3],xmm13[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm11 = xmm12[0,1],xmm11[2,3],xmm12[4,5,6,7]
; AVX-NEXT:    vpblendw {{.*#+}} xmm10 = xmm11[0,1,2,3],xmm10[4,5,6,7]
; AVX-NEXT:    vbroadcastss {{.*#+}} xmm11 = [0,0,3,11,0,0,3,11,0,0,3,11,0,0,3,11]
; AVX-NEXT:    vpshufb %xmm11, %xmm8, %xmm12
; AVX-NEXT:    vpshufb %xmm11, %xmm2, %xmm11
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm11[0],xmm12[0],xmm11[1],xmm12[1],xmm11[2],xmm12[2],xmm11[3],xmm12[3]
; AVX-NEXT:    vbroadcastss {{.*#+}} xmm12 = [3,11,0,0,3,11,0,0,3,11,0,0,3,11,0,0]
; AVX-NEXT:    vpshufb %xmm12, %xmm4, %xmm13
; AVX-NEXT:    vpshufb %xmm12, %xmm5, %xmm12
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm12[0],xmm13[0],xmm12[1],xmm13[1],xmm12[2],xmm13[2],xmm12[3],xmm13[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm11 = xmm12[0,1,2,3,4,5],xmm11[6,7]
; AVX-NEXT:    vmovd {{.*#+}} xmm12 = [0,0,3,11,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vpshufb %xmm12, %xmm7, %xmm13
; AVX-NEXT:    vpshufb %xmm12, %xmm6, %xmm12
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm12[0],xmm13[0],xmm12[1],xmm13[1],xmm12[2],xmm13[2],xmm12[3],xmm13[3]
; AVX-NEXT:    vmovd {{.*#+}} xmm13 = [3,11,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vpshufb %xmm13, %xmm3, %xmm14
; AVX-NEXT:    vpshufb %xmm13, %xmm1, %xmm13
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm13[0],xmm14[0],xmm13[1],xmm14[1],xmm13[2],xmm14[2],xmm13[3],xmm14[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm12 = xmm13[0,1],xmm12[2,3],xmm13[4,5,6,7]
; AVX-NEXT:    vpblendw {{.*#+}} xmm11 = xmm12[0,1,2,3],xmm11[4,5,6,7]
; AVX-NEXT:    vbroadcastss {{.*#+}} xmm12 = [0,0,4,12,0,0,4,12,0,0,4,12,0,0,4,12]
; AVX-NEXT:    vpshufb %xmm12, %xmm8, %xmm13
; AVX-NEXT:    vpshufb %xmm12, %xmm2, %xmm12
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm12[0],xmm13[0],xmm12[1],xmm13[1],xmm12[2],xmm13[2],xmm12[3],xmm13[3]
; AVX-NEXT:    vbroadcastss {{.*#+}} xmm13 = [4,12,0,0,4,12,0,0,4,12,0,0,4,12,0,0]
; AVX-NEXT:    vpshufb %xmm13, %xmm4, %xmm14
; AVX-NEXT:    vpshufb %xmm13, %xmm5, %xmm13
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm13[0],xmm14[0],xmm13[1],xmm14[1],xmm13[2],xmm14[2],xmm13[3],xmm14[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm12 = xmm13[0,1,2,3,4,5],xmm12[6,7]
; AVX-NEXT:    vmovd {{.*#+}} xmm13 = [0,0,4,12,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vpshufb %xmm13, %xmm7, %xmm14
; AVX-NEXT:    vpshufb %xmm13, %xmm6, %xmm13
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm13[0],xmm14[0],xmm13[1],xmm14[1],xmm13[2],xmm14[2],xmm13[3],xmm14[3]
; AVX-NEXT:    vmovd {{.*#+}} xmm14 = [4,12,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vpshufb %xmm14, %xmm3, %xmm15
; AVX-NEXT:    vpshufb %xmm14, %xmm1, %xmm14
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm14[0],xmm15[0],xmm14[1],xmm15[1],xmm14[2],xmm15[2],xmm14[3],xmm15[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm13 = xmm14[0,1],xmm13[2,3],xmm14[4,5,6,7]
; AVX-NEXT:    vpblendw {{.*#+}} xmm12 = xmm13[0,1,2,3],xmm12[4,5,6,7]
; AVX-NEXT:    vbroadcastss {{.*#+}} xmm13 = [0,0,5,13,0,0,5,13,0,0,5,13,0,0,5,13]
; AVX-NEXT:    vpshufb %xmm13, %xmm8, %xmm14
; AVX-NEXT:    vpshufb %xmm13, %xmm2, %xmm13
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm13[0],xmm14[0],xmm13[1],xmm14[1],xmm13[2],xmm14[2],xmm13[3],xmm14[3]
; AVX-NEXT:    vbroadcastss {{.*#+}} xmm14 = [5,13,0,0,5,13,0,0,5,13,0,0,5,13,0,0]
; AVX-NEXT:    vpshufb %xmm14, %xmm4, %xmm15
; AVX-NEXT:    vpshufb %xmm14, %xmm5, %xmm14
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm14[0],xmm15[0],xmm14[1],xmm15[1],xmm14[2],xmm15[2],xmm14[3],xmm15[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm13 = xmm14[0,1,2,3,4,5],xmm13[6,7]
; AVX-NEXT:    vmovd {{.*#+}} xmm14 = [0,0,5,13,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vpshufb %xmm14, %xmm7, %xmm15
; AVX-NEXT:    vpshufb %xmm14, %xmm6, %xmm14
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm14[0],xmm15[0],xmm14[1],xmm15[1],xmm14[2],xmm15[2],xmm14[3],xmm15[3]
; AVX-NEXT:    vmovd {{.*#+}} xmm15 = [5,13,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vpshufb %xmm15, %xmm3, %xmm0
; AVX-NEXT:    vpshufb %xmm15, %xmm1, %xmm15
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm15[0],xmm0[0],xmm15[1],xmm0[1],xmm15[2],xmm0[2],xmm15[3],xmm0[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0,1],xmm14[2,3],xmm0[4,5,6,7]
; AVX-NEXT:    vpblendw {{.*#+}} xmm13 = xmm0[0,1,2,3],xmm13[4,5,6,7]
; AVX-NEXT:    vbroadcastss {{.*#+}} xmm0 = [0,0,6,14,0,0,6,14,0,0,6,14,0,0,6,14]
; AVX-NEXT:    vpshufb %xmm0, %xmm8, %xmm14
; AVX-NEXT:    vpshufb %xmm0, %xmm2, %xmm0
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm14[0],xmm0[1],xmm14[1],xmm0[2],xmm14[2],xmm0[3],xmm14[3]
; AVX-NEXT:    vbroadcastss {{.*#+}} xmm14 = [6,14,0,0,6,14,0,0,6,14,0,0,6,14,0,0]
; AVX-NEXT:    vpshufb %xmm14, %xmm4, %xmm15
; AVX-NEXT:    vpshufb %xmm14, %xmm5, %xmm14
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm14[0],xmm15[0],xmm14[1],xmm15[1],xmm14[2],xmm15[2],xmm14[3],xmm15[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm1 = xmm14[0,1,2,3,4,5],xmm0[6,7]
; AVX-NEXT:    vmovd {{.*#+}} xmm14 = [0,0,6,14,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vpshufb %xmm14, %xmm7, %xmm15
; AVX-NEXT:    vpshufb %xmm14, %xmm6, %xmm14
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm14[0],xmm15[0],xmm14[1],xmm15[1],xmm14[2],xmm15[2],xmm14[3],xmm15[3]
; AVX-NEXT:    vmovd {{.*#+}} xmm15 = [6,14,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vpshufb %xmm15, %xmm3, %xmm0
; AVX-NEXT:    vpshufb %xmm15, %xmm9, %xmm15
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm15[0],xmm0[0],xmm15[1],xmm0[1],xmm15[2],xmm0[2],xmm15[3],xmm0[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0,1],xmm14[2,3],xmm0[4,5,6,7]
; AVX-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0,1,2,3],xmm1[4,5,6,7]
; AVX-NEXT:    vbroadcastss {{.*#+}} xmm1 = [0,0,7,15,0,0,7,15,0,0,7,15,0,0,7,15]
; AVX-NEXT:    vpshufb %xmm1, %xmm8, %xmm8
; AVX-NEXT:    vpshufb %xmm1, %xmm2, %xmm1
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm8[0],xmm1[1],xmm8[1],xmm1[2],xmm8[2],xmm1[3],xmm8[3]
; AVX-NEXT:    vbroadcastss {{.*#+}} xmm2 = [7,15,0,0,7,15,0,0,7,15,0,0,7,15,0,0]
; AVX-NEXT:    vpshufb %xmm2, %xmm4, %xmm4
; AVX-NEXT:    vpshufb %xmm2, %xmm5, %xmm2
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm1 = xmm2[0,1,2,3,4,5],xmm1[6,7]
; AVX-NEXT:    vmovd {{.*#+}} xmm2 = [0,0,7,15,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vpshufb %xmm2, %xmm7, %xmm4
; AVX-NEXT:    vpshufb %xmm2, %xmm6, %xmm2
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3]
; AVX-NEXT:    vmovd {{.*#+}} xmm4 = [7,15,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vpshufb %xmm4, %xmm3, %xmm3
; AVX-NEXT:    vpshufb %xmm4, %xmm9, %xmm4
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm2 = xmm3[0,1],xmm2[2,3],xmm3[4,5,6,7]
; AVX-NEXT:    vpblendw {{.*#+}} xmm1 = xmm2[0,1,2,3],xmm1[4,5,6,7]
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm2, (%rsi)
; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX-NEXT:    vmovaps %xmm2, (%rdx)
; AVX-NEXT:    vmovdqa %xmm10, (%rcx)
; AVX-NEXT:    vmovdqa %xmm11, (%r8)
; AVX-NEXT:    vmovdqa %xmm12, (%r9)
; AVX-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX-NEXT:    vmovdqa %xmm13, (%rax)
; AVX-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX-NEXT:    vmovdqa %xmm0, (%rax)
; AVX-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX-NEXT:    vmovdqa %xmm1, (%rax)
; AVX-NEXT:    retq
;
; AVX2-LABEL: load_i8_stride8_vf16:
; AVX2:       # %bb.0:
; AVX2-NEXT:    vmovdqa 112(%rdi), %xmm8
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm0 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX2-NEXT:    vpshufb %xmm0, %xmm8, %xmm2
; AVX2-NEXT:    vmovdqa 96(%rdi), %xmm3
; AVX2-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],xmm2[3]
; AVX2-NEXT:    vmovdqa 80(%rdi), %xmm4
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm2 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX2-NEXT:    vpshufb %xmm2, %xmm4, %xmm6
; AVX2-NEXT:    vmovdqa 64(%rdi), %xmm5
; AVX2-NEXT:    vpshufb %xmm2, %xmm5, %xmm2
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm6[0],xmm2[1],xmm6[1],xmm2[2],xmm6[2],xmm2[3],xmm6[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm0 = xmm2[0,1,2],xmm0[3]
; AVX2-NEXT:    vmovdqa (%rdi), %xmm1
; AVX2-NEXT:    vmovdqa 16(%rdi), %xmm2
; AVX2-NEXT:    vmovdqa 32(%rdi), %xmm6
; AVX2-NEXT:    vmovdqa 48(%rdi), %xmm7
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm9 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX2-NEXT:    vpshufb %xmm9, %xmm7, %xmm10
; AVX2-NEXT:    vpshufb %xmm9, %xmm6, %xmm9
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm9[0],xmm10[0],xmm9[1],xmm10[1],xmm9[2],xmm10[2],xmm9[3],xmm10[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm10 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX2-NEXT:    vpshufb %xmm10, %xmm2, %xmm11
; AVX2-NEXT:    vpshufb %xmm10, %xmm1, %xmm10
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm10[0],xmm11[0],xmm10[1],xmm11[1],xmm10[2],xmm11[2],xmm10[3],xmm11[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm9 = xmm10[0],xmm9[1],xmm10[2,3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm0 = xmm9[0,1],xmm0[2,3]
; AVX2-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm9 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX2-NEXT:    vpshufb %xmm9, %xmm8, %xmm10
; AVX2-NEXT:    vpshufb %xmm9, %xmm3, %xmm9
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm9[0],xmm10[0],xmm9[1],xmm10[1],xmm9[2],xmm10[2],xmm9[3],xmm10[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm10 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX2-NEXT:    vpshufb %xmm10, %xmm4, %xmm11
; AVX2-NEXT:    vpshufb %xmm10, %xmm5, %xmm10
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm10[0],xmm11[0],xmm10[1],xmm11[1],xmm10[2],xmm11[2],xmm10[3],xmm11[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm9 = xmm10[0,1,2],xmm9[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm10 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX2-NEXT:    vpshufb %xmm10, %xmm7, %xmm11
; AVX2-NEXT:    vpshufb %xmm10, %xmm6, %xmm10
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm10[0],xmm11[0],xmm10[1],xmm11[1],xmm10[2],xmm11[2],xmm10[3],xmm11[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm11 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX2-NEXT:    vpshufb %xmm11, %xmm2, %xmm12
; AVX2-NEXT:    vpshufb %xmm11, %xmm1, %xmm11
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm11[0],xmm12[0],xmm11[1],xmm12[1],xmm11[2],xmm12[2],xmm11[3],xmm12[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm10 = xmm11[0],xmm10[1],xmm11[2,3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm0 = xmm10[0,1],xmm9[2,3]
; AVX2-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm10 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX2-NEXT:    vpshufb %xmm10, %xmm8, %xmm11
; AVX2-NEXT:    vpshufb %xmm10, %xmm3, %xmm10
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm10[0],xmm11[0],xmm10[1],xmm11[1],xmm10[2],xmm11[2],xmm10[3],xmm11[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm11 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX2-NEXT:    vpshufb %xmm11, %xmm4, %xmm12
; AVX2-NEXT:    vpshufb %xmm11, %xmm5, %xmm11
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm11[0],xmm12[0],xmm11[1],xmm12[1],xmm11[2],xmm12[2],xmm11[3],xmm12[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm10 = xmm11[0,1,2],xmm10[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm11 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX2-NEXT:    vpshufb %xmm11, %xmm7, %xmm12
; AVX2-NEXT:    vpshufb %xmm11, %xmm6, %xmm11
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm11[0],xmm12[0],xmm11[1],xmm12[1],xmm11[2],xmm12[2],xmm11[3],xmm12[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm12 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX2-NEXT:    vpshufb %xmm12, %xmm2, %xmm13
; AVX2-NEXT:    vpshufb %xmm12, %xmm1, %xmm12
; AVX2-NEXT:    vmovdqa %xmm1, %xmm9
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm12[0],xmm13[0],xmm12[1],xmm13[1],xmm12[2],xmm13[2],xmm12[3],xmm13[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm11 = xmm12[0],xmm11[1],xmm12[2,3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm10 = xmm11[0,1],xmm10[2,3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm11 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX2-NEXT:    vpshufb %xmm11, %xmm8, %xmm12
; AVX2-NEXT:    vpshufb %xmm11, %xmm3, %xmm11
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm11[0],xmm12[0],xmm11[1],xmm12[1],xmm11[2],xmm12[2],xmm11[3],xmm12[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm12 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX2-NEXT:    vpshufb %xmm12, %xmm4, %xmm13
; AVX2-NEXT:    vpshufb %xmm12, %xmm5, %xmm12
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm12[0],xmm13[0],xmm12[1],xmm13[1],xmm12[2],xmm13[2],xmm12[3],xmm13[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm11 = xmm12[0,1,2],xmm11[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm12 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX2-NEXT:    vpshufb %xmm12, %xmm7, %xmm13
; AVX2-NEXT:    vpshufb %xmm12, %xmm6, %xmm12
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm12[0],xmm13[0],xmm12[1],xmm13[1],xmm12[2],xmm13[2],xmm12[3],xmm13[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm13 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX2-NEXT:    vpshufb %xmm13, %xmm2, %xmm14
; AVX2-NEXT:    vpshufb %xmm13, %xmm1, %xmm13
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm13[0],xmm14[0],xmm13[1],xmm14[1],xmm13[2],xmm14[2],xmm13[3],xmm14[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm12 = xmm13[0],xmm12[1],xmm13[2,3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm11 = xmm12[0,1],xmm11[2,3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm12 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX2-NEXT:    vpshufb %xmm12, %xmm8, %xmm13
; AVX2-NEXT:    vpshufb %xmm12, %xmm3, %xmm12
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm12[0],xmm13[0],xmm12[1],xmm13[1],xmm12[2],xmm13[2],xmm12[3],xmm13[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm13 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX2-NEXT:    vpshufb %xmm13, %xmm4, %xmm14
; AVX2-NEXT:    vpshufb %xmm13, %xmm5, %xmm13
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm13[0],xmm14[0],xmm13[1],xmm14[1],xmm13[2],xmm14[2],xmm13[3],xmm14[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm12 = xmm13[0,1,2],xmm12[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm13 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX2-NEXT:    vpshufb %xmm13, %xmm7, %xmm14
; AVX2-NEXT:    vpshufb %xmm13, %xmm6, %xmm13
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm13[0],xmm14[0],xmm13[1],xmm14[1],xmm13[2],xmm14[2],xmm13[3],xmm14[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm14 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX2-NEXT:    vpshufb %xmm14, %xmm2, %xmm15
; AVX2-NEXT:    vpshufb %xmm14, %xmm1, %xmm14
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm14[0],xmm15[0],xmm14[1],xmm15[1],xmm14[2],xmm15[2],xmm14[3],xmm15[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm13 = xmm14[0],xmm13[1],xmm14[2,3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm12 = xmm13[0,1],xmm12[2,3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm13 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX2-NEXT:    vpshufb %xmm13, %xmm8, %xmm14
; AVX2-NEXT:    vpshufb %xmm13, %xmm3, %xmm13
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm13[0],xmm14[0],xmm13[1],xmm14[1],xmm13[2],xmm14[2],xmm13[3],xmm14[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm14 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX2-NEXT:    vpshufb %xmm14, %xmm4, %xmm15
; AVX2-NEXT:    vpshufb %xmm14, %xmm5, %xmm14
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm14[0],xmm15[0],xmm14[1],xmm15[1],xmm14[2],xmm15[2],xmm14[3],xmm15[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm13 = xmm14[0,1,2],xmm13[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm14 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX2-NEXT:    vpshufb %xmm14, %xmm7, %xmm15
; AVX2-NEXT:    vpshufb %xmm14, %xmm6, %xmm14
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm14[0],xmm15[0],xmm14[1],xmm15[1],xmm14[2],xmm15[2],xmm14[3],xmm15[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm15 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX2-NEXT:    vpshufb %xmm15, %xmm2, %xmm0
; AVX2-NEXT:    vpshufb %xmm15, %xmm1, %xmm15
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm15[0],xmm0[0],xmm15[1],xmm0[1],xmm15[2],xmm0[2],xmm15[3],xmm0[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm0 = xmm0[0],xmm14[1],xmm0[2,3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm13 = xmm0[0,1],xmm13[2,3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm0 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX2-NEXT:    vpshufb %xmm0, %xmm8, %xmm14
; AVX2-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm14[0],xmm0[1],xmm14[1],xmm0[2],xmm14[2],xmm0[3],xmm14[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm14 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX2-NEXT:    vpshufb %xmm14, %xmm4, %xmm15
; AVX2-NEXT:    vpshufb %xmm14, %xmm5, %xmm14
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm14[0],xmm15[0],xmm14[1],xmm15[1],xmm14[2],xmm15[2],xmm14[3],xmm15[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm1 = xmm14[0,1,2],xmm0[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm14 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX2-NEXT:    vpshufb %xmm14, %xmm7, %xmm15
; AVX2-NEXT:    vpshufb %xmm14, %xmm6, %xmm14
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm14[0],xmm15[0],xmm14[1],xmm15[1],xmm14[2],xmm15[2],xmm14[3],xmm15[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm15 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX2-NEXT:    vpshufb %xmm15, %xmm2, %xmm0
; AVX2-NEXT:    vpshufb %xmm15, %xmm9, %xmm15
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm15[0],xmm0[0],xmm15[1],xmm0[1],xmm15[2],xmm0[2],xmm15[3],xmm0[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm0 = xmm0[0],xmm14[1],xmm0[2,3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm1 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX2-NEXT:    vpshufb %xmm1, %xmm8, %xmm8
; AVX2-NEXT:    vpshufb %xmm1, %xmm3, %xmm1
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm8[0],xmm1[1],xmm8[1],xmm1[2],xmm8[2],xmm1[3],xmm8[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm3 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX2-NEXT:    vpshufb %xmm3, %xmm4, %xmm4
; AVX2-NEXT:    vpshufb %xmm3, %xmm5, %xmm3
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm1 = xmm3[0,1,2],xmm1[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm3 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX2-NEXT:    vpshufb %xmm3, %xmm7, %xmm4
; AVX2-NEXT:    vpshufb %xmm3, %xmm6, %xmm3
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm4 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX2-NEXT:    vpshufb %xmm4, %xmm2, %xmm2
; AVX2-NEXT:    vpshufb %xmm4, %xmm9, %xmm4
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm4[0],xmm2[0],xmm4[1],xmm2[1],xmm4[2],xmm2[2],xmm4[3],xmm2[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm2 = xmm2[0],xmm3[1],xmm2[2,3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3]
; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX2-NEXT:    vmovaps %xmm2, (%rsi)
; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX2-NEXT:    vmovaps %xmm2, (%rdx)
; AVX2-NEXT:    vmovdqa %xmm10, (%rcx)
; AVX2-NEXT:    vmovdqa %xmm11, (%r8)
; AVX2-NEXT:    vmovdqa %xmm12, (%r9)
; AVX2-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-NEXT:    vmovdqa %xmm13, (%rax)
; AVX2-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-NEXT:    vmovdqa %xmm0, (%rax)
; AVX2-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-NEXT:    vmovdqa %xmm1, (%rax)
; AVX2-NEXT:    retq
;
; AVX2-FP-LABEL: load_i8_stride8_vf16:
; AVX2-FP:       # %bb.0:
; AVX2-FP-NEXT:    vmovdqa 112(%rdi), %xmm8
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm0 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm8, %xmm2
; AVX2-FP-NEXT:    vmovdqa 96(%rdi), %xmm3
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],xmm2[3]
; AVX2-FP-NEXT:    vmovdqa 80(%rdi), %xmm4
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm2 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm4, %xmm6
; AVX2-FP-NEXT:    vmovdqa 64(%rdi), %xmm5
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm5, %xmm2
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm6[0],xmm2[1],xmm6[1],xmm2[2],xmm6[2],xmm2[3],xmm6[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm2[0,1,2],xmm0[3]
; AVX2-FP-NEXT:    vmovdqa (%rdi), %xmm1
; AVX2-FP-NEXT:    vmovdqa 16(%rdi), %xmm2
; AVX2-FP-NEXT:    vmovdqa 32(%rdi), %xmm6
; AVX2-FP-NEXT:    vmovdqa 48(%rdi), %xmm7
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm9 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX2-FP-NEXT:    vpshufb %xmm9, %xmm7, %xmm10
; AVX2-FP-NEXT:    vpshufb %xmm9, %xmm6, %xmm9
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm9[0],xmm10[0],xmm9[1],xmm10[1],xmm9[2],xmm10[2],xmm9[3],xmm10[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm10 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX2-FP-NEXT:    vpshufb %xmm10, %xmm2, %xmm11
; AVX2-FP-NEXT:    vpshufb %xmm10, %xmm1, %xmm10
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm10[0],xmm11[0],xmm10[1],xmm11[1],xmm10[2],xmm11[2],xmm10[3],xmm11[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm9 = xmm10[0],xmm9[1],xmm10[2,3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm9[0,1],xmm0[2,3]
; AVX2-FP-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm9 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX2-FP-NEXT:    vpshufb %xmm9, %xmm8, %xmm10
; AVX2-FP-NEXT:    vpshufb %xmm9, %xmm3, %xmm9
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm9[0],xmm10[0],xmm9[1],xmm10[1],xmm9[2],xmm10[2],xmm9[3],xmm10[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm10 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX2-FP-NEXT:    vpshufb %xmm10, %xmm4, %xmm11
; AVX2-FP-NEXT:    vpshufb %xmm10, %xmm5, %xmm10
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm10[0],xmm11[0],xmm10[1],xmm11[1],xmm10[2],xmm11[2],xmm10[3],xmm11[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm9 = xmm10[0,1,2],xmm9[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm10 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX2-FP-NEXT:    vpshufb %xmm10, %xmm7, %xmm11
; AVX2-FP-NEXT:    vpshufb %xmm10, %xmm6, %xmm10
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm10[0],xmm11[0],xmm10[1],xmm11[1],xmm10[2],xmm11[2],xmm10[3],xmm11[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm11 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX2-FP-NEXT:    vpshufb %xmm11, %xmm2, %xmm12
; AVX2-FP-NEXT:    vpshufb %xmm11, %xmm1, %xmm11
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm11[0],xmm12[0],xmm11[1],xmm12[1],xmm11[2],xmm12[2],xmm11[3],xmm12[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm10 = xmm11[0],xmm10[1],xmm11[2,3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm10[0,1],xmm9[2,3]
; AVX2-FP-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm10 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX2-FP-NEXT:    vpshufb %xmm10, %xmm8, %xmm11
; AVX2-FP-NEXT:    vpshufb %xmm10, %xmm3, %xmm10
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm10[0],xmm11[0],xmm10[1],xmm11[1],xmm10[2],xmm11[2],xmm10[3],xmm11[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm11 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX2-FP-NEXT:    vpshufb %xmm11, %xmm4, %xmm12
; AVX2-FP-NEXT:    vpshufb %xmm11, %xmm5, %xmm11
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm11[0],xmm12[0],xmm11[1],xmm12[1],xmm11[2],xmm12[2],xmm11[3],xmm12[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm10 = xmm11[0,1,2],xmm10[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm11 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX2-FP-NEXT:    vpshufb %xmm11, %xmm7, %xmm12
; AVX2-FP-NEXT:    vpshufb %xmm11, %xmm6, %xmm11
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm11[0],xmm12[0],xmm11[1],xmm12[1],xmm11[2],xmm12[2],xmm11[3],xmm12[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm12 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX2-FP-NEXT:    vpshufb %xmm12, %xmm2, %xmm13
; AVX2-FP-NEXT:    vpshufb %xmm12, %xmm1, %xmm12
; AVX2-FP-NEXT:    vmovdqa %xmm1, %xmm9
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm12[0],xmm13[0],xmm12[1],xmm13[1],xmm12[2],xmm13[2],xmm12[3],xmm13[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm11 = xmm12[0],xmm11[1],xmm12[2,3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm10 = xmm11[0,1],xmm10[2,3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm11 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX2-FP-NEXT:    vpshufb %xmm11, %xmm8, %xmm12
; AVX2-FP-NEXT:    vpshufb %xmm11, %xmm3, %xmm11
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm11[0],xmm12[0],xmm11[1],xmm12[1],xmm11[2],xmm12[2],xmm11[3],xmm12[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm12 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX2-FP-NEXT:    vpshufb %xmm12, %xmm4, %xmm13
; AVX2-FP-NEXT:    vpshufb %xmm12, %xmm5, %xmm12
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm12[0],xmm13[0],xmm12[1],xmm13[1],xmm12[2],xmm13[2],xmm12[3],xmm13[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm11 = xmm12[0,1,2],xmm11[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm12 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX2-FP-NEXT:    vpshufb %xmm12, %xmm7, %xmm13
; AVX2-FP-NEXT:    vpshufb %xmm12, %xmm6, %xmm12
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm12[0],xmm13[0],xmm12[1],xmm13[1],xmm12[2],xmm13[2],xmm12[3],xmm13[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm13 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX2-FP-NEXT:    vpshufb %xmm13, %xmm2, %xmm14
; AVX2-FP-NEXT:    vpshufb %xmm13, %xmm1, %xmm13
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm13[0],xmm14[0],xmm13[1],xmm14[1],xmm13[2],xmm14[2],xmm13[3],xmm14[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm12 = xmm13[0],xmm12[1],xmm13[2,3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm11 = xmm12[0,1],xmm11[2,3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm12 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX2-FP-NEXT:    vpshufb %xmm12, %xmm8, %xmm13
; AVX2-FP-NEXT:    vpshufb %xmm12, %xmm3, %xmm12
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm12[0],xmm13[0],xmm12[1],xmm13[1],xmm12[2],xmm13[2],xmm12[3],xmm13[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm13 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX2-FP-NEXT:    vpshufb %xmm13, %xmm4, %xmm14
; AVX2-FP-NEXT:    vpshufb %xmm13, %xmm5, %xmm13
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm13[0],xmm14[0],xmm13[1],xmm14[1],xmm13[2],xmm14[2],xmm13[3],xmm14[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm12 = xmm13[0,1,2],xmm12[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm13 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX2-FP-NEXT:    vpshufb %xmm13, %xmm7, %xmm14
; AVX2-FP-NEXT:    vpshufb %xmm13, %xmm6, %xmm13
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm13[0],xmm14[0],xmm13[1],xmm14[1],xmm13[2],xmm14[2],xmm13[3],xmm14[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm14 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX2-FP-NEXT:    vpshufb %xmm14, %xmm2, %xmm15
; AVX2-FP-NEXT:    vpshufb %xmm14, %xmm1, %xmm14
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm14[0],xmm15[0],xmm14[1],xmm15[1],xmm14[2],xmm15[2],xmm14[3],xmm15[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm13 = xmm14[0],xmm13[1],xmm14[2,3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm12 = xmm13[0,1],xmm12[2,3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm13 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX2-FP-NEXT:    vpshufb %xmm13, %xmm8, %xmm14
; AVX2-FP-NEXT:    vpshufb %xmm13, %xmm3, %xmm13
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm13[0],xmm14[0],xmm13[1],xmm14[1],xmm13[2],xmm14[2],xmm13[3],xmm14[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm14 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX2-FP-NEXT:    vpshufb %xmm14, %xmm4, %xmm15
; AVX2-FP-NEXT:    vpshufb %xmm14, %xmm5, %xmm14
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm14[0],xmm15[0],xmm14[1],xmm15[1],xmm14[2],xmm15[2],xmm14[3],xmm15[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm13 = xmm14[0,1,2],xmm13[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm14 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX2-FP-NEXT:    vpshufb %xmm14, %xmm7, %xmm15
; AVX2-FP-NEXT:    vpshufb %xmm14, %xmm6, %xmm14
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm14[0],xmm15[0],xmm14[1],xmm15[1],xmm14[2],xmm15[2],xmm14[3],xmm15[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm15 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX2-FP-NEXT:    vpshufb %xmm15, %xmm2, %xmm0
; AVX2-FP-NEXT:    vpshufb %xmm15, %xmm1, %xmm15
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm15[0],xmm0[0],xmm15[1],xmm0[1],xmm15[2],xmm0[2],xmm15[3],xmm0[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm0[0],xmm14[1],xmm0[2,3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm13 = xmm0[0,1],xmm13[2,3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm0 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm8, %xmm14
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm14[0],xmm0[1],xmm14[1],xmm0[2],xmm14[2],xmm0[3],xmm14[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm14 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX2-FP-NEXT:    vpshufb %xmm14, %xmm4, %xmm15
; AVX2-FP-NEXT:    vpshufb %xmm14, %xmm5, %xmm14
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm14[0],xmm15[0],xmm14[1],xmm15[1],xmm14[2],xmm15[2],xmm14[3],xmm15[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm1 = xmm14[0,1,2],xmm0[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm14 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX2-FP-NEXT:    vpshufb %xmm14, %xmm7, %xmm15
; AVX2-FP-NEXT:    vpshufb %xmm14, %xmm6, %xmm14
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm14[0],xmm15[0],xmm14[1],xmm15[1],xmm14[2],xmm15[2],xmm14[3],xmm15[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm15 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX2-FP-NEXT:    vpshufb %xmm15, %xmm2, %xmm0
; AVX2-FP-NEXT:    vpshufb %xmm15, %xmm9, %xmm15
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm15[0],xmm0[0],xmm15[1],xmm0[1],xmm15[2],xmm0[2],xmm15[3],xmm0[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm0[0],xmm14[1],xmm0[2,3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm1 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX2-FP-NEXT:    vpshufb %xmm1, %xmm8, %xmm8
; AVX2-FP-NEXT:    vpshufb %xmm1, %xmm3, %xmm1
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm8[0],xmm1[1],xmm8[1],xmm1[2],xmm8[2],xmm1[3],xmm8[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm3 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm4, %xmm4
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm5, %xmm3
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm1 = xmm3[0,1,2],xmm1[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm3 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm7, %xmm4
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm6, %xmm3
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm4 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX2-FP-NEXT:    vpshufb %xmm4, %xmm2, %xmm2
; AVX2-FP-NEXT:    vpshufb %xmm4, %xmm9, %xmm4
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm4[0],xmm2[0],xmm4[1],xmm2[1],xmm4[2],xmm2[2],xmm4[3],xmm2[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm2 = xmm2[0],xmm3[1],xmm2[2,3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3]
; AVX2-FP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX2-FP-NEXT:    vmovaps %xmm2, (%rsi)
; AVX2-FP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX2-FP-NEXT:    vmovaps %xmm2, (%rdx)
; AVX2-FP-NEXT:    vmovdqa %xmm10, (%rcx)
; AVX2-FP-NEXT:    vmovdqa %xmm11, (%r8)
; AVX2-FP-NEXT:    vmovdqa %xmm12, (%r9)
; AVX2-FP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FP-NEXT:    vmovdqa %xmm13, (%rax)
; AVX2-FP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FP-NEXT:    vmovdqa %xmm0, (%rax)
; AVX2-FP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FP-NEXT:    vmovdqa %xmm1, (%rax)
; AVX2-FP-NEXT:    retq
;
; AVX2-FCP-LABEL: load_i8_stride8_vf16:
; AVX2-FCP:       # %bb.0:
; AVX2-FCP-NEXT:    vmovdqa 112(%rdi), %xmm8
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm0 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX2-FCP-NEXT:    vpshufb %xmm0, %xmm8, %xmm2
; AVX2-FCP-NEXT:    vmovdqa 96(%rdi), %xmm3
; AVX2-FCP-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],xmm2[3]
; AVX2-FCP-NEXT:    vmovdqa 80(%rdi), %xmm4
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm2 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX2-FCP-NEXT:    vpshufb %xmm2, %xmm4, %xmm6
; AVX2-FCP-NEXT:    vmovdqa 64(%rdi), %xmm5
; AVX2-FCP-NEXT:    vpshufb %xmm2, %xmm5, %xmm2
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm6[0],xmm2[1],xmm6[1],xmm2[2],xmm6[2],xmm2[3],xmm6[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm2[0,1,2],xmm0[3]
; AVX2-FCP-NEXT:    vmovdqa (%rdi), %xmm1
; AVX2-FCP-NEXT:    vmovdqa 16(%rdi), %xmm2
; AVX2-FCP-NEXT:    vmovdqa 32(%rdi), %xmm6
; AVX2-FCP-NEXT:    vmovdqa 48(%rdi), %xmm7
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm9 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX2-FCP-NEXT:    vpshufb %xmm9, %xmm7, %xmm10
; AVX2-FCP-NEXT:    vpshufb %xmm9, %xmm6, %xmm9
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm9[0],xmm10[0],xmm9[1],xmm10[1],xmm9[2],xmm10[2],xmm9[3],xmm10[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm10 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX2-FCP-NEXT:    vpshufb %xmm10, %xmm2, %xmm11
; AVX2-FCP-NEXT:    vpshufb %xmm10, %xmm1, %xmm10
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm10[0],xmm11[0],xmm10[1],xmm11[1],xmm10[2],xmm11[2],xmm10[3],xmm11[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm9 = xmm10[0],xmm9[1],xmm10[2,3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm9[0,1],xmm0[2,3]
; AVX2-FCP-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm9 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX2-FCP-NEXT:    vpshufb %xmm9, %xmm8, %xmm10
; AVX2-FCP-NEXT:    vpshufb %xmm9, %xmm3, %xmm9
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm9[0],xmm10[0],xmm9[1],xmm10[1],xmm9[2],xmm10[2],xmm9[3],xmm10[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm10 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX2-FCP-NEXT:    vpshufb %xmm10, %xmm4, %xmm11
; AVX2-FCP-NEXT:    vpshufb %xmm10, %xmm5, %xmm10
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm10[0],xmm11[0],xmm10[1],xmm11[1],xmm10[2],xmm11[2],xmm10[3],xmm11[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm9 = xmm10[0,1,2],xmm9[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm10 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX2-FCP-NEXT:    vpshufb %xmm10, %xmm7, %xmm11
; AVX2-FCP-NEXT:    vpshufb %xmm10, %xmm6, %xmm10
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm10[0],xmm11[0],xmm10[1],xmm11[1],xmm10[2],xmm11[2],xmm10[3],xmm11[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm11 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX2-FCP-NEXT:    vpshufb %xmm11, %xmm2, %xmm12
; AVX2-FCP-NEXT:    vpshufb %xmm11, %xmm1, %xmm11
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm11[0],xmm12[0],xmm11[1],xmm12[1],xmm11[2],xmm12[2],xmm11[3],xmm12[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm10 = xmm11[0],xmm10[1],xmm11[2,3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm10[0,1],xmm9[2,3]
; AVX2-FCP-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm10 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX2-FCP-NEXT:    vpshufb %xmm10, %xmm8, %xmm11
; AVX2-FCP-NEXT:    vpshufb %xmm10, %xmm3, %xmm10
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm10[0],xmm11[0],xmm10[1],xmm11[1],xmm10[2],xmm11[2],xmm10[3],xmm11[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm11 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX2-FCP-NEXT:    vpshufb %xmm11, %xmm4, %xmm12
; AVX2-FCP-NEXT:    vpshufb %xmm11, %xmm5, %xmm11
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm11[0],xmm12[0],xmm11[1],xmm12[1],xmm11[2],xmm12[2],xmm11[3],xmm12[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm10 = xmm11[0,1,2],xmm10[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm11 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX2-FCP-NEXT:    vpshufb %xmm11, %xmm7, %xmm12
; AVX2-FCP-NEXT:    vpshufb %xmm11, %xmm6, %xmm11
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm11[0],xmm12[0],xmm11[1],xmm12[1],xmm11[2],xmm12[2],xmm11[3],xmm12[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm12 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX2-FCP-NEXT:    vpshufb %xmm12, %xmm2, %xmm13
; AVX2-FCP-NEXT:    vpshufb %xmm12, %xmm1, %xmm12
; AVX2-FCP-NEXT:    vmovdqa %xmm1, %xmm9
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm12[0],xmm13[0],xmm12[1],xmm13[1],xmm12[2],xmm13[2],xmm12[3],xmm13[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm11 = xmm12[0],xmm11[1],xmm12[2,3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm10 = xmm11[0,1],xmm10[2,3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm11 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX2-FCP-NEXT:    vpshufb %xmm11, %xmm8, %xmm12
; AVX2-FCP-NEXT:    vpshufb %xmm11, %xmm3, %xmm11
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm11[0],xmm12[0],xmm11[1],xmm12[1],xmm11[2],xmm12[2],xmm11[3],xmm12[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm12 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX2-FCP-NEXT:    vpshufb %xmm12, %xmm4, %xmm13
; AVX2-FCP-NEXT:    vpshufb %xmm12, %xmm5, %xmm12
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm12[0],xmm13[0],xmm12[1],xmm13[1],xmm12[2],xmm13[2],xmm12[3],xmm13[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm11 = xmm12[0,1,2],xmm11[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm12 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX2-FCP-NEXT:    vpshufb %xmm12, %xmm7, %xmm13
; AVX2-FCP-NEXT:    vpshufb %xmm12, %xmm6, %xmm12
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm12[0],xmm13[0],xmm12[1],xmm13[1],xmm12[2],xmm13[2],xmm12[3],xmm13[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm13 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX2-FCP-NEXT:    vpshufb %xmm13, %xmm2, %xmm14
; AVX2-FCP-NEXT:    vpshufb %xmm13, %xmm1, %xmm13
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm13[0],xmm14[0],xmm13[1],xmm14[1],xmm13[2],xmm14[2],xmm13[3],xmm14[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm12 = xmm13[0],xmm12[1],xmm13[2,3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm11 = xmm12[0,1],xmm11[2,3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm12 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX2-FCP-NEXT:    vpshufb %xmm12, %xmm8, %xmm13
; AVX2-FCP-NEXT:    vpshufb %xmm12, %xmm3, %xmm12
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm12[0],xmm13[0],xmm12[1],xmm13[1],xmm12[2],xmm13[2],xmm12[3],xmm13[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm13 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX2-FCP-NEXT:    vpshufb %xmm13, %xmm4, %xmm14
; AVX2-FCP-NEXT:    vpshufb %xmm13, %xmm5, %xmm13
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm13[0],xmm14[0],xmm13[1],xmm14[1],xmm13[2],xmm14[2],xmm13[3],xmm14[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm12 = xmm13[0,1,2],xmm12[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm13 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX2-FCP-NEXT:    vpshufb %xmm13, %xmm7, %xmm14
; AVX2-FCP-NEXT:    vpshufb %xmm13, %xmm6, %xmm13
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm13[0],xmm14[0],xmm13[1],xmm14[1],xmm13[2],xmm14[2],xmm13[3],xmm14[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm14 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX2-FCP-NEXT:    vpshufb %xmm14, %xmm2, %xmm15
; AVX2-FCP-NEXT:    vpshufb %xmm14, %xmm1, %xmm14
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm14[0],xmm15[0],xmm14[1],xmm15[1],xmm14[2],xmm15[2],xmm14[3],xmm15[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm13 = xmm14[0],xmm13[1],xmm14[2,3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm12 = xmm13[0,1],xmm12[2,3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm13 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX2-FCP-NEXT:    vpshufb %xmm13, %xmm8, %xmm14
; AVX2-FCP-NEXT:    vpshufb %xmm13, %xmm3, %xmm13
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm13[0],xmm14[0],xmm13[1],xmm14[1],xmm13[2],xmm14[2],xmm13[3],xmm14[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm14 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX2-FCP-NEXT:    vpshufb %xmm14, %xmm4, %xmm15
; AVX2-FCP-NEXT:    vpshufb %xmm14, %xmm5, %xmm14
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm14[0],xmm15[0],xmm14[1],xmm15[1],xmm14[2],xmm15[2],xmm14[3],xmm15[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm13 = xmm14[0,1,2],xmm13[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm14 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX2-FCP-NEXT:    vpshufb %xmm14, %xmm7, %xmm15
; AVX2-FCP-NEXT:    vpshufb %xmm14, %xmm6, %xmm14
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm14[0],xmm15[0],xmm14[1],xmm15[1],xmm14[2],xmm15[2],xmm14[3],xmm15[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm15 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX2-FCP-NEXT:    vpshufb %xmm15, %xmm2, %xmm0
; AVX2-FCP-NEXT:    vpshufb %xmm15, %xmm1, %xmm15
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm15[0],xmm0[0],xmm15[1],xmm0[1],xmm15[2],xmm0[2],xmm15[3],xmm0[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm0[0],xmm14[1],xmm0[2,3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm13 = xmm0[0,1],xmm13[2,3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm0 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX2-FCP-NEXT:    vpshufb %xmm0, %xmm8, %xmm14
; AVX2-FCP-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm14[0],xmm0[1],xmm14[1],xmm0[2],xmm14[2],xmm0[3],xmm14[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm14 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX2-FCP-NEXT:    vpshufb %xmm14, %xmm4, %xmm15
; AVX2-FCP-NEXT:    vpshufb %xmm14, %xmm5, %xmm14
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm14[0],xmm15[0],xmm14[1],xmm15[1],xmm14[2],xmm15[2],xmm14[3],xmm15[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm1 = xmm14[0,1,2],xmm0[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm14 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX2-FCP-NEXT:    vpshufb %xmm14, %xmm7, %xmm15
; AVX2-FCP-NEXT:    vpshufb %xmm14, %xmm6, %xmm14
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm14[0],xmm15[0],xmm14[1],xmm15[1],xmm14[2],xmm15[2],xmm14[3],xmm15[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm15 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX2-FCP-NEXT:    vpshufb %xmm15, %xmm2, %xmm0
; AVX2-FCP-NEXT:    vpshufb %xmm15, %xmm9, %xmm15
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm15[0],xmm0[0],xmm15[1],xmm0[1],xmm15[2],xmm0[2],xmm15[3],xmm0[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm0[0],xmm14[1],xmm0[2,3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm1 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX2-FCP-NEXT:    vpshufb %xmm1, %xmm8, %xmm8
; AVX2-FCP-NEXT:    vpshufb %xmm1, %xmm3, %xmm1
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm8[0],xmm1[1],xmm8[1],xmm1[2],xmm8[2],xmm1[3],xmm8[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm3 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX2-FCP-NEXT:    vpshufb %xmm3, %xmm4, %xmm4
; AVX2-FCP-NEXT:    vpshufb %xmm3, %xmm5, %xmm3
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm1 = xmm3[0,1,2],xmm1[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm3 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX2-FCP-NEXT:    vpshufb %xmm3, %xmm7, %xmm4
; AVX2-FCP-NEXT:    vpshufb %xmm3, %xmm6, %xmm3
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm4 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX2-FCP-NEXT:    vpshufb %xmm4, %xmm2, %xmm2
; AVX2-FCP-NEXT:    vpshufb %xmm4, %xmm9, %xmm4
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm4[0],xmm2[0],xmm4[1],xmm2[1],xmm4[2],xmm2[2],xmm4[3],xmm2[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm2 = xmm2[0],xmm3[1],xmm2[2,3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3]
; AVX2-FCP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX2-FCP-NEXT:    vmovaps %xmm2, (%rsi)
; AVX2-FCP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX2-FCP-NEXT:    vmovaps %xmm2, (%rdx)
; AVX2-FCP-NEXT:    vmovdqa %xmm10, (%rcx)
; AVX2-FCP-NEXT:    vmovdqa %xmm11, (%r8)
; AVX2-FCP-NEXT:    vmovdqa %xmm12, (%r9)
; AVX2-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FCP-NEXT:    vmovdqa %xmm13, (%rax)
; AVX2-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FCP-NEXT:    vmovdqa %xmm0, (%rax)
; AVX2-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FCP-NEXT:    vmovdqa %xmm1, (%rax)
; AVX2-FCP-NEXT:    retq
;
; AVX512-LABEL: load_i8_stride8_vf16:
; AVX512:       # %bb.0:
; AVX512-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX512-NEXT:    vpbroadcastd {{.*#+}} xmm2 = [0,0,0,8,0,0,0,8,0,0,0,8,0,0,0,8]
; AVX512-NEXT:    vmovdqa 112(%rdi), %xmm0
; AVX512-NEXT:    vpshufb %xmm2, %xmm0, %xmm3
; AVX512-NEXT:    vmovdqa 96(%rdi), %xmm1
; AVX512-NEXT:    vpshufb %xmm2, %xmm1, %xmm2
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
; AVX512-NEXT:    vpbroadcastd {{.*#+}} xmm5 = [0,8,0,0,0,8,0,0,0,8,0,0,0,8,0,0]
; AVX512-NEXT:    vmovdqa 80(%rdi), %xmm2
; AVX512-NEXT:    vpshufb %xmm5, %xmm2, %xmm6
; AVX512-NEXT:    vmovdqa 64(%rdi), %xmm3
; AVX512-NEXT:    vpshufb %xmm5, %xmm3, %xmm5
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3]
; AVX512-NEXT:    vpblendd {{.*#+}} xmm4 = xmm5[0,1,2],xmm4[3]
; AVX512-NEXT:    vmovdqa64 (%rdi), %zmm5
; AVX512-NEXT:    vpmovqb %zmm5, %xmm6
; AVX512-NEXT:    vpblendd {{.*#+}} xmm4 = xmm6[0,1],xmm4[2,3]
; AVX512-NEXT:    vpbroadcastd {{.*#+}} xmm6 = [0,0,1,9,0,0,1,9,0,0,1,9,0,0,1,9]
; AVX512-NEXT:    vpshufb %xmm6, %xmm0, %xmm7
; AVX512-NEXT:    vpshufb %xmm6, %xmm1, %xmm6
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm6[0],xmm7[0],xmm6[1],xmm7[1],xmm6[2],xmm7[2],xmm6[3],xmm7[3]
; AVX512-NEXT:    vpbroadcastd {{.*#+}} xmm7 = [1,9,0,0,1,9,0,0,1,9,0,0,1,9,0,0]
; AVX512-NEXT:    vpshufb %xmm7, %xmm2, %xmm8
; AVX512-NEXT:    vpshufb %xmm7, %xmm3, %xmm7
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm7 = xmm7[0],xmm8[0],xmm7[1],xmm8[1],xmm7[2],xmm8[2],xmm7[3],xmm8[3]
; AVX512-NEXT:    vpblendd {{.*#+}} xmm6 = xmm7[0,1,2],xmm6[3]
; AVX512-NEXT:    vpsrlq $8, %zmm5, %zmm7
; AVX512-NEXT:    vpmovqb %zmm7, %xmm7
; AVX512-NEXT:    vpblendd {{.*#+}} xmm6 = xmm7[0,1],xmm6[2,3]
; AVX512-NEXT:    vpbroadcastd {{.*#+}} xmm7 = [0,0,2,10,0,0,2,10,0,0,2,10,0,0,2,10]
; AVX512-NEXT:    vpshufb %xmm7, %xmm0, %xmm8
; AVX512-NEXT:    vpshufb %xmm7, %xmm1, %xmm7
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm7 = xmm7[0],xmm8[0],xmm7[1],xmm8[1],xmm7[2],xmm8[2],xmm7[3],xmm8[3]
; AVX512-NEXT:    vpbroadcastd {{.*#+}} xmm8 = [2,10,0,0,2,10,0,0,2,10,0,0,2,10,0,0]
; AVX512-NEXT:    vpshufb %xmm8, %xmm2, %xmm9
; AVX512-NEXT:    vpshufb %xmm8, %xmm3, %xmm8
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm9[0],xmm8[1],xmm9[1],xmm8[2],xmm9[2],xmm8[3],xmm9[3]
; AVX512-NEXT:    vpblendd {{.*#+}} xmm7 = xmm8[0,1,2],xmm7[3]
; AVX512-NEXT:    vpsrlq $16, %zmm5, %zmm8
; AVX512-NEXT:    vpmovqb %zmm8, %xmm8
; AVX512-NEXT:    vpblendd {{.*#+}} xmm7 = xmm8[0,1],xmm7[2,3]
; AVX512-NEXT:    vpbroadcastd {{.*#+}} xmm8 = [0,0,3,11,0,0,3,11,0,0,3,11,0,0,3,11]
; AVX512-NEXT:    vpshufb %xmm8, %xmm0, %xmm9
; AVX512-NEXT:    vpshufb %xmm8, %xmm1, %xmm8
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm9[0],xmm8[1],xmm9[1],xmm8[2],xmm9[2],xmm8[3],xmm9[3]
; AVX512-NEXT:    vpbroadcastd {{.*#+}} xmm9 = [3,11,0,0,3,11,0,0,3,11,0,0,3,11,0,0]
; AVX512-NEXT:    vpshufb %xmm9, %xmm2, %xmm10
; AVX512-NEXT:    vpshufb %xmm9, %xmm3, %xmm9
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm9[0],xmm10[0],xmm9[1],xmm10[1],xmm9[2],xmm10[2],xmm9[3],xmm10[3]
; AVX512-NEXT:    vpblendd {{.*#+}} xmm8 = xmm9[0,1,2],xmm8[3]
; AVX512-NEXT:    vpsrlq $24, %zmm5, %zmm9
; AVX512-NEXT:    vpmovqb %zmm9, %xmm9
; AVX512-NEXT:    vpblendd {{.*#+}} xmm8 = xmm9[0,1],xmm8[2,3]
; AVX512-NEXT:    vpbroadcastd {{.*#+}} xmm9 = [0,0,4,12,0,0,4,12,0,0,4,12,0,0,4,12]
; AVX512-NEXT:    vpshufb %xmm9, %xmm0, %xmm10
; AVX512-NEXT:    vpshufb %xmm9, %xmm1, %xmm9
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm9[0],xmm10[0],xmm9[1],xmm10[1],xmm9[2],xmm10[2],xmm9[3],xmm10[3]
; AVX512-NEXT:    vpbroadcastd {{.*#+}} xmm10 = [4,12,0,0,4,12,0,0,4,12,0,0,4,12,0,0]
; AVX512-NEXT:    vpshufb %xmm10, %xmm2, %xmm11
; AVX512-NEXT:    vpshufb %xmm10, %xmm3, %xmm10
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm10[0],xmm11[0],xmm10[1],xmm11[1],xmm10[2],xmm11[2],xmm10[3],xmm11[3]
; AVX512-NEXT:    vpblendd {{.*#+}} xmm9 = xmm10[0,1,2],xmm9[3]
; AVX512-NEXT:    vpsrlq $32, %zmm5, %zmm10
; AVX512-NEXT:    vpmovqb %zmm10, %xmm10
; AVX512-NEXT:    vpblendd {{.*#+}} xmm9 = xmm10[0,1],xmm9[2,3]
; AVX512-NEXT:    vpbroadcastd {{.*#+}} xmm10 = [0,0,5,13,0,0,5,13,0,0,5,13,0,0,5,13]
; AVX512-NEXT:    vpshufb %xmm10, %xmm0, %xmm11
; AVX512-NEXT:    vpshufb %xmm10, %xmm1, %xmm10
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm10[0],xmm11[0],xmm10[1],xmm11[1],xmm10[2],xmm11[2],xmm10[3],xmm11[3]
; AVX512-NEXT:    vpbroadcastd {{.*#+}} xmm11 = [5,13,0,0,5,13,0,0,5,13,0,0,5,13,0,0]
; AVX512-NEXT:    vpshufb %xmm11, %xmm2, %xmm12
; AVX512-NEXT:    vpshufb %xmm11, %xmm3, %xmm11
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm11[0],xmm12[0],xmm11[1],xmm12[1],xmm11[2],xmm12[2],xmm11[3],xmm12[3]
; AVX512-NEXT:    vpblendd {{.*#+}} xmm10 = xmm11[0,1,2],xmm10[3]
; AVX512-NEXT:    vpsrlq $40, %zmm5, %zmm11
; AVX512-NEXT:    vpmovqb %zmm11, %xmm11
; AVX512-NEXT:    vpblendd {{.*#+}} xmm10 = xmm11[0,1],xmm10[2,3]
; AVX512-NEXT:    vpbroadcastd {{.*#+}} xmm11 = [0,0,6,14,0,0,6,14,0,0,6,14,0,0,6,14]
; AVX512-NEXT:    vpshufb %xmm11, %xmm0, %xmm12
; AVX512-NEXT:    vpshufb %xmm11, %xmm1, %xmm11
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm11[0],xmm12[0],xmm11[1],xmm12[1],xmm11[2],xmm12[2],xmm11[3],xmm12[3]
; AVX512-NEXT:    vpbroadcastd {{.*#+}} xmm12 = [6,14,0,0,6,14,0,0,6,14,0,0,6,14,0,0]
; AVX512-NEXT:    vpshufb %xmm12, %xmm2, %xmm13
; AVX512-NEXT:    vpshufb %xmm12, %xmm3, %xmm12
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm12[0],xmm13[0],xmm12[1],xmm13[1],xmm12[2],xmm13[2],xmm12[3],xmm13[3]
; AVX512-NEXT:    vpblendd {{.*#+}} xmm11 = xmm12[0,1,2],xmm11[3]
; AVX512-NEXT:    vpsrlq $48, %zmm5, %zmm12
; AVX512-NEXT:    vpmovqb %zmm12, %xmm12
; AVX512-NEXT:    vpblendd {{.*#+}} xmm11 = xmm12[0,1],xmm11[2,3]
; AVX512-NEXT:    vpbroadcastd {{.*#+}} xmm12 = [0,0,7,15,0,0,7,15,0,0,7,15,0,0,7,15]
; AVX512-NEXT:    vpshufb %xmm12, %xmm0, %xmm0
; AVX512-NEXT:    vpshufb %xmm12, %xmm1, %xmm1
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
; AVX512-NEXT:    vpbroadcastd {{.*#+}} xmm1 = [7,15,0,0,7,15,0,0,7,15,0,0,7,15,0,0]
; AVX512-NEXT:    vpshufb %xmm1, %xmm2, %xmm2
; AVX512-NEXT:    vpshufb %xmm1, %xmm3, %xmm1
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
; AVX512-NEXT:    vpblendd {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[3]
; AVX512-NEXT:    vpsrlq $56, %zmm5, %zmm1
; AVX512-NEXT:    vpmovqb %zmm1, %xmm1
; AVX512-NEXT:    vpblendd {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3]
; AVX512-NEXT:    vmovdqa %xmm4, (%rsi)
; AVX512-NEXT:    vmovdqa %xmm6, (%rdx)
; AVX512-NEXT:    vmovdqa %xmm7, (%rcx)
; AVX512-NEXT:    vmovdqa %xmm8, (%r8)
; AVX512-NEXT:    vmovdqa %xmm9, (%r9)
; AVX512-NEXT:    vmovdqa %xmm10, (%r11)
; AVX512-NEXT:    vmovdqa %xmm11, (%r10)
; AVX512-NEXT:    vmovdqa %xmm0, (%rax)
; AVX512-NEXT:    vzeroupper
; AVX512-NEXT:    retq
;
; AVX512-FCP-LABEL: load_i8_stride8_vf16:
; AVX512-FCP:       # %bb.0:
; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm2 = [0,0,0,8,0,0,0,8,0,0,0,8,0,0,0,8]
; AVX512-FCP-NEXT:    vmovdqa 112(%rdi), %xmm0
; AVX512-FCP-NEXT:    vpshufb %xmm2, %xmm0, %xmm3
; AVX512-FCP-NEXT:    vmovdqa 96(%rdi), %xmm1
; AVX512-FCP-NEXT:    vpshufb %xmm2, %xmm1, %xmm2
; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm5 = [0,8,0,0,0,8,0,0,0,8,0,0,0,8,0,0]
; AVX512-FCP-NEXT:    vmovdqa 80(%rdi), %xmm2
; AVX512-FCP-NEXT:    vpshufb %xmm5, %xmm2, %xmm6
; AVX512-FCP-NEXT:    vmovdqa 64(%rdi), %xmm3
; AVX512-FCP-NEXT:    vpshufb %xmm5, %xmm3, %xmm5
; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} xmm4 = xmm5[0,1,2],xmm4[3]
; AVX512-FCP-NEXT:    vmovdqa64 (%rdi), %zmm5
; AVX512-FCP-NEXT:    vpmovqb %zmm5, %xmm6
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} xmm4 = xmm6[0,1],xmm4[2,3]
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm6 = [0,0,1,9,0,0,1,9,0,0,1,9,0,0,1,9]
; AVX512-FCP-NEXT:    vpshufb %xmm6, %xmm0, %xmm7
; AVX512-FCP-NEXT:    vpshufb %xmm6, %xmm1, %xmm6
; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm6[0],xmm7[0],xmm6[1],xmm7[1],xmm6[2],xmm7[2],xmm6[3],xmm7[3]
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm7 = [1,9,0,0,1,9,0,0,1,9,0,0,1,9,0,0]
; AVX512-FCP-NEXT:    vpshufb %xmm7, %xmm2, %xmm8
; AVX512-FCP-NEXT:    vpshufb %xmm7, %xmm3, %xmm7
; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm7 = xmm7[0],xmm8[0],xmm7[1],xmm8[1],xmm7[2],xmm8[2],xmm7[3],xmm8[3]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} xmm6 = xmm7[0,1,2],xmm6[3]
; AVX512-FCP-NEXT:    vpsrlq $8, %zmm5, %zmm7
; AVX512-FCP-NEXT:    vpmovqb %zmm7, %xmm7
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} xmm6 = xmm7[0,1],xmm6[2,3]
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm7 = [0,0,2,10,0,0,2,10,0,0,2,10,0,0,2,10]
; AVX512-FCP-NEXT:    vpshufb %xmm7, %xmm0, %xmm8
; AVX512-FCP-NEXT:    vpshufb %xmm7, %xmm1, %xmm7
; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm7 = xmm7[0],xmm8[0],xmm7[1],xmm8[1],xmm7[2],xmm8[2],xmm7[3],xmm8[3]
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm8 = [2,10,0,0,2,10,0,0,2,10,0,0,2,10,0,0]
; AVX512-FCP-NEXT:    vpshufb %xmm8, %xmm2, %xmm9
; AVX512-FCP-NEXT:    vpshufb %xmm8, %xmm3, %xmm8
; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm9[0],xmm8[1],xmm9[1],xmm8[2],xmm9[2],xmm8[3],xmm9[3]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} xmm7 = xmm8[0,1,2],xmm7[3]
; AVX512-FCP-NEXT:    vpsrlq $16, %zmm5, %zmm8
; AVX512-FCP-NEXT:    vpmovqb %zmm8, %xmm8
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} xmm7 = xmm8[0,1],xmm7[2,3]
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm8 = [0,0,3,11,0,0,3,11,0,0,3,11,0,0,3,11]
; AVX512-FCP-NEXT:    vpshufb %xmm8, %xmm0, %xmm9
; AVX512-FCP-NEXT:    vpshufb %xmm8, %xmm1, %xmm8
; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm9[0],xmm8[1],xmm9[1],xmm8[2],xmm9[2],xmm8[3],xmm9[3]
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm9 = [3,11,0,0,3,11,0,0,3,11,0,0,3,11,0,0]
; AVX512-FCP-NEXT:    vpshufb %xmm9, %xmm2, %xmm10
; AVX512-FCP-NEXT:    vpshufb %xmm9, %xmm3, %xmm9
; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm9[0],xmm10[0],xmm9[1],xmm10[1],xmm9[2],xmm10[2],xmm9[3],xmm10[3]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} xmm8 = xmm9[0,1,2],xmm8[3]
; AVX512-FCP-NEXT:    vpsrlq $24, %zmm5, %zmm9
; AVX512-FCP-NEXT:    vpmovqb %zmm9, %xmm9
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} xmm8 = xmm9[0,1],xmm8[2,3]
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm9 = [0,0,4,12,0,0,4,12,0,0,4,12,0,0,4,12]
; AVX512-FCP-NEXT:    vpshufb %xmm9, %xmm0, %xmm10
; AVX512-FCP-NEXT:    vpshufb %xmm9, %xmm1, %xmm9
; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm9[0],xmm10[0],xmm9[1],xmm10[1],xmm9[2],xmm10[2],xmm9[3],xmm10[3]
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm10 = [4,12,0,0,4,12,0,0,4,12,0,0,4,12,0,0]
; AVX512-FCP-NEXT:    vpshufb %xmm10, %xmm2, %xmm11
; AVX512-FCP-NEXT:    vpshufb %xmm10, %xmm3, %xmm10
; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm10[0],xmm11[0],xmm10[1],xmm11[1],xmm10[2],xmm11[2],xmm10[3],xmm11[3]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} xmm9 = xmm10[0,1,2],xmm9[3]
; AVX512-FCP-NEXT:    vpsrlq $32, %zmm5, %zmm10
; AVX512-FCP-NEXT:    vpmovqb %zmm10, %xmm10
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} xmm9 = xmm10[0,1],xmm9[2,3]
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm10 = [0,0,5,13,0,0,5,13,0,0,5,13,0,0,5,13]
; AVX512-FCP-NEXT:    vpshufb %xmm10, %xmm0, %xmm11
; AVX512-FCP-NEXT:    vpshufb %xmm10, %xmm1, %xmm10
; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm10[0],xmm11[0],xmm10[1],xmm11[1],xmm10[2],xmm11[2],xmm10[3],xmm11[3]
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm11 = [5,13,0,0,5,13,0,0,5,13,0,0,5,13,0,0]
; AVX512-FCP-NEXT:    vpshufb %xmm11, %xmm2, %xmm12
; AVX512-FCP-NEXT:    vpshufb %xmm11, %xmm3, %xmm11
; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm11[0],xmm12[0],xmm11[1],xmm12[1],xmm11[2],xmm12[2],xmm11[3],xmm12[3]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} xmm10 = xmm11[0,1,2],xmm10[3]
; AVX512-FCP-NEXT:    vpsrlq $40, %zmm5, %zmm11
; AVX512-FCP-NEXT:    vpmovqb %zmm11, %xmm11
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} xmm10 = xmm11[0,1],xmm10[2,3]
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm11 = [0,0,6,14,0,0,6,14,0,0,6,14,0,0,6,14]
; AVX512-FCP-NEXT:    vpshufb %xmm11, %xmm0, %xmm12
; AVX512-FCP-NEXT:    vpshufb %xmm11, %xmm1, %xmm11
; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm11[0],xmm12[0],xmm11[1],xmm12[1],xmm11[2],xmm12[2],xmm11[3],xmm12[3]
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm12 = [6,14,0,0,6,14,0,0,6,14,0,0,6,14,0,0]
; AVX512-FCP-NEXT:    vpshufb %xmm12, %xmm2, %xmm13
; AVX512-FCP-NEXT:    vpshufb %xmm12, %xmm3, %xmm12
; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm12[0],xmm13[0],xmm12[1],xmm13[1],xmm12[2],xmm13[2],xmm12[3],xmm13[3]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} xmm11 = xmm12[0,1,2],xmm11[3]
; AVX512-FCP-NEXT:    vpsrlq $48, %zmm5, %zmm12
; AVX512-FCP-NEXT:    vpmovqb %zmm12, %xmm12
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} xmm11 = xmm12[0,1],xmm11[2,3]
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm12 = [0,0,7,15,0,0,7,15,0,0,7,15,0,0,7,15]
; AVX512-FCP-NEXT:    vpshufb %xmm12, %xmm0, %xmm0
; AVX512-FCP-NEXT:    vpshufb %xmm12, %xmm1, %xmm1
; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm1 = [7,15,0,0,7,15,0,0,7,15,0,0,7,15,0,0]
; AVX512-FCP-NEXT:    vpshufb %xmm1, %xmm2, %xmm2
; AVX512-FCP-NEXT:    vpshufb %xmm1, %xmm3, %xmm1
; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[3]
; AVX512-FCP-NEXT:    vpsrlq $56, %zmm5, %zmm1
; AVX512-FCP-NEXT:    vpmovqb %zmm1, %xmm1
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3]
; AVX512-FCP-NEXT:    vmovdqa %xmm4, (%rsi)
; AVX512-FCP-NEXT:    vmovdqa %xmm6, (%rdx)
; AVX512-FCP-NEXT:    vmovdqa %xmm7, (%rcx)
; AVX512-FCP-NEXT:    vmovdqa %xmm8, (%r8)
; AVX512-FCP-NEXT:    vmovdqa %xmm9, (%r9)
; AVX512-FCP-NEXT:    vmovdqa %xmm10, (%r11)
; AVX512-FCP-NEXT:    vmovdqa %xmm11, (%r10)
; AVX512-FCP-NEXT:    vmovdqa %xmm0, (%rax)
; AVX512-FCP-NEXT:    vzeroupper
; AVX512-FCP-NEXT:    retq
;
; AVX512DQ-LABEL: load_i8_stride8_vf16:
; AVX512DQ:       # %bb.0:
; AVX512DQ-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512DQ-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX512DQ-NEXT:    vpbroadcastd {{.*#+}} xmm2 = [0,0,0,8,0,0,0,8,0,0,0,8,0,0,0,8]
; AVX512DQ-NEXT:    vmovdqa 112(%rdi), %xmm0
; AVX512DQ-NEXT:    vpshufb %xmm2, %xmm0, %xmm3
; AVX512DQ-NEXT:    vmovdqa 96(%rdi), %xmm1
; AVX512DQ-NEXT:    vpshufb %xmm2, %xmm1, %xmm2
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
; AVX512DQ-NEXT:    vpbroadcastd {{.*#+}} xmm5 = [0,8,0,0,0,8,0,0,0,8,0,0,0,8,0,0]
; AVX512DQ-NEXT:    vmovdqa 80(%rdi), %xmm2
; AVX512DQ-NEXT:    vpshufb %xmm5, %xmm2, %xmm6
; AVX512DQ-NEXT:    vmovdqa 64(%rdi), %xmm3
; AVX512DQ-NEXT:    vpshufb %xmm5, %xmm3, %xmm5
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} xmm4 = xmm5[0,1,2],xmm4[3]
; AVX512DQ-NEXT:    vmovdqa64 (%rdi), %zmm5
; AVX512DQ-NEXT:    vpmovqb %zmm5, %xmm6
; AVX512DQ-NEXT:    vpblendd {{.*#+}} xmm4 = xmm6[0,1],xmm4[2,3]
; AVX512DQ-NEXT:    vpbroadcastd {{.*#+}} xmm6 = [0,0,1,9,0,0,1,9,0,0,1,9,0,0,1,9]
; AVX512DQ-NEXT:    vpshufb %xmm6, %xmm0, %xmm7
; AVX512DQ-NEXT:    vpshufb %xmm6, %xmm1, %xmm6
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm6[0],xmm7[0],xmm6[1],xmm7[1],xmm6[2],xmm7[2],xmm6[3],xmm7[3]
; AVX512DQ-NEXT:    vpbroadcastd {{.*#+}} xmm7 = [1,9,0,0,1,9,0,0,1,9,0,0,1,9,0,0]
; AVX512DQ-NEXT:    vpshufb %xmm7, %xmm2, %xmm8
; AVX512DQ-NEXT:    vpshufb %xmm7, %xmm3, %xmm7
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm7 = xmm7[0],xmm8[0],xmm7[1],xmm8[1],xmm7[2],xmm8[2],xmm7[3],xmm8[3]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} xmm6 = xmm7[0,1,2],xmm6[3]
; AVX512DQ-NEXT:    vpsrlq $8, %zmm5, %zmm7
; AVX512DQ-NEXT:    vpmovqb %zmm7, %xmm7
; AVX512DQ-NEXT:    vpblendd {{.*#+}} xmm6 = xmm7[0,1],xmm6[2,3]
; AVX512DQ-NEXT:    vpbroadcastd {{.*#+}} xmm7 = [0,0,2,10,0,0,2,10,0,0,2,10,0,0,2,10]
; AVX512DQ-NEXT:    vpshufb %xmm7, %xmm0, %xmm8
; AVX512DQ-NEXT:    vpshufb %xmm7, %xmm1, %xmm7
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm7 = xmm7[0],xmm8[0],xmm7[1],xmm8[1],xmm7[2],xmm8[2],xmm7[3],xmm8[3]
; AVX512DQ-NEXT:    vpbroadcastd {{.*#+}} xmm8 = [2,10,0,0,2,10,0,0,2,10,0,0,2,10,0,0]
; AVX512DQ-NEXT:    vpshufb %xmm8, %xmm2, %xmm9
; AVX512DQ-NEXT:    vpshufb %xmm8, %xmm3, %xmm8
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm9[0],xmm8[1],xmm9[1],xmm8[2],xmm9[2],xmm8[3],xmm9[3]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} xmm7 = xmm8[0,1,2],xmm7[3]
; AVX512DQ-NEXT:    vpsrlq $16, %zmm5, %zmm8
; AVX512DQ-NEXT:    vpmovqb %zmm8, %xmm8
; AVX512DQ-NEXT:    vpblendd {{.*#+}} xmm7 = xmm8[0,1],xmm7[2,3]
; AVX512DQ-NEXT:    vpbroadcastd {{.*#+}} xmm8 = [0,0,3,11,0,0,3,11,0,0,3,11,0,0,3,11]
; AVX512DQ-NEXT:    vpshufb %xmm8, %xmm0, %xmm9
; AVX512DQ-NEXT:    vpshufb %xmm8, %xmm1, %xmm8
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm9[0],xmm8[1],xmm9[1],xmm8[2],xmm9[2],xmm8[3],xmm9[3]
; AVX512DQ-NEXT:    vpbroadcastd {{.*#+}} xmm9 = [3,11,0,0,3,11,0,0,3,11,0,0,3,11,0,0]
; AVX512DQ-NEXT:    vpshufb %xmm9, %xmm2, %xmm10
; AVX512DQ-NEXT:    vpshufb %xmm9, %xmm3, %xmm9
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm9[0],xmm10[0],xmm9[1],xmm10[1],xmm9[2],xmm10[2],xmm9[3],xmm10[3]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} xmm8 = xmm9[0,1,2],xmm8[3]
; AVX512DQ-NEXT:    vpsrlq $24, %zmm5, %zmm9
; AVX512DQ-NEXT:    vpmovqb %zmm9, %xmm9
; AVX512DQ-NEXT:    vpblendd {{.*#+}} xmm8 = xmm9[0,1],xmm8[2,3]
; AVX512DQ-NEXT:    vpbroadcastd {{.*#+}} xmm9 = [0,0,4,12,0,0,4,12,0,0,4,12,0,0,4,12]
; AVX512DQ-NEXT:    vpshufb %xmm9, %xmm0, %xmm10
; AVX512DQ-NEXT:    vpshufb %xmm9, %xmm1, %xmm9
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm9[0],xmm10[0],xmm9[1],xmm10[1],xmm9[2],xmm10[2],xmm9[3],xmm10[3]
; AVX512DQ-NEXT:    vpbroadcastd {{.*#+}} xmm10 = [4,12,0,0,4,12,0,0,4,12,0,0,4,12,0,0]
; AVX512DQ-NEXT:    vpshufb %xmm10, %xmm2, %xmm11
; AVX512DQ-NEXT:    vpshufb %xmm10, %xmm3, %xmm10
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm10[0],xmm11[0],xmm10[1],xmm11[1],xmm10[2],xmm11[2],xmm10[3],xmm11[3]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} xmm9 = xmm10[0,1,2],xmm9[3]
; AVX512DQ-NEXT:    vpsrlq $32, %zmm5, %zmm10
; AVX512DQ-NEXT:    vpmovqb %zmm10, %xmm10
; AVX512DQ-NEXT:    vpblendd {{.*#+}} xmm9 = xmm10[0,1],xmm9[2,3]
; AVX512DQ-NEXT:    vpbroadcastd {{.*#+}} xmm10 = [0,0,5,13,0,0,5,13,0,0,5,13,0,0,5,13]
; AVX512DQ-NEXT:    vpshufb %xmm10, %xmm0, %xmm11
; AVX512DQ-NEXT:    vpshufb %xmm10, %xmm1, %xmm10
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm10[0],xmm11[0],xmm10[1],xmm11[1],xmm10[2],xmm11[2],xmm10[3],xmm11[3]
; AVX512DQ-NEXT:    vpbroadcastd {{.*#+}} xmm11 = [5,13,0,0,5,13,0,0,5,13,0,0,5,13,0,0]
; AVX512DQ-NEXT:    vpshufb %xmm11, %xmm2, %xmm12
; AVX512DQ-NEXT:    vpshufb %xmm11, %xmm3, %xmm11
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm11[0],xmm12[0],xmm11[1],xmm12[1],xmm11[2],xmm12[2],xmm11[3],xmm12[3]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} xmm10 = xmm11[0,1,2],xmm10[3]
; AVX512DQ-NEXT:    vpsrlq $40, %zmm5, %zmm11
; AVX512DQ-NEXT:    vpmovqb %zmm11, %xmm11
; AVX512DQ-NEXT:    vpblendd {{.*#+}} xmm10 = xmm11[0,1],xmm10[2,3]
; AVX512DQ-NEXT:    vpbroadcastd {{.*#+}} xmm11 = [0,0,6,14,0,0,6,14,0,0,6,14,0,0,6,14]
; AVX512DQ-NEXT:    vpshufb %xmm11, %xmm0, %xmm12
; AVX512DQ-NEXT:    vpshufb %xmm11, %xmm1, %xmm11
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm11[0],xmm12[0],xmm11[1],xmm12[1],xmm11[2],xmm12[2],xmm11[3],xmm12[3]
; AVX512DQ-NEXT:    vpbroadcastd {{.*#+}} xmm12 = [6,14,0,0,6,14,0,0,6,14,0,0,6,14,0,0]
; AVX512DQ-NEXT:    vpshufb %xmm12, %xmm2, %xmm13
; AVX512DQ-NEXT:    vpshufb %xmm12, %xmm3, %xmm12
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm12[0],xmm13[0],xmm12[1],xmm13[1],xmm12[2],xmm13[2],xmm12[3],xmm13[3]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} xmm11 = xmm12[0,1,2],xmm11[3]
; AVX512DQ-NEXT:    vpsrlq $48, %zmm5, %zmm12
; AVX512DQ-NEXT:    vpmovqb %zmm12, %xmm12
; AVX512DQ-NEXT:    vpblendd {{.*#+}} xmm11 = xmm12[0,1],xmm11[2,3]
; AVX512DQ-NEXT:    vpbroadcastd {{.*#+}} xmm12 = [0,0,7,15,0,0,7,15,0,0,7,15,0,0,7,15]
; AVX512DQ-NEXT:    vpshufb %xmm12, %xmm0, %xmm0
; AVX512DQ-NEXT:    vpshufb %xmm12, %xmm1, %xmm1
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
; AVX512DQ-NEXT:    vpbroadcastd {{.*#+}} xmm1 = [7,15,0,0,7,15,0,0,7,15,0,0,7,15,0,0]
; AVX512DQ-NEXT:    vpshufb %xmm1, %xmm2, %xmm2
; AVX512DQ-NEXT:    vpshufb %xmm1, %xmm3, %xmm1
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[3]
; AVX512DQ-NEXT:    vpsrlq $56, %zmm5, %zmm1
; AVX512DQ-NEXT:    vpmovqb %zmm1, %xmm1
; AVX512DQ-NEXT:    vpblendd {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3]
; AVX512DQ-NEXT:    vmovdqa %xmm4, (%rsi)
; AVX512DQ-NEXT:    vmovdqa %xmm6, (%rdx)
; AVX512DQ-NEXT:    vmovdqa %xmm7, (%rcx)
; AVX512DQ-NEXT:    vmovdqa %xmm8, (%r8)
; AVX512DQ-NEXT:    vmovdqa %xmm9, (%r9)
; AVX512DQ-NEXT:    vmovdqa %xmm10, (%r11)
; AVX512DQ-NEXT:    vmovdqa %xmm11, (%r10)
; AVX512DQ-NEXT:    vmovdqa %xmm0, (%rax)
; AVX512DQ-NEXT:    vzeroupper
; AVX512DQ-NEXT:    retq
;
; AVX512DQ-FCP-LABEL: load_i8_stride8_vf16:
; AVX512DQ-FCP:       # %bb.0:
; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm2 = [0,0,0,8,0,0,0,8,0,0,0,8,0,0,0,8]
; AVX512DQ-FCP-NEXT:    vmovdqa 112(%rdi), %xmm0
; AVX512DQ-FCP-NEXT:    vpshufb %xmm2, %xmm0, %xmm3
; AVX512DQ-FCP-NEXT:    vmovdqa 96(%rdi), %xmm1
; AVX512DQ-FCP-NEXT:    vpshufb %xmm2, %xmm1, %xmm2
; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm5 = [0,8,0,0,0,8,0,0,0,8,0,0,0,8,0,0]
; AVX512DQ-FCP-NEXT:    vmovdqa 80(%rdi), %xmm2
; AVX512DQ-FCP-NEXT:    vpshufb %xmm5, %xmm2, %xmm6
; AVX512DQ-FCP-NEXT:    vmovdqa 64(%rdi), %xmm3
; AVX512DQ-FCP-NEXT:    vpshufb %xmm5, %xmm3, %xmm5
; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} xmm4 = xmm5[0,1,2],xmm4[3]
; AVX512DQ-FCP-NEXT:    vmovdqa64 (%rdi), %zmm5
; AVX512DQ-FCP-NEXT:    vpmovqb %zmm5, %xmm6
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} xmm4 = xmm6[0,1],xmm4[2,3]
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm6 = [0,0,1,9,0,0,1,9,0,0,1,9,0,0,1,9]
; AVX512DQ-FCP-NEXT:    vpshufb %xmm6, %xmm0, %xmm7
; AVX512DQ-FCP-NEXT:    vpshufb %xmm6, %xmm1, %xmm6
; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm6[0],xmm7[0],xmm6[1],xmm7[1],xmm6[2],xmm7[2],xmm6[3],xmm7[3]
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm7 = [1,9,0,0,1,9,0,0,1,9,0,0,1,9,0,0]
; AVX512DQ-FCP-NEXT:    vpshufb %xmm7, %xmm2, %xmm8
; AVX512DQ-FCP-NEXT:    vpshufb %xmm7, %xmm3, %xmm7
; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm7 = xmm7[0],xmm8[0],xmm7[1],xmm8[1],xmm7[2],xmm8[2],xmm7[3],xmm8[3]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} xmm6 = xmm7[0,1,2],xmm6[3]
; AVX512DQ-FCP-NEXT:    vpsrlq $8, %zmm5, %zmm7
; AVX512DQ-FCP-NEXT:    vpmovqb %zmm7, %xmm7
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} xmm6 = xmm7[0,1],xmm6[2,3]
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm7 = [0,0,2,10,0,0,2,10,0,0,2,10,0,0,2,10]
; AVX512DQ-FCP-NEXT:    vpshufb %xmm7, %xmm0, %xmm8
; AVX512DQ-FCP-NEXT:    vpshufb %xmm7, %xmm1, %xmm7
; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm7 = xmm7[0],xmm8[0],xmm7[1],xmm8[1],xmm7[2],xmm8[2],xmm7[3],xmm8[3]
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm8 = [2,10,0,0,2,10,0,0,2,10,0,0,2,10,0,0]
; AVX512DQ-FCP-NEXT:    vpshufb %xmm8, %xmm2, %xmm9
; AVX512DQ-FCP-NEXT:    vpshufb %xmm8, %xmm3, %xmm8
; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm9[0],xmm8[1],xmm9[1],xmm8[2],xmm9[2],xmm8[3],xmm9[3]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} xmm7 = xmm8[0,1,2],xmm7[3]
; AVX512DQ-FCP-NEXT:    vpsrlq $16, %zmm5, %zmm8
; AVX512DQ-FCP-NEXT:    vpmovqb %zmm8, %xmm8
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} xmm7 = xmm8[0,1],xmm7[2,3]
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm8 = [0,0,3,11,0,0,3,11,0,0,3,11,0,0,3,11]
; AVX512DQ-FCP-NEXT:    vpshufb %xmm8, %xmm0, %xmm9
; AVX512DQ-FCP-NEXT:    vpshufb %xmm8, %xmm1, %xmm8
; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm9[0],xmm8[1],xmm9[1],xmm8[2],xmm9[2],xmm8[3],xmm9[3]
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm9 = [3,11,0,0,3,11,0,0,3,11,0,0,3,11,0,0]
; AVX512DQ-FCP-NEXT:    vpshufb %xmm9, %xmm2, %xmm10
; AVX512DQ-FCP-NEXT:    vpshufb %xmm9, %xmm3, %xmm9
; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm9[0],xmm10[0],xmm9[1],xmm10[1],xmm9[2],xmm10[2],xmm9[3],xmm10[3]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} xmm8 = xmm9[0,1,2],xmm8[3]
; AVX512DQ-FCP-NEXT:    vpsrlq $24, %zmm5, %zmm9
; AVX512DQ-FCP-NEXT:    vpmovqb %zmm9, %xmm9
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} xmm8 = xmm9[0,1],xmm8[2,3]
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm9 = [0,0,4,12,0,0,4,12,0,0,4,12,0,0,4,12]
; AVX512DQ-FCP-NEXT:    vpshufb %xmm9, %xmm0, %xmm10
; AVX512DQ-FCP-NEXT:    vpshufb %xmm9, %xmm1, %xmm9
; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm9[0],xmm10[0],xmm9[1],xmm10[1],xmm9[2],xmm10[2],xmm9[3],xmm10[3]
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm10 = [4,12,0,0,4,12,0,0,4,12,0,0,4,12,0,0]
; AVX512DQ-FCP-NEXT:    vpshufb %xmm10, %xmm2, %xmm11
; AVX512DQ-FCP-NEXT:    vpshufb %xmm10, %xmm3, %xmm10
; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm10[0],xmm11[0],xmm10[1],xmm11[1],xmm10[2],xmm11[2],xmm10[3],xmm11[3]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} xmm9 = xmm10[0,1,2],xmm9[3]
; AVX512DQ-FCP-NEXT:    vpsrlq $32, %zmm5, %zmm10
; AVX512DQ-FCP-NEXT:    vpmovqb %zmm10, %xmm10
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} xmm9 = xmm10[0,1],xmm9[2,3]
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm10 = [0,0,5,13,0,0,5,13,0,0,5,13,0,0,5,13]
; AVX512DQ-FCP-NEXT:    vpshufb %xmm10, %xmm0, %xmm11
; AVX512DQ-FCP-NEXT:    vpshufb %xmm10, %xmm1, %xmm10
; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm10[0],xmm11[0],xmm10[1],xmm11[1],xmm10[2],xmm11[2],xmm10[3],xmm11[3]
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm11 = [5,13,0,0,5,13,0,0,5,13,0,0,5,13,0,0]
; AVX512DQ-FCP-NEXT:    vpshufb %xmm11, %xmm2, %xmm12
; AVX512DQ-FCP-NEXT:    vpshufb %xmm11, %xmm3, %xmm11
; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm11[0],xmm12[0],xmm11[1],xmm12[1],xmm11[2],xmm12[2],xmm11[3],xmm12[3]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} xmm10 = xmm11[0,1,2],xmm10[3]
; AVX512DQ-FCP-NEXT:    vpsrlq $40, %zmm5, %zmm11
; AVX512DQ-FCP-NEXT:    vpmovqb %zmm11, %xmm11
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} xmm10 = xmm11[0,1],xmm10[2,3]
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm11 = [0,0,6,14,0,0,6,14,0,0,6,14,0,0,6,14]
; AVX512DQ-FCP-NEXT:    vpshufb %xmm11, %xmm0, %xmm12
; AVX512DQ-FCP-NEXT:    vpshufb %xmm11, %xmm1, %xmm11
; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm11[0],xmm12[0],xmm11[1],xmm12[1],xmm11[2],xmm12[2],xmm11[3],xmm12[3]
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm12 = [6,14,0,0,6,14,0,0,6,14,0,0,6,14,0,0]
; AVX512DQ-FCP-NEXT:    vpshufb %xmm12, %xmm2, %xmm13
; AVX512DQ-FCP-NEXT:    vpshufb %xmm12, %xmm3, %xmm12
; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm12[0],xmm13[0],xmm12[1],xmm13[1],xmm12[2],xmm13[2],xmm12[3],xmm13[3]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} xmm11 = xmm12[0,1,2],xmm11[3]
; AVX512DQ-FCP-NEXT:    vpsrlq $48, %zmm5, %zmm12
; AVX512DQ-FCP-NEXT:    vpmovqb %zmm12, %xmm12
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} xmm11 = xmm12[0,1],xmm11[2,3]
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm12 = [0,0,7,15,0,0,7,15,0,0,7,15,0,0,7,15]
; AVX512DQ-FCP-NEXT:    vpshufb %xmm12, %xmm0, %xmm0
; AVX512DQ-FCP-NEXT:    vpshufb %xmm12, %xmm1, %xmm1
; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm1 = [7,15,0,0,7,15,0,0,7,15,0,0,7,15,0,0]
; AVX512DQ-FCP-NEXT:    vpshufb %xmm1, %xmm2, %xmm2
; AVX512DQ-FCP-NEXT:    vpshufb %xmm1, %xmm3, %xmm1
; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[3]
; AVX512DQ-FCP-NEXT:    vpsrlq $56, %zmm5, %zmm1
; AVX512DQ-FCP-NEXT:    vpmovqb %zmm1, %xmm1
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3]
; AVX512DQ-FCP-NEXT:    vmovdqa %xmm4, (%rsi)
; AVX512DQ-FCP-NEXT:    vmovdqa %xmm6, (%rdx)
; AVX512DQ-FCP-NEXT:    vmovdqa %xmm7, (%rcx)
; AVX512DQ-FCP-NEXT:    vmovdqa %xmm8, (%r8)
; AVX512DQ-FCP-NEXT:    vmovdqa %xmm9, (%r9)
; AVX512DQ-FCP-NEXT:    vmovdqa %xmm10, (%r11)
; AVX512DQ-FCP-NEXT:    vmovdqa %xmm11, (%r10)
; AVX512DQ-FCP-NEXT:    vmovdqa %xmm0, (%rax)
; AVX512DQ-FCP-NEXT:    vzeroupper
; AVX512DQ-FCP-NEXT:    retq
;
; AVX512BW-LABEL: load_i8_stride8_vf16:
; AVX512BW:       # %bb.0:
; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm2 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX512BW-NEXT:    vmovdqa 112(%rdi), %xmm0
; AVX512BW-NEXT:    vpshufb %xmm2, %xmm0, %xmm3
; AVX512BW-NEXT:    vmovdqa 96(%rdi), %xmm1
; AVX512BW-NEXT:    vpshufb %xmm2, %xmm1, %xmm2
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm5 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX512BW-NEXT:    vmovdqa 80(%rdi), %xmm2
; AVX512BW-NEXT:    vpshufb %xmm5, %xmm2, %xmm6
; AVX512BW-NEXT:    vmovdqa 64(%rdi), %xmm3
; AVX512BW-NEXT:    vpshufb %xmm5, %xmm3, %xmm5
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3]
; AVX512BW-NEXT:    vpblendd {{.*#+}} xmm4 = xmm5[0,1,2],xmm4[3]
; AVX512BW-NEXT:    vmovdqa64 (%rdi), %zmm5
; AVX512BW-NEXT:    vpmovqb %zmm5, %xmm6
; AVX512BW-NEXT:    vpblendd {{.*#+}} xmm4 = xmm6[0,1],xmm4[2,3]
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm6 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX512BW-NEXT:    vpshufb %xmm6, %xmm0, %xmm7
; AVX512BW-NEXT:    vpshufb %xmm6, %xmm1, %xmm6
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm6[0],xmm7[0],xmm6[1],xmm7[1],xmm6[2],xmm7[2],xmm6[3],xmm7[3]
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm7 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX512BW-NEXT:    vpshufb %xmm7, %xmm2, %xmm8
; AVX512BW-NEXT:    vpshufb %xmm7, %xmm3, %xmm7
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm7 = xmm7[0],xmm8[0],xmm7[1],xmm8[1],xmm7[2],xmm8[2],xmm7[3],xmm8[3]
; AVX512BW-NEXT:    vpblendd {{.*#+}} xmm6 = xmm7[0,1,2],xmm6[3]
; AVX512BW-NEXT:    vpsrlq $8, %zmm5, %zmm7
; AVX512BW-NEXT:    vpmovqb %zmm7, %xmm7
; AVX512BW-NEXT:    vpblendd {{.*#+}} xmm6 = xmm7[0,1],xmm6[2,3]
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm7 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX512BW-NEXT:    vpshufb %xmm7, %xmm0, %xmm8
; AVX512BW-NEXT:    vpshufb %xmm7, %xmm1, %xmm7
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm7 = xmm7[0],xmm8[0],xmm7[1],xmm8[1],xmm7[2],xmm8[2],xmm7[3],xmm8[3]
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm8 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX512BW-NEXT:    vpshufb %xmm8, %xmm2, %xmm9
; AVX512BW-NEXT:    vpshufb %xmm8, %xmm3, %xmm8
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm9[0],xmm8[1],xmm9[1],xmm8[2],xmm9[2],xmm8[3],xmm9[3]
; AVX512BW-NEXT:    vpblendd {{.*#+}} xmm7 = xmm8[0,1,2],xmm7[3]
; AVX512BW-NEXT:    vpsrlq $16, %zmm5, %zmm8
; AVX512BW-NEXT:    vpmovqb %zmm8, %xmm8
; AVX512BW-NEXT:    vpblendd {{.*#+}} xmm7 = xmm8[0,1],xmm7[2,3]
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm8 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX512BW-NEXT:    vpshufb %xmm8, %xmm0, %xmm9
; AVX512BW-NEXT:    vpshufb %xmm8, %xmm1, %xmm8
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm9[0],xmm8[1],xmm9[1],xmm8[2],xmm9[2],xmm8[3],xmm9[3]
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm9 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX512BW-NEXT:    vpshufb %xmm9, %xmm2, %xmm10
; AVX512BW-NEXT:    vpshufb %xmm9, %xmm3, %xmm9
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm9[0],xmm10[0],xmm9[1],xmm10[1],xmm9[2],xmm10[2],xmm9[3],xmm10[3]
; AVX512BW-NEXT:    vpblendd {{.*#+}} xmm8 = xmm9[0,1,2],xmm8[3]
; AVX512BW-NEXT:    vpsrlq $24, %zmm5, %zmm9
; AVX512BW-NEXT:    vpmovqb %zmm9, %xmm9
; AVX512BW-NEXT:    vpblendd {{.*#+}} xmm8 = xmm9[0,1],xmm8[2,3]
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm9 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX512BW-NEXT:    vpshufb %xmm9, %xmm0, %xmm10
; AVX512BW-NEXT:    vpshufb %xmm9, %xmm1, %xmm9
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm9[0],xmm10[0],xmm9[1],xmm10[1],xmm9[2],xmm10[2],xmm9[3],xmm10[3]
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm10 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX512BW-NEXT:    vpshufb %xmm10, %xmm2, %xmm11
; AVX512BW-NEXT:    vpshufb %xmm10, %xmm3, %xmm10
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm10[0],xmm11[0],xmm10[1],xmm11[1],xmm10[2],xmm11[2],xmm10[3],xmm11[3]
; AVX512BW-NEXT:    vpblendd {{.*#+}} xmm9 = xmm10[0,1,2],xmm9[3]
; AVX512BW-NEXT:    vpsrlq $32, %zmm5, %zmm10
; AVX512BW-NEXT:    vpmovqb %zmm10, %xmm10
; AVX512BW-NEXT:    vpblendd {{.*#+}} xmm9 = xmm10[0,1],xmm9[2,3]
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm10 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX512BW-NEXT:    vpshufb %xmm10, %xmm0, %xmm11
; AVX512BW-NEXT:    vpshufb %xmm10, %xmm1, %xmm10
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm10[0],xmm11[0],xmm10[1],xmm11[1],xmm10[2],xmm11[2],xmm10[3],xmm11[3]
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm11 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX512BW-NEXT:    vpshufb %xmm11, %xmm2, %xmm12
; AVX512BW-NEXT:    vpshufb %xmm11, %xmm3, %xmm11
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm11[0],xmm12[0],xmm11[1],xmm12[1],xmm11[2],xmm12[2],xmm11[3],xmm12[3]
; AVX512BW-NEXT:    vpblendd {{.*#+}} xmm10 = xmm11[0,1,2],xmm10[3]
; AVX512BW-NEXT:    vpsrlq $40, %zmm5, %zmm11
; AVX512BW-NEXT:    vpmovqb %zmm11, %xmm11
; AVX512BW-NEXT:    vpblendd {{.*#+}} xmm10 = xmm11[0,1],xmm10[2,3]
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm11 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX512BW-NEXT:    vpshufb %xmm11, %xmm0, %xmm12
; AVX512BW-NEXT:    vpshufb %xmm11, %xmm1, %xmm11
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm11[0],xmm12[0],xmm11[1],xmm12[1],xmm11[2],xmm12[2],xmm11[3],xmm12[3]
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm12 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX512BW-NEXT:    vpshufb %xmm12, %xmm2, %xmm13
; AVX512BW-NEXT:    vpshufb %xmm12, %xmm3, %xmm12
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm12[0],xmm13[0],xmm12[1],xmm13[1],xmm12[2],xmm13[2],xmm12[3],xmm13[3]
; AVX512BW-NEXT:    vpblendd {{.*#+}} xmm11 = xmm12[0,1,2],xmm11[3]
; AVX512BW-NEXT:    vpsrlq $48, %zmm5, %zmm12
; AVX512BW-NEXT:    vpmovqb %zmm12, %xmm12
; AVX512BW-NEXT:    vpblendd {{.*#+}} xmm11 = xmm12[0,1],xmm11[2,3]
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm12 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX512BW-NEXT:    vpshufb %xmm12, %xmm0, %xmm0
; AVX512BW-NEXT:    vpshufb %xmm12, %xmm1, %xmm1
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm1 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX512BW-NEXT:    vpshufb %xmm1, %xmm2, %xmm2
; AVX512BW-NEXT:    vpshufb %xmm1, %xmm3, %xmm1
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
; AVX512BW-NEXT:    vpblendd {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[3]
; AVX512BW-NEXT:    vpsrlq $56, %zmm5, %zmm1
; AVX512BW-NEXT:    vpmovqb %zmm1, %xmm1
; AVX512BW-NEXT:    vpblendd {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3]
; AVX512BW-NEXT:    vmovdqa %xmm4, (%rsi)
; AVX512BW-NEXT:    vmovdqa %xmm6, (%rdx)
; AVX512BW-NEXT:    vmovdqa %xmm7, (%rcx)
; AVX512BW-NEXT:    vmovdqa %xmm8, (%r8)
; AVX512BW-NEXT:    vmovdqa %xmm9, (%r9)
; AVX512BW-NEXT:    vmovdqa %xmm10, (%r11)
; AVX512BW-NEXT:    vmovdqa %xmm11, (%r10)
; AVX512BW-NEXT:    vmovdqa %xmm0, (%rax)
; AVX512BW-NEXT:    vzeroupper
; AVX512BW-NEXT:    retq
;
; AVX512BW-FCP-LABEL: load_i8_stride8_vf16:
; AVX512BW-FCP:       # %bb.0:
; AVX512BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX512BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm2 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX512BW-FCP-NEXT:    vmovdqa 112(%rdi), %xmm0
; AVX512BW-FCP-NEXT:    vpshufb %xmm2, %xmm0, %xmm3
; AVX512BW-FCP-NEXT:    vmovdqa 96(%rdi), %xmm1
; AVX512BW-FCP-NEXT:    vpshufb %xmm2, %xmm1, %xmm2
; AVX512BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
; AVX512BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm5 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX512BW-FCP-NEXT:    vmovdqa 80(%rdi), %xmm2
; AVX512BW-FCP-NEXT:    vpshufb %xmm5, %xmm2, %xmm6
; AVX512BW-FCP-NEXT:    vmovdqa 64(%rdi), %xmm3
; AVX512BW-FCP-NEXT:    vpshufb %xmm5, %xmm3, %xmm5
; AVX512BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} xmm4 = xmm5[0,1,2],xmm4[3]
; AVX512BW-FCP-NEXT:    vmovdqa64 (%rdi), %zmm5
; AVX512BW-FCP-NEXT:    vpmovqb %zmm5, %xmm6
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} xmm4 = xmm6[0,1],xmm4[2,3]
; AVX512BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm6 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX512BW-FCP-NEXT:    vpshufb %xmm6, %xmm0, %xmm7
; AVX512BW-FCP-NEXT:    vpshufb %xmm6, %xmm1, %xmm6
; AVX512BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm6[0],xmm7[0],xmm6[1],xmm7[1],xmm6[2],xmm7[2],xmm6[3],xmm7[3]
; AVX512BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm7 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX512BW-FCP-NEXT:    vpshufb %xmm7, %xmm2, %xmm8
; AVX512BW-FCP-NEXT:    vpshufb %xmm7, %xmm3, %xmm7
; AVX512BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm7 = xmm7[0],xmm8[0],xmm7[1],xmm8[1],xmm7[2],xmm8[2],xmm7[3],xmm8[3]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} xmm6 = xmm7[0,1,2],xmm6[3]
; AVX512BW-FCP-NEXT:    vpsrlq $8, %zmm5, %zmm7
; AVX512BW-FCP-NEXT:    vpmovqb %zmm7, %xmm7
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} xmm6 = xmm7[0,1],xmm6[2,3]
; AVX512BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm7 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX512BW-FCP-NEXT:    vpshufb %xmm7, %xmm0, %xmm8
; AVX512BW-FCP-NEXT:    vpshufb %xmm7, %xmm1, %xmm7
; AVX512BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm7 = xmm7[0],xmm8[0],xmm7[1],xmm8[1],xmm7[2],xmm8[2],xmm7[3],xmm8[3]
; AVX512BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm8 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX512BW-FCP-NEXT:    vpshufb %xmm8, %xmm2, %xmm9
; AVX512BW-FCP-NEXT:    vpshufb %xmm8, %xmm3, %xmm8
; AVX512BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm9[0],xmm8[1],xmm9[1],xmm8[2],xmm9[2],xmm8[3],xmm9[3]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} xmm7 = xmm8[0,1,2],xmm7[3]
; AVX512BW-FCP-NEXT:    vpsrlq $16, %zmm5, %zmm8
; AVX512BW-FCP-NEXT:    vpmovqb %zmm8, %xmm8
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} xmm7 = xmm8[0,1],xmm7[2,3]
; AVX512BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm8 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX512BW-FCP-NEXT:    vpshufb %xmm8, %xmm0, %xmm9
; AVX512BW-FCP-NEXT:    vpshufb %xmm8, %xmm1, %xmm8
; AVX512BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm9[0],xmm8[1],xmm9[1],xmm8[2],xmm9[2],xmm8[3],xmm9[3]
; AVX512BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm9 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX512BW-FCP-NEXT:    vpshufb %xmm9, %xmm2, %xmm10
; AVX512BW-FCP-NEXT:    vpshufb %xmm9, %xmm3, %xmm9
; AVX512BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm9[0],xmm10[0],xmm9[1],xmm10[1],xmm9[2],xmm10[2],xmm9[3],xmm10[3]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} xmm8 = xmm9[0,1,2],xmm8[3]
; AVX512BW-FCP-NEXT:    vpsrlq $24, %zmm5, %zmm9
; AVX512BW-FCP-NEXT:    vpmovqb %zmm9, %xmm9
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} xmm8 = xmm9[0,1],xmm8[2,3]
; AVX512BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm9 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX512BW-FCP-NEXT:    vpshufb %xmm9, %xmm0, %xmm10
; AVX512BW-FCP-NEXT:    vpshufb %xmm9, %xmm1, %xmm9
; AVX512BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm9[0],xmm10[0],xmm9[1],xmm10[1],xmm9[2],xmm10[2],xmm9[3],xmm10[3]
; AVX512BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm10 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX512BW-FCP-NEXT:    vpshufb %xmm10, %xmm2, %xmm11
; AVX512BW-FCP-NEXT:    vpshufb %xmm10, %xmm3, %xmm10
; AVX512BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm10[0],xmm11[0],xmm10[1],xmm11[1],xmm10[2],xmm11[2],xmm10[3],xmm11[3]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} xmm9 = xmm10[0,1,2],xmm9[3]
; AVX512BW-FCP-NEXT:    vpsrlq $32, %zmm5, %zmm10
; AVX512BW-FCP-NEXT:    vpmovqb %zmm10, %xmm10
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} xmm9 = xmm10[0,1],xmm9[2,3]
; AVX512BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm10 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX512BW-FCP-NEXT:    vpshufb %xmm10, %xmm0, %xmm11
; AVX512BW-FCP-NEXT:    vpshufb %xmm10, %xmm1, %xmm10
; AVX512BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm10[0],xmm11[0],xmm10[1],xmm11[1],xmm10[2],xmm11[2],xmm10[3],xmm11[3]
; AVX512BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm11 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX512BW-FCP-NEXT:    vpshufb %xmm11, %xmm2, %xmm12
; AVX512BW-FCP-NEXT:    vpshufb %xmm11, %xmm3, %xmm11
; AVX512BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm11[0],xmm12[0],xmm11[1],xmm12[1],xmm11[2],xmm12[2],xmm11[3],xmm12[3]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} xmm10 = xmm11[0,1,2],xmm10[3]
; AVX512BW-FCP-NEXT:    vpsrlq $40, %zmm5, %zmm11
; AVX512BW-FCP-NEXT:    vpmovqb %zmm11, %xmm11
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} xmm10 = xmm11[0,1],xmm10[2,3]
; AVX512BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm11 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX512BW-FCP-NEXT:    vpshufb %xmm11, %xmm0, %xmm12
; AVX512BW-FCP-NEXT:    vpshufb %xmm11, %xmm1, %xmm11
; AVX512BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm11[0],xmm12[0],xmm11[1],xmm12[1],xmm11[2],xmm12[2],xmm11[3],xmm12[3]
; AVX512BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm12 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX512BW-FCP-NEXT:    vpshufb %xmm12, %xmm2, %xmm13
; AVX512BW-FCP-NEXT:    vpshufb %xmm12, %xmm3, %xmm12
; AVX512BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm12[0],xmm13[0],xmm12[1],xmm13[1],xmm12[2],xmm13[2],xmm12[3],xmm13[3]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} xmm11 = xmm12[0,1,2],xmm11[3]
; AVX512BW-FCP-NEXT:    vpsrlq $48, %zmm5, %zmm12
; AVX512BW-FCP-NEXT:    vpmovqb %zmm12, %xmm12
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} xmm11 = xmm12[0,1],xmm11[2,3]
; AVX512BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm12 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX512BW-FCP-NEXT:    vpshufb %xmm12, %xmm0, %xmm0
; AVX512BW-FCP-NEXT:    vpshufb %xmm12, %xmm1, %xmm1
; AVX512BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
; AVX512BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm1 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX512BW-FCP-NEXT:    vpshufb %xmm1, %xmm2, %xmm2
; AVX512BW-FCP-NEXT:    vpshufb %xmm1, %xmm3, %xmm1
; AVX512BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[3]
; AVX512BW-FCP-NEXT:    vpsrlq $56, %zmm5, %zmm1
; AVX512BW-FCP-NEXT:    vpmovqb %zmm1, %xmm1
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3]
; AVX512BW-FCP-NEXT:    vmovdqa %xmm4, (%rsi)
; AVX512BW-FCP-NEXT:    vmovdqa %xmm6, (%rdx)
; AVX512BW-FCP-NEXT:    vmovdqa %xmm7, (%rcx)
; AVX512BW-FCP-NEXT:    vmovdqa %xmm8, (%r8)
; AVX512BW-FCP-NEXT:    vmovdqa %xmm9, (%r9)
; AVX512BW-FCP-NEXT:    vmovdqa %xmm10, (%r11)
; AVX512BW-FCP-NEXT:    vmovdqa %xmm11, (%r10)
; AVX512BW-FCP-NEXT:    vmovdqa %xmm0, (%rax)
; AVX512BW-FCP-NEXT:    vzeroupper
; AVX512BW-FCP-NEXT:    retq
;
; AVX512DQ-BW-LABEL: load_i8_stride8_vf16:
; AVX512DQ-BW:       # %bb.0:
; AVX512DQ-BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-BW-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512DQ-BW-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm2 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX512DQ-BW-NEXT:    vmovdqa 112(%rdi), %xmm0
; AVX512DQ-BW-NEXT:    vpshufb %xmm2, %xmm0, %xmm3
; AVX512DQ-BW-NEXT:    vmovdqa 96(%rdi), %xmm1
; AVX512DQ-BW-NEXT:    vpshufb %xmm2, %xmm1, %xmm2
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm5 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX512DQ-BW-NEXT:    vmovdqa 80(%rdi), %xmm2
; AVX512DQ-BW-NEXT:    vpshufb %xmm5, %xmm2, %xmm6
; AVX512DQ-BW-NEXT:    vmovdqa 64(%rdi), %xmm3
; AVX512DQ-BW-NEXT:    vpshufb %xmm5, %xmm3, %xmm5
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} xmm4 = xmm5[0,1,2],xmm4[3]
; AVX512DQ-BW-NEXT:    vmovdqa64 (%rdi), %zmm5
; AVX512DQ-BW-NEXT:    vpmovqb %zmm5, %xmm6
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} xmm4 = xmm6[0,1],xmm4[2,3]
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm6 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX512DQ-BW-NEXT:    vpshufb %xmm6, %xmm0, %xmm7
; AVX512DQ-BW-NEXT:    vpshufb %xmm6, %xmm1, %xmm6
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm6[0],xmm7[0],xmm6[1],xmm7[1],xmm6[2],xmm7[2],xmm6[3],xmm7[3]
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm7 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX512DQ-BW-NEXT:    vpshufb %xmm7, %xmm2, %xmm8
; AVX512DQ-BW-NEXT:    vpshufb %xmm7, %xmm3, %xmm7
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm7 = xmm7[0],xmm8[0],xmm7[1],xmm8[1],xmm7[2],xmm8[2],xmm7[3],xmm8[3]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} xmm6 = xmm7[0,1,2],xmm6[3]
; AVX512DQ-BW-NEXT:    vpsrlq $8, %zmm5, %zmm7
; AVX512DQ-BW-NEXT:    vpmovqb %zmm7, %xmm7
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} xmm6 = xmm7[0,1],xmm6[2,3]
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm7 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX512DQ-BW-NEXT:    vpshufb %xmm7, %xmm0, %xmm8
; AVX512DQ-BW-NEXT:    vpshufb %xmm7, %xmm1, %xmm7
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm7 = xmm7[0],xmm8[0],xmm7[1],xmm8[1],xmm7[2],xmm8[2],xmm7[3],xmm8[3]
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm8 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX512DQ-BW-NEXT:    vpshufb %xmm8, %xmm2, %xmm9
; AVX512DQ-BW-NEXT:    vpshufb %xmm8, %xmm3, %xmm8
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm9[0],xmm8[1],xmm9[1],xmm8[2],xmm9[2],xmm8[3],xmm9[3]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} xmm7 = xmm8[0,1,2],xmm7[3]
; AVX512DQ-BW-NEXT:    vpsrlq $16, %zmm5, %zmm8
; AVX512DQ-BW-NEXT:    vpmovqb %zmm8, %xmm8
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} xmm7 = xmm8[0,1],xmm7[2,3]
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm8 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX512DQ-BW-NEXT:    vpshufb %xmm8, %xmm0, %xmm9
; AVX512DQ-BW-NEXT:    vpshufb %xmm8, %xmm1, %xmm8
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm9[0],xmm8[1],xmm9[1],xmm8[2],xmm9[2],xmm8[3],xmm9[3]
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm9 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX512DQ-BW-NEXT:    vpshufb %xmm9, %xmm2, %xmm10
; AVX512DQ-BW-NEXT:    vpshufb %xmm9, %xmm3, %xmm9
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm9[0],xmm10[0],xmm9[1],xmm10[1],xmm9[2],xmm10[2],xmm9[3],xmm10[3]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} xmm8 = xmm9[0,1,2],xmm8[3]
; AVX512DQ-BW-NEXT:    vpsrlq $24, %zmm5, %zmm9
; AVX512DQ-BW-NEXT:    vpmovqb %zmm9, %xmm9
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} xmm8 = xmm9[0,1],xmm8[2,3]
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm9 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX512DQ-BW-NEXT:    vpshufb %xmm9, %xmm0, %xmm10
; AVX512DQ-BW-NEXT:    vpshufb %xmm9, %xmm1, %xmm9
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm9[0],xmm10[0],xmm9[1],xmm10[1],xmm9[2],xmm10[2],xmm9[3],xmm10[3]
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm10 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX512DQ-BW-NEXT:    vpshufb %xmm10, %xmm2, %xmm11
; AVX512DQ-BW-NEXT:    vpshufb %xmm10, %xmm3, %xmm10
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm10[0],xmm11[0],xmm10[1],xmm11[1],xmm10[2],xmm11[2],xmm10[3],xmm11[3]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} xmm9 = xmm10[0,1,2],xmm9[3]
; AVX512DQ-BW-NEXT:    vpsrlq $32, %zmm5, %zmm10
; AVX512DQ-BW-NEXT:    vpmovqb %zmm10, %xmm10
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} xmm9 = xmm10[0,1],xmm9[2,3]
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm10 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX512DQ-BW-NEXT:    vpshufb %xmm10, %xmm0, %xmm11
; AVX512DQ-BW-NEXT:    vpshufb %xmm10, %xmm1, %xmm10
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm10[0],xmm11[0],xmm10[1],xmm11[1],xmm10[2],xmm11[2],xmm10[3],xmm11[3]
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm11 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX512DQ-BW-NEXT:    vpshufb %xmm11, %xmm2, %xmm12
; AVX512DQ-BW-NEXT:    vpshufb %xmm11, %xmm3, %xmm11
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm11[0],xmm12[0],xmm11[1],xmm12[1],xmm11[2],xmm12[2],xmm11[3],xmm12[3]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} xmm10 = xmm11[0,1,2],xmm10[3]
; AVX512DQ-BW-NEXT:    vpsrlq $40, %zmm5, %zmm11
; AVX512DQ-BW-NEXT:    vpmovqb %zmm11, %xmm11
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} xmm10 = xmm11[0,1],xmm10[2,3]
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm11 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX512DQ-BW-NEXT:    vpshufb %xmm11, %xmm0, %xmm12
; AVX512DQ-BW-NEXT:    vpshufb %xmm11, %xmm1, %xmm11
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm11[0],xmm12[0],xmm11[1],xmm12[1],xmm11[2],xmm12[2],xmm11[3],xmm12[3]
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm12 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX512DQ-BW-NEXT:    vpshufb %xmm12, %xmm2, %xmm13
; AVX512DQ-BW-NEXT:    vpshufb %xmm12, %xmm3, %xmm12
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm12[0],xmm13[0],xmm12[1],xmm13[1],xmm12[2],xmm13[2],xmm12[3],xmm13[3]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} xmm11 = xmm12[0,1,2],xmm11[3]
; AVX512DQ-BW-NEXT:    vpsrlq $48, %zmm5, %zmm12
; AVX512DQ-BW-NEXT:    vpmovqb %zmm12, %xmm12
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} xmm11 = xmm12[0,1],xmm11[2,3]
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm12 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX512DQ-BW-NEXT:    vpshufb %xmm12, %xmm0, %xmm0
; AVX512DQ-BW-NEXT:    vpshufb %xmm12, %xmm1, %xmm1
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm1 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX512DQ-BW-NEXT:    vpshufb %xmm1, %xmm2, %xmm2
; AVX512DQ-BW-NEXT:    vpshufb %xmm1, %xmm3, %xmm1
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[3]
; AVX512DQ-BW-NEXT:    vpsrlq $56, %zmm5, %zmm1
; AVX512DQ-BW-NEXT:    vpmovqb %zmm1, %xmm1
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3]
; AVX512DQ-BW-NEXT:    vmovdqa %xmm4, (%rsi)
; AVX512DQ-BW-NEXT:    vmovdqa %xmm6, (%rdx)
; AVX512DQ-BW-NEXT:    vmovdqa %xmm7, (%rcx)
; AVX512DQ-BW-NEXT:    vmovdqa %xmm8, (%r8)
; AVX512DQ-BW-NEXT:    vmovdqa %xmm9, (%r9)
; AVX512DQ-BW-NEXT:    vmovdqa %xmm10, (%r11)
; AVX512DQ-BW-NEXT:    vmovdqa %xmm11, (%r10)
; AVX512DQ-BW-NEXT:    vmovdqa %xmm0, (%rax)
; AVX512DQ-BW-NEXT:    vzeroupper
; AVX512DQ-BW-NEXT:    retq
;
; AVX512DQ-BW-FCP-LABEL: load_i8_stride8_vf16:
; AVX512DQ-BW-FCP:       # %bb.0:
; AVX512DQ-BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512DQ-BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm2 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 112(%rdi), %xmm0
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm2, %xmm0, %xmm3
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 96(%rdi), %xmm1
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm2, %xmm1, %xmm2
; AVX512DQ-BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm5 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 80(%rdi), %xmm2
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm5, %xmm2, %xmm6
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 64(%rdi), %xmm3
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm5, %xmm3, %xmm5
; AVX512DQ-BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} xmm4 = xmm5[0,1,2],xmm4[3]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 (%rdi), %zmm5
; AVX512DQ-BW-FCP-NEXT:    vpmovqb %zmm5, %xmm6
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} xmm4 = xmm6[0,1],xmm4[2,3]
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm6 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm6, %xmm0, %xmm7
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm6, %xmm1, %xmm6
; AVX512DQ-BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm6[0],xmm7[0],xmm6[1],xmm7[1],xmm6[2],xmm7[2],xmm6[3],xmm7[3]
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm7 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm7, %xmm2, %xmm8
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm7, %xmm3, %xmm7
; AVX512DQ-BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm7 = xmm7[0],xmm8[0],xmm7[1],xmm8[1],xmm7[2],xmm8[2],xmm7[3],xmm8[3]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} xmm6 = xmm7[0,1,2],xmm6[3]
; AVX512DQ-BW-FCP-NEXT:    vpsrlq $8, %zmm5, %zmm7
; AVX512DQ-BW-FCP-NEXT:    vpmovqb %zmm7, %xmm7
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} xmm6 = xmm7[0,1],xmm6[2,3]
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm7 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm7, %xmm0, %xmm8
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm7, %xmm1, %xmm7
; AVX512DQ-BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm7 = xmm7[0],xmm8[0],xmm7[1],xmm8[1],xmm7[2],xmm8[2],xmm7[3],xmm8[3]
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm8 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm8, %xmm2, %xmm9
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm8, %xmm3, %xmm8
; AVX512DQ-BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm9[0],xmm8[1],xmm9[1],xmm8[2],xmm9[2],xmm8[3],xmm9[3]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} xmm7 = xmm8[0,1,2],xmm7[3]
; AVX512DQ-BW-FCP-NEXT:    vpsrlq $16, %zmm5, %zmm8
; AVX512DQ-BW-FCP-NEXT:    vpmovqb %zmm8, %xmm8
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} xmm7 = xmm8[0,1],xmm7[2,3]
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm8 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm8, %xmm0, %xmm9
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm8, %xmm1, %xmm8
; AVX512DQ-BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm9[0],xmm8[1],xmm9[1],xmm8[2],xmm9[2],xmm8[3],xmm9[3]
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm9 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm9, %xmm2, %xmm10
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm9, %xmm3, %xmm9
; AVX512DQ-BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm9[0],xmm10[0],xmm9[1],xmm10[1],xmm9[2],xmm10[2],xmm9[3],xmm10[3]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} xmm8 = xmm9[0,1,2],xmm8[3]
; AVX512DQ-BW-FCP-NEXT:    vpsrlq $24, %zmm5, %zmm9
; AVX512DQ-BW-FCP-NEXT:    vpmovqb %zmm9, %xmm9
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} xmm8 = xmm9[0,1],xmm8[2,3]
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm9 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm9, %xmm0, %xmm10
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm9, %xmm1, %xmm9
; AVX512DQ-BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm9[0],xmm10[0],xmm9[1],xmm10[1],xmm9[2],xmm10[2],xmm9[3],xmm10[3]
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm10 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm10, %xmm2, %xmm11
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm10, %xmm3, %xmm10
; AVX512DQ-BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm10[0],xmm11[0],xmm10[1],xmm11[1],xmm10[2],xmm11[2],xmm10[3],xmm11[3]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} xmm9 = xmm10[0,1,2],xmm9[3]
; AVX512DQ-BW-FCP-NEXT:    vpsrlq $32, %zmm5, %zmm10
; AVX512DQ-BW-FCP-NEXT:    vpmovqb %zmm10, %xmm10
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} xmm9 = xmm10[0,1],xmm9[2,3]
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm10 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm10, %xmm0, %xmm11
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm10, %xmm1, %xmm10
; AVX512DQ-BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm10[0],xmm11[0],xmm10[1],xmm11[1],xmm10[2],xmm11[2],xmm10[3],xmm11[3]
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm11 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm11, %xmm2, %xmm12
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm11, %xmm3, %xmm11
; AVX512DQ-BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm11[0],xmm12[0],xmm11[1],xmm12[1],xmm11[2],xmm12[2],xmm11[3],xmm12[3]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} xmm10 = xmm11[0,1,2],xmm10[3]
; AVX512DQ-BW-FCP-NEXT:    vpsrlq $40, %zmm5, %zmm11
; AVX512DQ-BW-FCP-NEXT:    vpmovqb %zmm11, %xmm11
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} xmm10 = xmm11[0,1],xmm10[2,3]
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm11 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm11, %xmm0, %xmm12
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm11, %xmm1, %xmm11
; AVX512DQ-BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm11[0],xmm12[0],xmm11[1],xmm12[1],xmm11[2],xmm12[2],xmm11[3],xmm12[3]
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm12 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm12, %xmm2, %xmm13
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm12, %xmm3, %xmm12
; AVX512DQ-BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm12[0],xmm13[0],xmm12[1],xmm13[1],xmm12[2],xmm13[2],xmm12[3],xmm13[3]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} xmm11 = xmm12[0,1,2],xmm11[3]
; AVX512DQ-BW-FCP-NEXT:    vpsrlq $48, %zmm5, %zmm12
; AVX512DQ-BW-FCP-NEXT:    vpmovqb %zmm12, %xmm12
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} xmm11 = xmm12[0,1],xmm11[2,3]
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm12 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm12, %xmm0, %xmm0
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm12, %xmm1, %xmm1
; AVX512DQ-BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm1 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm1, %xmm2, %xmm2
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm1, %xmm3, %xmm1
; AVX512DQ-BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[3]
; AVX512DQ-BW-FCP-NEXT:    vpsrlq $56, %zmm5, %zmm1
; AVX512DQ-BW-FCP-NEXT:    vpmovqb %zmm1, %xmm1
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa %xmm4, (%rsi)
; AVX512DQ-BW-FCP-NEXT:    vmovdqa %xmm6, (%rdx)
; AVX512DQ-BW-FCP-NEXT:    vmovdqa %xmm7, (%rcx)
; AVX512DQ-BW-FCP-NEXT:    vmovdqa %xmm8, (%r8)
; AVX512DQ-BW-FCP-NEXT:    vmovdqa %xmm9, (%r9)
; AVX512DQ-BW-FCP-NEXT:    vmovdqa %xmm10, (%r11)
; AVX512DQ-BW-FCP-NEXT:    vmovdqa %xmm11, (%r10)
; AVX512DQ-BW-FCP-NEXT:    vmovdqa %xmm0, (%rax)
; AVX512DQ-BW-FCP-NEXT:    vzeroupper
; AVX512DQ-BW-FCP-NEXT:    retq
  %wide.vec = load <128 x i8>, ptr %in.vec, align 64
  %strided.vec0 = shufflevector <128 x i8> %wide.vec, <128 x i8> poison, <16 x i32> <i32 0, i32 8, i32 16, i32 24, i32 32, i32 40, i32 48, i32 56, i32 64, i32 72, i32 80, i32 88, i32 96, i32 104, i32 112, i32 120>
  %strided.vec1 = shufflevector <128 x i8> %wide.vec, <128 x i8> poison, <16 x i32> <i32 1, i32 9, i32 17, i32 25, i32 33, i32 41, i32 49, i32 57, i32 65, i32 73, i32 81, i32 89, i32 97, i32 105, i32 113, i32 121>
  %strided.vec2 = shufflevector <128 x i8> %wide.vec, <128 x i8> poison, <16 x i32> <i32 2, i32 10, i32 18, i32 26, i32 34, i32 42, i32 50, i32 58, i32 66, i32 74, i32 82, i32 90, i32 98, i32 106, i32 114, i32 122>
  %strided.vec3 = shufflevector <128 x i8> %wide.vec, <128 x i8> poison, <16 x i32> <i32 3, i32 11, i32 19, i32 27, i32 35, i32 43, i32 51, i32 59, i32 67, i32 75, i32 83, i32 91, i32 99, i32 107, i32 115, i32 123>
  %strided.vec4 = shufflevector <128 x i8> %wide.vec, <128 x i8> poison, <16 x i32> <i32 4, i32 12, i32 20, i32 28, i32 36, i32 44, i32 52, i32 60, i32 68, i32 76, i32 84, i32 92, i32 100, i32 108, i32 116, i32 124>
  %strided.vec5 = shufflevector <128 x i8> %wide.vec, <128 x i8> poison, <16 x i32> <i32 5, i32 13, i32 21, i32 29, i32 37, i32 45, i32 53, i32 61, i32 69, i32 77, i32 85, i32 93, i32 101, i32 109, i32 117, i32 125>
  %strided.vec6 = shufflevector <128 x i8> %wide.vec, <128 x i8> poison, <16 x i32> <i32 6, i32 14, i32 22, i32 30, i32 38, i32 46, i32 54, i32 62, i32 70, i32 78, i32 86, i32 94, i32 102, i32 110, i32 118, i32 126>
  %strided.vec7 = shufflevector <128 x i8> %wide.vec, <128 x i8> poison, <16 x i32> <i32 7, i32 15, i32 23, i32 31, i32 39, i32 47, i32 55, i32 63, i32 71, i32 79, i32 87, i32 95, i32 103, i32 111, i32 119, i32 127>
  store <16 x i8> %strided.vec0, ptr %out.vec0, align 64
  store <16 x i8> %strided.vec1, ptr %out.vec1, align 64
  store <16 x i8> %strided.vec2, ptr %out.vec2, align 64
  store <16 x i8> %strided.vec3, ptr %out.vec3, align 64
  store <16 x i8> %strided.vec4, ptr %out.vec4, align 64
  store <16 x i8> %strided.vec5, ptr %out.vec5, align 64
  store <16 x i8> %strided.vec6, ptr %out.vec6, align 64
  store <16 x i8> %strided.vec7, ptr %out.vec7, align 64
  ret void
}

define void @load_i8_stride8_vf32(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr %out.vec2, ptr %out.vec3, ptr %out.vec4, ptr %out.vec5, ptr %out.vec6, ptr %out.vec7) nounwind {
; SSE-LABEL: load_i8_stride8_vf32:
; SSE:       # %bb.0:
; SSE-NEXT:    subq $904, %rsp # imm = 0x388
; SSE-NEXT:    movdqa 64(%rdi), %xmm5
; SSE-NEXT:    movdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa 80(%rdi), %xmm10
; SSE-NEXT:    movdqa %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa 96(%rdi), %xmm12
; SSE-NEXT:    movdqa %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa 128(%rdi), %xmm6
; SSE-NEXT:    movdqa 144(%rdi), %xmm13
; SSE-NEXT:    movdqa 160(%rdi), %xmm11
; SSE-NEXT:    movdqa 176(%rdi), %xmm14
; SSE-NEXT:    movdqa 192(%rdi), %xmm2
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa 208(%rdi), %xmm7
; SSE-NEXT:    movdqa 224(%rdi), %xmm8
; SSE-NEXT:    movdqa 240(%rdi), %xmm9
; SSE-NEXT:    movdqa {{.*#+}} xmm4 = [255,0,0,0,255,0,0,0]
; SSE-NEXT:    movdqa %xmm9, %xmm0
; SSE-NEXT:    pand %xmm4, %xmm0
; SSE-NEXT:    movdqa %xmm8, %xmm1
; SSE-NEXT:    pand %xmm4, %xmm1
; SSE-NEXT:    packuswb %xmm0, %xmm1
; SSE-NEXT:    packuswb %xmm1, %xmm0
; SSE-NEXT:    movdqa %xmm7, %xmm1
; SSE-NEXT:    pand %xmm4, %xmm1
; SSE-NEXT:    pand %xmm4, %xmm2
; SSE-NEXT:    packuswb %xmm1, %xmm2
; SSE-NEXT:    packuswb %xmm2, %xmm2
; SSE-NEXT:    packuswb %xmm0, %xmm2
; SSE-NEXT:    movdqa %xmm14, %xmm0
; SSE-NEXT:    pand %xmm4, %xmm0
; SSE-NEXT:    movdqa %xmm11, %xmm1
; SSE-NEXT:    pand %xmm4, %xmm1
; SSE-NEXT:    packuswb %xmm0, %xmm1
; SSE-NEXT:    movdqa %xmm13, %xmm0
; SSE-NEXT:    pand %xmm4, %xmm0
; SSE-NEXT:    movdqa %xmm6, %xmm3
; SSE-NEXT:    pand %xmm4, %xmm3
; SSE-NEXT:    packuswb %xmm0, %xmm3
; SSE-NEXT:    movdqa 112(%rdi), %xmm15
; SSE-NEXT:    movdqa %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    packuswb %xmm1, %xmm0
; SSE-NEXT:    packuswb %xmm3, %xmm3
; SSE-NEXT:    packuswb %xmm0, %xmm3
; SSE-NEXT:    shufps {{.*#+}} xmm3 = xmm3[0,3],xmm2[0,3]
; SSE-NEXT:    movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm15, %xmm0
; SSE-NEXT:    pand %xmm4, %xmm0
; SSE-NEXT:    movdqa %xmm12, %xmm1
; SSE-NEXT:    pand %xmm4, %xmm1
; SSE-NEXT:    packuswb %xmm0, %xmm1
; SSE-NEXT:    movdqa %xmm10, %xmm0
; SSE-NEXT:    pand %xmm4, %xmm0
; SSE-NEXT:    movdqa %xmm5, %xmm2
; SSE-NEXT:    pand %xmm4, %xmm2
; SSE-NEXT:    packuswb %xmm0, %xmm2
; SSE-NEXT:    packuswb %xmm1, %xmm0
; SSE-NEXT:    packuswb %xmm2, %xmm2
; SSE-NEXT:    packuswb %xmm0, %xmm2
; SSE-NEXT:    movdqa 48(%rdi), %xmm15
; SSE-NEXT:    movdqa %xmm15, %xmm0
; SSE-NEXT:    pand %xmm4, %xmm0
; SSE-NEXT:    movdqa 32(%rdi), %xmm1
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pand %xmm4, %xmm1
; SSE-NEXT:    packuswb %xmm0, %xmm1
; SSE-NEXT:    movdqa 16(%rdi), %xmm10
; SSE-NEXT:    movdqa %xmm10, %xmm0
; SSE-NEXT:    pand %xmm4, %xmm0
; SSE-NEXT:    movdqa (%rdi), %xmm12
; SSE-NEXT:    pand %xmm12, %xmm4
; SSE-NEXT:    packuswb %xmm0, %xmm4
; SSE-NEXT:    packuswb %xmm1, %xmm0
; SSE-NEXT:    packuswb %xmm4, %xmm4
; SSE-NEXT:    packuswb %xmm0, %xmm4
; SSE-NEXT:    shufps {{.*#+}} xmm4 = xmm4[0,3],xmm2[0,3]
; SSE-NEXT:    movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pxor %xmm5, %xmm5
; SSE-NEXT:    movdqa %xmm13, %xmm0
; SSE-NEXT:    movdqa %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm13, %xmm1
; SSE-NEXT:    punpckhbw {{.*#+}} xmm1 = xmm1[8],xmm5[8],xmm1[9],xmm5[9],xmm1[10],xmm5[10],xmm1[11],xmm5[11],xmm1[12],xmm5[12],xmm1[13],xmm5[13],xmm1[14],xmm5[14],xmm1[15],xmm5[15]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm5[0],xmm0[1],xmm5[1],xmm0[2],xmm5[2],xmm0[3],xmm5[3],xmm0[4],xmm5[4],xmm0[5],xmm5[5],xmm0[6],xmm5[6],xmm0[7],xmm5[7]
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa {{.*#+}} xmm13 = [65535,0,65535,65535,65535,65535,65535,65535]
; SSE-NEXT:    movdqa %xmm13, %xmm1
; SSE-NEXT:    pandn %xmm0, %xmm1
; SSE-NEXT:    movdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm6, %xmm2
; SSE-NEXT:    punpckhbw {{.*#+}} xmm2 = xmm2[8],xmm5[8],xmm2[9],xmm5[9],xmm2[10],xmm5[10],xmm2[11],xmm5[11],xmm2[12],xmm5[12],xmm2[13],xmm5[13],xmm2[14],xmm5[14],xmm2[15],xmm5[15]
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklbw {{.*#+}} xmm6 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3],xmm6[4],xmm5[4],xmm6[5],xmm5[5],xmm6[6],xmm5[6],xmm6[7],xmm5[7]
; SSE-NEXT:    movdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklwd {{.*#+}} xmm6 = xmm6[0],xmm2[0],xmm6[1],xmm2[1],xmm6[2],xmm2[2],xmm6[3],xmm2[3]
; SSE-NEXT:    movdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[1,1,1,1]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    pand %xmm13, %xmm0
; SSE-NEXT:    por %xmm1, %xmm0
; SSE-NEXT:    movdqa %xmm14, %xmm3
; SSE-NEXT:    movdqa %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm14, %xmm1
; SSE-NEXT:    punpckhbw {{.*#+}} xmm1 = xmm1[8],xmm5[8],xmm1[9],xmm5[9],xmm1[10],xmm5[10],xmm1[11],xmm5[11],xmm1[12],xmm5[12],xmm1[13],xmm5[13],xmm1[14],xmm5[14],xmm1[15],xmm5[15]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm5[0],xmm3[1],xmm5[1],xmm3[2],xmm5[2],xmm3[3],xmm5[3],xmm3[4],xmm5[4],xmm3[5],xmm5[5],xmm3[6],xmm5[6],xmm3[7],xmm5[7]
; SSE-NEXT:    movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklwd {{.*#+}} xmm3 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3]
; SSE-NEXT:    packuswb %xmm3, %xmm3
; SSE-NEXT:    movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[0,0,2,3]
; SSE-NEXT:    movdqa {{.*#+}} xmm14 = [65535,65535,65535,0,65535,65535,65535,65535]
; SSE-NEXT:    movdqa %xmm14, %xmm4
; SSE-NEXT:    pandn %xmm1, %xmm4
; SSE-NEXT:    movdqa %xmm11, %xmm2
; SSE-NEXT:    movdqa %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm11, %xmm1
; SSE-NEXT:    punpckhbw {{.*#+}} xmm1 = xmm1[8],xmm5[8],xmm1[9],xmm5[9],xmm1[10],xmm5[10],xmm1[11],xmm5[11],xmm1[12],xmm5[12],xmm1[13],xmm5[13],xmm1[14],xmm5[14],xmm1[15],xmm5[15]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm5[0],xmm2[1],xmm5[1],xmm2[2],xmm5[2],xmm2[3],xmm5[3],xmm2[4],xmm5[4],xmm2[5],xmm5[5],xmm2[6],xmm5[6],xmm2[7],xmm5[7]
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[0,1,1,3]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm14, %xmm1
; SSE-NEXT:    por %xmm4, %xmm1
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,1,1]
; SSE-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
; SSE-NEXT:    movdqa %xmm9, %xmm1
; SSE-NEXT:    movdqa %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm9, %xmm2
; SSE-NEXT:    punpckhbw {{.*#+}} xmm2 = xmm2[8],xmm5[8],xmm2[9],xmm5[9],xmm2[10],xmm5[10],xmm2[11],xmm5[11],xmm2[12],xmm5[12],xmm2[13],xmm5[13],xmm2[14],xmm5[14],xmm2[15],xmm5[15]
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm5[0],xmm1[1],xmm5[1],xmm1[2],xmm5[2],xmm1[3],xmm5[3],xmm1[4],xmm5[4],xmm1[5],xmm5[5],xmm1[6],xmm5[6],xmm1[7],xmm5[7]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[0,1,2,2]
; SSE-NEXT:    movdqa {{.*#+}} xmm11 = [65535,65535,65535,65535,65535,65535,65535,0]
; SSE-NEXT:    movdqa %xmm11, %xmm4
; SSE-NEXT:    pandn %xmm1, %xmm4
; SSE-NEXT:    movdqa %xmm8, %xmm2
; SSE-NEXT:    movdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm8, %xmm1
; SSE-NEXT:    punpckhbw {{.*#+}} xmm1 = xmm1[8],xmm5[8],xmm1[9],xmm5[9],xmm1[10],xmm5[10],xmm1[11],xmm5[11],xmm1[12],xmm5[12],xmm1[13],xmm5[13],xmm1[14],xmm5[14],xmm1[15],xmm5[15]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm5[0],xmm2[1],xmm5[1],xmm2[2],xmm5[2],xmm2[3],xmm5[3],xmm2[4],xmm5[4],xmm2[5],xmm5[5],xmm2[6],xmm5[6],xmm2[7],xmm5[7]
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[0,1,1,3]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm11, %xmm1
; SSE-NEXT:    por %xmm4, %xmm1
; SSE-NEXT:    movdqa %xmm7, %xmm4
; SSE-NEXT:    movdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm7, %xmm2
; SSE-NEXT:    punpckhbw {{.*#+}} xmm2 = xmm2[8],xmm5[8],xmm2[9],xmm5[9],xmm2[10],xmm5[10],xmm2[11],xmm5[11],xmm2[12],xmm5[12],xmm2[13],xmm5[13],xmm2[14],xmm5[14],xmm2[15],xmm5[15]
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklbw {{.*#+}} xmm4 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3],xmm4[4],xmm5[4],xmm4[5],xmm5[5],xmm4[6],xmm5[6],xmm4[7],xmm5[7]
; SSE-NEXT:    movdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklwd {{.*#+}} xmm4 = xmm4[0],xmm2[0],xmm4[1],xmm2[1],xmm4[2],xmm2[2],xmm4[3],xmm2[3]
; SSE-NEXT:    packuswb %xmm4, %xmm2
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa {{.*#+}} xmm9 = [65535,65535,65535,65535,65535,0,65535,65535]
; SSE-NEXT:    movdqa %xmm9, %xmm4
; SSE-NEXT:    pandn %xmm2, %xmm4
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
; SSE-NEXT:    movdqa %xmm7, %xmm2
; SSE-NEXT:    punpckhbw {{.*#+}} xmm2 = xmm2[8],xmm5[8],xmm2[9],xmm5[9],xmm2[10],xmm5[10],xmm2[11],xmm5[11],xmm2[12],xmm5[12],xmm2[13],xmm5[13],xmm2[14],xmm5[14],xmm2[15],xmm5[15]
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklbw {{.*#+}} xmm7 = xmm7[0],xmm5[0],xmm7[1],xmm5[1],xmm7[2],xmm5[2],xmm7[3],xmm5[3],xmm7[4],xmm5[4],xmm7[5],xmm5[5],xmm7[6],xmm5[6],xmm7[7],xmm5[7]
; SSE-NEXT:    movdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklwd {{.*#+}} xmm7 = xmm7[0],xmm2[0],xmm7[1],xmm2[1],xmm7[2],xmm2[2],xmm7[3],xmm2[3]
; SSE-NEXT:    movdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm8 = xmm7[1,1,1,1]
; SSE-NEXT:    packuswb %xmm8, %xmm8
; SSE-NEXT:    pand %xmm9, %xmm8
; SSE-NEXT:    por %xmm4, %xmm8
; SSE-NEXT:    pshufd {{.*#+}} xmm4 = xmm8[2,2,2,2]
; SSE-NEXT:    punpckhdq {{.*#+}} xmm4 = xmm4[2],xmm1[2],xmm4[3],xmm1[3]
; SSE-NEXT:    movsd {{.*#+}} xmm4 = xmm0[0],xmm4[1]
; SSE-NEXT:    movapd %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm10, (%rsp) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm10, %xmm0
; SSE-NEXT:    punpckhbw {{.*#+}} xmm0 = xmm0[8],xmm5[8],xmm0[9],xmm5[9],xmm0[10],xmm5[10],xmm0[11],xmm5[11],xmm0[12],xmm5[12],xmm0[13],xmm5[13],xmm0[14],xmm5[14],xmm0[15],xmm5[15]
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklbw {{.*#+}} xmm10 = xmm10[0],xmm5[0],xmm10[1],xmm5[1],xmm10[2],xmm5[2],xmm10[3],xmm5[3],xmm10[4],xmm5[4],xmm10[5],xmm5[5],xmm10[6],xmm5[6],xmm10[7],xmm5[7]
; SSE-NEXT:    movdqa %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm10, %xmm8
; SSE-NEXT:    punpcklwd {{.*#+}} xmm8 = xmm8[0],xmm0[0],xmm8[1],xmm0[1],xmm8[2],xmm0[2],xmm8[3],xmm0[3]
; SSE-NEXT:    packuswb %xmm8, %xmm8
; SSE-NEXT:    movdqa %xmm13, %xmm1
; SSE-NEXT:    pandn %xmm8, %xmm1
; SSE-NEXT:    movdqa %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm12, %xmm2
