Google Git
Sign in
llvm / llvm-project / llvm / e7e1a7ad6f753bcb646a0a80f8ac71809ba0829d / . / test / CodeGen / X86 / vector-interleaved-store-i8-stride-6.ll
blob: bb052c6fa70d946a3911f8b0c37271457166ebc7 [file] [log] [blame]
; 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,FALLBACK0
; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx | FileCheck %s --check-prefixes=AVX,AVX1,AVX1-ONLY,FALLBACK1
; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx2 | FileCheck %s --check-prefixes=AVX,AVX1,AVX2,AVX2-ONLY,AVX2-SLOW,FALLBACK2
; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx2,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX,AVX1,AVX2,AVX2-ONLY,AVX2-FAST,FALLBACK3
; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx2,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX,AVX1,AVX2,AVX2-ONLY,AVX2-FAST-PERLANE,FALLBACK4
; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl | FileCheck %s --check-prefixes=AVX,AVX2,AVX512,AVX512F,AVX512F-SLOW,AVX512F-ONLY-SLOW,FALLBACK5
; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX,AVX2,AVX512,AVX512F,AVX512F-FAST,AVX512F-ONLY-FAST,FALLBACK6
; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+avx512dq | FileCheck %s --check-prefixes=AVX,AVX2,AVX512,AVX512F,AVX512F-SLOW,AVX512DQ-SLOW,FALLBACK7
; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+avx512dq,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX,AVX2,AVX512,AVX512F,AVX512F-FAST,AVX512DQ-FAST,FALLBACK8
; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+avx512bw | FileCheck %s --check-prefixes=AVX,AVX2,AVX512,AVX512BW,AVX512BW-SLOW,AVX512BW-ONLY-SLOW,FALLBACK9
; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+avx512bw,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX,AVX2,AVX512,AVX512BW,AVX512BW-FAST,AVX512BW-ONLY-FAST,FALLBACK10
; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+avx512dq,+avx512bw | FileCheck %s --check-prefixes=AVX,AVX2,AVX512,AVX512BW,AVX512BW-SLOW,AVX512DQBW-SLOW,FALLBACK11
; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+avx512dq,+avx512bw,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX,AVX2,AVX512,AVX512BW,AVX512BW-FAST,AVX512DQBW-FAST,FALLBACK12
; These patterns are produced by LoopVectorizer for interleaved stores.
define void @store_i8_stride6_vf2(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecptr2, ptr %in.vecptr3, ptr %in.vecptr4, ptr %in.vecptr5, ptr %out.vec) nounwind {
; SSE-LABEL: store_i8_stride6_vf2:
; SSE: # %bb.0:
; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT: movdqa (%rdi), %xmm0
; SSE-NEXT: movdqa (%rdx), %xmm1
; SSE-NEXT: movdqa (%r8), %xmm2
; SSE-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
; SSE-NEXT: punpcklwd {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1],xmm1[2],mem[2],xmm1[3],mem[3]
; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
; SSE-NEXT: pxor %xmm1, %xmm1
; SSE-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3],xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
; SSE-NEXT: pshufhw {{.*#+}} xmm1 = xmm0[0,1,2,3,5,7,6,7]
; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[2,2,2,2]
; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[0,2,1,3,4,5,6,7]
; SSE-NEXT: pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,6,6,7]
; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,2,2,1]
; SSE-NEXT: packuswb %xmm1, %xmm0
; SSE-NEXT: movdqa {{.*#+}} xmm1 = [65535,65535,0,65535,65535,0,65535,65535]
; SSE-NEXT: pand %xmm1, %xmm0
; SSE-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm2[0,0,0,0]
; SSE-NEXT: pandn %xmm2, %xmm1
; SSE-NEXT: por %xmm0, %xmm1
; SSE-NEXT: movdqa {{.*#+}} xmm0 = [255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255]
; SSE-NEXT: pand %xmm0, %xmm1
; SSE-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0],mem[0],xmm2[1],mem[1],xmm2[2],mem[2],xmm2[3],mem[3],xmm2[4],mem[4],xmm2[5],mem[5],xmm2[6],mem[6],xmm2[7],mem[7]
; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm2[0,0,0,0]
; SSE-NEXT: pandn %xmm2, %xmm0
; SSE-NEXT: por %xmm1, %xmm0
; SSE-NEXT: movq %xmm0, (%rax)
; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[2,3,2,3]
; SSE-NEXT: movd %xmm0, 8(%rax)
; SSE-NEXT: retq
;
; AVX-LABEL: store_i8_stride6_vf2:
; AVX: # %bb.0:
; AVX-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX-NEXT: vmovdqa (%rdi), %xmm0
; AVX-NEXT: vmovdqa (%rdx), %xmm1
; AVX-NEXT: vmovdqa (%r9), %xmm2
; AVX-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
; AVX-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1],xmm1[2],mem[2],xmm1[3],mem[3]
; AVX-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
; AVX-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3],xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7]
; AVX-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,4,8,12,1],zero,xmm0[2,6,10,14,3],zero,xmm0[u,u,u,u]
; AVX-NEXT: vpshufb {{.*#+}} xmm1 = zero,zero,zero,zero,zero,xmm2[0],zero,zero,zero,zero,zero,xmm2[1,u,u,u,u]
; AVX-NEXT: vpor %xmm1, %xmm0, %xmm0
; AVX-NEXT: vpextrd $2, %xmm0, 8(%rax)
; AVX-NEXT: vmovq %xmm0, (%rax)
; AVX-NEXT: retq
%in.vec0 = load <2 x i8>, ptr %in.vecptr0, align 64
%in.vec1 = load <2 x i8>, ptr %in.vecptr1, align 64
%in.vec2 = load <2 x i8>, ptr %in.vecptr2, align 64
%in.vec3 = load <2 x i8>, ptr %in.vecptr3, align 64
%in.vec4 = load <2 x i8>, ptr %in.vecptr4, align 64
%in.vec5 = load <2 x i8>, ptr %in.vecptr5, align 64
%1 = shufflevector <2 x i8> %in.vec0, <2 x i8> %in.vec1, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
%2 = shufflevector <2 x i8> %in.vec2, <2 x i8> %in.vec3, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
%3 = shufflevector <2 x i8> %in.vec4, <2 x i8> %in.vec5, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
%4 = shufflevector <4 x i8> %1, <4 x i8> %2, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
%5 = shufflevector <4 x i8> %3, <4 x i8> poison, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 undef, i32 undef, i32 undef, i32 undef>
%6 = shufflevector <8 x i8> %4, <8 x i8> %5, <12 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11>
%interleaved.vec = shufflevector <12 x i8> %6, <12 x i8> poison, <12 x i32> <i32 0, i32 2, i32 4, i32 6, i32 8, i32 10, i32 1, i32 3, i32 5, i32 7, i32 9, i32 11>
store <12 x i8> %interleaved.vec, ptr %out.vec, align 64
ret void
}
define void @store_i8_stride6_vf4(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecptr2, ptr %in.vecptr3, ptr %in.vecptr4, ptr %in.vecptr5, ptr %out.vec) nounwind {
; SSE-LABEL: store_i8_stride6_vf4:
; SSE: # %bb.0:
; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT: movdqa (%rdi), %xmm1
; SSE-NEXT: movdqa (%rdx), %xmm2
; SSE-NEXT: movdqa (%r8), %xmm0
; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]
; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],mem[0],xmm2[1],mem[1]
; SSE-NEXT: pxor %xmm3, %xmm3
; 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: pshufd {{.*#+}} xmm4 = xmm2[0,2,2,3]
; SSE-NEXT: pshuflw {{.*#+}} xmm4 = xmm4[0,1,0,2,4,5,6,7]
; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm2[0,2,1,3]
; SSE-NEXT: pshuflw {{.*#+}} xmm5 = xmm5[1,3,2,3,4,5,6,7]
; SSE-NEXT: pshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,4,5,4,6]
; SSE-NEXT: packuswb %xmm5, %xmm4
; SSE-NEXT: movdqa {{.*#+}} xmm5 = [65535,0,65535,65535,0,65535,65535,0]
; SSE-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1],xmm1[2],xmm3[2],xmm1[3],xmm3[3],xmm1[4],xmm3[4],xmm1[5],xmm3[5],xmm1[6],xmm3[6],xmm1[7],xmm3[7]
; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm1[0,1,1,3]
; SSE-NEXT: pshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,6,6,7]
; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm1[0,2,2,0]
; SSE-NEXT: pshuflw {{.*#+}} xmm6 = xmm6[0,2,2,3,4,5,6,7]
; SSE-NEXT: pshufhw {{.*#+}} xmm6 = xmm6[0,1,2,3,4,5,7,5]
; SSE-NEXT: packuswb %xmm3, %xmm6
; SSE-NEXT: pand %xmm5, %xmm6
; SSE-NEXT: pandn %xmm4, %xmm5
; SSE-NEXT: por %xmm6, %xmm5
; SSE-NEXT: movdqa {{.*#+}} xmm3 = [65535,65535,0,65535,65535,0,65535,65535]
; SSE-NEXT: pand %xmm3, %xmm5
; SSE-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3],xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7]
; SSE-NEXT: pshufd {{.*#+}} xmm4 = xmm0[0,0,0,0]
; SSE-NEXT: pandn %xmm4, %xmm3
; SSE-NEXT: por %xmm5, %xmm3
; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm2[0,1,1,3]
; SSE-NEXT: pshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,5,7,6,7]
; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[3,1,2,3]
; SSE-NEXT: pshuflw {{.*#+}} xmm1 = xmm1[0,1,3,1,4,5,6,7]
; SSE-NEXT: packuswb %xmm2, %xmm1
; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,3,2,3]
; SSE-NEXT: movdqa {{.*#+}} xmm2 = [65535,0,0,65535,65535,65535,65535,65535]
; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,1,1]
; SSE-NEXT: pand %xmm2, %xmm0
; SSE-NEXT: pandn %xmm1, %xmm2
; SSE-NEXT: por %xmm0, %xmm2
; SSE-NEXT: movq %xmm2, 16(%rax)
; SSE-NEXT: movdqa %xmm3, (%rax)
; SSE-NEXT: retq
;
; AVX1-ONLY-LABEL: store_i8_stride6_vf4:
; AVX1-ONLY: # %bb.0:
; AVX1-ONLY-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX1-ONLY-NEXT: vmovdqa (%rdi), %xmm0
; AVX1-ONLY-NEXT: vmovdqa (%rdx), %xmm1
; AVX1-ONLY-NEXT: vmovdqa (%r8), %xmm2
; AVX1-ONLY-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
; AVX1-ONLY-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]
; AVX1-ONLY-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
; AVX1-ONLY-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm2[0],mem[0],xmm2[1],mem[1]
; AVX1-ONLY-NEXT: vpshufb {{.*#+}} xmm2 = xmm0[0,4,8,12],zero,zero,xmm0[1,5,9,13],zero,zero,xmm0[2,6,10,14]
; AVX1-ONLY-NEXT: vpshufb {{.*#+}} xmm3 = zero,zero,zero,zero,xmm1[0,4],zero,zero,zero,zero,xmm1[1,5],zero,zero,zero,zero
; AVX1-ONLY-NEXT: vpor %xmm3, %xmm2, %xmm2
; AVX1-ONLY-NEXT: vpshufb {{.*#+}} xmm0 = zero,zero,xmm0[3,7,11,15],zero,zero,xmm0[u,u,u,u,u,u,u,u]
; AVX1-ONLY-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[2,6],zero,zero,zero,zero,xmm1[3,7,u,u,u,u,u,u,u,u]
; AVX1-ONLY-NEXT: vpor %xmm1, %xmm0, %xmm0
; AVX1-ONLY-NEXT: vmovq %xmm0, 16(%rax)
; AVX1-ONLY-NEXT: vmovdqa %xmm2, (%rax)
; AVX1-ONLY-NEXT: retq
;
; AVX2-ONLY-LABEL: store_i8_stride6_vf4:
; AVX2-ONLY: # %bb.0:
; AVX2-ONLY-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX2-ONLY-NEXT: vmovdqa (%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovdqa (%rdx), %xmm1
; AVX2-ONLY-NEXT: vmovdqa (%r8), %xmm2
; AVX2-ONLY-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
; AVX2-ONLY-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]
; AVX2-ONLY-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
; AVX2-ONLY-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm2[0],mem[0],xmm2[1],mem[1]
; AVX2-ONLY-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm0
; AVX2-ONLY-NEXT: vpshufb {{.*#+}} ymm1 = ymm0[0,4,8,12],zero,zero,ymm0[1,5,9,13],zero,zero,ymm0[2,6,10,14,18,22],zero,zero,zero,zero,ymm0[19,23],zero,zero,zero,zero,zero,zero,zero,zero
; AVX2-ONLY-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,0,1]
; AVX2-ONLY-NEXT: vpshufb {{.*#+}} ymm0 = zero,zero,zero,zero,ymm0[0,4],zero,zero,zero,zero,ymm0[1,5],zero,zero,zero,zero,zero,zero,ymm0[19,23,27,31],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; AVX2-ONLY-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX2-ONLY-NEXT: vextracti128 $1, %ymm0, %xmm1
; AVX2-ONLY-NEXT: vmovq %xmm1, 16(%rax)
; AVX2-ONLY-NEXT: vmovdqa %xmm0, (%rax)
; AVX2-ONLY-NEXT: vzeroupper
; AVX2-ONLY-NEXT: retq
;
; AVX512F-LABEL: store_i8_stride6_vf4:
; AVX512F: # %bb.0:
; AVX512F-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512F-NEXT: vmovdqa (%rdi), %xmm0
; AVX512F-NEXT: vmovdqa (%rdx), %xmm1
; AVX512F-NEXT: vmovdqa (%r8), %xmm2
; AVX512F-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
; AVX512F-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]
; AVX512F-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
; AVX512F-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm2[0],mem[0],xmm2[1],mem[1]
; AVX512F-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm0
; AVX512F-NEXT: vpshufb {{.*#+}} ymm1 = ymm0[0,4,8,12],zero,zero,ymm0[1,5,9,13],zero,zero,ymm0[2,6,10,14,18,22],zero,zero,zero,zero,ymm0[19,23,u,u,u,u,u,u,u,u]
; AVX512F-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,0,1]
; AVX512F-NEXT: vpshufb {{.*#+}} ymm0 = zero,zero,zero,zero,ymm0[0,4],zero,zero,zero,zero,ymm0[1,5],zero,zero,zero,zero,zero,zero,ymm0[19,23,27,31],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; AVX512F-NEXT: vpor %ymm0, %ymm1, %ymm0
; AVX512F-NEXT: vextracti128 $1, %ymm0, %xmm1
; AVX512F-NEXT: vmovq %xmm1, 16(%rax)
; AVX512F-NEXT: vmovdqa %xmm0, (%rax)
; AVX512F-NEXT: vzeroupper
; AVX512F-NEXT: retq
;
; AVX512BW-LABEL: store_i8_stride6_vf4:
; AVX512BW: # %bb.0:
; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-NEXT: vmovdqa (%rdi), %xmm0
; AVX512BW-NEXT: vmovdqa (%rdx), %xmm1
; AVX512BW-NEXT: vmovdqa (%r8), %xmm2
; AVX512BW-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
; AVX512BW-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]
; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
; AVX512BW-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm2[0],mem[0],xmm2[1],mem[1]
; AVX512BW-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm0
; AVX512BW-NEXT: vpshufb {{.*#+}} ymm1 = ymm0[0,4,8,12,u,u,1,5,9,13,u,u,2,6,10,14,18,22,u,u,u,u,19,23,u,u,u,u,u,u,u,u]
; AVX512BW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,0,1]
; AVX512BW-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[u,u,u,u,0,4,u,u,u,u,1,5,u,u,u,u,u,u,19,23,27,31,u,u,u,u,u,u,u,u,u,u]
; AVX512BW-NEXT: movw $1572, %cx # imm = 0x624
; AVX512BW-NEXT: kmovd %ecx, %k1
; AVX512BW-NEXT: vmovdqu16 %ymm0, %ymm1 {%k1}
; AVX512BW-NEXT: vextracti128 $1, %ymm1, %xmm0
; AVX512BW-NEXT: vmovq %xmm0, 16(%rax)
; AVX512BW-NEXT: vmovdqa %xmm1, (%rax)
; AVX512BW-NEXT: vzeroupper
; AVX512BW-NEXT: retq
%in.vec0 = load <4 x i8>, ptr %in.vecptr0, align 64
%in.vec1 = load <4 x i8>, ptr %in.vecptr1, align 64
%in.vec2 = load <4 x i8>, ptr %in.vecptr2, align 64
%in.vec3 = load <4 x i8>, ptr %in.vecptr3, align 64
%in.vec4 = load <4 x i8>, ptr %in.vecptr4, align 64
%in.vec5 = load <4 x i8>, ptr %in.vecptr5, align 64
%1 = shufflevector <4 x i8> %in.vec0, <4 x i8> %in.vec1, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
%2 = shufflevector <4 x i8> %in.vec2, <4 x i8> %in.vec3, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
%3 = shufflevector <4 x i8> %in.vec4, <4 x i8> %in.vec5, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
%4 = shufflevector <8 x i8> %1, <8 x i8> %2, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
%5 = shufflevector <8 x i8> %3, <8 x i8> poison, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
%6 = shufflevector <16 x i8> %4, <16 x i8> %5, <24 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23>
%interleaved.vec = shufflevector <24 x i8> %6, <24 x i8> poison, <24 x i32> <i32 0, i32 4, i32 8, i32 12, i32 16, i32 20, i32 1, i32 5, i32 9, i32 13, i32 17, i32 21, i32 2, i32 6, i32 10, i32 14, i32 18, i32 22, i32 3, i32 7, i32 11, i32 15, i32 19, i32 23>
store <24 x i8> %interleaved.vec, ptr %out.vec, align 64
ret void
}
define void @store_i8_stride6_vf8(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecptr2, ptr %in.vecptr3, ptr %in.vecptr4, ptr %in.vecptr5, ptr %out.vec) nounwind {
; SSE-LABEL: store_i8_stride6_vf8:
; SSE: # %bb.0:
; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT: movq {{.*#+}} xmm0 = mem[0],zero
; SSE-NEXT: movq {{.*#+}} xmm1 = mem[0],zero
; SSE-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3],xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
; SSE-NEXT: movq {{.*#+}} xmm1 = mem[0],zero
; SSE-NEXT: movq {{.*#+}} xmm2 = mem[0],zero
; SSE-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3],xmm1[4],xmm2[4],xmm1[5],xmm2[5],xmm1[6],xmm2[6],xmm1[7],xmm2[7]
; SSE-NEXT: movq {{.*#+}} xmm2 = mem[0],zero
; SSE-NEXT: movq {{.*#+}} xmm4 = mem[0],zero
; SSE-NEXT: pxor %xmm5, %xmm5
; SSE-NEXT: movdqa %xmm2, %xmm3
; SSE-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3],xmm2[4],xmm4[4],xmm2[5],xmm4[5],xmm2[6],xmm4[6],xmm2[7],xmm4[7]
; 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: 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, %xmm5
; SSE-NEXT: punpcklwd {{.*#+}} xmm5 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]
; SSE-NEXT: packuswb %xmm5, %xmm5
; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm5[0,0,2,3]
; SSE-NEXT: movdqa {{.*#+}} xmm6 = [65535,65535,0,65535,65535,0,65535,65535]
; SSE-NEXT: movdqa %xmm6, %xmm7
; SSE-NEXT: pandn %xmm5, %xmm7
; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm0[0,0,1,1]
; SSE-NEXT: movdqa {{.*#+}} xmm8 = [65535,0,65535,65535,0,65535,65535,0]
; SSE-NEXT: pand %xmm8, %xmm5
; SSE-NEXT: pshuflw {{.*#+}} xmm9 = xmm1[1,0,2,2,4,5,6,7]
; SSE-NEXT: pshufd {{.*#+}} xmm9 = xmm9[0,1,0,1]
; SSE-NEXT: movdqa %xmm8, %xmm10
; SSE-NEXT: pandn %xmm9, %xmm10
; SSE-NEXT: por %xmm5, %xmm10
; SSE-NEXT: pand %xmm6, %xmm10
; SSE-NEXT: por %xmm7, %xmm10
; SSE-NEXT: punpckhwd {{.*#+}} xmm3 = xmm3[4],xmm4[4],xmm3[5],xmm4[5],xmm3[6],xmm4[6],xmm3[7],xmm4[7]
; SSE-NEXT: packuswb %xmm3, %xmm3
; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm3[0,1,3,3]
; SSE-NEXT: pshufd {{.*#+}} xmm4 = xmm0[3,3,3,3]
; SSE-NEXT: movdqa %xmm6, %xmm5
; SSE-NEXT: pandn %xmm4, %xmm5
; SSE-NEXT: pshufhw {{.*#+}} xmm4 = xmm1[0,1,2,3,5,6,7,7]
; SSE-NEXT: pshufd {{.*#+}} xmm4 = xmm4[2,2,2,3]
; SSE-NEXT: pand %xmm6, %xmm4
; SSE-NEXT: por %xmm5, %xmm4
; SSE-NEXT: pand %xmm8, %xmm4
; SSE-NEXT: pandn %xmm3, %xmm8
; SSE-NEXT: por %xmm4, %xmm8
; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,2,2]
; SSE-NEXT: pand %xmm6, %xmm0
; SSE-NEXT: pshuflw {{.*#+}} xmm1 = xmm1[3,3,3,3,4,5,6,7]
; SSE-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,4,4,4]
; SSE-NEXT: pandn %xmm1, %xmm6
; SSE-NEXT: por %xmm0, %xmm6
; SSE-NEXT: movdqa {{.*#+}} xmm0 = [0,65535,65535,0,65535,65535,0,65535]
; SSE-NEXT: pand %xmm0, %xmm6
; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm2[1,1,2,2]
; SSE-NEXT: pandn %xmm1, %xmm0
; SSE-NEXT: por %xmm6, %xmm0
; SSE-NEXT: movdqa %xmm8, 32(%rax)
; SSE-NEXT: movdqa %xmm0, 16(%rax)
; SSE-NEXT: movdqa %xmm10, (%rax)
; SSE-NEXT: retq
;
; AVX1-ONLY-LABEL: store_i8_stride6_vf8:
; AVX1-ONLY: # %bb.0:
; AVX1-ONLY-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX1-ONLY-NEXT: vmovq {{.*#+}} xmm0 = mem[0],zero
; AVX1-ONLY-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero
; AVX1-ONLY-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0]
; AVX1-ONLY-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero
; AVX1-ONLY-NEXT: vmovq {{.*#+}} xmm2 = mem[0],zero
; AVX1-ONLY-NEXT: vpunpcklqdq {{.*#+}} xmm1 = xmm2[0],xmm1[0]
; AVX1-ONLY-NEXT: vmovq {{.*#+}} xmm2 = mem[0],zero
; AVX1-ONLY-NEXT: vmovq {{.*#+}} xmm3 = mem[0],zero
; AVX1-ONLY-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm3[0],xmm2[0]
; AVX1-ONLY-NEXT: vpshufb {{.*#+}} xmm3 = xmm1[u,u],zero,zero,xmm1[3,11,u,u],zero,zero,xmm1[4,12,u,u],zero,zero
; AVX1-ONLY-NEXT: vpshufb {{.*#+}} xmm4 = xmm0[u,u,3,11],zero,zero,xmm0[u,u,4,12],zero,zero,xmm0[u,u,5,13]
; AVX1-ONLY-NEXT: vpor %xmm3, %xmm4, %xmm3
; AVX1-ONLY-NEXT: vpshufb {{.*#+}} xmm4 = xmm2[2,10,u,u,u,u,3,11,u,u,u,u,4,12,u,u]
; AVX1-ONLY-NEXT: vpblendw {{.*#+}} xmm3 = xmm4[0],xmm3[1,2],xmm4[3],xmm3[4,5],xmm4[6],xmm3[7]
; AVX1-ONLY-NEXT: vpshufb {{.*#+}} xmm4 = zero,zero,xmm1[0,8,u,u],zero,zero,xmm1[1,9,u,u],zero,zero,xmm1[2,10]
; AVX1-ONLY-NEXT: vpshufb {{.*#+}} xmm5 = xmm0[0,8],zero,zero,xmm0[u,u,1,9],zero,zero,xmm0[u,u,2,10],zero,zero
; AVX1-ONLY-NEXT: vpor %xmm4, %xmm5, %xmm4
; AVX1-ONLY-NEXT: vpshufb {{.*#+}} xmm5 = xmm2[u,u,u,u,0,8,u,u,u,u,1,9,u,u,u,u]
; AVX1-ONLY-NEXT: vpblendw {{.*#+}} xmm4 = xmm4[0,1],xmm5[2],xmm4[3,4],xmm5[5],xmm4[6,7]
; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm3, %ymm4, %ymm3
; AVX1-ONLY-NEXT: vpshufb {{.*#+}} xmm0 = zero,zero,xmm0[u,u,6,14],zero,zero,xmm0[u,u,7,15],zero,zero,xmm0[u,u]
; AVX1-ONLY-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[5,13,u,u],zero,zero,xmm1[6,14,u,u],zero,zero,xmm1[7,15,u,u]
; AVX1-ONLY-NEXT: vpor %xmm0, %xmm1, %xmm0
; AVX1-ONLY-NEXT: vpshufb {{.*#+}} xmm1 = xmm2[u,u,5,13,u,u,u,u,6,14,u,u,u,u,7,15]
; AVX1-ONLY-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0],xmm1[1],xmm0[2,3],xmm1[4],xmm0[5,6],xmm1[7]
; AVX1-ONLY-NEXT: vmovdqa %xmm0, 32(%rax)
; AVX1-ONLY-NEXT: vmovaps %ymm3, (%rax)
; AVX1-ONLY-NEXT: vzeroupper
; AVX1-ONLY-NEXT: retq
;
; AVX2-ONLY-LABEL: store_i8_stride6_vf8:
; AVX2-ONLY: # %bb.0:
; AVX2-ONLY-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX2-ONLY-NEXT: vmovq {{.*#+}} xmm0 = mem[0],zero
; AVX2-ONLY-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero
; AVX2-ONLY-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0]
; AVX2-ONLY-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero
; AVX2-ONLY-NEXT: vmovq {{.*#+}} xmm2 = mem[0],zero
; AVX2-ONLY-NEXT: vpunpcklqdq {{.*#+}} xmm1 = xmm2[0],xmm1[0]
; AVX2-ONLY-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm2
; AVX2-ONLY-NEXT: vmovq {{.*#+}} xmm3 = mem[0],zero
; AVX2-ONLY-NEXT: vmovq {{.*#+}} xmm4 = mem[0],zero
; AVX2-ONLY-NEXT: vpunpcklqdq {{.*#+}} xmm3 = xmm4[0],xmm3[0]
; AVX2-ONLY-NEXT: vpshufb {{.*#+}} ymm4 = ymm2[0,8,u,u,u,u,1,9,u,u,u,u,2,10,u,u,u,u,u,u,19,27,u,u,u,u,20,28,u,u,u,u]
; AVX2-ONLY-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,3,0,1]
; AVX2-ONLY-NEXT: vpshufb {{.*#+}} ymm2 = ymm2[u,u,0,8,u,u,u,u,1,9,u,u,u,u,2,10,u,u,19,27,u,u,u,u,20,28,u,u,u,u,21,29]
; AVX2-ONLY-NEXT: vpblendw {{.*#+}} ymm2 = ymm4[0],ymm2[1],ymm4[2,3],ymm2[4],ymm4[5,6],ymm2[7],ymm4[8],ymm2[9],ymm4[10,11],ymm2[12],ymm4[13,14],ymm2[15]
; AVX2-ONLY-NEXT: vpshufb {{.*#+}} xmm4 = xmm3[2,10,1,9,0,8,3,11,u,u,u,u,4,12,u,u]
; AVX2-ONLY-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,0,1]
; AVX2-ONLY-NEXT: vmovdqa {{.*#+}} ymm5 = [255,255,255,255,0,0,255,255,255,255,0,0,255,255,255,255,0,0,255,255,255,255,0,0,255,255,255,255,0,0,255,255]
; AVX2-ONLY-NEXT: vpblendvb %ymm5, %ymm2, %ymm4, %ymm2
; AVX2-ONLY-NEXT: vpshufb {{.*#+}} xmm0 = zero,zero,xmm0[u,u,6,14],zero,zero,xmm0[u,u,7,15],zero,zero,xmm0[u,u]
; AVX2-ONLY-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[5,13,u,u],zero,zero,xmm1[6,14,u,u],zero,zero,xmm1[7,15,u,u]
; AVX2-ONLY-NEXT: vpor %xmm0, %xmm1, %xmm0
; AVX2-ONLY-NEXT: vpshufb {{.*#+}} xmm1 = xmm3[u,u,5,13,u,u,u,u,6,14,u,u,u,u,7,15]
; AVX2-ONLY-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0],xmm1[1],xmm0[2,3],xmm1[4],xmm0[5,6],xmm1[7]
; AVX2-ONLY-NEXT: vmovdqa %xmm0, 32(%rax)
; AVX2-ONLY-NEXT: vmovdqa %ymm2, (%rax)
; AVX2-ONLY-NEXT: vzeroupper
; AVX2-ONLY-NEXT: retq
;
; AVX512F-LABEL: store_i8_stride6_vf8:
; AVX512F: # %bb.0:
; AVX512F-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512F-NEXT: vmovq {{.*#+}} xmm0 = mem[0],zero
; AVX512F-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero
; AVX512F-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0]
; AVX512F-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero
; AVX512F-NEXT: vmovq {{.*#+}} xmm2 = mem[0],zero
; AVX512F-NEXT: vpunpcklqdq {{.*#+}} xmm1 = xmm2[0],xmm1[0]
; AVX512F-NEXT: vmovq {{.*#+}} xmm2 = mem[0],zero
; AVX512F-NEXT: vmovq {{.*#+}} xmm3 = mem[0],zero
; AVX512F-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm3[0],xmm2[0]
; AVX512F-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm3
; AVX512F-NEXT: vpshufb {{.*#+}} ymm4 = ymm3[0,8,u,u,u,u,1,9,u,u,u,u,2,10,u,u,u,u,u,u,19,27,u,u,u,u,20,28,u,u,u,u]
; AVX512F-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,3,0,1]
; AVX512F-NEXT: vpshufb {{.*#+}} ymm3 = ymm3[u,u,0,8,u,u,u,u,1,9,u,u,u,u,2,10,u,u,19,27,u,u,u,u,20,28,u,u,u,u,21,29]
; AVX512F-NEXT: vpblendw {{.*#+}} ymm3 = ymm4[0],ymm3[1],ymm4[2,3],ymm3[4],ymm4[5,6],ymm3[7],ymm4[8],ymm3[9],ymm4[10,11],ymm3[12],ymm4[13,14],ymm3[15]
; AVX512F-NEXT: vpshufb {{.*#+}} xmm4 = xmm2[2,10,1,9,0,8,3,11,u,u,u,u,4,12,u,u]
; AVX512F-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,0,1]
; AVX512F-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm3, %ymm4
; AVX512F-NEXT: vpshufb {{.*#+}} xmm0 = zero,zero,xmm0[u,u,6,14],zero,zero,xmm0[u,u,7,15],zero,zero,xmm0[u,u]
; AVX512F-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[5,13,u,u],zero,zero,xmm1[6,14,u,u],zero,zero,xmm1[7,15,u,u]
; AVX512F-NEXT: vpor %xmm0, %xmm1, %xmm0
; AVX512F-NEXT: vpshufb {{.*#+}} xmm1 = xmm2[u,u,5,13,u,u,u,u,6,14,u,u,u,u,7,15]
; AVX512F-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0],xmm1[1],xmm0[2,3],xmm1[4],xmm0[5,6],xmm1[7]
; AVX512F-NEXT: vinserti32x4 $2, %xmm0, %zmm4, %zmm1
; AVX512F-NEXT: vmovdqa %xmm0, 32(%rax)
; AVX512F-NEXT: vmovdqa %ymm1, (%rax)
; AVX512F-NEXT: vzeroupper
; AVX512F-NEXT: retq
;
; AVX512BW-LABEL: store_i8_stride6_vf8:
; AVX512BW: # %bb.0:
; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-NEXT: vmovq {{.*#+}} xmm0 = mem[0],zero
; AVX512BW-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero
; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0]
; AVX512BW-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero
; AVX512BW-NEXT: vmovq {{.*#+}} xmm2 = mem[0],zero
; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} xmm1 = xmm2[0],xmm1[0]
; AVX512BW-NEXT: vmovq {{.*#+}} xmm2 = mem[0],zero
; AVX512BW-NEXT: vmovq {{.*#+}} xmm3 = mem[0],zero
; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm3[0],xmm2[0]
; AVX512BW-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm3
; AVX512BW-NEXT: vpshufb {{.*#+}} ymm4 = ymm3[0,8,u,u,u,u,1,9,u,u,u,u,2,10,u,u,u,u,u,u,19,27,u,u,u,u,20,28,u,u,u,u]
; AVX512BW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,3,0,1]
; AVX512BW-NEXT: vpshufb {{.*#+}} ymm3 = ymm3[u,u,0,8,u,u,u,u,1,9,u,u,u,u,2,10,u,u,19,27,u,u,u,u,20,28,u,u,u,u,21,29]
; AVX512BW-NEXT: vpblendw {{.*#+}} ymm3 = ymm4[0],ymm3[1],ymm4[2,3],ymm3[4],ymm4[5,6],ymm3[7],ymm4[8],ymm3[9],ymm4[10,11],ymm3[12],ymm4[13,14],ymm3[15]
; AVX512BW-NEXT: vpshufb {{.*#+}} xmm4 = xmm2[2,10,1,9,0,8,3,11,u,u,u,u,4,12,u,u]
; AVX512BW-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,0,1]
; AVX512BW-NEXT: movw $18724, %cx # imm = 0x4924
; AVX512BW-NEXT: kmovd %ecx, %k1
; AVX512BW-NEXT: vmovdqu16 %ymm4, %ymm3 {%k1}
; AVX512BW-NEXT: vpshufb {{.*#+}} xmm0 = zero,zero,xmm0[u,u,6,14],zero,zero,xmm0[u,u,7,15],zero,zero,xmm0[u,u]
; AVX512BW-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[5,13,u,u],zero,zero,xmm1[6,14,u,u],zero,zero,xmm1[7,15,u,u]
; AVX512BW-NEXT: vpor %xmm0, %xmm1, %xmm0
; AVX512BW-NEXT: vpshufb {{.*#+}} xmm1 = xmm2[u,u,5,13,u,u,u,u,6,14,u,u,u,u,7,15]
; AVX512BW-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0],xmm1[1],xmm0[2,3],xmm1[4],xmm0[5,6],xmm1[7]
; AVX512BW-NEXT: vinserti32x4 $2, %xmm0, %zmm3, %zmm1
; AVX512BW-NEXT: vmovdqa %xmm0, 32(%rax)
; AVX512BW-NEXT: vmovdqa %ymm1, (%rax)
; AVX512BW-NEXT: vzeroupper
; AVX512BW-NEXT: retq
%in.vec0 = load <8 x i8>, ptr %in.vecptr0, align 64
%in.vec1 = load <8 x i8>, ptr %in.vecptr1, align 64
%in.vec2 = load <8 x i8>, ptr %in.vecptr2, align 64
%in.vec3 = load <8 x i8>, ptr %in.vecptr3, align 64
%in.vec4 = load <8 x i8>, ptr %in.vecptr4, align 64
%in.vec5 = load <8 x i8>, ptr %in.vecptr5, align 64
%1 = shufflevector <8 x i8> %in.vec0, <8 x i8> %in.vec1, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
%2 = shufflevector <8 x i8> %in.vec2, <8 x i8> %in.vec3, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
%3 = shufflevector <8 x i8> %in.vec4, <8 x i8> %in.vec5, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
%4 = shufflevector <16 x i8> %1, <16 x i8> %2, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
%5 = shufflevector <16 x i8> %3, <16 x i8> poison, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
%6 = shufflevector <32 x i8> %4, <32 x i8> %5, <48 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47>
%interleaved.vec = shufflevector <48 x i8> %6, <48 x i8> poison, <48 x i32> <i32 0, i32 8, i32 16, i32 24, i32 32, i32 40, i32 1, i32 9, i32 17, i32 25, i32 33, i32 41, i32 2, i32 10, i32 18, i32 26, i32 34, i32 42, i32 3, i32 11, i32 19, i32 27, i32 35, i32 43, i32 4, i32 12, i32 20, i32 28, i32 36, i32 44, i32 5, i32 13, i32 21, i32 29, i32 37, i32 45, i32 6, i32 14, i32 22, i32 30, i32 38, i32 46, i32 7, i32 15, i32 23, i32 31, i32 39, i32 47>
store <48 x i8> %interleaved.vec, ptr %out.vec, align 64
ret void
}
define void @store_i8_stride6_vf16(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecptr2, ptr %in.vecptr3, ptr %in.vecptr4, ptr %in.vecptr5, ptr %out.vec) nounwind {
; SSE-LABEL: store_i8_stride6_vf16:
; SSE: # %bb.0:
; SSE-NEXT: movdqa (%rdi), %xmm11
; SSE-NEXT: movdqa (%rsi), %xmm8
; SSE-NEXT: movdqa (%rdx), %xmm13
; SSE-NEXT: movdqa (%rcx), %xmm2
; SSE-NEXT: movdqa (%r8), %xmm9
; SSE-NEXT: movdqa (%r9), %xmm12
; SSE-NEXT: movdqa %xmm11, %xmm0
; SSE-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm8[0],xmm0[1],xmm8[1],xmm0[2],xmm8[2],xmm0[3],xmm8[3],xmm0[4],xmm8[4],xmm0[5],xmm8[5],xmm0[6],xmm8[6],xmm0[7],xmm8[7]
; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,0,1,1]
; SSE-NEXT: movdqa {{.*#+}} xmm4 = [65535,0,65535,65535,0,65535,65535,0]
; SSE-NEXT: pand %xmm4, %xmm0
; SSE-NEXT: movdqa %xmm13, %xmm7
; SSE-NEXT: punpcklbw {{.*#+}} xmm7 = xmm7[0],xmm2[0],xmm7[1],xmm2[1],xmm7[2],xmm2[2],xmm7[3],xmm2[3],xmm7[4],xmm2[4],xmm7[5],xmm2[5],xmm7[6],xmm2[6],xmm7[7],xmm2[7]
; SSE-NEXT: pshuflw {{.*#+}} xmm1 = xmm7[1,0,2,2,4,5,6,7]
; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,1,0,1]
; SSE-NEXT: movdqa %xmm4, %xmm3
; SSE-NEXT: pandn %xmm1, %xmm3
; SSE-NEXT: por %xmm0, %xmm3
; SSE-NEXT: movdqa {{.*#+}} xmm1 = [65535,65535,0,65535,65535,0,65535,65535]
; SSE-NEXT: pand %xmm1, %xmm3
; SSE-NEXT: movdqa %xmm9, %xmm6
; SSE-NEXT: punpcklbw {{.*#+}} xmm6 = xmm6[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm6[0,0,0,0]
; SSE-NEXT: movdqa %xmm1, %xmm10
; SSE-NEXT: pandn %xmm0, %xmm10
; SSE-NEXT: por %xmm3, %xmm10
; SSE-NEXT: movdqa {{.*#+}} xmm3 = [255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255]
; SSE-NEXT: pand %xmm3, %xmm10
; SSE-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm12[0],xmm0[1],xmm12[1],xmm0[2],xmm12[2],xmm0[3],xmm12[3],xmm0[4],xmm12[4],xmm0[5],xmm12[5],xmm0[6],xmm12[6],xmm0[7],xmm12[7]
; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT: pshufd {{.*#+}} xmm14 = xmm0[0,0,0,0]
; SSE-NEXT: movdqa %xmm3, %xmm0
; SSE-NEXT: pandn %xmm14, %xmm0
; SSE-NEXT: por %xmm10, %xmm0
; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT: punpckhbw {{.*#+}} xmm11 = xmm11[8],xmm8[8],xmm11[9],xmm8[9],xmm11[10],xmm8[10],xmm11[11],xmm8[11],xmm11[12],xmm8[12],xmm11[13],xmm8[13],xmm11[14],xmm8[14],xmm11[15],xmm8[15]
; SSE-NEXT: pshufd {{.*#+}} xmm8 = xmm11[1,1,2,2]
; SSE-NEXT: pand %xmm1, %xmm8
; SSE-NEXT: punpckhbw {{.*#+}} xmm13 = xmm13[8],xmm2[8],xmm13[9],xmm2[9],xmm13[10],xmm2[10],xmm13[11],xmm2[11],xmm13[12],xmm2[12],xmm13[13],xmm2[13],xmm13[14],xmm2[14],xmm13[15],xmm2[15]
; SSE-NEXT: pshuflw {{.*#+}} xmm2 = xmm13[3,3,3,3,4,5,6,7]
; SSE-NEXT: pshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,4,4,4]
; SSE-NEXT: movdqa %xmm1, %xmm10
; SSE-NEXT: pandn %xmm2, %xmm10
; SSE-NEXT: por %xmm8, %xmm10
; SSE-NEXT: movdqa {{.*#+}} xmm8 = [0,65535,65535,0,65535,65535,0,65535]
; SSE-NEXT: pand %xmm8, %xmm10
; SSE-NEXT: punpckhbw {{.*#+}} xmm9 = xmm9[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm9[1,1,2,2]
; SSE-NEXT: movdqa %xmm8, %xmm14
; SSE-NEXT: pandn %xmm2, %xmm14
; SSE-NEXT: por %xmm10, %xmm14
; SSE-NEXT: movdqa {{.*#+}} xmm10 = [255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255]
; SSE-NEXT: punpckhbw {{.*#+}} xmm2 = xmm2[8],xmm12[8],xmm2[9],xmm12[9],xmm2[10],xmm12[10],xmm2[11],xmm12[11],xmm2[12],xmm12[12],xmm2[13],xmm12[13],xmm2[14],xmm12[14],xmm2[15],xmm12[15]
; SSE-NEXT: pshufd {{.*#+}} xmm15 = xmm2[1,1,2,2]
; SSE-NEXT: movdqa %xmm10, %xmm12
; SSE-NEXT: pandn %xmm15, %xmm12
; SSE-NEXT: pand %xmm10, %xmm14
; SSE-NEXT: por %xmm14, %xmm12
; SSE-NEXT: pshufd {{.*#+}} xmm14 = xmm11[3,3,3,3]
; SSE-NEXT: movdqa %xmm1, %xmm15
; SSE-NEXT: pandn %xmm14, %xmm15
; SSE-NEXT: pshufhw {{.*#+}} xmm14 = xmm13[0,1,2,3,5,6,7,7]
; SSE-NEXT: pshufd {{.*#+}} xmm14 = xmm14[2,2,2,3]
; SSE-NEXT: pand %xmm1, %xmm14
; SSE-NEXT: por %xmm15, %xmm14
; SSE-NEXT: pshufd {{.*#+}} xmm15 = xmm9[2,2,3,3]
; SSE-NEXT: movdqa %xmm4, %xmm5
; SSE-NEXT: pandn %xmm15, %xmm5
; SSE-NEXT: pand %xmm4, %xmm14
; SSE-NEXT: por %xmm14, %xmm5
; SSE-NEXT: movdqa {{.*#+}} xmm14 = [255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0]
; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm2[2,2,3,3]
; SSE-NEXT: movdqa %xmm14, %xmm15
; SSE-NEXT: pandn %xmm0, %xmm15
; SSE-NEXT: pand %xmm14, %xmm5
; SSE-NEXT: por %xmm5, %xmm15
; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm13[1,0,2,2,4,5,6,7]
; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,1,0,1]
; SSE-NEXT: movdqa %xmm4, %xmm5
; SSE-NEXT: pandn %xmm0, %xmm5
; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm11[0,0,1,1]
; SSE-NEXT: pand %xmm4, %xmm0
; SSE-NEXT: por %xmm0, %xmm5
; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm9[0,0,0,0]
; SSE-NEXT: movdqa %xmm1, %xmm9
; SSE-NEXT: pandn %xmm0, %xmm9
; SSE-NEXT: pand %xmm1, %xmm5
; SSE-NEXT: por %xmm5, %xmm9
; SSE-NEXT: pand %xmm3, %xmm9
; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm2[0,0,0,0]
; SSE-NEXT: pandn %xmm0, %xmm3
; SSE-NEXT: por %xmm9, %xmm3
; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm5[3,3,3,3]
; SSE-NEXT: movdqa %xmm1, %xmm2
; SSE-NEXT: pandn %xmm0, %xmm2
; SSE-NEXT: pshufhw {{.*#+}} xmm0 = xmm7[0,1,2,3,5,6,7,7]
; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[2,2,2,3]
; SSE-NEXT: pand %xmm1, %xmm0
; SSE-NEXT: por %xmm2, %xmm0
; SSE-NEXT: pand %xmm4, %xmm0
; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm6[2,2,3,3]
; SSE-NEXT: pandn %xmm2, %xmm4
; SSE-NEXT: por %xmm0, %xmm4
; SSE-NEXT: pand %xmm14, %xmm4
; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm9[2,2,3,3]
; SSE-NEXT: pandn %xmm0, %xmm14
; SSE-NEXT: por %xmm4, %xmm14
; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm5[1,1,2,2]
; SSE-NEXT: pand %xmm1, %xmm0
; SSE-NEXT: pshuflw {{.*#+}} xmm2 = xmm7[3,3,3,3,4,5,6,7]
; SSE-NEXT: pshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,4,4,4]
; SSE-NEXT: pandn %xmm2, %xmm1
; SSE-NEXT: por %xmm0, %xmm1
; SSE-NEXT: pand %xmm8, %xmm1
; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm6[1,1,2,2]
; SSE-NEXT: pandn %xmm0, %xmm8
; SSE-NEXT: por %xmm1, %xmm8
; SSE-NEXT: pand %xmm10, %xmm8
; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm9[1,1,2,2]
; SSE-NEXT: pandn %xmm0, %xmm10
; SSE-NEXT: por %xmm8, %xmm10
; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT: movdqa %xmm10, 16(%rax)
; SSE-NEXT: movdqa %xmm14, 32(%rax)
; SSE-NEXT: movdqa %xmm3, 48(%rax)
; SSE-NEXT: movdqa %xmm15, 80(%rax)
; SSE-NEXT: movdqa %xmm12, 64(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT: movaps %xmm0, (%rax)
; SSE-NEXT: retq
;
; AVX1-ONLY-LABEL: store_i8_stride6_vf16:
; AVX1-ONLY: # %bb.0:
; AVX1-ONLY-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX1-ONLY-NEXT: vmovdqa (%rdi), %xmm1
; AVX1-ONLY-NEXT: vmovdqa (%rsi), %xmm2
; AVX1-ONLY-NEXT: vmovdqa (%rdx), %xmm3
; AVX1-ONLY-NEXT: vmovdqa (%rcx), %xmm4
; AVX1-ONLY-NEXT: vmovdqa (%r8), %xmm5
; AVX1-ONLY-NEXT: vmovdqa (%r9), %xmm6
; AVX1-ONLY-NEXT: vpunpcklbw {{.*#+}} xmm7 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3],xmm1[4],xmm2[4],xmm1[5],xmm2[5],xmm1[6],xmm2[6],xmm1[7],xmm2[7]
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm0 = xmm7[1,1,2,2]
; AVX1-ONLY-NEXT: vpunpcklbw {{.*#+}} xmm8 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3],xmm3[4],xmm4[4],xmm3[5],xmm4[5],xmm3[6],xmm4[6],xmm3[7],xmm4[7]
; AVX1-ONLY-NEXT: vpshuflw {{.*#+}} xmm9 = xmm8[3,3,3,3,4,5,6,7]
; AVX1-ONLY-NEXT: vpshufhw {{.*#+}} xmm9 = xmm9[0,1,2,3,4,4,4,4]
; AVX1-ONLY-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0,1],xmm9[2],xmm0[3,4],xmm9[5],xmm0[6,7]
; AVX1-ONLY-NEXT: vpunpcklbw {{.*#+}} xmm9 = 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]
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm10 = xmm9[1,1,2,2]
; AVX1-ONLY-NEXT: vpblendw {{.*#+}} xmm0 = xmm10[0],xmm0[1,2],xmm10[3],xmm0[4,5],xmm10[6],xmm0[7]
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm10 = xmm7[0,0,1,1]
; AVX1-ONLY-NEXT: vpshuflw {{.*#+}} xmm11 = xmm8[1,0,2,2,4,5,6,7]
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm11 = xmm11[0,1,0,1]
; AVX1-ONLY-NEXT: vpblendw {{.*#+}} xmm10 = xmm10[0],xmm11[1],xmm10[2,3],xmm11[4],xmm10[5,6],xmm11[7]
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm11 = xmm9[0,0,0,0]
; AVX1-ONLY-NEXT: vpblendw {{.*#+}} xmm10 = xmm10[0,1],xmm11[2],xmm10[3,4],xmm11[5],xmm10[6,7]
; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm0, %ymm10, %ymm0
; AVX1-ONLY-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm1[8],xmm2[8],xmm1[9],xmm2[9],xmm1[10],xmm2[10],xmm1[11],xmm2[11],xmm1[12],xmm2[12],xmm1[13],xmm2[13],xmm1[14],xmm2[14],xmm1[15],xmm2[15]
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm2 = xmm1[0,0,1,1]
; AVX1-ONLY-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm3[8],xmm4[8],xmm3[9],xmm4[9],xmm3[10],xmm4[10],xmm3[11],xmm4[11],xmm3[12],xmm4[12],xmm3[13],xmm4[13],xmm3[14],xmm4[14],xmm3[15],xmm4[15]
; AVX1-ONLY-NEXT: vpshuflw {{.*#+}} xmm4 = xmm3[1,0,2,2,4,5,6,7]
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm4 = xmm4[0,1,0,1]
; AVX1-ONLY-NEXT: vpblendw {{.*#+}} xmm2 = xmm2[0],xmm4[1],xmm2[2,3],xmm4[4],xmm2[5,6],xmm4[7]
; AVX1-ONLY-NEXT: vpunpckhbw {{.*#+}} xmm4 = xmm5[8],xmm6[8],xmm5[9],xmm6[9],xmm5[10],xmm6[10],xmm5[11],xmm6[11],xmm5[12],xmm6[12],xmm5[13],xmm6[13],xmm5[14],xmm6[14],xmm5[15],xmm6[15]
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm5 = xmm4[0,0,0,0]
; AVX1-ONLY-NEXT: vpblendw {{.*#+}} xmm2 = xmm2[0,1],xmm5[2],xmm2[3,4],xmm5[5],xmm2[6,7]
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm5 = xmm7[3,3,3,3]
; AVX1-ONLY-NEXT: vpshufhw {{.*#+}} xmm6 = xmm8[0,1,2,3,5,6,7,7]
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm6 = xmm6[2,2,2,3]
; AVX1-ONLY-NEXT: vpblendw {{.*#+}} xmm5 = xmm6[0,1],xmm5[2],xmm6[3,4],xmm5[5],xmm6[6,7]
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm6 = xmm9[2,2,3,3]
; AVX1-ONLY-NEXT: vpblendw {{.*#+}} xmm5 = xmm5[0],xmm6[1],xmm5[2,3],xmm6[4],xmm5[5,6],xmm6[7]
; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm2, %ymm5, %ymm2
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm5 = xmm1[3,3,3,3]
; AVX1-ONLY-NEXT: vpshufhw {{.*#+}} xmm6 = xmm3[0,1,2,3,5,6,7,7]
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm6 = xmm6[2,2,2,3]
; AVX1-ONLY-NEXT: vpblendw {{.*#+}} xmm5 = xmm6[0,1],xmm5[2],xmm6[3,4],xmm5[5],xmm6[6,7]
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm6 = xmm4[2,2,3,3]
; AVX1-ONLY-NEXT: vpblendw {{.*#+}} xmm5 = xmm5[0],xmm6[1],xmm5[2,3],xmm6[4],xmm5[5,6],xmm6[7]
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[1,1,2,2]
; AVX1-ONLY-NEXT: vpshuflw {{.*#+}} xmm3 = xmm3[3,3,3,3,4,5,6,7]
; AVX1-ONLY-NEXT: vpshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,4,4,4]
; AVX1-ONLY-NEXT: vpblendw {{.*#+}} xmm1 = xmm1[0,1],xmm3[2],xmm1[3,4],xmm3[5],xmm1[6,7]
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm3 = xmm4[1,1,2,2]
; AVX1-ONLY-NEXT: vpblendw {{.*#+}} xmm1 = xmm3[0],xmm1[1,2],xmm3[3],xmm1[4,5],xmm3[6],xmm1[7]
; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm5, %ymm1, %ymm1
; AVX1-ONLY-NEXT: vmovaps %ymm1, 64(%rax)
; AVX1-ONLY-NEXT: vmovaps %ymm2, 32(%rax)
; AVX1-ONLY-NEXT: vmovaps %ymm0, (%rax)
; AVX1-ONLY-NEXT: vzeroupper
; AVX1-ONLY-NEXT: retq
;
; AVX2-ONLY-LABEL: store_i8_stride6_vf16:
; AVX2-ONLY: # %bb.0:
; AVX2-ONLY-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX2-ONLY-NEXT: vmovdqa (%rdi), %xmm0
; AVX2-ONLY-NEXT: vmovdqa (%rdx), %xmm1
; AVX2-ONLY-NEXT: vmovdqa (%r8), %xmm2
; AVX2-ONLY-NEXT: vinserti128 $1, (%rsi), %ymm0, %ymm0
; AVX2-ONLY-NEXT: vinserti128 $1, (%rcx), %ymm1, %ymm1
; AVX2-ONLY-NEXT: vinserti128 $1, (%r9), %ymm2, %ymm2
; AVX2-ONLY-NEXT: vpermq {{.*#+}} ymm3 = ymm1[0,2,0,2]
; AVX2-ONLY-NEXT: vpshufb {{.*#+}} ymm3 = zero,zero,ymm3[0,8],zero,zero,zero,zero,ymm3[1,9],zero,zero,zero,zero,ymm3[2,10],zero,zero,zero,zero,ymm3[19,27],zero,zero,zero,zero,ymm3[20,28],zero,zero,zero,zero
; AVX2-ONLY-NEXT: vpermq {{.*#+}} ymm4 = ymm0[0,2,0,2]
; AVX2-ONLY-NEXT: vpshufb {{.*#+}} ymm4 = ymm4[0,8],zero,zero,zero,zero,ymm4[1,9],zero,zero,zero,zero,ymm4[2,10],zero,zero,zero,zero,ymm4[19,27],zero,zero,zero,zero,ymm4[20,28],zero,zero,zero,zero,ymm4[21,29]
; AVX2-ONLY-NEXT: vpor %ymm3, %ymm4, %ymm3
; AVX2-ONLY-NEXT: vpermq {{.*#+}} ymm4 = ymm2[0,2,0,2]
; AVX2-ONLY-NEXT: vpshufb {{.*#+}} ymm4 = ymm4[u,u,u,u,0,8,u,u,u,u,1,9,u,u,u,u,18,26,u,u,u,u,19,27,u,u,u,u,20,28,u,u]
; AVX2-ONLY-NEXT: vmovdqa {{.*#+}} ymm5 = [255,255,255,255,0,0,255,255,255,255,0,0,255,255,255,255,0,0,255,255,255,255,0,0,255,255,255,255,0,0,255,255]
; AVX2-ONLY-NEXT: vpblendvb %ymm5, %ymm3, %ymm4, %ymm3
; AVX2-ONLY-NEXT: vpermq {{.*#+}} ymm4 = ymm0[0,2,1,3]
; AVX2-ONLY-NEXT: vpshufb {{.*#+}} ymm4 = zero,zero,zero,zero,ymm4[6,14],zero,zero,zero,zero,ymm4[7,15],zero,zero,zero,zero,ymm4[16,24],zero,zero,zero,zero,ymm4[17,25],zero,zero,zero,zero,ymm4[18,26],zero,zero
; AVX2-ONLY-NEXT: vpermq {{.*#+}} ymm5 = ymm1[0,2,1,3]
; AVX2-ONLY-NEXT: vpshufb {{.*#+}} ymm5 = ymm5[5,13],zero,zero,zero,zero,ymm5[6,14],zero,zero,zero,zero,ymm5[7,15],zero,zero,zero,zero,ymm5[16,24],zero,zero,zero,zero,ymm5[17,25],zero,zero,zero,zero,ymm5[18,26]
; AVX2-ONLY-NEXT: vpor %ymm4, %ymm5, %ymm4
; AVX2-ONLY-NEXT: vpermq {{.*#+}} ymm5 = ymm2[0,2,1,3]
; AVX2-ONLY-NEXT: vpshufb {{.*#+}} ymm5 = ymm5[u,u,5,13,u,u,u,u,6,14,u,u,u,u,7,15,u,u,u,u,16,24,u,u,u,u,17,25,u,u,u,u]
; AVX2-ONLY-NEXT: vmovdqa {{.*#+}} ymm6 = [255,255,0,0,255,255,255,255,0,0,255,255,255,255,0,0,255,255,255,255,0,0,255,255,255,255,0,0,255,255,255,255]
; AVX2-ONLY-NEXT: vpblendvb %ymm6, %ymm4, %ymm5, %ymm4
; AVX2-ONLY-NEXT: vpermq {{.*#+}} ymm1 = ymm1[1,3,1,3]
; AVX2-ONLY-NEXT: vpshufb {{.*#+}} ymm1 = zero,zero,zero,zero,ymm1[3,11],zero,zero,zero,zero,ymm1[4,12],zero,zero,zero,zero,ymm1[21,29],zero,zero,zero,zero,ymm1[22,30],zero,zero,zero,zero,ymm1[23,31],zero,zero
; AVX2-ONLY-NEXT: vpermq {{.*#+}} ymm0 = ymm0[1,3,1,3]
; AVX2-ONLY-NEXT: vpshufb {{.*#+}} ymm0 = zero,zero,ymm0[3,11],zero,zero,zero,zero,ymm0[4,12],zero,zero,zero,zero,ymm0[5,13],zero,zero,zero,zero,ymm0[22,30],zero,zero,zero,zero,ymm0[23,31],zero,zero,zero,zero
; AVX2-ONLY-NEXT: vpor %ymm1, %ymm0, %ymm0
; AVX2-ONLY-NEXT: vpermq {{.*#+}} ymm1 = ymm2[1,3,1,3]
; AVX2-ONLY-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[2,10,u,u,u,u,3,11,u,u,u,u,4,12,u,u,u,u,21,29,u,u,u,u,22,30,u,u,u,u,23,31]
; AVX2-ONLY-NEXT: vmovdqa {{.*#+}} ymm2 = [0,0,255,255,255,255,0,0,255,255,255,255,0,0,255,255,255,255,0,0,255,255,255,255,0,0,255,255,255,255,0,0]
; AVX2-ONLY-NEXT: vpblendvb %ymm2, %ymm0, %ymm1, %ymm0
; AVX2-ONLY-NEXT: vmovdqa %ymm0, 64(%rax)
; AVX2-ONLY-NEXT: vmovdqa %ymm4, 32(%rax)
; AVX2-ONLY-NEXT: vmovdqa %ymm3, (%rax)
; AVX2-ONLY-NEXT: vzeroupper
; AVX2-ONLY-NEXT: retq
;
; AVX512F-LABEL: store_i8_stride6_vf16:
; AVX512F: # %bb.0:
; AVX512F-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512F-NEXT: vmovdqa (%rdi), %xmm0
; AVX512F-NEXT: vmovdqa (%rdx), %xmm1
; AVX512F-NEXT: vmovdqa (%r8), %xmm2
; AVX512F-NEXT: vinserti128 $1, (%rsi), %ymm0, %ymm0
; AVX512F-NEXT: vinserti128 $1, (%rcx), %ymm1, %ymm1
; AVX512F-NEXT: vinserti32x4 $1, (%r9), %zmm2, %zmm2
; AVX512F-NEXT: vpermq {{.*#+}} ymm3 = ymm1[0,2,0,2]
; AVX512F-NEXT: vpshufb {{.*#+}} ymm3 = zero,zero,ymm3[0,8,u,u],zero,zero,ymm3[1,9,u,u],zero,zero,ymm3[2,10,u,u],zero,zero,ymm3[19,27,u,u],zero,zero,ymm3[20,28,u,u],zero,zero
; AVX512F-NEXT: vpermq {{.*#+}} ymm4 = ymm0[0,2,1,3]
; AVX512F-NEXT: vpshufb {{.*#+}} ymm4 = zero,zero,ymm4[u,u,6,14],zero,zero,ymm4[u,u,7,15],zero,zero,ymm4[u,u,16,24],zero,zero,ymm4[u,u,17,25],zero,zero,ymm4[u,u,18,26],zero,zero
; AVX512F-NEXT: vinserti64x4 $1, %ymm4, %zmm3, %zmm3
; AVX512F-NEXT: vpermq {{.*#+}} ymm4 = ymm0[0,2,0,2]
; AVX512F-NEXT: vpshufb {{.*#+}} ymm4 = ymm4[0,8],zero,zero,ymm4[u,u,1,9],zero,zero,ymm4[u,u,2,10],zero,zero,ymm4[u,u,19,27],zero,zero,ymm4[u,u,20,28],zero,zero,ymm4[u,u,21,29]
; AVX512F-NEXT: vpermq {{.*#+}} ymm5 = ymm1[0,2,1,3]
; AVX512F-NEXT: vpshufb {{.*#+}} ymm5 = ymm5[5,13,u,u],zero,zero,ymm5[6,14,u,u],zero,zero,ymm5[7,15,u,u],zero,zero,ymm5[16,24,u,u],zero,zero,ymm5[17,25,u,u],zero,zero,ymm5[18,26]
; AVX512F-NEXT: vinserti64x4 $1, %ymm5, %zmm4, %zmm4
; AVX512F-NEXT: vporq %zmm3, %zmm4, %zmm3
; AVX512F-NEXT: vpermq {{.*#+}} ymm4 = ymm2[0,2,0,2]
; AVX512F-NEXT: vpshufb {{.*#+}} ymm4 = zero,zero,zero,zero,ymm4[0,8],zero,zero,zero,zero,ymm4[1,9],zero,zero,zero,zero,ymm4[18,26],zero,zero,zero,zero,ymm4[19,27],zero,zero,zero,zero,ymm4[20,28],zero,zero
; AVX512F-NEXT: vpermq {{.*#+}} ymm5 = ymm2[0,2,1,3]
; AVX512F-NEXT: vpshufb {{.*#+}} ymm5 = zero,zero,ymm5[5,13],zero,zero,zero,zero,ymm5[6,14],zero,zero,zero,zero,ymm5[7,15],zero,zero,zero,zero,ymm5[16,24],zero,zero,zero,zero,ymm5[17,25],zero,zero,zero,zero
; AVX512F-NEXT: vinserti64x4 $1, %ymm5, %zmm4, %zmm4
; AVX512F-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm3, %zmm4
; AVX512F-NEXT: vpermq {{.*#+}} ymm1 = ymm1[1,3,1,3]
; AVX512F-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[u,u],zero,zero,ymm1[3,11,u,u],zero,zero,ymm1[4,12,u,u],zero,zero,ymm1[21,29,u,u],zero,zero,ymm1[22,30,u,u],zero,zero,ymm1[23,31,u,u]
; AVX512F-NEXT: vpermq {{.*#+}} ymm0 = ymm0[1,3,1,3]
; AVX512F-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[u,u,3,11],zero,zero,ymm0[u,u,4,12],zero,zero,ymm0[u,u,5,13],zero,zero,ymm0[u,u,22,30],zero,zero,ymm0[u,u,23,31],zero,zero,ymm0[u,u]
; AVX512F-NEXT: vpor %ymm1, %ymm0, %ymm0
; AVX512F-NEXT: vpermq {{.*#+}} ymm1 = ymm2[1,3,1,3]
; AVX512F-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[2,10],zero,zero,zero,zero,ymm1[3,11],zero,zero,zero,zero,ymm1[4,12],zero,zero,zero,zero,ymm1[21,29],zero,zero,zero,zero,ymm1[22,30],zero,zero,zero,zero,ymm1[23,31]
; AVX512F-NEXT: vpternlogq $248, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm1
; AVX512F-NEXT: vmovdqa %ymm1, 64(%rax)
; AVX512F-NEXT: vmovdqa64 %zmm4, (%rax)
; AVX512F-NEXT: vzeroupper
; AVX512F-NEXT: retq
;
; AVX512BW-LABEL: store_i8_stride6_vf16:
; AVX512BW: # %bb.0:
; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-NEXT: vmovdqa (%rdi), %xmm0
; AVX512BW-NEXT: vmovdqa (%rdx), %xmm1
; AVX512BW-NEXT: vmovdqa (%r8), %xmm2
; AVX512BW-NEXT: vinserti128 $1, (%rsi), %ymm0, %ymm0
; AVX512BW-NEXT: vinserti128 $1, (%rcx), %ymm1, %ymm1
; AVX512BW-NEXT: vinserti32x4 $1, (%r9), %zmm2, %zmm2
; AVX512BW-NEXT: vpermq {{.*#+}} ymm3 = ymm1[0,2,0,2]
; AVX512BW-NEXT: vpermq {{.*#+}} ymm4 = ymm0[0,2,1,3]
; AVX512BW-NEXT: vinserti64x4 $1, %ymm4, %zmm3, %zmm3
; AVX512BW-NEXT: vpshufb {{.*#+}} zmm3 = zmm3[u,u,0,8,u,u,u,u,1,9,u,u,u,u,2,10,u,u,u,u,19,27,u,u,u,u,20,28,u,u,u,u,u,u,u,u,38,46,u,u,u,u,39,47,u,u,u,u,48,56,u,u,u,u,49,57,u,u,u,u,50,58,u,u]
; AVX512BW-NEXT: vpermq {{.*#+}} ymm4 = ymm0[0,2,0,2]
; AVX512BW-NEXT: vpermq {{.*#+}} ymm5 = ymm1[0,2,1,3]
; AVX512BW-NEXT: vinserti64x4 $1, %ymm5, %zmm4, %zmm4
; AVX512BW-NEXT: vpshufb {{.*#+}} zmm4 = zmm4[0,8,u,u,u,u,1,9,u,u,u,u,2,10,u,u,u,u,19,27,u,u,u,u,20,28,u,u,u,u,21,29,37,45,u,u,u,u,38,46,u,u,u,u,39,47,u,u,u,u,48,56,u,u,u,u,49,57,u,u,u,u,50,58]
; AVX512BW-NEXT: movl $1227105426, %ecx # imm = 0x49242492
; AVX512BW-NEXT: kmovd %ecx, %k1
; AVX512BW-NEXT: vmovdqu16 %zmm3, %zmm4 {%k1}
; AVX512BW-NEXT: vpermq {{.*#+}} ymm3 = ymm2[0,2,0,2]
; AVX512BW-NEXT: vpermq {{.*#+}} ymm5 = ymm2[0,2,1,3]
; AVX512BW-NEXT: vinserti64x4 $1, %ymm5, %zmm3, %zmm3
; AVX512BW-NEXT: vpshufb {{.*#+}} zmm3 = zmm3[u,u,u,u,0,8,u,u,u,u,1,9,u,u,u,u,18,26,u,u,u,u,19,27,u,u,u,u,20,28,u,u,u,u,37,45,u,u,u,u,38,46,u,u,u,u,39,47,u,u,u,u,48,56,u,u,u,u,49,57,u,u,u,u]
; AVX512BW-NEXT: movl $613566756, %ecx # imm = 0x24924924
; AVX512BW-NEXT: kmovd %ecx, %k1
; AVX512BW-NEXT: vmovdqu16 %zmm3, %zmm4 {%k1}
; AVX512BW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[1,3,1,3]
; AVX512BW-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,3,11,u,u,u,u,4,12,u,u,u,u,21,29,u,u,u,u,22,30,u,u,u,u,23,31,u,u]
; AVX512BW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[1,3,1,3]
; AVX512BW-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[u,u,3,11,u,u,u,u,4,12,u,u,u,u,5,13,u,u,u,u,22,30,u,u,u,u,23,31,u,u,u,u]
; AVX512BW-NEXT: movw $18724, %cx # imm = 0x4924
; AVX512BW-NEXT: kmovd %ecx, %k1
; AVX512BW-NEXT: vmovdqu16 %ymm1, %ymm0 {%k1}
; AVX512BW-NEXT: vpermq {{.*#+}} ymm1 = ymm2[1,3,1,3]
; AVX512BW-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[2,10,u,u,u,u,3,11,u,u,u,u,4,12,u,u,u,u,21,29,u,u,u,u,22,30,u,u,u,u,23,31]
; AVX512BW-NEXT: movw $-28087, %cx # imm = 0x9249
; AVX512BW-NEXT: kmovd %ecx, %k1
; AVX512BW-NEXT: vmovdqu16 %ymm1, %ymm0 {%k1}
; AVX512BW-NEXT: vmovdqa %ymm0, 64(%rax)
; AVX512BW-NEXT: vmovdqa64 %zmm4, (%rax)
; AVX512BW-NEXT: vzeroupper
; AVX512BW-NEXT: retq
%in.vec0 = load <16 x i8>, ptr %in.vecptr0, align 64
%in.vec1 = load <16 x i8>, ptr %in.vecptr1, align 64
%in.vec2 = load <16 x i8>, ptr %in.vecptr2, align 64
%in.vec3 = load <16 x i8>, ptr %in.vecptr3, align 64
%in.vec4 = load <16 x i8>, ptr %in.vecptr4, align 64
%in.vec5 = load <16 x i8>, ptr %in.vecptr5, align 64
%1 = shufflevector <16 x i8> %in.vec0, <16 x i8> %in.vec1, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
%2 = shufflevector <16 x i8> %in.vec2, <16 x i8> %in.vec3, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
%3 = shufflevector <16 x i8> %in.vec4, <16 x i8> %in.vec5, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
%4 = shufflevector <32 x i8> %1, <32 x i8> %2, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
%5 = shufflevector <32 x i8> %3, <32 x i8> poison, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
%6 = shufflevector <64 x i8> %4, <64 x i8> %5, <96 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 64, i32 65, i32 66, i32 67, i32 68, i32 69, i32 70, i32 71, i32 72, i32 73, i32 74, i32 75, i32 76, i32 77, i32 78, i32 79, i32 80, i32 81, i32 82, i32 83, i32 84, i32 85, i32 86, i32 87, i32 88, i32 89, i32 90, i32 91, i32 92, i32 93, i32 94, i32 95>
