; 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_stride8_vf2(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecptr2, ptr %in.vecptr3, ptr %in.vecptr4, ptr %in.vecptr5, ptr %in.vecptr6, ptr %in.vecptr7, ptr %out.vec) nounwind { ; SSE-LABEL: store_i8_stride8_vf2: ; SSE: # %bb.0: ; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax ; SSE-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; SSE-NEXT: movq {{[0-9]+}}(%rsp), %r11 ; SSE-NEXT: movdqa (%rdi), %xmm0 ; SSE-NEXT: movdqa (%rdx), %xmm1 ; SSE-NEXT: movdqa (%r8), %xmm2 ; SSE-NEXT: movdqa (%r11), %xmm3 ; 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: punpcklwd {{.*#+}} xmm2 = xmm2[0],mem[0],xmm2[1],mem[1],xmm2[2],mem[2],xmm2[3],mem[3] ; SSE-NEXT: punpcklwd {{.*#+}} xmm3 = xmm3[0],mem[0],xmm3[1],mem[1],xmm3[2],mem[2],xmm3[3],mem[3] ; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1] ; SSE-NEXT: pxor %xmm1, %xmm1 ; 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: pshuflw {{.*#+}} xmm3 = xmm2[0,2,2,3,4,5,6,7] ; SSE-NEXT: pshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,6,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm3[0,1,0,2] ; SSE-NEXT: pshuflw {{.*#+}} xmm2 = xmm2[3,1,2,3,4,5,6,7] ; SSE-NEXT: pshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,7,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm2[0,1,2,0] ; SSE-NEXT: pshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,7,6,5,4] ; SSE-NEXT: packuswb %xmm2, %xmm3 ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm3[1,3,2,3] ; 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: pshuflw {{.*#+}} xmm1 = xmm0[0,2,2,3,4,5,6,7] ; SSE-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,6,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,2,2,3] ; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[3,1,2,3,4,5,6,7] ; SSE-NEXT: pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,7,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3] ; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[1,0,3,2,4,5,6,7] ; SSE-NEXT: packuswb %xmm0, %xmm1 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm1[0,2,2,3] ; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1] ; SSE-NEXT: movdqa %xmm0, (%rax) ; SSE-NEXT: retq ; ; AVX-LABEL: store_i8_stride8_vf2: ; AVX: # %bb.0: ; AVX-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX-NEXT: movq {{[0-9]+}}(%rsp), %r11 ; AVX-NEXT: vmovdqa (%rdi), %xmm0 ; AVX-NEXT: vmovdqa (%rdx), %xmm1 ; AVX-NEXT: vmovdqa (%r8), %xmm2 ; AVX-NEXT: vmovdqa (%r11), %xmm3 ; 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: vpunpcklwd {{.*#+}} xmm1 = xmm2[0],mem[0],xmm2[1],mem[1],xmm2[2],mem[2],xmm2[3],mem[3] ; AVX-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm3[0],mem[0],xmm3[1],mem[1],xmm3[2],mem[2],xmm3[3],mem[3] ; AVX-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1] ; AVX-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] ; AVX-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,4,8,12,1,5,9,13,2,6,10,14,3,7,11,15] ; AVX-NEXT: vmovdqa %xmm0, (%rax) ; AVX-NEXT: retq ; ; AVX2-LABEL: store_i8_stride8_vf2: ; AVX2: # %bb.0: ; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %r11 ; AVX2-NEXT: vmovdqa (%rdi), %xmm0 ; AVX2-NEXT: vmovdqa (%rdx), %xmm1 ; AVX2-NEXT: vmovdqa (%r8), %xmm2 ; AVX2-NEXT: vmovdqa (%r11), %xmm3 ; 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: vpunpcklwd {{.*#+}} xmm1 = xmm2[0],mem[0],xmm2[1],mem[1],xmm2[2],mem[2],xmm2[3],mem[3] ; AVX2-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm3[0],mem[0],xmm3[1],mem[1],xmm3[2],mem[2],xmm3[3],mem[3] ; AVX2-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1] ; AVX2-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] ; AVX2-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,4,8,12,1,5,9,13,2,6,10,14,3,7,11,15] ; AVX2-NEXT: vmovdqa %xmm0, (%rax) ; AVX2-NEXT: retq ; ; AVX2-FP-LABEL: store_i8_stride8_vf2: ; AVX2-FP: # %bb.0: ; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %r11 ; AVX2-FP-NEXT: vmovdqa (%rdi), %xmm0 ; AVX2-FP-NEXT: vmovdqa (%rdx), %xmm1 ; AVX2-FP-NEXT: vmovdqa (%r8), %xmm2 ; AVX2-FP-NEXT: vmovdqa (%r11), %xmm3 ; 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: vpunpcklwd {{.*#+}} xmm1 = xmm2[0],mem[0],xmm2[1],mem[1],xmm2[2],mem[2],xmm2[3],mem[3] ; AVX2-FP-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm3[0],mem[0],xmm3[1],mem[1],xmm3[2],mem[2],xmm3[3],mem[3] ; AVX2-FP-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1] ; AVX2-FP-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] ; AVX2-FP-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,4,8,12,1,5,9,13,2,6,10,14,3,7,11,15] ; AVX2-FP-NEXT: vmovdqa %xmm0, (%rax) ; AVX2-FP-NEXT: retq ; ; AVX2-FCP-LABEL: store_i8_stride8_vf2: ; AVX2-FCP: # %bb.0: ; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r11 ; AVX2-FCP-NEXT: vmovdqa (%rdi), %xmm0 ; AVX2-FCP-NEXT: vmovdqa (%rdx), %xmm1 ; AVX2-FCP-NEXT: vmovdqa (%r8), %xmm2 ; AVX2-FCP-NEXT: vmovdqa (%r11), %xmm3 ; 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: vpunpcklwd {{.*#+}} xmm1 = xmm2[0],mem[0],xmm2[1],mem[1],xmm2[2],mem[2],xmm2[3],mem[3] ; AVX2-FCP-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm3[0],mem[0],xmm3[1],mem[1],xmm3[2],mem[2],xmm3[3],mem[3] ; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1] ; AVX2-FCP-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] ; AVX2-FCP-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,4,8,12,1,5,9,13,2,6,10,14,3,7,11,15] ; AVX2-FCP-NEXT: vmovdqa %xmm0, (%rax) ; AVX2-FCP-NEXT: retq ; ; AVX512-LABEL: store_i8_stride8_vf2: ; AVX512: # %bb.0: ; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %r11 ; AVX512-NEXT: vmovdqa (%rdi), %xmm0 ; AVX512-NEXT: vmovdqa (%rdx), %xmm1 ; AVX512-NEXT: vmovdqa (%r8), %xmm2 ; AVX512-NEXT: vmovdqa (%r11), %xmm3 ; 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: vpunpcklwd {{.*#+}} xmm1 = xmm2[0],mem[0],xmm2[1],mem[1],xmm2[2],mem[2],xmm2[3],mem[3] ; AVX512-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm3[0],mem[0],xmm3[1],mem[1],xmm3[2],mem[2],xmm3[3],mem[3] ; AVX512-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1] ; 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: vpshufb {{.*#+}} xmm0 = xmm0[0,4,8,12,1,5,9,13,2,6,10,14,3,7,11,15] ; AVX512-NEXT: vmovdqa %xmm0, (%rax) ; AVX512-NEXT: retq ; ; AVX512-FCP-LABEL: store_i8_stride8_vf2: ; AVX512-FCP: # %bb.0: ; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r11 ; AVX512-FCP-NEXT: vmovdqa (%rdi), %xmm0 ; AVX512-FCP-NEXT: vmovdqa (%rdx), %xmm1 ; AVX512-FCP-NEXT: vmovdqa (%r8), %xmm2 ; AVX512-FCP-NEXT: vmovdqa (%r11), %xmm3 ; 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: vpunpcklwd {{.*#+}} xmm1 = xmm2[0],mem[0],xmm2[1],mem[1],xmm2[2],mem[2],xmm2[3],mem[3] ; AVX512-FCP-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm3[0],mem[0],xmm3[1],mem[1],xmm3[2],mem[2],xmm3[3],mem[3] ; AVX512-FCP-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1] ; AVX512-FCP-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-FCP-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,4,8,12,1,5,9,13,2,6,10,14,3,7,11,15] ; AVX512-FCP-NEXT: vmovdqa %xmm0, (%rax) ; AVX512-FCP-NEXT: retq ; ; AVX512DQ-LABEL: store_i8_stride8_vf2: ; AVX512DQ: # %bb.0: ; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %r11 ; AVX512DQ-NEXT: vmovdqa (%rdi), %xmm0 ; AVX512DQ-NEXT: vmovdqa (%rdx), %xmm1 ; AVX512DQ-NEXT: vmovdqa (%r8), %xmm2 ; AVX512DQ-NEXT: vmovdqa (%r11), %xmm3 ; 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: vpunpcklwd {{.*#+}} xmm1 = xmm2[0],mem[0],xmm2[1],mem[1],xmm2[2],mem[2],xmm2[3],mem[3] ; AVX512DQ-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm3[0],mem[0],xmm3[1],mem[1],xmm3[2],mem[2],xmm3[3],mem[3] ; AVX512DQ-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1] ; 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: vpshufb {{.*#+}} xmm0 = xmm0[0,4,8,12,1,5,9,13,2,6,10,14,3,7,11,15] ; AVX512DQ-NEXT: vmovdqa %xmm0, (%rax) ; AVX512DQ-NEXT: retq ; ; AVX512DQ-FCP-LABEL: store_i8_stride8_vf2: ; AVX512DQ-FCP: # %bb.0: ; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r11 ; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %xmm0 ; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %xmm1 ; AVX512DQ-FCP-NEXT: vmovdqa (%r8), %xmm2 ; AVX512DQ-FCP-NEXT: vmovdqa (%r11), %xmm3 ; 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: vpunpcklwd {{.*#+}} xmm1 = xmm2[0],mem[0],xmm2[1],mem[1],xmm2[2],mem[2],xmm2[3],mem[3] ; AVX512DQ-FCP-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm3[0],mem[0],xmm3[1],mem[1],xmm3[2],mem[2],xmm3[3],mem[3] ; AVX512DQ-FCP-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1] ; AVX512DQ-FCP-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-FCP-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,4,8,12,1,5,9,13,2,6,10,14,3,7,11,15] ; AVX512DQ-FCP-NEXT: vmovdqa %xmm0, (%rax) ; AVX512DQ-FCP-NEXT: retq ; ; AVX512BW-LABEL: store_i8_stride8_vf2: ; AVX512BW: # %bb.0: ; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %r11 ; AVX512BW-NEXT: vmovdqa (%rdi), %xmm0 ; AVX512BW-NEXT: vmovdqa (%rdx), %xmm1 ; AVX512BW-NEXT: vmovdqa (%r8), %xmm2 ; AVX512BW-NEXT: vmovdqa (%r11), %xmm3 ; 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: vpunpcklwd {{.*#+}} xmm1 = xmm2[0],mem[0],xmm2[1],mem[1],xmm2[2],mem[2],xmm2[3],mem[3] ; AVX512BW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm3[0],mem[0],xmm3[1],mem[1],xmm3[2],mem[2],xmm3[3],mem[3] ; AVX512BW-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1] ; AVX512BW-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] ; AVX512BW-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,4,8,12,1,5,9,13,2,6,10,14,3,7,11,15] ; AVX512BW-NEXT: vmovdqa %xmm0, (%rax) ; AVX512BW-NEXT: retq ; ; AVX512BW-FCP-LABEL: store_i8_stride8_vf2: ; AVX512BW-FCP: # %bb.0: ; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r11 ; AVX512BW-FCP-NEXT: vmovdqa (%rdi), %xmm0 ; AVX512BW-FCP-NEXT: vmovdqa (%rdx), %xmm1 ; AVX512BW-FCP-NEXT: vmovdqa (%r8), %xmm2 ; AVX512BW-FCP-NEXT: vmovdqa (%r11), %xmm3 ; 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: vpunpcklwd {{.*#+}} xmm1 = xmm2[0],mem[0],xmm2[1],mem[1],xmm2[2],mem[2],xmm2[3],mem[3] ; AVX512BW-FCP-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm3[0],mem[0],xmm3[1],mem[1],xmm3[2],mem[2],xmm3[3],mem[3] ; AVX512BW-FCP-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1] ; AVX512BW-FCP-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] ; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,4,8,12,1,5,9,13,2,6,10,14,3,7,11,15] ; AVX512BW-FCP-NEXT: vmovdqa %xmm0, (%rax) ; AVX512BW-FCP-NEXT: retq ; ; AVX512DQ-BW-LABEL: store_i8_stride8_vf2: ; AVX512DQ-BW: # %bb.0: ; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %r11 ; AVX512DQ-BW-NEXT: vmovdqa (%rdi), %xmm0 ; AVX512DQ-BW-NEXT: vmovdqa (%rdx), %xmm1 ; AVX512DQ-BW-NEXT: vmovdqa (%r8), %xmm2 ; AVX512DQ-BW-NEXT: vmovdqa (%r11), %xmm3 ; 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: vpunpcklwd {{.*#+}} xmm1 = xmm2[0],mem[0],xmm2[1],mem[1],xmm2[2],mem[2],xmm2[3],mem[3] ; AVX512DQ-BW-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm3[0],mem[0],xmm3[1],mem[1],xmm3[2],mem[2],xmm3[3],mem[3] ; AVX512DQ-BW-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1] ; AVX512DQ-BW-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-BW-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,4,8,12,1,5,9,13,2,6,10,14,3,7,11,15] ; AVX512DQ-BW-NEXT: vmovdqa %xmm0, (%rax) ; AVX512DQ-BW-NEXT: retq ; ; AVX512DQ-BW-FCP-LABEL: store_i8_stride8_vf2: ; AVX512DQ-BW-FCP: # %bb.0: ; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r11 ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdi), %xmm0 ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdx), %xmm1 ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%r8), %xmm2 ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%r11), %xmm3 ; 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: vpunpcklwd {{.*#+}} xmm1 = xmm2[0],mem[0],xmm2[1],mem[1],xmm2[2],mem[2],xmm2[3],mem[3] ; AVX512DQ-BW-FCP-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm3[0],mem[0],xmm3[1],mem[1],xmm3[2],mem[2],xmm3[3],mem[3] ; AVX512DQ-BW-FCP-NEXT: vpunpckldq {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1] ; AVX512DQ-BW-FCP-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-BW-FCP-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,4,8,12,1,5,9,13,2,6,10,14,3,7,11,15] ; AVX512DQ-BW-FCP-NEXT: vmovdqa %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 %in.vec6 = load <2 x i8>, ptr %in.vecptr6, align 64 %in.vec7 = load <2 x i8>, ptr %in.vecptr7, 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 <2 x i8> %in.vec6, <2 x i8> %in.vec7, <4 x i32> %5 = shufflevector <4 x i8> %1, <4 x i8> %2, <8 x i32> %6 = shufflevector <4 x i8> %3, <4 x i8> %4, <8 x i32> %7 = shufflevector <8 x i8> %5, <8 x i8> %6, <16 x i32> %interleaved.vec = shufflevector <16 x i8> %7, <16 x i8> poison, <16 x i32> store <16 x i8> %interleaved.vec, ptr %out.vec, align 64 ret void } define void @store_i8_stride8_vf4(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecptr2, ptr %in.vecptr3, ptr %in.vecptr4, ptr %in.vecptr5, ptr %in.vecptr6, ptr %in.vecptr7, ptr %out.vec) nounwind { ; SSE-LABEL: store_i8_stride8_vf4: ; SSE: # %bb.0: ; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax ; SSE-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; SSE-NEXT: movq {{[0-9]+}}(%rsp), %r11 ; SSE-NEXT: movdqa (%rdi), %xmm0 ; SSE-NEXT: movdqa (%rdx), %xmm1 ; SSE-NEXT: movdqa (%r8), %xmm2 ; SSE-NEXT: movdqa (%r11), %xmm3 ; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1] ; 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: punpckldq {{.*#+}} xmm3 = xmm3[0],mem[0],xmm3[1],mem[1] ; SSE-NEXT: pxor %xmm6, %xmm6 ; SSE-NEXT: punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm6[0],xmm3[1],xmm6[1],xmm3[2],xmm6[2],xmm3[3],xmm6[3],xmm3[4],xmm6[4],xmm3[5],xmm6[5],xmm3[6],xmm6[6],xmm3[7],xmm6[7] ; SSE-NEXT: pshufd {{.*#+}} xmm4 = xmm3[0,1,2,0] ; SSE-NEXT: pshufhw {{.*#+}} xmm5 = xmm4[0,1,2,3,4,5,7,5] ; SSE-NEXT: pshufhw {{.*#+}} xmm7 = xmm4[0,1,2,3,4,5,6,4] ; SSE-NEXT: packuswb %xmm5, %xmm7 ; SSE-NEXT: movdqa {{.*#+}} xmm4 = [65535,65535,65535,0,65535,65535,65535,0] ; SSE-NEXT: movdqa %xmm4, %xmm5 ; SSE-NEXT: pandn %xmm7, %xmm5 ; 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 {{.*#+}} xmm7 = xmm2[0,1,2,0] ; SSE-NEXT: pshufhw {{.*#+}} xmm8 = xmm7[0,1,2,3,7,5,6,7] ; SSE-NEXT: pshufhw {{.*#+}} xmm7 = xmm7[0,1,2,3,6,4,6,7] ; SSE-NEXT: packuswb %xmm8, %xmm7 ; SSE-NEXT: pand %xmm4, %xmm7 ; SSE-NEXT: por %xmm5, %xmm7 ; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm7[1,3,2,3] ; SSE-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm6[0],xmm1[1],xmm6[1],xmm1[2],xmm6[2],xmm1[3],xmm6[3],xmm1[4],xmm6[4],xmm1[5],xmm6[5],xmm1[6],xmm6[6],xmm1[7],xmm6[7] ; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm1[0,2,2,3] ; SSE-NEXT: pshuflw {{.*#+}} xmm8 = xmm5[0,1,1,3,4,5,6,7] ; SSE-NEXT: pshuflw {{.*#+}} xmm9 = xmm5[0,1,0,2,4,5,6,7] ; SSE-NEXT: packuswb %xmm8, %xmm9 ; SSE-NEXT: movdqa {{.*#+}} xmm5 = [65535,0,65535,65535,65535,0,65535,65535] ; SSE-NEXT: movdqa %xmm5, %xmm8 ; SSE-NEXT: pandn %xmm9, %xmm8 ; SSE-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm6[0],xmm0[1],xmm6[1],xmm0[2],xmm6[2],xmm0[3],xmm6[3],xmm0[4],xmm6[4],xmm0[5],xmm6[5],xmm0[6],xmm6[6],xmm0[7],xmm6[7] ; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm0[0,2,2,3] ; SSE-NEXT: pshuflw {{.*#+}} xmm9 = xmm6[1,3,2,3,4,5,6,7] ; SSE-NEXT: pshuflw {{.*#+}} xmm6 = xmm6[0,2,2,3,4,5,6,7] ; SSE-NEXT: packuswb %xmm9, %xmm6 ; SSE-NEXT: pand %xmm5, %xmm6 ; SSE-NEXT: por %xmm8, %xmm6 ; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm6[0,2,2,3] ; SSE-NEXT: punpckldq {{.*#+}} xmm6 = xmm6[0],xmm7[0],xmm6[1],xmm7[1] ; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm3[0,1,1,3] ; SSE-NEXT: pshufhw {{.*#+}} xmm7 = xmm3[0,1,2,3,4,5,5,7] ; SSE-NEXT: pshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,5,4,6] ; SSE-NEXT: packuswb %xmm7, %xmm3 ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm2[0,1,1,3] ; SSE-NEXT: pshufhw {{.*#+}} xmm7 = xmm2[0,1,2,3,5,7,6,7] ; SSE-NEXT: pshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,6,6,7] ; SSE-NEXT: packuswb %xmm7, %xmm2 ; SSE-NEXT: pand %xmm4, %xmm2 ; SSE-NEXT: pandn %xmm3, %xmm4 ; SSE-NEXT: por %xmm2, %xmm4 ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm4[1,3,2,3] ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[3,1,2,3] ; SSE-NEXT: pshuflw {{.*#+}} xmm3 = xmm1[0,1,3,1,4,5,6,7] ; SSE-NEXT: pshuflw {{.*#+}} xmm1 = xmm1[0,1,2,0,4,5,6,7] ; SSE-NEXT: packuswb %xmm3, %xmm1 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[3,1,2,3] ; SSE-NEXT: pshuflw {{.*#+}} xmm3 = xmm0[3,1,2,3,4,5,6,7] ; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[2,0,2,3,4,5,6,7] ; SSE-NEXT: packuswb %xmm3, %xmm0 ; SSE-NEXT: pand %xmm5, %xmm0 ; SSE-NEXT: pandn %xmm1, %xmm5 ; SSE-NEXT: por %xmm0, %xmm5 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm5[0,2,2,3] ; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1] ; SSE-NEXT: movdqa %xmm0, 16(%rax) ; SSE-NEXT: movdqa %xmm6, (%rax) ; SSE-NEXT: retq ; ; AVX-LABEL: store_i8_stride8_vf4: ; AVX: # %bb.0: ; AVX-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX-NEXT: movq {{[0-9]+}}(%rsp), %r11 ; AVX-NEXT: vmovdqa (%rdi), %xmm0 ; AVX-NEXT: vmovdqa (%rdx), %xmm1 ; AVX-NEXT: vmovdqa (%r8), %xmm2 ; AVX-NEXT: vmovdqa (%r11), %xmm3 ; 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: vpunpckldq {{.*#+}} xmm2 = xmm3[0],mem[0],xmm3[1],mem[1] ; AVX-NEXT: vpunpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm2[0] ; AVX-NEXT: vmovq {{.*#+}} xmm2 = [2,6,10,14,3,7,11,15,0,0,0,0,0,0,0,0] ; AVX-NEXT: vpshufb %xmm2, %xmm1, %xmm3 ; AVX-NEXT: vpshufb %xmm2, %xmm0, %xmm2 ; AVX-NEXT: vpunpckldq {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1] ; AVX-NEXT: vmovq {{.*#+}} xmm3 = [0,4,8,12,1,5,9,13,0,0,0,0,0,0,0,0] ; AVX-NEXT: vpshufb %xmm3, %xmm1, %xmm1 ; AVX-NEXT: vpshufb %xmm3, %xmm0, %xmm0 ; AVX-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] ; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm0 ; AVX-NEXT: vmovaps %ymm0, (%rax) ; AVX-NEXT: vzeroupper ; AVX-NEXT: retq ; ; AVX2-LABEL: store_i8_stride8_vf4: ; AVX2: # %bb.0: ; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %r11 ; AVX2-NEXT: vmovdqa (%rdi), %xmm0 ; AVX2-NEXT: vmovdqa (%rdx), %xmm1 ; AVX2-NEXT: vmovdqa (%r8), %xmm2 ; AVX2-NEXT: vmovdqa (%r11), %xmm3 ; 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: vpunpckldq {{.*#+}} xmm2 = xmm2[0],mem[0],xmm2[1],mem[1] ; AVX2-NEXT: vpunpckldq {{.*#+}} xmm3 = xmm3[0],mem[0],xmm3[1],mem[1] ; AVX2-NEXT: vinserti128 $1, %xmm3, %ymm1, %ymm1 ; AVX2-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm0 ; AVX2-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2] ; AVX2-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12,1,5,9,13,2,6,10,14,3,7,11,15,16,20,24,28,17,21,25,29,18,22,26,30,19,23,27,31] ; AVX2-NEXT: vpmovsxbd {{.*#+}} ymm1 = [0,4,1,5,2,6,3,7] ; AVX2-NEXT: vpermd %ymm0, %ymm1, %ymm0 ; AVX2-NEXT: vmovdqa %ymm0, (%rax) ; AVX2-NEXT: vzeroupper ; AVX2-NEXT: retq ; ; AVX2-FP-LABEL: store_i8_stride8_vf4: ; AVX2-FP: # %bb.0: ; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %r11 ; AVX2-FP-NEXT: vmovdqa (%rdi), %xmm0 ; AVX2-FP-NEXT: vmovdqa (%rdx), %xmm1 ; AVX2-FP-NEXT: vmovdqa (%r8), %xmm2 ; AVX2-FP-NEXT: vmovdqa (%r11), %xmm3 ; 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: vpunpckldq {{.*#+}} xmm2 = xmm2[0],mem[0],xmm2[1],mem[1] ; AVX2-FP-NEXT: vpunpckldq {{.*#+}} xmm3 = xmm3[0],mem[0],xmm3[1],mem[1] ; AVX2-FP-NEXT: vinserti128 $1, %xmm3, %ymm1, %ymm1 ; AVX2-FP-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm0 ; AVX2-FP-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2] ; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12,1,5,9,13,2,6,10,14,3,7,11,15,16,20,24,28,17,21,25,29,18,22,26,30,19,23,27,31] ; AVX2-FP-NEXT: vpmovsxbd {{.*#+}} ymm1 = [0,4,1,5,2,6,3,7] ; AVX2-FP-NEXT: vpermd %ymm0, %ymm1, %ymm0 ; AVX2-FP-NEXT: vmovdqa %ymm0, (%rax) ; AVX2-FP-NEXT: vzeroupper ; AVX2-FP-NEXT: retq ; ; AVX2-FCP-LABEL: store_i8_stride8_vf4: ; AVX2-FCP: # %bb.0: ; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r11 ; AVX2-FCP-NEXT: vmovdqa (%rdi), %xmm0 ; AVX2-FCP-NEXT: vmovdqa (%rdx), %xmm1 ; AVX2-FCP-NEXT: vmovdqa (%r8), %xmm2 ; AVX2-FCP-NEXT: vmovdqa (%r11), %xmm3 ; 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: vpunpckldq {{.*#+}} xmm2 = xmm2[0],mem[0],xmm2[1],mem[1] ; AVX2-FCP-NEXT: vpunpckldq {{.*#+}} xmm3 = xmm3[0],mem[0],xmm3[1],mem[1] ; AVX2-FCP-NEXT: vinserti128 $1, %xmm3, %ymm1, %ymm1 ; AVX2-FCP-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm0 ; AVX2-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2] ; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12,1,5,9,13,2,6,10,14,3,7,11,15,16,20,24,28,17,21,25,29,18,22,26,30,19,23,27,31] ; AVX2-FCP-NEXT: vpmovsxbd {{.*#+}} ymm1 = [0,4,1,5,2,6,3,7] ; AVX2-FCP-NEXT: vpermd %ymm0, %ymm1, %ymm0 ; AVX2-FCP-NEXT: vmovdqa %ymm0, (%rax) ; AVX2-FCP-NEXT: vzeroupper ; AVX2-FCP-NEXT: retq ; ; AVX512-LABEL: store_i8_stride8_vf4: ; AVX512: # %bb.0: ; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %r11 ; AVX512-NEXT: vmovdqa (%rdi), %xmm0 ; AVX512-NEXT: vmovdqa (%rdx), %xmm1 ; AVX512-NEXT: vmovdqa (%r8), %xmm2 ; AVX512-NEXT: vmovdqa (%r11), %xmm3 ; 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: vpunpckldq {{.*#+}} xmm2 = xmm2[0],mem[0],xmm2[1],mem[1] ; AVX512-NEXT: vpunpckldq {{.*#+}} xmm3 = xmm3[0],mem[0],xmm3[1],mem[1] ; AVX512-NEXT: vinserti128 $1, %xmm3, %ymm1, %ymm1 ; AVX512-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm0 ; AVX512-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2] ; AVX512-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12,1,5,9,13,2,6,10,14,3,7,11,15,16,20,24,28,17,21,25,29,18,22,26,30,19,23,27,31] ; AVX512-NEXT: vpmovsxbd {{.*#+}} ymm1 = [0,4,1,5,2,6,3,7] ; AVX512-NEXT: vpermd %ymm0, %ymm1, %ymm0 ; AVX512-NEXT: vmovdqa %ymm0, (%rax) ; AVX512-NEXT: vzeroupper ; AVX512-NEXT: retq ; ; AVX512-FCP-LABEL: store_i8_stride8_vf4: ; AVX512-FCP: # %bb.0: ; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r11 ; AVX512-FCP-NEXT: vmovdqa (%rdi), %xmm0 ; AVX512-FCP-NEXT: vmovdqa (%rdx), %xmm1 ; AVX512-FCP-NEXT: vmovdqa (%r8), %xmm2 ; AVX512-FCP-NEXT: vmovdqa (%r11), %xmm3 ; 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: vpunpckldq {{.*#+}} xmm2 = xmm2[0],mem[0],xmm2[1],mem[1] ; AVX512-FCP-NEXT: vpunpckldq {{.*#+}} xmm3 = xmm3[0],mem[0],xmm3[1],mem[1] ; AVX512-FCP-NEXT: vinserti128 $1, %xmm3, %ymm1, %ymm1 ; AVX512-FCP-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm0 ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2] ; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12,1,5,9,13,2,6,10,14,3,7,11,15,16,20,24,28,17,21,25,29,18,22,26,30,19,23,27,31] ; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm1 = [0,4,1,5,2,6,3,7] ; AVX512-FCP-NEXT: vpermd %ymm0, %ymm1, %ymm0 ; AVX512-FCP-NEXT: vmovdqa %ymm0, (%rax) ; AVX512-FCP-NEXT: vzeroupper ; AVX512-FCP-NEXT: retq ; ; AVX512DQ-LABEL: store_i8_stride8_vf4: ; AVX512DQ: # %bb.0: ; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %r11 ; AVX512DQ-NEXT: vmovdqa (%rdi), %xmm0 ; AVX512DQ-NEXT: vmovdqa (%rdx), %xmm1 ; AVX512DQ-NEXT: vmovdqa (%r8), %xmm2 ; AVX512DQ-NEXT: vmovdqa (%r11), %xmm3 ; 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: vpunpckldq {{.*#+}} xmm2 = xmm2[0],mem[0],xmm2[1],mem[1] ; AVX512DQ-NEXT: vpunpckldq {{.*#+}} xmm3 = xmm3[0],mem[0],xmm3[1],mem[1] ; AVX512DQ-NEXT: vinserti128 $1, %xmm3, %ymm1, %ymm1 ; AVX512DQ-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm0 ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2] ; AVX512DQ-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12,1,5,9,13,2,6,10,14,3,7,11,15,16,20,24,28,17,21,25,29,18,22,26,30,19,23,27,31] ; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} ymm1 = [0,4,1,5,2,6,3,7] ; AVX512DQ-NEXT: vpermd %ymm0, %ymm1, %ymm0 ; AVX512DQ-NEXT: vmovdqa %ymm0, (%rax) ; AVX512DQ-NEXT: vzeroupper ; AVX512DQ-NEXT: retq ; ; AVX512DQ-FCP-LABEL: store_i8_stride8_vf4: ; AVX512DQ-FCP: # %bb.0: ; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r11 ; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %xmm0 ; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %xmm1 ; AVX512DQ-FCP-NEXT: vmovdqa (%r8), %xmm2 ; AVX512DQ-FCP-NEXT: vmovdqa (%r11), %xmm3 ; 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: vpunpckldq {{.*#+}} xmm2 = xmm2[0],mem[0],xmm2[1],mem[1] ; AVX512DQ-FCP-NEXT: vpunpckldq {{.*#+}} xmm3 = xmm3[0],mem[0],xmm3[1],mem[1] ; AVX512DQ-FCP-NEXT: vinserti128 $1, %xmm3, %ymm1, %ymm1 ; AVX512DQ-FCP-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm0 ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2] ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12,1,5,9,13,2,6,10,14,3,7,11,15,16,20,24,28,17,21,25,29,18,22,26,30,19,23,27,31] ; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm1 = [0,4,1,5,2,6,3,7] ; AVX512DQ-FCP-NEXT: vpermd %ymm0, %ymm1, %ymm0 ; AVX512DQ-FCP-NEXT: vmovdqa %ymm0, (%rax) ; AVX512DQ-FCP-NEXT: vzeroupper ; AVX512DQ-FCP-NEXT: retq ; ; AVX512BW-LABEL: store_i8_stride8_vf4: ; AVX512BW: # %bb.0: ; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %r11 ; AVX512BW-NEXT: vmovdqa (%rdi), %xmm0 ; AVX512BW-NEXT: vmovdqa (%rdx), %xmm1 ; AVX512BW-NEXT: vmovdqa (%r8), %xmm2 ; AVX512BW-NEXT: vmovdqa (%r11), %xmm3 ; 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: vpunpckldq {{.*#+}} xmm2 = xmm2[0],mem[0],xmm2[1],mem[1] ; AVX512BW-NEXT: vpunpckldq {{.*#+}} xmm3 = xmm3[0],mem[0],xmm3[1],mem[1] ; AVX512BW-NEXT: vinserti128 $1, %xmm3, %ymm1, %ymm1 ; AVX512BW-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm0 ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2] ; AVX512BW-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12,1,5,9,13,2,6,10,14,3,7,11,15,16,20,24,28,17,21,25,29,18,22,26,30,19,23,27,31] ; AVX512BW-NEXT: vpmovsxbd {{.*#+}} ymm1 = [0,4,1,5,2,6,3,7] ; AVX512BW-NEXT: vpermd %ymm0, %ymm1, %ymm0 ; AVX512BW-NEXT: vmovdqa %ymm0, (%rax) ; AVX512BW-NEXT: vzeroupper ; AVX512BW-NEXT: retq ; ; AVX512BW-FCP-LABEL: store_i8_stride8_vf4: ; AVX512BW-FCP: # %bb.0: ; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r11 ; AVX512BW-FCP-NEXT: vmovdqa (%rdi), %xmm0 ; AVX512BW-FCP-NEXT: vmovdqa (%rdx), %xmm1 ; AVX512BW-FCP-NEXT: vmovdqa (%r8), %xmm2 ; AVX512BW-FCP-NEXT: vmovdqa (%r11), %xmm3 ; 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: vpunpckldq {{.*#+}} xmm2 = xmm2[0],mem[0],xmm2[1],mem[1] ; AVX512BW-FCP-NEXT: vpunpckldq {{.*#+}} xmm3 = xmm3[0],mem[0],xmm3[1],mem[1] ; AVX512BW-FCP-NEXT: vinserti128 $1, %xmm3, %ymm1, %ymm1 ; AVX512BW-FCP-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm0 ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2] ; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12,1,5,9,13,2,6,10,14,3,7,11,15,16,20,24,28,17,21,25,29,18,22,26,30,19,23,27,31] ; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm1 = [0,4,1,5,2,6,3,7] ; AVX512BW-FCP-NEXT: vpermd %ymm0, %ymm1, %ymm0 ; AVX512BW-FCP-NEXT: vmovdqa %ymm0, (%rax) ; AVX512BW-FCP-NEXT: vzeroupper ; AVX512BW-FCP-NEXT: retq ; ; AVX512DQ-BW-LABEL: store_i8_stride8_vf4: ; AVX512DQ-BW: # %bb.0: ; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %r11 ; AVX512DQ-BW-NEXT: vmovdqa (%rdi), %xmm0 ; AVX512DQ-BW-NEXT: vmovdqa (%rdx), %xmm1 ; AVX512DQ-BW-NEXT: vmovdqa (%r8), %xmm2 ; AVX512DQ-BW-NEXT: vmovdqa (%r11), %xmm3 ; 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: vpunpckldq {{.*#+}} xmm2 = xmm2[0],mem[0],xmm2[1],mem[1] ; AVX512DQ-BW-NEXT: vpunpckldq {{.*#+}} xmm3 = xmm3[0],mem[0],xmm3[1],mem[1] ; AVX512DQ-BW-NEXT: vinserti128 $1, %xmm3, %ymm1, %ymm1 ; AVX512DQ-BW-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm0 ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2] ; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12,1,5,9,13,2,6,10,14,3,7,11,15,16,20,24,28,17,21,25,29,18,22,26,30,19,23,27,31] ; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} ymm1 = [0,4,1,5,2,6,3,7] ; AVX512DQ-BW-NEXT: vpermd %ymm0, %ymm1, %ymm0 ; AVX512DQ-BW-NEXT: vmovdqa %ymm0, (%rax) ; AVX512DQ-BW-NEXT: vzeroupper ; AVX512DQ-BW-NEXT: retq ; ; AVX512DQ-BW-FCP-LABEL: store_i8_stride8_vf4: ; AVX512DQ-BW-FCP: # %bb.0: ; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r11 ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdi), %xmm0 ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdx), %xmm1 ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%r8), %xmm2 ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%r11), %xmm3 ; 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: vpunpckldq {{.*#+}} xmm2 = xmm2[0],mem[0],xmm2[1],mem[1] ; AVX512DQ-BW-FCP-NEXT: vpunpckldq {{.*#+}} xmm3 = xmm3[0],mem[0],xmm3[1],mem[1] ; AVX512DQ-BW-FCP-NEXT: vinserti128 $1, %xmm3, %ymm1, %ymm1 ; AVX512DQ-BW-FCP-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm0 ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[2],ymm1[2] ; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12,1,5,9,13,2,6,10,14,3,7,11,15,16,20,24,28,17,21,25,29,18,22,26,30,19,23,27,31] ; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm1 = [0,4,1,5,2,6,3,7] ; AVX512DQ-BW-FCP-NEXT: vpermd %ymm0, %ymm1, %ymm0 ; AVX512DQ-BW-FCP-NEXT: vmovdqa %ymm0, (%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 %in.vec6 = load <4 x i8>, ptr %in.vecptr6, align 64 %in.vec7 = load <4 x i8>, ptr %in.vecptr7, 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 <4 x i8> %in.vec6, <4 x i8> %in.vec7, <8 x i32> %5 = shufflevector <8 x i8> %1, <8 x i8> %2, <16 x i32> %6 = shufflevector <8 x i8> %3, <8 x i8> %4, <16 x i32> %7 = shufflevector <16 x i8> %5, <16 x i8> %6, <32 x i32> %interleaved.vec = shufflevector <32 x i8> %7, <32 x i8> poison, <32 x i32> store <32 x i8> %interleaved.vec, ptr %out.vec, align 64 ret void } define void @store_i8_stride8_vf8(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecptr2, ptr %in.vecptr3, ptr %in.vecptr4, ptr %in.vecptr5, ptr %in.vecptr6, ptr %in.vecptr7, ptr %out.vec) nounwind { ; SSE-LABEL: store_i8_stride8_vf8: ; SSE: # %bb.0: ; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax ; SSE-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; SSE-NEXT: movq {{[0-9]+}}(%rsp), %r11 ; SSE-NEXT: movq {{.*#+}} xmm0 = mem[0],zero ; SSE-NEXT: movq {{.*#+}} xmm1 = mem[0],zero ; SSE-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3],xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7] ; SSE-NEXT: movq {{.*#+}} xmm1 = mem[0],zero ; SSE-NEXT: movq {{.*#+}} xmm2 = mem[0],zero ; SSE-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3],xmm1[4],xmm2[4],xmm1[5],xmm2[5],xmm1[6],xmm2[6],xmm1[7],xmm2[7] ; SSE-NEXT: movq {{.*#+}} xmm2 = mem[0],zero ; SSE-NEXT: movq {{.*#+}} xmm3 = mem[0],zero ; 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: movq {{.*#+}} xmm5 = mem[0],zero ; SSE-NEXT: movq {{.*#+}} xmm3 = mem[0],zero ; SSE-NEXT: punpcklbw {{.*#+}} xmm5 = xmm5[0],xmm3[0],xmm5[1],xmm3[1],xmm5[2],xmm3[2],xmm5[3],xmm3[3],xmm5[4],xmm3[4],xmm5[5],xmm3[5],xmm5[6],xmm3[6],xmm5[7],xmm3[7] ; SSE-NEXT: pshuflw {{.*#+}} xmm3 = xmm5[0,0,2,1,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm4 = xmm3[0,0,2,1] ; SSE-NEXT: movdqa {{.*#+}} xmm3 = [65535,65535,65535,0,65535,65535,65535,0] ; SSE-NEXT: movdqa %xmm3, %xmm6 ; SSE-NEXT: pandn %xmm4, %xmm6 ; SSE-NEXT: pshuflw {{.*#+}} xmm4 = xmm2[0,1,1,3,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm4 = xmm4[0,0,2,1] ; SSE-NEXT: pand %xmm3, %xmm4 ; SSE-NEXT: por %xmm6, %xmm4 ; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm4[1,3,2,3] ; SSE-NEXT: pshuflw {{.*#+}} xmm4 = xmm1[0,0,2,1,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm4[0,1,1,3] ; SSE-NEXT: movdqa {{.*#+}} xmm4 = [65535,0,65535,65535,65535,0,65535,65535] ; SSE-NEXT: movdqa %xmm4, %xmm8 ; SSE-NEXT: pandn %xmm6, %xmm8 ; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm0[0,0,0,0] ; SSE-NEXT: pshufhw {{.*#+}} xmm6 = xmm6[0,1,2,3,5,5,5,5] ; SSE-NEXT: pand %xmm4, %xmm6 ; SSE-NEXT: por %xmm8, %xmm6 ; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm6[0,2,2,3] ; SSE-NEXT: punpckldq {{.*#+}} xmm6 = xmm6[0],xmm7[0],xmm6[1],xmm7[1] ; SSE-NEXT: pshuflw {{.*#+}} xmm7 = xmm5[0,2,2,3,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm7[0,0,2,1] ; SSE-NEXT: movdqa %xmm3, %xmm8 ; SSE-NEXT: pandn %xmm7, %xmm8 ; SSE-NEXT: pshuflw {{.*#+}} xmm7 = xmm2[2,1,3,3,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm7[0,0,2,1] ; SSE-NEXT: pand %xmm3, %xmm7 ; SSE-NEXT: por %xmm8, %xmm7 ; SSE-NEXT: pshufd {{.*#+}} xmm8 = xmm7[1,3,2,3] ; SSE-NEXT: pshuflw {{.*#+}} xmm7 = xmm1[0,2,2,3,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm7[0,1,1,3] ; SSE-NEXT: movdqa %xmm4, %xmm9 ; SSE-NEXT: pandn %xmm7, %xmm9 ; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm0[1,1,1,1] ; SSE-NEXT: pshufhw {{.*#+}} xmm7 = xmm7[0,1,2,3,5,5,5,5] ; SSE-NEXT: pand %xmm4, %xmm7 ; SSE-NEXT: por %xmm9, %xmm7 ; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm7[0,2,2,3] ; SSE-NEXT: punpckldq {{.*#+}} xmm7 = xmm7[0],xmm8[0],xmm7[1],xmm8[1] ; SSE-NEXT: pshufhw {{.*#+}} xmm8 = xmm5[0,1,2,3,4,4,6,5] ; SSE-NEXT: pshufd {{.*#+}} xmm8 = xmm8[0,2,2,3] ; SSE-NEXT: movdqa %xmm3, %xmm9 ; SSE-NEXT: pandn %xmm8, %xmm9 ; SSE-NEXT: pshufhw {{.*#+}} xmm8 = xmm2[0,1,2,3,4,5,5,7] ; SSE-NEXT: pshufd {{.*#+}} xmm8 = xmm8[0,2,2,3] ; SSE-NEXT: pand %xmm3, %xmm8 ; SSE-NEXT: por %xmm9, %xmm8 ; SSE-NEXT: pshufd {{.*#+}} xmm8 = xmm8[1,3,2,3] ; SSE-NEXT: pshufhw {{.*#+}} xmm9 = xmm1[0,1,2,3,4,4,6,5] ; SSE-NEXT: pshufd {{.*#+}} xmm9 = xmm9[2,1,3,3] ; SSE-NEXT: movdqa %xmm4, %xmm10 ; SSE-NEXT: pandn %xmm9, %xmm10 ; SSE-NEXT: pshufd {{.*#+}} xmm9 = xmm0[2,2,2,2] ; SSE-NEXT: pshufhw {{.*#+}} xmm9 = xmm9[0,1,2,3,5,5,5,5] ; SSE-NEXT: pand %xmm4, %xmm9 ; SSE-NEXT: por %xmm10, %xmm9 ; SSE-NEXT: pshufd {{.*#+}} xmm9 = xmm9[0,2,2,3] ; SSE-NEXT: punpckldq {{.*#+}} xmm9 = xmm9[0],xmm8[0],xmm9[1],xmm8[1] ; SSE-NEXT: pshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,4,6,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm5[0,2,2,3] ; SSE-NEXT: pshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,6,5,7,7] ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm2[0,2,2,3] ; SSE-NEXT: pand %xmm3, %xmm2 ; SSE-NEXT: pandn %xmm5, %xmm3 ; SSE-NEXT: por %xmm2, %xmm3 ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm3[1,3,2,3] ; SSE-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,6,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[2,1,3,3] ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[3,3,3,3] ; SSE-NEXT: pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,5,5,5,5] ; SSE-NEXT: pand %xmm4, %xmm0 ; SSE-NEXT: pandn %xmm1, %xmm4 ; SSE-NEXT: por %xmm0, %xmm4 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm4[0,2,2,3] ; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1] ; SSE-NEXT: movdqa %xmm0, 48(%rax) ; SSE-NEXT: movdqa %xmm9, 32(%rax) ; SSE-NEXT: movdqa %xmm7, 16(%rax) ; SSE-NEXT: movdqa %xmm6, (%rax) ; SSE-NEXT: retq ; ; AVX-LABEL: store_i8_stride8_vf8: ; AVX: # %bb.0: ; AVX-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX-NEXT: movq {{[0-9]+}}(%rsp), %r11 ; AVX-NEXT: 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: vmovq {{.*#+}} xmm3 = mem[0],zero ; AVX-NEXT: vmovq {{.*#+}} xmm4 = mem[0],zero ; AVX-NEXT: vpunpcklqdq {{.*#+}} xmm3 = xmm4[0],xmm3[0] ; AVX-NEXT: vmovq {{.*#+}} xmm4 = [0,0,2,10,0,0,3,11,0,0,0,0,0,0,0,0] ; AVX-NEXT: vpshufb %xmm4, %xmm3, %xmm5 ; AVX-NEXT: vpshufb %xmm4, %xmm2, %xmm4 ; AVX-NEXT: vpunpcklwd {{.*#+}} xmm4 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3] ; AVX-NEXT: vmovq {{.*#+}} xmm5 = [2,10,0,0,3,11,0,0,0,0,0,0,0,0,0,0] ; AVX-NEXT: vpshufb %xmm5, %xmm1, %xmm6 ; AVX-NEXT: vpshufb %xmm5, %xmm0, %xmm5 ; AVX-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3] ; AVX-NEXT: vpblendw {{.*#+}} xmm4 = xmm5[0,1],xmm4[2,3],xmm5[4,5],xmm4[6,7] ; AVX-NEXT: vmovq {{.*#+}} xmm5 = [0,0,0,8,0,0,1,9,0,0,0,0,0,0,0,0] ; AVX-NEXT: vpshufb %xmm5, %xmm3, %xmm6 ; AVX-NEXT: vpshufb %xmm5, %xmm2, %xmm5 ; AVX-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3] ; AVX-NEXT: vmovq {{.*#+}} xmm6 = [0,8,0,0,1,9,0,0,0,0,0,0,0,0,0,0] ; AVX-NEXT: vpshufb %xmm6, %xmm1, %xmm7 ; AVX-NEXT: vpshufb %xmm6, %xmm0, %xmm6 ; AVX-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm6[0],xmm7[0],xmm6[1],xmm7[1],xmm6[2],xmm7[2],xmm6[3],xmm7[3] ; AVX-NEXT: vpblendw {{.*#+}} xmm5 = xmm6[0,1],xmm5[2,3],xmm6[4,5],xmm5[6,7] ; AVX-NEXT: vinsertf128 $1, %xmm4, %ymm5, %ymm4 ; AVX-NEXT: vmovq {{.*#+}} xmm5 = [0,0,6,14,0,0,7,15,0,0,0,0,0,0,0,0] ; AVX-NEXT: vpshufb %xmm5, %xmm3, %xmm6 ; AVX-NEXT: vpshufb %xmm5, %xmm2, %xmm5 ; AVX-NEXT: vpunpcklwd {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3] ; AVX-NEXT: vmovq {{.*#+}} xmm6 = [6,14,0,0,7,15,0,0,0,0,0,0,0,0,0,0] ; AVX-NEXT: vpshufb %xmm6, %xmm1, %xmm7 ; AVX-NEXT: vpshufb %xmm6, %xmm0, %xmm6 ; AVX-NEXT: vpunpcklwd {{.*#+}} xmm6 = xmm6[0],xmm7[0],xmm6[1],xmm7[1],xmm6[2],xmm7[2],xmm6[3],xmm7[3] ; AVX-NEXT: vpblendw {{.*#+}} xmm5 = xmm6[0,1],xmm5[2,3],xmm6[4,5],xmm5[6,7] ; AVX-NEXT: vmovq {{.*#+}} xmm6 = [0,0,4,12,0,0,5,13,0,0,0,0,0,0,0,0] ; AVX-NEXT: vpshufb %xmm6, %xmm3, %xmm3 ; AVX-NEXT: vpshufb %xmm6, %xmm2, %xmm2 ; AVX-NEXT: vpunpcklwd {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3] ; AVX-NEXT: vmovq {{.*#+}} xmm3 = [4,12,0,0,5,13,0,0,0,0,0,0,0,0,0,0] ; AVX-NEXT: vpshufb %xmm3, %xmm1, %xmm1 ; AVX-NEXT: vpshufb %xmm3, %xmm0, %xmm0 ; AVX-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3] ; AVX-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0,1],xmm2[2,3],xmm0[4,5],xmm2[6,7] ; AVX-NEXT: vinsertf128 $1, %xmm5, %ymm0, %ymm0 ; AVX-NEXT: vmovaps %ymm0, 32(%rax) ; AVX-NEXT: vmovaps %ymm4, (%rax) ; AVX-NEXT: vzeroupper ; AVX-NEXT: retq ; ; AVX2-LABEL: store_i8_stride8_vf8: ; AVX2: # %bb.0: ; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %r11 ; AVX2-NEXT: 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: vmovq {{.*#+}} xmm2 = mem[0],zero ; AVX2-NEXT: vmovq {{.*#+}} xmm3 = mem[0],zero ; AVX2-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm3[0],xmm2[0] ; AVX2-NEXT: vmovq {{.*#+}} xmm3 = mem[0],zero ; AVX2-NEXT: vmovq {{.*#+}} xmm4 = mem[0],zero ; AVX2-NEXT: vpunpcklqdq {{.*#+}} xmm3 = xmm4[0],xmm3[0] ; AVX2-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm4 ; AVX2-NEXT: vinserti128 $1, %xmm3, %ymm2, %ymm5 ; AVX2-NEXT: vinserti128 $1, %xmm2, %ymm3, %ymm2 ; AVX2-NEXT: vpshufb {{.*#+}} ymm3 = zero,zero,zero,zero,zero,zero,ymm2[0,8],zero,zero,zero,zero,zero,zero,ymm2[1,9],zero,zero,zero,zero,ymm2[18,26],zero,zero,zero,zero,zero,zero,ymm2[19,27],zero,zero ; AVX2-NEXT: vpshufb {{.*#+}} ymm6 = zero,zero,zero,zero,ymm5[0,8],zero,zero,zero,zero,zero,zero,ymm5[1,9],zero,zero,zero,zero,zero,zero,zero,zero,ymm5[18,26],zero,zero,zero,zero,zero,zero,ymm5[19,27] ; AVX2-NEXT: vpor %ymm3, %ymm6, %ymm3 ; AVX2-NEXT: vinserti128 $1, %xmm0, %ymm1, %ymm0 ; AVX2-NEXT: vpshufb {{.*#+}} ymm1 = zero,zero,ymm0[0,8],zero,zero,zero,zero,zero,zero,ymm0[1,9],zero,zero,zero,zero,ymm0[18,26],zero,zero,zero,zero,zero,zero,ymm0[19,27],zero,zero,zero,zero,zero,zero ; AVX2-NEXT: vpshufb {{.*#+}} ymm6 = ymm4[0,8],zero,zero,zero,zero,zero,zero,ymm4[1,9],zero,zero,zero,zero,zero,zero,zero,zero,ymm4[18,26],zero,zero,zero,zero,zero,zero,ymm4[19,27],zero,zero,zero,zero ; AVX2-NEXT: vpor %ymm1, %ymm6, %ymm1 ; AVX2-NEXT: vpor %ymm3, %ymm1, %ymm1 ; AVX2-NEXT: vpshufb {{.*#+}} ymm2 = zero,zero,zero,zero,zero,zero,ymm2[4,12],zero,zero,zero,zero,zero,zero,ymm2[5,13],zero,zero,zero,zero,ymm2[22,30],zero,zero,zero,zero,zero,zero,ymm2[23,31],zero,zero ; AVX2-NEXT: vpshufb {{.*#+}} ymm3 = zero,zero,zero,zero,ymm5[4,12],zero,zero,zero,zero,zero,zero,ymm5[5,13],zero,zero,zero,zero,zero,zero,zero,zero,ymm5[22,30],zero,zero,zero,zero,zero,zero,ymm5[23,31] ; AVX2-NEXT: vpor %ymm2, %ymm3, %ymm2 ; AVX2-NEXT: vpshufb {{.*#+}} ymm0 = zero,zero,ymm0[4,12],zero,zero,zero,zero,zero,zero,ymm0[5,13],zero,zero,zero,zero,ymm0[22,30],zero,zero,zero,zero,zero,zero,ymm0[23,31],zero,zero,zero,zero,zero,zero ; AVX2-NEXT: vpshufb {{.*#+}} ymm3 = ymm4[4,12],zero,zero,zero,zero,zero,zero,ymm4[5,13],zero,zero,zero,zero,zero,zero,zero,zero,ymm4[22,30],zero,zero,zero,zero,zero,zero,ymm4[23,31],zero,zero,zero,zero ; AVX2-NEXT: vpor %ymm0, %ymm3, %ymm0 ; AVX2-NEXT: vpor %ymm2, %ymm0, %ymm0 ; AVX2-NEXT: vmovdqa %ymm0, 32(%rax) ; AVX2-NEXT: vmovdqa %ymm1, (%rax) ; AVX2-NEXT: vzeroupper ; AVX2-NEXT: retq ; ; AVX2-FP-LABEL: store_i8_stride8_vf8: ; AVX2-FP: # %bb.0: ; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %r11 ; AVX2-FP-NEXT: 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: vmovq {{.*#+}} xmm2 = mem[0],zero ; AVX2-FP-NEXT: vmovq {{.*#+}} xmm3 = mem[0],zero ; AVX2-FP-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm3[0],xmm2[0] ; 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: vinserti128 $1, %xmm1, %ymm0, %ymm4 ; AVX2-FP-NEXT: vinserti128 $1, %xmm3, %ymm2, %ymm5 ; AVX2-FP-NEXT: vinserti128 $1, %xmm2, %ymm3, %ymm2 ; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm3 = zero,zero,zero,zero,zero,zero,ymm2[0,8],zero,zero,zero,zero,zero,zero,ymm2[1,9],zero,zero,zero,zero,ymm2[18,26],zero,zero,zero,zero,zero,zero,ymm2[19,27],zero,zero ; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm6 = zero,zero,zero,zero,ymm5[0,8],zero,zero,zero,zero,zero,zero,ymm5[1,9],zero,zero,zero,zero,zero,zero,zero,zero,ymm5[18,26],zero,zero,zero,zero,zero,zero,ymm5[19,27] ; AVX2-FP-NEXT: vpor %ymm3, %ymm6, %ymm3 ; AVX2-FP-NEXT: vinserti128 $1, %xmm0, %ymm1, %ymm0 ; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm1 = zero,zero,ymm0[0,8],zero,zero,zero,zero,zero,zero,ymm0[1,9],zero,zero,zero,zero,ymm0[18,26],zero,zero,zero,zero,zero,zero,ymm0[19,27],zero,zero,zero,zero,zero,zero ; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm6 = ymm4[0,8],zero,zero,zero,zero,zero,zero,ymm4[1,9],zero,zero,zero,zero,zero,zero,zero,zero,ymm4[18,26],zero,zero,zero,zero,zero,zero,ymm4[19,27],zero,zero,zero,zero ; AVX2-FP-NEXT: vpor %ymm1, %ymm6, %ymm1 ; AVX2-FP-NEXT: vpor %ymm3, %ymm1, %ymm1 ; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm2 = zero,zero,zero,zero,zero,zero,ymm2[4,12],zero,zero,zero,zero,zero,zero,ymm2[5,13],zero,zero,zero,zero,ymm2[22,30],zero,zero,zero,zero,zero,zero,ymm2[23,31],zero,zero ; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm3 = zero,zero,zero,zero,ymm5[4,12],zero,zero,zero,zero,zero,zero,ymm5[5,13],zero,zero,zero,zero,zero,zero,zero,zero,ymm5[22,30],zero,zero,zero,zero,zero,zero,ymm5[23,31] ; AVX2-FP-NEXT: vpor %ymm2, %ymm3, %ymm2 ; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm0 = zero,zero,ymm0[4,12],zero,zero,zero,zero,zero,zero,ymm0[5,13],zero,zero,zero,zero,ymm0[22,30],zero,zero,zero,zero,zero,zero,ymm0[23,31],zero,zero,zero,zero,zero,zero ; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm3 = ymm4[4,12],zero,zero,zero,zero,zero,zero,ymm4[5,13],zero,zero,zero,zero,zero,zero,zero,zero,ymm4[22,30],zero,zero,zero,zero,zero,zero,ymm4[23,31],zero,zero,zero,zero ; AVX2-FP-NEXT: vpor %ymm0, %ymm3, %ymm0 ; AVX2-FP-NEXT: vpor %ymm2, %ymm0, %ymm0 ; AVX2-FP-NEXT: vmovdqa %ymm0, 32(%rax) ; AVX2-FP-NEXT: vmovdqa %ymm1, (%rax) ; AVX2-FP-NEXT: vzeroupper ; AVX2-FP-NEXT: retq ; ; AVX2-FCP-LABEL: store_i8_stride8_vf8: ; AVX2-FCP: # %bb.0: ; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r11 ; AVX2-FCP-NEXT: 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: vmovq {{.*#+}} xmm2 = mem[0],zero ; AVX2-FCP-NEXT: vmovq {{.*#+}} xmm3 = mem[0],zero ; AVX2-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm3[0],xmm2[0] ; 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: vinserti128 $1, %xmm1, %ymm0, %ymm0 ; AVX2-FCP-NEXT: vinserti128 $1, %xmm3, %ymm2, %ymm1 ; AVX2-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [0,2,4,6,0,2,4,6] ; AVX2-FCP-NEXT: # ymm2 = mem[0,1,0,1] ; AVX2-FCP-NEXT: vpermd %ymm1, %ymm2, %ymm3 ; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm4 = [u,u,u,u,0,4,8,12,u,u,u,u,1,5,9,13,u,u,u,u,2,6,10,14,u,u,u,u,3,7,11,15] ; AVX2-FCP-NEXT: vpshufb %ymm4, %ymm3, %ymm3 ; AVX2-FCP-NEXT: vpermd %ymm0, %ymm2, %ymm2 ; AVX2-FCP-NEXT: vpmovsxdq {{.*#+}} ymm5 = [201851904,218694913,235537922,252380931] ; AVX2-FCP-NEXT: vpshufb %ymm5, %ymm2, %ymm2 ; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0],ymm3[1],ymm2[2],ymm3[3],ymm2[4],ymm3[5],ymm2[6],ymm3[7] ; AVX2-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [1,3,5,7,1,3,5,7] ; AVX2-FCP-NEXT: # ymm3 = mem[0,1,0,1] ; AVX2-FCP-NEXT: vpermd %ymm1, %ymm3, %ymm1 ; AVX2-FCP-NEXT: vpshufb %ymm4, %ymm1, %ymm1 ; AVX2-FCP-NEXT: vpermd %ymm0, %ymm3, %ymm0 ; AVX2-FCP-NEXT: vpshufb %ymm5, %ymm0, %ymm0 ; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2],ymm1[3],ymm0[4],ymm1[5],ymm0[6],ymm1[7] ; AVX2-FCP-NEXT: vmovdqa %ymm0, 32(%rax) ; AVX2-FCP-NEXT: vmovdqa %ymm2, (%rax) ; AVX2-FCP-NEXT: vzeroupper ; AVX2-FCP-NEXT: retq ; ; AVX512-LABEL: store_i8_stride8_vf8: ; AVX512: # %bb.0: ; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %r11 ; AVX512-NEXT: 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: vmovq {{.*#+}} xmm3 = mem[0],zero ; AVX512-NEXT: vmovq {{.*#+}} xmm4 = mem[0],zero ; AVX512-NEXT: vpunpcklqdq {{.*#+}} xmm3 = xmm4[0],xmm3[0] ; AVX512-NEXT: vinserti128 $1, %xmm2, %ymm3, %ymm4 ; AVX512-NEXT: vmovdqa {{.*#+}} ymm5 = [128,128,4,12,128,128,4,12,128,128,5,13,128,128,5,13,22,30,128,128,22,30,128,128,23,31,128,128,23,31,128,128] ; AVX512-NEXT: vpshufb %ymm5, %ymm4, %ymm6 ; AVX512-NEXT: vinserti128 $1, %xmm0, %ymm1, %ymm7 ; AVX512-NEXT: vpshufb %ymm5, %ymm7, %ymm5 ; AVX512-NEXT: vpblendd {{.*#+}} ymm5 = ymm5[0],ymm6[1],ymm5[2],ymm6[3],ymm5[4],ymm6[5],ymm5[6],ymm6[7] ; AVX512-NEXT: vmovdqa {{.*#+}} ymm6 = [128,128,0,8,128,128,0,8,128,128,1,9,128,128,1,9,18,26,128,128,18,26,128,128,19,27,128,128,19,27,128,128] ; AVX512-NEXT: vpshufb %ymm6, %ymm4, %ymm4 ; AVX512-NEXT: vpshufb %ymm6, %ymm7, %ymm6 ; AVX512-NEXT: vpblendd {{.*#+}} ymm4 = ymm6[0],ymm4[1],ymm6[2],ymm4[3],ymm6[4],ymm4[5],ymm6[6],ymm4[7] ; AVX512-NEXT: vinserti64x4 $1, %ymm5, %zmm4, %zmm4 ; AVX512-NEXT: vinserti128 $1, %xmm3, %ymm2, %ymm2 ; AVX512-NEXT: vmovdqa {{.*#+}} ymm3 = [4,12,128,128,4,12,128,128,5,13,128,128,5,13,128,128,128,128,22,30,128,128,22,30,128,128,23,31,128,128,23,31] ; AVX512-NEXT: vpshufb %ymm3, %ymm2, %ymm5 ; AVX512-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm0 ; AVX512-NEXT: vpshufb %ymm3, %ymm0, %ymm1 ; AVX512-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0],ymm5[1],ymm1[2],ymm5[3],ymm1[4],ymm5[5],ymm1[6],ymm5[7] ; AVX512-NEXT: vmovdqa {{.*#+}} ymm3 = [0,8,128,128,0,8,128,128,1,9,128,128,1,9,128,128,128,128,18,26,128,128,18,26,128,128,19,27,128,128,19,27] ; AVX512-NEXT: vpshufb %ymm3, %ymm2, %ymm2 ; AVX512-NEXT: vpshufb %ymm3, %ymm0, %ymm0 ; AVX512-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0],ymm2[1],ymm0[2],ymm2[3],ymm0[4],ymm2[5],ymm0[6],ymm2[7] ; AVX512-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0 ; AVX512-NEXT: vpord %zmm4, %zmm0, %zmm0 ; AVX512-NEXT: vmovdqa64 %zmm0, (%rax) ; AVX512-NEXT: vzeroupper ; AVX512-NEXT: retq ; ; AVX512-FCP-LABEL: store_i8_stride8_vf8: ; AVX512-FCP: # %bb.0: ; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r11 ; AVX512-FCP-NEXT: 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: vmovq {{.*#+}} xmm3 = mem[0],zero ; AVX512-FCP-NEXT: vmovq {{.*#+}} xmm4 = mem[0],zero ; AVX512-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm3 = xmm4[0],xmm3[0] ; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [1,3,9,11,1,3,9,11] ; AVX512-FCP-NEXT: # ymm4 = mem[0,1,0,1] ; AVX512-FCP-NEXT: vmovdqa %ymm2, %ymm5 ; AVX512-FCP-NEXT: vpermt2d %ymm3, %ymm4, %ymm5 ; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm6 = [u,u,u,u,0,4,8,12,u,u,u,u,1,5,9,13,u,u,u,u,2,6,10,14,u,u,u,u,3,7,11,15] ; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm5, %ymm5 ; AVX512-FCP-NEXT: vpermi2d %ymm1, %ymm0, %ymm4 ; AVX512-FCP-NEXT: vpmovsxdq {{.*#+}} ymm7 = [201851904,218694913,235537922,252380931] ; AVX512-FCP-NEXT: vpshufb %ymm7, %ymm4, %ymm4 ; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0],ymm5[1],ymm4[2],ymm5[3],ymm4[4],ymm5[5],ymm4[6],ymm5[7] ; AVX512-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm5 = [0,2,8,10,0,2,8,10] ; AVX512-FCP-NEXT: # ymm5 = mem[0,1,0,1] ; AVX512-FCP-NEXT: vpermt2d %ymm3, %ymm5, %ymm2 ; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm2, %ymm2 ; AVX512-FCP-NEXT: vpermt2d %ymm1, %ymm5, %ymm0 ; AVX512-FCP-NEXT: vpshufb %ymm7, %ymm0, %ymm0 ; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0],ymm2[1],ymm0[2],ymm2[3],ymm0[4],ymm2[5],ymm0[6],ymm2[7] ; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm0, %zmm0 ; AVX512-FCP-NEXT: vmovdqa64 %zmm0, (%rax) ; AVX512-FCP-NEXT: vzeroupper ; AVX512-FCP-NEXT: retq ; ; AVX512DQ-LABEL: store_i8_stride8_vf8: ; AVX512DQ: # %bb.0: ; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %r11 ; AVX512DQ-NEXT: 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: vmovq {{.*#+}} xmm3 = mem[0],zero ; AVX512DQ-NEXT: vmovq {{.*#+}} xmm4 = mem[0],zero ; AVX512DQ-NEXT: vpunpcklqdq {{.