; SSE-NEXT:    punpckhbw {{.*#+}} xmm2 = xmm2[8],xmm5[8],xmm2[9],xmm5[9],xmm2[10],xmm5[10],xmm2[11],xmm5[11],xmm2[12],xmm5[12],xmm2[13],xmm5[13],xmm2[14],xmm5[14],xmm2[15],xmm5[15]
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklbw {{.*#+}} xmm12 = xmm12[0],xmm5[0],xmm12[1],xmm5[1],xmm12[2],xmm5[2],xmm12[3],xmm5[3],xmm12[4],xmm5[4],xmm12[5],xmm5[5],xmm12[6],xmm5[6],xmm12[7],xmm5[7]
; SSE-NEXT:    movdqa %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklwd {{.*#+}} xmm12 = xmm12[0],xmm2[0],xmm12[1],xmm2[1],xmm12[2],xmm2[2],xmm12[3],xmm2[3]
; SSE-NEXT:    movdqa %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm6 = xmm12[1,1,1,1]
; SSE-NEXT:    packuswb %xmm6, %xmm6
; SSE-NEXT:    pand %xmm13, %xmm6
; SSE-NEXT:    por %xmm1, %xmm6
; SSE-NEXT:    movdqa %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm15, %xmm1
; SSE-NEXT:    punpckhbw {{.*#+}} xmm1 = xmm1[8],xmm5[8],xmm1[9],xmm5[9],xmm1[10],xmm5[10],xmm1[11],xmm5[11],xmm1[12],xmm5[12],xmm1[13],xmm5[13],xmm1[14],xmm5[14],xmm1[15],xmm5[15]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklbw {{.*#+}} xmm15 = xmm15[0],xmm5[0],xmm15[1],xmm5[1],xmm15[2],xmm5[2],xmm15[3],xmm5[3],xmm15[4],xmm5[4],xmm15[5],xmm5[5],xmm15[6],xmm5[6],xmm15[7],xmm5[7]
; SSE-NEXT:    movdqa %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm15, %xmm4
; SSE-NEXT:    punpcklwd {{.*#+}} xmm4 = xmm4[0],xmm1[0],xmm4[1],xmm1[1],xmm4[2],xmm1[2],xmm4[3],xmm1[3]
; SSE-NEXT:    packuswb %xmm4, %xmm4
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm4[0,0,2,3]
; SSE-NEXT:    movdqa %xmm14, %xmm12
; SSE-NEXT:    pandn %xmm1, %xmm12
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; SSE-NEXT:    movdqa %xmm2, %xmm1
; SSE-NEXT:    punpckhbw {{.*#+}} xmm1 = xmm1[8],xmm5[8],xmm1[9],xmm5[9],xmm1[10],xmm5[10],xmm1[11],xmm5[11],xmm1[12],xmm5[12],xmm1[13],xmm5[13],xmm1[14],xmm5[14],xmm1[15],xmm5[15]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm5[0],xmm2[1],xmm5[1],xmm2[2],xmm5[2],xmm2[3],xmm5[3],xmm2[4],xmm5[4],xmm2[5],xmm5[5],xmm2[6],xmm5[6],xmm2[7],xmm5[7]
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[0,1,1,3]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm14, %xmm1
; SSE-NEXT:    por %xmm12, %xmm1
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,1,1]
; SSE-NEXT:    punpckldq {{.*#+}} xmm6 = xmm6[0],xmm1[0],xmm6[1],xmm1[1]
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movdqa %xmm1, %xmm2
; SSE-NEXT:    punpckhbw {{.*#+}} xmm2 = xmm2[8],xmm5[8],xmm2[9],xmm5[9],xmm2[10],xmm5[10],xmm2[11],xmm5[11],xmm2[12],xmm5[12],xmm2[13],xmm5[13],xmm2[14],xmm5[14],xmm2[15],xmm5[15]
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm5[0],xmm1[1],xmm5[1],xmm1[2],xmm5[2],xmm1[3],xmm5[3],xmm1[4],xmm5[4],xmm1[5],xmm5[5],xmm1[6],xmm5[6],xmm1[7],xmm5[7]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
; SSE-NEXT:    packuswb %xmm1, %xmm0
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,1,2,2]
; SSE-NEXT:    movdqa %xmm11, %xmm7
; SSE-NEXT:    pandn %xmm1, %xmm7
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload
; SSE-NEXT:    movdqa %xmm12, %xmm1
; SSE-NEXT:    punpckhbw {{.*#+}} xmm1 = xmm1[8],xmm5[8],xmm1[9],xmm5[9],xmm1[10],xmm5[10],xmm1[11],xmm5[11],xmm1[12],xmm5[12],xmm1[13],xmm5[13],xmm1[14],xmm5[14],xmm1[15],xmm5[15]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklbw {{.*#+}} xmm12 = xmm12[0],xmm5[0],xmm12[1],xmm5[1],xmm12[2],xmm5[2],xmm12[3],xmm5[3],xmm12[4],xmm5[4],xmm12[5],xmm5[5],xmm12[6],xmm5[6],xmm12[7],xmm5[7]
; SSE-NEXT:    movdqa %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklwd {{.*#+}} xmm12 = xmm12[0],xmm1[0],xmm12[1],xmm1[1],xmm12[2],xmm1[2],xmm12[3],xmm1[3]
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm12[0,1,1,3]
; SSE-NEXT:    packuswb %xmm1, %xmm3
; SSE-NEXT:    pand %xmm11, %xmm3
; SSE-NEXT:    por %xmm7, %xmm3
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movdqa %xmm0, %xmm2
; SSE-NEXT:    punpckhbw {{.*#+}} xmm2 = xmm2[8],xmm5[8],xmm2[9],xmm5[9],xmm2[10],xmm5[10],xmm2[11],xmm5[11],xmm2[12],xmm5[12],xmm2[13],xmm5[13],xmm2[14],xmm5[14],xmm2[15],xmm5[15]
; SSE-NEXT:    movdqa %xmm2, %xmm7
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm0, %xmm2
; SSE-NEXT:    punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm5[0],xmm2[1],xmm5[1],xmm2[2],xmm5[2],xmm2[3],xmm5[3],xmm2[4],xmm5[4],xmm2[5],xmm5[5],xmm2[6],xmm5[6],xmm2[7],xmm5[7]
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movdqa %xmm0, %xmm1
; SSE-NEXT:    punpckhbw {{.*#+}} xmm1 = xmm1[8],xmm5[8],xmm1[9],xmm5[9],xmm1[10],xmm5[10],xmm1[11],xmm5[11],xmm1[12],xmm5[12],xmm1[13],xmm5[13],xmm1[14],xmm5[14],xmm1[15],xmm5[15]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm5[0],xmm0[1],xmm5[1],xmm0[2],xmm5[2],xmm0[3],xmm5[3],xmm0[4],xmm5[4],xmm0[5],xmm5[5],xmm0[6],xmm5[6],xmm0[7],xmm5[7]
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm2, %xmm5
; SSE-NEXT:    punpcklwd {{.*#+}} xmm5 = xmm5[0],xmm7[0],xmm5[1],xmm7[1],xmm5[2],xmm7[2],xmm5[3],xmm7[3]
; SSE-NEXT:    packuswb %xmm5, %xmm2
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm9, %xmm7
; SSE-NEXT:    pandn %xmm2, %xmm7
; SSE-NEXT:    movdqa %xmm0, %xmm5
; SSE-NEXT:    punpcklwd {{.*#+}} xmm5 = xmm5[0],xmm1[0],xmm5[1],xmm1[1],xmm5[2],xmm1[2],xmm5[3],xmm1[3]
; SSE-NEXT:    pshufd {{.*#+}} xmm15 = xmm5[1,1,1,1]
; SSE-NEXT:    packuswb %xmm15, %xmm15
; SSE-NEXT:    pand %xmm9, %xmm15
; SSE-NEXT:    por %xmm7, %xmm15
; SSE-NEXT:    pshufd {{.*#+}} xmm7 = xmm15[2,2,2,2]
; SSE-NEXT:    punpckhdq {{.*#+}} xmm7 = xmm7[2],xmm3[2],xmm7[3],xmm3[3]
; SSE-NEXT:    movsd {{.*#+}} xmm7 = xmm6[0],xmm7[1]
; SSE-NEXT:    movapd %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa {{.*#+}} xmm2 = [255,255,255,255,255,255,255,255]
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; SSE-NEXT:    pand %xmm2, %xmm10
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm10[0,2,2,3]
; SSE-NEXT:    movdqa %xmm10, %xmm3
; SSE-NEXT:    pshuflw {{.*#+}} xmm0 = xmm0[0,1,1,3,4,5,6,7]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    movdqa %xmm13, %xmm7
; SSE-NEXT:    pandn %xmm0, %xmm7
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; SSE-NEXT:    pand %xmm2, %xmm10
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm10[0,2,2,3]
; SSE-NEXT:    movdqa %xmm10, %xmm6
; SSE-NEXT:    pshuflw {{.*#+}} xmm1 = xmm0[1,3,2,3,4,5,6,7]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm13, %xmm1
; SSE-NEXT:    por %xmm7, %xmm1
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    pand %xmm2, %xmm0
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm7 = xmm0[0,1,2,0]
; SSE-NEXT:    pshufhw {{.*#+}} xmm7 = xmm7[0,1,2,3,4,5,7,5]
; SSE-NEXT:    packuswb %xmm7, %xmm7
; SSE-NEXT:    movdqa %xmm14, %xmm15
; SSE-NEXT:    pandn %xmm7, %xmm15
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    pand %xmm2, %xmm0
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm7 = xmm0[0,1,2,0]
; SSE-NEXT:    pshufhw {{.*#+}} xmm7 = xmm7[0,1,2,3,7,5,6,7]
; SSE-NEXT:    packuswb %xmm7, %xmm7
; SSE-NEXT:    pand %xmm14, %xmm7
; SSE-NEXT:    por %xmm15, %xmm7
; SSE-NEXT:    pshufd {{.*#+}} xmm7 = xmm7[1,1,1,1]
; SSE-NEXT:    punpckldq {{.*#+}} xmm1 = xmm1[0],xmm7[0],xmm1[1],xmm7[1]
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    pand %xmm2, %xmm0
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm7 = xmm0[0,1,2,0]
; SSE-NEXT:    pshufhw {{.*#+}} xmm7 = xmm7[0,1,2,3,4,5,7,5]
; SSE-NEXT:    packuswb %xmm7, %xmm7
; SSE-NEXT:    movdqa %xmm11, %xmm15
; SSE-NEXT:    pandn %xmm7, %xmm15
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    pand %xmm2, %xmm0
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm7 = xmm0[0,1,2,0]
; SSE-NEXT:    pshufhw {{.*#+}} xmm7 = xmm7[0,1,2,3,7,5,6,7]
; SSE-NEXT:    packuswb %xmm7, %xmm7
; SSE-NEXT:    pand %xmm11, %xmm7
; SSE-NEXT:    por %xmm15, %xmm7
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    pand %xmm2, %xmm0
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm15 = xmm0[0,2,2,3]
; SSE-NEXT:    pshuflw {{.*#+}} xmm15 = xmm15[0,1,1,3,4,5,6,7]
; SSE-NEXT:    packuswb %xmm15, %xmm15
; SSE-NEXT:    movdqa %xmm9, %xmm10
; SSE-NEXT:    pandn %xmm15, %xmm10
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    pand %xmm2, %xmm0
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm15 = xmm0[0,2,2,3]
; SSE-NEXT:    pshuflw {{.*#+}} xmm15 = xmm15[1,3,2,3,4,5,6,7]
; SSE-NEXT:    packuswb %xmm15, %xmm15
; SSE-NEXT:    pand %xmm9, %xmm15
; SSE-NEXT:    por %xmm10, %xmm15
; SSE-NEXT:    pshufd {{.*#+}} xmm10 = xmm15[2,2,2,2]
; SSE-NEXT:    punpckhdq {{.*#+}} xmm10 = xmm10[2],xmm7[2],xmm10[3],xmm7[3]
; SSE-NEXT:    movsd {{.*#+}} xmm10 = xmm1[0],xmm10[1]
; SSE-NEXT:    movapd %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa (%rsp), %xmm0 # 16-byte Reload
; SSE-NEXT:    pand %xmm2, %xmm0
; SSE-NEXT:    movdqa %xmm0, (%rsp) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
; SSE-NEXT:    pshuflw {{.*#+}} xmm0 = xmm0[0,1,1,3,4,5,6,7]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    movdqa %xmm13, %xmm7
; SSE-NEXT:    pandn %xmm0, %xmm7
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    pand %xmm2, %xmm0
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
; SSE-NEXT:    pshuflw {{.*#+}} xmm0 = xmm0[1,3,2,3,4,5,6,7]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    pand %xmm13, %xmm0
; SSE-NEXT:    por %xmm7, %xmm0
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    pand %xmm2, %xmm1
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm7 = xmm1[0,1,2,0]
; SSE-NEXT:    pshufhw {{.*#+}} xmm7 = xmm7[0,1,2,3,4,5,7,5]
; SSE-NEXT:    packuswb %xmm7, %xmm7
; SSE-NEXT:    movdqa %xmm14, %xmm10
; SSE-NEXT:    pandn %xmm7, %xmm10
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    pand %xmm2, %xmm1
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm7 = xmm1[0,1,2,0]
; SSE-NEXT:    pshufhw {{.*#+}} xmm7 = xmm7[0,1,2,3,7,5,6,7]
; SSE-NEXT:    packuswb %xmm7, %xmm7
; SSE-NEXT:    pand %xmm14, %xmm7
; SSE-NEXT:    por %xmm10, %xmm7
; SSE-NEXT:    pshufd {{.*#+}} xmm7 = xmm7[1,1,1,1]
; SSE-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm7[0],xmm0[1],xmm7[1]
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    pand %xmm2, %xmm1
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm7 = xmm1[0,1,2,0]
; SSE-NEXT:    pshufhw {{.*#+}} xmm7 = xmm7[0,1,2,3,4,5,7,5]
; SSE-NEXT:    packuswb %xmm7, %xmm7
; SSE-NEXT:    movdqa %xmm11, %xmm10
; SSE-NEXT:    pandn %xmm7, %xmm10
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    pand %xmm2, %xmm1
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm7 = xmm1[0,1,2,0]
; SSE-NEXT:    pshufhw {{.*#+}} xmm7 = xmm7[0,1,2,3,7,5,6,7]
; SSE-NEXT:    packuswb %xmm7, %xmm7
; SSE-NEXT:    pand %xmm11, %xmm7
; SSE-NEXT:    por %xmm10, %xmm7
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    pand %xmm2, %xmm1
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm10 = xmm1[0,2,2,3]
; SSE-NEXT:    pshuflw {{.*#+}} xmm10 = xmm10[0,1,1,3,4,5,6,7]
; SSE-NEXT:    packuswb %xmm10, %xmm10
; SSE-NEXT:    movdqa %xmm9, %xmm15
; SSE-NEXT:    pandn %xmm10, %xmm15
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    pand %xmm2, %xmm1
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[0,2,2,3]
; SSE-NEXT:    pshuflw {{.*#+}} xmm1 = xmm1[1,3,2,3,4,5,6,7]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm9, %xmm1
; SSE-NEXT:    por %xmm15, %xmm1
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[2,2,2,2]
; SSE-NEXT:    punpckhdq {{.*#+}} xmm1 = xmm1[2],xmm7[2],xmm1[3],xmm7[3]
; SSE-NEXT:    movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]
; SSE-NEXT:    movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd $229, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
; SSE-NEXT:    # xmm0 = mem[1,1,2,3]
; SSE-NEXT:    movdqa %xmm13, %xmm1
; SSE-NEXT:    pandn %xmm0, %xmm1
; SSE-NEXT:    pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
; SSE-NEXT:    # xmm0 = mem[3,3,3,3]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    pand %xmm13, %xmm0
; SSE-NEXT:    por %xmm1, %xmm0
; SSE-NEXT:    movdqa %xmm14, %xmm1
; SSE-NEXT:    pandn {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    pshufd $250, {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Folded Reload
; SSE-NEXT:    # xmm7 = mem[2,2,3,3]
; SSE-NEXT:    packuswb %xmm7, %xmm7
; SSE-NEXT:    pand %xmm14, %xmm7
; SSE-NEXT:    por %xmm1, %xmm7
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm7[1,1,1,1]
; SSE-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
; SSE-NEXT:    movdqa %xmm11, %xmm1
; SSE-NEXT:    pandn {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    pshufd $250, {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Folded Reload
; SSE-NEXT:    # xmm7 = mem[2,2,3,3]
; SSE-NEXT:    packuswb %xmm7, %xmm7
; SSE-NEXT:    pand %xmm11, %xmm7
; SSE-NEXT:    por %xmm1, %xmm7
; SSE-NEXT:    pshufd $244, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[0,1,3,3]
; SSE-NEXT:    movdqa %xmm9, %xmm10
; SSE-NEXT:    pandn %xmm1, %xmm10
; SSE-NEXT:    pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[3,3,3,3]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm9, %xmm1
; SSE-NEXT:    por %xmm10, %xmm1
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[2,2,2,2]
; SSE-NEXT:    punpckhdq {{.*#+}} xmm1 = xmm1[2],xmm7[2],xmm1[3],xmm7[3]
; SSE-NEXT:    movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]
; SSE-NEXT:    movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm8[1,1,2,3]
; SSE-NEXT:    movdqa %xmm13, %xmm1
; SSE-NEXT:    pandn %xmm0, %xmm1
; SSE-NEXT:    pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
; SSE-NEXT:    # xmm0 = mem[3,3,3,3]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    pand %xmm13, %xmm0
; SSE-NEXT:    por %xmm1, %xmm0
; SSE-NEXT:    movdqa %xmm14, %xmm1
; SSE-NEXT:    pandn %xmm4, %xmm1
; SSE-NEXT:    pshufd $250, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Folded Reload
; SSE-NEXT:    # xmm4 = mem[2,2,3,3]
; SSE-NEXT:    packuswb %xmm4, %xmm4
; SSE-NEXT:    pand %xmm14, %xmm4
; SSE-NEXT:    por %xmm1, %xmm4
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm4[1,1,1,1]
; SSE-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
; SSE-NEXT:    movdqa %xmm11, %xmm1
; SSE-NEXT:    pandn {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    pshufd {{.*#+}} xmm4 = xmm12[2,2,3,3]
; SSE-NEXT:    packuswb %xmm4, %xmm4
; SSE-NEXT:    pand %xmm11, %xmm4
; SSE-NEXT:    por %xmm1, %xmm4
; SSE-NEXT:    pshufd $244, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[0,1,3,3]
; SSE-NEXT:    movdqa %xmm9, %xmm7
; SSE-NEXT:    pandn %xmm1, %xmm7
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm5[3,3,3,3]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm9, %xmm1
; SSE-NEXT:    por %xmm7, %xmm1
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[2,2,2,2]
; SSE-NEXT:    punpckhdq {{.*#+}} xmm1 = xmm1[2],xmm4[2],xmm1[3],xmm4[3]
; SSE-NEXT:    movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]
; SSE-NEXT:    movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[3,1,2,3]
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshuflw {{.*#+}} xmm0 = xmm0[0,1,2,0,4,5,6,7]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    movdqa %xmm13, %xmm1
; SSE-NEXT:    pandn %xmm0, %xmm1
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[3,1,2,3]
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshuflw {{.*#+}} xmm0 = xmm0[2,0,2,3,4,5,6,7]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    pand %xmm13, %xmm0
; SSE-NEXT:    por %xmm1, %xmm0
; SSE-NEXT:    pshufd $212, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[0,1,1,3]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,5,4,6]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    movdqa %xmm14, %xmm4
; SSE-NEXT:    pandn %xmm1, %xmm4
; SSE-NEXT:    pshufd $212, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[0,1,1,3]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,6,6,7]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm14, %xmm1
; SSE-NEXT:    por %xmm4, %xmm1
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,1,1]
; SSE-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
; SSE-NEXT:    pshufd $212, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[0,1,1,3]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,5,4,6]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    movdqa %xmm11, %xmm4
; SSE-NEXT:    pandn %xmm1, %xmm4
; SSE-NEXT:    pshufd $212, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[0,1,1,3]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,6,6,7]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm11, %xmm1
; SSE-NEXT:    por %xmm4, %xmm1
; SSE-NEXT:    pshufd $231, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
; SSE-NEXT:    # xmm2 = mem[3,1,2,3]
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshuflw {{.*#+}} xmm4 = xmm2[0,1,2,0,4,5,6,7]
; SSE-NEXT:    packuswb %xmm4, %xmm4
; SSE-NEXT:    movdqa %xmm9, %xmm5
; SSE-NEXT:    pandn %xmm4, %xmm5
; SSE-NEXT:    pshufd $231, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
; SSE-NEXT:    # xmm2 = mem[3,1,2,3]
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshuflw {{.*#+}} xmm4 = xmm2[2,0,2,3,4,5,6,7]
; SSE-NEXT:    packuswb %xmm4, %xmm4
; SSE-NEXT:    pand %xmm9, %xmm4
; SSE-NEXT:    por %xmm5, %xmm4
; SSE-NEXT:    pshufd {{.*#+}} xmm6 = xmm4[2,2,2,2]
; SSE-NEXT:    punpckhdq {{.*#+}} xmm6 = xmm6[2],xmm1[2],xmm6[3],xmm1[3]
; SSE-NEXT:    movsd {{.*#+}} xmm6 = xmm0[0],xmm6[1]
; SSE-NEXT:    pshufd $231, (%rsp), %xmm0 # 16-byte Folded Reload
; SSE-NEXT:    # xmm0 = mem[3,1,2,3]
; SSE-NEXT:    movdqa %xmm0, (%rsp) # 16-byte Spill
; SSE-NEXT:    pshuflw {{.*#+}} xmm0 = xmm0[0,1,2,0,4,5,6,7]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    movdqa %xmm13, %xmm1
; SSE-NEXT:    pandn %xmm0, %xmm1
; SSE-NEXT:    pshufd $231, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
; SSE-NEXT:    # xmm0 = mem[3,1,2,3]
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshuflw {{.*#+}} xmm0 = xmm0[2,0,2,3,4,5,6,7]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    pand %xmm13, %xmm0
; SSE-NEXT:    por %xmm1, %xmm0
; SSE-NEXT:    pshufd $212, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[0,1,1,3]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,5,4,6]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    movdqa %xmm14, %xmm4
; SSE-NEXT:    pandn %xmm1, %xmm4
; SSE-NEXT:    pshufd $212, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[0,1,1,3]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,6,6,7]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm14, %xmm1
; SSE-NEXT:    por %xmm4, %xmm1
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,1,1]
; SSE-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
; SSE-NEXT:    pshufd $212, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[0,1,1,3]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,5,4,6]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    movdqa %xmm11, %xmm4
; SSE-NEXT:    pandn %xmm1, %xmm4
; SSE-NEXT:    pshufd $212, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[0,1,1,3]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,6,6,7]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm11, %xmm1
; SSE-NEXT:    por %xmm4, %xmm1
; SSE-NEXT:    pshufd $231, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
; SSE-NEXT:    # xmm2 = mem[3,1,2,3]
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshuflw {{.*#+}} xmm4 = xmm2[0,1,2,0,4,5,6,7]
; SSE-NEXT:    packuswb %xmm4, %xmm4
; SSE-NEXT:    movdqa %xmm9, %xmm5
; SSE-NEXT:    pandn %xmm4, %xmm5
; SSE-NEXT:    pshufd $231, {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Folded Reload
; SSE-NEXT:    # xmm12 = mem[3,1,2,3]
; SSE-NEXT:    pshuflw {{.*#+}} xmm4 = xmm12[2,0,2,3,4,5,6,7]
; SSE-NEXT:    packuswb %xmm4, %xmm4
; SSE-NEXT:    pand %xmm9, %xmm4
; SSE-NEXT:    por %xmm5, %xmm4
; SSE-NEXT:    pshufd {{.*#+}} xmm10 = xmm4[2,2,2,2]
; SSE-NEXT:    punpckhdq {{.*#+}} xmm10 = xmm10[2],xmm1[2],xmm10[3],xmm1[3]
; SSE-NEXT:    movsd {{.*#+}} xmm10 = xmm0[0],xmm10[1]
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    punpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
; SSE-NEXT:    # xmm0 = xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7]
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; SSE-NEXT:    punpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
; SSE-NEXT:    # xmm2 = xmm2[4],mem[4],xmm2[5],mem[5],xmm2[6],mem[6],xmm2[7],mem[7]
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm13, %xmm1
; SSE-NEXT:    pandn %xmm0, %xmm1
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,1,1]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    pand %xmm13, %xmm0
; SSE-NEXT:    por %xmm1, %xmm0
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    punpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = xmm1[4],mem[4],xmm1[5],mem[5],xmm1[6],mem[6],xmm1[7],mem[7]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[0,0,2,3]
; SSE-NEXT:    movdqa %xmm14, %xmm4
; SSE-NEXT:    pandn %xmm1, %xmm4
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    punpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = xmm1[4],mem[4],xmm1[5],mem[5],xmm1[6],mem[6],xmm1[7],mem[7]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[0,1,1,3]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm14, %xmm1
; SSE-NEXT:    por %xmm4, %xmm1
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,1,1]
; SSE-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    punpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = xmm1[4],mem[4],xmm1[5],mem[5],xmm1[6],mem[6],xmm1[7],mem[7]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[0,1,2,2]
; SSE-NEXT:    movdqa %xmm11, %xmm4
; SSE-NEXT:    pandn %xmm1, %xmm4
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    punpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = xmm1[4],mem[4],xmm1[5],mem[5],xmm1[6],mem[6],xmm1[7],mem[7]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[0,1,1,3]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm11, %xmm1
; SSE-NEXT:    por %xmm4, %xmm1
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; SSE-NEXT:    punpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload
; SSE-NEXT:    # xmm3 = xmm3[4],mem[4],xmm3[5],mem[5],xmm3[6],mem[6],xmm3[7],mem[7]
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; SSE-NEXT:    punpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
; SSE-NEXT:    # xmm2 = xmm2[4],mem[4],xmm2[5],mem[5],xmm2[6],mem[6],xmm2[7],mem[7]
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    packuswb %xmm3, %xmm3
; SSE-NEXT:    movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm9, %xmm4
; SSE-NEXT:    pandn %xmm3, %xmm4
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm2[1,1,1,1]
; SSE-NEXT:    packuswb %xmm5, %xmm5
; SSE-NEXT:    pand %xmm9, %xmm5
; SSE-NEXT:    por %xmm4, %xmm5
; SSE-NEXT:    pshufd {{.*#+}} xmm8 = xmm5[2,2,2,2]
; SSE-NEXT:    punpckhdq {{.*#+}} xmm8 = xmm8[2],xmm1[2],xmm8[3],xmm1[3]
; SSE-NEXT:    movsd {{.*#+}} xmm8 = xmm0[0],xmm8[1]
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    punpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
; SSE-NEXT:    # xmm0 = xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7]
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; SSE-NEXT:    punpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
; SSE-NEXT:    # xmm2 = xmm2[4],mem[4],xmm2[5],mem[5],xmm2[6],mem[6],xmm2[7],mem[7]
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm13, %xmm1
; SSE-NEXT:    pandn %xmm0, %xmm1
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,1,1]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    pand %xmm13, %xmm0
; SSE-NEXT:    por %xmm1, %xmm0
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    punpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = xmm1[4],mem[4],xmm1[5],mem[5],xmm1[6],mem[6],xmm1[7],mem[7]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[0,0,2,3]
; SSE-NEXT:    movdqa %xmm14, %xmm4
; SSE-NEXT:    pandn %xmm1, %xmm4
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    punpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = xmm1[4],mem[4],xmm1[5],mem[5],xmm1[6],mem[6],xmm1[7],mem[7]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[0,1,1,3]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm14, %xmm1
; SSE-NEXT:    por %xmm4, %xmm1
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,1,1]
; SSE-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    punpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = xmm1[4],mem[4],xmm1[5],mem[5],xmm1[6],mem[6],xmm1[7],mem[7]
; SSE-NEXT:    packuswb %xmm1, %xmm7
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm7[0,1,2,2]
; SSE-NEXT:    movdqa %xmm11, %xmm4
; SSE-NEXT:    pandn %xmm1, %xmm4
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    punpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = xmm1[4],mem[4],xmm1[5],mem[5],xmm1[6],mem[6],xmm1[7],mem[7]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[0,1,1,3]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm11, %xmm1
; SSE-NEXT:    por %xmm4, %xmm1
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; SSE-NEXT:    punpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload
; SSE-NEXT:    # xmm3 = xmm3[4],mem[4],xmm3[5],mem[5],xmm3[6],mem[6],xmm3[7],mem[7]
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; SSE-NEXT:    punpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
; SSE-NEXT:    # xmm2 = xmm2[4],mem[4],xmm2[5],mem[5],xmm2[6],mem[6],xmm2[7],mem[7]
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    packuswb %xmm3, %xmm5
; SSE-NEXT:    movdqa %xmm9, %xmm4
; SSE-NEXT:    pandn %xmm5, %xmm4
; SSE-NEXT:    pshufd {{.*#+}} xmm15 = xmm2[1,1,1,1]
; SSE-NEXT:    packuswb %xmm15, %xmm15
; SSE-NEXT:    pand %xmm9, %xmm15
; SSE-NEXT:    por %xmm4, %xmm15
; SSE-NEXT:    pshufd {{.*#+}} xmm4 = xmm15[2,2,2,2]
; SSE-NEXT:    punpckhdq {{.*#+}} xmm4 = xmm4[2],xmm1[2],xmm4[3],xmm1[3]
; SSE-NEXT:    movsd {{.*#+}} xmm4 = xmm0[0],xmm4[1]
; SSE-NEXT:    pshuflw $116, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
; SSE-NEXT:    # xmm0 = mem[0,1,3,1,4,5,6,7]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    movdqa %xmm13, %xmm1
; SSE-NEXT:    pandn %xmm0, %xmm1
; SSE-NEXT:    pshuflw $231, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
; SSE-NEXT:    # xmm0 = mem[3,1,2,3,4,5,6,7]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    pand %xmm13, %xmm0
; SSE-NEXT:    por %xmm1, %xmm0
; SSE-NEXT:    pshufhw $212, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[0,1,2,3,4,5,5,7]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    movdqa %xmm14, %xmm15
; SSE-NEXT:    pandn %xmm1, %xmm15
; SSE-NEXT:    pshufhw $237, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[0,1,2,3,5,7,6,7]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm14, %xmm1
; SSE-NEXT:    por %xmm15, %xmm1
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,1,1]
; SSE-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
; SSE-NEXT:    pshufhw $212, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[0,1,2,3,4,5,5,7]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    movdqa %xmm11, %xmm15
; SSE-NEXT:    pandn %xmm1, %xmm15
; SSE-NEXT:    pshufhw $237, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[0,1,2,3,5,7,6,7]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm11, %xmm1
; SSE-NEXT:    por %xmm15, %xmm1
; SSE-NEXT:    pshuflw $116, {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Folded Reload
; SSE-NEXT:    # xmm15 = mem[0,1,3,1,4,5,6,7]
; SSE-NEXT:    packuswb %xmm15, %xmm15
; SSE-NEXT:    movdqa %xmm9, %xmm2
; SSE-NEXT:    pandn %xmm15, %xmm2
; SSE-NEXT:    pshuflw $231, {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Folded Reload
; SSE-NEXT:    # xmm15 = mem[3,1,2,3,4,5,6,7]
; SSE-NEXT:    packuswb %xmm15, %xmm15
; SSE-NEXT:    pand %xmm9, %xmm15
; SSE-NEXT:    por %xmm2, %xmm15
; SSE-NEXT:    pshufd {{.*#+}} xmm15 = xmm15[2,2,2,2]
; SSE-NEXT:    punpckhdq {{.*#+}} xmm15 = xmm15[2],xmm1[2],xmm15[3],xmm1[3]
; SSE-NEXT:    movsd {{.*#+}} xmm15 = xmm0[0],xmm15[1]
; SSE-NEXT:    pshuflw $116, (%rsp), %xmm0 # 16-byte Folded Reload
; SSE-NEXT:    # xmm0 = mem[0,1,3,1,4,5,6,7]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    movdqa %xmm13, %xmm1
; SSE-NEXT:    pandn %xmm0, %xmm1
; SSE-NEXT:    pshuflw $231, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
; SSE-NEXT:    # xmm0 = mem[3,1,2,3,4,5,6,7]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    pand %xmm13, %xmm0
; SSE-NEXT:    por %xmm1, %xmm0
; SSE-NEXT:    pshufhw $212, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[0,1,2,3,4,5,5,7]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    movdqa %xmm14, %xmm2
; SSE-NEXT:    pandn %xmm1, %xmm2
; SSE-NEXT:    pshufhw $237, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[0,1,2,3,5,7,6,7]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm14, %xmm1
; SSE-NEXT:    por %xmm2, %xmm1
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,1,1]
; SSE-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
; SSE-NEXT:    pshufhw $212, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[0,1,2,3,4,5,5,7]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    movdqa %xmm11, %xmm2
; SSE-NEXT:    pandn %xmm1, %xmm2
; SSE-NEXT:    pshufhw $237, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[0,1,2,3,5,7,6,7]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm11, %xmm1
; SSE-NEXT:    por %xmm2, %xmm1
; SSE-NEXT:    pshuflw $116, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
; SSE-NEXT:    # xmm2 = mem[0,1,3,1,4,5,6,7]
; SSE-NEXT:    packuswb %xmm2, %xmm2
; SSE-NEXT:    movdqa %xmm9, %xmm3
; SSE-NEXT:    pandn %xmm2, %xmm3
; SSE-NEXT:    pshuflw {{.*#+}} xmm2 = xmm12[3,1,2,3,4,5,6,7]
; SSE-NEXT:    packuswb %xmm2, %xmm2
; SSE-NEXT:    pand %xmm9, %xmm2
; SSE-NEXT:    por %xmm3, %xmm2
; SSE-NEXT:    pshufd {{.*#+}} xmm12 = xmm2[2,2,2,2]
; SSE-NEXT:    punpckhdq {{.*#+}} xmm12 = xmm12[2],xmm1[2],xmm12[3],xmm1[3]
; SSE-NEXT:    movsd {{.*#+}} xmm12 = xmm0[0],xmm12[1]
; SSE-NEXT:    pshufd $229, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
; SSE-NEXT:    # xmm0 = mem[1,1,2,3]
; SSE-NEXT:    movdqa %xmm13, %xmm1
; SSE-NEXT:    pandn %xmm0, %xmm1
; SSE-NEXT:    pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
; SSE-NEXT:    # xmm0 = mem[3,3,3,3]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    pand %xmm13, %xmm0
; SSE-NEXT:    por %xmm1, %xmm0
; SSE-NEXT:    movdqa %xmm14, %xmm1
; SSE-NEXT:    pandn {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    pshufd $250, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
; SSE-NEXT:    # xmm2 = mem[2,2,3,3]
; SSE-NEXT:    packuswb %xmm2, %xmm2
; SSE-NEXT:    pand %xmm14, %xmm2
; SSE-NEXT:    por %xmm1, %xmm2
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[1,1,1,1]
; SSE-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
; SSE-NEXT:    movdqa %xmm11, %xmm1
; SSE-NEXT:    pandn %xmm7, %xmm1
; SSE-NEXT:    pshufd $250, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
; SSE-NEXT:    # xmm2 = mem[2,2,3,3]
; SSE-NEXT:    packuswb %xmm2, %xmm2
; SSE-NEXT:    pand %xmm11, %xmm2
; SSE-NEXT:    por %xmm1, %xmm2
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm5[0,1,3,3]
; SSE-NEXT:    movdqa %xmm9, %xmm3
; SSE-NEXT:    pandn %xmm1, %xmm3
; SSE-NEXT:    pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[3,3,3,3]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm9, %xmm1
; SSE-NEXT:    por %xmm3, %xmm1
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm1[2,2,2,2]
; SSE-NEXT:    punpckhdq {{.*#+}} xmm5 = xmm5[2],xmm2[2],xmm5[3],xmm2[3]
; SSE-NEXT:    movsd {{.*#+}} xmm5 = xmm0[0],xmm5[1]
; SSE-NEXT:    pshufd $229, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
; SSE-NEXT:    # xmm0 = mem[1,1,2,3]
; SSE-NEXT:    pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[3,3,3,3]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm13, %xmm1
; SSE-NEXT:    pandn %xmm0, %xmm13
; SSE-NEXT:    por %xmm1, %xmm13
; SSE-NEXT:    pshufd $250, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
; SSE-NEXT:    # xmm0 = mem[2,2,3,3]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    pand %xmm14, %xmm0
; SSE-NEXT:    pandn {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Folded Reload
; SSE-NEXT:    por %xmm0, %xmm14
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm14[1,1,1,1]
; SSE-NEXT:    punpckldq {{.*#+}} xmm13 = xmm13[0],xmm0[0],xmm13[1],xmm0[1]
; SSE-NEXT:    pshufd $250, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
; SSE-NEXT:    # xmm0 = mem[2,2,3,3]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    pand %xmm11, %xmm0
; SSE-NEXT:    pandn {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Folded Reload
; SSE-NEXT:    por %xmm0, %xmm11
; SSE-NEXT:    pshufd $244, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
; SSE-NEXT:    # xmm0 = mem[0,1,3,3]
; SSE-NEXT:    pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[3,3,3,3]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm9, %xmm1
; SSE-NEXT:    pandn %xmm0, %xmm9
; SSE-NEXT:    por %xmm1, %xmm9
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm9[2,2,2,2]
; SSE-NEXT:    punpckhdq {{.*#+}} xmm0 = xmm0[2],xmm11[2],xmm0[3],xmm11[3]
; SSE-NEXT:    movsd {{.*#+}} xmm0 = xmm13[0],xmm0[1]
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, (%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 16(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, (%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 16(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, (%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 16(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, (%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movaps %xmm1, 16(%r8)
; SSE-NEXT:    movapd %xmm10, (%r9)
; SSE-NEXT:    movapd %xmm6, 16(%r9)
; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT:    movapd %xmm4, (%rax)
; SSE-NEXT:    movapd %xmm8, 16(%rax)
; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT:    movapd %xmm12, (%rax)
; SSE-NEXT:    movapd %xmm15, 16(%rax)
; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT:    movapd %xmm0, 16(%rax)
; SSE-NEXT:    movapd %xmm5, (%rax)
; SSE-NEXT:    addq $904, %rsp # imm = 0x388
; SSE-NEXT:    retq
;
; AVX-LABEL: load_i8_stride8_vf32:
; AVX:       # %bb.0:
; AVX-NEXT:    subq $360, %rsp # imm = 0x168
; AVX-NEXT:    vbroadcastss {{.*#+}} xmm10 = [0,0,0,8,0,0,0,8,0,0,0,8,0,0,0,8]
; AVX-NEXT:    vmovdqa 112(%rdi), %xmm4
; AVX-NEXT:    vpshufb %xmm10, %xmm4, %xmm0
; AVX-NEXT:    vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovdqa 96(%rdi), %xmm1
; AVX-NEXT:    vpshufb %xmm10, %xmm1, %xmm3
; AVX-NEXT:    vmovdqa %xmm1, %xmm7
; AVX-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm3[0],xmm0[0],xmm3[1],xmm0[1],xmm3[2],xmm0[2],xmm3[3],xmm0[3]
; AVX-NEXT:    vbroadcastss {{.*#+}} xmm12 = [0,8,0,0,0,8,0,0,0,8,0,0,0,8,0,0]
; AVX-NEXT:    vmovdqa 80(%rdi), %xmm1
; AVX-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vpshufb %xmm12, %xmm1, %xmm3
; AVX-NEXT:    vmovdqa 64(%rdi), %xmm5
; AVX-NEXT:    vpshufb %xmm12, %xmm5, %xmm6
; AVX-NEXT:    vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm6[0],xmm3[0],xmm6[1],xmm3[1],xmm6[2],xmm3[2],xmm6[3],xmm3[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm8 = xmm3[0,1,2,3,4,5],xmm0[6,7]
; AVX-NEXT:    vmovd {{.*#+}} xmm14 = [0,0,0,8,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vmovdqa (%rdi), %xmm0
; AVX-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovdqa 16(%rdi), %xmm1
; AVX-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovdqa 32(%rdi), %xmm2
; AVX-NEXT:    vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovdqa 48(%rdi), %xmm3
; AVX-NEXT:    vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vpshufb %xmm14, %xmm3, %xmm9
; AVX-NEXT:    vpshufb %xmm14, %xmm2, %xmm11
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm11[0],xmm9[0],xmm11[1],xmm9[1],xmm11[2],xmm9[2],xmm11[3],xmm9[3]
; AVX-NEXT:    vmovd {{.*#+}} xmm3 = [0,8,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vpshufb %xmm3, %xmm1, %xmm11
; AVX-NEXT:    vpshufb %xmm3, %xmm0, %xmm13
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm13[0],xmm11[0],xmm13[1],xmm11[1],xmm13[2],xmm11[2],xmm13[3],xmm11[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm9 = xmm11[0,1],xmm9[2,3],xmm11[4,5,6,7]
; AVX-NEXT:    vpblendw {{.*#+}} xmm8 = xmm9[0,1,2,3],xmm8[4,5,6,7]
; AVX-NEXT:    vmovdqa 240(%rdi), %xmm0
; AVX-NEXT:    vmovdqa %xmm0, (%rsp) # 16-byte Spill
; AVX-NEXT:    vpshufb %xmm10, %xmm0, %xmm11
; AVX-NEXT:    vmovdqa 224(%rdi), %xmm0
; AVX-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vpshufb %xmm10, %xmm0, %xmm10
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm10[0],xmm11[0],xmm10[1],xmm11[1],xmm10[2],xmm11[2],xmm10[3],xmm11[3]
; AVX-NEXT:    vmovdqa 208(%rdi), %xmm0
; AVX-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vpshufb %xmm12, %xmm0, %xmm15
; AVX-NEXT:    vmovdqa 192(%rdi), %xmm0
; AVX-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vpshufb %xmm12, %xmm0, %xmm12
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm12[0],xmm15[0],xmm12[1],xmm15[1],xmm12[2],xmm15[2],xmm12[3],xmm15[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm2 = xmm12[0,1,2,3,4,5],xmm13[6,7]
; AVX-NEXT:    vmovdqa 176(%rdi), %xmm0
; AVX-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vpshufb %xmm14, %xmm0, %xmm15
; AVX-NEXT:    vmovdqa 160(%rdi), %xmm0
; AVX-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vpshufb %xmm14, %xmm0, %xmm14
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm14[0],xmm15[0],xmm14[1],xmm15[1],xmm14[2],xmm15[2],xmm14[3],xmm15[3]
; AVX-NEXT:    vmovdqa 144(%rdi), %xmm15
; AVX-NEXT:    vpshufb %xmm3, %xmm15, %xmm0
; AVX-NEXT:    vmovdqa %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovdqa 128(%rdi), %xmm6
; AVX-NEXT:    vpshufb %xmm3, %xmm6, %xmm3
; AVX-NEXT:    vmovdqa %xmm6, %xmm14
; AVX-NEXT:    vmovdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm3[0],xmm0[0],xmm3[1],xmm0[1],xmm3[2],xmm0[2],xmm3[3],xmm0[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3],xmm0[4,5,6,7]
; AVX-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm1
; AVX-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm8[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vbroadcastss {{.*#+}} xmm8 = [0,0,1,9,0,0,1,9,0,0,1,9,0,0,1,9]
; AVX-NEXT:    vpshufb %xmm8, %xmm4, %xmm1
; AVX-NEXT:    vpshufb %xmm8, %xmm7, %xmm2
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX-NEXT:    vbroadcastss {{.*#+}} xmm2 = [1,9,0,0,1,9,0,0,1,9,0,0,1,9,0,0]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm2, %xmm6, %xmm3
; AVX-NEXT:    vpshufb %xmm2, %xmm5, %xmm4
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm1 = xmm3[0,1,2,3,4,5],xmm1[6,7]
; AVX-NEXT:    vmovd {{.*#+}} xmm3 = [0,0,1,9,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm3, %xmm13, %xmm4
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm3, %xmm9, %xmm5
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX-NEXT:    vmovd {{.*#+}} xmm5 = [1,9,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm5, %xmm11, %xmm0
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm5, %xmm7, %xmm7
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm7[0],xmm0[0],xmm7[1],xmm0[1],xmm7[2],xmm0[2],xmm7[3],xmm0[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0,1],xmm4[2,3],xmm0[4,5,6,7]
; AVX-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0,1,2,3],xmm1[4,5,6,7]
; AVX-NEXT:    vmovdqa (%rsp), %xmm12 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm8, %xmm12, %xmm1
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm8, %xmm10, %xmm4
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm4[0],xmm1[0],xmm4[1],xmm1[1],xmm4[2],xmm1[2],xmm4[3],xmm1[3]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm2, %xmm4, %xmm4
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm2, %xmm7, %xmm2
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm1 = xmm2[0,1,2,3,4,5],xmm1[6,7]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm3, %xmm2, %xmm2
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm3, %xmm4, %xmm3
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
; AVX-NEXT:    vpshufb %xmm5, %xmm15, %xmm3
; AVX-NEXT:    vpshufb %xmm5, %xmm14, %xmm4
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm2 = xmm3[0,1],xmm2[2,3],xmm3[4,5,6,7]
; AVX-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm1
; AVX-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm2
; AVX-NEXT:    vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vbroadcastss {{.*#+}} xmm0 = [0,0,2,10,0,0,2,10,0,0,2,10,0,0,2,10]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm0, %xmm14, %xmm1
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm0, %xmm15, %xmm2
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX-NEXT:    vbroadcastss {{.*#+}} xmm2 = [2,10,0,0,2,10,0,0,2,10,0,0,2,10,0,0]
; AVX-NEXT:    vpshufb %xmm2, %xmm6, %xmm3
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm2, %xmm8, %xmm4
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm1 = xmm3[0,1,2,3,4,5],xmm1[6,7]
; AVX-NEXT:    vmovd {{.*#+}} xmm3 = [0,0,2,10,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vpshufb %xmm3, %xmm13, %xmm4
; AVX-NEXT:    vpshufb %xmm3, %xmm9, %xmm5
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX-NEXT:    vmovd {{.*#+}} xmm5 = [2,10,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vpshufb %xmm5, %xmm11, %xmm6
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm5, %xmm13, %xmm7
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm4 = xmm6[0,1],xmm4[2,3],xmm6[4,5,6,7]
; AVX-NEXT:    vpblendw {{.*#+}} xmm1 = xmm4[0,1,2,3],xmm1[4,5,6,7]
; AVX-NEXT:    vpshufb %xmm0, %xmm12, %xmm4
; AVX-NEXT:    vpshufb %xmm0, %xmm10, %xmm0
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm4[0],xmm0[1],xmm4[1],xmm0[2],xmm4[2],xmm0[3],xmm4[3]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm2, %xmm10, %xmm4
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm2, %xmm11, %xmm2
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm0 = xmm2[0,1,2,3,4,5],xmm0[6,7]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm3, %xmm12, %xmm2
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm3, %xmm4, %xmm3
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm5, %xmm3, %xmm3
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm5, %xmm4, %xmm4
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm2 = xmm3[0,1],xmm2[2,3],xmm3[4,5,6,7]
; AVX-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
; AVX-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm2
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3,4,5],ymm0[6,7]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vbroadcastss {{.*#+}} xmm0 = [0,0,3,11,0,0,3,11,0,0,3,11,0,0,3,11]
; AVX-NEXT:    vpshufb %xmm0, %xmm14, %xmm1
; AVX-NEXT:    vpshufb %xmm0, %xmm15, %xmm2
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX-NEXT:    vbroadcastss {{.*#+}} xmm2 = [3,11,0,0,3,11,0,0,3,11,0,0,3,11,0,0]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm2, %xmm3, %xmm3
; AVX-NEXT:    vpshufb %xmm2, %xmm8, %xmm4
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm1 = xmm3[0,1,2,3,4,5],xmm1[6,7]
; AVX-NEXT:    vmovd {{.*#+}} xmm3 = [0,0,3,11,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm3, %xmm9, %xmm4
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm3, %xmm5, %xmm5
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX-NEXT:    vmovd {{.*#+}} xmm5 = [3,11,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm5, %xmm6, %xmm6
; AVX-NEXT:    vmovdqa %xmm13, %xmm8
; AVX-NEXT:    vpshufb %xmm5, %xmm13, %xmm7
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm4 = xmm6[0,1],xmm4[2,3],xmm6[4,5,6,7]
; AVX-NEXT:    vpblendw {{.*#+}} xmm1 = xmm4[0,1,2,3],xmm1[4,5,6,7]
; AVX-NEXT:    vmovdqa (%rsp), %xmm4 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm0, %xmm4, %xmm4
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm0, %xmm6, %xmm0
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm4[0],xmm0[1],xmm4[1],xmm0[2],xmm4[2],xmm0[3],xmm4[3]
; AVX-NEXT:    vpshufb %xmm2, %xmm10, %xmm4
; AVX-NEXT:    vpshufb %xmm2, %xmm11, %xmm2
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm0 = xmm2[0,1,2,3,4,5],xmm0[6,7]
; AVX-NEXT:    vpshufb %xmm3, %xmm12, %xmm2
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm3, %xmm14, %xmm3
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm5, %xmm12, %xmm3
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm5, %xmm15, %xmm4
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm2 = xmm3[0,1],xmm2[2,3],xmm3[4,5,6,7]
; AVX-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
; AVX-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm2
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3,4,5],ymm0[6,7]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vbroadcastss {{.*#+}} xmm0 = [0,0,4,12,0,0,4,12,0,0,4,12,0,0,4,12]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm0, %xmm10, %xmm1
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm0, %xmm2, %xmm2
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX-NEXT:    vbroadcastss {{.*#+}} xmm2 = [4,12,0,0,4,12,0,0,4,12,0,0,4,12,0,0]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm2, %xmm11, %xmm3
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm2, %xmm13, %xmm4
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm1 = xmm3[0,1,2,3,4,5],xmm1[6,7]
; AVX-NEXT:    vmovd {{.*#+}} xmm3 = [0,0,4,12,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vpshufb %xmm3, %xmm9, %xmm4
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm3, %xmm5, %xmm5
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX-NEXT:    vmovd {{.*#+}} xmm5 = [4,12,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm5, %xmm6, %xmm6
; AVX-NEXT:    vpshufb %xmm5, %xmm8, %xmm7
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm4 = xmm6[0,1],xmm4[2,3],xmm6[4,5,6,7]
; AVX-NEXT:    vpblendw {{.*#+}} xmm1 = xmm4[0,1,2,3],xmm1[4,5,6,7]
; AVX-NEXT:    vmovdqa (%rsp), %xmm9 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm0, %xmm9, %xmm4
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm0, %xmm6, %xmm0
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm4[0],xmm0[1],xmm4[1],xmm0[2],xmm4[2],xmm0[3],xmm4[3]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm2, %xmm4, %xmm4
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm2, %xmm6, %xmm2
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm0 = xmm2[0,1,2,3,4,5],xmm0[6,7]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm3, %xmm2, %xmm2
; AVX-NEXT:    vpshufb %xmm3, %xmm14, %xmm3
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
; AVX-NEXT:    vpshufb %xmm5, %xmm12, %xmm3
; AVX-NEXT:    vpshufb %xmm5, %xmm15, %xmm4
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm2 = xmm3[0,1],xmm2[2,3],xmm3[4,5,6,7]
; AVX-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
; AVX-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm2
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3,4,5],ymm0[6,7]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vbroadcastss {{.*#+}} xmm0 = [0,0,5,13,0,0,5,13,0,0,5,13,0,0,5,13]
; AVX-NEXT:    vpshufb %xmm0, %xmm10, %xmm1
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm0, %xmm8, %xmm2
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX-NEXT:    vbroadcastss {{.*#+}} xmm2 = [5,13,0,0,5,13,0,0,5,13,0,0,5,13,0,0]
; AVX-NEXT:    vpshufb %xmm2, %xmm11, %xmm3
; AVX-NEXT:    vpshufb %xmm2, %xmm13, %xmm4
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm1 = xmm3[0,1,2,3,4,5],xmm1[6,7]
; AVX-NEXT:    vmovd {{.*#+}} xmm3 = [0,0,5,13,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm3, %xmm4, %xmm4
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm3, %xmm11, %xmm5
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX-NEXT:    vmovd {{.*#+}} xmm5 = [5,13,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm5, %xmm10, %xmm6
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm5, %xmm14, %xmm7
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm4 = xmm6[0,1],xmm4[2,3],xmm6[4,5,6,7]
; AVX-NEXT:    vpblendw {{.*#+}} xmm1 = xmm4[0,1,2,3],xmm1[4,5,6,7]
; AVX-NEXT:    vpshufb %xmm0, %xmm9, %xmm4
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm0, %xmm15, %xmm0
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm4[0],xmm0[1],xmm4[1],xmm0[2],xmm4[2],xmm0[3],xmm4[3]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm2, %xmm9, %xmm4
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm2, %xmm12, %xmm2
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm0 = xmm2[0,1,2,3,4,5],xmm0[6,7]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm3, %xmm13, %xmm2
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm3, %xmm4, %xmm3
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm5, %xmm3, %xmm3
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm5, %xmm4, %xmm4
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm2 = xmm3[0,1],xmm2[2,3],xmm3[4,5,6,7]
; AVX-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
; AVX-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm2
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3,4,5],ymm0[6,7]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vbroadcastss {{.*#+}} xmm0 = [0,0,6,14,0,0,6,14,0,0,6,14,0,0,6,14]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm0, %xmm1, %xmm1
; AVX-NEXT:    vpshufb %xmm0, %xmm8, %xmm2
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX-NEXT:    vbroadcastss {{.*#+}} xmm2 = [6,14,0,0,6,14,0,0,6,14,0,0,6,14,0,0]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm2, %xmm3, %xmm3
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm2, %xmm4, %xmm4
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm1 = xmm3[0,1,2,3,4,5],xmm1[6,7]
; AVX-NEXT:    vmovd {{.*#+}} xmm3 = [0,0,6,14,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm3, %xmm8, %xmm4
; AVX-NEXT:    vpshufb %xmm3, %xmm11, %xmm5
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX-NEXT:    vmovd {{.*#+}} xmm5 = [6,14,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vpshufb %xmm5, %xmm10, %xmm6
; AVX-NEXT:    vpshufb %xmm5, %xmm14, %xmm7
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm4 = xmm6[0,1],xmm4[2,3],xmm6[4,5,6,7]
; AVX-NEXT:    vpblendw {{.*#+}} xmm1 = xmm4[0,1,2,3],xmm1[4,5,6,7]
; AVX-NEXT:    vmovdqa (%rsp), %xmm14 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm0, %xmm14, %xmm4
; AVX-NEXT:    vmovdqa %xmm15, %xmm10
; AVX-NEXT:    vpshufb %xmm0, %xmm15, %xmm0
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm4[0],xmm0[1],xmm4[1],xmm0[2],xmm4[2],xmm0[3],xmm4[3]
; AVX-NEXT:    vpshufb %xmm2, %xmm9, %xmm4
; AVX-NEXT:    vpshufb %xmm2, %xmm12, %xmm2
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm0 = xmm2[0,1,2,3,4,5],xmm0[6,7]
; AVX-NEXT:    vpshufb %xmm3, %xmm13, %xmm2
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm3, %xmm15, %xmm3
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm5, %xmm9, %xmm3
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm5, %xmm11, %xmm4
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm2 = xmm3[0,1],xmm2[2,3],xmm3[4,5,6,7]
; AVX-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
; AVX-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm2
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3,4,5],ymm0[6,7]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vbroadcastss {{.*#+}} xmm0 = [0,0,7,15,0,0,7,15,0,0,7,15,0,0,7,15]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm0, %xmm1, %xmm1
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm0, %xmm2, %xmm2
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX-NEXT:    vbroadcastss {{.*#+}} xmm2 = [7,15,0,0,7,15,0,0,7,15,0,0,7,15,0,0]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm2, %xmm3, %xmm3
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm2, %xmm4, %xmm4
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm1 = xmm3[0,1,2,3,4,5],xmm1[6,7]
; AVX-NEXT:    vmovd {{.*#+}} xmm3 = [0,0,7,15,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vpshufb %xmm3, %xmm8, %xmm4
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm3, %xmm5, %xmm5
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX-NEXT:    vmovd {{.*#+}} xmm5 = [7,15,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm5, %xmm6, %xmm6
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm5, %xmm7, %xmm7
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm4 = xmm6[0,1],xmm4[2,3],xmm6[4,5,6,7]
; AVX-NEXT:    vpblendw {{.*#+}} xmm1 = xmm4[0,1,2,3],xmm1[4,5,6,7]
; AVX-NEXT:    vpshufb %xmm0, %xmm14, %xmm4
; AVX-NEXT:    vpshufb %xmm0, %xmm10, %xmm0
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm4[0],xmm0[1],xmm4[1],xmm0[2],xmm4[2],xmm0[3],xmm4[3]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm2, %xmm4, %xmm4
; AVX-NEXT:    vpshufb %xmm2, %xmm12, %xmm2
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm0 = xmm2[0,1,2,3,4,5],xmm0[6,7]
; AVX-NEXT:    vpshufb %xmm3, %xmm13, %xmm2
; AVX-NEXT:    vpshufb %xmm3, %xmm15, %xmm3
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
; AVX-NEXT:    vpshufb %xmm5, %xmm9, %xmm3
; AVX-NEXT:    vpshufb %xmm5, %xmm11, %xmm4
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm2 = xmm3[0,1],xmm2[2,3],xmm3[4,5,6,7]
; AVX-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
; AVX-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm2
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3,4,5],ymm0[6,7]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, (%rsi)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, (%rdx)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, (%rcx)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, (%r8)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, (%r9)
; AVX-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, (%rax)
; AVX-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, (%rax)
; AVX-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX-NEXT:    vmovaps %ymm0, (%rax)
; AVX-NEXT:    addq $360, %rsp # imm = 0x168
; AVX-NEXT:    vzeroupper
; AVX-NEXT:    retq
;
; AVX2-LABEL: load_i8_stride8_vf32:
; AVX2:       # %bb.0:
; AVX2-NEXT:    subq $360, %rsp # imm = 0x168
; AVX2-NEXT:    vmovdqa 112(%rdi), %xmm0
; AVX2-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm10 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX2-NEXT:    vpshufb %xmm10, %xmm0, %xmm0
; AVX2-NEXT:    vmovdqa 96(%rdi), %xmm4
; AVX2-NEXT:    vpshufb %xmm10, %xmm4, %xmm2
; AVX2-NEXT:    vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
; AVX2-NEXT:    vmovdqa 80(%rdi), %xmm1
; AVX2-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm12 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX2-NEXT:    vpshufb %xmm12, %xmm1, %xmm2
; AVX2-NEXT:    vmovdqa 64(%rdi), %xmm5
; AVX2-NEXT:    vpshufb %xmm12, %xmm5, %xmm6
; AVX2-NEXT:    vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm6[0],xmm2[0],xmm6[1],xmm2[1],xmm6[2],xmm2[2],xmm6[3],xmm2[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm8 = xmm2[0,1,2],xmm0[3]
; AVX2-NEXT:    vmovdqa (%rdi), %xmm0
; AVX2-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-NEXT:    vmovdqa 16(%rdi), %xmm7
; AVX2-NEXT:    vmovdqa 32(%rdi), %xmm2
; AVX2-NEXT:    vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-NEXT:    vmovdqa 48(%rdi), %xmm3
; AVX2-NEXT:    vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm14 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX2-NEXT:    vpshufb %xmm14, %xmm3, %xmm9
; AVX2-NEXT:    vpshufb %xmm14, %xmm2, %xmm11
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm11[0],xmm9[0],xmm11[1],xmm9[1],xmm11[2],xmm9[2],xmm11[3],xmm9[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm3 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX2-NEXT:    vpshufb %xmm3, %xmm7, %xmm11
; AVX2-NEXT:    vmovdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-NEXT:    vpshufb %xmm3, %xmm0, %xmm13
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm13[0],xmm11[0],xmm13[1],xmm11[1],xmm13[2],xmm11[2],xmm13[3],xmm11[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm9 = xmm11[0],xmm9[1],xmm11[2,3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm8 = xmm9[0,1],xmm8[2,3]
; AVX2-NEXT:    vmovdqa 240(%rdi), %xmm0
; AVX2-NEXT:    vmovdqa %xmm0, (%rsp) # 16-byte Spill
; AVX2-NEXT:    vpshufb %xmm10, %xmm0, %xmm11
; AVX2-NEXT:    vmovdqa 224(%rdi), %xmm0
; AVX2-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-NEXT:    vpshufb %xmm10, %xmm0, %xmm10
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm10[0],xmm11[0],xmm10[1],xmm11[1],xmm10[2],xmm11[2],xmm10[3],xmm11[3]
; AVX2-NEXT:    vmovdqa 208(%rdi), %xmm11
; AVX2-NEXT:    vpshufb %xmm12, %xmm11, %xmm15
; AVX2-NEXT:    vmovdqa %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-NEXT:    vmovdqa 192(%rdi), %xmm0
; AVX2-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-NEXT:    vpshufb %xmm12, %xmm0, %xmm12
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm12[0],xmm15[0],xmm12[1],xmm15[1],xmm12[2],xmm15[2],xmm12[3],xmm15[3]
; AVX2-NEXT:    vinserti128 $1, %xmm13, %ymm0, %ymm13
; AVX2-NEXT:    vinserti128 $1, %xmm12, %ymm0, %ymm12
; AVX2-NEXT:    vpblendd {{.*#+}} ymm2 = ymm12[0,1,2,3,4,5,6],ymm13[7]
; AVX2-NEXT:    vmovdqa 176(%rdi), %xmm0
; AVX2-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-NEXT:    vpshufb %xmm14, %xmm0, %xmm15
; AVX2-NEXT:    vmovdqa 160(%rdi), %xmm0
; AVX2-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-NEXT:    vpshufb %xmm14, %xmm0, %xmm14
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm14[0],xmm15[0],xmm14[1],xmm15[1],xmm14[2],xmm15[2],xmm14[3],xmm15[3]
; AVX2-NEXT:    vmovdqa 144(%rdi), %xmm0
; AVX2-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-NEXT:    vpshufb %xmm3, %xmm0, %xmm0
; AVX2-NEXT:    vmovdqa 128(%rdi), %xmm13
; AVX2-NEXT:    vpshufb %xmm3, %xmm13, %xmm3
; AVX2-NEXT:    vmovdqa %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm3[0],xmm0[0],xmm3[1],xmm0[1],xmm3[2],xmm0[2],xmm3[3],xmm0[3]
; AVX2-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm1
; AVX2-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX2-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5],ymm0[6,7]
; AVX2-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm2[6,7]
; AVX2-NEXT:    vpblendd {{.*#+}} ymm0 = ymm8[0,1,2,3],ymm0[4,5,6,7]
; AVX2-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm8 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm8, %xmm0, %xmm1
; AVX2-NEXT:    vpshufb %xmm8, %xmm4, %xmm2
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm2 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm2, %xmm6, %xmm3
; AVX2-NEXT:    vpshufb %xmm2, %xmm5, %xmm4
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm1 = xmm3[0,1,2],xmm1[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm3 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm3, %xmm15, %xmm4
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm3, %xmm9, %xmm5
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm5 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX2-NEXT:    vpshufb %xmm5, %xmm7, %xmm0
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm5, %xmm14, %xmm7
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm7[0],xmm0[0],xmm7[1],xmm0[1],xmm7[2],xmm0[2],xmm7[3],xmm0[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm0 = xmm0[0],xmm4[1],xmm0[2,3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]
; AVX2-NEXT:    vmovdqa (%rsp), %xmm12 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm8, %xmm12, %xmm1
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm8, %xmm10, %xmm4
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm4[0],xmm1[0],xmm4[1],xmm1[1],xmm4[2],xmm1[2],xmm4[3],xmm1[3]
; AVX2-NEXT:    vpshufb %xmm2, %xmm11, %xmm4
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm2, %xmm11, %xmm2
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3]
; AVX2-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm1
; AVX2-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX2-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5,6],ymm1[7]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm3, %xmm2, %xmm2
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm3, %xmm4, %xmm3
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm5, %xmm3, %xmm3
; AVX2-NEXT:    vpshufb %xmm5, %xmm13, %xmm4
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX2-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX2-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX2-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4],ymm2[5],ymm3[6,7]
; AVX2-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]
; AVX2-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX2-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm0 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm0, %xmm1, %xmm1
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm0, %xmm13, %xmm2
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm2 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX2-NEXT:    vpshufb %xmm2, %xmm6, %xmm3
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm2, %xmm8, %xmm4
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm1 = xmm3[0,1,2],xmm1[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm3 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX2-NEXT:    vpshufb %xmm3, %xmm15, %xmm4
; AVX2-NEXT:    vpshufb %xmm3, %xmm9, %xmm5
; AVX2-NEXT:    vmovdqa %xmm9, %xmm15
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm5 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm5, %xmm6, %xmm6
; AVX2-NEXT:    vpshufb %xmm5, %xmm14, %xmm7
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm4 = xmm6[0],xmm4[1],xmm6[2,3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm1 = xmm4[0,1],xmm1[2,3]
; AVX2-NEXT:    vpshufb %xmm0, %xmm12, %xmm4
; AVX2-NEXT:    vpshufb %xmm0, %xmm10, %xmm0
; AVX2-NEXT:    vmovdqa %xmm10, %xmm14
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm4[0],xmm0[1],xmm4[1],xmm0[2],xmm4[2],xmm0[3],xmm4[3]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm2, %xmm10, %xmm4
; AVX2-NEXT:    vpshufb %xmm2, %xmm11, %xmm2
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3]
; AVX2-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX2-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX2-NEXT:    vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3,4,5,6],ymm0[7]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm3, %xmm12, %xmm2
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm3, %xmm4, %xmm3
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm5, %xmm3, %xmm3
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm5, %xmm4, %xmm4
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX2-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX2-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX2-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4],ymm2[5],ymm3[6,7]
; AVX2-NEXT:    vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3,4,5],ymm0[6,7]
; AVX2-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm0 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm0, %xmm1, %xmm1
; AVX2-NEXT:    vpshufb %xmm0, %xmm13, %xmm2
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm2 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm2, %xmm3, %xmm3
; AVX2-NEXT:    vpshufb %xmm2, %xmm8, %xmm4
; AVX2-NEXT:    vmovdqa %xmm8, %xmm13
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm1 = xmm3[0,1,2],xmm1[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm3 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm3, %xmm9, %xmm4
; AVX2-NEXT:    vpshufb %xmm3, %xmm15, %xmm5
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm5 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm5, %xmm11, %xmm6
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm5, %xmm15, %xmm7
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm4 = xmm6[0],xmm4[1],xmm6[2,3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm1 = xmm4[0,1],xmm1[2,3]
; AVX2-NEXT:    vmovdqa (%rsp), %xmm4 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm0, %xmm4, %xmm4
; AVX2-NEXT:    vpshufb %xmm0, %xmm14, %xmm0
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm4[0],xmm0[1],xmm4[1],xmm0[2],xmm4[2],xmm0[3],xmm4[3]
; AVX2-NEXT:    vpshufb %xmm2, %xmm10, %xmm4
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm2, %xmm8, %xmm2
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3]
; AVX2-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX2-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX2-NEXT:    vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3,4,5,6],ymm0[7]
; AVX2-NEXT:    vpshufb %xmm3, %xmm12, %xmm2
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm3, %xmm14, %xmm3
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm5, %xmm3, %xmm3
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm5, %xmm4, %xmm4
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX2-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX2-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX2-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4],ymm2[5],ymm3[6,7]
; AVX2-NEXT:    vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3,4,5],ymm0[6,7]
; AVX2-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm0 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm0, %xmm10, %xmm1
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm0, %xmm2, %xmm2
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm2 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm2, %xmm12, %xmm3
; AVX2-NEXT:    vpshufb %xmm2, %xmm13, %xmm4
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm1 = xmm3[0,1,2],xmm1[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm3 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX2-NEXT:    vpshufb %xmm3, %xmm9, %xmm4
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm3, %xmm13, %xmm5
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm5 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX2-NEXT:    vpshufb %xmm5, %xmm11, %xmm6
; AVX2-NEXT:    vpshufb %xmm5, %xmm15, %xmm7
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm4 = xmm6[0],xmm4[1],xmm6[2,3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm1 = xmm4[0,1],xmm1[2,3]
; AVX2-NEXT:    vmovdqa (%rsp), %xmm9 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm0, %xmm9, %xmm4
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm0, %xmm6, %xmm0
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm4[0],xmm0[1],xmm4[1],xmm0[2],xmm4[2],xmm0[3],xmm4[3]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm2, %xmm4, %xmm4
; AVX2-NEXT:    vpshufb %xmm2, %xmm8, %xmm2
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3]
; AVX2-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX2-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX2-NEXT:    vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3,4,5,6],ymm0[7]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm3, %xmm15, %xmm2
; AVX2-NEXT:    vpshufb %xmm3, %xmm14, %xmm3
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm5, %xmm14, %xmm3
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm5, %xmm4, %xmm4
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX2-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX2-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX2-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4],ymm2[5],ymm3[6,7]
; AVX2-NEXT:    vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3,4,5],ymm0[6,7]
; AVX2-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm0 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX2-NEXT:    vpshufb %xmm0, %xmm10, %xmm1
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm0, %xmm8, %xmm2
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm2 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX2-NEXT:    vmovdqa %xmm12, %xmm11
; AVX2-NEXT:    vpshufb %xmm2, %xmm12, %xmm3
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm2, %xmm4, %xmm4
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm1 = xmm3[0,1,2],xmm1[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm3 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm3, %xmm4, %xmm4
; AVX2-NEXT:    vpshufb %xmm3, %xmm13, %xmm5
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm5 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm5, %xmm6, %xmm6
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm5, %xmm7, %xmm7
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm4 = xmm6[0],xmm4[1],xmm6[2,3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm1 = xmm4[0,1],xmm1[2,3]
; AVX2-NEXT:    vpshufb %xmm0, %xmm9, %xmm4
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm0, %xmm9, %xmm0
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm4[0],xmm0[1],xmm4[1],xmm0[2],xmm4[2],xmm0[3],xmm4[3]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm2, %xmm13, %xmm4
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm2, %xmm12, %xmm2
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3]
; AVX2-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX2-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX2-NEXT:    vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3,4,5,6],ymm0[7]
; AVX2-NEXT:    vpshufb %xmm3, %xmm15, %xmm2
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm3, %xmm4, %xmm3
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
; AVX2-NEXT:    vpshufb %xmm5, %xmm14, %xmm3
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm5, %xmm10, %xmm4
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX2-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX2-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX2-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4],ymm2[5],ymm3[6,7]
; AVX2-NEXT:    vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3,4,5],ymm0[6,7]
; AVX2-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm0 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm0, %xmm1, %xmm1
; AVX2-NEXT:    vpshufb %xmm0, %xmm8, %xmm2
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm2 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX2-NEXT:    vpshufb %xmm2, %xmm11, %xmm3
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm2, %xmm14, %xmm4
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm1 = xmm3[0,1,2],xmm1[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm3 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm3, %xmm11, %xmm4
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm3, %xmm5, %xmm5
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm5 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm5, %xmm6, %xmm6
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm5, %xmm8, %xmm7
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm4 = xmm6[0],xmm4[1],xmm6[2,3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm1 = xmm4[0,1],xmm1[2,3]
; AVX2-NEXT:    vmovdqa (%rsp), %xmm15 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm0, %xmm15, %xmm4
; AVX2-NEXT:    vpshufb %xmm0, %xmm9, %xmm0
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm4[0],xmm0[1],xmm4[1],xmm0[2],xmm4[2],xmm0[3],xmm4[3]
; AVX2-NEXT:    vpshufb %xmm2, %xmm13, %xmm4
; AVX2-NEXT:    vpshufb %xmm2, %xmm12, %xmm2
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3]
; AVX2-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX2-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX2-NEXT:    vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3,4,5,6],ymm0[7]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm3, %xmm13, %xmm2
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm3, %xmm9, %xmm3
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm5, %xmm12, %xmm3
; AVX2-NEXT:    vpshufb %xmm5, %xmm10, %xmm4
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX2-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX2-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX2-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4],ymm2[5],ymm3[6,7]
; AVX2-NEXT:    vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3,4,5],ymm0[6,7]
; AVX2-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm0 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm0, %xmm1, %xmm1
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm0, %xmm2, %xmm2
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm2 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm2, %xmm3, %xmm3
; AVX2-NEXT:    vpshufb %xmm2, %xmm14, %xmm4
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm1 = xmm3[0,1,2],xmm1[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm3 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX2-NEXT:    vpshufb %xmm3, %xmm11, %xmm4
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm3, %xmm5, %xmm5
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm5 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm5, %xmm6, %xmm6
; AVX2-NEXT:    vpshufb %xmm5, %xmm8, %xmm7
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm4 = xmm6[0],xmm4[1],xmm6[2,3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm1 = xmm4[0,1],xmm1[2,3]
; AVX2-NEXT:    vpshufb %xmm0, %xmm15, %xmm4
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm0, %xmm6, %xmm0
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm4[0],xmm0[1],xmm4[1],xmm0[2],xmm4[2],xmm0[3],xmm4[3]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm2, %xmm4, %xmm4
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm2, %xmm6, %xmm2
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3]
; AVX2-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX2-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX2-NEXT:    vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3,4,5,6],ymm0[7]
; AVX2-NEXT:    vpshufb %xmm3, %xmm13, %xmm2
; AVX2-NEXT:    vpshufb %xmm3, %xmm9, %xmm3
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
; AVX2-NEXT:    vpshufb %xmm5, %xmm12, %xmm3
; AVX2-NEXT:    vpshufb %xmm5, %xmm10, %xmm4
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX2-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX2-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX2-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4],ymm2[5],ymm3[6,7]
; AVX2-NEXT:    vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3,4,5],ymm0[6,7]
; AVX2-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, (%rsi)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, (%rdx)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, (%rcx)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, (%r8)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, (%r9)
; AVX2-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, (%rax)
; AVX2-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, (%rax)
; AVX2-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-NEXT:    vmovdqa %ymm0, (%rax)
; AVX2-NEXT:    addq $360, %rsp # imm = 0x168
; AVX2-NEXT:    vzeroupper
; AVX2-NEXT:    retq
;
; AVX2-FP-LABEL: load_i8_stride8_vf32:
; AVX2-FP:       # %bb.0:
; AVX2-FP-NEXT:    subq $360, %rsp # imm = 0x168
; AVX2-FP-NEXT:    vmovdqa 112(%rdi), %xmm0
; AVX2-FP-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm10 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX2-FP-NEXT:    vpshufb %xmm10, %xmm0, %xmm0
; AVX2-FP-NEXT:    vmovdqa 96(%rdi), %xmm4
; AVX2-FP-NEXT:    vpshufb %xmm10, %xmm4, %xmm2
; AVX2-FP-NEXT:    vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
; AVX2-FP-NEXT:    vmovdqa 80(%rdi), %xmm1
; AVX2-FP-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm12 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX2-FP-NEXT:    vpshufb %xmm12, %xmm1, %xmm2
; AVX2-FP-NEXT:    vmovdqa 64(%rdi), %xmm5
; AVX2-FP-NEXT:    vpshufb %xmm12, %xmm5, %xmm6
; AVX2-FP-NEXT:    vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm6[0],xmm2[0],xmm6[1],xmm2[1],xmm6[2],xmm2[2],xmm6[3],xmm2[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm8 = xmm2[0,1,2],xmm0[3]
; AVX2-FP-NEXT:    vmovdqa (%rdi), %xmm0
; AVX2-FP-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FP-NEXT:    vmovdqa 16(%rdi), %xmm7
; AVX2-FP-NEXT:    vmovdqa 32(%rdi), %xmm2
; AVX2-FP-NEXT:    vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FP-NEXT:    vmovdqa 48(%rdi), %xmm3
; AVX2-FP-NEXT:    vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm14 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX2-FP-NEXT:    vpshufb %xmm14, %xmm3, %xmm9
; AVX2-FP-NEXT:    vpshufb %xmm14, %xmm2, %xmm11
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm11[0],xmm9[0],xmm11[1],xmm9[1],xmm11[2],xmm9[2],xmm11[3],xmm9[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm3 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm7, %xmm11
; AVX2-FP-NEXT:    vmovdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm0, %xmm13
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm13[0],xmm11[0],xmm13[1],xmm11[1],xmm13[2],xmm11[2],xmm13[3],xmm11[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm9 = xmm11[0],xmm9[1],xmm11[2,3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm8 = xmm9[0,1],xmm8[2,3]
; AVX2-FP-NEXT:    vmovdqa 240(%rdi), %xmm0
; AVX2-FP-NEXT:    vmovdqa %xmm0, (%rsp) # 16-byte Spill
; AVX2-FP-NEXT:    vpshufb %xmm10, %xmm0, %xmm11
; AVX2-FP-NEXT:    vmovdqa 224(%rdi), %xmm0
; AVX2-FP-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FP-NEXT:    vpshufb %xmm10, %xmm0, %xmm10
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm10[0],xmm11[0],xmm10[1],xmm11[1],xmm10[2],xmm11[2],xmm10[3],xmm11[3]
; AVX2-FP-NEXT:    vmovdqa 208(%rdi), %xmm11
; AVX2-FP-NEXT:    vpshufb %xmm12, %xmm11, %xmm15
; AVX2-FP-NEXT:    vmovdqa %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FP-NEXT:    vmovdqa 192(%rdi), %xmm0
; AVX2-FP-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FP-NEXT:    vpshufb %xmm12, %xmm0, %xmm12
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm12[0],xmm15[0],xmm12[1],xmm15[1],xmm12[2],xmm15[2],xmm12[3],xmm15[3]
; AVX2-FP-NEXT:    vinserti128 $1, %xmm13, %ymm0, %ymm13
; AVX2-FP-NEXT:    vinserti128 $1, %xmm12, %ymm0, %ymm12
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm2 = ymm12[0,1,2,3,4,5,6],ymm13[7]
; AVX2-FP-NEXT:    vmovdqa 176(%rdi), %xmm0
; AVX2-FP-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FP-NEXT:    vpshufb %xmm14, %xmm0, %xmm15
; AVX2-FP-NEXT:    vmovdqa 160(%rdi), %xmm0
; AVX2-FP-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FP-NEXT:    vpshufb %xmm14, %xmm0, %xmm14
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm14[0],xmm15[0],xmm14[1],xmm15[1],xmm14[2],xmm15[2],xmm14[3],xmm15[3]
; AVX2-FP-NEXT:    vmovdqa 144(%rdi), %xmm0
; AVX2-FP-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm0, %xmm0
; AVX2-FP-NEXT:    vmovdqa 128(%rdi), %xmm13
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm13, %xmm3
; AVX2-FP-NEXT:    vmovdqa %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm3[0],xmm0[0],xmm3[1],xmm0[1],xmm3[2],xmm0[2],xmm3[3],xmm0[3]
; AVX2-FP-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm1
; AVX2-FP-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5],ymm0[6,7]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm2[6,7]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm8[0,1,2,3],ymm0[4,5,6,7]
; AVX2-FP-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm8 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm8, %xmm0, %xmm1
; AVX2-FP-NEXT:    vpshufb %xmm8, %xmm4, %xmm2
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm2 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm6, %xmm3
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm5, %xmm4
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm1 = xmm3[0,1,2],xmm1[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm3 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm15, %xmm4
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm9, %xmm5
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm5 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX2-FP-NEXT:    vpshufb %xmm5, %xmm7, %xmm0
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm5, %xmm14, %xmm7
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm7[0],xmm0[0],xmm7[1],xmm0[1],xmm7[2],xmm0[2],xmm7[3],xmm0[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm0[0],xmm4[1],xmm0[2,3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]
; AVX2-FP-NEXT:    vmovdqa (%rsp), %xmm12 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm8, %xmm12, %xmm1
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm8, %xmm10, %xmm4
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm4[0],xmm1[0],xmm4[1],xmm1[1],xmm4[2],xmm1[2],xmm4[3],xmm1[3]
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm11, %xmm4
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm11, %xmm2
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3]
; AVX2-FP-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm1
; AVX2-FP-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5,6],ymm1[7]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm2, %xmm2
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm4, %xmm3
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm5, %xmm3, %xmm3
; AVX2-FP-NEXT:    vpshufb %xmm5, %xmm13, %xmm4
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX2-FP-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX2-FP-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4],ymm2[5],ymm3[6,7]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX2-FP-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm0 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm1, %xmm1
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm13, %xmm2
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm2 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm6, %xmm3
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm8, %xmm4
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm1 = xmm3[0,1,2],xmm1[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm3 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm15, %xmm4
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm9, %xmm5
; AVX2-FP-NEXT:    vmovdqa %xmm9, %xmm15
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm5 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm5, %xmm6, %xmm6
; AVX2-FP-NEXT:    vpshufb %xmm5, %xmm14, %xmm7
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm4 = xmm6[0],xmm4[1],xmm6[2,3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm1 = xmm4[0,1],xmm1[2,3]
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm12, %xmm4
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm10, %xmm0
; AVX2-FP-NEXT:    vmovdqa %xmm10, %xmm14
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm4[0],xmm0[1],xmm4[1],xmm0[2],xmm4[2],xmm0[3],xmm4[3]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm10, %xmm4
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm11, %xmm2
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3]
; AVX2-FP-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX2-FP-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3,4,5,6],ymm0[7]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm12, %xmm2
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm4, %xmm3
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm5, %xmm3, %xmm3
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm5, %xmm4, %xmm4
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX2-FP-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX2-FP-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4],ymm2[5],ymm3[6,7]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3,4,5],ymm0[6,7]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-FP-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm0 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm1, %xmm1
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm13, %xmm2
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm2 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm3, %xmm3
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm8, %xmm4
; AVX2-FP-NEXT:    vmovdqa %xmm8, %xmm13
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm1 = xmm3[0,1,2],xmm1[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm3 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm9, %xmm4
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm15, %xmm5
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm5 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm5, %xmm11, %xmm6
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm5, %xmm15, %xmm7
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm4 = xmm6[0],xmm4[1],xmm6[2,3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm1 = xmm4[0,1],xmm1[2,3]
; AVX2-FP-NEXT:    vmovdqa (%rsp), %xmm4 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm4, %xmm4
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm14, %xmm0
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm4[0],xmm0[1],xmm4[1],xmm0[2],xmm4[2],xmm0[3],xmm4[3]
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm10, %xmm4
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm8, %xmm2
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3]
; AVX2-FP-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX2-FP-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3,4,5,6],ymm0[7]
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm12, %xmm2
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm14, %xmm3
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm5, %xmm3, %xmm3
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm5, %xmm4, %xmm4
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX2-FP-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX2-FP-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4],ymm2[5],ymm3[6,7]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3,4,5],ymm0[6,7]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-FP-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm0 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm10, %xmm1
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm2, %xmm2
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm2 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm12, %xmm3
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm13, %xmm4
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm1 = xmm3[0,1,2],xmm1[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm3 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm9, %xmm4
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm13, %xmm5
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm5 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX2-FP-NEXT:    vpshufb %xmm5, %xmm11, %xmm6
; AVX2-FP-NEXT:    vpshufb %xmm5, %xmm15, %xmm7
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm4 = xmm6[0],xmm4[1],xmm6[2,3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm1 = xmm4[0,1],xmm1[2,3]
; AVX2-FP-NEXT:    vmovdqa (%rsp), %xmm9 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm9, %xmm4
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm6, %xmm0
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm4[0],xmm0[1],xmm4[1],xmm0[2],xmm4[2],xmm0[3],xmm4[3]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm4, %xmm4
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm8, %xmm2
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3]
; AVX2-FP-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX2-FP-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3,4,5,6],ymm0[7]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm15, %xmm2
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm14, %xmm3
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm5, %xmm14, %xmm3
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm5, %xmm4, %xmm4
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX2-FP-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX2-FP-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4],ymm2[5],ymm3[6,7]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3,4,5],ymm0[6,7]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-FP-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm0 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm10, %xmm1
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm8, %xmm2
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm2 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX2-FP-NEXT:    vmovdqa %xmm12, %xmm11
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm12, %xmm3
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm4, %xmm4
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm1 = xmm3[0,1,2],xmm1[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm3 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm4, %xmm4
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm13, %xmm5
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm5 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm5, %xmm6, %xmm6
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm5, %xmm7, %xmm7
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm4 = xmm6[0],xmm4[1],xmm6[2,3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm1 = xmm4[0,1],xmm1[2,3]
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm9, %xmm4
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm9, %xmm0
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm4[0],xmm0[1],xmm4[1],xmm0[2],xmm4[2],xmm0[3],xmm4[3]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm13, %xmm4
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm12, %xmm2
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3]
; AVX2-FP-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX2-FP-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3,4,5,6],ymm0[7]
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm15, %xmm2
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm4, %xmm3
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
; AVX2-FP-NEXT:    vpshufb %xmm5, %xmm14, %xmm3
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm5, %xmm10, %xmm4
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX2-FP-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX2-FP-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4],ymm2[5],ymm3[6,7]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3,4,5],ymm0[6,7]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-FP-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm0 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm1, %xmm1
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm8, %xmm2
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm2 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm11, %xmm3
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm14, %xmm4
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm1 = xmm3[0,1,2],xmm1[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm3 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm11, %xmm4
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm5, %xmm5
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm5 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm5, %xmm6, %xmm6
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm5, %xmm8, %xmm7
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm4 = xmm6[0],xmm4[1],xmm6[2,3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm1 = xmm4[0,1],xmm1[2,3]
; AVX2-FP-NEXT:    vmovdqa (%rsp), %xmm15 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm15, %xmm4
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm9, %xmm0
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm4[0],xmm0[1],xmm4[1],xmm0[2],xmm4[2],xmm0[3],xmm4[3]
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm13, %xmm4
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm12, %xmm2
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3]
; AVX2-FP-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX2-FP-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3,4,5,6],ymm0[7]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm13, %xmm2
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm9, %xmm3
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm5, %xmm12, %xmm3
; AVX2-FP-NEXT:    vpshufb %xmm5, %xmm10, %xmm4
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX2-FP-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX2-FP-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4],ymm2[5],ymm3[6,7]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3,4,5],ymm0[6,7]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-FP-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm0 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm1, %xmm1
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm2, %xmm2
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm2 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm3, %xmm3
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm14, %xmm4
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm1 = xmm3[0,1,2],xmm1[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm3 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm11, %xmm4
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm5, %xmm5
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm5 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm5, %xmm6, %xmm6
; AVX2-FP-NEXT:    vpshufb %xmm5, %xmm8, %xmm7
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm4 = xmm6[0],xmm4[1],xmm6[2,3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm1 = xmm4[0,1],xmm1[2,3]
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm15, %xmm4
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm6, %xmm0
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm4[0],xmm0[1],xmm4[1],xmm0[2],xmm4[2],xmm0[3],xmm4[3]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm4, %xmm4
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm6, %xmm2
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3]
; AVX2-FP-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX2-FP-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3,4,5,6],ymm0[7]
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm13, %xmm2
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm9, %xmm3
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
; AVX2-FP-NEXT:    vpshufb %xmm5, %xmm12, %xmm3
; AVX2-FP-NEXT:    vpshufb %xmm5, %xmm10, %xmm4
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX2-FP-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX2-FP-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4],ymm2[5],ymm3[6,7]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3,4,5],ymm0[6,7]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, (%rsi)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, (%rdx)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, (%rcx)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, (%r8)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, (%r9)
; AVX2-FP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, (%rax)
; AVX2-FP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, (%rax)
; AVX2-FP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FP-NEXT:    vmovdqa %ymm0, (%rax)
; AVX2-FP-NEXT:    addq $360, %rsp # imm = 0x168
; AVX2-FP-NEXT:    vzeroupper
; AVX2-FP-NEXT:    retq
;
; AVX2-FCP-LABEL: load_i8_stride8_vf32:
; AVX2-FCP:       # %bb.0:
; AVX2-FCP-NEXT:    subq $248, %rsp
; AVX2-FCP-NEXT:    vmovdqa 192(%rdi), %ymm4
; AVX2-FCP-NEXT:    vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovdqa 224(%rdi), %ymm7
; AVX2-FCP-NEXT:    vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovdqa 112(%rdi), %xmm8
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm0 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX2-FCP-NEXT:    vpshufb %xmm0, %xmm8, %xmm1
; AVX2-FCP-NEXT:    vmovdqa 96(%rdi), %xmm13
; AVX2-FCP-NEXT:    vpshufb %xmm0, %xmm13, %xmm0
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
; AVX2-FCP-NEXT:    vmovdqa 80(%rdi), %xmm3
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm1 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX2-FCP-NEXT:    vpshufb %xmm1, %xmm3, %xmm2
; AVX2-FCP-NEXT:    vmovdqa %xmm3, %xmm10
; AVX2-FCP-NEXT:    vmovdqa 64(%rdi), %xmm3
; AVX2-FCP-NEXT:    vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FCP-NEXT:    vpshufb %xmm1, %xmm3, %xmm1
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[3]
; AVX2-FCP-NEXT:    vmovdqa (%rdi), %xmm2
; AVX2-FCP-NEXT:    vmovdqa 16(%rdi), %xmm3
; AVX2-FCP-NEXT:    vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FCP-NEXT:    vmovdqa 32(%rdi), %xmm5
; AVX2-FCP-NEXT:    vmovdqa 48(%rdi), %xmm15
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm1 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX2-FCP-NEXT:    vpshufb %xmm1, %xmm15, %xmm9
; AVX2-FCP-NEXT:    vpshufb %xmm1, %xmm5, %xmm1
; AVX2-FCP-NEXT:    vmovdqa %xmm5, %xmm6
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm9[0],xmm1[1],xmm9[1],xmm1[2],xmm9[2],xmm1[3],xmm9[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm9 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX2-FCP-NEXT:    vpshufb %xmm9, %xmm3, %xmm11
; AVX2-FCP-NEXT:    vpshufb %xmm9, %xmm2, %xmm9
; AVX2-FCP-NEXT:    vmovdqa %xmm2, %xmm5
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm9[0],xmm11[0],xmm9[1],xmm11[1],xmm9[2],xmm11[2],xmm9[3],xmm11[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm1 = xmm9[0],xmm1[1],xmm9[2,3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm3 = xmm1[0,1],xmm0[2,3]
; AVX2-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm1 = [0,2,2,3,0,2,4,6]
; AVX2-FCP-NEXT:    vpermd %ymm7, %ymm1, %ymm9
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm2 = ymm9[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,16,20,24,28]
; AVX2-FCP-NEXT:    vpermd %ymm4, %ymm1, %ymm0
; AVX2-FCP-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm11 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,16,20,24,28,u,u,u,u]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm2 = ymm11[0,1,2,3,4,5,6],ymm2[7]
; AVX2-FCP-NEXT:    vmovdqa 128(%rdi), %ymm0
; AVX2-FCP-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovdqa 160(%rdi), %ymm4
; AVX2-FCP-NEXT:    vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vpermd %ymm4, %ymm1, %ymm11
; AVX2-FCP-NEXT:    vpermd %ymm0, %ymm1, %ymm14
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm0 = ymm11[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,16,20,24,28,u,u,u,u,u,u,u,u]
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm12 = ymm14[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,16,20,24,28,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm12[0,1,2,3,4],ymm0[5],ymm12[6,7]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm2[6,7]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm3[0,1,2,3],ymm0[4,5,6,7]
; AVX2-FCP-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm0 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX2-FCP-NEXT:    vpshufb %xmm0, %xmm8, %xmm2
; AVX2-FCP-NEXT:    vpshufb %xmm0, %xmm13, %xmm0
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],xmm2[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm2 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX2-FCP-NEXT:    vmovdqa %xmm10, %xmm4
; AVX2-FCP-NEXT:    vmovdqa %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FCP-NEXT:    vpshufb %xmm2, %xmm10, %xmm3
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm2, %xmm1, %xmm2
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm2[0,1,2],xmm0[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm2 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX2-FCP-NEXT:    vpshufb %xmm2, %xmm15, %xmm3
; AVX2-FCP-NEXT:    vmovdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FCP-NEXT:    vpshufb %xmm2, %xmm6, %xmm2
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm3 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm3, %xmm7, %xmm12
; AVX2-FCP-NEXT:    vpshufb %xmm3, %xmm5, %xmm3
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm12[0],xmm3[1],xmm12[1],xmm3[2],xmm12[2],xmm3[3],xmm12[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm2 = xmm3[0],xmm2[1],xmm3[2,3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm2[0,1],xmm0[2,3]
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm2 = ymm9[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,17,21,25,29]
; AVX2-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm3 = ymm10[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,17,21,25,29,u,u,u,u]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4,5,6],ymm2[7]
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm3 = ymm11[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,17,21,25,29,u,u,u,u,u,u,u,u]
; AVX2-FCP-NEXT:    vmovdqu %ymm11, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm12 = ymm14[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,17,21,25,29,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm3 = ymm12[0,1,2,3,4],ymm3[5],ymm12[6,7]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4,5],ymm2[6,7]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
; AVX2-FCP-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm0 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX2-FCP-NEXT:    vpshufb %xmm0, %xmm8, %xmm2
; AVX2-FCP-NEXT:    vpshufb %xmm0, %xmm13, %xmm0
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],xmm2[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm2 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX2-FCP-NEXT:    vpshufb %xmm2, %xmm4, %xmm3
; AVX2-FCP-NEXT:    vpshufb %xmm2, %xmm1, %xmm2
; AVX2-FCP-NEXT:    vmovdqa %xmm1, %xmm4
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm2[0,1,2],xmm0[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm2 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX2-FCP-NEXT:    vpshufb %xmm2, %xmm15, %xmm3
; AVX2-FCP-NEXT:    vpshufb %xmm2, %xmm6, %xmm2
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm3 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX2-FCP-NEXT:    vpshufb %xmm3, %xmm7, %xmm12
; AVX2-FCP-NEXT:    vpshufb %xmm3, %xmm5, %xmm3
; AVX2-FCP-NEXT:    vmovdqa %xmm5, %xmm1
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm12[0],xmm3[1],xmm12[1],xmm3[2],xmm12[2],xmm3[3],xmm12[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm2 = xmm3[0],xmm2[1],xmm3[2,3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm2[0,1],xmm0[2,3]
; AVX2-FCP-NEXT:    vmovdqa %ymm9, %ymm7
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm2 = ymm9[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,22,26,30]
; AVX2-FCP-NEXT:    vmovdqa %ymm10, %ymm9
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm3 = ymm10[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,22,26,30,u,u,u,u]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4,5,6],ymm2[7]
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm3 = ymm11[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,22,26,30,u,u,u,u,u,u,u,u]
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm12 = ymm14[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,22,26,30,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm3 = ymm12[0,1,2,3,4],ymm3[5],ymm12[6,7]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4,5],ymm2[6,7]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
; AVX2-FCP-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm0 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX2-FCP-NEXT:    vmovdqa %xmm8, %xmm5
; AVX2-FCP-NEXT:    vpshufb %xmm0, %xmm8, %xmm2
; AVX2-FCP-NEXT:    vmovdqa %xmm13, %xmm8
; AVX2-FCP-NEXT:    vpshufb %xmm0, %xmm13, %xmm0
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],xmm2[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm2 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm2, %xmm6, %xmm3
; AVX2-FCP-NEXT:    vpshufb %xmm2, %xmm4, %xmm2
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm2[0,1,2],xmm0[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm2 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX2-FCP-NEXT:    vpshufb %xmm2, %xmm15, %xmm3
; AVX2-FCP-NEXT:    vmovdqa %xmm15, %xmm10
; AVX2-FCP-NEXT:    vmovdqa %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm2, %xmm11, %xmm2
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm3 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm3, %xmm13, %xmm12
; AVX2-FCP-NEXT:    vpshufb %xmm3, %xmm1, %xmm3
; AVX2-FCP-NEXT:    vmovdqa %xmm1, %xmm4
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm12[0],xmm3[1],xmm12[1],xmm3[2],xmm12[2],xmm3[3],xmm12[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm2 = xmm3[0],xmm2[1],xmm3[2,3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm2[0,1],xmm0[2,3]
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm2 = ymm7[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,19,23,27,31]
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm3 = ymm9[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,19,23,27,31,u,u,u,u]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4,5,6],ymm2[7]
; AVX2-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm3 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,19,23,27,31,u,u,u,u,u,u,u,u]
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm1 = ymm14[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,19,23,27,31,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3,4],ymm3[5],ymm1[6,7]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5],ymm2[6,7]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX2-FCP-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm0 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX2-FCP-NEXT:    vpshufb %xmm0, %xmm5, %xmm1
; AVX2-FCP-NEXT:    vpshufb %xmm0, %xmm8, %xmm0
; AVX2-FCP-NEXT:    vmovdqa %xmm8, %xmm9
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm1 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX2-FCP-NEXT:    vpshufb %xmm1, %xmm6, %xmm2
; AVX2-FCP-NEXT:    vmovdqa %xmm6, %xmm8
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm1, %xmm15, %xmm1
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm1 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX2-FCP-NEXT:    vpshufb %xmm1, %xmm10, %xmm2
; AVX2-FCP-NEXT:    vpshufb %xmm1, %xmm11, %xmm1
; AVX2-FCP-NEXT:    vmovdqa %xmm11, %xmm14
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm2 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX2-FCP-NEXT:    vpshufb %xmm2, %xmm13, %xmm3
; AVX2-FCP-NEXT:    vmovdqa %xmm13, %xmm6
; AVX2-FCP-NEXT:    vpshufb %xmm2, %xmm4, %xmm2
; AVX2-FCP-NEXT:    vmovdqa %xmm4, %xmm10
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm1 = xmm2[0],xmm1[1],xmm2[2,3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm1 = xmm1[0,1],xmm0[2,3]
; AVX2-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm2 = [1,3,2,3,1,3,5,7]
; AVX2-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm7 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm3 = ymm7[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,16,20,24,28]
; AVX2-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm12 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm4 = ymm12[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,16,20,24,28,u,u,u,u]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm3 = ymm4[0,1,2,3,4,5,6],ymm3[7]
; AVX2-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm11 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm13 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm2 = ymm11[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,16,20,24,28,u,u,u,u,u,u,u,u]
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm4 = ymm13[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,16,20,24,28,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm2 = ymm4[0,1,2,3,4],ymm2[5],ymm4[6,7]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm2[4,5,6,7]
; AVX2-FCP-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm1 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX2-FCP-NEXT:    vmovdqa %xmm5, %xmm0
; AVX2-FCP-NEXT:    vpshufb %xmm1, %xmm5, %xmm2
; AVX2-FCP-NEXT:    vpshufb %xmm1, %xmm9, %xmm1
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm2 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX2-FCP-NEXT:    vpshufb %xmm2, %xmm8, %xmm3
; AVX2-FCP-NEXT:    vpshufb %xmm2, %xmm15, %xmm2
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm1 = xmm2[0,1,2],xmm1[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm2 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm2, %xmm5, %xmm3
; AVX2-FCP-NEXT:    vpshufb %xmm2, %xmm14, %xmm2
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm3 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX2-FCP-NEXT:    vpshufb %xmm3, %xmm6, %xmm4
; AVX2-FCP-NEXT:    vpshufb %xmm3, %xmm10, %xmm3
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm2 = xmm3[0],xmm2[1],xmm3[2,3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3]
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm2 = ymm7[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,17,21,25,29]
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm3 = ymm12[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,17,21,25,29,u,u,u,u]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4,5,6],ymm2[7]
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm3 = ymm11[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,17,21,25,29,u,u,u,u,u,u,u,u]
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm4 = ymm13[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,17,21,25,29,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm3 = ymm4[0,1,2,3,4],ymm3[5],ymm4[6,7]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4,5],ymm2[6,7]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
; AVX2-FCP-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm1 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX2-FCP-NEXT:    vpshufb %xmm1, %xmm0, %xmm3
; AVX2-FCP-NEXT:    vmovdqa %xmm0, %xmm2
; AVX2-FCP-NEXT:    vpshufb %xmm1, %xmm9, %xmm1
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1],xmm1[2],xmm3[2],xmm1[3],xmm3[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm3 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX2-FCP-NEXT:    vpshufb %xmm3, %xmm8, %xmm4
; AVX2-FCP-NEXT:    vpshufb %xmm3, %xmm15, %xmm3
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm1 = xmm3[0,1,2],xmm1[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm3 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX2-FCP-NEXT:    vpshufb %xmm3, %xmm5, %xmm4
; AVX2-FCP-NEXT:    vpshufb %xmm3, %xmm14, %xmm3
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm4 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX2-FCP-NEXT:    vpshufb %xmm4, %xmm6, %xmm0
; AVX2-FCP-NEXT:    vpshufb %xmm4, %xmm10, %xmm4
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm4[0],xmm0[0],xmm4[1],xmm0[1],xmm4[2],xmm0[2],xmm4[3],xmm0[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm0[0],xmm3[1],xmm0[2,3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm1 = ymm7[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,22,26,30]
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm3 = ymm12[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,22,26,30,u,u,u,u]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm3[0,1,2,3,4,5,6],ymm1[7]
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm3 = ymm11[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,22,26,30,u,u,u,u,u,u,u,u]
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm4 = ymm13[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,22,26,30,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm3 = ymm4[0,1,2,3,4],ymm3[5],ymm4[6,7]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm3[0,1,2,3,4,5],ymm1[6,7]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm0 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX2-FCP-NEXT:    vpshufb %xmm0, %xmm2, %xmm3
; AVX2-FCP-NEXT:    vpshufb %xmm0, %xmm9, %xmm0
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm3 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX2-FCP-NEXT:    vpshufb %xmm3, %xmm8, %xmm4
; AVX2-FCP-NEXT:    vpshufb %xmm3, %xmm15, %xmm3
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm3[0,1,2],xmm0[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm3 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX2-FCP-NEXT:    vpshufb %xmm3, %xmm5, %xmm4
; AVX2-FCP-NEXT:    vpshufb %xmm3, %xmm14, %xmm3
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm4 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX2-FCP-NEXT:    vpshufb %xmm4, %xmm6, %xmm6
; AVX2-FCP-NEXT:    vpshufb %xmm4, %xmm10, %xmm4
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm4[0],xmm6[0],xmm4[1],xmm6[1],xmm4[2],xmm6[2],xmm4[3],xmm6[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm3 = xmm4[0],xmm3[1],xmm4[2,3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm3[0,1],xmm0[2,3]
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm3 = ymm7[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,19,23,27,31]
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm4 = ymm12[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,19,23,27,31,u,u,u,u]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm3 = ymm4[0,1,2,3,4,5,6],ymm3[7]
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm4 = ymm11[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,19,23,27,31,u,u,u,u,u,u,u,u]
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm5 = ymm13[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,19,23,27,31,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm4 = ymm5[0,1,2,3,4],ymm4[5],ymm5[6,7]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm3 = ymm4[0,1,2,3,4,5],ymm3[6,7]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm3[4,5,6,7]
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm2, (%rsi)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm2, (%rdx)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm2, (%rcx)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm2, (%r8)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm2, (%r9)
; AVX2-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm2, (%rax)
; AVX2-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FCP-NEXT:    vmovdqa %ymm1, (%rax)
; AVX2-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FCP-NEXT:    vmovdqa %ymm0, (%rax)
; AVX2-FCP-NEXT:    addq $248, %rsp
; AVX2-FCP-NEXT:    vzeroupper
; AVX2-FCP-NEXT:    retq
;
; AVX512-LABEL: load_i8_stride8_vf32:
; AVX512:       # %bb.0:
; AVX512-NEXT:    vmovdqa64 (%rdi), %zmm16
; AVX512-NEXT:    vmovdqa64 128(%rdi), %zmm0
; AVX512-NEXT:    vpmovqb %zmm0, %xmm0
; AVX512-NEXT:    vmovdqa 240(%rdi), %xmm2
; AVX512-NEXT:    vpbroadcastd {{.*#+}} xmm3 = [0,0,0,8,0,0,0,8,0,0,0,8,0,0,0,8]
; AVX512-NEXT:    vpshufb %xmm3, %xmm2, %xmm1
; AVX512-NEXT:    vmovdqa %xmm2, %xmm12
; AVX512-NEXT:    vmovdqa 224(%rdi), %xmm4
; AVX512-NEXT:    vpshufb %xmm3, %xmm4, %xmm2
; AVX512-NEXT:    vmovdqa %xmm4, %xmm7
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX512-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm1
; AVX512-NEXT:    vmovdqa 208(%rdi), %xmm5
; AVX512-NEXT:    vpbroadcastd {{.*#+}} xmm4 = [0,8,0,0,0,8,0,0,0,8,0,0,0,8,0,0]
; AVX512-NEXT:    vpshufb %xmm4, %xmm5, %xmm2
; AVX512-NEXT:    vmovdqa %xmm5, %xmm9
; AVX512-NEXT:    vmovdqa 192(%rdi), %xmm5
; AVX512-NEXT:    vpshufb %xmm4, %xmm5, %xmm6
; AVX512-NEXT:    vmovdqa %xmm5, %xmm8
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm6[0],xmm2[0],xmm6[1],xmm2[1],xmm6[2],xmm2[2],xmm6[3],xmm2[3]
; AVX512-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX512-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5,6],ymm1[7]
; AVX512-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX512-NEXT:    vmovdqa 128(%rdi), %ymm2
; AVX512-NEXT:    vpmovqb %ymm2, %xmm2
; AVX512-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX512-NEXT:    vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3,4],ymm0[5],ymm2[6,7]
; AVX512-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
; AVX512-NEXT:    vmovdqa 112(%rdi), %xmm1
; AVX512-NEXT:    vpshufb %xmm3, %xmm1, %xmm6
; AVX512-NEXT:    vmovdqa %xmm1, %xmm5
; AVX512-NEXT:    vmovdqa 96(%rdi), %xmm1
; AVX512-NEXT:    vpshufb %xmm3, %xmm1, %xmm3
; AVX512-NEXT:    vmovdqa64 %xmm1, %xmm18
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm3[0],xmm6[0],xmm3[1],xmm6[1],xmm3[2],xmm6[2],xmm3[3],xmm6[3]
; AVX512-NEXT:    vmovdqa 80(%rdi), %xmm1
; AVX512-NEXT:    vpshufb %xmm4, %xmm1, %xmm11
; AVX512-NEXT:    vmovdqa64 %xmm1, %xmm21
; AVX512-NEXT:    vmovdqa 64(%rdi), %xmm3
; AVX512-NEXT:    vpshufb %xmm4, %xmm3, %xmm4
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm4[0],xmm11[0],xmm4[1],xmm11[1],xmm4[2],xmm11[2],xmm4[3],xmm11[3]
; AVX512-NEXT:    vpblendd {{.*#+}} xmm4 = xmm4[0,1,2],xmm10[3]
; AVX512-NEXT:    vpmovqb %zmm16, %xmm10
; AVX512-NEXT:    vpblendd {{.*#+}} xmm4 = xmm10[0,1],xmm4[2,3]
; AVX512-NEXT:    vpblendd {{.*#+}} ymm0 = ymm4[0,1,2,3],ymm0[4,5,6,7]
; AVX512-NEXT:    vmovdqa64 %ymm0, %ymm17
; AVX512-NEXT:    vmovdqa 160(%rdi), %xmm0
; AVX512-NEXT:    vmovdqa 176(%rdi), %xmm6
; AVX512-NEXT:    vpbroadcastd {{.*#+}} xmm2 = [0,0,1,9,0,0,1,9,0,0,1,9,0,0,1,9]
; AVX512-NEXT:    vpshufb %xmm2, %xmm12, %xmm4
; AVX512-NEXT:    vmovdqa64 %xmm12, %xmm19
; AVX512-NEXT:    vpshufb %xmm2, %xmm7, %xmm12
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm12[0],xmm4[0],xmm12[1],xmm4[1],xmm12[2],xmm4[2],xmm12[3],xmm4[3]
; AVX512-NEXT:    vinserti128 $1, %xmm4, %ymm0, %ymm4
; AVX512-NEXT:    vpbroadcastd {{.*#+}} xmm14 = [1,9,0,0,1,9,0,0,1,9,0,0,1,9,0,0]
; AVX512-NEXT:    vpshufb %xmm14, %xmm9, %xmm12
; AVX512-NEXT:    vmovdqa64 %xmm9, %xmm22
; AVX512-NEXT:    vpshufb %xmm14, %xmm8, %xmm13
; AVX512-NEXT:    vmovdqa64 %xmm8, %xmm20
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm13[0],xmm12[0],xmm13[1],xmm12[1],xmm13[2],xmm12[2],xmm13[3],xmm12[3]
; AVX512-NEXT:    vinserti128 $1, %xmm12, %ymm0, %ymm12
; AVX512-NEXT:    vpblendd {{.*#+}} ymm4 = ymm12[0,1,2,3,4,5,6],ymm4[7]
; AVX512-NEXT:    vmovd {{.*#+}} xmm12 = [0,0,1,9,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX512-NEXT:    vpshufb %xmm12, %xmm6, %xmm13
; AVX512-NEXT:    vpshufb %xmm12, %xmm0, %xmm12
; AVX512-NEXT:    vmovdqa %xmm0, %xmm10
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm15 = xmm12[0],xmm13[0],xmm12[1],xmm13[1],xmm12[2],xmm13[2],xmm12[3],xmm13[3]