%interleaved.vec = shufflevector <96 x i8> %6, <96 x i8> poison, <96 x i32> <i32 0, i32 16, i32 32, i32 48, i32 64, i32 80, i32 1, i32 17, i32 33, i32 49, i32 65, i32 81, i32 2, i32 18, i32 34, i32 50, i32 66, i32 82, i32 3, i32 19, i32 35, i32 51, i32 67, i32 83, i32 4, i32 20, i32 36, i32 52, i32 68, i32 84, i32 5, i32 21, i32 37, i32 53, i32 69, i32 85, i32 6, i32 22, i32 38, i32 54, i32 70, i32 86, i32 7, i32 23, i32 39, i32 55, i32 71, i32 87, i32 8, i32 24, i32 40, i32 56, i32 72, i32 88, i32 9, i32 25, i32 41, i32 57, i32 73, i32 89, i32 10, i32 26, i32 42, i32 58, i32 74, i32 90, i32 11, i32 27, i32 43, i32 59, i32 75, i32 91, i32 12, i32 28, i32 44, i32 60, i32 76, i32 92, i32 13, i32 29, i32 45, i32 61, i32 77, i32 93, i32 14, i32 30, i32 46, i32 62, i32 78, i32 94, i32 15, i32 31, i32 47, i32 63, i32 79, i32 95>
store <96 x i8> %interleaved.vec, ptr %out.vec, align 64
ret void
}
define void @store_i8_stride6_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecptr2, ptr %in.vecptr3, ptr %in.vecptr4, ptr %in.vecptr5, ptr %out.vec) nounwind {
; SSE-LABEL: store_i8_stride6_vf32:
; SSE: # %bb.0:
; SSE-NEXT: subq $200, %rsp
; SSE-NEXT: movdqa 16(%rdi), %xmm8
; SSE-NEXT: movdqa 16(%rsi), %xmm5
; SSE-NEXT: movdqa 16(%rdx), %xmm12
; SSE-NEXT: movdqa 16(%rcx), %xmm4
; SSE-NEXT: movdqa 16(%r8), %xmm11
; SSE-NEXT: movdqa 16(%r9), %xmm0
; 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: pshufd {{.*#+}} xmm2 = xmm1[0,0,1,1]
; SSE-NEXT: movdqa {{.*#+}} xmm9 = [65535,0,65535,65535,0,65535,65535,0]
; SSE-NEXT: pand %xmm9, %xmm2
; SSE-NEXT: movdqa %xmm12, %xmm1
; SSE-NEXT: punpckhbw {{.*#+}} xmm1 = xmm1[8],xmm4[8],xmm1[9],xmm4[9],xmm1[10],xmm4[10],xmm1[11],xmm4[11],xmm1[12],xmm4[12],xmm1[13],xmm4[13],xmm1[14],xmm4[14],xmm1[15],xmm4[15]
; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT: pshuflw {{.*#+}} xmm3 = xmm1[1,0,2,2,4,5,6,7]
; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm3[0,1,0,1]
; SSE-NEXT: movdqa %xmm9, %xmm6
; SSE-NEXT: pandn %xmm3, %xmm6
; SSE-NEXT: por %xmm2, %xmm6
; SSE-NEXT: movdqa {{.*#+}} xmm2 = [65535,65535,0,65535,65535,0,65535,65535]
; SSE-NEXT: pand %xmm2, %xmm6
; SSE-NEXT: movdqa %xmm11, %xmm1
; SSE-NEXT: punpckhbw {{.*#+}} xmm1 = xmm1[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm1[0,0,0,0]
; SSE-NEXT: movdqa %xmm2, %xmm7
; SSE-NEXT: pandn %xmm3, %xmm7
; SSE-NEXT: por %xmm6, %xmm7
; SSE-NEXT: movdqa {{.*#+}} xmm10 = [255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255]
; SSE-NEXT: pand %xmm10, %xmm7
; SSE-NEXT: punpckhbw {{.*#+}} xmm1 = xmm1[8],xmm0[8],xmm1[9],xmm0[9],xmm1[10],xmm0[10],xmm1[11],xmm0[11],xmm1[12],xmm0[12],xmm1[13],xmm0[13],xmm1[14],xmm0[14],xmm1[15],xmm0[15]
; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm1[0,0,0,0]
; SSE-NEXT: movdqa %xmm10, %xmm1
; SSE-NEXT: pandn %xmm6, %xmm1
; SSE-NEXT: por %xmm7, %xmm1
; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT: punpcklbw {{.*#+}} xmm8 = xmm8[0],xmm5[0],xmm8[1],xmm5[1],xmm8[2],xmm5[2],xmm8[3],xmm5[3],xmm8[4],xmm5[4],xmm8[5],xmm5[5],xmm8[6],xmm5[6],xmm8[7],xmm5[7]
; SSE-NEXT: movdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm8[3,3,3,3]
; SSE-NEXT: movdqa %xmm2, %xmm6
; SSE-NEXT: pandn %xmm5, %xmm6
; SSE-NEXT: punpcklbw {{.*#+}} xmm12 = xmm12[0],xmm4[0],xmm12[1],xmm4[1],xmm12[2],xmm4[2],xmm12[3],xmm4[3],xmm12[4],xmm4[4],xmm12[5],xmm4[5],xmm12[6],xmm4[6],xmm12[7],xmm4[7]
; SSE-NEXT: movdqa %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT: pshufhw {{.*#+}} xmm4 = xmm12[0,1,2,3,5,6,7,7]
; SSE-NEXT: pshufd {{.*#+}} xmm4 = xmm4[2,2,2,3]
; SSE-NEXT: pand %xmm2, %xmm4
; SSE-NEXT: por %xmm6, %xmm4
; SSE-NEXT: pand %xmm9, %xmm4
; SSE-NEXT: punpcklbw {{.*#+}} xmm11 = xmm11[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
; SSE-NEXT: movdqa %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm11[2,2,3,3]
; SSE-NEXT: movdqa %xmm9, %xmm6
; SSE-NEXT: pandn %xmm5, %xmm6
; SSE-NEXT: por %xmm4, %xmm6
; SSE-NEXT: movdqa {{.*#+}} xmm7 = [255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0]
; SSE-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[2,2,3,3]
; SSE-NEXT: movdqa %xmm7, %xmm1
; SSE-NEXT: pandn %xmm0, %xmm1
; SSE-NEXT: movdqa (%rdi), %xmm0
; SSE-NEXT: pand %xmm7, %xmm6
; SSE-NEXT: por %xmm6, %xmm1
; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT: movdqa (%rsi), %xmm12
; SSE-NEXT: movdqa %xmm0, %xmm1
; SSE-NEXT: movdqa %xmm0, %xmm8
; SSE-NEXT: punpckhbw {{.*#+}} xmm1 = xmm1[8],xmm12[8],xmm1[9],xmm12[9],xmm1[10],xmm12[10],xmm1[11],xmm12[11],xmm1[12],xmm12[12],xmm1[13],xmm12[13],xmm1[14],xmm12[14],xmm1[15],xmm12[15]
; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm1[3,3,3,3]
; SSE-NEXT: movdqa %xmm1, %xmm3
; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT: movdqa %xmm2, %xmm4
; SSE-NEXT: pandn %xmm0, %xmm4
; SSE-NEXT: movdqa (%rdx), %xmm14
; SSE-NEXT: movdqa (%rcx), %xmm6
; 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: pshufhw {{.*#+}} xmm0 = xmm1[0,1,2,3,5,6,7,7]
; SSE-NEXT: movdqa %xmm1, %xmm11
; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[2,2,2,3]
; SSE-NEXT: pand %xmm2, %xmm0
; SSE-NEXT: por %xmm4, %xmm0
; SSE-NEXT: movdqa (%r8), %xmm13
; SSE-NEXT: movdqa %xmm13, %xmm5
; SSE-NEXT: punpckhbw {{.*#+}} xmm5 = xmm5[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
; SSE-NEXT: pshufd {{.*#+}} xmm4 = xmm5[2,2,3,3]
; SSE-NEXT: movdqa %xmm5, (%rsp) # 16-byte Spill
; SSE-NEXT: movdqa %xmm9, %xmm15
; SSE-NEXT: pandn %xmm4, %xmm15
; SSE-NEXT: pand %xmm9, %xmm0
; SSE-NEXT: por %xmm0, %xmm15
; SSE-NEXT: movdqa (%r9), %xmm0
; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT: punpckhbw {{.*#+}} xmm4 = xmm4[8],xmm0[8],xmm4[9],xmm0[9],xmm4[10],xmm0[10],xmm4[11],xmm0[11],xmm4[12],xmm0[12],xmm4[13],xmm0[13],xmm4[14],xmm0[14],xmm4[15],xmm0[15]
; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm4[2,2,3,3]
; SSE-NEXT: movdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT: movdqa %xmm7, %xmm1
; SSE-NEXT: pandn %xmm0, %xmm1
; SSE-NEXT: pand %xmm7, %xmm15
; SSE-NEXT: por %xmm15, %xmm1
; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm11[3,3,3,3,4,5,6,7]
; SSE-NEXT: pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,4,4,4]
; SSE-NEXT: movdqa %xmm2, %xmm15
; SSE-NEXT: pandn %xmm0, %xmm15
; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm3[1,1,2,2]
; SSE-NEXT: pand %xmm2, %xmm0
; SSE-NEXT: por %xmm0, %xmm15
; SSE-NEXT: movdqa {{.*#+}} xmm1 = [0,65535,65535,0,65535,65535,0,65535]
; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm5[1,1,2,2]
; SSE-NEXT: movdqa %xmm1, %xmm0
; SSE-NEXT: pandn %xmm3, %xmm0
; SSE-NEXT: pand %xmm1, %xmm15
; SSE-NEXT: por %xmm15, %xmm0
; SSE-NEXT: movdqa {{.*#+}} xmm15 = [255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255]
; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm4[1,1,2,2]
; SSE-NEXT: movdqa %xmm15, %xmm11
; SSE-NEXT: pandn %xmm3, %xmm11
; SSE-NEXT: pand %xmm15, %xmm0
; SSE-NEXT: por %xmm0, %xmm11
; SSE-NEXT: movdqa %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT: punpcklbw {{.*#+}} xmm8 = xmm8[0],xmm12[0],xmm8[1],xmm12[1],xmm8[2],xmm12[2],xmm8[3],xmm12[3],xmm8[4],xmm12[4],xmm8[5],xmm12[5],xmm8[6],xmm12[6],xmm8[7],xmm12[7]
; 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: pshuflw {{.*#+}} xmm0 = xmm14[3,3,3,3,4,5,6,7]
; SSE-NEXT: movdqa %xmm14, %xmm4
; SSE-NEXT: movdqa %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT: pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,4,4,4]
; SSE-NEXT: movdqa %xmm2, %xmm3
; SSE-NEXT: pandn %xmm0, %xmm3
; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm8[1,1,2,2]
; SSE-NEXT: movdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT: pand %xmm2, %xmm0
; SSE-NEXT: por %xmm0, %xmm3
; SSE-NEXT: punpcklbw {{.*#+}} xmm13 = xmm13[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm13[1,1,2,2]
; SSE-NEXT: movdqa %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT: movdqa %xmm1, %xmm5
; SSE-NEXT: pandn %xmm0, %xmm5
; SSE-NEXT: pand %xmm1, %xmm3
; SSE-NEXT: por %xmm3, %xmm5
; SSE-NEXT: punpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload
; SSE-NEXT: # xmm3 = xmm3[0],mem[0],xmm3[1],mem[1],xmm3[2],mem[2],xmm3[3],mem[3],xmm3[4],mem[4],xmm3[5],mem[5],xmm3[6],mem[6],xmm3[7],mem[7]
; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm3[1,1,2,2]
; SSE-NEXT: movdqa %xmm3, %xmm6
; SSE-NEXT: movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT: movdqa %xmm15, %xmm14
; SSE-NEXT: pandn %xmm0, %xmm14
; SSE-NEXT: pand %xmm15, %xmm5
; SSE-NEXT: por %xmm5, %xmm14
; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm4[1,0,2,2,4,5,6,7]
; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,1,0,1]
; SSE-NEXT: movdqa %xmm9, %xmm3
; SSE-NEXT: pandn %xmm0, %xmm3
; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm8[0,0,1,1]
; SSE-NEXT: pand %xmm9, %xmm0
; SSE-NEXT: por %xmm0, %xmm3
; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm13[0,0,0,0]
; SSE-NEXT: movdqa %xmm2, %xmm4
; SSE-NEXT: pandn %xmm0, %xmm4
; SSE-NEXT: pand %xmm2, %xmm3
; SSE-NEXT: por %xmm3, %xmm4
; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm6[0,0,0,0]
; SSE-NEXT: movdqa %xmm10, %xmm12
; SSE-NEXT: pandn %xmm0, %xmm12
; SSE-NEXT: pand %xmm10, %xmm4
; SSE-NEXT: por %xmm4, %xmm12
; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm4[3,3,3,3]
; SSE-NEXT: movdqa %xmm2, %xmm3
; SSE-NEXT: pandn %xmm0, %xmm3
; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; SSE-NEXT: pshufhw {{.*#+}} xmm0 = xmm5[0,1,2,3,5,6,7,7]
; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[2,2,2,3]
; SSE-NEXT: pand %xmm2, %xmm0
; SSE-NEXT: por %xmm3, %xmm0
; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm6[2,2,3,3]
; SSE-NEXT: movdqa %xmm9, %xmm8
; SSE-NEXT: pandn %xmm3, %xmm8
; SSE-NEXT: pand %xmm9, %xmm0
; SSE-NEXT: por %xmm0, %xmm8
; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm13[2,2,3,3]
; SSE-NEXT: movdqa %xmm7, %xmm11
; SSE-NEXT: pandn %xmm0, %xmm11
; SSE-NEXT: pand %xmm7, %xmm8
; SSE-NEXT: por %xmm8, %xmm11
; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm5[3,3,3,3,4,5,6,7]
; SSE-NEXT: pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,4,4,4]
; SSE-NEXT: movdqa %xmm2, %xmm3
; SSE-NEXT: pandn %xmm0, %xmm3
; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm4[1,1,2,2]
; SSE-NEXT: pand %xmm2, %xmm0
; SSE-NEXT: por %xmm0, %xmm3
; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm6[1,1,2,2]
; SSE-NEXT: movdqa %xmm1, %xmm8
; SSE-NEXT: pandn %xmm0, %xmm8
; SSE-NEXT: pand %xmm1, %xmm3
; SSE-NEXT: por %xmm3, %xmm8
; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm13[1,1,2,2]
; SSE-NEXT: movdqa %xmm15, %xmm6
; SSE-NEXT: pandn %xmm0, %xmm6
; SSE-NEXT: pand %xmm15, %xmm8
; SSE-NEXT: por %xmm8, %xmm6
; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm3[3,3,3,3,4,5,6,7]
; SSE-NEXT: pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,4,4,4]
; SSE-NEXT: movdqa %xmm2, %xmm8
; SSE-NEXT: pandn %xmm0, %xmm8
; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload
; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm13[1,1,2,2]
; SSE-NEXT: pand %xmm2, %xmm0
; SSE-NEXT: por %xmm0, %xmm8
; SSE-NEXT: pand %xmm1, %xmm8
; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload
; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm4[1,1,2,2]
; SSE-NEXT: pandn %xmm0, %xmm1
; SSE-NEXT: por %xmm8, %xmm1
; SSE-NEXT: pand %xmm15, %xmm1
; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm5[1,1,2,2]
; SSE-NEXT: pandn %xmm0, %xmm15
; SSE-NEXT: por %xmm1, %xmm15
; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm3[1,0,2,2,4,5,6,7]
; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,1,0,1]
; SSE-NEXT: movdqa %xmm9, %xmm1
; SSE-NEXT: pandn %xmm0, %xmm1
; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm13[0,0,1,1]
; SSE-NEXT: pand %xmm9, %xmm0
; SSE-NEXT: por %xmm0, %xmm1
; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm4[0,0,0,0]
; SSE-NEXT: movdqa %xmm2, %xmm8
; SSE-NEXT: pandn %xmm0, %xmm8
; SSE-NEXT: pand %xmm2, %xmm1
; SSE-NEXT: por %xmm1, %xmm8
; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm5[0,0,0,0]
; SSE-NEXT: movdqa %xmm10, %xmm0
; SSE-NEXT: pandn %xmm1, %xmm0
; SSE-NEXT: pand %xmm10, %xmm8
; SSE-NEXT: por %xmm8, %xmm0
; SSE-NEXT: pshuflw $161, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT: # xmm1 = mem[1,0,2,2,4,5,6,7]
; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,1,0,1]
; SSE-NEXT: movdqa %xmm9, %xmm8
; SSE-NEXT: pandn %xmm1, %xmm8
; SSE-NEXT: pshufd $80, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT: # xmm1 = mem[0,0,1,1]
; SSE-NEXT: pand %xmm9, %xmm1
; SSE-NEXT: por %xmm1, %xmm8
; SSE-NEXT: pshufd $0, (%rsp), %xmm1 # 16-byte Folded Reload
; SSE-NEXT: # xmm1 = mem[0,0,0,0]
; SSE-NEXT: movdqa %xmm2, %xmm13
; SSE-NEXT: pandn %xmm1, %xmm13
; SSE-NEXT: pand %xmm2, %xmm8
; SSE-NEXT: por %xmm8, %xmm13
; SSE-NEXT: pand %xmm10, %xmm13
; SSE-NEXT: pshufd $0, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT: # xmm1 = mem[0,0,0,0]
; SSE-NEXT: pandn %xmm1, %xmm10
; SSE-NEXT: por %xmm13, %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: pshufhw $249, {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Folded Reload
; SSE-NEXT: # xmm8 = mem[0,1,2,3,5,6,7,7]
; SSE-NEXT: pshufd {{.*#+}} xmm8 = xmm8[2,2,2,3]
; SSE-NEXT: pand %xmm2, %xmm8
; SSE-NEXT: pandn %xmm1, %xmm2
; SSE-NEXT: por %xmm8, %xmm2
; SSE-NEXT: pand %xmm9, %xmm2
; SSE-NEXT: pshufd $250, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT: # xmm1 = mem[2,2,3,3]
; SSE-NEXT: pandn %xmm1, %xmm9
; SSE-NEXT: por %xmm2, %xmm9
; SSE-NEXT: pand %xmm7, %xmm9
; SSE-NEXT: pshufd $250, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT: # xmm1 = mem[2,2,3,3]
; SSE-NEXT: pandn %xmm1, %xmm7
; SSE-NEXT: por %xmm9, %xmm7
; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT: movdqa %xmm7, 32(%rax)
; SSE-NEXT: movdqa %xmm10, 48(%rax)
; SSE-NEXT: movdqa %xmm0, 96(%rax)
; SSE-NEXT: movdqa %xmm15, 112(%rax)
; SSE-NEXT: movdqa %xmm6, 160(%rax)
; SSE-NEXT: movdqa %xmm11, 176(%rax)
; SSE-NEXT: movdqa %xmm12, (%rax)
; SSE-NEXT: movdqa %xmm14, 16(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT: movaps %xmm0, 64(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT: movaps %xmm0, 80(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT: movaps %xmm0, 128(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT: movaps %xmm0, 144(%rax)
; SSE-NEXT: addq $200, %rsp
; SSE-NEXT: retq
;
; AVX1-ONLY-LABEL: store_i8_stride6_vf32:
; AVX1-ONLY: # %bb.0:
; AVX1-ONLY-NEXT: vmovdqa 16(%rsi), %xmm0
; AVX1-ONLY-NEXT: vmovdqa 16(%rdi), %xmm1
; AVX1-ONLY-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3],xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
; AVX1-ONLY-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[3,3,3,3]
; AVX1-ONLY-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm1[8],xmm0[8],xmm1[9],xmm0[9],xmm1[10],xmm0[10],xmm1[11],xmm0[11],xmm1[12],xmm0[12],xmm1[13],xmm0[13],xmm1[14],xmm0[14],xmm1[15],xmm0[15]
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[0,0,1,1]
; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm1, %ymm2, %ymm1
; AVX1-ONLY-NEXT: vmovaps {{.*#+}} ymm10 = [65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535]
; AVX1-ONLY-NEXT: vandnps %ymm1, %ymm10, %ymm2
; AVX1-ONLY-NEXT: vmovdqa 16(%rcx), %xmm1
; AVX1-ONLY-NEXT: vmovdqa 16(%rdx), %xmm3
; AVX1-ONLY-NEXT: vpunpcklbw {{.*#+}} xmm13 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3],xmm3[4],xmm1[4],xmm3[5],xmm1[5],xmm3[6],xmm1[6],xmm3[7],xmm1[7]
; AVX1-ONLY-NEXT: vpshufhw {{.*#+}} xmm4 = xmm13[0,1,2,3,5,6,7,7]
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm4 = xmm4[2,2,2,3]
; AVX1-ONLY-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm3[8],xmm1[8],xmm3[9],xmm1[9],xmm3[10],xmm1[10],xmm3[11],xmm1[11],xmm3[12],xmm1[12],xmm3[13],xmm1[13],xmm3[14],xmm1[14],xmm3[15],xmm1[15]
; AVX1-ONLY-NEXT: vpshuflw {{.*#+}} xmm3 = xmm1[1,0,2,2,4,5,6,7]
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm3 = xmm3[0,1,0,1]
; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm3, %ymm4, %ymm3
; AVX1-ONLY-NEXT: vandps %ymm3, %ymm10, %ymm3
; AVX1-ONLY-NEXT: vorps %ymm2, %ymm3, %ymm2
; AVX1-ONLY-NEXT: vextractf128 $1, %ymm2, %xmm3
; AVX1-ONLY-NEXT: vmovdqa 16(%r8), %xmm12
; AVX1-ONLY-NEXT: vpshufb {{.*#+}} xmm4 = zero,zero,zero,zero,xmm12[8,u],zero,zero,zero,zero,xmm12[9,u],zero,zero,zero,zero
; AVX1-ONLY-NEXT: vpblendw {{.*#+}} xmm3 = xmm3[0,1],xmm4[2],xmm3[3,4],xmm4[5],xmm3[6,7]
; AVX1-ONLY-NEXT: vpshufb {{.*#+}} xmm3 = xmm3[0,1,2,3,4],zero,xmm3[6,7,8,9,10],zero,xmm3[12,13,14,15]
; AVX1-ONLY-NEXT: vmovdqa 16(%r9), %xmm14
; AVX1-ONLY-NEXT: vpshufb {{.*#+}} xmm4 = zero,zero,zero,zero,zero,xmm14[8],zero,zero,zero,zero,zero,xmm14[9],zero,zero,zero,zero
; AVX1-ONLY-NEXT: vpor %xmm4, %xmm3, %xmm3
; AVX1-ONLY-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vpshufb {{.*#+}} xmm3 = zero,zero,xmm12[5,u],zero,zero,zero,zero,xmm12[6,u],zero,zero,zero,zero,xmm12[7,u]
; AVX1-ONLY-NEXT: vpblendw {{.*#+}} xmm2 = xmm2[0],xmm3[1],xmm2[2,3],xmm3[4],xmm2[5,6],xmm3[7]
; AVX1-ONLY-NEXT: vmovdqa {{.*#+}} xmm8 = [0,1,2,128,4,5,6,7,8,128,10,11,12,13,14,128]
; AVX1-ONLY-NEXT: vpshufb %xmm8, %xmm2, %xmm2
; AVX1-ONLY-NEXT: vpshufb {{.*#+}} xmm3 = zero,zero,zero,xmm14[5],zero,zero,zero,zero,zero,xmm14[6],zero,zero,zero,zero,zero,xmm14[7]
; AVX1-ONLY-NEXT: vpor %xmm3, %xmm2, %xmm2
; AVX1-ONLY-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm2 = xmm0[1,1,2,2]
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[3,3,3,3]
; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm0, %ymm2, %ymm0
; AVX1-ONLY-NEXT: vandps %ymm0, %ymm10, %ymm0
; AVX1-ONLY-NEXT: vpshuflw {{.*#+}} xmm2 = xmm1[3,3,3,3,4,5,6,7]
; AVX1-ONLY-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,4,4,4]
; AVX1-ONLY-NEXT: vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,5,6,7,7]
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[2,2,2,3]
; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm1, %ymm2, %ymm1
; AVX1-ONLY-NEXT: vandnps %ymm1, %ymm10, %ymm1
; AVX1-ONLY-NEXT: vorps %ymm1, %ymm0, %ymm0
; AVX1-ONLY-NEXT: vpshufb {{.*#+}} xmm1 = xmm12[10,u],zero,zero,zero,zero,xmm12[11,u],zero,zero,zero,zero,xmm12[12,u],zero,zero
; AVX1-ONLY-NEXT: vpblendw {{.*#+}} xmm1 = xmm1[0],xmm0[1,2],xmm1[3],xmm0[4,5],xmm1[6],xmm0[7]
; AVX1-ONLY-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[0],zero,xmm1[2,3,4,5,6],zero,xmm1[8,9,10,11,12],zero,xmm1[14,15]
; AVX1-ONLY-NEXT: vpshufb {{.*#+}} xmm2 = zero,xmm14[10],zero,zero,zero,zero,zero,xmm14[11],zero,zero,zero,zero,zero,xmm14[12],zero,zero
; AVX1-ONLY-NEXT: vpor %xmm2, %xmm1, %xmm1
; AVX1-ONLY-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vextractf128 $1, %ymm0, %xmm0
; AVX1-ONLY-NEXT: vmovdqa {{.*#+}} xmm9 = <128,128,13,u,128,128,128,128,14,u,128,128,128,128,15,u>
; AVX1-ONLY-NEXT: vpshufb %xmm9, %xmm12, %xmm1
; AVX1-ONLY-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0],xmm1[1],xmm0[2,3],xmm1[4],xmm0[5,6],xmm1[7]
; AVX1-ONLY-NEXT: vpshufb %xmm8, %xmm0, %xmm0
; AVX1-ONLY-NEXT: vmovdqa {{.*#+}} xmm11 = [128,128,128,13,128,128,128,128,128,14,128,128,128,128,128,15]
; AVX1-ONLY-NEXT: vpshufb %xmm11, %xmm14, %xmm1
; AVX1-ONLY-NEXT: vpor %xmm1, %xmm0, %xmm0
; AVX1-ONLY-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovdqa (%rsi), %xmm3
; AVX1-ONLY-NEXT: vmovdqa (%rdi), %xmm2
; AVX1-ONLY-NEXT: vpunpckhbw {{.*#+}} xmm4 = xmm2[8],xmm3[8],xmm2[9],xmm3[9],xmm2[10],xmm3[10],xmm2[11],xmm3[11],xmm2[12],xmm3[12],xmm2[13],xmm3[13],xmm2[14],xmm3[14],xmm2[15],xmm3[15]
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm0 = xmm4[1,1,2,2]
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm1 = xmm4[3,3,3,3]
; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
; AVX1-ONLY-NEXT: vmovdqa (%rcx), %xmm7
; AVX1-ONLY-NEXT: vmovdqa (%rdx), %xmm6
; AVX1-ONLY-NEXT: vpunpckhbw {{.*#+}} xmm5 = xmm6[8],xmm7[8],xmm6[9],xmm7[9],xmm6[10],xmm7[10],xmm6[11],xmm7[11],xmm6[12],xmm7[12],xmm6[13],xmm7[13],xmm6[14],xmm7[14],xmm6[15],xmm7[15]
; AVX1-ONLY-NEXT: vpshuflw {{.*#+}} xmm1 = xmm5[3,3,3,3,4,5,6,7]
; AVX1-ONLY-NEXT: vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,4,4,4]
; AVX1-ONLY-NEXT: vpshufhw {{.*#+}} xmm15 = xmm5[0,1,2,3,5,6,7,7]
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm15 = xmm15[2,2,2,3]
; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm15, %ymm1, %ymm1
; AVX1-ONLY-NEXT: vandps %ymm0, %ymm10, %ymm0
; AVX1-ONLY-NEXT: vandnps %ymm1, %ymm10, %ymm1
; AVX1-ONLY-NEXT: vorps %ymm1, %ymm0, %ymm15
; AVX1-ONLY-NEXT: vmovdqa (%r8), %xmm1
; AVX1-ONLY-NEXT: vpshufb %xmm9, %xmm1, %xmm0
; AVX1-ONLY-NEXT: vextractf128 $1, %ymm15, %xmm9
; AVX1-ONLY-NEXT: vpblendw {{.*#+}} xmm0 = xmm9[0],xmm0[1],xmm9[2,3],xmm0[4],xmm9[5,6],xmm0[7]
; AVX1-ONLY-NEXT: vpshufb %xmm8, %xmm0, %xmm9
; AVX1-ONLY-NEXT: vmovdqa (%r9), %xmm0
; AVX1-ONLY-NEXT: vpshufb %xmm11, %xmm0, %xmm8
; AVX1-ONLY-NEXT: vpor %xmm8, %xmm9, %xmm8
; AVX1-ONLY-NEXT: vmovdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vpshufb {{.*#+}} xmm8 = xmm1[10,u],zero,zero,zero,zero,xmm1[11,u],zero,zero,zero,zero,xmm1[12,u],zero,zero
; AVX1-ONLY-NEXT: vpblendw {{.*#+}} xmm8 = xmm8[0],xmm15[1,2],xmm8[3],xmm15[4,5],xmm8[6],xmm15[7]
; AVX1-ONLY-NEXT: vpshufb {{.*#+}} xmm8 = xmm8[0],zero,xmm8[2,3,4,5,6],zero,xmm8[8,9,10,11,12],zero,xmm8[14,15]
; AVX1-ONLY-NEXT: vpshufb {{.*#+}} xmm9 = zero,xmm0[10],zero,zero,zero,zero,zero,xmm0[11],zero,zero,zero,zero,zero,xmm0[12],zero,zero
; AVX1-ONLY-NEXT: vpor %xmm9, %xmm8, %xmm8
; AVX1-ONLY-NEXT: vmovdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm8 = xmm9[0,0,1,1]
; AVX1-ONLY-NEXT: vshufps {{.*#+}} xmm10 = xmm9[1,1,2,2]
; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm10, %ymm8, %ymm8
; AVX1-ONLY-NEXT: vpshuflw {{.*#+}} xmm10 = xmm13[1,0,2,2,4,5,6,7]
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm10 = xmm10[0,1,0,1]
; AVX1-ONLY-NEXT: vpshuflw {{.*#+}} xmm11 = xmm13[3,3,3,3,4,5,6,7]
; AVX1-ONLY-NEXT: vpshufhw {{.*#+}} xmm11 = xmm11[0,1,2,3,4,4,4,4]
; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm11, %ymm10, %ymm10
; AVX1-ONLY-NEXT: vmovaps {{.*#+}} ymm15 = [65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535]
; AVX1-ONLY-NEXT: vandps %ymm15, %ymm8, %ymm8
; AVX1-ONLY-NEXT: vandnps %ymm10, %ymm15, %ymm10
; AVX1-ONLY-NEXT: vorps %ymm10, %ymm8, %ymm10
; AVX1-ONLY-NEXT: vpshufb {{.*#+}} xmm8 = zero,zero,zero,zero,xmm12[0,u],zero,zero,zero,zero,xmm12[1,u],zero,zero,zero,zero
; AVX1-ONLY-NEXT: vpblendw {{.*#+}} xmm8 = xmm10[0,1],xmm8[2],xmm10[3,4],xmm8[5],xmm10[6,7]
; AVX1-ONLY-NEXT: vmovdqa {{.*#+}} xmm13 = [0,1,2,3,4,128,6,7,8,9,10,128,12,13,14,15]
; AVX1-ONLY-NEXT: vpshufb %xmm13, %xmm8, %xmm11
; AVX1-ONLY-NEXT: vpshufb {{.*#+}} xmm9 = zero,zero,zero,zero,zero,xmm14[0],zero,zero,zero,zero,zero,xmm14[1],zero,zero,zero,zero
; AVX1-ONLY-NEXT: vpor %xmm9, %xmm11, %xmm11
; AVX1-ONLY-NEXT: vextractf128 $1, %ymm10, %xmm9
; AVX1-ONLY-NEXT: vmovdqa {{.*#+}} xmm10 = <2,u,128,128,128,128,3,u,128,128,128,128,4,u,128,128>
; AVX1-ONLY-NEXT: vpshufb %xmm10, %xmm12, %xmm12
; AVX1-ONLY-NEXT: vpblendw {{.*#+}} xmm9 = xmm12[0],xmm9[1,2],xmm12[3],xmm9[4,5],xmm12[6],xmm9[7]
; AVX1-ONLY-NEXT: vmovdqa {{.*#+}} xmm8 = [0,128,2,3,4,5,6,128,8,9,10,11,12,128,14,15]
; AVX1-ONLY-NEXT: vpshufb %xmm8, %xmm9, %xmm9
; AVX1-ONLY-NEXT: vmovdqa {{.*#+}} xmm12 = [128,2,128,128,128,128,128,3,128,128,128,128,128,4,128,128]
; AVX1-ONLY-NEXT: vpshufb %xmm12, %xmm14, %xmm14
; AVX1-ONLY-NEXT: vpor %xmm14, %xmm9, %xmm9
; AVX1-ONLY-NEXT: vpunpcklbw {{.*#+}} xmm3 = 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]
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm2 = xmm3[0,0,1,1]
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm14 = xmm3[1,1,2,2]
; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm14, %ymm2, %ymm2
; AVX1-ONLY-NEXT: vpunpcklbw {{.*#+}} xmm6 = xmm6[0],xmm7[0],xmm6[1],xmm7[1],xmm6[2],xmm7[2],xmm6[3],xmm7[3],xmm6[4],xmm7[4],xmm6[5],xmm7[5],xmm6[6],xmm7[6],xmm6[7],xmm7[7]
; AVX1-ONLY-NEXT: vpshuflw {{.*#+}} xmm7 = xmm6[1,0,2,2,4,5,6,7]
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm7 = xmm7[0,1,0,1]
; AVX1-ONLY-NEXT: vpshuflw {{.*#+}} xmm14 = xmm6[3,3,3,3,4,5,6,7]
; AVX1-ONLY-NEXT: vpshufhw {{.*#+}} xmm14 = xmm14[0,1,2,3,4,4,4,4]
; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm14, %ymm7, %ymm7
; AVX1-ONLY-NEXT: vandps %ymm2, %ymm15, %ymm2
; AVX1-ONLY-NEXT: vandnps %ymm7, %ymm15, %ymm7
; AVX1-ONLY-NEXT: vorps %ymm7, %ymm2, %ymm7
; AVX1-ONLY-NEXT: vpshufb %xmm10, %xmm1, %xmm2
; AVX1-ONLY-NEXT: vextractf128 $1, %ymm7, %xmm10
; AVX1-ONLY-NEXT: vpblendw {{.*#+}} xmm2 = xmm2[0],xmm10[1,2],xmm2[3],xmm10[4,5],xmm2[6],xmm10[7]
; AVX1-ONLY-NEXT: vpshufb %xmm8, %xmm2, %xmm2
; AVX1-ONLY-NEXT: vpshufb %xmm12, %xmm0, %xmm10
; AVX1-ONLY-NEXT: vpor %xmm2, %xmm10, %xmm2
; AVX1-ONLY-NEXT: vpshufb {{.*#+}} xmm10 = zero,zero,zero,zero,xmm1[0,u],zero,zero,zero,zero,xmm1[1,u],zero,zero,zero,zero
; AVX1-ONLY-NEXT: vpblendw {{.*#+}} xmm7 = xmm7[0,1],xmm10[2],xmm7[3,4],xmm10[5],xmm7[6,7]
; AVX1-ONLY-NEXT: vpshufb %xmm13, %xmm7, %xmm7
; AVX1-ONLY-NEXT: vpshufb {{.*#+}} xmm8 = zero,zero,zero,zero,zero,xmm0[0],zero,zero,zero,zero,zero,xmm0[1],zero,zero,zero,zero
; AVX1-ONLY-NEXT: vpor %xmm7, %xmm8, %xmm7
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm3 = xmm3[3,3,3,3]
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm4 = xmm4[0,0,1,1]
; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm4, %ymm3, %ymm3
; AVX1-ONLY-NEXT: vpshufhw {{.*#+}} xmm4 = xmm6[0,1,2,3,5,6,7,7]
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm4 = xmm4[2,2,2,3]
; AVX1-ONLY-NEXT: vpshuflw {{.*#+}} xmm5 = xmm5[1,0,2,2,4,5,6,7]
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm5 = xmm5[0,1,0,1]
; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm5, %ymm4, %ymm4
; AVX1-ONLY-NEXT: vmovaps {{.*#+}} ymm5 = [65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535]
; AVX1-ONLY-NEXT: vandnps %ymm3, %ymm5, %ymm3
; AVX1-ONLY-NEXT: vandps %ymm5, %ymm4, %ymm4
; AVX1-ONLY-NEXT: vorps %ymm3, %ymm4, %ymm3
; AVX1-ONLY-NEXT: vpshufb {{.*#+}} xmm4 = zero,zero,xmm1[5,u],zero,zero,zero,zero,xmm1[6,u],zero,zero,zero,zero,xmm1[7,u]
; AVX1-ONLY-NEXT: vpblendw {{.*#+}} xmm4 = xmm3[0],xmm4[1],xmm3[2,3],xmm4[4],xmm3[5,6],xmm4[7]
; AVX1-ONLY-NEXT: vpshufb {{.*#+}} xmm4 = xmm4[0,1,2],zero,xmm4[4,5,6,7,8],zero,xmm4[10,11,12,13,14],zero
; AVX1-ONLY-NEXT: vpshufb {{.*#+}} xmm5 = zero,zero,zero,xmm0[5],zero,zero,zero,zero,zero,xmm0[6],zero,zero,zero,zero,zero,xmm0[7]
; AVX1-ONLY-NEXT: vpor %xmm5, %xmm4, %xmm4
; AVX1-ONLY-NEXT: vpshufb {{.*#+}} xmm1 = zero,zero,zero,zero,xmm1[8,u],zero,zero,zero,zero,xmm1[9,u],zero,zero,zero,zero
; AVX1-ONLY-NEXT: vextractf128 $1, %ymm3, %xmm3
; AVX1-ONLY-NEXT: vpblendw {{.*#+}} xmm1 = xmm3[0,1],xmm1[2],xmm3[3,4],xmm1[5],xmm3[6,7]
; AVX1-ONLY-NEXT: vpshufb %xmm13, %xmm1, %xmm1
; AVX1-ONLY-NEXT: vpshufb {{.*#+}} xmm0 = zero,zero,zero,zero,zero,xmm0[8],zero,zero,zero,zero,zero,xmm0[9],zero,zero,zero,zero
; AVX1-ONLY-NEXT: vpor %xmm0, %xmm1, %xmm0
; AVX1-ONLY-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX1-ONLY-NEXT: vmovdqa %xmm0, 48(%rax)
; AVX1-ONLY-NEXT: vmovdqa %xmm4, 32(%rax)
; AVX1-ONLY-NEXT: vmovdqa %xmm7, (%rax)
; AVX1-ONLY-NEXT: vmovdqa %xmm2, 16(%rax)
; AVX1-ONLY-NEXT: vmovdqa %xmm9, 112(%rax)
; AVX1-ONLY-NEXT: vmovdqa %xmm11, 96(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm0, 64(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm0, 80(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm0, 176(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm0, 160(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm0, 128(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm0, 144(%rax)
; AVX1-ONLY-NEXT: vzeroupper
; AVX1-ONLY-NEXT: retq
;
; AVX2-SLOW-LABEL: store_i8_stride6_vf32:
; AVX2-SLOW: # %bb.0:
; AVX2-SLOW-NEXT: vmovdqa (%rdi), %ymm0
; AVX2-SLOW-NEXT: vmovdqa (%rsi), %ymm2
; AVX2-SLOW-NEXT: vmovdqa (%rdx), %ymm1
; AVX2-SLOW-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT: vmovdqa (%rcx), %ymm3
; AVX2-SLOW-NEXT: vmovdqa (%r8), %ymm4
; AVX2-SLOW-NEXT: vmovdqa (%rcx), %xmm6
; AVX2-SLOW-NEXT: vpbroadcastq {{.*#+}} xmm5 = [5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10]
; AVX2-SLOW-NEXT: vpshufb %xmm5, %xmm6, %xmm8
; AVX2-SLOW-NEXT: vmovdqa %xmm6, %xmm7
; AVX2-SLOW-NEXT: vmovdqa (%rdx), %xmm6
; AVX2-SLOW-NEXT: vpshufb %xmm5, %xmm6, %xmm5
; AVX2-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm5 = xmm5[0],xmm8[0],xmm5[1],xmm8[1],xmm5[2],xmm8[2],xmm5[3],xmm8[3],xmm5[4],xmm8[4],xmm5[5],xmm8[5],xmm5[6],xmm8[6],xmm5[7],xmm8[7]
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,0,0,1]
; AVX2-SLOW-NEXT: vmovdqa (%rsi), %xmm9
; AVX2-SLOW-NEXT: vpbroadcastq {{.*#+}} xmm8 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
; AVX2-SLOW-NEXT: vpshufb %xmm8, %xmm9, %xmm10
; AVX2-SLOW-NEXT: vmovdqa (%rdi), %xmm11
; AVX2-SLOW-NEXT: vpshufb %xmm8, %xmm11, %xmm8
; AVX2-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm8 = xmm8[8],xmm10[8],xmm8[9],xmm10[9],xmm8[10],xmm10[10],xmm8[11],xmm10[11],xmm8[12],xmm10[12],xmm8[13],xmm10[13],xmm8[14],xmm10[14],xmm8[15],xmm10[15]
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm8 = ymm8[0,0,0,1]
; AVX2-SLOW-NEXT: vmovdqa {{.*#+}} ymm10 = <255,255,u,u,0,0,255,255,u,u,0,0,255,255,u,u,0,0,255,255,u,u,0,0,255,255,u,u,0,0,255,255>
; AVX2-SLOW-NEXT: vpblendvb %ymm10, %ymm5, %ymm8, %ymm8
; AVX2-SLOW-NEXT: vmovdqa (%r8), %xmm5
; AVX2-SLOW-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-SLOW-NEXT: vpshufb {{.*#+}} xmm12 = xmm5[6,u,5,u,8,u,7,u,9,u,9,u,9,u,9,u]
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm12 = ymm12[0,0,0,1]
; AVX2-SLOW-NEXT: vmovdqa {{.*#+}} ymm13 = [255,255,0,0,255,255,255,255,0,0,255,255,255,255,0,0,255,255,255,255,0,0,255,255,255,255,0,0,255,255,255,255]
; AVX2-SLOW-NEXT: vpblendvb %ymm13, %ymm8, %ymm12, %ymm14
; AVX2-SLOW-NEXT: vpbroadcastq {{.*#+}} ymm8 = [5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10]
; AVX2-SLOW-NEXT: vpshufb %ymm8, %ymm3, %ymm12
; AVX2-SLOW-NEXT: vmovdqa %ymm3, %ymm5
; AVX2-SLOW-NEXT: vpshufb %ymm8, %ymm1, %ymm8
; AVX2-SLOW-NEXT: vpunpcklbw {{.*#+}} ymm12 = ymm8[0],ymm12[0],ymm8[1],ymm12[1],ymm8[2],ymm12[2],ymm8[3],ymm12[3],ymm8[4],ymm12[4],ymm8[5],ymm12[5],ymm8[6],ymm12[6],ymm8[7],ymm12[7],ymm8[16],ymm12[16],ymm8[17],ymm12[17],ymm8[18],ymm12[18],ymm8[19],ymm12[19],ymm8[20],ymm12[20],ymm8[21],ymm12[21],ymm8[22],ymm12[22],ymm8[23],ymm12[23]
; AVX2-SLOW-NEXT: vpbroadcastq {{.*#+}} ymm8 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
; AVX2-SLOW-NEXT: vpshufb %ymm8, %ymm2, %ymm15
; AVX2-SLOW-NEXT: vpshufb %ymm8, %ymm0, %ymm8
; AVX2-SLOW-NEXT: vmovdqa %ymm0, %ymm1
; AVX2-SLOW-NEXT: vpunpcklbw {{.*#+}} ymm15 = ymm8[0],ymm15[0],ymm8[1],ymm15[1],ymm8[2],ymm15[2],ymm8[3],ymm15[3],ymm8[4],ymm15[4],ymm8[5],ymm15[5],ymm8[6],ymm15[6],ymm8[7],ymm15[7],ymm8[16],ymm15[16],ymm8[17],ymm15[17],ymm8[18],ymm15[18],ymm8[19],ymm15[19],ymm8[20],ymm15[20],ymm8[21],ymm15[21],ymm8[22],ymm15[22],ymm8[23],ymm15[23]
; AVX2-SLOW-NEXT: vmovdqa (%r9), %ymm8
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,2,2,3]
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,2,2,3]
; AVX2-SLOW-NEXT: vpblendvb %ymm10, %ymm12, %ymm15, %ymm10
; AVX2-SLOW-NEXT: vmovdqa (%r9), %xmm12
; AVX2-SLOW-NEXT: vpshufb {{.*#+}} ymm15 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,u,21,u,24,u,23,u,u,u,25,u,u,u,u,u]
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,2,2,3]
; AVX2-SLOW-NEXT: vpblendvb %ymm13, %ymm10, %ymm15, %ymm13
; AVX2-SLOW-NEXT: vpshufb {{.*#+}} xmm10 = xmm12[u,6,u,5,u,8,u,7,u,9,u,9,u,9,u,9]
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,0,0,1]
; AVX2-SLOW-NEXT: vmovdqa {{.*#+}} ymm15 = [255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255]
; AVX2-SLOW-NEXT: vpblendvb %ymm15, %ymm14, %ymm10, %ymm0
; AVX2-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT: vpshufb {{.*#+}} ymm14 = ymm8[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,u,21,u,24,u,23,u,u,u,25,u,u,u,u]
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,2,2,3]
; AVX2-SLOW-NEXT: vpblendvb %ymm15, %ymm13, %ymm14, %ymm0
; AVX2-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm14 = xmm11[0],xmm9[0],xmm11[1],xmm9[1],xmm11[2],xmm9[2],xmm11[3],xmm9[3],xmm11[4],xmm9[4],xmm11[5],xmm9[5],xmm11[6],xmm9[6],xmm11[7],xmm9[7]
; AVX2-SLOW-NEXT: vpshuflw {{.*#+}} xmm14 = xmm14[0,3,2,1,4,5,6,7]
; AVX2-SLOW-NEXT: vpshufhw {{.*#+}} xmm14 = xmm14[0,1,2,3,4,5,6,5]
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm14 = ymm14[0,0,0,1]
; AVX2-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm15 = xmm6[0],xmm7[0],xmm6[1],xmm7[1],xmm6[2],xmm7[2],xmm6[3],xmm7[3],xmm6[4],xmm7[4],xmm6[5],xmm7[5],xmm6[6],xmm7[6],xmm6[7],xmm7[7]
; AVX2-SLOW-NEXT: vmovdqa %xmm7, %xmm13
; AVX2-SLOW-NEXT: vpshuflw {{.*#+}} xmm15 = xmm15[1,0,3,2,4,5,6,7]
; AVX2-SLOW-NEXT: vpshufhw {{.*#+}} xmm15 = xmm15[0,1,2,3,4,4,4,4]
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm15 = ymm15[0,0,0,1]
; AVX2-SLOW-NEXT: vmovdqa {{.*#+}} ymm0 = <255,255,0,0,u,u,255,255,0,0,u,u,255,255,0,0,u,u,255,255,0,0,u,u,255,255,0,0,u,u,255,255>
; AVX2-SLOW-NEXT: vpblendvb %ymm0, %ymm14, %ymm15, %ymm14
; AVX2-SLOW-NEXT: vmovdqa %ymm2, %ymm7
; AVX2-SLOW-NEXT: vmovdqa %ymm1, %ymm2
; AVX2-SLOW-NEXT: vpunpcklbw {{.*#+}} ymm15 = ymm1[0],ymm7[0],ymm1[1],ymm7[1],ymm1[2],ymm7[2],ymm1[3],ymm7[3],ymm1[4],ymm7[4],ymm1[5],ymm7[5],ymm1[6],ymm7[6],ymm1[7],ymm7[7],ymm1[16],ymm7[16],ymm1[17],ymm7[17],ymm1[18],ymm7[18],ymm1[19],ymm7[19],ymm1[20],ymm7[20],ymm1[21],ymm7[21],ymm1[22],ymm7[22],ymm1[23],ymm7[23]
; AVX2-SLOW-NEXT: vpshuflw {{.*#+}} ymm15 = ymm15[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
; AVX2-SLOW-NEXT: vpshufhw {{.*#+}} ymm15 = ymm15[0,1,2,3,4,5,6,5,8,9,10,11,12,13,14,13]
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,2,2,3]
; AVX2-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-SLOW-NEXT: vpunpcklbw {{.*#+}} ymm10 = ymm1[0],ymm3[0],ymm1[1],ymm3[1],ymm1[2],ymm3[2],ymm1[3],ymm3[3],ymm1[4],ymm3[4],ymm1[5],ymm3[5],ymm1[6],ymm3[6],ymm1[7],ymm3[7],ymm1[16],ymm3[16],ymm1[17],ymm3[17],ymm1[18],ymm3[18],ymm1[19],ymm3[19],ymm1[20],ymm3[20],ymm1[21],ymm3[21],ymm1[22],ymm3[22],ymm1[23],ymm3[23]
; AVX2-SLOW-NEXT: vpshuflw {{.*#+}} ymm10 = ymm10[1,0,3,2,4,5,6,7,9,8,11,10,12,13,14,15]
; AVX2-SLOW-NEXT: vpshufhw {{.*#+}} ymm10 = ymm10[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12]
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,2,2,3]
; AVX2-SLOW-NEXT: vpblendvb %ymm0, %ymm15, %ymm10, %ymm0
; AVX2-SLOW-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-SLOW-NEXT: vpshufb {{.*#+}} xmm10 = xmm3[2,u,1,u,0,u,3,u,4,u,4,u,4,u,4,u]
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,0,0,1]
; AVX2-SLOW-NEXT: vmovdqa {{.*#+}} ymm15 = [255,255,255,255,0,0,255,255,255,255,0,0,255,255,255,255,0,0,255,255,255,255,0,0,255,255,255,255,0,0,255,255]
; AVX2-SLOW-NEXT: vpblendvb %ymm15, %ymm14, %ymm10, %ymm10
; AVX2-SLOW-NEXT: vpshufb {{.*#+}} ymm14 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,u,17,u,16,u,19,u,u,u,u,u,20,u,u,u]
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,2,2,3]
; AVX2-SLOW-NEXT: vpblendvb %ymm15, %ymm0, %ymm14, %ymm0
; AVX2-SLOW-NEXT: vpshufb {{.*#+}} xmm14 = xmm12[u,2,u,1,u,0,u,3,u,4,u,4,u,4,u,4]
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm14 = ymm14[0,0,0,1]
; AVX2-SLOW-NEXT: vmovdqa {{.*#+}} ymm15 = [255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255]
; AVX2-SLOW-NEXT: vpblendvb %ymm15, %ymm10, %ymm14, %ymm14
; AVX2-SLOW-NEXT: vpshufb {{.*#+}} ymm10 = ymm8[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,u,17,u,16,u,19,u,u,u,u,u,20,u,u]
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,2,2,3]
; AVX2-SLOW-NEXT: vpblendvb %ymm15, %ymm0, %ymm10, %ymm15
; AVX2-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm11[8],xmm9[8],xmm11[9],xmm9[9],xmm11[10],xmm9[10],xmm11[11],xmm9[11],xmm11[12],xmm9[12],xmm11[13],xmm9[13],xmm11[14],xmm9[14],xmm11[15],xmm9[15]
; AVX2-SLOW-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[8,9,6,7,12,13,10,11,14,15,14,15,14,15,14,15]
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,0,0,1]
; AVX2-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm6 = xmm6[8],xmm13[8],xmm6[9],xmm13[9],xmm6[10],xmm13[10],xmm6[11],xmm13[11],xmm6[12],xmm13[12],xmm6[13],xmm13[13],xmm6[14],xmm13[14],xmm6[15],xmm13[15]
; AVX2-SLOW-NEXT: vpshufb {{.*#+}} xmm6 = xmm6[10,11,8,9,6,7,12,13,14,15,14,15,14,15,14,15]
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,0,0,1]
; AVX2-SLOW-NEXT: vmovdqa {{.*#+}} ymm9 = <u,u,255,255,0,0,u,u,255,255,0,0,u,u,255,255,0,0,u,u,255,255,0,0,u,u,255,255,0,0,u,u>
; AVX2-SLOW-NEXT: vpblendvb %ymm9, %ymm0, %ymm6, %ymm0
; AVX2-SLOW-NEXT: vpunpckhbw {{.*#+}} ymm2 = ymm2[8],ymm7[8],ymm2[9],ymm7[9],ymm2[10],ymm7[10],ymm2[11],ymm7[11],ymm2[12],ymm7[12],ymm2[13],ymm7[13],ymm2[14],ymm7[14],ymm2[15],ymm7[15],ymm2[24],ymm7[24],ymm2[25],ymm7[25],ymm2[26],ymm7[26],ymm2[27],ymm7[27],ymm2[28],ymm7[28],ymm2[29],ymm7[29],ymm2[30],ymm7[30],ymm2[31],ymm7[31]
; AVX2-SLOW-NEXT: vpunpckhbw {{.*#+}} ymm1 = ymm1[8],ymm5[8],ymm1[9],ymm5[9],ymm1[10],ymm5[10],ymm1[11],ymm5[11],ymm1[12],ymm5[12],ymm1[13],ymm5[13],ymm1[14],ymm5[14],ymm1[15],ymm5[15],ymm1[24],ymm5[24],ymm1[25],ymm5[25],ymm1[26],ymm5[26],ymm1[27],ymm5[27],ymm1[28],ymm5[28],ymm1[29],ymm5[29],ymm1[30],ymm5[30],ymm1[31],ymm5[31]
; AVX2-SLOW-NEXT: vpshufb {{.*#+}} ymm2 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,22,23,28,29,26,27,30,31,30,31,30,31,30,31]
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
; AVX2-SLOW-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,26,27,24,25,22,23,28,29,30,31,30,31,30,31,30,31]
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
; AVX2-SLOW-NEXT: vpblendvb %ymm9, %ymm2, %ymm1, %ymm1
; AVX2-SLOW-NEXT: vpshufb {{.*#+}} xmm2 = xmm3[10,u,13,u,12,u,11,u,14,u,13,u,14,u,15,u]
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,0,0,1]
; AVX2-SLOW-NEXT: vmovdqa {{.*#+}} ymm3 = [0,0,255,255,255,255,0,0,255,255,255,255,0,0,255,255,255,255,0,0,255,255,255,255,0,0,255,255,255,255,0,0]
; AVX2-SLOW-NEXT: vpblendvb %ymm3, %ymm0, %ymm2, %ymm0
; AVX2-SLOW-NEXT: vpshufb {{.*#+}} ymm2 = ymm4[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,26,u,29,u,28,u,27,u,30,u,u,u,u,u,31,u]
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
; AVX2-SLOW-NEXT: vpblendvb %ymm3, %ymm1, %ymm2, %ymm1
; AVX2-SLOW-NEXT: vpshufb {{.*#+}} xmm2 = xmm12[u,10,u,13,u,12,u,11,u,14,u,13,u,14,u,15]
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,0,0,1]
; AVX2-SLOW-NEXT: vmovdqa {{.*#+}} ymm3 = [255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0]
; AVX2-SLOW-NEXT: vpblendvb %ymm3, %ymm0, %ymm2, %ymm0
; AVX2-SLOW-NEXT: vpshufb {{.*#+}} ymm2 = ymm8[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,26,u,29,u,28,u,27,u,30,u,u,u,u,u,31]
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
; AVX2-SLOW-NEXT: vpblendvb %ymm3, %ymm1, %ymm2, %ymm1
; AVX2-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX2-SLOW-NEXT: vmovaps %ymm2, 128(%rax)
; AVX2-SLOW-NEXT: vmovdqa %ymm1, 160(%rax)
; AVX2-SLOW-NEXT: vmovdqa %ymm0, 64(%rax)
; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-SLOW-NEXT: vmovaps %ymm0, 32(%rax)
; AVX2-SLOW-NEXT: vmovdqa %ymm15, 96(%rax)
; AVX2-SLOW-NEXT: vmovdqa %ymm14, (%rax)
; AVX2-SLOW-NEXT: vzeroupper
; AVX2-SLOW-NEXT: retq
;
; AVX2-FAST-LABEL: store_i8_stride6_vf32:
; AVX2-FAST: # %bb.0:
; AVX2-FAST-NEXT: vmovdqa (%rdi), %ymm1
; AVX2-FAST-NEXT: vmovdqa (%rsi), %ymm3
; AVX2-FAST-NEXT: vmovdqa (%rdx), %ymm2
; AVX2-FAST-NEXT: vmovdqa (%rcx), %ymm4
; AVX2-FAST-NEXT: vmovdqa (%r8), %ymm6
; AVX2-FAST-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT: vmovdqa (%rcx), %xmm0
; AVX2-FAST-NEXT: vpbroadcastq {{.*#+}} xmm7 = [5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10]
; AVX2-FAST-NEXT: vpshufb %xmm7, %xmm0, %xmm8
; AVX2-FAST-NEXT: vmovdqa (%rdx), %xmm5
; AVX2-FAST-NEXT: vpshufb %xmm7, %xmm5, %xmm7
; AVX2-FAST-NEXT: vpunpcklbw {{.*#+}} xmm7 = xmm7[0],xmm8[0],xmm7[1],xmm8[1],xmm7[2],xmm8[2],xmm7[3],xmm8[3],xmm7[4],xmm8[4],xmm7[5],xmm8[5],xmm7[6],xmm8[6],xmm7[7],xmm8[7]
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm8 = ymm7[0,0,0,1]
; AVX2-FAST-NEXT: vmovdqa (%rsi), %xmm7
; AVX2-FAST-NEXT: vpbroadcastq {{.*#+}} xmm10 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
; AVX2-FAST-NEXT: vpshufb %xmm10, %xmm7, %xmm11
; AVX2-FAST-NEXT: vmovdqa %xmm7, %xmm9
; AVX2-FAST-NEXT: vmovdqa (%rdi), %xmm7
; AVX2-FAST-NEXT: vpshufb %xmm10, %xmm7, %xmm10
; AVX2-FAST-NEXT: vpunpckhbw {{.*#+}} xmm10 = xmm10[8],xmm11[8],xmm10[9],xmm11[9],xmm10[10],xmm11[10],xmm10[11],xmm11[11],xmm10[12],xmm11[12],xmm10[13],xmm11[13],xmm10[14],xmm11[14],xmm10[15],xmm11[15]
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,0,0,1]
; AVX2-FAST-NEXT: vmovdqa {{.*#+}} ymm11 = <255,255,u,u,0,0,255,255,u,u,0,0,255,255,u,u,0,0,255,255,u,u,0,0,255,255,u,u,0,0,255,255>
; AVX2-FAST-NEXT: vpblendvb %ymm11, %ymm8, %ymm10, %ymm10
; AVX2-FAST-NEXT: vmovdqa (%r8), %xmm8
; AVX2-FAST-NEXT: vmovdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-NEXT: vpshufb {{.*#+}} xmm12 = xmm8[6,u,5,u,8,u,7,u,9,u,9,u,9,u,9,u]
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm12 = ymm12[0,0,0,1]
; AVX2-FAST-NEXT: vmovdqa {{.*#+}} ymm13 = [255,255,0,0,255,255,255,255,0,0,255,255,255,255,0,0,255,255,255,255,0,0,255,255,255,255,0,0,255,255,255,255]
; AVX2-FAST-NEXT: vpblendvb %ymm13, %ymm10, %ymm12, %ymm14
; AVX2-FAST-NEXT: vpbroadcastq {{.*#+}} ymm10 = [5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10]
; AVX2-FAST-NEXT: vpshufb %ymm10, %ymm4, %ymm12
; AVX2-FAST-NEXT: vpshufb %ymm10, %ymm2, %ymm10
; AVX2-FAST-NEXT: vpunpcklbw {{.*#+}} ymm12 = ymm10[0],ymm12[0],ymm10[1],ymm12[1],ymm10[2],ymm12[2],ymm10[3],ymm12[3],ymm10[4],ymm12[4],ymm10[5],ymm12[5],ymm10[6],ymm12[6],ymm10[7],ymm12[7],ymm10[16],ymm12[16],ymm10[17],ymm12[17],ymm10[18],ymm12[18],ymm10[19],ymm12[19],ymm10[20],ymm12[20],ymm10[21],ymm12[21],ymm10[22],ymm12[22],ymm10[23],ymm12[23]
; AVX2-FAST-NEXT: vpbroadcastq {{.*#+}} ymm10 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
; AVX2-FAST-NEXT: vpshufb %ymm10, %ymm3, %ymm15
; AVX2-FAST-NEXT: vpshufb %ymm10, %ymm1, %ymm10
; AVX2-FAST-NEXT: vpunpcklbw {{.*#+}} ymm15 = ymm10[0],ymm15[0],ymm10[1],ymm15[1],ymm10[2],ymm15[2],ymm10[3],ymm15[3],ymm10[4],ymm15[4],ymm10[5],ymm15[5],ymm10[6],ymm15[6],ymm10[7],ymm15[7],ymm10[16],ymm15[16],ymm10[17],ymm15[17],ymm10[18],ymm15[18],ymm10[19],ymm15[19],ymm10[20],ymm15[20],ymm10[21],ymm15[21],ymm10[22],ymm15[22],ymm10[23],ymm15[23]
; AVX2-FAST-NEXT: vmovdqa (%r9), %ymm10
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,2,2,3]
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,2,2,3]
; AVX2-FAST-NEXT: vpblendvb %ymm11, %ymm12, %ymm15, %ymm11
; AVX2-FAST-NEXT: vmovdqa (%r9), %xmm12
; AVX2-FAST-NEXT: vpshufb {{.*#+}} ymm15 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,u,21,u,24,u,23,u,u,u,25,u,u,u,u,u]
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,2,2,3]
; AVX2-FAST-NEXT: vpblendvb %ymm13, %ymm11, %ymm15, %ymm13
; AVX2-FAST-NEXT: vpshufb {{.*#+}} xmm11 = xmm12[u,6,u,5,u,8,u,7,u,9,u,9,u,9,u,9]
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm11 = ymm11[0,0,0,1]
; AVX2-FAST-NEXT: vmovdqa {{.*#+}} ymm15 = [255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255]
; AVX2-FAST-NEXT: vpblendvb %ymm15, %ymm14, %ymm11, %ymm6
; AVX2-FAST-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT: vpshufb {{.*#+}} ymm14 = ymm10[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,u,21,u,24,u,23,u,u,u,25,u,u,u,u]
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,2,2,3]
; AVX2-FAST-NEXT: vpblendvb %ymm15, %ymm13, %ymm14, %ymm6
; AVX2-FAST-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT: vpunpckhbw {{.*#+}} xmm14 = xmm7[8],xmm9[8],xmm7[9],xmm9[9],xmm7[10],xmm9[10],xmm7[11],xmm9[11],xmm7[12],xmm9[12],xmm7[13],xmm9[13],xmm7[14],xmm9[14],xmm7[15],xmm9[15]
; AVX2-FAST-NEXT: vmovdqa %xmm9, %xmm13
; AVX2-FAST-NEXT: vpshufb {{.*#+}} xmm14 = xmm14[8,9,6,7,12,13,10,11,14,15,14,15,14,15,14,15]
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm14 = ymm14[0,0,0,1]
; AVX2-FAST-NEXT: vmovdqa %xmm0, %xmm9
; AVX2-FAST-NEXT: vmovdqa %xmm5, %xmm8
; AVX2-FAST-NEXT: vpunpckhbw {{.*#+}} xmm15 = xmm5[8],xmm0[8],xmm5[9],xmm0[9],xmm5[10],xmm0[10],xmm5[11],xmm0[11],xmm5[12],xmm0[12],xmm5[13],xmm0[13],xmm5[14],xmm0[14],xmm5[15],xmm0[15]
; AVX2-FAST-NEXT: vpshufb {{.*#+}} xmm15 = xmm15[10,11,8,9,6,7,12,13,14,15,14,15,14,15,14,15]
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm15 = ymm15[0,0,0,1]
; AVX2-FAST-NEXT: vmovdqa {{.*#+}} ymm0 = <u,u,255,255,0,0,u,u,255,255,0,0,u,u,255,255,0,0,u,u,255,255,0,0,u,u,255,255,0,0,u,u>
; AVX2-FAST-NEXT: vpblendvb %ymm0, %ymm14, %ymm15, %ymm14
; AVX2-FAST-NEXT: vpunpckhbw {{.*#+}} ymm15 = ymm1[8],ymm3[8],ymm1[9],ymm3[9],ymm1[10],ymm3[10],ymm1[11],ymm3[11],ymm1[12],ymm3[12],ymm1[13],ymm3[13],ymm1[14],ymm3[14],ymm1[15],ymm3[15],ymm1[24],ymm3[24],ymm1[25],ymm3[25],ymm1[26],ymm3[26],ymm1[27],ymm3[27],ymm1[28],ymm3[28],ymm1[29],ymm3[29],ymm1[30],ymm3[30],ymm1[31],ymm3[31]
; AVX2-FAST-NEXT: vpshufb {{.*#+}} ymm15 = ymm15[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,22,23,28,29,26,27,30,31,30,31,30,31,30,31]
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,2,2,3]
; AVX2-FAST-NEXT: vpunpckhbw {{.*#+}} ymm11 = ymm2[8],ymm4[8],ymm2[9],ymm4[9],ymm2[10],ymm4[10],ymm2[11],ymm4[11],ymm2[12],ymm4[12],ymm2[13],ymm4[13],ymm2[14],ymm4[14],ymm2[15],ymm4[15],ymm2[24],ymm4[24],ymm2[25],ymm4[25],ymm2[26],ymm4[26],ymm2[27],ymm4[27],ymm2[28],ymm4[28],ymm2[29],ymm4[29],ymm2[30],ymm4[30],ymm2[31],ymm4[31]
; AVX2-FAST-NEXT: vpshufb {{.*#+}} ymm11 = ymm11[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,26,27,24,25,22,23,28,29,30,31,30,31,30,31,30,31]
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,2,2,3]
; AVX2-FAST-NEXT: vpblendvb %ymm0, %ymm15, %ymm11, %ymm5
; AVX2-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-FAST-NEXT: vpshufb {{.*#+}} xmm11 = xmm6[10,u,13,u,12,u,11,u,14,u,13,u,14,u,15,u]
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm11 = ymm11[0,0,0,1]
; AVX2-FAST-NEXT: vmovdqa {{.*#+}} ymm15 = [0,0,255,255,255,255,0,0,255,255,255,255,0,0,255,255,255,255,0,0,255,255,255,255,0,0,255,255,255,255,0,0]
; AVX2-FAST-NEXT: vpblendvb %ymm15, %ymm14, %ymm11, %ymm11
; AVX2-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-NEXT: vpshufb {{.*#+}} ymm14 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,26,u,29,u,28,u,27,u,30,u,u,u,u,u,31,u]
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,2,2,3]
; AVX2-FAST-NEXT: vpblendvb %ymm15, %ymm5, %ymm14, %ymm5
; AVX2-FAST-NEXT: vpshufb {{.*#+}} xmm14 = xmm12[u,10,u,13,u,12,u,11,u,14,u,13,u,14,u,15]
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm14 = ymm14[0,0,0,1]
; AVX2-FAST-NEXT: vmovdqa {{.*#+}} ymm15 = [255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0]
; AVX2-FAST-NEXT: vpblendvb %ymm15, %ymm11, %ymm14, %ymm11
; AVX2-FAST-NEXT: vpshufb {{.*#+}} ymm14 = ymm10[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,26,u,29,u,28,u,27,u,30,u,u,u,u,u,31]
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,2,2,3]
; AVX2-FAST-NEXT: vpblendvb %ymm15, %ymm5, %ymm14, %ymm14
; AVX2-FAST-NEXT: vpunpcklbw {{.*#+}} xmm7 = xmm7[0],xmm13[0],xmm7[1],xmm13[1],xmm7[2],xmm13[2],xmm7[3],xmm13[3],xmm7[4],xmm13[4],xmm7[5],xmm13[5],xmm7[6],xmm13[6],xmm7[7],xmm13[7]
; AVX2-FAST-NEXT: vpshufb {{.*#+}} xmm7 = xmm7[0,1,6,7,4,5,2,3,8,9,10,11,12,13,10,11]
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,0,0,1]
; AVX2-FAST-NEXT: vpunpcklbw {{.*#+}} xmm5 = xmm8[0],xmm9[0],xmm8[1],xmm9[1],xmm8[2],xmm9[2],xmm8[3],xmm9[3],xmm8[4],xmm9[4],xmm8[5],xmm9[5],xmm8[6],xmm9[6],xmm8[7],xmm9[7]
; AVX2-FAST-NEXT: vpshufb {{.*#+}} xmm5 = xmm5[2,3,0,1,6,7,4,5,8,9,8,9,8,9,8,9]
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,0,0,1]
; AVX2-FAST-NEXT: vmovdqa {{.*#+}} ymm8 = <255,255,0,0,u,u,255,255,0,0,u,u,255,255,0,0,u,u,255,255,0,0,u,u,255,255,0,0,u,u,255,255>
; AVX2-FAST-NEXT: vpblendvb %ymm8, %ymm7, %ymm5, %ymm5
; AVX2-FAST-NEXT: vpunpcklbw {{.*#+}} ymm1 = ymm1[0],ymm3[0],ymm1[1],ymm3[1],ymm1[2],ymm3[2],ymm1[3],ymm3[3],ymm1[4],ymm3[4],ymm1[5],ymm3[5],ymm1[6],ymm3[6],ymm1[7],ymm3[7],ymm1[16],ymm3[16],ymm1[17],ymm3[17],ymm1[18],ymm3[18],ymm1[19],ymm3[19],ymm1[20],ymm3[20],ymm1[21],ymm3[21],ymm1[22],ymm3[22],ymm1[23],ymm3[23]
; AVX2-FAST-NEXT: vpunpcklbw {{.*#+}} ymm2 = ymm2[0],ymm4[0],ymm2[1],ymm4[1],ymm2[2],ymm4[2],ymm2[3],ymm4[3],ymm2[4],ymm4[4],ymm2[5],ymm4[5],ymm2[6],ymm4[6],ymm2[7],ymm4[7],ymm2[16],ymm4[16],ymm2[17],ymm4[17],ymm2[18],ymm4[18],ymm2[19],ymm4[19],ymm2[20],ymm4[20],ymm2[21],ymm4[21],ymm2[22],ymm4[22],ymm2[23],ymm4[23]
; AVX2-FAST-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,16,17,22,23,20,21,18,19,24,25,26,27,28,29,26,27]
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
; AVX2-FAST-NEXT: vpshufb {{.*#+}} ymm2 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,16,17,22,23,20,21,24,25,24,25,24,25,24,25]
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
; AVX2-FAST-NEXT: vpblendvb %ymm8, %ymm1, %ymm2, %ymm1
; AVX2-FAST-NEXT: vpshufb {{.*#+}} xmm2 = xmm6[2,u,1,u,0,u,3,u,4,u,4,u,4,u,4,u]
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,0,0,1]
; AVX2-FAST-NEXT: vmovdqa {{.*#+}} ymm3 = [255,255,255,255,0,0,255,255,255,255,0,0,255,255,255,255,0,0,255,255,255,255,0,0,255,255,255,255,0,0,255,255]
; AVX2-FAST-NEXT: vpblendvb %ymm3, %ymm5, %ymm2, %ymm2
; AVX2-FAST-NEXT: vpshufb {{.*#+}} ymm4 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,u,17,u,16,u,19,u,u,u,u,u,20,u,u,u]
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,2,3]
; AVX2-FAST-NEXT: vpblendvb %ymm3, %ymm1, %ymm4, %ymm1
; AVX2-FAST-NEXT: vpshufb {{.*#+}} xmm3 = xmm12[u,2,u,1,u,0,u,3,u,4,u,4,u,4,u,4]
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,0,1]
; AVX2-FAST-NEXT: vmovdqa {{.*#+}} ymm4 = [255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255]
