; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py ; RUN: llc < %s -mtriple=x86_64-- -mattr=+sse2 | FileCheck %s --check-prefixes=SSE ; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx | FileCheck %s --check-prefixes=AVX ; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx2 | FileCheck %s --check-prefixes=AVX2 ; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx2,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX2-FP ; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx2,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX2-FCP ; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl | FileCheck %s --check-prefixes=AVX512 ; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX512-FCP ; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+avx512dq | FileCheck %s --check-prefixes=AVX512DQ ; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+avx512dq,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX512DQ-FCP ; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+avx512bw | FileCheck %s --check-prefixes=AVX512BW ; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+avx512bw,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX512BW-FCP ; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+avx512dq,+avx512bw | FileCheck %s --check-prefixes=AVX512DQ-BW ; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+avx512dq,+avx512bw,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX512DQ-BW-FCP ; These patterns are produced by LoopVectorizer for interleaved 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 ; ; AVX2-LABEL: store_i8_stride6_vf2: ; AVX2: # %bb.0: ; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX2-NEXT: vmovdqa (%rdi), %xmm0 ; AVX2-NEXT: vmovdqa (%rdx), %xmm1 ; AVX2-NEXT: vmovdqa (%r9), %xmm2 ; AVX2-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3] ; AVX2-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1],xmm1[2],mem[2],xmm1[3],mem[3] ; AVX2-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] ; AVX2-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] ; AVX2-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,4,8,12,1],zero,xmm0[2,6,10,14,3],zero,xmm0[u,u,u,u] ; AVX2-NEXT: vpshufb {{.*#+}} xmm1 = zero,zero,zero,zero,zero,xmm2[0],zero,zero,zero,zero,zero,xmm2[1,u,u,u,u] ; AVX2-NEXT: vpor %xmm1, %xmm0, %xmm0 ; AVX2-NEXT: vpextrd $2, %xmm0, 8(%rax) ; AVX2-NEXT: vmovq %xmm0, (%rax) ; AVX2-NEXT: retq ; ; AVX2-FP-LABEL: store_i8_stride6_vf2: ; AVX2-FP: # %bb.0: ; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX2-FP-NEXT: vmovdqa (%rdi), %xmm0 ; AVX2-FP-NEXT: vmovdqa (%rdx), %xmm1 ; AVX2-FP-NEXT: vmovdqa (%r9), %xmm2 ; AVX2-FP-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3] ; AVX2-FP-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1],xmm1[2],mem[2],xmm1[3],mem[3] ; AVX2-FP-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] ; AVX2-FP-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] ; AVX2-FP-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,4,8,12,1],zero,xmm0[2,6,10,14,3],zero,xmm0[u,u,u,u] ; AVX2-FP-NEXT: vpshufb {{.*#+}} xmm1 = zero,zero,zero,zero,zero,xmm2[0],zero,zero,zero,zero,zero,xmm2[1,u,u,u,u] ; AVX2-FP-NEXT: vpor %xmm1, %xmm0, %xmm0 ; AVX2-FP-NEXT: vpextrd $2, %xmm0, 8(%rax) ; AVX2-FP-NEXT: vmovq %xmm0, (%rax) ; AVX2-FP-NEXT: retq ; ; AVX2-FCP-LABEL: store_i8_stride6_vf2: ; AVX2-FCP: # %bb.0: ; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX2-FCP-NEXT: vmovdqa (%rdi), %xmm0 ; AVX2-FCP-NEXT: vmovdqa (%rdx), %xmm1 ; AVX2-FCP-NEXT: vmovdqa (%r9), %xmm2 ; AVX2-FCP-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3] ; AVX2-FCP-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1],xmm1[2],mem[2],xmm1[3],mem[3] ; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] ; AVX2-FCP-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] ; AVX2-FCP-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,4,8,12,1],zero,xmm0[2,6,10,14,3],zero,xmm0[u,u,u,u] ; AVX2-FCP-NEXT: vpshufb {{.*#+}} xmm1 = zero,zero,zero,zero,zero,xmm2[0],zero,zero,zero,zero,zero,xmm2[1,u,u,u,u] ; AVX2-FCP-NEXT: vpor %xmm1, %xmm0, %xmm0 ; AVX2-FCP-NEXT: vpextrd $2, %xmm0, 8(%rax) ; AVX2-FCP-NEXT: vmovq %xmm0, (%rax) ; AVX2-FCP-NEXT: retq ; ; AVX512-LABEL: store_i8_stride6_vf2: ; AVX512: # %bb.0: ; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512-NEXT: vmovdqa (%rdi), %xmm0 ; AVX512-NEXT: vmovdqa (%rdx), %xmm1 ; AVX512-NEXT: vmovdqa (%r9), %xmm2 ; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3] ; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1],xmm1[2],mem[2],xmm1[3],mem[3] ; AVX512-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] ; AVX512-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] ; AVX512-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,4,8,12,1],zero,xmm0[2,6,10,14,3],zero,xmm0[u,u,u,u] ; AVX512-NEXT: vpshufb {{.*#+}} xmm1 = zero,zero,zero,zero,zero,xmm2[0],zero,zero,zero,zero,zero,xmm2[1,u,u,u,u] ; AVX512-NEXT: vpor %xmm1, %xmm0, %xmm0 ; AVX512-NEXT: vpextrd $2, %xmm0, 8(%rax) ; AVX512-NEXT: vmovq %xmm0, (%rax) ; AVX512-NEXT: retq ; ; AVX512-FCP-LABEL: store_i8_stride6_vf2: ; AVX512-FCP: # %bb.0: ; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512-FCP-NEXT: vmovdqa (%rdi), %xmm0 ; AVX512-FCP-NEXT: vmovdqa (%rdx), %xmm1 ; AVX512-FCP-NEXT: vmovdqa (%r9), %xmm2 ; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3] ; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1],xmm1[2],mem[2],xmm1[3],mem[3] ; AVX512-FCP-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] ; AVX512-FCP-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] ; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,4,8,12,1],zero,xmm0[2,6,10,14,3],zero,xmm0[u,u,u,u] ; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm1 = zero,zero,zero,zero,zero,xmm2[0],zero,zero,zero,zero,zero,xmm2[1,u,u,u,u] ; AVX512-FCP-NEXT: vpor %xmm1, %xmm0, %xmm0 ; AVX512-FCP-NEXT: vpextrd $2, %xmm0, 8(%rax) ; AVX512-FCP-NEXT: vmovq %xmm0, (%rax) ; AVX512-FCP-NEXT: retq ; ; AVX512DQ-LABEL: store_i8_stride6_vf2: ; AVX512DQ: # %bb.0: ; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512DQ-NEXT: vmovdqa (%rdi), %xmm0 ; AVX512DQ-NEXT: vmovdqa (%rdx), %xmm1 ; AVX512DQ-NEXT: vmovdqa (%r9), %xmm2 ; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3] ; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1],xmm1[2],mem[2],xmm1[3],mem[3] ; AVX512DQ-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] ; AVX512DQ-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] ; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,4,8,12,1],zero,xmm0[2,6,10,14,3],zero,xmm0[u,u,u,u] ; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm1 = zero,zero,zero,zero,zero,xmm2[0],zero,zero,zero,zero,zero,xmm2[1,u,u,u,u] ; AVX512DQ-NEXT: vpor %xmm1, %xmm0, %xmm0 ; AVX512DQ-NEXT: vpextrd $2, %xmm0, 8(%rax) ; AVX512DQ-NEXT: vmovq %xmm0, (%rax) ; AVX512DQ-NEXT: retq ; ; AVX512DQ-FCP-LABEL: store_i8_stride6_vf2: ; AVX512DQ-FCP: # %bb.0: ; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %xmm0 ; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %xmm1 ; AVX512DQ-FCP-NEXT: vmovdqa (%r9), %xmm2 ; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3] ; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1],xmm1[2],mem[2],xmm1[3],mem[3] ; AVX512DQ-FCP-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] ; AVX512DQ-FCP-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] ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,4,8,12,1],zero,xmm0[2,6,10,14,3],zero,xmm0[u,u,u,u] ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm1 = zero,zero,zero,zero,zero,xmm2[0],zero,zero,zero,zero,zero,xmm2[1,u,u,u,u] ; AVX512DQ-FCP-NEXT: vpor %xmm1, %xmm0, %xmm0 ; AVX512DQ-FCP-NEXT: vpextrd $2, %xmm0, 8(%rax) ; AVX512DQ-FCP-NEXT: vmovq %xmm0, (%rax) ; AVX512DQ-FCP-NEXT: retq ; ; AVX512BW-LABEL: store_i8_stride6_vf2: ; AVX512BW: # %bb.0: ; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512BW-NEXT: vmovdqa (%rdi), %xmm0 ; AVX512BW-NEXT: vmovdqa (%rdx), %xmm1 ; AVX512BW-NEXT: vmovdqa (%r9), %xmm2 ; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3] ; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1],xmm1[2],mem[2],xmm1[3],mem[3] ; AVX512BW-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] ; AVX512BW-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] ; AVX512BW-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,4,8,12,1],zero,xmm0[2,6,10,14,3],zero,xmm0[u,u,u,u] ; AVX512BW-NEXT: vpshufb {{.*#+}} xmm1 = zero,zero,zero,zero,zero,xmm2[0],zero,zero,zero,zero,zero,xmm2[1,u,u,u,u] ; AVX512BW-NEXT: vpor %xmm1, %xmm0, %xmm0 ; AVX512BW-NEXT: vpextrd $2, %xmm0, 8(%rax) ; AVX512BW-NEXT: vmovq %xmm0, (%rax) ; AVX512BW-NEXT: retq ; ; AVX512BW-FCP-LABEL: store_i8_stride6_vf2: ; AVX512BW-FCP: # %bb.0: ; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512BW-FCP-NEXT: vmovdqa (%rdi), %xmm0 ; AVX512BW-FCP-NEXT: vmovdqa (%rdx), %xmm1 ; AVX512BW-FCP-NEXT: vmovdqa (%r9), %xmm2 ; AVX512BW-FCP-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3] ; AVX512BW-FCP-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1],xmm1[2],mem[2],xmm1[3],mem[3] ; AVX512BW-FCP-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] ; AVX512BW-FCP-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] ; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,4,8,12,1],zero,xmm0[2,6,10,14,3],zero,xmm0[u,u,u,u] ; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm1 = zero,zero,zero,zero,zero,xmm2[0],zero,zero,zero,zero,zero,xmm2[1,u,u,u,u] ; AVX512BW-FCP-NEXT: vpor %xmm1, %xmm0, %xmm0 ; AVX512BW-FCP-NEXT: vpextrd $2, %xmm0, 8(%rax) ; AVX512BW-FCP-NEXT: vmovq %xmm0, (%rax) ; AVX512BW-FCP-NEXT: retq ; ; AVX512DQ-BW-LABEL: store_i8_stride6_vf2: ; AVX512DQ-BW: # %bb.0: ; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512DQ-BW-NEXT: vmovdqa (%rdi), %xmm0 ; AVX512DQ-BW-NEXT: vmovdqa (%rdx), %xmm1 ; AVX512DQ-BW-NEXT: vmovdqa (%r9), %xmm2 ; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3] ; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1],xmm1[2],mem[2],xmm1[3],mem[3] ; AVX512DQ-BW-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] ; AVX512DQ-BW-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] ; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,4,8,12,1],zero,xmm0[2,6,10,14,3],zero,xmm0[u,u,u,u] ; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm1 = zero,zero,zero,zero,zero,xmm2[0],zero,zero,zero,zero,zero,xmm2[1,u,u,u,u] ; AVX512DQ-BW-NEXT: vpor %xmm1, %xmm0, %xmm0 ; AVX512DQ-BW-NEXT: vpextrd $2, %xmm0, 8(%rax) ; AVX512DQ-BW-NEXT: vmovq %xmm0, (%rax) ; AVX512DQ-BW-NEXT: retq ; ; AVX512DQ-BW-FCP-LABEL: store_i8_stride6_vf2: ; AVX512DQ-BW-FCP: # %bb.0: ; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdi), %xmm0 ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdx), %xmm1 ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%r9), %xmm2 ; AVX512DQ-BW-FCP-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3] ; AVX512DQ-BW-FCP-NEXT: vpunpcklwd {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1],xmm1[2],mem[2],xmm1[3],mem[3] ; AVX512DQ-BW-FCP-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] ; AVX512DQ-BW-FCP-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] ; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,4,8,12,1],zero,xmm0[2,6,10,14,3],zero,xmm0[u,u,u,u] ; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm1 = zero,zero,zero,zero,zero,xmm2[0],zero,zero,zero,zero,zero,xmm2[1,u,u,u,u] ; AVX512DQ-BW-FCP-NEXT: vpor %xmm1, %xmm0, %xmm0 ; AVX512DQ-BW-FCP-NEXT: vpextrd $2, %xmm0, 8(%rax) ; AVX512DQ-BW-FCP-NEXT: vmovq %xmm0, (%rax) ; AVX512DQ-BW-FCP-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> %2 = shufflevector <2 x i8> %in.vec2, <2 x i8> %in.vec3, <4 x i32> %3 = shufflevector <2 x i8> %in.vec4, <2 x i8> %in.vec5, <4 x i32> %4 = shufflevector <4 x i8> %1, <4 x i8> %2, <8 x i32> %5 = shufflevector <4 x i8> %3, <4 x i8> poison, <8 x i32> %6 = shufflevector <8 x i8> %4, <8 x i8> %5, <12 x i32> %interleaved.vec = shufflevector <12 x i8> %6, <12 x i8> poison, <12 x i32> 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 ; ; AVX-LABEL: store_i8_stride6_vf4: ; AVX: # %bb.0: ; AVX-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX-NEXT: vmovdqa (%rdi), %xmm0 ; AVX-NEXT: vmovdqa (%rdx), %xmm1 ; AVX-NEXT: vmovdqa (%r8), %xmm2 ; AVX-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1] ; AVX-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1] ; AVX-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0] ; AVX-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm2[0],mem[0],xmm2[1],mem[1] ; AVX-NEXT: vpshufb {{.*#+}} xmm2 = xmm0[0,4,8,12],zero,zero,xmm0[1,5,9,13],zero,zero,xmm0[2,6,10,14] ; AVX-NEXT: vpshufb {{.*#+}} xmm3 = zero,zero,zero,zero,xmm1[0,4],zero,zero,zero,zero,xmm1[1,5],zero,zero,zero,zero ; AVX-NEXT: vpor %xmm3, %xmm2, %xmm2 ; AVX-NEXT: vpshufb {{.*#+}} xmm0 = zero,zero,xmm0[3,7,11,15],zero,zero,xmm0[u,u,u,u,u,u,u,u] ; AVX-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[2,6],zero,zero,zero,zero,xmm1[3,7,u,u,u,u,u,u,u,u] ; AVX-NEXT: vpor %xmm1, %xmm0, %xmm0 ; AVX-NEXT: vmovq %xmm0, 16(%rax) ; AVX-NEXT: vmovdqa %xmm2, (%rax) ; AVX-NEXT: retq ; ; AVX2-LABEL: store_i8_stride6_vf4: ; AVX2: # %bb.0: ; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX2-NEXT: vmovdqa (%rdi), %xmm0 ; AVX2-NEXT: vmovdqa (%rdx), %xmm1 ; AVX2-NEXT: vmovdqa (%r8), %xmm2 ; AVX2-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1] ; AVX2-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1] ; AVX2-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0] ; AVX2-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm2[0],mem[0],xmm2[1],mem[1] ; AVX2-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm0 ; AVX2-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-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,0,1] ; AVX2-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-NEXT: vpor %ymm0, %ymm1, %ymm0 ; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1 ; AVX2-NEXT: vmovq %xmm1, 16(%rax) ; AVX2-NEXT: vmovdqa %xmm0, (%rax) ; AVX2-NEXT: vzeroupper ; AVX2-NEXT: retq ; ; AVX2-FP-LABEL: store_i8_stride6_vf4: ; AVX2-FP: # %bb.0: ; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX2-FP-NEXT: vmovdqa (%rdi), %xmm0 ; AVX2-FP-NEXT: vmovdqa (%rdx), %xmm1 ; AVX2-FP-NEXT: vmovdqa (%r8), %xmm2 ; AVX2-FP-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1] ; AVX2-FP-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1] ; AVX2-FP-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0] ; AVX2-FP-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm2[0],mem[0],xmm2[1],mem[1] ; AVX2-FP-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm0 ; AVX2-FP-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-FP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,0,1] ; AVX2-FP-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-FP-NEXT: vpor %ymm0, %ymm1, %ymm0 ; AVX2-FP-NEXT: vextracti128 $1, %ymm0, %xmm1 ; AVX2-FP-NEXT: vmovq %xmm1, 16(%rax) ; AVX2-FP-NEXT: vmovdqa %xmm0, (%rax) ; AVX2-FP-NEXT: vzeroupper ; AVX2-FP-NEXT: retq ; ; AVX2-FCP-LABEL: store_i8_stride6_vf4: ; AVX2-FCP: # %bb.0: ; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX2-FCP-NEXT: vmovdqa (%rdi), %xmm0 ; AVX2-FCP-NEXT: vmovdqa (%rdx), %xmm1 ; AVX2-FCP-NEXT: vmovdqa (%r8), %xmm2 ; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1] ; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1] ; AVX2-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0] ; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm2[0],mem[0],xmm2[1],mem[1] ; AVX2-FCP-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm0 ; AVX2-FCP-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-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,0,1] ; AVX2-FCP-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-FCP-NEXT: vpor %ymm0, %ymm1, %ymm0 ; AVX2-FCP-NEXT: vextracti128 $1, %ymm0, %xmm1 ; AVX2-FCP-NEXT: vmovq %xmm1, 16(%rax) ; AVX2-FCP-NEXT: vmovdqa %xmm0, (%rax) ; AVX2-FCP-NEXT: vzeroupper ; AVX2-FCP-NEXT: retq ; ; AVX512-LABEL: store_i8_stride6_vf4: ; AVX512: # %bb.0: ; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512-NEXT: vmovdqa (%rdi), %xmm0 ; AVX512-NEXT: vmovdqa (%rdx), %xmm1 ; AVX512-NEXT: vmovdqa (%r8), %xmm2 ; AVX512-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1] ; AVX512-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1] ; AVX512-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0] ; AVX512-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm2[0],mem[0],xmm2[1],mem[1] ; AVX512-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm0 ; AVX512-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] ; AVX512-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,0,1] ; AVX512-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 ; AVX512-NEXT: vpor %ymm0, %ymm1, %ymm0 ; AVX512-NEXT: vextracti128 $1, %ymm0, %xmm1 ; AVX512-NEXT: vmovq %xmm1, 16(%rax) ; AVX512-NEXT: vmovdqa %xmm0, (%rax) ; AVX512-NEXT: vzeroupper ; AVX512-NEXT: retq ; ; AVX512-FCP-LABEL: store_i8_stride6_vf4: ; AVX512-FCP: # %bb.0: ; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512-FCP-NEXT: vmovdqa (%rdi), %xmm0 ; AVX512-FCP-NEXT: vmovdqa (%rdx), %xmm1 ; AVX512-FCP-NEXT: vmovdqa (%r8), %xmm2 ; AVX512-FCP-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1] ; AVX512-FCP-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1] ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0] ; AVX512-FCP-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm2[0],mem[0],xmm2[1],mem[1] ; AVX512-FCP-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm0 ; AVX512-FCP-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] ; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,0,1] ; AVX512-FCP-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 ; AVX512-FCP-NEXT: vpor %ymm0, %ymm1, %ymm0 ; AVX512-FCP-NEXT: vextracti128 $1, %ymm0, %xmm1 ; AVX512-FCP-NEXT: vmovq %xmm1, 16(%rax) ; AVX512-FCP-NEXT: vmovdqa %xmm0, (%rax) ; AVX512-FCP-NEXT: vzeroupper ; AVX512-FCP-NEXT: retq ; ; AVX512DQ-LABEL: store_i8_stride6_vf4: ; AVX512DQ: # %bb.0: ; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512DQ-NEXT: vmovdqa (%rdi), %xmm0 ; AVX512DQ-NEXT: vmovdqa (%rdx), %xmm1 ; AVX512DQ-NEXT: vmovdqa (%r8), %xmm2 ; AVX512DQ-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1] ; AVX512DQ-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1] ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0] ; AVX512DQ-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm2[0],mem[0],xmm2[1],mem[1] ; AVX512DQ-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm0 ; AVX512DQ-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] ; AVX512DQ-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,0,1] ; AVX512DQ-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 ; AVX512DQ-NEXT: vpor %ymm0, %ymm1, %ymm0 ; AVX512DQ-NEXT: vextracti128 $1, %ymm0, %xmm1 ; AVX512DQ-NEXT: vmovq %xmm1, 16(%rax) ; AVX512DQ-NEXT: vmovdqa %xmm0, (%rax) ; AVX512DQ-NEXT: vzeroupper ; AVX512DQ-NEXT: retq ; ; AVX512DQ-FCP-LABEL: store_i8_stride6_vf4: ; AVX512DQ-FCP: # %bb.0: ; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %xmm0 ; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %xmm1 ; AVX512DQ-FCP-NEXT: vmovdqa (%r8), %xmm2 ; AVX512DQ-FCP-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1] ; AVX512DQ-FCP-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1] ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0] ; AVX512DQ-FCP-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm2[0],mem[0],xmm2[1],mem[1] ; AVX512DQ-FCP-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm0 ; AVX512DQ-FCP-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] ; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,0,1] ; AVX512DQ-FCP-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 ; AVX512DQ-FCP-NEXT: vpor %ymm0, %ymm1, %ymm0 ; AVX512DQ-FCP-NEXT: vextracti128 $1, %ymm0, %xmm1 ; AVX512DQ-FCP-NEXT: vmovq %xmm1, 16(%rax) ; AVX512DQ-FCP-NEXT: vmovdqa %xmm0, (%rax) ; AVX512DQ-FCP-NEXT: vzeroupper ; AVX512DQ-FCP-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],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 ; AVX512BW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,0,1] ; AVX512BW-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 ; AVX512BW-NEXT: vpor %ymm1, %ymm0, %ymm0 ; AVX512BW-NEXT: vextracti128 $1, %ymm0, %xmm1 ; AVX512BW-NEXT: vmovq %xmm1, 16(%rax) ; AVX512BW-NEXT: vmovdqa %xmm0, (%rax) ; AVX512BW-NEXT: vzeroupper ; AVX512BW-NEXT: retq ; ; AVX512BW-FCP-LABEL: store_i8_stride6_vf4: ; AVX512BW-FCP: # %bb.0: ; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512BW-FCP-NEXT: vmovdqa (%rdi), %xmm0 ; AVX512BW-FCP-NEXT: vmovdqa (%rdx), %xmm1 ; AVX512BW-FCP-NEXT: vmovdqa (%r8), %xmm2 ; AVX512BW-FCP-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1] ; AVX512BW-FCP-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1] ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0] ; AVX512BW-FCP-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm2[0],mem[0],xmm2[1],mem[1] ; AVX512BW-FCP-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm0 ; AVX512BW-FCP-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 ; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,0,1] ; AVX512BW-FCP-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 ; AVX512BW-FCP-NEXT: vpor %ymm1, %ymm0, %ymm0 ; AVX512BW-FCP-NEXT: vextracti128 $1, %ymm0, %xmm1 ; AVX512BW-FCP-NEXT: vmovq %xmm1, 16(%rax) ; AVX512BW-FCP-NEXT: vmovdqa %xmm0, (%rax) ; AVX512BW-FCP-NEXT: vzeroupper ; AVX512BW-FCP-NEXT: retq ; ; AVX512DQ-BW-LABEL: store_i8_stride6_vf4: ; AVX512DQ-BW: # %bb.0: ; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512DQ-BW-NEXT: vmovdqa (%rdi), %xmm0 ; AVX512DQ-BW-NEXT: vmovdqa (%rdx), %xmm1 ; AVX512DQ-BW-NEXT: vmovdqa (%r8), %xmm2 ; AVX512DQ-BW-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1] ; AVX512DQ-BW-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1] ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0] ; AVX512DQ-BW-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm2[0],mem[0],xmm2[1],mem[1] ; AVX512DQ-BW-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm0 ; AVX512DQ-BW-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 ; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,0,1] ; AVX512DQ-BW-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 ; AVX512DQ-BW-NEXT: vpor %ymm1, %ymm0, %ymm0 ; AVX512DQ-BW-NEXT: vextracti128 $1, %ymm0, %xmm1 ; AVX512DQ-BW-NEXT: vmovq %xmm1, 16(%rax) ; AVX512DQ-BW-NEXT: vmovdqa %xmm0, (%rax) ; AVX512DQ-BW-NEXT: vzeroupper ; AVX512DQ-BW-NEXT: retq ; ; AVX512DQ-BW-FCP-LABEL: store_i8_stride6_vf4: ; AVX512DQ-BW-FCP: # %bb.0: ; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdi), %xmm0 ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdx), %xmm1 ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%r8), %xmm2 ; AVX512DQ-BW-FCP-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1] ; AVX512DQ-BW-FCP-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1] ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0] ; AVX512DQ-BW-FCP-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm2[0],mem[0],xmm2[1],mem[1] ; AVX512DQ-BW-FCP-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm0 ; AVX512DQ-BW-FCP-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 ; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,3,0,1] ; AVX512DQ-BW-FCP-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 ; AVX512DQ-BW-FCP-NEXT: vpor %ymm1, %ymm0, %ymm0 ; AVX512DQ-BW-FCP-NEXT: vextracti128 $1, %ymm0, %xmm1 ; AVX512DQ-BW-FCP-NEXT: vmovq %xmm1, 16(%rax) ; AVX512DQ-BW-FCP-NEXT: vmovdqa %xmm0, (%rax) ; AVX512DQ-BW-FCP-NEXT: vzeroupper ; AVX512DQ-BW-FCP-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> %2 = shufflevector <4 x i8> %in.vec2, <4 x i8> %in.vec3, <8 x i32> %3 = shufflevector <4 x i8> %in.vec4, <4 x i8> %in.vec5, <8 x i32> %4 = shufflevector <8 x i8> %1, <8 x i8> %2, <16 x i32> %5 = shufflevector <8 x i8> %3, <8 x i8> poison, <16 x i32> %6 = shufflevector <16 x i8> %4, <16 x i8> %5, <24 x i32> %interleaved.vec = shufflevector <24 x i8> %6, <24 x i8> poison, <24 x i32> 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 {{.*#+}} xmm2 = mem[0],zero ; SSE-NEXT: movq {{.*#+}} xmm1 = mem[0],zero ; SSE-NEXT: punpcklbw {{.*#+}} xmm2 = 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] ; SSE-NEXT: movq {{.*#+}} xmm1 = mem[0],zero ; SSE-NEXT: movq {{.*#+}} xmm7 = mem[0],zero ; SSE-NEXT: pxor %xmm4, %xmm4 ; SSE-NEXT: movdqa %xmm1, %xmm3 ; SSE-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm7[0],xmm1[1],xmm7[1],xmm1[2],xmm7[2],xmm1[3],xmm7[3],xmm1[4],xmm7[4],xmm1[5],xmm7[5],xmm1[6],xmm7[6],xmm1[7],xmm7[7] ; SSE-NEXT: punpcklbw {{.*#+}} xmm7 = 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] ; SSE-NEXT: punpcklbw {{.*#+}} xmm3 = 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] ; SSE-NEXT: movdqa %xmm3, %xmm4 ; SSE-NEXT: punpcklwd {{.*#+}} xmm4 = xmm4[0],xmm7[0],xmm4[1],xmm7[1],xmm4[2],xmm7[2],xmm4[3],xmm7[3] ; SSE-NEXT: packuswb %xmm4, %xmm4 ; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm4[0,0,2,3] ; SSE-NEXT: movdqa {{.*#+}} xmm4 = [65535,65535,0,65535,65535,0,65535,65535] ; SSE-NEXT: movdqa %xmm4, %xmm8 ; SSE-NEXT: pandn %xmm5, %xmm8 ; SSE-NEXT: pshufd {{.*#+}} xmm9 = xmm0[0,0,1,1] ; SSE-NEXT: movdqa {{.*#+}} xmm5 = [65535,0,65535,65535,0,65535,65535,0] ; SSE-NEXT: pand %xmm5, %xmm9 ; SSE-NEXT: pshuflw {{.*#+}} xmm6 = xmm2[1,0,2,2,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm10 = xmm6[0,1,0,1] ; SSE-NEXT: movdqa %xmm5, %xmm6 ; SSE-NEXT: pandn %xmm10, %xmm6 ; SSE-NEXT: por %xmm9, %xmm6 ; SSE-NEXT: pand %xmm4, %xmm6 ; SSE-NEXT: por %xmm8, %xmm6 ; SSE-NEXT: punpckhwd {{.*#+}} xmm3 = xmm3[4],xmm7[4],xmm3[5],xmm7[5],xmm3[6],xmm7[6],xmm3[7],xmm7[7] ; SSE-NEXT: packuswb %xmm3, %xmm3 ; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm3[0,1,3,3] ; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm0[3,3,3,3] ; SSE-NEXT: movdqa %xmm4, %xmm8 ; SSE-NEXT: pandn %xmm7, %xmm8 ; SSE-NEXT: pshufhw {{.*#+}} xmm7 = xmm2[0,1,2,3,5,6,7,7] ; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm7[2,2,2,3] ; SSE-NEXT: pand %xmm4, %xmm7 ; SSE-NEXT: por %xmm8, %xmm7 ; SSE-NEXT: pand %xmm5, %xmm7 ; SSE-NEXT: pandn %xmm3, %xmm5 ; SSE-NEXT: por %xmm7, %xmm5 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,2,2] ; SSE-NEXT: pand %xmm4, %xmm0 ; SSE-NEXT: pshuflw {{.*#+}} xmm2 = xmm2[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, %xmm4 ; SSE-NEXT: por %xmm0, %xmm4 ; SSE-NEXT: movdqa {{.*#+}} xmm0 = [0,65535,65535,0,65535,65535,0,65535] ; SSE-NEXT: pand %xmm0, %xmm4 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,2,2] ; SSE-NEXT: pandn %xmm1, %xmm0 ; SSE-NEXT: por %xmm4, %xmm0 ; SSE-NEXT: movdqa %xmm5, 32(%rax) ; SSE-NEXT: movdqa %xmm0, 16(%rax) ; SSE-NEXT: movdqa %xmm6, (%rax) ; SSE-NEXT: retq ; ; AVX-LABEL: store_i8_stride6_vf8: ; AVX: # %bb.0: ; AVX-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX-NEXT: vmovq {{.*#+}} xmm0 = mem[0],zero ; AVX-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero ; AVX-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0] ; AVX-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero ; AVX-NEXT: vmovq {{.*#+}} xmm2 = mem[0],zero ; AVX-NEXT: vpunpcklqdq {{.*#+}} xmm1 = xmm2[0],xmm1[0] ; AVX-NEXT: vmovq {{.*#+}} xmm2 = mem[0],zero ; AVX-NEXT: vmovq {{.*#+}} xmm3 = mem[0],zero ; AVX-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm3[0],xmm2[0] ; AVX-NEXT: vpshufb {{.*#+}} xmm3 = xmm1[u,u],zero,zero,xmm1[3,11,u,u],zero,zero,xmm1[4,12,u,u],zero,zero ; AVX-NEXT: vpshufb {{.*#+}} xmm4 = xmm0[u,u,3,11],zero,zero,xmm0[u,u,4,12],zero,zero,xmm0[u,u,5,13] ; AVX-NEXT: vpor %xmm3, %xmm4, %xmm3 ; AVX-NEXT: vpshufb {{.*#+}} xmm4 = xmm2[2,10,u,u,u,u,3,11,u,u,u,u,4,12,u,u] ; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm4[0],xmm3[1,2],xmm4[3],xmm3[4,5],xmm4[6],xmm3[7] ; AVX-NEXT: vpshufb {{.*#+}} xmm4 = zero,zero,xmm1[0,8,u,u],zero,zero,xmm1[1,9,u,u],zero,zero,xmm1[2,10] ; AVX-NEXT: vpshufb {{.*#+}} xmm5 = xmm0[0,8],zero,zero,xmm0[u,u,1,9],zero,zero,xmm0[u,u,2,10],zero,zero ; AVX-NEXT: vpor %xmm4, %xmm5, %xmm4 ; AVX-NEXT: vpshufb {{.*#+}} xmm5 = xmm2[u,u,u,u,0,8,u,u,u,u,1,9,u,u,u,u] ; AVX-NEXT: vpblendw {{.*#+}} xmm4 = xmm4[0,1],xmm5[2],xmm4[3,4],xmm5[5],xmm4[6,7] ; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm4, %ymm3 ; AVX-NEXT: vpshufb {{.*#+}} xmm0 = zero,zero,xmm0[u,u,6,14],zero,zero,xmm0[u,u,7,15],zero,zero,xmm0[u,u] ; AVX-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[5,13,u,u],zero,zero,xmm1[6,14,u,u],zero,zero,xmm1[7,15,u,u] ; AVX-NEXT: vpor %xmm0, %xmm1, %xmm0 ; AVX-NEXT: vpshufb {{.*#+}} xmm1 = xmm2[u,u,5,13,u,u,u,u,6,14,u,u,u,u,7,15] ; AVX-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0],xmm1[1],xmm0[2,3],xmm1[4],xmm0[5,6],xmm1[7] ; AVX-NEXT: vmovdqa %xmm0, 32(%rax) ; AVX-NEXT: vmovaps %ymm3, (%rax) ; AVX-NEXT: vzeroupper ; AVX-NEXT: retq ; ; AVX2-LABEL: store_i8_stride6_vf8: ; AVX2: # %bb.0: ; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX2-NEXT: vmovq {{.*#+}} xmm0 = mem[0],zero ; AVX2-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero ; AVX2-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0] ; AVX2-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero ; AVX2-NEXT: vmovq {{.*#+}} xmm2 = mem[0],zero ; AVX2-NEXT: vpunpcklqdq {{.*#+}} xmm1 = xmm2[0],xmm1[0] ; AVX2-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm2 ; AVX2-NEXT: vmovq {{.*#+}} xmm3 = mem[0],zero ; AVX2-NEXT: vmovq {{.*#+}} xmm4 = mem[0],zero ; AVX2-NEXT: vpunpcklqdq {{.*#+}} xmm3 = xmm4[0],xmm3[0] ; AVX2-NEXT: vmovdqa {{.*#+}} ymm4 = [0,8,0,8,u,u,1,9,1,9,u,u,2,10,2,10,u,u,3,11,3,11,u,u,4,12,4,12,u,u,5,13] ; AVX2-NEXT: vpshufb %ymm4, %ymm2, %ymm5 ; AVX2-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,3,0,1] ; AVX2-NEXT: vpshufb %ymm4, %ymm2, %ymm2 ; AVX2-NEXT: vpblendw {{.*#+}} ymm2 = ymm5[0],ymm2[1],ymm5[2,3],ymm2[4],ymm5[5,6],ymm2[7],ymm5[8],ymm2[9],ymm5[10,11],ymm2[12],ymm5[13,14],ymm2[15] ; AVX2-NEXT: vpshufb {{.*#+}} xmm4 = xmm3[2,10,1,9,0,8,3,11,u,u,u,u,4,12,u,u] ; AVX2-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,0,1] ; AVX2-NEXT: vpmovsxbw {{.*#+}} ymm5 = [65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535] ; AVX2-NEXT: vpblendvb %ymm5, %ymm2, %ymm4, %ymm2 ; AVX2-NEXT: vpshufb {{.*#+}} xmm0 = zero,zero,xmm0[u,u,6,14],zero,zero,xmm0[u,u,7,15],zero,zero,xmm0[u,u] ; AVX2-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[5,13,u,u],zero,zero,xmm1[6,14,u,u],zero,zero,xmm1[7,15,u,u] ; AVX2-NEXT: vpor %xmm0, %xmm1, %xmm0 ; AVX2-NEXT: vpshufb {{.*#+}} xmm1 = xmm3[u,u,5,13,u,u,u,u,6,14,u,u,u,u,7,15] ; AVX2-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0],xmm1[1],xmm0[2,3],xmm1[4],xmm0[5,6],xmm1[7] ; AVX2-NEXT: vmovdqa %xmm0, 32(%rax) ; AVX2-NEXT: vmovdqa %ymm2, (%rax) ; AVX2-NEXT: vzeroupper ; AVX2-NEXT: retq ; ; AVX2-FP-LABEL: store_i8_stride6_vf8: ; AVX2-FP: # %bb.0: ; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX2-FP-NEXT: vmovq {{.*#+}} xmm0 = mem[0],zero ; AVX2-FP-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero ; AVX2-FP-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0] ; AVX2-FP-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero ; AVX2-FP-NEXT: vmovq {{.*#+}} xmm2 = mem[0],zero ; AVX2-FP-NEXT: vpunpcklqdq {{.*#+}} xmm1 = xmm2[0],xmm1[0] ; AVX2-FP-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm2 ; AVX2-FP-NEXT: vmovq {{.*#+}} xmm3 = mem[0],zero ; AVX2-FP-NEXT: vmovq {{.*#+}} xmm4 = mem[0],zero ; AVX2-FP-NEXT: vpunpcklqdq {{.*#+}} xmm3 = xmm4[0],xmm3[0] ; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm4 = [0,8,0,8,u,u,1,9,1,9,u,u,2,10,2,10,u,u,3,11,3,11,u,u,4,12,4,12,u,u,5,13] ; AVX2-FP-NEXT: vpshufb %ymm4, %ymm2, %ymm5 ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,3,0,1] ; AVX2-FP-NEXT: vpshufb %ymm4, %ymm2, %ymm2 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm2 = ymm5[0],ymm2[1],ymm5[2,3],ymm2[4],ymm5[5,6],ymm2[7],ymm5[8],ymm2[9],ymm5[10,11],ymm2[12],ymm5[13,14],ymm2[15] ; AVX2-FP-NEXT: vpshufb {{.*#+}} xmm4 = xmm3[2,10,1,9,0,8,3,11,u,u,u,u,4,12,u,u] ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,0,1] ; AVX2-FP-NEXT: vpmovsxbw {{.*#+}} ymm5 = [65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535] ; AVX2-FP-NEXT: vpblendvb %ymm5, %ymm2, %ymm4, %ymm2 ; AVX2-FP-NEXT: vpshufb {{.*#+}} xmm0 = zero,zero,xmm0[u,u,6,14],zero,zero,xmm0[u,u,7,15],zero,zero,xmm0[u,u] ; AVX2-FP-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[5,13,u,u],zero,zero,xmm1[6,14,u,u],zero,zero,xmm1[7,15,u,u] ; AVX2-FP-NEXT: vpor %xmm0, %xmm1, %xmm0 ; AVX2-FP-NEXT: vpshufb {{.*#+}} xmm1 = xmm3[u,u,5,13,u,u,u,u,6,14,u,u,u,u,7,15] ; AVX2-FP-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0],xmm1[1],xmm0[2,3],xmm1[4],xmm0[5,6],xmm1[7] ; AVX2-FP-NEXT: vmovdqa %xmm0, 32(%rax) ; AVX2-FP-NEXT: vmovdqa %ymm2, (%rax) ; AVX2-FP-NEXT: vzeroupper ; AVX2-FP-NEXT: retq ; ; AVX2-FCP-LABEL: store_i8_stride6_vf8: ; AVX2-FCP: # %bb.0: ; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX2-FCP-NEXT: vmovq {{.*#+}} xmm0 = mem[0],zero ; AVX2-FCP-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero ; AVX2-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0] ; AVX2-FCP-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero ; AVX2-FCP-NEXT: vmovq {{.*#+}} xmm2 = mem[0],zero ; AVX2-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm1 = xmm2[0],xmm1[0] ; AVX2-FCP-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm2 ; AVX2-FCP-NEXT: vmovq {{.*#+}} xmm3 = mem[0],zero ; AVX2-FCP-NEXT: vmovq {{.*#+}} xmm4 = mem[0],zero ; AVX2-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm3 = xmm4[0],xmm3[0] ; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm4 = [0,8,0,8,u,u,1,9,1,9,u,u,2,10,2,10,u,u,3,11,3,11,u,u,4,12,4,12,u,u,5,13] ; AVX2-FCP-NEXT: vpshufb %ymm4, %ymm2, %ymm5 ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,3,0,1] ; AVX2-FCP-NEXT: vpshufb %ymm4, %ymm2, %ymm2 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm5[0],ymm2[1],ymm5[2,3],ymm2[4],ymm5[5,6],ymm2[7],ymm5[8],ymm2[9],ymm5[10,11],ymm2[12],ymm5[13,14],ymm2[15] ; AVX2-FCP-NEXT: vpshufb {{.*#+}} xmm4 = xmm3[2,10,1,9,0,8,3,11,u,u,u,u,4,12,u,u] ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,0,1] ; AVX2-FCP-NEXT: vpmovsxbw {{.*#+}} ymm5 = [65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535] ; AVX2-FCP-NEXT: vpblendvb %ymm5, %ymm2, %ymm4, %ymm2 ; AVX2-FCP-NEXT: vpshufb {{.*#+}} xmm0 = zero,zero,xmm0[u,u,6,14],zero,zero,xmm0[u,u,7,15],zero,zero,xmm0[u,u] ; AVX2-FCP-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[5,13,u,u],zero,zero,xmm1[6,14,u,u],zero,zero,xmm1[7,15,u,u] ; AVX2-FCP-NEXT: vpor %xmm0, %xmm1, %xmm0 ; AVX2-FCP-NEXT: vpshufb {{.*#+}} xmm1 = xmm3[u,u,5,13,u,u,u,u,6,14,u,u,u,u,7,15] ; AVX2-FCP-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0],xmm1[1],xmm0[2,3],xmm1[4],xmm0[5,6],xmm1[7] ; AVX2-FCP-NEXT: vmovdqa %xmm0, 32(%rax) ; AVX2-FCP-NEXT: vmovdqa %ymm2, (%rax) ; AVX2-FCP-NEXT: vzeroupper ; AVX2-FCP-NEXT: retq ; ; AVX512-LABEL: store_i8_stride6_vf8: ; AVX512: # %bb.0: ; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512-NEXT: vmovq {{.*#+}} xmm0 = mem[0],zero ; AVX512-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero ; AVX512-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0] ; AVX512-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero ; AVX512-NEXT: vmovq {{.*#+}} xmm2 = mem[0],zero ; AVX512-NEXT: vpunpcklqdq {{.*#+}} xmm1 = xmm2[0],xmm1[0] ; AVX512-NEXT: vmovq {{.*#+}} xmm2 = mem[0],zero ; AVX512-NEXT: vmovq {{.*#+}} xmm3 = mem[0],zero ; AVX512-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm3[0],xmm2[0] ; AVX512-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm3 ; AVX512-NEXT: vmovdqa {{.*#+}} ymm4 = [0,8,0,8,u,u,1,9,1,9,u,u,2,10,2,10,u,u,3,11,3,11,u,u,4,12,4,12,u,u,5,13] ; AVX512-NEXT: vpshufb %ymm4, %ymm3, %ymm5 ; AVX512-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,3,0,1] ; AVX512-NEXT: vpshufb %ymm4, %ymm3, %ymm3 ; AVX512-NEXT: vpblendw {{.*#+}} ymm3 = ymm5[0],ymm3[1],ymm5[2,3],ymm3[4],ymm5[5,6],ymm3[7],ymm5[8],ymm3[9],ymm5[10,11],ymm3[12],ymm5[13,14],ymm3[15] ; AVX512-NEXT: vpshufb {{.*#+}} xmm4 = xmm2[2,10,1,9,0,8,3,11,u,u,u,u,4,12,u,u] ; AVX512-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,0,1] ; AVX512-NEXT: vpternlogq {{.*#+}} ymm4 = ymm4 ^ (mem & (ymm4 ^ ymm3)) ; AVX512-NEXT: vpshufb {{.*#+}} xmm0 = zero,zero,xmm0[u,u,6,14],zero,zero,xmm0[u,u,7,15],zero,zero,xmm0[u,u] ; AVX512-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[5,13,u,u],zero,zero,xmm1[6,14,u,u],zero,zero,xmm1[7,15,u,u] ; AVX512-NEXT: vpor %xmm0, %xmm1, %xmm0 ; AVX512-NEXT: vpshufb {{.*#+}} xmm1 = xmm2[u,u,5,13,u,u,u,u,6,14,u,u,u,u,7,15] ; AVX512-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0],xmm1[1],xmm0[2,3],xmm1[4],xmm0[5,6],xmm1[7] ; AVX512-NEXT: vinserti32x4 $2, %xmm0, %zmm4, %zmm1 ; AVX512-NEXT: vmovdqa %xmm0, 32(%rax) ; AVX512-NEXT: vmovdqa %ymm1, (%rax) ; AVX512-NEXT: vzeroupper ; AVX512-NEXT: retq ; ; AVX512-FCP-LABEL: store_i8_stride6_vf8: ; AVX512-FCP: # %bb.0: ; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512-FCP-NEXT: vmovq {{.*#+}} xmm0 = mem[0],zero ; AVX512-FCP-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0] ; AVX512-FCP-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero ; AVX512-FCP-NEXT: vmovq {{.*#+}} xmm2 = mem[0],zero ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm1 = xmm2[0],xmm1[0] ; AVX512-FCP-NEXT: vmovq {{.*#+}} xmm2 = mem[0],zero ; AVX512-FCP-NEXT: vmovq {{.*#+}} xmm3 = mem[0],zero ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm3[0],xmm2[0] ; AVX512-FCP-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm3 ; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm4 = [0,8,0,8,u,u,1,9,1,9,u,u,2,10,2,10,u,u,3,11,3,11,u,u,4,12,4,12,u,u,5,13] ; AVX512-FCP-NEXT: vpshufb %ymm4, %ymm3, %ymm5 ; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,3,0,1] ; AVX512-FCP-NEXT: vpshufb %ymm4, %ymm3, %ymm3 ; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm5[0],ymm3[1],ymm5[2,3],ymm3[4],ymm5[5,6],ymm3[7],ymm5[8],ymm3[9],ymm5[10,11],ymm3[12],ymm5[13,14],ymm3[15] ; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm4 = xmm2[2,10,1,9,0,8,3,11,u,u,u,u,4,12,u,u] ; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,0,1] ; AVX512-FCP-NEXT: vpternlogq {{.*#+}} ymm4 = ymm4 ^ (mem & (ymm4 ^ ymm3)) ; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm0 = zero,zero,xmm0[u,u,6,14],zero,zero,xmm0[u,u,7,15],zero,zero,xmm0[u,u] ; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[5,13,u,u],zero,zero,xmm1[6,14,u,u],zero,zero,xmm1[7,15,u,u] ; AVX512-FCP-NEXT: vpor %xmm0, %xmm1, %xmm0 ; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm1 = xmm2[u,u,5,13,u,u,u,u,6,14,u,u,u,u,7,15] ; AVX512-FCP-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0],xmm1[1],xmm0[2,3],xmm1[4],xmm0[5,6],xmm1[7] ; AVX512-FCP-NEXT: vinserti32x4 $2, %xmm0, %zmm4, %zmm1 ; AVX512-FCP-NEXT: vmovdqa %xmm0, 32(%rax) ; AVX512-FCP-NEXT: vmovdqa %ymm1, (%rax) ; AVX512-FCP-NEXT: vzeroupper ; AVX512-FCP-NEXT: retq ; ; AVX512DQ-LABEL: store_i8_stride6_vf8: ; AVX512DQ: # %bb.0: ; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512DQ-NEXT: vmovq {{.*#+}} xmm0 = mem[0],zero ; AVX512DQ-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0] ; AVX512DQ-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero ; AVX512DQ-NEXT: vmovq {{.*#+}} xmm2 = mem[0],zero ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} xmm1 = xmm2[0],xmm1[0] ; AVX512DQ-NEXT: vmovq {{.*#+}} xmm2 = mem[0],zero ; AVX512DQ-NEXT: vmovq {{.*#+}} xmm3 = mem[0],zero ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm3[0],xmm2[0] ; AVX512DQ-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm3 ; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm4 = [0,8,0,8,u,u,1,9,1,9,u,u,2,10,2,10,u,u,3,11,3,11,u,u,4,12,4,12,u,u,5,13] ; AVX512DQ-NEXT: vpshufb %ymm4, %ymm3, %ymm5 ; AVX512DQ-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,3,0,1] ; AVX512DQ-NEXT: vpshufb %ymm4, %ymm3, %ymm3 ; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm3 = ymm5[0],ymm3[1],ymm5[2,3],ymm3[4],ymm5[5,6],ymm3[7],ymm5[8],ymm3[9],ymm5[10,11],ymm3[12],ymm5[13,14],ymm3[15] ; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm4 = xmm2[2,10,1,9,0,8,3,11,u,u,u,u,4,12,u,u] ; AVX512DQ-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,0,1] ; AVX512DQ-NEXT: vpternlogq {{.*#+}} ymm4 = ymm4 ^ (mem & (ymm4 ^ ymm3)) ; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm0 = zero,zero,xmm0[u,u,6,14],zero,zero,xmm0[u,u,7,15],zero,zero,xmm0[u,u] ; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[5,13,u,u],zero,zero,xmm1[6,14,u,u],zero,zero,xmm1[7,15,u,u] ; AVX512DQ-NEXT: vpor %xmm0, %xmm1, %xmm0 ; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm1 = xmm2[u,u,5,13,u,u,u,u,6,14,u,u,u,u,7,15] ; AVX512DQ-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0],xmm1[1],xmm0[2,3],xmm1[4],xmm0[5,6],xmm1[7] ; AVX512DQ-NEXT: vinserti32x4 $2, %xmm0, %zmm4, %zmm1 ; AVX512DQ-NEXT: vmovdqa %xmm0, 32(%rax) ; AVX512DQ-NEXT: vmovdqa %ymm1, (%rax) ; AVX512DQ-NEXT: vzeroupper ; AVX512DQ-NEXT: retq ; ; AVX512DQ-FCP-LABEL: store_i8_stride6_vf8: ; AVX512DQ-FCP: # %bb.0: ; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512DQ-FCP-NEXT: vmovq {{.*#+}} xmm0 = mem[0],zero ; AVX512DQ-FCP-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0] ; AVX512DQ-FCP-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero ; AVX512DQ-FCP-NEXT: vmovq {{.*#+}} xmm2 = mem[0],zero ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm1 = xmm2[0],xmm1[0] ; AVX512DQ-FCP-NEXT: vmovq {{.*#+}} xmm2 = mem[0],zero ; AVX512DQ-FCP-NEXT: vmovq {{.*#+}} xmm3 = mem[0],zero ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm3[0],xmm2[0] ; AVX512DQ-FCP-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm3 ; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm4 = [0,8,0,8,u,u,1,9,1,9,u,u,2,10,2,10,u,u,3,11,3,11,u,u,4,12,4,12,u,u,5,13] ; AVX512DQ-FCP-NEXT: vpshufb %ymm4, %ymm3, %ymm5 ; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,3,0,1] ; AVX512DQ-FCP-NEXT: vpshufb %ymm4, %ymm3, %ymm3 ; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm5[0],ymm3[1],ymm5[2,3],ymm3[4],ymm5[5,6],ymm3[7],ymm5[8],ymm3[9],ymm5[10,11],ymm3[12],ymm5[13,14],ymm3[15] ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm4 = xmm2[2,10,1,9,0,8,3,11,u,u,u,u,4,12,u,u] ; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,0,1] ; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} ymm4 = ymm4 ^ (mem & (ymm4 ^ ymm3)) ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm0 = zero,zero,xmm0[u,u,6,14],zero,zero,xmm0[u,u,7,15],zero,zero,xmm0[u,u] ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[5,13,u,u],zero,zero,xmm1[6,14,u,u],zero,zero,xmm1[7,15,u,u] ; AVX512DQ-FCP-NEXT: vpor %xmm0, %xmm1, %xmm0 ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm1 = xmm2[u,u,5,13,u,u,u,u,6,14,u,u,u,u,7,15] ; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0],xmm1[1],xmm0[2,3],xmm1[4],xmm0[5,6],xmm1[7] ; AVX512DQ-FCP-NEXT: vinserti32x4 $2, %xmm0, %zmm4, %zmm1 ; AVX512DQ-FCP-NEXT: vmovdqa %xmm0, 32(%rax) ; AVX512DQ-FCP-NEXT: vmovdqa %ymm1, (%rax) ; AVX512DQ-FCP-NEXT: vzeroupper ; AVX512DQ-FCP-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: vmovdqa {{.*#+}} ymm4 = [0,8,0,8,u,u,1,9,1,9,u,u,2,10,2,10,u,u,3,11,3,11,u,u,4,12,4,12,u,u,5,13] ; AVX512BW-NEXT: vpshufb %ymm4, %ymm3, %ymm5 ; AVX512BW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,3,0,1] ; AVX512BW-NEXT: vpshufb %ymm4, %ymm3, %ymm3 ; AVX512BW-NEXT: vpblendw {{.*#+}} ymm3 = ymm5[0],ymm3[1],ymm5[2,3],ymm3[4],ymm5[5,6],ymm3[7],ymm5[8],ymm3[9],ymm5[10,11],ymm3[12],ymm5[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 ; ; AVX512BW-FCP-LABEL: store_i8_stride6_vf8: ; AVX512BW-FCP: # %bb.0: ; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512BW-FCP-NEXT: vmovq {{.*#+}} xmm0 = mem[0],zero ; AVX512BW-FCP-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0] ; AVX512BW-FCP-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero ; AVX512BW-FCP-NEXT: vmovq {{.*#+}} xmm2 = mem[0],zero ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm1 = xmm2[0],xmm1[0] ; AVX512BW-FCP-NEXT: vmovq {{.*#+}} xmm2 = mem[0],zero ; AVX512BW-FCP-NEXT: vmovq {{.*#+}} xmm3 = mem[0],zero ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm3[0],xmm2[0] ; AVX512BW-FCP-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm3 ; AVX512BW-FCP-NEXT: vmovdqa {{.*#+}} ymm4 = [0,8,0,8,u,u,1,9,1,9,u,u,2,10,2,10,u,u,3,11,3,11,u,u,4,12,4,12,u,u,5,13] ; AVX512BW-FCP-NEXT: vpshufb %ymm4, %ymm3, %ymm5 ; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,3,0,1] ; AVX512BW-FCP-NEXT: vpshufb %ymm4, %ymm3, %ymm3 ; AVX512BW-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm5[0],ymm3[1],ymm5[2,3],ymm3[4],ymm5[5,6],ymm3[7],ymm5[8],ymm3[9],ymm5[10,11],ymm3[12],ymm5[13,14],ymm3[15] ; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm4 = xmm2[2,10,1,9,0,8,3,11,u,u,u,u,4,12,u,u] ; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,0,1] ; AVX512BW-FCP-NEXT: movw $18724, %cx # imm = 0x4924 ; AVX512BW-FCP-NEXT: kmovd %ecx, %k1 ; AVX512BW-FCP-NEXT: vmovdqu16 %ymm4, %ymm3 {%k1} ; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm0 = zero,zero,xmm0[u,u,6,14],zero,zero,xmm0[u,u,7,15],zero,zero,xmm0[u,u] ; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[5,13,u,u],zero,zero,xmm1[6,14,u,u],zero,zero,xmm1[7,15,u,u] ; AVX512BW-FCP-NEXT: vpor %xmm0, %xmm1, %xmm0 ; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm1 = xmm2[u,u,5,13,u,u,u,u,6,14,u,u,u,u,7,15] ; AVX512BW-FCP-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0],xmm1[1],xmm0[2,3],xmm1[4],xmm0[5,6],xmm1[7] ; AVX512BW-FCP-NEXT: vinserti32x4 $2, %xmm0, %zmm3, %zmm1 ; AVX512BW-FCP-NEXT: vmovdqa %xmm0, 32(%rax) ; AVX512BW-FCP-NEXT: vmovdqa %ymm1, (%rax) ; AVX512BW-FCP-NEXT: vzeroupper ; AVX512BW-FCP-NEXT: retq ; ; AVX512DQ-BW-LABEL: store_i8_stride6_vf8: ; AVX512DQ-BW: # %bb.0: ; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512DQ-BW-NEXT: vmovq {{.*#+}} xmm0 = mem[0],zero ; AVX512DQ-BW-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0] ; AVX512DQ-BW-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero ; AVX512DQ-BW-NEXT: vmovq {{.*#+}} xmm2 = mem[0],zero ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} xmm1 = xmm2[0],xmm1[0] ; AVX512DQ-BW-NEXT: vmovq {{.*#+}} xmm2 = mem[0],zero ; AVX512DQ-BW-NEXT: vmovq {{.*#+}} xmm3 = mem[0],zero ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm3[0],xmm2[0] ; AVX512DQ-BW-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm3 ; AVX512DQ-BW-NEXT: vmovdqa {{.*#+}} ymm4 = [0,8,0,8,u,u,1,9,1,9,u,u,2,10,2,10,u,u,3,11,3,11,u,u,4,12,4,12,u,u,5,13] ; AVX512DQ-BW-NEXT: vpshufb %ymm4, %ymm3, %ymm5 ; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,3,0,1] ; AVX512DQ-BW-NEXT: vpshufb %ymm4, %ymm3, %ymm3 ; AVX512DQ-BW-NEXT: vpblendw {{.*#+}} ymm3 = ymm5[0],ymm3[1],ymm5[2,3],ymm3[4],ymm5[5,6],ymm3[7],ymm5[8],ymm3[9],ymm5[10,11],ymm3[12],ymm5[13,14],ymm3[15] ; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm4 = xmm2[2,10,1,9,0,8,3,11,u,u,u,u,4,12,u,u] ; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,0,1] ; AVX512DQ-BW-NEXT: movw $18724, %cx # imm = 0x4924 ; AVX512DQ-BW-NEXT: kmovd %ecx, %k1 ; AVX512DQ-BW-NEXT: vmovdqu16 %ymm4, %ymm3 {%k1} ; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm0 = zero,zero,xmm0[u,u,6,14],zero,zero,xmm0[u,u,7,15],zero,zero,xmm0[u,u] ; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[5,13,u,u],zero,zero,xmm1[6,14,u,u],zero,zero,xmm1[7,15,u,u] ; AVX512DQ-BW-NEXT: vpor %xmm0, %xmm1, %xmm0 ; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} xmm1 = xmm2[u,u,5,13,u,u,u,u,6,14,u,u,u,u,7,15] ; AVX512DQ-BW-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0],xmm1[1],xmm0[2,3],xmm1[4],xmm0[5,6],xmm1[7] ; AVX512DQ-BW-NEXT: vinserti32x4 $2, %xmm0, %zmm3, %zmm1 ; AVX512DQ-BW-NEXT: vmovdqa %xmm0, 32(%rax) ; AVX512DQ-BW-NEXT: vmovdqa %ymm1, (%rax) ; AVX512DQ-BW-NEXT: vzeroupper ; AVX512DQ-BW-NEXT: retq ; ; AVX512DQ-BW-FCP-LABEL: store_i8_stride6_vf8: ; AVX512DQ-BW-FCP: # %bb.0: ; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512DQ-BW-FCP-NEXT: vmovq {{.*#+}} xmm0 = mem[0],zero ; AVX512DQ-BW-FCP-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0] ; AVX512DQ-BW-FCP-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero ; AVX512DQ-BW-FCP-NEXT: vmovq {{.*#+}} xmm2 = mem[0],zero ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm1 = xmm2[0],xmm1[0] ; AVX512DQ-BW-FCP-NEXT: vmovq {{.*#+}} xmm2 = mem[0],zero ; AVX512DQ-BW-FCP-NEXT: vmovq {{.*#+}} xmm3 = mem[0],zero ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm3[0],xmm2[0] ; AVX512DQ-BW-FCP-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm3 ; AVX512DQ-BW-FCP-NEXT: vmovdqa {{.*#+}} ymm4 = [0,8,0,8,u,u,1,9,1,9,u,u,2,10,2,10,u,u,3,11,3,11,u,u,4,12,4,12,u,u,5,13] ; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm4, %ymm3, %ymm5 ; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,3,0,1] ; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm4, %ymm3, %ymm3 ; AVX512DQ-BW-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm5[0],ymm3[1],ymm5[2,3],ymm3[4],ymm5[5,6],ymm3[7],ymm5[8],ymm3[9],ymm5[10,11],ymm3[12],ymm5[13,14],ymm3[15] ; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm4 = xmm2[2,10,1,9,0,8,3,11,u,u,u,u,4,12,u,u] ; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,0,1] ; AVX512DQ-BW-FCP-NEXT: movw $18724, %cx # imm = 0x4924 ; AVX512DQ-BW-FCP-NEXT: kmovd %ecx, %k1 ; AVX512DQ-BW-FCP-NEXT: vmovdqu16 %ymm4, %ymm3 {%k1} ; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm0 = zero,zero,xmm0[u,u,6,14],zero,zero,xmm0[u,u,7,15],zero,zero,xmm0[u,u] ; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[5,13,u,u],zero,zero,xmm1[6,14,u,u],zero,zero,xmm1[7,15,u,u] ; AVX512DQ-BW-FCP-NEXT: vpor %xmm0, %xmm1, %xmm0 ; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} xmm1 = xmm2[u,u,5,13,u,u,u,u,6,14,u,u,u,u,7,15] ; AVX512DQ-BW-FCP-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0],xmm1[1],xmm0[2,3],xmm1[4],xmm0[5,6],xmm1[7] ; AVX512DQ-BW-FCP-NEXT: vinserti32x4 $2, %xmm0, %zmm3, %zmm1 ; AVX512DQ-BW-FCP-NEXT: vmovdqa %xmm0, 32(%rax) ; AVX512DQ-BW-FCP-NEXT: vmovdqa %ymm1, (%rax) ; AVX512DQ-BW-FCP-NEXT: vzeroupper ; AVX512DQ-BW-FCP-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> %2 = shufflevector <8 x i8> %in.vec2, <8 x i8> %in.vec3, <16 x i32> %3 = shufflevector <8 x i8> %in.vec4, <8 x i8> %in.vec5, <16 x i32> %4 = shufflevector <16 x i8> %1, <16 x i8> %2, <32 x i32> %5 = shufflevector <16 x i8> %3, <16 x i8> poison, <32 x i32> %6 = shufflevector <32 x i8> %4, <32 x i8> %5, <48 x i32> %interleaved.vec = shufflevector <48 x i8> %6, <48 x i8> poison, <48 x i32> 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), %xmm12 ; SSE-NEXT: movdqa (%rsi), %xmm8 ; SSE-NEXT: movdqa (%rdx), %xmm13 ; SSE-NEXT: movdqa (%rcx), %xmm2 ; SSE-NEXT: movdqa (%r8), %xmm11 ; SSE-NEXT: movdqa (%r9), %xmm10 ; SSE-NEXT: movdqa %xmm12, %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 %xmm11, %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, %xmm9 ; SSE-NEXT: pandn %xmm0, %xmm9 ; SSE-NEXT: por %xmm3, %xmm9 ; SSE-NEXT: movdqa {{.*#+}} xmm3 = [255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255] ; SSE-NEXT: pand %xmm3, %xmm9 ; SSE-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm10[0],xmm0[1],xmm10[1],xmm0[2],xmm10[2],xmm0[3],xmm10[3],xmm0[4],xmm10[4],xmm0[5],xmm10[5],xmm0[6],xmm10[6],xmm0[7],xmm10[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 %xmm9, %xmm0 ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: punpckhbw {{.*#+}} xmm12 = 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] ; SSE-NEXT: pshufd {{.*#+}} xmm8 = xmm12[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, %xmm9 ; SSE-NEXT: pandn %xmm2, %xmm9 ; SSE-NEXT: por %xmm8, %xmm9 ; SSE-NEXT: movdqa {{.*#+}} xmm8 = [0,65535,65535,0,65535,65535,0,65535] ; SSE-NEXT: pand %xmm8, %xmm9 ; SSE-NEXT: punpckhbw {{.*#+}} xmm11 = xmm11[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15] ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm11[1,1,2,2] ; SSE-NEXT: movdqa %xmm8, %xmm14 ; SSE-NEXT: pandn %xmm2, %xmm14 ; SSE-NEXT: por %xmm9, %xmm14 ; SSE-NEXT: movdqa {{.*#+}} xmm9 = [255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255] ; SSE-NEXT: punpckhbw {{.*#+}} xmm2 = xmm2[8],xmm10[8],xmm2[9],xmm10[9],xmm2[10],xmm10[10],xmm2[11],xmm10[11],xmm2[12],xmm10[12],xmm2[13],xmm10[13],xmm2[14],xmm10[14],xmm2[15],xmm10[15] ; SSE-NEXT: pshufd {{.*#+}} xmm15 = xmm2[1,1,2,2] ; SSE-NEXT: movdqa %xmm9, %xmm10 ; SSE-NEXT: pandn %xmm15, %xmm10 ; SSE-NEXT: pand %xmm9, %xmm14 ; SSE-NEXT: por %xmm14, %xmm10 ; SSE-NEXT: pshufd {{.*#+}} xmm14 = xmm12[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 = xmm11[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 = xmm12[0,0,1,1] ; SSE-NEXT: pand %xmm4, %xmm0 ; SSE-NEXT: por %xmm0, %xmm5 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm11[0,0,0,0] ; SSE-NEXT: movdqa %xmm1, %xmm11 ; SSE-NEXT: pandn %xmm0, %xmm11 ; SSE-NEXT: pand %xmm1, %xmm5 ; SSE-NEXT: por %xmm5, %xmm11 ; SSE-NEXT: pand %xmm3, %xmm11 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm2[0,0,0,0] ; SSE-NEXT: pandn %xmm0, %xmm3 ; SSE-NEXT: por %xmm11, %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), %xmm11 # 16-byte Reload ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm11[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 %xmm9, %xmm8 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm11[1,1,2,2] ; SSE-NEXT: pandn %xmm0, %xmm9 ; SSE-NEXT: por %xmm8, %xmm9 ; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax ; SSE-NEXT: movdqa %xmm9, 16(%rax) ; SSE-NEXT: movdqa %xmm14, 32(%rax) ; SSE-NEXT: movdqa %xmm3, 48(%rax) ; SSE-NEXT: movdqa %xmm15, 80(%rax) ; SSE-NEXT: movdqa %xmm10, 64(%rax) ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload ; SSE-NEXT: movaps %xmm0, (%rax) ; SSE-NEXT: retq ; ; AVX-LABEL: store_i8_stride6_vf16: ; AVX: # %bb.0: ; AVX-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX-NEXT: vmovdqa (%rdi), %xmm1 ; AVX-NEXT: vmovdqa (%rsi), %xmm2 ; AVX-NEXT: vmovdqa (%rdx), %xmm3 ; AVX-NEXT: vmovdqa (%rcx), %xmm4 ; AVX-NEXT: vmovdqa (%r8), %xmm5 ; AVX-NEXT: vmovdqa (%r9), %xmm6 ; AVX-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] ; AVX-NEXT: vpshufd {{.*#+}} xmm0 = xmm7[1,1,2,2] ; AVX-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] ; AVX-NEXT: vpshuflw {{.*#+}} xmm9 = xmm8[3,3,3,3,4,5,6,7] ; AVX-NEXT: vpshufhw {{.*#+}} xmm9 = xmm9[0,1,2,3,4,4,4,4] ; AVX-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0,1],xmm9[2],xmm0[3,4],xmm9[5],xmm0[6,7] ; AVX-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] ; AVX-NEXT: vpshufd {{.