; AVX512-NEXT:    vmovdqa 128(%rdi), %xmm8
; AVX512-NEXT:    vmovdqa 144(%rdi), %xmm9
; AVX512-NEXT:    vmovd {{.*#+}} xmm1 = [1,9,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX512-NEXT:    vpshufb %xmm1, %xmm9, %xmm0
; AVX512-NEXT:    vmovdqa %xmm9, %xmm13
; AVX512-NEXT:    vpshufb %xmm1, %xmm8, %xmm1
; AVX512-NEXT:    vmovdqa64 %xmm8, %xmm26
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
; AVX512-NEXT:    vinserti128 $1, %xmm15, %ymm0, %ymm1
; AVX512-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX512-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5],ymm0[6,7]
; AVX512-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm4[6,7]
; AVX512-NEXT:    vpshufb %xmm2, %xmm5, %xmm1
; AVX512-NEXT:    vmovdqa64 %xmm5, %xmm24
; AVX512-NEXT:    vmovdqa64 %xmm18, %xmm5
; AVX512-NEXT:    vpshufb %xmm2, %xmm5, %xmm2
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX512-NEXT:    vmovdqa64 %xmm21, %xmm4
; AVX512-NEXT:    vpshufb %xmm14, %xmm4, %xmm2
; AVX512-NEXT:    vpshufb %xmm14, %xmm3, %xmm4
; AVX512-NEXT:    vmovdqa %xmm3, %xmm11
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm4[0],xmm2[0],xmm4[1],xmm2[1],xmm4[2],xmm2[2],xmm4[3],xmm2[3]
; AVX512-NEXT:    vpblendd {{.*#+}} xmm1 = xmm2[0,1,2],xmm1[3]
; AVX512-NEXT:    vpsrlq $8, %zmm16, %zmm2
; AVX512-NEXT:    vpmovqb %zmm2, %xmm2
; AVX512-NEXT:    vpblendd {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3]
; AVX512-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX512-NEXT:    vmovdqa64 %ymm0, %ymm18
; AVX512-NEXT:    vpbroadcastd {{.*#+}} xmm0 = [0,0,2,10,0,0,2,10,0,0,2,10,0,0,2,10]
; AVX512-NEXT:    vmovdqa64 %xmm19, %xmm3
; AVX512-NEXT:    vpshufb %xmm0, %xmm3, %xmm1
; AVX512-NEXT:    vpshufb %xmm0, %xmm7, %xmm2
; AVX512-NEXT:    vmovdqa %xmm7, %xmm9
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX512-NEXT:    vpbroadcastd {{.*#+}} xmm2 = [2,10,0,0,2,10,0,0,2,10,0,0,2,10,0,0]
; AVX512-NEXT:    vmovdqa64 %xmm22, %xmm8
; AVX512-NEXT:    vpshufb %xmm2, %xmm8, %xmm4
; AVX512-NEXT:    vmovdqa64 %xmm20, %xmm7
; AVX512-NEXT:    vpshufb %xmm2, %xmm7, %xmm14
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm14[0],xmm4[0],xmm14[1],xmm4[1],xmm14[2],xmm4[2],xmm14[3],xmm4[3]
; AVX512-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm1
; AVX512-NEXT:    vinserti128 $1, %xmm4, %ymm0, %ymm4
; AVX512-NEXT:    vpblendd {{.*#+}} ymm1 = ymm4[0,1,2,3,4,5,6],ymm1[7]
; AVX512-NEXT:    vmovd {{.*#+}} xmm4 = [0,0,2,10,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX512-NEXT:    vpshufb %xmm4, %xmm6, %xmm14
; AVX512-NEXT:    vmovdqa %xmm6, %xmm12
; AVX512-NEXT:    vpshufb %xmm4, %xmm10, %xmm4
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm4[0],xmm14[0],xmm4[1],xmm14[1],xmm4[2],xmm14[2],xmm4[3],xmm14[3]
; AVX512-NEXT:    vmovd {{.*#+}} xmm14 = [2,10,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX512-NEXT:    vpshufb %xmm14, %xmm13, %xmm15
; AVX512-NEXT:    vmovdqa64 %xmm26, %xmm6
; AVX512-NEXT:    vpshufb %xmm14, %xmm6, %xmm14
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm14[0],xmm15[0],xmm14[1],xmm15[1],xmm14[2],xmm15[2],xmm14[3],xmm15[3]
; AVX512-NEXT:    vinserti128 $1, %xmm4, %ymm0, %ymm4
; AVX512-NEXT:    vinserti128 $1, %xmm14, %ymm0, %ymm14
; AVX512-NEXT:    vpblendd {{.*#+}} ymm4 = ymm14[0,1,2,3,4],ymm4[5],ymm14[6,7]
; AVX512-NEXT:    vpblendd {{.*#+}} ymm1 = ymm4[0,1,2,3,4,5],ymm1[6,7]
; AVX512-NEXT:    vmovdqa64 %xmm24, %xmm4
; AVX512-NEXT:    vpshufb %xmm0, %xmm4, %xmm4
; AVX512-NEXT:    vpshufb %xmm0, %xmm5, %xmm0
; AVX512-NEXT:    vmovdqa64 %xmm5, %xmm20
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm4[0],xmm0[1],xmm4[1],xmm0[2],xmm4[2],xmm0[3],xmm4[3]
; AVX512-NEXT:    vmovdqa64 %xmm21, %xmm6
; AVX512-NEXT:    vpshufb %xmm2, %xmm6, %xmm4
; AVX512-NEXT:    vpshufb %xmm2, %xmm11, %xmm2
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3]
; AVX512-NEXT:    vpblendd {{.*#+}} xmm0 = xmm2[0,1,2],xmm0[3]
; AVX512-NEXT:    vpsrlq $16, %zmm16, %zmm2
; AVX512-NEXT:    vpmovqb %zmm2, %xmm2
; AVX512-NEXT:    vpblendd {{.*#+}} xmm0 = xmm2[0,1],xmm0[2,3]
; AVX512-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX512-NEXT:    vmovdqa64 %ymm0, %ymm19
; AVX512-NEXT:    vpbroadcastd {{.*#+}} xmm0 = [0,0,3,11,0,0,3,11,0,0,3,11,0,0,3,11]
; AVX512-NEXT:    vpshufb %xmm0, %xmm3, %xmm1
; AVX512-NEXT:    vmovdqa64 %xmm3, %xmm25
; AVX512-NEXT:    vpshufb %xmm0, %xmm9, %xmm2
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX512-NEXT:    vpbroadcastd {{.*#+}} xmm2 = [3,11,0,0,3,11,0,0,3,11,0,0,3,11,0,0]
; AVX512-NEXT:    vpshufb %xmm2, %xmm8, %xmm4
; AVX512-NEXT:    vpshufb %xmm2, %xmm7, %xmm14
; AVX512-NEXT:    vmovdqa64 %xmm7, %xmm22
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm14[0],xmm4[0],xmm14[1],xmm4[1],xmm14[2],xmm4[2],xmm14[3],xmm4[3]
; AVX512-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm1
; AVX512-NEXT:    vinserti128 $1, %xmm4, %ymm0, %ymm4
; AVX512-NEXT:    vpblendd {{.*#+}} ymm1 = ymm4[0,1,2,3,4,5,6],ymm1[7]
; AVX512-NEXT:    vmovd {{.*#+}} xmm4 = [0,0,3,11,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX512-NEXT:    vpshufb %xmm4, %xmm12, %xmm14
; AVX512-NEXT:    vmovdqa64 %xmm12, %xmm27
; AVX512-NEXT:    vpshufb %xmm4, %xmm10, %xmm4
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm4[0],xmm14[0],xmm4[1],xmm14[1],xmm4[2],xmm14[2],xmm4[3],xmm14[3]
; AVX512-NEXT:    vmovd {{.*#+}} xmm14 = [3,11,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX512-NEXT:    vpshufb %xmm14, %xmm13, %xmm15
; AVX512-NEXT:    vmovdqa64 %xmm26, %xmm12
; AVX512-NEXT:    vpshufb %xmm14, %xmm12, %xmm14
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm14[0],xmm15[0],xmm14[1],xmm15[1],xmm14[2],xmm15[2],xmm14[3],xmm15[3]
; AVX512-NEXT:    vinserti128 $1, %xmm4, %ymm0, %ymm4
; AVX512-NEXT:    vinserti128 $1, %xmm14, %ymm0, %ymm14
; AVX512-NEXT:    vpblendd {{.*#+}} ymm4 = ymm14[0,1,2,3,4],ymm4[5],ymm14[6,7]
; AVX512-NEXT:    vpblendd {{.*#+}} ymm1 = ymm4[0,1,2,3,4,5],ymm1[6,7]
; AVX512-NEXT:    vmovdqa64 %xmm24, %xmm5
; AVX512-NEXT:    vpshufb %xmm0, %xmm5, %xmm4
; AVX512-NEXT:    vmovdqa64 %xmm20, %xmm15
; AVX512-NEXT:    vpshufb %xmm0, %xmm15, %xmm0
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm4[0],xmm0[1],xmm4[1],xmm0[2],xmm4[2],xmm0[3],xmm4[3]
; AVX512-NEXT:    vpshufb %xmm2, %xmm6, %xmm4
; AVX512-NEXT:    vpshufb %xmm2, %xmm11, %xmm2
; AVX512-NEXT:    vmovdqa %xmm11, %xmm6
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3]
; AVX512-NEXT:    vpblendd {{.*#+}} xmm0 = xmm2[0,1,2],xmm0[3]
; AVX512-NEXT:    vpsrlq $24, %zmm16, %zmm2
; AVX512-NEXT:    vpmovqb %zmm2, %xmm2
; AVX512-NEXT:    vpblendd {{.*#+}} xmm0 = xmm2[0,1],xmm0[2,3]
; AVX512-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX512-NEXT:    vmovdqa64 %ymm0, %ymm20
; AVX512-NEXT:    vpbroadcastd {{.*#+}} xmm3 = [0,0,4,12,0,0,4,12,0,0,4,12,0,0,4,12]
; AVX512-NEXT:    vmovdqa64 %xmm25, %xmm11
; AVX512-NEXT:    vpshufb %xmm3, %xmm11, %xmm1
; AVX512-NEXT:    vpshufb %xmm3, %xmm9, %xmm2
; AVX512-NEXT:    vmovdqa %xmm9, %xmm7
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX512-NEXT:    vpbroadcastd {{.*#+}} xmm2 = [4,12,0,0,4,12,0,0,4,12,0,0,4,12,0,0]
; AVX512-NEXT:    vpshufb %xmm2, %xmm8, %xmm4
; AVX512-NEXT:    vmovdqa64 %xmm22, %xmm9
; AVX512-NEXT:    vpshufb %xmm2, %xmm9, %xmm14
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm14[0],xmm4[0],xmm14[1],xmm4[1],xmm14[2],xmm4[2],xmm14[3],xmm4[3]
; AVX512-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm1
; AVX512-NEXT:    vinserti128 $1, %xmm4, %ymm0, %ymm4
; AVX512-NEXT:    vpblendd {{.*#+}} ymm1 = ymm4[0,1,2,3,4,5,6],ymm1[7]
; AVX512-NEXT:    vmovd {{.*#+}} xmm4 = [0,0,4,12,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX512-NEXT:    vmovdqa64 %xmm27, %xmm0
; AVX512-NEXT:    vpshufb %xmm4, %xmm0, %xmm14
; AVX512-NEXT:    vpshufb %xmm4, %xmm10, %xmm4
; AVX512-NEXT:    vmovdqa64 %xmm10, %xmm22
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm4[0],xmm14[0],xmm4[1],xmm14[1],xmm4[2],xmm14[2],xmm4[3],xmm14[3]
; AVX512-NEXT:    vmovd {{.*#+}} xmm14 = [4,12,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX512-NEXT:    vpshufb %xmm14, %xmm13, %xmm0
; AVX512-NEXT:    vmovdqa64 %xmm13, %xmm25
; AVX512-NEXT:    vpshufb %xmm14, %xmm12, %xmm14
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm14[0],xmm0[0],xmm14[1],xmm0[1],xmm14[2],xmm0[2],xmm14[3],xmm0[3]
; AVX512-NEXT:    vinserti128 $1, %xmm4, %ymm0, %ymm4
; AVX512-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX512-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm4[5],ymm0[6,7]
; AVX512-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
; AVX512-NEXT:    vpshufb %xmm3, %xmm5, %xmm1
; AVX512-NEXT:    vpshufb %xmm3, %xmm15, %xmm3
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3]
; AVX512-NEXT:    vmovdqa64 %xmm21, %xmm10
; AVX512-NEXT:    vpshufb %xmm2, %xmm10, %xmm3
; AVX512-NEXT:    vpshufb %xmm2, %xmm6, %xmm2
; AVX512-NEXT:    vmovdqa64 %xmm6, %xmm23
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
; AVX512-NEXT:    vpblendd {{.*#+}} xmm1 = xmm2[0,1,2],xmm1[3]
; AVX512-NEXT:    vpsrlq $32, %zmm16, %zmm2
; AVX512-NEXT:    vpmovqb %zmm2, %xmm2
; AVX512-NEXT:    vpblendd {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3]
; AVX512-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX512-NEXT:    vmovdqa64 %ymm0, %ymm21
; AVX512-NEXT:    vpbroadcastd {{.*#+}} xmm5 = [0,0,5,13,0,0,5,13,0,0,5,13,0,0,5,13]
; AVX512-NEXT:    vpshufb %xmm5, %xmm11, %xmm1
; AVX512-NEXT:    vpshufb %xmm5, %xmm7, %xmm2
; AVX512-NEXT:    vmovdqa %xmm7, %xmm13
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX512-NEXT:    vpbroadcastd {{.*#+}} xmm2 = [5,13,0,0,5,13,0,0,5,13,0,0,5,13,0,0]
; AVX512-NEXT:    vpshufb %xmm2, %xmm8, %xmm3
; AVX512-NEXT:    vmovdqa64 %xmm8, %xmm26
; AVX512-NEXT:    vpshufb %xmm2, %xmm9, %xmm4
; AVX512-NEXT:    vmovdqa %xmm9, %xmm14
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX512-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm1
; AVX512-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX512-NEXT:    vpblendd {{.*#+}} ymm1 = ymm3[0,1,2,3,4,5,6],ymm1[7]
; AVX512-NEXT:    vmovd {{.*#+}} xmm3 = [0,0,5,13,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX512-NEXT:    vmovdqa64 %xmm27, %xmm6
; AVX512-NEXT:    vpshufb %xmm3, %xmm6, %xmm4
; AVX512-NEXT:    vmovdqa64 %xmm22, %xmm7
; AVX512-NEXT:    vpshufb %xmm3, %xmm7, %xmm3
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
; AVX512-NEXT:    vmovd {{.*#+}} xmm4 = [5,13,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX512-NEXT:    vmovdqa64 %xmm25, %xmm9
; AVX512-NEXT:    vpshufb %xmm4, %xmm9, %xmm0
; AVX512-NEXT:    vpshufb %xmm4, %xmm12, %xmm4
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm4[0],xmm0[0],xmm4[1],xmm0[1],xmm4[2],xmm0[2],xmm4[3],xmm0[3]
; AVX512-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX512-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX512-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm3[5],ymm0[6,7]
; AVX512-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
; AVX512-NEXT:    vmovdqa64 %xmm24, %xmm8
; AVX512-NEXT:    vpshufb %xmm5, %xmm8, %xmm1
; AVX512-NEXT:    vpshufb %xmm5, %xmm15, %xmm3
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3]
; AVX512-NEXT:    vpshufb %xmm2, %xmm10, %xmm3
; AVX512-NEXT:    vmovdqa64 %xmm10, %xmm24
; AVX512-NEXT:    vmovdqa64 %xmm23, %xmm4
; AVX512-NEXT:    vpshufb %xmm2, %xmm4, %xmm2
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
; AVX512-NEXT:    vpblendd {{.*#+}} xmm1 = xmm2[0,1,2],xmm1[3]
; AVX512-NEXT:    vpsrlq $40, %zmm16, %zmm2
; AVX512-NEXT:    vpmovqb %zmm2, %xmm2
; AVX512-NEXT:    vpblendd {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3]
; AVX512-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX512-NEXT:    vmovdqa64 %ymm0, %ymm22
; AVX512-NEXT:    vpbroadcastd {{.*#+}} xmm10 = [0,0,6,14,0,0,6,14,0,0,6,14,0,0,6,14]
; AVX512-NEXT:    vpshufb %xmm10, %xmm11, %xmm1
; AVX512-NEXT:    vmovdqa64 %xmm11, %xmm23
; AVX512-NEXT:    vpshufb %xmm10, %xmm13, %xmm2
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX512-NEXT:    vpbroadcastd {{.*#+}} xmm2 = [6,14,0,0,6,14,0,0,6,14,0,0,6,14,0,0]
; AVX512-NEXT:    vmovdqa64 %xmm26, %xmm11
; AVX512-NEXT:    vpshufb %xmm2, %xmm11, %xmm3
; AVX512-NEXT:    vpshufb %xmm2, %xmm14, %xmm5
; AVX512-NEXT:    vmovdqa64 %xmm14, %xmm25
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm5[0],xmm3[0],xmm5[1],xmm3[1],xmm5[2],xmm3[2],xmm5[3],xmm3[3]
; AVX512-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm1
; AVX512-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX512-NEXT:    vpblendd {{.*#+}} ymm1 = ymm3[0,1,2,3,4,5,6],ymm1[7]
; AVX512-NEXT:    vmovd {{.*#+}} xmm3 = [0,0,6,14,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX512-NEXT:    vpshufb %xmm3, %xmm6, %xmm5
; AVX512-NEXT:    vpshufb %xmm3, %xmm7, %xmm3
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm5[0],xmm3[1],xmm5[1],xmm3[2],xmm5[2],xmm3[3],xmm5[3]
; AVX512-NEXT:    vmovd {{.*#+}} xmm5 = [6,14,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX512-NEXT:    vpshufb %xmm5, %xmm9, %xmm0
; AVX512-NEXT:    vmovdqa %xmm9, %xmm14
; AVX512-NEXT:    vpshufb %xmm5, %xmm12, %xmm5
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm5[0],xmm0[0],xmm5[1],xmm0[1],xmm5[2],xmm0[2],xmm5[3],xmm0[3]
; AVX512-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX512-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX512-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm3[5],ymm0[6,7]
; AVX512-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
; AVX512-NEXT:    vpshufb %xmm10, %xmm8, %xmm1
; AVX512-NEXT:    vmovdqa %xmm8, %xmm9
; AVX512-NEXT:    vpshufb %xmm10, %xmm15, %xmm3
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3]
; AVX512-NEXT:    vmovdqa64 %xmm24, %xmm10
; AVX512-NEXT:    vpshufb %xmm2, %xmm10, %xmm3
; AVX512-NEXT:    vpshufb %xmm2, %xmm4, %xmm2
; AVX512-NEXT:    vmovdqa %xmm4, %xmm8
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
; AVX512-NEXT:    vpblendd {{.*#+}} xmm1 = xmm2[0,1,2],xmm1[3]
; AVX512-NEXT:    vpsrlq $48, %zmm16, %zmm2
; AVX512-NEXT:    vpmovqb %zmm2, %xmm2
; AVX512-NEXT:    vpblendd {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3]
; AVX512-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX512-NEXT:    vpbroadcastd {{.*#+}} xmm1 = [0,0,7,15,0,0,7,15,0,0,7,15,0,0,7,15]
; AVX512-NEXT:    vmovdqa64 %xmm23, %xmm2
; AVX512-NEXT:    vpshufb %xmm1, %xmm2, %xmm2
; AVX512-NEXT:    vpshufb %xmm1, %xmm13, %xmm3
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
; AVX512-NEXT:    vpbroadcastd {{.*#+}} xmm3 = [7,15,0,0,7,15,0,0,7,15,0,0,7,15,0,0]
; AVX512-NEXT:    vpshufb %xmm3, %xmm11, %xmm4
; AVX512-NEXT:    vmovdqa64 %xmm25, %xmm5
; AVX512-NEXT:    vpshufb %xmm3, %xmm5, %xmm5
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX512-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX512-NEXT:    vinserti128 $1, %xmm4, %ymm0, %ymm4
; AVX512-NEXT:    vpblendd {{.*#+}} ymm2 = ymm4[0,1,2,3,4,5,6],ymm2[7]
; AVX512-NEXT:    vmovd {{.*#+}} xmm4 = [0,0,7,15,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX512-NEXT:    vpshufb %xmm4, %xmm6, %xmm5
; AVX512-NEXT:    vpshufb %xmm4, %xmm7, %xmm4
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3]
; AVX512-NEXT:    vmovd {{.*#+}} xmm5 = [7,15,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX512-NEXT:    vpshufb %xmm5, %xmm14, %xmm7
; AVX512-NEXT:    vpshufb %xmm5, %xmm12, %xmm5
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm7[0],xmm5[1],xmm7[1],xmm5[2],xmm7[2],xmm5[3],xmm7[3]
; AVX512-NEXT:    vinserti128 $1, %xmm4, %ymm0, %ymm4
; AVX512-NEXT:    vinserti128 $1, %xmm5, %ymm0, %ymm5
; AVX512-NEXT:    vpblendd {{.*#+}} ymm4 = ymm5[0,1,2,3,4],ymm4[5],ymm5[6,7]
; AVX512-NEXT:    vpblendd {{.*#+}} ymm2 = ymm4[0,1,2,3,4,5],ymm2[6,7]
; AVX512-NEXT:    vpshufb %xmm1, %xmm9, %xmm4
; AVX512-NEXT:    vpshufb %xmm1, %xmm15, %xmm1
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm4[0],xmm1[1],xmm4[1],xmm1[2],xmm4[2],xmm1[3],xmm4[3]
; AVX512-NEXT:    vpshufb %xmm3, %xmm10, %xmm4
; AVX512-NEXT:    vpshufb %xmm3, %xmm8, %xmm3
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
; AVX512-NEXT:    vpblendd {{.*#+}} xmm1 = xmm3[0,1,2],xmm1[3]
; AVX512-NEXT:    vpsrlq $56, %zmm16, %zmm3
; AVX512-NEXT:    vpmovqb %zmm3, %xmm3
; AVX512-NEXT:    vpblendd {{.*#+}} xmm1 = xmm3[0,1],xmm1[2,3]
; AVX512-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
; AVX512-NEXT:    vmovdqa64 %ymm17, (%rsi)
; AVX512-NEXT:    vmovdqa64 %ymm18, (%rdx)
; AVX512-NEXT:    vmovdqa64 %ymm19, (%rcx)
; AVX512-NEXT:    vmovdqa64 %ymm20, (%r8)
; AVX512-NEXT:    vmovdqa64 %ymm21, (%r9)
; AVX512-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512-NEXT:    vmovdqa64 %ymm22, (%rax)
; AVX512-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512-NEXT:    vmovdqa %ymm0, (%rax)
; AVX512-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512-NEXT:    vmovdqa %ymm1, (%rax)
; AVX512-NEXT:    vzeroupper
; AVX512-NEXT:    retq
;
; AVX512-FCP-LABEL: load_i8_stride8_vf32:
; AVX512-FCP:       # %bb.0:
; AVX512-FCP-NEXT:    vmovdqa64 (%rdi), %zmm18
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm14 = [0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12]
; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm0 = [0,2,2,3,0,2,4,6]
; AVX512-FCP-NEXT:    vmovdqa64 224(%rdi), %ymm16
; AVX512-FCP-NEXT:    vpermd %ymm16, %ymm0, %ymm4
; AVX512-FCP-NEXT:    vpshufb %ymm14, %ymm4, %ymm1
; AVX512-FCP-NEXT:    vmovdqa64 192(%rdi), %ymm17
; AVX512-FCP-NEXT:    vpermd %ymm17, %ymm0, %ymm6
; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm2 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,16,20,24,28,u,u,u,u]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5,6],ymm1[7]
; AVX512-FCP-NEXT:    vmovdqa 160(%rdi), %ymm13
; AVX512-FCP-NEXT:    vpermd %ymm13, %ymm0, %ymm9
; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm2 = ymm9[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,16,20,24,28,u,u,u,u,u,u,u,u]
; AVX512-FCP-NEXT:    vmovdqa 128(%rdi), %ymm15
; AVX512-FCP-NEXT:    vpermd %ymm15, %ymm0, %ymm11
; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm0 = ymm11[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,16,20,24,28,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm2[5],ymm0[6,7]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm3 = [0,0,0,8,0,0,0,8,0,0,0,8,0,0,0,8]
; AVX512-FCP-NEXT:    vmovdqa 112(%rdi), %xmm1
; AVX512-FCP-NEXT:    vpshufb %xmm3, %xmm1, %xmm5
; AVX512-FCP-NEXT:    vmovdqa %xmm1, %xmm2
; AVX512-FCP-NEXT:    vmovdqa 96(%rdi), %xmm1
; AVX512-FCP-NEXT:    vpshufb %xmm3, %xmm1, %xmm3
; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm7 = xmm3[0],xmm5[0],xmm3[1],xmm5[1],xmm3[2],xmm5[2],xmm3[3],xmm5[3]
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm8 = [0,8,0,0,0,8,0,0,0,8,0,0,0,8,0,0]
; AVX512-FCP-NEXT:    vmovdqa 80(%rdi), %xmm12
; AVX512-FCP-NEXT:    vpshufb %xmm8, %xmm12, %xmm10
; AVX512-FCP-NEXT:    vmovdqa 64(%rdi), %xmm5
; AVX512-FCP-NEXT:    vpshufb %xmm8, %xmm5, %xmm8
; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm10[0],xmm8[1],xmm10[1],xmm8[2],xmm10[2],xmm8[3],xmm10[3]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} xmm7 = xmm8[0,1,2],xmm7[3]
; AVX512-FCP-NEXT:    vpmovqb %zmm18, %xmm8
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} xmm7 = xmm8[0,1],xmm7[2,3]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm7[0,1,2,3],ymm0[4,5,6,7]
; AVX512-FCP-NEXT:    vmovdqa64 %ymm0, %ymm22
; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm0 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,17,21,25,29]
; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm7 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,17,21,25,29,u,u,u,u]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm7[0,1,2,3,4,5,6],ymm0[7]
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm3 = [1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13]
; AVX512-FCP-NEXT:    vpshufb %ymm3, %ymm9, %ymm7
; AVX512-FCP-NEXT:    vmovdqa64 %ymm3, %ymm27
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm3 = [1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13]
; AVX512-FCP-NEXT:    vpshufb %ymm3, %ymm11, %ymm8
; AVX512-FCP-NEXT:    vmovdqa64 %ymm3, %ymm28
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm7 = ymm8[0,1,2,3,4],ymm7[5],ymm8[6,7]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm7[0,1,2,3,4,5],ymm0[6,7]
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm7 = [0,0,1,9,0,0,1,9,0,0,1,9,0,0,1,9]
; AVX512-FCP-NEXT:    vpshufb %xmm7, %xmm2, %xmm8
; AVX512-FCP-NEXT:    vpshufb %xmm7, %xmm1, %xmm7
; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm7 = xmm7[0],xmm8[0],xmm7[1],xmm8[1],xmm7[2],xmm8[2],xmm7[3],xmm8[3]
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm8 = [1,9,0,0,1,9,0,0,1,9,0,0,1,9,0,0]
; AVX512-FCP-NEXT:    vpshufb %xmm8, %xmm12, %xmm10
; AVX512-FCP-NEXT:    vpshufb %xmm8, %xmm5, %xmm8
; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm10[0],xmm8[1],xmm10[1],xmm8[2],xmm10[2],xmm8[3],xmm10[3]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} xmm7 = xmm8[0,1,2],xmm7[3]
; AVX512-FCP-NEXT:    vpsrlq $8, %zmm18, %zmm8
; AVX512-FCP-NEXT:    vpmovqb %zmm8, %xmm8
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} xmm7 = xmm8[0,1],xmm7[2,3]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm7[0,1,2,3],ymm0[4,5,6,7]
; AVX512-FCP-NEXT:    vmovdqa64 %ymm0, %ymm20
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm3 = [2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14]
; AVX512-FCP-NEXT:    vpshufb %ymm3, %ymm4, %ymm0
; AVX512-FCP-NEXT:    vmovdqa64 %ymm3, %ymm29
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm3 = [2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14]
; AVX512-FCP-NEXT:    vpshufb %ymm3, %ymm6, %ymm7
; AVX512-FCP-NEXT:    vmovdqa64 %ymm3, %ymm30
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm7[0,1,2,3,4,5,6],ymm0[7]
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm3 = [2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14]
; AVX512-FCP-NEXT:    vpshufb %ymm3, %ymm9, %ymm7
; AVX512-FCP-NEXT:    vmovdqa64 %ymm3, %ymm31
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm3 = [2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14]
; AVX512-FCP-NEXT:    vpshufb %ymm3, %ymm11, %ymm8
; AVX512-FCP-NEXT:    vmovdqa64 %ymm3, %ymm26
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm7 = ymm8[0,1,2,3,4],ymm7[5],ymm8[6,7]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm7[0,1,2,3,4,5],ymm0[6,7]
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm7 = [0,0,2,10,0,0,2,10,0,0,2,10,0,0,2,10]
; AVX512-FCP-NEXT:    vpshufb %xmm7, %xmm2, %xmm8
; AVX512-FCP-NEXT:    vpshufb %xmm7, %xmm1, %xmm7
; AVX512-FCP-NEXT:    vmovdqa %xmm1, %xmm3
; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm7 = xmm7[0],xmm8[0],xmm7[1],xmm8[1],xmm7[2],xmm8[2],xmm7[3],xmm8[3]
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm8 = [2,10,0,0,2,10,0,0,2,10,0,0,2,10,0,0]
; AVX512-FCP-NEXT:    vpshufb %xmm8, %xmm12, %xmm10
; AVX512-FCP-NEXT:    vpshufb %xmm8, %xmm5, %xmm8
; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm10[0],xmm8[1],xmm10[1],xmm8[2],xmm10[2],xmm8[3],xmm10[3]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} xmm7 = xmm8[0,1,2],xmm7[3]
; AVX512-FCP-NEXT:    vpsrlq $16, %zmm18, %zmm8
; AVX512-FCP-NEXT:    vpmovqb %zmm8, %xmm8
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} xmm7 = xmm8[0,1],xmm7[2,3]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm7[0,1,2,3],ymm0[4,5,6,7]
; AVX512-FCP-NEXT:    vmovdqa64 %ymm0, %ymm21
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm1 = [3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15]
; AVX512-FCP-NEXT:    vpshufb %ymm1, %ymm4, %ymm0
; AVX512-FCP-NEXT:    vmovdqa64 %ymm1, %ymm25
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm1 = [3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15]
; AVX512-FCP-NEXT:    vpshufb %ymm1, %ymm6, %ymm4
; AVX512-FCP-NEXT:    vmovdqa64 %ymm1, %ymm24
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm4[0,1,2,3,4,5,6],ymm0[7]
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm1 = [3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15]
; AVX512-FCP-NEXT:    vpshufb %ymm1, %ymm9, %ymm4
; AVX512-FCP-NEXT:    vmovdqa64 %ymm1, %ymm23
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm8 = [3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15]
; AVX512-FCP-NEXT:    vpshufb %ymm8, %ymm11, %ymm6
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm4 = ymm6[0,1,2,3,4],ymm4[5],ymm6[6,7]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm4[0,1,2,3,4,5],ymm0[6,7]
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm4 = [0,0,3,11,0,0,3,11,0,0,3,11,0,0,3,11]
; AVX512-FCP-NEXT:    vpshufb %xmm4, %xmm2, %xmm6
; AVX512-FCP-NEXT:    vpshufb %xmm4, %xmm3, %xmm4
; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm4[0],xmm6[0],xmm4[1],xmm6[1],xmm4[2],xmm6[2],xmm4[3],xmm6[3]
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm6 = [3,11,0,0,3,11,0,0,3,11,0,0,3,11,0,0]
; AVX512-FCP-NEXT:    vpshufb %xmm6, %xmm12, %xmm9
; AVX512-FCP-NEXT:    vpshufb %xmm6, %xmm5, %xmm6
; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm6[0],xmm9[0],xmm6[1],xmm9[1],xmm6[2],xmm9[2],xmm6[3],xmm9[3]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} xmm4 = xmm6[0,1,2],xmm4[3]
; AVX512-FCP-NEXT:    vpsrlq $24, %zmm18, %zmm6
; AVX512-FCP-NEXT:    vpmovqb %zmm6, %xmm6
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} xmm4 = xmm6[0,1],xmm4[2,3]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm4[0,1,2,3],ymm0[4,5,6,7]
; AVX512-FCP-NEXT:    vmovdqa64 %ymm0, %ymm19
; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm0 = [1,3,2,3,1,3,5,7]
; AVX512-FCP-NEXT:    vpermd %ymm16, %ymm0, %ymm11
; AVX512-FCP-NEXT:    vpshufb %ymm14, %ymm11, %ymm4
; AVX512-FCP-NEXT:    vpermd %ymm17, %ymm0, %ymm9
; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm14 = ymm9[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,16,20,24,28,u,u,u,u]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm4 = ymm14[0,1,2,3,4,5,6],ymm4[7]
; AVX512-FCP-NEXT:    vpermd %ymm13, %ymm0, %ymm13
; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm14 = ymm13[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,16,20,24,28,u,u,u,u,u,u,u,u]
; AVX512-FCP-NEXT:    vpermd %ymm15, %ymm0, %ymm1
; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm15 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,16,20,24,28,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm14 = ymm15[0,1,2,3,4],ymm14[5],ymm15[6,7]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm4 = ymm14[0,1,2,3,4,5],ymm4[6,7]
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm14 = [0,0,4,12,0,0,4,12,0,0,4,12,0,0,4,12]
; AVX512-FCP-NEXT:    vmovdqa %xmm2, %xmm6
; AVX512-FCP-NEXT:    vpshufb %xmm14, %xmm2, %xmm15
; AVX512-FCP-NEXT:    vpshufb %xmm14, %xmm3, %xmm14
; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm14[0],xmm15[0],xmm14[1],xmm15[1],xmm14[2],xmm15[2],xmm14[3],xmm15[3]
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm15 = [4,12,0,0,4,12,0,0,4,12,0,0,4,12,0,0]
; AVX512-FCP-NEXT:    vpshufb %xmm15, %xmm12, %xmm0
; AVX512-FCP-NEXT:    vpshufb %xmm15, %xmm5, %xmm15
; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm15[0],xmm0[0],xmm15[1],xmm0[1],xmm15[2],xmm0[2],xmm15[3],xmm0[3]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm0[0,1,2],xmm14[3]
; AVX512-FCP-NEXT:    vpsrlq $32, %zmm18, %zmm14
; AVX512-FCP-NEXT:    vpmovqb %zmm14, %xmm14
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm14[0,1],xmm0[2,3]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm15 = ymm0[0,1,2,3],ymm4[4,5,6,7]
; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm0 = ymm11[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,17,21,25,29]
; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm4 = ymm9[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,17,21,25,29,u,u,u,u]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm4[0,1,2,3,4,5,6],ymm0[7]
; AVX512-FCP-NEXT:    vmovdqa64 %ymm27, %ymm2
; AVX512-FCP-NEXT:    vpshufb %ymm2, %ymm13, %ymm4
; AVX512-FCP-NEXT:    vmovdqa64 %ymm28, %ymm2
; AVX512-FCP-NEXT:    vpshufb %ymm2, %ymm1, %ymm14
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm4 = ymm14[0,1,2,3,4],ymm4[5],ymm14[6,7]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm2 = ymm4[0,1,2,3,4,5],ymm0[6,7]
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm4 = [0,0,5,13,0,0,5,13,0,0,5,13,0,0,5,13]
; AVX512-FCP-NEXT:    vpshufb %xmm4, %xmm6, %xmm14
; AVX512-FCP-NEXT:    vmovdqa %xmm3, %xmm7
; AVX512-FCP-NEXT:    vpshufb %xmm4, %xmm3, %xmm4
; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm4[0],xmm14[0],xmm4[1],xmm14[1],xmm4[2],xmm14[2],xmm4[3],xmm14[3]
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm14 = [5,13,0,0,5,13,0,0,5,13,0,0,5,13,0,0]
; AVX512-FCP-NEXT:    vpshufb %xmm14, %xmm12, %xmm0
; AVX512-FCP-NEXT:    vpshufb %xmm14, %xmm5, %xmm14
; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm14[0],xmm0[0],xmm14[1],xmm0[1],xmm14[2],xmm0[2],xmm14[3],xmm0[3]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm0[0,1,2],xmm4[3]
; AVX512-FCP-NEXT:    vpsrlq $40, %zmm18, %zmm4
; AVX512-FCP-NEXT:    vpmovqb %zmm4, %xmm4
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm4[0,1],xmm0[2,3]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm10 = ymm0[0,1,2,3],ymm2[4,5,6,7]
; AVX512-FCP-NEXT:    vmovdqa64 %ymm29, %ymm0
; AVX512-FCP-NEXT:    vpshufb %ymm0, %ymm11, %ymm0
; AVX512-FCP-NEXT:    vmovdqa64 %ymm30, %ymm2
; AVX512-FCP-NEXT:    vpshufb %ymm2, %ymm9, %ymm2
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3,4,5,6],ymm0[7]
; AVX512-FCP-NEXT:    vmovdqa64 %ymm31, %ymm2
; AVX512-FCP-NEXT:    vpshufb %ymm2, %ymm13, %ymm2
; AVX512-FCP-NEXT:    vmovdqa64 %ymm26, %ymm4
; AVX512-FCP-NEXT:    vpshufb %ymm4, %ymm1, %ymm14
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm2 = ymm14[0,1,2,3,4],ymm2[5],ymm14[6,7]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm4 = ymm2[0,1,2,3,4,5],ymm0[6,7]
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm2 = [0,0,6,14,0,0,6,14,0,0,6,14,0,0,6,14]
; AVX512-FCP-NEXT:    vpshufb %xmm2, %xmm6, %xmm14
; AVX512-FCP-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm14[0],xmm2[1],xmm14[1],xmm2[2],xmm14[2],xmm2[3],xmm14[3]
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm14 = [6,14,0,0,6,14,0,0,6,14,0,0,6,14,0,0]
; AVX512-FCP-NEXT:    vpshufb %xmm14, %xmm12, %xmm0
; AVX512-FCP-NEXT:    vpshufb %xmm14, %xmm5, %xmm14
; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm14[0],xmm0[0],xmm14[1],xmm0[1],xmm14[2],xmm0[2],xmm14[3],xmm0[3]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm0[0,1,2],xmm2[3]
; AVX512-FCP-NEXT:    vpsrlq $48, %zmm18, %zmm2
; AVX512-FCP-NEXT:    vpmovqb %zmm2, %xmm2
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm2[0,1],xmm0[2,3]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm4[4,5,6,7]
; AVX512-FCP-NEXT:    vmovdqa64 %ymm25, %ymm2
; AVX512-FCP-NEXT:    vpshufb %ymm2, %ymm11, %ymm2
; AVX512-FCP-NEXT:    vmovdqa64 %ymm24, %ymm3
; AVX512-FCP-NEXT:    vpshufb %ymm3, %ymm9, %ymm3
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4,5,6],ymm2[7]
; AVX512-FCP-NEXT:    vmovdqa64 %ymm23, %ymm3
; AVX512-FCP-NEXT:    vpshufb %ymm3, %ymm13, %ymm3
; AVX512-FCP-NEXT:    vpshufb %ymm8, %ymm1, %ymm1
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3,4],ymm3[5],ymm1[6,7]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5],ymm2[6,7]
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm2 = [0,0,7,15,0,0,7,15,0,0,7,15,0,0,7,15]
; AVX512-FCP-NEXT:    vpshufb %xmm2, %xmm6, %xmm3
; AVX512-FCP-NEXT:    vpshufb %xmm2, %xmm7, %xmm2
; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm3 = [7,15,0,0,7,15,0,0,7,15,0,0,7,15,0,0]
; AVX512-FCP-NEXT:    vpshufb %xmm3, %xmm12, %xmm8
; AVX512-FCP-NEXT:    vpshufb %xmm3, %xmm5, %xmm3
; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm8[0],xmm3[1],xmm8[1],xmm3[2],xmm8[2],xmm3[3],xmm8[3]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} xmm2 = xmm3[0,1,2],xmm2[3]
; AVX512-FCP-NEXT:    vpsrlq $56, %zmm18, %zmm3
; AVX512-FCP-NEXT:    vpmovqb %zmm3, %xmm3
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} xmm2 = xmm3[0,1],xmm2[2,3]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512-FCP-NEXT:    vmovdqa64 %ymm22, (%rsi)
; AVX512-FCP-NEXT:    vmovdqa64 %ymm20, (%rdx)
; AVX512-FCP-NEXT:    vmovdqa64 %ymm21, (%rcx)
; AVX512-FCP-NEXT:    vmovdqa64 %ymm19, (%r8)
; AVX512-FCP-NEXT:    vmovdqa %ymm15, (%r9)
; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512-FCP-NEXT:    vmovdqa %ymm10, (%rax)
; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512-FCP-NEXT:    vmovdqa %ymm0, (%rax)
; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512-FCP-NEXT:    vmovdqa %ymm1, (%rax)
; AVX512-FCP-NEXT:    vzeroupper
; AVX512-FCP-NEXT:    retq
;
; AVX512DQ-LABEL: load_i8_stride8_vf32:
; AVX512DQ:       # %bb.0:
; AVX512DQ-NEXT:    vmovdqa64 (%rdi), %zmm16
; AVX512DQ-NEXT:    vmovdqa64 128(%rdi), %zmm0
; AVX512DQ-NEXT:    vpmovqb %zmm0, %xmm0
; AVX512DQ-NEXT:    vmovdqa 240(%rdi), %xmm2
; AVX512DQ-NEXT:    vpbroadcastd {{.*#+}} xmm3 = [0,0,0,8,0,0,0,8,0,0,0,8,0,0,0,8]
; AVX512DQ-NEXT:    vpshufb %xmm3, %xmm2, %xmm1
; AVX512DQ-NEXT:    vmovdqa %xmm2, %xmm12
; AVX512DQ-NEXT:    vmovdqa 224(%rdi), %xmm4
; AVX512DQ-NEXT:    vpshufb %xmm3, %xmm4, %xmm2
; AVX512DQ-NEXT:    vmovdqa %xmm4, %xmm7
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX512DQ-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm1
; AVX512DQ-NEXT:    vmovdqa 208(%rdi), %xmm5
; AVX512DQ-NEXT:    vpbroadcastd {{.*#+}} xmm4 = [0,8,0,0,0,8,0,0,0,8,0,0,0,8,0,0]
; AVX512DQ-NEXT:    vpshufb %xmm4, %xmm5, %xmm2
; AVX512DQ-NEXT:    vmovdqa %xmm5, %xmm9
; AVX512DQ-NEXT:    vmovdqa 192(%rdi), %xmm5
; AVX512DQ-NEXT:    vpshufb %xmm4, %xmm5, %xmm6
; AVX512DQ-NEXT:    vmovdqa %xmm5, %xmm8
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm6[0],xmm2[0],xmm6[1],xmm2[1],xmm6[2],xmm2[2],xmm6[3],xmm2[3]
; AVX512DQ-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5,6],ymm1[7]
; AVX512DQ-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX512DQ-NEXT:    vmovdqa 128(%rdi), %ymm2
; AVX512DQ-NEXT:    vpmovqb %ymm2, %xmm2
; AVX512DQ-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3,4],ymm0[5],ymm2[6,7]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
; AVX512DQ-NEXT:    vmovdqa 112(%rdi), %xmm1
; AVX512DQ-NEXT:    vpshufb %xmm3, %xmm1, %xmm6
; AVX512DQ-NEXT:    vmovdqa %xmm1, %xmm5
; AVX512DQ-NEXT:    vmovdqa 96(%rdi), %xmm1
; AVX512DQ-NEXT:    vpshufb %xmm3, %xmm1, %xmm3
; AVX512DQ-NEXT:    vmovdqa64 %xmm1, %xmm18
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm3[0],xmm6[0],xmm3[1],xmm6[1],xmm3[2],xmm6[2],xmm3[3],xmm6[3]
; AVX512DQ-NEXT:    vmovdqa 80(%rdi), %xmm1
; AVX512DQ-NEXT:    vpshufb %xmm4, %xmm1, %xmm11
; AVX512DQ-NEXT:    vmovdqa64 %xmm1, %xmm21
; AVX512DQ-NEXT:    vmovdqa 64(%rdi), %xmm3
; AVX512DQ-NEXT:    vpshufb %xmm4, %xmm3, %xmm4
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm4[0],xmm11[0],xmm4[1],xmm11[1],xmm4[2],xmm11[2],xmm4[3],xmm11[3]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} xmm4 = xmm4[0,1,2],xmm10[3]
; AVX512DQ-NEXT:    vpmovqb %zmm16, %xmm10
; AVX512DQ-NEXT:    vpblendd {{.*#+}} xmm4 = xmm10[0,1],xmm4[2,3]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm0 = ymm4[0,1,2,3],ymm0[4,5,6,7]
; AVX512DQ-NEXT:    vmovdqa64 %ymm0, %ymm17
; AVX512DQ-NEXT:    vmovdqa 160(%rdi), %xmm0
; AVX512DQ-NEXT:    vmovdqa 176(%rdi), %xmm6
; AVX512DQ-NEXT:    vpbroadcastd {{.*#+}} xmm2 = [0,0,1,9,0,0,1,9,0,0,1,9,0,0,1,9]
; AVX512DQ-NEXT:    vpshufb %xmm2, %xmm12, %xmm4
; AVX512DQ-NEXT:    vmovdqa64 %xmm12, %xmm19
; AVX512DQ-NEXT:    vpshufb %xmm2, %xmm7, %xmm12
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm12[0],xmm4[0],xmm12[1],xmm4[1],xmm12[2],xmm4[2],xmm12[3],xmm4[3]
; AVX512DQ-NEXT:    vinserti128 $1, %xmm4, %ymm0, %ymm4
; AVX512DQ-NEXT:    vpbroadcastd {{.*#+}} xmm14 = [1,9,0,0,1,9,0,0,1,9,0,0,1,9,0,0]
; AVX512DQ-NEXT:    vpshufb %xmm14, %xmm9, %xmm12
; AVX512DQ-NEXT:    vmovdqa64 %xmm9, %xmm22
; AVX512DQ-NEXT:    vpshufb %xmm14, %xmm8, %xmm13
; AVX512DQ-NEXT:    vmovdqa64 %xmm8, %xmm20
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm13[0],xmm12[0],xmm13[1],xmm12[1],xmm13[2],xmm12[2],xmm13[3],xmm12[3]
; AVX512DQ-NEXT:    vinserti128 $1, %xmm12, %ymm0, %ymm12
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm4 = ymm12[0,1,2,3,4,5,6],ymm4[7]
; AVX512DQ-NEXT:    vmovd {{.*#+}} xmm12 = [0,0,1,9,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX512DQ-NEXT:    vpshufb %xmm12, %xmm6, %xmm13
; AVX512DQ-NEXT:    vpshufb %xmm12, %xmm0, %xmm12
; AVX512DQ-NEXT:    vmovdqa %xmm0, %xmm10
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm15 = xmm12[0],xmm13[0],xmm12[1],xmm13[1],xmm12[2],xmm13[2],xmm12[3],xmm13[3]
; AVX512DQ-NEXT:    vmovdqa 128(%rdi), %xmm8
; AVX512DQ-NEXT:    vmovdqa 144(%rdi), %xmm9
; AVX512DQ-NEXT:    vmovd {{.*#+}} xmm1 = [1,9,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX512DQ-NEXT:    vpshufb %xmm1, %xmm9, %xmm0
; AVX512DQ-NEXT:    vmovdqa %xmm9, %xmm13
; AVX512DQ-NEXT:    vpshufb %xmm1, %xmm8, %xmm1
; AVX512DQ-NEXT:    vmovdqa64 %xmm8, %xmm26
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
; AVX512DQ-NEXT:    vinserti128 $1, %xmm15, %ymm0, %ymm1
; AVX512DQ-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5],ymm0[6,7]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm4[6,7]
; AVX512DQ-NEXT:    vpshufb %xmm2, %xmm5, %xmm1
; AVX512DQ-NEXT:    vmovdqa64 %xmm5, %xmm24
; AVX512DQ-NEXT:    vmovdqa64 %xmm18, %xmm5
; AVX512DQ-NEXT:    vpshufb %xmm2, %xmm5, %xmm2
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX512DQ-NEXT:    vmovdqa64 %xmm21, %xmm4
; AVX512DQ-NEXT:    vpshufb %xmm14, %xmm4, %xmm2
; AVX512DQ-NEXT:    vpshufb %xmm14, %xmm3, %xmm4
; AVX512DQ-NEXT:    vmovdqa %xmm3, %xmm11
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm4[0],xmm2[0],xmm4[1],xmm2[1],xmm4[2],xmm2[2],xmm4[3],xmm2[3]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} xmm1 = xmm2[0,1,2],xmm1[3]
; AVX512DQ-NEXT:    vpsrlq $8, %zmm16, %zmm2
; AVX512DQ-NEXT:    vpmovqb %zmm2, %xmm2
; AVX512DQ-NEXT:    vpblendd {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX512DQ-NEXT:    vmovdqa64 %ymm0, %ymm18
; AVX512DQ-NEXT:    vpbroadcastd {{.*#+}} xmm0 = [0,0,2,10,0,0,2,10,0,0,2,10,0,0,2,10]
; AVX512DQ-NEXT:    vmovdqa64 %xmm19, %xmm3
; AVX512DQ-NEXT:    vpshufb %xmm0, %xmm3, %xmm1
; AVX512DQ-NEXT:    vpshufb %xmm0, %xmm7, %xmm2
; AVX512DQ-NEXT:    vmovdqa %xmm7, %xmm9
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX512DQ-NEXT:    vpbroadcastd {{.*#+}} xmm2 = [2,10,0,0,2,10,0,0,2,10,0,0,2,10,0,0]
; AVX512DQ-NEXT:    vmovdqa64 %xmm22, %xmm8
; AVX512DQ-NEXT:    vpshufb %xmm2, %xmm8, %xmm4
; AVX512DQ-NEXT:    vmovdqa64 %xmm20, %xmm7
; AVX512DQ-NEXT:    vpshufb %xmm2, %xmm7, %xmm14
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm14[0],xmm4[0],xmm14[1],xmm4[1],xmm14[2],xmm4[2],xmm14[3],xmm4[3]
; AVX512DQ-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm1
; AVX512DQ-NEXT:    vinserti128 $1, %xmm4, %ymm0, %ymm4
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm1 = ymm4[0,1,2,3,4,5,6],ymm1[7]
; AVX512DQ-NEXT:    vmovd {{.*#+}} xmm4 = [0,0,2,10,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX512DQ-NEXT:    vpshufb %xmm4, %xmm6, %xmm14
; AVX512DQ-NEXT:    vmovdqa %xmm6, %xmm12
; AVX512DQ-NEXT:    vpshufb %xmm4, %xmm10, %xmm4
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm4[0],xmm14[0],xmm4[1],xmm14[1],xmm4[2],xmm14[2],xmm4[3],xmm14[3]
; AVX512DQ-NEXT:    vmovd {{.*#+}} xmm14 = [2,10,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX512DQ-NEXT:    vpshufb %xmm14, %xmm13, %xmm15
; AVX512DQ-NEXT:    vmovdqa64 %xmm26, %xmm6
; AVX512DQ-NEXT:    vpshufb %xmm14, %xmm6, %xmm14
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm14[0],xmm15[0],xmm14[1],xmm15[1],xmm14[2],xmm15[2],xmm14[3],xmm15[3]
; AVX512DQ-NEXT:    vinserti128 $1, %xmm4, %ymm0, %ymm4
; AVX512DQ-NEXT:    vinserti128 $1, %xmm14, %ymm0, %ymm14
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm4 = ymm14[0,1,2,3,4],ymm4[5],ymm14[6,7]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm1 = ymm4[0,1,2,3,4,5],ymm1[6,7]
; AVX512DQ-NEXT:    vmovdqa64 %xmm24, %xmm4
; AVX512DQ-NEXT:    vpshufb %xmm0, %xmm4, %xmm4
; AVX512DQ-NEXT:    vpshufb %xmm0, %xmm5, %xmm0
; AVX512DQ-NEXT:    vmovdqa64 %xmm5, %xmm20
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm4[0],xmm0[1],xmm4[1],xmm0[2],xmm4[2],xmm0[3],xmm4[3]
; AVX512DQ-NEXT:    vmovdqa64 %xmm21, %xmm6
; AVX512DQ-NEXT:    vpshufb %xmm2, %xmm6, %xmm4
; AVX512DQ-NEXT:    vpshufb %xmm2, %xmm11, %xmm2
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} xmm0 = xmm2[0,1,2],xmm0[3]
; AVX512DQ-NEXT:    vpsrlq $16, %zmm16, %zmm2
; AVX512DQ-NEXT:    vpmovqb %zmm2, %xmm2
; AVX512DQ-NEXT:    vpblendd {{.*#+}} xmm0 = xmm2[0,1],xmm0[2,3]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-NEXT:    vmovdqa64 %ymm0, %ymm19
; AVX512DQ-NEXT:    vpbroadcastd {{.*#+}} xmm0 = [0,0,3,11,0,0,3,11,0,0,3,11,0,0,3,11]
; AVX512DQ-NEXT:    vpshufb %xmm0, %xmm3, %xmm1
; AVX512DQ-NEXT:    vmovdqa64 %xmm3, %xmm25
; AVX512DQ-NEXT:    vpshufb %xmm0, %xmm9, %xmm2
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX512DQ-NEXT:    vpbroadcastd {{.*#+}} xmm2 = [3,11,0,0,3,11,0,0,3,11,0,0,3,11,0,0]
; AVX512DQ-NEXT:    vpshufb %xmm2, %xmm8, %xmm4
; AVX512DQ-NEXT:    vpshufb %xmm2, %xmm7, %xmm14
; AVX512DQ-NEXT:    vmovdqa64 %xmm7, %xmm22
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm14[0],xmm4[0],xmm14[1],xmm4[1],xmm14[2],xmm4[2],xmm14[3],xmm4[3]
; AVX512DQ-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm1
; AVX512DQ-NEXT:    vinserti128 $1, %xmm4, %ymm0, %ymm4
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm1 = ymm4[0,1,2,3,4,5,6],ymm1[7]
; AVX512DQ-NEXT:    vmovd {{.*#+}} xmm4 = [0,0,3,11,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX512DQ-NEXT:    vpshufb %xmm4, %xmm12, %xmm14
; AVX512DQ-NEXT:    vmovdqa64 %xmm12, %xmm27
; AVX512DQ-NEXT:    vpshufb %xmm4, %xmm10, %xmm4
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm4[0],xmm14[0],xmm4[1],xmm14[1],xmm4[2],xmm14[2],xmm4[3],xmm14[3]
; AVX512DQ-NEXT:    vmovd {{.*#+}} xmm14 = [3,11,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX512DQ-NEXT:    vpshufb %xmm14, %xmm13, %xmm15
; AVX512DQ-NEXT:    vmovdqa64 %xmm26, %xmm12
; AVX512DQ-NEXT:    vpshufb %xmm14, %xmm12, %xmm14
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm14[0],xmm15[0],xmm14[1],xmm15[1],xmm14[2],xmm15[2],xmm14[3],xmm15[3]
; AVX512DQ-NEXT:    vinserti128 $1, %xmm4, %ymm0, %ymm4
; AVX512DQ-NEXT:    vinserti128 $1, %xmm14, %ymm0, %ymm14
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm4 = ymm14[0,1,2,3,4],ymm4[5],ymm14[6,7]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm1 = ymm4[0,1,2,3,4,5],ymm1[6,7]
; AVX512DQ-NEXT:    vmovdqa64 %xmm24, %xmm5
; AVX512DQ-NEXT:    vpshufb %xmm0, %xmm5, %xmm4
; AVX512DQ-NEXT:    vmovdqa64 %xmm20, %xmm15
; AVX512DQ-NEXT:    vpshufb %xmm0, %xmm15, %xmm0
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm4[0],xmm0[1],xmm4[1],xmm0[2],xmm4[2],xmm0[3],xmm4[3]
; AVX512DQ-NEXT:    vpshufb %xmm2, %xmm6, %xmm4
; AVX512DQ-NEXT:    vpshufb %xmm2, %xmm11, %xmm2
; AVX512DQ-NEXT:    vmovdqa %xmm11, %xmm6
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} xmm0 = xmm2[0,1,2],xmm0[3]
; AVX512DQ-NEXT:    vpsrlq $24, %zmm16, %zmm2
; AVX512DQ-NEXT:    vpmovqb %zmm2, %xmm2
; AVX512DQ-NEXT:    vpblendd {{.*#+}} xmm0 = xmm2[0,1],xmm0[2,3]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-NEXT:    vmovdqa64 %ymm0, %ymm20
; AVX512DQ-NEXT:    vpbroadcastd {{.*#+}} xmm3 = [0,0,4,12,0,0,4,12,0,0,4,12,0,0,4,12]
; AVX512DQ-NEXT:    vmovdqa64 %xmm25, %xmm11
; AVX512DQ-NEXT:    vpshufb %xmm3, %xmm11, %xmm1
; AVX512DQ-NEXT:    vpshufb %xmm3, %xmm9, %xmm2
; AVX512DQ-NEXT:    vmovdqa %xmm9, %xmm7
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX512DQ-NEXT:    vpbroadcastd {{.*#+}} xmm2 = [4,12,0,0,4,12,0,0,4,12,0,0,4,12,0,0]
; AVX512DQ-NEXT:    vpshufb %xmm2, %xmm8, %xmm4
; AVX512DQ-NEXT:    vmovdqa64 %xmm22, %xmm9
; AVX512DQ-NEXT:    vpshufb %xmm2, %xmm9, %xmm14
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm14[0],xmm4[0],xmm14[1],xmm4[1],xmm14[2],xmm4[2],xmm14[3],xmm4[3]
; AVX512DQ-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm1
; AVX512DQ-NEXT:    vinserti128 $1, %xmm4, %ymm0, %ymm4
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm1 = ymm4[0,1,2,3,4,5,6],ymm1[7]
; AVX512DQ-NEXT:    vmovd {{.*#+}} xmm4 = [0,0,4,12,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX512DQ-NEXT:    vmovdqa64 %xmm27, %xmm0
; AVX512DQ-NEXT:    vpshufb %xmm4, %xmm0, %xmm14
; AVX512DQ-NEXT:    vpshufb %xmm4, %xmm10, %xmm4
; AVX512DQ-NEXT:    vmovdqa64 %xmm10, %xmm22
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm4[0],xmm14[0],xmm4[1],xmm14[1],xmm4[2],xmm14[2],xmm4[3],xmm14[3]
; AVX512DQ-NEXT:    vmovd {{.*#+}} xmm14 = [4,12,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX512DQ-NEXT:    vpshufb %xmm14, %xmm13, %xmm0
; AVX512DQ-NEXT:    vmovdqa64 %xmm13, %xmm25
; AVX512DQ-NEXT:    vpshufb %xmm14, %xmm12, %xmm14
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm14[0],xmm0[0],xmm14[1],xmm0[1],xmm14[2],xmm0[2],xmm14[3],xmm0[3]
; AVX512DQ-NEXT:    vinserti128 $1, %xmm4, %ymm0, %ymm4
; AVX512DQ-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm4[5],ymm0[6,7]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
; AVX512DQ-NEXT:    vpshufb %xmm3, %xmm5, %xmm1
; AVX512DQ-NEXT:    vpshufb %xmm3, %xmm15, %xmm3
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3]
; AVX512DQ-NEXT:    vmovdqa64 %xmm21, %xmm10
; AVX512DQ-NEXT:    vpshufb %xmm2, %xmm10, %xmm3
; AVX512DQ-NEXT:    vpshufb %xmm2, %xmm6, %xmm2
; AVX512DQ-NEXT:    vmovdqa64 %xmm6, %xmm23
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} xmm1 = xmm2[0,1,2],xmm1[3]
; AVX512DQ-NEXT:    vpsrlq $32, %zmm16, %zmm2
; AVX512DQ-NEXT:    vpmovqb %zmm2, %xmm2
; AVX512DQ-NEXT:    vpblendd {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX512DQ-NEXT:    vmovdqa64 %ymm0, %ymm21
; AVX512DQ-NEXT:    vpbroadcastd {{.*#+}} xmm5 = [0,0,5,13,0,0,5,13,0,0,5,13,0,0,5,13]
; AVX512DQ-NEXT:    vpshufb %xmm5, %xmm11, %xmm1
; AVX512DQ-NEXT:    vpshufb %xmm5, %xmm7, %xmm2
; AVX512DQ-NEXT:    vmovdqa %xmm7, %xmm13
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX512DQ-NEXT:    vpbroadcastd {{.*#+}} xmm2 = [5,13,0,0,5,13,0,0,5,13,0,0,5,13,0,0]
; AVX512DQ-NEXT:    vpshufb %xmm2, %xmm8, %xmm3
; AVX512DQ-NEXT:    vmovdqa64 %xmm8, %xmm26
; AVX512DQ-NEXT:    vpshufb %xmm2, %xmm9, %xmm4
; AVX512DQ-NEXT:    vmovdqa %xmm9, %xmm14
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX512DQ-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm1
; AVX512DQ-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm1 = ymm3[0,1,2,3,4,5,6],ymm1[7]
; AVX512DQ-NEXT:    vmovd {{.*#+}} xmm3 = [0,0,5,13,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX512DQ-NEXT:    vmovdqa64 %xmm27, %xmm6
; AVX512DQ-NEXT:    vpshufb %xmm3, %xmm6, %xmm4
; AVX512DQ-NEXT:    vmovdqa64 %xmm22, %xmm7
; AVX512DQ-NEXT:    vpshufb %xmm3, %xmm7, %xmm3
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
; AVX512DQ-NEXT:    vmovd {{.*#+}} xmm4 = [5,13,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX512DQ-NEXT:    vmovdqa64 %xmm25, %xmm9
; AVX512DQ-NEXT:    vpshufb %xmm4, %xmm9, %xmm0
; AVX512DQ-NEXT:    vpshufb %xmm4, %xmm12, %xmm4
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm4[0],xmm0[0],xmm4[1],xmm0[1],xmm4[2],xmm0[2],xmm4[3],xmm0[3]
; AVX512DQ-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX512DQ-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm3[5],ymm0[6,7]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
; AVX512DQ-NEXT:    vmovdqa64 %xmm24, %xmm8
; AVX512DQ-NEXT:    vpshufb %xmm5, %xmm8, %xmm1
; AVX512DQ-NEXT:    vpshufb %xmm5, %xmm15, %xmm3
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3]
; AVX512DQ-NEXT:    vpshufb %xmm2, %xmm10, %xmm3
; AVX512DQ-NEXT:    vmovdqa64 %xmm10, %xmm24
; AVX512DQ-NEXT:    vmovdqa64 %xmm23, %xmm4
; AVX512DQ-NEXT:    vpshufb %xmm2, %xmm4, %xmm2
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} xmm1 = xmm2[0,1,2],xmm1[3]
; AVX512DQ-NEXT:    vpsrlq $40, %zmm16, %zmm2
; AVX512DQ-NEXT:    vpmovqb %zmm2, %xmm2
; AVX512DQ-NEXT:    vpblendd {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX512DQ-NEXT:    vmovdqa64 %ymm0, %ymm22
; AVX512DQ-NEXT:    vpbroadcastd {{.*#+}} xmm10 = [0,0,6,14,0,0,6,14,0,0,6,14,0,0,6,14]
; AVX512DQ-NEXT:    vpshufb %xmm10, %xmm11, %xmm1
; AVX512DQ-NEXT:    vmovdqa64 %xmm11, %xmm23
; AVX512DQ-NEXT:    vpshufb %xmm10, %xmm13, %xmm2
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX512DQ-NEXT:    vpbroadcastd {{.*#+}} xmm2 = [6,14,0,0,6,14,0,0,6,14,0,0,6,14,0,0]
; AVX512DQ-NEXT:    vmovdqa64 %xmm26, %xmm11
; AVX512DQ-NEXT:    vpshufb %xmm2, %xmm11, %xmm3
; AVX512DQ-NEXT:    vpshufb %xmm2, %xmm14, %xmm5
; AVX512DQ-NEXT:    vmovdqa64 %xmm14, %xmm25
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm5[0],xmm3[0],xmm5[1],xmm3[1],xmm5[2],xmm3[2],xmm5[3],xmm3[3]
; AVX512DQ-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm1
; AVX512DQ-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm1 = ymm3[0,1,2,3,4,5,6],ymm1[7]
; AVX512DQ-NEXT:    vmovd {{.*#+}} xmm3 = [0,0,6,14,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX512DQ-NEXT:    vpshufb %xmm3, %xmm6, %xmm5
; AVX512DQ-NEXT:    vpshufb %xmm3, %xmm7, %xmm3
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm5[0],xmm3[1],xmm5[1],xmm3[2],xmm5[2],xmm3[3],xmm5[3]
; AVX512DQ-NEXT:    vmovd {{.*#+}} xmm5 = [6,14,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX512DQ-NEXT:    vpshufb %xmm5, %xmm9, %xmm0
; AVX512DQ-NEXT:    vmovdqa %xmm9, %xmm14
; AVX512DQ-NEXT:    vpshufb %xmm5, %xmm12, %xmm5
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm5[0],xmm0[0],xmm5[1],xmm0[1],xmm5[2],xmm0[2],xmm5[3],xmm0[3]
; AVX512DQ-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX512DQ-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm3[5],ymm0[6,7]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
; AVX512DQ-NEXT:    vpshufb %xmm10, %xmm8, %xmm1
; AVX512DQ-NEXT:    vmovdqa %xmm8, %xmm9
; AVX512DQ-NEXT:    vpshufb %xmm10, %xmm15, %xmm3
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3]
; AVX512DQ-NEXT:    vmovdqa64 %xmm24, %xmm10
; AVX512DQ-NEXT:    vpshufb %xmm2, %xmm10, %xmm3
; AVX512DQ-NEXT:    vpshufb %xmm2, %xmm4, %xmm2
; AVX512DQ-NEXT:    vmovdqa %xmm4, %xmm8
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} xmm1 = xmm2[0,1,2],xmm1[3]
; AVX512DQ-NEXT:    vpsrlq $48, %zmm16, %zmm2
; AVX512DQ-NEXT:    vpmovqb %zmm2, %xmm2
; AVX512DQ-NEXT:    vpblendd {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX512DQ-NEXT:    vpbroadcastd {{.*#+}} xmm1 = [0,0,7,15,0,0,7,15,0,0,7,15,0,0,7,15]
; AVX512DQ-NEXT:    vmovdqa64 %xmm23, %xmm2
; AVX512DQ-NEXT:    vpshufb %xmm1, %xmm2, %xmm2
; AVX512DQ-NEXT:    vpshufb %xmm1, %xmm13, %xmm3
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
; AVX512DQ-NEXT:    vpbroadcastd {{.*#+}} xmm3 = [7,15,0,0,7,15,0,0,7,15,0,0,7,15,0,0]
; AVX512DQ-NEXT:    vpshufb %xmm3, %xmm11, %xmm4
; AVX512DQ-NEXT:    vmovdqa64 %xmm25, %xmm5
; AVX512DQ-NEXT:    vpshufb %xmm3, %xmm5, %xmm5
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX512DQ-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX512DQ-NEXT:    vinserti128 $1, %xmm4, %ymm0, %ymm4
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm2 = ymm4[0,1,2,3,4,5,6],ymm2[7]
; AVX512DQ-NEXT:    vmovd {{.*#+}} xmm4 = [0,0,7,15,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX512DQ-NEXT:    vpshufb %xmm4, %xmm6, %xmm5
; AVX512DQ-NEXT:    vpshufb %xmm4, %xmm7, %xmm4
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3]
; AVX512DQ-NEXT:    vmovd {{.*#+}} xmm5 = [7,15,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX512DQ-NEXT:    vpshufb %xmm5, %xmm14, %xmm7
; AVX512DQ-NEXT:    vpshufb %xmm5, %xmm12, %xmm5
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm7[0],xmm5[1],xmm7[1],xmm5[2],xmm7[2],xmm5[3],xmm7[3]
; AVX512DQ-NEXT:    vinserti128 $1, %xmm4, %ymm0, %ymm4
; AVX512DQ-NEXT:    vinserti128 $1, %xmm5, %ymm0, %ymm5
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm4 = ymm5[0,1,2,3,4],ymm4[5],ymm5[6,7]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm2 = ymm4[0,1,2,3,4,5],ymm2[6,7]
; AVX512DQ-NEXT:    vpshufb %xmm1, %xmm9, %xmm4
; AVX512DQ-NEXT:    vpshufb %xmm1, %xmm15, %xmm1
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm4[0],xmm1[1],xmm4[1],xmm1[2],xmm4[2],xmm1[3],xmm4[3]
; AVX512DQ-NEXT:    vpshufb %xmm3, %xmm10, %xmm4
; AVX512DQ-NEXT:    vpshufb %xmm3, %xmm8, %xmm3
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} xmm1 = xmm3[0,1,2],xmm1[3]
; AVX512DQ-NEXT:    vpsrlq $56, %zmm16, %zmm3
; AVX512DQ-NEXT:    vpmovqb %zmm3, %xmm3
; AVX512DQ-NEXT:    vpblendd {{.*#+}} xmm1 = xmm3[0,1],xmm1[2,3]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
; AVX512DQ-NEXT:    vmovdqa64 %ymm17, (%rsi)
; AVX512DQ-NEXT:    vmovdqa64 %ymm18, (%rdx)
; AVX512DQ-NEXT:    vmovdqa64 %ymm19, (%rcx)
; AVX512DQ-NEXT:    vmovdqa64 %ymm20, (%r8)
; AVX512DQ-NEXT:    vmovdqa64 %ymm21, (%r9)
; AVX512DQ-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-NEXT:    vmovdqa64 %ymm22, (%rax)
; AVX512DQ-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-NEXT:    vmovdqa %ymm0, (%rax)
; AVX512DQ-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-NEXT:    vmovdqa %ymm1, (%rax)
; AVX512DQ-NEXT:    vzeroupper
; AVX512DQ-NEXT:    retq
;
; AVX512DQ-FCP-LABEL: load_i8_stride8_vf32:
; AVX512DQ-FCP:       # %bb.0:
; AVX512DQ-FCP-NEXT:    vmovdqa64 (%rdi), %zmm18
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm14 = [0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12]
; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm0 = [0,2,2,3,0,2,4,6]
; AVX512DQ-FCP-NEXT:    vmovdqa64 224(%rdi), %ymm16
; AVX512DQ-FCP-NEXT:    vpermd %ymm16, %ymm0, %ymm4
; AVX512DQ-FCP-NEXT:    vpshufb %ymm14, %ymm4, %ymm1
; AVX512DQ-FCP-NEXT:    vmovdqa64 192(%rdi), %ymm17
; AVX512DQ-FCP-NEXT:    vpermd %ymm17, %ymm0, %ymm6
; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm2 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,16,20,24,28,u,u,u,u]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5,6],ymm1[7]
; AVX512DQ-FCP-NEXT:    vmovdqa 160(%rdi), %ymm13
; AVX512DQ-FCP-NEXT:    vpermd %ymm13, %ymm0, %ymm9
; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm2 = ymm9[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,16,20,24,28,u,u,u,u,u,u,u,u]
; AVX512DQ-FCP-NEXT:    vmovdqa 128(%rdi), %ymm15
; AVX512DQ-FCP-NEXT:    vpermd %ymm15, %ymm0, %ymm11
; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm0 = ymm11[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,16,20,24,28,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm2[5],ymm0[6,7]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm3 = [0,0,0,8,0,0,0,8,0,0,0,8,0,0,0,8]
; AVX512DQ-FCP-NEXT:    vmovdqa 112(%rdi), %xmm1
; AVX512DQ-FCP-NEXT:    vpshufb %xmm3, %xmm1, %xmm5
; AVX512DQ-FCP-NEXT:    vmovdqa %xmm1, %xmm2
; AVX512DQ-FCP-NEXT:    vmovdqa 96(%rdi), %xmm1
; AVX512DQ-FCP-NEXT:    vpshufb %xmm3, %xmm1, %xmm3
; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm7 = xmm3[0],xmm5[0],xmm3[1],xmm5[1],xmm3[2],xmm5[2],xmm3[3],xmm5[3]
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm8 = [0,8,0,0,0,8,0,0,0,8,0,0,0,8,0,0]
; AVX512DQ-FCP-NEXT:    vmovdqa 80(%rdi), %xmm12
; AVX512DQ-FCP-NEXT:    vpshufb %xmm8, %xmm12, %xmm10
; AVX512DQ-FCP-NEXT:    vmovdqa 64(%rdi), %xmm5
; AVX512DQ-FCP-NEXT:    vpshufb %xmm8, %xmm5, %xmm8
; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm10[0],xmm8[1],xmm10[1],xmm8[2],xmm10[2],xmm8[3],xmm10[3]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} xmm7 = xmm8[0,1,2],xmm7[3]
; AVX512DQ-FCP-NEXT:    vpmovqb %zmm18, %xmm8
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} xmm7 = xmm8[0,1],xmm7[2,3]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm7[0,1,2,3],ymm0[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm0, %ymm22
; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm0 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,17,21,25,29]
; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm7 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,17,21,25,29,u,u,u,u]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm7[0,1,2,3,4,5,6],ymm0[7]
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm3 = [1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13]
; AVX512DQ-FCP-NEXT:    vpshufb %ymm3, %ymm9, %ymm7
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm3, %ymm27
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm3 = [1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13]
; AVX512DQ-FCP-NEXT:    vpshufb %ymm3, %ymm11, %ymm8
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm3, %ymm28
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm7 = ymm8[0,1,2,3,4],ymm7[5],ymm8[6,7]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm7[0,1,2,3,4,5],ymm0[6,7]
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm7 = [0,0,1,9,0,0,1,9,0,0,1,9,0,0,1,9]
; AVX512DQ-FCP-NEXT:    vpshufb %xmm7, %xmm2, %xmm8
; AVX512DQ-FCP-NEXT:    vpshufb %xmm7, %xmm1, %xmm7
; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm7 = xmm7[0],xmm8[0],xmm7[1],xmm8[1],xmm7[2],xmm8[2],xmm7[3],xmm8[3]
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm8 = [1,9,0,0,1,9,0,0,1,9,0,0,1,9,0,0]
; AVX512DQ-FCP-NEXT:    vpshufb %xmm8, %xmm12, %xmm10
; AVX512DQ-FCP-NEXT:    vpshufb %xmm8, %xmm5, %xmm8
; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm10[0],xmm8[1],xmm10[1],xmm8[2],xmm10[2],xmm8[3],xmm10[3]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} xmm7 = xmm8[0,1,2],xmm7[3]
; AVX512DQ-FCP-NEXT:    vpsrlq $8, %zmm18, %zmm8
; AVX512DQ-FCP-NEXT:    vpmovqb %zmm8, %xmm8
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} xmm7 = xmm8[0,1],xmm7[2,3]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm7[0,1,2,3],ymm0[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm0, %ymm20
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm3 = [2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14]
; AVX512DQ-FCP-NEXT:    vpshufb %ymm3, %ymm4, %ymm0
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm3, %ymm29
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm3 = [2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14]
; AVX512DQ-FCP-NEXT:    vpshufb %ymm3, %ymm6, %ymm7
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm3, %ymm30
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm7[0,1,2,3,4,5,6],ymm0[7]
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm3 = [2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14]
; AVX512DQ-FCP-NEXT:    vpshufb %ymm3, %ymm9, %ymm7
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm3, %ymm31
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm3 = [2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14]
; AVX512DQ-FCP-NEXT:    vpshufb %ymm3, %ymm11, %ymm8
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm3, %ymm26
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm7 = ymm8[0,1,2,3,4],ymm7[5],ymm8[6,7]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm7[0,1,2,3,4,5],ymm0[6,7]
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm7 = [0,0,2,10,0,0,2,10,0,0,2,10,0,0,2,10]
; AVX512DQ-FCP-NEXT:    vpshufb %xmm7, %xmm2, %xmm8
; AVX512DQ-FCP-NEXT:    vpshufb %xmm7, %xmm1, %xmm7
; AVX512DQ-FCP-NEXT:    vmovdqa %xmm1, %xmm3
; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm7 = xmm7[0],xmm8[0],xmm7[1],xmm8[1],xmm7[2],xmm8[2],xmm7[3],xmm8[3]
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm8 = [2,10,0,0,2,10,0,0,2,10,0,0,2,10,0,0]
; AVX512DQ-FCP-NEXT:    vpshufb %xmm8, %xmm12, %xmm10
; AVX512DQ-FCP-NEXT:    vpshufb %xmm8, %xmm5, %xmm8
; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm10[0],xmm8[1],xmm10[1],xmm8[2],xmm10[2],xmm8[3],xmm10[3]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} xmm7 = xmm8[0,1,2],xmm7[3]
; AVX512DQ-FCP-NEXT:    vpsrlq $16, %zmm18, %zmm8
; AVX512DQ-FCP-NEXT:    vpmovqb %zmm8, %xmm8
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} xmm7 = xmm8[0,1],xmm7[2,3]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm7[0,1,2,3],ymm0[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm0, %ymm21
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm1 = [3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15]
; AVX512DQ-FCP-NEXT:    vpshufb %ymm1, %ymm4, %ymm0
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm1, %ymm25
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm1 = [3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15]
; AVX512DQ-FCP-NEXT:    vpshufb %ymm1, %ymm6, %ymm4
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm1, %ymm24
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm4[0,1,2,3,4,5,6],ymm0[7]
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm1 = [3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15]
; AVX512DQ-FCP-NEXT:    vpshufb %ymm1, %ymm9, %ymm4
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm1, %ymm23
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm8 = [3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15]
; AVX512DQ-FCP-NEXT:    vpshufb %ymm8, %ymm11, %ymm6
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm4 = ymm6[0,1,2,3,4],ymm4[5],ymm6[6,7]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm4[0,1,2,3,4,5],ymm0[6,7]
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm4 = [0,0,3,11,0,0,3,11,0,0,3,11,0,0,3,11]
; AVX512DQ-FCP-NEXT:    vpshufb %xmm4, %xmm2, %xmm6
; AVX512DQ-FCP-NEXT:    vpshufb %xmm4, %xmm3, %xmm4
; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm4[0],xmm6[0],xmm4[1],xmm6[1],xmm4[2],xmm6[2],xmm4[3],xmm6[3]
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm6 = [3,11,0,0,3,11,0,0,3,11,0,0,3,11,0,0]
; AVX512DQ-FCP-NEXT:    vpshufb %xmm6, %xmm12, %xmm9
; AVX512DQ-FCP-NEXT:    vpshufb %xmm6, %xmm5, %xmm6
; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm6[0],xmm9[0],xmm6[1],xmm9[1],xmm6[2],xmm9[2],xmm6[3],xmm9[3]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} xmm4 = xmm6[0,1,2],xmm4[3]
; AVX512DQ-FCP-NEXT:    vpsrlq $24, %zmm18, %zmm6
; AVX512DQ-FCP-NEXT:    vpmovqb %zmm6, %xmm6
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} xmm4 = xmm6[0,1],xmm4[2,3]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm4[0,1,2,3],ymm0[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm0, %ymm19
; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm0 = [1,3,2,3,1,3,5,7]
; AVX512DQ-FCP-NEXT:    vpermd %ymm16, %ymm0, %ymm11
; AVX512DQ-FCP-NEXT:    vpshufb %ymm14, %ymm11, %ymm4
; AVX512DQ-FCP-NEXT:    vpermd %ymm17, %ymm0, %ymm9
; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm14 = ymm9[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,16,20,24,28,u,u,u,u]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm4 = ymm14[0,1,2,3,4,5,6],ymm4[7]
; AVX512DQ-FCP-NEXT:    vpermd %ymm13, %ymm0, %ymm13
; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm14 = ymm13[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,16,20,24,28,u,u,u,u,u,u,u,u]
; AVX512DQ-FCP-NEXT:    vpermd %ymm15, %ymm0, %ymm1
; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm15 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,16,20,24,28,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm14 = ymm15[0,1,2,3,4],ymm14[5],ymm15[6,7]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm4 = ymm14[0,1,2,3,4,5],ymm4[6,7]
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm14 = [0,0,4,12,0,0,4,12,0,0,4,12,0,0,4,12]
; AVX512DQ-FCP-NEXT:    vmovdqa %xmm2, %xmm6
; AVX512DQ-FCP-NEXT:    vpshufb %xmm14, %xmm2, %xmm15
; AVX512DQ-FCP-NEXT:    vpshufb %xmm14, %xmm3, %xmm14
; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm14[0],xmm15[0],xmm14[1],xmm15[1],xmm14[2],xmm15[2],xmm14[3],xmm15[3]
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm15 = [4,12,0,0,4,12,0,0,4,12,0,0,4,12,0,0]
; AVX512DQ-FCP-NEXT:    vpshufb %xmm15, %xmm12, %xmm0
; AVX512DQ-FCP-NEXT:    vpshufb %xmm15, %xmm5, %xmm15
; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm15[0],xmm0[0],xmm15[1],xmm0[1],xmm15[2],xmm0[2],xmm15[3],xmm0[3]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm0[0,1,2],xmm14[3]
; AVX512DQ-FCP-NEXT:    vpsrlq $32, %zmm18, %zmm14
; AVX512DQ-FCP-NEXT:    vpmovqb %zmm14, %xmm14
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm14[0,1],xmm0[2,3]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm15 = ymm0[0,1,2,3],ymm4[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm0 = ymm11[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,17,21,25,29]
; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm4 = ymm9[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,17,21,25,29,u,u,u,u]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm4[0,1,2,3,4,5,6],ymm0[7]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm27, %ymm2
; AVX512DQ-FCP-NEXT:    vpshufb %ymm2, %ymm13, %ymm4
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm28, %ymm2
; AVX512DQ-FCP-NEXT:    vpshufb %ymm2, %ymm1, %ymm14
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm4 = ymm14[0,1,2,3,4],ymm4[5],ymm14[6,7]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm2 = ymm4[0,1,2,3,4,5],ymm0[6,7]
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm4 = [0,0,5,13,0,0,5,13,0,0,5,13,0,0,5,13]
; AVX512DQ-FCP-NEXT:    vpshufb %xmm4, %xmm6, %xmm14
; AVX512DQ-FCP-NEXT:    vmovdqa %xmm3, %xmm7
; AVX512DQ-FCP-NEXT:    vpshufb %xmm4, %xmm3, %xmm4
; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm4[0],xmm14[0],xmm4[1],xmm14[1],xmm4[2],xmm14[2],xmm4[3],xmm14[3]
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm14 = [5,13,0,0,5,13,0,0,5,13,0,0,5,13,0,0]
; AVX512DQ-FCP-NEXT:    vpshufb %xmm14, %xmm12, %xmm0
; AVX512DQ-FCP-NEXT:    vpshufb %xmm14, %xmm5, %xmm14
; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm14[0],xmm0[0],xmm14[1],xmm0[1],xmm14[2],xmm0[2],xmm14[3],xmm0[3]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm0[0,1,2],xmm4[3]
; AVX512DQ-FCP-NEXT:    vpsrlq $40, %zmm18, %zmm4
; AVX512DQ-FCP-NEXT:    vpmovqb %zmm4, %xmm4
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm4[0,1],xmm0[2,3]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm10 = ymm0[0,1,2,3],ymm2[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm29, %ymm0
; AVX512DQ-FCP-NEXT:    vpshufb %ymm0, %ymm11, %ymm0
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm30, %ymm2
; AVX512DQ-FCP-NEXT:    vpshufb %ymm2, %ymm9, %ymm2
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3,4,5,6],ymm0[7]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm31, %ymm2
; AVX512DQ-FCP-NEXT:    vpshufb %ymm2, %ymm13, %ymm2
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm26, %ymm4
; AVX512DQ-FCP-NEXT:    vpshufb %ymm4, %ymm1, %ymm14
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm2 = ymm14[0,1,2,3,4],ymm2[5],ymm14[6,7]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm4 = ymm2[0,1,2,3,4,5],ymm0[6,7]
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm2 = [0,0,6,14,0,0,6,14,0,0,6,14,0,0,6,14]
; AVX512DQ-FCP-NEXT:    vpshufb %xmm2, %xmm6, %xmm14
; AVX512DQ-FCP-NEXT:    vpshufb %xmm2, %xmm3, %xmm2
; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm14[0],xmm2[1],xmm14[1],xmm2[2],xmm14[2],xmm2[3],xmm14[3]
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm14 = [6,14,0,0,6,14,0,0,6,14,0,0,6,14,0,0]
; AVX512DQ-FCP-NEXT:    vpshufb %xmm14, %xmm12, %xmm0
; AVX512DQ-FCP-NEXT:    vpshufb %xmm14, %xmm5, %xmm14
; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm14[0],xmm0[0],xmm14[1],xmm0[1],xmm14[2],xmm0[2],xmm14[3],xmm0[3]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm0[0,1,2],xmm2[3]
; AVX512DQ-FCP-NEXT:    vpsrlq $48, %zmm18, %zmm2
; AVX512DQ-FCP-NEXT:    vpmovqb %zmm2, %xmm2
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm2[0,1],xmm0[2,3]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm4[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm25, %ymm2
; AVX512DQ-FCP-NEXT:    vpshufb %ymm2, %ymm11, %ymm2
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm24, %ymm3
; AVX512DQ-FCP-NEXT:    vpshufb %ymm3, %ymm9, %ymm3
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4,5,6],ymm2[7]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm23, %ymm3
; AVX512DQ-FCP-NEXT:    vpshufb %ymm3, %ymm13, %ymm3
; AVX512DQ-FCP-NEXT:    vpshufb %ymm8, %ymm1, %ymm1
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3,4],ymm3[5],ymm1[6,7]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5],ymm2[6,7]
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm2 = [0,0,7,15,0,0,7,15,0,0,7,15,0,0,7,15]
; AVX512DQ-FCP-NEXT:    vpshufb %xmm2, %xmm6, %xmm3
; AVX512DQ-FCP-NEXT:    vpshufb %xmm2, %xmm7, %xmm2
; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm3 = [7,15,0,0,7,15,0,0,7,15,0,0,7,15,0,0]
; AVX512DQ-FCP-NEXT:    vpshufb %xmm3, %xmm12, %xmm8
; AVX512DQ-FCP-NEXT:    vpshufb %xmm3, %xmm5, %xmm3
; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm8[0],xmm3[1],xmm8[1],xmm3[2],xmm8[2],xmm3[3],xmm8[3]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} xmm2 = xmm3[0,1,2],xmm2[3]
; AVX512DQ-FCP-NEXT:    vpsrlq $56, %zmm18, %zmm3
; AVX512DQ-FCP-NEXT:    vpmovqb %zmm3, %xmm3
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} xmm2 = xmm3[0,1],xmm2[2,3]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm22, (%rsi)
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm20, (%rdx)
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm21, (%rcx)
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm19, (%r8)
; AVX512DQ-FCP-NEXT:    vmovdqa %ymm15, (%r9)
; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-FCP-NEXT:    vmovdqa %ymm10, (%rax)
; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-FCP-NEXT:    vmovdqa %ymm0, (%rax)
; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-FCP-NEXT:    vmovdqa %ymm1, (%rax)
; AVX512DQ-FCP-NEXT:    vzeroupper
; AVX512DQ-FCP-NEXT:    retq
;
; AVX512BW-LABEL: load_i8_stride8_vf32:
; AVX512BW:       # %bb.0:
; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX512BW-NEXT:    vmovdqa64 (%rdi), %zmm16
; AVX512BW-NEXT:    vmovdqa64 128(%rdi), %zmm0
; AVX512BW-NEXT:    vpmovqb %zmm0, %xmm0
; AVX512BW-NEXT:    vmovdqa 240(%rdi), %xmm2
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm4 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX512BW-NEXT:    vpshufb %xmm4, %xmm2, %xmm1
; AVX512BW-NEXT:    vmovdqa 224(%rdi), %xmm6
; AVX512BW-NEXT:    vpshufb %xmm4, %xmm6, %xmm3
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3]
; AVX512BW-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm1
; AVX512BW-NEXT:    vmovdqa 208(%rdi), %xmm7
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm5 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX512BW-NEXT:    vpshufb %xmm5, %xmm7, %xmm3
; AVX512BW-NEXT:    vmovdqa 192(%rdi), %xmm9
; AVX512BW-NEXT:    vpshufb %xmm5, %xmm9, %xmm8
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm8[0],xmm3[0],xmm8[1],xmm3[1],xmm8[2],xmm3[2],xmm8[3],xmm3[3]
; AVX512BW-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm3[0,1,2,3,4,5,6],ymm1[7]
; AVX512BW-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX512BW-NEXT:    vmovdqa 128(%rdi), %ymm3
; AVX512BW-NEXT:    vpmovqb %ymm3, %xmm3
; AVX512BW-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm3[0,1,2,3,4],ymm0[5],ymm3[6,7]
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
; AVX512BW-NEXT:    vmovdqa 112(%rdi), %xmm1
; AVX512BW-NEXT:    vpshufb %xmm4, %xmm1, %xmm8
; AVX512BW-NEXT:    vmovdqa 96(%rdi), %xmm3
; AVX512BW-NEXT:    vpshufb %xmm4, %xmm3, %xmm4
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm4[0],xmm8[0],xmm4[1],xmm8[1],xmm4[2],xmm8[2],xmm4[3],xmm8[3]
; AVX512BW-NEXT:    vmovdqa 80(%rdi), %xmm4
; AVX512BW-NEXT:    vpshufb %xmm5, %xmm4, %xmm11
; AVX512BW-NEXT:    vmovdqa 64(%rdi), %xmm8
; AVX512BW-NEXT:    vpshufb %xmm5, %xmm8, %xmm5
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm11[0],xmm5[1],xmm11[1],xmm5[2],xmm11[2],xmm5[3],xmm11[3]
; AVX512BW-NEXT:    vpblendd {{.*#+}} xmm5 = xmm5[0,1,2],xmm10[3]
; AVX512BW-NEXT:    vpmovqb %zmm16, %xmm10
; AVX512BW-NEXT:    vpblendd {{.*#+}} xmm5 = xmm10[0,1],xmm5[2,3]
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm5[0,1,2,3],ymm0[4,5,6,7]
; AVX512BW-NEXT:    vmovdqa64 %ymm0, %ymm20
; AVX512BW-NEXT:    vmovdqa64 128(%rdi), %xmm19
; AVX512BW-NEXT:    vmovdqa 144(%rdi), %xmm11
; AVX512BW-NEXT:    vmovdqa 160(%rdi), %xmm12
; AVX512BW-NEXT:    vmovdqa 176(%rdi), %xmm13
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm0 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX512BW-NEXT:    vpshufb %xmm0, %xmm2, %xmm5
; AVX512BW-NEXT:    vpshufb %xmm0, %xmm6, %xmm14
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm14[0],xmm5[0],xmm14[1],xmm5[1],xmm14[2],xmm5[2],xmm14[3],xmm5[3]
; AVX512BW-NEXT:    vinserti128 $1, %xmm5, %ymm0, %ymm5
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm14 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX512BW-NEXT:    vpshufb %xmm14, %xmm7, %xmm15
; AVX512BW-NEXT:    vpshufb %xmm14, %xmm9, %xmm17
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm15 = xmm17[0],xmm15[0],xmm17[1],xmm15[1],xmm17[2],xmm15[2],xmm17[3],xmm15[3]
; AVX512BW-NEXT:    vinserti128 $1, %xmm15, %ymm0, %ymm15
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm5 = ymm15[0,1,2,3,4,5,6],ymm5[7]
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm15 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX512BW-NEXT:    vpshufb %xmm15, %xmm13, %xmm17
; AVX512BW-NEXT:    vpshufb %xmm15, %xmm12, %xmm15
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm15 = xmm15[0],xmm17[0],xmm15[1],xmm17[1],xmm15[2],xmm17[2],xmm15[3],xmm17[3]
; AVX512BW-NEXT:    vinserti128 $1, %xmm15, %ymm0, %ymm15
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm17 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX512BW-NEXT:    vpshufb %xmm17, %xmm11, %xmm18
; AVX512BW-NEXT:    vpshufb %xmm17, %xmm19, %xmm17
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm17 = xmm17[0],xmm18[0],xmm17[1],xmm18[1],xmm17[2],xmm18[2],xmm17[3],xmm18[3]
; AVX512BW-NEXT:    vinserti32x4 $1, %xmm17, %ymm0, %ymm10
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm10 = ymm10[0,1,2,3,4],ymm15[5],ymm10[6,7]
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm5 = ymm10[0,1,2,3,4,5],ymm5[6,7]
; AVX512BW-NEXT:    vpshufb %xmm0, %xmm1, %xmm10
; AVX512BW-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm10[0],xmm0[1],xmm10[1],xmm0[2],xmm10[2],xmm0[3],xmm10[3]
; AVX512BW-NEXT:    vpshufb %xmm14, %xmm4, %xmm10
; AVX512BW-NEXT:    vpshufb %xmm14, %xmm8, %xmm14
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm14[0],xmm10[0],xmm14[1],xmm10[1],xmm14[2],xmm10[2],xmm14[3],xmm10[3]
; AVX512BW-NEXT:    vpblendd {{.*#+}} xmm0 = xmm10[0,1,2],xmm0[3]
; AVX512BW-NEXT:    vpsrlq $8, %zmm16, %zmm10
; AVX512BW-NEXT:    vpmovqb %zmm10, %xmm10
; AVX512BW-NEXT:    vpblendd {{.*#+}} xmm0 = xmm10[0,1],xmm0[2,3]
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm5[4,5,6,7]
; AVX512BW-NEXT:    vmovdqa64 %ymm0, %ymm21
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm0 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX512BW-NEXT:    vpshufb %xmm0, %xmm2, %xmm5
; AVX512BW-NEXT:    vpshufb %xmm0, %xmm6, %xmm10
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm10[0],xmm5[0],xmm10[1],xmm5[1],xmm10[2],xmm5[2],xmm10[3],xmm5[3]
; AVX512BW-NEXT:    vinserti128 $1, %xmm5, %ymm0, %ymm5
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm10 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX512BW-NEXT:    vpshufb %xmm10, %xmm7, %xmm14
; AVX512BW-NEXT:    vpshufb %xmm10, %xmm9, %xmm15
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3]
; AVX512BW-NEXT:    vinserti128 $1, %xmm14, %ymm0, %ymm14
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm5 = ymm14[0,1,2,3,4,5,6],ymm5[7]
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm14 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX512BW-NEXT:    vpshufb %xmm14, %xmm13, %xmm15
; AVX512BW-NEXT:    vpshufb %xmm14, %xmm12, %xmm14
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm14[0],xmm15[0],xmm14[1],xmm15[1],xmm14[2],xmm15[2],xmm14[3],xmm15[3]
; AVX512BW-NEXT:    vinserti128 $1, %xmm14, %ymm0, %ymm14
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm15 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX512BW-NEXT:    vpshufb %xmm15, %xmm11, %xmm17
; AVX512BW-NEXT:    vpshufb %xmm15, %xmm19, %xmm15
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm15 = xmm15[0],xmm17[0],xmm15[1],xmm17[1],xmm15[2],xmm17[2],xmm15[3],xmm17[3]
; AVX512BW-NEXT:    vinserti128 $1, %xmm15, %ymm0, %ymm15
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm14 = ymm15[0,1,2,3,4],ymm14[5],ymm15[6,7]
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm5 = ymm14[0,1,2,3,4,5],ymm5[6,7]
; AVX512BW-NEXT:    vpshufb %xmm0, %xmm1, %xmm14
; AVX512BW-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm14[0],xmm0[1],xmm14[1],xmm0[2],xmm14[2],xmm0[3],xmm14[3]
; AVX512BW-NEXT:    vpshufb %xmm10, %xmm4, %xmm14
; AVX512BW-NEXT:    vpshufb %xmm10, %xmm8, %xmm10
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm10[0],xmm14[0],xmm10[1],xmm14[1],xmm10[2],xmm14[2],xmm10[3],xmm14[3]
; AVX512BW-NEXT:    vpblendd {{.*#+}} xmm0 = xmm10[0,1,2],xmm0[3]
; AVX512BW-NEXT:    vpsrlq $16, %zmm16, %zmm10
; AVX512BW-NEXT:    vpmovqb %zmm10, %xmm10
; AVX512BW-NEXT:    vpblendd {{.*#+}} xmm0 = xmm10[0,1],xmm0[2,3]
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm5[4,5,6,7]
; AVX512BW-NEXT:    vmovdqa64 %ymm0, %ymm22
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm0 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX512BW-NEXT:    vpshufb %xmm0, %xmm2, %xmm5
; AVX512BW-NEXT:    vpshufb %xmm0, %xmm6, %xmm10
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm10[0],xmm5[0],xmm10[1],xmm5[1],xmm10[2],xmm5[2],xmm10[3],xmm5[3]
; AVX512BW-NEXT:    vinserti128 $1, %xmm5, %ymm0, %ymm5
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm10 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX512BW-NEXT:    vpshufb %xmm10, %xmm7, %xmm14
; AVX512BW-NEXT:    vpshufb %xmm10, %xmm9, %xmm15
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3]
; AVX512BW-NEXT:    vinserti128 $1, %xmm14, %ymm0, %ymm14
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm5 = ymm14[0,1,2,3,4,5,6],ymm5[7]
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm14 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX512BW-NEXT:    vpshufb %xmm14, %xmm13, %xmm15
; AVX512BW-NEXT:    vpshufb %xmm14, %xmm12, %xmm14
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm14[0],xmm15[0],xmm14[1],xmm15[1],xmm14[2],xmm15[2],xmm14[3],xmm15[3]
; AVX512BW-NEXT:    vinserti128 $1, %xmm14, %ymm0, %ymm14
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm15 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX512BW-NEXT:    vpshufb %xmm15, %xmm11, %xmm17
; AVX512BW-NEXT:    vpshufb %xmm15, %xmm19, %xmm15
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm15 = xmm15[0],xmm17[0],xmm15[1],xmm17[1],xmm15[2],xmm17[2],xmm15[3],xmm17[3]
; AVX512BW-NEXT:    vinserti128 $1, %xmm15, %ymm0, %ymm15
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm14 = ymm15[0,1,2,3,4],ymm14[5],ymm15[6,7]
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm5 = ymm14[0,1,2,3,4,5],ymm5[6,7]
; AVX512BW-NEXT:    vpshufb %xmm0, %xmm1, %xmm14
; AVX512BW-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm14[0],xmm0[1],xmm14[1],xmm0[2],xmm14[2],xmm0[3],xmm14[3]
; AVX512BW-NEXT:    vpshufb %xmm10, %xmm4, %xmm14
; AVX512BW-NEXT:    vpshufb %xmm10, %xmm8, %xmm10
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm10[0],xmm14[0],xmm10[1],xmm14[1],xmm10[2],xmm14[2],xmm10[3],xmm14[3]
; AVX512BW-NEXT:    vpblendd {{.*#+}} xmm0 = xmm10[0,1,2],xmm0[3]
; AVX512BW-NEXT:    vpsrlq $24, %zmm16, %zmm10
; AVX512BW-NEXT:    vpmovqb %zmm10, %xmm10
; AVX512BW-NEXT:    vpblendd {{.*#+}} xmm0 = xmm10[0,1],xmm0[2,3]
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm5[4,5,6,7]
; AVX512BW-NEXT:    vmovdqa64 %ymm0, %ymm23
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm0 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX512BW-NEXT:    vpshufb %xmm0, %xmm2, %xmm5
; AVX512BW-NEXT:    vpshufb %xmm0, %xmm6, %xmm10
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm10[0],xmm5[0],xmm10[1],xmm5[1],xmm10[2],xmm5[2],xmm10[3],xmm5[3]
; AVX512BW-NEXT:    vinserti128 $1, %xmm5, %ymm0, %ymm5
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm10 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX512BW-NEXT:    vpshufb %xmm10, %xmm7, %xmm14
; AVX512BW-NEXT:    vpshufb %xmm10, %xmm9, %xmm17
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm17[0],xmm14[0],xmm17[1],xmm14[1],xmm17[2],xmm14[2],xmm17[3],xmm14[3]
; AVX512BW-NEXT:    vinserti128 $1, %xmm14, %ymm0, %ymm14
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm5 = ymm14[0,1,2,3,4,5,6],ymm5[7]
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm14 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX512BW-NEXT:    vpshufb %xmm14, %xmm13, %xmm17
; AVX512BW-NEXT:    vpshufb %xmm14, %xmm12, %xmm14
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm14[0],xmm17[0],xmm14[1],xmm17[1],xmm14[2],xmm17[2],xmm14[3],xmm17[3]
; AVX512BW-NEXT:    vinserti128 $1, %xmm14, %ymm0, %ymm14
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm17 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX512BW-NEXT:    vpshufb %xmm17, %xmm11, %xmm18
; AVX512BW-NEXT:    vpshufb %xmm17, %xmm19, %xmm17
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm17 = xmm17[0],xmm18[0],xmm17[1],xmm18[1],xmm17[2],xmm18[2],xmm17[3],xmm18[3]
; AVX512BW-NEXT:    vinserti32x4 $1, %xmm17, %ymm0, %ymm15
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm14 = ymm15[0,1,2,3,4],ymm14[5],ymm15[6,7]
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm5 = ymm14[0,1,2,3,4,5],ymm5[6,7]
; AVX512BW-NEXT:    vpshufb %xmm0, %xmm1, %xmm14
; AVX512BW-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm14[0],xmm0[1],xmm14[1],xmm0[2],xmm14[2],xmm0[3],xmm14[3]
; AVX512BW-NEXT:    vpshufb %xmm10, %xmm4, %xmm14
; AVX512BW-NEXT:    vpshufb %xmm10, %xmm8, %xmm10
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm10[0],xmm14[0],xmm10[1],xmm14[1],xmm10[2],xmm14[2],xmm10[3],xmm14[3]
; AVX512BW-NEXT:    vpblendd {{.*#+}} xmm0 = xmm10[0,1,2],xmm0[3]
; AVX512BW-NEXT:    vpsrlq $32, %zmm16, %zmm10
; AVX512BW-NEXT:    vpmovqb %zmm10, %xmm10
; AVX512BW-NEXT:    vpblendd {{.*#+}} xmm0 = xmm10[0,1],xmm0[2,3]
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm5[4,5,6,7]
; AVX512BW-NEXT:    vmovdqa64 %ymm0, %ymm24
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm0 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX512BW-NEXT:    vpshufb %xmm0, %xmm2, %xmm5
; AVX512BW-NEXT:    vpshufb %xmm0, %xmm6, %xmm10
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm10[0],xmm5[0],xmm10[1],xmm5[1],xmm10[2],xmm5[2],xmm10[3],xmm5[3]
; AVX512BW-NEXT:    vinserti128 $1, %xmm5, %ymm0, %ymm5
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm10 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX512BW-NEXT:    vpshufb %xmm10, %xmm7, %xmm15
; AVX512BW-NEXT:    vpshufb %xmm10, %xmm9, %xmm17
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm15 = xmm17[0],xmm15[0],xmm17[1],xmm15[1],xmm17[2],xmm15[2],xmm17[3],xmm15[3]
; AVX512BW-NEXT:    vinserti128 $1, %xmm15, %ymm0, %ymm15
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm5 = ymm15[0,1,2,3,4,5,6],ymm5[7]
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm15 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX512BW-NEXT:    vpshufb %xmm15, %xmm13, %xmm17
; AVX512BW-NEXT:    vpshufb %xmm15, %xmm12, %xmm15
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm15 = xmm15[0],xmm17[0],xmm15[1],xmm17[1],xmm15[2],xmm17[2],xmm15[3],xmm17[3]
; AVX512BW-NEXT:    vinserti128 $1, %xmm15, %ymm0, %ymm15
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm17 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX512BW-NEXT:    vpshufb %xmm17, %xmm11, %xmm18
; AVX512BW-NEXT:    vpshufb %xmm17, %xmm19, %xmm17
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm17 = xmm17[0],xmm18[0],xmm17[1],xmm18[1],xmm17[2],xmm18[2],xmm17[3],xmm18[3]
; AVX512BW-NEXT:    vinserti32x4 $1, %xmm17, %ymm0, %ymm14
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm14 = ymm14[0,1,2,3,4],ymm15[5],ymm14[6,7]
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm5 = ymm14[0,1,2,3,4,5],ymm5[6,7]
; AVX512BW-NEXT:    vpshufb %xmm0, %xmm1, %xmm14
; AVX512BW-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm14[0],xmm0[1],xmm14[1],xmm0[2],xmm14[2],xmm0[3],xmm14[3]
; AVX512BW-NEXT:    vpshufb %xmm10, %xmm4, %xmm14
; AVX512BW-NEXT:    vpshufb %xmm10, %xmm8, %xmm10
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm10[0],xmm14[0],xmm10[1],xmm14[1],xmm10[2],xmm14[2],xmm10[3],xmm14[3]
; AVX512BW-NEXT:    vpblendd {{.*#+}} xmm0 = xmm10[0,1,2],xmm0[3]
; AVX512BW-NEXT:    vpsrlq $40, %zmm16, %zmm10
; AVX512BW-NEXT:    vpmovqb %zmm10, %xmm10
; AVX512BW-NEXT:    vpblendd {{.*#+}} xmm0 = xmm10[0,1],xmm0[2,3]
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm5[4,5,6,7]
; AVX512BW-NEXT:    vmovdqa64 %ymm0, %ymm25
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm0 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX512BW-NEXT:    vpshufb %xmm0, %xmm2, %xmm10
; AVX512BW-NEXT:    vpshufb %xmm0, %xmm6, %xmm14
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm14[0],xmm10[0],xmm14[1],xmm10[1],xmm14[2],xmm10[2],xmm14[3],xmm10[3]
; AVX512BW-NEXT:    vinserti128 $1, %xmm10, %ymm0, %ymm10
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm14 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX512BW-NEXT:    vpshufb %xmm14, %xmm7, %xmm15
; AVX512BW-NEXT:    vpshufb %xmm14, %xmm9, %xmm17
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm15 = xmm17[0],xmm15[0],xmm17[1],xmm15[1],xmm17[2],xmm15[2],xmm17[3],xmm15[3]
; AVX512BW-NEXT:    vinserti128 $1, %xmm15, %ymm0, %ymm15
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm10 = ymm15[0,1,2,3,4,5,6],ymm10[7]
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm15 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX512BW-NEXT:    vpshufb %xmm15, %xmm13, %xmm17
; AVX512BW-NEXT:    vpshufb %xmm15, %xmm12, %xmm15
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm15 = xmm15[0],xmm17[0],xmm15[1],xmm17[1],xmm15[2],xmm17[2],xmm15[3],xmm17[3]
; AVX512BW-NEXT:    vinserti128 $1, %xmm15, %ymm0, %ymm15
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm17 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX512BW-NEXT:    vpshufb %xmm17, %xmm11, %xmm18
; AVX512BW-NEXT:    vpshufb %xmm17, %xmm19, %xmm17
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm17 = xmm17[0],xmm18[0],xmm17[1],xmm18[1],xmm17[2],xmm18[2],xmm17[3],xmm18[3]
; AVX512BW-NEXT:    vinserti32x4 $1, %xmm17, %ymm0, %ymm5
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm5 = ymm5[0,1,2,3,4],ymm15[5],ymm5[6,7]
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm5 = ymm5[0,1,2,3,4,5],ymm10[6,7]
; AVX512BW-NEXT:    vpshufb %xmm0, %xmm1, %xmm10
; AVX512BW-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm10[0],xmm0[1],xmm10[1],xmm0[2],xmm10[2],xmm0[3],xmm10[3]
; AVX512BW-NEXT:    vpshufb %xmm14, %xmm4, %xmm10
; AVX512BW-NEXT:    vpshufb %xmm14, %xmm8, %xmm14
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm14[0],xmm10[0],xmm14[1],xmm10[1],xmm14[2],xmm10[2],xmm14[3],xmm10[3]
; AVX512BW-NEXT:    vpblendd {{.*#+}} xmm0 = xmm10[0,1,2],xmm0[3]
; AVX512BW-NEXT:    vpsrlq $48, %zmm16, %zmm10
; AVX512BW-NEXT:    vpmovqb %zmm10, %xmm10
; AVX512BW-NEXT:    vpblendd {{.*#+}} xmm0 = xmm10[0,1],xmm0[2,3]
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm5[4,5,6,7]
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm5 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX512BW-NEXT:    vpshufb %xmm5, %xmm2, %xmm2
; AVX512BW-NEXT:    vpshufb %xmm5, %xmm6, %xmm6
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm6[0],xmm2[0],xmm6[1],xmm2[1],xmm6[2],xmm2[2],xmm6[3],xmm2[3]
; AVX512BW-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm6 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX512BW-NEXT:    vpshufb %xmm6, %xmm7, %xmm7
; AVX512BW-NEXT:    vpshufb %xmm6, %xmm9, %xmm9
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm7 = xmm9[0],xmm7[0],xmm9[1],xmm7[1],xmm9[2],xmm7[2],xmm9[3],xmm7[3]
; AVX512BW-NEXT:    vinserti128 $1, %xmm7, %ymm0, %ymm7
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm2 = ymm7[0,1,2,3,4,5,6],ymm2[7]
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm7 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX512BW-NEXT:    vpshufb %xmm7, %xmm13, %xmm9
; AVX512BW-NEXT:    vpshufb %xmm7, %xmm12, %xmm7
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm7 = xmm7[0],xmm9[0],xmm7[1],xmm9[1],xmm7[2],xmm9[2],xmm7[3],xmm9[3]
; AVX512BW-NEXT:    vinserti128 $1, %xmm7, %ymm0, %ymm7
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm9 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX512BW-NEXT:    vpshufb %xmm9, %xmm11, %xmm10
; AVX512BW-NEXT:    vpshufb %xmm9, %xmm19, %xmm9
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm9[0],xmm10[0],xmm9[1],xmm10[1],xmm9[2],xmm10[2],xmm9[3],xmm10[3]
; AVX512BW-NEXT:    vinserti128 $1, %xmm9, %ymm0, %ymm9
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm7 = ymm9[0,1,2,3,4],ymm7[5],ymm9[6,7]
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm2 = ymm7[0,1,2,3,4,5],ymm2[6,7]
; AVX512BW-NEXT:    vpshufb %xmm5, %xmm1, %xmm1
; AVX512BW-NEXT:    vpshufb %xmm5, %xmm3, %xmm3
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3]
; AVX512BW-NEXT:    vpshufb %xmm6, %xmm4, %xmm3
; AVX512BW-NEXT:    vpshufb %xmm6, %xmm8, %xmm4
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX512BW-NEXT:    vpblendd {{.*#+}} xmm1 = xmm3[0,1,2],xmm1[3]
; AVX512BW-NEXT:    vpsrlq $56, %zmm16, %zmm3
; AVX512BW-NEXT:    vpmovqb %zmm3, %xmm3
; AVX512BW-NEXT:    vpblendd {{.*#+}} xmm1 = xmm3[0,1],xmm1[2,3]
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
; AVX512BW-NEXT:    vmovdqa64 %ymm20, (%rsi)
; AVX512BW-NEXT:    vmovdqa64 %ymm21, (%rdx)
; AVX512BW-NEXT:    vmovdqa64 %ymm22, (%rcx)
; AVX512BW-NEXT:    vmovdqa64 %ymm23, (%r8)
; AVX512BW-NEXT:    vmovdqa64 %ymm24, (%r9)
; AVX512BW-NEXT:    vmovdqa64 %ymm25, (%r11)
; AVX512BW-NEXT:    vmovdqa %ymm0, (%r10)
; AVX512BW-NEXT:    vmovdqa %ymm1, (%rax)
; AVX512BW-NEXT:    vzeroupper
; AVX512BW-NEXT:    retq
;
; AVX512BW-FCP-LABEL: load_i8_stride8_vf32:
; AVX512BW-FCP:       # %bb.0:
; AVX512BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX512BW-FCP-NEXT:    vmovdqa64 (%rdi), %zmm0
; AVX512BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm12 = [0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12]
; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm1 = [0,2,2,3,0,2,4,6]
; AVX512BW-FCP-NEXT:    vmovdqa 224(%rdi), %ymm9
; AVX512BW-FCP-NEXT:    vpermd %ymm9, %ymm1, %ymm4
; AVX512BW-FCP-NEXT:    vpshufb %ymm12, %ymm4, %ymm2
; AVX512BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm30 = [0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12]
; AVX512BW-FCP-NEXT:    vmovdqa 192(%rdi), %ymm11
; AVX512BW-FCP-NEXT:    vpermd %ymm11, %ymm1, %ymm26
; AVX512BW-FCP-NEXT:    vpshufb %ymm30, %ymm26, %ymm3
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4,5,6],ymm2[7]
; AVX512BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm31 = [0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12]
; AVX512BW-FCP-NEXT:    vmovdqa 160(%rdi), %ymm13
; AVX512BW-FCP-NEXT:    vpermd %ymm13, %ymm1, %ymm27
; AVX512BW-FCP-NEXT:    vpshufb %ymm31, %ymm27, %ymm3
; AVX512BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm14 = [0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12]
; AVX512BW-FCP-NEXT:    vmovdqa 128(%rdi), %ymm15
; AVX512BW-FCP-NEXT:    vpermd %ymm15, %ymm1, %ymm28
; AVX512BW-FCP-NEXT:    vpshufb %ymm14, %ymm28, %ymm1
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3,4],ymm3[5],ymm1[6,7]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm6 = ymm1[0,1,2,3,4,5],ymm2[6,7]
; AVX512BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm3 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX512BW-FCP-NEXT:    vmovdqa 112(%rdi), %xmm1
; AVX512BW-FCP-NEXT:    vpshufb %xmm3, %xmm1, %xmm5
; AVX512BW-FCP-NEXT:    vmovdqa 96(%rdi), %xmm2
; AVX512BW-FCP-NEXT:    vpshufb %xmm3, %xmm2, %xmm3
; AVX512BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm7 = xmm3[0],xmm5[0],xmm3[1],xmm5[1],xmm3[2],xmm5[2],xmm3[3],xmm5[3]
; AVX512BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm16 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX512BW-FCP-NEXT:    vmovdqa 80(%rdi), %xmm3
; AVX512BW-FCP-NEXT:    vpshufb %xmm16, %xmm3, %xmm17
; AVX512BW-FCP-NEXT:    vmovdqa 64(%rdi), %xmm5
; AVX512BW-FCP-NEXT:    vpshufb %xmm16, %xmm5, %xmm16
; AVX512BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm16[0],xmm17[0],xmm16[1],xmm17[1],xmm16[2],xmm17[2],xmm16[3],xmm17[3]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} xmm7 = xmm8[0,1,2],xmm7[3]
; AVX512BW-FCP-NEXT:    vpmovqb %zmm0, %xmm8
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} xmm7 = xmm8[0,1],xmm7[2,3]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]
; AVX512BW-FCP-NEXT:    vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm16 = [1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13]
; AVX512BW-FCP-NEXT:    vpshufb %ymm16, %ymm4, %ymm6
; AVX512BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm17 = [1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13]
; AVX512BW-FCP-NEXT:    vpshufb %ymm17, %ymm26, %ymm7
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm6 = ymm7[0,1,2,3,4,5,6],ymm6[7]
; AVX512BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm18 = [1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13]
; AVX512BW-FCP-NEXT:    vpshufb %ymm18, %ymm27, %ymm7
; AVX512BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm19 = [1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13]
; AVX512BW-FCP-NEXT:    vpshufb %ymm19, %ymm28, %ymm8
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm7 = ymm8[0,1,2,3,4],ymm7[5],ymm8[6,7]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm6 = ymm7[0,1,2,3,4,5],ymm6[6,7]
; AVX512BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm7 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX512BW-FCP-NEXT:    vpshufb %xmm7, %xmm1, %xmm8
; AVX512BW-FCP-NEXT:    vpshufb %xmm7, %xmm2, %xmm7
; AVX512BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm7 = xmm7[0],xmm8[0],xmm7[1],xmm8[1],xmm7[2],xmm8[2],xmm7[3],xmm8[3]
; AVX512BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm8 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX512BW-FCP-NEXT:    vpshufb %xmm8, %xmm3, %xmm20
; AVX512BW-FCP-NEXT:    vpshufb %xmm8, %xmm5, %xmm8
; AVX512BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm20[0],xmm8[1],xmm20[1],xmm8[2],xmm20[2],xmm8[3],xmm20[3]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} xmm7 = xmm8[0,1,2],xmm7[3]
; AVX512BW-FCP-NEXT:    vpsrlq $8, %zmm0, %zmm8
; AVX512BW-FCP-NEXT:    vpmovqb %zmm8, %xmm8
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} xmm7 = xmm8[0,1],xmm7[2,3]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]
; AVX512BW-FCP-NEXT:    vmovdqa64 %ymm6, %ymm29
; AVX512BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm20 = [2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14]
; AVX512BW-FCP-NEXT:    vpshufb %ymm20, %ymm4, %ymm7
; AVX512BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm21 = [2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14]
; AVX512BW-FCP-NEXT:    vpshufb %ymm21, %ymm26, %ymm8
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm7 = ymm8[0,1,2,3,4,5,6],ymm7[7]
; AVX512BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm22 = [2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14]
; AVX512BW-FCP-NEXT:    vpshufb %ymm22, %ymm27, %ymm8
; AVX512BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm23 = [2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14]
; AVX512BW-FCP-NEXT:    vpshufb %ymm23, %ymm28, %ymm10
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm8 = ymm10[0,1,2,3,4],ymm8[5],ymm10[6,7]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm7 = ymm8[0,1,2,3,4,5],ymm7[6,7]
; AVX512BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm8 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX512BW-FCP-NEXT:    vpshufb %xmm8, %xmm1, %xmm10
; AVX512BW-FCP-NEXT:    vpshufb %xmm8, %xmm2, %xmm8
; AVX512BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm10[0],xmm8[1],xmm10[1],xmm8[2],xmm10[2],xmm8[3],xmm10[3]
; AVX512BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm10 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX512BW-FCP-NEXT:    vpshufb %xmm10, %xmm3, %xmm24
; AVX512BW-FCP-NEXT:    vpshufb %xmm10, %xmm5, %xmm10
; AVX512BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm10[0],xmm24[0],xmm10[1],xmm24[1],xmm10[2],xmm24[2],xmm10[3],xmm24[3]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} xmm8 = xmm10[0,1,2],xmm8[3]
; AVX512BW-FCP-NEXT:    vpsrlq $16, %zmm0, %zmm10
; AVX512BW-FCP-NEXT:    vpmovqb %zmm10, %xmm10
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} xmm8 = xmm10[0,1],xmm8[2,3]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm7 = ymm8[0,1,2,3],ymm7[4,5,6,7]
; AVX512BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm24 = [3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15]
; AVX512BW-FCP-NEXT:    vpshufb %ymm24, %ymm4, %ymm4
; AVX512BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm25 = [3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15]
; AVX512BW-FCP-NEXT:    vpshufb %ymm25, %ymm26, %ymm8
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm4 = ymm8[0,1,2,3,4,5,6],ymm4[7]
; AVX512BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm26 = [3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15]
; AVX512BW-FCP-NEXT:    vpshufb %ymm26, %ymm27, %ymm8
; AVX512BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm27 = [3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15]
; AVX512BW-FCP-NEXT:    vpshufb %ymm27, %ymm28, %ymm10
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm8 = ymm10[0,1,2,3,4],ymm8[5],ymm10[6,7]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm4 = ymm8[0,1,2,3,4,5],ymm4[6,7]
; AVX512BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm8 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX512BW-FCP-NEXT:    vpshufb %xmm8, %xmm1, %xmm10
; AVX512BW-FCP-NEXT:    vpshufb %xmm8, %xmm2, %xmm8
; AVX512BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm10[0],xmm8[1],xmm10[1],xmm8[2],xmm10[2],xmm8[3],xmm10[3]
; AVX512BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm10 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX512BW-FCP-NEXT:    vpshufb %xmm10, %xmm3, %xmm28
; AVX512BW-FCP-NEXT:    vpshufb %xmm10, %xmm5, %xmm10
; AVX512BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm10[0],xmm28[0],xmm10[1],xmm28[1],xmm10[2],xmm28[2],xmm10[3],xmm28[3]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} xmm8 = xmm10[0,1,2],xmm8[3]
; AVX512BW-FCP-NEXT:    vpsrlq $24, %zmm0, %zmm10
; AVX512BW-FCP-NEXT:    vpmovqb %zmm10, %xmm10
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} xmm8 = xmm10[0,1],xmm8[2,3]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm4 = ymm8[0,1,2,3],ymm4[4,5,6,7]
; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm8 = [1,3,2,3,1,3,5,7]
; AVX512BW-FCP-NEXT:    vpermd %ymm9, %ymm8, %ymm9
; AVX512BW-FCP-NEXT:    vpshufb %ymm12, %ymm9, %ymm10
; AVX512BW-FCP-NEXT:    vpermd %ymm11, %ymm8, %ymm11
; AVX512BW-FCP-NEXT:    vpshufb %ymm30, %ymm11, %ymm12
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm12 = ymm12[0,1,2,3,4,5,6],ymm10[7]
; AVX512BW-FCP-NEXT:    vpermd %ymm13, %ymm8, %ymm10
; AVX512BW-FCP-NEXT:    vpshufb %ymm31, %ymm10, %ymm13
; AVX512BW-FCP-NEXT:    vpermd %ymm15, %ymm8, %ymm15
; AVX512BW-FCP-NEXT:    vpshufb %ymm14, %ymm15, %ymm8
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm8 = ymm8[0,1,2,3,4],ymm13[5],ymm8[6,7]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm8 = ymm8[0,1,2,3,4,5],ymm12[6,7]
; AVX512BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm12 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX512BW-FCP-NEXT:    vpshufb %xmm12, %xmm1, %xmm13
; AVX512BW-FCP-NEXT:    vpshufb %xmm12, %xmm2, %xmm12
; AVX512BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm12[0],xmm13[0],xmm12[1],xmm13[1],xmm12[2],xmm13[2],xmm12[3],xmm13[3]
; AVX512BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm13 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX512BW-FCP-NEXT:    vpshufb %xmm13, %xmm3, %xmm14
; AVX512BW-FCP-NEXT:    vpshufb %xmm13, %xmm5, %xmm13
; AVX512BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm13[0],xmm14[0],xmm13[1],xmm14[1],xmm13[2],xmm14[2],xmm13[3],xmm14[3]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} xmm12 = xmm13[0,1,2],xmm12[3]
; AVX512BW-FCP-NEXT:    vpsrlq $32, %zmm0, %zmm13
; AVX512BW-FCP-NEXT:    vpmovqb %zmm13, %xmm13
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} xmm12 = xmm13[0,1],xmm12[2,3]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm8 = ymm12[0,1,2,3],ymm8[4,5,6,7]
; AVX512BW-FCP-NEXT:    vpshufb %ymm16, %ymm9, %ymm12
; AVX512BW-FCP-NEXT:    vpshufb %ymm17, %ymm11, %ymm13
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm12 = ymm13[0,1,2,3,4,5,6],ymm12[7]
; AVX512BW-FCP-NEXT:    vpshufb %ymm18, %ymm10, %ymm13
; AVX512BW-FCP-NEXT:    vpshufb %ymm19, %ymm15, %ymm14
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm13 = ymm14[0,1,2,3,4],ymm13[5],ymm14[6,7]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm12 = ymm13[0,1,2,3,4,5],ymm12[6,7]
; AVX512BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm13 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX512BW-FCP-NEXT:    vpshufb %xmm13, %xmm1, %xmm14
; AVX512BW-FCP-NEXT:    vpshufb %xmm13, %xmm2, %xmm13
; AVX512BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm13[0],xmm14[0],xmm13[1],xmm14[1],xmm13[2],xmm14[2],xmm13[3],xmm14[3]
; AVX512BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm14 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX512BW-FCP-NEXT:    vpshufb %xmm14, %xmm3, %xmm16
; AVX512BW-FCP-NEXT:    vpshufb %xmm14, %xmm5, %xmm14
; AVX512BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm14[0],xmm16[0],xmm14[1],xmm16[1],xmm14[2],xmm16[2],xmm14[3],xmm16[3]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} xmm13 = xmm14[0,1,2],xmm13[3]
; AVX512BW-FCP-NEXT:    vpsrlq $40, %zmm0, %zmm14
; AVX512BW-FCP-NEXT:    vpmovqb %zmm14, %xmm14
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} xmm13 = xmm14[0,1],xmm13[2,3]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm12 = ymm13[0,1,2,3],ymm12[4,5,6,7]
; AVX512BW-FCP-NEXT:    vpshufb %ymm20, %ymm9, %ymm13
; AVX512BW-FCP-NEXT:    vpshufb %ymm21, %ymm11, %ymm14
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm13 = ymm14[0,1,2,3,4,5,6],ymm13[7]
; AVX512BW-FCP-NEXT:    vpshufb %ymm22, %ymm10, %ymm14
; AVX512BW-FCP-NEXT:    vpshufb %ymm23, %ymm15, %ymm6
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm6 = ymm6[0,1,2,3,4],ymm14[5],ymm6[6,7]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm6 = ymm6[0,1,2,3,4,5],ymm13[6,7]
; AVX512BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm13 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX512BW-FCP-NEXT:    vpshufb %xmm13, %xmm1, %xmm14
; AVX512BW-FCP-NEXT:    vpshufb %xmm13, %xmm2, %xmm13
; AVX512BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm13[0],xmm14[0],xmm13[1],xmm14[1],xmm13[2],xmm14[2],xmm13[3],xmm14[3]
; AVX512BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm14 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX512BW-FCP-NEXT:    vpshufb %xmm14, %xmm3, %xmm16
; AVX512BW-FCP-NEXT:    vpshufb %xmm14, %xmm5, %xmm14
; AVX512BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm14[0],xmm16[0],xmm14[1],xmm16[1],xmm14[2],xmm16[2],xmm14[3],xmm16[3]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} xmm13 = xmm14[0,1,2],xmm13[3]
; AVX512BW-FCP-NEXT:    vpsrlq $48, %zmm0, %zmm14
; AVX512BW-FCP-NEXT:    vpmovqb %zmm14, %xmm14
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} xmm13 = xmm14[0,1],xmm13[2,3]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm6 = ymm13[0,1,2,3],ymm6[4,5,6,7]
; AVX512BW-FCP-NEXT:    vpshufb %ymm24, %ymm9, %ymm9
; AVX512BW-FCP-NEXT:    vpshufb %ymm25, %ymm11, %ymm11
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm9 = ymm11[0,1,2,3,4,5,6],ymm9[7]
; AVX512BW-FCP-NEXT:    vpshufb %ymm26, %ymm10, %ymm10
; AVX512BW-FCP-NEXT:    vpshufb %ymm27, %ymm15, %ymm11
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm10 = ymm11[0,1,2,3,4],ymm10[5],ymm11[6,7]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm9 = ymm10[0,1,2,3,4,5],ymm9[6,7]
; AVX512BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm10 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX512BW-FCP-NEXT:    vpshufb %xmm10, %xmm1, %xmm1
; AVX512BW-FCP-NEXT:    vpshufb %xmm10, %xmm2, %xmm2
; AVX512BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX512BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm2 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX512BW-FCP-NEXT:    vpshufb %xmm2, %xmm3, %xmm3
; AVX512BW-FCP-NEXT:    vpshufb %xmm2, %xmm5, %xmm2
; AVX512BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} xmm1 = xmm2[0,1,2],xmm1[3]
; AVX512BW-FCP-NEXT:    vpsrlq $56, %zmm0, %zmm0
; AVX512BW-FCP-NEXT:    vpmovqb %zmm0, %xmm0
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm9[4,5,6,7]
; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX512BW-FCP-NEXT:    vmovaps %ymm1, (%rsi)
; AVX512BW-FCP-NEXT:    vmovdqa64 %ymm29, (%rdx)
; AVX512BW-FCP-NEXT:    vmovdqa %ymm7, (%rcx)
; AVX512BW-FCP-NEXT:    vmovdqa %ymm4, (%r8)
; AVX512BW-FCP-NEXT:    vmovdqa %ymm8, (%r9)
; AVX512BW-FCP-NEXT:    vmovdqa %ymm12, (%r11)
; AVX512BW-FCP-NEXT:    vmovdqa %ymm6, (%r10)
; AVX512BW-FCP-NEXT:    vmovdqa %ymm0, (%rax)
; AVX512BW-FCP-NEXT:    vzeroupper
; AVX512BW-FCP-NEXT:    retq
;
; AVX512DQ-BW-LABEL: load_i8_stride8_vf32:
; AVX512DQ-BW:       # %bb.0:
; AVX512DQ-BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-BW-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512DQ-BW-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX512DQ-BW-NEXT:    vmovdqa64 (%rdi), %zmm16
; AVX512DQ-BW-NEXT:    vmovdqa64 128(%rdi), %zmm0
; AVX512DQ-BW-NEXT:    vpmovqb %zmm0, %xmm0
; AVX512DQ-BW-NEXT:    vmovdqa 240(%rdi), %xmm2
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm4 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX512DQ-BW-NEXT:    vpshufb %xmm4, %xmm2, %xmm1
; AVX512DQ-BW-NEXT:    vmovdqa 224(%rdi), %xmm6
; AVX512DQ-BW-NEXT:    vpshufb %xmm4, %xmm6, %xmm3
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3]
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm1
; AVX512DQ-BW-NEXT:    vmovdqa 208(%rdi), %xmm7
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm5 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX512DQ-BW-NEXT:    vpshufb %xmm5, %xmm7, %xmm3
; AVX512DQ-BW-NEXT:    vmovdqa 192(%rdi), %xmm9
; AVX512DQ-BW-NEXT:    vpshufb %xmm5, %xmm9, %xmm8
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm8[0],xmm3[0],xmm8[1],xmm3[1],xmm8[2],xmm3[2],xmm8[3],xmm3[3]
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm3[0,1,2,3,4,5,6],ymm1[7]
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX512DQ-BW-NEXT:    vmovdqa 128(%rdi), %ymm3
; AVX512DQ-BW-NEXT:    vpmovqb %ymm3, %xmm3
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm3[0,1,2,3,4],ymm0[5],ymm3[6,7]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
; AVX512DQ-BW-NEXT:    vmovdqa 112(%rdi), %xmm1
; AVX512DQ-BW-NEXT:    vpshufb %xmm4, %xmm1, %xmm8
; AVX512DQ-BW-NEXT:    vmovdqa 96(%rdi), %xmm3
; AVX512DQ-BW-NEXT:    vpshufb %xmm4, %xmm3, %xmm4
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm4[0],xmm8[0],xmm4[1],xmm8[1],xmm4[2],xmm8[2],xmm4[3],xmm8[3]
; AVX512DQ-BW-NEXT:    vmovdqa 80(%rdi), %xmm4
; AVX512DQ-BW-NEXT:    vpshufb %xmm5, %xmm4, %xmm11
; AVX512DQ-BW-NEXT:    vmovdqa 64(%rdi), %xmm8
; AVX512DQ-BW-NEXT:    vpshufb %xmm5, %xmm8, %xmm5
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm11[0],xmm5[1],xmm11[1],xmm5[2],xmm11[2],xmm5[3],xmm11[3]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} xmm5 = xmm5[0,1,2],xmm10[3]
; AVX512DQ-BW-NEXT:    vpmovqb %zmm16, %xmm10
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} xmm5 = xmm10[0,1],xmm5[2,3]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm5[0,1,2,3],ymm0[4,5,6,7]
; AVX512DQ-BW-NEXT:    vmovdqa64 %ymm0, %ymm20
; AVX512DQ-BW-NEXT:    vmovdqa64 128(%rdi), %xmm19
; AVX512DQ-BW-NEXT:    vmovdqa 144(%rdi), %xmm11
; AVX512DQ-BW-NEXT:    vmovdqa 160(%rdi), %xmm12
; AVX512DQ-BW-NEXT:    vmovdqa 176(%rdi), %xmm13
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm0 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX512DQ-BW-NEXT:    vpshufb %xmm0, %xmm2, %xmm5
; AVX512DQ-BW-NEXT:    vpshufb %xmm0, %xmm6, %xmm14
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm14[0],xmm5[0],xmm14[1],xmm5[1],xmm14[2],xmm5[2],xmm14[3],xmm5[3]
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm5, %ymm0, %ymm5
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm14 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX512DQ-BW-NEXT:    vpshufb %xmm14, %xmm7, %xmm15
; AVX512DQ-BW-NEXT:    vpshufb %xmm14, %xmm9, %xmm17
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm15 = xmm17[0],xmm15[0],xmm17[1],xmm15[1],xmm17[2],xmm15[2],xmm17[3],xmm15[3]
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm15, %ymm0, %ymm15
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm5 = ymm15[0,1,2,3,4,5,6],ymm5[7]
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm15 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX512DQ-BW-NEXT:    vpshufb %xmm15, %xmm13, %xmm17
; AVX512DQ-BW-NEXT:    vpshufb %xmm15, %xmm12, %xmm15
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm15 = xmm15[0],xmm17[0],xmm15[1],xmm17[1],xmm15[2],xmm17[2],xmm15[3],xmm17[3]
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm15, %ymm0, %ymm15
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm17 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX512DQ-BW-NEXT:    vpshufb %xmm17, %xmm11, %xmm18
; AVX512DQ-BW-NEXT:    vpshufb %xmm17, %xmm19, %xmm17
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm17 = xmm17[0],xmm18[0],xmm17[1],xmm18[1],xmm17[2],xmm18[2],xmm17[3],xmm18[3]
; AVX512DQ-BW-NEXT:    vinserti32x4 $1, %xmm17, %ymm0, %ymm10
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm10 = ymm10[0,1,2,3,4],ymm15[5],ymm10[6,7]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm5 = ymm10[0,1,2,3,4,5],ymm5[6,7]
; AVX512DQ-BW-NEXT:    vpshufb %xmm0, %xmm1, %xmm10
; AVX512DQ-BW-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm10[0],xmm0[1],xmm10[1],xmm0[2],xmm10[2],xmm0[3],xmm10[3]
; AVX512DQ-BW-NEXT:    vpshufb %xmm14, %xmm4, %xmm10
; AVX512DQ-BW-NEXT:    vpshufb %xmm14, %xmm8, %xmm14
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm14[0],xmm10[0],xmm14[1],xmm10[1],xmm14[2],xmm10[2],xmm14[3],xmm10[3]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} xmm0 = xmm10[0,1,2],xmm0[3]
; AVX512DQ-BW-NEXT:    vpsrlq $8, %zmm16, %zmm10
; AVX512DQ-BW-NEXT:    vpmovqb %zmm10, %xmm10
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} xmm0 = xmm10[0,1],xmm0[2,3]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm5[4,5,6,7]
; AVX512DQ-BW-NEXT:    vmovdqa64 %ymm0, %ymm21
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm0 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX512DQ-BW-NEXT:    vpshufb %xmm0, %xmm2, %xmm5
; AVX512DQ-BW-NEXT:    vpshufb %xmm0, %xmm6, %xmm10
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm10[0],xmm5[0],xmm10[1],xmm5[1],xmm10[2],xmm5[2],xmm10[3],xmm5[3]
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm5, %ymm0, %ymm5
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm10 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX512DQ-BW-NEXT:    vpshufb %xmm10, %xmm7, %xmm14
; AVX512DQ-BW-NEXT:    vpshufb %xmm10, %xmm9, %xmm15
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3]
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm14, %ymm0, %ymm14
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm5 = ymm14[0,1,2,3,4,5,6],ymm5[7]
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm14 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX512DQ-BW-NEXT:    vpshufb %xmm14, %xmm13, %xmm15
; AVX512DQ-BW-NEXT:    vpshufb %xmm14, %xmm12, %xmm14
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm14[0],xmm15[0],xmm14[1],xmm15[1],xmm14[2],xmm15[2],xmm14[3],xmm15[3]
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm14, %ymm0, %ymm14
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm15 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX512DQ-BW-NEXT:    vpshufb %xmm15, %xmm11, %xmm17
; AVX512DQ-BW-NEXT:    vpshufb %xmm15, %xmm19, %xmm15
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm15 = xmm15[0],xmm17[0],xmm15[1],xmm17[1],xmm15[2],xmm17[2],xmm15[3],xmm17[3]
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm15, %ymm0, %ymm15
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm14 = ymm15[0,1,2,3,4],ymm14[5],ymm15[6,7]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm5 = ymm14[0,1,2,3,4,5],ymm5[6,7]
; AVX512DQ-BW-NEXT:    vpshufb %xmm0, %xmm1, %xmm14
; AVX512DQ-BW-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm14[0],xmm0[1],xmm14[1],xmm0[2],xmm14[2],xmm0[3],xmm14[3]
; AVX512DQ-BW-NEXT:    vpshufb %xmm10, %xmm4, %xmm14
; AVX512DQ-BW-NEXT:    vpshufb %xmm10, %xmm8, %xmm10
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm10[0],xmm14[0],xmm10[1],xmm14[1],xmm10[2],xmm14[2],xmm10[3],xmm14[3]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} xmm0 = xmm10[0,1,2],xmm0[3]
; AVX512DQ-BW-NEXT:    vpsrlq $16, %zmm16, %zmm10
; AVX512DQ-BW-NEXT:    vpmovqb %zmm10, %xmm10
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} xmm0 = xmm10[0,1],xmm0[2,3]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm5[4,5,6,7]
; AVX512DQ-BW-NEXT:    vmovdqa64 %ymm0, %ymm22
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm0 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX512DQ-BW-NEXT:    vpshufb %xmm0, %xmm2, %xmm5
; AVX512DQ-BW-NEXT:    vpshufb %xmm0, %xmm6, %xmm10
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm10[0],xmm5[0],xmm10[1],xmm5[1],xmm10[2],xmm5[2],xmm10[3],xmm5[3]
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm5, %ymm0, %ymm5
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm10 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX512DQ-BW-NEXT:    vpshufb %xmm10, %xmm7, %xmm14
; AVX512DQ-BW-NEXT:    vpshufb %xmm10, %xmm9, %xmm15
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3]
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm14, %ymm0, %ymm14
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm5 = ymm14[0,1,2,3,4,5,6],ymm5[7]
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm14 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX512DQ-BW-NEXT:    vpshufb %xmm14, %xmm13, %xmm15
; AVX512DQ-BW-NEXT:    vpshufb %xmm14, %xmm12, %xmm14
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm14[0],xmm15[0],xmm14[1],xmm15[1],xmm14[2],xmm15[2],xmm14[3],xmm15[3]
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm14, %ymm0, %ymm14
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm15 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX512DQ-BW-NEXT:    vpshufb %xmm15, %xmm11, %xmm17
; AVX512DQ-BW-NEXT:    vpshufb %xmm15, %xmm19, %xmm15
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm15 = xmm15[0],xmm17[0],xmm15[1],xmm17[1],xmm15[2],xmm17[2],xmm15[3],xmm17[3]
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm15, %ymm0, %ymm15
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm14 = ymm15[0,1,2,3,4],ymm14[5],ymm15[6,7]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm5 = ymm14[0,1,2,3,4,5],ymm5[6,7]
; AVX512DQ-BW-NEXT:    vpshufb %xmm0, %xmm1, %xmm14
; AVX512DQ-BW-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm14[0],xmm0[1],xmm14[1],xmm0[2],xmm14[2],xmm0[3],xmm14[3]
; AVX512DQ-BW-NEXT:    vpshufb %xmm10, %xmm4, %xmm14
; AVX512DQ-BW-NEXT:    vpshufb %xmm10, %xmm8, %xmm10
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm10[0],xmm14[0],xmm10[1],xmm14[1],xmm10[2],xmm14[2],xmm10[3],xmm14[3]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} xmm0 = xmm10[0,1,2],xmm0[3]
; AVX512DQ-BW-NEXT:    vpsrlq $24, %zmm16, %zmm10
; AVX512DQ-BW-NEXT:    vpmovqb %zmm10, %xmm10
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} xmm0 = xmm10[0,1],xmm0[2,3]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm5[4,5,6,7]
; AVX512DQ-BW-NEXT:    vmovdqa64 %ymm0, %ymm23
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm0 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX512DQ-BW-NEXT:    vpshufb %xmm0, %xmm2, %xmm5
; AVX512DQ-BW-NEXT:    vpshufb %xmm0, %xmm6, %xmm10
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm10[0],xmm5[0],xmm10[1],xmm5[1],xmm10[2],xmm5[2],xmm10[3],xmm5[3]
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm5, %ymm0, %ymm5
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm10 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX512DQ-BW-NEXT:    vpshufb %xmm10, %xmm7, %xmm14
; AVX512DQ-BW-NEXT:    vpshufb %xmm10, %xmm9, %xmm17
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm17[0],xmm14[0],xmm17[1],xmm14[1],xmm17[2],xmm14[2],xmm17[3],xmm14[3]
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm14, %ymm0, %ymm14
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm5 = ymm14[0,1,2,3,4,5,6],ymm5[7]
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm14 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX512DQ-BW-NEXT:    vpshufb %xmm14, %xmm13, %xmm17
; AVX512DQ-BW-NEXT:    vpshufb %xmm14, %xmm12, %xmm14
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm14[0],xmm17[0],xmm14[1],xmm17[1],xmm14[2],xmm17[2],xmm14[3],xmm17[3]
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm14, %ymm0, %ymm14