; AVX2-FAST-NEXT: vpblendvb %ymm4, %ymm2, %ymm3, %ymm2
; AVX2-FAST-NEXT: vpshufb {{.*#+}} ymm3 = ymm10[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,u,17,u,16,u,19,u,u,u,u,u,20,u,u]
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3]
; AVX2-FAST-NEXT: vpblendvb %ymm4, %ymm1, %ymm3, %ymm1
; AVX2-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-NEXT: vmovaps %ymm0, 128(%rax)
; AVX2-FAST-NEXT: vmovdqa %ymm1, 96(%rax)
; AVX2-FAST-NEXT: vmovdqa %ymm14, 160(%rax)
; AVX2-FAST-NEXT: vmovdqa %ymm11, 64(%rax)
; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-NEXT: vmovaps %ymm0, 32(%rax)
; AVX2-FAST-NEXT: vmovdqa %ymm2, (%rax)
; AVX2-FAST-NEXT: vzeroupper
; AVX2-FAST-NEXT: retq
;
; AVX2-FAST-PERLANE-LABEL: store_i8_stride6_vf32:
; AVX2-FAST-PERLANE: # %bb.0:
; AVX2-FAST-PERLANE-NEXT: vmovdqa (%rdi), %ymm1
; AVX2-FAST-PERLANE-NEXT: vmovdqa (%rsi), %ymm3
; AVX2-FAST-PERLANE-NEXT: vmovdqa (%rdx), %ymm2
; AVX2-FAST-PERLANE-NEXT: vmovdqa (%rcx), %ymm4
; AVX2-FAST-PERLANE-NEXT: vmovdqa (%r8), %ymm6
; AVX2-FAST-PERLANE-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT: vmovdqa (%rcx), %xmm0
; AVX2-FAST-PERLANE-NEXT: vpbroadcastq {{.*#+}} xmm7 = [5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10]
; AVX2-FAST-PERLANE-NEXT: vpshufb %xmm7, %xmm0, %xmm8
; AVX2-FAST-PERLANE-NEXT: vmovdqa (%rdx), %xmm5
; AVX2-FAST-PERLANE-NEXT: vpshufb %xmm7, %xmm5, %xmm7
; AVX2-FAST-PERLANE-NEXT: vpunpcklbw {{.*#+}} xmm7 = xmm7[0],xmm8[0],xmm7[1],xmm8[1],xmm7[2],xmm8[2],xmm7[3],xmm8[3],xmm7[4],xmm8[4],xmm7[5],xmm8[5],xmm7[6],xmm8[6],xmm7[7],xmm8[7]
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm8 = ymm7[0,0,0,1]
; AVX2-FAST-PERLANE-NEXT: vmovdqa (%rsi), %xmm7
; AVX2-FAST-PERLANE-NEXT: vpbroadcastq {{.*#+}} xmm10 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
; AVX2-FAST-PERLANE-NEXT: vpshufb %xmm10, %xmm7, %xmm11
; AVX2-FAST-PERLANE-NEXT: vmovdqa %xmm7, %xmm9
; AVX2-FAST-PERLANE-NEXT: vmovdqa (%rdi), %xmm7
; AVX2-FAST-PERLANE-NEXT: vpshufb %xmm10, %xmm7, %xmm10
; AVX2-FAST-PERLANE-NEXT: vpunpckhbw {{.*#+}} xmm10 = xmm10[8],xmm11[8],xmm10[9],xmm11[9],xmm10[10],xmm11[10],xmm10[11],xmm11[11],xmm10[12],xmm11[12],xmm10[13],xmm11[13],xmm10[14],xmm11[14],xmm10[15],xmm11[15]
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,0,0,1]
; AVX2-FAST-PERLANE-NEXT: vmovdqa {{.*#+}} ymm11 = <255,255,u,u,0,0,255,255,u,u,0,0,255,255,u,u,0,0,255,255,u,u,0,0,255,255,u,u,0,0,255,255>
; AVX2-FAST-PERLANE-NEXT: vpblendvb %ymm11, %ymm8, %ymm10, %ymm10
; AVX2-FAST-PERLANE-NEXT: vmovdqa (%r8), %xmm8
; AVX2-FAST-PERLANE-NEXT: vmovdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-PERLANE-NEXT: vpshufb {{.*#+}} xmm12 = xmm8[6,u,5,u,8,u,7,u,9,u,9,u,9,u,9,u]
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm12 = ymm12[0,0,0,1]
; AVX2-FAST-PERLANE-NEXT: vmovdqa {{.*#+}} ymm13 = [255,255,0,0,255,255,255,255,0,0,255,255,255,255,0,0,255,255,255,255,0,0,255,255,255,255,0,0,255,255,255,255]
; AVX2-FAST-PERLANE-NEXT: vpblendvb %ymm13, %ymm10, %ymm12, %ymm14
; AVX2-FAST-PERLANE-NEXT: vpbroadcastq {{.*#+}} ymm10 = [5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10]
; AVX2-FAST-PERLANE-NEXT: vpshufb %ymm10, %ymm4, %ymm12
; AVX2-FAST-PERLANE-NEXT: vpshufb %ymm10, %ymm2, %ymm10
; AVX2-FAST-PERLANE-NEXT: vpunpcklbw {{.*#+}} ymm12 = ymm10[0],ymm12[0],ymm10[1],ymm12[1],ymm10[2],ymm12[2],ymm10[3],ymm12[3],ymm10[4],ymm12[4],ymm10[5],ymm12[5],ymm10[6],ymm12[6],ymm10[7],ymm12[7],ymm10[16],ymm12[16],ymm10[17],ymm12[17],ymm10[18],ymm12[18],ymm10[19],ymm12[19],ymm10[20],ymm12[20],ymm10[21],ymm12[21],ymm10[22],ymm12[22],ymm10[23],ymm12[23]
; AVX2-FAST-PERLANE-NEXT: vpbroadcastq {{.*#+}} ymm10 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
; AVX2-FAST-PERLANE-NEXT: vpshufb %ymm10, %ymm3, %ymm15
; AVX2-FAST-PERLANE-NEXT: vpshufb %ymm10, %ymm1, %ymm10
; AVX2-FAST-PERLANE-NEXT: vpunpcklbw {{.*#+}} ymm15 = ymm10[0],ymm15[0],ymm10[1],ymm15[1],ymm10[2],ymm15[2],ymm10[3],ymm15[3],ymm10[4],ymm15[4],ymm10[5],ymm15[5],ymm10[6],ymm15[6],ymm10[7],ymm15[7],ymm10[16],ymm15[16],ymm10[17],ymm15[17],ymm10[18],ymm15[18],ymm10[19],ymm15[19],ymm10[20],ymm15[20],ymm10[21],ymm15[21],ymm10[22],ymm15[22],ymm10[23],ymm15[23]
; AVX2-FAST-PERLANE-NEXT: vmovdqa (%r9), %ymm10
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,2,2,3]
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,2,2,3]
; AVX2-FAST-PERLANE-NEXT: vpblendvb %ymm11, %ymm12, %ymm15, %ymm11
; AVX2-FAST-PERLANE-NEXT: vmovdqa (%r9), %xmm12
; AVX2-FAST-PERLANE-NEXT: vpshufb {{.*#+}} ymm15 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,u,21,u,24,u,23,u,u,u,25,u,u,u,u,u]
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,2,2,3]
; AVX2-FAST-PERLANE-NEXT: vpblendvb %ymm13, %ymm11, %ymm15, %ymm13
; AVX2-FAST-PERLANE-NEXT: vpshufb {{.*#+}} xmm11 = xmm12[u,6,u,5,u,8,u,7,u,9,u,9,u,9,u,9]
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm11 = ymm11[0,0,0,1]
; AVX2-FAST-PERLANE-NEXT: vmovdqa {{.*#+}} ymm15 = [255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255]
; AVX2-FAST-PERLANE-NEXT: vpblendvb %ymm15, %ymm14, %ymm11, %ymm6
; AVX2-FAST-PERLANE-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT: vpshufb {{.*#+}} ymm14 = ymm10[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,22,u,21,u,24,u,23,u,u,u,25,u,u,u,u]
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,2,2,3]
; AVX2-FAST-PERLANE-NEXT: vpblendvb %ymm15, %ymm13, %ymm14, %ymm6
; AVX2-FAST-PERLANE-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT: vpunpckhbw {{.*#+}} xmm14 = xmm7[8],xmm9[8],xmm7[9],xmm9[9],xmm7[10],xmm9[10],xmm7[11],xmm9[11],xmm7[12],xmm9[12],xmm7[13],xmm9[13],xmm7[14],xmm9[14],xmm7[15],xmm9[15]
; AVX2-FAST-PERLANE-NEXT: vmovdqa %xmm9, %xmm13
; AVX2-FAST-PERLANE-NEXT: vpshufb {{.*#+}} xmm14 = xmm14[8,9,6,7,12,13,10,11,14,15,14,15,14,15,14,15]
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm14 = ymm14[0,0,0,1]
; AVX2-FAST-PERLANE-NEXT: vmovdqa %xmm0, %xmm9
; AVX2-FAST-PERLANE-NEXT: vmovdqa %xmm5, %xmm8
; AVX2-FAST-PERLANE-NEXT: vpunpckhbw {{.*#+}} xmm15 = xmm5[8],xmm0[8],xmm5[9],xmm0[9],xmm5[10],xmm0[10],xmm5[11],xmm0[11],xmm5[12],xmm0[12],xmm5[13],xmm0[13],xmm5[14],xmm0[14],xmm5[15],xmm0[15]
; AVX2-FAST-PERLANE-NEXT: vpshufb {{.*#+}} xmm15 = xmm15[10,11,8,9,6,7,12,13,14,15,14,15,14,15,14,15]
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm15 = ymm15[0,0,0,1]
; AVX2-FAST-PERLANE-NEXT: vmovdqa {{.*#+}} ymm0 = <u,u,255,255,0,0,u,u,255,255,0,0,u,u,255,255,0,0,u,u,255,255,0,0,u,u,255,255,0,0,u,u>
; AVX2-FAST-PERLANE-NEXT: vpblendvb %ymm0, %ymm14, %ymm15, %ymm14
; AVX2-FAST-PERLANE-NEXT: vpunpckhbw {{.*#+}} ymm15 = ymm1[8],ymm3[8],ymm1[9],ymm3[9],ymm1[10],ymm3[10],ymm1[11],ymm3[11],ymm1[12],ymm3[12],ymm1[13],ymm3[13],ymm1[14],ymm3[14],ymm1[15],ymm3[15],ymm1[24],ymm3[24],ymm1[25],ymm3[25],ymm1[26],ymm3[26],ymm1[27],ymm3[27],ymm1[28],ymm3[28],ymm1[29],ymm3[29],ymm1[30],ymm3[30],ymm1[31],ymm3[31]
; AVX2-FAST-PERLANE-NEXT: vpshufb {{.*#+}} ymm15 = ymm15[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,22,23,28,29,26,27,30,31,30,31,30,31,30,31]
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,2,2,3]
; AVX2-FAST-PERLANE-NEXT: vpunpckhbw {{.*#+}} ymm11 = ymm2[8],ymm4[8],ymm2[9],ymm4[9],ymm2[10],ymm4[10],ymm2[11],ymm4[11],ymm2[12],ymm4[12],ymm2[13],ymm4[13],ymm2[14],ymm4[14],ymm2[15],ymm4[15],ymm2[24],ymm4[24],ymm2[25],ymm4[25],ymm2[26],ymm4[26],ymm2[27],ymm4[27],ymm2[28],ymm4[28],ymm2[29],ymm4[29],ymm2[30],ymm4[30],ymm2[31],ymm4[31]
; AVX2-FAST-PERLANE-NEXT: vpshufb {{.*#+}} ymm11 = ymm11[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,26,27,24,25,22,23,28,29,30,31,30,31,30,31,30,31]
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,2,2,3]
; AVX2-FAST-PERLANE-NEXT: vpblendvb %ymm0, %ymm15, %ymm11, %ymm5
; AVX2-FAST-PERLANE-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-FAST-PERLANE-NEXT: vpshufb {{.*#+}} xmm11 = xmm6[10,u,13,u,12,u,11,u,14,u,13,u,14,u,15,u]
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm11 = ymm11[0,0,0,1]
; AVX2-FAST-PERLANE-NEXT: vmovdqa {{.*#+}} ymm15 = [0,0,255,255,255,255,0,0,255,255,255,255,0,0,255,255,255,255,0,0,255,255,255,255,0,0,255,255,255,255,0,0]
; AVX2-FAST-PERLANE-NEXT: vpblendvb %ymm15, %ymm14, %ymm11, %ymm11
; AVX2-FAST-PERLANE-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT: vpshufb {{.*#+}} ymm14 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,26,u,29,u,28,u,27,u,30,u,u,u,u,u,31,u]
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,2,2,3]
; AVX2-FAST-PERLANE-NEXT: vpblendvb %ymm15, %ymm5, %ymm14, %ymm5
; AVX2-FAST-PERLANE-NEXT: vpshufb {{.*#+}} xmm14 = xmm12[u,10,u,13,u,12,u,11,u,14,u,13,u,14,u,15]
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm14 = ymm14[0,0,0,1]
; AVX2-FAST-PERLANE-NEXT: vmovdqa {{.*#+}} ymm15 = [255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0]
; AVX2-FAST-PERLANE-NEXT: vpblendvb %ymm15, %ymm11, %ymm14, %ymm11
; AVX2-FAST-PERLANE-NEXT: vpshufb {{.*#+}} ymm14 = ymm10[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,26,u,29,u,28,u,27,u,30,u,u,u,u,u,31]
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,2,2,3]
; AVX2-FAST-PERLANE-NEXT: vpblendvb %ymm15, %ymm5, %ymm14, %ymm14
; AVX2-FAST-PERLANE-NEXT: vpunpcklbw {{.*#+}} xmm7 = xmm7[0],xmm13[0],xmm7[1],xmm13[1],xmm7[2],xmm13[2],xmm7[3],xmm13[3],xmm7[4],xmm13[4],xmm7[5],xmm13[5],xmm7[6],xmm13[6],xmm7[7],xmm13[7]
; AVX2-FAST-PERLANE-NEXT: vpshufb {{.*#+}} xmm7 = xmm7[0,1,6,7,4,5,2,3,8,9,10,11,12,13,10,11]
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,0,0,1]
; AVX2-FAST-PERLANE-NEXT: vpunpcklbw {{.*#+}} xmm5 = xmm8[0],xmm9[0],xmm8[1],xmm9[1],xmm8[2],xmm9[2],xmm8[3],xmm9[3],xmm8[4],xmm9[4],xmm8[5],xmm9[5],xmm8[6],xmm9[6],xmm8[7],xmm9[7]
; AVX2-FAST-PERLANE-NEXT: vpshufb {{.*#+}} xmm5 = xmm5[2,3,0,1,6,7,4,5,8,9,8,9,8,9,8,9]
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,0,0,1]
; AVX2-FAST-PERLANE-NEXT: vmovdqa {{.*#+}} ymm8 = <255,255,0,0,u,u,255,255,0,0,u,u,255,255,0,0,u,u,255,255,0,0,u,u,255,255,0,0,u,u,255,255>
; AVX2-FAST-PERLANE-NEXT: vpblendvb %ymm8, %ymm7, %ymm5, %ymm5
; AVX2-FAST-PERLANE-NEXT: vpunpcklbw {{.*#+}} ymm1 = ymm1[0],ymm3[0],ymm1[1],ymm3[1],ymm1[2],ymm3[2],ymm1[3],ymm3[3],ymm1[4],ymm3[4],ymm1[5],ymm3[5],ymm1[6],ymm3[6],ymm1[7],ymm3[7],ymm1[16],ymm3[16],ymm1[17],ymm3[17],ymm1[18],ymm3[18],ymm1[19],ymm3[19],ymm1[20],ymm3[20],ymm1[21],ymm3[21],ymm1[22],ymm3[22],ymm1[23],ymm3[23]
; AVX2-FAST-PERLANE-NEXT: vpunpcklbw {{.*#+}} ymm2 = ymm2[0],ymm4[0],ymm2[1],ymm4[1],ymm2[2],ymm4[2],ymm2[3],ymm4[3],ymm2[4],ymm4[4],ymm2[5],ymm4[5],ymm2[6],ymm4[6],ymm2[7],ymm4[7],ymm2[16],ymm4[16],ymm2[17],ymm4[17],ymm2[18],ymm4[18],ymm2[19],ymm4[19],ymm2[20],ymm4[20],ymm2[21],ymm4[21],ymm2[22],ymm4[22],ymm2[23],ymm4[23]
; AVX2-FAST-PERLANE-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,16,17,22,23,20,21,18,19,24,25,26,27,28,29,26,27]
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
; AVX2-FAST-PERLANE-NEXT: vpshufb {{.*#+}} ymm2 = ymm2[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,19,16,17,22,23,20,21,24,25,24,25,24,25,24,25]
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
; AVX2-FAST-PERLANE-NEXT: vpblendvb %ymm8, %ymm1, %ymm2, %ymm1
; AVX2-FAST-PERLANE-NEXT: vpshufb {{.*#+}} xmm2 = xmm6[2,u,1,u,0,u,3,u,4,u,4,u,4,u,4,u]
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,0,0,1]
; AVX2-FAST-PERLANE-NEXT: vmovdqa {{.*#+}} ymm3 = [255,255,255,255,0,0,255,255,255,255,0,0,255,255,255,255,0,0,255,255,255,255,0,0,255,255,255,255,0,0,255,255]
; AVX2-FAST-PERLANE-NEXT: vpblendvb %ymm3, %ymm5, %ymm2, %ymm2
; AVX2-FAST-PERLANE-NEXT: vpshufb {{.*#+}} ymm4 = ymm0[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,u,17,u,16,u,19,u,u,u,u,u,20,u,u,u]
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,2,3]
; AVX2-FAST-PERLANE-NEXT: vpblendvb %ymm3, %ymm1, %ymm4, %ymm1
; AVX2-FAST-PERLANE-NEXT: vpshufb {{.*#+}} xmm3 = xmm12[u,2,u,1,u,0,u,3,u,4,u,4,u,4,u,4]
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,0,1]
; AVX2-FAST-PERLANE-NEXT: vmovdqa {{.*#+}} ymm4 = [255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255]
; AVX2-FAST-PERLANE-NEXT: vpblendvb %ymm4, %ymm2, %ymm3, %ymm2
; AVX2-FAST-PERLANE-NEXT: vpshufb {{.*#+}} ymm3 = ymm10[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,18,u,17,u,16,u,19,u,u,u,u,u,20,u,u]
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3]
; AVX2-FAST-PERLANE-NEXT: vpblendvb %ymm4, %ymm1, %ymm3, %ymm1
; AVX2-FAST-PERLANE-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 128(%rax)
; AVX2-FAST-PERLANE-NEXT: vmovdqa %ymm1, 96(%rax)
; AVX2-FAST-PERLANE-NEXT: vmovdqa %ymm14, 160(%rax)
; AVX2-FAST-PERLANE-NEXT: vmovdqa %ymm11, 64(%rax)
; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 32(%rax)
; AVX2-FAST-PERLANE-NEXT: vmovdqa %ymm2, (%rax)
; AVX2-FAST-PERLANE-NEXT: vzeroupper
; AVX2-FAST-PERLANE-NEXT: retq
;
; AVX512F-SLOW-LABEL: store_i8_stride6_vf32:
; AVX512F-SLOW: # %bb.0:
; AVX512F-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512F-SLOW-NEXT: vmovdqa (%rdi), %ymm11
; AVX512F-SLOW-NEXT: vmovdqa (%rsi), %ymm13
; AVX512F-SLOW-NEXT: vmovdqa (%rdx), %ymm10
; AVX512F-SLOW-NEXT: vmovdqa (%rcx), %ymm12
; AVX512F-SLOW-NEXT: vmovdqa (%r8), %ymm8
; AVX512F-SLOW-NEXT: vmovdqa (%r9), %ymm9
; AVX512F-SLOW-NEXT: vmovdqa (%rcx), %xmm1
; AVX512F-SLOW-NEXT: vmovdqa (%rdx), %xmm2
; AVX512F-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm2[8],xmm1[8],xmm2[9],xmm1[9],xmm2[10],xmm1[10],xmm2[11],xmm1[11],xmm2[12],xmm1[12],xmm2[13],xmm1[13],xmm2[14],xmm1[14],xmm2[15],xmm1[15]
; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[10,11,8,9,6,7,12,13,14,15,14,15,14,15,14,15]
; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,0,0,1]
; AVX512F-SLOW-NEXT: vpunpcklbw {{.*#+}} ymm3 = ymm10[0],ymm12[0],ymm10[1],ymm12[1],ymm10[2],ymm12[2],ymm10[3],ymm12[3],ymm10[4],ymm12[4],ymm10[5],ymm12[5],ymm10[6],ymm12[6],ymm10[7],ymm12[7],ymm10[16],ymm12[16],ymm10[17],ymm12[17],ymm10[18],ymm12[18],ymm10[19],ymm12[19],ymm10[20],ymm12[20],ymm10[21],ymm12[21],ymm10[22],ymm12[22],ymm10[23],ymm12[23]
; AVX512F-SLOW-NEXT: vprold $16, %ymm3, %ymm3
; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3]
; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm3, %zmm0, %zmm5
; AVX512F-SLOW-NEXT: vmovdqa (%rsi), %xmm3
; AVX512F-SLOW-NEXT: vmovdqa (%rdi), %xmm4
; AVX512F-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm4[8],xmm3[8],xmm4[9],xmm3[9],xmm4[10],xmm3[10],xmm4[11],xmm3[11],xmm4[12],xmm3[12],xmm4[13],xmm3[13],xmm4[14],xmm3[14],xmm4[15],xmm3[15]
; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[8,9,6,7,12,13,10,11,14,15,14,15,14,15,14,15]
; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,0,0,1]
; AVX512F-SLOW-NEXT: vpunpcklbw {{.*#+}} ymm6 = ymm11[0],ymm13[0],ymm11[1],ymm13[1],ymm11[2],ymm13[2],ymm11[3],ymm13[3],ymm11[4],ymm13[4],ymm11[5],ymm13[5],ymm11[6],ymm13[6],ymm11[7],ymm13[7],ymm11[16],ymm13[16],ymm11[17],ymm13[17],ymm11[18],ymm13[18],ymm11[19],ymm13[19],ymm11[20],ymm13[20],ymm11[21],ymm13[21],ymm11[22],ymm13[22],ymm11[23],ymm13[23]
; AVX512F-SLOW-NEXT: vpshuflw {{.*#+}} ymm6 = ymm6[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} ymm6 = ymm6[0,1,2,3,4,5,6,5,8,9,10,11,12,13,14,13]
; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,2,2,3]
; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm6, %zmm0, %zmm14
; AVX512F-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm17 = [65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535]
; AVX512F-SLOW-NEXT: vpternlogq $226, %zmm5, %zmm17, %zmm14
; AVX512F-SLOW-NEXT: vmovdqa (%r9), %xmm5
; AVX512F-SLOW-NEXT: vmovdqa (%r8), %xmm6
; AVX512F-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm7 = xmm6[8],xmm5[8],xmm6[9],xmm5[9],xmm6[10],xmm5[10],xmm6[11],xmm5[11],xmm6[12],xmm5[12],xmm6[13],xmm5[13],xmm6[14],xmm5[14],xmm6[15],xmm5[15]
; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} xmm7 = xmm7[4,5,10,11,8,9,6,7,12,13,10,11,12,13,14,15]
; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,0,0,1]
; AVX512F-SLOW-NEXT: vpunpcklbw {{.*#+}} ymm15 = ymm8[0],ymm9[0],ymm8[1],ymm9[1],ymm8[2],ymm9[2],ymm8[3],ymm9[3],ymm8[4],ymm9[4],ymm8[5],ymm9[5],ymm8[6],ymm9[6],ymm8[7],ymm9[7],ymm8[16],ymm9[16],ymm8[17],ymm9[17],ymm8[18],ymm9[18],ymm8[19],ymm9[19],ymm8[20],ymm9[20],ymm8[21],ymm9[21],ymm8[22],ymm9[22],ymm8[23],ymm9[23]
; AVX512F-SLOW-NEXT: vpshuflw {{.*#+}} ymm15 = ymm15[2,1,0,3,4,5,6,7,10,9,8,11,12,13,14,15]
; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} ymm15 = ymm15[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12]
; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,2,2,3]
; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm15, %zmm7, %zmm7
; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm14, %zmm7
; AVX512F-SLOW-NEXT: vpbroadcastq {{.*#+}} ymm14 = [5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10]
; AVX512F-SLOW-NEXT: vpshufb %ymm14, %ymm12, %ymm15
; AVX512F-SLOW-NEXT: vpshufb %ymm14, %ymm10, %ymm14
; AVX512F-SLOW-NEXT: vpunpcklbw {{.*#+}} ymm14 = ymm14[0],ymm15[0],ymm14[1],ymm15[1],ymm14[2],ymm15[2],ymm14[3],ymm15[3],ymm14[4],ymm15[4],ymm14[5],ymm15[5],ymm14[6],ymm15[6],ymm14[7],ymm15[7],ymm14[16],ymm15[16],ymm14[17],ymm15[17],ymm14[18],ymm15[18],ymm14[19],ymm15[19],ymm14[20],ymm15[20],ymm14[21],ymm15[21],ymm14[22],ymm15[22],ymm14[23],ymm15[23]
; AVX512F-SLOW-NEXT: vpunpckhbw {{.*#+}} ymm15 = ymm11[8],ymm13[8],ymm11[9],ymm13[9],ymm11[10],ymm13[10],ymm11[11],ymm13[11],ymm11[12],ymm13[12],ymm11[13],ymm13[13],ymm11[14],ymm13[14],ymm11[15],ymm13[15],ymm11[24],ymm13[24],ymm11[25],ymm13[25],ymm11[26],ymm13[26],ymm11[27],ymm13[27],ymm11[28],ymm13[28],ymm11[29],ymm13[29],ymm11[30],ymm13[30],ymm11[31],ymm13[31]
; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm15 = ymm15[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,22,23,28,29,26,27,30,31,30,31,30,31,30,31]
; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm15, %zmm14, %zmm14
; AVX512F-SLOW-NEXT: vpermq {{.*#+}} zmm14 = zmm14[2,2,2,3,6,6,6,7]
; AVX512F-SLOW-NEXT: vbroadcasti64x4 {{.*#+}} zmm15 = [65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535]
; AVX512F-SLOW-NEXT: # zmm15 = mem[0,1,2,3,0,1,2,3]
; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm15, %zmm15, %zmm16
; AVX512F-SLOW-NEXT: vpandq %zmm16, %zmm14, %zmm14
; AVX512F-SLOW-NEXT: vpbroadcastq {{.*#+}} ymm0 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
; AVX512F-SLOW-NEXT: vpshufb %ymm0, %ymm13, %ymm13
; AVX512F-SLOW-NEXT: vpshufb %ymm0, %ymm11, %ymm0
; AVX512F-SLOW-NEXT: vpunpcklbw {{.*#+}} ymm0 = ymm0[0],ymm13[0],ymm0[1],ymm13[1],ymm0[2],ymm13[2],ymm0[3],ymm13[3],ymm0[4],ymm13[4],ymm0[5],ymm13[5],ymm0[6],ymm13[6],ymm0[7],ymm13[7],ymm0[16],ymm13[16],ymm0[17],ymm13[17],ymm0[18],ymm13[18],ymm0[19],ymm13[19],ymm0[20],ymm13[20],ymm0[21],ymm13[21],ymm0[22],ymm13[22],ymm0[23],ymm13[23]
; AVX512F-SLOW-NEXT: vpunpckhbw {{.*#+}} ymm10 = ymm10[8],ymm12[8],ymm10[9],ymm12[9],ymm10[10],ymm12[10],ymm10[11],ymm12[11],ymm10[12],ymm12[12],ymm10[13],ymm12[13],ymm10[14],ymm12[14],ymm10[15],ymm12[15],ymm10[24],ymm12[24],ymm10[25],ymm12[25],ymm10[26],ymm12[26],ymm10[27],ymm12[27],ymm10[28],ymm12[28],ymm10[29],ymm12[29],ymm10[30],ymm12[30],ymm10[31],ymm12[31]
; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm10 = ymm10[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,26,27,24,25,22,23,28,29,30,31,30,31,30,31,30,31]
; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm10, %zmm0, %zmm0
; AVX512F-SLOW-NEXT: vpermq {{.*#+}} zmm0 = zmm0[2,2,2,3,6,6,6,7]
; AVX512F-SLOW-NEXT: vpternlogq $186, %zmm14, %zmm15, %zmm0
; AVX512F-SLOW-NEXT: vpbroadcastq {{.*#+}} ymm10 = [6,5,8,7,0,9,0,0,6,5,8,7,0,9,0,0,6,5,8,7,0,9,0,0,6,5,8,7,0,9,0,0]
; AVX512F-SLOW-NEXT: vpshufb %ymm10, %ymm9, %ymm11
; AVX512F-SLOW-NEXT: vpshufb %ymm10, %ymm8, %ymm10
; AVX512F-SLOW-NEXT: vpunpcklbw {{.*#+}} ymm10 = ymm10[0],ymm11[0],ymm10[1],ymm11[1],ymm10[2],ymm11[2],ymm10[3],ymm11[3],ymm10[4],ymm11[4],ymm10[5],ymm11[5],ymm10[6],ymm11[6],ymm10[7],ymm11[7],ymm10[16],ymm11[16],ymm10[17],ymm11[17],ymm10[18],ymm11[18],ymm10[19],ymm11[19],ymm10[20],ymm11[20],ymm10[21],ymm11[21],ymm10[22],ymm11[22],ymm10[23],ymm11[23]
; AVX512F-SLOW-NEXT: vpunpckhbw {{.*#+}} ymm8 = ymm8[8],ymm9[8],ymm8[9],ymm9[9],ymm8[10],ymm9[10],ymm8[11],ymm9[11],ymm8[12],ymm9[12],ymm8[13],ymm9[13],ymm8[14],ymm9[14],ymm8[15],ymm9[15],ymm8[24],ymm9[24],ymm8[25],ymm9[25],ymm8[26],ymm9[26],ymm8[27],ymm9[27],ymm8[28],ymm9[28],ymm8[29],ymm9[29],ymm8[30],ymm9[30],ymm8[31],ymm9[31]
; AVX512F-SLOW-NEXT: vpshufb {{.*#+}} ymm8 = ymm8[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,20,21,26,27,24,25,22,23,28,29,26,27,28,29,30,31]
; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm8, %zmm10, %zmm8
; AVX512F-SLOW-NEXT: vpermq {{.*#+}} zmm8 = zmm8[2,2,2,3,6,6,6,7]
; AVX512F-SLOW-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm8
; AVX512F-SLOW-NEXT: vpbroadcastq {{.*#+}} xmm0 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
; AVX512F-SLOW-NEXT: vpshufb %xmm0, %xmm3, %xmm9
; AVX512F-SLOW-NEXT: vpshufb %xmm0, %xmm4, %xmm0
; AVX512F-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm0[8],xmm9[8],xmm0[9],xmm9[9],xmm0[10],xmm9[10],xmm0[11],xmm9[11],xmm0[12],xmm9[12],xmm0[13],xmm9[13],xmm0[14],xmm9[14],xmm0[15],xmm9[15]
; AVX512F-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm9 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3],xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
; AVX512F-SLOW-NEXT: vprold $16, %xmm9, %xmm9
; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm9, %zmm0
; AVX512F-SLOW-NEXT: vpermq {{.*#+}} zmm0 = zmm0[0,0,0,1,4,4,4,5]
; AVX512F-SLOW-NEXT: vpbroadcastq {{.*#+}} xmm9 = [5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10]
; AVX512F-SLOW-NEXT: vpshufb %xmm9, %xmm1, %xmm1
; AVX512F-SLOW-NEXT: vpshufb %xmm9, %xmm2, %xmm2
; AVX512F-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3],xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
; AVX512F-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm2 = 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]
; AVX512F-SLOW-NEXT: vpshuflw {{.*#+}} xmm2 = xmm2[0,3,2,1,4,5,6,7]
; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,5,6,5]
; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm1
; AVX512F-SLOW-NEXT: vpermq {{.*#+}} zmm1 = zmm1[0,0,0,1,4,4,4,5]
; AVX512F-SLOW-NEXT: vpternlogq $228, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm1
; AVX512F-SLOW-NEXT: vpbroadcastq {{.*#+}} xmm0 = [6,5,8,7,0,9,0,0,6,5,8,7,0,9,0,0]
; AVX512F-SLOW-NEXT: vpshufb %xmm0, %xmm5, %xmm2
; AVX512F-SLOW-NEXT: vpshufb %xmm0, %xmm6, %xmm0
; AVX512F-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],xmm2[3],xmm0[4],xmm2[4],xmm0[5],xmm2[5],xmm0[6],xmm2[6],xmm0[7],xmm2[7]
; AVX512F-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm2 = 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]
; AVX512F-SLOW-NEXT: vpshuflw {{.*#+}} xmm2 = xmm2[2,1,0,3,4,5,6,7]
; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,4,4,4]
; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm2, %zmm0
; AVX512F-SLOW-NEXT: vpermq {{.*#+}} zmm0 = zmm0[0,0,0,1,4,4,4,5]
; AVX512F-SLOW-NEXT: vpternlogq $184, %zmm1, %zmm17, %zmm0
; AVX512F-SLOW-NEXT: vmovdqa64 %zmm0, (%rax)
; AVX512F-SLOW-NEXT: vmovdqa64 %zmm8, 128(%rax)
; AVX512F-SLOW-NEXT: vmovdqa64 %zmm7, 64(%rax)
; AVX512F-SLOW-NEXT: vzeroupper
; AVX512F-SLOW-NEXT: retq
;
; AVX512F-FAST-LABEL: store_i8_stride6_vf32:
; AVX512F-FAST: # %bb.0:
; AVX512F-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512F-FAST-NEXT: vmovdqa (%rdi), %ymm11
; AVX512F-FAST-NEXT: vmovdqa (%rsi), %ymm13
; AVX512F-FAST-NEXT: vmovdqa (%rdx), %ymm10
; AVX512F-FAST-NEXT: vmovdqa (%rcx), %ymm12
; AVX512F-FAST-NEXT: vmovdqa (%r8), %ymm8
; AVX512F-FAST-NEXT: vmovdqa (%r9), %ymm9
; AVX512F-FAST-NEXT: vmovdqa (%rcx), %xmm0
; AVX512F-FAST-NEXT: vmovdqa (%rdx), %xmm1
; AVX512F-FAST-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm1[8],xmm0[8],xmm1[9],xmm0[9],xmm1[10],xmm0[10],xmm1[11],xmm0[11],xmm1[12],xmm0[12],xmm1[13],xmm0[13],xmm1[14],xmm0[14],xmm1[15],xmm0[15]
; AVX512F-FAST-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[10,11,8,9,6,7,12,13,14,15,14,15,14,15,14,15]
; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,0,0,1]
; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} ymm3 = ymm10[0],ymm12[0],ymm10[1],ymm12[1],ymm10[2],ymm12[2],ymm10[3],ymm12[3],ymm10[4],ymm12[4],ymm10[5],ymm12[5],ymm10[6],ymm12[6],ymm10[7],ymm12[7],ymm10[16],ymm12[16],ymm10[17],ymm12[17],ymm10[18],ymm12[18],ymm10[19],ymm12[19],ymm10[20],ymm12[20],ymm10[21],ymm12[21],ymm10[22],ymm12[22],ymm10[23],ymm12[23]
; AVX512F-FAST-NEXT: vprold $16, %ymm3, %ymm3
; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3]
; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm3
; AVX512F-FAST-NEXT: vmovdqa (%rsi), %xmm4
; AVX512F-FAST-NEXT: vmovdqa (%rdi), %xmm5
; AVX512F-FAST-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm5[8],xmm4[8],xmm5[9],xmm4[9],xmm5[10],xmm4[10],xmm5[11],xmm4[11],xmm5[12],xmm4[12],xmm5[13],xmm4[13],xmm5[14],xmm4[14],xmm5[15],xmm4[15]
; AVX512F-FAST-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[8,9,6,7,12,13,10,11,14,15,14,15,14,15,14,15]
; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,0,0,1]
; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} ymm6 = ymm11[0],ymm13[0],ymm11[1],ymm13[1],ymm11[2],ymm13[2],ymm11[3],ymm13[3],ymm11[4],ymm13[4],ymm11[5],ymm13[5],ymm11[6],ymm13[6],ymm11[7],ymm13[7],ymm11[16],ymm13[16],ymm11[17],ymm13[17],ymm11[18],ymm13[18],ymm11[19],ymm13[19],ymm11[20],ymm13[20],ymm11[21],ymm13[21],ymm11[22],ymm13[22],ymm11[23],ymm13[23]
; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm6 = ymm6[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,16,17,22,23,20,21,18,19,24,25,26,27,28,29,26,27]
; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,2,2,3]
; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm6, %zmm2, %zmm14
; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm17 = [65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535]
; AVX512F-FAST-NEXT: vpternlogq $226, %zmm3, %zmm17, %zmm14
; AVX512F-FAST-NEXT: vmovdqa (%r9), %xmm3
; AVX512F-FAST-NEXT: vmovdqa (%r8), %xmm6
; AVX512F-FAST-NEXT: vpunpckhbw {{.*#+}} xmm7 = xmm6[8],xmm3[8],xmm6[9],xmm3[9],xmm6[10],xmm3[10],xmm6[11],xmm3[11],xmm6[12],xmm3[12],xmm6[13],xmm3[13],xmm6[14],xmm3[14],xmm6[15],xmm3[15]
; AVX512F-FAST-NEXT: vpshufb {{.*#+}} xmm7 = xmm7[4,5,10,11,8,9,6,7,12,13,10,11,12,13,14,15]
; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,0,0,1]
; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} ymm15 = ymm8[0],ymm9[0],ymm8[1],ymm9[1],ymm8[2],ymm9[2],ymm8[3],ymm9[3],ymm8[4],ymm9[4],ymm8[5],ymm9[5],ymm8[6],ymm9[6],ymm8[7],ymm9[7],ymm8[16],ymm9[16],ymm8[17],ymm9[17],ymm8[18],ymm9[18],ymm8[19],ymm9[19],ymm8[20],ymm9[20],ymm8[21],ymm9[21],ymm8[22],ymm9[22],ymm8[23],ymm9[23]
; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm15 = ymm15[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,20,21,18,19,16,17,22,23,24,25,24,25,24,25,24,25]
; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,2,2,3]
; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm15, %zmm7, %zmm7
; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm14, %zmm7
; AVX512F-FAST-NEXT: vpbroadcastq {{.*#+}} ymm14 = [5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10]
; AVX512F-FAST-NEXT: vpshufb %ymm14, %ymm12, %ymm15
; AVX512F-FAST-NEXT: vpshufb %ymm14, %ymm10, %ymm14
; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} ymm14 = ymm14[0],ymm15[0],ymm14[1],ymm15[1],ymm14[2],ymm15[2],ymm14[3],ymm15[3],ymm14[4],ymm15[4],ymm14[5],ymm15[5],ymm14[6],ymm15[6],ymm14[7],ymm15[7],ymm14[16],ymm15[16],ymm14[17],ymm15[17],ymm14[18],ymm15[18],ymm14[19],ymm15[19],ymm14[20],ymm15[20],ymm14[21],ymm15[21],ymm14[22],ymm15[22],ymm14[23],ymm15[23]
; AVX512F-FAST-NEXT: vpunpckhbw {{.*#+}} ymm15 = ymm11[8],ymm13[8],ymm11[9],ymm13[9],ymm11[10],ymm13[10],ymm11[11],ymm13[11],ymm11[12],ymm13[12],ymm11[13],ymm13[13],ymm11[14],ymm13[14],ymm11[15],ymm13[15],ymm11[24],ymm13[24],ymm11[25],ymm13[25],ymm11[26],ymm13[26],ymm11[27],ymm13[27],ymm11[28],ymm13[28],ymm11[29],ymm13[29],ymm11[30],ymm13[30],ymm11[31],ymm13[31]
; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm15 = ymm15[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,24,25,22,23,28,29,26,27,30,31,30,31,30,31,30,31]
; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm15, %zmm14, %zmm14
; AVX512F-FAST-NEXT: vpermq {{.*#+}} zmm14 = zmm14[2,2,2,3,6,6,6,7]
; AVX512F-FAST-NEXT: vbroadcasti64x4 {{.*#+}} zmm15 = [65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535]
; AVX512F-FAST-NEXT: # zmm15 = mem[0,1,2,3,0,1,2,3]
; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm15, %zmm15, %zmm16
; AVX512F-FAST-NEXT: vpandq %zmm16, %zmm14, %zmm14
; AVX512F-FAST-NEXT: vpbroadcastq {{.*#+}} ymm2 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
; AVX512F-FAST-NEXT: vpshufb %ymm2, %ymm13, %ymm13
; AVX512F-FAST-NEXT: vpshufb %ymm2, %ymm11, %ymm2
; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} ymm2 = ymm2[0],ymm13[0],ymm2[1],ymm13[1],ymm2[2],ymm13[2],ymm2[3],ymm13[3],ymm2[4],ymm13[4],ymm2[5],ymm13[5],ymm2[6],ymm13[6],ymm2[7],ymm13[7],ymm2[16],ymm13[16],ymm2[17],ymm13[17],ymm2[18],ymm13[18],ymm2[19],ymm13[19],ymm2[20],ymm13[20],ymm2[21],ymm13[21],ymm2[22],ymm13[22],ymm2[23],ymm13[23]
; AVX512F-FAST-NEXT: vpunpckhbw {{.*#+}} ymm10 = ymm10[8],ymm12[8],ymm10[9],ymm12[9],ymm10[10],ymm12[10],ymm10[11],ymm12[11],ymm10[12],ymm12[12],ymm10[13],ymm12[13],ymm10[14],ymm12[14],ymm10[15],ymm12[15],ymm10[24],ymm12[24],ymm10[25],ymm12[25],ymm10[26],ymm12[26],ymm10[27],ymm12[27],ymm10[28],ymm12[28],ymm10[29],ymm12[29],ymm10[30],ymm12[30],ymm10[31],ymm12[31]
; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm10 = ymm10[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,26,27,24,25,22,23,28,29,30,31,30,31,30,31,30,31]
; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm10, %zmm2, %zmm2
; AVX512F-FAST-NEXT: vpermq {{.*#+}} zmm2 = zmm2[2,2,2,3,6,6,6,7]
; AVX512F-FAST-NEXT: vpternlogq $186, %zmm14, %zmm15, %zmm2
; AVX512F-FAST-NEXT: vpbroadcastq {{.*#+}} ymm10 = [6,5,8,7,0,9,0,0,6,5,8,7,0,9,0,0,6,5,8,7,0,9,0,0,6,5,8,7,0,9,0,0]
; AVX512F-FAST-NEXT: vpshufb %ymm10, %ymm9, %ymm11
; AVX512F-FAST-NEXT: vpshufb %ymm10, %ymm8, %ymm10
; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} ymm10 = ymm10[0],ymm11[0],ymm10[1],ymm11[1],ymm10[2],ymm11[2],ymm10[3],ymm11[3],ymm10[4],ymm11[4],ymm10[5],ymm11[5],ymm10[6],ymm11[6],ymm10[7],ymm11[7],ymm10[16],ymm11[16],ymm10[17],ymm11[17],ymm10[18],ymm11[18],ymm10[19],ymm11[19],ymm10[20],ymm11[20],ymm10[21],ymm11[21],ymm10[22],ymm11[22],ymm10[23],ymm11[23]
; AVX512F-FAST-NEXT: vpunpckhbw {{.*#+}} ymm8 = ymm8[8],ymm9[8],ymm8[9],ymm9[9],ymm8[10],ymm9[10],ymm8[11],ymm9[11],ymm8[12],ymm9[12],ymm8[13],ymm9[13],ymm8[14],ymm9[14],ymm8[15],ymm9[15],ymm8[24],ymm9[24],ymm8[25],ymm9[25],ymm8[26],ymm9[26],ymm8[27],ymm9[27],ymm8[28],ymm9[28],ymm8[29],ymm9[29],ymm8[30],ymm9[30],ymm8[31],ymm9[31]
; AVX512F-FAST-NEXT: vpshufb {{.*#+}} ymm8 = ymm8[u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,20,21,26,27,24,25,22,23,28,29,26,27,28,29,30,31]
; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm8, %zmm10, %zmm8
; AVX512F-FAST-NEXT: vpermq {{.*#+}} zmm8 = zmm8[2,2,2,3,6,6,6,7]
; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm8
; AVX512F-FAST-NEXT: vpbroadcastq {{.*#+}} xmm2 = [5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10]
; AVX512F-FAST-NEXT: vpshufb %xmm2, %xmm0, %xmm9
; AVX512F-FAST-NEXT: vpshufb %xmm2, %xmm1, %xmm2
; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm2[0],xmm9[0],xmm2[1],xmm9[1],xmm2[2],xmm9[2],xmm2[3],xmm9[3],xmm2[4],xmm9[4],xmm2[5],xmm9[5],xmm2[6],xmm9[6],xmm2[7],xmm9[7]
; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} xmm9 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3],xmm5[4],xmm4[4],xmm5[5],xmm4[5],xmm5[6],xmm4[6],xmm5[7],xmm4[7]
; AVX512F-FAST-NEXT: vpshufb {{.*#+}} xmm9 = xmm9[0,1,6,7,4,5,2,3,8,9,10,11,12,13,10,11]
; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm2, %zmm9, %zmm2
; AVX512F-FAST-NEXT: vpermq {{.*#+}} zmm2 = zmm2[0,0,0,1,4,4,4,5]
; AVX512F-FAST-NEXT: vpbroadcastq {{.*#+}} xmm9 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
; AVX512F-FAST-NEXT: vpshufb %xmm9, %xmm4, %xmm4
; AVX512F-FAST-NEXT: vpshufb %xmm9, %xmm5, %xmm5
; AVX512F-FAST-NEXT: vpunpckhbw {{.*#+}} xmm4 = xmm5[8],xmm4[8],xmm5[9],xmm4[9],xmm5[10],xmm4[10],xmm5[11],xmm4[11],xmm5[12],xmm4[12],xmm5[13],xmm4[13],xmm5[14],xmm4[14],xmm5[15],xmm4[15]
; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3],xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
; AVX512F-FAST-NEXT: vprold $16, %xmm0, %xmm0
; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm4, %zmm0, %zmm0
; AVX512F-FAST-NEXT: vpermq {{.*#+}} zmm0 = zmm0[0,0,0,1,4,4,4,5]
; AVX512F-FAST-NEXT: vpternlogq $216, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm2, %zmm0
; AVX512F-FAST-NEXT: vpbroadcastq {{.*#+}} xmm1 = [6,5,8,7,0,9,0,0,6,5,8,7,0,9,0,0]
; AVX512F-FAST-NEXT: vpshufb %xmm1, %xmm3, %xmm2
; AVX512F-FAST-NEXT: vpshufb %xmm1, %xmm6, %xmm1
; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3],xmm1[4],xmm2[4],xmm1[5],xmm2[5],xmm1[6],xmm2[6],xmm1[7],xmm2[7]
; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm6[0],xmm3[0],xmm6[1],xmm3[1],xmm6[2],xmm3[2],xmm6[3],xmm3[3],xmm6[4],xmm3[4],xmm6[5],xmm3[5],xmm6[6],xmm3[6],xmm6[7],xmm3[7]
; AVX512F-FAST-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[4,5,2,3,0,1,6,7,8,9,8,9,8,9,8,9]
; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm1, %zmm2, %zmm1
; AVX512F-FAST-NEXT: vpermq {{.*#+}} zmm1 = zmm1[0,0,0,1,4,4,4,5]
; AVX512F-FAST-NEXT: vpternlogq $184, %zmm0, %zmm17, %zmm1
; AVX512F-FAST-NEXT: vmovdqa64 %zmm1, (%rax)
; AVX512F-FAST-NEXT: vmovdqa64 %zmm8, 128(%rax)
; AVX512F-FAST-NEXT: vmovdqa64 %zmm7, 64(%rax)
; AVX512F-FAST-NEXT: vzeroupper
; AVX512F-FAST-NEXT: retq
;
; AVX512BW-SLOW-LABEL: store_i8_stride6_vf32:
; AVX512BW-SLOW: # %bb.0:
; AVX512BW-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-SLOW-NEXT: vmovdqa (%rdi), %ymm4
; AVX512BW-SLOW-NEXT: vmovdqa (%rsi), %ymm5
; AVX512BW-SLOW-NEXT: vmovdqa (%rdx), %ymm2
; AVX512BW-SLOW-NEXT: vmovdqa (%rcx), %ymm3
; AVX512BW-SLOW-NEXT: vmovdqa (%r8), %ymm0
; AVX512BW-SLOW-NEXT: vmovdqa (%r9), %ymm1
; AVX512BW-SLOW-NEXT: vmovdqa (%rsi), %xmm7
; AVX512BW-SLOW-NEXT: vpbroadcastq {{.*#+}} xmm6 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
; AVX512BW-SLOW-NEXT: vpshufb %xmm6, %xmm7, %xmm8
; AVX512BW-SLOW-NEXT: vmovdqa (%rdi), %xmm9
; AVX512BW-SLOW-NEXT: vpshufb %xmm6, %xmm9, %xmm6
; AVX512BW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm6 = xmm6[8],xmm8[8],xmm6[9],xmm8[9],xmm6[10],xmm8[10],xmm6[11],xmm8[11],xmm6[12],xmm8[12],xmm6[13],xmm8[13],xmm6[14],xmm8[14],xmm6[15],xmm8[15]
; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,0,0,1]
; AVX512BW-SLOW-NEXT: vmovdqa (%rcx), %xmm8
; AVX512BW-SLOW-NEXT: vpbroadcastq {{.*#+}} xmm10 = [5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10]
; AVX512BW-SLOW-NEXT: vpshufb %xmm10, %xmm8, %xmm11
; AVX512BW-SLOW-NEXT: vmovdqa (%rdx), %xmm12
; AVX512BW-SLOW-NEXT: vpshufb %xmm10, %xmm12, %xmm10
; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm10 = xmm10[0],xmm11[0],xmm10[1],xmm11[1],xmm10[2],xmm11[2],xmm10[3],xmm11[3],xmm10[4],xmm11[4],xmm10[5],xmm11[5],xmm10[6],xmm11[6],xmm10[7],xmm11[7]
; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,0,0,1]
; AVX512BW-SLOW-NEXT: movw $18724, %cx # imm = 0x4924
; AVX512BW-SLOW-NEXT: kmovd %ecx, %k1
; AVX512BW-SLOW-NEXT: vmovdqu16 %ymm6, %ymm10 {%k1}
; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm10, %zmm0, %zmm6
; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm10 = xmm9[0],xmm7[0],xmm9[1],xmm7[1],xmm9[2],xmm7[2],xmm9[3],xmm7[3],xmm9[4],xmm7[4],xmm9[5],xmm7[5],xmm9[6],xmm7[6],xmm9[7],xmm7[7]
; AVX512BW-SLOW-NEXT: vmovdqa {{.*#+}} ymm11 = [0,3,2,1,0,3,2,1,0,3,2,1,4,5,6,5]
; AVX512BW-SLOW-NEXT: vpermw %ymm10, %ymm11, %ymm10
; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm11 = xmm12[0],xmm8[0],xmm12[1],xmm8[1],xmm12[2],xmm8[2],xmm12[3],xmm8[3],xmm12[4],xmm8[4],xmm12[5],xmm8[5],xmm12[6],xmm8[6],xmm12[7],xmm8[7]
; AVX512BW-SLOW-NEXT: vprold $16, %xmm11, %xmm11
; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm11 = ymm11[0,0,0,1]
; AVX512BW-SLOW-NEXT: movw $9362, %cx # imm = 0x2492
; AVX512BW-SLOW-NEXT: kmovd %ecx, %k2
; AVX512BW-SLOW-NEXT: vmovdqu16 %ymm11, %ymm10 {%k2}
; AVX512BW-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm6 = zmm10[0,1,2,3],zmm6[4,5,6,7]
; AVX512BW-SLOW-NEXT: vmovdqa (%r9), %xmm10
; AVX512BW-SLOW-NEXT: vpbroadcastq {{.*#+}} xmm11 = [6,5,8,7,0,9,0,0,6,5,8,7,0,9,0,0]
; AVX512BW-SLOW-NEXT: vpshufb %xmm11, %xmm10, %xmm13
; AVX512BW-SLOW-NEXT: vmovdqa (%r8), %xmm14
; AVX512BW-SLOW-NEXT: vpshufb %xmm11, %xmm14, %xmm11
; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm11 = xmm11[0],xmm13[0],xmm11[1],xmm13[1],xmm11[2],xmm13[2],xmm11[3],xmm13[3],xmm11[4],xmm13[4],xmm11[5],xmm13[5],xmm11[6],xmm13[6],xmm11[7],xmm13[7]
; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm11 = ymm11[0,0,0,1]
; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm13 = xmm14[0],xmm10[0],xmm14[1],xmm10[1],xmm14[2],xmm10[2],xmm14[3],xmm10[3],xmm14[4],xmm10[4],xmm14[5],xmm10[5],xmm14[6],xmm10[6],xmm14[7],xmm10[7]
; AVX512BW-SLOW-NEXT: vmovdqa {{.*#+}} ymm15 = [2,1,0,3,2,1,0,3,2,1,0,3,4,4,4,4]
; AVX512BW-SLOW-NEXT: vpermw %ymm13, %ymm15, %ymm13
; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm11, %zmm13, %zmm11
; AVX512BW-SLOW-NEXT: movl $613566756, %ecx # imm = 0x24924924
; AVX512BW-SLOW-NEXT: kmovd %ecx, %k3
; AVX512BW-SLOW-NEXT: vmovdqu16 %zmm11, %zmm6 {%k3}
; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} ymm11 = ymm4[0],ymm5[0],ymm4[1],ymm5[1],ymm4[2],ymm5[2],ymm4[3],ymm5[3],ymm4[4],ymm5[4],ymm4[5],ymm5[5],ymm4[6],ymm5[6],ymm4[7],ymm5[7],ymm4[16],ymm5[16],ymm4[17],ymm5[17],ymm4[18],ymm5[18],ymm4[19],ymm5[19],ymm4[20],ymm5[20],ymm4[21],ymm5[21],ymm4[22],ymm5[22],ymm4[23],ymm5[23]
; AVX512BW-SLOW-NEXT: vmovdqa {{.*#+}} ymm13 = [8,11,10,9,8,11,10,9,8,11,10,9,12,13,14,13]
; AVX512BW-SLOW-NEXT: vpermw %ymm11, %ymm13, %ymm11
; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} ymm13 = ymm2[0],ymm3[0],ymm2[1],ymm3[1],ymm2[2],ymm3[2],ymm2[3],ymm3[3],ymm2[4],ymm3[4],ymm2[5],ymm3[5],ymm2[6],ymm3[6],ymm2[7],ymm3[7],ymm2[16],ymm3[16],ymm2[17],ymm3[17],ymm2[18],ymm3[18],ymm2[19],ymm3[19],ymm2[20],ymm3[20],ymm2[21],ymm3[21],ymm2[22],ymm3[22],ymm2[23],ymm3[23]
; AVX512BW-SLOW-NEXT: vprold $16, %ymm13, %ymm13
; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm13 = ymm13[2,2,2,3]
; AVX512BW-SLOW-NEXT: vmovdqu16 %ymm13, %ymm11 {%k2}
; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm11, %zmm0, %zmm11
; AVX512BW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm8 = xmm12[8],xmm8[8],xmm12[9],xmm8[9],xmm12[10],xmm8[10],xmm12[11],xmm8[11],xmm12[12],xmm8[12],xmm12[13],xmm8[13],xmm12[14],xmm8[14],xmm12[15],xmm8[15]
; AVX512BW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm7 = xmm9[8],xmm7[8],xmm9[9],xmm7[9],xmm9[10],xmm7[10],xmm9[11],xmm7[11],xmm9[12],xmm7[12],xmm9[13],xmm7[13],xmm9[14],xmm7[14],xmm9[15],xmm7[15]
; AVX512BW-SLOW-NEXT: vmovdqa {{.*#+}} ymm9 = [4,3,6,5,4,3,6,5,4,3,6,5,7,7,7,7]
; AVX512BW-SLOW-NEXT: vpermw %ymm7, %ymm9, %ymm7
; AVX512BW-SLOW-NEXT: vmovdqa {{.*#+}} ymm9 = [5,4,3,6,5,4,3,6,5,4,3,6,7,7,7,7]
; AVX512BW-SLOW-NEXT: vpermw %ymm8, %ymm9, %ymm7 {%k1}
; AVX512BW-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm7 = zmm7[0,1,2,3],zmm11[4,5,6,7]
; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} ymm8 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[4],ymm1[4],ymm0[5],ymm1[5],ymm0[6],ymm1[6],ymm0[7],ymm1[7],ymm0[16],ymm1[16],ymm0[17],ymm1[17],ymm0[18],ymm1[18],ymm0[19],ymm1[19],ymm0[20],ymm1[20],ymm0[21],ymm1[21],ymm0[22],ymm1[22],ymm0[23],ymm1[23]
; AVX512BW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm9 = xmm14[8],xmm10[8],xmm14[9],xmm10[9],xmm14[10],xmm10[10],xmm14[11],xmm10[11],xmm14[12],xmm10[12],xmm14[13],xmm10[13],xmm14[14],xmm10[14],xmm14[15],xmm10[15]
; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm8, %zmm9, %zmm8
; AVX512BW-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm9 = [2,5,4,3,2,5,4,3,2,5,4,3,6,5,6,7,26,25,24,27,26,25,24,27,26,25,24,27,28,28,28,28]
; AVX512BW-SLOW-NEXT: movl $1227133513, %ecx # imm = 0x49249249
; AVX512BW-SLOW-NEXT: kmovd %ecx, %k2
; AVX512BW-SLOW-NEXT: vpermw %zmm8, %zmm9, %zmm7 {%k2}
; AVX512BW-SLOW-NEXT: vpunpckhbw {{.*#+}} ymm8 = ymm2[8],ymm3[8],ymm2[9],ymm3[9],ymm2[10],ymm3[10],ymm2[11],ymm3[11],ymm2[12],ymm3[12],ymm2[13],ymm3[13],ymm2[14],ymm3[14],ymm2[15],ymm3[15],ymm2[24],ymm3[24],ymm2[25],ymm3[25],ymm2[26],ymm3[26],ymm2[27],ymm3[27],ymm2[28],ymm3[28],ymm2[29],ymm3[29],ymm2[30],ymm3[30],ymm2[31],ymm3[31]
; AVX512BW-SLOW-NEXT: vpunpckhbw {{.*#+}} ymm9 = ymm4[8],ymm5[8],ymm4[9],ymm5[9],ymm4[10],ymm5[10],ymm4[11],ymm5[11],ymm4[12],ymm5[12],ymm4[13],ymm5[13],ymm4[14],ymm5[14],ymm4[15],ymm5[15],ymm4[24],ymm5[24],ymm4[25],ymm5[25],ymm4[26],ymm5[26],ymm4[27],ymm5[27],ymm4[28],ymm5[28],ymm4[29],ymm5[29],ymm4[30],ymm5[30],ymm4[31],ymm5[31]
; AVX512BW-SLOW-NEXT: vmovdqa {{.*#+}} ymm10 = [12,11,14,13,12,11,14,13,12,11,14,13,15,15,15,15]
; AVX512BW-SLOW-NEXT: vpermw %ymm9, %ymm10, %ymm9
; AVX512BW-SLOW-NEXT: vmovdqa {{.*#+}} ymm10 = [13,12,11,14,13,12,11,14,13,12,11,14,15,15,15,15]
; AVX512BW-SLOW-NEXT: vpermw %ymm8, %ymm10, %ymm9 {%k1}
; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm9, %zmm0, %zmm8
; AVX512BW-SLOW-NEXT: vpbroadcastq {{.*#+}} ymm9 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
; AVX512BW-SLOW-NEXT: vpshufb %ymm9, %ymm5, %ymm5
; AVX512BW-SLOW-NEXT: vpshufb %ymm9, %ymm4, %ymm4
; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} ymm4 = ymm4[0],ymm5[0],ymm4[1],ymm5[1],ymm4[2],ymm5[2],ymm4[3],ymm5[3],ymm4[4],ymm5[4],ymm4[5],ymm5[5],ymm4[6],ymm5[6],ymm4[7],ymm5[7],ymm4[16],ymm5[16],ymm4[17],ymm5[17],ymm4[18],ymm5[18],ymm4[19],ymm5[19],ymm4[20],ymm5[20],ymm4[21],ymm5[21],ymm4[22],ymm5[22],ymm4[23],ymm5[23]
; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,2,3]
; AVX512BW-SLOW-NEXT: vpbroadcastq {{.*#+}} ymm5 = [5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10]
; AVX512BW-SLOW-NEXT: vpshufb %ymm5, %ymm3, %ymm3
; AVX512BW-SLOW-NEXT: vpshufb %ymm5, %ymm2, %ymm2
; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} ymm2 = ymm2[0],ymm3[0],ymm2[1],ymm3[1],ymm2[2],ymm3[2],ymm2[3],ymm3[3],ymm2[4],ymm3[4],ymm2[5],ymm3[5],ymm2[6],ymm3[6],ymm2[7],ymm3[7],ymm2[16],ymm3[16],ymm2[17],ymm3[17],ymm2[18],ymm3[18],ymm2[19],ymm3[19],ymm2[20],ymm3[20],ymm2[21],ymm3[21],ymm2[22],ymm3[22],ymm2[23],ymm3[23]
; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
; AVX512BW-SLOW-NEXT: vmovdqu16 %ymm4, %ymm2 {%k1}
; AVX512BW-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm2[0,1,2,3],zmm8[4,5,6,7]
; AVX512BW-SLOW-NEXT: vpbroadcastq {{.*#+}} ymm3 = [6,5,8,7,0,9,0,0,6,5,8,7,0,9,0,0,6,5,8,7,0,9,0,0,6,5,8,7,0,9,0,0]
; AVX512BW-SLOW-NEXT: vpshufb %ymm3, %ymm1, %ymm4
; AVX512BW-SLOW-NEXT: vpshufb %ymm3, %ymm0, %ymm3
; AVX512BW-SLOW-NEXT: vpunpcklbw {{.*#+}} ymm3 = ymm3[0],ymm4[0],ymm3[1],ymm4[1],ymm3[2],ymm4[2],ymm3[3],ymm4[3],ymm3[4],ymm4[4],ymm3[5],ymm4[5],ymm3[6],ymm4[6],ymm3[7],ymm4[7],ymm3[16],ymm4[16],ymm3[17],ymm4[17],ymm3[18],ymm4[18],ymm3[19],ymm4[19],ymm3[20],ymm4[20],ymm3[21],ymm4[21],ymm3[22],ymm4[22],ymm3[23],ymm4[23]
; AVX512BW-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3]
; AVX512BW-SLOW-NEXT: vpunpckhbw {{.*#+}} ymm0 = ymm0[8],ymm1[8],ymm0[9],ymm1[9],ymm0[10],ymm1[10],ymm0[11],ymm1[11],ymm0[12],ymm1[12],ymm0[13],ymm1[13],ymm0[14],ymm1[14],ymm0[15],ymm1[15],ymm0[24],ymm1[24],ymm0[25],ymm1[25],ymm0[26],ymm1[26],ymm0[27],ymm1[27],ymm0[28],ymm1[28],ymm0[29],ymm1[29],ymm0[30],ymm1[30],ymm0[31],ymm1[31]
; AVX512BW-SLOW-NEXT: vmovdqa {{.*#+}} ymm1 = [10,13,12,11,10,13,12,11,10,13,12,11,14,13,14,15]
; AVX512BW-SLOW-NEXT: vpermw %ymm0, %ymm1, %ymm0
; AVX512BW-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm3, %zmm0
; AVX512BW-SLOW-NEXT: movl $-1840700270, %ecx # imm = 0x92492492
; AVX512BW-SLOW-NEXT: kmovd %ecx, %k1
; AVX512BW-SLOW-NEXT: vmovdqu16 %zmm0, %zmm2 {%k1}
; AVX512BW-SLOW-NEXT: vmovdqa64 %zmm2, 128(%rax)
; AVX512BW-SLOW-NEXT: vmovdqa64 %zmm7, 64(%rax)
; AVX512BW-SLOW-NEXT: vmovdqa64 %zmm6, (%rax)
; AVX512BW-SLOW-NEXT: vzeroupper
; AVX512BW-SLOW-NEXT: retq
;
; AVX512BW-FAST-LABEL: store_i8_stride6_vf32:
; AVX512BW-FAST: # %bb.0:
; AVX512BW-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-FAST-NEXT: vmovdqa (%rdi), %ymm4
; AVX512BW-FAST-NEXT: vmovdqa (%rsi), %ymm5
; AVX512BW-FAST-NEXT: vmovdqa (%rdx), %ymm2
; AVX512BW-FAST-NEXT: vmovdqa (%rcx), %ymm3
; AVX512BW-FAST-NEXT: vmovdqa (%r8), %ymm0
; AVX512BW-FAST-NEXT: vmovdqa (%r9), %ymm1
; AVX512BW-FAST-NEXT: vmovdqa (%rcx), %xmm7
; AVX512BW-FAST-NEXT: vmovdqa (%rdx), %xmm8
; AVX512BW-FAST-NEXT: vpunpcklbw {{.*#+}} xmm6 = 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]
; AVX512BW-FAST-NEXT: vmovdqa (%rsi), %xmm9
; AVX512BW-FAST-NEXT: vmovdqa (%rdi), %xmm10
; AVX512BW-FAST-NEXT: vpunpcklbw {{.*#+}} xmm11 = xmm10[0],xmm9[0],xmm10[1],xmm9[1],xmm10[2],xmm9[2],xmm10[3],xmm9[3],xmm10[4],xmm9[4],xmm10[5],xmm9[5],xmm10[6],xmm9[6],xmm10[7],xmm9[7]
; AVX512BW-FAST-NEXT: vmovdqa {{.*#+}} ymm12 = [0,3,2,1,0,3,2,1,0,3,2,1,4,5,6,5]
; AVX512BW-FAST-NEXT: vpermw %ymm11, %ymm12, %ymm11
; AVX512BW-FAST-NEXT: vmovdqa {{.*#+}} ymm12 = [1,0,3,2,1,0,3,2,1,0,3,2,5,4,7,6]
; AVX512BW-FAST-NEXT: movw $9362, %cx # imm = 0x2492
; AVX512BW-FAST-NEXT: kmovd %ecx, %k2
; AVX512BW-FAST-NEXT: vpermw %ymm6, %ymm12, %ymm11 {%k2}
; AVX512BW-FAST-NEXT: vpbroadcastq {{.*#+}} xmm6 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
; AVX512BW-FAST-NEXT: vpshufb %xmm6, %xmm9, %xmm12
; AVX512BW-FAST-NEXT: vpshufb %xmm6, %xmm10, %xmm6
; AVX512BW-FAST-NEXT: vpunpckhbw {{.*#+}} xmm6 = xmm6[8],xmm12[8],xmm6[9],xmm12[9],xmm6[10],xmm12[10],xmm6[11],xmm12[11],xmm6[12],xmm12[12],xmm6[13],xmm12[13],xmm6[14],xmm12[14],xmm6[15],xmm12[15]
; AVX512BW-FAST-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,0,0,1]
; AVX512BW-FAST-NEXT: vpbroadcastq {{.*#+}} xmm12 = [5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10]
; AVX512BW-FAST-NEXT: vpshufb %xmm12, %xmm7, %xmm13
; AVX512BW-FAST-NEXT: vpshufb %xmm12, %xmm8, %xmm12
; AVX512BW-FAST-NEXT: vpunpcklbw {{.*#+}} xmm12 = xmm12[0],xmm13[0],xmm12[1],xmm13[1],xmm12[2],xmm13[2],xmm12[3],xmm13[3],xmm12[4],xmm13[4],xmm12[5],xmm13[5],xmm12[6],xmm13[6],xmm12[7],xmm13[7]
; AVX512BW-FAST-NEXT: vpermq {{.*#+}} ymm12 = ymm12[0,0,0,1]
; AVX512BW-FAST-NEXT: movw $18724, %cx # imm = 0x4924
; AVX512BW-FAST-NEXT: kmovd %ecx, %k1
; AVX512BW-FAST-NEXT: vmovdqu16 %ymm6, %ymm12 {%k1}
; AVX512BW-FAST-NEXT: vinserti64x4 $1, %ymm12, %zmm0, %zmm6
; AVX512BW-FAST-NEXT: vshufi64x2 {{.*#+}} zmm6 = zmm11[0,1,2,3],zmm6[4,5,6,7]
; AVX512BW-FAST-NEXT: vmovdqa (%r9), %xmm11
; AVX512BW-FAST-NEXT: vpbroadcastq {{.*#+}} xmm12 = [6,5,8,7,0,9,0,0,6,5,8,7,0,9,0,0]
; AVX512BW-FAST-NEXT: vpshufb %xmm12, %xmm11, %xmm13
; AVX512BW-FAST-NEXT: vmovdqa (%r8), %xmm14
; AVX512BW-FAST-NEXT: vpshufb %xmm12, %xmm14, %xmm12
; AVX512BW-FAST-NEXT: vpunpcklbw {{.*#+}} xmm12 = xmm12[0],xmm13[0],xmm12[1],xmm13[1],xmm12[2],xmm13[2],xmm12[3],xmm13[3],xmm12[4],xmm13[4],xmm12[5],xmm13[5],xmm12[6],xmm13[6],xmm12[7],xmm13[7]
; AVX512BW-FAST-NEXT: vpermq {{.*#+}} ymm12 = ymm12[0,0,0,1]
; AVX512BW-FAST-NEXT: vpunpcklbw {{.*#+}} xmm13 = xmm14[0],xmm11[0],xmm14[1],xmm11[1],xmm14[2],xmm11[2],xmm14[3],xmm11[3],xmm14[4],xmm11[4],xmm14[5],xmm11[5],xmm14[6],xmm11[6],xmm14[7],xmm11[7]
; AVX512BW-FAST-NEXT: vmovdqa {{.*#+}} ymm15 = [2,1,0,3,2,1,0,3,2,1,0,3,4,4,4,4]
; AVX512BW-FAST-NEXT: vpermw %ymm13, %ymm15, %ymm13
; AVX512BW-FAST-NEXT: vinserti64x4 $1, %ymm12, %zmm13, %zmm12
; AVX512BW-FAST-NEXT: movl $613566756, %ecx # imm = 0x24924924
; AVX512BW-FAST-NEXT: kmovd %ecx, %k3
; AVX512BW-FAST-NEXT: vmovdqu16 %zmm12, %zmm6 {%k3}
; AVX512BW-FAST-NEXT: vpunpcklbw {{.*#+}} ymm12 = ymm2[0],ymm3[0],ymm2[1],ymm3[1],ymm2[2],ymm3[2],ymm2[3],ymm3[3],ymm2[4],ymm3[4],ymm2[5],ymm3[5],ymm2[6],ymm3[6],ymm2[7],ymm3[7],ymm2[16],ymm3[16],ymm2[17],ymm3[17],ymm2[18],ymm3[18],ymm2[19],ymm3[19],ymm2[20],ymm3[20],ymm2[21],ymm3[21],ymm2[22],ymm3[22],ymm2[23],ymm3[23]
; AVX512BW-FAST-NEXT: vpunpcklbw {{.*#+}} ymm13 = ymm4[0],ymm5[0],ymm4[1],ymm5[1],ymm4[2],ymm5[2],ymm4[3],ymm5[3],ymm4[4],ymm5[4],ymm4[5],ymm5[5],ymm4[6],ymm5[6],ymm4[7],ymm5[7],ymm4[16],ymm5[16],ymm4[17],ymm5[17],ymm4[18],ymm5[18],ymm4[19],ymm5[19],ymm4[20],ymm5[20],ymm4[21],ymm5[21],ymm4[22],ymm5[22],ymm4[23],ymm5[23]
; AVX512BW-FAST-NEXT: vmovdqa {{.*#+}} ymm15 = [8,11,10,9,8,11,10,9,8,11,10,9,12,13,14,13]
; AVX512BW-FAST-NEXT: vpermw %ymm13, %ymm15, %ymm13
; AVX512BW-FAST-NEXT: vmovdqa {{.*#+}} ymm15 = [9,8,11,10,9,8,11,10,9,8,11,10,13,12,15,14]
; AVX512BW-FAST-NEXT: vpermw %ymm12, %ymm15, %ymm13 {%k2}
; AVX512BW-FAST-NEXT: vinserti64x4 $1, %ymm13, %zmm0, %zmm12
; AVX512BW-FAST-NEXT: vpunpckhbw {{.*#+}} xmm7 = 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]
; AVX512BW-FAST-NEXT: vpunpckhbw {{.*#+}} xmm8 = xmm10[8],xmm9[8],xmm10[9],xmm9[9],xmm10[10],xmm9[10],xmm10[11],xmm9[11],xmm10[12],xmm9[12],xmm10[13],xmm9[13],xmm10[14],xmm9[14],xmm10[15],xmm9[15]
; AVX512BW-FAST-NEXT: vmovdqa {{.*#+}} ymm9 = [4,3,6,5,4,3,6,5,4,3,6,5,7,7,7,7]
; AVX512BW-FAST-NEXT: vpermw %ymm8, %ymm9, %ymm8
; AVX512BW-FAST-NEXT: vmovdqa {{.*#+}} ymm9 = [5,4,3,6,5,4,3,6,5,4,3,6,7,7,7,7]
; AVX512BW-FAST-NEXT: vpermw %ymm7, %ymm9, %ymm8 {%k1}
; AVX512BW-FAST-NEXT: vshufi64x2 {{.*#+}} zmm7 = zmm8[0,1,2,3],zmm12[4,5,6,7]
; AVX512BW-FAST-NEXT: vpunpcklbw {{.*#+}} ymm8 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[4],ymm1[4],ymm0[5],ymm1[5],ymm0[6],ymm1[6],ymm0[7],ymm1[7],ymm0[16],ymm1[16],ymm0[17],ymm1[17],ymm0[18],ymm1[18],ymm0[19],ymm1[19],ymm0[20],ymm1[20],ymm0[21],ymm1[21],ymm0[22],ymm1[22],ymm0[23],ymm1[23]
; AVX512BW-FAST-NEXT: vpunpckhbw {{.*#+}} xmm9 = xmm14[8],xmm11[8],xmm14[9],xmm11[9],xmm14[10],xmm11[10],xmm14[11],xmm11[11],xmm14[12],xmm11[12],xmm14[13],xmm11[13],xmm14[14],xmm11[14],xmm14[15],xmm11[15]