*#+}} xmm10 = xmm9[1,1,2,2] ; AVX-NEXT: vpblendw {{.*#+}} xmm0 = xmm10[0],xmm0[1,2],xmm10[3],xmm0[4,5],xmm10[6],xmm0[7] ; AVX-NEXT: vpshufd {{.*#+}} xmm10 = xmm7[0,0,1,1] ; AVX-NEXT: vpshuflw {{.*#+}} xmm11 = xmm8[1,0,2,2,4,5,6,7] ; AVX-NEXT: vpshufd {{.*#+}} xmm11 = xmm11[0,1,0,1] ; AVX-NEXT: vpblendw {{.*#+}} xmm10 = xmm10[0],xmm11[1],xmm10[2,3],xmm11[4],xmm10[5,6],xmm11[7] ; AVX-NEXT: vpshufd {{.*#+}} xmm11 = xmm9[0,0,0,0] ; AVX-NEXT: vpblendw {{.*#+}} xmm10 = xmm10[0,1],xmm11[2],xmm10[3,4],xmm11[5],xmm10[6,7] ; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm10, %ymm0 ; AVX-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] ; AVX-NEXT: vpshufd {{.*#+}} xmm2 = xmm1[0,0,1,1] ; AVX-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] ; AVX-NEXT: vpshuflw {{.*#+}} xmm4 = xmm3[1,0,2,2,4,5,6,7] ; AVX-NEXT: vpshufd {{.*#+}} xmm4 = xmm4[0,1,0,1] ; AVX-NEXT: vpblendw {{.*#+}} xmm2 = xmm2[0],xmm4[1],xmm2[2,3],xmm4[4],xmm2[5,6],xmm4[7] ; AVX-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] ; AVX-NEXT: vpshufd {{.*#+}} xmm5 = xmm4[0,0,0,0] ; AVX-NEXT: vpblendw {{.*#+}} xmm2 = xmm2[0,1],xmm5[2],xmm2[3,4],xmm5[5],xmm2[6,7] ; AVX-NEXT: vpshufd {{.*#+}} xmm5 = xmm7[3,3,3,3] ; AVX-NEXT: vpshufhw {{.*#+}} xmm6 = xmm8[0,1,2,3,5,6,7,7] ; AVX-NEXT: vpshufd {{.*#+}} xmm6 = xmm6[2,2,2,3] ; AVX-NEXT: vpblendw {{.*#+}} xmm5 = xmm6[0,1],xmm5[2],xmm6[3,4],xmm5[5],xmm6[6,7] ; AVX-NEXT: vpshufd {{.*#+}} xmm6 = xmm9[2,2,3,3] ; AVX-NEXT: vpblendw {{.*#+}} xmm5 = xmm5[0],xmm6[1],xmm5[2,3],xmm6[4],xmm5[5,6],xmm6[7] ; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm5, %ymm2 ; AVX-NEXT: vpshufd {{.*#+}} xmm5 = xmm1[3,3,3,3] ; AVX-NEXT: vpshufhw {{.*#+}} xmm6 = xmm3[0,1,2,3,5,6,7,7] ; AVX-NEXT: vpshufd {{.*#+}} xmm6 = xmm6[2,2,2,3] ; AVX-NEXT: vpblendw {{.*#+}} xmm5 = xmm6[0,1],xmm5[2],xmm6[3,4],xmm5[5],xmm6[6,7] ; AVX-NEXT: vpshufd {{.*#+}} xmm6 = xmm4[2,2,3,3] ; AVX-NEXT: vpblendw {{.*#+}} xmm5 = xmm5[0],xmm6[1],xmm5[2,3],xmm6[4],xmm5[5,6],xmm6[7] ; AVX-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[1,1,2,2] ; AVX-NEXT: vpshuflw {{.*#+}} xmm3 = xmm3[3,3,3,3,4,5,6,7] ; AVX-NEXT: vpshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,4,4,4] ; AVX-NEXT: vpblendw {{.*#+}} xmm1 = xmm1[0,1],xmm3[2],xmm1[3,4],xmm3[5],xmm1[6,7] ; AVX-NEXT: vpshufd {{.*#+}} xmm3 = xmm4[1,1,2,2] ; AVX-NEXT: vpblendw {{.*#+}} xmm1 = xmm3[0],xmm1[1,2],xmm3[3],xmm1[4,5],xmm3[6],xmm1[7] ; AVX-NEXT: vinsertf128 $1, %xmm5, %ymm1, %ymm1 ; AVX-NEXT: vmovaps %ymm1, 64(%rax) ; AVX-NEXT: vmovaps %ymm2, 32(%rax) ; AVX-NEXT: vmovaps %ymm0, (%rax) ; AVX-NEXT: vzeroupper ; AVX-NEXT: retq ; ; AVX2-LABEL: store_i8_stride6_vf16: ; AVX2: # %bb.0: ; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX2-NEXT: vmovdqa (%rdi), %xmm0 ; AVX2-NEXT: vmovdqa (%rdx), %xmm1 ; AVX2-NEXT: vmovdqa (%r8), %xmm2 ; AVX2-NEXT: vinserti128 $1, (%rsi), %ymm0, %ymm0 ; AVX2-NEXT: vinserti128 $1, (%rcx), %ymm1, %ymm1 ; AVX2-NEXT: vinserti128 $1, (%r9), %ymm2, %ymm2 ; AVX2-NEXT: vpermq {{.*#+}} ymm3 = ymm1[0,2,0,2] ; AVX2-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-NEXT: vpermq {{.*#+}} ymm4 = ymm0[0,2,0,2] ; AVX2-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-NEXT: vpor %ymm3, %ymm4, %ymm3 ; AVX2-NEXT: vpermq {{.*#+}} ymm4 = ymm2[0,2,0,2] ; AVX2-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-NEXT: vpmovsxbw {{.*#+}} ymm5 = [65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535] ; AVX2-NEXT: vpblendvb %ymm5, %ymm3, %ymm4, %ymm3 ; AVX2-NEXT: vpermq {{.*#+}} ymm4 = ymm0[0,2,1,3] ; AVX2-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-NEXT: vpermq {{.*#+}} ymm5 = ymm1[0,2,1,3] ; AVX2-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-NEXT: vpor %ymm4, %ymm5, %ymm4 ; AVX2-NEXT: vpermq {{.*#+}} ymm5 = ymm2[0,2,1,3] ; AVX2-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-NEXT: vpmovsxbw {{.*#+}} ymm6 = [65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535] ; AVX2-NEXT: vpblendvb %ymm6, %ymm4, %ymm5, %ymm4 ; AVX2-NEXT: vpermq {{.*#+}} ymm1 = ymm1[1,3,1,3] ; AVX2-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-NEXT: vpermq {{.*#+}} ymm0 = ymm0[1,3,1,3] ; AVX2-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-NEXT: vpor %ymm1, %ymm0, %ymm0 ; AVX2-NEXT: vpermq {{.*#+}} ymm1 = ymm2[1,3,1,3] ; AVX2-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-NEXT: vpmovsxbw {{.*#+}} ymm2 = [0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0] ; AVX2-NEXT: vpblendvb %ymm2, %ymm0, %ymm1, %ymm0 ; AVX2-NEXT: vmovdqa %ymm0, 64(%rax) ; AVX2-NEXT: vmovdqa %ymm4, 32(%rax) ; AVX2-NEXT: vmovdqa %ymm3, (%rax) ; AVX2-NEXT: vzeroupper ; AVX2-NEXT: retq ; ; AVX2-FP-LABEL: store_i8_stride6_vf16: ; AVX2-FP: # %bb.0: ; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX2-FP-NEXT: vmovdqa (%rdi), %xmm0 ; AVX2-FP-NEXT: vmovdqa (%rdx), %xmm1 ; AVX2-FP-NEXT: vmovdqa (%r8), %xmm2 ; AVX2-FP-NEXT: vinserti128 $1, (%rsi), %ymm0, %ymm0 ; AVX2-FP-NEXT: vinserti128 $1, (%rcx), %ymm1, %ymm1 ; AVX2-FP-NEXT: vinserti128 $1, (%r9), %ymm2, %ymm2 ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm3 = ymm1[0,2,0,2] ; AVX2-FP-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-FP-NEXT: vpermq {{.*#+}} ymm4 = ymm0[0,2,0,2] ; AVX2-FP-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-FP-NEXT: vpor %ymm3, %ymm4, %ymm3 ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm4 = ymm2[0,2,0,2] ; AVX2-FP-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-FP-NEXT: vpmovsxbw {{.*#+}} ymm5 = [65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535] ; AVX2-FP-NEXT: vpblendvb %ymm5, %ymm3, %ymm4, %ymm3 ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm4 = ymm0[0,2,1,3] ; AVX2-FP-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-FP-NEXT: vpermq {{.*#+}} ymm5 = ymm1[0,2,1,3] ; AVX2-FP-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-FP-NEXT: vpor %ymm4, %ymm5, %ymm4 ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm5 = ymm2[0,2,1,3] ; AVX2-FP-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-FP-NEXT: vpmovsxbw {{.*#+}} ymm6 = [65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535] ; AVX2-FP-NEXT: vpblendvb %ymm6, %ymm4, %ymm5, %ymm4 ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[1,3,1,3] ; AVX2-FP-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-FP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[1,3,1,3] ; AVX2-FP-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-FP-NEXT: vpor %ymm1, %ymm0, %ymm0 ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm1 = ymm2[1,3,1,3] ; AVX2-FP-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-FP-NEXT: vpmovsxbw {{.*#+}} ymm2 = [0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0] ; AVX2-FP-NEXT: vpblendvb %ymm2, %ymm0, %ymm1, %ymm0 ; AVX2-FP-NEXT: vmovdqa %ymm0, 64(%rax) ; AVX2-FP-NEXT: vmovdqa %ymm4, 32(%rax) ; AVX2-FP-NEXT: vmovdqa %ymm3, (%rax) ; AVX2-FP-NEXT: vzeroupper ; AVX2-FP-NEXT: retq ; ; AVX2-FCP-LABEL: store_i8_stride6_vf16: ; AVX2-FCP: # %bb.0: ; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX2-FCP-NEXT: vmovdqa (%rdi), %xmm0 ; AVX2-FCP-NEXT: vmovdqa (%rdx), %xmm1 ; AVX2-FCP-NEXT: vmovdqa (%r8), %xmm2 ; AVX2-FCP-NEXT: vinserti128 $1, (%rsi), %ymm0, %ymm0 ; AVX2-FCP-NEXT: vinserti128 $1, (%rcx), %ymm1, %ymm1 ; AVX2-FCP-NEXT: vinserti128 $1, (%r9), %ymm2, %ymm2 ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm1[0,2,0,2] ; AVX2-FCP-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-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm0[0,2,0,2] ; AVX2-FCP-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-FCP-NEXT: vpor %ymm3, %ymm4, %ymm3 ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm2[0,2,0,2] ; AVX2-FCP-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-FCP-NEXT: vpmovsxbw {{.*#+}} ymm5 = [65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535] ; AVX2-FCP-NEXT: vpblendvb %ymm5, %ymm3, %ymm4, %ymm3 ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm0[0,2,1,3] ; AVX2-FCP-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-FCP-NEXT: vpermq {{.*#+}} ymm5 = ymm1[0,2,1,3] ; AVX2-FCP-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-FCP-NEXT: vpor %ymm4, %ymm5, %ymm4 ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm5 = ymm2[0,2,1,3] ; AVX2-FCP-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-FCP-NEXT: vpmovsxbw {{.*#+}} ymm6 = [65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535] ; AVX2-FCP-NEXT: vpblendvb %ymm6, %ymm4, %ymm5, %ymm4 ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[1,3,1,3] ; AVX2-FCP-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-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[1,3,1,3] ; AVX2-FCP-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-FCP-NEXT: vpor %ymm1, %ymm0, %ymm0 ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm2[1,3,1,3] ; AVX2-FCP-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-FCP-NEXT: vpmovsxbw {{.*#+}} ymm2 = [0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0] ; AVX2-FCP-NEXT: vpblendvb %ymm2, %ymm0, %ymm1, %ymm0 ; AVX2-FCP-NEXT: vmovdqa %ymm0, 64(%rax) ; AVX2-FCP-NEXT: vmovdqa %ymm4, 32(%rax) ; AVX2-FCP-NEXT: vmovdqa %ymm3, (%rax) ; AVX2-FCP-NEXT: vzeroupper ; AVX2-FCP-NEXT: retq ; ; AVX512-LABEL: store_i8_stride6_vf16: ; AVX512: # %bb.0: ; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512-NEXT: vmovdqa (%rdi), %xmm0 ; AVX512-NEXT: vmovdqa (%rdx), %xmm1 ; AVX512-NEXT: vmovdqa (%r8), %xmm2 ; AVX512-NEXT: vinserti128 $1, (%rcx), %ymm1, %ymm1 ; AVX512-NEXT: vpermq {{.*#+}} ymm3 = ymm1[0,2,0,2] ; AVX512-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 ; AVX512-NEXT: vinserti128 $1, (%rsi), %ymm0, %ymm0 ; AVX512-NEXT: vpermq {{.*#+}} ymm4 = ymm0[0,2,1,3] ; AVX512-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 ; AVX512-NEXT: vinserti64x4 $1, %ymm4, %zmm3, %zmm3 ; AVX512-NEXT: vpermq {{.*#+}} ymm4 = ymm0[0,2,0,2] ; AVX512-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] ; AVX512-NEXT: vpermq {{.*#+}} ymm5 = ymm1[0,2,1,3] ; AVX512-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] ; AVX512-NEXT: vinserti64x4 $1, %ymm5, %zmm4, %zmm4 ; AVX512-NEXT: vporq %zmm3, %zmm4, %zmm3 ; AVX512-NEXT: vinserti128 $1, (%r9), %ymm2, %ymm2 ; AVX512-NEXT: vpermq {{.*#+}} ymm4 = ymm2[0,2,0,2] ; AVX512-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 ; AVX512-NEXT: vpermq {{.*#+}} ymm5 = ymm2[0,2,1,3] ; AVX512-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 ; AVX512-NEXT: vinserti64x4 $1, %ymm5, %zmm4, %zmm4 ; AVX512-NEXT: vpternlogq {{.*#+}} zmm4 = zmm4 | (zmm3 & mem) ; AVX512-NEXT: vpermq {{.*#+}} ymm1 = ymm1[1,3,1,3] ; AVX512-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] ; AVX512-NEXT: vpermq {{.*#+}} ymm0 = ymm0[1,3,1,3] ; AVX512-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] ; AVX512-NEXT: vpor %ymm1, %ymm0, %ymm0 ; AVX512-NEXT: vpermq {{.*#+}} ymm1 = ymm2[1,3,1,3] ; AVX512-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] ; AVX512-NEXT: vpternlogq {{.*#+}} ymm1 = ymm1 | (ymm0 & mem) ; AVX512-NEXT: vmovdqa %ymm1, 64(%rax) ; AVX512-NEXT: vmovdqa64 %zmm4, (%rax) ; AVX512-NEXT: vzeroupper ; AVX512-NEXT: retq ; ; AVX512-FCP-LABEL: store_i8_stride6_vf16: ; AVX512-FCP: # %bb.0: ; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512-FCP-NEXT: vmovdqa (%rdi), %xmm0 ; AVX512-FCP-NEXT: vmovdqa (%rdx), %xmm1 ; AVX512-FCP-NEXT: vmovdqa (%r8), %xmm2 ; AVX512-FCP-NEXT: vinserti128 $1, (%rcx), %ymm1, %ymm1 ; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm1[0,2,0,2] ; AVX512-FCP-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 ; AVX512-FCP-NEXT: vinserti128 $1, (%rsi), %ymm0, %ymm0 ; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm0[0,2,1,3] ; AVX512-FCP-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 ; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm3, %zmm3 ; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm0[0,2,0,2] ; AVX512-FCP-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] ; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm5 = ymm1[0,2,1,3] ; AVX512-FCP-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] ; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm5, %zmm4, %zmm4 ; AVX512-FCP-NEXT: vporq %zmm3, %zmm4, %zmm3 ; AVX512-FCP-NEXT: vinserti128 $1, (%r9), %ymm2, %ymm2 ; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm2[0,2,0,2] ; AVX512-FCP-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 ; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm5 = ymm2[0,2,1,3] ; AVX512-FCP-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 ; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm5, %zmm4, %zmm4 ; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm4 = zmm4 | (zmm3 & mem) ; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[1,3,1,3] ; AVX512-FCP-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] ; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[1,3,1,3] ; AVX512-FCP-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] ; AVX512-FCP-NEXT: vpor %ymm1, %ymm0, %ymm0 ; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm2[1,3,1,3] ; AVX512-FCP-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] ; AVX512-FCP-NEXT: vpternlogq {{.*#+}} ymm1 = ymm1 | (ymm0 & mem) ; AVX512-FCP-NEXT: vmovdqa %ymm1, 64(%rax) ; AVX512-FCP-NEXT: vmovdqa64 %zmm4, (%rax) ; AVX512-FCP-NEXT: vzeroupper ; AVX512-FCP-NEXT: retq ; ; AVX512DQ-LABEL: store_i8_stride6_vf16: ; AVX512DQ: # %bb.0: ; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512DQ-NEXT: vmovdqa (%rdi), %xmm0 ; AVX512DQ-NEXT: vmovdqa (%rdx), %xmm1 ; AVX512DQ-NEXT: vmovdqa (%r8), %xmm2 ; AVX512DQ-NEXT: vinserti128 $1, (%rcx), %ymm1, %ymm1 ; AVX512DQ-NEXT: vpermq {{.*#+}} ymm3 = ymm1[0,2,0,2] ; AVX512DQ-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 ; AVX512DQ-NEXT: vinserti128 $1, (%rsi), %ymm0, %ymm0 ; AVX512DQ-NEXT: vpermq {{.*#+}} ymm4 = ymm0[0,2,1,3] ; AVX512DQ-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 ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm4, %zmm3, %zmm3 ; AVX512DQ-NEXT: vpermq {{.*#+}} ymm4 = ymm0[0,2,0,2] ; AVX512DQ-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] ; AVX512DQ-NEXT: vpermq {{.*#+}} ymm5 = ymm1[0,2,1,3] ; AVX512DQ-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] ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm5, %zmm4, %zmm4 ; AVX512DQ-NEXT: vporq %zmm3, %zmm4, %zmm3 ; AVX512DQ-NEXT: vinserti128 $1, (%r9), %ymm2, %ymm2 ; AVX512DQ-NEXT: vpermq {{.*#+}} ymm4 = ymm2[0,2,0,2] ; AVX512DQ-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 ; AVX512DQ-NEXT: vpermq {{.*#+}} ymm5 = ymm2[0,2,1,3] ; AVX512DQ-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 ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm5, %zmm4, %zmm4 ; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm4 = zmm4 | (zmm3 & mem) ; AVX512DQ-NEXT: vpermq {{.*#+}} ymm1 = ymm1[1,3,1,3] ; AVX512DQ-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] ; AVX512DQ-NEXT: vpermq {{.*#+}} ymm0 = ymm0[1,3,1,3] ; AVX512DQ-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] ; AVX512DQ-NEXT: vpor %ymm1, %ymm0, %ymm0 ; AVX512DQ-NEXT: vpermq {{.*#+}} ymm1 = ymm2[1,3,1,3] ; AVX512DQ-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] ; AVX512DQ-NEXT: vpternlogq {{.*#+}} ymm1 = ymm1 | (ymm0 & mem) ; AVX512DQ-NEXT: vmovdqa %ymm1, 64(%rax) ; AVX512DQ-NEXT: vmovdqa64 %zmm4, (%rax) ; AVX512DQ-NEXT: vzeroupper ; AVX512DQ-NEXT: retq ; ; AVX512DQ-FCP-LABEL: store_i8_stride6_vf16: ; AVX512DQ-FCP: # %bb.0: ; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %xmm0 ; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %xmm1 ; AVX512DQ-FCP-NEXT: vmovdqa (%r8), %xmm2 ; AVX512DQ-FCP-NEXT: vinserti128 $1, (%rcx), %ymm1, %ymm1 ; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm1[0,2,0,2] ; AVX512DQ-FCP-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 ; AVX512DQ-FCP-NEXT: vinserti128 $1, (%rsi), %ymm0, %ymm0 ; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm0[0,2,1,3] ; AVX512DQ-FCP-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 ; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm3, %zmm3 ; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm0[0,2,0,2] ; AVX512DQ-FCP-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] ; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm5 = ymm1[0,2,1,3] ; AVX512DQ-FCP-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] ; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm5, %zmm4, %zmm4 ; AVX512DQ-FCP-NEXT: vporq %zmm3, %zmm4, %zmm3 ; AVX512DQ-FCP-NEXT: vinserti128 $1, (%r9), %ymm2, %ymm2 ; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm2[0,2,0,2] ; AVX512DQ-FCP-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 ; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm5 = ymm2[0,2,1,3] ; AVX512DQ-FCP-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 ; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm5, %zmm4, %zmm4 ; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm4 = zmm4 | (zmm3 & mem) ; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[1,3,1,3] ; AVX512DQ-FCP-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] ; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[1,3,1,3] ; AVX512DQ-FCP-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] ; AVX512DQ-FCP-NEXT: vpor %ymm1, %ymm0, %ymm0 ; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm2[1,3,1,3] ; AVX512DQ-FCP-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] ; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} ymm1 = ymm1 | (ymm0 & mem) ; AVX512DQ-FCP-NEXT: vmovdqa %ymm1, 64(%rax) ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm4, (%rax) ; AVX512DQ-FCP-NEXT: vzeroupper ; AVX512DQ-FCP-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: vpermq {{.*#+}} ymm3 = ymm0[0,2,0,2] ; AVX512BW-NEXT: vinserti128 $1, (%rcx), %ymm1, %ymm1 ; AVX512BW-NEXT: vpermq {{.*#+}} ymm4 = ymm1[0,2,1,3] ; AVX512BW-NEXT: vinserti64x4 $1, %ymm4, %zmm3, %zmm3 ; AVX512BW-NEXT: vpshufb {{.*#+}} zmm3 = zmm3[0,8],zero,zero,zero,zero,zmm3[1,9],zero,zero,zero,zero,zmm3[2,10],zero,zero,zero,zero,zmm3[19,27],zero,zero,zero,zero,zmm3[20,28],zero,zero,zero,zero,zmm3[21,29,37,45],zero,zero,zero,zero,zmm3[38,46],zero,zero,zero,zero,zmm3[39,47],zero,zero,zero,zero,zmm3[48,56],zero,zero,zero,zero,zmm3[49,57],zero,zero,zero,zero,zmm3[50,58] ; AVX512BW-NEXT: vpermq {{.*#+}} ymm4 = ymm1[0,2,0,2] ; AVX512BW-NEXT: vpermq {{.*#+}} ymm5 = ymm0[0,2,1,3] ; AVX512BW-NEXT: vinserti64x4 $1, %ymm5, %zmm4, %zmm4 ; AVX512BW-NEXT: vpshufb {{.*#+}} zmm4 = zero,zero,zmm4[0,8],zero,zero,zero,zero,zmm4[1,9],zero,zero,zero,zero,zmm4[2,10],zero,zero,zero,zero,zmm4[19,27],zero,zero,zero,zero,zmm4[20,28],zero,zero,zero,zero,zero,zero,zero,zero,zmm4[38,46],zero,zero,zero,zero,zmm4[39,47],zero,zero,zero,zero,zmm4[48,56],zero,zero,zero,zero,zmm4[49,57],zero,zero,zero,zero,zmm4[50,58],zero,zero ; AVX512BW-NEXT: vporq %zmm3, %zmm4, %zmm3 ; AVX512BW-NEXT: vinserti128 $1, (%r9), %ymm2, %ymm2 ; AVX512BW-NEXT: vpermq {{.*#+}} ymm4 = ymm2[0,2,0,2] ; AVX512BW-NEXT: vpermq {{.*#+}} ymm5 = ymm2[0,2,1,3] ; AVX512BW-NEXT: vinserti64x4 $1, %ymm5, %zmm4, %zmm4 ; AVX512BW-NEXT: vpshufb {{.*#+}} zmm4 = zmm4[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 %zmm4, %zmm3 {%k1} ; AVX512BW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[1,3,1,3] ; AVX512BW-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 ; AVX512BW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[1,3,1,3] ; AVX512BW-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 ; AVX512BW-NEXT: vpor %ymm0, %ymm1, %ymm0 ; 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 %zmm3, (%rax) ; AVX512BW-NEXT: vzeroupper ; AVX512BW-NEXT: retq ; ; AVX512BW-FCP-LABEL: store_i8_stride6_vf16: ; AVX512BW-FCP: # %bb.0: ; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512BW-FCP-NEXT: vmovdqa (%rdi), %xmm0 ; AVX512BW-FCP-NEXT: vmovdqa (%rdx), %xmm1 ; AVX512BW-FCP-NEXT: vmovdqa (%r8), %xmm2 ; AVX512BW-FCP-NEXT: vinserti128 $1, (%rcx), %ymm1, %ymm1 ; AVX512BW-FCP-NEXT: vinserti128 $1, (%rsi), %ymm0, %ymm0 ; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm3 = [0,2,0,2,8,10,9,11] ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm4 ; AVX512BW-FCP-NEXT: vpermt2q %zmm1, %zmm3, %zmm4 ; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} zmm4 = zmm4[0,8],zero,zero,zero,zero,zmm4[1,9],zero,zero,zero,zero,zmm4[2,10],zero,zero,zero,zero,zmm4[19,27],zero,zero,zero,zero,zmm4[20,28],zero,zero,zero,zero,zmm4[21,29,37,45],zero,zero,zero,zero,zmm4[38,46],zero,zero,zero,zero,zmm4[39,47],zero,zero,zero,zero,zmm4[48,56],zero,zero,zero,zero,zmm4[49,57],zero,zero,zero,zero,zmm4[50,58] ; AVX512BW-FCP-NEXT: vpermi2q %zmm0, %zmm1, %zmm3 ; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} zmm3 = zero,zero,zmm3[0,8],zero,zero,zero,zero,zmm3[1,9],zero,zero,zero,zero,zmm3[2,10],zero,zero,zero,zero,zmm3[19,27],zero,zero,zero,zero,zmm3[20,28],zero,zero,zero,zero,zero,zero,zero,zero,zmm3[38,46],zero,zero,zero,zero,zmm3[39,47],zero,zero,zero,zero,zmm3[48,56],zero,zero,zero,zero,zmm3[49,57],zero,zero,zero,zero,zmm3[50,58],zero,zero ; AVX512BW-FCP-NEXT: vporq %zmm4, %zmm3, %zmm3 ; AVX512BW-FCP-NEXT: vinserti128 $1, (%r9), %ymm2, %ymm2 ; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm4 = [0,2,0,2,0,2,1,3] ; AVX512BW-FCP-NEXT: vpermq %zmm2, %zmm4, %zmm4 ; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} zmm4 = zmm4[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-FCP-NEXT: movl $613566756, %ecx # imm = 0x24924924 ; AVX512BW-FCP-NEXT: kmovd %ecx, %k1 ; AVX512BW-FCP-NEXT: vmovdqu16 %zmm4, %zmm3 {%k1} ; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[1,3,1,3] ; AVX512BW-FCP-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 ; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[1,3,1,3] ; AVX512BW-FCP-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 ; AVX512BW-FCP-NEXT: vpor %ymm0, %ymm1, %ymm0 ; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm2[1,3,1,3] ; AVX512BW-FCP-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-FCP-NEXT: movw $-28087, %cx # imm = 0x9249 ; AVX512BW-FCP-NEXT: kmovd %ecx, %k1 ; AVX512BW-FCP-NEXT: vmovdqu16 %ymm1, %ymm0 {%k1} ; AVX512BW-FCP-NEXT: vmovdqa %ymm0, 64(%rax) ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, (%rax) ; AVX512BW-FCP-NEXT: vzeroupper ; AVX512BW-FCP-NEXT: retq ; ; AVX512DQ-BW-LABEL: store_i8_stride6_vf16: ; AVX512DQ-BW: # %bb.0: ; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512DQ-BW-NEXT: vmovdqa (%rdi), %xmm0 ; AVX512DQ-BW-NEXT: vmovdqa (%rdx), %xmm1 ; AVX512DQ-BW-NEXT: vmovdqa (%r8), %xmm2 ; AVX512DQ-BW-NEXT: vinserti128 $1, (%rsi), %ymm0, %ymm0 ; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm3 = ymm0[0,2,0,2] ; AVX512DQ-BW-NEXT: vinserti128 $1, (%rcx), %ymm1, %ymm1 ; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm4 = ymm1[0,2,1,3] ; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm4, %zmm3, %zmm3 ; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} zmm3 = zmm3[0,8],zero,zero,zero,zero,zmm3[1,9],zero,zero,zero,zero,zmm3[2,10],zero,zero,zero,zero,zmm3[19,27],zero,zero,zero,zero,zmm3[20,28],zero,zero,zero,zero,zmm3[21,29,37,45],zero,zero,zero,zero,zmm3[38,46],zero,zero,zero,zero,zmm3[39,47],zero,zero,zero,zero,zmm3[48,56],zero,zero,zero,zero,zmm3[49,57],zero,zero,zero,zero,zmm3[50,58] ; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm4 = ymm1[0,2,0,2] ; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm5 = ymm0[0,2,1,3] ; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm5, %zmm4, %zmm4 ; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} zmm4 = zero,zero,zmm4[0,8],zero,zero,zero,zero,zmm4[1,9],zero,zero,zero,zero,zmm4[2,10],zero,zero,zero,zero,zmm4[19,27],zero,zero,zero,zero,zmm4[20,28],zero,zero,zero,zero,zero,zero,zero,zero,zmm4[38,46],zero,zero,zero,zero,zmm4[39,47],zero,zero,zero,zero,zmm4[48,56],zero,zero,zero,zero,zmm4[49,57],zero,zero,zero,zero,zmm4[50,58],zero,zero ; AVX512DQ-BW-NEXT: vporq %zmm3, %zmm4, %zmm3 ; AVX512DQ-BW-NEXT: vinserti128 $1, (%r9), %ymm2, %ymm2 ; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm4 = ymm2[0,2,0,2] ; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm5 = ymm2[0,2,1,3] ; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm5, %zmm4, %zmm4 ; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} zmm4 = zmm4[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] ; AVX512DQ-BW-NEXT: movl $613566756, %ecx # imm = 0x24924924 ; AVX512DQ-BW-NEXT: kmovd %ecx, %k1 ; AVX512DQ-BW-NEXT: vmovdqu16 %zmm4, %zmm3 {%k1} ; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[1,3,1,3] ; AVX512DQ-BW-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 ; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[1,3,1,3] ; AVX512DQ-BW-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 ; AVX512DQ-BW-NEXT: vpor %ymm0, %ymm1, %ymm0 ; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm1 = ymm2[1,3,1,3] ; AVX512DQ-BW-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] ; AVX512DQ-BW-NEXT: movw $-28087, %cx # imm = 0x9249 ; AVX512DQ-BW-NEXT: kmovd %ecx, %k1 ; AVX512DQ-BW-NEXT: vmovdqu16 %ymm1, %ymm0 {%k1} ; AVX512DQ-BW-NEXT: vmovdqa %ymm0, 64(%rax) ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, (%rax) ; AVX512DQ-BW-NEXT: vzeroupper ; AVX512DQ-BW-NEXT: retq ; ; AVX512DQ-BW-FCP-LABEL: store_i8_stride6_vf16: ; AVX512DQ-BW-FCP: # %bb.0: ; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdi), %xmm0 ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdx), %xmm1 ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%r8), %xmm2 ; AVX512DQ-BW-FCP-NEXT: vinserti128 $1, (%rcx), %ymm1, %ymm1 ; AVX512DQ-BW-FCP-NEXT: vinserti128 $1, (%rsi), %ymm0, %ymm0 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm3 = [0,2,0,2,8,10,9,11] ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm4 ; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm1, %zmm3, %zmm4 ; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} zmm4 = zmm4[0,8],zero,zero,zero,zero,zmm4[1,9],zero,zero,zero,zero,zmm4[2,10],zero,zero,zero,zero,zmm4[19,27],zero,zero,zero,zero,zmm4[20,28],zero,zero,zero,zero,zmm4[21,29,37,45],zero,zero,zero,zero,zmm4[38,46],zero,zero,zero,zero,zmm4[39,47],zero,zero,zero,zero,zmm4[48,56],zero,zero,zero,zero,zmm4[49,57],zero,zero,zero,zero,zmm4[50,58] ; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm0, %zmm1, %zmm3 ; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} zmm3 = zero,zero,zmm3[0,8],zero,zero,zero,zero,zmm3[1,9],zero,zero,zero,zero,zmm3[2,10],zero,zero,zero,zero,zmm3[19,27],zero,zero,zero,zero,zmm3[20,28],zero,zero,zero,zero,zero,zero,zero,zero,zmm3[38,46],zero,zero,zero,zero,zmm3[39,47],zero,zero,zero,zero,zmm3[48,56],zero,zero,zero,zero,zmm3[49,57],zero,zero,zero,zero,zmm3[50,58],zero,zero ; AVX512DQ-BW-FCP-NEXT: vporq %zmm4, %zmm3, %zmm3 ; AVX512DQ-BW-FCP-NEXT: vinserti128 $1, (%r9), %ymm2, %ymm2 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm4 = [0,2,0,2,0,2,1,3] ; AVX512DQ-BW-FCP-NEXT: vpermq %zmm2, %zmm4, %zmm4 ; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} zmm4 = zmm4[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] ; AVX512DQ-BW-FCP-NEXT: movl $613566756, %ecx # imm = 0x24924924 ; AVX512DQ-BW-FCP-NEXT: kmovd %ecx, %k1 ; AVX512DQ-BW-FCP-NEXT: vmovdqu16 %zmm4, %zmm3 {%k1} ; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[1,3,1,3] ; AVX512DQ-BW-FCP-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 ; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[1,3,1,3] ; AVX512DQ-BW-FCP-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 ; AVX512DQ-BW-FCP-NEXT: vpor %ymm0, %ymm1, %ymm0 ; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm2[1,3,1,3] ; AVX512DQ-BW-FCP-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] ; AVX512DQ-BW-FCP-NEXT: movw $-28087, %cx # imm = 0x9249 ; AVX512DQ-BW-FCP-NEXT: kmovd %ecx, %k1 ; AVX512DQ-BW-FCP-NEXT: vmovdqu16 %ymm1, %ymm0 {%k1} ; AVX512DQ-BW-FCP-NEXT: vmovdqa %ymm0, 64(%rax) ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, (%rax) ; AVX512DQ-BW-FCP-NEXT: vzeroupper ; AVX512DQ-BW-FCP-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> %2 = shufflevector <16 x i8> %in.vec2, <16 x i8> %in.vec3, <32 x i32> %3 = shufflevector <16 x i8> %in.vec4, <16 x i8> %in.vec5, <32 x i32> %4 = shufflevector <32 x i8> %1, <32 x i8> %2, <64 x i32> %5 = shufflevector <32 x i8> %3, <32 x i8> poison, <64 x i32> %6 = shufflevector <64 x i8> %4, <64 x i8> %5, <96 x i32> %interleaved.vec = shufflevector <96 x i8> %6, <96 x i8> poison, <96 x i32> 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 {{.*#+}} xmm3 = [65535,65535,0,65535,65535,0,65535,65535] ; SSE-NEXT: pand %xmm3, %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 {{.*#+}} xmm2 = xmm1[0,0,0,0] ; SSE-NEXT: movdqa %xmm3, %xmm7 ; SSE-NEXT: pandn %xmm2, %xmm7 ; SSE-NEXT: por %xmm6, %xmm7 ; SSE-NEXT: movdqa {{.*#+}} xmm2 = [255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255] ; SSE-NEXT: pand %xmm2, %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 %xmm2, %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 %xmm3, %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 %xmm3, %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 {{.*#+}} xmm10 = [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 %xmm10, %xmm1 ; SSE-NEXT: pandn %xmm0, %xmm1 ; SSE-NEXT: movdqa (%rdi), %xmm0 ; SSE-NEXT: pand %xmm10, %xmm6 ; SSE-NEXT: por %xmm6, %xmm1 ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: movdqa (%rsi), %xmm14 ; SSE-NEXT: movdqa %xmm0, %xmm1 ; SSE-NEXT: movdqa %xmm0, %xmm8 ; SSE-NEXT: punpckhbw {{.*#+}} xmm1 = xmm1[8],xmm14[8],xmm1[9],xmm14[9],xmm1[10],xmm14[10],xmm1[11],xmm14[11],xmm1[12],xmm14[12],xmm1[13],xmm14[13],xmm1[14],xmm14[14],xmm1[15],xmm14[15] ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm1[3,3,3,3] ; SSE-NEXT: movdqa %xmm1, %xmm2 ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: movdqa %xmm3, %xmm6 ; SSE-NEXT: pandn %xmm0, %xmm6 ; SSE-NEXT: movdqa (%rdx), %xmm11 ; SSE-NEXT: movdqa (%rcx), %xmm12 ; SSE-NEXT: movdqa %xmm11, %xmm1 ; 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: pshufhw {{.*#+}} xmm0 = xmm1[0,1,2,3,5,6,7,7] ; SSE-NEXT: movdqa %xmm1, %xmm5 ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[2,2,2,3] ; SSE-NEXT: pand %xmm3, %xmm0 ; SSE-NEXT: por %xmm6, %xmm0 ; SSE-NEXT: movdqa (%r8), %xmm13 ; SSE-NEXT: movdqa %xmm13, %xmm7 ; SSE-NEXT: punpckhbw {{.*#+}} xmm7 = xmm7[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15] ; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm7[2,2,3,3] ; SSE-NEXT: movdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: movdqa %xmm9, %xmm15 ; SSE-NEXT: pandn %xmm6, %xmm15 ; SSE-NEXT: pand %xmm9, %xmm0 ; SSE-NEXT: por %xmm0, %xmm15 ; SSE-NEXT: movdqa (%r9), %xmm6 ; SSE-NEXT: punpckhbw {{.*#+}} xmm4 = xmm4[8],xmm6[8],xmm4[9],xmm6[9],xmm4[10],xmm6[10],xmm4[11],xmm6[11],xmm4[12],xmm6[12],xmm4[13],xmm6[13],xmm4[14],xmm6[14],xmm4[15],xmm6[15] ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm4[2,2,3,3] ; SSE-NEXT: movdqa %xmm4, (%rsp) # 16-byte Spill ; SSE-NEXT: movdqa %xmm10, %xmm1 ; SSE-NEXT: pandn %xmm0, %xmm1 ; SSE-NEXT: pand %xmm10, %xmm15 ; SSE-NEXT: por %xmm15, %xmm1 ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; 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 %xmm3, %xmm15 ; SSE-NEXT: pandn %xmm0, %xmm15 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm2[1,1,2,2] ; SSE-NEXT: pand %xmm3, %xmm0 ; SSE-NEXT: por %xmm0, %xmm15 ; SSE-NEXT: movdqa {{.*#+}} xmm1 = [0,65535,65535,0,65535,65535,0,65535] ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm7[1,1,2,2] ; SSE-NEXT: movdqa %xmm1, %xmm0 ; SSE-NEXT: pandn %xmm2, %xmm0 ; SSE-NEXT: pand %xmm1, %xmm15 ; SSE-NEXT: por %xmm15, %xmm0 ; SSE-NEXT: movdqa {{.*#+}} xmm5 = [255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255] ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm4[1,1,2,2] ; SSE-NEXT: movdqa %xmm5, %xmm15 ; SSE-NEXT: pandn %xmm2, %xmm15 ; SSE-NEXT: pand %xmm5, %xmm0 ; SSE-NEXT: por %xmm0, %xmm15 ; SSE-NEXT: movdqa %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: punpcklbw {{.*#+}} xmm8 = xmm8[0],xmm14[0],xmm8[1],xmm14[1],xmm8[2],xmm14[2],xmm8[3],xmm14[3],xmm8[4],xmm14[4],xmm8[5],xmm14[5],xmm8[6],xmm14[6],xmm8[7],xmm14[7] ; SSE-NEXT: punpcklbw {{.*#+}} xmm11 = xmm11[0],xmm12[0],xmm11[1],xmm12[1],xmm11[2],xmm12[2],xmm11[3],xmm12[3],xmm11[4],xmm12[4],xmm11[5],xmm12[5],xmm11[6],xmm12[6],xmm11[7],xmm12[7] ; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm11[3,3,3,3,4,5,6,7] ; SSE-NEXT: movdqa %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; 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 = xmm8[1,1,2,2] ; SSE-NEXT: movdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: pand %xmm3, %xmm0 ; SSE-NEXT: por %xmm0, %xmm2 ; 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, %xmm4 ; SSE-NEXT: pandn %xmm0, %xmm4 ; SSE-NEXT: pand %xmm1, %xmm2 ; SSE-NEXT: por %xmm2, %xmm4 ; SSE-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm6[0],xmm2[1],xmm6[1],xmm2[2],xmm6[2],xmm2[3],xmm6[3],xmm2[4],xmm6[4],xmm2[5],xmm6[5],xmm2[6],xmm6[6],xmm2[7],xmm6[7] ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm2[1,1,2,2] ; SSE-NEXT: movdqa %xmm2, %xmm6 ; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: movdqa %xmm5, %xmm12 ; SSE-NEXT: pandn %xmm0, %xmm12 ; SSE-NEXT: pand %xmm5, %xmm4 ; SSE-NEXT: por %xmm4, %xmm12 ; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm11[1,0,2,2,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,1,0,1] ; SSE-NEXT: movdqa %xmm9, %xmm2 ; SSE-NEXT: pandn %xmm0, %xmm2 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm8[0,0,1,1] ; SSE-NEXT: pand %xmm9, %xmm0 ; SSE-NEXT: por %xmm0, %xmm2 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm13[0,0,0,0] ; SSE-NEXT: movdqa %xmm3, %xmm4 ; SSE-NEXT: pandn %xmm0, %xmm4 ; SSE-NEXT: pand %xmm3, %xmm2 ; SSE-NEXT: por %xmm2, %xmm4 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm6[0,0,0,0] ; SSE-NEXT: movdqa {{.*#+}} xmm7 = [255,255,255,255,255,0,255,255,255,255,255,0,255,255,255,255] ; SSE-NEXT: movdqa %xmm7, %xmm11 ; SSE-NEXT: pandn %xmm0, %xmm11 ; SSE-NEXT: pand %xmm7, %xmm4 ; SSE-NEXT: por %xmm4, %xmm11 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm4[3,3,3,3] ; SSE-NEXT: movdqa %xmm3, %xmm2 ; SSE-NEXT: pandn %xmm0, %xmm2 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm13 # 16-byte Reload ; SSE-NEXT: pshufhw {{.*#+}} xmm0 = xmm13[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 %xmm2, %xmm0 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm14[2,2,3,3] ; SSE-NEXT: movdqa %xmm9, %xmm8 ; SSE-NEXT: pandn %xmm2, %xmm8 ; SSE-NEXT: pand %xmm9, %xmm0 ; SSE-NEXT: por %xmm0, %xmm8 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm15[2,2,3,3] ; SSE-NEXT: movdqa %xmm10, %xmm6 ; SSE-NEXT: pandn %xmm0, %xmm6 ; SSE-NEXT: pand %xmm10, %xmm8 ; SSE-NEXT: por %xmm8, %xmm6 ; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm13[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 = xmm4[1,1,2,2] ; SSE-NEXT: pand %xmm3, %xmm0 ; SSE-NEXT: por %xmm0, %xmm2 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm14[1,1,2,2] ; SSE-NEXT: movdqa %xmm1, %xmm8 ; SSE-NEXT: pandn %xmm0, %xmm8 ; SSE-NEXT: pand %xmm1, %xmm2 ; SSE-NEXT: por %xmm2, %xmm8 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm15[1,1,2,2] ; SSE-NEXT: movdqa %xmm5, %xmm15 ; SSE-NEXT: pandn %xmm0, %xmm15 ; SSE-NEXT: pand %xmm5, %xmm8 ; SSE-NEXT: por %xmm8, %xmm15 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload ; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm2[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, %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 %xmm3, %xmm0 ; SSE-NEXT: por %xmm0, %xmm8 ; SSE-NEXT: pand %xmm1, %xmm8 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm14[1,1,2,2] ; SSE-NEXT: pandn %xmm0, %xmm1 ; SSE-NEXT: por %xmm8, %xmm1 ; SSE-NEXT: pand %xmm5, %xmm1 ; 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, %xmm5 ; SSE-NEXT: por %xmm1, %xmm5 ; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm2[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 = xmm14[0,0,0,0] ; SSE-NEXT: movdqa %xmm3, %xmm8 ; SSE-NEXT: pandn %xmm0, %xmm8 ; SSE-NEXT: pand %xmm3, %xmm1 ; SSE-NEXT: por %xmm1, %xmm8 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm4[0,0,0,0] ; SSE-NEXT: movdqa %xmm7, %xmm0 ; SSE-NEXT: pandn %xmm1, %xmm0 ; SSE-NEXT: pand %xmm7, %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, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload ; SSE-NEXT: # xmm1 = mem[0,0,0,0] ; SSE-NEXT: movdqa %xmm3, %xmm13 ; SSE-NEXT: pandn %xmm1, %xmm13 ; SSE-NEXT: pand %xmm3, %xmm8 ; SSE-NEXT: por %xmm8, %xmm13 ; SSE-NEXT: pand %xmm7, %xmm13 ; SSE-NEXT: pshufd $0, (%rsp), %xmm1 # 16-byte Folded Reload ; SSE-NEXT: # xmm1 = mem[0,0,0,0] ; SSE-NEXT: pandn %xmm1, %xmm7 ; SSE-NEXT: por %xmm13, %xmm7 ; SSE-NEXT: pshufd $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload ; SSE-NEXT: # xmm1 = mem[3,3,3,3] ; SSE-NEXT: 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 %xmm3, %xmm8 ; SSE-NEXT: pandn %xmm1, %xmm3 ; SSE-NEXT: por %xmm8, %xmm3 ; SSE-NEXT: pand %xmm9, %xmm3 ; 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 %xmm3, %xmm9 ; SSE-NEXT: pand %xmm10, %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, %xmm10 ; SSE-NEXT: por %xmm9, %xmm10 ; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax ; SSE-NEXT: movdqa %xmm10, 32(%rax) ; SSE-NEXT: movdqa %xmm7, 48(%rax) ; SSE-NEXT: movdqa %xmm0, 96(%rax) ; SSE-NEXT: movdqa %xmm5, 112(%rax) ; SSE-NEXT: movdqa %xmm15, 160(%rax) ; SSE-NEXT: movdqa %xmm6, 176(%rax) ; SSE-NEXT: movdqa %xmm11, (%rax) ; SSE-NEXT: movdqa %xmm12, 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 ; ; AVX-LABEL: store_i8_stride6_vf32: ; AVX: # %bb.0: ; AVX-NEXT: vmovdqa 16(%rsi), %xmm0 ; AVX-NEXT: vmovdqa 16(%rdi), %xmm1 ; AVX-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] ; AVX-NEXT: vpshufd {{.*#+}} xmm2 = xmm11[3,3,3,3] ; AVX-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] ; AVX-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[0,0,1,1] ; AVX-NEXT: vinsertf128 $1, %xmm1, %ymm2, %ymm1 ; AVX-NEXT: vmovaps {{.*#+}} ymm10 = [65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535] ; AVX-NEXT: vandnps %ymm1, %ymm10, %ymm2 ; AVX-NEXT: vmovdqa 16(%rcx), %xmm1 ; AVX-NEXT: vmovdqa 16(%rdx), %xmm3 ; AVX-NEXT: vpunpcklbw {{.*#+}} xmm14 = 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] ; AVX-NEXT: vpshufhw {{.*#+}} xmm4 = xmm14[0,1,2,3,5,6,7,7] ; AVX-NEXT: vpshufd {{.*#+}} xmm4 = xmm4[2,2,2,3] ; AVX-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] ; AVX-NEXT: vpshuflw {{.*#+}} xmm3 = xmm1[1,0,2,2,4,5,6,7] ; AVX-NEXT: vpshufd {{.*#+}} xmm3 = xmm3[0,1,0,1] ; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm4, %ymm3 ; AVX-NEXT: vandps %ymm3, %ymm10, %ymm3 ; AVX-NEXT: vorps %ymm2, %ymm3, %ymm2 ; AVX-NEXT: vextractf128 $1, %ymm2, %xmm3 ; AVX-NEXT: vmovdqa 16(%r8), %xmm12 ; AVX-NEXT: vpshufb {{.*#+}} xmm4 = zero,zero,zero,zero,xmm12[8,u],zero,zero,zero,zero,xmm12[9,u],zero,zero,zero,zero ; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm3[0,1],xmm4[2],xmm3[3,4],xmm4[5],xmm3[6,7] ; AVX-NEXT: vpshufb {{.*#+}} xmm3 = xmm3[0,1,2,3,4],zero,xmm3[6,7,8,9,10],zero,xmm3[12,13,14,15] ; AVX-NEXT: vmovdqa 16(%r9), %xmm13 ; AVX-NEXT: vpshufb {{.*#+}} xmm4 = zero,zero,zero,zero,zero,xmm13[8],zero,zero,zero,zero,zero,xmm13[9],zero,zero,zero,zero ; AVX-NEXT: vpor %xmm4, %xmm3, %xmm3 ; AVX-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX-NEXT: vpshufb {{.*#+}} xmm3 = zero,zero,xmm12[5,u],zero,zero,zero,zero,xmm12[6,u],zero,zero,zero,zero,xmm12[7,u] ; AVX-NEXT: vpblendw {{.*#+}} xmm2 = xmm2[0],xmm3[1],xmm2[2,3],xmm3[4],xmm2[5,6],xmm3[7] ; AVX-NEXT: vmovdqa {{.*#+}} xmm9 = [0,1,2,128,4,5,6,7,8,128,10,11,12,13,14,128] ; AVX-NEXT: vpshufb %xmm9, %xmm2, %xmm2 ; AVX-NEXT: vpshufb {{.*#+}} xmm3 = zero,zero,zero,xmm13[5],zero,zero,zero,zero,zero,xmm13[6],zero,zero,zero,zero,zero,xmm13[7] ; AVX-NEXT: vpor %xmm3, %xmm2, %xmm2 ; AVX-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX-NEXT: vpshufd {{.*#+}} xmm2 = xmm0[1,1,2,2] ; AVX-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[3,3,3,3] ; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm2, %ymm0 ; AVX-NEXT: vandps %ymm0, %ymm10, %ymm0 ; AVX-NEXT: vpshuflw {{.*#+}} xmm2 = xmm1[3,3,3,3,4,5,6,7] ; AVX-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,4,4,4] ; AVX-NEXT: vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,5,6,7,7] ; AVX-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[2,2,2,3] ; AVX-NEXT: vinsertf128 $1, %xmm1, %ymm2, %ymm1 ; AVX-NEXT: vandnps %ymm1, %ymm10, %ymm1 ; AVX-NEXT: vorps %ymm1, %ymm0, %ymm0 ; AVX-NEXT: vpshufb {{.*#+}} xmm1 = xmm12[10,u],zero,zero,zero,zero,xmm12[11,u],zero,zero,zero,zero,xmm12[12,u],zero,zero ; AVX-NEXT: vpblendw {{.*#+}} xmm1 = xmm1[0],xmm0[1,2],xmm1[3],xmm0[4,5],xmm1[6],xmm0[7] ; AVX-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[0],zero,xmm1[2,3,4,5,6],zero,xmm1[8,9,10,11,12],zero,xmm1[14,15] ; AVX-NEXT: vpshufb {{.*#+}} xmm3 = zero,xmm13[10],zero,zero,zero,zero,zero,xmm13[11],zero,zero,zero,zero,zero,xmm13[12],zero,zero ; AVX-NEXT: vpor %xmm3, %xmm1, %xmm1 ; AVX-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX-NEXT: vextractf128 $1, %ymm0, %xmm0 ; AVX-NEXT: vmovdqa {{.*#+}} xmm2 = [128,128,13,u,128,128,128,128,14,u,128,128,128,128,15,u] ; AVX-NEXT: vpshufb %xmm2, %xmm12, %xmm1 ; AVX-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0],xmm1[1],xmm0[2,3],xmm1[4],xmm0[5,6],xmm1[7] ; AVX-NEXT: vpshufb %xmm9, %xmm0, %xmm0 ; AVX-NEXT: vpshufb {{.*#+}} xmm1 = zero,zero,zero,xmm13[13],zero,zero,zero,zero,zero,xmm13[14],zero,zero,zero,zero,zero,xmm13[15] ; AVX-NEXT: vpor %xmm1, %xmm0, %xmm0 ; AVX-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX-NEXT: vmovdqa (%rsi), %xmm8 ; AVX-NEXT: vmovdqa (%rdi), %xmm6 ; AVX-NEXT: vpunpckhbw {{.*#+}} xmm5 = 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] ; AVX-NEXT: vpshufd {{.*#+}} xmm0 = xmm5[1,1,2,2] ; AVX-NEXT: vpshufd {{.*#+}} xmm1 = xmm5[3,3,3,3] ; AVX-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0 ; AVX-NEXT: vmovdqa (%rcx), %xmm7 ; AVX-NEXT: vmovdqa (%rdx), %xmm4 ; AVX-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm4[8],xmm7[8],xmm4[9],xmm7[9],xmm4[10],xmm7[10],xmm4[11],xmm7[11],xmm4[12],xmm7[12],xmm4[13],xmm7[13],xmm4[14],xmm7[14],xmm4[15],xmm7[15] ; AVX-NEXT: vpshuflw {{.*#+}} xmm1 = xmm3[3,3,3,3,4,5,6,7] ; AVX-NEXT: vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,4,4,4] ; AVX-NEXT: vpshufhw {{.*#+}} xmm15 = xmm3[0,1,2,3,5,6,7,7] ; AVX-NEXT: vpshufd {{.*#+}} xmm15 = xmm15[2,2,2,3] ; AVX-NEXT: vinsertf128 $1, %xmm15, %ymm1, %ymm1 ; AVX-NEXT: vandps %ymm0, %ymm10, %ymm0 ; AVX-NEXT: vandnps %ymm1, %ymm10, %ymm1 ; AVX-NEXT: vorps %ymm1, %ymm0, %ymm15 ; AVX-NEXT: vmovdqa (%r8), %xmm1 ; AVX-NEXT: vpshufb %xmm2, %xmm1, %xmm0 ; AVX-NEXT: vextractf128 $1, %ymm15, %xmm2 ; AVX-NEXT: vpblendw {{.*#+}} xmm0 = xmm2[0],xmm0[1],xmm2[2,3],xmm0[4],xmm2[5,6],xmm0[7] ; AVX-NEXT: vpshufb %xmm9, %xmm0, %xmm2 ; AVX-NEXT: vmovdqa (%r9), %xmm0 ; AVX-NEXT: vpshufb {{.*#+}} xmm9 = zero,zero,zero,xmm0[13],zero,zero,zero,zero,zero,xmm0[14],zero,zero,zero,zero,zero,xmm0[15] ; AVX-NEXT: vpor %xmm2, %xmm9, %xmm2 ; AVX-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX-NEXT: vpshufb {{.*#+}} xmm2 = xmm1[10,u],zero,zero,zero,zero,xmm1[11,u],zero,zero,zero,zero,xmm1[12,u],zero,zero ; AVX-NEXT: vpblendw {{.*#+}} xmm2 = xmm2[0],xmm15[1,2],xmm2[3],xmm15[4,5],xmm2[6],xmm15[7] ; AVX-NEXT: vpshufb {{.*#+}} xmm2 = xmm2[0],zero,xmm2[2,3,4,5,6],zero,xmm2[8,9,10,11,12],zero,xmm2[14,15] ; AVX-NEXT: vpshufb {{.*#+}} xmm9 = zero,xmm0[10],zero,zero,zero,zero,zero,xmm0[11],zero,zero,zero,zero,zero,xmm0[12],zero,zero ; AVX-NEXT: vpor %xmm2, %xmm9, %xmm2 ; AVX-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX-NEXT: vpshufd {{.*#+}} xmm2 = xmm11[0,0,1,1] ; AVX-NEXT: vpshufd {{.*#+}} xmm9 = xmm11[1,1,2,2] ; AVX-NEXT: vinsertf128 $1, %xmm9, %ymm2, %ymm9 ; AVX-NEXT: vpshuflw {{.*#+}} xmm2 = xmm14[1,0,2,2,4,5,6,7] ; AVX-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[0,1,0,1] ; AVX-NEXT: vpshuflw {{.*#+}} xmm11 = xmm14[3,3,3,3,4,5,6,7] ; AVX-NEXT: vpshufhw {{.*#+}} xmm11 = xmm11[0,1,2,3,4,4,4,4] ; AVX-NEXT: vinsertf128 $1, %xmm11, %ymm2, %ymm11 ; AVX-NEXT: vmovaps {{.*#+}} ymm2 = [65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535] ; AVX-NEXT: vandps %ymm2, %ymm9, %ymm9 ; AVX-NEXT: vandnps %ymm11, %ymm2, %ymm11 ; AVX-NEXT: vorps %ymm11, %ymm9, %ymm15 ; AVX-NEXT: vpshufb {{.*#+}} xmm11 = zero,zero,zero,zero,xmm12[0,u],zero,zero,zero,zero,xmm12[1,u],zero,zero,zero,zero ; AVX-NEXT: vpblendw {{.*#+}} xmm11 = xmm15[0,1],xmm11[2],xmm15[3,4],xmm11[5],xmm15[6,7] ; AVX-NEXT: vmovdqa {{.*#+}} xmm14 = [0,1,2,3,4,128,6,7,8,9,10,128,12,13,14,15] ; AVX-NEXT: vpshufb %xmm14, %xmm11, %xmm11 ; AVX-NEXT: vpshufb {{.*#+}} xmm10 = zero,zero,zero,zero,zero,xmm13[0],zero,zero,zero,zero,zero,xmm13[1],zero,zero,zero,zero ; AVX-NEXT: vpor %xmm10, %xmm11, %xmm11 ; AVX-NEXT: vextractf128 $1, %ymm15, %xmm10 ; AVX-NEXT: vmovdqa {{.*#+}} xmm15 = [2,u,128,128,128,128,3,u,128,128,128,128,4,u,128,128] ; AVX-NEXT: vpshufb %xmm15, %xmm12, %xmm12 ; AVX-NEXT: vpblendw {{.*#+}} xmm10 = xmm12[0],xmm10[1,2],xmm12[3],xmm10[4,5],xmm12[6],xmm10[7] ; AVX-NEXT: vmovdqa {{.*#+}} xmm9 = [0,128,2,3,4,5,6,128,8,9,10,11,12,128,14,15] ; AVX-NEXT: vpshufb %xmm9, %xmm10, %xmm10 ; AVX-NEXT: vmovdqa {{.*#+}} xmm12 = [128,2,128,128,128,128,128,3,128,128,128,128,128,4,128,128] ; AVX-NEXT: vpshufb %xmm12, %xmm13, %xmm13 ; AVX-NEXT: vpor %xmm13, %xmm10, %xmm10 ; AVX-NEXT: vpunpcklbw {{.*#+}} xmm6 = xmm6[0],xmm8[0],xmm6[1],xmm8[1],xmm6[2],xmm8[2],xmm6[3],xmm8[3],xmm6[4],xmm8[4],xmm6[5],xmm8[5],xmm6[6],xmm8[6],xmm6[7],xmm8[7] ; AVX-NEXT: vpshufd {{.*#+}} xmm8 = xmm6[0,0,1,1] ; AVX-NEXT: vpshufd {{.*#+}} xmm13 = xmm6[1,1,2,2] ; AVX-NEXT: vinsertf128 $1, %xmm13, %ymm8, %ymm8 ; AVX-NEXT: vpunpcklbw {{.*#+}} xmm4 = xmm4[0],xmm7[0],xmm4[1],xmm7[1],xmm4[2],xmm7[2],xmm4[3],xmm7[3],xmm4[4],xmm7[4],xmm4[5],xmm7[5],xmm4[6],xmm7[6],xmm4[7],xmm7[7] ; AVX-NEXT: vpshuflw {{.*#+}} xmm7 = xmm4[1,0,2,2,4,5,6,7] ; AVX-NEXT: vpshufd {{.*#+}} xmm7 = xmm7[0,1,0,1] ; AVX-NEXT: vpshuflw {{.*#+}} xmm13 = xmm4[3,3,3,3,4,5,6,7] ; AVX-NEXT: vpshufhw {{.*#+}} xmm13 = xmm13[0,1,2,3,4,4,4,4] ; AVX-NEXT: vinsertf128 $1, %xmm13, %ymm7, %ymm7 ; AVX-NEXT: vandps %ymm2, %ymm8, %ymm8 ; AVX-NEXT: vandnps %ymm7, %ymm2, %ymm2 ; AVX-NEXT: vorps %ymm2, %ymm8, %ymm7 ; AVX-NEXT: vpshufb %xmm15, %xmm1, %xmm2 ; AVX-NEXT: vextractf128 $1, %ymm7, %xmm8 ; AVX-NEXT: vpblendw {{.*#+}} xmm2 = xmm2[0],xmm8[1,2],xmm2[3],xmm8[4,5],xmm2[6],xmm8[7] ; AVX-NEXT: vpshufb %xmm9, %xmm2, %xmm2 ; AVX-NEXT: vpshufb %xmm12, %xmm0, %xmm8 ; AVX-NEXT: vpor %xmm2, %xmm8, %xmm2 ; AVX-NEXT: vpshufb {{.*#+}} xmm8 = zero,zero,zero,zero,xmm1[0,u],zero,zero,zero,zero,xmm1[1,u],zero,zero,zero,zero ; AVX-NEXT: vpblendw {{.*#+}} xmm7 = xmm7[0,1],xmm8[2],xmm7[3,4],xmm8[5],xmm7[6,7] ; AVX-NEXT: vpshufb %xmm14, %xmm7, %xmm7 ; AVX-NEXT: vpshufb {{.*#+}} xmm8 = zero,zero,zero,zero,zero,xmm0[0],zero,zero,zero,zero,zero,xmm0[1],zero,zero,zero,zero ; AVX-NEXT: vpor %xmm7, %xmm8, %xmm7 ; AVX-NEXT: vpshufd {{.*#+}} xmm6 = xmm6[3,3,3,3] ; AVX-NEXT: vpshufd {{.*#+}} xmm5 = xmm5[0,0,1,1] ; AVX-NEXT: vinsertf128 $1, %xmm5, %ymm6, %ymm5 ; AVX-NEXT: vpshufhw {{.*#+}} xmm4 = xmm4[0,1,2,3,5,6,7,7] ; AVX-NEXT: vpshufd {{.*#+}} xmm4 = xmm4[2,2,2,3] ; AVX-NEXT: vpshuflw {{.*#+}} xmm3 = xmm3[1,0,2,2,4,5,6,7] ; AVX-NEXT: vpshufd {{.*#+}} xmm3 = xmm3[0,1,0,1] ; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm4, %ymm3 ; AVX-NEXT: vmovaps {{.*#+}} ymm6 = [65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535] ; AVX-NEXT: vandnps %ymm5, %ymm6, %ymm4 ; AVX-NEXT: vandps %ymm6, %ymm3, %ymm3 ; AVX-NEXT: vorps %ymm4, %ymm3, %ymm3 ; AVX-NEXT: vpshufb {{.*#+}} xmm4 = zero,zero,xmm1[5,u],zero,zero,zero,zero,xmm1[6,u],zero,zero,zero,zero,xmm1[7,u] ; AVX-NEXT: vpblendw {{.*#+}} xmm4 = xmm3[0],xmm4[1],xmm3[2,3],xmm4[4],xmm3[5,6],xmm4[7] ; AVX-NEXT: vpshufb {{.*#+}} xmm4 = xmm4[0,1,2],zero,xmm4[4,5,6,7,8],zero,xmm4[10,11,12,13,14],zero ; AVX-NEXT: vpshufb {{.*#+}} xmm5 = zero,zero,zero,xmm0[5],zero,zero,zero,zero,zero,xmm0[6],zero,zero,zero,zero,zero,xmm0[7] ; AVX-NEXT: vpor %xmm5, %xmm4, %xmm4 ; AVX-NEXT: vpshufb {{.*#+}} xmm1 = zero,zero,zero,zero,xmm1[8,u],zero,zero,zero,zero,xmm1[9,u],zero,zero,zero,zero ; AVX-NEXT: vextractf128 $1, %ymm3, %xmm3 ; AVX-NEXT: vpblendw {{.*#+}} xmm1 = xmm3[0,1],xmm1[2],xmm3[3,4],xmm1[5],xmm3[6,7] ; AVX-NEXT: vpshufb %xmm14, %xmm1, %xmm1 ; AVX-NEXT: vpshufb {{.*#+}} xmm0 = zero,zero,zero,zero,zero,xmm0[8],zero,zero,zero,zero,zero,xmm0[9],zero,zero,zero,zero ; AVX-NEXT: vpor %xmm0, %xmm1, %xmm0 ; AVX-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX-NEXT: vmovdqa %xmm0, 48(%rax) ; AVX-NEXT: vmovdqa %xmm4, 32(%rax) ; AVX-NEXT: vmovdqa %xmm7, (%rax) ; AVX-NEXT: vmovdqa %xmm2, 16(%rax) ; AVX-NEXT: vmovdqa %xmm10, 112(%rax) ; AVX-NEXT: vmovdqa %xmm11, 96(%rax) ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload ; AVX-NEXT: vmovaps %xmm0, 64(%rax) ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload ; AVX-NEXT: vmovaps %xmm0, 80(%rax) ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload ; AVX-NEXT: vmovaps %xmm0, 176(%rax) ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload ; AVX-NEXT: vmovaps %xmm0, 160(%rax) ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload ; AVX-NEXT: vmovaps %xmm0, 128(%rax) ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload ; AVX-NEXT: vmovaps %xmm0, 144(%rax) ; AVX-NEXT: vzeroupper ; AVX-NEXT: retq ; ; AVX2-LABEL: store_i8_stride6_vf32: ; AVX2: # %bb.0: ; AVX2-NEXT: pushq %rax ; AVX2-NEXT: vmovdqa (%rdi), %ymm1 ; AVX2-NEXT: vmovdqa (%rsi), %ymm3 ; AVX2-NEXT: vmovdqa (%rdx), %ymm0 ; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-NEXT: vmovdqa (%rcx), %ymm2 ; AVX2-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-NEXT: vmovdqa (%r8), %ymm4 ; AVX2-NEXT: vpbroadcastq {{.*#+}} ymm7 = [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-NEXT: vmovdqa (%rcx), %xmm6 ; AVX2-NEXT: vpshufb %xmm7, %xmm6, %xmm5 ; AVX2-NEXT: vmovdqa (%rdx), %xmm8 ; AVX2-NEXT: vpshufb %xmm7, %xmm8, %xmm9 ; AVX2-NEXT: vpunpcklbw {{.*#+}} xmm5 = 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] ; AVX2-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,0,0,1] ; AVX2-NEXT: vmovdqa (%rsi), %xmm11 ; AVX2-NEXT: vpbroadcastq {{.*#+}} xmm9 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0] ; AVX2-NEXT: vpshufb %xmm9, %xmm11, %xmm10 ; AVX2-NEXT: vmovdqa (%rdi), %xmm13 ; AVX2-NEXT: vpshufb %xmm9, %xmm13, %xmm9 ; AVX2-NEXT: vpunpckhbw {{.*#+}} xmm9 = xmm9[8],xmm10[8],xmm9[9],xmm10[9],xmm9[10],xmm10[10],xmm9[11],xmm10[11],xmm9[12],xmm10[12],xmm9[13],xmm10[13],xmm9[14],xmm10[14],xmm9[15],xmm10[15] ; AVX2-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,0,0,1] ; AVX2-NEXT: vpmovsxbw {{.*#+}} ymm10 = [65535,0,0,65535,0,0,65535,0,0,65535,0,0,65535,0,0,65535] ; AVX2-NEXT: vpblendvb %ymm10, %ymm5, %ymm9, %ymm9 ; AVX2-NEXT: vmovdqa (%r8), %xmm5 ; AVX2-NEXT: vpshufb {{.*#+}} xmm12 = xmm5[6,u,5,u,8,u,7,u,9,u,9,u,9,u,9,u] ; AVX2-NEXT: vpermq {{.*#+}} ymm12 = ymm12[0,0,0,1] ; AVX2-NEXT: vpmovsxbw {{.*#+}} ymm14 = [65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535] ; AVX2-NEXT: vpblendvb %ymm14, %ymm9, %ymm12, %ymm9 ; AVX2-NEXT: vpshufb %ymm7, %ymm2, %ymm12 ; AVX2-NEXT: vpshufb %ymm7, %ymm0, %ymm7 ; AVX2-NEXT: vpunpcklbw {{.*#+}} ymm12 = 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-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] ; AVX2-NEXT: vpshufb %ymm7, %ymm3, %ymm15 ; AVX2-NEXT: vmovdqa %ymm1, %ymm2 ; AVX2-NEXT: vpshufb %ymm7, %ymm1, %ymm7 ; AVX2-NEXT: vpunpcklbw {{.*#+}} ymm15 = ymm7[0],ymm15[0],ymm7[1],ymm15[1],ymm7[2],ymm15[2],ymm7[3],ymm15[3],ymm7[4],ymm15[4],ymm7[5],ymm15[5],ymm7[6],ymm15[6],ymm7[7],ymm15[7],ymm7[16],ymm15[16],ymm7[17],ymm15[17],ymm7[18],ymm15[18],ymm7[19],ymm15[19],ymm7[20],ymm15[20],ymm7[21],ymm15[21],ymm7[22],ymm15[22],ymm7[23],ymm15[23] ; AVX2-NEXT: vmovdqa (%r9), %ymm7 ; AVX2-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,2,2,3] ; AVX2-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,2,2,3] ; AVX2-NEXT: vpblendvb %ymm10, %ymm12, %ymm15, %ymm12 ; AVX2-NEXT: vmovdqa (%r9), %xmm10 ; AVX2-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-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,2,2,3] ; AVX2-NEXT: vpblendvb %ymm14, %ymm12, %ymm15, %ymm12 ; AVX2-NEXT: vpshufb {{.*#+}} xmm14 = xmm10[u,6,u,5,u,8,u,7,u,9,u,9,u,9,u,9] ; AVX2-NEXT: vpermq {{.*#+}} ymm14 = ymm14[0,0,0,1] ; AVX2-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-NEXT: vpblendvb %ymm15, %ymm9, %ymm14, %ymm0 ; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-NEXT: vpshufb {{.*#+}} ymm14 = ymm7[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-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,2,2,3] ; AVX2-NEXT: vpblendvb %ymm15, %ymm12, %ymm14, %ymm0 ; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-NEXT: vpunpcklbw {{.*#+}} xmm14 = xmm13[0],xmm11[0],xmm13[1],xmm11[1],xmm13[2],xmm11[2],xmm13[3],xmm11[3],xmm13[4],xmm11[4],xmm13[5],xmm11[5],xmm13[6],xmm11[6],xmm13[7],xmm11[7] ; AVX2-NEXT: vpshuflw {{.*#+}} xmm14 = xmm14[0,3,2,1,4,5,6,7] ; AVX2-NEXT: vpshufhw {{.*#+}} xmm14 = xmm14[0,1,2,3,4,5,6,5] ; AVX2-NEXT: vpermq {{.*#+}} ymm14 = ymm14[0,0,0,1] ; AVX2-NEXT: vpunpcklbw {{.*#+}} xmm15 = xmm8[0],xmm6[0],xmm8[1],xmm6[1],xmm8[2],xmm6[2],xmm8[3],xmm6[3],xmm8[4],xmm6[4],xmm8[5],xmm6[5],xmm8[6],xmm6[6],xmm8[7],xmm6[7] ; AVX2-NEXT: vpshuflw {{.*#+}} xmm15 = xmm15[1,0,3,2,4,5,6,7] ; AVX2-NEXT: vpshufhw {{.*#+}} xmm15 = xmm15[0,1,2,3,4,4,4,4] ; AVX2-NEXT: vpermq {{.*#+}} ymm15 = ymm15[0,0,0,1] ; AVX2-NEXT: vpmovsxbw {{.*#+}} ymm0 = [65535,0,0,65535,0,0,65535,0,0,65535,0,0,65535,0,0,65535] ; AVX2-NEXT: vpblendvb %ymm0, %ymm14, %ymm15, %ymm14 ; AVX2-NEXT: vpunpcklbw {{.*#+}} ymm15 = 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-NEXT: vmovdqa %ymm3, %ymm12 ; AVX2-NEXT: vpshuflw {{.*#+}} ymm15 = ymm15[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15] ; AVX2-NEXT: vpshufhw {{.*#+}} ymm15 = ymm15[0,1,2,3,4,5,6,5,8,9,10,11,12,13,14,13] ; AVX2-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,2,2,3] ; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload ; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload ; AVX2-NEXT: vpunpcklbw {{.*#+}} ymm9 = ymm3[0],ymm1[0],ymm3[1],ymm1[1],ymm3[2],ymm1[2],ymm3[3],ymm1[3],ymm3[4],ymm1[4],ymm3[5],ymm1[5],ymm3[6],ymm1[6],ymm3[7],ymm1[7],ymm3[16],ymm1[16],ymm3[17],ymm1[17],ymm3[18],ymm1[18],ymm3[19],ymm1[19],ymm3[20],ymm1[20],ymm3[21],ymm1[21],ymm3[22],ymm1[22],ymm3[23],ymm1[23] ; AVX2-NEXT: vpshuflw {{.