*#+}} xmm3 = xmm4[0],xmm3[0] ; AVX512DQ-NEXT: vinserti128 $1, %xmm2, %ymm3, %ymm4 ; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm5 = [128,128,4,12,128,128,4,12,128,128,5,13,128,128,5,13,22,30,128,128,22,30,128,128,23,31,128,128,23,31,128,128] ; AVX512DQ-NEXT: vpshufb %ymm5, %ymm4, %ymm6 ; AVX512DQ-NEXT: vinserti128 $1, %xmm0, %ymm1, %ymm7 ; AVX512DQ-NEXT: vpshufb %ymm5, %ymm7, %ymm5 ; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm5 = ymm5[0],ymm6[1],ymm5[2],ymm6[3],ymm5[4],ymm6[5],ymm5[6],ymm6[7] ; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm6 = [128,128,0,8,128,128,0,8,128,128,1,9,128,128,1,9,18,26,128,128,18,26,128,128,19,27,128,128,19,27,128,128] ; AVX512DQ-NEXT: vpshufb %ymm6, %ymm4, %ymm4 ; AVX512DQ-NEXT: vpshufb %ymm6, %ymm7, %ymm6 ; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm4 = ymm6[0],ymm4[1],ymm6[2],ymm4[3],ymm6[4],ymm4[5],ymm6[6],ymm4[7] ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm5, %zmm4, %zmm4 ; AVX512DQ-NEXT: vinserti128 $1, %xmm3, %ymm2, %ymm2 ; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm3 = [4,12,128,128,4,12,128,128,5,13,128,128,5,13,128,128,128,128,22,30,128,128,22,30,128,128,23,31,128,128,23,31] ; AVX512DQ-NEXT: vpshufb %ymm3, %ymm2, %ymm5 ; AVX512DQ-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm0 ; AVX512DQ-NEXT: vpshufb %ymm3, %ymm0, %ymm1 ; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0],ymm5[1],ymm1[2],ymm5[3],ymm1[4],ymm5[5],ymm1[6],ymm5[7] ; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm3 = [0,8,128,128,0,8,128,128,1,9,128,128,1,9,128,128,128,128,18,26,128,128,18,26,128,128,19,27,128,128,19,27] ; AVX512DQ-NEXT: vpshufb %ymm3, %ymm2, %ymm2 ; AVX512DQ-NEXT: vpshufb %ymm3, %ymm0, %ymm0 ; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0],ymm2[1],ymm0[2],ymm2[3],ymm0[4],ymm2[5],ymm0[6],ymm2[7] ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0 ; AVX512DQ-NEXT: vpord %zmm4, %zmm0, %zmm0 ; AVX512DQ-NEXT: vmovdqa64 %zmm0, (%rax) ; AVX512DQ-NEXT: vzeroupper ; AVX512DQ-NEXT: retq ; ; AVX512DQ-FCP-LABEL: store_i8_stride8_vf8: ; AVX512DQ-FCP: # %bb.0: ; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r11 ; AVX512DQ-FCP-NEXT: 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: vmovq {{.*#+}} xmm3 = mem[0],zero ; AVX512DQ-FCP-NEXT: vmovq {{.*#+}} xmm4 = mem[0],zero ; AVX512DQ-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm3 = xmm4[0],xmm3[0] ; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [1,3,9,11,1,3,9,11] ; AVX512DQ-FCP-NEXT: # ymm4 = mem[0,1,0,1] ; AVX512DQ-FCP-NEXT: vmovdqa %ymm2, %ymm5 ; AVX512DQ-FCP-NEXT: vpermt2d %ymm3, %ymm4, %ymm5 ; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm6 = [u,u,u,u,0,4,8,12,u,u,u,u,1,5,9,13,u,u,u,u,2,6,10,14,u,u,u,u,3,7,11,15] ; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm5, %ymm5 ; AVX512DQ-FCP-NEXT: vpermi2d %ymm1, %ymm0, %ymm4 ; AVX512DQ-FCP-NEXT: vpmovsxdq {{.*#+}} ymm7 = [201851904,218694913,235537922,252380931] ; AVX512DQ-FCP-NEXT: vpshufb %ymm7, %ymm4, %ymm4 ; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0],ymm5[1],ymm4[2],ymm5[3],ymm4[4],ymm5[5],ymm4[6],ymm5[7] ; AVX512DQ-FCP-NEXT: vbroadcasti128 {{.*#+}} ymm5 = [0,2,8,10,0,2,8,10] ; AVX512DQ-FCP-NEXT: # ymm5 = mem[0,1,0,1] ; AVX512DQ-FCP-NEXT: vpermt2d %ymm3, %ymm5, %ymm2 ; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm2, %ymm2 ; AVX512DQ-FCP-NEXT: vpermt2d %ymm1, %ymm5, %ymm0 ; AVX512DQ-FCP-NEXT: vpshufb %ymm7, %ymm0, %ymm0 ; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0],ymm2[1],ymm0[2],ymm2[3],ymm0[4],ymm2[5],ymm0[6],ymm2[7] ; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm0, %zmm0 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, (%rax) ; AVX512DQ-FCP-NEXT: vzeroupper ; AVX512DQ-FCP-NEXT: retq ; ; AVX512BW-LABEL: store_i8_stride8_vf8: ; AVX512BW: # %bb.0: ; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %r11 ; AVX512BW-NEXT: 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: vmovq {{.*#+}} xmm3 = mem[0],zero ; AVX512BW-NEXT: vmovq {{.*#+}} xmm4 = mem[0],zero ; AVX512BW-NEXT: vpunpcklqdq {{.*#+}} xmm3 = xmm4[0],xmm3[0] ; AVX512BW-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm4 ; AVX512BW-NEXT: vinserti64x4 $1, %ymm4, %zmm4, %zmm4 ; AVX512BW-NEXT: vpshufb {{.*#+}} zmm4 = zmm4[0,8],zero,zero,zero,zero,zero,zero,zmm4[1,9],zero,zero,zero,zero,zero,zero,zero,zero,zmm4[18,26],zero,zero,zero,zero,zero,zero,zmm4[19,27],zero,zero,zero,zero,zmm4[36,44],zero,zero,zero,zero,zero,zero,zmm4[37,45],zero,zero,zero,zero,zero,zero,zero,zero,zmm4[54,62],zero,zero,zero,zero,zero,zero,zmm4[55,63],zero,zero,zero,zero ; AVX512BW-NEXT: vinserti128 $1, %xmm0, %ymm1, %ymm0 ; AVX512BW-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm0 ; AVX512BW-NEXT: vpshufb {{.*#+}} zmm0 = zero,zero,zmm0[0,8],zero,zero,zero,zero,zero,zero,zmm0[1,9],zero,zero,zero,zero,zmm0[18,26],zero,zero,zero,zero,zero,zero,zmm0[19,27],zero,zero,zero,zero,zero,zero,zero,zero,zmm0[36,44],zero,zero,zero,zero,zero,zero,zmm0[37,45],zero,zero,zero,zero,zmm0[54,62],zero,zero,zero,zero,zero,zero,zmm0[55,63],zero,zero,zero,zero,zero,zero ; AVX512BW-NEXT: vporq %zmm4, %zmm0, %zmm0 ; AVX512BW-NEXT: vinserti128 $1, %xmm3, %ymm2, %ymm1 ; AVX512BW-NEXT: vinserti64x4 $1, %ymm1, %zmm1, %zmm1 ; AVX512BW-NEXT: vpshufb {{.*#+}} zmm1 = zero,zero,zero,zero,zmm1[0,8],zero,zero,zero,zero,zero,zero,zmm1[1,9],zero,zero,zero,zero,zero,zero,zero,zero,zmm1[18,26],zero,zero,zero,zero,zero,zero,zmm1[19,27],zero,zero,zero,zero,zmm1[36,44],zero,zero,zero,zero,zero,zero,zmm1[37,45],zero,zero,zero,zero,zero,zero,zero,zero,zmm1[54,62],zero,zero,zero,zero,zero,zero,zmm1[55,63] ; AVX512BW-NEXT: vinserti128 $1, %xmm2, %ymm3, %ymm2 ; AVX512BW-NEXT: vinserti64x4 $1, %ymm2, %zmm2, %zmm2 ; AVX512BW-NEXT: vpshufb {{.*#+}} zmm2 = zero,zero,zero,zero,zero,zero,zmm2[0,8],zero,zero,zero,zero,zero,zero,zmm2[1,9],zero,zero,zero,zero,zmm2[18,26],zero,zero,zero,zero,zero,zero,zmm2[19,27],zero,zero,zero,zero,zero,zero,zero,zero,zmm2[36,44],zero,zero,zero,zero,zero,zero,zmm2[37,45],zero,zero,zero,zero,zmm2[54,62],zero,zero,zero,zero,zero,zero,zmm2[55,63],zero,zero ; AVX512BW-NEXT: vpternlogq {{.*#+}} zmm2 = zmm2 | zmm0 | zmm1 ; AVX512BW-NEXT: vmovdqa64 %zmm2, (%rax) ; AVX512BW-NEXT: vzeroupper ; AVX512BW-NEXT: retq ; ; AVX512BW-FCP-LABEL: store_i8_stride8_vf8: ; AVX512BW-FCP: # %bb.0: ; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r11 ; AVX512BW-FCP-NEXT: 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: vmovq {{.*#+}} xmm3 = mem[0],zero ; AVX512BW-FCP-NEXT: vmovq {{.*#+}} xmm4 = mem[0],zero ; AVX512BW-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm3 = xmm4[0],xmm3[0] ; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} zmm4 = [0,2,16,18,0,2,16,18,1,3,17,19,1,3,17,19] ; AVX512BW-FCP-NEXT: vpermt2d %zmm1, %zmm4, %zmm0 ; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} zmm0 = zmm0[0,4,8,12],zero,zero,zero,zero,zmm0[1,5,9,13],zero,zero,zero,zero,zmm0[18,22,26,30],zero,zero,zero,zero,zmm0[19,23,27,31],zero,zero,zero,zero,zmm0[32,36,40,44],zero,zero,zero,zero,zmm0[33,37,41,45],zero,zero,zero,zero,zmm0[50,54,58,62],zero,zero,zero,zero,zmm0[51,55,59,63],zero,zero,zero,zero ; AVX512BW-FCP-NEXT: vpermt2d %zmm3, %zmm4, %zmm2 ; AVX512BW-FCP-NEXT: vpshufb {{.*#+}} zmm1 = zero,zero,zero,zero,zmm2[0,4,8,12],zero,zero,zero,zero,zmm2[1,5,9,13],zero,zero,zero,zero,zmm2[18,22,26,30],zero,zero,zero,zero,zmm2[19,23,27,31],zero,zero,zero,zero,zmm2[32,36,40,44],zero,zero,zero,zero,zmm2[33,37,41,45],zero,zero,zero,zero,zmm2[50,54,58,62],zero,zero,zero,zero,zmm2[51,55,59,63] ; AVX512BW-FCP-NEXT: vporq %zmm0, %zmm1, %zmm0 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, (%rax) ; AVX512BW-FCP-NEXT: vzeroupper ; AVX512BW-FCP-NEXT: retq ; ; AVX512DQ-BW-LABEL: store_i8_stride8_vf8: ; AVX512DQ-BW: # %bb.0: ; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %r11 ; AVX512DQ-BW-NEXT: 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: vmovq {{.*#+}} xmm3 = mem[0],zero ; AVX512DQ-BW-NEXT: vmovq {{.*#+}} xmm4 = mem[0],zero ; AVX512DQ-BW-NEXT: vpunpcklqdq {{.*#+}} xmm3 = xmm4[0],xmm3[0] ; AVX512DQ-BW-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm4 ; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm4, %zmm4, %zmm4 ; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} zmm4 = zmm4[0,8],zero,zero,zero,zero,zero,zero,zmm4[1,9],zero,zero,zero,zero,zero,zero,zero,zero,zmm4[18,26],zero,zero,zero,zero,zero,zero,zmm4[19,27],zero,zero,zero,zero,zmm4[36,44],zero,zero,zero,zero,zero,zero,zmm4[37,45],zero,zero,zero,zero,zero,zero,zero,zero,zmm4[54,62],zero,zero,zero,zero,zero,zero,zmm4[55,63],zero,zero,zero,zero ; AVX512DQ-BW-NEXT: vinserti128 $1, %xmm0, %ymm1, %ymm0 ; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm0 ; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} zmm0 = zero,zero,zmm0[0,8],zero,zero,zero,zero,zero,zero,zmm0[1,9],zero,zero,zero,zero,zmm0[18,26],zero,zero,zero,zero,zero,zero,zmm0[19,27],zero,zero,zero,zero,zero,zero,zero,zero,zmm0[36,44],zero,zero,zero,zero,zero,zero,zmm0[37,45],zero,zero,zero,zero,zmm0[54,62],zero,zero,zero,zero,zero,zero,zmm0[55,63],zero,zero,zero,zero,zero,zero ; AVX512DQ-BW-NEXT: vporq %zmm4, %zmm0, %zmm0 ; AVX512DQ-BW-NEXT: vinserti128 $1, %xmm3, %ymm2, %ymm1 ; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm1, %zmm1, %zmm1 ; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} zmm1 = zero,zero,zero,zero,zmm1[0,8],zero,zero,zero,zero,zero,zero,zmm1[1,9],zero,zero,zero,zero,zero,zero,zero,zero,zmm1[18,26],zero,zero,zero,zero,zero,zero,zmm1[19,27],zero,zero,zero,zero,zmm1[36,44],zero,zero,zero,zero,zero,zero,zmm1[37,45],zero,zero,zero,zero,zero,zero,zero,zero,zmm1[54,62],zero,zero,zero,zero,zero,zero,zmm1[55,63] ; AVX512DQ-BW-NEXT: vinserti128 $1, %xmm2, %ymm3, %ymm2 ; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm2, %zmm2, %zmm2 ; AVX512DQ-BW-NEXT: vpshufb {{.*#+}} zmm2 = zero,zero,zero,zero,zero,zero,zmm2[0,8],zero,zero,zero,zero,zero,zero,zmm2[1,9],zero,zero,zero,zero,zmm2[18,26],zero,zero,zero,zero,zero,zero,zmm2[19,27],zero,zero,zero,zero,zero,zero,zero,zero,zmm2[36,44],zero,zero,zero,zero,zero,zero,zmm2[37,45],zero,zero,zero,zero,zmm2[54,62],zero,zero,zero,zero,zero,zero,zmm2[55,63],zero,zero ; AVX512DQ-BW-NEXT: vpternlogq {{.*#+}} zmm2 = zmm2 | zmm0 | zmm1 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, (%rax) ; AVX512DQ-BW-NEXT: vzeroupper ; AVX512DQ-BW-NEXT: retq ; ; AVX512DQ-BW-FCP-LABEL: store_i8_stride8_vf8: ; AVX512DQ-BW-FCP: # %bb.0: ; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r11 ; AVX512DQ-BW-FCP-NEXT: 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: vmovq {{.*#+}} xmm3 = mem[0],zero ; AVX512DQ-BW-FCP-NEXT: vmovq {{.*#+}} xmm4 = mem[0],zero ; AVX512DQ-BW-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm3 = xmm4[0],xmm3[0] ; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} zmm4 = [0,2,16,18,0,2,16,18,1,3,17,19,1,3,17,19] ; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm1, %zmm4, %zmm0 ; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} zmm0 = zmm0[0,4,8,12],zero,zero,zero,zero,zmm0[1,5,9,13],zero,zero,zero,zero,zmm0[18,22,26,30],zero,zero,zero,zero,zmm0[19,23,27,31],zero,zero,zero,zero,zmm0[32,36,40,44],zero,zero,zero,zero,zmm0[33,37,41,45],zero,zero,zero,zero,zmm0[50,54,58,62],zero,zero,zero,zero,zmm0[51,55,59,63],zero,zero,zero,zero ; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm3, %zmm4, %zmm2 ; AVX512DQ-BW-FCP-NEXT: vpshufb {{.*#+}} zmm1 = zero,zero,zero,zero,zmm2[0,4,8,12],zero,zero,zero,zero,zmm2[1,5,9,13],zero,zero,zero,zero,zmm2[18,22,26,30],zero,zero,zero,zero,zmm2[19,23,27,31],zero,zero,zero,zero,zmm2[32,36,40,44],zero,zero,zero,zero,zmm2[33,37,41,45],zero,zero,zero,zero,zmm2[50,54,58,62],zero,zero,zero,zero,zmm2[51,55,59,63] ; AVX512DQ-BW-FCP-NEXT: vporq %zmm0, %zmm1, %zmm0 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, (%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 %in.vec6 = load <8 x i8>, ptr %in.vecptr6, align 64 %in.vec7 = load <8 x i8>, ptr %in.vecptr7, 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 <8 x i8> %in.vec6, <8 x i8> %in.vec7, <16 x i32> %5 = shufflevector <16 x i8> %1, <16 x i8> %2, <32 x i32> %6 = shufflevector <16 x i8> %3, <16 x i8> %4, <32 x i32> %7 = shufflevector <32 x i8> %5, <32 x i8> %6, <64 x i32> %interleaved.vec = shufflevector <64 x i8> %7, <64 x i8> poison, <64 x i32> store <64 x i8> %interleaved.vec, ptr %out.vec, align 64 ret void } define void @store_i8_stride8_vf16(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecptr2, ptr %in.vecptr3, ptr %in.vecptr4, ptr %in.vecptr5, ptr %in.vecptr6, ptr %in.vecptr7, ptr %out.vec) nounwind { ; SSE-LABEL: store_i8_stride8_vf16: ; SSE: # %bb.0: ; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax ; SSE-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; SSE-NEXT: movdqa (%rdi), %xmm10 ; SSE-NEXT: movdqa (%rsi), %xmm9 ; SSE-NEXT: movdqa %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: movdqa (%rdx), %xmm1 ; SSE-NEXT: movdqa (%rcx), %xmm6 ; SSE-NEXT: movdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: movdqa (%r8), %xmm3 ; SSE-NEXT: movdqa (%r9), %xmm11 ; SSE-NEXT: movdqa (%r10), %xmm4 ; SSE-NEXT: movdqa (%rax), %xmm13 ; SSE-NEXT: movdqa %xmm4, %xmm12 ; SSE-NEXT: punpcklbw {{.*#+}} xmm12 = xmm12[0],xmm13[0],xmm12[1],xmm13[1],xmm12[2],xmm13[2],xmm12[3],xmm13[3],xmm12[4],xmm13[4],xmm12[5],xmm13[5],xmm12[6],xmm13[6],xmm12[7],xmm13[7] ; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm12[0,0,2,1,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,0,2,1] ; SSE-NEXT: movdqa {{.*#+}} xmm2 = [65535,65535,65535,0,65535,65535,65535,0] ; SSE-NEXT: movdqa %xmm2, %xmm5 ; SSE-NEXT: pandn %xmm0, %xmm5 ; SSE-NEXT: movdqa %xmm3, %xmm14 ; SSE-NEXT: punpcklbw {{.*#+}} xmm14 = 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] ; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm14[0,1,1,3,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm0[0,0,2,1] ; SSE-NEXT: pand %xmm2, %xmm7 ; SSE-NEXT: por %xmm5, %xmm7 ; SSE-NEXT: movdqa %xmm1, %xmm15 ; SSE-NEXT: punpcklbw {{.*#+}} xmm15 = xmm15[0],xmm6[0],xmm15[1],xmm6[1],xmm15[2],xmm6[2],xmm15[3],xmm6[3],xmm15[4],xmm6[4],xmm15[5],xmm6[5],xmm15[6],xmm6[6],xmm15[7],xmm6[7] ; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm15[0,0,2,1,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm0[0,1,1,3] ; SSE-NEXT: movdqa {{.*#+}} xmm0 = [65535,0,65535,65535,65535,0,65535,65535] ; SSE-NEXT: movdqa %xmm0, %xmm8 ; SSE-NEXT: pandn %xmm5, %xmm8 ; SSE-NEXT: movdqa %xmm10, %xmm6 ; SSE-NEXT: punpcklbw {{.*#+}} xmm6 = xmm6[0],xmm9[0],xmm6[1],xmm9[1],xmm6[2],xmm9[2],xmm6[3],xmm9[3],xmm6[4],xmm9[4],xmm6[5],xmm9[5],xmm6[6],xmm9[6],xmm6[7],xmm9[7] ; SSE-NEXT: pshufd {{.*#+}} xmm9 = xmm6[0,0,0,0] ; SSE-NEXT: pshufhw {{.*#+}} xmm9 = xmm9[0,1,2,3,5,5,5,5] ; SSE-NEXT: pand %xmm0, %xmm9 ; SSE-NEXT: por %xmm8, %xmm9 ; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm7[1,3,2,3] ; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm9[0,2,2,3] ; SSE-NEXT: punpckldq {{.*#+}} xmm5 = xmm5[0],xmm7[0],xmm5[1],xmm7[1] ; SSE-NEXT: movdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: pshuflw {{.*#+}} xmm7 = xmm12[0,2,2,3,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm7[0,0,2,1] ; SSE-NEXT: movdqa %xmm2, %xmm8 ; SSE-NEXT: pandn %xmm7, %xmm8 ; SSE-NEXT: pshuflw {{.*#+}} xmm7 = xmm14[2,1,3,3,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm7[0,0,2,1] ; SSE-NEXT: pand %xmm2, %xmm7 ; SSE-NEXT: por %xmm8, %xmm7 ; SSE-NEXT: pshuflw {{.*#+}} xmm8 = xmm15[0,2,2,3,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm8 = xmm8[0,1,1,3] ; SSE-NEXT: movdqa %xmm0, %xmm9 ; SSE-NEXT: pandn %xmm8, %xmm9 ; SSE-NEXT: pshufd {{.*#+}} xmm8 = xmm6[1,1,1,1] ; SSE-NEXT: pshufhw {{.*#+}} xmm8 = xmm8[0,1,2,3,5,5,5,5] ; SSE-NEXT: pand %xmm0, %xmm8 ; SSE-NEXT: por %xmm9, %xmm8 ; SSE-NEXT: pshufd {{.*#+}} xmm9 = xmm7[1,3,2,3] ; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm8[0,2,2,3] ; SSE-NEXT: punpckldq {{.*#+}} xmm7 = xmm7[0],xmm9[0],xmm7[1],xmm9[1] ; SSE-NEXT: pshufhw {{.*#+}} xmm8 = xmm12[0,1,2,3,4,6,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm8 = xmm8[0,2,2,3] ; SSE-NEXT: movdqa %xmm2, %xmm9 ; SSE-NEXT: pandn %xmm8, %xmm9 ; SSE-NEXT: pshufhw {{.*#+}} xmm8 = xmm14[0,1,2,3,6,5,7,7] ; SSE-NEXT: pshufd {{.*#+}} xmm8 = xmm8[0,2,2,3] ; SSE-NEXT: pand %xmm2, %xmm8 ; SSE-NEXT: por %xmm9, %xmm8 ; SSE-NEXT: pshufhw {{.*#+}} xmm9 = xmm15[0,1,2,3,4,6,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm9 = xmm9[2,1,3,3] ; SSE-NEXT: movdqa %xmm0, %xmm5 ; SSE-NEXT: pandn %xmm9, %xmm5 ; SSE-NEXT: pshufd {{.*#+}} xmm9 = xmm6[3,3,3,3] ; SSE-NEXT: pshufhw {{.*#+}} xmm9 = xmm9[0,1,2,3,5,5,5,5] ; SSE-NEXT: pand %xmm0, %xmm9 ; SSE-NEXT: por %xmm5, %xmm9 ; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm8[1,3,2,3] ; SSE-NEXT: pshufd {{.*#+}} xmm8 = xmm9[0,2,2,3] ; SSE-NEXT: punpckldq {{.*#+}} xmm8 = xmm8[0],xmm5[0],xmm8[1],xmm5[1] ; SSE-NEXT: pshufhw {{.*#+}} xmm5 = xmm12[0,1,2,3,4,4,6,5] ; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm5[0,2,2,3] ; SSE-NEXT: movdqa %xmm2, %xmm9 ; SSE-NEXT: pandn %xmm5, %xmm9 ; SSE-NEXT: pshufhw {{.*#+}} xmm5 = xmm14[0,1,2,3,4,5,5,7] ; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm5[0,2,2,3] ; SSE-NEXT: pand %xmm2, %xmm5 ; SSE-NEXT: por %xmm9, %xmm5 ; SSE-NEXT: pshufhw {{.*#+}} xmm9 = xmm15[0,1,2,3,4,4,6,5] ; SSE-NEXT: pshufd {{.*#+}} xmm9 = xmm9[2,1,3,3] ; SSE-NEXT: movdqa %xmm0, %xmm12 ; SSE-NEXT: pandn %xmm9, %xmm12 ; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm6[2,2,2,2] ; SSE-NEXT: pshufhw {{.*#+}} xmm6 = xmm6[0,1,2,3,5,5,5,5] ; SSE-NEXT: pand %xmm0, %xmm6 ; SSE-NEXT: por %xmm12, %xmm6 ; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm5[1,3,2,3] ; SSE-NEXT: pshufd {{.*#+}} xmm12 = xmm6[0,2,2,3] ; SSE-NEXT: punpckldq {{.*#+}} xmm12 = xmm12[0],xmm5[0],xmm12[1],xmm5[1] ; SSE-NEXT: punpckhbw {{.*#+}} xmm4 = xmm4[8],xmm13[8],xmm4[9],xmm13[9],xmm4[10],xmm13[10],xmm4[11],xmm13[11],xmm4[12],xmm13[12],xmm4[13],xmm13[13],xmm4[14],xmm13[14],xmm4[15],xmm13[15] ; SSE-NEXT: pshuflw {{.*#+}} xmm5 = xmm4[0,0,2,1,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm5[0,0,2,1] ; SSE-NEXT: movdqa %xmm2, %xmm6 ; SSE-NEXT: pandn %xmm5, %xmm6 ; SSE-NEXT: punpckhbw {{.*#+}} xmm3 = xmm3[8],xmm11[8],xmm3[9],xmm11[9],xmm3[10],xmm11[10],xmm3[11],xmm11[11],xmm3[12],xmm11[12],xmm3[13],xmm11[13],xmm3[14],xmm11[14],xmm3[15],xmm11[15] ; SSE-NEXT: pshuflw {{.*#+}} xmm5 = xmm3[0,1,1,3,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm5[0,0,2,1] ; SSE-NEXT: pand %xmm2, %xmm5 ; SSE-NEXT: por %xmm6, %xmm5 ; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm5[1,3,2,3] ; SSE-NEXT: punpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload ; SSE-NEXT: # xmm1 = xmm1[8],mem[8],xmm1[9],mem[9],xmm1[10],mem[10],xmm1[11],mem[11],xmm1[12],mem[12],xmm1[13],mem[13],xmm1[14],mem[14],xmm1[15],mem[15] ; SSE-NEXT: pshuflw {{.*#+}} xmm6 = xmm1[0,0,2,1,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm6[0,1,1,3] ; SSE-NEXT: movdqa %xmm0, %xmm9 ; SSE-NEXT: pandn %xmm6, %xmm9 ; SSE-NEXT: punpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Folded Reload ; SSE-NEXT: # xmm10 = xmm10[8],mem[8],xmm10[9],mem[9],xmm10[10],mem[10],xmm10[11],mem[11],xmm10[12],mem[12],xmm10[13],mem[13],xmm10[14],mem[14],xmm10[15],mem[15] ; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm10[0,0,0,0] ; SSE-NEXT: pshufhw {{.*#+}} xmm6 = xmm6[0,1,2,3,5,5,5,5] ; SSE-NEXT: pand %xmm0, %xmm6 ; SSE-NEXT: por %xmm9, %xmm6 ; SSE-NEXT: pshufd {{.*#+}} xmm9 = xmm6[0,2,2,3] ; SSE-NEXT: punpckldq {{.*#+}} xmm9 = xmm9[0],xmm5[0],xmm9[1],xmm5[1] ; SSE-NEXT: pshuflw {{.*#+}} xmm5 = xmm4[0,2,2,3,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm5[0,0,2,1] ; SSE-NEXT: movdqa %xmm2, %xmm6 ; SSE-NEXT: pandn %xmm5, %xmm6 ; SSE-NEXT: pshuflw {{.*#+}} xmm5 = xmm3[2,1,3,3,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm5[0,0,2,1] ; SSE-NEXT: pand %xmm2, %xmm5 ; SSE-NEXT: por %xmm6, %xmm5 ; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm5[1,3,2,3] ; SSE-NEXT: pshuflw {{.*#+}} xmm5 = xmm1[0,2,2,3,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm5[0,1,1,3] ; SSE-NEXT: movdqa %xmm0, %xmm11 ; SSE-NEXT: pandn %xmm5, %xmm11 ; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm10[1,1,1,1] ; SSE-NEXT: pshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,5,5,5,5] ; SSE-NEXT: pand %xmm0, %xmm5 ; SSE-NEXT: por %xmm11, %xmm5 ; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm5[0,2,2,3] ; SSE-NEXT: punpckldq {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1] ; SSE-NEXT: pshufhw {{.*#+}} xmm6 = xmm4[0,1,2,3,4,6,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm6[0,2,2,3] ; SSE-NEXT: movdqa %xmm2, %xmm11 ; SSE-NEXT: pandn %xmm6, %xmm11 ; SSE-NEXT: pshufhw {{.*#+}} xmm6 = xmm3[0,1,2,3,6,5,7,7] ; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm6[0,2,2,3] ; SSE-NEXT: pand %xmm2, %xmm6 ; SSE-NEXT: por %xmm11, %xmm6 ; SSE-NEXT: pshufhw {{.*#+}} xmm11 = xmm1[0,1,2,3,4,6,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm11 = xmm11[2,1,3,3] ; SSE-NEXT: movdqa %xmm0, %xmm13 ; SSE-NEXT: pandn %xmm11, %xmm13 ; SSE-NEXT: pshufd {{.*#+}} xmm11 = xmm10[3,3,3,3] ; SSE-NEXT: pshufhw {{.*#+}} xmm11 = xmm11[0,1,2,3,5,5,5,5] ; SSE-NEXT: pand %xmm0, %xmm11 ; SSE-NEXT: por %xmm13, %xmm11 ; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm6[1,3,2,3] ; SSE-NEXT: pshufd {{.*#+}} xmm11 = xmm11[0,2,2,3] ; SSE-NEXT: punpckldq {{.*#+}} xmm11 = xmm11[0],xmm6[0],xmm11[1],xmm6[1] ; SSE-NEXT: pshufhw {{.*#+}} xmm4 = xmm4[0,1,2,3,4,4,6,5] ; SSE-NEXT: pshufd {{.*#+}} xmm4 = xmm4[0,2,2,3] ; SSE-NEXT: pshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,5,5,7] ; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm3[0,2,2,3] ; SSE-NEXT: pand %xmm2, %xmm3 ; SSE-NEXT: pandn %xmm4, %xmm2 ; SSE-NEXT: por %xmm3, %xmm2 ; SSE-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,4,6,5] ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[2,1,3,3] ; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm10[2,2,2,2] ; SSE-NEXT: pshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,5,5,5,5] ; SSE-NEXT: pand %xmm0, %xmm3 ; SSE-NEXT: pandn %xmm1, %xmm0 ; SSE-NEXT: por %xmm3, %xmm0 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm2[1,3,2,3] ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3] ; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] ; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax ; SSE-NEXT: movdqa %xmm0, 96(%rax) ; SSE-NEXT: movdqa %xmm11, 112(%rax) ; SSE-NEXT: movdqa %xmm5, 80(%rax) ; SSE-NEXT: movdqa %xmm9, 64(%rax) ; SSE-NEXT: movdqa %xmm12, 32(%rax) ; SSE-NEXT: movdqa %xmm8, 48(%rax) ; SSE-NEXT: movdqa %xmm7, 16(%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_stride8_vf16: ; AVX: # %bb.0: ; AVX-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX-NEXT: movq {{[0-9]+}}(%rsp), %r11 ; AVX-NEXT: vmovdqa (%rdi), %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: vmovdqa (%r11), %xmm8 ; AVX-NEXT: vmovdqa (%r10), %xmm9 ; AVX-NEXT: vpunpcklbw {{.*#+}} xmm7 = xmm8[0],xmm9[0],xmm8[1],xmm9[1],xmm8[2],xmm9[2],xmm8[3],xmm9[3],xmm8[4],xmm9[4],xmm8[5],xmm9[5],xmm8[6],xmm9[6],xmm8[7],xmm9[7] ; AVX-NEXT: vpshuflw {{.*#+}} xmm0 = xmm7[0,2,2,3,4,5,6,7] ; AVX-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,0,2,1] ; AVX-NEXT: vpunpcklbw {{.*#+}} xmm10 = 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 {{.*#+}} xmm11 = xmm10[2,1,3,3,4,5,6,7] ; AVX-NEXT: vpshufd {{.*#+}} xmm11 = xmm11[0,0,2,1] ; AVX-NEXT: vpblendw {{.*#+}} xmm0 = xmm11[0,1,2],xmm0[3],xmm11[4,5,6],xmm0[7] ; AVX-NEXT: vpshuflw {{.*#+}} xmm11 = xmm7[0,0,2,1,4,5,6,7] ; AVX-NEXT: vpshufd {{.*#+}} xmm11 = xmm11[0,0,2,1] ; AVX-NEXT: vpshuflw {{.*#+}} xmm12 = xmm10[0,1,1,3,4,5,6,7] ; AVX-NEXT: vpshufd {{.*#+}} xmm12 = xmm12[0,0,2,1] ; AVX-NEXT: vpblendw {{.*#+}} xmm11 = xmm12[0,1,2],xmm11[3],xmm12[4,5,6],xmm11[7] ; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm11, %ymm0 ; AVX-NEXT: vpunpcklbw {{.*#+}} xmm11 = 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 {{.*#+}} xmm12 = xmm11[0,2,2,3,4,5,6,7] ; AVX-NEXT: vpmovzxdq {{.*#+}} xmm12 = xmm12[0],zero,xmm12[1],zero ; AVX-NEXT: vpunpcklbw {{.*#+}} xmm13 = 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 {{.*#+}} xmm14 = xmm13[1,1,1,1] ; AVX-NEXT: vpmovzxwq {{.*#+}} xmm14 = xmm14[0],zero,zero,zero,xmm14[1],zero,zero,zero ; AVX-NEXT: vpblendw {{.*#+}} xmm12 = xmm14[0],xmm12[1],xmm14[2,3,4],xmm12[5],xmm14[6,7] ; AVX-NEXT: vpmovzxwq {{.*#+}} xmm14 = xmm13[0],zero,zero,zero,xmm13[1],zero,zero,zero ; AVX-NEXT: vpshuflw {{.*#+}} xmm15 = xmm11[0,0,2,1,4,5,6,7] ; AVX-NEXT: vpmovzxdq {{.*#+}} xmm15 = xmm15[0],zero,xmm15[1],zero ; AVX-NEXT: vpblendw {{.*#+}} xmm14 = xmm14[0],xmm15[1],xmm14[2,3,4],xmm15[5],xmm14[6,7] ; AVX-NEXT: vinsertf128 $1, %xmm12, %ymm14, %ymm12 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm12[0],ymm0[1],ymm12[2],ymm0[3],ymm12[4],ymm0[5],ymm12[6],ymm0[7] ; AVX-NEXT: vpshufhw {{.*#+}} xmm12 = xmm7[0,1,2,3,4,6,6,7] ; AVX-NEXT: vpshufd {{.*#+}} xmm12 = xmm12[0,2,2,3] ; AVX-NEXT: vpshufhw {{.*#+}} xmm14 = xmm10[0,1,2,3,6,5,7,7] ; AVX-NEXT: vpshufd {{.*#+}} xmm14 = xmm14[0,2,2,3] ; AVX-NEXT: vpblendw {{.*#+}} xmm12 = xmm14[0,1,2],xmm12[3],xmm14[4,5,6],xmm12[7] ; AVX-NEXT: vpshufhw {{.*#+}} xmm7 = xmm7[0,1,2,3,4,4,6,5] ; AVX-NEXT: vpshufd {{.*#+}} xmm7 = xmm7[0,2,2,3] ; AVX-NEXT: vpshufhw {{.*#+}} xmm10 = xmm10[0,1,2,3,4,5,5,7] ; AVX-NEXT: vpshufd {{.*#+}} xmm10 = xmm10[0,2,2,3] ; AVX-NEXT: vpblendw {{.*#+}} xmm7 = xmm10[0,1,2],xmm7[3],xmm10[4,5,6],xmm7[7] ; AVX-NEXT: vinsertf128 $1, %xmm12, %ymm7, %ymm7 ; AVX-NEXT: vpshufhw {{.*#+}} xmm10 = xmm11[0,1,2,3,4,6,6,7] ; AVX-NEXT: vpshufd {{.*#+}} xmm10 = xmm10[2,1,3,3] ; AVX-NEXT: vpshufd {{.*#+}} xmm12 = xmm13[3,3,3,3] ; AVX-NEXT: vpmovzxwq {{.*#+}} xmm12 = xmm12[0],zero,zero,zero,xmm12[1],zero,zero,zero ; AVX-NEXT: vpblendw {{.*#+}} xmm10 = xmm12[0],xmm10[1],xmm12[2,3,4],xmm10[5],xmm12[6,7] ; AVX-NEXT: vpshufhw {{.*#+}} xmm11 = xmm11[0,1,2,3,4,4,6,5] ; AVX-NEXT: vpshufd {{.*#+}} xmm11 = xmm11[2,1,3,3] ; AVX-NEXT: vpshufd {{.*#+}} xmm12 = xmm13[2,3,2,3] ; AVX-NEXT: vpmovzxwq {{.*#+}} xmm12 = xmm12[0],zero,zero,zero,xmm12[1],zero,zero,zero ; AVX-NEXT: vpblendw {{.*#+}} xmm11 = xmm12[0],xmm11[1],xmm12[2,3,4],xmm11[5],xmm12[6,7] ; AVX-NEXT: vinsertf128 $1, %xmm10, %ymm11, %ymm10 ; AVX-NEXT: vblendps {{.*#+}} ymm7 = ymm10[0],ymm7[1],ymm10[2],ymm7[3],ymm10[4],ymm7[5],ymm10[6],ymm7[7] ; AVX-NEXT: vpunpckhbw {{.*#+}} xmm8 = xmm8[8],xmm9[8],xmm8[9],xmm9[9],xmm8[10],xmm9[10],xmm8[11],xmm9[11],xmm8[12],xmm9[12],xmm8[13],xmm9[13],xmm8[14],xmm9[14],xmm8[15],xmm9[15] ; AVX-NEXT: vpshuflw {{.*#+}} xmm9 = xmm8[0,2,2,3,4,5,6,7] ; AVX-NEXT: vpshufd {{.*#+}} xmm9 = xmm9[0,0,2,1] ; AVX-NEXT: vpunpckhbw {{.*#+}} 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] ; AVX-NEXT: vpshuflw {{.*#+}} xmm6 = xmm5[2,1,3,3,4,5,6,7] ; AVX-NEXT: vpshufd {{.*#+}} xmm6 = xmm6[0,0,2,1] ; AVX-NEXT: vpblendw {{.*#+}} xmm6 = xmm6[0,1,2],xmm9[3],xmm6[4,5,6],xmm9[7] ; AVX-NEXT: vpshuflw {{.*#+}} xmm9 = xmm8[0,0,2,1,4,5,6,7] ; AVX-NEXT: vpshufd {{.*#+}} xmm9 = xmm9[0,0,2,1] ; AVX-NEXT: vpshuflw {{.*#+}} xmm10 = xmm5[0,1,1,3,4,5,6,7] ; AVX-NEXT: vpshufd {{.*#+}} xmm10 = xmm10[0,0,2,1] ; AVX-NEXT: vpblendw {{.*#+}} xmm9 = xmm10[0,1,2],xmm9[3],xmm10[4,5,6],xmm9[7] ; AVX-NEXT: vinsertf128 $1, %xmm6, %ymm9, %ymm6 ; 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[0,2,2,3,4,5,6,7] ; AVX-NEXT: vpmovzxdq {{.*#+}} xmm4 = xmm4[0],zero,xmm4[1],zero ; 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[1,1,1,1] ; AVX-NEXT: vpmovzxwq {{.*#+}} xmm2 = xmm2[0],zero,zero,zero,xmm2[1],zero,zero,zero ; AVX-NEXT: vpblendw {{.*#+}} xmm2 = xmm2[0],xmm4[1],xmm2[2,3,4],xmm4[5],xmm2[6,7] ; AVX-NEXT: vpmovzxwq {{.*#+}} xmm4 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero ; AVX-NEXT: vpshuflw {{.*#+}} xmm9 = xmm3[0,0,2,1,4,5,6,7] ; AVX-NEXT: vpmovzxdq {{.*#+}} xmm9 = xmm9[0],zero,xmm9[1],zero ; AVX-NEXT: vpblendw {{.*#+}} xmm4 = xmm4[0],xmm9[1],xmm4[2,3,4],xmm9[5],xmm4[6,7] ; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm4, %ymm2 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0],ymm6[1],ymm2[2],ymm6[3],ymm2[4],ymm6[5],ymm2[6],ymm6[7] ; AVX-NEXT: vpshufhw {{.*#+}} xmm4 = xmm8[0,1,2,3,4,6,6,7] ; AVX-NEXT: vpshufd {{.*#+}} xmm4 = xmm4[0,2,2,3] ; AVX-NEXT: vpshufhw {{.*#+}} xmm6 = xmm5[0,1,2,3,6,5,7,7] ; AVX-NEXT: vpshufd {{.*#+}} xmm6 = xmm6[0,2,2,3] ; AVX-NEXT: vpblendw {{.*#+}} xmm4 = xmm6[0,1,2],xmm4[3],xmm6[4,5,6],xmm4[7] ; AVX-NEXT: vpshufhw {{.*#+}} xmm6 = xmm8[0,1,2,3,4,4,6,5] ; AVX-NEXT: vpshufd {{.*#+}} xmm6 = xmm6[0,2,2,3] ; AVX-NEXT: vpshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,4,5,5,7] ; AVX-NEXT: vpshufd {{.*#+}} xmm5 = xmm5[0,2,2,3] ; AVX-NEXT: vpblendw {{.*#+}} xmm5 = xmm5[0,1,2],xmm6[3],xmm5[4,5,6],xmm6[7] ; AVX-NEXT: vinsertf128 $1, %xmm4, %ymm5, %ymm4 ; AVX-NEXT: vpshufhw {{.*#+}} xmm5 = xmm3[0,1,2,3,4,6,6,7] ; AVX-NEXT: vpshufd {{.*#+}} xmm5 = xmm5[2,1,3,3] ; AVX-NEXT: vpshufd {{.*#+}} xmm6 = xmm1[3,3,3,3] ; AVX-NEXT: vpmovzxwq {{.*#+}} xmm6 = xmm6[0],zero,zero,zero,xmm6[1],zero,zero,zero ; AVX-NEXT: vpblendw {{.*#+}} xmm5 = xmm6[0],xmm5[1],xmm6[2,3,4],xmm5[5],xmm6[6,7] ; AVX-NEXT: vpshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,4,6,5] ; AVX-NEXT: vpshufd {{.*#+}} xmm3 = xmm3[2,1,3,3] ; AVX-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[2,3,2,3] ; AVX-NEXT: vpmovzxwq {{.*#+}} xmm1 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero ; AVX-NEXT: vpblendw {{.*#+}} xmm1 = xmm1[0],xmm3[1],xmm1[2,3,4],xmm3[5],xmm1[6,7] ; AVX-NEXT: vinsertf128 $1, %xmm5, %ymm1, %ymm1 ; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm4[1],ymm1[2],ymm4[3],ymm1[4],ymm4[5],ymm1[6],ymm4[7] ; AVX-NEXT: vmovaps %ymm1, 96(%rax) ; AVX-NEXT: vmovaps %ymm2, 64(%rax) ; AVX-NEXT: vmovaps %ymm7, 32(%rax) ; AVX-NEXT: vmovaps %ymm0, (%rax) ; AVX-NEXT: vzeroupper ; AVX-NEXT: retq ; ; AVX2-LABEL: store_i8_stride8_vf16: ; AVX2: # %bb.0: ; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %r11 ; AVX2-NEXT: vmovdqa (%rdi), %xmm0 ; AVX2-NEXT: vmovdqa (%rdx), %xmm1 ; AVX2-NEXT: vmovdqa (%r8), %xmm2 ; AVX2-NEXT: vmovdqa (%r11), %xmm3 ; AVX2-NEXT: vinserti128 $1, (%rsi), %ymm0, %ymm0 ; AVX2-NEXT: vinserti128 $1, (%rcx), %ymm1, %ymm1 ; AVX2-NEXT: vinserti128 $1, (%r9), %ymm2, %ymm4 ; AVX2-NEXT: vinserti128 $1, (%r10), %ymm3, %ymm3 ; AVX2-NEXT: vpermq {{.*#+}} ymm5 = ymm3[0,2,0,2] ; AVX2-NEXT: vmovdqa {{.*#+}} ymm6 = [u,u,u,u,u,u,0,8,u,u,u,u,u,u,1,9,u,u,u,u,u,u,2,10,u,u,u,u,u,u,3,11] ; AVX2-NEXT: vpshufb %ymm6, %ymm5, %ymm2 ; AVX2-NEXT: vpermq {{.*#+}} ymm7 = ymm4[0,2,0,2] ; AVX2-NEXT: vpmovsxwd {{.*#+}} ymm8 = [0,2048,0,2305,0,2562,0,2819] ; AVX2-NEXT: vpshufb %ymm8, %ymm7, %ymm9 ; AVX2-NEXT: vpblendw {{.*#+}} ymm2 = ymm9[0,1,2],ymm2[3],ymm9[4,5,6],ymm2[7],ymm9[8,9,10],ymm2[11],ymm9[12,13,14],ymm2[15] ; AVX2-NEXT: vpermq {{.*#+}} ymm9 = ymm1[0,2,0,2] ; AVX2-NEXT: vpmovsxdq {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm10 ; AVX2-NEXT: vpshufb %ymm10, %ymm9, %ymm11 ; AVX2-NEXT: vpermq {{.*#+}} ymm12 = ymm0[0,2,0,2] ; AVX2-NEXT: vpmovsxwq {{.*#+}} ymm13 = [2048,2305,2562,2819] ; AVX2-NEXT: vpshufb %ymm13, %ymm12, %ymm14 ; AVX2-NEXT: vpblendw {{.*#+}} ymm11 = ymm14[0],ymm11[1],ymm14[2,3,4],ymm11[5],ymm14[6,7,8],ymm11[9],ymm14[10,11,12],ymm11[13],ymm14[14,15] ; AVX2-NEXT: vpblendd {{.*#+}} ymm2 = ymm11[0],ymm2[1],ymm11[2],ymm2[3],ymm11[4],ymm2[5],ymm11[6],ymm2[7] ; AVX2-NEXT: vmovdqa {{.*#+}} ymm11 = [u,u,u,u,u,u,4,12,u,u,u,u,u,u,5,13,u,u,u,u,u,u,6,14,u,u,u,u,u,u,7,15] ; AVX2-NEXT: vpshufb %ymm11, %ymm5, %ymm5 ; AVX2-NEXT: vpmovsxwd {{.*#+}} ymm14 = [0,3076,0,3333,0,3590,0,3847] ; AVX2-NEXT: vpshufb %ymm14, %ymm7, %ymm7 ; AVX2-NEXT: vpblendw {{.*#+}} ymm5 = ymm7[0,1,2],ymm5[3],ymm7[4,5,6],ymm5[7],ymm7[8,9,10],ymm5[11],ymm7[12,13,14],ymm5[15] ; AVX2-NEXT: vpmovsxdq {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm7 ; AVX2-NEXT: vpshufb %ymm7, %ymm9, %ymm9 ; AVX2-NEXT: vpmovsxwq {{.*#+}} ymm15 = [3076,3333,3590,3847] ; AVX2-NEXT: vpshufb %ymm15, %ymm12, %ymm12 ; AVX2-NEXT: vpblendw {{.*#+}} ymm9 = ymm12[0],ymm9[1],ymm12[2,3,4],ymm9[5],ymm12[6,7,8],ymm9[9],ymm12[10,11,12],ymm9[13],ymm12[14,15] ; AVX2-NEXT: vpblendd {{.*#+}} ymm5 = ymm9[0],ymm5[1],ymm9[2],ymm5[3],ymm9[4],ymm5[5],ymm9[6],ymm5[7] ; AVX2-NEXT: vpermq {{.*#+}} ymm3 = ymm3[1,3,1,3] ; AVX2-NEXT: vpshufb %ymm6, %ymm3, %ymm6 ; AVX2-NEXT: vpermq {{.*#+}} ymm4 = ymm4[1,3,1,3] ; AVX2-NEXT: vpshufb %ymm8, %ymm4, %ymm8 ; AVX2-NEXT: vpblendw {{.*#+}} ymm6 = ymm8[0,1,2],ymm6[3],ymm8[4,5,6],ymm6[7],ymm8[8,9,10],ymm6[11],ymm8[12,13,14],ymm6[15] ; AVX2-NEXT: vpermq {{.*#+}} ymm1 = ymm1[1,3,1,3] ; AVX2-NEXT: vpshufb %ymm10, %ymm1, %ymm8 ; AVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[1,3,1,3] ; AVX2-NEXT: vpshufb %ymm13, %ymm0, %ymm9 ; AVX2-NEXT: vpblendw {{.*#+}} ymm8 = ymm9[0],ymm8[1],ymm9[2,3,4],ymm8[5],ymm9[6,7,8],ymm8[9],ymm9[10,11,12],ymm8[13],ymm9[14,15] ; AVX2-NEXT: vpblendd {{.*#+}} ymm6 = ymm8[0],ymm6[1],ymm8[2],ymm6[3],ymm8[4],ymm6[5],ymm8[6],ymm6[7] ; AVX2-NEXT: vpshufb %ymm11, %ymm3, %ymm3 ; AVX2-NEXT: vpshufb %ymm14, %ymm4, %ymm4 ; AVX2-NEXT: vpblendw {{.*#+}} ymm3 = ymm4[0,1,2],ymm3[3],ymm4[4,5,6],ymm3[7],ymm4[8,9,10],ymm3[11],ymm4[12,13,14],ymm3[15] ; AVX2-NEXT: vpshufb %ymm7, %ymm1, %ymm1 ; AVX2-NEXT: vpshufb %ymm15, %ymm0, %ymm0 ; AVX2-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2,3,4],ymm1[5],ymm0[6,7,8],ymm1[9],ymm0[10,11,12],ymm1[13],ymm0[14,15] ; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0],ymm3[1],ymm0[2],ymm3[3],ymm0[4],ymm3[5],ymm0[6],ymm3[7] ; AVX2-NEXT: vmovdqa %ymm0, 96(%rax) ; AVX2-NEXT: vmovdqa %ymm6, 64(%rax) ; AVX2-NEXT: vmovdqa %ymm5, 32(%rax) ; AVX2-NEXT: vmovdqa %ymm2, (%rax) ; AVX2-NEXT: vzeroupper ; AVX2-NEXT: retq ; ; AVX2-FP-LABEL: store_i8_stride8_vf16: ; AVX2-FP: # %bb.0: ; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %r11 ; AVX2-FP-NEXT: vmovdqa (%rdi), %xmm0 ; AVX2-FP-NEXT: vmovdqa (%rdx), %xmm1 ; AVX2-FP-NEXT: vmovdqa (%r8), %xmm2 ; AVX2-FP-NEXT: vmovdqa (%r11), %xmm3 ; AVX2-FP-NEXT: vinserti128 $1, (%rsi), %ymm0, %ymm0 ; AVX2-FP-NEXT: vinserti128 $1, (%rcx), %ymm1, %ymm1 ; AVX2-FP-NEXT: vinserti128 $1, (%r9), %ymm2, %ymm4 ; AVX2-FP-NEXT: vinserti128 $1, (%r10), %ymm3, %ymm3 ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm5 = ymm3[0,2,0,2] ; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm6 = [u,u,u,u,u,u,0,8,u,u,u,u,u,u,1,9,u,u,u,u,u,u,2,10,u,u,u,u,u,u,3,11] ; AVX2-FP-NEXT: vpshufb %ymm6, %ymm5, %ymm2 ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm7 = ymm4[0,2,0,2] ; AVX2-FP-NEXT: vpmovsxwd {{.*#+}} ymm8 = [0,2048,0,2305,0,2562,0,2819] ; AVX2-FP-NEXT: vpshufb %ymm8, %ymm7, %ymm9 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm2 = ymm9[0,1,2],ymm2[3],ymm9[4,5,6],ymm2[7],ymm9[8,9,10],ymm2[11],ymm9[12,13,14],ymm2[15] ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm9 = ymm1[0,2,0,2] ; AVX2-FP-NEXT: vpmovsxdq {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm10 ; AVX2-FP-NEXT: vpshufb %ymm10, %ymm9, %ymm11 ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm12 = ymm0[0,2,0,2] ; AVX2-FP-NEXT: vpmovsxwq {{.*#+}} ymm13 = [2048,2305,2562,2819] ; AVX2-FP-NEXT: vpshufb %ymm13, %ymm12, %ymm14 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm11 = ymm14[0],ymm11[1],ymm14[2,3,4],ymm11[5],ymm14[6,7,8],ymm11[9],ymm14[10,11,12],ymm11[13],ymm14[14,15] ; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm2 = ymm11[0],ymm2[1],ymm11[2],ymm2[3],ymm11[4],ymm2[5],ymm11[6],ymm2[7] ; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm11 = [u,u,u,u,u,u,4,12,u,u,u,u,u,u,5,13,u,u,u,u,u,u,6,14,u,u,u,u,u,u,7,15] ; AVX2-FP-NEXT: vpshufb %ymm11, %ymm5, %ymm5 ; AVX2-FP-NEXT: vpmovsxwd {{.*#+}} ymm14 = [0,3076,0,3333,0,3590,0,3847] ; AVX2-FP-NEXT: vpshufb %ymm14, %ymm7, %ymm7 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm5 = ymm7[0,1,2],ymm5[3],ymm7[4,5,6],ymm5[7],ymm7[8,9,10],ymm5[11],ymm7[12,13,14],ymm5[15] ; AVX2-FP-NEXT: vpmovsxdq {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm7 ; AVX2-FP-NEXT: vpshufb %ymm7, %ymm9, %ymm9 ; AVX2-FP-NEXT: vpmovsxwq {{.*#+}} ymm15 = [3076,3333,3590,3847] ; AVX2-FP-NEXT: vpshufb %ymm15, %ymm12, %ymm12 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm9 = ymm12[0],ymm9[1],ymm12[2,3,4],ymm9[5],ymm12[6,7,8],ymm9[9],ymm12[10,11,12],ymm9[13],ymm12[14,15] ; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm5 = ymm9[0],ymm5[1],ymm9[2],ymm5[3],ymm9[4],ymm5[5],ymm9[6],ymm5[7] ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm3 = ymm3[1,3,1,3] ; AVX2-FP-NEXT: vpshufb %ymm6, %ymm3, %ymm6 ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[1,3,1,3] ; AVX2-FP-NEXT: vpshufb %ymm8, %ymm4, %ymm8 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm6 = ymm8[0,1,2],ymm6[3],ymm8[4,5,6],ymm6[7],ymm8[8,9,10],ymm6[11],ymm8[12,13,14],ymm6[15] ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[1,3,1,3] ; AVX2-FP-NEXT: vpshufb %ymm10, %ymm1, %ymm8 ; AVX2-FP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[1,3,1,3] ; AVX2-FP-NEXT: vpshufb %ymm13, %ymm0, %ymm9 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm8 = ymm9[0],ymm8[1],ymm9[2,3,4],ymm8[5],ymm9[6,7,8],ymm8[9],ymm9[10,11,12],ymm8[13],ymm9[14,15] ; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm6 = ymm8[0],ymm6[1],ymm8[2],ymm6[3],ymm8[4],ymm6[5],ymm8[6],ymm6[7] ; AVX2-FP-NEXT: vpshufb %ymm11, %ymm3, %ymm3 ; AVX2-FP-NEXT: vpshufb %ymm14, %ymm4, %ymm4 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm3 = ymm4[0,1,2],ymm3[3],ymm4[4,5,6],ymm3[7],ymm4[8,9,10],ymm3[11],ymm4[12,13,14],ymm3[15] ; AVX2-FP-NEXT: vpshufb %ymm7, %ymm1, %ymm1 ; AVX2-FP-NEXT: vpshufb %ymm15, %ymm0, %ymm0 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2,3,4],ymm1[5],ymm0[6,7,8],ymm1[9],ymm0[10,11,12],ymm1[13],ymm0[14,15] ; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0],ymm3[1],ymm0[2],ymm3[3],ymm0[4],ymm3[5],ymm0[6],ymm3[7] ; AVX2-FP-NEXT: vmovdqa %ymm0, 96(%rax) ; AVX2-FP-NEXT: vmovdqa %ymm6, 64(%rax) ; AVX2-FP-NEXT: vmovdqa %ymm5, 32(%rax) ; AVX2-FP-NEXT: vmovdqa %ymm2, (%rax) ; AVX2-FP-NEXT: vzeroupper ; AVX2-FP-NEXT: retq ; ; AVX2-FCP-LABEL: store_i8_stride8_vf16: ; AVX2-FCP: # %bb.0: ; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r11 ; AVX2-FCP-NEXT: vmovdqa (%rdi), %xmm0 ; AVX2-FCP-NEXT: vmovdqa (%rdx), %xmm1 ; AVX2-FCP-NEXT: vmovdqa (%r8), %xmm2 ; AVX2-FCP-NEXT: vmovdqa (%r11), %xmm3 ; AVX2-FCP-NEXT: vinserti128 $1, (%rsi), %ymm0, %ymm0 ; AVX2-FCP-NEXT: vinserti128 $1, (%rcx), %ymm1, %ymm1 ; AVX2-FCP-NEXT: vinserti128 $1, (%r9), %ymm2, %ymm4 ; AVX2-FCP-NEXT: vinserti128 $1, (%r10), %ymm3, %ymm3 ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm5 = ymm3[0,2,0,2] ; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm6 = [u,u,u,u,u,u,0,8,u,u,u,u,u,u,1,9,u,u,u,u,u,u,2,10,u,u,u,u,u,u,3,11] ; AVX2-FCP-NEXT: vpshufb %ymm6, %ymm5, %ymm2 ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm7 = ymm4[0,2,0,2] ; AVX2-FCP-NEXT: vpmovsxwd {{.*#+}} ymm8 = [0,2048,0,2305,0,2562,0,2819] ; AVX2-FCP-NEXT: vpshufb %ymm8, %ymm7, %ymm9 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm9[0,1,2],ymm2[3],ymm9[4,5,6],ymm2[7],ymm9[8,9,10],ymm2[11],ymm9[12,13,14],ymm2[15] ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm9 = ymm1[0,2,0,2] ; AVX2-FCP-NEXT: vpmovsxdq {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm10 ; AVX2-FCP-NEXT: vpshufb %ymm10, %ymm9, %ymm11 ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm12 = ymm0[0,2,0,2] ; AVX2-FCP-NEXT: vpmovsxwq {{.*#+}} ymm13 = [2048,2305,2562,2819] ; AVX2-FCP-NEXT: vpshufb %ymm13, %ymm12, %ymm14 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm11 = ymm14[0],ymm11[1],ymm14[2,3,4],ymm11[5],ymm14[6,7,8],ymm11[9],ymm14[10,11,12],ymm11[13],ymm14[14,15] ; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm11[0],ymm2[1],ymm11[2],ymm2[3],ymm11[4],ymm2[5],ymm11[6],ymm2[7] ; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm11 = [u,u,u,u,u,u,4,12,u,u,u,u,u,u,5,13,u,u,u,u,u,u,6,14,u,u,u,u,u,u,7,15] ; AVX2-FCP-NEXT: vpshufb %ymm11, %ymm5, %ymm5 ; AVX2-FCP-NEXT: vpmovsxwd {{.*#+}} ymm14 = [0,3076,0,3333,0,3590,0,3847] ; AVX2-FCP-NEXT: vpshufb %ymm14, %ymm7, %ymm7 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm5 = ymm7[0,1,2],ymm5[3],ymm7[4,5,6],ymm5[7],ymm7[8,9,10],ymm5[11],ymm7[12,13,14],ymm5[15] ; AVX2-FCP-NEXT: vpmovsxdq {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm7 ; AVX2-FCP-NEXT: vpshufb %ymm7, %ymm9, %ymm9 ; AVX2-FCP-NEXT: vpmovsxwq {{.*#+}} ymm15 = [3076,3333,3590,3847] ; AVX2-FCP-NEXT: vpshufb %ymm15, %ymm12, %ymm12 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm9 = ymm12[0],ymm9[1],ymm12[2,3,4],ymm9[5],ymm12[6,7,8],ymm9[9],ymm12[10,11,12],ymm9[13],ymm12[14,15] ; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm5 = ymm9[0],ymm5[1],ymm9[2],ymm5[3],ymm9[4],ymm5[5],ymm9[6],ymm5[7] ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm3[1,3,1,3] ; AVX2-FCP-NEXT: vpshufb %ymm6, %ymm3, %ymm6 ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[1,3,1,3] ; AVX2-FCP-NEXT: vpshufb %ymm8, %ymm4, %ymm8 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm6 = ymm8[0,1,2],ymm6[3],ymm8[4,5,6],ymm6[7],ymm8[8,9,10],ymm6[11],ymm8[12,13,14],ymm6[15] ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[1,3,1,3] ; AVX2-FCP-NEXT: vpshufb %ymm10, %ymm1, %ymm8 ; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[1,3,1,3] ; AVX2-FCP-NEXT: vpshufb %ymm13, %ymm0, %ymm9 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm8 = ymm9[0],ymm8[1],ymm9[2,3,4],ymm8[5],ymm9[6,7,8],ymm8[9],ymm9[10,11,12],ymm8[13],ymm9[14,15] ; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm8[0],ymm6[1],ymm8[2],ymm6[3],ymm8[4],ymm6[5],ymm8[6],ymm6[7] ; AVX2-FCP-NEXT: vpshufb %ymm11, %ymm3, %ymm3 ; AVX2-FCP-NEXT: vpshufb %ymm14, %ymm4, %ymm4 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm4[0,1,2],ymm3[3],ymm4[4,5,6],ymm3[7],ymm4[8,9,10],ymm3[11],ymm4[12,13,14],ymm3[15] ; AVX2-FCP-NEXT: vpshufb %ymm7, %ymm1, %ymm1 ; AVX2-FCP-NEXT: vpshufb %ymm15, %ymm0, %ymm0 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2,3,4],ymm1[5],ymm0[6,7,8],ymm1[9],ymm0[10,11,12],ymm1[13],ymm0[14,15] ; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0],ymm3[1],ymm0[2],ymm3[3],ymm0[4],ymm3[5],ymm0[6],ymm3[7] ; AVX2-FCP-NEXT: vmovdqa %ymm0, 96(%rax) ; AVX2-FCP-NEXT: vmovdqa %ymm6, 64(%rax) ; AVX2-FCP-NEXT: vmovdqa %ymm5, 32(%rax) ; AVX2-FCP-NEXT: vmovdqa %ymm2, (%rax) ; AVX2-FCP-NEXT: vzeroupper ; AVX2-FCP-NEXT: retq ; ; AVX512-LABEL: store_i8_stride8_vf16: ; AVX512: # %bb.0: ; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %r11 ; AVX512-NEXT: vmovdqa (%rdi), %xmm1 ; AVX512-NEXT: vmovdqa (%rdx), %xmm2 ; AVX512-NEXT: vmovdqa (%r8), %xmm3 ; AVX512-NEXT: vmovdqa (%r11), %xmm0 ; AVX512-NEXT: vinserti128 $1, (%r10), %ymm0, %ymm4 ; AVX512-NEXT: vpermq {{.*#+}} ymm5 = ymm4[0,2,0,2] ; AVX512-NEXT: vmovdqa {{.*#+}} ymm0 = [u,u,u,u,u,u,4,12,u,u,u,u,u,u,5,13,u,u,u,u,u,u,6,14,u,u,u,u,u,u,7,15] ; AVX512-NEXT: vpshufb %ymm0, %ymm5, %ymm6 ; AVX512-NEXT: vinserti128 $1, (%r9), %ymm3, %ymm3 ; AVX512-NEXT: vpermq {{.*#+}} ymm7 = ymm3[0,2,0,2] ; AVX512-NEXT: vpmovsxwd {{.*#+}} ymm8 = [0,3076,0,3333,0,3590,0,3847] ; AVX512-NEXT: vpshufb %ymm8, %ymm7, %ymm9 ; AVX512-NEXT: vpblendw {{.*#+}} ymm6 = ymm9[0,1,2],ymm6[3],ymm9[4,5,6],ymm6[7],ymm9[8,9,10],ymm6[11],ymm9[12,13,14],ymm6[15] ; AVX512-NEXT: vinserti128 $1, (%rcx), %ymm2, %ymm2 ; AVX512-NEXT: vpermq {{.*#+}} ymm9 = ymm2[0,2,0,2] ; AVX512-NEXT: vpmovsxdq {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm10 ; AVX512-NEXT: vpshufb %ymm10, %ymm9, %ymm11 ; AVX512-NEXT: vinserti128 $1, (%rsi), %ymm1, %ymm1 ; AVX512-NEXT: vpermq {{.*#+}} ymm12 = ymm1[0,2,0,2] ; AVX512-NEXT: vpmovsxwq {{.*#+}} ymm13 = [3076,3333,3590,3847] ; AVX512-NEXT: vpshufb %ymm13, %ymm12, %ymm14 ; AVX512-NEXT: vpblendw {{.*#+}} ymm11 = ymm14[0],ymm11[1],ymm14[2,3,4],ymm11[5],ymm14[6,7,8],ymm11[9],ymm14[10,11,12],ymm11[13],ymm14[14,15] ; AVX512-NEXT: vpblendd {{.*#+}} ymm6 = ymm11[0],ymm6[1],ymm11[2],ymm6[3],ymm11[4],ymm6[5],ymm11[6],ymm6[7] ; AVX512-NEXT: vmovdqa {{.*#+}} ymm11 = [u,u,u,u,u,u,0,8,u,u,u,u,u,u,1,9,u,u,u,u,u,u,2,10,u,u,u,u,u,u,3,11] ; AVX512-NEXT: vpshufb %ymm11, %ymm5, %ymm5 ; AVX512-NEXT: vpmovsxwd {{.*#+}} ymm14 = [0,2048,0,2305,0,2562,0,2819] ; AVX512-NEXT: vpshufb %ymm14, %ymm7, %ymm7 ; AVX512-NEXT: vpblendw {{.*#+}} ymm5 = ymm7[0,1,2],ymm5[3],ymm7[4,5,6],ymm5[7],ymm7[8,9,10],ymm5[11],ymm7[12,13,14],ymm5[15] ; AVX512-NEXT: vpmovsxdq {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm7 ; AVX512-NEXT: vpshufb %ymm7, %ymm9, %ymm9 ; AVX512-NEXT: vpmovsxwq {{.*#+}} ymm15 = [2048,2305,2562,2819] ; AVX512-NEXT: vpshufb %ymm15, %ymm12, %ymm12 ; AVX512-NEXT: vpblendw {{.*#+}} ymm9 = ymm12[0],ymm9[1],ymm12[2,3,4],ymm9[5],ymm12[6,7,8],ymm9[9],ymm12[10,11,12],ymm9[13],ymm12[14,15] ; AVX512-NEXT: vpblendd {{.*#+}} ymm5 = ymm9[0],ymm5[1],ymm9[2],ymm5[3],ymm9[4],ymm5[5],ymm9[6],ymm5[7] ; AVX512-NEXT: vinserti64x4 $1, %ymm6, %zmm5, %zmm5 ; AVX512-NEXT: vpermq {{.*#+}} ymm4 = ymm4[1,3,1,3] ; AVX512-NEXT: vpshufb %ymm0, %ymm4, %ymm0 ; AVX512-NEXT: vpermq {{.*#+}} ymm3 = ymm3[1,3,1,3] ; AVX512-NEXT: vpshufb %ymm8, %ymm3, %ymm6 ; AVX512-NEXT: vpblendw {{.*#+}} ymm0 = ymm6[0,1,2],ymm0[3],ymm6[4,5,6],ymm0[7],ymm6[8,9,10],ymm0[11],ymm6[12,13,14],ymm0[15] ; AVX512-NEXT: vpermq {{.*#+}} ymm2 = ymm2[1,3,1,3] ; AVX512-NEXT: vpshufb %ymm10, %ymm2, %ymm6 ; AVX512-NEXT: vpermq {{.*#+}} ymm1 = ymm1[1,3,1,3] ; AVX512-NEXT: vpshufb %ymm13, %ymm1, %ymm8 ; AVX512-NEXT: vpblendw {{.*#+}} ymm6 = ymm8[0],ymm6[1],ymm8[2,3,4],ymm6[5],ymm8[6,7,8],ymm6[9],ymm8[10,11,12],ymm6[13],ymm8[14,15] ; AVX512-NEXT: vpblendd {{.*#+}} ymm0 = ymm6[0],ymm0[1],ymm6[2],ymm0[3],ymm6[4],ymm0[5],ymm6[6],ymm0[7] ; AVX512-NEXT: vpshufb %ymm11, %ymm4, %ymm4 ; AVX512-NEXT: vpshufb %ymm14, %ymm3, %ymm3 ; AVX512-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0,1,2],ymm4[3],ymm3[4,5,6],ymm4[7],ymm3[8,9,10],ymm4[11],ymm3[12,13,14],ymm4[15] ; AVX512-NEXT: vpshufb %ymm7, %ymm2, %ymm2 ; AVX512-NEXT: vpshufb %ymm15, %ymm1, %ymm1 ; AVX512-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3,4],ymm2[5],ymm1[6,7,8],ymm2[9],ymm1[10,11,12],ymm2[13],ymm1[14,15] ; AVX512-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0],ymm3[1],ymm1[2],ymm3[3],ymm1[4],ymm3[5],ymm1[6],ymm3[7] ; AVX512-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0 ; AVX512-NEXT: vmovdqa64 %zmm0, 64(%rax) ; AVX512-NEXT: vmovdqa64 %zmm5, (%rax) ; AVX512-NEXT: vzeroupper ; AVX512-NEXT: retq ; ; AVX512-FCP-LABEL: store_i8_stride8_vf16: ; AVX512-FCP: # %bb.0: ; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r11 ; AVX512-FCP-NEXT: vmovdqa (%rdi), %xmm1 ; AVX512-FCP-NEXT: vmovdqa (%rdx), %xmm2 ; AVX512-FCP-NEXT: vmovdqa (%r8), %xmm3 ; AVX512-FCP-NEXT: vmovdqa (%r11), %xmm0 ; AVX512-FCP-NEXT: vinserti128 $1, (%r10), %ymm0, %ymm4 ; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm5 = ymm4[0,2,0,2] ; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm0 = [u,u,u,u,u,u,4,12,u,u,u,u,u,u,5,13,u,u,u,u,u,u,6,14,u,u,u,u,u,u,7,15] ; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm5, %ymm6 ; AVX512-FCP-NEXT: vinserti128 $1, (%r9), %ymm3, %ymm3 ; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm7 = ymm3[0,2,0,2] ; AVX512-FCP-NEXT: vpmovsxwd {{.*#+}} ymm8 = [0,3076,0,3333,0,3590,0,3847] ; AVX512-FCP-NEXT: vpshufb %ymm8, %ymm7, %ymm9 ; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm6 = ymm9[0,1,2],ymm6[3],ymm9[4,5,6],ymm6[7],ymm9[8,9,10],ymm6[11],ymm9[12,13,14],ymm6[15] ; AVX512-FCP-NEXT: vinserti128 $1, (%rcx), %ymm2, %ymm2 ; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm9 = ymm2[0,2,0,2] ; AVX512-FCP-NEXT: vpmovsxdq {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm10 ; AVX512-FCP-NEXT: vpshufb %ymm10, %ymm9, %ymm11 ; AVX512-FCP-NEXT: vinserti128 $1, (%rsi), %ymm1, %ymm1 ; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm12 = ymm1[0,2,0,2] ; AVX512-FCP-NEXT: vpmovsxwq {{.*#+}} ymm13 = [3076,3333,3590,3847] ; AVX512-FCP-NEXT: vpshufb %ymm13, %ymm12, %ymm14 ; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm11 = ymm14[0],ymm11[1],ymm14[2,3,4],ymm11[5],ymm14[6,7,8],ymm11[9],ymm14[10,11,12],ymm11[13],ymm14[14,15] ; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm11[0],ymm6[1],ymm11[2],ymm6[3],ymm11[4],ymm6[5],ymm11[6],ymm6[7] ; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm11 = [u,u,u,u,u,u,0,8,u,u,u,u,u,u,1,9,u,u,u,u,u,u,2,10,u,u,u,u,u,u,3,11] ; AVX512-FCP-NEXT: vpshufb %ymm11, %ymm5, %ymm5 ; AVX512-FCP-NEXT: vpmovsxwd {{.*#+}} ymm14 = [0,2048,0,2305,0,2562,0,2819] ; AVX512-FCP-NEXT: vpshufb %ymm14, %ymm7, %ymm7 ; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm5 = ymm7[0,1,2],ymm5[3],ymm7[4,5,6],ymm5[7],ymm7[8,9,10],ymm5[11],ymm7[12,13,14],ymm5[15] ; AVX512-FCP-NEXT: vpmovsxdq {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm7 ; AVX512-FCP-NEXT: vpshufb %ymm7, %ymm9, %ymm9 ; AVX512-FCP-NEXT: vpmovsxwq {{.