; AVX512BW-FAST-NEXT: vinserti64x4 $1, %ymm8, %zmm9, %zmm8
; AVX512BW-FAST-NEXT: vmovdqa64 {{.*#+}} zmm9 = [2,5,4,3,2,5,4,3,2,5,4,3,6,5,6,7,26,25,24,27,26,25,24,27,26,25,24,27,28,28,28,28]
; AVX512BW-FAST-NEXT: movl $1227133513, %ecx # imm = 0x49249249
; AVX512BW-FAST-NEXT: kmovd %ecx, %k2
; AVX512BW-FAST-NEXT: vpermw %zmm8, %zmm9, %zmm7 {%k2}
; AVX512BW-FAST-NEXT: vpunpckhbw {{.*#+}} ymm8 = ymm2[8],ymm3[8],ymm2[9],ymm3[9],ymm2[10],ymm3[10],ymm2[11],ymm3[11],ymm2[12],ymm3[12],ymm2[13],ymm3[13],ymm2[14],ymm3[14],ymm2[15],ymm3[15],ymm2[24],ymm3[24],ymm2[25],ymm3[25],ymm2[26],ymm3[26],ymm2[27],ymm3[27],ymm2[28],ymm3[28],ymm2[29],ymm3[29],ymm2[30],ymm3[30],ymm2[31],ymm3[31]
; AVX512BW-FAST-NEXT: vpunpckhbw {{.*#+}} ymm9 = ymm4[8],ymm5[8],ymm4[9],ymm5[9],ymm4[10],ymm5[10],ymm4[11],ymm5[11],ymm4[12],ymm5[12],ymm4[13],ymm5[13],ymm4[14],ymm5[14],ymm4[15],ymm5[15],ymm4[24],ymm5[24],ymm4[25],ymm5[25],ymm4[26],ymm5[26],ymm4[27],ymm5[27],ymm4[28],ymm5[28],ymm4[29],ymm5[29],ymm4[30],ymm5[30],ymm4[31],ymm5[31]
; AVX512BW-FAST-NEXT: vmovdqa {{.*#+}} ymm10 = [12,11,14,13,12,11,14,13,12,11,14,13,15,15,15,15]
; AVX512BW-FAST-NEXT: vpermw %ymm9, %ymm10, %ymm9
; AVX512BW-FAST-NEXT: vmovdqa {{.*#+}} ymm10 = [13,12,11,14,13,12,11,14,13,12,11,14,15,15,15,15]
; AVX512BW-FAST-NEXT: vpermw %ymm8, %ymm10, %ymm9 {%k1}
; AVX512BW-FAST-NEXT: vinserti64x4 $1, %ymm9, %zmm0, %zmm8
; AVX512BW-FAST-NEXT: vpbroadcastq {{.*#+}} ymm9 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
; AVX512BW-FAST-NEXT: vpshufb %ymm9, %ymm5, %ymm5
; AVX512BW-FAST-NEXT: vpshufb %ymm9, %ymm4, %ymm4
; AVX512BW-FAST-NEXT: vpunpcklbw {{.*#+}} ymm4 = ymm4[0],ymm5[0],ymm4[1],ymm5[1],ymm4[2],ymm5[2],ymm4[3],ymm5[3],ymm4[4],ymm5[4],ymm4[5],ymm5[5],ymm4[6],ymm5[6],ymm4[7],ymm5[7],ymm4[16],ymm5[16],ymm4[17],ymm5[17],ymm4[18],ymm5[18],ymm4[19],ymm5[19],ymm4[20],ymm5[20],ymm4[21],ymm5[21],ymm4[22],ymm5[22],ymm4[23],ymm5[23]
; AVX512BW-FAST-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,2,3]
; AVX512BW-FAST-NEXT: vpbroadcastq {{.*#+}} ymm5 = [5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10]
; AVX512BW-FAST-NEXT: vpshufb %ymm5, %ymm3, %ymm3
; AVX512BW-FAST-NEXT: vpshufb %ymm5, %ymm2, %ymm2
; AVX512BW-FAST-NEXT: vpunpcklbw {{.*#+}} ymm2 = ymm2[0],ymm3[0],ymm2[1],ymm3[1],ymm2[2],ymm3[2],ymm2[3],ymm3[3],ymm2[4],ymm3[4],ymm2[5],ymm3[5],ymm2[6],ymm3[6],ymm2[7],ymm3[7],ymm2[16],ymm3[16],ymm2[17],ymm3[17],ymm2[18],ymm3[18],ymm2[19],ymm3[19],ymm2[20],ymm3[20],ymm2[21],ymm3[21],ymm2[22],ymm3[22],ymm2[23],ymm3[23]
; AVX512BW-FAST-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
; AVX512BW-FAST-NEXT: vmovdqu16 %ymm4, %ymm2 {%k1}
; AVX512BW-FAST-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm2[0,1,2,3],zmm8[4,5,6,7]
; AVX512BW-FAST-NEXT: vpbroadcastq {{.*#+}} ymm3 = [6,5,8,7,0,9,0,0,6,5,8,7,0,9,0,0,6,5,8,7,0,9,0,0,6,5,8,7,0,9,0,0]
; AVX512BW-FAST-NEXT: vpshufb %ymm3, %ymm1, %ymm4
; AVX512BW-FAST-NEXT: vpshufb %ymm3, %ymm0, %ymm3
; AVX512BW-FAST-NEXT: vpunpcklbw {{.*#+}} ymm3 = ymm3[0],ymm4[0],ymm3[1],ymm4[1],ymm3[2],ymm4[2],ymm3[3],ymm4[3],ymm3[4],ymm4[4],ymm3[5],ymm4[5],ymm3[6],ymm4[6],ymm3[7],ymm4[7],ymm3[16],ymm4[16],ymm3[17],ymm4[17],ymm3[18],ymm4[18],ymm3[19],ymm4[19],ymm3[20],ymm4[20],ymm3[21],ymm4[21],ymm3[22],ymm4[22],ymm3[23],ymm4[23]
; AVX512BW-FAST-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3]
; AVX512BW-FAST-NEXT: vpunpckhbw {{.*#+}} ymm0 = ymm0[8],ymm1[8],ymm0[9],ymm1[9],ymm0[10],ymm1[10],ymm0[11],ymm1[11],ymm0[12],ymm1[12],ymm0[13],ymm1[13],ymm0[14],ymm1[14],ymm0[15],ymm1[15],ymm0[24],ymm1[24],ymm0[25],ymm1[25],ymm0[26],ymm1[26],ymm0[27],ymm1[27],ymm0[28],ymm1[28],ymm0[29],ymm1[29],ymm0[30],ymm1[30],ymm0[31],ymm1[31]
; AVX512BW-FAST-NEXT: vmovdqa {{.*#+}} ymm1 = [10,13,12,11,10,13,12,11,10,13,12,11,14,13,14,15]
; AVX512BW-FAST-NEXT: vpermw %ymm0, %ymm1, %ymm0
; AVX512BW-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm3, %zmm0
; AVX512BW-FAST-NEXT: movl $-1840700270, %ecx # imm = 0x92492492
; AVX512BW-FAST-NEXT: kmovd %ecx, %k1
; AVX512BW-FAST-NEXT: vmovdqu16 %zmm0, %zmm2 {%k1}
; AVX512BW-FAST-NEXT: vmovdqa64 %zmm2, 128(%rax)
; AVX512BW-FAST-NEXT: vmovdqa64 %zmm7, 64(%rax)
; AVX512BW-FAST-NEXT: vmovdqa64 %zmm6, (%rax)
; AVX512BW-FAST-NEXT: vzeroupper
; AVX512BW-FAST-NEXT: retq
%in.vec0 = load <32 x i8>, ptr %in.vecptr0, align 64
%in.vec1 = load <32 x i8>, ptr %in.vecptr1, align 64
%in.vec2 = load <32 x i8>, ptr %in.vecptr2, align 64
%in.vec3 = load <32 x i8>, ptr %in.vecptr3, align 64
%in.vec4 = load <32 x i8>, ptr %in.vecptr4, align 64
%in.vec5 = load <32 x i8>, ptr %in.vecptr5, align 64
%1 = shufflevector <32 x i8> %in.vec0, <32 x i8> %in.vec1, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
%2 = shufflevector <32 x i8> %in.vec2, <32 x i8> %in.vec3, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
%3 = shufflevector <32 x i8> %in.vec4, <32 x i8> %in.vec5, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
%4 = shufflevector <64 x i8> %1, <64 x i8> %2, <128 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 64, i32 65, i32 66, i32 67, i32 68, i32 69, i32 70, i32 71, i32 72, i32 73, i32 74, i32 75, i32 76, i32 77, i32 78, i32 79, i32 80, i32 81, i32 82, i32 83, i32 84, i32 85, i32 86, i32 87, i32 88, i32 89, i32 90, i32 91, i32 92, i32 93, i32 94, i32 95, i32 96, i32 97, i32 98, i32 99, i32 100, i32 101, i32 102, i32 103, i32 104, i32 105, i32 106, i32 107, i32 108, i32 109, i32 110, i32 111, i32 112, i32 113, i32 114, i32 115, i32 116, i32 117, i32 118, i32 119, i32 120, i32 121, i32 122, i32 123, i32 124, i32 125, i32 126, i32 127>
%5 = shufflevector <64 x i8> %3, <64 x i8> poison, <128 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
%6 = shufflevector <128 x i8> %4, <128 x i8> %5, <192 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 64, i32 65, i32 66, i32 67, i32 68, i32 69, i32 70, i32 71, i32 72, i32 73, i32 74, i32 75, i32 76, i32 77, i32 78, i32 79, i32 80, i32 81, i32 82, i32 83, i32 84, i32 85, i32 86, i32 87, i32 88, i32 89, i32 90, i32 91, i32 92, i32 93, i32 94, i32 95, i32 96, i32 97, i32 98, i32 99, i32 100, i32 101, i32 102, i32 103, i32 104, i32 105, i32 106, i32 107, i32 108, i32 109, i32 110, i32 111, i32 112, i32 113, i32 114, i32 115, i32 116, i32 117, i32 118, i32 119, i32 120, i32 121, i32 122, i32 123, i32 124, i32 125, i32 126, i32 127, i32 128, i32 129, i32 130, i32 131, i32 132, i32 133, i32 134, i32 135, i32 136, i32 137, i32 138, i32 139, i32 140, i32 141, i32 142, i32 143, i32 144, i32 145, i32 146, i32 147, i32 148, i32 149, i32 150, i32 151, i32 152, i32 153, i32 154, i32 155, i32 156, i32 157, i32 158, i32 159, i32 160, i32 161, i32 162, i32 163, i32 164, i32 165, i32 166, i32 167, i32 168, i32 169, i32 170, i32 171, i32 172, i32 173, i32 174, i32 175, i32 176, i32 177, i32 178, i32 179, i32 180, i32 181, i32 182, i32 183, i32 184, i32 185, i32 186, i32 187, i32 188, i32 189, i32 190, i32 191>
%interleaved.vec = shufflevector <192 x i8> %6, <192 x i8> poison, <192 x i32> <i32 0, i32 32, i32 64, i32 96, i32 128, i32 160, i32 1, i32 33, i32 65, i32 97, i32 129, i32 161, i32 2, i32 34, i32 66, i32 98, i32 130, i32 162, i32 3, i32 35, i32 67, i32 99, i32 131, i32 163, i32 4, i32 36, i32 68, i32 100, i32 132, i32 164, i32 5, i32 37, i32 69, i32 101, i32 133, i32 165, i32 6, i32 38, i32 70, i32 102, i32 134, i32 166, i32 7, i32 39, i32 71, i32 103, i32 135, i32 167, i32 8, i32 40, i32 72, i32 104, i32 136, i32 168, i32 9, i32 41, i32 73, i32 105, i32 137, i32 169, i32 10, i32 42, i32 74, i32 106, i32 138, i32 170, i32 11, i32 43, i32 75, i32 107, i32 139, i32 171, i32 12, i32 44, i32 76, i32 108, i32 140, i32 172, i32 13, i32 45, i32 77, i32 109, i32 141, i32 173, i32 14, i32 46, i32 78, i32 110, i32 142, i32 174, i32 15, i32 47, i32 79, i32 111, i32 143, i32 175, i32 16, i32 48, i32 80, i32 112, i32 144, i32 176, i32 17, i32 49, i32 81, i32 113, i32 145, i32 177, i32 18, i32 50, i32 82, i32 114, i32 146, i32 178, i32 19, i32 51, i32 83, i32 115, i32 147, i32 179, i32 20, i32 52, i32 84, i32 116, i32 148, i32 180, i32 21, i32 53, i32 85, i32 117, i32 149, i32 181, i32 22, i32 54, i32 86, i32 118, i32 150, i32 182, i32 23, i32 55, i32 87, i32 119, i32 151, i32 183, i32 24, i32 56, i32 88, i32 120, i32 152, i32 184, i32 25, i32 57, i32 89, i32 121, i32 153, i32 185, i32 26, i32 58, i32 90, i32 122, i32 154, i32 186, i32 27, i32 59, i32 91, i32 123, i32 155, i32 187, i32 28, i32 60, i32 92, i32 124, i32 156, i32 188, i32 29, i32 61, i32 93, i32 125, i32 157, i32 189, i32 30, i32 62, i32 94, i32 126, i32 158, i32 190, i32 31, i32 63, i32 95, i32 127, i32 159, i32 191>
store <192 x i8> %interleaved.vec, ptr %out.vec, align 64
ret void
}
define void @store_i8_stride6_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecptr2, ptr %in.vecptr3, ptr %in.vecptr4, ptr %in.vecptr5, ptr %out.vec) nounwind {
; SSE-LABEL: store_i8_stride6_vf64:
; SSE: # %bb.0:
; SSE-NEXT: subq $184, %rsp
; SSE-NEXT: movdqa (%rdi), %xmm1
; SSE-NEXT: movdqa (%rsi), %xmm5
; SSE-NEXT: movdqa (%rdx), %xmm4
; SSE-NEXT: movdqa (%rcx), %xmm9
; SSE-NEXT: movdqa (%r8), %xmm6
; SSE-NEXT: movdqa (%r9), %xmm2
; SSE-NEXT: movdqa %xmm1, %xmm8
; SSE-NEXT: punpcklbw {{.*#+}} xmm8 = xmm8[0],xmm5[0],xmm8[1],xmm5[1],xmm8[2],xmm5[2],xmm8[3],xmm5[3],xmm8[4],xmm5[4],xmm8[5],xmm5[5],xmm8[6],xmm5[6],xmm8[7],xmm5[7]
; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm8[0,0,1,1]
; SSE-NEXT: movdqa {{.*#+}} xmm14 = [65535,0,65535,65535,0,65535,65535,0]
; SSE-NEXT: pand %xmm14, %xmm3
; SSE-NEXT: movdqa %xmm4, %xmm10
; SSE-NEXT: punpcklbw {{.*#+}} xmm10 = xmm10[0],xmm9[0],xmm10[1],xmm9[1],xmm10[2],xmm9[2],xmm10[3],xmm9[3],xmm10[4],xmm9[4],xmm10[5],xmm9[5],xmm10[6],xmm9[6],xmm10[7],xmm9[7]
; SSE-NEXT: pshuflw {{.*#+}} xmm7 = xmm10[1,0,2,2,4,5,6,7]
; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm7[0,1,0,1]
; SSE-NEXT: movdqa %xmm14, %xmm12
; SSE-NEXT: pandn %xmm7, %xmm12
; SSE-NEXT: por %xmm3, %xmm12
; SSE-NEXT: movdqa {{.*#+}} xmm3 = [65535,65535,0,65535,65535,0,65535,65535]
; SSE-NEXT: pand %xmm3, %xmm12
; SSE-NEXT: movdqa %xmm6, %xmm11
; SSE-NEXT: punpcklbw {{.*#+}} xmm11 = xmm11[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm11[0,0,0,0]
; SSE-NEXT: movdqa %xmm3, %xmm13
; SSE-NEXT: pandn %xmm7, %xmm13
; SSE-NEXT: por %xmm12, %xmm13
; SSE-NEXT: movdqa {{.*#+}} xmm12 = [255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255]
; SSE-NEXT: pand %xmm12, %xmm13
; SSE-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],xmm2[3],xmm0[4],xmm2[4],xmm0[5],xmm2[5],xmm0[6],xmm2[6],xmm0[7],xmm2[7]
; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm0[0,0,0,0]
; SSE-NEXT: pandn %xmm7, %xmm12
; SSE-NEXT: por %xmm13, %xmm12
; SSE-NEXT: movdqa %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT: pshuflw {{.*#+}} xmm7 = xmm10[3,3,3,3,4,5,6,7]
; SSE-NEXT: pshufhw {{.*#+}} xmm7 = xmm7[0,1,2,3,4,4,4,4]
; SSE-NEXT: movdqa %xmm3, %xmm13
; SSE-NEXT: pandn %xmm7, %xmm13
; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm8[1,1,2,2]
; SSE-NEXT: pand %xmm3, %xmm7
; SSE-NEXT: por %xmm7, %xmm13
; SSE-NEXT: movdqa {{.*#+}} xmm12 = [0,65535,65535,0,65535,65535,0,65535]
; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm11[1,1,2,2]
; SSE-NEXT: movdqa %xmm12, %xmm15
; SSE-NEXT: pandn %xmm7, %xmm15
; SSE-NEXT: pand %xmm12, %xmm13
; SSE-NEXT: por %xmm13, %xmm15
; SSE-NEXT: movdqa {{.*#+}} xmm13 = [255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255]
; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm0[1,1,2,2]
; SSE-NEXT: movdqa %xmm13, %xmm12
; SSE-NEXT: pandn %xmm7, %xmm12
; SSE-NEXT: pand %xmm13, %xmm15
; SSE-NEXT: por %xmm15, %xmm12
; SSE-NEXT: movdqa %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm8[3,3,3,3]
; SSE-NEXT: movdqa %xmm3, %xmm8
; SSE-NEXT: pandn %xmm7, %xmm8
; SSE-NEXT: pshufhw {{.*#+}} xmm7 = xmm10[0,1,2,3,5,6,7,7]
; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm7[2,2,2,3]
; SSE-NEXT: pand %xmm3, %xmm7
; SSE-NEXT: por %xmm8, %xmm7
; SSE-NEXT: pshufd {{.*#+}} xmm8 = xmm11[2,2,3,3]
; SSE-NEXT: movdqa %xmm14, %xmm10
; SSE-NEXT: pandn %xmm8, %xmm10
; SSE-NEXT: pand %xmm14, %xmm7
; SSE-NEXT: por %xmm7, %xmm10
; SSE-NEXT: movdqa {{.*#+}} xmm15 = [255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0]
; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm0[2,2,3,3]
; SSE-NEXT: movdqa %xmm15, %xmm0
; SSE-NEXT: pandn %xmm7, %xmm0
; SSE-NEXT: pand %xmm15, %xmm10
; SSE-NEXT: por %xmm10, %xmm0
; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; 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: pshufd {{.*#+}} xmm5 = xmm1[0,0,1,1]
; SSE-NEXT: pand %xmm14, %xmm5
; SSE-NEXT: punpckhbw {{.*#+}} xmm4 = xmm4[8],xmm9[8],xmm4[9],xmm9[9],xmm4[10],xmm9[10],xmm4[11],xmm9[11],xmm4[12],xmm9[12],xmm4[13],xmm9[13],xmm4[14],xmm9[14],xmm4[15],xmm9[15]
; SSE-NEXT: pshuflw {{.*#+}} xmm7 = xmm4[1,0,2,2,4,5,6,7]
; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm7[0,1,0,1]
; SSE-NEXT: movdqa %xmm14, %xmm8
; SSE-NEXT: pandn %xmm7, %xmm8
; SSE-NEXT: por %xmm5, %xmm8
; SSE-NEXT: pand %xmm3, %xmm8
; SSE-NEXT: punpckhbw {{.*#+}} xmm6 = xmm6[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm6[0,0,0,0]
; SSE-NEXT: movdqa %xmm3, %xmm7
; SSE-NEXT: pandn %xmm5, %xmm7
; SSE-NEXT: por %xmm8, %xmm7
; SSE-NEXT: movdqa {{.*#+}} xmm8 = [255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255]
; SSE-NEXT: pand %xmm8, %xmm7
; SSE-NEXT: punpckhbw {{.*#+}} xmm2 = xmm2[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm2[0,0,0,0]
; SSE-NEXT: movdqa %xmm8, %xmm0
; SSE-NEXT: movdqa %xmm8, %xmm12
; SSE-NEXT: pandn %xmm5, %xmm0
; SSE-NEXT: por %xmm7, %xmm0
; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT: pshuflw {{.*#+}} xmm5 = xmm4[3,3,3,3,4,5,6,7]
; SSE-NEXT: pshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,4,4,4,4]
; SSE-NEXT: movdqa %xmm3, %xmm7
; SSE-NEXT: pandn %xmm5, %xmm7
; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm1[1,1,2,2]
; SSE-NEXT: pand %xmm3, %xmm5
; SSE-NEXT: por %xmm5, %xmm7
; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm6[1,1,2,2]
; SSE-NEXT: movdqa {{.*#+}} xmm0 = [0,65535,65535,0,65535,65535,0,65535]
; SSE-NEXT: movdqa %xmm0, %xmm8
; SSE-NEXT: pandn %xmm5, %xmm8
; SSE-NEXT: pand %xmm0, %xmm7
; SSE-NEXT: por %xmm7, %xmm8
; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm2[1,1,2,2]
; SSE-NEXT: movdqa %xmm13, %xmm0
; SSE-NEXT: pandn %xmm5, %xmm0
; SSE-NEXT: pand %xmm13, %xmm8
; SSE-NEXT: por %xmm8, %xmm0
; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[3,3,3,3]
; SSE-NEXT: movdqa %xmm3, %xmm5
; SSE-NEXT: pandn %xmm1, %xmm5
; SSE-NEXT: movdqa 16(%rdx), %xmm11
; SSE-NEXT: pshufhw {{.*#+}} xmm4 = xmm4[0,1,2,3,5,6,7,7]
; SSE-NEXT: pshufd {{.*#+}} xmm4 = xmm4[2,2,2,3]
; SSE-NEXT: pand %xmm3, %xmm4
; SSE-NEXT: por %xmm5, %xmm4
; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm6[2,2,3,3]
; SSE-NEXT: movdqa %xmm14, %xmm6
; SSE-NEXT: pandn %xmm5, %xmm6
; SSE-NEXT: movdqa 16(%rcx), %xmm5
; SSE-NEXT: pand %xmm14, %xmm4
; SSE-NEXT: por %xmm4, %xmm6
; SSE-NEXT: movdqa %xmm15, %xmm0
; SSE-NEXT: pand %xmm15, %xmm6
; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm2[2,2,3,3]
; SSE-NEXT: pandn %xmm2, %xmm0
; SSE-NEXT: por %xmm6, %xmm0
; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT: movdqa %xmm11, %xmm9
; SSE-NEXT: punpcklbw {{.*#+}} xmm9 = xmm9[0],xmm5[0],xmm9[1],xmm5[1],xmm9[2],xmm5[2],xmm9[3],xmm5[3],xmm9[4],xmm5[4],xmm9[5],xmm5[5],xmm9[6],xmm5[6],xmm9[7],xmm5[7]
; SSE-NEXT: pshuflw {{.*#+}} xmm2 = xmm9[1,0,2,2,4,5,6,7]
; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm2[0,1,0,1]
; SSE-NEXT: movdqa %xmm14, %xmm6
; SSE-NEXT: pandn %xmm2, %xmm6
; SSE-NEXT: movdqa 16(%rdi), %xmm4
; SSE-NEXT: movdqa 16(%rsi), %xmm8
; SSE-NEXT: movdqa %xmm4, %xmm13
; SSE-NEXT: punpcklbw {{.*#+}} xmm13 = xmm13[0],xmm8[0],xmm13[1],xmm8[1],xmm13[2],xmm8[2],xmm13[3],xmm8[3],xmm13[4],xmm8[4],xmm13[5],xmm8[5],xmm13[6],xmm8[6],xmm13[7],xmm8[7]
; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm13[0,0,1,1]
; SSE-NEXT: pand %xmm14, %xmm2
; SSE-NEXT: por %xmm2, %xmm6
; SSE-NEXT: movdqa 16(%r8), %xmm2
; SSE-NEXT: movdqa %xmm2, %xmm10
; SSE-NEXT: punpcklbw {{.*#+}} xmm10 = xmm10[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm10[0,0,0,0]
; SSE-NEXT: movdqa %xmm3, %xmm15
; SSE-NEXT: pandn %xmm7, %xmm15
; SSE-NEXT: pand %xmm3, %xmm6
; SSE-NEXT: por %xmm6, %xmm15
; SSE-NEXT: movdqa 16(%r9), %xmm6
; 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: pshufd {{.*#+}} xmm7 = xmm1[0,0,0,0]
; SSE-NEXT: movdqa %xmm12, %xmm0
; SSE-NEXT: pandn %xmm7, %xmm12
; SSE-NEXT: pand %xmm0, %xmm15
; SSE-NEXT: por %xmm15, %xmm12
; SSE-NEXT: movdqa %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT: pshuflw {{.*#+}} xmm7 = xmm9[3,3,3,3,4,5,6,7]
; SSE-NEXT: pshufhw {{.*#+}} xmm7 = xmm7[0,1,2,3,4,4,4,4]
; SSE-NEXT: movdqa %xmm3, %xmm15
; SSE-NEXT: pandn %xmm7, %xmm15
; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm13[1,1,2,2]
; SSE-NEXT: pand %xmm3, %xmm7
; SSE-NEXT: por %xmm7, %xmm15
; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm10[1,1,2,2]
; SSE-NEXT: movdqa {{.*#+}} xmm0 = [0,65535,65535,0,65535,65535,0,65535]
; SSE-NEXT: movdqa %xmm0, %xmm12
; SSE-NEXT: pandn %xmm7, %xmm12
; SSE-NEXT: pand %xmm0, %xmm15
; SSE-NEXT: por %xmm15, %xmm12
; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm1[1,1,2,2]
; SSE-NEXT: movdqa {{.*#+}} xmm0 = [255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255]
; SSE-NEXT: movdqa %xmm0, %xmm15
; SSE-NEXT: pandn %xmm7, %xmm15
; SSE-NEXT: pand %xmm0, %xmm12
; SSE-NEXT: por %xmm12, %xmm15
; SSE-NEXT: movdqa %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm13[3,3,3,3]
; SSE-NEXT: movdqa %xmm3, %xmm12
; SSE-NEXT: pandn %xmm7, %xmm12
; SSE-NEXT: pshufhw {{.*#+}} xmm7 = xmm9[0,1,2,3,5,6,7,7]
; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm7[2,2,2,3]
; SSE-NEXT: pand %xmm3, %xmm7
; SSE-NEXT: por %xmm12, %xmm7
; SSE-NEXT: pshufd {{.*#+}} xmm9 = xmm10[2,2,3,3]
; SSE-NEXT: movdqa %xmm14, %xmm10
; SSE-NEXT: pandn %xmm9, %xmm10
; SSE-NEXT: pand %xmm14, %xmm7
; SSE-NEXT: por %xmm7, %xmm10
; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm1[2,2,3,3]
; SSE-NEXT: movdqa {{.*#+}} xmm1 = [255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0]
; SSE-NEXT: movdqa %xmm1, %xmm0
; SSE-NEXT: pandn %xmm7, %xmm0
; SSE-NEXT: pand %xmm1, %xmm10
; SSE-NEXT: movdqa %xmm1, %xmm9
; SSE-NEXT: por %xmm10, %xmm0
; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT: punpckhbw {{.*#+}} xmm4 = xmm4[8],xmm8[8],xmm4[9],xmm8[9],xmm4[10],xmm8[10],xmm4[11],xmm8[11],xmm4[12],xmm8[12],xmm4[13],xmm8[13],xmm4[14],xmm8[14],xmm4[15],xmm8[15]
; SSE-NEXT: punpckhbw {{.*#+}} xmm11 = xmm11[8],xmm5[8],xmm11[9],xmm5[9],xmm11[10],xmm5[10],xmm11[11],xmm5[11],xmm11[12],xmm5[12],xmm11[13],xmm5[13],xmm11[14],xmm5[14],xmm11[15],xmm5[15]
; SSE-NEXT: pshuflw {{.*#+}} xmm5 = xmm11[1,0,2,2,4,5,6,7]
; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm5[0,1,0,1]
; SSE-NEXT: movdqa %xmm14, %xmm7
; SSE-NEXT: pandn %xmm5, %xmm7
; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm4[0,0,1,1]
; SSE-NEXT: pand %xmm14, %xmm5
; SSE-NEXT: por %xmm5, %xmm7
; SSE-NEXT: punpckhbw {{.*#+}} xmm2 = xmm2[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm2[0,0,0,0]
; SSE-NEXT: movdqa %xmm3, %xmm8
; SSE-NEXT: pandn %xmm5, %xmm8
; SSE-NEXT: pand %xmm3, %xmm7
; SSE-NEXT: por %xmm7, %xmm8
; SSE-NEXT: punpckhbw {{.*#+}} xmm5 = xmm5[8],xmm6[8],xmm5[9],xmm6[9],xmm5[10],xmm6[10],xmm5[11],xmm6[11],xmm5[12],xmm6[12],xmm5[13],xmm6[13],xmm5[14],xmm6[14],xmm5[15],xmm6[15]
; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm5[0,0,0,0]
; SSE-NEXT: movdqa {{.*#+}} xmm0 = [255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255]
; SSE-NEXT: movdqa %xmm0, %xmm7
; SSE-NEXT: pandn %xmm6, %xmm7
; SSE-NEXT: pand %xmm0, %xmm8
; SSE-NEXT: movdqa %xmm0, %xmm15
; SSE-NEXT: por %xmm8, %xmm7
; SSE-NEXT: movdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT: pshuflw {{.*#+}} xmm6 = xmm11[3,3,3,3,4,5,6,7]
; SSE-NEXT: pshufhw {{.*#+}} xmm6 = xmm6[0,1,2,3,4,4,4,4]
; SSE-NEXT: movdqa %xmm3, %xmm7
; SSE-NEXT: pandn %xmm6, %xmm7
; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm4[1,1,2,2]
; SSE-NEXT: pand %xmm3, %xmm6
; SSE-NEXT: por %xmm6, %xmm7
; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm2[1,1,2,2]
; SSE-NEXT: movdqa {{.*#+}} xmm0 = [0,65535,65535,0,65535,65535,0,65535]
; SSE-NEXT: movdqa %xmm0, %xmm8
; SSE-NEXT: pandn %xmm6, %xmm8
; SSE-NEXT: pand %xmm0, %xmm7
; SSE-NEXT: por %xmm7, %xmm8
; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm5[1,1,2,2]
; SSE-NEXT: movdqa {{.*#+}} xmm0 = [255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255]
; SSE-NEXT: movdqa %xmm0, %xmm7
; SSE-NEXT: pandn %xmm6, %xmm7
; SSE-NEXT: pand %xmm0, %xmm8
; SSE-NEXT: por %xmm8, %xmm7
; SSE-NEXT: movdqa %xmm7, (%rsp) # 16-byte Spill
; SSE-NEXT: pshufd {{.*#+}} xmm4 = xmm4[3,3,3,3]
; SSE-NEXT: movdqa %xmm3, %xmm6
; SSE-NEXT: pandn %xmm4, %xmm6
; SSE-NEXT: pshufhw {{.*#+}} xmm1 = xmm11[0,1,2,3,5,6,7,7]
; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[2,2,2,3]
; SSE-NEXT: pand %xmm3, %xmm1
; SSE-NEXT: por %xmm6, %xmm1
; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm2[2,2,3,3]
; SSE-NEXT: movdqa %xmm14, %xmm4
; SSE-NEXT: pandn %xmm2, %xmm4
; SSE-NEXT: pand %xmm14, %xmm1
; SSE-NEXT: por %xmm1, %xmm4
; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm5[2,2,3,3]
; SSE-NEXT: movdqa %xmm9, %xmm0
; SSE-NEXT: pandn %xmm1, %xmm0
; SSE-NEXT: pand %xmm9, %xmm4
; SSE-NEXT: por %xmm4, %xmm0
; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT: movdqa 32(%rdx), %xmm1
; SSE-NEXT: movdqa 32(%rcx), %xmm5
; SSE-NEXT: movdqa %xmm1, %xmm9
; SSE-NEXT: punpcklbw {{.*#+}} xmm9 = xmm9[0],xmm5[0],xmm9[1],xmm5[1],xmm9[2],xmm5[2],xmm9[3],xmm5[3],xmm9[4],xmm5[4],xmm9[5],xmm5[5],xmm9[6],xmm5[6],xmm9[7],xmm5[7]
; SSE-NEXT: pshuflw {{.*#+}} xmm2 = xmm9[1,0,2,2,4,5,6,7]
; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm2[0,1,0,1]
; SSE-NEXT: movdqa %xmm14, %xmm6
; SSE-NEXT: pandn %xmm2, %xmm6
; SSE-NEXT: movdqa 32(%rdi), %xmm4
; SSE-NEXT: movdqa 32(%rsi), %xmm8
; SSE-NEXT: movdqa %xmm4, %xmm13
; SSE-NEXT: punpcklbw {{.*#+}} xmm13 = xmm13[0],xmm8[0],xmm13[1],xmm8[1],xmm13[2],xmm8[2],xmm13[3],xmm8[3],xmm13[4],xmm8[4],xmm13[5],xmm8[5],xmm13[6],xmm8[6],xmm13[7],xmm8[7]
; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm13[0,0,1,1]
; SSE-NEXT: pand %xmm14, %xmm2
; SSE-NEXT: por %xmm2, %xmm6
; SSE-NEXT: movdqa 32(%r8), %xmm2
; SSE-NEXT: movdqa %xmm2, %xmm10
; SSE-NEXT: punpcklbw {{.*#+}} xmm10 = xmm10[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm10[0,0,0,0]
; SSE-NEXT: movdqa %xmm3, %xmm12
; SSE-NEXT: pandn %xmm7, %xmm12
; SSE-NEXT: pand %xmm3, %xmm6
; SSE-NEXT: por %xmm6, %xmm12
; SSE-NEXT: movdqa 32(%r9), %xmm6
; 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: pshufd {{.*#+}} xmm7 = xmm0[0,0,0,0]
; SSE-NEXT: movdqa %xmm15, %xmm11
; SSE-NEXT: pandn %xmm7, %xmm15
; SSE-NEXT: pand %xmm11, %xmm12
; SSE-NEXT: por %xmm12, %xmm15
; SSE-NEXT: movdqa %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT: pshuflw {{.*#+}} xmm7 = xmm9[3,3,3,3,4,5,6,7]
; SSE-NEXT: pshufhw {{.*#+}} xmm7 = xmm7[0,1,2,3,4,4,4,4]
; SSE-NEXT: movdqa %xmm3, %xmm12
; SSE-NEXT: pandn %xmm7, %xmm12
; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm13[1,1,2,2]
; SSE-NEXT: pand %xmm3, %xmm7
; SSE-NEXT: por %xmm7, %xmm12
; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm10[1,1,2,2]
; SSE-NEXT: movdqa {{.*#+}} xmm11 = [0,65535,65535,0,65535,65535,0,65535]
; SSE-NEXT: movdqa %xmm11, %xmm15
; SSE-NEXT: pandn %xmm7, %xmm15
; SSE-NEXT: pand %xmm11, %xmm12
; SSE-NEXT: por %xmm12, %xmm15
; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm0[1,1,2,2]
; SSE-NEXT: movdqa {{.*#+}} xmm11 = [255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255]
; SSE-NEXT: movdqa %xmm11, %xmm12
; SSE-NEXT: pandn %xmm7, %xmm12
; SSE-NEXT: pand %xmm11, %xmm15
; SSE-NEXT: por %xmm15, %xmm12
; SSE-NEXT: movdqa %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm13[3,3,3,3]
; SSE-NEXT: movdqa %xmm3, %xmm12
; SSE-NEXT: pandn %xmm7, %xmm12
; SSE-NEXT: pshufhw {{.*#+}} xmm7 = xmm9[0,1,2,3,5,6,7,7]
; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm7[2,2,2,3]
; SSE-NEXT: pand %xmm3, %xmm7
; SSE-NEXT: por %xmm12, %xmm7
; SSE-NEXT: pshufd {{.*#+}} xmm9 = xmm10[2,2,3,3]
; SSE-NEXT: movdqa %xmm14, %xmm10
; SSE-NEXT: pandn %xmm9, %xmm10
; SSE-NEXT: pand %xmm14, %xmm7
; SSE-NEXT: por %xmm7, %xmm10
; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm0[2,2,3,3]
; SSE-NEXT: movdqa {{.*#+}} xmm11 = [255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0]
; SSE-NEXT: movdqa %xmm11, %xmm0
; SSE-NEXT: pandn %xmm7, %xmm0
; SSE-NEXT: pand %xmm11, %xmm10
; SSE-NEXT: por %xmm10, %xmm0
; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT: punpckhbw {{.*#+}} xmm4 = xmm4[8],xmm8[8],xmm4[9],xmm8[9],xmm4[10],xmm8[10],xmm4[11],xmm8[11],xmm4[12],xmm8[12],xmm4[13],xmm8[13],xmm4[14],xmm8[14],xmm4[15],xmm8[15]
; 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: pshuflw {{.*#+}} xmm5 = xmm1[1,0,2,2,4,5,6,7]
; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm5[0,1,0,1]
; SSE-NEXT: movdqa %xmm14, %xmm7
; SSE-NEXT: pandn %xmm5, %xmm7
; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm4[0,0,1,1]
; SSE-NEXT: pand %xmm14, %xmm5
; SSE-NEXT: por %xmm5, %xmm7
; SSE-NEXT: punpckhbw {{.*#+}} xmm2 = xmm2[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm2[0,0,0,0]
; SSE-NEXT: movdqa %xmm3, %xmm8
; SSE-NEXT: pandn %xmm5, %xmm8
; SSE-NEXT: pand %xmm3, %xmm7
; SSE-NEXT: por %xmm7, %xmm8
; SSE-NEXT: punpckhbw {{.*#+}} xmm5 = xmm5[8],xmm6[8],xmm5[9],xmm6[9],xmm5[10],xmm6[10],xmm5[11],xmm6[11],xmm5[12],xmm6[12],xmm5[13],xmm6[13],xmm5[14],xmm6[14],xmm5[15],xmm6[15]
; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm5[0,0,0,0]
; SSE-NEXT: movdqa {{.*#+}} xmm0 = [255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255]
; SSE-NEXT: movdqa %xmm0, %xmm7
; SSE-NEXT: pandn %xmm6, %xmm7
; SSE-NEXT: pand %xmm0, %xmm8
; SSE-NEXT: movdqa %xmm0, %xmm15
; SSE-NEXT: por %xmm8, %xmm7
; SSE-NEXT: movdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT: pshuflw {{.*#+}} xmm6 = xmm1[3,3,3,3,4,5,6,7]
; SSE-NEXT: pshufhw {{.*#+}} xmm6 = xmm6[0,1,2,3,4,4,4,4]
; SSE-NEXT: movdqa %xmm3, %xmm7
; SSE-NEXT: pandn %xmm6, %xmm7
; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm4[1,1,2,2]
; SSE-NEXT: pand %xmm3, %xmm6
; SSE-NEXT: por %xmm6, %xmm7
; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm2[1,1,2,2]
; SSE-NEXT: movdqa {{.*#+}} xmm0 = [0,65535,65535,0,65535,65535,0,65535]
; SSE-NEXT: movdqa %xmm0, %xmm8
; SSE-NEXT: pandn %xmm6, %xmm8
; SSE-NEXT: pand %xmm0, %xmm7
; SSE-NEXT: por %xmm7, %xmm8
; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm5[1,1,2,2]
; SSE-NEXT: movdqa {{.*#+}} xmm0 = [255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255]
; SSE-NEXT: movdqa %xmm0, %xmm7
; SSE-NEXT: pandn %xmm6, %xmm7
; SSE-NEXT: pand %xmm0, %xmm8
; SSE-NEXT: por %xmm8, %xmm7
; SSE-NEXT: movdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT: pshufd {{.*#+}} xmm4 = xmm4[3,3,3,3]
; SSE-NEXT: movdqa %xmm3, %xmm6
; SSE-NEXT: pandn %xmm4, %xmm6
; SSE-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,5,6,7,7]
; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[2,2,2,3]
; SSE-NEXT: pand %xmm3, %xmm1
; SSE-NEXT: por %xmm6, %xmm1
; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm2[2,2,3,3]
; SSE-NEXT: movdqa %xmm14, %xmm4
; SSE-NEXT: pandn %xmm2, %xmm4
; SSE-NEXT: pand %xmm14, %xmm1
; SSE-NEXT: por %xmm1, %xmm4
; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm5[2,2,3,3]
; SSE-NEXT: movdqa %xmm11, %xmm0
; SSE-NEXT: pandn %xmm1, %xmm11
; SSE-NEXT: pand %xmm0, %xmm4
; SSE-NEXT: por %xmm4, %xmm11
; SSE-NEXT: movdqa 48(%rdx), %xmm10
; SSE-NEXT: movdqa 48(%rcx), %xmm0
; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT: movdqa %xmm10, %xmm8
; SSE-NEXT: punpcklbw {{.*#+}} xmm8 = xmm8[0],xmm0[0],xmm8[1],xmm0[1],xmm8[2],xmm0[2],xmm8[3],xmm0[3],xmm8[4],xmm0[4],xmm8[5],xmm0[5],xmm8[6],xmm0[6],xmm8[7],xmm0[7]
; SSE-NEXT: pshuflw {{.*#+}} xmm1 = xmm8[1,0,2,2,4,5,6,7]
; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,1,0,1]
; SSE-NEXT: movdqa %xmm14, %xmm12
; SSE-NEXT: pandn %xmm1, %xmm12
; SSE-NEXT: movdqa 48(%rdi), %xmm6
; SSE-NEXT: movdqa 48(%rsi), %xmm9
; SSE-NEXT: movdqa %xmm6, %xmm4
; SSE-NEXT: punpcklbw {{.*#+}} xmm4 = xmm4[0],xmm9[0],xmm4[1],xmm9[1],xmm4[2],xmm9[2],xmm4[3],xmm9[3],xmm4[4],xmm9[4],xmm4[5],xmm9[5],xmm4[6],xmm9[6],xmm4[7],xmm9[7]
; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm4[0,0,1,1]
; SSE-NEXT: pand %xmm14, %xmm1
; SSE-NEXT: por %xmm1, %xmm12
; SSE-NEXT: movdqa 48(%r8), %xmm7
; SSE-NEXT: movdqa %xmm7, %xmm2
; SSE-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm2[0,0,0,0]
; SSE-NEXT: movdqa %xmm3, %xmm13
; SSE-NEXT: pandn %xmm1, %xmm13
; SSE-NEXT: pand %xmm3, %xmm12
; SSE-NEXT: por %xmm12, %xmm13
; SSE-NEXT: movdqa 48(%r9), %xmm0
; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; SSE-NEXT: punpcklbw {{.*#+}} xmm5 = xmm5[0],xmm0[0],xmm5[1],xmm0[1],xmm5[2],xmm0[2],xmm5[3],xmm0[3],xmm5[4],xmm0[4],xmm5[5],xmm0[5],xmm5[6],xmm0[6],xmm5[7],xmm0[7]
; SSE-NEXT: pshufd {{.*#+}} xmm12 = xmm5[0,0,0,0]
; SSE-NEXT: movdqa %xmm15, %xmm0
; SSE-NEXT: pandn %xmm12, %xmm15
; SSE-NEXT: pand %xmm0, %xmm13
; SSE-NEXT: por %xmm13, %xmm15
; SSE-NEXT: pshuflw {{.*#+}} xmm12 = xmm8[3,3,3,3,4,5,6,7]
; SSE-NEXT: pshufhw {{.*#+}} xmm12 = xmm12[0,1,2,3,4,4,4,4]
; SSE-NEXT: movdqa %xmm3, %xmm13
; SSE-NEXT: pandn %xmm12, %xmm13
; SSE-NEXT: pshufd {{.*#+}} xmm12 = xmm4[1,1,2,2]
; SSE-NEXT: pand %xmm3, %xmm12
; SSE-NEXT: por %xmm12, %xmm13
; SSE-NEXT: pshufd {{.*#+}} xmm12 = xmm2[1,1,2,2]
; SSE-NEXT: movdqa {{.*#+}} xmm1 = [0,65535,65535,0,65535,65535,0,65535]
; SSE-NEXT: movdqa %xmm1, %xmm0
; SSE-NEXT: pandn %xmm12, %xmm0
; SSE-NEXT: pand %xmm1, %xmm13
; SSE-NEXT: por %xmm13, %xmm0
; SSE-NEXT: pshufd {{.*#+}} xmm12 = xmm5[1,1,2,2]
; SSE-NEXT: movdqa {{.*#+}} xmm1 = [255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255]
; SSE-NEXT: movdqa %xmm1, %xmm13
; SSE-NEXT: pandn %xmm12, %xmm13
; SSE-NEXT: pand %xmm1, %xmm0
; SSE-NEXT: movdqa %xmm1, %xmm12
; SSE-NEXT: por %xmm0, %xmm13
; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm4[3,3,3,3]
; SSE-NEXT: movdqa %xmm3, %xmm4
; SSE-NEXT: pandn %xmm0, %xmm4
; SSE-NEXT: pshufhw {{.*#+}} xmm0 = xmm8[0,1,2,3,5,6,7,7]
; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[2,2,2,3]
; SSE-NEXT: pand %xmm3, %xmm0
; SSE-NEXT: por %xmm4, %xmm0
; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm2[2,2,3,3]
; SSE-NEXT: movdqa %xmm14, %xmm4
; SSE-NEXT: pandn %xmm2, %xmm4
; SSE-NEXT: pand %xmm14, %xmm0
; SSE-NEXT: por %xmm0, %xmm4
; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm5[2,2,3,3]
; SSE-NEXT: movdqa {{.*#+}} xmm5 = [255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0]
; SSE-NEXT: movdqa %xmm5, %xmm8
; SSE-NEXT: pandn %xmm0, %xmm8
; SSE-NEXT: pand %xmm5, %xmm4
; SSE-NEXT: por %xmm4, %xmm8
; SSE-NEXT: punpckhbw {{.*#+}} xmm6 = xmm6[8],xmm9[8],xmm6[9],xmm9[9],xmm6[10],xmm9[10],xmm6[11],xmm9[11],xmm6[12],xmm9[12],xmm6[13],xmm9[13],xmm6[14],xmm9[14],xmm6[15],xmm9[15]
; SSE-NEXT: punpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Folded Reload
; SSE-NEXT: # xmm10 = xmm10[8],mem[8],xmm10[9],mem[9],xmm10[10],mem[10],xmm10[11],mem[11],xmm10[12],mem[12],xmm10[13],mem[13],xmm10[14],mem[14],xmm10[15],mem[15]
; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm10[1,0,2,2,4,5,6,7]
; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,1,0,1]
; SSE-NEXT: movdqa %xmm14, %xmm1
; SSE-NEXT: pandn %xmm0, %xmm1
; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm6[0,0,1,1]
; SSE-NEXT: pand %xmm14, %xmm0
; SSE-NEXT: por %xmm0, %xmm1
; SSE-NEXT: punpckhbw {{.*#+}} xmm7 = xmm7[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm7[0,0,0,0]
; SSE-NEXT: movdqa %xmm3, %xmm2
; SSE-NEXT: pandn %xmm0, %xmm2
; SSE-NEXT: pand %xmm3, %xmm1
; SSE-NEXT: por %xmm1, %xmm2
; SSE-NEXT: movdqa {{.*#+}} xmm4 = [255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255]
; SSE-NEXT: pand %xmm4, %xmm2
; SSE-NEXT: punpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
; SSE-NEXT: # xmm1 = xmm1[8],mem[8],xmm1[9],mem[9],xmm1[10],mem[10],xmm1[11],mem[11],xmm1[12],mem[12],xmm1[13],mem[13],xmm1[14],mem[14],xmm1[15],mem[15]
; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm1[0,0,0,0]
; SSE-NEXT: pandn %xmm0, %xmm4
; SSE-NEXT: por %xmm2, %xmm4
; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm10[3,3,3,3,4,5,6,7]
; SSE-NEXT: pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,4,4,4]
; SSE-NEXT: movdqa %xmm3, %xmm2
; SSE-NEXT: pandn %xmm0, %xmm2
; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm6[1,1,2,2]
; SSE-NEXT: pand %xmm3, %xmm0
; SSE-NEXT: por %xmm0, %xmm2
; SSE-NEXT: movdqa {{.*#+}} xmm9 = [0,65535,65535,0,65535,65535,0,65535]
; SSE-NEXT: pand %xmm9, %xmm2
; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm7[1,1,2,2]
; SSE-NEXT: pandn %xmm0, %xmm9
; SSE-NEXT: por %xmm2, %xmm9
; SSE-NEXT: movdqa %xmm12, %xmm2
; SSE-NEXT: pand %xmm12, %xmm9
; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,2,2]
; SSE-NEXT: pandn %xmm0, %xmm2
; SSE-NEXT: por %xmm9, %xmm2
; SSE-NEXT: movdqa %xmm2, %xmm9
; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm6[3,3,3,3]
; SSE-NEXT: pshufhw {{.*#+}} xmm2 = xmm10[0,1,2,3,5,6,7,7]
; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm2[2,2,2,3]
; SSE-NEXT: pand %xmm3, %xmm2
; SSE-NEXT: pandn %xmm0, %xmm3
; SSE-NEXT: por %xmm2, %xmm3
; SSE-NEXT: pand %xmm14, %xmm3
; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm7[2,2,3,3]
; SSE-NEXT: pandn %xmm0, %xmm14
; SSE-NEXT: por %xmm3, %xmm14
; SSE-NEXT: movdqa %xmm5, %xmm2
; SSE-NEXT: pand %xmm5, %xmm14
; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,2,3,3]
; SSE-NEXT: pandn %xmm0, %xmm2
; SSE-NEXT: por %xmm14, %xmm2
; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax
; SSE-NEXT: movdqa %xmm2, 368(%rax)
; SSE-NEXT: movdqa %xmm9, 352(%rax)
; SSE-NEXT: movdqa %xmm4, 336(%rax)
; SSE-NEXT: movdqa %xmm8, 320(%rax)
; SSE-NEXT: movdqa %xmm13, 304(%rax)
; SSE-NEXT: movdqa %xmm15, 288(%rax)
; SSE-NEXT: movdqa %xmm11, 272(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT: movaps %xmm0, 256(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT: movaps %xmm0, 240(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT: movaps %xmm0, 224(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT: movaps %xmm0, 208(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT: movaps %xmm0, 192(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT: movaps %xmm0, 176(%rax)
; SSE-NEXT: movaps (%rsp), %xmm0 # 16-byte Reload
; SSE-NEXT: movaps %xmm0, 160(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT: movaps %xmm0, 144(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT: movaps %xmm0, 128(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT: movaps %xmm0, 112(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT: movaps %xmm0, 96(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT: movaps %xmm0, 80(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT: movaps %xmm0, 64(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT: movaps %xmm0, 48(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT: movaps %xmm0, 32(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT: movaps %xmm0, 16(%rax)
; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; SSE-NEXT: movaps %xmm0, (%rax)
; SSE-NEXT: addq $184, %rsp
; SSE-NEXT: retq
;
; AVX1-ONLY-LABEL: store_i8_stride6_vf64:
; AVX1-ONLY: # %bb.0:
; AVX1-ONLY-NEXT: subq $200, %rsp
; AVX1-ONLY-NEXT: vmovdqa 48(%rsi), %xmm0
; AVX1-ONLY-NEXT: vmovdqa 48(%rdi), %xmm1
; AVX1-ONLY-NEXT: vpunpcklbw {{.*#+}} xmm6 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3],xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm2 = xmm6[3,3,3,3]
; AVX1-ONLY-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm1[8],xmm0[8],xmm1[9],xmm0[9],xmm1[10],xmm0[10],xmm1[11],xmm0[11],xmm1[12],xmm0[12],xmm1[13],xmm0[13],xmm1[14],xmm0[14],xmm1[15],xmm0[15]
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[0,0,1,1]
; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm1, %ymm2, %ymm1
; AVX1-ONLY-NEXT: vmovaps {{.*#+}} ymm13 = [65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535]
; AVX1-ONLY-NEXT: vandnps %ymm1, %ymm13, %ymm2
; AVX1-ONLY-NEXT: vmovdqa 48(%rcx), %xmm1
; AVX1-ONLY-NEXT: vmovdqa 48(%rdx), %xmm3
; AVX1-ONLY-NEXT: vpunpcklbw {{.*#+}} xmm9 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3],xmm3[4],xmm1[4],xmm3[5],xmm1[5],xmm3[6],xmm1[6],xmm3[7],xmm1[7]
; AVX1-ONLY-NEXT: vpshufhw {{.*#+}} xmm4 = xmm9[0,1,2,3,5,6,7,7]
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm4 = xmm4[2,2,2,3]
; AVX1-ONLY-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm3[8],xmm1[8],xmm3[9],xmm1[9],xmm3[10],xmm1[10],xmm3[11],xmm1[11],xmm3[12],xmm1[12],xmm3[13],xmm1[13],xmm3[14],xmm1[14],xmm3[15],xmm1[15]
; AVX1-ONLY-NEXT: vpshuflw {{.*#+}} xmm3 = xmm1[1,0,2,2,4,5,6,7]
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm3 = xmm3[0,1,0,1]
; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm3, %ymm4, %ymm3
; AVX1-ONLY-NEXT: vandps %ymm3, %ymm13, %ymm3
; AVX1-ONLY-NEXT: vorps %ymm2, %ymm3, %ymm3
; AVX1-ONLY-NEXT: vextractf128 $1, %ymm3, %xmm4
; AVX1-ONLY-NEXT: vmovdqa 48(%r8), %xmm2
; AVX1-ONLY-NEXT: vpshufb {{.*#+}} xmm5 = zero,zero,zero,zero,xmm2[8,u],zero,zero,zero,zero,xmm2[9,u],zero,zero,zero,zero
; AVX1-ONLY-NEXT: vpblendw {{.*#+}} xmm4 = xmm4[0,1],xmm5[2],xmm4[3,4],xmm5[5],xmm4[6,7]
; AVX1-ONLY-NEXT: vpshufb {{.*#+}} xmm4 = xmm4[0,1,2,3,4],zero,xmm4[6,7,8,9,10],zero,xmm4[12,13,14,15]
; AVX1-ONLY-NEXT: vmovdqa 48(%r9), %xmm5
; AVX1-ONLY-NEXT: vpshufb {{.*#+}} xmm7 = zero,zero,zero,zero,zero,xmm5[8],zero,zero,zero,zero,zero,xmm5[9],zero,zero,zero,zero
; AVX1-ONLY-NEXT: vpor %xmm7, %xmm4, %xmm4
; AVX1-ONLY-NEXT: vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vpshufb {{.*#+}} xmm4 = zero,zero,xmm2[5,u],zero,zero,zero,zero,xmm2[6,u],zero,zero,zero,zero,xmm2[7,u]
; AVX1-ONLY-NEXT: vpblendw {{.*#+}} xmm3 = xmm3[0],xmm4[1],xmm3[2,3],xmm4[4],xmm3[5,6],xmm4[7]
; AVX1-ONLY-NEXT: vmovdqa {{.*#+}} xmm15 = [0,1,2,128,4,5,6,7,8,128,10,11,12,13,14,128]
; AVX1-ONLY-NEXT: vpshufb %xmm15, %xmm3, %xmm3
; AVX1-ONLY-NEXT: vpshufb {{.*#+}} xmm4 = zero,zero,zero,xmm5[5],zero,zero,zero,zero,zero,xmm5[6],zero,zero,zero,zero,zero,xmm5[7]
; AVX1-ONLY-NEXT: vpor %xmm4, %xmm3, %xmm3
; AVX1-ONLY-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm3 = xmm0[1,1,2,2]
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[3,3,3,3]
; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm0, %ymm3, %ymm0
; AVX1-ONLY-NEXT: vandps %ymm0, %ymm13, %ymm0
; AVX1-ONLY-NEXT: vpshuflw {{.*#+}} xmm3 = xmm1[3,3,3,3,4,5,6,7]
; AVX1-ONLY-NEXT: vpshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,4,4,4]
; AVX1-ONLY-NEXT: vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,5,6,7,7]
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[2,2,2,3]
; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm1, %ymm3, %ymm1
; AVX1-ONLY-NEXT: vandnps %ymm1, %ymm13, %ymm1
; AVX1-ONLY-NEXT: vorps %ymm1, %ymm0, %ymm0
; AVX1-ONLY-NEXT: vextractf128 $1, %ymm0, %xmm1
; AVX1-ONLY-NEXT: vpshufb {{.*#+}} xmm3 = zero,zero,xmm2[13,u],zero,zero,zero,zero,xmm2[14,u],zero,zero,zero,zero,xmm2[15,u]
; AVX1-ONLY-NEXT: vpblendw {{.*#+}} xmm1 = xmm1[0],xmm3[1],xmm1[2,3],xmm3[4],xmm1[5,6],xmm3[7]
; AVX1-ONLY-NEXT: vpshufb %xmm15, %xmm1, %xmm1
; AVX1-ONLY-NEXT: vpshufb {{.*#+}} xmm3 = zero,zero,zero,xmm5[13],zero,zero,zero,zero,zero,xmm5[14],zero,zero,zero,zero,zero,xmm5[15]
; AVX1-ONLY-NEXT: vpor %xmm3, %xmm1, %xmm1
; AVX1-ONLY-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vpshufb {{.*#+}} xmm1 = xmm2[10,u],zero,zero,zero,zero,xmm2[11,u],zero,zero,zero,zero,xmm2[12,u],zero,zero
; AVX1-ONLY-NEXT: vpblendw {{.*#+}} xmm0 = xmm1[0],xmm0[1,2],xmm1[3],xmm0[4,5],xmm1[6],xmm0[7]
; AVX1-ONLY-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0],zero,xmm0[2,3,4,5,6],zero,xmm0[8,9,10,11,12],zero,xmm0[14,15]
; AVX1-ONLY-NEXT: vmovdqa {{.*#+}} xmm12 = [128,10,128,128,128,128,128,11,128,128,128,128,128,12,128,128]
; AVX1-ONLY-NEXT: vpshufb %xmm12, %xmm5, %xmm1
; AVX1-ONLY-NEXT: vpor %xmm1, %xmm0, %xmm0
; AVX1-ONLY-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovdqa 32(%rsi), %xmm10
; AVX1-ONLY-NEXT: vmovdqa 32(%rdi), %xmm14
; AVX1-ONLY-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm14[8],xmm10[8],xmm14[9],xmm10[9],xmm14[10],xmm10[10],xmm14[11],xmm10[11],xmm14[12],xmm10[12],xmm14[13],xmm10[13],xmm14[14],xmm10[14],xmm14[15],xmm10[15]
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm0 = xmm1[1,1,2,2]
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm3 = xmm1[3,3,3,3]
; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm0
; AVX1-ONLY-NEXT: vmovdqa 32(%rcx), %xmm4
; AVX1-ONLY-NEXT: vmovdqa 32(%rdx), %xmm7
; AVX1-ONLY-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm7[8],xmm4[8],xmm7[9],xmm4[9],xmm7[10],xmm4[10],xmm7[11],xmm4[11],xmm7[12],xmm4[12],xmm7[13],xmm4[13],xmm7[14],xmm4[14],xmm7[15],xmm4[15]
; AVX1-ONLY-NEXT: vpshuflw {{.*#+}} xmm8 = xmm3[3,3,3,3,4,5,6,7]
; AVX1-ONLY-NEXT: vpshufhw {{.*#+}} xmm8 = xmm8[0,1,2,3,4,4,4,4]
; AVX1-ONLY-NEXT: vpshufhw {{.*#+}} xmm11 = xmm3[0,1,2,3,5,6,7,7]
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm11 = xmm11[2,2,2,3]
; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm11, %ymm8, %ymm8
; AVX1-ONLY-NEXT: vandps %ymm0, %ymm13, %ymm0
; AVX1-ONLY-NEXT: vandnps %ymm8, %ymm13, %ymm8
; AVX1-ONLY-NEXT: vorps %ymm0, %ymm8, %ymm8
; AVX1-ONLY-NEXT: vmovdqa 32(%r8), %xmm11
; AVX1-ONLY-NEXT: vextractf128 $1, %ymm8, %xmm0
; AVX1-ONLY-NEXT: vpshufb {{.*#+}} xmm13 = zero,zero,xmm11[13,u],zero,zero,zero,zero,xmm11[14,u],zero,zero,zero,zero,xmm11[15,u]
; AVX1-ONLY-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0],xmm13[1],xmm0[2,3],xmm13[4],xmm0[5,6],xmm13[7]
; AVX1-ONLY-NEXT: vpshufb %xmm15, %xmm0, %xmm13
; AVX1-ONLY-NEXT: vmovdqa 32(%r9), %xmm0
; AVX1-ONLY-NEXT: vpshufb {{.*#+}} xmm15 = zero,zero,zero,xmm0[13],zero,zero,zero,zero,zero,xmm0[14],zero,zero,zero,zero,zero,xmm0[15]
; AVX1-ONLY-NEXT: vpor %xmm15, %xmm13, %xmm13
; AVX1-ONLY-NEXT: vmovdqa %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vpshufb {{.*#+}} xmm13 = xmm11[10,u],zero,zero,zero,zero,xmm11[11,u],zero,zero,zero,zero,xmm11[12,u],zero,zero
; AVX1-ONLY-NEXT: vpblendw {{.*#+}} xmm8 = xmm13[0],xmm8[1,2],xmm13[3],xmm8[4,5],xmm13[6],xmm8[7]
; AVX1-ONLY-NEXT: vmovdqa {{.*#+}} xmm15 = [0,128,2,3,4,5,6,128,8,9,10,11,12,128,14,15]
; AVX1-ONLY-NEXT: vpshufb %xmm15, %xmm8, %xmm8
; AVX1-ONLY-NEXT: vpshufb %xmm12, %xmm0, %xmm13
; AVX1-ONLY-NEXT: vpor %xmm13, %xmm8, %xmm8
; AVX1-ONLY-NEXT: vmovdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm8 = xmm6[0,0,1,1]
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm6 = xmm6[1,1,2,2]
; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm6, %ymm8, %ymm6
; AVX1-ONLY-NEXT: vpshuflw {{.*#+}} xmm8 = xmm9[1,0,2,2,4,5,6,7]
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm8 = xmm8[0,1,0,1]
; AVX1-ONLY-NEXT: vpshuflw {{.*#+}} xmm9 = xmm9[3,3,3,3,4,5,6,7]
; AVX1-ONLY-NEXT: vpshufhw {{.*#+}} xmm9 = xmm9[0,1,2,3,4,4,4,4]
; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm9, %ymm8, %ymm8
; AVX1-ONLY-NEXT: vmovaps {{.*#+}} ymm9 = [65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535]
; AVX1-ONLY-NEXT: vandps %ymm6, %ymm9, %ymm6
; AVX1-ONLY-NEXT: vandnps %ymm8, %ymm9, %ymm8
; AVX1-ONLY-NEXT: vorps %ymm6, %ymm8, %ymm6
; AVX1-ONLY-NEXT: vextractf128 $1, %ymm6, %xmm8
; AVX1-ONLY-NEXT: vpshufb {{.*#+}} xmm9 = xmm2[2,u],zero,zero,zero,zero,xmm2[3,u],zero,zero,zero,zero,xmm2[4,u],zero,zero
; AVX1-ONLY-NEXT: vpblendw {{.*#+}} xmm8 = xmm9[0],xmm8[1,2],xmm9[3],xmm8[4,5],xmm9[6],xmm8[7]
; AVX1-ONLY-NEXT: vpshufb %xmm15, %xmm8, %xmm8
; AVX1-ONLY-NEXT: vmovdqa {{.*#+}} xmm13 = [128,2,128,128,128,128,128,3,128,128,128,128,128,4,128,128]
; AVX1-ONLY-NEXT: vpshufb %xmm13, %xmm5, %xmm9
; AVX1-ONLY-NEXT: vpor %xmm9, %xmm8, %xmm8
; AVX1-ONLY-NEXT: vmovdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovdqa {{.*#+}} xmm8 = <128,128,128,128,0,u,128,128,128,128,1,u,128,128,128,128>
; AVX1-ONLY-NEXT: vpshufb %xmm8, %xmm2, %xmm2
; AVX1-ONLY-NEXT: vpblendw {{.*#+}} xmm2 = xmm6[0,1],xmm2[2],xmm6[3,4],xmm2[5],xmm6[6,7]
; AVX1-ONLY-NEXT: vmovdqa {{.*#+}} xmm12 = [0,1,2,3,4,128,6,7,8,9,10,128,12,13,14,15]
; AVX1-ONLY-NEXT: vpshufb %xmm12, %xmm2, %xmm2
; AVX1-ONLY-NEXT: vmovdqa {{.*#+}} xmm6 = [128,128,128,128,128,0,128,128,128,128,128,1,128,128,128,128]
; AVX1-ONLY-NEXT: vpshufb %xmm6, %xmm5, %xmm5
; AVX1-ONLY-NEXT: vmovdqa %xmm6, %xmm9
; AVX1-ONLY-NEXT: vpor %xmm5, %xmm2, %xmm2
; AVX1-ONLY-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm14[0],xmm10[0],xmm14[1],xmm10[1],xmm14[2],xmm10[2],xmm14[3],xmm10[3],xmm14[4],xmm10[4],xmm14[5],xmm10[5],xmm14[6],xmm10[6],xmm14[7],xmm10[7]
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm5 = xmm2[0,0,1,1]
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm6 = xmm2[1,1,2,2]
; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm6, %ymm5, %ymm5
; AVX1-ONLY-NEXT: vpunpcklbw {{.*#+}} xmm4 = xmm7[0],xmm4[0],xmm7[1],xmm4[1],xmm7[2],xmm4[2],xmm7[3],xmm4[3],xmm7[4],xmm4[4],xmm7[5],xmm4[5],xmm7[6],xmm4[6],xmm7[7],xmm4[7]
; AVX1-ONLY-NEXT: vpshuflw {{.*#+}} xmm6 = xmm4[1,0,2,2,4,5,6,7]
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm6 = xmm6[0,1,0,1]
; AVX1-ONLY-NEXT: vpshuflw {{.*#+}} xmm7 = xmm4[3,3,3,3,4,5,6,7]
; AVX1-ONLY-NEXT: vpshufhw {{.*#+}} xmm7 = xmm7[0,1,2,3,4,4,4,4]
; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm7, %ymm6, %ymm6
; AVX1-ONLY-NEXT: vmovaps {{.*#+}} ymm7 = [65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535]
; AVX1-ONLY-NEXT: vandps %ymm7, %ymm5, %ymm5
; AVX1-ONLY-NEXT: vandnps %ymm6, %ymm7, %ymm6
; AVX1-ONLY-NEXT: vorps %ymm6, %ymm5, %ymm5
; AVX1-ONLY-NEXT: vextractf128 $1, %ymm5, %xmm6
; AVX1-ONLY-NEXT: vpshufb {{.*#+}} xmm7 = xmm11[2,u],zero,zero,zero,zero,xmm11[3,u],zero,zero,zero,zero,xmm11[4,u],zero,zero
; AVX1-ONLY-NEXT: vpblendw {{.*#+}} xmm6 = xmm7[0],xmm6[1,2],xmm7[3],xmm6[4,5],xmm7[6],xmm6[7]
; AVX1-ONLY-NEXT: vpshufb %xmm15, %xmm6, %xmm6
; AVX1-ONLY-NEXT: vmovdqa %xmm15, %xmm14
; AVX1-ONLY-NEXT: vpshufb %xmm13, %xmm0, %xmm7
; AVX1-ONLY-NEXT: vpor %xmm7, %xmm6, %xmm6
; AVX1-ONLY-NEXT: vmovdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vpshufb %xmm8, %xmm11, %xmm6
; AVX1-ONLY-NEXT: vpblendw {{.*#+}} xmm5 = xmm5[0,1],xmm6[2],xmm5[3,4],xmm6[5],xmm5[6,7]
; AVX1-ONLY-NEXT: vmovdqa %xmm12, %xmm7
; AVX1-ONLY-NEXT: vpshufb %xmm12, %xmm5, %xmm5
; AVX1-ONLY-NEXT: vpshufb %xmm9, %xmm0, %xmm6
; AVX1-ONLY-NEXT: vpor %xmm6, %xmm5, %xmm5
; AVX1-ONLY-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[3,3,3,3]
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[0,0,1,1]
; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm1, %ymm2, %ymm1
; AVX1-ONLY-NEXT: vpshufhw {{.*#+}} xmm2 = xmm4[0,1,2,3,5,6,7,7]
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[2,2,2,3]
; AVX1-ONLY-NEXT: vpshuflw {{.*#+}} xmm3 = xmm3[1,0,2,2,4,5,6,7]
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm3 = xmm3[0,1,0,1]
; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm3, %ymm2, %ymm2
; AVX1-ONLY-NEXT: vmovaps {{.*#+}} ymm12 = [65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535]
; AVX1-ONLY-NEXT: vandnps %ymm1, %ymm12, %ymm1
; AVX1-ONLY-NEXT: vandps %ymm2, %ymm12, %ymm2
; AVX1-ONLY-NEXT: vmovaps %ymm12, %ymm15
; AVX1-ONLY-NEXT: vorps %ymm1, %ymm2, %ymm1
; AVX1-ONLY-NEXT: vextractf128 $1, %ymm1, %xmm2
; AVX1-ONLY-NEXT: vmovdqa {{.*#+}} xmm5 = <128,128,128,128,8,u,128,128,128,128,9,u,128,128,128,128>
; AVX1-ONLY-NEXT: vpshufb %xmm5, %xmm11, %xmm3
; AVX1-ONLY-NEXT: vpblendw {{.*#+}} xmm2 = xmm2[0,1],xmm3[2],xmm2[3,4],xmm3[5],xmm2[6,7]
; AVX1-ONLY-NEXT: vpshufb %xmm7, %xmm2, %xmm2
; AVX1-ONLY-NEXT: vmovdqa %xmm7, %xmm13
; AVX1-ONLY-NEXT: vmovdqa {{.*#+}} xmm6 = [128,128,128,128,128,8,128,128,128,128,128,9,128,128,128,128]
; AVX1-ONLY-NEXT: vpshufb %xmm6, %xmm0, %xmm3
; AVX1-ONLY-NEXT: vpor %xmm3, %xmm2, %xmm2
; AVX1-ONLY-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovdqa {{.*#+}} xmm7 = <128,128,5,u,128,128,128,128,6,u,128,128,128,128,7,u>
; AVX1-ONLY-NEXT: vpshufb %xmm7, %xmm11, %xmm2
; AVX1-ONLY-NEXT: vpblendw {{.*#+}} xmm1 = xmm1[0],xmm2[1],xmm1[2,3],xmm2[4],xmm1[5,6],xmm2[7]
; AVX1-ONLY-NEXT: vmovdqa {{.*#+}} xmm12 = [0,1,2,128,4,5,6,7,8,128,10,11,12,13,14,128]
; AVX1-ONLY-NEXT: vpshufb %xmm12, %xmm1, %xmm1
; AVX1-ONLY-NEXT: vmovdqa {{.*#+}} xmm8 = [128,128,128,5,128,128,128,128,128,6,128,128,128,128,128,7]
; AVX1-ONLY-NEXT: vpshufb %xmm8, %xmm0, %xmm0
; AVX1-ONLY-NEXT: vpor %xmm0, %xmm1, %xmm0
; AVX1-ONLY-NEXT: vmovdqa %xmm0, (%rsp) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovdqa 16(%rsi), %xmm0
; AVX1-ONLY-NEXT: vmovdqa 16(%rdi), %xmm1
; AVX1-ONLY-NEXT: vpunpcklbw {{.*#+}} xmm11 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3],xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
; AVX1-ONLY-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm1[8],xmm0[8],xmm1[9],xmm0[9],xmm1[10],xmm0[10],xmm1[11],xmm0[11],xmm1[12],xmm0[12],xmm1[13],xmm0[13],xmm1[14],xmm0[14],xmm1[15],xmm0[15]
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm1 = xmm11[3,3,3,3]
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm2 = xmm0[0,0,1,1]
; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm2, %ymm1, %ymm1
; AVX1-ONLY-NEXT: vmovdqa 16(%rcx), %xmm2
; AVX1-ONLY-NEXT: vmovdqa 16(%rdx), %xmm3
; AVX1-ONLY-NEXT: vpunpcklbw {{.*#+}} xmm10 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3],xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
; AVX1-ONLY-NEXT: vpunpckhbw {{.*#+}} xmm2 = 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]
; AVX1-ONLY-NEXT: vpshufhw {{.*#+}} xmm3 = xmm10[0,1,2,3,5,6,7,7]
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm3 = xmm3[2,2,2,3]
; AVX1-ONLY-NEXT: vpshuflw {{.*#+}} xmm4 = xmm2[1,0,2,2,4,5,6,7]