*#+}} ymm9 = ymm9[1,0,3,2,4,5,6,7,9,8,11,10,12,13,14,15] ; AVX2-NEXT: vpshufhw {{.*#+}} ymm9 = ymm9[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12] ; AVX2-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,2,3] ; AVX2-NEXT: vpblendvb %ymm0, %ymm15, %ymm9, %ymm0 ; AVX2-NEXT: vpshufb {{.*#+}} xmm9 = xmm5[2,u,1,u,0,u,3,u,4,u,4,u,4,u,4,u] ; AVX2-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,0,0,1] ; AVX2-NEXT: vpmovsxbw {{.*#+}} ymm15 = [65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535] ; AVX2-NEXT: vpblendvb %ymm15, %ymm14, %ymm9, %ymm9 ; AVX2-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-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,2,2,3] ; AVX2-NEXT: vpblendvb %ymm15, %ymm0, %ymm14, %ymm0 ; AVX2-NEXT: vpshufb {{.*#+}} xmm14 = xmm10[u,2,u,1,u,0,u,3,u,4,u,4,u,4,u,4] ; AVX2-NEXT: vpermq {{.*#+}} ymm14 = ymm14[0,0,0,1] ; AVX2-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-NEXT: vpblendvb %ymm15, %ymm9, %ymm14, %ymm14 ; AVX2-NEXT: vpshufb {{.*#+}} ymm9 = ymm7[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-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,2,3] ; AVX2-NEXT: vpblendvb %ymm15, %ymm0, %ymm9, %ymm15 ; AVX2-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm13[8],xmm11[8],xmm13[9],xmm11[9],xmm13[10],xmm11[10],xmm13[11],xmm11[11],xmm13[12],xmm11[12],xmm13[13],xmm11[13],xmm13[14],xmm11[14],xmm13[15],xmm11[15] ; AVX2-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[8,9,6,7,12,13,10,11,14,15,14,15,14,15,14,15] ; AVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,0,0,1] ; AVX2-NEXT: vpunpckhbw {{.*#+}} xmm6 = xmm8[8],xmm6[8],xmm8[9],xmm6[9],xmm8[10],xmm6[10],xmm8[11],xmm6[11],xmm8[12],xmm6[12],xmm8[13],xmm6[13],xmm8[14],xmm6[14],xmm8[15],xmm6[15] ; AVX2-NEXT: vpshufb {{.*#+}} xmm6 = xmm6[10,11,8,9,6,7,12,13,14,15,14,15,14,15,14,15] ; AVX2-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,0,0,1] ; AVX2-NEXT: vpmovsxbw {{.*#+}} ymm8 = [0,65535,0,0,65535,0,0,65535,0,0,65535,0,0,65535,0,0] ; AVX2-NEXT: vpblendvb %ymm8, %ymm0, %ymm6, %ymm0 ; AVX2-NEXT: vpunpckhbw {{.*#+}} ymm2 = ymm2[8],ymm12[8],ymm2[9],ymm12[9],ymm2[10],ymm12[10],ymm2[11],ymm12[11],ymm2[12],ymm12[12],ymm2[13],ymm12[13],ymm2[14],ymm12[14],ymm2[15],ymm12[15],ymm2[24],ymm12[24],ymm2[25],ymm12[25],ymm2[26],ymm12[26],ymm2[27],ymm12[27],ymm2[28],ymm12[28],ymm2[29],ymm12[29],ymm2[30],ymm12[30],ymm2[31],ymm12[31] ; AVX2-NEXT: vpunpckhbw {{.*#+}} ymm1 = ymm3[8],ymm1[8],ymm3[9],ymm1[9],ymm3[10],ymm1[10],ymm3[11],ymm1[11],ymm3[12],ymm1[12],ymm3[13],ymm1[13],ymm3[14],ymm1[14],ymm3[15],ymm1[15],ymm3[24],ymm1[24],ymm3[25],ymm1[25],ymm3[26],ymm1[26],ymm3[27],ymm1[27],ymm3[28],ymm1[28],ymm3[29],ymm1[29],ymm3[30],ymm1[30],ymm3[31],ymm1[31] ; AVX2-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-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3] ; AVX2-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-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3] ; AVX2-NEXT: vpblendvb %ymm8, %ymm2, %ymm1, %ymm1 ; AVX2-NEXT: vpshufb {{.*#+}} xmm2 = xmm5[10,u,13,u,12,u,11,u,14,u,13,u,14,u,15,u] ; AVX2-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,0,0,1] ; AVX2-NEXT: vpmovsxbw {{.*#+}} ymm3 = [0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0] ; AVX2-NEXT: vpblendvb %ymm3, %ymm0, %ymm2, %ymm0 ; AVX2-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-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3] ; AVX2-NEXT: vpblendvb %ymm3, %ymm1, %ymm2, %ymm1 ; AVX2-NEXT: vpshufb {{.*#+}} xmm2 = xmm10[u,10,u,13,u,12,u,11,u,14,u,13,u,14,u,15] ; AVX2-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,0,0,1] ; AVX2-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-NEXT: vpblendvb %ymm3, %ymm0, %ymm2, %ymm0 ; AVX2-NEXT: vpshufb {{.*#+}} ymm2 = ymm7[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-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3] ; AVX2-NEXT: vpblendvb %ymm3, %ymm1, %ymm2, %ymm1 ; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload ; AVX2-NEXT: vmovaps %ymm2, 128(%rax) ; AVX2-NEXT: vmovdqa %ymm1, 160(%rax) ; AVX2-NEXT: vmovdqa %ymm0, 64(%rax) ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload ; AVX2-NEXT: vmovaps %ymm0, 32(%rax) ; AVX2-NEXT: vmovdqa %ymm15, 96(%rax) ; AVX2-NEXT: vmovdqa %ymm14, (%rax) ; AVX2-NEXT: popq %rax ; AVX2-NEXT: vzeroupper ; AVX2-NEXT: retq ; ; AVX2-FP-LABEL: store_i8_stride6_vf32: ; AVX2-FP: # %bb.0: ; AVX2-FP-NEXT: subq $40, %rsp ; AVX2-FP-NEXT: vmovdqa (%rdi), %ymm2 ; AVX2-FP-NEXT: vmovdqa (%rsi), %ymm6 ; AVX2-FP-NEXT: vmovdqa (%rdx), %ymm3 ; AVX2-FP-NEXT: vmovdqa (%rcx), %ymm4 ; AVX2-FP-NEXT: vmovdqa (%r8), %ymm0 ; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FP-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] ; AVX2-FP-NEXT: vmovdqa (%rcx), %xmm1 ; AVX2-FP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX2-FP-NEXT: vpshufb %xmm9, %xmm1, %xmm7 ; AVX2-FP-NEXT: vmovdqa (%rdx), %xmm1 ; AVX2-FP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX2-FP-NEXT: vpshufb %xmm9, %xmm1, %xmm8 ; AVX2-FP-NEXT: vpunpcklbw {{.*#+}} xmm7 = 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] ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,0,0,1] ; AVX2-FP-NEXT: vmovdqa (%rsi), %xmm10 ; AVX2-FP-NEXT: vpbroadcastq {{.*#+}} xmm11 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0] ; AVX2-FP-NEXT: vpshufb %xmm11, %xmm10, %xmm12 ; AVX2-FP-NEXT: vmovdqa (%rdi), %xmm5 ; AVX2-FP-NEXT: vpshufb %xmm11, %xmm5, %xmm11 ; AVX2-FP-NEXT: vpunpckhbw {{.*#+}} xmm11 = xmm11[8],xmm12[8],xmm11[9],xmm12[9],xmm11[10],xmm12[10],xmm11[11],xmm12[11],xmm11[12],xmm12[12],xmm11[13],xmm12[13],xmm11[14],xmm12[14],xmm11[15],xmm12[15] ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm11 = ymm11[0,0,0,1] ; AVX2-FP-NEXT: vpmovsxbw {{.*#+}} ymm12 = [65535,0,0,65535,0,0,65535,0,0,65535,0,0,65535,0,0,65535] ; AVX2-FP-NEXT: vpblendvb %ymm12, %ymm7, %ymm11, %ymm11 ; AVX2-FP-NEXT: vmovdqa (%r8), %xmm7 ; AVX2-FP-NEXT: vpshufb {{.*#+}} xmm13 = xmm7[6,u,5,u,8,u,7,u,9,u,9,u,9,u,9,u] ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm13 = ymm13[0,0,0,1] ; AVX2-FP-NEXT: vpmovsxbw {{.*#+}} ymm14 = [65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535] ; AVX2-FP-NEXT: vpblendvb %ymm14, %ymm11, %ymm13, %ymm11 ; AVX2-FP-NEXT: vpshufb %ymm9, %ymm4, %ymm13 ; AVX2-FP-NEXT: vpshufb %ymm9, %ymm3, %ymm9 ; AVX2-FP-NEXT: vpunpcklbw {{.*#+}} ymm13 = ymm9[0],ymm13[0],ymm9[1],ymm13[1],ymm9[2],ymm13[2],ymm9[3],ymm13[3],ymm9[4],ymm13[4],ymm9[5],ymm13[5],ymm9[6],ymm13[6],ymm9[7],ymm13[7],ymm9[16],ymm13[16],ymm9[17],ymm13[17],ymm9[18],ymm13[18],ymm9[19],ymm13[19],ymm9[20],ymm13[20],ymm9[21],ymm13[21],ymm9[22],ymm13[22],ymm9[23],ymm13[23] ; AVX2-FP-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] ; AVX2-FP-NEXT: vpshufb %ymm9, %ymm6, %ymm15 ; AVX2-FP-NEXT: vpshufb %ymm9, %ymm2, %ymm9 ; AVX2-FP-NEXT: vmovdqa %ymm2, %ymm8 ; AVX2-FP-NEXT: vpunpcklbw {{.*#+}} ymm15 = ymm9[0],ymm15[0],ymm9[1],ymm15[1],ymm9[2],ymm15[2],ymm9[3],ymm15[3],ymm9[4],ymm15[4],ymm9[5],ymm15[5],ymm9[6],ymm15[6],ymm9[7],ymm15[7],ymm9[16],ymm15[16],ymm9[17],ymm15[17],ymm9[18],ymm15[18],ymm9[19],ymm15[19],ymm9[20],ymm15[20],ymm9[21],ymm15[21],ymm9[22],ymm15[22],ymm9[23],ymm15[23] ; AVX2-FP-NEXT: vmovdqa (%r9), %ymm1 ; AVX2-FP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm13 = ymm13[2,2,2,3] ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,2,2,3] ; AVX2-FP-NEXT: vpblendvb %ymm12, %ymm13, %ymm15, %ymm13 ; AVX2-FP-NEXT: vmovdqa (%r9), %xmm12 ; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm15 = ymm0[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-FP-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,2,2,3] ; AVX2-FP-NEXT: vpblendvb %ymm14, %ymm13, %ymm15, %ymm13 ; AVX2-FP-NEXT: vpshufb {{.*#+}} xmm14 = xmm12[u,6,u,5,u,8,u,7,u,9,u,9,u,9,u,9] ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm14 = ymm14[0,0,0,1] ; AVX2-FP-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-FP-NEXT: vpblendvb %ymm15, %ymm11, %ymm14, %ymm0 ; AVX2-FP-NEXT: vmovdqu %ymm0, (%rsp) # 32-byte Spill ; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm14 = ymm1[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-FP-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,2,2,3] ; AVX2-FP-NEXT: vpblendvb %ymm15, %ymm13, %ymm14, %ymm0 ; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FP-NEXT: vpunpckhbw {{.*#+}} xmm14 = xmm5[8],xmm10[8],xmm5[9],xmm10[9],xmm5[10],xmm10[10],xmm5[11],xmm10[11],xmm5[12],xmm10[12],xmm5[13],xmm10[13],xmm5[14],xmm10[14],xmm5[15],xmm10[15] ; AVX2-FP-NEXT: vmovdqa %xmm10, %xmm13 ; AVX2-FP-NEXT: vpshufb {{.*#+}} xmm14 = xmm14[8,9,6,7,12,13,10,11,14,15,14,15,14,15,14,15] ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm14 = ymm14[0,0,0,1] ; AVX2-FP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload ; AVX2-FP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload ; AVX2-FP-NEXT: vpunpckhbw {{.*#+}} xmm15 = 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] ; AVX2-FP-NEXT: vpshufb {{.*#+}} xmm15 = xmm15[10,11,8,9,6,7,12,13,14,15,14,15,14,15,14,15] ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm15 = ymm15[0,0,0,1] ; AVX2-FP-NEXT: vpmovsxbw {{.*#+}} ymm0 = [0,65535,0,0,65535,0,0,65535,0,0,65535,0,0,65535,0,0] ; AVX2-FP-NEXT: vpblendvb %ymm0, %ymm14, %ymm15, %ymm14 ; AVX2-FP-NEXT: vpunpckhbw {{.*#+}} ymm15 = ymm8[8],ymm6[8],ymm8[9],ymm6[9],ymm8[10],ymm6[10],ymm8[11],ymm6[11],ymm8[12],ymm6[12],ymm8[13],ymm6[13],ymm8[14],ymm6[14],ymm8[15],ymm6[15],ymm8[24],ymm6[24],ymm8[25],ymm6[25],ymm8[26],ymm6[26],ymm8[27],ymm6[27],ymm8[28],ymm6[28],ymm8[29],ymm6[29],ymm8[30],ymm6[30],ymm8[31],ymm6[31] ; AVX2-FP-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-FP-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,2,2,3] ; AVX2-FP-NEXT: vpunpckhbw {{.*#+}} ymm11 = ymm3[8],ymm4[8],ymm3[9],ymm4[9],ymm3[10],ymm4[10],ymm3[11],ymm4[11],ymm3[12],ymm4[12],ymm3[13],ymm4[13],ymm3[14],ymm4[14],ymm3[15],ymm4[15],ymm3[24],ymm4[24],ymm3[25],ymm4[25],ymm3[26],ymm4[26],ymm3[27],ymm4[27],ymm3[28],ymm4[28],ymm3[29],ymm4[29],ymm3[30],ymm4[30],ymm3[31],ymm4[31] ; AVX2-FP-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-FP-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,2,2,3] ; AVX2-FP-NEXT: vpblendvb %ymm0, %ymm15, %ymm11, %ymm0 ; AVX2-FP-NEXT: vmovdqa %xmm7, %xmm9 ; AVX2-FP-NEXT: vpshufb {{.*#+}} xmm11 = xmm7[10,u,13,u,12,u,11,u,14,u,13,u,14,u,15,u] ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm11 = ymm11[0,0,0,1] ; AVX2-FP-NEXT: vpmovsxbw {{.*#+}} ymm15 = [0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0] ; AVX2-FP-NEXT: vpblendvb %ymm15, %ymm14, %ymm11, %ymm11 ; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload ; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm14 = ymm7[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-FP-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,2,2,3] ; AVX2-FP-NEXT: vpblendvb %ymm15, %ymm0, %ymm14, %ymm0 ; AVX2-FP-NEXT: vpshufb {{.*#+}} xmm14 = xmm12[u,10,u,13,u,12,u,11,u,14,u,13,u,14,u,15] ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm14 = ymm14[0,0,0,1] ; AVX2-FP-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-FP-NEXT: vpblendvb %ymm15, %ymm11, %ymm14, %ymm14 ; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload ; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm11 = 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-FP-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,2,2,3] ; AVX2-FP-NEXT: vpblendvb %ymm15, %ymm0, %ymm11, %ymm15 ; AVX2-FP-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm5[0],xmm13[0],xmm5[1],xmm13[1],xmm5[2],xmm13[2],xmm5[3],xmm13[3],xmm5[4],xmm13[4],xmm5[5],xmm13[5],xmm5[6],xmm13[6],xmm5[7],xmm13[7] ; AVX2-FP-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,1,6,7,4,5,2,3,8,9,10,11,12,13,10,11] ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,0,0,1] ; AVX2-FP-NEXT: vpunpcklbw {{.*#+}} xmm5 = 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] ; AVX2-FP-NEXT: vpshufb {{.*#+}} xmm5 = xmm5[2,3,0,1,6,7,4,5,8,9,8,9,8,9,8,9] ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,0,0,1] ; AVX2-FP-NEXT: vpmovsxbw {{.*#+}} ymm11 = [65535,0,0,65535,0,0,65535,0,0,65535,0,0,65535,0,0,65535] ; AVX2-FP-NEXT: vpblendvb %ymm11, %ymm0, %ymm5, %ymm0 ; AVX2-FP-NEXT: vpunpcklbw {{.*#+}} ymm2 = ymm8[0],ymm6[0],ymm8[1],ymm6[1],ymm8[2],ymm6[2],ymm8[3],ymm6[3],ymm8[4],ymm6[4],ymm8[5],ymm6[5],ymm8[6],ymm6[6],ymm8[7],ymm6[7],ymm8[16],ymm6[16],ymm8[17],ymm6[17],ymm8[18],ymm6[18],ymm8[19],ymm6[19],ymm8[20],ymm6[20],ymm8[21],ymm6[21],ymm8[22],ymm6[22],ymm8[23],ymm6[23] ; AVX2-FP-NEXT: vpunpcklbw {{.*#+}} ymm1 = 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] ; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm2 = ymm2[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-FP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3] ; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[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-FP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3] ; AVX2-FP-NEXT: vpblendvb %ymm11, %ymm2, %ymm1, %ymm1 ; AVX2-FP-NEXT: vpshufb {{.*#+}} xmm2 = xmm9[2,u,1,u,0,u,3,u,4,u,4,u,4,u,4,u] ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,0,0,1] ; AVX2-FP-NEXT: vpmovsxbw {{.*#+}} ymm3 = [65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535] ; AVX2-FP-NEXT: vpblendvb %ymm3, %ymm0, %ymm2, %ymm0 ; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm2 = ymm7[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-FP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3] ; AVX2-FP-NEXT: vpblendvb %ymm3, %ymm1, %ymm2, %ymm1 ; AVX2-FP-NEXT: vpshufb {{.*#+}} xmm2 = xmm12[u,2,u,1,u,0,u,3,u,4,u,4,u,4,u,4] ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,0,0,1] ; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm3 = [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-FP-NEXT: vpblendvb %ymm3, %ymm0, %ymm2, %ymm0 ; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm2 = 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-FP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3] ; AVX2-FP-NEXT: vpblendvb %ymm3, %ymm1, %ymm2, %ymm1 ; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload ; AVX2-FP-NEXT: vmovaps %ymm2, 128(%rax) ; AVX2-FP-NEXT: vmovdqa %ymm1, 96(%rax) ; AVX2-FP-NEXT: vmovdqa %ymm15, 160(%rax) ; AVX2-FP-NEXT: vmovdqa %ymm14, 64(%rax) ; AVX2-FP-NEXT: vmovups (%rsp), %ymm1 # 32-byte Reload ; AVX2-FP-NEXT: vmovaps %ymm1, 32(%rax) ; AVX2-FP-NEXT: vmovdqa %ymm0, (%rax) ; AVX2-FP-NEXT: addq $40, %rsp ; AVX2-FP-NEXT: vzeroupper ; AVX2-FP-NEXT: retq ; ; AVX2-FCP-LABEL: store_i8_stride6_vf32: ; AVX2-FCP: # %bb.0: ; AVX2-FCP-NEXT: subq $40, %rsp ; AVX2-FCP-NEXT: vmovdqa (%rdi), %ymm2 ; AVX2-FCP-NEXT: vmovdqa (%rsi), %ymm6 ; AVX2-FCP-NEXT: vmovdqa (%rdx), %ymm3 ; AVX2-FCP-NEXT: vmovdqa (%rcx), %ymm4 ; AVX2-FCP-NEXT: vmovdqa (%r8), %ymm0 ; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FCP-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] ; AVX2-FCP-NEXT: vmovdqa (%rcx), %xmm1 ; AVX2-FCP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX2-FCP-NEXT: vpshufb %xmm9, %xmm1, %xmm7 ; AVX2-FCP-NEXT: vmovdqa (%rdx), %xmm1 ; AVX2-FCP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX2-FCP-NEXT: vpshufb %xmm9, %xmm1, %xmm8 ; AVX2-FCP-NEXT: vpunpcklbw {{.*#+}} xmm7 = 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] ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,0,0,1] ; AVX2-FCP-NEXT: vmovdqa (%rsi), %xmm10 ; AVX2-FCP-NEXT: vpbroadcastq {{.*#+}} xmm11 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0] ; AVX2-FCP-NEXT: vpshufb %xmm11, %xmm10, %xmm12 ; AVX2-FCP-NEXT: vmovdqa (%rdi), %xmm5 ; AVX2-FCP-NEXT: vpshufb %xmm11, %xmm5, %xmm11 ; AVX2-FCP-NEXT: vpunpckhbw {{.*#+}} xmm11 = xmm11[8],xmm12[8],xmm11[9],xmm12[9],xmm11[10],xmm12[10],xmm11[11],xmm12[11],xmm11[12],xmm12[12],xmm11[13],xmm12[13],xmm11[14],xmm12[14],xmm11[15],xmm12[15] ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm11 = ymm11[0,0,0,1] ; AVX2-FCP-NEXT: vpmovsxbw {{.*#+}} ymm12 = [65535,0,0,65535,0,0,65535,0,0,65535,0,0,65535,0,0,65535] ; AVX2-FCP-NEXT: vpblendvb %ymm12, %ymm7, %ymm11, %ymm11 ; AVX2-FCP-NEXT: vmovdqa (%r8), %xmm7 ; AVX2-FCP-NEXT: vpshufb {{.*#+}} xmm13 = xmm7[6,u,5,u,8,u,7,u,9,u,9,u,9,u,9,u] ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm13 = ymm13[0,0,0,1] ; AVX2-FCP-NEXT: vpmovsxbw {{.*#+}} ymm14 = [65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535] ; AVX2-FCP-NEXT: vpblendvb %ymm14, %ymm11, %ymm13, %ymm11 ; AVX2-FCP-NEXT: vpshufb %ymm9, %ymm4, %ymm13 ; AVX2-FCP-NEXT: vpshufb %ymm9, %ymm3, %ymm9 ; AVX2-FCP-NEXT: vpunpcklbw {{.*#+}} ymm13 = ymm9[0],ymm13[0],ymm9[1],ymm13[1],ymm9[2],ymm13[2],ymm9[3],ymm13[3],ymm9[4],ymm13[4],ymm9[5],ymm13[5],ymm9[6],ymm13[6],ymm9[7],ymm13[7],ymm9[16],ymm13[16],ymm9[17],ymm13[17],ymm9[18],ymm13[18],ymm9[19],ymm13[19],ymm9[20],ymm13[20],ymm9[21],ymm13[21],ymm9[22],ymm13[22],ymm9[23],ymm13[23] ; AVX2-FCP-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] ; AVX2-FCP-NEXT: vpshufb %ymm9, %ymm6, %ymm15 ; AVX2-FCP-NEXT: vpshufb %ymm9, %ymm2, %ymm9 ; AVX2-FCP-NEXT: vmovdqa %ymm2, %ymm8 ; AVX2-FCP-NEXT: vpunpcklbw {{.*#+}} ymm15 = ymm9[0],ymm15[0],ymm9[1],ymm15[1],ymm9[2],ymm15[2],ymm9[3],ymm15[3],ymm9[4],ymm15[4],ymm9[5],ymm15[5],ymm9[6],ymm15[6],ymm9[7],ymm15[7],ymm9[16],ymm15[16],ymm9[17],ymm15[17],ymm9[18],ymm15[18],ymm9[19],ymm15[19],ymm9[20],ymm15[20],ymm9[21],ymm15[21],ymm9[22],ymm15[22],ymm9[23],ymm15[23] ; AVX2-FCP-NEXT: vmovdqa (%r9), %ymm1 ; AVX2-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm13 = ymm13[2,2,2,3] ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,2,2,3] ; AVX2-FCP-NEXT: vpblendvb %ymm12, %ymm13, %ymm15, %ymm13 ; AVX2-FCP-NEXT: vmovdqa (%r9), %xmm12 ; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm15 = ymm0[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-FCP-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,2,2,3] ; AVX2-FCP-NEXT: vpblendvb %ymm14, %ymm13, %ymm15, %ymm13 ; AVX2-FCP-NEXT: vpshufb {{.*#+}} xmm14 = xmm12[u,6,u,5,u,8,u,7,u,9,u,9,u,9,u,9] ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm14 = ymm14[0,0,0,1] ; AVX2-FCP-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-FCP-NEXT: vpblendvb %ymm15, %ymm11, %ymm14, %ymm0 ; AVX2-FCP-NEXT: vmovdqu %ymm0, (%rsp) # 32-byte Spill ; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm14 = ymm1[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-FCP-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,2,2,3] ; AVX2-FCP-NEXT: vpblendvb %ymm15, %ymm13, %ymm14, %ymm0 ; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FCP-NEXT: vpunpckhbw {{.*#+}} xmm14 = xmm5[8],xmm10[8],xmm5[9],xmm10[9],xmm5[10],xmm10[10],xmm5[11],xmm10[11],xmm5[12],xmm10[12],xmm5[13],xmm10[13],xmm5[14],xmm10[14],xmm5[15],xmm10[15] ; AVX2-FCP-NEXT: vmovdqa %xmm10, %xmm13 ; AVX2-FCP-NEXT: vpshufb {{.*#+}} xmm14 = xmm14[8,9,6,7,12,13,10,11,14,15,14,15,14,15,14,15] ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm14 = ymm14[0,0,0,1] ; AVX2-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload ; AVX2-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload ; AVX2-FCP-NEXT: vpunpckhbw {{.*#+}} xmm15 = 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] ; AVX2-FCP-NEXT: vpshufb {{.*#+}} xmm15 = xmm15[10,11,8,9,6,7,12,13,14,15,14,15,14,15,14,15] ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm15 = ymm15[0,0,0,1] ; AVX2-FCP-NEXT: vpmovsxbw {{.*#+}} ymm0 = [0,65535,0,0,65535,0,0,65535,0,0,65535,0,0,65535,0,0] ; AVX2-FCP-NEXT: vpblendvb %ymm0, %ymm14, %ymm15, %ymm14 ; AVX2-FCP-NEXT: vpunpckhbw {{.*#+}} ymm15 = ymm8[8],ymm6[8],ymm8[9],ymm6[9],ymm8[10],ymm6[10],ymm8[11],ymm6[11],ymm8[12],ymm6[12],ymm8[13],ymm6[13],ymm8[14],ymm6[14],ymm8[15],ymm6[15],ymm8[24],ymm6[24],ymm8[25],ymm6[25],ymm8[26],ymm6[26],ymm8[27],ymm6[27],ymm8[28],ymm6[28],ymm8[29],ymm6[29],ymm8[30],ymm6[30],ymm8[31],ymm6[31] ; AVX2-FCP-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-FCP-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,2,2,3] ; AVX2-FCP-NEXT: vpunpckhbw {{.*#+}} ymm11 = ymm3[8],ymm4[8],ymm3[9],ymm4[9],ymm3[10],ymm4[10],ymm3[11],ymm4[11],ymm3[12],ymm4[12],ymm3[13],ymm4[13],ymm3[14],ymm4[14],ymm3[15],ymm4[15],ymm3[24],ymm4[24],ymm3[25],ymm4[25],ymm3[26],ymm4[26],ymm3[27],ymm4[27],ymm3[28],ymm4[28],ymm3[29],ymm4[29],ymm3[30],ymm4[30],ymm3[31],ymm4[31] ; AVX2-FCP-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-FCP-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,2,2,3] ; AVX2-FCP-NEXT: vpblendvb %ymm0, %ymm15, %ymm11, %ymm0 ; AVX2-FCP-NEXT: vmovdqa %xmm7, %xmm9 ; AVX2-FCP-NEXT: vpshufb {{.*#+}} xmm11 = xmm7[10,u,13,u,12,u,11,u,14,u,13,u,14,u,15,u] ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm11 = ymm11[0,0,0,1] ; AVX2-FCP-NEXT: vpmovsxbw {{.*#+}} ymm15 = [0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0] ; AVX2-FCP-NEXT: vpblendvb %ymm15, %ymm14, %ymm11, %ymm11 ; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload ; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm14 = ymm7[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-FCP-NEXT: vpermq {{.*#+}} ymm14 = ymm14[2,2,2,3] ; AVX2-FCP-NEXT: vpblendvb %ymm15, %ymm0, %ymm14, %ymm0 ; AVX2-FCP-NEXT: vpshufb {{.*#+}} xmm14 = xmm12[u,10,u,13,u,12,u,11,u,14,u,13,u,14,u,15] ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm14 = ymm14[0,0,0,1] ; AVX2-FCP-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-FCP-NEXT: vpblendvb %ymm15, %ymm11, %ymm14, %ymm14 ; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload ; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm11 = 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-FCP-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,2,2,3] ; AVX2-FCP-NEXT: vpblendvb %ymm15, %ymm0, %ymm11, %ymm15 ; AVX2-FCP-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm5[0],xmm13[0],xmm5[1],xmm13[1],xmm5[2],xmm13[2],xmm5[3],xmm13[3],xmm5[4],xmm13[4],xmm5[5],xmm13[5],xmm5[6],xmm13[6],xmm5[7],xmm13[7] ; AVX2-FCP-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,1,6,7,4,5,2,3,8,9,10,11,12,13,10,11] ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,0,0,1] ; AVX2-FCP-NEXT: vpunpcklbw {{.*#+}} xmm5 = 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] ; AVX2-FCP-NEXT: vpshufb {{.*#+}} xmm5 = xmm5[2,3,0,1,6,7,4,5,8,9,8,9,8,9,8,9] ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,0,0,1] ; AVX2-FCP-NEXT: vpmovsxbw {{.*#+}} ymm11 = [65535,0,0,65535,0,0,65535,0,0,65535,0,0,65535,0,0,65535] ; AVX2-FCP-NEXT: vpblendvb %ymm11, %ymm0, %ymm5, %ymm0 ; AVX2-FCP-NEXT: vpunpcklbw {{.*#+}} ymm2 = ymm8[0],ymm6[0],ymm8[1],ymm6[1],ymm8[2],ymm6[2],ymm8[3],ymm6[3],ymm8[4],ymm6[4],ymm8[5],ymm6[5],ymm8[6],ymm6[6],ymm8[7],ymm6[7],ymm8[16],ymm6[16],ymm8[17],ymm6[17],ymm8[18],ymm6[18],ymm8[19],ymm6[19],ymm8[20],ymm6[20],ymm8[21],ymm6[21],ymm8[22],ymm6[22],ymm8[23],ymm6[23] ; AVX2-FCP-NEXT: vpunpcklbw {{.*#+}} ymm1 = 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] ; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm2[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-FCP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3] ; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm1[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-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3] ; AVX2-FCP-NEXT: vpblendvb %ymm11, %ymm2, %ymm1, %ymm1 ; AVX2-FCP-NEXT: vpshufb {{.*#+}} xmm2 = xmm9[2,u,1,u,0,u,3,u,4,u,4,u,4,u,4,u] ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,0,0,1] ; AVX2-FCP-NEXT: vpmovsxbw {{.*#+}} ymm3 = [65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535] ; AVX2-FCP-NEXT: vpblendvb %ymm3, %ymm0, %ymm2, %ymm0 ; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm7[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-FCP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3] ; AVX2-FCP-NEXT: vpblendvb %ymm3, %ymm1, %ymm2, %ymm1 ; AVX2-FCP-NEXT: vpshufb {{.*#+}} xmm2 = xmm12[u,2,u,1,u,0,u,3,u,4,u,4,u,4,u,4] ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,0,0,1] ; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm3 = [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-FCP-NEXT: vpblendvb %ymm3, %ymm0, %ymm2, %ymm0 ; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm2 = 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-FCP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3] ; AVX2-FCP-NEXT: vpblendvb %ymm3, %ymm1, %ymm2, %ymm1 ; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload ; AVX2-FCP-NEXT: vmovaps %ymm2, 128(%rax) ; AVX2-FCP-NEXT: vmovdqa %ymm1, 96(%rax) ; AVX2-FCP-NEXT: vmovdqa %ymm15, 160(%rax) ; AVX2-FCP-NEXT: vmovdqa %ymm14, 64(%rax) ; AVX2-FCP-NEXT: vmovups (%rsp), %ymm1 # 32-byte Reload ; AVX2-FCP-NEXT: vmovaps %ymm1, 32(%rax) ; AVX2-FCP-NEXT: vmovdqa %ymm0, (%rax) ; AVX2-FCP-NEXT: addq $40, %rsp ; AVX2-FCP-NEXT: vzeroupper ; AVX2-FCP-NEXT: retq ; ; AVX512-LABEL: store_i8_stride6_vf32: ; AVX512: # %bb.0: ; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512-NEXT: vmovdqa (%rdi), %ymm2 ; AVX512-NEXT: vmovdqa (%rsi), %ymm3 ; AVX512-NEXT: vmovdqa (%rdx), %ymm4 ; AVX512-NEXT: vmovdqa (%rcx), %ymm5 ; AVX512-NEXT: vmovdqa (%r8), %ymm0 ; AVX512-NEXT: vmovdqa (%r9), %ymm1 ; AVX512-NEXT: vmovdqa (%rcx), %xmm7 ; AVX512-NEXT: vmovdqa (%rdx), %xmm8 ; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm6 = 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] ; AVX512-NEXT: vpshufb {{.*#+}} xmm6 = xmm6[10,11,8,9,6,7,12,13,14,15,14,15,14,15,14,15] ; AVX512-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,0,0,1] ; AVX512-NEXT: vpunpcklbw {{.*#+}} ymm9 = 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] ; AVX512-NEXT: vprold $16, %ymm9, %ymm9 ; AVX512-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,2,3] ; AVX512-NEXT: vinserti64x4 $1, %ymm9, %zmm6, %zmm6 ; AVX512-NEXT: vmovdqa (%rsi), %xmm9 ; AVX512-NEXT: vmovdqa (%rdi), %xmm10 ; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm11 = 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] ; AVX512-NEXT: vpshufb {{.*#+}} xmm11 = xmm11[8,9,6,7,12,13,10,11,14,15,14,15,14,15,14,15] ; AVX512-NEXT: vpermq {{.*#+}} ymm11 = ymm11[0,0,0,1] ; AVX512-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] ; AVX512-NEXT: vpshuflw {{.*#+}} ymm12 = ymm12[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15] ; AVX512-NEXT: vpshufhw {{.*#+}} ymm12 = ymm12[0,1,2,3,4,5,6,5,8,9,10,11,12,13,14,13] ; AVX512-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,2,2,3] ; AVX512-NEXT: vinserti64x4 $1, %ymm12, %zmm11, %zmm13 ; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm14 = [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] ; AVX512-NEXT: vpternlogq {{.*#+}} zmm13 = zmm6 ^ (zmm14 & (zmm13 ^ zmm6)) ; AVX512-NEXT: vmovdqa (%r9), %xmm11 ; AVX512-NEXT: vmovdqa (%r8), %xmm12 ; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm6 = xmm12[8],xmm11[8],xmm12[9],xmm11[9],xmm12[10],xmm11[10],xmm12[11],xmm11[11],xmm12[12],xmm11[12],xmm12[13],xmm11[13],xmm12[14],xmm11[14],xmm12[15],xmm11[15] ; AVX512-NEXT: vpshufb {{.*#+}} xmm6 = xmm6[4,5,10,11,8,9,6,7,12,13,10,11,12,13,14,15] ; AVX512-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,0,0,1] ; AVX512-NEXT: vpunpcklbw {{.*#+}} ymm15 = 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] ; AVX512-NEXT: vpshuflw {{.*#+}} ymm15 = ymm15[2,1,0,3,4,5,6,7,10,9,8,11,12,13,14,15] ; AVX512-NEXT: vpshufhw {{.*#+}} ymm15 = ymm15[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12] ; AVX512-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,2,2,3] ; AVX512-NEXT: vinserti64x4 $1, %ymm15, %zmm6, %zmm6 ; AVX512-NEXT: vpternlogq {{.*#+}} zmm6 = zmm6 ^ (mem & (zmm6 ^ zmm13)) ; AVX512-NEXT: vpbroadcastq {{.*#+}} xmm13 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0] ; AVX512-NEXT: vpshufb %xmm13, %xmm9, %xmm15 ; AVX512-NEXT: vpshufb %xmm13, %xmm10, %xmm13 ; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm13 = xmm13[8],xmm15[8],xmm13[9],xmm15[9],xmm13[10],xmm15[10],xmm13[11],xmm15[11],xmm13[12],xmm15[12],xmm13[13],xmm15[13],xmm13[14],xmm15[14],xmm13[15],xmm15[15] ; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm15 = 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] ; AVX512-NEXT: vprold $16, %xmm15, %xmm15 ; AVX512-NEXT: vinserti64x4 $1, %ymm13, %zmm15, %zmm13 ; AVX512-NEXT: vpermq {{.*#+}} zmm15 = zmm13[0,0,0,1,4,4,4,5] ; AVX512-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] ; AVX512-NEXT: vpshufb %xmm13, %xmm7, %xmm7 ; AVX512-NEXT: vpshufb %xmm13, %xmm8, %xmm8 ; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm7 = 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] ; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm8 = 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] ; AVX512-NEXT: vpshuflw {{.*#+}} xmm8 = xmm8[0,3,2,1,4,5,6,7] ; AVX512-NEXT: vpshufhw {{.*#+}} xmm8 = xmm8[0,1,2,3,4,5,6,5] ; AVX512-NEXT: vinserti64x4 $1, %ymm7, %zmm8, %zmm7 ; AVX512-NEXT: vpermq {{.*#+}} zmm8 = zmm7[0,0,0,1,4,4,4,5] ; AVX512-NEXT: vpternlogq {{.*#+}} zmm8 = zmm15 ^ (mem & (zmm8 ^ zmm15)) ; AVX512-NEXT: vpbroadcastq {{.*#+}} ymm7 = [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] ; AVX512-NEXT: vpshufb %xmm7, %xmm11, %xmm9 ; AVX512-NEXT: vpshufb %xmm7, %xmm12, %xmm10 ; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm9 = 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] ; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm10 = xmm12[0],xmm11[0],xmm12[1],xmm11[1],xmm12[2],xmm11[2],xmm12[3],xmm11[3],xmm12[4],xmm11[4],xmm12[5],xmm11[5],xmm12[6],xmm11[6],xmm12[7],xmm11[7] ; AVX512-NEXT: vpshuflw {{.*#+}} xmm10 = xmm10[2,1,0,3,4,5,6,7] ; AVX512-NEXT: vpshufhw {{.*#+}} xmm10 = xmm10[0,1,2,3,4,4,4,4] ; AVX512-NEXT: vinserti64x4 $1, %ymm9, %zmm10, %zmm9 ; AVX512-NEXT: vpermq {{.*#+}} zmm9 = zmm9[0,0,0,1,4,4,4,5] ; AVX512-NEXT: vpternlogq {{.*#+}} zmm9 = zmm9 ^ (zmm14 & (zmm9 ^ zmm8)) ; AVX512-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] ; AVX512-NEXT: vpshufb %ymm8, %ymm3, %ymm10 ; AVX512-NEXT: vpshufb %ymm8, %ymm2, %ymm8 ; AVX512-NEXT: vpunpcklbw {{.*#+}} ymm8 = ymm8[0],ymm10[0],ymm8[1],ymm10[1],ymm8[2],ymm10[2],ymm8[3],ymm10[3],ymm8[4],ymm10[4],ymm8[5],ymm10[5],ymm8[6],ymm10[6],ymm8[7],ymm10[7],ymm8[16],ymm10[16],ymm8[17],ymm10[17],ymm8[18],ymm10[18],ymm8[19],ymm10[19],ymm8[20],ymm10[20],ymm8[21],ymm10[21],ymm8[22],ymm10[22],ymm8[23],ymm10[23] ; AVX512-NEXT: vpunpckhbw {{.*#+}} ymm10 = 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] ; AVX512-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] ; AVX512-NEXT: vinserti64x4 $1, %ymm10, %zmm8, %zmm8 ; AVX512-NEXT: vpermq {{.*#+}} zmm8 = zmm8[2,2,2,3,6,6,6,7] ; AVX512-NEXT: vpshufb %ymm13, %ymm5, %ymm5 ; AVX512-NEXT: vpshufb %ymm13, %ymm4, %ymm4 ; AVX512-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] ; AVX512-NEXT: vpunpckhbw {{.*#+}} ymm2 = 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] ; AVX512-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] ; AVX512-NEXT: vinserti64x4 $1, %ymm2, %zmm4, %zmm2 ; AVX512-NEXT: vpermq {{.*#+}} zmm2 = zmm2[2,2,2,3,6,6,6,7] ; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm3 = [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] ; AVX512-NEXT: # zmm3 = mem[0,1,2,3,0,1,2,3] ; AVX512-NEXT: vpternlogq {{.*#+}} zmm3 = zmm8 ^ (zmm3 & (zmm2 ^ zmm8)) ; AVX512-NEXT: vpshufb %ymm7, %ymm1, %ymm2 ; AVX512-NEXT: vpshufb %ymm7, %ymm0, %ymm4 ; AVX512-NEXT: vpunpcklbw {{.*#+}} ymm2 = ymm4[0],ymm2[0],ymm4[1],ymm2[1],ymm4[2],ymm2[2],ymm4[3],ymm2[3],ymm4[4],ymm2[4],ymm4[5],ymm2[5],ymm4[6],ymm2[6],ymm4[7],ymm2[7],ymm4[16],ymm2[16],ymm4[17],ymm2[17],ymm4[18],ymm2[18],ymm4[19],ymm2[19],ymm4[20],ymm2[20],ymm4[21],ymm2[21],ymm4[22],ymm2[22],ymm4[23],ymm2[23] ; AVX512-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] ; AVX512-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[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] ; AVX512-NEXT: vinserti64x4 $1, %ymm0, %zmm2, %zmm0 ; AVX512-NEXT: vpermq {{.*#+}} zmm0 = zmm0[2,2,2,3,6,6,6,7] ; AVX512-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (mem & (zmm0 ^ zmm3)) ; AVX512-NEXT: vmovdqa64 %zmm0, 128(%rax) ; AVX512-NEXT: vmovdqa64 %zmm9, (%rax) ; AVX512-NEXT: vmovdqa64 %zmm6, 64(%rax) ; AVX512-NEXT: vzeroupper ; AVX512-NEXT: retq ; ; AVX512-FCP-LABEL: store_i8_stride6_vf32: ; AVX512-FCP: # %bb.0: ; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512-FCP-NEXT: vmovdqa (%rdi), %ymm2 ; AVX512-FCP-NEXT: vmovdqa (%rsi), %ymm3 ; AVX512-FCP-NEXT: vmovdqa (%rdx), %ymm4 ; AVX512-FCP-NEXT: vmovdqa (%rcx), %ymm5 ; AVX512-FCP-NEXT: vmovdqa (%r8), %ymm0 ; AVX512-FCP-NEXT: vmovdqa (%r9), %ymm1 ; AVX512-FCP-NEXT: vpbroadcastq {{.*#+}} ymm6 = [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] ; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm3, %ymm7 ; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm2, %ymm6 ; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} ymm6 = ymm6[0],ymm7[0],ymm6[1],ymm7[1],ymm6[2],ymm7[2],ymm6[3],ymm7[3],ymm6[4],ymm7[4],ymm6[5],ymm7[5],ymm6[6],ymm7[6],ymm6[7],ymm7[7],ymm6[16],ymm7[16],ymm6[17],ymm7[17],ymm6[18],ymm7[18],ymm6[19],ymm7[19],ymm6[20],ymm7[20],ymm6[21],ymm7[21],ymm6[22],ymm7[22],ymm6[23],ymm7[23] ; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} ymm7 = 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] ; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm7 = ymm7[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] ; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm7, %zmm6, %zmm6 ; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm6 = zmm6[2,2,2,3,6,6,6,7] ; AVX512-FCP-NEXT: vpbroadcastq {{.*#+}} ymm7 = [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] ; AVX512-FCP-NEXT: vpshufb %ymm7, %ymm5, %ymm8 ; AVX512-FCP-NEXT: vpshufb %ymm7, %ymm4, %ymm9 ; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} ymm8 = ymm9[0],ymm8[0],ymm9[1],ymm8[1],ymm9[2],ymm8[2],ymm9[3],ymm8[3],ymm9[4],ymm8[4],ymm9[5],ymm8[5],ymm9[6],ymm8[6],ymm9[7],ymm8[7],ymm9[16],ymm8[16],ymm9[17],ymm8[17],ymm9[18],ymm8[18],ymm9[19],ymm8[19],ymm9[20],ymm8[20],ymm9[21],ymm8[21],ymm9[22],ymm8[22],ymm9[23],ymm8[23] ; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} ymm9 = 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] ; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm9 = ymm9[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] ; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm9, %zmm8, %zmm8 ; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm8 = zmm8[2,2,2,3,6,6,6,7] ; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm9 = [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] ; AVX512-FCP-NEXT: # zmm9 = mem[0,1,2,3,0,1,2,3] ; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm9 = zmm6 ^ (zmm9 & (zmm8 ^ zmm6)) ; AVX512-FCP-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] ; AVX512-FCP-NEXT: vpshufb %ymm10, %ymm1, %ymm6 ; AVX512-FCP-NEXT: vpshufb %ymm10, %ymm0, %ymm8 ; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} ymm6 = ymm8[0],ymm6[0],ymm8[1],ymm6[1],ymm8[2],ymm6[2],ymm8[3],ymm6[3],ymm8[4],ymm6[4],ymm8[5],ymm6[5],ymm8[6],ymm6[6],ymm8[7],ymm6[7],ymm8[16],ymm6[16],ymm8[17],ymm6[17],ymm8[18],ymm6[18],ymm8[19],ymm6[19],ymm8[20],ymm6[20],ymm8[21],ymm6[21],ymm8[22],ymm6[22],ymm8[23],ymm6[23] ; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} ymm8 = 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] ; AVX512-FCP-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] ; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm8, %zmm6, %zmm6 ; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm6 = zmm6[2,2,2,3,6,6,6,7] ; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm6 = zmm6 ^ (mem & (zmm6 ^ zmm9)) ; AVX512-FCP-NEXT: vmovdqa (%rcx), %xmm9 ; AVX512-FCP-NEXT: vpshufb %xmm7, %xmm9, %xmm8 ; AVX512-FCP-NEXT: vmovdqa (%rdx), %xmm11 ; AVX512-FCP-NEXT: vpshufb %xmm7, %xmm11, %xmm7 ; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm12 = 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] ; AVX512-FCP-NEXT: vmovdqa (%rsi), %xmm7 ; AVX512-FCP-NEXT: vmovdqa (%rdi), %xmm8 ; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm13 = 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] ; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm13 = xmm13[0,1,6,7,4,5,2,3,8,9,10,11,12,13,10,11] ; AVX512-FCP-NEXT: vinserti32x4 $2, %xmm12, %zmm13, %zmm12 ; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm12 = zmm12[0,0,0,1,4,4,4,5] ; AVX512-FCP-NEXT: vpbroadcastq {{.*#+}} xmm13 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0] ; AVX512-FCP-NEXT: vpshufb %xmm13, %xmm7, %xmm14 ; AVX512-FCP-NEXT: vpshufb %xmm13, %xmm8, %xmm13 ; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm13 = xmm13[8],xmm14[8],xmm13[9],xmm14[9],xmm13[10],xmm14[10],xmm13[11],xmm14[11],xmm13[12],xmm14[12],xmm13[13],xmm14[13],xmm13[14],xmm14[14],xmm13[15],xmm14[15] ; AVX512-FCP-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] ; AVX512-FCP-NEXT: vprold $16, %xmm14, %xmm14 ; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm13, %zmm14, %zmm13 ; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm14 = zmm13[0,0,0,1,4,4,4,5] ; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm14 = zmm14 ^ (mem & (zmm14 ^ zmm12)) ; AVX512-FCP-NEXT: vmovdqa (%r9), %xmm12 ; AVX512-FCP-NEXT: vpshufb %xmm10, %xmm12, %xmm15 ; AVX512-FCP-NEXT: vmovdqa (%r8), %xmm13 ; AVX512-FCP-NEXT: vpshufb %xmm10, %xmm13, %xmm10 ; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm10 = xmm10[0],xmm15[0],xmm10[1],xmm15[1],xmm10[2],xmm15[2],xmm10[3],xmm15[3],xmm10[4],xmm15[4],xmm10[5],xmm15[5],xmm10[6],xmm15[6],xmm10[7],xmm15[7] ; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm15 = xmm13[0],xmm12[0],xmm13[1],xmm12[1],xmm13[2],xmm12[2],xmm13[3],xmm12[3],xmm13[4],xmm12[4],xmm13[5],xmm12[5],xmm13[6],xmm12[6],xmm13[7],xmm12[7] ; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm15 = xmm15[4,5,2,3,0,1,6,7,8,9,8,9,8,9,8,9] ; AVX512-FCP-NEXT: vinserti32x4 $2, %xmm10, %zmm15, %zmm10 ; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm10 = zmm10[0,0,0,1,4,4,4,5] ; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm15 = [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] ; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm10 = zmm10 ^ (zmm15 & (zmm10 ^ zmm14)) ; AVX512-FCP-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] ; AVX512-FCP-NEXT: vprold $16, %ymm4, %ymm4 ; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm5 = 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] ; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm5 = xmm5[10,11,8,9,6,7,12,13,14,15,14,15,14,15,14,15] ; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm9 = [0,0,0,1,10,10,10,11] ; AVX512-FCP-NEXT: vpermt2q %zmm4, %zmm9, %zmm5 ; AVX512-FCP-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] ; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm2[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] ; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm3 = 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] ; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm3 = xmm3[8,9,6,7,12,13,10,11,14,15,14,15,14,15,14,15] ; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm9, %zmm3 ; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm3 = zmm5 ^ (zmm15 & (zmm3 ^ zmm5)) ; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} ymm0 = 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] ; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[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] ; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm13[8],xmm12[8],xmm13[9],xmm12[9],xmm13[10],xmm12[10],xmm13[11],xmm12[11],xmm13[12],xmm12[12],xmm13[13],xmm12[13],xmm13[14],xmm12[14],xmm13[15],xmm12[15] ; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[4,5,10,11,8,9,6,7,12,13,10,11,12,13,14,15] ; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm9, %zmm1 ; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm1 = zmm1 ^ (mem & (zmm1 ^ zmm3)) ; AVX512-FCP-NEXT: vmovdqa64 %zmm1, 64(%rax) ; AVX512-FCP-NEXT: vmovdqa64 %zmm10, (%rax) ; AVX512-FCP-NEXT: vmovdqa64 %zmm6, 128(%rax) ; AVX512-FCP-NEXT: vzeroupper ; AVX512-FCP-NEXT: retq ; ; AVX512DQ-LABEL: store_i8_stride6_vf32: ; AVX512DQ: # %bb.0: ; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512DQ-NEXT: vmovdqa (%rdi), %ymm2 ; AVX512DQ-NEXT: vmovdqa (%rsi), %ymm3 ; AVX512DQ-NEXT: vmovdqa (%rdx), %ymm4 ; AVX512DQ-NEXT: vmovdqa (%rcx), %ymm5 ; AVX512DQ-NEXT: vmovdqa (%r8), %ymm0 ; AVX512DQ-NEXT: vmovdqa (%r9), %ymm1 ; AVX512DQ-NEXT: vmovdqa (%rcx), %xmm7 ; AVX512DQ-NEXT: vmovdqa (%rdx), %xmm8 ; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm6 = 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] ; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm6 = xmm6[10,11,8,9,6,7,12,13,14,15,14,15,14,15,14,15] ; AVX512DQ-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,0,0,1] ; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} ymm9 = 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] ; AVX512DQ-NEXT: vprold $16, %ymm9, %ymm9 ; AVX512DQ-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,2,3] ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm9, %zmm6, %zmm6 ; AVX512DQ-NEXT: vmovdqa (%rsi), %xmm9 ; AVX512DQ-NEXT: vmovdqa (%rdi), %xmm10 ; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm11 = 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] ; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm11 = xmm11[8,9,6,7,12,13,10,11,14,15,14,15,14,15,14,15] ; AVX512DQ-NEXT: vpermq {{.*#+}} ymm11 = ymm11[0,0,0,1] ; AVX512DQ-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] ; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm12 = ymm12[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15] ; AVX512DQ-NEXT: vpshufhw {{.*#+}} ymm12 = ymm12[0,1,2,3,4,5,6,5,8,9,10,11,12,13,14,13] ; AVX512DQ-NEXT: vpermq {{.*#+}} ymm12 = ymm12[2,2,2,3] ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm12, %zmm11, %zmm13 ; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm14 = [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] ; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm13 = zmm6 ^ (zmm14 & (zmm13 ^ zmm6)) ; AVX512DQ-NEXT: vmovdqa (%r9), %xmm11 ; AVX512DQ-NEXT: vmovdqa (%r8), %xmm12 ; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm6 = xmm12[8],xmm11[8],xmm12[9],xmm11[9],xmm12[10],xmm11[10],xmm12[11],xmm11[11],xmm12[12],xmm11[12],xmm12[13],xmm11[13],xmm12[14],xmm11[14],xmm12[15],xmm11[15] ; AVX512DQ-NEXT: vpshufb {{.*#+}} xmm6 = xmm6[4,5,10,11,8,9,6,7,12,13,10,11,12,13,14,15] ; AVX512DQ-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,0,0,1] ; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} ymm15 = 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] ; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm15 = ymm15[2,1,0,3,4,5,6,7,10,9,8,11,12,13,14,15] ; AVX512DQ-NEXT: vpshufhw {{.*#+}} ymm15 = ymm15[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12] ; AVX512DQ-NEXT: vpermq {{.*#+}} ymm15 = ymm15[2,2,2,3] ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm15, %zmm6, %zmm6 ; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm6 = zmm6 ^ (mem & (zmm6 ^ zmm13)) ; AVX512DQ-NEXT: vpbroadcastq {{.*#+}} xmm13 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0] ; AVX512DQ-NEXT: vpshufb %xmm13, %xmm9, %xmm15 ; AVX512DQ-NEXT: vpshufb %xmm13, %xmm10, %xmm13 ; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm13 = xmm13[8],xmm15[8],xmm13[9],xmm15[9],xmm13[10],xmm15[10],xmm13[11],xmm15[11],xmm13[12],xmm15[12],xmm13[13],xmm15[13],xmm13[14],xmm15[14],xmm13[15],xmm15[15] ; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm15 = 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] ; AVX512DQ-NEXT: vprold $16, %xmm15, %xmm15 ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm13, %zmm15, %zmm13 ; AVX512DQ-NEXT: vpermq {{.*#+}} zmm15 = zmm13[0,0,0,1,4,4,4,5] ; AVX512DQ-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] ; AVX512DQ-NEXT: vpshufb %xmm13, %xmm7, %xmm7 ; AVX512DQ-NEXT: vpshufb %xmm13, %xmm8, %xmm8 ; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm7 = 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] ; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm8 = 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] ; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm8 = xmm8[0,3,2,1,4,5,6,7] ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm8 = xmm8[0,1,2,3,4,5,6,5] ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm7, %zmm8, %zmm7 ; AVX512DQ-NEXT: vpermq {{.*#+}} zmm8 = zmm7[0,0,0,1,4,4,4,5] ; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm8 = zmm15 ^ (mem & (zmm8 ^ zmm15)) ; AVX512DQ-NEXT: vpbroadcastq {{.*#+}} ymm7 = [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] ; AVX512DQ-NEXT: vpshufb %xmm7, %xmm11, %xmm9 ; AVX512DQ-NEXT: vpshufb %xmm7, %xmm12, %xmm10 ; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm9 = 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] ; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm10 = xmm12[0],xmm11[0],xmm12[1],xmm11[1],xmm12[2],xmm11[2],xmm12[3],xmm11[3],xmm12[4],xmm11[4],xmm12[5],xmm11[5],xmm12[6],xmm11[6],xmm12[7],xmm11[7] ; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm10 = xmm10[2,1,0,3,4,5,6,7] ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm10 = xmm10[0,1,2,3,4,4,4,4] ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm9, %zmm10, %zmm9 ; AVX512DQ-NEXT: vpermq {{.*#+}} zmm9 = zmm9[0,0,0,1,4,4,4,5] ; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm9 = zmm9 ^ (zmm14 & (zmm9 ^ zmm8)) ; AVX512DQ-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] ; AVX512DQ-NEXT: vpshufb %ymm8, %ymm3, %ymm10 ; AVX512DQ-NEXT: vpshufb %ymm8, %ymm2, %ymm8 ; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} ymm8 = ymm8[0],ymm10[0],ymm8[1],ymm10[1],ymm8[2],ymm10[2],ymm8[3],ymm10[3],ymm8[4],ymm10[4],ymm8[5],ymm10[5],ymm8[6],ymm10[6],ymm8[7],ymm10[7],ymm8[16],ymm10[16],ymm8[17],ymm10[17],ymm8[18],ymm10[18],ymm8[19],ymm10[19],ymm8[20],ymm10[20],ymm8[21],ymm10[21],ymm8[22],ymm10[22],ymm8[23],ymm10[23] ; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} ymm10 = 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] ; AVX512DQ-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] ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm10, %zmm8, %zmm8 ; AVX512DQ-NEXT: vpermq {{.*#+}} zmm8 = zmm8[2,2,2,3,6,6,6,7] ; AVX512DQ-NEXT: vpshufb %ymm13, %ymm5, %ymm5 ; AVX512DQ-NEXT: vpshufb %ymm13, %ymm4, %ymm4 ; AVX512DQ-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] ; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} ymm2 = 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] ; AVX512DQ-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] ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm2, %zmm4, %zmm2 ; AVX512DQ-NEXT: vpermq {{.*#+}} zmm2 = zmm2[2,2,2,3,6,6,6,7] ; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm3 = [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] ; AVX512DQ-NEXT: # zmm3 = mem[0,1,2,3,0,1,2,3] ; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm3 = zmm8 ^ (zmm3 & (zmm2 ^ zmm8)) ; AVX512DQ-NEXT: vpshufb %ymm7, %ymm1, %ymm2 ; AVX512DQ-NEXT: vpshufb %ymm7, %ymm0, %ymm4 ; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} ymm2 = ymm4[0],ymm2[0],ymm4[1],ymm2[1],ymm4[2],ymm2[2],ymm4[3],ymm2[3],ymm4[4],ymm2[4],ymm4[5],ymm2[5],ymm4[6],ymm2[6],ymm4[7],ymm2[7],ymm4[16],ymm2[16],ymm4[17],ymm2[17],ymm4[18],ymm2[18],ymm4[19],ymm2[19],ymm4[20],ymm2[20],ymm4[21],ymm2[21],ymm4[22],ymm2[22],ymm4[23],ymm2[23] ; AVX512DQ-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] ; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[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] ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm0, %zmm2, %zmm0 ; AVX512DQ-NEXT: vpermq {{.*#+}} zmm0 = zmm0[2,2,2,3,6,6,6,7] ; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (mem & (zmm0 ^ zmm3)) ; AVX512DQ-NEXT: vmovdqa64 %zmm0, 128(%rax) ; AVX512DQ-NEXT: vmovdqa64 %zmm9, (%rax) ; AVX512DQ-NEXT: vmovdqa64 %zmm6, 64(%rax) ; AVX512DQ-NEXT: vzeroupper ; AVX512DQ-NEXT: retq ; ; AVX512DQ-FCP-LABEL: store_i8_stride6_vf32: ; AVX512DQ-FCP: # %bb.0: ; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %ymm2 ; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %ymm3 ; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %ymm4 ; AVX512DQ-FCP-NEXT: vmovdqa (%rcx), %ymm5 ; AVX512DQ-FCP-NEXT: vmovdqa (%r8), %ymm0 ; AVX512DQ-FCP-NEXT: vmovdqa (%r9), %ymm1 ; AVX512DQ-FCP-NEXT: vpbroadcastq {{.*#+}} ymm6 = [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] ; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm3, %ymm7 ; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm2, %ymm6 ; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} ymm6 = ymm6[0],ymm7[0],ymm6[1],ymm7[1],ymm6[2],ymm7[2],ymm6[3],ymm7[3],ymm6[4],ymm7[4],ymm6[5],ymm7[5],ymm6[6],ymm7[6],ymm6[7],ymm7[7],ymm6[16],ymm7[16],ymm6[17],ymm7[17],ymm6[18],ymm7[18],ymm6[19],ymm7[19],ymm6[20],ymm7[20],ymm6[21],ymm7[21],ymm6[22],ymm7[22],ymm6[23],ymm7[23] ; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} ymm7 = 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] ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm7 = ymm7[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] ; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm7, %zmm6, %zmm6 ; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm6 = zmm6[2,2,2,3,6,6,6,7] ; AVX512DQ-FCP-NEXT: vpbroadcastq {{.*#+}} ymm7 = [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] ; AVX512DQ-FCP-NEXT: vpshufb %ymm7, %ymm5, %ymm8 ; AVX512DQ-FCP-NEXT: vpshufb %ymm7, %ymm4, %ymm9 ; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} ymm8 = ymm9[0],ymm8[0],ymm9[1],ymm8[1],ymm9[2],ymm8[2],ymm9[3],ymm8[3],ymm9[4],ymm8[4],ymm9[5],ymm8[5],ymm9[6],ymm8[6],ymm9[7],ymm8[7],ymm9[16],ymm8[16],ymm9[17],ymm8[17],ymm9[18],ymm8[18],ymm9[19],ymm8[19],ymm9[20],ymm8[20],ymm9[21],ymm8[21],ymm9[22],ymm8[22],ymm9[23],ymm8[23] ; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} ymm9 = 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] ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm9 = ymm9[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] ; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm9, %zmm8, %zmm8 ; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm8 = zmm8[2,2,2,3,6,6,6,7] ; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm9 = [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] ; AVX512DQ-FCP-NEXT: # zmm9 = mem[0,1,2,3,0,1,2,3] ; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm9 = zmm6 ^ (zmm9 & (zmm8 ^ zmm6)) ; AVX512DQ-FCP-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] ; AVX512DQ-FCP-NEXT: vpshufb %ymm10, %ymm1, %ymm6 ; AVX512DQ-FCP-NEXT: vpshufb %ymm10, %ymm0, %ymm8 ; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} ymm6 = ymm8[0],ymm6[0],ymm8[1],ymm6[1],ymm8[2],ymm6[2],ymm8[3],ymm6[3],ymm8[4],ymm6[4],ymm8[5],ymm6[5],ymm8[6],ymm6[6],ymm8[7],ymm6[7],ymm8[16],ymm6[16],ymm8[17],ymm6[17],ymm8[18],ymm6[18],ymm8[19],ymm6[19],ymm8[20],ymm6[20],ymm8[21],ymm6[21],ymm8[22],ymm6[22],ymm8[23],ymm6[23] ; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} ymm8 = 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] ; AVX512DQ-FCP-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] ; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm8, %zmm6, %zmm6 ; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm6 = zmm6[2,2,2,3,6,6,6,7] ; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm6 = zmm6 ^ (mem & (zmm6 ^ zmm9)) ; AVX512DQ-FCP-NEXT: vmovdqa (%rcx), %xmm9 ; AVX512DQ-FCP-NEXT: vpshufb %xmm7, %xmm9, %xmm8 ; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %xmm11 ; AVX512DQ-FCP-NEXT: vpshufb %xmm7, %xmm11, %xmm7 ; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm12 = 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] ; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %xmm7 ; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %xmm8 ; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm13 = 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] ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm13 = xmm13[0,1,6,7,4,5,2,3,8,9,10,11,12,13,10,11] ; AVX512DQ-FCP-NEXT: vinserti32x4 $2, %xmm12, %zmm13, %zmm12 ; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm12 = zmm12[0,0,0,1,4,4,4,5] ; AVX512DQ-FCP-NEXT: vpbroadcastq {{.*#+}} xmm13 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0] ; AVX512DQ-FCP-NEXT: vpshufb %xmm13, %xmm7, %xmm14 ; AVX512DQ-FCP-NEXT: vpshufb %xmm13, %xmm8, %xmm13 ; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm13 = xmm13[8],xmm14[8],xmm13[9],xmm14[9],xmm13[10],xmm14[10],xmm13[11],xmm14[11],xmm13[12],xmm14[12],xmm13[13],xmm14[13],xmm13[14],xmm14[14],xmm13[15],xmm14[15] ; AVX512DQ-FCP-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] ; AVX512DQ-FCP-NEXT: vprold $16, %xmm14, %xmm14 ; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm13, %zmm14, %zmm13 ; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm14 = zmm13[0,0,0,1,4,4,4,5] ; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm14 = zmm14 ^ (mem & (zmm14 ^ zmm12)) ; AVX512DQ-FCP-NEXT: vmovdqa (%r9), %xmm12 ; AVX512DQ-FCP-NEXT: vpshufb %xmm10, %xmm12, %xmm15 ; AVX512DQ-FCP-NEXT: vmovdqa (%r8), %xmm13 ; AVX512DQ-FCP-NEXT: vpshufb %xmm10, %xmm13, %xmm10 ; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm10 = xmm10[0],xmm15[0],xmm10[1],xmm15[1],xmm10[2],xmm15[2],xmm10[3],xmm15[3],xmm10[4],xmm15[4],xmm10[5],xmm15[5],xmm10[6],xmm15[6],xmm10[7],xmm15[7] ; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm15 = xmm13[0],xmm12[0],xmm13[1],xmm12[1],xmm13[2],xmm12[2],xmm13[3],xmm12[3],xmm13[4],xmm12[4],xmm13[5],xmm12[5],xmm13[6],xmm12[6],xmm13[7],xmm12[7] ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm15 = xmm15[4,5,2,3,0,1,6,7,8,9,8,9,8,9,8,9] ; AVX512DQ-FCP-NEXT: vinserti32x4 $2, %xmm10, %zmm15, %zmm10 ; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm10 = zmm10[0,0,0,1,4,4,4,5] ; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm15 = [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] ; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm10 = zmm10 ^ (zmm15 & (zmm10 ^ zmm14)) ; AVX512DQ-FCP-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] ; AVX512DQ-FCP-NEXT: vprold $16, %ymm4, %ymm4 ; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm5 = 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] ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm5 = xmm5[10,11,8,9,6,7,12,13,14,15,14,15,14,15,14,15] ; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm9 = [0,0,0,1,10,10,10,11] ; AVX512DQ-FCP-NEXT: vpermt2q %zmm4, %zmm9, %zmm5 ; AVX512DQ-FCP-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] ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm2[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] ; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm3 = 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] ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm3 = xmm3[8,9,6,7,12,13,10,11,14,15,14,15,14,15,14,15] ; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm9, %zmm3 ; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm3 = zmm5 ^ (zmm15 & (zmm3 ^ zmm5)) ; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} ymm0 = 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] ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[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] ; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm13[8],xmm12[8],xmm13[9],xmm12[9],xmm13[10],xmm12[10],xmm13[11],xmm12[11],xmm13[12],xmm12[12],xmm13[13],xmm12[13],xmm13[14],xmm12[14],xmm13[15],xmm12[15] ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[4,5,10,11,8,9,6,7,12,13,10,11,12,13,14,15] ; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm9, %zmm1 ; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm1 = zmm1 ^ (mem & (zmm1 ^ zmm3)) ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, 64(%rax) ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm10, (%rax) ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, 128(%rax) ; AVX512DQ-FCP-NEXT: vzeroupper ; AVX512DQ-FCP-NEXT: retq ; ; AVX512BW-LABEL: store_i8_stride6_vf32: ; AVX512BW: # %bb.0: ; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512BW-NEXT: vmovdqa (%rdi), %ymm2 ; AVX512BW-NEXT: vmovdqa (%rsi), %ymm3 ; AVX512BW-NEXT: vmovdqa (%rdx), %ymm4 ; AVX512BW-NEXT: vmovdqa (%rcx), %ymm5 ; AVX512BW-NEXT: vmovdqa (%r8), %ymm0 ; AVX512BW-NEXT: vmovdqa (%r9), %ymm1 ; AVX512BW-NEXT: vpbroadcastq {{.