*#+}} ymm15 = [2048,2305,2562,2819] ; AVX512-FCP-NEXT: vpshufb %ymm15, %ymm12, %ymm12 ; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm9 = ymm12[0],ymm9[1],ymm12[2,3,4],ymm9[5],ymm12[6,7,8],ymm9[9],ymm12[10,11,12],ymm9[13],ymm12[14,15] ; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm5 = ymm9[0],ymm5[1],ymm9[2],ymm5[3],ymm9[4],ymm5[5],ymm9[6],ymm5[7] ; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm6, %zmm5, %zmm5 ; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[1,3,1,3] ; AVX512-FCP-NEXT: vpshufb %ymm0, %ymm4, %ymm0 ; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm3[1,3,1,3] ; AVX512-FCP-NEXT: vpshufb %ymm8, %ymm3, %ymm6 ; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm6[0,1,2],ymm0[3],ymm6[4,5,6],ymm0[7],ymm6[8,9,10],ymm0[11],ymm6[12,13,14],ymm0[15] ; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[1,3,1,3] ; AVX512-FCP-NEXT: vpshufb %ymm10, %ymm2, %ymm6 ; AVX512-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[1,3,1,3] ; AVX512-FCP-NEXT: vpshufb %ymm13, %ymm1, %ymm8 ; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm6 = ymm8[0],ymm6[1],ymm8[2,3,4],ymm6[5],ymm8[6,7,8],ymm6[9],ymm8[10,11,12],ymm6[13],ymm8[14,15] ; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm6[0],ymm0[1],ymm6[2],ymm0[3],ymm6[4],ymm0[5],ymm6[6],ymm0[7] ; AVX512-FCP-NEXT: vpshufb %ymm11, %ymm4, %ymm4 ; AVX512-FCP-NEXT: vpshufb %ymm14, %ymm3, %ymm3 ; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0,1,2],ymm4[3],ymm3[4,5,6],ymm4[7],ymm3[8,9,10],ymm4[11],ymm3[12,13,14],ymm4[15] ; AVX512-FCP-NEXT: vpshufb %ymm7, %ymm2, %ymm2 ; AVX512-FCP-NEXT: vpshufb %ymm15, %ymm1, %ymm1 ; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3,4],ymm2[5],ymm1[6,7,8],ymm2[9],ymm1[10,11,12],ymm2[13],ymm1[14,15] ; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0],ymm3[1],ymm1[2],ymm3[3],ymm1[4],ymm3[5],ymm1[6],ymm3[7] ; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0 ; AVX512-FCP-NEXT: vmovdqa64 %zmm0, 64(%rax) ; AVX512-FCP-NEXT: vmovdqa64 %zmm5, (%rax) ; AVX512-FCP-NEXT: vzeroupper ; AVX512-FCP-NEXT: retq ; ; AVX512DQ-LABEL: store_i8_stride8_vf16: ; AVX512DQ: # %bb.0: ; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %r11 ; AVX512DQ-NEXT: vmovdqa (%rdi), %xmm1 ; AVX512DQ-NEXT: vmovdqa (%rdx), %xmm2 ; AVX512DQ-NEXT: vmovdqa (%r8), %xmm3 ; AVX512DQ-NEXT: vmovdqa (%r11), %xmm0 ; AVX512DQ-NEXT: vinserti128 $1, (%r10), %ymm0, %ymm4 ; AVX512DQ-NEXT: vpermq {{.*#+}} ymm5 = ymm4[0,2,0,2] ; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm0 = [u,u,u,u,u,u,4,12,u,u,u,u,u,u,5,13,u,u,u,u,u,u,6,14,u,u,u,u,u,u,7,15] ; AVX512DQ-NEXT: vpshufb %ymm0, %ymm5, %ymm6 ; AVX512DQ-NEXT: vinserti128 $1, (%r9), %ymm3, %ymm3 ; AVX512DQ-NEXT: vpermq {{.*#+}} ymm7 = ymm3[0,2,0,2] ; AVX512DQ-NEXT: vpmovsxwd {{.*#+}} ymm8 = [0,3076,0,3333,0,3590,0,3847] ; AVX512DQ-NEXT: vpshufb %ymm8, %ymm7, %ymm9 ; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm6 = ymm9[0,1,2],ymm6[3],ymm9[4,5,6],ymm6[7],ymm9[8,9,10],ymm6[11],ymm9[12,13,14],ymm6[15] ; AVX512DQ-NEXT: vinserti128 $1, (%rcx), %ymm2, %ymm2 ; AVX512DQ-NEXT: vpermq {{.*#+}} ymm9 = ymm2[0,2,0,2] ; AVX512DQ-NEXT: vpmovsxdq {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm10 ; AVX512DQ-NEXT: vpshufb %ymm10, %ymm9, %ymm11 ; AVX512DQ-NEXT: vinserti128 $1, (%rsi), %ymm1, %ymm1 ; AVX512DQ-NEXT: vpermq {{.*#+}} ymm12 = ymm1[0,2,0,2] ; AVX512DQ-NEXT: vpmovsxwq {{.*#+}} ymm13 = [3076,3333,3590,3847] ; AVX512DQ-NEXT: vpshufb %ymm13, %ymm12, %ymm14 ; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm11 = ymm14[0],ymm11[1],ymm14[2,3,4],ymm11[5],ymm14[6,7,8],ymm11[9],ymm14[10,11,12],ymm11[13],ymm14[14,15] ; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm6 = ymm11[0],ymm6[1],ymm11[2],ymm6[3],ymm11[4],ymm6[5],ymm11[6],ymm6[7] ; AVX512DQ-NEXT: vmovdqa {{.*#+}} ymm11 = [u,u,u,u,u,u,0,8,u,u,u,u,u,u,1,9,u,u,u,u,u,u,2,10,u,u,u,u,u,u,3,11] ; AVX512DQ-NEXT: vpshufb %ymm11, %ymm5, %ymm5 ; AVX512DQ-NEXT: vpmovsxwd {{.*#+}} ymm14 = [0,2048,0,2305,0,2562,0,2819] ; AVX512DQ-NEXT: vpshufb %ymm14, %ymm7, %ymm7 ; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm5 = ymm7[0,1,2],ymm5[3],ymm7[4,5,6],ymm5[7],ymm7[8,9,10],ymm5[11],ymm7[12,13,14],ymm5[15] ; AVX512DQ-NEXT: vpmovsxdq {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm7 ; AVX512DQ-NEXT: vpshufb %ymm7, %ymm9, %ymm9 ; AVX512DQ-NEXT: vpmovsxwq {{.*#+}} ymm15 = [2048,2305,2562,2819] ; AVX512DQ-NEXT: vpshufb %ymm15, %ymm12, %ymm12 ; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm9 = ymm12[0],ymm9[1],ymm12[2,3,4],ymm9[5],ymm12[6,7,8],ymm9[9],ymm12[10,11,12],ymm9[13],ymm12[14,15] ; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm5 = ymm9[0],ymm5[1],ymm9[2],ymm5[3],ymm9[4],ymm5[5],ymm9[6],ymm5[7] ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm6, %zmm5, %zmm5 ; AVX512DQ-NEXT: vpermq {{.*#+}} ymm4 = ymm4[1,3,1,3] ; AVX512DQ-NEXT: vpshufb %ymm0, %ymm4, %ymm0 ; AVX512DQ-NEXT: vpermq {{.*#+}} ymm3 = ymm3[1,3,1,3] ; AVX512DQ-NEXT: vpshufb %ymm8, %ymm3, %ymm6 ; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm0 = ymm6[0,1,2],ymm0[3],ymm6[4,5,6],ymm0[7],ymm6[8,9,10],ymm0[11],ymm6[12,13,14],ymm0[15] ; AVX512DQ-NEXT: vpermq {{.*#+}} ymm2 = ymm2[1,3,1,3] ; AVX512DQ-NEXT: vpshufb %ymm10, %ymm2, %ymm6 ; AVX512DQ-NEXT: vpermq {{.*#+}} ymm1 = ymm1[1,3,1,3] ; AVX512DQ-NEXT: vpshufb %ymm13, %ymm1, %ymm8 ; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm6 = ymm8[0],ymm6[1],ymm8[2,3,4],ymm6[5],ymm8[6,7,8],ymm6[9],ymm8[10,11,12],ymm6[13],ymm8[14,15] ; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm0 = ymm6[0],ymm0[1],ymm6[2],ymm0[3],ymm6[4],ymm0[5],ymm6[6],ymm0[7] ; AVX512DQ-NEXT: vpshufb %ymm11, %ymm4, %ymm4 ; AVX512DQ-NEXT: vpshufb %ymm14, %ymm3, %ymm3 ; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0,1,2],ymm4[3],ymm3[4,5,6],ymm4[7],ymm3[8,9,10],ymm4[11],ymm3[12,13,14],ymm4[15] ; AVX512DQ-NEXT: vpshufb %ymm7, %ymm2, %ymm2 ; AVX512DQ-NEXT: vpshufb %ymm15, %ymm1, %ymm1 ; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3,4],ymm2[5],ymm1[6,7,8],ymm2[9],ymm1[10,11,12],ymm2[13],ymm1[14,15] ; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0],ymm3[1],ymm1[2],ymm3[3],ymm1[4],ymm3[5],ymm1[6],ymm3[7] ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0 ; AVX512DQ-NEXT: vmovdqa64 %zmm0, 64(%rax) ; AVX512DQ-NEXT: vmovdqa64 %zmm5, (%rax) ; AVX512DQ-NEXT: vzeroupper ; AVX512DQ-NEXT: retq ; ; AVX512DQ-FCP-LABEL: store_i8_stride8_vf16: ; AVX512DQ-FCP: # %bb.0: ; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r11 ; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %xmm1 ; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %xmm2 ; AVX512DQ-FCP-NEXT: vmovdqa (%r8), %xmm3 ; AVX512DQ-FCP-NEXT: vmovdqa (%r11), %xmm0 ; AVX512DQ-FCP-NEXT: vinserti128 $1, (%r10), %ymm0, %ymm4 ; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm5 = ymm4[0,2,0,2] ; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm0 = [u,u,u,u,u,u,4,12,u,u,u,u,u,u,5,13,u,u,u,u,u,u,6,14,u,u,u,u,u,u,7,15] ; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm5, %ymm6 ; AVX512DQ-FCP-NEXT: vinserti128 $1, (%r9), %ymm3, %ymm3 ; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm7 = ymm3[0,2,0,2] ; AVX512DQ-FCP-NEXT: vpmovsxwd {{.*#+}} ymm8 = [0,3076,0,3333,0,3590,0,3847] ; AVX512DQ-FCP-NEXT: vpshufb %ymm8, %ymm7, %ymm9 ; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm6 = ymm9[0,1,2],ymm6[3],ymm9[4,5,6],ymm6[7],ymm9[8,9,10],ymm6[11],ymm9[12,13,14],ymm6[15] ; AVX512DQ-FCP-NEXT: vinserti128 $1, (%rcx), %ymm2, %ymm2 ; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm9 = ymm2[0,2,0,2] ; AVX512DQ-FCP-NEXT: vpmovsxdq {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm10 ; AVX512DQ-FCP-NEXT: vpshufb %ymm10, %ymm9, %ymm11 ; AVX512DQ-FCP-NEXT: vinserti128 $1, (%rsi), %ymm1, %ymm1 ; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm12 = ymm1[0,2,0,2] ; AVX512DQ-FCP-NEXT: vpmovsxwq {{.*#+}} ymm13 = [3076,3333,3590,3847] ; AVX512DQ-FCP-NEXT: vpshufb %ymm13, %ymm12, %ymm14 ; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm11 = ymm14[0],ymm11[1],ymm14[2,3,4],ymm11[5],ymm14[6,7,8],ymm11[9],ymm14[10,11,12],ymm11[13],ymm14[14,15] ; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm11[0],ymm6[1],ymm11[2],ymm6[3],ymm11[4],ymm6[5],ymm11[6],ymm6[7] ; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm11 = [u,u,u,u,u,u,0,8,u,u,u,u,u,u,1,9,u,u,u,u,u,u,2,10,u,u,u,u,u,u,3,11] ; AVX512DQ-FCP-NEXT: vpshufb %ymm11, %ymm5, %ymm5 ; AVX512DQ-FCP-NEXT: vpmovsxwd {{.*#+}} ymm14 = [0,2048,0,2305,0,2562,0,2819] ; AVX512DQ-FCP-NEXT: vpshufb %ymm14, %ymm7, %ymm7 ; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm5 = ymm7[0,1,2],ymm5[3],ymm7[4,5,6],ymm5[7],ymm7[8,9,10],ymm5[11],ymm7[12,13,14],ymm5[15] ; AVX512DQ-FCP-NEXT: vpmovsxdq {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm7 ; AVX512DQ-FCP-NEXT: vpshufb %ymm7, %ymm9, %ymm9 ; AVX512DQ-FCP-NEXT: vpmovsxwq {{.*#+}} ymm15 = [2048,2305,2562,2819] ; AVX512DQ-FCP-NEXT: vpshufb %ymm15, %ymm12, %ymm12 ; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm9 = ymm12[0],ymm9[1],ymm12[2,3,4],ymm9[5],ymm12[6,7,8],ymm9[9],ymm12[10,11,12],ymm9[13],ymm12[14,15] ; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm5 = ymm9[0],ymm5[1],ymm9[2],ymm5[3],ymm9[4],ymm5[5],ymm9[6],ymm5[7] ; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm6, %zmm5, %zmm5 ; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[1,3,1,3] ; AVX512DQ-FCP-NEXT: vpshufb %ymm0, %ymm4, %ymm0 ; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm3[1,3,1,3] ; AVX512DQ-FCP-NEXT: vpshufb %ymm8, %ymm3, %ymm6 ; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm6[0,1,2],ymm0[3],ymm6[4,5,6],ymm0[7],ymm6[8,9,10],ymm0[11],ymm6[12,13,14],ymm0[15] ; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[1,3,1,3] ; AVX512DQ-FCP-NEXT: vpshufb %ymm10, %ymm2, %ymm6 ; AVX512DQ-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[1,3,1,3] ; AVX512DQ-FCP-NEXT: vpshufb %ymm13, %ymm1, %ymm8 ; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm6 = ymm8[0],ymm6[1],ymm8[2,3,4],ymm6[5],ymm8[6,7,8],ymm6[9],ymm8[10,11,12],ymm6[13],ymm8[14,15] ; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm6[0],ymm0[1],ymm6[2],ymm0[3],ymm6[4],ymm0[5],ymm6[6],ymm0[7] ; AVX512DQ-FCP-NEXT: vpshufb %ymm11, %ymm4, %ymm4 ; AVX512DQ-FCP-NEXT: vpshufb %ymm14, %ymm3, %ymm3 ; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0,1,2],ymm4[3],ymm3[4,5,6],ymm4[7],ymm3[8,9,10],ymm4[11],ymm3[12,13,14],ymm4[15] ; AVX512DQ-FCP-NEXT: vpshufb %ymm7, %ymm2, %ymm2 ; AVX512DQ-FCP-NEXT: vpshufb %ymm15, %ymm1, %ymm1 ; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3,4],ymm2[5],ymm1[6,7,8],ymm2[9],ymm1[10,11,12],ymm2[13],ymm1[14,15] ; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0],ymm3[1],ymm1[2],ymm3[3],ymm1[4],ymm3[5],ymm1[6],ymm3[7] ; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0 ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, 64(%rax) ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm5, (%rax) ; AVX512DQ-FCP-NEXT: vzeroupper ; AVX512DQ-FCP-NEXT: retq ; ; AVX512BW-LABEL: store_i8_stride8_vf16: ; AVX512BW: # %bb.0: ; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %r11 ; AVX512BW-NEXT: vmovdqa (%rdi), %xmm0 ; AVX512BW-NEXT: vmovdqa (%rdx), %xmm1 ; AVX512BW-NEXT: vmovdqa (%r8), %xmm2 ; AVX512BW-NEXT: vmovdqa (%r11), %xmm3 ; AVX512BW-NEXT: vinserti128 $1, (%rsi), %ymm0, %ymm0 ; AVX512BW-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm0 ; AVX512BW-NEXT: vpermq {{.*#+}} zmm4 = zmm0[0,2,0,2,4,6,4,6] ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm5 = [0,8,128,128,128,128,128,128,1,9,128,128,128,128,128,128,18,26,128,128,128,128,128,128,19,27,128,128,128,128,128,128,36,44,128,128,128,128,128,128,37,45,128,128,128,128,128,128,54,62,128,128,128,128,128,128,55,63,128,128,128,128,128,128] ; AVX512BW-NEXT: vpshufb %zmm5, %zmm4, %zmm4 ; AVX512BW-NEXT: vinserti128 $1, (%rcx), %ymm1, %ymm1 ; AVX512BW-NEXT: vinserti64x4 $1, %ymm1, %zmm1, %zmm1 ; AVX512BW-NEXT: vpermq {{.*#+}} zmm6 = zmm1[0,2,0,2,4,6,4,6] ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm7 = [128,128,0,8,128,128,128,128,128,128,1,9,128,128,128,128,128,128,18,26,128,128,128,128,128,128,19,27,128,128,128,128,128,128,36,44,128,128,128,128,128,128,37,45,128,128,128,128,128,128,54,62,128,128,128,128,128,128,55,63,128,128,128,128] ; AVX512BW-NEXT: vpshufb %zmm7, %zmm6, %zmm6 ; AVX512BW-NEXT: vporq %zmm4, %zmm6, %zmm4 ; AVX512BW-NEXT: vinserti128 $1, (%r9), %ymm2, %ymm2 ; AVX512BW-NEXT: vinserti64x4 $1, %ymm2, %zmm2, %zmm2 ; AVX512BW-NEXT: vpermq {{.*#+}} zmm6 = zmm2[0,2,0,2,4,6,4,6] ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm8 = [128,128,128,128,0,8,128,128,128,128,128,128,1,9,128,128,128,128,128,128,18,26,128,128,128,128,128,128,19,27,128,128,128,128,128,128,36,44,128,128,128,128,128,128,37,45,128,128,128,128,128,128,54,62,128,128,128,128,128,128,55,63,128,128] ; AVX512BW-NEXT: vpshufb %zmm8, %zmm6, %zmm6 ; AVX512BW-NEXT: vinserti128 $1, (%r10), %ymm3, %ymm3 ; AVX512BW-NEXT: vinserti64x4 $1, %ymm3, %zmm3, %zmm3 ; AVX512BW-NEXT: vpermq {{.*#+}} zmm9 = zmm3[0,2,0,2,4,6,4,6] ; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm10 = [128,128,128,128,128,128,0,8,128,128,128,128,128,128,1,9,128,128,128,128,128,128,18,26,128,128,128,128,128,128,19,27,128,128,128,128,128,128,36,44,128,128,128,128,128,128,37,45,128,128,128,128,128,128,54,62,128,128,128,128,128,128,55,63] ; AVX512BW-NEXT: vpshufb %zmm10, %zmm9, %zmm9 ; AVX512BW-NEXT: vpternlogq {{.*#+}} zmm9 = zmm9 | zmm4 | zmm6 ; AVX512BW-NEXT: vpermq {{.*#+}} zmm0 = zmm0[1,3,1,3,5,7,5,7] ; AVX512BW-NEXT: vpshufb %zmm5, %zmm0, %zmm0 ; AVX512BW-NEXT: vpermq {{.*#+}} zmm1 = zmm1[1,3,1,3,5,7,5,7] ; AVX512BW-NEXT: vpshufb %zmm7, %zmm1, %zmm1 ; AVX512BW-NEXT: vporq %zmm0, %zmm1, %zmm0 ; AVX512BW-NEXT: vpermq {{.*#+}} zmm1 = zmm2[1,3,1,3,5,7,5,7] ; AVX512BW-NEXT: vpshufb %zmm8, %zmm1, %zmm1 ; AVX512BW-NEXT: vpermq {{.*#+}} zmm2 = zmm3[1,3,1,3,5,7,5,7] ; AVX512BW-NEXT: vpshufb %zmm10, %zmm2, %zmm2 ; AVX512BW-NEXT: vpternlogq {{.*#+}} zmm2 = zmm2 | zmm0 | zmm1 ; AVX512BW-NEXT: vmovdqa64 %zmm2, 64(%rax) ; AVX512BW-NEXT: vmovdqa64 %zmm9, (%rax) ; AVX512BW-NEXT: vzeroupper ; AVX512BW-NEXT: retq ; ; AVX512BW-FCP-LABEL: store_i8_stride8_vf16: ; AVX512BW-FCP: # %bb.0: ; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r11 ; AVX512BW-FCP-NEXT: vmovdqa (%rdi), %xmm0 ; AVX512BW-FCP-NEXT: vmovdqa (%rdx), %xmm1 ; AVX512BW-FCP-NEXT: vmovdqa (%r8), %xmm2 ; AVX512BW-FCP-NEXT: vmovdqa (%r11), %xmm3 ; AVX512BW-FCP-NEXT: vinserti128 $1, (%rsi), %ymm0, %ymm0 ; AVX512BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm4 = [0,2,0,2,0,2,0,2] ; AVX512BW-FCP-NEXT: # zmm4 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3] ; AVX512BW-FCP-NEXT: vpermq %zmm0, %zmm4, %zmm5 ; AVX512BW-FCP-NEXT: vmovdqa64 {{.*#+}} zmm6 = [0,8,128,128,128,128,128,128,1,9,128,128,128,128,128,128,18,26,128,128,128,128,128,128,19,27,128,128,128,128,128,128,36,44,128,128,128,128,128,128,37,45,128,128,128,128,128,128,54,62,128,128,128,128,128,128,55,63,128,128,128,128,128,128] ; AVX512BW-FCP-NEXT: vpshufb %zmm6, %zmm5, %zmm5 ; AVX512BW-FCP-NEXT: vinserti128 $1, (%rcx), %ymm1, %ymm1 ; AVX512BW-FCP-NEXT: vpermq %zmm1, %zmm4, %zmm7 ; AVX512BW-FCP-NEXT: vmovdqa64 {{.*#+}} zmm8 = [128,128,0,8,128,128,128,128,128,128,1,9,128,128,128,128,128,128,18,26,128,128,128,128,128,128,19,27,128,128,128,128,128,128,36,44,128,128,128,128,128,128,37,45,128,128,128,128,128,128,54,62,128,128,128,128,128,128,55,63,128,128,128,128] ; AVX512BW-FCP-NEXT: vpshufb %zmm8, %zmm7, %zmm7 ; AVX512BW-FCP-NEXT: vporq %zmm5, %zmm7, %zmm5 ; AVX512BW-FCP-NEXT: vinserti128 $1, (%r9), %ymm2, %ymm2 ; AVX512BW-FCP-NEXT: vpermq %zmm2, %zmm4, %zmm7 ; AVX512BW-FCP-NEXT: vmovdqa64 {{.*#+}} zmm9 = [128,128,128,128,0,8,128,128,128,128,128,128,1,9,128,128,128,128,128,128,18,26,128,128,128,128,128,128,19,27,128,128,128,128,128,128,36,44,128,128,128,128,128,128,37,45,128,128,128,128,128,128,54,62,128,128,128,128,128,128,55,63,128,128] ; AVX512BW-FCP-NEXT: vpshufb %zmm9, %zmm7, %zmm7 ; AVX512BW-FCP-NEXT: vinserti128 $1, (%r10), %ymm3, %ymm3 ; AVX512BW-FCP-NEXT: vpermq %zmm3, %zmm4, %zmm4 ; AVX512BW-FCP-NEXT: vmovdqa64 {{.*#+}} zmm10 = [128,128,128,128,128,128,0,8,128,128,128,128,128,128,1,9,128,128,128,128,128,128,18,26,128,128,128,128,128,128,19,27,128,128,128,128,128,128,36,44,128,128,128,128,128,128,37,45,128,128,128,128,128,128,54,62,128,128,128,128,128,128,55,63] ; AVX512BW-FCP-NEXT: vpshufb %zmm10, %zmm4, %zmm4 ; AVX512BW-FCP-NEXT: vpternlogq {{.*#+}} zmm4 = zmm4 | zmm5 | zmm7 ; AVX512BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm5 = [1,3,1,3,1,3,1,3] ; AVX512BW-FCP-NEXT: # zmm5 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3] ; AVX512BW-FCP-NEXT: vpermq %zmm0, %zmm5, %zmm0 ; AVX512BW-FCP-NEXT: vpshufb %zmm6, %zmm0, %zmm0 ; AVX512BW-FCP-NEXT: vpermq %zmm1, %zmm5, %zmm1 ; AVX512BW-FCP-NEXT: vpshufb %zmm8, %zmm1, %zmm1 ; AVX512BW-FCP-NEXT: vporq %zmm0, %zmm1, %zmm0 ; AVX512BW-FCP-NEXT: vpermq %zmm2, %zmm5, %zmm1 ; AVX512BW-FCP-NEXT: vpshufb %zmm9, %zmm1, %zmm1 ; AVX512BW-FCP-NEXT: vpermq %zmm3, %zmm5, %zmm2 ; AVX512BW-FCP-NEXT: vpshufb %zmm10, %zmm2, %zmm2 ; AVX512BW-FCP-NEXT: vpternlogq {{.*#+}} zmm2 = zmm2 | zmm0 | zmm1 ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, 64(%rax) ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm4, (%rax) ; AVX512BW-FCP-NEXT: vzeroupper ; AVX512BW-FCP-NEXT: retq ; ; AVX512DQ-BW-LABEL: store_i8_stride8_vf16: ; AVX512DQ-BW: # %bb.0: ; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %r11 ; AVX512DQ-BW-NEXT: vmovdqa (%rdi), %xmm0 ; AVX512DQ-BW-NEXT: vmovdqa (%rdx), %xmm1 ; AVX512DQ-BW-NEXT: vmovdqa (%r8), %xmm2 ; AVX512DQ-BW-NEXT: vmovdqa (%r11), %xmm3 ; AVX512DQ-BW-NEXT: vinserti128 $1, (%rsi), %ymm0, %ymm0 ; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm0 ; AVX512DQ-BW-NEXT: vpermq {{.*#+}} zmm4 = zmm0[0,2,0,2,4,6,4,6] ; AVX512DQ-BW-NEXT: vmovdqa64 {{.*#+}} zmm5 = [0,8,128,128,128,128,128,128,1,9,128,128,128,128,128,128,18,26,128,128,128,128,128,128,19,27,128,128,128,128,128,128,36,44,128,128,128,128,128,128,37,45,128,128,128,128,128,128,54,62,128,128,128,128,128,128,55,63,128,128,128,128,128,128] ; AVX512DQ-BW-NEXT: vpshufb %zmm5, %zmm4, %zmm4 ; AVX512DQ-BW-NEXT: vinserti128 $1, (%rcx), %ymm1, %ymm1 ; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm1, %zmm1, %zmm1 ; AVX512DQ-BW-NEXT: vpermq {{.*#+}} zmm6 = zmm1[0,2,0,2,4,6,4,6] ; AVX512DQ-BW-NEXT: vmovdqa64 {{.*#+}} zmm7 = [128,128,0,8,128,128,128,128,128,128,1,9,128,128,128,128,128,128,18,26,128,128,128,128,128,128,19,27,128,128,128,128,128,128,36,44,128,128,128,128,128,128,37,45,128,128,128,128,128,128,54,62,128,128,128,128,128,128,55,63,128,128,128,128] ; AVX512DQ-BW-NEXT: vpshufb %zmm7, %zmm6, %zmm6 ; AVX512DQ-BW-NEXT: vporq %zmm4, %zmm6, %zmm4 ; AVX512DQ-BW-NEXT: vinserti128 $1, (%r9), %ymm2, %ymm2 ; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm2, %zmm2, %zmm2 ; AVX512DQ-BW-NEXT: vpermq {{.*#+}} zmm6 = zmm2[0,2,0,2,4,6,4,6] ; AVX512DQ-BW-NEXT: vmovdqa64 {{.*#+}} zmm8 = [128,128,128,128,0,8,128,128,128,128,128,128,1,9,128,128,128,128,128,128,18,26,128,128,128,128,128,128,19,27,128,128,128,128,128,128,36,44,128,128,128,128,128,128,37,45,128,128,128,128,128,128,54,62,128,128,128,128,128,128,55,63,128,128] ; AVX512DQ-BW-NEXT: vpshufb %zmm8, %zmm6, %zmm6 ; AVX512DQ-BW-NEXT: vinserti128 $1, (%r10), %ymm3, %ymm3 ; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm3, %zmm3, %zmm3 ; AVX512DQ-BW-NEXT: vpermq {{.*#+}} zmm9 = zmm3[0,2,0,2,4,6,4,6] ; AVX512DQ-BW-NEXT: vmovdqa64 {{.*#+}} zmm10 = [128,128,128,128,128,128,0,8,128,128,128,128,128,128,1,9,128,128,128,128,128,128,18,26,128,128,128,128,128,128,19,27,128,128,128,128,128,128,36,44,128,128,128,128,128,128,37,45,128,128,128,128,128,128,54,62,128,128,128,128,128,128,55,63] ; AVX512DQ-BW-NEXT: vpshufb %zmm10, %zmm9, %zmm9 ; AVX512DQ-BW-NEXT: vpternlogq {{.*#+}} zmm9 = zmm9 | zmm4 | zmm6 ; AVX512DQ-BW-NEXT: vpermq {{.*#+}} zmm0 = zmm0[1,3,1,3,5,7,5,7] ; AVX512DQ-BW-NEXT: vpshufb %zmm5, %zmm0, %zmm0 ; AVX512DQ-BW-NEXT: vpermq {{.*#+}} zmm1 = zmm1[1,3,1,3,5,7,5,7] ; AVX512DQ-BW-NEXT: vpshufb %zmm7, %zmm1, %zmm1 ; AVX512DQ-BW-NEXT: vporq %zmm0, %zmm1, %zmm0 ; AVX512DQ-BW-NEXT: vpermq {{.*#+}} zmm1 = zmm2[1,3,1,3,5,7,5,7] ; AVX512DQ-BW-NEXT: vpshufb %zmm8, %zmm1, %zmm1 ; AVX512DQ-BW-NEXT: vpermq {{.*#+}} zmm2 = zmm3[1,3,1,3,5,7,5,7] ; AVX512DQ-BW-NEXT: vpshufb %zmm10, %zmm2, %zmm2 ; AVX512DQ-BW-NEXT: vpternlogq {{.*#+}} zmm2 = zmm2 | zmm0 | zmm1 ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, 64(%rax) ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm9, (%rax) ; AVX512DQ-BW-NEXT: vzeroupper ; AVX512DQ-BW-NEXT: retq ; ; AVX512DQ-BW-FCP-LABEL: store_i8_stride8_vf16: ; AVX512DQ-BW-FCP: # %bb.0: ; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r11 ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdi), %xmm0 ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdx), %xmm1 ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%r8), %xmm2 ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%r11), %xmm3 ; AVX512DQ-BW-FCP-NEXT: vinserti128 $1, (%rsi), %ymm0, %ymm0 ; AVX512DQ-BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm4 = [0,2,0,2,0,2,0,2] ; AVX512DQ-BW-FCP-NEXT: # zmm4 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3] ; AVX512DQ-BW-FCP-NEXT: vpermq %zmm0, %zmm4, %zmm5 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 {{.*#+}} zmm6 = [0,8,128,128,128,128,128,128,1,9,128,128,128,128,128,128,18,26,128,128,128,128,128,128,19,27,128,128,128,128,128,128,36,44,128,128,128,128,128,128,37,45,128,128,128,128,128,128,54,62,128,128,128,128,128,128,55,63,128,128,128,128,128,128] ; AVX512DQ-BW-FCP-NEXT: vpshufb %zmm6, %zmm5, %zmm5 ; AVX512DQ-BW-FCP-NEXT: vinserti128 $1, (%rcx), %ymm1, %ymm1 ; AVX512DQ-BW-FCP-NEXT: vpermq %zmm1, %zmm4, %zmm7 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 {{.*#+}} zmm8 = [128,128,0,8,128,128,128,128,128,128,1,9,128,128,128,128,128,128,18,26,128,128,128,128,128,128,19,27,128,128,128,128,128,128,36,44,128,128,128,128,128,128,37,45,128,128,128,128,128,128,54,62,128,128,128,128,128,128,55,63,128,128,128,128] ; AVX512DQ-BW-FCP-NEXT: vpshufb %zmm8, %zmm7, %zmm7 ; AVX512DQ-BW-FCP-NEXT: vporq %zmm5, %zmm7, %zmm5 ; AVX512DQ-BW-FCP-NEXT: vinserti128 $1, (%r9), %ymm2, %ymm2 ; AVX512DQ-BW-FCP-NEXT: vpermq %zmm2, %zmm4, %zmm7 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 {{.*#+}} zmm9 = [128,128,128,128,0,8,128,128,128,128,128,128,1,9,128,128,128,128,128,128,18,26,128,128,128,128,128,128,19,27,128,128,128,128,128,128,36,44,128,128,128,128,128,128,37,45,128,128,128,128,128,128,54,62,128,128,128,128,128,128,55,63,128,128] ; AVX512DQ-BW-FCP-NEXT: vpshufb %zmm9, %zmm7, %zmm7 ; AVX512DQ-BW-FCP-NEXT: vinserti128 $1, (%r10), %ymm3, %ymm3 ; AVX512DQ-BW-FCP-NEXT: vpermq %zmm3, %zmm4, %zmm4 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 {{.*#+}} zmm10 = [128,128,128,128,128,128,0,8,128,128,128,128,128,128,1,9,128,128,128,128,128,128,18,26,128,128,128,128,128,128,19,27,128,128,128,128,128,128,36,44,128,128,128,128,128,128,37,45,128,128,128,128,128,128,54,62,128,128,128,128,128,128,55,63] ; AVX512DQ-BW-FCP-NEXT: vpshufb %zmm10, %zmm4, %zmm4 ; AVX512DQ-BW-FCP-NEXT: vpternlogq {{.*#+}} zmm4 = zmm4 | zmm5 | zmm7 ; AVX512DQ-BW-FCP-NEXT: vbroadcasti32x4 {{.*#+}} zmm5 = [1,3,1,3,1,3,1,3] ; AVX512DQ-BW-FCP-NEXT: # zmm5 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3] ; AVX512DQ-BW-FCP-NEXT: vpermq %zmm0, %zmm5, %zmm0 ; AVX512DQ-BW-FCP-NEXT: vpshufb %zmm6, %zmm0, %zmm0 ; AVX512DQ-BW-FCP-NEXT: vpermq %zmm1, %zmm5, %zmm1 ; AVX512DQ-BW-FCP-NEXT: vpshufb %zmm8, %zmm1, %zmm1 ; AVX512DQ-BW-FCP-NEXT: vporq %zmm0, %zmm1, %zmm0 ; AVX512DQ-BW-FCP-NEXT: vpermq %zmm2, %zmm5, %zmm1 ; AVX512DQ-BW-FCP-NEXT: vpshufb %zmm9, %zmm1, %zmm1 ; AVX512DQ-BW-FCP-NEXT: vpermq %zmm3, %zmm5, %zmm2 ; AVX512DQ-BW-FCP-NEXT: vpshufb %zmm10, %zmm2, %zmm2 ; AVX512DQ-BW-FCP-NEXT: vpternlogq {{.*#+}} zmm2 = zmm2 | zmm0 | zmm1 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, 64(%rax) ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm4, (%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 %in.vec6 = load <16 x i8>, ptr %in.vecptr6, align 64 %in.vec7 = load <16 x i8>, ptr %in.vecptr7, 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 <16 x i8> %in.vec6, <16 x i8> %in.vec7, <32 x i32> %5 = shufflevector <32 x i8> %1, <32 x i8> %2, <64 x i32> %6 = shufflevector <32 x i8> %3, <32 x i8> %4, <64 x i32> %7 = shufflevector <64 x i8> %5, <64 x i8> %6, <128 x i32> %interleaved.vec = shufflevector <128 x i8> %7, <128 x i8> poison, <128 x i32> store <128 x i8> %interleaved.vec, ptr %out.vec, align 64 ret void } define void @store_i8_stride8_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecptr2, ptr %in.vecptr3, ptr %in.vecptr4, ptr %in.vecptr5, ptr %in.vecptr6, ptr %in.vecptr7, ptr %out.vec) nounwind { ; SSE-LABEL: store_i8_stride8_vf32: ; SSE: # %bb.0: ; SSE-NEXT: subq $232, %rsp ; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax ; SSE-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; SSE-NEXT: movdqa (%rdi), %xmm5 ; SSE-NEXT: movdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: movdqa (%rsi), %xmm4 ; SSE-NEXT: movdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: movdqa (%rdx), %xmm1 ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: movdqa (%rcx), %xmm8 ; SSE-NEXT: movdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: movdqa (%r8), %xmm13 ; SSE-NEXT: movdqa (%r9), %xmm12 ; SSE-NEXT: movdqa (%r10), %xmm14 ; SSE-NEXT: movdqa (%rax), %xmm11 ; SSE-NEXT: movdqa %xmm14, %xmm2 ; SSE-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm11[0],xmm2[1],xmm11[1],xmm2[2],xmm11[2],xmm2[3],xmm11[3],xmm2[4],xmm11[4],xmm2[5],xmm11[5],xmm2[6],xmm11[6],xmm2[7],xmm11[7] ; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm2[0,2,2,3,4,5,6,7] ; SSE-NEXT: movdqa %xmm2, %xmm15 ; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm0[0,0,2,1] ; SSE-NEXT: movdqa {{.*#+}} xmm9 = [65535,65535,65535,0,65535,65535,65535,0] ; SSE-NEXT: movdqa %xmm9, %xmm6 ; SSE-NEXT: pandn %xmm2, %xmm6 ; SSE-NEXT: movdqa %xmm13, %xmm3 ; SSE-NEXT: punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm12[0],xmm3[1],xmm12[1],xmm3[2],xmm12[2],xmm3[3],xmm12[3],xmm3[4],xmm12[4],xmm3[5],xmm12[5],xmm3[6],xmm12[6],xmm3[7],xmm12[7] ; SSE-NEXT: pshuflw {{.*#+}} xmm2 = xmm3[2,1,3,3,4,5,6,7] ; SSE-NEXT: movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm2[0,0,2,1] ; SSE-NEXT: pand %xmm9, %xmm7 ; SSE-NEXT: por %xmm6, %xmm7 ; SSE-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm8[0],xmm1[1],xmm8[1],xmm1[2],xmm8[2],xmm1[3],xmm8[3],xmm1[4],xmm8[4],xmm1[5],xmm8[5],xmm1[6],xmm8[6],xmm1[7],xmm8[7] ; SSE-NEXT: pshuflw {{.*#+}} xmm2 = xmm1[0,2,2,3,4,5,6,7] ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm2[0,1,1,3] ; SSE-NEXT: movdqa {{.*#+}} xmm8 = [65535,0,65535,65535,65535,0,65535,65535] ; SSE-NEXT: movdqa %xmm8, %xmm10 ; SSE-NEXT: pandn %xmm6, %xmm10 ; SSE-NEXT: punpcklbw {{.*#+}} xmm5 = 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] ; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm5[1,1,1,1] ; SSE-NEXT: movdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: pshufhw {{.*#+}} xmm6 = xmm6[0,1,2,3,5,5,5,5] ; SSE-NEXT: pand %xmm8, %xmm6 ; SSE-NEXT: por %xmm10, %xmm6 ; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm7[1,3,2,3] ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm6[0,2,2,3] ; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm7[0],xmm0[1],xmm7[1] ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: pshuflw {{.*#+}} xmm6 = xmm15[0,0,2,1,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm6[0,0,2,1] ; SSE-NEXT: movdqa %xmm9, %xmm7 ; SSE-NEXT: pandn %xmm6, %xmm7 ; SSE-NEXT: pshuflw {{.*#+}} xmm6 = xmm3[0,1,1,3,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm6[0,0,2,1] ; SSE-NEXT: pand %xmm9, %xmm6 ; SSE-NEXT: por %xmm7, %xmm6 ; SSE-NEXT: pshuflw {{.*#+}} xmm7 = xmm1[0,0,2,1,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm7[0,1,1,3] ; SSE-NEXT: movdqa %xmm8, %xmm10 ; SSE-NEXT: pandn %xmm7, %xmm10 ; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm5[0,0,0,0] ; SSE-NEXT: pshufhw {{.*#+}} xmm7 = xmm7[0,1,2,3,5,5,5,5] ; SSE-NEXT: pand %xmm8, %xmm7 ; SSE-NEXT: por %xmm10, %xmm7 ; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm6[1,3,2,3] ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm7[0,2,2,3] ; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm6[0],xmm0[1],xmm6[1] ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: punpckhbw {{.*#+}} xmm14 = 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] ; SSE-NEXT: pshuflw {{.*#+}} xmm5 = xmm14[0,2,2,3,4,5,6,7] ; SSE-NEXT: movdqa %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm5[0,0,2,1] ; SSE-NEXT: movdqa %xmm9, %xmm6 ; SSE-NEXT: pandn %xmm5, %xmm6 ; SSE-NEXT: punpckhbw {{.*#+}} xmm13 = 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] ; SSE-NEXT: pshuflw {{.*#+}} xmm4 = xmm13[2,1,3,3,4,5,6,7] ; SSE-NEXT: movdqa %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: pshufd {{.*#+}} xmm4 = xmm4[0,0,2,1] ; SSE-NEXT: pand %xmm9, %xmm4 ; SSE-NEXT: por %xmm6, %xmm4 ; SSE-NEXT: pshufd {{.*#+}} xmm4 = xmm4[1,3,2,3] ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload ; SSE-NEXT: punpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Folded Reload ; SSE-NEXT: # xmm12 = xmm12[8],mem[8],xmm12[9],mem[9],xmm12[10],mem[10],xmm12[11],mem[11],xmm12[12],mem[12],xmm12[13],mem[13],xmm12[14],mem[14],xmm12[15],mem[15] ; SSE-NEXT: pshuflw {{.*#+}} xmm3 = xmm12[0,2,2,3,4,5,6,7] ; SSE-NEXT: movdqa %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm3[0,1,1,3] ; SSE-NEXT: movdqa %xmm8, %xmm5 ; SSE-NEXT: pandn %xmm3, %xmm5 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload ; SSE-NEXT: punpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Folded Reload ; SSE-NEXT: # xmm11 = xmm11[8],mem[8],xmm11[9],mem[9],xmm11[10],mem[10],xmm11[11],mem[11],xmm11[12],mem[12],xmm11[13],mem[13],xmm11[14],mem[14],xmm11[15],mem[15] ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm11[1,1,1,1] ; SSE-NEXT: movdqa %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,5,5,5,5] ; SSE-NEXT: pand %xmm8, %xmm1 ; SSE-NEXT: por %xmm5, %xmm1 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm1[0,2,2,3] ; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm4[0],xmm0[1],xmm4[1] ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: pshuflw {{.*#+}} xmm1 = xmm14[0,0,2,1,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,0,2,1] ; SSE-NEXT: movdqa %xmm9, %xmm3 ; SSE-NEXT: pandn %xmm1, %xmm3 ; SSE-NEXT: pshuflw {{.*#+}} xmm1 = xmm13[0,1,1,3,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,0,2,1] ; SSE-NEXT: pand %xmm9, %xmm1 ; SSE-NEXT: por %xmm3, %xmm1 ; SSE-NEXT: pshuflw {{.*#+}} xmm3 = xmm12[0,0,2,1,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm3[0,1,1,3] ; SSE-NEXT: movdqa %xmm8, %xmm4 ; SSE-NEXT: pandn %xmm3, %xmm4 ; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm11[0,0,0,0] ; SSE-NEXT: pshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,5,5,5,5] ; SSE-NEXT: pand %xmm8, %xmm3 ; SSE-NEXT: por %xmm4, %xmm3 ; SSE-NEXT: movdqa 16(%r10), %xmm10 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,3,2,3] ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm3[0,2,2,3] ; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: movdqa 16(%rax), %xmm1 ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: movdqa %xmm10, %xmm5 ; SSE-NEXT: punpcklbw {{.*#+}} xmm5 = 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] ; SSE-NEXT: pshuflw {{.*#+}} xmm1 = xmm5[0,2,2,3,4,5,6,7] ; SSE-NEXT: movdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,0,2,1] ; SSE-NEXT: movdqa %xmm9, %xmm3 ; SSE-NEXT: pandn %xmm1, %xmm3 ; SSE-NEXT: movdqa 16(%r8), %xmm12 ; SSE-NEXT: movdqa 16(%r9), %xmm11 ; SSE-NEXT: movdqa %xmm12, %xmm4 ; SSE-NEXT: punpcklbw {{.*#+}} xmm4 = xmm4[0],xmm11[0],xmm4[1],xmm11[1],xmm4[2],xmm11[2],xmm4[3],xmm11[3],xmm4[4],xmm11[4],xmm4[5],xmm11[5],xmm4[6],xmm11[6],xmm4[7],xmm11[7] ; SSE-NEXT: pshuflw {{.*#+}} xmm1 = xmm4[2,1,3,3,4,5,6,7] ; SSE-NEXT: movdqa %xmm4, (%rsp) # 16-byte Spill ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm1[0,0,2,1] ; SSE-NEXT: pand %xmm9, %xmm0 ; SSE-NEXT: por %xmm3, %xmm0 ; SSE-NEXT: movdqa 16(%rdx), %xmm13 ; SSE-NEXT: movdqa 16(%rcx), %xmm7 ; SSE-NEXT: movdqa %xmm13, %xmm3 ; SSE-NEXT: punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm7[0],xmm3[1],xmm7[1],xmm3[2],xmm7[2],xmm3[3],xmm7[3],xmm3[4],xmm7[4],xmm3[5],xmm7[5],xmm3[6],xmm7[6],xmm3[7],xmm7[7] ; SSE-NEXT: pshuflw {{.*#+}} xmm1 = xmm3[0,2,2,3,4,5,6,7] ; SSE-NEXT: movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,1,1,3] ; SSE-NEXT: movdqa %xmm8, %xmm2 ; SSE-NEXT: pandn %xmm1, %xmm2 ; SSE-NEXT: movdqa 16(%rdi), %xmm14 ; SSE-NEXT: movdqa 16(%rsi), %xmm6 ; SSE-NEXT: movdqa %xmm14, %xmm1 ; SSE-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm6[0],xmm1[1],xmm6[1],xmm1[2],xmm6[2],xmm1[3],xmm6[3],xmm1[4],xmm6[4],xmm1[5],xmm6[5],xmm1[6],xmm6[6],xmm1[7],xmm6[7] ; SSE-NEXT: pshufd {{.*#+}} xmm15 = xmm1[1,1,1,1] ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: pshufhw {{.*#+}} xmm15 = xmm15[0,1,2,3,5,5,5,5] ; SSE-NEXT: pand %xmm8, %xmm15 ; SSE-NEXT: por %xmm2, %xmm15 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,3,2,3] ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm15[0,2,2,3] ; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1] ; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm5[0,0,2,1,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,0,2,1] ; SSE-NEXT: movdqa %xmm9, %xmm2 ; SSE-NEXT: pandn %xmm0, %xmm2 ; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm4[0,1,1,3,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,0,2,1] ; SSE-NEXT: pand %xmm9, %xmm0 ; SSE-NEXT: por %xmm2, %xmm0 ; SSE-NEXT: pshuflw {{.*#+}} xmm2 = xmm3[0,0,2,1,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm2[0,1,1,3] ; SSE-NEXT: movdqa %xmm8, %xmm15 ; SSE-NEXT: pandn %xmm2, %xmm15 ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm1[0,0,0,0] ; SSE-NEXT: pshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,5,5,5,5] ; SSE-NEXT: pand %xmm8, %xmm2 ; SSE-NEXT: por %xmm15, %xmm2 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,3,2,3] ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm2[0,2,2,3] ; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1] ; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: punpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Folded Reload ; SSE-NEXT: # xmm10 = xmm10[8],mem[8],xmm10[9],mem[9],xmm10[10],mem[10],xmm10[11],mem[11],xmm10[12],mem[12],xmm10[13],mem[13],xmm10[14],mem[14],xmm10[15],mem[15] ; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm10[0,2,2,3,4,5,6,7] ; SSE-NEXT: movdqa %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,0,2,1] ; SSE-NEXT: movdqa %xmm9, %xmm2 ; SSE-NEXT: pandn %xmm0, %xmm2 ; SSE-NEXT: movdqa %xmm12, %xmm15 ; SSE-NEXT: punpckhbw {{.*#+}} xmm15 = xmm15[8],xmm11[8],xmm15[9],xmm11[9],xmm15[10],xmm11[10],xmm15[11],xmm11[11],xmm15[12],xmm11[12],xmm15[13],xmm11[13],xmm15[14],xmm11[14],xmm15[15],xmm11[15] ; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm15[2,1,3,3,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,0,2,1] ; SSE-NEXT: pand %xmm9, %xmm0 ; SSE-NEXT: por %xmm2, %xmm0 ; SSE-NEXT: punpckhbw {{.*#+}} xmm13 = xmm13[8],xmm7[8],xmm13[9],xmm7[9],xmm13[10],xmm7[10],xmm13[11],xmm7[11],xmm13[12],xmm7[12],xmm13[13],xmm7[13],xmm13[14],xmm7[14],xmm13[15],xmm7[15] ; SSE-NEXT: pshuflw {{.*#+}} xmm2 = xmm13[0,2,2,3,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm2[0,1,1,3] ; SSE-NEXT: movdqa %xmm8, %xmm3 ; SSE-NEXT: pandn %xmm2, %xmm3 ; SSE-NEXT: punpckhbw {{.*#+}} xmm14 = xmm14[8],xmm6[8],xmm14[9],xmm6[9],xmm14[10],xmm6[10],xmm14[11],xmm6[11],xmm14[12],xmm6[12],xmm14[13],xmm6[13],xmm14[14],xmm6[14],xmm14[15],xmm6[15] ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm14[1,1,1,1] ; SSE-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,5,5,5,5] ; SSE-NEXT: pand %xmm8, %xmm1 ; SSE-NEXT: por %xmm3, %xmm1 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,3,2,3] ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,2,2,3] ; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1] ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm10[0,0,2,1,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,0,2,1] ; SSE-NEXT: movdqa %xmm9, %xmm1 ; SSE-NEXT: pandn %xmm0, %xmm1 ; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm15[0,1,1,3,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,0,2,1] ; SSE-NEXT: pand %xmm9, %xmm0 ; SSE-NEXT: por %xmm1, %xmm0 ; SSE-NEXT: pshuflw {{.*#+}} xmm1 = xmm13[0,0,2,1,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,1,1,3] ; SSE-NEXT: movdqa %xmm8, %xmm2 ; SSE-NEXT: pandn %xmm1, %xmm2 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm14[0,0,0,0] ; SSE-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,5,5,5,5] ; SSE-NEXT: pand %xmm8, %xmm1 ; SSE-NEXT: por %xmm2, %xmm1 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,3,2,3] ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,2,2,3] ; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1] ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload ; SSE-NEXT: pshufhw {{.*#+}} xmm0 = xmm3[0,1,2,3,4,6,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3] ; SSE-NEXT: movdqa %xmm9, %xmm1 ; SSE-NEXT: pandn %xmm0, %xmm1 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload ; SSE-NEXT: pshufhw {{.*#+}} xmm0 = xmm4[0,1,2,3,6,5,7,7] ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3] ; SSE-NEXT: pand %xmm9, %xmm0 ; SSE-NEXT: por %xmm1, %xmm0 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload ; SSE-NEXT: pshufhw {{.*#+}} xmm1 = xmm5[0,1,2,3,4,6,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[2,1,3,3] ; SSE-NEXT: movdqa %xmm8, %xmm2 ; SSE-NEXT: pandn %xmm1, %xmm2 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm10[3,3,3,3] ; SSE-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,5,5,5,5] ; SSE-NEXT: pand %xmm8, %xmm1 ; SSE-NEXT: por %xmm2, %xmm1 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,3,2,3] ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,2,2,3] ; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1] ; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: pshufhw {{.*#+}} xmm0 = xmm3[0,1,2,3,4,4,6,5] ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3] ; SSE-NEXT: movdqa %xmm9, %xmm1 ; SSE-NEXT: pandn %xmm0, %xmm1 ; SSE-NEXT: pshufhw {{.*#+}} xmm0 = xmm4[0,1,2,3,4,5,5,7] ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3] ; SSE-NEXT: pand %xmm9, %xmm0 ; SSE-NEXT: por %xmm1, %xmm0 ; SSE-NEXT: pshufhw {{.*#+}} xmm1 = xmm5[0,1,2,3,4,4,6,5] ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[2,1,3,3] ; SSE-NEXT: movdqa %xmm8, %xmm2 ; SSE-NEXT: pandn %xmm1, %xmm2 ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm10[2,2,2,2] ; SSE-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,5,5,5,5] ; SSE-NEXT: pand %xmm8, %xmm1 ; SSE-NEXT: por %xmm2, %xmm1 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,3,2,3] ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,2,2,3] ; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1] ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload ; SSE-NEXT: pshufhw {{.*#+}} xmm0 = xmm4[0,1,2,3,4,6,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3] ; SSE-NEXT: movdqa %xmm9, %xmm2 ; SSE-NEXT: pandn %xmm0, %xmm2 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload ; SSE-NEXT: pshufhw {{.*#+}} xmm0 = xmm5[0,1,2,3,6,5,7,7] ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3] ; SSE-NEXT: pand %xmm9, %xmm0 ; SSE-NEXT: por %xmm2, %xmm0 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload ; SSE-NEXT: pshufhw {{.*#+}} xmm2 = xmm10[0,1,2,3,4,6,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm2[2,1,3,3] ; SSE-NEXT: movdqa %xmm8, %xmm3 ; SSE-NEXT: pandn %xmm2, %xmm3 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm12 # 16-byte Reload ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm12[3,3,3,3] ; SSE-NEXT: pshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,5,5,5,5] ; SSE-NEXT: pand %xmm8, %xmm2 ; SSE-NEXT: por %xmm3, %xmm2 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,3,2,3] ; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm2[0,2,2,3] ; SSE-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm0[0],xmm3[1],xmm0[1] ; SSE-NEXT: pshufhw {{.*#+}} xmm0 = xmm4[0,1,2,3,4,4,6,5] ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3] ; SSE-NEXT: movdqa %xmm9, %xmm2 ; SSE-NEXT: pandn %xmm0, %xmm2 ; SSE-NEXT: pshufhw {{.*#+}} xmm0 = xmm5[0,1,2,3,4,5,5,7] ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3] ; SSE-NEXT: pand %xmm9, %xmm0 ; SSE-NEXT: por %xmm2, %xmm0 ; SSE-NEXT: pshufhw {{.*#+}} xmm2 = xmm10[0,1,2,3,4,4,6,5] ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm2[2,1,3,3] ; SSE-NEXT: movdqa %xmm8, %xmm4 ; SSE-NEXT: pandn %xmm2, %xmm4 ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm12[2,2,2,2] ; SSE-NEXT: pshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,5,5,5,5] ; SSE-NEXT: pand %xmm8, %xmm2 ; SSE-NEXT: por %xmm4, %xmm2 ; SSE-NEXT: pshufd {{.*#+}} xmm4 = xmm0[1,3,2,3] ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm2[0,2,2,3] ; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm4[0],xmm0[1],xmm4[1] ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Reload ; SSE-NEXT: pshufhw {{.*#+}} xmm2 = xmm10[0,1,2,3,4,6,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm2[0,2,2,3] ; SSE-NEXT: movdqa %xmm9, %xmm4 ; SSE-NEXT: pandn %xmm2, %xmm4 ; SSE-NEXT: movdqa (%rsp), %xmm6 # 16-byte Reload ; SSE-NEXT: pshufhw {{.*#+}} xmm2 = xmm6[0,1,2,3,6,5,7,7] ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm2[0,2,2,3] ; SSE-NEXT: pand %xmm9, %xmm2 ; SSE-NEXT: por %xmm4, %xmm2 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload ; SSE-NEXT: pshufhw {{.*#+}} xmm4 = xmm7[0,1,2,3,4,6,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm4 = xmm4[2,1,3,3] ; SSE-NEXT: movdqa %xmm8, %xmm5 ; SSE-NEXT: pandn %xmm4, %xmm5 ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload ; SSE-NEXT: pshufd {{.*#+}} xmm4 = xmm11[3,3,3,3] ; SSE-NEXT: pshufhw {{.*#+}} xmm4 = xmm4[0,1,2,3,5,5,5,5] ; SSE-NEXT: pand %xmm8, %xmm4 ; SSE-NEXT: por %xmm5, %xmm4 ; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm2[1,3,2,3] ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm4[0,2,2,3] ; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm5[0],xmm2[1],xmm5[1] ; SSE-NEXT: pshufhw {{.*#+}} xmm4 = xmm10[0,1,2,3,4,4,6,5] ; SSE-NEXT: pshufd {{.*#+}} xmm4 = xmm4[0,2,2,3] ; SSE-NEXT: movdqa %xmm9, %xmm5 ; SSE-NEXT: pandn %xmm4, %xmm5 ; SSE-NEXT: pshufhw {{.*#+}} xmm4 = xmm6[0,1,2,3,4,5,5,7] ; SSE-NEXT: pshufd {{.*#+}} xmm4 = xmm4[0,2,2,3] ; SSE-NEXT: pand %xmm9, %xmm4 ; SSE-NEXT: por %xmm5, %xmm4 ; SSE-NEXT: pshufhw {{.*#+}} xmm5 = xmm7[0,1,2,3,4,4,6,5] ; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm5[2,1,3,3] ; SSE-NEXT: movdqa %xmm8, %xmm10 ; SSE-NEXT: pandn %xmm5, %xmm10 ; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm11[2,2,2,2] ; SSE-NEXT: pshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,5,5,5,5] ; SSE-NEXT: pand %xmm8, %xmm5 ; SSE-NEXT: por %xmm10, %xmm5 ; SSE-NEXT: pshufd {{.*#+}} xmm10 = xmm4[1,3,2,3] ; SSE-NEXT: pshufd {{.*#+}} xmm4 = xmm5[0,2,2,3] ; SSE-NEXT: punpckldq {{.*#+}} xmm4 = xmm4[0],xmm10[0],xmm4[1],xmm10[1] ; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload ; SSE-NEXT: pshufhw {{.*#+}} xmm5 = xmm6[0,1,2,3,4,6,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm5[0,2,2,3] ; SSE-NEXT: movdqa %xmm9, %xmm10 ; SSE-NEXT: pandn %xmm5, %xmm10 ; SSE-NEXT: pshufhw {{.*#+}} xmm5 = xmm15[0,1,2,3,6,5,7,7] ; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm5[0,2,2,3] ; SSE-NEXT: pand %xmm9, %xmm5 ; SSE-NEXT: por %xmm10, %xmm5 ; SSE-NEXT: pshufhw {{.*#+}} xmm10 = xmm13[0,1,2,3,4,6,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm10 = xmm10[2,1,3,3] ; SSE-NEXT: movdqa %xmm8, %xmm12 ; SSE-NEXT: pandn %xmm10, %xmm12 ; SSE-NEXT: pshufd {{.*#+}} xmm10 = xmm14[3,3,3,3] ; SSE-NEXT: pshufhw {{.*#+}} xmm10 = xmm10[0,1,2,3,5,5,5,5] ; SSE-NEXT: pand %xmm8, %xmm10 ; SSE-NEXT: por %xmm12, %xmm10 ; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm5[1,3,2,3] ; SSE-NEXT: pshufd {{.*#+}} xmm10 = xmm10[0,2,2,3] ; SSE-NEXT: punpckldq {{.*#+}} xmm10 = xmm10[0],xmm5[0],xmm10[1],xmm5[1] ; SSE-NEXT: pshufhw {{.*#+}} xmm5 = xmm6[0,1,2,3,4,4,6,5] ; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm5[0,2,2,3] ; SSE-NEXT: pshufhw {{.*#+}} xmm11 = xmm15[0,1,2,3,4,5,5,7] ; SSE-NEXT: pshufd {{.*#+}} xmm11 = xmm11[0,2,2,3] ; SSE-NEXT: pand %xmm9, %xmm11 ; SSE-NEXT: pandn %xmm5, %xmm9 ; SSE-NEXT: por %xmm11, %xmm9 ; SSE-NEXT: pshufhw {{.*#+}} xmm5 = xmm13[0,1,2,3,4,4,6,5] ; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm5[2,1,3,3] ; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm14[2,2,2,2] ; SSE-NEXT: pshufhw {{.*#+}} xmm6 = xmm6[0,1,2,3,5,5,5,5] ; SSE-NEXT: pand %xmm8, %xmm6 ; SSE-NEXT: pandn %xmm5, %xmm8 ; SSE-NEXT: por %xmm6, %xmm8 ; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm9[1,3,2,3] ; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm8[0,2,2,3] ; SSE-NEXT: punpckldq {{.*#+}} xmm6 = xmm6[0],xmm5[0],xmm6[1],xmm5[1] ; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax ; SSE-NEXT: movdqa %xmm6, 224(%rax) ; SSE-NEXT: movdqa %xmm10, 240(%rax) ; SSE-NEXT: movdqa %xmm4, 160(%rax) ; SSE-NEXT: movdqa %xmm2, 176(%rax) ; SSE-NEXT: movdqa %xmm0, 96(%rax) ; SSE-NEXT: movdqa %xmm3, 112(%rax) ; SSE-NEXT: movdqa %xmm1, 32(%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, 192(%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, 128(%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, 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, (%rax) ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload ; SSE-NEXT: movaps %xmm0, 16(%rax) ; SSE-NEXT: addq $232, %rsp ; SSE-NEXT: retq ; ; AVX-LABEL: store_i8_stride8_vf32: ; AVX: # %bb.0: ; AVX-NEXT: subq $56, %rsp ; AVX-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX-NEXT: vmovdqa (%r10), %xmm0 ; AVX-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX-NEXT: vmovdqa (%rax), %xmm2 ; AVX-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3],xmm2[4],xmm0[4],xmm2[5],xmm0[5],xmm2[6],xmm0[6],xmm2[7],xmm0[7] ; AVX-NEXT: vpshuflw {{.*#+}} xmm1 = xmm0[0,0,2,1,4,5,6,7] ; AVX-NEXT: vpshuflw {{.*#+}} xmm3 = xmm0[0,2,2,3,4,5,6,7] ; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm1, %ymm1 ; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm1[0,0,2,1,4,4,6,5] ; AVX-NEXT: vbroadcastsd {{.*#+}} ymm9 = [65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0] ; AVX-NEXT: vandnps %ymm1, %ymm9, %ymm1 ; AVX-NEXT: vmovdqa (%r9), %xmm5 ; AVX-NEXT: vmovdqa (%r8), %xmm7 ; AVX-NEXT: vpunpcklbw {{.*#+}} xmm12 = 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] ; AVX-NEXT: vpshuflw {{.*#+}} xmm3 = xmm12[0,1,1,3,4,5,6,7] ; AVX-NEXT: vpshuflw {{.*#+}} xmm4 = xmm12[2,1,3,3,4,5,6,7] ; AVX-NEXT: vinsertf128 $1, %xmm4, %ymm3, %ymm3 ; AVX-NEXT: vshufps {{.*#+}} ymm3 = ymm3[0,0,2,1,4,4,6,5] ; AVX-NEXT: vandps %ymm3, %ymm9, %ymm3 ; AVX-NEXT: vorps %ymm1, %ymm3, %ymm8 ; AVX-NEXT: vmovdqa (%rsi), %xmm3 ; AVX-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX-NEXT: vmovdqa (%rdi), %xmm1 ; AVX-NEXT: vpunpcklbw {{.*#+}} xmm3 = 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] ; AVX-NEXT: vpmovzxwq {{.*#+}} xmm4 = xmm3[0],zero,zero,zero,xmm3[1],zero,zero,zero ; AVX-NEXT: vpshufd {{.*#+}} xmm10 = xmm3[1,1,1,1] ; AVX-NEXT: vpmovzxwq {{.*#+}} xmm10 = xmm10[0],zero,zero,zero,xmm10[1],zero,zero,zero ; AVX-NEXT: vinsertf128 $1, %xmm10, %ymm4, %ymm13 ; AVX-NEXT: vmovdqa (%rcx), %xmm10 ; AVX-NEXT: vmovdqa (%rdx), %xmm11 ; AVX-NEXT: vpunpcklbw {{.*#+}} xmm15 = 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] ; AVX-NEXT: vpshuflw {{.*#+}} xmm4 = xmm15[0,0,2,1,4,5,6,7] ; AVX-NEXT: vpmovzxdq {{.*#+}} xmm4 = xmm4[0],zero,xmm4[1],zero ; AVX-NEXT: vpshuflw {{.*#+}} xmm14 = xmm15[0,2,2,3,4,5,6,7] ; AVX-NEXT: vpmovzxdq {{.*#+}} xmm14 = xmm14[0],zero,xmm14[1],zero ; AVX-NEXT: vinsertf128 $1, %xmm14, %ymm4, %ymm14 ; AVX-NEXT: vbroadcastsd {{.*#+}} ymm6 = [65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535] ; AVX-NEXT: vandps %ymm6, %ymm13, %ymm13 ; AVX-NEXT: vandnps %ymm14, %ymm6, %ymm14 ; AVX-NEXT: vorps %ymm14, %ymm13, %ymm13 ; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm13[0],ymm8[1],ymm13[2],ymm8[3],ymm13[4],ymm8[5],ymm13[6],ymm8[7] ; AVX-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX-NEXT: vpshufhw {{.*#+}} xmm8 = xmm0[0,1,2,3,4,4,6,5] ; AVX-NEXT: vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,6,6,7] ; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm8, %ymm0 ; AVX-NEXT: vpshufhw {{.*#+}} xmm8 = xmm12[0,1,2,3,4,5,5,7] ; AVX-NEXT: vpshufhw {{.*#+}} xmm12 = xmm12[0,1,2,3,6,5,7,7] ; AVX-NEXT: vinsertf128 $1, %xmm12, %ymm8, %ymm8 ; AVX-NEXT: vmovdqa 16(%r10), %xmm4 ; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm0[0,2,2,3,4,6,6,7] ; AVX-NEXT: vandnps %ymm0, %ymm9, %ymm0 ; AVX-NEXT: vshufps {{.*#+}} ymm8 = ymm8[0,2,2,3,4,6,6,7] ; AVX-NEXT: vandps %ymm9, %ymm8, %ymm8 ; AVX-NEXT: vorps %ymm0, %ymm8, %ymm0 ; AVX-NEXT: vpshufhw {{.*#+}} xmm8 = xmm15[0,1,2,3,4,4,6,5] ; AVX-NEXT: vpshufhw {{.*#+}} xmm12 = xmm15[0,1,2,3,4,6,6,7] ; AVX-NEXT: vinsertf128 $1, %xmm12, %ymm8, %ymm12 ; AVX-NEXT: vpshufd {{.*#+}} xmm8 = xmm3[2,3,2,3] ; AVX-NEXT: vpmovzxwq {{.*#+}} xmm8 = xmm8[0],zero,zero,zero,xmm8[1],zero,zero,zero ; AVX-NEXT: vpshufd {{.*#+}} xmm3 = xmm3[3,3,3,3] ; AVX-NEXT: vpmovzxwq {{.*#+}} xmm3 = xmm3[0],zero,zero,zero,xmm3[1],zero,zero,zero ; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm8, %ymm3 ; AVX-NEXT: vmovdqa 16(%rax), %xmm8 ; AVX-NEXT: vmovdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX-NEXT: vshufps {{.*#+}} ymm12 = ymm12[2,1,3,3,6,5,7,7] ; AVX-NEXT: vandnps %ymm12, %ymm6, %ymm12 ; AVX-NEXT: vandps %ymm6, %ymm3, %ymm3 ; AVX-NEXT: vorps %ymm3, %ymm12, %ymm3 ; AVX-NEXT: vmovdqa 16(%r9), %xmm13 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm3[0],ymm0[1],ymm3[2],ymm0[3],ymm3[4],ymm0[5],ymm3[6],ymm0[7] ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX-NEXT: vpunpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm2 # 16-byte Folded Reload ; AVX-NEXT: # xmm2 = xmm2[8],mem[8],xmm2[9],mem[9],xmm2[10],mem[10],xmm2[11],mem[11],xmm2[12],mem[12],xmm2[13],mem[13],xmm2[14],mem[14],xmm2[15],mem[15] ; AVX-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX-NEXT: vpshufhw {{.*#+}} xmm0 = xmm2[0,1,2,3,4,4,6,5] ; AVX-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,6,6,7] ; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm0 ; AVX-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm7[8],xmm5[8],xmm7[9],xmm5[9],xmm7[10],xmm5[10],xmm7[11],xmm5[11],xmm7[12],xmm5[12],xmm7[13],xmm5[13],xmm7[14],xmm5[14],xmm7[15],xmm5[15] ; AVX-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX-NEXT: vpshufhw {{.