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm17 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX512DQ-BW-NEXT:    vpshufb %xmm17, %xmm11, %xmm18
; AVX512DQ-BW-NEXT:    vpshufb %xmm17, %xmm19, %xmm17
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm17 = xmm17[0],xmm18[0],xmm17[1],xmm18[1],xmm17[2],xmm18[2],xmm17[3],xmm18[3]
; AVX512DQ-BW-NEXT:    vinserti32x4 $1, %xmm17, %ymm0, %ymm15
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm14 = ymm15[0,1,2,3,4],ymm14[5],ymm15[6,7]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm5 = ymm14[0,1,2,3,4,5],ymm5[6,7]
; AVX512DQ-BW-NEXT:    vpshufb %xmm0, %xmm1, %xmm14
; AVX512DQ-BW-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm14[0],xmm0[1],xmm14[1],xmm0[2],xmm14[2],xmm0[3],xmm14[3]
; AVX512DQ-BW-NEXT:    vpshufb %xmm10, %xmm4, %xmm14
; AVX512DQ-BW-NEXT:    vpshufb %xmm10, %xmm8, %xmm10
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm10[0],xmm14[0],xmm10[1],xmm14[1],xmm10[2],xmm14[2],xmm10[3],xmm14[3]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} xmm0 = xmm10[0,1,2],xmm0[3]
; AVX512DQ-BW-NEXT:    vpsrlq $32, %zmm16, %zmm10
; AVX512DQ-BW-NEXT:    vpmovqb %zmm10, %xmm10
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} xmm0 = xmm10[0,1],xmm0[2,3]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm5[4,5,6,7]
; AVX512DQ-BW-NEXT:    vmovdqa64 %ymm0, %ymm24
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm0 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX512DQ-BW-NEXT:    vpshufb %xmm0, %xmm2, %xmm5
; AVX512DQ-BW-NEXT:    vpshufb %xmm0, %xmm6, %xmm10
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm10[0],xmm5[0],xmm10[1],xmm5[1],xmm10[2],xmm5[2],xmm10[3],xmm5[3]
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm5, %ymm0, %ymm5
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm10 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX512DQ-BW-NEXT:    vpshufb %xmm10, %xmm7, %xmm15
; AVX512DQ-BW-NEXT:    vpshufb %xmm10, %xmm9, %xmm17
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm15 = xmm17[0],xmm15[0],xmm17[1],xmm15[1],xmm17[2],xmm15[2],xmm17[3],xmm15[3]
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm15, %ymm0, %ymm15
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm5 = ymm15[0,1,2,3,4,5,6],ymm5[7]
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm15 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX512DQ-BW-NEXT:    vpshufb %xmm15, %xmm13, %xmm17
; AVX512DQ-BW-NEXT:    vpshufb %xmm15, %xmm12, %xmm15
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm15 = xmm15[0],xmm17[0],xmm15[1],xmm17[1],xmm15[2],xmm17[2],xmm15[3],xmm17[3]
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm15, %ymm0, %ymm15
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm17 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX512DQ-BW-NEXT:    vpshufb %xmm17, %xmm11, %xmm18
; AVX512DQ-BW-NEXT:    vpshufb %xmm17, %xmm19, %xmm17
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm17 = xmm17[0],xmm18[0],xmm17[1],xmm18[1],xmm17[2],xmm18[2],xmm17[3],xmm18[3]
; AVX512DQ-BW-NEXT:    vinserti32x4 $1, %xmm17, %ymm0, %ymm14
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm14 = ymm14[0,1,2,3,4],ymm15[5],ymm14[6,7]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm5 = ymm14[0,1,2,3,4,5],ymm5[6,7]
; AVX512DQ-BW-NEXT:    vpshufb %xmm0, %xmm1, %xmm14
; AVX512DQ-BW-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm14[0],xmm0[1],xmm14[1],xmm0[2],xmm14[2],xmm0[3],xmm14[3]
; AVX512DQ-BW-NEXT:    vpshufb %xmm10, %xmm4, %xmm14
; AVX512DQ-BW-NEXT:    vpshufb %xmm10, %xmm8, %xmm10
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm10[0],xmm14[0],xmm10[1],xmm14[1],xmm10[2],xmm14[2],xmm10[3],xmm14[3]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} xmm0 = xmm10[0,1,2],xmm0[3]
; AVX512DQ-BW-NEXT:    vpsrlq $40, %zmm16, %zmm10
; AVX512DQ-BW-NEXT:    vpmovqb %zmm10, %xmm10
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} xmm0 = xmm10[0,1],xmm0[2,3]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm5[4,5,6,7]
; AVX512DQ-BW-NEXT:    vmovdqa64 %ymm0, %ymm25
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm0 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX512DQ-BW-NEXT:    vpshufb %xmm0, %xmm2, %xmm10
; AVX512DQ-BW-NEXT:    vpshufb %xmm0, %xmm6, %xmm14
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm14[0],xmm10[0],xmm14[1],xmm10[1],xmm14[2],xmm10[2],xmm14[3],xmm10[3]
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm10, %ymm0, %ymm10
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm14 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX512DQ-BW-NEXT:    vpshufb %xmm14, %xmm7, %xmm15
; AVX512DQ-BW-NEXT:    vpshufb %xmm14, %xmm9, %xmm17
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm15 = xmm17[0],xmm15[0],xmm17[1],xmm15[1],xmm17[2],xmm15[2],xmm17[3],xmm15[3]
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm15, %ymm0, %ymm15
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm10 = ymm15[0,1,2,3,4,5,6],ymm10[7]
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm15 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX512DQ-BW-NEXT:    vpshufb %xmm15, %xmm13, %xmm17
; AVX512DQ-BW-NEXT:    vpshufb %xmm15, %xmm12, %xmm15
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm15 = xmm15[0],xmm17[0],xmm15[1],xmm17[1],xmm15[2],xmm17[2],xmm15[3],xmm17[3]
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm15, %ymm0, %ymm15
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm17 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX512DQ-BW-NEXT:    vpshufb %xmm17, %xmm11, %xmm18
; AVX512DQ-BW-NEXT:    vpshufb %xmm17, %xmm19, %xmm17
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm17 = xmm17[0],xmm18[0],xmm17[1],xmm18[1],xmm17[2],xmm18[2],xmm17[3],xmm18[3]
; AVX512DQ-BW-NEXT:    vinserti32x4 $1, %xmm17, %ymm0, %ymm5
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm5 = ymm5[0,1,2,3,4],ymm15[5],ymm5[6,7]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm5 = ymm5[0,1,2,3,4,5],ymm10[6,7]
; AVX512DQ-BW-NEXT:    vpshufb %xmm0, %xmm1, %xmm10
; AVX512DQ-BW-NEXT:    vpshufb %xmm0, %xmm3, %xmm0
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm10[0],xmm0[1],xmm10[1],xmm0[2],xmm10[2],xmm0[3],xmm10[3]
; AVX512DQ-BW-NEXT:    vpshufb %xmm14, %xmm4, %xmm10
; AVX512DQ-BW-NEXT:    vpshufb %xmm14, %xmm8, %xmm14
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm14[0],xmm10[0],xmm14[1],xmm10[1],xmm14[2],xmm10[2],xmm14[3],xmm10[3]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} xmm0 = xmm10[0,1,2],xmm0[3]
; AVX512DQ-BW-NEXT:    vpsrlq $48, %zmm16, %zmm10
; AVX512DQ-BW-NEXT:    vpmovqb %zmm10, %xmm10
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} xmm0 = xmm10[0,1],xmm0[2,3]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm5[4,5,6,7]
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm5 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX512DQ-BW-NEXT:    vpshufb %xmm5, %xmm2, %xmm2
; AVX512DQ-BW-NEXT:    vpshufb %xmm5, %xmm6, %xmm6
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm6[0],xmm2[0],xmm6[1],xmm2[1],xmm6[2],xmm2[2],xmm6[3],xmm2[3]
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm6 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX512DQ-BW-NEXT:    vpshufb %xmm6, %xmm7, %xmm7
; AVX512DQ-BW-NEXT:    vpshufb %xmm6, %xmm9, %xmm9
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm7 = xmm9[0],xmm7[0],xmm9[1],xmm7[1],xmm9[2],xmm7[2],xmm9[3],xmm7[3]
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm7, %ymm0, %ymm7
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm2 = ymm7[0,1,2,3,4,5,6],ymm2[7]
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm7 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX512DQ-BW-NEXT:    vpshufb %xmm7, %xmm13, %xmm9
; AVX512DQ-BW-NEXT:    vpshufb %xmm7, %xmm12, %xmm7
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm7 = xmm7[0],xmm9[0],xmm7[1],xmm9[1],xmm7[2],xmm9[2],xmm7[3],xmm9[3]
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm7, %ymm0, %ymm7
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm9 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX512DQ-BW-NEXT:    vpshufb %xmm9, %xmm11, %xmm10
; AVX512DQ-BW-NEXT:    vpshufb %xmm9, %xmm19, %xmm9
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm9[0],xmm10[0],xmm9[1],xmm10[1],xmm9[2],xmm10[2],xmm9[3],xmm10[3]
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm9, %ymm0, %ymm9
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm7 = ymm9[0,1,2,3,4],ymm7[5],ymm9[6,7]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm2 = ymm7[0,1,2,3,4,5],ymm2[6,7]
; AVX512DQ-BW-NEXT:    vpshufb %xmm5, %xmm1, %xmm1
; AVX512DQ-BW-NEXT:    vpshufb %xmm5, %xmm3, %xmm3
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3]
; AVX512DQ-BW-NEXT:    vpshufb %xmm6, %xmm4, %xmm3
; AVX512DQ-BW-NEXT:    vpshufb %xmm6, %xmm8, %xmm4
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} xmm1 = xmm3[0,1,2],xmm1[3]
; AVX512DQ-BW-NEXT:    vpsrlq $56, %zmm16, %zmm3
; AVX512DQ-BW-NEXT:    vpmovqb %zmm3, %xmm3
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} xmm1 = xmm3[0,1],xmm1[2,3]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
; AVX512DQ-BW-NEXT:    vmovdqa64 %ymm20, (%rsi)
; AVX512DQ-BW-NEXT:    vmovdqa64 %ymm21, (%rdx)
; AVX512DQ-BW-NEXT:    vmovdqa64 %ymm22, (%rcx)
; AVX512DQ-BW-NEXT:    vmovdqa64 %ymm23, (%r8)
; AVX512DQ-BW-NEXT:    vmovdqa64 %ymm24, (%r9)
; AVX512DQ-BW-NEXT:    vmovdqa64 %ymm25, (%r11)
; AVX512DQ-BW-NEXT:    vmovdqa %ymm0, (%r10)
; AVX512DQ-BW-NEXT:    vmovdqa %ymm1, (%rax)
; AVX512DQ-BW-NEXT:    vzeroupper
; AVX512DQ-BW-NEXT:    retq
;
; AVX512DQ-BW-FCP-LABEL: load_i8_stride8_vf32:
; AVX512DQ-BW-FCP:       # %bb.0:
; AVX512DQ-BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r10
; AVX512DQ-BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r11
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 (%rdi), %zmm0
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm12 = [0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12]
; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm1 = [0,2,2,3,0,2,4,6]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 224(%rdi), %ymm9
; AVX512DQ-BW-FCP-NEXT:    vpermd %ymm9, %ymm1, %ymm4
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm12, %ymm4, %ymm2
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm30 = [0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 192(%rdi), %ymm11
; AVX512DQ-BW-FCP-NEXT:    vpermd %ymm11, %ymm1, %ymm26
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm30, %ymm26, %ymm3
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4,5,6],ymm2[7]
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm31 = [0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 160(%rdi), %ymm13
; AVX512DQ-BW-FCP-NEXT:    vpermd %ymm13, %ymm1, %ymm27
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm31, %ymm27, %ymm3
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm14 = [0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 128(%rdi), %ymm15
; AVX512DQ-BW-FCP-NEXT:    vpermd %ymm15, %ymm1, %ymm28
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm14, %ymm28, %ymm1
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3,4],ymm3[5],ymm1[6,7]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm6 = ymm1[0,1,2,3,4,5],ymm2[6,7]
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm3 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 112(%rdi), %xmm1
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm3, %xmm1, %xmm5
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 96(%rdi), %xmm2
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm3, %xmm2, %xmm3
; AVX512DQ-BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm7 = xmm3[0],xmm5[0],xmm3[1],xmm5[1],xmm3[2],xmm5[2],xmm3[3],xmm5[3]
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm16 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 80(%rdi), %xmm3
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm16, %xmm3, %xmm17
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 64(%rdi), %xmm5
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm16, %xmm5, %xmm16
; AVX512DQ-BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm16[0],xmm17[0],xmm16[1],xmm17[1],xmm16[2],xmm17[2],xmm16[3],xmm17[3]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} xmm7 = xmm8[0,1,2],xmm7[3]
; AVX512DQ-BW-FCP-NEXT:    vpmovqb %zmm0, %xmm8
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} xmm7 = xmm8[0,1],xmm7[2,3]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm16 = [1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm16, %ymm4, %ymm6
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm17 = [1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm17, %ymm26, %ymm7
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm6 = ymm7[0,1,2,3,4,5,6],ymm6[7]
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm18 = [1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm18, %ymm27, %ymm7
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm19 = [1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm19, %ymm28, %ymm8
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm7 = ymm8[0,1,2,3,4],ymm7[5],ymm8[6,7]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm6 = ymm7[0,1,2,3,4,5],ymm6[6,7]
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm7 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm7, %xmm1, %xmm8
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm7, %xmm2, %xmm7
; AVX512DQ-BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm7 = xmm7[0],xmm8[0],xmm7[1],xmm8[1],xmm7[2],xmm8[2],xmm7[3],xmm8[3]
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm8 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm8, %xmm3, %xmm20
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm8, %xmm5, %xmm8
; AVX512DQ-BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm20[0],xmm8[1],xmm20[1],xmm8[2],xmm20[2],xmm8[3],xmm20[3]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} xmm7 = xmm8[0,1,2],xmm7[3]
; AVX512DQ-BW-FCP-NEXT:    vpsrlq $8, %zmm0, %zmm8
; AVX512DQ-BW-FCP-NEXT:    vpmovqb %zmm8, %xmm8
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} xmm7 = xmm8[0,1],xmm7[2,3]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %ymm6, %ymm29
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm20 = [2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm20, %ymm4, %ymm7
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm21 = [2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm21, %ymm26, %ymm8
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm7 = ymm8[0,1,2,3,4,5,6],ymm7[7]
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm22 = [2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm22, %ymm27, %ymm8
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm23 = [2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm23, %ymm28, %ymm10
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm8 = ymm10[0,1,2,3,4],ymm8[5],ymm10[6,7]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm7 = ymm8[0,1,2,3,4,5],ymm7[6,7]
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm8 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm8, %xmm1, %xmm10
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm8, %xmm2, %xmm8
; AVX512DQ-BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm10[0],xmm8[1],xmm10[1],xmm8[2],xmm10[2],xmm8[3],xmm10[3]
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm10 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm10, %xmm3, %xmm24
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm10, %xmm5, %xmm10
; AVX512DQ-BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm10[0],xmm24[0],xmm10[1],xmm24[1],xmm10[2],xmm24[2],xmm10[3],xmm24[3]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} xmm8 = xmm10[0,1,2],xmm8[3]
; AVX512DQ-BW-FCP-NEXT:    vpsrlq $16, %zmm0, %zmm10
; AVX512DQ-BW-FCP-NEXT:    vpmovqb %zmm10, %xmm10
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} xmm8 = xmm10[0,1],xmm8[2,3]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm7 = ymm8[0,1,2,3],ymm7[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm24 = [3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm24, %ymm4, %ymm4
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm25 = [3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm25, %ymm26, %ymm8
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm4 = ymm8[0,1,2,3,4,5,6],ymm4[7]
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm26 = [3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm26, %ymm27, %ymm8
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm27 = [3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm27, %ymm28, %ymm10
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm8 = ymm10[0,1,2,3,4],ymm8[5],ymm10[6,7]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm4 = ymm8[0,1,2,3,4,5],ymm4[6,7]
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm8 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm8, %xmm1, %xmm10
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm8, %xmm2, %xmm8
; AVX512DQ-BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm10[0],xmm8[1],xmm10[1],xmm8[2],xmm10[2],xmm8[3],xmm10[3]
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm10 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm10, %xmm3, %xmm28
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm10, %xmm5, %xmm10
; AVX512DQ-BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm10[0],xmm28[0],xmm10[1],xmm28[1],xmm10[2],xmm28[2],xmm10[3],xmm28[3]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} xmm8 = xmm10[0,1,2],xmm8[3]
; AVX512DQ-BW-FCP-NEXT:    vpsrlq $24, %zmm0, %zmm10
; AVX512DQ-BW-FCP-NEXT:    vpmovqb %zmm10, %xmm10
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} xmm8 = xmm10[0,1],xmm8[2,3]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm4 = ymm8[0,1,2,3],ymm4[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm8 = [1,3,2,3,1,3,5,7]
; AVX512DQ-BW-FCP-NEXT:    vpermd %ymm9, %ymm8, %ymm9
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm12, %ymm9, %ymm10
; AVX512DQ-BW-FCP-NEXT:    vpermd %ymm11, %ymm8, %ymm11
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm30, %ymm11, %ymm12
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm12 = ymm12[0,1,2,3,4,5,6],ymm10[7]
; AVX512DQ-BW-FCP-NEXT:    vpermd %ymm13, %ymm8, %ymm10
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm31, %ymm10, %ymm13
; AVX512DQ-BW-FCP-NEXT:    vpermd %ymm15, %ymm8, %ymm15
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm14, %ymm15, %ymm8
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm8 = ymm8[0,1,2,3,4],ymm13[5],ymm8[6,7]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm8 = ymm8[0,1,2,3,4,5],ymm12[6,7]
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm12 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm12, %xmm1, %xmm13
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm12, %xmm2, %xmm12
; AVX512DQ-BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm12[0],xmm13[0],xmm12[1],xmm13[1],xmm12[2],xmm13[2],xmm12[3],xmm13[3]
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm13 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm13, %xmm3, %xmm14
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm13, %xmm5, %xmm13
; AVX512DQ-BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm13[0],xmm14[0],xmm13[1],xmm14[1],xmm13[2],xmm14[2],xmm13[3],xmm14[3]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} xmm12 = xmm13[0,1,2],xmm12[3]
; AVX512DQ-BW-FCP-NEXT:    vpsrlq $32, %zmm0, %zmm13
; AVX512DQ-BW-FCP-NEXT:    vpmovqb %zmm13, %xmm13
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} xmm12 = xmm13[0,1],xmm12[2,3]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm8 = ymm12[0,1,2,3],ymm8[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm16, %ymm9, %ymm12
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm17, %ymm11, %ymm13
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm12 = ymm13[0,1,2,3,4,5,6],ymm12[7]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm18, %ymm10, %ymm13
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm19, %ymm15, %ymm14
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm13 = ymm14[0,1,2,3,4],ymm13[5],ymm14[6,7]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm12 = ymm13[0,1,2,3,4,5],ymm12[6,7]
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm13 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm13, %xmm1, %xmm14
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm13, %xmm2, %xmm13
; AVX512DQ-BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm13[0],xmm14[0],xmm13[1],xmm14[1],xmm13[2],xmm14[2],xmm13[3],xmm14[3]
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm14 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm14, %xmm3, %xmm16
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm14, %xmm5, %xmm14
; AVX512DQ-BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm14[0],xmm16[0],xmm14[1],xmm16[1],xmm14[2],xmm16[2],xmm14[3],xmm16[3]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} xmm13 = xmm14[0,1,2],xmm13[3]
; AVX512DQ-BW-FCP-NEXT:    vpsrlq $40, %zmm0, %zmm14
; AVX512DQ-BW-FCP-NEXT:    vpmovqb %zmm14, %xmm14
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} xmm13 = xmm14[0,1],xmm13[2,3]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm12 = ymm13[0,1,2,3],ymm12[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm20, %ymm9, %ymm13
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm21, %ymm11, %ymm14
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm13 = ymm14[0,1,2,3,4,5,6],ymm13[7]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm22, %ymm10, %ymm14
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm23, %ymm15, %ymm6
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm6 = ymm6[0,1,2,3,4],ymm14[5],ymm6[6,7]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm6 = ymm6[0,1,2,3,4,5],ymm13[6,7]
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm13 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm13, %xmm1, %xmm14
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm13, %xmm2, %xmm13
; AVX512DQ-BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm13[0],xmm14[0],xmm13[1],xmm14[1],xmm13[2],xmm14[2],xmm13[3],xmm14[3]
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm14 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm14, %xmm3, %xmm16
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm14, %xmm5, %xmm14
; AVX512DQ-BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm14[0],xmm16[0],xmm14[1],xmm16[1],xmm14[2],xmm16[2],xmm14[3],xmm16[3]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} xmm13 = xmm14[0,1,2],xmm13[3]
; AVX512DQ-BW-FCP-NEXT:    vpsrlq $48, %zmm0, %zmm14
; AVX512DQ-BW-FCP-NEXT:    vpmovqb %zmm14, %xmm14
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} xmm13 = xmm14[0,1],xmm13[2,3]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm6 = ymm13[0,1,2,3],ymm6[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm24, %ymm9, %ymm9
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm25, %ymm11, %ymm11
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm9 = ymm11[0,1,2,3,4,5,6],ymm9[7]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm26, %ymm10, %ymm10
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm27, %ymm15, %ymm11
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm10 = ymm11[0,1,2,3,4],ymm10[5],ymm11[6,7]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm9 = ymm10[0,1,2,3,4,5],ymm9[6,7]
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm10 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm10, %xmm1, %xmm1
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm10, %xmm2, %xmm2
; AVX512DQ-BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm2 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm2, %xmm3, %xmm3
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm2, %xmm5, %xmm2
; AVX512DQ-BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} xmm1 = xmm2[0,1,2],xmm1[3]
; AVX512DQ-BW-FCP-NEXT:    vpsrlq $56, %zmm0, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vpmovqb %zmm0, %xmm0
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm9[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovaps %ymm1, (%rsi)
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %ymm29, (%rdx)
; AVX512DQ-BW-FCP-NEXT:    vmovdqa %ymm7, (%rcx)
; AVX512DQ-BW-FCP-NEXT:    vmovdqa %ymm4, (%r8)
; AVX512DQ-BW-FCP-NEXT:    vmovdqa %ymm8, (%r9)
; AVX512DQ-BW-FCP-NEXT:    vmovdqa %ymm12, (%r11)
; AVX512DQ-BW-FCP-NEXT:    vmovdqa %ymm6, (%r10)
; AVX512DQ-BW-FCP-NEXT:    vmovdqa %ymm0, (%rax)
; AVX512DQ-BW-FCP-NEXT:    vzeroupper
; AVX512DQ-BW-FCP-NEXT:    retq
  %wide.vec = load <256 x i8>, ptr %in.vec, align 64
  %strided.vec0 = shufflevector <256 x i8> %wide.vec, <256 x i8> poison, <32 x i32> <i32 0, i32 8, i32 16, i32 24, i32 32, i32 40, i32 48, i32 56, i32 64, i32 72, i32 80, i32 88, i32 96, i32 104, i32 112, i32 120, i32 128, i32 136, i32 144, i32 152, i32 160, i32 168, i32 176, i32 184, i32 192, i32 200, i32 208, i32 216, i32 224, i32 232, i32 240, i32 248>
  %strided.vec1 = shufflevector <256 x i8> %wide.vec, <256 x i8> poison, <32 x i32> <i32 1, i32 9, i32 17, i32 25, i32 33, i32 41, i32 49, i32 57, i32 65, i32 73, i32 81, i32 89, i32 97, i32 105, i32 113, i32 121, i32 129, i32 137, i32 145, i32 153, i32 161, i32 169, i32 177, i32 185, i32 193, i32 201, i32 209, i32 217, i32 225, i32 233, i32 241, i32 249>
  %strided.vec2 = shufflevector <256 x i8> %wide.vec, <256 x i8> poison, <32 x i32> <i32 2, i32 10, i32 18, i32 26, i32 34, i32 42, i32 50, i32 58, i32 66, i32 74, i32 82, i32 90, i32 98, i32 106, i32 114, i32 122, i32 130, i32 138, i32 146, i32 154, i32 162, i32 170, i32 178, i32 186, i32 194, i32 202, i32 210, i32 218, i32 226, i32 234, i32 242, i32 250>
  %strided.vec3 = shufflevector <256 x i8> %wide.vec, <256 x i8> poison, <32 x i32> <i32 3, i32 11, i32 19, i32 27, i32 35, i32 43, i32 51, i32 59, i32 67, i32 75, i32 83, i32 91, i32 99, i32 107, i32 115, i32 123, i32 131, i32 139, i32 147, i32 155, i32 163, i32 171, i32 179, i32 187, i32 195, i32 203, i32 211, i32 219, i32 227, i32 235, i32 243, i32 251>
  %strided.vec4 = shufflevector <256 x i8> %wide.vec, <256 x i8> poison, <32 x i32> <i32 4, i32 12, i32 20, i32 28, i32 36, i32 44, i32 52, i32 60, i32 68, i32 76, i32 84, i32 92, i32 100, i32 108, i32 116, i32 124, i32 132, i32 140, i32 148, i32 156, i32 164, i32 172, i32 180, i32 188, i32 196, i32 204, i32 212, i32 220, i32 228, i32 236, i32 244, i32 252>
  %strided.vec5 = shufflevector <256 x i8> %wide.vec, <256 x i8> poison, <32 x i32> <i32 5, i32 13, i32 21, i32 29, i32 37, i32 45, i32 53, i32 61, i32 69, i32 77, i32 85, i32 93, i32 101, i32 109, i32 117, i32 125, i32 133, i32 141, i32 149, i32 157, i32 165, i32 173, i32 181, i32 189, i32 197, i32 205, i32 213, i32 221, i32 229, i32 237, i32 245, i32 253>
  %strided.vec6 = shufflevector <256 x i8> %wide.vec, <256 x i8> poison, <32 x i32> <i32 6, i32 14, i32 22, i32 30, i32 38, i32 46, i32 54, i32 62, i32 70, i32 78, i32 86, i32 94, i32 102, i32 110, i32 118, i32 126, i32 134, i32 142, i32 150, i32 158, i32 166, i32 174, i32 182, i32 190, i32 198, i32 206, i32 214, i32 222, i32 230, i32 238, i32 246, i32 254>
  %strided.vec7 = shufflevector <256 x i8> %wide.vec, <256 x i8> poison, <32 x i32> <i32 7, i32 15, i32 23, i32 31, i32 39, i32 47, i32 55, i32 63, i32 71, i32 79, i32 87, i32 95, i32 103, i32 111, i32 119, i32 127, i32 135, i32 143, i32 151, i32 159, i32 167, i32 175, i32 183, i32 191, i32 199, i32 207, i32 215, i32 223, i32 231, i32 239, i32 247, i32 255>
  store <32 x i8> %strided.vec0, ptr %out.vec0, align 64
  store <32 x i8> %strided.vec1, ptr %out.vec1, align 64
  store <32 x i8> %strided.vec2, ptr %out.vec2, align 64
  store <32 x i8> %strided.vec3, ptr %out.vec3, align 64
  store <32 x i8> %strided.vec4, ptr %out.vec4, align 64
  store <32 x i8> %strided.vec5, ptr %out.vec5, align 64
  store <32 x i8> %strided.vec6, ptr %out.vec6, align 64
  store <32 x i8> %strided.vec7, ptr %out.vec7, align 64
  ret void
}

define void @load_i8_stride8_vf64(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr %out.vec2, ptr %out.vec3, ptr %out.vec4, ptr %out.vec5, ptr %out.vec6, ptr %out.vec7) nounwind {
; SSE-LABEL: load_i8_stride8_vf64:
; SSE:       # %bb.0:
; SSE-NEXT:    subq $2040, %rsp # imm = 0x7F8
; SSE-NEXT:    movdqa 64(%rdi), %xmm6
; SSE-NEXT:    movdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa 80(%rdi), %xmm8
; SSE-NEXT:    movdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa 96(%rdi), %xmm11
; SSE-NEXT:    movdqa %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa 128(%rdi), %xmm5
; SSE-NEXT:    movdqa 144(%rdi), %xmm10
; SSE-NEXT:    movdqa 160(%rdi), %xmm7
; SSE-NEXT:    movdqa 176(%rdi), %xmm13
; SSE-NEXT:    movdqa 192(%rdi), %xmm3
; SSE-NEXT:    movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa 208(%rdi), %xmm2
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa 224(%rdi), %xmm9
; SSE-NEXT:    movdqa 240(%rdi), %xmm12
; SSE-NEXT:    movdqa {{.*#+}} xmm4 = [255,0,0,0,255,0,0,0]
; SSE-NEXT:    movdqa %xmm12, %xmm0
; SSE-NEXT:    pand %xmm4, %xmm0
; SSE-NEXT:    movdqa %xmm9, %xmm1
; SSE-NEXT:    pand %xmm4, %xmm1
; SSE-NEXT:    packuswb %xmm0, %xmm1
; SSE-NEXT:    packuswb %xmm1, %xmm0
; SSE-NEXT:    movdqa %xmm2, %xmm1
; SSE-NEXT:    pand %xmm4, %xmm1
; SSE-NEXT:    movdqa %xmm3, %xmm2
; SSE-NEXT:    pand %xmm4, %xmm2
; SSE-NEXT:    packuswb %xmm1, %xmm2
; SSE-NEXT:    packuswb %xmm2, %xmm2
; SSE-NEXT:    packuswb %xmm0, %xmm2
; SSE-NEXT:    movdqa %xmm13, %xmm0
; SSE-NEXT:    pand %xmm4, %xmm0
; SSE-NEXT:    movdqa %xmm7, %xmm1
; SSE-NEXT:    pand %xmm4, %xmm1
; SSE-NEXT:    packuswb %xmm0, %xmm1
; SSE-NEXT:    movdqa %xmm10, %xmm0
; SSE-NEXT:    pand %xmm4, %xmm0
; SSE-NEXT:    movdqa %xmm5, %xmm3
; SSE-NEXT:    pand %xmm4, %xmm3
; SSE-NEXT:    packuswb %xmm0, %xmm3
; SSE-NEXT:    movdqa 112(%rdi), %xmm14
; SSE-NEXT:    movdqa %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    packuswb %xmm1, %xmm0
; SSE-NEXT:    packuswb %xmm3, %xmm3
; SSE-NEXT:    packuswb %xmm0, %xmm3
; SSE-NEXT:    shufps {{.*#+}} xmm3 = xmm3[0,3],xmm2[0,3]
; SSE-NEXT:    movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm14, %xmm0
; SSE-NEXT:    pand %xmm4, %xmm0
; SSE-NEXT:    movdqa %xmm11, %xmm1
; SSE-NEXT:    pand %xmm4, %xmm1
; SSE-NEXT:    packuswb %xmm0, %xmm1
; SSE-NEXT:    movdqa %xmm8, %xmm0
; SSE-NEXT:    pand %xmm4, %xmm0
; SSE-NEXT:    movdqa %xmm6, %xmm2
; SSE-NEXT:    pand %xmm4, %xmm2
; SSE-NEXT:    packuswb %xmm0, %xmm2
; SSE-NEXT:    packuswb %xmm1, %xmm0
; SSE-NEXT:    packuswb %xmm2, %xmm2
; SSE-NEXT:    packuswb %xmm0, %xmm2
; SSE-NEXT:    movdqa 48(%rdi), %xmm14
; SSE-NEXT:    movdqa %xmm14, %xmm0
; SSE-NEXT:    pand %xmm4, %xmm0
; SSE-NEXT:    movdqa 32(%rdi), %xmm15
; SSE-NEXT:    movdqa %xmm15, %xmm1
; SSE-NEXT:    pand %xmm4, %xmm1
; SSE-NEXT:    packuswb %xmm0, %xmm1
; SSE-NEXT:    movdqa 16(%rdi), %xmm8
; SSE-NEXT:    movdqa %xmm8, %xmm0
; SSE-NEXT:    pand %xmm4, %xmm0
; SSE-NEXT:    movdqa (%rdi), %xmm11
; SSE-NEXT:    movdqa %xmm11, %xmm3
; SSE-NEXT:    pand %xmm4, %xmm3
; SSE-NEXT:    packuswb %xmm0, %xmm3
; SSE-NEXT:    packuswb %xmm1, %xmm0
; SSE-NEXT:    packuswb %xmm3, %xmm3
; SSE-NEXT:    packuswb %xmm0, %xmm3
; SSE-NEXT:    shufps {{.*#+}} xmm3 = xmm3[0,3],xmm2[0,3]
; SSE-NEXT:    movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa 496(%rdi), %xmm0
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pand %xmm4, %xmm0
; SSE-NEXT:    movdqa 480(%rdi), %xmm1
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pand %xmm4, %xmm1
; SSE-NEXT:    packuswb %xmm0, %xmm1
; SSE-NEXT:    movdqa 464(%rdi), %xmm0
; SSE-NEXT:    movdqa %xmm0, (%rsp) # 16-byte Spill
; SSE-NEXT:    pand %xmm4, %xmm0
; SSE-NEXT:    movdqa 448(%rdi), %xmm2
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pand %xmm4, %xmm2
; SSE-NEXT:    packuswb %xmm0, %xmm2
; SSE-NEXT:    packuswb %xmm1, %xmm0
; SSE-NEXT:    packuswb %xmm2, %xmm2
; SSE-NEXT:    packuswb %xmm0, %xmm2
; SSE-NEXT:    movdqa 432(%rdi), %xmm0
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pand %xmm4, %xmm0
; SSE-NEXT:    movdqa 416(%rdi), %xmm1
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pand %xmm4, %xmm1
; SSE-NEXT:    packuswb %xmm0, %xmm1
; SSE-NEXT:    movdqa 400(%rdi), %xmm0
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pand %xmm4, %xmm0
; SSE-NEXT:    movdqa 384(%rdi), %xmm3
; SSE-NEXT:    movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pand %xmm4, %xmm3
; SSE-NEXT:    packuswb %xmm0, %xmm3
; SSE-NEXT:    packuswb %xmm1, %xmm0
; SSE-NEXT:    packuswb %xmm3, %xmm3
; SSE-NEXT:    packuswb %xmm0, %xmm3
; SSE-NEXT:    shufps {{.*#+}} xmm3 = xmm3[0,3],xmm2[0,3]
; SSE-NEXT:    movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa 368(%rdi), %xmm0
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pand %xmm4, %xmm0
; SSE-NEXT:    movdqa 352(%rdi), %xmm1
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pand %xmm4, %xmm1
; SSE-NEXT:    packuswb %xmm0, %xmm1
; SSE-NEXT:    movdqa 336(%rdi), %xmm0
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pand %xmm4, %xmm0
; SSE-NEXT:    movdqa 320(%rdi), %xmm2
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pand %xmm4, %xmm2
; SSE-NEXT:    packuswb %xmm0, %xmm2
; SSE-NEXT:    packuswb %xmm1, %xmm0
; SSE-NEXT:    packuswb %xmm2, %xmm2
; SSE-NEXT:    packuswb %xmm0, %xmm2
; SSE-NEXT:    movdqa 304(%rdi), %xmm0
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pand %xmm4, %xmm0
; SSE-NEXT:    movdqa 288(%rdi), %xmm1
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pand %xmm4, %xmm1
; SSE-NEXT:    packuswb %xmm0, %xmm1
; SSE-NEXT:    movdqa 272(%rdi), %xmm0
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pand %xmm4, %xmm0
; SSE-NEXT:    movdqa 256(%rdi), %xmm3
; SSE-NEXT:    movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pand %xmm3, %xmm4
; SSE-NEXT:    packuswb %xmm0, %xmm4
; SSE-NEXT:    packuswb %xmm1, %xmm0
; SSE-NEXT:    packuswb %xmm4, %xmm4
; SSE-NEXT:    packuswb %xmm0, %xmm4
; SSE-NEXT:    shufps {{.*#+}} xmm4 = xmm4[0,3],xmm2[0,3]
; SSE-NEXT:    movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pxor %xmm6, %xmm6
; SSE-NEXT:    movdqa %xmm10, %xmm0
; SSE-NEXT:    movdqa %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm10, %xmm1
; SSE-NEXT:    punpckhbw {{.*#+}} xmm1 = xmm1[8],xmm6[8],xmm1[9],xmm6[9],xmm1[10],xmm6[10],xmm1[11],xmm6[11],xmm1[12],xmm6[12],xmm1[13],xmm6[13],xmm1[14],xmm6[14],xmm1[15],xmm6[15]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm6[0],xmm0[1],xmm6[1],xmm0[2],xmm6[2],xmm0[3],xmm6[3],xmm0[4],xmm6[4],xmm0[5],xmm6[5],xmm0[6],xmm6[6],xmm0[7],xmm6[7]
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa {{.*#+}} xmm10 = [65535,0,65535,65535,65535,65535,65535,65535]
; SSE-NEXT:    movdqa %xmm10, %xmm1
; SSE-NEXT:    pandn %xmm0, %xmm1
; SSE-NEXT:    movdqa %xmm5, %xmm0
; SSE-NEXT:    movdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm5, %xmm2
; SSE-NEXT:    punpckhbw {{.*#+}} xmm2 = xmm2[8],xmm6[8],xmm2[9],xmm6[9],xmm2[10],xmm6[10],xmm2[11],xmm6[11],xmm2[12],xmm6[12],xmm2[13],xmm6[13],xmm2[14],xmm6[14],xmm2[15],xmm6[15]
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm6[0],xmm0[1],xmm6[1],xmm0[2],xmm6[2],xmm0[3],xmm6[3],xmm0[4],xmm6[4],xmm0[5],xmm6[5],xmm0[6],xmm6[6],xmm0[7],xmm6[7]
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],xmm2[3]
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm0[1,1,1,1]
; SSE-NEXT:    packuswb %xmm5, %xmm5
; SSE-NEXT:    pand %xmm10, %xmm5
; SSE-NEXT:    por %xmm1, %xmm5
; SSE-NEXT:    movdqa %xmm13, %xmm2
; SSE-NEXT:    movdqa %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm13, %xmm1
; SSE-NEXT:    punpckhbw {{.*#+}} xmm1 = xmm1[8],xmm6[8],xmm1[9],xmm6[9],xmm1[10],xmm6[10],xmm1[11],xmm6[11],xmm1[12],xmm6[12],xmm1[13],xmm6[13],xmm1[14],xmm6[14],xmm1[15],xmm6[15]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm6[0],xmm2[1],xmm6[1],xmm2[2],xmm6[2],xmm2[3],xmm6[3],xmm2[4],xmm6[4],xmm2[5],xmm6[5],xmm2[6],xmm6[6],xmm2[7],xmm6[7]
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; SSE-NEXT:    packuswb %xmm2, %xmm2
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[0,0,2,3]
; SSE-NEXT:    movdqa {{.*#+}} xmm13 = [65535,65535,65535,0,65535,65535,65535,65535]
; SSE-NEXT:    movdqa %xmm13, %xmm2
; SSE-NEXT:    pandn %xmm1, %xmm2
; SSE-NEXT:    movdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm7, %xmm1
; SSE-NEXT:    punpckhbw {{.*#+}} xmm1 = xmm1[8],xmm6[8],xmm1[9],xmm6[9],xmm1[10],xmm6[10],xmm1[11],xmm6[11],xmm1[12],xmm6[12],xmm1[13],xmm6[13],xmm1[14],xmm6[14],xmm1[15],xmm6[15]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklbw {{.*#+}} xmm7 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3],xmm7[4],xmm6[4],xmm7[5],xmm6[5],xmm7[6],xmm6[6],xmm7[7],xmm6[7]
; SSE-NEXT:    movdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklwd {{.*#+}} xmm7 = xmm7[0],xmm1[0],xmm7[1],xmm1[1],xmm7[2],xmm1[2],xmm7[3],xmm1[3]
; SSE-NEXT:    movdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm7[0,1,1,3]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm13, %xmm1
; SSE-NEXT:    por %xmm2, %xmm1
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,1,1]
; SSE-NEXT:    punpckldq {{.*#+}} xmm5 = xmm5[0],xmm1[0],xmm5[1],xmm1[1]
; SSE-NEXT:    movdqa %xmm12, %xmm1
; SSE-NEXT:    movdqa %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm12, %xmm2
; SSE-NEXT:    punpckhbw {{.*#+}} xmm2 = xmm2[8],xmm6[8],xmm2[9],xmm6[9],xmm2[10],xmm6[10],xmm2[11],xmm6[11],xmm2[12],xmm6[12],xmm2[13],xmm6[13],xmm2[14],xmm6[14],xmm2[15],xmm6[15]
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm6[0],xmm1[1],xmm6[1],xmm1[2],xmm6[2],xmm1[3],xmm6[3],xmm1[4],xmm6[4],xmm1[5],xmm6[5],xmm1[6],xmm6[6],xmm1[7],xmm6[7]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[0,1,2,2]
; SSE-NEXT:    movdqa {{.*#+}} xmm12 = [65535,65535,65535,65535,65535,65535,65535,0]
; SSE-NEXT:    movdqa %xmm12, %xmm2
; SSE-NEXT:    pandn %xmm1, %xmm2
; SSE-NEXT:    movdqa %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm9, %xmm1
; SSE-NEXT:    punpckhbw {{.*#+}} xmm1 = xmm1[8],xmm6[8],xmm1[9],xmm6[9],xmm1[10],xmm6[10],xmm1[11],xmm6[11],xmm1[12],xmm6[12],xmm1[13],xmm6[13],xmm1[14],xmm6[14],xmm1[15],xmm6[15]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklbw {{.*#+}} xmm9 = xmm9[0],xmm6[0],xmm9[1],xmm6[1],xmm9[2],xmm6[2],xmm9[3],xmm6[3],xmm9[4],xmm6[4],xmm9[5],xmm6[5],xmm9[6],xmm6[6],xmm9[7],xmm6[7]
; SSE-NEXT:    movdqa %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklwd {{.*#+}} xmm9 = xmm9[0],xmm1[0],xmm9[1],xmm1[1],xmm9[2],xmm1[2],xmm9[3],xmm1[3]
; SSE-NEXT:    movdqa %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm9[0,1,1,3]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm12, %xmm1
; SSE-NEXT:    por %xmm2, %xmm1
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; SSE-NEXT:    movdqa %xmm2, %xmm3
; SSE-NEXT:    punpckhbw {{.*#+}} xmm3 = xmm3[8],xmm6[8],xmm3[9],xmm6[9],xmm3[10],xmm6[10],xmm3[11],xmm6[11],xmm3[12],xmm6[12],xmm3[13],xmm6[13],xmm3[14],xmm6[14],xmm3[15],xmm6[15]
; SSE-NEXT:    movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm6[0],xmm2[1],xmm6[1],xmm2[2],xmm6[2],xmm2[3],xmm6[3],xmm2[4],xmm6[4],xmm2[5],xmm6[5],xmm2[6],xmm6[6],xmm2[7],xmm6[7]
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
; SSE-NEXT:    packuswb %xmm2, %xmm3
; SSE-NEXT:    movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa {{.*#+}} xmm9 = [65535,65535,65535,65535,65535,0,65535,65535]
; SSE-NEXT:    movdqa %xmm9, %xmm2
; SSE-NEXT:    pandn %xmm3, %xmm2
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; SSE-NEXT:    movdqa %xmm3, %xmm4
; SSE-NEXT:    punpckhbw {{.*#+}} xmm4 = xmm4[8],xmm6[8],xmm4[9],xmm6[9],xmm4[10],xmm6[10],xmm4[11],xmm6[11],xmm4[12],xmm6[12],xmm4[13],xmm6[13],xmm4[14],xmm6[14],xmm4[15],xmm6[15]
; SSE-NEXT:    movdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm6[0],xmm3[1],xmm6[1],xmm3[2],xmm6[2],xmm3[3],xmm6[3],xmm3[4],xmm6[4],xmm3[5],xmm6[5],xmm3[6],xmm6[6],xmm3[7],xmm6[7]
; SSE-NEXT:    movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklwd {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
; SSE-NEXT:    movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,1,1]
; SSE-NEXT:    packuswb %xmm3, %xmm3
; SSE-NEXT:    pand %xmm9, %xmm3
; SSE-NEXT:    por %xmm2, %xmm3
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[2,2,2,2]
; SSE-NEXT:    punpckhdq {{.*#+}} xmm2 = xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; SSE-NEXT:    movsd {{.*#+}} xmm2 = xmm5[0],xmm2[1]
; SSE-NEXT:    movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm8, %xmm1
; SSE-NEXT:    punpckhbw {{.*#+}} xmm1 = xmm1[8],xmm6[8],xmm1[9],xmm6[9],xmm1[10],xmm6[10],xmm1[11],xmm6[11],xmm1[12],xmm6[12],xmm1[13],xmm6[13],xmm1[14],xmm6[14],xmm1[15],xmm6[15]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklbw {{.*#+}} xmm8 = xmm8[0],xmm6[0],xmm8[1],xmm6[1],xmm8[2],xmm6[2],xmm8[3],xmm6[3],xmm8[4],xmm6[4],xmm8[5],xmm6[5],xmm8[6],xmm6[6],xmm8[7],xmm6[7]
; SSE-NEXT:    movdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklwd {{.*#+}} xmm8 = xmm8[0],xmm1[0],xmm8[1],xmm1[1],xmm8[2],xmm1[2],xmm8[3],xmm1[3]
; SSE-NEXT:    packuswb %xmm8, %xmm8
; SSE-NEXT:    movdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm10, %xmm1
; SSE-NEXT:    pandn %xmm8, %xmm1
; SSE-NEXT:    movdqa %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm11, %xmm2
; SSE-NEXT:    punpckhbw {{.*#+}} xmm2 = xmm2[8],xmm6[8],xmm2[9],xmm6[9],xmm2[10],xmm6[10],xmm2[11],xmm6[11],xmm2[12],xmm6[12],xmm2[13],xmm6[13],xmm2[14],xmm6[14],xmm2[15],xmm6[15]
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklbw {{.*#+}} xmm11 = xmm11[0],xmm6[0],xmm11[1],xmm6[1],xmm11[2],xmm6[2],xmm11[3],xmm6[3],xmm11[4],xmm6[4],xmm11[5],xmm6[5],xmm11[6],xmm6[6],xmm11[7],xmm6[7]
; SSE-NEXT:    movdqa %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklwd {{.*#+}} xmm11 = xmm11[0],xmm2[0],xmm11[1],xmm2[1],xmm11[2],xmm2[2],xmm11[3],xmm2[3]
; SSE-NEXT:    movdqa %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm11[1,1,1,1]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    pand %xmm10, %xmm0
; SSE-NEXT:    por %xmm1, %xmm0
; SSE-NEXT:    movdqa %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm14, %xmm1
; SSE-NEXT:    punpckhbw {{.*#+}} xmm1 = xmm1[8],xmm6[8],xmm1[9],xmm6[9],xmm1[10],xmm6[10],xmm1[11],xmm6[11],xmm1[12],xmm6[12],xmm1[13],xmm6[13],xmm1[14],xmm6[14],xmm1[15],xmm6[15]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklbw {{.*#+}} xmm14 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3],xmm14[4],xmm6[4],xmm14[5],xmm6[5],xmm14[6],xmm6[6],xmm14[7],xmm6[7]
; SSE-NEXT:    movdqa %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklwd {{.*#+}} xmm14 = xmm14[0],xmm1[0],xmm14[1],xmm1[1],xmm14[2],xmm1[2],xmm14[3],xmm1[3]
; SSE-NEXT:    packuswb %xmm14, %xmm14
; SSE-NEXT:    movdqa %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm14[0,0,2,3]
; SSE-NEXT:    movdqa %xmm13, %xmm2
; SSE-NEXT:    pandn %xmm1, %xmm2
; SSE-NEXT:    movdqa %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm15, %xmm1
; SSE-NEXT:    punpckhbw {{.*#+}} xmm1 = xmm1[8],xmm6[8],xmm1[9],xmm6[9],xmm1[10],xmm6[10],xmm1[11],xmm6[11],xmm1[12],xmm6[12],xmm1[13],xmm6[13],xmm1[14],xmm6[14],xmm1[15],xmm6[15]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklbw {{.*#+}} xmm15 = xmm15[0],xmm6[0],xmm15[1],xmm6[1],xmm15[2],xmm6[2],xmm15[3],xmm6[3],xmm15[4],xmm6[4],xmm15[5],xmm6[5],xmm15[6],xmm6[6],xmm15[7],xmm6[7]
; SSE-NEXT:    movdqa %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklwd {{.*#+}} xmm15 = xmm15[0],xmm1[0],xmm15[1],xmm1[1],xmm15[2],xmm1[2],xmm15[3],xmm1[3]
; SSE-NEXT:    movdqa %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm15[0,1,1,3]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm13, %xmm1
; SSE-NEXT:    por %xmm2, %xmm1
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,1,1]
; SSE-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movdqa %xmm1, %xmm2
; SSE-NEXT:    punpckhbw {{.*#+}} xmm2 = xmm2[8],xmm6[8],xmm2[9],xmm6[9],xmm2[10],xmm6[10],xmm2[11],xmm6[11],xmm2[12],xmm6[12],xmm2[13],xmm6[13],xmm2[14],xmm6[14],xmm2[15],xmm6[15]
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm6[0],xmm1[1],xmm6[1],xmm1[2],xmm6[2],xmm1[3],xmm6[3],xmm1[4],xmm6[4],xmm1[5],xmm6[5],xmm1[6],xmm6[6],xmm1[7],xmm6[7]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[0,1,2,2]
; SSE-NEXT:    movdqa %xmm12, %xmm2
; SSE-NEXT:    pandn %xmm1, %xmm2
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; SSE-NEXT:    movdqa %xmm3, %xmm1
; SSE-NEXT:    punpckhbw {{.*#+}} xmm1 = xmm1[8],xmm6[8],xmm1[9],xmm6[9],xmm1[10],xmm6[10],xmm1[11],xmm6[11],xmm1[12],xmm6[12],xmm1[13],xmm6[13],xmm1[14],xmm6[14],xmm1[15],xmm6[15]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm6[0],xmm3[1],xmm6[1],xmm3[2],xmm6[2],xmm3[3],xmm6[3],xmm3[4],xmm6[4],xmm3[5],xmm6[5],xmm3[6],xmm6[6],xmm3[7],xmm6[7]
; SSE-NEXT:    movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklwd {{.*#+}} xmm3 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3]
; SSE-NEXT:    movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[0,1,1,3]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm12, %xmm1
; SSE-NEXT:    por %xmm2, %xmm1
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; SSE-NEXT:    movdqa %xmm2, %xmm3
; SSE-NEXT:    punpckhbw {{.*#+}} xmm3 = xmm3[8],xmm6[8],xmm3[9],xmm6[9],xmm3[10],xmm6[10],xmm3[11],xmm6[11],xmm3[12],xmm6[12],xmm3[13],xmm6[13],xmm3[14],xmm6[14],xmm3[15],xmm6[15]
; SSE-NEXT:    movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm6[0],xmm2[1],xmm6[1],xmm2[2],xmm6[2],xmm2[3],xmm6[3],xmm2[4],xmm6[4],xmm2[5],xmm6[5],xmm2[6],xmm6[6],xmm2[7],xmm6[7]
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
; SSE-NEXT:    packuswb %xmm2, %xmm3
; SSE-NEXT:    movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm9, %xmm2
; SSE-NEXT:    pandn %xmm3, %xmm2
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; SSE-NEXT:    movdqa %xmm3, %xmm4
; SSE-NEXT:    punpckhbw {{.*#+}} xmm4 = xmm4[8],xmm6[8],xmm4[9],xmm6[9],xmm4[10],xmm6[10],xmm4[11],xmm6[11],xmm4[12],xmm6[12],xmm4[13],xmm6[13],xmm4[14],xmm6[14],xmm4[15],xmm6[15]
; SSE-NEXT:    movdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm6[0],xmm3[1],xmm6[1],xmm3[2],xmm6[2],xmm3[3],xmm6[3],xmm3[4],xmm6[4],xmm3[5],xmm6[5],xmm3[6],xmm6[6],xmm3[7],xmm6[7]
; SSE-NEXT:    movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklwd {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
; SSE-NEXT:    movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,1,1]
; SSE-NEXT:    packuswb %xmm3, %xmm3
; SSE-NEXT:    pand %xmm9, %xmm3
; SSE-NEXT:    por %xmm2, %xmm3
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[2,2,2,2]
; SSE-NEXT:    punpckhdq {{.*#+}} xmm2 = xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; SSE-NEXT:    movsd {{.*#+}} xmm2 = xmm0[0],xmm2[1]
; SSE-NEXT:    movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
; SSE-NEXT:    movdqa %xmm14, %xmm0
; SSE-NEXT:    punpckhbw {{.*#+}} xmm0 = xmm0[8],xmm6[8],xmm0[9],xmm6[9],xmm0[10],xmm6[10],xmm0[11],xmm6[11],xmm0[12],xmm6[12],xmm0[13],xmm6[13],xmm0[14],xmm6[14],xmm0[15],xmm6[15]
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklbw {{.*#+}} xmm14 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3],xmm14[4],xmm6[4],xmm14[5],xmm6[5],xmm14[6],xmm6[6],xmm14[7],xmm6[7]
; SSE-NEXT:    movdqa %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklwd {{.*#+}} xmm14 = xmm14[0],xmm0[0],xmm14[1],xmm0[1],xmm14[2],xmm0[2],xmm14[3],xmm0[3]
; SSE-NEXT:    packuswb %xmm14, %xmm14
; SSE-NEXT:    movdqa %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm10, %xmm1
; SSE-NEXT:    pandn %xmm14, %xmm1
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movdqa %xmm0, %xmm2
; SSE-NEXT:    punpckhbw {{.*#+}} xmm2 = xmm2[8],xmm6[8],xmm2[9],xmm6[9],xmm2[10],xmm6[10],xmm2[11],xmm6[11],xmm2[12],xmm6[12],xmm2[13],xmm6[13],xmm2[14],xmm6[14],xmm2[15],xmm6[15]
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm6[0],xmm0[1],xmm6[1],xmm0[2],xmm6[2],xmm0[3],xmm6[3],xmm0[4],xmm6[4],xmm0[5],xmm6[5],xmm0[6],xmm6[6],xmm0[7],xmm6[7]
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],xmm2[3]
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,1,1]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    pand %xmm10, %xmm0
; SSE-NEXT:    por %xmm1, %xmm0
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
; SSE-NEXT:    movdqa %xmm14, %xmm1
; SSE-NEXT:    punpckhbw {{.*#+}} xmm1 = xmm1[8],xmm6[8],xmm1[9],xmm6[9],xmm1[10],xmm6[10],xmm1[11],xmm6[11],xmm1[12],xmm6[12],xmm1[13],xmm6[13],xmm1[14],xmm6[14],xmm1[15],xmm6[15]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklbw {{.*#+}} xmm14 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3],xmm14[4],xmm6[4],xmm14[5],xmm6[5],xmm14[6],xmm6[6],xmm14[7],xmm6[7]
; SSE-NEXT:    movdqa %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklwd {{.*#+}} xmm14 = xmm14[0],xmm1[0],xmm14[1],xmm1[1],xmm14[2],xmm1[2],xmm14[3],xmm1[3]
; SSE-NEXT:    packuswb %xmm14, %xmm14
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm14[0,0,2,3]
; SSE-NEXT:    movdqa %xmm13, %xmm2
; SSE-NEXT:    pandn %xmm1, %xmm2
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; SSE-NEXT:    movdqa %xmm3, %xmm1
; SSE-NEXT:    punpckhbw {{.*#+}} xmm1 = xmm1[8],xmm6[8],xmm1[9],xmm6[9],xmm1[10],xmm6[10],xmm1[11],xmm6[11],xmm1[12],xmm6[12],xmm1[13],xmm6[13],xmm1[14],xmm6[14],xmm1[15],xmm6[15]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm6[0],xmm3[1],xmm6[1],xmm3[2],xmm6[2],xmm3[3],xmm6[3],xmm3[4],xmm6[4],xmm3[5],xmm6[5],xmm3[6],xmm6[6],xmm3[7],xmm6[7]
; SSE-NEXT:    movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklwd {{.*#+}} xmm3 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3]
; SSE-NEXT:    movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[0,1,1,3]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm13, %xmm1
; SSE-NEXT:    por %xmm2, %xmm1
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,1,1]
; SSE-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movdqa %xmm1, %xmm2
; SSE-NEXT:    punpckhbw {{.*#+}} xmm2 = xmm2[8],xmm6[8],xmm2[9],xmm6[9],xmm2[10],xmm6[10],xmm2[11],xmm6[11],xmm2[12],xmm6[12],xmm2[13],xmm6[13],xmm2[14],xmm6[14],xmm2[15],xmm6[15]
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm6[0],xmm1[1],xmm6[1],xmm1[2],xmm6[2],xmm1[3],xmm6[3],xmm1[4],xmm6[4],xmm1[5],xmm6[5],xmm1[6],xmm6[6],xmm1[7],xmm6[7]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[0,1,2,2]
; SSE-NEXT:    movdqa %xmm12, %xmm2
; SSE-NEXT:    pandn %xmm1, %xmm2
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; SSE-NEXT:    movdqa %xmm3, %xmm1
; SSE-NEXT:    punpckhbw {{.*#+}} xmm1 = xmm1[8],xmm6[8],xmm1[9],xmm6[9],xmm1[10],xmm6[10],xmm1[11],xmm6[11],xmm1[12],xmm6[12],xmm1[13],xmm6[13],xmm1[14],xmm6[14],xmm1[15],xmm6[15]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm6[0],xmm3[1],xmm6[1],xmm3[2],xmm6[2],xmm3[3],xmm6[3],xmm3[4],xmm6[4],xmm3[5],xmm6[5],xmm3[6],xmm6[6],xmm3[7],xmm6[7]
; SSE-NEXT:    movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklwd {{.*#+}} xmm3 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3]
; SSE-NEXT:    movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[0,1,1,3]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm12, %xmm1
; SSE-NEXT:    por %xmm2, %xmm1
; SSE-NEXT:    movdqa (%rsp), %xmm2 # 16-byte Reload
; SSE-NEXT:    movdqa %xmm2, %xmm3
; SSE-NEXT:    punpckhbw {{.*#+}} xmm3 = xmm3[8],xmm6[8],xmm3[9],xmm6[9],xmm3[10],xmm6[10],xmm3[11],xmm6[11],xmm3[12],xmm6[12],xmm3[13],xmm6[13],xmm3[14],xmm6[14],xmm3[15],xmm6[15]
; SSE-NEXT:    movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm6[0],xmm2[1],xmm6[1],xmm2[2],xmm6[2],xmm2[3],xmm6[3],xmm2[4],xmm6[4],xmm2[5],xmm6[5],xmm2[6],xmm6[6],xmm2[7],xmm6[7]
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
; SSE-NEXT:    packuswb %xmm2, %xmm3
; SSE-NEXT:    movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm9, %xmm2
; SSE-NEXT:    pandn %xmm3, %xmm2
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; SSE-NEXT:    movdqa %xmm3, %xmm4
; SSE-NEXT:    punpckhbw {{.*#+}} xmm4 = xmm4[8],xmm6[8],xmm4[9],xmm6[9],xmm4[10],xmm6[10],xmm4[11],xmm6[11],xmm4[12],xmm6[12],xmm4[13],xmm6[13],xmm4[14],xmm6[14],xmm4[15],xmm6[15]
; SSE-NEXT:    movdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm6[0],xmm3[1],xmm6[1],xmm3[2],xmm6[2],xmm3[3],xmm6[3],xmm3[4],xmm6[4],xmm3[5],xmm6[5],xmm3[6],xmm6[6],xmm3[7],xmm6[7]
; SSE-NEXT:    movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklwd {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
; SSE-NEXT:    movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,1,1]
; SSE-NEXT:    packuswb %xmm3, %xmm3
; SSE-NEXT:    pand %xmm9, %xmm3
; SSE-NEXT:    por %xmm2, %xmm3
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[2,2,2,2]
; SSE-NEXT:    punpckhdq {{.*#+}} xmm2 = xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; SSE-NEXT:    movsd {{.*#+}} xmm2 = xmm0[0],xmm2[1]
; SSE-NEXT:    movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; SSE-NEXT:    movdqa %xmm8, %xmm0
; SSE-NEXT:    punpckhbw {{.*#+}} xmm0 = xmm0[8],xmm6[8],xmm0[9],xmm6[9],xmm0[10],xmm6[10],xmm0[11],xmm6[11],xmm0[12],xmm6[12],xmm0[13],xmm6[13],xmm0[14],xmm6[14],xmm0[15],xmm6[15]
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklbw {{.*#+}} xmm8 = xmm8[0],xmm6[0],xmm8[1],xmm6[1],xmm8[2],xmm6[2],xmm8[3],xmm6[3],xmm8[4],xmm6[4],xmm8[5],xmm6[5],xmm8[6],xmm6[6],xmm8[7],xmm6[7]
; SSE-NEXT:    movdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklwd {{.*#+}} xmm8 = xmm8[0],xmm0[0],xmm8[1],xmm0[1],xmm8[2],xmm0[2],xmm8[3],xmm0[3]
; SSE-NEXT:    packuswb %xmm8, %xmm8
; SSE-NEXT:    movdqa %xmm10, %xmm1
; SSE-NEXT:    pandn %xmm8, %xmm1
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    movdqa %xmm0, %xmm2
; SSE-NEXT:    punpckhbw {{.*#+}} xmm2 = xmm2[8],xmm6[8],xmm2[9],xmm6[9],xmm2[10],xmm6[10],xmm2[11],xmm6[11],xmm2[12],xmm6[12],xmm2[13],xmm6[13],xmm2[14],xmm6[14],xmm2[15],xmm6[15]
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm6[0],xmm0[1],xmm6[1],xmm0[2],xmm6[2],xmm0[3],xmm6[3],xmm0[4],xmm6[4],xmm0[5],xmm6[5],xmm0[6],xmm6[6],xmm0[7],xmm6[7]
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],xmm2[3]
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,1,1]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    pand %xmm10, %xmm0
; SSE-NEXT:    por %xmm1, %xmm0
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; SSE-NEXT:    movdqa %xmm4, %xmm1
; SSE-NEXT:    punpckhbw {{.*#+}} xmm1 = xmm1[8],xmm6[8],xmm1[9],xmm6[9],xmm1[10],xmm6[10],xmm1[11],xmm6[11],xmm1[12],xmm6[12],xmm1[13],xmm6[13],xmm1[14],xmm6[14],xmm1[15],xmm6[15]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklbw {{.*#+}} xmm4 = xmm4[0],xmm6[0],xmm4[1],xmm6[1],xmm4[2],xmm6[2],xmm4[3],xmm6[3],xmm4[4],xmm6[4],xmm4[5],xmm6[5],xmm4[6],xmm6[6],xmm4[7],xmm6[7]
; SSE-NEXT:    movdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklwd {{.*#+}} xmm4 = xmm4[0],xmm1[0],xmm4[1],xmm1[1],xmm4[2],xmm1[2],xmm4[3],xmm1[3]
; SSE-NEXT:    packuswb %xmm4, %xmm4
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm4[0,0,2,3]
; SSE-NEXT:    movdqa %xmm13, %xmm5
; SSE-NEXT:    pandn %xmm1, %xmm5
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; SSE-NEXT:    movdqa %xmm2, %xmm1
; SSE-NEXT:    punpckhbw {{.*#+}} xmm1 = xmm1[8],xmm6[8],xmm1[9],xmm6[9],xmm1[10],xmm6[10],xmm1[11],xmm6[11],xmm1[12],xmm6[12],xmm1[13],xmm6[13],xmm1[14],xmm6[14],xmm1[15],xmm6[15]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm6[0],xmm2[1],xmm6[1],xmm2[2],xmm6[2],xmm2[3],xmm6[3],xmm2[4],xmm6[4],xmm2[5],xmm6[5],xmm2[6],xmm6[6],xmm2[7],xmm6[7]
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[0,1,1,3]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm13, %xmm1
; SSE-NEXT:    por %xmm5, %xmm1
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,1,1]
; SSE-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    movdqa %xmm1, %xmm2
; SSE-NEXT:    punpckhbw {{.*#+}} xmm2 = xmm2[8],xmm6[8],xmm2[9],xmm6[9],xmm2[10],xmm6[10],xmm2[11],xmm6[11],xmm2[12],xmm6[12],xmm2[13],xmm6[13],xmm2[14],xmm6[14],xmm2[15],xmm6[15]
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm6[0],xmm1[1],xmm6[1],xmm1[2],xmm6[2],xmm1[3],xmm6[3],xmm1[4],xmm6[4],xmm1[5],xmm6[5],xmm1[6],xmm6[6],xmm1[7],xmm6[7]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[0,1,2,2]
; SSE-NEXT:    movdqa %xmm12, %xmm5
; SSE-NEXT:    pandn %xmm1, %xmm5
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
; SSE-NEXT:    movdqa %xmm11, %xmm1
; SSE-NEXT:    punpckhbw {{.*#+}} xmm1 = xmm1[8],xmm6[8],xmm1[9],xmm6[9],xmm1[10],xmm6[10],xmm1[11],xmm6[11],xmm1[12],xmm6[12],xmm1[13],xmm6[13],xmm1[14],xmm6[14],xmm1[15],xmm6[15]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklbw {{.*#+}} xmm11 = xmm11[0],xmm6[0],xmm11[1],xmm6[1],xmm11[2],xmm6[2],xmm11[3],xmm6[3],xmm11[4],xmm6[4],xmm11[5],xmm6[5],xmm11[6],xmm6[6],xmm11[7],xmm6[7]
; SSE-NEXT:    movdqa %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklwd {{.*#+}} xmm11 = xmm11[0],xmm1[0],xmm11[1],xmm1[1],xmm11[2],xmm1[2],xmm11[3],xmm1[3]
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm11[0,1,1,3]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm12, %xmm1
; SSE-NEXT:    por %xmm5, %xmm1
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; SSE-NEXT:    movdqa %xmm5, %xmm2
; SSE-NEXT:    punpckhbw {{.*#+}} xmm2 = xmm2[8],xmm6[8],xmm2[9],xmm6[9],xmm2[10],xmm6[10],xmm2[11],xmm6[11],xmm2[12],xmm6[12],xmm2[13],xmm6[13],xmm2[14],xmm6[14],xmm2[15],xmm6[15]
; SSE-NEXT:    movdqa %xmm2, %xmm15
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklbw {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3],xmm5[4],xmm6[4],xmm5[5],xmm6[5],xmm5[6],xmm6[6],xmm5[7],xmm6[7]
; SSE-NEXT:    movdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; SSE-NEXT:    movdqa %xmm3, %xmm2
; SSE-NEXT:    punpckhbw {{.*#+}} xmm2 = xmm2[8],xmm6[8],xmm2[9],xmm6[9],xmm2[10],xmm6[10],xmm2[11],xmm6[11],xmm2[12],xmm6[12],xmm2[13],xmm6[13],xmm2[14],xmm6[14],xmm2[15],xmm6[15]
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm3, %xmm7
; SSE-NEXT:    punpcklbw {{.*#+}} xmm7 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3],xmm7[4],xmm6[4],xmm7[5],xmm6[5],xmm7[6],xmm6[6],xmm7[7],xmm6[7]
; SSE-NEXT:    movdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    punpcklwd {{.*#+}} xmm5 = xmm5[0],xmm15[0],xmm5[1],xmm15[1],xmm5[2],xmm15[2],xmm5[3],xmm15[3]
; SSE-NEXT:    packuswb %xmm5, %xmm6
; SSE-NEXT:    movdqa %xmm9, %xmm5
; SSE-NEXT:    pandn %xmm6, %xmm5
; SSE-NEXT:    punpcklwd {{.*#+}} xmm7 = xmm7[0],xmm2[0],xmm7[1],xmm2[1],xmm7[2],xmm2[2],xmm7[3],xmm2[3]
; SSE-NEXT:    movdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm15 = xmm7[1,1,1,1]
; SSE-NEXT:    packuswb %xmm15, %xmm15
; SSE-NEXT:    pand %xmm9, %xmm15
; SSE-NEXT:    por %xmm5, %xmm15
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm15[2,2,2,2]
; SSE-NEXT:    punpckhdq {{.*#+}} xmm5 = xmm5[2],xmm1[2],xmm5[3],xmm1[3]
; SSE-NEXT:    movsd {{.*#+}} xmm5 = xmm0[0],xmm5[1]
; SSE-NEXT:    movapd %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa {{.*#+}} xmm2 = [255,255,255,255,255,255,255,255]
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    pand %xmm2, %xmm0
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
; SSE-NEXT:    pshuflw {{.*#+}} xmm0 = xmm0[0,1,1,3,4,5,6,7]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    movdqa %xmm10, %xmm5
; SSE-NEXT:    pandn %xmm0, %xmm5
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    pand %xmm2, %xmm0
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
; SSE-NEXT:    pshuflw {{.*#+}} xmm0 = xmm0[1,3,2,3,4,5,6,7]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    pand %xmm10, %xmm0
; SSE-NEXT:    por %xmm5, %xmm0
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    pand %xmm2, %xmm1
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm1[0,1,2,0]
; SSE-NEXT:    pshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,4,5,7,5]
; SSE-NEXT:    packuswb %xmm5, %xmm5
; SSE-NEXT:    movdqa %xmm13, %xmm15
; SSE-NEXT:    pandn %xmm5, %xmm15
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    pand %xmm2, %xmm1
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm1[0,1,2,0]
; SSE-NEXT:    pshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,7,5,6,7]
; SSE-NEXT:    packuswb %xmm5, %xmm5
; SSE-NEXT:    pand %xmm13, %xmm5
; SSE-NEXT:    por %xmm15, %xmm5
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[1,1,1,1]
; SSE-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm5[0],xmm0[1],xmm5[1]
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    pand %xmm2, %xmm1
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm1[0,1,2,0]
; SSE-NEXT:    pshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,4,5,7,5]
; SSE-NEXT:    packuswb %xmm5, %xmm5
; SSE-NEXT:    movdqa %xmm12, %xmm15
; SSE-NEXT:    pandn %xmm5, %xmm15
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    pand %xmm2, %xmm1
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm1[0,1,2,0]
; SSE-NEXT:    pshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,7,5,6,7]
; SSE-NEXT:    packuswb %xmm5, %xmm5
; SSE-NEXT:    pand %xmm12, %xmm5
; SSE-NEXT:    por %xmm15, %xmm5
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    pand %xmm2, %xmm1
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm15 = xmm1[0,2,2,3]
; SSE-NEXT:    pshuflw {{.*#+}} xmm15 = xmm15[0,1,1,3,4,5,6,7]
; SSE-NEXT:    packuswb %xmm15, %xmm15
; SSE-NEXT:    movdqa %xmm9, %xmm7
; SSE-NEXT:    pandn %xmm15, %xmm7
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    pand %xmm2, %xmm1
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm15 = xmm1[0,2,2,3]
; SSE-NEXT:    pshuflw {{.*#+}} xmm15 = xmm15[1,3,2,3,4,5,6,7]
; SSE-NEXT:    packuswb %xmm15, %xmm15
; SSE-NEXT:    pand %xmm9, %xmm15
; SSE-NEXT:    por %xmm7, %xmm15
; SSE-NEXT:    pshufd {{.*#+}} xmm7 = xmm15[2,2,2,2]
; SSE-NEXT:    punpckhdq {{.*#+}} xmm7 = xmm7[2],xmm5[2],xmm7[3],xmm5[3]
; SSE-NEXT:    movsd {{.*#+}} xmm7 = xmm0[0],xmm7[1]
; SSE-NEXT:    movapd %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    pand %xmm2, %xmm0
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
; SSE-NEXT:    pshuflw {{.*#+}} xmm0 = xmm0[0,1,1,3,4,5,6,7]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    movdqa %xmm10, %xmm5
; SSE-NEXT:    pandn %xmm0, %xmm5
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    pand %xmm2, %xmm0
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
; SSE-NEXT:    pshuflw {{.*#+}} xmm1 = xmm0[1,3,2,3,4,5,6,7]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm10, %xmm1
; SSE-NEXT:    por %xmm5, %xmm1
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    pand %xmm2, %xmm0
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm0[0,1,2,0]
; SSE-NEXT:    pshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,4,5,7,5]
; SSE-NEXT:    packuswb %xmm5, %xmm5
; SSE-NEXT:    movdqa %xmm13, %xmm7
; SSE-NEXT:    pandn %xmm5, %xmm7
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    pand %xmm2, %xmm0
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm0[0,1,2,0]
; SSE-NEXT:    pshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,7,5,6,7]
; SSE-NEXT:    packuswb %xmm5, %xmm5
; SSE-NEXT:    pand %xmm13, %xmm5
; SSE-NEXT:    por %xmm7, %xmm5
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[1,1,1,1]
; SSE-NEXT:    punpckldq {{.*#+}} xmm1 = xmm1[0],xmm5[0],xmm1[1],xmm5[1]
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    pand %xmm2, %xmm0
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm0[0,1,2,0]
; SSE-NEXT:    pshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,4,5,7,5]
; SSE-NEXT:    packuswb %xmm5, %xmm5
; SSE-NEXT:    movdqa %xmm12, %xmm7
; SSE-NEXT:    pandn %xmm5, %xmm7
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    pand %xmm2, %xmm0
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm0[0,1,2,0]
; SSE-NEXT:    pshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,7,5,6,7]
; SSE-NEXT:    packuswb %xmm5, %xmm5
; SSE-NEXT:    pand %xmm12, %xmm5
; SSE-NEXT:    por %xmm7, %xmm5
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    pand %xmm2, %xmm0
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm7 = xmm0[0,2,2,3]
; SSE-NEXT:    pshuflw {{.*#+}} xmm7 = xmm7[0,1,1,3,4,5,6,7]
; SSE-NEXT:    packuswb %xmm7, %xmm7
; SSE-NEXT:    movdqa %xmm9, %xmm15
; SSE-NEXT:    pandn %xmm7, %xmm15
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    pand %xmm2, %xmm0
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm7 = xmm0[0,2,2,3]
; SSE-NEXT:    pshuflw {{.*#+}} xmm7 = xmm7[1,3,2,3,4,5,6,7]
; SSE-NEXT:    packuswb %xmm7, %xmm7
; SSE-NEXT:    pand %xmm9, %xmm7
; SSE-NEXT:    por %xmm15, %xmm7
; SSE-NEXT:    pshufd {{.*#+}} xmm7 = xmm7[2,2,2,2]
; SSE-NEXT:    punpckhdq {{.*#+}} xmm7 = xmm7[2],xmm5[2],xmm7[3],xmm5[3]
; SSE-NEXT:    movsd {{.*#+}} xmm7 = xmm1[0],xmm7[1]
; SSE-NEXT:    movapd %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    pand %xmm2, %xmm0
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
; SSE-NEXT:    pshuflw {{.*#+}} xmm0 = xmm0[0,1,1,3,4,5,6,7]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    movdqa %xmm10, %xmm5
; SSE-NEXT:    pandn %xmm0, %xmm5
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    pand %xmm2, %xmm0
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
; SSE-NEXT:    pshuflw {{.*#+}} xmm0 = xmm0[1,3,2,3,4,5,6,7]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    pand %xmm10, %xmm0
; SSE-NEXT:    por %xmm5, %xmm0
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    pand %xmm2, %xmm1
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm1[0,1,2,0]
; SSE-NEXT:    pshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,4,5,7,5]
; SSE-NEXT:    packuswb %xmm5, %xmm5
; SSE-NEXT:    movdqa %xmm13, %xmm7
; SSE-NEXT:    pandn %xmm5, %xmm7
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    pand %xmm2, %xmm1
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm1[0,1,2,0]
; SSE-NEXT:    pshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,7,5,6,7]
; SSE-NEXT:    packuswb %xmm5, %xmm5
; SSE-NEXT:    pand %xmm13, %xmm5
; SSE-NEXT:    por %xmm7, %xmm5
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[1,1,1,1]
; SSE-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm5[0],xmm0[1],xmm5[1]
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    pand %xmm2, %xmm1
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm1[0,1,2,0]
; SSE-NEXT:    pshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,4,5,7,5]
; SSE-NEXT:    packuswb %xmm5, %xmm5
; SSE-NEXT:    movdqa %xmm12, %xmm7
; SSE-NEXT:    pandn %xmm5, %xmm7
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    pand %xmm2, %xmm1
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm1[0,1,2,0]
; SSE-NEXT:    pshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,7,5,6,7]
; SSE-NEXT:    packuswb %xmm5, %xmm5
; SSE-NEXT:    pand %xmm12, %xmm5
; SSE-NEXT:    por %xmm7, %xmm5
; SSE-NEXT:    movdqa (%rsp), %xmm1 # 16-byte Reload
; SSE-NEXT:    pand %xmm2, %xmm1
; SSE-NEXT:    movdqa %xmm1, (%rsp) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm7 = xmm1[0,2,2,3]
; SSE-NEXT:    pshuflw {{.*#+}} xmm7 = xmm7[0,1,1,3,4,5,6,7]
; SSE-NEXT:    packuswb %xmm7, %xmm7
; SSE-NEXT:    movdqa %xmm9, %xmm15
; SSE-NEXT:    pandn %xmm7, %xmm15
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    pand %xmm2, %xmm1
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm7 = xmm1[0,2,2,3]
; SSE-NEXT:    pshuflw {{.*#+}} xmm7 = xmm7[1,3,2,3,4,5,6,7]
; SSE-NEXT:    packuswb %xmm7, %xmm7
; SSE-NEXT:    pand %xmm9, %xmm7
; SSE-NEXT:    por %xmm15, %xmm7
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm7[2,2,2,2]
; SSE-NEXT:    punpckhdq {{.*#+}} xmm1 = xmm1[2],xmm5[2],xmm1[3],xmm5[3]
; SSE-NEXT:    movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]
; SSE-NEXT:    movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    pand %xmm2, %xmm0
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
; SSE-NEXT:    pshuflw {{.*#+}} xmm0 = xmm0[0,1,1,3,4,5,6,7]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    movdqa %xmm10, %xmm5
; SSE-NEXT:    pandn %xmm0, %xmm5
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    pand %xmm2, %xmm0
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
; SSE-NEXT:    pshuflw {{.*#+}} xmm0 = xmm0[1,3,2,3,4,5,6,7]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    pand %xmm10, %xmm0
; SSE-NEXT:    por %xmm5, %xmm0
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    pand %xmm2, %xmm1
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm1[0,1,2,0]
; SSE-NEXT:    pshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,4,5,7,5]
; SSE-NEXT:    packuswb %xmm5, %xmm5
; SSE-NEXT:    movdqa %xmm13, %xmm7
; SSE-NEXT:    pandn %xmm5, %xmm7
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    pand %xmm2, %xmm1
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm1[0,1,2,0]
; SSE-NEXT:    pshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,7,5,6,7]
; SSE-NEXT:    packuswb %xmm5, %xmm5
; SSE-NEXT:    pand %xmm13, %xmm5
; SSE-NEXT:    por %xmm7, %xmm5
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm5[1,1,1,1]
; SSE-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm5[0],xmm0[1],xmm5[1]
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    pand %xmm2, %xmm1
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm1[0,1,2,0]
; SSE-NEXT:    pshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,4,5,7,5]
; SSE-NEXT:    packuswb %xmm5, %xmm5
; SSE-NEXT:    movdqa %xmm12, %xmm7
; SSE-NEXT:    pandn %xmm5, %xmm7
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    pand %xmm2, %xmm1
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm1[0,1,2,0]
; SSE-NEXT:    pshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,7,5,6,7]
; SSE-NEXT:    packuswb %xmm5, %xmm5
; SSE-NEXT:    pand %xmm12, %xmm5
; SSE-NEXT:    por %xmm7, %xmm5
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    pand %xmm2, %xmm1
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm7 = xmm1[0,2,2,3]
; SSE-NEXT:    pshuflw {{.*#+}} xmm7 = xmm7[0,1,1,3,4,5,6,7]
; SSE-NEXT:    packuswb %xmm7, %xmm7
; SSE-NEXT:    movdqa %xmm9, %xmm15
; SSE-NEXT:    pandn %xmm7, %xmm15
; SSE-NEXT:    movdqa %xmm3, %xmm1
; SSE-NEXT:    pand %xmm2, %xmm1
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[0,2,2,3]
; SSE-NEXT:    pshuflw {{.*#+}} xmm1 = xmm1[1,3,2,3,4,5,6,7]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm9, %xmm1
; SSE-NEXT:    por %xmm15, %xmm1
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[2,2,2,2]
; SSE-NEXT:    punpckhdq {{.*#+}} xmm1 = xmm1[2],xmm5[2],xmm1[3],xmm5[3]
; SSE-NEXT:    movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]
; SSE-NEXT:    movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd $229, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
; SSE-NEXT:    # xmm0 = mem[1,1,2,3]
; SSE-NEXT:    movdqa %xmm10, %xmm1
; SSE-NEXT:    pandn %xmm0, %xmm1
; SSE-NEXT:    pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
; SSE-NEXT:    # xmm0 = mem[3,3,3,3]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    pand %xmm10, %xmm0
; SSE-NEXT:    por %xmm1, %xmm0
; SSE-NEXT:    movdqa %xmm13, %xmm1
; SSE-NEXT:    pandn {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    pshufd $250, {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Folded Reload
; SSE-NEXT:    # xmm5 = mem[2,2,3,3]
; SSE-NEXT:    packuswb %xmm5, %xmm5
; SSE-NEXT:    pand %xmm13, %xmm5
; SSE-NEXT:    por %xmm1, %xmm5
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm5[1,1,1,1]
; SSE-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
; SSE-NEXT:    movdqa %xmm12, %xmm1
; SSE-NEXT:    pandn {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    pshufd $250, {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Folded Reload
; SSE-NEXT:    # xmm5 = mem[2,2,3,3]
; SSE-NEXT:    packuswb %xmm5, %xmm5
; SSE-NEXT:    pand %xmm12, %xmm5
; SSE-NEXT:    por %xmm1, %xmm5
; SSE-NEXT:    pshufd $244, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[0,1,3,3]
; SSE-NEXT:    movdqa %xmm9, %xmm7
; SSE-NEXT:    pandn %xmm1, %xmm7
; SSE-NEXT:    pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[3,3,3,3]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm9, %xmm1
; SSE-NEXT:    por %xmm7, %xmm1
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[2,2,2,2]
; SSE-NEXT:    punpckhdq {{.*#+}} xmm1 = xmm1[2],xmm5[2],xmm1[3],xmm5[3]
; SSE-NEXT:    movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]
; SSE-NEXT:    movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd $229, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
; SSE-NEXT:    # xmm0 = mem[1,1,2,3]
; SSE-NEXT:    movdqa %xmm10, %xmm1
; SSE-NEXT:    pandn %xmm0, %xmm1
; SSE-NEXT:    pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
; SSE-NEXT:    # xmm0 = mem[3,3,3,3]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    pand %xmm10, %xmm0
; SSE-NEXT:    por %xmm1, %xmm0
; SSE-NEXT:    movdqa %xmm13, %xmm1
; SSE-NEXT:    pandn {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    pshufd $250, {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Folded Reload
; SSE-NEXT:    # xmm5 = mem[2,2,3,3]
; SSE-NEXT:    packuswb %xmm5, %xmm5
; SSE-NEXT:    pand %xmm13, %xmm5
; SSE-NEXT:    por %xmm1, %xmm5
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm5[1,1,1,1]
; SSE-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
; SSE-NEXT:    movdqa %xmm12, %xmm1
; SSE-NEXT:    pandn {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    pshufd $250, {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Folded Reload
; SSE-NEXT:    # xmm5 = mem[2,2,3,3]
; SSE-NEXT:    packuswb %xmm5, %xmm5
; SSE-NEXT:    pand %xmm12, %xmm5
; SSE-NEXT:    por %xmm1, %xmm5
; SSE-NEXT:    pshufd $244, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[0,1,3,3]
; SSE-NEXT:    movdqa %xmm9, %xmm7
; SSE-NEXT:    pandn %xmm1, %xmm7
; SSE-NEXT:    pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[3,3,3,3]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm9, %xmm1
; SSE-NEXT:    por %xmm7, %xmm1
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[2,2,2,2]
; SSE-NEXT:    punpckhdq {{.*#+}} xmm1 = xmm1[2],xmm5[2],xmm1[3],xmm5[3]
; SSE-NEXT:    movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]
; SSE-NEXT:    movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd $229, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
; SSE-NEXT:    # xmm0 = mem[1,1,2,3]
; SSE-NEXT:    movdqa %xmm10, %xmm1
; SSE-NEXT:    pandn %xmm0, %xmm1
; SSE-NEXT:    pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
; SSE-NEXT:    # xmm0 = mem[3,3,3,3]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    pand %xmm10, %xmm0
; SSE-NEXT:    por %xmm1, %xmm0
; SSE-NEXT:    movdqa %xmm13, %xmm1
; SSE-NEXT:    pandn %xmm14, %xmm1
; SSE-NEXT:    pshufd $250, {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Folded Reload
; SSE-NEXT:    # xmm5 = mem[2,2,3,3]
; SSE-NEXT:    packuswb %xmm5, %xmm5
; SSE-NEXT:    pand %xmm13, %xmm5
; SSE-NEXT:    por %xmm1, %xmm5
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm5[1,1,1,1]
; SSE-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
; SSE-NEXT:    movdqa %xmm12, %xmm1
; SSE-NEXT:    pandn {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    pshufd $250, {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Folded Reload
; SSE-NEXT:    # xmm5 = mem[2,2,3,3]
; SSE-NEXT:    packuswb %xmm5, %xmm5
; SSE-NEXT:    pand %xmm12, %xmm5
; SSE-NEXT:    por %xmm1, %xmm5
; SSE-NEXT:    pshufd $244, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[0,1,3,3]
; SSE-NEXT:    movdqa %xmm9, %xmm7
; SSE-NEXT:    pandn %xmm1, %xmm7
; SSE-NEXT:    pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[3,3,3,3]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm9, %xmm1
; SSE-NEXT:    por %xmm7, %xmm1
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[2,2,2,2]
; SSE-NEXT:    punpckhdq {{.*#+}} xmm1 = xmm1[2],xmm5[2],xmm1[3],xmm5[3]
; SSE-NEXT:    movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]
; SSE-NEXT:    movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm8[1,1,2,3]
; SSE-NEXT:    movdqa %xmm10, %xmm1
; SSE-NEXT:    pandn %xmm0, %xmm1
; SSE-NEXT:    pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
; SSE-NEXT:    # xmm0 = mem[3,3,3,3]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    pand %xmm10, %xmm0
; SSE-NEXT:    por %xmm1, %xmm0
; SSE-NEXT:    movdqa %xmm13, %xmm1
; SSE-NEXT:    pandn %xmm4, %xmm1
; SSE-NEXT:    pshufd $250, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
; SSE-NEXT:    # xmm2 = mem[2,2,3,3]
; SSE-NEXT:    packuswb %xmm2, %xmm2
; SSE-NEXT:    pand %xmm13, %xmm2
; SSE-NEXT:    por %xmm1, %xmm2
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[1,1,1,1]
; SSE-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
; SSE-NEXT:    movdqa %xmm12, %xmm1
; SSE-NEXT:    pandn {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm11[2,2,3,3]
; SSE-NEXT:    packuswb %xmm2, %xmm2
; SSE-NEXT:    pand %xmm12, %xmm2
; SSE-NEXT:    por %xmm1, %xmm2
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm6[0,1,3,3]
; SSE-NEXT:    movdqa %xmm9, %xmm3
; SSE-NEXT:    pandn %xmm1, %xmm3
; SSE-NEXT:    pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[3,3,3,3]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm9, %xmm1
; SSE-NEXT:    por %xmm3, %xmm1
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[2,2,2,2]
; SSE-NEXT:    punpckhdq {{.*#+}} xmm1 = xmm1[2],xmm2[2],xmm1[3],xmm2[3]
; SSE-NEXT:    movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]
; SSE-NEXT:    movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd $231, {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Folded Reload
; SSE-NEXT:    # xmm7 = mem[3,1,2,3]
; SSE-NEXT:    pshuflw {{.*#+}} xmm0 = xmm7[0,1,2,0,4,5,6,7]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    movdqa %xmm10, %xmm1
; SSE-NEXT:    pandn %xmm0, %xmm1
; SSE-NEXT:    pshufd $231, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
; SSE-NEXT:    # xmm0 = mem[3,1,2,3]
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshuflw {{.*#+}} xmm0 = xmm0[2,0,2,3,4,5,6,7]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    pand %xmm10, %xmm0
; SSE-NEXT:    por %xmm1, %xmm0
; SSE-NEXT:    pshufd $212, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[0,1,1,3]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,5,4,6]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    movdqa %xmm13, %xmm2
; SSE-NEXT:    pandn %xmm1, %xmm2
; SSE-NEXT:    pshufd $212, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[0,1,1,3]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,6,6,7]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm13, %xmm1
; SSE-NEXT:    por %xmm2, %xmm1
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,1,1]
; SSE-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
; SSE-NEXT:    pshufd $212, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[0,1,1,3]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,5,4,6]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    movdqa %xmm12, %xmm2
; SSE-NEXT:    pandn %xmm1, %xmm2
; SSE-NEXT:    pshufd $212, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[0,1,1,3]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,6,6,7]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm12, %xmm1
; SSE-NEXT:    por %xmm2, %xmm1
; SSE-NEXT:    pshufd $231, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
; SSE-NEXT:    # xmm2 = mem[3,1,2,3]
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshuflw {{.*#+}} xmm2 = xmm2[0,1,2,0,4,5,6,7]
; SSE-NEXT:    packuswb %xmm2, %xmm2
; SSE-NEXT:    movdqa %xmm9, %xmm3
; SSE-NEXT:    pandn %xmm2, %xmm3
; SSE-NEXT:    pshufd $231, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
; SSE-NEXT:    # xmm2 = mem[3,1,2,3]
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshuflw {{.*#+}} xmm2 = xmm2[2,0,2,3,4,5,6,7]
; SSE-NEXT:    packuswb %xmm2, %xmm2
; SSE-NEXT:    pand %xmm9, %xmm2
; SSE-NEXT:    por %xmm3, %xmm2
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[2,2,2,2]
; SSE-NEXT:    punpckhdq {{.*#+}} xmm2 = xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; SSE-NEXT:    movsd {{.*#+}} xmm2 = xmm0[0],xmm2[1]
; SSE-NEXT:    movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd $231, {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Folded Reload
; SSE-NEXT:    # xmm14 = mem[3,1,2,3]
; SSE-NEXT:    pshuflw {{.*#+}} xmm0 = xmm14[0,1,2,0,4,5,6,7]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    movdqa %xmm10, %xmm1
; SSE-NEXT:    pandn %xmm0, %xmm1
; SSE-NEXT:    pshufd $231, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
; SSE-NEXT:    # xmm0 = mem[3,1,2,3]
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshuflw {{.*#+}} xmm0 = xmm0[2,0,2,3,4,5,6,7]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    pand %xmm10, %xmm0
; SSE-NEXT:    por %xmm1, %xmm0
; SSE-NEXT:    pshufd $212, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[0,1,1,3]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,5,4,6]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    movdqa %xmm13, %xmm2
; SSE-NEXT:    pandn %xmm1, %xmm2
; SSE-NEXT:    pshufd $212, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[0,1,1,3]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,6,6,7]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm13, %xmm1
; SSE-NEXT:    por %xmm2, %xmm1
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,1,1]
; SSE-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
; SSE-NEXT:    pshufd $212, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[0,1,1,3]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,5,4,6]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    movdqa %xmm12, %xmm2
; SSE-NEXT:    pandn %xmm1, %xmm2
; SSE-NEXT:    pshufd $212, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[0,1,1,3]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,6,6,7]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm12, %xmm1
; SSE-NEXT:    por %xmm2, %xmm1
; SSE-NEXT:    pshufd $231, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
; SSE-NEXT:    # xmm2 = mem[3,1,2,3]
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshuflw {{.*#+}} xmm2 = xmm2[0,1,2,0,4,5,6,7]
; SSE-NEXT:    packuswb %xmm2, %xmm2
; SSE-NEXT:    movdqa %xmm9, %xmm3
; SSE-NEXT:    pandn %xmm2, %xmm3
; SSE-NEXT:    pshufd $231, {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Folded Reload
; SSE-NEXT:    # xmm11 = mem[3,1,2,3]
; SSE-NEXT:    pshuflw {{.*#+}} xmm2 = xmm11[2,0,2,3,4,5,6,7]
; SSE-NEXT:    packuswb %xmm2, %xmm2
; SSE-NEXT:    pand %xmm9, %xmm2
; SSE-NEXT:    por %xmm3, %xmm2
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[2,2,2,2]
; SSE-NEXT:    punpckhdq {{.*#+}} xmm2 = xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; SSE-NEXT:    movsd {{.*#+}} xmm2 = xmm0[0],xmm2[1]
; SSE-NEXT:    movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd $231, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
; SSE-NEXT:    # xmm0 = mem[3,1,2,3]
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshuflw {{.*#+}} xmm0 = xmm0[0,1,2,0,4,5,6,7]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    movdqa %xmm10, %xmm1
; SSE-NEXT:    pandn %xmm0, %xmm1
; SSE-NEXT:    pshufd $231, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
; SSE-NEXT:    # xmm0 = mem[3,1,2,3]
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshuflw {{.*#+}} xmm0 = xmm0[2,0,2,3,4,5,6,7]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    pand %xmm10, %xmm0
; SSE-NEXT:    por %xmm1, %xmm0
; SSE-NEXT:    pshufd $212, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[0,1,1,3]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,5,4,6]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    movdqa %xmm13, %xmm2
; SSE-NEXT:    pandn %xmm1, %xmm2
; SSE-NEXT:    pshufd $212, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[0,1,1,3]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,6,6,7]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm13, %xmm1
; SSE-NEXT:    por %xmm2, %xmm1
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,1,1]
; SSE-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
; SSE-NEXT:    pshufd $212, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[0,1,1,3]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,5,4,6]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    movdqa %xmm12, %xmm2
; SSE-NEXT:    pandn %xmm1, %xmm2
; SSE-NEXT:    pshufd $212, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[0,1,1,3]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,6,6,7]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm12, %xmm1
; SSE-NEXT:    por %xmm2, %xmm1
; SSE-NEXT:    pshufd $231, (%rsp), %xmm2 # 16-byte Folded Reload
; SSE-NEXT:    # xmm2 = mem[3,1,2,3]
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshuflw {{.*#+}} xmm2 = xmm2[0,1,2,0,4,5,6,7]
; SSE-NEXT:    packuswb %xmm2, %xmm2
; SSE-NEXT:    movdqa %xmm9, %xmm3
; SSE-NEXT:    pandn %xmm2, %xmm3
; SSE-NEXT:    pshufd $231, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
; SSE-NEXT:    # xmm2 = mem[3,1,2,3]
; SSE-NEXT:    movdqa %xmm2, (%rsp) # 16-byte Spill
; SSE-NEXT:    pshuflw {{.*#+}} xmm2 = xmm2[2,0,2,3,4,5,6,7]
; SSE-NEXT:    packuswb %xmm2, %xmm2
; SSE-NEXT:    pand %xmm9, %xmm2
; SSE-NEXT:    por %xmm3, %xmm2
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[2,2,2,2]
; SSE-NEXT:    punpckhdq {{.*#+}} xmm2 = xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; SSE-NEXT:    movsd {{.*#+}} xmm2 = xmm0[0],xmm2[1]
; SSE-NEXT:    movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd $231, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
; SSE-NEXT:    # xmm0 = mem[3,1,2,3]
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshuflw {{.*#+}} xmm0 = xmm0[0,1,2,0,4,5,6,7]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    movdqa %xmm10, %xmm1
; SSE-NEXT:    pandn %xmm0, %xmm1
; SSE-NEXT:    pshufd $231, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
; SSE-NEXT:    # xmm0 = mem[3,1,2,3]
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshuflw {{.*#+}} xmm0 = xmm0[2,0,2,3,4,5,6,7]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    pand %xmm10, %xmm0
; SSE-NEXT:    por %xmm1, %xmm0
; SSE-NEXT:    pshufd $212, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[0,1,1,3]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,5,4,6]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    movdqa %xmm13, %xmm2
; SSE-NEXT:    pandn %xmm1, %xmm2
; SSE-NEXT:    pshufd $212, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[0,1,1,3]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,6,6,7]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm13, %xmm1
; SSE-NEXT:    por %xmm2, %xmm1
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,1,1]
; SSE-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
; SSE-NEXT:    pshufd $212, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[0,1,1,3]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,5,4,6]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    movdqa %xmm12, %xmm2
; SSE-NEXT:    pandn %xmm1, %xmm2
; SSE-NEXT:    pshufd $212, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[0,1,1,3]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,6,6,7]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm12, %xmm1
; SSE-NEXT:    por %xmm2, %xmm1
; SSE-NEXT:    pshufd $231, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
; SSE-NEXT:    # xmm2 = mem[3,1,2,3]
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshuflw {{.*#+}} xmm2 = xmm2[0,1,2,0,4,5,6,7]
; SSE-NEXT:    packuswb %xmm2, %xmm2
; SSE-NEXT:    movdqa %xmm9, %xmm3
; SSE-NEXT:    pandn %xmm2, %xmm3
; SSE-NEXT:    pshufd $231, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
; SSE-NEXT:    # xmm2 = mem[3,1,2,3]
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshuflw {{.*#+}} xmm2 = xmm2[2,0,2,3,4,5,6,7]
; SSE-NEXT:    packuswb %xmm2, %xmm2
; SSE-NEXT:    pand %xmm9, %xmm2
; SSE-NEXT:    por %xmm3, %xmm2
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[2,2,2,2]
; SSE-NEXT:    punpckhdq {{.*#+}} xmm2 = xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; SSE-NEXT:    movsd {{.*#+}} xmm2 = xmm0[0],xmm2[1]
; SSE-NEXT:    movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; SSE-NEXT:    punpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
; SSE-NEXT:    # xmm2 = xmm2[4],mem[4],xmm2[5],mem[5],xmm2[6],mem[6],xmm2[7],mem[7]
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    punpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
; SSE-NEXT:    # xmm0 = xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7]
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    packuswb %xmm2, %xmm2
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm10, %xmm1
; SSE-NEXT:    pandn %xmm2, %xmm1
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,1,1]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    pand %xmm10, %xmm0
; SSE-NEXT:    por %xmm1, %xmm0
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    punpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = xmm1[4],mem[4],xmm1[5],mem[5],xmm1[6],mem[6],xmm1[7],mem[7]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[0,0,2,3]
; SSE-NEXT:    movdqa %xmm13, %xmm2
; SSE-NEXT:    pandn %xmm1, %xmm2
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    punpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = xmm1[4],mem[4],xmm1[5],mem[5],xmm1[6],mem[6],xmm1[7],mem[7]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[0,1,1,3]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm13, %xmm1
; SSE-NEXT:    por %xmm2, %xmm1
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,1,1]
; SSE-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    punpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = xmm1[4],mem[4],xmm1[5],mem[5],xmm1[6],mem[6],xmm1[7],mem[7]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[0,1,2,2]
; SSE-NEXT:    movdqa %xmm12, %xmm2
; SSE-NEXT:    pandn %xmm1, %xmm2
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    punpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = xmm1[4],mem[4],xmm1[5],mem[5],xmm1[6],mem[6],xmm1[7],mem[7]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[0,1,1,3]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm12, %xmm1
; SSE-NEXT:    por %xmm2, %xmm1
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; SSE-NEXT:    punpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
; SSE-NEXT:    # xmm2 = xmm2[4],mem[4],xmm2[5],mem[5],xmm2[6],mem[6],xmm2[7],mem[7]
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; SSE-NEXT:    punpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Folded Reload
; SSE-NEXT:    # xmm4 = xmm4[4],mem[4],xmm4[5],mem[5],xmm4[6],mem[6],xmm4[7],mem[7]
; SSE-NEXT:    movdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    packuswb %xmm2, %xmm3
; SSE-NEXT:    movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm9, %xmm2
; SSE-NEXT:    pandn %xmm3, %xmm2
; SSE-NEXT:    pshufd {{.*#+}} xmm3 = xmm4[1,1,1,1]
; SSE-NEXT:    packuswb %xmm3, %xmm3
; SSE-NEXT:    pand %xmm9, %xmm3
; SSE-NEXT:    por %xmm2, %xmm3
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[2,2,2,2]
; SSE-NEXT:    punpckhdq {{.*#+}} xmm2 = xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; SSE-NEXT:    movsd {{.*#+}} xmm2 = xmm0[0],xmm2[1]
; SSE-NEXT:    movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; SSE-NEXT:    punpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
; SSE-NEXT:    # xmm2 = xmm2[4],mem[4],xmm2[5],mem[5],xmm2[6],mem[6],xmm2[7],mem[7]
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    punpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
; SSE-NEXT:    # xmm0 = xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7]
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    packuswb %xmm2, %xmm2
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm10, %xmm1
; SSE-NEXT:    pandn %xmm2, %xmm1
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,1,1]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    pand %xmm10, %xmm0
; SSE-NEXT:    por %xmm1, %xmm0
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    punpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = xmm1[4],mem[4],xmm1[5],mem[5],xmm1[6],mem[6],xmm1[7],mem[7]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[0,0,2,3]
; SSE-NEXT:    movdqa %xmm13, %xmm2
; SSE-NEXT:    pandn %xmm1, %xmm2
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    punpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = xmm1[4],mem[4],xmm1[5],mem[5],xmm1[6],mem[6],xmm1[7],mem[7]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[0,1,1,3]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm13, %xmm1
; SSE-NEXT:    por %xmm2, %xmm1
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,1,1]
; SSE-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    punpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = xmm1[4],mem[4],xmm1[5],mem[5],xmm1[6],mem[6],xmm1[7],mem[7]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[0,1,2,2]
; SSE-NEXT:    movdqa %xmm12, %xmm2
; SSE-NEXT:    pandn %xmm1, %xmm2
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    punpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = xmm1[4],mem[4],xmm1[5],mem[5],xmm1[6],mem[6],xmm1[7],mem[7]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[0,1,1,3]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm12, %xmm1
; SSE-NEXT:    por %xmm2, %xmm1
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; SSE-NEXT:    punpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
; SSE-NEXT:    # xmm2 = xmm2[4],mem[4],xmm2[5],mem[5],xmm2[6],mem[6],xmm2[7],mem[7]
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; SSE-NEXT:    punpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload
; SSE-NEXT:    # xmm3 = xmm3[4],mem[4],xmm3[5],mem[5],xmm3[6],mem[6],xmm3[7],mem[7]
; SSE-NEXT:    movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    packuswb %xmm2, %xmm8
; SSE-NEXT:    movdqa %xmm9, %xmm2
; SSE-NEXT:    pandn %xmm8, %xmm2
; SSE-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,1,1]
; SSE-NEXT:    packuswb %xmm3, %xmm3
; SSE-NEXT:    pand %xmm9, %xmm3
; SSE-NEXT:    por %xmm2, %xmm3
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[2,2,2,2]
; SSE-NEXT:    punpckhdq {{.*#+}} xmm2 = xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; SSE-NEXT:    movsd {{.*#+}} xmm2 = xmm0[0],xmm2[1]
; SSE-NEXT:    movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    punpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
; SSE-NEXT:    # xmm0 = xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7]
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; SSE-NEXT:    punpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
; SSE-NEXT:    # xmm2 = xmm2[4],mem[4],xmm2[5],mem[5],xmm2[6],mem[6],xmm2[7],mem[7]
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm10, %xmm1
; SSE-NEXT:    pandn %xmm0, %xmm1
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,1,1]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    pand %xmm10, %xmm0
; SSE-NEXT:    por %xmm1, %xmm0
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    punpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = xmm1[4],mem[4],xmm1[5],mem[5],xmm1[6],mem[6],xmm1[7],mem[7]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[0,0,2,3]
; SSE-NEXT:    movdqa %xmm13, %xmm2
; SSE-NEXT:    pandn %xmm1, %xmm2
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    punpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = xmm1[4],mem[4],xmm1[5],mem[5],xmm1[6],mem[6],xmm1[7],mem[7]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[0,1,1,3]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm13, %xmm1
; SSE-NEXT:    por %xmm2, %xmm1
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,1,1]
; SSE-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    punpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = xmm1[4],mem[4],xmm1[5],mem[5],xmm1[6],mem[6],xmm1[7],mem[7]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[0,1,2,2]
; SSE-NEXT:    movdqa %xmm12, %xmm2
; SSE-NEXT:    pandn %xmm1, %xmm2
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    punpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = xmm1[4],mem[4],xmm1[5],mem[5],xmm1[6],mem[6],xmm1[7],mem[7]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[0,1,1,3]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm12, %xmm1
; SSE-NEXT:    por %xmm2, %xmm1
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; SSE-NEXT:    punpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
; SSE-NEXT:    # xmm2 = xmm2[4],mem[4],xmm2[5],mem[5],xmm2[6],mem[6],xmm2[7],mem[7]
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; SSE-NEXT:    punpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Folded Reload
; SSE-NEXT:    # xmm4 = xmm4[4],mem[4],xmm4[5],mem[5],xmm4[6],mem[6],xmm4[7],mem[7]
; SSE-NEXT:    movdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    packuswb %xmm2, %xmm3
; SSE-NEXT:    movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm9, %xmm2
; SSE-NEXT:    pandn %xmm3, %xmm2
; SSE-NEXT:    pshufd {{.*#+}} xmm3 = xmm4[1,1,1,1]
; SSE-NEXT:    packuswb %xmm3, %xmm3
; SSE-NEXT:    pand %xmm9, %xmm3
; SSE-NEXT:    por %xmm2, %xmm3
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[2,2,2,2]
; SSE-NEXT:    punpckhdq {{.*#+}} xmm2 = xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; SSE-NEXT:    movsd {{.*#+}} xmm2 = xmm0[0],xmm2[1]
; SSE-NEXT:    movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; SSE-NEXT:    punpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
; SSE-NEXT:    # xmm2 = xmm2[4],mem[4],xmm2[5],mem[5],xmm2[6],mem[6],xmm2[7],mem[7]
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT:    punpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
; SSE-NEXT:    # xmm0 = xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7]
; SSE-NEXT:    movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    packuswb %xmm2, %xmm2
; SSE-NEXT:    movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    movdqa %xmm10, %xmm1
; SSE-NEXT:    pandn %xmm2, %xmm1
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,1,1]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    pand %xmm10, %xmm0
; SSE-NEXT:    por %xmm1, %xmm0
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    punpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = xmm1[4],mem[4],xmm1[5],mem[5],xmm1[6],mem[6],xmm1[7],mem[7]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[0,0,2,3]
; SSE-NEXT:    movdqa %xmm13, %xmm2
; SSE-NEXT:    pandn %xmm1, %xmm2
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    punpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = xmm1[4],mem[4],xmm1[5],mem[5],xmm1[6],mem[6],xmm1[7],mem[7]
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[0,1,1,3]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm13, %xmm1
; SSE-NEXT:    por %xmm2, %xmm1
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,1,1]
; SSE-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; SSE-NEXT:    punpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = xmm1[4],mem[4],xmm1[5],mem[5],xmm1[6],mem[6],xmm1[7],mem[7]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[0,1,2,2]
; SSE-NEXT:    movdqa %xmm12, %xmm2
; SSE-NEXT:    pandn %xmm1, %xmm2
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; SSE-NEXT:    punpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Folded Reload
; SSE-NEXT:    # xmm6 = xmm6[4],mem[4],xmm6[5],mem[5],xmm6[6],mem[6],xmm6[7],mem[7]
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm6[0,1,1,3]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm12, %xmm1
; SSE-NEXT:    por %xmm2, %xmm1
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; SSE-NEXT:    punpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
; SSE-NEXT:    # xmm2 = xmm2[4],mem[4],xmm2[5],mem[5],xmm2[6],mem[6],xmm2[7],mem[7]
; SSE-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; SSE-NEXT:    punpckhwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload
; SSE-NEXT:    # xmm3 = xmm3[4],mem[4],xmm3[5],mem[5],xmm3[6],mem[6],xmm3[7],mem[7]
; SSE-NEXT:    movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT:    packuswb %xmm2, %xmm4
; SSE-NEXT:    movdqa %xmm9, %xmm2
; SSE-NEXT:    pandn %xmm4, %xmm2
; SSE-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,1,1]
; SSE-NEXT:    packuswb %xmm3, %xmm3
; SSE-NEXT:    pand %xmm9, %xmm3
; SSE-NEXT:    por %xmm2, %xmm3
; SSE-NEXT:    pshufd {{.*#+}} xmm15 = xmm3[2,2,2,2]
; SSE-NEXT:    punpckhdq {{.*#+}} xmm15 = xmm15[2],xmm1[2],xmm15[3],xmm1[3]
; SSE-NEXT:    movsd {{.*#+}} xmm15 = xmm0[0],xmm15[1]
; SSE-NEXT:    pshuflw {{.*#+}} xmm0 = xmm14[0,1,3,1,4,5,6,7]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    movdqa %xmm10, %xmm1
; SSE-NEXT:    pandn %xmm0, %xmm1
; SSE-NEXT:    pshuflw $231, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
; SSE-NEXT:    # xmm0 = mem[3,1,2,3,4,5,6,7]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    pand %xmm10, %xmm0
; SSE-NEXT:    por %xmm1, %xmm0
; SSE-NEXT:    pshufhw $212, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[0,1,2,3,4,5,5,7]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    movdqa %xmm13, %xmm2
; SSE-NEXT:    pandn %xmm1, %xmm2
; SSE-NEXT:    pshufhw $237, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[0,1,2,3,5,7,6,7]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm13, %xmm1
; SSE-NEXT:    por %xmm2, %xmm1
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,1,1]
; SSE-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
; SSE-NEXT:    pshufhw $212, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[0,1,2,3,4,5,5,7]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    movdqa %xmm12, %xmm2
; SSE-NEXT:    pandn %xmm1, %xmm2
; SSE-NEXT:    pshufhw $237, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[0,1,2,3,5,7,6,7]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm12, %xmm1
; SSE-NEXT:    por %xmm2, %xmm1
; SSE-NEXT:    pshuflw $116, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
; SSE-NEXT:    # xmm2 = mem[0,1,3,1,4,5,6,7]
; SSE-NEXT:    packuswb %xmm2, %xmm2
; SSE-NEXT:    movdqa %xmm9, %xmm3
; SSE-NEXT:    pandn %xmm2, %xmm3
; SSE-NEXT:    pshuflw {{.*#+}} xmm2 = xmm11[3,1,2,3,4,5,6,7]
; SSE-NEXT:    packuswb %xmm2, %xmm2
; SSE-NEXT:    pand %xmm9, %xmm2
; SSE-NEXT:    por %xmm3, %xmm2
; SSE-NEXT:    pshufd {{.*#+}} xmm14 = xmm2[2,2,2,2]
; SSE-NEXT:    punpckhdq {{.*#+}} xmm14 = xmm14[2],xmm1[2],xmm14[3],xmm1[3]
; SSE-NEXT:    movsd {{.*#+}} xmm14 = xmm0[0],xmm14[1]
; SSE-NEXT:    pshuflw {{.*#+}} xmm0 = xmm7[0,1,3,1,4,5,6,7]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    movdqa %xmm10, %xmm1
; SSE-NEXT:    pandn %xmm0, %xmm1
; SSE-NEXT:    pshuflw $231, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
; SSE-NEXT:    # xmm0 = mem[3,1,2,3,4,5,6,7]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    pand %xmm10, %xmm0
; SSE-NEXT:    por %xmm1, %xmm0
; SSE-NEXT:    pshufhw $212, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[0,1,2,3,4,5,5,7]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    movdqa %xmm13, %xmm2
; SSE-NEXT:    pandn %xmm1, %xmm2
; SSE-NEXT:    pshufhw $237, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[0,1,2,3,5,7,6,7]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm13, %xmm1
; SSE-NEXT:    por %xmm2, %xmm1
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,1,1]
; SSE-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
; SSE-NEXT:    pshufhw $212, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[0,1,2,3,4,5,5,7]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    movdqa %xmm12, %xmm2
; SSE-NEXT:    pandn %xmm1, %xmm2
; SSE-NEXT:    pshufhw $237, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[0,1,2,3,5,7,6,7]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm12, %xmm1
; SSE-NEXT:    por %xmm2, %xmm1
; SSE-NEXT:    pshuflw $116, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
; SSE-NEXT:    # xmm2 = mem[0,1,3,1,4,5,6,7]
; SSE-NEXT:    packuswb %xmm2, %xmm2
; SSE-NEXT:    movdqa %xmm9, %xmm3
; SSE-NEXT:    pandn %xmm2, %xmm3
; SSE-NEXT:    pshuflw $231, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
; SSE-NEXT:    # xmm2 = mem[3,1,2,3,4,5,6,7]
; SSE-NEXT:    packuswb %xmm2, %xmm2
; SSE-NEXT:    pand %xmm9, %xmm2
; SSE-NEXT:    por %xmm3, %xmm2
; SSE-NEXT:    pshufd {{.*#+}} xmm11 = xmm2[2,2,2,2]
; SSE-NEXT:    punpckhdq {{.*#+}} xmm11 = xmm11[2],xmm1[2],xmm11[3],xmm1[3]
; SSE-NEXT:    movsd {{.*#+}} xmm11 = xmm0[0],xmm11[1]
; SSE-NEXT:    pshuflw $116, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
; SSE-NEXT:    # xmm0 = mem[0,1,3,1,4,5,6,7]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    movdqa %xmm10, %xmm1
; SSE-NEXT:    pandn %xmm0, %xmm1
; SSE-NEXT:    pshuflw $231, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
; SSE-NEXT:    # xmm0 = mem[3,1,2,3,4,5,6,7]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    pand %xmm10, %xmm0
; SSE-NEXT:    por %xmm1, %xmm0
; SSE-NEXT:    pshufhw $212, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[0,1,2,3,4,5,5,7]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    movdqa %xmm13, %xmm2
; SSE-NEXT:    pandn %xmm1, %xmm2
; SSE-NEXT:    pshufhw $237, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[0,1,2,3,5,7,6,7]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm13, %xmm1
; SSE-NEXT:    por %xmm2, %xmm1
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,1,1]
; SSE-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
; SSE-NEXT:    pshufhw $212, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[0,1,2,3,4,5,5,7]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    movdqa %xmm12, %xmm2
; SSE-NEXT:    pandn %xmm1, %xmm2
; SSE-NEXT:    pshufhw $237, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[0,1,2,3,5,7,6,7]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm12, %xmm1
; SSE-NEXT:    por %xmm2, %xmm1
; SSE-NEXT:    pshuflw $116, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
; SSE-NEXT:    # xmm2 = mem[0,1,3,1,4,5,6,7]
; SSE-NEXT:    packuswb %xmm2, %xmm2
; SSE-NEXT:    movdqa %xmm9, %xmm3
; SSE-NEXT:    pandn %xmm2, %xmm3
; SSE-NEXT:    pshuflw $231, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
; SSE-NEXT:    # xmm2 = mem[3,1,2,3,4,5,6,7]
; SSE-NEXT:    packuswb %xmm2, %xmm2
; SSE-NEXT:    pand %xmm9, %xmm2
; SSE-NEXT:    por %xmm3, %xmm2
; SSE-NEXT:    pshufd {{.*#+}} xmm7 = xmm2[2,2,2,2]
; SSE-NEXT:    punpckhdq {{.*#+}} xmm7 = xmm7[2],xmm1[2],xmm7[3],xmm1[3]
; SSE-NEXT:    movsd {{.*#+}} xmm7 = xmm0[0],xmm7[1]
; SSE-NEXT:    pshuflw $116, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
; SSE-NEXT:    # xmm0 = mem[0,1,3,1,4,5,6,7]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    movdqa %xmm10, %xmm1
; SSE-NEXT:    pandn %xmm0, %xmm1
; SSE-NEXT:    pshuflw $231, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
; SSE-NEXT:    # xmm0 = mem[3,1,2,3,4,5,6,7]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    pand %xmm10, %xmm0
; SSE-NEXT:    por %xmm1, %xmm0
; SSE-NEXT:    pshufhw $212, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[0,1,2,3,4,5,5,7]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    movdqa %xmm13, %xmm2
; SSE-NEXT:    pandn %xmm1, %xmm2
; SSE-NEXT:    pshufhw $237, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[0,1,2,3,5,7,6,7]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm13, %xmm1
; SSE-NEXT:    por %xmm2, %xmm1
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,1,1]
; SSE-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
; SSE-NEXT:    pshufhw $212, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[0,1,2,3,4,5,5,7]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    movdqa %xmm12, %xmm2
; SSE-NEXT:    pandn %xmm1, %xmm2
; SSE-NEXT:    pshufhw $237, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[0,1,2,3,5,7,6,7]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm12, %xmm1
; SSE-NEXT:    por %xmm2, %xmm1
; SSE-NEXT:    pshuflw $116, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
; SSE-NEXT:    # xmm2 = mem[0,1,3,1,4,5,6,7]
; SSE-NEXT:    packuswb %xmm2, %xmm2
; SSE-NEXT:    movdqa %xmm9, %xmm3
; SSE-NEXT:    pandn %xmm2, %xmm3
; SSE-NEXT:    pshuflw $231, (%rsp), %xmm2 # 16-byte Folded Reload
; SSE-NEXT:    # xmm2 = mem[3,1,2,3,4,5,6,7]
; SSE-NEXT:    packuswb %xmm2, %xmm2
; SSE-NEXT:    pand %xmm9, %xmm2
; SSE-NEXT:    por %xmm3, %xmm2
; SSE-NEXT:    pshufd {{.*#+}} xmm5 = xmm2[2,2,2,2]
; SSE-NEXT:    punpckhdq {{.*#+}} xmm5 = xmm5[2],xmm1[2],xmm5[3],xmm1[3]
; SSE-NEXT:    movsd {{.*#+}} xmm5 = xmm0[0],xmm5[1]
; SSE-NEXT:    pshufd $229, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
; SSE-NEXT:    # xmm0 = mem[1,1,2,3]
; SSE-NEXT:    movdqa %xmm10, %xmm1
; SSE-NEXT:    pandn %xmm0, %xmm1
; SSE-NEXT:    pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
; SSE-NEXT:    # xmm0 = mem[3,3,3,3]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    pand %xmm10, %xmm0
; SSE-NEXT:    por %xmm1, %xmm0
; SSE-NEXT:    movdqa %xmm13, %xmm1
; SSE-NEXT:    pandn {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    pshufd $250, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
; SSE-NEXT:    # xmm2 = mem[2,2,3,3]
; SSE-NEXT:    packuswb %xmm2, %xmm2
; SSE-NEXT:    pand %xmm13, %xmm2
; SSE-NEXT:    por %xmm1, %xmm2
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[1,1,1,1]
; SSE-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
; SSE-NEXT:    movdqa %xmm12, %xmm1
; SSE-NEXT:    pandn {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    pshufd $250, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
; SSE-NEXT:    # xmm2 = mem[2,2,3,3]
; SSE-NEXT:    packuswb %xmm2, %xmm2
; SSE-NEXT:    pand %xmm12, %xmm2
; SSE-NEXT:    por %xmm1, %xmm2
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm8[0,1,3,3]
; SSE-NEXT:    movdqa %xmm9, %xmm3
; SSE-NEXT:    pandn %xmm1, %xmm3
; SSE-NEXT:    pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[3,3,3,3]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm9, %xmm1
; SSE-NEXT:    por %xmm3, %xmm1
; SSE-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[2,2,2,2]
; SSE-NEXT:    punpckhdq {{.*#+}} xmm3 = xmm3[2],xmm2[2],xmm3[3],xmm2[3]
; SSE-NEXT:    movsd {{.*#+}} xmm3 = xmm0[0],xmm3[1]
; SSE-NEXT:    pshufd $229, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
; SSE-NEXT:    # xmm0 = mem[1,1,2,3]
; SSE-NEXT:    movdqa %xmm10, %xmm1
; SSE-NEXT:    pandn %xmm0, %xmm1
; SSE-NEXT:    pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
; SSE-NEXT:    # xmm0 = mem[3,3,3,3]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    pand %xmm10, %xmm0
; SSE-NEXT:    por %xmm1, %xmm0
; SSE-NEXT:    movdqa %xmm13, %xmm1
; SSE-NEXT:    pandn {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    pshufd $250, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
; SSE-NEXT:    # xmm2 = mem[2,2,3,3]
; SSE-NEXT:    packuswb %xmm2, %xmm2
; SSE-NEXT:    pand %xmm13, %xmm2
; SSE-NEXT:    por %xmm1, %xmm2
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[1,1,1,1]
; SSE-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
; SSE-NEXT:    movdqa %xmm12, %xmm8
; SSE-NEXT:    pandn {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Folded Reload
; SSE-NEXT:    pshufd $250, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
; SSE-NEXT:    # xmm2 = mem[2,2,3,3]
; SSE-NEXT:    packuswb %xmm2, %xmm1
; SSE-NEXT:    pand %xmm12, %xmm1
; SSE-NEXT:    por %xmm8, %xmm1
; SSE-NEXT:    pshufd $244, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
; SSE-NEXT:    # xmm2 = mem[0,1,3,3]
; SSE-NEXT:    movdqa %xmm9, %xmm8
; SSE-NEXT:    pandn %xmm2, %xmm8
; SSE-NEXT:    pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload
; SSE-NEXT:    # xmm2 = mem[3,3,3,3]
; SSE-NEXT:    packuswb %xmm2, %xmm2
; SSE-NEXT:    pand %xmm9, %xmm2
; SSE-NEXT:    por %xmm8, %xmm2
; SSE-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[2,2,2,2]
; SSE-NEXT:    punpckhdq {{.*#+}} xmm2 = xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; SSE-NEXT:    movsd {{.*#+}} xmm2 = xmm0[0],xmm2[1]
; SSE-NEXT:    pshufd $229, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
; SSE-NEXT:    # xmm0 = mem[1,1,2,3]
; SSE-NEXT:    movdqa %xmm10, %xmm1
; SSE-NEXT:    pandn %xmm0, %xmm1
; SSE-NEXT:    pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
; SSE-NEXT:    # xmm0 = mem[3,3,3,3]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    pand %xmm10, %xmm0
; SSE-NEXT:    por %xmm1, %xmm0
; SSE-NEXT:    movdqa %xmm13, %xmm1
; SSE-NEXT:    pandn {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    pshufd $250, {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Folded Reload
; SSE-NEXT:    # xmm8 = mem[2,2,3,3]
; SSE-NEXT:    packuswb %xmm8, %xmm8
; SSE-NEXT:    pand %xmm13, %xmm8
; SSE-NEXT:    por %xmm1, %xmm8
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm8[1,1,1,1]
; SSE-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
; SSE-NEXT:    movdqa %xmm12, %xmm1
; SSE-NEXT:    pandn {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    pshufd {{.*#+}} xmm8 = xmm6[2,2,3,3]
; SSE-NEXT:    packuswb %xmm8, %xmm8
; SSE-NEXT:    pand %xmm12, %xmm8
; SSE-NEXT:    por %xmm1, %xmm8
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm4[0,1,3,3]
; SSE-NEXT:    movdqa %xmm9, %xmm4
; SSE-NEXT:    pandn %xmm1, %xmm4
; SSE-NEXT:    pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT:    # xmm1 = mem[3,3,3,3]
; SSE-NEXT:    packuswb %xmm1, %xmm1
; SSE-NEXT:    pand %xmm9, %xmm1
; SSE-NEXT:    por %xmm4, %xmm1
; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[2,2,2,2]
; SSE-NEXT:    punpckhdq {{.*#+}} xmm1 = xmm1[2],xmm8[2],xmm1[3],xmm8[3]
; SSE-NEXT:    movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]
; SSE-NEXT:    pshufd $229, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
; SSE-NEXT:    # xmm0 = mem[1,1,2,3]
; SSE-NEXT:    pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Folded Reload
; SSE-NEXT:    # xmm4 = mem[3,3,3,3]
; SSE-NEXT:    packuswb %xmm4, %xmm4
; SSE-NEXT:    pand %xmm10, %xmm4
; SSE-NEXT:    pandn %xmm0, %xmm10
; SSE-NEXT:    por %xmm4, %xmm10
; SSE-NEXT:    pshufd $250, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
; SSE-NEXT:    # xmm0 = mem[2,2,3,3]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    pand %xmm13, %xmm0
; SSE-NEXT:    pandn {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Folded Reload
; SSE-NEXT:    por %xmm0, %xmm13
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm13[1,1,1,1]
; SSE-NEXT:    punpckldq {{.*#+}} xmm10 = xmm10[0],xmm0[0],xmm10[1],xmm0[1]
; SSE-NEXT:    pshufd $250, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
; SSE-NEXT:    # xmm0 = mem[2,2,3,3]
; SSE-NEXT:    packuswb %xmm0, %xmm0
; SSE-NEXT:    pand %xmm12, %xmm0
; SSE-NEXT:    pandn {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Folded Reload
; SSE-NEXT:    por %xmm0, %xmm12
; SSE-NEXT:    pshufd $244, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
; SSE-NEXT:    # xmm0 = mem[0,1,3,3]
; SSE-NEXT:    pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Folded Reload
; SSE-NEXT:    # xmm4 = mem[3,3,3,3]
; SSE-NEXT:    packuswb %xmm4, %xmm4
; SSE-NEXT:    pand %xmm9, %xmm4
; SSE-NEXT:    pandn %xmm0, %xmm9
; SSE-NEXT:    por %xmm4, %xmm9
; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm9[2,2,2,2]
; SSE-NEXT:    punpckhdq {{.*#+}} xmm0 = xmm0[2],xmm12[2],xmm0[3],xmm12[3]
; SSE-NEXT:    movsd {{.*#+}} xmm0 = xmm10[0],xmm0[1]
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; SSE-NEXT:    movaps %xmm4, 32(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; SSE-NEXT:    movaps %xmm4, 48(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; SSE-NEXT:    movaps %xmm4, (%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; SSE-NEXT:    movaps %xmm4, 16(%rsi)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; SSE-NEXT:    movaps %xmm4, 32(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; SSE-NEXT:    movaps %xmm4, 48(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; SSE-NEXT:    movaps %xmm4, (%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; SSE-NEXT:    movaps %xmm4, 16(%rdx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; SSE-NEXT:    movaps %xmm4, 32(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; SSE-NEXT:    movaps %xmm4, 48(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; SSE-NEXT:    movaps %xmm4, (%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; SSE-NEXT:    movaps %xmm4, 16(%rcx)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; SSE-NEXT:    movaps %xmm4, 32(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; SSE-NEXT:    movaps %xmm4, 48(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; SSE-NEXT:    movaps %xmm4, (%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; SSE-NEXT:    movaps %xmm4, 16(%r8)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; SSE-NEXT:    movaps %xmm4, 32(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; SSE-NEXT:    movaps %xmm4, 48(%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; SSE-NEXT:    movaps %xmm4, (%r9)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; SSE-NEXT:    movaps %xmm4, 16(%r9)
; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT:    movapd %xmm15, 32(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; SSE-NEXT:    movaps %xmm4, 48(%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; SSE-NEXT:    movaps %xmm4, (%rax)
; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; SSE-NEXT:    movaps %xmm4, 16(%rax)
; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT:    movapd %xmm5, 48(%rax)
; SSE-NEXT:    movapd %xmm7, 32(%rax)
; SSE-NEXT:    movapd %xmm11, 16(%rax)
; SSE-NEXT:    movapd %xmm14, (%rax)
; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT:    movapd %xmm0, 48(%rax)