*#+}} ymm6 = [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-NEXT: vpshufb %ymm6, %ymm3, %ymm7 ; AVX512BW-NEXT: vpshufb %ymm6, %ymm2, %ymm6 ; AVX512BW-NEXT: vpunpcklbw {{.*#+}} ymm6 = ymm6[0],ymm7[0],ymm6[1],ymm7[1],ymm6[2],ymm7[2],ymm6[3],ymm7[3],ymm6[4],ymm7[4],ymm6[5],ymm7[5],ymm6[6],ymm7[6],ymm6[7],ymm7[7],ymm6[16],ymm7[16],ymm6[17],ymm7[17],ymm6[18],ymm7[18],ymm6[19],ymm7[19],ymm6[20],ymm7[20],ymm6[21],ymm7[21],ymm6[22],ymm7[22],ymm6[23],ymm7[23] ; AVX512BW-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,2,2,3] ; AVX512BW-NEXT: vpbroadcastq {{.*#+}} ymm7 = [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-NEXT: vpshufb %ymm7, %ymm5, %ymm8 ; AVX512BW-NEXT: vpshufb %ymm7, %ymm4, %ymm9 ; AVX512BW-NEXT: vpunpcklbw {{.*#+}} ymm8 = ymm9[0],ymm8[0],ymm9[1],ymm8[1],ymm9[2],ymm8[2],ymm9[3],ymm8[3],ymm9[4],ymm8[4],ymm9[5],ymm8[5],ymm9[6],ymm8[6],ymm9[7],ymm8[7],ymm9[16],ymm8[16],ymm9[17],ymm8[17],ymm9[18],ymm8[18],ymm9[19],ymm8[19],ymm9[20],ymm8[20],ymm9[21],ymm8[21],ymm9[22],ymm8[22],ymm9[23],ymm8[23] ; AVX512BW-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,2,3] ; AVX512BW-NEXT: movw $18724, %r10w # imm = 0x4924 ; AVX512BW-NEXT: kmovd %r10d, %k1 ; AVX512BW-NEXT: vmovdqu16 %ymm6, %ymm8 {%k1} ; AVX512BW-NEXT: vpbroadcastq {{.*#+}} ymm9 = [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-NEXT: vpshufb %ymm9, %ymm1, %ymm6 ; AVX512BW-NEXT: vpshufb %ymm9, %ymm0, %ymm10 ; AVX512BW-NEXT: vpunpcklbw {{.*#+}} ymm6 = ymm10[0],ymm6[0],ymm10[1],ymm6[1],ymm10[2],ymm6[2],ymm10[3],ymm6[3],ymm10[4],ymm6[4],ymm10[5],ymm6[5],ymm10[6],ymm6[6],ymm10[7],ymm6[7],ymm10[16],ymm6[16],ymm10[17],ymm6[17],ymm10[18],ymm6[18],ymm10[19],ymm6[19],ymm10[20],ymm6[20],ymm10[21],ymm6[21],ymm10[22],ymm6[22],ymm10[23],ymm6[23] ; AVX512BW-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,2,2,3] ; AVX512BW-NEXT: movw $9362, %r10w # imm = 0x2492 ; AVX512BW-NEXT: kmovd %r10d, %k2 ; AVX512BW-NEXT: vmovdqu16 %ymm6, %ymm8 {%k2} ; AVX512BW-NEXT: vpunpckhbw {{.*#+}} ymm6 = 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-NEXT: vpunpckhbw {{.*#+}} ymm10 = 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-NEXT: vpmovsxbw {{.*#+}} ymm11 = [0,11,27,0,12,28,0,13,29,0,14,30,0,15,31,0] ; AVX512BW-NEXT: vpermi2w %ymm6, %ymm10, %ymm11 ; AVX512BW-NEXT: vinserti64x4 $1, %ymm11, %zmm0, %zmm6 ; AVX512BW-NEXT: vpmovsxbw {{.*#+}} ymm10 = [42,17,18,43,20,21,44,23,24,45,26,27,46,29,30,47] ; AVX512BW-NEXT: vpunpckhbw {{.*#+}} ymm11 = 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-NEXT: vpermi2w %zmm11, %zmm6, %zmm10 ; AVX512BW-NEXT: vinserti64x4 $1, %ymm10, %zmm8, %zmm6 ; AVX512BW-NEXT: vmovdqa (%rsi), %xmm8 ; AVX512BW-NEXT: vpbroadcastq {{.*#+}} xmm10 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0] ; AVX512BW-NEXT: vpshufb %xmm10, %xmm8, %xmm11 ; AVX512BW-NEXT: vmovdqa (%rdi), %xmm12 ; AVX512BW-NEXT: vpshufb %xmm10, %xmm12, %xmm10 ; AVX512BW-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] ; AVX512BW-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,0,0,1] ; AVX512BW-NEXT: vmovdqa (%rcx), %xmm11 ; AVX512BW-NEXT: vpshufb %xmm7, %xmm11, %xmm13 ; AVX512BW-NEXT: vmovdqa (%rdx), %xmm14 ; AVX512BW-NEXT: vpshufb %xmm7, %xmm14, %xmm7 ; AVX512BW-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] ; AVX512BW-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,0,0,1] ; AVX512BW-NEXT: vmovdqu16 %ymm10, %ymm7 {%k1} ; AVX512BW-NEXT: vmovdqa (%r9), %xmm10 ; AVX512BW-NEXT: vpshufb %xmm9, %xmm10, %xmm13 ; AVX512BW-NEXT: vmovdqa (%r8), %xmm15 ; AVX512BW-NEXT: vpshufb %xmm9, %xmm15, %xmm9 ; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm9 = xmm9[0],xmm13[0],xmm9[1],xmm13[1],xmm9[2],xmm13[2],xmm9[3],xmm13[3],xmm9[4],xmm13[4],xmm9[5],xmm13[5],xmm9[6],xmm13[6],xmm9[7],xmm13[7] ; AVX512BW-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,0,0,1] ; AVX512BW-NEXT: vmovdqu16 %ymm9, %ymm7 {%k2} ; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm9 = 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-NEXT: vpunpcklbw {{.*#+}} xmm13 = 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-NEXT: vpmovsxbw {{.*#+}} ymm16 = [0,16,0,1,17,0,2,18,0,3,19,0,4,20,0,5] ; AVX512BW-NEXT: vpermi2w %ymm9, %ymm13, %ymm16 ; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm9 = 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-NEXT: vpmovsxbw {{.*#+}} ymm13 = [0,1,16,3,4,17,6,7,18,9,10,19,12,13,20,15] ; AVX512BW-NEXT: vpermi2w %ymm9, %ymm16, %ymm13 ; AVX512BW-NEXT: vinserti64x4 $1, %ymm7, %zmm13, %zmm7 ; AVX512BW-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-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-NEXT: vpmovsxbw {{.*#+}} ymm11 = [0,3,19,0,4,20,0,5,21,0,6,22,0,7,23,0] ; AVX512BW-NEXT: vpermi2w %ymm9, %ymm8, %ymm11 ; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm8 = xmm15[8],xmm10[8],xmm15[9],xmm10[9],xmm15[10],xmm10[10],xmm15[11],xmm10[11],xmm15[12],xmm10[12],xmm15[13],xmm10[13],xmm15[14],xmm10[14],xmm15[15],xmm10[15] ; AVX512BW-NEXT: vpmovsxbw {{.*#+}} ymm9 = [18,1,2,19,4,5,20,7,8,21,10,11,22,13,14,23] ; AVX512BW-NEXT: vpermi2w %ymm8, %ymm11, %ymm9 ; AVX512BW-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-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-NEXT: vpmovsxbw {{.*#+}} ymm3 = [8,24,0,9,25,0,10,26,0,11,27,0,12,28,0,13] ; AVX512BW-NEXT: vpermi2w %ymm4, %ymm2, %ymm3 ; AVX512BW-NEXT: vinserti64x4 $1, %ymm3, %zmm0, %zmm2 ; AVX512BW-NEXT: vpmovsxbw {{.*#+}} ymm3 = [16,17,40,19,20,41,22,23,42,25,26,43,28,29,44,31] ; AVX512BW-NEXT: vpunpcklbw {{.*#+}} ymm0 = 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-NEXT: vpermi2w %zmm0, %zmm2, %zmm3 ; AVX512BW-NEXT: vinserti64x4 $1, %ymm3, %zmm9, %zmm0 ; AVX512BW-NEXT: vmovdqa64 %zmm0, 64(%rax) ; AVX512BW-NEXT: vmovdqa64 %zmm7, (%rax) ; AVX512BW-NEXT: vmovdqa64 %zmm6, 128(%rax) ; AVX512BW-NEXT: vzeroupper ; AVX512BW-NEXT: retq ; ; AVX512BW-FCP-LABEL: store_i8_stride6_vf32: ; AVX512BW-FCP: # %bb.0: ; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512BW-FCP-NEXT: vmovdqa (%rdi), %ymm2 ; AVX512BW-FCP-NEXT: vmovdqa (%rsi), %ymm3 ; AVX512BW-FCP-NEXT: vmovdqa (%rdx), %ymm4 ; AVX512BW-FCP-NEXT: vmovdqa (%rcx), %ymm5 ; AVX512BW-FCP-NEXT: vmovdqa (%r8), %ymm0 ; AVX512BW-FCP-NEXT: vmovdqa (%r9), %ymm1 ; AVX512BW-FCP-NEXT: vpbroadcastq {{.*#+}} ymm6 = [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-FCP-NEXT: vpshufb %ymm6, %ymm3, %ymm7 ; AVX512BW-FCP-NEXT: vpshufb %ymm6, %ymm2, %ymm6 ; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} ymm6 = ymm6[0],ymm7[0],ymm6[1],ymm7[1],ymm6[2],ymm7[2],ymm6[3],ymm7[3],ymm6[4],ymm7[4],ymm6[5],ymm7[5],ymm6[6],ymm7[6],ymm6[7],ymm7[7],ymm6[16],ymm7[16],ymm6[17],ymm7[17],ymm6[18],ymm7[18],ymm6[19],ymm7[19],ymm6[20],ymm7[20],ymm6[21],ymm7[21],ymm6[22],ymm7[22],ymm6[23],ymm7[23] ; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,2,2,3] ; AVX512BW-FCP-NEXT: vpbroadcastq {{.*#+}} ymm7 = [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-FCP-NEXT: vpshufb %ymm7, %ymm5, %ymm8 ; AVX512BW-FCP-NEXT: vpshufb %ymm7, %ymm4, %ymm9 ; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} ymm8 = ymm9[0],ymm8[0],ymm9[1],ymm8[1],ymm9[2],ymm8[2],ymm9[3],ymm8[3],ymm9[4],ymm8[4],ymm9[5],ymm8[5],ymm9[6],ymm8[6],ymm9[7],ymm8[7],ymm9[16],ymm8[16],ymm9[17],ymm8[17],ymm9[18],ymm8[18],ymm9[19],ymm8[19],ymm9[20],ymm8[20],ymm9[21],ymm8[21],ymm9[22],ymm8[22],ymm9[23],ymm8[23] ; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,2,3] ; AVX512BW-FCP-NEXT: movw $18724, %r10w # imm = 0x4924 ; AVX512BW-FCP-NEXT: kmovd %r10d, %k1 ; AVX512BW-FCP-NEXT: vmovdqu16 %ymm6, %ymm8 {%k1} ; AVX512BW-FCP-NEXT: vpbroadcastq {{.*#+}} ymm9 = [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-FCP-NEXT: vpshufb %ymm9, %ymm1, %ymm6 ; AVX512BW-FCP-NEXT: vpshufb %ymm9, %ymm0, %ymm10 ; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} ymm6 = ymm10[0],ymm6[0],ymm10[1],ymm6[1],ymm10[2],ymm6[2],ymm10[3],ymm6[3],ymm10[4],ymm6[4],ymm10[5],ymm6[5],ymm10[6],ymm6[6],ymm10[7],ymm6[7],ymm10[16],ymm6[16],ymm10[17],ymm6[17],ymm10[18],ymm6[18],ymm10[19],ymm6[19],ymm10[20],ymm6[20],ymm10[21],ymm6[21],ymm10[22],ymm6[22],ymm10[23],ymm6[23] ; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,2,2,3] ; AVX512BW-FCP-NEXT: movw $9362, %r10w # imm = 0x2492 ; AVX512BW-FCP-NEXT: kmovd %r10d, %k2 ; AVX512BW-FCP-NEXT: vmovdqu16 %ymm6, %ymm8 {%k2} ; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} ymm6 = 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-FCP-NEXT: vpunpckhbw {{.*#+}} ymm10 = 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-FCP-NEXT: vpmovsxbw {{.*#+}} ymm11 = [0,11,27,0,12,28,0,13,29,0,14,30,0,15,31,0] ; AVX512BW-FCP-NEXT: vpermi2w %ymm6, %ymm10, %ymm11 ; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm11, %zmm0, %zmm6 ; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm10 = [42,17,18,43,20,21,44,23,24,45,26,27,46,29,30,47] ; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} ymm11 = 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-FCP-NEXT: vpermi2w %zmm11, %zmm6, %zmm10 ; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm10, %zmm8, %zmm6 ; AVX512BW-FCP-NEXT: vmovdqa (%rsi), %xmm8 ; AVX512BW-FCP-NEXT: vpbroadcastq {{.*#+}} xmm10 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0] ; AVX512BW-FCP-NEXT: vpshufb %xmm10, %xmm8, %xmm11 ; AVX512BW-FCP-NEXT: vmovdqa (%rdi), %xmm12 ; AVX512BW-FCP-NEXT: vpshufb %xmm10, %xmm12, %xmm10 ; AVX512BW-FCP-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] ; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,0,0,1] ; AVX512BW-FCP-NEXT: vmovdqa (%rcx), %xmm11 ; AVX512BW-FCP-NEXT: vpshufb %xmm7, %xmm11, %xmm13 ; AVX512BW-FCP-NEXT: vmovdqa (%rdx), %xmm14 ; AVX512BW-FCP-NEXT: vpshufb %xmm7, %xmm14, %xmm7 ; AVX512BW-FCP-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] ; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,0,0,1] ; AVX512BW-FCP-NEXT: vmovdqu16 %ymm10, %ymm7 {%k1} ; AVX512BW-FCP-NEXT: vmovdqa (%r9), %xmm10 ; AVX512BW-FCP-NEXT: vpshufb %xmm9, %xmm10, %xmm13 ; AVX512BW-FCP-NEXT: vmovdqa (%r8), %xmm15 ; AVX512BW-FCP-NEXT: vpshufb %xmm9, %xmm15, %xmm9 ; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm9 = xmm9[0],xmm13[0],xmm9[1],xmm13[1],xmm9[2],xmm13[2],xmm9[3],xmm13[3],xmm9[4],xmm13[4],xmm9[5],xmm13[5],xmm9[6],xmm13[6],xmm9[7],xmm13[7] ; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,0,0,1] ; AVX512BW-FCP-NEXT: vmovdqu16 %ymm9, %ymm7 {%k2} ; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm9 = 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-FCP-NEXT: vpunpcklbw {{.*#+}} xmm13 = 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-FCP-NEXT: vpmovsxbw {{.*#+}} ymm16 = [0,16,0,1,17,0,2,18,0,3,19,0,4,20,0,5] ; AVX512BW-FCP-NEXT: vpermi2w %ymm9, %ymm13, %ymm16 ; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm9 = 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-FCP-NEXT: vpmovsxbw {{.*#+}} ymm13 = [0,1,16,3,4,17,6,7,18,9,10,19,12,13,20,15] ; AVX512BW-FCP-NEXT: vpermi2w %ymm9, %ymm16, %ymm13 ; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm7, %zmm13, %zmm7 ; AVX512BW-FCP-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-FCP-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-FCP-NEXT: vpmovsxbw {{.*#+}} ymm11 = [0,3,19,0,4,20,0,5,21,0,6,22,0,7,23,0] ; AVX512BW-FCP-NEXT: vpermi2w %ymm9, %ymm8, %ymm11 ; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm8 = xmm15[8],xmm10[8],xmm15[9],xmm10[9],xmm15[10],xmm10[10],xmm15[11],xmm10[11],xmm15[12],xmm10[12],xmm15[13],xmm10[13],xmm15[14],xmm10[14],xmm15[15],xmm10[15] ; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm9 = [18,1,2,19,4,5,20,7,8,21,10,11,22,13,14,23] ; AVX512BW-FCP-NEXT: vpermi2w %ymm8, %ymm11, %ymm9 ; AVX512BW-FCP-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-FCP-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-FCP-NEXT: vpmovsxbw {{.*#+}} ymm3 = [8,24,0,9,25,0,10,26,0,11,27,0,12,28,0,13] ; AVX512BW-FCP-NEXT: vpermi2w %ymm4, %ymm2, %ymm3 ; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm3, %zmm0, %zmm2 ; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm3 = [16,17,40,19,20,41,22,23,42,25,26,43,28,29,44,31] ; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} ymm0 = 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-FCP-NEXT: vpermi2w %zmm0, %zmm2, %zmm3 ; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm3, %zmm9, %zmm0 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, 64(%rax) ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm7, (%rax) ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm6, 128(%rax) ; AVX512BW-FCP-NEXT: vzeroupper ; AVX512BW-FCP-NEXT: retq ; ; AVX512DQ-BW-LABEL: store_i8_stride6_vf32: ; AVX512DQ-BW: # %bb.0: ; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512DQ-BW-NEXT: vmovdqa (%rdi), %ymm2 ; AVX512DQ-BW-NEXT: vmovdqa (%rsi), %ymm3 ; AVX512DQ-BW-NEXT: vmovdqa (%rdx), %ymm4 ; AVX512DQ-BW-NEXT: vmovdqa (%rcx), %ymm5 ; AVX512DQ-BW-NEXT: vmovdqa (%r8), %ymm0 ; AVX512DQ-BW-NEXT: vmovdqa (%r9), %ymm1 ; AVX512DQ-BW-NEXT: vpbroadcastq {{.*#+}} ymm6 = [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] ; AVX512DQ-BW-NEXT: vpshufb %ymm6, %ymm3, %ymm7 ; AVX512DQ-BW-NEXT: vpshufb %ymm6, %ymm2, %ymm6 ; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} ymm6 = ymm6[0],ymm7[0],ymm6[1],ymm7[1],ymm6[2],ymm7[2],ymm6[3],ymm7[3],ymm6[4],ymm7[4],ymm6[5],ymm7[5],ymm6[6],ymm7[6],ymm6[7],ymm7[7],ymm6[16],ymm7[16],ymm6[17],ymm7[17],ymm6[18],ymm7[18],ymm6[19],ymm7[19],ymm6[20],ymm7[20],ymm6[21],ymm7[21],ymm6[22],ymm7[22],ymm6[23],ymm7[23] ; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,2,2,3] ; AVX512DQ-BW-NEXT: vpbroadcastq {{.*#+}} ymm7 = [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] ; AVX512DQ-BW-NEXT: vpshufb %ymm7, %ymm5, %ymm8 ; AVX512DQ-BW-NEXT: vpshufb %ymm7, %ymm4, %ymm9 ; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} ymm8 = ymm9[0],ymm8[0],ymm9[1],ymm8[1],ymm9[2],ymm8[2],ymm9[3],ymm8[3],ymm9[4],ymm8[4],ymm9[5],ymm8[5],ymm9[6],ymm8[6],ymm9[7],ymm8[7],ymm9[16],ymm8[16],ymm9[17],ymm8[17],ymm9[18],ymm8[18],ymm9[19],ymm8[19],ymm9[20],ymm8[20],ymm9[21],ymm8[21],ymm9[22],ymm8[22],ymm9[23],ymm8[23] ; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,2,3] ; AVX512DQ-BW-NEXT: movw $18724, %r10w # imm = 0x4924 ; AVX512DQ-BW-NEXT: kmovd %r10d, %k1 ; AVX512DQ-BW-NEXT: vmovdqu16 %ymm6, %ymm8 {%k1} ; AVX512DQ-BW-NEXT: vpbroadcastq {{.*#+}} ymm9 = [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] ; AVX512DQ-BW-NEXT: vpshufb %ymm9, %ymm1, %ymm6 ; AVX512DQ-BW-NEXT: vpshufb %ymm9, %ymm0, %ymm10 ; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} ymm6 = ymm10[0],ymm6[0],ymm10[1],ymm6[1],ymm10[2],ymm6[2],ymm10[3],ymm6[3],ymm10[4],ymm6[4],ymm10[5],ymm6[5],ymm10[6],ymm6[6],ymm10[7],ymm6[7],ymm10[16],ymm6[16],ymm10[17],ymm6[17],ymm10[18],ymm6[18],ymm10[19],ymm6[19],ymm10[20],ymm6[20],ymm10[21],ymm6[21],ymm10[22],ymm6[22],ymm10[23],ymm6[23] ; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,2,2,3] ; AVX512DQ-BW-NEXT: movw $9362, %r10w # imm = 0x2492 ; AVX512DQ-BW-NEXT: kmovd %r10d, %k2 ; AVX512DQ-BW-NEXT: vmovdqu16 %ymm6, %ymm8 {%k2} ; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} ymm6 = 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] ; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} ymm10 = 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] ; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} ymm11 = [0,11,27,0,12,28,0,13,29,0,14,30,0,15,31,0] ; AVX512DQ-BW-NEXT: vpermi2w %ymm6, %ymm10, %ymm11 ; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm11, %zmm0, %zmm6 ; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} ymm10 = [42,17,18,43,20,21,44,23,24,45,26,27,46,29,30,47] ; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} ymm11 = 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] ; AVX512DQ-BW-NEXT: vpermi2w %zmm11, %zmm6, %zmm10 ; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm10, %zmm8, %zmm6 ; AVX512DQ-BW-NEXT: vmovdqa (%rsi), %xmm8 ; AVX512DQ-BW-NEXT: vpbroadcastq {{.*#+}} xmm10 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0] ; AVX512DQ-BW-NEXT: vpshufb %xmm10, %xmm8, %xmm11 ; AVX512DQ-BW-NEXT: vmovdqa (%rdi), %xmm12 ; AVX512DQ-BW-NEXT: vpshufb %xmm10, %xmm12, %xmm10 ; AVX512DQ-BW-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] ; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,0,0,1] ; AVX512DQ-BW-NEXT: vmovdqa (%rcx), %xmm11 ; AVX512DQ-BW-NEXT: vpshufb %xmm7, %xmm11, %xmm13 ; AVX512DQ-BW-NEXT: vmovdqa (%rdx), %xmm14 ; AVX512DQ-BW-NEXT: vpshufb %xmm7, %xmm14, %xmm7 ; AVX512DQ-BW-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] ; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,0,0,1] ; AVX512DQ-BW-NEXT: vmovdqu16 %ymm10, %ymm7 {%k1} ; AVX512DQ-BW-NEXT: vmovdqa (%r9), %xmm10 ; AVX512DQ-BW-NEXT: vpshufb %xmm9, %xmm10, %xmm13 ; AVX512DQ-BW-NEXT: vmovdqa (%r8), %xmm15 ; AVX512DQ-BW-NEXT: vpshufb %xmm9, %xmm15, %xmm9 ; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm9 = xmm9[0],xmm13[0],xmm9[1],xmm13[1],xmm9[2],xmm13[2],xmm9[3],xmm13[3],xmm9[4],xmm13[4],xmm9[5],xmm13[5],xmm9[6],xmm13[6],xmm9[7],xmm13[7] ; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,0,0,1] ; AVX512DQ-BW-NEXT: vmovdqu16 %ymm9, %ymm7 {%k2} ; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm9 = 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] ; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm13 = 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] ; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} ymm16 = [0,16,0,1,17,0,2,18,0,3,19,0,4,20,0,5] ; AVX512DQ-BW-NEXT: vpermi2w %ymm9, %ymm13, %ymm16 ; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm9 = 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] ; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} ymm13 = [0,1,16,3,4,17,6,7,18,9,10,19,12,13,20,15] ; AVX512DQ-BW-NEXT: vpermi2w %ymm9, %ymm16, %ymm13 ; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm7, %zmm13, %zmm7 ; AVX512DQ-BW-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] ; AVX512DQ-BW-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] ; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} ymm11 = [0,3,19,0,4,20,0,5,21,0,6,22,0,7,23,0] ; AVX512DQ-BW-NEXT: vpermi2w %ymm9, %ymm8, %ymm11 ; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm8 = xmm15[8],xmm10[8],xmm15[9],xmm10[9],xmm15[10],xmm10[10],xmm15[11],xmm10[11],xmm15[12],xmm10[12],xmm15[13],xmm10[13],xmm15[14],xmm10[14],xmm15[15],xmm10[15] ; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} ymm9 = [18,1,2,19,4,5,20,7,8,21,10,11,22,13,14,23] ; AVX512DQ-BW-NEXT: vpermi2w %ymm8, %ymm11, %ymm9 ; AVX512DQ-BW-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] ; AVX512DQ-BW-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] ; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} ymm3 = [8,24,0,9,25,0,10,26,0,11,27,0,12,28,0,13] ; AVX512DQ-BW-NEXT: vpermi2w %ymm4, %ymm2, %ymm3 ; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm3, %zmm0, %zmm2 ; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} ymm3 = [16,17,40,19,20,41,22,23,42,25,26,43,28,29,44,31] ; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} ymm0 = 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] ; AVX512DQ-BW-NEXT: vpermi2w %zmm0, %zmm2, %zmm3 ; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm3, %zmm9, %zmm0 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, 64(%rax) ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm7, (%rax) ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm6, 128(%rax) ; AVX512DQ-BW-NEXT: vzeroupper ; AVX512DQ-BW-NEXT: retq ; ; AVX512DQ-BW-FCP-LABEL: store_i8_stride6_vf32: ; AVX512DQ-BW-FCP: # %bb.0: ; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdi), %ymm2 ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rsi), %ymm3 ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdx), %ymm4 ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rcx), %ymm5 ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%r8), %ymm0 ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%r9), %ymm1 ; AVX512DQ-BW-FCP-NEXT: vpbroadcastq {{.*#+}} ymm6 = [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] ; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm6, %ymm3, %ymm7 ; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm6, %ymm2, %ymm6 ; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} ymm6 = ymm6[0],ymm7[0],ymm6[1],ymm7[1],ymm6[2],ymm7[2],ymm6[3],ymm7[3],ymm6[4],ymm7[4],ymm6[5],ymm7[5],ymm6[6],ymm7[6],ymm6[7],ymm7[7],ymm6[16],ymm7[16],ymm6[17],ymm7[17],ymm6[18],ymm7[18],ymm6[19],ymm7[19],ymm6[20],ymm7[20],ymm6[21],ymm7[21],ymm6[22],ymm7[22],ymm6[23],ymm7[23] ; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,2,2,3] ; AVX512DQ-BW-FCP-NEXT: vpbroadcastq {{.*#+}} ymm7 = [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] ; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm7, %ymm5, %ymm8 ; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm7, %ymm4, %ymm9 ; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} ymm8 = ymm9[0],ymm8[0],ymm9[1],ymm8[1],ymm9[2],ymm8[2],ymm9[3],ymm8[3],ymm9[4],ymm8[4],ymm9[5],ymm8[5],ymm9[6],ymm8[6],ymm9[7],ymm8[7],ymm9[16],ymm8[16],ymm9[17],ymm8[17],ymm9[18],ymm8[18],ymm9[19],ymm8[19],ymm9[20],ymm8[20],ymm9[21],ymm8[21],ymm9[22],ymm8[22],ymm9[23],ymm8[23] ; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,2,3] ; AVX512DQ-BW-FCP-NEXT: movw $18724, %r10w # imm = 0x4924 ; AVX512DQ-BW-FCP-NEXT: kmovd %r10d, %k1 ; AVX512DQ-BW-FCP-NEXT: vmovdqu16 %ymm6, %ymm8 {%k1} ; AVX512DQ-BW-FCP-NEXT: vpbroadcastq {{.*#+}} ymm9 = [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] ; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm9, %ymm1, %ymm6 ; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm9, %ymm0, %ymm10 ; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} ymm6 = ymm10[0],ymm6[0],ymm10[1],ymm6[1],ymm10[2],ymm6[2],ymm10[3],ymm6[3],ymm10[4],ymm6[4],ymm10[5],ymm6[5],ymm10[6],ymm6[6],ymm10[7],ymm6[7],ymm10[16],ymm6[16],ymm10[17],ymm6[17],ymm10[18],ymm6[18],ymm10[19],ymm6[19],ymm10[20],ymm6[20],ymm10[21],ymm6[21],ymm10[22],ymm6[22],ymm10[23],ymm6[23] ; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,2,2,3] ; AVX512DQ-BW-FCP-NEXT: movw $9362, %r10w # imm = 0x2492 ; AVX512DQ-BW-FCP-NEXT: kmovd %r10d, %k2 ; AVX512DQ-BW-FCP-NEXT: vmovdqu16 %ymm6, %ymm8 {%k2} ; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} ymm6 = 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] ; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} ymm10 = 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] ; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm11 = [0,11,27,0,12,28,0,13,29,0,14,30,0,15,31,0] ; AVX512DQ-BW-FCP-NEXT: vpermi2w %ymm6, %ymm10, %ymm11 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm11, %zmm0, %zmm6 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm10 = [42,17,18,43,20,21,44,23,24,45,26,27,46,29,30,47] ; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} ymm11 = 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] ; AVX512DQ-BW-FCP-NEXT: vpermi2w %zmm11, %zmm6, %zmm10 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm10, %zmm8, %zmm6 ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rsi), %xmm8 ; AVX512DQ-BW-FCP-NEXT: vpbroadcastq {{.*#+}} xmm10 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0] ; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm10, %xmm8, %xmm11 ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdi), %xmm12 ; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm10, %xmm12, %xmm10 ; AVX512DQ-BW-FCP-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] ; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm10 = ymm10[0,0,0,1] ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rcx), %xmm11 ; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm7, %xmm11, %xmm13 ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdx), %xmm14 ; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm7, %xmm14, %xmm7 ; AVX512DQ-BW-FCP-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] ; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,0,0,1] ; AVX512DQ-BW-FCP-NEXT: vmovdqu16 %ymm10, %ymm7 {%k1} ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%r9), %xmm10 ; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm9, %xmm10, %xmm13 ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%r8), %xmm15 ; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm9, %xmm15, %xmm9 ; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm9 = xmm9[0],xmm13[0],xmm9[1],xmm13[1],xmm9[2],xmm13[2],xmm9[3],xmm13[3],xmm9[4],xmm13[4],xmm9[5],xmm13[5],xmm9[6],xmm13[6],xmm9[7],xmm13[7] ; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,0,0,1] ; AVX512DQ-BW-FCP-NEXT: vmovdqu16 %ymm9, %ymm7 {%k2} ; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm9 = 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] ; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm13 = 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] ; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm16 = [0,16,0,1,17,0,2,18,0,3,19,0,4,20,0,5] ; AVX512DQ-BW-FCP-NEXT: vpermi2w %ymm9, %ymm13, %ymm16 ; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm9 = 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] ; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm13 = [0,1,16,3,4,17,6,7,18,9,10,19,12,13,20,15] ; AVX512DQ-BW-FCP-NEXT: vpermi2w %ymm9, %ymm16, %ymm13 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm7, %zmm13, %zmm7 ; AVX512DQ-BW-FCP-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] ; AVX512DQ-BW-FCP-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] ; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm11 = [0,3,19,0,4,20,0,5,21,0,6,22,0,7,23,0] ; AVX512DQ-BW-FCP-NEXT: vpermi2w %ymm9, %ymm8, %ymm11 ; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm8 = xmm15[8],xmm10[8],xmm15[9],xmm10[9],xmm15[10],xmm10[10],xmm15[11],xmm10[11],xmm15[12],xmm10[12],xmm15[13],xmm10[13],xmm15[14],xmm10[14],xmm15[15],xmm10[15] ; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm9 = [18,1,2,19,4,5,20,7,8,21,10,11,22,13,14,23] ; AVX512DQ-BW-FCP-NEXT: vpermi2w %ymm8, %ymm11, %ymm9 ; AVX512DQ-BW-FCP-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] ; AVX512DQ-BW-FCP-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] ; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm3 = [8,24,0,9,25,0,10,26,0,11,27,0,12,28,0,13] ; AVX512DQ-BW-FCP-NEXT: vpermi2w %ymm4, %ymm2, %ymm3 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm3, %zmm0, %zmm2 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm3 = [16,17,40,19,20,41,22,23,42,25,26,43,28,29,44,31] ; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} ymm0 = 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] ; AVX512DQ-BW-FCP-NEXT: vpermi2w %zmm0, %zmm2, %zmm3 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm3, %zmm9, %zmm0 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, 64(%rax) ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm7, (%rax) ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm6, 128(%rax) ; AVX512DQ-BW-FCP-NEXT: vzeroupper ; AVX512DQ-BW-FCP-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> %2 = shufflevector <32 x i8> %in.vec2, <32 x i8> %in.vec3, <64 x i32> %3 = shufflevector <32 x i8> %in.vec4, <32 x i8> %in.vec5, <64 x i32> %4 = shufflevector <64 x i8> %1, <64 x i8> %2, <128 x i32> %5 = shufflevector <64 x i8> %3, <64 x i8> poison, <128 x i32> %6 = shufflevector <128 x i8> %4, <128 x i8> %5, <192 x i32> %interleaved.vec = shufflevector <192 x i8> %6, <192 x i8> poison, <192 x i32> 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 {{.*#+}} xmm12 = [65535,0,65535,65535,0,65535,65535,0] ; SSE-NEXT: pand %xmm12, %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: 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 {{.*#+}} xmm15 = xmm15[0],xmm2[0],xmm15[1],xmm2[1],xmm15[2],xmm2[2],xmm15[3],xmm2[3],xmm15[4],xmm2[4],xmm15[5],xmm2[5],xmm15[6],xmm2[6],xmm15[7],xmm2[7] ; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm15[0,0,0,0] ; SSE-NEXT: movdqa %xmm12, %xmm0 ; SSE-NEXT: pandn %xmm7, %xmm0 ; SSE-NEXT: por %xmm13, %xmm0 ; SSE-NEXT: movdqa %xmm0, {{[-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 {{.*#+}} xmm0 = [0,65535,65535,0,65535,65535,0,65535] ; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm11[1,1,2,2] ; SSE-NEXT: movdqa %xmm0, %xmm14 ; SSE-NEXT: pandn %xmm7, %xmm14 ; SSE-NEXT: pand %xmm0, %xmm13 ; SSE-NEXT: por %xmm13, %xmm14 ; SSE-NEXT: movdqa {{.*#+}} xmm13 = [255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255] ; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm15[1,1,2,2] ; SSE-NEXT: movdqa %xmm13, %xmm0 ; SSE-NEXT: pandn %xmm7, %xmm0 ; SSE-NEXT: pand %xmm13, %xmm14 ; SSE-NEXT: por %xmm14, %xmm0 ; SSE-NEXT: movdqa %xmm0, {{[-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 {{.*#+}} xmm0 = [65535,0,65535,65535,0,65535,65535,0] ; SSE-NEXT: movdqa %xmm0, %xmm10 ; SSE-NEXT: pandn %xmm8, %xmm10 ; SSE-NEXT: pand %xmm0, %xmm7 ; SSE-NEXT: movdqa %xmm0, %xmm11 ; SSE-NEXT: por %xmm7, %xmm10 ; SSE-NEXT: movdqa {{.*#+}} xmm14 = [255,255,255,0,255,255,255,255,255,0,255,255,255,255,255,0] ; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm15[2,2,3,3] ; SSE-NEXT: movdqa %xmm14, %xmm0 ; SSE-NEXT: pandn %xmm7, %xmm0 ; SSE-NEXT: pand %xmm14, %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 %xmm11, %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 %xmm11, %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: pand %xmm12, %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 %xmm12, %xmm0 ; 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), %xmm1 ; 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 %xmm11, %xmm6 ; SSE-NEXT: pandn %xmm5, %xmm6 ; SSE-NEXT: movdqa 16(%rcx), %xmm5 ; SSE-NEXT: pand %xmm11, %xmm4 ; SSE-NEXT: por %xmm4, %xmm6 ; SSE-NEXT: pand %xmm14, %xmm6 ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm2[2,2,3,3] ; SSE-NEXT: movdqa %xmm14, %xmm0 ; 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 %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 %xmm11, %xmm0 ; SSE-NEXT: movdqa %xmm11, %xmm6 ; SSE-NEXT: pandn %xmm2, %xmm6 ; SSE-NEXT: movdqa 16(%rdi), %xmm2 ; SSE-NEXT: movdqa 16(%rsi), %xmm8 ; SSE-NEXT: movdqa %xmm2, %xmm11 ; SSE-NEXT: punpcklbw {{.*#+}} xmm11 = xmm11[0],xmm8[0],xmm11[1],xmm8[1],xmm11[2],xmm8[2],xmm11[3],xmm8[3],xmm11[4],xmm8[4],xmm11[5],xmm8[5],xmm11[6],xmm8[6],xmm11[7],xmm8[7] ; SSE-NEXT: pshufd {{.*#+}} xmm4 = xmm11[0,0,1,1] ; SSE-NEXT: pand %xmm0, %xmm4 ; SSE-NEXT: por %xmm4, %xmm6 ; SSE-NEXT: movdqa 16(%r8), %xmm4 ; SSE-NEXT: movdqa %xmm4, %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, %xmm13 ; SSE-NEXT: pandn %xmm7, %xmm13 ; SSE-NEXT: pand %xmm3, %xmm6 ; SSE-NEXT: por %xmm6, %xmm13 ; SSE-NEXT: movdqa 16(%r9), %xmm6 ; SSE-NEXT: punpcklbw {{.*#+}} xmm7 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3],xmm7[4],xmm6[4],xmm7[5],xmm6[5],xmm7[6],xmm6[6],xmm7[7],xmm6[7] ; SSE-NEXT: pshufd {{.*#+}} xmm15 = xmm7[0,0,0,0] ; SSE-NEXT: movdqa %xmm12, %xmm0 ; SSE-NEXT: pandn %xmm15, %xmm0 ; SSE-NEXT: pand %xmm12, %xmm13 ; SSE-NEXT: por %xmm13, %xmm0 ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: pshuflw {{.*#+}} xmm13 = xmm9[3,3,3,3,4,5,6,7] ; SSE-NEXT: pshufhw {{.*#+}} xmm13 = xmm13[0,1,2,3,4,4,4,4] ; SSE-NEXT: movdqa %xmm3, %xmm15 ; SSE-NEXT: pandn %xmm13, %xmm15 ; SSE-NEXT: pshufd {{.*#+}} xmm13 = xmm11[1,1,2,2] ; SSE-NEXT: pand %xmm3, %xmm13 ; SSE-NEXT: por %xmm13, %xmm15 ; SSE-NEXT: pshufd {{.*#+}} xmm13 = 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 %xmm13, %xmm12 ; SSE-NEXT: pand %xmm0, %xmm15 ; SSE-NEXT: por %xmm15, %xmm12 ; SSE-NEXT: pshufd {{.*#+}} xmm13 = xmm7[1,1,2,2] ; SSE-NEXT: movdqa {{.*#+}} xmm15 = [255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255] ; SSE-NEXT: movdqa %xmm15, %xmm0 ; SSE-NEXT: pandn %xmm13, %xmm0 ; SSE-NEXT: pand %xmm15, %xmm12 ; SSE-NEXT: movdqa %xmm15, %xmm13 ; SSE-NEXT: por %xmm12, %xmm0 ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: pshufd {{.*#+}} xmm11 = xmm11[3,3,3,3] ; SSE-NEXT: movdqa %xmm3, %xmm12 ; SSE-NEXT: pandn %xmm11, %xmm12 ; SSE-NEXT: pshufhw {{.*#+}} xmm9 = xmm9[0,1,2,3,5,6,7,7] ; SSE-NEXT: pshufd {{.*#+}} xmm9 = xmm9[2,2,2,3] ; SSE-NEXT: pand %xmm3, %xmm9 ; SSE-NEXT: por %xmm12, %xmm9 ; SSE-NEXT: pshufd {{.*#+}} xmm10 = xmm10[2,2,3,3] ; SSE-NEXT: movdqa {{.*#+}} xmm0 = [65535,0,65535,65535,0,65535,65535,0] ; SSE-NEXT: movdqa %xmm0, %xmm11 ; SSE-NEXT: pandn %xmm10, %xmm11 ; SSE-NEXT: pand %xmm0, %xmm9 ; SSE-NEXT: movdqa %xmm0, %xmm10 ; SSE-NEXT: por %xmm9, %xmm11 ; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm7[2,2,3,3] ; SSE-NEXT: movdqa %xmm14, %xmm0 ; SSE-NEXT: pandn %xmm7, %xmm0 ; SSE-NEXT: pand %xmm14, %xmm11 ; SSE-NEXT: por %xmm11, %xmm0 ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: punpckhbw {{.*#+}} xmm2 = xmm2[8],xmm8[8],xmm2[9],xmm8[9],xmm2[10],xmm8[10],xmm2[11],xmm8[11],xmm2[12],xmm8[12],xmm2[13],xmm8[13],xmm2[14],xmm8[14],xmm2[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 %xmm10, %xmm7 ; SSE-NEXT: pandn %xmm5, %xmm7 ; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm2[0,0,1,1] ; SSE-NEXT: pand %xmm10, %xmm5 ; SSE-NEXT: por %xmm5, %xmm7 ; SSE-NEXT: punpckhbw {{.*#+}} xmm4 = xmm4[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15] ; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm4[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 = xmm2[1,1,2,2] ; SSE-NEXT: pand %xmm3, %xmm6 ; SSE-NEXT: por %xmm6, %xmm7 ; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm4[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 %xmm13, %xmm0 ; SSE-NEXT: pandn %xmm6, %xmm0 ; SSE-NEXT: pand %xmm13, %xmm8 ; SSE-NEXT: por %xmm8, %xmm0 ; SSE-NEXT: movdqa %xmm0, (%rsp) # 16-byte Spill ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm2[3,3,3,3] ; SSE-NEXT: movdqa %xmm3, %xmm6 ; SSE-NEXT: pandn %xmm2, %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 = xmm4[2,2,3,3] ; SSE-NEXT: movdqa %xmm10, %xmm4 ; SSE-NEXT: pandn %xmm2, %xmm4 ; SSE-NEXT: pand %xmm10, %xmm1 ; SSE-NEXT: por %xmm1, %xmm4 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm5[2,2,3,3] ; SSE-NEXT: movdqa %xmm14, %xmm0 ; SSE-NEXT: pandn %xmm1, %xmm0 ; SSE-NEXT: pand %xmm14, %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 %xmm10, %xmm6 ; SSE-NEXT: pandn %xmm2, %xmm6 ; SSE-NEXT: movdqa 32(%rdi), %xmm2 ; SSE-NEXT: movdqa 32(%rsi), %xmm8 ; SSE-NEXT: movdqa %xmm2, %xmm11 ; SSE-NEXT: punpcklbw {{.*#+}} xmm11 = xmm11[0],xmm8[0],xmm11[1],xmm8[1],xmm11[2],xmm8[2],xmm11[3],xmm8[3],xmm11[4],xmm8[4],xmm11[5],xmm8[5],xmm11[6],xmm8[6],xmm11[7],xmm8[7] ; SSE-NEXT: pshufd {{.*#+}} xmm4 = xmm11[0,0,1,1] ; SSE-NEXT: pand %xmm10, %xmm4 ; SSE-NEXT: por %xmm4, %xmm6 ; SSE-NEXT: movdqa 32(%r8), %xmm4 ; SSE-NEXT: movdqa %xmm4, %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 {{.*#+}} xmm7 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3],xmm7[4],xmm6[4],xmm7[5],xmm6[5],xmm7[6],xmm6[6],xmm7[7],xmm6[7] ; SSE-NEXT: pshufd {{.*#+}} xmm13 = xmm7[0,0,0,0] ; SSE-NEXT: movdqa %xmm15, %xmm0 ; SSE-NEXT: pandn %xmm13, %xmm0 ; SSE-NEXT: pand %xmm15, %xmm12 ; SSE-NEXT: por %xmm12, %xmm0 ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: pshuflw {{.*#+}} xmm12 = xmm9[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 = xmm11[1,1,2,2] ; SSE-NEXT: pand %xmm3, %xmm12 ; SSE-NEXT: por %xmm12, %xmm13 ; SSE-NEXT: pshufd {{.*#+}} xmm12 = xmm10[1,1,2,2] ; SSE-NEXT: movdqa {{.*#+}} xmm0 = [0,65535,65535,0,65535,65535,0,65535] ; SSE-NEXT: movdqa %xmm0, %xmm15 ; SSE-NEXT: pandn %xmm12, %xmm15 ; SSE-NEXT: pand %xmm0, %xmm13 ; SSE-NEXT: por %xmm13, %xmm15 ; SSE-NEXT: pshufd {{.*#+}} xmm12 = xmm7[1,1,2,2] ; SSE-NEXT: movdqa {{.*#+}} xmm13 = [255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255] ; SSE-NEXT: movdqa %xmm13, %xmm0 ; SSE-NEXT: pandn %xmm12, %xmm0 ; SSE-NEXT: pand %xmm13, %xmm15 ; SSE-NEXT: por %xmm15, %xmm0 ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: pshufd {{.*#+}} xmm11 = xmm11[3,3,3,3] ; SSE-NEXT: movdqa %xmm3, %xmm12 ; SSE-NEXT: pandn %xmm11, %xmm12 ; SSE-NEXT: pshufhw {{.*#+}} xmm9 = xmm9[0,1,2,3,5,6,7,7] ; SSE-NEXT: pshufd {{.*#+}} xmm9 = xmm9[2,2,2,3] ; SSE-NEXT: pand %xmm3, %xmm9 ; SSE-NEXT: por %xmm12, %xmm9 ; SSE-NEXT: pshufd {{.*#+}} xmm10 = xmm10[2,2,3,3] ; SSE-NEXT: movdqa {{.*#+}} xmm0 = [65535,0,65535,65535,0,65535,65535,0] ; SSE-NEXT: movdqa %xmm0, %xmm11 ; SSE-NEXT: pandn %xmm10, %xmm11 ; SSE-NEXT: pand %xmm0, %xmm9 ; SSE-NEXT: movdqa %xmm0, %xmm10 ; SSE-NEXT: por %xmm9, %xmm11 ; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm7[2,2,3,3] ; SSE-NEXT: movdqa %xmm14, %xmm0 ; SSE-NEXT: pandn %xmm7, %xmm0 ; SSE-NEXT: pand %xmm14, %xmm11 ; SSE-NEXT: por %xmm11, %xmm0 ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: punpckhbw {{.*#+}} xmm2 = xmm2[8],xmm8[8],xmm2[9],xmm8[9],xmm2[10],xmm8[10],xmm2[11],xmm8[11],xmm2[12],xmm8[12],xmm2[13],xmm8[13],xmm2[14],xmm8[14],xmm2[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 %xmm10, %xmm7 ; SSE-NEXT: pandn %xmm5, %xmm7 ; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm2[0,0,1,1] ; SSE-NEXT: pand %xmm10, %xmm5 ; SSE-NEXT: por %xmm5, %xmm7 ; SSE-NEXT: punpckhbw {{.*#+}} xmm4 = xmm4[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15] ; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm4[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 = xmm2[1,1,2,2] ; SSE-NEXT: pand %xmm3, %xmm6 ; SSE-NEXT: por %xmm6, %xmm7 ; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm4[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 %xmm13, %xmm7 ; SSE-NEXT: pandn %xmm6, %xmm7 ; SSE-NEXT: pand %xmm13, %xmm8 ; SSE-NEXT: por %xmm8, %xmm7 ; SSE-NEXT: movdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm2[3,3,3,3] ; SSE-NEXT: movdqa %xmm3, %xmm6 ; SSE-NEXT: pandn %xmm2, %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 = xmm4[2,2,3,3] ; SSE-NEXT: movdqa %xmm10, %xmm4 ; SSE-NEXT: pandn %xmm2, %xmm4 ; SSE-NEXT: pand %xmm10, %xmm1 ; SSE-NEXT: por %xmm1, %xmm4 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm5[2,2,3,3] ; SSE-NEXT: movdqa %xmm14, %xmm0 ; SSE-NEXT: pandn %xmm1, %xmm0 ; SSE-NEXT: pand %xmm14, %xmm4 ; SSE-NEXT: por %xmm4, %xmm0 ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: movdqa 48(%rdx), %xmm9 ; SSE-NEXT: movdqa 48(%rcx), %xmm0 ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: movdqa %xmm9, %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 %xmm10, %xmm0 ; SSE-NEXT: movdqa %xmm10, %xmm4 ; SSE-NEXT: pandn %xmm1, %xmm4 ; SSE-NEXT: movdqa 48(%rdi), %xmm6 ; SSE-NEXT: movdqa 48(%rsi), %xmm10 ; SSE-NEXT: movdqa %xmm6, %xmm1 ; SSE-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm10[0],xmm1[1],xmm10[1],xmm1[2],xmm10[2],xmm1[3],xmm10[3],xmm1[4],xmm10[4],xmm1[5],xmm10[5],xmm1[6],xmm10[6],xmm1[7],xmm10[7] ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm1[0,0,1,1] ; SSE-NEXT: pand %xmm0, %xmm2 ; SSE-NEXT: por %xmm2, %xmm4 ; 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 {{.*#+}} xmm11 = xmm2[0,0,0,0] ; SSE-NEXT: movdqa %xmm3, %xmm12 ; SSE-NEXT: pandn %xmm11, %xmm12 ; SSE-NEXT: pand %xmm3, %xmm4 ; SSE-NEXT: por %xmm4, %xmm12 ; SSE-NEXT: movdqa 48(%r9), %xmm11 ; SSE-NEXT: punpcklbw {{.*#+}} xmm5 = xmm5[0],xmm11[0],xmm5[1],xmm11[1],xmm5[2],xmm11[2],xmm5[3],xmm11[3],xmm5[4],xmm11[4],xmm5[5],xmm11[5],xmm5[6],xmm11[6],xmm5[7],xmm11[7] ; SSE-NEXT: pshufd {{.*#+}} xmm13 = xmm5[0,0,0,0] ; SSE-NEXT: movdqa %xmm15, %xmm0 ; SSE-NEXT: pandn %xmm13, %xmm15 ; SSE-NEXT: pand %xmm0, %xmm12 ; SSE-NEXT: por %xmm12, %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 = xmm1[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 {{.*#+}} xmm4 = [0,65535,65535,0,65535,65535,0,65535] ; SSE-NEXT: movdqa %xmm4, %xmm0 ; SSE-NEXT: pandn %xmm12, %xmm0 ; SSE-NEXT: pand %xmm4, %xmm13 ; SSE-NEXT: por %xmm13, %xmm0 ; SSE-NEXT: pshufd {{.*#+}} xmm12 = xmm5[1,1,2,2] ; SSE-NEXT: movdqa {{.*#+}} xmm4 = [255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255] ; SSE-NEXT: movdqa %xmm4, %xmm13 ; SSE-NEXT: pandn %xmm12, %xmm13 ; SSE-NEXT: pand %xmm4, %xmm0 ; SSE-NEXT: por %xmm0, %xmm13 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm1[3,3,3,3] ; SSE-NEXT: movdqa %xmm3, %xmm1 ; SSE-NEXT: pandn %xmm0, %xmm1 ; 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 %xmm1, %xmm0 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm2[2,2,3,3] ; SSE-NEXT: movdqa {{.*#+}} xmm12 = [65535,0,65535,65535,0,65535,65535,0] ; SSE-NEXT: movdqa %xmm12, %xmm2 ; SSE-NEXT: pandn %xmm1, %xmm2 ; SSE-NEXT: pand %xmm12, %xmm0 ; SSE-NEXT: por %xmm0, %xmm2 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm5[2,2,3,3] ; SSE-NEXT: movdqa %xmm14, %xmm8 ; SSE-NEXT: pandn %xmm0, %xmm8 ; SSE-NEXT: pand %xmm14, %xmm2 ; SSE-NEXT: por %xmm2, %xmm8 ; SSE-NEXT: punpckhbw {{.*#+}} xmm6 = xmm6[8],xmm10[8],xmm6[9],xmm10[9],xmm6[10],xmm10[10],xmm6[11],xmm10[11],xmm6[12],xmm10[12],xmm6[13],xmm10[13],xmm6[14],xmm10[14],xmm6[15],xmm10[15] ; SSE-NEXT: punpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Folded Reload ; SSE-NEXT: # xmm9 = xmm9[8],mem[8],xmm9[9],mem[9],xmm9[10],mem[10],xmm9[11],mem[11],xmm9[12],mem[12],xmm9[13],mem[13],xmm9[14],mem[14],xmm9[15],mem[15] ; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm9[1,0,2,2,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,1,0,1] ; SSE-NEXT: movdqa %xmm12, %xmm1 ; SSE-NEXT: pandn %xmm0, %xmm1 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm6[0,0,1,1] ; SSE-NEXT: pand %xmm12, %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 {{.*#+}} xmm1 = xmm1[8],xmm11[8],xmm1[9],xmm11[9],xmm1[10],xmm11[10],xmm1[11],xmm11[11],xmm1[12],xmm11[12],xmm1[13],xmm11[13],xmm1[14],xmm11[14],xmm1[15],xmm11[15] ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm1[0,0,0,0] ; SSE-NEXT: pandn %xmm0, %xmm4 ; SSE-NEXT: por %xmm2, %xmm4 ; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm9[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 {{.*#+}} xmm10 = [0,65535,65535,0,65535,65535,0,65535] ; SSE-NEXT: pand %xmm10, %xmm2 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm7[1,1,2,2] ; SSE-NEXT: pandn %xmm0, %xmm10 ; SSE-NEXT: por %xmm2, %xmm10 ; SSE-NEXT: movdqa {{.*#+}} xmm2 = [255,0,255,255,255,255,255,0,255,255,255,255,255,0,255,255] ; SSE-NEXT: pand %xmm2, %xmm10 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,2,2] ; SSE-NEXT: pandn %xmm0, %xmm2 ; SSE-NEXT: por %xmm10, %xmm2 ; SSE-NEXT: movdqa %xmm2, %xmm10 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm6[3,3,3,3] ; SSE-NEXT: pshufhw {{.*#+}} xmm2 = xmm9[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: movdqa %xmm12, %xmm2 ; SSE-NEXT: pand %xmm12, %xmm3 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm7[2,2,3,3] ; SSE-NEXT: pandn %xmm0, %xmm2 ; SSE-NEXT: por %xmm3, %xmm2 ; SSE-NEXT: pand %xmm14, %xmm2 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,2,3,3] ; SSE-NEXT: pandn %xmm0, %xmm14 ; SSE-NEXT: por %xmm2, %xmm14 ; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax ; SSE-NEXT: movdqa %xmm14, 368(%rax) ; SSE-NEXT: movdqa %xmm10, 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: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload ; SSE-NEXT: movaps %xmm0, 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 ; ; AVX-LABEL: store_i8_stride6_vf64: ; AVX: # %bb.0: ; AVX-NEXT: subq $200, %rsp ; AVX-NEXT: vmovdqa 48(%rsi), %xmm1 ; AVX-NEXT: vmovdqa 48(%rdi), %xmm2 ; AVX-NEXT: vpunpcklbw {{.*#+}} xmm0 = 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] ; AVX-NEXT: vpshufd {{.*#+}} xmm3 = xmm0[3,3,3,3] ; AVX-NEXT: vpunpckhbw {{.*#+}} xmm1 = 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] ; AVX-NEXT: vpshufd {{.*#+}} xmm2 = xmm1[0,0,1,1] ; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm3, %ymm2 ; AVX-NEXT: vmovaps {{.*#+}} ymm10 = [65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535] ; AVX-NEXT: vandnps %ymm2, %ymm10, %ymm2 ; AVX-NEXT: vmovdqa 48(%rcx), %xmm4 ; AVX-NEXT: vmovdqa 48(%rdx), %xmm5 ; AVX-NEXT: vpunpcklbw {{.*#+}} xmm3 = 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] ; AVX-NEXT: vpshufhw {{.*#+}} xmm6 = xmm3[0,1,2,3,5,6,7,7] ; AVX-NEXT: vpshufd {{.*#+}} xmm6 = xmm6[2,2,2,3] ; AVX-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] ; AVX-NEXT: vpshuflw {{.*#+}} xmm5 = xmm4[1,0,2,2,4,5,6,7] ; AVX-NEXT: vpshufd {{.*#+}} xmm5 = xmm5[0,1,0,1] ; AVX-NEXT: vinsertf128 $1, %xmm5, %ymm6, %ymm5 ; AVX-NEXT: vandps %ymm5, %ymm10, %ymm5 ; AVX-NEXT: vorps %ymm2, %ymm5, %ymm6 ; AVX-NEXT: vextractf128 $1, %ymm6, %xmm5 ; AVX-NEXT: vmovdqa 48(%r8), %xmm2 ; AVX-NEXT: vpshufb {{.*#+}} xmm7 = zero,zero,zero,zero,xmm2[8,u],zero,zero,zero,zero,xmm2[9,u],zero,zero,zero,zero ; AVX-NEXT: vpblendw {{.*#+}} xmm5 = xmm5[0,1],xmm7[2],xmm5[3,4],xmm7[5],xmm5[6,7] ; AVX-NEXT: vpshufb {{.*#+}} xmm7 = xmm5[0,1,2,3,4],zero,xmm5[6,7,8,9,10],zero,xmm5[12,13,14,15] ; AVX-NEXT: vmovdqa 48(%r9), %xmm5 ; AVX-NEXT: vpshufb {{.*#+}} xmm8 = zero,zero,zero,zero,zero,xmm5[8],zero,zero,zero,zero,zero,xmm5[9],zero,zero,zero,zero ; AVX-NEXT: vpor %xmm7, %xmm8, %xmm7 ; AVX-NEXT: vmovdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX-NEXT: vpshufb {{.*#+}} xmm7 = zero,zero,xmm2[5,u],zero,zero,zero,zero,xmm2[6,u],zero,zero,zero,zero,xmm2[7,u] ; AVX-NEXT: vpblendw {{.*#+}} xmm6 = xmm6[0],xmm7[1],xmm6[2,3],xmm7[4],xmm6[5,6],xmm7[7] ; AVX-NEXT: vmovdqa {{.*#+}} xmm8 = [0,1,2,128,4,5,6,7,8,128,10,11,12,13,14,128] ; AVX-NEXT: vpshufb %xmm8, %xmm6, %xmm6 ; AVX-NEXT: vpshufb {{.*#+}} xmm7 = zero,zero,zero,xmm5[5],zero,zero,zero,zero,zero,xmm5[6],zero,zero,zero,zero,zero,xmm5[7] ; AVX-NEXT: vpor %xmm7, %xmm6, %xmm6 ; AVX-NEXT: vmovdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX-NEXT: vpshufd {{.*#+}} xmm6 = xmm1[1,1,2,2] ; AVX-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[3,3,3,3] ; AVX-NEXT: vinsertf128 $1, %xmm1, %ymm6, %ymm1 ; AVX-NEXT: vandps %ymm1, %ymm10, %ymm1 ; AVX-NEXT: vpshuflw {{.*#+}} xmm6 = xmm4[3,3,3,3,4,5,6,7] ; AVX-NEXT: vpshufhw {{.*#+}} xmm6 = xmm6[0,1,2,3,4,4,4,4] ; AVX-NEXT: vpshufhw {{.*#+}} xmm4 = xmm4[0,1,2,3,5,6,7,7] ; AVX-NEXT: vpshufd {{.*#+}} xmm4 = xmm4[2,2,2,3] ; AVX-NEXT: vinsertf128 $1, %xmm4, %ymm6, %ymm4 ; AVX-NEXT: vandnps %ymm4, %ymm10, %ymm4 ; AVX-NEXT: vorps %ymm4, %ymm1, %ymm1 ; AVX-NEXT: vextractf128 $1, %ymm1, %xmm4 ; AVX-NEXT: vmovdqa {{.*#+}} xmm15 = [128,128,13,u,128,128,128,128,14,u,128,128,128,128,15,u] ; AVX-NEXT: vpshufb %xmm15, %xmm2, %xmm6 ; AVX-NEXT: vpblendw {{.*#+}} xmm4 = xmm4[0],xmm6[1],xmm4[2,3],xmm6[4],xmm4[5,6],xmm6[7] ; AVX-NEXT: vpshufb %xmm8, %xmm4, %xmm4 ; AVX-NEXT: vpshufb {{.*#+}} xmm6 = zero,zero,zero,xmm5[13],zero,zero,zero,zero,zero,xmm5[14],zero,zero,zero,zero,zero,xmm5[15] ; AVX-NEXT: vpor %xmm6, %xmm4, %xmm4 ; AVX-NEXT: vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX-NEXT: vpshufb {{.*#+}} xmm4 = xmm2[10,u],zero,zero,zero,zero,xmm2[11,u],zero,zero,zero,zero,xmm2[12,u],zero,zero ; AVX-NEXT: vpblendw {{.*#+}} xmm1 = xmm4[0],xmm1[1,2],xmm4[3],xmm1[4,5],xmm4[6],xmm1[7] ; AVX-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[0],zero,xmm1[2,3,4,5,6],zero,xmm1[8,9,10,11,12],zero,xmm1[14,15] ; AVX-NEXT: vmovdqa {{.*#+}} xmm13 = [128,10,128,128,128,128,128,11,128,128,128,128,128,12,128,128] ; AVX-NEXT: vpshufb %xmm13, %xmm5, %xmm4 ; AVX-NEXT: vpor %xmm4, %xmm1, %xmm1 ; AVX-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX-NEXT: vmovdqa 32(%rsi), %xmm6 ; AVX-NEXT: vmovdqa 32(%rdi), %xmm11 ; AVX-NEXT: vpunpckhbw {{.*#+}} xmm7 = xmm11[8],xmm6[8],xmm11[9],xmm6[9],xmm11[10],xmm6[10],xmm11[11],xmm6[11],xmm11[12],xmm6[12],xmm11[13],xmm6[13],xmm11[14],xmm6[14],xmm11[15],xmm6[15] ; AVX-NEXT: vpshufd {{.*#+}} xmm1 = xmm7[1,1,2,2] ; AVX-NEXT: vpshufd {{.*#+}} xmm4 = xmm7[3,3,3,3] ; AVX-NEXT: vinsertf128 $1, %xmm4, %ymm1, %ymm1 ; AVX-NEXT: vmovdqa 32(%rcx), %xmm12 ; AVX-NEXT: vmovdqa 32(%rdx), %xmm14 ; AVX-NEXT: vpunpckhbw {{.*#+}} xmm4 = 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] ; AVX-NEXT: vpshuflw {{.*#+}} xmm8 = xmm4[3,3,3,3,4,5,6,7] ; AVX-NEXT: vpshufhw {{.*#+}} xmm8 = xmm8[0,1,2,3,4,4,4,4] ; AVX-NEXT: vpshufhw {{.*#+}} xmm9 = xmm4[0,1,2,3,5,6,7,7] ; AVX-NEXT: vpshufd {{.*#+}} xmm9 = xmm9[2,2,2,3] ; AVX-NEXT: vinsertf128 $1, %xmm9, %ymm8, %ymm8 ; AVX-NEXT: vandps %ymm1, %ymm10, %ymm1 ; AVX-NEXT: vandnps %ymm8, %ymm10, %ymm8 ; AVX-NEXT: vorps %ymm1, %ymm8, %ymm9 ; AVX-NEXT: vmovdqa 32(%r8), %xmm1 ; AVX-NEXT: vextractf128 $1, %ymm9, %xmm8 ; AVX-NEXT: vpshufb %xmm15, %xmm1, %xmm10 ; AVX-NEXT: vpblendw {{.*#+}} xmm8 = xmm8[0],xmm10[1],xmm8[2,3],xmm10[4],xmm8[5,6],xmm10[7] ; AVX-NEXT: vpshufb {{.*#+}} xmm10 = xmm8[0,1,2],zero,xmm8[4,5,6,7,8],zero,xmm8[10,11,12,13,14],zero ; AVX-NEXT: vmovdqa 32(%r9), %xmm8 ; AVX-NEXT: vpshufb {{.*#+}} xmm15 = zero,zero,zero,xmm8[13],zero,zero,zero,zero,zero,xmm8[14],zero,zero,zero,zero,zero,xmm8[15] ; AVX-NEXT: vpor %xmm15, %xmm10, %xmm10 ; AVX-NEXT: vmovdqa %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX-NEXT: vpshufb {{.*#+}} xmm10 = xmm1[10,u],zero,zero,zero,zero,xmm1[11,u],zero,zero,zero,zero,xmm1[12,u],zero,zero ; AVX-NEXT: vpblendw {{.*#+}} xmm9 = xmm10[0],xmm9[1,2],xmm10[3],xmm9[4,5],xmm10[6],xmm9[7] ; AVX-NEXT: vmovdqa {{.*#+}} xmm15 = [0,128,2,3,4,5,6,128,8,9,10,11,12,128,14,15] ; AVX-NEXT: vpshufb %xmm15, %xmm9, %xmm9 ; AVX-NEXT: vpshufb %xmm13, %xmm8, %xmm10 ; AVX-NEXT: vpor %xmm10, %xmm9, %xmm9 ; AVX-NEXT: vmovdqa %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX-NEXT: vpshufd {{.*#+}} xmm9 = xmm0[0,0,1,1] ; AVX-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[1,1,2,2] ; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm9, %ymm0 ; AVX-NEXT: vpshuflw {{.*#+}} xmm9 = xmm3[1,0,2,2,4,5,6,7] ; AVX-NEXT: vpshufd {{.*#+}} xmm9 = xmm9[0,1,0,1] ; AVX-NEXT: vpshuflw {{.*#+}} xmm3 = xmm3[3,3,3,3,4,5,6,7] ; AVX-NEXT: vpshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,4,4,4] ; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm9, %ymm3 ; AVX-NEXT: vmovaps {{.*#+}} ymm9 = [65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535] ; AVX-NEXT: vandps %ymm0, %ymm9, %ymm0 ; AVX-NEXT: vandnps %ymm3, %ymm9, %ymm3 ; AVX-NEXT: vorps %ymm3, %ymm0, %ymm0 ; AVX-NEXT: vextractf128 $1, %ymm0, %xmm3 ; AVX-NEXT: vpshufb {{.*#+}} xmm9 = xmm2[2,u],zero,zero,zero,zero,xmm2[3,u],zero,zero,zero,zero,xmm2[4,u],zero,zero ; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm9[0],xmm3[1,2],xmm9[3],xmm3[4,5],xmm9[6],xmm3[7] ; AVX-NEXT: vpshufb %xmm15, %xmm3, %xmm3 ; AVX-NEXT: vmovdqa {{.*#+}} xmm13 = [128,2,128,128,128,128,128,3,128,128,128,128,128,4,128,128] ; AVX-NEXT: vpshufb %xmm13, %xmm5, %xmm9 ; AVX-NEXT: vpor %xmm3, %xmm9, %xmm3 ; AVX-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX-NEXT: vmovdqa {{.*#+}} xmm10 = [128,128,128,128,0,u,128,128,128,128,1,u,128,128,128,128] ; AVX-NEXT: vpshufb %xmm10, %xmm2, %xmm2 ; AVX-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0,1],xmm2[2],xmm0[3,4],xmm2[5],xmm0[6,7] ; AVX-NEXT: vmovdqa {{.*#+}} xmm9 = [0,1,2,3,4,128,6,7,8,9,10,128,12,13,14,15] ; AVX-NEXT: vpshufb %xmm9, %xmm0, %xmm0 ; AVX-NEXT: vmovdqa {{.*#+}} xmm15 = [128,128,128,128,128,0,128,128,128,128,128,1,128,128,128,128] ; AVX-NEXT: vpshufb %xmm15, %xmm5, %xmm2 ; AVX-NEXT: vpor %xmm2, %xmm0, %xmm0 ; AVX-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm11[0],xmm6[0],xmm11[1],xmm6[1],xmm11[2],xmm6[2],xmm11[3],xmm6[3],xmm11[4],xmm6[4],xmm11[5],xmm6[5],xmm11[6],xmm6[6],xmm11[7],xmm6[7] ; AVX-NEXT: vpshufd {{.*#+}} xmm2 = xmm0[0,0,1,1] ; AVX-NEXT: vpshufd {{.*#+}} xmm3 = xmm0[1,1,2,2] ; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm2, %ymm2 ; AVX-NEXT: vpunpcklbw {{.*#+}} xmm3 = 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] ; AVX-NEXT: vpshuflw {{.*#+}} xmm5 = xmm3[1,0,2,2,4,5,6,7] ; AVX-NEXT: vpshufd {{.*#+}} xmm5 = xmm5[0,1,0,1] ; AVX-NEXT: vpshuflw {{.*#+}} xmm6 = xmm3[3,3,3,3,4,5,6,7] ; AVX-NEXT: vpshufhw {{.*#+}} xmm6 = xmm6[0,1,2,3,4,4,4,4] ; AVX-NEXT: vinsertf128 $1, %xmm6, %ymm5, %ymm5 ; AVX-NEXT: vmovaps {{.*#+}} ymm6 = [65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535] ; AVX-NEXT: vandps %ymm6, %ymm2, %ymm2 ; AVX-NEXT: vandnps %ymm5, %ymm6, %ymm5 ; AVX-NEXT: vorps %ymm5, %ymm2, %ymm2 ; AVX-NEXT: vextractf128 $1, %ymm2, %xmm5 ; AVX-NEXT: vpshufb {{.*#+}} xmm6 = xmm1[2,u],zero,zero,zero,zero,xmm1[3,u],zero,zero,zero,zero,xmm1[4,u],zero,zero ; AVX-NEXT: vpblendw {{.*#+}} xmm5 = xmm6[0],xmm5[1,2],xmm6[3],xmm5[4,5],xmm6[6],xmm5[7] ; AVX-NEXT: vpshufb {{.*#+}} xmm5 = xmm5[0],zero,xmm5[2,3,4,5,6],zero,xmm5[8,9,10,11,12],zero,xmm5[14,15] ; AVX-NEXT: vpshufb %xmm13, %xmm8, %xmm6 ; AVX-NEXT: vpor %xmm6, %xmm5, %xmm5 ; AVX-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX-NEXT: vpshufb %xmm10, %xmm1, %xmm5 ; AVX-NEXT: vpblendw {{.*#+}} xmm2 = xmm2[0,1],xmm5[2],xmm2[3,4],xmm5[5],xmm2[6,7] ; AVX-NEXT: vpshufb %xmm9, %xmm2, %xmm2 ; AVX-NEXT: vpshufb %xmm15, %xmm8, %xmm5 ; AVX-NEXT: vpor %xmm5, %xmm2, %xmm2 ; AVX-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[3,3,3,3] ; AVX-NEXT: vpshufd {{.*#+}} xmm2 = xmm7[0,0,1,1] ; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm0 ; AVX-NEXT: vpshufhw {{.