*#+}} xmm2 = xmm3[0,1,2,3,4,5,5,7] ; AVX-NEXT: vpshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,6,5,7,7] ; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm2, %ymm2 ; AVX-NEXT: vmovdqa 16(%r8), %xmm7 ; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm0[0,2,2,3,4,6,6,7] ; AVX-NEXT: vandnps %ymm0, %ymm9, %ymm0 ; AVX-NEXT: vshufps {{.*#+}} ymm2 = ymm2[0,2,2,3,4,6,6,7] ; AVX-NEXT: vandps %ymm2, %ymm9, %ymm2 ; AVX-NEXT: vmovaps %ymm9, %ymm5 ; AVX-NEXT: vorps %ymm0, %ymm2, %ymm0 ; AVX-NEXT: vpunpckhbw {{.*#+}} xmm3 = 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] ; AVX-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX-NEXT: vpshufhw {{.*#+}} xmm2 = xmm3[0,1,2,3,4,4,6,5] ; AVX-NEXT: vpshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,6,6,7] ; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm2, %ymm2 ; AVX-NEXT: vshufps {{.*#+}} ymm2 = ymm2[2,1,3,3,6,5,7,7] ; AVX-NEXT: vandnps %ymm2, %ymm6, %ymm2 ; AVX-NEXT: vpunpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm3 # 16-byte Folded Reload ; AVX-NEXT: # xmm3 = 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] ; AVX-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX-NEXT: vpshufd {{.*#+}} xmm1 = xmm3[2,3,2,3] ; AVX-NEXT: vpmovzxwq {{.*#+}} xmm1 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero ; AVX-NEXT: vpshufd {{.*#+}} xmm3 = xmm3[3,3,3,3] ; AVX-NEXT: vpmovzxwq {{.*#+}} xmm3 = xmm3[0],zero,zero,zero,xmm3[1],zero,zero,zero ; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm1, %ymm1 ; AVX-NEXT: vandps %ymm6, %ymm1, %ymm1 ; AVX-NEXT: vorps %ymm2, %ymm1, %ymm1 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[2],ymm0[3],ymm1[4],ymm0[5],ymm1[6],ymm0[7] ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX-NEXT: vpunpcklbw {{.*#+}} xmm11 = xmm8[0],xmm4[0],xmm8[1],xmm4[1],xmm8[2],xmm4[2],xmm8[3],xmm4[3],xmm8[4],xmm4[4],xmm8[5],xmm4[5],xmm8[6],xmm4[6],xmm8[7],xmm4[7] ; AVX-NEXT: vmovdqa %xmm4, %xmm8 ; AVX-NEXT: vpshufhw {{.*#+}} xmm0 = xmm11[0,1,2,3,4,4,6,5] ; AVX-NEXT: vpshufhw {{.*#+}} xmm1 = xmm11[0,1,2,3,4,6,6,7] ; AVX-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0 ; AVX-NEXT: vpunpcklbw {{.*#+}} xmm12 = 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] ; AVX-NEXT: vmovdqa %xmm13, %xmm9 ; AVX-NEXT: vpshufhw {{.*#+}} xmm1 = xmm12[0,1,2,3,4,5,5,7] ; AVX-NEXT: vpshufhw {{.*#+}} xmm2 = xmm12[0,1,2,3,6,5,7,7] ; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm1, %ymm1 ; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm0[0,2,2,3,4,6,6,7] ; AVX-NEXT: vandnps %ymm0, %ymm5, %ymm0 ; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm1[0,2,2,3,4,6,6,7] ; AVX-NEXT: vandps %ymm5, %ymm1, %ymm1 ; AVX-NEXT: vorps %ymm0, %ymm1, %ymm5 ; AVX-NEXT: vmovdqa 16(%rcx), %xmm4 ; AVX-NEXT: vmovdqa 16(%rdx), %xmm3 ; AVX-NEXT: vpunpcklbw {{.*#+}} xmm10 = 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: vpshufhw {{.*#+}} xmm1 = xmm10[0,1,2,3,4,4,6,5] ; AVX-NEXT: vpshufhw {{.*#+}} xmm2 = xmm10[0,1,2,3,4,6,6,7] ; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm1, %ymm0 ; AVX-NEXT: vmovdqa 16(%rsi), %xmm2 ; AVX-NEXT: vmovdqa 16(%rdi), %xmm1 ; AVX-NEXT: vpunpcklbw {{.*#+}} xmm14 = 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 {{.*#+}} xmm13 = xmm14[2,3,2,3] ; AVX-NEXT: vpmovzxwq {{.*#+}} xmm13 = xmm13[0],zero,zero,zero,xmm13[1],zero,zero,zero ; AVX-NEXT: vpshufd {{.*#+}} xmm15 = xmm14[3,3,3,3] ; AVX-NEXT: vpmovzxwq {{.*#+}} xmm15 = xmm15[0],zero,zero,zero,xmm15[1],zero,zero,zero ; AVX-NEXT: vinsertf128 $1, %xmm15, %ymm13, %ymm13 ; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm0[2,1,3,3,6,5,7,7] ; AVX-NEXT: vandnps %ymm0, %ymm6, %ymm0 ; AVX-NEXT: vandps %ymm6, %ymm13, %ymm13 ; AVX-NEXT: vorps %ymm0, %ymm13, %ymm0 ; AVX-NEXT: vblendps {{.*#+}} ymm15 = ymm0[0],ymm5[1],ymm0[2],ymm5[3],ymm0[4],ymm5[5],ymm0[6],ymm5[7] ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload ; AVX-NEXT: vpunpckhbw {{.*#+}} xmm8 = xmm0[8],xmm8[8],xmm0[9],xmm8[9],xmm0[10],xmm8[10],xmm0[11],xmm8[11],xmm0[12],xmm8[12],xmm0[13],xmm8[13],xmm0[14],xmm8[14],xmm0[15],xmm8[15] ; AVX-NEXT: vpshufhw {{.*#+}} xmm0 = xmm8[0,1,2,3,4,4,6,5] ; AVX-NEXT: vpshufhw {{.*#+}} xmm5 = xmm8[0,1,2,3,4,6,6,7] ; AVX-NEXT: vinsertf128 $1, %xmm5, %ymm0, %ymm0 ; AVX-NEXT: vpunpckhbw {{.*#+}} xmm5 = xmm7[8],xmm9[8],xmm7[9],xmm9[9],xmm7[10],xmm9[10],xmm7[11],xmm9[11],xmm7[12],xmm9[12],xmm7[13],xmm9[13],xmm7[14],xmm9[14],xmm7[15],xmm9[15] ; AVX-NEXT: vpshufhw {{.*#+}} xmm7 = xmm5[0,1,2,3,4,5,5,7] ; AVX-NEXT: vpshufhw {{.*#+}} xmm13 = xmm5[0,1,2,3,6,5,7,7] ; AVX-NEXT: vinsertf128 $1, %xmm13, %ymm7, %ymm7 ; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm0[0,2,2,3,4,6,6,7] ; AVX-NEXT: vbroadcastsd {{.*#+}} ymm9 = [65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0] ; AVX-NEXT: vandnps %ymm0, %ymm9, %ymm0 ; AVX-NEXT: vshufps {{.*#+}} ymm7 = ymm7[0,2,2,3,4,6,6,7] ; AVX-NEXT: vandps %ymm7, %ymm9, %ymm7 ; AVX-NEXT: vorps %ymm0, %ymm7, %ymm7 ; AVX-NEXT: vpunpckhbw {{.*#+}} xmm0 = 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: vpshufhw {{.*#+}} xmm3 = xmm0[0,1,2,3,4,4,6,5] ; AVX-NEXT: vpshufhw {{.*#+}} xmm4 = xmm0[0,1,2,3,4,6,6,7] ; AVX-NEXT: vinsertf128 $1, %xmm4, %ymm3, %ymm3 ; AVX-NEXT: vpunpckhbw {{.*#+}} xmm2 = 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 {{.*#+}} xmm1 = xmm2[2,3,2,3] ; AVX-NEXT: vpmovzxwq {{.*#+}} xmm1 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero ; AVX-NEXT: vpshufd {{.*#+}} xmm4 = xmm2[3,3,3,3] ; AVX-NEXT: vpmovzxwq {{.*#+}} xmm4 = xmm4[0],zero,zero,zero,xmm4[1],zero,zero,zero ; AVX-NEXT: vinsertf128 $1, %xmm4, %ymm1, %ymm1 ; AVX-NEXT: vshufps {{.*#+}} ymm3 = ymm3[2,1,3,3,6,5,7,7] ; AVX-NEXT: vandnps %ymm3, %ymm6, %ymm3 ; AVX-NEXT: vandps %ymm6, %ymm1, %ymm1 ; AVX-NEXT: vorps %ymm3, %ymm1, %ymm1 ; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm7[1],ymm1[2],ymm7[3],ymm1[4],ymm7[5],ymm1[6],ymm7[7] ; AVX-NEXT: vpshuflw {{.*#+}} xmm3 = xmm8[0,0,2,1,4,5,6,7] ; AVX-NEXT: vpshuflw {{.*#+}} xmm4 = xmm8[0,2,2,3,4,5,6,7] ; AVX-NEXT: vinsertf128 $1, %xmm4, %ymm3, %ymm3 ; AVX-NEXT: vpshuflw {{.*#+}} xmm4 = xmm5[0,1,1,3,4,5,6,7] ; AVX-NEXT: vpshuflw {{.*#+}} xmm5 = xmm5[2,1,3,3,4,5,6,7] ; AVX-NEXT: vinsertf128 $1, %xmm5, %ymm4, %ymm4 ; AVX-NEXT: vshufps {{.*#+}} ymm3 = ymm3[0,0,2,1,4,4,6,5] ; AVX-NEXT: vandnps %ymm3, %ymm9, %ymm3 ; AVX-NEXT: vshufps {{.*#+}} ymm4 = ymm4[0,0,2,1,4,4,6,5] ; AVX-NEXT: vandps %ymm4, %ymm9, %ymm4 ; AVX-NEXT: vorps %ymm3, %ymm4, %ymm3 ; AVX-NEXT: vpmovzxwq {{.*#+}} xmm4 = xmm2[0],zero,zero,zero,xmm2[1],zero,zero,zero ; AVX-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[1,1,1,1] ; AVX-NEXT: vpmovzxwq {{.*#+}} xmm2 = xmm2[0],zero,zero,zero,xmm2[1],zero,zero,zero ; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm4, %ymm2 ; AVX-NEXT: vpshuflw {{.*#+}} xmm4 = xmm0[0,0,2,1,4,5,6,7] ; AVX-NEXT: vpmovzxdq {{.*#+}} xmm4 = xmm4[0],zero,xmm4[1],zero ; AVX-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[0,2,2,3,4,5,6,7] ; AVX-NEXT: vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero ; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm4, %ymm0 ; AVX-NEXT: vandps %ymm6, %ymm2, %ymm2 ; AVX-NEXT: vandnps %ymm0, %ymm6, %ymm0 ; AVX-NEXT: vorps %ymm0, %ymm2, %ymm0 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm3[1],ymm0[2],ymm3[3],ymm0[4],ymm3[5],ymm0[6],ymm3[7] ; AVX-NEXT: vpshuflw {{.*#+}} xmm2 = xmm11[0,0,2,1,4,5,6,7] ; AVX-NEXT: vpshuflw {{.*#+}} xmm3 = xmm11[0,2,2,3,4,5,6,7] ; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm2, %ymm2 ; AVX-NEXT: vpshuflw {{.*#+}} xmm3 = xmm12[0,1,1,3,4,5,6,7] ; AVX-NEXT: vpshuflw {{.*#+}} xmm4 = xmm12[2,1,3,3,4,5,6,7] ; AVX-NEXT: vinsertf128 $1, %xmm4, %ymm3, %ymm3 ; AVX-NEXT: vshufps {{.*#+}} ymm2 = ymm2[0,0,2,1,4,4,6,5] ; AVX-NEXT: vandnps %ymm2, %ymm9, %ymm2 ; AVX-NEXT: vshufps {{.*#+}} ymm3 = ymm3[0,0,2,1,4,4,6,5] ; AVX-NEXT: vandps %ymm3, %ymm9, %ymm3 ; AVX-NEXT: vorps %ymm2, %ymm3, %ymm2 ; AVX-NEXT: vpmovzxwq {{.*#+}} xmm3 = xmm14[0],zero,zero,zero,xmm14[1],zero,zero,zero ; AVX-NEXT: vpshufd {{.*#+}} xmm4 = xmm14[1,1,1,1] ; AVX-NEXT: vpmovzxwq {{.*#+}} xmm4 = xmm4[0],zero,zero,zero,xmm4[1],zero,zero,zero ; AVX-NEXT: vinsertf128 $1, %xmm4, %ymm3, %ymm3 ; AVX-NEXT: vpshuflw {{.*#+}} xmm4 = xmm10[0,0,2,1,4,5,6,7] ; AVX-NEXT: vpmovzxdq {{.*#+}} xmm4 = xmm4[0],zero,xmm4[1],zero ; AVX-NEXT: vpshuflw {{.*#+}} xmm5 = xmm10[0,2,2,3,4,5,6,7] ; AVX-NEXT: vpmovzxdq {{.*#+}} xmm5 = xmm5[0],zero,xmm5[1],zero ; AVX-NEXT: vinsertf128 $1, %xmm5, %ymm4, %ymm4 ; AVX-NEXT: vandps %ymm6, %ymm3, %ymm3 ; AVX-NEXT: vandnps %ymm4, %ymm6, %ymm4 ; AVX-NEXT: vorps %ymm4, %ymm3, %ymm3 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm3[0],ymm2[1],ymm3[2],ymm2[3],ymm3[4],ymm2[5],ymm3[6],ymm2[7] ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload ; AVX-NEXT: vpshuflw {{.*#+}} xmm3 = xmm4[0,0,2,1,4,5,6,7] ; AVX-NEXT: vpshuflw {{.*#+}} xmm4 = xmm4[0,2,2,3,4,5,6,7] ; AVX-NEXT: vinsertf128 $1, %xmm4, %ymm3, %ymm3 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload ; AVX-NEXT: vpshuflw {{.*#+}} xmm4 = xmm5[0,1,1,3,4,5,6,7] ; AVX-NEXT: vpshuflw {{.*#+}} xmm5 = xmm5[2,1,3,3,4,5,6,7] ; AVX-NEXT: vinsertf128 $1, %xmm5, %ymm4, %ymm4 ; AVX-NEXT: vshufps {{.*#+}} ymm3 = ymm3[0,0,2,1,4,4,6,5] ; AVX-NEXT: vandnps %ymm3, %ymm9, %ymm3 ; AVX-NEXT: vshufps {{.*#+}} ymm4 = ymm4[0,0,2,1,4,4,6,5] ; AVX-NEXT: vandps %ymm4, %ymm9, %ymm4 ; AVX-NEXT: vorps %ymm3, %ymm4, %ymm3 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload ; AVX-NEXT: vpmovzxwq {{.*#+}} xmm4 = xmm5[0],zero,zero,zero,xmm5[1],zero,zero,zero ; AVX-NEXT: vpshufd {{.*#+}} xmm5 = xmm5[1,1,1,1] ; AVX-NEXT: vpmovzxwq {{.*#+}} xmm5 = xmm5[0],zero,zero,zero,xmm5[1],zero,zero,zero ; AVX-NEXT: vinsertf128 $1, %xmm5, %ymm4, %ymm4 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload ; AVX-NEXT: vpshuflw {{.*#+}} xmm5 = xmm7[0,0,2,1,4,5,6,7] ; AVX-NEXT: vpmovzxdq {{.*#+}} xmm5 = xmm5[0],zero,xmm5[1],zero ; AVX-NEXT: vpshuflw {{.*#+}} xmm7 = xmm7[0,2,2,3,4,5,6,7] ; AVX-NEXT: vpmovzxdq {{.*#+}} xmm7 = xmm7[0],zero,xmm7[1],zero ; AVX-NEXT: vinsertf128 $1, %xmm7, %ymm5, %ymm5 ; AVX-NEXT: vandps %ymm6, %ymm4, %ymm4 ; AVX-NEXT: vandnps %ymm5, %ymm6, %ymm5 ; AVX-NEXT: vorps %ymm5, %ymm4, %ymm4 ; AVX-NEXT: vblendps {{.*#+}} ymm3 = ymm4[0],ymm3[1],ymm4[2],ymm3[3],ymm4[4],ymm3[5],ymm4[6],ymm3[7] ; AVX-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX-NEXT: vmovaps %ymm3, 64(%rax) ; AVX-NEXT: vmovaps %ymm2, 128(%rax) ; AVX-NEXT: vmovaps %ymm0, 192(%rax) ; AVX-NEXT: vmovaps %ymm1, 224(%rax) ; AVX-NEXT: vmovaps %ymm15, 160(%rax) ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload ; AVX-NEXT: vmovaps %ymm0, 96(%rax) ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload ; AVX-NEXT: vmovaps %ymm0, 32(%rax) ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload ; AVX-NEXT: vmovaps %ymm0, (%rax) ; AVX-NEXT: addq $56, %rsp ; AVX-NEXT: vzeroupper ; AVX-NEXT: retq ; ; AVX2-LABEL: store_i8_stride8_vf32: ; AVX2: # %bb.0: ; AVX2-NEXT: subq $40, %rsp ; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX2-NEXT: vmovdqa (%r10), %xmm3 ; AVX2-NEXT: vmovdqa (%rax), %xmm4 ; AVX2-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3],xmm4[4],xmm3[4],xmm4[5],xmm3[5],xmm4[6],xmm3[6],xmm4[7],xmm3[7] ; AVX2-NEXT: vpshuflw {{.*#+}} xmm0 = xmm2[0,0,2,1,4,5,6,7] ; AVX2-NEXT: vpshuflw {{.*#+}} xmm1 = xmm2[0,2,2,3,4,5,6,7] ; AVX2-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm1 ; AVX2-NEXT: vmovdqa (%r9), %xmm6 ; AVX2-NEXT: vmovdqa (%r8), %xmm0 ; AVX2-NEXT: vpunpcklbw {{.*#+}} xmm11 = xmm0[0],xmm6[0],xmm0[1],xmm6[1],xmm0[2],xmm6[2],xmm0[3],xmm6[3],xmm0[4],xmm6[4],xmm0[5],xmm6[5],xmm0[6],xmm6[6],xmm0[7],xmm6[7] ; AVX2-NEXT: vpshuflw {{.*#+}} xmm5 = xmm11[0,1,1,3,4,5,6,7] ; AVX2-NEXT: vpshuflw {{.*#+}} xmm7 = xmm11[2,1,3,3,4,5,6,7] ; AVX2-NEXT: vinserti128 $1, %xmm7, %ymm5, %ymm5 ; AVX2-NEXT: vpblendw {{.*#+}} ymm12 = ymm5[0],ymm1[1],ymm5[2],ymm1[3],ymm5[4,5,6,7,8],ymm1[9],ymm5[10],ymm1[11],ymm5[12,13,14,15] ; AVX2-NEXT: vmovdqa (%rsi), %xmm5 ; AVX2-NEXT: vmovdqa (%rdi), %xmm1 ; AVX2-NEXT: vpunpcklbw {{.*#+}} xmm13 = xmm1[0],xmm5[0],xmm1[1],xmm5[1],xmm1[2],xmm5[2],xmm1[3],xmm5[3],xmm1[4],xmm5[4],xmm1[5],xmm5[5],xmm1[6],xmm5[6],xmm1[7],xmm5[7] ; AVX2-NEXT: vpmovzxwq {{.*#+}} ymm9 = xmm13[0],zero,zero,zero,xmm13[1],zero,zero,zero,xmm13[2],zero,zero,zero,xmm13[3],zero,zero,zero ; AVX2-NEXT: vmovdqa (%rcx), %xmm8 ; AVX2-NEXT: vmovdqa (%rdx), %xmm10 ; AVX2-NEXT: vpunpcklbw {{.*#+}} xmm14 = xmm10[0],xmm8[0],xmm10[1],xmm8[1],xmm10[2],xmm8[2],xmm10[3],xmm8[3],xmm10[4],xmm8[4],xmm10[5],xmm8[5],xmm10[6],xmm8[6],xmm10[7],xmm8[7] ; AVX2-NEXT: vpshufb {{.*#+}} xmm15 = xmm14[0,1,0,1,4,5,2,3,0,1,4,5,4,5,6,7] ; AVX2-NEXT: vpmovzxdq {{.*#+}} ymm15 = xmm15[0],zero,xmm15[1],zero,xmm15[2],zero,xmm15[3],zero ; AVX2-NEXT: vpblendw {{.*#+}} ymm15 = ymm9[0],ymm15[1],ymm9[2,3,4],ymm15[5],ymm9[6,7,8],ymm15[9],ymm9[10,11,12],ymm15[13],ymm9[14,15] ; AVX2-NEXT: vmovaps 16(%r10), %xmm7 ; AVX2-NEXT: vmovaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX2-NEXT: vpshufd {{.*#+}} ymm12 = ymm12[0,0,2,1,4,4,6,5] ; AVX2-NEXT: vpblendd {{.*#+}} ymm7 = ymm15[0],ymm12[1],ymm15[2],ymm12[3],ymm15[4],ymm12[5],ymm15[6],ymm12[7] ; AVX2-NEXT: vmovdqu %ymm7, (%rsp) # 32-byte Spill ; AVX2-NEXT: vpshufhw {{.*#+}} xmm12 = xmm2[0,1,2,3,4,4,6,5] ; AVX2-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,6,6,7] ; AVX2-NEXT: vinserti128 $1, %xmm2, %ymm12, %ymm2 ; AVX2-NEXT: vpshufhw {{.*#+}} xmm12 = xmm11[0,1,2,3,4,5,5,7] ; AVX2-NEXT: vpshufhw {{.*#+}} xmm11 = xmm11[0,1,2,3,6,5,7,7] ; AVX2-NEXT: vinserti128 $1, %xmm11, %ymm12, %ymm11 ; AVX2-NEXT: vmovdqa 16(%rax), %xmm12 ; AVX2-NEXT: vpblendw {{.*#+}} ymm2 = ymm11[0,1,2,3,4],ymm2[5],ymm11[6],ymm2[7],ymm11[8,9,10,11,12],ymm2[13],ymm11[14],ymm2[15] ; AVX2-NEXT: vpshufhw {{.*#+}} xmm11 = xmm14[0,1,2,3,4,4,6,5] ; AVX2-NEXT: vpshufhw {{.*#+}} xmm14 = xmm14[0,1,2,3,4,6,6,7] ; AVX2-NEXT: vinserti128 $1, %xmm14, %ymm11, %ymm11 ; AVX2-NEXT: vmovdqa 16(%r9), %xmm14 ; AVX2-NEXT: vpshufd {{.*#+}} ymm11 = ymm11[2,1,3,3,6,5,7,7] ; AVX2-NEXT: vpshufd {{.*#+}} xmm13 = xmm13[2,3,2,3] ; AVX2-NEXT: vpmovzxwq {{.*#+}} ymm13 = xmm13[0],zero,zero,zero,xmm13[1],zero,zero,zero,xmm13[2],zero,zero,zero,xmm13[3],zero,zero,zero ; AVX2-NEXT: vpblendw {{.*#+}} ymm11 = ymm13[0],ymm11[1],ymm13[2,3,4],ymm11[5],ymm13[6,7,8],ymm11[9],ymm13[10,11,12],ymm11[13],ymm13[14,15] ; AVX2-NEXT: vmovdqa 16(%r8), %xmm15 ; AVX2-NEXT: vpshufd {{.*#+}} ymm2 = ymm2[0,2,2,3,4,6,6,7] ; AVX2-NEXT: vpblendd {{.*#+}} ymm2 = ymm11[0],ymm2[1],ymm11[2],ymm2[3],ymm11[4],ymm2[5],ymm11[6],ymm2[7] ; AVX2-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm4[8],xmm3[8],xmm4[9],xmm3[9],xmm4[10],xmm3[10],xmm4[11],xmm3[11],xmm4[12],xmm3[12],xmm4[13],xmm3[13],xmm4[14],xmm3[14],xmm4[15],xmm3[15] ; AVX2-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX2-NEXT: vpshufhw {{.*#+}} xmm4 = xmm2[0,1,2,3,4,4,6,5] ; AVX2-NEXT: vpshufhw {{.*#+}} xmm11 = xmm2[0,1,2,3,4,6,6,7] ; AVX2-NEXT: vinserti128 $1, %xmm11, %ymm4, %ymm11 ; AVX2-NEXT: vpunpckhbw {{.*#+}} xmm9 = xmm0[8],xmm6[8],xmm0[9],xmm6[9],xmm0[10],xmm6[10],xmm0[11],xmm6[11],xmm0[12],xmm6[12],xmm0[13],xmm6[13],xmm0[14],xmm6[14],xmm0[15],xmm6[15] ; AVX2-NEXT: vpshufhw {{.*#+}} xmm0 = xmm9[0,1,2,3,4,5,5,7] ; AVX2-NEXT: vpshufhw {{.*#+}} xmm6 = xmm9[0,1,2,3,6,5,7,7] ; AVX2-NEXT: vinserti128 $1, %xmm6, %ymm0, %ymm0 ; AVX2-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm11[5],ymm0[6],ymm11[7],ymm0[8,9,10,11,12],ymm11[13],ymm0[14],ymm11[15] ; AVX2-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[0,2,2,3,4,6,6,7] ; AVX2-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm10[8],xmm8[8],xmm10[9],xmm8[9],xmm10[10],xmm8[10],xmm10[11],xmm8[11],xmm10[12],xmm8[12],xmm10[13],xmm8[13],xmm10[14],xmm8[14],xmm10[15],xmm8[15] ; AVX2-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX2-NEXT: vpshufhw {{.*#+}} xmm8 = xmm2[0,1,2,3,4,4,6,5] ; AVX2-NEXT: vpshufhw {{.*#+}} xmm10 = xmm2[0,1,2,3,4,6,6,7] ; AVX2-NEXT: vinserti128 $1, %xmm10, %ymm8, %ymm8 ; AVX2-NEXT: vpshufd {{.*#+}} ymm8 = ymm8[2,1,3,3,6,5,7,7] ; AVX2-NEXT: vpunpckhbw {{.*#+}} 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] ; AVX2-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX2-NEXT: vpshufd {{.*#+}} xmm5 = xmm1[2,3,2,3] ; AVX2-NEXT: vpmovzxwq {{.*#+}} ymm5 = xmm5[0],zero,zero,zero,xmm5[1],zero,zero,zero,xmm5[2],zero,zero,zero,xmm5[3],zero,zero,zero ; AVX2-NEXT: vpblendw {{.*#+}} ymm5 = ymm5[0],ymm8[1],ymm5[2,3,4],ymm8[5],ymm5[6,7,8],ymm8[9],ymm5[10,11,12],ymm8[13],ymm5[14,15] ; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm5[0],ymm0[1],ymm5[2],ymm0[3],ymm5[4],ymm0[5],ymm5[6],ymm0[7] ; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload ; AVX2-NEXT: vpunpcklbw {{.*#+}} xmm10 = xmm12[0],xmm3[0],xmm12[1],xmm3[1],xmm12[2],xmm3[2],xmm12[3],xmm3[3],xmm12[4],xmm3[4],xmm12[5],xmm3[5],xmm12[6],xmm3[6],xmm12[7],xmm3[7] ; AVX2-NEXT: vpshufhw {{.*#+}} xmm0 = xmm10[0,1,2,3,4,4,6,5] ; AVX2-NEXT: vpshufhw {{.*#+}} xmm8 = xmm10[0,1,2,3,4,6,6,7] ; AVX2-NEXT: vinserti128 $1, %xmm8, %ymm0, %ymm0 ; AVX2-NEXT: vpunpcklbw {{.*#+}} xmm11 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3],xmm15[4],xmm14[4],xmm15[5],xmm14[5],xmm15[6],xmm14[6],xmm15[7],xmm14[7] ; AVX2-NEXT: vpshufhw {{.*#+}} xmm8 = xmm11[0,1,2,3,4,5,5,7] ; AVX2-NEXT: vpshufhw {{.*#+}} xmm13 = xmm11[0,1,2,3,6,5,7,7] ; AVX2-NEXT: vinserti128 $1, %xmm13, %ymm8, %ymm8 ; AVX2-NEXT: vpblendw {{.*#+}} ymm8 = ymm8[0,1,2,3,4],ymm0[5],ymm8[6],ymm0[7],ymm8[8,9,10,11,12],ymm0[13],ymm8[14],ymm0[15] ; AVX2-NEXT: vmovdqa 16(%rcx), %xmm7 ; AVX2-NEXT: vmovdqa 16(%rdx), %xmm4 ; AVX2-NEXT: vpunpcklbw {{.*#+}} xmm13 = 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] ; AVX2-NEXT: vpshufhw {{.*#+}} xmm2 = xmm13[0,1,2,3,4,4,6,5] ; AVX2-NEXT: vpshufhw {{.*#+}} xmm5 = xmm13[0,1,2,3,4,6,6,7] ; AVX2-NEXT: vinserti128 $1, %xmm5, %ymm2, %ymm2 ; AVX2-NEXT: vpshufd {{.*#+}} ymm5 = ymm2[2,1,3,3,6,5,7,7] ; AVX2-NEXT: vmovdqa 16(%rsi), %xmm1 ; AVX2-NEXT: vmovdqa 16(%rdi), %xmm0 ; AVX2-NEXT: vpunpcklbw {{.*#+}} xmm2 = 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] ; AVX2-NEXT: vpshufd {{.*#+}} xmm6 = xmm2[2,3,2,3] ; AVX2-NEXT: vpmovzxwq {{.*#+}} ymm6 = xmm6[0],zero,zero,zero,xmm6[1],zero,zero,zero,xmm6[2],zero,zero,zero,xmm6[3],zero,zero,zero ; AVX2-NEXT: vpblendw {{.*#+}} ymm5 = ymm6[0],ymm5[1],ymm6[2,3,4],ymm5[5],ymm6[6,7,8],ymm5[9],ymm6[10,11,12],ymm5[13],ymm6[14,15] ; AVX2-NEXT: vpshufd {{.*#+}} ymm6 = ymm8[0,2,2,3,4,6,6,7] ; AVX2-NEXT: vpblendd {{.*#+}} ymm8 = ymm5[0],ymm6[1],ymm5[2],ymm6[3],ymm5[4],ymm6[5],ymm5[6],ymm6[7] ; AVX2-NEXT: vpunpckhbw {{.*#+}} xmm5 = xmm12[8],xmm3[8],xmm12[9],xmm3[9],xmm12[10],xmm3[10],xmm12[11],xmm3[11],xmm12[12],xmm3[12],xmm12[13],xmm3[13],xmm12[14],xmm3[14],xmm12[15],xmm3[15] ; AVX2-NEXT: vpshufhw {{.*#+}} xmm6 = xmm5[0,1,2,3,4,4,6,5] ; AVX2-NEXT: vpshufhw {{.*#+}} xmm12 = xmm5[0,1,2,3,4,6,6,7] ; AVX2-NEXT: vinserti128 $1, %xmm12, %ymm6, %ymm6 ; AVX2-NEXT: vpunpckhbw {{.*#+}} xmm12 = xmm15[8],xmm14[8],xmm15[9],xmm14[9],xmm15[10],xmm14[10],xmm15[11],xmm14[11],xmm15[12],xmm14[12],xmm15[13],xmm14[13],xmm15[14],xmm14[14],xmm15[15],xmm14[15] ; AVX2-NEXT: vpshufhw {{.*#+}} xmm14 = xmm12[0,1,2,3,4,5,5,7] ; AVX2-NEXT: vpshufhw {{.*#+}} xmm15 = xmm12[0,1,2,3,6,5,7,7] ; AVX2-NEXT: vinserti128 $1, %xmm15, %ymm14, %ymm14 ; AVX2-NEXT: vpblendw {{.*#+}} ymm6 = ymm14[0,1,2,3,4],ymm6[5],ymm14[6],ymm6[7],ymm14[8,9,10,11,12],ymm6[13],ymm14[14],ymm6[15] ; AVX2-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] ; AVX2-NEXT: vpshufhw {{.*#+}} xmm4 = xmm3[0,1,2,3,4,4,6,5] ; AVX2-NEXT: vpshufhw {{.*#+}} xmm14 = xmm3[0,1,2,3,4,6,6,7] ; AVX2-NEXT: vinserti128 $1, %xmm14, %ymm4, %ymm4 ; AVX2-NEXT: vpunpckhbw {{.*#+}} xmm1 = 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] ; AVX2-NEXT: vpshufd {{.*#+}} ymm0 = ymm4[2,1,3,3,6,5,7,7] ; AVX2-NEXT: vpshufd {{.*#+}} xmm4 = xmm1[2,3,2,3] ; AVX2-NEXT: vpmovzxwq {{.*#+}} ymm4 = xmm4[0],zero,zero,zero,xmm4[1],zero,zero,zero,xmm4[2],zero,zero,zero,xmm4[3],zero,zero,zero ; AVX2-NEXT: vpblendw {{.*#+}} ymm0 = ymm4[0],ymm0[1],ymm4[2,3,4],ymm0[5],ymm4[6,7,8],ymm0[9],ymm4[10,11,12],ymm0[13],ymm4[14,15] ; AVX2-NEXT: vpshufd {{.*#+}} ymm4 = ymm6[0,2,2,3,4,6,6,7] ; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0],ymm4[1],ymm0[2],ymm4[3],ymm0[4],ymm4[5],ymm0[6],ymm4[7] ; AVX2-NEXT: vpshuflw {{.*#+}} xmm4 = xmm5[0,0,2,1,4,5,6,7] ; AVX2-NEXT: vpshuflw {{.*#+}} xmm5 = xmm5[0,2,2,3,4,5,6,7] ; AVX2-NEXT: vinserti128 $1, %xmm5, %ymm4, %ymm4 ; AVX2-NEXT: vpshuflw {{.*#+}} xmm5 = xmm12[0,1,1,3,4,5,6,7] ; AVX2-NEXT: vpshuflw {{.*#+}} xmm6 = xmm12[2,1,3,3,4,5,6,7] ; AVX2-NEXT: vinserti128 $1, %xmm6, %ymm5, %ymm5 ; AVX2-NEXT: vpblendw {{.*#+}} ymm4 = ymm5[0],ymm4[1],ymm5[2],ymm4[3],ymm5[4,5,6,7,8],ymm4[9],ymm5[10],ymm4[11],ymm5[12,13,14,15] ; AVX2-NEXT: vpmovzxwq {{.*#+}} ymm1 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero,xmm1[2],zero,zero,zero,xmm1[3],zero,zero,zero ; AVX2-NEXT: vmovdqa {{.*#+}} xmm5 = [0,1,0,1,4,5,2,3,0,1,4,5,4,5,6,7] ; AVX2-NEXT: vpshufb %xmm5, %xmm3, %xmm3 ; AVX2-NEXT: vpmovzxdq {{.*#+}} ymm3 = xmm3[0],zero,xmm3[1],zero,xmm3[2],zero,xmm3[3],zero ; AVX2-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0],ymm3[1],ymm1[2,3,4],ymm3[5],ymm1[6,7,8],ymm3[9],ymm1[10,11,12],ymm3[13],ymm1[14,15] ; AVX2-NEXT: vpshufd {{.*#+}} ymm3 = ymm4[0,0,2,1,4,4,6,5] ; AVX2-NEXT: vpblendd {{.*#+}} ymm12 = ymm1[0],ymm3[1],ymm1[2],ymm3[3],ymm1[4],ymm3[5],ymm1[6],ymm3[7] ; AVX2-NEXT: vpshuflw {{.*#+}} xmm1 = xmm10[0,0,2,1,4,5,6,7] ; AVX2-NEXT: vpshuflw {{.*#+}} xmm3 = xmm10[0,2,2,3,4,5,6,7] ; AVX2-NEXT: vinserti128 $1, %xmm3, %ymm1, %ymm1 ; AVX2-NEXT: vpshuflw {{.*#+}} xmm3 = xmm11[0,1,1,3,4,5,6,7] ; AVX2-NEXT: vpshuflw {{.*#+}} xmm4 = xmm11[2,1,3,3,4,5,6,7] ; AVX2-NEXT: vinserti128 $1, %xmm4, %ymm3, %ymm3 ; AVX2-NEXT: vpblendw {{.*#+}} ymm1 = ymm3[0],ymm1[1],ymm3[2],ymm1[3],ymm3[4,5,6,7,8],ymm1[9],ymm3[10],ymm1[11],ymm3[12,13,14,15] ; AVX2-NEXT: vpmovzxwq {{.*#+}} ymm2 = xmm2[0],zero,zero,zero,xmm2[1],zero,zero,zero,xmm2[2],zero,zero,zero,xmm2[3],zero,zero,zero ; AVX2-NEXT: vpshufb %xmm5, %xmm13, %xmm3 ; AVX2-NEXT: vpmovzxdq {{.*#+}} ymm3 = xmm3[0],zero,xmm3[1],zero,xmm3[2],zero,xmm3[3],zero ; AVX2-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0],ymm3[1],ymm2[2,3,4],ymm3[5],ymm2[6,7,8],ymm3[9],ymm2[10,11,12],ymm3[13],ymm2[14,15] ; AVX2-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[0,0,2,1,4,4,6,5] ; AVX2-NEXT: vpblendd {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2],ymm1[3],ymm2[4],ymm1[5],ymm2[6],ymm1[7] ; AVX2-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload ; AVX2-NEXT: vpshuflw {{.*#+}} xmm2 = xmm3[0,0,2,1,4,5,6,7] ; AVX2-NEXT: vpshuflw {{.*#+}} xmm3 = xmm3[0,2,2,3,4,5,6,7] ; AVX2-NEXT: vinserti128 $1, %xmm3, %ymm2, %ymm2 ; AVX2-NEXT: vpshuflw {{.*#+}} xmm3 = xmm9[0,1,1,3,4,5,6,7] ; AVX2-NEXT: vpshuflw {{.*#+}} xmm4 = xmm9[2,1,3,3,4,5,6,7] ; AVX2-NEXT: vinserti128 $1, %xmm4, %ymm3, %ymm3 ; AVX2-NEXT: vpblendw {{.*#+}} ymm2 = ymm3[0],ymm2[1],ymm3[2],ymm2[3],ymm3[4,5,6,7,8],ymm2[9],ymm3[10],ymm2[11],ymm3[12,13,14,15] ; AVX2-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload ; AVX2-NEXT: vpshufb %xmm5, %xmm3, %xmm3 ; AVX2-NEXT: vpmovzxwq {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 16-byte Folded Reload ; AVX2-NEXT: # ymm4 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero ; AVX2-NEXT: vpmovzxdq {{.*#+}} ymm3 = xmm3[0],zero,xmm3[1],zero,xmm3[2],zero,xmm3[3],zero ; AVX2-NEXT: vpblendw {{.*#+}} ymm3 = ymm4[0],ymm3[1],ymm4[2,3,4],ymm3[5],ymm4[6,7,8],ymm3[9],ymm4[10,11,12],ymm3[13],ymm4[14,15] ; AVX2-NEXT: vpshufd {{.*#+}} ymm2 = ymm2[0,0,2,1,4,4,6,5] ; AVX2-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0],ymm2[1],ymm3[2],ymm2[3],ymm3[4],ymm2[5],ymm3[6],ymm2[7] ; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX2-NEXT: vmovdqa %ymm2, 64(%rax) ; AVX2-NEXT: vmovdqa %ymm1, 128(%rax) ; AVX2-NEXT: vmovdqa %ymm12, 192(%rax) ; AVX2-NEXT: vmovdqa %ymm0, 224(%rax) ; AVX2-NEXT: vmovdqa %ymm8, 160(%rax) ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload ; AVX2-NEXT: vmovaps %ymm0, 96(%rax) ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload ; AVX2-NEXT: vmovaps %ymm0, 32(%rax) ; AVX2-NEXT: vmovups (%rsp), %ymm0 # 32-byte Reload ; AVX2-NEXT: vmovaps %ymm0, (%rax) ; AVX2-NEXT: addq $40, %rsp ; AVX2-NEXT: vzeroupper ; AVX2-NEXT: retq ; ; AVX2-FP-LABEL: store_i8_stride8_vf32: ; AVX2-FP: # %bb.0: ; AVX2-FP-NEXT: subq $72, %rsp ; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX2-FP-NEXT: vmovdqa (%r10), %xmm5 ; AVX2-FP-NEXT: vmovdqa (%rax), %xmm6 ; AVX2-FP-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3],xmm6[4],xmm5[4],xmm6[5],xmm5[5],xmm6[6],xmm5[6],xmm6[7],xmm5[7] ; AVX2-FP-NEXT: vinserti128 $1, %xmm1, %ymm1, %ymm8 ; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm3 = ymm8[0,1,0,1,0,1,0,1,8,9,10,11,4,5,2,3,16,17,20,21,16,17,20,21,24,25,26,27,20,21,22,23] ; AVX2-FP-NEXT: vmovdqa (%r9), %xmm1 ; AVX2-FP-NEXT: vmovdqa (%r8), %xmm2 ; AVX2-FP-NEXT: vpunpcklbw {{.*#+}} xmm4 = 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: vinserti128 $1, %xmm4, %ymm4, %ymm9 ; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm4 = ymm9[0,1,2,3,0,1,2,3,8,9,10,11,2,3,6,7,20,21,18,19,20,21,18,19,24,25,26,27,22,23,22,23] ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm11 = ymm4[0,1,2],ymm3[3],ymm4[4,5,6],ymm3[7],ymm4[8,9,10],ymm3[11],ymm4[12,13,14],ymm3[15] ; AVX2-FP-NEXT: vmovdqa (%rcx), %xmm7 ; AVX2-FP-NEXT: vmovdqa (%rsi), %xmm3 ; AVX2-FP-NEXT: vmovdqa (%rdi), %xmm10 ; AVX2-FP-NEXT: vpunpcklbw {{.*#+}} xmm12 = xmm10[0],xmm3[0],xmm10[1],xmm3[1],xmm10[2],xmm3[2],xmm10[3],xmm3[3],xmm10[4],xmm3[4],xmm10[5],xmm3[5],xmm10[6],xmm3[6],xmm10[7],xmm3[7] ; AVX2-FP-NEXT: vpshufb {{.*#+}} xmm4 = xmm12[4,5,u,u,u,u,u,u,6,7,u,u,u,u,u,u] ; AVX2-FP-NEXT: vpmovzxwq {{.*#+}} xmm13 = xmm12[0],zero,zero,zero,xmm12[1],zero,zero,zero ; AVX2-FP-NEXT: vinserti128 $1, %xmm4, %ymm13, %ymm13 ; AVX2-FP-NEXT: vmovdqa (%rdx), %xmm15 ; AVX2-FP-NEXT: vpunpcklbw {{.*#+}} xmm4 = xmm15[0],xmm7[0],xmm15[1],xmm7[1],xmm15[2],xmm7[2],xmm15[3],xmm7[3],xmm15[4],xmm7[4],xmm15[5],xmm7[5],xmm15[6],xmm7[6],xmm15[7],xmm7[7] ; AVX2-FP-NEXT: vinserti128 $1, %xmm4, %ymm4, %ymm0 ; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm14 = ymm0[0,1,0,1,4,5,2,3,4,5,2,3,12,13,14,15,16,17,20,21,20,21,22,23,20,21,22,23,28,29,30,31] ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm13 = ymm13[0],ymm14[1],ymm13[2,3,4],ymm14[5],ymm13[6,7,8],ymm14[9],ymm13[10,11,12],ymm14[13],ymm13[14,15] ; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm4 = ymm13[0],ymm11[1],ymm13[2],ymm11[3],ymm13[4],ymm11[5],ymm13[6],ymm11[7] ; AVX2-FP-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm14 = [0,1,2,3,8,9,8,9,8,9,8,9,12,13,10,11,0,1,2,3,8,9,12,13,8,9,12,13,12,13,14,15] ; AVX2-FP-NEXT: vpshufb %ymm14, %ymm8, %ymm8 ; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm13 = [0,1,2,3,8,9,10,11,8,9,10,11,10,11,14,15,0,1,2,3,12,13,10,11,12,13,10,11,14,15,14,15] ; AVX2-FP-NEXT: vpshufb %ymm13, %ymm9, %ymm9 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm8 = ymm9[0,1,2],ymm8[3],ymm9[4,5,6],ymm8[7],ymm9[8,9,10],ymm8[11],ymm9[12,13,14],ymm8[15] ; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm4 = [8,9,8,9,4,5,6,7,12,13,10,11,12,13,10,11,8,9,12,13,4,5,6,7,12,13,14,15,12,13,14,15] ; AVX2-FP-NEXT: vpshufb %ymm4, %ymm0, %ymm0 ; AVX2-FP-NEXT: vinserti128 $1, %xmm12, %ymm12, %ymm9 ; AVX2-FP-NEXT: vpmovsxwq {{.*#+}} ymm12 = [2312,2826,3340,3854] ; AVX2-FP-NEXT: vpshufb %ymm12, %ymm9, %ymm9 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm0 = ymm9[0],ymm0[1],ymm9[2,3,4],ymm0[5],ymm9[6,7,8],ymm0[9],ymm9[10,11,12],ymm0[13],ymm9[14,15] ; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0],ymm8[1],ymm0[2],ymm8[3],ymm0[4],ymm8[5],ymm0[6],ymm8[7] ; AVX2-FP-NEXT: vmovdqu %ymm0, (%rsp) # 32-byte Spill ; AVX2-FP-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm6[8],xmm5[8],xmm6[9],xmm5[9],xmm6[10],xmm5[10],xmm6[11],xmm5[11],xmm6[12],xmm5[12],xmm6[13],xmm5[13],xmm6[14],xmm5[14],xmm6[15],xmm5[15] ; AVX2-FP-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] ; AVX2-FP-NEXT: vinserti128 $1, %xmm0, %ymm0, %ymm0 ; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FP-NEXT: vinserti128 $1, %xmm1, %ymm1, %ymm1 ; AVX2-FP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FP-NEXT: vpshufb %ymm14, %ymm0, %ymm0 ; AVX2-FP-NEXT: vpshufb %ymm13, %ymm1, %ymm1 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3],ymm1[4,5,6],ymm0[7],ymm1[8,9,10],ymm0[11],ymm1[12,13,14],ymm0[15] ; AVX2-FP-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm15[8],xmm7[8],xmm15[9],xmm7[9],xmm15[10],xmm7[10],xmm15[11],xmm7[11],xmm15[12],xmm7[12],xmm15[13],xmm7[13],xmm15[14],xmm7[14],xmm15[15],xmm7[15] ; AVX2-FP-NEXT: vinserti128 $1, %xmm1, %ymm1, %ymm2 ; AVX2-FP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FP-NEXT: vpunpckhbw {{.*#+}} xmm11 = xmm10[8],xmm3[8],xmm10[9],xmm3[9],xmm10[10],xmm3[10],xmm10[11],xmm3[11],xmm10[12],xmm3[12],xmm10[13],xmm3[13],xmm10[14],xmm3[14],xmm10[15],xmm3[15] ; AVX2-FP-NEXT: vinserti128 $1, %xmm11, %ymm11, %ymm1 ; AVX2-FP-NEXT: vpshufb %ymm12, %ymm1, %ymm1 ; AVX2-FP-NEXT: vpshufb %ymm4, %ymm2, %ymm2 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3,4],ymm2[5],ymm1[6,7,8],ymm2[9],ymm1[10,11,12],ymm2[13],ymm1[14,15] ; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[2],ymm0[3],ymm1[4],ymm0[5],ymm1[6],ymm0[7] ; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FP-NEXT: vmovdqa 16(%r10), %xmm8 ; AVX2-FP-NEXT: vmovdqa 16(%rax), %xmm6 ; AVX2-FP-NEXT: vpunpcklbw {{.*#+}} xmm1 = 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] ; AVX2-FP-NEXT: vinserti128 $1, %xmm1, %ymm1, %ymm12 ; AVX2-FP-NEXT: vmovdqa 16(%r9), %xmm7 ; AVX2-FP-NEXT: vmovdqa 16(%r8), %xmm5 ; AVX2-FP-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm5[0],xmm7[0],xmm5[1],xmm7[1],xmm5[2],xmm7[2],xmm5[3],xmm7[3],xmm5[4],xmm7[4],xmm5[5],xmm7[5],xmm5[6],xmm7[6],xmm5[7],xmm7[7] ; AVX2-FP-NEXT: vinserti128 $1, %xmm3, %ymm3, %ymm10 ; AVX2-FP-NEXT: vpshufb %ymm14, %ymm12, %ymm3 ; AVX2-FP-NEXT: vpshufb %ymm13, %ymm10, %ymm4 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm15 = ymm4[0,1,2],ymm3[3],ymm4[4,5,6],ymm3[7],ymm4[8,9,10],ymm3[11],ymm4[12,13,14],ymm3[15] ; AVX2-FP-NEXT: vmovdqa 16(%rcx), %xmm4 ; AVX2-FP-NEXT: vmovdqa 16(%rdx), %xmm2 ; AVX2-FP-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3],xmm2[4],xmm4[4],xmm2[5],xmm4[5],xmm2[6],xmm4[6],xmm2[7],xmm4[7] ; AVX2-FP-NEXT: vinserti128 $1, %xmm3, %ymm3, %ymm9 ; AVX2-FP-NEXT: vmovdqa 16(%rsi), %xmm1 ; AVX2-FP-NEXT: vmovdqa 16(%rdi), %xmm0 ; AVX2-FP-NEXT: vpunpcklbw {{.*#+}} xmm3 = 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] ; AVX2-FP-NEXT: vinserti128 $1, %xmm3, %ymm3, %ymm14 ; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm14 = ymm14[8,9,u,u,u,u,u,u,10,11,u,u,u,u,u,u,28,29,u,u,u,u,u,u,30,31,u,u,u,u,u,u] ; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm13 = ymm9[8,9,8,9,4,5,6,7,12,13,10,11,12,13,10,11,24,25,28,29,20,21,22,23,28,29,30,31,28,29,30,31] ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm13 = ymm14[0],ymm13[1],ymm14[2,3,4],ymm13[5],ymm14[6,7,8],ymm13[9],ymm14[10,11,12],ymm13[13],ymm14[14,15] ; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm15 = ymm13[0],ymm15[1],ymm13[2],ymm15[3],ymm13[4],ymm15[5],ymm13[6],ymm15[7] ; AVX2-FP-NEXT: vpunpckhbw {{.*#+}} xmm6 = xmm6[8],xmm8[8],xmm6[9],xmm8[9],xmm6[10],xmm8[10],xmm6[11],xmm8[11],xmm6[12],xmm8[12],xmm6[13],xmm8[13],xmm6[14],xmm8[14],xmm6[15],xmm8[15] ; AVX2-FP-NEXT: vpunpckhbw {{.*#+}} xmm5 = xmm5[8],xmm7[8],xmm5[9],xmm7[9],xmm5[10],xmm7[10],xmm5[11],xmm7[11],xmm5[12],xmm7[12],xmm5[13],xmm7[13],xmm5[14],xmm7[14],xmm5[15],xmm7[15] ; AVX2-FP-NEXT: vinserti128 $1, %xmm6, %ymm6, %ymm6 ; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm7 = ymm6[0,1,2,3,8,9,8,9,8,9,8,9,12,13,10,11,16,17,18,19,24,25,28,29,24,25,28,29,28,29,30,31] ; AVX2-FP-NEXT: vinserti128 $1, %xmm5, %ymm5, %ymm5 ; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm8 = ymm5[0,1,2,3,8,9,10,11,8,9,10,11,10,11,14,15,16,17,18,19,28,29,26,27,28,29,26,27,30,31,30,31] ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm7 = ymm8[0,1,2],ymm7[3],ymm8[4,5,6],ymm7[7],ymm8[8,9,10],ymm7[11],ymm8[12,13,14],ymm7[15] ; AVX2-FP-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm2[8],xmm4[8],xmm2[9],xmm4[9],xmm2[10],xmm4[10],xmm2[11],xmm4[11],xmm2[12],xmm4[12],xmm2[13],xmm4[13],xmm2[14],xmm4[14],xmm2[15],xmm4[15] ; AVX2-FP-NEXT: vpunpckhbw {{.*#+}} xmm1 = 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] ; AVX2-FP-NEXT: vinserti128 $1, %xmm1, %ymm1, %ymm0 ; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[8,9,u,u,u,u,u,u,10,11,u,u,u,u,u,u,28,29,u,u,u,u,u,u,30,31,u,u,u,u,u,u] ; AVX2-FP-NEXT: vinserti128 $1, %xmm2, %ymm2, %ymm2 ; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm4 = ymm2[8,9,8,9,4,5,6,7,12,13,10,11,12,13,10,11,24,25,28,29,20,21,22,23,28,29,30,31,28,29,30,31] ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm4[1],ymm0[2,3,4],ymm4[5],ymm0[6,7,8],ymm4[9],ymm0[10,11,12],ymm4[13],ymm0[14,15] ; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0],ymm7[1],ymm0[2],ymm7[3],ymm0[4],ymm7[5],ymm0[6],ymm7[7] ; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm7 = [0,1,0,1,0,1,0,1,8,9,10,11,4,5,2,3,0,1,4,5,0,1,4,5,8,9,10,11,4,5,6,7] ; AVX2-FP-NEXT: vpshufb %ymm7, %ymm6, %ymm4 ; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm6 = [0,1,2,3,0,1,2,3,8,9,10,11,2,3,6,7,4,5,2,3,4,5,2,3,8,9,10,11,6,7,6,7] ; AVX2-FP-NEXT: vpshufb %ymm6, %ymm5, %ymm5 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm4 = ymm5[0,1,2],ymm4[3],ymm5[4,5,6],ymm4[7],ymm5[8,9,10],ymm4[11],ymm5[12,13,14],ymm4[15] ; AVX2-FP-NEXT: vpmovsxwq {{.*#+}} xmm8 = [1284,1798] ; AVX2-FP-NEXT: vpshufb %xmm8, %xmm1, %xmm5 ; AVX2-FP-NEXT: vpmovzxwq {{.*#+}} xmm1 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero ; AVX2-FP-NEXT: vinserti128 $1, %xmm5, %ymm1, %ymm1 ; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm5 = [0,1,0,1,4,5,2,3,4,5,2,3,12,13,14,15,0,1,4,5,4,5,6,7,4,5,6,7,12,13,14,15] ; AVX2-FP-NEXT: vpshufb %ymm5, %ymm2, %ymm2 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3,4],ymm2[5],ymm1[6,7,8],ymm2[9],ymm1[10,11,12],ymm2[13],ymm1[14,15] ; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0],ymm4[1],ymm1[2],ymm4[3],ymm1[4],ymm4[5],ymm1[6],ymm4[7] ; AVX2-FP-NEXT: vpshufb %ymm7, %ymm12, %ymm2 ; AVX2-FP-NEXT: vpshufb %ymm6, %ymm10, %ymm4 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm2 = ymm4[0,1,2],ymm2[3],ymm4[4,5,6],ymm2[7],ymm4[8,9,10],ymm2[11],ymm4[12,13,14],ymm2[15] ; AVX2-FP-NEXT: vpshufb %xmm8, %xmm3, %xmm4 ; AVX2-FP-NEXT: vpmovzxwq {{.*#+}} xmm3 = xmm3[0],zero,zero,zero,xmm3[1],zero,zero,zero ; AVX2-FP-NEXT: vinserti128 $1, %xmm4, %ymm3, %ymm3 ; AVX2-FP-NEXT: vpshufb %ymm5, %ymm9, %ymm4 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0],ymm4[1],ymm3[2,3,4],ymm4[5],ymm3[6,7,8],ymm4[9],ymm3[10,11,12],ymm4[13],ymm3[14,15] ; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0],ymm2[1],ymm3[2],ymm2[3],ymm3[4],ymm2[5],ymm3[6],ymm2[7] ; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload ; AVX2-FP-NEXT: vpshufb %ymm7, %ymm3, %ymm3 ; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload ; AVX2-FP-NEXT: vpshufb %ymm6, %ymm4, %ymm4 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm3 = ymm4[0,1,2],ymm3[3],ymm4[4,5,6],ymm3[7],ymm4[8,9,10],ymm3[11],ymm4[12,13,14],ymm3[15] ; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload ; AVX2-FP-NEXT: vpshufb %ymm5, %ymm4, %ymm4 ; AVX2-FP-NEXT: vpshufb %xmm8, %xmm11, %xmm5 ; AVX2-FP-NEXT: vpmovzxwq {{.*#+}} xmm6 = xmm11[0],zero,zero,zero,xmm11[1],zero,zero,zero ; AVX2-FP-NEXT: vinserti128 $1, %xmm5, %ymm6, %ymm5 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm4 = ymm5[0],ymm4[1],ymm5[2,3,4],ymm4[5],ymm5[6,7,8],ymm4[9],ymm5[10,11,12],ymm4[13],ymm5[14,15] ; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm3 = ymm4[0],ymm3[1],ymm4[2],ymm3[3],ymm4[4],ymm3[5],ymm4[6],ymm3[7] ; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX2-FP-NEXT: vmovdqa %ymm3, 64(%rax) ; AVX2-FP-NEXT: vmovdqa %ymm2, 128(%rax) ; AVX2-FP-NEXT: vmovdqa %ymm1, 192(%rax) ; AVX2-FP-NEXT: vmovdqa %ymm0, 224(%rax) ; AVX2-FP-NEXT: vmovdqa %ymm15, 160(%rax) ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload ; AVX2-FP-NEXT: vmovaps %ymm0, 96(%rax) ; AVX2-FP-NEXT: vmovups (%rsp), %ymm0 # 32-byte Reload ; AVX2-FP-NEXT: vmovaps %ymm0, 32(%rax) ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload ; AVX2-FP-NEXT: vmovaps %ymm0, (%rax) ; AVX2-FP-NEXT: addq $72, %rsp ; AVX2-FP-NEXT: vzeroupper ; AVX2-FP-NEXT: retq ; ; AVX2-FCP-LABEL: store_i8_stride8_vf32: ; AVX2-FCP: # %bb.0: ; AVX2-FCP-NEXT: subq $72, %rsp ; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX2-FCP-NEXT: vmovdqa (%r10), %xmm5 ; AVX2-FCP-NEXT: vmovdqa (%rax), %xmm6 ; AVX2-FCP-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3],xmm6[4],xmm5[4],xmm6[5],xmm5[5],xmm6[6],xmm5[6],xmm6[7],xmm5[7] ; AVX2-FCP-NEXT: vinserti128 $1, %xmm1, %ymm1, %ymm8 ; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm3 = ymm8[0,1,0,1,0,1,0,1,8,9,10,11,4,5,2,3,16,17,20,21,16,17,20,21,24,25,26,27,20,21,22,23] ; AVX2-FCP-NEXT: vmovdqa (%r9), %xmm1 ; AVX2-FCP-NEXT: vmovdqa (%r8), %xmm2 ; AVX2-FCP-NEXT: vpunpcklbw {{.*#+}} xmm4 = 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: vinserti128 $1, %xmm4, %ymm4, %ymm9 ; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm4 = ymm9[0,1,2,3,0,1,2,3,8,9,10,11,2,3,6,7,20,21,18,19,20,21,18,19,24,25,26,27,22,23,22,23] ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm11 = ymm4[0,1,2],ymm3[3],ymm4[4,5,6],ymm3[7],ymm4[8,9,10],ymm3[11],ymm4[12,13,14],ymm3[15] ; AVX2-FCP-NEXT: vmovdqa (%rcx), %xmm7 ; AVX2-FCP-NEXT: vmovdqa (%rsi), %xmm3 ; AVX2-FCP-NEXT: vmovdqa (%rdi), %xmm10 ; AVX2-FCP-NEXT: vpunpcklbw {{.*#+}} xmm12 = xmm10[0],xmm3[0],xmm10[1],xmm3[1],xmm10[2],xmm3[2],xmm10[3],xmm3[3],xmm10[4],xmm3[4],xmm10[5],xmm3[5],xmm10[6],xmm3[6],xmm10[7],xmm3[7] ; AVX2-FCP-NEXT: vpshufb {{.*#+}} xmm4 = xmm12[4,5,u,u,u,u,u,u,6,7,u,u,u,u,u,u] ; AVX2-FCP-NEXT: vpmovzxwq {{.*#+}} xmm13 = xmm12[0],zero,zero,zero,xmm12[1],zero,zero,zero ; AVX2-FCP-NEXT: vinserti128 $1, %xmm4, %ymm13, %ymm13 ; AVX2-FCP-NEXT: vmovdqa (%rdx), %xmm15 ; AVX2-FCP-NEXT: vpunpcklbw {{.*#+}} xmm4 = xmm15[0],xmm7[0],xmm15[1],xmm7[1],xmm15[2],xmm7[2],xmm15[3],xmm7[3],xmm15[4],xmm7[4],xmm15[5],xmm7[5],xmm15[6],xmm7[6],xmm15[7],xmm7[7] ; AVX2-FCP-NEXT: vinserti128 $1, %xmm4, %ymm4, %ymm0 ; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm14 = ymm0[0,1,0,1,4,5,2,3,4,5,2,3,12,13,14,15,16,17,20,21,20,21,22,23,20,21,22,23,28,29,30,31] ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm13 = ymm13[0],ymm14[1],ymm13[2,3,4],ymm14[5],ymm13[6,7,8],ymm14[9],ymm13[10,11,12],ymm14[13],ymm13[14,15] ; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm13[0],ymm11[1],ymm13[2],ymm11[3],ymm13[4],ymm11[5],ymm13[6],ymm11[7] ; AVX2-FCP-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm14 = [0,1,2,3,8,9,8,9,8,9,8,9,12,13,10,11,0,1,2,3,8,9,12,13,8,9,12,13,12,13,14,15] ; AVX2-FCP-NEXT: vpshufb %ymm14, %ymm8, %ymm8 ; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm13 = [0,1,2,3,8,9,10,11,8,9,10,11,10,11,14,15,0,1,2,3,12,13,10,11,12,13,10,11,14,15,14,15] ; AVX2-FCP-NEXT: vpshufb %ymm13, %ymm9, %ymm9 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm8 = ymm9[0,1,2],ymm8[3],ymm9[4,5,6],ymm8[7],ymm9[8,9,10],ymm8[11],ymm9[12,13,14],ymm8[15] ; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm4 = [8,9,8,9,4,5,6,7,12,13,10,11,12,13,10,11,8,9,12,13,4,5,6,7,12,13,14,15,12,13,14,15] ; AVX2-FCP-NEXT: vpshufb %ymm4, %ymm0, %ymm0 ; AVX2-FCP-NEXT: vinserti128 $1, %xmm12, %ymm12, %ymm9 ; AVX2-FCP-NEXT: vpmovsxwq {{.*#+}} ymm12 = [2312,2826,3340,3854] ; AVX2-FCP-NEXT: vpshufb %ymm12, %ymm9, %ymm9 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm9[0],ymm0[1],ymm9[2,3,4],ymm0[5],ymm9[6,7,8],ymm0[9],ymm9[10,11,12],ymm0[13],ymm9[14,15] ; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0],ymm8[1],ymm0[2],ymm8[3],ymm0[4],ymm8[5],ymm0[6],ymm8[7] ; AVX2-FCP-NEXT: vmovdqu %ymm0, (%rsp) # 32-byte Spill ; AVX2-FCP-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm6[8],xmm5[8],xmm6[9],xmm5[9],xmm6[10],xmm5[10],xmm6[11],xmm5[11],xmm6[12],xmm5[12],xmm6[13],xmm5[13],xmm6[14],xmm5[14],xmm6[15],xmm5[15] ; AVX2-FCP-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] ; AVX2-FCP-NEXT: vinserti128 $1, %xmm0, %ymm0, %ymm0 ; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FCP-NEXT: vinserti128 $1, %xmm1, %ymm1, %ymm1 ; AVX2-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FCP-NEXT: vpshufb %ymm14, %ymm0, %ymm0 ; AVX2-FCP-NEXT: vpshufb %ymm13, %ymm1, %ymm1 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3],ymm1[4,5,6],ymm0[7],ymm1[8,9,10],ymm0[11],ymm1[12,13,14],ymm0[15] ; AVX2-FCP-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm15[8],xmm7[8],xmm15[9],xmm7[9],xmm15[10],xmm7[10],xmm15[11],xmm7[11],xmm15[12],xmm7[12],xmm15[13],xmm7[13],xmm15[14],xmm7[14],xmm15[15],xmm7[15] ; AVX2-FCP-NEXT: vinserti128 $1, %xmm1, %ymm1, %ymm2 ; AVX2-FCP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FCP-NEXT: vpunpckhbw {{.*#+}} xmm11 = xmm10[8],xmm3[8],xmm10[9],xmm3[9],xmm10[10],xmm3[10],xmm10[11],xmm3[11],xmm10[12],xmm3[12],xmm10[13],xmm3[13],xmm10[14],xmm3[14],xmm10[15],xmm3[15] ; AVX2-FCP-NEXT: vinserti128 $1, %xmm11, %ymm11, %ymm1 ; AVX2-FCP-NEXT: vpshufb %ymm12, %ymm1, %ymm1 ; AVX2-FCP-NEXT: vpshufb %ymm4, %ymm2, %ymm2 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3,4],ymm2[5],ymm1[6,7,8],ymm2[9],ymm1[10,11,12],ymm2[13],ymm1[14,15] ; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[2],ymm0[3],ymm1[4],ymm0[5],ymm1[6],ymm0[7] ; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FCP-NEXT: vmovdqa 16(%r10), %xmm8 ; AVX2-FCP-NEXT: vmovdqa 16(%rax), %xmm6 ; AVX2-FCP-NEXT: vpunpcklbw {{.*#+}} xmm1 = 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] ; AVX2-FCP-NEXT: vinserti128 $1, %xmm1, %ymm1, %ymm12 ; AVX2-FCP-NEXT: vmovdqa 16(%r9), %xmm7 ; AVX2-FCP-NEXT: vmovdqa 16(%r8), %xmm5 ; AVX2-FCP-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm5[0],xmm7[0],xmm5[1],xmm7[1],xmm5[2],xmm7[2],xmm5[3],xmm7[3],xmm5[4],xmm7[4],xmm5[5],xmm7[5],xmm5[6],xmm7[6],xmm5[7],xmm7[7] ; AVX2-FCP-NEXT: vinserti128 $1, %xmm3, %ymm3, %ymm10 ; AVX2-FCP-NEXT: vpshufb %ymm14, %ymm12, %ymm3 ; AVX2-FCP-NEXT: vpshufb %ymm13, %ymm10, %ymm4 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm15 = ymm4[0,1,2],ymm3[3],ymm4[4,5,6],ymm3[7],ymm4[8,9,10],ymm3[11],ymm4[12,13,14],ymm3[15] ; AVX2-FCP-NEXT: vmovdqa 16(%rcx), %xmm4 ; AVX2-FCP-NEXT: vmovdqa 16(%rdx), %xmm2 ; AVX2-FCP-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3],xmm2[4],xmm4[4],xmm2[5],xmm4[5],xmm2[6],xmm4[6],xmm2[7],xmm4[7] ; AVX2-FCP-NEXT: vinserti128 $1, %xmm3, %ymm3, %ymm9 ; AVX2-FCP-NEXT: vmovdqa 16(%rsi), %xmm1 ; AVX2-FCP-NEXT: vmovdqa 16(%rdi), %xmm0 ; AVX2-FCP-NEXT: vpunpcklbw {{.*#+}} xmm3 = 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] ; AVX2-FCP-NEXT: vinserti128 $1, %xmm3, %ymm3, %ymm14 ; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm14 = ymm14[8,9,u,u,u,u,u,u,10,11,u,u,u,u,u,u,28,29,u,u,u,u,u,u,30,31,u,u,u,u,u,u] ; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm13 = ymm9[8,9,8,9,4,5,6,7,12,13,10,11,12,13,10,11,24,25,28,29,20,21,22,23,28,29,30,31,28,29,30,31] ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm13 = ymm14[0],ymm13[1],ymm14[2,3,4],ymm13[5],ymm14[6,7,8],ymm13[9],ymm14[10,11,12],ymm13[13],ymm14[14,15] ; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm15 = ymm13[0],ymm15[1],ymm13[2],ymm15[3],ymm13[4],ymm15[5],ymm13[6],ymm15[7] ; AVX2-FCP-NEXT: vpunpckhbw {{.*#+}} xmm6 = xmm6[8],xmm8[8],xmm6[9],xmm8[9],xmm6[10],xmm8[10],xmm6[11],xmm8[11],xmm6[12],xmm8[12],xmm6[13],xmm8[13],xmm6[14],xmm8[14],xmm6[15],xmm8[15] ; AVX2-FCP-NEXT: vpunpckhbw {{.*#+}} xmm5 = xmm5[8],xmm7[8],xmm5[9],xmm7[9],xmm5[10],xmm7[10],xmm5[11],xmm7[11],xmm5[12],xmm7[12],xmm5[13],xmm7[13],xmm5[14],xmm7[14],xmm5[15],xmm7[15] ; AVX2-FCP-NEXT: vinserti128 $1, %xmm6, %ymm6, %ymm6 ; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm7 = ymm6[0,1,2,3,8,9,8,9,8,9,8,9,12,13,10,11,16,17,18,19,24,25,28,29,24,25,28,29,28,29,30,31] ; AVX2-FCP-NEXT: vinserti128 $1, %xmm5, %ymm5, %ymm5 ; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm8 = ymm5[0,1,2,3,8,9,10,11,8,9,10,11,10,11,14,15,16,17,18,19,28,29,26,27,28,29,26,27,30,31,30,31] ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm7 = ymm8[0,1,2],ymm7[3],ymm8[4,5,6],ymm7[7],ymm8[8,9,10],ymm7[11],ymm8[12,13,14],ymm7[15] ; AVX2-FCP-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm2[8],xmm4[8],xmm2[9],xmm4[9],xmm2[10],xmm4[10],xmm2[11],xmm4[11],xmm2[12],xmm4[12],xmm2[13],xmm4[13],xmm2[14],xmm4[14],xmm2[15],xmm4[15] ; AVX2-FCP-NEXT: vpunpckhbw {{.*#+}} xmm1 = 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] ; AVX2-FCP-NEXT: vinserti128 $1, %xmm1, %ymm1, %ymm0 ; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[8,9,u,u,u,u,u,u,10,11,u,u,u,u,u,u,28,29,u,u,u,u,u,u,30,31,u,u,u,u,u,u] ; AVX2-FCP-NEXT: vinserti128 $1, %xmm2, %ymm2, %ymm2 ; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm4 = ymm2[8,9,8,9,4,5,6,7,12,13,10,11,12,13,10,11,24,25,28,29,20,21,22,23,28,29,30,31,28,29,30,31] ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm4[1],ymm0[2,3,4],ymm4[5],ymm0[6,7,8],ymm4[9],ymm0[10,11,12],ymm4[13],ymm0[14,15] ; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0],ymm7[1],ymm0[2],ymm7[3],ymm0[4],ymm7[5],ymm0[6],ymm7[7] ; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm7 = [0,1,0,1,0,1,0,1,8,9,10,11,4,5,2,3,0,1,4,5,0,1,4,5,8,9,10,11,4,5,6,7] ; AVX2-FCP-NEXT: vpshufb %ymm7, %ymm6, %ymm4 ; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm6 = [0,1,2,3,0,1,2,3,8,9,10,11,2,3,6,7,4,5,2,3,4,5,2,3,8,9,10,11,6,7,6,7] ; AVX2-FCP-NEXT: vpshufb %ymm6, %ymm5, %ymm5 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm4 = ymm5[0,1,2],ymm4[3],ymm5[4,5,6],ymm4[7],ymm5[8,9,10],ymm4[11],ymm5[12,13,14],ymm4[15] ; AVX2-FCP-NEXT: vpmovsxwq {{.*#+}} xmm8 = [1284,1798] ; AVX2-FCP-NEXT: vpshufb %xmm8, %xmm1, %xmm5 ; AVX2-FCP-NEXT: vpmovzxwq {{.*#+}} xmm1 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero ; AVX2-FCP-NEXT: vinserti128 $1, %xmm5, %ymm1, %ymm1 ; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm5 = [0,1,0,1,4,5,2,3,4,5,2,3,12,13,14,15,0,1,4,5,4,5,6,7,4,5,6,7,12,13,14,15] ; AVX2-FCP-NEXT: vpshufb %ymm5, %ymm2, %ymm2 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3,4],ymm2[5],ymm1[6,7,8],ymm2[9],ymm1[10,11,12],ymm2[13],ymm1[14,15] ; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0],ymm4[1],ymm1[2],ymm4[3],ymm1[4],ymm4[5],ymm1[6],ymm4[7] ; AVX2-FCP-NEXT: vpshufb %ymm7, %ymm12, %ymm2 ; AVX2-FCP-NEXT: vpshufb %ymm6, %ymm10, %ymm4 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm4[0,1,2],ymm2[3],ymm4[4,5,6],ymm2[7],ymm4[8,9,10],ymm2[11],ymm4[12,13,14],ymm2[15] ; AVX2-FCP-NEXT: vpshufb %xmm8, %xmm3, %xmm4 ; AVX2-FCP-NEXT: vpmovzxwq {{.*#+}} xmm3 = xmm3[0],zero,zero,zero,xmm3[1],zero,zero,zero ; AVX2-FCP-NEXT: vinserti128 $1, %xmm4, %ymm3, %ymm3 ; AVX2-FCP-NEXT: vpshufb %ymm5, %ymm9, %ymm4 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0],ymm4[1],ymm3[2,3,4],ymm4[5],ymm3[6,7,8],ymm4[9],ymm3[10,11,12],ymm4[13],ymm3[14,15] ; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0],ymm2[1],ymm3[2],ymm2[3],ymm3[4],ymm2[5],ymm3[6],ymm2[7] ; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload ; AVX2-FCP-NEXT: vpshufb %ymm7, %ymm3, %ymm3 ; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload ; AVX2-FCP-NEXT: vpshufb %ymm6, %ymm4, %ymm4 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm4[0,1,2],ymm3[3],ymm4[4,5,6],ymm3[7],ymm4[8,9,10],ymm3[11],ymm4[12,13,14],ymm3[15] ; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload ; AVX2-FCP-NEXT: vpshufb %ymm5, %ymm4, %ymm4 ; AVX2-FCP-NEXT: vpshufb %xmm8, %xmm11, %xmm5 ; AVX2-FCP-NEXT: vpmovzxwq {{.*#+}} xmm6 = xmm11[0],zero,zero,zero,xmm11[1],zero,zero,zero ; AVX2-FCP-NEXT: vinserti128 $1, %xmm5, %ymm6, %ymm5 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm4 = ymm5[0],ymm4[1],ymm5[2,3,4],ymm4[5],ymm5[6,7,8],ymm4[9],ymm5[10,11,12],ymm4[13],ymm5[14,15] ; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm3 = ymm4[0],ymm3[1],ymm4[2],ymm3[3],ymm4[4],ymm3[5],ymm4[6],ymm3[7] ; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX2-FCP-NEXT: vmovdqa %ymm3, 64(%rax) ; AVX2-FCP-NEXT: vmovdqa %ymm2, 128(%rax) ; AVX2-FCP-NEXT: vmovdqa %ymm1, 192(%rax) ; AVX2-FCP-NEXT: vmovdqa %ymm0, 224(%rax) ; AVX2-FCP-NEXT: vmovdqa %ymm15, 160(%rax) ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload ; AVX2-FCP-NEXT: vmovaps %ymm0, 96(%rax) ; AVX2-FCP-NEXT: vmovups (%rsp), %ymm0 # 32-byte Reload ; AVX2-FCP-NEXT: vmovaps %ymm0, 32(%rax) ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload ; AVX2-FCP-NEXT: vmovaps %ymm0, (%rax) ; AVX2-FCP-NEXT: addq $72, %rsp ; AVX2-FCP-NEXT: vzeroupper ; AVX2-FCP-NEXT: retq ; ; AVX512-LABEL: store_i8_stride8_vf32: ; AVX512: # %bb.0: ; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX512-NEXT: vmovdqa (%r10), %xmm1 ; AVX512-NEXT: vmovdqa 16(%r10), %xmm12 ; AVX512-NEXT: vmovdqa (%rax), %xmm5 ; AVX512-NEXT: vmovdqa 16(%rax), %xmm13 ; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm5[8],xmm1[8],xmm5[9],xmm1[9],xmm5[10],xmm1[10],xmm5[11],xmm1[11],xmm5[12],xmm1[12],xmm5[13],xmm1[13],xmm5[14],xmm1[14],xmm5[15],xmm1[15] ; AVX512-NEXT: vmovdqa64 %xmm1, %xmm22 ; AVX512-NEXT: vpshufhw {{.