; SSE-NEXT:    movapd %xmm1, 32(%rax)
; SSE-NEXT:    movapd %xmm2, 16(%rax)
; SSE-NEXT:    movapd %xmm3, (%rax)
; SSE-NEXT:    addq $2040, %rsp # imm = 0x7F8
; SSE-NEXT:    retq
;
; AVX-LABEL: load_i8_stride8_vf64:
; AVX:       # %bb.0:
; AVX-NEXT:    subq $808, %rsp # imm = 0x328
; AVX-NEXT:    vbroadcastss {{.*#+}} xmm3 = [0,0,0,8,0,0,0,8,0,0,0,8,0,0,0,8]
; AVX-NEXT:    vmovdqa 368(%rdi), %xmm0
; AVX-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vpshufb %xmm3, %xmm0, %xmm0
; AVX-NEXT:    vmovdqa 352(%rdi), %xmm13
; AVX-NEXT:    vpshufb %xmm3, %xmm13, %xmm1
; AVX-NEXT:    vmovdqa %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
; AVX-NEXT:    vbroadcastss {{.*#+}} xmm1 = [0,8,0,0,0,8,0,0,0,8,0,0,0,8,0,0]
; AVX-NEXT:    vmovdqa 336(%rdi), %xmm4
; AVX-NEXT:    vpshufb %xmm1, %xmm4, %xmm2
; AVX-NEXT:    vmovdqa %xmm4, %xmm15
; AVX-NEXT:    vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovdqa 320(%rdi), %xmm4
; AVX-NEXT:    vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vpshufb %xmm1, %xmm4, %xmm4
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm4[0],xmm2[0],xmm4[1],xmm2[1],xmm4[2],xmm2[2],xmm4[3],xmm2[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm4 = xmm2[0,1,2,3,4,5],xmm0[6,7]
; AVX-NEXT:    vmovd {{.*#+}} xmm2 = [0,0,0,8,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vmovdqa 304(%rdi), %xmm12
; AVX-NEXT:    vpshufb %xmm2, %xmm12, %xmm0
; AVX-NEXT:    vmovdqa %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovdqa 288(%rdi), %xmm5
; AVX-NEXT:    vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vpshufb %xmm2, %xmm5, %xmm5
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm0[0],xmm5[1],xmm0[1],xmm5[2],xmm0[2],xmm5[3],xmm0[3]
; AVX-NEXT:    vmovd {{.*#+}} xmm9 = [0,8,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vmovdqa 272(%rdi), %xmm11
; AVX-NEXT:    vpshufb %xmm9, %xmm11, %xmm6
; AVX-NEXT:    vmovdqa %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovdqa 256(%rdi), %xmm10
; AVX-NEXT:    vpshufb %xmm9, %xmm10, %xmm7
; AVX-NEXT:    vmovdqa %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm5 = xmm6[0,1],xmm5[2,3],xmm6[4,5,6,7]
; AVX-NEXT:    vpblendw {{.*#+}} xmm4 = xmm5[0,1,2,3],xmm4[4,5,6,7]
; AVX-NEXT:    vmovdqa 496(%rdi), %xmm0
; AVX-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vpshufb %xmm3, %xmm0, %xmm5
; AVX-NEXT:    vmovdqa 480(%rdi), %xmm0
; AVX-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vpshufb %xmm3, %xmm0, %xmm6
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX-NEXT:    vmovdqa 464(%rdi), %xmm0
; AVX-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vpshufb %xmm1, %xmm0, %xmm6
; AVX-NEXT:    vmovdqa 448(%rdi), %xmm0
; AVX-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vpshufb %xmm1, %xmm0, %xmm7
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm5 = xmm6[0,1,2,3,4,5],xmm5[6,7]
; AVX-NEXT:    vmovdqa 432(%rdi), %xmm0
; AVX-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vpshufb %xmm2, %xmm0, %xmm6
; AVX-NEXT:    vmovdqa 416(%rdi), %xmm14
; AVX-NEXT:    vpshufb %xmm2, %xmm14, %xmm7
; AVX-NEXT:    vmovdqa %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
; AVX-NEXT:    vmovdqa 400(%rdi), %xmm0
; AVX-NEXT:    vmovdqa %xmm0, (%rsp) # 16-byte Spill
; AVX-NEXT:    vpshufb %xmm9, %xmm0, %xmm7
; AVX-NEXT:    vmovdqa 384(%rdi), %xmm0
; AVX-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vpshufb %xmm9, %xmm0, %xmm8
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm7 = xmm8[0],xmm7[0],xmm8[1],xmm7[1],xmm8[2],xmm7[2],xmm8[3],xmm7[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm6 = xmm7[0,1],xmm6[2,3],xmm7[4,5,6,7]
; AVX-NEXT:    vinsertf128 $1, %xmm5, %ymm0, %ymm5
; AVX-NEXT:    vinsertf128 $1, %xmm6, %ymm0, %ymm6
; AVX-NEXT:    vblendps {{.*#+}} ymm5 = ymm6[0,1,2,3,4,5],ymm5[6,7]
; AVX-NEXT:    vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm5[4,5,6,7]
; AVX-NEXT:    vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovdqa 112(%rdi), %xmm0
; AVX-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vpshufb %xmm3, %xmm0, %xmm4
; AVX-NEXT:    vmovdqa 96(%rdi), %xmm0
; AVX-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vpshufb %xmm3, %xmm0, %xmm5
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX-NEXT:    vmovdqa 80(%rdi), %xmm0
; AVX-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vpshufb %xmm1, %xmm0, %xmm5
; AVX-NEXT:    vmovdqa 64(%rdi), %xmm0
; AVX-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vpshufb %xmm1, %xmm0, %xmm6
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm4 = xmm5[0,1,2,3,4,5],xmm4[6,7]
; AVX-NEXT:    vmovdqa 32(%rdi), %xmm0
; AVX-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovdqa 48(%rdi), %xmm5
; AVX-NEXT:    vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vpshufb %xmm2, %xmm5, %xmm5
; AVX-NEXT:    vpshufb %xmm2, %xmm0, %xmm6
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX-NEXT:    vmovdqa (%rdi), %xmm0
; AVX-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vmovdqa 16(%rdi), %xmm6
; AVX-NEXT:    vmovdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vpshufb %xmm9, %xmm6, %xmm6
; AVX-NEXT:    vpshufb %xmm9, %xmm0, %xmm7
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm5 = xmm6[0,1],xmm5[2,3],xmm6[4,5,6,7]
; AVX-NEXT:    vpblendw {{.*#+}} xmm4 = xmm5[0,1,2,3],xmm4[4,5,6,7]
; AVX-NEXT:    vmovdqa 240(%rdi), %xmm0
; AVX-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vpshufb %xmm3, %xmm0, %xmm5
; AVX-NEXT:    vmovdqa 224(%rdi), %xmm0
; AVX-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vpshufb %xmm3, %xmm0, %xmm3
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm5[0],xmm3[1],xmm5[1],xmm3[2],xmm5[2],xmm3[3],xmm5[3]
; AVX-NEXT:    vmovdqa 208(%rdi), %xmm0
; AVX-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vpshufb %xmm1, %xmm0, %xmm5
; AVX-NEXT:    vmovdqa 192(%rdi), %xmm0
; AVX-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vpshufb %xmm1, %xmm0, %xmm1
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm5[0],xmm1[1],xmm5[1],xmm1[2],xmm5[2],xmm1[3],xmm5[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,5],xmm3[6,7]
; AVX-NEXT:    vmovdqa 176(%rdi), %xmm0
; AVX-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vpshufb %xmm2, %xmm0, %xmm3
; AVX-NEXT:    vmovdqa 160(%rdi), %xmm0
; AVX-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vpshufb %xmm2, %xmm0, %xmm2
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
; AVX-NEXT:    vmovdqa 144(%rdi), %xmm0
; AVX-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vpshufb %xmm9, %xmm0, %xmm3
; AVX-NEXT:    vmovdqa 128(%rdi), %xmm0
; AVX-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX-NEXT:    vpshufb %xmm9, %xmm0, %xmm0
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0,1],xmm2[2,3],xmm0[4,5,6,7]
; AVX-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm1
; AVX-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm4[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vbroadcastss {{.*#+}} xmm6 = [0,0,1,9,0,0,1,9,0,0,1,9,0,0,1,9]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm6, %xmm0, %xmm0
; AVX-NEXT:    vpshufb %xmm6, %xmm13, %xmm1
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
; AVX-NEXT:    vbroadcastss {{.*#+}} xmm1 = [1,9,0,0,1,9,0,0,1,9,0,0,1,9,0,0]
; AVX-NEXT:    vpshufb %xmm1, %xmm15, %xmm3
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm1, %xmm13, %xmm4
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm4 = xmm3[0,1,2,3,4,5],xmm0[6,7]
; AVX-NEXT:    vmovd {{.*#+}} xmm2 = [0,0,1,9,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vpshufb %xmm2, %xmm12, %xmm3
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm2, %xmm7, %xmm5
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm3[0],xmm5[1],xmm3[1],xmm5[2],xmm3[2],xmm5[3],xmm3[3]
; AVX-NEXT:    vmovd {{.*#+}} xmm3 = [1,9,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vpshufb %xmm3, %xmm11, %xmm0
; AVX-NEXT:    vpshufb %xmm3, %xmm10, %xmm15
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm15[0],xmm0[0],xmm15[1],xmm0[1],xmm15[2],xmm0[2],xmm15[3],xmm0[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0,1],xmm5[2,3],xmm0[4,5,6,7]
; AVX-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0,1,2,3],xmm4[4,5,6,7]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm6, %xmm12, %xmm4
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm6, %xmm9, %xmm5
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm1, %xmm10, %xmm5
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm1, %xmm11, %xmm15
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm15[0],xmm5[0],xmm15[1],xmm5[1],xmm15[2],xmm5[2],xmm15[3],xmm5[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm4 = xmm5[0,1,2,3,4,5],xmm4[6,7]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm2, %xmm8, %xmm5
; AVX-NEXT:    vpshufb %xmm2, %xmm14, %xmm15
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm15[0],xmm5[0],xmm15[1],xmm5[1],xmm15[2],xmm5[2],xmm15[3],xmm5[3]
; AVX-NEXT:    vmovdqa (%rsp), %xmm14 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm3, %xmm14, %xmm15
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm3, %xmm14, %xmm14
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm14[0],xmm15[0],xmm14[1],xmm15[1],xmm14[2],xmm15[2],xmm14[3],xmm15[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm5 = xmm14[0,1],xmm5[2,3],xmm14[4,5,6,7]
; AVX-NEXT:    vinsertf128 $1, %xmm4, %ymm0, %ymm4
; AVX-NEXT:    vinsertf128 $1, %xmm5, %ymm0, %ymm5
; AVX-NEXT:    vblendps {{.*#+}} ymm4 = ymm5[0,1,2,3,4,5],ymm4[6,7]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm4[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm6, %xmm0, %xmm0
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm6, %xmm4, %xmm4
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm4[0],xmm0[0],xmm4[1],xmm0[1],xmm4[2],xmm0[2],xmm4[3],xmm0[3]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm1, %xmm4, %xmm4
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm1, %xmm5, %xmm5
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm0 = xmm4[0,1,2,3,4,5],xmm0[6,7]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm2, %xmm4, %xmm4
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm2, %xmm5, %xmm5
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm3, %xmm5, %xmm5
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm3, %xmm14, %xmm14
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm14[0],xmm5[0],xmm14[1],xmm5[1],xmm14[2],xmm5[2],xmm14[3],xmm5[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm4 = xmm5[0,1],xmm4[2,3],xmm5[4,5,6,7]
; AVX-NEXT:    vpblendw {{.*#+}} xmm0 = xmm4[0,1,2,3],xmm0[4,5,6,7]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm6, %xmm4, %xmm4
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm6, %xmm5, %xmm5
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm1, %xmm5, %xmm5
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm1, %xmm6, %xmm1
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm5[0],xmm1[1],xmm5[1],xmm1[2],xmm5[2],xmm1[3],xmm5[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,5],xmm4[6,7]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm2, %xmm4, %xmm4
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm2, %xmm5, %xmm2
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm3, %xmm4, %xmm4
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm3, %xmm5, %xmm3
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm2 = xmm3[0,1],xmm2[2,3],xmm3[4,5,6,7]
; AVX-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm1
; AVX-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm2
; AVX-NEXT:    vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vbroadcastss {{.*#+}} xmm0 = [0,0,2,10,0,0,2,10,0,0,2,10,0,0,2,10]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm0, %xmm1, %xmm1
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm0, %xmm2, %xmm2
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX-NEXT:    vbroadcastss {{.*#+}} xmm1 = [2,10,0,0,2,10,0,0,2,10,0,0,2,10,0,0]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm1, %xmm3, %xmm3
; AVX-NEXT:    vpshufb %xmm1, %xmm13, %xmm4
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm4 = xmm3[0,1,2,3,4,5],xmm2[6,7]
; AVX-NEXT:    vmovd {{.*#+}} xmm2 = [0,0,2,10,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm2, %xmm3, %xmm3
; AVX-NEXT:    vpshufb %xmm2, %xmm7, %xmm5
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm3[0],xmm5[1],xmm3[1],xmm5[2],xmm3[2],xmm5[3],xmm3[3]
; AVX-NEXT:    vmovd {{.*#+}} xmm3 = [2,10,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm3, %xmm7, %xmm6
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm3, %xmm13, %xmm14
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm5 = xmm6[0,1],xmm5[2,3],xmm6[4,5,6,7]
; AVX-NEXT:    vpblendw {{.*#+}} xmm4 = xmm5[0,1,2,3],xmm4[4,5,6,7]
; AVX-NEXT:    vpshufb %xmm0, %xmm12, %xmm5
; AVX-NEXT:    vpshufb %xmm0, %xmm9, %xmm6
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX-NEXT:    vpshufb %xmm1, %xmm10, %xmm6
; AVX-NEXT:    vpshufb %xmm1, %xmm11, %xmm14
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm5 = xmm6[0,1,2,3,4,5],xmm5[6,7]
; AVX-NEXT:    vpshufb %xmm2, %xmm8, %xmm6
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm2, %xmm9, %xmm14
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3]
; AVX-NEXT:    vmovdqa (%rsp), %xmm10 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm3, %xmm10, %xmm14
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm3, %xmm11, %xmm15
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm6 = xmm14[0,1],xmm6[2,3],xmm14[4,5,6,7]
; AVX-NEXT:    vinsertf128 $1, %xmm5, %ymm0, %ymm5
; AVX-NEXT:    vinsertf128 $1, %xmm6, %ymm0, %ymm6
; AVX-NEXT:    vblendps {{.*#+}} ymm5 = ymm6[0,1,2,3,4,5],ymm5[6,7]
; AVX-NEXT:    vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm5[4,5,6,7]
; AVX-NEXT:    vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm0, %xmm12, %xmm4
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm0, %xmm8, %xmm5
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm1, %xmm5, %xmm5
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm1, %xmm6, %xmm6
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm4 = xmm5[0,1,2,3,4,5],xmm4[6,7]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm2, %xmm5, %xmm5
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm2, %xmm6, %xmm6
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm3, %xmm6, %xmm6
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm3, %xmm13, %xmm14
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm5 = xmm6[0,1],xmm5[2,3],xmm6[4,5,6,7]
; AVX-NEXT:    vpblendw {{.*#+}} xmm4 = xmm5[0,1,2,3],xmm4[4,5,6,7]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm0, %xmm5, %xmm5
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm0, %xmm6, %xmm0
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm5[0],xmm0[1],xmm5[1],xmm0[2],xmm5[2],xmm0[3],xmm5[3]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm1, %xmm5, %xmm5
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm1, %xmm6, %xmm1
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm5[0],xmm1[1],xmm5[1],xmm1[2],xmm5[2],xmm1[3],xmm5[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm0 = xmm1[0,1,2,3,4,5],xmm0[6,7]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm2, %xmm1, %xmm1
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm2, %xmm5, %xmm2
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm3, %xmm2, %xmm2
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm3, %xmm5, %xmm3
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3],xmm2[4,5,6,7]
; AVX-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
; AVX-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm1
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm4[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vbroadcastss {{.*#+}} xmm0 = [0,0,3,11,0,0,3,11,0,0,3,11,0,0,3,11]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm0, %xmm13, %xmm1
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm0, %xmm2, %xmm2
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX-NEXT:    vbroadcastss {{.*#+}} xmm1 = [3,11,0,0,3,11,0,0,3,11,0,0,3,11,0,0]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm1, %xmm3, %xmm3
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm1, %xmm4, %xmm4
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm4 = xmm3[0,1,2,3,4,5],xmm2[6,7]
; AVX-NEXT:    vmovd {{.*#+}} xmm2 = [0,0,3,11,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm2, %xmm3, %xmm3
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm2, %xmm5, %xmm5
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm3[0],xmm5[1],xmm3[1],xmm5[2],xmm3[2],xmm5[3],xmm3[3]
; AVX-NEXT:    vmovd {{.*#+}} xmm3 = [3,11,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vpshufb %xmm3, %xmm7, %xmm6
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm3, %xmm7, %xmm14
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm5 = xmm6[0,1],xmm5[2,3],xmm6[4,5,6,7]
; AVX-NEXT:    vpblendw {{.*#+}} xmm4 = xmm5[0,1,2,3],xmm4[4,5,6,7]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm0, %xmm5, %xmm5
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm0, %xmm6, %xmm6
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm1, %xmm6, %xmm6
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm1, %xmm7, %xmm14
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm5 = xmm6[0,1,2,3,4,5],xmm5[6,7]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm2, %xmm6, %xmm6
; AVX-NEXT:    vpshufb %xmm2, %xmm9, %xmm14
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3]
; AVX-NEXT:    vpshufb %xmm3, %xmm10, %xmm14
; AVX-NEXT:    vpshufb %xmm3, %xmm11, %xmm15
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm6 = xmm14[0,1],xmm6[2,3],xmm14[4,5,6,7]
; AVX-NEXT:    vinsertf128 $1, %xmm5, %ymm0, %ymm5
; AVX-NEXT:    vinsertf128 $1, %xmm6, %ymm0, %ymm6
; AVX-NEXT:    vblendps {{.*#+}} ymm5 = ymm6[0,1,2,3,4,5],ymm5[6,7]
; AVX-NEXT:    vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm5[4,5,6,7]
; AVX-NEXT:    vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vpshufb %xmm0, %xmm12, %xmm4
; AVX-NEXT:    vpshufb %xmm0, %xmm8, %xmm5
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm1, %xmm9, %xmm5
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm1, %xmm10, %xmm6
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm4 = xmm5[0,1,2,3,4,5],xmm4[6,7]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm2, %xmm12, %xmm5
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm2, %xmm11, %xmm6
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm3, %xmm6, %xmm6
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm3, %xmm7, %xmm14
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm5 = xmm6[0,1],xmm5[2,3],xmm6[4,5,6,7]
; AVX-NEXT:    vpblendw {{.*#+}} xmm4 = xmm5[0,1,2,3],xmm4[4,5,6,7]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm0, %xmm5, %xmm5
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm0, %xmm6, %xmm0
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm5[0],xmm0[1],xmm5[1],xmm0[2],xmm5[2],xmm0[3],xmm5[3]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm1, %xmm5, %xmm5
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm1, %xmm6, %xmm1
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm5[0],xmm1[1],xmm5[1],xmm1[2],xmm5[2],xmm1[3],xmm5[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm0 = xmm1[0,1,2,3,4,5],xmm0[6,7]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm2, %xmm1, %xmm1
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm2, %xmm5, %xmm2
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm3, %xmm2, %xmm2
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm3, %xmm5, %xmm3
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3],xmm2[4,5,6,7]
; AVX-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
; AVX-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm1
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm4[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vbroadcastss {{.*#+}} xmm0 = [0,0,4,12,0,0,4,12,0,0,4,12,0,0,4,12]
; AVX-NEXT:    vpshufb %xmm0, %xmm13, %xmm1
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm0, %xmm2, %xmm2
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX-NEXT:    vbroadcastss {{.*#+}} xmm1 = [4,12,0,0,4,12,0,0,4,12,0,0,4,12,0,0]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm1, %xmm3, %xmm3
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm1, %xmm4, %xmm4
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm4 = xmm3[0,1,2,3,4,5],xmm2[6,7]
; AVX-NEXT:    vmovd {{.*#+}} xmm2 = [0,0,4,12,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm2, %xmm7, %xmm3
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm2, %xmm5, %xmm5
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm3[0],xmm5[1],xmm3[1],xmm5[2],xmm3[2],xmm5[3],xmm3[3]
; AVX-NEXT:    vmovd {{.*#+}} xmm3 = [4,12,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm3, %xmm6, %xmm6
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm3, %xmm8, %xmm14
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm5 = xmm6[0,1],xmm5[2,3],xmm6[4,5,6,7]
; AVX-NEXT:    vpblendw {{.*#+}} xmm4 = xmm5[0,1,2,3],xmm4[4,5,6,7]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm0, %xmm5, %xmm5
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm0, %xmm6, %xmm6
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm1, %xmm6, %xmm6
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm1, %xmm13, %xmm14
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm5 = xmm6[0,1,2,3,4,5],xmm5[6,7]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm2, %xmm6, %xmm6
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm2, %xmm13, %xmm14
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3]
; AVX-NEXT:    vmovdqa (%rsp), %xmm13 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm3, %xmm13, %xmm14
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm3, %xmm13, %xmm15
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm6 = xmm14[0,1],xmm6[2,3],xmm14[4,5,6,7]
; AVX-NEXT:    vinsertf128 $1, %xmm5, %ymm0, %ymm5
; AVX-NEXT:    vinsertf128 $1, %xmm6, %ymm0, %ymm6
; AVX-NEXT:    vblendps {{.*#+}} ymm5 = ymm6[0,1,2,3,4,5],ymm5[6,7]
; AVX-NEXT:    vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm5[4,5,6,7]
; AVX-NEXT:    vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm0, %xmm4, %xmm4
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm0, %xmm5, %xmm5
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX-NEXT:    vpshufb %xmm1, %xmm9, %xmm5
; AVX-NEXT:    vpshufb %xmm1, %xmm10, %xmm6
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm4 = xmm5[0,1,2,3,4,5],xmm4[6,7]
; AVX-NEXT:    vpshufb %xmm2, %xmm12, %xmm5
; AVX-NEXT:    vpshufb %xmm2, %xmm11, %xmm6
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm3, %xmm10, %xmm6
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm3, %xmm9, %xmm14
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm5 = xmm6[0,1],xmm5[2,3],xmm6[4,5,6,7]
; AVX-NEXT:    vpblendw {{.*#+}} xmm4 = xmm5[0,1,2,3],xmm4[4,5,6,7]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm0, %xmm11, %xmm5
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm0, %xmm12, %xmm0
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm5[0],xmm0[1],xmm5[1],xmm0[2],xmm5[2],xmm0[3],xmm5[3]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm1, %xmm13, %xmm5
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm1, %xmm6, %xmm1
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm5[0],xmm1[1],xmm5[1],xmm1[2],xmm5[2],xmm1[3],xmm5[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm0 = xmm1[0,1,2,3,4,5],xmm0[6,7]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm2, %xmm1, %xmm1
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm2, %xmm5, %xmm2
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm3, %xmm2, %xmm2
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm3, %xmm5, %xmm3
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3],xmm2[4,5,6,7]
; AVX-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
; AVX-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm1
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm4[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vbroadcastss {{.*#+}} xmm0 = [0,0,5,13,0,0,5,13,0,0,5,13,0,0,5,13]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm0, %xmm1, %xmm1
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm0, %xmm2, %xmm2
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX-NEXT:    vbroadcastss {{.*#+}} xmm1 = [5,13,0,0,5,13,0,0,5,13,0,0,5,13,0,0]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm1, %xmm3, %xmm3
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm1, %xmm4, %xmm4
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm4 = xmm3[0,1,2,3,4,5],xmm2[6,7]
; AVX-NEXT:    vmovd {{.*#+}} xmm2 = [0,0,5,13,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vpshufb %xmm2, %xmm7, %xmm3
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm2, %xmm5, %xmm5
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm3[0],xmm5[1],xmm3[1],xmm5[2],xmm3[2],xmm5[3],xmm3[3]
; AVX-NEXT:    vmovd {{.*#+}} xmm3 = [5,13,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm3, %xmm6, %xmm6
; AVX-NEXT:    vpshufb %xmm3, %xmm8, %xmm14
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm5 = xmm6[0,1],xmm5[2,3],xmm6[4,5,6,7]
; AVX-NEXT:    vpblendw {{.*#+}} xmm4 = xmm5[0,1,2,3],xmm4[4,5,6,7]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm0, %xmm5, %xmm5
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm0, %xmm6, %xmm6
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm1, %xmm6, %xmm6
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm1, %xmm7, %xmm14
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm5 = xmm6[0,1,2,3,4,5],xmm5[6,7]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm2, %xmm6, %xmm6
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm2, %xmm7, %xmm14
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3]
; AVX-NEXT:    vmovdqa (%rsp), %xmm7 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm3, %xmm7, %xmm14
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm3, %xmm7, %xmm15
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm6 = xmm14[0,1],xmm6[2,3],xmm14[4,5,6,7]
; AVX-NEXT:    vinsertf128 $1, %xmm5, %ymm0, %ymm5
; AVX-NEXT:    vinsertf128 $1, %xmm6, %ymm0, %ymm6
; AVX-NEXT:    vblendps {{.*#+}} ymm5 = ymm6[0,1,2,3,4,5],ymm5[6,7]
; AVX-NEXT:    vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm5[4,5,6,7]
; AVX-NEXT:    vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm0, %xmm4, %xmm4
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm0, %xmm5, %xmm5
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm1, %xmm5, %xmm5
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm1, %xmm6, %xmm6
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm4 = xmm5[0,1,2,3,4,5],xmm4[6,7]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm2, %xmm5, %xmm5
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm2, %xmm6, %xmm6
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX-NEXT:    vpshufb %xmm3, %xmm10, %xmm6
; AVX-NEXT:    vpshufb %xmm3, %xmm9, %xmm14
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm5 = xmm6[0,1],xmm5[2,3],xmm6[4,5,6,7]
; AVX-NEXT:    vpblendw {{.*#+}} xmm4 = xmm5[0,1,2,3],xmm4[4,5,6,7]
; AVX-NEXT:    vpshufb %xmm0, %xmm11, %xmm5
; AVX-NEXT:    vpshufb %xmm0, %xmm12, %xmm0
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm5[0],xmm0[1],xmm5[1],xmm0[2],xmm5[2],xmm0[3],xmm5[3]
; AVX-NEXT:    vpshufb %xmm1, %xmm13, %xmm5
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm1, %xmm13, %xmm1
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm5[0],xmm1[1],xmm5[1],xmm1[2],xmm5[2],xmm1[3],xmm5[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm0 = xmm1[0,1,2,3,4,5],xmm0[6,7]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm2, %xmm7, %xmm1
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm2, %xmm8, %xmm2
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm3, %xmm9, %xmm2
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm3, %xmm10, %xmm3
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3],xmm2[4,5,6,7]
; AVX-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
; AVX-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm1
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm4[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vbroadcastss {{.*#+}} xmm0 = [0,0,6,14,0,0,6,14,0,0,6,14,0,0,6,14]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm0, %xmm1, %xmm1
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm0, %xmm11, %xmm2
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX-NEXT:    vbroadcastss {{.*#+}} xmm1 = [6,14,0,0,6,14,0,0,6,14,0,0,6,14,0,0]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm1, %xmm12, %xmm3
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm1, %xmm4, %xmm4
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm4 = xmm3[0,1,2,3,4,5],xmm2[6,7]
; AVX-NEXT:    vmovd {{.*#+}} xmm2 = [0,0,6,14,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm2, %xmm3, %xmm3
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm2, %xmm5, %xmm5
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm3[0],xmm5[1],xmm3[1],xmm5[2],xmm3[2],xmm5[3],xmm3[3]
; AVX-NEXT:    vmovd {{.*#+}} xmm3 = [6,14,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm3, %xmm6, %xmm6
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm3, %xmm14, %xmm14
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm5 = xmm6[0,1],xmm5[2,3],xmm6[4,5,6,7]
; AVX-NEXT:    vpblendw {{.*#+}} xmm4 = xmm5[0,1,2,3],xmm4[4,5,6,7]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm0, %xmm5, %xmm5
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm0, %xmm6, %xmm6
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm1, %xmm6, %xmm6
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm1, %xmm14, %xmm14
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm5 = xmm6[0,1,2,3,4,5],xmm5[6,7]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm2, %xmm6, %xmm6
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm2, %xmm14, %xmm14
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3]
; AVX-NEXT:    vmovdqa (%rsp), %xmm14 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm3, %xmm14, %xmm14
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm3, %xmm15, %xmm15
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm6 = xmm14[0,1],xmm6[2,3],xmm14[4,5,6,7]
; AVX-NEXT:    vinsertf128 $1, %xmm5, %ymm0, %ymm5
; AVX-NEXT:    vinsertf128 $1, %xmm6, %ymm0, %ymm6
; AVX-NEXT:    vblendps {{.*#+}} ymm5 = ymm6[0,1,2,3,4,5],ymm5[6,7]
; AVX-NEXT:    vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm5[4,5,6,7]
; AVX-NEXT:    vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm0, %xmm4, %xmm4
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm0, %xmm5, %xmm5
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm1, %xmm5, %xmm5
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm1, %xmm6, %xmm6
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm4 = xmm5[0,1,2,3,4,5],xmm4[6,7]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm2, %xmm5, %xmm5
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm2, %xmm6, %xmm6
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm3, %xmm6, %xmm6
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm3, %xmm14, %xmm14
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm5 = xmm6[0,1],xmm5[2,3],xmm6[4,5,6,7]
; AVX-NEXT:    vpblendw {{.*#+}} xmm4 = xmm5[0,1,2,3],xmm4[4,5,6,7]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm0, %xmm5, %xmm5
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm0, %xmm6, %xmm0
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm5[0],xmm0[1],xmm5[1],xmm0[2],xmm5[2],xmm0[3],xmm5[3]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm1, %xmm5, %xmm5
; AVX-NEXT:    vpshufb %xmm1, %xmm13, %xmm1
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm5[0],xmm1[1],xmm5[1],xmm1[2],xmm5[2],xmm1[3],xmm5[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm0 = xmm1[0,1,2,3,4,5],xmm0[6,7]
; AVX-NEXT:    vpshufb %xmm2, %xmm7, %xmm1
; AVX-NEXT:    vpshufb %xmm2, %xmm8, %xmm2
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX-NEXT:    vpshufb %xmm3, %xmm9, %xmm2
; AVX-NEXT:    vpshufb %xmm3, %xmm10, %xmm3
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3],xmm2[4,5,6,7]
; AVX-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
; AVX-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm1
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
; AVX-NEXT:    vblendps {{.*#+}} ymm5 = ymm4[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vbroadcastss {{.*#+}} xmm3 = [0,0,7,15,0,0,7,15,0,0,7,15,0,0,7,15]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm3, %xmm0, %xmm0
; AVX-NEXT:    vpshufb %xmm3, %xmm11, %xmm1
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
; AVX-NEXT:    vbroadcastss {{.*#+}} xmm2 = [7,15,0,0,7,15,0,0,7,15,0,0,7,15,0,0]
; AVX-NEXT:    vpshufb %xmm2, %xmm12, %xmm1
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm2, %xmm4, %xmm4
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm4[0],xmm1[0],xmm4[1],xmm1[1],xmm4[2],xmm1[2],xmm4[3],xmm1[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm4 = xmm1[0,1,2,3,4,5],xmm0[6,7]
; AVX-NEXT:    vmovd {{.*#+}} xmm1 = [0,0,7,15,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm1, %xmm0, %xmm0
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm1, %xmm6, %xmm6
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm6[0],xmm0[0],xmm6[1],xmm0[1],xmm6[2],xmm0[2],xmm6[3],xmm0[3]
; AVX-NEXT:    vmovd {{.*#+}} xmm0 = [7,15,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm0, %xmm11, %xmm14
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm0, %xmm11, %xmm15
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm6 = xmm14[0,1],xmm6[2,3],xmm14[4,5,6,7]
; AVX-NEXT:    vpblendw {{.*#+}} xmm4 = xmm6[0,1,2,3],xmm4[4,5,6,7]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm3, %xmm6, %xmm6
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm3, %xmm11, %xmm14
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm2, %xmm11, %xmm14
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm2, %xmm11, %xmm15
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm6 = xmm14[0,1,2,3,4,5],xmm6[6,7]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm1, %xmm11, %xmm14
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm1, %xmm11, %xmm15
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3]
; AVX-NEXT:    vmovdqa (%rsp), %xmm11 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm0, %xmm11, %xmm15
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm0, %xmm11, %xmm13
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm13[0],xmm15[0],xmm13[1],xmm15[1],xmm13[2],xmm15[2],xmm13[3],xmm15[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm13 = xmm13[0,1],xmm14[2,3],xmm13[4,5,6,7]
; AVX-NEXT:    vinsertf128 $1, %xmm6, %ymm0, %ymm6
; AVX-NEXT:    vinsertf128 $1, %xmm13, %ymm0, %ymm13
; AVX-NEXT:    vblendps {{.*#+}} ymm6 = ymm13[0,1,2,3,4,5],ymm6[6,7]
; AVX-NEXT:    vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm6[4,5,6,7]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm3, %xmm6, %xmm6
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm3, %xmm11, %xmm13
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm13[0],xmm6[0],xmm13[1],xmm6[1],xmm13[2],xmm6[2],xmm13[3],xmm6[3]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm2, %xmm11, %xmm13
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm2, %xmm11, %xmm14
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm14[0],xmm13[0],xmm14[1],xmm13[1],xmm14[2],xmm13[2],xmm14[3],xmm13[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm6 = xmm13[0,1,2,3,4,5],xmm6[6,7]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm1, %xmm11, %xmm13
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm1, %xmm11, %xmm14
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm14[0],xmm13[0],xmm14[1],xmm13[1],xmm14[2],xmm13[2],xmm14[3],xmm13[3]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm0, %xmm11, %xmm14
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm0, %xmm11, %xmm15
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm13 = xmm14[0,1],xmm13[2,3],xmm14[4,5,6,7]
; AVX-NEXT:    vpblendw {{.*#+}} xmm6 = xmm13[0,1,2,3],xmm6[4,5,6,7]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm3, %xmm11, %xmm13
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm3, %xmm11, %xmm3
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm13[0],xmm3[1],xmm13[1],xmm3[2],xmm13[2],xmm3[3],xmm13[3]
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm2, %xmm11, %xmm12
; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
; AVX-NEXT:    vpshufb %xmm2, %xmm11, %xmm2
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm12[0],xmm2[1],xmm12[1],xmm2[2],xmm12[2],xmm2[3],xmm12[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,5],xmm3[6,7]
; AVX-NEXT:    vpshufb %xmm1, %xmm7, %xmm3
; AVX-NEXT:    vpshufb %xmm1, %xmm8, %xmm1
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1],xmm1[2],xmm3[2],xmm1[3],xmm3[3]
; AVX-NEXT:    vpshufb %xmm0, %xmm9, %xmm3
; AVX-NEXT:    vpshufb %xmm0, %xmm10, %xmm0
; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3]
; AVX-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3],xmm0[4,5,6,7]
; AVX-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm1
; AVX-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm6[0,1,2,3],ymm0[4,5,6,7]
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, (%rsi)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 32(%rsi)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, (%rdx)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 32(%rdx)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, (%rcx)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 32(%rcx)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, (%r8)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 32(%r8)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, (%r9)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 32(%r9)
; AVX-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, (%rax)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 32(%rax)
; AVX-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX-NEXT:    vmovaps %ymm5, (%rax)
; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX-NEXT:    vmovaps %ymm1, 32(%rax)
; AVX-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX-NEXT:    vmovaps %ymm0, (%rax)
; AVX-NEXT:    vmovaps %ymm4, 32(%rax)
; AVX-NEXT:    addq $808, %rsp # imm = 0x328
; AVX-NEXT:    vzeroupper
; AVX-NEXT:    retq
;
; AVX2-LABEL: load_i8_stride8_vf64:
; AVX2:       # %bb.0:
; AVX2-NEXT:    subq $840, %rsp # imm = 0x348
; AVX2-NEXT:    vmovdqa 368(%rdi), %xmm0
; AVX2-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm2 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX2-NEXT:    vpshufb %xmm2, %xmm0, %xmm0
; AVX2-NEXT:    vmovdqa 352(%rdi), %xmm13
; AVX2-NEXT:    vpshufb %xmm2, %xmm13, %xmm1
; AVX2-NEXT:    vmovdqa %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
; AVX2-NEXT:    vmovdqa 336(%rdi), %xmm4
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm3 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX2-NEXT:    vpshufb %xmm3, %xmm4, %xmm1
; AVX2-NEXT:    vmovdqa %xmm4, %xmm15
; AVX2-NEXT:    vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-NEXT:    vmovdqa 320(%rdi), %xmm4
; AVX2-NEXT:    vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-NEXT:    vpshufb %xmm3, %xmm4, %xmm4
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm4[0],xmm1[0],xmm4[1],xmm1[1],xmm4[2],xmm1[2],xmm4[3],xmm1[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm4 = xmm1[0,1,2],xmm0[3]
; AVX2-NEXT:    vmovdqa 304(%rdi), %xmm0
; AVX2-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm9 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX2-NEXT:    vpshufb %xmm9, %xmm0, %xmm1
; AVX2-NEXT:    vmovdqa 288(%rdi), %xmm11
; AVX2-NEXT:    vpshufb %xmm9, %xmm11, %xmm5
; AVX2-NEXT:    vmovdqa %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm1[0],xmm5[1],xmm1[1],xmm5[2],xmm1[2],xmm5[3],xmm1[3]
; AVX2-NEXT:    vmovdqa 272(%rdi), %xmm10
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm1 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX2-NEXT:    vpshufb %xmm1, %xmm10, %xmm6
; AVX2-NEXT:    vmovdqa %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-NEXT:    vmovdqa 256(%rdi), %xmm8
; AVX2-NEXT:    vpshufb %xmm1, %xmm8, %xmm7
; AVX2-NEXT:    vmovdqa %xmm8, %xmm12
; AVX2-NEXT:    vmovdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm5 = xmm6[0],xmm5[1],xmm6[2,3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm4 = xmm5[0,1],xmm4[2,3]
; AVX2-NEXT:    vmovdqa 496(%rdi), %xmm0
; AVX2-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-NEXT:    vpshufb %xmm2, %xmm0, %xmm5
; AVX2-NEXT:    vmovdqa 480(%rdi), %xmm0
; AVX2-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-NEXT:    vpshufb %xmm2, %xmm0, %xmm6
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX2-NEXT:    vmovdqa 464(%rdi), %xmm0
; AVX2-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-NEXT:    vpshufb %xmm3, %xmm0, %xmm6
; AVX2-NEXT:    vmovdqa 448(%rdi), %xmm0
; AVX2-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-NEXT:    vpshufb %xmm3, %xmm0, %xmm7
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
; AVX2-NEXT:    vinserti128 $1, %xmm5, %ymm0, %ymm5
; AVX2-NEXT:    vinserti128 $1, %xmm6, %ymm0, %ymm6
; AVX2-NEXT:    vpblendd {{.*#+}} ymm5 = ymm6[0,1,2,3,4,5,6],ymm5[7]
; AVX2-NEXT:    vmovdqa 432(%rdi), %xmm0
; AVX2-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-NEXT:    vpshufb %xmm9, %xmm0, %xmm6
; AVX2-NEXT:    vmovdqa 416(%rdi), %xmm0
; AVX2-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-NEXT:    vpshufb %xmm9, %xmm0, %xmm7
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
; AVX2-NEXT:    vmovdqa 400(%rdi), %xmm14
; AVX2-NEXT:    vpshufb %xmm1, %xmm14, %xmm7
; AVX2-NEXT:    vmovdqa %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-NEXT:    vmovdqa 384(%rdi), %xmm0
; AVX2-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-NEXT:    vpshufb %xmm1, %xmm0, %xmm8
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm7 = xmm8[0],xmm7[0],xmm8[1],xmm7[1],xmm8[2],xmm7[2],xmm8[3],xmm7[3]
; AVX2-NEXT:    vinserti128 $1, %xmm6, %ymm0, %ymm6
; AVX2-NEXT:    vinserti128 $1, %xmm7, %ymm0, %ymm7
; AVX2-NEXT:    vpblendd {{.*#+}} ymm6 = ymm7[0,1,2,3,4],ymm6[5],ymm7[6,7]
; AVX2-NEXT:    vpblendd {{.*#+}} ymm5 = ymm6[0,1,2,3,4,5],ymm5[6,7]
; AVX2-NEXT:    vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm5[4,5,6,7]
; AVX2-NEXT:    vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovdqa 112(%rdi), %xmm8
; AVX2-NEXT:    vpshufb %xmm2, %xmm8, %xmm4
; AVX2-NEXT:    vmovdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-NEXT:    vmovdqa 96(%rdi), %xmm0
; AVX2-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-NEXT:    vpshufb %xmm2, %xmm0, %xmm5
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX2-NEXT:    vmovdqa 80(%rdi), %xmm0
; AVX2-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-NEXT:    vpshufb %xmm3, %xmm0, %xmm5
; AVX2-NEXT:    vmovdqa 64(%rdi), %xmm0
; AVX2-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-NEXT:    vpshufb %xmm3, %xmm0, %xmm6
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm4 = xmm5[0,1,2],xmm4[3]
; AVX2-NEXT:    vmovdqa 32(%rdi), %xmm0
; AVX2-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-NEXT:    vmovdqa 48(%rdi), %xmm5
; AVX2-NEXT:    vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-NEXT:    vpshufb %xmm9, %xmm5, %xmm5
; AVX2-NEXT:    vpshufb %xmm9, %xmm0, %xmm6
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX2-NEXT:    vmovdqa (%rdi), %xmm0
; AVX2-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-NEXT:    vmovdqa 16(%rdi), %xmm6
; AVX2-NEXT:    vmovdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-NEXT:    vpshufb %xmm1, %xmm6, %xmm6
; AVX2-NEXT:    vpshufb %xmm1, %xmm0, %xmm7
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm5 = xmm6[0],xmm5[1],xmm6[2,3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm4 = xmm5[0,1],xmm4[2,3]
; AVX2-NEXT:    vmovdqa 240(%rdi), %xmm0
; AVX2-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-NEXT:    vpshufb %xmm2, %xmm0, %xmm5
; AVX2-NEXT:    vmovdqa 224(%rdi), %xmm0
; AVX2-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-NEXT:    vpshufb %xmm2, %xmm0, %xmm2
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm5[0],xmm2[1],xmm5[1],xmm2[2],xmm5[2],xmm2[3],xmm5[3]
; AVX2-NEXT:    vmovdqa 208(%rdi), %xmm0
; AVX2-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-NEXT:    vpshufb %xmm3, %xmm0, %xmm5
; AVX2-NEXT:    vmovdqa 192(%rdi), %xmm0
; AVX2-NEXT:    vmovdqa %xmm0, (%rsp) # 16-byte Spill
; AVX2-NEXT:    vpshufb %xmm3, %xmm0, %xmm3
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm5[0],xmm3[1],xmm5[1],xmm3[2],xmm5[2],xmm3[3],xmm5[3]
; AVX2-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX2-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX2-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4,5,6],ymm2[7]
; AVX2-NEXT:    vmovdqa 176(%rdi), %xmm0
; AVX2-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-NEXT:    vpshufb %xmm9, %xmm0, %xmm3
; AVX2-NEXT:    vmovdqa 160(%rdi), %xmm0
; AVX2-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-NEXT:    vpshufb %xmm9, %xmm0, %xmm0
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3]
; AVX2-NEXT:    vmovdqa 144(%rdi), %xmm3
; AVX2-NEXT:    vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-NEXT:    vpshufb %xmm1, %xmm3, %xmm3
; AVX2-NEXT:    vmovdqa 128(%rdi), %xmm5
; AVX2-NEXT:    vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-NEXT:    vpshufb %xmm1, %xmm5, %xmm1
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1],xmm1[2],xmm3[2],xmm1[3],xmm3[3]
; AVX2-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX2-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm1
; AVX2-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3,4],ymm0[5],ymm1[6,7]
; AVX2-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm2[6,7]
; AVX2-NEXT:    vpblendd {{.*#+}} ymm0 = ymm4[0,1,2,3],ymm0[4,5,6,7]
; AVX2-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm2 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm2, %xmm0, %xmm0
; AVX2-NEXT:    vpshufb %xmm2, %xmm13, %xmm1
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm6 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX2-NEXT:    vpshufb %xmm6, %xmm15, %xmm3
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm6, %xmm13, %xmm4
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm4 = xmm3[0,1,2],xmm1[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm3 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm3, %xmm0, %xmm1
; AVX2-NEXT:    vpshufb %xmm3, %xmm11, %xmm5
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm1[0],xmm5[1],xmm1[1],xmm5[2],xmm1[2],xmm5[3],xmm1[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm1 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX2-NEXT:    vpshufb %xmm1, %xmm10, %xmm0
; AVX2-NEXT:    vpshufb %xmm1, %xmm12, %xmm15
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm15[0],xmm0[0],xmm15[1],xmm0[1],xmm15[2],xmm0[2],xmm15[3],xmm0[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm0 = xmm0[0],xmm5[1],xmm0[2,3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm0 = xmm0[0,1],xmm4[2,3]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm2, %xmm9, %xmm4
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm2, %xmm10, %xmm5
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm6, %xmm11, %xmm5
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm6, %xmm12, %xmm15
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm15[0],xmm5[0],xmm15[1],xmm5[1],xmm15[2],xmm5[2],xmm15[3],xmm5[3]
; AVX2-NEXT:    vinserti128 $1, %xmm4, %ymm0, %ymm4
; AVX2-NEXT:    vinserti128 $1, %xmm5, %ymm0, %ymm5
; AVX2-NEXT:    vpblendd {{.*#+}} ymm4 = ymm5[0,1,2,3,4,5,6],ymm4[7]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm3, %xmm5, %xmm5
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm3, %xmm7, %xmm15
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm15[0],xmm5[0],xmm15[1],xmm5[1],xmm15[2],xmm5[2],xmm15[3],xmm5[3]
; AVX2-NEXT:    vpshufb %xmm1, %xmm14, %xmm15
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm1, %xmm7, %xmm14
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm14[0],xmm15[0],xmm14[1],xmm15[1],xmm14[2],xmm15[2],xmm14[3],xmm15[3]
; AVX2-NEXT:    vinserti128 $1, %xmm5, %ymm0, %ymm5
; AVX2-NEXT:    vinserti128 $1, %xmm14, %ymm0, %ymm14
; AVX2-NEXT:    vpblendd {{.*#+}} ymm5 = ymm14[0,1,2,3,4],ymm5[5],ymm14[6,7]
; AVX2-NEXT:    vpblendd {{.*#+}} ymm4 = ymm5[0,1,2,3,4,5],ymm4[6,7]
; AVX2-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm4[4,5,6,7]
; AVX2-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vpshufb %xmm2, %xmm8, %xmm0
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm2, %xmm8, %xmm4
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm4[0],xmm0[0],xmm4[1],xmm0[1],xmm4[2],xmm0[2],xmm4[3],xmm0[3]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm6, %xmm7, %xmm4
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm6, %xmm5, %xmm5
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm0 = xmm4[0,1,2],xmm0[3]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm3, %xmm4, %xmm4
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm3, %xmm5, %xmm5
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm1, %xmm5, %xmm5
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm1, %xmm14, %xmm14
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm14[0],xmm5[0],xmm14[1],xmm5[1],xmm14[2],xmm5[2],xmm14[3],xmm5[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm4 = xmm5[0],xmm4[1],xmm5[2,3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm0 = xmm4[0,1],xmm0[2,3]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm2, %xmm4, %xmm4
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm2, %xmm5, %xmm2
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm6, %xmm4, %xmm4
; AVX2-NEXT:    vmovdqa (%rsp), %xmm5 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm6, %xmm5, %xmm5
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX2-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX2-NEXT:    vinserti128 $1, %xmm4, %ymm0, %ymm4
; AVX2-NEXT:    vpblendd {{.*#+}} ymm2 = ymm4[0,1,2,3,4,5,6],ymm2[7]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm3, %xmm4, %xmm4
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm3, %xmm5, %xmm3
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm1, %xmm4, %xmm4
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm1, %xmm5, %xmm1
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm4[0],xmm1[1],xmm4[1],xmm1[2],xmm4[2],xmm1[3],xmm4[3]
; AVX2-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX2-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm1
; AVX2-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3,4],ymm3[5],ymm1[6,7]
; AVX2-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5],ymm2[6,7]
; AVX2-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX2-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm2 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm2, %xmm0, %xmm0
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm2, %xmm1, %xmm1
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm3 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm3, %xmm1, %xmm1
; AVX2-NEXT:    vpshufb %xmm3, %xmm13, %xmm4
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm4[0],xmm1[0],xmm4[1],xmm1[1],xmm4[2],xmm1[2],xmm4[3],xmm1[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm4 = xmm1[0,1,2],xmm0[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm0 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm0, %xmm13, %xmm1
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm0, %xmm5, %xmm5
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm1[0],xmm5[1],xmm1[1],xmm5[2],xmm1[2],xmm5[3],xmm1[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm1 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm1, %xmm6, %xmm6
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm1, %xmm14, %xmm14
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm5 = xmm6[0],xmm5[1],xmm6[2,3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm4 = xmm5[0,1],xmm4[2,3]
; AVX2-NEXT:    vpshufb %xmm2, %xmm9, %xmm5
; AVX2-NEXT:    vpshufb %xmm2, %xmm10, %xmm6
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX2-NEXT:    vpshufb %xmm3, %xmm11, %xmm6
; AVX2-NEXT:    vpshufb %xmm3, %xmm12, %xmm14
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3]
; AVX2-NEXT:    vinserti128 $1, %xmm5, %ymm0, %ymm5
; AVX2-NEXT:    vinserti128 $1, %xmm6, %ymm0, %ymm6
; AVX2-NEXT:    vpblendd {{.*#+}} ymm5 = ymm6[0,1,2,3,4,5,6],ymm5[7]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm0, %xmm9, %xmm6
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm0, %xmm10, %xmm14
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm1, %xmm11, %xmm14
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm1, %xmm11, %xmm15
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3]
; AVX2-NEXT:    vinserti128 $1, %xmm6, %ymm0, %ymm6
; AVX2-NEXT:    vinserti128 $1, %xmm14, %ymm0, %ymm14
; AVX2-NEXT:    vpblendd {{.*#+}} ymm6 = ymm14[0,1,2,3,4],ymm6[5],ymm14[6,7]
; AVX2-NEXT:    vpblendd {{.*#+}} ymm5 = ymm6[0,1,2,3,4,5],ymm5[6,7]
; AVX2-NEXT:    vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm5[4,5,6,7]
; AVX2-NEXT:    vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm2, %xmm4, %xmm4
; AVX2-NEXT:    vpshufb %xmm2, %xmm8, %xmm5
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX2-NEXT:    vpshufb %xmm3, %xmm7, %xmm5
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm3, %xmm8, %xmm6
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm4 = xmm5[0,1,2],xmm4[3]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm0, %xmm5, %xmm5
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm0, %xmm7, %xmm6
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm1, %xmm6, %xmm6
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm1, %xmm11, %xmm14
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm5 = xmm6[0],xmm5[1],xmm6[2,3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm4 = xmm5[0,1],xmm4[2,3]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm2, %xmm5, %xmm5
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm2, %xmm6, %xmm2
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm5[0],xmm2[1],xmm5[1],xmm2[2],xmm5[2],xmm2[3],xmm5[3]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm3, %xmm5, %xmm5
; AVX2-NEXT:    vmovdqa (%rsp), %xmm6 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm3, %xmm6, %xmm3
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm5[0],xmm3[1],xmm5[1],xmm3[2],xmm5[2],xmm3[3],xmm5[3]
; AVX2-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX2-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX2-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4,5,6],ymm2[7]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm0, %xmm3, %xmm3
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm0, %xmm5, %xmm0
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm1, %xmm3, %xmm3
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm1, %xmm5, %xmm1
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1],xmm1[2],xmm3[2],xmm1[3],xmm3[3]
; AVX2-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX2-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm1
; AVX2-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3,4],ymm0[5],ymm1[6,7]
; AVX2-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm2[6,7]
; AVX2-NEXT:    vpblendd {{.*#+}} ymm0 = ymm4[0,1,2,3],ymm0[4,5,6,7]
; AVX2-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm2 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm2, %xmm11, %xmm0
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm2, %xmm1, %xmm1
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm3 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm3, %xmm1, %xmm1
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm3, %xmm4, %xmm4
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm4[0],xmm1[0],xmm4[1],xmm1[1],xmm4[2],xmm1[2],xmm4[3],xmm1[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm4 = xmm1[0,1,2],xmm0[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm0 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX2-NEXT:    vpshufb %xmm0, %xmm13, %xmm1
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm0, %xmm12, %xmm5
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm1[0],xmm5[1],xmm1[1],xmm5[2],xmm1[2],xmm5[3],xmm1[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm1 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm1, %xmm13, %xmm6
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm1, %xmm14, %xmm14
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm5 = xmm6[0],xmm5[1],xmm6[2,3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm4 = xmm5[0,1],xmm4[2,3]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm2, %xmm5, %xmm5
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm2, %xmm6, %xmm6
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm3, %xmm6, %xmm6
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm3, %xmm14, %xmm14
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3]
; AVX2-NEXT:    vinserti128 $1, %xmm5, %ymm0, %ymm5
; AVX2-NEXT:    vinserti128 $1, %xmm6, %ymm0, %ymm6
; AVX2-NEXT:    vpblendd {{.*#+}} ymm5 = ymm6[0,1,2,3,4,5,6],ymm5[7]
; AVX2-NEXT:    vpshufb %xmm0, %xmm9, %xmm6
; AVX2-NEXT:    vpshufb %xmm0, %xmm10, %xmm14
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm1, %xmm9, %xmm14
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm1, %xmm10, %xmm15
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3]
; AVX2-NEXT:    vinserti128 $1, %xmm6, %ymm0, %ymm6
; AVX2-NEXT:    vinserti128 $1, %xmm14, %ymm0, %ymm14
; AVX2-NEXT:    vpblendd {{.*#+}} ymm6 = ymm14[0,1,2,3,4],ymm6[5],ymm14[6,7]
; AVX2-NEXT:    vpblendd {{.*#+}} ymm5 = ymm6[0,1,2,3,4,5],ymm5[6,7]
; AVX2-NEXT:    vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm5[4,5,6,7]
; AVX2-NEXT:    vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm2, %xmm4, %xmm4
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm2, %xmm5, %xmm5
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm3, %xmm5, %xmm5
; AVX2-NEXT:    vpshufb %xmm3, %xmm8, %xmm6
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm4 = xmm5[0,1,2],xmm4[3]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm0, %xmm8, %xmm5
; AVX2-NEXT:    vpshufb %xmm0, %xmm7, %xmm6
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm1, %xmm6, %xmm6
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm1, %xmm7, %xmm14
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm5 = xmm6[0],xmm5[1],xmm6[2,3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm4 = xmm5[0,1],xmm4[2,3]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm2, %xmm5, %xmm5
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm2, %xmm6, %xmm2
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm5[0],xmm2[1],xmm5[1],xmm2[2],xmm5[2],xmm2[3],xmm5[3]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm3, %xmm5, %xmm5
; AVX2-NEXT:    vmovdqa (%rsp), %xmm6 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm3, %xmm6, %xmm3
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm5[0],xmm3[1],xmm5[1],xmm3[2],xmm5[2],xmm3[3],xmm5[3]
; AVX2-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX2-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX2-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4,5,6],ymm2[7]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm0, %xmm3, %xmm3
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm0, %xmm5, %xmm0
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm1, %xmm3, %xmm3
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm1, %xmm5, %xmm1
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1],xmm1[2],xmm3[2],xmm1[3],xmm3[3]
; AVX2-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX2-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm1
; AVX2-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3,4],ymm0[5],ymm1[6,7]
; AVX2-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm2[6,7]
; AVX2-NEXT:    vpblendd {{.*#+}} ymm0 = ymm4[0,1,2,3],ymm0[4,5,6,7]
; AVX2-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm2 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX2-NEXT:    vpshufb %xmm2, %xmm11, %xmm0
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm2, %xmm1, %xmm1
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm3 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm3, %xmm1, %xmm1
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm3, %xmm4, %xmm4
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm4[0],xmm1[0],xmm4[1],xmm1[1],xmm4[2],xmm1[2],xmm4[3],xmm1[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm4 = xmm1[0,1,2],xmm0[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm0 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm0, %xmm1, %xmm1
; AVX2-NEXT:    vpshufb %xmm0, %xmm12, %xmm5
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm1[0],xmm5[1],xmm1[1],xmm5[2],xmm1[2],xmm5[3],xmm1[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm1 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX2-NEXT:    vpshufb %xmm1, %xmm13, %xmm6
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm1, %xmm7, %xmm14
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm5 = xmm6[0],xmm5[1],xmm6[2,3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm4 = xmm5[0,1],xmm4[2,3]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm2, %xmm5, %xmm5
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm2, %xmm6, %xmm6
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm3, %xmm6, %xmm6
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm3, %xmm7, %xmm14
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3]
; AVX2-NEXT:    vinserti128 $1, %xmm5, %ymm0, %ymm5
; AVX2-NEXT:    vinserti128 $1, %xmm6, %ymm0, %ymm6
; AVX2-NEXT:    vpblendd {{.*#+}} ymm5 = ymm6[0,1,2,3,4,5,6],ymm5[7]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm0, %xmm6, %xmm6
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm0, %xmm7, %xmm14
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3]
; AVX2-NEXT:    vpshufb %xmm1, %xmm9, %xmm14
; AVX2-NEXT:    vpshufb %xmm1, %xmm10, %xmm15
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3]
; AVX2-NEXT:    vinserti128 $1, %xmm6, %ymm0, %ymm6
; AVX2-NEXT:    vinserti128 $1, %xmm14, %ymm0, %ymm14
; AVX2-NEXT:    vpblendd {{.*#+}} ymm6 = ymm14[0,1,2,3,4],ymm6[5],ymm14[6,7]
; AVX2-NEXT:    vpblendd {{.*#+}} ymm5 = ymm6[0,1,2,3,4,5],ymm5[6,7]
; AVX2-NEXT:    vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm5[4,5,6,7]
; AVX2-NEXT:    vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm2, %xmm7, %xmm4
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm2, %xmm5, %xmm5
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm3, %xmm5, %xmm5
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm3, %xmm6, %xmm6
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm4 = xmm5[0,1,2],xmm4[3]
; AVX2-NEXT:    vpshufb %xmm0, %xmm8, %xmm5
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm0, %xmm6, %xmm6
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm1, %xmm9, %xmm6
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm1, %xmm8, %xmm14
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm5 = xmm6[0],xmm5[1],xmm6[2,3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm4 = xmm5[0,1],xmm4[2,3]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm2, %xmm10, %xmm5
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm2, %xmm11, %xmm2
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm5[0],xmm2[1],xmm5[1],xmm2[2],xmm5[2],xmm2[3],xmm5[3]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm3, %xmm12, %xmm5
; AVX2-NEXT:    vmovdqa (%rsp), %xmm13 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm3, %xmm13, %xmm3
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm5[0],xmm3[1],xmm5[1],xmm3[2],xmm5[2],xmm3[3],xmm5[3]
; AVX2-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX2-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX2-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4,5,6],ymm2[7]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm0, %xmm3, %xmm3
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm0, %xmm5, %xmm0
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm1, %xmm3, %xmm3
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm1, %xmm5, %xmm1
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1],xmm1[2],xmm3[2],xmm1[3],xmm3[3]
; AVX2-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX2-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm1
; AVX2-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3,4],ymm0[5],ymm1[6,7]
; AVX2-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm2[6,7]
; AVX2-NEXT:    vpblendd {{.*#+}} ymm0 = ymm4[0,1,2,3],ymm0[4,5,6,7]
; AVX2-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm2 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm2, %xmm0, %xmm0
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm2, %xmm1, %xmm1
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm3 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm3, %xmm1, %xmm1
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm3, %xmm4, %xmm4
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm4[0],xmm1[0],xmm4[1],xmm1[1],xmm4[2],xmm1[2],xmm4[3],xmm1[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm4 = xmm1[0,1,2],xmm0[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm0 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm0, %xmm1, %xmm1
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm0, %xmm5, %xmm5
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm1[0],xmm5[1],xmm1[1],xmm5[2],xmm1[2],xmm5[3],xmm1[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm1 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm1, %xmm6, %xmm6
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm1, %xmm14, %xmm14
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm5 = xmm6[0],xmm5[1],xmm6[2,3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm4 = xmm5[0,1],xmm4[2,3]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm2, %xmm5, %xmm5
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm2, %xmm6, %xmm6
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm3, %xmm6, %xmm6
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm3, %xmm14, %xmm14
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3]
; AVX2-NEXT:    vinserti128 $1, %xmm5, %ymm0, %ymm5
; AVX2-NEXT:    vinserti128 $1, %xmm6, %ymm0, %ymm6
; AVX2-NEXT:    vpblendd {{.*#+}} ymm5 = ymm6[0,1,2,3,4,5,6],ymm5[7]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm0, %xmm6, %xmm6
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm0, %xmm14, %xmm14
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm1, %xmm14, %xmm14
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm1, %xmm15, %xmm15
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3]
; AVX2-NEXT:    vinserti128 $1, %xmm6, %ymm0, %ymm6
; AVX2-NEXT:    vinserti128 $1, %xmm14, %ymm0, %ymm14
; AVX2-NEXT:    vpblendd {{.*#+}} ymm6 = ymm14[0,1,2,3,4],ymm6[5],ymm14[6,7]
; AVX2-NEXT:    vpblendd {{.*#+}} ymm5 = ymm6[0,1,2,3,4,5],ymm5[6,7]
; AVX2-NEXT:    vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm5[4,5,6,7]
; AVX2-NEXT:    vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vpshufb %xmm2, %xmm7, %xmm4
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm2, %xmm5, %xmm5
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm3, %xmm5, %xmm5
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm3, %xmm6, %xmm6
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm4 = xmm5[0,1,2],xmm4[3]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm0, %xmm5, %xmm5
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm0, %xmm6, %xmm6
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX2-NEXT:    vpshufb %xmm1, %xmm9, %xmm6
; AVX2-NEXT:    vpshufb %xmm1, %xmm8, %xmm14
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm5 = xmm6[0],xmm5[1],xmm6[2,3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm4 = xmm5[0,1],xmm4[2,3]
; AVX2-NEXT:    vpshufb %xmm2, %xmm10, %xmm5
; AVX2-NEXT:    vpshufb %xmm2, %xmm11, %xmm2
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm5[0],xmm2[1],xmm5[1],xmm2[2],xmm5[2],xmm2[3],xmm5[3]
; AVX2-NEXT:    vpshufb %xmm3, %xmm12, %xmm5
; AVX2-NEXT:    vpshufb %xmm3, %xmm13, %xmm3
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm5[0],xmm3[1],xmm5[1],xmm3[2],xmm5[2],xmm3[3],xmm5[3]
; AVX2-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX2-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX2-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4,5,6],ymm2[7]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm0, %xmm13, %xmm3
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm0, %xmm7, %xmm0
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm1, %xmm8, %xmm3
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm1, %xmm9, %xmm1
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1],xmm1[2],xmm3[2],xmm1[3],xmm3[3]
; AVX2-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX2-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm1
; AVX2-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3,4],ymm0[5],ymm1[6,7]
; AVX2-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm2[6,7]
; AVX2-NEXT:    vpblendd {{.*#+}} ymm0 = ymm4[0,1,2,3],ymm0[4,5,6,7]
; AVX2-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm2 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm2, %xmm0, %xmm0
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm2, %xmm10, %xmm1
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm3 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm3, %xmm11, %xmm1
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm3, %xmm4, %xmm4
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm4[0],xmm1[0],xmm4[1],xmm1[1],xmm4[2],xmm1[2],xmm4[3],xmm1[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm4 = xmm1[0,1,2],xmm0[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm0 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm0, %xmm1, %xmm1
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm0, %xmm5, %xmm5
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm1[0],xmm5[1],xmm1[1],xmm5[2],xmm1[2],xmm5[3],xmm1[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm1 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm1, %xmm6, %xmm6
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm1, %xmm12, %xmm14
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm5 = xmm6[0],xmm5[1],xmm6[2,3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm4 = xmm5[0,1],xmm4[2,3]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm2, %xmm5, %xmm5
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm2, %xmm6, %xmm6
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm3, %xmm6, %xmm6
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm3, %xmm14, %xmm14
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3]
; AVX2-NEXT:    vinserti128 $1, %xmm5, %ymm0, %ymm5
; AVX2-NEXT:    vinserti128 $1, %xmm6, %ymm0, %ymm6
; AVX2-NEXT:    vpblendd {{.*#+}} ymm5 = ymm6[0,1,2,3,4,5,6],ymm5[7]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm0, %xmm6, %xmm6
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm0, %xmm14, %xmm14
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm1, %xmm14, %xmm14
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm1, %xmm15, %xmm15
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3]
; AVX2-NEXT:    vinserti128 $1, %xmm6, %ymm0, %ymm6
; AVX2-NEXT:    vinserti128 $1, %xmm14, %ymm0, %ymm14
; AVX2-NEXT:    vpblendd {{.*#+}} ymm6 = ymm14[0,1,2,3,4],ymm6[5],ymm14[6,7]
; AVX2-NEXT:    vpblendd {{.*#+}} ymm5 = ymm6[0,1,2,3,4,5],ymm5[6,7]
; AVX2-NEXT:    vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm5[4,5,6,7]
; AVX2-NEXT:    vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm2, %xmm4, %xmm4
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm2, %xmm5, %xmm5
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm3, %xmm5, %xmm5
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm3, %xmm6, %xmm6
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm4 = xmm5[0,1,2],xmm4[3]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm0, %xmm5, %xmm5
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm0, %xmm6, %xmm6
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm1, %xmm6, %xmm6
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm1, %xmm14, %xmm14
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm5 = xmm6[0],xmm5[1],xmm6[2,3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm4 = xmm5[0,1],xmm4[2,3]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm2, %xmm5, %xmm5
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm2, %xmm6, %xmm2
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm5[0],xmm2[1],xmm5[1],xmm2[2],xmm5[2],xmm2[3],xmm5[3]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm3, %xmm5, %xmm5
; AVX2-NEXT:    vmovdqa (%rsp), %xmm6 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm3, %xmm6, %xmm3
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm5[0],xmm3[1],xmm5[1],xmm3[2],xmm5[2],xmm3[3],xmm5[3]
; AVX2-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX2-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX2-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4,5,6],ymm2[7]
; AVX2-NEXT:    vpshufb %xmm0, %xmm13, %xmm3
; AVX2-NEXT:    vmovdqa %xmm13, %xmm5
; AVX2-NEXT:    vpshufb %xmm0, %xmm7, %xmm0
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3]
; AVX2-NEXT:    vpshufb %xmm1, %xmm8, %xmm3
; AVX2-NEXT:    vpshufb %xmm1, %xmm9, %xmm1
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1],xmm1[2],xmm3[2],xmm1[3],xmm3[3]
; AVX2-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX2-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm1
; AVX2-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3,4],ymm0[5],ymm1[6,7]
; AVX2-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm2[6,7]
; AVX2-NEXT:    vpblendd {{.*#+}} ymm0 = ymm4[0,1,2,3],ymm0[4,5,6,7]
; AVX2-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm3 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm3, %xmm0, %xmm0
; AVX2-NEXT:    vpshufb %xmm3, %xmm10, %xmm1
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm0 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX2-NEXT:    vpshufb %xmm0, %xmm11, %xmm2
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm0, %xmm4, %xmm4
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm4[0],xmm2[0],xmm4[1],xmm2[1],xmm4[2],xmm2[2],xmm4[3],xmm2[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm4 = xmm2[0,1,2],xmm1[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm1 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm1, %xmm2, %xmm2
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm1, %xmm6, %xmm6
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm6[0],xmm2[0],xmm6[1],xmm2[1],xmm6[2],xmm2[2],xmm6[3],xmm2[3]
; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm2 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm2, %xmm10, %xmm14
; AVX2-NEXT:    vpshufb %xmm2, %xmm12, %xmm15
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm6 = xmm14[0],xmm6[1],xmm14[2,3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm4 = xmm6[0,1],xmm4[2,3]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm3, %xmm6, %xmm6
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm3, %xmm10, %xmm14
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm0, %xmm10, %xmm14
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm0, %xmm10, %xmm15
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3]
; AVX2-NEXT:    vinserti128 $1, %xmm6, %ymm0, %ymm6
; AVX2-NEXT:    vinserti128 $1, %xmm14, %ymm0, %ymm14
; AVX2-NEXT:    vpblendd {{.*#+}} ymm6 = ymm14[0,1,2,3,4,5,6],ymm6[7]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm1, %xmm10, %xmm14
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm1, %xmm10, %xmm15
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm2, %xmm10, %xmm15
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm2, %xmm10, %xmm13
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm13[0],xmm15[0],xmm13[1],xmm15[1],xmm13[2],xmm15[2],xmm13[3],xmm15[3]
; AVX2-NEXT:    vinserti128 $1, %xmm14, %ymm0, %ymm14
; AVX2-NEXT:    vinserti128 $1, %xmm13, %ymm0, %ymm13
; AVX2-NEXT:    vpblendd {{.*#+}} ymm13 = ymm13[0,1,2,3,4],ymm14[5],ymm13[6,7]
; AVX2-NEXT:    vpblendd {{.*#+}} ymm6 = ymm13[0,1,2,3,4,5],ymm6[6,7]
; AVX2-NEXT:    vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm6[4,5,6,7]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm3, %xmm6, %xmm6
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm3, %xmm10, %xmm13
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm13[0],xmm6[0],xmm13[1],xmm6[1],xmm13[2],xmm6[2],xmm13[3],xmm6[3]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm0, %xmm10, %xmm13
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm0, %xmm10, %xmm14
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm14[0],xmm13[0],xmm14[1],xmm13[1],xmm14[2],xmm13[2],xmm14[3],xmm13[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm6 = xmm13[0,1,2],xmm6[3]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm1, %xmm10, %xmm13
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm1, %xmm10, %xmm14
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm14[0],xmm13[0],xmm14[1],xmm13[1],xmm14[2],xmm13[2],xmm14[3],xmm13[3]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm2, %xmm10, %xmm14
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm2, %xmm10, %xmm15
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm13 = xmm14[0],xmm13[1],xmm14[2,3]
; AVX2-NEXT:    vpblendd {{.*#+}} xmm6 = xmm13[0,1],xmm6[2,3]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm3, %xmm10, %xmm13
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm3, %xmm10, %xmm3
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm13[0],xmm3[1],xmm13[1],xmm3[2],xmm13[2],xmm3[3],xmm13[3]
; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm0, %xmm10, %xmm12
; AVX2-NEXT:    vmovdqa (%rsp), %xmm10 # 16-byte Reload
; AVX2-NEXT:    vpshufb %xmm0, %xmm10, %xmm0
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm12[0],xmm0[1],xmm12[1],xmm0[2],xmm12[2],xmm0[3],xmm12[3]
; AVX2-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX2-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX2-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5,6],ymm3[7]
; AVX2-NEXT:    vpshufb %xmm1, %xmm5, %xmm3
; AVX2-NEXT:    vpshufb %xmm1, %xmm7, %xmm1
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1],xmm1[2],xmm3[2],xmm1[3],xmm3[3]
; AVX2-NEXT:    vpshufb %xmm2, %xmm8, %xmm3
; AVX2-NEXT:    vpshufb %xmm2, %xmm9, %xmm2
; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
; AVX2-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm1
; AVX2-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX2-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3,4],ymm1[5],ymm2[6,7]
; AVX2-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
; AVX2-NEXT:    vpblendd {{.*#+}} ymm0 = ymm6[0,1,2,3],ymm0[4,5,6,7]
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, (%rsi)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 32(%rsi)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, (%rdx)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 32(%rdx)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, (%rcx)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 32(%rcx)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, (%r8)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 32(%r8)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, (%r9)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 32(%r9)
; AVX2-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, (%rax)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 32(%rax)
; AVX2-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, (%rax)
; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-NEXT:    vmovaps %ymm1, 32(%rax)
; AVX2-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-NEXT:    vmovdqa %ymm0, (%rax)
; AVX2-NEXT:    vmovdqa %ymm4, 32(%rax)
; AVX2-NEXT:    addq $840, %rsp # imm = 0x348
; AVX2-NEXT:    vzeroupper
; AVX2-NEXT:    retq
;
; AVX2-FP-LABEL: load_i8_stride8_vf64:
; AVX2-FP:       # %bb.0:
; AVX2-FP-NEXT:    subq $840, %rsp # imm = 0x348
; AVX2-FP-NEXT:    vmovdqa 368(%rdi), %xmm0
; AVX2-FP-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm2 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm0, %xmm0
; AVX2-FP-NEXT:    vmovdqa 352(%rdi), %xmm13
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm13, %xmm1
; AVX2-FP-NEXT:    vmovdqa %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
; AVX2-FP-NEXT:    vmovdqa 336(%rdi), %xmm4
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm3 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm4, %xmm1
; AVX2-FP-NEXT:    vmovdqa %xmm4, %xmm15
; AVX2-FP-NEXT:    vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FP-NEXT:    vmovdqa 320(%rdi), %xmm4
; AVX2-FP-NEXT:    vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm4, %xmm4
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm4[0],xmm1[0],xmm4[1],xmm1[1],xmm4[2],xmm1[2],xmm4[3],xmm1[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm4 = xmm1[0,1,2],xmm0[3]
; AVX2-FP-NEXT:    vmovdqa 304(%rdi), %xmm0
; AVX2-FP-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm9 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX2-FP-NEXT:    vpshufb %xmm9, %xmm0, %xmm1
; AVX2-FP-NEXT:    vmovdqa 288(%rdi), %xmm11
; AVX2-FP-NEXT:    vpshufb %xmm9, %xmm11, %xmm5
; AVX2-FP-NEXT:    vmovdqa %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm1[0],xmm5[1],xmm1[1],xmm5[2],xmm1[2],xmm5[3],xmm1[3]
; AVX2-FP-NEXT:    vmovdqa 272(%rdi), %xmm10
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm1 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX2-FP-NEXT:    vpshufb %xmm1, %xmm10, %xmm6
; AVX2-FP-NEXT:    vmovdqa %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FP-NEXT:    vmovdqa 256(%rdi), %xmm8
; AVX2-FP-NEXT:    vpshufb %xmm1, %xmm8, %xmm7
; AVX2-FP-NEXT:    vmovdqa %xmm8, %xmm12
; AVX2-FP-NEXT:    vmovdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm5 = xmm6[0],xmm5[1],xmm6[2,3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm4 = xmm5[0,1],xmm4[2,3]
; AVX2-FP-NEXT:    vmovdqa 496(%rdi), %xmm0
; AVX2-FP-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm0, %xmm5
; AVX2-FP-NEXT:    vmovdqa 480(%rdi), %xmm0
; AVX2-FP-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm0, %xmm6
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX2-FP-NEXT:    vmovdqa 464(%rdi), %xmm0
; AVX2-FP-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm0, %xmm6
; AVX2-FP-NEXT:    vmovdqa 448(%rdi), %xmm0
; AVX2-FP-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm0, %xmm7
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
; AVX2-FP-NEXT:    vinserti128 $1, %xmm5, %ymm0, %ymm5
; AVX2-FP-NEXT:    vinserti128 $1, %xmm6, %ymm0, %ymm6
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm5 = ymm6[0,1,2,3,4,5,6],ymm5[7]
; AVX2-FP-NEXT:    vmovdqa 432(%rdi), %xmm0
; AVX2-FP-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FP-NEXT:    vpshufb %xmm9, %xmm0, %xmm6
; AVX2-FP-NEXT:    vmovdqa 416(%rdi), %xmm0
; AVX2-FP-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FP-NEXT:    vpshufb %xmm9, %xmm0, %xmm7
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
; AVX2-FP-NEXT:    vmovdqa 400(%rdi), %xmm14
; AVX2-FP-NEXT:    vpshufb %xmm1, %xmm14, %xmm7
; AVX2-FP-NEXT:    vmovdqa %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FP-NEXT:    vmovdqa 384(%rdi), %xmm0
; AVX2-FP-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FP-NEXT:    vpshufb %xmm1, %xmm0, %xmm8
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm7 = xmm8[0],xmm7[0],xmm8[1],xmm7[1],xmm8[2],xmm7[2],xmm8[3],xmm7[3]
; AVX2-FP-NEXT:    vinserti128 $1, %xmm6, %ymm0, %ymm6
; AVX2-FP-NEXT:    vinserti128 $1, %xmm7, %ymm0, %ymm7
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm6 = ymm7[0,1,2,3,4],ymm6[5],ymm7[6,7]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm5 = ymm6[0,1,2,3,4,5],ymm5[6,7]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm5[4,5,6,7]
; AVX2-FP-NEXT:    vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovdqa 112(%rdi), %xmm8
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm8, %xmm4
; AVX2-FP-NEXT:    vmovdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FP-NEXT:    vmovdqa 96(%rdi), %xmm0
; AVX2-FP-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm0, %xmm5
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX2-FP-NEXT:    vmovdqa 80(%rdi), %xmm0
; AVX2-FP-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm0, %xmm5
; AVX2-FP-NEXT:    vmovdqa 64(%rdi), %xmm0
; AVX2-FP-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm0, %xmm6
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm4 = xmm5[0,1,2],xmm4[3]
; AVX2-FP-NEXT:    vmovdqa 32(%rdi), %xmm0
; AVX2-FP-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FP-NEXT:    vmovdqa 48(%rdi), %xmm5
; AVX2-FP-NEXT:    vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FP-NEXT:    vpshufb %xmm9, %xmm5, %xmm5
; AVX2-FP-NEXT:    vpshufb %xmm9, %xmm0, %xmm6
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX2-FP-NEXT:    vmovdqa (%rdi), %xmm0
; AVX2-FP-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FP-NEXT:    vmovdqa 16(%rdi), %xmm6
; AVX2-FP-NEXT:    vmovdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FP-NEXT:    vpshufb %xmm1, %xmm6, %xmm6
; AVX2-FP-NEXT:    vpshufb %xmm1, %xmm0, %xmm7
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm5 = xmm6[0],xmm5[1],xmm6[2,3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm4 = xmm5[0,1],xmm4[2,3]
; AVX2-FP-NEXT:    vmovdqa 240(%rdi), %xmm0
; AVX2-FP-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm0, %xmm5
; AVX2-FP-NEXT:    vmovdqa 224(%rdi), %xmm0
; AVX2-FP-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm0, %xmm2
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm5[0],xmm2[1],xmm5[1],xmm2[2],xmm5[2],xmm2[3],xmm5[3]
; AVX2-FP-NEXT:    vmovdqa 208(%rdi), %xmm0
; AVX2-FP-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm0, %xmm5
; AVX2-FP-NEXT:    vmovdqa 192(%rdi), %xmm0
; AVX2-FP-NEXT:    vmovdqa %xmm0, (%rsp) # 16-byte Spill
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm0, %xmm3
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm5[0],xmm3[1],xmm5[1],xmm3[2],xmm5[2],xmm3[3],xmm5[3]
; AVX2-FP-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX2-FP-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4,5,6],ymm2[7]
; AVX2-FP-NEXT:    vmovdqa 176(%rdi), %xmm0
; AVX2-FP-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FP-NEXT:    vpshufb %xmm9, %xmm0, %xmm3
; AVX2-FP-NEXT:    vmovdqa 160(%rdi), %xmm0
; AVX2-FP-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FP-NEXT:    vpshufb %xmm9, %xmm0, %xmm0
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3]
; AVX2-FP-NEXT:    vmovdqa 144(%rdi), %xmm3
; AVX2-FP-NEXT:    vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FP-NEXT:    vpshufb %xmm1, %xmm3, %xmm3
; AVX2-FP-NEXT:    vmovdqa 128(%rdi), %xmm5
; AVX2-FP-NEXT:    vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FP-NEXT:    vpshufb %xmm1, %xmm5, %xmm1
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1],xmm1[2],xmm3[2],xmm1[3],xmm3[3]
; AVX2-FP-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX2-FP-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm1
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3,4],ymm0[5],ymm1[6,7]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm2[6,7]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm4[0,1,2,3],ymm0[4,5,6,7]
; AVX2-FP-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm2 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm0, %xmm0
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm13, %xmm1
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm6 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX2-FP-NEXT:    vpshufb %xmm6, %xmm15, %xmm3
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm6, %xmm13, %xmm4
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm4 = xmm3[0,1,2],xmm1[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm3 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm0, %xmm1
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm11, %xmm5
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm1[0],xmm5[1],xmm1[1],xmm5[2],xmm1[2],xmm5[3],xmm1[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm1 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX2-FP-NEXT:    vpshufb %xmm1, %xmm10, %xmm0
; AVX2-FP-NEXT:    vpshufb %xmm1, %xmm12, %xmm15
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm15[0],xmm0[0],xmm15[1],xmm0[1],xmm15[2],xmm0[2],xmm15[3],xmm0[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm0[0],xmm5[1],xmm0[2,3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm0[0,1],xmm4[2,3]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm9, %xmm4
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm10, %xmm5
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm6, %xmm11, %xmm5
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm6, %xmm12, %xmm15
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm15[0],xmm5[0],xmm15[1],xmm5[1],xmm15[2],xmm5[2],xmm15[3],xmm5[3]
; AVX2-FP-NEXT:    vinserti128 $1, %xmm4, %ymm0, %ymm4
; AVX2-FP-NEXT:    vinserti128 $1, %xmm5, %ymm0, %ymm5
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm4 = ymm5[0,1,2,3,4,5,6],ymm4[7]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm5, %xmm5
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm7, %xmm15
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm15[0],xmm5[0],xmm15[1],xmm5[1],xmm15[2],xmm5[2],xmm15[3],xmm5[3]
; AVX2-FP-NEXT:    vpshufb %xmm1, %xmm14, %xmm15
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm1, %xmm7, %xmm14
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm14[0],xmm15[0],xmm14[1],xmm15[1],xmm14[2],xmm15[2],xmm14[3],xmm15[3]
; AVX2-FP-NEXT:    vinserti128 $1, %xmm5, %ymm0, %ymm5
; AVX2-FP-NEXT:    vinserti128 $1, %xmm14, %ymm0, %ymm14
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm5 = ymm14[0,1,2,3,4],ymm5[5],ymm14[6,7]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm4 = ymm5[0,1,2,3,4,5],ymm4[6,7]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm4[4,5,6,7]
; AVX2-FP-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm8, %xmm0
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm8, %xmm4
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm4[0],xmm0[0],xmm4[1],xmm0[1],xmm4[2],xmm0[2],xmm4[3],xmm0[3]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm6, %xmm7, %xmm4
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm6, %xmm5, %xmm5
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm4[0,1,2],xmm0[3]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm4, %xmm4
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm5, %xmm5
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm1, %xmm5, %xmm5
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm1, %xmm14, %xmm14
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm14[0],xmm5[0],xmm14[1],xmm5[1],xmm14[2],xmm5[2],xmm14[3],xmm5[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm4 = xmm5[0],xmm4[1],xmm5[2,3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm4[0,1],xmm0[2,3]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm4, %xmm4
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm5, %xmm2
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm6, %xmm4, %xmm4
; AVX2-FP-NEXT:    vmovdqa (%rsp), %xmm5 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm6, %xmm5, %xmm5
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX2-FP-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX2-FP-NEXT:    vinserti128 $1, %xmm4, %ymm0, %ymm4
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm2 = ymm4[0,1,2,3,4,5,6],ymm2[7]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm4, %xmm4
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm5, %xmm3
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm1, %xmm4, %xmm4
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm1, %xmm5, %xmm1
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm4[0],xmm1[1],xmm4[1],xmm1[2],xmm4[2],xmm1[3],xmm4[3]
; AVX2-FP-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX2-FP-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm1
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3,4],ymm3[5],ymm1[6,7]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5],ymm2[6,7]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX2-FP-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm2 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm0, %xmm0
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm1, %xmm1
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm3 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm1, %xmm1
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm13, %xmm4
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm4[0],xmm1[0],xmm4[1],xmm1[1],xmm4[2],xmm1[2],xmm4[3],xmm1[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm4 = xmm1[0,1,2],xmm0[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm0 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm13, %xmm1
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm5, %xmm5
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm1[0],xmm5[1],xmm1[1],xmm5[2],xmm1[2],xmm5[3],xmm1[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm1 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm1, %xmm6, %xmm6
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm1, %xmm14, %xmm14
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm5 = xmm6[0],xmm5[1],xmm6[2,3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm4 = xmm5[0,1],xmm4[2,3]
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm9, %xmm5
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm10, %xmm6
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm11, %xmm6
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm12, %xmm14
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3]
; AVX2-FP-NEXT:    vinserti128 $1, %xmm5, %ymm0, %ymm5
; AVX2-FP-NEXT:    vinserti128 $1, %xmm6, %ymm0, %ymm6
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm5 = ymm6[0,1,2,3,4,5,6],ymm5[7]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm9, %xmm6
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm10, %xmm14
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm1, %xmm11, %xmm14
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm1, %xmm11, %xmm15
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3]
; AVX2-FP-NEXT:    vinserti128 $1, %xmm6, %ymm0, %ymm6
; AVX2-FP-NEXT:    vinserti128 $1, %xmm14, %ymm0, %ymm14
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm6 = ymm14[0,1,2,3,4],ymm6[5],ymm14[6,7]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm5 = ymm6[0,1,2,3,4,5],ymm5[6,7]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm5[4,5,6,7]
; AVX2-FP-NEXT:    vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm4, %xmm4
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm8, %xmm5
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm7, %xmm5
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm8, %xmm6
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm4 = xmm5[0,1,2],xmm4[3]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm5, %xmm5
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm7, %xmm6
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm1, %xmm6, %xmm6
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm1, %xmm11, %xmm14
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm5 = xmm6[0],xmm5[1],xmm6[2,3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm4 = xmm5[0,1],xmm4[2,3]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm5, %xmm5
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm6, %xmm2
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm5[0],xmm2[1],xmm5[1],xmm2[2],xmm5[2],xmm2[3],xmm5[3]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm5, %xmm5
; AVX2-FP-NEXT:    vmovdqa (%rsp), %xmm6 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm6, %xmm3
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm5[0],xmm3[1],xmm5[1],xmm3[2],xmm5[2],xmm3[3],xmm5[3]
; AVX2-FP-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX2-FP-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4,5,6],ymm2[7]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm3, %xmm3
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm5, %xmm0
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm1, %xmm3, %xmm3
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm1, %xmm5, %xmm1
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1],xmm1[2],xmm3[2],xmm1[3],xmm3[3]
; AVX2-FP-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX2-FP-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm1
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3,4],ymm0[5],ymm1[6,7]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm2[6,7]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm4[0,1,2,3],ymm0[4,5,6,7]
; AVX2-FP-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm2 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm11, %xmm0
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm1, %xmm1
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm3 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm1, %xmm1
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm4, %xmm4
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm4[0],xmm1[0],xmm4[1],xmm1[1],xmm4[2],xmm1[2],xmm4[3],xmm1[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm4 = xmm1[0,1,2],xmm0[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm0 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm13, %xmm1
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm12, %xmm5
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm1[0],xmm5[1],xmm1[1],xmm5[2],xmm1[2],xmm5[3],xmm1[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm1 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm1, %xmm13, %xmm6
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm1, %xmm14, %xmm14
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm5 = xmm6[0],xmm5[1],xmm6[2,3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm4 = xmm5[0,1],xmm4[2,3]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm5, %xmm5
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm6, %xmm6
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm6, %xmm6
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm14, %xmm14
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3]
; AVX2-FP-NEXT:    vinserti128 $1, %xmm5, %ymm0, %ymm5
; AVX2-FP-NEXT:    vinserti128 $1, %xmm6, %ymm0, %ymm6
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm5 = ymm6[0,1,2,3,4,5,6],ymm5[7]
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm9, %xmm6
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm10, %xmm14
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm1, %xmm9, %xmm14
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm1, %xmm10, %xmm15
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3]
; AVX2-FP-NEXT:    vinserti128 $1, %xmm6, %ymm0, %ymm6
; AVX2-FP-NEXT:    vinserti128 $1, %xmm14, %ymm0, %ymm14
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm6 = ymm14[0,1,2,3,4],ymm6[5],ymm14[6,7]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm5 = ymm6[0,1,2,3,4,5],ymm5[6,7]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm5[4,5,6,7]
; AVX2-FP-NEXT:    vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm4, %xmm4
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm5, %xmm5
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm5, %xmm5
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm8, %xmm6
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm4 = xmm5[0,1,2],xmm4[3]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm8, %xmm5
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm7, %xmm6
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm1, %xmm6, %xmm6
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm1, %xmm7, %xmm14
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm5 = xmm6[0],xmm5[1],xmm6[2,3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm4 = xmm5[0,1],xmm4[2,3]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm5, %xmm5
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm6, %xmm2
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm5[0],xmm2[1],xmm5[1],xmm2[2],xmm5[2],xmm2[3],xmm5[3]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm5, %xmm5
; AVX2-FP-NEXT:    vmovdqa (%rsp), %xmm6 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm6, %xmm3
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm5[0],xmm3[1],xmm5[1],xmm3[2],xmm5[2],xmm3[3],xmm5[3]
; AVX2-FP-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX2-FP-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4,5,6],ymm2[7]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm3, %xmm3
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm5, %xmm0
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm1, %xmm3, %xmm3
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm1, %xmm5, %xmm1
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1],xmm1[2],xmm3[2],xmm1[3],xmm3[3]
; AVX2-FP-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX2-FP-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm1
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3,4],ymm0[5],ymm1[6,7]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm2[6,7]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm4[0,1,2,3],ymm0[4,5,6,7]
; AVX2-FP-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm2 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm11, %xmm0
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm1, %xmm1
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm3 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm1, %xmm1
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm4, %xmm4
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm4[0],xmm1[0],xmm4[1],xmm1[1],xmm4[2],xmm1[2],xmm4[3],xmm1[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm4 = xmm1[0,1,2],xmm0[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm0 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm1, %xmm1
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm12, %xmm5
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm1[0],xmm5[1],xmm1[1],xmm5[2],xmm1[2],xmm5[3],xmm1[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm1 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX2-FP-NEXT:    vpshufb %xmm1, %xmm13, %xmm6
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm1, %xmm7, %xmm14
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm5 = xmm6[0],xmm5[1],xmm6[2,3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm4 = xmm5[0,1],xmm4[2,3]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm5, %xmm5
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm6, %xmm6
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm6, %xmm6
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm7, %xmm14
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3]
; AVX2-FP-NEXT:    vinserti128 $1, %xmm5, %ymm0, %ymm5
; AVX2-FP-NEXT:    vinserti128 $1, %xmm6, %ymm0, %ymm6
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm5 = ymm6[0,1,2,3,4,5,6],ymm5[7]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm6, %xmm6
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm7, %xmm14
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3]
; AVX2-FP-NEXT:    vpshufb %xmm1, %xmm9, %xmm14
; AVX2-FP-NEXT:    vpshufb %xmm1, %xmm10, %xmm15
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3]
; AVX2-FP-NEXT:    vinserti128 $1, %xmm6, %ymm0, %ymm6
; AVX2-FP-NEXT:    vinserti128 $1, %xmm14, %ymm0, %ymm14
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm6 = ymm14[0,1,2,3,4],ymm6[5],ymm14[6,7]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm5 = ymm6[0,1,2,3,4,5],ymm5[6,7]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm5[4,5,6,7]
; AVX2-FP-NEXT:    vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm7, %xmm4
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm5, %xmm5
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm5, %xmm5
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm6, %xmm6
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm4 = xmm5[0,1,2],xmm4[3]
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm8, %xmm5
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm6, %xmm6
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm1, %xmm9, %xmm6
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm1, %xmm8, %xmm14
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm5 = xmm6[0],xmm5[1],xmm6[2,3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm4 = xmm5[0,1],xmm4[2,3]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm10, %xmm5
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm11, %xmm2
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm5[0],xmm2[1],xmm5[1],xmm2[2],xmm5[2],xmm2[3],xmm5[3]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm12, %xmm5
; AVX2-FP-NEXT:    vmovdqa (%rsp), %xmm13 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm13, %xmm3
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm5[0],xmm3[1],xmm5[1],xmm3[2],xmm5[2],xmm3[3],xmm5[3]
; AVX2-FP-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX2-FP-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4,5,6],ymm2[7]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm3, %xmm3
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm5, %xmm0
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm1, %xmm3, %xmm3
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm1, %xmm5, %xmm1
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1],xmm1[2],xmm3[2],xmm1[3],xmm3[3]
; AVX2-FP-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX2-FP-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm1
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3,4],ymm0[5],ymm1[6,7]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm2[6,7]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm4[0,1,2,3],ymm0[4,5,6,7]
; AVX2-FP-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm2 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm0, %xmm0
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm1, %xmm1
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm3 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm1, %xmm1
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm4, %xmm4
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm4[0],xmm1[0],xmm4[1],xmm1[1],xmm4[2],xmm1[2],xmm4[3],xmm1[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm4 = xmm1[0,1,2],xmm0[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm0 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm1, %xmm1
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm5, %xmm5
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm1[0],xmm5[1],xmm1[1],xmm5[2],xmm1[2],xmm5[3],xmm1[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm1 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm1, %xmm6, %xmm6
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm1, %xmm14, %xmm14
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm5 = xmm6[0],xmm5[1],xmm6[2,3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm4 = xmm5[0,1],xmm4[2,3]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm5, %xmm5
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm6, %xmm6
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm6, %xmm6
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm14, %xmm14
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3]
; AVX2-FP-NEXT:    vinserti128 $1, %xmm5, %ymm0, %ymm5
; AVX2-FP-NEXT:    vinserti128 $1, %xmm6, %ymm0, %ymm6
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm5 = ymm6[0,1,2,3,4,5,6],ymm5[7]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm6, %xmm6
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm14, %xmm14
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm1, %xmm14, %xmm14
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm1, %xmm15, %xmm15
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3]
; AVX2-FP-NEXT:    vinserti128 $1, %xmm6, %ymm0, %ymm6
; AVX2-FP-NEXT:    vinserti128 $1, %xmm14, %ymm0, %ymm14
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm6 = ymm14[0,1,2,3,4],ymm6[5],ymm14[6,7]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm5 = ymm6[0,1,2,3,4,5],ymm5[6,7]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm5[4,5,6,7]
; AVX2-FP-NEXT:    vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm7, %xmm4
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm5, %xmm5
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm5, %xmm5
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm6, %xmm6
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm4 = xmm5[0,1,2],xmm4[3]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm5, %xmm5
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm6, %xmm6
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX2-FP-NEXT:    vpshufb %xmm1, %xmm9, %xmm6
; AVX2-FP-NEXT:    vpshufb %xmm1, %xmm8, %xmm14
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm5 = xmm6[0],xmm5[1],xmm6[2,3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm4 = xmm5[0,1],xmm4[2,3]
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm10, %xmm5
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm11, %xmm2
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm5[0],xmm2[1],xmm5[1],xmm2[2],xmm5[2],xmm2[3],xmm5[3]
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm12, %xmm5
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm13, %xmm3
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm5[0],xmm3[1],xmm5[1],xmm3[2],xmm5[2],xmm3[3],xmm5[3]
; AVX2-FP-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX2-FP-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4,5,6],ymm2[7]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm13, %xmm3
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm7, %xmm0
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm1, %xmm8, %xmm3
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm1, %xmm9, %xmm1
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1],xmm1[2],xmm3[2],xmm1[3],xmm3[3]
; AVX2-FP-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX2-FP-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm1
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3,4],ymm0[5],ymm1[6,7]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm2[6,7]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm4[0,1,2,3],ymm0[4,5,6,7]
; AVX2-FP-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm2 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm0, %xmm0
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm10, %xmm1
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm3 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm11, %xmm1
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm4, %xmm4
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm4[0],xmm1[0],xmm4[1],xmm1[1],xmm4[2],xmm1[2],xmm4[3],xmm1[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm4 = xmm1[0,1,2],xmm0[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm0 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm1, %xmm1
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm5, %xmm5
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm1[0],xmm5[1],xmm1[1],xmm5[2],xmm1[2],xmm5[3],xmm1[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm1 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm1, %xmm6, %xmm6
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm1, %xmm12, %xmm14
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm5 = xmm6[0],xmm5[1],xmm6[2,3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm4 = xmm5[0,1],xmm4[2,3]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm5, %xmm5
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm6, %xmm6
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm6, %xmm6
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm14, %xmm14
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3]
; AVX2-FP-NEXT:    vinserti128 $1, %xmm5, %ymm0, %ymm5
; AVX2-FP-NEXT:    vinserti128 $1, %xmm6, %ymm0, %ymm6
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm5 = ymm6[0,1,2,3,4,5,6],ymm5[7]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm6, %xmm6
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm14, %xmm14
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm1, %xmm14, %xmm14
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm1, %xmm15, %xmm15
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3]
; AVX2-FP-NEXT:    vinserti128 $1, %xmm6, %ymm0, %ymm6
; AVX2-FP-NEXT:    vinserti128 $1, %xmm14, %ymm0, %ymm14
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm6 = ymm14[0,1,2,3,4],ymm6[5],ymm14[6,7]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm5 = ymm6[0,1,2,3,4,5],ymm5[6,7]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm5[4,5,6,7]
; AVX2-FP-NEXT:    vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm4, %xmm4
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm5, %xmm5
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm5, %xmm5
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm6, %xmm6
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm4 = xmm5[0,1,2],xmm4[3]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm5, %xmm5
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm6, %xmm6
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm1, %xmm6, %xmm6
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm1, %xmm14, %xmm14
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm5 = xmm6[0],xmm5[1],xmm6[2,3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm4 = xmm5[0,1],xmm4[2,3]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm5, %xmm5
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm6, %xmm2
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm5[0],xmm2[1],xmm5[1],xmm2[2],xmm5[2],xmm2[3],xmm5[3]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm5, %xmm5
; AVX2-FP-NEXT:    vmovdqa (%rsp), %xmm6 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm6, %xmm3
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm5[0],xmm3[1],xmm5[1],xmm3[2],xmm5[2],xmm3[3],xmm5[3]
; AVX2-FP-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX2-FP-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4,5,6],ymm2[7]
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm13, %xmm3
; AVX2-FP-NEXT:    vmovdqa %xmm13, %xmm5
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm7, %xmm0
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3]
; AVX2-FP-NEXT:    vpshufb %xmm1, %xmm8, %xmm3
; AVX2-FP-NEXT:    vpshufb %xmm1, %xmm9, %xmm1
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1],xmm1[2],xmm3[2],xmm1[3],xmm3[3]
; AVX2-FP-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX2-FP-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm1
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3,4],ymm0[5],ymm1[6,7]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm2[6,7]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm4[0,1,2,3],ymm0[4,5,6,7]
; AVX2-FP-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm3 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm0, %xmm0
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm10, %xmm1
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm0 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm11, %xmm2
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm4, %xmm4
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm4[0],xmm2[0],xmm4[1],xmm2[1],xmm4[2],xmm2[2],xmm4[3],xmm2[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm4 = xmm2[0,1,2],xmm1[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm1 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm1, %xmm2, %xmm2
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm1, %xmm6, %xmm6
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm6[0],xmm2[0],xmm6[1],xmm2[1],xmm6[2],xmm2[2],xmm6[3],xmm2[3]
; AVX2-FP-NEXT:    vpbroadcastw {{.*#+}} xmm2 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm10, %xmm14
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm12, %xmm15
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm6 = xmm14[0],xmm6[1],xmm14[2,3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm4 = xmm6[0,1],xmm4[2,3]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm6, %xmm6
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm10, %xmm14
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm10, %xmm14
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm10, %xmm15
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3]
; AVX2-FP-NEXT:    vinserti128 $1, %xmm6, %ymm0, %ymm6
; AVX2-FP-NEXT:    vinserti128 $1, %xmm14, %ymm0, %ymm14
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm6 = ymm14[0,1,2,3,4,5,6],ymm6[7]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm1, %xmm10, %xmm14
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm1, %xmm10, %xmm15
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm10, %xmm15
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm10, %xmm13
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm13[0],xmm15[0],xmm13[1],xmm15[1],xmm13[2],xmm15[2],xmm13[3],xmm15[3]
; AVX2-FP-NEXT:    vinserti128 $1, %xmm14, %ymm0, %ymm14
; AVX2-FP-NEXT:    vinserti128 $1, %xmm13, %ymm0, %ymm13
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm13 = ymm13[0,1,2,3,4],ymm14[5],ymm13[6,7]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm6 = ymm13[0,1,2,3,4,5],ymm6[6,7]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm6[4,5,6,7]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm6, %xmm6