*#+}} xmm2 = xmm3[0,1,2,3,5,6,7,7] ; AVX-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[2,2,2,3] ; AVX-NEXT: vpshuflw {{.*#+}} xmm3 = xmm4[1,0,2,2,4,5,6,7] ; AVX-NEXT: vpshufd {{.*#+}} xmm3 = xmm3[0,1,0,1] ; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm2, %ymm2 ; AVX-NEXT: vmovaps {{.*#+}} ymm15 = [65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535] ; AVX-NEXT: vandnps %ymm0, %ymm15, %ymm0 ; AVX-NEXT: vandps %ymm2, %ymm15, %ymm2 ; AVX-NEXT: vorps %ymm0, %ymm2, %ymm0 ; AVX-NEXT: vextractf128 $1, %ymm0, %xmm2 ; AVX-NEXT: vmovdqa {{.*#+}} xmm5 = [128,128,128,128,8,u,128,128,128,128,9,u,128,128,128,128] ; AVX-NEXT: vpshufb %xmm5, %xmm1, %xmm3 ; AVX-NEXT: vpblendw {{.*#+}} xmm2 = xmm2[0,1],xmm3[2],xmm2[3,4],xmm3[5],xmm2[6,7] ; AVX-NEXT: vpshufb %xmm9, %xmm2, %xmm2 ; AVX-NEXT: vmovdqa %xmm9, %xmm14 ; AVX-NEXT: vmovdqa {{.*#+}} xmm6 = [128,128,128,128,128,8,128,128,128,128,128,9,128,128,128,128] ; AVX-NEXT: vpshufb %xmm6, %xmm8, %xmm3 ; AVX-NEXT: vpor %xmm3, %xmm2, %xmm2 ; AVX-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX-NEXT: vmovdqa {{.*#+}} xmm7 = [128,128,5,u,128,128,128,128,6,u,128,128,128,128,7,u] ; AVX-NEXT: vpshufb %xmm7, %xmm1, %xmm1 ; AVX-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0],xmm1[1],xmm0[2,3],xmm1[4],xmm0[5,6],xmm1[7] ; AVX-NEXT: vmovdqa {{.*#+}} xmm13 = [0,1,2,128,4,5,6,7,8,128,10,11,12,13,14,128] ; AVX-NEXT: vpshufb %xmm13, %xmm0, %xmm0 ; AVX-NEXT: vmovdqa {{.*#+}} xmm9 = [128,128,128,5,128,128,128,128,128,6,128,128,128,128,128,7] ; AVX-NEXT: vpshufb %xmm9, %xmm8, %xmm1 ; AVX-NEXT: vpor %xmm1, %xmm0, %xmm0 ; AVX-NEXT: vmovdqa %xmm0, (%rsp) # 16-byte Spill ; AVX-NEXT: vmovdqa 16(%rsi), %xmm0 ; AVX-NEXT: vmovdqa 16(%rdi), %xmm1 ; AVX-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] ; AVX-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] ; AVX-NEXT: vpshufd {{.*#+}} xmm1 = xmm11[3,3,3,3] ; AVX-NEXT: vpshufd {{.*#+}} xmm2 = xmm0[0,0,1,1] ; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm1, %ymm1 ; AVX-NEXT: vmovdqa 16(%rcx), %xmm2 ; AVX-NEXT: vmovdqa 16(%rdx), %xmm3 ; AVX-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] ; AVX-NEXT: vpunpckhbw {{.*#+}} xmm4 = 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] ; AVX-NEXT: vpshufhw {{.*#+}} xmm2 = xmm10[0,1,2,3,5,6,7,7] ; AVX-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[2,2,2,3] ; AVX-NEXT: vpshuflw {{.*#+}} xmm3 = xmm4[1,0,2,2,4,5,6,7] ; AVX-NEXT: vpshufd {{.*#+}} xmm3 = xmm3[0,1,0,1] ; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm2, %ymm2 ; AVX-NEXT: vandnps %ymm1, %ymm15, %ymm1 ; AVX-NEXT: vandps %ymm2, %ymm15, %ymm2 ; AVX-NEXT: vorps %ymm1, %ymm2, %ymm1 ; AVX-NEXT: vmovdqa 16(%r8), %xmm3 ; AVX-NEXT: vextractf128 $1, %ymm1, %xmm2 ; AVX-NEXT: vpshufb %xmm5, %xmm3, %xmm5 ; AVX-NEXT: vpblendw {{.*#+}} xmm2 = xmm2[0,1],xmm5[2],xmm2[3,4],xmm5[5],xmm2[6,7] ; AVX-NEXT: vpshufb %xmm14, %xmm2, %xmm5 ; AVX-NEXT: vmovdqa 16(%r9), %xmm2 ; AVX-NEXT: vpshufb %xmm6, %xmm2, %xmm6 ; AVX-NEXT: vpor %xmm6, %xmm5, %xmm5 ; AVX-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX-NEXT: vpshufb %xmm7, %xmm3, %xmm5 ; AVX-NEXT: vpblendw {{.*#+}} xmm1 = xmm1[0],xmm5[1],xmm1[2,3],xmm5[4],xmm1[5,6],xmm5[7] ; AVX-NEXT: vpshufb %xmm13, %xmm1, %xmm1 ; AVX-NEXT: vpshufb %xmm9, %xmm2, %xmm5 ; AVX-NEXT: vpor %xmm5, %xmm1, %xmm1 ; AVX-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,2] ; AVX-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[3,3,3,3] ; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm1, %ymm0 ; AVX-NEXT: vpshuflw {{.*#+}} xmm1 = xmm4[3,3,3,3,4,5,6,7] ; AVX-NEXT: vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,4,4,4] ; AVX-NEXT: vpshufhw {{.*#+}} xmm4 = xmm4[0,1,2,3,5,6,7,7] ; AVX-NEXT: vpshufd {{.*#+}} xmm4 = xmm4[2,2,2,3] ; AVX-NEXT: vinsertf128 $1, %xmm4, %ymm1, %ymm1 ; AVX-NEXT: vandps %ymm0, %ymm15, %ymm0 ; AVX-NEXT: vandnps %ymm1, %ymm15, %ymm1 ; AVX-NEXT: vorps %ymm1, %ymm0, %ymm0 ; AVX-NEXT: vextractf128 $1, %ymm0, %xmm1 ; AVX-NEXT: vmovdqa {{.*#+}} xmm14 = [128,128,13,u,128,128,128,128,14,u,128,128,128,128,15,u] ; AVX-NEXT: vpshufb %xmm14, %xmm3, %xmm4 ; AVX-NEXT: vpblendw {{.*#+}} xmm1 = xmm1[0],xmm4[1],xmm1[2,3],xmm4[4],xmm1[5,6],xmm4[7] ; AVX-NEXT: vpshufb %xmm13, %xmm1, %xmm1 ; AVX-NEXT: vmovdqa {{.*#+}} xmm15 = [128,128,128,13,128,128,128,128,128,14,128,128,128,128,128,15] ; AVX-NEXT: vpshufb %xmm15, %xmm2, %xmm4 ; AVX-NEXT: vpor %xmm4, %xmm1, %xmm1 ; AVX-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX-NEXT: vmovdqa {{.*#+}} xmm13 = [10,u,128,128,128,128,11,u,128,128,128,128,12,u,128,128] ; AVX-NEXT: vpshufb %xmm13, %xmm3, %xmm1 ; AVX-NEXT: vpblendw {{.*#+}} xmm0 = xmm1[0],xmm0[1,2],xmm1[3],xmm0[4,5],xmm1[6],xmm0[7] ; AVX-NEXT: vmovdqa {{.*#+}} xmm13 = [0,128,2,3,4,5,6,128,8,9,10,11,12,128,14,15] ; AVX-NEXT: vpshufb %xmm13, %xmm0, %xmm0 ; AVX-NEXT: vpshufb {{.*#+}} xmm1 = zero,xmm2[10],zero,zero,zero,zero,zero,xmm2[11],zero,zero,zero,zero,zero,xmm2[12],zero,zero ; AVX-NEXT: vpor %xmm1, %xmm0, %xmm0 ; AVX-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX-NEXT: vmovdqa (%rsi), %xmm8 ; AVX-NEXT: vmovdqa (%rdi), %xmm7 ; AVX-NEXT: vpunpckhbw {{.*#+}} xmm1 = 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] ; AVX-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX-NEXT: vpshufd {{.*#+}} xmm0 = xmm1[1,1,2,2] ; AVX-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[3,3,3,3] ; AVX-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0 ; AVX-NEXT: vmovdqa (%rcx), %xmm6 ; AVX-NEXT: vmovdqa (%rdx), %xmm5 ; AVX-NEXT: vpunpckhbw {{.*#+}} xmm9 = 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] ; AVX-NEXT: vpshuflw {{.*#+}} xmm1 = xmm9[3,3,3,3,4,5,6,7] ; AVX-NEXT: vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,4,4,4] ; AVX-NEXT: vpshufhw {{.*#+}} xmm12 = xmm9[0,1,2,3,5,6,7,7] ; AVX-NEXT: vpshufd {{.*#+}} xmm12 = xmm12[2,2,2,3] ; AVX-NEXT: vinsertf128 $1, %xmm12, %ymm1, %ymm1 ; AVX-NEXT: vmovaps {{.*#+}} ymm4 = [65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535] ; AVX-NEXT: vandps %ymm4, %ymm0, %ymm0 ; AVX-NEXT: vandnps %ymm1, %ymm4, %ymm1 ; AVX-NEXT: vorps %ymm1, %ymm0, %ymm12 ; AVX-NEXT: vmovdqa (%r8), %xmm1 ; AVX-NEXT: vpshufb %xmm14, %xmm1, %xmm0 ; AVX-NEXT: vextractf128 $1, %ymm12, %xmm14 ; AVX-NEXT: vpblendw {{.*#+}} xmm0 = xmm14[0],xmm0[1],xmm14[2,3],xmm0[4],xmm14[5,6],xmm0[7] ; AVX-NEXT: vpshufb {{.*#+}} xmm14 = xmm0[0,1,2],zero,xmm0[4,5,6,7,8],zero,xmm0[10,11,12,13,14],zero ; AVX-NEXT: vmovdqa (%r9), %xmm0 ; AVX-NEXT: vpshufb %xmm15, %xmm0, %xmm15 ; AVX-NEXT: vpor %xmm15, %xmm14, %xmm4 ; AVX-NEXT: vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX-NEXT: vpshufb {{.*#+}} xmm15 = xmm1[10,u],zero,zero,zero,zero,xmm1[11,u],zero,zero,zero,zero,xmm1[12,u],zero,zero ; AVX-NEXT: vpblendw {{.*#+}} xmm12 = xmm15[0],xmm12[1,2],xmm15[3],xmm12[4,5],xmm15[6],xmm12[7] ; AVX-NEXT: vmovdqa %xmm13, %xmm4 ; AVX-NEXT: vpshufb %xmm13, %xmm12, %xmm12 ; AVX-NEXT: vpshufb {{.*#+}} xmm15 = zero,xmm0[10],zero,zero,zero,zero,zero,xmm0[11],zero,zero,zero,zero,zero,xmm0[12],zero,zero ; AVX-NEXT: vpor %xmm15, %xmm12, %xmm12 ; AVX-NEXT: vmovdqa %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX-NEXT: vpshufd {{.*#+}} xmm15 = xmm11[0,0,1,1] ; AVX-NEXT: vpshufd {{.*#+}} xmm11 = xmm11[1,1,2,2] ; AVX-NEXT: vinsertf128 $1, %xmm11, %ymm15, %ymm11 ; AVX-NEXT: vpshuflw {{.*#+}} xmm15 = xmm10[1,0,2,2,4,5,6,7] ; AVX-NEXT: vpshufd {{.*#+}} xmm15 = xmm15[0,1,0,1] ; AVX-NEXT: vpshuflw {{.*#+}} xmm10 = xmm10[3,3,3,3,4,5,6,7] ; AVX-NEXT: vpshufhw {{.*#+}} xmm10 = xmm10[0,1,2,3,4,4,4,4] ; AVX-NEXT: vinsertf128 $1, %xmm10, %ymm15, %ymm10 ; AVX-NEXT: vmovaps {{.*#+}} ymm13 = [65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535] ; AVX-NEXT: vandps %ymm13, %ymm11, %ymm11 ; AVX-NEXT: vandnps %ymm10, %ymm13, %ymm10 ; AVX-NEXT: vorps %ymm10, %ymm11, %ymm11 ; AVX-NEXT: vextractf128 $1, %ymm11, %xmm10 ; AVX-NEXT: vmovdqa {{.*#+}} xmm12 = [2,u,128,128,128,128,3,u,128,128,128,128,4,u,128,128] ; AVX-NEXT: vpshufb %xmm12, %xmm3, %xmm15 ; AVX-NEXT: vpblendw {{.*#+}} xmm10 = xmm15[0],xmm10[1,2],xmm15[3],xmm10[4,5],xmm15[6],xmm10[7] ; AVX-NEXT: vpshufb %xmm4, %xmm10, %xmm10 ; AVX-NEXT: vmovdqa {{.*#+}} xmm14 = [128,2,128,128,128,128,128,3,128,128,128,128,128,4,128,128] ; AVX-NEXT: vpshufb %xmm14, %xmm2, %xmm15 ; AVX-NEXT: vpor %xmm15, %xmm10, %xmm4 ; AVX-NEXT: vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX-NEXT: vmovdqa {{.*#+}} xmm15 = [128,128,128,128,0,u,128,128,128,128,1,u,128,128,128,128] ; AVX-NEXT: vpshufb %xmm15, %xmm3, %xmm3 ; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm11[0,1],xmm3[2],xmm11[3,4],xmm3[5],xmm11[6,7] ; AVX-NEXT: vmovdqa {{.*#+}} xmm11 = [0,1,2,3,4,128,6,7,8,9,10,128,12,13,14,15] ; AVX-NEXT: vpshufb %xmm11, %xmm3, %xmm3 ; AVX-NEXT: vmovdqa {{.*#+}} xmm10 = [128,128,128,128,128,0,128,128,128,128,128,1,128,128,128,128] ; AVX-NEXT: vpshufb %xmm10, %xmm2, %xmm2 ; AVX-NEXT: vpor %xmm2, %xmm3, %xmm2 ; AVX-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] ; AVX-NEXT: vpshufd {{.*#+}} xmm3 = xmm7[0,0,1,1] ; AVX-NEXT: vpshufd {{.*#+}} xmm8 = xmm7[1,1,2,2] ; AVX-NEXT: vinsertf128 $1, %xmm8, %ymm3, %ymm3 ; AVX-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] ; AVX-NEXT: vpshuflw {{.*#+}} xmm6 = xmm4[1,0,2,2,4,5,6,7] ; AVX-NEXT: vpshufd {{.*#+}} xmm6 = xmm6[0,1,0,1] ; AVX-NEXT: vpshuflw {{.*#+}} xmm8 = xmm4[3,3,3,3,4,5,6,7] ; AVX-NEXT: vpshufhw {{.*#+}} xmm8 = xmm8[0,1,2,3,4,4,4,4] ; AVX-NEXT: vinsertf128 $1, %xmm8, %ymm6, %ymm6 ; AVX-NEXT: vandps %ymm3, %ymm13, %ymm3 ; AVX-NEXT: vandnps %ymm6, %ymm13, %ymm6 ; AVX-NEXT: vorps %ymm6, %ymm3, %ymm6 ; AVX-NEXT: vpshufb %xmm12, %xmm1, %xmm3 ; AVX-NEXT: vextractf128 $1, %ymm6, %xmm8 ; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm3[0],xmm8[1,2],xmm3[3],xmm8[4,5],xmm3[6],xmm8[7] ; AVX-NEXT: vpshufb {{.*#+}} xmm3 = xmm3[0],zero,xmm3[2,3,4,5,6],zero,xmm3[8,9,10,11,12],zero,xmm3[14,15] ; AVX-NEXT: vpshufb %xmm14, %xmm0, %xmm8 ; AVX-NEXT: vpor %xmm3, %xmm8, %xmm3 ; AVX-NEXT: vpshufb %xmm15, %xmm1, %xmm8 ; AVX-NEXT: vpblendw {{.*#+}} xmm6 = xmm6[0,1],xmm8[2],xmm6[3,4],xmm8[5],xmm6[6,7] ; AVX-NEXT: vpshufb %xmm11, %xmm6, %xmm6 ; AVX-NEXT: vpshufb %xmm10, %xmm0, %xmm8 ; AVX-NEXT: vpor %xmm6, %xmm8, %xmm6 ; AVX-NEXT: vpshufd {{.*#+}} xmm7 = xmm7[3,3,3,3] ; AVX-NEXT: vpermilps $80, {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Folded Reload ; AVX-NEXT: # xmm8 = mem[0,0,1,1] ; AVX-NEXT: vinsertf128 $1, %xmm8, %ymm7, %ymm7 ; AVX-NEXT: vpshufhw {{.*#+}} xmm4 = xmm4[0,1,2,3,5,6,7,7] ; AVX-NEXT: vpshufd {{.*#+}} xmm4 = xmm4[2,2,2,3] ; AVX-NEXT: vpshuflw {{.*#+}} xmm5 = xmm9[1,0,2,2,4,5,6,7] ; AVX-NEXT: vpshufd {{.*#+}} xmm5 = xmm5[0,1,0,1] ; AVX-NEXT: vinsertf128 $1, %xmm5, %ymm4, %ymm4 ; AVX-NEXT: vmovaps {{.*#+}} ymm8 = [65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535] ; AVX-NEXT: vandnps %ymm7, %ymm8, %ymm5 ; AVX-NEXT: vandps %ymm4, %ymm8, %ymm4 ; AVX-NEXT: vorps %ymm5, %ymm4, %ymm4 ; AVX-NEXT: vpshufb {{.*#+}} xmm5 = zero,zero,zero,zero,xmm1[8,u],zero,zero,zero,zero,xmm1[9,u],zero,zero,zero,zero ; AVX-NEXT: vextractf128 $1, %ymm4, %xmm7 ; AVX-NEXT: vpblendw {{.*#+}} xmm5 = xmm7[0,1],xmm5[2],xmm7[3,4],xmm5[5],xmm7[6,7] ; AVX-NEXT: vpshufb %xmm11, %xmm5, %xmm5 ; AVX-NEXT: vpshufb {{.*#+}} xmm7 = zero,zero,zero,zero,zero,xmm0[8],zero,zero,zero,zero,zero,xmm0[9],zero,zero,zero,zero ; AVX-NEXT: vpor %xmm7, %xmm5, %xmm5 ; AVX-NEXT: vpshufb {{.*#+}} xmm1 = zero,zero,xmm1[5,u],zero,zero,zero,zero,xmm1[6,u],zero,zero,zero,zero,xmm1[7,u] ; AVX-NEXT: vpblendw {{.*#+}} xmm1 = xmm4[0],xmm1[1],xmm4[2,3],xmm1[4],xmm4[5,6],xmm1[7] ; AVX-NEXT: vpshufb {{.*#+}} xmm1 = xmm1[0,1,2],zero,xmm1[4,5,6,7,8],zero,xmm1[10,11,12,13,14],zero ; AVX-NEXT: vpshufb {{.*#+}} xmm0 = zero,zero,zero,xmm0[5],zero,zero,zero,zero,zero,xmm0[6],zero,zero,zero,zero,zero,xmm0[7] ; AVX-NEXT: vpor %xmm0, %xmm1, %xmm0 ; AVX-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX-NEXT: vmovdqa %xmm0, 32(%rax) ; AVX-NEXT: vmovdqa %xmm5, 48(%rax) ; AVX-NEXT: vmovdqa %xmm6, (%rax) ; AVX-NEXT: vmovdqa %xmm3, 16(%rax) ; AVX-NEXT: vmovdqa %xmm2, 96(%rax) ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload ; AVX-NEXT: vmovaps %xmm0, 112(%rax) ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload ; AVX-NEXT: vmovaps %xmm0, 64(%rax) ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload ; AVX-NEXT: vmovaps %xmm0, 80(%rax) ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload ; AVX-NEXT: vmovaps %xmm0, 160(%rax) ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload ; AVX-NEXT: vmovaps %xmm0, 176(%rax) ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload ; AVX-NEXT: vmovaps %xmm0, 128(%rax) ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload ; AVX-NEXT: vmovaps %xmm0, 144(%rax) ; AVX-NEXT: vmovaps (%rsp), %xmm0 # 16-byte Reload ; AVX-NEXT: vmovaps %xmm0, 224(%rax) ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload ; AVX-NEXT: vmovaps %xmm0, 240(%rax) ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload ; AVX-NEXT: vmovaps %xmm0, 192(%rax) ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload ; AVX-NEXT: vmovaps %xmm0, 208(%rax) ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload ; AVX-NEXT: vmovaps %xmm0, 288(%rax) ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload ; AVX-NEXT: vmovaps %xmm0, 304(%rax) ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload ; AVX-NEXT: vmovaps %xmm0, 256(%rax) ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload ; AVX-NEXT: vmovaps %xmm0, 272(%rax) ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload ; AVX-NEXT: vmovaps %xmm0, 352(%rax) ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload ; AVX-NEXT: vmovaps %xmm0, 368(%rax) ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload ; AVX-NEXT: vmovaps %xmm0, 320(%rax) ; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload ; AVX-NEXT: vmovaps %xmm0, 336(%rax) ; AVX-NEXT: addq $200, %rsp ; AVX-NEXT: vzeroupper ; AVX-NEXT: retq ; ; AVX2-LABEL: store_i8_stride6_vf64: ; AVX2: # %bb.0: ; AVX2-NEXT: subq $664, %rsp # imm = 0x298 ; AVX2-NEXT: vmovdqa 32(%rdx), %ymm6 ; AVX2-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-NEXT: vmovdqa 32(%rcx), %ymm7 ; AVX2-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-NEXT: vpbroadcastq {{.*#+}} ymm0 = [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-NEXT: vmovdqa (%rcx), %xmm8 ; AVX2-NEXT: vmovdqa 32(%rcx), %xmm5 ; AVX2-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX2-NEXT: vpshufb %xmm0, %xmm8, %xmm1 ; AVX2-NEXT: vmovdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX2-NEXT: vmovdqa (%rdx), %xmm11 ; AVX2-NEXT: vmovdqa 32(%rdx), %xmm9 ; AVX2-NEXT: vmovdqa %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX2-NEXT: vpshufb %xmm0, %xmm11, %xmm2 ; AVX2-NEXT: vmovdqa %xmm11, (%rsp) # 16-byte Spill ; AVX2-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] ; AVX2-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,0,1] ; AVX2-NEXT: vmovdqa (%rsi), %xmm15 ; AVX2-NEXT: vmovdqa 32(%rsi), %xmm14 ; AVX2-NEXT: vpbroadcastq {{.*#+}} xmm3 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0] ; AVX2-NEXT: vpshufb %xmm3, %xmm15, %xmm2 ; AVX2-NEXT: vmovdqa %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX2-NEXT: vmovdqa (%rdi), %xmm10 ; AVX2-NEXT: vmovdqa 32(%rdi), %xmm13 ; AVX2-NEXT: vpshufb %xmm3, %xmm10, %xmm4 ; AVX2-NEXT: vmovdqa %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX2-NEXT: vpunpckhbw {{.*#+}} xmm2 = 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] ; AVX2-NEXT: vpermq {{.*#+}} ymm4 = ymm2[0,0,0,1] ; AVX2-NEXT: vpmovsxbw {{.*#+}} ymm2 = [65535,0,0,65535,0,0,65535,0,0,65535,0,0,65535,0,0,65535] ; AVX2-NEXT: vpblendvb %ymm2, %ymm1, %ymm4, %ymm1 ; AVX2-NEXT: vpshufb %xmm0, %xmm5, %xmm4 ; AVX2-NEXT: vpshufb %xmm0, %xmm9, %xmm5 ; AVX2-NEXT: vpunpcklbw {{.*#+}} xmm4 = 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] ; AVX2-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,0,1] ; AVX2-NEXT: vpshufb %xmm3, %xmm14, %xmm5 ; AVX2-NEXT: vmovdqa %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX2-NEXT: vpshufb %xmm3, %xmm13, %xmm3 ; AVX2-NEXT: vmovdqa %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX2-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm3[8],xmm5[8],xmm3[9],xmm5[9],xmm3[10],xmm5[10],xmm3[11],xmm5[11],xmm3[12],xmm5[12],xmm3[13],xmm5[13],xmm3[14],xmm5[14],xmm3[15],xmm5[15] ; AVX2-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,0,1] ; AVX2-NEXT: vpblendvb %ymm2, %ymm4, %ymm3, %ymm3 ; AVX2-NEXT: vpshufb %ymm0, %ymm7, %ymm4 ; AVX2-NEXT: vpshufb %ymm0, %ymm6, %ymm5 ; AVX2-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-NEXT: vmovdqa 32(%rdi), %ymm7 ; AVX2-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-NEXT: vmovdqa 32(%rsi), %ymm6 ; AVX2-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-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-NEXT: vpshufb %ymm5, %ymm6, %ymm6 ; AVX2-NEXT: vpshufb %ymm5, %ymm7, %ymm7 ; AVX2-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-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,2,3] ; AVX2-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,2,2,3] ; AVX2-NEXT: vpblendvb %ymm2, %ymm4, %ymm6, %ymm4 ; AVX2-NEXT: vmovdqa (%rdx), %ymm7 ; AVX2-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-NEXT: vmovdqa (%rcx), %ymm6 ; AVX2-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-NEXT: vpshufb %ymm0, %ymm6, %ymm6 ; AVX2-NEXT: vpshufb %ymm0, %ymm7, %ymm0 ; AVX2-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] ; AVX2-NEXT: vmovdqa (%rdi), %ymm7 ; AVX2-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-NEXT: vmovdqa (%rsi), %ymm12 ; AVX2-NEXT: vpshufb %ymm5, %ymm12, %ymm6 ; AVX2-NEXT: vpshufb %ymm5, %ymm7, %ymm5 ; AVX2-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-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3] ; AVX2-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,2,2,3] ; AVX2-NEXT: vpblendvb %ymm2, %ymm0, %ymm5, %ymm0 ; AVX2-NEXT: vmovdqa (%r8), %xmm7 ; AVX2-NEXT: vpmovsxbw {{.*#+}} xmm2 = [6,5,8,7,9,9,9,9] ; AVX2-NEXT: vpshufb %xmm2, %xmm7, %xmm5 ; AVX2-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,0,0,1] ; AVX2-NEXT: vpmovsxbw {{.*#+}} ymm6 = [65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535] ; AVX2-NEXT: vpblendvb %ymm6, %ymm1, %ymm5, %ymm1 ; AVX2-NEXT: vmovdqa 32(%r8), %xmm5 ; AVX2-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX2-NEXT: vpshufb %xmm2, %xmm5, %xmm2 ; AVX2-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,0,0,1] ; AVX2-NEXT: vpblendvb %ymm6, %ymm3, %ymm2, %ymm2 ; AVX2-NEXT: vmovdqa 32(%r8), %ymm5 ; AVX2-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-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-NEXT: # ymm3 = mem[0,1,0,1] ; AVX2-NEXT: vpshufb %ymm3, %ymm5, %ymm5 ; AVX2-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,2,2,3] ; AVX2-NEXT: vpblendvb %ymm6, %ymm4, %ymm5, %ymm4 ; AVX2-NEXT: vmovdqa (%r8), %ymm5 ; AVX2-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-NEXT: vpshufb %ymm3, %ymm5, %ymm3 ; AVX2-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3] ; AVX2-NEXT: vpblendvb %ymm6, %ymm0, %ymm3, %ymm0 ; AVX2-NEXT: vmovdqa (%r9), %xmm5 ; AVX2-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX2-NEXT: vmovdqa {{.*#+}} xmm3 = [u,6,u,5,u,8,u,7,u,9,u,9,u,9,u,9] ; AVX2-NEXT: vpshufb %xmm3, %xmm5, %xmm5 ; AVX2-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,0,0,1] ; AVX2-NEXT: vmovdqa {{.*#+}} ymm6 = [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-NEXT: vpblendvb %ymm6, %ymm1, %ymm5, %ymm1 ; AVX2-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-NEXT: vmovdqa 32(%r9), %xmm9 ; AVX2-NEXT: vpshufb %xmm3, %xmm9, %xmm1 ; AVX2-NEXT: vmovdqa %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX2-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,0,1] ; AVX2-NEXT: vpblendvb %ymm6, %ymm2, %ymm1, %ymm1 ; AVX2-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-NEXT: vmovdqa 32(%r9), %ymm2 ; AVX2-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-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] ; AVX2-NEXT: # ymm1 = mem[0,1,0,1] ; AVX2-NEXT: vpshufb %ymm1, %ymm2, %ymm2 ; AVX2-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3] ; AVX2-NEXT: vpblendvb %ymm6, %ymm4, %ymm2, %ymm2 ; AVX2-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-NEXT: vmovdqa (%r9), %ymm2 ; AVX2-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-NEXT: vpshufb %ymm1, %ymm2, %ymm1 ; AVX2-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3] ; AVX2-NEXT: vpblendvb %ymm6, %ymm0, %ymm1, %ymm0 ; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm10[0],xmm15[0],xmm10[1],xmm15[1],xmm10[2],xmm15[2],xmm10[3],xmm15[3],xmm10[4],xmm15[4],xmm10[5],xmm15[5],xmm10[6],xmm15[6],xmm10[7],xmm15[7] ; AVX2-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[0,3,2,1,4,5,6,7] ; AVX2-NEXT: vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,5,6,5] ; AVX2-NEXT: vpermq {{.*#+}} ymm1 = ymm0[0,0,0,1] ; AVX2-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm11[0],xmm8[0],xmm11[1],xmm8[1],xmm11[2],xmm8[2],xmm11[3],xmm8[3],xmm11[4],xmm8[4],xmm11[5],xmm8[5],xmm11[6],xmm8[6],xmm11[7],xmm8[7] ; AVX2-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[1,0,3,2,4,5,6,7] ; AVX2-NEXT: vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,4,4,4] ; AVX2-NEXT: vpermq {{.*#+}} ymm2 = ymm0[0,0,0,1] ; AVX2-NEXT: vpmovsxbw {{.*#+}} ymm0 = [65535,0,0,65535,0,0,65535,0,0,65535,0,0,65535,0,0,65535] ; AVX2-NEXT: vpblendvb %ymm0, %ymm1, %ymm2, %ymm11 ; AVX2-NEXT: vpunpcklbw {{.*#+}} xmm1 = 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-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[0,3,2,1,4,5,6,7] ; AVX2-NEXT: vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,5,6,5] ; AVX2-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,0,1] ; AVX2-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload ; AVX2-NEXT: vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm2 # 16-byte Folded Reload ; AVX2-NEXT: # 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] ; AVX2-NEXT: vpshuflw {{.*#+}} xmm2 = xmm2[1,0,3,2,4,5,6,7] ; AVX2-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,4,4,4] ; AVX2-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,0,0,1] ; AVX2-NEXT: vpblendvb %ymm0, %ymm1, %ymm2, %ymm6 ; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload ; AVX2-NEXT: vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload ; AVX2-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-NEXT: vpshuflw {{.*#+}} ymm1 = ymm1[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15] ; AVX2-NEXT: vpshufhw {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5,6,5,8,9,10,11,12,13,14,13] ; AVX2-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3] ; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Reload ; AVX2-NEXT: vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm2 # 32-byte Folded Reload ; AVX2-NEXT: # ymm2 = ymm15[0],mem[0],ymm15[1],mem[1],ymm15[2],mem[2],ymm15[3],mem[3],ymm15[4],mem[4],ymm15[5],mem[5],ymm15[6],mem[6],ymm15[7],mem[7],ymm15[16],mem[16],ymm15[17],mem[17],ymm15[18],mem[18],ymm15[19],mem[19],ymm15[20],mem[20],ymm15[21],mem[21],ymm15[22],mem[22],ymm15[23],mem[23] ; AVX2-NEXT: vpshuflw {{.*#+}} ymm2 = ymm2[1,0,3,2,4,5,6,7,9,8,11,10,12,13,14,15] ; AVX2-NEXT: vpshufhw {{.*#+}} ymm2 = ymm2[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12] ; AVX2-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3] ; AVX2-NEXT: vpblendvb %ymm0, %ymm1, %ymm2, %ymm1 ; AVX2-NEXT: vmovdqa %ymm12, %ymm13 ; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Reload ; AVX2-NEXT: vpunpcklbw {{.*#+}} ymm2 = ymm12[0],ymm13[0],ymm12[1],ymm13[1],ymm12[2],ymm13[2],ymm12[3],ymm13[3],ymm12[4],ymm13[4],ymm12[5],ymm13[5],ymm12[6],ymm13[6],ymm12[7],ymm13[7],ymm12[16],ymm13[16],ymm12[17],ymm13[17],ymm12[18],ymm13[18],ymm12[19],ymm13[19],ymm12[20],ymm13[20],ymm12[21],ymm13[21],ymm12[22],ymm13[22],ymm12[23],ymm13[23] ; AVX2-NEXT: vpshuflw {{.*#+}} ymm2 = ymm2[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15] ; AVX2-NEXT: vpshufhw {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5,6,5,8,9,10,11,12,13,14,13] ; AVX2-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3] ; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload ; AVX2-NEXT: vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %ymm3, %ymm3 # 32-byte Folded Reload ; AVX2-NEXT: # ymm3 = ymm3[0],mem[0],ymm3[1],mem[1],ymm3[2],mem[2],ymm3[3],mem[3],ymm3[4],mem[4],ymm3[5],mem[5],ymm3[6],mem[6],ymm3[7],mem[7],ymm3[16],mem[16],ymm3[17],mem[17],ymm3[18],mem[18],ymm3[19],mem[19],ymm3[20],mem[20],ymm3[21],mem[21],ymm3[22],mem[22],ymm3[23],mem[23] ; AVX2-NEXT: vpshuflw {{.*#+}} ymm3 = ymm3[1,0,3,2,4,5,6,7,9,8,11,10,12,13,14,15] ; AVX2-NEXT: vpshufhw {{.*#+}} ymm3 = ymm3[0,1,2,3,4,4,4,4,8,9,10,11,12,12,12,12] ; AVX2-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3] ; AVX2-NEXT: vpblendvb %ymm0, %ymm2, %ymm3, %ymm0 ; AVX2-NEXT: vpmovsxbw {{.*#+}} xmm2 = [2,1,0,3,4,4,4,4] ; AVX2-NEXT: vmovdqa %xmm7, %xmm14 ; AVX2-NEXT: vpshufb %xmm2, %xmm7, %xmm3 ; AVX2-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,0,1] ; AVX2-NEXT: vpmovsxbw {{.*#+}} ymm4 = [65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535] ; AVX2-NEXT: vpblendvb %ymm4, %ymm11, %ymm3, %ymm3 ; AVX2-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload ; AVX2-NEXT: vpshufb %xmm2, %xmm7, %xmm2 ; AVX2-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,0,0,1] ; AVX2-NEXT: vpblendvb %ymm4, %ymm6, %ymm2, %ymm2 ; AVX2-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] ; AVX2-NEXT: # ymm6 = mem[0,1,0,1] ; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload ; AVX2-NEXT: vpshufb %ymm6, %ymm5, %ymm11 ; AVX2-NEXT: vpermq {{.*#+}} ymm11 = ymm11[2,2,2,3] ; AVX2-NEXT: vpblendvb %ymm4, %ymm1, %ymm11, %ymm1 ; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload ; AVX2-NEXT: vpshufb %ymm6, %ymm8, %ymm6 ; AVX2-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,2,2,3] ; AVX2-NEXT: vpblendvb %ymm4, %ymm0, %ymm6, %ymm5 ; AVX2-NEXT: vmovdqa {{.*#+}} xmm4 = [u,2,u,1,u,0,u,3,u,4,u,4,u,4,u,4] ; AVX2-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload ; AVX2-NEXT: vpshufb %xmm4, %xmm11, %xmm6 ; AVX2-NEXT: vpermq {{.*#+}} ymm6 = ymm6[0,0,0,1] ; AVX2-NEXT: vmovdqa {{.*#+}} ymm0 = [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-NEXT: vpblendvb %ymm0, %ymm3, %ymm6, %ymm3 ; AVX2-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-NEXT: vpshufb %xmm4, %xmm9, %xmm3 ; AVX2-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,0,1] ; AVX2-NEXT: vpblendvb %ymm0, %ymm2, %ymm3, %ymm2 ; AVX2-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-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-NEXT: # ymm2 = mem[0,1,0,1] ; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload ; AVX2-NEXT: vpshufb %ymm2, %ymm9, %ymm3 ; AVX2-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3] ; AVX2-NEXT: vpblendvb %ymm0, %ymm1, %ymm3, %ymm1 ; AVX2-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload ; AVX2-NEXT: vpshufb %ymm2, %ymm10, %ymm1 ; AVX2-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3] ; AVX2-NEXT: vpblendvb %ymm0, %ymm5, %ymm1, %ymm0 ; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload ; AVX2-NEXT: vpunpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload ; AVX2-NEXT: # xmm0 = 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-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload ; AVX2-NEXT: vpunpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload ; AVX2-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] ; AVX2-NEXT: vmovdqa {{.*#+}} xmm5 = [8,9,6,7,12,13,10,11,14,15,14,15,14,15,14,15] ; AVX2-NEXT: vpshufb %xmm5, %xmm0, %xmm0 ; AVX2-NEXT: vpermq {{.*#+}} ymm2 = ymm0[0,0,0,1] ; AVX2-NEXT: vmovdqa {{.*#+}} xmm3 = [10,11,8,9,6,7,12,13,14,15,14,15,14,15,14,15] ; AVX2-NEXT: vpshufb %xmm3, %xmm1, %xmm1 ; AVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm1[0,0,0,1] ; AVX2-NEXT: vpmovsxbw {{.*#+}} ymm1 = [0,65535,0,0,65535,0,0,65535,0,0,65535,0,0,65535,0,0] ; AVX2-NEXT: vpblendvb %ymm1, %ymm2, %ymm0, %ymm6 ; AVX2-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload ; AVX2-NEXT: vpunpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload ; AVX2-NEXT: # xmm0 = 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-NEXT: vpshufb %xmm5, %xmm0, %xmm0 ; AVX2-NEXT: vmovdqa (%rsp), %xmm2 # 16-byte Reload ; AVX2-NEXT: vpunpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm5 # 16-byte Folded Reload ; AVX2-NEXT: # xmm5 = 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-NEXT: vpshufb %xmm3, %xmm5, %xmm3 ; AVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,0,0,1] ; AVX2-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,0,1] ; AVX2-NEXT: vpblendvb %ymm1, %ymm0, %ymm3, %ymm4 ; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload ; AVX2-NEXT: vpunpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm3 # 32-byte Folded Reload ; AVX2-NEXT: # ymm3 = 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-NEXT: vpunpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm5 # 32-byte Folded Reload ; AVX2-NEXT: # ymm5 = ymm15[8],mem[8],ymm15[9],mem[9],ymm15[10],mem[10],ymm15[11],mem[11],ymm15[12],mem[12],ymm15[13],mem[13],ymm15[14],mem[14],ymm15[15],mem[15],ymm15[24],mem[24],ymm15[25],mem[25],ymm15[26],mem[26],ymm15[27],mem[27],ymm15[28],mem[28],ymm15[29],mem[29],ymm15[30],mem[30],ymm15[31],mem[31] ; AVX2-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [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-NEXT: # ymm0 = mem[0,1,0,1] ; AVX2-NEXT: vpshufb %ymm0, %ymm3, %ymm3 ; AVX2-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3] ; AVX2-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-NEXT: # ymm2 = mem[0,1,0,1] ; AVX2-NEXT: vpshufb %ymm2, %ymm5, %ymm5 ; AVX2-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,2,2,3] ; AVX2-NEXT: vpblendvb %ymm1, %ymm3, %ymm5, %ymm3 ; AVX2-NEXT: vpunpckhbw {{.*#+}} ymm5 = ymm12[8],ymm13[8],ymm12[9],ymm13[9],ymm12[10],ymm13[10],ymm12[11],ymm13[11],ymm12[12],ymm13[12],ymm12[13],ymm13[13],ymm12[14],ymm13[14],ymm12[15],ymm13[15],ymm12[24],ymm13[24],ymm12[25],ymm13[25],ymm12[26],ymm13[26],ymm12[27],ymm13[27],ymm12[28],ymm13[28],ymm12[29],ymm13[29],ymm12[30],ymm13[30],ymm12[31],ymm13[31] ; AVX2-NEXT: vpshufb %ymm0, %ymm5, %ymm0 ; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload ; AVX2-NEXT: vpunpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm5 # 32-byte Folded Reload ; AVX2-NEXT: # ymm5 = ymm5[8],mem[8],ymm5[9],mem[9],ymm5[10],mem[10],ymm5[11],mem[11],ymm5[12],mem[12],ymm5[13],mem[13],ymm5[14],mem[14],ymm5[15],mem[15],ymm5[24],mem[24],ymm5[25],mem[25],ymm5[26],mem[26],ymm5[27],mem[27],ymm5[28],mem[28],ymm5[29],mem[29],ymm5[30],mem[30],ymm5[31],mem[31] ; AVX2-NEXT: vpshufb %ymm2, %ymm5, %ymm2 ; AVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3] ; AVX2-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3] ; AVX2-NEXT: vpblendvb %ymm1, %ymm0, %ymm2, %ymm0 ; AVX2-NEXT: vpmovsxbw {{.*#+}} xmm1 = [10,13,12,11,14,13,14,15] ; AVX2-NEXT: vpshufb %xmm1, %xmm7, %xmm2 ; AVX2-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,0,0,1] ; AVX2-NEXT: vpmovsxbw {{.*#+}} ymm5 = [0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0] ; AVX2-NEXT: vpblendvb %ymm5, %ymm6, %ymm2, %ymm2 ; AVX2-NEXT: vpshufb %xmm1, %xmm14, %xmm1 ; AVX2-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,0,1] ; AVX2-NEXT: vpblendvb %ymm5, %ymm4, %ymm1, %ymm1 ; AVX2-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] ; AVX2-NEXT: # ymm4 = mem[0,1,0,1] ; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload ; AVX2-NEXT: vpshufb %ymm4, %ymm6, %ymm6 ; AVX2-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,2,2,3] ; AVX2-NEXT: vpblendvb %ymm5, %ymm3, %ymm6, %ymm3 ; AVX2-NEXT: vpshufb %ymm4, %ymm8, %ymm4 ; AVX2-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,2,3] ; AVX2-NEXT: vpblendvb %ymm5, %ymm0, %ymm4, %ymm0 ; AVX2-NEXT: vmovdqa {{.*#+}} xmm4 = [u,10,u,13,u,12,u,11,u,14,u,13,u,14,u,15] ; AVX2-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload ; AVX2-NEXT: vpshufb %xmm4, %xmm5, %xmm5 ; AVX2-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,0,0,1] ; AVX2-NEXT: vmovdqa {{.*#+}} ymm6 = [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-NEXT: vpblendvb %ymm6, %ymm2, %ymm5, %ymm2 ; AVX2-NEXT: vpshufb %xmm4, %xmm11, %xmm4 ; AVX2-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,0,1] ; AVX2-NEXT: vpblendvb %ymm6, %ymm1, %ymm4, %ymm1 ; AVX2-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-NEXT: # ymm4 = mem[0,1,0,1] ; AVX2-NEXT: vpshufb %ymm4, %ymm9, %ymm5 ; AVX2-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,2,2,3] ; AVX2-NEXT: vpblendvb %ymm6, %ymm3, %ymm5, %ymm3 ; AVX2-NEXT: vpshufb %ymm4, %ymm10, %ymm4 ; AVX2-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,2,3] ; AVX2-NEXT: vpblendvb %ymm6, %ymm0, %ymm4, %ymm0 ; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX2-NEXT: vmovdqa %ymm0, 160(%rax) ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload ; AVX2-NEXT: vmovaps %ymm0, 128(%rax) ; AVX2-NEXT: vmovdqa %ymm3, 352(%rax) ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload ; AVX2-NEXT: vmovaps %ymm0, 320(%rax) ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload ; AVX2-NEXT: vmovaps %ymm0, 96(%rax) ; AVX2-NEXT: vmovdqa %ymm1, 64(%rax) ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload ; AVX2-NEXT: vmovaps %ymm0, 224(%rax) ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload ; AVX2-NEXT: vmovaps %ymm0, 288(%rax) ; AVX2-NEXT: vmovdqa %ymm2, 256(%rax) ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload ; AVX2-NEXT: vmovaps %ymm0, 32(%rax) ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload ; AVX2-NEXT: vmovaps %ymm0, 192(%rax) ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload ; AVX2-NEXT: vmovaps %ymm0, (%rax) ; AVX2-NEXT: addq $664, %rsp # imm = 0x298 ; AVX2-NEXT: vzeroupper ; AVX2-NEXT: retq ; ; AVX2-FP-LABEL: store_i8_stride6_vf64: ; AVX2-FP: # %bb.0: ; AVX2-FP-NEXT: subq $664, %rsp # imm = 0x298 ; AVX2-FP-NEXT: vmovdqa 32(%rdx), %ymm6 ; AVX2-FP-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FP-NEXT: vmovdqa 32(%rcx), %ymm7 ; AVX2-FP-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FP-NEXT: vpbroadcastq {{.*#+}} ymm0 = [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-FP-NEXT: vmovdqa (%rcx), %xmm1 ; AVX2-FP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX2-FP-NEXT: vmovdqa 32(%rcx), %xmm14 ; AVX2-FP-NEXT: vpshufb %xmm0, %xmm1, %xmm1 ; AVX2-FP-NEXT: vmovdqa (%rdx), %xmm2 ; AVX2-FP-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX2-FP-NEXT: vmovdqa 32(%rdx), %xmm8 ; AVX2-FP-NEXT: vpshufb %xmm0, %xmm2, %xmm2 ; AVX2-FP-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] ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,0,1] ; AVX2-FP-NEXT: vmovdqa (%rsi), %xmm12 ; AVX2-FP-NEXT: vmovdqa 32(%rsi), %xmm9 ; AVX2-FP-NEXT: vpbroadcastq {{.*#+}} xmm3 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0] ; AVX2-FP-NEXT: vpshufb %xmm3, %xmm12, %xmm2 ; AVX2-FP-NEXT: vmovdqa %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX2-FP-NEXT: vmovdqa (%rdi), %xmm13 ; AVX2-FP-NEXT: vmovdqa 32(%rdi), %xmm11 ; AVX2-FP-NEXT: vpshufb %xmm3, %xmm13, %xmm4 ; AVX2-FP-NEXT: vmovdqa %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX2-FP-NEXT: vpunpckhbw {{.*#+}} xmm2 = 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] ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm4 = ymm2[0,0,0,1] ; AVX2-FP-NEXT: vpmovsxbw {{.*#+}} ymm2 = [65535,0,0,65535,0,0,65535,0,0,65535,0,0,65535,0,0,65535] ; AVX2-FP-NEXT: vpblendvb %ymm2, %ymm1, %ymm4, %ymm1 ; AVX2-FP-NEXT: vpshufb %xmm0, %xmm14, %xmm4 ; AVX2-FP-NEXT: vmovdqa %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX2-FP-NEXT: vpshufb %xmm0, %xmm8, %xmm5 ; AVX2-FP-NEXT: vmovdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX2-FP-NEXT: vpunpcklbw {{.*#+}} xmm4 = 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] ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,0,1] ; AVX2-FP-NEXT: vpshufb %xmm3, %xmm9, %xmm5 ; AVX2-FP-NEXT: vmovdqa %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX2-FP-NEXT: vpshufb %xmm3, %xmm11, %xmm3 ; AVX2-FP-NEXT: vmovdqa %xmm11, (%rsp) # 16-byte Spill ; AVX2-FP-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm3[8],xmm5[8],xmm3[9],xmm5[9],xmm3[10],xmm5[10],xmm3[11],xmm5[11],xmm3[12],xmm5[12],xmm3[13],xmm5[13],xmm3[14],xmm5[14],xmm3[15],xmm5[15] ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,0,1] ; AVX2-FP-NEXT: vpblendvb %ymm2, %ymm4, %ymm3, %ymm3 ; AVX2-FP-NEXT: vpshufb %ymm0, %ymm7, %ymm4 ; AVX2-FP-NEXT: vpshufb %ymm0, %ymm6, %ymm5 ; AVX2-FP-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-FP-NEXT: vmovdqa 32(%rdi), %ymm7 ; AVX2-FP-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FP-NEXT: vmovdqa 32(%rsi), %ymm6 ; AVX2-FP-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FP-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-FP-NEXT: vpshufb %ymm5, %ymm6, %ymm6 ; AVX2-FP-NEXT: vpshufb %ymm5, %ymm7, %ymm7 ; AVX2-FP-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-FP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,2,3] ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,2,2,3] ; AVX2-FP-NEXT: vpblendvb %ymm2, %ymm4, %ymm6, %ymm4 ; AVX2-FP-NEXT: vmovdqa (%rdx), %ymm15 ; AVX2-FP-NEXT: vmovdqa (%rcx), %ymm6 ; AVX2-FP-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FP-NEXT: vpshufb %ymm0, %ymm6, %ymm6 ; AVX2-FP-NEXT: vpshufb %ymm0, %ymm15, %ymm0 ; AVX2-FP-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] ; AVX2-FP-NEXT: vmovdqa (%rdi), %ymm10 ; AVX2-FP-NEXT: vmovdqa (%rsi), %ymm6 ; AVX2-FP-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FP-NEXT: vpshufb %ymm5, %ymm6, %ymm6 ; AVX2-FP-NEXT: vpshufb %ymm5, %ymm10, %ymm5 ; AVX2-FP-NEXT: vmovdqu %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FP-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-FP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3] ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,2,2,3] ; AVX2-FP-NEXT: vpblendvb %ymm2, %ymm0, %ymm5, %ymm0 ; AVX2-FP-NEXT: vmovdqa (%r8), %xmm5 ; AVX2-FP-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX2-FP-NEXT: vpmovsxbw {{.*#+}} xmm2 = [6,5,8,7,9,9,9,9] ; AVX2-FP-NEXT: vpshufb %xmm2, %xmm5, %xmm5 ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,0,0,1] ; AVX2-FP-NEXT: vpmovsxbw {{.*#+}} ymm6 = [65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535] ; AVX2-FP-NEXT: vpblendvb %ymm6, %ymm1, %ymm5, %ymm1 ; AVX2-FP-NEXT: vmovdqa 32(%r8), %xmm7 ; AVX2-FP-NEXT: vpshufb %xmm2, %xmm7, %xmm2 ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,0,0,1] ; AVX2-FP-NEXT: vpblendvb %ymm6, %ymm3, %ymm2, %ymm2 ; AVX2-FP-NEXT: vmovdqa 32(%r8), %ymm5 ; AVX2-FP-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FP-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-FP-NEXT: # ymm3 = mem[0,1,0,1] ; AVX2-FP-NEXT: vpshufb %ymm3, %ymm5, %ymm5 ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,2,2,3] ; AVX2-FP-NEXT: vpblendvb %ymm6, %ymm4, %ymm5, %ymm4 ; AVX2-FP-NEXT: vmovdqa (%r8), %ymm5 ; AVX2-FP-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FP-NEXT: vpshufb %ymm3, %ymm5, %ymm3 ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3] ; AVX2-FP-NEXT: vpblendvb %ymm6, %ymm0, %ymm3, %ymm6 ; AVX2-FP-NEXT: vmovdqa (%r9), %xmm0 ; AVX2-FP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX2-FP-NEXT: vmovdqa {{.*#+}} xmm3 = [u,6,u,5,u,8,u,7,u,9,u,9,u,9,u,9] ; AVX2-FP-NEXT: vpshufb %xmm3, %xmm0, %xmm5 ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,0,0,1] ; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm0 = [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-FP-NEXT: vpblendvb %ymm0, %ymm1, %ymm5, %ymm1 ; AVX2-FP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FP-NEXT: vmovdqa 32(%r9), %xmm1 ; AVX2-FP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX2-FP-NEXT: vpshufb %xmm3, %xmm1, %xmm1 ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,0,1] ; AVX2-FP-NEXT: vpblendvb %ymm0, %ymm2, %ymm1, %ymm1 ; AVX2-FP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FP-NEXT: vmovdqa 32(%r9), %ymm2 ; AVX2-FP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FP-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] ; AVX2-FP-NEXT: # ymm1 = mem[0,1,0,1] ; AVX2-FP-NEXT: vpshufb %ymm1, %ymm2, %ymm2 ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3] ; AVX2-FP-NEXT: vpblendvb %ymm0, %ymm4, %ymm2, %ymm2 ; AVX2-FP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FP-NEXT: vmovdqa (%r9), %ymm2 ; AVX2-FP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FP-NEXT: vpshufb %ymm1, %ymm2, %ymm1 ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3] ; AVX2-FP-NEXT: vpblendvb %ymm0, %ymm6, %ymm1, %ymm0 ; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FP-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-FP-NEXT: vmovdqa {{.*#+}} xmm2 = [8,9,6,7,12,13,10,11,14,15,14,15,14,15,14,15] ; AVX2-FP-NEXT: vpshufb %xmm2, %xmm0, %xmm0 ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,0,0,1] ; AVX2-FP-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm8[8],xmm14[8],xmm8[9],xmm14[9],xmm8[10],xmm14[10],xmm8[11],xmm14[11],xmm8[12],xmm14[12],xmm8[13],xmm14[13],xmm8[14],xmm14[14],xmm8[15],xmm14[15] ; AVX2-FP-NEXT: vmovdqa {{.*#+}} xmm3 = [10,11,8,9,6,7,12,13,14,15,14,15,14,15,14,15] ; AVX2-FP-NEXT: vpshufb %xmm3, %xmm1, %xmm1 ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,0,1] ; AVX2-FP-NEXT: vpmovsxbw {{.*#+}} ymm4 = [0,65535,0,0,65535,0,0,65535,0,0,65535,0,0,65535,0,0] ; AVX2-FP-NEXT: vpblendvb %ymm4, %ymm0, %ymm1, %ymm5 ; AVX2-FP-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm13[8],xmm12[8],xmm13[9],xmm12[9],xmm13[10],xmm12[10],xmm13[11],xmm12[11],xmm13[12],xmm12[12],xmm13[13],xmm12[13],xmm13[14],xmm12[14],xmm13[15],xmm12[15] ; AVX2-FP-NEXT: vpshufb %xmm2, %xmm0, %xmm0 ; AVX2-FP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload ; AVX2-FP-NEXT: vpunpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload ; AVX2-FP-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] ; AVX2-FP-NEXT: vpshufb %xmm3, %xmm1, %xmm1 ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,0,0,1] ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,0,1] ; AVX2-FP-NEXT: vpblendvb %ymm4, %ymm0, %ymm1, %ymm2 ; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload ; AVX2-FP-NEXT: vpunpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload ; AVX2-FP-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-FP-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [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-FP-NEXT: # ymm1 = mem[0,1,0,1] ; AVX2-FP-NEXT: vpshufb %ymm1, %ymm0, %ymm0 ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm3 = ymm0[2,2,2,3] ; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload ; AVX2-FP-NEXT: vpunpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm6 # 32-byte Folded Reload ; AVX2-FP-NEXT: # ymm6 = 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-FP-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [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-FP-NEXT: # ymm0 = mem[0,1,0,1] ; AVX2-FP-NEXT: vpshufb %ymm0, %ymm6, %ymm6 ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,2,2,3] ; AVX2-FP-NEXT: vpblendvb %ymm4, %ymm3, %ymm6, %ymm3 ; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload ; AVX2-FP-NEXT: vpunpckhbw {{.*#+}} ymm6 = ymm10[8],ymm9[8],ymm10[9],ymm9[9],ymm10[10],ymm9[10],ymm10[11],ymm9[11],ymm10[12],ymm9[12],ymm10[13],ymm9[13],ymm10[14],ymm9[14],ymm10[15],ymm9[15],ymm10[24],ymm9[24],ymm10[25],ymm9[25],ymm10[26],ymm9[26],ymm10[27],ymm9[27],ymm10[28],ymm9[28],ymm10[29],ymm9[29],ymm10[30],ymm9[30],ymm10[31],ymm9[31] ; AVX2-FP-NEXT: vpshufb %ymm1, %ymm6, %ymm1 ; AVX2-FP-NEXT: vmovdqa %ymm15, %ymm12 ; AVX2-FP-NEXT: vpunpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm6 # 32-byte Folded Reload ; AVX2-FP-NEXT: # ymm6 = ymm15[8],mem[8],ymm15[9],mem[9],ymm15[10],mem[10],ymm15[11],mem[11],ymm15[12],mem[12],ymm15[13],mem[13],ymm15[14],mem[14],ymm15[15],mem[15],ymm15[24],mem[24],ymm15[25],mem[25],ymm15[26],mem[26],ymm15[27],mem[27],ymm15[28],mem[28],ymm15[29],mem[29],ymm15[30],mem[30],ymm15[31],mem[31] ; AVX2-FP-NEXT: vpshufb %ymm0, %ymm6, %ymm0 ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3] ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3] ; AVX2-FP-NEXT: vpblendvb %ymm4, %ymm1, %ymm0, %ymm0 ; AVX2-FP-NEXT: vpmovsxbw {{.*#+}} xmm1 = [10,13,12,11,14,13,14,15] ; AVX2-FP-NEXT: vmovdqa %xmm7, %xmm15 ; AVX2-FP-NEXT: vpshufb %xmm1, %xmm7, %xmm4 ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,0,1] ; AVX2-FP-NEXT: vpmovsxbw {{.*#+}} ymm6 = [0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0] ; AVX2-FP-NEXT: vpblendvb %ymm6, %ymm5, %ymm4, %ymm4 ; AVX2-FP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload ; AVX2-FP-NEXT: vpshufb %xmm1, %xmm7, %xmm1 ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,0,1] ; AVX2-FP-NEXT: vpblendvb %ymm6, %ymm2, %ymm1, %ymm1 ; AVX2-FP-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-FP-NEXT: # ymm2 = mem[0,1,0,1] ; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload ; AVX2-FP-NEXT: vpshufb %ymm2, %ymm10, %ymm5 ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,2,2,3] ; AVX2-FP-NEXT: vpblendvb %ymm6, %ymm3, %ymm5, %ymm5 ; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload ; AVX2-FP-NEXT: vpshufb %ymm2, %ymm11, %ymm2 ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3] ; AVX2-FP-NEXT: vpblendvb %ymm6, %ymm0, %ymm2, %ymm0 ; AVX2-FP-NEXT: vmovdqa {{.*#+}} xmm2 = [u,10,u,13,u,12,u,11,u,14,u,13,u,14,u,15] ; AVX2-FP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload ; AVX2-FP-NEXT: vpshufb %xmm2, %xmm3, %xmm3 ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,0,1] ; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm6 = [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-FP-NEXT: vpblendvb %ymm6, %ymm4, %ymm3, %ymm3 ; AVX2-FP-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload ; AVX2-FP-NEXT: vpshufb %xmm2, %xmm8, %xmm2 ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,0,0,1] ; AVX2-FP-NEXT: vpblendvb %ymm6, %ymm1, %ymm2, %ymm1 ; AVX2-FP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FP-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-FP-NEXT: # ymm4 = mem[0,1,0,1] ; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload ; AVX2-FP-NEXT: vpshufb %ymm4, %ymm13, %ymm1 ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3] ; AVX2-FP-NEXT: vpblendvb %ymm6, %ymm5, %ymm1, %ymm1 ; AVX2-FP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Reload ; AVX2-FP-NEXT: vpshufb %ymm4, %ymm14, %ymm4 ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,2,3] ; AVX2-FP-NEXT: vpblendvb %ymm6, %ymm0, %ymm4, %ymm0 ; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload ; AVX2-FP-NEXT: vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm4 # 16-byte Folded Reload ; AVX2-FP-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-FP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload ; AVX2-FP-NEXT: vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm5 # 16-byte Folded Reload ; AVX2-FP-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-FP-NEXT: vmovdqa {{.*#+}} xmm6 = [0,1,6,7,4,5,2,3,8,9,10,11,12,13,10,11] ; AVX2-FP-NEXT: vpshufb %xmm6, %xmm4, %xmm4 ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,0,1] ; AVX2-FP-NEXT: vmovdqa {{.*#+}} xmm1 = [2,3,0,1,6,7,4,5,8,9,8,9,8,9,8,9] ; AVX2-FP-NEXT: vpshufb %xmm1, %xmm5, %xmm5 ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,0,0,1] ; AVX2-FP-NEXT: vpmovsxbw {{.*#+}} ymm0 = [65535,0,0,65535,0,0,65535,0,0,65535,0,0,65535,0,0,65535] ; AVX2-FP-NEXT: vpblendvb %ymm0, %ymm4, %ymm5, %ymm4 ; AVX2-FP-NEXT: vmovdqa (%rsp), %xmm2 # 16-byte Reload ; AVX2-FP-NEXT: vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm5 # 16-byte Folded Reload ; AVX2-FP-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-FP-NEXT: vpshufb %xmm6, %xmm5, %xmm5 ; AVX2-FP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload ; AVX2-FP-NEXT: vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm6 # 16-byte Folded Reload ; AVX2-FP-NEXT: # xmm6 = 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-FP-NEXT: vpshufb %xmm1, %xmm6, %xmm1 ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,0,0,1] ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,0,1] ; AVX2-FP-NEXT: vpblendvb %ymm0, %ymm5, %ymm1, %ymm3 ; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload ; AVX2-FP-NEXT: vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm5 # 32-byte Folded Reload ; AVX2-FP-NEXT: # ymm5 = 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-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload ; AVX2-FP-NEXT: vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm6 # 32-byte Folded Reload ; AVX2-FP-NEXT: # ymm6 = 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-FP-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [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-FP-NEXT: # ymm1 = mem[0,1,0,1] ; AVX2-FP-NEXT: vpshufb %ymm1, %ymm5, %ymm5 ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,2,2,3] ; AVX2-FP-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [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-FP-NEXT: # ymm2 = mem[0,1,0,1] ; AVX2-FP-NEXT: vpshufb %ymm2, %ymm6, %ymm6 ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,2,2,3] ; AVX2-FP-NEXT: vpblendvb %ymm0, %ymm5, %ymm6, %ymm5 ; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload ; AVX2-FP-NEXT: vpunpcklbw {{.*#+}} ymm6 = ymm6[0],ymm9[0],ymm6[1],ymm9[1],ymm6[2],ymm9[2],ymm6[3],ymm9[3],ymm6[4],ymm9[4],ymm6[5],ymm9[5],ymm6[6],ymm9[6],ymm6[7],ymm9[7],ymm6[16],ymm9[16],ymm6[17],ymm9[17],ymm6[18],ymm9[18],ymm6[19],ymm9[19],ymm6[20],ymm9[20],ymm6[21],ymm9[21],ymm6[22],ymm9[22],ymm6[23],ymm9[23] ; AVX2-FP-NEXT: vpshufb %ymm1, %ymm6, %ymm1 ; AVX2-FP-NEXT: vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %ymm12, %ymm6 # 32-byte Folded Reload ; AVX2-FP-NEXT: # ymm6 = 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-FP-NEXT: vpshufb %ymm2, %ymm6, %ymm2 ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3] ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3] ; AVX2-FP-NEXT: vpblendvb %ymm0, %ymm1, %ymm2, %ymm0 ; AVX2-FP-NEXT: vpmovsxbw {{.*#+}} xmm1 = [2,1,0,3,4,4,4,4] ; AVX2-FP-NEXT: vpshufb %xmm1, %xmm7, %xmm2 ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,0,0,1] ; AVX2-FP-NEXT: vpmovsxbw {{.*#+}} ymm6 = [65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535] ; AVX2-FP-NEXT: vpblendvb %ymm6, %ymm4, %ymm2, %ymm2 ; AVX2-FP-NEXT: vpshufb %xmm1, %xmm15, %xmm1 ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,0,1] ; AVX2-FP-NEXT: vpblendvb %ymm6, %ymm3, %ymm1, %ymm1 ; AVX2-FP-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [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-FP-NEXT: # ymm3 = mem[0,1,0,1] ; AVX2-FP-NEXT: vpshufb %ymm3, %ymm10, %ymm4 ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,2,3] ; AVX2-FP-NEXT: vpblendvb %ymm6, %ymm5, %ymm4, %ymm4 ; AVX2-FP-NEXT: vpshufb %ymm3, %ymm11, %ymm3 ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3] ; AVX2-FP-NEXT: vpblendvb %ymm6, %ymm0, %ymm3, %ymm0 ; AVX2-FP-NEXT: vmovdqa {{.*#+}} xmm3 = [u,2,u,1,u,0,u,3,u,4,u,4,u,4,u,4] ; AVX2-FP-NEXT: vpshufb %xmm3, %xmm8, %xmm5 ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,0,0,1] ; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm6 = [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-FP-NEXT: vpblendvb %ymm6, %ymm2, %ymm5, %ymm2 ; AVX2-FP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload ; AVX2-FP-NEXT: vpshufb %xmm3, %xmm5, %xmm3 ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,0,1] ; AVX2-FP-NEXT: vpblendvb %ymm6, %ymm1, %ymm3, %ymm1 ; AVX2-FP-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [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-FP-NEXT: # ymm3 = mem[0,1,0,1] ; AVX2-FP-NEXT: vpshufb %ymm3, %ymm13, %ymm5 ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,2,2,3] ; AVX2-FP-NEXT: vpblendvb %ymm6, %ymm4, %ymm5, %ymm4 ; AVX2-FP-NEXT: vpshufb %ymm3, %ymm14, %ymm3 ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3] ; AVX2-FP-NEXT: vpblendvb %ymm6, %ymm0, %ymm3, %ymm0 ; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX2-FP-NEXT: vmovdqa %ymm0, 96(%rax) ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload ; AVX2-FP-NEXT: vmovaps %ymm0, 160(%rax) ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload ; AVX2-FP-NEXT: vmovaps %ymm0, 128(%rax) ; AVX2-FP-NEXT: vmovdqa %ymm4, 288(%rax) ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload ; AVX2-FP-NEXT: vmovaps %ymm0, 352(%rax) ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload ; AVX2-FP-NEXT: vmovaps %ymm0, 320(%rax) ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload ; AVX2-FP-NEXT: vmovaps %ymm0, 64(%rax) ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload ; AVX2-FP-NEXT: vmovaps %ymm0, 224(%rax) ; AVX2-FP-NEXT: vmovdqa %ymm1, 192(%rax) ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload ; AVX2-FP-NEXT: vmovaps %ymm0, 256(%rax) ; AVX2-FP-NEXT: vmovdqa %ymm2, (%rax) ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload ; AVX2-FP-NEXT: vmovaps %ymm0, 32(%rax) ; AVX2-FP-NEXT: addq $664, %rsp # imm = 0x298 ; AVX2-FP-NEXT: vzeroupper ; AVX2-FP-NEXT: retq ; ; AVX2-FCP-LABEL: store_i8_stride6_vf64: ; AVX2-FCP: # %bb.0: ; AVX2-FCP-NEXT: subq $664, %rsp # imm = 0x298 ; AVX2-FCP-NEXT: vmovdqa 32(%rdx), %ymm6 ; AVX2-FCP-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FCP-NEXT: vmovdqa 32(%rcx), %ymm7 ; AVX2-FCP-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FCP-NEXT: vpbroadcastq {{.*#+}} ymm0 = [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-FCP-NEXT: vmovdqa (%rcx), %xmm1 ; AVX2-FCP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX2-FCP-NEXT: vmovdqa 32(%rcx), %xmm14 ; AVX2-FCP-NEXT: vpshufb %xmm0, %xmm1, %xmm1 ; AVX2-FCP-NEXT: vmovdqa (%rdx), %xmm2 ; AVX2-FCP-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX2-FCP-NEXT: vmovdqa 32(%rdx), %xmm8 ; AVX2-FCP-NEXT: vpshufb %xmm0, %xmm2, %xmm2 ; AVX2-FCP-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] ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,0,1] ; AVX2-FCP-NEXT: vmovdqa (%rsi), %xmm12 ; AVX2-FCP-NEXT: vmovdqa 32(%rsi), %xmm9 ; AVX2-FCP-NEXT: vpbroadcastq {{.*#+}} xmm3 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0] ; AVX2-FCP-NEXT: vpshufb %xmm3, %xmm12, %xmm2 ; AVX2-FCP-NEXT: vmovdqa %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX2-FCP-NEXT: vmovdqa (%rdi), %xmm13 ; AVX2-FCP-NEXT: vmovdqa 32(%rdi), %xmm11 ; AVX2-FCP-NEXT: vpshufb %xmm3, %xmm13, %xmm4 ; AVX2-FCP-NEXT: vmovdqa %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX2-FCP-NEXT: vpunpckhbw {{.