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm4 = xmm4[0,1,0,1]
; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm4, %ymm3, %ymm3
; AVX1-ONLY-NEXT: vandnps %ymm1, %ymm15, %ymm1
; AVX1-ONLY-NEXT: vandps %ymm3, %ymm15, %ymm3
; AVX1-ONLY-NEXT: vorps %ymm1, %ymm3, %ymm4
; AVX1-ONLY-NEXT: vmovdqa 16(%r8), %xmm3
; AVX1-ONLY-NEXT: vextractf128 $1, %ymm4, %xmm1
; AVX1-ONLY-NEXT: vpshufb %xmm5, %xmm3, %xmm5
; AVX1-ONLY-NEXT: vpblendw {{.*#+}} xmm1 = xmm1[0,1],xmm5[2],xmm1[3,4],xmm5[5],xmm1[6,7]
; AVX1-ONLY-NEXT: vpshufb %xmm13, %xmm1, %xmm5
; AVX1-ONLY-NEXT: vmovdqa 16(%r9), %xmm1
; AVX1-ONLY-NEXT: vpshufb %xmm6, %xmm1, %xmm6
; AVX1-ONLY-NEXT: vpor %xmm6, %xmm5, %xmm5
; AVX1-ONLY-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vpshufb %xmm7, %xmm3, %xmm5
; AVX1-ONLY-NEXT: vpblendw {{.*#+}} xmm4 = xmm4[0],xmm5[1],xmm4[2,3],xmm5[4],xmm4[5,6],xmm5[7]
; AVX1-ONLY-NEXT: vpshufb %xmm12, %xmm4, %xmm4
; AVX1-ONLY-NEXT: vpshufb %xmm8, %xmm1, %xmm5
; AVX1-ONLY-NEXT: vpor %xmm5, %xmm4, %xmm4
; AVX1-ONLY-NEXT: vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm4 = xmm0[1,1,2,2]
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[3,3,3,3]
; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm0, %ymm4, %ymm0
; AVX1-ONLY-NEXT: vpshuflw {{.*#+}} xmm4 = xmm2[3,3,3,3,4,5,6,7]
; AVX1-ONLY-NEXT: vpshufhw {{.*#+}} xmm4 = xmm4[0,1,2,3,4,4,4,4]
; AVX1-ONLY-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,5,6,7,7]
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[2,2,2,3]
; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm2, %ymm4, %ymm2
; AVX1-ONLY-NEXT: vandps %ymm0, %ymm15, %ymm0
; AVX1-ONLY-NEXT: vandnps %ymm2, %ymm15, %ymm2
; AVX1-ONLY-NEXT: vorps %ymm2, %ymm0, %ymm0
; AVX1-ONLY-NEXT: vextractf128 $1, %ymm0, %xmm2
; AVX1-ONLY-NEXT: vmovdqa {{.*#+}} xmm13 = <128,128,13,u,128,128,128,128,14,u,128,128,128,128,15,u>
; AVX1-ONLY-NEXT: vpshufb %xmm13, %xmm3, %xmm4
; AVX1-ONLY-NEXT: vpblendw {{.*#+}} xmm2 = xmm2[0],xmm4[1],xmm2[2,3],xmm4[4],xmm2[5,6],xmm4[7]
; AVX1-ONLY-NEXT: vpshufb %xmm12, %xmm2, %xmm2
; AVX1-ONLY-NEXT: vmovdqa %xmm12, %xmm13
; AVX1-ONLY-NEXT: vmovdqa {{.*#+}} xmm12 = [128,128,128,13,128,128,128,128,128,14,128,128,128,128,128,15]
; AVX1-ONLY-NEXT: vpshufb %xmm12, %xmm1, %xmm4
; AVX1-ONLY-NEXT: vpor %xmm4, %xmm2, %xmm2
; AVX1-ONLY-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vpshufb {{.*#+}} xmm2 = xmm3[10,u],zero,zero,zero,zero,xmm3[11,u],zero,zero,zero,zero,xmm3[12,u],zero,zero
; AVX1-ONLY-NEXT: vpblendw {{.*#+}} xmm0 = xmm2[0],xmm0[1,2],xmm2[3],xmm0[4,5],xmm2[6],xmm0[7]
; AVX1-ONLY-NEXT: vmovdqa %xmm14, %xmm4
; AVX1-ONLY-NEXT: vpshufb %xmm14, %xmm0, %xmm0
; AVX1-ONLY-NEXT: vpshufb {{.*#+}} xmm2 = zero,xmm1[10],zero,zero,zero,zero,zero,xmm1[11],zero,zero,zero,zero,zero,xmm1[12],zero,zero
; AVX1-ONLY-NEXT: vpor %xmm2, %xmm0, %xmm0
; AVX1-ONLY-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovdqa (%rsi), %xmm6
; AVX1-ONLY-NEXT: vmovdqa (%rdi), %xmm5
; AVX1-ONLY-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm5[8],xmm6[8],xmm5[9],xmm6[9],xmm5[10],xmm6[10],xmm5[11],xmm6[11],xmm5[12],xmm6[12],xmm5[13],xmm6[13],xmm5[14],xmm6[14],xmm5[15],xmm6[15]
; AVX1-ONLY-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm0 = xmm2[1,1,2,2]
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[3,3,3,3]
; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm0
; AVX1-ONLY-NEXT: vmovdqa (%rcx), %xmm8
; AVX1-ONLY-NEXT: vmovdqa (%rdx), %xmm7
; AVX1-ONLY-NEXT: vpunpckhbw {{.*#+}} xmm9 = xmm7[8],xmm8[8],xmm7[9],xmm8[9],xmm7[10],xmm8[10],xmm7[11],xmm8[11],xmm7[12],xmm8[12],xmm7[13],xmm8[13],xmm7[14],xmm8[14],xmm7[15],xmm8[15]
; AVX1-ONLY-NEXT: vpshuflw {{.*#+}} xmm2 = xmm9[3,3,3,3,4,5,6,7]
; AVX1-ONLY-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,4,4,4]
; AVX1-ONLY-NEXT: vpshufhw {{.*#+}} xmm14 = xmm9[0,1,2,3,5,6,7,7]
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm14 = xmm14[2,2,2,3]
; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm14, %ymm2, %ymm2
; AVX1-ONLY-NEXT: vandps %ymm0, %ymm15, %ymm0
; AVX1-ONLY-NEXT: vandnps %ymm2, %ymm15, %ymm2
; AVX1-ONLY-NEXT: vorps %ymm2, %ymm0, %ymm14
; AVX1-ONLY-NEXT: vmovdqa (%r8), %xmm2
; AVX1-ONLY-NEXT: vpshufb {{.*#+}} xmm0 = zero,zero,xmm2[13,u],zero,zero,zero,zero,xmm2[14,u],zero,zero,zero,zero,xmm2[15,u]
; AVX1-ONLY-NEXT: vextractf128 $1, %ymm14, %xmm15
; AVX1-ONLY-NEXT: vpblendw {{.*#+}} xmm0 = xmm15[0],xmm0[1],xmm15[2,3],xmm0[4],xmm15[5,6],xmm0[7]
; AVX1-ONLY-NEXT: vpshufb %xmm13, %xmm0, %xmm15
; AVX1-ONLY-NEXT: vmovdqa (%r9), %xmm0
; AVX1-ONLY-NEXT: vpshufb %xmm12, %xmm0, %xmm13
; AVX1-ONLY-NEXT: vpor %xmm13, %xmm15, %xmm12
; AVX1-ONLY-NEXT: vmovdqa %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vpshufb {{.*#+}} xmm13 = xmm2[10,u],zero,zero,zero,zero,xmm2[11,u],zero,zero,zero,zero,xmm2[12,u],zero,zero
; AVX1-ONLY-NEXT: vpblendw {{.*#+}} xmm13 = xmm13[0],xmm14[1,2],xmm13[3],xmm14[4,5],xmm13[6],xmm14[7]
; AVX1-ONLY-NEXT: vpshufb %xmm4, %xmm13, %xmm13
; AVX1-ONLY-NEXT: vpshufb {{.*#+}} xmm14 = zero,xmm0[10],zero,zero,zero,zero,zero,xmm0[11],zero,zero,zero,zero,zero,xmm0[12],zero,zero
; AVX1-ONLY-NEXT: vpor %xmm14, %xmm13, %xmm12
; AVX1-ONLY-NEXT: vmovdqa %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm13 = xmm11[0,0,1,1]
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm11 = xmm11[1,1,2,2]
; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm11, %ymm13, %ymm11
; AVX1-ONLY-NEXT: vpshuflw {{.*#+}} xmm13 = xmm10[1,0,2,2,4,5,6,7]
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm13 = xmm13[0,1,0,1]
; AVX1-ONLY-NEXT: vpshuflw {{.*#+}} xmm10 = xmm10[3,3,3,3,4,5,6,7]
; AVX1-ONLY-NEXT: vpshufhw {{.*#+}} xmm10 = xmm10[0,1,2,3,4,4,4,4]
; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm10, %ymm13, %ymm10
; AVX1-ONLY-NEXT: vmovaps {{.*#+}} ymm12 = [65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535]
; AVX1-ONLY-NEXT: vandps %ymm12, %ymm11, %ymm11
; AVX1-ONLY-NEXT: vandnps %ymm10, %ymm12, %ymm10
; AVX1-ONLY-NEXT: vorps %ymm10, %ymm11, %ymm11
; AVX1-ONLY-NEXT: vextractf128 $1, %ymm11, %xmm10
; AVX1-ONLY-NEXT: vmovdqa {{.*#+}} xmm15 = <2,u,128,128,128,128,3,u,128,128,128,128,4,u,128,128>
; AVX1-ONLY-NEXT: vpshufb %xmm15, %xmm3, %xmm13
; AVX1-ONLY-NEXT: vpblendw {{.*#+}} xmm10 = xmm13[0],xmm10[1,2],xmm13[3],xmm10[4,5],xmm13[6],xmm10[7]
; AVX1-ONLY-NEXT: vpshufb %xmm4, %xmm10, %xmm10
; AVX1-ONLY-NEXT: vmovdqa {{.*#+}} xmm14 = [128,2,128,128,128,128,128,3,128,128,128,128,128,4,128,128]
; AVX1-ONLY-NEXT: vpshufb %xmm14, %xmm1, %xmm13
; AVX1-ONLY-NEXT: vpor %xmm13, %xmm10, %xmm4
; AVX1-ONLY-NEXT: vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX1-ONLY-NEXT: vmovdqa {{.*#+}} xmm13 = <128,128,128,128,0,u,128,128,128,128,1,u,128,128,128,128>
; AVX1-ONLY-NEXT: vpshufb %xmm13, %xmm3, %xmm3
; AVX1-ONLY-NEXT: vpblendw {{.*#+}} xmm3 = xmm11[0,1],xmm3[2],xmm11[3,4],xmm3[5],xmm11[6,7]
; AVX1-ONLY-NEXT: vmovdqa {{.*#+}} xmm11 = [0,1,2,3,4,128,6,7,8,9,10,128,12,13,14,15]
; AVX1-ONLY-NEXT: vpshufb %xmm11, %xmm3, %xmm3
; AVX1-ONLY-NEXT: vmovdqa {{.*#+}} xmm10 = [128,128,128,128,128,0,128,128,128,128,128,1,128,128,128,128]
; AVX1-ONLY-NEXT: vpshufb %xmm10, %xmm1, %xmm1
; AVX1-ONLY-NEXT: vpor %xmm1, %xmm3, %xmm1
; AVX1-ONLY-NEXT: vpunpcklbw {{.*#+}} xmm4 = 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]
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm3 = xmm4[0,0,1,1]
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm5 = xmm4[1,1,2,2]
; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm5, %ymm3, %ymm3
; AVX1-ONLY-NEXT: vpunpcklbw {{.*#+}} xmm5 = xmm7[0],xmm8[0],xmm7[1],xmm8[1],xmm7[2],xmm8[2],xmm7[3],xmm8[3],xmm7[4],xmm8[4],xmm7[5],xmm8[5],xmm7[6],xmm8[6],xmm7[7],xmm8[7]
; AVX1-ONLY-NEXT: vpshuflw {{.*#+}} xmm6 = xmm5[1,0,2,2,4,5,6,7]
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm6 = xmm6[0,1,0,1]
; AVX1-ONLY-NEXT: vpshuflw {{.*#+}} xmm7 = xmm5[3,3,3,3,4,5,6,7]
; AVX1-ONLY-NEXT: vpshufhw {{.*#+}} xmm7 = xmm7[0,1,2,3,4,4,4,4]
; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm7, %ymm6, %ymm6
; AVX1-ONLY-NEXT: vandps %ymm3, %ymm12, %ymm3
; AVX1-ONLY-NEXT: vandnps %ymm6, %ymm12, %ymm6
; AVX1-ONLY-NEXT: vorps %ymm6, %ymm3, %ymm6
; AVX1-ONLY-NEXT: vpshufb %xmm15, %xmm2, %xmm3
; AVX1-ONLY-NEXT: vextractf128 $1, %ymm6, %xmm7
; AVX1-ONLY-NEXT: vpblendw {{.*#+}} xmm3 = xmm3[0],xmm7[1,2],xmm3[3],xmm7[4,5],xmm3[6],xmm7[7]
; AVX1-ONLY-NEXT: vpshufb {{.*#+}} xmm3 = xmm3[0],zero,xmm3[2,3,4,5,6],zero,xmm3[8,9,10,11,12],zero,xmm3[14,15]
; AVX1-ONLY-NEXT: vpshufb %xmm14, %xmm0, %xmm7
; AVX1-ONLY-NEXT: vpor %xmm7, %xmm3, %xmm3
; AVX1-ONLY-NEXT: vpshufb %xmm13, %xmm2, %xmm7
; AVX1-ONLY-NEXT: vpblendw {{.*#+}} xmm6 = xmm6[0,1],xmm7[2],xmm6[3,4],xmm7[5],xmm6[6,7]
; AVX1-ONLY-NEXT: vpshufb %xmm11, %xmm6, %xmm6
; AVX1-ONLY-NEXT: vpshufb %xmm10, %xmm0, %xmm7
; AVX1-ONLY-NEXT: vpor %xmm7, %xmm6, %xmm6
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm4 = xmm4[3,3,3,3]
; AVX1-ONLY-NEXT: vpermilps $80, {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Folded Reload
; AVX1-ONLY-NEXT: # xmm7 = mem[0,0,1,1]
; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm7, %ymm4, %ymm4
; AVX1-ONLY-NEXT: vpshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,5,6,7,7]
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm5 = xmm5[2,2,2,3]
; AVX1-ONLY-NEXT: vpshuflw {{.*#+}} xmm7 = xmm9[1,0,2,2,4,5,6,7]
; AVX1-ONLY-NEXT: vpshufd {{.*#+}} xmm7 = xmm7[0,1,0,1]
; AVX1-ONLY-NEXT: vinsertf128 $1, %xmm7, %ymm5, %ymm5
; AVX1-ONLY-NEXT: vmovaps {{.*#+}} ymm7 = [65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535]
; AVX1-ONLY-NEXT: vandnps %ymm4, %ymm7, %ymm4
; AVX1-ONLY-NEXT: vandps %ymm7, %ymm5, %ymm5
; AVX1-ONLY-NEXT: vorps %ymm4, %ymm5, %ymm4
; AVX1-ONLY-NEXT: vpshufb {{.*#+}} xmm5 = zero,zero,zero,zero,xmm2[8,u],zero,zero,zero,zero,xmm2[9,u],zero,zero,zero,zero
; AVX1-ONLY-NEXT: vextractf128 $1, %ymm4, %xmm7
; AVX1-ONLY-NEXT: vpblendw {{.*#+}} xmm5 = xmm7[0,1],xmm5[2],xmm7[3,4],xmm5[5],xmm7[6,7]
; AVX1-ONLY-NEXT: vpshufb %xmm11, %xmm5, %xmm5
; AVX1-ONLY-NEXT: vpshufb {{.*#+}} xmm7 = zero,zero,zero,zero,zero,xmm0[8],zero,zero,zero,zero,zero,xmm0[9],zero,zero,zero,zero
; AVX1-ONLY-NEXT: vpor %xmm7, %xmm5, %xmm5
; AVX1-ONLY-NEXT: vpshufb {{.*#+}} xmm2 = zero,zero,xmm2[5,u],zero,zero,zero,zero,xmm2[6,u],zero,zero,zero,zero,xmm2[7,u]
; AVX1-ONLY-NEXT: vpblendw {{.*#+}} xmm2 = xmm4[0],xmm2[1],xmm4[2,3],xmm2[4],xmm4[5,6],xmm2[7]
; AVX1-ONLY-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[0,1,2],zero,xmm2[4,5,6,7,8],zero,xmm2[10,11,12,13,14],zero
; AVX1-ONLY-NEXT: vpshufb {{.*#+}} xmm0 = zero,zero,zero,xmm0[5],zero,zero,zero,zero,zero,xmm0[6],zero,zero,zero,zero,zero,xmm0[7]
; AVX1-ONLY-NEXT: vpor %xmm0, %xmm2, %xmm0
; AVX1-ONLY-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX1-ONLY-NEXT: vmovdqa %xmm0, 32(%rax)
; AVX1-ONLY-NEXT: vmovdqa %xmm5, 48(%rax)
; AVX1-ONLY-NEXT: vmovdqa %xmm6, (%rax)
; AVX1-ONLY-NEXT: vmovdqa %xmm3, 16(%rax)
; AVX1-ONLY-NEXT: vmovdqa %xmm1, 96(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm0, 112(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm0, 64(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm0, 80(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm0, 160(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm0, 176(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm0, 128(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm0, 144(%rax)
; AVX1-ONLY-NEXT: vmovaps (%rsp), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm0, 224(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm0, 240(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm0, 192(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm0, 208(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm0, 288(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm0, 304(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm0, 256(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm0, 272(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm0, 352(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm0, 368(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm0, 320(%rax)
; AVX1-ONLY-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX1-ONLY-NEXT: vmovaps %xmm0, 336(%rax)
; AVX1-ONLY-NEXT: addq $200, %rsp
; AVX1-ONLY-NEXT: vzeroupper
; AVX1-ONLY-NEXT: retq
;
; AVX2-SLOW-LABEL: store_i8_stride6_vf64:
; AVX2-SLOW: # %bb.0:
; AVX2-SLOW-NEXT: subq $696, %rsp # imm = 0x2B8
; AVX2-SLOW-NEXT: vmovdqa 32(%rdx), %ymm6
; AVX2-SLOW-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT: vmovdqa 32(%rcx), %ymm5
; AVX2-SLOW-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT: vmovdqa (%rcx), %xmm0
; AVX2-SLOW-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-SLOW-NEXT: vmovdqa 32(%rcx), %xmm7
; AVX2-SLOW-NEXT: vmovdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-SLOW-NEXT: vpbroadcastq {{.*#+}} xmm2 = [5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10]
; AVX2-SLOW-NEXT: vpshufb %xmm2, %xmm0, %xmm0
; AVX2-SLOW-NEXT: vmovdqa (%rdx), %xmm1
; AVX2-SLOW-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-SLOW-NEXT: vmovdqa 32(%rdx), %xmm11
; AVX2-SLOW-NEXT: vpshufb %xmm2, %xmm1, %xmm1
; AVX2-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3],xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,0,0,1]
; AVX2-SLOW-NEXT: vmovdqa (%rsi), %xmm14
; AVX2-SLOW-NEXT: vmovdqa 32(%rsi), %xmm12
; AVX2-SLOW-NEXT: vpbroadcastq {{.*#+}} xmm3 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
; AVX2-SLOW-NEXT: vpshufb %xmm3, %xmm14, %xmm1
; AVX2-SLOW-NEXT: vmovdqa %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-SLOW-NEXT: vmovdqa (%rdi), %xmm13
; AVX2-SLOW-NEXT: vmovdqa 32(%rdi), %xmm9
; AVX2-SLOW-NEXT: vpshufb %xmm3, %xmm13, %xmm4
; AVX2-SLOW-NEXT: vmovdqa %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm4[8],xmm1[8],xmm4[9],xmm1[9],xmm4[10],xmm1[10],xmm4[11],xmm1[11],xmm4[12],xmm1[12],xmm4[13],xmm1[13],xmm4[14],xmm1[14],xmm4[15],xmm1[15]
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm4 = ymm1[0,0,0,1]
; AVX2-SLOW-NEXT: vmovdqa {{.*#+}} ymm1 = <255,255,u,u,0,0,255,255,u,u,0,0,255,255,u,u,0,0,255,255,u,u,0,0,255,255,u,u,0,0,255,255>
; AVX2-SLOW-NEXT: vpblendvb %ymm1, %ymm0, %ymm4, %ymm0
; AVX2-SLOW-NEXT: vpshufb %xmm2, %xmm7, %xmm4
; AVX2-SLOW-NEXT: vpshufb %xmm2, %xmm11, %xmm2
; AVX2-SLOW-NEXT: vmovdqa %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3],xmm2[4],xmm4[4],xmm2[5],xmm4[5],xmm2[6],xmm4[6],xmm2[7],xmm4[7]
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,0,0,1]
; AVX2-SLOW-NEXT: vpshufb %xmm3, %xmm12, %xmm4
; AVX2-SLOW-NEXT: vmovdqa %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-SLOW-NEXT: vpshufb %xmm3, %xmm9, %xmm3
; AVX2-SLOW-NEXT: vmovdqa %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm3[8],xmm4[8],xmm3[9],xmm4[9],xmm3[10],xmm4[10],xmm3[11],xmm4[11],xmm3[12],xmm4[12],xmm3[13],xmm4[13],xmm3[14],xmm4[14],xmm3[15],xmm4[15]
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,0,1]
; AVX2-SLOW-NEXT: vpblendvb %ymm1, %ymm2, %ymm3, %ymm2
; AVX2-SLOW-NEXT: vpbroadcastq {{.*#+}} ymm3 = [5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10]
; AVX2-SLOW-NEXT: vpshufb %ymm3, %ymm5, %ymm4
; AVX2-SLOW-NEXT: vpshufb %ymm3, %ymm6, %ymm5
; AVX2-SLOW-NEXT: vpunpcklbw {{.*#+}} ymm4 = ymm5[0],ymm4[0],ymm5[1],ymm4[1],ymm5[2],ymm4[2],ymm5[3],ymm4[3],ymm5[4],ymm4[4],ymm5[5],ymm4[5],ymm5[6],ymm4[6],ymm5[7],ymm4[7],ymm5[16],ymm4[16],ymm5[17],ymm4[17],ymm5[18],ymm4[18],ymm5[19],ymm4[19],ymm5[20],ymm4[20],ymm5[21],ymm4[21],ymm5[22],ymm4[22],ymm5[23],ymm4[23]
; AVX2-SLOW-NEXT: vmovdqa 32(%rdi), %ymm7
; AVX2-SLOW-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT: vmovdqa 32(%rsi), %ymm6
; AVX2-SLOW-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT: vpbroadcastq {{.*#+}} ymm5 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
; AVX2-SLOW-NEXT: vpshufb %ymm5, %ymm6, %ymm6
; AVX2-SLOW-NEXT: vpshufb %ymm5, %ymm7, %ymm7
; AVX2-SLOW-NEXT: vpunpcklbw {{.*#+}} ymm6 = ymm7[0],ymm6[0],ymm7[1],ymm6[1],ymm7[2],ymm6[2],ymm7[3],ymm6[3],ymm7[4],ymm6[4],ymm7[5],ymm6[5],ymm7[6],ymm6[6],ymm7[7],ymm6[7],ymm7[16],ymm6[16],ymm7[17],ymm6[17],ymm7[18],ymm6[18],ymm7[19],ymm6[19],ymm7[20],ymm6[20],ymm7[21],ymm6[21],ymm7[22],ymm6[22],ymm7[23],ymm6[23]
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,2,3]
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,2,2,3]
; AVX2-SLOW-NEXT: vpblendvb %ymm1, %ymm4, %ymm6, %ymm4
; AVX2-SLOW-NEXT: vmovdqa (%rdx), %ymm7
; AVX2-SLOW-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT: vmovdqa (%rcx), %ymm6
; AVX2-SLOW-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT: vpshufb %ymm3, %ymm6, %ymm6
; AVX2-SLOW-NEXT: vpshufb %ymm3, %ymm7, %ymm3
; AVX2-SLOW-NEXT: vpunpcklbw {{.*#+}} ymm3 = ymm3[0],ymm6[0],ymm3[1],ymm6[1],ymm3[2],ymm6[2],ymm3[3],ymm6[3],ymm3[4],ymm6[4],ymm3[5],ymm6[5],ymm3[6],ymm6[6],ymm3[7],ymm6[7],ymm3[16],ymm6[16],ymm3[17],ymm6[17],ymm3[18],ymm6[18],ymm3[19],ymm6[19],ymm3[20],ymm6[20],ymm3[21],ymm6[21],ymm3[22],ymm6[22],ymm3[23],ymm6[23]
; AVX2-SLOW-NEXT: vmovdqa (%rdi), %ymm7
; AVX2-SLOW-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT: vmovdqa (%rsi), %ymm6
; AVX2-SLOW-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT: vpshufb %ymm5, %ymm6, %ymm6
; AVX2-SLOW-NEXT: vpshufb %ymm5, %ymm7, %ymm5
; AVX2-SLOW-NEXT: vpunpcklbw {{.*#+}} ymm5 = ymm5[0],ymm6[0],ymm5[1],ymm6[1],ymm5[2],ymm6[2],ymm5[3],ymm6[3],ymm5[4],ymm6[4],ymm5[5],ymm6[5],ymm5[6],ymm6[6],ymm5[7],ymm6[7],ymm5[16],ymm6[16],ymm5[17],ymm6[17],ymm5[18],ymm6[18],ymm5[19],ymm6[19],ymm5[20],ymm6[20],ymm5[21],ymm6[21],ymm5[22],ymm6[22],ymm5[23],ymm6[23]
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3]
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,2,2,3]
; AVX2-SLOW-NEXT: vpblendvb %ymm1, %ymm3, %ymm5, %ymm1
; AVX2-SLOW-NEXT: vmovdqa (%r8), %xmm6
; AVX2-SLOW-NEXT: vmovdqa {{.*#+}} xmm3 = <6,u,5,u,8,u,7,u,9,u,9,u,9,u,9,u>
; AVX2-SLOW-NEXT: vpshufb %xmm3, %xmm6, %xmm5
; AVX2-SLOW-NEXT: vmovdqa %xmm6, %xmm7
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,0,0,1]
; AVX2-SLOW-NEXT: vmovdqa {{.*#+}} ymm8 = [255,255,0,0,255,255,255,255,0,0,255,255,255,255,0,0,255,255,255,255,0,0,255,255,255,255,0,0,255,255,255,255]
; AVX2-SLOW-NEXT: vpblendvb %ymm8, %ymm0, %ymm5, %ymm0
; AVX2-SLOW-NEXT: vmovdqa 32(%r8), %xmm5
; AVX2-SLOW-NEXT: vmovdqa %xmm5, (%rsp) # 16-byte Spill
; AVX2-SLOW-NEXT: vpshufb %xmm3, %xmm5, %xmm3
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,0,1]
; AVX2-SLOW-NEXT: vpblendvb %ymm8, %ymm2, %ymm3, %ymm3
; AVX2-SLOW-NEXT: vmovdqa 32(%r8), %ymm5
; AVX2-SLOW-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [6,0,5,0,8,0,7,0,0,0,9,0,0,0,0,0,6,0,5,0,8,0,7,0,0,0,9,0,0,0,0,0]
; AVX2-SLOW-NEXT: # ymm2 = mem[0,1,0,1]
; AVX2-SLOW-NEXT: vpshufb %ymm2, %ymm5, %ymm6
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,2,2,3]
; AVX2-SLOW-NEXT: vpblendvb %ymm8, %ymm4, %ymm6, %ymm10
; AVX2-SLOW-NEXT: vmovdqa (%r8), %ymm4
; AVX2-SLOW-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT: vpshufb %ymm2, %ymm4, %ymm2
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
; AVX2-SLOW-NEXT: vpblendvb %ymm8, %ymm1, %ymm2, %ymm1
; AVX2-SLOW-NEXT: vmovdqa (%r9), %xmm2
; AVX2-SLOW-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-SLOW-NEXT: vmovdqa {{.*#+}} xmm4 = <u,6,u,5,u,8,u,7,u,9,u,9,u,9,u,9>
; AVX2-SLOW-NEXT: vpshufb %xmm4, %xmm2, %xmm2
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,0,0,1]
; AVX2-SLOW-NEXT: vmovdqa {{.*#+}} ymm15 = [255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255]
; AVX2-SLOW-NEXT: vpblendvb %ymm15, %ymm0, %ymm2, %ymm0
; AVX2-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT: vmovdqa 32(%r9), %xmm2
; AVX2-SLOW-NEXT: vpshufb %xmm4, %xmm2, %xmm0
; AVX2-SLOW-NEXT: vmovdqa %xmm2, %xmm6
; AVX2-SLOW-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,0,0,1]
; AVX2-SLOW-NEXT: vpblendvb %ymm15, %ymm3, %ymm0, %ymm0
; AVX2-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT: vmovdqa 32(%r9), %ymm2
; AVX2-SLOW-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [0,6,0,5,0,8,0,7,0,0,0,9,0,0,0,0,0,6,0,5,0,8,0,7,0,0,0,9,0,0,0,0]
; AVX2-SLOW-NEXT: # ymm0 = mem[0,1,0,1]
; AVX2-SLOW-NEXT: vpshufb %ymm0, %ymm2, %ymm3
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3]
; AVX2-SLOW-NEXT: vpblendvb %ymm15, %ymm10, %ymm3, %ymm2
; AVX2-SLOW-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT: vmovdqa (%r9), %ymm2
; AVX2-SLOW-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT: vpshufb %ymm0, %ymm2, %ymm0
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3]
; AVX2-SLOW-NEXT: vpblendvb %ymm15, %ymm1, %ymm0, %ymm0
; AVX2-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm13[0],xmm14[0],xmm13[1],xmm14[1],xmm13[2],xmm14[2],xmm13[3],xmm14[3],xmm13[4],xmm14[4],xmm13[5],xmm14[5],xmm13[6],xmm14[6],xmm13[7],xmm14[7]
; AVX2-SLOW-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[0,3,2,1,4,5,6,7]
; AVX2-SLOW-NEXT: vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,5,6,5]
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,0,0,1]
; AVX2-SLOW-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
; AVX2-SLOW-NEXT: vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload
; AVX2-SLOW-NEXT: # xmm1 = xmm1[0],mem[0],xmm1[1],mem[1],xmm1[2],mem[2],xmm1[3],mem[3],xmm1[4],mem[4],xmm1[5],mem[5],xmm1[6],mem[6],xmm1[7],mem[7]
; AVX2-SLOW-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[1,0,3,2,4,5,6,7]
; AVX2-SLOW-NEXT: vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,4,4,4]
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,0,1]
; AVX2-SLOW-NEXT: vmovdqa {{.*#+}} ymm10 = <255,255,0,0,u,u,255,255,0,0,u,u,255,255,0,0,u,u,255,255,0,0,u,u,255,255,0,0,u,u,255,255>
; AVX2-SLOW-NEXT: vpblendvb %ymm10, %ymm0, %ymm1, %ymm2
; AVX2-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm9[0],xmm12[0],xmm9[1],xmm12[1],xmm9[2],xmm12[2],xmm9[3],xmm12[3],xmm9[4],xmm12[4],xmm9[5],xmm12[5],xmm9[6],xmm12[6],xmm9[7],xmm12[7]
; AVX2-SLOW-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[0,3,2,1,4,5,6,7]
; AVX2-SLOW-NEXT: vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,5,6,5]
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,0,0,1]
; AVX2-SLOW-NEXT: vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm11, %xmm1 # 16-byte Folded Reload
; AVX2-SLOW-NEXT: # xmm1 = xmm11[0],mem[0],xmm11[1],mem[1],xmm11[2],mem[2],xmm11[3],mem[3],xmm11[4],mem[4],xmm11[5],mem[5],xmm11[6],mem[6],xmm11[7],mem[7]
; AVX2-SLOW-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[1,0,3,2,4,5,6,7]
; AVX2-SLOW-NEXT: vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,4,4,4]
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,0,1]
; AVX2-SLOW-NEXT: vpblendvb %ymm10, %ymm0, %ymm1, %ymm5
; AVX2-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-SLOW-NEXT: vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-SLOW-NEXT: # ymm0 = ymm0[0],mem[0],ymm0[1],mem[1],ymm0[2],mem[2],ymm0[3],mem[3],ymm0[4],mem[4],ymm0[5],mem[5],ymm0[6],mem[6],ymm0[7],mem[7],ymm0[16],mem[16],ymm0[17],mem[17],ymm0[18],mem[18],ymm0[19],mem[19],ymm0[20],mem[20],ymm0[21],mem[21],ymm0[22],mem[22],ymm0[23],mem[23]
; AVX2-SLOW-NEXT: vpshuflw {{.*#+}} ymm0 = ymm0[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
; AVX2-SLOW-NEXT: vpshufhw {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5,6,5,8,9,10,11,12,13,14,13]
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3]
; AVX2-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload
; AVX2-SLOW-NEXT: vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload
; AVX2-SLOW-NEXT: # ymm1 = ymm1[0],mem[0],ymm1[1],mem[1],ymm1[2],mem[2],ymm1[3],mem[3],ymm1[4],mem[4],ymm1[5],mem[5],ymm1[6],mem[6],ymm1[7],mem[7],ymm1[16],mem[16],ymm1[17],mem[17],ymm1[18],mem[18],ymm1[19],mem[19],ymm1[20],mem[20],ymm1[21],mem[21],ymm1[22],mem[22],ymm1[23],mem[23]
; AVX2-SLOW-NEXT: vpshuflw {{.*#+}} ymm1 = ymm1[1,0,3,2,4,5,6,7,9,8,11,10,12,13,14,15]
; AVX2-SLOW-NEXT: vpshufhw {{.*#+}} ymm1 = ymm1[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12]
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
; AVX2-SLOW-NEXT: vpblendvb %ymm10, %ymm0, %ymm1, %ymm3
; AVX2-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Reload
; AVX2-SLOW-NEXT: vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %ymm12, %ymm1 # 32-byte Folded Reload
; AVX2-SLOW-NEXT: # ymm1 = ymm12[0],mem[0],ymm12[1],mem[1],ymm12[2],mem[2],ymm12[3],mem[3],ymm12[4],mem[4],ymm12[5],mem[5],ymm12[6],mem[6],ymm12[7],mem[7],ymm12[16],mem[16],ymm12[17],mem[17],ymm12[18],mem[18],ymm12[19],mem[19],ymm12[20],mem[20],ymm12[21],mem[21],ymm12[22],mem[22],ymm12[23],mem[23]
; AVX2-SLOW-NEXT: vpshuflw {{.*#+}} ymm1 = ymm1[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
; AVX2-SLOW-NEXT: vpshufhw {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5,6,5,8,9,10,11,12,13,14,13]
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
; AVX2-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload
; AVX2-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX2-SLOW-NEXT: vpunpcklbw {{.*#+}} ymm15 = ymm8[0],ymm11[0],ymm8[1],ymm11[1],ymm8[2],ymm11[2],ymm8[3],ymm11[3],ymm8[4],ymm11[4],ymm8[5],ymm11[5],ymm8[6],ymm11[6],ymm8[7],ymm11[7],ymm8[16],ymm11[16],ymm8[17],ymm11[17],ymm8[18],ymm11[18],ymm8[19],ymm11[19],ymm8[20],ymm11[20],ymm8[21],ymm11[21],ymm8[22],ymm11[22],ymm8[23],ymm11[23]
; AVX2-SLOW-NEXT: vpshuflw {{.*#+}} ymm15 = ymm15[1,0,3,2,4,5,6,7,9,8,11,10,12,13,14,15]
; AVX2-SLOW-NEXT: vpshufhw {{.*#+}} ymm15 = ymm15[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12]
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,2,2,3]
; AVX2-SLOW-NEXT: vpblendvb %ymm10, %ymm1, %ymm15, %ymm1
; AVX2-SLOW-NEXT: vmovdqa {{.*#+}} xmm10 = <2,u,1,u,0,u,3,u,4,u,4,u,4,u,4,u>
; AVX2-SLOW-NEXT: vmovdqa %xmm7, %xmm13
; AVX2-SLOW-NEXT: vpshufb %xmm10, %xmm7, %xmm15
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm15 = ymm15[0,0,0,1]
; AVX2-SLOW-NEXT: vmovdqa {{.*#+}} ymm0 = [255,255,255,255,0,0,255,255,255,255,0,0,255,255,255,255,0,0,255,255,255,255,0,0,255,255,255,255,0,0,255,255]
; AVX2-SLOW-NEXT: vpblendvb %ymm0, %ymm2, %ymm15, %ymm15
; AVX2-SLOW-NEXT: vmovdqa (%rsp), %xmm4 # 16-byte Reload
; AVX2-SLOW-NEXT: vpshufb %xmm10, %xmm4, %xmm10
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,0,0,1]
; AVX2-SLOW-NEXT: vpblendvb %ymm0, %ymm5, %ymm10, %ymm10
; AVX2-SLOW-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [2,0,1,0,0,0,3,0,0,0,0,0,4,0,0,0,2,0,1,0,0,0,3,0,0,0,0,0,4,0,0,0]
; AVX2-SLOW-NEXT: # ymm2 = mem[0,1,0,1]
; AVX2-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX2-SLOW-NEXT: vpshufb %ymm2, %ymm5, %ymm14
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,2,2,3]
; AVX2-SLOW-NEXT: vpblendvb %ymm0, %ymm3, %ymm14, %ymm3
; AVX2-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload
; AVX2-SLOW-NEXT: vpshufb %ymm2, %ymm9, %ymm2
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
; AVX2-SLOW-NEXT: vpblendvb %ymm0, %ymm1, %ymm2, %ymm0
; AVX2-SLOW-NEXT: vmovdqa {{.*#+}} xmm1 = <u,2,u,1,u,0,u,3,u,4,u,4,u,4,u,4>
; AVX2-SLOW-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload
; AVX2-SLOW-NEXT: vpshufb %xmm1, %xmm9, %xmm2
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,0,0,1]
; AVX2-SLOW-NEXT: vmovdqa {{.*#+}} ymm14 = [255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255]
; AVX2-SLOW-NEXT: vpblendvb %ymm14, %ymm15, %ymm2, %ymm2
; AVX2-SLOW-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT: vpshufb %xmm1, %xmm6, %xmm1
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,0,1]
; AVX2-SLOW-NEXT: vpblendvb %ymm14, %ymm10, %ymm1, %ymm1
; AVX2-SLOW-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT: vbroadcasti128 {{.*#+}} ymm10 = [0,2,0,1,0,0,0,3,0,0,0,0,0,4,0,0,0,2,0,1,0,0,0,3,0,0,0,0,0,4,0,0]
; AVX2-SLOW-NEXT: # ymm10 = mem[0,1,0,1]
; AVX2-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload
; AVX2-SLOW-NEXT: vpshufb %ymm10, %ymm6, %ymm1
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
; AVX2-SLOW-NEXT: vpblendvb %ymm14, %ymm3, %ymm1, %ymm1
; AVX2-SLOW-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
; AVX2-SLOW-NEXT: vpshufb %ymm10, %ymm7, %ymm3
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3]
; AVX2-SLOW-NEXT: vpblendvb %ymm14, %ymm0, %ymm3, %ymm0
; AVX2-SLOW-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-SLOW-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-SLOW-NEXT: vpunpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm3 # 16-byte Folded Reload
; AVX2-SLOW-NEXT: # xmm3 = xmm0[8],mem[8],xmm0[9],mem[9],xmm0[10],mem[10],xmm0[11],mem[11],xmm0[12],mem[12],xmm0[13],mem[13],xmm0[14],mem[14],xmm0[15],mem[15]
; AVX2-SLOW-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-SLOW-NEXT: vpunpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm10 # 16-byte Folded Reload
; AVX2-SLOW-NEXT: # xmm10 = xmm0[8],mem[8],xmm0[9],mem[9],xmm0[10],mem[10],xmm0[11],mem[11],xmm0[12],mem[12],xmm0[13],mem[13],xmm0[14],mem[14],xmm0[15],mem[15]
; AVX2-SLOW-NEXT: vmovdqa {{.*#+}} xmm14 = [8,9,6,7,12,13,10,11,14,15,14,15,14,15,14,15]
; AVX2-SLOW-NEXT: vpshufb %xmm14, %xmm3, %xmm3
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,0,1]
; AVX2-SLOW-NEXT: vmovdqa {{.*#+}} xmm1 = [10,11,8,9,6,7,12,13,14,15,14,15,14,15,14,15]
; AVX2-SLOW-NEXT: vpshufb %xmm1, %xmm10, %xmm10
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,0,0,1]
; AVX2-SLOW-NEXT: vmovdqa {{.*#+}} ymm0 = <u,u,255,255,0,0,u,u,255,255,0,0,u,u,255,255,0,0,u,u,255,255,0,0,u,u,255,255,0,0,u,u>
; AVX2-SLOW-NEXT: vpblendvb %ymm0, %ymm3, %ymm10, %ymm15
; AVX2-SLOW-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX2-SLOW-NEXT: vpunpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm10 # 16-byte Folded Reload
; AVX2-SLOW-NEXT: # xmm10 = xmm2[8],mem[8],xmm2[9],mem[9],xmm2[10],mem[10],xmm2[11],mem[11],xmm2[12],mem[12],xmm2[13],mem[13],xmm2[14],mem[14],xmm2[15],mem[15]
; AVX2-SLOW-NEXT: vpshufb %xmm14, %xmm10, %xmm10
; AVX2-SLOW-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX2-SLOW-NEXT: vpunpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm14 # 16-byte Folded Reload
; AVX2-SLOW-NEXT: # xmm14 = xmm2[8],mem[8],xmm2[9],mem[9],xmm2[10],mem[10],xmm2[11],mem[11],xmm2[12],mem[12],xmm2[13],mem[13],xmm2[14],mem[14],xmm2[15],mem[15]
; AVX2-SLOW-NEXT: vpshufb %xmm1, %xmm14, %xmm1
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,0,0,1]
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,0,1]
; AVX2-SLOW-NEXT: vpblendvb %ymm0, %ymm10, %ymm1, %ymm1
; AVX2-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX2-SLOW-NEXT: vpunpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm10 # 32-byte Folded Reload
; AVX2-SLOW-NEXT: # ymm10 = ymm2[8],mem[8],ymm2[9],mem[9],ymm2[10],mem[10],ymm2[11],mem[11],ymm2[12],mem[12],ymm2[13],mem[13],ymm2[14],mem[14],ymm2[15],mem[15],ymm2[24],mem[24],ymm2[25],mem[25],ymm2[26],mem[26],ymm2[27],mem[27],ymm2[28],mem[28],ymm2[29],mem[29],ymm2[30],mem[30],ymm2[31],mem[31]
; AVX2-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX2-SLOW-NEXT: vpunpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm14 # 32-byte Folded Reload
; AVX2-SLOW-NEXT: # ymm14 = ymm2[8],mem[8],ymm2[9],mem[9],ymm2[10],mem[10],ymm2[11],mem[11],ymm2[12],mem[12],ymm2[13],mem[13],ymm2[14],mem[14],ymm2[15],mem[15],ymm2[24],mem[24],ymm2[25],mem[25],ymm2[26],mem[26],ymm2[27],mem[27],ymm2[28],mem[28],ymm2[29],mem[29],ymm2[30],mem[30],ymm2[31],mem[31]
; AVX2-SLOW-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [24,25,22,23,28,29,26,27,30,31,30,31,30,31,30,31,24,25,22,23,28,29,26,27,30,31,30,31,30,31,30,31]
; AVX2-SLOW-NEXT: # ymm2 = mem[0,1,0,1]
; AVX2-SLOW-NEXT: vpshufb %ymm2, %ymm10, %ymm10
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,2,2,3]
; AVX2-SLOW-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [26,27,24,25,22,23,28,29,30,31,30,31,30,31,30,31,26,27,24,25,22,23,28,29,30,31,30,31,30,31,30,31]
; AVX2-SLOW-NEXT: # ymm3 = mem[0,1,0,1]
; AVX2-SLOW-NEXT: vpshufb %ymm3, %ymm14, %ymm14
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,2,2,3]
; AVX2-SLOW-NEXT: vpblendvb %ymm0, %ymm10, %ymm14, %ymm10
; AVX2-SLOW-NEXT: vpunpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %ymm12, %ymm14 # 32-byte Folded Reload
; AVX2-SLOW-NEXT: # ymm14 = ymm12[8],mem[8],ymm12[9],mem[9],ymm12[10],mem[10],ymm12[11],mem[11],ymm12[12],mem[12],ymm12[13],mem[13],ymm12[14],mem[14],ymm12[15],mem[15],ymm12[24],mem[24],ymm12[25],mem[25],ymm12[26],mem[26],ymm12[27],mem[27],ymm12[28],mem[28],ymm12[29],mem[29],ymm12[30],mem[30],ymm12[31],mem[31]
; AVX2-SLOW-NEXT: vpshufb %ymm2, %ymm14, %ymm2
; AVX2-SLOW-NEXT: vpunpckhbw {{.*#+}} ymm14 = ymm8[8],ymm11[8],ymm8[9],ymm11[9],ymm8[10],ymm11[10],ymm8[11],ymm11[11],ymm8[12],ymm11[12],ymm8[13],ymm11[13],ymm8[14],ymm11[14],ymm8[15],ymm11[15],ymm8[24],ymm11[24],ymm8[25],ymm11[25],ymm8[26],ymm11[26],ymm8[27],ymm11[27],ymm8[28],ymm11[28],ymm8[29],ymm11[29],ymm8[30],ymm11[30],ymm8[31],ymm11[31]
; AVX2-SLOW-NEXT: vpshufb %ymm3, %ymm14, %ymm3
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3]
; AVX2-SLOW-NEXT: vpblendvb %ymm0, %ymm2, %ymm3, %ymm0
; AVX2-SLOW-NEXT: vmovdqa {{.*#+}} xmm2 = <10,u,13,u,12,u,11,u,14,u,13,u,14,u,15,u>
; AVX2-SLOW-NEXT: vpshufb %xmm2, %xmm4, %xmm3
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,0,1]
; AVX2-SLOW-NEXT: vmovdqa {{.*#+}} ymm12 = [0,0,255,255,255,255,0,0,255,255,255,255,0,0,255,255,255,255,0,0,255,255,255,255,0,0,255,255,255,255,0,0]
; AVX2-SLOW-NEXT: vpblendvb %ymm12, %ymm15, %ymm3, %ymm3
; AVX2-SLOW-NEXT: vpshufb %xmm2, %xmm13, %xmm2
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,0,0,1]
; AVX2-SLOW-NEXT: vpblendvb %ymm12, %ymm1, %ymm2, %ymm1
; AVX2-SLOW-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [10,0,13,0,12,0,11,0,14,0,0,0,0,0,15,0,10,0,13,0,12,0,11,0,14,0,0,0,0,0,15,0]
; AVX2-SLOW-NEXT: # ymm2 = mem[0,1,0,1]
; AVX2-SLOW-NEXT: vpshufb %ymm2, %ymm5, %ymm4
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,2,3]
; AVX2-SLOW-NEXT: vpblendvb %ymm12, %ymm10, %ymm4, %ymm4
; AVX2-SLOW-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload
; AVX2-SLOW-NEXT: vpshufb %ymm2, %ymm5, %ymm2
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
; AVX2-SLOW-NEXT: vpblendvb %ymm12, %ymm0, %ymm2, %ymm0
; AVX2-SLOW-NEXT: vmovdqa {{.*#+}} xmm2 = <u,10,u,13,u,12,u,11,u,14,u,13,u,14,u,15>
; AVX2-SLOW-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-SLOW-NEXT: vpshufb %xmm2, %xmm5, %xmm8
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm8 = ymm8[0,0,0,1]
; AVX2-SLOW-NEXT: vmovdqa {{.*#+}} ymm10 = [255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0]
; AVX2-SLOW-NEXT: vpblendvb %ymm10, %ymm3, %ymm8, %ymm3
; AVX2-SLOW-NEXT: vpshufb %xmm2, %xmm9, %xmm2
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,0,0,1]
; AVX2-SLOW-NEXT: vpblendvb %ymm10, %ymm1, %ymm2, %ymm1
; AVX2-SLOW-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [0,10,0,13,0,12,0,11,0,14,0,0,0,0,0,15,0,10,0,13,0,12,0,11,0,14,0,0,0,0,0,15]
; AVX2-SLOW-NEXT: # ymm2 = mem[0,1,0,1]
; AVX2-SLOW-NEXT: vpshufb %ymm2, %ymm6, %ymm8
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,2,3]
; AVX2-SLOW-NEXT: vpblendvb %ymm10, %ymm4, %ymm8, %ymm4
; AVX2-SLOW-NEXT: vpshufb %ymm2, %ymm7, %ymm2
; AVX2-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
; AVX2-SLOW-NEXT: vpblendvb %ymm10, %ymm0, %ymm2, %ymm0
; AVX2-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX2-SLOW-NEXT: vmovdqa %ymm0, 160(%rax)
; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-SLOW-NEXT: vmovaps %ymm0, 128(%rax)
; AVX2-SLOW-NEXT: vmovdqa %ymm4, 352(%rax)
; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-SLOW-NEXT: vmovaps %ymm0, 320(%rax)
; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-SLOW-NEXT: vmovaps %ymm0, 96(%rax)
; AVX2-SLOW-NEXT: vmovdqa %ymm1, 64(%rax)
; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-SLOW-NEXT: vmovaps %ymm0, 224(%rax)
; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-SLOW-NEXT: vmovaps %ymm0, 288(%rax)
; AVX2-SLOW-NEXT: vmovdqa %ymm3, 256(%rax)
; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-SLOW-NEXT: vmovaps %ymm0, 32(%rax)
; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-SLOW-NEXT: vmovaps %ymm0, 192(%rax)
; AVX2-SLOW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-SLOW-NEXT: vmovaps %ymm0, (%rax)
; AVX2-SLOW-NEXT: addq $696, %rsp # imm = 0x2B8
; AVX2-SLOW-NEXT: vzeroupper
; AVX2-SLOW-NEXT: retq
;
; AVX2-FAST-LABEL: store_i8_stride6_vf64:
; AVX2-FAST: # %bb.0:
; AVX2-FAST-NEXT: subq $648, %rsp # imm = 0x288
; AVX2-FAST-NEXT: vmovdqa 32(%rdx), %ymm6
; AVX2-FAST-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT: vmovdqa 32(%rcx), %ymm5
; AVX2-FAST-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT: vmovdqa (%rcx), %xmm0
; AVX2-FAST-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-NEXT: vmovdqa 32(%rcx), %xmm9
; AVX2-FAST-NEXT: vpbroadcastq {{.*#+}} xmm2 = [5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10]
; AVX2-FAST-NEXT: vpshufb %xmm2, %xmm0, %xmm0
; AVX2-FAST-NEXT: vmovdqa (%rdx), %xmm1
; AVX2-FAST-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-NEXT: vmovdqa 32(%rdx), %xmm8
; AVX2-FAST-NEXT: vpshufb %xmm2, %xmm1, %xmm1
; AVX2-FAST-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3],xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,0,0,1]
; AVX2-FAST-NEXT: vmovdqa (%rsi), %xmm1
; AVX2-FAST-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-NEXT: vmovdqa 32(%rsi), %xmm10
; AVX2-FAST-NEXT: vpbroadcastq {{.*#+}} xmm3 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
; AVX2-FAST-NEXT: vpshufb %xmm3, %xmm1, %xmm1
; AVX2-FAST-NEXT: vmovdqa (%rdi), %xmm4
; AVX2-FAST-NEXT: vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-NEXT: vmovdqa 32(%rdi), %xmm11
; AVX2-FAST-NEXT: vpshufb %xmm3, %xmm4, %xmm4
; AVX2-FAST-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm4[8],xmm1[8],xmm4[9],xmm1[9],xmm4[10],xmm1[10],xmm4[11],xmm1[11],xmm4[12],xmm1[12],xmm4[13],xmm1[13],xmm4[14],xmm1[14],xmm4[15],xmm1[15]
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm4 = ymm1[0,0,0,1]
; AVX2-FAST-NEXT: vmovdqa {{.*#+}} ymm1 = <255,255,u,u,0,0,255,255,u,u,0,0,255,255,u,u,0,0,255,255,u,u,0,0,255,255,u,u,0,0,255,255>
; AVX2-FAST-NEXT: vpblendvb %ymm1, %ymm0, %ymm4, %ymm13
; AVX2-FAST-NEXT: vpshufb %xmm2, %xmm9, %xmm4
; AVX2-FAST-NEXT: vmovdqa %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-NEXT: vpshufb %xmm2, %xmm8, %xmm2
; AVX2-FAST-NEXT: vmovdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3],xmm2[4],xmm4[4],xmm2[5],xmm4[5],xmm2[6],xmm4[6],xmm2[7],xmm4[7]
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,0,0,1]
; AVX2-FAST-NEXT: vpshufb %xmm3, %xmm10, %xmm4
; AVX2-FAST-NEXT: vmovdqa %xmm10, (%rsp) # 16-byte Spill
; AVX2-FAST-NEXT: vpshufb %xmm3, %xmm11, %xmm3
; AVX2-FAST-NEXT: vmovdqa %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm3[8],xmm4[8],xmm3[9],xmm4[9],xmm3[10],xmm4[10],xmm3[11],xmm4[11],xmm3[12],xmm4[12],xmm3[13],xmm4[13],xmm3[14],xmm4[14],xmm3[15],xmm4[15]
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,0,1]
; AVX2-FAST-NEXT: vpblendvb %ymm1, %ymm2, %ymm3, %ymm2
; AVX2-FAST-NEXT: vpbroadcastq {{.*#+}} ymm3 = [5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10]
; AVX2-FAST-NEXT: vpshufb %ymm3, %ymm5, %ymm4
; AVX2-FAST-NEXT: vpshufb %ymm3, %ymm6, %ymm5
; AVX2-FAST-NEXT: vpunpcklbw {{.*#+}} ymm4 = ymm5[0],ymm4[0],ymm5[1],ymm4[1],ymm5[2],ymm4[2],ymm5[3],ymm4[3],ymm5[4],ymm4[4],ymm5[5],ymm4[5],ymm5[6],ymm4[6],ymm5[7],ymm4[7],ymm5[16],ymm4[16],ymm5[17],ymm4[17],ymm5[18],ymm4[18],ymm5[19],ymm4[19],ymm5[20],ymm4[20],ymm5[21],ymm4[21],ymm5[22],ymm4[22],ymm5[23],ymm4[23]
; AVX2-FAST-NEXT: vmovdqa 32(%rdi), %ymm7
; AVX2-FAST-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT: vmovdqa 32(%rsi), %ymm0
; AVX2-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT: vpbroadcastq {{.*#+}} ymm5 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
; AVX2-FAST-NEXT: vpshufb %ymm5, %ymm0, %ymm6
; AVX2-FAST-NEXT: vpshufb %ymm5, %ymm7, %ymm7
; AVX2-FAST-NEXT: vpunpcklbw {{.*#+}} ymm6 = ymm7[0],ymm6[0],ymm7[1],ymm6[1],ymm7[2],ymm6[2],ymm7[3],ymm6[3],ymm7[4],ymm6[4],ymm7[5],ymm6[5],ymm7[6],ymm6[6],ymm7[7],ymm6[7],ymm7[16],ymm6[16],ymm7[17],ymm6[17],ymm7[18],ymm6[18],ymm7[19],ymm6[19],ymm7[20],ymm6[20],ymm7[21],ymm6[21],ymm7[22],ymm6[22],ymm7[23],ymm6[23]
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,2,3]
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,2,2,3]
; AVX2-FAST-NEXT: vpblendvb %ymm1, %ymm4, %ymm6, %ymm4
; AVX2-FAST-NEXT: vmovdqa (%rdx), %ymm0
; AVX2-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT: vmovdqa (%rcx), %ymm12
; AVX2-FAST-NEXT: vpshufb %ymm3, %ymm12, %ymm6
; AVX2-FAST-NEXT: vmovdqa %ymm12, %ymm14
; AVX2-FAST-NEXT: vpshufb %ymm3, %ymm0, %ymm3
; AVX2-FAST-NEXT: vpunpcklbw {{.*#+}} ymm3 = ymm3[0],ymm6[0],ymm3[1],ymm6[1],ymm3[2],ymm6[2],ymm3[3],ymm6[3],ymm3[4],ymm6[4],ymm3[5],ymm6[5],ymm3[6],ymm6[6],ymm3[7],ymm6[7],ymm3[16],ymm6[16],ymm3[17],ymm6[17],ymm3[18],ymm6[18],ymm3[19],ymm6[19],ymm3[20],ymm6[20],ymm3[21],ymm6[21],ymm3[22],ymm6[22],ymm3[23],ymm6[23]
; AVX2-FAST-NEXT: vmovdqa (%rdi), %ymm7
; AVX2-FAST-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT: vmovdqa (%rsi), %ymm12
; AVX2-FAST-NEXT: vpshufb %ymm5, %ymm12, %ymm6
; AVX2-FAST-NEXT: vpshufb %ymm5, %ymm7, %ymm5
; AVX2-FAST-NEXT: vpunpcklbw {{.*#+}} ymm5 = ymm5[0],ymm6[0],ymm5[1],ymm6[1],ymm5[2],ymm6[2],ymm5[3],ymm6[3],ymm5[4],ymm6[4],ymm5[5],ymm6[5],ymm5[6],ymm6[6],ymm5[7],ymm6[7],ymm5[16],ymm6[16],ymm5[17],ymm6[17],ymm5[18],ymm6[18],ymm5[19],ymm6[19],ymm5[20],ymm6[20],ymm5[21],ymm6[21],ymm5[22],ymm6[22],ymm5[23],ymm6[23]
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3]
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,2,2,3]
; AVX2-FAST-NEXT: vpblendvb %ymm1, %ymm3, %ymm5, %ymm1
; AVX2-FAST-NEXT: vmovdqa (%r8), %xmm15
; AVX2-FAST-NEXT: vmovdqa {{.*#+}} xmm3 = <6,u,5,u,8,u,7,u,9,u,9,u,9,u,9,u>
; AVX2-FAST-NEXT: vpshufb %xmm3, %xmm15, %xmm5
; AVX2-FAST-NEXT: vmovdqa %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,0,0,1]
; AVX2-FAST-NEXT: vmovdqa {{.*#+}} ymm6 = [255,255,0,0,255,255,255,255,0,0,255,255,255,255,0,0,255,255,255,255,0,0,255,255,255,255,0,0,255,255,255,255]
; AVX2-FAST-NEXT: vpblendvb %ymm6, %ymm13, %ymm5, %ymm0
; AVX2-FAST-NEXT: vmovdqa 32(%r8), %xmm5
; AVX2-FAST-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-NEXT: vpshufb %xmm3, %xmm5, %xmm3
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,0,1]
; AVX2-FAST-NEXT: vpblendvb %ymm6, %ymm2, %ymm3, %ymm2
; AVX2-FAST-NEXT: vmovdqa 32(%r8), %ymm5
; AVX2-FAST-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [6,0,5,0,8,0,7,0,0,0,9,0,0,0,0,0,6,0,5,0,8,0,7,0,0,0,9,0,0,0,0,0]
; AVX2-FAST-NEXT: # ymm3 = mem[0,1,0,1]
; AVX2-FAST-NEXT: vpshufb %ymm3, %ymm5, %ymm5
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,2,2,3]
; AVX2-FAST-NEXT: vpblendvb %ymm6, %ymm4, %ymm5, %ymm4
; AVX2-FAST-NEXT: vmovdqa (%r8), %ymm5
; AVX2-FAST-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT: vpshufb %ymm3, %ymm5, %ymm3
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3]
; AVX2-FAST-NEXT: vpblendvb %ymm6, %ymm1, %ymm3, %ymm1
; AVX2-FAST-NEXT: vmovdqa (%r9), %xmm3
; AVX2-FAST-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-NEXT: vmovdqa {{.*#+}} xmm6 = <u,6,u,5,u,8,u,7,u,9,u,9,u,9,u,9>
; AVX2-FAST-NEXT: vpshufb %xmm6, %xmm3, %xmm5
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,0,0,1]
; AVX2-FAST-NEXT: vmovdqa {{.*#+}} ymm3 = [255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255]
; AVX2-FAST-NEXT: vpblendvb %ymm3, %ymm0, %ymm5, %ymm0
; AVX2-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT: vmovdqa 32(%r9), %xmm5
; AVX2-FAST-NEXT: vpshufb %xmm6, %xmm5, %xmm0
; AVX2-FAST-NEXT: vmovdqa %xmm5, %xmm6
; AVX2-FAST-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,0,0,1]
; AVX2-FAST-NEXT: vpblendvb %ymm3, %ymm2, %ymm0, %ymm0
; AVX2-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT: vmovdqa 32(%r9), %ymm2
; AVX2-FAST-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [0,6,0,5,0,8,0,7,0,0,0,9,0,0,0,0,0,6,0,5,0,8,0,7,0,0,0,9,0,0,0,0]
; AVX2-FAST-NEXT: # ymm0 = mem[0,1,0,1]
; AVX2-FAST-NEXT: vpshufb %ymm0, %ymm2, %ymm2
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
; AVX2-FAST-NEXT: vpblendvb %ymm3, %ymm4, %ymm2, %ymm2
; AVX2-FAST-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT: vmovdqa (%r9), %ymm2
; AVX2-FAST-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT: vpshufb %ymm0, %ymm2, %ymm0
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3]
; AVX2-FAST-NEXT: vpblendvb %ymm3, %ymm1, %ymm0, %ymm0
; AVX2-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm11[8],xmm10[8],xmm11[9],xmm10[9],xmm11[10],xmm10[10],xmm11[11],xmm10[11],xmm11[12],xmm10[12],xmm11[13],xmm10[13],xmm11[14],xmm10[14],xmm11[15],xmm10[15]
; AVX2-FAST-NEXT: vmovdqa {{.*#+}} xmm1 = [8,9,6,7,12,13,10,11,14,15,14,15,14,15,14,15]
; AVX2-FAST-NEXT: vpshufb %xmm1, %xmm0, %xmm0
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm2 = ymm0[0,0,0,1]
; AVX2-FAST-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm8[8],xmm9[8],xmm8[9],xmm9[9],xmm8[10],xmm9[10],xmm8[11],xmm9[11],xmm8[12],xmm9[12],xmm8[13],xmm9[13],xmm8[14],xmm9[14],xmm8[15],xmm9[15]
; AVX2-FAST-NEXT: vmovdqa {{.*#+}} xmm0 = [10,11,8,9,6,7,12,13,14,15,14,15,14,15,14,15]
; AVX2-FAST-NEXT: vpshufb %xmm0, %xmm3, %xmm3
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,0,1]
; AVX2-FAST-NEXT: vmovdqa {{.*#+}} ymm5 = <u,u,255,255,0,0,u,u,255,255,0,0,u,u,255,255,0,0,u,u,255,255,0,0,u,u,255,255,0,0,u,u>
; AVX2-FAST-NEXT: vpblendvb %ymm5, %ymm2, %ymm3, %ymm8
; AVX2-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX2-FAST-NEXT: vpunpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm2 # 16-byte Folded Reload
; AVX2-FAST-NEXT: # xmm2 = xmm2[8],mem[8],xmm2[9],mem[9],xmm2[10],mem[10],xmm2[11],mem[11],xmm2[12],mem[12],xmm2[13],mem[13],xmm2[14],mem[14],xmm2[15],mem[15]
; AVX2-FAST-NEXT: vpshufb %xmm1, %xmm2, %xmm1
; AVX2-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX2-FAST-NEXT: vpunpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm2 # 16-byte Folded Reload
; AVX2-FAST-NEXT: # xmm2 = xmm2[8],mem[8],xmm2[9],mem[9],xmm2[10],mem[10],xmm2[11],mem[11],xmm2[12],mem[12],xmm2[13],mem[13],xmm2[14],mem[14],xmm2[15],mem[15]
; AVX2-FAST-NEXT: vpshufb %xmm0, %xmm2, %xmm0
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,0,1]
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,0,0,1]
; AVX2-FAST-NEXT: vpblendvb %ymm5, %ymm1, %ymm0, %ymm3
; AVX2-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-NEXT: vpunpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-FAST-NEXT: # ymm0 = ymm0[8],mem[8],ymm0[9],mem[9],ymm0[10],mem[10],ymm0[11],mem[11],ymm0[12],mem[12],ymm0[13],mem[13],ymm0[14],mem[14],ymm0[15],mem[15],ymm0[24],mem[24],ymm0[25],mem[25],ymm0[26],mem[26],ymm0[27],mem[27],ymm0[28],mem[28],ymm0[29],mem[29],ymm0[30],mem[30],ymm0[31],mem[31]
; AVX2-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [24,25,22,23,28,29,26,27,30,31,30,31,30,31,30,31,24,25,22,23,28,29,26,27,30,31,30,31,30,31,30,31]
; AVX2-FAST-NEXT: # ymm4 = mem[0,1,0,1]
; AVX2-FAST-NEXT: vpshufb %ymm4, %ymm0, %ymm0
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm1 = ymm0[2,2,2,3]
; AVX2-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-NEXT: vpunpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-FAST-NEXT: # ymm0 = ymm0[8],mem[8],ymm0[9],mem[9],ymm0[10],mem[10],ymm0[11],mem[11],ymm0[12],mem[12],ymm0[13],mem[13],ymm0[14],mem[14],ymm0[15],mem[15],ymm0[24],mem[24],ymm0[25],mem[25],ymm0[26],mem[26],ymm0[27],mem[27],ymm0[28],mem[28],ymm0[29],mem[29],ymm0[30],mem[30],ymm0[31],mem[31]
; AVX2-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [26,27,24,25,22,23,28,29,30,31,30,31,30,31,30,31,26,27,24,25,22,23,28,29,30,31,30,31,30,31,30,31]
; AVX2-FAST-NEXT: # ymm2 = mem[0,1,0,1]
; AVX2-FAST-NEXT: vpshufb %ymm2, %ymm0, %ymm0
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3]
; AVX2-FAST-NEXT: vpblendvb %ymm5, %ymm1, %ymm0, %ymm1
; AVX2-FAST-NEXT: vmovdqa %ymm12, %ymm13
; AVX2-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-NEXT: vpunpckhbw {{.*#+}} ymm0 = ymm0[8],ymm12[8],ymm0[9],ymm12[9],ymm0[10],ymm12[10],ymm0[11],ymm12[11],ymm0[12],ymm12[12],ymm0[13],ymm12[13],ymm0[14],ymm12[14],ymm0[15],ymm12[15],ymm0[24],ymm12[24],ymm0[25],ymm12[25],ymm0[26],ymm12[26],ymm0[27],ymm12[27],ymm0[28],ymm12[28],ymm0[29],ymm12[29],ymm0[30],ymm12[30],ymm0[31],ymm12[31]
; AVX2-FAST-NEXT: vpshufb %ymm4, %ymm0, %ymm0
; AVX2-FAST-NEXT: vmovdqa %ymm14, %ymm12
; AVX2-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
; AVX2-FAST-NEXT: vpunpckhbw {{.*#+}} ymm4 = ymm7[8],ymm14[8],ymm7[9],ymm14[9],ymm7[10],ymm14[10],ymm7[11],ymm14[11],ymm7[12],ymm14[12],ymm7[13],ymm14[13],ymm7[14],ymm14[14],ymm7[15],ymm14[15],ymm7[24],ymm14[24],ymm7[25],ymm14[25],ymm7[26],ymm14[26],ymm7[27],ymm14[27],ymm7[28],ymm14[28],ymm7[29],ymm14[29],ymm7[30],ymm14[30],ymm7[31],ymm14[31]
; AVX2-FAST-NEXT: vpshufb %ymm2, %ymm4, %ymm2
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3]
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
; AVX2-FAST-NEXT: vpblendvb %ymm5, %ymm0, %ymm2, %ymm0
; AVX2-FAST-NEXT: vmovdqa {{.*#+}} xmm2 = <10,u,13,u,12,u,11,u,14,u,13,u,14,u,15,u>
; AVX2-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
; AVX2-FAST-NEXT: vpshufb %xmm2, %xmm14, %xmm4
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,0,1]
; AVX2-FAST-NEXT: vmovdqa {{.*#+}} ymm5 = [0,0,255,255,255,255,0,0,255,255,255,255,0,0,255,255,255,255,0,0,255,255,255,255,0,0,255,255,255,255,0,0]
; AVX2-FAST-NEXT: vpblendvb %ymm5, %ymm8, %ymm4, %ymm4
; AVX2-FAST-NEXT: vpshufb %xmm2, %xmm15, %xmm2
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,0,0,1]
; AVX2-FAST-NEXT: vpblendvb %ymm5, %ymm3, %ymm2, %ymm2
; AVX2-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [10,0,13,0,12,0,11,0,14,0,0,0,0,0,15,0,10,0,13,0,12,0,11,0,14,0,0,0,0,0,15,0]
; AVX2-FAST-NEXT: # ymm3 = mem[0,1,0,1]
; AVX2-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload
; AVX2-FAST-NEXT: vpshufb %ymm3, %ymm9, %ymm15