*#+}} xmm1 = xmm0[0,1,2,3,4,4,6,5] ; AVX512-NEXT: vpshufhw {{.*#+}} xmm2 = xmm0[0,1,2,3,4,6,6,7] ; AVX512-NEXT: vinserti128 $1, %xmm2, %ymm1, %ymm1 ; AVX512-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[0,2,2,3,4,6,6,7] ; AVX512-NEXT: vmovdqa (%r9), %xmm6 ; AVX512-NEXT: vmovdqa 16(%r9), %xmm14 ; AVX512-NEXT: vmovdqa (%r8), %xmm7 ; AVX512-NEXT: vmovdqa 16(%r8), %xmm15 ; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm2 = 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-NEXT: vpshufhw {{.*#+}} xmm4 = xmm2[0,1,2,3,4,5,5,7] ; AVX512-NEXT: vpshufhw {{.*#+}} xmm8 = xmm2[0,1,2,3,6,5,7,7] ; AVX512-NEXT: vinserti128 $1, %xmm8, %ymm4, %ymm4 ; AVX512-NEXT: vpshufd {{.*#+}} ymm4 = ymm4[0,2,2,3,4,6,6,7] ; AVX512-NEXT: vpblendw {{.*#+}} ymm1 = ymm4[0,1,2],ymm1[3],ymm4[4,5,6],ymm1[7],ymm4[8,9,10],ymm1[11],ymm4[12,13,14],ymm1[15] ; AVX512-NEXT: vpshuflw {{.*#+}} xmm4 = xmm0[0,0,2,1,4,5,6,7] ; AVX512-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[0,2,2,3,4,5,6,7] ; AVX512-NEXT: vinserti128 $1, %xmm0, %ymm4, %ymm0 ; AVX512-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[0,0,2,1,4,4,6,5] ; AVX512-NEXT: vpshuflw {{.*#+}} xmm4 = xmm2[0,1,1,3,4,5,6,7] ; AVX512-NEXT: vpshuflw {{.*#+}} xmm2 = xmm2[2,1,3,3,4,5,6,7] ; AVX512-NEXT: vinserti128 $1, %xmm2, %ymm4, %ymm2 ; AVX512-NEXT: vpshufd {{.*#+}} ymm2 = ymm2[0,0,2,1,4,4,6,5] ; AVX512-NEXT: vpblendw {{.*#+}} ymm0 = ymm2[0,1,2],ymm0[3],ymm2[4,5,6],ymm0[7],ymm2[8,9,10],ymm0[11],ymm2[12,13,14],ymm0[15] ; AVX512-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm17 ; AVX512-NEXT: vmovdqa (%rsi), %xmm1 ; AVX512-NEXT: vmovdqa (%rdi), %xmm2 ; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm2[8],xmm1[8],xmm2[9],xmm1[9],xmm2[10],xmm1[10],xmm2[11],xmm1[11],xmm2[12],xmm1[12],xmm2[13],xmm1[13],xmm2[14],xmm1[14],xmm2[15],xmm1[15] ; AVX512-NEXT: vmovdqa64 %xmm2, %xmm20 ; AVX512-NEXT: vmovdqa64 %xmm1, %xmm21 ; AVX512-NEXT: vmovdqa (%rcx), %xmm8 ; AVX512-NEXT: vmovdqa (%rdx), %xmm9 ; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm9[8],xmm8[8],xmm9[9],xmm8[9],xmm9[10],xmm8[10],xmm9[11],xmm8[11],xmm9[12],xmm8[12],xmm9[13],xmm8[13],xmm9[14],xmm8[14],xmm9[15],xmm8[15] ; AVX512-NEXT: vpshufhw {{.*#+}} xmm4 = xmm1[0,1,2,3,4,4,6,5] ; AVX512-NEXT: vpshufhw {{.*#+}} xmm11 = xmm1[0,1,2,3,4,6,6,7] ; AVX512-NEXT: vinserti128 $1, %xmm11, %ymm4, %ymm4 ; AVX512-NEXT: vpmovzxwq {{.*#+}} ymm11 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero ; AVX512-NEXT: vmovdqa {{.*#+}} xmm3 = [0,1,0,1,4,5,2,3,0,1,4,5,4,5,6,7] ; AVX512-NEXT: vpshufb %xmm3, %xmm1, %xmm1 ; AVX512-NEXT: vpmovzxdq {{.*#+}} ymm1 = xmm1[0],zero,xmm1[1],zero,xmm1[2],zero,xmm1[3],zero ; AVX512-NEXT: vpblendw {{.*#+}} ymm1 = ymm11[0],ymm1[1],ymm11[2,3,4],ymm1[5],ymm11[6,7,8],ymm1[9],ymm11[10,11,12],ymm1[13],ymm11[14,15] ; AVX512-NEXT: vpshufd {{.*#+}} ymm4 = ymm4[2,1,3,3,6,5,7,7] ; AVX512-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,3,2,3] ; AVX512-NEXT: vpmovzxwq {{.*#+}} ymm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero ; AVX512-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm4[1],ymm0[2,3,4],ymm4[5],ymm0[6,7,8],ymm4[9],ymm0[10,11,12],ymm4[13],ymm0[14,15] ; AVX512-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm19 ; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm0 = 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-NEXT: vpshufhw {{.*#+}} xmm1 = xmm0[0,1,2,3,4,4,6,5] ; AVX512-NEXT: vpshufhw {{.*#+}} xmm4 = xmm0[0,1,2,3,4,6,6,7] ; AVX512-NEXT: vinserti128 $1, %xmm4, %ymm1, %ymm1 ; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm4 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3],xmm15[4],xmm14[4],xmm15[5],xmm14[5],xmm15[6],xmm14[6],xmm15[7],xmm14[7] ; AVX512-NEXT: vpshufhw {{.*#+}} xmm11 = xmm4[0,1,2,3,4,5,5,7] ; AVX512-NEXT: vpshufhw {{.*#+}} xmm10 = xmm4[0,1,2,3,6,5,7,7] ; AVX512-NEXT: vinserti128 $1, %xmm10, %ymm11, %ymm10 ; AVX512-NEXT: vpshuflw {{.*#+}} xmm11 = xmm0[0,0,2,1,4,5,6,7] ; AVX512-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[0,2,2,3,4,5,6,7] ; AVX512-NEXT: vinserti128 $1, %xmm0, %ymm11, %ymm0 ; AVX512-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[0,2,2,3,4,6,6,7] ; AVX512-NEXT: vpshufd {{.*#+}} ymm10 = ymm10[0,2,2,3,4,6,6,7] ; AVX512-NEXT: vpblendw {{.*#+}} ymm10 = ymm10[0,1,2],ymm1[3],ymm10[4,5,6],ymm1[7],ymm10[8,9,10],ymm1[11],ymm10[12,13,14],ymm1[15] ; AVX512-NEXT: vpshuflw {{.*#+}} xmm1 = xmm4[0,1,1,3,4,5,6,7] ; AVX512-NEXT: vpshuflw {{.*#+}} xmm4 = xmm4[2,1,3,3,4,5,6,7] ; AVX512-NEXT: vinserti128 $1, %xmm4, %ymm1, %ymm1 ; AVX512-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[0,0,2,1,4,4,6,5] ; AVX512-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[0,0,2,1,4,4,6,5] ; AVX512-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3],ymm1[4,5,6],ymm0[7],ymm1[8,9,10],ymm0[11],ymm1[12,13,14],ymm0[15] ; AVX512-NEXT: vmovdqa 16(%rcx), %xmm2 ; AVX512-NEXT: vinserti64x4 $1, %ymm10, %zmm0, %zmm18 ; AVX512-NEXT: vmovdqa 16(%rdx), %xmm11 ; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm11[0],xmm2[0],xmm11[1],xmm2[1],xmm11[2],xmm2[2],xmm11[3],xmm2[3],xmm11[4],xmm2[4],xmm11[5],xmm2[5],xmm11[6],xmm2[6],xmm11[7],xmm2[7] ; AVX512-NEXT: vpshufhw {{.*#+}} xmm4 = xmm0[0,1,2,3,4,4,6,5] ; AVX512-NEXT: vpshufhw {{.*#+}} xmm10 = xmm0[0,1,2,3,4,6,6,7] ; AVX512-NEXT: vinserti32x4 $1, %xmm10, %ymm4, %ymm16 ; AVX512-NEXT: vpshufb %xmm3, %xmm0, %xmm0 ; AVX512-NEXT: vmovdqa64 %xmm3, %xmm23 ; AVX512-NEXT: vpmovzxdq {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero ; AVX512-NEXT: vmovdqa 16(%rsi), %xmm4 ; AVX512-NEXT: vmovdqa 16(%rdi), %xmm1 ; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm10 = xmm1[0],xmm4[0],xmm1[1],xmm4[1],xmm1[2],xmm4[2],xmm1[3],xmm4[3],xmm1[4],xmm4[4],xmm1[5],xmm4[5],xmm1[6],xmm4[6],xmm1[7],xmm4[7] ; AVX512-NEXT: vpmovzxwq {{.*#+}} ymm3 = xmm10[0],zero,zero,zero,xmm10[1],zero,zero,zero,xmm10[2],zero,zero,zero,xmm10[3],zero,zero,zero ; AVX512-NEXT: vpblendw {{.*#+}} ymm0 = ymm3[0],ymm0[1],ymm3[2,3,4],ymm0[5],ymm3[6,7,8],ymm0[9],ymm3[10,11,12],ymm0[13],ymm3[14,15] ; AVX512-NEXT: vpshufd {{.*#+}} xmm3 = xmm10[2,3,2,3] ; AVX512-NEXT: vpshufd {{.*#+}} ymm10 = ymm16[2,1,3,3,6,5,7,7] ; AVX512-NEXT: vpmovzxwq {{.*#+}} ymm3 = xmm3[0],zero,zero,zero,xmm3[1],zero,zero,zero,xmm3[2],zero,zero,zero,xmm3[3],zero,zero,zero ; AVX512-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0],ymm10[1],ymm3[2,3,4],ymm10[5],ymm3[6,7,8],ymm10[9],ymm3[10,11,12],ymm10[13],ymm3[14,15] ; AVX512-NEXT: vinserti64x4 $1, %ymm3, %zmm0, %zmm16 ; AVX512-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] ; AVX512-NEXT: vpshufhw {{.*#+}} xmm3 = xmm0[0,1,2,3,4,4,6,5] ; AVX512-NEXT: vpshufhw {{.*#+}} xmm10 = xmm0[0,1,2,3,4,6,6,7] ; AVX512-NEXT: vinserti128 $1, %xmm10, %ymm3, %ymm3 ; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm10 = xmm15[8],xmm14[8],xmm15[9],xmm14[9],xmm15[10],xmm14[10],xmm15[11],xmm14[11],xmm15[12],xmm14[12],xmm15[13],xmm14[13],xmm15[14],xmm14[14],xmm15[15],xmm14[15] ; AVX512-NEXT: vpshufhw {{.*#+}} xmm12 = xmm10[0,1,2,3,4,5,5,7] ; AVX512-NEXT: vpshufhw {{.*#+}} xmm13 = xmm10[0,1,2,3,6,5,7,7] ; AVX512-NEXT: vinserti128 $1, %xmm13, %ymm12, %ymm12 ; AVX512-NEXT: vpshuflw {{.*#+}} xmm13 = xmm0[0,0,2,1,4,5,6,7] ; AVX512-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[0,2,2,3,4,5,6,7] ; AVX512-NEXT: vinserti128 $1, %xmm0, %ymm13, %ymm0 ; AVX512-NEXT: vpshufd {{.*#+}} ymm3 = ymm3[0,2,2,3,4,6,6,7] ; AVX512-NEXT: vpshufd {{.*#+}} ymm12 = ymm12[0,2,2,3,4,6,6,7] ; AVX512-NEXT: vpblendw {{.*#+}} ymm3 = ymm12[0,1,2],ymm3[3],ymm12[4,5,6],ymm3[7],ymm12[8,9,10],ymm3[11],ymm12[12,13,14],ymm3[15] ; AVX512-NEXT: vpshuflw {{.*#+}} xmm12 = xmm10[0,1,1,3,4,5,6,7] ; AVX512-NEXT: vpshuflw {{.*#+}} xmm10 = xmm10[2,1,3,3,4,5,6,7] ; AVX512-NEXT: vinserti128 $1, %xmm10, %ymm12, %ymm10 ; AVX512-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[0,0,2,1,4,4,6,5] ; AVX512-NEXT: vpshufd {{.*#+}} ymm10 = ymm10[0,0,2,1,4,4,6,5] ; AVX512-NEXT: vpblendw {{.*#+}} ymm0 = ymm10[0,1,2],ymm0[3],ymm10[4,5,6],ymm0[7],ymm10[8,9,10],ymm0[11],ymm10[12,13,14],ymm0[15] ; AVX512-NEXT: vinserti64x4 $1, %ymm3, %zmm0, %zmm12 ; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm11[8],xmm2[8],xmm11[9],xmm2[9],xmm11[10],xmm2[10],xmm11[11],xmm2[11],xmm11[12],xmm2[12],xmm11[13],xmm2[13],xmm11[14],xmm2[14],xmm11[15],xmm2[15] ; AVX512-NEXT: vpshufhw {{.*#+}} xmm2 = xmm0[0,1,2,3,4,4,6,5] ; AVX512-NEXT: vpshufhw {{.*#+}} xmm3 = xmm0[0,1,2,3,4,6,6,7] ; AVX512-NEXT: vinserti128 $1, %xmm3, %ymm2, %ymm2 ; AVX512-NEXT: vpunpckhbw {{.*#+}} 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] ; AVX512-NEXT: vpmovzxwq {{.*#+}} ymm3 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero,xmm1[2],zero,zero,zero,xmm1[3],zero,zero,zero ; AVX512-NEXT: vmovdqa64 %xmm23, %xmm10 ; AVX512-NEXT: vpshufb %xmm10, %xmm0, %xmm0 ; AVX512-NEXT: vpmovzxdq {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero ; AVX512-NEXT: vpblendw {{.*#+}} ymm0 = ymm3[0],ymm0[1],ymm3[2,3,4],ymm0[5],ymm3[6,7,8],ymm0[9],ymm3[10,11,12],ymm0[13],ymm3[14,15] ; AVX512-NEXT: vpshufd {{.*#+}} ymm2 = ymm2[2,1,3,3,6,5,7,7] ; AVX512-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[2,3,2,3] ; AVX512-NEXT: vpmovzxwq {{.*#+}} ymm1 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero,xmm1[2],zero,zero,zero,xmm1[3],zero,zero,zero ; AVX512-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3,4],ymm2[5],ymm1[6,7,8],ymm2[9],ymm1[10,11,12],ymm2[13],ymm1[14,15] ; AVX512-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0 ; AVX512-NEXT: vmovdqa64 %xmm22, %xmm1 ; AVX512-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-NEXT: vpshufhw {{.*#+}} xmm2 = xmm1[0,1,2,3,4,4,6,5] ; AVX512-NEXT: vpshufhw {{.*#+}} xmm3 = xmm1[0,1,2,3,4,6,6,7] ; AVX512-NEXT: vinserti128 $1, %xmm3, %ymm2, %ymm2 ; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm3 = 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: vpshufhw {{.*#+}} xmm4 = xmm3[0,1,2,3,4,5,5,7] ; AVX512-NEXT: vpshufhw {{.*#+}} xmm5 = xmm3[0,1,2,3,6,5,7,7] ; AVX512-NEXT: vinserti128 $1, %xmm5, %ymm4, %ymm4 ; AVX512-NEXT: vpshuflw {{.*#+}} xmm5 = xmm1[0,0,2,1,4,5,6,7] ; AVX512-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[0,2,2,3,4,5,6,7] ; AVX512-NEXT: vinserti128 $1, %xmm1, %ymm5, %ymm1 ; AVX512-NEXT: vpshufd {{.*#+}} ymm2 = ymm2[0,2,2,3,4,6,6,7] ; AVX512-NEXT: vpshufd {{.*#+}} ymm4 = ymm4[0,2,2,3,4,6,6,7] ; AVX512-NEXT: vpblendw {{.*#+}} ymm2 = ymm4[0,1,2],ymm2[3],ymm4[4,5,6],ymm2[7],ymm4[8,9,10],ymm2[11],ymm4[12,13,14],ymm2[15] ; AVX512-NEXT: vpshuflw {{.*#+}} xmm4 = xmm3[0,1,1,3,4,5,6,7] ; AVX512-NEXT: vpshuflw {{.*#+}} xmm3 = xmm3[2,1,3,3,4,5,6,7] ; AVX512-NEXT: vinserti128 $1, %xmm3, %ymm4, %ymm3 ; AVX512-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[0,0,2,1,4,4,6,5] ; AVX512-NEXT: vpshufd {{.*#+}} ymm3 = ymm3[0,0,2,1,4,4,6,5] ; AVX512-NEXT: vpblendw {{.*#+}} ymm1 = ymm3[0,1,2],ymm1[3],ymm3[4,5,6],ymm1[7],ymm3[8,9,10],ymm1[11],ymm3[12,13,14],ymm1[15] ; AVX512-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm1 ; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm2 = 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] ; AVX512-NEXT: vpshufhw {{.*#+}} xmm3 = xmm2[0,1,2,3,4,4,6,5] ; AVX512-NEXT: vpshufhw {{.*#+}} xmm4 = xmm2[0,1,2,3,4,6,6,7] ; AVX512-NEXT: vinserti128 $1, %xmm4, %ymm3, %ymm3 ; AVX512-NEXT: vmovdqa64 %xmm20, %xmm4 ; AVX512-NEXT: vmovdqa64 %xmm21, %xmm5 ; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm4 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3],xmm4[4],xmm5[4],xmm4[5],xmm5[5],xmm4[6],xmm5[6],xmm4[7],xmm5[7] ; AVX512-NEXT: vpshufb %xmm10, %xmm2, %xmm2 ; AVX512-NEXT: vpmovzxwq {{.*#+}} ymm5 = xmm4[0],zero,zero,zero,xmm4[1],zero,zero,zero,xmm4[2],zero,zero,zero,xmm4[3],zero,zero,zero ; AVX512-NEXT: vpmovzxdq {{.*#+}} ymm2 = xmm2[0],zero,xmm2[1],zero,xmm2[2],zero,xmm2[3],zero ; AVX512-NEXT: vpblendw {{.*#+}} ymm2 = ymm5[0],ymm2[1],ymm5[2,3,4],ymm2[5],ymm5[6,7,8],ymm2[9],ymm5[10,11,12],ymm2[13],ymm5[14,15] ; AVX512-NEXT: vpshufd {{.*#+}} ymm3 = ymm3[2,1,3,3,6,5,7,7] ; AVX512-NEXT: vpshufd {{.*#+}} xmm4 = xmm4[2,3,2,3] ; AVX512-NEXT: vpmovzxwq {{.*#+}} ymm4 = xmm4[0],zero,zero,zero,xmm4[1],zero,zero,zero,xmm4[2],zero,zero,zero,xmm4[3],zero,zero,zero ; AVX512-NEXT: vpblendw {{.*#+}} ymm3 = ymm4[0],ymm3[1],ymm4[2,3,4],ymm3[5],ymm4[6,7,8],ymm3[9],ymm4[10,11,12],ymm3[13],ymm4[14,15] ; AVX512-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm2 ; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512-NEXT: movw $-21846, %cx # imm = 0xAAAA ; AVX512-NEXT: kmovw %ecx, %k1 ; AVX512-NEXT: vmovdqa32 %zmm17, %zmm19 {%k1} ; AVX512-NEXT: vmovdqa32 %zmm18, %zmm16 {%k1} ; AVX512-NEXT: vmovdqa32 %zmm12, %zmm0 {%k1} ; AVX512-NEXT: vmovdqa32 %zmm1, %zmm2 {%k1} ; AVX512-NEXT: vmovdqa64 %zmm2, (%rax) ; AVX512-NEXT: vmovdqa64 %zmm0, 192(%rax) ; AVX512-NEXT: vmovdqa64 %zmm16, 128(%rax) ; AVX512-NEXT: vmovdqa64 %zmm19, 64(%rax) ; AVX512-NEXT: vzeroupper ; AVX512-NEXT: retq ; ; AVX512-FCP-LABEL: store_i8_stride8_vf32: ; AVX512-FCP: # %bb.0: ; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX512-FCP-NEXT: vmovdqa (%r10), %xmm1 ; AVX512-FCP-NEXT: vmovdqa (%rax), %xmm2 ; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm2[8],xmm1[8],xmm2[9],xmm1[9],xmm2[10],xmm1[10],xmm2[11],xmm1[11],xmm2[12],xmm1[12],xmm2[13],xmm1[13],xmm2[14],xmm1[14],xmm2[15],xmm1[15] ; AVX512-FCP-NEXT: vmovdqa64 %xmm2, %xmm21 ; AVX512-FCP-NEXT: vmovdqa64 %xmm1, %xmm22 ; AVX512-FCP-NEXT: vinserti128 $1, %xmm0, %ymm0, %ymm0 ; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm1 = ymm0[0,1,2,3,8,9,8,9,8,9,8,9,12,13,10,11,16,17,18,19,24,25,28,29,24,25,28,29,28,29,30,31] ; AVX512-FCP-NEXT: vmovdqa (%r9), %xmm3 ; AVX512-FCP-NEXT: vmovdqa (%r8), %xmm4 ; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm4[8],xmm3[8],xmm4[9],xmm3[9],xmm4[10],xmm3[10],xmm4[11],xmm3[11],xmm4[12],xmm3[12],xmm4[13],xmm3[13],xmm4[14],xmm3[14],xmm4[15],xmm3[15] ; AVX512-FCP-NEXT: vmovdqa64 %xmm4, %xmm24 ; AVX512-FCP-NEXT: vmovdqa64 %xmm3, %xmm25 ; AVX512-FCP-NEXT: vinserti128 $1, %xmm2, %ymm2, %ymm2 ; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm4 = [0,1,2,3,8,9,10,11,8,9,10,11,10,11,14,15,0,1,2,3,12,13,10,11,12,13,10,11,14,15,14,15] ; AVX512-FCP-NEXT: vpshufb %ymm4, %ymm2, %ymm3 ; AVX512-FCP-NEXT: vmovdqa64 %ymm4, %ymm20 ; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm3[0,1,2],ymm1[3],ymm3[4,5,6],ymm1[7],ymm3[8,9,10],ymm1[11],ymm3[12,13,14],ymm1[15] ; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm3 = [0,1,0,1,0,1,0,1,8,9,10,11,4,5,2,3,0,1,4,5,0,1,4,5,8,9,10,11,4,5,6,7] ; AVX512-FCP-NEXT: vpshufb %ymm3, %ymm0, %ymm0 ; AVX512-FCP-NEXT: vmovdqa64 %ymm3, %ymm26 ; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm3 = [0,1,2,3,0,1,2,3,8,9,10,11,2,3,6,7,4,5,2,3,4,5,2,3,8,9,10,11,6,7,6,7] ; AVX512-FCP-NEXT: vpshufb %ymm3, %ymm2, %ymm2 ; AVX512-FCP-NEXT: vmovdqa64 %ymm3, %ymm27 ; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm2[0,1,2],ymm0[3],ymm2[4,5,6],ymm0[7],ymm2[8,9,10],ymm0[11],ymm2[12,13,14],ymm0[15] ; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm23 ; AVX512-FCP-NEXT: vmovdqa (%rcx), %xmm1 ; AVX512-FCP-NEXT: vmovdqa (%rdx), %xmm10 ; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm10[8],xmm1[8],xmm10[9],xmm1[9],xmm10[10],xmm1[10],xmm10[11],xmm1[11],xmm10[12],xmm1[12],xmm10[13],xmm1[13],xmm10[14],xmm1[14],xmm10[15],xmm1[15] ; AVX512-FCP-NEXT: vmovdqa64 %xmm1, %xmm30 ; AVX512-FCP-NEXT: vinserti128 $1, %xmm0, %ymm0, %ymm0 ; AVX512-FCP-NEXT: vmovdqa (%rsi), %xmm12 ; AVX512-FCP-NEXT: vmovdqa (%rdi), %xmm13 ; 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: vinserti128 $1, %xmm1, %ymm1, %ymm2 ; AVX512-FCP-NEXT: vpmovsxwq {{.*#+}} ymm14 = [2312,2826,3340,3854] ; AVX512-FCP-NEXT: vpshufb %ymm14, %ymm2, %ymm2 ; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm4 = [8,9,8,9,4,5,6,7,12,13,10,11,12,13,10,11,8,9,12,13,4,5,6,7,12,13,14,15,12,13,14,15] ; AVX512-FCP-NEXT: vpshufb %ymm4, %ymm0, %ymm3 ; AVX512-FCP-NEXT: vmovdqa64 %ymm4, %ymm28 ; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0],ymm3[1],ymm2[2,3,4],ymm3[5],ymm2[6,7,8],ymm3[9],ymm2[10,11,12],ymm3[13],ymm2[14,15] ; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm3 = [0,1,0,1,4,5,2,3,4,5,2,3,12,13,14,15,0,1,4,5,4,5,6,7,4,5,6,7,12,13,14,15] ; AVX512-FCP-NEXT: vpshufb %ymm3, %ymm0, %ymm0 ; AVX512-FCP-NEXT: vmovdqa64 %ymm3, %ymm29 ; AVX512-FCP-NEXT: vpmovzxwq {{.*#+}} xmm3 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero ; AVX512-FCP-NEXT: vpmovsxwq {{.*#+}} xmm9 = [1284,1798] ; AVX512-FCP-NEXT: vpshufb %xmm9, %xmm1, %xmm1 ; AVX512-FCP-NEXT: vinserti128 $1, %xmm1, %ymm3, %ymm1 ; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[2,3,4],ymm0[5],ymm1[6,7,8],ymm0[9],ymm1[10,11,12],ymm0[13],ymm1[14,15] ; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm0, %zmm19 ; AVX512-FCP-NEXT: vmovdqa 16(%r10), %xmm8 ; AVX512-FCP-NEXT: vmovdqa 16(%rax), %xmm11 ; AVX512-FCP-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] ; AVX512-FCP-NEXT: vinserti128 $1, %xmm0, %ymm0, %ymm0 ; AVX512-FCP-NEXT: vmovdqa 16(%r9), %xmm7 ; AVX512-FCP-NEXT: vmovdqa 16(%r8), %xmm6 ; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm6[0],xmm7[0],xmm6[1],xmm7[1],xmm6[2],xmm7[2],xmm6[3],xmm7[3],xmm6[4],xmm7[4],xmm6[5],xmm7[5],xmm6[6],xmm7[6],xmm6[7],xmm7[7] ; AVX512-FCP-NEXT: vinserti128 $1, %xmm1, %ymm1, %ymm1 ; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm2 = [0,1,2,3,8,9,8,9,8,9,10,11,10,11,10,11,0,1,2,3,12,13,12,13,12,13,10,11,14,15,14,15] ; AVX512-FCP-NEXT: vpshufb %ymm2, %ymm0, %ymm3 ; AVX512-FCP-NEXT: vpshufb %ymm2, %ymm1, %ymm2 ; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0,1,2],ymm3[3],ymm2[4,5,6],ymm3[7],ymm2[8,9,10],ymm3[11],ymm2[12,13,14],ymm3[15] ; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm3 = [0,1,2,3,0,1,0,1,8,9,10,11,2,3,2,3,4,5,2,3,4,5,4,5,8,9,10,11,6,7,6,7] ; AVX512-FCP-NEXT: vpshufb %ymm3, %ymm0, %ymm0 ; AVX512-FCP-NEXT: vpshufb %ymm3, %ymm1, %ymm1 ; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3],ymm1[4,5,6],ymm0[7],ymm1[8,9,10],ymm0[11],ymm1[12,13,14],ymm0[15] ; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm0, %zmm17 ; AVX512-FCP-NEXT: vmovdqa 16(%rsi), %xmm5 ; AVX512-FCP-NEXT: vmovdqa 16(%rdi), %xmm4 ; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3],xmm4[4],xmm5[4],xmm4[5],xmm5[5],xmm4[6],xmm5[6],xmm4[7],xmm5[7] ; AVX512-FCP-NEXT: vinserti128 $1, %xmm0, %ymm0, %ymm1 ; AVX512-FCP-NEXT: vpshufb %ymm14, %ymm1, %ymm1 ; AVX512-FCP-NEXT: vmovdqa64 %ymm14, %ymm16 ; AVX512-FCP-NEXT: vmovdqa 16(%rcx), %xmm3 ; AVX512-FCP-NEXT: vmovdqa 16(%rdx), %xmm2 ; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm15 = 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-FCP-NEXT: vinserti128 $1, %xmm15, %ymm15, %ymm15 ; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm14 = ymm15[u,u,8,9,u,u,u,u,u,u,10,11,u,u,u,u,u,u,28,29,u,u,u,u,u,u,30,31,u,u,u,u] ; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0],ymm14[1],ymm1[2,3,4],ymm14[5],ymm1[6,7,8],ymm14[9],ymm1[10,11,12],ymm14[13],ymm1[14,15] ; AVX512-FCP-NEXT: vpmovzxwq {{.*#+}} xmm14 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero ; AVX512-FCP-NEXT: vpshufb %xmm9, %xmm0, %xmm0 ; AVX512-FCP-NEXT: vmovdqa64 %xmm9, %xmm31 ; AVX512-FCP-NEXT: vinserti128 $1, %xmm0, %ymm14, %ymm0 ; AVX512-FCP-NEXT: vpshufb {{.*#+}} ymm14 = ymm15[u,u,0,1,u,u,u,u,u,u,2,3,u,u,u,u,u,u,20,21,u,u,u,u,u,u,22,23,u,u,u,u] ; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm14[1],ymm0[2,3,4],ymm14[5],ymm0[6,7,8],ymm14[9],ymm0[10,11,12],ymm14[13],ymm0[14,15] ; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm18 ; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm11[8],xmm8[8],xmm11[9],xmm8[9],xmm11[10],xmm8[10],xmm11[11],xmm8[11],xmm11[12],xmm8[12],xmm11[13],xmm8[13],xmm11[14],xmm8[14],xmm11[15],xmm8[15] ; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm6[8],xmm7[8],xmm6[9],xmm7[9],xmm6[10],xmm7[10],xmm6[11],xmm7[11],xmm6[12],xmm7[12],xmm6[13],xmm7[13],xmm6[14],xmm7[14],xmm6[15],xmm7[15] ; AVX512-FCP-NEXT: vinserti128 $1, %xmm0, %ymm0, %ymm0 ; AVX512-FCP-NEXT: vinserti128 $1, %xmm1, %ymm1, %ymm1 ; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm8 = [0,1,2,3,8,9,8,9,8,9,8,9,12,13,10,11,0,1,2,3,8,9,12,13,8,9,12,13,12,13,14,15] ; AVX512-FCP-NEXT: vpshufb %ymm8, %ymm0, %ymm6 ; AVX512-FCP-NEXT: vmovdqa64 %ymm20, %ymm11 ; AVX512-FCP-NEXT: vpshufb %ymm11, %ymm1, %ymm7 ; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm6 = ymm7[0,1,2],ymm6[3],ymm7[4,5,6],ymm6[7],ymm7[8,9,10],ymm6[11],ymm7[12,13,14],ymm6[15] ; AVX512-FCP-NEXT: vmovdqa64 %ymm26, %ymm7 ; AVX512-FCP-NEXT: vpshufb %ymm7, %ymm0, %ymm0 ; AVX512-FCP-NEXT: vmovdqa64 %ymm27, %ymm14 ; AVX512-FCP-NEXT: vpshufb %ymm14, %ymm1, %ymm1 ; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3],ymm1[4,5,6],ymm0[7],ymm1[8,9,10],ymm0[11],ymm1[12,13,14],ymm0[15] ; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm6, %zmm0, %zmm20 ; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm1 = 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-FCP-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm4[8],xmm5[8],xmm4[9],xmm5[9],xmm4[10],xmm5[10],xmm4[11],xmm5[11],xmm4[12],xmm5[12],xmm4[13],xmm5[13],xmm4[14],xmm5[14],xmm4[15],xmm5[15] ; AVX512-FCP-NEXT: vinserti128 $1, %xmm1, %ymm1, %ymm1 ; AVX512-FCP-NEXT: vinserti128 $1, %xmm2, %ymm2, %ymm3 ; AVX512-FCP-NEXT: vmovdqa64 %ymm16, %ymm15 ; AVX512-FCP-NEXT: vpshufb %ymm15, %ymm3, %ymm3 ; AVX512-FCP-NEXT: vmovdqa64 %ymm28, %ymm6 ; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm1, %ymm4 ; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0],ymm4[1],ymm3[2,3,4],ymm4[5],ymm3[6,7,8],ymm4[9],ymm3[10,11,12],ymm4[13],ymm3[14,15] ; AVX512-FCP-NEXT: vmovdqa64 %ymm29, %ymm9 ; AVX512-FCP-NEXT: vpshufb %ymm9, %ymm1, %ymm1 ; AVX512-FCP-NEXT: vpmovzxwq {{.*#+}} xmm4 = xmm2[0],zero,zero,zero,xmm2[1],zero,zero,zero ; AVX512-FCP-NEXT: vmovdqa64 %xmm31, %xmm0 ; AVX512-FCP-NEXT: vpshufb %xmm0, %xmm2, %xmm2 ; AVX512-FCP-NEXT: vinserti128 $1, %xmm2, %ymm4, %ymm2 ; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2,3,4],ymm1[5],ymm2[6,7,8],ymm1[9],ymm2[10,11,12],ymm1[13],ymm2[14,15] ; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm3, %zmm1, %zmm1 ; AVX512-FCP-NEXT: vmovdqa64 %xmm21, %xmm2 ; AVX512-FCP-NEXT: vmovdqa64 %xmm22, %xmm3 ; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} 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] ; AVX512-FCP-NEXT: vmovdqa64 %xmm24, %xmm3 ; AVX512-FCP-NEXT: vmovdqa64 %xmm25, %xmm4 ; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} 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] ; AVX512-FCP-NEXT: vinserti128 $1, %xmm2, %ymm2, %ymm2 ; AVX512-FCP-NEXT: vpshufb %ymm8, %ymm2, %ymm4 ; AVX512-FCP-NEXT: vinserti128 $1, %xmm3, %ymm3, %ymm3 ; AVX512-FCP-NEXT: vpshufb %ymm11, %ymm3, %ymm5 ; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm4 = ymm5[0,1,2],ymm4[3],ymm5[4,5,6],ymm4[7],ymm5[8,9,10],ymm4[11],ymm5[12,13,14],ymm4[15] ; AVX512-FCP-NEXT: vpshufb %ymm7, %ymm2, %ymm2 ; AVX512-FCP-NEXT: vpshufb %ymm14, %ymm3, %ymm3 ; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm3[0,1,2],ymm2[3],ymm3[4,5,6],ymm2[7],ymm3[8,9,10],ymm2[11],ymm3[12,13,14],ymm2[15] ; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm2, %zmm2 ; AVX512-FCP-NEXT: vmovdqa64 %xmm30, %xmm3 ; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm10[0],xmm3[0],xmm10[1],xmm3[1],xmm10[2],xmm3[2],xmm10[3],xmm3[3],xmm10[4],xmm3[4],xmm10[5],xmm3[5],xmm10[6],xmm3[6],xmm10[7],xmm3[7] ; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm4 = 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: vinserti128 $1, %xmm4, %ymm4, %ymm5 ; AVX512-FCP-NEXT: vpshufb %ymm15, %ymm5, %ymm5 ; AVX512-FCP-NEXT: vinserti128 $1, %xmm3, %ymm3, %ymm3 ; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm3, %ymm6 ; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm5 = ymm5[0],ymm6[1],ymm5[2,3,4],ymm6[5],ymm5[6,7,8],ymm6[9],ymm5[10,11,12],ymm6[13],ymm5[14,15] ; AVX512-FCP-NEXT: vpshufb %ymm9, %ymm3, %ymm3 ; AVX512-FCP-NEXT: vpshufb %xmm0, %xmm4, %xmm6 ; AVX512-FCP-NEXT: vpmovzxwq {{.*#+}} xmm4 = xmm4[0],zero,zero,zero,xmm4[1],zero,zero,zero ; AVX512-FCP-NEXT: vinserti128 $1, %xmm6, %ymm4, %ymm4 ; AVX512-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm4[0],ymm3[1],ymm4[2,3,4],ymm3[5],ymm4[6,7,8],ymm3[9],ymm4[10,11,12],ymm3[13],ymm4[14,15] ; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm5, %zmm3, %zmm3 ; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512-FCP-NEXT: movw $-21846, %cx # imm = 0xAAAA ; AVX512-FCP-NEXT: kmovw %ecx, %k1 ; AVX512-FCP-NEXT: vmovdqa32 %zmm23, %zmm19 {%k1} ; AVX512-FCP-NEXT: vmovdqa32 %zmm17, %zmm18 {%k1} ; AVX512-FCP-NEXT: vmovdqa32 %zmm20, %zmm1 {%k1} ; AVX512-FCP-NEXT: vmovdqa32 %zmm2, %zmm3 {%k1} ; AVX512-FCP-NEXT: vmovdqa64 %zmm3, (%rax) ; AVX512-FCP-NEXT: vmovdqa64 %zmm1, 192(%rax) ; AVX512-FCP-NEXT: vmovdqa64 %zmm18, 128(%rax) ; AVX512-FCP-NEXT: vmovdqa64 %zmm19, 64(%rax) ; AVX512-FCP-NEXT: vzeroupper ; AVX512-FCP-NEXT: retq ; ; AVX512DQ-LABEL: store_i8_stride8_vf32: ; AVX512DQ: # %bb.0: ; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX512DQ-NEXT: vmovdqa (%r10), %xmm0 ; AVX512DQ-NEXT: vmovdqa 16(%r10), %xmm10 ; AVX512DQ-NEXT: vmovdqa (%rax), %xmm2 ; AVX512DQ-NEXT: vmovdqa 16(%rax), %xmm11 ; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm2[8],xmm0[8],xmm2[9],xmm0[9],xmm2[10],xmm0[10],xmm2[11],xmm0[11],xmm2[12],xmm0[12],xmm2[13],xmm0[13],xmm2[14],xmm0[14],xmm2[15],xmm0[15] ; AVX512DQ-NEXT: vmovdqa64 %xmm2, %xmm23 ; AVX512DQ-NEXT: vmovdqa64 %xmm0, %xmm24 ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm0 = xmm1[0,1,2,3,4,4,6,5] ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm4 = xmm1[0,1,2,3,4,6,6,7] ; AVX512DQ-NEXT: vinserti128 $1, %xmm4, %ymm0, %ymm0 ; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[0,2,2,3,4,6,6,7] ; AVX512DQ-NEXT: vmovdqa (%r9), %xmm2 ; AVX512DQ-NEXT: vmovdqa 16(%r9), %xmm12 ; AVX512DQ-NEXT: vmovdqa (%r8), %xmm6 ; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm5 = xmm6[8],xmm2[8],xmm6[9],xmm2[9],xmm6[10],xmm2[10],xmm6[11],xmm2[11],xmm6[12],xmm2[12],xmm6[13],xmm2[13],xmm6[14],xmm2[14],xmm6[15],xmm2[15] ; AVX512DQ-NEXT: vmovdqa64 %xmm2, %xmm25 ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm7 = xmm5[0,1,2,3,4,5,5,7] ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm8 = xmm5[0,1,2,3,6,5,7,7] ; AVX512DQ-NEXT: vinserti128 $1, %xmm8, %ymm7, %ymm7 ; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm7 = ymm7[0,2,2,3,4,6,6,7] ; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm2 = ymm7[0,1,2],ymm0[3],ymm7[4,5,6],ymm0[7],ymm7[8,9,10],ymm0[11],ymm7[12,13,14],ymm0[15] ; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm7 = xmm1[0,0,2,1,4,5,6,7] ; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[0,2,2,3,4,5,6,7] ; AVX512DQ-NEXT: vinserti128 $1, %xmm1, %ymm7, %ymm13 ; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm1 = xmm5[0,1,1,3,4,5,6,7] ; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm5 = xmm5[2,1,3,3,4,5,6,7] ; AVX512DQ-NEXT: vinserti128 $1, %xmm5, %ymm1, %ymm5 ; AVX512DQ-NEXT: vmovdqa (%rsi), %xmm0 ; AVX512DQ-NEXT: vmovdqa (%rdi), %xmm1 ; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm14 = 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: vmovdqa64 %xmm1, %xmm21 ; AVX512DQ-NEXT: vmovdqa64 %xmm0, %xmm22 ; AVX512DQ-NEXT: vmovdqa (%rcx), %xmm7 ; AVX512DQ-NEXT: vmovdqa (%rdx), %xmm8 ; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm15 = 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: vpshufhw {{.*#+}} xmm0 = xmm15[0,1,2,3,4,4,6,5] ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm1 = xmm15[0,1,2,3,4,6,6,7] ; AVX512DQ-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm0 ; AVX512DQ-NEXT: vpmovzxwq {{.*#+}} ymm1 = xmm14[0],zero,zero,zero,xmm14[1],zero,zero,zero,xmm14[2],zero,zero,zero,xmm14[3],zero,zero,zero ; AVX512DQ-NEXT: vmovdqa {{.*#+}} xmm4 = [0,1,0,1,4,5,2,3,0,1,4,5,4,5,6,7] ; AVX512DQ-NEXT: vpshufb %xmm4, %xmm15, %xmm15 ; AVX512DQ-NEXT: vpmovzxdq {{.*#+}} ymm15 = xmm15[0],zero,xmm15[1],zero,xmm15[2],zero,xmm15[3],zero ; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0],ymm15[1],ymm1[2,3,4],ymm15[5],ymm1[6,7,8],ymm15[9],ymm1[10,11,12],ymm15[13],ymm1[14,15] ; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[2,1,3,3,6,5,7,7] ; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm14 = xmm14[2,3,2,3] ; AVX512DQ-NEXT: vpmovzxwq {{.*#+}} ymm14 = xmm14[0],zero,zero,zero,xmm14[1],zero,zero,zero,xmm14[2],zero,zero,zero,xmm14[3],zero,zero,zero ; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm0 = ymm14[0],ymm0[1],ymm14[2,3,4],ymm0[5],ymm14[6,7,8],ymm0[9],ymm14[10,11,12],ymm0[13],ymm14[14,15] ; AVX512DQ-NEXT: vmovdqa 16(%r8), %xmm14 ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm16 ; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm0 = ymm13[0,0,2,1,4,4,6,5] ; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm1 = ymm5[0,0,2,1,4,4,6,5] ; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3],ymm1[4,5,6],ymm0[7],ymm1[8,9,10],ymm0[11],ymm1[12,13,14],ymm0[15] ; AVX512DQ-NEXT: movw $-21846, %ax # imm = 0xAAAA ; AVX512DQ-NEXT: kmovw %eax, %k1 ; AVX512DQ-NEXT: vinserti32x8 $1, %ymm2, %zmm0, %zmm16 {%k1} ; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm0 = 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-NEXT: vpshufhw {{.*#+}} xmm1 = xmm0[0,1,2,3,4,4,6,5] ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm2 = xmm0[0,1,2,3,4,6,6,7] ; AVX512DQ-NEXT: vinserti32x4 $1, %xmm2, %ymm1, %ymm20 ; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm2 = 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] ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm5 = xmm2[0,1,2,3,4,5,5,7] ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm13 = xmm2[0,1,2,3,6,5,7,7] ; AVX512DQ-NEXT: vinserti32x4 $1, %xmm13, %ymm5, %ymm18 ; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm5 = xmm0[0,0,2,1,4,5,6,7] ; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[0,2,2,3,4,5,6,7] ; AVX512DQ-NEXT: vinserti32x4 $1, %xmm0, %ymm5, %ymm19 ; AVX512DQ-NEXT: vmovdqa 16(%rcx), %xmm15 ; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm0 = xmm2[0,1,1,3,4,5,6,7] ; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm2 = xmm2[2,1,3,3,4,5,6,7] ; AVX512DQ-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm2 ; AVX512DQ-NEXT: vmovdqa 16(%rdx), %xmm3 ; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm5 = xmm3[0],xmm15[0],xmm3[1],xmm15[1],xmm3[2],xmm15[2],xmm3[3],xmm15[3],xmm3[4],xmm15[4],xmm3[5],xmm15[5],xmm3[6],xmm15[6],xmm3[7],xmm15[7] ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm13 = xmm5[0,1,2,3,4,4,6,5] ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm9 = xmm5[0,1,2,3,4,6,6,7] ; AVX512DQ-NEXT: vinserti128 $1, %xmm9, %ymm13, %ymm9 ; AVX512DQ-NEXT: vpshufb %xmm4, %xmm5, %xmm5 ; AVX512DQ-NEXT: vmovdqa64 %xmm4, %xmm26 ; AVX512DQ-NEXT: vpmovzxdq {{.*#+}} ymm1 = xmm5[0],zero,xmm5[1],zero,xmm5[2],zero,xmm5[3],zero ; AVX512DQ-NEXT: vmovdqa 16(%rsi), %xmm5 ; AVX512DQ-NEXT: vmovdqa 16(%rdi), %xmm13 ; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm13[0],xmm5[0],xmm13[1],xmm5[1],xmm13[2],xmm5[2],xmm13[3],xmm5[3],xmm13[4],xmm5[4],xmm13[5],xmm5[5],xmm13[6],xmm5[6],xmm13[7],xmm5[7] ; AVX512DQ-NEXT: vpmovzxwq {{.*#+}} ymm4 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero ; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm1 = ymm4[0],ymm1[1],ymm4[2,3,4],ymm1[5],ymm4[6,7,8],ymm1[9],ymm4[10,11,12],ymm1[13],ymm4[14,15] ; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,3,2,3] ; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm4 = ymm9[2,1,3,3,6,5,7,7] ; AVX512DQ-NEXT: vpmovzxwq {{.*#+}} ymm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero ; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm4[1],ymm0[2,3,4],ymm4[5],ymm0[6,7,8],ymm4[9],ymm0[10,11,12],ymm4[13],ymm0[14,15] ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm17 ; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm0 = ymm20[0,2,2,3,4,6,6,7] ; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm1 = ymm18[0,2,2,3,4,6,6,7] ; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3],ymm1[4,5,6],ymm0[7],ymm1[8,9,10],ymm0[11],ymm1[12,13,14],ymm0[15] ; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm1 = ymm19[0,0,2,1,4,4,6,5] ; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm2 = ymm2[0,0,2,1,4,4,6,5] ; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0,1,2],ymm1[3],ymm2[4,5,6],ymm1[7],ymm2[8,9,10],ymm1[11],ymm2[12,13,14],ymm1[15] ; AVX512DQ-NEXT: vinserti32x8 $1, %ymm0, %zmm1, %zmm17 {%k1} ; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm11[8],xmm10[8],xmm11[9],xmm10[9],xmm11[10],xmm10[10],xmm11[11],xmm10[11],xmm11[12],xmm10[12],xmm11[13],xmm10[13],xmm11[14],xmm10[14],xmm11[15],xmm10[15] ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm1 = xmm0[0,1,2,3,4,4,6,5] ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm2 = xmm0[0,1,2,3,4,6,6,7] ; AVX512DQ-NEXT: vinserti128 $1, %xmm2, %ymm1, %ymm1 ; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm2 = 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] ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm4 = xmm2[0,1,2,3,4,5,5,7] ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm9 = xmm2[0,1,2,3,6,5,7,7] ; AVX512DQ-NEXT: vinserti128 $1, %xmm9, %ymm4, %ymm4 ; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm9 = xmm0[0,0,2,1,4,5,6,7] ; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[0,2,2,3,4,5,6,7] ; AVX512DQ-NEXT: vinserti128 $1, %xmm0, %ymm9, %ymm0 ; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm9 = xmm2[0,1,1,3,4,5,6,7] ; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm2 = xmm2[2,1,3,3,4,5,6,7] ; AVX512DQ-NEXT: vinserti128 $1, %xmm2, %ymm9, %ymm2 ; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm3[8],xmm15[8],xmm3[9],xmm15[9],xmm3[10],xmm15[10],xmm3[11],xmm15[11],xmm3[12],xmm15[12],xmm3[13],xmm15[13],xmm3[14],xmm15[14],xmm3[15],xmm15[15] ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm9 = xmm3[0,1,2,3,4,4,6,5] ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm10 = xmm3[0,1,2,3,4,6,6,7] ; AVX512DQ-NEXT: vinserti128 $1, %xmm10, %ymm9, %ymm9 ; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[0,2,2,3,4,6,6,7] ; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm4 = ymm4[0,2,2,3,4,6,6,7] ; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm1 = ymm4[0,1,2],ymm1[3],ymm4[4,5,6],ymm1[7],ymm4[8,9,10],ymm1[11],ymm4[12,13,14],ymm1[15] ; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm4 = xmm13[8],xmm5[8],xmm13[9],xmm5[9],xmm13[10],xmm5[10],xmm13[11],xmm5[11],xmm13[12],xmm5[12],xmm13[13],xmm5[13],xmm13[14],xmm5[14],xmm13[15],xmm5[15] ; AVX512DQ-NEXT: vpmovzxwq {{.*#+}} ymm5 = xmm4[0],zero,zero,zero,xmm4[1],zero,zero,zero,xmm4[2],zero,zero,zero,xmm4[3],zero,zero,zero ; AVX512DQ-NEXT: vmovdqa64 %xmm26, %xmm10 ; AVX512DQ-NEXT: vpshufb %xmm10, %xmm3, %xmm3 ; AVX512DQ-NEXT: vpmovzxdq {{.*#+}} ymm3 = xmm3[0],zero,xmm3[1],zero,xmm3[2],zero,xmm3[3],zero ; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm3 = ymm5[0],ymm3[1],ymm5[2,3,4],ymm3[5],ymm5[6,7,8],ymm3[9],ymm5[10,11,12],ymm3[13],ymm5[14,15] ; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm5 = ymm9[2,1,3,3,6,5,7,7] ; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm4 = xmm4[2,3,2,3] ; AVX512DQ-NEXT: vpmovzxwq {{.*#+}} ymm4 = xmm4[0],zero,zero,zero,xmm4[1],zero,zero,zero,xmm4[2],zero,zero,zero,xmm4[3],zero,zero,zero ; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm4 = ymm4[0],ymm5[1],ymm4[2,3,4],ymm5[5],ymm4[6,7,8],ymm5[9],ymm4[10,11,12],ymm5[13],ymm4[14,15] ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm4, %zmm3, %zmm3 ; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[0,0,2,1,4,4,6,5] ; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm2 = ymm2[0,0,2,1,4,4,6,5] ; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm0 = ymm2[0,1,2],ymm0[3],ymm2[4,5,6],ymm0[7],ymm2[8,9,10],ymm0[11],ymm2[12,13,14],ymm0[15] ; AVX512DQ-NEXT: vinserti32x8 $1, %ymm1, %zmm0, %zmm3 {%k1} ; AVX512DQ-NEXT: vmovdqa64 %xmm23, %xmm0 ; AVX512DQ-NEXT: vmovdqa64 %xmm24, %xmm1 ; 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: vpshufhw {{.*#+}} xmm1 = xmm0[0,1,2,3,4,4,6,5] ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm2 = xmm0[0,1,2,3,4,6,6,7] ; AVX512DQ-NEXT: vinserti128 $1, %xmm2, %ymm1, %ymm1 ; AVX512DQ-NEXT: vmovdqa64 %xmm25, %xmm2 ; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm6[0],xmm2[0],xmm6[1],xmm2[1],xmm6[2],xmm2[2],xmm6[3],xmm2[3],xmm6[4],xmm2[4],xmm6[5],xmm2[5],xmm6[6],xmm2[6],xmm6[7],xmm2[7] ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm4 = xmm2[0,1,2,3,4,5,5,7] ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm5 = xmm2[0,1,2,3,6,5,7,7] ; AVX512DQ-NEXT: vinserti128 $1, %xmm5, %ymm4, %ymm4 ; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm5 = xmm0[0,0,2,1,4,5,6,7] ; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[0,2,2,3,4,5,6,7] ; AVX512DQ-NEXT: vinserti128 $1, %xmm0, %ymm5, %ymm0 ; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm5 = xmm2[0,1,1,3,4,5,6,7] ; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm2 = xmm2[2,1,3,3,4,5,6,7] ; AVX512DQ-NEXT: vinserti128 $1, %xmm2, %ymm5, %ymm2 ; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm5 = 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: vpshufhw {{.*#+}} xmm6 = xmm5[0,1,2,3,4,4,6,5] ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm7 = xmm5[0,1,2,3,4,6,6,7] ; AVX512DQ-NEXT: vinserti128 $1, %xmm7, %ymm6, %ymm6 ; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[0,2,2,3,4,6,6,7] ; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm4 = ymm4[0,2,2,3,4,6,6,7] ; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm1 = ymm4[0,1,2],ymm1[3],ymm4[4,5,6],ymm1[7],ymm4[8,9,10],ymm1[11],ymm4[12,13,14],ymm1[15] ; AVX512DQ-NEXT: vmovdqa64 %xmm21, %xmm4 ; AVX512DQ-NEXT: vmovdqa64 %xmm22, %xmm7 ; AVX512DQ-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] ; AVX512DQ-NEXT: vpshufb %xmm10, %xmm5, %xmm5 ; AVX512DQ-NEXT: vpmovzxwq {{.*#+}} ymm7 = xmm4[0],zero,zero,zero,xmm4[1],zero,zero,zero,xmm4[2],zero,zero,zero,xmm4[3],zero,zero,zero ; AVX512DQ-NEXT: vpmovzxdq {{.*#+}} ymm5 = xmm5[0],zero,xmm5[1],zero,xmm5[2],zero,xmm5[3],zero ; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm5 = ymm7[0],ymm5[1],ymm7[2,3,4],ymm5[5],ymm7[6,7,8],ymm5[9],ymm7[10,11,12],ymm5[13],ymm7[14,15] ; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm6 = ymm6[2,1,3,3,6,5,7,7] ; AVX512DQ-NEXT: vpshufd {{.*#+}} xmm4 = xmm4[2,3,2,3] ; AVX512DQ-NEXT: vpmovzxwq {{.*#+}} ymm4 = xmm4[0],zero,zero,zero,xmm4[1],zero,zero,zero,xmm4[2],zero,zero,zero,xmm4[3],zero,zero,zero ; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm4 = ymm4[0],ymm6[1],ymm4[2,3,4],ymm6[5],ymm4[6,7,8],ymm6[9],ymm4[10,11,12],ymm6[13],ymm4[14,15] ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm4, %zmm5, %zmm4 ; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[0,0,2,1,4,4,6,5] ; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm2 = ymm2[0,0,2,1,4,4,6,5] ; AVX512DQ-NEXT: vpblendw {{.*#+}} ymm0 = ymm2[0,1,2],ymm0[3],ymm2[4,5,6],ymm0[7],ymm2[8,9,10],ymm0[11],ymm2[12,13,14],ymm0[15] ; AVX512DQ-NEXT: vinserti32x8 $1, %ymm1, %zmm0, %zmm4 {%k1} ; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512DQ-NEXT: vmovdqa64 %zmm4, (%rax) ; AVX512DQ-NEXT: vmovdqa64 %zmm3, 192(%rax) ; AVX512DQ-NEXT: vmovdqa64 %zmm17, 128(%rax) ; AVX512DQ-NEXT: vmovdqa64 %zmm16, 64(%rax) ; AVX512DQ-NEXT: vzeroupper ; AVX512DQ-NEXT: retq ; ; AVX512DQ-FCP-LABEL: store_i8_stride8_vf32: ; AVX512DQ-FCP: # %bb.0: ; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX512DQ-FCP-NEXT: vmovdqa (%rcx), %xmm1 ; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %xmm2 ; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm2[8],xmm1[8],xmm2[9],xmm1[9],xmm2[10],xmm1[10],xmm2[11],xmm1[11],xmm2[12],xmm1[12],xmm2[13],xmm1[13],xmm2[14],xmm1[14],xmm2[15],xmm1[15] ; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm2, %xmm19 ; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm1, %xmm20 ; AVX512DQ-FCP-NEXT: vinserti128 $1, %xmm0, %ymm0, %ymm0 ; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm2 = [8,9,8,9,4,5,6,7,12,13,10,11,12,13,10,11,8,9,12,13,4,5,6,7,12,13,14,15,12,13,14,15] ; AVX512DQ-FCP-NEXT: vpshufb %ymm2, %ymm0, %ymm1 ; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm2, %ymm18 ; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %xmm3 ; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %xmm4 ; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm4[8],xmm3[8],xmm4[9],xmm3[9],xmm4[10],xmm3[10],xmm4[11],xmm3[11],xmm4[12],xmm3[12],xmm4[13],xmm3[13],xmm4[14],xmm3[14],xmm4[15],xmm3[15] ; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm4, %xmm21 ; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm3, %xmm22 ; AVX512DQ-FCP-NEXT: vinserti128 $1, %xmm2, %ymm2, %ymm3 ; AVX512DQ-FCP-NEXT: vpmovsxwq {{.*#+}} ymm11 = [2312,2826,3340,3854] ; AVX512DQ-FCP-NEXT: vpshufb %ymm11, %ymm3, %ymm3 ; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm3[0],ymm1[1],ymm3[2,3,4],ymm1[5],ymm3[6,7,8],ymm1[9],ymm3[10,11,12],ymm1[13],ymm3[14,15] ; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm3 = [0,1,0,1,4,5,2,3,4,5,2,3,12,13,14,15,0,1,4,5,4,5,6,7,4,5,6,7,12,13,14,15] ; AVX512DQ-FCP-NEXT: vpshufb %ymm3, %ymm0, %ymm0 ; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm3, %ymm23 ; AVX512DQ-FCP-NEXT: vpmovsxwq {{.*#+}} xmm8 = [1284,1798] ; AVX512DQ-FCP-NEXT: vpshufb %xmm8, %xmm2, %xmm3 ; AVX512DQ-FCP-NEXT: vpmovzxwq {{.*#+}} xmm2 = xmm2[0],zero,zero,zero,xmm2[1],zero,zero,zero ; AVX512DQ-FCP-NEXT: vinserti128 $1, %xmm3, %ymm2, %ymm2 ; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm2[0],ymm0[1],ymm2[2,3,4],ymm0[5],ymm2[6,7,8],ymm0[9],ymm2[10,11,12],ymm0[13],ymm2[14,15] ; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm17 ; AVX512DQ-FCP-NEXT: vmovdqa (%r10), %xmm1 ; AVX512DQ-FCP-NEXT: vmovdqa (%rax), %xmm14 ; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm14[8],xmm1[8],xmm14[9],xmm1[9],xmm14[10],xmm1[10],xmm14[11],xmm1[11],xmm14[12],xmm1[12],xmm14[13],xmm1[13],xmm14[14],xmm1[14],xmm14[15],xmm1[15] ; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm1, %xmm28 ; AVX512DQ-FCP-NEXT: vinserti128 $1, %xmm0, %ymm0, %ymm0 ; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm2 = [0,1,2,3,8,9,8,9,8,9,8,9,12,13,10,11,0,1,2,3,8,9,12,13,8,9,12,13,12,13,14,15] ; AVX512DQ-FCP-NEXT: vpshufb %ymm2, %ymm0, %ymm1 ; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm2, %ymm24 ; AVX512DQ-FCP-NEXT: vmovdqa (%r9), %xmm12 ; AVX512DQ-FCP-NEXT: vmovdqa (%r8), %xmm13 ; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm2 = 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: vinserti128 $1, %xmm2, %ymm2, %ymm2 ; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm4 = [0,1,2,3,8,9,10,11,8,9,10,11,10,11,14,15,0,1,2,3,12,13,10,11,12,13,10,11,14,15,14,15] ; AVX512DQ-FCP-NEXT: vpshufb %ymm4, %ymm2, %ymm3 ; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm4, %ymm25 ; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm3[0,1,2],ymm1[3],ymm3[4,5,6],ymm1[7],ymm3[8,9,10],ymm1[11],ymm3[12,13,14],ymm1[15] ; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm3 = [0,1,0,1,0,1,0,1,8,9,10,11,4,5,2,3,0,1,4,5,0,1,4,5,8,9,10,11,4,5,6,7] ; AVX512DQ-FCP-NEXT: vpshufb %ymm3, %ymm0, %ymm0 ; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm3, %ymm26 ; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm3 = [0,1,2,3,0,1,2,3,8,9,10,11,2,3,6,7,4,5,2,3,4,5,2,3,8,9,10,11,6,7,6,7] ; AVX512DQ-FCP-NEXT: vpshufb %ymm3, %ymm2, %ymm2 ; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm3, %ymm27 ; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm2[0,1,2],ymm0[3],ymm2[4,5,6],ymm0[7],ymm2[8,9,10],ymm0[11],ymm2[12,13,14],ymm0[15] ; AVX512DQ-FCP-NEXT: movw $-21846, %r11w # imm = 0xAAAA ; AVX512DQ-FCP-NEXT: kmovw %r11d, %k1 ; AVX512DQ-FCP-NEXT: vinserti32x8 $1, %ymm1, %zmm0, %zmm17 {%k1} ; AVX512DQ-FCP-NEXT: vmovdqa 16(%rsi), %xmm10 ; AVX512DQ-FCP-NEXT: vmovdqa 16(%rdi), %xmm9 ; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm9[0],xmm10[0],xmm9[1],xmm10[1],xmm9[2],xmm10[2],xmm9[3],xmm10[3],xmm9[4],xmm10[4],xmm9[5],xmm10[5],xmm9[6],xmm10[6],xmm9[7],xmm10[7] ; AVX512DQ-FCP-NEXT: vinserti128 $1, %xmm2, %ymm2, %ymm0 ; AVX512DQ-FCP-NEXT: vpshufb %ymm11, %ymm0, %ymm3 ; AVX512DQ-FCP-NEXT: vmovdqa 16(%rcx), %xmm7 ; AVX512DQ-FCP-NEXT: vmovdqa 16(%rdx), %xmm6 ; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm4 = xmm6[0],xmm7[0],xmm6[1],xmm7[1],xmm6[2],xmm7[2],xmm6[3],xmm7[3],xmm6[4],xmm7[4],xmm6[5],xmm7[5],xmm6[6],xmm7[6],xmm6[7],xmm7[7] ; AVX512DQ-FCP-NEXT: vinserti128 $1, %xmm4, %ymm4, %ymm4 ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm5 = ymm4[u,u,8,9,u,u,u,u,u,u,10,11,u,u,u,u,u,u,28,29,u,u,u,u,u,u,30,31,u,u,u,u] ; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0],ymm5[1],ymm3[2,3,4],ymm5[5],ymm3[6,7,8],ymm5[9],ymm3[10,11,12],ymm5[13],ymm3[14,15] ; AVX512DQ-FCP-NEXT: vpmovzxwq {{.*#+}} xmm5 = xmm2[0],zero,zero,zero,xmm2[1],zero,zero,zero ; AVX512DQ-FCP-NEXT: vpshufb %xmm8, %xmm2, %xmm2 ; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm8, %xmm29 ; AVX512DQ-FCP-NEXT: vinserti128 $1, %xmm2, %ymm5, %ymm2 ; AVX512DQ-FCP-NEXT: vpshufb {{.*#+}} ymm4 = ymm4[u,u,0,1,u,u,u,u,u,u,2,3,u,u,u,u,u,u,20,21,u,u,u,u,u,u,22,23,u,u,u,u] ; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0],ymm4[1],ymm2[2,3,4],ymm4[5],ymm2[6,7,8],ymm4[9],ymm2[10,11,12],ymm4[13],ymm2[14,15] ; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm16 ; AVX512DQ-FCP-NEXT: vmovdqa 16(%r10), %xmm8 ; AVX512DQ-FCP-NEXT: vmovdqa 16(%rax), %xmm4 ; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm4[0],xmm8[0],xmm4[1],xmm8[1],xmm4[2],xmm8[2],xmm4[3],xmm8[3],xmm4[4],xmm8[4],xmm4[5],xmm8[5],xmm4[6],xmm8[6],xmm4[7],xmm8[7] ; AVX512DQ-FCP-NEXT: vinserti128 $1, %xmm2, %ymm2, %ymm0 ; AVX512DQ-FCP-NEXT: vmovdqa 16(%r9), %xmm5 ; AVX512DQ-FCP-NEXT: vmovdqa 16(%r8), %xmm3 ; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm15 = xmm3[0],xmm5[0],xmm3[1],xmm5[1],xmm3[2],xmm5[2],xmm3[3],xmm5[3],xmm3[4],xmm5[4],xmm3[5],xmm5[5],xmm3[6],xmm5[6],xmm3[7],xmm5[7] ; AVX512DQ-FCP-NEXT: vinserti128 $1, %xmm15, %ymm15, %ymm15 ; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm2 = [0,1,2,3,8,9,8,9,8,9,10,11,10,11,10,11,0,1,2,3,12,13,12,13,12,13,10,11,14,15,14,15] ; AVX512DQ-FCP-NEXT: vpshufb %ymm2, %ymm0, %ymm1 ; AVX512DQ-FCP-NEXT: vpshufb %ymm2, %ymm15, %ymm2 ; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0,1,2],ymm1[3],ymm2[4,5,6],ymm1[7],ymm2[8,9,10],ymm1[11],ymm2[12,13,14],ymm1[15] ; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm2 = [0,1,2,3,0,1,0,1,8,9,10,11,2,3,2,3,4,5,2,3,4,5,4,5,8,9,10,11,6,7,6,7] ; AVX512DQ-FCP-NEXT: vpshufb %ymm2, %ymm0, %ymm0 ; AVX512DQ-FCP-NEXT: vpshufb %ymm2, %ymm15, %ymm2 ; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm2[0,1,2],ymm0[3],ymm2[4,5,6],ymm0[7],ymm2[8,9,10],ymm0[11],ymm2[12,13,14],ymm0[15] ; AVX512DQ-FCP-NEXT: vinserti32x8 $1, %ymm1, %zmm0, %zmm16 {%k1} ; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm6[8],xmm7[8],xmm6[9],xmm7[9],xmm6[10],xmm7[10],xmm6[11],xmm7[11],xmm6[12],xmm7[12],xmm6[13],xmm7[13],xmm6[14],xmm7[14],xmm6[15],xmm7[15] ; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm1 = 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] ; AVX512DQ-FCP-NEXT: vinserti128 $1, %xmm0, %ymm0, %ymm0 ; AVX512DQ-FCP-NEXT: vinserti128 $1, %xmm1, %ymm1, %ymm2 ; AVX512DQ-FCP-NEXT: vmovdqa %ymm11, %ymm10 ; AVX512DQ-FCP-NEXT: vpshufb %ymm11, %ymm2, %ymm2 ; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm18, %ymm11 ; AVX512DQ-FCP-NEXT: vpshufb %ymm11, %ymm0, %ymm6 ; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0],ymm6[1],ymm2[2,3,4],ymm6[5],ymm2[6,7,8],ymm6[9],ymm2[10,11,12],ymm6[13],ymm2[14,15] ; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm23, %ymm9 ; AVX512DQ-FCP-NEXT: vpshufb %ymm9, %ymm0, %ymm0 ; AVX512DQ-FCP-NEXT: vpmovzxwq {{.*#+}} xmm6 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero ; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm29, %xmm15 ; AVX512DQ-FCP-NEXT: vpshufb %xmm15, %xmm1, %xmm1 ; AVX512DQ-FCP-NEXT: vinserti128 $1, %xmm1, %ymm6, %ymm1 ; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[2,3,4],ymm0[5],ymm1[6,7,8],ymm0[9],ymm1[10,11,12],ymm0[13],ymm1[14,15] ; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm0, %zmm0 ; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm4[8],xmm8[8],xmm4[9],xmm8[9],xmm4[10],xmm8[10],xmm4[11],xmm8[11],xmm4[12],xmm8[12],xmm4[13],xmm8[13],xmm4[14],xmm8[14],xmm4[15],xmm8[15] ; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm2 = 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] ; AVX512DQ-FCP-NEXT: vinserti128 $1, %xmm1, %ymm1, %ymm1 ; AVX512DQ-FCP-NEXT: vinserti128 $1, %xmm2, %ymm2, %ymm2 ; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm24, %ymm5 ; AVX512DQ-FCP-NEXT: vpshufb %ymm5, %ymm1, %ymm3 ; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm25, %ymm6 ; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm2, %ymm4 ; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm4[0,1,2],ymm3[3],ymm4[4,5,6],ymm3[7],ymm4[8,9,10],ymm3[11],ymm4[12,13,14],ymm3[15] ; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm26, %ymm8 ; AVX512DQ-FCP-NEXT: vpshufb %ymm8, %ymm1, %ymm1 ; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm27, %ymm7 ; AVX512DQ-FCP-NEXT: vpshufb %ymm7, %ymm2, %ymm2 ; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0,1,2],ymm1[3],ymm2[4,5,6],ymm1[7],ymm2[8,9,10],ymm1[11],ymm2[12,13,14],ymm1[15] ; AVX512DQ-FCP-NEXT: vinserti32x8 $1, %ymm3, %zmm1, %zmm0 {%k1} ; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm19, %xmm1 ; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm20, %xmm2 ; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3],xmm1[4],xmm2[4],xmm1[5],xmm2[5],xmm1[6],xmm2[6],xmm1[7],xmm2[7] ; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm21, %xmm2 ; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm22, %xmm3 ; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} 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] ; AVX512DQ-FCP-NEXT: vinserti128 $1, %xmm2, %ymm2, %ymm3 ; AVX512DQ-FCP-NEXT: vpshufb %ymm10, %ymm3, %ymm3 ; AVX512DQ-FCP-NEXT: vinserti128 $1, %xmm1, %ymm1, %ymm1 ; AVX512DQ-FCP-NEXT: vpshufb %ymm11, %ymm1, %ymm4 ; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0],ymm4[1],ymm3[2,3,4],ymm4[5],ymm3[6,7,8],ymm4[9],ymm3[10,11,12],ymm4[13],ymm3[14,15] ; AVX512DQ-FCP-NEXT: vpshufb %ymm9, %ymm1, %ymm1 ; AVX512DQ-FCP-NEXT: vpshufb %xmm15, %xmm2, %xmm4 ; AVX512DQ-FCP-NEXT: vpmovzxwq {{.