*#+}} xmm2 = 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] ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm2[0,0,0,1] ; AVX2-FCP-NEXT: vpmovsxbw {{.*#+}} ymm2 = [65535,0,0,65535,0,0,65535,0,0,65535,0,0,65535,0,0,65535] ; AVX2-FCP-NEXT: vpblendvb %ymm2, %ymm1, %ymm4, %ymm1 ; AVX2-FCP-NEXT: vpshufb %xmm0, %xmm14, %xmm4 ; AVX2-FCP-NEXT: vmovdqa %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX2-FCP-NEXT: vpshufb %xmm0, %xmm8, %xmm5 ; AVX2-FCP-NEXT: vmovdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX2-FCP-NEXT: vpunpcklbw {{.*#+}} xmm4 = 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] ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,0,1] ; AVX2-FCP-NEXT: vpshufb %xmm3, %xmm9, %xmm5 ; AVX2-FCP-NEXT: vmovdqa %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX2-FCP-NEXT: vpshufb %xmm3, %xmm11, %xmm3 ; AVX2-FCP-NEXT: vmovdqa %xmm11, (%rsp) # 16-byte Spill ; AVX2-FCP-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm3[8],xmm5[8],xmm3[9],xmm5[9],xmm3[10],xmm5[10],xmm3[11],xmm5[11],xmm3[12],xmm5[12],xmm3[13],xmm5[13],xmm3[14],xmm5[14],xmm3[15],xmm5[15] ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,0,1] ; AVX2-FCP-NEXT: vpblendvb %ymm2, %ymm4, %ymm3, %ymm3 ; AVX2-FCP-NEXT: vpshufb %ymm0, %ymm7, %ymm4 ; AVX2-FCP-NEXT: vpshufb %ymm0, %ymm6, %ymm5 ; AVX2-FCP-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-FCP-NEXT: vmovdqa 32(%rdi), %ymm7 ; AVX2-FCP-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FCP-NEXT: vmovdqa 32(%rsi), %ymm6 ; AVX2-FCP-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FCP-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-FCP-NEXT: vpshufb %ymm5, %ymm6, %ymm6 ; AVX2-FCP-NEXT: vpshufb %ymm5, %ymm7, %ymm7 ; AVX2-FCP-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-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,2,3] ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,2,2,3] ; AVX2-FCP-NEXT: vpblendvb %ymm2, %ymm4, %ymm6, %ymm4 ; AVX2-FCP-NEXT: vmovdqa (%rdx), %ymm15 ; AVX2-FCP-NEXT: vmovdqa (%rcx), %ymm6 ; AVX2-FCP-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FCP-NEXT: vpshufb %ymm0, %ymm6, %ymm6 ; AVX2-FCP-NEXT: vpshufb %ymm0, %ymm15, %ymm0 ; AVX2-FCP-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] ; AVX2-FCP-NEXT: vmovdqa (%rdi), %ymm10 ; AVX2-FCP-NEXT: vmovdqa (%rsi), %ymm6 ; AVX2-FCP-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FCP-NEXT: vpshufb %ymm5, %ymm6, %ymm6 ; AVX2-FCP-NEXT: vpshufb %ymm5, %ymm10, %ymm5 ; AVX2-FCP-NEXT: vmovdqu %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FCP-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-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3] ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,2,2,3] ; AVX2-FCP-NEXT: vpblendvb %ymm2, %ymm0, %ymm5, %ymm0 ; AVX2-FCP-NEXT: vmovdqa (%r8), %xmm5 ; AVX2-FCP-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX2-FCP-NEXT: vpmovsxbw {{.*#+}} xmm2 = [6,5,8,7,9,9,9,9] ; AVX2-FCP-NEXT: vpshufb %xmm2, %xmm5, %xmm5 ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,0,0,1] ; AVX2-FCP-NEXT: vpmovsxbw {{.*#+}} ymm6 = [65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535] ; AVX2-FCP-NEXT: vpblendvb %ymm6, %ymm1, %ymm5, %ymm1 ; AVX2-FCP-NEXT: vmovdqa 32(%r8), %xmm7 ; AVX2-FCP-NEXT: vpshufb %xmm2, %xmm7, %xmm2 ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,0,0,1] ; AVX2-FCP-NEXT: vpblendvb %ymm6, %ymm3, %ymm2, %ymm2 ; AVX2-FCP-NEXT: vmovdqa 32(%r8), %ymm5 ; AVX2-FCP-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FCP-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-FCP-NEXT: # ymm3 = mem[0,1,0,1] ; AVX2-FCP-NEXT: vpshufb %ymm3, %ymm5, %ymm5 ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,2,2,3] ; AVX2-FCP-NEXT: vpblendvb %ymm6, %ymm4, %ymm5, %ymm4 ; AVX2-FCP-NEXT: vmovdqa (%r8), %ymm5 ; AVX2-FCP-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FCP-NEXT: vpshufb %ymm3, %ymm5, %ymm3 ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3] ; AVX2-FCP-NEXT: vpblendvb %ymm6, %ymm0, %ymm3, %ymm6 ; AVX2-FCP-NEXT: vmovdqa (%r9), %xmm0 ; AVX2-FCP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX2-FCP-NEXT: vmovdqa {{.*#+}} xmm3 = [u,6,u,5,u,8,u,7,u,9,u,9,u,9,u,9] ; AVX2-FCP-NEXT: vpshufb %xmm3, %xmm0, %xmm5 ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,0,0,1] ; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm0 = [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-FCP-NEXT: vpblendvb %ymm0, %ymm1, %ymm5, %ymm1 ; AVX2-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FCP-NEXT: vmovdqa 32(%r9), %xmm1 ; AVX2-FCP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX2-FCP-NEXT: vpshufb %xmm3, %xmm1, %xmm1 ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,0,1] ; AVX2-FCP-NEXT: vpblendvb %ymm0, %ymm2, %ymm1, %ymm1 ; AVX2-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FCP-NEXT: vmovdqa 32(%r9), %ymm2 ; AVX2-FCP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FCP-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] ; AVX2-FCP-NEXT: # ymm1 = mem[0,1,0,1] ; AVX2-FCP-NEXT: vpshufb %ymm1, %ymm2, %ymm2 ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3] ; AVX2-FCP-NEXT: vpblendvb %ymm0, %ymm4, %ymm2, %ymm2 ; AVX2-FCP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FCP-NEXT: vmovdqa (%r9), %ymm2 ; AVX2-FCP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FCP-NEXT: vpshufb %ymm1, %ymm2, %ymm1 ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3] ; AVX2-FCP-NEXT: vpblendvb %ymm0, %ymm6, %ymm1, %ymm0 ; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FCP-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-FCP-NEXT: vmovdqa {{.*#+}} xmm2 = [8,9,6,7,12,13,10,11,14,15,14,15,14,15,14,15] ; AVX2-FCP-NEXT: vpshufb %xmm2, %xmm0, %xmm0 ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,0,0,1] ; AVX2-FCP-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm8[8],xmm14[8],xmm8[9],xmm14[9],xmm8[10],xmm14[10],xmm8[11],xmm14[11],xmm8[12],xmm14[12],xmm8[13],xmm14[13],xmm8[14],xmm14[14],xmm8[15],xmm14[15] ; AVX2-FCP-NEXT: vmovdqa {{.*#+}} xmm3 = [10,11,8,9,6,7,12,13,14,15,14,15,14,15,14,15] ; AVX2-FCP-NEXT: vpshufb %xmm3, %xmm1, %xmm1 ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,0,1] ; AVX2-FCP-NEXT: vpmovsxbw {{.*#+}} ymm4 = [0,65535,0,0,65535,0,0,65535,0,0,65535,0,0,65535,0,0] ; AVX2-FCP-NEXT: vpblendvb %ymm4, %ymm0, %ymm1, %ymm5 ; AVX2-FCP-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm13[8],xmm12[8],xmm13[9],xmm12[9],xmm13[10],xmm12[10],xmm13[11],xmm12[11],xmm13[12],xmm12[12],xmm13[13],xmm12[13],xmm13[14],xmm12[14],xmm13[15],xmm12[15] ; AVX2-FCP-NEXT: vpshufb %xmm2, %xmm0, %xmm0 ; AVX2-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload ; AVX2-FCP-NEXT: vpunpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload ; AVX2-FCP-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] ; AVX2-FCP-NEXT: vpshufb %xmm3, %xmm1, %xmm1 ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,0,0,1] ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,0,1] ; AVX2-FCP-NEXT: vpblendvb %ymm4, %ymm0, %ymm1, %ymm2 ; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload ; AVX2-FCP-NEXT: vpunpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload ; AVX2-FCP-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-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [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-FCP-NEXT: # ymm1 = mem[0,1,0,1] ; AVX2-FCP-NEXT: vpshufb %ymm1, %ymm0, %ymm0 ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm0[2,2,2,3] ; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload ; AVX2-FCP-NEXT: vpunpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm6 # 32-byte Folded Reload ; AVX2-FCP-NEXT: # ymm6 = 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-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [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-FCP-NEXT: # ymm0 = mem[0,1,0,1] ; AVX2-FCP-NEXT: vpshufb %ymm0, %ymm6, %ymm6 ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,2,2,3] ; AVX2-FCP-NEXT: vpblendvb %ymm4, %ymm3, %ymm6, %ymm3 ; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Reload ; AVX2-FCP-NEXT: vpunpckhbw {{.*#+}} ymm6 = ymm10[8],ymm9[8],ymm10[9],ymm9[9],ymm10[10],ymm9[10],ymm10[11],ymm9[11],ymm10[12],ymm9[12],ymm10[13],ymm9[13],ymm10[14],ymm9[14],ymm10[15],ymm9[15],ymm10[24],ymm9[24],ymm10[25],ymm9[25],ymm10[26],ymm9[26],ymm10[27],ymm9[27],ymm10[28],ymm9[28],ymm10[29],ymm9[29],ymm10[30],ymm9[30],ymm10[31],ymm9[31] ; AVX2-FCP-NEXT: vpshufb %ymm1, %ymm6, %ymm1 ; AVX2-FCP-NEXT: vmovdqa %ymm15, %ymm12 ; AVX2-FCP-NEXT: vpunpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm6 # 32-byte Folded Reload ; AVX2-FCP-NEXT: # ymm6 = ymm15[8],mem[8],ymm15[9],mem[9],ymm15[10],mem[10],ymm15[11],mem[11],ymm15[12],mem[12],ymm15[13],mem[13],ymm15[14],mem[14],ymm15[15],mem[15],ymm15[24],mem[24],ymm15[25],mem[25],ymm15[26],mem[26],ymm15[27],mem[27],ymm15[28],mem[28],ymm15[29],mem[29],ymm15[30],mem[30],ymm15[31],mem[31] ; AVX2-FCP-NEXT: vpshufb %ymm0, %ymm6, %ymm0 ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3] ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3] ; AVX2-FCP-NEXT: vpblendvb %ymm4, %ymm1, %ymm0, %ymm0 ; AVX2-FCP-NEXT: vpmovsxbw {{.*#+}} xmm1 = [10,13,12,11,14,13,14,15] ; AVX2-FCP-NEXT: vmovdqa %xmm7, %xmm15 ; AVX2-FCP-NEXT: vpshufb %xmm1, %xmm7, %xmm4 ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,0,1] ; AVX2-FCP-NEXT: vpmovsxbw {{.*#+}} ymm6 = [0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0] ; AVX2-FCP-NEXT: vpblendvb %ymm6, %ymm5, %ymm4, %ymm4 ; AVX2-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload ; AVX2-FCP-NEXT: vpshufb %xmm1, %xmm7, %xmm1 ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,0,1] ; AVX2-FCP-NEXT: vpblendvb %ymm6, %ymm2, %ymm1, %ymm1 ; AVX2-FCP-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-FCP-NEXT: # ymm2 = mem[0,1,0,1] ; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload ; AVX2-FCP-NEXT: vpshufb %ymm2, %ymm10, %ymm5 ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,2,2,3] ; AVX2-FCP-NEXT: vpblendvb %ymm6, %ymm3, %ymm5, %ymm5 ; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload ; AVX2-FCP-NEXT: vpshufb %ymm2, %ymm11, %ymm2 ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3] ; AVX2-FCP-NEXT: vpblendvb %ymm6, %ymm0, %ymm2, %ymm0 ; AVX2-FCP-NEXT: vmovdqa {{.*#+}} xmm2 = [u,10,u,13,u,12,u,11,u,14,u,13,u,14,u,15] ; AVX2-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload ; AVX2-FCP-NEXT: vpshufb %xmm2, %xmm3, %xmm3 ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,0,1] ; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm6 = [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-FCP-NEXT: vpblendvb %ymm6, %ymm4, %ymm3, %ymm3 ; AVX2-FCP-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload ; AVX2-FCP-NEXT: vpshufb %xmm2, %xmm8, %xmm2 ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,0,0,1] ; AVX2-FCP-NEXT: vpblendvb %ymm6, %ymm1, %ymm2, %ymm1 ; AVX2-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FCP-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-FCP-NEXT: # ymm4 = mem[0,1,0,1] ; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload ; AVX2-FCP-NEXT: vpshufb %ymm4, %ymm13, %ymm1 ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3] ; AVX2-FCP-NEXT: vpblendvb %ymm6, %ymm5, %ymm1, %ymm1 ; AVX2-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Reload ; AVX2-FCP-NEXT: vpshufb %ymm4, %ymm14, %ymm4 ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,2,3] ; AVX2-FCP-NEXT: vpblendvb %ymm6, %ymm0, %ymm4, %ymm0 ; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload ; AVX2-FCP-NEXT: vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm4 # 16-byte Folded Reload ; AVX2-FCP-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-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload ; AVX2-FCP-NEXT: vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm5 # 16-byte Folded Reload ; AVX2-FCP-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-FCP-NEXT: vmovdqa {{.*#+}} xmm6 = [0,1,6,7,4,5,2,3,8,9,10,11,12,13,10,11] ; AVX2-FCP-NEXT: vpshufb %xmm6, %xmm4, %xmm4 ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,0,0,1] ; AVX2-FCP-NEXT: vmovdqa {{.*#+}} xmm1 = [2,3,0,1,6,7,4,5,8,9,8,9,8,9,8,9] ; AVX2-FCP-NEXT: vpshufb %xmm1, %xmm5, %xmm5 ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,0,0,1] ; AVX2-FCP-NEXT: vpmovsxbw {{.*#+}} ymm0 = [65535,0,0,65535,0,0,65535,0,0,65535,0,0,65535,0,0,65535] ; AVX2-FCP-NEXT: vpblendvb %ymm0, %ymm4, %ymm5, %ymm4 ; AVX2-FCP-NEXT: vmovdqa (%rsp), %xmm2 # 16-byte Reload ; AVX2-FCP-NEXT: vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm5 # 16-byte Folded Reload ; AVX2-FCP-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-FCP-NEXT: vpshufb %xmm6, %xmm5, %xmm5 ; AVX2-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload ; AVX2-FCP-NEXT: vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm6 # 16-byte Folded Reload ; AVX2-FCP-NEXT: # xmm6 = 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-FCP-NEXT: vpshufb %xmm1, %xmm6, %xmm1 ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,0,0,1] ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,0,1] ; AVX2-FCP-NEXT: vpblendvb %ymm0, %ymm5, %ymm1, %ymm3 ; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload ; AVX2-FCP-NEXT: vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm5 # 32-byte Folded Reload ; AVX2-FCP-NEXT: # ymm5 = 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-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload ; AVX2-FCP-NEXT: vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm6 # 32-byte Folded Reload ; AVX2-FCP-NEXT: # ymm6 = 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-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [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-FCP-NEXT: # ymm1 = mem[0,1,0,1] ; AVX2-FCP-NEXT: vpshufb %ymm1, %ymm5, %ymm5 ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,2,2,3] ; AVX2-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [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-FCP-NEXT: # ymm2 = mem[0,1,0,1] ; AVX2-FCP-NEXT: vpshufb %ymm2, %ymm6, %ymm6 ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,2,2,3] ; AVX2-FCP-NEXT: vpblendvb %ymm0, %ymm5, %ymm6, %ymm5 ; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload ; AVX2-FCP-NEXT: vpunpcklbw {{.*#+}} ymm6 = ymm6[0],ymm9[0],ymm6[1],ymm9[1],ymm6[2],ymm9[2],ymm6[3],ymm9[3],ymm6[4],ymm9[4],ymm6[5],ymm9[5],ymm6[6],ymm9[6],ymm6[7],ymm9[7],ymm6[16],ymm9[16],ymm6[17],ymm9[17],ymm6[18],ymm9[18],ymm6[19],ymm9[19],ymm6[20],ymm9[20],ymm6[21],ymm9[21],ymm6[22],ymm9[22],ymm6[23],ymm9[23] ; AVX2-FCP-NEXT: vpshufb %ymm1, %ymm6, %ymm1 ; AVX2-FCP-NEXT: vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %ymm12, %ymm6 # 32-byte Folded Reload ; AVX2-FCP-NEXT: # ymm6 = 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-FCP-NEXT: vpshufb %ymm2, %ymm6, %ymm2 ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3] ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[2,2,2,3] ; AVX2-FCP-NEXT: vpblendvb %ymm0, %ymm1, %ymm2, %ymm0 ; AVX2-FCP-NEXT: vpmovsxbw {{.*#+}} xmm1 = [2,1,0,3,4,4,4,4] ; AVX2-FCP-NEXT: vpshufb %xmm1, %xmm7, %xmm2 ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,0,0,1] ; AVX2-FCP-NEXT: vpmovsxbw {{.*#+}} ymm6 = [65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535] ; AVX2-FCP-NEXT: vpblendvb %ymm6, %ymm4, %ymm2, %ymm2 ; AVX2-FCP-NEXT: vpshufb %xmm1, %xmm15, %xmm1 ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,0,0,1] ; AVX2-FCP-NEXT: vpblendvb %ymm6, %ymm3, %ymm1, %ymm1 ; AVX2-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [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-FCP-NEXT: # ymm3 = mem[0,1,0,1] ; AVX2-FCP-NEXT: vpshufb %ymm3, %ymm10, %ymm4 ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,2,3] ; AVX2-FCP-NEXT: vpblendvb %ymm6, %ymm5, %ymm4, %ymm4 ; AVX2-FCP-NEXT: vpshufb %ymm3, %ymm11, %ymm3 ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3] ; AVX2-FCP-NEXT: vpblendvb %ymm6, %ymm0, %ymm3, %ymm0 ; AVX2-FCP-NEXT: vmovdqa {{.*#+}} xmm3 = [u,2,u,1,u,0,u,3,u,4,u,4,u,4,u,4] ; AVX2-FCP-NEXT: vpshufb %xmm3, %xmm8, %xmm5 ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,0,0,1] ; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm6 = [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-FCP-NEXT: vpblendvb %ymm6, %ymm2, %ymm5, %ymm2 ; AVX2-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload ; AVX2-FCP-NEXT: vpshufb %xmm3, %xmm5, %xmm3 ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,0,0,1] ; AVX2-FCP-NEXT: vpblendvb %ymm6, %ymm1, %ymm3, %ymm1 ; AVX2-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [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-FCP-NEXT: # ymm3 = mem[0,1,0,1] ; AVX2-FCP-NEXT: vpshufb %ymm3, %ymm13, %ymm5 ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,2,2,3] ; AVX2-FCP-NEXT: vpblendvb %ymm6, %ymm4, %ymm5, %ymm4 ; AVX2-FCP-NEXT: vpshufb %ymm3, %ymm14, %ymm3 ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3] ; AVX2-FCP-NEXT: vpblendvb %ymm6, %ymm0, %ymm3, %ymm0 ; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX2-FCP-NEXT: vmovdqa %ymm0, 96(%rax) ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload ; AVX2-FCP-NEXT: vmovaps %ymm0, 160(%rax) ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload ; AVX2-FCP-NEXT: vmovaps %ymm0, 128(%rax) ; AVX2-FCP-NEXT: vmovdqa %ymm4, 288(%rax) ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload ; AVX2-FCP-NEXT: vmovaps %ymm0, 352(%rax) ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload ; AVX2-FCP-NEXT: vmovaps %ymm0, 320(%rax) ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload ; AVX2-FCP-NEXT: vmovaps %ymm0, 64(%rax) ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload ; AVX2-FCP-NEXT: vmovaps %ymm0, 224(%rax) ; AVX2-FCP-NEXT: vmovdqa %ymm1, 192(%rax) ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload ; AVX2-FCP-NEXT: vmovaps %ymm0, 256(%rax) ; AVX2-FCP-NEXT: vmovdqa %ymm2, (%rax) ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload ; AVX2-FCP-NEXT: vmovaps %ymm0, 32(%rax) ; AVX2-FCP-NEXT: addq $664, %rsp # imm = 0x298 ; AVX2-FCP-NEXT: vzeroupper ; AVX2-FCP-NEXT: retq ; ; AVX512-LABEL: store_i8_stride6_vf64: ; AVX512: # %bb.0: ; AVX512-NEXT: subq $456, %rsp # imm = 0x1C8 ; AVX512-NEXT: vmovdqa (%rsi), %xmm6 ; AVX512-NEXT: vmovdqa 32(%rsi), %xmm3 ; AVX512-NEXT: vmovdqa (%rdi), %xmm7 ; AVX512-NEXT: vmovdqa 32(%rdi), %xmm2 ; AVX512-NEXT: vmovdqa (%r8), %xmm10 ; AVX512-NEXT: vmovdqa 32(%r8), %xmm4 ; AVX512-NEXT: vmovdqa (%r9), %xmm12 ; AVX512-NEXT: vmovdqa 32(%r9), %xmm8 ; AVX512-NEXT: vmovdqa 32(%rcx), %ymm13 ; AVX512-NEXT: vpbroadcastq {{.*#+}} xmm9 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0] ; AVX512-NEXT: vpshufb %xmm9, %xmm3, %xmm0 ; AVX512-NEXT: vpshufb %xmm9, %xmm2, %xmm1 ; AVX512-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] ; AVX512-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] ; AVX512-NEXT: vmovdqa64 %xmm3, %xmm30 ; AVX512-NEXT: vmovdqa64 %xmm2, %xmm31 ; AVX512-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[0,3,2,1,4,5,6,7] ; AVX512-NEXT: vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,5,6,5] ; AVX512-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill ; AVX512-NEXT: vmovdqa {{.*#+}} xmm0 = [6,u,5,u,8,u,7,u,9,u,9,u,9,u,9,u] ; AVX512-NEXT: vpshufb %xmm0, %xmm4, %xmm2 ; AVX512-NEXT: vmovdqa {{.*#+}} xmm1 = [2,u,1,u,0,u,3,u,4,u,4,u,4,u,4,u] ; AVX512-NEXT: vpshufb %xmm1, %xmm4, %xmm3 ; AVX512-NEXT: vmovdqa64 %xmm4, %xmm29 ; AVX512-NEXT: vinserti32x4 $2, %xmm2, %zmm3, %zmm2 ; AVX512-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill ; AVX512-NEXT: vmovdqa {{.*#+}} xmm2 = [u,6,u,5,u,8,u,7,u,9,u,9,u,9,u,9] ; AVX512-NEXT: vpshufb %xmm2, %xmm8, %xmm4 ; AVX512-NEXT: vmovdqa {{.*#+}} xmm3 = [u,2,u,1,u,0,u,3,u,4,u,4,u,4,u,4] ; AVX512-NEXT: vpshufb %xmm3, %xmm8, %xmm5 ; AVX512-NEXT: vmovdqa64 %xmm8, %xmm27 ; AVX512-NEXT: vinserti32x4 $2, %xmm4, %zmm5, %zmm4 ; AVX512-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill ; AVX512-NEXT: vmovdqa 32(%rdx), %ymm8 ; AVX512-NEXT: vpshufb %xmm9, %xmm6, %xmm4 ; AVX512-NEXT: vpshufb %xmm9, %xmm7, %xmm5 ; AVX512-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] ; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm5 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3],xmm7[4],xmm6[4],xmm7[5],xmm6[5],xmm7[6],xmm6[6],xmm7[7],xmm6[7] ; AVX512-NEXT: vmovdqa64 %xmm7, %xmm18 ; AVX512-NEXT: vmovdqa64 %xmm6, %xmm20 ; AVX512-NEXT: vpshuflw {{.*#+}} xmm5 = xmm5[0,3,2,1,4,5,6,7] ; AVX512-NEXT: vpshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,4,5,6,5] ; AVX512-NEXT: vinserti64x4 $1, %ymm4, %zmm5, %zmm4 ; AVX512-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill ; AVX512-NEXT: vmovdqa 32(%rsi), %ymm4 ; AVX512-NEXT: vmovdqa 32(%rdi), %ymm11 ; AVX512-NEXT: vpshufb %xmm0, %xmm10, %xmm0 ; AVX512-NEXT: vpshufb %xmm1, %xmm10, %xmm1 ; AVX512-NEXT: vmovdqa64 %xmm10, %xmm22 ; AVX512-NEXT: vinserti32x4 $2, %xmm0, %zmm1, %zmm0 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill ; AVX512-NEXT: vpshufb %xmm2, %xmm12, %xmm0 ; AVX512-NEXT: vpshufb %xmm3, %xmm12, %xmm1 ; AVX512-NEXT: vmovdqa64 %xmm12, %xmm21 ; AVX512-NEXT: vinserti32x4 $2, %xmm0, %zmm1, %zmm0 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill ; AVX512-NEXT: vpbroadcastq {{.*#+}} ymm15 = [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] ; AVX512-NEXT: vpshufb %ymm15, %ymm4, %ymm0 ; AVX512-NEXT: vpshufb %ymm15, %ymm11, %ymm1 ; AVX512-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] ; AVX512-NEXT: vpunpckhbw {{.*#+}} ymm1 = ymm11[8],ymm4[8],ymm11[9],ymm4[9],ymm11[10],ymm4[10],ymm11[11],ymm4[11],ymm11[12],ymm4[12],ymm11[13],ymm4[13],ymm11[14],ymm4[14],ymm11[15],ymm4[15],ymm11[24],ymm4[24],ymm11[25],ymm4[25],ymm11[26],ymm4[26],ymm11[27],ymm4[27],ymm11[28],ymm4[28],ymm11[29],ymm4[29],ymm11[30],ymm4[30],ymm11[31],ymm4[31] ; AVX512-NEXT: vmovdqa64 %ymm4, %ymm19 ; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm6 = [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] ; AVX512-NEXT: # ymm6 = mem[0,1,0,1] ; AVX512-NEXT: vpshufb %ymm6, %ymm1, %ymm1 ; AVX512-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0 ; AVX512-NEXT: vmovdqu64 %zmm0, (%rsp) # 64-byte Spill ; AVX512-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] ; AVX512-NEXT: vpshufb %ymm12, %ymm13, %ymm0 ; AVX512-NEXT: vpshufb %ymm12, %ymm8, %ymm1 ; AVX512-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] ; AVX512-NEXT: vpunpckhbw {{.*#+}} ymm1 = ymm8[8],ymm13[8],ymm8[9],ymm13[9],ymm8[10],ymm13[10],ymm8[11],ymm13[11],ymm8[12],ymm13[12],ymm8[13],ymm13[13],ymm8[14],ymm13[14],ymm8[15],ymm13[15],ymm8[24],ymm13[24],ymm8[25],ymm13[25],ymm8[26],ymm13[26],ymm8[27],ymm13[27],ymm8[28],ymm13[28],ymm8[29],ymm13[29],ymm8[30],ymm13[30],ymm8[31],ymm13[31] ; AVX512-NEXT: vmovdqa64 %ymm13, %ymm17 ; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm7 = [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] ; AVX512-NEXT: # ymm7 = mem[0,1,0,1] ; AVX512-NEXT: vpshufb %ymm7, %ymm1, %ymm1 ; AVX512-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill ; AVX512-NEXT: vmovdqa 32(%r8), %ymm13 ; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm5 = [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] ; AVX512-NEXT: # ymm5 = mem[0,1,0,1] ; AVX512-NEXT: vpshufb %ymm5, %ymm13, %ymm0 ; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [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] ; AVX512-NEXT: # ymm4 = mem[0,1,0,1] ; AVX512-NEXT: vpshufb %ymm4, %ymm13, %ymm1 ; AVX512-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0 ; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill ; AVX512-NEXT: vmovdqa 32(%r9), %ymm14 ; AVX512-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] ; AVX512-NEXT: # ymm3 = mem[0,1,0,1] ; AVX512-NEXT: vpshufb %ymm3, %ymm14, %ymm0 ; AVX512-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] ; AVX512-NEXT: # ymm2 = mem[0,1,0,1] ; AVX512-NEXT: vpshufb %ymm2, %ymm14, %ymm1 ; AVX512-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm28 ; AVX512-NEXT: vmovdqa (%rsi), %ymm1 ; AVX512-NEXT: vmovdqa (%rdi), %ymm0 ; AVX512-NEXT: vpshufb %ymm15, %ymm1, %ymm10 ; AVX512-NEXT: vpshufb %ymm15, %ymm0, %ymm15 ; AVX512-NEXT: vpunpcklbw {{.*#+}} ymm10 = ymm15[0],ymm10[0],ymm15[1],ymm10[1],ymm15[2],ymm10[2],ymm15[3],ymm10[3],ymm15[4],ymm10[4],ymm15[5],ymm10[5],ymm15[6],ymm10[6],ymm15[7],ymm10[7],ymm15[16],ymm10[16],ymm15[17],ymm10[17],ymm15[18],ymm10[18],ymm15[19],ymm10[19],ymm15[20],ymm10[20],ymm15[21],ymm10[21],ymm15[22],ymm10[22],ymm15[23],ymm10[23] ; AVX512-NEXT: vpunpckhbw {{.*#+}} ymm15 = 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] ; AVX512-NEXT: vpshufb %ymm6, %ymm15, %ymm6 ; AVX512-NEXT: vinserti64x4 $1, %ymm6, %zmm10, %zmm26 ; AVX512-NEXT: vmovdqa (%rcx), %ymm10 ; AVX512-NEXT: vmovdqa (%rdx), %ymm15 ; AVX512-NEXT: vpshufb %ymm12, %ymm10, %ymm6 ; AVX512-NEXT: vpshufb %ymm12, %ymm15, %ymm9 ; AVX512-NEXT: vpunpcklbw {{.*#+}} ymm6 = ymm9[0],ymm6[0],ymm9[1],ymm6[1],ymm9[2],ymm6[2],ymm9[3],ymm6[3],ymm9[4],ymm6[4],ymm9[5],ymm6[5],ymm9[6],ymm6[6],ymm9[7],ymm6[7],ymm9[16],ymm6[16],ymm9[17],ymm6[17],ymm9[18],ymm6[18],ymm9[19],ymm6[19],ymm9[20],ymm6[20],ymm9[21],ymm6[21],ymm9[22],ymm6[22],ymm9[23],ymm6[23] ; AVX512-NEXT: vpunpckhbw {{.*#+}} ymm9 = ymm15[8],ymm10[8],ymm15[9],ymm10[9],ymm15[10],ymm10[10],ymm15[11],ymm10[11],ymm15[12],ymm10[12],ymm15[13],ymm10[13],ymm15[14],ymm10[14],ymm15[15],ymm10[15],ymm15[24],ymm10[24],ymm15[25],ymm10[25],ymm15[26],ymm10[26],ymm15[27],ymm10[27],ymm15[28],ymm10[28],ymm15[29],ymm10[29],ymm15[30],ymm10[30],ymm15[31],ymm10[31] ; AVX512-NEXT: vpshufb %ymm7, %ymm9, %ymm9 ; AVX512-NEXT: vinserti64x4 $1, %ymm9, %zmm6, %zmm24 ; AVX512-NEXT: vmovdqa (%r8), %ymm6 ; AVX512-NEXT: vpshufb %ymm5, %ymm6, %ymm5 ; AVX512-NEXT: vpshufb %ymm4, %ymm6, %ymm4 ; AVX512-NEXT: vinserti64x4 $1, %ymm5, %zmm4, %zmm23 ; AVX512-NEXT: vmovdqa (%r9), %ymm4 ; AVX512-NEXT: vpshufb %ymm3, %ymm4, %ymm3 ; AVX512-NEXT: vpshufb %ymm2, %ymm4, %ymm2 ; AVX512-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm25 ; AVX512-NEXT: vpunpcklbw {{.*#+}} ymm2 = ymm15[0],ymm10[0],ymm15[1],ymm10[1],ymm15[2],ymm10[2],ymm15[3],ymm10[3],ymm15[4],ymm10[4],ymm15[5],ymm10[5],ymm15[6],ymm10[6],ymm15[7],ymm10[7],ymm15[16],ymm10[16],ymm15[17],ymm10[17],ymm15[18],ymm10[18],ymm15[19],ymm10[19],ymm15[20],ymm10[20],ymm15[21],ymm10[21],ymm15[22],ymm10[22],ymm15[23],ymm10[23] ; AVX512-NEXT: vmovdqa64 %ymm2, %ymm16 ; AVX512-NEXT: vmovdqa64 %xmm18, %xmm2 ; AVX512-NEXT: vmovdqa64 %xmm20, %xmm3 ; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm5 = 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] ; AVX512-NEXT: vpunpcklbw {{.*#+}} ymm0 = 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] ; AVX512-NEXT: vmovdqa64 %ymm0, %ymm18 ; AVX512-NEXT: vmovdqa (%rcx), %xmm15 ; AVX512-NEXT: vmovdqa (%rdx), %xmm2 ; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm2[8],xmm15[8],xmm2[9],xmm15[9],xmm2[10],xmm15[10],xmm2[11],xmm15[11],xmm2[12],xmm15[12],xmm2[13],xmm15[13],xmm2[14],xmm15[14],xmm2[15],xmm15[15] ; AVX512-NEXT: vmovdqa {{.*#+}} xmm3 = [10,11,8,9,6,7,12,13,14,15,14,15,14,15,14,15] ; AVX512-NEXT: vpshufb %xmm3, %xmm1, %xmm10 ; AVX512-NEXT: vmovdqa 32(%rcx), %xmm1 ; AVX512-NEXT: vmovdqa 32(%rdx), %xmm0 ; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm9 = xmm0[8],xmm1[8],xmm0[9],xmm1[9],xmm0[10],xmm1[10],xmm0[11],xmm1[11],xmm0[12],xmm1[12],xmm0[13],xmm1[13],xmm0[14],xmm1[14],xmm0[15],xmm1[15] ; AVX512-NEXT: vpshufb %xmm3, %xmm9, %xmm3 ; AVX512-NEXT: vmovdqa64 %ymm3, %ymm20 ; AVX512-NEXT: vmovdqa64 %ymm17, %ymm3 ; AVX512-NEXT: vpunpcklbw {{.*#+}} ymm3 = ymm8[0],ymm3[0],ymm8[1],ymm3[1],ymm8[2],ymm3[2],ymm8[3],ymm3[3],ymm8[4],ymm3[4],ymm8[5],ymm3[5],ymm8[6],ymm3[6],ymm8[7],ymm3[7],ymm8[16],ymm3[16],ymm8[17],ymm3[17],ymm8[18],ymm3[18],ymm8[19],ymm3[19],ymm8[20],ymm3[20],ymm8[21],ymm3[21],ymm8[22],ymm3[22],ymm8[23],ymm3[23] ; AVX512-NEXT: vmovdqa64 %ymm3, %ymm17 ; AVX512-NEXT: vmovdqa64 %xmm30, %xmm3 ; AVX512-NEXT: vmovdqa64 %xmm31, %xmm7 ; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm7[8],xmm3[8],xmm7[9],xmm3[9],xmm7[10],xmm3[10],xmm7[11],xmm3[11],xmm7[12],xmm3[12],xmm7[13],xmm3[13],xmm7[14],xmm3[14],xmm7[15],xmm3[15] ; AVX512-NEXT: vmovdqa {{.*#+}} xmm8 = [8,9,6,7,12,13,10,11,14,15,14,15,14,15,14,15] ; AVX512-NEXT: vpshufb %xmm8, %xmm5, %xmm9 ; AVX512-NEXT: vpshufb %xmm8, %xmm3, %xmm3 ; AVX512-NEXT: vmovdqa64 %ymm3, %ymm31 ; AVX512-NEXT: vmovdqa64 %ymm19, %ymm3 ; AVX512-NEXT: vpunpcklbw {{.*#+}} ymm3 = ymm11[0],ymm3[0],ymm11[1],ymm3[1],ymm11[2],ymm3[2],ymm11[3],ymm3[3],ymm11[4],ymm3[4],ymm11[5],ymm3[5],ymm11[6],ymm3[6],ymm11[7],ymm3[7],ymm11[16],ymm3[16],ymm11[17],ymm3[17],ymm11[18],ymm3[18],ymm11[19],ymm3[19],ymm11[20],ymm3[20],ymm11[21],ymm3[21],ymm11[22],ymm3[22],ymm11[23],ymm3[23] ; AVX512-NEXT: vmovdqa64 %ymm3, %ymm19 ; AVX512-NEXT: vmovdqa {{.*#+}} xmm3 = [10,u,13,u,12,u,11,u,14,u,13,u,14,u,15,u] ; AVX512-NEXT: vmovdqa64 %xmm22, %xmm5 ; AVX512-NEXT: vpshufb %xmm3, %xmm5, %xmm7 ; AVX512-NEXT: vmovdqa64 %xmm29, %xmm5 ; AVX512-NEXT: vpshufb %xmm3, %xmm5, %xmm3 ; AVX512-NEXT: vmovdqa64 %ymm3, %ymm22 ; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm11 = [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] ; AVX512-NEXT: # ymm11 = mem[0,1,0,1] ; AVX512-NEXT: vpshufb %ymm11, %ymm6, %ymm5 ; AVX512-NEXT: vpshufb %ymm11, %ymm13, %ymm6 ; AVX512-NEXT: vmovdqa {{.*#+}} xmm11 = [u,10,u,13,u,12,u,11,u,14,u,13,u,14,u,15] ; AVX512-NEXT: vmovdqa64 %xmm21, %xmm3 ; AVX512-NEXT: vpshufb %xmm11, %xmm3, %xmm13 ; AVX512-NEXT: vmovdqa64 %xmm27, %xmm3 ; AVX512-NEXT: vpshufb %xmm11, %xmm3, %xmm11 ; AVX512-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [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] ; AVX512-NEXT: # ymm3 = mem[0,1,0,1] ; AVX512-NEXT: vpshufb %ymm3, %ymm4, %ymm4 ; AVX512-NEXT: vpshufb %ymm3, %ymm14, %ymm3 ; AVX512-NEXT: vpshufb %xmm12, %xmm1, %xmm14 ; AVX512-NEXT: vpshufb %xmm12, %xmm0, %xmm8 ; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm8 = xmm8[0],xmm14[0],xmm8[1],xmm14[1],xmm8[2],xmm14[2],xmm8[3],xmm14[3],xmm8[4],xmm14[4],xmm8[5],xmm14[5],xmm8[6],xmm14[6],xmm8[7],xmm14[7] ; AVX512-NEXT: vpunpcklbw {{.*#+}} 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] ; AVX512-NEXT: vprold $16, %xmm0, %xmm0 ; AVX512-NEXT: vinserti64x4 $1, %ymm8, %zmm0, %zmm0 ; AVX512-NEXT: vpshufb %xmm12, %xmm15, %xmm1 ; AVX512-NEXT: vpshufb %xmm12, %xmm2, %xmm8 ; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm8[0],xmm1[0],xmm8[1],xmm1[1],xmm8[2],xmm1[2],xmm8[3],xmm1[3],xmm8[4],xmm1[4],xmm8[5],xmm1[5],xmm8[6],xmm1[6],xmm8[7],xmm1[7] ; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm2[0],xmm15[0],xmm2[1],xmm15[1],xmm2[2],xmm15[2],xmm2[3],xmm15[3],xmm2[4],xmm15[4],xmm2[5],xmm15[5],xmm2[6],xmm15[6],xmm2[7],xmm15[7] ; AVX512-NEXT: vprold $16, %xmm2, %xmm2 ; AVX512-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm30 ; AVX512-NEXT: vpermq {{.*#+}} ymm2 = ymm10[0,0,0,1] ; AVX512-NEXT: vprold $16, %ymm16, %ymm8 ; AVX512-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,0,0,1] ; AVX512-NEXT: vmovdqa64 %ymm18, %ymm1 ; AVX512-NEXT: vpshuflw {{.*#+}} ymm10 = ymm1[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15] ; AVX512-NEXT: vpshufhw {{.*#+}} ymm10 = ymm10[0,1,2,3,4,5,6,5,8,9,10,11,12,13,14,13] ; AVX512-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,2,2,3] ; AVX512-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,0,0,1] ; AVX512-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,2,2,3] ; AVX512-NEXT: vpermq {{.*#+}} ymm12 = ymm13[0,0,0,1] ; AVX512-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,2,3] ; AVX512-NEXT: vpermq {{.*#+}} ymm13 = ymm20[0,0,0,1] ; AVX512-NEXT: vprold $16, %ymm17, %ymm14 ; AVX512-NEXT: vpermq {{.*#+}} ymm15 = ymm31[0,0,0,1] ; AVX512-NEXT: vmovdqa64 %ymm19, %ymm1 ; AVX512-NEXT: vpshuflw {{.*#+}} ymm1 = ymm1[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15] ; AVX512-NEXT: vpshufhw {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5,6,5,8,9,10,11,12,13,14,13] ; AVX512-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3] ; AVX512-NEXT: vpermq {{.*#+}} ymm31 = ymm22[0,0,0,1] ; AVX512-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,2,2,3] ; AVX512-NEXT: vpermq {{.*#+}} ymm11 = ymm11[0,0,0,1] ; AVX512-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3] ; AVX512-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,2,3] ; AVX512-NEXT: vinserti64x4 $1, %ymm8, %zmm2, %zmm2 ; AVX512-NEXT: vinserti64x4 $1, %ymm10, %zmm9, %zmm8 ; AVX512-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] ; AVX512-NEXT: vpternlogq {{.*#+}} zmm8 = zmm2 ^ (zmm9 & (zmm8 ^ zmm2)) ; AVX512-NEXT: vinserti64x4 $1, %ymm5, %zmm7, %zmm2 ; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm5 = [0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535] ; AVX512-NEXT: vpternlogq {{.*#+}} zmm2 = zmm2 ^ (zmm5 & (zmm2 ^ zmm8)) ; AVX512-NEXT: vpermq {{.*#+}} ymm7 = ymm14[2,2,2,3] ; AVX512-NEXT: vinserti64x4 $1, %ymm7, %zmm13, %zmm7 ; AVX512-NEXT: vinserti64x4 $1, %ymm1, %zmm15, %zmm1 ; AVX512-NEXT: vpternlogq {{.*#+}} zmm1 = zmm7 ^ (zmm9 & (zmm1 ^ zmm7)) ; AVX512-NEXT: vinserti64x4 $1, %ymm6, %zmm31, %zmm6 ; AVX512-NEXT: vpternlogq {{.*#+}} zmm6 = zmm6 ^ (zmm5 & (zmm6 ^ zmm1)) ; AVX512-NEXT: vinserti64x4 $1, %ymm4, %zmm12, %zmm1 ; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm4 = [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] ; AVX512-NEXT: vpternlogq {{.*#+}} zmm1 = zmm1 ^ (zmm4 & (zmm1 ^ zmm2)) ; AVX512-NEXT: vinserti64x4 $1, %ymm3, %zmm11, %zmm2 ; AVX512-NEXT: vpternlogq {{.*#+}} zmm2 = zmm2 ^ (zmm4 & (zmm2 ^ zmm6)) ; AVX512-NEXT: vpermq {{.*#+}} zmm0 = zmm0[0,0,0,1,4,4,4,5] ; AVX512-NEXT: vpermq $64, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Folded Reload ; AVX512-NEXT: # zmm3 = mem[0,0,0,1,4,4,4,5] ; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm4 = [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] ; AVX512-NEXT: vpternlogq {{.*#+}} zmm3 = zmm0 ^ (zmm4 & (zmm3 ^ zmm0)) ; AVX512-NEXT: vpermq $64, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Folded Reload ; AVX512-NEXT: # zmm0 = mem[0,0,0,1,4,4,4,5] ; AVX512-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (zmm9 & (zmm0 ^ zmm3)) ; AVX512-NEXT: vpermq $64, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Folded Reload ; AVX512-NEXT: # zmm3 = mem[0,0,0,1,4,4,4,5] ; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm5 = [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] ; AVX512-NEXT: vpternlogq {{.*#+}} zmm3 = zmm3 ^ (zmm5 & (zmm3 ^ zmm0)) ; AVX512-NEXT: vpermq {{.*#+}} zmm0 = zmm30[0,0,0,1,4,4,4,5] ; AVX512-NEXT: vpermq $64, {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Folded Reload ; AVX512-NEXT: # zmm6 = mem[0,0,0,1,4,4,4,5] ; AVX512-NEXT: vpternlogq {{.*#+}} zmm6 = zmm0 ^ (zmm4 & (zmm6 ^ zmm0)) ; AVX512-NEXT: vpermq $64, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Folded Reload ; AVX512-NEXT: # zmm0 = mem[0,0,0,1,4,4,4,5] ; AVX512-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (zmm9 & (zmm0 ^ zmm6)) ; AVX512-NEXT: vpermq $64, {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Folded Reload ; AVX512-NEXT: # zmm6 = mem[0,0,0,1,4,4,4,5] ; AVX512-NEXT: vpternlogq {{.*#+}} zmm6 = zmm6 ^ (zmm5 & (zmm6 ^ zmm0)) ; AVX512-NEXT: vpermq $234, (%rsp), %zmm0 # 64-byte Folded Reload ; AVX512-NEXT: # zmm0 = mem[2,2,2,3,6,6,6,7] ; AVX512-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Folded Reload ; AVX512-NEXT: # zmm5 = mem[2,2,2,3,6,6,6,7] ; AVX512-NEXT: vpternlogq {{.*#+}} zmm5 = zmm0 ^ (zmm9 & (zmm5 ^ zmm0)) ; AVX512-NEXT: vpermq {{.*#+}} zmm0 = zmm26[2,2,2,3,6,6,6,7] ; AVX512-NEXT: vpermq {{.*#+}} zmm7 = zmm24[2,2,2,3,6,6,6,7] ; AVX512-NEXT: vpternlogq {{.*#+}} zmm7 = zmm0 ^ (zmm9 & (zmm7 ^ zmm0)) ; AVX512-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Folded Reload ; AVX512-NEXT: # zmm0 = mem[2,2,2,3,6,6,6,7] ; AVX512-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (zmm4 & (zmm0 ^ zmm5)) ; AVX512-NEXT: vpermq {{.*#+}} zmm5 = zmm23[2,2,2,3,6,6,6,7] ; AVX512-NEXT: vpternlogq {{.*#+}} zmm5 = zmm5 ^ (zmm4 & (zmm5 ^ zmm7)) ; AVX512-NEXT: vpermq {{.*#+}} zmm4 = zmm28[2,2,2,3,6,6,6,7] ; AVX512-NEXT: vmovdqa64 {{.*#+}} zmm7 = [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] ; AVX512-NEXT: vpternlogq {{.*#+}} zmm4 = zmm4 ^ (zmm7 & (zmm4 ^ zmm0)) ; AVX512-NEXT: vpermq {{.*#+}} zmm0 = zmm25[2,2,2,3,6,6,6,7] ; AVX512-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (zmm7 & (zmm0 ^ zmm5)) ; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512-NEXT: vmovdqa64 %zmm0, 128(%rax) ; AVX512-NEXT: vmovdqa64 %zmm4, 320(%rax) ; AVX512-NEXT: vmovdqa64 %zmm6, (%rax) ; AVX512-NEXT: vmovdqa64 %zmm3, 192(%rax) ; AVX512-NEXT: vmovdqa64 %zmm2, 256(%rax) ; AVX512-NEXT: vmovdqa64 %zmm1, 64(%rax) ; AVX512-NEXT: addq $456, %rsp # imm = 0x1C8 ; AVX512-NEXT: vzeroupper ; AVX512-NEXT: retq ; ; AVX512-FCP-LABEL: store_i8_stride6_vf64: ; AVX512-FCP: # %bb.0: ; AVX512-FCP-NEXT: subq $280, %rsp # imm = 0x118 ; AVX512-FCP-NEXT: vmovdqa (%rcx), %ymm1 ; AVX512-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX512-FCP-NEXT: vmovdqa (%rdx), %ymm0 ; AVX512-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} ymm0 = 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] ; AVX512-FCP-NEXT: vprold $16, %ymm0, %ymm18 ; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm23 = [0,0,0,1,10,10,10,11] ; AVX512-FCP-NEXT: vmovdqa (%rsi), %ymm6 ; AVX512-FCP-NEXT: vmovdqa (%rdi), %ymm7 ; AVX512-FCP-NEXT: vmovdqa (%r8), %ymm11 ; AVX512-FCP-NEXT: vmovdqa 32(%r8), %xmm8 ; AVX512-FCP-NEXT: vmovdqa (%r9), %ymm10 ; AVX512-FCP-NEXT: vmovdqa 32(%r9), %xmm9 ; AVX512-FCP-NEXT: vmovdqa 32(%r8), %ymm4 ; AVX512-FCP-NEXT: vmovdqa 32(%r9), %ymm12 ; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [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] ; AVX512-FCP-NEXT: # ymm0 = mem[0,1,0,1] ; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm4, %ymm2 ; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [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] ; AVX512-FCP-NEXT: # ymm1 = mem[0,1,0,1] ; AVX512-FCP-NEXT: vpshufb %ymm1, %ymm4, %ymm3 ; AVX512-FCP-NEXT: vmovdqa64 %ymm4, %ymm21 ; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm3, %zmm2 ; AVX512-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill ; AVX512-FCP-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] ; AVX512-FCP-NEXT: # ymm3 = mem[0,1,0,1] ; AVX512-FCP-NEXT: vpshufb %ymm3, %ymm12, %ymm2 ; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [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] ; AVX512-FCP-NEXT: # ymm4 = mem[0,1,0,1] ; AVX512-FCP-NEXT: vpshufb %ymm4, %ymm12, %ymm5 ; AVX512-FCP-NEXT: vmovdqa64 %ymm12, %ymm22 ; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm5, %zmm2 ; AVX512-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill ; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm11, %ymm0 ; AVX512-FCP-NEXT: vpshufb %ymm1, %ymm11, %ymm1 ; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm30 ; AVX512-FCP-NEXT: vpshufb %ymm3, %ymm10, %ymm0 ; AVX512-FCP-NEXT: vpshufb %ymm4, %ymm10, %ymm1 ; AVX512-FCP-NEXT: vmovdqa64 %ymm10, %ymm19 ; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill ; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm0 = xmm8[6,u,5,u,8,u,7,u,9,u,9,u,9,u,9,u] ; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm1 = xmm8[2,u,1,u,0,u,3,u,4,u,4,u,4,u,4,u] ; AVX512-FCP-NEXT: vmovdqa64 %xmm8, %xmm16 ; AVX512-FCP-NEXT: vinserti32x4 $2, %xmm0, %zmm1, %zmm0 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill ; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm0 = xmm9[u,6,u,5,u,8,u,7,u,9,u,9,u,9,u,9] ; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm1 = xmm9[u,2,u,1,u,0,u,3,u,4,u,4,u,4,u,4] ; AVX512-FCP-NEXT: vmovdqa64 %xmm9, %xmm17 ; AVX512-FCP-NEXT: vinserti32x4 $2, %xmm0, %zmm1, %zmm0 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill ; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} ymm0 = 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] ; AVX512-FCP-NEXT: vmovdqa64 %ymm7, %ymm29 ; AVX512-FCP-NEXT: vmovdqa64 %ymm6, %ymm28 ; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm15 = [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] ; AVX512-FCP-NEXT: # ymm15 = mem[0,1,0,1] ; AVX512-FCP-NEXT: vpshufb %ymm15, %ymm0, %ymm0 ; AVX512-FCP-NEXT: vmovdqa (%rsi), %xmm2 ; AVX512-FCP-NEXT: vmovdqa (%rdi), %xmm1 ; AVX512-FCP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX512-FCP-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] ; AVX512-FCP-NEXT: vmovdqa64 %xmm2, %xmm31 ; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm13 = [8,9,6,7,12,13,10,11,14,15,14,15,14,15,14,15] ; AVX512-FCP-NEXT: vpshufb %xmm13, %xmm1, %xmm1 ; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm23, %zmm1 ; AVX512-FCP-NEXT: vmovdqa (%rcx), %xmm2 ; AVX512-FCP-NEXT: vmovdqa (%rdx), %xmm3 ; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm0 = 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] ; AVX512-FCP-NEXT: vmovdqa64 %xmm3, %xmm27 ; AVX512-FCP-NEXT: vmovdqa64 %xmm2, %xmm26 ; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm6 = [10,11,8,9,6,7,12,13,14,15,14,15,14,15,14,15] ; AVX512-FCP-NEXT: vpshufb %xmm6, %xmm0, %xmm0 ; AVX512-FCP-NEXT: vpermt2q %zmm18, %zmm23, %zmm0 ; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm18 = [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] ; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm1 = zmm0 ^ (zmm18 & (zmm1 ^ zmm0)) ; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm12 = [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] ; AVX512-FCP-NEXT: # ymm12 = mem[0,1,0,1] ; AVX512-FCP-NEXT: vpshufb %ymm12, %ymm11, %ymm2 ; AVX512-FCP-NEXT: vmovdqa (%r8), %xmm10 ; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm8 = [10,u,13,u,12,u,11,u,14,u,13,u,14,u,15,u] ; AVX512-FCP-NEXT: vpshufb %xmm8, %xmm10, %xmm0 ; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm23, %zmm0 ; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm24 = [0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535] ; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (zmm24 & (zmm0 ^ zmm1)) ; AVX512-FCP-NEXT: vmovdqa 32(%rcx), %ymm7 ; AVX512-FCP-NEXT: vmovdqa 32(%rdx), %ymm5 ; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} ymm4 = ymm5[0],ymm7[0],ymm5[1],ymm7[1],ymm5[2],ymm7[2],ymm5[3],ymm7[3],ymm5[4],ymm7[4],ymm5[5],ymm7[5],ymm5[6],ymm7[6],ymm5[7],ymm7[7],ymm5[16],ymm7[16],ymm5[17],ymm7[17],ymm5[18],ymm7[18],ymm5[19],ymm7[19],ymm5[20],ymm7[20],ymm5[21],ymm7[21],ymm5[22],ymm7[22],ymm5[23],ymm7[23] ; AVX512-FCP-NEXT: vprold $16, %ymm4, %ymm25 ; AVX512-FCP-NEXT: vmovdqa 32(%rcx), %xmm4 ; AVX512-FCP-NEXT: vmovdqa 32(%rdx), %xmm11 ; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm9 = xmm11[8],xmm4[8],xmm11[9],xmm4[9],xmm11[10],xmm4[10],xmm11[11],xmm4[11],xmm11[12],xmm4[12],xmm11[13],xmm4[13],xmm11[14],xmm4[14],xmm11[15],xmm4[15] ; AVX512-FCP-NEXT: vpshufb %xmm6, %xmm9, %xmm9 ; AVX512-FCP-NEXT: vmovdqa 32(%rsi), %ymm2 ; AVX512-FCP-NEXT: vmovdqa 32(%rdi), %ymm1 ; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} ymm6 = ymm1[0],ymm2[0],ymm1[1],ymm2[1],ymm1[2],ymm2[2],ymm1[3],ymm2[3],ymm1[4],ymm2[4],ymm1[5],ymm2[5],ymm1[6],ymm2[6],ymm1[7],ymm2[7],ymm1[16],ymm2[16],ymm1[17],ymm2[17],ymm1[18],ymm2[18],ymm1[19],ymm2[19],ymm1[20],ymm2[20],ymm1[21],ymm2[21],ymm1[22],ymm2[22],ymm1[23],ymm2[23] ; AVX512-FCP-NEXT: vpshufb %ymm15, %ymm6, %ymm3 ; AVX512-FCP-NEXT: vmovdqa 32(%rsi), %xmm6 ; AVX512-FCP-NEXT: vmovdqa 32(%rdi), %xmm15 ; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm14 = xmm15[8],xmm6[8],xmm15[9],xmm6[9],xmm15[10],xmm6[10],xmm15[11],xmm6[11],xmm15[12],xmm6[12],xmm15[13],xmm6[13],xmm15[14],xmm6[14],xmm15[15],xmm6[15] ; AVX512-FCP-NEXT: vmovdqa64 %xmm6, %xmm20 ; AVX512-FCP-NEXT: vpshufb %xmm13, %xmm14, %xmm13 ; AVX512-FCP-NEXT: vpermt2q %zmm3, %zmm23, %zmm13 ; AVX512-FCP-NEXT: vpermt2q %zmm25, %zmm23, %zmm9 ; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm13 = zmm9 ^ (zmm18 & (zmm13 ^ zmm9)) ; AVX512-FCP-NEXT: vmovdqa64 %ymm21, %ymm3 ; AVX512-FCP-NEXT: vpshufb %ymm12, %ymm3, %ymm3 ; AVX512-FCP-NEXT: vmovdqa64 %xmm16, %xmm6 ; AVX512-FCP-NEXT: vpshufb %xmm8, %xmm6, %xmm9 ; AVX512-FCP-NEXT: vpermt2q %zmm3, %zmm23, %zmm9 ; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm9 = zmm9 ^ (zmm24 & (zmm9 ^ zmm13)) ; AVX512-FCP-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] ; AVX512-FCP-NEXT: # ymm8 = mem[0,1,0,1] ; AVX512-FCP-NEXT: vmovdqa64 %ymm19, %ymm3 ; AVX512-FCP-NEXT: vpshufb %ymm8, %ymm3, %ymm12 ; AVX512-FCP-NEXT: vmovdqa (%r9), %xmm13 ; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm14 = [u,10,u,13,u,12,u,11,u,14,u,13,u,14,u,15] ; AVX512-FCP-NEXT: vpshufb %xmm14, %xmm13, %xmm3 ; AVX512-FCP-NEXT: vpermt2q %zmm12, %zmm23, %zmm3 ; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm12 = [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] ; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm3 = zmm3 ^ (zmm12 & (zmm3 ^ zmm0)) ; AVX512-FCP-NEXT: vmovdqa64 %ymm22, %ymm0 ; AVX512-FCP-NEXT: vpshufb %ymm8, %ymm0, %ymm0 ; AVX512-FCP-NEXT: vmovdqa64 %xmm17, %xmm6 ; AVX512-FCP-NEXT: vpshufb %xmm14, %xmm6, %xmm8 ; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm23, %zmm8 ; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm8 = zmm8 ^ (zmm12 & (zmm8 ^ zmm9)) ; AVX512-FCP-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] ; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm2, %ymm9 ; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm1, %ymm12 ; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} ymm9 = ymm12[0],ymm9[0],ymm12[1],ymm9[1],ymm12[2],ymm9[2],ymm12[3],ymm9[3],ymm12[4],ymm9[4],ymm12[5],ymm9[5],ymm12[6],ymm9[6],ymm12[7],ymm9[7],ymm12[16],ymm9[16],ymm12[17],ymm9[17],ymm12[18],ymm9[18],ymm12[19],ymm9[19],ymm12[20],ymm9[20],ymm12[21],ymm9[21],ymm12[22],ymm9[22],ymm12[23],ymm9[23] ; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} ymm1 = ymm1[8],ymm2[8],ymm1[9],ymm2[9],ymm1[10],ymm2[10],ymm1[11],ymm2[11],ymm1[12],ymm2[12],ymm1[13],ymm2[13],ymm1[14],ymm2[14],ymm1[15],ymm2[15],ymm1[24],ymm2[24],ymm1[25],ymm2[25],ymm1[26],ymm2[26],ymm1[27],ymm2[27],ymm1[28],ymm2[28],ymm1[29],ymm2[29],ymm1[30],ymm2[30],ymm1[31],ymm2[31] ; AVX512-FCP-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] ; AVX512-FCP-NEXT: # ymm2 = mem[0,1,0,1] ; AVX512-FCP-NEXT: vpshufb %ymm2, %ymm1, %ymm1 ; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm14 = [2,2,2,3,10,10,10,11] ; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm14, %zmm9 ; AVX512-FCP-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] ; AVX512-FCP-NEXT: vpshufb %ymm12, %ymm7, %ymm1 ; AVX512-FCP-NEXT: vpshufb %ymm12, %ymm5, %ymm6 ; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} ymm1 = ymm6[0],ymm1[0],ymm6[1],ymm1[1],ymm6[2],ymm1[2],ymm6[3],ymm1[3],ymm6[4],ymm1[4],ymm6[5],ymm1[5],ymm6[6],ymm1[6],ymm6[7],ymm1[7],ymm6[16],ymm1[16],ymm6[17],ymm1[17],ymm6[18],ymm1[18],ymm6[19],ymm1[19],ymm6[20],ymm1[20],ymm6[21],ymm1[21],ymm6[22],ymm1[22],ymm6[23],ymm1[23] ; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} ymm5 = ymm5[8],ymm7[8],ymm5[9],ymm7[9],ymm5[10],ymm7[10],ymm5[11],ymm7[11],ymm5[12],ymm7[12],ymm5[13],ymm7[13],ymm5[14],ymm7[14],ymm5[15],ymm7[15],ymm5[24],ymm7[24],ymm5[25],ymm7[25],ymm5[26],ymm7[26],ymm5[27],ymm7[27],ymm5[28],ymm7[28],ymm5[29],ymm7[29],ymm5[30],ymm7[30],ymm5[31],ymm7[31] ; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm6 = [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] ; AVX512-FCP-NEXT: # ymm6 = mem[0,1,0,1] ; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm5, %ymm5 ; AVX512-FCP-NEXT: vpermt2q %zmm5, %zmm14, %zmm1 ; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm1 = zmm9 ^ (zmm18 & (zmm1 ^ zmm9)) ; AVX512-FCP-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Folded Reload ; AVX512-FCP-NEXT: # zmm9 = mem[2,2,2,3,6,6,6,7] ; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm22 = [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] ; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm9 = zmm9 ^ (zmm22 & (zmm9 ^ zmm1)) ; AVX512-FCP-NEXT: vmovdqa64 %ymm28, %ymm5 ; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm5, %ymm1 ; AVX512-FCP-NEXT: vmovdqa64 %ymm29, %ymm7 ; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm7, %ymm0 ; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} ymm0 = 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] ; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} ymm1 = ymm7[8],ymm5[8],ymm7[9],ymm5[9],ymm7[10],ymm5[10],ymm7[11],ymm5[11],ymm7[12],ymm5[12],ymm7[13],ymm5[13],ymm7[14],ymm5[14],ymm7[15],ymm5[15],ymm7[24],ymm5[24],ymm7[25],ymm5[25],ymm7[26],ymm5[26],ymm7[27],ymm5[27],ymm7[28],ymm5[28],ymm7[29],ymm5[29],ymm7[30],ymm5[30],ymm7[31],ymm5[31] ; AVX512-FCP-NEXT: vpshufb %ymm2, %ymm1, %ymm1 ; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm14, %zmm0 ; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload ; AVX512-FCP-NEXT: vpshufb %ymm12, %ymm7, %ymm1 ; AVX512-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload ; AVX512-FCP-NEXT: vpshufb %ymm12, %ymm5, %ymm2 ; AVX512-FCP-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] ; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} ymm2 = ymm5[8],ymm7[8],ymm5[9],ymm7[9],ymm5[10],ymm7[10],ymm5[11],ymm7[11],ymm5[12],ymm7[12],ymm5[13],ymm7[13],ymm5[14],ymm7[14],ymm5[15],ymm7[15],ymm5[24],ymm7[24],ymm5[25],ymm7[25],ymm5[26],ymm7[26],ymm5[27],ymm7[27],ymm5[28],ymm7[28],ymm5[29],ymm7[29],ymm5[30],ymm7[30],ymm5[31],ymm7[31] ; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm2, %ymm2 ; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm14, %zmm1 ; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm1 = zmm0 ^ (zmm18 & (zmm1 ^ zmm0)) ; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm2 = zmm30[2,2,2,3,6,6,6,7] ; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm2 = zmm2 ^ (zmm22 & (zmm2 ^ zmm1)) ; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm0 = xmm10[6,u,5,u,8,u,7,u,9,u,9,u,9,u,9,u] ; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm1 = xmm10[2,u,1,u,0,u,3,u,4,u,4,u,4,u,4,u] ; AVX512-FCP-NEXT: vinserti32x4 $2, %xmm0, %zmm1, %zmm0 ; AVX512-FCP-NEXT: vpshufb %xmm12, %xmm4, %xmm1 ; AVX512-FCP-NEXT: vpshufb %xmm12, %xmm11, %xmm5 ; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm5[0],xmm1[0],xmm5[1],xmm1[1],xmm5[2],xmm1[2],xmm5[3],xmm1[3],xmm5[4],xmm1[4],xmm5[5],xmm1[5],xmm5[6],xmm1[6],xmm5[7],xmm1[7] ; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm4 = xmm11[0],xmm4[0],xmm11[1],xmm4[1],xmm11[2],xmm4[2],xmm11[3],xmm4[3],xmm11[4],xmm4[4],xmm11[5],xmm4[5],xmm11[6],xmm4[6],xmm11[7],xmm4[7] ; AVX512-FCP-NEXT: vprold $16, %xmm4, %xmm4 ; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,0,0,1,8,8,8,9] ; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm5, %zmm4 ; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm1 = xmm13[u,6,u,5,u,8,u,7,u,9,u,9,u,9,u,9] ; AVX512-FCP-NEXT: vpshufb {{.*#+}} xmm6 = xmm13[u,2,u,1,u,0,u,3,u,4,u,4,u,4,u,4] ; AVX512-FCP-NEXT: vpbroadcastq {{.*#+}} xmm7 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0] ; AVX512-FCP-NEXT: vinserti32x4 $2, %xmm1, %zmm6, %zmm1 ; AVX512-FCP-NEXT: vmovdqa64 %xmm20, %xmm11 ; AVX512-FCP-NEXT: vpshufb %xmm7, %xmm11, %xmm6 ; AVX512-FCP-NEXT: vpshufb %xmm7, %xmm15, %xmm10 ; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm6 = xmm10[8],xmm6[8],xmm10[9],xmm6[9],xmm10[10],xmm6[10],xmm10[11],xmm6[11],xmm10[12],xmm6[12],xmm10[13],xmm6[13],xmm10[14],xmm6[14],xmm10[15],xmm6[15] ; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm10 = xmm15[0],xmm11[0],xmm15[1],xmm11[1],xmm15[2],xmm11[2],xmm15[3],xmm11[3],xmm15[4],xmm11[4],xmm15[5],xmm11[5],xmm15[6],xmm11[6],xmm15[7],xmm11[7] ; AVX512-FCP-NEXT: vmovdqa {{.*#+}} xmm11 = [0,1,6,7,4,5,2,3,8,9,10,11,12,13,10,11] ; AVX512-FCP-NEXT: vpshufb %xmm11, %xmm10, %xmm10 ; AVX512-FCP-NEXT: vpermt2q %zmm6, %zmm5, %zmm10 ; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm10 = zmm4 ^ (zmm22 & (zmm10 ^ zmm4)) ; AVX512-FCP-NEXT: vpermq $64, {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Folded Reload ; AVX512-FCP-NEXT: # zmm4 = mem[0,0,0,1,4,4,4,5] ; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm4 = zmm4 ^ (zmm18 & (zmm4 ^ zmm10)) ; AVX512-FCP-NEXT: vmovdqa64 %xmm26, %xmm14 ; AVX512-FCP-NEXT: vpshufb %xmm12, %xmm14, %xmm6 ; AVX512-FCP-NEXT: vmovdqa64 %xmm27, %xmm13 ; AVX512-FCP-NEXT: vpshufb %xmm12, %xmm13, %xmm10 ; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm6 = xmm10[0],xmm6[0],xmm10[1],xmm6[1],xmm10[2],xmm6[2],xmm10[3],xmm6[3],xmm10[4],xmm6[4],xmm10[5],xmm6[5],xmm10[6],xmm6[6],xmm10[7],xmm6[7] ; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm10 = 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] ; AVX512-FCP-NEXT: vprold $16, %xmm10, %xmm10 ; AVX512-FCP-NEXT: vpermt2q %zmm6, %zmm5, %zmm10 ; AVX512-FCP-NEXT: vmovdqa64 %xmm31, %xmm13 ; AVX512-FCP-NEXT: vpshufb %xmm7, %xmm13, %xmm6 ; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload ; AVX512-FCP-NEXT: vpshufb %xmm7, %xmm12, %xmm7 ; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm6 = xmm7[8],xmm6[8],xmm7[9],xmm6[9],xmm7[10],xmm6[10],xmm7[11],xmm6[11],xmm7[12],xmm6[12],xmm7[13],xmm6[13],xmm7[14],xmm6[14],xmm7[15],xmm6[15] ; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm7 = 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] ; AVX512-FCP-NEXT: vpshufb %xmm11, %xmm7, %xmm7 ; AVX512-FCP-NEXT: vpermt2q %zmm6, %zmm5, %zmm7 ; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm7 = zmm10 ^ (zmm22 & (zmm7 ^ zmm10)) ; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm0 = zmm0[0,0,0,1,4,4,4,5] ; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (zmm18 & (zmm0 ^ zmm7)) ; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512-FCP-NEXT: vmovdqa64 %zmm8, 256(%rax) ; AVX512-FCP-NEXT: vmovdqa64 %zmm3, 64(%rax) ; AVX512-FCP-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Folded Reload ; AVX512-FCP-NEXT: # zmm3 = mem[2,2,2,3,6,6,6,7] ; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm5 = [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] ; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm3 = zmm3 ^ (zmm5 & (zmm3 ^ zmm9)) ; AVX512-FCP-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Folded Reload ; AVX512-FCP-NEXT: # zmm6 = mem[2,2,2,3,6,6,6,7] ; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm6 = zmm6 ^ (zmm5 & (zmm6 ^ zmm2)) ; AVX512-FCP-NEXT: vpermq $64, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Folded Reload ; AVX512-FCP-NEXT: # zmm2 = mem[0,0,0,1,4,4,4,5] ; AVX512-FCP-NEXT: vmovdqa64 {{.