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm10, %xmm13
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm13[0],xmm6[0],xmm13[1],xmm6[1],xmm13[2],xmm6[2],xmm13[3],xmm6[3]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm10, %xmm13
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm10, %xmm14
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm14[0],xmm13[0],xmm14[1],xmm13[1],xmm14[2],xmm13[2],xmm14[3],xmm13[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm6 = xmm13[0,1,2],xmm6[3]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm1, %xmm10, %xmm13
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm1, %xmm10, %xmm14
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm14[0],xmm13[0],xmm14[1],xmm13[1],xmm14[2],xmm13[2],xmm14[3],xmm13[3]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm10, %xmm14
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm10, %xmm15
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm13 = xmm14[0],xmm13[1],xmm14[2,3]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} xmm6 = xmm13[0,1],xmm6[2,3]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm10, %xmm13
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm3, %xmm10, %xmm3
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm13[0],xmm3[1],xmm13[1],xmm3[2],xmm13[2],xmm3[3],xmm13[3]
; AVX2-FP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm10, %xmm12
; AVX2-FP-NEXT:    vmovdqa (%rsp), %xmm10 # 16-byte Reload
; AVX2-FP-NEXT:    vpshufb %xmm0, %xmm10, %xmm0
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm12[0],xmm0[1],xmm12[1],xmm0[2],xmm12[2],xmm0[3],xmm12[3]
; AVX2-FP-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX2-FP-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5,6],ymm3[7]
; AVX2-FP-NEXT:    vpshufb %xmm1, %xmm5, %xmm3
; AVX2-FP-NEXT:    vpshufb %xmm1, %xmm7, %xmm1
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1],xmm1[2],xmm3[2],xmm1[3],xmm3[3]
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm8, %xmm3
; AVX2-FP-NEXT:    vpshufb %xmm2, %xmm9, %xmm2
; AVX2-FP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
; AVX2-FP-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm1
; AVX2-FP-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3,4],ymm1[5],ymm2[6,7]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
; AVX2-FP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm6[0,1,2,3],ymm0[4,5,6,7]
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, (%rsi)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 32(%rsi)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, (%rdx)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 32(%rdx)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, (%rcx)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 32(%rcx)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, (%r8)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 32(%r8)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, (%r9)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 32(%r9)
; AVX2-FP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, (%rax)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 32(%rax)
; AVX2-FP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, (%rax)
; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FP-NEXT:    vmovaps %ymm1, 32(%rax)
; AVX2-FP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FP-NEXT:    vmovdqa %ymm0, (%rax)
; AVX2-FP-NEXT:    vmovdqa %ymm4, 32(%rax)
; AVX2-FP-NEXT:    addq $840, %rsp # imm = 0x348
; AVX2-FP-NEXT:    vzeroupper
; AVX2-FP-NEXT:    retq
;
; AVX2-FCP-LABEL: load_i8_stride8_vf64:
; AVX2-FCP:       # %bb.0:
; AVX2-FCP-NEXT:    subq $904, %rsp # imm = 0x388
; AVX2-FCP-NEXT:    vmovdqa 368(%rdi), %xmm0
; AVX2-FCP-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm1 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX2-FCP-NEXT:    vpshufb %xmm1, %xmm0, %xmm0
; AVX2-FCP-NEXT:    vmovdqa 352(%rdi), %xmm2
; AVX2-FCP-NEXT:    vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FCP-NEXT:    vpshufb %xmm1, %xmm2, %xmm2
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
; AVX2-FCP-NEXT:    vmovdqa 336(%rdi), %xmm15
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm3 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX2-FCP-NEXT:    vpshufb %xmm3, %xmm15, %xmm2
; AVX2-FCP-NEXT:    vmovdqa %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FCP-NEXT:    vmovdqa 320(%rdi), %xmm4
; AVX2-FCP-NEXT:    vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FCP-NEXT:    vpshufb %xmm3, %xmm4, %xmm4
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm4[0],xmm2[0],xmm4[1],xmm2[1],xmm4[2],xmm2[2],xmm4[3],xmm2[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm2[0,1,2],xmm0[3]
; AVX2-FCP-NEXT:    vmovdqa 304(%rdi), %xmm14
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm2 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX2-FCP-NEXT:    vpshufb %xmm2, %xmm14, %xmm4
; AVX2-FCP-NEXT:    vmovdqa %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FCP-NEXT:    vmovdqa 288(%rdi), %xmm5
; AVX2-FCP-NEXT:    vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FCP-NEXT:    vpshufb %xmm2, %xmm5, %xmm5
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX2-FCP-NEXT:    vmovdqa 272(%rdi), %xmm5
; AVX2-FCP-NEXT:    vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm8 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX2-FCP-NEXT:    vpshufb %xmm8, %xmm5, %xmm5
; AVX2-FCP-NEXT:    vmovdqa 256(%rdi), %xmm6
; AVX2-FCP-NEXT:    vmovdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FCP-NEXT:    vpshufb %xmm8, %xmm6, %xmm6
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX2-FCP-NEXT:    vmovdqa 448(%rdi), %ymm6
; AVX2-FCP-NEXT:    vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm4 = xmm5[0],xmm4[1],xmm5[2,3]
; AVX2-FCP-NEXT:    vmovdqa 480(%rdi), %ymm5
; AVX2-FCP-NEXT:    vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm9 = xmm4[0,1],xmm0[2,3]
; AVX2-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm0 = [0,2,2,3,0,2,4,6]
; AVX2-FCP-NEXT:    vpermd %ymm5, %ymm0, %ymm4
; AVX2-FCP-NEXT:    vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm4 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,16,20,24,28]
; AVX2-FCP-NEXT:    vpermd %ymm6, %ymm0, %ymm5
; AVX2-FCP-NEXT:    vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm7 = [0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12]
; AVX2-FCP-NEXT:    vpshufb %ymm7, %ymm5, %ymm5
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm10 = ymm5[0,1,2,3,4,5,6],ymm4[7]
; AVX2-FCP-NEXT:    vmovdqa 384(%rdi), %ymm5
; AVX2-FCP-NEXT:    vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovdqa 416(%rdi), %ymm4
; AVX2-FCP-NEXT:    vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vpermd %ymm4, %ymm0, %ymm13
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm11 = ymm13[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,16,20,24,28,u,u,u,u,u,u,u,u]
; AVX2-FCP-NEXT:    vmovdqu %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vpermd %ymm5, %ymm0, %ymm5
; AVX2-FCP-NEXT:    vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm6 = [0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12]
; AVX2-FCP-NEXT:    vpshufb %ymm6, %ymm5, %ymm12
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm11 = ymm12[0,1,2,3,4],ymm11[5],ymm12[6,7]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm10 = ymm11[0,1,2,3,4,5],ymm10[6,7]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm4 = ymm9[0,1,2,3],ymm10[4,5,6,7]
; AVX2-FCP-NEXT:    vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovdqa 112(%rdi), %xmm4
; AVX2-FCP-NEXT:    vmovdqa %xmm4, (%rsp) # 16-byte Spill
; AVX2-FCP-NEXT:    vpshufb %xmm1, %xmm4, %xmm9
; AVX2-FCP-NEXT:    vmovdqa 96(%rdi), %xmm4
; AVX2-FCP-NEXT:    vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FCP-NEXT:    vpshufb %xmm1, %xmm4, %xmm1
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm9[0],xmm1[1],xmm9[1],xmm1[2],xmm9[2],xmm1[3],xmm9[3]
; AVX2-FCP-NEXT:    vmovdqa 80(%rdi), %xmm4
; AVX2-FCP-NEXT:    vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FCP-NEXT:    vpshufb %xmm3, %xmm4, %xmm9
; AVX2-FCP-NEXT:    vmovdqa 64(%rdi), %xmm4
; AVX2-FCP-NEXT:    vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FCP-NEXT:    vpshufb %xmm3, %xmm4, %xmm3
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm9[0],xmm3[1],xmm9[1],xmm3[2],xmm9[2],xmm3[3],xmm9[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm1 = xmm3[0,1,2],xmm1[3]
; AVX2-FCP-NEXT:    vmovdqa 32(%rdi), %xmm4
; AVX2-FCP-NEXT:    vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FCP-NEXT:    vmovdqa 48(%rdi), %xmm3
; AVX2-FCP-NEXT:    vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FCP-NEXT:    vpshufb %xmm2, %xmm3, %xmm3
; AVX2-FCP-NEXT:    vpshufb %xmm2, %xmm4, %xmm2
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
; AVX2-FCP-NEXT:    vmovdqa (%rdi), %xmm4
; AVX2-FCP-NEXT:    vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FCP-NEXT:    vmovdqa 16(%rdi), %xmm3
; AVX2-FCP-NEXT:    vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FCP-NEXT:    vpshufb %xmm8, %xmm3, %xmm3
; AVX2-FCP-NEXT:    vpshufb %xmm8, %xmm4, %xmm8
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm8[0],xmm3[0],xmm8[1],xmm3[1],xmm8[2],xmm3[2],xmm8[3],xmm3[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm2 = xmm3[0],xmm2[1],xmm3[2,3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm8 = xmm2[0,1],xmm1[2,3]
; AVX2-FCP-NEXT:    vmovdqa 192(%rdi), %ymm2
; AVX2-FCP-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovdqa 224(%rdi), %ymm1
; AVX2-FCP-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vpermd %ymm1, %ymm0, %ymm1
; AVX2-FCP-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,16,20,24,28]
; AVX2-FCP-NEXT:    vpermd %ymm2, %ymm0, %ymm2
; AVX2-FCP-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vpshufb %ymm7, %ymm2, %ymm10
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm10 = ymm10[0,1,2,3,4,5,6],ymm1[7]
; AVX2-FCP-NEXT:    vmovdqa 128(%rdi), %ymm1
; AVX2-FCP-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovdqa 160(%rdi), %ymm2
; AVX2-FCP-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vpermd %ymm2, %ymm0, %ymm2
; AVX2-FCP-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vpermd %ymm1, %ymm0, %ymm1
; AVX2-FCP-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm0 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,16,20,24,28,u,u,u,u,u,u,u,u]
; AVX2-FCP-NEXT:    vpshufb %ymm6, %ymm1, %ymm12
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm12[0,1,2,3,4],ymm0[5],ymm12[6,7]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm10[6,7]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm8[0,1,2,3],ymm0[4,5,6,7]
; AVX2-FCP-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm8 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm8, %xmm6, %xmm0
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm8, %xmm7, %xmm10
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm10[0],xmm0[0],xmm10[1],xmm0[1],xmm10[2],xmm0[2],xmm10[3],xmm0[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm10 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX2-FCP-NEXT:    vpshufb %xmm10, %xmm15, %xmm12
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm10, %xmm2, %xmm11
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm11[0],xmm12[0],xmm11[1],xmm12[1],xmm11[2],xmm12[2],xmm11[3],xmm12[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm1 = xmm11[0,1,2],xmm0[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm11 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX2-FCP-NEXT:    vpshufb %xmm11, %xmm14, %xmm12
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm11, %xmm15, %xmm14
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm14[0],xmm12[0],xmm14[1],xmm12[1],xmm14[2],xmm12[2],xmm14[3],xmm12[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm14 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm14, %xmm0, %xmm0
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm14, %xmm4, %xmm9
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm9[0],xmm0[0],xmm9[1],xmm0[1],xmm9[2],xmm0[2],xmm9[3],xmm0[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm0[0],xmm12[1],xmm0[2,3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]
; AVX2-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm1 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,17,21,25,29]
; AVX2-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm9 = ymm9[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,17,21,25,29,u,u,u,u]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm9[0,1,2,3,4,5,6],ymm1[7]
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm9 = ymm13[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,17,21,25,29,u,u,u,u,u,u,u,u]
; AVX2-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Reload
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm12 = ymm12[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,17,21,25,29,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm9 = ymm12[0,1,2,3,4],ymm9[5],ymm12[6,7]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm9[0,1,2,3,4,5],ymm1[6,7]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX2-FCP-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovdqa (%rsp), %xmm0 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm8, %xmm0, %xmm0
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm8, %xmm13, %xmm1
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm10, %xmm1, %xmm1
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm10, %xmm5, %xmm8
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm8[0],xmm1[0],xmm8[1],xmm1[1],xmm8[2],xmm1[2],xmm8[3],xmm1[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[3]
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm11, %xmm1, %xmm1
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm11, %xmm8, %xmm8
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm8[0],xmm1[0],xmm8[1],xmm1[1],xmm8[2],xmm1[2],xmm8[3],xmm1[3]
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm14, %xmm8, %xmm8
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm14, %xmm9, %xmm9
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm9[0],xmm8[0],xmm9[1],xmm8[1],xmm9[2],xmm8[2],xmm9[3],xmm8[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm1 = xmm8[0],xmm1[1],xmm8[2,3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3]
; AVX2-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,17,21,25,29]
; AVX2-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm8 = ymm8[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,17,21,25,29,u,u,u,u]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm8[0,1,2,3,4,5,6],ymm1[7]
; AVX2-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm8 = ymm8[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,17,21,25,29,u,u,u,u,u,u,u,u]
; AVX2-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm9 = ymm9[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,17,21,25,29,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm8 = ymm9[0,1,2,3,4],ymm8[5],ymm9[6,7]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm8[0,1,2,3,4,5],ymm1[6,7]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX2-FCP-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm8 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX2-FCP-NEXT:    vpshufb %xmm8, %xmm6, %xmm0
; AVX2-FCP-NEXT:    vpshufb %xmm8, %xmm7, %xmm1
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm1 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm1, %xmm6, %xmm9
; AVX2-FCP-NEXT:    vpshufb %xmm1, %xmm2, %xmm10
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm10[0],xmm9[0],xmm10[1],xmm9[1],xmm10[2],xmm9[2],xmm10[3],xmm9[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm9[0,1,2],xmm0[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm9 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm9, %xmm2, %xmm10
; AVX2-FCP-NEXT:    vpshufb %xmm9, %xmm15, %xmm11
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm11[0],xmm10[0],xmm11[1],xmm10[1],xmm11[2],xmm10[2],xmm11[3],xmm10[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm11 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm11, %xmm7, %xmm12
; AVX2-FCP-NEXT:    vpshufb %xmm11, %xmm4, %xmm14
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm14[0],xmm12[0],xmm14[1],xmm12[1],xmm14[2],xmm12[2],xmm14[3],xmm12[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm10 = xmm12[0],xmm10[1],xmm12[2,3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm10[0,1],xmm0[2,3]
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm10 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,22,26,30]
; AVX2-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm12 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,22,26,30,u,u,u,u]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm10 = ymm12[0,1,2,3,4,5,6],ymm10[7]
; AVX2-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm12 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,22,26,30,u,u,u,u,u,u,u,u]
; AVX2-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm14 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,22,26,30,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm12 = ymm14[0,1,2,3,4],ymm12[5],ymm14[6,7]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm10 = ymm12[0,1,2,3,4,5],ymm10[6,7]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm10[4,5,6,7]
; AVX2-FCP-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovdqa (%rsp), %xmm4 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm8, %xmm4, %xmm0
; AVX2-FCP-NEXT:    vpshufb %xmm8, %xmm13, %xmm8
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm8[0],xmm0[0],xmm8[1],xmm0[1],xmm8[2],xmm0[2],xmm8[3],xmm0[3]
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm1, %xmm13, %xmm8
; AVX2-FCP-NEXT:    vpshufb %xmm1, %xmm5, %xmm1
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm8[0],xmm1[1],xmm8[1],xmm1[2],xmm8[2],xmm1[3],xmm8[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[3]
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm9, %xmm15, %xmm1
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm9, %xmm3, %xmm8
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm8[0],xmm1[0],xmm8[1],xmm1[1],xmm8[2],xmm1[2],xmm8[3],xmm1[3]
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm11, %xmm5, %xmm8
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm11, %xmm5, %xmm9
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm9[0],xmm8[0],xmm9[1],xmm8[1],xmm9[2],xmm8[2],xmm9[3],xmm8[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm1 = xmm8[0],xmm1[1],xmm8[2,3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3]
; AVX2-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,22,26,30]
; AVX2-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm8 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,22,26,30,u,u,u,u]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm8[0,1,2,3,4,5,6],ymm1[7]
; AVX2-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm8 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,22,26,30,u,u,u,u,u,u,u,u]
; AVX2-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm9 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,22,26,30,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm8 = ymm9[0,1,2,3,4],ymm8[5],ymm9[6,7]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm8[0,1,2,3,4,5],ymm1[6,7]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX2-FCP-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm5 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm5, %xmm0, %xmm0
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm5, %xmm1, %xmm1
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm1 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX2-FCP-NEXT:    vpshufb %xmm1, %xmm6, %xmm9
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm1, %xmm5, %xmm10
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm10[0],xmm9[0],xmm10[1],xmm9[1],xmm10[2],xmm9[2],xmm10[3],xmm9[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm9[0,1,2],xmm0[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm9 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX2-FCP-NEXT:    vpshufb %xmm9, %xmm2, %xmm10
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm9, %xmm8, %xmm11
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm11[0],xmm10[0],xmm11[1],xmm10[1],xmm11[2],xmm10[2],xmm11[3],xmm10[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm11 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX2-FCP-NEXT:    vpshufb %xmm11, %xmm7, %xmm12
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm11, %xmm2, %xmm14
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm14[0],xmm12[0],xmm14[1],xmm12[1],xmm14[2],xmm12[2],xmm14[3],xmm12[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm10 = xmm12[0],xmm10[1],xmm12[2,3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm10[0,1],xmm0[2,3]
; AVX2-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm10 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,19,23,27,31]
; AVX2-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm12 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,19,23,27,31,u,u,u,u]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm10 = ymm12[0,1,2,3,4,5,6],ymm10[7]
; AVX2-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm14 = [3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15]
; AVX2-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload
; AVX2-FCP-NEXT:    vpshufb %ymm14, %ymm6, %ymm7
; AVX2-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm2 = [3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15]
; AVX2-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload
; AVX2-FCP-NEXT:    vpshufb %ymm2, %ymm6, %ymm6
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm6 = ymm6[0,1,2,3,4],ymm7[5],ymm6[6,7]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm6 = ymm6[0,1,2,3,4,5],ymm10[6,7]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm6[4,5,6,7]
; AVX2-FCP-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm6 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX2-FCP-NEXT:    vpshufb %xmm6, %xmm4, %xmm0
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm6, %xmm4, %xmm6
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm6[0],xmm0[0],xmm6[1],xmm0[1],xmm6[2],xmm0[2],xmm6[3],xmm0[3]
; AVX2-FCP-NEXT:    vpshufb %xmm1, %xmm13, %xmm6
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm1, %xmm13, %xmm1
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm6[0],xmm1[1],xmm6[1],xmm1[2],xmm6[2],xmm1[3],xmm6[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[3]
; AVX2-FCP-NEXT:    vpshufb %xmm9, %xmm15, %xmm1
; AVX2-FCP-NEXT:    vpshufb %xmm9, %xmm3, %xmm6
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm6[0],xmm1[0],xmm6[1],xmm1[1],xmm6[2],xmm1[2],xmm6[3],xmm1[3]
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm11, %xmm3, %xmm6
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm11, %xmm3, %xmm7
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm1 = xmm6[0],xmm1[1],xmm6[2,3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3]
; AVX2-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,19,23,27,31]
; AVX2-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm4 = ymm3[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,19,23,27,31,u,u,u,u]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm4[0,1,2,3,4,5,6],ymm1[7]
; AVX2-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-FCP-NEXT:    vpshufb %ymm14, %ymm3, %ymm3
; AVX2-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload
; AVX2-FCP-NEXT:    vpshufb %ymm2, %ymm4, %ymm2
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3,4],ymm3[5],ymm2[6,7]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX2-FCP-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm1 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm1, %xmm0, %xmm0
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm1, %xmm14, %xmm1
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm6 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm6, %xmm1, %xmm1
; AVX2-FCP-NEXT:    vpshufb %xmm6, %xmm5, %xmm2
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm7 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm7, %xmm1, %xmm1
; AVX2-FCP-NEXT:    vpshufb %xmm7, %xmm8, %xmm2
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm8 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm8, %xmm2, %xmm2
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm8, %xmm3, %xmm3
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm1 = xmm2[0],xmm1[1],xmm2[2,3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm9 = xmm1[0,1],xmm0[2,3]
; AVX2-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm0 = [1,3,2,3,1,3,5,7]
; AVX2-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm1 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm2 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm15 = [0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12]
; AVX2-FCP-NEXT:    vpshufb %ymm15, %ymm1, %ymm3
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm4 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,16,20,24,28,u,u,u,u]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm10 = ymm4[0,1,2,3,4,5,6],ymm3[7]
; AVX2-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm1 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm2 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm11 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,16,20,24,28,u,u,u,u,u,u,u,u]
; AVX2-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm1 = [0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12]
; AVX2-FCP-NEXT:    vpshufb %ymm1, %ymm2, %ymm12
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm11 = ymm12[0,1,2,3,4],ymm11[5],ymm12[6,7]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm10 = ymm11[0,1,2,3,4,5],ymm10[6,7]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm9[0,1,2,3],ymm10[4,5,6,7]
; AVX2-FCP-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovdqa (%rsp), %xmm1 # 16-byte Reload
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm3 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX2-FCP-NEXT:    vpshufb %xmm3, %xmm1, %xmm9
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm3, %xmm2, %xmm5
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm9[0],xmm5[1],xmm9[1],xmm5[2],xmm9[2],xmm5[3],xmm9[3]
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm6, %xmm2, %xmm9
; AVX2-FCP-NEXT:    vpshufb %xmm6, %xmm13, %xmm6
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm6[0],xmm9[0],xmm6[1],xmm9[1],xmm6[2],xmm9[2],xmm6[3],xmm9[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm5 = xmm6[0,1,2],xmm5[3]
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm7, %xmm3, %xmm6
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm7, %xmm4, %xmm7
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm8, %xmm4, %xmm7
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm8, %xmm9, %xmm8
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm7 = xmm8[0],xmm7[0],xmm8[1],xmm7[1],xmm8[2],xmm7[2],xmm8[3],xmm7[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm6 = xmm7[0],xmm6[1],xmm7[2,3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm8 = xmm6[0,1],xmm5[2,3]
; AVX2-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm5 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vpshufb %ymm15, %ymm5, %ymm7
; AVX2-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm5 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm9 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,16,20,24,28,u,u,u,u]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm9 = ymm9[0,1,2,3,4,5,6],ymm7[7]
; AVX2-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm5 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-FCP-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm10 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,16,20,24,28,u,u,u,u,u,u,u,u]
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm11 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,16,20,24,28,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm10 = ymm11[0,1,2,3,4],ymm10[5],ymm11[6,7]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm9 = ymm10[0,1,2,3,4,5],ymm9[6,7]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm8[0,1,2,3],ymm9[4,5,6,7]
; AVX2-FCP-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm8 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm8, %xmm0, %xmm9
; AVX2-FCP-NEXT:    vpshufb %xmm8, %xmm14, %xmm10
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm10[0],xmm9[0],xmm10[1],xmm9[1],xmm10[2],xmm9[2],xmm10[3],xmm9[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm10 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm10, %xmm5, %xmm11
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm10, %xmm6, %xmm12
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm12[0],xmm11[0],xmm12[1],xmm11[1],xmm12[2],xmm11[2],xmm12[3],xmm11[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm9 = xmm11[0,1,2],xmm9[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm11 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm11, %xmm0, %xmm12
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm11, %xmm7, %xmm13
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm13[0],xmm12[0],xmm13[1],xmm12[1],xmm13[2],xmm12[2],xmm13[3],xmm12[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm13 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm13, %xmm0, %xmm14
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm13, %xmm0, %xmm15
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm12 = xmm14[0],xmm12[1],xmm14[2,3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm9 = xmm12[0,1],xmm9[2,3]
; AVX2-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm12 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,17,21,25,29]
; AVX2-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm14 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,17,21,25,29,u,u,u,u]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm12 = ymm14[0,1,2,3,4,5,6],ymm12[7]
; AVX2-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm14 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,17,21,25,29,u,u,u,u,u,u,u,u]
; AVX2-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm15 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,17,21,25,29,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm14 = ymm15[0,1,2,3,4],ymm14[5],ymm15[6,7]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm12 = ymm14[0,1,2,3,4,5],ymm12[6,7]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm9[0,1,2,3],ymm12[4,5,6,7]
; AVX2-FCP-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vpshufb %xmm8, %xmm1, %xmm9
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm8, %xmm0, %xmm8
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm9[0],xmm8[1],xmm9[1],xmm8[2],xmm9[2],xmm8[3],xmm9[3]
; AVX2-FCP-NEXT:    vmovdqa %xmm2, %xmm1
; AVX2-FCP-NEXT:    vpshufb %xmm10, %xmm2, %xmm9
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm10, %xmm2, %xmm10
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm10[0],xmm9[0],xmm10[1],xmm9[1],xmm10[2],xmm9[2],xmm10[3],xmm9[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm8 = xmm9[0,1,2],xmm8[3]
; AVX2-FCP-NEXT:    vpshufb %xmm11, %xmm3, %xmm9
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm11, %xmm3, %xmm10
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm10[0],xmm9[0],xmm10[1],xmm9[1],xmm10[2],xmm9[2],xmm10[3],xmm9[3]
; AVX2-FCP-NEXT:    vpshufb %xmm13, %xmm4, %xmm10
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm13, %xmm4, %xmm11
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm11[0],xmm10[0],xmm11[1],xmm10[1],xmm11[2],xmm10[2],xmm11[3],xmm10[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm9 = xmm10[0],xmm9[1],xmm10[2,3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm8 = xmm9[0,1],xmm8[2,3]
; AVX2-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm9 = ymm9[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,17,21,25,29]
; AVX2-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm10 = ymm10[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,17,21,25,29,u,u,u,u]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm9 = ymm10[0,1,2,3,4,5,6],ymm9[7]
; AVX2-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm10 = ymm10[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,17,21,25,29,u,u,u,u,u,u,u,u]
; AVX2-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm11 = ymm11[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,17,21,25,29,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm10 = ymm11[0,1,2,3,4],ymm10[5],ymm11[6,7]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm9 = ymm10[0,1,2,3,4,5],ymm9[6,7]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm8 = ymm8[0,1,2,3],ymm9[4,5,6,7]
; AVX2-FCP-NEXT:    vmovdqu %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm8 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm8, %xmm9, %xmm9
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm8, %xmm10, %xmm10
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm10[0],xmm9[0],xmm10[1],xmm9[1],xmm10[2],xmm9[2],xmm10[3],xmm9[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm10 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX2-FCP-NEXT:    vpshufb %xmm10, %xmm5, %xmm11
; AVX2-FCP-NEXT:    vpshufb %xmm10, %xmm6, %xmm12
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm12[0],xmm11[0],xmm12[1],xmm11[1],xmm12[2],xmm11[2],xmm12[3],xmm11[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm9 = xmm11[0,1,2],xmm9[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm11 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm11, %xmm6, %xmm12
; AVX2-FCP-NEXT:    vpshufb %xmm11, %xmm7, %xmm13
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm13[0],xmm12[0],xmm13[1],xmm12[1],xmm13[2],xmm12[2],xmm13[3],xmm12[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm13 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm13, %xmm7, %xmm14
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm13, %xmm5, %xmm15
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm12 = xmm14[0],xmm12[1],xmm14[2,3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm9 = xmm12[0,1],xmm9[2,3]
; AVX2-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm12 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,22,26,30]
; AVX2-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm14 = ymm5[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,22,26,30,u,u,u,u]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm12 = ymm14[0,1,2,3,4,5,6],ymm12[7]
; AVX2-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm5 = [2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14]
; AVX2-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Reload
; AVX2-FCP-NEXT:    vpshufb %ymm5, %ymm14, %ymm14
; AVX2-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm0 = [2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14]
; AVX2-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; AVX2-FCP-NEXT:    vpshufb %ymm0, %ymm15, %ymm15
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm14 = ymm15[0,1,2,3,4],ymm14[5],ymm15[6,7]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm12 = ymm14[0,1,2,3,4,5],ymm12[6,7]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm9 = ymm9[0,1,2,3],ymm12[4,5,6,7]
; AVX2-FCP-NEXT:    vmovdqu %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vmovdqa (%rsp), %xmm9 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm8, %xmm9, %xmm9
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm8, %xmm0, %xmm8
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm9[0],xmm8[1],xmm9[1],xmm8[2],xmm9[2],xmm8[3],xmm9[3]
; AVX2-FCP-NEXT:    vpshufb %xmm10, %xmm1, %xmm9
; AVX2-FCP-NEXT:    vpshufb %xmm10, %xmm2, %xmm10
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm10[0],xmm9[0],xmm10[1],xmm9[1],xmm10[2],xmm9[2],xmm10[3],xmm9[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm8 = xmm9[0,1,2],xmm8[3]
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm11, %xmm0, %xmm9
; AVX2-FCP-NEXT:    vpshufb %xmm11, %xmm3, %xmm10
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm10[0],xmm9[0],xmm10[1],xmm9[1],xmm10[2],xmm9[2],xmm10[3],xmm9[3]
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm13, %xmm0, %xmm10
; AVX2-FCP-NEXT:    vpshufb %xmm13, %xmm4, %xmm11
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm11[0],xmm10[0],xmm11[1],xmm10[1],xmm11[2],xmm10[2],xmm11[3],xmm10[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm9 = xmm10[0],xmm9[1],xmm10[2,3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm8 = xmm9[0,1],xmm8[2,3]
; AVX2-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm9 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,22,26,30]
; AVX2-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm10 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,22,26,30,u,u,u,u]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm9 = ymm10[0,1,2,3,4,5,6],ymm9[7]
; AVX2-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vpshufb %ymm5, %ymm1, %ymm10
; AVX2-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vpshufb {{.*#+}} ymm11 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,22,26,30,u,u,u,u,u,u,u,u,u,u,u,u]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm10 = ymm11[0,1,2,3,4],ymm10[5],ymm11[6,7]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm9 = ymm10[0,1,2,3,4,5],ymm9[6,7]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm8[0,1,2,3],ymm9[4,5,6,7]
; AVX2-FCP-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm9 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm9, %xmm1, %xmm10
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm9, %xmm1, %xmm11
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm11[0],xmm10[0],xmm11[1],xmm10[1],xmm11[2],xmm10[2],xmm11[3],xmm10[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm11 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm11, %xmm1, %xmm12
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm11, %xmm1, %xmm13
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm13[0],xmm12[0],xmm13[1],xmm12[1],xmm13[2],xmm12[2],xmm13[3],xmm12[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm10 = xmm12[0,1,2],xmm10[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm12 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX2-FCP-NEXT:    vpshufb %xmm12, %xmm6, %xmm13
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm12, %xmm1, %xmm14
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm14[0],xmm13[0],xmm14[1],xmm13[1],xmm14[2],xmm13[2],xmm14[3],xmm13[3]
; AVX2-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm14 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX2-FCP-NEXT:    vpshufb %xmm14, %xmm7, %xmm15
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm14, %xmm1, %xmm8
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm15[0],xmm8[1],xmm15[1],xmm8[2],xmm15[2],xmm8[3],xmm15[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm8 = xmm8[0],xmm13[1],xmm8[2,3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm8 = xmm8[0,1],xmm10[2,3]
; AVX2-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm5 = [3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15]
; AVX2-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vpshufb %ymm5, %ymm1, %ymm2
; AVX2-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm6 = [3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15]
; AVX2-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-FCP-NEXT:    vpshufb %ymm6, %ymm1, %ymm1
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5,6],ymm2[7]
; AVX2-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm7 = [3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15]
; AVX2-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX2-FCP-NEXT:    vpshufb %ymm7, %ymm2, %ymm2
; AVX2-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm10 = [3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15]
; AVX2-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload
; AVX2-FCP-NEXT:    vpshufb %ymm10, %ymm3, %ymm3
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4],ymm2[5],ymm3[6,7]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm8[0,1,2,3],ymm1[4,5,6,7]
; AVX2-FCP-NEXT:    vmovdqa (%rsp), %xmm2 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm9, %xmm2, %xmm2
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm9, %xmm3, %xmm3
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm11, %xmm3, %xmm3
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm11, %xmm4, %xmm4
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm2 = xmm3[0,1,2],xmm2[3]
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm12, %xmm3, %xmm3
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm12, %xmm4, %xmm4
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm14, %xmm4, %xmm4
; AVX2-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX2-FCP-NEXT:    vpshufb %xmm14, %xmm8, %xmm8
; AVX2-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm8[0],xmm4[0],xmm8[1],xmm4[1],xmm8[2],xmm4[2],xmm8[3],xmm4[3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm3 = xmm4[0],xmm3[1],xmm4[2,3]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} xmm2 = xmm3[0,1],xmm2[2,3]
; AVX2-FCP-NEXT:    vpshufb %ymm5, %ymm0, %ymm3
; AVX2-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FCP-NEXT:    vpshufb %ymm6, %ymm0, %ymm4
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm3 = ymm4[0,1,2,3,4,5,6],ymm3[7]
; AVX2-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FCP-NEXT:    vpshufb %ymm7, %ymm0, %ymm4
; AVX2-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FCP-NEXT:    vpshufb %ymm10, %ymm0, %ymm0
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm4[5],ymm0[6,7]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm3[6,7]
; AVX2-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm2, (%rsi)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm2, 32(%rsi)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm2, (%rdx)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm2, 32(%rdx)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm2, (%rcx)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm2, 32(%rcx)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm2, (%r8)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm2, 32(%r8)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm2, (%r9)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm2, 32(%r9)
; AVX2-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm2, (%rax)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm2, 32(%rax)
; AVX2-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm2, (%rax)
; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX2-FCP-NEXT:    vmovaps %ymm2, 32(%rax)
; AVX2-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX2-FCP-NEXT:    vmovdqa %ymm0, (%rax)
; AVX2-FCP-NEXT:    vmovdqa %ymm1, 32(%rax)
; AVX2-FCP-NEXT:    addq $904, %rsp # imm = 0x388
; AVX2-FCP-NEXT:    vzeroupper
; AVX2-FCP-NEXT:    retq
;
; AVX512-LABEL: load_i8_stride8_vf64:
; AVX512:       # %bb.0:
; AVX512-NEXT:    subq $520, %rsp # imm = 0x208
; AVX512-NEXT:    vmovdqa64 256(%rdi), %zmm17
; AVX512-NEXT:    vmovdqa64 384(%rdi), %zmm0
; AVX512-NEXT:    vpmovqb %zmm0, %xmm2
; AVX512-NEXT:    vmovdqa 496(%rdi), %xmm3
; AVX512-NEXT:    vpbroadcastd {{.*#+}} xmm0 = [0,0,0,8,0,0,0,8,0,0,0,8,0,0,0,8]
; AVX512-NEXT:    vpshufb %xmm0, %xmm3, %xmm1
; AVX512-NEXT:    vmovdqa %xmm3, %xmm7
; AVX512-NEXT:    vmovdqa 480(%rdi), %xmm4
; AVX512-NEXT:    vpshufb %xmm0, %xmm4, %xmm3
; AVX512-NEXT:    vmovdqa %xmm4, %xmm8
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3]
; AVX512-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm3
; AVX512-NEXT:    vmovdqa 464(%rdi), %xmm5
; AVX512-NEXT:    vpbroadcastd {{.*#+}} xmm1 = [0,8,0,0,0,8,0,0,0,8,0,0,0,8,0,0]
; AVX512-NEXT:    vpshufb %xmm1, %xmm5, %xmm4
; AVX512-NEXT:    vmovdqa %xmm5, %xmm11
; AVX512-NEXT:    vmovdqa 448(%rdi), %xmm15
; AVX512-NEXT:    vpshufb %xmm1, %xmm15, %xmm5
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX512-NEXT:    vinserti128 $1, %xmm4, %ymm0, %ymm4
; AVX512-NEXT:    vpblendd {{.*#+}} ymm3 = ymm4[0,1,2,3,4,5,6],ymm3[7]
; AVX512-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX512-NEXT:    vmovdqa 384(%rdi), %ymm4
; AVX512-NEXT:    vpmovqb %ymm4, %xmm4
; AVX512-NEXT:    vinserti128 $1, %xmm4, %ymm0, %ymm4
; AVX512-NEXT:    vpblendd {{.*#+}} ymm2 = ymm4[0,1,2,3,4],ymm2[5],ymm4[6,7]
; AVX512-NEXT:    vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
; AVX512-NEXT:    vmovdqa 368(%rdi), %xmm4
; AVX512-NEXT:    vpshufb %xmm0, %xmm4, %xmm3
; AVX512-NEXT:    vmovdqa64 %xmm4, %xmm18
; AVX512-NEXT:    vmovdqa 352(%rdi), %xmm5
; AVX512-NEXT:    vpshufb %xmm0, %xmm5, %xmm4
; AVX512-NEXT:    vmovdqa %xmm5, %xmm14
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX512-NEXT:    vmovdqa 336(%rdi), %xmm4
; AVX512-NEXT:    vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512-NEXT:    vpshufb %xmm1, %xmm4, %xmm4
; AVX512-NEXT:    vmovdqa 320(%rdi), %xmm6
; AVX512-NEXT:    vpshufb %xmm1, %xmm6, %xmm5
; AVX512-NEXT:    vmovdqa64 %xmm6, %xmm19
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX512-NEXT:    vpblendd {{.*#+}} xmm3 = xmm4[0,1,2],xmm3[3]
; AVX512-NEXT:    vpmovqb %zmm17, %xmm4
; AVX512-NEXT:    vpblendd {{.*#+}} xmm3 = xmm4[0,1],xmm3[2,3]
; AVX512-NEXT:    vinserti32x4 $2, %xmm3, %zmm0, %zmm3
; AVX512-NEXT:    movb $-64, %al
; AVX512-NEXT:    kmovw %eax, %k1
; AVX512-NEXT:    vinserti64x4 $1, %ymm2, %zmm0, %zmm3 {%k1}
; AVX512-NEXT:    vmovdqa 240(%rdi), %xmm4
; AVX512-NEXT:    vpshufb %xmm0, %xmm4, %xmm2
; AVX512-NEXT:    vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512-NEXT:    vmovdqa64 %xmm4, %xmm25
; AVX512-NEXT:    vmovdqa 224(%rdi), %xmm4
; AVX512-NEXT:    vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512-NEXT:    vpshufb %xmm0, %xmm4, %xmm4
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm4[0],xmm2[0],xmm4[1],xmm2[1],xmm4[2],xmm2[2],xmm4[3],xmm2[3]
; AVX512-NEXT:    vmovdqa 208(%rdi), %xmm5
; AVX512-NEXT:    vpshufb %xmm1, %xmm5, %xmm4
; AVX512-NEXT:    vmovdqa64 %xmm5, %xmm29
; AVX512-NEXT:    vmovdqa 192(%rdi), %xmm10
; AVX512-NEXT:    vpshufb %xmm1, %xmm10, %xmm5
; AVX512-NEXT:    vmovdqa %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX512-NEXT:    vmovdqa64 128(%rdi), %zmm5
; AVX512-NEXT:    vpmovqb %zmm5, %xmm5
; AVX512-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX512-NEXT:    vinserti128 $1, %xmm4, %ymm0, %ymm4
; AVX512-NEXT:    vpblendd {{.*#+}} ymm2 = ymm4[0,1,2,3,4,5,6],ymm2[7]
; AVX512-NEXT:    vinserti128 $1, %xmm5, %ymm0, %ymm4
; AVX512-NEXT:    vmovdqa 128(%rdi), %ymm5
; AVX512-NEXT:    vpmovqb %ymm5, %xmm5
; AVX512-NEXT:    vinserti128 $1, %xmm5, %ymm0, %ymm5
; AVX512-NEXT:    vpblendd {{.*#+}} ymm4 = ymm5[0,1,2,3,4],ymm4[5],ymm5[6,7]
; AVX512-NEXT:    vpblendd {{.*#+}} ymm2 = ymm4[0,1,2,3,4,5],ymm2[6,7]
; AVX512-NEXT:    vmovdqa 112(%rdi), %xmm4
; AVX512-NEXT:    vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512-NEXT:    vpshufb %xmm0, %xmm4, %xmm4
; AVX512-NEXT:    vmovdqa 96(%rdi), %xmm5
; AVX512-NEXT:    vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512-NEXT:    vpshufb %xmm0, %xmm5, %xmm0
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm4[0],xmm0[1],xmm4[1],xmm0[2],xmm4[2],xmm0[3],xmm4[3]
; AVX512-NEXT:    vmovdqa 80(%rdi), %xmm5
; AVX512-NEXT:    vpshufb %xmm1, %xmm5, %xmm4
; AVX512-NEXT:    vmovdqa64 %xmm5, %xmm30
; AVX512-NEXT:    vmovdqa 64(%rdi), %xmm5
; AVX512-NEXT:    vpshufb %xmm1, %xmm5, %xmm1
; AVX512-NEXT:    vmovdqa64 %xmm5, %xmm28
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm4[0],xmm1[1],xmm4[1],xmm1[2],xmm4[2],xmm1[3],xmm4[3]
; AVX512-NEXT:    vpblendd {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[3]
; AVX512-NEXT:    vmovdqa64 (%rdi), %zmm16
; AVX512-NEXT:    vpmovqb %zmm16, %xmm1
; AVX512-NEXT:    vpblendd {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3]
; AVX512-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
; AVX512-NEXT:    vinserti64x4 $0, %ymm0, %zmm3, %zmm0
; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpbroadcastd {{.*#+}} xmm9 = [0,0,1,9,0,0,1,9,0,0,1,9,0,0,1,9]
; AVX512-NEXT:    vpshufb %xmm9, %xmm7, %xmm0
; AVX512-NEXT:    vmovdqa64 %xmm7, %xmm24
; AVX512-NEXT:    vpshufb %xmm9, %xmm8, %xmm2
; AVX512-NEXT:    vmovdqa64 %xmm8, %xmm21
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
; AVX512-NEXT:    vpbroadcastd {{.*#+}} xmm8 = [1,9,0,0,1,9,0,0,1,9,0,0,1,9,0,0]
; AVX512-NEXT:    vpshufb %xmm8, %xmm11, %xmm3
; AVX512-NEXT:    vmovdqa64 %xmm11, %xmm27
; AVX512-NEXT:    vpshufb %xmm8, %xmm15, %xmm4
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX512-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX512-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX512-NEXT:    vpblendd {{.*#+}} ymm4 = ymm3[0,1,2,3,4,5,6],ymm2[7]
; AVX512-NEXT:    vmovdqa 416(%rdi), %xmm0
; AVX512-NEXT:    vmovdqa 432(%rdi), %xmm11
; AVX512-NEXT:    vmovd {{.*#+}} xmm3 = [0,0,1,9,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX512-NEXT:    vpshufb %xmm3, %xmm11, %xmm2
; AVX512-NEXT:    vpshufb %xmm3, %xmm0, %xmm5
; AVX512-NEXT:    vmovdqa64 %xmm0, %xmm22
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm2[0],xmm5[1],xmm2[1],xmm5[2],xmm2[2],xmm5[3],xmm2[3]
; AVX512-NEXT:    vmovdqa 384(%rdi), %xmm0
; AVX512-NEXT:    vmovdqa 400(%rdi), %xmm12
; AVX512-NEXT:    vmovd {{.*#+}} xmm2 = [1,9,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX512-NEXT:    vpshufb %xmm2, %xmm12, %xmm6
; AVX512-NEXT:    vpshufb %xmm2, %xmm0, %xmm7
; AVX512-NEXT:    vmovdqa64 %xmm0, %xmm31
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
; AVX512-NEXT:    vinserti128 $1, %xmm5, %ymm0, %ymm5
; AVX512-NEXT:    vinserti128 $1, %xmm6, %ymm0, %ymm6
; AVX512-NEXT:    vpblendd {{.*#+}} ymm5 = ymm6[0,1,2,3,4],ymm5[5],ymm6[6,7]
; AVX512-NEXT:    vpblendd {{.*#+}} ymm4 = ymm5[0,1,2,3,4,5],ymm4[6,7]
; AVX512-NEXT:    vmovdqa64 %xmm18, %xmm0
; AVX512-NEXT:    vpshufb %xmm9, %xmm0, %xmm5
; AVX512-NEXT:    vmovdqa64 %xmm18, %xmm23
; AVX512-NEXT:    vpshufb %xmm9, %xmm14, %xmm6
; AVX512-NEXT:    vmovdqa64 %xmm14, %xmm26
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX512-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX512-NEXT:    vpshufb %xmm8, %xmm0, %xmm6
; AVX512-NEXT:    vmovdqa64 %xmm19, %xmm14
; AVX512-NEXT:    vpshufb %xmm8, %xmm14, %xmm7
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
; AVX512-NEXT:    vpblendd {{.*#+}} xmm5 = xmm6[0,1,2],xmm5[3]
; AVX512-NEXT:    vpsrlq $8, %zmm17, %zmm6
; AVX512-NEXT:    vpmovqb %zmm6, %xmm6
; AVX512-NEXT:    vpblendd {{.*#+}} xmm5 = xmm6[0,1],xmm5[2,3]
; AVX512-NEXT:    vinserti32x4 $2, %xmm5, %zmm0, %zmm19
; AVX512-NEXT:    vinserti64x4 $1, %ymm4, %zmm0, %zmm19 {%k1}
; AVX512-NEXT:    vmovdqa64 %xmm25, %xmm0
; AVX512-NEXT:    vpshufb %xmm9, %xmm0, %xmm4
; AVX512-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX512-NEXT:    vpshufb %xmm9, %xmm0, %xmm5
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX512-NEXT:    vmovdqa64 %xmm29, %xmm0
; AVX512-NEXT:    vpshufb %xmm8, %xmm0, %xmm5
; AVX512-NEXT:    vpshufb %xmm8, %xmm10, %xmm6
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX512-NEXT:    vinserti128 $1, %xmm4, %ymm0, %ymm4
; AVX512-NEXT:    vinserti128 $1, %xmm5, %ymm0, %ymm5
; AVX512-NEXT:    vpblendd {{.*#+}} ymm1 = ymm5[0,1,2,3,4,5,6],ymm4[7]
; AVX512-NEXT:    vmovdqa 160(%rdi), %xmm0
; AVX512-NEXT:    vmovdqa 176(%rdi), %xmm5
; AVX512-NEXT:    vpshufb %xmm3, %xmm5, %xmm4
; AVX512-NEXT:    vmovdqa64 %xmm5, %xmm25
; AVX512-NEXT:    vpshufb %xmm3, %xmm0, %xmm3
; AVX512-NEXT:    vmovdqa64 %xmm0, %xmm20
; AVX512-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
; AVX512-NEXT:    vmovdqa 128(%rdi), %xmm4
; AVX512-NEXT:    vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512-NEXT:    vmovdqa 144(%rdi), %xmm0
; AVX512-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512-NEXT:    vpshufb %xmm2, %xmm0, %xmm0
; AVX512-NEXT:    vpshufb %xmm2, %xmm4, %xmm2
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
; AVX512-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm2
; AVX512-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX512-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm2[5],ymm0[6,7]
; AVX512-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
; AVX512-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX512-NEXT:    vpshufb %xmm9, %xmm1, %xmm1
; AVX512-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX512-NEXT:    vpshufb %xmm9, %xmm2, %xmm2
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX512-NEXT:    vmovdqa64 %xmm30, %xmm7
; AVX512-NEXT:    vpshufb %xmm8, %xmm7, %xmm2
; AVX512-NEXT:    vmovdqa64 %xmm28, %xmm3
; AVX512-NEXT:    vpshufb %xmm8, %xmm3, %xmm3
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
; AVX512-NEXT:    vpblendd {{.*#+}} xmm1 = xmm2[0,1,2],xmm1[3]
; AVX512-NEXT:    vpsrlq $8, %zmm16, %zmm2
; AVX512-NEXT:    vpmovqb %zmm2, %xmm2
; AVX512-NEXT:    vpblendd {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3]
; AVX512-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX512-NEXT:    vinserti64x4 $0, %ymm0, %zmm19, %zmm0
; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpbroadcastd {{.*#+}} xmm4 = [0,0,2,10,0,0,2,10,0,0,2,10,0,0,2,10]
; AVX512-NEXT:    vmovdqa64 %xmm24, %xmm13
; AVX512-NEXT:    vmovdqa64 %xmm24, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512-NEXT:    vpshufb %xmm4, %xmm13, %xmm1
; AVX512-NEXT:    vmovdqa64 %xmm21, %xmm10
; AVX512-NEXT:    vmovdqa64 %xmm21, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512-NEXT:    vpshufb %xmm4, %xmm10, %xmm2
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX512-NEXT:    vpbroadcastd {{.*#+}} xmm1 = [2,10,0,0,2,10,0,0,2,10,0,0,2,10,0,0]
; AVX512-NEXT:    vmovdqa64 %xmm27, %xmm0
; AVX512-NEXT:    vpshufb %xmm1, %xmm0, %xmm3
; AVX512-NEXT:    vpshufb %xmm1, %xmm15, %xmm8
; AVX512-NEXT:    vmovdqa64 %xmm15, %xmm18
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm8[0],xmm3[0],xmm8[1],xmm3[1],xmm8[2],xmm3[2],xmm8[3],xmm3[3]
; AVX512-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX512-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX512-NEXT:    vpblendd {{.*#+}} ymm8 = ymm3[0,1,2,3,4,5,6],ymm2[7]
; AVX512-NEXT:    vmovd {{.*#+}} xmm3 = [0,0,2,10,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX512-NEXT:    vpshufb %xmm3, %xmm11, %xmm2
; AVX512-NEXT:    vmovdqa64 %xmm11, %xmm24
; AVX512-NEXT:    vmovdqa64 %xmm22, %xmm11
; AVX512-NEXT:    vpshufb %xmm3, %xmm11, %xmm9
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm9[0],xmm2[0],xmm9[1],xmm2[1],xmm9[2],xmm2[2],xmm9[3],xmm2[3]
; AVX512-NEXT:    vmovd {{.*#+}} xmm2 = [2,10,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX512-NEXT:    vpshufb %xmm2, %xmm12, %xmm0
; AVX512-NEXT:    vmovdqa64 %xmm12, %xmm30
; AVX512-NEXT:    vmovdqa64 %xmm31, %xmm5
; AVX512-NEXT:    vpshufb %xmm2, %xmm5, %xmm15
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm15[0],xmm0[0],xmm15[1],xmm0[1],xmm15[2],xmm0[2],xmm15[3],xmm0[3]
; AVX512-NEXT:    vinserti128 $1, %xmm9, %ymm0, %ymm9
; AVX512-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX512-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm9[5],ymm0[6,7]
; AVX512-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm8[6,7]
; AVX512-NEXT:    vmovdqa64 %xmm23, %xmm12
; AVX512-NEXT:    vmovdqa64 %xmm23, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512-NEXT:    vpshufb %xmm4, %xmm12, %xmm8
; AVX512-NEXT:    vmovdqa64 %xmm26, %xmm5
; AVX512-NEXT:    vpshufb %xmm4, %xmm5, %xmm9
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm9[0],xmm8[0],xmm9[1],xmm8[1],xmm9[2],xmm8[2],xmm9[3],xmm8[3]
; AVX512-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX512-NEXT:    vpshufb %xmm1, %xmm5, %xmm9
; AVX512-NEXT:    vpshufb %xmm1, %xmm14, %xmm15
; AVX512-NEXT:    vmovdqa64 %xmm14, %xmm21
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm15[0],xmm9[0],xmm15[1],xmm9[1],xmm15[2],xmm9[2],xmm15[3],xmm9[3]
; AVX512-NEXT:    vpblendd {{.*#+}} xmm8 = xmm9[0,1,2],xmm8[3]
; AVX512-NEXT:    vpsrlq $16, %zmm17, %zmm9
; AVX512-NEXT:    vpmovqb %zmm9, %xmm9
; AVX512-NEXT:    vpblendd {{.*#+}} xmm8 = xmm9[0,1],xmm8[2,3]
; AVX512-NEXT:    vinserti32x4 $2, %xmm8, %zmm0, %zmm8
; AVX512-NEXT:    vinserti64x4 $1, %ymm0, %zmm0, %zmm8 {%k1}
; AVX512-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX512-NEXT:    vpshufb %xmm4, %xmm0, %xmm0
; AVX512-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
; AVX512-NEXT:    vpshufb %xmm4, %xmm14, %xmm9
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm9[0],xmm0[0],xmm9[1],xmm0[1],xmm9[2],xmm0[2],xmm9[3],xmm0[3]
; AVX512-NEXT:    vmovdqa64 %xmm29, %xmm6
; AVX512-NEXT:    vpshufb %xmm1, %xmm6, %xmm9
; AVX512-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX512-NEXT:    vpshufb %xmm1, %xmm15, %xmm15
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm15[0],xmm9[0],xmm15[1],xmm9[1],xmm15[2],xmm9[2],xmm15[3],xmm9[3]
; AVX512-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX512-NEXT:    vinserti128 $1, %xmm9, %ymm0, %ymm9
; AVX512-NEXT:    vpblendd {{.*#+}} ymm0 = ymm9[0,1,2,3,4,5,6],ymm0[7]
; AVX512-NEXT:    vmovdqa64 %xmm25, %xmm9
; AVX512-NEXT:    vpshufb %xmm3, %xmm9, %xmm9
; AVX512-NEXT:    vmovdqa64 %xmm20, %xmm15
; AVX512-NEXT:    vpshufb %xmm3, %xmm15, %xmm3
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm9[0],xmm3[1],xmm9[1],xmm3[2],xmm9[2],xmm3[3],xmm9[3]
; AVX512-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
; AVX512-NEXT:    vpshufb %xmm2, %xmm9, %xmm9
; AVX512-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX512-NEXT:    vpshufb %xmm2, %xmm15, %xmm2
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm9[0],xmm2[1],xmm9[1],xmm2[2],xmm9[2],xmm2[3],xmm9[3]
; AVX512-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX512-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX512-NEXT:    vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3,4],ymm3[5],ymm2[6,7]
; AVX512-NEXT:    vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3,4,5],ymm0[6,7]
; AVX512-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX512-NEXT:    vpshufb %xmm4, %xmm2, %xmm2
; AVX512-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX512-NEXT:    vpshufb %xmm4, %xmm3, %xmm3
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
; AVX512-NEXT:    vpshufb %xmm1, %xmm7, %xmm3
; AVX512-NEXT:    vmovdqa64 %xmm7, %xmm29
; AVX512-NEXT:    vmovdqa64 %xmm28, %xmm7
; AVX512-NEXT:    vmovdqa64 %xmm28, (%rsp) # 16-byte Spill
; AVX512-NEXT:    vpshufb %xmm1, %xmm7, %xmm1
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1],xmm1[2],xmm3[2],xmm1[3],xmm3[3]
; AVX512-NEXT:    vpblendd {{.*#+}} xmm1 = xmm1[0,1,2],xmm2[3]
; AVX512-NEXT:    vpsrlq $16, %zmm16, %zmm2
; AVX512-NEXT:    vpmovqb %zmm2, %xmm2
; AVX512-NEXT:    vpblendd {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3]
; AVX512-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX512-NEXT:    vinserti64x4 $0, %ymm0, %zmm8, %zmm0
; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpbroadcastd {{.*#+}} xmm0 = [0,0,3,11,0,0,3,11,0,0,3,11,0,0,3,11]
; AVX512-NEXT:    vpshufb %xmm0, %xmm13, %xmm1
; AVX512-NEXT:    vpshufb %xmm0, %xmm10, %xmm2
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX512-NEXT:    vpbroadcastd {{.*#+}} xmm1 = [3,11,0,0,3,11,0,0,3,11,0,0,3,11,0,0]
; AVX512-NEXT:    vmovdqa64 %xmm27, %xmm10
; AVX512-NEXT:    vmovdqa64 %xmm27, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512-NEXT:    vpshufb %xmm1, %xmm10, %xmm3
; AVX512-NEXT:    vmovdqa64 %xmm18, %xmm4
; AVX512-NEXT:    vpshufb %xmm1, %xmm4, %xmm4
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX512-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX512-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX512-NEXT:    vpblendd {{.*#+}} ymm4 = ymm3[0,1,2,3,4,5,6],ymm2[7]
; AVX512-NEXT:    vmovd {{.*#+}} xmm2 = [0,0,3,11,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX512-NEXT:    vmovdqa64 %xmm24, %xmm3
; AVX512-NEXT:    vpshufb %xmm2, %xmm3, %xmm3
; AVX512-NEXT:    vpshufb %xmm2, %xmm11, %xmm8
; AVX512-NEXT:    vmovdqa64 %xmm22, %xmm28
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm3[0],xmm8[1],xmm3[1],xmm8[2],xmm3[2],xmm8[3],xmm3[3]
; AVX512-NEXT:    vmovd {{.*#+}} xmm3 = [3,11,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX512-NEXT:    vmovdqa64 %xmm30, %xmm9
; AVX512-NEXT:    vpshufb %xmm3, %xmm9, %xmm9
; AVX512-NEXT:    vmovdqa64 %xmm31, %xmm11
; AVX512-NEXT:    vpshufb %xmm3, %xmm11, %xmm15
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm15[0],xmm9[0],xmm15[1],xmm9[1],xmm15[2],xmm9[2],xmm15[3],xmm9[3]
; AVX512-NEXT:    vinserti128 $1, %xmm8, %ymm0, %ymm8
; AVX512-NEXT:    vinserti128 $1, %xmm9, %ymm0, %ymm9
; AVX512-NEXT:    vpblendd {{.*#+}} ymm8 = ymm9[0,1,2,3,4],ymm8[5],ymm9[6,7]
; AVX512-NEXT:    vpblendd {{.*#+}} ymm4 = ymm8[0,1,2,3,4,5],ymm4[6,7]
; AVX512-NEXT:    vpshufb %xmm0, %xmm12, %xmm8
; AVX512-NEXT:    vmovdqa64 %xmm26, %xmm13
; AVX512-NEXT:    vmovdqa64 %xmm26, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512-NEXT:    vpshufb %xmm0, %xmm13, %xmm9
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm9[0],xmm8[0],xmm9[1],xmm8[1],xmm9[2],xmm8[2],xmm9[3],xmm8[3]
; AVX512-NEXT:    vpshufb %xmm1, %xmm5, %xmm9
; AVX512-NEXT:    vmovdqa64 %xmm21, %xmm5
; AVX512-NEXT:    vpshufb %xmm1, %xmm5, %xmm15
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm15[0],xmm9[0],xmm15[1],xmm9[1],xmm15[2],xmm9[2],xmm15[3],xmm9[3]
; AVX512-NEXT:    vpblendd {{.*#+}} xmm8 = xmm9[0,1,2],xmm8[3]
; AVX512-NEXT:    vpsrlq $24, %zmm17, %zmm9
; AVX512-NEXT:    vpmovqb %zmm9, %xmm9
; AVX512-NEXT:    vpblendd {{.*#+}} xmm8 = xmm9[0,1],xmm8[2,3]
; AVX512-NEXT:    vinserti32x4 $2, %xmm8, %zmm0, %zmm8
; AVX512-NEXT:    vinserti64x4 $1, %ymm4, %zmm0, %zmm8 {%k1}
; AVX512-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX512-NEXT:    vpshufb %xmm0, %xmm4, %xmm4
; AVX512-NEXT:    vpshufb %xmm0, %xmm14, %xmm9
; AVX512-NEXT:    vmovdqa64 %xmm14, %xmm22
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm9[0],xmm4[0],xmm9[1],xmm4[1],xmm9[2],xmm4[2],xmm9[3],xmm4[3]
; AVX512-NEXT:    vpshufb %xmm1, %xmm6, %xmm9
; AVX512-NEXT:    vmovdqa64 %xmm6, %xmm26
; AVX512-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX512-NEXT:    vpshufb %xmm1, %xmm6, %xmm15
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm15[0],xmm9[0],xmm15[1],xmm9[1],xmm15[2],xmm9[2],xmm15[3],xmm9[3]
; AVX512-NEXT:    vinserti128 $1, %xmm4, %ymm0, %ymm4
; AVX512-NEXT:    vinserti128 $1, %xmm9, %ymm0, %ymm9
; AVX512-NEXT:    vpblendd {{.*#+}} ymm4 = ymm9[0,1,2,3,4,5,6],ymm4[7]
; AVX512-NEXT:    vmovdqa64 %xmm25, %xmm14
; AVX512-NEXT:    vpshufb %xmm2, %xmm14, %xmm9
; AVX512-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX512-NEXT:    vpshufb %xmm2, %xmm6, %xmm2
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm9[0],xmm2[1],xmm9[1],xmm2[2],xmm9[2],xmm2[3],xmm9[3]
; AVX512-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX512-NEXT:    vpshufb %xmm3, %xmm6, %xmm9
; AVX512-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX512-NEXT:    vpshufb %xmm3, %xmm6, %xmm3
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm9[0],xmm3[1],xmm9[1],xmm3[2],xmm9[2],xmm3[3],xmm9[3]
; AVX512-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX512-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX512-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4],ymm2[5],ymm3[6,7]
; AVX512-NEXT:    vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm4[6,7]
; AVX512-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX512-NEXT:    vpshufb %xmm0, %xmm3, %xmm3
; AVX512-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX512-NEXT:    vpshufb %xmm0, %xmm6, %xmm0
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3]
; AVX512-NEXT:    vmovdqa64 %xmm29, %xmm3
; AVX512-NEXT:    vpshufb %xmm1, %xmm3, %xmm3
; AVX512-NEXT:    vpshufb %xmm1, %xmm7, %xmm1
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1],xmm1[2],xmm3[2],xmm1[3],xmm3[3]
; AVX512-NEXT:    vpblendd {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[3]
; AVX512-NEXT:    vpsrlq $24, %zmm16, %zmm1
; AVX512-NEXT:    vpmovqb %zmm1, %xmm1
; AVX512-NEXT:    vpblendd {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3]
; AVX512-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
; AVX512-NEXT:    vinserti64x4 $0, %ymm0, %zmm8, %zmm0
; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpbroadcastd {{.*#+}} xmm0 = [0,0,4,12,0,0,4,12,0,0,4,12,0,0,4,12]
; AVX512-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX512-NEXT:    vpshufb %xmm0, %xmm1, %xmm1
; AVX512-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX512-NEXT:    vpshufb %xmm0, %xmm2, %xmm2
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX512-NEXT:    vpbroadcastd {{.*#+}} xmm1 = [4,12,0,0,4,12,0,0,4,12,0,0,4,12,0,0]
; AVX512-NEXT:    vpshufb %xmm1, %xmm10, %xmm3
; AVX512-NEXT:    vmovdqa64 %xmm18, %xmm10
; AVX512-NEXT:    vmovdqa64 %xmm18, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512-NEXT:    vpshufb %xmm1, %xmm10, %xmm4
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX512-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX512-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX512-NEXT:    vpblendd {{.*#+}} ymm4 = ymm3[0,1,2,3,4,5,6],ymm2[7]
; AVX512-NEXT:    vmovd {{.*#+}} xmm2 = [0,0,4,12,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX512-NEXT:    vmovdqa64 %xmm24, %xmm11
; AVX512-NEXT:    vpshufb %xmm2, %xmm11, %xmm3
; AVX512-NEXT:    vmovdqa64 %xmm28, %xmm7
; AVX512-NEXT:    vpshufb %xmm2, %xmm7, %xmm8
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm3[0],xmm8[1],xmm3[1],xmm8[2],xmm3[2],xmm8[3],xmm3[3]
; AVX512-NEXT:    vmovd {{.*#+}} xmm3 = [4,12,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX512-NEXT:    vmovdqa64 %xmm30, %xmm7
; AVX512-NEXT:    vpshufb %xmm3, %xmm7, %xmm9
; AVX512-NEXT:    vmovdqa64 %xmm31, %xmm12
; AVX512-NEXT:    vpshufb %xmm3, %xmm12, %xmm15
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm15[0],xmm9[0],xmm15[1],xmm9[1],xmm15[2],xmm9[2],xmm15[3],xmm9[3]
; AVX512-NEXT:    vinserti128 $1, %xmm8, %ymm0, %ymm8
; AVX512-NEXT:    vinserti128 $1, %xmm9, %ymm0, %ymm9
; AVX512-NEXT:    vpblendd {{.*#+}} ymm8 = ymm9[0,1,2,3,4],ymm8[5],ymm9[6,7]
; AVX512-NEXT:    vpblendd {{.*#+}} ymm4 = ymm8[0,1,2,3,4,5],ymm4[6,7]
; AVX512-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
; AVX512-NEXT:    vpshufb %xmm0, %xmm7, %xmm8
; AVX512-NEXT:    vpshufb %xmm0, %xmm13, %xmm9
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm9[0],xmm8[0],xmm9[1],xmm8[1],xmm9[2],xmm8[2],xmm9[3],xmm8[3]
; AVX512-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
; AVX512-NEXT:    vpshufb %xmm1, %xmm7, %xmm9
; AVX512-NEXT:    vpshufb %xmm1, %xmm5, %xmm15
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm15[0],xmm9[0],xmm15[1],xmm9[1],xmm15[2],xmm9[2],xmm15[3],xmm9[3]
; AVX512-NEXT:    vpblendd {{.*#+}} xmm8 = xmm9[0,1,2],xmm8[3]
; AVX512-NEXT:    vpsrlq $32, %zmm17, %zmm9
; AVX512-NEXT:    vpmovqb %zmm9, %xmm9
; AVX512-NEXT:    vpblendd {{.*#+}} xmm8 = xmm9[0,1],xmm8[2,3]
; AVX512-NEXT:    vinserti32x4 $2, %xmm8, %zmm0, %zmm8
; AVX512-NEXT:    vinserti64x4 $1, %ymm4, %zmm0, %zmm8 {%k1}
; AVX512-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
; AVX512-NEXT:    vpshufb %xmm0, %xmm13, %xmm4
; AVX512-NEXT:    vmovdqa64 %xmm22, %xmm5
; AVX512-NEXT:    vpshufb %xmm0, %xmm5, %xmm9
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm9[0],xmm4[0],xmm9[1],xmm4[1],xmm9[2],xmm4[2],xmm9[3],xmm4[3]
; AVX512-NEXT:    vmovdqa64 %xmm26, %xmm5
; AVX512-NEXT:    vpshufb %xmm1, %xmm5, %xmm9
; AVX512-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX512-NEXT:    vpshufb %xmm1, %xmm5, %xmm15
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm15[0],xmm9[0],xmm15[1],xmm9[1],xmm15[2],xmm9[2],xmm15[3],xmm9[3]
; AVX512-NEXT:    vinserti128 $1, %xmm4, %ymm0, %ymm4
; AVX512-NEXT:    vinserti128 $1, %xmm9, %ymm0, %ymm9
; AVX512-NEXT:    vpblendd {{.*#+}} ymm4 = ymm9[0,1,2,3,4,5,6],ymm4[7]
; AVX512-NEXT:    vpshufb %xmm2, %xmm14, %xmm9
; AVX512-NEXT:    vmovdqa64 %xmm25, %xmm20
; AVX512-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX512-NEXT:    vpshufb %xmm2, %xmm5, %xmm2
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm9[0],xmm2[1],xmm9[1],xmm2[2],xmm9[2],xmm2[3],xmm9[3]
; AVX512-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX512-NEXT:    vpshufb %xmm3, %xmm5, %xmm9
; AVX512-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
; AVX512-NEXT:    vpshufb %xmm3, %xmm14, %xmm3
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm9[0],xmm3[1],xmm9[1],xmm3[2],xmm9[2],xmm3[3],xmm9[3]
; AVX512-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX512-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX512-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4],ymm2[5],ymm3[6,7]
; AVX512-NEXT:    vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm4[6,7]
; AVX512-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
; AVX512-NEXT:    vpshufb %xmm0, %xmm7, %xmm3
; AVX512-NEXT:    vpshufb %xmm0, %xmm6, %xmm0
; AVX512-NEXT:    vmovdqa64 %xmm6, %xmm25
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3]
; AVX512-NEXT:    vmovdqa64 %xmm29, %xmm3
; AVX512-NEXT:    vpshufb %xmm1, %xmm3, %xmm3
; AVX512-NEXT:    vmovdqa (%rsp), %xmm4 # 16-byte Reload
; AVX512-NEXT:    vpshufb %xmm1, %xmm4, %xmm1
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1],xmm1[2],xmm3[2],xmm1[3],xmm3[3]
; AVX512-NEXT:    vpblendd {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[3]
; AVX512-NEXT:    vpsrlq $32, %zmm16, %zmm1
; AVX512-NEXT:    vmovdqa64 %zmm16, %zmm18
; AVX512-NEXT:    vpmovqb %zmm1, %xmm1
; AVX512-NEXT:    vpblendd {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3]
; AVX512-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
; AVX512-NEXT:    vinserti64x4 $0, %ymm0, %zmm8, %zmm0
; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpbroadcastd {{.*#+}} xmm0 = [0,0,5,13,0,0,5,13,0,0,5,13,0,0,5,13]
; AVX512-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX512-NEXT:    vpshufb %xmm0, %xmm1, %xmm1
; AVX512-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX512-NEXT:    vpshufb %xmm0, %xmm2, %xmm2
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX512-NEXT:    vpbroadcastd {{.*#+}} xmm1 = [5,13,0,0,5,13,0,0,5,13,0,0,5,13,0,0]
; AVX512-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX512-NEXT:    vpshufb %xmm1, %xmm3, %xmm3
; AVX512-NEXT:    vpshufb %xmm1, %xmm10, %xmm4
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX512-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX512-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX512-NEXT:    vpblendd {{.*#+}} ymm4 = ymm3[0,1,2,3,4,5,6],ymm2[7]
; AVX512-NEXT:    vmovd {{.*#+}} xmm2 = [0,0,5,13,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX512-NEXT:    vpshufb %xmm2, %xmm11, %xmm3
; AVX512-NEXT:    vmovdqa64 %xmm24, %xmm27
; AVX512-NEXT:    vmovdqa64 %xmm28, %xmm5
; AVX512-NEXT:    vpshufb %xmm2, %xmm5, %xmm8
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm3[0],xmm8[1],xmm3[1],xmm8[2],xmm3[2],xmm8[3],xmm3[3]
; AVX512-NEXT:    vmovd {{.*#+}} xmm3 = [5,13,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX512-NEXT:    vmovdqa64 %xmm30, %xmm10
; AVX512-NEXT:    vpshufb %xmm3, %xmm10, %xmm9
; AVX512-NEXT:    vpshufb %xmm3, %xmm12, %xmm15
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm15[0],xmm9[0],xmm15[1],xmm9[1],xmm15[2],xmm9[2],xmm15[3],xmm9[3]
; AVX512-NEXT:    vinserti128 $1, %xmm8, %ymm0, %ymm8
; AVX512-NEXT:    vinserti128 $1, %xmm9, %ymm0, %ymm9
; AVX512-NEXT:    vpblendd {{.*#+}} ymm8 = ymm9[0,1,2,3,4],ymm8[5],ymm9[6,7]
; AVX512-NEXT:    vpblendd {{.*#+}} ymm4 = ymm8[0,1,2,3,4,5],ymm4[6,7]
; AVX512-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX512-NEXT:    vpshufb %xmm0, %xmm6, %xmm8
; AVX512-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX512-NEXT:    vpshufb %xmm0, %xmm6, %xmm9
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm9[0],xmm8[0],xmm9[1],xmm8[1],xmm9[2],xmm8[2],xmm9[3],xmm8[3]
; AVX512-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX512-NEXT:    vpshufb %xmm1, %xmm6, %xmm9
; AVX512-NEXT:    vmovdqa64 %xmm21, %xmm11
; AVX512-NEXT:    vpshufb %xmm1, %xmm11, %xmm15
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm15[0],xmm9[0],xmm15[1],xmm9[1],xmm15[2],xmm9[2],xmm15[3],xmm9[3]
; AVX512-NEXT:    vpblendd {{.*#+}} xmm8 = xmm9[0,1,2],xmm8[3]
; AVX512-NEXT:    vpsrlq $40, %zmm17, %zmm9
; AVX512-NEXT:    vmovdqa64 %zmm17, %zmm23
; AVX512-NEXT:    vpmovqb %zmm9, %xmm9
; AVX512-NEXT:    vpblendd {{.*#+}} xmm8 = xmm9[0,1],xmm8[2,3]
; AVX512-NEXT:    vinserti32x4 $2, %xmm8, %zmm0, %zmm8
; AVX512-NEXT:    vinserti64x4 $1, %ymm4, %zmm0, %zmm8 {%k1}
; AVX512-NEXT:    vpshufb %xmm0, %xmm13, %xmm4
; AVX512-NEXT:    vmovdqa64 %xmm13, %xmm30
; AVX512-NEXT:    vmovdqa64 %xmm22, %xmm6
; AVX512-NEXT:    vpshufb %xmm0, %xmm6, %xmm9
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm9[0],xmm4[0],xmm9[1],xmm4[1],xmm9[2],xmm4[2],xmm9[3],xmm4[3]
; AVX512-NEXT:    vmovdqa64 %xmm26, %xmm13
; AVX512-NEXT:    vpshufb %xmm1, %xmm13, %xmm9
; AVX512-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX512-NEXT:    vpshufb %xmm1, %xmm6, %xmm15
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm15[0],xmm9[0],xmm15[1],xmm9[1],xmm15[2],xmm9[2],xmm15[3],xmm9[3]
; AVX512-NEXT:    vinserti128 $1, %xmm4, %ymm0, %ymm4
; AVX512-NEXT:    vinserti128 $1, %xmm9, %ymm0, %ymm9
; AVX512-NEXT:    vpblendd {{.*#+}} ymm4 = ymm9[0,1,2,3,4,5,6],ymm4[7]
; AVX512-NEXT:    vmovdqa64 %xmm20, %xmm9
; AVX512-NEXT:    vpshufb %xmm2, %xmm9, %xmm9
; AVX512-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload
; AVX512-NEXT:    vpshufb %xmm2, %xmm12, %xmm2
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm9[0],xmm2[1],xmm9[1],xmm2[2],xmm9[2],xmm2[3],xmm9[3]
; AVX512-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX512-NEXT:    vpshufb %xmm3, %xmm15, %xmm9
; AVX512-NEXT:    vmovdqa64 %xmm15, %xmm28
; AVX512-NEXT:    vpshufb %xmm3, %xmm14, %xmm3
; AVX512-NEXT:    vmovdqa64 %xmm14, %xmm16
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm9[0],xmm3[1],xmm9[1],xmm3[2],xmm9[2],xmm3[3],xmm9[3]
; AVX512-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX512-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX512-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4],ymm2[5],ymm3[6,7]
; AVX512-NEXT:    vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm4[6,7]
; AVX512-NEXT:    vpshufb %xmm0, %xmm7, %xmm3
; AVX512-NEXT:    vmovdqa64 %xmm7, %xmm24
; AVX512-NEXT:    vmovdqa64 %xmm25, %xmm4
; AVX512-NEXT:    vpshufb %xmm0, %xmm4, %xmm0
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3]
; AVX512-NEXT:    vmovdqa64 %xmm29, %xmm7
; AVX512-NEXT:    vpshufb %xmm1, %xmm7, %xmm3
; AVX512-NEXT:    vmovdqa (%rsp), %xmm4 # 16-byte Reload
; AVX512-NEXT:    vpshufb %xmm1, %xmm4, %xmm1
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1],xmm1[2],xmm3[2],xmm1[3],xmm3[3]
; AVX512-NEXT:    vpblendd {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[3]
; AVX512-NEXT:    vpsrlq $40, %zmm18, %zmm1
; AVX512-NEXT:    vmovdqa64 %zmm18, %zmm26
; AVX512-NEXT:    vpmovqb %zmm1, %xmm1
; AVX512-NEXT:    vpblendd {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3]
; AVX512-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
; AVX512-NEXT:    vinserti64x4 $0, %ymm0, %zmm8, %zmm0
; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-NEXT:    vpbroadcastd {{.*#+}} xmm0 = [0,0,6,14,0,0,6,14,0,0,6,14,0,0,6,14]
; AVX512-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX512-NEXT:    vpshufb %xmm0, %xmm2, %xmm1
; AVX512-NEXT:    vmovdqa64 %xmm2, %xmm29
; AVX512-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX512-NEXT:    vpshufb %xmm0, %xmm2, %xmm2
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX512-NEXT:    vpbroadcastd {{.*#+}} xmm1 = [6,14,0,0,6,14,0,0,6,14,0,0,6,14,0,0]
; AVX512-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX512-NEXT:    vpshufb %xmm1, %xmm3, %xmm3
; AVX512-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX512-NEXT:    vpshufb %xmm1, %xmm4, %xmm4
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX512-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX512-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX512-NEXT:    vpblendd {{.*#+}} ymm4 = ymm3[0,1,2,3,4,5,6],ymm2[7]
; AVX512-NEXT:    vmovd {{.*#+}} xmm2 = [0,0,6,14,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX512-NEXT:    vmovdqa64 %xmm27, %xmm3
; AVX512-NEXT:    vpshufb %xmm2, %xmm3, %xmm3
; AVX512-NEXT:    vpshufb %xmm2, %xmm5, %xmm8
; AVX512-NEXT:    vmovdqa64 %xmm5, %xmm21
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm3[0],xmm8[1],xmm3[1],xmm8[2],xmm3[2],xmm8[3],xmm3[3]
; AVX512-NEXT:    vmovd {{.*#+}} xmm3 = [6,14,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX512-NEXT:    vpshufb %xmm3, %xmm10, %xmm9
; AVX512-NEXT:    vmovdqa64 %xmm10, %xmm18
; AVX512-NEXT:    vmovdqa64 %xmm31, %xmm5
; AVX512-NEXT:    vpshufb %xmm3, %xmm5, %xmm15
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm15[0],xmm9[0],xmm15[1],xmm9[1],xmm15[2],xmm9[2],xmm15[3],xmm9[3]
; AVX512-NEXT:    vinserti128 $1, %xmm8, %ymm0, %ymm8
; AVX512-NEXT:    vinserti128 $1, %xmm9, %ymm0, %ymm9
; AVX512-NEXT:    vpblendd {{.*#+}} ymm8 = ymm9[0,1,2,3,4],ymm8[5],ymm9[6,7]
; AVX512-NEXT:    vpblendd {{.*#+}} ymm4 = ymm8[0,1,2,3,4,5],ymm4[6,7]
; AVX512-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX512-NEXT:    vpshufb %xmm0, %xmm5, %xmm8
; AVX512-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
; AVX512-NEXT:    vpshufb %xmm0, %xmm9, %xmm9
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm9[0],xmm8[0],xmm9[1],xmm8[1],xmm9[2],xmm8[2],xmm9[3],xmm8[3]
; AVX512-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; AVX512-NEXT:    vpshufb %xmm1, %xmm10, %xmm9
; AVX512-NEXT:    vpshufb %xmm1, %xmm11, %xmm15
; AVX512-NEXT:    vmovdqa64 %xmm11, %xmm22
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm15[0],xmm9[0],xmm15[1],xmm9[1],xmm15[2],xmm9[2],xmm15[3],xmm9[3]
; AVX512-NEXT:    vpblendd {{.*#+}} xmm8 = xmm9[0,1,2],xmm8[3]
; AVX512-NEXT:    vpsrlq $48, %zmm17, %zmm9
; AVX512-NEXT:    vpmovqb %zmm9, %xmm9
; AVX512-NEXT:    vpblendd {{.*#+}} xmm8 = xmm9[0,1],xmm8[2,3]
; AVX512-NEXT:    vinserti32x4 $2, %xmm8, %zmm0, %zmm8
; AVX512-NEXT:    vinserti64x4 $1, %ymm4, %zmm0, %zmm8 {%k1}
; AVX512-NEXT:    vmovdqa64 %xmm30, %xmm4
; AVX512-NEXT:    vpshufb %xmm0, %xmm4, %xmm4
; AVX512-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
; AVX512-NEXT:    vpshufb %xmm0, %xmm9, %xmm9
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm9[0],xmm4[0],xmm9[1],xmm4[1],xmm9[2],xmm4[2],xmm9[3],xmm4[3]
; AVX512-NEXT:    vpshufb %xmm1, %xmm13, %xmm9
; AVX512-NEXT:    vmovdqa64 %xmm13, %xmm19
; AVX512-NEXT:    vpshufb %xmm1, %xmm6, %xmm15
; AVX512-NEXT:    vmovdqa64 %xmm6, %xmm17
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm15[0],xmm9[0],xmm15[1],xmm9[1],xmm15[2],xmm9[2],xmm15[3],xmm9[3]
; AVX512-NEXT:    vinserti128 $1, %xmm4, %ymm0, %ymm4
; AVX512-NEXT:    vinserti128 $1, %xmm9, %ymm0, %ymm9
; AVX512-NEXT:    vpblendd {{.*#+}} ymm4 = ymm9[0,1,2,3,4,5,6],ymm4[7]
; AVX512-NEXT:    vmovdqa64 %xmm20, %xmm6
; AVX512-NEXT:    vpshufb %xmm2, %xmm6, %xmm9
; AVX512-NEXT:    vpshufb %xmm2, %xmm12, %xmm2
; AVX512-NEXT:    vmovdqa64 %xmm12, %xmm20
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm9[0],xmm2[1],xmm9[1],xmm2[2],xmm9[2],xmm2[3],xmm9[3]
; AVX512-NEXT:    vmovdqa64 %xmm28, %xmm14
; AVX512-NEXT:    vpshufb %xmm3, %xmm14, %xmm9
; AVX512-NEXT:    vmovdqa64 %xmm16, %xmm13
; AVX512-NEXT:    vpshufb %xmm3, %xmm13, %xmm3
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm9[0],xmm3[1],xmm9[1],xmm3[2],xmm9[2],xmm3[3],xmm9[3]
; AVX512-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX512-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX512-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4],ymm2[5],ymm3[6,7]
; AVX512-NEXT:    vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm4[6,7]
; AVX512-NEXT:    vmovdqa64 %xmm24, %xmm15
; AVX512-NEXT:    vpshufb %xmm0, %xmm15, %xmm3
; AVX512-NEXT:    vmovdqa64 %xmm25, %xmm12
; AVX512-NEXT:    vpshufb %xmm0, %xmm12, %xmm0
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3]
; AVX512-NEXT:    vpshufb %xmm1, %xmm7, %xmm3
; AVX512-NEXT:    vmovdqa64 %xmm7, %xmm16
; AVX512-NEXT:    vmovdqa (%rsp), %xmm7 # 16-byte Reload
; AVX512-NEXT:    vpshufb %xmm1, %xmm7, %xmm1
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1],xmm1[2],xmm3[2],xmm1[3],xmm3[3]
; AVX512-NEXT:    vpblendd {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[3]
; AVX512-NEXT:    vpsrlq $48, %zmm26, %zmm1
; AVX512-NEXT:    vpmovqb %zmm1, %xmm1
; AVX512-NEXT:    vpblendd {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3]
; AVX512-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
; AVX512-NEXT:    vinserti64x4 $0, %ymm0, %zmm8, %zmm24
; AVX512-NEXT:    vpbroadcastd {{.*#+}} xmm0 = [0,0,7,15,0,0,7,15,0,0,7,15,0,0,7,15]
; AVX512-NEXT:    vmovdqa64 %xmm29, %xmm1
; AVX512-NEXT:    vpshufb %xmm0, %xmm1, %xmm1
; AVX512-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX512-NEXT:    vpshufb %xmm0, %xmm2, %xmm2
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX512-NEXT:    vpbroadcastd {{.*#+}} xmm1 = [7,15,0,0,7,15,0,0,7,15,0,0,7,15,0,0]
; AVX512-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX512-NEXT:    vpshufb %xmm1, %xmm3, %xmm3
; AVX512-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX512-NEXT:    vpshufb %xmm1, %xmm4, %xmm4
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX512-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX512-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX512-NEXT:    vpblendd {{.*#+}} ymm4 = ymm3[0,1,2,3,4,5,6],ymm2[7]
; AVX512-NEXT:    vmovd {{.*#+}} xmm2 = [0,0,7,15,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX512-NEXT:    vmovdqa64 %xmm27, %xmm3
; AVX512-NEXT:    vpshufb %xmm2, %xmm3, %xmm3
; AVX512-NEXT:    vmovdqa64 %xmm21, %xmm8
; AVX512-NEXT:    vpshufb %xmm2, %xmm8, %xmm8
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm3[0],xmm8[1],xmm3[1],xmm8[2],xmm3[2],xmm8[3],xmm3[3]
; AVX512-NEXT:    vmovd {{.*#+}} xmm3 = [7,15,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX512-NEXT:    vmovdqa64 %xmm18, %xmm9
; AVX512-NEXT:    vpshufb %xmm3, %xmm9, %xmm9
; AVX512-NEXT:    vmovdqa64 %xmm31, %xmm11
; AVX512-NEXT:    vpshufb %xmm3, %xmm11, %xmm11
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm11[0],xmm9[0],xmm11[1],xmm9[1],xmm11[2],xmm9[2],xmm11[3],xmm9[3]
; AVX512-NEXT:    vinserti128 $1, %xmm8, %ymm0, %ymm8
; AVX512-NEXT:    vinserti128 $1, %xmm9, %ymm0, %ymm9
; AVX512-NEXT:    vpblendd {{.*#+}} ymm8 = ymm9[0,1,2,3,4],ymm8[5],ymm9[6,7]
; AVX512-NEXT:    vpblendd {{.*#+}} ymm4 = ymm8[0,1,2,3,4,5],ymm4[6,7]
; AVX512-NEXT:    vpshufb %xmm0, %xmm5, %xmm8
; AVX512-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX512-NEXT:    vpshufb %xmm0, %xmm5, %xmm9
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm9[0],xmm8[0],xmm9[1],xmm8[1],xmm9[2],xmm8[2],xmm9[3],xmm8[3]
; AVX512-NEXT:    vpshufb %xmm1, %xmm10, %xmm9
; AVX512-NEXT:    vmovdqa64 %xmm22, %xmm5
; AVX512-NEXT:    vpshufb %xmm1, %xmm5, %xmm11
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm11[0],xmm9[0],xmm11[1],xmm9[1],xmm11[2],xmm9[2],xmm11[3],xmm9[3]
; AVX512-NEXT:    vpblendd {{.*#+}} xmm8 = xmm9[0,1,2],xmm8[3]
; AVX512-NEXT:    vpsrlq $56, %zmm23, %zmm9
; AVX512-NEXT:    vpmovqb %zmm9, %xmm9
; AVX512-NEXT:    vpblendd {{.*#+}} xmm8 = xmm9[0,1],xmm8[2,3]
; AVX512-NEXT:    vinserti32x4 $2, %xmm8, %zmm0, %zmm8
; AVX512-NEXT:    vinserti64x4 $1, %ymm4, %zmm0, %zmm8 {%k1}
; AVX512-NEXT:    vmovdqa64 %xmm30, %xmm4
; AVX512-NEXT:    vpshufb %xmm0, %xmm4, %xmm4
; AVX512-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX512-NEXT:    vpshufb %xmm0, %xmm5, %xmm9
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm9[0],xmm4[0],xmm9[1],xmm4[1],xmm9[2],xmm4[2],xmm9[3],xmm4[3]
; AVX512-NEXT:    vmovdqa64 %xmm19, %xmm5
; AVX512-NEXT:    vpshufb %xmm1, %xmm5, %xmm9
; AVX512-NEXT:    vmovdqa64 %xmm17, %xmm5
; AVX512-NEXT:    vpshufb %xmm1, %xmm5, %xmm11
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm11[0],xmm9[0],xmm11[1],xmm9[1],xmm11[2],xmm9[2],xmm11[3],xmm9[3]
; AVX512-NEXT:    vinserti128 $1, %xmm4, %ymm0, %ymm4
; AVX512-NEXT:    vinserti128 $1, %xmm9, %ymm0, %ymm9
; AVX512-NEXT:    vpblendd {{.*#+}} ymm4 = ymm9[0,1,2,3,4,5,6],ymm4[7]
; AVX512-NEXT:    vpshufb %xmm2, %xmm6, %xmm6
; AVX512-NEXT:    vmovdqa64 %xmm20, %xmm5
; AVX512-NEXT:    vpshufb %xmm2, %xmm5, %xmm2
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm6[0],xmm2[1],xmm6[1],xmm2[2],xmm6[2],xmm2[3],xmm6[3]
; AVX512-NEXT:    vpshufb %xmm3, %xmm14, %xmm5
; AVX512-NEXT:    vpshufb %xmm3, %xmm13, %xmm3
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm5[0],xmm3[1],xmm5[1],xmm3[2],xmm5[2],xmm3[3],xmm5[3]
; AVX512-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX512-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX512-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4],ymm2[5],ymm3[6,7]
; AVX512-NEXT:    vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm4[6,7]
; AVX512-NEXT:    vpshufb %xmm0, %xmm15, %xmm3
; AVX512-NEXT:    vpshufb %xmm0, %xmm12, %xmm0
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3]
; AVX512-NEXT:    vmovdqa64 %xmm16, %xmm3
; AVX512-NEXT:    vpshufb %xmm1, %xmm3, %xmm3
; AVX512-NEXT:    vpshufb %xmm1, %xmm7, %xmm1
; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1],xmm1[2],xmm3[2],xmm1[3],xmm3[3]
; AVX512-NEXT:    vpblendd {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[3]
; AVX512-NEXT:    vpsrlq $56, %zmm26, %zmm1
; AVX512-NEXT:    vpmovqb %zmm1, %xmm1
; AVX512-NEXT:    vpblendd {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3]
; AVX512-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
; AVX512-NEXT:    vinserti64x4 $0, %ymm0, %zmm8, %zmm0
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm1, (%rsi)
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm1, (%rdx)
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm1, (%rcx)
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm1, (%r8)
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm1, (%r9)
; AVX512-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512-NEXT:    vmovaps %zmm1, (%rax)
; AVX512-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512-NEXT:    vmovdqa64 %zmm24, (%rax)
; AVX512-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512-NEXT:    vmovdqa64 %zmm0, (%rax)
; AVX512-NEXT:    addq $520, %rsp # imm = 0x208
; AVX512-NEXT:    vzeroupper
; AVX512-NEXT:    retq
;
; AVX512-FCP-LABEL: load_i8_stride8_vf64:
; AVX512-FCP:       # %bb.0:
; AVX512-FCP-NEXT:    subq $440, %rsp # imm = 0x1B8
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm2 = [0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12]
; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm0 = [0,2,2,3,0,2,4,6]
; AVX512-FCP-NEXT:    vmovdqa 480(%rdi), %ymm1
; AVX512-FCP-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512-FCP-NEXT:    vpermd %ymm1, %ymm0, %ymm3
; AVX512-FCP-NEXT:    vpshufb %ymm2, %ymm3, %ymm1
; AVX512-FCP-NEXT:    vmovdqa64 %ymm3, %ymm19
; AVX512-FCP-NEXT:    vmovdqa %ymm2, %ymm14
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm3 = [0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12]
; AVX512-FCP-NEXT:    vmovdqa 448(%rdi), %ymm2
; AVX512-FCP-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512-FCP-NEXT:    vpermd %ymm2, %ymm0, %ymm4
; AVX512-FCP-NEXT:    vpshufb %ymm3, %ymm4, %ymm2
; AVX512-FCP-NEXT:    vmovdqa %ymm3, %ymm9
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5,6],ymm1[7]
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm2 = [0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12]
; AVX512-FCP-NEXT:    vmovdqa 416(%rdi), %ymm3
; AVX512-FCP-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512-FCP-NEXT:    vpermd %ymm3, %ymm0, %ymm12
; AVX512-FCP-NEXT:    vpshufb %ymm2, %ymm12, %ymm3
; AVX512-FCP-NEXT:    vmovdqa %ymm2, %ymm11
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm10 = [0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12]
; AVX512-FCP-NEXT:    vmovdqa 384(%rdi), %ymm5
; AVX512-FCP-NEXT:    vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512-FCP-NEXT:    vpermd %ymm5, %ymm0, %ymm13
; AVX512-FCP-NEXT:    vpshufb %ymm10, %ymm13, %ymm5
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm3 = ymm5[0,1,2,3,4],ymm3[5],ymm5[6,7]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm3[0,1,2,3,4,5],ymm1[6,7]
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm3 = [0,0,0,8,0,0,0,8,0,0,0,8,0,0,0,8]
; AVX512-FCP-NEXT:    vmovdqa 368(%rdi), %xmm2
; AVX512-FCP-NEXT:    vpshufb %xmm3, %xmm2, %xmm5
; AVX512-FCP-NEXT:    vmovdqa64 %xmm2, %xmm30
; AVX512-FCP-NEXT:    vmovdqa 352(%rdi), %xmm2
; AVX512-FCP-NEXT:    vpshufb %xmm3, %xmm2, %xmm6
; AVX512-FCP-NEXT:    vmovdqa64 %xmm2, %xmm25
; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm5 = [0,8,0,0,0,8,0,0,0,8,0,0,0,8,0,0]
; AVX512-FCP-NEXT:    vmovdqa 336(%rdi), %xmm2
; AVX512-FCP-NEXT:    vpshufb %xmm5, %xmm2, %xmm7
; AVX512-FCP-NEXT:    vmovdqa64 %xmm2, %xmm24
; AVX512-FCP-NEXT:    vmovdqa 320(%rdi), %xmm2
; AVX512-FCP-NEXT:    vpshufb %xmm5, %xmm2, %xmm8
; AVX512-FCP-NEXT:    vmovdqa %xmm2, %xmm15
; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm7 = xmm8[0],xmm7[0],xmm8[1],xmm7[1],xmm8[2],xmm7[2],xmm8[3],xmm7[3]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} xmm6 = xmm7[0,1,2],xmm6[3]
; AVX512-FCP-NEXT:    vmovdqa64 256(%rdi), %zmm2
; AVX512-FCP-NEXT:    vpmovqb %zmm2, %xmm7
; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, %zmm28
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} xmm6 = xmm7[0,1],xmm6[2,3]
; AVX512-FCP-NEXT:    vinserti32x4 $2, %xmm6, %zmm0, %zmm16
; AVX512-FCP-NEXT:    movb $-64, %al
; AVX512-FCP-NEXT:    kmovw %eax, %k1
; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm16 {%k1}
; AVX512-FCP-NEXT:    vmovdqa 224(%rdi), %ymm1
; AVX512-FCP-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512-FCP-NEXT:    vpermd %ymm1, %ymm0, %ymm2
; AVX512-FCP-NEXT:    vpshufb %ymm14, %ymm2, %ymm1
; AVX512-FCP-NEXT:    vmovdqa64 %ymm2, %ymm23
; AVX512-FCP-NEXT:    vmovdqa 192(%rdi), %ymm2
; AVX512-FCP-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512-FCP-NEXT:    vpermd %ymm2, %ymm0, %ymm2
; AVX512-FCP-NEXT:    vpshufb %ymm9, %ymm2, %ymm6
; AVX512-FCP-NEXT:    vmovdqa64 %ymm2, %ymm27
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm6 = ymm6[0,1,2,3,4,5,6],ymm1[7]
; AVX512-FCP-NEXT:    vmovdqa 160(%rdi), %ymm1
; AVX512-FCP-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512-FCP-NEXT:    vmovdqa 128(%rdi), %ymm2
; AVX512-FCP-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512-FCP-NEXT:    vpermd %ymm1, %ymm0, %ymm1
; AVX512-FCP-NEXT:    vpshufb %ymm11, %ymm1, %ymm7
; AVX512-FCP-NEXT:    vmovdqa64 %ymm1, %ymm18
; AVX512-FCP-NEXT:    vpermd %ymm2, %ymm0, %ymm0
; AVX512-FCP-NEXT:    vpshufb %ymm10, %ymm0, %ymm8
; AVX512-FCP-NEXT:    vmovdqa64 %ymm0, %ymm17
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm7 = ymm8[0,1,2,3,4],ymm7[5],ymm8[6,7]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm7[0,1,2,3,4,5],ymm6[6,7]
; AVX512-FCP-NEXT:    vmovdqa 112(%rdi), %xmm10
; AVX512-FCP-NEXT:    vpshufb %xmm3, %xmm10, %xmm7
; AVX512-FCP-NEXT:    vmovdqa 96(%rdi), %xmm0
; AVX512-FCP-NEXT:    vpshufb %xmm3, %xmm0, %xmm3
; AVX512-FCP-NEXT:    vmovdqa64 %xmm0, %xmm20
; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm7[0],xmm3[1],xmm7[1],xmm3[2],xmm7[2],xmm3[3],xmm7[3]
; AVX512-FCP-NEXT:    vmovdqa 80(%rdi), %xmm2
; AVX512-FCP-NEXT:    vpshufb %xmm5, %xmm2, %xmm0
; AVX512-FCP-NEXT:    vmovdqa64 %xmm2, %xmm31
; AVX512-FCP-NEXT:    vmovdqa 64(%rdi), %xmm2
; AVX512-FCP-NEXT:    vpshufb %xmm5, %xmm2, %xmm5
; AVX512-FCP-NEXT:    vmovdqa64 %xmm2, %xmm22
; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm5[0],xmm0[0],xmm5[1],xmm0[1],xmm5[2],xmm0[2],xmm5[3],xmm0[3]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm0[0,1,2],xmm3[3]
; AVX512-FCP-NEXT:    vmovdqa64 (%rdi), %zmm29
; AVX512-FCP-NEXT:    vpmovqb %zmm29, %xmm3
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm3[0,1],xmm0[2,3]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm16, %zmm0
; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm6 = [1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13]
; AVX512-FCP-NEXT:    vmovdqa64 %ymm19, %ymm11
; AVX512-FCP-NEXT:    vpshufb %ymm6, %ymm11, %ymm0
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm7 = [1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13]
; AVX512-FCP-NEXT:    vpshufb %ymm7, %ymm4, %ymm1
; AVX512-FCP-NEXT:    vmovdqa64 %ymm4, %ymm26
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5,6],ymm0[7]
; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm1 = ymm12[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,17,21,25,29,u,u,u,u,u,u,u,u]
; AVX512-FCP-NEXT:    vmovdqa64 %ymm12, %ymm19
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm2 = [1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13]
; AVX512-FCP-NEXT:    vpshufb %ymm2, %ymm13, %ymm3
; AVX512-FCP-NEXT:    vmovdqa64 %ymm13, %ymm21
; AVX512-FCP-NEXT:    vmovdqa64 %ymm2, %ymm16
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm3[0,1,2,3,4],ymm1[5],ymm3[6,7]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm2 = ymm1[0,1,2,3,4,5],ymm0[6,7]
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm1 = [0,0,1,9,0,0,1,9,0,0,1,9,0,0,1,9]
; AVX512-FCP-NEXT:    vmovdqa64 %xmm30, %xmm13
; AVX512-FCP-NEXT:    vpshufb %xmm1, %xmm13, %xmm3
; AVX512-FCP-NEXT:    vmovdqa64 %xmm25, %xmm14
; AVX512-FCP-NEXT:    vpshufb %xmm1, %xmm14, %xmm5
; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm5[0],xmm3[0],xmm5[1],xmm3[1],xmm5[2],xmm3[2],xmm5[3],xmm3[3]
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm5 = [1,9,0,0,1,9,0,0,1,9,0,0,1,9,0,0]
; AVX512-FCP-NEXT:    vmovdqa64 %xmm24, %xmm9
; AVX512-FCP-NEXT:    vpshufb %xmm5, %xmm9, %xmm0
; AVX512-FCP-NEXT:    vmovdqa %xmm15, %xmm12
; AVX512-FCP-NEXT:    vpshufb %xmm5, %xmm15, %xmm15
; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm15[0],xmm0[0],xmm15[1],xmm0[1],xmm15[2],xmm0[2],xmm15[3],xmm0[3]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm0[0,1,2],xmm3[3]
; AVX512-FCP-NEXT:    vpsrlq $8, %zmm28, %zmm3
; AVX512-FCP-NEXT:    vpmovqb %zmm3, %xmm3
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm3[0,1],xmm0[2,3]
; AVX512-FCP-NEXT:    vinserti32x4 $2, %xmm0, %zmm0, %zmm0
; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm2, %zmm0, %zmm0 {%k1}
; AVX512-FCP-NEXT:    vmovdqa64 %ymm23, %ymm8
; AVX512-FCP-NEXT:    vpshufb %ymm6, %ymm8, %ymm2
; AVX512-FCP-NEXT:    vmovdqa64 %ymm27, %ymm4
; AVX512-FCP-NEXT:    vpshufb %ymm7, %ymm4, %ymm3
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4,5,6],ymm2[7]
; AVX512-FCP-NEXT:    vmovdqa64 %ymm18, %ymm7
; AVX512-FCP-NEXT:    vpshufb {{.*#+}} ymm3 = ymm7[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,17,21,25,29,u,u,u,u,u,u,u,u]
; AVX512-FCP-NEXT:    vmovdqa64 %ymm17, %ymm6
; AVX512-FCP-NEXT:    vmovdqa64 %ymm16, %ymm15
; AVX512-FCP-NEXT:    vpshufb %ymm15, %ymm6, %ymm15
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm3 = ymm15[0,1,2,3,4],ymm3[5],ymm15[6,7]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4,5],ymm2[6,7]
; AVX512-FCP-NEXT:    vpshufb %xmm1, %xmm10, %xmm3
; AVX512-FCP-NEXT:    vmovdqa64 %xmm10, %xmm16
; AVX512-FCP-NEXT:    vmovdqa64 %xmm20, %xmm10
; AVX512-FCP-NEXT:    vpshufb %xmm1, %xmm10, %xmm1
; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1],xmm1[2],xmm3[2],xmm1[3],xmm3[3]
; AVX512-FCP-NEXT:    vmovdqa64 %xmm31, %xmm15
; AVX512-FCP-NEXT:    vpshufb %xmm5, %xmm15, %xmm3
; AVX512-FCP-NEXT:    vmovdqa64 %xmm22, %xmm15
; AVX512-FCP-NEXT:    vpshufb %xmm5, %xmm15, %xmm5
; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm5[0],xmm3[0],xmm5[1],xmm3[1],xmm5[2],xmm3[2],xmm5[3],xmm3[3]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} xmm1 = xmm3[0,1,2],xmm1[3]
; AVX512-FCP-NEXT:    vpsrlq $8, %zmm29, %zmm3
; AVX512-FCP-NEXT:    vpmovqb %zmm3, %xmm3
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} xmm1 = xmm3[0,1],xmm1[2,3]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm0, %zmm0
; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm1 = [2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14]
; AVX512-FCP-NEXT:    vpshufb %ymm1, %ymm11, %ymm0
; AVX512-FCP-NEXT:    vmovdqa64 %ymm11, %ymm20
; AVX512-FCP-NEXT:    vmovdqa64 %ymm1, %ymm23
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm2 = [2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14]
; AVX512-FCP-NEXT:    vmovdqa64 %ymm26, %ymm11
; AVX512-FCP-NEXT:    vpshufb %ymm2, %ymm11, %ymm1
; AVX512-FCP-NEXT:    vmovdqa64 %ymm2, %ymm24
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5,6],ymm0[7]
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm2 = [2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14]
; AVX512-FCP-NEXT:    vmovdqa64 %ymm19, %ymm1
; AVX512-FCP-NEXT:    vpshufb %ymm2, %ymm1, %ymm1
; AVX512-FCP-NEXT:    vmovdqa64 %ymm2, %ymm25
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm3 = [2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14]
; AVX512-FCP-NEXT:    vmovdqa64 %ymm21, %ymm2
; AVX512-FCP-NEXT:    vpshufb %ymm3, %ymm2, %ymm2
; AVX512-FCP-NEXT:    vmovdqa64 %ymm3, %ymm26
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3,4],ymm1[5],ymm2[6,7]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm1 = [0,0,2,10,0,0,2,10,0,0,2,10,0,0,2,10]
; AVX512-FCP-NEXT:    vpshufb %xmm1, %xmm13, %xmm2
; AVX512-FCP-NEXT:    vpshufb %xmm1, %xmm14, %xmm3
; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm3 = [2,10,0,0,2,10,0,0,2,10,0,0,2,10,0,0]
; AVX512-FCP-NEXT:    vpshufb %xmm3, %xmm9, %xmm5
; AVX512-FCP-NEXT:    vmovdqa64 %xmm9, %xmm17
; AVX512-FCP-NEXT:    vpshufb %xmm3, %xmm12, %xmm15
; AVX512-FCP-NEXT:    vmovdqa %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm15[0],xmm5[0],xmm15[1],xmm5[1],xmm15[2],xmm5[2],xmm15[3],xmm5[3]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} xmm2 = xmm5[0,1,2],xmm2[3]
; AVX512-FCP-NEXT:    vpsrlq $16, %zmm28, %zmm5
; AVX512-FCP-NEXT:    vpmovqb %zmm5, %xmm5
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} xmm2 = xmm5[0,1],xmm2[2,3]
; AVX512-FCP-NEXT:    vinserti32x4 $2, %xmm2, %zmm0, %zmm2
; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm0, %zmm0, %zmm2 {%k1}
; AVX512-FCP-NEXT:    vmovdqa64 %ymm23, %ymm0
; AVX512-FCP-NEXT:    vpshufb %ymm0, %ymm8, %ymm0
; AVX512-FCP-NEXT:    vmovdqa %ymm8, %ymm9
; AVX512-FCP-NEXT:    vmovdqa64 %ymm24, %ymm5
; AVX512-FCP-NEXT:    vpshufb %ymm5, %ymm4, %ymm5
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm5[0,1,2,3,4,5,6],ymm0[7]
; AVX512-FCP-NEXT:    vmovdqa64 %ymm25, %ymm4
; AVX512-FCP-NEXT:    vpshufb %ymm4, %ymm7, %ymm5
; AVX512-FCP-NEXT:    vmovdqa64 %ymm18, %ymm25
; AVX512-FCP-NEXT:    vmovdqa64 %ymm26, %ymm4
; AVX512-FCP-NEXT:    vpshufb %ymm4, %ymm6, %ymm15
; AVX512-FCP-NEXT:    vmovdqa64 %ymm6, %ymm26
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm5 = ymm15[0,1,2,3,4],ymm5[5],ymm15[6,7]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm5[0,1,2,3,4,5],ymm0[6,7]
; AVX512-FCP-NEXT:    vmovdqa64 %xmm16, %xmm6
; AVX512-FCP-NEXT:    vpshufb %xmm1, %xmm6, %xmm5
; AVX512-FCP-NEXT:    vpshufb %xmm1, %xmm10, %xmm1
; AVX512-FCP-NEXT:    vmovdqa64 %xmm10, %xmm16
; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm5[0],xmm1[1],xmm5[1],xmm1[2],xmm5[2],xmm1[3],xmm5[3]
; AVX512-FCP-NEXT:    vmovdqa64 %xmm31, %xmm7
; AVX512-FCP-NEXT:    vpshufb %xmm3, %xmm7, %xmm5
; AVX512-FCP-NEXT:    vmovdqa64 %xmm22, %xmm8
; AVX512-FCP-NEXT:    vpshufb %xmm3, %xmm8, %xmm3
; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm5[0],xmm3[1],xmm5[1],xmm3[2],xmm5[2],xmm3[3],xmm5[3]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} xmm1 = xmm3[0,1,2],xmm1[3]
; AVX512-FCP-NEXT:    vpsrlq $16, %zmm29, %zmm3
; AVX512-FCP-NEXT:    vpmovqb %zmm3, %xmm3
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} xmm1 = xmm3[0,1],xmm1[2,3]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm2, %zmm0
; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm1 = [3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15]
; AVX512-FCP-NEXT:    vmovdqa64 %ymm20, %ymm0
; AVX512-FCP-NEXT:    vpshufb %ymm1, %ymm0, %ymm0
; AVX512-FCP-NEXT:    vmovdqa64 %ymm1, %ymm18
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm2 = [3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15]
; AVX512-FCP-NEXT:    vpshufb %ymm2, %ymm11, %ymm1
; AVX512-FCP-NEXT:    vmovdqa %ymm2, %ymm11
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5,6],ymm0[7]
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm10 = [3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15]
; AVX512-FCP-NEXT:    vmovdqa64 %ymm19, %ymm1
; AVX512-FCP-NEXT:    vpshufb %ymm10, %ymm1, %ymm1
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm4 = [3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15]
; AVX512-FCP-NEXT:    vmovdqa64 %ymm21, %ymm2
; AVX512-FCP-NEXT:    vpshufb %ymm4, %ymm2, %ymm2
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3,4],ymm1[5],ymm2[6,7]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm1 = [0,0,3,11,0,0,3,11,0,0,3,11,0,0,3,11]
; AVX512-FCP-NEXT:    vpshufb %xmm1, %xmm13, %xmm2
; AVX512-FCP-NEXT:    vpshufb %xmm1, %xmm14, %xmm3
; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm3 = [3,11,0,0,3,11,0,0,3,11,0,0,3,11,0,0]
; AVX512-FCP-NEXT:    vmovdqa64 %xmm17, %xmm15
; AVX512-FCP-NEXT:    vpshufb %xmm3, %xmm15, %xmm5
; AVX512-FCP-NEXT:    vpshufb %xmm3, %xmm12, %xmm12
; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm12[0],xmm5[0],xmm12[1],xmm5[1],xmm12[2],xmm5[2],xmm12[3],xmm5[3]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} xmm2 = xmm5[0,1,2],xmm2[3]
; AVX512-FCP-NEXT:    vpsrlq $24, %zmm28, %zmm5
; AVX512-FCP-NEXT:    vmovdqa64 %zmm28, %zmm24
; AVX512-FCP-NEXT:    vpmovqb %zmm5, %xmm5
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} xmm2 = xmm5[0,1],xmm2[2,3]
; AVX512-FCP-NEXT:    vinserti32x4 $2, %xmm2, %zmm0, %zmm2
; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm0, %zmm0, %zmm2 {%k1}
; AVX512-FCP-NEXT:    vmovdqa64 %ymm18, %ymm0
; AVX512-FCP-NEXT:    vpshufb %ymm0, %ymm9, %ymm0
; AVX512-FCP-NEXT:    vmovdqa64 %ymm27, %ymm5
; AVX512-FCP-NEXT:    vpshufb %ymm11, %ymm5, %ymm5
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm5[0,1,2,3,4,5,6],ymm0[7]
; AVX512-FCP-NEXT:    vmovdqa64 %ymm25, %ymm5
; AVX512-FCP-NEXT:    vpshufb %ymm10, %ymm5, %ymm5
; AVX512-FCP-NEXT:    vmovdqa64 %ymm26, %ymm9
; AVX512-FCP-NEXT:    vpshufb %ymm4, %ymm9, %ymm4
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3,4],ymm5[5],ymm4[6,7]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm4[0,1,2,3,4,5],ymm0[6,7]
; AVX512-FCP-NEXT:    vpshufb %xmm1, %xmm6, %xmm4
; AVX512-FCP-NEXT:    vmovdqa64 %xmm6, %xmm21
; AVX512-FCP-NEXT:    vmovdqa64 %xmm16, %xmm5
; AVX512-FCP-NEXT:    vpshufb %xmm1, %xmm5, %xmm1
; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm4[0],xmm1[1],xmm4[1],xmm1[2],xmm4[2],xmm1[3],xmm4[3]
; AVX512-FCP-NEXT:    vpshufb %xmm3, %xmm7, %xmm4
; AVX512-FCP-NEXT:    vmovdqa64 %xmm31, %xmm16
; AVX512-FCP-NEXT:    vpshufb %xmm3, %xmm8, %xmm3
; AVX512-FCP-NEXT:    vmovdqa64 %xmm22, %xmm6
; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} xmm1 = xmm3[0,1,2],xmm1[3]
; AVX512-FCP-NEXT:    vpsrlq $24, %zmm29, %zmm3
; AVX512-FCP-NEXT:    vpmovqb %zmm3, %xmm3
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} xmm1 = xmm3[0,1],xmm1[2,3]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm2, %zmm0
; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm0 = [1,3,2,3,1,3,5,7]
; AVX512-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm2 # 32-byte Folded Reload
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm17 = [0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12]
; AVX512-FCP-NEXT:    vmovdqa64 %ymm17, %ymm1
; AVX512-FCP-NEXT:    vpshufb %ymm1, %ymm2, %ymm1
; AVX512-FCP-NEXT:    vmovdqa64 %ymm2, %ymm28
; AVX512-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm3 # 32-byte Folded Reload
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm18 = [0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12]
; AVX512-FCP-NEXT:    vmovdqa64 %ymm18, %ymm2
; AVX512-FCP-NEXT:    vpshufb %ymm2, %ymm3, %ymm2
; AVX512-FCP-NEXT:    vmovdqa %ymm3, %ymm4
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5,6],ymm1[7]
; AVX512-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm2 # 32-byte Folded Reload
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm7 = [0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12]
; AVX512-FCP-NEXT:    vpshufb %ymm7, %ymm2, %ymm9
; AVX512-FCP-NEXT:    vmovdqa64 %ymm2, %ymm19
; AVX512-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm2 # 32-byte Folded Reload
; AVX512-FCP-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm8 = [0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12]
; AVX512-FCP-NEXT:    vpshufb %ymm8, %ymm2, %ymm10
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm9 = ymm10[0,1,2,3,4],ymm9[5],ymm10[6,7]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm9[0,1,2,3,4,5],ymm1[6,7]
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm11 = [0,0,4,12,0,0,4,12,0,0,4,12,0,0,4,12]
; AVX512-FCP-NEXT:    vpshufb %xmm11, %xmm13, %xmm9
; AVX512-FCP-NEXT:    vmovdqa64 %xmm30, %xmm20
; AVX512-FCP-NEXT:    vpshufb %xmm11, %xmm14, %xmm10
; AVX512-FCP-NEXT:    vmovdqa64 %xmm14, %xmm22
; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm10[0],xmm9[0],xmm10[1],xmm9[1],xmm10[2],xmm9[2],xmm10[3],xmm9[3]
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm12 = [4,12,0,0,4,12,0,0,4,12,0,0,4,12,0,0]
; AVX512-FCP-NEXT:    vpshufb %xmm12, %xmm15, %xmm10
; AVX512-FCP-NEXT:    vmovdqa64 %xmm15, %xmm23
; AVX512-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX512-FCP-NEXT:    vpshufb %xmm12, %xmm2, %xmm13
; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm13[0],xmm10[0],xmm13[1],xmm10[1],xmm13[2],xmm10[2],xmm13[3],xmm10[3]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} xmm9 = xmm10[0,1,2],xmm9[3]
; AVX512-FCP-NEXT:    vpsrlq $32, %zmm24, %zmm10
; AVX512-FCP-NEXT:    vmovdqa64 %zmm24, %zmm25
; AVX512-FCP-NEXT:    vpmovqb %zmm10, %xmm10
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} xmm9 = xmm10[0,1],xmm9[2,3]
; AVX512-FCP-NEXT:    vinserti32x4 $2, %xmm9, %zmm0, %zmm13
; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm13 {%k1}
; AVX512-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm2 # 32-byte Folded Reload
; AVX512-FCP-NEXT:    vmovdqa64 %ymm17, %ymm1
; AVX512-FCP-NEXT:    vpshufb %ymm1, %ymm2, %ymm1
; AVX512-FCP-NEXT:    vmovdqa64 %ymm2, %ymm27
; AVX512-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm3 # 32-byte Folded Reload
; AVX512-FCP-NEXT:    vmovdqa64 %ymm18, %ymm2
; AVX512-FCP-NEXT:    vpshufb %ymm2, %ymm3, %ymm15
; AVX512-FCP-NEXT:    vmovdqa64 %ymm3, %ymm26
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm15 = ymm15[0,1,2,3,4,5,6],ymm1[7]
; AVX512-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm2 # 32-byte Folded Reload
; AVX512-FCP-NEXT:    vpshufb %ymm7, %ymm2, %ymm1
; AVX512-FCP-NEXT:    vmovdqa64 %ymm2, %ymm31
; AVX512-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX512-FCP-NEXT:    vpshufb %ymm8, %ymm0, %ymm14
; AVX512-FCP-NEXT:    vmovdqa64 %ymm0, %ymm30
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm14[0,1,2,3,4],ymm1[5],ymm14[6,7]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5],ymm15[6,7]
; AVX512-FCP-NEXT:    vmovdqa64 %xmm21, %xmm10
; AVX512-FCP-NEXT:    vpshufb %xmm11, %xmm10, %xmm14
; AVX512-FCP-NEXT:    vpshufb %xmm11, %xmm5, %xmm11
; AVX512-FCP-NEXT:    vmovdqa64 %xmm5, %xmm24
; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm11[0],xmm14[0],xmm11[1],xmm14[1],xmm11[2],xmm14[2],xmm11[3],xmm14[3]
; AVX512-FCP-NEXT:    vmovdqa64 %xmm16, %xmm8
; AVX512-FCP-NEXT:    vpshufb %xmm12, %xmm8, %xmm14
; AVX512-FCP-NEXT:    vpshufb %xmm12, %xmm6, %xmm12
; AVX512-FCP-NEXT:    vmovdqa64 %xmm6, %xmm21
; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm12[0],xmm14[0],xmm12[1],xmm14[1],xmm12[2],xmm14[2],xmm12[3],xmm14[3]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} xmm11 = xmm12[0,1,2],xmm11[3]
; AVX512-FCP-NEXT:    vpsrlq $32, %zmm29, %zmm12
; AVX512-FCP-NEXT:    vpmovqb %zmm12, %xmm12
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} xmm11 = xmm12[0,1],xmm11[2,3]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm11[0,1,2,3],ymm1[4,5,6,7]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm13, %zmm0
; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm3 = [1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13]
; AVX512-FCP-NEXT:    vmovdqa64 %ymm28, %ymm0
; AVX512-FCP-NEXT:    vpshufb %ymm3, %ymm0, %ymm1
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm5 = [1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13]
; AVX512-FCP-NEXT:    vpshufb %ymm5, %ymm4, %ymm11
; AVX512-FCP-NEXT:    vmovdqa64 %ymm4, %ymm16
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm11[0,1,2,3,4,5,6],ymm1[7]
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm6 = [1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13]
; AVX512-FCP-NEXT:    vmovdqa64 %ymm19, %ymm0
; AVX512-FCP-NEXT:    vpshufb %ymm6, %ymm0, %ymm11
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm18 = [1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13]
; AVX512-FCP-NEXT:    vmovdqa64 %ymm18, %ymm7
; AVX512-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX512-FCP-NEXT:    vpshufb %ymm7, %ymm2, %ymm12
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm11 = ymm12[0,1,2,3,4],ymm11[5],ymm12[6,7]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm11[0,1,2,3,4,5],ymm1[6,7]
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm11 = [0,0,5,13,0,0,5,13,0,0,5,13,0,0,5,13]
; AVX512-FCP-NEXT:    vmovdqa64 %xmm20, %xmm2
; AVX512-FCP-NEXT:    vpshufb %xmm11, %xmm2, %xmm12
; AVX512-FCP-NEXT:    vmovdqa64 %xmm20, %xmm19
; AVX512-FCP-NEXT:    vmovdqa64 %xmm22, %xmm4
; AVX512-FCP-NEXT:    vpshufb %xmm11, %xmm4, %xmm13
; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm13[0],xmm12[0],xmm13[1],xmm12[1],xmm13[2],xmm12[2],xmm13[3],xmm12[3]
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm13 = [5,13,0,0,5,13,0,0,5,13,0,0,5,13,0,0]
; AVX512-FCP-NEXT:    vmovdqa64 %xmm23, %xmm9
; AVX512-FCP-NEXT:    vpshufb %xmm13, %xmm9, %xmm14
; AVX512-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
; AVX512-FCP-NEXT:    vpshufb %xmm13, %xmm7, %xmm15
; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} xmm12 = xmm14[0,1,2],xmm12[3]
; AVX512-FCP-NEXT:    vpsrlq $40, %zmm25, %zmm14
; AVX512-FCP-NEXT:    vpmovqb %zmm14, %xmm14
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} xmm12 = xmm14[0,1],xmm12[2,3]
; AVX512-FCP-NEXT:    vinserti32x4 $2, %xmm12, %zmm0, %zmm12
; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm12 {%k1}
; AVX512-FCP-NEXT:    vmovdqa64 %ymm27, %ymm1
; AVX512-FCP-NEXT:    vpshufb %ymm3, %ymm1, %ymm1
; AVX512-FCP-NEXT:    vmovdqa64 %ymm26, %ymm2
; AVX512-FCP-NEXT:    vpshufb %ymm5, %ymm2, %ymm14
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm14[0,1,2,3,4,5,6],ymm1[7]
; AVX512-FCP-NEXT:    vmovdqa64 %ymm31, %ymm3
; AVX512-FCP-NEXT:    vpshufb %ymm6, %ymm3, %ymm14
; AVX512-FCP-NEXT:    vmovdqa64 %ymm18, %ymm3
; AVX512-FCP-NEXT:    vmovdqa64 %ymm30, %ymm2
; AVX512-FCP-NEXT:    vpshufb %ymm3, %ymm2, %ymm15
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm14 = ymm15[0,1,2,3,4],ymm14[5],ymm15[6,7]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm14[0,1,2,3,4,5],ymm1[6,7]
; AVX512-FCP-NEXT:    vpshufb %xmm11, %xmm10, %xmm14
; AVX512-FCP-NEXT:    vmovdqa64 %xmm10, %xmm20
; AVX512-FCP-NEXT:    vmovdqa64 %xmm24, %xmm2
; AVX512-FCP-NEXT:    vpshufb %xmm11, %xmm2, %xmm11
; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm11[0],xmm14[0],xmm11[1],xmm14[1],xmm11[2],xmm14[2],xmm11[3],xmm14[3]
; AVX512-FCP-NEXT:    vpshufb %xmm13, %xmm8, %xmm14
; AVX512-FCP-NEXT:    vmovdqa64 %xmm8, %xmm22
; AVX512-FCP-NEXT:    vmovdqa64 %xmm21, %xmm2
; AVX512-FCP-NEXT:    vpshufb %xmm13, %xmm2, %xmm13
; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm13[0],xmm14[0],xmm13[1],xmm14[1],xmm13[2],xmm14[2],xmm13[3],xmm14[3]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} xmm11 = xmm13[0,1,2],xmm11[3]
; AVX512-FCP-NEXT:    vpsrlq $40, %zmm29, %zmm13
; AVX512-FCP-NEXT:    vmovdqa64 %zmm29, %zmm18
; AVX512-FCP-NEXT:    vpmovqb %zmm13, %xmm13
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} xmm11 = xmm13[0,1],xmm11[2,3]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm11[0,1,2,3],ymm1[4,5,6,7]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm12, %zmm29
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm3 = [2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14]
; AVX512-FCP-NEXT:    vmovdqa64 %ymm28, %ymm10
; AVX512-FCP-NEXT:    vpshufb %ymm3, %ymm10, %ymm1
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm21 = [2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14]
; AVX512-FCP-NEXT:    vmovdqa64 %ymm16, %ymm5
; AVX512-FCP-NEXT:    vmovdqa64 %ymm21, %ymm6
; AVX512-FCP-NEXT:    vpshufb %ymm6, %ymm5, %ymm11
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm11[0,1,2,3,4,5,6],ymm1[7]
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm6 = [2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14]
; AVX512-FCP-NEXT:    vpshufb %ymm6, %ymm0, %ymm11
; AVX512-FCP-NEXT:    vmovdqa64 %ymm0, %ymm17
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm8 = [2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14]
; AVX512-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX512-FCP-NEXT:    vpshufb %ymm8, %ymm0, %ymm12
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm11 = ymm12[0,1,2,3,4],ymm11[5],ymm12[6,7]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm11[0,1,2,3,4,5],ymm1[6,7]
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm11 = [0,0,6,14,0,0,6,14,0,0,6,14,0,0,6,14]
; AVX512-FCP-NEXT:    vmovdqa64 %xmm19, %xmm0
; AVX512-FCP-NEXT:    vpshufb %xmm11, %xmm0, %xmm12
; AVX512-FCP-NEXT:    vpshufb %xmm11, %xmm4, %xmm13
; AVX512-FCP-NEXT:    vmovdqa64 %xmm4, %xmm23
; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm13[0],xmm12[0],xmm13[1],xmm12[1],xmm13[2],xmm12[2],xmm13[3],xmm12[3]
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm13 = [6,14,0,0,6,14,0,0,6,14,0,0,6,14,0,0]
; AVX512-FCP-NEXT:    vpshufb %xmm13, %xmm9, %xmm14
; AVX512-FCP-NEXT:    vmovdqa64 %xmm9, %xmm28
; AVX512-FCP-NEXT:    vpshufb %xmm13, %xmm7, %xmm15
; AVX512-FCP-NEXT:    vmovdqa64 %xmm7, %xmm16
; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} xmm12 = xmm14[0,1,2],xmm12[3]
; AVX512-FCP-NEXT:    vpsrlq $48, %zmm25, %zmm14
; AVX512-FCP-NEXT:    vpmovqb %zmm14, %xmm14
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} xmm12 = xmm14[0,1],xmm12[2,3]
; AVX512-FCP-NEXT:    vinserti32x4 $2, %xmm12, %zmm0, %zmm12
; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm12 {%k1}
; AVX512-FCP-NEXT:    vmovdqa64 %ymm27, %ymm0
; AVX512-FCP-NEXT:    vpshufb %ymm3, %ymm0, %ymm1
; AVX512-FCP-NEXT:    vmovdqa64 %ymm26, %ymm9
; AVX512-FCP-NEXT:    vmovdqa64 %ymm21, %ymm3
; AVX512-FCP-NEXT:    vpshufb %ymm3, %ymm9, %ymm14
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm14[0,1,2,3,4,5,6],ymm1[7]
; AVX512-FCP-NEXT:    vmovdqa64 %ymm31, %ymm3
; AVX512-FCP-NEXT:    vpshufb %ymm6, %ymm3, %ymm14
; AVX512-FCP-NEXT:    vmovdqa64 %ymm30, %ymm3
; AVX512-FCP-NEXT:    vpshufb %ymm8, %ymm3, %ymm15
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm14 = ymm15[0,1,2,3,4],ymm14[5],ymm15[6,7]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm14[0,1,2,3,4,5],ymm1[6,7]
; AVX512-FCP-NEXT:    vmovdqa64 %xmm20, %xmm8
; AVX512-FCP-NEXT:    vpshufb %xmm11, %xmm8, %xmm14
; AVX512-FCP-NEXT:    vmovdqa64 %xmm24, %xmm15
; AVX512-FCP-NEXT:    vpshufb %xmm11, %xmm15, %xmm11
; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm11[0],xmm14[0],xmm11[1],xmm14[1],xmm11[2],xmm14[2],xmm11[3],xmm14[3]
; AVX512-FCP-NEXT:    vmovdqa64 %xmm22, %xmm3
; AVX512-FCP-NEXT:    vpshufb %xmm13, %xmm3, %xmm14
; AVX512-FCP-NEXT:    vpshufb %xmm13, %xmm2, %xmm13
; AVX512-FCP-NEXT:    vmovdqa64 %xmm2, %xmm20
; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm13[0],xmm14[0],xmm13[1],xmm14[1],xmm13[2],xmm14[2],xmm13[3],xmm14[3]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} xmm11 = xmm13[0,1,2],xmm11[3]
; AVX512-FCP-NEXT:    vpsrlq $48, %zmm18, %zmm13
; AVX512-FCP-NEXT:    vpmovqb %zmm13, %xmm13
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} xmm11 = xmm13[0,1],xmm11[2,3]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm11[0,1,2,3],ymm1[4,5,6,7]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm12, %zmm21
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm13 = [3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15]
; AVX512-FCP-NEXT:    vpshufb %ymm13, %ymm10, %ymm6
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm14 = [3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15]
; AVX512-FCP-NEXT:    vpshufb %ymm14, %ymm5, %ymm5
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm5 = ymm5[0,1,2,3,4,5,6],ymm6[7]
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm1 = [3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15]
; AVX512-FCP-NEXT:    vmovdqa64 %ymm17, %ymm2
; AVX512-FCP-NEXT:    vpshufb %ymm1, %ymm2, %ymm3
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm7 = [3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15]
; AVX512-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX512-FCP-NEXT:    vpshufb %ymm7, %ymm2, %ymm4
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm3 = ymm4[0,1,2,3,4],ymm3[5],ymm4[6,7]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm5[6,7]
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm4 = [0,0,7,15,0,0,7,15,0,0,7,15,0,0,7,15]
; AVX512-FCP-NEXT:    vmovdqa64 %xmm19, %xmm2
; AVX512-FCP-NEXT:    vpshufb %xmm4, %xmm2, %xmm5