*#+}} xmm2 = xmm2[0],zero,zero,zero,xmm2[1],zero,zero,zero ; AVX512DQ-FCP-NEXT: vinserti128 $1, %xmm4, %ymm2, %ymm2 ; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2,3,4],ymm1[5],ymm2[6,7,8],ymm1[9],ymm2[10,11,12],ymm1[13],ymm2[14,15] ; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm3, %zmm1, %zmm1 ; AVX512DQ-FCP-NEXT: vmovdqa64 %xmm28, %xmm2 ; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm14[0],xmm2[0],xmm14[1],xmm2[1],xmm14[2],xmm2[2],xmm14[3],xmm2[3],xmm14[4],xmm2[4],xmm14[5],xmm2[5],xmm14[6],xmm2[6],xmm14[7],xmm2[7] ; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm3 = 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: vinserti128 $1, %xmm2, %ymm2, %ymm2 ; AVX512DQ-FCP-NEXT: vpshufb %ymm5, %ymm2, %ymm4 ; AVX512DQ-FCP-NEXT: vinserti128 $1, %xmm3, %ymm3, %ymm3 ; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm3, %ymm5 ; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm4 = ymm5[0,1,2],ymm4[3],ymm5[4,5,6],ymm4[7],ymm5[8,9,10],ymm4[11],ymm5[12,13,14],ymm4[15] ; AVX512DQ-FCP-NEXT: vpshufb %ymm8, %ymm2, %ymm2 ; AVX512DQ-FCP-NEXT: vpshufb %ymm7, %ymm3, %ymm3 ; AVX512DQ-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm3[0,1,2],ymm2[3],ymm3[4,5,6],ymm2[7],ymm3[8,9,10],ymm2[11],ymm3[12,13,14],ymm2[15] ; AVX512DQ-FCP-NEXT: vinserti32x8 $1, %ymm4, %zmm2, %zmm1 {%k1} ; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, (%rax) ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, 192(%rax) ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm16, 128(%rax) ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm17, 64(%rax) ; AVX512DQ-FCP-NEXT: vzeroupper ; AVX512DQ-FCP-NEXT: retq ; ; AVX512BW-LABEL: store_i8_stride8_vf32: ; AVX512BW: # %bb.0: ; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %r11 ; AVX512BW-NEXT: vmovdqa (%r11), %xmm0 ; AVX512BW-NEXT: vmovdqa 16(%r11), %xmm1 ; AVX512BW-NEXT: vmovdqa (%r10), %xmm2 ; AVX512BW-NEXT: vmovdqa 16(%r10), %xmm3 ; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm4 = xmm2[8],xmm0[8],xmm2[9],xmm0[9],xmm2[10],xmm0[10],xmm2[11],xmm0[11],xmm2[12],xmm0[12],xmm2[13],xmm0[13],xmm2[14],xmm0[14],xmm2[15],xmm0[15] ; AVX512BW-NEXT: vmovdqa (%r9), %xmm5 ; AVX512BW-NEXT: vmovdqa 16(%r9), %xmm6 ; AVX512BW-NEXT: vmovdqa (%r8), %xmm7 ; AVX512BW-NEXT: vmovdqa 16(%r8), %xmm8 ; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm9 = xmm7[8],xmm5[8],xmm7[9],xmm5[9],xmm7[10],xmm5[10],xmm7[11],xmm5[11],xmm7[12],xmm5[12],xmm7[13],xmm5[13],xmm7[14],xmm5[14],xmm7[15],xmm5[15] ; AVX512BW-NEXT: vpmovsxbw {{.*#+}} zmm10 = [0,1,0,32,4,5,1,33,2,1,2,34,4,5,3,35,0,1,4,36,4,5,5,37,0,1,6,38,6,5,7,39] ; AVX512BW-NEXT: vpermt2w %zmm4, %zmm10, %zmm9 ; AVX512BW-NEXT: vmovdqa (%rcx), %xmm4 ; AVX512BW-NEXT: vmovdqa 16(%rcx), %xmm11 ; AVX512BW-NEXT: vmovdqa (%rdx), %xmm12 ; AVX512BW-NEXT: vmovdqa 16(%rdx), %xmm13 ; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm14 = xmm12[8],xmm4[8],xmm12[9],xmm4[9],xmm12[10],xmm4[10],xmm12[11],xmm4[11],xmm12[12],xmm4[12],xmm12[13],xmm4[13],xmm12[14],xmm4[14],xmm12[15],xmm4[15] ; AVX512BW-NEXT: vmovdqa (%rsi), %xmm15 ; AVX512BW-NEXT: vmovdqa64 16(%rsi), %xmm16 ; AVX512BW-NEXT: vmovdqa64 (%rdi), %xmm17 ; AVX512BW-NEXT: vmovdqa64 16(%rdi), %xmm18 ; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm19 = xmm17[8],xmm15[8],xmm17[9],xmm15[9],xmm17[10],xmm15[10],xmm17[11],xmm15[11],xmm17[12],xmm15[12],xmm17[13],xmm15[13],xmm17[14],xmm15[14],xmm17[15],xmm15[15] ; AVX512BW-NEXT: vpmovsxbw {{.*#+}} zmm20 = [0,32,0,0,1,33,0,0,2,34,0,0,3,35,0,0,4,36,0,0,5,37,0,0,6,38,0,0,7,39,0,0] ; AVX512BW-NEXT: vpermt2w %zmm14, %zmm20, %zmm19 ; AVX512BW-NEXT: movw $-21846, %cx # imm = 0xAAAA ; AVX512BW-NEXT: kmovd %ecx, %k1 ; AVX512BW-NEXT: vmovdqa32 %zmm9, %zmm19 {%k1} ; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm9 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3],xmm3[4],xmm1[4],xmm3[5],xmm1[5],xmm3[6],xmm1[6],xmm3[7],xmm1[7] ; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm14 = 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-NEXT: vpermt2w %zmm9, %zmm10, %zmm14 ; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm9 = 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] ; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm21 = xmm18[0],xmm16[0],xmm18[1],xmm16[1],xmm18[2],xmm16[2],xmm18[3],xmm16[3],xmm18[4],xmm16[4],xmm18[5],xmm16[5],xmm18[6],xmm16[6],xmm18[7],xmm16[7] ; AVX512BW-NEXT: vpermt2w %zmm9, %zmm20, %zmm21 ; AVX512BW-NEXT: vmovdqa32 %zmm14, %zmm21 {%k1} ; AVX512BW-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] ; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm3 = 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-NEXT: vpermt2w %zmm1, %zmm10, %zmm3 ; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm1 = 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] ; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm6 = xmm18[8],xmm16[8],xmm18[9],xmm16[9],xmm18[10],xmm16[10],xmm18[11],xmm16[11],xmm18[12],xmm16[12],xmm18[13],xmm16[13],xmm18[14],xmm16[14],xmm18[15],xmm16[15] ; AVX512BW-NEXT: vpermt2w %zmm1, %zmm20, %zmm6 ; AVX512BW-NEXT: vmovdqa32 %zmm3, %zmm6 {%k1} ; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3],xmm2[4],xmm0[4],xmm2[5],xmm0[5],xmm2[6],xmm0[6],xmm2[7],xmm0[7] ; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm1 = 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: vpermt2w %zmm0, %zmm10, %zmm1 ; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm0 = 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] ; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm17[0],xmm15[0],xmm17[1],xmm15[1],xmm17[2],xmm15[2],xmm17[3],xmm15[3],xmm17[4],xmm15[4],xmm17[5],xmm15[5],xmm17[6],xmm15[6],xmm17[7],xmm15[7] ; AVX512BW-NEXT: vpermt2w %zmm0, %zmm20, %zmm2 ; AVX512BW-NEXT: vmovdqa32 %zmm1, %zmm2 {%k1} ; AVX512BW-NEXT: vmovdqa64 %zmm2, (%rax) ; AVX512BW-NEXT: vmovdqa64 %zmm6, 192(%rax) ; AVX512BW-NEXT: vmovdqa64 %zmm21, 128(%rax) ; AVX512BW-NEXT: vmovdqa64 %zmm19, 64(%rax) ; AVX512BW-NEXT: vzeroupper ; AVX512BW-NEXT: retq ; ; AVX512BW-FCP-LABEL: store_i8_stride8_vf32: ; AVX512BW-FCP: # %bb.0: ; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r11 ; AVX512BW-FCP-NEXT: vmovdqa (%r11), %xmm0 ; AVX512BW-FCP-NEXT: vmovdqa 16(%r11), %xmm1 ; AVX512BW-FCP-NEXT: vmovdqa (%r10), %xmm2 ; AVX512BW-FCP-NEXT: vmovdqa 16(%r10), %xmm3 ; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm4 = xmm2[8],xmm0[8],xmm2[9],xmm0[9],xmm2[10],xmm0[10],xmm2[11],xmm0[11],xmm2[12],xmm0[12],xmm2[13],xmm0[13],xmm2[14],xmm0[14],xmm2[15],xmm0[15] ; AVX512BW-FCP-NEXT: vmovdqa (%r9), %xmm5 ; AVX512BW-FCP-NEXT: vmovdqa 16(%r9), %xmm6 ; AVX512BW-FCP-NEXT: vmovdqa (%r8), %xmm7 ; AVX512BW-FCP-NEXT: vmovdqa 16(%r8), %xmm8 ; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm9 = xmm7[8],xmm5[8],xmm7[9],xmm5[9],xmm7[10],xmm5[10],xmm7[11],xmm5[11],xmm7[12],xmm5[12],xmm7[13],xmm5[13],xmm7[14],xmm5[14],xmm7[15],xmm5[15] ; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm10 = [0,1,0,32,4,5,1,33,2,1,2,34,4,5,3,35,0,1,4,36,4,5,5,37,0,1,6,38,6,5,7,39] ; AVX512BW-FCP-NEXT: vpermt2w %zmm4, %zmm10, %zmm9 ; AVX512BW-FCP-NEXT: vmovdqa (%rcx), %xmm4 ; AVX512BW-FCP-NEXT: vmovdqa 16(%rcx), %xmm11 ; AVX512BW-FCP-NEXT: vmovdqa (%rdx), %xmm12 ; AVX512BW-FCP-NEXT: vmovdqa 16(%rdx), %xmm13 ; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm14 = xmm12[8],xmm4[8],xmm12[9],xmm4[9],xmm12[10],xmm4[10],xmm12[11],xmm4[11],xmm12[12],xmm4[12],xmm12[13],xmm4[13],xmm12[14],xmm4[14],xmm12[15],xmm4[15] ; AVX512BW-FCP-NEXT: vmovdqa (%rsi), %xmm15 ; AVX512BW-FCP-NEXT: vmovdqa64 16(%rsi), %xmm16 ; AVX512BW-FCP-NEXT: vmovdqa64 (%rdi), %xmm17 ; AVX512BW-FCP-NEXT: vmovdqa64 16(%rdi), %xmm18 ; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm19 = xmm17[8],xmm15[8],xmm17[9],xmm15[9],xmm17[10],xmm15[10],xmm17[11],xmm15[11],xmm17[12],xmm15[12],xmm17[13],xmm15[13],xmm17[14],xmm15[14],xmm17[15],xmm15[15] ; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm20 = [0,32,0,0,1,33,0,0,2,34,0,0,3,35,0,0,4,36,0,0,5,37,0,0,6,38,0,0,7,39,0,0] ; AVX512BW-FCP-NEXT: vpermt2w %zmm14, %zmm20, %zmm19 ; AVX512BW-FCP-NEXT: movw $-21846, %cx # imm = 0xAAAA ; AVX512BW-FCP-NEXT: kmovd %ecx, %k1 ; AVX512BW-FCP-NEXT: vmovdqa32 %zmm9, %zmm19 {%k1} ; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm9 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3],xmm3[4],xmm1[4],xmm3[5],xmm1[5],xmm3[6],xmm1[6],xmm3[7],xmm1[7] ; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm14 = 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: vpermt2w %zmm9, %zmm10, %zmm14 ; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm9 = 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] ; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm21 = xmm18[0],xmm16[0],xmm18[1],xmm16[1],xmm18[2],xmm16[2],xmm18[3],xmm16[3],xmm18[4],xmm16[4],xmm18[5],xmm16[5],xmm18[6],xmm16[6],xmm18[7],xmm16[7] ; AVX512BW-FCP-NEXT: vpermt2w %zmm9, %zmm20, %zmm21 ; AVX512BW-FCP-NEXT: vmovdqa32 %zmm14, %zmm21 {%k1} ; AVX512BW-FCP-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] ; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm3 = 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 %zmm1, %zmm10, %zmm3 ; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm1 = 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] ; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm6 = xmm18[8],xmm16[8],xmm18[9],xmm16[9],xmm18[10],xmm16[10],xmm18[11],xmm16[11],xmm18[12],xmm16[12],xmm18[13],xmm16[13],xmm18[14],xmm16[14],xmm18[15],xmm16[15] ; AVX512BW-FCP-NEXT: vpermt2w %zmm1, %zmm20, %zmm6 ; AVX512BW-FCP-NEXT: vmovdqa32 %zmm3, %zmm6 {%k1} ; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3],xmm2[4],xmm0[4],xmm2[5],xmm0[5],xmm2[6],xmm0[6],xmm2[7],xmm0[7] ; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm1 = 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-FCP-NEXT: vpermt2w %zmm0, %zmm10, %zmm1 ; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm0 = 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] ; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm17[0],xmm15[0],xmm17[1],xmm15[1],xmm17[2],xmm15[2],xmm17[3],xmm15[3],xmm17[4],xmm15[4],xmm17[5],xmm15[5],xmm17[6],xmm15[6],xmm17[7],xmm15[7] ; AVX512BW-FCP-NEXT: vpermt2w %zmm0, %zmm20, %zmm2 ; AVX512BW-FCP-NEXT: vmovdqa32 %zmm1, %zmm2 {%k1} ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, (%rax) ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm6, 192(%rax) ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm21, 128(%rax) ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm19, 64(%rax) ; AVX512BW-FCP-NEXT: vzeroupper ; AVX512BW-FCP-NEXT: retq ; ; AVX512DQ-BW-LABEL: store_i8_stride8_vf32: ; AVX512DQ-BW: # %bb.0: ; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %r11 ; AVX512DQ-BW-NEXT: vmovdqa (%r11), %xmm0 ; AVX512DQ-BW-NEXT: vmovdqa 16(%r11), %xmm1 ; AVX512DQ-BW-NEXT: vmovdqa (%r10), %xmm2 ; AVX512DQ-BW-NEXT: vmovdqa 16(%r10), %xmm3 ; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm4 = xmm2[8],xmm0[8],xmm2[9],xmm0[9],xmm2[10],xmm0[10],xmm2[11],xmm0[11],xmm2[12],xmm0[12],xmm2[13],xmm0[13],xmm2[14],xmm0[14],xmm2[15],xmm0[15] ; AVX512DQ-BW-NEXT: vmovdqa (%r9), %xmm5 ; AVX512DQ-BW-NEXT: vmovdqa 16(%r9), %xmm6 ; AVX512DQ-BW-NEXT: vmovdqa (%r8), %xmm7 ; AVX512DQ-BW-NEXT: vmovdqa 16(%r8), %xmm8 ; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm9 = xmm7[8],xmm5[8],xmm7[9],xmm5[9],xmm7[10],xmm5[10],xmm7[11],xmm5[11],xmm7[12],xmm5[12],xmm7[13],xmm5[13],xmm7[14],xmm5[14],xmm7[15],xmm5[15] ; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} zmm10 = [0,1,0,32,4,5,1,33,2,1,2,34,4,5,3,35,0,1,4,36,4,5,5,37,0,1,6,38,6,5,7,39] ; AVX512DQ-BW-NEXT: vpermt2w %zmm4, %zmm10, %zmm9 ; AVX512DQ-BW-NEXT: vmovdqa (%rcx), %xmm4 ; AVX512DQ-BW-NEXT: vmovdqa 16(%rcx), %xmm11 ; AVX512DQ-BW-NEXT: vmovdqa (%rdx), %xmm12 ; AVX512DQ-BW-NEXT: vmovdqa 16(%rdx), %xmm13 ; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm14 = xmm12[8],xmm4[8],xmm12[9],xmm4[9],xmm12[10],xmm4[10],xmm12[11],xmm4[11],xmm12[12],xmm4[12],xmm12[13],xmm4[13],xmm12[14],xmm4[14],xmm12[15],xmm4[15] ; AVX512DQ-BW-NEXT: vmovdqa (%rsi), %xmm15 ; AVX512DQ-BW-NEXT: vmovdqa64 16(%rsi), %xmm16 ; AVX512DQ-BW-NEXT: vmovdqa64 (%rdi), %xmm17 ; AVX512DQ-BW-NEXT: vmovdqa64 16(%rdi), %xmm18 ; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm19 = xmm17[8],xmm15[8],xmm17[9],xmm15[9],xmm17[10],xmm15[10],xmm17[11],xmm15[11],xmm17[12],xmm15[12],xmm17[13],xmm15[13],xmm17[14],xmm15[14],xmm17[15],xmm15[15] ; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} zmm20 = [0,32,0,0,1,33,0,0,2,34,0,0,3,35,0,0,4,36,0,0,5,37,0,0,6,38,0,0,7,39,0,0] ; AVX512DQ-BW-NEXT: vpermt2w %zmm14, %zmm20, %zmm19 ; AVX512DQ-BW-NEXT: movw $-21846, %cx # imm = 0xAAAA ; AVX512DQ-BW-NEXT: kmovd %ecx, %k1 ; AVX512DQ-BW-NEXT: vmovdqa32 %zmm9, %zmm19 {%k1} ; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm9 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3],xmm3[4],xmm1[4],xmm3[5],xmm1[5],xmm3[6],xmm1[6],xmm3[7],xmm1[7] ; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm14 = 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-NEXT: vpermt2w %zmm9, %zmm10, %zmm14 ; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm9 = 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] ; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm21 = xmm18[0],xmm16[0],xmm18[1],xmm16[1],xmm18[2],xmm16[2],xmm18[3],xmm16[3],xmm18[4],xmm16[4],xmm18[5],xmm16[5],xmm18[6],xmm16[6],xmm18[7],xmm16[7] ; AVX512DQ-BW-NEXT: vpermt2w %zmm9, %zmm20, %zmm21 ; AVX512DQ-BW-NEXT: vmovdqa32 %zmm14, %zmm21 {%k1} ; AVX512DQ-BW-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] ; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm3 = 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-NEXT: vpermt2w %zmm1, %zmm10, %zmm3 ; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm1 = 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] ; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm6 = xmm18[8],xmm16[8],xmm18[9],xmm16[9],xmm18[10],xmm16[10],xmm18[11],xmm16[11],xmm18[12],xmm16[12],xmm18[13],xmm16[13],xmm18[14],xmm16[14],xmm18[15],xmm16[15] ; AVX512DQ-BW-NEXT: vpermt2w %zmm1, %zmm20, %zmm6 ; AVX512DQ-BW-NEXT: vmovdqa32 %zmm3, %zmm6 {%k1} ; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3],xmm2[4],xmm0[4],xmm2[5],xmm0[5],xmm2[6],xmm0[6],xmm2[7],xmm0[7] ; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm1 = 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: vpermt2w %zmm0, %zmm10, %zmm1 ; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm0 = 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] ; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm17[0],xmm15[0],xmm17[1],xmm15[1],xmm17[2],xmm15[2],xmm17[3],xmm15[3],xmm17[4],xmm15[4],xmm17[5],xmm15[5],xmm17[6],xmm15[6],xmm17[7],xmm15[7] ; AVX512DQ-BW-NEXT: vpermt2w %zmm0, %zmm20, %zmm2 ; AVX512DQ-BW-NEXT: vmovdqa32 %zmm1, %zmm2 {%k1} ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, (%rax) ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm6, 192(%rax) ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm21, 128(%rax) ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm19, 64(%rax) ; AVX512DQ-BW-NEXT: vzeroupper ; AVX512DQ-BW-NEXT: retq ; ; AVX512DQ-BW-FCP-LABEL: store_i8_stride8_vf32: ; AVX512DQ-BW-FCP: # %bb.0: ; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r11 ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%r11), %xmm0 ; AVX512DQ-BW-FCP-NEXT: vmovdqa 16(%r11), %xmm1 ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%r10), %xmm2 ; AVX512DQ-BW-FCP-NEXT: vmovdqa 16(%r10), %xmm3 ; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm4 = xmm2[8],xmm0[8],xmm2[9],xmm0[9],xmm2[10],xmm0[10],xmm2[11],xmm0[11],xmm2[12],xmm0[12],xmm2[13],xmm0[13],xmm2[14],xmm0[14],xmm2[15],xmm0[15] ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%r9), %xmm5 ; AVX512DQ-BW-FCP-NEXT: vmovdqa 16(%r9), %xmm6 ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%r8), %xmm7 ; AVX512DQ-BW-FCP-NEXT: vmovdqa 16(%r8), %xmm8 ; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm9 = xmm7[8],xmm5[8],xmm7[9],xmm5[9],xmm7[10],xmm5[10],xmm7[11],xmm5[11],xmm7[12],xmm5[12],xmm7[13],xmm5[13],xmm7[14],xmm5[14],xmm7[15],xmm5[15] ; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm10 = [0,1,0,32,4,5,1,33,2,1,2,34,4,5,3,35,0,1,4,36,4,5,5,37,0,1,6,38,6,5,7,39] ; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm4, %zmm10, %zmm9 ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rcx), %xmm4 ; AVX512DQ-BW-FCP-NEXT: vmovdqa 16(%rcx), %xmm11 ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdx), %xmm12 ; AVX512DQ-BW-FCP-NEXT: vmovdqa 16(%rdx), %xmm13 ; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm14 = xmm12[8],xmm4[8],xmm12[9],xmm4[9],xmm12[10],xmm4[10],xmm12[11],xmm4[11],xmm12[12],xmm4[12],xmm12[13],xmm4[13],xmm12[14],xmm4[14],xmm12[15],xmm4[15] ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rsi), %xmm15 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 16(%rsi), %xmm16 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rdi), %xmm17 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 16(%rdi), %xmm18 ; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm19 = xmm17[8],xmm15[8],xmm17[9],xmm15[9],xmm17[10],xmm15[10],xmm17[11],xmm15[11],xmm17[12],xmm15[12],xmm17[13],xmm15[13],xmm17[14],xmm15[14],xmm17[15],xmm15[15] ; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm20 = [0,32,0,0,1,33,0,0,2,34,0,0,3,35,0,0,4,36,0,0,5,37,0,0,6,38,0,0,7,39,0,0] ; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm14, %zmm20, %zmm19 ; AVX512DQ-BW-FCP-NEXT: movw $-21846, %cx # imm = 0xAAAA ; AVX512DQ-BW-FCP-NEXT: kmovd %ecx, %k1 ; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm9, %zmm19 {%k1} ; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm9 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3],xmm3[4],xmm1[4],xmm3[5],xmm1[5],xmm3[6],xmm1[6],xmm3[7],xmm1[7] ; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm14 = 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: vpermt2w %zmm9, %zmm10, %zmm14 ; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm9 = 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] ; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm21 = xmm18[0],xmm16[0],xmm18[1],xmm16[1],xmm18[2],xmm16[2],xmm18[3],xmm16[3],xmm18[4],xmm16[4],xmm18[5],xmm16[5],xmm18[6],xmm16[6],xmm18[7],xmm16[7] ; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm9, %zmm20, %zmm21 ; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm14, %zmm21 {%k1} ; AVX512DQ-BW-FCP-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] ; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm3 = 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 %zmm1, %zmm10, %zmm3 ; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm1 = 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] ; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm6 = xmm18[8],xmm16[8],xmm18[9],xmm16[9],xmm18[10],xmm16[10],xmm18[11],xmm16[11],xmm18[12],xmm16[12],xmm18[13],xmm16[13],xmm18[14],xmm16[14],xmm18[15],xmm16[15] ; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm1, %zmm20, %zmm6 ; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm3, %zmm6 {%k1} ; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3],xmm2[4],xmm0[4],xmm2[5],xmm0[5],xmm2[6],xmm0[6],xmm2[7],xmm0[7] ; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm1 = 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-FCP-NEXT: vpermt2w %zmm0, %zmm10, %zmm1 ; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm0 = 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] ; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm17[0],xmm15[0],xmm17[1],xmm15[1],xmm17[2],xmm15[2],xmm17[3],xmm15[3],xmm17[4],xmm15[4],xmm17[5],xmm15[5],xmm17[6],xmm15[6],xmm17[7],xmm15[7] ; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm0, %zmm20, %zmm2 ; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm1, %zmm2 {%k1} ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, (%rax) ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm6, 192(%rax) ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm21, 128(%rax) ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm19, 64(%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 %in.vec6 = load <32 x i8>, ptr %in.vecptr6, align 64 %in.vec7 = load <32 x i8>, ptr %in.vecptr7, 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 <32 x i8> %in.vec6, <32 x i8> %in.vec7, <64 x i32> %5 = shufflevector <64 x i8> %1, <64 x i8> %2, <128 x i32> %6 = shufflevector <64 x i8> %3, <64 x i8> %4, <128 x i32> %7 = shufflevector <128 x i8> %5, <128 x i8> %6, <256 x i32> %interleaved.vec = shufflevector <256 x i8> %7, <256 x i8> poison, <256 x i32> store <256 x i8> %interleaved.vec, ptr %out.vec, align 64 ret void } define void @store_i8_stride8_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecptr2, ptr %in.vecptr3, ptr %in.vecptr4, ptr %in.vecptr5, ptr %in.vecptr6, ptr %in.vecptr7, ptr %out.vec) nounwind { ; SSE-LABEL: store_i8_stride8_vf64: ; SSE: # %bb.0: ; SSE-NEXT: subq $312, %rsp # imm = 0x138 ; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax ; SSE-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; SSE-NEXT: movdqa (%rdi), %xmm3 ; SSE-NEXT: movdqa (%rsi), %xmm5 ; SSE-NEXT: movdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: movdqa (%rdx), %xmm4 ; SSE-NEXT: movdqa (%rcx), %xmm8 ; SSE-NEXT: movdqa (%r8), %xmm6 ; SSE-NEXT: movdqa (%r9), %xmm9 ; SSE-NEXT: movdqa (%r10), %xmm7 ; SSE-NEXT: movdqa (%rax), %xmm10 ; SSE-NEXT: movdqa %xmm7, %xmm0 ; 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: pshuflw {{.*#+}} xmm1 = xmm0[0,0,2,1,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm1[0,0,2,1] ; SSE-NEXT: movdqa {{.*#+}} xmm13 = [65535,65535,65535,0,65535,65535,65535,0] ; SSE-NEXT: movdqa %xmm13, %xmm12 ; SSE-NEXT: pandn %xmm2, %xmm12 ; SSE-NEXT: movdqa %xmm6, %xmm11 ; SSE-NEXT: punpcklbw {{.*#+}} xmm11 = 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] ; SSE-NEXT: pshuflw {{.*#+}} xmm2 = xmm11[0,1,1,3,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm14 = xmm2[0,0,2,1] ; SSE-NEXT: pand %xmm13, %xmm14 ; SSE-NEXT: por %xmm12, %xmm14 ; SSE-NEXT: movdqa %xmm4, %xmm12 ; SSE-NEXT: punpcklbw {{.*#+}} xmm12 = 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] ; SSE-NEXT: pshuflw {{.*#+}} xmm2 = xmm12[0,0,2,1,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm2[0,1,1,3] ; SSE-NEXT: movdqa {{.*#+}} xmm1 = [65535,0,65535,65535,65535,0,65535,65535] ; SSE-NEXT: movdqa %xmm1, %xmm15 ; SSE-NEXT: pandn %xmm2, %xmm15 ; SSE-NEXT: movdqa %xmm3, %xmm2 ; SSE-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm5[0],xmm2[1],xmm5[1],xmm2[2],xmm5[2],xmm2[3],xmm5[3],xmm2[4],xmm5[4],xmm2[5],xmm5[5],xmm2[6],xmm5[6],xmm2[7],xmm5[7] ; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm2[0,0,0,0] ; SSE-NEXT: pshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,5,5,5,5] ; SSE-NEXT: pand %xmm1, %xmm5 ; SSE-NEXT: por %xmm15, %xmm5 ; SSE-NEXT: pshufd {{.*#+}} xmm14 = xmm14[1,3,2,3] ; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm5[0,2,2,3] ; SSE-NEXT: punpckldq {{.*#+}} xmm5 = xmm5[0],xmm14[0],xmm5[1],xmm14[1] ; SSE-NEXT: movdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: pshuflw {{.*#+}} xmm5 = xmm0[0,2,2,3,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm5[0,0,2,1] ; SSE-NEXT: movdqa %xmm13, %xmm14 ; SSE-NEXT: pandn %xmm5, %xmm14 ; SSE-NEXT: pshuflw {{.*#+}} xmm5 = xmm11[2,1,3,3,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm5[0,0,2,1] ; SSE-NEXT: pand %xmm13, %xmm5 ; SSE-NEXT: por %xmm14, %xmm5 ; SSE-NEXT: pshuflw {{.*#+}} xmm14 = xmm12[0,2,2,3,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm14 = xmm14[0,1,1,3] ; SSE-NEXT: movdqa %xmm1, %xmm15 ; SSE-NEXT: pandn %xmm14, %xmm15 ; SSE-NEXT: pshufd {{.*#+}} xmm14 = xmm2[1,1,1,1] ; SSE-NEXT: pshufhw {{.*#+}} xmm14 = xmm14[0,1,2,3,5,5,5,5] ; SSE-NEXT: pand %xmm1, %xmm14 ; SSE-NEXT: por %xmm15, %xmm14 ; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm5[1,3,2,3] ; SSE-NEXT: pshufd {{.*#+}} xmm14 = xmm14[0,2,2,3] ; SSE-NEXT: punpckldq {{.*#+}} xmm14 = xmm14[0],xmm5[0],xmm14[1],xmm5[1] ; SSE-NEXT: movdqa %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: pshufhw {{.*#+}} xmm5 = xmm0[0,1,2,3,4,4,6,5] ; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm5[0,2,2,3] ; SSE-NEXT: movdqa %xmm13, %xmm14 ; SSE-NEXT: pandn %xmm5, %xmm14 ; SSE-NEXT: pshufhw {{.*#+}} xmm5 = xmm11[0,1,2,3,4,5,5,7] ; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm5[0,2,2,3] ; SSE-NEXT: pand %xmm13, %xmm5 ; SSE-NEXT: por %xmm14, %xmm5 ; SSE-NEXT: pshufhw {{.*#+}} xmm14 = xmm12[0,1,2,3,4,4,6,5] ; SSE-NEXT: pshufd {{.*#+}} xmm14 = xmm14[2,1,3,3] ; SSE-NEXT: movdqa %xmm1, %xmm15 ; SSE-NEXT: pandn %xmm14, %xmm15 ; SSE-NEXT: pshufd {{.*#+}} xmm14 = xmm2[2,2,2,2] ; SSE-NEXT: pshufhw {{.*#+}} xmm14 = xmm14[0,1,2,3,5,5,5,5] ; SSE-NEXT: pand %xmm1, %xmm14 ; SSE-NEXT: por %xmm15, %xmm14 ; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm5[1,3,2,3] ; SSE-NEXT: pshufd {{.*#+}} xmm14 = xmm14[0,2,2,3] ; SSE-NEXT: punpckldq {{.*#+}} xmm14 = xmm14[0],xmm5[0],xmm14[1],xmm5[1] ; SSE-NEXT: movdqa %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,6,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3] ; SSE-NEXT: movdqa %xmm13, %xmm5 ; SSE-NEXT: pandn %xmm0, %xmm5 ; SSE-NEXT: pshufhw {{.*#+}} xmm0 = xmm11[0,1,2,3,6,5,7,7] ; SSE-NEXT: pshufd {{.*#+}} xmm11 = xmm0[0,2,2,3] ; SSE-NEXT: pand %xmm13, %xmm11 ; SSE-NEXT: por %xmm5, %xmm11 ; SSE-NEXT: pshufhw {{.*#+}} xmm0 = xmm12[0,1,2,3,4,6,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[2,1,3,3] ; SSE-NEXT: movdqa %xmm1, %xmm5 ; SSE-NEXT: pandn %xmm0, %xmm5 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm2[3,3,3,3] ; SSE-NEXT: pshufhw {{.*#+}} xmm12 = xmm0[0,1,2,3,5,5,5,5] ; SSE-NEXT: pand %xmm1, %xmm12 ; SSE-NEXT: por %xmm5, %xmm12 ; SSE-NEXT: movdqa 16(%r8), %xmm0 ; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm11[1,3,2,3] ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm12[0,2,2,3] ; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm5[0],xmm2[1],xmm5[1] ; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: punpckhbw {{.*#+}} xmm7 = xmm7[8],xmm10[8],xmm7[9],xmm10[9],xmm7[10],xmm10[10],xmm7[11],xmm10[11],xmm7[12],xmm10[12],xmm7[13],xmm10[13],xmm7[14],xmm10[14],xmm7[15],xmm10[15] ; SSE-NEXT: pshuflw {{.*#+}} xmm5 = xmm7[0,0,2,1,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm5[0,0,2,1] ; SSE-NEXT: movdqa %xmm13, %xmm10 ; SSE-NEXT: pandn %xmm5, %xmm10 ; SSE-NEXT: punpckhbw {{.*#+}} xmm6 = xmm6[8],xmm9[8],xmm6[9],xmm9[9],xmm6[10],xmm9[10],xmm6[11],xmm9[11],xmm6[12],xmm9[12],xmm6[13],xmm9[13],xmm6[14],xmm9[14],xmm6[15],xmm9[15] ; SSE-NEXT: pshuflw {{.*#+}} xmm5 = xmm6[0,1,1,3,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm5[0,0,2,1] ; SSE-NEXT: pand %xmm13, %xmm5 ; SSE-NEXT: por %xmm10, %xmm5 ; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm5[1,3,2,3] ; SSE-NEXT: punpckhbw {{.*#+}} xmm4 = xmm4[8],xmm8[8],xmm4[9],xmm8[9],xmm4[10],xmm8[10],xmm4[11],xmm8[11],xmm4[12],xmm8[12],xmm4[13],xmm8[13],xmm4[14],xmm8[14],xmm4[15],xmm8[15] ; SSE-NEXT: pshuflw {{.*#+}} xmm8 = xmm4[0,0,2,1,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm8 = xmm8[0,1,1,3] ; SSE-NEXT: movdqa %xmm1, %xmm9 ; SSE-NEXT: pandn %xmm8, %xmm9 ; SSE-NEXT: punpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload ; SSE-NEXT: # xmm3 = xmm3[8],mem[8],xmm3[9],mem[9],xmm3[10],mem[10],xmm3[11],mem[11],xmm3[12],mem[12],xmm3[13],mem[13],xmm3[14],mem[14],xmm3[15],mem[15] ; SSE-NEXT: pshufd {{.*#+}} xmm8 = xmm3[0,0,0,0] ; SSE-NEXT: pshufhw {{.*#+}} xmm8 = xmm8[0,1,2,3,5,5,5,5] ; SSE-NEXT: pand %xmm1, %xmm8 ; SSE-NEXT: por %xmm9, %xmm8 ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm8[0,2,2,3] ; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm5[0],xmm2[1],xmm5[1] ; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: pshuflw {{.*#+}} xmm5 = xmm7[0,2,2,3,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm5[0,0,2,1] ; SSE-NEXT: movdqa %xmm13, %xmm8 ; SSE-NEXT: pandn %xmm5, %xmm8 ; SSE-NEXT: pshuflw {{.*#+}} xmm5 = xmm6[2,1,3,3,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm5[0,0,2,1] ; SSE-NEXT: pand %xmm13, %xmm5 ; SSE-NEXT: por %xmm8, %xmm5 ; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm5[1,3,2,3] ; SSE-NEXT: pshuflw {{.*#+}} xmm8 = xmm4[0,2,2,3,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm8 = xmm8[0,1,1,3] ; SSE-NEXT: movdqa %xmm1, %xmm9 ; SSE-NEXT: pandn %xmm8, %xmm9 ; SSE-NEXT: pshufd {{.*#+}} xmm8 = xmm3[1,1,1,1] ; SSE-NEXT: pshufhw {{.*#+}} xmm8 = xmm8[0,1,2,3,5,5,5,5] ; SSE-NEXT: pand %xmm1, %xmm8 ; SSE-NEXT: por %xmm9, %xmm8 ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm8[0,2,2,3] ; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm5[0],xmm2[1],xmm5[1] ; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: pshufhw {{.*#+}} xmm5 = xmm7[0,1,2,3,4,4,6,5] ; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm5[0,2,2,3] ; SSE-NEXT: movdqa %xmm13, %xmm8 ; SSE-NEXT: pandn %xmm5, %xmm8 ; SSE-NEXT: pshufhw {{.*#+}} xmm5 = xmm6[0,1,2,3,4,5,5,7] ; SSE-NEXT: pshufd {{.*#+}} xmm9 = xmm5[0,2,2,3] ; SSE-NEXT: pand %xmm13, %xmm9 ; SSE-NEXT: por %xmm8, %xmm9 ; SSE-NEXT: pshufhw {{.*#+}} xmm5 = xmm4[0,1,2,3,4,4,6,5] ; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm5[2,1,3,3] ; SSE-NEXT: movdqa %xmm1, %xmm8 ; SSE-NEXT: pandn %xmm5, %xmm8 ; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm3[2,2,2,2] ; SSE-NEXT: pshufhw {{.*#+}} xmm10 = xmm5[0,1,2,3,5,5,5,5] ; SSE-NEXT: pand %xmm1, %xmm10 ; SSE-NEXT: por %xmm8, %xmm10 ; SSE-NEXT: movdqa 16(%r10), %xmm5 ; SSE-NEXT: pshufd {{.*#+}} xmm8 = xmm9[1,3,2,3] ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm10[0,2,2,3] ; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm8[0],xmm2[1],xmm8[1] ; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: pshufhw {{.*#+}} xmm7 = xmm7[0,1,2,3,4,6,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm7[0,2,2,3] ; SSE-NEXT: movdqa %xmm13, %xmm8 ; SSE-NEXT: pandn %xmm7, %xmm8 ; SSE-NEXT: movdqa 16(%rax), %xmm7 ; SSE-NEXT: movdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: pshufhw {{.*#+}} xmm6 = xmm6[0,1,2,3,6,5,7,7] ; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm6[0,2,2,3] ; SSE-NEXT: pand %xmm13, %xmm6 ; SSE-NEXT: por %xmm8, %xmm6 ; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm6[1,3,2,3] ; SSE-NEXT: pshufhw {{.*#+}} xmm4 = xmm4[0,1,2,3,4,6,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm4 = xmm4[2,1,3,3] ; SSE-NEXT: movdqa %xmm1, %xmm8 ; SSE-NEXT: pandn %xmm4, %xmm8 ; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm3[3,3,3,3] ; SSE-NEXT: pshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,5,5,5,5] ; SSE-NEXT: pand %xmm1, %xmm3 ; SSE-NEXT: por %xmm8, %xmm3 ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm3[0,2,2,3] ; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm6[0],xmm2[1],xmm6[1] ; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: movdqa %xmm5, %xmm10 ; SSE-NEXT: punpcklbw {{.*#+}} xmm10 = xmm10[0],xmm7[0],xmm10[1],xmm7[1],xmm10[2],xmm7[2],xmm10[3],xmm7[3],xmm10[4],xmm7[4],xmm10[5],xmm7[5],xmm10[6],xmm7[6],xmm10[7],xmm7[7] ; SSE-NEXT: pshuflw {{.*#+}} xmm3 = xmm10[0,0,2,1,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm3[0,0,2,1] ; SSE-NEXT: movdqa %xmm13, %xmm4 ; SSE-NEXT: pandn %xmm3, %xmm4 ; SSE-NEXT: movdqa 16(%r9), %xmm6 ; SSE-NEXT: movdqa %xmm0, %xmm11 ; SSE-NEXT: punpcklbw {{.*#+}} xmm11 = xmm11[0],xmm6[0],xmm11[1],xmm6[1],xmm11[2],xmm6[2],xmm11[3],xmm6[3],xmm11[4],xmm6[4],xmm11[5],xmm6[5],xmm11[6],xmm6[6],xmm11[7],xmm6[7] ; SSE-NEXT: pshuflw {{.*#+}} xmm3 = xmm11[0,1,1,3,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm14 = xmm3[0,0,2,1] ; SSE-NEXT: pand %xmm13, %xmm14 ; SSE-NEXT: por %xmm4, %xmm14 ; SSE-NEXT: movdqa 16(%rdx), %xmm3 ; SSE-NEXT: movdqa 16(%rcx), %xmm8 ; SSE-NEXT: movdqa %xmm3, %xmm12 ; SSE-NEXT: punpcklbw {{.*#+}} xmm12 = 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] ; SSE-NEXT: pshuflw {{.*#+}} xmm4 = xmm12[0,0,2,1,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm4 = xmm4[0,1,1,3] ; SSE-NEXT: movdqa %xmm1, %xmm15 ; SSE-NEXT: pandn %xmm4, %xmm15 ; SSE-NEXT: movdqa 16(%rdi), %xmm4 ; SSE-NEXT: movdqa 16(%rsi), %xmm9 ; SSE-NEXT: movdqa %xmm4, %xmm2 ; SSE-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm9[0],xmm2[1],xmm9[1],xmm2[2],xmm9[2],xmm2[3],xmm9[3],xmm2[4],xmm9[4],xmm2[5],xmm9[5],xmm2[6],xmm9[6],xmm2[7],xmm9[7] ; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm2[0,0,0,0] ; SSE-NEXT: pshufhw {{.*#+}} xmm7 = xmm7[0,1,2,3,5,5,5,5] ; SSE-NEXT: pand %xmm1, %xmm7 ; SSE-NEXT: por %xmm15, %xmm7 ; SSE-NEXT: pshufd {{.*#+}} xmm14 = xmm14[1,3,2,3] ; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm7[0,2,2,3] ; SSE-NEXT: punpckldq {{.*#+}} xmm7 = xmm7[0],xmm14[0],xmm7[1],xmm14[1] ; SSE-NEXT: movdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: pshuflw {{.*#+}} xmm7 = xmm10[0,2,2,3,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm7[0,0,2,1] ; SSE-NEXT: movdqa %xmm13, %xmm14 ; SSE-NEXT: pandn %xmm7, %xmm14 ; SSE-NEXT: pshuflw {{.*#+}} xmm7 = xmm11[2,1,3,3,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm7[0,0,2,1] ; SSE-NEXT: pand %xmm13, %xmm7 ; SSE-NEXT: por %xmm14, %xmm7 ; SSE-NEXT: pshuflw {{.*#+}} xmm14 = xmm12[0,2,2,3,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm14 = xmm14[0,1,1,3] ; SSE-NEXT: movdqa %xmm1, %xmm15 ; SSE-NEXT: pandn %xmm14, %xmm15 ; SSE-NEXT: pshufd {{.*#+}} xmm14 = xmm2[1,1,1,1] ; SSE-NEXT: pshufhw {{.*#+}} xmm14 = xmm14[0,1,2,3,5,5,5,5] ; SSE-NEXT: pand %xmm1, %xmm14 ; SSE-NEXT: por %xmm15, %xmm14 ; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm7[1,3,2,3] ; SSE-NEXT: pshufd {{.*#+}} xmm14 = xmm14[0,2,2,3] ; SSE-NEXT: punpckldq {{.*#+}} xmm14 = xmm14[0],xmm7[0],xmm14[1],xmm7[1] ; SSE-NEXT: movdqa %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: pshufhw {{.*#+}} xmm7 = xmm10[0,1,2,3,4,4,6,5] ; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm7[0,2,2,3] ; SSE-NEXT: movdqa %xmm13, %xmm14 ; SSE-NEXT: pandn %xmm7, %xmm14 ; SSE-NEXT: pshufhw {{.*#+}} xmm7 = xmm11[0,1,2,3,4,5,5,7] ; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm7[0,2,2,3] ; SSE-NEXT: pand %xmm13, %xmm7 ; SSE-NEXT: por %xmm14, %xmm7 ; SSE-NEXT: pshufhw {{.*#+}} xmm14 = xmm12[0,1,2,3,4,4,6,5] ; SSE-NEXT: pshufd {{.*#+}} xmm14 = xmm14[2,1,3,3] ; SSE-NEXT: movdqa %xmm1, %xmm15 ; SSE-NEXT: pandn %xmm14, %xmm15 ; SSE-NEXT: pshufd {{.*#+}} xmm14 = xmm2[2,2,2,2] ; SSE-NEXT: pshufhw {{.*#+}} xmm14 = xmm14[0,1,2,3,5,5,5,5] ; SSE-NEXT: pand %xmm1, %xmm14 ; SSE-NEXT: por %xmm15, %xmm14 ; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm7[1,3,2,3] ; SSE-NEXT: pshufd {{.*#+}} xmm14 = xmm14[0,2,2,3] ; SSE-NEXT: punpckldq {{.*#+}} xmm14 = xmm14[0],xmm7[0],xmm14[1],xmm7[1] ; SSE-NEXT: movdqa %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: pshufhw {{.*#+}} xmm7 = xmm10[0,1,2,3,4,6,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm7[0,2,2,3] ; SSE-NEXT: movdqa %xmm13, %xmm10 ; SSE-NEXT: pandn %xmm7, %xmm10 ; SSE-NEXT: pshufhw {{.*#+}} xmm7 = xmm11[0,1,2,3,6,5,7,7] ; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm7[0,2,2,3] ; SSE-NEXT: pand %xmm13, %xmm7 ; SSE-NEXT: por %xmm10, %xmm7 ; SSE-NEXT: pshufhw {{.*#+}} xmm10 = xmm12[0,1,2,3,4,6,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm10 = xmm10[2,1,3,3] ; SSE-NEXT: movdqa %xmm1, %xmm11 ; SSE-NEXT: pandn %xmm10, %xmm11 ; SSE-NEXT: pshufd {{.*#+}} xmm10 = xmm2[3,3,3,3] ; SSE-NEXT: pshufhw {{.*#+}} xmm10 = xmm10[0,1,2,3,5,5,5,5] ; SSE-NEXT: pand %xmm1, %xmm10 ; SSE-NEXT: por %xmm11, %xmm10 ; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm7[1,3,2,3] ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm10[0,2,2,3] ; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm7[0],xmm2[1],xmm7[1] ; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: punpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Folded Reload ; SSE-NEXT: # xmm5 = xmm5[8],mem[8],xmm5[9],mem[9],xmm5[10],mem[10],xmm5[11],mem[11],xmm5[12],mem[12],xmm5[13],mem[13],xmm5[14],mem[14],xmm5[15],mem[15] ; SSE-NEXT: pshuflw {{.*#+}} xmm7 = xmm5[0,0,2,1,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm7[0,0,2,1] ; SSE-NEXT: movdqa %xmm13, %xmm10 ; SSE-NEXT: pandn %xmm7, %xmm10 ; SSE-NEXT: punpckhbw {{.*#+}} xmm0 = xmm0[8],xmm6[8],xmm0[9],xmm6[9],xmm0[10],xmm6[10],xmm0[11],xmm6[11],xmm0[12],xmm6[12],xmm0[13],xmm6[13],xmm0[14],xmm6[14],xmm0[15],xmm6[15] ; SSE-NEXT: pshuflw {{.*#+}} xmm6 = xmm0[0,1,1,3,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm6[0,0,2,1] ; SSE-NEXT: pand %xmm13, %xmm6 ; SSE-NEXT: por %xmm10, %xmm6 ; SSE-NEXT: punpckhbw {{.*#+}} xmm3 = xmm3[8],xmm8[8],xmm3[9],xmm8[9],xmm3[10],xmm8[10],xmm3[11],xmm8[11],xmm3[12],xmm8[12],xmm3[13],xmm8[13],xmm3[14],xmm8[14],xmm3[15],xmm8[15] ; SSE-NEXT: pshuflw {{.*#+}} xmm7 = xmm3[0,0,2,1,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm7[0,1,1,3] ; SSE-NEXT: movdqa %xmm1, %xmm8 ; SSE-NEXT: pandn %xmm7, %xmm8 ; 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: pshufd {{.*#+}} xmm7 = xmm4[0,0,0,0] ; SSE-NEXT: pshufhw {{.*#+}} xmm7 = xmm7[0,1,2,3,5,5,5,5] ; SSE-NEXT: pand %xmm1, %xmm7 ; SSE-NEXT: por %xmm8, %xmm7 ; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm6[1,3,2,3] ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm7[0,2,2,3] ; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm6[0],xmm2[1],xmm6[1] ; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: pshuflw {{.*#+}} xmm6 = xmm5[0,2,2,3,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm6[0,0,2,1] ; SSE-NEXT: movdqa %xmm13, %xmm7 ; SSE-NEXT: pandn %xmm6, %xmm7 ; SSE-NEXT: pshuflw {{.*#+}} xmm6 = xmm0[2,1,3,3,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm6[0,0,2,1] ; SSE-NEXT: pand %xmm13, %xmm6 ; SSE-NEXT: por %xmm7, %xmm6 ; SSE-NEXT: pshuflw {{.*#+}} xmm7 = xmm3[0,2,2,3,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm7[0,1,1,3] ; SSE-NEXT: movdqa %xmm1, %xmm8 ; SSE-NEXT: pandn %xmm7, %xmm8 ; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm4[1,1,1,1] ; SSE-NEXT: pshufhw {{.*#+}} xmm7 = xmm7[0,1,2,3,5,5,5,5] ; SSE-NEXT: pand %xmm1, %xmm7 ; SSE-NEXT: por %xmm8, %xmm7 ; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm6[1,3,2,3] ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm7[0,2,2,3] ; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm6[0],xmm2[1],xmm6[1] ; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: pshufhw {{.*#+}} xmm6 = xmm5[0,1,2,3,4,4,6,5] ; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm6[0,2,2,3] ; SSE-NEXT: movdqa %xmm13, %xmm7 ; SSE-NEXT: pandn %xmm6, %xmm7 ; SSE-NEXT: pshufhw {{.*#+}} xmm6 = xmm0[0,1,2,3,4,5,5,7] ; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm6[0,2,2,3] ; SSE-NEXT: pand %xmm13, %xmm6 ; SSE-NEXT: por %xmm7, %xmm6 ; SSE-NEXT: pshufhw {{.*#+}} xmm7 = xmm3[0,1,2,3,4,4,6,5] ; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm7[2,1,3,3] ; SSE-NEXT: movdqa %xmm1, %xmm8 ; SSE-NEXT: pandn %xmm7, %xmm8 ; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm4[2,2,2,2] ; SSE-NEXT: pshufhw {{.*#+}} xmm7 = xmm7[0,1,2,3,5,5,5,5] ; SSE-NEXT: pand %xmm1, %xmm7 ; SSE-NEXT: por %xmm8, %xmm7 ; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm6[1,3,2,3] ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm7[0,2,2,3] ; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm6[0],xmm2[1],xmm6[1] ; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: pshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,4,6,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm5[0,2,2,3] ; SSE-NEXT: movdqa %xmm13, %xmm6 ; SSE-NEXT: pandn %xmm5, %xmm6 ; SSE-NEXT: pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,6,5,7,7] ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3] ; SSE-NEXT: pand %xmm13, %xmm0 ; SSE-NEXT: por %xmm6, %xmm0 ; SSE-NEXT: pshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,6,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm3[2,1,3,3] ; SSE-NEXT: movdqa %xmm1, %xmm5 ; SSE-NEXT: pandn %xmm3, %xmm5 ; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm4[3,3,3,3] ; SSE-NEXT: pshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,5,5,5,5] ; SSE-NEXT: pand %xmm1, %xmm3 ; SSE-NEXT: por %xmm5, %xmm3 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,3,2,3] ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm3[0,2,2,3] ; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1] ; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: movdqa 32(%r10), %xmm0 ; SSE-NEXT: movdqa 32(%rax), %xmm2 ; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: movdqa %xmm0, %xmm10 ; SSE-NEXT: punpcklbw {{.*#+}} xmm10 = xmm10[0],xmm2[0],xmm10[1],xmm2[1],xmm10[2],xmm2[2],xmm10[3],xmm2[3],xmm10[4],xmm2[4],xmm10[5],xmm2[5],xmm10[6],xmm2[6],xmm10[7],xmm2[7] ; SSE-NEXT: pshuflw {{.*#+}} xmm3 = xmm10[0,0,2,1,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm3[0,0,2,1] ; SSE-NEXT: movdqa %xmm13, %xmm4 ; SSE-NEXT: pandn %xmm3, %xmm4 ; SSE-NEXT: movdqa 32(%r8), %xmm3 ; SSE-NEXT: movdqa 32(%r9), %xmm7 ; SSE-NEXT: movdqa %xmm3, %xmm11 ; SSE-NEXT: punpcklbw {{.*#+}} xmm11 = xmm11[0],xmm7[0],xmm11[1],xmm7[1],xmm11[2],xmm7[2],xmm11[3],xmm7[3],xmm11[4],xmm7[4],xmm11[5],xmm7[5],xmm11[6],xmm7[6],xmm11[7],xmm7[7] ; SSE-NEXT: pshuflw {{.*#+}} xmm5 = xmm11[0,1,1,3,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm14 = xmm5[0,0,2,1] ; SSE-NEXT: pand %xmm13, %xmm14 ; SSE-NEXT: por %xmm4, %xmm14 ; SSE-NEXT: movdqa 32(%rdx), %xmm4 ; SSE-NEXT: movdqa 32(%rcx), %xmm8 ; SSE-NEXT: movdqa %xmm4, %xmm12 ; SSE-NEXT: punpcklbw {{.*#+}} xmm12 = 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] ; SSE-NEXT: pshuflw {{.*#+}} xmm5 = xmm12[0,0,2,1,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm5[0,1,1,3] ; SSE-NEXT: movdqa %xmm1, %xmm15 ; SSE-NEXT: pandn %xmm5, %xmm15 ; SSE-NEXT: movdqa 32(%rdi), %xmm5 ; SSE-NEXT: movdqa 32(%rsi), %xmm9 ; SSE-NEXT: movdqa %xmm5, %xmm2 ; SSE-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm9[0],xmm2[1],xmm9[1],xmm2[2],xmm9[2],xmm2[3],xmm9[3],xmm2[4],xmm9[4],xmm2[5],xmm9[5],xmm2[6],xmm9[6],xmm2[7],xmm9[7] ; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm2[0,0,0,0] ; SSE-NEXT: pshufhw {{.*#+}} xmm6 = xmm6[0,1,2,3,5,5,5,5] ; SSE-NEXT: pand %xmm1, %xmm6 ; SSE-NEXT: por %xmm15, %xmm6 ; SSE-NEXT: pshufd {{.*#+}} xmm14 = xmm14[1,3,2,3] ; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm6[0,2,2,3] ; SSE-NEXT: punpckldq {{.*#+}} xmm6 = xmm6[0],xmm14[0],xmm6[1],xmm14[1] ; SSE-NEXT: movdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: pshuflw {{.*#+}} xmm6 = xmm10[0,2,2,3,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm6[0,0,2,1] ; SSE-NEXT: movdqa %xmm13, %xmm14 ; SSE-NEXT: pandn %xmm6, %xmm14 ; SSE-NEXT: pshuflw {{.*#+}} xmm6 = xmm11[2,1,3,3,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm6[0,0,2,1] ; SSE-NEXT: pand %xmm13, %xmm6 ; SSE-NEXT: por %xmm14, %xmm6 ; SSE-NEXT: pshuflw {{.*#+}} xmm14 = xmm12[0,2,2,3,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm14 = xmm14[0,1,1,3] ; SSE-NEXT: movdqa %xmm1, %xmm15 ; SSE-NEXT: pandn %xmm14, %xmm15 ; SSE-NEXT: pshufd {{.*#+}} xmm14 = xmm2[1,1,1,1] ; SSE-NEXT: pshufhw {{.*#+}} xmm14 = xmm14[0,1,2,3,5,5,5,5] ; SSE-NEXT: pand %xmm1, %xmm14 ; SSE-NEXT: por %xmm15, %xmm14 ; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm6[1,3,2,3] ; SSE-NEXT: pshufd {{.*#+}} xmm14 = xmm14[0,2,2,3] ; SSE-NEXT: punpckldq {{.*#+}} xmm14 = xmm14[0],xmm6[0],xmm14[1],xmm6[1] ; SSE-NEXT: movdqa %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: pshufhw {{.*#+}} xmm6 = xmm10[0,1,2,3,4,4,6,5] ; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm6[0,2,2,3] ; SSE-NEXT: movdqa %xmm13, %xmm14 ; SSE-NEXT: pandn %xmm6, %xmm14 ; SSE-NEXT: pshufhw {{.*#+}} xmm6 = xmm11[0,1,2,3,4,5,5,7] ; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm6[0,2,2,3] ; SSE-NEXT: pand %xmm13, %xmm6 ; SSE-NEXT: por %xmm14, %xmm6 ; SSE-NEXT: pshufhw {{.*#+}} xmm14 = xmm12[0,1,2,3,4,4,6,5] ; SSE-NEXT: pshufd {{.*#+}} xmm14 = xmm14[2,1,3,3] ; SSE-NEXT: movdqa %xmm1, %xmm15 ; SSE-NEXT: pandn %xmm14, %xmm15 ; SSE-NEXT: pshufd {{.*#+}} xmm14 = xmm2[2,2,2,2] ; SSE-NEXT: pshufhw {{.*#+}} xmm14 = xmm14[0,1,2,3,5,5,5,5] ; SSE-NEXT: pand %xmm1, %xmm14 ; SSE-NEXT: por %xmm15, %xmm14 ; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm6[1,3,2,3] ; SSE-NEXT: pshufd {{.*#+}} xmm14 = xmm14[0,2,2,3] ; SSE-NEXT: punpckldq {{.*#+}} xmm14 = xmm14[0],xmm6[0],xmm14[1],xmm6[1] ; SSE-NEXT: movdqa %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: pshufhw {{.*#+}} xmm6 = xmm10[0,1,2,3,4,6,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm6[0,2,2,3] ; SSE-NEXT: movdqa %xmm13, %xmm10 ; SSE-NEXT: pandn %xmm6, %xmm10 ; SSE-NEXT: pshufhw {{.*#+}} xmm6 = xmm11[0,1,2,3,6,5,7,7] ; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm6[0,2,2,3] ; SSE-NEXT: pand %xmm13, %xmm6 ; SSE-NEXT: por %xmm10, %xmm6 ; SSE-NEXT: pshufhw {{.*#+}} xmm10 = xmm12[0,1,2,3,4,6,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm10 = xmm10[2,1,3,3] ; SSE-NEXT: movdqa %xmm1, %xmm11 ; SSE-NEXT: pandn %xmm10, %xmm11 ; SSE-NEXT: pshufd {{.*#+}} xmm10 = xmm2[3,3,3,3] ; SSE-NEXT: pshufhw {{.*#+}} xmm10 = xmm10[0,1,2,3,5,5,5,5] ; SSE-NEXT: pand %xmm1, %xmm10 ; SSE-NEXT: por %xmm11, %xmm10 ; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm6[1,3,2,3] ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm10[0,2,2,3] ; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm6[0],xmm2[1],xmm6[1] ; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: punpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload ; SSE-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] ; SSE-NEXT: pshuflw {{.*#+}} xmm6 = xmm0[0,0,2,1,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm6[0,0,2,1] ; SSE-NEXT: movdqa %xmm13, %xmm10 ; SSE-NEXT: pandn %xmm6, %xmm10 ; SSE-NEXT: punpckhbw {{.*#+}} xmm3 = xmm3[8],xmm7[8],xmm3[9],xmm7[9],xmm3[10],xmm7[10],xmm3[11],xmm7[11],xmm3[12],xmm7[12],xmm3[13],xmm7[13],xmm3[14],xmm7[14],xmm3[15],xmm7[15] ; SSE-NEXT: pshuflw {{.*#+}} xmm6 = xmm3[0,1,1,3,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm6[0,0,2,1] ; SSE-NEXT: pand %xmm13, %xmm6 ; SSE-NEXT: por %xmm10, %xmm6 ; SSE-NEXT: punpckhbw {{.*#+}} xmm4 = xmm4[8],xmm8[8],xmm4[9],xmm8[9],xmm4[10],xmm8[10],xmm4[11],xmm8[11],xmm4[12],xmm8[12],xmm4[13],xmm8[13],xmm4[14],xmm8[14],xmm4[15],xmm8[15] ; SSE-NEXT: pshuflw {{.*#+}} xmm7 = xmm4[0,0,2,1,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm7[0,1,1,3] ; SSE-NEXT: movdqa %xmm1, %xmm8 ; SSE-NEXT: pandn %xmm7, %xmm8 ; SSE-NEXT: punpckhbw {{.*#+}} xmm5 = xmm5[8],xmm9[8],xmm5[9],xmm9[9],xmm5[10],xmm9[10],xmm5[11],xmm9[11],xmm5[12],xmm9[12],xmm5[13],xmm9[13],xmm5[14],xmm9[14],xmm5[15],xmm9[15] ; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm5[0,0,0,0] ; SSE-NEXT: pshufhw {{.*#+}} xmm7 = xmm7[0,1,2,3,5,5,5,5] ; SSE-NEXT: pand %xmm1, %xmm7 ; SSE-NEXT: por %xmm8, %xmm7 ; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm6[1,3,2,3] ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm7[0,2,2,3] ; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm6[0],xmm2[1],xmm6[1] ; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: pshuflw {{.*#+}} xmm6 = xmm0[0,2,2,3,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm6[0,0,2,1] ; SSE-NEXT: movdqa %xmm13, %xmm7 ; SSE-NEXT: pandn %xmm6, %xmm7 ; SSE-NEXT: pshuflw {{.*#+}} xmm6 = xmm3[2,1,3,3,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm6[0,0,2,1] ; SSE-NEXT: pand %xmm13, %xmm6 ; SSE-NEXT: por %xmm7, %xmm6 ; SSE-NEXT: pshuflw {{.*#+}} xmm7 = xmm4[0,2,2,3,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm7[0,1,1,3] ; SSE-NEXT: movdqa %xmm1, %xmm8 ; SSE-NEXT: pandn %xmm7, %xmm8 ; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm5[1,1,1,1] ; SSE-NEXT: pshufhw {{.*#+}} xmm7 = xmm7[0,1,2,3,5,5,5,5] ; SSE-NEXT: pand %xmm1, %xmm7 ; SSE-NEXT: por %xmm8, %xmm7 ; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm6[1,3,2,3] ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm7[0,2,2,3] ; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm6[0],xmm2[1],xmm6[1] ; SSE-NEXT: movdqa %xmm2, (%rsp) # 16-byte Spill ; SSE-NEXT: pshufhw {{.*#+}} xmm6 = xmm0[0,1,2,3,4,4,6,5] ; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm6[0,2,2,3] ; SSE-NEXT: movdqa %xmm13, %xmm7 ; SSE-NEXT: pandn %xmm6, %xmm7 ; SSE-NEXT: pshufhw {{.*#+}} xmm6 = xmm3[0,1,2,3,4,5,5,7] ; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm6[0,2,2,3] ; SSE-NEXT: pand %xmm13, %xmm6 ; SSE-NEXT: por %xmm7, %xmm6 ; SSE-NEXT: pshufhw {{.