*#+}} zmm5 = [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] ; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm2 = zmm2 ^ (zmm5 & (zmm2 ^ zmm4)) ; AVX512-FCP-NEXT: vpermq {{.*#+}} zmm1 = zmm1[0,0,0,1,4,4,4,5] ; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm1 = zmm1 ^ (zmm5 & (zmm1 ^ zmm0)) ; AVX512-FCP-NEXT: vmovdqa64 %zmm1, (%rax) ; AVX512-FCP-NEXT: vmovdqa64 %zmm2, 192(%rax) ; AVX512-FCP-NEXT: vmovdqa64 %zmm6, 128(%rax) ; AVX512-FCP-NEXT: vmovdqa64 %zmm3, 320(%rax) ; AVX512-FCP-NEXT: addq $280, %rsp # imm = 0x118 ; AVX512-FCP-NEXT: vzeroupper ; AVX512-FCP-NEXT: retq ; ; AVX512DQ-LABEL: store_i8_stride6_vf64: ; AVX512DQ: # %bb.0: ; AVX512DQ-NEXT: subq $456, %rsp # imm = 0x1C8 ; AVX512DQ-NEXT: vmovdqa (%rsi), %xmm6 ; AVX512DQ-NEXT: vmovdqa 32(%rsi), %xmm3 ; AVX512DQ-NEXT: vmovdqa (%rdi), %xmm7 ; AVX512DQ-NEXT: vmovdqa 32(%rdi), %xmm2 ; AVX512DQ-NEXT: vmovdqa (%r8), %xmm10 ; AVX512DQ-NEXT: vmovdqa 32(%r8), %xmm4 ; AVX512DQ-NEXT: vmovdqa (%r9), %xmm12 ; AVX512DQ-NEXT: vmovdqa 32(%r9), %xmm8 ; AVX512DQ-NEXT: vmovdqa 32(%rcx), %ymm13 ; AVX512DQ-NEXT: vpbroadcastq {{.*#+}} xmm9 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0] ; AVX512DQ-NEXT: vpshufb %xmm9, %xmm3, %xmm0 ; AVX512DQ-NEXT: vpshufb %xmm9, %xmm2, %xmm1 ; AVX512DQ-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] ; AVX512DQ-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] ; AVX512DQ-NEXT: vmovdqa64 %xmm3, %xmm30 ; AVX512DQ-NEXT: vmovdqa64 %xmm2, %xmm31 ; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[0,3,2,1,4,5,6,7] ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,5,6,5] ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill ; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm0 = [6,u,5,u,8,u,7,u,9,u,9,u,9,u,9,u] ; AVX512DQ-NEXT: vpshufb %xmm0, %xmm4, %xmm2 ; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm1 = [2,u,1,u,0,u,3,u,4,u,4,u,4,u,4,u] ; AVX512DQ-NEXT: vpshufb %xmm1, %xmm4, %xmm3 ; AVX512DQ-NEXT: vmovdqa64 %xmm4, %xmm29 ; AVX512DQ-NEXT: vinserti32x4 $2, %xmm2, %zmm3, %zmm2 ; AVX512DQ-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill ; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm2 = [u,6,u,5,u,8,u,7,u,9,u,9,u,9,u,9] ; AVX512DQ-NEXT: vpshufb %xmm2, %xmm8, %xmm4 ; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm3 = [u,2,u,1,u,0,u,3,u,4,u,4,u,4,u,4] ; AVX512DQ-NEXT: vpshufb %xmm3, %xmm8, %xmm5 ; AVX512DQ-NEXT: vmovdqa64 %xmm8, %xmm27 ; AVX512DQ-NEXT: vinserti32x4 $2, %xmm4, %zmm5, %zmm4 ; AVX512DQ-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill ; AVX512DQ-NEXT: vmovdqa 32(%rdx), %ymm8 ; AVX512DQ-NEXT: vpshufb %xmm9, %xmm6, %xmm4 ; AVX512DQ-NEXT: vpshufb %xmm9, %xmm7, %xmm5 ; AVX512DQ-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] ; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm5 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3],xmm7[4],xmm6[4],xmm7[5],xmm6[5],xmm7[6],xmm6[6],xmm7[7],xmm6[7] ; AVX512DQ-NEXT: vmovdqa64 %xmm7, %xmm18 ; AVX512DQ-NEXT: vmovdqa64 %xmm6, %xmm20 ; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm5 = xmm5[0,3,2,1,4,5,6,7] ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,4,5,6,5] ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm4, %zmm5, %zmm4 ; AVX512DQ-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill ; AVX512DQ-NEXT: vmovdqa 32(%rsi), %ymm4 ; AVX512DQ-NEXT: vmovdqa 32(%rdi), %ymm11 ; AVX512DQ-NEXT: vpshufb %xmm0, %xmm10, %xmm0 ; AVX512DQ-NEXT: vpshufb %xmm1, %xmm10, %xmm1 ; AVX512DQ-NEXT: vmovdqa64 %xmm10, %xmm22 ; AVX512DQ-NEXT: vinserti32x4 $2, %xmm0, %zmm1, %zmm0 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill ; AVX512DQ-NEXT: vpshufb %xmm2, %xmm12, %xmm0 ; AVX512DQ-NEXT: vpshufb %xmm3, %xmm12, %xmm1 ; AVX512DQ-NEXT: vmovdqa64 %xmm12, %xmm21 ; AVX512DQ-NEXT: vinserti32x4 $2, %xmm0, %zmm1, %zmm0 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill ; AVX512DQ-NEXT: vpbroadcastq {{.*#+}} ymm15 = [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] ; AVX512DQ-NEXT: vpshufb %ymm15, %ymm4, %ymm0 ; AVX512DQ-NEXT: vpshufb %ymm15, %ymm11, %ymm1 ; AVX512DQ-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] ; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} ymm1 = ymm11[8],ymm4[8],ymm11[9],ymm4[9],ymm11[10],ymm4[10],ymm11[11],ymm4[11],ymm11[12],ymm4[12],ymm11[13],ymm4[13],ymm11[14],ymm4[14],ymm11[15],ymm4[15],ymm11[24],ymm4[24],ymm11[25],ymm4[25],ymm11[26],ymm4[26],ymm11[27],ymm4[27],ymm11[28],ymm4[28],ymm11[29],ymm4[29],ymm11[30],ymm4[30],ymm11[31],ymm4[31] ; AVX512DQ-NEXT: vmovdqa64 %ymm4, %ymm19 ; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm6 = [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] ; AVX512DQ-NEXT: # ymm6 = mem[0,1,0,1] ; AVX512DQ-NEXT: vpshufb %ymm6, %ymm1, %ymm1 ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, (%rsp) # 64-byte Spill ; AVX512DQ-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] ; AVX512DQ-NEXT: vpshufb %ymm12, %ymm13, %ymm0 ; AVX512DQ-NEXT: vpshufb %ymm12, %ymm8, %ymm1 ; AVX512DQ-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] ; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} ymm1 = ymm8[8],ymm13[8],ymm8[9],ymm13[9],ymm8[10],ymm13[10],ymm8[11],ymm13[11],ymm8[12],ymm13[12],ymm8[13],ymm13[13],ymm8[14],ymm13[14],ymm8[15],ymm13[15],ymm8[24],ymm13[24],ymm8[25],ymm13[25],ymm8[26],ymm13[26],ymm8[27],ymm13[27],ymm8[28],ymm13[28],ymm8[29],ymm13[29],ymm8[30],ymm13[30],ymm8[31],ymm13[31] ; AVX512DQ-NEXT: vmovdqa64 %ymm13, %ymm17 ; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm7 = [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] ; AVX512DQ-NEXT: # ymm7 = mem[0,1,0,1] ; AVX512DQ-NEXT: vpshufb %ymm7, %ymm1, %ymm1 ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill ; AVX512DQ-NEXT: vmovdqa 32(%r8), %ymm13 ; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm5 = [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] ; AVX512DQ-NEXT: # ymm5 = mem[0,1,0,1] ; AVX512DQ-NEXT: vpshufb %ymm5, %ymm13, %ymm0 ; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [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] ; AVX512DQ-NEXT: # ymm4 = mem[0,1,0,1] ; AVX512DQ-NEXT: vpshufb %ymm4, %ymm13, %ymm1 ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0 ; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill ; AVX512DQ-NEXT: vmovdqa 32(%r9), %ymm14 ; AVX512DQ-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] ; AVX512DQ-NEXT: # ymm3 = mem[0,1,0,1] ; AVX512DQ-NEXT: vpshufb %ymm3, %ymm14, %ymm0 ; AVX512DQ-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] ; AVX512DQ-NEXT: # ymm2 = mem[0,1,0,1] ; AVX512DQ-NEXT: vpshufb %ymm2, %ymm14, %ymm1 ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm28 ; AVX512DQ-NEXT: vmovdqa (%rsi), %ymm1 ; AVX512DQ-NEXT: vmovdqa (%rdi), %ymm0 ; AVX512DQ-NEXT: vpshufb %ymm15, %ymm1, %ymm10 ; AVX512DQ-NEXT: vpshufb %ymm15, %ymm0, %ymm15 ; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} ymm10 = ymm15[0],ymm10[0],ymm15[1],ymm10[1],ymm15[2],ymm10[2],ymm15[3],ymm10[3],ymm15[4],ymm10[4],ymm15[5],ymm10[5],ymm15[6],ymm10[6],ymm15[7],ymm10[7],ymm15[16],ymm10[16],ymm15[17],ymm10[17],ymm15[18],ymm10[18],ymm15[19],ymm10[19],ymm15[20],ymm10[20],ymm15[21],ymm10[21],ymm15[22],ymm10[22],ymm15[23],ymm10[23] ; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} ymm15 = 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] ; AVX512DQ-NEXT: vpshufb %ymm6, %ymm15, %ymm6 ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm6, %zmm10, %zmm26 ; AVX512DQ-NEXT: vmovdqa (%rcx), %ymm10 ; AVX512DQ-NEXT: vmovdqa (%rdx), %ymm15 ; AVX512DQ-NEXT: vpshufb %ymm12, %ymm10, %ymm6 ; AVX512DQ-NEXT: vpshufb %ymm12, %ymm15, %ymm9 ; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} ymm6 = ymm9[0],ymm6[0],ymm9[1],ymm6[1],ymm9[2],ymm6[2],ymm9[3],ymm6[3],ymm9[4],ymm6[4],ymm9[5],ymm6[5],ymm9[6],ymm6[6],ymm9[7],ymm6[7],ymm9[16],ymm6[16],ymm9[17],ymm6[17],ymm9[18],ymm6[18],ymm9[19],ymm6[19],ymm9[20],ymm6[20],ymm9[21],ymm6[21],ymm9[22],ymm6[22],ymm9[23],ymm6[23] ; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} ymm9 = ymm15[8],ymm10[8],ymm15[9],ymm10[9],ymm15[10],ymm10[10],ymm15[11],ymm10[11],ymm15[12],ymm10[12],ymm15[13],ymm10[13],ymm15[14],ymm10[14],ymm15[15],ymm10[15],ymm15[24],ymm10[24],ymm15[25],ymm10[25],ymm15[26],ymm10[26],ymm15[27],ymm10[27],ymm15[28],ymm10[28],ymm15[29],ymm10[29],ymm15[30],ymm10[30],ymm15[31],ymm10[31] ; AVX512DQ-NEXT: vpshufb %ymm7, %ymm9, %ymm9 ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm9, %zmm6, %zmm24 ; AVX512DQ-NEXT: vmovdqa (%r8), %ymm6 ; AVX512DQ-NEXT: vpshufb %ymm5, %ymm6, %ymm5 ; AVX512DQ-NEXT: vpshufb %ymm4, %ymm6, %ymm4 ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm5, %zmm4, %zmm23 ; AVX512DQ-NEXT: vmovdqa (%r9), %ymm4 ; AVX512DQ-NEXT: vpshufb %ymm3, %ymm4, %ymm3 ; AVX512DQ-NEXT: vpshufb %ymm2, %ymm4, %ymm2 ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm25 ; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} ymm2 = ymm15[0],ymm10[0],ymm15[1],ymm10[1],ymm15[2],ymm10[2],ymm15[3],ymm10[3],ymm15[4],ymm10[4],ymm15[5],ymm10[5],ymm15[6],ymm10[6],ymm15[7],ymm10[7],ymm15[16],ymm10[16],ymm15[17],ymm10[17],ymm15[18],ymm10[18],ymm15[19],ymm10[19],ymm15[20],ymm10[20],ymm15[21],ymm10[21],ymm15[22],ymm10[22],ymm15[23],ymm10[23] ; AVX512DQ-NEXT: vmovdqa64 %ymm2, %ymm16 ; AVX512DQ-NEXT: vmovdqa64 %xmm18, %xmm2 ; AVX512DQ-NEXT: vmovdqa64 %xmm20, %xmm3 ; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm5 = 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] ; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} ymm0 = 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] ; AVX512DQ-NEXT: vmovdqa64 %ymm0, %ymm18 ; AVX512DQ-NEXT: vmovdqa (%rcx), %xmm15 ; AVX512DQ-NEXT: vmovdqa (%rdx), %xmm2 ; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm2[8],xmm15[8],xmm2[9],xmm15[9],xmm2[10],xmm15[10],xmm2[11],xmm15[11],xmm2[12],xmm15[12],xmm2[13],xmm15[13],xmm2[14],xmm15[14],xmm2[15],xmm15[15] ; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm3 = [10,11,8,9,6,7,12,13,14,15,14,15,14,15,14,15] ; AVX512DQ-NEXT: vpshufb %xmm3, %xmm1, %xmm10 ; AVX512DQ-NEXT: vmovdqa 32(%rcx), %xmm1 ; AVX512DQ-NEXT: vmovdqa 32(%rdx), %xmm0 ; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm9 = xmm0[8],xmm1[8],xmm0[9],xmm1[9],xmm0[10],xmm1[10],xmm0[11],xmm1[11],xmm0[12],xmm1[12],xmm0[13],xmm1[13],xmm0[14],xmm1[14],xmm0[15],xmm1[15] ; AVX512DQ-NEXT: vpshufb %xmm3, %xmm9, %xmm3 ; AVX512DQ-NEXT: vmovdqa64 %ymm3, %ymm20 ; AVX512DQ-NEXT: vmovdqa64 %ymm17, %ymm3 ; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} ymm3 = ymm8[0],ymm3[0],ymm8[1],ymm3[1],ymm8[2],ymm3[2],ymm8[3],ymm3[3],ymm8[4],ymm3[4],ymm8[5],ymm3[5],ymm8[6],ymm3[6],ymm8[7],ymm3[7],ymm8[16],ymm3[16],ymm8[17],ymm3[17],ymm8[18],ymm3[18],ymm8[19],ymm3[19],ymm8[20],ymm3[20],ymm8[21],ymm3[21],ymm8[22],ymm3[22],ymm8[23],ymm3[23] ; AVX512DQ-NEXT: vmovdqa64 %ymm3, %ymm17 ; AVX512DQ-NEXT: vmovdqa64 %xmm30, %xmm3 ; AVX512DQ-NEXT: vmovdqa64 %xmm31, %xmm7 ; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm7[8],xmm3[8],xmm7[9],xmm3[9],xmm7[10],xmm3[10],xmm7[11],xmm3[11],xmm7[12],xmm3[12],xmm7[13],xmm3[13],xmm7[14],xmm3[14],xmm7[15],xmm3[15] ; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm8 = [8,9,6,7,12,13,10,11,14,15,14,15,14,15,14,15] ; AVX512DQ-NEXT: vpshufb %xmm8, %xmm5, %xmm9 ; AVX512DQ-NEXT: vpshufb %xmm8, %xmm3, %xmm3 ; AVX512DQ-NEXT: vmovdqa64 %ymm3, %ymm31 ; AVX512DQ-NEXT: vmovdqa64 %ymm19, %ymm3 ; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} ymm3 = ymm11[0],ymm3[0],ymm11[1],ymm3[1],ymm11[2],ymm3[2],ymm11[3],ymm3[3],ymm11[4],ymm3[4],ymm11[5],ymm3[5],ymm11[6],ymm3[6],ymm11[7],ymm3[7],ymm11[16],ymm3[16],ymm11[17],ymm3[17],ymm11[18],ymm3[18],ymm11[19],ymm3[19],ymm11[20],ymm3[20],ymm11[21],ymm3[21],ymm11[22],ymm3[22],ymm11[23],ymm3[23] ; AVX512DQ-NEXT: vmovdqa64 %ymm3, %ymm19 ; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm3 = [10,u,13,u,12,u,11,u,14,u,13,u,14,u,15,u] ; AVX512DQ-NEXT: vmovdqa64 %xmm22, %xmm5 ; AVX512DQ-NEXT: vpshufb %xmm3, %xmm5, %xmm7 ; AVX512DQ-NEXT: vmovdqa64 %xmm29, %xmm5 ; AVX512DQ-NEXT: vpshufb %xmm3, %xmm5, %xmm3 ; AVX512DQ-NEXT: vmovdqa64 %ymm3, %ymm22 ; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm11 = [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] ; AVX512DQ-NEXT: # ymm11 = mem[0,1,0,1] ; AVX512DQ-NEXT: vpshufb %ymm11, %ymm6, %ymm5 ; AVX512DQ-NEXT: vpshufb %ymm11, %ymm13, %ymm6 ; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm11 = [u,10,u,13,u,12,u,11,u,14,u,13,u,14,u,15] ; AVX512DQ-NEXT: vmovdqa64 %xmm21, %xmm3 ; AVX512DQ-NEXT: vpshufb %xmm11, %xmm3, %xmm13 ; AVX512DQ-NEXT: vmovdqa64 %xmm27, %xmm3 ; AVX512DQ-NEXT: vpshufb %xmm11, %xmm3, %xmm11 ; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [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] ; AVX512DQ-NEXT: # ymm3 = mem[0,1,0,1] ; AVX512DQ-NEXT: vpshufb %ymm3, %ymm4, %ymm4 ; AVX512DQ-NEXT: vpshufb %ymm3, %ymm14, %ymm3 ; AVX512DQ-NEXT: vpshufb %xmm12, %xmm1, %xmm14 ; AVX512DQ-NEXT: vpshufb %xmm12, %xmm0, %xmm8 ; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm8 = xmm8[0],xmm14[0],xmm8[1],xmm14[1],xmm8[2],xmm14[2],xmm8[3],xmm14[3],xmm8[4],xmm14[4],xmm8[5],xmm14[5],xmm8[6],xmm14[6],xmm8[7],xmm14[7] ; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} 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] ; AVX512DQ-NEXT: vprold $16, %xmm0, %xmm0 ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm8, %zmm0, %zmm0 ; AVX512DQ-NEXT: vpshufb %xmm12, %xmm15, %xmm1 ; AVX512DQ-NEXT: vpshufb %xmm12, %xmm2, %xmm8 ; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm8[0],xmm1[0],xmm8[1],xmm1[1],xmm8[2],xmm1[2],xmm8[3],xmm1[3],xmm8[4],xmm1[4],xmm8[5],xmm1[5],xmm8[6],xmm1[6],xmm8[7],xmm1[7] ; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm2[0],xmm15[0],xmm2[1],xmm15[1],xmm2[2],xmm15[2],xmm2[3],xmm15[3],xmm2[4],xmm15[4],xmm2[5],xmm15[5],xmm2[6],xmm15[6],xmm2[7],xmm15[7] ; AVX512DQ-NEXT: vprold $16, %xmm2, %xmm2 ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm2, %zmm30 ; AVX512DQ-NEXT: vpermq {{.*#+}} ymm2 = ymm10[0,0,0,1] ; AVX512DQ-NEXT: vprold $16, %ymm16, %ymm8 ; AVX512DQ-NEXT: vpermq {{.*#+}} ymm9 = ymm9[0,0,0,1] ; AVX512DQ-NEXT: vmovdqa64 %ymm18, %ymm1 ; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm10 = ymm1[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15] ; AVX512DQ-NEXT: vpshufhw {{.*#+}} ymm10 = ymm10[0,1,2,3,4,5,6,5,8,9,10,11,12,13,14,13] ; AVX512DQ-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,2,2,3] ; AVX512DQ-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,0,0,1] ; AVX512DQ-NEXT: vpermq {{.*#+}} ymm5 = ymm5[2,2,2,3] ; AVX512DQ-NEXT: vpermq {{.*#+}} ymm12 = ymm13[0,0,0,1] ; AVX512DQ-NEXT: vpermq {{.*#+}} ymm4 = ymm4[2,2,2,3] ; AVX512DQ-NEXT: vpermq {{.*#+}} ymm13 = ymm20[0,0,0,1] ; AVX512DQ-NEXT: vprold $16, %ymm17, %ymm14 ; AVX512DQ-NEXT: vpermq {{.*#+}} ymm15 = ymm31[0,0,0,1] ; AVX512DQ-NEXT: vmovdqa64 %ymm19, %ymm1 ; AVX512DQ-NEXT: vpshuflw {{.*#+}} ymm1 = ymm1[0,3,2,1,4,5,6,7,8,11,10,9,12,13,14,15] ; AVX512DQ-NEXT: vpshufhw {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5,6,5,8,9,10,11,12,13,14,13] ; AVX512DQ-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3] ; AVX512DQ-NEXT: vpermq {{.*#+}} ymm31 = ymm22[0,0,0,1] ; AVX512DQ-NEXT: vpermq {{.*#+}} ymm6 = ymm6[2,2,2,3] ; AVX512DQ-NEXT: vpermq {{.*#+}} ymm11 = ymm11[0,0,0,1] ; AVX512DQ-NEXT: vpermq {{.*#+}} ymm3 = ymm3[2,2,2,3] ; AVX512DQ-NEXT: vpermq {{.*#+}} ymm8 = ymm8[2,2,2,3] ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm8, %zmm2, %zmm2 ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm10, %zmm9, %zmm8 ; AVX512DQ-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] ; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm8 = zmm2 ^ (zmm9 & (zmm8 ^ zmm2)) ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm5, %zmm7, %zmm2 ; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm5 = [0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535] ; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm2 = zmm2 ^ (zmm5 & (zmm2 ^ zmm8)) ; AVX512DQ-NEXT: vpermq {{.*#+}} ymm7 = ymm14[2,2,2,3] ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm7, %zmm13, %zmm7 ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm15, %zmm1 ; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm1 = zmm7 ^ (zmm9 & (zmm1 ^ zmm7)) ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm6, %zmm31, %zmm6 ; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm6 = zmm6 ^ (zmm5 & (zmm6 ^ zmm1)) ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm4, %zmm12, %zmm1 ; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm4 = [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] ; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm1 = zmm1 ^ (zmm4 & (zmm1 ^ zmm2)) ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm3, %zmm11, %zmm2 ; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm2 = zmm2 ^ (zmm4 & (zmm2 ^ zmm6)) ; AVX512DQ-NEXT: vpermq {{.*#+}} zmm0 = zmm0[0,0,0,1,4,4,4,5] ; AVX512DQ-NEXT: vpermq $64, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Folded Reload ; AVX512DQ-NEXT: # zmm3 = mem[0,0,0,1,4,4,4,5] ; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm4 = [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] ; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm3 = zmm0 ^ (zmm4 & (zmm3 ^ zmm0)) ; AVX512DQ-NEXT: vpermq $64, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Folded Reload ; AVX512DQ-NEXT: # zmm0 = mem[0,0,0,1,4,4,4,5] ; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (zmm9 & (zmm0 ^ zmm3)) ; AVX512DQ-NEXT: vpermq $64, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Folded Reload ; AVX512DQ-NEXT: # zmm3 = mem[0,0,0,1,4,4,4,5] ; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm5 = [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] ; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm3 = zmm3 ^ (zmm5 & (zmm3 ^ zmm0)) ; AVX512DQ-NEXT: vpermq {{.*#+}} zmm0 = zmm30[0,0,0,1,4,4,4,5] ; AVX512DQ-NEXT: vpermq $64, {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Folded Reload ; AVX512DQ-NEXT: # zmm6 = mem[0,0,0,1,4,4,4,5] ; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm6 = zmm0 ^ (zmm4 & (zmm6 ^ zmm0)) ; AVX512DQ-NEXT: vpermq $64, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Folded Reload ; AVX512DQ-NEXT: # zmm0 = mem[0,0,0,1,4,4,4,5] ; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (zmm9 & (zmm0 ^ zmm6)) ; AVX512DQ-NEXT: vpermq $64, {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Folded Reload ; AVX512DQ-NEXT: # zmm6 = mem[0,0,0,1,4,4,4,5] ; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm6 = zmm6 ^ (zmm5 & (zmm6 ^ zmm0)) ; AVX512DQ-NEXT: vpermq $234, (%rsp), %zmm0 # 64-byte Folded Reload ; AVX512DQ-NEXT: # zmm0 = mem[2,2,2,3,6,6,6,7] ; AVX512DQ-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm5 # 64-byte Folded Reload ; AVX512DQ-NEXT: # zmm5 = mem[2,2,2,3,6,6,6,7] ; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm5 = zmm0 ^ (zmm9 & (zmm5 ^ zmm0)) ; AVX512DQ-NEXT: vpermq {{.*#+}} zmm0 = zmm26[2,2,2,3,6,6,6,7] ; AVX512DQ-NEXT: vpermq {{.*#+}} zmm7 = zmm24[2,2,2,3,6,6,6,7] ; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm7 = zmm0 ^ (zmm9 & (zmm7 ^ zmm0)) ; AVX512DQ-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Folded Reload ; AVX512DQ-NEXT: # zmm0 = mem[2,2,2,3,6,6,6,7] ; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (zmm4 & (zmm0 ^ zmm5)) ; AVX512DQ-NEXT: vpermq {{.*#+}} zmm5 = zmm23[2,2,2,3,6,6,6,7] ; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm5 = zmm5 ^ (zmm4 & (zmm5 ^ zmm7)) ; AVX512DQ-NEXT: vpermq {{.*#+}} zmm4 = zmm28[2,2,2,3,6,6,6,7] ; AVX512DQ-NEXT: vmovdqa64 {{.*#+}} zmm7 = [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] ; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm4 = zmm4 ^ (zmm7 & (zmm4 ^ zmm0)) ; AVX512DQ-NEXT: vpermq {{.*#+}} zmm0 = zmm25[2,2,2,3,6,6,6,7] ; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (zmm7 & (zmm0 ^ zmm5)) ; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512DQ-NEXT: vmovdqa64 %zmm0, 128(%rax) ; AVX512DQ-NEXT: vmovdqa64 %zmm4, 320(%rax) ; AVX512DQ-NEXT: vmovdqa64 %zmm6, (%rax) ; AVX512DQ-NEXT: vmovdqa64 %zmm3, 192(%rax) ; AVX512DQ-NEXT: vmovdqa64 %zmm2, 256(%rax) ; AVX512DQ-NEXT: vmovdqa64 %zmm1, 64(%rax) ; AVX512DQ-NEXT: addq $456, %rsp # imm = 0x1C8 ; AVX512DQ-NEXT: vzeroupper ; AVX512DQ-NEXT: retq ; ; AVX512DQ-FCP-LABEL: store_i8_stride6_vf64: ; AVX512DQ-FCP: # %bb.0: ; AVX512DQ-FCP-NEXT: subq $280, %rsp # imm = 0x118 ; AVX512DQ-FCP-NEXT: vmovdqa (%rcx), %ymm1 ; AVX512DQ-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %ymm0 ; AVX512DQ-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} ymm0 = 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] ; AVX512DQ-FCP-NEXT: vprold $16, %ymm0, %ymm18 ; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm23 = [0,0,0,1,10,10,10,11] ; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %ymm6 ; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %ymm7 ; AVX512DQ-FCP-NEXT: vmovdqa (%r8), %ymm11 ; AVX512DQ-FCP-NEXT: vmovdqa 32(%r8), %xmm8 ; AVX512DQ-FCP-NEXT: vmovdqa (%r9), %ymm10 ; AVX512DQ-FCP-NEXT: vmovdqa 32(%r9), %xmm9 ; AVX512DQ-FCP-NEXT: vmovdqa 32(%r8), %ymm4 ; AVX512DQ-FCP-NEXT: vmovdqa 32(%r9), %ymm12 ; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm0 = [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] ; AVX512DQ-FCP-NEXT: # ymm0 = mem[0,1,0,1] ; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm4, %ymm2 ; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm1 = [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] ; AVX512DQ-FCP-NEXT: # ymm1 = mem[0,1,0,1] ; AVX512DQ-FCP-NEXT: vpshufb %ymm1, %ymm4, %ymm3 ; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm4, %ymm21 ; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm3, %zmm2 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill ; AVX512DQ-FCP-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] ; AVX512DQ-FCP-NEXT: # ymm3 = mem[0,1,0,1] ; AVX512DQ-FCP-NEXT: vpshufb %ymm3, %ymm12, %ymm2 ; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [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] ; AVX512DQ-FCP-NEXT: # ymm4 = mem[0,1,0,1] ; AVX512DQ-FCP-NEXT: vpshufb %ymm4, %ymm12, %ymm5 ; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm12, %ymm22 ; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm5, %zmm2 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill ; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm11, %ymm0 ; AVX512DQ-FCP-NEXT: vpshufb %ymm1, %ymm11, %ymm1 ; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm30 ; AVX512DQ-FCP-NEXT: vpshufb %ymm3, %ymm10, %ymm0 ; AVX512DQ-FCP-NEXT: vpshufb %ymm4, %ymm10, %ymm1 ; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm10, %ymm19 ; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm0 = xmm8[6,u,5,u,8,u,7,u,9,u,9,u,9,u,9,u] ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm1 = xmm8[2,u,1,u,0,u,3,u,4,u,4,u,4,u,4,u] ; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm8, %xmm16 ; AVX512DQ-FCP-NEXT: vinserti32x4 $2, %xmm0, %zmm1, %zmm0 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm0 = xmm9[u,6,u,5,u,8,u,7,u,9,u,9,u,9,u,9] ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm1 = xmm9[u,2,u,1,u,0,u,3,u,4,u,4,u,4,u,4] ; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm9, %xmm17 ; AVX512DQ-FCP-NEXT: vinserti32x4 $2, %xmm0, %zmm1, %zmm0 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill ; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} ymm0 = 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] ; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm7, %ymm29 ; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm6, %ymm28 ; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm15 = [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] ; AVX512DQ-FCP-NEXT: # ymm15 = mem[0,1,0,1] ; AVX512DQ-FCP-NEXT: vpshufb %ymm15, %ymm0, %ymm0 ; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %xmm2 ; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %xmm1 ; AVX512DQ-FCP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX512DQ-FCP-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] ; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm2, %xmm31 ; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm13 = [8,9,6,7,12,13,10,11,14,15,14,15,14,15,14,15] ; AVX512DQ-FCP-NEXT: vpshufb %xmm13, %xmm1, %xmm1 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm23, %zmm1 ; AVX512DQ-FCP-NEXT: vmovdqa (%rcx), %xmm2 ; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %xmm3 ; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm0 = 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] ; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm3, %xmm27 ; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm2, %xmm26 ; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm6 = [10,11,8,9,6,7,12,13,14,15,14,15,14,15,14,15] ; AVX512DQ-FCP-NEXT: vpshufb %xmm6, %xmm0, %xmm0 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm18, %zmm23, %zmm0 ; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm18 = [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] ; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm1 = zmm0 ^ (zmm18 & (zmm1 ^ zmm0)) ; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm12 = [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] ; AVX512DQ-FCP-NEXT: # ymm12 = mem[0,1,0,1] ; AVX512DQ-FCP-NEXT: vpshufb %ymm12, %ymm11, %ymm2 ; AVX512DQ-FCP-NEXT: vmovdqa (%r8), %xmm10 ; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm8 = [10,u,13,u,12,u,11,u,14,u,13,u,14,u,15,u] ; AVX512DQ-FCP-NEXT: vpshufb %xmm8, %xmm10, %xmm0 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm23, %zmm0 ; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm24 = [0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535,65535,0,65535] ; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (zmm24 & (zmm0 ^ zmm1)) ; AVX512DQ-FCP-NEXT: vmovdqa 32(%rcx), %ymm7 ; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdx), %ymm5 ; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} ymm4 = ymm5[0],ymm7[0],ymm5[1],ymm7[1],ymm5[2],ymm7[2],ymm5[3],ymm7[3],ymm5[4],ymm7[4],ymm5[5],ymm7[5],ymm5[6],ymm7[6],ymm5[7],ymm7[7],ymm5[16],ymm7[16],ymm5[17],ymm7[17],ymm5[18],ymm7[18],ymm5[19],ymm7[19],ymm5[20],ymm7[20],ymm5[21],ymm7[21],ymm5[22],ymm7[22],ymm5[23],ymm7[23] ; AVX512DQ-FCP-NEXT: vprold $16, %ymm4, %ymm25 ; AVX512DQ-FCP-NEXT: vmovdqa 32(%rcx), %xmm4 ; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdx), %xmm11 ; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm9 = xmm11[8],xmm4[8],xmm11[9],xmm4[9],xmm11[10],xmm4[10],xmm11[11],xmm4[11],xmm11[12],xmm4[12],xmm11[13],xmm4[13],xmm11[14],xmm4[14],xmm11[15],xmm4[15] ; AVX512DQ-FCP-NEXT: vpshufb %xmm6, %xmm9, %xmm9 ; AVX512DQ-FCP-NEXT: vmovdqa 32(%rsi), %ymm2 ; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdi), %ymm1 ; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} ymm6 = ymm1[0],ymm2[0],ymm1[1],ymm2[1],ymm1[2],ymm2[2],ymm1[3],ymm2[3],ymm1[4],ymm2[4],ymm1[5],ymm2[5],ymm1[6],ymm2[6],ymm1[7],ymm2[7],ymm1[16],ymm2[16],ymm1[17],ymm2[17],ymm1[18],ymm2[18],ymm1[19],ymm2[19],ymm1[20],ymm2[20],ymm1[21],ymm2[21],ymm1[22],ymm2[22],ymm1[23],ymm2[23] ; AVX512DQ-FCP-NEXT: vpshufb %ymm15, %ymm6, %ymm3 ; AVX512DQ-FCP-NEXT: vmovdqa 32(%rsi), %xmm6 ; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdi), %xmm15 ; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm14 = xmm15[8],xmm6[8],xmm15[9],xmm6[9],xmm15[10],xmm6[10],xmm15[11],xmm6[11],xmm15[12],xmm6[12],xmm15[13],xmm6[13],xmm15[14],xmm6[14],xmm15[15],xmm6[15] ; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm6, %xmm20 ; AVX512DQ-FCP-NEXT: vpshufb %xmm13, %xmm14, %xmm13 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm3, %zmm23, %zmm13 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm25, %zmm23, %zmm9 ; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm13 = zmm9 ^ (zmm18 & (zmm13 ^ zmm9)) ; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm21, %ymm3 ; AVX512DQ-FCP-NEXT: vpshufb %ymm12, %ymm3, %ymm3 ; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm16, %xmm6 ; AVX512DQ-FCP-NEXT: vpshufb %xmm8, %xmm6, %xmm9 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm3, %zmm23, %zmm9 ; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm9 = zmm9 ^ (zmm24 & (zmm9 ^ zmm13)) ; AVX512DQ-FCP-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] ; AVX512DQ-FCP-NEXT: # ymm8 = mem[0,1,0,1] ; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm19, %ymm3 ; AVX512DQ-FCP-NEXT: vpshufb %ymm8, %ymm3, %ymm12 ; AVX512DQ-FCP-NEXT: vmovdqa (%r9), %xmm13 ; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm14 = [u,10,u,13,u,12,u,11,u,14,u,13,u,14,u,15] ; AVX512DQ-FCP-NEXT: vpshufb %xmm14, %xmm13, %xmm3 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm12, %zmm23, %zmm3 ; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm12 = [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] ; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm3 = zmm3 ^ (zmm12 & (zmm3 ^ zmm0)) ; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm22, %ymm0 ; AVX512DQ-FCP-NEXT: vpshufb %ymm8, %ymm0, %ymm0 ; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm17, %xmm6 ; AVX512DQ-FCP-NEXT: vpshufb %xmm14, %xmm6, %xmm8 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm23, %zmm8 ; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm8 = zmm8 ^ (zmm12 & (zmm8 ^ zmm9)) ; AVX512DQ-FCP-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] ; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm2, %ymm9 ; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm1, %ymm12 ; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} ymm9 = ymm12[0],ymm9[0],ymm12[1],ymm9[1],ymm12[2],ymm9[2],ymm12[3],ymm9[3],ymm12[4],ymm9[4],ymm12[5],ymm9[5],ymm12[6],ymm9[6],ymm12[7],ymm9[7],ymm12[16],ymm9[16],ymm12[17],ymm9[17],ymm12[18],ymm9[18],ymm12[19],ymm9[19],ymm12[20],ymm9[20],ymm12[21],ymm9[21],ymm12[22],ymm9[22],ymm12[23],ymm9[23] ; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} ymm1 = ymm1[8],ymm2[8],ymm1[9],ymm2[9],ymm1[10],ymm2[10],ymm1[11],ymm2[11],ymm1[12],ymm2[12],ymm1[13],ymm2[13],ymm1[14],ymm2[14],ymm1[15],ymm2[15],ymm1[24],ymm2[24],ymm1[25],ymm2[25],ymm1[26],ymm2[26],ymm1[27],ymm2[27],ymm1[28],ymm2[28],ymm1[29],ymm2[29],ymm1[30],ymm2[30],ymm1[31],ymm2[31] ; AVX512DQ-FCP-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] ; AVX512DQ-FCP-NEXT: # ymm2 = mem[0,1,0,1] ; AVX512DQ-FCP-NEXT: vpshufb %ymm2, %ymm1, %ymm1 ; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm14 = [2,2,2,3,10,10,10,11] ; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm14, %zmm9 ; AVX512DQ-FCP-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] ; AVX512DQ-FCP-NEXT: vpshufb %ymm12, %ymm7, %ymm1 ; AVX512DQ-FCP-NEXT: vpshufb %ymm12, %ymm5, %ymm6 ; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} ymm1 = ymm6[0],ymm1[0],ymm6[1],ymm1[1],ymm6[2],ymm1[2],ymm6[3],ymm1[3],ymm6[4],ymm1[4],ymm6[5],ymm1[5],ymm6[6],ymm1[6],ymm6[7],ymm1[7],ymm6[16],ymm1[16],ymm6[17],ymm1[17],ymm6[18],ymm1[18],ymm6[19],ymm1[19],ymm6[20],ymm1[20],ymm6[21],ymm1[21],ymm6[22],ymm1[22],ymm6[23],ymm1[23] ; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} ymm5 = ymm5[8],ymm7[8],ymm5[9],ymm7[9],ymm5[10],ymm7[10],ymm5[11],ymm7[11],ymm5[12],ymm7[12],ymm5[13],ymm7[13],ymm5[14],ymm7[14],ymm5[15],ymm7[15],ymm5[24],ymm7[24],ymm5[25],ymm7[25],ymm5[26],ymm7[26],ymm5[27],ymm7[27],ymm5[28],ymm7[28],ymm5[29],ymm7[29],ymm5[30],ymm7[30],ymm5[31],ymm7[31] ; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm6 = [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] ; AVX512DQ-FCP-NEXT: # ymm6 = mem[0,1,0,1] ; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm5, %ymm5 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm5, %zmm14, %zmm1 ; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm1 = zmm9 ^ (zmm18 & (zmm1 ^ zmm9)) ; AVX512DQ-FCP-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm9 # 64-byte Folded Reload ; AVX512DQ-FCP-NEXT: # zmm9 = mem[2,2,2,3,6,6,6,7] ; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm22 = [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] ; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm9 = zmm9 ^ (zmm22 & (zmm9 ^ zmm1)) ; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm28, %ymm5 ; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm5, %ymm1 ; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm29, %ymm7 ; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm7, %ymm0 ; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} ymm0 = 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] ; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} ymm1 = ymm7[8],ymm5[8],ymm7[9],ymm5[9],ymm7[10],ymm5[10],ymm7[11],ymm5[11],ymm7[12],ymm5[12],ymm7[13],ymm5[13],ymm7[14],ymm5[14],ymm7[15],ymm5[15],ymm7[24],ymm5[24],ymm7[25],ymm5[25],ymm7[26],ymm5[26],ymm7[27],ymm5[27],ymm7[28],ymm5[28],ymm7[29],ymm5[29],ymm7[30],ymm5[30],ymm7[31],ymm5[31] ; AVX512DQ-FCP-NEXT: vpshufb %ymm2, %ymm1, %ymm1 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm14, %zmm0 ; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload ; AVX512DQ-FCP-NEXT: vpshufb %ymm12, %ymm7, %ymm1 ; AVX512DQ-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload ; AVX512DQ-FCP-NEXT: vpshufb %ymm12, %ymm5, %ymm2 ; AVX512DQ-FCP-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] ; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} ymm2 = ymm5[8],ymm7[8],ymm5[9],ymm7[9],ymm5[10],ymm7[10],ymm5[11],ymm7[11],ymm5[12],ymm7[12],ymm5[13],ymm7[13],ymm5[14],ymm7[14],ymm5[15],ymm7[15],ymm5[24],ymm7[24],ymm5[25],ymm7[25],ymm5[26],ymm7[26],ymm5[27],ymm7[27],ymm5[28],ymm7[28],ymm5[29],ymm7[29],ymm5[30],ymm7[30],ymm5[31],ymm7[31] ; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm2, %ymm2 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm14, %zmm1 ; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm1 = zmm0 ^ (zmm18 & (zmm1 ^ zmm0)) ; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm2 = zmm30[2,2,2,3,6,6,6,7] ; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm2 = zmm2 ^ (zmm22 & (zmm2 ^ zmm1)) ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm0 = xmm10[6,u,5,u,8,u,7,u,9,u,9,u,9,u,9,u] ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm1 = xmm10[2,u,1,u,0,u,3,u,4,u,4,u,4,u,4,u] ; AVX512DQ-FCP-NEXT: vinserti32x4 $2, %xmm0, %zmm1, %zmm0 ; AVX512DQ-FCP-NEXT: vpshufb %xmm12, %xmm4, %xmm1 ; AVX512DQ-FCP-NEXT: vpshufb %xmm12, %xmm11, %xmm5 ; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm5[0],xmm1[0],xmm5[1],xmm1[1],xmm5[2],xmm1[2],xmm5[3],xmm1[3],xmm5[4],xmm1[4],xmm5[5],xmm1[5],xmm5[6],xmm1[6],xmm5[7],xmm1[7] ; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm4 = xmm11[0],xmm4[0],xmm11[1],xmm4[1],xmm11[2],xmm4[2],xmm11[3],xmm4[3],xmm11[4],xmm4[4],xmm11[5],xmm4[5],xmm11[6],xmm4[6],xmm11[7],xmm4[7] ; AVX512DQ-FCP-NEXT: vprold $16, %xmm4, %xmm4 ; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,0,0,1,8,8,8,9] ; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm5, %zmm4 ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm1 = xmm13[u,6,u,5,u,8,u,7,u,9,u,9,u,9,u,9] ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} xmm6 = xmm13[u,2,u,1,u,0,u,3,u,4,u,4,u,4,u,4] ; AVX512DQ-FCP-NEXT: vpbroadcastq {{.*#+}} xmm7 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0] ; AVX512DQ-FCP-NEXT: vinserti32x4 $2, %xmm1, %zmm6, %zmm1 ; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm20, %xmm11 ; AVX512DQ-FCP-NEXT: vpshufb %xmm7, %xmm11, %xmm6 ; AVX512DQ-FCP-NEXT: vpshufb %xmm7, %xmm15, %xmm10 ; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm6 = xmm10[8],xmm6[8],xmm10[9],xmm6[9],xmm10[10],xmm6[10],xmm10[11],xmm6[11],xmm10[12],xmm6[12],xmm10[13],xmm6[13],xmm10[14],xmm6[14],xmm10[15],xmm6[15] ; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm10 = xmm15[0],xmm11[0],xmm15[1],xmm11[1],xmm15[2],xmm11[2],xmm15[3],xmm11[3],xmm15[4],xmm11[4],xmm15[5],xmm11[5],xmm15[6],xmm11[6],xmm15[7],xmm11[7] ; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} xmm11 = [0,1,6,7,4,5,2,3,8,9,10,11,12,13,10,11] ; AVX512DQ-FCP-NEXT: vpshufb %xmm11, %xmm10, %xmm10 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm6, %zmm5, %zmm10 ; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm10 = zmm4 ^ (zmm22 & (zmm10 ^ zmm4)) ; AVX512DQ-FCP-NEXT: vpermq $64, {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Folded Reload ; AVX512DQ-FCP-NEXT: # zmm4 = mem[0,0,0,1,4,4,4,5] ; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm4 = zmm4 ^ (zmm18 & (zmm4 ^ zmm10)) ; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm26, %xmm14 ; AVX512DQ-FCP-NEXT: vpshufb %xmm12, %xmm14, %xmm6 ; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm27, %xmm13 ; AVX512DQ-FCP-NEXT: vpshufb %xmm12, %xmm13, %xmm10 ; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm6 = xmm10[0],xmm6[0],xmm10[1],xmm6[1],xmm10[2],xmm6[2],xmm10[3],xmm6[3],xmm10[4],xmm6[4],xmm10[5],xmm6[5],xmm10[6],xmm6[6],xmm10[7],xmm6[7] ; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm10 = 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] ; AVX512DQ-FCP-NEXT: vprold $16, %xmm10, %xmm10 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm6, %zmm5, %zmm10 ; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm31, %xmm13 ; AVX512DQ-FCP-NEXT: vpshufb %xmm7, %xmm13, %xmm6 ; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload ; AVX512DQ-FCP-NEXT: vpshufb %xmm7, %xmm12, %xmm7 ; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm6 = xmm7[8],xmm6[8],xmm7[9],xmm6[9],xmm7[10],xmm6[10],xmm7[11],xmm6[11],xmm7[12],xmm6[12],xmm7[13],xmm6[13],xmm7[14],xmm6[14],xmm7[15],xmm6[15] ; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm7 = 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] ; AVX512DQ-FCP-NEXT: vpshufb %xmm11, %xmm7, %xmm7 ; AVX512DQ-FCP-NEXT: vpermt2q %zmm6, %zmm5, %zmm7 ; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm7 = zmm10 ^ (zmm22 & (zmm7 ^ zmm10)) ; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm0 = zmm0[0,0,0,1,4,4,4,5] ; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (zmm18 & (zmm0 ^ zmm7)) ; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm8, 256(%rax) ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, 64(%rax) ; AVX512DQ-FCP-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Folded Reload ; AVX512DQ-FCP-NEXT: # zmm3 = mem[2,2,2,3,6,6,6,7] ; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm5 = [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] ; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm3 = zmm3 ^ (zmm5 & (zmm3 ^ zmm9)) ; AVX512DQ-FCP-NEXT: vpermq $234, {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Folded Reload ; AVX512DQ-FCP-NEXT: # zmm6 = mem[2,2,2,3,6,6,6,7] ; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm6 = zmm6 ^ (zmm5 & (zmm6 ^ zmm2)) ; AVX512DQ-FCP-NEXT: vpermq $64, {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Folded Reload ; AVX512DQ-FCP-NEXT: # zmm2 = mem[0,0,0,1,4,4,4,5] ; AVX512DQ-FCP-NEXT: vmovdqa64 {{.*#+}} zmm5 = [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] ; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm2 = zmm2 ^ (zmm5 & (zmm2 ^ zmm4)) ; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} zmm1 = zmm1[0,0,0,1,4,4,4,5] ; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm1 = zmm1 ^ (zmm5 & (zmm1 ^ zmm0)) ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, (%rax) ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, 192(%rax) ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, 128(%rax) ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, 320(%rax) ; AVX512DQ-FCP-NEXT: addq $280, %rsp # imm = 0x118 ; AVX512DQ-FCP-NEXT: vzeroupper ; AVX512DQ-FCP-NEXT: retq ; ; AVX512BW-LABEL: store_i8_stride6_vf64: ; AVX512BW: # %bb.0: ; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512BW-NEXT: vmovdqa64 (%r8), %zmm2 ; AVX512BW-NEXT: vmovdqa64 (%r9), %zmm1 ; AVX512BW-NEXT: vmovdqa (%rsi), %xmm15 ; AVX512BW-NEXT: vmovdqa 32(%rsi), %xmm4 ; AVX512BW-NEXT: vpbroadcastq {{.*#+}} xmm5 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0] ; AVX512BW-NEXT: vpshufb %xmm5, %xmm4, %xmm0 ; AVX512BW-NEXT: vmovdqa64 (%rdi), %xmm18 ; AVX512BW-NEXT: vmovdqa 32(%rdi), %xmm6 ; AVX512BW-NEXT: vpshufb %xmm5, %xmm6, %xmm3 ; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm3[8],xmm0[8],xmm3[9],xmm0[9],xmm3[10],xmm0[10],xmm3[11],xmm0[11],xmm3[12],xmm0[12],xmm3[13],xmm0[13],xmm3[14],xmm0[14],xmm3[15],xmm0[15] ; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm6[0],xmm4[0],xmm6[1],xmm4[1],xmm6[2],xmm4[2],xmm6[3],xmm4[3],xmm6[4],xmm4[4],xmm6[5],xmm4[5],xmm6[6],xmm4[6],xmm6[7],xmm4[7] ; AVX512BW-NEXT: vpmovsxbw {{.*#+}} zmm7 = [0,3,2,1,0,3,2,1,0,3,2,1,4,5,6,5,32,33,34,35,32,33,34,35,32,33,34,35,36,37,38,39] ; AVX512BW-NEXT: vpermt2w %zmm0, %zmm7, %zmm3 ; AVX512BW-NEXT: vpbroadcastq {{.*#+}} ymm0 = [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-NEXT: vmovdqa64 (%rcx), %xmm16 ; AVX512BW-NEXT: vmovdqa 32(%rcx), %xmm10 ; AVX512BW-NEXT: vpshufb %xmm0, %xmm10, %xmm8 ; AVX512BW-NEXT: vmovdqa64 (%rdx), %xmm17 ; AVX512BW-NEXT: vmovdqa 32(%rdx), %xmm11 ; AVX512BW-NEXT: vpshufb %xmm0, %xmm11, %xmm9 ; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm8 = xmm9[0],xmm8[0],xmm9[1],xmm8[1],xmm9[2],xmm8[2],xmm9[3],xmm8[3],xmm9[4],xmm8[4],xmm9[5],xmm8[5],xmm9[6],xmm8[6],xmm9[7],xmm8[7] ; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm9 = xmm11[0],xmm10[0],xmm11[1],xmm10[1],xmm11[2],xmm10[2],xmm11[3],xmm10[3],xmm11[4],xmm10[4],xmm11[5],xmm10[5],xmm11[6],xmm10[6],xmm11[7],xmm10[7] ; AVX512BW-NEXT: vprold $16, %xmm9, %xmm9 ; AVX512BW-NEXT: vinserti64x4 $1, %ymm8, %zmm9, %zmm8 ; AVX512BW-NEXT: vpermq {{.*#+}} zmm8 = zmm8[0,0,0,1,4,4,4,5] ; AVX512BW-NEXT: movl $-1840700270, %r10d # imm = 0x92492492 ; AVX512BW-NEXT: kmovd %r10d, %k1 ; AVX512BW-NEXT: vmovdqu16 %zmm8, %zmm3 {%k1} ; AVX512BW-NEXT: vmovdqa 32(%r8), %xmm9 ; AVX512BW-NEXT: vpmovzxbw {{.*#+}} xmm8 = xmm9[0],zero,xmm9[1],zero,xmm9[2],zero,xmm9[3],zero,xmm9[4],zero,xmm9[5],zero,xmm9[6],zero,xmm9[7],zero ; AVX512BW-NEXT: vpmovsxbw {{.*#+}} ymm13 = [0,1,16,3,4,17,6,7,18,9,10,19,12,13,20,15] ; AVX512BW-NEXT: vmovdqa %ymm3, %ymm12 ; AVX512BW-NEXT: vpermt2w %ymm8, %ymm13, %ymm12 ; AVX512BW-NEXT: vpmovsxbw {{.*#+}} ymm8 = [16,37,18,19,38,21,22,39,24,25,32,27,28,33,30,31] ; AVX512BW-NEXT: vpshufd {{.*#+}} xmm14 = xmm9[2,1,2,3] ; AVX512BW-NEXT: vpmovzxbw {{.*#+}} xmm14 = xmm14[0],zero,xmm14[1],zero,xmm14[2],zero,xmm14[3],zero,xmm14[4],zero,xmm14[5],zero,xmm14[6],zero,xmm14[7],zero ; AVX512BW-NEXT: vpermt2w %zmm14, %zmm8, %zmm3 ; AVX512BW-NEXT: vinserti64x4 $1, %ymm3, %zmm12, %zmm3 ; AVX512BW-NEXT: vmovdqa 32(%r9), %xmm12 ; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm14 = xmm12[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7] ; AVX512BW-NEXT: vpshufd {{.*#+}} xmm19 = xmm12[2,1,2,3] ; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm19 = xmm19[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7] ; AVX512BW-NEXT: vpmovsxbw {{.*#+}} zmm20 = [2,1,0,3,2,1,0,3,2,1,0,3,4,4,4,4,38,37,32,39,38,37,32,39,38,37,32,39,33,33,33,33] ; AVX512BW-NEXT: vpermt2w %zmm19, %zmm20, %zmm14 ; AVX512BW-NEXT: movabsq $585610922974906400, %r10 # imm = 0x820820820820820 ; AVX512BW-NEXT: kmovq %r10, %k2 ; AVX512BW-NEXT: vmovdqu8 %zmm14, %zmm3 {%k2} ; AVX512BW-NEXT: vpshufb %xmm5, %xmm15, %xmm14 ; AVX512BW-NEXT: vpshufb %xmm5, %xmm18, %xmm5 ; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm5 = xmm5[8],xmm14[8],xmm5[9],xmm14[9],xmm5[10],xmm14[10],xmm5[11],xmm14[11],xmm5[12],xmm14[12],xmm5[13],xmm14[13],xmm5[14],xmm14[14],xmm5[15],xmm14[15] ; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm14 = xmm18[0],xmm15[0],xmm18[1],xmm15[1],xmm18[2],xmm15[2],xmm18[3],xmm15[3],xmm18[4],xmm15[4],xmm18[5],xmm15[5],xmm18[6],xmm15[6],xmm18[7],xmm15[7] ; AVX512BW-NEXT: vpermt2w %zmm5, %zmm7, %zmm14 ; AVX512BW-NEXT: vpshufb %xmm0, %xmm16, %xmm5 ; AVX512BW-NEXT: vpshufb %xmm0, %xmm17, %xmm7 ; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm5 = xmm7[0],xmm5[0],xmm7[1],xmm5[1],xmm7[2],xmm5[2],xmm7[3],xmm5[3],xmm7[4],xmm5[4],xmm7[5],xmm5[5],xmm7[6],xmm5[6],xmm7[7],xmm5[7] ; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm7 = xmm17[0],xmm16[0],xmm17[1],xmm16[1],xmm17[2],xmm16[2],xmm17[3],xmm16[3],xmm17[4],xmm16[4],xmm17[5],xmm16[5],xmm17[6],xmm16[6],xmm17[7],xmm16[7] ; AVX512BW-NEXT: vprold $16, %xmm7, %xmm7 ; AVX512BW-NEXT: vinserti64x4 $1, %ymm5, %zmm7, %zmm5 ; AVX512BW-NEXT: vpermq {{.*#+}} zmm5 = zmm5[0,0,0,1,4,4,4,5] ; AVX512BW-NEXT: vmovdqu16 %zmm5, %zmm14 {%k1} ; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm5 = mem[0,1,2,3,0,1,2,3] ; AVX512BW-NEXT: vpmovzxbw {{.*#+}} xmm7 = 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-NEXT: vpermi2w %ymm7, %ymm14, %ymm13 ; AVX512BW-NEXT: vpshufd {{.*#+}} xmm7 = xmm5[2,1,2,3] ; AVX512BW-NEXT: vpmovzxbw {{.*#+}} xmm7 = xmm7[0],zero,xmm7[1],zero,xmm7[2],zero,xmm7[3],zero,xmm7[4],zero,xmm7[5],zero,xmm7[6],zero,xmm7[7],zero ; AVX512BW-NEXT: vpermt2w %zmm7, %zmm8, %zmm14 ; AVX512BW-NEXT: vinserti64x4 $1, %ymm14, %zmm13, %zmm7 ; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = mem[0,1,2,3,0,1,2,3] ; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm13 = xmm8[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7] ; AVX512BW-NEXT: vpshufd {{.*#+}} xmm14 = xmm8[2,1,2,3] ; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm14 = xmm14[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7] ; AVX512BW-NEXT: vpermt2w %zmm14, %zmm20, %zmm13 ; AVX512BW-NEXT: vmovdqu8 %zmm13, %zmm7 {%k2} ; AVX512BW-NEXT: vmovdqa (%rcx), %ymm13 ; AVX512BW-NEXT: vmovdqa (%rdx), %ymm14 ; AVX512BW-NEXT: vpunpcklbw {{.*#+}} ymm19 = ymm14[0],ymm13[0],ymm14[1],ymm13[1],ymm14[2],ymm13[2],ymm14[3],ymm13[3],ymm14[4],ymm13[4],ymm14[5],ymm13[5],ymm14[6],ymm13[6],ymm14[7],ymm13[7],ymm14[16],ymm13[16],ymm14[17],ymm13[17],ymm14[18],ymm13[18],ymm14[19],ymm13[19],ymm14[20],ymm13[20],ymm14[21],ymm13[21],ymm14[22],ymm13[22],ymm14[23],ymm13[23] ; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm20 = xmm17[8],xmm16[8],xmm17[9],xmm16[9],xmm17[10],xmm16[10],xmm17[11],xmm16[11],xmm17[12],xmm16[12],xmm17[13],xmm16[13],xmm17[14],xmm16[14],xmm17[15],xmm16[15] ; AVX512BW-NEXT: vpmovsxbw {{.*#+}} zmm21 = [5,4,3,6,5,4,3,6,5,4,3,6,7,7,7,7,41,40,43,42,41,40,43,42,41,40,43,42,45,44,47,46] ; AVX512BW-NEXT: vpermt2w %zmm19, %zmm21, %zmm20 ; AVX512BW-NEXT: vmovdqa64 (%rsi), %ymm16 ; AVX512BW-NEXT: vmovdqa64 (%rdi), %ymm17 ; AVX512BW-NEXT: vpunpcklbw {{.*#+}} ymm19 = 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-NEXT: vpunpckhbw {{.*#+}} xmm15 = xmm18[8],xmm15[8],xmm18[9],xmm15[9],xmm18[10],xmm15[10],xmm18[11],xmm15[11],xmm18[12],xmm15[12],xmm18[13],xmm15[13],xmm18[14],xmm15[14],xmm18[15],xmm15[15] ; AVX512BW-NEXT: vpmovsxbw {{.*#+}} zmm18 = [4,3,6,5,4,3,6,5,4,3,6,5,7,7,7,7,40,43,42,41,40,43,42,41,40,43,42,41,44,45,46,45] ; AVX512BW-NEXT: vpermt2w %zmm19, %zmm18, %zmm15 ; AVX512BW-NEXT: movl $613566756, %r10d # imm = 0x24924924 ; AVX512BW-NEXT: kmovd %r10d, %k2 ; AVX512BW-NEXT: vmovdqu16 %zmm20, %zmm15 {%k2} ; AVX512BW-NEXT: vextracti64x4 $1, %zmm15, %ymm19 ; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} ymm20 = [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-NEXT: # ymm20 = mem[0,1,2,3,0,1,2,3] ; AVX512BW-NEXT: vpshufb %ymm20, %ymm5, %ymm22 ; AVX512BW-NEXT: vpermq {{.*#+}} ymm22 = ymm22[2,2,2,3] ; AVX512BW-NEXT: movw $18724, %r10w # imm = 0x4924 ; AVX512BW-NEXT: kmovd %r10d, %k3 ; AVX512BW-NEXT: vmovdqu16 %ymm22, %ymm19 {%k3} ; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm22 = xmm5[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15] ; AVX512BW-NEXT: vpmovsxbw {{.*#+}} ymm23 = [18,1,2,19,4,5,20,7,8,21,10,11,22,13,14,23] ; AVX512BW-NEXT: vpermt2w %ymm22, %ymm23, %ymm15 ; AVX512BW-NEXT: vinserti64x4 $1, %ymm19, %zmm15, %zmm15 ; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} ymm19 = [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-NEXT: # ymm19 = mem[0,1,2,3,0,1,2,3] ; AVX512BW-NEXT: vpshufb %ymm19, %ymm8, %ymm22 ; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm24 = xmm8[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15] ; AVX512BW-NEXT: vpmovsxbw {{.*#+}} zmm25 = [2,5,4,3,2,5,4,3,2,5,4,3,6,5,6,7,40,41,42,43,40,41,42,43,40,41,42,43,44,45,46,47] ; AVX512BW-NEXT: vpermt2w %zmm22, %zmm25, %zmm24 ; AVX512BW-NEXT: movabsq $2342443691899625602, %r10 # imm = 0x2082082082082082 ; AVX512BW-NEXT: kmovq %r10, %k4 ; AVX512BW-NEXT: vmovdqu8 %zmm24, %zmm15 {%k4} ; AVX512BW-NEXT: vmovdqa64 32(%rcx), %ymm22 ; AVX512BW-NEXT: vmovdqa64 32(%rdx), %ymm24 ; AVX512BW-NEXT: vpunpcklbw {{.*#+}} ymm26 = ymm24[0],ymm22[0],ymm24[1],ymm22[1],ymm24[2],ymm22[2],ymm24[3],ymm22[3],ymm24[4],ymm22[4],ymm24[5],ymm22[5],ymm24[6],ymm22[6],ymm24[7],ymm22[7],ymm24[16],ymm22[16],ymm24[17],ymm22[17],ymm24[18],ymm22[18],ymm24[19],ymm22[19],ymm24[20],ymm22[20],ymm24[21],ymm22[21],ymm24[22],ymm22[22],ymm24[23],ymm22[23] ; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm10 = 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] ; AVX512BW-NEXT: vpermt2w %zmm26, %zmm21, %zmm10 ; AVX512BW-NEXT: vmovdqa 32(%rsi), %ymm11 ; AVX512BW-NEXT: vmovdqa64 32(%rdi), %ymm21 ; AVX512BW-NEXT: vpunpcklbw {{.*#+}} ymm26 = ymm21[0],ymm11[0],ymm21[1],ymm11[1],ymm21[2],ymm11[2],ymm21[3],ymm11[3],ymm21[4],ymm11[4],ymm21[5],ymm11[5],ymm21[6],ymm11[6],ymm21[7],ymm11[7],ymm21[16],ymm11[16],ymm21[17],ymm11[17],ymm21[18],ymm11[18],ymm21[19],ymm11[19],ymm21[20],ymm11[20],ymm21[21],ymm11[21],ymm21[22],ymm11[22],ymm21[23],ymm11[23] ; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm4 = xmm6[8],xmm4[8],xmm6[9],xmm4[9],xmm6[10],xmm4[10],xmm6[11],xmm4[11],xmm6[12],xmm4[12],xmm6[13],xmm4[13],xmm6[14],xmm4[14],xmm6[15],xmm4[15] ; AVX512BW-NEXT: vpermt2w %zmm26, %zmm18, %zmm4 ; AVX512BW-NEXT: vmovdqu16 %zmm10, %zmm4 {%k2} ; AVX512BW-NEXT: vextracti64x4 $1, %zmm4, %ymm6 ; AVX512BW-NEXT: vmovdqa 32(%r8), %ymm10 ; AVX512BW-NEXT: vpshufb %ymm20, %ymm10, %ymm10 ; AVX512BW-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,2,2,3] ; AVX512BW-NEXT: vmovdqu16 %ymm10, %ymm6 {%k3} ; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm9 = xmm9[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15] ; AVX512BW-NEXT: vpermt2w %ymm9, %ymm23, %ymm4 ; AVX512BW-NEXT: vinserti64x4 $1, %ymm6, %zmm4, %zmm4 ; AVX512BW-NEXT: vmovdqa 32(%r9), %ymm6 ; AVX512BW-NEXT: vpshufb %ymm19, %ymm6, %ymm6 ; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm9 = xmm12[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15] ; AVX512BW-NEXT: vpermt2w %zmm6, %zmm25, %zmm9 ; AVX512BW-NEXT: vmovdqu8 %zmm9, %zmm4 {%k4} ; AVX512BW-NEXT: vpbroadcastq {{.*#+}} ymm6 = [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-NEXT: vpshufb %ymm6, %ymm11, %ymm9 ; AVX512BW-NEXT: vpshufb %ymm6, %ymm21, %ymm10 ; AVX512BW-NEXT: vpunpcklbw {{.*#+}} ymm9 = ymm10[0],ymm9[0],ymm10[1],ymm9[1],ymm10[2],ymm9[2],ymm10[3],ymm9[3],ymm10[4],ymm9[4],ymm10[5],ymm9[5],ymm10[6],ymm9[6],ymm10[7],ymm9[7],ymm10[16],ymm9[16],ymm10[17],ymm9[17],ymm10[18],ymm9[18],ymm10[19],ymm9[19],ymm10[20],ymm9[20],ymm10[21],ymm9[21],ymm10[22],ymm9[22],ymm10[23],ymm9[23] ; AVX512BW-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,2,3] ; AVX512BW-NEXT: vpunpckhbw {{.*#+}} ymm10 = ymm21[8],ymm11[8],ymm21[9],ymm11[9],ymm21[10],ymm11[10],ymm21[11],ymm11[11],ymm21[12],ymm11[12],ymm21[13],ymm11[13],ymm21[14],ymm11[14],ymm21[15],ymm11[15],ymm21[24],ymm11[24],ymm21[25],ymm11[25],ymm21[26],ymm11[26],ymm21[27],ymm11[27],ymm21[28],ymm11[28],ymm21[29],ymm11[29],ymm21[30],ymm11[30],ymm21[31],ymm11[31] ; AVX512BW-NEXT: vpmovsxbw {{.*#+}} ymm11 = [12,11,14,13,12,11,14,13,12,11,14,13,15,15,15,15] ; AVX512BW-NEXT: vpermw %ymm10, %ymm11, %ymm10 ; AVX512BW-NEXT: vinserti64x4 $1, %ymm10, %zmm9, %zmm9 ; AVX512BW-NEXT: vpshufb %ymm0, %ymm22, %ymm10 ; AVX512BW-NEXT: vpshufb %ymm0, %ymm24, %ymm12 ; AVX512BW-NEXT: vpunpcklbw {{.*#+}} ymm10 = ymm12[0],ymm10[0],ymm12[1],ymm10[1],ymm12[2],ymm10[2],ymm12[3],ymm10[3],ymm12[4],ymm10[4],ymm12[5],ymm10[5],ymm12[6],ymm10[6],ymm12[7],ymm10[7],ymm12[16],ymm10[16],ymm12[17],ymm10[17],ymm12[18],ymm10[18],ymm12[19],ymm10[19],ymm12[20],ymm10[20],ymm12[21],ymm10[21],ymm12[22],ymm10[22],ymm12[23],ymm10[23] ; AVX512BW-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,2,2,3] ; AVX512BW-NEXT: vpunpckhbw {{.*#+}} ymm12 = ymm24[8],ymm22[8],ymm24[9],ymm22[9],ymm24[10],ymm22[10],ymm24[11],ymm22[11],ymm24[12],ymm22[12],ymm24[13],ymm22[13],ymm24[14],ymm22[14],ymm24[15],ymm22[15],ymm24[24],ymm22[24],ymm24[25],ymm22[25],ymm24[26],ymm22[26],ymm24[27],ymm22[27],ymm24[28],ymm22[28],ymm24[29],ymm22[29],ymm24[30],ymm22[30],ymm24[31],ymm22[31] ; AVX512BW-NEXT: vpmovsxbw {{.*#+}} ymm18 = [13,12,11,14,13,12,11,14,13,12,11,14,15,15,15,15] ; AVX512BW-NEXT: vpermw %ymm12, %ymm18, %ymm12 ; AVX512BW-NEXT: vinserti64x4 $1, %ymm12, %zmm10, %zmm10 ; AVX512BW-NEXT: vmovdqu16 %zmm9, %zmm10 {%k2} ; AVX512BW-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm2[4,5,6,7,4,5,6,7] ; AVX512BW-NEXT: vpmovsxbw {{.*#+}} zmm9 = [6,5,8,7,0,9,0,0,6,5,8,7,0,9,0,0,10,13,12,11,14,0,0,15,10,13,12,11,14,0,0,15] ; AVX512BW-NEXT: vpshufb %zmm9, %zmm2, %zmm2 ; AVX512BW-NEXT: vpermq {{.*#+}} zmm2 = zmm2[2,2,2,3,6,6,6,7] ; AVX512BW-NEXT: vmovdqu16 %zmm2, %zmm10 {%k1} ; AVX512BW-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm1[4,5,6,7,4,5,6,7] ; AVX512BW-NEXT: vpshufb {{.*#+}} zmm1 = zmm1[u,6,u,5,u,8,u,7,u,u,u,9,u,u,u,u,u,22,u,21,u,24,u,23,u,u,u,25,u,u,u,u,u,42,u,45,u,44,u,43,u,46,u,u,u,u,u,47,u,58,u,61,u,60,u,59,u,62,u,u,u,u,u,63] ; AVX512BW-NEXT: vpermq {{.*#+}} zmm1 = zmm1[2,2,2,3,6,6,6,7] ; AVX512BW-NEXT: movabsq $-9076969306111049208, %rcx # imm = 0x8208208208208208 ; AVX512BW-NEXT: kmovq %rcx, %k3 ; AVX512BW-NEXT: vmovdqu8 %zmm1, %zmm10 {%k3} ; AVX512BW-NEXT: vpshufb %ymm6, %ymm16, %ymm1 ; AVX512BW-NEXT: vpshufb %ymm6, %ymm17, %ymm2 ; AVX512BW-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-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3] ; AVX512BW-NEXT: vpunpckhbw {{.*#+}} ymm2 = 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-NEXT: vpermw %ymm2, %ymm11, %ymm2 ; AVX512BW-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm1 ; AVX512BW-NEXT: vpshufb %ymm0, %ymm13, %ymm2 ; AVX512BW-NEXT: vpshufb %ymm0, %ymm14, %ymm0 ; AVX512BW-NEXT: vpunpcklbw {{.*#+}} ymm0 = ymm0[0],ymm2[0],ymm0[1],ymm2[1],ymm0[2],ymm2[2],ymm0[3],ymm2[3],ymm0[4],ymm2[4],ymm0[5],ymm2[5],ymm0[6],ymm2[6],ymm0[7],ymm2[7],ymm0[16],ymm2[16],ymm0[17],ymm2[17],ymm0[18],ymm2[18],ymm0[19],ymm2[19],ymm0[20],ymm2[20],ymm0[21],ymm2[21],ymm0[22],ymm2[22],ymm0[23],ymm2[23] ; AVX512BW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3] ; AVX512BW-NEXT: vpunpckhbw {{.