; AVX512-FCP-NEXT:    vmovdqa64 %xmm23, %xmm2
; AVX512-FCP-NEXT:    vpshufb %xmm4, %xmm2, %xmm6
; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX512-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm6 = [7,15,0,0,7,15,0,0,7,15,0,0,7,15,0,0]
; AVX512-FCP-NEXT:    vmovdqa64 %xmm28, %xmm2
; AVX512-FCP-NEXT:    vpshufb %xmm6, %xmm2, %xmm11
; AVX512-FCP-NEXT:    vmovdqa64 %xmm16, %xmm2
; AVX512-FCP-NEXT:    vpshufb %xmm6, %xmm2, %xmm12
; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm12[0],xmm11[0],xmm12[1],xmm11[1],xmm12[2],xmm11[2],xmm12[3],xmm11[3]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} xmm5 = xmm11[0,1,2],xmm5[3]
; AVX512-FCP-NEXT:    vpsrlq $56, %zmm25, %zmm11
; AVX512-FCP-NEXT:    vpmovqb %zmm11, %xmm11
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} xmm5 = xmm11[0,1],xmm5[2,3]
; AVX512-FCP-NEXT:    vinserti32x4 $2, %xmm5, %zmm0, %zmm5
; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm3, %zmm0, %zmm5 {%k1}
; AVX512-FCP-NEXT:    vpshufb %ymm13, %ymm0, %ymm3
; AVX512-FCP-NEXT:    vpshufb %ymm14, %ymm9, %ymm9
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm3 = ymm9[0,1,2,3,4,5,6],ymm3[7]
; AVX512-FCP-NEXT:    vmovdqa64 %ymm31, %ymm0
; AVX512-FCP-NEXT:    vpshufb %ymm1, %ymm0, %ymm2
; AVX512-FCP-NEXT:    vmovdqa64 %ymm30, %ymm0
; AVX512-FCP-NEXT:    vpshufb %ymm7, %ymm0, %ymm0
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm2[5],ymm0[6,7]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm3[6,7]
; AVX512-FCP-NEXT:    vpshufb %xmm4, %xmm8, %xmm2
; AVX512-FCP-NEXT:    vpshufb %xmm4, %xmm15, %xmm3
; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
; AVX512-FCP-NEXT:    vmovdqa64 %xmm22, %xmm1
; AVX512-FCP-NEXT:    vpshufb %xmm6, %xmm1, %xmm3
; AVX512-FCP-NEXT:    vmovdqa64 %xmm20, %xmm1
; AVX512-FCP-NEXT:    vpshufb %xmm6, %xmm1, %xmm4
; AVX512-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} xmm2 = xmm3[0,1,2],xmm2[3]
; AVX512-FCP-NEXT:    vpsrlq $56, %zmm18, %zmm3
; AVX512-FCP-NEXT:    vpmovqb %zmm3, %xmm3
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} xmm2 = xmm3[0,1],xmm2[2,3]
; AVX512-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]
; AVX512-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm5, %zmm0
; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovaps %zmm1, (%rsi)
; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovaps %zmm1, (%rdx)
; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovaps %zmm1, (%rcx)
; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovaps %zmm1, (%r8)
; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512-FCP-NEXT:    vmovaps %zmm1, (%r9)
; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512-FCP-NEXT:    vmovdqa64 %zmm29, (%rax)
; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512-FCP-NEXT:    vmovdqa64 %zmm21, (%rax)
; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, (%rax)
; AVX512-FCP-NEXT:    addq $440, %rsp # imm = 0x1B8
; AVX512-FCP-NEXT:    vzeroupper
; AVX512-FCP-NEXT:    retq
;
; AVX512DQ-LABEL: load_i8_stride8_vf64:
; AVX512DQ:       # %bb.0:
; AVX512DQ-NEXT:    subq $520, %rsp # imm = 0x208
; AVX512DQ-NEXT:    vmovdqa64 256(%rdi), %zmm17
; AVX512DQ-NEXT:    vmovdqa64 384(%rdi), %zmm0
; AVX512DQ-NEXT:    vpmovqb %zmm0, %xmm2
; AVX512DQ-NEXT:    vmovdqa 496(%rdi), %xmm3
; AVX512DQ-NEXT:    vpbroadcastd {{.*#+}} xmm0 = [0,0,0,8,0,0,0,8,0,0,0,8,0,0,0,8]
; AVX512DQ-NEXT:    vpshufb %xmm0, %xmm3, %xmm1
; AVX512DQ-NEXT:    vmovdqa %xmm3, %xmm7
; AVX512DQ-NEXT:    vmovdqa 480(%rdi), %xmm4
; AVX512DQ-NEXT:    vpshufb %xmm0, %xmm4, %xmm3
; AVX512DQ-NEXT:    vmovdqa %xmm4, %xmm8
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3]
; AVX512DQ-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm3
; AVX512DQ-NEXT:    vmovdqa 464(%rdi), %xmm5
; AVX512DQ-NEXT:    vpbroadcastd {{.*#+}} xmm1 = [0,8,0,0,0,8,0,0,0,8,0,0,0,8,0,0]
; AVX512DQ-NEXT:    vpshufb %xmm1, %xmm5, %xmm4
; AVX512DQ-NEXT:    vmovdqa %xmm5, %xmm11
; AVX512DQ-NEXT:    vmovdqa 448(%rdi), %xmm15
; AVX512DQ-NEXT:    vpshufb %xmm1, %xmm15, %xmm5
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX512DQ-NEXT:    vinserti128 $1, %xmm4, %ymm0, %ymm4
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm3 = ymm4[0,1,2,3,4,5,6],ymm3[7]
; AVX512DQ-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX512DQ-NEXT:    vmovdqa 384(%rdi), %ymm4
; AVX512DQ-NEXT:    vpmovqb %ymm4, %xmm4
; AVX512DQ-NEXT:    vinserti128 $1, %xmm4, %ymm0, %ymm4
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm2 = ymm4[0,1,2,3,4],ymm2[5],ymm4[6,7]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm3[6,7]
; AVX512DQ-NEXT:    vmovdqa 368(%rdi), %xmm4
; AVX512DQ-NEXT:    vpshufb %xmm0, %xmm4, %xmm3
; AVX512DQ-NEXT:    vmovdqa64 %xmm4, %xmm18
; AVX512DQ-NEXT:    vmovdqa 352(%rdi), %xmm5
; AVX512DQ-NEXT:    vpshufb %xmm0, %xmm5, %xmm4
; AVX512DQ-NEXT:    vmovdqa %xmm5, %xmm14
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX512DQ-NEXT:    vmovdqa 336(%rdi), %xmm4
; AVX512DQ-NEXT:    vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-NEXT:    vpshufb %xmm1, %xmm4, %xmm4
; AVX512DQ-NEXT:    vmovdqa 320(%rdi), %xmm6
; AVX512DQ-NEXT:    vpshufb %xmm1, %xmm6, %xmm5
; AVX512DQ-NEXT:    vmovdqa64 %xmm6, %xmm19
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} xmm3 = xmm4[0,1,2],xmm3[3]
; AVX512DQ-NEXT:    vpmovqb %zmm17, %xmm4
; AVX512DQ-NEXT:    vpblendd {{.*#+}} xmm3 = xmm4[0,1],xmm3[2,3]
; AVX512DQ-NEXT:    vinserti32x4 $2, %xmm3, %zmm0, %zmm3
; AVX512DQ-NEXT:    movb $-64, %al
; AVX512DQ-NEXT:    kmovw %eax, %k1
; AVX512DQ-NEXT:    vinserti64x4 $1, %ymm2, %zmm0, %zmm3 {%k1}
; AVX512DQ-NEXT:    vmovdqa 240(%rdi), %xmm4
; AVX512DQ-NEXT:    vpshufb %xmm0, %xmm4, %xmm2
; AVX512DQ-NEXT:    vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-NEXT:    vmovdqa64 %xmm4, %xmm25
; AVX512DQ-NEXT:    vmovdqa 224(%rdi), %xmm4
; AVX512DQ-NEXT:    vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-NEXT:    vpshufb %xmm0, %xmm4, %xmm4
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm4[0],xmm2[0],xmm4[1],xmm2[1],xmm4[2],xmm2[2],xmm4[3],xmm2[3]
; AVX512DQ-NEXT:    vmovdqa 208(%rdi), %xmm5
; AVX512DQ-NEXT:    vpshufb %xmm1, %xmm5, %xmm4
; AVX512DQ-NEXT:    vmovdqa64 %xmm5, %xmm29
; AVX512DQ-NEXT:    vmovdqa 192(%rdi), %xmm10
; AVX512DQ-NEXT:    vpshufb %xmm1, %xmm10, %xmm5
; AVX512DQ-NEXT:    vmovdqa %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX512DQ-NEXT:    vmovdqa64 128(%rdi), %zmm5
; AVX512DQ-NEXT:    vpmovqb %zmm5, %xmm5
; AVX512DQ-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX512DQ-NEXT:    vinserti128 $1, %xmm4, %ymm0, %ymm4
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm2 = ymm4[0,1,2,3,4,5,6],ymm2[7]
; AVX512DQ-NEXT:    vinserti128 $1, %xmm5, %ymm0, %ymm4
; AVX512DQ-NEXT:    vmovdqa 128(%rdi), %ymm5
; AVX512DQ-NEXT:    vpmovqb %ymm5, %xmm5
; AVX512DQ-NEXT:    vinserti128 $1, %xmm5, %ymm0, %ymm5
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm4 = ymm5[0,1,2,3,4],ymm4[5],ymm5[6,7]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm2 = ymm4[0,1,2,3,4,5],ymm2[6,7]
; AVX512DQ-NEXT:    vmovdqa 112(%rdi), %xmm4
; AVX512DQ-NEXT:    vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-NEXT:    vpshufb %xmm0, %xmm4, %xmm4
; AVX512DQ-NEXT:    vmovdqa 96(%rdi), %xmm5
; AVX512DQ-NEXT:    vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-NEXT:    vpshufb %xmm0, %xmm5, %xmm0
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm4[0],xmm0[1],xmm4[1],xmm0[2],xmm4[2],xmm0[3],xmm4[3]
; AVX512DQ-NEXT:    vmovdqa 80(%rdi), %xmm5
; AVX512DQ-NEXT:    vpshufb %xmm1, %xmm5, %xmm4
; AVX512DQ-NEXT:    vmovdqa64 %xmm5, %xmm30
; AVX512DQ-NEXT:    vmovdqa 64(%rdi), %xmm5
; AVX512DQ-NEXT:    vpshufb %xmm1, %xmm5, %xmm1
; AVX512DQ-NEXT:    vmovdqa64 %xmm5, %xmm28
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm4[0],xmm1[1],xmm4[1],xmm1[2],xmm4[2],xmm1[3],xmm4[3]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[3]
; AVX512DQ-NEXT:    vmovdqa64 (%rdi), %zmm16
; AVX512DQ-NEXT:    vpmovqb %zmm16, %xmm1
; AVX512DQ-NEXT:    vpblendd {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm0, %zmm3, %zmm0
; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpbroadcastd {{.*#+}} xmm9 = [0,0,1,9,0,0,1,9,0,0,1,9,0,0,1,9]
; AVX512DQ-NEXT:    vpshufb %xmm9, %xmm7, %xmm0
; AVX512DQ-NEXT:    vmovdqa64 %xmm7, %xmm24
; AVX512DQ-NEXT:    vpshufb %xmm9, %xmm8, %xmm2
; AVX512DQ-NEXT:    vmovdqa64 %xmm8, %xmm21
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
; AVX512DQ-NEXT:    vpbroadcastd {{.*#+}} xmm8 = [1,9,0,0,1,9,0,0,1,9,0,0,1,9,0,0]
; AVX512DQ-NEXT:    vpshufb %xmm8, %xmm11, %xmm3
; AVX512DQ-NEXT:    vmovdqa64 %xmm11, %xmm27
; AVX512DQ-NEXT:    vpshufb %xmm8, %xmm15, %xmm4
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX512DQ-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX512DQ-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm4 = ymm3[0,1,2,3,4,5,6],ymm2[7]
; AVX512DQ-NEXT:    vmovdqa 416(%rdi), %xmm0
; AVX512DQ-NEXT:    vmovdqa 432(%rdi), %xmm11
; AVX512DQ-NEXT:    vmovd {{.*#+}} xmm3 = [0,0,1,9,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX512DQ-NEXT:    vpshufb %xmm3, %xmm11, %xmm2
; AVX512DQ-NEXT:    vpshufb %xmm3, %xmm0, %xmm5
; AVX512DQ-NEXT:    vmovdqa64 %xmm0, %xmm22
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm2[0],xmm5[1],xmm2[1],xmm5[2],xmm2[2],xmm5[3],xmm2[3]
; AVX512DQ-NEXT:    vmovdqa 384(%rdi), %xmm0
; AVX512DQ-NEXT:    vmovdqa 400(%rdi), %xmm12
; AVX512DQ-NEXT:    vmovd {{.*#+}} xmm2 = [1,9,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX512DQ-NEXT:    vpshufb %xmm2, %xmm12, %xmm6
; AVX512DQ-NEXT:    vpshufb %xmm2, %xmm0, %xmm7
; AVX512DQ-NEXT:    vmovdqa64 %xmm0, %xmm31
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
; AVX512DQ-NEXT:    vinserti128 $1, %xmm5, %ymm0, %ymm5
; AVX512DQ-NEXT:    vinserti128 $1, %xmm6, %ymm0, %ymm6
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm5 = ymm6[0,1,2,3,4],ymm5[5],ymm6[6,7]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm4 = ymm5[0,1,2,3,4,5],ymm4[6,7]
; AVX512DQ-NEXT:    vmovdqa64 %xmm18, %xmm0
; AVX512DQ-NEXT:    vpshufb %xmm9, %xmm0, %xmm5
; AVX512DQ-NEXT:    vmovdqa64 %xmm18, %xmm23
; AVX512DQ-NEXT:    vpshufb %xmm9, %xmm14, %xmm6
; AVX512DQ-NEXT:    vmovdqa64 %xmm14, %xmm26
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX512DQ-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX512DQ-NEXT:    vpshufb %xmm8, %xmm0, %xmm6
; AVX512DQ-NEXT:    vmovdqa64 %xmm19, %xmm14
; AVX512DQ-NEXT:    vpshufb %xmm8, %xmm14, %xmm7
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} xmm5 = xmm6[0,1,2],xmm5[3]
; AVX512DQ-NEXT:    vpsrlq $8, %zmm17, %zmm6
; AVX512DQ-NEXT:    vpmovqb %zmm6, %xmm6
; AVX512DQ-NEXT:    vpblendd {{.*#+}} xmm5 = xmm6[0,1],xmm5[2,3]
; AVX512DQ-NEXT:    vinserti32x4 $2, %xmm5, %zmm0, %zmm19
; AVX512DQ-NEXT:    vinserti64x4 $1, %ymm4, %zmm0, %zmm19 {%k1}
; AVX512DQ-NEXT:    vmovdqa64 %xmm25, %xmm0
; AVX512DQ-NEXT:    vpshufb %xmm9, %xmm0, %xmm4
; AVX512DQ-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX512DQ-NEXT:    vpshufb %xmm9, %xmm0, %xmm5
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX512DQ-NEXT:    vmovdqa64 %xmm29, %xmm0
; AVX512DQ-NEXT:    vpshufb %xmm8, %xmm0, %xmm5
; AVX512DQ-NEXT:    vpshufb %xmm8, %xmm10, %xmm6
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX512DQ-NEXT:    vinserti128 $1, %xmm4, %ymm0, %ymm4
; AVX512DQ-NEXT:    vinserti128 $1, %xmm5, %ymm0, %ymm5
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm1 = ymm5[0,1,2,3,4,5,6],ymm4[7]
; AVX512DQ-NEXT:    vmovdqa 160(%rdi), %xmm0
; AVX512DQ-NEXT:    vmovdqa 176(%rdi), %xmm5
; AVX512DQ-NEXT:    vpshufb %xmm3, %xmm5, %xmm4
; AVX512DQ-NEXT:    vmovdqa64 %xmm5, %xmm25
; AVX512DQ-NEXT:    vpshufb %xmm3, %xmm0, %xmm3
; AVX512DQ-NEXT:    vmovdqa64 %xmm0, %xmm20
; AVX512DQ-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
; AVX512DQ-NEXT:    vmovdqa 128(%rdi), %xmm4
; AVX512DQ-NEXT:    vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-NEXT:    vmovdqa 144(%rdi), %xmm0
; AVX512DQ-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-NEXT:    vpshufb %xmm2, %xmm0, %xmm0
; AVX512DQ-NEXT:    vpshufb %xmm2, %xmm4, %xmm2
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3]
; AVX512DQ-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm2
; AVX512DQ-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm2[5],ymm0[6,7]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
; AVX512DQ-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX512DQ-NEXT:    vpshufb %xmm9, %xmm1, %xmm1
; AVX512DQ-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX512DQ-NEXT:    vpshufb %xmm9, %xmm2, %xmm2
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX512DQ-NEXT:    vmovdqa64 %xmm30, %xmm7
; AVX512DQ-NEXT:    vpshufb %xmm8, %xmm7, %xmm2
; AVX512DQ-NEXT:    vmovdqa64 %xmm28, %xmm3
; AVX512DQ-NEXT:    vpshufb %xmm8, %xmm3, %xmm3
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} xmm1 = xmm2[0,1,2],xmm1[3]
; AVX512DQ-NEXT:    vpsrlq $8, %zmm16, %zmm2
; AVX512DQ-NEXT:    vpmovqb %zmm2, %xmm2
; AVX512DQ-NEXT:    vpblendd {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm0, %zmm19, %zmm0
; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpbroadcastd {{.*#+}} xmm4 = [0,0,2,10,0,0,2,10,0,0,2,10,0,0,2,10]
; AVX512DQ-NEXT:    vmovdqa64 %xmm24, %xmm13
; AVX512DQ-NEXT:    vmovdqa64 %xmm24, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-NEXT:    vpshufb %xmm4, %xmm13, %xmm1
; AVX512DQ-NEXT:    vmovdqa64 %xmm21, %xmm10
; AVX512DQ-NEXT:    vmovdqa64 %xmm21, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-NEXT:    vpshufb %xmm4, %xmm10, %xmm2
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX512DQ-NEXT:    vpbroadcastd {{.*#+}} xmm1 = [2,10,0,0,2,10,0,0,2,10,0,0,2,10,0,0]
; AVX512DQ-NEXT:    vmovdqa64 %xmm27, %xmm0
; AVX512DQ-NEXT:    vpshufb %xmm1, %xmm0, %xmm3
; AVX512DQ-NEXT:    vpshufb %xmm1, %xmm15, %xmm8
; AVX512DQ-NEXT:    vmovdqa64 %xmm15, %xmm18
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm8[0],xmm3[0],xmm8[1],xmm3[1],xmm8[2],xmm3[2],xmm8[3],xmm3[3]
; AVX512DQ-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX512DQ-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm8 = ymm3[0,1,2,3,4,5,6],ymm2[7]
; AVX512DQ-NEXT:    vmovd {{.*#+}} xmm3 = [0,0,2,10,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX512DQ-NEXT:    vpshufb %xmm3, %xmm11, %xmm2
; AVX512DQ-NEXT:    vmovdqa64 %xmm11, %xmm24
; AVX512DQ-NEXT:    vmovdqa64 %xmm22, %xmm11
; AVX512DQ-NEXT:    vpshufb %xmm3, %xmm11, %xmm9
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm9[0],xmm2[0],xmm9[1],xmm2[1],xmm9[2],xmm2[2],xmm9[3],xmm2[3]
; AVX512DQ-NEXT:    vmovd {{.*#+}} xmm2 = [2,10,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX512DQ-NEXT:    vpshufb %xmm2, %xmm12, %xmm0
; AVX512DQ-NEXT:    vmovdqa64 %xmm12, %xmm30
; AVX512DQ-NEXT:    vmovdqa64 %xmm31, %xmm5
; AVX512DQ-NEXT:    vpshufb %xmm2, %xmm5, %xmm15
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm15[0],xmm0[0],xmm15[1],xmm0[1],xmm15[2],xmm0[2],xmm15[3],xmm0[3]
; AVX512DQ-NEXT:    vinserti128 $1, %xmm9, %ymm0, %ymm9
; AVX512DQ-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm9[5],ymm0[6,7]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm8[6,7]
; AVX512DQ-NEXT:    vmovdqa64 %xmm23, %xmm12
; AVX512DQ-NEXT:    vmovdqa64 %xmm23, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-NEXT:    vpshufb %xmm4, %xmm12, %xmm8
; AVX512DQ-NEXT:    vmovdqa64 %xmm26, %xmm5
; AVX512DQ-NEXT:    vpshufb %xmm4, %xmm5, %xmm9
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm9[0],xmm8[0],xmm9[1],xmm8[1],xmm9[2],xmm8[2],xmm9[3],xmm8[3]
; AVX512DQ-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX512DQ-NEXT:    vpshufb %xmm1, %xmm5, %xmm9
; AVX512DQ-NEXT:    vpshufb %xmm1, %xmm14, %xmm15
; AVX512DQ-NEXT:    vmovdqa64 %xmm14, %xmm21
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm15[0],xmm9[0],xmm15[1],xmm9[1],xmm15[2],xmm9[2],xmm15[3],xmm9[3]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} xmm8 = xmm9[0,1,2],xmm8[3]
; AVX512DQ-NEXT:    vpsrlq $16, %zmm17, %zmm9
; AVX512DQ-NEXT:    vpmovqb %zmm9, %xmm9
; AVX512DQ-NEXT:    vpblendd {{.*#+}} xmm8 = xmm9[0,1],xmm8[2,3]
; AVX512DQ-NEXT:    vinserti32x4 $2, %xmm8, %zmm0, %zmm8
; AVX512DQ-NEXT:    vinserti64x4 $1, %ymm0, %zmm0, %zmm8 {%k1}
; AVX512DQ-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX512DQ-NEXT:    vpshufb %xmm4, %xmm0, %xmm0
; AVX512DQ-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
; AVX512DQ-NEXT:    vpshufb %xmm4, %xmm14, %xmm9
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm9[0],xmm0[0],xmm9[1],xmm0[1],xmm9[2],xmm0[2],xmm9[3],xmm0[3]
; AVX512DQ-NEXT:    vmovdqa64 %xmm29, %xmm6
; AVX512DQ-NEXT:    vpshufb %xmm1, %xmm6, %xmm9
; AVX512DQ-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX512DQ-NEXT:    vpshufb %xmm1, %xmm15, %xmm15
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm15[0],xmm9[0],xmm15[1],xmm9[1],xmm15[2],xmm9[2],xmm15[3],xmm9[3]
; AVX512DQ-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX512DQ-NEXT:    vinserti128 $1, %xmm9, %ymm0, %ymm9
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm0 = ymm9[0,1,2,3,4,5,6],ymm0[7]
; AVX512DQ-NEXT:    vmovdqa64 %xmm25, %xmm9
; AVX512DQ-NEXT:    vpshufb %xmm3, %xmm9, %xmm9
; AVX512DQ-NEXT:    vmovdqa64 %xmm20, %xmm15
; AVX512DQ-NEXT:    vpshufb %xmm3, %xmm15, %xmm3
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm9[0],xmm3[1],xmm9[1],xmm3[2],xmm9[2],xmm3[3],xmm9[3]
; AVX512DQ-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
; AVX512DQ-NEXT:    vpshufb %xmm2, %xmm9, %xmm9
; AVX512DQ-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX512DQ-NEXT:    vpshufb %xmm2, %xmm15, %xmm2
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm9[0],xmm2[1],xmm9[1],xmm2[2],xmm9[2],xmm2[3],xmm9[3]
; AVX512DQ-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX512DQ-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3,4],ymm3[5],ymm2[6,7]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3,4,5],ymm0[6,7]
; AVX512DQ-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX512DQ-NEXT:    vpshufb %xmm4, %xmm2, %xmm2
; AVX512DQ-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX512DQ-NEXT:    vpshufb %xmm4, %xmm3, %xmm3
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
; AVX512DQ-NEXT:    vpshufb %xmm1, %xmm7, %xmm3
; AVX512DQ-NEXT:    vmovdqa64 %xmm7, %xmm29
; AVX512DQ-NEXT:    vmovdqa64 %xmm28, %xmm7
; AVX512DQ-NEXT:    vmovdqa64 %xmm28, (%rsp) # 16-byte Spill
; AVX512DQ-NEXT:    vpshufb %xmm1, %xmm7, %xmm1
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1],xmm1[2],xmm3[2],xmm1[3],xmm3[3]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} xmm1 = xmm1[0,1,2],xmm2[3]
; AVX512DQ-NEXT:    vpsrlq $16, %zmm16, %zmm2
; AVX512DQ-NEXT:    vpmovqb %zmm2, %xmm2
; AVX512DQ-NEXT:    vpblendd {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm0, %zmm8, %zmm0
; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpbroadcastd {{.*#+}} xmm0 = [0,0,3,11,0,0,3,11,0,0,3,11,0,0,3,11]
; AVX512DQ-NEXT:    vpshufb %xmm0, %xmm13, %xmm1
; AVX512DQ-NEXT:    vpshufb %xmm0, %xmm10, %xmm2
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX512DQ-NEXT:    vpbroadcastd {{.*#+}} xmm1 = [3,11,0,0,3,11,0,0,3,11,0,0,3,11,0,0]
; AVX512DQ-NEXT:    vmovdqa64 %xmm27, %xmm10
; AVX512DQ-NEXT:    vmovdqa64 %xmm27, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-NEXT:    vpshufb %xmm1, %xmm10, %xmm3
; AVX512DQ-NEXT:    vmovdqa64 %xmm18, %xmm4
; AVX512DQ-NEXT:    vpshufb %xmm1, %xmm4, %xmm4
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX512DQ-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX512DQ-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm4 = ymm3[0,1,2,3,4,5,6],ymm2[7]
; AVX512DQ-NEXT:    vmovd {{.*#+}} xmm2 = [0,0,3,11,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX512DQ-NEXT:    vmovdqa64 %xmm24, %xmm3
; AVX512DQ-NEXT:    vpshufb %xmm2, %xmm3, %xmm3
; AVX512DQ-NEXT:    vpshufb %xmm2, %xmm11, %xmm8
; AVX512DQ-NEXT:    vmovdqa64 %xmm22, %xmm28
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm3[0],xmm8[1],xmm3[1],xmm8[2],xmm3[2],xmm8[3],xmm3[3]
; AVX512DQ-NEXT:    vmovd {{.*#+}} xmm3 = [3,11,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX512DQ-NEXT:    vmovdqa64 %xmm30, %xmm9
; AVX512DQ-NEXT:    vpshufb %xmm3, %xmm9, %xmm9
; AVX512DQ-NEXT:    vmovdqa64 %xmm31, %xmm11
; AVX512DQ-NEXT:    vpshufb %xmm3, %xmm11, %xmm15
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm15[0],xmm9[0],xmm15[1],xmm9[1],xmm15[2],xmm9[2],xmm15[3],xmm9[3]
; AVX512DQ-NEXT:    vinserti128 $1, %xmm8, %ymm0, %ymm8
; AVX512DQ-NEXT:    vinserti128 $1, %xmm9, %ymm0, %ymm9
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm8 = ymm9[0,1,2,3,4],ymm8[5],ymm9[6,7]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm4 = ymm8[0,1,2,3,4,5],ymm4[6,7]
; AVX512DQ-NEXT:    vpshufb %xmm0, %xmm12, %xmm8
; AVX512DQ-NEXT:    vmovdqa64 %xmm26, %xmm13
; AVX512DQ-NEXT:    vmovdqa64 %xmm26, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-NEXT:    vpshufb %xmm0, %xmm13, %xmm9
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm9[0],xmm8[0],xmm9[1],xmm8[1],xmm9[2],xmm8[2],xmm9[3],xmm8[3]
; AVX512DQ-NEXT:    vpshufb %xmm1, %xmm5, %xmm9
; AVX512DQ-NEXT:    vmovdqa64 %xmm21, %xmm5
; AVX512DQ-NEXT:    vpshufb %xmm1, %xmm5, %xmm15
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm15[0],xmm9[0],xmm15[1],xmm9[1],xmm15[2],xmm9[2],xmm15[3],xmm9[3]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} xmm8 = xmm9[0,1,2],xmm8[3]
; AVX512DQ-NEXT:    vpsrlq $24, %zmm17, %zmm9
; AVX512DQ-NEXT:    vpmovqb %zmm9, %xmm9
; AVX512DQ-NEXT:    vpblendd {{.*#+}} xmm8 = xmm9[0,1],xmm8[2,3]
; AVX512DQ-NEXT:    vinserti32x4 $2, %xmm8, %zmm0, %zmm8
; AVX512DQ-NEXT:    vinserti64x4 $1, %ymm4, %zmm0, %zmm8 {%k1}
; AVX512DQ-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX512DQ-NEXT:    vpshufb %xmm0, %xmm4, %xmm4
; AVX512DQ-NEXT:    vpshufb %xmm0, %xmm14, %xmm9
; AVX512DQ-NEXT:    vmovdqa64 %xmm14, %xmm22
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm9[0],xmm4[0],xmm9[1],xmm4[1],xmm9[2],xmm4[2],xmm9[3],xmm4[3]
; AVX512DQ-NEXT:    vpshufb %xmm1, %xmm6, %xmm9
; AVX512DQ-NEXT:    vmovdqa64 %xmm6, %xmm26
; AVX512DQ-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX512DQ-NEXT:    vpshufb %xmm1, %xmm6, %xmm15
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm15[0],xmm9[0],xmm15[1],xmm9[1],xmm15[2],xmm9[2],xmm15[3],xmm9[3]
; AVX512DQ-NEXT:    vinserti128 $1, %xmm4, %ymm0, %ymm4
; AVX512DQ-NEXT:    vinserti128 $1, %xmm9, %ymm0, %ymm9
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm4 = ymm9[0,1,2,3,4,5,6],ymm4[7]
; AVX512DQ-NEXT:    vmovdqa64 %xmm25, %xmm14
; AVX512DQ-NEXT:    vpshufb %xmm2, %xmm14, %xmm9
; AVX512DQ-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX512DQ-NEXT:    vpshufb %xmm2, %xmm6, %xmm2
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm9[0],xmm2[1],xmm9[1],xmm2[2],xmm9[2],xmm2[3],xmm9[3]
; AVX512DQ-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX512DQ-NEXT:    vpshufb %xmm3, %xmm6, %xmm9
; AVX512DQ-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX512DQ-NEXT:    vpshufb %xmm3, %xmm6, %xmm3
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm9[0],xmm3[1],xmm9[1],xmm3[2],xmm9[2],xmm3[3],xmm9[3]
; AVX512DQ-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX512DQ-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4],ymm2[5],ymm3[6,7]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm4[6,7]
; AVX512DQ-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX512DQ-NEXT:    vpshufb %xmm0, %xmm3, %xmm3
; AVX512DQ-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX512DQ-NEXT:    vpshufb %xmm0, %xmm6, %xmm0
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3]
; AVX512DQ-NEXT:    vmovdqa64 %xmm29, %xmm3
; AVX512DQ-NEXT:    vpshufb %xmm1, %xmm3, %xmm3
; AVX512DQ-NEXT:    vpshufb %xmm1, %xmm7, %xmm1
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1],xmm1[2],xmm3[2],xmm1[3],xmm3[3]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[3]
; AVX512DQ-NEXT:    vpsrlq $24, %zmm16, %zmm1
; AVX512DQ-NEXT:    vpmovqb %zmm1, %xmm1
; AVX512DQ-NEXT:    vpblendd {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm0, %zmm8, %zmm0
; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpbroadcastd {{.*#+}} xmm0 = [0,0,4,12,0,0,4,12,0,0,4,12,0,0,4,12]
; AVX512DQ-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX512DQ-NEXT:    vpshufb %xmm0, %xmm1, %xmm1
; AVX512DQ-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX512DQ-NEXT:    vpshufb %xmm0, %xmm2, %xmm2
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX512DQ-NEXT:    vpbroadcastd {{.*#+}} xmm1 = [4,12,0,0,4,12,0,0,4,12,0,0,4,12,0,0]
; AVX512DQ-NEXT:    vpshufb %xmm1, %xmm10, %xmm3
; AVX512DQ-NEXT:    vmovdqa64 %xmm18, %xmm10
; AVX512DQ-NEXT:    vmovdqa64 %xmm18, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-NEXT:    vpshufb %xmm1, %xmm10, %xmm4
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX512DQ-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX512DQ-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm4 = ymm3[0,1,2,3,4,5,6],ymm2[7]
; AVX512DQ-NEXT:    vmovd {{.*#+}} xmm2 = [0,0,4,12,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX512DQ-NEXT:    vmovdqa64 %xmm24, %xmm11
; AVX512DQ-NEXT:    vpshufb %xmm2, %xmm11, %xmm3
; AVX512DQ-NEXT:    vmovdqa64 %xmm28, %xmm7
; AVX512DQ-NEXT:    vpshufb %xmm2, %xmm7, %xmm8
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm3[0],xmm8[1],xmm3[1],xmm8[2],xmm3[2],xmm8[3],xmm3[3]
; AVX512DQ-NEXT:    vmovd {{.*#+}} xmm3 = [4,12,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX512DQ-NEXT:    vmovdqa64 %xmm30, %xmm7
; AVX512DQ-NEXT:    vpshufb %xmm3, %xmm7, %xmm9
; AVX512DQ-NEXT:    vmovdqa64 %xmm31, %xmm12
; AVX512DQ-NEXT:    vpshufb %xmm3, %xmm12, %xmm15
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm15[0],xmm9[0],xmm15[1],xmm9[1],xmm15[2],xmm9[2],xmm15[3],xmm9[3]
; AVX512DQ-NEXT:    vinserti128 $1, %xmm8, %ymm0, %ymm8
; AVX512DQ-NEXT:    vinserti128 $1, %xmm9, %ymm0, %ymm9
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm8 = ymm9[0,1,2,3,4],ymm8[5],ymm9[6,7]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm4 = ymm8[0,1,2,3,4,5],ymm4[6,7]
; AVX512DQ-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
; AVX512DQ-NEXT:    vpshufb %xmm0, %xmm7, %xmm8
; AVX512DQ-NEXT:    vpshufb %xmm0, %xmm13, %xmm9
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm9[0],xmm8[0],xmm9[1],xmm8[1],xmm9[2],xmm8[2],xmm9[3],xmm8[3]
; AVX512DQ-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
; AVX512DQ-NEXT:    vpshufb %xmm1, %xmm7, %xmm9
; AVX512DQ-NEXT:    vpshufb %xmm1, %xmm5, %xmm15
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm15[0],xmm9[0],xmm15[1],xmm9[1],xmm15[2],xmm9[2],xmm15[3],xmm9[3]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} xmm8 = xmm9[0,1,2],xmm8[3]
; AVX512DQ-NEXT:    vpsrlq $32, %zmm17, %zmm9
; AVX512DQ-NEXT:    vpmovqb %zmm9, %xmm9
; AVX512DQ-NEXT:    vpblendd {{.*#+}} xmm8 = xmm9[0,1],xmm8[2,3]
; AVX512DQ-NEXT:    vinserti32x4 $2, %xmm8, %zmm0, %zmm8
; AVX512DQ-NEXT:    vinserti64x4 $1, %ymm4, %zmm0, %zmm8 {%k1}
; AVX512DQ-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
; AVX512DQ-NEXT:    vpshufb %xmm0, %xmm13, %xmm4
; AVX512DQ-NEXT:    vmovdqa64 %xmm22, %xmm5
; AVX512DQ-NEXT:    vpshufb %xmm0, %xmm5, %xmm9
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm9[0],xmm4[0],xmm9[1],xmm4[1],xmm9[2],xmm4[2],xmm9[3],xmm4[3]
; AVX512DQ-NEXT:    vmovdqa64 %xmm26, %xmm5
; AVX512DQ-NEXT:    vpshufb %xmm1, %xmm5, %xmm9
; AVX512DQ-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX512DQ-NEXT:    vpshufb %xmm1, %xmm5, %xmm15
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm15[0],xmm9[0],xmm15[1],xmm9[1],xmm15[2],xmm9[2],xmm15[3],xmm9[3]
; AVX512DQ-NEXT:    vinserti128 $1, %xmm4, %ymm0, %ymm4
; AVX512DQ-NEXT:    vinserti128 $1, %xmm9, %ymm0, %ymm9
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm4 = ymm9[0,1,2,3,4,5,6],ymm4[7]
; AVX512DQ-NEXT:    vpshufb %xmm2, %xmm14, %xmm9
; AVX512DQ-NEXT:    vmovdqa64 %xmm25, %xmm20
; AVX512DQ-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX512DQ-NEXT:    vpshufb %xmm2, %xmm5, %xmm2
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm9[0],xmm2[1],xmm9[1],xmm2[2],xmm9[2],xmm2[3],xmm9[3]
; AVX512DQ-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX512DQ-NEXT:    vpshufb %xmm3, %xmm5, %xmm9
; AVX512DQ-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
; AVX512DQ-NEXT:    vpshufb %xmm3, %xmm14, %xmm3
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm9[0],xmm3[1],xmm9[1],xmm3[2],xmm9[2],xmm3[3],xmm9[3]
; AVX512DQ-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX512DQ-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4],ymm2[5],ymm3[6,7]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm4[6,7]
; AVX512DQ-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
; AVX512DQ-NEXT:    vpshufb %xmm0, %xmm7, %xmm3
; AVX512DQ-NEXT:    vpshufb %xmm0, %xmm6, %xmm0
; AVX512DQ-NEXT:    vmovdqa64 %xmm6, %xmm25
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3]
; AVX512DQ-NEXT:    vmovdqa64 %xmm29, %xmm3
; AVX512DQ-NEXT:    vpshufb %xmm1, %xmm3, %xmm3
; AVX512DQ-NEXT:    vmovdqa (%rsp), %xmm4 # 16-byte Reload
; AVX512DQ-NEXT:    vpshufb %xmm1, %xmm4, %xmm1
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1],xmm1[2],xmm3[2],xmm1[3],xmm3[3]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[3]
; AVX512DQ-NEXT:    vpsrlq $32, %zmm16, %zmm1
; AVX512DQ-NEXT:    vmovdqa64 %zmm16, %zmm18
; AVX512DQ-NEXT:    vpmovqb %zmm1, %xmm1
; AVX512DQ-NEXT:    vpblendd {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm0, %zmm8, %zmm0
; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpbroadcastd {{.*#+}} xmm0 = [0,0,5,13,0,0,5,13,0,0,5,13,0,0,5,13]
; AVX512DQ-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX512DQ-NEXT:    vpshufb %xmm0, %xmm1, %xmm1
; AVX512DQ-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX512DQ-NEXT:    vpshufb %xmm0, %xmm2, %xmm2
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX512DQ-NEXT:    vpbroadcastd {{.*#+}} xmm1 = [5,13,0,0,5,13,0,0,5,13,0,0,5,13,0,0]
; AVX512DQ-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX512DQ-NEXT:    vpshufb %xmm1, %xmm3, %xmm3
; AVX512DQ-NEXT:    vpshufb %xmm1, %xmm10, %xmm4
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX512DQ-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX512DQ-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm4 = ymm3[0,1,2,3,4,5,6],ymm2[7]
; AVX512DQ-NEXT:    vmovd {{.*#+}} xmm2 = [0,0,5,13,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX512DQ-NEXT:    vpshufb %xmm2, %xmm11, %xmm3
; AVX512DQ-NEXT:    vmovdqa64 %xmm24, %xmm27
; AVX512DQ-NEXT:    vmovdqa64 %xmm28, %xmm5
; AVX512DQ-NEXT:    vpshufb %xmm2, %xmm5, %xmm8
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm3[0],xmm8[1],xmm3[1],xmm8[2],xmm3[2],xmm8[3],xmm3[3]
; AVX512DQ-NEXT:    vmovd {{.*#+}} xmm3 = [5,13,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX512DQ-NEXT:    vmovdqa64 %xmm30, %xmm10
; AVX512DQ-NEXT:    vpshufb %xmm3, %xmm10, %xmm9
; AVX512DQ-NEXT:    vpshufb %xmm3, %xmm12, %xmm15
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm15[0],xmm9[0],xmm15[1],xmm9[1],xmm15[2],xmm9[2],xmm15[3],xmm9[3]
; AVX512DQ-NEXT:    vinserti128 $1, %xmm8, %ymm0, %ymm8
; AVX512DQ-NEXT:    vinserti128 $1, %xmm9, %ymm0, %ymm9
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm8 = ymm9[0,1,2,3,4],ymm8[5],ymm9[6,7]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm4 = ymm8[0,1,2,3,4,5],ymm4[6,7]
; AVX512DQ-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX512DQ-NEXT:    vpshufb %xmm0, %xmm6, %xmm8
; AVX512DQ-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX512DQ-NEXT:    vpshufb %xmm0, %xmm6, %xmm9
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm9[0],xmm8[0],xmm9[1],xmm8[1],xmm9[2],xmm8[2],xmm9[3],xmm8[3]
; AVX512DQ-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX512DQ-NEXT:    vpshufb %xmm1, %xmm6, %xmm9
; AVX512DQ-NEXT:    vmovdqa64 %xmm21, %xmm11
; AVX512DQ-NEXT:    vpshufb %xmm1, %xmm11, %xmm15
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm15[0],xmm9[0],xmm15[1],xmm9[1],xmm15[2],xmm9[2],xmm15[3],xmm9[3]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} xmm8 = xmm9[0,1,2],xmm8[3]
; AVX512DQ-NEXT:    vpsrlq $40, %zmm17, %zmm9
; AVX512DQ-NEXT:    vmovdqa64 %zmm17, %zmm23
; AVX512DQ-NEXT:    vpmovqb %zmm9, %xmm9
; AVX512DQ-NEXT:    vpblendd {{.*#+}} xmm8 = xmm9[0,1],xmm8[2,3]
; AVX512DQ-NEXT:    vinserti32x4 $2, %xmm8, %zmm0, %zmm8
; AVX512DQ-NEXT:    vinserti64x4 $1, %ymm4, %zmm0, %zmm8 {%k1}
; AVX512DQ-NEXT:    vpshufb %xmm0, %xmm13, %xmm4
; AVX512DQ-NEXT:    vmovdqa64 %xmm13, %xmm30
; AVX512DQ-NEXT:    vmovdqa64 %xmm22, %xmm6
; AVX512DQ-NEXT:    vpshufb %xmm0, %xmm6, %xmm9
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm9[0],xmm4[0],xmm9[1],xmm4[1],xmm9[2],xmm4[2],xmm9[3],xmm4[3]
; AVX512DQ-NEXT:    vmovdqa64 %xmm26, %xmm13
; AVX512DQ-NEXT:    vpshufb %xmm1, %xmm13, %xmm9
; AVX512DQ-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX512DQ-NEXT:    vpshufb %xmm1, %xmm6, %xmm15
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm15[0],xmm9[0],xmm15[1],xmm9[1],xmm15[2],xmm9[2],xmm15[3],xmm9[3]
; AVX512DQ-NEXT:    vinserti128 $1, %xmm4, %ymm0, %ymm4
; AVX512DQ-NEXT:    vinserti128 $1, %xmm9, %ymm0, %ymm9
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm4 = ymm9[0,1,2,3,4,5,6],ymm4[7]
; AVX512DQ-NEXT:    vmovdqa64 %xmm20, %xmm9
; AVX512DQ-NEXT:    vpshufb %xmm2, %xmm9, %xmm9
; AVX512DQ-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload
; AVX512DQ-NEXT:    vpshufb %xmm2, %xmm12, %xmm2
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm9[0],xmm2[1],xmm9[1],xmm2[2],xmm9[2],xmm2[3],xmm9[3]
; AVX512DQ-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX512DQ-NEXT:    vpshufb %xmm3, %xmm15, %xmm9
; AVX512DQ-NEXT:    vmovdqa64 %xmm15, %xmm28
; AVX512DQ-NEXT:    vpshufb %xmm3, %xmm14, %xmm3
; AVX512DQ-NEXT:    vmovdqa64 %xmm14, %xmm16
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm9[0],xmm3[1],xmm9[1],xmm3[2],xmm9[2],xmm3[3],xmm9[3]
; AVX512DQ-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX512DQ-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4],ymm2[5],ymm3[6,7]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm4[6,7]
; AVX512DQ-NEXT:    vpshufb %xmm0, %xmm7, %xmm3
; AVX512DQ-NEXT:    vmovdqa64 %xmm7, %xmm24
; AVX512DQ-NEXT:    vmovdqa64 %xmm25, %xmm4
; AVX512DQ-NEXT:    vpshufb %xmm0, %xmm4, %xmm0
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3]
; AVX512DQ-NEXT:    vmovdqa64 %xmm29, %xmm7
; AVX512DQ-NEXT:    vpshufb %xmm1, %xmm7, %xmm3
; AVX512DQ-NEXT:    vmovdqa (%rsp), %xmm4 # 16-byte Reload
; AVX512DQ-NEXT:    vpshufb %xmm1, %xmm4, %xmm1
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1],xmm1[2],xmm3[2],xmm1[3],xmm3[3]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[3]
; AVX512DQ-NEXT:    vpsrlq $40, %zmm18, %zmm1
; AVX512DQ-NEXT:    vmovdqa64 %zmm18, %zmm26
; AVX512DQ-NEXT:    vpmovqb %zmm1, %xmm1
; AVX512DQ-NEXT:    vpblendd {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm0, %zmm8, %zmm0
; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-NEXT:    vpbroadcastd {{.*#+}} xmm0 = [0,0,6,14,0,0,6,14,0,0,6,14,0,0,6,14]
; AVX512DQ-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX512DQ-NEXT:    vpshufb %xmm0, %xmm2, %xmm1
; AVX512DQ-NEXT:    vmovdqa64 %xmm2, %xmm29
; AVX512DQ-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX512DQ-NEXT:    vpshufb %xmm0, %xmm2, %xmm2
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX512DQ-NEXT:    vpbroadcastd {{.*#+}} xmm1 = [6,14,0,0,6,14,0,0,6,14,0,0,6,14,0,0]
; AVX512DQ-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX512DQ-NEXT:    vpshufb %xmm1, %xmm3, %xmm3
; AVX512DQ-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX512DQ-NEXT:    vpshufb %xmm1, %xmm4, %xmm4
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX512DQ-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX512DQ-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm4 = ymm3[0,1,2,3,4,5,6],ymm2[7]
; AVX512DQ-NEXT:    vmovd {{.*#+}} xmm2 = [0,0,6,14,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX512DQ-NEXT:    vmovdqa64 %xmm27, %xmm3
; AVX512DQ-NEXT:    vpshufb %xmm2, %xmm3, %xmm3
; AVX512DQ-NEXT:    vpshufb %xmm2, %xmm5, %xmm8
; AVX512DQ-NEXT:    vmovdqa64 %xmm5, %xmm21
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm3[0],xmm8[1],xmm3[1],xmm8[2],xmm3[2],xmm8[3],xmm3[3]
; AVX512DQ-NEXT:    vmovd {{.*#+}} xmm3 = [6,14,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX512DQ-NEXT:    vpshufb %xmm3, %xmm10, %xmm9
; AVX512DQ-NEXT:    vmovdqa64 %xmm10, %xmm18
; AVX512DQ-NEXT:    vmovdqa64 %xmm31, %xmm5
; AVX512DQ-NEXT:    vpshufb %xmm3, %xmm5, %xmm15
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm15[0],xmm9[0],xmm15[1],xmm9[1],xmm15[2],xmm9[2],xmm15[3],xmm9[3]
; AVX512DQ-NEXT:    vinserti128 $1, %xmm8, %ymm0, %ymm8
; AVX512DQ-NEXT:    vinserti128 $1, %xmm9, %ymm0, %ymm9
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm8 = ymm9[0,1,2,3,4],ymm8[5],ymm9[6,7]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm4 = ymm8[0,1,2,3,4,5],ymm4[6,7]
; AVX512DQ-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX512DQ-NEXT:    vpshufb %xmm0, %xmm5, %xmm8
; AVX512DQ-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
; AVX512DQ-NEXT:    vpshufb %xmm0, %xmm9, %xmm9
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm9[0],xmm8[0],xmm9[1],xmm8[1],xmm9[2],xmm8[2],xmm9[3],xmm8[3]
; AVX512DQ-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; AVX512DQ-NEXT:    vpshufb %xmm1, %xmm10, %xmm9
; AVX512DQ-NEXT:    vpshufb %xmm1, %xmm11, %xmm15
; AVX512DQ-NEXT:    vmovdqa64 %xmm11, %xmm22
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm15[0],xmm9[0],xmm15[1],xmm9[1],xmm15[2],xmm9[2],xmm15[3],xmm9[3]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} xmm8 = xmm9[0,1,2],xmm8[3]
; AVX512DQ-NEXT:    vpsrlq $48, %zmm17, %zmm9
; AVX512DQ-NEXT:    vpmovqb %zmm9, %xmm9
; AVX512DQ-NEXT:    vpblendd {{.*#+}} xmm8 = xmm9[0,1],xmm8[2,3]
; AVX512DQ-NEXT:    vinserti32x4 $2, %xmm8, %zmm0, %zmm8
; AVX512DQ-NEXT:    vinserti64x4 $1, %ymm4, %zmm0, %zmm8 {%k1}
; AVX512DQ-NEXT:    vmovdqa64 %xmm30, %xmm4
; AVX512DQ-NEXT:    vpshufb %xmm0, %xmm4, %xmm4
; AVX512DQ-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
; AVX512DQ-NEXT:    vpshufb %xmm0, %xmm9, %xmm9
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm9[0],xmm4[0],xmm9[1],xmm4[1],xmm9[2],xmm4[2],xmm9[3],xmm4[3]
; AVX512DQ-NEXT:    vpshufb %xmm1, %xmm13, %xmm9
; AVX512DQ-NEXT:    vmovdqa64 %xmm13, %xmm19
; AVX512DQ-NEXT:    vpshufb %xmm1, %xmm6, %xmm15
; AVX512DQ-NEXT:    vmovdqa64 %xmm6, %xmm17
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm15[0],xmm9[0],xmm15[1],xmm9[1],xmm15[2],xmm9[2],xmm15[3],xmm9[3]
; AVX512DQ-NEXT:    vinserti128 $1, %xmm4, %ymm0, %ymm4
; AVX512DQ-NEXT:    vinserti128 $1, %xmm9, %ymm0, %ymm9
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm4 = ymm9[0,1,2,3,4,5,6],ymm4[7]
; AVX512DQ-NEXT:    vmovdqa64 %xmm20, %xmm6
; AVX512DQ-NEXT:    vpshufb %xmm2, %xmm6, %xmm9
; AVX512DQ-NEXT:    vpshufb %xmm2, %xmm12, %xmm2
; AVX512DQ-NEXT:    vmovdqa64 %xmm12, %xmm20
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm9[0],xmm2[1],xmm9[1],xmm2[2],xmm9[2],xmm2[3],xmm9[3]
; AVX512DQ-NEXT:    vmovdqa64 %xmm28, %xmm14
; AVX512DQ-NEXT:    vpshufb %xmm3, %xmm14, %xmm9
; AVX512DQ-NEXT:    vmovdqa64 %xmm16, %xmm13
; AVX512DQ-NEXT:    vpshufb %xmm3, %xmm13, %xmm3
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm9[0],xmm3[1],xmm9[1],xmm3[2],xmm9[2],xmm3[3],xmm9[3]
; AVX512DQ-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX512DQ-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4],ymm2[5],ymm3[6,7]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm4[6,7]
; AVX512DQ-NEXT:    vmovdqa64 %xmm24, %xmm15
; AVX512DQ-NEXT:    vpshufb %xmm0, %xmm15, %xmm3
; AVX512DQ-NEXT:    vmovdqa64 %xmm25, %xmm12
; AVX512DQ-NEXT:    vpshufb %xmm0, %xmm12, %xmm0
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3]
; AVX512DQ-NEXT:    vpshufb %xmm1, %xmm7, %xmm3
; AVX512DQ-NEXT:    vmovdqa64 %xmm7, %xmm16
; AVX512DQ-NEXT:    vmovdqa (%rsp), %xmm7 # 16-byte Reload
; AVX512DQ-NEXT:    vpshufb %xmm1, %xmm7, %xmm1
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1],xmm1[2],xmm3[2],xmm1[3],xmm3[3]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[3]
; AVX512DQ-NEXT:    vpsrlq $48, %zmm26, %zmm1
; AVX512DQ-NEXT:    vpmovqb %zmm1, %xmm1
; AVX512DQ-NEXT:    vpblendd {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm0, %zmm8, %zmm24
; AVX512DQ-NEXT:    vpbroadcastd {{.*#+}} xmm0 = [0,0,7,15,0,0,7,15,0,0,7,15,0,0,7,15]
; AVX512DQ-NEXT:    vmovdqa64 %xmm29, %xmm1
; AVX512DQ-NEXT:    vpshufb %xmm0, %xmm1, %xmm1
; AVX512DQ-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX512DQ-NEXT:    vpshufb %xmm0, %xmm2, %xmm2
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX512DQ-NEXT:    vpbroadcastd {{.*#+}} xmm1 = [7,15,0,0,7,15,0,0,7,15,0,0,7,15,0,0]
; AVX512DQ-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX512DQ-NEXT:    vpshufb %xmm1, %xmm3, %xmm3
; AVX512DQ-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX512DQ-NEXT:    vpshufb %xmm1, %xmm4, %xmm4
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX512DQ-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX512DQ-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm4 = ymm3[0,1,2,3,4,5,6],ymm2[7]
; AVX512DQ-NEXT:    vmovd {{.*#+}} xmm2 = [0,0,7,15,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX512DQ-NEXT:    vmovdqa64 %xmm27, %xmm3
; AVX512DQ-NEXT:    vpshufb %xmm2, %xmm3, %xmm3
; AVX512DQ-NEXT:    vmovdqa64 %xmm21, %xmm8
; AVX512DQ-NEXT:    vpshufb %xmm2, %xmm8, %xmm8
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm8[0],xmm3[0],xmm8[1],xmm3[1],xmm8[2],xmm3[2],xmm8[3],xmm3[3]
; AVX512DQ-NEXT:    vmovd {{.*#+}} xmm3 = [7,15,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
; AVX512DQ-NEXT:    vmovdqa64 %xmm18, %xmm9
; AVX512DQ-NEXT:    vpshufb %xmm3, %xmm9, %xmm9
; AVX512DQ-NEXT:    vmovdqa64 %xmm31, %xmm11
; AVX512DQ-NEXT:    vpshufb %xmm3, %xmm11, %xmm11
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm11[0],xmm9[0],xmm11[1],xmm9[1],xmm11[2],xmm9[2],xmm11[3],xmm9[3]
; AVX512DQ-NEXT:    vinserti128 $1, %xmm8, %ymm0, %ymm8
; AVX512DQ-NEXT:    vinserti128 $1, %xmm9, %ymm0, %ymm9
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm8 = ymm9[0,1,2,3,4],ymm8[5],ymm9[6,7]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm4 = ymm8[0,1,2,3,4,5],ymm4[6,7]
; AVX512DQ-NEXT:    vpshufb %xmm0, %xmm5, %xmm8
; AVX512DQ-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX512DQ-NEXT:    vpshufb %xmm0, %xmm5, %xmm9
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm8 = xmm9[0],xmm8[0],xmm9[1],xmm8[1],xmm9[2],xmm8[2],xmm9[3],xmm8[3]
; AVX512DQ-NEXT:    vpshufb %xmm1, %xmm10, %xmm9
; AVX512DQ-NEXT:    vmovdqa64 %xmm22, %xmm5
; AVX512DQ-NEXT:    vpshufb %xmm1, %xmm5, %xmm11
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm11[0],xmm9[0],xmm11[1],xmm9[1],xmm11[2],xmm9[2],xmm11[3],xmm9[3]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} xmm8 = xmm9[0,1,2],xmm8[3]
; AVX512DQ-NEXT:    vpsrlq $56, %zmm23, %zmm9
; AVX512DQ-NEXT:    vpmovqb %zmm9, %xmm9
; AVX512DQ-NEXT:    vpblendd {{.*#+}} xmm8 = xmm9[0,1],xmm8[2,3]
; AVX512DQ-NEXT:    vinserti32x4 $2, %xmm8, %zmm0, %zmm8
; AVX512DQ-NEXT:    vinserti64x4 $1, %ymm4, %zmm0, %zmm8 {%k1}
; AVX512DQ-NEXT:    vmovdqa64 %xmm30, %xmm4
; AVX512DQ-NEXT:    vpshufb %xmm0, %xmm4, %xmm4
; AVX512DQ-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX512DQ-NEXT:    vpshufb %xmm0, %xmm5, %xmm9
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm9[0],xmm4[0],xmm9[1],xmm4[1],xmm9[2],xmm4[2],xmm9[3],xmm4[3]
; AVX512DQ-NEXT:    vmovdqa64 %xmm19, %xmm5
; AVX512DQ-NEXT:    vpshufb %xmm1, %xmm5, %xmm9
; AVX512DQ-NEXT:    vmovdqa64 %xmm17, %xmm5
; AVX512DQ-NEXT:    vpshufb %xmm1, %xmm5, %xmm11
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm11[0],xmm9[0],xmm11[1],xmm9[1],xmm11[2],xmm9[2],xmm11[3],xmm9[3]
; AVX512DQ-NEXT:    vinserti128 $1, %xmm4, %ymm0, %ymm4
; AVX512DQ-NEXT:    vinserti128 $1, %xmm9, %ymm0, %ymm9
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm4 = ymm9[0,1,2,3,4,5,6],ymm4[7]
; AVX512DQ-NEXT:    vpshufb %xmm2, %xmm6, %xmm6
; AVX512DQ-NEXT:    vmovdqa64 %xmm20, %xmm5
; AVX512DQ-NEXT:    vpshufb %xmm2, %xmm5, %xmm2
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm6[0],xmm2[1],xmm6[1],xmm2[2],xmm6[2],xmm2[3],xmm6[3]
; AVX512DQ-NEXT:    vpshufb %xmm3, %xmm14, %xmm5
; AVX512DQ-NEXT:    vpshufb %xmm3, %xmm13, %xmm3
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm5[0],xmm3[1],xmm5[1],xmm3[2],xmm5[2],xmm3[3],xmm5[3]
; AVX512DQ-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX512DQ-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4],ymm2[5],ymm3[6,7]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm4[6,7]
; AVX512DQ-NEXT:    vpshufb %xmm0, %xmm15, %xmm3
; AVX512DQ-NEXT:    vpshufb %xmm0, %xmm12, %xmm0
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3]
; AVX512DQ-NEXT:    vmovdqa64 %xmm16, %xmm3
; AVX512DQ-NEXT:    vpshufb %xmm1, %xmm3, %xmm3
; AVX512DQ-NEXT:    vpshufb %xmm1, %xmm7, %xmm1
; AVX512DQ-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1],xmm1[2],xmm3[2],xmm1[3],xmm3[3]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[3]
; AVX512DQ-NEXT:    vpsrlq $56, %zmm26, %zmm1
; AVX512DQ-NEXT:    vpmovqb %zmm1, %xmm1
; AVX512DQ-NEXT:    vpblendd {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3]
; AVX512DQ-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]
; AVX512DQ-NEXT:    vinserti64x4 $0, %ymm0, %zmm8, %zmm0
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm1, (%rsi)
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm1, (%rdx)
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm1, (%rcx)
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm1, (%r8)
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm1, (%r9)
; AVX512DQ-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-NEXT:    vmovaps %zmm1, (%rax)
; AVX512DQ-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-NEXT:    vmovdqa64 %zmm24, (%rax)
; AVX512DQ-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-NEXT:    vmovdqa64 %zmm0, (%rax)
; AVX512DQ-NEXT:    addq $520, %rsp # imm = 0x208
; AVX512DQ-NEXT:    vzeroupper
; AVX512DQ-NEXT:    retq
;
; AVX512DQ-FCP-LABEL: load_i8_stride8_vf64:
; AVX512DQ-FCP:       # %bb.0:
; AVX512DQ-FCP-NEXT:    subq $440, %rsp # imm = 0x1B8
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm2 = [0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12]
; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm0 = [0,2,2,3,0,2,4,6]
; AVX512DQ-FCP-NEXT:    vmovdqa 480(%rdi), %ymm1
; AVX512DQ-FCP-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FCP-NEXT:    vpermd %ymm1, %ymm0, %ymm3
; AVX512DQ-FCP-NEXT:    vpshufb %ymm2, %ymm3, %ymm1
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm3, %ymm19
; AVX512DQ-FCP-NEXT:    vmovdqa %ymm2, %ymm14
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm3 = [0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12]
; AVX512DQ-FCP-NEXT:    vmovdqa 448(%rdi), %ymm2
; AVX512DQ-FCP-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FCP-NEXT:    vpermd %ymm2, %ymm0, %ymm4
; AVX512DQ-FCP-NEXT:    vpshufb %ymm3, %ymm4, %ymm2
; AVX512DQ-FCP-NEXT:    vmovdqa %ymm3, %ymm9
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5,6],ymm1[7]
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm2 = [0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12]
; AVX512DQ-FCP-NEXT:    vmovdqa 416(%rdi), %ymm3
; AVX512DQ-FCP-NEXT:    vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FCP-NEXT:    vpermd %ymm3, %ymm0, %ymm12
; AVX512DQ-FCP-NEXT:    vpshufb %ymm2, %ymm12, %ymm3
; AVX512DQ-FCP-NEXT:    vmovdqa %ymm2, %ymm11
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm10 = [0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12]
; AVX512DQ-FCP-NEXT:    vmovdqa 384(%rdi), %ymm5
; AVX512DQ-FCP-NEXT:    vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FCP-NEXT:    vpermd %ymm5, %ymm0, %ymm13
; AVX512DQ-FCP-NEXT:    vpshufb %ymm10, %ymm13, %ymm5
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm3 = ymm5[0,1,2,3,4],ymm3[5],ymm5[6,7]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm3[0,1,2,3,4,5],ymm1[6,7]
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm3 = [0,0,0,8,0,0,0,8,0,0,0,8,0,0,0,8]
; AVX512DQ-FCP-NEXT:    vmovdqa 368(%rdi), %xmm2
; AVX512DQ-FCP-NEXT:    vpshufb %xmm3, %xmm2, %xmm5
; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm2, %xmm30
; AVX512DQ-FCP-NEXT:    vmovdqa 352(%rdi), %xmm2
; AVX512DQ-FCP-NEXT:    vpshufb %xmm3, %xmm2, %xmm6
; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm2, %xmm25
; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm5 = [0,8,0,0,0,8,0,0,0,8,0,0,0,8,0,0]
; AVX512DQ-FCP-NEXT:    vmovdqa 336(%rdi), %xmm2
; AVX512DQ-FCP-NEXT:    vpshufb %xmm5, %xmm2, %xmm7
; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm2, %xmm24
; AVX512DQ-FCP-NEXT:    vmovdqa 320(%rdi), %xmm2
; AVX512DQ-FCP-NEXT:    vpshufb %xmm5, %xmm2, %xmm8
; AVX512DQ-FCP-NEXT:    vmovdqa %xmm2, %xmm15
; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm7 = xmm8[0],xmm7[0],xmm8[1],xmm7[1],xmm8[2],xmm7[2],xmm8[3],xmm7[3]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} xmm6 = xmm7[0,1,2],xmm6[3]
; AVX512DQ-FCP-NEXT:    vmovdqa64 256(%rdi), %zmm2
; AVX512DQ-FCP-NEXT:    vpmovqb %zmm2, %xmm7
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm2, %zmm28
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} xmm6 = xmm7[0,1],xmm6[2,3]
; AVX512DQ-FCP-NEXT:    vinserti32x4 $2, %xmm6, %zmm0, %zmm16
; AVX512DQ-FCP-NEXT:    movb $-64, %al
; AVX512DQ-FCP-NEXT:    kmovw %eax, %k1
; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm16 {%k1}
; AVX512DQ-FCP-NEXT:    vmovdqa 224(%rdi), %ymm1
; AVX512DQ-FCP-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FCP-NEXT:    vpermd %ymm1, %ymm0, %ymm2
; AVX512DQ-FCP-NEXT:    vpshufb %ymm14, %ymm2, %ymm1
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm2, %ymm23
; AVX512DQ-FCP-NEXT:    vmovdqa 192(%rdi), %ymm2
; AVX512DQ-FCP-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FCP-NEXT:    vpermd %ymm2, %ymm0, %ymm2
; AVX512DQ-FCP-NEXT:    vpshufb %ymm9, %ymm2, %ymm6
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm2, %ymm27
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm6 = ymm6[0,1,2,3,4,5,6],ymm1[7]
; AVX512DQ-FCP-NEXT:    vmovdqa 160(%rdi), %ymm1
; AVX512DQ-FCP-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FCP-NEXT:    vmovdqa 128(%rdi), %ymm2
; AVX512DQ-FCP-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FCP-NEXT:    vpermd %ymm1, %ymm0, %ymm1
; AVX512DQ-FCP-NEXT:    vpshufb %ymm11, %ymm1, %ymm7
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm1, %ymm18
; AVX512DQ-FCP-NEXT:    vpermd %ymm2, %ymm0, %ymm0
; AVX512DQ-FCP-NEXT:    vpshufb %ymm10, %ymm0, %ymm8
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm0, %ymm17
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm7 = ymm8[0,1,2,3,4],ymm7[5],ymm8[6,7]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm7[0,1,2,3,4,5],ymm6[6,7]
; AVX512DQ-FCP-NEXT:    vmovdqa 112(%rdi), %xmm10
; AVX512DQ-FCP-NEXT:    vpshufb %xmm3, %xmm10, %xmm7
; AVX512DQ-FCP-NEXT:    vmovdqa 96(%rdi), %xmm0
; AVX512DQ-FCP-NEXT:    vpshufb %xmm3, %xmm0, %xmm3
; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm0, %xmm20
; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm7[0],xmm3[1],xmm7[1],xmm3[2],xmm7[2],xmm3[3],xmm7[3]
; AVX512DQ-FCP-NEXT:    vmovdqa 80(%rdi), %xmm2
; AVX512DQ-FCP-NEXT:    vpshufb %xmm5, %xmm2, %xmm0
; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm2, %xmm31
; AVX512DQ-FCP-NEXT:    vmovdqa 64(%rdi), %xmm2
; AVX512DQ-FCP-NEXT:    vpshufb %xmm5, %xmm2, %xmm5
; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm2, %xmm22
; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm5[0],xmm0[0],xmm5[1],xmm0[1],xmm5[2],xmm0[2],xmm5[3],xmm0[3]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm0[0,1,2],xmm3[3]
; AVX512DQ-FCP-NEXT:    vmovdqa64 (%rdi), %zmm29
; AVX512DQ-FCP-NEXT:    vpmovqb %zmm29, %xmm3
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm3[0,1],xmm0[2,3]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm16, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm6 = [1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm19, %ymm11
; AVX512DQ-FCP-NEXT:    vpshufb %ymm6, %ymm11, %ymm0
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm7 = [1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13]
; AVX512DQ-FCP-NEXT:    vpshufb %ymm7, %ymm4, %ymm1
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm4, %ymm26
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5,6],ymm0[7]
; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm1 = ymm12[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,17,21,25,29,u,u,u,u,u,u,u,u]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm12, %ymm19
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm2 = [1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13]
; AVX512DQ-FCP-NEXT:    vpshufb %ymm2, %ymm13, %ymm3
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm13, %ymm21
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm2, %ymm16
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm3[0,1,2,3,4],ymm1[5],ymm3[6,7]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm2 = ymm1[0,1,2,3,4,5],ymm0[6,7]
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm1 = [0,0,1,9,0,0,1,9,0,0,1,9,0,0,1,9]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm30, %xmm13
; AVX512DQ-FCP-NEXT:    vpshufb %xmm1, %xmm13, %xmm3
; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm25, %xmm14
; AVX512DQ-FCP-NEXT:    vpshufb %xmm1, %xmm14, %xmm5
; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm5[0],xmm3[0],xmm5[1],xmm3[1],xmm5[2],xmm3[2],xmm5[3],xmm3[3]
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm5 = [1,9,0,0,1,9,0,0,1,9,0,0,1,9,0,0]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm24, %xmm9
; AVX512DQ-FCP-NEXT:    vpshufb %xmm5, %xmm9, %xmm0
; AVX512DQ-FCP-NEXT:    vmovdqa %xmm15, %xmm12
; AVX512DQ-FCP-NEXT:    vpshufb %xmm5, %xmm15, %xmm15
; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm15[0],xmm0[0],xmm15[1],xmm0[1],xmm15[2],xmm0[2],xmm15[3],xmm0[3]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm0[0,1,2],xmm3[3]
; AVX512DQ-FCP-NEXT:    vpsrlq $8, %zmm28, %zmm3
; AVX512DQ-FCP-NEXT:    vpmovqb %zmm3, %xmm3
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} xmm0 = xmm3[0,1],xmm0[2,3]
; AVX512DQ-FCP-NEXT:    vinserti32x4 $2, %xmm0, %zmm0, %zmm0
; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm2, %zmm0, %zmm0 {%k1}
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm23, %ymm8
; AVX512DQ-FCP-NEXT:    vpshufb %ymm6, %ymm8, %ymm2
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm27, %ymm4
; AVX512DQ-FCP-NEXT:    vpshufb %ymm7, %ymm4, %ymm3
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4,5,6],ymm2[7]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm18, %ymm7
; AVX512DQ-FCP-NEXT:    vpshufb {{.*#+}} ymm3 = ymm7[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,17,21,25,29,u,u,u,u,u,u,u,u]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm17, %ymm6
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm16, %ymm15
; AVX512DQ-FCP-NEXT:    vpshufb %ymm15, %ymm6, %ymm15
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm3 = ymm15[0,1,2,3,4],ymm3[5],ymm15[6,7]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4,5],ymm2[6,7]
; AVX512DQ-FCP-NEXT:    vpshufb %xmm1, %xmm10, %xmm3
; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm10, %xmm16
; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm20, %xmm10
; AVX512DQ-FCP-NEXT:    vpshufb %xmm1, %xmm10, %xmm1
; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1],xmm1[2],xmm3[2],xmm1[3],xmm3[3]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm31, %xmm15
; AVX512DQ-FCP-NEXT:    vpshufb %xmm5, %xmm15, %xmm3
; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm22, %xmm15
; AVX512DQ-FCP-NEXT:    vpshufb %xmm5, %xmm15, %xmm5
; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm5[0],xmm3[0],xmm5[1],xmm3[1],xmm5[2],xmm3[2],xmm5[3],xmm3[3]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} xmm1 = xmm3[0,1,2],xmm1[3]
; AVX512DQ-FCP-NEXT:    vpsrlq $8, %zmm29, %zmm3
; AVX512DQ-FCP-NEXT:    vpmovqb %zmm3, %xmm3
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} xmm1 = xmm3[0,1],xmm1[2,3]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm0, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm1 = [2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14]
; AVX512DQ-FCP-NEXT:    vpshufb %ymm1, %ymm11, %ymm0
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm11, %ymm20
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm1, %ymm23
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm2 = [2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm26, %ymm11
; AVX512DQ-FCP-NEXT:    vpshufb %ymm2, %ymm11, %ymm1
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm2, %ymm24
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5,6],ymm0[7]
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm2 = [2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm19, %ymm1
; AVX512DQ-FCP-NEXT:    vpshufb %ymm2, %ymm1, %ymm1
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm2, %ymm25
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm3 = [2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm21, %ymm2
; AVX512DQ-FCP-NEXT:    vpshufb %ymm3, %ymm2, %ymm2
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm3, %ymm26
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3,4],ymm1[5],ymm2[6,7]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm1 = [0,0,2,10,0,0,2,10,0,0,2,10,0,0,2,10]
; AVX512DQ-FCP-NEXT:    vpshufb %xmm1, %xmm13, %xmm2
; AVX512DQ-FCP-NEXT:    vpshufb %xmm1, %xmm14, %xmm3
; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm3 = [2,10,0,0,2,10,0,0,2,10,0,0,2,10,0,0]
; AVX512DQ-FCP-NEXT:    vpshufb %xmm3, %xmm9, %xmm5
; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm9, %xmm17
; AVX512DQ-FCP-NEXT:    vpshufb %xmm3, %xmm12, %xmm15
; AVX512DQ-FCP-NEXT:    vmovdqa %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm15[0],xmm5[0],xmm15[1],xmm5[1],xmm15[2],xmm5[2],xmm15[3],xmm5[3]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} xmm2 = xmm5[0,1,2],xmm2[3]
; AVX512DQ-FCP-NEXT:    vpsrlq $16, %zmm28, %zmm5
; AVX512DQ-FCP-NEXT:    vpmovqb %zmm5, %xmm5
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} xmm2 = xmm5[0,1],xmm2[2,3]
; AVX512DQ-FCP-NEXT:    vinserti32x4 $2, %xmm2, %zmm0, %zmm2
; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm0, %zmm0, %zmm2 {%k1}
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm23, %ymm0
; AVX512DQ-FCP-NEXT:    vpshufb %ymm0, %ymm8, %ymm0
; AVX512DQ-FCP-NEXT:    vmovdqa %ymm8, %ymm9
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm24, %ymm5
; AVX512DQ-FCP-NEXT:    vpshufb %ymm5, %ymm4, %ymm5
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm5[0,1,2,3,4,5,6],ymm0[7]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm25, %ymm4
; AVX512DQ-FCP-NEXT:    vpshufb %ymm4, %ymm7, %ymm5
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm18, %ymm25
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm26, %ymm4
; AVX512DQ-FCP-NEXT:    vpshufb %ymm4, %ymm6, %ymm15
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm6, %ymm26
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm5 = ymm15[0,1,2,3,4],ymm5[5],ymm15[6,7]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm5[0,1,2,3,4,5],ymm0[6,7]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm16, %xmm6
; AVX512DQ-FCP-NEXT:    vpshufb %xmm1, %xmm6, %xmm5
; AVX512DQ-FCP-NEXT:    vpshufb %xmm1, %xmm10, %xmm1
; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm10, %xmm16
; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm5[0],xmm1[1],xmm5[1],xmm1[2],xmm5[2],xmm1[3],xmm5[3]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm31, %xmm7
; AVX512DQ-FCP-NEXT:    vpshufb %xmm3, %xmm7, %xmm5
; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm22, %xmm8
; AVX512DQ-FCP-NEXT:    vpshufb %xmm3, %xmm8, %xmm3
; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm5[0],xmm3[1],xmm5[1],xmm3[2],xmm5[2],xmm3[3],xmm5[3]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} xmm1 = xmm3[0,1,2],xmm1[3]
; AVX512DQ-FCP-NEXT:    vpsrlq $16, %zmm29, %zmm3
; AVX512DQ-FCP-NEXT:    vpmovqb %zmm3, %xmm3
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} xmm1 = xmm3[0,1],xmm1[2,3]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm2, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm1 = [3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm20, %ymm0
; AVX512DQ-FCP-NEXT:    vpshufb %ymm1, %ymm0, %ymm0
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm1, %ymm18
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm2 = [3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15]
; AVX512DQ-FCP-NEXT:    vpshufb %ymm2, %ymm11, %ymm1
; AVX512DQ-FCP-NEXT:    vmovdqa %ymm2, %ymm11
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5,6],ymm0[7]
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm10 = [3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm19, %ymm1
; AVX512DQ-FCP-NEXT:    vpshufb %ymm10, %ymm1, %ymm1
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm4 = [3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm21, %ymm2
; AVX512DQ-FCP-NEXT:    vpshufb %ymm4, %ymm2, %ymm2
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3,4],ymm1[5],ymm2[6,7]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm1 = [0,0,3,11,0,0,3,11,0,0,3,11,0,0,3,11]
; AVX512DQ-FCP-NEXT:    vpshufb %xmm1, %xmm13, %xmm2
; AVX512DQ-FCP-NEXT:    vpshufb %xmm1, %xmm14, %xmm3
; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm3 = [3,11,0,0,3,11,0,0,3,11,0,0,3,11,0,0]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm17, %xmm15
; AVX512DQ-FCP-NEXT:    vpshufb %xmm3, %xmm15, %xmm5
; AVX512DQ-FCP-NEXT:    vpshufb %xmm3, %xmm12, %xmm12
; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm12[0],xmm5[0],xmm12[1],xmm5[1],xmm12[2],xmm5[2],xmm12[3],xmm5[3]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} xmm2 = xmm5[0,1,2],xmm2[3]
; AVX512DQ-FCP-NEXT:    vpsrlq $24, %zmm28, %zmm5
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm28, %zmm24
; AVX512DQ-FCP-NEXT:    vpmovqb %zmm5, %xmm5
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} xmm2 = xmm5[0,1],xmm2[2,3]
; AVX512DQ-FCP-NEXT:    vinserti32x4 $2, %xmm2, %zmm0, %zmm2
; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm0, %zmm0, %zmm2 {%k1}
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm18, %ymm0
; AVX512DQ-FCP-NEXT:    vpshufb %ymm0, %ymm9, %ymm0
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm27, %ymm5
; AVX512DQ-FCP-NEXT:    vpshufb %ymm11, %ymm5, %ymm5
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm5[0,1,2,3,4,5,6],ymm0[7]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm25, %ymm5
; AVX512DQ-FCP-NEXT:    vpshufb %ymm10, %ymm5, %ymm5
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm26, %ymm9
; AVX512DQ-FCP-NEXT:    vpshufb %ymm4, %ymm9, %ymm4
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3,4],ymm5[5],ymm4[6,7]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm4[0,1,2,3,4,5],ymm0[6,7]
; AVX512DQ-FCP-NEXT:    vpshufb %xmm1, %xmm6, %xmm4
; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm6, %xmm21
; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm16, %xmm5
; AVX512DQ-FCP-NEXT:    vpshufb %xmm1, %xmm5, %xmm1
; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm4[0],xmm1[1],xmm4[1],xmm1[2],xmm4[2],xmm1[3],xmm4[3]
; AVX512DQ-FCP-NEXT:    vpshufb %xmm3, %xmm7, %xmm4
; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm31, %xmm16
; AVX512DQ-FCP-NEXT:    vpshufb %xmm3, %xmm8, %xmm3
; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm22, %xmm6
; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} xmm1 = xmm3[0,1,2],xmm1[3]
; AVX512DQ-FCP-NEXT:    vpsrlq $24, %zmm29, %zmm3
; AVX512DQ-FCP-NEXT:    vpmovqb %zmm3, %xmm3
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} xmm1 = xmm3[0,1],xmm1[2,3]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm2, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm0 = [1,3,2,3,1,3,5,7]
; AVX512DQ-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm2 # 32-byte Folded Reload
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm17 = [0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm17, %ymm1
; AVX512DQ-FCP-NEXT:    vpshufb %ymm1, %ymm2, %ymm1
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm2, %ymm28
; AVX512DQ-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm3 # 32-byte Folded Reload
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm18 = [0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm18, %ymm2
; AVX512DQ-FCP-NEXT:    vpshufb %ymm2, %ymm3, %ymm2
; AVX512DQ-FCP-NEXT:    vmovdqa %ymm3, %ymm4
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5,6],ymm1[7]
; AVX512DQ-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm2 # 32-byte Folded Reload
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm7 = [0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12]
; AVX512DQ-FCP-NEXT:    vpshufb %ymm7, %ymm2, %ymm9
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm2, %ymm19
; AVX512DQ-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm2 # 32-byte Folded Reload
; AVX512DQ-FCP-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm8 = [0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12]
; AVX512DQ-FCP-NEXT:    vpshufb %ymm8, %ymm2, %ymm10
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm9 = ymm10[0,1,2,3,4],ymm9[5],ymm10[6,7]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm9[0,1,2,3,4,5],ymm1[6,7]
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm11 = [0,0,4,12,0,0,4,12,0,0,4,12,0,0,4,12]
; AVX512DQ-FCP-NEXT:    vpshufb %xmm11, %xmm13, %xmm9
; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm30, %xmm20
; AVX512DQ-FCP-NEXT:    vpshufb %xmm11, %xmm14, %xmm10
; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm14, %xmm22
; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm10[0],xmm9[0],xmm10[1],xmm9[1],xmm10[2],xmm9[2],xmm10[3],xmm9[3]
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm12 = [4,12,0,0,4,12,0,0,4,12,0,0,4,12,0,0]
; AVX512DQ-FCP-NEXT:    vpshufb %xmm12, %xmm15, %xmm10
; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm15, %xmm23
; AVX512DQ-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX512DQ-FCP-NEXT:    vpshufb %xmm12, %xmm2, %xmm13
; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm13[0],xmm10[0],xmm13[1],xmm10[1],xmm13[2],xmm10[2],xmm13[3],xmm10[3]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} xmm9 = xmm10[0,1,2],xmm9[3]
; AVX512DQ-FCP-NEXT:    vpsrlq $32, %zmm24, %zmm10
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm24, %zmm25
; AVX512DQ-FCP-NEXT:    vpmovqb %zmm10, %xmm10
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} xmm9 = xmm10[0,1],xmm9[2,3]
; AVX512DQ-FCP-NEXT:    vinserti32x4 $2, %xmm9, %zmm0, %zmm13
; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm13 {%k1}
; AVX512DQ-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm2 # 32-byte Folded Reload
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm17, %ymm1
; AVX512DQ-FCP-NEXT:    vpshufb %ymm1, %ymm2, %ymm1
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm2, %ymm27
; AVX512DQ-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm3 # 32-byte Folded Reload
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm18, %ymm2
; AVX512DQ-FCP-NEXT:    vpshufb %ymm2, %ymm3, %ymm15
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm3, %ymm26
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm15 = ymm15[0,1,2,3,4,5,6],ymm1[7]
; AVX512DQ-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm2 # 32-byte Folded Reload
; AVX512DQ-FCP-NEXT:    vpshufb %ymm7, %ymm2, %ymm1
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm2, %ymm31
; AVX512DQ-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX512DQ-FCP-NEXT:    vpshufb %ymm8, %ymm0, %ymm14
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm0, %ymm30
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm14[0,1,2,3,4],ymm1[5],ymm14[6,7]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5],ymm15[6,7]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm21, %xmm10
; AVX512DQ-FCP-NEXT:    vpshufb %xmm11, %xmm10, %xmm14
; AVX512DQ-FCP-NEXT:    vpshufb %xmm11, %xmm5, %xmm11
; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm5, %xmm24
; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm11[0],xmm14[0],xmm11[1],xmm14[1],xmm11[2],xmm14[2],xmm11[3],xmm14[3]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm16, %xmm8
; AVX512DQ-FCP-NEXT:    vpshufb %xmm12, %xmm8, %xmm14
; AVX512DQ-FCP-NEXT:    vpshufb %xmm12, %xmm6, %xmm12
; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm6, %xmm21
; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm12[0],xmm14[0],xmm12[1],xmm14[1],xmm12[2],xmm14[2],xmm12[3],xmm14[3]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} xmm11 = xmm12[0,1,2],xmm11[3]
; AVX512DQ-FCP-NEXT:    vpsrlq $32, %zmm29, %zmm12
; AVX512DQ-FCP-NEXT:    vpmovqb %zmm12, %xmm12
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} xmm11 = xmm12[0,1],xmm11[2,3]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm11[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm13, %zmm0
; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm3 = [1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm28, %ymm0
; AVX512DQ-FCP-NEXT:    vpshufb %ymm3, %ymm0, %ymm1
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm5 = [1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13]
; AVX512DQ-FCP-NEXT:    vpshufb %ymm5, %ymm4, %ymm11
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm4, %ymm16
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm11[0,1,2,3,4,5,6],ymm1[7]
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm6 = [1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm19, %ymm0
; AVX512DQ-FCP-NEXT:    vpshufb %ymm6, %ymm0, %ymm11
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm18 = [1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm18, %ymm7
; AVX512DQ-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX512DQ-FCP-NEXT:    vpshufb %ymm7, %ymm2, %ymm12
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm11 = ymm12[0,1,2,3,4],ymm11[5],ymm12[6,7]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm11[0,1,2,3,4,5],ymm1[6,7]
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm11 = [0,0,5,13,0,0,5,13,0,0,5,13,0,0,5,13]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm20, %xmm2
; AVX512DQ-FCP-NEXT:    vpshufb %xmm11, %xmm2, %xmm12
; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm20, %xmm19
; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm22, %xmm4
; AVX512DQ-FCP-NEXT:    vpshufb %xmm11, %xmm4, %xmm13
; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm13[0],xmm12[0],xmm13[1],xmm12[1],xmm13[2],xmm12[2],xmm13[3],xmm12[3]
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm13 = [5,13,0,0,5,13,0,0,5,13,0,0,5,13,0,0]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm23, %xmm9
; AVX512DQ-FCP-NEXT:    vpshufb %xmm13, %xmm9, %xmm14
; AVX512DQ-FCP-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
; AVX512DQ-FCP-NEXT:    vpshufb %xmm13, %xmm7, %xmm15
; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} xmm12 = xmm14[0,1,2],xmm12[3]
; AVX512DQ-FCP-NEXT:    vpsrlq $40, %zmm25, %zmm14
; AVX512DQ-FCP-NEXT:    vpmovqb %zmm14, %xmm14
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} xmm12 = xmm14[0,1],xmm12[2,3]
; AVX512DQ-FCP-NEXT:    vinserti32x4 $2, %xmm12, %zmm0, %zmm12
; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm12 {%k1}
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm27, %ymm1
; AVX512DQ-FCP-NEXT:    vpshufb %ymm3, %ymm1, %ymm1
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm26, %ymm2
; AVX512DQ-FCP-NEXT:    vpshufb %ymm5, %ymm2, %ymm14
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm14[0,1,2,3,4,5,6],ymm1[7]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm31, %ymm3
; AVX512DQ-FCP-NEXT:    vpshufb %ymm6, %ymm3, %ymm14
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm18, %ymm3
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm30, %ymm2
; AVX512DQ-FCP-NEXT:    vpshufb %ymm3, %ymm2, %ymm15
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm14 = ymm15[0,1,2,3,4],ymm14[5],ymm15[6,7]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm14[0,1,2,3,4,5],ymm1[6,7]
; AVX512DQ-FCP-NEXT:    vpshufb %xmm11, %xmm10, %xmm14
; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm10, %xmm20
; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm24, %xmm2
; AVX512DQ-FCP-NEXT:    vpshufb %xmm11, %xmm2, %xmm11
; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm11[0],xmm14[0],xmm11[1],xmm14[1],xmm11[2],xmm14[2],xmm11[3],xmm14[3]
; AVX512DQ-FCP-NEXT:    vpshufb %xmm13, %xmm8, %xmm14
; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm8, %xmm22
; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm21, %xmm2
; AVX512DQ-FCP-NEXT:    vpshufb %xmm13, %xmm2, %xmm13
; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm13[0],xmm14[0],xmm13[1],xmm14[1],xmm13[2],xmm14[2],xmm13[3],xmm14[3]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} xmm11 = xmm13[0,1,2],xmm11[3]
; AVX512DQ-FCP-NEXT:    vpsrlq $40, %zmm29, %zmm13
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm29, %zmm18
; AVX512DQ-FCP-NEXT:    vpmovqb %zmm13, %xmm13
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} xmm11 = xmm13[0,1],xmm11[2,3]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm11[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm12, %zmm29
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm3 = [2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm28, %ymm10
; AVX512DQ-FCP-NEXT:    vpshufb %ymm3, %ymm10, %ymm1
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm21 = [2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm16, %ymm5
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm21, %ymm6
; AVX512DQ-FCP-NEXT:    vpshufb %ymm6, %ymm5, %ymm11
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm11[0,1,2,3,4,5,6],ymm1[7]
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm6 = [2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14]
; AVX512DQ-FCP-NEXT:    vpshufb %ymm6, %ymm0, %ymm11
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm0, %ymm17
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm8 = [2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14]
; AVX512DQ-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX512DQ-FCP-NEXT:    vpshufb %ymm8, %ymm0, %ymm12
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm11 = ymm12[0,1,2,3,4],ymm11[5],ymm12[6,7]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm11[0,1,2,3,4,5],ymm1[6,7]
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm11 = [0,0,6,14,0,0,6,14,0,0,6,14,0,0,6,14]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm19, %xmm0
; AVX512DQ-FCP-NEXT:    vpshufb %xmm11, %xmm0, %xmm12
; AVX512DQ-FCP-NEXT:    vpshufb %xmm11, %xmm4, %xmm13
; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm4, %xmm23
; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm13[0],xmm12[0],xmm13[1],xmm12[1],xmm13[2],xmm12[2],xmm13[3],xmm12[3]
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm13 = [6,14,0,0,6,14,0,0,6,14,0,0,6,14,0,0]
; AVX512DQ-FCP-NEXT:    vpshufb %xmm13, %xmm9, %xmm14
; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm9, %xmm28
; AVX512DQ-FCP-NEXT:    vpshufb %xmm13, %xmm7, %xmm15
; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm7, %xmm16
; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} xmm12 = xmm14[0,1,2],xmm12[3]
; AVX512DQ-FCP-NEXT:    vpsrlq $48, %zmm25, %zmm14
; AVX512DQ-FCP-NEXT:    vpmovqb %zmm14, %xmm14
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} xmm12 = xmm14[0,1],xmm12[2,3]
; AVX512DQ-FCP-NEXT:    vinserti32x4 $2, %xmm12, %zmm0, %zmm12
; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm12 {%k1}
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm27, %ymm0
; AVX512DQ-FCP-NEXT:    vpshufb %ymm3, %ymm0, %ymm1
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm26, %ymm9
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm21, %ymm3
; AVX512DQ-FCP-NEXT:    vpshufb %ymm3, %ymm9, %ymm14
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm14[0,1,2,3,4,5,6],ymm1[7]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm31, %ymm3
; AVX512DQ-FCP-NEXT:    vpshufb %ymm6, %ymm3, %ymm14
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm30, %ymm3
; AVX512DQ-FCP-NEXT:    vpshufb %ymm8, %ymm3, %ymm15
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm14 = ymm15[0,1,2,3,4],ymm14[5],ymm15[6,7]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm14[0,1,2,3,4,5],ymm1[6,7]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm20, %xmm8
; AVX512DQ-FCP-NEXT:    vpshufb %xmm11, %xmm8, %xmm14
; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm24, %xmm15
; AVX512DQ-FCP-NEXT:    vpshufb %xmm11, %xmm15, %xmm11
; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm11[0],xmm14[0],xmm11[1],xmm14[1],xmm11[2],xmm14[2],xmm11[3],xmm14[3]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm22, %xmm3
; AVX512DQ-FCP-NEXT:    vpshufb %xmm13, %xmm3, %xmm14
; AVX512DQ-FCP-NEXT:    vpshufb %xmm13, %xmm2, %xmm13
; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm2, %xmm20
; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm13[0],xmm14[0],xmm13[1],xmm14[1],xmm13[2],xmm14[2],xmm13[3],xmm14[3]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} xmm11 = xmm13[0,1,2],xmm11[3]
; AVX512DQ-FCP-NEXT:    vpsrlq $48, %zmm18, %zmm13
; AVX512DQ-FCP-NEXT:    vpmovqb %zmm13, %xmm13
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} xmm11 = xmm13[0,1],xmm11[2,3]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm11[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm1, %zmm12, %zmm21
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm13 = [3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15]
; AVX512DQ-FCP-NEXT:    vpshufb %ymm13, %ymm10, %ymm6
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm14 = [3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15]
; AVX512DQ-FCP-NEXT:    vpshufb %ymm14, %ymm5, %ymm5
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm5 = ymm5[0,1,2,3,4,5,6],ymm6[7]
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm1 = [3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm17, %ymm2
; AVX512DQ-FCP-NEXT:    vpshufb %ymm1, %ymm2, %ymm3
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm7 = [3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15]
; AVX512DQ-FCP-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX512DQ-FCP-NEXT:    vpshufb %ymm7, %ymm2, %ymm4
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm3 = ymm4[0,1,2,3,4],ymm3[5],ymm4[6,7]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm5[6,7]
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm4 = [0,0,7,15,0,0,7,15,0,0,7,15,0,0,7,15]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm19, %xmm2
; AVX512DQ-FCP-NEXT:    vpshufb %xmm4, %xmm2, %xmm5
; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm23, %xmm2
; AVX512DQ-FCP-NEXT:    vpshufb %xmm4, %xmm2, %xmm6
; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX512DQ-FCP-NEXT:    vpbroadcastd {{.*#+}} xmm6 = [7,15,0,0,7,15,0,0,7,15,0,0,7,15,0,0]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm28, %xmm2
; AVX512DQ-FCP-NEXT:    vpshufb %xmm6, %xmm2, %xmm11
; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm16, %xmm2
; AVX512DQ-FCP-NEXT:    vpshufb %xmm6, %xmm2, %xmm12
; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm12[0],xmm11[0],xmm12[1],xmm11[1],xmm12[2],xmm11[2],xmm12[3],xmm11[3]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} xmm5 = xmm11[0,1,2],xmm5[3]
; AVX512DQ-FCP-NEXT:    vpsrlq $56, %zmm25, %zmm11
; AVX512DQ-FCP-NEXT:    vpmovqb %zmm11, %xmm11
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} xmm5 = xmm11[0,1],xmm5[2,3]
; AVX512DQ-FCP-NEXT:    vinserti32x4 $2, %xmm5, %zmm0, %zmm5
; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm3, %zmm0, %zmm5 {%k1}
; AVX512DQ-FCP-NEXT:    vpshufb %ymm13, %ymm0, %ymm3
; AVX512DQ-FCP-NEXT:    vpshufb %ymm14, %ymm9, %ymm9
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm3 = ymm9[0,1,2,3,4,5,6],ymm3[7]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm31, %ymm0
; AVX512DQ-FCP-NEXT:    vpshufb %ymm1, %ymm0, %ymm2
; AVX512DQ-FCP-NEXT:    vmovdqa64 %ymm30, %ymm0
; AVX512DQ-FCP-NEXT:    vpshufb %ymm7, %ymm0, %ymm0
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm2[5],ymm0[6,7]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm3[6,7]
; AVX512DQ-FCP-NEXT:    vpshufb %xmm4, %xmm8, %xmm2
; AVX512DQ-FCP-NEXT:    vpshufb %xmm4, %xmm15, %xmm3
; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm22, %xmm1
; AVX512DQ-FCP-NEXT:    vpshufb %xmm6, %xmm1, %xmm3
; AVX512DQ-FCP-NEXT:    vmovdqa64 %xmm20, %xmm1
; AVX512DQ-FCP-NEXT:    vpshufb %xmm6, %xmm1, %xmm4
; AVX512DQ-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} xmm2 = xmm3[0,1,2],xmm2[3]
; AVX512DQ-FCP-NEXT:    vpsrlq $56, %zmm18, %zmm3
; AVX512DQ-FCP-NEXT:    vpmovqb %zmm3, %xmm3
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} xmm2 = xmm3[0,1],xmm2[2,3]
; AVX512DQ-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]
; AVX512DQ-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm5, %zmm0
; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovaps %zmm1, (%rsi)
; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovaps %zmm1, (%rdx)
; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovaps %zmm1, (%rcx)
; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovaps %zmm1, (%r8)
; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-FCP-NEXT:    vmovaps %zmm1, (%r9)
; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm29, (%rax)
; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm21, (%rax)
; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, (%rax)
; AVX512DQ-FCP-NEXT:    addq $440, %rsp # imm = 0x1B8
; AVX512DQ-FCP-NEXT:    vzeroupper
; AVX512DQ-FCP-NEXT:    retq
;
; AVX512BW-LABEL: load_i8_stride8_vf64:
; AVX512BW:       # %bb.0:
; AVX512BW-NEXT:    subq $744, %rsp # imm = 0x2E8
; AVX512BW-NEXT:    vmovdqa64 (%rdi), %zmm5
; AVX512BW-NEXT:    vmovdqa64 128(%rdi), %zmm0
; AVX512BW-NEXT:    vpmovqb %zmm0, %xmm0
; AVX512BW-NEXT:    vmovdqa64 256(%rdi), %zmm1
; AVX512BW-NEXT:    vmovdqa64 384(%rdi), %zmm2
; AVX512BW-NEXT:    vpmovqb %zmm2, %xmm2
; AVX512BW-NEXT:    vmovdqa 496(%rdi), %xmm4
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm12 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX512BW-NEXT:    vpshufb %xmm12, %xmm4, %xmm3
; AVX512BW-NEXT:    vmovdqa64 %xmm4, %xmm24
; AVX512BW-NEXT:    vmovdqa 480(%rdi), %xmm6
; AVX512BW-NEXT:    vpshufb %xmm12, %xmm6, %xmm4
; AVX512BW-NEXT:    vmovdqa64 %xmm6, %xmm25
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX512BW-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX512BW-NEXT:    vmovdqa 464(%rdi), %xmm6
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm19 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX512BW-NEXT:    vpshufb %xmm19, %xmm6, %xmm4
; AVX512BW-NEXT:    vmovdqa64 %xmm6, %xmm26
; AVX512BW-NEXT:    vmovdqa 448(%rdi), %xmm7
; AVX512BW-NEXT:    vpshufb %xmm19, %xmm7, %xmm6
; AVX512BW-NEXT:    vmovdqa64 %xmm7, %xmm30
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm6[0],xmm4[0],xmm6[1],xmm4[1],xmm6[2],xmm4[2],xmm6[3],xmm4[3]
; AVX512BW-NEXT:    vinserti128 $1, %xmm4, %ymm0, %ymm4
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm3 = ymm4[0,1,2,3,4,5,6],ymm3[7]
; AVX512BW-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX512BW-NEXT:    vmovdqa 384(%rdi), %ymm4
; AVX512BW-NEXT:    vpmovqb %ymm4, %xmm4
; AVX512BW-NEXT:    vinserti128 $1, %xmm4, %ymm0, %ymm4
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm2 = ymm4[0,1,2,3,4],ymm2[5],ymm4[6,7]
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm7 = ymm2[0,1,2,3,4,5],ymm3[6,7]
; AVX512BW-NEXT:    vmovdqa 368(%rdi), %xmm2
; AVX512BW-NEXT:    vpshufb %xmm12, %xmm2, %xmm4
; AVX512BW-NEXT:    vmovdqa64 %xmm2, %xmm31
; AVX512BW-NEXT:    vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512BW-NEXT:    vmovdqa64 352(%rdi), %xmm27
; AVX512BW-NEXT:    vpshufb %xmm12, %xmm27, %xmm6
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm6[0],xmm4[0],xmm6[1],xmm4[1],xmm6[2],xmm4[2],xmm6[3],xmm4[3]
; AVX512BW-NEXT:    vmovdqa 336(%rdi), %xmm2
; AVX512BW-NEXT:    vpshufb %xmm19, %xmm2, %xmm11
; AVX512BW-NEXT:    vmovdqa64 %xmm2, %xmm22
; AVX512BW-NEXT:    vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512BW-NEXT:    vmovdqa 320(%rdi), %xmm2
; AVX512BW-NEXT:    vpshufb %xmm19, %xmm2, %xmm15
; AVX512BW-NEXT:    vmovdqa %xmm2, %xmm9
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm15[0],xmm11[0],xmm15[1],xmm11[1],xmm15[2],xmm11[2],xmm15[3],xmm11[3]
; AVX512BW-NEXT:    vpblendd {{.*#+}} xmm10 = xmm11[0,1,2],xmm10[3]
; AVX512BW-NEXT:    vpmovqb %zmm1, %xmm11
; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm15
; AVX512BW-NEXT:    vpblendd {{.*#+}} xmm10 = xmm11[0,1],xmm10[2,3]
; AVX512BW-NEXT:    vinserti32x4 $2, %xmm10, %zmm0, %zmm20
; AVX512BW-NEXT:    movb $-64, %al
; AVX512BW-NEXT:    kmovd %eax, %k1
; AVX512BW-NEXT:    vinserti64x4 $1, %ymm7, %zmm0, %zmm20 {%k1}
; AVX512BW-NEXT:    vmovdqa64 240(%rdi), %xmm28
; AVX512BW-NEXT:    vpshufb %xmm12, %xmm28, %xmm7
; AVX512BW-NEXT:    vmovdqa 224(%rdi), %xmm1
; AVX512BW-NEXT:    vpshufb %xmm12, %xmm1, %xmm10
; AVX512BW-NEXT:    vmovdqa64 %xmm1, %xmm18
; AVX512BW-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm7 = xmm10[0],xmm7[0],xmm10[1],xmm7[1],xmm10[2],xmm7[2],xmm10[3],xmm7[3]
; AVX512BW-NEXT:    vinserti128 $1, %xmm7, %ymm0, %ymm7
; AVX512BW-NEXT:    vmovdqa64 208(%rdi), %xmm17
; AVX512BW-NEXT:    vpshufb %xmm19, %xmm17, %xmm10
; AVX512BW-NEXT:    vmovdqa64 %xmm17, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512BW-NEXT:    vmovdqa 192(%rdi), %xmm8
; AVX512BW-NEXT:    vpshufb %xmm19, %xmm8, %xmm16
; AVX512BW-NEXT:    vmovdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm16[0],xmm10[0],xmm16[1],xmm10[1],xmm16[2],xmm10[2],xmm16[3],xmm10[3]
; AVX512BW-NEXT:    vinserti128 $1, %xmm10, %ymm0, %ymm10
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm7 = ymm10[0,1,2,3,4,5,6],ymm7[7]
; AVX512BW-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX512BW-NEXT:    vmovdqa 128(%rdi), %ymm10
; AVX512BW-NEXT:    vpmovqb %ymm10, %xmm10
; AVX512BW-NEXT:    vinserti128 $1, %xmm10, %ymm0, %ymm10
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm10[0,1,2,3,4],ymm0[5],ymm10[6,7]
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm0[0,1,2,3,4,5],ymm7[6,7]
; AVX512BW-NEXT:    vmovdqa 112(%rdi), %xmm0
; AVX512BW-NEXT:    vpshufb %xmm12, %xmm0, %xmm16
; AVX512BW-NEXT:    vmovdqa %xmm0, %xmm10
; AVX512BW-NEXT:    vmovdqa 96(%rdi), %xmm0
; AVX512BW-NEXT:    vpshufb %xmm12, %xmm0, %xmm12
; AVX512BW-NEXT:    vmovdqa %xmm0, %xmm14
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm12[0],xmm16[0],xmm12[1],xmm16[1],xmm12[2],xmm16[2],xmm12[3],xmm16[3]
; AVX512BW-NEXT:    vmovdqa 80(%rdi), %xmm2
; AVX512BW-NEXT:    vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512BW-NEXT:    vpshufb %xmm19, %xmm2, %xmm21
; AVX512BW-NEXT:    vmovdqa 64(%rdi), %xmm2
; AVX512BW-NEXT:    vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512BW-NEXT:    vpshufb %xmm19, %xmm2, %xmm19
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm19[0],xmm21[0],xmm19[1],xmm21[1],xmm19[2],xmm21[2],xmm19[3],xmm21[3]
; AVX512BW-NEXT:    vpblendd {{.*#+}} xmm0 = xmm13[0,1,2],xmm0[3]
; AVX512BW-NEXT:    vpmovqb %zmm5, %xmm13
; AVX512BW-NEXT:    vpblendd {{.*#+}} xmm0 = xmm13[0,1],xmm0[2,3]
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm0, %zmm20, %zmm0
; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vmovdqa 160(%rdi), %xmm6
; AVX512BW-NEXT:    vmovdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512BW-NEXT:    vmovdqa 384(%rdi), %xmm7
; AVX512BW-NEXT:    vmovdqa 400(%rdi), %xmm3
; AVX512BW-NEXT:    vmovdqa64 416(%rdi), %xmm20
; AVX512BW-NEXT:    vmovdqa64 432(%rdi), %xmm29
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm4 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX512BW-NEXT:    vmovdqa64 %xmm24, %xmm19
; AVX512BW-NEXT:    vpshufb %xmm4, %xmm24, %xmm1
; AVX512BW-NEXT:    vmovdqa64 %xmm25, %xmm11
; AVX512BW-NEXT:    vpshufb %xmm4, %xmm25, %xmm13
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm13[0],xmm1[0],xmm13[1],xmm1[1],xmm13[2],xmm1[2],xmm13[3],xmm1[3]
; AVX512BW-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm1
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm13 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX512BW-NEXT:    vmovdqa64 %xmm26, %xmm12
; AVX512BW-NEXT:    vpshufb %xmm13, %xmm26, %xmm24
; AVX512BW-NEXT:    vmovdqa64 %xmm30, %xmm16
; AVX512BW-NEXT:    vmovdqa64 %xmm30, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512BW-NEXT:    vpshufb %xmm13, %xmm30, %xmm25
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm24 = xmm25[0],xmm24[0],xmm25[1],xmm24[1],xmm25[2],xmm24[2],xmm25[3],xmm24[3]
; AVX512BW-NEXT:    vinserti32x4 $1, %xmm24, %ymm0, %ymm2
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5,6],ymm1[7]
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm26 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX512BW-NEXT:    vpshufb %xmm26, %xmm29, %xmm24
; AVX512BW-NEXT:    vpshufb %xmm26, %xmm20, %xmm25
; AVX512BW-NEXT:    vmovdqa64 %xmm20, %xmm23
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm24 = xmm25[0],xmm24[0],xmm25[1],xmm24[1],xmm25[2],xmm24[2],xmm25[3],xmm24[3]
; AVX512BW-NEXT:    vinserti32x4 $1, %xmm24, %ymm0, %ymm2
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm30 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX512BW-NEXT:    vpshufb %xmm30, %xmm3, %xmm24
; AVX512BW-NEXT:    vmovdqa64 %xmm3, %xmm21
; AVX512BW-NEXT:    vpshufb %xmm30, %xmm7, %xmm25
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm24 = xmm25[0],xmm24[0],xmm25[1],xmm24[1],xmm25[2],xmm24[2],xmm25[3],xmm24[3]
; AVX512BW-NEXT:    vinserti32x4 $1, %xmm24, %ymm0, %ymm3
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4],ymm2[5],ymm3[6,7]
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]
; AVX512BW-NEXT:    vpshufb %xmm4, %xmm31, %xmm2
; AVX512BW-NEXT:    vpshufb %xmm4, %xmm27, %xmm3
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
; AVX512BW-NEXT:    vpshufb %xmm13, %xmm22, %xmm3
; AVX512BW-NEXT:    vpshufb %xmm13, %xmm9, %xmm24
; AVX512BW-NEXT:    vmovdqa64 %xmm9, %xmm31
; AVX512BW-NEXT:    vmovdqa %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm24[0],xmm3[0],xmm24[1],xmm3[1],xmm24[2],xmm3[2],xmm24[3],xmm3[3]
; AVX512BW-NEXT:    vpblendd {{.*#+}} xmm2 = xmm3[0,1,2],xmm2[3]
; AVX512BW-NEXT:    vpsrlq $8, %zmm15, %zmm3
; AVX512BW-NEXT:    vmovdqa64 %zmm15, %zmm22
; AVX512BW-NEXT:    vpmovqb %zmm3, %xmm3
; AVX512BW-NEXT:    vpblendd {{.*#+}} xmm2 = xmm3[0,1],xmm2[2,3]
; AVX512BW-NEXT:    vinserti32x4 $2, %xmm2, %zmm0, %zmm2
; AVX512BW-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm2 {%k1}
; AVX512BW-NEXT:    vmovdqa64 %xmm28, %xmm9
; AVX512BW-NEXT:    vpshufb %xmm4, %xmm28, %xmm1
; AVX512BW-NEXT:    vpshufb %xmm4, %xmm18, %xmm3
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3]
; AVX512BW-NEXT:    vpshufb %xmm13, %xmm17, %xmm3
; AVX512BW-NEXT:    vpshufb %xmm13, %xmm8, %xmm24
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm24[0],xmm3[0],xmm24[1],xmm3[1],xmm24[2],xmm3[2],xmm24[3],xmm3[3]
; AVX512BW-NEXT:    vmovdqa 176(%rdi), %xmm0
; AVX512BW-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512BW-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm1
; AVX512BW-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm3[0,1,2,3,4,5,6],ymm1[7]
; AVX512BW-NEXT:    vpshufb %xmm26, %xmm0, %xmm3
; AVX512BW-NEXT:    vpshufb %xmm26, %xmm6, %xmm25
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm25[0],xmm3[0],xmm25[1],xmm3[1],xmm25[2],xmm3[2],xmm25[3],xmm3[3]
; AVX512BW-NEXT:    vmovdqa 128(%rdi), %xmm6
; AVX512BW-NEXT:    vmovdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512BW-NEXT:    vmovdqa 144(%rdi), %xmm0
; AVX512BW-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512BW-NEXT:    vpshufb %xmm30, %xmm0, %xmm0
; AVX512BW-NEXT:    vpshufb %xmm30, %xmm6, %xmm30
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm30[0],xmm0[0],xmm30[1],xmm0[1],xmm30[2],xmm0[2],xmm30[3],xmm0[3]
; AVX512BW-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX512BW-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm3[5],ymm0[6,7]
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
; AVX512BW-NEXT:    vpshufb %xmm4, %xmm10, %xmm1
; AVX512BW-NEXT:    vmovdqa %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512BW-NEXT:    vpshufb %xmm4, %xmm14, %xmm3
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3]
; AVX512BW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX512BW-NEXT:    vpshufb %xmm13, %xmm15, %xmm3
; AVX512BW-NEXT:    vmovdqa64 {{[-0-9]+}}(%r{{[sb]}}p), %xmm28 # 16-byte Reload
; AVX512BW-NEXT:    vpshufb %xmm13, %xmm28, %xmm4
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX512BW-NEXT:    vpblendd {{.*#+}} xmm1 = xmm3[0,1,2],xmm1[3]
; AVX512BW-NEXT:    vpsrlq $8, %zmm5, %zmm3
; AVX512BW-NEXT:    vpmovqb %zmm3, %xmm3
; AVX512BW-NEXT:    vpblendd {{.*#+}} xmm1 = xmm3[0,1],xmm1[2,3]
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm0, %zmm2, %zmm0
; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm1 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX512BW-NEXT:    vpshufb %xmm1, %xmm19, %xmm2
; AVX512BW-NEXT:    vmovdqa64 %xmm19, %xmm20
; AVX512BW-NEXT:    vpshufb %xmm1, %xmm11, %xmm3
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm3 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX512BW-NEXT:    vpshufb %xmm3, %xmm12, %xmm4
; AVX512BW-NEXT:    vmovdqa64 %xmm12, %xmm25
; AVX512BW-NEXT:    vmovdqa %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512BW-NEXT:    vpshufb %xmm3, %xmm16, %xmm13
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm13[0],xmm4[0],xmm13[1],xmm4[1],xmm13[2],xmm4[2],xmm13[3],xmm4[3]
; AVX512BW-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX512BW-NEXT:    vinserti128 $1, %xmm4, %ymm0, %ymm4
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm2 = ymm4[0,1,2,3,4,5,6],ymm2[7]
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm4 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX512BW-NEXT:    vmovdqa64 %xmm29, %xmm24
; AVX512BW-NEXT:    vpshufb %xmm4, %xmm29, %xmm13
; AVX512BW-NEXT:    vpshufb %xmm4, %xmm23, %xmm30
; AVX512BW-NEXT:    vmovdqa64 %xmm23, %xmm16
; AVX512BW-NEXT:    vmovdqa64 %xmm23, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm30[0],xmm13[0],xmm30[1],xmm13[1],xmm30[2],xmm13[2],xmm30[3],xmm13[3]
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm30 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX512BW-NEXT:    vmovdqa64 %xmm21, %xmm18
; AVX512BW-NEXT:    vpshufb %xmm30, %xmm21, %xmm0
; AVX512BW-NEXT:    vmovdqa64 %xmm7, %xmm17
; AVX512BW-NEXT:    vmovdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512BW-NEXT:    vpshufb %xmm30, %xmm7, %xmm19
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm19[0],xmm0[0],xmm19[1],xmm0[1],xmm19[2],xmm0[2],xmm19[3],xmm0[3]
; AVX512BW-NEXT:    vinserti128 $1, %xmm13, %ymm0, %ymm13
; AVX512BW-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm13[5],ymm0[6,7]
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm2[6,7]
; AVX512BW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload
; AVX512BW-NEXT:    vpshufb %xmm1, %xmm12, %xmm2
; AVX512BW-NEXT:    vpshufb %xmm1, %xmm27, %xmm13
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm13[0],xmm2[0],xmm13[1],xmm2[1],xmm13[2],xmm2[2],xmm13[3],xmm2[3]
; AVX512BW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX512BW-NEXT:    vpshufb %xmm3, %xmm8, %xmm13
; AVX512BW-NEXT:    vpshufb %xmm3, %xmm31, %xmm19
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm19[0],xmm13[0],xmm19[1],xmm13[1],xmm19[2],xmm13[2],xmm19[3],xmm13[3]
; AVX512BW-NEXT:    vpblendd {{.*#+}} xmm2 = xmm13[0,1,2],xmm2[3]
; AVX512BW-NEXT:    vmovdqa64 %zmm22, %zmm6
; AVX512BW-NEXT:    vmovdqu64 %zmm22, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpsrlq $16, %zmm22, %zmm13
; AVX512BW-NEXT:    vpmovqb %zmm13, %xmm13
; AVX512BW-NEXT:    vpblendd {{.*#+}} xmm2 = xmm13[0,1],xmm2[2,3]
; AVX512BW-NEXT:    vinserti32x4 $2, %xmm2, %zmm0, %zmm2
; AVX512BW-NEXT:    vinserti64x4 $1, %ymm0, %zmm0, %zmm2 {%k1}
; AVX512BW-NEXT:    vpshufb %xmm1, %xmm9, %xmm0
; AVX512BW-NEXT:    vmovdqa64 %xmm9, %xmm26
; AVX512BW-NEXT:    vmovdqa %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512BW-NEXT:    vmovdqa64 {{[-0-9]+}}(%r{{[sb]}}p), %xmm21 # 16-byte Reload
; AVX512BW-NEXT:    vpshufb %xmm1, %xmm21, %xmm13
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm13[0],xmm0[0],xmm13[1],xmm0[1],xmm13[2],xmm0[2],xmm13[3],xmm0[3]
; AVX512BW-NEXT:    vmovdqa64 {{[-0-9]+}}(%r{{[sb]}}p), %xmm22 # 16-byte Reload
; AVX512BW-NEXT:    vpshufb %xmm3, %xmm22, %xmm13
; AVX512BW-NEXT:    vmovdqa64 {{[-0-9]+}}(%r{{[sb]}}p), %xmm29 # 16-byte Reload
; AVX512BW-NEXT:    vpshufb %xmm3, %xmm29, %xmm19
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm19[0],xmm13[0],xmm19[1],xmm13[1],xmm19[2],xmm13[2],xmm19[3],xmm13[3]
; AVX512BW-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX512BW-NEXT:    vinserti128 $1, %xmm13, %ymm0, %ymm13
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm13[0,1,2,3,4,5,6],ymm0[7]
; AVX512BW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
; AVX512BW-NEXT:    vpshufb %xmm4, %xmm9, %xmm13
; AVX512BW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
; AVX512BW-NEXT:    vpshufb %xmm4, %xmm7, %xmm4
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm4[0],xmm13[0],xmm4[1],xmm13[1],xmm4[2],xmm13[2],xmm4[3],xmm13[3]
; AVX512BW-NEXT:    vmovdqa64 {{[-0-9]+}}(%r{{[sb]}}p), %xmm31 # 16-byte Reload
; AVX512BW-NEXT:    vpshufb %xmm30, %xmm31, %xmm13
; AVX512BW-NEXT:    vmovdqa64 {{[-0-9]+}}(%r{{[sb]}}p), %xmm23 # 16-byte Reload
; AVX512BW-NEXT:    vpshufb %xmm30, %xmm23, %xmm19
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm19[0],xmm13[0],xmm19[1],xmm13[1],xmm19[2],xmm13[2],xmm19[3],xmm13[3]
; AVX512BW-NEXT:    vinserti128 $1, %xmm4, %ymm0, %ymm4
; AVX512BW-NEXT:    vinserti128 $1, %xmm13, %ymm0, %ymm13
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm4 = ymm13[0,1,2,3,4],ymm4[5],ymm13[6,7]
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm4[0,1,2,3,4,5],ymm0[6,7]
; AVX512BW-NEXT:    vpshufb %xmm1, %xmm10, %xmm4
; AVX512BW-NEXT:    vpshufb %xmm1, %xmm14, %xmm1
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm4[0],xmm1[1],xmm4[1],xmm1[2],xmm4[2],xmm1[3],xmm4[3]
; AVX512BW-NEXT:    vpshufb %xmm3, %xmm15, %xmm4
; AVX512BW-NEXT:    vpshufb %xmm3, %xmm28, %xmm3
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
; AVX512BW-NEXT:    vpblendd {{.*#+}} xmm1 = xmm3[0,1,2],xmm1[3]
; AVX512BW-NEXT:    vpsrlq $16, %zmm5, %zmm3
; AVX512BW-NEXT:    vmovdqa64 %zmm5, %zmm28
; AVX512BW-NEXT:    vpmovqb %zmm3, %xmm3
; AVX512BW-NEXT:    vpblendd {{.*#+}} xmm1 = xmm3[0,1],xmm1[2,3]
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm0, %zmm2, %zmm0
; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm0 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX512BW-NEXT:    vmovdqa64 %xmm20, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512BW-NEXT:    vpshufb %xmm0, %xmm20, %xmm1
; AVX512BW-NEXT:    vpshufb %xmm0, %xmm11, %xmm2
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm2 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX512BW-NEXT:    vpshufb %xmm2, %xmm25, %xmm3
; AVX512BW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX512BW-NEXT:    vpshufb %xmm2, %xmm4, %xmm4
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX512BW-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm1
; AVX512BW-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm3[0,1,2,3,4,5,6],ymm1[7]
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm3 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX512BW-NEXT:    vpshufb %xmm3, %xmm24, %xmm4
; AVX512BW-NEXT:    vmovdqa64 %xmm24, (%rsp) # 16-byte Spill
; AVX512BW-NEXT:    vpshufb %xmm3, %xmm16, %xmm19
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm19[0],xmm4[0],xmm19[1],xmm4[1],xmm19[2],xmm4[2],xmm19[3],xmm4[3]
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm19 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX512BW-NEXT:    vpshufb %xmm19, %xmm18, %xmm30
; AVX512BW-NEXT:    vpshufb %xmm19, %xmm17, %xmm13
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm13[0],xmm30[0],xmm13[1],xmm30[1],xmm13[2],xmm30[2],xmm13[3],xmm30[3]
; AVX512BW-NEXT:    vinserti128 $1, %xmm4, %ymm0, %ymm4
; AVX512BW-NEXT:    vinserti128 $1, %xmm13, %ymm0, %ymm13
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm4 = ymm13[0,1,2,3,4],ymm4[5],ymm13[6,7]
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm4[0,1,2,3,4,5],ymm1[6,7]
; AVX512BW-NEXT:    vmovdqa %xmm12, %xmm10
; AVX512BW-NEXT:    vpshufb %xmm0, %xmm12, %xmm4
; AVX512BW-NEXT:    vpshufb %xmm0, %xmm27, %xmm13
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm13[0],xmm4[0],xmm13[1],xmm4[1],xmm13[2],xmm4[2],xmm13[3],xmm4[3]
; AVX512BW-NEXT:    vpshufb %xmm2, %xmm8, %xmm13
; AVX512BW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX512BW-NEXT:    vpshufb %xmm2, %xmm8, %xmm30
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm30[0],xmm13[0],xmm30[1],xmm13[1],xmm30[2],xmm13[2],xmm30[3],xmm13[3]
; AVX512BW-NEXT:    vpblendd {{.*#+}} xmm4 = xmm13[0,1,2],xmm4[3]
; AVX512BW-NEXT:    vpsrlq $24, %zmm6, %zmm13
; AVX512BW-NEXT:    vpmovqb %zmm13, %xmm13
; AVX512BW-NEXT:    vpblendd {{.*#+}} xmm4 = xmm13[0,1],xmm4[2,3]
; AVX512BW-NEXT:    vinserti32x4 $2, %xmm4, %zmm0, %zmm4
; AVX512BW-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm4 {%k1}
; AVX512BW-NEXT:    vpshufb %xmm0, %xmm26, %xmm1
; AVX512BW-NEXT:    vpshufb %xmm0, %xmm21, %xmm13
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm13[0],xmm1[0],xmm13[1],xmm1[1],xmm13[2],xmm1[2],xmm13[3],xmm1[3]
; AVX512BW-NEXT:    vpshufb %xmm2, %xmm22, %xmm13
; AVX512BW-NEXT:    vmovdqa64 %xmm22, %xmm21
; AVX512BW-NEXT:    vpshufb %xmm2, %xmm29, %xmm30
; AVX512BW-NEXT:    vmovdqa64 %xmm29, %xmm22
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm30[0],xmm13[0],xmm30[1],xmm13[1],xmm30[2],xmm13[2],xmm30[3],xmm13[3]
; AVX512BW-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm1
; AVX512BW-NEXT:    vinserti128 $1, %xmm13, %ymm0, %ymm13
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm13[0,1,2,3,4,5,6],ymm1[7]
; AVX512BW-NEXT:    vpshufb %xmm3, %xmm9, %xmm13
; AVX512BW-NEXT:    vmovdqa64 %xmm9, %xmm25
; AVX512BW-NEXT:    vpshufb %xmm3, %xmm7, %xmm3
; AVX512BW-NEXT:    vmovdqa %xmm7, %xmm9
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm13[0],xmm3[1],xmm13[1],xmm3[2],xmm13[2],xmm3[3],xmm13[3]
; AVX512BW-NEXT:    vpshufb %xmm19, %xmm31, %xmm13
; AVX512BW-NEXT:    vpshufb %xmm19, %xmm23, %xmm19
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm19[0],xmm13[0],xmm19[1],xmm13[1],xmm19[2],xmm13[2],xmm19[3],xmm13[3]
; AVX512BW-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX512BW-NEXT:    vinserti128 $1, %xmm13, %ymm0, %ymm13
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm3 = ymm13[0,1,2,3,4],ymm3[5],ymm13[6,7]
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm3[0,1,2,3,4,5],ymm1[6,7]
; AVX512BW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX512BW-NEXT:    vpshufb %xmm0, %xmm5, %xmm3
; AVX512BW-NEXT:    vpshufb %xmm0, %xmm14, %xmm0
; AVX512BW-NEXT:    vmovdqa64 %xmm14, %xmm23
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3]
; AVX512BW-NEXT:    vpshufb %xmm2, %xmm15, %xmm3
; AVX512BW-NEXT:    vmovdqa64 {{[-0-9]+}}(%r{{[sb]}}p), %xmm29 # 16-byte Reload
; AVX512BW-NEXT:    vpshufb %xmm2, %xmm29, %xmm2
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
; AVX512BW-NEXT:    vpblendd {{.*#+}} xmm0 = xmm2[0,1,2],xmm0[3]
; AVX512BW-NEXT:    vmovdqa64 %zmm28, %zmm17
; AVX512BW-NEXT:    vpsrlq $24, %zmm28, %zmm2
; AVX512BW-NEXT:    vpmovqb %zmm2, %xmm2
; AVX512BW-NEXT:    vpblendd {{.*#+}} xmm0 = xmm2[0,1],xmm0[2,3]
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm0, %zmm4, %zmm0
; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm0 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX512BW-NEXT:    vpshufb %xmm0, %xmm20, %xmm1
; AVX512BW-NEXT:    vpshufb %xmm0, %xmm11, %xmm3
; AVX512BW-NEXT:    vmovdqa64 %xmm11, %xmm28
; AVX512BW-NEXT:    vmovdqa %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3]
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm3 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX512BW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload
; AVX512BW-NEXT:    vpshufb %xmm3, %xmm12, %xmm4
; AVX512BW-NEXT:    vmovdqa64 {{[-0-9]+}}(%r{{[sb]}}p), %xmm16 # 16-byte Reload
; AVX512BW-NEXT:    vpshufb %xmm3, %xmm16, %xmm13
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm13[0],xmm4[0],xmm13[1],xmm4[1],xmm13[2],xmm4[2],xmm13[3],xmm4[3]
; AVX512BW-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm1
; AVX512BW-NEXT:    vinserti128 $1, %xmm4, %ymm0, %ymm4
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm4[0,1,2,3,4,5,6],ymm1[7]
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm4 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX512BW-NEXT:    vpshufb %xmm4, %xmm24, %xmm13
; AVX512BW-NEXT:    vmovdqa64 {{[-0-9]+}}(%r{{[sb]}}p), %xmm26 # 16-byte Reload
; AVX512BW-NEXT:    vpshufb %xmm4, %xmm26, %xmm19
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm19[0],xmm13[0],xmm19[1],xmm13[1],xmm19[2],xmm13[2],xmm19[3],xmm13[3]
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm19 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX512BW-NEXT:    vpshufb %xmm19, %xmm18, %xmm30
; AVX512BW-NEXT:    vmovdqa64 %xmm18, %xmm24
; AVX512BW-NEXT:    vmovdqa64 {{[-0-9]+}}(%r{{[sb]}}p), %xmm18 # 16-byte Reload
; AVX512BW-NEXT:    vpshufb %xmm19, %xmm18, %xmm2
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm30[0],xmm2[1],xmm30[1],xmm2[2],xmm30[2],xmm2[3],xmm30[3]
; AVX512BW-NEXT:    vinserti128 $1, %xmm13, %ymm0, %ymm13
; AVX512BW-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3,4],ymm13[5],ymm2[6,7]
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]
; AVX512BW-NEXT:    vmovdqa %xmm10, %xmm6
; AVX512BW-NEXT:    vpshufb %xmm0, %xmm10, %xmm2
; AVX512BW-NEXT:    vmovdqa64 %xmm27, %xmm7
; AVX512BW-NEXT:    vpshufb %xmm0, %xmm27, %xmm13
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm13[0],xmm2[0],xmm13[1],xmm2[1],xmm13[2],xmm2[2],xmm13[3],xmm2[3]
; AVX512BW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; AVX512BW-NEXT:    vpshufb %xmm3, %xmm10, %xmm13
; AVX512BW-NEXT:    vmovdqa64 %xmm8, %xmm27
; AVX512BW-NEXT:    vpshufb %xmm3, %xmm8, %xmm30
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm30[0],xmm13[0],xmm30[1],xmm13[1],xmm30[2],xmm13[2],xmm30[3],xmm13[3]
; AVX512BW-NEXT:    vpblendd {{.*#+}} xmm2 = xmm13[0,1,2],xmm2[3]
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512BW-NEXT:    vpsrlq $32, %zmm10, %zmm13
; AVX512BW-NEXT:    vpmovqb %zmm13, %xmm13
; AVX512BW-NEXT:    vpblendd {{.*#+}} xmm2 = xmm13[0,1],xmm2[2,3]
; AVX512BW-NEXT:    vinserti32x4 $2, %xmm2, %zmm0, %zmm2
; AVX512BW-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm2 {%k1}
; AVX512BW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
; AVX512BW-NEXT:    vpshufb %xmm0, %xmm14, %xmm1
; AVX512BW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
; AVX512BW-NEXT:    vpshufb %xmm0, %xmm11, %xmm13
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm13[0],xmm1[0],xmm13[1],xmm1[1],xmm13[2],xmm1[2],xmm13[3],xmm1[3]
; AVX512BW-NEXT:    vpshufb %xmm3, %xmm21, %xmm13
; AVX512BW-NEXT:    vpshufb %xmm3, %xmm22, %xmm30
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm30[0],xmm13[0],xmm30[1],xmm13[1],xmm30[2],xmm13[2],xmm30[3],xmm13[3]
; AVX512BW-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm1
; AVX512BW-NEXT:    vinserti128 $1, %xmm13, %ymm0, %ymm13
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm13[0,1,2,3,4,5,6],ymm1[7]
; AVX512BW-NEXT:    vpshufb %xmm4, %xmm25, %xmm13
; AVX512BW-NEXT:    vpshufb %xmm4, %xmm9, %xmm4
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm4[0],xmm13[0],xmm4[1],xmm13[1],xmm4[2],xmm13[2],xmm4[3],xmm13[3]
; AVX512BW-NEXT:    vpshufb %xmm19, %xmm31, %xmm13
; AVX512BW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX512BW-NEXT:    vpshufb %xmm19, %xmm8, %xmm19
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm19[0],xmm13[0],xmm19[1],xmm13[1],xmm19[2],xmm13[2],xmm19[3],xmm13[3]
; AVX512BW-NEXT:    vinserti128 $1, %xmm4, %ymm0, %ymm4
; AVX512BW-NEXT:    vinserti128 $1, %xmm13, %ymm0, %ymm13
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm4 = ymm13[0,1,2,3,4],ymm4[5],ymm13[6,7]
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm4[0,1,2,3,4,5],ymm1[6,7]
; AVX512BW-NEXT:    vpshufb %xmm0, %xmm5, %xmm4
; AVX512BW-NEXT:    vpshufb %xmm0, %xmm23, %xmm0
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm4[0],xmm0[1],xmm4[1],xmm0[2],xmm4[2],xmm0[3],xmm4[3]
; AVX512BW-NEXT:    vpshufb %xmm3, %xmm15, %xmm4
; AVX512BW-NEXT:    vpshufb %xmm3, %xmm29, %xmm3
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
; AVX512BW-NEXT:    vpblendd {{.*#+}} xmm0 = xmm3[0,1,2],xmm0[3]
; AVX512BW-NEXT:    vmovdqa64 %zmm17, %zmm20
; AVX512BW-NEXT:    vmovdqu64 %zmm17, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpsrlq $32, %zmm17, %zmm3
; AVX512BW-NEXT:    vpmovqb %zmm3, %xmm3
; AVX512BW-NEXT:    vpblendd {{.*#+}} xmm0 = xmm3[0,1],xmm0[2,3]
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm0, %zmm2, %zmm0
; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm0 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX512BW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX512BW-NEXT:    vpshufb %xmm0, %xmm1, %xmm1
; AVX512BW-NEXT:    vpshufb %xmm0, %xmm28, %xmm2
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm2 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX512BW-NEXT:    vpshufb %xmm2, %xmm12, %xmm3
; AVX512BW-NEXT:    vpshufb %xmm2, %xmm16, %xmm4
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX512BW-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm1
; AVX512BW-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm3[0,1,2,3,4,5,6],ymm1[7]
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm3 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX512BW-NEXT:    vmovdqa (%rsp), %xmm15 # 16-byte Reload
; AVX512BW-NEXT:    vpshufb %xmm3, %xmm15, %xmm4
; AVX512BW-NEXT:    vpshufb %xmm3, %xmm26, %xmm13
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm13[0],xmm4[0],xmm13[1],xmm4[1],xmm13[2],xmm4[2],xmm13[3],xmm4[3]
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm13 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX512BW-NEXT:    vpshufb %xmm13, %xmm24, %xmm19
; AVX512BW-NEXT:    vmovdqa64 %xmm24, %xmm26
; AVX512BW-NEXT:    vmovdqa64 %xmm24, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512BW-NEXT:    vpshufb %xmm13, %xmm18, %xmm30
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm19 = xmm30[0],xmm19[0],xmm30[1],xmm19[1],xmm30[2],xmm19[2],xmm30[3],xmm19[3]
; AVX512BW-NEXT:    vinserti128 $1, %xmm4, %ymm0, %ymm4
; AVX512BW-NEXT:    vinserti32x4 $1, %xmm19, %ymm0, %ymm5
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm4 = ymm5[0,1,2,3,4],ymm4[5],ymm5[6,7]
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm4[0,1,2,3,4,5],ymm1[6,7]
; AVX512BW-NEXT:    vpshufb %xmm0, %xmm6, %xmm4
; AVX512BW-NEXT:    vmovdqa64 %xmm6, %xmm24
; AVX512BW-NEXT:    vpshufb %xmm0, %xmm7, %xmm5
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX512BW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX512BW-NEXT:    vpshufb %xmm2, %xmm5, %xmm5
; AVX512BW-NEXT:    vpshufb %xmm2, %xmm27, %xmm19
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm19[0],xmm5[0],xmm19[1],xmm5[1],xmm19[2],xmm5[2],xmm19[3],xmm5[3]
; AVX512BW-NEXT:    vpblendd {{.*#+}} xmm4 = xmm5[0,1,2],xmm4[3]
; AVX512BW-NEXT:    vpsrlq $40, %zmm10, %zmm5
; AVX512BW-NEXT:    vpmovqb %zmm5, %xmm5
; AVX512BW-NEXT:    vpblendd {{.*#+}} xmm4 = xmm5[0,1],xmm4[2,3]
; AVX512BW-NEXT:    vinserti32x4 $2, %xmm4, %zmm0, %zmm4
; AVX512BW-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm4 {%k1}
; AVX512BW-NEXT:    vpshufb %xmm0, %xmm14, %xmm1
; AVX512BW-NEXT:    vpshufb %xmm0, %xmm11, %xmm5
; AVX512BW-NEXT:    vmovdqa64 %xmm11, %xmm17
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm5[0],xmm1[0],xmm5[1],xmm1[1],xmm5[2],xmm1[2],xmm5[3],xmm1[3]
; AVX512BW-NEXT:    vpshufb %xmm2, %xmm21, %xmm5
; AVX512BW-NEXT:    vpshufb %xmm2, %xmm22, %xmm19
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm19[0],xmm5[0],xmm19[1],xmm5[1],xmm19[2],xmm5[2],xmm19[3],xmm5[3]
; AVX512BW-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm1
; AVX512BW-NEXT:    vinserti128 $1, %xmm5, %ymm0, %ymm5
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm5[0,1,2,3,4,5,6],ymm1[7]
; AVX512BW-NEXT:    vpshufb %xmm3, %xmm25, %xmm5
; AVX512BW-NEXT:    vpshufb %xmm3, %xmm9, %xmm3
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm5[0],xmm3[1],xmm5[1],xmm3[2],xmm5[2],xmm3[3],xmm5[3]
; AVX512BW-NEXT:    vmovdqa64 {{[-0-9]+}}(%r{{[sb]}}p), %xmm31 # 16-byte Reload
; AVX512BW-NEXT:    vpshufb %xmm13, %xmm31, %xmm5
; AVX512BW-NEXT:    vpshufb %xmm13, %xmm8, %xmm13
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm13[0],xmm5[0],xmm13[1],xmm5[1],xmm13[2],xmm5[2],xmm13[3],xmm5[3]
; AVX512BW-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX512BW-NEXT:    vinserti128 $1, %xmm5, %ymm0, %ymm5
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm3 = ymm5[0,1,2,3,4],ymm3[5],ymm5[6,7]
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm3[0,1,2,3,4,5],ymm1[6,7]
; AVX512BW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
; AVX512BW-NEXT:    vpshufb %xmm0, %xmm11, %xmm3
; AVX512BW-NEXT:    vpshufb %xmm0, %xmm23, %xmm0
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3]
; AVX512BW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX512BW-NEXT:    vpshufb %xmm2, %xmm6, %xmm3
; AVX512BW-NEXT:    vpshufb %xmm2, %xmm29, %xmm2
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
; AVX512BW-NEXT:    vpblendd {{.*#+}} xmm0 = xmm2[0,1,2],xmm0[3]
; AVX512BW-NEXT:    vpsrlq $40, %zmm20, %zmm2
; AVX512BW-NEXT:    vpmovqb %zmm2, %xmm2
; AVX512BW-NEXT:    vpblendd {{.*#+}} xmm0 = xmm2[0,1],xmm0[2,3]
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm0, %zmm4, %zmm0
; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm0 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX512BW-NEXT:    vmovdqa64 {{[-0-9]+}}(%r{{[sb]}}p), %xmm28 # 16-byte Reload
; AVX512BW-NEXT:    vpshufb %xmm0, %xmm28, %xmm1
; AVX512BW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX512BW-NEXT:    vpshufb %xmm0, %xmm8, %xmm2
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm2 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX512BW-NEXT:    vpshufb %xmm2, %xmm12, %xmm4
; AVX512BW-NEXT:    vpshufb %xmm2, %xmm16, %xmm5
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX512BW-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm1
; AVX512BW-NEXT:    vinserti128 $1, %xmm4, %ymm0, %ymm4
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm4[0,1,2,3,4,5,6],ymm1[7]
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm4 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX512BW-NEXT:    vpshufb %xmm4, %xmm15, %xmm5
; AVX512BW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX512BW-NEXT:    vpshufb %xmm4, %xmm15, %xmm13
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm13[0],xmm5[0],xmm13[1],xmm5[1],xmm13[2],xmm5[2],xmm13[3],xmm5[3]
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm13 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX512BW-NEXT:    vpshufb %xmm13, %xmm26, %xmm19
; AVX512BW-NEXT:    vpshufb %xmm13, %xmm18, %xmm30
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm19 = xmm30[0],xmm19[0],xmm30[1],xmm19[1],xmm30[2],xmm19[2],xmm30[3],xmm19[3]
; AVX512BW-NEXT:    vinserti128 $1, %xmm5, %ymm0, %ymm5
; AVX512BW-NEXT:    vinserti32x4 $1, %xmm19, %ymm0, %ymm3
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm3 = ymm3[0,1,2,3,4],ymm5[5],ymm3[6,7]
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm3[0,1,2,3,4,5],ymm1[6,7]
; AVX512BW-NEXT:    vpshufb %xmm0, %xmm24, %xmm3
; AVX512BW-NEXT:    vpshufb %xmm0, %xmm7, %xmm5
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm5[0],xmm3[0],xmm5[1],xmm3[1],xmm5[2],xmm3[2],xmm5[3],xmm3[3]
; AVX512BW-NEXT:    vmovdqa64 {{[-0-9]+}}(%r{{[sb]}}p), %xmm27 # 16-byte Reload
; AVX512BW-NEXT:    vpshufb %xmm2, %xmm27, %xmm5
; AVX512BW-NEXT:    vmovdqa64 {{[-0-9]+}}(%r{{[sb]}}p), %xmm30 # 16-byte Reload
; AVX512BW-NEXT:    vpshufb %xmm2, %xmm30, %xmm19
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm19[0],xmm5[0],xmm19[1],xmm5[1],xmm19[2],xmm5[2],xmm19[3],xmm5[3]
; AVX512BW-NEXT:    vpblendd {{.*#+}} xmm3 = xmm5[0,1,2],xmm3[3]
; AVX512BW-NEXT:    vpsrlq $48, %zmm10, %zmm5
; AVX512BW-NEXT:    vpmovqb %zmm5, %xmm5
; AVX512BW-NEXT:    vpblendd {{.*#+}} xmm3 = xmm5[0,1],xmm3[2,3]
; AVX512BW-NEXT:    vinserti32x4 $2, %xmm3, %zmm0, %zmm3
; AVX512BW-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm3 {%k1}
; AVX512BW-NEXT:    vpshufb %xmm0, %xmm14, %xmm1
; AVX512BW-NEXT:    vpshufb %xmm0, %xmm17, %xmm5
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm5[0],xmm1[0],xmm5[1],xmm1[1],xmm5[2],xmm1[2],xmm5[3],xmm1[3]
; AVX512BW-NEXT:    vpshufb %xmm2, %xmm21, %xmm5
; AVX512BW-NEXT:    vpshufb %xmm2, %xmm22, %xmm19
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm19[0],xmm5[0],xmm19[1],xmm5[1],xmm19[2],xmm5[2],xmm19[3],xmm5[3]
; AVX512BW-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm1
; AVX512BW-NEXT:    vinserti128 $1, %xmm5, %ymm0, %ymm5
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm5[0,1,2,3,4,5,6],ymm1[7]