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,2,2,3]
; AVX2-FAST-NEXT: vpblendvb %ymm5, %ymm1, %ymm15, %ymm15
; AVX2-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX2-FAST-NEXT: vpshufb %ymm3, %ymm8, %ymm1
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
; AVX2-FAST-NEXT: vpblendvb %ymm5, %ymm0, %ymm1, %ymm3
; AVX2-FAST-NEXT: vmovdqa {{.*#+}} xmm1 = <u,10,u,13,u,12,u,11,u,14,u,13,u,14,u,15>
; AVX2-FAST-NEXT: vpshufb %xmm1, %xmm6, %xmm0
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,0,0,1]
; AVX2-FAST-NEXT: vmovdqa {{.*#+}} ymm5 = [255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0]
; AVX2-FAST-NEXT: vpblendvb %ymm5, %ymm4, %ymm0, %ymm0
; AVX2-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-FAST-NEXT: vpshufb %xmm1, %xmm6, %xmm1
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,0,1]
; AVX2-FAST-NEXT: vpblendvb %ymm5, %ymm2, %ymm1, %ymm0
; AVX2-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [0,10,0,13,0,12,0,11,0,14,0,0,0,0,0,15,0,10,0,13,0,12,0,11,0,14,0,0,0,0,0,15]
; AVX2-FAST-NEXT: # ymm4 = mem[0,1,0,1]
; AVX2-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload
; AVX2-FAST-NEXT: vpshufb %ymm4, %ymm10, %ymm2
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
; AVX2-FAST-NEXT: vpblendvb %ymm5, %ymm15, %ymm2, %ymm0
; AVX2-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload
; AVX2-FAST-NEXT: vpshufb %ymm4, %ymm11, %ymm4
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,2,3]
; AVX2-FAST-NEXT: vpblendvb %ymm5, %ymm3, %ymm4, %ymm0
; AVX2-FAST-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-FAST-NEXT: vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm4 # 16-byte Folded Reload
; AVX2-FAST-NEXT: # xmm4 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3],xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7]
; AVX2-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-FAST-NEXT: vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm5 # 16-byte Folded Reload
; AVX2-FAST-NEXT: # xmm5 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3],xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7]
; AVX2-FAST-NEXT: vmovdqa {{.*#+}} xmm15 = [0,1,6,7,4,5,2,3,8,9,10,11,12,13,10,11]
; AVX2-FAST-NEXT: vpshufb %xmm15, %xmm4, %xmm4
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,0,1]
; AVX2-FAST-NEXT: vmovdqa {{.*#+}} xmm1 = [2,3,0,1,6,7,4,5,8,9,8,9,8,9,8,9]
; AVX2-FAST-NEXT: vpshufb %xmm1, %xmm5, %xmm5
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,0,0,1]
; AVX2-FAST-NEXT: vmovdqa {{.*#+}} ymm0 = <255,255,0,0,u,u,255,255,0,0,u,u,255,255,0,0,u,u,255,255,0,0,u,u,255,255,0,0,u,u,255,255>
; AVX2-FAST-NEXT: vpblendvb %ymm0, %ymm4, %ymm5, %ymm4
; AVX2-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX2-FAST-NEXT: vpunpcklbw (%rsp), %xmm2, %xmm5 # 16-byte Folded Reload
; AVX2-FAST-NEXT: # xmm5 = xmm2[0],mem[0],xmm2[1],mem[1],xmm2[2],mem[2],xmm2[3],mem[3],xmm2[4],mem[4],xmm2[5],mem[5],xmm2[6],mem[6],xmm2[7],mem[7]
; AVX2-FAST-NEXT: vpshufb %xmm15, %xmm5, %xmm5
; AVX2-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX2-FAST-NEXT: vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm15 # 16-byte Folded Reload
; AVX2-FAST-NEXT: # xmm15 = xmm2[0],mem[0],xmm2[1],mem[1],xmm2[2],mem[2],xmm2[3],mem[3],xmm2[4],mem[4],xmm2[5],mem[5],xmm2[6],mem[6],xmm2[7],mem[7]
; AVX2-FAST-NEXT: vpshufb %xmm1, %xmm15, %xmm1
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,0,0,1]
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,0,1]
; AVX2-FAST-NEXT: vpblendvb %ymm0, %ymm5, %ymm1, %ymm1
; AVX2-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX2-FAST-NEXT: vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm5 # 32-byte Folded Reload
; AVX2-FAST-NEXT: # ymm5 = ymm2[0],mem[0],ymm2[1],mem[1],ymm2[2],mem[2],ymm2[3],mem[3],ymm2[4],mem[4],ymm2[5],mem[5],ymm2[6],mem[6],ymm2[7],mem[7],ymm2[16],mem[16],ymm2[17],mem[17],ymm2[18],mem[18],ymm2[19],mem[19],ymm2[20],mem[20],ymm2[21],mem[21],ymm2[22],mem[22],ymm2[23],mem[23]
; AVX2-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX2-FAST-NEXT: vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm15 # 32-byte Folded Reload
; AVX2-FAST-NEXT: # ymm15 = ymm2[0],mem[0],ymm2[1],mem[1],ymm2[2],mem[2],ymm2[3],mem[3],ymm2[4],mem[4],ymm2[5],mem[5],ymm2[6],mem[6],ymm2[7],mem[7],ymm2[16],mem[16],ymm2[17],mem[17],ymm2[18],mem[18],ymm2[19],mem[19],ymm2[20],mem[20],ymm2[21],mem[21],ymm2[22],mem[22],ymm2[23],mem[23]
; AVX2-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [16,17,22,23,20,21,18,19,24,25,26,27,28,29,26,27,16,17,22,23,20,21,18,19,24,25,26,27,28,29,26,27]
; AVX2-FAST-NEXT: # ymm2 = mem[0,1,0,1]
; AVX2-FAST-NEXT: vpshufb %ymm2, %ymm5, %ymm5
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,2,2,3]
; AVX2-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [18,19,16,17,22,23,20,21,24,25,24,25,24,25,24,25,18,19,16,17,22,23,20,21,24,25,24,25,24,25,24,25]
; AVX2-FAST-NEXT: # ymm3 = mem[0,1,0,1]
; AVX2-FAST-NEXT: vpshufb %ymm3, %ymm15, %ymm15
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,2,2,3]
; AVX2-FAST-NEXT: vpblendvb %ymm0, %ymm5, %ymm15, %ymm5
; AVX2-FAST-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; AVX2-FAST-NEXT: vpunpcklbw {{.*#+}} ymm15 = ymm15[0],ymm13[0],ymm15[1],ymm13[1],ymm15[2],ymm13[2],ymm15[3],ymm13[3],ymm15[4],ymm13[4],ymm15[5],ymm13[5],ymm15[6],ymm13[6],ymm15[7],ymm13[7],ymm15[16],ymm13[16],ymm15[17],ymm13[17],ymm15[18],ymm13[18],ymm15[19],ymm13[19],ymm15[20],ymm13[20],ymm15[21],ymm13[21],ymm15[22],ymm13[22],ymm15[23],ymm13[23]
; AVX2-FAST-NEXT: vpshufb %ymm2, %ymm15, %ymm2
; AVX2-FAST-NEXT: vpunpcklbw {{.*#+}} ymm15 = ymm7[0],ymm12[0],ymm7[1],ymm12[1],ymm7[2],ymm12[2],ymm7[3],ymm12[3],ymm7[4],ymm12[4],ymm7[5],ymm12[5],ymm7[6],ymm12[6],ymm7[7],ymm12[7],ymm7[16],ymm12[16],ymm7[17],ymm12[17],ymm7[18],ymm12[18],ymm7[19],ymm12[19],ymm7[20],ymm12[20],ymm7[21],ymm12[21],ymm7[22],ymm12[22],ymm7[23],ymm12[23]
; AVX2-FAST-NEXT: vpshufb %ymm3, %ymm15, %ymm3
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3]
; AVX2-FAST-NEXT: vpblendvb %ymm0, %ymm2, %ymm3, %ymm0
; AVX2-FAST-NEXT: vmovdqa {{.*#+}} xmm2 = <2,u,1,u,0,u,3,u,4,u,4,u,4,u,4,u>
; AVX2-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-FAST-NEXT: vpshufb %xmm2, %xmm3, %xmm3
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,0,1]
; AVX2-FAST-NEXT: vmovdqa {{.*#+}} ymm15 = [255,255,255,255,0,0,255,255,255,255,0,0,255,255,255,255,0,0,255,255,255,255,0,0,255,255,255,255,0,0,255,255]
; AVX2-FAST-NEXT: vpblendvb %ymm15, %ymm4, %ymm3, %ymm3
; AVX2-FAST-NEXT: vpshufb %xmm2, %xmm14, %xmm2
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,0,0,1]
; AVX2-FAST-NEXT: vpblendvb %ymm15, %ymm1, %ymm2, %ymm1
; AVX2-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [2,0,1,0,0,0,3,0,0,0,0,0,4,0,0,0,2,0,1,0,0,0,3,0,0,0,0,0,4,0,0,0]
; AVX2-FAST-NEXT: # ymm2 = mem[0,1,0,1]
; AVX2-FAST-NEXT: vpshufb %ymm2, %ymm9, %ymm4
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,2,3]
; AVX2-FAST-NEXT: vpblendvb %ymm15, %ymm5, %ymm4, %ymm4
; AVX2-FAST-NEXT: vpshufb %ymm2, %ymm8, %ymm2
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
; AVX2-FAST-NEXT: vpblendvb %ymm15, %ymm0, %ymm2, %ymm0
; AVX2-FAST-NEXT: vmovdqa {{.*#+}} xmm2 = <u,2,u,1,u,0,u,3,u,4,u,4,u,4,u,4>
; AVX2-FAST-NEXT: vpshufb %xmm2, %xmm6, %xmm5
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,0,0,1]
; AVX2-FAST-NEXT: vmovdqa {{.*#+}} ymm13 = [255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255]
; AVX2-FAST-NEXT: vpblendvb %ymm13, %ymm3, %ymm5, %ymm3
; AVX2-FAST-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-FAST-NEXT: vpshufb %xmm2, %xmm5, %xmm2
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,0,0,1]
; AVX2-FAST-NEXT: vpblendvb %ymm13, %ymm1, %ymm2, %ymm1
; AVX2-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [0,2,0,1,0,0,0,3,0,0,0,0,0,4,0,0,0,2,0,1,0,0,0,3,0,0,0,0,0,4,0,0]
; AVX2-FAST-NEXT: # ymm2 = mem[0,1,0,1]
; AVX2-FAST-NEXT: vpshufb %ymm2, %ymm10, %ymm5
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,2,2,3]
; AVX2-FAST-NEXT: vpblendvb %ymm13, %ymm4, %ymm5, %ymm4
; AVX2-FAST-NEXT: vpshufb %ymm2, %ymm11, %ymm2
; AVX2-FAST-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
; AVX2-FAST-NEXT: vpblendvb %ymm13, %ymm0, %ymm2, %ymm0
; AVX2-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX2-FAST-NEXT: vmovdqa %ymm0, 96(%rax)
; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-NEXT: vmovaps %ymm0, 160(%rax)
; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-NEXT: vmovaps %ymm0, 128(%rax)
; AVX2-FAST-NEXT: vmovdqa %ymm4, 288(%rax)
; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-NEXT: vmovaps %ymm0, 352(%rax)
; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-NEXT: vmovaps %ymm0, 320(%rax)
; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-NEXT: vmovaps %ymm0, 64(%rax)
; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-NEXT: vmovaps %ymm0, 224(%rax)
; AVX2-FAST-NEXT: vmovdqa %ymm1, 192(%rax)
; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-NEXT: vmovaps %ymm0, 256(%rax)
; AVX2-FAST-NEXT: vmovdqa %ymm3, (%rax)
; AVX2-FAST-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-NEXT: vmovaps %ymm0, 32(%rax)
; AVX2-FAST-NEXT: addq $648, %rsp # imm = 0x288
; AVX2-FAST-NEXT: vzeroupper
; AVX2-FAST-NEXT: retq
;
; AVX2-FAST-PERLANE-LABEL: store_i8_stride6_vf64:
; AVX2-FAST-PERLANE: # %bb.0:
; AVX2-FAST-PERLANE-NEXT: subq $648, %rsp # imm = 0x288
; AVX2-FAST-PERLANE-NEXT: vmovdqa 32(%rdx), %ymm6
; AVX2-FAST-PERLANE-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT: vmovdqa 32(%rcx), %ymm5
; AVX2-FAST-PERLANE-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT: vmovdqa (%rcx), %xmm0
; AVX2-FAST-PERLANE-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-PERLANE-NEXT: vmovdqa 32(%rcx), %xmm9
; AVX2-FAST-PERLANE-NEXT: vpbroadcastq {{.*#+}} xmm2 = [5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10]
; AVX2-FAST-PERLANE-NEXT: vpshufb %xmm2, %xmm0, %xmm0
; AVX2-FAST-PERLANE-NEXT: vmovdqa (%rdx), %xmm1
; AVX2-FAST-PERLANE-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-PERLANE-NEXT: vmovdqa 32(%rdx), %xmm8
; AVX2-FAST-PERLANE-NEXT: vpshufb %xmm2, %xmm1, %xmm1
; AVX2-FAST-PERLANE-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3],xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,0,0,1]
; AVX2-FAST-PERLANE-NEXT: vmovdqa (%rsi), %xmm1
; AVX2-FAST-PERLANE-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-PERLANE-NEXT: vmovdqa 32(%rsi), %xmm10
; AVX2-FAST-PERLANE-NEXT: vpbroadcastq {{.*#+}} xmm3 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
; AVX2-FAST-PERLANE-NEXT: vpshufb %xmm3, %xmm1, %xmm1
; AVX2-FAST-PERLANE-NEXT: vmovdqa (%rdi), %xmm4
; AVX2-FAST-PERLANE-NEXT: vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-PERLANE-NEXT: vmovdqa 32(%rdi), %xmm11
; AVX2-FAST-PERLANE-NEXT: vpshufb %xmm3, %xmm4, %xmm4
; AVX2-FAST-PERLANE-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm4[8],xmm1[8],xmm4[9],xmm1[9],xmm4[10],xmm1[10],xmm4[11],xmm1[11],xmm4[12],xmm1[12],xmm4[13],xmm1[13],xmm4[14],xmm1[14],xmm4[15],xmm1[15]
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm4 = ymm1[0,0,0,1]
; AVX2-FAST-PERLANE-NEXT: vmovdqa {{.*#+}} ymm1 = <255,255,u,u,0,0,255,255,u,u,0,0,255,255,u,u,0,0,255,255,u,u,0,0,255,255,u,u,0,0,255,255>
; AVX2-FAST-PERLANE-NEXT: vpblendvb %ymm1, %ymm0, %ymm4, %ymm13
; AVX2-FAST-PERLANE-NEXT: vpshufb %xmm2, %xmm9, %xmm4
; AVX2-FAST-PERLANE-NEXT: vmovdqa %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-PERLANE-NEXT: vpshufb %xmm2, %xmm8, %xmm2
; AVX2-FAST-PERLANE-NEXT: vmovdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-PERLANE-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3],xmm2[4],xmm4[4],xmm2[5],xmm4[5],xmm2[6],xmm4[6],xmm2[7],xmm4[7]
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,0,0,1]
; AVX2-FAST-PERLANE-NEXT: vpshufb %xmm3, %xmm10, %xmm4
; AVX2-FAST-PERLANE-NEXT: vmovdqa %xmm10, (%rsp) # 16-byte Spill
; AVX2-FAST-PERLANE-NEXT: vpshufb %xmm3, %xmm11, %xmm3
; AVX2-FAST-PERLANE-NEXT: vmovdqa %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-PERLANE-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm3[8],xmm4[8],xmm3[9],xmm4[9],xmm3[10],xmm4[10],xmm3[11],xmm4[11],xmm3[12],xmm4[12],xmm3[13],xmm4[13],xmm3[14],xmm4[14],xmm3[15],xmm4[15]
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,0,1]
; AVX2-FAST-PERLANE-NEXT: vpblendvb %ymm1, %ymm2, %ymm3, %ymm2
; AVX2-FAST-PERLANE-NEXT: vpbroadcastq {{.*#+}} ymm3 = [5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10]
; AVX2-FAST-PERLANE-NEXT: vpshufb %ymm3, %ymm5, %ymm4
; AVX2-FAST-PERLANE-NEXT: vpshufb %ymm3, %ymm6, %ymm5
; AVX2-FAST-PERLANE-NEXT: vpunpcklbw {{.*#+}} ymm4 = ymm5[0],ymm4[0],ymm5[1],ymm4[1],ymm5[2],ymm4[2],ymm5[3],ymm4[3],ymm5[4],ymm4[4],ymm5[5],ymm4[5],ymm5[6],ymm4[6],ymm5[7],ymm4[7],ymm5[16],ymm4[16],ymm5[17],ymm4[17],ymm5[18],ymm4[18],ymm5[19],ymm4[19],ymm5[20],ymm4[20],ymm5[21],ymm4[21],ymm5[22],ymm4[22],ymm5[23],ymm4[23]
; AVX2-FAST-PERLANE-NEXT: vmovdqa 32(%rdi), %ymm7
; AVX2-FAST-PERLANE-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT: vmovdqa 32(%rsi), %ymm0
; AVX2-FAST-PERLANE-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT: vpbroadcastq {{.*#+}} ymm5 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
; AVX2-FAST-PERLANE-NEXT: vpshufb %ymm5, %ymm0, %ymm6
; AVX2-FAST-PERLANE-NEXT: vpshufb %ymm5, %ymm7, %ymm7
; AVX2-FAST-PERLANE-NEXT: vpunpcklbw {{.*#+}} ymm6 = ymm7[0],ymm6[0],ymm7[1],ymm6[1],ymm7[2],ymm6[2],ymm7[3],ymm6[3],ymm7[4],ymm6[4],ymm7[5],ymm6[5],ymm7[6],ymm6[6],ymm7[7],ymm6[7],ymm7[16],ymm6[16],ymm7[17],ymm6[17],ymm7[18],ymm6[18],ymm7[19],ymm6[19],ymm7[20],ymm6[20],ymm7[21],ymm6[21],ymm7[22],ymm6[22],ymm7[23],ymm6[23]
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,2,3]
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,2,2,3]
; AVX2-FAST-PERLANE-NEXT: vpblendvb %ymm1, %ymm4, %ymm6, %ymm4
; AVX2-FAST-PERLANE-NEXT: vmovdqa (%rdx), %ymm0
; AVX2-FAST-PERLANE-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT: vmovdqa (%rcx), %ymm12
; AVX2-FAST-PERLANE-NEXT: vpshufb %ymm3, %ymm12, %ymm6
; AVX2-FAST-PERLANE-NEXT: vmovdqa %ymm12, %ymm14
; AVX2-FAST-PERLANE-NEXT: vpshufb %ymm3, %ymm0, %ymm3
; AVX2-FAST-PERLANE-NEXT: vpunpcklbw {{.*#+}} ymm3 = ymm3[0],ymm6[0],ymm3[1],ymm6[1],ymm3[2],ymm6[2],ymm3[3],ymm6[3],ymm3[4],ymm6[4],ymm3[5],ymm6[5],ymm3[6],ymm6[6],ymm3[7],ymm6[7],ymm3[16],ymm6[16],ymm3[17],ymm6[17],ymm3[18],ymm6[18],ymm3[19],ymm6[19],ymm3[20],ymm6[20],ymm3[21],ymm6[21],ymm3[22],ymm6[22],ymm3[23],ymm6[23]
; AVX2-FAST-PERLANE-NEXT: vmovdqa (%rdi), %ymm7
; AVX2-FAST-PERLANE-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT: vmovdqa (%rsi), %ymm12
; AVX2-FAST-PERLANE-NEXT: vpshufb %ymm5, %ymm12, %ymm6
; AVX2-FAST-PERLANE-NEXT: vpshufb %ymm5, %ymm7, %ymm5
; AVX2-FAST-PERLANE-NEXT: vpunpcklbw {{.*#+}} ymm5 = ymm5[0],ymm6[0],ymm5[1],ymm6[1],ymm5[2],ymm6[2],ymm5[3],ymm6[3],ymm5[4],ymm6[4],ymm5[5],ymm6[5],ymm5[6],ymm6[6],ymm5[7],ymm6[7],ymm5[16],ymm6[16],ymm5[17],ymm6[17],ymm5[18],ymm6[18],ymm5[19],ymm6[19],ymm5[20],ymm6[20],ymm5[21],ymm6[21],ymm5[22],ymm6[22],ymm5[23],ymm6[23]
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3]
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,2,2,3]
; AVX2-FAST-PERLANE-NEXT: vpblendvb %ymm1, %ymm3, %ymm5, %ymm1
; AVX2-FAST-PERLANE-NEXT: vmovdqa (%r8), %xmm15
; AVX2-FAST-PERLANE-NEXT: vmovdqa {{.*#+}} xmm3 = <6,u,5,u,8,u,7,u,9,u,9,u,9,u,9,u>
; AVX2-FAST-PERLANE-NEXT: vpshufb %xmm3, %xmm15, %xmm5
; AVX2-FAST-PERLANE-NEXT: vmovdqa %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,0,0,1]
; AVX2-FAST-PERLANE-NEXT: vmovdqa {{.*#+}} ymm6 = [255,255,0,0,255,255,255,255,0,0,255,255,255,255,0,0,255,255,255,255,0,0,255,255,255,255,0,0,255,255,255,255]
; AVX2-FAST-PERLANE-NEXT: vpblendvb %ymm6, %ymm13, %ymm5, %ymm0
; AVX2-FAST-PERLANE-NEXT: vmovdqa 32(%r8), %xmm5
; AVX2-FAST-PERLANE-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-PERLANE-NEXT: vpshufb %xmm3, %xmm5, %xmm3
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,0,1]
; AVX2-FAST-PERLANE-NEXT: vpblendvb %ymm6, %ymm2, %ymm3, %ymm2
; AVX2-FAST-PERLANE-NEXT: vmovdqa 32(%r8), %ymm5
; AVX2-FAST-PERLANE-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [6,0,5,0,8,0,7,0,0,0,9,0,0,0,0,0,6,0,5,0,8,0,7,0,0,0,9,0,0,0,0,0]
; AVX2-FAST-PERLANE-NEXT: # ymm3 = mem[0,1,0,1]
; AVX2-FAST-PERLANE-NEXT: vpshufb %ymm3, %ymm5, %ymm5
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,2,2,3]
; AVX2-FAST-PERLANE-NEXT: vpblendvb %ymm6, %ymm4, %ymm5, %ymm4
; AVX2-FAST-PERLANE-NEXT: vmovdqa (%r8), %ymm5
; AVX2-FAST-PERLANE-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT: vpshufb %ymm3, %ymm5, %ymm3
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3]
; AVX2-FAST-PERLANE-NEXT: vpblendvb %ymm6, %ymm1, %ymm3, %ymm1
; AVX2-FAST-PERLANE-NEXT: vmovdqa (%r9), %xmm3
; AVX2-FAST-PERLANE-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-PERLANE-NEXT: vmovdqa {{.*#+}} xmm6 = <u,6,u,5,u,8,u,7,u,9,u,9,u,9,u,9>
; AVX2-FAST-PERLANE-NEXT: vpshufb %xmm6, %xmm3, %xmm5
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,0,0,1]
; AVX2-FAST-PERLANE-NEXT: vmovdqa {{.*#+}} ymm3 = [255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255]
; AVX2-FAST-PERLANE-NEXT: vpblendvb %ymm3, %ymm0, %ymm5, %ymm0
; AVX2-FAST-PERLANE-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT: vmovdqa 32(%r9), %xmm5
; AVX2-FAST-PERLANE-NEXT: vpshufb %xmm6, %xmm5, %xmm0
; AVX2-FAST-PERLANE-NEXT: vmovdqa %xmm5, %xmm6
; AVX2-FAST-PERLANE-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,0,0,1]
; AVX2-FAST-PERLANE-NEXT: vpblendvb %ymm3, %ymm2, %ymm0, %ymm0
; AVX2-FAST-PERLANE-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT: vmovdqa 32(%r9), %ymm2
; AVX2-FAST-PERLANE-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [0,6,0,5,0,8,0,7,0,0,0,9,0,0,0,0,0,6,0,5,0,8,0,7,0,0,0,9,0,0,0,0]
; AVX2-FAST-PERLANE-NEXT: # ymm0 = mem[0,1,0,1]
; AVX2-FAST-PERLANE-NEXT: vpshufb %ymm0, %ymm2, %ymm2
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
; AVX2-FAST-PERLANE-NEXT: vpblendvb %ymm3, %ymm4, %ymm2, %ymm2
; AVX2-FAST-PERLANE-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT: vmovdqa (%r9), %ymm2
; AVX2-FAST-PERLANE-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT: vpshufb %ymm0, %ymm2, %ymm0
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3]
; AVX2-FAST-PERLANE-NEXT: vpblendvb %ymm3, %ymm1, %ymm0, %ymm0
; AVX2-FAST-PERLANE-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm11[8],xmm10[8],xmm11[9],xmm10[9],xmm11[10],xmm10[10],xmm11[11],xmm10[11],xmm11[12],xmm10[12],xmm11[13],xmm10[13],xmm11[14],xmm10[14],xmm11[15],xmm10[15]
; AVX2-FAST-PERLANE-NEXT: vmovdqa {{.*#+}} xmm1 = [8,9,6,7,12,13,10,11,14,15,14,15,14,15,14,15]
; AVX2-FAST-PERLANE-NEXT: vpshufb %xmm1, %xmm0, %xmm0
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm2 = ymm0[0,0,0,1]
; AVX2-FAST-PERLANE-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm8[8],xmm9[8],xmm8[9],xmm9[9],xmm8[10],xmm9[10],xmm8[11],xmm9[11],xmm8[12],xmm9[12],xmm8[13],xmm9[13],xmm8[14],xmm9[14],xmm8[15],xmm9[15]
; AVX2-FAST-PERLANE-NEXT: vmovdqa {{.*#+}} xmm0 = [10,11,8,9,6,7,12,13,14,15,14,15,14,15,14,15]
; AVX2-FAST-PERLANE-NEXT: vpshufb %xmm0, %xmm3, %xmm3
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,0,1]
; AVX2-FAST-PERLANE-NEXT: vmovdqa {{.*#+}} ymm5 = <u,u,255,255,0,0,u,u,255,255,0,0,u,u,255,255,0,0,u,u,255,255,0,0,u,u,255,255,0,0,u,u>
; AVX2-FAST-PERLANE-NEXT: vpblendvb %ymm5, %ymm2, %ymm3, %ymm8
; AVX2-FAST-PERLANE-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX2-FAST-PERLANE-NEXT: vpunpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm2 # 16-byte Folded Reload
; AVX2-FAST-PERLANE-NEXT: # xmm2 = xmm2[8],mem[8],xmm2[9],mem[9],xmm2[10],mem[10],xmm2[11],mem[11],xmm2[12],mem[12],xmm2[13],mem[13],xmm2[14],mem[14],xmm2[15],mem[15]
; AVX2-FAST-PERLANE-NEXT: vpshufb %xmm1, %xmm2, %xmm1
; AVX2-FAST-PERLANE-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX2-FAST-PERLANE-NEXT: vpunpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm2 # 16-byte Folded Reload
; AVX2-FAST-PERLANE-NEXT: # xmm2 = xmm2[8],mem[8],xmm2[9],mem[9],xmm2[10],mem[10],xmm2[11],mem[11],xmm2[12],mem[12],xmm2[13],mem[13],xmm2[14],mem[14],xmm2[15],mem[15]
; AVX2-FAST-PERLANE-NEXT: vpshufb %xmm0, %xmm2, %xmm0
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,0,1]
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,0,0,1]
; AVX2-FAST-PERLANE-NEXT: vpblendvb %ymm5, %ymm1, %ymm0, %ymm3
; AVX2-FAST-PERLANE-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT: vpunpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-FAST-PERLANE-NEXT: # ymm0 = ymm0[8],mem[8],ymm0[9],mem[9],ymm0[10],mem[10],ymm0[11],mem[11],ymm0[12],mem[12],ymm0[13],mem[13],ymm0[14],mem[14],ymm0[15],mem[15],ymm0[24],mem[24],ymm0[25],mem[25],ymm0[26],mem[26],ymm0[27],mem[27],ymm0[28],mem[28],ymm0[29],mem[29],ymm0[30],mem[30],ymm0[31],mem[31]
; AVX2-FAST-PERLANE-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [24,25,22,23,28,29,26,27,30,31,30,31,30,31,30,31,24,25,22,23,28,29,26,27,30,31,30,31,30,31,30,31]
; AVX2-FAST-PERLANE-NEXT: # ymm4 = mem[0,1,0,1]
; AVX2-FAST-PERLANE-NEXT: vpshufb %ymm4, %ymm0, %ymm0
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm1 = ymm0[2,2,2,3]
; AVX2-FAST-PERLANE-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT: vpunpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload
; AVX2-FAST-PERLANE-NEXT: # ymm0 = ymm0[8],mem[8],ymm0[9],mem[9],ymm0[10],mem[10],ymm0[11],mem[11],ymm0[12],mem[12],ymm0[13],mem[13],ymm0[14],mem[14],ymm0[15],mem[15],ymm0[24],mem[24],ymm0[25],mem[25],ymm0[26],mem[26],ymm0[27],mem[27],ymm0[28],mem[28],ymm0[29],mem[29],ymm0[30],mem[30],ymm0[31],mem[31]
; AVX2-FAST-PERLANE-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [26,27,24,25,22,23,28,29,30,31,30,31,30,31,30,31,26,27,24,25,22,23,28,29,30,31,30,31,30,31,30,31]
; AVX2-FAST-PERLANE-NEXT: # ymm2 = mem[0,1,0,1]
; AVX2-FAST-PERLANE-NEXT: vpshufb %ymm2, %ymm0, %ymm0
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3]
; AVX2-FAST-PERLANE-NEXT: vpblendvb %ymm5, %ymm1, %ymm0, %ymm1
; AVX2-FAST-PERLANE-NEXT: vmovdqa %ymm12, %ymm13
; AVX2-FAST-PERLANE-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT: vpunpckhbw {{.*#+}} ymm0 = ymm0[8],ymm12[8],ymm0[9],ymm12[9],ymm0[10],ymm12[10],ymm0[11],ymm12[11],ymm0[12],ymm12[12],ymm0[13],ymm12[13],ymm0[14],ymm12[14],ymm0[15],ymm12[15],ymm0[24],ymm12[24],ymm0[25],ymm12[25],ymm0[26],ymm12[26],ymm0[27],ymm12[27],ymm0[28],ymm12[28],ymm0[29],ymm12[29],ymm0[30],ymm12[30],ymm0[31],ymm12[31]
; AVX2-FAST-PERLANE-NEXT: vpshufb %ymm4, %ymm0, %ymm0
; AVX2-FAST-PERLANE-NEXT: vmovdqa %ymm14, %ymm12
; AVX2-FAST-PERLANE-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT: vpunpckhbw {{.*#+}} ymm4 = ymm7[8],ymm14[8],ymm7[9],ymm14[9],ymm7[10],ymm14[10],ymm7[11],ymm14[11],ymm7[12],ymm14[12],ymm7[13],ymm14[13],ymm7[14],ymm14[14],ymm7[15],ymm14[15],ymm7[24],ymm14[24],ymm7[25],ymm14[25],ymm7[26],ymm14[26],ymm7[27],ymm14[27],ymm7[28],ymm14[28],ymm7[29],ymm14[29],ymm7[30],ymm14[30],ymm7[31],ymm14[31]
; AVX2-FAST-PERLANE-NEXT: vpshufb %ymm2, %ymm4, %ymm2
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3]
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
; AVX2-FAST-PERLANE-NEXT: vpblendvb %ymm5, %ymm0, %ymm2, %ymm0
; AVX2-FAST-PERLANE-NEXT: vmovdqa {{.*#+}} xmm2 = <10,u,13,u,12,u,11,u,14,u,13,u,14,u,15,u>
; AVX2-FAST-PERLANE-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload
; AVX2-FAST-PERLANE-NEXT: vpshufb %xmm2, %xmm14, %xmm4
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,0,1]
; AVX2-FAST-PERLANE-NEXT: vmovdqa {{.*#+}} ymm5 = [0,0,255,255,255,255,0,0,255,255,255,255,0,0,255,255,255,255,0,0,255,255,255,255,0,0,255,255,255,255,0,0]
; AVX2-FAST-PERLANE-NEXT: vpblendvb %ymm5, %ymm8, %ymm4, %ymm4
; AVX2-FAST-PERLANE-NEXT: vpshufb %xmm2, %xmm15, %xmm2
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,0,0,1]
; AVX2-FAST-PERLANE-NEXT: vpblendvb %ymm5, %ymm3, %ymm2, %ymm2
; AVX2-FAST-PERLANE-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [10,0,13,0,12,0,11,0,14,0,0,0,0,0,15,0,10,0,13,0,12,0,11,0,14,0,0,0,0,0,15,0]
; AVX2-FAST-PERLANE-NEXT: # ymm3 = mem[0,1,0,1]
; AVX2-FAST-PERLANE-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT: vpshufb %ymm3, %ymm9, %ymm15
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,2,2,3]
; AVX2-FAST-PERLANE-NEXT: vpblendvb %ymm5, %ymm1, %ymm15, %ymm15
; AVX2-FAST-PERLANE-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT: vpshufb %ymm3, %ymm8, %ymm1
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
; AVX2-FAST-PERLANE-NEXT: vpblendvb %ymm5, %ymm0, %ymm1, %ymm3
; AVX2-FAST-PERLANE-NEXT: vmovdqa {{.*#+}} xmm1 = <u,10,u,13,u,12,u,11,u,14,u,13,u,14,u,15>
; AVX2-FAST-PERLANE-NEXT: vpshufb %xmm1, %xmm6, %xmm0
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,0,0,1]
; AVX2-FAST-PERLANE-NEXT: vmovdqa {{.*#+}} ymm5 = [255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0]
; AVX2-FAST-PERLANE-NEXT: vpblendvb %ymm5, %ymm4, %ymm0, %ymm0
; AVX2-FAST-PERLANE-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload
; AVX2-FAST-PERLANE-NEXT: vpshufb %xmm1, %xmm6, %xmm1
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,0,1]
; AVX2-FAST-PERLANE-NEXT: vpblendvb %ymm5, %ymm2, %ymm1, %ymm0
; AVX2-FAST-PERLANE-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [0,10,0,13,0,12,0,11,0,14,0,0,0,0,0,15,0,10,0,13,0,12,0,11,0,14,0,0,0,0,0,15]
; AVX2-FAST-PERLANE-NEXT: # ymm4 = mem[0,1,0,1]
; AVX2-FAST-PERLANE-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT: vpshufb %ymm4, %ymm10, %ymm2
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
; AVX2-FAST-PERLANE-NEXT: vpblendvb %ymm5, %ymm15, %ymm2, %ymm0
; AVX2-FAST-PERLANE-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT: vpshufb %ymm4, %ymm11, %ymm4
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,2,3]
; AVX2-FAST-PERLANE-NEXT: vpblendvb %ymm5, %ymm3, %ymm4, %ymm0
; AVX2-FAST-PERLANE-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
; AVX2-FAST-PERLANE-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-FAST-PERLANE-NEXT: vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm4 # 16-byte Folded Reload
; AVX2-FAST-PERLANE-NEXT: # xmm4 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3],xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7]
; AVX2-FAST-PERLANE-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
; AVX2-FAST-PERLANE-NEXT: vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm5 # 16-byte Folded Reload
; AVX2-FAST-PERLANE-NEXT: # xmm5 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3],xmm0[4],mem[4],xmm0[5],mem[5],xmm0[6],mem[6],xmm0[7],mem[7]
; AVX2-FAST-PERLANE-NEXT: vmovdqa {{.*#+}} xmm15 = [0,1,6,7,4,5,2,3,8,9,10,11,12,13,10,11]
; AVX2-FAST-PERLANE-NEXT: vpshufb %xmm15, %xmm4, %xmm4
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,0,1]
; AVX2-FAST-PERLANE-NEXT: vmovdqa {{.*#+}} xmm1 = [2,3,0,1,6,7,4,5,8,9,8,9,8,9,8,9]
; AVX2-FAST-PERLANE-NEXT: vpshufb %xmm1, %xmm5, %xmm5
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,0,0,1]
; AVX2-FAST-PERLANE-NEXT: vmovdqa {{.*#+}} ymm0 = <255,255,0,0,u,u,255,255,0,0,u,u,255,255,0,0,u,u,255,255,0,0,u,u,255,255,0,0,u,u,255,255>
; AVX2-FAST-PERLANE-NEXT: vpblendvb %ymm0, %ymm4, %ymm5, %ymm4
; AVX2-FAST-PERLANE-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX2-FAST-PERLANE-NEXT: vpunpcklbw (%rsp), %xmm2, %xmm5 # 16-byte Folded Reload
; AVX2-FAST-PERLANE-NEXT: # xmm5 = xmm2[0],mem[0],xmm2[1],mem[1],xmm2[2],mem[2],xmm2[3],mem[3],xmm2[4],mem[4],xmm2[5],mem[5],xmm2[6],mem[6],xmm2[7],mem[7]
; AVX2-FAST-PERLANE-NEXT: vpshufb %xmm15, %xmm5, %xmm5
; AVX2-FAST-PERLANE-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
; AVX2-FAST-PERLANE-NEXT: vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm15 # 16-byte Folded Reload
; AVX2-FAST-PERLANE-NEXT: # xmm15 = xmm2[0],mem[0],xmm2[1],mem[1],xmm2[2],mem[2],xmm2[3],mem[3],xmm2[4],mem[4],xmm2[5],mem[5],xmm2[6],mem[6],xmm2[7],mem[7]
; AVX2-FAST-PERLANE-NEXT: vpshufb %xmm1, %xmm15, %xmm1
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,0,0,1]
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,0,1]
; AVX2-FAST-PERLANE-NEXT: vpblendvb %ymm0, %ymm5, %ymm1, %ymm1
; AVX2-FAST-PERLANE-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT: vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm5 # 32-byte Folded Reload
; AVX2-FAST-PERLANE-NEXT: # ymm5 = ymm2[0],mem[0],ymm2[1],mem[1],ymm2[2],mem[2],ymm2[3],mem[3],ymm2[4],mem[4],ymm2[5],mem[5],ymm2[6],mem[6],ymm2[7],mem[7],ymm2[16],mem[16],ymm2[17],mem[17],ymm2[18],mem[18],ymm2[19],mem[19],ymm2[20],mem[20],ymm2[21],mem[21],ymm2[22],mem[22],ymm2[23],mem[23]
; AVX2-FAST-PERLANE-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT: vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm15 # 32-byte Folded Reload
; AVX2-FAST-PERLANE-NEXT: # ymm15 = ymm2[0],mem[0],ymm2[1],mem[1],ymm2[2],mem[2],ymm2[3],mem[3],ymm2[4],mem[4],ymm2[5],mem[5],ymm2[6],mem[6],ymm2[7],mem[7],ymm2[16],mem[16],ymm2[17],mem[17],ymm2[18],mem[18],ymm2[19],mem[19],ymm2[20],mem[20],ymm2[21],mem[21],ymm2[22],mem[22],ymm2[23],mem[23]
; AVX2-FAST-PERLANE-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [16,17,22,23,20,21,18,19,24,25,26,27,28,29,26,27,16,17,22,23,20,21,18,19,24,25,26,27,28,29,26,27]
; AVX2-FAST-PERLANE-NEXT: # ymm2 = mem[0,1,0,1]
; AVX2-FAST-PERLANE-NEXT: vpshufb %ymm2, %ymm5, %ymm5
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,2,2,3]
; AVX2-FAST-PERLANE-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [18,19,16,17,22,23,20,21,24,25,24,25,24,25,24,25,18,19,16,17,22,23,20,21,24,25,24,25,24,25,24,25]
; AVX2-FAST-PERLANE-NEXT: # ymm3 = mem[0,1,0,1]
; AVX2-FAST-PERLANE-NEXT: vpshufb %ymm3, %ymm15, %ymm15
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,2,2,3]
; AVX2-FAST-PERLANE-NEXT: vpblendvb %ymm0, %ymm5, %ymm15, %ymm5
; AVX2-FAST-PERLANE-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT: vpunpcklbw {{.*#+}} ymm15 = ymm15[0],ymm13[0],ymm15[1],ymm13[1],ymm15[2],ymm13[2],ymm15[3],ymm13[3],ymm15[4],ymm13[4],ymm15[5],ymm13[5],ymm15[6],ymm13[6],ymm15[7],ymm13[7],ymm15[16],ymm13[16],ymm15[17],ymm13[17],ymm15[18],ymm13[18],ymm15[19],ymm13[19],ymm15[20],ymm13[20],ymm15[21],ymm13[21],ymm15[22],ymm13[22],ymm15[23],ymm13[23]
; AVX2-FAST-PERLANE-NEXT: vpshufb %ymm2, %ymm15, %ymm2
; AVX2-FAST-PERLANE-NEXT: vpunpcklbw {{.*#+}} ymm15 = ymm7[0],ymm12[0],ymm7[1],ymm12[1],ymm7[2],ymm12[2],ymm7[3],ymm12[3],ymm7[4],ymm12[4],ymm7[5],ymm12[5],ymm7[6],ymm12[6],ymm7[7],ymm12[7],ymm7[16],ymm12[16],ymm7[17],ymm12[17],ymm7[18],ymm12[18],ymm7[19],ymm12[19],ymm7[20],ymm12[20],ymm7[21],ymm12[21],ymm7[22],ymm12[22],ymm7[23],ymm12[23]
; AVX2-FAST-PERLANE-NEXT: vpshufb %ymm3, %ymm15, %ymm3
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3]
; AVX2-FAST-PERLANE-NEXT: vpblendvb %ymm0, %ymm2, %ymm3, %ymm0
; AVX2-FAST-PERLANE-NEXT: vmovdqa {{.*#+}} xmm2 = <2,u,1,u,0,u,3,u,4,u,4,u,4,u,4,u>
; AVX2-FAST-PERLANE-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload
; AVX2-FAST-PERLANE-NEXT: vpshufb %xmm2, %xmm3, %xmm3
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,0,1]
; AVX2-FAST-PERLANE-NEXT: vmovdqa {{.*#+}} ymm15 = [255,255,255,255,0,0,255,255,255,255,0,0,255,255,255,255,0,0,255,255,255,255,0,0,255,255,255,255,0,0,255,255]
; AVX2-FAST-PERLANE-NEXT: vpblendvb %ymm15, %ymm4, %ymm3, %ymm3
; AVX2-FAST-PERLANE-NEXT: vpshufb %xmm2, %xmm14, %xmm2
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,0,0,1]
; AVX2-FAST-PERLANE-NEXT: vpblendvb %ymm15, %ymm1, %ymm2, %ymm1
; AVX2-FAST-PERLANE-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [2,0,1,0,0,0,3,0,0,0,0,0,4,0,0,0,2,0,1,0,0,0,3,0,0,0,0,0,4,0,0,0]
; AVX2-FAST-PERLANE-NEXT: # ymm2 = mem[0,1,0,1]
; AVX2-FAST-PERLANE-NEXT: vpshufb %ymm2, %ymm9, %ymm4
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,2,3]
; AVX2-FAST-PERLANE-NEXT: vpblendvb %ymm15, %ymm5, %ymm4, %ymm4
; AVX2-FAST-PERLANE-NEXT: vpshufb %ymm2, %ymm8, %ymm2
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
; AVX2-FAST-PERLANE-NEXT: vpblendvb %ymm15, %ymm0, %ymm2, %ymm0
; AVX2-FAST-PERLANE-NEXT: vmovdqa {{.*#+}} xmm2 = <u,2,u,1,u,0,u,3,u,4,u,4,u,4,u,4>
; AVX2-FAST-PERLANE-NEXT: vpshufb %xmm2, %xmm6, %xmm5
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,0,0,1]
; AVX2-FAST-PERLANE-NEXT: vmovdqa {{.*#+}} ymm13 = [255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255]
; AVX2-FAST-PERLANE-NEXT: vpblendvb %ymm13, %ymm3, %ymm5, %ymm3
; AVX2-FAST-PERLANE-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload
; AVX2-FAST-PERLANE-NEXT: vpshufb %xmm2, %xmm5, %xmm2
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,0,0,1]
; AVX2-FAST-PERLANE-NEXT: vpblendvb %ymm13, %ymm1, %ymm2, %ymm1
; AVX2-FAST-PERLANE-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [0,2,0,1,0,0,0,3,0,0,0,0,0,4,0,0,0,2,0,1,0,0,0,3,0,0,0,0,0,4,0,0]
; AVX2-FAST-PERLANE-NEXT: # ymm2 = mem[0,1,0,1]
; AVX2-FAST-PERLANE-NEXT: vpshufb %ymm2, %ymm10, %ymm5
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,2,2,3]
; AVX2-FAST-PERLANE-NEXT: vpblendvb %ymm13, %ymm4, %ymm5, %ymm4
; AVX2-FAST-PERLANE-NEXT: vpshufb %ymm2, %ymm11, %ymm2
; AVX2-FAST-PERLANE-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
; AVX2-FAST-PERLANE-NEXT: vpblendvb %ymm13, %ymm0, %ymm2, %ymm0
; AVX2-FAST-PERLANE-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX2-FAST-PERLANE-NEXT: vmovdqa %ymm0, 96(%rax)
; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 160(%rax)
; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 128(%rax)
; AVX2-FAST-PERLANE-NEXT: vmovdqa %ymm4, 288(%rax)
; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 352(%rax)
; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 320(%rax)
; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 64(%rax)
; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 224(%rax)
; AVX2-FAST-PERLANE-NEXT: vmovdqa %ymm1, 192(%rax)
; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 256(%rax)
; AVX2-FAST-PERLANE-NEXT: vmovdqa %ymm3, (%rax)
; AVX2-FAST-PERLANE-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
; AVX2-FAST-PERLANE-NEXT: vmovaps %ymm0, 32(%rax)
; AVX2-FAST-PERLANE-NEXT: addq $648, %rsp # imm = 0x288
; AVX2-FAST-PERLANE-NEXT: vzeroupper
; AVX2-FAST-PERLANE-NEXT: retq
;
; AVX512F-SLOW-LABEL: store_i8_stride6_vf64:
; AVX512F-SLOW: # %bb.0:
; AVX512F-SLOW-NEXT: subq $264, %rsp # imm = 0x108
; AVX512F-SLOW-NEXT: vmovdqa (%rsi), %xmm5
; AVX512F-SLOW-NEXT: vmovdqa 32(%rsi), %xmm8
; AVX512F-SLOW-NEXT: vmovdqa (%rdi), %xmm9
; AVX512F-SLOW-NEXT: vmovdqa 32(%rdi), %xmm2
; AVX512F-SLOW-NEXT: vmovdqa (%r9), %xmm3
; AVX512F-SLOW-NEXT: vmovdqa 32(%r9), %xmm12
; AVX512F-SLOW-NEXT: vmovdqa 32(%rcx), %ymm6
; AVX512F-SLOW-NEXT: vmovdqa 32(%rdx), %ymm7
; AVX512F-SLOW-NEXT: vmovdqa 32(%rsi), %ymm13
; AVX512F-SLOW-NEXT: vmovdqa 32(%rdi), %ymm4
; AVX512F-SLOW-NEXT: vpbroadcastq {{.*#+}} xmm10 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
; AVX512F-SLOW-NEXT: vpshufb %xmm10, %xmm8, %xmm0
; AVX512F-SLOW-NEXT: vpshufb %xmm10, %xmm2, %xmm1
; AVX512F-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm1[8],xmm0[8],xmm1[9],xmm0[9],xmm1[10],xmm0[10],xmm1[11],xmm0[11],xmm1[12],xmm0[12],xmm1[13],xmm0[13],xmm1[14],xmm0[14],xmm1[15],xmm0[15]
; AVX512F-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm2[0],xmm8[0],xmm2[1],xmm8[1],xmm2[2],xmm8[2],xmm2[3],xmm8[3],xmm2[4],xmm8[4],xmm2[5],xmm8[5],xmm2[6],xmm8[6],xmm2[7],xmm8[7]
; AVX512F-SLOW-NEXT: vmovdqa64 %xmm2, %xmm26
; AVX512F-SLOW-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[0,3,2,1,4,5,6,7]
; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,5,6,5]
; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
; AVX512F-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} xmm11 = <u,6,u,5,u,8,u,7,u,9,u,9,u,9,u,9>
; AVX512F-SLOW-NEXT: vmovdqa %xmm12, %xmm2
; AVX512F-SLOW-NEXT: vpshufb %xmm11, %xmm12, %xmm0
; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} xmm12 = <u,2,u,1,u,0,u,3,u,4,u,4,u,4,u,4>
; AVX512F-SLOW-NEXT: vpshufb %xmm12, %xmm2, %xmm1
; AVX512F-SLOW-NEXT: vmovdqa64 %xmm2, %xmm27
; AVX512F-SLOW-NEXT: vinserti32x4 $2, %xmm0, %zmm1, %zmm0
; AVX512F-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-SLOW-NEXT: vpshufb %xmm10, %xmm5, %xmm0
; AVX512F-SLOW-NEXT: vpshufb %xmm10, %xmm9, %xmm1
; AVX512F-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm1[8],xmm0[8],xmm1[9],xmm0[9],xmm1[10],xmm0[10],xmm1[11],xmm0[11],xmm1[12],xmm0[12],xmm1[13],xmm0[13],xmm1[14],xmm0[14],xmm1[15],xmm0[15]
; AVX512F-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm9[0],xmm5[0],xmm9[1],xmm5[1],xmm9[2],xmm5[2],xmm9[3],xmm5[3],xmm9[4],xmm5[4],xmm9[5],xmm5[5],xmm9[6],xmm5[6],xmm9[7],xmm5[7]
; AVX512F-SLOW-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[0,3,2,1,4,5,6,7]
; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,5,6,5]
; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
; AVX512F-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-SLOW-NEXT: vpshufb %xmm11, %xmm3, %xmm0
; AVX512F-SLOW-NEXT: vpshufb %xmm12, %xmm3, %xmm1
; AVX512F-SLOW-NEXT: vmovdqa64 %xmm3, %xmm28
; AVX512F-SLOW-NEXT: vinserti32x4 $2, %xmm0, %zmm1, %zmm0
; AVX512F-SLOW-NEXT: vmovdqu64 %zmm0, (%rsp) # 64-byte Spill
; AVX512F-SLOW-NEXT: vpbroadcastq {{.*#+}} ymm11 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
; AVX512F-SLOW-NEXT: vpshufb %ymm11, %ymm13, %ymm0
; AVX512F-SLOW-NEXT: vpshufb %ymm11, %ymm4, %ymm1
; AVX512F-SLOW-NEXT: vpunpcklbw {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[4],ymm0[4],ymm1[5],ymm0[5],ymm1[6],ymm0[6],ymm1[7],ymm0[7],ymm1[16],ymm0[16],ymm1[17],ymm0[17],ymm1[18],ymm0[18],ymm1[19],ymm0[19],ymm1[20],ymm0[20],ymm1[21],ymm0[21],ymm1[22],ymm0[22],ymm1[23],ymm0[23]
; AVX512F-SLOW-NEXT: vpunpckhbw {{.*#+}} ymm1 = ymm4[8],ymm13[8],ymm4[9],ymm13[9],ymm4[10],ymm13[10],ymm4[11],ymm13[11],ymm4[12],ymm13[12],ymm4[13],ymm13[13],ymm4[14],ymm13[14],ymm4[15],ymm13[15],ymm4[24],ymm13[24],ymm4[25],ymm13[25],ymm4[26],ymm13[26],ymm4[27],ymm13[27],ymm4[28],ymm13[28],ymm4[29],ymm13[29],ymm4[30],ymm13[30],ymm4[31],ymm13[31]
; AVX512F-SLOW-NEXT: vmovdqa64 %ymm4, %ymm29
; AVX512F-SLOW-NEXT: vmovdqa %ymm13, %ymm4
; AVX512F-SLOW-NEXT: vbroadcasti128 {{.*#+}} ymm12 = [24,25,22,23,28,29,26,27,30,31,30,31,30,31,30,31,24,25,22,23,28,29,26,27,30,31,30,31,30,31,30,31]
; AVX512F-SLOW-NEXT: # ymm12 = mem[0,1,0,1]
; AVX512F-SLOW-NEXT: vpshufb %ymm12, %ymm1, %ymm1
; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
; AVX512F-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-SLOW-NEXT: vpbroadcastq {{.*#+}} ymm13 = [5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10]
; AVX512F-SLOW-NEXT: vpshufb %ymm13, %ymm6, %ymm0
; AVX512F-SLOW-NEXT: vpshufb %ymm13, %ymm7, %ymm1
; AVX512F-SLOW-NEXT: vpunpcklbw {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[4],ymm0[4],ymm1[5],ymm0[5],ymm1[6],ymm0[6],ymm1[7],ymm0[7],ymm1[16],ymm0[16],ymm1[17],ymm0[17],ymm1[18],ymm0[18],ymm1[19],ymm0[19],ymm1[20],ymm0[20],ymm1[21],ymm0[21],ymm1[22],ymm0[22],ymm1[23],ymm0[23]
; AVX512F-SLOW-NEXT: vpunpckhbw {{.*#+}} ymm1 = ymm7[8],ymm6[8],ymm7[9],ymm6[9],ymm7[10],ymm6[10],ymm7[11],ymm6[11],ymm7[12],ymm6[12],ymm7[13],ymm6[13],ymm7[14],ymm6[14],ymm7[15],ymm6[15],ymm7[24],ymm6[24],ymm7[25],ymm6[25],ymm7[26],ymm6[26],ymm7[27],ymm6[27],ymm7[28],ymm6[28],ymm7[29],ymm6[29],ymm7[30],ymm6[30],ymm7[31],ymm6[31]
; AVX512F-SLOW-NEXT: vbroadcasti128 {{.*#+}} ymm14 = [26,27,24,25,22,23,28,29,30,31,30,31,30,31,30,31,26,27,24,25,22,23,28,29,30,31,30,31,30,31,30,31]
; AVX512F-SLOW-NEXT: # ymm14 = mem[0,1,0,1]
; AVX512F-SLOW-NEXT: vpshufb %ymm14, %ymm1, %ymm1
; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
; AVX512F-SLOW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-SLOW-NEXT: vmovdqa 32(%r9), %ymm10
; AVX512F-SLOW-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [0,10,0,13,0,12,0,11,0,14,0,0,0,0,0,15,0,10,0,13,0,12,0,11,0,14,0,0,0,0,0,15]
; AVX512F-SLOW-NEXT: # ymm3 = mem[0,1,0,1]
; AVX512F-SLOW-NEXT: vpshufb %ymm3, %ymm10, %ymm0
; AVX512F-SLOW-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [0,6,0,5,0,8,0,7,0,0,0,9,0,0,0,0,0,6,0,5,0,8,0,7,0,0,0,9,0,0,0,0]
; AVX512F-SLOW-NEXT: # ymm2 = mem[0,1,0,1]
; AVX512F-SLOW-NEXT: vpshufb %ymm2, %ymm10, %ymm15
; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm15, %zmm25
; AVX512F-SLOW-NEXT: vmovdqa (%rsi), %ymm15
; AVX512F-SLOW-NEXT: vmovdqa (%rdi), %ymm1
; AVX512F-SLOW-NEXT: vpshufb %ymm11, %ymm15, %ymm0
; AVX512F-SLOW-NEXT: vpshufb %ymm11, %ymm1, %ymm11
; AVX512F-SLOW-NEXT: vpunpcklbw {{.*#+}} ymm0 = ymm11[0],ymm0[0],ymm11[1],ymm0[1],ymm11[2],ymm0[2],ymm11[3],ymm0[3],ymm11[4],ymm0[4],ymm11[5],ymm0[5],ymm11[6],ymm0[6],ymm11[7],ymm0[7],ymm11[16],ymm0[16],ymm11[17],ymm0[17],ymm11[18],ymm0[18],ymm11[19],ymm0[19],ymm11[20],ymm0[20],ymm11[21],ymm0[21],ymm11[22],ymm0[22],ymm11[23],ymm0[23]
; AVX512F-SLOW-NEXT: vpunpckhbw {{.*#+}} ymm11 = ymm1[8],ymm15[8],ymm1[9],ymm15[9],ymm1[10],ymm15[10],ymm1[11],ymm15[11],ymm1[12],ymm15[12],ymm1[13],ymm15[13],ymm1[14],ymm15[14],ymm1[15],ymm15[15],ymm1[24],ymm15[24],ymm1[25],ymm15[25],ymm1[26],ymm15[26],ymm1[27],ymm15[27],ymm1[28],ymm15[28],ymm1[29],ymm15[29],ymm1[30],ymm15[30],ymm1[31],ymm15[31]
; AVX512F-SLOW-NEXT: vpshufb %ymm12, %ymm11, %ymm11
; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm11, %zmm0, %zmm24
; AVX512F-SLOW-NEXT: vmovdqa (%rcx), %ymm0
; AVX512F-SLOW-NEXT: vmovdqa (%rdx), %ymm11
; AVX512F-SLOW-NEXT: vpshufb %ymm13, %ymm0, %ymm12
; AVX512F-SLOW-NEXT: vpshufb %ymm13, %ymm11, %ymm13
; AVX512F-SLOW-NEXT: vpunpcklbw {{.*#+}} ymm12 = ymm13[0],ymm12[0],ymm13[1],ymm12[1],ymm13[2],ymm12[2],ymm13[3],ymm12[3],ymm13[4],ymm12[4],ymm13[5],ymm12[5],ymm13[6],ymm12[6],ymm13[7],ymm12[7],ymm13[16],ymm12[16],ymm13[17],ymm12[17],ymm13[18],ymm12[18],ymm13[19],ymm12[19],ymm13[20],ymm12[20],ymm13[21],ymm12[21],ymm13[22],ymm12[22],ymm13[23],ymm12[23]
; AVX512F-SLOW-NEXT: vpunpckhbw {{.*#+}} ymm13 = ymm11[8],ymm0[8],ymm11[9],ymm0[9],ymm11[10],ymm0[10],ymm11[11],ymm0[11],ymm11[12],ymm0[12],ymm11[13],ymm0[13],ymm11[14],ymm0[14],ymm11[15],ymm0[15],ymm11[24],ymm0[24],ymm11[25],ymm0[25],ymm11[26],ymm0[26],ymm11[27],ymm0[27],ymm11[28],ymm0[28],ymm11[29],ymm0[29],ymm11[30],ymm0[30],ymm11[31],ymm0[31]
; AVX512F-SLOW-NEXT: vpshufb %ymm14, %ymm13, %ymm13
; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm13, %zmm12, %zmm23
; AVX512F-SLOW-NEXT: vmovdqa (%r9), %ymm12
; AVX512F-SLOW-NEXT: vpshufb %ymm3, %ymm12, %ymm3
; AVX512F-SLOW-NEXT: vpshufb %ymm2, %ymm12, %ymm2
; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm22
; AVX512F-SLOW-NEXT: vpunpcklbw {{.*#+}} ymm0 = ymm11[0],ymm0[0],ymm11[1],ymm0[1],ymm11[2],ymm0[2],ymm11[3],ymm0[3],ymm11[4],ymm0[4],ymm11[5],ymm0[5],ymm11[6],ymm0[6],ymm11[7],ymm0[7],ymm11[16],ymm0[16],ymm11[17],ymm0[17],ymm11[18],ymm0[18],ymm11[19],ymm0[19],ymm11[20],ymm0[20],ymm11[21],ymm0[21],ymm11[22],ymm0[22],ymm11[23],ymm0[23]
; AVX512F-SLOW-NEXT: vmovdqa64 %ymm0, %ymm19
; AVX512F-SLOW-NEXT: vmovdqa (%rcx), %xmm13
; AVX512F-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm9[8],xmm5[8],xmm9[9],xmm5[9],xmm9[10],xmm5[10],xmm9[11],xmm5[11],xmm9[12],xmm5[12],xmm9[13],xmm5[13],xmm9[14],xmm5[14],xmm9[15],xmm5[15]
; AVX512F-SLOW-NEXT: vmovdqa (%rdx), %xmm14
; AVX512F-SLOW-NEXT: vpunpcklbw {{.*#+}} ymm0 = ymm1[0],ymm15[0],ymm1[1],ymm15[1],ymm1[2],ymm15[2],ymm1[3],ymm15[3],ymm1[4],ymm15[4],ymm1[5],ymm15[5],ymm1[6],ymm15[6],ymm1[7],ymm15[7],ymm1[16],ymm15[16],ymm1[17],ymm15[17],ymm1[18],ymm15[18],ymm1[19],ymm15[19],ymm1[20],ymm15[20],ymm1[21],ymm15[21],ymm1[22],ymm15[22],ymm1[23],ymm15[23]
; AVX512F-SLOW-NEXT: vmovdqa64 %ymm0, %ymm18
; AVX512F-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm14[8],xmm13[8],xmm14[9],xmm13[9],xmm14[10],xmm13[10],xmm14[11],xmm13[11],xmm14[12],xmm13[12],xmm14[13],xmm13[13],xmm14[14],xmm13[14],xmm14[15],xmm13[15]
; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} xmm5 = [10,11,8,9,6,7,12,13,14,15,14,15,14,15,14,15]
; AVX512F-SLOW-NEXT: vpshufb %xmm5, %xmm1, %xmm15
; AVX512F-SLOW-NEXT: vmovdqa 32(%rcx), %xmm3
; AVX512F-SLOW-NEXT: vmovdqa 32(%rdx), %xmm0
; AVX512F-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm9 = xmm0[8],xmm3[8],xmm0[9],xmm3[9],xmm0[10],xmm3[10],xmm0[11],xmm3[11],xmm0[12],xmm3[12],xmm0[13],xmm3[13],xmm0[14],xmm3[14],xmm0[15],xmm3[15]
; AVX512F-SLOW-NEXT: vpshufb %xmm5, %xmm9, %xmm1
; AVX512F-SLOW-NEXT: vmovdqa64 %ymm1, %ymm30
; AVX512F-SLOW-NEXT: vpunpcklbw {{.*#+}} ymm1 = ymm7[0],ymm6[0],ymm7[1],ymm6[1],ymm7[2],ymm6[2],ymm7[3],ymm6[3],ymm7[4],ymm6[4],ymm7[5],ymm6[5],ymm7[6],ymm6[6],ymm7[7],ymm6[7],ymm7[16],ymm6[16],ymm7[17],ymm6[17],ymm7[18],ymm6[18],ymm7[19],ymm6[19],ymm7[20],ymm6[20],ymm7[21],ymm6[21],ymm7[22],ymm6[22],ymm7[23],ymm6[23]
; AVX512F-SLOW-NEXT: vmovdqa64 %ymm1, %ymm31
; AVX512F-SLOW-NEXT: vmovdqa64 %xmm26, %xmm1
; AVX512F-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm6 = xmm1[8],xmm8[8],xmm1[9],xmm8[9],xmm1[10],xmm8[10],xmm1[11],xmm8[11],xmm1[12],xmm8[12],xmm1[13],xmm8[13],xmm1[14],xmm8[14],xmm1[15],xmm8[15]
; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} xmm7 = [8,9,6,7,12,13,10,11,14,15,14,15,14,15,14,15]
; AVX512F-SLOW-NEXT: vpshufb %xmm7, %xmm2, %xmm8
; AVX512F-SLOW-NEXT: vpshufb %xmm7, %xmm6, %xmm1
; AVX512F-SLOW-NEXT: vmovdqa64 %ymm1, %ymm21
; AVX512F-SLOW-NEXT: vmovdqa64 %ymm29, %ymm1
; AVX512F-SLOW-NEXT: vpunpcklbw {{.*#+}} ymm1 = ymm1[0],ymm4[0],ymm1[1],ymm4[1],ymm1[2],ymm4[2],ymm1[3],ymm4[3],ymm1[4],ymm4[4],ymm1[5],ymm4[5],ymm1[6],ymm4[6],ymm1[7],ymm4[7],ymm1[16],ymm4[16],ymm1[17],ymm4[17],ymm1[18],ymm4[18],ymm1[19],ymm4[19],ymm1[20],ymm4[20],ymm1[21],ymm4[21],ymm1[22],ymm4[22],ymm1[23],ymm4[23]
; AVX512F-SLOW-NEXT: vmovdqa64 %ymm1, %ymm29
; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} xmm4 = <u,10,u,13,u,12,u,11,u,14,u,13,u,14,u,15>
; AVX512F-SLOW-NEXT: vmovdqa64 %xmm28, %xmm1
; AVX512F-SLOW-NEXT: vpshufb %xmm4, %xmm1, %xmm7
; AVX512F-SLOW-NEXT: vmovdqa64 %xmm27, %xmm1
; AVX512F-SLOW-NEXT: vpshufb %xmm4, %xmm1, %xmm1
; AVX512F-SLOW-NEXT: vmovdqa64 %ymm1, %ymm28
; AVX512F-SLOW-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [0,2,0,1,0,0,0,3,0,0,0,0,0,4,0,0,0,2,0,1,0,0,0,3,0,0,0,0,0,4,0,0]
; AVX512F-SLOW-NEXT: # ymm4 = mem[0,1,0,1]
; AVX512F-SLOW-NEXT: vpshufb %ymm4, %ymm12, %ymm12
; AVX512F-SLOW-NEXT: vpshufb %ymm4, %ymm10, %ymm1
; AVX512F-SLOW-NEXT: vmovdqa64 %ymm1, %ymm20
; AVX512F-SLOW-NEXT: vpbroadcastq {{.*#+}} xmm10 = [5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10]
; AVX512F-SLOW-NEXT: vpshufb %xmm10, %xmm3, %xmm1
; AVX512F-SLOW-NEXT: vpshufb %xmm10, %xmm0, %xmm9
; AVX512F-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm9[0],xmm1[0],xmm9[1],xmm1[1],xmm9[2],xmm1[2],xmm9[3],xmm1[3],xmm9[4],xmm1[4],xmm9[5],xmm1[5],xmm9[6],xmm1[6],xmm9[7],xmm1[7]
; AVX512F-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3],xmm0[4],xmm3[4],xmm0[5],xmm3[5],xmm0[6],xmm3[6],xmm0[7],xmm3[7]
; AVX512F-SLOW-NEXT: vprold $16, %xmm0, %xmm0
; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm26
; AVX512F-SLOW-NEXT: vpshufb %xmm10, %xmm13, %xmm1
; AVX512F-SLOW-NEXT: vpshufb %xmm10, %xmm14, %xmm3
; AVX512F-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3],xmm3[4],xmm1[4],xmm3[5],xmm1[5],xmm3[6],xmm1[6],xmm3[7],xmm1[7]
; AVX512F-SLOW-NEXT: vmovdqa (%r8), %xmm3
; AVX512F-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm9 = xmm14[0],xmm13[0],xmm14[1],xmm13[1],xmm14[2],xmm13[2],xmm14[3],xmm13[3],xmm14[4],xmm13[4],xmm14[5],xmm13[5],xmm14[6],xmm13[6],xmm14[7],xmm13[7]
; AVX512F-SLOW-NEXT: vmovdqa 32(%r8), %xmm10
; AVX512F-SLOW-NEXT: vprold $16, %xmm9, %xmm9
; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} xmm13 = <10,u,13,u,12,u,11,u,14,u,13,u,14,u,15,u>
; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm9, %zmm27
; AVX512F-SLOW-NEXT: vpshufb %xmm13, %xmm3, %xmm9
; AVX512F-SLOW-NEXT: vpshufb %xmm13, %xmm10, %xmm13
; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} xmm14 = <6,u,5,u,8,u,7,u,9,u,9,u,9,u,9,u>
; AVX512F-SLOW-NEXT: vpshufb %xmm14, %xmm10, %xmm11
; AVX512F-SLOW-NEXT: vmovdqa64 %ymm11, %ymm17
; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} xmm1 = <2,u,1,u,0,u,3,u,4,u,4,u,4,u,4,u>
; AVX512F-SLOW-NEXT: vpshufb %xmm1, %xmm10, %xmm11
; AVX512F-SLOW-NEXT: vpshufb %xmm14, %xmm3, %xmm14
; AVX512F-SLOW-NEXT: vpshufb %xmm1, %xmm3, %xmm0
; AVX512F-SLOW-NEXT: vmovdqa64 %ymm0, %ymm16
; AVX512F-SLOW-NEXT: vmovdqa (%r8), %ymm3
; AVX512F-SLOW-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [2,0,1,0,0,0,3,0,0,0,0,0,4,0,0,0,2,0,1,0,0,0,3,0,0,0,0,0,4,0,0,0]
; AVX512F-SLOW-NEXT: # ymm2 = mem[0,1,0,1]
; AVX512F-SLOW-NEXT: vpshufb %ymm2, %ymm3, %ymm1
; AVX512F-SLOW-NEXT: vmovdqa 32(%r8), %ymm0
; AVX512F-SLOW-NEXT: vpshufb %ymm2, %ymm0, %ymm5
; AVX512F-SLOW-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [10,0,13,0,12,0,11,0,14,0,0,0,0,0,15,0,10,0,13,0,12,0,11,0,14,0,0,0,0,0,15,0]
; AVX512F-SLOW-NEXT: # ymm4 = mem[0,1,0,1]
; AVX512F-SLOW-NEXT: vpshufb %ymm4, %ymm0, %ymm6
; AVX512F-SLOW-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [6,0,5,0,8,0,7,0,0,0,9,0,0,0,0,0,6,0,5,0,8,0,7,0,0,0,9,0,0,0,0,0]
; AVX512F-SLOW-NEXT: # ymm2 = mem[0,1,0,1]
; AVX512F-SLOW-NEXT: vpshufb %ymm2, %ymm0, %ymm0
; AVX512F-SLOW-NEXT: vpshufb %ymm4, %ymm3, %ymm4
; AVX512F-SLOW-NEXT: vpshufb %ymm2, %ymm3, %ymm2
; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm3 = ymm15[0,0,0,1]
; AVX512F-SLOW-NEXT: vprold $16, %ymm19, %ymm15
; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm8 = ymm8[0,0,0,1]
; AVX512F-SLOW-NEXT: vmovdqa64 %ymm18, %ymm10
; AVX512F-SLOW-NEXT: vpshuflw {{.*#+}} ymm10 = ymm10[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} ymm10 = ymm10[0,1,2,3,4,5,6,5,8,9,10,11,12,13,14,13]
; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,2,2,3]
; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3]
; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,0,0,1]
; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,0,0,1]
; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,2,2,3]
; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,2,2,3]
; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm15, %zmm3, %zmm3
; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm10, %zmm8, %zmm8
; AVX512F-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm10 = [65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535]
; AVX512F-SLOW-NEXT: vpternlogq $226, %zmm3, %zmm10, %zmm8
; AVX512F-SLOW-NEXT: vextracti64x4 $1, %zmm8, %ymm3
; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} ymm15 = [65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535]
; AVX512F-SLOW-NEXT: vpternlogq $184, %ymm3, %ymm15, %ymm1
; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} ymm3 = [0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0]
; AVX512F-SLOW-NEXT: vpternlogq $184, %ymm8, %ymm3, %ymm9
; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm1
; AVX512F-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm9[0,1,2,3],zmm1[4,5,6,7]
; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm12, %zmm7, %zmm7
; AVX512F-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm8 = [255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255]
; AVX512F-SLOW-NEXT: vpternlogq $184, %zmm1, %zmm8, %zmm7
; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm1 = ymm30[0,0,0,1]
; AVX512F-SLOW-NEXT: vprold $16, %ymm31, %ymm9
; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,2,3]
; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm9, %zmm1, %zmm1
; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm9 = ymm21[0,0,0,1]
; AVX512F-SLOW-NEXT: vmovdqa64 %ymm29, %ymm12
; AVX512F-SLOW-NEXT: vpshuflw {{.*#+}} ymm12 = ymm12[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15]
; AVX512F-SLOW-NEXT: vpshufhw {{.*#+}} ymm12 = ymm12[0,1,2,3,4,5,6,5,8,9,10,11,12,13,14,13]
; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,2,2,3]
; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,2,2,3]
; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm13 = ymm13[0,0,0,1]
; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm28 = ymm28[0,0,0,1]
; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm29 = ymm20[2,2,2,3]
; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm30 = ymm17[0,0,0,1]
; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm31 = ymm11[0,0,0,1]
; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm14 = ymm14[0,0,0,1]
; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm11 = ymm16[0,0,0,1]
; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,2,2,3]
; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3]
; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,2,3]
; AVX512F-SLOW-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm12, %zmm9, %zmm9
; AVX512F-SLOW-NEXT: vpternlogq $226, %zmm1, %zmm10, %zmm9
; AVX512F-SLOW-NEXT: vextracti64x4 $1, %zmm9, %ymm1
; AVX512F-SLOW-NEXT: vpternlogq $184, %ymm1, %ymm15, %ymm5
; AVX512F-SLOW-NEXT: vpternlogq $184, %ymm9, %ymm3, %ymm13
; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm5, %zmm0, %zmm1
; AVX512F-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm13[0,1,2,3],zmm1[4,5,6,7]
; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm29, %zmm28, %zmm5
; AVX512F-SLOW-NEXT: vpternlogq $184, %zmm1, %zmm8, %zmm5
; AVX512F-SLOW-NEXT: vpermq {{.*#+}} zmm1 = zmm26[0,0,0,1,4,4,4,5]
; AVX512F-SLOW-NEXT: vpermq $64, {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Folded Reload
; AVX512F-SLOW-NEXT: # zmm8 = mem[0,0,0,1,4,4,4,5]
; AVX512F-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm9 = [65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0]
; AVX512F-SLOW-NEXT: vpternlogq $226, %zmm1, %zmm9, %zmm8
; AVX512F-SLOW-NEXT: vpermq {{.*#+}} zmm1 = zmm27[0,0,0,1,4,4,4,5]
; AVX512F-SLOW-NEXT: vpermq $64, {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Folded Reload
; AVX512F-SLOW-NEXT: # zmm12 = mem[0,0,0,1,4,4,4,5]
; AVX512F-SLOW-NEXT: vpternlogq $226, %zmm1, %zmm9, %zmm12
; AVX512F-SLOW-NEXT: vextracti64x4 $1, %zmm8, %ymm1
; AVX512F-SLOW-NEXT: vmovdqa {{.*#+}} ymm9 = [65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535]
; AVX512F-SLOW-NEXT: vpternlogq $184, %ymm1, %ymm9, %ymm30
; AVX512F-SLOW-NEXT: vpternlogq $184, %ymm8, %ymm15, %ymm31
; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm30, %zmm0, %zmm1
; AVX512F-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm31[0,1,2,3],zmm1[4,5,6,7]
; AVX512F-SLOW-NEXT: vpermq $64, {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Folded Reload
; AVX512F-SLOW-NEXT: # zmm8 = mem[0,0,0,1,4,4,4,5]
; AVX512F-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm13 = [255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255]
; AVX512F-SLOW-NEXT: vpternlogq $184, %zmm1, %zmm13, %zmm8
; AVX512F-SLOW-NEXT: vpternlogq $184, %ymm12, %ymm15, %ymm11
; AVX512F-SLOW-NEXT: vextracti64x4 $1, %zmm12, %ymm1
; AVX512F-SLOW-NEXT: vpternlogq $184, %ymm1, %ymm9, %ymm14
; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm14, %zmm0, %zmm1
; AVX512F-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm11[0,1,2,3],zmm1[4,5,6,7]
; AVX512F-SLOW-NEXT: vpermq $64, (%rsp), %zmm11 # 64-byte Folded Reload
; AVX512F-SLOW-NEXT: # zmm11 = mem[0,0,0,1,4,4,4,5]
; AVX512F-SLOW-NEXT: vpternlogq $184, %zmm1, %zmm13, %zmm11
; AVX512F-SLOW-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Folded Reload
; AVX512F-SLOW-NEXT: # zmm1 = mem[2,2,2,3,6,6,6,7]
; AVX512F-SLOW-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm12 # 64-byte Folded Reload
; AVX512F-SLOW-NEXT: # zmm12 = mem[2,2,2,3,6,6,6,7]
; AVX512F-SLOW-NEXT: vpternlogq $226, %zmm1, %zmm10, %zmm12
; AVX512F-SLOW-NEXT: vpermq {{.*#+}} zmm1 = zmm24[2,2,2,3,6,6,6,7]
; AVX512F-SLOW-NEXT: vpermq {{.*#+}} zmm13 = zmm23[2,2,2,3,6,6,6,7]
; AVX512F-SLOW-NEXT: vpternlogq $226, %zmm1, %zmm10, %zmm13
; AVX512F-SLOW-NEXT: vextracti64x4 $1, %zmm12, %ymm1
; AVX512F-SLOW-NEXT: vpternlogq $184, %ymm1, %ymm3, %ymm6
; AVX512F-SLOW-NEXT: vextracti64x4 $1, %zmm13, %ymm1
; AVX512F-SLOW-NEXT: vpternlogq $184, %ymm1, %ymm3, %ymm4
; AVX512F-SLOW-NEXT: vpternlogq $184, %ymm12, %ymm9, %ymm0
; AVX512F-SLOW-NEXT: vpternlogq $184, %ymm13, %ymm9, %ymm2
; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm6, %zmm0, %zmm1
; AVX512F-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm1[4,5,6,7]
; AVX512F-SLOW-NEXT: vpermq {{.*#+}} zmm1 = zmm25[2,2,2,3,6,6,6,7]
; AVX512F-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm3 = [255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0]
; AVX512F-SLOW-NEXT: vpternlogq $184, %zmm0, %zmm3, %zmm1
; AVX512F-SLOW-NEXT: vinserti64x4 $1, %ymm4, %zmm0, %zmm0
; AVX512F-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm2[0,1,2,3],zmm0[4,5,6,7]
; AVX512F-SLOW-NEXT: vpermq {{.*#+}} zmm2 = zmm22[2,2,2,3,6,6,6,7]
; AVX512F-SLOW-NEXT: vpternlogq $184, %zmm0, %zmm3, %zmm2
; AVX512F-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512F-SLOW-NEXT: vmovdqa64 %zmm2, 128(%rax)
; AVX512F-SLOW-NEXT: vmovdqa64 %zmm1, 320(%rax)
; AVX512F-SLOW-NEXT: vmovdqa64 %zmm11, (%rax)
; AVX512F-SLOW-NEXT: vmovdqa64 %zmm8, 192(%rax)
; AVX512F-SLOW-NEXT: vmovdqa64 %zmm5, 256(%rax)
; AVX512F-SLOW-NEXT: vmovdqa64 %zmm7, 64(%rax)
; AVX512F-SLOW-NEXT: addq $264, %rsp # imm = 0x108
; AVX512F-SLOW-NEXT: vzeroupper
; AVX512F-SLOW-NEXT: retq
;
; AVX512F-FAST-LABEL: store_i8_stride6_vf64:
; AVX512F-FAST: # %bb.0:
; AVX512F-FAST-NEXT: subq $200, %rsp
; AVX512F-FAST-NEXT: vmovdqa (%rcx), %ymm4
; AVX512F-FAST-NEXT: vmovdqa (%rdx), %ymm10
; AVX512F-FAST-NEXT: vmovdqa (%rsi), %ymm3
; AVX512F-FAST-NEXT: vmovdqa 32(%rcx), %ymm2
; AVX512F-FAST-NEXT: vmovdqa 32(%rdx), %ymm6
; AVX512F-FAST-NEXT: vmovdqa 32(%rsi), %ymm7
; AVX512F-FAST-NEXT: vmovdqa 32(%rdi), %ymm8
; AVX512F-FAST-NEXT: vmovdqa 32(%r9), %ymm12
; AVX512F-FAST-NEXT: vpbroadcastq {{.*#+}} ymm5 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
; AVX512F-FAST-NEXT: vpshufb %ymm5, %ymm7, %ymm0
; AVX512F-FAST-NEXT: vpshufb %ymm5, %ymm8, %ymm1
; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[4],ymm0[4],ymm1[5],ymm0[5],ymm1[6],ymm0[6],ymm1[7],ymm0[7],ymm1[16],ymm0[16],ymm1[17],ymm0[17],ymm1[18],ymm0[18],ymm1[19],ymm0[19],ymm1[20],ymm0[20],ymm1[21],ymm0[21],ymm1[22],ymm0[22],ymm1[23],ymm0[23]
; AVX512F-FAST-NEXT: vpunpckhbw {{.*#+}} ymm1 = ymm8[8],ymm7[8],ymm8[9],ymm7[9],ymm8[10],ymm7[10],ymm8[11],ymm7[11],ymm8[12],ymm7[12],ymm8[13],ymm7[13],ymm8[14],ymm7[14],ymm8[15],ymm7[15],ymm8[24],ymm7[24],ymm8[25],ymm7[25],ymm8[26],ymm7[26],ymm8[27],ymm7[27],ymm8[28],ymm7[28],ymm8[29],ymm7[29],ymm8[30],ymm7[30],ymm8[31],ymm7[31]
; AVX512F-FAST-NEXT: vmovdqa64 %ymm8, %ymm26
; AVX512F-FAST-NEXT: vmovdqa64 %ymm7, %ymm27
; AVX512F-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm8 = [24,25,22,23,28,29,26,27,30,31,30,31,30,31,30,31,24,25,22,23,28,29,26,27,30,31,30,31,30,31,30,31]
; AVX512F-FAST-NEXT: # ymm8 = mem[0,1,0,1]
; AVX512F-FAST-NEXT: vpshufb %ymm8, %ymm1, %ymm1
; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
; AVX512F-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-FAST-NEXT: vpbroadcastq {{.*#+}} ymm9 = [5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10]
; AVX512F-FAST-NEXT: vpshufb %ymm9, %ymm2, %ymm0
; AVX512F-FAST-NEXT: vpshufb %ymm9, %ymm6, %ymm1
; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[4],ymm0[4],ymm1[5],ymm0[5],ymm1[6],ymm0[6],ymm1[7],ymm0[7],ymm1[16],ymm0[16],ymm1[17],ymm0[17],ymm1[18],ymm0[18],ymm1[19],ymm0[19],ymm1[20],ymm0[20],ymm1[21],ymm0[21],ymm1[22],ymm0[22],ymm1[23],ymm0[23]
; AVX512F-FAST-NEXT: vpunpckhbw {{.*#+}} ymm1 = ymm6[8],ymm2[8],ymm6[9],ymm2[9],ymm6[10],ymm2[10],ymm6[11],ymm2[11],ymm6[12],ymm2[12],ymm6[13],ymm2[13],ymm6[14],ymm2[14],ymm6[15],ymm2[15],ymm6[24],ymm2[24],ymm6[25],ymm2[25],ymm6[26],ymm2[26],ymm6[27],ymm2[27],ymm6[28],ymm2[28],ymm6[29],ymm2[29],ymm6[30],ymm2[30],ymm6[31],ymm2[31]
; AVX512F-FAST-NEXT: vmovdqa64 %ymm6, %ymm28
; AVX512F-FAST-NEXT: vmovdqa64 %ymm2, %ymm29
; AVX512F-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm11 = [26,27,24,25,22,23,28,29,30,31,30,31,30,31,30,31,26,27,24,25,22,23,28,29,30,31,30,31,30,31,30,31]
; AVX512F-FAST-NEXT: # ymm11 = mem[0,1,0,1]
; AVX512F-FAST-NEXT: vpshufb %ymm11, %ymm1, %ymm1
; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
; AVX512F-FAST-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [0,10,0,13,0,12,0,11,0,14,0,0,0,0,0,15,0,10,0,13,0,12,0,11,0,14,0,0,0,0,0,15]
; AVX512F-FAST-NEXT: # ymm0 = mem[0,1,0,1]
; AVX512F-FAST-NEXT: vmovdqa %ymm12, %ymm6
; AVX512F-FAST-NEXT: vpshufb %ymm0, %ymm12, %ymm2
; AVX512F-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [0,6,0,5,0,8,0,7,0,0,0,9,0,0,0,0,0,6,0,5,0,8,0,7,0,0,0,9,0,0,0,0]
; AVX512F-FAST-NEXT: # ymm1 = mem[0,1,0,1]
; AVX512F-FAST-NEXT: vpshufb %ymm1, %ymm12, %ymm12
; AVX512F-FAST-NEXT: vmovdqa64 %ymm6, %ymm30
; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm2, %zmm12, %zmm2
; AVX512F-FAST-NEXT: vmovdqu64 %zmm2, (%rsp) # 64-byte Spill
; AVX512F-FAST-NEXT: vmovdqa (%rdi), %ymm15
; AVX512F-FAST-NEXT: vmovdqa %ymm3, %ymm6
; AVX512F-FAST-NEXT: vpshufb %ymm5, %ymm3, %ymm2
; AVX512F-FAST-NEXT: vpshufb %ymm5, %ymm15, %ymm5
; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} ymm2 = ymm5[0],ymm2[0],ymm5[1],ymm2[1],ymm5[2],ymm2[2],ymm5[3],ymm2[3],ymm5[4],ymm2[4],ymm5[5],ymm2[5],ymm5[6],ymm2[6],ymm5[7],ymm2[7],ymm5[16],ymm2[16],ymm5[17],ymm2[17],ymm5[18],ymm2[18],ymm5[19],ymm2[19],ymm5[20],ymm2[20],ymm5[21],ymm2[21],ymm5[22],ymm2[22],ymm5[23],ymm2[23]
; AVX512F-FAST-NEXT: vpunpckhbw {{.*#+}} ymm5 = ymm15[8],ymm3[8],ymm15[9],ymm3[9],ymm15[10],ymm3[10],ymm15[11],ymm3[11],ymm15[12],ymm3[12],ymm15[13],ymm3[13],ymm15[14],ymm3[14],ymm15[15],ymm3[15],ymm15[24],ymm3[24],ymm15[25],ymm3[25],ymm15[26],ymm3[26],ymm15[27],ymm3[27],ymm15[28],ymm3[28],ymm15[29],ymm3[29],ymm15[30],ymm3[30],ymm15[31],ymm3[31]
; AVX512F-FAST-NEXT: vpshufb %ymm8, %ymm5, %ymm8
; AVX512F-FAST-NEXT: vmovdqa 32(%rsi), %xmm7
; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm8, %zmm2, %zmm2
; AVX512F-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-FAST-NEXT: vmovdqa 32(%rdi), %xmm8
; AVX512F-FAST-NEXT: vpshufb %ymm9, %ymm4, %ymm2
; AVX512F-FAST-NEXT: vpshufb %ymm9, %ymm10, %ymm9
; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} ymm2 = ymm9[0],ymm2[0],ymm9[1],ymm2[1],ymm9[2],ymm2[2],ymm9[3],ymm2[3],ymm9[4],ymm2[4],ymm9[5],ymm2[5],ymm9[6],ymm2[6],ymm9[7],ymm2[7],ymm9[16],ymm2[16],ymm9[17],ymm2[17],ymm9[18],ymm2[18],ymm9[19],ymm2[19],ymm9[20],ymm2[20],ymm9[21],ymm2[21],ymm9[22],ymm2[22],ymm9[23],ymm2[23]
; AVX512F-FAST-NEXT: vpunpckhbw {{.*#+}} ymm9 = ymm10[8],ymm4[8],ymm10[9],ymm4[9],ymm10[10],ymm4[10],ymm10[11],ymm4[11],ymm10[12],ymm4[12],ymm10[13],ymm4[13],ymm10[14],ymm4[14],ymm10[15],ymm4[15],ymm10[24],ymm4[24],ymm10[25],ymm4[25],ymm10[26],ymm4[26],ymm10[27],ymm4[27],ymm10[28],ymm4[28],ymm10[29],ymm4[29],ymm10[30],ymm4[30],ymm10[31],ymm4[31]
; AVX512F-FAST-NEXT: vmovdqa %ymm10, %ymm5
; AVX512F-FAST-NEXT: vpshufb %ymm11, %ymm9, %ymm11
; AVX512F-FAST-NEXT: vmovdqa (%r9), %ymm10
; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm11, %zmm2, %zmm2
; AVX512F-FAST-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
; AVX512F-FAST-NEXT: vpshufb %ymm0, %ymm10, %ymm0
; AVX512F-FAST-NEXT: vpshufb %ymm1, %ymm10, %ymm1
; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm24
; AVX512F-FAST-NEXT: vpbroadcastq {{.*#+}} xmm12 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
; AVX512F-FAST-NEXT: vpshufb %xmm12, %xmm7, %xmm0
; AVX512F-FAST-NEXT: vpshufb %xmm12, %xmm8, %xmm1
; AVX512F-FAST-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm1[8],xmm0[8],xmm1[9],xmm0[9],xmm1[10],xmm0[10],xmm1[11],xmm0[11],xmm1[12],xmm0[12],xmm1[13],xmm0[13],xmm1[14],xmm0[14],xmm1[15],xmm0[15]
; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} xmm1 = 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]
; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm13 = [0,1,6,7,4,5,2,3,8,9,10,11,12,13,10,11]
; AVX512F-FAST-NEXT: vpshufb %xmm13, %xmm1, %xmm1
; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm0, %zmm1, %zmm25
; AVX512F-FAST-NEXT: vmovdqa 32(%r9), %xmm11
; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm14 = <u,6,u,5,u,8,u,7,u,9,u,9,u,9,u,9>
; AVX512F-FAST-NEXT: vpshufb %xmm14, %xmm11, %xmm0
; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm3 = <u,2,u,1,u,0,u,3,u,4,u,4,u,4,u,4>
; AVX512F-FAST-NEXT: vpshufb %xmm3, %xmm11, %xmm1
; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm0, %zmm1, %zmm23
; AVX512F-FAST-NEXT: vmovdqa (%rsi), %xmm2
; AVX512F-FAST-NEXT: vmovdqa (%rdi), %xmm1
; AVX512F-FAST-NEXT: vpshufb %xmm12, %xmm2, %xmm0
; AVX512F-FAST-NEXT: vpshufb %xmm12, %xmm1, %xmm12
; AVX512F-FAST-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm12[8],xmm0[8],xmm12[9],xmm0[9],xmm12[10],xmm0[10],xmm12[11],xmm0[11],xmm12[12],xmm0[12],xmm12[13],xmm0[13],xmm12[14],xmm0[14],xmm12[15],xmm0[15]
; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} xmm12 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3],xmm1[4],xmm2[4],xmm1[5],xmm2[5],xmm1[6],xmm2[6],xmm1[7],xmm2[7]
; AVX512F-FAST-NEXT: vpshufb %xmm13, %xmm12, %xmm12
; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm0, %zmm12, %zmm21
; AVX512F-FAST-NEXT: vmovdqa (%r9), %xmm12
; AVX512F-FAST-NEXT: vpshufb %xmm14, %xmm12, %xmm0
; AVX512F-FAST-NEXT: vpshufb %xmm3, %xmm12, %xmm3
; AVX512F-FAST-NEXT: vinserti32x4 $2, %xmm0, %zmm3, %zmm22
; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} ymm0 = ymm5[0],ymm4[0],ymm5[1],ymm4[1],ymm5[2],ymm4[2],ymm5[3],ymm4[3],ymm5[4],ymm4[4],ymm5[5],ymm4[5],ymm5[6],ymm4[6],ymm5[7],ymm4[7],ymm5[16],ymm4[16],ymm5[17],ymm4[17],ymm5[18],ymm4[18],ymm5[19],ymm4[19],ymm5[20],ymm4[20],ymm5[21],ymm4[21],ymm5[22],ymm4[22],ymm5[23],ymm4[23]
; AVX512F-FAST-NEXT: vmovdqa64 %ymm0, %ymm20
; AVX512F-FAST-NEXT: vmovdqa (%rcx), %xmm14
; AVX512F-FAST-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm1[8],xmm2[8],xmm1[9],xmm2[9],xmm1[10],xmm2[10],xmm1[11],xmm2[11],xmm1[12],xmm2[12],xmm1[13],xmm2[13],xmm1[14],xmm2[14],xmm1[15],xmm2[15]
; AVX512F-FAST-NEXT: vmovdqa (%rdx), %xmm4
; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} ymm5 = ymm15[0],ymm6[0],ymm15[1],ymm6[1],ymm15[2],ymm6[2],ymm15[3],ymm6[3],ymm15[4],ymm6[4],ymm15[5],ymm6[5],ymm15[6],ymm6[6],ymm15[7],ymm6[7],ymm15[16],ymm6[16],ymm15[17],ymm6[17],ymm15[18],ymm6[18],ymm15[19],ymm6[19],ymm15[20],ymm6[20],ymm15[21],ymm6[21],ymm15[22],ymm6[22],ymm15[23],ymm6[23]
; AVX512F-FAST-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm4[8],xmm14[8],xmm4[9],xmm14[9],xmm4[10],xmm14[10],xmm4[11],xmm14[11],xmm4[12],xmm14[12],xmm4[13],xmm14[13],xmm4[14],xmm14[14],xmm4[15],xmm14[15]
; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm15 = [10,11,8,9,6,7,12,13,14,15,14,15,14,15,14,15]
; AVX512F-FAST-NEXT: vpshufb %xmm15, %xmm0, %xmm9
; AVX512F-FAST-NEXT: vmovdqa 32(%rcx), %xmm3
; AVX512F-FAST-NEXT: vmovdqa 32(%rdx), %xmm2
; AVX512F-FAST-NEXT: vpunpckhbw {{.*#+}} xmm6 = xmm2[8],xmm3[8],xmm2[9],xmm3[9],xmm2[10],xmm3[10],xmm2[11],xmm3[11],xmm2[12],xmm3[12],xmm2[13],xmm3[13],xmm2[14],xmm3[14],xmm2[15],xmm3[15]
; AVX512F-FAST-NEXT: vpshufb %xmm15, %xmm6, %xmm0
; AVX512F-FAST-NEXT: vmovdqa64 %ymm0, %ymm31
; AVX512F-FAST-NEXT: vmovdqa64 %ymm28, %ymm0
; AVX512F-FAST-NEXT: vmovdqa64 %ymm29, %ymm6
; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} ymm0 = ymm0[0],ymm6[0],ymm0[1],ymm6[1],ymm0[2],ymm6[2],ymm0[3],ymm6[3],ymm0[4],ymm6[4],ymm0[5],ymm6[5],ymm0[6],ymm6[6],ymm0[7],ymm6[7],ymm0[16],ymm6[16],ymm0[17],ymm6[17],ymm0[18],ymm6[18],ymm0[19],ymm6[19],ymm0[20],ymm6[20],ymm0[21],ymm6[21],ymm0[22],ymm6[22],ymm0[23],ymm6[23]
; AVX512F-FAST-NEXT: vmovdqa64 %ymm0, %ymm28
; AVX512F-FAST-NEXT: vpunpckhbw {{.*#+}} xmm7 = 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]
; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm8 = [8,9,6,7,12,13,10,11,14,15,14,15,14,15,14,15]
; AVX512F-FAST-NEXT: vpshufb %xmm8, %xmm1, %xmm13
; AVX512F-FAST-NEXT: vpshufb %xmm8, %xmm7, %xmm0
; AVX512F-FAST-NEXT: vmovdqa64 %ymm0, %ymm29
; AVX512F-FAST-NEXT: vmovdqa64 %ymm26, %ymm0
; AVX512F-FAST-NEXT: vmovdqa64 %ymm27, %ymm1
; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} ymm6 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[4],ymm1[4],ymm0[5],ymm1[5],ymm0[6],ymm1[6],ymm0[7],ymm1[7],ymm0[16],ymm1[16],ymm0[17],ymm1[17],ymm0[18],ymm1[18],ymm0[19],ymm1[19],ymm0[20],ymm1[20],ymm0[21],ymm1[21],ymm0[22],ymm1[22],ymm0[23],ymm1[23]
; AVX512F-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm7 = [16,17,22,23,20,21,18,19,24,25,26,27,28,29,26,27,16,17,22,23,20,21,18,19,24,25,26,27,28,29,26,27]
; AVX512F-FAST-NEXT: # ymm7 = mem[0,1,0,1]
; AVX512F-FAST-NEXT: vpshufb %ymm7, %ymm5, %ymm8
; AVX512F-FAST-NEXT: vpshufb %ymm7, %ymm6, %ymm0
; AVX512F-FAST-NEXT: vmovdqa64 %ymm0, %ymm18
; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm7 = <u,10,u,13,u,12,u,11,u,14,u,13,u,14,u,15>
; AVX512F-FAST-NEXT: vpshufb %xmm7, %xmm12, %xmm12
; AVX512F-FAST-NEXT: vpshufb %xmm7, %xmm11, %xmm0
; AVX512F-FAST-NEXT: vmovdqa64 %ymm0, %ymm19
; AVX512F-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm11 = [0,2,0,1,0,0,0,3,0,0,0,0,0,4,0,0,0,2,0,1,0,0,0,3,0,0,0,0,0,4,0,0]
; AVX512F-FAST-NEXT: # ymm11 = mem[0,1,0,1]
; AVX512F-FAST-NEXT: vpshufb %ymm11, %ymm10, %ymm10
; AVX512F-FAST-NEXT: vmovdqa64 %ymm30, %ymm0
; AVX512F-FAST-NEXT: vpshufb %ymm11, %ymm0, %ymm0
; AVX512F-FAST-NEXT: vmovdqa64 %ymm0, %ymm30
; AVX512F-FAST-NEXT: vmovdqa (%r8), %ymm1
; AVX512F-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm5 = [2,0,1,0,0,0,3,0,0,0,0,0,4,0,0,0,2,0,1,0,0,0,3,0,0,0,0,0,4,0,0,0]
; AVX512F-FAST-NEXT: # ymm5 = mem[0,1,0,1]
; AVX512F-FAST-NEXT: vpshufb %ymm5, %ymm1, %ymm7
; AVX512F-FAST-NEXT: vmovdqa 32(%r8), %ymm0
; AVX512F-FAST-NEXT: vpshufb %ymm5, %ymm0, %ymm11
; AVX512F-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm6 = [10,0,13,0,12,0,11,0,14,0,0,0,0,0,15,0,10,0,13,0,12,0,11,0,14,0,0,0,0,0,15,0]
; AVX512F-FAST-NEXT: # ymm6 = mem[0,1,0,1]
; AVX512F-FAST-NEXT: vpshufb %ymm6, %ymm0, %ymm15
; AVX512F-FAST-NEXT: vmovdqa64 %ymm15, %ymm17
; AVX512F-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm5 = [6,0,5,0,8,0,7,0,0,0,9,0,0,0,0,0,6,0,5,0,8,0,7,0,0,0,9,0,0,0,0,0]
; AVX512F-FAST-NEXT: # ymm5 = mem[0,1,0,1]
; AVX512F-FAST-NEXT: vpshufb %ymm5, %ymm0, %ymm15
; AVX512F-FAST-NEXT: vpshufb %ymm6, %ymm1, %ymm6
; AVX512F-FAST-NEXT: vpshufb %ymm5, %ymm1, %ymm0
; AVX512F-FAST-NEXT: vmovdqa64 %ymm0, %ymm16
; AVX512F-FAST-NEXT: vpbroadcastq {{.*#+}} xmm5 = [5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10]
; AVX512F-FAST-NEXT: vpshufb %xmm5, %xmm3, %xmm0
; AVX512F-FAST-NEXT: vpshufb %xmm5, %xmm2, %xmm1
; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3],xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} xmm1 = 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]
; AVX512F-FAST-NEXT: vprold $16, %xmm1, %xmm1
; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm26
; AVX512F-FAST-NEXT: vpshufb %xmm5, %xmm14, %xmm1
; AVX512F-FAST-NEXT: vpshufb %xmm5, %xmm4, %xmm2
; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3],xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
; AVX512F-FAST-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm4[0],xmm14[0],xmm4[1],xmm14[1],xmm4[2],xmm14[2],xmm4[3],xmm14[3],xmm4[4],xmm14[4],xmm4[5],xmm14[5],xmm4[6],xmm14[6],xmm4[7],xmm14[7]
; AVX512F-FAST-NEXT: vprold $16, %xmm2, %xmm2
; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm27
; AVX512F-FAST-NEXT: vmovdqa (%r8), %xmm2
; AVX512F-FAST-NEXT: vmovdqa 32(%r8), %xmm3
; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm4 = <10,u,13,u,12,u,11,u,14,u,13,u,14,u,15,u>
; AVX512F-FAST-NEXT: vpshufb %xmm4, %xmm2, %xmm5
; AVX512F-FAST-NEXT: vpshufb %xmm4, %xmm3, %xmm4
; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm14 = <6,u,5,u,8,u,7,u,9,u,9,u,9,u,9,u>
; AVX512F-FAST-NEXT: vpshufb %xmm14, %xmm3, %xmm0
; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} xmm1 = <2,u,1,u,0,u,3,u,4,u,4,u,4,u,4,u>
; AVX512F-FAST-NEXT: vpshufb %xmm1, %xmm3, %xmm3
; AVX512F-FAST-NEXT: vpshufb %xmm14, %xmm2, %xmm14
; AVX512F-FAST-NEXT: vpshufb %xmm1, %xmm2, %xmm1
; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm2 = ymm9[0,0,0,1]
; AVX512F-FAST-NEXT: vprold $16, %ymm20, %ymm9
; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm13 = ymm13[0,0,0,1]
; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,2,3]
; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,2,2,3]
; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,0,0,1]
; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm12 = ymm12[0,0,0,1]
; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,2,2,3]
; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,2,3]
; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm9, %zmm2, %zmm2
; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm8, %zmm13, %zmm8
; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm9 = [65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535]
; AVX512F-FAST-NEXT: vpternlogq $226, %zmm2, %zmm9, %zmm8
; AVX512F-FAST-NEXT: vextracti64x4 $1, %zmm8, %ymm2
; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm13 = [65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535]
; AVX512F-FAST-NEXT: vpternlogq $184, %ymm2, %ymm13, %ymm7
; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm2 = [0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0]
; AVX512F-FAST-NEXT: vpternlogq $184, %ymm8, %ymm2, %ymm5
; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm7, %zmm0, %zmm7
; AVX512F-FAST-NEXT: vshufi64x2 {{.*#+}} zmm5 = zmm5[0,1,2,3],zmm7[4,5,6,7]
; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm10, %zmm12, %zmm7
; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm8 = [255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255]
; AVX512F-FAST-NEXT: vpternlogq $184, %zmm5, %zmm8, %zmm7
; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm5 = ymm31[0,0,0,1]
; AVX512F-FAST-NEXT: vprold $16, %ymm28, %ymm10
; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,2,2,3]
; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm10, %zmm5, %zmm5
; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm10 = ymm29[0,0,0,1]
; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm12 = ymm18[2,2,2,3]
; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,2,2,3]
; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,0,1]
; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm28 = ymm19[0,0,0,1]
; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm29 = ymm30[2,2,2,3]
; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm30 = ymm17[2,2,2,3]
; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,2,2,3]
; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,2,2,3]
; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm31 = ymm16[2,2,2,3]
; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,0,0,1]
; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,0,1]
; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm14 = ymm14[0,0,0,1]
; AVX512F-FAST-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,0,1]
; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm12, %zmm10, %zmm10
; AVX512F-FAST-NEXT: vpternlogq $226, %zmm5, %zmm9, %zmm10
; AVX512F-FAST-NEXT: vextracti64x4 $1, %zmm10, %ymm5
; AVX512F-FAST-NEXT: vpternlogq $184, %ymm5, %ymm13, %ymm11
; AVX512F-FAST-NEXT: vpternlogq $184, %ymm10, %ymm2, %ymm4
; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm11, %zmm0, %zmm5
; AVX512F-FAST-NEXT: vshufi64x2 {{.*#+}} zmm4 = zmm4[0,1,2,3],zmm5[4,5,6,7]
; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm29, %zmm28, %zmm5
; AVX512F-FAST-NEXT: vpternlogq $184, %zmm4, %zmm8, %zmm5
; AVX512F-FAST-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Folded Reload
; AVX512F-FAST-NEXT: # zmm4 = mem[2,2,2,3,6,6,6,7]
; AVX512F-FAST-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Folded Reload
; AVX512F-FAST-NEXT: # zmm8 = mem[2,2,2,3,6,6,6,7]
; AVX512F-FAST-NEXT: vpternlogq $226, %zmm4, %zmm9, %zmm8
; AVX512F-FAST-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Folded Reload
; AVX512F-FAST-NEXT: # zmm4 = mem[2,2,2,3,6,6,6,7]
; AVX512F-FAST-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm10 # 64-byte Folded Reload
; AVX512F-FAST-NEXT: # zmm10 = mem[2,2,2,3,6,6,6,7]
; AVX512F-FAST-NEXT: vpternlogq $226, %zmm4, %zmm9, %zmm10
; AVX512F-FAST-NEXT: vextracti64x4 $1, %zmm8, %ymm4
; AVX512F-FAST-NEXT: vpternlogq $184, %ymm4, %ymm2, %ymm30
; AVX512F-FAST-NEXT: vextracti64x4 $1, %zmm10, %ymm4
; AVX512F-FAST-NEXT: vpternlogq $184, %ymm4, %ymm2, %ymm6
; AVX512F-FAST-NEXT: vmovdqa {{.*#+}} ymm2 = [65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535]
; AVX512F-FAST-NEXT: vpternlogq $184, %ymm8, %ymm2, %ymm15
; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm30, %zmm0, %zmm4
; AVX512F-FAST-NEXT: vshufi64x2 {{.*#+}} zmm4 = zmm15[0,1,2,3],zmm4[4,5,6,7]
; AVX512F-FAST-NEXT: vpermq $234, (%rsp), %zmm8 # 64-byte Folded Reload
; AVX512F-FAST-NEXT: # zmm8 = mem[2,2,2,3,6,6,6,7]
; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm9 = [255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0]
; AVX512F-FAST-NEXT: vpternlogq $184, %zmm4, %zmm9, %zmm8
; AVX512F-FAST-NEXT: vpternlogq $184, %ymm10, %ymm2, %ymm31
; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm6, %zmm0, %zmm4
; AVX512F-FAST-NEXT: vshufi64x2 {{.*#+}} zmm4 = zmm31[0,1,2,3],zmm4[4,5,6,7]
; AVX512F-FAST-NEXT: vpermq {{.*#+}} zmm6 = zmm24[2,2,2,3,6,6,6,7]
; AVX512F-FAST-NEXT: vpternlogq $184, %zmm4, %zmm9, %zmm6
; AVX512F-FAST-NEXT: vpermq {{.*#+}} zmm4 = zmm26[0,0,0,1,4,4,4,5]
; AVX512F-FAST-NEXT: vpermq {{.*#+}} zmm9 = zmm25[0,0,0,1,4,4,4,5]
; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm10 = [65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0]
; AVX512F-FAST-NEXT: vpternlogq $226, %zmm4, %zmm10, %zmm9
; AVX512F-FAST-NEXT: vpermq {{.*#+}} zmm4 = zmm27[0,0,0,1,4,4,4,5]
; AVX512F-FAST-NEXT: vpermq {{.*#+}} zmm11 = zmm21[0,0,0,1,4,4,4,5]
; AVX512F-FAST-NEXT: vpternlogq $226, %zmm4, %zmm10, %zmm11
; AVX512F-FAST-NEXT: vextracti64x4 $1, %zmm9, %ymm4
; AVX512F-FAST-NEXT: vpternlogq $184, %ymm4, %ymm2, %ymm0
; AVX512F-FAST-NEXT: vextracti64x4 $1, %zmm11, %ymm4
; AVX512F-FAST-NEXT: vpternlogq $184, %ymm4, %ymm2, %ymm14
; AVX512F-FAST-NEXT: vpternlogq $184, %ymm9, %ymm13, %ymm3
; AVX512F-FAST-NEXT: vpternlogq $184, %ymm11, %ymm13, %ymm1
; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm0
; AVX512F-FAST-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm3[0,1,2,3],zmm0[4,5,6,7]
; AVX512F-FAST-NEXT: vpermq {{.*#+}} zmm2 = zmm23[0,0,0,1,4,4,4,5]
; AVX512F-FAST-NEXT: vmovdqa64 {{.*#+}} zmm3 = [255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255]
; AVX512F-FAST-NEXT: vpternlogq $184, %zmm0, %zmm3, %zmm2
; AVX512F-FAST-NEXT: vinserti64x4 $1, %ymm14, %zmm0, %zmm0
; AVX512F-FAST-NEXT: vshufi64x2 {{.*#+}} zmm0 = zmm1[0,1,2,3],zmm0[4,5,6,7]
; AVX512F-FAST-NEXT: vpermq {{.*#+}} zmm1 = zmm22[0,0,0,1,4,4,4,5]
; AVX512F-FAST-NEXT: vpternlogq $184, %zmm0, %zmm3, %zmm1
; AVX512F-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512F-FAST-NEXT: vmovdqa64 %zmm1, (%rax)
; AVX512F-FAST-NEXT: vmovdqa64 %zmm2, 192(%rax)
; AVX512F-FAST-NEXT: vmovdqa64 %zmm6, 128(%rax)
; AVX512F-FAST-NEXT: vmovdqa64 %zmm8, 320(%rax)
; AVX512F-FAST-NEXT: vmovdqa64 %zmm5, 256(%rax)
; AVX512F-FAST-NEXT: vmovdqa64 %zmm7, 64(%rax)
; AVX512F-FAST-NEXT: addq $200, %rsp
; AVX512F-FAST-NEXT: vzeroupper
; AVX512F-FAST-NEXT: retq
;
; AVX512BW-ONLY-SLOW-LABEL: store_i8_stride6_vf64:
; AVX512BW-ONLY-SLOW: # %bb.0:
; AVX512BW-ONLY-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 (%r8), %zmm14
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 (%r9), %zmm12
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 (%rsi), %ymm16
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 (%rdi), %ymm17
; AVX512BW-ONLY-SLOW-NEXT: vpunpcklbw {{.*#+}} ymm0 = ymm17[0],ymm16[0],ymm17[1],ymm16[1],ymm17[2],ymm16[2],ymm17[3],ymm16[3],ymm17[4],ymm16[4],ymm17[5],ymm16[5],ymm17[6],ymm16[6],ymm17[7],ymm16[7],ymm17[16],ymm16[16],ymm17[17],ymm16[17],ymm17[18],ymm16[18],ymm17[19],ymm16[19],ymm17[20],ymm16[20],ymm17[21],ymm16[21],ymm17[22],ymm16[22],ymm17[23],ymm16[23]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa (%rsi), %xmm1
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa 32(%rsi), %xmm9
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa (%rdi), %xmm3
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa 32(%rdi), %xmm11
; AVX512BW-ONLY-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm3[8],xmm1[8],xmm3[9],xmm1[9],xmm3[10],xmm1[10],xmm3[11],xmm1[11],xmm3[12],xmm1[12],xmm3[13],xmm1[13],xmm3[14],xmm1[14],xmm3[15],xmm1[15]
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm2, %zmm0
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm7 = [4,3,6,5,4,3,6,5,4,3,6,5,7,7,7,7,24,27,26,25,24,27,26,25,24,27,26,25,28,29,30,29]
; AVX512BW-ONLY-SLOW-NEXT: vpermw %zmm0, %zmm7, %zmm0
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa (%rcx), %xmm2
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa 32(%rcx), %xmm8
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa (%rdx), %xmm4
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa 32(%rdx), %xmm10
; AVX512BW-ONLY-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm5 = xmm4[8],xmm2[8],xmm4[9],xmm2[9],xmm4[10],xmm2[10],xmm4[11],xmm2[11],xmm4[12],xmm2[12],xmm4[13],xmm2[13],xmm4[14],xmm2[14],xmm4[15],xmm2[15]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} ymm20 = [5,4,3,6,5,4,3,6,5,4,3,6,7,7,7,7]
; AVX512BW-ONLY-SLOW-NEXT: vpermw %ymm5, %ymm20, %ymm5
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 (%rcx), %ymm18
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 (%rdx), %ymm19
; AVX512BW-ONLY-SLOW-NEXT: vpunpcklbw {{.*#+}} ymm6 = ymm19[0],ymm18[0],ymm19[1],ymm18[1],ymm19[2],ymm18[2],ymm19[3],ymm18[3],ymm19[4],ymm18[4],ymm19[5],ymm18[5],ymm19[6],ymm18[6],ymm19[7],ymm18[7],ymm19[16],ymm18[16],ymm19[17],ymm18[17],ymm19[18],ymm18[18],ymm19[19],ymm18[19],ymm19[20],ymm18[20],ymm19[21],ymm18[21],ymm19[22],ymm18[22],ymm19[23],ymm18[23]
; AVX512BW-ONLY-SLOW-NEXT: vprold $16, %ymm6, %ymm6
; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,2,2,3]
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm6, %zmm5, %zmm5
; AVX512BW-ONLY-SLOW-NEXT: movl $613566756, %r10d # imm = 0x24924924
; AVX512BW-ONLY-SLOW-NEXT: kmovd %r10d, %k1
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu16 %zmm5, %zmm0 {%k1}
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa (%r8), %xmm5
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa 32(%r8), %xmm13
; AVX512BW-ONLY-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm6 = xmm5[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} ymm23 = [2,5,4,3,2,5,4,3,2,5,4,3,6,5,6,7]
; AVX512BW-ONLY-SLOW-NEXT: vpermw %ymm6, %ymm23, %ymm6
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 (%r8), %ymm21
; AVX512BW-ONLY-SLOW-NEXT: vbroadcasti32x4 {{.*#+}} ymm24 = [2,0,1,0,0,0,3,0,0,0,0,0,4,0,0,0,2,0,1,0,0,0,3,0,0,0,0,0,4,0,0,0]
; AVX512BW-ONLY-SLOW-NEXT: # ymm24 = mem[0,1,2,3,0,1,2,3]
; AVX512BW-ONLY-SLOW-NEXT: vpshufb %ymm24, %ymm21, %ymm15
; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,2,2,3]
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm15, %zmm6, %zmm6
; AVX512BW-ONLY-SLOW-NEXT: movl $1227133513, %r10d # imm = 0x49249249
; AVX512BW-ONLY-SLOW-NEXT: kmovd %r10d, %k2
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu16 %zmm6, %zmm0 {%k2}
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa (%r9), %xmm6
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa 32(%r9), %xmm15
; AVX512BW-ONLY-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm22 = xmm6[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
; AVX512BW-ONLY-SLOW-NEXT: vpermw %ymm22, %ymm23, %ymm25
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 (%r9), %ymm22
; AVX512BW-ONLY-SLOW-NEXT: vbroadcasti32x4 {{.*#+}} ymm26 = [0,2,0,1,0,0,0,3,0,0,0,0,0,4,0,0,0,2,0,1,0,0,0,3,0,0,0,0,0,4,0,0]
; AVX512BW-ONLY-SLOW-NEXT: # ymm26 = mem[0,1,2,3,0,1,2,3]
; AVX512BW-ONLY-SLOW-NEXT: vpshufb %ymm26, %ymm22, %ymm27
; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm27 = ymm27[2,2,2,3]
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm27, %zmm25, %zmm25
; AVX512BW-ONLY-SLOW-NEXT: movabsq $2342443691899625602, %r10 # imm = 0x2082082082082082
; AVX512BW-ONLY-SLOW-NEXT: kmovq %r10, %k3
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu8 %zmm25, %zmm0 {%k3}
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 32(%rsi), %ymm25
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 32(%rdi), %ymm27
; AVX512BW-ONLY-SLOW-NEXT: vpunpcklbw {{.*#+}} ymm28 = ymm27[0],ymm25[0],ymm27[1],ymm25[1],ymm27[2],ymm25[2],ymm27[3],ymm25[3],ymm27[4],ymm25[4],ymm27[5],ymm25[5],ymm27[6],ymm25[6],ymm27[7],ymm25[7],ymm27[16],ymm25[16],ymm27[17],ymm25[17],ymm27[18],ymm25[18],ymm27[19],ymm25[19],ymm27[20],ymm25[20],ymm27[21],ymm25[21],ymm27[22],ymm25[22],ymm27[23],ymm25[23]
; AVX512BW-ONLY-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm29 = xmm11[8],xmm9[8],xmm11[9],xmm9[9],xmm11[10],xmm9[10],xmm11[11],xmm9[11],xmm11[12],xmm9[12],xmm11[13],xmm9[13],xmm11[14],xmm9[14],xmm11[15],xmm9[15]
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm28, %zmm29, %zmm28
; AVX512BW-ONLY-SLOW-NEXT: vpermw %zmm28, %zmm7, %zmm7
; AVX512BW-ONLY-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm28 = xmm10[8],xmm8[8],xmm10[9],xmm8[9],xmm10[10],xmm8[10],xmm10[11],xmm8[11],xmm10[12],xmm8[12],xmm10[13],xmm8[13],xmm10[14],xmm8[14],xmm10[15],xmm8[15]
; AVX512BW-ONLY-SLOW-NEXT: vpermw %ymm28, %ymm20, %ymm20
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 32(%rcx), %ymm28
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 32(%rdx), %ymm29
; AVX512BW-ONLY-SLOW-NEXT: vpunpcklbw {{.*#+}} ymm30 = ymm29[0],ymm28[0],ymm29[1],ymm28[1],ymm29[2],ymm28[2],ymm29[3],ymm28[3],ymm29[4],ymm28[4],ymm29[5],ymm28[5],ymm29[6],ymm28[6],ymm29[7],ymm28[7],ymm29[16],ymm28[16],ymm29[17],ymm28[17],ymm29[18],ymm28[18],ymm29[19],ymm28[19],ymm29[20],ymm28[20],ymm29[21],ymm28[21],ymm29[22],ymm28[22],ymm29[23],ymm28[23]
; AVX512BW-ONLY-SLOW-NEXT: vprold $16, %ymm30, %ymm30
; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm30 = ymm30[2,2,2,3]
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm30, %zmm20, %zmm20
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu16 %zmm20, %zmm7 {%k1}
; AVX512BW-ONLY-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm20 = xmm13[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
; AVX512BW-ONLY-SLOW-NEXT: vpermw %ymm20, %ymm23, %ymm20
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 32(%r8), %ymm30
; AVX512BW-ONLY-SLOW-NEXT: vpshufb %ymm24, %ymm30, %ymm24
; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm24 = ymm24[2,2,2,3]
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm24, %zmm20, %zmm20
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu16 %zmm20, %zmm7 {%k2}
; AVX512BW-ONLY-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm20 = xmm15[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
; AVX512BW-ONLY-SLOW-NEXT: vpermw %ymm20, %ymm23, %ymm20
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 32(%r9), %ymm23
; AVX512BW-ONLY-SLOW-NEXT: vpshufb %ymm26, %ymm23, %ymm24
; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm24 = ymm24[2,2,2,3]
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm24, %zmm20, %zmm20
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu8 %zmm20, %zmm7 {%k3}
; AVX512BW-ONLY-SLOW-NEXT: vpbroadcastq {{.*#+}} ymm24 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
; AVX512BW-ONLY-SLOW-NEXT: vpshufb %ymm24, %ymm25, %ymm20
; AVX512BW-ONLY-SLOW-NEXT: vpshufb %ymm24, %ymm27, %ymm26
; AVX512BW-ONLY-SLOW-NEXT: vpunpcklbw {{.*#+}} ymm20 = ymm26[0],ymm20[0],ymm26[1],ymm20[1],ymm26[2],ymm20[2],ymm26[3],ymm20[3],ymm26[4],ymm20[4],ymm26[5],ymm20[5],ymm26[6],ymm20[6],ymm26[7],ymm20[7],ymm26[16],ymm20[16],ymm26[17],ymm20[17],ymm26[18],ymm20[18],ymm26[19],ymm20[19],ymm26[20],ymm20[20],ymm26[21],ymm20[21],ymm26[22],ymm20[22],ymm26[23],ymm20[23]
; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm20 = ymm20[2,2,2,3]
; AVX512BW-ONLY-SLOW-NEXT: vpunpckhbw {{.*#+}} ymm25 = ymm27[8],ymm25[8],ymm27[9],ymm25[9],ymm27[10],ymm25[10],ymm27[11],ymm25[11],ymm27[12],ymm25[12],ymm27[13],ymm25[13],ymm27[14],ymm25[14],ymm27[15],ymm25[15],ymm27[24],ymm25[24],ymm27[25],ymm25[25],ymm27[26],ymm25[26],ymm27[27],ymm25[27],ymm27[28],ymm25[28],ymm27[29],ymm25[29],ymm27[30],ymm25[30],ymm27[31],ymm25[31]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} ymm26 = [12,11,14,13,12,11,14,13,12,11,14,13,15,15,15,15]
; AVX512BW-ONLY-SLOW-NEXT: vpermw %ymm25, %ymm26, %ymm25
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm25, %zmm20, %zmm25
; AVX512BW-ONLY-SLOW-NEXT: vpbroadcastq {{.*#+}} ymm27 = [5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10]
; AVX512BW-ONLY-SLOW-NEXT: vpshufb %ymm27, %ymm28, %ymm20
; AVX512BW-ONLY-SLOW-NEXT: vpshufb %ymm27, %ymm29, %ymm31
; AVX512BW-ONLY-SLOW-NEXT: vpunpcklbw {{.*#+}} ymm20 = ymm31[0],ymm20[0],ymm31[1],ymm20[1],ymm31[2],ymm20[2],ymm31[3],ymm20[3],ymm31[4],ymm20[4],ymm31[5],ymm20[5],ymm31[6],ymm20[6],ymm31[7],ymm20[7],ymm31[16],ymm20[16],ymm31[17],ymm20[17],ymm31[18],ymm20[18],ymm31[19],ymm20[19],ymm31[20],ymm20[20],ymm31[21],ymm20[21],ymm31[22],ymm20[22],ymm31[23],ymm20[23]
; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm20 = ymm20[2,2,2,3]
; AVX512BW-ONLY-SLOW-NEXT: vpunpckhbw {{.*#+}} ymm28 = ymm29[8],ymm28[8],ymm29[9],ymm28[9],ymm29[10],ymm28[10],ymm29[11],ymm28[11],ymm29[12],ymm28[12],ymm29[13],ymm28[13],ymm29[14],ymm28[14],ymm29[15],ymm28[15],ymm29[24],ymm28[24],ymm29[25],ymm28[25],ymm29[26],ymm28[26],ymm29[27],ymm28[27],ymm29[28],ymm28[28],ymm29[29],ymm28[29],ymm29[30],ymm28[30],ymm29[31],ymm28[31]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} ymm29 = [13,12,11,14,13,12,11,14,13,12,11,14,15,15,15,15]
; AVX512BW-ONLY-SLOW-NEXT: vpermw %ymm28, %ymm29, %ymm28
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm28, %zmm20, %zmm20
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu16 %zmm25, %zmm20 {%k1}
; AVX512BW-ONLY-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm25 = zmm30[0,1,2,3],zmm14[4,5,6,7]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm28 = <u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,6,u,5,u,8,u,7,u,u,u,9,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,10,u,13,u,12,u,11,u,14,u,u,u,u,u,15,u>
; AVX512BW-ONLY-SLOW-NEXT: vpshufb %zmm28, %zmm25, %zmm25
; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} zmm25 = zmm25[2,2,2,3,6,6,6,7]
; AVX512BW-ONLY-SLOW-NEXT: movl $-1840700270, %ecx # imm = 0x92492492
; AVX512BW-ONLY-SLOW-NEXT: kmovd %ecx, %k2
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu16 %zmm25, %zmm20 {%k2}
; AVX512BW-ONLY-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm23 = zmm23[0,1,2,3],zmm12[4,5,6,7]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm25 = <u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,6,u,5,u,8,u,7,u,u,u,9,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,10,u,13,u,12,u,11,u,14,u,u,u,u,u,15>
; AVX512BW-ONLY-SLOW-NEXT: vpshufb %zmm25, %zmm23, %zmm23
; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} zmm23 = zmm23[2,2,2,3,6,6,6,7]
; AVX512BW-ONLY-SLOW-NEXT: movabsq $-9076969306111049208, %rcx # imm = 0x8208208208208208
; AVX512BW-ONLY-SLOW-NEXT: kmovq %rcx, %k3
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu8 %zmm23, %zmm20 {%k3}
; AVX512BW-ONLY-SLOW-NEXT: vpshufb %ymm24, %ymm16, %ymm23
; AVX512BW-ONLY-SLOW-NEXT: vpshufb %ymm24, %ymm17, %ymm24
; AVX512BW-ONLY-SLOW-NEXT: vpunpcklbw {{.*#+}} ymm23 = ymm24[0],ymm23[0],ymm24[1],ymm23[1],ymm24[2],ymm23[2],ymm24[3],ymm23[3],ymm24[4],ymm23[4],ymm24[5],ymm23[5],ymm24[6],ymm23[6],ymm24[7],ymm23[7],ymm24[16],ymm23[16],ymm24[17],ymm23[17],ymm24[18],ymm23[18],ymm24[19],ymm23[19],ymm24[20],ymm23[20],ymm24[21],ymm23[21],ymm24[22],ymm23[22],ymm24[23],ymm23[23]
; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm23 = ymm23[2,2,2,3]
; AVX512BW-ONLY-SLOW-NEXT: vpunpckhbw {{.*#+}} ymm16 = ymm17[8],ymm16[8],ymm17[9],ymm16[9],ymm17[10],ymm16[10],ymm17[11],ymm16[11],ymm17[12],ymm16[12],ymm17[13],ymm16[13],ymm17[14],ymm16[14],ymm17[15],ymm16[15],ymm17[24],ymm16[24],ymm17[25],ymm16[25],ymm17[26],ymm16[26],ymm17[27],ymm16[27],ymm17[28],ymm16[28],ymm17[29],ymm16[29],ymm17[30],ymm16[30],ymm17[31],ymm16[31]
; AVX512BW-ONLY-SLOW-NEXT: vpermw %ymm16, %ymm26, %ymm16
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm16, %zmm23, %zmm17
; AVX512BW-ONLY-SLOW-NEXT: vpshufb %ymm27, %ymm18, %ymm16
; AVX512BW-ONLY-SLOW-NEXT: vpshufb %ymm27, %ymm19, %ymm23
; AVX512BW-ONLY-SLOW-NEXT: vpunpcklbw {{.*#+}} ymm16 = ymm23[0],ymm16[0],ymm23[1],ymm16[1],ymm23[2],ymm16[2],ymm23[3],ymm16[3],ymm23[4],ymm16[4],ymm23[5],ymm16[5],ymm23[6],ymm16[6],ymm23[7],ymm16[7],ymm23[16],ymm16[16],ymm23[17],ymm16[17],ymm23[18],ymm16[18],ymm23[19],ymm16[19],ymm23[20],ymm16[20],ymm23[21],ymm16[21],ymm23[22],ymm16[22],ymm23[23],ymm16[23]
; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm16 = ymm16[2,2,2,3]
; AVX512BW-ONLY-SLOW-NEXT: vpunpckhbw {{.*#+}} ymm18 = ymm19[8],ymm18[8],ymm19[9],ymm18[9],ymm19[10],ymm18[10],ymm19[11],ymm18[11],ymm19[12],ymm18[12],ymm19[13],ymm18[13],ymm19[14],ymm18[14],ymm19[15],ymm18[15],ymm19[24],ymm18[24],ymm19[25],ymm18[25],ymm19[26],ymm18[26],ymm19[27],ymm18[27],ymm19[28],ymm18[28],ymm19[29],ymm18[29],ymm19[30],ymm18[30],ymm19[31],ymm18[31]
; AVX512BW-ONLY-SLOW-NEXT: vpermw %ymm18, %ymm29, %ymm18
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm18, %zmm16, %zmm16
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu16 %zmm17, %zmm16 {%k1}
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm21, %zmm14, %zmm14
; AVX512BW-ONLY-SLOW-NEXT: vpshufb %zmm28, %zmm14, %zmm14
; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} zmm14 = zmm14[2,2,2,3,6,6,6,7]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu16 %zmm14, %zmm16 {%k2}
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm22, %zmm12, %zmm12
; AVX512BW-ONLY-SLOW-NEXT: vpshufb %zmm25, %zmm12, %zmm12
; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} zmm12 = zmm12[2,2,2,3,6,6,6,7]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu8 %zmm12, %zmm16 {%k3}
; AVX512BW-ONLY-SLOW-NEXT: vpbroadcastq {{.*#+}} xmm12 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
; AVX512BW-ONLY-SLOW-NEXT: vpshufb %xmm12, %xmm9, %xmm14
; AVX512BW-ONLY-SLOW-NEXT: vpshufb %xmm12, %xmm11, %xmm17
; AVX512BW-ONLY-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm14 = xmm17[8],xmm14[8],xmm17[9],xmm14[9],xmm17[10],xmm14[10],xmm17[11],xmm14[11],xmm17[12],xmm14[12],xmm17[13],xmm14[13],xmm17[14],xmm14[14],xmm17[15],xmm14[15]
; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm14 = ymm14[0,0,0,1]
; AVX512BW-ONLY-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm9 = xmm11[0],xmm9[0],xmm11[1],xmm9[1],xmm11[2],xmm9[2],xmm11[3],xmm9[3],xmm11[4],xmm9[4],xmm11[5],xmm9[5],xmm11[6],xmm9[6],xmm11[7],xmm9[7]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa {{.*#+}} ymm11 = [0,3,2,1,0,3,2,1,0,3,2,1,4,5,6,5]
; AVX512BW-ONLY-SLOW-NEXT: vpermw %ymm9, %ymm11, %ymm9
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm14, %zmm9, %zmm9
; AVX512BW-ONLY-SLOW-NEXT: vpbroadcastq {{.*#+}} xmm14 = [5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10]
; AVX512BW-ONLY-SLOW-NEXT: vpshufb %xmm14, %xmm8, %xmm17
; AVX512BW-ONLY-SLOW-NEXT: vpshufb %xmm14, %xmm10, %xmm18
; AVX512BW-ONLY-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm17 = xmm18[0],xmm17[0],xmm18[1],xmm17[1],xmm18[2],xmm17[2],xmm18[3],xmm17[3],xmm18[4],xmm17[4],xmm18[5],xmm17[5],xmm18[6],xmm17[6],xmm18[7],xmm17[7]
; AVX512BW-ONLY-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm8 = xmm10[0],xmm8[0],xmm10[1],xmm8[1],xmm10[2],xmm8[2],xmm10[3],xmm8[3],xmm10[4],xmm8[4],xmm10[5],xmm8[5],xmm10[6],xmm8[6],xmm10[7],xmm8[7]
; AVX512BW-ONLY-SLOW-NEXT: vprold $16, %xmm8, %xmm8
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm17, %zmm8, %zmm8
; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} zmm8 = zmm8[0,0,0,1,4,4,4,5]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu16 %zmm8, %zmm9 {%k2}
; AVX512BW-ONLY-SLOW-NEXT: vpmovzxbw {{.*#+}} xmm8 = xmm13[0],zero,xmm13[1],zero,xmm13[2],zero,xmm13[3],zero,xmm13[4],zero,xmm13[5],zero,xmm13[6],zero,xmm13[7],zero
; AVX512BW-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm10 = xmm13[2,1,2,3]
; AVX512BW-ONLY-SLOW-NEXT: vpmovzxbw {{.*#+}} xmm10 = xmm10[0],zero,xmm10[1],zero,xmm10[2],zero,xmm10[3],zero,xmm10[4],zero,xmm10[5],zero,xmm10[6],zero,xmm10[7],zero
; AVX512BW-ONLY-SLOW-NEXT: vinserti32x4 $2, %xmm10, %zmm8, %zmm8
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm10 = [2,1,0,3,2,1,0,3,2,1,0,3,4,4,4,4,22,21,16,23,22,21,16,23,22,21,16,23,17,17,17,17]
; AVX512BW-ONLY-SLOW-NEXT: vpermw %zmm8, %zmm10, %zmm9 {%k1}
; AVX512BW-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm8 = xmm15[2,1,2,3]
; AVX512BW-ONLY-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm8 = xmm8[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
; AVX512BW-ONLY-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm13 = xmm15[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
; AVX512BW-ONLY-SLOW-NEXT: vinserti32x4 $2, %xmm8, %zmm13, %zmm8
; AVX512BW-ONLY-SLOW-NEXT: vpermw %zmm8, %zmm10, %zmm8
; AVX512BW-ONLY-SLOW-NEXT: movabsq $585610922974906400, %rcx # imm = 0x820820820820820
; AVX512BW-ONLY-SLOW-NEXT: kmovq %rcx, %k3
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu8 %zmm8, %zmm9 {%k3}
; AVX512BW-ONLY-SLOW-NEXT: vpshufb %xmm12, %xmm1, %xmm8
; AVX512BW-ONLY-SLOW-NEXT: vpshufb %xmm12, %xmm3, %xmm12
; AVX512BW-ONLY-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm8 = xmm12[8],xmm8[8],xmm12[9],xmm8[9],xmm12[10],xmm8[10],xmm12[11],xmm8[11],xmm12[12],xmm8[12],xmm12[13],xmm8[13],xmm12[14],xmm8[14],xmm12[15],xmm8[15]
; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} ymm8 = ymm8[0,0,0,1]
; AVX512BW-ONLY-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3],xmm3[4],xmm1[4],xmm3[5],xmm1[5],xmm3[6],xmm1[6],xmm3[7],xmm1[7]
; AVX512BW-ONLY-SLOW-NEXT: vpermw %ymm1, %ymm11, %ymm1
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm8, %zmm1, %zmm1
; AVX512BW-ONLY-SLOW-NEXT: vpshufb %xmm14, %xmm2, %xmm3
; AVX512BW-ONLY-SLOW-NEXT: vpshufb %xmm14, %xmm4, %xmm8
; AVX512BW-ONLY-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm8[0],xmm3[0],xmm8[1],xmm3[1],xmm8[2],xmm3[2],xmm8[3],xmm3[3],xmm8[4],xmm3[4],xmm8[5],xmm3[5],xmm8[6],xmm3[6],xmm8[7],xmm3[7]
; AVX512BW-ONLY-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm2 = 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]
; AVX512BW-ONLY-SLOW-NEXT: vprold $16, %xmm2, %xmm2
; AVX512BW-ONLY-SLOW-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm2
; AVX512BW-ONLY-SLOW-NEXT: vpermq {{.*#+}} zmm2 = zmm2[0,0,0,1,4,4,4,5]
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu16 %zmm2, %zmm1 {%k2}
; AVX512BW-ONLY-SLOW-NEXT: vpmovzxbw {{.*#+}} xmm2 = xmm5[0],zero,xmm5[1],zero,xmm5[2],zero,xmm5[3],zero,xmm5[4],zero,xmm5[5],zero,xmm5[6],zero,xmm5[7],zero
; AVX512BW-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm3 = xmm5[2,1,2,3]
; AVX512BW-ONLY-SLOW-NEXT: vpmovzxbw {{.*#+}} xmm3 = xmm3[0],zero,xmm3[1],zero,xmm3[2],zero,xmm3[3],zero,xmm3[4],zero,xmm3[5],zero,xmm3[6],zero,xmm3[7],zero
; AVX512BW-ONLY-SLOW-NEXT: vinserti32x4 $2, %xmm3, %zmm2, %zmm2
; AVX512BW-ONLY-SLOW-NEXT: vpermw %zmm2, %zmm10, %zmm1 {%k1}
; AVX512BW-ONLY-SLOW-NEXT: vpshufd {{.*#+}} xmm2 = xmm6[2,1,2,3]
; AVX512BW-ONLY-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm2[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
; AVX512BW-ONLY-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm6[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
; AVX512BW-ONLY-SLOW-NEXT: vinserti32x4 $2, %xmm2, %zmm3, %zmm2
; AVX512BW-ONLY-SLOW-NEXT: vpermw %zmm2, %zmm10, %zmm2
; AVX512BW-ONLY-SLOW-NEXT: vmovdqu8 %zmm2, %zmm1 {%k3}
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm1, (%rax)
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm9, 192(%rax)
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm16, 128(%rax)