*#+}} ymm2 = ymm14[8],ymm13[8],ymm14[9],ymm13[9],ymm14[10],ymm13[10],ymm14[11],ymm13[11],ymm14[12],ymm13[12],ymm14[13],ymm13[13],ymm14[14],ymm13[14],ymm14[15],ymm13[15],ymm14[24],ymm13[24],ymm14[25],ymm13[25],ymm14[26],ymm13[26],ymm14[27],ymm13[27],ymm14[28],ymm13[28],ymm14[29],ymm13[29],ymm14[30],ymm13[30],ymm14[31],ymm13[31] ; AVX512BW-NEXT: vpermw %ymm2, %ymm18, %ymm2 ; AVX512BW-NEXT: vinserti64x4 $1, %ymm2, %zmm0, %zmm0 ; AVX512BW-NEXT: vmovdqu16 %zmm1, %zmm0 {%k2} ; AVX512BW-NEXT: vpshufb %zmm9, %zmm5, %zmm1 ; AVX512BW-NEXT: vpermq {{.*#+}} zmm1 = zmm1[2,2,2,3,6,6,6,7] ; AVX512BW-NEXT: vmovdqu16 %zmm1, %zmm0 {%k1} ; AVX512BW-NEXT: vpshufb {{.*#+}} zmm1 = zmm8[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,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,58,u,61,u,60,u,59,u,62,u,u,u,u,u,63] ; AVX512BW-NEXT: vpermq {{.*#+}} zmm1 = zmm1[2,2,2,3,6,6,6,7] ; AVX512BW-NEXT: vmovdqu8 %zmm1, %zmm0 {%k3} ; AVX512BW-NEXT: vmovdqa64 %zmm0, 128(%rax) ; AVX512BW-NEXT: vmovdqa64 %zmm10, 320(%rax) ; AVX512BW-NEXT: vmovdqa64 %zmm4, 256(%rax) ; AVX512BW-NEXT: vmovdqa64 %zmm15, 64(%rax) ; AVX512BW-NEXT: vmovdqa64 %zmm7, (%rax) ; AVX512BW-NEXT: vmovdqa64 %zmm3, 192(%rax) ; AVX512BW-NEXT: vzeroupper ; AVX512BW-NEXT: retq ; ; AVX512BW-FCP-LABEL: store_i8_stride6_vf64: ; AVX512BW-FCP: # %bb.0: ; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512BW-FCP-NEXT: vmovdqa64 (%r8), %zmm2 ; AVX512BW-FCP-NEXT: vmovdqa64 (%r9), %zmm1 ; AVX512BW-FCP-NEXT: vpbroadcastq {{.*#+}} ymm0 = [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-FCP-NEXT: vmovdqa (%rcx), %xmm15 ; AVX512BW-FCP-NEXT: vmovdqa 32(%rcx), %xmm6 ; AVX512BW-FCP-NEXT: vpshufb %xmm0, %xmm6, %xmm3 ; AVX512BW-FCP-NEXT: vmovdqa64 (%rdx), %xmm16 ; AVX512BW-FCP-NEXT: vmovdqa 32(%rdx), %xmm8 ; AVX512BW-FCP-NEXT: vpshufb %xmm0, %xmm8, %xmm4 ; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm3 = 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] ; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm4 = xmm8[0],xmm6[0],xmm8[1],xmm6[1],xmm8[2],xmm6[2],xmm8[3],xmm6[3],xmm8[4],xmm6[4],xmm8[5],xmm6[5],xmm8[6],xmm6[6],xmm8[7],xmm6[7] ; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm5 = [1,0,3,2,1,0,3,2,1,0,3,2,5,4,7,6,32,33,34,35,32,33,34,35,32,33,34,35,36,37,38,39] ; AVX512BW-FCP-NEXT: vpermt2w %zmm3, %zmm5, %zmm4 ; AVX512BW-FCP-NEXT: vmovdqa64 (%rsi), %xmm17 ; AVX512BW-FCP-NEXT: vmovdqa 32(%rsi), %xmm9 ; AVX512BW-FCP-NEXT: vpbroadcastq {{.*#+}} xmm7 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0] ; AVX512BW-FCP-NEXT: vpshufb %xmm7, %xmm9, %xmm3 ; AVX512BW-FCP-NEXT: vmovdqa64 (%rdi), %xmm18 ; AVX512BW-FCP-NEXT: vmovdqa 32(%rdi), %xmm10 ; AVX512BW-FCP-NEXT: vpshufb %xmm7, %xmm10, %xmm11 ; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm11[8],xmm3[8],xmm11[9],xmm3[9],xmm11[10],xmm3[10],xmm11[11],xmm3[11],xmm11[12],xmm3[12],xmm11[13],xmm3[13],xmm11[14],xmm3[14],xmm11[15],xmm3[15] ; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm12 = 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-FCP-NEXT: vpmovsxbw {{.*#+}} zmm13 = [0,3,2,1,0,3,2,1,0,3,2,1,4,5,6,5,32,33,34,35,32,33,34,35,32,33,34,35,36,37,38,39] ; AVX512BW-FCP-NEXT: vpermt2w %zmm3, %zmm13, %zmm12 ; AVX512BW-FCP-NEXT: movl $-1840700270, %r10d # imm = 0x92492492 ; AVX512BW-FCP-NEXT: kmovd %r10d, %k1 ; AVX512BW-FCP-NEXT: vmovdqu16 %zmm4, %zmm12 {%k1} ; AVX512BW-FCP-NEXT: vmovdqa 32(%r8), %xmm11 ; AVX512BW-FCP-NEXT: vpmovzxbw {{.*#+}} xmm3 = xmm11[0],zero,xmm11[1],zero,xmm11[2],zero,xmm11[3],zero,xmm11[4],zero,xmm11[5],zero,xmm11[6],zero,xmm11[7],zero ; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm14 = [0,1,16,3,4,17,6,7,18,9,10,19,12,13,20,15] ; AVX512BW-FCP-NEXT: vmovdqa %ymm12, %ymm4 ; AVX512BW-FCP-NEXT: vpermt2w %ymm3, %ymm14, %ymm4 ; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm19 = [16,37,18,19,38,21,22,39,24,25,32,27,28,33,30,31] ; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} xmm20 = [8,9,0,0,0,5,6,7] ; AVX512BW-FCP-NEXT: vpshufb %xmm20, %xmm11, %xmm3 ; AVX512BW-FCP-NEXT: vpermt2w %zmm3, %zmm19, %zmm12 ; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm12, %zmm4, %zmm3 ; AVX512BW-FCP-NEXT: vmovdqa 32(%r9), %xmm12 ; AVX512BW-FCP-NEXT: vmovdqa64 {{.*#+}} xmm21 = [u,8,u,9,u,10,u,11,u,4,u,5,u,6,u,7] ; AVX512BW-FCP-NEXT: vpshufb %xmm21, %xmm12, %xmm4 ; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm22 = xmm12[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7] ; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm23 = [2,1,0,3,2,1,0,3,2,1,0,3,4,4,4,4,38,37,32,39,38,37,32,39,38,37,32,39,33,33,33,33] ; AVX512BW-FCP-NEXT: vpermt2w %zmm4, %zmm23, %zmm22 ; AVX512BW-FCP-NEXT: movabsq $585610922974906400, %r10 # imm = 0x820820820820820 ; AVX512BW-FCP-NEXT: kmovq %r10, %k2 ; AVX512BW-FCP-NEXT: vmovdqu8 %zmm22, %zmm3 {%k2} ; AVX512BW-FCP-NEXT: vpshufb %xmm0, %xmm15, %xmm4 ; AVX512BW-FCP-NEXT: vpshufb %xmm0, %xmm16, %xmm22 ; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm4 = xmm22[0],xmm4[0],xmm22[1],xmm4[1],xmm22[2],xmm4[2],xmm22[3],xmm4[3],xmm22[4],xmm4[4],xmm22[5],xmm4[5],xmm22[6],xmm4[6],xmm22[7],xmm4[7] ; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm22 = xmm16[0],xmm15[0],xmm16[1],xmm15[1],xmm16[2],xmm15[2],xmm16[3],xmm15[3],xmm16[4],xmm15[4],xmm16[5],xmm15[5],xmm16[6],xmm15[6],xmm16[7],xmm15[7] ; AVX512BW-FCP-NEXT: vpermt2w %zmm4, %zmm5, %zmm22 ; AVX512BW-FCP-NEXT: vpshufb %xmm7, %xmm17, %xmm4 ; AVX512BW-FCP-NEXT: vpshufb %xmm7, %xmm18, %xmm5 ; AVX512BW-FCP-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] ; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm5 = 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-FCP-NEXT: vpermt2w %zmm4, %zmm13, %zmm5 ; AVX512BW-FCP-NEXT: vmovdqu16 %zmm22, %zmm5 {%k1} ; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm4 = mem[0,1,2,3,0,1,2,3] ; AVX512BW-FCP-NEXT: vpmovzxbw {{.*#+}} xmm7 = xmm4[0],zero,xmm4[1],zero,xmm4[2],zero,xmm4[3],zero,xmm4[4],zero,xmm4[5],zero,xmm4[6],zero,xmm4[7],zero ; AVX512BW-FCP-NEXT: vpermi2w %ymm7, %ymm5, %ymm14 ; AVX512BW-FCP-NEXT: vpshufb %xmm20, %xmm4, %xmm7 ; AVX512BW-FCP-NEXT: vpermt2w %zmm7, %zmm19, %zmm5 ; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm5, %zmm14, %zmm5 ; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = mem[0,1,2,3,0,1,2,3] ; AVX512BW-FCP-NEXT: vpshufb %xmm21, %xmm7, %xmm13 ; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm14 = xmm7[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7] ; AVX512BW-FCP-NEXT: vpermt2w %zmm13, %zmm23, %zmm14 ; AVX512BW-FCP-NEXT: vmovdqu8 %zmm14, %zmm5 {%k2} ; AVX512BW-FCP-NEXT: vmovdqa (%rcx), %ymm13 ; AVX512BW-FCP-NEXT: vmovdqa (%rdx), %ymm14 ; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} ymm19 = ymm14[0],ymm13[0],ymm14[1],ymm13[1],ymm14[2],ymm13[2],ymm14[3],ymm13[3],ymm14[4],ymm13[4],ymm14[5],ymm13[5],ymm14[6],ymm13[6],ymm14[7],ymm13[7],ymm14[16],ymm13[16],ymm14[17],ymm13[17],ymm14[18],ymm13[18],ymm14[19],ymm13[19],ymm14[20],ymm13[20],ymm14[21],ymm13[21],ymm14[22],ymm13[22],ymm14[23],ymm13[23] ; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm20 = xmm16[8],xmm15[8],xmm16[9],xmm15[9],xmm16[10],xmm15[10],xmm16[11],xmm15[11],xmm16[12],xmm15[12],xmm16[13],xmm15[13],xmm16[14],xmm15[14],xmm16[15],xmm15[15] ; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm21 = [5,4,3,6,5,4,3,6,5,4,3,6,7,7,7,7,41,40,43,42,41,40,43,42,41,40,43,42,45,44,47,46] ; AVX512BW-FCP-NEXT: vpermt2w %zmm19, %zmm21, %zmm20 ; AVX512BW-FCP-NEXT: vmovdqa (%rsi), %ymm15 ; AVX512BW-FCP-NEXT: vmovdqa64 (%rdi), %ymm16 ; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} ymm19 = ymm16[0],ymm15[0],ymm16[1],ymm15[1],ymm16[2],ymm15[2],ymm16[3],ymm15[3],ymm16[4],ymm15[4],ymm16[5],ymm15[5],ymm16[6],ymm15[6],ymm16[7],ymm15[7],ymm16[16],ymm15[16],ymm16[17],ymm15[17],ymm16[18],ymm15[18],ymm16[19],ymm15[19],ymm16[20],ymm15[20],ymm16[21],ymm15[21],ymm16[22],ymm15[22],ymm16[23],ymm15[23] ; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm17 = xmm18[8],xmm17[8],xmm18[9],xmm17[9],xmm18[10],xmm17[10],xmm18[11],xmm17[11],xmm18[12],xmm17[12],xmm18[13],xmm17[13],xmm18[14],xmm17[14],xmm18[15],xmm17[15] ; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm18 = [4,3,6,5,4,3,6,5,4,3,6,5,7,7,7,7,40,43,42,41,40,43,42,41,40,43,42,41,44,45,46,45] ; AVX512BW-FCP-NEXT: vpermt2w %zmm19, %zmm18, %zmm17 ; AVX512BW-FCP-NEXT: movl $613566756, %r10d # imm = 0x24924924 ; AVX512BW-FCP-NEXT: kmovd %r10d, %k2 ; AVX512BW-FCP-NEXT: vmovdqu16 %zmm20, %zmm17 {%k2} ; AVX512BW-FCP-NEXT: vextracti64x4 $1, %zmm17, %ymm19 ; AVX512BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} ymm20 = [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-FCP-NEXT: # ymm20 = mem[0,1,2,3,0,1,2,3] ; AVX512BW-FCP-NEXT: vpshufb %ymm20, %ymm4, %ymm22 ; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm22 = ymm22[2,2,2,3] ; AVX512BW-FCP-NEXT: movw $18724, %r10w # imm = 0x4924 ; AVX512BW-FCP-NEXT: kmovd %r10d, %k3 ; AVX512BW-FCP-NEXT: vmovdqu16 %ymm22, %ymm19 {%k3} ; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm22 = xmm4[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15] ; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm23 = [18,1,2,19,4,5,20,7,8,21,10,11,22,13,14,23] ; AVX512BW-FCP-NEXT: vpermt2w %ymm22, %ymm23, %ymm17 ; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm19, %zmm17, %zmm17 ; AVX512BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} ymm19 = [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-FCP-NEXT: # ymm19 = mem[0,1,2,3,0,1,2,3] ; AVX512BW-FCP-NEXT: vpshufb %ymm19, %ymm7, %ymm22 ; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm24 = xmm7[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15] ; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm25 = [2,5,4,3,2,5,4,3,2,5,4,3,6,5,6,7,40,41,42,43,40,41,42,43,40,41,42,43,44,45,46,47] ; AVX512BW-FCP-NEXT: vpermt2w %zmm22, %zmm25, %zmm24 ; AVX512BW-FCP-NEXT: movabsq $2342443691899625602, %r10 # imm = 0x2082082082082082 ; AVX512BW-FCP-NEXT: kmovq %r10, %k4 ; AVX512BW-FCP-NEXT: vmovdqu8 %zmm24, %zmm17 {%k4} ; AVX512BW-FCP-NEXT: vmovdqa64 32(%rcx), %ymm22 ; AVX512BW-FCP-NEXT: vmovdqa64 32(%rdx), %ymm24 ; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} ymm26 = ymm24[0],ymm22[0],ymm24[1],ymm22[1],ymm24[2],ymm22[2],ymm24[3],ymm22[3],ymm24[4],ymm22[4],ymm24[5],ymm22[5],ymm24[6],ymm22[6],ymm24[7],ymm22[7],ymm24[16],ymm22[16],ymm24[17],ymm22[17],ymm24[18],ymm22[18],ymm24[19],ymm22[19],ymm24[20],ymm22[20],ymm24[21],ymm22[21],ymm24[22],ymm22[22],ymm24[23],ymm22[23] ; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm6 = xmm8[8],xmm6[8],xmm8[9],xmm6[9],xmm8[10],xmm6[10],xmm8[11],xmm6[11],xmm8[12],xmm6[12],xmm8[13],xmm6[13],xmm8[14],xmm6[14],xmm8[15],xmm6[15] ; AVX512BW-FCP-NEXT: vpermt2w %zmm26, %zmm21, %zmm6 ; AVX512BW-FCP-NEXT: vmovdqa 32(%rsi), %ymm8 ; AVX512BW-FCP-NEXT: vmovdqa64 32(%rdi), %ymm21 ; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} ymm26 = ymm21[0],ymm8[0],ymm21[1],ymm8[1],ymm21[2],ymm8[2],ymm21[3],ymm8[3],ymm21[4],ymm8[4],ymm21[5],ymm8[5],ymm21[6],ymm8[6],ymm21[7],ymm8[7],ymm21[16],ymm8[16],ymm21[17],ymm8[17],ymm21[18],ymm8[18],ymm21[19],ymm8[19],ymm21[20],ymm8[20],ymm21[21],ymm8[21],ymm21[22],ymm8[22],ymm21[23],ymm8[23] ; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm9 = 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-FCP-NEXT: vpermt2w %zmm26, %zmm18, %zmm9 ; AVX512BW-FCP-NEXT: vmovdqu16 %zmm6, %zmm9 {%k2} ; AVX512BW-FCP-NEXT: vextracti64x4 $1, %zmm9, %ymm6 ; AVX512BW-FCP-NEXT: vmovdqa 32(%r8), %ymm10 ; AVX512BW-FCP-NEXT: vpshufb %ymm20, %ymm10, %ymm10 ; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,2,2,3] ; AVX512BW-FCP-NEXT: vmovdqu16 %ymm10, %ymm6 {%k3} ; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm10 = xmm11[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15] ; AVX512BW-FCP-NEXT: vpermt2w %ymm10, %ymm23, %ymm9 ; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm6, %zmm9, %zmm6 ; AVX512BW-FCP-NEXT: vmovdqa 32(%r9), %ymm9 ; AVX512BW-FCP-NEXT: vpshufb %ymm19, %ymm9, %ymm9 ; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm10 = xmm12[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15] ; AVX512BW-FCP-NEXT: vpermt2w %zmm9, %zmm25, %zmm10 ; AVX512BW-FCP-NEXT: vmovdqu8 %zmm10, %zmm6 {%k4} ; AVX512BW-FCP-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-FCP-NEXT: vpshufb %ymm9, %ymm8, %ymm10 ; AVX512BW-FCP-NEXT: vpshufb %ymm9, %ymm21, %ymm11 ; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} ymm10 = ymm11[0],ymm10[0],ymm11[1],ymm10[1],ymm11[2],ymm10[2],ymm11[3],ymm10[3],ymm11[4],ymm10[4],ymm11[5],ymm10[5],ymm11[6],ymm10[6],ymm11[7],ymm10[7],ymm11[16],ymm10[16],ymm11[17],ymm10[17],ymm11[18],ymm10[18],ymm11[19],ymm10[19],ymm11[20],ymm10[20],ymm11[21],ymm10[21],ymm11[22],ymm10[22],ymm11[23],ymm10[23] ; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,2,2,3] ; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} ymm8 = ymm21[8],ymm8[8],ymm21[9],ymm8[9],ymm21[10],ymm8[10],ymm21[11],ymm8[11],ymm21[12],ymm8[12],ymm21[13],ymm8[13],ymm21[14],ymm8[14],ymm21[15],ymm8[15],ymm21[24],ymm8[24],ymm21[25],ymm8[25],ymm21[26],ymm8[26],ymm21[27],ymm8[27],ymm21[28],ymm8[28],ymm21[29],ymm8[29],ymm21[30],ymm8[30],ymm21[31],ymm8[31] ; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm11 = [12,11,14,13,12,11,14,13,12,11,14,13,15,15,15,15] ; AVX512BW-FCP-NEXT: vpermw %ymm8, %ymm11, %ymm8 ; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm8, %zmm10, %zmm8 ; AVX512BW-FCP-NEXT: vpshufb %ymm0, %ymm22, %ymm10 ; AVX512BW-FCP-NEXT: vpshufb %ymm0, %ymm24, %ymm12 ; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} ymm10 = ymm12[0],ymm10[0],ymm12[1],ymm10[1],ymm12[2],ymm10[2],ymm12[3],ymm10[3],ymm12[4],ymm10[4],ymm12[5],ymm10[5],ymm12[6],ymm10[6],ymm12[7],ymm10[7],ymm12[16],ymm10[16],ymm12[17],ymm10[17],ymm12[18],ymm10[18],ymm12[19],ymm10[19],ymm12[20],ymm10[20],ymm12[21],ymm10[21],ymm12[22],ymm10[22],ymm12[23],ymm10[23] ; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,2,2,3] ; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} ymm12 = ymm24[8],ymm22[8],ymm24[9],ymm22[9],ymm24[10],ymm22[10],ymm24[11],ymm22[11],ymm24[12],ymm22[12],ymm24[13],ymm22[13],ymm24[14],ymm22[14],ymm24[15],ymm22[15],ymm24[24],ymm22[24],ymm24[25],ymm22[25],ymm24[26],ymm22[26],ymm24[27],ymm22[27],ymm24[28],ymm22[28],ymm24[29],ymm22[29],ymm24[30],ymm22[30],ymm24[31],ymm22[31] ; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm18 = [13,12,11,14,13,12,11,14,13,12,11,14,15,15,15,15] ; AVX512BW-FCP-NEXT: vpermw %ymm12, %ymm18, %ymm12 ; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm12, %zmm10, %zmm10 ; AVX512BW-FCP-NEXT: vmovdqu16 %zmm8, %zmm10 {%k2} ; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm2[4,5,6,7,4,5,6,7] ; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm8 = [6,5,8,7,0,9,0,0,6,5,8,7,0,9,0,0,10,13,12,11,14,0,0,15,10,13,12,11,14,0,0,15] ; AVX512BW-FCP-NEXT: vpshufb %zmm8, %zmm2, %zmm2 ; AVX512BW-FCP-NEXT: vpermq {{.*#+}} zmm2 = zmm2[2,2,2,3,6,6,6,7] ; AVX512BW-FCP-NEXT: vmovdqu16 %zmm2, %zmm10 {%k1} ; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm1[4,5,6,7,4,5,6,7] ; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} zmm1 = zmm1[u,6,u,5,u,8,u,7,u,u,u,9,u,u,u,u,u,22,u,21,u,24,u,23,u,u,u,25,u,u,u,u,u,42,u,45,u,44,u,43,u,46,u,u,u,u,u,47,u,58,u,61,u,60,u,59,u,62,u,u,u,u,u,63] ; AVX512BW-FCP-NEXT: vpermq {{.*#+}} zmm1 = zmm1[2,2,2,3,6,6,6,7] ; AVX512BW-FCP-NEXT: movabsq $-9076969306111049208, %rcx # imm = 0x8208208208208208 ; AVX512BW-FCP-NEXT: kmovq %rcx, %k3 ; AVX512BW-FCP-NEXT: vmovdqu8 %zmm1, %zmm10 {%k3} ; AVX512BW-FCP-NEXT: vpshufb %ymm9, %ymm15, %ymm1 ; AVX512BW-FCP-NEXT: vpshufb %ymm9, %ymm16, %ymm2 ; AVX512BW-FCP-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-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3] ; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} ymm2 = ymm16[8],ymm15[8],ymm16[9],ymm15[9],ymm16[10],ymm15[10],ymm16[11],ymm15[11],ymm16[12],ymm15[12],ymm16[13],ymm15[13],ymm16[14],ymm15[14],ymm16[15],ymm15[15],ymm16[24],ymm15[24],ymm16[25],ymm15[25],ymm16[26],ymm15[26],ymm16[27],ymm15[27],ymm16[28],ymm15[28],ymm16[29],ymm15[29],ymm16[30],ymm15[30],ymm16[31],ymm15[31] ; AVX512BW-FCP-NEXT: vpermw %ymm2, %ymm11, %ymm2 ; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm1 ; AVX512BW-FCP-NEXT: vpshufb %ymm0, %ymm13, %ymm2 ; AVX512BW-FCP-NEXT: vpshufb %ymm0, %ymm14, %ymm0 ; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} ymm0 = ymm0[0],ymm2[0],ymm0[1],ymm2[1],ymm0[2],ymm2[2],ymm0[3],ymm2[3],ymm0[4],ymm2[4],ymm0[5],ymm2[5],ymm0[6],ymm2[6],ymm0[7],ymm2[7],ymm0[16],ymm2[16],ymm0[17],ymm2[17],ymm0[18],ymm2[18],ymm0[19],ymm2[19],ymm0[20],ymm2[20],ymm0[21],ymm2[21],ymm0[22],ymm2[22],ymm0[23],ymm2[23] ; AVX512BW-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3] ; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} ymm2 = ymm14[8],ymm13[8],ymm14[9],ymm13[9],ymm14[10],ymm13[10],ymm14[11],ymm13[11],ymm14[12],ymm13[12],ymm14[13],ymm13[13],ymm14[14],ymm13[14],ymm14[15],ymm13[15],ymm14[24],ymm13[24],ymm14[25],ymm13[25],ymm14[26],ymm13[26],ymm14[27],ymm13[27],ymm14[28],ymm13[28],ymm14[29],ymm13[29],ymm14[30],ymm13[30],ymm14[31],ymm13[31] ; AVX512BW-FCP-NEXT: vpermw %ymm2, %ymm18, %ymm2 ; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm0, %zmm0 ; AVX512BW-FCP-NEXT: vmovdqu16 %zmm1, %zmm0 {%k2} ; AVX512BW-FCP-NEXT: vpshufb %zmm8, %zmm4, %zmm1 ; AVX512BW-FCP-NEXT: vpermq {{.*#+}} zmm1 = zmm1[2,2,2,3,6,6,6,7] ; AVX512BW-FCP-NEXT: vmovdqu16 %zmm1, %zmm0 {%k1} ; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} zmm1 = zmm7[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,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,58,u,61,u,60,u,59,u,62,u,u,u,u,u,63] ; AVX512BW-FCP-NEXT: vpermq {{.*#+}} zmm1 = zmm1[2,2,2,3,6,6,6,7] ; AVX512BW-FCP-NEXT: vmovdqu8 %zmm1, %zmm0 {%k3} ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, 128(%rax) ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm10, 320(%rax) ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm6, 256(%rax) ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm17, 64(%rax) ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm5, (%rax) ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, 192(%rax) ; AVX512BW-FCP-NEXT: vzeroupper ; AVX512BW-FCP-NEXT: retq ; ; AVX512DQ-BW-LABEL: store_i8_stride6_vf64: ; AVX512DQ-BW: # %bb.0: ; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512DQ-BW-NEXT: vmovdqa64 (%r8), %zmm2 ; AVX512DQ-BW-NEXT: vmovdqa64 (%r9), %zmm1 ; AVX512DQ-BW-NEXT: vmovdqa (%rsi), %xmm15 ; AVX512DQ-BW-NEXT: vmovdqa 32(%rsi), %xmm4 ; AVX512DQ-BW-NEXT: vpbroadcastq {{.*#+}} xmm5 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0] ; AVX512DQ-BW-NEXT: vpshufb %xmm5, %xmm4, %xmm0 ; AVX512DQ-BW-NEXT: vmovdqa64 (%rdi), %xmm18 ; AVX512DQ-BW-NEXT: vmovdqa 32(%rdi), %xmm6 ; AVX512DQ-BW-NEXT: vpshufb %xmm5, %xmm6, %xmm3 ; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm3[8],xmm0[8],xmm3[9],xmm0[9],xmm3[10],xmm0[10],xmm3[11],xmm0[11],xmm3[12],xmm0[12],xmm3[13],xmm0[13],xmm3[14],xmm0[14],xmm3[15],xmm0[15] ; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm6[0],xmm4[0],xmm6[1],xmm4[1],xmm6[2],xmm4[2],xmm6[3],xmm4[3],xmm6[4],xmm4[4],xmm6[5],xmm4[5],xmm6[6],xmm4[6],xmm6[7],xmm4[7] ; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} zmm7 = [0,3,2,1,0,3,2,1,0,3,2,1,4,5,6,5,32,33,34,35,32,33,34,35,32,33,34,35,36,37,38,39] ; AVX512DQ-BW-NEXT: vpermt2w %zmm0, %zmm7, %zmm3 ; AVX512DQ-BW-NEXT: vpbroadcastq {{.*#+}} ymm0 = [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] ; AVX512DQ-BW-NEXT: vmovdqa64 (%rcx), %xmm16 ; AVX512DQ-BW-NEXT: vmovdqa 32(%rcx), %xmm10 ; AVX512DQ-BW-NEXT: vpshufb %xmm0, %xmm10, %xmm8 ; AVX512DQ-BW-NEXT: vmovdqa64 (%rdx), %xmm17 ; AVX512DQ-BW-NEXT: vmovdqa 32(%rdx), %xmm11 ; AVX512DQ-BW-NEXT: vpshufb %xmm0, %xmm11, %xmm9 ; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm8 = xmm9[0],xmm8[0],xmm9[1],xmm8[1],xmm9[2],xmm8[2],xmm9[3],xmm8[3],xmm9[4],xmm8[4],xmm9[5],xmm8[5],xmm9[6],xmm8[6],xmm9[7],xmm8[7] ; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm9 = xmm11[0],xmm10[0],xmm11[1],xmm10[1],xmm11[2],xmm10[2],xmm11[3],xmm10[3],xmm11[4],xmm10[4],xmm11[5],xmm10[5],xmm11[6],xmm10[6],xmm11[7],xmm10[7] ; AVX512DQ-BW-NEXT: vprold $16, %xmm9, %xmm9 ; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm8, %zmm9, %zmm8 ; AVX512DQ-BW-NEXT: vpermq {{.*#+}} zmm8 = zmm8[0,0,0,1,4,4,4,5] ; AVX512DQ-BW-NEXT: movl $-1840700270, %r10d # imm = 0x92492492 ; AVX512DQ-BW-NEXT: kmovd %r10d, %k1 ; AVX512DQ-BW-NEXT: vmovdqu16 %zmm8, %zmm3 {%k1} ; AVX512DQ-BW-NEXT: vmovdqa 32(%r8), %xmm9 ; AVX512DQ-BW-NEXT: vpmovzxbw {{.*#+}} xmm8 = xmm9[0],zero,xmm9[1],zero,xmm9[2],zero,xmm9[3],zero,xmm9[4],zero,xmm9[5],zero,xmm9[6],zero,xmm9[7],zero ; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} ymm13 = [0,1,16,3,4,17,6,7,18,9,10,19,12,13,20,15] ; AVX512DQ-BW-NEXT: vmovdqa %ymm3, %ymm12 ; AVX512DQ-BW-NEXT: vpermt2w %ymm8, %ymm13, %ymm12 ; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} ymm8 = [16,37,18,19,38,21,22,39,24,25,32,27,28,33,30,31] ; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} xmm14 = xmm9[2,1,2,3] ; AVX512DQ-BW-NEXT: vpmovzxbw {{.*#+}} xmm14 = xmm14[0],zero,xmm14[1],zero,xmm14[2],zero,xmm14[3],zero,xmm14[4],zero,xmm14[5],zero,xmm14[6],zero,xmm14[7],zero ; AVX512DQ-BW-NEXT: vpermt2w %zmm14, %zmm8, %zmm3 ; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm3, %zmm12, %zmm3 ; AVX512DQ-BW-NEXT: vmovdqa 32(%r9), %xmm12 ; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm14 = xmm12[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7] ; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} xmm19 = xmm12[2,1,2,3] ; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm19 = xmm19[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7] ; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} zmm20 = [2,1,0,3,2,1,0,3,2,1,0,3,4,4,4,4,38,37,32,39,38,37,32,39,38,37,32,39,33,33,33,33] ; AVX512DQ-BW-NEXT: vpermt2w %zmm19, %zmm20, %zmm14 ; AVX512DQ-BW-NEXT: movabsq $585610922974906400, %r10 # imm = 0x820820820820820 ; AVX512DQ-BW-NEXT: kmovq %r10, %k2 ; AVX512DQ-BW-NEXT: vmovdqu8 %zmm14, %zmm3 {%k2} ; AVX512DQ-BW-NEXT: vpshufb %xmm5, %xmm15, %xmm14 ; AVX512DQ-BW-NEXT: vpshufb %xmm5, %xmm18, %xmm5 ; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm5 = xmm5[8],xmm14[8],xmm5[9],xmm14[9],xmm5[10],xmm14[10],xmm5[11],xmm14[11],xmm5[12],xmm14[12],xmm5[13],xmm14[13],xmm5[14],xmm14[14],xmm5[15],xmm14[15] ; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm14 = xmm18[0],xmm15[0],xmm18[1],xmm15[1],xmm18[2],xmm15[2],xmm18[3],xmm15[3],xmm18[4],xmm15[4],xmm18[5],xmm15[5],xmm18[6],xmm15[6],xmm18[7],xmm15[7] ; AVX512DQ-BW-NEXT: vpermt2w %zmm5, %zmm7, %zmm14 ; AVX512DQ-BW-NEXT: vpshufb %xmm0, %xmm16, %xmm5 ; AVX512DQ-BW-NEXT: vpshufb %xmm0, %xmm17, %xmm7 ; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm5 = xmm7[0],xmm5[0],xmm7[1],xmm5[1],xmm7[2],xmm5[2],xmm7[3],xmm5[3],xmm7[4],xmm5[4],xmm7[5],xmm5[5],xmm7[6],xmm5[6],xmm7[7],xmm5[7] ; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm7 = xmm17[0],xmm16[0],xmm17[1],xmm16[1],xmm17[2],xmm16[2],xmm17[3],xmm16[3],xmm17[4],xmm16[4],xmm17[5],xmm16[5],xmm17[6],xmm16[6],xmm17[7],xmm16[7] ; AVX512DQ-BW-NEXT: vprold $16, %xmm7, %xmm7 ; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm5, %zmm7, %zmm5 ; AVX512DQ-BW-NEXT: vpermq {{.*#+}} zmm5 = zmm5[0,0,0,1,4,4,4,5] ; AVX512DQ-BW-NEXT: vmovdqu16 %zmm5, %zmm14 {%k1} ; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm5 = mem[0,1,2,3,0,1,2,3] ; AVX512DQ-BW-NEXT: vpmovzxbw {{.*#+}} xmm7 = xmm5[0],zero,xmm5[1],zero,xmm5[2],zero,xmm5[3],zero,xmm5[4],zero,xmm5[5],zero,xmm5[6],zero,xmm5[7],zero ; AVX512DQ-BW-NEXT: vpermi2w %ymm7, %ymm14, %ymm13 ; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} xmm7 = xmm5[2,1,2,3] ; AVX512DQ-BW-NEXT: vpmovzxbw {{.*#+}} xmm7 = xmm7[0],zero,xmm7[1],zero,xmm7[2],zero,xmm7[3],zero,xmm7[4],zero,xmm7[5],zero,xmm7[6],zero,xmm7[7],zero ; AVX512DQ-BW-NEXT: vpermt2w %zmm7, %zmm8, %zmm14 ; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm14, %zmm13, %zmm7 ; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = mem[0,1,2,3,0,1,2,3] ; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm13 = xmm8[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7] ; AVX512DQ-BW-NEXT: vpshufd {{.*#+}} xmm14 = xmm8[2,1,2,3] ; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm14 = xmm14[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7] ; AVX512DQ-BW-NEXT: vpermt2w %zmm14, %zmm20, %zmm13 ; AVX512DQ-BW-NEXT: vmovdqu8 %zmm13, %zmm7 {%k2} ; AVX512DQ-BW-NEXT: vmovdqa (%rcx), %ymm13 ; AVX512DQ-BW-NEXT: vmovdqa (%rdx), %ymm14 ; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} ymm19 = ymm14[0],ymm13[0],ymm14[1],ymm13[1],ymm14[2],ymm13[2],ymm14[3],ymm13[3],ymm14[4],ymm13[4],ymm14[5],ymm13[5],ymm14[6],ymm13[6],ymm14[7],ymm13[7],ymm14[16],ymm13[16],ymm14[17],ymm13[17],ymm14[18],ymm13[18],ymm14[19],ymm13[19],ymm14[20],ymm13[20],ymm14[21],ymm13[21],ymm14[22],ymm13[22],ymm14[23],ymm13[23] ; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm20 = xmm17[8],xmm16[8],xmm17[9],xmm16[9],xmm17[10],xmm16[10],xmm17[11],xmm16[11],xmm17[12],xmm16[12],xmm17[13],xmm16[13],xmm17[14],xmm16[14],xmm17[15],xmm16[15] ; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} zmm21 = [5,4,3,6,5,4,3,6,5,4,3,6,7,7,7,7,41,40,43,42,41,40,43,42,41,40,43,42,45,44,47,46] ; AVX512DQ-BW-NEXT: vpermt2w %zmm19, %zmm21, %zmm20 ; AVX512DQ-BW-NEXT: vmovdqa64 (%rsi), %ymm16 ; AVX512DQ-BW-NEXT: vmovdqa64 (%rdi), %ymm17 ; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} ymm19 = 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] ; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm15 = xmm18[8],xmm15[8],xmm18[9],xmm15[9],xmm18[10],xmm15[10],xmm18[11],xmm15[11],xmm18[12],xmm15[12],xmm18[13],xmm15[13],xmm18[14],xmm15[14],xmm18[15],xmm15[15] ; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} zmm18 = [4,3,6,5,4,3,6,5,4,3,6,5,7,7,7,7,40,43,42,41,40,43,42,41,40,43,42,41,44,45,46,45] ; AVX512DQ-BW-NEXT: vpermt2w %zmm19, %zmm18, %zmm15 ; AVX512DQ-BW-NEXT: movl $613566756, %r10d # imm = 0x24924924 ; AVX512DQ-BW-NEXT: kmovd %r10d, %k2 ; AVX512DQ-BW-NEXT: vmovdqu16 %zmm20, %zmm15 {%k2} ; AVX512DQ-BW-NEXT: vextracti64x4 $1, %zmm15, %ymm19 ; AVX512DQ-BW-NEXT: vbroadcasti32x4 {{.*#+}} ymm20 = [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] ; AVX512DQ-BW-NEXT: # ymm20 = mem[0,1,2,3,0,1,2,3] ; AVX512DQ-BW-NEXT: vpshufb %ymm20, %ymm5, %ymm22 ; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm22 = ymm22[2,2,2,3] ; AVX512DQ-BW-NEXT: movw $18724, %r10w # imm = 0x4924 ; AVX512DQ-BW-NEXT: kmovd %r10d, %k3 ; AVX512DQ-BW-NEXT: vmovdqu16 %ymm22, %ymm19 {%k3} ; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm22 = xmm5[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15] ; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} ymm23 = [18,1,2,19,4,5,20,7,8,21,10,11,22,13,14,23] ; AVX512DQ-BW-NEXT: vpermt2w %ymm22, %ymm23, %ymm15 ; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm19, %zmm15, %zmm15 ; AVX512DQ-BW-NEXT: vbroadcasti32x4 {{.*#+}} ymm19 = [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] ; AVX512DQ-BW-NEXT: # ymm19 = mem[0,1,2,3,0,1,2,3] ; AVX512DQ-BW-NEXT: vpshufb %ymm19, %ymm8, %ymm22 ; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm24 = xmm8[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15] ; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} zmm25 = [2,5,4,3,2,5,4,3,2,5,4,3,6,5,6,7,40,41,42,43,40,41,42,43,40,41,42,43,44,45,46,47] ; AVX512DQ-BW-NEXT: vpermt2w %zmm22, %zmm25, %zmm24 ; AVX512DQ-BW-NEXT: movabsq $2342443691899625602, %r10 # imm = 0x2082082082082082 ; AVX512DQ-BW-NEXT: kmovq %r10, %k4 ; AVX512DQ-BW-NEXT: vmovdqu8 %zmm24, %zmm15 {%k4} ; AVX512DQ-BW-NEXT: vmovdqa64 32(%rcx), %ymm22 ; AVX512DQ-BW-NEXT: vmovdqa64 32(%rdx), %ymm24 ; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} ymm26 = ymm24[0],ymm22[0],ymm24[1],ymm22[1],ymm24[2],ymm22[2],ymm24[3],ymm22[3],ymm24[4],ymm22[4],ymm24[5],ymm22[5],ymm24[6],ymm22[6],ymm24[7],ymm22[7],ymm24[16],ymm22[16],ymm24[17],ymm22[17],ymm24[18],ymm22[18],ymm24[19],ymm22[19],ymm24[20],ymm22[20],ymm24[21],ymm22[21],ymm24[22],ymm22[22],ymm24[23],ymm22[23] ; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm10 = 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] ; AVX512DQ-BW-NEXT: vpermt2w %zmm26, %zmm21, %zmm10 ; AVX512DQ-BW-NEXT: vmovdqa 32(%rsi), %ymm11 ; AVX512DQ-BW-NEXT: vmovdqa64 32(%rdi), %ymm21 ; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} ymm26 = ymm21[0],ymm11[0],ymm21[1],ymm11[1],ymm21[2],ymm11[2],ymm21[3],ymm11[3],ymm21[4],ymm11[4],ymm21[5],ymm11[5],ymm21[6],ymm11[6],ymm21[7],ymm11[7],ymm21[16],ymm11[16],ymm21[17],ymm11[17],ymm21[18],ymm11[18],ymm21[19],ymm11[19],ymm21[20],ymm11[20],ymm21[21],ymm11[21],ymm21[22],ymm11[22],ymm21[23],ymm11[23] ; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm4 = xmm6[8],xmm4[8],xmm6[9],xmm4[9],xmm6[10],xmm4[10],xmm6[11],xmm4[11],xmm6[12],xmm4[12],xmm6[13],xmm4[13],xmm6[14],xmm4[14],xmm6[15],xmm4[15] ; AVX512DQ-BW-NEXT: vpermt2w %zmm26, %zmm18, %zmm4 ; AVX512DQ-BW-NEXT: vmovdqu16 %zmm10, %zmm4 {%k2} ; AVX512DQ-BW-NEXT: vextracti64x4 $1, %zmm4, %ymm6 ; AVX512DQ-BW-NEXT: vmovdqa 32(%r8), %ymm10 ; AVX512DQ-BW-NEXT: vpshufb %ymm20, %ymm10, %ymm10 ; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,2,2,3] ; AVX512DQ-BW-NEXT: vmovdqu16 %ymm10, %ymm6 {%k3} ; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm9 = xmm9[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15] ; AVX512DQ-BW-NEXT: vpermt2w %ymm9, %ymm23, %ymm4 ; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm6, %zmm4, %zmm4 ; AVX512DQ-BW-NEXT: vmovdqa 32(%r9), %ymm6 ; AVX512DQ-BW-NEXT: vpshufb %ymm19, %ymm6, %ymm6 ; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm9 = xmm12[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15] ; AVX512DQ-BW-NEXT: vpermt2w %zmm6, %zmm25, %zmm9 ; AVX512DQ-BW-NEXT: vmovdqu8 %zmm9, %zmm4 {%k4} ; AVX512DQ-BW-NEXT: vpbroadcastq {{.*#+}} ymm6 = [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] ; AVX512DQ-BW-NEXT: vpshufb %ymm6, %ymm11, %ymm9 ; AVX512DQ-BW-NEXT: vpshufb %ymm6, %ymm21, %ymm10 ; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} ymm9 = ymm10[0],ymm9[0],ymm10[1],ymm9[1],ymm10[2],ymm9[2],ymm10[3],ymm9[3],ymm10[4],ymm9[4],ymm10[5],ymm9[5],ymm10[6],ymm9[6],ymm10[7],ymm9[7],ymm10[16],ymm9[16],ymm10[17],ymm9[17],ymm10[18],ymm9[18],ymm10[19],ymm9[19],ymm10[20],ymm9[20],ymm10[21],ymm9[21],ymm10[22],ymm9[22],ymm10[23],ymm9[23] ; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm9 = ymm9[2,2,2,3] ; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} ymm10 = ymm21[8],ymm11[8],ymm21[9],ymm11[9],ymm21[10],ymm11[10],ymm21[11],ymm11[11],ymm21[12],ymm11[12],ymm21[13],ymm11[13],ymm21[14],ymm11[14],ymm21[15],ymm11[15],ymm21[24],ymm11[24],ymm21[25],ymm11[25],ymm21[26],ymm11[26],ymm21[27],ymm11[27],ymm21[28],ymm11[28],ymm21[29],ymm11[29],ymm21[30],ymm11[30],ymm21[31],ymm11[31] ; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} ymm11 = [12,11,14,13,12,11,14,13,12,11,14,13,15,15,15,15] ; AVX512DQ-BW-NEXT: vpermw %ymm10, %ymm11, %ymm10 ; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm10, %zmm9, %zmm9 ; AVX512DQ-BW-NEXT: vpshufb %ymm0, %ymm22, %ymm10 ; AVX512DQ-BW-NEXT: vpshufb %ymm0, %ymm24, %ymm12 ; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} ymm10 = ymm12[0],ymm10[0],ymm12[1],ymm10[1],ymm12[2],ymm10[2],ymm12[3],ymm10[3],ymm12[4],ymm10[4],ymm12[5],ymm10[5],ymm12[6],ymm10[6],ymm12[7],ymm10[7],ymm12[16],ymm10[16],ymm12[17],ymm10[17],ymm12[18],ymm10[18],ymm12[19],ymm10[19],ymm12[20],ymm10[20],ymm12[21],ymm10[21],ymm12[22],ymm10[22],ymm12[23],ymm10[23] ; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,2,2,3] ; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} ymm12 = ymm24[8],ymm22[8],ymm24[9],ymm22[9],ymm24[10],ymm22[10],ymm24[11],ymm22[11],ymm24[12],ymm22[12],ymm24[13],ymm22[13],ymm24[14],ymm22[14],ymm24[15],ymm22[15],ymm24[24],ymm22[24],ymm24[25],ymm22[25],ymm24[26],ymm22[26],ymm24[27],ymm22[27],ymm24[28],ymm22[28],ymm24[29],ymm22[29],ymm24[30],ymm22[30],ymm24[31],ymm22[31] ; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} ymm18 = [13,12,11,14,13,12,11,14,13,12,11,14,15,15,15,15] ; AVX512DQ-BW-NEXT: vpermw %ymm12, %ymm18, %ymm12 ; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm12, %zmm10, %zmm10 ; AVX512DQ-BW-NEXT: vmovdqu16 %zmm9, %zmm10 {%k2} ; AVX512DQ-BW-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm2[4,5,6,7,4,5,6,7] ; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} zmm9 = [6,5,8,7,0,9,0,0,6,5,8,7,0,9,0,0,10,13,12,11,14,0,0,15,10,13,12,11,14,0,0,15] ; AVX512DQ-BW-NEXT: vpshufb %zmm9, %zmm2, %zmm2 ; AVX512DQ-BW-NEXT: vpermq {{.*#+}} zmm2 = zmm2[2,2,2,3,6,6,6,7] ; AVX512DQ-BW-NEXT: vmovdqu16 %zmm2, %zmm10 {%k1} ; AVX512DQ-BW-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm1[4,5,6,7,4,5,6,7] ; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} zmm1 = zmm1[u,6,u,5,u,8,u,7,u,u,u,9,u,u,u,u,u,22,u,21,u,24,u,23,u,u,u,25,u,u,u,u,u,42,u,45,u,44,u,43,u,46,u,u,u,u,u,47,u,58,u,61,u,60,u,59,u,62,u,u,u,u,u,63] ; AVX512DQ-BW-NEXT: vpermq {{.*#+}} zmm1 = zmm1[2,2,2,3,6,6,6,7] ; AVX512DQ-BW-NEXT: movabsq $-9076969306111049208, %rcx # imm = 0x8208208208208208 ; AVX512DQ-BW-NEXT: kmovq %rcx, %k3 ; AVX512DQ-BW-NEXT: vmovdqu8 %zmm1, %zmm10 {%k3} ; AVX512DQ-BW-NEXT: vpshufb %ymm6, %ymm16, %ymm1 ; AVX512DQ-BW-NEXT: vpshufb %ymm6, %ymm17, %ymm2 ; AVX512DQ-BW-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] ; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3] ; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} ymm2 = 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] ; AVX512DQ-BW-NEXT: vpermw %ymm2, %ymm11, %ymm2 ; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm1 ; AVX512DQ-BW-NEXT: vpshufb %ymm0, %ymm13, %ymm2 ; AVX512DQ-BW-NEXT: vpshufb %ymm0, %ymm14, %ymm0 ; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} ymm0 = ymm0[0],ymm2[0],ymm0[1],ymm2[1],ymm0[2],ymm2[2],ymm0[3],ymm2[3],ymm0[4],ymm2[4],ymm0[5],ymm2[5],ymm0[6],ymm2[6],ymm0[7],ymm2[7],ymm0[16],ymm2[16],ymm0[17],ymm2[17],ymm0[18],ymm2[18],ymm0[19],ymm2[19],ymm0[20],ymm2[20],ymm0[21],ymm2[21],ymm0[22],ymm2[22],ymm0[23],ymm2[23] ; AVX512DQ-BW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3] ; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} ymm2 = ymm14[8],ymm13[8],ymm14[9],ymm13[9],ymm14[10],ymm13[10],ymm14[11],ymm13[11],ymm14[12],ymm13[12],ymm14[13],ymm13[13],ymm14[14],ymm13[14],ymm14[15],ymm13[15],ymm14[24],ymm13[24],ymm14[25],ymm13[25],ymm14[26],ymm13[26],ymm14[27],ymm13[27],ymm14[28],ymm13[28],ymm14[29],ymm13[29],ymm14[30],ymm13[30],ymm14[31],ymm13[31] ; AVX512DQ-BW-NEXT: vpermw %ymm2, %ymm18, %ymm2 ; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm2, %zmm0, %zmm0 ; AVX512DQ-BW-NEXT: vmovdqu16 %zmm1, %zmm0 {%k2} ; AVX512DQ-BW-NEXT: vpshufb %zmm9, %zmm5, %zmm1 ; AVX512DQ-BW-NEXT: vpermq {{.*#+}} zmm1 = zmm1[2,2,2,3,6,6,6,7] ; AVX512DQ-BW-NEXT: vmovdqu16 %zmm1, %zmm0 {%k1} ; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} zmm1 = zmm8[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,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,58,u,61,u,60,u,59,u,62,u,u,u,u,u,63] ; AVX512DQ-BW-NEXT: vpermq {{.*#+}} zmm1 = zmm1[2,2,2,3,6,6,6,7] ; AVX512DQ-BW-NEXT: vmovdqu8 %zmm1, %zmm0 {%k3} ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, 128(%rax) ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm10, 320(%rax) ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm4, 256(%rax) ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm15, 64(%rax) ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm7, (%rax) ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, 192(%rax) ; AVX512DQ-BW-NEXT: vzeroupper ; AVX512DQ-BW-NEXT: retq ; ; AVX512DQ-BW-FCP-LABEL: store_i8_stride6_vf64: ; AVX512DQ-BW-FCP: # %bb.0: ; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%r8), %zmm2 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%r9), %zmm1 ; AVX512DQ-BW-FCP-NEXT: vpbroadcastq {{.*#+}} ymm0 = [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] ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rcx), %xmm15 ; AVX512DQ-BW-FCP-NEXT: vmovdqa 32(%rcx), %xmm6 ; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm0, %xmm6, %xmm3 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rdx), %xmm16 ; AVX512DQ-BW-FCP-NEXT: vmovdqa 32(%rdx), %xmm8 ; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm0, %xmm8, %xmm4 ; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm3 = 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] ; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm4 = xmm8[0],xmm6[0],xmm8[1],xmm6[1],xmm8[2],xmm6[2],xmm8[3],xmm6[3],xmm8[4],xmm6[4],xmm8[5],xmm6[5],xmm8[6],xmm6[6],xmm8[7],xmm6[7] ; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm5 = [1,0,3,2,1,0,3,2,1,0,3,2,5,4,7,6,32,33,34,35,32,33,34,35,32,33,34,35,36,37,38,39] ; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm3, %zmm5, %zmm4 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rsi), %xmm17 ; AVX512DQ-BW-FCP-NEXT: vmovdqa 32(%rsi), %xmm9 ; AVX512DQ-BW-FCP-NEXT: vpbroadcastq {{.*#+}} xmm7 = [8,7,6,9,0,0,10,0,8,7,6,9,0,0,10,0] ; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm7, %xmm9, %xmm3 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rdi), %xmm18 ; AVX512DQ-BW-FCP-NEXT: vmovdqa 32(%rdi), %xmm10 ; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm7, %xmm10, %xmm11 ; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm11[8],xmm3[8],xmm11[9],xmm3[9],xmm11[10],xmm3[10],xmm11[11],xmm3[11],xmm11[12],xmm3[12],xmm11[13],xmm3[13],xmm11[14],xmm3[14],xmm11[15],xmm3[15] ; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm12 = 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] ; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm13 = [0,3,2,1,0,3,2,1,0,3,2,1,4,5,6,5,32,33,34,35,32,33,34,35,32,33,34,35,36,37,38,39] ; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm3, %zmm13, %zmm12 ; AVX512DQ-BW-FCP-NEXT: movl $-1840700270, %r10d # imm = 0x92492492 ; AVX512DQ-BW-FCP-NEXT: kmovd %r10d, %k1 ; AVX512DQ-BW-FCP-NEXT: vmovdqu16 %zmm4, %zmm12 {%k1} ; AVX512DQ-BW-FCP-NEXT: vmovdqa 32(%r8), %xmm11 ; AVX512DQ-BW-FCP-NEXT: vpmovzxbw {{.*#+}} xmm3 = xmm11[0],zero,xmm11[1],zero,xmm11[2],zero,xmm11[3],zero,xmm11[4],zero,xmm11[5],zero,xmm11[6],zero,xmm11[7],zero ; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm14 = [0,1,16,3,4,17,6,7,18,9,10,19,12,13,20,15] ; AVX512DQ-BW-FCP-NEXT: vmovdqa %ymm12, %ymm4 ; AVX512DQ-BW-FCP-NEXT: vpermt2w %ymm3, %ymm14, %ymm4 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm19 = [16,37,18,19,38,21,22,39,24,25,32,27,28,33,30,31] ; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} xmm20 = [8,9,0,0,0,5,6,7] ; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm20, %xmm11, %xmm3 ; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm3, %zmm19, %zmm12 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm12, %zmm4, %zmm3 ; AVX512DQ-BW-FCP-NEXT: vmovdqa 32(%r9), %xmm12 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 {{.*#+}} xmm21 = [u,8,u,9,u,10,u,11,u,4,u,5,u,6,u,7] ; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm21, %xmm12, %xmm4 ; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm22 = xmm12[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7] ; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm23 = [2,1,0,3,2,1,0,3,2,1,0,3,4,4,4,4,38,37,32,39,38,37,32,39,38,37,32,39,33,33,33,33] ; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm4, %zmm23, %zmm22 ; AVX512DQ-BW-FCP-NEXT: movabsq $585610922974906400, %r10 # imm = 0x820820820820820 ; AVX512DQ-BW-FCP-NEXT: kmovq %r10, %k2 ; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm22, %zmm3 {%k2} ; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm0, %xmm15, %xmm4 ; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm0, %xmm16, %xmm22 ; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm4 = xmm22[0],xmm4[0],xmm22[1],xmm4[1],xmm22[2],xmm4[2],xmm22[3],xmm4[3],xmm22[4],xmm4[4],xmm22[5],xmm4[5],xmm22[6],xmm4[6],xmm22[7],xmm4[7] ; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm22 = xmm16[0],xmm15[0],xmm16[1],xmm15[1],xmm16[2],xmm15[2],xmm16[3],xmm15[3],xmm16[4],xmm15[4],xmm16[5],xmm15[5],xmm16[6],xmm15[6],xmm16[7],xmm15[7] ; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm4, %zmm5, %zmm22 ; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm7, %xmm17, %xmm4 ; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm7, %xmm18, %xmm5 ; AVX512DQ-BW-FCP-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] ; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm5 = 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] ; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm4, %zmm13, %zmm5 ; AVX512DQ-BW-FCP-NEXT: vmovdqu16 %zmm22, %zmm5 {%k1} ; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm4 = mem[0,1,2,3,0,1,2,3] ; AVX512DQ-BW-FCP-NEXT: vpmovzxbw {{.*#+}} xmm7 = xmm4[0],zero,xmm4[1],zero,xmm4[2],zero,xmm4[3],zero,xmm4[4],zero,xmm4[5],zero,xmm4[6],zero,xmm4[7],zero ; AVX512DQ-BW-FCP-NEXT: vpermi2w %ymm7, %ymm5, %ymm14 ; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm20, %xmm4, %xmm7 ; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm7, %zmm19, %zmm5 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm5, %zmm14, %zmm5 ; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = mem[0,1,2,3,0,1,2,3] ; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm21, %xmm7, %xmm13 ; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm14 = xmm7[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7] ; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm13, %zmm23, %zmm14 ; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm14, %zmm5 {%k2} ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rcx), %ymm13 ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdx), %ymm14 ; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} ymm19 = ymm14[0],ymm13[0],ymm14[1],ymm13[1],ymm14[2],ymm13[2],ymm14[3],ymm13[3],ymm14[4],ymm13[4],ymm14[5],ymm13[5],ymm14[6],ymm13[6],ymm14[7],ymm13[7],ymm14[16],ymm13[16],ymm14[17],ymm13[17],ymm14[18],ymm13[18],ymm14[19],ymm13[19],ymm14[20],ymm13[20],ymm14[21],ymm13[21],ymm14[22],ymm13[22],ymm14[23],ymm13[23] ; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm20 = xmm16[8],xmm15[8],xmm16[9],xmm15[9],xmm16[10],xmm15[10],xmm16[11],xmm15[11],xmm16[12],xmm15[12],xmm16[13],xmm15[13],xmm16[14],xmm15[14],xmm16[15],xmm15[15] ; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm21 = [5,4,3,6,5,4,3,6,5,4,3,6,7,7,7,7,41,40,43,42,41,40,43,42,41,40,43,42,45,44,47,46] ; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm19, %zmm21, %zmm20 ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rsi), %ymm15 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rdi), %ymm16 ; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} ymm19 = ymm16[0],ymm15[0],ymm16[1],ymm15[1],ymm16[2],ymm15[2],ymm16[3],ymm15[3],ymm16[4],ymm15[4],ymm16[5],ymm15[5],ymm16[6],ymm15[6],ymm16[7],ymm15[7],ymm16[16],ymm15[16],ymm16[17],ymm15[17],ymm16[18],ymm15[18],ymm16[19],ymm15[19],ymm16[20],ymm15[20],ymm16[21],ymm15[21],ymm16[22],ymm15[22],ymm16[23],ymm15[23] ; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm17 = xmm18[8],xmm17[8],xmm18[9],xmm17[9],xmm18[10],xmm17[10],xmm18[11],xmm17[11],xmm18[12],xmm17[12],xmm18[13],xmm17[13],xmm18[14],xmm17[14],xmm18[15],xmm17[15] ; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm18 = [4,3,6,5,4,3,6,5,4,3,6,5,7,7,7,7,40,43,42,41,40,43,42,41,40,43,42,41,44,45,46,45] ; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm19, %zmm18, %zmm17 ; AVX512DQ-BW-FCP-NEXT: movl $613566756, %r10d # imm = 0x24924924 ; AVX512DQ-BW-FCP-NEXT: kmovd %r10d, %k2 ; AVX512DQ-BW-FCP-NEXT: vmovdqu16 %zmm20, %zmm17 {%k2} ; AVX512DQ-BW-FCP-NEXT: vextracti64x4 $1, %zmm17, %ymm19 ; AVX512DQ-BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} ymm20 = [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] ; AVX512DQ-BW-FCP-NEXT: # ymm20 = mem[0,1,2,3,0,1,2,3] ; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm20, %ymm4, %ymm22 ; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm22 = ymm22[2,2,2,3] ; AVX512DQ-BW-FCP-NEXT: movw $18724, %r10w # imm = 0x4924 ; AVX512DQ-BW-FCP-NEXT: kmovd %r10d, %k3 ; AVX512DQ-BW-FCP-NEXT: vmovdqu16 %ymm22, %ymm19 {%k3} ; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm22 = xmm4[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15] ; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm23 = [18,1,2,19,4,5,20,7,8,21,10,11,22,13,14,23] ; AVX512DQ-BW-FCP-NEXT: vpermt2w %ymm22, %ymm23, %ymm17 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm19, %zmm17, %zmm17 ; AVX512DQ-BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} ymm19 = [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] ; AVX512DQ-BW-FCP-NEXT: # ymm19 = mem[0,1,2,3,0,1,2,3] ; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm19, %ymm7, %ymm22 ; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm24 = xmm7[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15] ; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm25 = [2,5,4,3,2,5,4,3,2,5,4,3,6,5,6,7,40,41,42,43,40,41,42,43,40,41,42,43,44,45,46,47] ; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm22, %zmm25, %zmm24 ; AVX512DQ-BW-FCP-NEXT: movabsq $2342443691899625602, %r10 # imm = 0x2082082082082082 ; AVX512DQ-BW-FCP-NEXT: kmovq %r10, %k4 ; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm24, %zmm17 {%k4} ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 32(%rcx), %ymm22 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 32(%rdx), %ymm24 ; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} ymm26 = ymm24[0],ymm22[0],ymm24[1],ymm22[1],ymm24[2],ymm22[2],ymm24[3],ymm22[3],ymm24[4],ymm22[4],ymm24[5],ymm22[5],ymm24[6],ymm22[6],ymm24[7],ymm22[7],ymm24[16],ymm22[16],ymm24[17],ymm22[17],ymm24[18],ymm22[18],ymm24[19],ymm22[19],ymm24[20],ymm22[20],ymm24[21],ymm22[21],ymm24[22],ymm22[22],ymm24[23],ymm22[23] ; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm6 = xmm8[8],xmm6[8],xmm8[9],xmm6[9],xmm8[10],xmm6[10],xmm8[11],xmm6[11],xmm8[12],xmm6[12],xmm8[13],xmm6[13],xmm8[14],xmm6[14],xmm8[15],xmm6[15] ; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm26, %zmm21, %zmm6 ; AVX512DQ-BW-FCP-NEXT: vmovdqa 32(%rsi), %ymm8 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 32(%rdi), %ymm21 ; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} ymm26 = ymm21[0],ymm8[0],ymm21[1],ymm8[1],ymm21[2],ymm8[2],ymm21[3],ymm8[3],ymm21[4],ymm8[4],ymm21[5],ymm8[5],ymm21[6],ymm8[6],ymm21[7],ymm8[7],ymm21[16],ymm8[16],ymm21[17],ymm8[17],ymm21[18],ymm8[18],ymm21[19],ymm8[19],ymm21[20],ymm8[20],ymm21[21],ymm8[21],ymm21[22],ymm8[22],ymm21[23],ymm8[23] ; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm9 = 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] ; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm26, %zmm18, %zmm9 ; AVX512DQ-BW-FCP-NEXT: vmovdqu16 %zmm6, %zmm9 {%k2} ; AVX512DQ-BW-FCP-NEXT: vextracti64x4 $1, %zmm9, %ymm6 ; AVX512DQ-BW-FCP-NEXT: vmovdqa 32(%r8), %ymm10 ; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm20, %ymm10, %ymm10 ; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,2,2,3] ; AVX512DQ-BW-FCP-NEXT: vmovdqu16 %ymm10, %ymm6 {%k3} ; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm10 = xmm11[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15] ; AVX512DQ-BW-FCP-NEXT: vpermt2w %ymm10, %ymm23, %ymm9 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm6, %zmm9, %zmm6 ; AVX512DQ-BW-FCP-NEXT: vmovdqa 32(%r9), %ymm9 ; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm19, %ymm9, %ymm9 ; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm10 = xmm12[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15] ; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm9, %zmm25, %zmm10 ; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm10, %zmm6 {%k4} ; AVX512DQ-BW-FCP-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] ; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm9, %ymm8, %ymm10 ; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm9, %ymm21, %ymm11 ; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} ymm10 = ymm11[0],ymm10[0],ymm11[1],ymm10[1],ymm11[2],ymm10[2],ymm11[3],ymm10[3],ymm11[4],ymm10[4],ymm11[5],ymm10[5],ymm11[6],ymm10[6],ymm11[7],ymm10[7],ymm11[16],ymm10[16],ymm11[17],ymm10[17],ymm11[18],ymm10[18],ymm11[19],ymm10[19],ymm11[20],ymm10[20],ymm11[21],ymm10[21],ymm11[22],ymm10[22],ymm11[23],ymm10[23] ; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,2,2,3] ; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} ymm8 = ymm21[8],ymm8[8],ymm21[9],ymm8[9],ymm21[10],ymm8[10],ymm21[11],ymm8[11],ymm21[12],ymm8[12],ymm21[13],ymm8[13],ymm21[14],ymm8[14],ymm21[15],ymm8[15],ymm21[24],ymm8[24],ymm21[25],ymm8[25],ymm21[26],ymm8[26],ymm21[27],ymm8[27],ymm21[28],ymm8[28],ymm21[29],ymm8[29],ymm21[30],ymm8[30],ymm21[31],ymm8[31] ; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm11 = [12,11,14,13,12,11,14,13,12,11,14,13,15,15,15,15] ; AVX512DQ-BW-FCP-NEXT: vpermw %ymm8, %ymm11, %ymm8 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm8, %zmm10, %zmm8 ; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm0, %ymm22, %ymm10 ; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm0, %ymm24, %ymm12 ; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} ymm10 = ymm12[0],ymm10[0],ymm12[1],ymm10[1],ymm12[2],ymm10[2],ymm12[3],ymm10[3],ymm12[4],ymm10[4],ymm12[5],ymm10[5],ymm12[6],ymm10[6],ymm12[7],ymm10[7],ymm12[16],ymm10[16],ymm12[17],ymm10[17],ymm12[18],ymm10[18],ymm12[19],ymm10[19],ymm12[20],ymm10[20],ymm12[21],ymm10[21],ymm12[22],ymm10[22],ymm12[23],ymm10[23] ; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm10 = ymm10[2,2,2,3] ; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} ymm12 = ymm24[8],ymm22[8],ymm24[9],ymm22[9],ymm24[10],ymm22[10],ymm24[11],ymm22[11],ymm24[12],ymm22[12],ymm24[13],ymm22[13],ymm24[14],ymm22[14],ymm24[15],ymm22[15],ymm24[24],ymm22[24],ymm24[25],ymm22[25],ymm24[26],ymm22[26],ymm24[27],ymm22[27],ymm24[28],ymm22[28],ymm24[29],ymm22[29],ymm24[30],ymm22[30],ymm24[31],ymm22[31] ; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} ymm18 = [13,12,11,14,13,12,11,14,13,12,11,14,15,15,15,15] ; AVX512DQ-BW-FCP-NEXT: vpermw %ymm12, %ymm18, %ymm12 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm12, %zmm10, %zmm10 ; AVX512DQ-BW-FCP-NEXT: vmovdqu16 %zmm8, %zmm10 {%k2} ; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm2[4,5,6,7,4,5,6,7] ; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm8 = [6,5,8,7,0,9,0,0,6,5,8,7,0,9,0,0,10,13,12,11,14,0,0,15,10,13,12,11,14,0,0,15] ; AVX512DQ-BW-FCP-NEXT: vpshufb %zmm8, %zmm2, %zmm2 ; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} zmm2 = zmm2[2,2,2,3,6,6,6,7] ; AVX512DQ-BW-FCP-NEXT: vmovdqu16 %zmm2, %zmm10 {%k1} ; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm1 = zmm1[4,5,6,7,4,5,6,7] ; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} zmm1 = zmm1[u,6,u,5,u,8,u,7,u,u,u,9,u,u,u,u,u,22,u,21,u,24,u,23,u,u,u,25,u,u,u,u,u,42,u,45,u,44,u,43,u,46,u,u,u,u,u,47,u,58,u,61,u,60,u,59,u,62,u,u,u,u,u,63] ; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} zmm1 = zmm1[2,2,2,3,6,6,6,7] ; AVX512DQ-BW-FCP-NEXT: movabsq $-9076969306111049208, %rcx # imm = 0x8208208208208208 ; AVX512DQ-BW-FCP-NEXT: kmovq %rcx, %k3 ; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm1, %zmm10 {%k3} ; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm9, %ymm15, %ymm1 ; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm9, %ymm16, %ymm2 ; AVX512DQ-BW-FCP-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] ; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[2,2,2,3] ; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} ymm2 = ymm16[8],ymm15[8],ymm16[9],ymm15[9],ymm16[10],ymm15[10],ymm16[11],ymm15[11],ymm16[12],ymm15[12],ymm16[13],ymm15[13],ymm16[14],ymm15[14],ymm16[15],ymm15[15],ymm16[24],ymm15[24],ymm16[25],ymm15[25],ymm16[26],ymm15[26],ymm16[27],ymm15[27],ymm16[28],ymm15[28],ymm16[29],ymm15[29],ymm16[30],ymm15[30],ymm16[31],ymm15[31] ; AVX512DQ-BW-FCP-NEXT: vpermw %ymm2, %ymm11, %ymm2 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm1 ; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm0, %ymm13, %ymm2 ; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm0, %ymm14, %ymm0 ; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} ymm0 = ymm0[0],ymm2[0],ymm0[1],ymm2[1],ymm0[2],ymm2[2],ymm0[3],ymm2[3],ymm0[4],ymm2[4],ymm0[5],ymm2[5],ymm0[6],ymm2[6],ymm0[7],ymm2[7],ymm0[16],ymm2[16],ymm0[17],ymm2[17],ymm0[18],ymm2[18],ymm0[19],ymm2[19],ymm0[20],ymm2[20],ymm0[21],ymm2[21],ymm0[22],ymm2[22],ymm0[23],ymm2[23] ; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[2,2,2,3] ; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} ymm2 = ymm14[8],ymm13[8],ymm14[9],ymm13[9],ymm14[10],ymm13[10],ymm14[11],ymm13[11],ymm14[12],ymm13[12],ymm14[13],ymm13[13],ymm14[14],ymm13[14],ymm14[15],ymm13[15],ymm14[24],ymm13[24],ymm14[25],ymm13[25],ymm14[26],ymm13[26],ymm14[27],ymm13[27],ymm14[28],ymm13[28],ymm14[29],ymm13[29],ymm14[30],ymm13[30],ymm14[31],ymm13[31] ; AVX512DQ-BW-FCP-NEXT: vpermw %ymm2, %ymm18, %ymm2 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm0, %zmm0 ; AVX512DQ-BW-FCP-NEXT: vmovdqu16 %zmm1, %zmm0 {%k2} ; AVX512DQ-BW-FCP-NEXT: vpshufb %zmm8, %zmm4, %zmm1 ; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} zmm1 = zmm1[2,2,2,3,6,6,6,7] ; AVX512DQ-BW-FCP-NEXT: vmovdqu16 %zmm1, %zmm0 {%k1} ; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} zmm1 = zmm7[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,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,58,u,61,u,60,u,59,u,62,u,u,u,u,u,63] ; AVX512DQ-BW-FCP-NEXT: vpermq {{.*#+}} zmm1 = zmm1[2,2,2,3,6,6,6,7] ; AVX512DQ-BW-FCP-NEXT: vmovdqu8 %zmm1, %zmm0 {%k3} ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, 128(%rax) ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm10, 320(%rax) ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm6, 256(%rax) ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm17, 64(%rax) ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm5, (%rax) ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, 192(%rax) ; AVX512DQ-BW-FCP-NEXT: vzeroupper ; AVX512DQ-BW-FCP-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> %2 = shufflevector <64 x i8> %in.vec2, <64 x i8> %in.vec3, <128 x i32> %3 = shufflevector <64 x i8> %in.vec4, <64 x i8> %in.vec5, <128 x i32> %4 = shufflevector <128 x i8> %1, <128 x i8> %2, <256 x i32> %5 = shufflevector <128 x i8> %3, <128 x i8> poison, <256 x i32> %6 = shufflevector <256 x i8> %4, <256 x i8> %5, <384 x i32> %interleaved.vec = shufflevector <384 x i8> %6, <384 x i8> poison, <384 x i32> store <384 x i8> %interleaved.vec, ptr %out.vec, align 64 ret void }