; AVX512BW-NEXT:    vpshufb %xmm4, %xmm25, %xmm5
; AVX512BW-NEXT:    vpshufb %xmm4, %xmm9, %xmm4
; AVX512BW-NEXT:    vmovdqa64 %xmm9, %xmm18
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3]
; AVX512BW-NEXT:    vpshufb %xmm13, %xmm31, %xmm5
; AVX512BW-NEXT:    vmovdqa64 {{[-0-9]+}}(%r{{[sb]}}p), %xmm20 # 16-byte Reload
; AVX512BW-NEXT:    vpshufb %xmm13, %xmm20, %xmm13
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm13[0],xmm5[0],xmm13[1],xmm5[1],xmm13[2],xmm5[2],xmm13[3],xmm5[3]
; AVX512BW-NEXT:    vinserti128 $1, %xmm4, %ymm0, %ymm4
; AVX512BW-NEXT:    vinserti128 $1, %xmm5, %ymm0, %ymm5
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm4 = ymm5[0,1,2,3,4],ymm4[5],ymm5[6,7]
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm4[0,1,2,3,4,5],ymm1[6,7]
; AVX512BW-NEXT:    vpshufb %xmm0, %xmm11, %xmm4
; AVX512BW-NEXT:    vpshufb %xmm0, %xmm23, %xmm0
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm4[0],xmm0[1],xmm4[1],xmm0[2],xmm4[2],xmm0[3],xmm4[3]
; AVX512BW-NEXT:    vpshufb %xmm2, %xmm6, %xmm4
; AVX512BW-NEXT:    vmovdqa64 %xmm6, %xmm26
; AVX512BW-NEXT:    vpshufb %xmm2, %xmm29, %xmm2
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3]
; AVX512BW-NEXT:    vpblendd {{.*#+}} xmm0 = xmm2[0,1,2],xmm0[3]
; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload
; AVX512BW-NEXT:    vpsrlq $48, %zmm19, %zmm2
; AVX512BW-NEXT:    vpmovqb %zmm2, %xmm2
; AVX512BW-NEXT:    vpblendd {{.*#+}} xmm0 = xmm2[0,1],xmm0[2,3]
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm0, %zmm3, %zmm0
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm1 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX512BW-NEXT:    vpshufb %xmm1, %xmm28, %xmm2
; AVX512BW-NEXT:    vpshufb %xmm1, %xmm8, %xmm3
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm3 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX512BW-NEXT:    vpshufb %xmm3, %xmm12, %xmm4
; AVX512BW-NEXT:    vpshufb %xmm3, %xmm16, %xmm5
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX512BW-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX512BW-NEXT:    vinserti128 $1, %xmm4, %ymm0, %ymm4
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm2 = ymm4[0,1,2,3,4,5,6],ymm2[7]
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm4 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX512BW-NEXT:    vmovdqa (%rsp), %xmm5 # 16-byte Reload
; AVX512BW-NEXT:    vpshufb %xmm4, %xmm5, %xmm5
; AVX512BW-NEXT:    vpshufb %xmm4, %xmm15, %xmm8
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm8[0],xmm5[0],xmm8[1],xmm5[1],xmm8[2],xmm5[2],xmm8[3],xmm5[3]
; AVX512BW-NEXT:    vpbroadcastw {{.*#+}} xmm8 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX512BW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX512BW-NEXT:    vpshufb %xmm8, %xmm6, %xmm9
; AVX512BW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX512BW-NEXT:    vpshufb %xmm8, %xmm6, %xmm13
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm13[0],xmm9[0],xmm13[1],xmm9[1],xmm13[2],xmm9[2],xmm13[3],xmm9[3]
; AVX512BW-NEXT:    vinserti128 $1, %xmm5, %ymm0, %ymm5
; AVX512BW-NEXT:    vinserti128 $1, %xmm9, %ymm0, %ymm9
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm5 = ymm9[0,1,2,3,4],ymm5[5],ymm9[6,7]
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm2 = ymm5[0,1,2,3,4,5],ymm2[6,7]
; AVX512BW-NEXT:    vpshufb %xmm1, %xmm24, %xmm5
; AVX512BW-NEXT:    vpshufb %xmm1, %xmm7, %xmm9
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm9[0],xmm5[0],xmm9[1],xmm5[1],xmm9[2],xmm5[2],xmm9[3],xmm5[3]
; AVX512BW-NEXT:    vpshufb %xmm3, %xmm27, %xmm9
; AVX512BW-NEXT:    vpshufb %xmm3, %xmm30, %xmm6
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm6[0],xmm9[0],xmm6[1],xmm9[1],xmm6[2],xmm9[2],xmm6[3],xmm9[3]
; AVX512BW-NEXT:    vpblendd {{.*#+}} xmm5 = xmm6[0,1,2],xmm5[3]
; AVX512BW-NEXT:    vpsrlq $56, %zmm10, %zmm6
; AVX512BW-NEXT:    vpmovqb %zmm6, %xmm6
; AVX512BW-NEXT:    vpblendd {{.*#+}} xmm5 = xmm6[0,1],xmm5[2,3]
; AVX512BW-NEXT:    vinserti32x4 $2, %xmm5, %zmm0, %zmm5
; AVX512BW-NEXT:    vinserti64x4 $1, %ymm2, %zmm0, %zmm5 {%k1}
; AVX512BW-NEXT:    vpshufb %xmm1, %xmm14, %xmm2
; AVX512BW-NEXT:    vpshufb %xmm1, %xmm17, %xmm6
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm6[0],xmm2[0],xmm6[1],xmm2[1],xmm6[2],xmm2[2],xmm6[3],xmm2[3]
; AVX512BW-NEXT:    vpshufb %xmm3, %xmm21, %xmm6
; AVX512BW-NEXT:    vpshufb %xmm3, %xmm22, %xmm9
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm9[0],xmm6[0],xmm9[1],xmm6[1],xmm9[2],xmm6[2],xmm9[3],xmm6[3]
; AVX512BW-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX512BW-NEXT:    vinserti128 $1, %xmm6, %ymm0, %ymm6
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm2 = ymm6[0,1,2,3,4,5,6],ymm2[7]
; AVX512BW-NEXT:    vpshufb %xmm4, %xmm25, %xmm6
; AVX512BW-NEXT:    vpshufb %xmm4, %xmm18, %xmm4
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm4[0],xmm6[0],xmm4[1],xmm6[1],xmm4[2],xmm6[2],xmm4[3],xmm6[3]
; AVX512BW-NEXT:    vpshufb %xmm8, %xmm31, %xmm6
; AVX512BW-NEXT:    vpshufb %xmm8, %xmm20, %xmm8
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm8[0],xmm6[0],xmm8[1],xmm6[1],xmm8[2],xmm6[2],xmm8[3],xmm6[3]
; AVX512BW-NEXT:    vinserti128 $1, %xmm4, %ymm0, %ymm4
; AVX512BW-NEXT:    vinserti128 $1, %xmm6, %ymm0, %ymm6
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm4 = ymm6[0,1,2,3,4],ymm4[5],ymm6[6,7]
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm2 = ymm4[0,1,2,3,4,5],ymm2[6,7]
; AVX512BW-NEXT:    vpshufb %xmm1, %xmm11, %xmm4
; AVX512BW-NEXT:    vpshufb %xmm1, %xmm23, %xmm1
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm4[0],xmm1[1],xmm4[1],xmm1[2],xmm4[2],xmm1[3],xmm4[3]
; AVX512BW-NEXT:    vpshufb %xmm3, %xmm26, %xmm4
; AVX512BW-NEXT:    vpshufb %xmm3, %xmm29, %xmm3
; AVX512BW-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
; AVX512BW-NEXT:    vpblendd {{.*#+}} xmm1 = xmm3[0,1,2],xmm1[3]
; AVX512BW-NEXT:    vpsrlq $56, %zmm19, %zmm3
; AVX512BW-NEXT:    vpmovqb %zmm3, %xmm3
; AVX512BW-NEXT:    vpblendd {{.*#+}} xmm1 = xmm3[0,1],xmm1[2,3]
; AVX512BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
; AVX512BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm5, %zmm1
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm2, (%rsi)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm2, (%rdx)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm2, (%rcx)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm2, (%r8)
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm2, (%r9)
; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512BW-NEXT:    vmovaps %zmm2, (%rax)
; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-NEXT:    vmovdqa64 %zmm0, (%rax)
; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-NEXT:    vmovdqa64 %zmm1, (%rax)
; AVX512BW-NEXT:    addq $744, %rsp # imm = 0x2E8
; AVX512BW-NEXT:    vzeroupper
; AVX512BW-NEXT:    retq
;
; AVX512BW-FCP-LABEL: load_i8_stride8_vf64:
; AVX512BW-FCP:       # %bb.0:
; AVX512BW-FCP-NEXT:    subq $328, %rsp # imm = 0x148
; AVX512BW-FCP-NEXT:    vmovdqa64 256(%rdi), %zmm18
; AVX512BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm2 = [0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12]
; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm0 = [0,2,2,3,0,2,4,6]
; AVX512BW-FCP-NEXT:    vmovdqa 480(%rdi), %ymm1
; AVX512BW-FCP-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512BW-FCP-NEXT:    vpermd %ymm1, %ymm0, %ymm30
; AVX512BW-FCP-NEXT:    vpshufb %ymm2, %ymm30, %ymm1
; AVX512BW-FCP-NEXT:    vmovdqa %ymm2, %ymm11
; AVX512BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm3 = [0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12]
; AVX512BW-FCP-NEXT:    vmovdqa 448(%rdi), %ymm2
; AVX512BW-FCP-NEXT:    vmovdqu %ymm2, (%rsp) # 32-byte Spill
; AVX512BW-FCP-NEXT:    vpermd %ymm2, %ymm0, %ymm31
; AVX512BW-FCP-NEXT:    vpshufb %ymm3, %ymm31, %ymm2
; AVX512BW-FCP-NEXT:    vmovdqa %ymm3, %ymm8
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5,6],ymm1[7]
; AVX512BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm3 = [0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12]
; AVX512BW-FCP-NEXT:    vmovdqa 416(%rdi), %ymm2
; AVX512BW-FCP-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512BW-FCP-NEXT:    vpermd %ymm2, %ymm0, %ymm19
; AVX512BW-FCP-NEXT:    vpshufb %ymm3, %ymm19, %ymm2
; AVX512BW-FCP-NEXT:    vmovdqa %ymm3, %ymm9
; AVX512BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm5 = [0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12]
; AVX512BW-FCP-NEXT:    vmovdqa64 384(%rdi), %ymm29
; AVX512BW-FCP-NEXT:    vpermd %ymm29, %ymm0, %ymm14
; AVX512BW-FCP-NEXT:    vpshufb %ymm5, %ymm14, %ymm3
; AVX512BW-FCP-NEXT:    vmovdqa %ymm5, %ymm10
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4],ymm2[5],ymm3[6,7]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]
; AVX512BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm7 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX512BW-FCP-NEXT:    vmovdqa64 368(%rdi), %xmm21
; AVX512BW-FCP-NEXT:    vpshufb %xmm7, %xmm21, %xmm2
; AVX512BW-FCP-NEXT:    vmovdqa 352(%rdi), %xmm4
; AVX512BW-FCP-NEXT:    vpshufb %xmm7, %xmm4, %xmm3
; AVX512BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
; AVX512BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm23 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX512BW-FCP-NEXT:    vmovdqa 336(%rdi), %xmm12
; AVX512BW-FCP-NEXT:    vpshufb %xmm23, %xmm12, %xmm5
; AVX512BW-FCP-NEXT:    vmovdqa64 320(%rdi), %xmm28
; AVX512BW-FCP-NEXT:    vpshufb %xmm23, %xmm28, %xmm6
; AVX512BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} xmm2 = xmm5[0,1,2],xmm2[3]
; AVX512BW-FCP-NEXT:    vpmovqb %zmm18, %xmm5
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} xmm2 = xmm5[0,1],xmm2[2,3]
; AVX512BW-FCP-NEXT:    vinserti32x4 $2, %xmm2, %zmm0, %zmm20
; AVX512BW-FCP-NEXT:    movb $-64, %al
; AVX512BW-FCP-NEXT:    kmovd %eax, %k1
; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm20 {%k1}
; AVX512BW-FCP-NEXT:    vmovdqa 224(%rdi), %ymm1
; AVX512BW-FCP-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512BW-FCP-NEXT:    vpermd %ymm1, %ymm0, %ymm17
; AVX512BW-FCP-NEXT:    vpshufb %ymm11, %ymm17, %ymm1
; AVX512BW-FCP-NEXT:    vmovdqa 192(%rdi), %ymm2
; AVX512BW-FCP-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512BW-FCP-NEXT:    vpermd %ymm2, %ymm0, %ymm3
; AVX512BW-FCP-NEXT:    vpshufb %ymm8, %ymm3, %ymm5
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm5 = ymm5[0,1,2,3,4,5,6],ymm1[7]
; AVX512BW-FCP-NEXT:    vmovdqa 160(%rdi), %ymm1
; AVX512BW-FCP-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512BW-FCP-NEXT:    vpermd %ymm1, %ymm0, %ymm2
; AVX512BW-FCP-NEXT:    vpshufb %ymm9, %ymm2, %ymm6
; AVX512BW-FCP-NEXT:    vmovdqa64 128(%rdi), %ymm27
; AVX512BW-FCP-NEXT:    vpermd %ymm27, %ymm0, %ymm16
; AVX512BW-FCP-NEXT:    vpshufb %ymm10, %ymm16, %ymm8
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm6 = ymm8[0,1,2,3,4],ymm6[5],ymm8[6,7]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm6[0,1,2,3,4,5],ymm5[6,7]
; AVX512BW-FCP-NEXT:    vmovdqa64 112(%rdi), %xmm26
; AVX512BW-FCP-NEXT:    vpshufb %xmm7, %xmm26, %xmm8
; AVX512BW-FCP-NEXT:    vmovdqa64 96(%rdi), %xmm24
; AVX512BW-FCP-NEXT:    vpshufb %xmm7, %xmm24, %xmm7
; AVX512BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm7[0],xmm8[0],xmm7[1],xmm8[1],xmm7[2],xmm8[2],xmm7[3],xmm8[3]
; AVX512BW-FCP-NEXT:    vmovdqa64 80(%rdi), %xmm22
; AVX512BW-FCP-NEXT:    vpshufb %xmm23, %xmm22, %xmm25
; AVX512BW-FCP-NEXT:    vmovdqa 64(%rdi), %xmm8
; AVX512BW-FCP-NEXT:    vpshufb %xmm23, %xmm8, %xmm23
; AVX512BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm15 = xmm23[0],xmm25[0],xmm23[1],xmm25[1],xmm23[2],xmm25[2],xmm23[3],xmm25[3]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} xmm15 = xmm15[0,1,2],xmm10[3]
; AVX512BW-FCP-NEXT:    vmovdqa64 (%rdi), %zmm10
; AVX512BW-FCP-NEXT:    vpmovqb %zmm10, %xmm13
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} xmm13 = xmm13[0,1],xmm15[2,3]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm13[0,1,2,3],ymm0[4,5,6,7]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm20, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm9 = [1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13]
; AVX512BW-FCP-NEXT:    vpshufb %ymm9, %ymm30, %ymm0
; AVX512BW-FCP-NEXT:    vmovdqa %ymm9, %ymm11
; AVX512BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm9 = [1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13]
; AVX512BW-FCP-NEXT:    vpshufb %ymm9, %ymm31, %ymm13
; AVX512BW-FCP-NEXT:    vmovdqa %ymm9, %ymm6
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm13[0,1,2,3,4,5,6],ymm0[7]
; AVX512BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm9 = [1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13]
; AVX512BW-FCP-NEXT:    vpshufb %ymm9, %ymm19, %ymm13
; AVX512BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm1 = [1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13]
; AVX512BW-FCP-NEXT:    vpshufb %ymm1, %ymm14, %ymm15
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm13 = ymm15[0,1,2,3,4],ymm13[5],ymm15[6,7]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm13[0,1,2,3,4,5],ymm0[6,7]
; AVX512BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm13 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX512BW-FCP-NEXT:    vmovdqa64 %xmm21, %xmm5
; AVX512BW-FCP-NEXT:    vpshufb %xmm13, %xmm21, %xmm15
; AVX512BW-FCP-NEXT:    vpshufb %xmm13, %xmm4, %xmm23
; AVX512BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm15 = xmm23[0],xmm15[0],xmm23[1],xmm15[1],xmm23[2],xmm15[2],xmm23[3],xmm15[3]
; AVX512BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm23 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX512BW-FCP-NEXT:    vmovdqa %xmm12, %xmm7
; AVX512BW-FCP-NEXT:    vpshufb %xmm23, %xmm12, %xmm25
; AVX512BW-FCP-NEXT:    vpshufb %xmm23, %xmm28, %xmm20
; AVX512BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm20[0],xmm25[0],xmm20[1],xmm25[1],xmm20[2],xmm25[2],xmm20[3],xmm25[3]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} xmm12 = xmm12[0,1,2],xmm15[3]
; AVX512BW-FCP-NEXT:    vpsrlq $8, %zmm18, %zmm15
; AVX512BW-FCP-NEXT:    vpmovqb %zmm15, %xmm15
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} xmm12 = xmm15[0,1],xmm12[2,3]
; AVX512BW-FCP-NEXT:    vinserti32x4 $2, %xmm12, %zmm0, %zmm12
; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm0, %zmm0, %zmm12 {%k1}
; AVX512BW-FCP-NEXT:    vpshufb %ymm11, %ymm17, %ymm0
; AVX512BW-FCP-NEXT:    vpshufb %ymm6, %ymm3, %ymm15
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm15[0,1,2,3,4,5,6],ymm0[7]
; AVX512BW-FCP-NEXT:    vpshufb %ymm9, %ymm2, %ymm15
; AVX512BW-FCP-NEXT:    vpshufb %ymm1, %ymm16, %ymm11
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm11 = ymm11[0,1,2,3,4],ymm15[5],ymm11[6,7]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm11[0,1,2,3,4,5],ymm0[6,7]
; AVX512BW-FCP-NEXT:    vpshufb %xmm13, %xmm26, %xmm11
; AVX512BW-FCP-NEXT:    vpshufb %xmm13, %xmm24, %xmm13
; AVX512BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm13[0],xmm11[0],xmm13[1],xmm11[1],xmm13[2],xmm11[2],xmm13[3],xmm11[3]
; AVX512BW-FCP-NEXT:    vpshufb %xmm23, %xmm22, %xmm13
; AVX512BW-FCP-NEXT:    vpshufb %xmm23, %xmm8, %xmm15
; AVX512BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm15[0],xmm13[0],xmm15[1],xmm13[1],xmm15[2],xmm13[2],xmm15[3],xmm13[3]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} xmm11 = xmm13[0,1,2],xmm11[3]
; AVX512BW-FCP-NEXT:    vpsrlq $8, %zmm10, %zmm13
; AVX512BW-FCP-NEXT:    vpmovqb %zmm13, %xmm13
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} xmm11 = xmm13[0,1],xmm11[2,3]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm11[0,1,2,3],ymm0[4,5,6,7]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm12, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm9 = [2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14]
; AVX512BW-FCP-NEXT:    vpshufb %ymm9, %ymm30, %ymm0
; AVX512BW-FCP-NEXT:    vmovdqa64 %ymm9, %ymm21
; AVX512BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm6 = [2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14]
; AVX512BW-FCP-NEXT:    vpshufb %ymm6, %ymm31, %ymm11
; AVX512BW-FCP-NEXT:    vmovdqa64 %ymm6, %ymm25
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm11[0,1,2,3,4,5,6],ymm0[7]
; AVX512BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm6 = [2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14]
; AVX512BW-FCP-NEXT:    vpshufb %ymm6, %ymm19, %ymm11
; AVX512BW-FCP-NEXT:    vmovdqa %ymm6, %ymm9
; AVX512BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm6 = [2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14]
; AVX512BW-FCP-NEXT:    vpshufb %ymm6, %ymm14, %ymm12
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm11 = ymm12[0,1,2,3,4],ymm11[5],ymm12[6,7]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm11[0,1,2,3,4,5],ymm0[6,7]
; AVX512BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm11 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX512BW-FCP-NEXT:    vpshufb %xmm11, %xmm5, %xmm12
; AVX512BW-FCP-NEXT:    vpshufb %xmm11, %xmm4, %xmm13
; AVX512BW-FCP-NEXT:    vmovdqa %xmm4, %xmm1
; AVX512BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm13[0],xmm12[0],xmm13[1],xmm12[1],xmm13[2],xmm12[2],xmm13[3],xmm12[3]
; AVX512BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm13 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX512BW-FCP-NEXT:    vpshufb %xmm13, %xmm7, %xmm15
; AVX512BW-FCP-NEXT:    vmovdqa64 %xmm7, %xmm23
; AVX512BW-FCP-NEXT:    vpshufb %xmm13, %xmm28, %xmm20
; AVX512BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm15 = xmm20[0],xmm15[0],xmm20[1],xmm15[1],xmm20[2],xmm15[2],xmm20[3],xmm15[3]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} xmm12 = xmm15[0,1,2],xmm12[3]
; AVX512BW-FCP-NEXT:    vpsrlq $16, %zmm18, %zmm15
; AVX512BW-FCP-NEXT:    vpmovqb %zmm15, %xmm15
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} xmm12 = xmm15[0,1],xmm12[2,3]
; AVX512BW-FCP-NEXT:    vinserti32x4 $2, %xmm12, %zmm0, %zmm12
; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm0, %zmm0, %zmm12 {%k1}
; AVX512BW-FCP-NEXT:    vpshufb %ymm21, %ymm17, %ymm0
; AVX512BW-FCP-NEXT:    vpshufb %ymm25, %ymm3, %ymm15
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm15[0,1,2,3,4,5,6],ymm0[7]
; AVX512BW-FCP-NEXT:    vpshufb %ymm9, %ymm2, %ymm15
; AVX512BW-FCP-NEXT:    vpshufb %ymm6, %ymm16, %ymm9
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm9 = ymm9[0,1,2,3,4],ymm15[5],ymm9[6,7]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm9[0,1,2,3,4,5],ymm0[6,7]
; AVX512BW-FCP-NEXT:    vpshufb %xmm11, %xmm26, %xmm9
; AVX512BW-FCP-NEXT:    vpshufb %xmm11, %xmm24, %xmm11
; AVX512BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm11[0],xmm9[0],xmm11[1],xmm9[1],xmm11[2],xmm9[2],xmm11[3],xmm9[3]
; AVX512BW-FCP-NEXT:    vpshufb %xmm13, %xmm22, %xmm11
; AVX512BW-FCP-NEXT:    vpshufb %xmm13, %xmm8, %xmm13
; AVX512BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm13[0],xmm11[0],xmm13[1],xmm11[1],xmm13[2],xmm11[2],xmm13[3],xmm11[3]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} xmm9 = xmm11[0,1,2],xmm9[3]
; AVX512BW-FCP-NEXT:    vpsrlq $16, %zmm10, %zmm11
; AVX512BW-FCP-NEXT:    vpmovqb %zmm11, %xmm11
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} xmm9 = xmm11[0,1],xmm9[2,3]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm9[0,1,2,3],ymm0[4,5,6,7]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm12, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm4 = [3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15]
; AVX512BW-FCP-NEXT:    vpshufb %ymm4, %ymm30, %ymm0
; AVX512BW-FCP-NEXT:    vmovdqa %ymm4, %ymm6
; AVX512BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm4 = [3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15]
; AVX512BW-FCP-NEXT:    vpshufb %ymm4, %ymm31, %ymm9
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm9[0,1,2,3,4,5,6],ymm0[7]
; AVX512BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm31 = [3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15]
; AVX512BW-FCP-NEXT:    vpshufb %ymm31, %ymm19, %ymm9
; AVX512BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm30 = [3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15]
; AVX512BW-FCP-NEXT:    vpshufb %ymm30, %ymm14, %ymm11
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm9 = ymm11[0,1,2,3,4],ymm9[5],ymm11[6,7]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm9[0,1,2,3,4,5],ymm0[6,7]
; AVX512BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm9 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX512BW-FCP-NEXT:    vmovdqa64 %xmm5, %xmm19
; AVX512BW-FCP-NEXT:    vpshufb %xmm9, %xmm5, %xmm11
; AVX512BW-FCP-NEXT:    vmovdqa64 %xmm1, %xmm20
; AVX512BW-FCP-NEXT:    vpshufb %xmm9, %xmm1, %xmm12
; AVX512BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm12[0],xmm11[0],xmm12[1],xmm11[1],xmm12[2],xmm11[2],xmm12[3],xmm11[3]
; AVX512BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm12 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX512BW-FCP-NEXT:    vpshufb %xmm12, %xmm7, %xmm14
; AVX512BW-FCP-NEXT:    vpshufb %xmm12, %xmm28, %xmm15
; AVX512BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} xmm11 = xmm14[0,1,2],xmm11[3]
; AVX512BW-FCP-NEXT:    vpsrlq $24, %zmm18, %zmm14
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm18, %zmm25
; AVX512BW-FCP-NEXT:    vpmovqb %zmm14, %xmm14
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} xmm11 = xmm14[0,1],xmm11[2,3]
; AVX512BW-FCP-NEXT:    vinserti32x4 $2, %xmm11, %zmm0, %zmm11
; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm0, %zmm0, %zmm11 {%k1}
; AVX512BW-FCP-NEXT:    vpshufb %ymm6, %ymm17, %ymm0
; AVX512BW-FCP-NEXT:    vpshufb %ymm4, %ymm3, %ymm3
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm3[0,1,2,3,4,5,6],ymm0[7]
; AVX512BW-FCP-NEXT:    vpshufb %ymm31, %ymm2, %ymm2
; AVX512BW-FCP-NEXT:    vpshufb %ymm30, %ymm16, %ymm1
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3,4],ymm2[5],ymm1[6,7]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
; AVX512BW-FCP-NEXT:    vpshufb %xmm9, %xmm26, %xmm1
; AVX512BW-FCP-NEXT:    vpshufb %xmm9, %xmm24, %xmm2
; AVX512BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX512BW-FCP-NEXT:    vpshufb %xmm12, %xmm22, %xmm2
; AVX512BW-FCP-NEXT:    vpshufb %xmm12, %xmm8, %xmm3
; AVX512BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} xmm1 = xmm2[0,1,2],xmm1[3]
; AVX512BW-FCP-NEXT:    vpsrlq $24, %zmm10, %zmm2
; AVX512BW-FCP-NEXT:    vpmovqb %zmm2, %xmm2
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm11, %zmm0
; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm3 = [1,3,2,3,1,3,5,7]
; AVX512BW-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm3, %ymm6 # 32-byte Folded Reload
; AVX512BW-FCP-NEXT:    vpermd (%rsp), %ymm3, %ymm7 # 32-byte Folded Reload
; AVX512BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm13 = [0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12]
; AVX512BW-FCP-NEXT:    vpshufb %ymm13, %ymm6, %ymm2
; AVX512BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm1 = [0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12]
; AVX512BW-FCP-NEXT:    vpshufb %ymm1, %ymm7, %ymm9
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm9 = ymm9[0,1,2,3,4,5,6],ymm2[7]
; AVX512BW-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm3, %ymm5 # 32-byte Folded Reload
; AVX512BW-FCP-NEXT:    vpermd %ymm29, %ymm3, %ymm14
; AVX512BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm0 = [0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12]
; AVX512BW-FCP-NEXT:    vpshufb %ymm0, %ymm5, %ymm11
; AVX512BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm4 = [0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12]
; AVX512BW-FCP-NEXT:    vpshufb %ymm4, %ymm14, %ymm12
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm11 = ymm12[0,1,2,3,4],ymm11[5],ymm12[6,7]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm9 = ymm11[0,1,2,3,4,5],ymm9[6,7]
; AVX512BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm11 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX512BW-FCP-NEXT:    vpshufb %xmm11, %xmm19, %xmm12
; AVX512BW-FCP-NEXT:    vmovdqa64 %xmm19, %xmm16
; AVX512BW-FCP-NEXT:    vpshufb %xmm11, %xmm20, %xmm15
; AVX512BW-FCP-NEXT:    vmovdqa64 %xmm20, %xmm17
; AVX512BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm15[0],xmm12[0],xmm15[1],xmm12[1],xmm15[2],xmm12[2],xmm15[3],xmm12[3]
; AVX512BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm20 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX512BW-FCP-NEXT:    vpshufb %xmm20, %xmm23, %xmm15
; AVX512BW-FCP-NEXT:    vpshufb %xmm20, %xmm28, %xmm29
; AVX512BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm15 = xmm29[0],xmm15[0],xmm29[1],xmm15[1],xmm29[2],xmm15[2],xmm29[3],xmm15[3]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} xmm12 = xmm15[0,1,2],xmm12[3]
; AVX512BW-FCP-NEXT:    vpsrlq $32, %zmm18, %zmm15
; AVX512BW-FCP-NEXT:    vpmovqb %zmm15, %xmm15
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} xmm12 = xmm15[0,1],xmm12[2,3]
; AVX512BW-FCP-NEXT:    vinserti32x4 $2, %xmm12, %zmm0, %zmm12
; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm9, %zmm0, %zmm12 {%k1}
; AVX512BW-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm3, %ymm19 # 32-byte Folded Reload
; AVX512BW-FCP-NEXT:    vpshufb %ymm13, %ymm19, %ymm9
; AVX512BW-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm3, %ymm18 # 32-byte Folded Reload
; AVX512BW-FCP-NEXT:    vpshufb %ymm1, %ymm18, %ymm15
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm9 = ymm15[0,1,2,3,4,5,6],ymm9[7]
; AVX512BW-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm3, %ymm15 # 32-byte Folded Reload
; AVX512BW-FCP-NEXT:    vpermd %ymm27, %ymm3, %ymm21
; AVX512BW-FCP-NEXT:    vpshufb %ymm0, %ymm15, %ymm3
; AVX512BW-FCP-NEXT:    vpshufb %ymm4, %ymm21, %ymm13
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm3 = ymm13[0,1,2,3,4],ymm3[5],ymm13[6,7]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm9[6,7]
; AVX512BW-FCP-NEXT:    vpshufb %xmm11, %xmm26, %xmm9
; AVX512BW-FCP-NEXT:    vpshufb %xmm11, %xmm24, %xmm11
; AVX512BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm11[0],xmm9[0],xmm11[1],xmm9[1],xmm11[2],xmm9[2],xmm11[3],xmm9[3]
; AVX512BW-FCP-NEXT:    vpshufb %xmm20, %xmm22, %xmm11
; AVX512BW-FCP-NEXT:    vpshufb %xmm20, %xmm8, %xmm13
; AVX512BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm13[0],xmm11[0],xmm13[1],xmm11[1],xmm13[2],xmm11[2],xmm13[3],xmm11[3]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} xmm9 = xmm11[0,1,2],xmm9[3]
; AVX512BW-FCP-NEXT:    vpsrlq $32, %zmm10, %zmm11
; AVX512BW-FCP-NEXT:    vpmovqb %zmm11, %xmm11
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} xmm9 = xmm11[0,1],xmm9[2,3]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm3 = ymm9[0,1,2,3],ymm3[4,5,6,7]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm3, %zmm12, %zmm29
; AVX512BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm4 = [1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13]
; AVX512BW-FCP-NEXT:    vpshufb %ymm4, %ymm6, %ymm9
; AVX512BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm0 = [1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13]
; AVX512BW-FCP-NEXT:    vpshufb %ymm0, %ymm7, %ymm11
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm9 = ymm11[0,1,2,3,4,5,6],ymm9[7]
; AVX512BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm2 = [1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13]
; AVX512BW-FCP-NEXT:    vpshufb %ymm2, %ymm5, %ymm11
; AVX512BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm1 = [1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13]
; AVX512BW-FCP-NEXT:    vpshufb %ymm1, %ymm14, %ymm12
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm11 = ymm12[0,1,2,3,4],ymm11[5],ymm12[6,7]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm9 = ymm11[0,1,2,3,4,5],ymm9[6,7]
; AVX512BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm11 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX512BW-FCP-NEXT:    vpshufb %xmm11, %xmm16, %xmm12
; AVX512BW-FCP-NEXT:    vpshufb %xmm11, %xmm17, %xmm13
; AVX512BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm13[0],xmm12[0],xmm13[1],xmm12[1],xmm13[2],xmm12[2],xmm13[3],xmm12[3]
; AVX512BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm13 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX512BW-FCP-NEXT:    vpshufb %xmm13, %xmm23, %xmm20
; AVX512BW-FCP-NEXT:    vpshufb %xmm13, %xmm28, %xmm27
; AVX512BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm27[0],xmm20[0],xmm27[1],xmm20[1],xmm27[2],xmm20[2],xmm27[3],xmm20[3]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} xmm3 = xmm3[0,1,2],xmm12[3]
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm25, %zmm27
; AVX512BW-FCP-NEXT:    vpsrlq $40, %zmm25, %zmm12
; AVX512BW-FCP-NEXT:    vpmovqb %zmm12, %xmm12
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} xmm3 = xmm12[0,1],xmm3[2,3]
; AVX512BW-FCP-NEXT:    vinserti32x4 $2, %xmm3, %zmm0, %zmm3
; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm9, %zmm0, %zmm3 {%k1}
; AVX512BW-FCP-NEXT:    vpshufb %ymm4, %ymm19, %ymm9
; AVX512BW-FCP-NEXT:    vpshufb %ymm0, %ymm18, %ymm12
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm9 = ymm12[0,1,2,3,4,5,6],ymm9[7]
; AVX512BW-FCP-NEXT:    vpshufb %ymm2, %ymm15, %ymm12
; AVX512BW-FCP-NEXT:    vpshufb %ymm1, %ymm21, %ymm4
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3,4],ymm12[5],ymm4[6,7]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm9[6,7]
; AVX512BW-FCP-NEXT:    vpshufb %xmm11, %xmm26, %xmm9
; AVX512BW-FCP-NEXT:    vpshufb %xmm11, %xmm24, %xmm11
; AVX512BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm11[0],xmm9[0],xmm11[1],xmm9[1],xmm11[2],xmm9[2],xmm11[3],xmm9[3]
; AVX512BW-FCP-NEXT:    vpshufb %xmm13, %xmm22, %xmm11
; AVX512BW-FCP-NEXT:    vpshufb %xmm13, %xmm8, %xmm12
; AVX512BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm12[0],xmm11[0],xmm12[1],xmm11[1],xmm12[2],xmm11[2],xmm12[3],xmm11[3]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} xmm9 = xmm11[0,1,2],xmm9[3]
; AVX512BW-FCP-NEXT:    vpsrlq $40, %zmm10, %zmm11
; AVX512BW-FCP-NEXT:    vpmovqb %zmm11, %xmm11
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} xmm9 = xmm11[0,1],xmm9[2,3]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm4 = ymm9[0,1,2,3],ymm4[4,5,6,7]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm4, %zmm3, %zmm20
; AVX512BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm0 = [2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14]
; AVX512BW-FCP-NEXT:    vpshufb %ymm0, %ymm6, %ymm3
; AVX512BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm1 = [2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14]
; AVX512BW-FCP-NEXT:    vpshufb %ymm1, %ymm7, %ymm4
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm3 = ymm4[0,1,2,3,4,5,6],ymm3[7]
; AVX512BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm2 = [2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14]
; AVX512BW-FCP-NEXT:    vpshufb %ymm2, %ymm5, %ymm4
; AVX512BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm25 = [2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14]
; AVX512BW-FCP-NEXT:    vpshufb %ymm25, %ymm14, %ymm9
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm4 = ymm9[0,1,2,3,4],ymm4[5],ymm9[6,7]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm3 = ymm4[0,1,2,3,4,5],ymm3[6,7]
; AVX512BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm4 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX512BW-FCP-NEXT:    vpshufb %xmm4, %xmm16, %xmm9
; AVX512BW-FCP-NEXT:    vpshufb %xmm4, %xmm17, %xmm11
; AVX512BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm11[0],xmm9[0],xmm11[1],xmm9[1],xmm11[2],xmm9[2],xmm11[3],xmm9[3]
; AVX512BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm11 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX512BW-FCP-NEXT:    vpshufb %xmm11, %xmm23, %xmm12
; AVX512BW-FCP-NEXT:    vpshufb %xmm11, %xmm28, %xmm13
; AVX512BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm13[0],xmm12[0],xmm13[1],xmm12[1],xmm13[2],xmm12[2],xmm13[3],xmm12[3]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} xmm9 = xmm12[0,1,2],xmm9[3]
; AVX512BW-FCP-NEXT:    vpsrlq $48, %zmm27, %zmm12
; AVX512BW-FCP-NEXT:    vpmovqb %zmm12, %xmm12
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} xmm9 = xmm12[0,1],xmm9[2,3]
; AVX512BW-FCP-NEXT:    vinserti32x4 $2, %xmm9, %zmm0, %zmm9
; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm3, %zmm0, %zmm9 {%k1}
; AVX512BW-FCP-NEXT:    vpshufb %ymm0, %ymm19, %ymm3
; AVX512BW-FCP-NEXT:    vpshufb %ymm1, %ymm18, %ymm12
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm3 = ymm12[0,1,2,3,4,5,6],ymm3[7]
; AVX512BW-FCP-NEXT:    vpshufb %ymm2, %ymm15, %ymm12
; AVX512BW-FCP-NEXT:    vpshufb %ymm25, %ymm21, %ymm13
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm12 = ymm13[0,1,2,3,4],ymm12[5],ymm13[6,7]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm3 = ymm12[0,1,2,3,4,5],ymm3[6,7]
; AVX512BW-FCP-NEXT:    vpshufb %xmm4, %xmm26, %xmm12
; AVX512BW-FCP-NEXT:    vpshufb %xmm4, %xmm24, %xmm4
; AVX512BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm4[0],xmm12[0],xmm4[1],xmm12[1],xmm4[2],xmm12[2],xmm4[3],xmm12[3]
; AVX512BW-FCP-NEXT:    vpshufb %xmm11, %xmm22, %xmm12
; AVX512BW-FCP-NEXT:    vpshufb %xmm11, %xmm8, %xmm11
; AVX512BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm11[0],xmm12[0],xmm11[1],xmm12[1],xmm11[2],xmm12[2],xmm11[3],xmm12[3]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} xmm4 = xmm11[0,1,2],xmm4[3]
; AVX512BW-FCP-NEXT:    vpsrlq $48, %zmm10, %zmm11
; AVX512BW-FCP-NEXT:    vpmovqb %zmm11, %xmm11
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} xmm4 = xmm11[0,1],xmm4[2,3]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm3 = ymm4[0,1,2,3],ymm3[4,5,6,7]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm3, %zmm9, %zmm3
; AVX512BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm12 = [3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15]
; AVX512BW-FCP-NEXT:    vpshufb %ymm12, %ymm6, %ymm0
; AVX512BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm6 = [3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15]
; AVX512BW-FCP-NEXT:    vpshufb %ymm6, %ymm7, %ymm1
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5,6],ymm0[7]
; AVX512BW-FCP-NEXT:    vpshufb %ymm31, %ymm5, %ymm1
; AVX512BW-FCP-NEXT:    vpshufb %ymm30, %ymm14, %ymm2
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3,4],ymm1[5],ymm2[6,7]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
; AVX512BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm1 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX512BW-FCP-NEXT:    vpshufb %xmm1, %xmm16, %xmm2
; AVX512BW-FCP-NEXT:    vpshufb %xmm1, %xmm17, %xmm4
; AVX512BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm4[0],xmm2[0],xmm4[1],xmm2[1],xmm4[2],xmm2[2],xmm4[3],xmm2[3]
; AVX512BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm4 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX512BW-FCP-NEXT:    vpshufb %xmm4, %xmm23, %xmm9
; AVX512BW-FCP-NEXT:    vpshufb %xmm4, %xmm28, %xmm11
; AVX512BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm11[0],xmm9[0],xmm11[1],xmm9[1],xmm11[2],xmm9[2],xmm11[3],xmm9[3]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} xmm2 = xmm9[0,1,2],xmm2[3]
; AVX512BW-FCP-NEXT:    vpsrlq $56, %zmm27, %zmm9
; AVX512BW-FCP-NEXT:    vpmovqb %zmm9, %xmm9
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} xmm2 = xmm9[0,1],xmm2[2,3]
; AVX512BW-FCP-NEXT:    vinserti32x4 $2, %xmm2, %zmm0, %zmm2
; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm0, %zmm0, %zmm2 {%k1}
; AVX512BW-FCP-NEXT:    vpshufb %ymm12, %ymm19, %ymm0
; AVX512BW-FCP-NEXT:    vpshufb %ymm6, %ymm18, %ymm9
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm9[0,1,2,3,4,5,6],ymm0[7]
; AVX512BW-FCP-NEXT:    vpshufb %ymm31, %ymm15, %ymm9
; AVX512BW-FCP-NEXT:    vpshufb %ymm30, %ymm21, %ymm11
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm9 = ymm11[0,1,2,3,4],ymm9[5],ymm11[6,7]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm9[0,1,2,3,4,5],ymm0[6,7]
; AVX512BW-FCP-NEXT:    vpshufb %xmm1, %xmm26, %xmm5
; AVX512BW-FCP-NEXT:    vpshufb %xmm1, %xmm24, %xmm1
; AVX512BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm5[0],xmm1[1],xmm5[1],xmm1[2],xmm5[2],xmm1[3],xmm5[3]
; AVX512BW-FCP-NEXT:    vpshufb %xmm4, %xmm22, %xmm5
; AVX512BW-FCP-NEXT:    vpshufb %xmm4, %xmm8, %xmm4
; AVX512BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} xmm1 = xmm4[0,1,2],xmm1[3]
; AVX512BW-FCP-NEXT:    vpsrlq $56, %zmm10, %zmm4
; AVX512BW-FCP-NEXT:    vpmovqb %zmm4, %xmm4
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} xmm1 = xmm4[0,1],xmm1[2,3]
; AVX512BW-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX512BW-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm2, %zmm0
; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovaps %zmm1, (%rsi)
; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovaps %zmm1, (%rdx)
; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovaps %zmm1, (%rcx)
; AVX512BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512BW-FCP-NEXT:    vmovaps %zmm1, (%r8)
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm29, (%r9)
; AVX512BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm20, (%rax)
; AVX512BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm3, (%rax)
; AVX512BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm0, (%rax)
; AVX512BW-FCP-NEXT:    addq $328, %rsp # imm = 0x148
; AVX512BW-FCP-NEXT:    vzeroupper
; AVX512BW-FCP-NEXT:    retq
;
; AVX512DQ-BW-LABEL: load_i8_stride8_vf64:
; AVX512DQ-BW:       # %bb.0:
; AVX512DQ-BW-NEXT:    subq $744, %rsp # imm = 0x2E8
; AVX512DQ-BW-NEXT:    vmovdqa64 (%rdi), %zmm5
; AVX512DQ-BW-NEXT:    vmovdqa64 128(%rdi), %zmm0
; AVX512DQ-BW-NEXT:    vpmovqb %zmm0, %xmm0
; AVX512DQ-BW-NEXT:    vmovdqa64 256(%rdi), %zmm1
; AVX512DQ-BW-NEXT:    vmovdqa64 384(%rdi), %zmm2
; AVX512DQ-BW-NEXT:    vpmovqb %zmm2, %xmm2
; AVX512DQ-BW-NEXT:    vmovdqa 496(%rdi), %xmm4
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm12 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX512DQ-BW-NEXT:    vpshufb %xmm12, %xmm4, %xmm3
; AVX512DQ-BW-NEXT:    vmovdqa64 %xmm4, %xmm24
; AVX512DQ-BW-NEXT:    vmovdqa 480(%rdi), %xmm6
; AVX512DQ-BW-NEXT:    vpshufb %xmm12, %xmm6, %xmm4
; AVX512DQ-BW-NEXT:    vmovdqa64 %xmm6, %xmm25
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX512DQ-BW-NEXT:    vmovdqa 464(%rdi), %xmm6
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm19 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX512DQ-BW-NEXT:    vpshufb %xmm19, %xmm6, %xmm4
; AVX512DQ-BW-NEXT:    vmovdqa64 %xmm6, %xmm26
; AVX512DQ-BW-NEXT:    vmovdqa 448(%rdi), %xmm7
; AVX512DQ-BW-NEXT:    vpshufb %xmm19, %xmm7, %xmm6
; AVX512DQ-BW-NEXT:    vmovdqa64 %xmm7, %xmm30
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm6[0],xmm4[0],xmm6[1],xmm4[1],xmm6[2],xmm4[2],xmm6[3],xmm4[3]
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm4, %ymm0, %ymm4
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm3 = ymm4[0,1,2,3,4,5,6],ymm3[7]
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX512DQ-BW-NEXT:    vmovdqa 384(%rdi), %ymm4
; AVX512DQ-BW-NEXT:    vpmovqb %ymm4, %xmm4
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm4, %ymm0, %ymm4
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm2 = ymm4[0,1,2,3,4],ymm2[5],ymm4[6,7]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm7 = ymm2[0,1,2,3,4,5],ymm3[6,7]
; AVX512DQ-BW-NEXT:    vmovdqa 368(%rdi), %xmm2
; AVX512DQ-BW-NEXT:    vpshufb %xmm12, %xmm2, %xmm4
; AVX512DQ-BW-NEXT:    vmovdqa64 %xmm2, %xmm31
; AVX512DQ-BW-NEXT:    vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 352(%rdi), %xmm27
; AVX512DQ-BW-NEXT:    vpshufb %xmm12, %xmm27, %xmm6
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm6[0],xmm4[0],xmm6[1],xmm4[1],xmm6[2],xmm4[2],xmm6[3],xmm4[3]
; AVX512DQ-BW-NEXT:    vmovdqa 336(%rdi), %xmm2
; AVX512DQ-BW-NEXT:    vpshufb %xmm19, %xmm2, %xmm11
; AVX512DQ-BW-NEXT:    vmovdqa64 %xmm2, %xmm22
; AVX512DQ-BW-NEXT:    vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa 320(%rdi), %xmm2
; AVX512DQ-BW-NEXT:    vpshufb %xmm19, %xmm2, %xmm15
; AVX512DQ-BW-NEXT:    vmovdqa %xmm2, %xmm9
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm15[0],xmm11[0],xmm15[1],xmm11[1],xmm15[2],xmm11[2],xmm15[3],xmm11[3]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} xmm10 = xmm11[0,1,2],xmm10[3]
; AVX512DQ-BW-NEXT:    vpmovqb %zmm1, %xmm11
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm1, %zmm15
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} xmm10 = xmm11[0,1],xmm10[2,3]
; AVX512DQ-BW-NEXT:    vinserti32x4 $2, %xmm10, %zmm0, %zmm20
; AVX512DQ-BW-NEXT:    movb $-64, %al
; AVX512DQ-BW-NEXT:    kmovd %eax, %k1
; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm7, %zmm0, %zmm20 {%k1}
; AVX512DQ-BW-NEXT:    vmovdqa64 240(%rdi), %xmm28
; AVX512DQ-BW-NEXT:    vpshufb %xmm12, %xmm28, %xmm7
; AVX512DQ-BW-NEXT:    vmovdqa 224(%rdi), %xmm1
; AVX512DQ-BW-NEXT:    vpshufb %xmm12, %xmm1, %xmm10
; AVX512DQ-BW-NEXT:    vmovdqa64 %xmm1, %xmm18
; AVX512DQ-BW-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm7 = xmm10[0],xmm7[0],xmm10[1],xmm7[1],xmm10[2],xmm7[2],xmm10[3],xmm7[3]
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm7, %ymm0, %ymm7
; AVX512DQ-BW-NEXT:    vmovdqa64 208(%rdi), %xmm17
; AVX512DQ-BW-NEXT:    vpshufb %xmm19, %xmm17, %xmm10
; AVX512DQ-BW-NEXT:    vmovdqa64 %xmm17, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa 192(%rdi), %xmm8
; AVX512DQ-BW-NEXT:    vpshufb %xmm19, %xmm8, %xmm16
; AVX512DQ-BW-NEXT:    vmovdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm16[0],xmm10[0],xmm16[1],xmm10[1],xmm16[2],xmm10[2],xmm16[3],xmm10[3]
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm10, %ymm0, %ymm10
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm7 = ymm10[0,1,2,3,4,5,6],ymm7[7]
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX512DQ-BW-NEXT:    vmovdqa 128(%rdi), %ymm10
; AVX512DQ-BW-NEXT:    vpmovqb %ymm10, %xmm10
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm10, %ymm0, %ymm10
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm10[0,1,2,3,4],ymm0[5],ymm10[6,7]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm0[0,1,2,3,4,5],ymm7[6,7]
; AVX512DQ-BW-NEXT:    vmovdqa 112(%rdi), %xmm0
; AVX512DQ-BW-NEXT:    vpshufb %xmm12, %xmm0, %xmm16
; AVX512DQ-BW-NEXT:    vmovdqa %xmm0, %xmm10
; AVX512DQ-BW-NEXT:    vmovdqa 96(%rdi), %xmm0
; AVX512DQ-BW-NEXT:    vpshufb %xmm12, %xmm0, %xmm12
; AVX512DQ-BW-NEXT:    vmovdqa %xmm0, %xmm14
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm12[0],xmm16[0],xmm12[1],xmm16[1],xmm12[2],xmm16[2],xmm12[3],xmm16[3]
; AVX512DQ-BW-NEXT:    vmovdqa 80(%rdi), %xmm2
; AVX512DQ-BW-NEXT:    vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-BW-NEXT:    vpshufb %xmm19, %xmm2, %xmm21
; AVX512DQ-BW-NEXT:    vmovdqa 64(%rdi), %xmm2
; AVX512DQ-BW-NEXT:    vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-BW-NEXT:    vpshufb %xmm19, %xmm2, %xmm19
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm19[0],xmm21[0],xmm19[1],xmm21[1],xmm19[2],xmm21[2],xmm19[3],xmm21[3]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} xmm0 = xmm13[0,1,2],xmm0[3]
; AVX512DQ-BW-NEXT:    vpmovqb %zmm5, %xmm13
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} xmm0 = xmm13[0,1],xmm0[2,3]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm0, %zmm20, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa 160(%rdi), %xmm6
; AVX512DQ-BW-NEXT:    vmovdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa 384(%rdi), %xmm7
; AVX512DQ-BW-NEXT:    vmovdqa 400(%rdi), %xmm3
; AVX512DQ-BW-NEXT:    vmovdqa64 416(%rdi), %xmm20
; AVX512DQ-BW-NEXT:    vmovdqa64 432(%rdi), %xmm29
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm4 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX512DQ-BW-NEXT:    vmovdqa64 %xmm24, %xmm19
; AVX512DQ-BW-NEXT:    vpshufb %xmm4, %xmm24, %xmm1
; AVX512DQ-BW-NEXT:    vmovdqa64 %xmm25, %xmm11
; AVX512DQ-BW-NEXT:    vpshufb %xmm4, %xmm25, %xmm13
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm13[0],xmm1[0],xmm13[1],xmm1[1],xmm13[2],xmm1[2],xmm13[3],xmm1[3]
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm1
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm13 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX512DQ-BW-NEXT:    vmovdqa64 %xmm26, %xmm12
; AVX512DQ-BW-NEXT:    vpshufb %xmm13, %xmm26, %xmm24
; AVX512DQ-BW-NEXT:    vmovdqa64 %xmm30, %xmm16
; AVX512DQ-BW-NEXT:    vmovdqa64 %xmm30, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-BW-NEXT:    vpshufb %xmm13, %xmm30, %xmm25
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm24 = xmm25[0],xmm24[0],xmm25[1],xmm24[1],xmm25[2],xmm24[2],xmm25[3],xmm24[3]
; AVX512DQ-BW-NEXT:    vinserti32x4 $1, %xmm24, %ymm0, %ymm2
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5,6],ymm1[7]
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm26 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX512DQ-BW-NEXT:    vpshufb %xmm26, %xmm29, %xmm24
; AVX512DQ-BW-NEXT:    vpshufb %xmm26, %xmm20, %xmm25
; AVX512DQ-BW-NEXT:    vmovdqa64 %xmm20, %xmm23
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm24 = xmm25[0],xmm24[0],xmm25[1],xmm24[1],xmm25[2],xmm24[2],xmm25[3],xmm24[3]
; AVX512DQ-BW-NEXT:    vinserti32x4 $1, %xmm24, %ymm0, %ymm2
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm30 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX512DQ-BW-NEXT:    vpshufb %xmm30, %xmm3, %xmm24
; AVX512DQ-BW-NEXT:    vmovdqa64 %xmm3, %xmm21
; AVX512DQ-BW-NEXT:    vpshufb %xmm30, %xmm7, %xmm25
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm24 = xmm25[0],xmm24[0],xmm25[1],xmm24[1],xmm25[2],xmm24[2],xmm25[3],xmm24[3]
; AVX512DQ-BW-NEXT:    vinserti32x4 $1, %xmm24, %ymm0, %ymm3
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4],ymm2[5],ymm3[6,7]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]
; AVX512DQ-BW-NEXT:    vpshufb %xmm4, %xmm31, %xmm2
; AVX512DQ-BW-NEXT:    vpshufb %xmm4, %xmm27, %xmm3
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
; AVX512DQ-BW-NEXT:    vpshufb %xmm13, %xmm22, %xmm3
; AVX512DQ-BW-NEXT:    vpshufb %xmm13, %xmm9, %xmm24
; AVX512DQ-BW-NEXT:    vmovdqa64 %xmm9, %xmm31
; AVX512DQ-BW-NEXT:    vmovdqa %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm24[0],xmm3[0],xmm24[1],xmm3[1],xmm24[2],xmm3[2],xmm24[3],xmm3[3]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} xmm2 = xmm3[0,1,2],xmm2[3]
; AVX512DQ-BW-NEXT:    vpsrlq $8, %zmm15, %zmm3
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm15, %zmm22
; AVX512DQ-BW-NEXT:    vpmovqb %zmm3, %xmm3
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} xmm2 = xmm3[0,1],xmm2[2,3]
; AVX512DQ-BW-NEXT:    vinserti32x4 $2, %xmm2, %zmm0, %zmm2
; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm2 {%k1}
; AVX512DQ-BW-NEXT:    vmovdqa64 %xmm28, %xmm9
; AVX512DQ-BW-NEXT:    vpshufb %xmm4, %xmm28, %xmm1
; AVX512DQ-BW-NEXT:    vpshufb %xmm4, %xmm18, %xmm3
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3]
; AVX512DQ-BW-NEXT:    vpshufb %xmm13, %xmm17, %xmm3
; AVX512DQ-BW-NEXT:    vpshufb %xmm13, %xmm8, %xmm24
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm24[0],xmm3[0],xmm24[1],xmm3[1],xmm24[2],xmm3[2],xmm24[3],xmm3[3]
; AVX512DQ-BW-NEXT:    vmovdqa 176(%rdi), %xmm0
; AVX512DQ-BW-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm1
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm3[0,1,2,3,4,5,6],ymm1[7]
; AVX512DQ-BW-NEXT:    vpshufb %xmm26, %xmm0, %xmm3
; AVX512DQ-BW-NEXT:    vpshufb %xmm26, %xmm6, %xmm25
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm25[0],xmm3[0],xmm25[1],xmm3[1],xmm25[2],xmm3[2],xmm25[3],xmm3[3]
; AVX512DQ-BW-NEXT:    vmovdqa 128(%rdi), %xmm6
; AVX512DQ-BW-NEXT:    vmovdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa 144(%rdi), %xmm0
; AVX512DQ-BW-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-BW-NEXT:    vpshufb %xmm30, %xmm0, %xmm0
; AVX512DQ-BW-NEXT:    vpshufb %xmm30, %xmm6, %xmm30
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm30[0],xmm0[0],xmm30[1],xmm0[1],xmm30[2],xmm0[2],xmm30[3],xmm0[3]
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm3[5],ymm0[6,7]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
; AVX512DQ-BW-NEXT:    vpshufb %xmm4, %xmm10, %xmm1
; AVX512DQ-BW-NEXT:    vmovdqa %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-BW-NEXT:    vpshufb %xmm4, %xmm14, %xmm3
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3]
; AVX512DQ-BW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX512DQ-BW-NEXT:    vpshufb %xmm13, %xmm15, %xmm3
; AVX512DQ-BW-NEXT:    vmovdqa64 {{[-0-9]+}}(%r{{[sb]}}p), %xmm28 # 16-byte Reload
; AVX512DQ-BW-NEXT:    vpshufb %xmm13, %xmm28, %xmm4
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} xmm1 = xmm3[0,1,2],xmm1[3]
; AVX512DQ-BW-NEXT:    vpsrlq $8, %zmm5, %zmm3
; AVX512DQ-BW-NEXT:    vpmovqb %zmm3, %xmm3
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} xmm1 = xmm3[0,1],xmm1[2,3]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm0, %zmm2, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm1 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX512DQ-BW-NEXT:    vpshufb %xmm1, %xmm19, %xmm2
; AVX512DQ-BW-NEXT:    vmovdqa64 %xmm19, %xmm20
; AVX512DQ-BW-NEXT:    vpshufb %xmm1, %xmm11, %xmm3
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm3 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX512DQ-BW-NEXT:    vpshufb %xmm3, %xmm12, %xmm4
; AVX512DQ-BW-NEXT:    vmovdqa64 %xmm12, %xmm25
; AVX512DQ-BW-NEXT:    vmovdqa %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-BW-NEXT:    vpshufb %xmm3, %xmm16, %xmm13
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm13[0],xmm4[0],xmm13[1],xmm4[1],xmm13[2],xmm4[2],xmm13[3],xmm4[3]
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm4, %ymm0, %ymm4
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm2 = ymm4[0,1,2,3,4,5,6],ymm2[7]
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm4 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX512DQ-BW-NEXT:    vmovdqa64 %xmm29, %xmm24
; AVX512DQ-BW-NEXT:    vpshufb %xmm4, %xmm29, %xmm13
; AVX512DQ-BW-NEXT:    vpshufb %xmm4, %xmm23, %xmm30
; AVX512DQ-BW-NEXT:    vmovdqa64 %xmm23, %xmm16
; AVX512DQ-BW-NEXT:    vmovdqa64 %xmm23, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm30[0],xmm13[0],xmm30[1],xmm13[1],xmm30[2],xmm13[2],xmm30[3],xmm13[3]
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm30 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX512DQ-BW-NEXT:    vmovdqa64 %xmm21, %xmm18
; AVX512DQ-BW-NEXT:    vpshufb %xmm30, %xmm21, %xmm0
; AVX512DQ-BW-NEXT:    vmovdqa64 %xmm7, %xmm17
; AVX512DQ-BW-NEXT:    vmovdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-BW-NEXT:    vpshufb %xmm30, %xmm7, %xmm19
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm19[0],xmm0[0],xmm19[1],xmm0[1],xmm19[2],xmm0[2],xmm19[3],xmm0[3]
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm13, %ymm0, %ymm13
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm13[5],ymm0[6,7]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm2[6,7]
; AVX512DQ-BW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload
; AVX512DQ-BW-NEXT:    vpshufb %xmm1, %xmm12, %xmm2
; AVX512DQ-BW-NEXT:    vpshufb %xmm1, %xmm27, %xmm13
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm13[0],xmm2[0],xmm13[1],xmm2[1],xmm13[2],xmm2[2],xmm13[3],xmm2[3]
; AVX512DQ-BW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX512DQ-BW-NEXT:    vpshufb %xmm3, %xmm8, %xmm13
; AVX512DQ-BW-NEXT:    vpshufb %xmm3, %xmm31, %xmm19
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm19[0],xmm13[0],xmm19[1],xmm13[1],xmm19[2],xmm13[2],xmm19[3],xmm13[3]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} xmm2 = xmm13[0,1,2],xmm2[3]
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm22, %zmm6
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm22, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpsrlq $16, %zmm22, %zmm13
; AVX512DQ-BW-NEXT:    vpmovqb %zmm13, %xmm13
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} xmm2 = xmm13[0,1],xmm2[2,3]
; AVX512DQ-BW-NEXT:    vinserti32x4 $2, %xmm2, %zmm0, %zmm2
; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm0, %zmm0, %zmm2 {%k1}
; AVX512DQ-BW-NEXT:    vpshufb %xmm1, %xmm9, %xmm0
; AVX512DQ-BW-NEXT:    vmovdqa64 %xmm9, %xmm26
; AVX512DQ-BW-NEXT:    vmovdqa %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-BW-NEXT:    vmovdqa64 {{[-0-9]+}}(%r{{[sb]}}p), %xmm21 # 16-byte Reload
; AVX512DQ-BW-NEXT:    vpshufb %xmm1, %xmm21, %xmm13
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm13[0],xmm0[0],xmm13[1],xmm0[1],xmm13[2],xmm0[2],xmm13[3],xmm0[3]
; AVX512DQ-BW-NEXT:    vmovdqa64 {{[-0-9]+}}(%r{{[sb]}}p), %xmm22 # 16-byte Reload
; AVX512DQ-BW-NEXT:    vpshufb %xmm3, %xmm22, %xmm13
; AVX512DQ-BW-NEXT:    vmovdqa64 {{[-0-9]+}}(%r{{[sb]}}p), %xmm29 # 16-byte Reload
; AVX512DQ-BW-NEXT:    vpshufb %xmm3, %xmm29, %xmm19
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm19[0],xmm13[0],xmm19[1],xmm13[1],xmm19[2],xmm13[2],xmm19[3],xmm13[3]
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm13, %ymm0, %ymm13
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm13[0,1,2,3,4,5,6],ymm0[7]
; AVX512DQ-BW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
; AVX512DQ-BW-NEXT:    vpshufb %xmm4, %xmm9, %xmm13
; AVX512DQ-BW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload
; AVX512DQ-BW-NEXT:    vpshufb %xmm4, %xmm7, %xmm4
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm4[0],xmm13[0],xmm4[1],xmm13[1],xmm4[2],xmm13[2],xmm4[3],xmm13[3]
; AVX512DQ-BW-NEXT:    vmovdqa64 {{[-0-9]+}}(%r{{[sb]}}p), %xmm31 # 16-byte Reload
; AVX512DQ-BW-NEXT:    vpshufb %xmm30, %xmm31, %xmm13
; AVX512DQ-BW-NEXT:    vmovdqa64 {{[-0-9]+}}(%r{{[sb]}}p), %xmm23 # 16-byte Reload
; AVX512DQ-BW-NEXT:    vpshufb %xmm30, %xmm23, %xmm19
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm19[0],xmm13[0],xmm19[1],xmm13[1],xmm19[2],xmm13[2],xmm19[3],xmm13[3]
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm4, %ymm0, %ymm4
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm13, %ymm0, %ymm13
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm4 = ymm13[0,1,2,3,4],ymm4[5],ymm13[6,7]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm4[0,1,2,3,4,5],ymm0[6,7]
; AVX512DQ-BW-NEXT:    vpshufb %xmm1, %xmm10, %xmm4
; AVX512DQ-BW-NEXT:    vpshufb %xmm1, %xmm14, %xmm1
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm4[0],xmm1[1],xmm4[1],xmm1[2],xmm4[2],xmm1[3],xmm4[3]
; AVX512DQ-BW-NEXT:    vpshufb %xmm3, %xmm15, %xmm4
; AVX512DQ-BW-NEXT:    vpshufb %xmm3, %xmm28, %xmm3
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} xmm1 = xmm3[0,1,2],xmm1[3]
; AVX512DQ-BW-NEXT:    vpsrlq $16, %zmm5, %zmm3
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm5, %zmm28
; AVX512DQ-BW-NEXT:    vpmovqb %zmm3, %xmm3
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} xmm1 = xmm3[0,1],xmm1[2,3]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm0, %zmm2, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm0 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX512DQ-BW-NEXT:    vmovdqa64 %xmm20, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-BW-NEXT:    vpshufb %xmm0, %xmm20, %xmm1
; AVX512DQ-BW-NEXT:    vpshufb %xmm0, %xmm11, %xmm2
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm2 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX512DQ-BW-NEXT:    vpshufb %xmm2, %xmm25, %xmm3
; AVX512DQ-BW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; AVX512DQ-BW-NEXT:    vpshufb %xmm2, %xmm4, %xmm4
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm1
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm3[0,1,2,3,4,5,6],ymm1[7]
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm3 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX512DQ-BW-NEXT:    vpshufb %xmm3, %xmm24, %xmm4
; AVX512DQ-BW-NEXT:    vmovdqa64 %xmm24, (%rsp) # 16-byte Spill
; AVX512DQ-BW-NEXT:    vpshufb %xmm3, %xmm16, %xmm19
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm19[0],xmm4[0],xmm19[1],xmm4[1],xmm19[2],xmm4[2],xmm19[3],xmm4[3]
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm19 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX512DQ-BW-NEXT:    vpshufb %xmm19, %xmm18, %xmm30
; AVX512DQ-BW-NEXT:    vpshufb %xmm19, %xmm17, %xmm13
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm13[0],xmm30[0],xmm13[1],xmm30[1],xmm13[2],xmm30[2],xmm13[3],xmm30[3]
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm4, %ymm0, %ymm4
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm13, %ymm0, %ymm13
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm4 = ymm13[0,1,2,3,4],ymm4[5],ymm13[6,7]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm4[0,1,2,3,4,5],ymm1[6,7]
; AVX512DQ-BW-NEXT:    vmovdqa %xmm12, %xmm10
; AVX512DQ-BW-NEXT:    vpshufb %xmm0, %xmm12, %xmm4
; AVX512DQ-BW-NEXT:    vpshufb %xmm0, %xmm27, %xmm13
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm13[0],xmm4[0],xmm13[1],xmm4[1],xmm13[2],xmm4[2],xmm13[3],xmm4[3]
; AVX512DQ-BW-NEXT:    vpshufb %xmm2, %xmm8, %xmm13
; AVX512DQ-BW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX512DQ-BW-NEXT:    vpshufb %xmm2, %xmm8, %xmm30
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm30[0],xmm13[0],xmm30[1],xmm13[1],xmm30[2],xmm13[2],xmm30[3],xmm13[3]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} xmm4 = xmm13[0,1,2],xmm4[3]
; AVX512DQ-BW-NEXT:    vpsrlq $24, %zmm6, %zmm13
; AVX512DQ-BW-NEXT:    vpmovqb %zmm13, %xmm13
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} xmm4 = xmm13[0,1],xmm4[2,3]
; AVX512DQ-BW-NEXT:    vinserti32x4 $2, %xmm4, %zmm0, %zmm4
; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm4 {%k1}
; AVX512DQ-BW-NEXT:    vpshufb %xmm0, %xmm26, %xmm1
; AVX512DQ-BW-NEXT:    vpshufb %xmm0, %xmm21, %xmm13
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm13[0],xmm1[0],xmm13[1],xmm1[1],xmm13[2],xmm1[2],xmm13[3],xmm1[3]
; AVX512DQ-BW-NEXT:    vpshufb %xmm2, %xmm22, %xmm13
; AVX512DQ-BW-NEXT:    vmovdqa64 %xmm22, %xmm21
; AVX512DQ-BW-NEXT:    vpshufb %xmm2, %xmm29, %xmm30
; AVX512DQ-BW-NEXT:    vmovdqa64 %xmm29, %xmm22
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm30[0],xmm13[0],xmm30[1],xmm13[1],xmm30[2],xmm13[2],xmm30[3],xmm13[3]
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm1
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm13, %ymm0, %ymm13
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm13[0,1,2,3,4,5,6],ymm1[7]
; AVX512DQ-BW-NEXT:    vpshufb %xmm3, %xmm9, %xmm13
; AVX512DQ-BW-NEXT:    vmovdqa64 %xmm9, %xmm25
; AVX512DQ-BW-NEXT:    vpshufb %xmm3, %xmm7, %xmm3
; AVX512DQ-BW-NEXT:    vmovdqa %xmm7, %xmm9
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm13[0],xmm3[1],xmm13[1],xmm3[2],xmm13[2],xmm3[3],xmm13[3]
; AVX512DQ-BW-NEXT:    vpshufb %xmm19, %xmm31, %xmm13
; AVX512DQ-BW-NEXT:    vpshufb %xmm19, %xmm23, %xmm19
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm19[0],xmm13[0],xmm19[1],xmm13[1],xmm19[2],xmm13[2],xmm19[3],xmm13[3]
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm13, %ymm0, %ymm13
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm3 = ymm13[0,1,2,3,4],ymm3[5],ymm13[6,7]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm3[0,1,2,3,4,5],ymm1[6,7]
; AVX512DQ-BW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX512DQ-BW-NEXT:    vpshufb %xmm0, %xmm5, %xmm3
; AVX512DQ-BW-NEXT:    vpshufb %xmm0, %xmm14, %xmm0
; AVX512DQ-BW-NEXT:    vmovdqa64 %xmm14, %xmm23
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3]
; AVX512DQ-BW-NEXT:    vpshufb %xmm2, %xmm15, %xmm3
; AVX512DQ-BW-NEXT:    vmovdqa64 {{[-0-9]+}}(%r{{[sb]}}p), %xmm29 # 16-byte Reload
; AVX512DQ-BW-NEXT:    vpshufb %xmm2, %xmm29, %xmm2
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} xmm0 = xmm2[0,1,2],xmm0[3]
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm28, %zmm17
; AVX512DQ-BW-NEXT:    vpsrlq $24, %zmm28, %zmm2
; AVX512DQ-BW-NEXT:    vpmovqb %zmm2, %xmm2
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} xmm0 = xmm2[0,1],xmm0[2,3]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm0, %zmm4, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm0 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX512DQ-BW-NEXT:    vpshufb %xmm0, %xmm20, %xmm1
; AVX512DQ-BW-NEXT:    vpshufb %xmm0, %xmm11, %xmm3
; AVX512DQ-BW-NEXT:    vmovdqa64 %xmm11, %xmm28
; AVX512DQ-BW-NEXT:    vmovdqa %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3]
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm3 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX512DQ-BW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload
; AVX512DQ-BW-NEXT:    vpshufb %xmm3, %xmm12, %xmm4
; AVX512DQ-BW-NEXT:    vmovdqa64 {{[-0-9]+}}(%r{{[sb]}}p), %xmm16 # 16-byte Reload
; AVX512DQ-BW-NEXT:    vpshufb %xmm3, %xmm16, %xmm13
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm13[0],xmm4[0],xmm13[1],xmm4[1],xmm13[2],xmm4[2],xmm13[3],xmm4[3]
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm1
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm4, %ymm0, %ymm4
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm4[0,1,2,3,4,5,6],ymm1[7]
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm4 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX512DQ-BW-NEXT:    vpshufb %xmm4, %xmm24, %xmm13
; AVX512DQ-BW-NEXT:    vmovdqa64 {{[-0-9]+}}(%r{{[sb]}}p), %xmm26 # 16-byte Reload
; AVX512DQ-BW-NEXT:    vpshufb %xmm4, %xmm26, %xmm19
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm19[0],xmm13[0],xmm19[1],xmm13[1],xmm19[2],xmm13[2],xmm19[3],xmm13[3]
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm19 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX512DQ-BW-NEXT:    vpshufb %xmm19, %xmm18, %xmm30
; AVX512DQ-BW-NEXT:    vmovdqa64 %xmm18, %xmm24
; AVX512DQ-BW-NEXT:    vmovdqa64 {{[-0-9]+}}(%r{{[sb]}}p), %xmm18 # 16-byte Reload
; AVX512DQ-BW-NEXT:    vpshufb %xmm19, %xmm18, %xmm2
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm30[0],xmm2[1],xmm30[1],xmm2[2],xmm30[2],xmm2[3],xmm30[3]
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm13, %ymm0, %ymm13
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3,4],ymm13[5],ymm2[6,7]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]
; AVX512DQ-BW-NEXT:    vmovdqa %xmm10, %xmm6
; AVX512DQ-BW-NEXT:    vpshufb %xmm0, %xmm10, %xmm2
; AVX512DQ-BW-NEXT:    vmovdqa64 %xmm27, %xmm7
; AVX512DQ-BW-NEXT:    vpshufb %xmm0, %xmm27, %xmm13
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm13[0],xmm2[0],xmm13[1],xmm2[1],xmm13[2],xmm2[2],xmm13[3],xmm2[3]
; AVX512DQ-BW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload
; AVX512DQ-BW-NEXT:    vpshufb %xmm3, %xmm10, %xmm13
; AVX512DQ-BW-NEXT:    vmovdqa64 %xmm8, %xmm27
; AVX512DQ-BW-NEXT:    vpshufb %xmm3, %xmm8, %xmm30
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm30[0],xmm13[0],xmm30[1],xmm13[1],xmm30[2],xmm13[2],xmm30[3],xmm13[3]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} xmm2 = xmm13[0,1,2],xmm2[3]
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpsrlq $32, %zmm10, %zmm13
; AVX512DQ-BW-NEXT:    vpmovqb %zmm13, %xmm13
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} xmm2 = xmm13[0,1],xmm2[2,3]
; AVX512DQ-BW-NEXT:    vinserti32x4 $2, %xmm2, %zmm0, %zmm2
; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm2 {%k1}
; AVX512DQ-BW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
; AVX512DQ-BW-NEXT:    vpshufb %xmm0, %xmm14, %xmm1
; AVX512DQ-BW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
; AVX512DQ-BW-NEXT:    vpshufb %xmm0, %xmm11, %xmm13
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm13[0],xmm1[0],xmm13[1],xmm1[1],xmm13[2],xmm1[2],xmm13[3],xmm1[3]
; AVX512DQ-BW-NEXT:    vpshufb %xmm3, %xmm21, %xmm13
; AVX512DQ-BW-NEXT:    vpshufb %xmm3, %xmm22, %xmm30
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm30[0],xmm13[0],xmm30[1],xmm13[1],xmm30[2],xmm13[2],xmm30[3],xmm13[3]
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm1
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm13, %ymm0, %ymm13
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm13[0,1,2,3,4,5,6],ymm1[7]
; AVX512DQ-BW-NEXT:    vpshufb %xmm4, %xmm25, %xmm13
; AVX512DQ-BW-NEXT:    vpshufb %xmm4, %xmm9, %xmm4
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm4[0],xmm13[0],xmm4[1],xmm13[1],xmm4[2],xmm13[2],xmm4[3],xmm13[3]
; AVX512DQ-BW-NEXT:    vpshufb %xmm19, %xmm31, %xmm13
; AVX512DQ-BW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX512DQ-BW-NEXT:    vpshufb %xmm19, %xmm8, %xmm19
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm19[0],xmm13[0],xmm19[1],xmm13[1],xmm19[2],xmm13[2],xmm19[3],xmm13[3]
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm4, %ymm0, %ymm4
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm13, %ymm0, %ymm13
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm4 = ymm13[0,1,2,3,4],ymm4[5],ymm13[6,7]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm4[0,1,2,3,4,5],ymm1[6,7]
; AVX512DQ-BW-NEXT:    vpshufb %xmm0, %xmm5, %xmm4
; AVX512DQ-BW-NEXT:    vpshufb %xmm0, %xmm23, %xmm0
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm4[0],xmm0[1],xmm4[1],xmm0[2],xmm4[2],xmm0[3],xmm4[3]
; AVX512DQ-BW-NEXT:    vpshufb %xmm3, %xmm15, %xmm4
; AVX512DQ-BW-NEXT:    vpshufb %xmm3, %xmm29, %xmm3
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} xmm0 = xmm3[0,1,2],xmm0[3]
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm17, %zmm20
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm17, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpsrlq $32, %zmm17, %zmm3
; AVX512DQ-BW-NEXT:    vpmovqb %zmm3, %xmm3
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} xmm0 = xmm3[0,1],xmm0[2,3]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm0, %zmm2, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm0 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX512DQ-BW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX512DQ-BW-NEXT:    vpshufb %xmm0, %xmm1, %xmm1
; AVX512DQ-BW-NEXT:    vpshufb %xmm0, %xmm28, %xmm2
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm2 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX512DQ-BW-NEXT:    vpshufb %xmm2, %xmm12, %xmm3
; AVX512DQ-BW-NEXT:    vpshufb %xmm2, %xmm16, %xmm4
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm1
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm3[0,1,2,3,4,5,6],ymm1[7]
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm3 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX512DQ-BW-NEXT:    vmovdqa (%rsp), %xmm15 # 16-byte Reload
; AVX512DQ-BW-NEXT:    vpshufb %xmm3, %xmm15, %xmm4
; AVX512DQ-BW-NEXT:    vpshufb %xmm3, %xmm26, %xmm13
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm13[0],xmm4[0],xmm13[1],xmm4[1],xmm13[2],xmm4[2],xmm13[3],xmm4[3]
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm13 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX512DQ-BW-NEXT:    vpshufb %xmm13, %xmm24, %xmm19
; AVX512DQ-BW-NEXT:    vmovdqa64 %xmm24, %xmm26
; AVX512DQ-BW-NEXT:    vmovdqa64 %xmm24, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX512DQ-BW-NEXT:    vpshufb %xmm13, %xmm18, %xmm30
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm19 = xmm30[0],xmm19[0],xmm30[1],xmm19[1],xmm30[2],xmm19[2],xmm30[3],xmm19[3]
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm4, %ymm0, %ymm4
; AVX512DQ-BW-NEXT:    vinserti32x4 $1, %xmm19, %ymm0, %ymm5
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm4 = ymm5[0,1,2,3,4],ymm4[5],ymm5[6,7]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm4[0,1,2,3,4,5],ymm1[6,7]
; AVX512DQ-BW-NEXT:    vpshufb %xmm0, %xmm6, %xmm4
; AVX512DQ-BW-NEXT:    vmovdqa64 %xmm6, %xmm24
; AVX512DQ-BW-NEXT:    vpshufb %xmm0, %xmm7, %xmm5
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX512DQ-BW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX512DQ-BW-NEXT:    vpshufb %xmm2, %xmm5, %xmm5
; AVX512DQ-BW-NEXT:    vpshufb %xmm2, %xmm27, %xmm19
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm19[0],xmm5[0],xmm19[1],xmm5[1],xmm19[2],xmm5[2],xmm19[3],xmm5[3]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} xmm4 = xmm5[0,1,2],xmm4[3]
; AVX512DQ-BW-NEXT:    vpsrlq $40, %zmm10, %zmm5
; AVX512DQ-BW-NEXT:    vpmovqb %zmm5, %xmm5
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} xmm4 = xmm5[0,1],xmm4[2,3]
; AVX512DQ-BW-NEXT:    vinserti32x4 $2, %xmm4, %zmm0, %zmm4
; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm4 {%k1}
; AVX512DQ-BW-NEXT:    vpshufb %xmm0, %xmm14, %xmm1
; AVX512DQ-BW-NEXT:    vpshufb %xmm0, %xmm11, %xmm5
; AVX512DQ-BW-NEXT:    vmovdqa64 %xmm11, %xmm17
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm5[0],xmm1[0],xmm5[1],xmm1[1],xmm5[2],xmm1[2],xmm5[3],xmm1[3]
; AVX512DQ-BW-NEXT:    vpshufb %xmm2, %xmm21, %xmm5
; AVX512DQ-BW-NEXT:    vpshufb %xmm2, %xmm22, %xmm19
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm19[0],xmm5[0],xmm19[1],xmm5[1],xmm19[2],xmm5[2],xmm19[3],xmm5[3]
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm1
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm5, %ymm0, %ymm5
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm5[0,1,2,3,4,5,6],ymm1[7]
; AVX512DQ-BW-NEXT:    vpshufb %xmm3, %xmm25, %xmm5
; AVX512DQ-BW-NEXT:    vpshufb %xmm3, %xmm9, %xmm3
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm5[0],xmm3[1],xmm5[1],xmm3[2],xmm5[2],xmm3[3],xmm5[3]
; AVX512DQ-BW-NEXT:    vmovdqa64 {{[-0-9]+}}(%r{{[sb]}}p), %xmm31 # 16-byte Reload
; AVX512DQ-BW-NEXT:    vpshufb %xmm13, %xmm31, %xmm5
; AVX512DQ-BW-NEXT:    vpshufb %xmm13, %xmm8, %xmm13
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm13[0],xmm5[0],xmm13[1],xmm5[1],xmm13[2],xmm5[2],xmm13[3],xmm5[3]
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm3
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm5, %ymm0, %ymm5
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm3 = ymm5[0,1,2,3,4],ymm3[5],ymm5[6,7]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm3[0,1,2,3,4,5],ymm1[6,7]
; AVX512DQ-BW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload
; AVX512DQ-BW-NEXT:    vpshufb %xmm0, %xmm11, %xmm3
; AVX512DQ-BW-NEXT:    vpshufb %xmm0, %xmm23, %xmm0
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3]
; AVX512DQ-BW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX512DQ-BW-NEXT:    vpshufb %xmm2, %xmm6, %xmm3
; AVX512DQ-BW-NEXT:    vpshufb %xmm2, %xmm29, %xmm2
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} xmm0 = xmm2[0,1,2],xmm0[3]
; AVX512DQ-BW-NEXT:    vpsrlq $40, %zmm20, %zmm2
; AVX512DQ-BW-NEXT:    vpmovqb %zmm2, %xmm2
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} xmm0 = xmm2[0,1],xmm0[2,3]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm0, %zmm4, %zmm0
; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm0 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX512DQ-BW-NEXT:    vmovdqa64 {{[-0-9]+}}(%r{{[sb]}}p), %xmm28 # 16-byte Reload
; AVX512DQ-BW-NEXT:    vpshufb %xmm0, %xmm28, %xmm1
; AVX512DQ-BW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload
; AVX512DQ-BW-NEXT:    vpshufb %xmm0, %xmm8, %xmm2
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm2 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX512DQ-BW-NEXT:    vpshufb %xmm2, %xmm12, %xmm4
; AVX512DQ-BW-NEXT:    vpshufb %xmm2, %xmm16, %xmm5
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm1
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm4, %ymm0, %ymm4
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm4[0,1,2,3,4,5,6],ymm1[7]
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm4 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX512DQ-BW-NEXT:    vpshufb %xmm4, %xmm15, %xmm5
; AVX512DQ-BW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload
; AVX512DQ-BW-NEXT:    vpshufb %xmm4, %xmm15, %xmm13
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm13[0],xmm5[0],xmm13[1],xmm5[1],xmm13[2],xmm5[2],xmm13[3],xmm5[3]
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm13 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX512DQ-BW-NEXT:    vpshufb %xmm13, %xmm26, %xmm19
; AVX512DQ-BW-NEXT:    vpshufb %xmm13, %xmm18, %xmm30
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm19 = xmm30[0],xmm19[0],xmm30[1],xmm19[1],xmm30[2],xmm19[2],xmm30[3],xmm19[3]
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm5, %ymm0, %ymm5
; AVX512DQ-BW-NEXT:    vinserti32x4 $1, %xmm19, %ymm0, %ymm3
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm3 = ymm3[0,1,2,3,4],ymm5[5],ymm3[6,7]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm3[0,1,2,3,4,5],ymm1[6,7]
; AVX512DQ-BW-NEXT:    vpshufb %xmm0, %xmm24, %xmm3
; AVX512DQ-BW-NEXT:    vpshufb %xmm0, %xmm7, %xmm5
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm5[0],xmm3[0],xmm5[1],xmm3[1],xmm5[2],xmm3[2],xmm5[3],xmm3[3]
; AVX512DQ-BW-NEXT:    vmovdqa64 {{[-0-9]+}}(%r{{[sb]}}p), %xmm27 # 16-byte Reload
; AVX512DQ-BW-NEXT:    vpshufb %xmm2, %xmm27, %xmm5
; AVX512DQ-BW-NEXT:    vmovdqa64 {{[-0-9]+}}(%r{{[sb]}}p), %xmm30 # 16-byte Reload
; AVX512DQ-BW-NEXT:    vpshufb %xmm2, %xmm30, %xmm19
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm19[0],xmm5[0],xmm19[1],xmm5[1],xmm19[2],xmm5[2],xmm19[3],xmm5[3]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} xmm3 = xmm5[0,1,2],xmm3[3]
; AVX512DQ-BW-NEXT:    vpsrlq $48, %zmm10, %zmm5
; AVX512DQ-BW-NEXT:    vpmovqb %zmm5, %xmm5
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} xmm3 = xmm5[0,1],xmm3[2,3]
; AVX512DQ-BW-NEXT:    vinserti32x4 $2, %xmm3, %zmm0, %zmm3
; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm3 {%k1}
; AVX512DQ-BW-NEXT:    vpshufb %xmm0, %xmm14, %xmm1
; AVX512DQ-BW-NEXT:    vpshufb %xmm0, %xmm17, %xmm5
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm5[0],xmm1[0],xmm5[1],xmm1[1],xmm5[2],xmm1[2],xmm5[3],xmm1[3]
; AVX512DQ-BW-NEXT:    vpshufb %xmm2, %xmm21, %xmm5
; AVX512DQ-BW-NEXT:    vpshufb %xmm2, %xmm22, %xmm19
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm19[0],xmm5[0],xmm19[1],xmm5[1],xmm19[2],xmm5[2],xmm19[3],xmm5[3]
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm1
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm5, %ymm0, %ymm5
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm5[0,1,2,3,4,5,6],ymm1[7]
; AVX512DQ-BW-NEXT:    vpshufb %xmm4, %xmm25, %xmm5
; AVX512DQ-BW-NEXT:    vpshufb %xmm4, %xmm9, %xmm4
; AVX512DQ-BW-NEXT:    vmovdqa64 %xmm9, %xmm18
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3]
; AVX512DQ-BW-NEXT:    vpshufb %xmm13, %xmm31, %xmm5
; AVX512DQ-BW-NEXT:    vmovdqa64 {{[-0-9]+}}(%r{{[sb]}}p), %xmm20 # 16-byte Reload
; AVX512DQ-BW-NEXT:    vpshufb %xmm13, %xmm20, %xmm13
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm13[0],xmm5[0],xmm13[1],xmm5[1],xmm13[2],xmm5[2],xmm13[3],xmm5[3]
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm4, %ymm0, %ymm4
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm5, %ymm0, %ymm5
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm4 = ymm5[0,1,2,3,4],ymm4[5],ymm5[6,7]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm4[0,1,2,3,4,5],ymm1[6,7]
; AVX512DQ-BW-NEXT:    vpshufb %xmm0, %xmm11, %xmm4
; AVX512DQ-BW-NEXT:    vpshufb %xmm0, %xmm23, %xmm0
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm4[0],xmm0[1],xmm4[1],xmm0[2],xmm4[2],xmm0[3],xmm4[3]
; AVX512DQ-BW-NEXT:    vpshufb %xmm2, %xmm6, %xmm4
; AVX512DQ-BW-NEXT:    vmovdqa64 %xmm6, %xmm26
; AVX512DQ-BW-NEXT:    vpshufb %xmm2, %xmm29, %xmm2
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} xmm0 = xmm2[0,1,2],xmm0[3]
; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vpsrlq $48, %zmm19, %zmm2
; AVX512DQ-BW-NEXT:    vpmovqb %zmm2, %xmm2
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} xmm0 = xmm2[0,1],xmm0[2,3]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm0, %zmm3, %zmm0
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm1 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX512DQ-BW-NEXT:    vpshufb %xmm1, %xmm28, %xmm2
; AVX512DQ-BW-NEXT:    vpshufb %xmm1, %xmm8, %xmm3
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm3 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX512DQ-BW-NEXT:    vpshufb %xmm3, %xmm12, %xmm4
; AVX512DQ-BW-NEXT:    vpshufb %xmm3, %xmm16, %xmm5
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm4, %ymm0, %ymm4
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm2 = ymm4[0,1,2,3,4,5,6],ymm2[7]
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm4 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX512DQ-BW-NEXT:    vmovdqa (%rsp), %xmm5 # 16-byte Reload
; AVX512DQ-BW-NEXT:    vpshufb %xmm4, %xmm5, %xmm5
; AVX512DQ-BW-NEXT:    vpshufb %xmm4, %xmm15, %xmm8
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm8[0],xmm5[0],xmm8[1],xmm5[1],xmm8[2],xmm5[2],xmm8[3],xmm5[3]
; AVX512DQ-BW-NEXT:    vpbroadcastw {{.*#+}} xmm8 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX512DQ-BW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX512DQ-BW-NEXT:    vpshufb %xmm8, %xmm6, %xmm9
; AVX512DQ-BW-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX512DQ-BW-NEXT:    vpshufb %xmm8, %xmm6, %xmm13
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm13[0],xmm9[0],xmm13[1],xmm9[1],xmm13[2],xmm9[2],xmm13[3],xmm9[3]
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm5, %ymm0, %ymm5
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm9, %ymm0, %ymm9
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm5 = ymm9[0,1,2,3,4],ymm5[5],ymm9[6,7]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm2 = ymm5[0,1,2,3,4,5],ymm2[6,7]
; AVX512DQ-BW-NEXT:    vpshufb %xmm1, %xmm24, %xmm5
; AVX512DQ-BW-NEXT:    vpshufb %xmm1, %xmm7, %xmm9
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm9[0],xmm5[0],xmm9[1],xmm5[1],xmm9[2],xmm5[2],xmm9[3],xmm5[3]
; AVX512DQ-BW-NEXT:    vpshufb %xmm3, %xmm27, %xmm9
; AVX512DQ-BW-NEXT:    vpshufb %xmm3, %xmm30, %xmm6
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm6[0],xmm9[0],xmm6[1],xmm9[1],xmm6[2],xmm9[2],xmm6[3],xmm9[3]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} xmm5 = xmm6[0,1,2],xmm5[3]
; AVX512DQ-BW-NEXT:    vpsrlq $56, %zmm10, %zmm6
; AVX512DQ-BW-NEXT:    vpmovqb %zmm6, %xmm6
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} xmm5 = xmm6[0,1],xmm5[2,3]
; AVX512DQ-BW-NEXT:    vinserti32x4 $2, %xmm5, %zmm0, %zmm5
; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm2, %zmm0, %zmm5 {%k1}
; AVX512DQ-BW-NEXT:    vpshufb %xmm1, %xmm14, %xmm2
; AVX512DQ-BW-NEXT:    vpshufb %xmm1, %xmm17, %xmm6
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm6[0],xmm2[0],xmm6[1],xmm2[1],xmm6[2],xmm2[2],xmm6[3],xmm2[3]
; AVX512DQ-BW-NEXT:    vpshufb %xmm3, %xmm21, %xmm6
; AVX512DQ-BW-NEXT:    vpshufb %xmm3, %xmm22, %xmm9
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm9[0],xmm6[0],xmm9[1],xmm6[1],xmm9[2],xmm6[2],xmm9[3],xmm6[3]
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm6, %ymm0, %ymm6
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm2 = ymm6[0,1,2,3,4,5,6],ymm2[7]
; AVX512DQ-BW-NEXT:    vpshufb %xmm4, %xmm25, %xmm6
; AVX512DQ-BW-NEXT:    vpshufb %xmm4, %xmm18, %xmm4
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm4[0],xmm6[0],xmm4[1],xmm6[1],xmm4[2],xmm6[2],xmm4[3],xmm6[3]
; AVX512DQ-BW-NEXT:    vpshufb %xmm8, %xmm31, %xmm6
; AVX512DQ-BW-NEXT:    vpshufb %xmm8, %xmm20, %xmm8
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm6 = xmm8[0],xmm6[0],xmm8[1],xmm6[1],xmm8[2],xmm6[2],xmm8[3],xmm6[3]
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm4, %ymm0, %ymm4
; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm6, %ymm0, %ymm6
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm4 = ymm6[0,1,2,3,4],ymm4[5],ymm6[6,7]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm2 = ymm4[0,1,2,3,4,5],ymm2[6,7]
; AVX512DQ-BW-NEXT:    vpshufb %xmm1, %xmm11, %xmm4
; AVX512DQ-BW-NEXT:    vpshufb %xmm1, %xmm23, %xmm1
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm4[0],xmm1[1],xmm4[1],xmm1[2],xmm4[2],xmm1[3],xmm4[3]
; AVX512DQ-BW-NEXT:    vpshufb %xmm3, %xmm26, %xmm4
; AVX512DQ-BW-NEXT:    vpshufb %xmm3, %xmm29, %xmm3
; AVX512DQ-BW-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} xmm1 = xmm3[0,1,2],xmm1[3]
; AVX512DQ-BW-NEXT:    vpsrlq $56, %zmm19, %zmm3
; AVX512DQ-BW-NEXT:    vpmovqb %zmm3, %xmm3
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} xmm1 = xmm3[0,1],xmm1[2,3]
; AVX512DQ-BW-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]
; AVX512DQ-BW-NEXT:    vinserti64x4 $0, %ymm1, %zmm5, %zmm1
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm2, (%rsi)
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm2, (%rdx)
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm2, (%rcx)
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm2, (%r8)
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm2, (%r9)
; AVX512DQ-BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-BW-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload
; AVX512DQ-BW-NEXT:    vmovaps %zmm2, (%rax)
; AVX512DQ-BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm0, (%rax)
; AVX512DQ-BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm1, (%rax)
; AVX512DQ-BW-NEXT:    addq $744, %rsp # imm = 0x2E8
; AVX512DQ-BW-NEXT:    vzeroupper
; AVX512DQ-BW-NEXT:    retq
;
; AVX512DQ-BW-FCP-LABEL: load_i8_stride8_vf64:
; AVX512DQ-BW-FCP:       # %bb.0:
; AVX512DQ-BW-FCP-NEXT:    subq $328, %rsp # imm = 0x148
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 256(%rdi), %zmm18
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm2 = [0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12]
; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm0 = [0,2,2,3,0,2,4,6]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 480(%rdi), %ymm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpermd %ymm1, %ymm0, %ymm30
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm2, %ymm30, %ymm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqa %ymm2, %ymm11
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm3 = [0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 448(%rdi), %ymm2
; AVX512DQ-BW-FCP-NEXT:    vmovdqu %ymm2, (%rsp) # 32-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpermd %ymm2, %ymm0, %ymm31
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm3, %ymm31, %ymm2
; AVX512DQ-BW-FCP-NEXT:    vmovdqa %ymm3, %ymm8
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5,6],ymm1[7]
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm3 = [0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 416(%rdi), %ymm2
; AVX512DQ-BW-FCP-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpermd %ymm2, %ymm0, %ymm19
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm3, %ymm19, %ymm2
; AVX512DQ-BW-FCP-NEXT:    vmovdqa %ymm3, %ymm9
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm5 = [0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 384(%rdi), %ymm29
; AVX512DQ-BW-FCP-NEXT:    vpermd %ymm29, %ymm0, %ymm14
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm5, %ymm14, %ymm3
; AVX512DQ-BW-FCP-NEXT:    vmovdqa %ymm5, %ymm10
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4],ymm2[5],ymm3[6,7]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm7 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 368(%rdi), %xmm21
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm7, %xmm21, %xmm2
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 352(%rdi), %xmm4
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm7, %xmm4, %xmm3
; AVX512DQ-BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm23 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 336(%rdi), %xmm12
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm23, %xmm12, %xmm5
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 320(%rdi), %xmm28
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm23, %xmm28, %xmm6
; AVX512DQ-BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} xmm2 = xmm5[0,1,2],xmm2[3]
; AVX512DQ-BW-FCP-NEXT:    vpmovqb %zmm18, %xmm5
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} xmm2 = xmm5[0,1],xmm2[2,3]
; AVX512DQ-BW-FCP-NEXT:    vinserti32x4 $2, %xmm2, %zmm0, %zmm20
; AVX512DQ-BW-FCP-NEXT:    movb $-64, %al
; AVX512DQ-BW-FCP-NEXT:    kmovd %eax, %k1
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm20 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 224(%rdi), %ymm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpermd %ymm1, %ymm0, %ymm17
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm11, %ymm17, %ymm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 192(%rdi), %ymm2
; AVX512DQ-BW-FCP-NEXT:    vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpermd %ymm2, %ymm0, %ymm3
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm8, %ymm3, %ymm5
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm5 = ymm5[0,1,2,3,4,5,6],ymm1[7]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 160(%rdi), %ymm1
; AVX512DQ-BW-FCP-NEXT:    vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpermd %ymm1, %ymm0, %ymm2
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm9, %ymm2, %ymm6
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 128(%rdi), %ymm27
; AVX512DQ-BW-FCP-NEXT:    vpermd %ymm27, %ymm0, %ymm16
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm10, %ymm16, %ymm8
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm6 = ymm8[0,1,2,3,4],ymm6[5],ymm8[6,7]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm6[0,1,2,3,4,5],ymm5[6,7]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 112(%rdi), %xmm26
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm7, %xmm26, %xmm8
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 96(%rdi), %xmm24
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm7, %xmm24, %xmm7
; AVX512DQ-BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm10 = xmm7[0],xmm8[0],xmm7[1],xmm8[1],xmm7[2],xmm8[2],xmm7[3],xmm8[3]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 80(%rdi), %xmm22
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm23, %xmm22, %xmm25
; AVX512DQ-BW-FCP-NEXT:    vmovdqa 64(%rdi), %xmm8
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm23, %xmm8, %xmm23
; AVX512DQ-BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm15 = xmm23[0],xmm25[0],xmm23[1],xmm25[1],xmm23[2],xmm25[2],xmm23[3],xmm25[3]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} xmm15 = xmm15[0,1,2],xmm10[3]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 (%rdi), %zmm10
; AVX512DQ-BW-FCP-NEXT:    vpmovqb %zmm10, %xmm13
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} xmm13 = xmm13[0,1],xmm15[2,3]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm13[0,1,2,3],ymm0[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm20, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm9 = [1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm9, %ymm30, %ymm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqa %ymm9, %ymm11
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm9 = [1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm9, %ymm31, %ymm13
; AVX512DQ-BW-FCP-NEXT:    vmovdqa %ymm9, %ymm6
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm13[0,1,2,3,4,5,6],ymm0[7]
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm9 = [1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm9, %ymm19, %ymm13
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm1 = [1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm1, %ymm14, %ymm15
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm13 = ymm15[0,1,2,3,4],ymm13[5],ymm15[6,7]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm13[0,1,2,3,4,5],ymm0[6,7]
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm13 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %xmm21, %xmm5
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm13, %xmm21, %xmm15
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm13, %xmm4, %xmm23
; AVX512DQ-BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm15 = xmm23[0],xmm15[0],xmm23[1],xmm15[1],xmm23[2],xmm15[2],xmm23[3],xmm15[3]
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm23 = [1,9,1,9,1,9,1,9,1,9,1,9,1,9,1,9]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa %xmm12, %xmm7
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm23, %xmm12, %xmm25
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm23, %xmm28, %xmm20
; AVX512DQ-BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm20[0],xmm25[0],xmm20[1],xmm25[1],xmm20[2],xmm25[2],xmm20[3],xmm25[3]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} xmm12 = xmm12[0,1,2],xmm15[3]
; AVX512DQ-BW-FCP-NEXT:    vpsrlq $8, %zmm18, %zmm15
; AVX512DQ-BW-FCP-NEXT:    vpmovqb %zmm15, %xmm15
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} xmm12 = xmm15[0,1],xmm12[2,3]
; AVX512DQ-BW-FCP-NEXT:    vinserti32x4 $2, %xmm12, %zmm0, %zmm12
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm0, %zmm0, %zmm12 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm11, %ymm17, %ymm0
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm6, %ymm3, %ymm15
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm15[0,1,2,3,4,5,6],ymm0[7]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm9, %ymm2, %ymm15
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm1, %ymm16, %ymm11
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm11 = ymm11[0,1,2,3,4],ymm15[5],ymm11[6,7]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm11[0,1,2,3,4,5],ymm0[6,7]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm13, %xmm26, %xmm11
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm13, %xmm24, %xmm13
; AVX512DQ-BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm13[0],xmm11[0],xmm13[1],xmm11[1],xmm13[2],xmm11[2],xmm13[3],xmm11[3]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm23, %xmm22, %xmm13
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm23, %xmm8, %xmm15
; AVX512DQ-BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm13 = xmm15[0],xmm13[0],xmm15[1],xmm13[1],xmm15[2],xmm13[2],xmm15[3],xmm13[3]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} xmm11 = xmm13[0,1,2],xmm11[3]
; AVX512DQ-BW-FCP-NEXT:    vpsrlq $8, %zmm10, %zmm13
; AVX512DQ-BW-FCP-NEXT:    vpmovqb %zmm13, %xmm13
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} xmm11 = xmm13[0,1],xmm11[2,3]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm11[0,1,2,3],ymm0[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm12, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm9 = [2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm9, %ymm30, %ymm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %ymm9, %ymm21
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm6 = [2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm6, %ymm31, %ymm11
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %ymm6, %ymm25
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm11[0,1,2,3,4,5,6],ymm0[7]
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm6 = [2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm6, %ymm19, %ymm11
; AVX512DQ-BW-FCP-NEXT:    vmovdqa %ymm6, %ymm9
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm6 = [2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm6, %ymm14, %ymm12
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm11 = ymm12[0,1,2,3,4],ymm11[5],ymm12[6,7]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm11[0,1,2,3,4,5],ymm0[6,7]
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm11 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm11, %xmm5, %xmm12
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm11, %xmm4, %xmm13
; AVX512DQ-BW-FCP-NEXT:    vmovdqa %xmm4, %xmm1
; AVX512DQ-BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm13[0],xmm12[0],xmm13[1],xmm12[1],xmm13[2],xmm12[2],xmm13[3],xmm12[3]
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm13 = [2,10,2,10,2,10,2,10,2,10,2,10,2,10,2,10]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm13, %xmm7, %xmm15
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %xmm7, %xmm23
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm13, %xmm28, %xmm20
; AVX512DQ-BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm15 = xmm20[0],xmm15[0],xmm20[1],xmm15[1],xmm20[2],xmm15[2],xmm20[3],xmm15[3]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} xmm12 = xmm15[0,1,2],xmm12[3]
; AVX512DQ-BW-FCP-NEXT:    vpsrlq $16, %zmm18, %zmm15
; AVX512DQ-BW-FCP-NEXT:    vpmovqb %zmm15, %xmm15
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} xmm12 = xmm15[0,1],xmm12[2,3]
; AVX512DQ-BW-FCP-NEXT:    vinserti32x4 $2, %xmm12, %zmm0, %zmm12
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm0, %zmm0, %zmm12 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm21, %ymm17, %ymm0
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm25, %ymm3, %ymm15
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm15[0,1,2,3,4,5,6],ymm0[7]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm9, %ymm2, %ymm15
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm6, %ymm16, %ymm9
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm9 = ymm9[0,1,2,3,4],ymm15[5],ymm9[6,7]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm9[0,1,2,3,4,5],ymm0[6,7]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm11, %xmm26, %xmm9
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm11, %xmm24, %xmm11
; AVX512DQ-BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm11[0],xmm9[0],xmm11[1],xmm9[1],xmm11[2],xmm9[2],xmm11[3],xmm9[3]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm13, %xmm22, %xmm11
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm13, %xmm8, %xmm13
; AVX512DQ-BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm13[0],xmm11[0],xmm13[1],xmm11[1],xmm13[2],xmm11[2],xmm13[3],xmm11[3]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} xmm9 = xmm11[0,1,2],xmm9[3]
; AVX512DQ-BW-FCP-NEXT:    vpsrlq $16, %zmm10, %zmm11
; AVX512DQ-BW-FCP-NEXT:    vpmovqb %zmm11, %xmm11
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} xmm9 = xmm11[0,1],xmm9[2,3]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm9[0,1,2,3],ymm0[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm12, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm4 = [3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm4, %ymm30, %ymm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqa %ymm4, %ymm6
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm4 = [3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm4, %ymm31, %ymm9
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm9[0,1,2,3,4,5,6],ymm0[7]
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm31 = [3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm31, %ymm19, %ymm9
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm30 = [3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm30, %ymm14, %ymm11
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm9 = ymm11[0,1,2,3,4],ymm9[5],ymm11[6,7]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm9[0,1,2,3,4,5],ymm0[6,7]
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm9 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %xmm5, %xmm19
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm9, %xmm5, %xmm11
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %xmm1, %xmm20
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm9, %xmm1, %xmm12
; AVX512DQ-BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm12[0],xmm11[0],xmm12[1],xmm11[1],xmm12[2],xmm11[2],xmm12[3],xmm11[3]
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm12 = [3,11,3,11,3,11,3,11,3,11,3,11,3,11,3,11]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm12, %xmm7, %xmm14
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm12, %xmm28, %xmm15
; AVX512DQ-BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm14 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} xmm11 = xmm14[0,1,2],xmm11[3]
; AVX512DQ-BW-FCP-NEXT:    vpsrlq $24, %zmm18, %zmm14
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm18, %zmm25
; AVX512DQ-BW-FCP-NEXT:    vpmovqb %zmm14, %xmm14
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} xmm11 = xmm14[0,1],xmm11[2,3]
; AVX512DQ-BW-FCP-NEXT:    vinserti32x4 $2, %xmm11, %zmm0, %zmm11
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm0, %zmm0, %zmm11 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm6, %ymm17, %ymm0
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm4, %ymm3, %ymm3
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm3[0,1,2,3,4,5,6],ymm0[7]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm31, %ymm2, %ymm2
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm30, %ymm16, %ymm1
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3,4],ymm2[5],ymm1[6,7]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm9, %xmm26, %xmm1
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm9, %xmm24, %xmm2
; AVX512DQ-BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm12, %xmm22, %xmm2
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm12, %xmm8, %xmm3
; AVX512DQ-BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} xmm1 = xmm2[0,1,2],xmm1[3]
; AVX512DQ-BW-FCP-NEXT:    vpsrlq $24, %zmm10, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vpmovqb %zmm2, %xmm2
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm11, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} ymm3 = [1,3,2,3,1,3,5,7]
; AVX512DQ-BW-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm3, %ymm6 # 32-byte Folded Reload
; AVX512DQ-BW-FCP-NEXT:    vpermd (%rsp), %ymm3, %ymm7 # 32-byte Folded Reload
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm13 = [0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm13, %ymm6, %ymm2
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm1 = [0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm1, %ymm7, %ymm9
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm9 = ymm9[0,1,2,3,4,5,6],ymm2[7]
; AVX512DQ-BW-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm3, %ymm5 # 32-byte Folded Reload
; AVX512DQ-BW-FCP-NEXT:    vpermd %ymm29, %ymm3, %ymm14
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm0 = [0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm0, %ymm5, %ymm11
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm4 = [0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12,0,4,8,12]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm4, %ymm14, %ymm12
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm11 = ymm12[0,1,2,3,4],ymm11[5],ymm12[6,7]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm9 = ymm11[0,1,2,3,4,5],ymm9[6,7]
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm11 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm11, %xmm19, %xmm12
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %xmm19, %xmm16
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm11, %xmm20, %xmm15
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %xmm20, %xmm17
; AVX512DQ-BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm15[0],xmm12[0],xmm15[1],xmm12[1],xmm15[2],xmm12[2],xmm15[3],xmm12[3]
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm20 = [4,12,4,12,4,12,4,12,4,12,4,12,4,12,4,12]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm20, %xmm23, %xmm15
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm20, %xmm28, %xmm29
; AVX512DQ-BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm15 = xmm29[0],xmm15[0],xmm29[1],xmm15[1],xmm29[2],xmm15[2],xmm29[3],xmm15[3]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} xmm12 = xmm15[0,1,2],xmm12[3]
; AVX512DQ-BW-FCP-NEXT:    vpsrlq $32, %zmm18, %zmm15
; AVX512DQ-BW-FCP-NEXT:    vpmovqb %zmm15, %xmm15
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} xmm12 = xmm15[0,1],xmm12[2,3]
; AVX512DQ-BW-FCP-NEXT:    vinserti32x4 $2, %xmm12, %zmm0, %zmm12
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm9, %zmm0, %zmm12 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm3, %ymm19 # 32-byte Folded Reload
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm13, %ymm19, %ymm9
; AVX512DQ-BW-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm3, %ymm18 # 32-byte Folded Reload
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm1, %ymm18, %ymm15
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm9 = ymm15[0,1,2,3,4,5,6],ymm9[7]
; AVX512DQ-BW-FCP-NEXT:    vpermd {{[-0-9]+}}(%r{{[sb]}}p), %ymm3, %ymm15 # 32-byte Folded Reload
; AVX512DQ-BW-FCP-NEXT:    vpermd %ymm27, %ymm3, %ymm21
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm0, %ymm15, %ymm3
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm4, %ymm21, %ymm13
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm3 = ymm13[0,1,2,3,4],ymm3[5],ymm13[6,7]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm9[6,7]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm11, %xmm26, %xmm9
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm11, %xmm24, %xmm11
; AVX512DQ-BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm11[0],xmm9[0],xmm11[1],xmm9[1],xmm11[2],xmm9[2],xmm11[3],xmm9[3]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm20, %xmm22, %xmm11
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm20, %xmm8, %xmm13
; AVX512DQ-BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm13[0],xmm11[0],xmm13[1],xmm11[1],xmm13[2],xmm11[2],xmm13[3],xmm11[3]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} xmm9 = xmm11[0,1,2],xmm9[3]
; AVX512DQ-BW-FCP-NEXT:    vpsrlq $32, %zmm10, %zmm11
; AVX512DQ-BW-FCP-NEXT:    vpmovqb %zmm11, %xmm11
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} xmm9 = xmm11[0,1],xmm9[2,3]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm3 = ymm9[0,1,2,3],ymm3[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm3, %zmm12, %zmm29
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm4 = [1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm4, %ymm6, %ymm9
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm0 = [1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm0, %ymm7, %ymm11
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm9 = ymm11[0,1,2,3,4,5,6],ymm9[7]
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm2 = [1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm2, %ymm5, %ymm11
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm1 = [1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13,1,5,9,13]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm1, %ymm14, %ymm12
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm11 = ymm12[0,1,2,3,4],ymm11[5],ymm12[6,7]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm9 = ymm11[0,1,2,3,4,5],ymm9[6,7]
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm11 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm11, %xmm16, %xmm12
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm11, %xmm17, %xmm13
; AVX512DQ-BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm13[0],xmm12[0],xmm13[1],xmm12[1],xmm13[2],xmm12[2],xmm13[3],xmm12[3]
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm13 = [5,13,5,13,5,13,5,13,5,13,5,13,5,13,5,13]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm13, %xmm23, %xmm20
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm13, %xmm28, %xmm27
; AVX512DQ-BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm3 = xmm27[0],xmm20[0],xmm27[1],xmm20[1],xmm27[2],xmm20[2],xmm27[3],xmm20[3]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} xmm3 = xmm3[0,1,2],xmm12[3]
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm25, %zmm27
; AVX512DQ-BW-FCP-NEXT:    vpsrlq $40, %zmm25, %zmm12
; AVX512DQ-BW-FCP-NEXT:    vpmovqb %zmm12, %xmm12
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} xmm3 = xmm12[0,1],xmm3[2,3]
; AVX512DQ-BW-FCP-NEXT:    vinserti32x4 $2, %xmm3, %zmm0, %zmm3
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm9, %zmm0, %zmm3 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm4, %ymm19, %ymm9
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm0, %ymm18, %ymm12
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm9 = ymm12[0,1,2,3,4,5,6],ymm9[7]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm2, %ymm15, %ymm12
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm1, %ymm21, %ymm4
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3,4],ymm12[5],ymm4[6,7]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm9[6,7]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm11, %xmm26, %xmm9
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm11, %xmm24, %xmm11
; AVX512DQ-BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm11[0],xmm9[0],xmm11[1],xmm9[1],xmm11[2],xmm9[2],xmm11[3],xmm9[3]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm13, %xmm22, %xmm11
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm13, %xmm8, %xmm12
; AVX512DQ-BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm12[0],xmm11[0],xmm12[1],xmm11[1],xmm12[2],xmm11[2],xmm12[3],xmm11[3]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} xmm9 = xmm11[0,1,2],xmm9[3]
; AVX512DQ-BW-FCP-NEXT:    vpsrlq $40, %zmm10, %zmm11
; AVX512DQ-BW-FCP-NEXT:    vpmovqb %zmm11, %xmm11
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} xmm9 = xmm11[0,1],xmm9[2,3]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm4 = ymm9[0,1,2,3],ymm4[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm4, %zmm3, %zmm20
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm0 = [2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm0, %ymm6, %ymm3
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm1 = [2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm1, %ymm7, %ymm4
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm3 = ymm4[0,1,2,3,4,5,6],ymm3[7]
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm2 = [2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm2, %ymm5, %ymm4
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm25 = [2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14,2,6,10,14]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm25, %ymm14, %ymm9
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm4 = ymm9[0,1,2,3,4],ymm4[5],ymm9[6,7]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm3 = ymm4[0,1,2,3,4,5],ymm3[6,7]
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm4 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm4, %xmm16, %xmm9
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm4, %xmm17, %xmm11
; AVX512DQ-BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm11[0],xmm9[0],xmm11[1],xmm9[1],xmm11[2],xmm9[2],xmm11[3],xmm9[3]
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm11 = [6,14,6,14,6,14,6,14,6,14,6,14,6,14,6,14]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm11, %xmm23, %xmm12
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm11, %xmm28, %xmm13
; AVX512DQ-BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm12 = xmm13[0],xmm12[0],xmm13[1],xmm12[1],xmm13[2],xmm12[2],xmm13[3],xmm12[3]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} xmm9 = xmm12[0,1,2],xmm9[3]
; AVX512DQ-BW-FCP-NEXT:    vpsrlq $48, %zmm27, %zmm12
; AVX512DQ-BW-FCP-NEXT:    vpmovqb %zmm12, %xmm12
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} xmm9 = xmm12[0,1],xmm9[2,3]
; AVX512DQ-BW-FCP-NEXT:    vinserti32x4 $2, %xmm9, %zmm0, %zmm9
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm3, %zmm0, %zmm9 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm0, %ymm19, %ymm3
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm1, %ymm18, %ymm12
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm3 = ymm12[0,1,2,3,4,5,6],ymm3[7]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm2, %ymm15, %ymm12
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm25, %ymm21, %ymm13
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm12 = ymm13[0,1,2,3,4],ymm12[5],ymm13[6,7]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm3 = ymm12[0,1,2,3,4,5],ymm3[6,7]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm4, %xmm26, %xmm12
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm4, %xmm24, %xmm4
; AVX512DQ-BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm4[0],xmm12[0],xmm4[1],xmm12[1],xmm4[2],xmm12[2],xmm4[3],xmm12[3]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm11, %xmm22, %xmm12
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm11, %xmm8, %xmm11
; AVX512DQ-BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm11 = xmm11[0],xmm12[0],xmm11[1],xmm12[1],xmm11[2],xmm12[2],xmm11[3],xmm12[3]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} xmm4 = xmm11[0,1,2],xmm4[3]
; AVX512DQ-BW-FCP-NEXT:    vpsrlq $48, %zmm10, %zmm11
; AVX512DQ-BW-FCP-NEXT:    vpmovqb %zmm11, %xmm11
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} xmm4 = xmm11[0,1],xmm4[2,3]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm3 = ymm4[0,1,2,3],ymm3[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm3, %zmm9, %zmm3
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm12 = [3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm12, %ymm6, %ymm0
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastd {{.*#+}} ymm6 = [3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15,3,7,11,15]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm6, %ymm7, %ymm1
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5,6],ymm0[7]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm31, %ymm5, %ymm1
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm30, %ymm14, %ymm2
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3,4],ymm1[5],ymm2[6,7]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm1 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm1, %xmm16, %xmm2
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm1, %xmm17, %xmm4
; AVX512DQ-BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm2 = xmm4[0],xmm2[0],xmm4[1],xmm2[1],xmm4[2],xmm2[2],xmm4[3],xmm2[3]
; AVX512DQ-BW-FCP-NEXT:    vpbroadcastw {{.*#+}} xmm4 = [7,15,7,15,7,15,7,15,7,15,7,15,7,15,7,15]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm4, %xmm23, %xmm9
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm4, %xmm28, %xmm11
; AVX512DQ-BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm9 = xmm11[0],xmm9[0],xmm11[1],xmm9[1],xmm11[2],xmm9[2],xmm11[3],xmm9[3]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} xmm2 = xmm9[0,1,2],xmm2[3]
; AVX512DQ-BW-FCP-NEXT:    vpsrlq $56, %zmm27, %zmm9
; AVX512DQ-BW-FCP-NEXT:    vpmovqb %zmm9, %xmm9
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} xmm2 = xmm9[0,1],xmm2[2,3]
; AVX512DQ-BW-FCP-NEXT:    vinserti32x4 $2, %xmm2, %zmm0, %zmm2
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm0, %zmm0, %zmm2 {%k1}
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm12, %ymm19, %ymm0
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm6, %ymm18, %ymm9
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm9[0,1,2,3,4,5,6],ymm0[7]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm31, %ymm15, %ymm9
; AVX512DQ-BW-FCP-NEXT:    vpshufb %ymm30, %ymm21, %ymm11
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm9 = ymm11[0,1,2,3,4],ymm9[5],ymm11[6,7]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm9[0,1,2,3,4,5],ymm0[6,7]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm1, %xmm26, %xmm5
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm1, %xmm24, %xmm1
; AVX512DQ-BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm1 = xmm1[0],xmm5[0],xmm1[1],xmm5[1],xmm1[2],xmm5[2],xmm1[3],xmm5[3]
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm4, %xmm22, %xmm5
; AVX512DQ-BW-FCP-NEXT:    vpshufb %xmm4, %xmm8, %xmm4
; AVX512DQ-BW-FCP-NEXT:    vpunpcklwd {{.*#+}} xmm4 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} xmm1 = xmm4[0,1,2],xmm1[3]
; AVX512DQ-BW-FCP-NEXT:    vpsrlq $56, %zmm10, %zmm4
; AVX512DQ-BW-FCP-NEXT:    vpmovqb %zmm4, %xmm4
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} xmm1 = xmm4[0,1],xmm1[2,3]
; AVX512DQ-BW-FCP-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $0, %ymm0, %zmm2, %zmm0
; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm1, (%rsi)
; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm1, (%rdx)
; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm1, (%rcx)
; AVX512DQ-BW-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload
; AVX512DQ-BW-FCP-NEXT:    vmovaps %zmm1, (%r8)
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm29, (%r9)
; AVX512DQ-BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm20, (%rax)
; AVX512DQ-BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm3, (%rax)
; AVX512DQ-BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax
; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm0, (%rax)
; AVX512DQ-BW-FCP-NEXT:    addq $328, %rsp # imm = 0x148
; AVX512DQ-BW-FCP-NEXT:    vzeroupper
; AVX512DQ-BW-FCP-NEXT:    retq
  %wide.vec = load <512 x i8>, ptr %in.vec, align 64
  %strided.vec0 = shufflevector <512 x i8> %wide.vec, <512 x i8> poison, <64 x i32> <i32 0, i32 8, i32 16, i32 24, i32 32, i32 40, i32 48, i32 56, i32 64, i32 72, i32 80, i32 88, i32 96, i32 104, i32 112, i32 120, i32 128, i32 136, i32 144, i32 152, i32 160, i32 168, i32 176, i32 184, i32 192, i32 200, i32 208, i32 216, i32 224, i32 232, i32 240, i32 248, i32 256, i32 264, i32 272, i32 280, i32 288, i32 296, i32 304, i32 312, i32 320, i32 328, i32 336, i32 344, i32 352, i32 360, i32 368, i32 376, i32 384, i32 392, i32 400, i32 408, i32 416, i32 424, i32 432, i32 440, i32 448, i32 456, i32 464, i32 472, i32 480, i32 488, i32 496, i32 504>
  %strided.vec1 = shufflevector <512 x i8> %wide.vec, <512 x i8> poison, <64 x i32> <i32 1, i32 9, i32 17, i32 25, i32 33, i32 41, i32 49, i32 57, i32 65, i32 73, i32 81, i32 89, i32 97, i32 105, i32 113, i32 121, i32 129, i32 137, i32 145, i32 153, i32 161, i32 169, i32 177, i32 185, i32 193, i32 201, i32 209, i32 217, i32 225, i32 233, i32 241, i32 249, i32 257, i32 265, i32 273, i32 281, i32 289, i32 297, i32 305, i32 313, i32 321, i32 329, i32 337, i32 345, i32 353, i32 361, i32 369, i32 377, i32 385, i32 393, i32 401, i32 409, i32 417, i32 425, i32 433, i32 441, i32 449, i32 457, i32 465, i32 473, i32 481, i32 489, i32 497, i32 505>
  %strided.vec2 = shufflevector <512 x i8> %wide.vec, <512 x i8> poison, <64 x i32> <i32 2, i32 10, i32 18, i32 26, i32 34, i32 42, i32 50, i32 58, i32 66, i32 74, i32 82, i32 90, i32 98, i32 106, i32 114, i32 122, i32 130, i32 138, i32 146, i32 154, i32 162, i32 170, i32 178, i32 186, i32 194, i32 202, i32 210, i32 218, i32 226, i32 234, i32 242, i32 250, i32 258, i32 266, i32 274, i32 282, i32 290, i32 298, i32 306, i32 314, i32 322, i32 330, i32 338, i32 346, i32 354, i32 362, i32 370, i32 378, i32 386, i32 394, i32 402, i32 410, i32 418, i32 426, i32 434, i32 442, i32 450, i32 458, i32 466, i32 474, i32 482, i32 490, i32 498, i32 506>
  %strided.vec3 = shufflevector <512 x i8> %wide.vec, <512 x i8> poison, <64 x i32> <i32 3, i32 11, i32 19, i32 27, i32 35, i32 43, i32 51, i32 59, i32 67, i32 75, i32 83, i32 91, i32 99, i32 107, i32 115, i32 123, i32 131, i32 139, i32 147, i32 155, i32 163, i32 171, i32 179, i32 187, i32 195, i32 203, i32 211, i32 219, i32 227, i32 235, i32 243, i32 251, i32 259, i32 267, i32 275, i32 283, i32 291, i32 299, i32 307, i32 315, i32 323, i32 331, i32 339, i32 347, i32 355, i32 363, i32 371, i32 379, i32 387, i32 395, i32 403, i32 411, i32 419, i32 427, i32 435, i32 443, i32 451, i32 459, i32 467, i32 475, i32 483, i32 491, i32 499, i32 507>
  %strided.vec4 = shufflevector <512 x i8> %wide.vec, <512 x i8> poison, <64 x i32> <i32 4, i32 12, i32 20, i32 28, i32 36, i32 44, i32 52, i32 60, i32 68, i32 76, i32 84, i32 92, i32 100, i32 108, i32 116, i32 124, i32 132, i32 140, i32 148, i32 156, i32 164, i32 172, i32 180, i32 188, i32 196, i32 204, i32 212, i32 220, i32 228, i32 236, i32 244, i32 252, i32 260, i32 268, i32 276, i32 284, i32 292, i32 300, i32 308, i32 316, i32 324, i32 332, i32 340, i32 348, i32 356, i32 364, i32 372, i32 380, i32 388, i32 396, i32 404, i32 412, i32 420, i32 428, i32 436, i32 444, i32 452, i32 460, i32 468, i32 476, i32 484, i32 492, i32 500, i32 508>
  %strided.vec5 = shufflevector <512 x i8> %wide.vec, <512 x i8> poison, <64 x i32> <i32 5, i32 13, i32 21, i32 29, i32 37, i32 45, i32 53, i32 61, i32 69, i32 77, i32 85, i32 93, i32 101, i32 109, i32 117, i32 125, i32 133, i32 141, i32 149, i32 157, i32 165, i32 173, i32 181, i32 189, i32 197, i32 205, i32 213, i32 221, i32 229, i32 237, i32 245, i32 253, i32 261, i32 269, i32 277, i32 285, i32 293, i32 301, i32 309, i32 317, i32 325, i32 333, i32 341, i32 349, i32 357, i32 365, i32 373, i32 381, i32 389, i32 397, i32 405, i32 413, i32 421, i32 429, i32 437, i32 445, i32 453, i32 461, i32 469, i32 477, i32 485, i32 493, i32 501, i32 509>
  %strided.vec6 = shufflevector <512 x i8> %wide.vec, <512 x i8> poison, <64 x i32> <i32 6, i32 14, i32 22, i32 30, i32 38, i32 46, i32 54, i32 62, i32 70, i32 78, i32 86, i32 94, i32 102, i32 110, i32 118, i32 126, i32 134, i32 142, i32 150, i32 158, i32 166, i32 174, i32 182, i32 190, i32 198, i32 206, i32 214, i32 222, i32 230, i32 238, i32 246, i32 254, i32 262, i32 270, i32 278, i32 286, i32 294, i32 302, i32 310, i32 318, i32 326, i32 334, i32 342, i32 350, i32 358, i32 366, i32 374, i32 382, i32 390, i32 398, i32 406, i32 414, i32 422, i32 430, i32 438, i32 446, i32 454, i32 462, i32 470, i32 478, i32 486, i32 494, i32 502, i32 510>
  %strided.vec7 = shufflevector <512 x i8> %wide.vec, <512 x i8> poison, <64 x i32> <i32 7, i32 15, i32 23, i32 31, i32 39, i32 47, i32 55, i32 63, i32 71, i32 79, i32 87, i32 95, i32 103, i32 111, i32 119, i32 127, i32 135, i32 143, i32 151, i32 159, i32 167, i32 175, i32 183, i32 191, i32 199, i32 207, i32 215, i32 223, i32 231, i32 239, i32 247, i32 255, i32 263, i32 271, i32 279, i32 287, i32 295, i32 303, i32 311, i32 319, i32 327, i32 335, i32 343, i32 351, i32 359, i32 367, i32 375, i32 383, i32 391, i32 399, i32 407, i32 415, i32 423, i32 431, i32 439, i32 447, i32 455, i32 463, i32 471, i32 479, i32 487, i32 495, i32 503, i32 511>
  store <64 x i8> %strided.vec0, ptr %out.vec0, align 64
  store <64 x i8> %strided.vec1, ptr %out.vec1, align 64
  store <64 x i8> %strided.vec2, ptr %out.vec2, align 64
  store <64 x i8> %strided.vec3, ptr %out.vec3, align 64
  store <64 x i8> %strided.vec4, ptr %out.vec4, align 64
  store <64 x i8> %strided.vec5, ptr %out.vec5, align 64
  store <64 x i8> %strided.vec6, ptr %out.vec6, align 64
  store <64 x i8> %strided.vec7, ptr %out.vec7, align 64
  ret void
}