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm20, 320(%rax)
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm7, 256(%rax)
; AVX512BW-ONLY-SLOW-NEXT: vmovdqa64 %zmm0, 64(%rax)
; AVX512BW-ONLY-SLOW-NEXT: vzeroupper
; AVX512BW-ONLY-SLOW-NEXT: retq
;
; AVX512BW-FAST-LABEL: store_i8_stride6_vf64:
; AVX512BW-FAST: # %bb.0:
; AVX512BW-FAST-NEXT: vmovdqa64 (%r8), %zmm9
; AVX512BW-FAST-NEXT: vmovdqa64 (%r9), %zmm10
; AVX512BW-FAST-NEXT: vmovdqa 32(%rsi), %ymm1
; AVX512BW-FAST-NEXT: vpbroadcastq {{.*#+}} ymm7 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
; AVX512BW-FAST-NEXT: vpshufb %ymm7, %ymm1, %ymm0
; AVX512BW-FAST-NEXT: vmovdqa 32(%rdi), %ymm2
; AVX512BW-FAST-NEXT: vpshufb %ymm7, %ymm2, %ymm3
; AVX512BW-FAST-NEXT: vpunpcklbw {{.*#+}} ymm0 = ymm3[0],ymm0[0],ymm3[1],ymm0[1],ymm3[2],ymm0[2],ymm3[3],ymm0[3],ymm3[4],ymm0[4],ymm3[5],ymm0[5],ymm3[6],ymm0[6],ymm3[7],ymm0[7],ymm3[16],ymm0[16],ymm3[17],ymm0[17],ymm3[18],ymm0[18],ymm3[19],ymm0[19],ymm3[20],ymm0[20],ymm3[21],ymm0[21],ymm3[22],ymm0[22],ymm3[23],ymm0[23]
; AVX512BW-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3]
; AVX512BW-FAST-NEXT: vpunpckhbw {{.*#+}} ymm3 = ymm2[8],ymm1[8],ymm2[9],ymm1[9],ymm2[10],ymm1[10],ymm2[11],ymm1[11],ymm2[12],ymm1[12],ymm2[13],ymm1[13],ymm2[14],ymm1[14],ymm2[15],ymm1[15],ymm2[24],ymm1[24],ymm2[25],ymm1[25],ymm2[26],ymm1[26],ymm2[27],ymm1[27],ymm2[28],ymm1[28],ymm2[29],ymm1[29],ymm2[30],ymm1[30],ymm2[31],ymm1[31]
; AVX512BW-FAST-NEXT: vmovdqa {{.*#+}} ymm8 = [12,11,14,13,12,11,14,13,12,11,14,13,15,15,15,15]
; AVX512BW-FAST-NEXT: vpermw %ymm3, %ymm8, %ymm3
; AVX512BW-FAST-NEXT: vinserti64x4 $1, %ymm3, %zmm0, %zmm3
; AVX512BW-FAST-NEXT: vmovdqa 32(%rcx), %ymm5
; AVX512BW-FAST-NEXT: vpbroadcastq {{.*#+}} ymm12 = [5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10]
; AVX512BW-FAST-NEXT: vpshufb %ymm12, %ymm5, %ymm0
; AVX512BW-FAST-NEXT: vmovdqa 32(%rdx), %ymm6
; AVX512BW-FAST-NEXT: vpshufb %ymm12, %ymm6, %ymm4
; AVX512BW-FAST-NEXT: vpunpcklbw {{.*#+}} ymm0 = ymm4[0],ymm0[0],ymm4[1],ymm0[1],ymm4[2],ymm0[2],ymm4[3],ymm0[3],ymm4[4],ymm0[4],ymm4[5],ymm0[5],ymm4[6],ymm0[6],ymm4[7],ymm0[7],ymm4[16],ymm0[16],ymm4[17],ymm0[17],ymm4[18],ymm0[18],ymm4[19],ymm0[19],ymm4[20],ymm0[20],ymm4[21],ymm0[21],ymm4[22],ymm0[22],ymm4[23],ymm0[23]
; AVX512BW-FAST-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3]
; AVX512BW-FAST-NEXT: vpunpckhbw {{.*#+}} ymm4 = ymm6[8],ymm5[8],ymm6[9],ymm5[9],ymm6[10],ymm5[10],ymm6[11],ymm5[11],ymm6[12],ymm5[12],ymm6[13],ymm5[13],ymm6[14],ymm5[14],ymm6[15],ymm5[15],ymm6[24],ymm5[24],ymm6[25],ymm5[25],ymm6[26],ymm5[26],ymm6[27],ymm5[27],ymm6[28],ymm5[28],ymm6[29],ymm5[29],ymm6[30],ymm5[30],ymm6[31],ymm5[31]
; AVX512BW-FAST-NEXT: vmovdqa {{.*#+}} ymm14 = [13,12,11,14,13,12,11,14,13,12,11,14,15,15,15,15]
; AVX512BW-FAST-NEXT: vpermw %ymm4, %ymm14, %ymm4
; AVX512BW-FAST-NEXT: vinserti64x4 $1, %ymm4, %zmm0, %zmm0
; AVX512BW-FAST-NEXT: movl $613566756, %eax # imm = 0x24924924
; AVX512BW-FAST-NEXT: kmovd %eax, %k1
; AVX512BW-FAST-NEXT: vmovdqu16 %zmm3, %zmm0 {%k1}
; AVX512BW-FAST-NEXT: vmovdqa 32(%r8), %ymm3
; AVX512BW-FAST-NEXT: vshufi64x2 {{.*#+}} zmm4 = zmm3[0,1,2,3],zmm9[4,5,6,7]
; AVX512BW-FAST-NEXT: vmovdqa64 {{.*#+}} zmm15 = <u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,6,u,5,u,8,u,7,u,u,u,9,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,10,u,13,u,12,u,11,u,14,u,u,u,u,u,15,u>
; AVX512BW-FAST-NEXT: vpshufb %zmm15, %zmm4, %zmm4
; AVX512BW-FAST-NEXT: vpermq {{.*#+}} zmm4 = zmm4[2,2,2,3,6,6,6,7]
; AVX512BW-FAST-NEXT: movl $-1840700270, %eax # imm = 0x92492492
; AVX512BW-FAST-NEXT: kmovd %eax, %k2
; AVX512BW-FAST-NEXT: vmovdqu16 %zmm4, %zmm0 {%k2}
; AVX512BW-FAST-NEXT: vmovdqa 32(%r9), %ymm4
; AVX512BW-FAST-NEXT: vshufi64x2 {{.*#+}} zmm11 = zmm4[0,1,2,3],zmm10[4,5,6,7]
; AVX512BW-FAST-NEXT: vmovdqa64 {{.*#+}} zmm16 = <u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,6,u,5,u,8,u,7,u,u,u,9,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,10,u,13,u,12,u,11,u,14,u,u,u,u,u,15>
; AVX512BW-FAST-NEXT: vpshufb %zmm16, %zmm11, %zmm11
; AVX512BW-FAST-NEXT: vpermq {{.*#+}} zmm11 = zmm11[2,2,2,3,6,6,6,7]
; AVX512BW-FAST-NEXT: movabsq $-9076969306111049208, %rax # imm = 0x8208208208208208
; AVX512BW-FAST-NEXT: kmovq %rax, %k3
; AVX512BW-FAST-NEXT: vmovdqu8 %zmm11, %zmm0 {%k3}
; AVX512BW-FAST-NEXT: vmovdqa (%rsi), %ymm11
; AVX512BW-FAST-NEXT: vpshufb %ymm7, %ymm11, %ymm17
; AVX512BW-FAST-NEXT: vmovdqa (%rdi), %ymm13
; AVX512BW-FAST-NEXT: vpshufb %ymm7, %ymm13, %ymm7
; AVX512BW-FAST-NEXT: vpunpcklbw {{.*#+}} ymm7 = ymm7[0],ymm17[0],ymm7[1],ymm17[1],ymm7[2],ymm17[2],ymm7[3],ymm17[3],ymm7[4],ymm17[4],ymm7[5],ymm17[5],ymm7[6],ymm17[6],ymm7[7],ymm17[7],ymm7[16],ymm17[16],ymm7[17],ymm17[17],ymm7[18],ymm17[18],ymm7[19],ymm17[19],ymm7[20],ymm17[20],ymm7[21],ymm17[21],ymm7[22],ymm17[22],ymm7[23],ymm17[23]
; AVX512BW-FAST-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,2,2,3]
; AVX512BW-FAST-NEXT: vpunpckhbw {{.*#+}} ymm17 = ymm13[8],ymm11[8],ymm13[9],ymm11[9],ymm13[10],ymm11[10],ymm13[11],ymm11[11],ymm13[12],ymm11[12],ymm13[13],ymm11[13],ymm13[14],ymm11[14],ymm13[15],ymm11[15],ymm13[24],ymm11[24],ymm13[25],ymm11[25],ymm13[26],ymm11[26],ymm13[27],ymm11[27],ymm13[28],ymm11[28],ymm13[29],ymm11[29],ymm13[30],ymm11[30],ymm13[31],ymm11[31]
; AVX512BW-FAST-NEXT: vpermw %ymm17, %ymm8, %ymm8
; AVX512BW-FAST-NEXT: vinserti64x4 $1, %ymm8, %zmm7, %zmm8
; AVX512BW-FAST-NEXT: vmovdqa64 (%rcx), %ymm17
; AVX512BW-FAST-NEXT: vpshufb %ymm12, %ymm17, %ymm7
; AVX512BW-FAST-NEXT: vmovdqa64 (%rdx), %ymm18
; AVX512BW-FAST-NEXT: vpshufb %ymm12, %ymm18, %ymm12
; AVX512BW-FAST-NEXT: vpunpcklbw {{.*#+}} ymm7 = ymm12[0],ymm7[0],ymm12[1],ymm7[1],ymm12[2],ymm7[2],ymm12[3],ymm7[3],ymm12[4],ymm7[4],ymm12[5],ymm7[5],ymm12[6],ymm7[6],ymm12[7],ymm7[7],ymm12[16],ymm7[16],ymm12[17],ymm7[17],ymm12[18],ymm7[18],ymm12[19],ymm7[19],ymm12[20],ymm7[20],ymm12[21],ymm7[21],ymm12[22],ymm7[22],ymm12[23],ymm7[23]
; AVX512BW-FAST-NEXT: vpermq {{.*#+}} ymm7 = ymm7[2,2,2,3]
; AVX512BW-FAST-NEXT: vpunpckhbw {{.*#+}} ymm12 = ymm18[8],ymm17[8],ymm18[9],ymm17[9],ymm18[10],ymm17[10],ymm18[11],ymm17[11],ymm18[12],ymm17[12],ymm18[13],ymm17[13],ymm18[14],ymm17[14],ymm18[15],ymm17[15],ymm18[24],ymm17[24],ymm18[25],ymm17[25],ymm18[26],ymm17[26],ymm18[27],ymm17[27],ymm18[28],ymm17[28],ymm18[29],ymm17[29],ymm18[30],ymm17[30],ymm18[31],ymm17[31]
; AVX512BW-FAST-NEXT: vpermw %ymm12, %ymm14, %ymm12
; AVX512BW-FAST-NEXT: vinserti64x4 $1, %ymm12, %zmm7, %zmm7
; AVX512BW-FAST-NEXT: vmovdqu16 %zmm8, %zmm7 {%k1}
; AVX512BW-FAST-NEXT: vmovdqa (%r8), %ymm8
; AVX512BW-FAST-NEXT: vinserti64x4 $1, %ymm8, %zmm9, %zmm9
; AVX512BW-FAST-NEXT: vpshufb %zmm15, %zmm9, %zmm9
; AVX512BW-FAST-NEXT: vpermq {{.*#+}} zmm9 = zmm9[2,2,2,3,6,6,6,7]
; AVX512BW-FAST-NEXT: vmovdqu16 %zmm9, %zmm7 {%k2}
; AVX512BW-FAST-NEXT: vmovdqa (%r9), %ymm9
; AVX512BW-FAST-NEXT: vinserti64x4 $1, %ymm9, %zmm10, %zmm10
; AVX512BW-FAST-NEXT: vpshufb %zmm16, %zmm10, %zmm10
; AVX512BW-FAST-NEXT: vpermq {{.*#+}} zmm10 = zmm10[2,2,2,3,6,6,6,7]
; AVX512BW-FAST-NEXT: vmovdqu8 %zmm10, %zmm7 {%k3}
; AVX512BW-FAST-NEXT: vmovdqa64 (%rcx), %xmm21
; AVX512BW-FAST-NEXT: vmovdqa 32(%rcx), %xmm12
; AVX512BW-FAST-NEXT: vpbroadcastq {{.*#+}} xmm23 = [5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10]
; AVX512BW-FAST-NEXT: vpshufb %xmm23, %xmm12, %xmm10
; AVX512BW-FAST-NEXT: vmovdqa64 (%rdx), %xmm22
; AVX512BW-FAST-NEXT: vmovdqa 32(%rdx), %xmm14
; AVX512BW-FAST-NEXT: vpshufb %xmm23, %xmm14, %xmm15
; AVX512BW-FAST-NEXT: vpunpcklbw {{.*#+}} xmm10 = xmm15[0],xmm10[0],xmm15[1],xmm10[1],xmm15[2],xmm10[2],xmm15[3],xmm10[3],xmm15[4],xmm10[4],xmm15[5],xmm10[5],xmm15[6],xmm10[6],xmm15[7],xmm10[7]
; AVX512BW-FAST-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,0,0,1]
; AVX512BW-FAST-NEXT: vpunpcklbw {{.*#+}} xmm15 = xmm14[0],xmm12[0],xmm14[1],xmm12[1],xmm14[2],xmm12[2],xmm14[3],xmm12[3],xmm14[4],xmm12[4],xmm14[5],xmm12[5],xmm14[6],xmm12[6],xmm14[7],xmm12[7]
; AVX512BW-FAST-NEXT: vmovdqa64 {{.*#+}} ymm24 = [1,0,3,2,1,0,3,2,1,0,3,2,5,4,7,6]
; AVX512BW-FAST-NEXT: vpermw %ymm15, %ymm24, %ymm15
; AVX512BW-FAST-NEXT: vinserti64x4 $1, %ymm10, %zmm15, %zmm16
; AVX512BW-FAST-NEXT: vmovdqa 32(%rsi), %xmm15
; AVX512BW-FAST-NEXT: vpbroadcastq {{.*#+}} xmm25 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
; AVX512BW-FAST-NEXT: vpshufb %xmm25, %xmm15, %xmm10
; AVX512BW-FAST-NEXT: vmovdqa64 32(%rdi), %xmm19
; AVX512BW-FAST-NEXT: vpshufb %xmm25, %xmm19, %xmm20
; AVX512BW-FAST-NEXT: vpunpckhbw {{.*#+}} xmm10 = xmm20[8],xmm10[8],xmm20[9],xmm10[9],xmm20[10],xmm10[10],xmm20[11],xmm10[11],xmm20[12],xmm10[12],xmm20[13],xmm10[13],xmm20[14],xmm10[14],xmm20[15],xmm10[15]
; AVX512BW-FAST-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,0,0,1]
; AVX512BW-FAST-NEXT: vpunpcklbw {{.*#+}} xmm20 = xmm19[0],xmm15[0],xmm19[1],xmm15[1],xmm19[2],xmm15[2],xmm19[3],xmm15[3],xmm19[4],xmm15[4],xmm19[5],xmm15[5],xmm19[6],xmm15[6],xmm19[7],xmm15[7]
; AVX512BW-FAST-NEXT: vmovdqa64 {{.*#+}} ymm26 = [0,3,2,1,0,3,2,1,0,3,2,1,4,5,6,5]
; AVX512BW-FAST-NEXT: vpermw %ymm20, %ymm26, %ymm20
; AVX512BW-FAST-NEXT: vinserti64x4 $1, %ymm10, %zmm20, %zmm10
; AVX512BW-FAST-NEXT: vmovdqu16 %zmm16, %zmm10 {%k2}
; AVX512BW-FAST-NEXT: vmovdqa64 32(%r8), %xmm16
; AVX512BW-FAST-NEXT: vmovdqa64 {{.*#+}} xmm27 = <8,u,9,u,u,u,u,u,u,u,5,u,6,u,7,u>
; AVX512BW-FAST-NEXT: vpshufb %xmm27, %xmm16, %xmm20
; AVX512BW-FAST-NEXT: vpmovzxbw {{.*#+}} xmm28 = xmm16[0],zero,xmm16[1],zero,xmm16[2],zero,xmm16[3],zero,xmm16[4],zero,xmm16[5],zero,xmm16[6],zero,xmm16[7],zero
; AVX512BW-FAST-NEXT: vinserti64x4 $1, %ymm20, %zmm28, %zmm20
; AVX512BW-FAST-NEXT: vmovdqa64 {{.*#+}} zmm28 = [2,1,0,3,2,1,0,3,2,1,0,3,4,4,4,4,22,21,16,23,22,21,16,23,22,21,16,23,17,17,17,17]
; AVX512BW-FAST-NEXT: vpermw %zmm20, %zmm28, %zmm10 {%k1}
; AVX512BW-FAST-NEXT: vmovdqa64 32(%r9), %xmm20
; AVX512BW-FAST-NEXT: vmovdqa64 {{.*#+}} xmm29 = <u,8,u,9,u,10,u,11,u,4,u,5,u,6,u,7>
; AVX512BW-FAST-NEXT: vpshufb %xmm29, %xmm20, %xmm30
; AVX512BW-FAST-NEXT: vpunpcklbw {{.*#+}} xmm31 = xmm20[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
; AVX512BW-FAST-NEXT: vinserti32x4 $2, %xmm30, %zmm31, %zmm30
; AVX512BW-FAST-NEXT: vmovdqa64 (%rsi), %xmm31
; AVX512BW-FAST-NEXT: vpermw %zmm30, %zmm28, %zmm30
; AVX512BW-FAST-NEXT: movabsq $585610922974906400, %rax # imm = 0x820820820820820
; AVX512BW-FAST-NEXT: kmovq %rax, %k3
; AVX512BW-FAST-NEXT: vmovdqu8 %zmm30, %zmm10 {%k3}
; AVX512BW-FAST-NEXT: vpshufb %xmm23, %xmm21, %xmm30
; AVX512BW-FAST-NEXT: vpshufb %xmm23, %xmm22, %xmm23
; AVX512BW-FAST-NEXT: vpunpcklbw {{.*#+}} xmm23 = xmm23[0],xmm30[0],xmm23[1],xmm30[1],xmm23[2],xmm30[2],xmm23[3],xmm30[3],xmm23[4],xmm30[4],xmm23[5],xmm30[5],xmm23[6],xmm30[6],xmm23[7],xmm30[7]
; AVX512BW-FAST-NEXT: vpunpcklbw {{.*#+}} xmm30 = xmm22[0],xmm21[0],xmm22[1],xmm21[1],xmm22[2],xmm21[2],xmm22[3],xmm21[3],xmm22[4],xmm21[4],xmm22[5],xmm21[5],xmm22[6],xmm21[6],xmm22[7],xmm21[7]
; AVX512BW-FAST-NEXT: vpermw %ymm30, %ymm24, %ymm24
; AVX512BW-FAST-NEXT: vmovdqa64 (%rdi), %xmm30
; AVX512BW-FAST-NEXT: vpermq {{.*#+}} ymm23 = ymm23[0,0,0,1]
; AVX512BW-FAST-NEXT: vinserti64x4 $1, %ymm23, %zmm24, %zmm24
; AVX512BW-FAST-NEXT: vpshufb %xmm25, %xmm31, %xmm23
; AVX512BW-FAST-NEXT: vpshufb %xmm25, %xmm30, %xmm25
; AVX512BW-FAST-NEXT: vpunpckhbw {{.*#+}} xmm23 = xmm25[8],xmm23[8],xmm25[9],xmm23[9],xmm25[10],xmm23[10],xmm25[11],xmm23[11],xmm25[12],xmm23[12],xmm25[13],xmm23[13],xmm25[14],xmm23[14],xmm25[15],xmm23[15]
; AVX512BW-FAST-NEXT: vpunpcklbw {{.*#+}} xmm25 = xmm30[0],xmm31[0],xmm30[1],xmm31[1],xmm30[2],xmm31[2],xmm30[3],xmm31[3],xmm30[4],xmm31[4],xmm30[5],xmm31[5],xmm30[6],xmm31[6],xmm30[7],xmm31[7]
; AVX512BW-FAST-NEXT: vpermw %ymm25, %ymm26, %ymm25
; AVX512BW-FAST-NEXT: vpermq {{.*#+}} ymm23 = ymm23[0,0,0,1]
; AVX512BW-FAST-NEXT: vinserti64x4 $1, %ymm23, %zmm25, %zmm23
; AVX512BW-FAST-NEXT: vmovdqa64 (%r8), %xmm25
; AVX512BW-FAST-NEXT: vmovdqu16 %zmm24, %zmm23 {%k2}
; AVX512BW-FAST-NEXT: vpshufb %xmm27, %xmm25, %xmm24
; AVX512BW-FAST-NEXT: vpmovzxbw {{.*#+}} xmm26 = xmm25[0],zero,xmm25[1],zero,xmm25[2],zero,xmm25[3],zero,xmm25[4],zero,xmm25[5],zero,xmm25[6],zero,xmm25[7],zero
; AVX512BW-FAST-NEXT: vinserti64x4 $1, %ymm24, %zmm26, %zmm24
; AVX512BW-FAST-NEXT: vmovdqa64 (%r9), %xmm26
; AVX512BW-FAST-NEXT: vpermw %zmm24, %zmm28, %zmm23 {%k1}
; AVX512BW-FAST-NEXT: vpshufb %xmm29, %xmm26, %xmm24
; AVX512BW-FAST-NEXT: vpunpcklbw {{.*#+}} xmm27 = xmm26[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
; AVX512BW-FAST-NEXT: vinserti32x4 $2, %xmm24, %zmm27, %zmm24
; AVX512BW-FAST-NEXT: vpermw %zmm24, %zmm28, %zmm24
; AVX512BW-FAST-NEXT: vmovdqu8 %zmm24, %zmm23 {%k3}
; AVX512BW-FAST-NEXT: vpunpcklbw {{.*#+}} ymm17 = ymm18[0],ymm17[0],ymm18[1],ymm17[1],ymm18[2],ymm17[2],ymm18[3],ymm17[3],ymm18[4],ymm17[4],ymm18[5],ymm17[5],ymm18[6],ymm17[6],ymm18[7],ymm17[7],ymm18[16],ymm17[16],ymm18[17],ymm17[17],ymm18[18],ymm17[18],ymm18[19],ymm17[19],ymm18[20],ymm17[20],ymm18[21],ymm17[21],ymm18[22],ymm17[22],ymm18[23],ymm17[23]
; AVX512BW-FAST-NEXT: vpunpckhbw {{.*#+}} xmm18 = xmm22[8],xmm21[8],xmm22[9],xmm21[9],xmm22[10],xmm21[10],xmm22[11],xmm21[11],xmm22[12],xmm21[12],xmm22[13],xmm21[13],xmm22[14],xmm21[14],xmm22[15],xmm21[15]
; AVX512BW-FAST-NEXT: vinserti64x4 $1, %ymm17, %zmm18, %zmm17
; AVX512BW-FAST-NEXT: vpunpcklbw {{.*#+}} ymm11 = ymm13[0],ymm11[0],ymm13[1],ymm11[1],ymm13[2],ymm11[2],ymm13[3],ymm11[3],ymm13[4],ymm11[4],ymm13[5],ymm11[5],ymm13[6],ymm11[6],ymm13[7],ymm11[7],ymm13[16],ymm11[16],ymm13[17],ymm11[17],ymm13[18],ymm11[18],ymm13[19],ymm11[19],ymm13[20],ymm11[20],ymm13[21],ymm11[21],ymm13[22],ymm11[22],ymm13[23],ymm11[23]
; AVX512BW-FAST-NEXT: vpunpckhbw {{.*#+}} xmm13 = xmm30[8],xmm31[8],xmm30[9],xmm31[9],xmm30[10],xmm31[10],xmm30[11],xmm31[11],xmm30[12],xmm31[12],xmm30[13],xmm31[13],xmm30[14],xmm31[14],xmm30[15],xmm31[15]
; AVX512BW-FAST-NEXT: vinserti64x4 $1, %ymm11, %zmm13, %zmm11
; AVX512BW-FAST-NEXT: vmovdqa64 {{.*#+}} zmm13 = [4,3,6,5,4,3,6,5,4,3,6,5,7,7,7,7,24,27,26,25,24,27,26,25,24,27,26,25,28,29,30,29]
; AVX512BW-FAST-NEXT: vpermw %zmm11, %zmm13, %zmm11
; AVX512BW-FAST-NEXT: vmovdqa64 {{.*#+}} zmm18 = [5,4,3,6,5,4,3,6,5,4,3,6,7,7,7,7,25,24,27,26,25,24,27,26,25,24,27,26,29,28,31,30]
; AVX512BW-FAST-NEXT: vpermw %zmm17, %zmm18, %zmm11 {%k1}
; AVX512BW-FAST-NEXT: vpunpcklbw {{.*#+}} ymm5 = ymm6[0],ymm5[0],ymm6[1],ymm5[1],ymm6[2],ymm5[2],ymm6[3],ymm5[3],ymm6[4],ymm5[4],ymm6[5],ymm5[5],ymm6[6],ymm5[6],ymm6[7],ymm5[7],ymm6[16],ymm5[16],ymm6[17],ymm5[17],ymm6[18],ymm5[18],ymm6[19],ymm5[19],ymm6[20],ymm5[20],ymm6[21],ymm5[21],ymm6[22],ymm5[22],ymm6[23],ymm5[23]
; AVX512BW-FAST-NEXT: vpunpckhbw {{.*#+}} xmm6 = xmm14[8],xmm12[8],xmm14[9],xmm12[9],xmm14[10],xmm12[10],xmm14[11],xmm12[11],xmm14[12],xmm12[12],xmm14[13],xmm12[13],xmm14[14],xmm12[14],xmm14[15],xmm12[15]
; AVX512BW-FAST-NEXT: vinserti64x4 $1, %ymm5, %zmm6, %zmm5
; AVX512BW-FAST-NEXT: vpunpcklbw {{.*#+}} ymm1 = ymm2[0],ymm1[0],ymm2[1],ymm1[1],ymm2[2],ymm1[2],ymm2[3],ymm1[3],ymm2[4],ymm1[4],ymm2[5],ymm1[5],ymm2[6],ymm1[6],ymm2[7],ymm1[7],ymm2[16],ymm1[16],ymm2[17],ymm1[17],ymm2[18],ymm1[18],ymm2[19],ymm1[19],ymm2[20],ymm1[20],ymm2[21],ymm1[21],ymm2[22],ymm1[22],ymm2[23],ymm1[23]
; AVX512BW-FAST-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm19[8],xmm15[8],xmm19[9],xmm15[9],xmm19[10],xmm15[10],xmm19[11],xmm15[11],xmm19[12],xmm15[12],xmm19[13],xmm15[13],xmm19[14],xmm15[14],xmm19[15],xmm15[15]
; AVX512BW-FAST-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm1
; AVX512BW-FAST-NEXT: vpermw %zmm1, %zmm13, %zmm1
; AVX512BW-FAST-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm25[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
; AVX512BW-FAST-NEXT: vpermw %zmm5, %zmm18, %zmm1 {%k1}
; AVX512BW-FAST-NEXT: vmovdqa {{.*#+}} ymm5 = [2,5,4,3,2,5,4,3,2,5,4,3,6,5,6,7]
; AVX512BW-FAST-NEXT: vpermw %ymm2, %ymm5, %ymm2
; AVX512BW-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm6 = [2,0,1,0,0,0,3,0,0,0,0,0,4,0,0,0,2,0,1,0,0,0,3,0,0,0,0,0,4,0,0,0]
; AVX512BW-FAST-NEXT: # ymm6 = mem[0,1,0,1]
; AVX512BW-FAST-NEXT: vpshufb %ymm6, %ymm8, %ymm8
; AVX512BW-FAST-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,2,3]
; AVX512BW-FAST-NEXT: vinserti64x4 $1, %ymm8, %zmm2, %zmm2
; AVX512BW-FAST-NEXT: movl $1227133513, %eax # imm = 0x49249249
; AVX512BW-FAST-NEXT: kmovd %eax, %k1
; AVX512BW-FAST-NEXT: vmovdqu16 %zmm2, %zmm11 {%k1}
; AVX512BW-FAST-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm26[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
; AVX512BW-FAST-NEXT: vpermw %ymm2, %ymm5, %ymm2
; AVX512BW-FAST-NEXT: vbroadcasti128 {{.*#+}} ymm8 = [0,2,0,1,0,0,0,3,0,0,0,0,0,4,0,0,0,2,0,1,0,0,0,3,0,0,0,0,0,4,0,0]
; AVX512BW-FAST-NEXT: # ymm8 = mem[0,1,0,1]
; AVX512BW-FAST-NEXT: vpshufb %ymm8, %ymm9, %ymm9
; AVX512BW-FAST-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,2,3]
; AVX512BW-FAST-NEXT: vinserti64x4 $1, %ymm9, %zmm2, %zmm2
; AVX512BW-FAST-NEXT: movabsq $2342443691899625602, %rax # imm = 0x2082082082082082
; AVX512BW-FAST-NEXT: kmovq %rax, %k2
; AVX512BW-FAST-NEXT: vmovdqu8 %zmm2, %zmm11 {%k2}
; AVX512BW-FAST-NEXT: vpshufb %ymm6, %ymm3, %ymm2
; AVX512BW-FAST-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm16[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
; AVX512BW-FAST-NEXT: vpermw %ymm3, %ymm5, %ymm3
; AVX512BW-FAST-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3]
; AVX512BW-FAST-NEXT: vinserti64x4 $1, %ymm2, %zmm3, %zmm2
; AVX512BW-FAST-NEXT: vmovdqu16 %zmm2, %zmm1 {%k1}
; AVX512BW-FAST-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm20[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
; AVX512BW-FAST-NEXT: vpermw %ymm2, %ymm5, %ymm2
; AVX512BW-FAST-NEXT: vpshufb %ymm8, %ymm4, %ymm3
; AVX512BW-FAST-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3]
; AVX512BW-FAST-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm2
; AVX512BW-FAST-NEXT: vmovdqu8 %zmm2, %zmm1 {%k2}
; AVX512BW-FAST-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512BW-FAST-NEXT: vmovdqa64 %zmm1, 256(%rax)
; AVX512BW-FAST-NEXT: vmovdqa64 %zmm11, 64(%rax)
; AVX512BW-FAST-NEXT: vmovdqa64 %zmm23, (%rax)
; AVX512BW-FAST-NEXT: vmovdqa64 %zmm10, 192(%rax)
; AVX512BW-FAST-NEXT: vmovdqa64 %zmm7, 128(%rax)
; AVX512BW-FAST-NEXT: vmovdqa64 %zmm0, 320(%rax)
; AVX512BW-FAST-NEXT: vzeroupper
; AVX512BW-FAST-NEXT: retq
;
; AVX512DQBW-SLOW-LABEL: store_i8_stride6_vf64:
; AVX512DQBW-SLOW: # %bb.0:
; AVX512DQBW-SLOW-NEXT: movq {{[0-9]+}}(%rsp), %rax
; AVX512DQBW-SLOW-NEXT: vmovdqa64 (%r8), %zmm14
; AVX512DQBW-SLOW-NEXT: vmovdqa64 (%r9), %zmm12
; AVX512DQBW-SLOW-NEXT: vmovdqa64 (%rsi), %ymm16
; AVX512DQBW-SLOW-NEXT: vmovdqa64 (%rdi), %ymm17
; AVX512DQBW-SLOW-NEXT: vpunpcklbw {{.*#+}} ymm0 = ymm17[0],ymm16[0],ymm17[1],ymm16[1],ymm17[2],ymm16[2],ymm17[3],ymm16[3],ymm17[4],ymm16[4],ymm17[5],ymm16[5],ymm17[6],ymm16[6],ymm17[7],ymm16[7],ymm17[16],ymm16[16],ymm17[17],ymm16[17],ymm17[18],ymm16[18],ymm17[19],ymm16[19],ymm17[20],ymm16[20],ymm17[21],ymm16[21],ymm17[22],ymm16[22],ymm17[23],ymm16[23]
; AVX512DQBW-SLOW-NEXT: vmovdqa (%rsi), %xmm1
; AVX512DQBW-SLOW-NEXT: vmovdqa 32(%rsi), %xmm9
; AVX512DQBW-SLOW-NEXT: vmovdqa (%rdi), %xmm3
; AVX512DQBW-SLOW-NEXT: vmovdqa 32(%rdi), %xmm11
; AVX512DQBW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm3[8],xmm1[8],xmm3[9],xmm1[9],xmm3[10],xmm1[10],xmm3[11],xmm1[11],xmm3[12],xmm1[12],xmm3[13],xmm1[13],xmm3[14],xmm1[14],xmm3[15],xmm1[15]
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $1, %ymm0, %zmm2, %zmm0
; AVX512DQBW-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm7 = [4,3,6,5,4,3,6,5,4,3,6,5,7,7,7,7,24,27,26,25,24,27,26,25,24,27,26,25,28,29,30,29]
; AVX512DQBW-SLOW-NEXT: vpermw %zmm0, %zmm7, %zmm0
; AVX512DQBW-SLOW-NEXT: vmovdqa (%rcx), %xmm2
; AVX512DQBW-SLOW-NEXT: vmovdqa 32(%rcx), %xmm8
; AVX512DQBW-SLOW-NEXT: vmovdqa (%rdx), %xmm4
; AVX512DQBW-SLOW-NEXT: vmovdqa 32(%rdx), %xmm10
; AVX512DQBW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm5 = xmm4[8],xmm2[8],xmm4[9],xmm2[9],xmm4[10],xmm2[10],xmm4[11],xmm2[11],xmm4[12],xmm2[12],xmm4[13],xmm2[13],xmm4[14],xmm2[14],xmm4[15],xmm2[15]
; AVX512DQBW-SLOW-NEXT: vmovdqa64 {{.*#+}} ymm20 = [5,4,3,6,5,4,3,6,5,4,3,6,7,7,7,7]
; AVX512DQBW-SLOW-NEXT: vpermw %ymm5, %ymm20, %ymm5
; AVX512DQBW-SLOW-NEXT: vmovdqa64 (%rcx), %ymm18
; AVX512DQBW-SLOW-NEXT: vmovdqa64 (%rdx), %ymm19
; AVX512DQBW-SLOW-NEXT: vpunpcklbw {{.*#+}} ymm6 = ymm19[0],ymm18[0],ymm19[1],ymm18[1],ymm19[2],ymm18[2],ymm19[3],ymm18[3],ymm19[4],ymm18[4],ymm19[5],ymm18[5],ymm19[6],ymm18[6],ymm19[7],ymm18[7],ymm19[16],ymm18[16],ymm19[17],ymm18[17],ymm19[18],ymm18[18],ymm19[19],ymm18[19],ymm19[20],ymm18[20],ymm19[21],ymm18[21],ymm19[22],ymm18[22],ymm19[23],ymm18[23]
; AVX512DQBW-SLOW-NEXT: vprold $16, %ymm6, %ymm6
; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,2,2,3]
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $1, %ymm6, %zmm5, %zmm5
; AVX512DQBW-SLOW-NEXT: movl $613566756, %r10d # imm = 0x24924924
; AVX512DQBW-SLOW-NEXT: kmovd %r10d, %k1
; AVX512DQBW-SLOW-NEXT: vmovdqu16 %zmm5, %zmm0 {%k1}
; AVX512DQBW-SLOW-NEXT: vmovdqa (%r8), %xmm5
; AVX512DQBW-SLOW-NEXT: vmovdqa 32(%r8), %xmm13
; AVX512DQBW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm6 = xmm5[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
; AVX512DQBW-SLOW-NEXT: vmovdqa64 {{.*#+}} ymm23 = [2,5,4,3,2,5,4,3,2,5,4,3,6,5,6,7]
; AVX512DQBW-SLOW-NEXT: vpermw %ymm6, %ymm23, %ymm6
; AVX512DQBW-SLOW-NEXT: vmovdqa64 (%r8), %ymm21
; AVX512DQBW-SLOW-NEXT: vbroadcasti64x2 {{.*#+}} ymm24 = [2,0,1,0,0,0,3,0,0,0,0,0,4,0,0,0,2,0,1,0,0,0,3,0,0,0,0,0,4,0,0,0]
; AVX512DQBW-SLOW-NEXT: # ymm24 = mem[0,1,0,1]
; AVX512DQBW-SLOW-NEXT: vpshufb %ymm24, %ymm21, %ymm15
; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,2,2,3]
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $1, %ymm15, %zmm6, %zmm6
; AVX512DQBW-SLOW-NEXT: movl $1227133513, %r10d # imm = 0x49249249
; AVX512DQBW-SLOW-NEXT: kmovd %r10d, %k2
; AVX512DQBW-SLOW-NEXT: vmovdqu16 %zmm6, %zmm0 {%k2}
; AVX512DQBW-SLOW-NEXT: vmovdqa (%r9), %xmm6
; AVX512DQBW-SLOW-NEXT: vmovdqa 32(%r9), %xmm15
; AVX512DQBW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm22 = xmm6[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
; AVX512DQBW-SLOW-NEXT: vpermw %ymm22, %ymm23, %ymm25
; AVX512DQBW-SLOW-NEXT: vmovdqa64 (%r9), %ymm22
; AVX512DQBW-SLOW-NEXT: vbroadcasti64x2 {{.*#+}} ymm26 = [0,2,0,1,0,0,0,3,0,0,0,0,0,4,0,0,0,2,0,1,0,0,0,3,0,0,0,0,0,4,0,0]
; AVX512DQBW-SLOW-NEXT: # ymm26 = mem[0,1,0,1]
; AVX512DQBW-SLOW-NEXT: vpshufb %ymm26, %ymm22, %ymm27
; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} ymm27 = ymm27[2,2,2,3]
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $1, %ymm27, %zmm25, %zmm25
; AVX512DQBW-SLOW-NEXT: movabsq $2342443691899625602, %r10 # imm = 0x2082082082082082
; AVX512DQBW-SLOW-NEXT: kmovq %r10, %k3
; AVX512DQBW-SLOW-NEXT: vmovdqu8 %zmm25, %zmm0 {%k3}
; AVX512DQBW-SLOW-NEXT: vmovdqa64 32(%rsi), %ymm25
; AVX512DQBW-SLOW-NEXT: vmovdqa64 32(%rdi), %ymm27
; AVX512DQBW-SLOW-NEXT: vpunpcklbw {{.*#+}} ymm28 = ymm27[0],ymm25[0],ymm27[1],ymm25[1],ymm27[2],ymm25[2],ymm27[3],ymm25[3],ymm27[4],ymm25[4],ymm27[5],ymm25[5],ymm27[6],ymm25[6],ymm27[7],ymm25[7],ymm27[16],ymm25[16],ymm27[17],ymm25[17],ymm27[18],ymm25[18],ymm27[19],ymm25[19],ymm27[20],ymm25[20],ymm27[21],ymm25[21],ymm27[22],ymm25[22],ymm27[23],ymm25[23]
; AVX512DQBW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm29 = xmm11[8],xmm9[8],xmm11[9],xmm9[9],xmm11[10],xmm9[10],xmm11[11],xmm9[11],xmm11[12],xmm9[12],xmm11[13],xmm9[13],xmm11[14],xmm9[14],xmm11[15],xmm9[15]
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $1, %ymm28, %zmm29, %zmm28
; AVX512DQBW-SLOW-NEXT: vpermw %zmm28, %zmm7, %zmm7
; AVX512DQBW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm28 = xmm10[8],xmm8[8],xmm10[9],xmm8[9],xmm10[10],xmm8[10],xmm10[11],xmm8[11],xmm10[12],xmm8[12],xmm10[13],xmm8[13],xmm10[14],xmm8[14],xmm10[15],xmm8[15]
; AVX512DQBW-SLOW-NEXT: vpermw %ymm28, %ymm20, %ymm20
; AVX512DQBW-SLOW-NEXT: vmovdqa64 32(%rcx), %ymm28
; AVX512DQBW-SLOW-NEXT: vmovdqa64 32(%rdx), %ymm29
; AVX512DQBW-SLOW-NEXT: vpunpcklbw {{.*#+}} ymm30 = ymm29[0],ymm28[0],ymm29[1],ymm28[1],ymm29[2],ymm28[2],ymm29[3],ymm28[3],ymm29[4],ymm28[4],ymm29[5],ymm28[5],ymm29[6],ymm28[6],ymm29[7],ymm28[7],ymm29[16],ymm28[16],ymm29[17],ymm28[17],ymm29[18],ymm28[18],ymm29[19],ymm28[19],ymm29[20],ymm28[20],ymm29[21],ymm28[21],ymm29[22],ymm28[22],ymm29[23],ymm28[23]
; AVX512DQBW-SLOW-NEXT: vprold $16, %ymm30, %ymm30
; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} ymm30 = ymm30[2,2,2,3]
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $1, %ymm30, %zmm20, %zmm20
; AVX512DQBW-SLOW-NEXT: vmovdqu16 %zmm20, %zmm7 {%k1}
; AVX512DQBW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm20 = xmm13[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
; AVX512DQBW-SLOW-NEXT: vpermw %ymm20, %ymm23, %ymm20
; AVX512DQBW-SLOW-NEXT: vmovdqa64 32(%r8), %ymm30
; AVX512DQBW-SLOW-NEXT: vpshufb %ymm24, %ymm30, %ymm24
; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} ymm24 = ymm24[2,2,2,3]
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $1, %ymm24, %zmm20, %zmm20
; AVX512DQBW-SLOW-NEXT: vmovdqu16 %zmm20, %zmm7 {%k2}
; AVX512DQBW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm20 = xmm15[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
; AVX512DQBW-SLOW-NEXT: vpermw %ymm20, %ymm23, %ymm20
; AVX512DQBW-SLOW-NEXT: vmovdqa64 32(%r9), %ymm23
; AVX512DQBW-SLOW-NEXT: vpshufb %ymm26, %ymm23, %ymm24
; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} ymm24 = ymm24[2,2,2,3]
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $1, %ymm24, %zmm20, %zmm20
; AVX512DQBW-SLOW-NEXT: vmovdqu8 %zmm20, %zmm7 {%k3}
; AVX512DQBW-SLOW-NEXT: vpbroadcastq {{.*#+}} ymm24 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
; AVX512DQBW-SLOW-NEXT: vpshufb %ymm24, %ymm25, %ymm20
; AVX512DQBW-SLOW-NEXT: vpshufb %ymm24, %ymm27, %ymm26
; AVX512DQBW-SLOW-NEXT: vpunpcklbw {{.*#+}} ymm20 = ymm26[0],ymm20[0],ymm26[1],ymm20[1],ymm26[2],ymm20[2],ymm26[3],ymm20[3],ymm26[4],ymm20[4],ymm26[5],ymm20[5],ymm26[6],ymm20[6],ymm26[7],ymm20[7],ymm26[16],ymm20[16],ymm26[17],ymm20[17],ymm26[18],ymm20[18],ymm26[19],ymm20[19],ymm26[20],ymm20[20],ymm26[21],ymm20[21],ymm26[22],ymm20[22],ymm26[23],ymm20[23]
; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} ymm20 = ymm20[2,2,2,3]
; AVX512DQBW-SLOW-NEXT: vpunpckhbw {{.*#+}} ymm25 = ymm27[8],ymm25[8],ymm27[9],ymm25[9],ymm27[10],ymm25[10],ymm27[11],ymm25[11],ymm27[12],ymm25[12],ymm27[13],ymm25[13],ymm27[14],ymm25[14],ymm27[15],ymm25[15],ymm27[24],ymm25[24],ymm27[25],ymm25[25],ymm27[26],ymm25[26],ymm27[27],ymm25[27],ymm27[28],ymm25[28],ymm27[29],ymm25[29],ymm27[30],ymm25[30],ymm27[31],ymm25[31]
; AVX512DQBW-SLOW-NEXT: vmovdqa64 {{.*#+}} ymm26 = [12,11,14,13,12,11,14,13,12,11,14,13,15,15,15,15]
; AVX512DQBW-SLOW-NEXT: vpermw %ymm25, %ymm26, %ymm25
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $1, %ymm25, %zmm20, %zmm25
; AVX512DQBW-SLOW-NEXT: vpbroadcastq {{.*#+}} ymm27 = [5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10]
; AVX512DQBW-SLOW-NEXT: vpshufb %ymm27, %ymm28, %ymm20
; AVX512DQBW-SLOW-NEXT: vpshufb %ymm27, %ymm29, %ymm31
; AVX512DQBW-SLOW-NEXT: vpunpcklbw {{.*#+}} ymm20 = ymm31[0],ymm20[0],ymm31[1],ymm20[1],ymm31[2],ymm20[2],ymm31[3],ymm20[3],ymm31[4],ymm20[4],ymm31[5],ymm20[5],ymm31[6],ymm20[6],ymm31[7],ymm20[7],ymm31[16],ymm20[16],ymm31[17],ymm20[17],ymm31[18],ymm20[18],ymm31[19],ymm20[19],ymm31[20],ymm20[20],ymm31[21],ymm20[21],ymm31[22],ymm20[22],ymm31[23],ymm20[23]
; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} ymm20 = ymm20[2,2,2,3]
; AVX512DQBW-SLOW-NEXT: vpunpckhbw {{.*#+}} ymm28 = ymm29[8],ymm28[8],ymm29[9],ymm28[9],ymm29[10],ymm28[10],ymm29[11],ymm28[11],ymm29[12],ymm28[12],ymm29[13],ymm28[13],ymm29[14],ymm28[14],ymm29[15],ymm28[15],ymm29[24],ymm28[24],ymm29[25],ymm28[25],ymm29[26],ymm28[26],ymm29[27],ymm28[27],ymm29[28],ymm28[28],ymm29[29],ymm28[29],ymm29[30],ymm28[30],ymm29[31],ymm28[31]
; AVX512DQBW-SLOW-NEXT: vmovdqa64 {{.*#+}} ymm29 = [13,12,11,14,13,12,11,14,13,12,11,14,15,15,15,15]
; AVX512DQBW-SLOW-NEXT: vpermw %ymm28, %ymm29, %ymm28
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $1, %ymm28, %zmm20, %zmm20
; AVX512DQBW-SLOW-NEXT: vmovdqu16 %zmm25, %zmm20 {%k1}
; AVX512DQBW-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm25 = zmm30[0,1,2,3],zmm14[4,5,6,7]
; AVX512DQBW-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm28 = <u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,6,u,5,u,8,u,7,u,u,u,9,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,10,u,13,u,12,u,11,u,14,u,u,u,u,u,15,u>
; AVX512DQBW-SLOW-NEXT: vpshufb %zmm28, %zmm25, %zmm25
; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} zmm25 = zmm25[2,2,2,3,6,6,6,7]
; AVX512DQBW-SLOW-NEXT: movl $-1840700270, %ecx # imm = 0x92492492
; AVX512DQBW-SLOW-NEXT: kmovd %ecx, %k2
; AVX512DQBW-SLOW-NEXT: vmovdqu16 %zmm25, %zmm20 {%k2}
; AVX512DQBW-SLOW-NEXT: vshufi64x2 {{.*#+}} zmm23 = zmm23[0,1,2,3],zmm12[4,5,6,7]
; AVX512DQBW-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm25 = <u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,6,u,5,u,8,u,7,u,u,u,9,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,10,u,13,u,12,u,11,u,14,u,u,u,u,u,15>
; AVX512DQBW-SLOW-NEXT: vpshufb %zmm25, %zmm23, %zmm23
; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} zmm23 = zmm23[2,2,2,3,6,6,6,7]
; AVX512DQBW-SLOW-NEXT: movabsq $-9076969306111049208, %rcx # imm = 0x8208208208208208
; AVX512DQBW-SLOW-NEXT: kmovq %rcx, %k3
; AVX512DQBW-SLOW-NEXT: vmovdqu8 %zmm23, %zmm20 {%k3}
; AVX512DQBW-SLOW-NEXT: vpshufb %ymm24, %ymm16, %ymm23
; AVX512DQBW-SLOW-NEXT: vpshufb %ymm24, %ymm17, %ymm24
; AVX512DQBW-SLOW-NEXT: vpunpcklbw {{.*#+}} ymm23 = ymm24[0],ymm23[0],ymm24[1],ymm23[1],ymm24[2],ymm23[2],ymm24[3],ymm23[3],ymm24[4],ymm23[4],ymm24[5],ymm23[5],ymm24[6],ymm23[6],ymm24[7],ymm23[7],ymm24[16],ymm23[16],ymm24[17],ymm23[17],ymm24[18],ymm23[18],ymm24[19],ymm23[19],ymm24[20],ymm23[20],ymm24[21],ymm23[21],ymm24[22],ymm23[22],ymm24[23],ymm23[23]
; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} ymm23 = ymm23[2,2,2,3]
; AVX512DQBW-SLOW-NEXT: vpunpckhbw {{.*#+}} ymm16 = ymm17[8],ymm16[8],ymm17[9],ymm16[9],ymm17[10],ymm16[10],ymm17[11],ymm16[11],ymm17[12],ymm16[12],ymm17[13],ymm16[13],ymm17[14],ymm16[14],ymm17[15],ymm16[15],ymm17[24],ymm16[24],ymm17[25],ymm16[25],ymm17[26],ymm16[26],ymm17[27],ymm16[27],ymm17[28],ymm16[28],ymm17[29],ymm16[29],ymm17[30],ymm16[30],ymm17[31],ymm16[31]
; AVX512DQBW-SLOW-NEXT: vpermw %ymm16, %ymm26, %ymm16
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $1, %ymm16, %zmm23, %zmm17
; AVX512DQBW-SLOW-NEXT: vpshufb %ymm27, %ymm18, %ymm16
; AVX512DQBW-SLOW-NEXT: vpshufb %ymm27, %ymm19, %ymm23
; AVX512DQBW-SLOW-NEXT: vpunpcklbw {{.*#+}} ymm16 = ymm23[0],ymm16[0],ymm23[1],ymm16[1],ymm23[2],ymm16[2],ymm23[3],ymm16[3],ymm23[4],ymm16[4],ymm23[5],ymm16[5],ymm23[6],ymm16[6],ymm23[7],ymm16[7],ymm23[16],ymm16[16],ymm23[17],ymm16[17],ymm23[18],ymm16[18],ymm23[19],ymm16[19],ymm23[20],ymm16[20],ymm23[21],ymm16[21],ymm23[22],ymm16[22],ymm23[23],ymm16[23]
; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} ymm16 = ymm16[2,2,2,3]
; AVX512DQBW-SLOW-NEXT: vpunpckhbw {{.*#+}} ymm18 = ymm19[8],ymm18[8],ymm19[9],ymm18[9],ymm19[10],ymm18[10],ymm19[11],ymm18[11],ymm19[12],ymm18[12],ymm19[13],ymm18[13],ymm19[14],ymm18[14],ymm19[15],ymm18[15],ymm19[24],ymm18[24],ymm19[25],ymm18[25],ymm19[26],ymm18[26],ymm19[27],ymm18[27],ymm19[28],ymm18[28],ymm19[29],ymm18[29],ymm19[30],ymm18[30],ymm19[31],ymm18[31]
; AVX512DQBW-SLOW-NEXT: vpermw %ymm18, %ymm29, %ymm18
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $1, %ymm18, %zmm16, %zmm16
; AVX512DQBW-SLOW-NEXT: vmovdqu16 %zmm17, %zmm16 {%k1}
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $1, %ymm21, %zmm14, %zmm14
; AVX512DQBW-SLOW-NEXT: vpshufb %zmm28, %zmm14, %zmm14
; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} zmm14 = zmm14[2,2,2,3,6,6,6,7]
; AVX512DQBW-SLOW-NEXT: vmovdqu16 %zmm14, %zmm16 {%k2}
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $1, %ymm22, %zmm12, %zmm12
; AVX512DQBW-SLOW-NEXT: vpshufb %zmm25, %zmm12, %zmm12
; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} zmm12 = zmm12[2,2,2,3,6,6,6,7]
; AVX512DQBW-SLOW-NEXT: vmovdqu8 %zmm12, %zmm16 {%k3}
; AVX512DQBW-SLOW-NEXT: vpbroadcastq {{.*#+}} xmm12 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0]
; AVX512DQBW-SLOW-NEXT: vpshufb %xmm12, %xmm9, %xmm14
; AVX512DQBW-SLOW-NEXT: vpshufb %xmm12, %xmm11, %xmm17
; AVX512DQBW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm14 = xmm17[8],xmm14[8],xmm17[9],xmm14[9],xmm17[10],xmm14[10],xmm17[11],xmm14[11],xmm17[12],xmm14[12],xmm17[13],xmm14[13],xmm17[14],xmm14[14],xmm17[15],xmm14[15]
; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} ymm14 = ymm14[0,0,0,1]
; AVX512DQBW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm9 = xmm11[0],xmm9[0],xmm11[1],xmm9[1],xmm11[2],xmm9[2],xmm11[3],xmm9[3],xmm11[4],xmm9[4],xmm11[5],xmm9[5],xmm11[6],xmm9[6],xmm11[7],xmm9[7]
; AVX512DQBW-SLOW-NEXT: vmovdqa {{.*#+}} ymm11 = [0,3,2,1,0,3,2,1,0,3,2,1,4,5,6,5]
; AVX512DQBW-SLOW-NEXT: vpermw %ymm9, %ymm11, %ymm9
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $1, %ymm14, %zmm9, %zmm9
; AVX512DQBW-SLOW-NEXT: vpbroadcastq {{.*#+}} xmm14 = [5,8,7,6,9,0,0,10,5,8,7,6,9,0,0,10]
; AVX512DQBW-SLOW-NEXT: vpshufb %xmm14, %xmm8, %xmm17
; AVX512DQBW-SLOW-NEXT: vpshufb %xmm14, %xmm10, %xmm18
; AVX512DQBW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm17 = xmm18[0],xmm17[0],xmm18[1],xmm17[1],xmm18[2],xmm17[2],xmm18[3],xmm17[3],xmm18[4],xmm17[4],xmm18[5],xmm17[5],xmm18[6],xmm17[6],xmm18[7],xmm17[7]
; AVX512DQBW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm8 = xmm10[0],xmm8[0],xmm10[1],xmm8[1],xmm10[2],xmm8[2],xmm10[3],xmm8[3],xmm10[4],xmm8[4],xmm10[5],xmm8[5],xmm10[6],xmm8[6],xmm10[7],xmm8[7]
; AVX512DQBW-SLOW-NEXT: vprold $16, %xmm8, %xmm8
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $1, %ymm17, %zmm8, %zmm8
; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} zmm8 = zmm8[0,0,0,1,4,4,4,5]
; AVX512DQBW-SLOW-NEXT: vmovdqu16 %zmm8, %zmm9 {%k2}
; AVX512DQBW-SLOW-NEXT: vpmovzxbw {{.*#+}} xmm8 = xmm13[0],zero,xmm13[1],zero,xmm13[2],zero,xmm13[3],zero,xmm13[4],zero,xmm13[5],zero,xmm13[6],zero,xmm13[7],zero
; AVX512DQBW-SLOW-NEXT: vpshufd {{.*#+}} xmm10 = xmm13[2,1,2,3]
; AVX512DQBW-SLOW-NEXT: vpmovzxbw {{.*#+}} xmm10 = xmm10[0],zero,xmm10[1],zero,xmm10[2],zero,xmm10[3],zero,xmm10[4],zero,xmm10[5],zero,xmm10[6],zero,xmm10[7],zero
; AVX512DQBW-SLOW-NEXT: vinserti32x4 $2, %xmm10, %zmm8, %zmm8
; AVX512DQBW-SLOW-NEXT: vmovdqa64 {{.*#+}} zmm10 = [2,1,0,3,2,1,0,3,2,1,0,3,4,4,4,4,22,21,16,23,22,21,16,23,22,21,16,23,17,17,17,17]
; AVX512DQBW-SLOW-NEXT: vpermw %zmm8, %zmm10, %zmm9 {%k1}
; AVX512DQBW-SLOW-NEXT: vpshufd {{.*#+}} xmm8 = xmm15[2,1,2,3]
; AVX512DQBW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm8 = xmm8[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
; AVX512DQBW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm13 = xmm15[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
; AVX512DQBW-SLOW-NEXT: vinserti32x4 $2, %xmm8, %zmm13, %zmm8
; AVX512DQBW-SLOW-NEXT: vpermw %zmm8, %zmm10, %zmm8
; AVX512DQBW-SLOW-NEXT: movabsq $585610922974906400, %rcx # imm = 0x820820820820820
; AVX512DQBW-SLOW-NEXT: kmovq %rcx, %k3
; AVX512DQBW-SLOW-NEXT: vmovdqu8 %zmm8, %zmm9 {%k3}
; AVX512DQBW-SLOW-NEXT: vpshufb %xmm12, %xmm1, %xmm8
; AVX512DQBW-SLOW-NEXT: vpshufb %xmm12, %xmm3, %xmm12
; AVX512DQBW-SLOW-NEXT: vpunpckhbw {{.*#+}} xmm8 = xmm12[8],xmm8[8],xmm12[9],xmm8[9],xmm12[10],xmm8[10],xmm12[11],xmm8[11],xmm12[12],xmm8[12],xmm12[13],xmm8[13],xmm12[14],xmm8[14],xmm12[15],xmm8[15]
; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} ymm8 = ymm8[0,0,0,1]
; AVX512DQBW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3],xmm3[4],xmm1[4],xmm3[5],xmm1[5],xmm3[6],xmm1[6],xmm3[7],xmm1[7]
; AVX512DQBW-SLOW-NEXT: vpermw %ymm1, %ymm11, %ymm1
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $1, %ymm8, %zmm1, %zmm1
; AVX512DQBW-SLOW-NEXT: vpshufb %xmm14, %xmm2, %xmm3
; AVX512DQBW-SLOW-NEXT: vpshufb %xmm14, %xmm4, %xmm8
; AVX512DQBW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm8[0],xmm3[0],xmm8[1],xmm3[1],xmm8[2],xmm3[2],xmm8[3],xmm3[3],xmm8[4],xmm3[4],xmm8[5],xmm3[5],xmm8[6],xmm3[6],xmm8[7],xmm3[7]
; AVX512DQBW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm2 = 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]
; AVX512DQBW-SLOW-NEXT: vprold $16, %xmm2, %xmm2
; AVX512DQBW-SLOW-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm2
; AVX512DQBW-SLOW-NEXT: vpermq {{.*#+}} zmm2 = zmm2[0,0,0,1,4,4,4,5]
; AVX512DQBW-SLOW-NEXT: vmovdqu16 %zmm2, %zmm1 {%k2}
; AVX512DQBW-SLOW-NEXT: vpmovzxbw {{.*#+}} xmm2 = xmm5[0],zero,xmm5[1],zero,xmm5[2],zero,xmm5[3],zero,xmm5[4],zero,xmm5[5],zero,xmm5[6],zero,xmm5[7],zero
; AVX512DQBW-SLOW-NEXT: vpshufd {{.*#+}} xmm3 = xmm5[2,1,2,3]
; AVX512DQBW-SLOW-NEXT: vpmovzxbw {{.*#+}} xmm3 = xmm3[0],zero,xmm3[1],zero,xmm3[2],zero,xmm3[3],zero,xmm3[4],zero,xmm3[5],zero,xmm3[6],zero,xmm3[7],zero
; AVX512DQBW-SLOW-NEXT: vinserti32x4 $2, %xmm3, %zmm2, %zmm2
; AVX512DQBW-SLOW-NEXT: vpermw %zmm2, %zmm10, %zmm1 {%k1}
; AVX512DQBW-SLOW-NEXT: vpshufd {{.*#+}} xmm2 = xmm6[2,1,2,3]
; AVX512DQBW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm2[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
; AVX512DQBW-SLOW-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm6[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
; AVX512DQBW-SLOW-NEXT: vinserti32x4 $2, %xmm2, %zmm3, %zmm2
; AVX512DQBW-SLOW-NEXT: vpermw %zmm2, %zmm10, %zmm2
; AVX512DQBW-SLOW-NEXT: vmovdqu8 %zmm2, %zmm1 {%k3}
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm1, (%rax)
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm9, 192(%rax)
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm16, 128(%rax)
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm20, 320(%rax)
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm7, 256(%rax)
; AVX512DQBW-SLOW-NEXT: vmovdqa64 %zmm0, 64(%rax)
; AVX512DQBW-SLOW-NEXT: vzeroupper
; AVX512DQBW-SLOW-NEXT: retq
%in.vec0 = load <64 x i8>, ptr %in.vecptr0, align 64
%in.vec1 = load <64 x i8>, ptr %in.vecptr1, align 64
%in.vec2 = load <64 x i8>, ptr %in.vecptr2, align 64
%in.vec3 = load <64 x i8>, ptr %in.vecptr3, align 64
%in.vec4 = load <64 x i8>, ptr %in.vecptr4, align 64
%in.vec5 = load <64 x i8>, ptr %in.vecptr5, align 64
%1 = shufflevector <64 x i8> %in.vec0, <64 x i8> %in.vec1, <128 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 64, i32 65, i32 66, i32 67, i32 68, i32 69, i32 70, i32 71, i32 72, i32 73, i32 74, i32 75, i32 76, i32 77, i32 78, i32 79, i32 80, i32 81, i32 82, i32 83, i32 84, i32 85, i32 86, i32 87, i32 88, i32 89, i32 90, i32 91, i32 92, i32 93, i32 94, i32 95, i32 96, i32 97, i32 98, i32 99, i32 100, i32 101, i32 102, i32 103, i32 104, i32 105, i32 106, i32 107, i32 108, i32 109, i32 110, i32 111, i32 112, i32 113, i32 114, i32 115, i32 116, i32 117, i32 118, i32 119, i32 120, i32 121, i32 122, i32 123, i32 124, i32 125, i32 126, i32 127>
%2 = shufflevector <64 x i8> %in.vec2, <64 x i8> %in.vec3, <128 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 64, i32 65, i32 66, i32 67, i32 68, i32 69, i32 70, i32 71, i32 72, i32 73, i32 74, i32 75, i32 76, i32 77, i32 78, i32 79, i32 80, i32 81, i32 82, i32 83, i32 84, i32 85, i32 86, i32 87, i32 88, i32 89, i32 90, i32 91, i32 92, i32 93, i32 94, i32 95, i32 96, i32 97, i32 98, i32 99, i32 100, i32 101, i32 102, i32 103, i32 104, i32 105, i32 106, i32 107, i32 108, i32 109, i32 110, i32 111, i32 112, i32 113, i32 114, i32 115, i32 116, i32 117, i32 118, i32 119, i32 120, i32 121, i32 122, i32 123, i32 124, i32 125, i32 126, i32 127>
%3 = shufflevector <64 x i8> %in.vec4, <64 x i8> %in.vec5, <128 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 64, i32 65, i32 66, i32 67, i32 68, i32 69, i32 70, i32 71, i32 72, i32 73, i32 74, i32 75, i32 76, i32 77, i32 78, i32 79, i32 80, i32 81, i32 82, i32 83, i32 84, i32 85, i32 86, i32 87, i32 88, i32 89, i32 90, i32 91, i32 92, i32 93, i32 94, i32 95, i32 96, i32 97, i32 98, i32 99, i32 100, i32 101, i32 102, i32 103, i32 104, i32 105, i32 106, i32 107, i32 108, i32 109, i32 110, i32 111, i32 112, i32 113, i32 114, i32 115, i32 116, i32 117, i32 118, i32 119, i32 120, i32 121, i32 122, i32 123, i32 124, i32 125, i32 126, i32 127>
%4 = shufflevector <128 x i8> %1, <128 x i8> %2, <256 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 64, i32 65, i32 66, i32 67, i32 68, i32 69, i32 70, i32 71, i32 72, i32 73, i32 74, i32 75, i32 76, i32 77, i32 78, i32 79, i32 80, i32 81, i32 82, i32 83, i32 84, i32 85, i32 86, i32 87, i32 88, i32 89, i32 90, i32 91, i32 92, i32 93, i32 94, i32 95, i32 96, i32 97, i32 98, i32 99, i32 100, i32 101, i32 102, i32 103, i32 104, i32 105, i32 106, i32 107, i32 108, i32 109, i32 110, i32 111, i32 112, i32 113, i32 114, i32 115, i32 116, i32 117, i32 118, i32 119, i32 120, i32 121, i32 122, i32 123, i32 124, i32 125, i32 126, i32 127, i32 128, i32 129, i32 130, i32 131, i32 132, i32 133, i32 134, i32 135, i32 136, i32 137, i32 138, i32 139, i32 140, i32 141, i32 142, i32 143, i32 144, i32 145, i32 146, i32 147, i32 148, i32 149, i32 150, i32 151, i32 152, i32 153, i32 154, i32 155, i32 156, i32 157, i32 158, i32 159, i32 160, i32 161, i32 162, i32 163, i32 164, i32 165, i32 166, i32 167, i32 168, i32 169, i32 170, i32 171, i32 172, i32 173, i32 174, i32 175, i32 176, i32 177, i32 178, i32 179, i32 180, i32 181, i32 182, i32 183, i32 184, i32 185, i32 186, i32 187, i32 188, i32 189, i32 190, i32 191, i32 192, i32 193, i32 194, i32 195, i32 196, i32 197, i32 198, i32 199, i32 200, i32 201, i32 202, i32 203, i32 204, i32 205, i32 206, i32 207, i32 208, i32 209, i32 210, i32 211, i32 212, i32 213, i32 214, i32 215, i32 216, i32 217, i32 218, i32 219, i32 220, i32 221, i32 222, i32 223, i32 224, i32 225, i32 226, i32 227, i32 228, i32 229, i32 230, i32 231, i32 232, i32 233, i32 234, i32 235, i32 236, i32 237, i32 238, i32 239, i32 240, i32 241, i32 242, i32 243, i32 244, i32 245, i32 246, i32 247, i32 248, i32 249, i32 250, i32 251, i32 252, i32 253, i32 254, i32 255>
%5 = shufflevector <128 x i8> %3, <128 x i8> poison, <256 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 64, i32 65, i32 66, i32 67, i32 68, i32 69, i32 70, i32 71, i32 72, i32 73, i32 74, i32 75, i32 76, i32 77, i32 78, i32 79, i32 80, i32 81, i32 82, i32 83, i32 84, i32 85, i32 86, i32 87, i32 88, i32 89, i32 90, i32 91, i32 92, i32 93, i32 94, i32 95, i32 96, i32 97, i32 98, i32 99, i32 100, i32 101, i32 102, i32 103, i32 104, i32 105, i32 106, i32 107, i32 108, i32 109, i32 110, i32 111, i32 112, i32 113, i32 114, i32 115, i32 116, i32 117, i32 118, i32 119, i32 120, i32 121, i32 122, i32 123, i32 124, i32 125, i32 126, i32 127, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
%6 = shufflevector <256 x i8> %4, <256 x i8> %5, <384 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 64, i32 65, i32 66, i32 67, i32 68, i32 69, i32 70, i32 71, i32 72, i32 73, i32 74, i32 75, i32 76, i32 77, i32 78, i32 79, i32 80, i32 81, i32 82, i32 83, i32 84, i32 85, i32 86, i32 87, i32 88, i32 89, i32 90, i32 91, i32 92, i32 93, i32 94, i32 95, i32 96, i32 97, i32 98, i32 99, i32 100, i32 101, i32 102, i32 103, i32 104, i32 105, i32 106, i32 107, i32 108, i32 109, i32 110, i32 111, i32 112, i32 113, i32 114, i32 115, i32 116, i32 117, i32 118, i32 119, i32 120, i32 121, i32 122, i32 123, i32 124, i32 125, i32 126, i32 127, i32 128, i32 129, i32 130, i32 131, i32 132, i32 133, i32 134, i32 135, i32 136, i32 137, i32 138, i32 139, i32 140, i32 141, i32 142, i32 143, i32 144, i32 145, i32 146, i32 147, i32 148, i32 149, i32 150, i32 151, i32 152, i32 153, i32 154, i32 155, i32 156, i32 157, i32 158, i32 159, i32 160, i32 161, i32 162, i32 163, i32 164, i32 165, i32 166, i32 167, i32 168, i32 169, i32 170, i32 171, i32 172, i32 173, i32 174, i32 175, i32 176, i32 177, i32 178, i32 179, i32 180, i32 181, i32 182, i32 183, i32 184, i32 185, i32 186, i32 187, i32 188, i32 189, i32 190, i32 191, i32 192, i32 193, i32 194, i32 195, i32 196, i32 197, i32 198, i32 199, i32 200, i32 201, i32 202, i32 203, i32 204, i32 205, i32 206, i32 207, i32 208, i32 209, i32 210, i32 211, i32 212, i32 213, i32 214, i32 215, i32 216, i32 217, i32 218, i32 219, i32 220, i32 221, i32 222, i32 223, i32 224, i32 225, i32 226, i32 227, i32 228, i32 229, i32 230, i32 231, i32 232, i32 233, i32 234, i32 235, i32 236, i32 237, i32 238, i32 239, i32 240, i32 241, i32 242, i32 243, i32 244, i32 245, i32 246, i32 247, i32 248, i32 249, i32 250, i32 251, i32 252, i32 253, i32 254, i32 255, i32 256, i32 257, i32 258, i32 259, i32 260, i32 261, i32 262, i32 263, i32 264, i32 265, i32 266, i32 267, i32 268, i32 269, i32 270, i32 271, i32 272, i32 273, i32 274, i32 275, i32 276, i32 277, i32 278, i32 279, i32 280, i32 281, i32 282, i32 283, i32 284, i32 285, i32 286, i32 287, i32 288, i32 289, i32 290, i32 291, i32 292, i32 293, i32 294, i32 295, i32 296, i32 297, i32 298, i32 299, i32 300, i32 301, i32 302, i32 303, i32 304, i32 305, i32 306, i32 307, i32 308, i32 309, i32 310, i32 311, i32 312, i32 313, i32 314, i32 315, i32 316, i32 317, i32 318, i32 319, i32 320, i32 321, i32 322, i32 323, i32 324, i32 325, i32 326, i32 327, i32 328, i32 329, i32 330, i32 331, i32 332, i32 333, i32 334, i32 335, i32 336, i32 337, i32 338, i32 339, i32 340, i32 341, i32 342, i32 343, i32 344, i32 345, i32 346, i32 347, i32 348, i32 349, i32 350, i32 351, i32 352, i32 353, i32 354, i32 355, i32 356, i32 357, i32 358, i32 359, i32 360, i32 361, i32 362, i32 363, i32 364, i32 365, i32 366, i32 367, i32 368, i32 369, i32 370, i32 371, i32 372, i32 373, i32 374, i32 375, i32 376, i32 377, i32 378, i32 379, i32 380, i32 381, i32 382, i32 383>
%interleaved.vec = shufflevector <384 x i8> %6, <384 x i8> poison, <384 x i32> <i32 0, i32 64, i32 128, i32 192, i32 256, i32 320, i32 1, i32 65, i32 129, i32 193, i32 257, i32 321, i32 2, i32 66, i32 130, i32 194, i32 258, i32 322, i32 3, i32 67, i32 131, i32 195, i32 259, i32 323, i32 4, i32 68, i32 132, i32 196, i32 260, i32 324, i32 5, i32 69, i32 133, i32 197, i32 261, i32 325, i32 6, i32 70, i32 134, i32 198, i32 262, i32 326, i32 7, i32 71, i32 135, i32 199, i32 263, i32 327, i32 8, i32 72, i32 136, i32 200, i32 264, i32 328, i32 9, i32 73, i32 137, i32 201, i32 265, i32 329, i32 10, i32 74, i32 138, i32 202, i32 266, i32 330, i32 11, i32 75, i32 139, i32 203, i32 267, i32 331, i32 12, i32 76, i32 140, i32 204, i32 268, i32 332, i32 13, i32 77, i32 141, i32 205, i32 269, i32 333, i32 14, i32 78, i32 142, i32 206, i32 270, i32 334, i32 15, i32 79, i32 143, i32 207, i32 271, i32 335, i32 16, i32 80, i32 144, i32 208, i32 272, i32 336, i32 17, i32 81, i32 145, i32 209, i32 273, i32 337, i32 18, i32 82, i32 146, i32 210, i32 274, i32 338, i32 19, i32 83, i32 147, i32 211, i32 275, i32 339, i32 20, i32 84, i32 148, i32 212, i32 276, i32 340, i32 21, i32 85, i32 149, i32 213, i32 277, i32 341, i32 22, i32 86, i32 150, i32 214, i32 278, i32 342, i32 23, i32 87, i32 151, i32 215, i32 279, i32 343, i32 24, i32 88, i32 152, i32 216, i32 280, i32 344, i32 25, i32 89, i32 153, i32 217, i32 281, i32 345, i32 26, i32 90, i32 154, i32 218, i32 282, i32 346, i32 27, i32 91, i32 155, i32 219, i32 283, i32 347, i32 28, i32 92, i32 156, i32 220, i32 284, i32 348, i32 29, i32 93, i32 157, i32 221, i32 285, i32 349, i32 30, i32 94, i32 158, i32 222, i32 286, i32 350, i32 31, i32 95, i32 159, i32 223, i32 287, i32 351, i32 32, i32 96, i32 160, i32 224, i32 288, i32 352, i32 33, i32 97, i32 161, i32 225, i32 289, i32 353, i32 34, i32 98, i32 162, i32 226, i32 290, i32 354, i32 35, i32 99, i32 163, i32 227, i32 291, i32 355, i32 36, i32 100, i32 164, i32 228, i32 292, i32 356, i32 37, i32 101, i32 165, i32 229, i32 293, i32 357, i32 38, i32 102, i32 166, i32 230, i32 294, i32 358, i32 39, i32 103, i32 167, i32 231, i32 295, i32 359, i32 40, i32 104, i32 168, i32 232, i32 296, i32 360, i32 41, i32 105, i32 169, i32 233, i32 297, i32 361, i32 42, i32 106, i32 170, i32 234, i32 298, i32 362, i32 43, i32 107, i32 171, i32 235, i32 299, i32 363, i32 44, i32 108, i32 172, i32 236, i32 300, i32 364, i32 45, i32 109, i32 173, i32 237, i32 301, i32 365, i32 46, i32 110, i32 174, i32 238, i32 302, i32 366, i32 47, i32 111, i32 175, i32 239, i32 303, i32 367, i32 48, i32 112, i32 176, i32 240, i32 304, i32 368, i32 49, i32 113, i32 177, i32 241, i32 305, i32 369, i32 50, i32 114, i32 178, i32 242, i32 306, i32 370, i32 51, i32 115, i32 179, i32 243, i32 307, i32 371, i32 52, i32 116, i32 180, i32 244, i32 308, i32 372, i32 53, i32 117, i32 181, i32 245, i32 309, i32 373, i32 54, i32 118, i32 182, i32 246, i32 310, i32 374, i32 55, i32 119, i32 183, i32 247, i32 311, i32 375, i32 56, i32 120, i32 184, i32 248, i32 312, i32 376, i32 57, i32 121, i32 185, i32 249, i32 313, i32 377, i32 58, i32 122, i32 186, i32 250, i32 314, i32 378, i32 59, i32 123, i32 187, i32 251, i32 315, i32 379, i32 60, i32 124, i32 188, i32 252, i32 316, i32 380, i32 61, i32 125, i32 189, i32 253, i32 317, i32 381, i32 62, i32 126, i32 190, i32 254, i32 318, i32 382, i32 63, i32 127, i32 191, i32 255, i32 319, i32 383>
store <384 x i8> %interleaved.vec, ptr %out.vec, align 64
ret void
}
;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
; AVX1: {{.*}}
; AVX2: {{.*}}
; AVX512: {{.*}}
; AVX512BW-ONLY-FAST: {{.*}}
; AVX512DQ-FAST: {{.*}}
; AVX512DQ-SLOW: {{.*}}
; AVX512DQBW-FAST: {{.*}}
; AVX512F-ONLY-FAST: {{.*}}
; AVX512F-ONLY-SLOW: {{.*}}
; FALLBACK0: {{.*}}
; FALLBACK1: {{.*}}
; FALLBACK10: {{.*}}
; FALLBACK11: {{.*}}
; FALLBACK12: {{.*}}
; FALLBACK2: {{.*}}
; FALLBACK3: {{.*}}
; FALLBACK4: {{.*}}
; FALLBACK5: {{.*}}
; FALLBACK6: {{.*}}
; FALLBACK7: {{.*}}
; FALLBACK8: {{.*}}
; FALLBACK9: {{.*}}