*#+}} xmm7 = xmm4[0,1,2,3,4,4,6,5] ; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm7[2,1,3,3] ; SSE-NEXT: movdqa %xmm1, %xmm8 ; SSE-NEXT: pandn %xmm7, %xmm8 ; SSE-NEXT: pshufd {{.*#+}} xmm7 = xmm5[2,2,2,2] ; SSE-NEXT: pshufhw {{.*#+}} xmm7 = xmm7[0,1,2,3,5,5,5,5] ; SSE-NEXT: pand %xmm1, %xmm7 ; SSE-NEXT: por %xmm8, %xmm7 ; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm6[1,3,2,3] ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm7[0,2,2,3] ; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm6[0],xmm2[1],xmm6[1] ; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,6,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3] ; SSE-NEXT: movdqa %xmm13, %xmm6 ; SSE-NEXT: pandn %xmm0, %xmm6 ; SSE-NEXT: pshufhw {{.*#+}} xmm0 = xmm3[0,1,2,3,6,5,7,7] ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3] ; SSE-NEXT: pand %xmm13, %xmm0 ; SSE-NEXT: por %xmm6, %xmm0 ; SSE-NEXT: pshufhw {{.*#+}} xmm3 = xmm4[0,1,2,3,4,6,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm3[2,1,3,3] ; SSE-NEXT: movdqa %xmm1, %xmm4 ; SSE-NEXT: pandn %xmm3, %xmm4 ; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm5[3,3,3,3] ; SSE-NEXT: pshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,5,5,5,5] ; SSE-NEXT: pand %xmm1, %xmm3 ; SSE-NEXT: por %xmm4, %xmm3 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,3,2,3] ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm3[0,2,2,3] ; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1] ; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: movdqa 48(%r10), %xmm9 ; SSE-NEXT: movdqa 48(%rax), %xmm0 ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: movdqa %xmm9, %xmm6 ; SSE-NEXT: punpcklbw {{.*#+}} xmm6 = xmm6[0],xmm0[0],xmm6[1],xmm0[1],xmm6[2],xmm0[2],xmm6[3],xmm0[3],xmm6[4],xmm0[4],xmm6[5],xmm0[5],xmm6[6],xmm0[6],xmm6[7],xmm0[7] ; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm6[0,0,2,1,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,0,2,1] ; SSE-NEXT: movdqa %xmm13, %xmm3 ; SSE-NEXT: pandn %xmm0, %xmm3 ; SSE-NEXT: movdqa 48(%r8), %xmm8 ; SSE-NEXT: movdqa 48(%r9), %xmm0 ; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: movdqa %xmm8, %xmm4 ; SSE-NEXT: punpcklbw {{.*#+}} xmm4 = xmm4[0],xmm0[0],xmm4[1],xmm0[1],xmm4[2],xmm0[2],xmm4[3],xmm0[3],xmm4[4],xmm0[4],xmm4[5],xmm0[5],xmm4[6],xmm0[6],xmm4[7],xmm0[7] ; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm4[0,1,1,3,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm10 = xmm0[0,0,2,1] ; SSE-NEXT: pand %xmm13, %xmm10 ; SSE-NEXT: por %xmm3, %xmm10 ; SSE-NEXT: movdqa 48(%rdx), %xmm7 ; SSE-NEXT: movdqa 48(%rcx), %xmm12 ; SSE-NEXT: movdqa %xmm7, %xmm3 ; SSE-NEXT: punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm12[0],xmm3[1],xmm12[1],xmm3[2],xmm12[2],xmm3[3],xmm12[3],xmm3[4],xmm12[4],xmm3[5],xmm12[5],xmm3[6],xmm12[6],xmm3[7],xmm12[7] ; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm3[0,0,2,1,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,1,1,3] ; SSE-NEXT: movdqa %xmm1, %xmm15 ; SSE-NEXT: pandn %xmm0, %xmm15 ; SSE-NEXT: movdqa 48(%rdi), %xmm5 ; SSE-NEXT: movdqa 48(%rsi), %xmm11 ; SSE-NEXT: movdqa %xmm5, %xmm0 ; SSE-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm11[0],xmm0[1],xmm11[1],xmm0[2],xmm11[2],xmm0[3],xmm11[3],xmm0[4],xmm11[4],xmm0[5],xmm11[5],xmm0[6],xmm11[6],xmm0[7],xmm11[7] ; SSE-NEXT: pshufd {{.*#+}} xmm14 = xmm0[0,0,0,0] ; SSE-NEXT: pshufhw {{.*#+}} xmm14 = xmm14[0,1,2,3,5,5,5,5] ; SSE-NEXT: pand %xmm1, %xmm14 ; SSE-NEXT: por %xmm15, %xmm14 ; SSE-NEXT: pshufd {{.*#+}} xmm10 = xmm10[1,3,2,3] ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm14[0,2,2,3] ; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm10[0],xmm2[1],xmm10[1] ; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; SSE-NEXT: pshuflw {{.*#+}} xmm10 = xmm6[0,2,2,3,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm10 = xmm10[0,0,2,1] ; SSE-NEXT: movdqa %xmm13, %xmm14 ; SSE-NEXT: pandn %xmm10, %xmm14 ; SSE-NEXT: pshuflw {{.*#+}} xmm10 = xmm4[2,1,3,3,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm10 = xmm10[0,0,2,1] ; SSE-NEXT: pand %xmm13, %xmm10 ; SSE-NEXT: por %xmm14, %xmm10 ; SSE-NEXT: pshuflw {{.*#+}} xmm14 = xmm3[0,2,2,3,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm14 = xmm14[0,1,1,3] ; SSE-NEXT: movdqa %xmm1, %xmm15 ; SSE-NEXT: pandn %xmm14, %xmm15 ; SSE-NEXT: pshufd {{.*#+}} xmm14 = xmm0[1,1,1,1] ; SSE-NEXT: pshufhw {{.*#+}} xmm14 = xmm14[0,1,2,3,5,5,5,5] ; SSE-NEXT: pand %xmm1, %xmm14 ; SSE-NEXT: por %xmm15, %xmm14 ; SSE-NEXT: pshufd {{.*#+}} xmm10 = xmm10[1,3,2,3] ; SSE-NEXT: pshufd {{.*#+}} xmm15 = xmm14[0,2,2,3] ; SSE-NEXT: punpckldq {{.*#+}} xmm15 = xmm15[0],xmm10[0],xmm15[1],xmm10[1] ; SSE-NEXT: pshufhw {{.*#+}} xmm10 = xmm6[0,1,2,3,4,4,6,5] ; SSE-NEXT: pshufd {{.*#+}} xmm10 = xmm10[0,2,2,3] ; SSE-NEXT: movdqa %xmm13, %xmm14 ; SSE-NEXT: pandn %xmm10, %xmm14 ; SSE-NEXT: pshufhw {{.*#+}} xmm10 = xmm4[0,1,2,3,4,5,5,7] ; SSE-NEXT: pshufd {{.*#+}} xmm10 = xmm10[0,2,2,3] ; SSE-NEXT: pand %xmm13, %xmm10 ; SSE-NEXT: por %xmm14, %xmm10 ; SSE-NEXT: pshufhw {{.*#+}} xmm14 = xmm3[0,1,2,3,4,4,6,5] ; SSE-NEXT: pshufd {{.*#+}} xmm14 = xmm14[2,1,3,3] ; SSE-NEXT: movdqa %xmm1, %xmm2 ; SSE-NEXT: pandn %xmm14, %xmm2 ; SSE-NEXT: pshufd {{.*#+}} xmm14 = xmm0[2,2,2,2] ; SSE-NEXT: pshufhw {{.*#+}} xmm14 = xmm14[0,1,2,3,5,5,5,5] ; SSE-NEXT: pand %xmm1, %xmm14 ; SSE-NEXT: por %xmm2, %xmm14 ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm10[1,3,2,3] ; SSE-NEXT: pshufd {{.*#+}} xmm10 = xmm14[0,2,2,3] ; SSE-NEXT: punpckldq {{.*#+}} xmm10 = xmm10[0],xmm2[0],xmm10[1],xmm2[1] ; SSE-NEXT: pshufhw {{.*#+}} xmm2 = xmm6[0,1,2,3,4,6,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm2[0,2,2,3] ; SSE-NEXT: movdqa %xmm13, %xmm6 ; SSE-NEXT: pandn %xmm2, %xmm6 ; SSE-NEXT: pshufhw {{.*#+}} xmm2 = xmm4[0,1,2,3,6,5,7,7] ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm2[0,2,2,3] ; SSE-NEXT: pand %xmm13, %xmm2 ; SSE-NEXT: por %xmm6, %xmm2 ; SSE-NEXT: pshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,6,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm3[2,1,3,3] ; SSE-NEXT: movdqa %xmm1, %xmm4 ; SSE-NEXT: pandn %xmm3, %xmm4 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[3,3,3,3] ; SSE-NEXT: pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,5,5,5,5] ; SSE-NEXT: pand %xmm1, %xmm0 ; SSE-NEXT: por %xmm4, %xmm0 ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm2[1,3,2,3] ; SSE-NEXT: pshufd {{.*#+}} xmm6 = xmm0[0,2,2,3] ; SSE-NEXT: punpckldq {{.*#+}} xmm6 = xmm6[0],xmm2[0],xmm6[1],xmm2[1] ; 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[0,0,2,1,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,0,2,1] ; SSE-NEXT: movdqa %xmm13, %xmm2 ; SSE-NEXT: pandn %xmm0, %xmm2 ; SSE-NEXT: punpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Folded Reload ; SSE-NEXT: # xmm8 = xmm8[8],mem[8],xmm8[9],mem[9],xmm8[10],mem[10],xmm8[11],mem[11],xmm8[12],mem[12],xmm8[13],mem[13],xmm8[14],mem[14],xmm8[15],mem[15] ; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm8[0,1,1,3,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,0,2,1] ; SSE-NEXT: pand %xmm13, %xmm0 ; SSE-NEXT: por %xmm2, %xmm0 ; SSE-NEXT: punpckhbw {{.*#+}} xmm7 = xmm7[8],xmm12[8],xmm7[9],xmm12[9],xmm7[10],xmm12[10],xmm7[11],xmm12[11],xmm7[12],xmm12[12],xmm7[13],xmm12[13],xmm7[14],xmm12[14],xmm7[15],xmm12[15] ; SSE-NEXT: pshuflw {{.*#+}} xmm2 = xmm7[0,0,2,1,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm2[0,1,1,3] ; SSE-NEXT: movdqa %xmm1, %xmm3 ; SSE-NEXT: pandn %xmm2, %xmm3 ; SSE-NEXT: punpckhbw {{.*#+}} xmm5 = xmm5[8],xmm11[8],xmm5[9],xmm11[9],xmm5[10],xmm11[10],xmm5[11],xmm11[11],xmm5[12],xmm11[12],xmm5[13],xmm11[13],xmm5[14],xmm11[14],xmm5[15],xmm11[15] ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm5[0,0,0,0] ; SSE-NEXT: pshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,5,5,5,5] ; SSE-NEXT: pand %xmm1, %xmm2 ; SSE-NEXT: por %xmm3, %xmm2 ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,3,2,3] ; SSE-NEXT: pshufd {{.*#+}} xmm4 = xmm2[0,2,2,3] ; SSE-NEXT: punpckldq {{.*#+}} xmm4 = xmm4[0],xmm0[0],xmm4[1],xmm0[1] ; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm9[0,2,2,3,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,0,2,1] ; SSE-NEXT: movdqa %xmm13, %xmm2 ; SSE-NEXT: pandn %xmm0, %xmm2 ; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm8[2,1,3,3,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,0,2,1] ; SSE-NEXT: pand %xmm13, %xmm0 ; SSE-NEXT: por %xmm2, %xmm0 ; SSE-NEXT: pshuflw {{.*#+}} xmm2 = xmm7[0,2,2,3,4,5,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm2[0,1,1,3] ; SSE-NEXT: movdqa %xmm1, %xmm3 ; SSE-NEXT: pandn %xmm2, %xmm3 ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm5[1,1,1,1] ; SSE-NEXT: pshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,5,5,5,5] ; SSE-NEXT: pand %xmm1, %xmm2 ; SSE-NEXT: por %xmm3, %xmm2 ; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm0[1,3,2,3] ; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm2[0,2,2,3] ; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1] ; SSE-NEXT: pshufhw {{.*#+}} xmm2 = xmm9[0,1,2,3,4,4,6,5] ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm2[0,2,2,3] ; SSE-NEXT: movdqa %xmm13, %xmm3 ; SSE-NEXT: pandn %xmm2, %xmm3 ; SSE-NEXT: pshufhw {{.*#+}} xmm2 = xmm8[0,1,2,3,4,5,5,7] ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm2[0,2,2,3] ; SSE-NEXT: pand %xmm13, %xmm2 ; SSE-NEXT: por %xmm3, %xmm2 ; SSE-NEXT: pshufhw {{.*#+}} xmm3 = xmm7[0,1,2,3,4,4,6,5] ; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm3[2,1,3,3] ; SSE-NEXT: movdqa %xmm1, %xmm11 ; SSE-NEXT: pandn %xmm3, %xmm11 ; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm5[2,2,2,2] ; SSE-NEXT: pshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,5,5,5,5] ; SSE-NEXT: pand %xmm1, %xmm3 ; SSE-NEXT: por %xmm11, %xmm3 ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm2[1,3,2,3] ; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm3[0,2,2,3] ; SSE-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm2[0],xmm3[1],xmm2[1] ; SSE-NEXT: pshufhw {{.*#+}} xmm2 = xmm9[0,1,2,3,4,6,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm2[0,2,2,3] ; SSE-NEXT: pshufhw {{.*#+}} xmm8 = xmm8[0,1,2,3,6,5,7,7] ; SSE-NEXT: pshufd {{.*#+}} xmm8 = xmm8[0,2,2,3] ; SSE-NEXT: pand %xmm13, %xmm8 ; SSE-NEXT: pandn %xmm2, %xmm13 ; SSE-NEXT: por %xmm8, %xmm13 ; SSE-NEXT: pshufhw {{.*#+}} xmm2 = xmm7[0,1,2,3,4,6,6,7] ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm2[2,1,3,3] ; SSE-NEXT: pshufd {{.*#+}} xmm5 = xmm5[3,3,3,3] ; SSE-NEXT: pshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,5,5,5,5] ; SSE-NEXT: pand %xmm1, %xmm5 ; SSE-NEXT: pandn %xmm2, %xmm1 ; SSE-NEXT: por %xmm5, %xmm1 ; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm13[1,3,2,3] ; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,2,2,3] ; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1] ; SSE-NEXT: movq {{[0-9]+}}(%rsp), %rax ; SSE-NEXT: movdqa %xmm1, 496(%rax) ; SSE-NEXT: movdqa %xmm3, 480(%rax) ; SSE-NEXT: movdqa %xmm0, 464(%rax) ; SSE-NEXT: movdqa %xmm4, 448(%rax) ; SSE-NEXT: movdqa %xmm6, 432(%rax) ; SSE-NEXT: movdqa %xmm10, 416(%rax) ; SSE-NEXT: movdqa %xmm15, 400(%rax) ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload ; SSE-NEXT: movaps %xmm0, 384(%rax) ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload ; SSE-NEXT: movaps %xmm0, 368(%rax) ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload ; SSE-NEXT: movaps %xmm0, 352(%rax) ; SSE-NEXT: movaps (%rsp), %xmm0 # 16-byte Reload ; SSE-NEXT: movaps %xmm0, 336(%rax) ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload ; SSE-NEXT: movaps %xmm0, 320(%rax) ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload ; SSE-NEXT: movaps %xmm0, 304(%rax) ; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload ; SSE-NEXT: movaps %xmm0, 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 {{[-0-9]+}}(%r{{[sb]}}p), %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 $312, %rsp # imm = 0x138 ; SSE-NEXT: retq ; ; AVX-LABEL: store_i8_stride8_vf64: ; AVX: # %bb.0: ; AVX-NEXT: subq $328, %rsp # imm = 0x148 ; AVX-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX-NEXT: vmovdqa (%r10), %xmm1 ; AVX-NEXT: vmovdqa %xmm1, (%rsp) # 16-byte Spill ; AVX-NEXT: vmovdqa (%rax), %xmm0 ; AVX-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX-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] ; AVX-NEXT: vpshufhw {{.*#+}} xmm1 = xmm0[0,1,2,3,4,4,6,5] ; AVX-NEXT: vpshufhw {{.*#+}} xmm2 = xmm0[0,1,2,3,4,6,6,7] ; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm1, %ymm1 ; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm1[0,2,2,3,4,6,6,7] ; AVX-NEXT: vbroadcastsd {{.*#+}} ymm9 = [65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0] ; AVX-NEXT: vandnps %ymm1, %ymm9, %ymm2 ; AVX-NEXT: vmovdqa (%r9), %xmm3 ; AVX-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX-NEXT: vmovdqa (%r8), %xmm1 ; AVX-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX-NEXT: vpunpcklbw {{.*#+}} 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] ; AVX-NEXT: vpshufhw {{.*#+}} xmm3 = xmm1[0,1,2,3,4,5,5,7] ; AVX-NEXT: vpshufhw {{.*#+}} xmm4 = xmm1[0,1,2,3,6,5,7,7] ; AVX-NEXT: vinsertf128 $1, %xmm4, %ymm3, %ymm3 ; AVX-NEXT: vshufps {{.*#+}} ymm3 = ymm3[0,2,2,3,4,6,6,7] ; AVX-NEXT: vandps %ymm3, %ymm9, %ymm3 ; AVX-NEXT: vorps %ymm2, %ymm3, %ymm3 ; AVX-NEXT: vmovdqa (%rcx), %xmm4 ; AVX-NEXT: vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX-NEXT: vmovdqa (%rdx), %xmm2 ; AVX-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3],xmm2[4],xmm4[4],xmm2[5],xmm4[5],xmm2[6],xmm4[6],xmm2[7],xmm4[7] ; AVX-NEXT: vpshufhw {{.*#+}} xmm4 = xmm2[0,1,2,3,4,4,6,5] ; AVX-NEXT: vpshufhw {{.*#+}} xmm5 = xmm2[0,1,2,3,4,6,6,7] ; AVX-NEXT: vinsertf128 $1, %xmm5, %ymm4, %ymm4 ; AVX-NEXT: vmovdqa (%rsi), %xmm6 ; AVX-NEXT: vmovdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX-NEXT: vmovdqa (%rdi), %xmm5 ; AVX-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX-NEXT: vpunpcklbw {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3],xmm5[4],xmm6[4],xmm5[5],xmm6[5],xmm5[6],xmm6[6],xmm5[7],xmm6[7] ; AVX-NEXT: vpshufd {{.*#+}} xmm6 = xmm5[2,3,2,3] ; AVX-NEXT: vpmovzxwq {{.*#+}} xmm6 = xmm6[0],zero,zero,zero,xmm6[1],zero,zero,zero ; AVX-NEXT: vpshufd {{.*#+}} xmm7 = xmm5[3,3,3,3] ; AVX-NEXT: vpmovzxwq {{.*#+}} xmm7 = xmm7[0],zero,zero,zero,xmm7[1],zero,zero,zero ; AVX-NEXT: vinsertf128 $1, %xmm7, %ymm6, %ymm6 ; AVX-NEXT: vshufps {{.*#+}} ymm4 = ymm4[2,1,3,3,6,5,7,7] ; AVX-NEXT: vbroadcastsd {{.*#+}} ymm14 = [65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535] ; AVX-NEXT: vandnps %ymm4, %ymm14, %ymm4 ; AVX-NEXT: vandps %ymm6, %ymm14, %ymm6 ; AVX-NEXT: vorps %ymm4, %ymm6, %ymm4 ; AVX-NEXT: vblendps {{.*#+}} ymm3 = ymm4[0],ymm3[1],ymm4[2],ymm3[3],ymm4[4],ymm3[5],ymm4[6],ymm3[7] ; AVX-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX-NEXT: vpshuflw {{.*#+}} xmm3 = xmm0[0,0,2,1,4,5,6,7] ; AVX-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[0,2,2,3,4,5,6,7] ; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm3, %ymm0 ; AVX-NEXT: vpshuflw {{.*#+}} xmm3 = xmm1[0,1,1,3,4,5,6,7] ; AVX-NEXT: vpshuflw {{.*#+}} xmm1 = xmm1[2,1,3,3,4,5,6,7] ; AVX-NEXT: vinsertf128 $1, %xmm1, %ymm3, %ymm3 ; AVX-NEXT: vmovdqa 48(%r10), %xmm1 ; AVX-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm0[0,0,2,1,4,4,6,5] ; AVX-NEXT: vandnps %ymm0, %ymm9, %ymm0 ; AVX-NEXT: vshufps {{.*#+}} ymm3 = ymm3[0,0,2,1,4,4,6,5] ; AVX-NEXT: vandps %ymm3, %ymm9, %ymm3 ; AVX-NEXT: vorps %ymm0, %ymm3, %ymm0 ; AVX-NEXT: vpmovzxwq {{.*#+}} xmm3 = xmm5[0],zero,zero,zero,xmm5[1],zero,zero,zero ; AVX-NEXT: vpshufd {{.*#+}} xmm4 = xmm5[1,1,1,1] ; AVX-NEXT: vpmovzxwq {{.*#+}} xmm4 = xmm4[0],zero,zero,zero,xmm4[1],zero,zero,zero ; AVX-NEXT: vinsertf128 $1, %xmm4, %ymm3, %ymm3 ; AVX-NEXT: vpshuflw {{.*#+}} xmm4 = xmm2[0,0,2,1,4,5,6,7] ; AVX-NEXT: vpmovzxdq {{.*#+}} xmm4 = xmm4[0],zero,xmm4[1],zero ; AVX-NEXT: vpshuflw {{.*#+}} xmm2 = xmm2[0,2,2,3,4,5,6,7] ; AVX-NEXT: vpmovzxdq {{.*#+}} xmm2 = xmm2[0],zero,xmm2[1],zero ; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm4, %ymm4 ; AVX-NEXT: vmovdqa 48(%rax), %xmm2 ; AVX-NEXT: vandps %ymm3, %ymm14, %ymm3 ; AVX-NEXT: vandnps %ymm4, %ymm14, %ymm4 ; AVX-NEXT: vorps %ymm4, %ymm3, %ymm3 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm3[0],ymm0[1],ymm3[2],ymm0[3],ymm3[4],ymm0[5],ymm3[6],ymm0[7] ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX-NEXT: vpunpckhbw {{.*#+}} xmm11 = 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: vpshuflw {{.*#+}} xmm0 = xmm11[0,0,2,1,4,5,6,7] ; AVX-NEXT: vpshuflw {{.*#+}} xmm3 = xmm11[0,2,2,3,4,5,6,7] ; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm4 ; AVX-NEXT: vmovdqa 48(%r9), %xmm12 ; AVX-NEXT: vmovdqa 48(%r8), %xmm3 ; AVX-NEXT: vpunpckhbw {{.*#+}} xmm8 = xmm3[8],xmm12[8],xmm3[9],xmm12[9],xmm3[10],xmm12[10],xmm3[11],xmm12[11],xmm3[12],xmm12[12],xmm3[13],xmm12[13],xmm3[14],xmm12[14],xmm3[15],xmm12[15] ; AVX-NEXT: vpshuflw {{.*#+}} xmm5 = xmm8[0,1,1,3,4,5,6,7] ; AVX-NEXT: vpshuflw {{.*#+}} xmm6 = xmm8[2,1,3,3,4,5,6,7] ; AVX-NEXT: vinsertf128 $1, %xmm6, %ymm5, %ymm5 ; AVX-NEXT: vshufps {{.*#+}} ymm4 = ymm4[0,0,2,1,4,4,6,5] ; AVX-NEXT: vandnps %ymm4, %ymm9, %ymm4 ; AVX-NEXT: vshufps {{.*#+}} ymm5 = ymm5[0,0,2,1,4,4,6,5] ; AVX-NEXT: vandps %ymm5, %ymm9, %ymm5 ; AVX-NEXT: vorps %ymm4, %ymm5, %ymm9 ; AVX-NEXT: vmovdqa 48(%rsi), %xmm4 ; AVX-NEXT: vmovdqa 48(%rdi), %xmm6 ; AVX-NEXT: vpunpckhbw {{.*#+}} xmm10 = 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] ; AVX-NEXT: vpshufd {{.*#+}} xmm5 = xmm10[1,1,1,1] ; AVX-NEXT: vpmovzxwq {{.*#+}} xmm5 = xmm5[0],zero,zero,zero,xmm5[1],zero,zero,zero ; AVX-NEXT: vpmovzxwq {{.*#+}} xmm7 = xmm10[0],zero,zero,zero,xmm10[1],zero,zero,zero ; AVX-NEXT: vinsertf128 $1, %xmm5, %ymm7, %ymm0 ; AVX-NEXT: vmovdqa 48(%rcx), %xmm5 ; AVX-NEXT: vmovdqa 48(%rdx), %xmm7 ; AVX-NEXT: vpunpckhbw {{.*#+}} xmm15 = xmm7[8],xmm5[8],xmm7[9],xmm5[9],xmm7[10],xmm5[10],xmm7[11],xmm5[11],xmm7[12],xmm5[12],xmm7[13],xmm5[13],xmm7[14],xmm5[14],xmm7[15],xmm5[15] ; AVX-NEXT: vpshuflw {{.*#+}} xmm1 = xmm15[0,0,2,1,4,5,6,7] ; AVX-NEXT: vpmovzxdq {{.*#+}} xmm1 = xmm1[0],zero,xmm1[1],zero ; AVX-NEXT: vpshuflw {{.*#+}} xmm13 = xmm15[0,2,2,3,4,5,6,7] ; AVX-NEXT: vpmovzxdq {{.*#+}} xmm13 = xmm13[0],zero,xmm13[1],zero ; AVX-NEXT: vinsertf128 $1, %xmm13, %ymm1, %ymm1 ; AVX-NEXT: vandps %ymm0, %ymm14, %ymm13 ; AVX-NEXT: vandnps %ymm1, %ymm14, %ymm1 ; AVX-NEXT: vorps %ymm1, %ymm13, %ymm1 ; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm9[1],ymm1[2],ymm9[3],ymm1[4],ymm9[5],ymm1[6],ymm9[7] ; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX-NEXT: vpshufhw {{.*#+}} xmm1 = xmm11[0,1,2,3,4,4,6,5] ; AVX-NEXT: vpshufhw {{.*#+}} xmm9 = xmm11[0,1,2,3,4,6,6,7] ; AVX-NEXT: vinsertf128 $1, %xmm9, %ymm1, %ymm1 ; AVX-NEXT: vpshufhw {{.*#+}} xmm9 = xmm8[0,1,2,3,4,5,5,7] ; AVX-NEXT: vpshufhw {{.*#+}} xmm8 = xmm8[0,1,2,3,6,5,7,7] ; AVX-NEXT: vinsertf128 $1, %xmm8, %ymm9, %ymm8 ; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm1[0,2,2,3,4,6,6,7] ; AVX-NEXT: vbroadcastsd {{.*#+}} ymm11 = [65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0] ; AVX-NEXT: vandnps %ymm1, %ymm11, %ymm1 ; AVX-NEXT: vshufps {{.*#+}} ymm8 = ymm8[0,2,2,3,4,6,6,7] ; AVX-NEXT: vandps %ymm11, %ymm8, %ymm8 ; AVX-NEXT: vorps %ymm1, %ymm8, %ymm1 ; AVX-NEXT: vpshufhw {{.*#+}} xmm8 = xmm15[0,1,2,3,4,4,6,5] ; AVX-NEXT: vpshufhw {{.*#+}} xmm9 = xmm15[0,1,2,3,4,6,6,7] ; AVX-NEXT: vinsertf128 $1, %xmm9, %ymm8, %ymm8 ; AVX-NEXT: vpshufd {{.*#+}} xmm9 = xmm10[2,3,2,3] ; AVX-NEXT: vpmovzxwq {{.*#+}} xmm9 = xmm9[0],zero,zero,zero,xmm9[1],zero,zero,zero ; AVX-NEXT: vpshufd {{.*#+}} xmm10 = xmm10[3,3,3,3] ; AVX-NEXT: vpmovzxwq {{.*#+}} xmm10 = xmm10[0],zero,zero,zero,xmm10[1],zero,zero,zero ; AVX-NEXT: vinsertf128 $1, %xmm10, %ymm9, %ymm9 ; AVX-NEXT: vshufps {{.*#+}} ymm8 = ymm8[2,1,3,3,6,5,7,7] ; AVX-NEXT: vandnps %ymm8, %ymm14, %ymm8 ; AVX-NEXT: vandps %ymm14, %ymm9, %ymm9 ; AVX-NEXT: vorps %ymm8, %ymm9, %ymm8 ; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm8[0],ymm1[1],ymm8[2],ymm1[3],ymm8[4],ymm1[5],ymm8[6],ymm1[7] ; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX-NEXT: vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm1 # 16-byte Folded Reload ; AVX-NEXT: # xmm1 = 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] ; AVX-NEXT: vpshuflw {{.*#+}} xmm2 = xmm1[0,0,2,1,4,5,6,7] ; AVX-NEXT: vpshuflw {{.*#+}} xmm8 = xmm1[0,2,2,3,4,5,6,7] ; AVX-NEXT: vinsertf128 $1, %xmm8, %ymm2, %ymm8 ; AVX-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm3[0],xmm12[0],xmm3[1],xmm12[1],xmm3[2],xmm12[2],xmm3[3],xmm12[3],xmm3[4],xmm12[4],xmm3[5],xmm12[5],xmm3[6],xmm12[6],xmm3[7],xmm12[7] ; AVX-NEXT: vpshuflw {{.*#+}} xmm0 = xmm2[0,1,1,3,4,5,6,7] ; AVX-NEXT: vpshuflw {{.*#+}} xmm3 = xmm2[2,1,3,3,4,5,6,7] ; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm0 ; AVX-NEXT: vshufps {{.*#+}} ymm3 = ymm8[0,0,2,1,4,4,6,5] ; AVX-NEXT: vmovaps %ymm11, %ymm12 ; AVX-NEXT: vandnps %ymm3, %ymm11, %ymm3 ; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm0[0,0,2,1,4,4,6,5] ; AVX-NEXT: vandps %ymm0, %ymm11, %ymm0 ; AVX-NEXT: vorps %ymm3, %ymm0, %ymm3 ; AVX-NEXT: vpunpcklbw {{.*#+}} xmm0 = 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] ; AVX-NEXT: vpshufd {{.*#+}} xmm4 = xmm0[1,1,1,1] ; AVX-NEXT: vpmovzxwq {{.*#+}} xmm4 = xmm4[0],zero,zero,zero,xmm4[1],zero,zero,zero ; AVX-NEXT: vpmovzxwq {{.*#+}} xmm6 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero ; AVX-NEXT: vinsertf128 $1, %xmm4, %ymm6, %ymm4 ; AVX-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] ; AVX-NEXT: vpshuflw {{.*#+}} xmm6 = xmm5[0,0,2,1,4,5,6,7] ; AVX-NEXT: vpmovzxdq {{.*#+}} xmm6 = xmm6[0],zero,xmm6[1],zero ; AVX-NEXT: vpshuflw {{.*#+}} xmm7 = xmm5[0,2,2,3,4,5,6,7] ; AVX-NEXT: vpmovzxdq {{.*#+}} xmm7 = xmm7[0],zero,xmm7[1],zero ; AVX-NEXT: vinsertf128 $1, %xmm7, %ymm6, %ymm6 ; AVX-NEXT: vandps %ymm4, %ymm14, %ymm4 ; AVX-NEXT: vandnps %ymm6, %ymm14, %ymm6 ; AVX-NEXT: vorps %ymm6, %ymm4, %ymm4 ; AVX-NEXT: vblendps {{.*#+}} ymm3 = ymm4[0],ymm3[1],ymm4[2],ymm3[3],ymm4[4],ymm3[5],ymm4[6],ymm3[7] ; AVX-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX-NEXT: vpshufhw {{.*#+}} xmm3 = xmm1[0,1,2,3,4,4,6,5] ; AVX-NEXT: vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,6,6,7] ; AVX-NEXT: vinsertf128 $1, %xmm1, %ymm3, %ymm1 ; AVX-NEXT: vpshufhw {{.*#+}} xmm3 = xmm2[0,1,2,3,4,5,5,7] ; AVX-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,6,5,7,7] ; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm3, %ymm2 ; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm1[0,2,2,3,4,6,6,7] ; AVX-NEXT: vandnps %ymm1, %ymm11, %ymm1 ; AVX-NEXT: vshufps {{.*#+}} ymm2 = ymm2[0,2,2,3,4,6,6,7] ; AVX-NEXT: vandps %ymm2, %ymm11, %ymm2 ; AVX-NEXT: vorps %ymm1, %ymm2, %ymm1 ; AVX-NEXT: vpshufhw {{.*#+}} xmm2 = xmm5[0,1,2,3,4,4,6,5] ; AVX-NEXT: vpshufhw {{.*#+}} xmm3 = xmm5[0,1,2,3,4,6,6,7] ; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm2, %ymm2 ; AVX-NEXT: vpshufd {{.*#+}} xmm3 = xmm0[2,3,2,3] ; AVX-NEXT: vpmovzxwq {{.*#+}} xmm3 = xmm3[0],zero,zero,zero,xmm3[1],zero,zero,zero ; AVX-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[3,3,3,3] ; AVX-NEXT: vpmovzxwq {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero ; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm3, %ymm0 ; AVX-NEXT: vshufps {{.*#+}} ymm2 = ymm2[2,1,3,3,6,5,7,7] ; AVX-NEXT: vandnps %ymm2, %ymm14, %ymm2 ; AVX-NEXT: vandps %ymm0, %ymm14, %ymm0 ; AVX-NEXT: vorps %ymm2, %ymm0, %ymm0 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2],ymm1[3],ymm0[4],ymm1[5],ymm0[6],ymm1[7] ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX-NEXT: vmovdqa 32(%r10), %xmm1 ; AVX-NEXT: vmovdqa 32(%rax), %xmm2 ; AVX-NEXT: vpunpckhbw {{.*#+}} xmm11 = 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: vpshuflw {{.*#+}} xmm0 = xmm11[0,0,2,1,4,5,6,7] ; AVX-NEXT: vpshuflw {{.*#+}} xmm3 = xmm11[0,2,2,3,4,5,6,7] ; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm4 ; AVX-NEXT: vmovdqa 32(%r9), %xmm0 ; AVX-NEXT: vmovdqa 32(%r8), %xmm3 ; AVX-NEXT: vpunpckhbw {{.*#+}} xmm8 = 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] ; AVX-NEXT: vpshuflw {{.*#+}} xmm5 = xmm8[0,1,1,3,4,5,6,7] ; AVX-NEXT: vpshuflw {{.*#+}} xmm6 = xmm8[2,1,3,3,4,5,6,7] ; AVX-NEXT: vinsertf128 $1, %xmm6, %ymm5, %ymm5 ; AVX-NEXT: vshufps {{.*#+}} ymm4 = ymm4[0,0,2,1,4,4,6,5] ; AVX-NEXT: vandnps %ymm4, %ymm12, %ymm4 ; AVX-NEXT: vshufps {{.*#+}} ymm5 = ymm5[0,0,2,1,4,4,6,5] ; AVX-NEXT: vandps %ymm5, %ymm12, %ymm5 ; AVX-NEXT: vorps %ymm4, %ymm5, %ymm9 ; AVX-NEXT: vmovdqa 32(%rsi), %xmm4 ; AVX-NEXT: vmovdqa 32(%rdi), %xmm6 ; AVX-NEXT: vpunpckhbw {{.*#+}} xmm10 = 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] ; AVX-NEXT: vpshufd {{.*#+}} xmm5 = xmm10[1,1,1,1] ; AVX-NEXT: vpmovzxwq {{.*#+}} xmm5 = xmm5[0],zero,zero,zero,xmm5[1],zero,zero,zero ; AVX-NEXT: vpmovzxwq {{.*#+}} xmm7 = xmm10[0],zero,zero,zero,xmm10[1],zero,zero,zero ; AVX-NEXT: vinsertf128 $1, %xmm5, %ymm7, %ymm13 ; AVX-NEXT: vmovdqa 32(%rcx), %xmm5 ; AVX-NEXT: vmovdqa 32(%rdx), %xmm7 ; AVX-NEXT: vpunpckhbw {{.*#+}} xmm14 = xmm7[8],xmm5[8],xmm7[9],xmm5[9],xmm7[10],xmm5[10],xmm7[11],xmm5[11],xmm7[12],xmm5[12],xmm7[13],xmm5[13],xmm7[14],xmm5[14],xmm7[15],xmm5[15] ; AVX-NEXT: vpshuflw {{.*#+}} xmm15 = xmm14[0,0,2,1,4,5,6,7] ; AVX-NEXT: vpmovzxdq {{.*#+}} xmm15 = xmm15[0],zero,xmm15[1],zero ; AVX-NEXT: vpshuflw {{.*#+}} xmm12 = xmm14[0,2,2,3,4,5,6,7] ; AVX-NEXT: vpmovzxdq {{.*#+}} xmm12 = xmm12[0],zero,xmm12[1],zero ; AVX-NEXT: vinsertf128 $1, %xmm12, %ymm15, %ymm12 ; AVX-NEXT: vbroadcastsd {{.*#+}} ymm15 = [65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535] ; AVX-NEXT: vandps %ymm15, %ymm13, %ymm13 ; AVX-NEXT: vandnps %ymm12, %ymm15, %ymm12 ; AVX-NEXT: vorps %ymm12, %ymm13, %ymm12 ; AVX-NEXT: vblendps {{.*#+}} ymm9 = ymm12[0],ymm9[1],ymm12[2],ymm9[3],ymm12[4],ymm9[5],ymm12[6],ymm9[7] ; AVX-NEXT: vmovups %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX-NEXT: vpshufhw {{.*#+}} xmm9 = xmm11[0,1,2,3,4,4,6,5] ; AVX-NEXT: vpshufhw {{.*#+}} xmm11 = xmm11[0,1,2,3,4,6,6,7] ; AVX-NEXT: vinsertf128 $1, %xmm11, %ymm9, %ymm9 ; AVX-NEXT: vpshufhw {{.*#+}} xmm11 = xmm8[0,1,2,3,4,5,5,7] ; AVX-NEXT: vpshufhw {{.*#+}} xmm8 = xmm8[0,1,2,3,6,5,7,7] ; AVX-NEXT: vinsertf128 $1, %xmm8, %ymm11, %ymm8 ; AVX-NEXT: vshufps {{.*#+}} ymm9 = ymm9[0,2,2,3,4,6,6,7] ; AVX-NEXT: vbroadcastsd {{.*#+}} ymm12 = [65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0] ; AVX-NEXT: vandnps %ymm9, %ymm12, %ymm9 ; AVX-NEXT: vshufps {{.*#+}} ymm8 = ymm8[0,2,2,3,4,6,6,7] ; AVX-NEXT: vandps %ymm12, %ymm8, %ymm8 ; AVX-NEXT: vorps %ymm9, %ymm8, %ymm8 ; AVX-NEXT: vpshufhw {{.*#+}} xmm9 = xmm14[0,1,2,3,4,4,6,5] ; AVX-NEXT: vpshufhw {{.*#+}} xmm11 = xmm14[0,1,2,3,4,6,6,7] ; AVX-NEXT: vinsertf128 $1, %xmm11, %ymm9, %ymm9 ; AVX-NEXT: vpshufd {{.*#+}} xmm11 = xmm10[2,3,2,3] ; AVX-NEXT: vpmovzxwq {{.*#+}} xmm11 = xmm11[0],zero,zero,zero,xmm11[1],zero,zero,zero ; AVX-NEXT: vpshufd {{.*#+}} xmm10 = xmm10[3,3,3,3] ; AVX-NEXT: vpmovzxwq {{.*#+}} xmm10 = xmm10[0],zero,zero,zero,xmm10[1],zero,zero,zero ; AVX-NEXT: vinsertf128 $1, %xmm10, %ymm11, %ymm10 ; AVX-NEXT: vshufps {{.*#+}} ymm9 = ymm9[2,1,3,3,6,5,7,7] ; AVX-NEXT: vandnps %ymm9, %ymm15, %ymm9 ; AVX-NEXT: vandps %ymm15, %ymm10, %ymm10 ; AVX-NEXT: vorps %ymm9, %ymm10, %ymm9 ; AVX-NEXT: vblendps {{.*#+}} ymm8 = ymm9[0],ymm8[1],ymm9[2],ymm8[3],ymm9[4],ymm8[5],ymm9[6],ymm8[7] ; AVX-NEXT: vmovups %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX-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] ; AVX-NEXT: vpshuflw {{.*#+}} xmm2 = xmm1[0,0,2,1,4,5,6,7] ; AVX-NEXT: vpshuflw {{.*#+}} xmm8 = xmm1[0,2,2,3,4,5,6,7] ; AVX-NEXT: vinsertf128 $1, %xmm8, %ymm2, %ymm8 ; AVX-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm3[0],xmm0[0],xmm3[1],xmm0[1],xmm3[2],xmm0[2],xmm3[3],xmm0[3],xmm3[4],xmm0[4],xmm3[5],xmm0[5],xmm3[6],xmm0[6],xmm3[7],xmm0[7] ; AVX-NEXT: vpshuflw {{.*#+}} xmm0 = xmm2[0,1,1,3,4,5,6,7] ; AVX-NEXT: vpshuflw {{.*#+}} xmm3 = xmm2[2,1,3,3,4,5,6,7] ; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm0 ; AVX-NEXT: vshufps {{.*#+}} ymm3 = ymm8[0,0,2,1,4,4,6,5] ; AVX-NEXT: vandnps %ymm3, %ymm12, %ymm3 ; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm0[0,0,2,1,4,4,6,5] ; AVX-NEXT: vandps %ymm0, %ymm12, %ymm0 ; AVX-NEXT: vorps %ymm3, %ymm0, %ymm3 ; AVX-NEXT: vpunpcklbw {{.*#+}} xmm0 = 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] ; AVX-NEXT: vpshufd {{.*#+}} xmm4 = xmm0[1,1,1,1] ; AVX-NEXT: vpmovzxwq {{.*#+}} xmm4 = xmm4[0],zero,zero,zero,xmm4[1],zero,zero,zero ; AVX-NEXT: vpmovzxwq {{.*#+}} xmm6 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero ; AVX-NEXT: vinsertf128 $1, %xmm4, %ymm6, %ymm4 ; AVX-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] ; AVX-NEXT: vpshuflw {{.*#+}} xmm6 = xmm5[0,0,2,1,4,5,6,7] ; AVX-NEXT: vpmovzxdq {{.*#+}} xmm6 = xmm6[0],zero,xmm6[1],zero ; AVX-NEXT: vpshuflw {{.*#+}} xmm7 = xmm5[0,2,2,3,4,5,6,7] ; AVX-NEXT: vpmovzxdq {{.*#+}} xmm7 = xmm7[0],zero,xmm7[1],zero ; AVX-NEXT: vinsertf128 $1, %xmm7, %ymm6, %ymm6 ; AVX-NEXT: vandps %ymm4, %ymm15, %ymm4 ; AVX-NEXT: vandnps %ymm6, %ymm15, %ymm6 ; AVX-NEXT: vorps %ymm6, %ymm4, %ymm4 ; AVX-NEXT: vblendps {{.*#+}} ymm3 = ymm4[0],ymm3[1],ymm4[2],ymm3[3],ymm4[4],ymm3[5],ymm4[6],ymm3[7] ; AVX-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX-NEXT: vpshufhw {{.*#+}} xmm3 = xmm1[0,1,2,3,4,4,6,5] ; AVX-NEXT: vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,6,6,7] ; AVX-NEXT: vinsertf128 $1, %xmm1, %ymm3, %ymm1 ; AVX-NEXT: vpshufhw {{.*#+}} xmm3 = xmm2[0,1,2,3,4,5,5,7] ; AVX-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,6,5,7,7] ; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm3, %ymm2 ; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm1[0,2,2,3,4,6,6,7] ; AVX-NEXT: vandnps %ymm1, %ymm12, %ymm1 ; AVX-NEXT: vshufps {{.*#+}} ymm2 = ymm2[0,2,2,3,4,6,6,7] ; AVX-NEXT: vandps %ymm2, %ymm12, %ymm2 ; AVX-NEXT: vorps %ymm1, %ymm2, %ymm1 ; AVX-NEXT: vpshufhw {{.*#+}} xmm2 = xmm5[0,1,2,3,4,4,6,5] ; AVX-NEXT: vpshufhw {{.*#+}} xmm3 = xmm5[0,1,2,3,4,6,6,7] ; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm2, %ymm2 ; AVX-NEXT: vpshufd {{.*#+}} xmm3 = xmm0[2,3,2,3] ; AVX-NEXT: vpmovzxwq {{.*#+}} xmm3 = xmm3[0],zero,zero,zero,xmm3[1],zero,zero,zero ; AVX-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[3,3,3,3] ; AVX-NEXT: vpmovzxwq {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero ; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm3, %ymm0 ; AVX-NEXT: vshufps {{.*#+}} ymm2 = ymm2[2,1,3,3,6,5,7,7] ; AVX-NEXT: vandnps %ymm2, %ymm15, %ymm2 ; AVX-NEXT: vandps %ymm0, %ymm15, %ymm0 ; AVX-NEXT: vorps %ymm2, %ymm0, %ymm0 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2],ymm1[3],ymm0[4],ymm1[5],ymm0[6],ymm1[7] ; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX-NEXT: vmovdqa 16(%r10), %xmm9 ; AVX-NEXT: vmovdqa 16(%rax), %xmm6 ; AVX-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm6[8],xmm9[8],xmm6[9],xmm9[9],xmm6[10],xmm9[10],xmm6[11],xmm9[11],xmm6[12],xmm9[12],xmm6[13],xmm9[13],xmm6[14],xmm9[14],xmm6[15],xmm9[15] ; AVX-NEXT: vpshuflw {{.*#+}} xmm0 = xmm1[0,0,2,1,4,5,6,7] ; AVX-NEXT: vpshuflw {{.*#+}} xmm2 = xmm1[0,2,2,3,4,5,6,7] ; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm2 ; AVX-NEXT: vmovdqa 16(%r9), %xmm8 ; AVX-NEXT: vmovdqa 16(%r8), %xmm7 ; AVX-NEXT: vpunpckhbw {{.*#+}} xmm0 = 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: vpshuflw {{.*#+}} xmm3 = xmm0[0,1,1,3,4,5,6,7] ; AVX-NEXT: vpshuflw {{.*#+}} xmm4 = xmm0[2,1,3,3,4,5,6,7] ; AVX-NEXT: vinsertf128 $1, %xmm4, %ymm3, %ymm3 ; AVX-NEXT: vshufps {{.*#+}} ymm2 = ymm2[0,0,2,1,4,4,6,5] ; AVX-NEXT: vandnps %ymm2, %ymm12, %ymm2 ; AVX-NEXT: vshufps {{.*#+}} ymm3 = ymm3[0,0,2,1,4,4,6,5] ; AVX-NEXT: vandps %ymm3, %ymm12, %ymm3 ; AVX-NEXT: vorps %ymm2, %ymm3, %ymm11 ; AVX-NEXT: vmovdqa 16(%rsi), %xmm5 ; AVX-NEXT: vmovdqa 16(%rdi), %xmm3 ; AVX-NEXT: vpunpckhbw {{.*#+}} xmm10 = 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] ; AVX-NEXT: vpshufd {{.*#+}} xmm2 = xmm10[1,1,1,1] ; AVX-NEXT: vpmovzxwq {{.*#+}} xmm2 = xmm2[0],zero,zero,zero,xmm2[1],zero,zero,zero ; AVX-NEXT: vpmovzxwq {{.*#+}} xmm4 = xmm10[0],zero,zero,zero,xmm10[1],zero,zero,zero ; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm4, %ymm12 ; AVX-NEXT: vmovdqa 16(%rcx), %xmm4 ; AVX-NEXT: vmovdqa 16(%rdx), %xmm2 ; AVX-NEXT: vpunpckhbw {{.*#+}} xmm13 = xmm2[8],xmm4[8],xmm2[9],xmm4[9],xmm2[10],xmm4[10],xmm2[11],xmm4[11],xmm2[12],xmm4[12],xmm2[13],xmm4[13],xmm2[14],xmm4[14],xmm2[15],xmm4[15] ; AVX-NEXT: vpshuflw {{.*#+}} xmm15 = xmm13[0,0,2,1,4,5,6,7] ; AVX-NEXT: vpmovzxdq {{.*#+}} xmm15 = xmm15[0],zero,xmm15[1],zero ; AVX-NEXT: vpshuflw {{.*#+}} xmm14 = xmm13[0,2,2,3,4,5,6,7] ; AVX-NEXT: vpmovzxdq {{.*#+}} xmm14 = xmm14[0],zero,xmm14[1],zero ; AVX-NEXT: vinsertf128 $1, %xmm14, %ymm15, %ymm14 ; AVX-NEXT: vbroadcastsd {{.*#+}} ymm15 = [65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535] ; AVX-NEXT: vandps %ymm15, %ymm12, %ymm12 ; AVX-NEXT: vandnps %ymm14, %ymm15, %ymm14 ; AVX-NEXT: vorps %ymm14, %ymm12, %ymm12 ; AVX-NEXT: vblendps {{.*#+}} ymm11 = ymm12[0],ymm11[1],ymm12[2],ymm11[3],ymm12[4],ymm11[5],ymm12[6],ymm11[7] ; AVX-NEXT: vpshufhw {{.*#+}} xmm12 = xmm1[0,1,2,3,4,4,6,5] ; AVX-NEXT: vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,6,6,7] ; AVX-NEXT: vinsertf128 $1, %xmm1, %ymm12, %ymm1 ; AVX-NEXT: vpshufhw {{.*#+}} xmm12 = xmm0[0,1,2,3,4,5,5,7] ; AVX-NEXT: vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,6,5,7,7] ; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm12, %ymm0 ; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm1[0,2,2,3,4,6,6,7] ; AVX-NEXT: vbroadcastsd {{.*#+}} ymm14 = [65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0] ; AVX-NEXT: vandnps %ymm1, %ymm14, %ymm1 ; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm0[0,2,2,3,4,6,6,7] ; AVX-NEXT: vandps %ymm0, %ymm14, %ymm0 ; AVX-NEXT: vorps %ymm1, %ymm0, %ymm0 ; AVX-NEXT: vpshufhw {{.*#+}} xmm1 = xmm13[0,1,2,3,4,4,6,5] ; AVX-NEXT: vpshufhw {{.*#+}} xmm12 = xmm13[0,1,2,3,4,6,6,7] ; AVX-NEXT: vinsertf128 $1, %xmm12, %ymm1, %ymm1 ; AVX-NEXT: vpshufd {{.*#+}} xmm12 = xmm10[2,3,2,3] ; AVX-NEXT: vpmovzxwq {{.*#+}} xmm12 = xmm12[0],zero,zero,zero,xmm12[1],zero,zero,zero ; AVX-NEXT: vpshufd {{.*#+}} xmm10 = xmm10[3,3,3,3] ; AVX-NEXT: vpmovzxwq {{.*#+}} xmm10 = xmm10[0],zero,zero,zero,xmm10[1],zero,zero,zero ; AVX-NEXT: vinsertf128 $1, %xmm10, %ymm12, %ymm10 ; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm1[2,1,3,3,6,5,7,7] ; AVX-NEXT: vandnps %ymm1, %ymm15, %ymm1 ; AVX-NEXT: vandps %ymm15, %ymm10, %ymm10 ; AVX-NEXT: vorps %ymm1, %ymm10, %ymm1 ; AVX-NEXT: vblendps {{.*#+}} ymm10 = ymm1[0],ymm0[1],ymm1[2],ymm0[3],ymm1[4],ymm0[5],ymm1[6],ymm0[7] ; AVX-NEXT: vpunpcklbw {{.*#+}} xmm6 = xmm6[0],xmm9[0],xmm6[1],xmm9[1],xmm6[2],xmm9[2],xmm6[3],xmm9[3],xmm6[4],xmm9[4],xmm6[5],xmm9[5],xmm6[6],xmm9[6],xmm6[7],xmm9[7] ; AVX-NEXT: vpshuflw {{.*#+}} xmm0 = xmm6[0,0,2,1,4,5,6,7] ; AVX-NEXT: vpshuflw {{.*#+}} xmm1 = xmm6[0,2,2,3,4,5,6,7] ; AVX-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0 ; 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: vpshuflw {{.*#+}} xmm1 = xmm7[0,1,1,3,4,5,6,7] ; AVX-NEXT: vpshuflw {{.*#+}} xmm8 = xmm7[2,1,3,3,4,5,6,7] ; AVX-NEXT: vinsertf128 $1, %xmm8, %ymm1, %ymm1 ; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm0[0,0,2,1,4,4,6,5] ; AVX-NEXT: vandnps %ymm0, %ymm14, %ymm0 ; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm1[0,0,2,1,4,4,6,5] ; AVX-NEXT: vandps %ymm1, %ymm14, %ymm1 ; AVX-NEXT: vorps %ymm0, %ymm1, %ymm1 ; AVX-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm3[0],xmm5[0],xmm3[1],xmm5[1],xmm3[2],xmm5[2],xmm3[3],xmm5[3],xmm3[4],xmm5[4],xmm3[5],xmm5[5],xmm3[6],xmm5[6],xmm3[7],xmm5[7] ; AVX-NEXT: vpshufd {{.*#+}} xmm3 = xmm0[1,1,1,1] ; AVX-NEXT: vpmovzxwq {{.*#+}} xmm3 = xmm3[0],zero,zero,zero,xmm3[1],zero,zero,zero ; AVX-NEXT: vpmovzxwq {{.*#+}} xmm5 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero ; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm5, %ymm3 ; AVX-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3],xmm2[4],xmm4[4],xmm2[5],xmm4[5],xmm2[6],xmm4[6],xmm2[7],xmm4[7] ; AVX-NEXT: vpshuflw {{.*#+}} xmm4 = xmm2[0,0,2,1,4,5,6,7] ; AVX-NEXT: vpmovzxdq {{.*#+}} xmm4 = xmm4[0],zero,xmm4[1],zero ; AVX-NEXT: vpshuflw {{.*#+}} xmm5 = xmm2[0,2,2,3,4,5,6,7] ; AVX-NEXT: vpmovzxdq {{.*#+}} xmm5 = xmm5[0],zero,xmm5[1],zero ; AVX-NEXT: vinsertf128 $1, %xmm5, %ymm4, %ymm4 ; AVX-NEXT: vandps %ymm3, %ymm15, %ymm3 ; AVX-NEXT: vandnps %ymm4, %ymm15, %ymm4 ; AVX-NEXT: vorps %ymm4, %ymm3, %ymm3 ; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm3[0],ymm1[1],ymm3[2],ymm1[3],ymm3[4],ymm1[5],ymm3[6],ymm1[7] ; AVX-NEXT: vpshufhw {{.*#+}} xmm3 = xmm6[0,1,2,3,4,4,6,5] ; AVX-NEXT: vpshufhw {{.*#+}} xmm4 = xmm6[0,1,2,3,4,6,6,7] ; AVX-NEXT: vinsertf128 $1, %xmm4, %ymm3, %ymm3 ; AVX-NEXT: vpshufhw {{.*#+}} xmm4 = xmm7[0,1,2,3,4,5,5,7] ; AVX-NEXT: vpshufhw {{.*#+}} xmm5 = xmm7[0,1,2,3,6,5,7,7] ; AVX-NEXT: vinsertf128 $1, %xmm5, %ymm4, %ymm4 ; AVX-NEXT: vshufps {{.*#+}} ymm3 = ymm3[0,2,2,3,4,6,6,7] ; AVX-NEXT: vandnps %ymm3, %ymm14, %ymm3 ; AVX-NEXT: vshufps {{.*#+}} ymm4 = ymm4[0,2,2,3,4,6,6,7] ; AVX-NEXT: vandps %ymm4, %ymm14, %ymm4 ; AVX-NEXT: vorps %ymm3, %ymm4, %ymm3 ; AVX-NEXT: vpshufhw {{.*#+}} xmm4 = xmm2[0,1,2,3,4,4,6,5] ; AVX-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,6,6,7] ; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm4, %ymm2 ; AVX-NEXT: vpshufd {{.*#+}} xmm4 = xmm0[2,3,2,3] ; AVX-NEXT: vpmovzxwq {{.*#+}} xmm4 = xmm4[0],zero,zero,zero,xmm4[1],zero,zero,zero ; AVX-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[3,3,3,3] ; AVX-NEXT: vpmovzxwq {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero ; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm4, %ymm0 ; AVX-NEXT: vshufps {{.*#+}} ymm2 = ymm2[2,1,3,3,6,5,7,7] ; AVX-NEXT: vandnps %ymm2, %ymm15, %ymm2 ; AVX-NEXT: vandps %ymm0, %ymm15, %ymm0 ; AVX-NEXT: vorps %ymm2, %ymm0, %ymm0 ; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm3[1],ymm0[2],ymm3[3],ymm0[4],ymm3[5],ymm0[6],ymm3[7] ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload ; AVX-NEXT: vpunpckhbw (%rsp), %xmm2, %xmm3 # 16-byte Folded Reload ; AVX-NEXT: # xmm3 = 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] ; AVX-NEXT: vpshuflw {{.*#+}} xmm2 = xmm3[0,0,2,1,4,5,6,7] ; AVX-NEXT: vpshuflw {{.*#+}} xmm4 = xmm3[0,2,2,3,4,5,6,7] ; AVX-NEXT: vinsertf128 $1, %xmm4, %ymm2, %ymm2 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload ; AVX-NEXT: vpunpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm4, %xmm4 # 16-byte Folded Reload ; AVX-NEXT: # xmm4 = xmm4[8],mem[8],xmm4[9],mem[9],xmm4[10],mem[10],xmm4[11],mem[11],xmm4[12],mem[12],xmm4[13],mem[13],xmm4[14],mem[14],xmm4[15],mem[15] ; AVX-NEXT: vpshuflw {{.*#+}} xmm5 = xmm4[0,1,1,3,4,5,6,7] ; AVX-NEXT: vpshuflw {{.*#+}} xmm6 = xmm4[2,1,3,3,4,5,6,7] ; AVX-NEXT: vinsertf128 $1, %xmm6, %ymm5, %ymm5 ; AVX-NEXT: vshufps {{.*#+}} ymm2 = ymm2[0,0,2,1,4,4,6,5] ; AVX-NEXT: vandnps %ymm2, %ymm14, %ymm2 ; AVX-NEXT: vshufps {{.*#+}} ymm5 = ymm5[0,0,2,1,4,4,6,5] ; AVX-NEXT: vandps %ymm5, %ymm14, %ymm5 ; AVX-NEXT: vorps %ymm2, %ymm5, %ymm5 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload ; AVX-NEXT: vpunpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm2 # 16-byte Folded Reload ; AVX-NEXT: # xmm2 = xmm2[8],mem[8],xmm2[9],mem[9],xmm2[10],mem[10],xmm2[11],mem[11],xmm2[12],mem[12],xmm2[13],mem[13],xmm2[14],mem[14],xmm2[15],mem[15] ; AVX-NEXT: vpshufd {{.*#+}} xmm6 = xmm2[1,1,1,1] ; AVX-NEXT: vpmovzxwq {{.*#+}} xmm6 = xmm6[0],zero,zero,zero,xmm6[1],zero,zero,zero ; AVX-NEXT: vpmovzxwq {{.*#+}} xmm7 = xmm2[0],zero,zero,zero,xmm2[1],zero,zero,zero ; AVX-NEXT: vinsertf128 $1, %xmm6, %ymm7, %ymm6 ; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload ; AVX-NEXT: vpunpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm7, %xmm7 # 16-byte Folded Reload ; AVX-NEXT: # xmm7 = xmm7[8],mem[8],xmm7[9],mem[9],xmm7[10],mem[10],xmm7[11],mem[11],xmm7[12],mem[12],xmm7[13],mem[13],xmm7[14],mem[14],xmm7[15],mem[15] ; AVX-NEXT: vpshuflw {{.*#+}} xmm8 = xmm7[0,0,2,1,4,5,6,7] ; AVX-NEXT: vpmovzxdq {{.*#+}} xmm8 = xmm8[0],zero,xmm8[1],zero ; AVX-NEXT: vpshuflw {{.*#+}} xmm9 = xmm7[0,2,2,3,4,5,6,7] ; AVX-NEXT: vpmovzxdq {{.*#+}} xmm9 = xmm9[0],zero,xmm9[1],zero ; AVX-NEXT: vinsertf128 $1, %xmm9, %ymm8, %ymm8 ; AVX-NEXT: vandps %ymm6, %ymm15, %ymm6 ; AVX-NEXT: vandnps %ymm8, %ymm15, %ymm8 ; AVX-NEXT: vorps %ymm6, %ymm8, %ymm6 ; AVX-NEXT: vblendps {{.*#+}} ymm5 = ymm6[0],ymm5[1],ymm6[2],ymm5[3],ymm6[4],ymm5[5],ymm6[6],ymm5[7] ; AVX-NEXT: vpshufhw {{.*#+}} xmm6 = xmm3[0,1,2,3,4,4,6,5] ; AVX-NEXT: vpshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,6,6,7] ; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm6, %ymm3 ; AVX-NEXT: vpshufhw {{.*#+}} xmm6 = xmm4[0,1,2,3,4,5,5,7] ; AVX-NEXT: vpshufhw {{.*#+}} xmm4 = xmm4[0,1,2,3,6,5,7,7] ; AVX-NEXT: vinsertf128 $1, %xmm4, %ymm6, %ymm4 ; AVX-NEXT: vshufps {{.*#+}} ymm3 = ymm3[0,2,2,3,4,6,6,7] ; AVX-NEXT: vandnps %ymm3, %ymm14, %ymm3 ; AVX-NEXT: vshufps {{.*#+}} ymm4 = ymm4[0,2,2,3,4,6,6,7] ; AVX-NEXT: vandps %ymm4, %ymm14, %ymm4 ; AVX-NEXT: vorps %ymm3, %ymm4, %ymm3 ; AVX-NEXT: vpshufhw {{.*#+}} xmm4 = xmm7[0,1,2,3,4,4,6,5] ; AVX-NEXT: vpshufhw {{.*#+}} xmm6 = xmm7[0,1,2,3,4,6,6,7] ; AVX-NEXT: vinsertf128 $1, %xmm6, %ymm4, %ymm4 ; AVX-NEXT: vpshufd {{.*#+}} xmm6 = xmm2[2,3,2,3] ; AVX-NEXT: vpmovzxwq {{.*#+}} xmm6 = xmm6[0],zero,zero,zero,xmm6[1],zero,zero,zero ; AVX-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[3,3,3,3] ; AVX-NEXT: vpmovzxwq {{.*#+}} xmm2 = xmm2[0],zero,zero,zero,xmm2[1],zero,zero,zero ; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm6, %ymm2 ; AVX-NEXT: vshufps {{.*#+}} ymm4 = ymm4[2,1,3,3,6,5,7,7] ; AVX-NEXT: vandnps %ymm4, %ymm15, %ymm4 ; AVX-NEXT: vandps %ymm2, %ymm15, %ymm2 ; AVX-NEXT: vorps %ymm4, %ymm2, %ymm2 ; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0],ymm3[1],ymm2[2],ymm3[3],ymm2[4],ymm3[5],ymm2[6],ymm3[7] ; AVX-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX-NEXT: vmovaps %ymm2, 96(%rax) ; AVX-NEXT: vmovaps %ymm5, 64(%rax) ; AVX-NEXT: vmovaps %ymm0, 160(%rax) ; AVX-NEXT: vmovaps %ymm1, 128(%rax) ; AVX-NEXT: vmovaps %ymm10, 224(%rax) ; AVX-NEXT: vmovaps %ymm11, 192(%rax) ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload ; AVX-NEXT: vmovaps %ymm0, 288(%rax) ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload ; AVX-NEXT: vmovaps %ymm0, 256(%rax) ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload ; AVX-NEXT: vmovaps %ymm0, 352(%rax) ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload ; AVX-NEXT: vmovaps %ymm0, 320(%rax) ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload ; AVX-NEXT: vmovaps %ymm0, 416(%rax) ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload ; AVX-NEXT: vmovaps %ymm0, 384(%rax) ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload ; AVX-NEXT: vmovaps %ymm0, 480(%rax) ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload ; AVX-NEXT: vmovaps %ymm0, 448(%rax) ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload ; AVX-NEXT: vmovaps %ymm0, (%rax) ; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload ; AVX-NEXT: vmovaps %ymm0, 32(%rax) ; AVX-NEXT: addq $328, %rsp # imm = 0x148 ; AVX-NEXT: vzeroupper ; AVX-NEXT: retq ; ; AVX2-LABEL: store_i8_stride8_vf64: ; AVX2: # %bb.0: ; AVX2-NEXT: subq $328, %rsp # imm = 0x148 ; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX2-NEXT: vmovdqa (%r10), %xmm0 ; AVX2-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX2-NEXT: vmovdqa 48(%r10), %xmm11 ; AVX2-NEXT: vmovdqa (%rax), %xmm1 ; AVX2-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX2-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3],xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7] ; AVX2-NEXT: vpshufhw {{.*#+}} xmm1 = xmm2[0,1,2,3,4,4,6,5] ; AVX2-NEXT: vpshufhw {{.*#+}} xmm3 = xmm2[0,1,2,3,4,6,6,7] ; AVX2-NEXT: vinserti128 $1, %xmm3, %ymm1, %ymm1 ; AVX2-NEXT: vmovdqa (%r9), %xmm0 ; AVX2-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX2-NEXT: vmovdqa (%r8), %xmm3 ; AVX2-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX2-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm3[0],xmm0[0],xmm3[1],xmm0[1],xmm3[2],xmm0[2],xmm3[3],xmm0[3],xmm3[4],xmm0[4],xmm3[5],xmm0[5],xmm3[6],xmm0[6],xmm3[7],xmm0[7] ; AVX2-NEXT: vpshufhw {{.*#+}} xmm4 = xmm3[0,1,2,3,4,5,5,7] ; AVX2-NEXT: vpshufhw {{.*#+}} xmm5 = xmm3[0,1,2,3,6,5,7,7] ; AVX2-NEXT: vinserti128 $1, %xmm5, %ymm4, %ymm4 ; AVX2-NEXT: vpblendw {{.*#+}} ymm5 = ymm4[0,1,2,3,4],ymm1[5],ymm4[6],ymm1[7],ymm4[8,9,10,11,12],ymm1[13],ymm4[14],ymm1[15] ; AVX2-NEXT: vmovdqa (%rcx), %xmm0 ; AVX2-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX2-NEXT: vmovdqa (%rdx), %xmm1 ; AVX2-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX2-NEXT: vpunpcklbw {{.*#+}} xmm4 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3],xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7] ; AVX2-NEXT: vpshufhw {{.*#+}} xmm1 = xmm4[0,1,2,3,4,4,6,5] ; AVX2-NEXT: vpshufhw {{.*#+}} xmm6 = xmm4[0,1,2,3,4,6,6,7] ; AVX2-NEXT: vinserti128 $1, %xmm6, %ymm1, %ymm1 ; AVX2-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[2,1,3,3,6,5,7,7] ; AVX2-NEXT: vmovdqa (%rsi), %xmm0 ; AVX2-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX2-NEXT: vmovdqa (%rdi), %xmm6 ; AVX2-NEXT: vmovdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX2-NEXT: vpunpcklbw {{.*#+}} xmm6 = xmm6[0],xmm0[0],xmm6[1],xmm0[1],xmm6[2],xmm0[2],xmm6[3],xmm0[3],xmm6[4],xmm0[4],xmm6[5],xmm0[5],xmm6[6],xmm0[6],xmm6[7],xmm0[7] ; AVX2-NEXT: vpshufd {{.*#+}} xmm7 = xmm6[2,3,2,3] ; AVX2-NEXT: vpmovzxwq {{.*#+}} ymm7 = xmm7[0],zero,zero,zero,xmm7[1],zero,zero,zero,xmm7[2],zero,zero,zero,xmm7[3],zero,zero,zero ; AVX2-NEXT: vpblendw {{.*#+}} ymm7 = ymm7[0],ymm1[1],ymm7[2,3,4],ymm1[5],ymm7[6,7,8],ymm1[9],ymm7[10,11,12],ymm1[13],ymm7[14,15] ; AVX2-NEXT: vmovdqa 48(%rax), %xmm1 ; AVX2-NEXT: vpshufd {{.*#+}} ymm5 = ymm5[0,2,2,3,4,6,6,7] ; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm7[0],ymm5[1],ymm7[2],ymm5[3],ymm7[4],ymm5[5],ymm7[6],ymm5[7] ; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-NEXT: vpshuflw {{.*#+}} xmm5 = xmm2[0,0,2,1,4,5,6,7] ; AVX2-NEXT: vpshuflw {{.*#+}} xmm2 = xmm2[0,2,2,3,4,5,6,7] ; AVX2-NEXT: vinserti128 $1, %xmm2, %ymm5, %ymm2 ; AVX2-NEXT: vpshuflw {{.*#+}} xmm5 = xmm3[0,1,1,3,4,5,6,7] ; AVX2-NEXT: vpshuflw {{.*#+}} xmm3 = xmm3[2,1,3,3,4,5,6,7] ; AVX2-NEXT: vinserti128 $1, %xmm3, %ymm5, %ymm3 ; AVX2-NEXT: vpblendw {{.*#+}} ymm2 = ymm3[0],ymm2[1],ymm3[2],ymm2[3],ymm3[4,5,6,7,8],ymm2[9],ymm3[10],ymm2[11],ymm3[12,13,14,15] ; AVX2-NEXT: vpshufd {{.*#+}} ymm2 = ymm2[0,0,2,1,4,4,6,5] ; AVX2-NEXT: vpmovzxwq {{.*#+}} ymm3 = xmm6[0],zero,zero,zero,xmm6[1],zero,zero,zero,xmm6[2],zero,zero,zero,xmm6[3],zero,zero,zero ; AVX2-NEXT: vmovdqa {{.*#+}} xmm14 = [0,1,0,1,4,5,2,3,0,1,4,5,4,5,6,7] ; AVX2-NEXT: vpshufb %xmm14, %xmm4, %xmm4 ; AVX2-NEXT: vpmovzxdq {{.*#+}} ymm4 = xmm4[0],zero,xmm4[1],zero,xmm4[2],zero,xmm4[3],zero ; AVX2-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0],ymm4[1],ymm3[2,3,4],ymm4[5],ymm3[6,7,8],ymm4[9],ymm3[10,11,12],ymm4[13],ymm3[14,15] ; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm3[0],ymm2[1],ymm3[2],ymm2[3],ymm3[4],ymm2[5],ymm3[6],ymm2[7] ; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-NEXT: vpunpckhbw {{.*#+}} xmm8 = 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] ; AVX2-NEXT: vpshuflw {{.*#+}} xmm2 = xmm8[0,0,2,1,4,5,6,7] ; AVX2-NEXT: vpshuflw {{.*#+}} xmm3 = xmm8[0,2,2,3,4,5,6,7] ; AVX2-NEXT: vinserti128 $1, %xmm3, %ymm2, %ymm2 ; AVX2-NEXT: vmovdqa 48(%r9), %xmm3 ; AVX2-NEXT: vmovdqa 48(%r8), %xmm5 ; AVX2-NEXT: vpunpckhbw {{.*#+}} xmm9 = xmm5[8],xmm3[8],xmm5[9],xmm3[9],xmm5[10],xmm3[10],xmm5[11],xmm3[11],xmm5[12],xmm3[12],xmm5[13],xmm3[13],xmm5[14],xmm3[14],xmm5[15],xmm3[15] ; AVX2-NEXT: vpshuflw {{.*#+}} xmm4 = xmm9[0,1,1,3,4,5,6,7] ; AVX2-NEXT: vpshuflw {{.*#+}} xmm6 = xmm9[2,1,3,3,4,5,6,7] ; AVX2-NEXT: vinserti128 $1, %xmm6, %ymm4, %ymm4 ; AVX2-NEXT: vpblendw {{.*#+}} ymm10 = ymm4[0],ymm2[1],ymm4[2],ymm2[3],ymm4[4,5,6,7,8],ymm2[9],ymm4[10],ymm2[11],ymm4[12,13,14,15] ; AVX2-NEXT: vmovdqa 48(%rcx), %xmm2 ; AVX2-NEXT: vmovdqa 48(%rdx), %xmm4 ; AVX2-NEXT: vpunpckhbw {{.*#+}} xmm12 = 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: vpshufb %xmm14, %xmm12, %xmm6 ; AVX2-NEXT: vpmovzxdq {{.*#+}} ymm13 = xmm6[0],zero,xmm6[1],zero,xmm6[2],zero,xmm6[3],zero ; AVX2-NEXT: vmovdqa 48(%rsi), %xmm6 ; AVX2-NEXT: vmovdqa 48(%rdi), %xmm7 ; AVX2-NEXT: vpunpckhbw {{.*#+}} xmm0 = 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] ; AVX2-NEXT: vpmovzxwq {{.*#+}} ymm15 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero ; AVX2-NEXT: vpblendw {{.*#+}} ymm13 = ymm15[0],ymm13[1],ymm15[2,3,4],ymm13[5],ymm15[6,7,8],ymm13[9],ymm15[10,11,12],ymm13[13],ymm15[14,15] ; AVX2-NEXT: vpshufd {{.*#+}} ymm10 = ymm10[0,0,2,1,4,4,6,5] ; AVX2-NEXT: vpblendd {{.*#+}} ymm10 = ymm13[0],ymm10[1],ymm13[2],ymm10[3],ymm13[4],ymm10[5],ymm13[6],ymm10[7] ; AVX2-NEXT: vmovdqu %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-NEXT: vpshufhw {{.*#+}} xmm10 = xmm8[0,1,2,3,4,4,6,5] ; AVX2-NEXT: vpshufhw {{.*#+}} xmm8 = xmm8[0,1,2,3,4,6,6,7] ; AVX2-NEXT: vinserti128 $1, %xmm8, %ymm10, %ymm8 ; AVX2-NEXT: vpshufhw {{.*#+}} xmm10 = xmm9[0,1,2,3,4,5,5,7] ; AVX2-NEXT: vpshufhw {{.*#+}} xmm9 = xmm9[0,1,2,3,6,5,7,7] ; AVX2-NEXT: vinserti128 $1, %xmm9, %ymm10, %ymm9 ; AVX2-NEXT: vpblendw {{.*#+}} ymm8 = ymm9[0,1,2,3,4],ymm8[5],ymm9[6],ymm8[7],ymm9[8,9,10,11,12],ymm8[13],ymm9[14],ymm8[15] ; AVX2-NEXT: vpshufhw {{.*#+}} xmm9 = xmm12[0,1,2,3,4,4,6,5] ; AVX2-NEXT: vpshufhw {{.*#+}} xmm10 = xmm12[0,1,2,3,4,6,6,7] ; AVX2-NEXT: vinserti128 $1, %xmm10, %ymm9, %ymm9 ; AVX2-NEXT: vpshufd {{.*#+}} ymm9 = ymm9[2,1,3,3,6,5,7,7] ; AVX2-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,3,2,3] ; AVX2-NEXT: vpmovzxwq {{.*#+}} ymm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero ; AVX2-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm9[1],ymm0[2,3,4],ymm9[5],ymm0[6,7,8],ymm9[9],ymm0[10,11,12],ymm9[13],ymm0[14,15] ; AVX2-NEXT: vpshufd {{.*#+}} ymm8 = ymm8[0,2,2,3,4,6,6,7] ; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0],ymm8[1],ymm0[2],ymm8[3],ymm0[4],ymm8[5],ymm0[6],ymm8[7] ; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm1[0],xmm11[0],xmm1[1],xmm11[1],xmm1[2],xmm11[2],xmm1[3],xmm11[3],xmm1[4],xmm11[4],xmm1[5],xmm11[5],xmm1[6],xmm11[6],xmm1[7],xmm11[7] ; AVX2-NEXT: vpshuflw {{.*#+}} xmm1 = xmm0[0,0,2,1,4,5,6,7] ; AVX2-NEXT: vpshuflw {{.*#+}} xmm8 = xmm0[0,2,2,3,4,5,6,7] ; AVX2-NEXT: vinserti128 $1, %xmm8, %ymm1, %ymm1 ; AVX2-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm5[0],xmm3[0],xmm5[1],xmm3[1],xmm5[2],xmm3[2],xmm5[3],xmm3[3],xmm5[4],xmm3[4],xmm5[5],xmm3[5],xmm5[6],xmm3[6],xmm5[7],xmm3[7] ; AVX2-NEXT: vpshuflw {{.*#+}} xmm5 = xmm3[0,1,1,3,4,5,6,7] ; AVX2-NEXT: vpshuflw {{.*#+}} xmm8 = xmm3[2,1,3,3,4,5,6,7] ; AVX2-NEXT: vinserti128 $1, %xmm8, %ymm5, %ymm5 ; AVX2-NEXT: vpblendw {{.*#+}} ymm1 = ymm5[0],ymm1[1],ymm5[2],ymm1[3],ymm5[4,5,6,7,8],ymm1[9],ymm5[10],ymm1[11],ymm5[12,13,14,15] ; AVX2-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] ; AVX2-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm4[0],xmm2[0],xmm4[1],xmm2[1],xmm4[2],xmm2[2],xmm4[3],xmm2[3],xmm4[4],xmm2[4],xmm4[5],xmm2[5],xmm4[6],xmm2[6],xmm4[7],xmm2[7] ; AVX2-NEXT: vpshufb %xmm14, %xmm2, %xmm4 ; AVX2-NEXT: vpmovzxdq {{.*#+}} ymm4 = xmm4[0],zero,xmm4[1],zero,xmm4[2],zero,xmm4[3],zero ; AVX2-NEXT: vpmovzxwq {{.*#+}} ymm6 = xmm5[0],zero,zero,zero,xmm5[1],zero,zero,zero,xmm5[2],zero,zero,zero,xmm5[3],zero,zero,zero ; AVX2-NEXT: vpblendw {{.*#+}} ymm4 = ymm6[0],ymm4[1],ymm6[2,3,4],ymm4[5],ymm6[6,7,8],ymm4[9],ymm6[10,11,12],ymm4[13],ymm6[14,15] ; AVX2-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[0,0,2,1,4,4,6,5] ; AVX2-NEXT: vpblendd {{.*#+}} ymm1 = ymm4[0],ymm1[1],ymm4[2],ymm1[3],ymm4[4],ymm1[5],ymm4[6],ymm1[7] ; AVX2-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-NEXT: vpshufhw {{.*#+}} xmm1 = xmm0[0,1,2,3,4,4,6,5] ; AVX2-NEXT: vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,6,6,7] ; AVX2-NEXT: vinserti128 $1, %xmm0, %ymm1, %ymm0 ; AVX2-NEXT: vpshufhw {{.*#+}} xmm1 = xmm3[0,1,2,3,4,5,5,7] ; AVX2-NEXT: vpshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,6,5,7,7] ; AVX2-NEXT: vinserti128 $1, %xmm3, %ymm1, %ymm1 ; AVX2-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0,1,2,3,4],ymm0[5],ymm1[6],ymm0[7],ymm1[8,9,10,11,12],ymm0[13],ymm1[14],ymm0[15] ; AVX2-NEXT: vpshufhw {{.*#+}} xmm1 = xmm2[0,1,2,3,4,4,6,5] ; AVX2-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,6,6,7] ; AVX2-NEXT: vinserti128 $1, %xmm2, %ymm1, %ymm1 ; AVX2-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[2,1,3,3,6,5,7,7] ; AVX2-NEXT: vpshufd {{.*#+}} xmm2 = xmm5[2,3,2,3] ; AVX2-NEXT: vpmovzxwq {{.*#+}} ymm2 = xmm2[0],zero,zero,zero,xmm2[1],zero,zero,zero,xmm2[2],zero,zero,zero,xmm2[3],zero,zero,zero ; AVX2-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2,3,4],ymm1[5],ymm2[6,7,8],ymm1[9],ymm2[10,11,12],ymm1[13],ymm2[14,15] ; AVX2-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[0,2,2,3,4,6,6,7] ; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[2],ymm0[3],ymm1[4],ymm0[5],ymm1[6],ymm0[7] ; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-NEXT: vmovdqa 32(%r10), %xmm0 ; AVX2-NEXT: vmovdqa 32(%rax), %xmm1 ; AVX2-NEXT: vpunpckhbw {{.*#+}} xmm8 = 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] ; AVX2-NEXT: vpshuflw {{.*#+}} xmm2 = xmm8[0,0,2,1,4,5,6,7] ; AVX2-NEXT: vpshuflw {{.*#+}} xmm3 = xmm8[0,2,2,3,4,5,6,7] ; AVX2-NEXT: vinserti128 $1, %xmm3, %ymm2, %ymm2 ; AVX2-NEXT: vmovdqa 32(%r9), %xmm3 ; AVX2-NEXT: vmovdqa 32(%r8), %xmm5 ; AVX2-NEXT: vpunpckhbw {{.*#+}} xmm9 = xmm5[8],xmm3[8],xmm5[9],xmm3[9],xmm5[10],xmm3[10],xmm5[11],xmm3[11],xmm5[12],xmm3[12],xmm5[13],xmm3[13],xmm5[14],xmm3[14],xmm5[15],xmm3[15] ; AVX2-NEXT: vpshuflw {{.*#+}} xmm4 = xmm9[0,1,1,3,4,5,6,7] ; AVX2-NEXT: vpshuflw {{.*#+}} xmm6 = xmm9[2,1,3,3,4,5,6,7] ; AVX2-NEXT: vinserti128 $1, %xmm6, %ymm4, %ymm4 ; AVX2-NEXT: vpblendw {{.*#+}} ymm10 = ymm4[0],ymm2[1],ymm4[2],ymm2[3],ymm4[4,5,6,7,8],ymm2[9],ymm4[10],ymm2[11],ymm4[12,13,14,15] ; AVX2-NEXT: vmovdqa 32(%rcx), %xmm2 ; AVX2-NEXT: vmovdqa 32(%rdx), %xmm4 ; AVX2-NEXT: vpunpckhbw {{.*#+}} xmm11 = 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: vpshufb %xmm14, %xmm11, %xmm6 ; AVX2-NEXT: vpmovzxdq {{.*#+}} ymm12 = xmm6[0],zero,xmm6[1],zero,xmm6[2],zero,xmm6[3],zero ; AVX2-NEXT: vmovdqa 32(%rsi), %xmm6 ; AVX2-NEXT: vmovdqa 32(%rdi), %xmm7 ; AVX2-NEXT: vpunpckhbw {{.*#+}} xmm13 = 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] ; AVX2-NEXT: vpmovzxwq {{.*#+}} ymm15 = xmm13[0],zero,zero,zero,xmm13[1],zero,zero,zero,xmm13[2],zero,zero,zero,xmm13[3],zero,zero,zero ; AVX2-NEXT: vpblendw {{.*#+}} ymm12 = ymm15[0],ymm12[1],ymm15[2,3,4],ymm12[5],ymm15[6,7,8],ymm12[9],ymm15[10,11,12],ymm12[13],ymm15[14,15] ; AVX2-NEXT: vpshufd {{.*#+}} ymm10 = ymm10[0,0,2,1,4,4,6,5] ; AVX2-NEXT: vpblendd {{.*#+}} ymm10 = ymm12[0],ymm10[1],ymm12[2],ymm10[3],ymm12[4],ymm10[5],ymm12[6],ymm10[7] ; AVX2-NEXT: vmovdqu %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-NEXT: vpshufhw {{.*#+}} xmm10 = xmm8[0,1,2,3,4,4,6,5] ; AVX2-NEXT: vpshufhw {{.*#+}} xmm8 = xmm8[0,1,2,3,4,6,6,7] ; AVX2-NEXT: vinserti128 $1, %xmm8, %ymm10, %ymm8 ; AVX2-NEXT: vpshufhw {{.*#+}} xmm10 = xmm9[0,1,2,3,4,5,5,7] ; AVX2-NEXT: vpshufhw {{.*#+}} xmm9 = xmm9[0,1,2,3,6,5,7,7] ; AVX2-NEXT: vinserti128 $1, %xmm9, %ymm10, %ymm9 ; AVX2-NEXT: vpblendw {{.*#+}} ymm8 = ymm9[0,1,2,3,4],ymm8[5],ymm9[6],ymm8[7],ymm9[8,9,10,11,12],ymm8[13],ymm9[14],ymm8[15] ; AVX2-NEXT: vpshufhw {{.*#+}} xmm9 = xmm11[0,1,2,3,4,4,6,5] ; AVX2-NEXT: vpshufhw {{.*#+}} xmm10 = xmm11[0,1,2,3,4,6,6,7] ; AVX2-NEXT: vinserti128 $1, %xmm10, %ymm9, %ymm9 ; AVX2-NEXT: vpshufd {{.*#+}} ymm9 = ymm9[2,1,3,3,6,5,7,7] ; AVX2-NEXT: vpshufd {{.*#+}} xmm10 = xmm13[2,3,2,3] ; AVX2-NEXT: vpmovzxwq {{.*#+}} ymm10 = xmm10[0],zero,zero,zero,xmm10[1],zero,zero,zero,xmm10[2],zero,zero,zero,xmm10[3],zero,zero,zero ; AVX2-NEXT: vpblendw {{.*#+}} ymm9 = ymm10[0],ymm9[1],ymm10[2,3,4],ymm9[5],ymm10[6,7,8],ymm9[9],ymm10[10,11,12],ymm9[13],ymm10[14,15] ; AVX2-NEXT: vpshufd {{.*#+}} ymm8 = ymm8[0,2,2,3,4,6,6,7] ; AVX2-NEXT: vpblendd {{.*#+}} ymm8 = ymm9[0],ymm8[1],ymm9[2],ymm8[3],ymm9[4],ymm8[5],ymm9[6],ymm8[7] ; AVX2-NEXT: vmovdqu %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3],xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7] ; AVX2-NEXT: vpshuflw {{.*#+}} xmm1 = xmm0[0,0,2,1,4,5,6,7] ; AVX2-NEXT: vpshuflw {{.*#+}} xmm8 = xmm0[0,2,2,3,4,5,6,7] ; AVX2-NEXT: vinserti128 $1, %xmm8, %ymm1, %ymm1 ; AVX2-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm5[0],xmm3[0],xmm5[1],xmm3[1],xmm5[2],xmm3[2],xmm5[3],xmm3[3],xmm5[4],xmm3[4],xmm5[5],xmm3[5],xmm5[6],xmm3[6],xmm5[7],xmm3[7] ; AVX2-NEXT: vpshuflw {{.*#+}} xmm5 = xmm3[0,1,1,3,4,5,6,7] ; AVX2-NEXT: vpshuflw {{.*#+}} xmm8 = xmm3[2,1,3,3,4,5,6,7] ; AVX2-NEXT: vinserti128 $1, %xmm8, %ymm5, %ymm5 ; AVX2-NEXT: vpblendw {{.*#+}} ymm1 = ymm5[0],ymm1[1],ymm5[2],ymm1[3],ymm5[4,5,6,7,8],ymm1[9],ymm5[10],ymm1[11],ymm5[12,13,14,15] ; AVX2-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] ; AVX2-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm4[0],xmm2[0],xmm4[1],xmm2[1],xmm4[2],xmm2[2],xmm4[3],xmm2[3],xmm4[4],xmm2[4],xmm4[5],xmm2[5],xmm4[6],xmm2[6],xmm4[7],xmm2[7] ; AVX2-NEXT: vpshufb %xmm14, %xmm2, %xmm4 ; AVX2-NEXT: vpmovzxdq {{.*#+}} ymm4 = xmm4[0],zero,xmm4[1],zero,xmm4[2],zero,xmm4[3],zero ; AVX2-NEXT: vpmovzxwq {{.*#+}} ymm6 = xmm5[0],zero,zero,zero,xmm5[1],zero,zero,zero,xmm5[2],zero,zero,zero,xmm5[3],zero,zero,zero ; AVX2-NEXT: vpblendw {{.*#+}} ymm4 = ymm6[0],ymm4[1],ymm6[2,3,4],ymm4[5],ymm6[6,7,8],ymm4[9],ymm6[10,11,12],ymm4[13],ymm6[14,15] ; AVX2-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[0,0,2,1,4,4,6,5] ; AVX2-NEXT: vpblendd {{.*#+}} ymm1 = ymm4[0],ymm1[1],ymm4[2],ymm1[3],ymm4[4],ymm1[5],ymm4[6],ymm1[7] ; AVX2-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-NEXT: vpshufhw {{.*#+}} xmm1 = xmm0[0,1,2,3,4,4,6,5] ; AVX2-NEXT: vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,6,6,7] ; AVX2-NEXT: vinserti128 $1, %xmm0, %ymm1, %ymm0 ; AVX2-NEXT: vpshufhw {{.*#+}} xmm1 = xmm3[0,1,2,3,4,5,5,7] ; AVX2-NEXT: vpshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,6,5,7,7] ; AVX2-NEXT: vinserti128 $1, %xmm3, %ymm1, %ymm1 ; AVX2-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0,1,2,3,4],ymm0[5],ymm1[6],ymm0[7],ymm1[8,9,10,11,12],ymm0[13],ymm1[14],ymm0[15] ; AVX2-NEXT: vpshufhw {{.*#+}} xmm1 = xmm2[0,1,2,3,4,4,6,5] ; AVX2-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,6,6,7] ; AVX2-NEXT: vinserti128 $1, %xmm2, %ymm1, %ymm1 ; AVX2-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[2,1,3,3,6,5,7,7] ; AVX2-NEXT: vpshufd {{.*#+}} xmm2 = xmm5[2,3,2,3] ; AVX2-NEXT: vpmovzxwq {{.*#+}} ymm2 = xmm2[0],zero,zero,zero,xmm2[1],zero,zero,zero,xmm2[2],zero,zero,zero,xmm2[3],zero,zero,zero ; AVX2-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2,3,4],ymm1[5],ymm2[6,7,8],ymm1[9],ymm2[10,11,12],ymm1[13],ymm2[14,15] ; AVX2-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[0,2,2,3,4,6,6,7] ; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[2],ymm0[3],ymm1[4],ymm0[5],ymm1[6],ymm0[7] ; AVX2-NEXT: vmovdqu %ymm0, (%rsp) # 32-byte Spill ; AVX2-NEXT: vmovdqa 16(%r10), %xmm12 ; AVX2-NEXT: vmovdqa 16(%rax), %xmm10 ; AVX2-NEXT: vpunpckhbw {{.*#+}} xmm8 = xmm10[8],xmm12[8],xmm10[9],xmm12[9],xmm10[10],xmm12[10],xmm10[11],xmm12[11],xmm10[12],xmm12[12],xmm10[13],xmm12[13],xmm10[14],xmm12[14],xmm10[15],xmm12[15] ; AVX2-NEXT: vpshuflw {{.*#+}} xmm0 = xmm8[0,0,2,1,4,5,6,7] ; AVX2-NEXT: vpshuflw {{.*#+}} xmm1 = xmm8[0,2,2,3,4,5,6,7] ; AVX2-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm0 ; AVX2-NEXT: vmovdqa 16(%r9), %xmm5 ; AVX2-NEXT: vmovdqa 16(%r8), %xmm4 ; AVX2-NEXT: vpunpckhbw {{.*#+}} xmm11 = xmm4[8],xmm5[8],xmm4[9],xmm5[9],xmm4[10],xmm5[10],xmm4[11],xmm5[11],xmm4[12],xmm5[12],xmm4[13],xmm5[13],xmm4[14],xmm5[14],xmm4[15],xmm5[15] ; AVX2-NEXT: vpshuflw {{.*#+}} xmm1 = xmm11[0,1,1,3,4,5,6,7] ; AVX2-NEXT: vpshuflw {{.*#+}} xmm3 = xmm11[2,1,3,3,4,5,6,7] ; AVX2-NEXT: vinserti128 $1, %xmm3, %ymm1, %ymm1 ; AVX2-NEXT: vpblendw {{.*#+}} ymm9 = ymm1[0],ymm0[1],ymm1[2],ymm0[3],ymm1[4,5,6,7,8],ymm0[9],ymm1[10],ymm0[11],ymm1[12,13,14,15] ; AVX2-NEXT: vmovdqa 16(%rcx), %xmm7 ; AVX2-NEXT: vmovdqa 16(%rdx), %xmm6 ; AVX2-NEXT: vpunpckhbw {{.*#+}} xmm15 = xmm6[8],xmm7[8],xmm6[9],xmm7[9],xmm6[10],xmm7[10],xmm6[11],xmm7[11],xmm6[12],xmm7[12],xmm6[13],xmm7[13],xmm6[14],xmm7[14],xmm6[15],xmm7[15] ; AVX2-NEXT: vpshufb %xmm14, %xmm15, %xmm0 ; AVX2-NEXT: vpmovzxdq {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero ; AVX2-NEXT: vmovdqa 16(%rsi), %xmm3 ; AVX2-NEXT: vmovdqa 16(%rdi), %xmm2 ; AVX2-NEXT: vpunpckhbw {{.*#+}} xmm1 = 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] ; AVX2-NEXT: vpmovzxwq {{.*#+}} ymm13 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero,xmm1[2],zero,zero,zero,xmm1[3],zero,zero,zero ; AVX2-NEXT: vpblendw {{.*#+}} ymm0 = ymm13[0],ymm0[1],ymm13[2,3,4],ymm0[5],ymm13[6,7,8],ymm0[9],ymm13[10,11,12],ymm0[13],ymm13[14,15] ; AVX2-NEXT: vpshufd {{.*#+}} ymm9 = ymm9[0,0,2,1,4,4,6,5] ; AVX2-NEXT: vpblendd {{.*#+}} ymm9 = ymm0[0],ymm9[1],ymm0[2],ymm9[3],ymm0[4],ymm9[5],ymm0[6],ymm9[7] ; AVX2-NEXT: vpshufhw {{.*#+}} xmm0 = xmm8[0,1,2,3,4,4,6,5] ; AVX2-NEXT: vpshufhw {{.*#+}} xmm8 = xmm8[0,1,2,3,4,6,6,7] ; AVX2-NEXT: vinserti128 $1, %xmm8, %ymm0, %ymm0 ; AVX2-NEXT: vpshufhw {{.*#+}} xmm8 = xmm11[0,1,2,3,4,5,5,7] ; AVX2-NEXT: vpshufhw {{.*#+}} xmm11 = xmm11[0,1,2,3,6,5,7,7] ; AVX2-NEXT: vinserti128 $1, %xmm11, %ymm8, %ymm8 ; AVX2-NEXT: vpblendw {{.*#+}} ymm0 = ymm8[0,1,2,3,4],ymm0[5],ymm8[6],ymm0[7],ymm8[8,9,10,11,12],ymm0[13],ymm8[14],ymm0[15] ; AVX2-NEXT: vpshufhw {{.*#+}} xmm8 = xmm15[0,1,2,3,4,4,6,5] ; AVX2-NEXT: vpshufhw {{.*#+}} xmm11 = xmm15[0,1,2,3,4,6,6,7] ; AVX2-NEXT: vinserti128 $1, %xmm11, %ymm8, %ymm8 ; AVX2-NEXT: vpshufd {{.*#+}} ymm8 = ymm8[2,1,3,3,6,5,7,7] ; AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[2,3,2,3] ; AVX2-NEXT: vpmovzxwq {{.*#+}} ymm1 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero,xmm1[2],zero,zero,zero,xmm1[3],zero,zero,zero ; AVX2-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0],ymm8[1],ymm1[2,3,4],ymm8[5],ymm1[6,7,8],ymm8[9],ymm1[10,11,12],ymm8[13],ymm1[14,15] ; AVX2-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[0,2,2,3,4,6,6,7] ; AVX2-NEXT: vpblendd {{.*#+}} ymm8 = ymm1[0],ymm0[1],ymm1[2],ymm0[3],ymm1[4],ymm0[5],ymm1[6],ymm0[7] ; AVX2-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm10[0],xmm12[0],xmm10[1],xmm12[1],xmm10[2],xmm12[2],xmm10[3],xmm12[3],xmm10[4],xmm12[4],xmm10[5],xmm12[5],xmm10[6],xmm12[6],xmm10[7],xmm12[7] ; AVX2-NEXT: vpshuflw {{.*#+}} xmm0 = xmm1[0,0,2,1,4,5,6,7] ; AVX2-NEXT: vpshuflw {{.*#+}} xmm10 = xmm1[0,2,2,3,4,5,6,7] ; AVX2-NEXT: vinserti128 $1, %xmm10, %ymm0, %ymm0 ; AVX2-NEXT: vpunpcklbw {{.*#+}} xmm4 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3],xmm4[4],xmm5[4],xmm4[5],xmm5[5],xmm4[6],xmm5[6],xmm4[7],xmm5[7] ; AVX2-NEXT: vpshuflw {{.*#+}} xmm5 = xmm4[0,1,1,3,4,5,6,7] ; AVX2-NEXT: vpshuflw {{.*#+}} xmm10 = xmm4[2,1,3,3,4,5,6,7] ; AVX2-NEXT: vinserti128 $1, %xmm10, %ymm5, %ymm5 ; AVX2-NEXT: vpblendw {{.*#+}} ymm0 = ymm5[0],ymm0[1],ymm5[2],ymm0[3],ymm5[4,5,6,7,8],ymm0[9],ymm5[10],ymm0[11],ymm5[12,13,14,15] ; AVX2-NEXT: vpunpcklbw {{.*#+}} 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] ; AVX2-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm6[0],xmm7[0],xmm6[1],xmm7[1],xmm6[2],xmm7[2],xmm6[3],xmm7[3],xmm6[4],xmm7[4],xmm6[5],xmm7[5],xmm6[6],xmm7[6],xmm6[7],xmm7[7] ; AVX2-NEXT: vpshufb %xmm14, %xmm3, %xmm5 ; AVX2-NEXT: vpmovzxdq {{.*#+}} ymm5 = xmm5[0],zero,xmm5[1],zero,xmm5[2],zero,xmm5[3],zero ; AVX2-NEXT: vpmovzxwq {{.*#+}} ymm6 = xmm2[0],zero,zero,zero,xmm2[1],zero,zero,zero,xmm2[2],zero,zero,zero,xmm2[3],zero,zero,zero ; AVX2-NEXT: vpblendw {{.*#+}} ymm5 = ymm6[0],ymm5[1],ymm6[2,3,4],ymm5[5],ymm6[6,7,8],ymm5[9],ymm6[10,11,12],ymm5[13],ymm6[14,15] ; AVX2-NEXT: vpshufd {{.*#+}} ymm0 = ymm0[0,0,2,1,4,4,6,5] ; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm5[0],ymm0[1],ymm5[2],ymm0[3],ymm5[4],ymm0[5],ymm5[6],ymm0[7] ; AVX2-NEXT: vpshufhw {{.*#+}} xmm5 = xmm1[0,1,2,3,4,4,6,5] ; AVX2-NEXT: vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,6,6,7] ; AVX2-NEXT: vinserti128 $1, %xmm1, %ymm5, %ymm1 ; AVX2-NEXT: vpshufhw {{.*#+}} xmm5 = xmm4[0,1,2,3,4,5,5,7] ; AVX2-NEXT: vpshufhw {{.*#+}} xmm4 = xmm4[0,1,2,3,6,5,7,7] ; AVX2-NEXT: vinserti128 $1, %xmm4, %ymm5, %ymm4 ; AVX2-NEXT: vpblendw {{.*#+}} ymm1 = ymm4[0,1,2,3,4],ymm1[5],ymm4[6],ymm1[7],ymm4[8,9,10,11,12],ymm1[13],ymm4[14],ymm1[15] ; AVX2-NEXT: vpshufhw {{.*#+}} xmm4 = xmm3[0,1,2,3,4,4,6,5] ; AVX2-NEXT: vpshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,4,6,6,7] ; AVX2-NEXT: vinserti128 $1, %xmm3, %ymm4, %ymm3 ; AVX2-NEXT: vpshufd {{.*#+}} ymm3 = ymm3[2,1,3,3,6,5,7,7] ; AVX2-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[2,3,2,3] ; AVX2-NEXT: vpmovzxwq {{.*#+}} ymm2 = xmm2[0],zero,zero,zero,xmm2[1],zero,zero,zero,xmm2[2],zero,zero,zero,xmm2[3],zero,zero,zero ; AVX2-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0],ymm3[1],ymm2[2,3,4],ymm3[5],ymm2[6,7,8],ymm3[9],ymm2[10,11,12],ymm3[13],ymm2[14,15] ; AVX2-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[0,2,2,3,4,6,6,7] ; AVX2-NEXT: vpblendd {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2],ymm1[3],ymm2[4],ymm1[5],ymm2[6],ymm1[7] ; AVX2-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload ; AVX2-NEXT: vpunpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm2 # 16-byte Folded Reload ; AVX2-NEXT: # xmm2 = xmm2[8],mem[8],xmm2[9],mem[9],xmm2[10],mem[10],xmm2[11],mem[11],xmm2[12],mem[12],xmm2[13],mem[13],xmm2[14],mem[14],xmm2[15],mem[15] ; AVX2-NEXT: vpshuflw {{.*#+}} xmm3 = xmm2[0,0,2,1,4,5,6,7] ; AVX2-NEXT: vpshuflw {{.*#+}} xmm4 = xmm2[0,2,2,3,4,5,6,7] ; AVX2-NEXT: vinserti128 $1, %xmm4, %ymm3, %ymm3 ; AVX2-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload ; AVX2-NEXT: vpunpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm4, %xmm4 # 16-byte Folded Reload ; AVX2-NEXT: # xmm4 = xmm4[8],mem[8],xmm4[9],mem[9],xmm4[10],mem[10],xmm4[11],mem[11],xmm4[12],mem[12],xmm4[13],mem[13],xmm4[14],mem[14],xmm4[15],mem[15] ; AVX2-NEXT: vpshuflw {{.*#+}} xmm5 = xmm4[0,1,1,3,4,5,6,7] ; AVX2-NEXT: vpshuflw {{.*#+}} xmm6 = xmm4[2,1,3,3,4,5,6,7] ; AVX2-NEXT: vinserti128 $1, %xmm6, %ymm5, %ymm5 ; AVX2-NEXT: vpblendw {{.*#+}} ymm3 = ymm5[0],ymm3[1],ymm5[2],ymm3[3],ymm5[4,5,6,7,8],ymm3[9],ymm5[10],ymm3[11],ymm5[12,13,14,15] ; AVX2-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload ; AVX2-NEXT: vpunpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm5, %xmm5 # 16-byte Folded Reload ; AVX2-NEXT: # xmm5 = xmm5[8],mem[8],xmm5[9],mem[9],xmm5[10],mem[10],xmm5[11],mem[11],xmm5[12],mem[12],xmm5[13],mem[13],xmm5[14],mem[14],xmm5[15],mem[15] ; AVX2-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload ; AVX2-NEXT: vpunpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm6, %xmm6 # 16-byte Folded Reload ; AVX2-NEXT: # xmm6 = xmm6[8],mem[8],xmm6[9],mem[9],xmm6[10],mem[10],xmm6[11],mem[11],xmm6[12],mem[12],xmm6[13],mem[13],xmm6[14],mem[14],xmm6[15],mem[15] ; AVX2-NEXT: vpshufb %xmm14, %xmm6, %xmm7 ; AVX2-NEXT: vpmovzxdq {{.*#+}} ymm7 = xmm7[0],zero,xmm7[1],zero,xmm7[2],zero,xmm7[3],zero ; AVX2-NEXT: vpmovzxwq {{.*#+}} ymm10 = xmm5[0],zero,zero,zero,xmm5[1],zero,zero,zero,xmm5[2],zero,zero,zero,xmm5[3],zero,zero,zero ; AVX2-NEXT: vpblendw {{.*#+}} ymm7 = ymm10[0],ymm7[1],ymm10[2,3,4],ymm7[5],ymm10[6,7,8],ymm7[9],ymm10[10,11,12],ymm7[13],ymm10[14,15] ; AVX2-NEXT: vpshufd {{.*#+}} ymm3 = ymm3[0,0,2,1,4,4,6,5] ; AVX2-NEXT: vpblendd {{.*#+}} ymm3 = ymm7[0],ymm3[1],ymm7[2],ymm3[3],ymm7[4],ymm3[5],ymm7[6],ymm3[7] ; AVX2-NEXT: vpshufhw {{.*#+}} xmm7 = xmm2[0,1,2,3,4,4,6,5] ; AVX2-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,6,6,7] ; AVX2-NEXT: vinserti128 $1, %xmm2, %ymm7, %ymm2 ; AVX2-NEXT: vpshufhw {{.*#+}} xmm7 = xmm4[0,1,2,3,4,5,5,7] ; AVX2-NEXT: vpshufhw {{.*#+}} xmm4 = xmm4[0,1,2,3,6,5,7,7] ; AVX2-NEXT: vinserti128 $1, %xmm4, %ymm7, %ymm4 ; AVX2-NEXT: vpblendw {{.*#+}} ymm2 = ymm4[0,1,2,3,4],ymm2[5],ymm4[6],ymm2[7],ymm4[8,9,10,11,12],ymm2[13],ymm4[14],ymm2[15] ; AVX2-NEXT: vpshufhw {{.*#+}} xmm4 = xmm6[0,1,2,3,4,4,6,5] ; AVX2-NEXT: vpshufhw {{.*#+}} xmm6 = xmm6[0,1,2,3,4,6,6,7] ; AVX2-NEXT: vinserti128 $1, %xmm6, %ymm4, %ymm4 ; AVX2-NEXT: vpshufd {{.*#+}} ymm4 = ymm4[2,1,3,3,6,5,7,7] ; AVX2-NEXT: vpshufd {{.*#+}} xmm5 = xmm5[2,3,2,3] ; AVX2-NEXT: vpmovzxwq {{.*#+}} ymm5 = xmm5[0],zero,zero,zero,xmm5[1],zero,zero,zero,xmm5[2],zero,zero,zero,xmm5[3],zero,zero,zero ; AVX2-NEXT: vpblendw {{.*#+}} ymm4 = ymm5[0],ymm4[1],ymm5[2,3,4],ymm4[5],ymm5[6,7,8],ymm4[9],ymm5[10,11,12],ymm4[13],ymm5[14,15] ; AVX2-NEXT: vpshufd {{.*#+}} ymm2 = ymm2[0,2,2,3,4,6,6,7] ; AVX2-NEXT: vpblendd {{.*#+}} ymm2 = ymm4[0],ymm2[1],ymm4[2],ymm2[3],ymm4[4],ymm2[5],ymm4[6],ymm2[7] ; AVX2-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX2-NEXT: vmovdqa %ymm2, 96(%rax) ; AVX2-NEXT: vmovdqa %ymm3, 64(%rax) ; AVX2-NEXT: vmovdqa %ymm1, 160(%rax) ; AVX2-NEXT: vmovdqa %ymm0, 128(%rax) ; AVX2-NEXT: vmovdqa %ymm8, 224(%rax) ; AVX2-NEXT: vmovdqa %ymm9, 192(%rax) ; AVX2-NEXT: vmovups (%rsp), %ymm0 # 32-byte Reload ; AVX2-NEXT: vmovaps %ymm0, 288(%rax) ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload ; AVX2-NEXT: vmovaps %ymm0, 256(%rax) ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload ; AVX2-NEXT: vmovaps %ymm0, 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, 416(%rax) ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload ; AVX2-NEXT: vmovaps %ymm0, 384(%rax) ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload ; AVX2-NEXT: vmovaps %ymm0, 480(%rax) ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload ; AVX2-NEXT: vmovaps %ymm0, 448(%rax) ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload ; AVX2-NEXT: vmovaps %ymm0, (%rax) ; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload ; AVX2-NEXT: vmovaps %ymm0, 32(%rax) ; AVX2-NEXT: addq $328, %rsp # imm = 0x148 ; AVX2-NEXT: vzeroupper ; AVX2-NEXT: retq ; ; AVX2-FP-LABEL: store_i8_stride8_vf64: ; AVX2-FP: # %bb.0: ; AVX2-FP-NEXT: subq $392, %rsp # imm = 0x188 ; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX2-FP-NEXT: vmovdqa (%r10), %xmm1 ; AVX2-FP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX2-FP-NEXT: vmovdqa (%rax), %xmm0 ; AVX2-FP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX2-FP-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] ; AVX2-FP-NEXT: vinserti128 $1, %xmm0, %ymm0, %ymm0 ; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm2 = ymm0[0,1,2,3,8,9,8,9,8,9,8,9,12,13,10,11,16,17,18,19,24,25,28,29,24,25,28,29,28,29,30,31] ; AVX2-FP-NEXT: vmovdqa (%r9), %xmm3 ; AVX2-FP-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX2-FP-NEXT: vmovdqa (%r8), %xmm1 ; AVX2-FP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX2-FP-NEXT: vpunpcklbw {{.*#+}} 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] ; AVX2-FP-NEXT: vinserti128 $1, %xmm1, %ymm1, %ymm1 ; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm3 = ymm1[0,1,2,3,8,9,10,11,8,9,10,11,10,11,14,15,16,17,18,19,28,29,26,27,28,29,26,27,30,31,30,31] ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0,1,2],ymm2[3],ymm3[4,5,6],ymm2[7],ymm3[8,9,10],ymm2[11],ymm3[12,13,14],ymm2[15] ; AVX2-FP-NEXT: vmovdqa (%rcx), %xmm4 ; AVX2-FP-NEXT: vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX2-FP-NEXT: vmovdqa (%rdx), %xmm2 ; AVX2-FP-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX2-FP-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3],xmm2[4],xmm4[4],xmm2[5],xmm4[5],xmm2[6],xmm4[6],xmm2[7],xmm4[7] ; AVX2-FP-NEXT: vinserti128 $1, %xmm2, %ymm2, %ymm2 ; AVX2-FP-NEXT: vmovdqa (%rsi), %xmm5 ; AVX2-FP-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX2-FP-NEXT: vmovdqa (%rdi), %xmm4 ; AVX2-FP-NEXT: vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX2-FP-NEXT: vpunpcklbw {{.*#+}} xmm4 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3],xmm4[4],xmm5[4],xmm4[5],xmm5[5],xmm4[6],xmm5[6],xmm4[7],xmm5[7] ; AVX2-FP-NEXT: vinserti128 $1, %xmm4, %ymm4, %ymm5 ; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm5 = ymm5[8,9,u,u,u,u,u,u,10,11,u,u,u,u,u,u,28,29,u,u,u,u,u,u,30,31,u,u,u,u,u,u] ; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm6 = ymm2[8,9,8,9,4,5,6,7,12,13,10,11,12,13,10,11,24,25,28,29,20,21,22,23,28,29,30,31,28,29,30,31] ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm5 = ymm5[0],ymm6[1],ymm5[2,3,4],ymm6[5],ymm5[6,7,8],ymm6[9],ymm5[10,11,12],ymm6[13],ymm5[14,15] ; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm3 = ymm5[0],ymm3[1],ymm5[2],ymm3[3],ymm5[4],ymm3[5],ymm5[6],ymm3[7] ; AVX2-FP-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm9 = [0,1,0,1,0,1,0,1,8,9,10,11,4,5,2,3,0,1,4,5,0,1,4,5,8,9,10,11,4,5,6,7] ; AVX2-FP-NEXT: vpshufb %ymm9, %ymm0, %ymm0 ; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm10 = [0,1,2,3,0,1,2,3,8,9,10,11,2,3,6,7,4,5,2,3,4,5,2,3,8,9,10,11,6,7,6,7] ; AVX2-FP-NEXT: vpshufb %ymm10, %ymm1, %ymm1 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3],ymm1[4,5,6],ymm0[7],ymm1[8,9,10],ymm0[11],ymm1[12,13,14],ymm0[15] ; AVX2-FP-NEXT: vpmovsxwq {{.*#+}} xmm13 = [1284,1798] ; AVX2-FP-NEXT: vpshufb %xmm13, %xmm4, %xmm1 ; AVX2-FP-NEXT: vpmovzxwq {{.*#+}} xmm3 = xmm4[0],zero,zero,zero,xmm4[1],zero,zero,zero ; AVX2-FP-NEXT: vinserti128 $1, %xmm1, %ymm3, %ymm1 ; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm12 = [0,1,0,1,4,5,2,3,4,5,2,3,12,13,14,15,0,1,4,5,4,5,6,7,4,5,6,7,12,13,14,15] ; AVX2-FP-NEXT: vpshufb %ymm12, %ymm2, %ymm2 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3,4],ymm2[5],ymm1[6,7,8],ymm2[9],ymm1[10,11,12],ymm2[13],ymm1[14,15] ; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[2],ymm0[3],ymm1[4],ymm0[5],ymm1[6],ymm0[7] ; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FP-NEXT: vmovdqa 48(%r10), %xmm8 ; AVX2-FP-NEXT: vmovdqa 48(%rax), %xmm3 ; AVX2-FP-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm3[8],xmm8[8],xmm3[9],xmm8[9],xmm3[10],xmm8[10],xmm3[11],xmm8[11],xmm3[12],xmm8[12],xmm3[13],xmm8[13],xmm3[14],xmm8[14],xmm3[15],xmm8[15] ; AVX2-FP-NEXT: vinserti128 $1, %xmm0, %ymm0, %ymm6 ; AVX2-FP-NEXT: vmovdqa 48(%r9), %xmm4 ; AVX2-FP-NEXT: vmovdqa 48(%r8), %xmm5 ; AVX2-FP-NEXT: vpunpckhbw {{.*#+}} xmm0 = 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] ; AVX2-FP-NEXT: vinserti128 $1, %xmm0, %ymm0, %ymm7 ; AVX2-FP-NEXT: vpshufb %ymm9, %ymm6, %ymm0 ; AVX2-FP-NEXT: vpshufb %ymm10, %ymm7, %ymm2 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm11 = ymm2[0,1,2],ymm0[3],ymm2[4,5,6],ymm0[7],ymm2[8,9,10],ymm0[11],ymm2[12,13,14],ymm0[15] ; AVX2-FP-NEXT: vmovdqa 48(%rsi), %xmm10 ; AVX2-FP-NEXT: vmovdqa 48(%rdi), %xmm9 ; AVX2-FP-NEXT: vpunpckhbw {{.*#+}} xmm0 = 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-FP-NEXT: vpshufb %xmm13, %xmm0, %xmm1 ; AVX2-FP-NEXT: vpmovzxwq {{.*#+}} xmm13 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero ; AVX2-FP-NEXT: vinserti128 $1, %xmm1, %ymm13, %ymm1 ; AVX2-FP-NEXT: vmovdqa 48(%rcx), %xmm13 ; AVX2-FP-NEXT: vmovdqa 48(%rdx), %xmm2 ; AVX2-FP-NEXT: vpunpckhbw {{.*#+}} xmm15 = xmm2[8],xmm13[8],xmm2[9],xmm13[9],xmm2[10],xmm13[10],xmm2[11],xmm13[11],xmm2[12],xmm13[12],xmm2[13],xmm13[13],xmm2[14],xmm13[14],xmm2[15],xmm13[15] ; AVX2-FP-NEXT: vinserti128 $1, %xmm15, %ymm15, %ymm15 ; AVX2-FP-NEXT: vpshufb %ymm12, %ymm15, %ymm14 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0],ymm14[1],ymm1[2,3,4],ymm14[5],ymm1[6,7,8],ymm14[9],ymm1[10,11,12],ymm14[13],ymm1[14,15] ; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0],ymm11[1],ymm1[2],ymm11[3],ymm1[4],ymm11[5],ymm1[6],ymm11[7] ; AVX2-FP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm12 = [0,1,2,3,8,9,8,9,8,9,8,9,12,13,10,11,0,1,2,3,8,9,12,13,8,9,12,13,12,13,14,15] ; AVX2-FP-NEXT: vpshufb %ymm12, %ymm6, %ymm1 ; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm11 = [0,1,2,3,8,9,10,11,8,9,10,11,10,11,14,15,0,1,2,3,12,13,10,11,12,13,10,11,14,15,14,15] ; AVX2-FP-NEXT: vpshufb %ymm11, %ymm7, %ymm6 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm1 = ymm6[0,1,2],ymm1[3],ymm6[4,5,6],ymm1[7],ymm6[8,9,10],ymm1[11],ymm6[12,13,14],ymm1[15] ; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm7 = [8,9,8,9,4,5,6,7,12,13,10,11,12,13,10,11,8,9,12,13,4,5,6,7,12,13,14,15,12,13,14,15] ; AVX2-FP-NEXT: vpshufb %ymm7, %ymm15, %ymm6 ; AVX2-FP-NEXT: vinserti128 $1, %xmm0, %ymm0, %ymm0 ; AVX2-FP-NEXT: vpmovsxwq {{.*#+}} ymm14 = [2312,2826,3340,3854] ; AVX2-FP-NEXT: vpshufb %ymm14, %ymm0, %ymm0 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm6[1],ymm0[2,3,4],ymm6[5],ymm0[6,7,8],ymm6[9],ymm0[10,11,12],ymm6[13],ymm0[14,15] ; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2],ymm1[3],ymm0[4],ymm1[5],ymm0[6],ymm1[7] ; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FP-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm3[0],xmm8[0],xmm3[1],xmm8[1],xmm3[2],xmm8[2],xmm3[3],xmm8[3],xmm3[4],xmm8[4],xmm3[5],xmm8[5],xmm3[6],xmm8[6],xmm3[7],xmm8[7] ; AVX2-FP-NEXT: vpunpcklbw {{.*#+}} xmm1 = 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: vinserti128 $1, %xmm0, %ymm0, %ymm0 ; AVX2-FP-NEXT: vinserti128 $1, %xmm1, %ymm1, %ymm1 ; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm15 = [0,1,0,1,0,1,0,1,8,9,10,11,4,5,2,3,0,1,4,5,0,1,4,5,8,9,10,11,4,5,6,7] ; AVX2-FP-NEXT: vpshufb %ymm15, %ymm0, %ymm3 ; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm8 = [0,1,2,3,0,1,2,3,8,9,10,11,2,3,6,7,4,5,2,3,4,5,2,3,8,9,10,11,6,7,6,7] ; AVX2-FP-NEXT: vpshufb %ymm8, %ymm1, %ymm4 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm3 = ymm4[0,1,2],ymm3[3],ymm4[4,5,6],ymm3[7],ymm4[8,9,10],ymm3[11],ymm4[12,13,14],ymm3[15] ; AVX2-FP-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm2[0],xmm13[0],xmm2[1],xmm13[1],xmm2[2],xmm13[2],xmm2[3],xmm13[3],xmm2[4],xmm13[4],xmm2[5],xmm13[5],xmm2[6],xmm13[6],xmm2[7],xmm13[7] ; AVX2-FP-NEXT: vpunpcklbw {{.*#+}} xmm4 = xmm9[0],xmm10[0],xmm9[1],xmm10[1],xmm9[2],xmm10[2],xmm9[3],xmm10[3],xmm9[4],xmm10[4],xmm9[5],xmm10[5],xmm9[6],xmm10[6],xmm9[7],xmm10[7] ; AVX2-FP-NEXT: vpmovsxwq {{.*#+}} xmm10 = [1284,1798] ; AVX2-FP-NEXT: vpshufb %xmm10, %xmm4, %xmm5 ; AVX2-FP-NEXT: vpmovzxwq {{.*#+}} xmm6 = xmm4[0],zero,zero,zero,xmm4[1],zero,zero,zero ; AVX2-FP-NEXT: vinserti128 $1, %xmm5, %ymm6, %ymm5 ; AVX2-FP-NEXT: vinserti128 $1, %xmm2, %ymm2, %ymm2 ; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm6 = ymm2[0,1,0,1,4,5,2,3,4,5,2,3,12,13,14,15,16,17,20,21,20,21,22,23,20,21,22,23,28,29,30,31] ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm5 = ymm5[0],ymm6[1],ymm5[2,3,4],ymm6[5],ymm5[6,7,8],ymm6[9],ymm5[10,11,12],ymm6[13],ymm5[14,15] ; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm3 = ymm5[0],ymm3[1],ymm5[2],ymm3[3],ymm5[4],ymm3[5],ymm5[6],ymm3[7] ; AVX2-FP-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FP-NEXT: vpshufb %ymm12, %ymm0, %ymm0 ; AVX2-FP-NEXT: vpshufb %ymm11, %ymm1, %ymm1 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3],ymm1[4,5,6],ymm0[7],ymm1[8,9,10],ymm0[11],ymm1[12,13,14],ymm0[15] ; AVX2-FP-NEXT: vpshufb %ymm7, %ymm2, %ymm1 ; AVX2-FP-NEXT: vinserti128 $1, %xmm4, %ymm4, %ymm2 ; AVX2-FP-NEXT: vpshufb %ymm14, %ymm2, %ymm2 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2,3,4],ymm1[5],ymm2[6,7,8],ymm1[9],ymm2[10,11,12],ymm1[13],ymm2[14,15] ; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[2],ymm0[3],ymm1[4],ymm0[5],ymm1[6],ymm0[7] ; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FP-NEXT: vmovdqa 32(%r10), %xmm1 ; AVX2-FP-NEXT: vmovdqa 32(%rax), %xmm3 ; AVX2-FP-NEXT: vpunpckhbw {{.*#+}} xmm0 = 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] ; AVX2-FP-NEXT: vinserti128 $1, %xmm0, %ymm0, %ymm6 ; AVX2-FP-NEXT: vmovdqa 32(%r9), %xmm4 ; AVX2-FP-NEXT: vmovdqa 32(%r8), %xmm5 ; AVX2-FP-NEXT: vpunpckhbw {{.*#+}} xmm0 = 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] ; AVX2-FP-NEXT: vinserti128 $1, %xmm0, %ymm0, %ymm7 ; AVX2-FP-NEXT: vpshufb %ymm15, %ymm6, %ymm0 ; AVX2-FP-NEXT: vpshufb %ymm8, %ymm7, %ymm2 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm8 = ymm2[0,1,2],ymm0[3],ymm2[4,5,6],ymm0[7],ymm2[8,9,10],ymm0[11],ymm2[12,13,14],ymm0[15] ; AVX2-FP-NEXT: vmovdqa 32(%rsi), %xmm0 ; AVX2-FP-NEXT: vmovdqa 32(%rdi), %xmm2 ; AVX2-FP-NEXT: vpunpckhbw {{.*#+}} xmm9 = xmm2[8],xmm0[8],xmm2[9],xmm0[9],xmm2[10],xmm0[10],xmm2[11],xmm0[11],xmm2[12],xmm0[12],xmm2[13],xmm0[13],xmm2[14],xmm0[14],xmm2[15],xmm0[15] ; AVX2-FP-NEXT: vpshufb %xmm10, %xmm9, %xmm10 ; AVX2-FP-NEXT: vpmovzxwq {{.*#+}} xmm11 = xmm9[0],zero,zero,zero,xmm9[1],zero,zero,zero ; AVX2-FP-NEXT: vinserti128 $1, %xmm10, %ymm11, %ymm10 ; AVX2-FP-NEXT: vmovdqa 32(%rcx), %xmm11 ; AVX2-FP-NEXT: vmovdqa 32(%rdx), %xmm13 ; AVX2-FP-NEXT: vpunpckhbw {{.*#+}} xmm14 = 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-FP-NEXT: vinserti128 $1, %xmm14, %ymm14, %ymm14 ; AVX2-FP-NEXT: vpshufb {{.*#+}} ymm15 = ymm14[0,1,0,1,4,5,2,3,4,5,2,3,12,13,14,15,16,17,20,21,20,21,22,23,20,21,22,23,28,29,30,31] ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm10 = ymm10[0],ymm15[1],ymm10[2,3,4],ymm15[5],ymm10[6,7,8],ymm15[9],ymm10[10,11,12],ymm15[13],ymm10[14,15] ; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm8 = ymm10[0],ymm8[1],ymm10[2],ymm8[3],ymm10[4],ymm8[5],ymm10[6],ymm8[7] ; AVX2-FP-NEXT: vmovdqu %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FP-NEXT: vmovdqa %ymm12, %ymm10 ; AVX2-FP-NEXT: vpshufb %ymm12, %ymm6, %ymm6 ; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm15 = [0,1,2,3,8,9,10,11,8,9,10,11,10,11,14,15,0,1,2,3,12,13,10,11,12,13,10,11,14,15,14,15] ; AVX2-FP-NEXT: vpshufb %ymm15, %ymm7, %ymm7 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm6 = ymm7[0,1,2],ymm6[3],ymm7[4,5,6],ymm6[7],ymm7[8,9,10],ymm6[11],ymm7[12,13,14],ymm6[15] ; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm12 = [8,9,8,9,4,5,6,7,12,13,10,11,12,13,10,11,8,9,12,13,4,5,6,7,12,13,14,15,12,13,14,15] ; AVX2-FP-NEXT: vpshufb %ymm12, %ymm14, %ymm7 ; AVX2-FP-NEXT: vinserti128 $1, %xmm9, %ymm9, %ymm8 ; AVX2-FP-NEXT: vpmovsxwq {{.*#+}} ymm14 = [2312,2826,3340,3854] ; AVX2-FP-NEXT: vpshufb %ymm14, %ymm8, %ymm8 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm7 = ymm8[0],ymm7[1],ymm8[2,3,4],ymm7[5],ymm8[6,7,8],ymm7[9],ymm8[10,11,12],ymm7[13],ymm8[14,15] ; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm6 = ymm7[0],ymm6[1],ymm7[2],ymm6[3],ymm7[4],ymm6[5],ymm7[6],ymm6[7] ; AVX2-FP-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FP-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3],xmm3[4],xmm1[4],xmm3[5],xmm1[5],xmm3[6],xmm1[6],xmm3[7],xmm1[7] ; AVX2-FP-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] ; AVX2-FP-NEXT: vinserti128 $1, %xmm1, %ymm1, %ymm1 ; AVX2-FP-NEXT: vinserti128 $1, %xmm3, %ymm3, %ymm3 ; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm9 = [0,1,0,1,0,1,0,1,8,9,10,11,4,5,2,3,0,1,4,5,0,1,4,5,8,9,10,11,4,5,6,7] ; AVX2-FP-NEXT: vpshufb %ymm9, %ymm1, %ymm4 ; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm7 = [0,1,2,3,0,1,2,3,8,9,10,11,2,3,6,7,4,5,2,3,4,5,2,3,8,9,10,11,6,7,6,7] ; AVX2-FP-NEXT: vpshufb %ymm7, %ymm3, %ymm5 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm4 = ymm5[0,1,2],ymm4[3],ymm5[4,5,6],ymm4[7],ymm5[8,9,10],ymm4[11],ymm5[12,13,14],ymm4[15] ; AVX2-FP-NEXT: vpunpcklbw {{.*#+}} xmm5 = 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-FP-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3],xmm2[4],xmm0[4],xmm2[5],xmm0[5],xmm2[6],xmm0[6],xmm2[7],xmm0[7] ; AVX2-FP-NEXT: vpmovsxwq {{.*#+}} xmm11 = [1284,1798] ; AVX2-FP-NEXT: vpshufb %xmm11, %xmm0, %xmm2 ; AVX2-FP-NEXT: vpmovzxwq {{.*#+}} xmm6 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero ; AVX2-FP-NEXT: vinserti128 $1, %xmm2, %ymm6, %ymm2 ; AVX2-FP-NEXT: vinserti128 $1, %xmm5, %ymm5, %ymm5 ; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm6 = [0,1,0,1,4,5,2,3,4,5,2,3,12,13,14,15,0,1,4,5,4,5,6,7,4,5,6,7,12,13,14,15] ; AVX2-FP-NEXT: vpshufb %ymm6, %ymm5, %ymm6 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0],ymm6[1],ymm2[2,3,4],ymm6[5],ymm2[6,7,8],ymm6[9],ymm2[10,11,12],ymm6[13],ymm2[14,15] ; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0],ymm4[1],ymm2[2],ymm4[3],ymm2[4],ymm4[5],ymm2[6],ymm4[7] ; AVX2-FP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FP-NEXT: vpshufb %ymm10, %ymm1, %ymm1 ; AVX2-FP-NEXT: vpshufb %ymm15, %ymm3, %ymm2 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0,1,2],ymm1[3],ymm2[4,5,6],ymm1[7],ymm2[8,9,10],ymm1[11],ymm2[12,13,14],ymm1[15] ; AVX2-FP-NEXT: vpshufb %ymm12, %ymm5, %ymm2 ; AVX2-FP-NEXT: vinserti128 $1, %xmm0, %ymm0, %ymm0 ; AVX2-FP-NEXT: vpshufb %ymm14, %ymm0, %ymm0 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm2[1],ymm0[2,3,4],ymm2[5],ymm0[6,7,8],ymm2[9],ymm0[10,11,12],ymm2[13],ymm0[14,15] ; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2],ymm1[3],ymm0[4],ymm1[5],ymm0[6],ymm1[7] ; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FP-NEXT: vmovdqa 16(%r10), %xmm4 ; AVX2-FP-NEXT: vmovdqa 16(%rax), %xmm2 ; AVX2-FP-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm2[8],xmm4[8],xmm2[9],xmm4[9],xmm2[10],xmm4[10],xmm2[11],xmm4[11],xmm2[12],xmm4[12],xmm2[13],xmm4[13],xmm2[14],xmm4[14],xmm2[15],xmm4[15] ; AVX2-FP-NEXT: vinserti128 $1, %xmm0, %ymm0, %ymm6 ; AVX2-FP-NEXT: vmovdqa 16(%r9), %xmm1 ; AVX2-FP-NEXT: vmovdqa 16(%r8), %xmm0 ; AVX2-FP-NEXT: vpunpckhbw {{.*#+}} xmm3 = 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] ; AVX2-FP-NEXT: vinserti128 $1, %xmm3, %ymm3, %ymm8 ; AVX2-FP-NEXT: vpshufb %ymm9, %ymm6, %ymm3 ; AVX2-FP-NEXT: vpshufb %ymm7, %ymm8, %ymm5 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm7 = ymm5[0,1,2],ymm3[3],ymm5[4,5,6],ymm3[7],ymm5[8,9,10],ymm3[11],ymm5[12,13,14],ymm3[15] ; AVX2-FP-NEXT: vmovdqa 16(%rsi), %xmm5 ; AVX2-FP-NEXT: vmovdqa 16(%rdi), %xmm3 ; AVX2-FP-NEXT: vpunpckhbw {{.*#+}} xmm9 = 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: vpshufb %xmm11, %xmm9, %xmm10 ; AVX2-FP-NEXT: vpmovzxwq {{.*#+}} xmm13 = xmm9[0],zero,zero,zero,xmm9[1],zero,zero,zero ; AVX2-FP-NEXT: vinserti128 $1, %xmm10, %ymm13, %ymm10 ; AVX2-FP-NEXT: vmovdqa 16(%rcx), %xmm13 ; AVX2-FP-NEXT: vmovdqa 16(%rdx), %xmm14 ; AVX2-FP-NEXT: vpunpckhbw {{.*#+}} xmm15 = xmm14[8],xmm13[8],xmm14[9],xmm13[9],xmm14[10],xmm13[10],xmm14[11],xmm13[11],xmm14[12],xmm13[12],xmm14[13],xmm13[13],xmm14[14],xmm13[14],xmm14[15],xmm13[15] ; AVX2-FP-NEXT: vinserti128 $1, %xmm15, %ymm15, %ymm15 ; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm11 = [0,1,0,1,4,5,2,3,4,5,2,3,12,13,14,15,0,1,4,5,4,5,6,7,4,5,6,7,12,13,14,15] ; AVX2-FP-NEXT: vpshufb %ymm11, %ymm15, %ymm11 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm10 = ymm10[0],ymm11[1],ymm10[2,3,4],ymm11[5],ymm10[6,7,8],ymm11[9],ymm10[10,11,12],ymm11[13],ymm10[14,15] ; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm7 = ymm10[0],ymm7[1],ymm10[2],ymm7[3],ymm10[4],ymm7[5],ymm10[6],ymm7[7] ; AVX2-FP-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm12 = [0,1,2,3,8,9,8,9,8,9,8,9,12,13,10,11,0,1,2,3,8,9,12,13,8,9,12,13,12,13,14,15] ; AVX2-FP-NEXT: vpshufb %ymm12, %ymm6, %ymm6 ; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm7 = [0,1,2,3,8,9,10,11,8,9,10,11,10,11,14,15,0,1,2,3,12,13,10,11,12,13,10,11,14,15,14,15] ; AVX2-FP-NEXT: vpshufb %ymm7, %ymm8, %ymm8 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm6 = ymm8[0,1,2],ymm6[3],ymm8[4,5,6],ymm6[7],ymm8[8,9,10],ymm6[11],ymm8[12,13,14],ymm6[15] ; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm11 = [8,9,8,9,4,5,6,7,12,13,10,11,12,13,10,11,8,9,12,13,4,5,6,7,12,13,14,15,12,13,14,15] ; AVX2-FP-NEXT: vpshufb %ymm11, %ymm15, %ymm8 ; AVX2-FP-NEXT: vinserti128 $1, %xmm9, %ymm9, %ymm9 ; AVX2-FP-NEXT: vpmovsxwq {{.*#+}} ymm10 = [2312,2826,3340,3854] ; AVX2-FP-NEXT: vpshufb %ymm10, %ymm9, %ymm9 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm8 = ymm9[0],ymm8[1],ymm9[2,3,4],ymm8[5],ymm9[6,7,8],ymm8[9],ymm9[10,11,12],ymm8[13],ymm9[14,15] ; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm6 = ymm8[0],ymm6[1],ymm8[2],ymm6[3],ymm8[4],ymm6[5],ymm8[6],ymm6[7] ; AVX2-FP-NEXT: vmovdqu %ymm6, (%rsp) # 32-byte Spill ; AVX2-FP-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3],xmm2[4],xmm4[4],xmm2[5],xmm4[5],xmm2[6],xmm4[6],xmm2[7],xmm4[7] ; AVX2-FP-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] ; AVX2-FP-NEXT: vinserti128 $1, %xmm2, %ymm2, %ymm1 ; AVX2-FP-NEXT: vinserti128 $1, %xmm0, %ymm0, %ymm2 ; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm6 = [0,1,0,1,0,1,0,1,8,9,10,11,4,5,2,3,0,1,4,5,0,1,4,5,8,9,10,11,4,5,6,7] ; AVX2-FP-NEXT: vpshufb %ymm6, %ymm1, %ymm0 ; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm9 = [0,1,2,3,0,1,2,3,8,9,10,11,2,3,6,7,4,5,2,3,4,5,2,3,8,9,10,11,6,7,6,7] ; AVX2-FP-NEXT: vpshufb %ymm9, %ymm2, %ymm4 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm0 = ymm4[0,1,2],ymm0[3],ymm4[4,5,6],ymm0[7],ymm4[8,9,10],ymm0[11],ymm4[12,13,14],ymm0[15] ; AVX2-FP-NEXT: vpunpcklbw {{.*#+}} xmm4 = xmm14[0],xmm13[0],xmm14[1],xmm13[1],xmm14[2],xmm13[2],xmm14[3],xmm13[3],xmm14[4],xmm13[4],xmm14[5],xmm13[5],xmm14[6],xmm13[6],xmm14[7],xmm13[7] ; AVX2-FP-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm3[0],xmm5[0],xmm3[1],xmm5[1],xmm3[2],xmm5[2],xmm3[3],xmm5[3],xmm3[4],xmm5[4],xmm3[5],xmm5[5],xmm3[6],xmm5[6],xmm3[7],xmm5[7] ; AVX2-FP-NEXT: vpmovsxwq {{.*#+}} xmm14 = [1284,1798] ; AVX2-FP-NEXT: vpshufb %xmm14, %xmm3, %xmm5 ; AVX2-FP-NEXT: vpmovzxwq {{.*#+}} xmm8 = xmm3[0],zero,zero,zero,xmm3[1],zero,zero,zero ; AVX2-FP-NEXT: vinserti128 $1, %xmm5, %ymm8, %ymm5 ; AVX2-FP-NEXT: vinserti128 $1, %xmm4, %ymm4, %ymm4 ; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm13 = [0,1,0,1,4,5,2,3,4,5,2,3,12,13,14,15,0,1,4,5,4,5,6,7,4,5,6,7,12,13,14,15] ; AVX2-FP-NEXT: vpshufb %ymm13, %ymm4, %ymm8 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm5 = ymm5[0],ymm8[1],ymm5[2,3,4],ymm8[5],ymm5[6,7,8],ymm8[9],ymm5[10,11,12],ymm8[13],ymm5[14,15] ; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm5[0],ymm0[1],ymm5[2],ymm0[3],ymm5[4],ymm0[5],ymm5[6],ymm0[7] ; AVX2-FP-NEXT: vpshufb %ymm12, %ymm1, %ymm1 ; AVX2-FP-NEXT: vmovdqa %ymm7, %ymm15 ; AVX2-FP-NEXT: vpshufb %ymm7, %ymm2, %ymm2 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0,1,2],ymm1[3],ymm2[4,5,6],ymm1[7],ymm2[8,9,10],ymm1[11],ymm2[12,13,14],ymm1[15] ; AVX2-FP-NEXT: vpshufb %ymm11, %ymm4, %ymm2 ; AVX2-FP-NEXT: vmovdqa %ymm11, %ymm7 ; AVX2-FP-NEXT: vinserti128 $1, %xmm3, %ymm3, %ymm3 ; AVX2-FP-NEXT: vpshufb %ymm10, %ymm3, %ymm3 ; AVX2-FP-NEXT: vmovdqa %ymm10, %ymm11 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm2 = ymm3[0],ymm2[1],ymm3[2,3,4],ymm2[5],ymm3[6,7,8],ymm2[9],ymm3[10,11,12],ymm2[13],ymm3[14,15] ; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2],ymm1[3],ymm2[4],ymm1[5],ymm2[6],ymm1[7] ; AVX2-FP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload ; AVX2-FP-NEXT: vpunpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm2 # 16-byte Folded Reload ; AVX2-FP-NEXT: # xmm2 = xmm2[8],mem[8],xmm2[9],mem[9],xmm2[10],mem[10],xmm2[11],mem[11],xmm2[12],mem[12],xmm2[13],mem[13],xmm2[14],mem[14],xmm2[15],mem[15] ; AVX2-FP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload ; AVX2-FP-NEXT: vpunpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm3 # 16-byte Folded Reload ; AVX2-FP-NEXT: # xmm3 = xmm3[8],mem[8],xmm3[9],mem[9],xmm3[10],mem[10],xmm3[11],mem[11],xmm3[12],mem[12],xmm3[13],mem[13],xmm3[14],mem[14],xmm3[15],mem[15] ; AVX2-FP-NEXT: vinserti128 $1, %xmm2, %ymm2, %ymm2 ; AVX2-FP-NEXT: vpshufb %ymm6, %ymm2, %ymm4 ; AVX2-FP-NEXT: vinserti128 $1, %xmm3, %ymm3, %ymm3 ; AVX2-FP-NEXT: vpshufb %ymm9, %ymm3, %ymm5 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm4 = ymm5[0,1,2],ymm4[3],ymm5[4,5,6],ymm4[7],ymm5[8,9,10],ymm4[11],ymm5[12,13,14],ymm4[15] ; AVX2-FP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload ; AVX2-FP-NEXT: vpunpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm5, %xmm5 # 16-byte Folded Reload ; AVX2-FP-NEXT: # xmm5 = xmm5[8],mem[8],xmm5[9],mem[9],xmm5[10],mem[10],xmm5[11],mem[11],xmm5[12],mem[12],xmm5[13],mem[13],xmm5[14],mem[14],xmm5[15],mem[15] ; AVX2-FP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload ; AVX2-FP-NEXT: vpunpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm8, %xmm8 # 16-byte Folded Reload ; AVX2-FP-NEXT: # xmm8 = xmm8[8],mem[8],xmm8[9],mem[9],xmm8[10],mem[10],xmm8[11],mem[11],xmm8[12],mem[12],xmm8[13],mem[13],xmm8[14],mem[14],xmm8[15],mem[15] ; AVX2-FP-NEXT: vpshufb %xmm14, %xmm8, %xmm9 ; AVX2-FP-NEXT: vpmovzxwq {{.*#+}} xmm10 = xmm8[0],zero,zero,zero,xmm8[1],zero,zero,zero ; AVX2-FP-NEXT: vinserti128 $1, %xmm9, %ymm10, %ymm9 ; AVX2-FP-NEXT: vinserti128 $1, %xmm5, %ymm5, %ymm5 ; AVX2-FP-NEXT: vpshufb %ymm13, %ymm5, %ymm10 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm9 = ymm9[0],ymm10[1],ymm9[2,3,4],ymm10[5],ymm9[6,7,8],ymm10[9],ymm9[10,11,12],ymm10[13],ymm9[14,15] ; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm4 = ymm9[0],ymm4[1],ymm9[2],ymm4[3],ymm9[4],ymm4[5],ymm9[6],ymm4[7] ; AVX2-FP-NEXT: vpshufb %ymm12, %ymm2, %ymm2 ; AVX2-FP-NEXT: vpshufb %ymm15, %ymm3, %ymm3 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm2 = ymm3[0,1,2],ymm2[3],ymm3[4,5,6],ymm2[7],ymm3[8,9,10],ymm2[11],ymm3[12,13,14],ymm2[15] ; AVX2-FP-NEXT: vpshufb %ymm7, %ymm5, %ymm3 ; AVX2-FP-NEXT: vinserti128 $1, %xmm8, %ymm8, %ymm5 ; AVX2-FP-NEXT: vpshufb %ymm11, %ymm5, %ymm5 ; AVX2-FP-NEXT: vpblendw {{.*#+}} ymm3 = ymm5[0],ymm3[1],ymm5[2,3,4],ymm3[5],ymm5[6,7,8],ymm3[9],ymm5[10,11,12],ymm3[13],ymm5[14,15] ; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0],ymm2[1],ymm3[2],ymm2[3],ymm3[4],ymm2[5],ymm3[6],ymm2[7] ; AVX2-FP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX2-FP-NEXT: vmovdqa %ymm2, 96(%rax) ; AVX2-FP-NEXT: vmovdqa %ymm4, 64(%rax) ; AVX2-FP-NEXT: vmovdqa %ymm1, 160(%rax) ; AVX2-FP-NEXT: vmovdqa %ymm0, 128(%rax) ; AVX2-FP-NEXT: vmovups (%rsp), %ymm0 # 32-byte Reload ; AVX2-FP-NEXT: vmovaps %ymm0, 224(%rax) ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload ; AVX2-FP-NEXT: vmovaps %ymm0, 192(%rax) ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload ; AVX2-FP-NEXT: vmovaps %ymm0, 288(%rax) ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload ; AVX2-FP-NEXT: vmovaps %ymm0, 256(%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, 416(%rax) ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload ; AVX2-FP-NEXT: vmovaps %ymm0, 384(%rax) ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload ; AVX2-FP-NEXT: vmovaps %ymm0, 480(%rax) ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload ; AVX2-FP-NEXT: vmovaps %ymm0, 448(%rax) ; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload ; AVX2-FP-NEXT: vmovaps %ymm0, (%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 $392, %rsp # imm = 0x188 ; AVX2-FP-NEXT: vzeroupper ; AVX2-FP-NEXT: retq ; ; AVX2-FCP-LABEL: store_i8_stride8_vf64: ; AVX2-FCP: # %bb.0: ; AVX2-FCP-NEXT: subq $392, %rsp # imm = 0x188 ; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX2-FCP-NEXT: vmovdqa (%r10), %xmm1 ; AVX2-FCP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX2-FCP-NEXT: vmovdqa (%rax), %xmm0 ; AVX2-FCP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX2-FCP-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] ; AVX2-FCP-NEXT: vinserti128 $1, %xmm0, %ymm0, %ymm0 ; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm2 = ymm0[0,1,2,3,8,9,8,9,8,9,8,9,12,13,10,11,16,17,18,19,24,25,28,29,24,25,28,29,28,29,30,31] ; AVX2-FCP-NEXT: vmovdqa (%r9), %xmm3 ; AVX2-FCP-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX2-FCP-NEXT: vmovdqa (%r8), %xmm1 ; AVX2-FCP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX2-FCP-NEXT: vpunpcklbw {{.*#+}} 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] ; AVX2-FCP-NEXT: vinserti128 $1, %xmm1, %ymm1, %ymm1 ; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm3 = ymm1[0,1,2,3,8,9,10,11,8,9,10,11,10,11,14,15,16,17,18,19,28,29,26,27,28,29,26,27,30,31,30,31] ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm3[0,1,2],ymm2[3],ymm3[4,5,6],ymm2[7],ymm3[8,9,10],ymm2[11],ymm3[12,13,14],ymm2[15] ; AVX2-FCP-NEXT: vmovdqa (%rcx), %xmm4 ; AVX2-FCP-NEXT: vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX2-FCP-NEXT: vmovdqa (%rdx), %xmm2 ; AVX2-FCP-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX2-FCP-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3],xmm2[4],xmm4[4],xmm2[5],xmm4[5],xmm2[6],xmm4[6],xmm2[7],xmm4[7] ; AVX2-FCP-NEXT: vinserti128 $1, %xmm2, %ymm2, %ymm2 ; AVX2-FCP-NEXT: vmovdqa (%rsi), %xmm5 ; AVX2-FCP-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX2-FCP-NEXT: vmovdqa (%rdi), %xmm4 ; AVX2-FCP-NEXT: vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX2-FCP-NEXT: vpunpcklbw {{.*#+}} xmm4 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3],xmm4[4],xmm5[4],xmm4[5],xmm5[5],xmm4[6],xmm5[6],xmm4[7],xmm5[7] ; AVX2-FCP-NEXT: vinserti128 $1, %xmm4, %ymm4, %ymm5 ; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm5 = ymm5[8,9,u,u,u,u,u,u,10,11,u,u,u,u,u,u,28,29,u,u,u,u,u,u,30,31,u,u,u,u,u,u] ; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm6 = ymm2[8,9,8,9,4,5,6,7,12,13,10,11,12,13,10,11,24,25,28,29,20,21,22,23,28,29,30,31,28,29,30,31] ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm5 = ymm5[0],ymm6[1],ymm5[2,3,4],ymm6[5],ymm5[6,7,8],ymm6[9],ymm5[10,11,12],ymm6[13],ymm5[14,15] ; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm3 = ymm5[0],ymm3[1],ymm5[2],ymm3[3],ymm5[4],ymm3[5],ymm5[6],ymm3[7] ; AVX2-FCP-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm9 = [0,1,0,1,0,1,0,1,8,9,10,11,4,5,2,3,0,1,4,5,0,1,4,5,8,9,10,11,4,5,6,7] ; AVX2-FCP-NEXT: vpshufb %ymm9, %ymm0, %ymm0 ; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm10 = [0,1,2,3,0,1,2,3,8,9,10,11,2,3,6,7,4,5,2,3,4,5,2,3,8,9,10,11,6,7,6,7] ; AVX2-FCP-NEXT: vpshufb %ymm10, %ymm1, %ymm1 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3],ymm1[4,5,6],ymm0[7],ymm1[8,9,10],ymm0[11],ymm1[12,13,14],ymm0[15] ; AVX2-FCP-NEXT: vpmovsxwq {{.*#+}} xmm13 = [1284,1798] ; AVX2-FCP-NEXT: vpshufb %xmm13, %xmm4, %xmm1 ; AVX2-FCP-NEXT: vpmovzxwq {{.*#+}} xmm3 = xmm4[0],zero,zero,zero,xmm4[1],zero,zero,zero ; AVX2-FCP-NEXT: vinserti128 $1, %xmm1, %ymm3, %ymm1 ; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm12 = [0,1,0,1,4,5,2,3,4,5,2,3,12,13,14,15,0,1,4,5,4,5,6,7,4,5,6,7,12,13,14,15] ; AVX2-FCP-NEXT: vpshufb %ymm12, %ymm2, %ymm2 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[2,3,4],ymm2[5],ymm1[6,7,8],ymm2[9],ymm1[10,11,12],ymm2[13],ymm1[14,15] ; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[2],ymm0[3],ymm1[4],ymm0[5],ymm1[6],ymm0[7] ; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FCP-NEXT: vmovdqa 48(%r10), %xmm8 ; AVX2-FCP-NEXT: vmovdqa 48(%rax), %xmm3 ; AVX2-FCP-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm3[8],xmm8[8],xmm3[9],xmm8[9],xmm3[10],xmm8[10],xmm3[11],xmm8[11],xmm3[12],xmm8[12],xmm3[13],xmm8[13],xmm3[14],xmm8[14],xmm3[15],xmm8[15] ; AVX2-FCP-NEXT: vinserti128 $1, %xmm0, %ymm0, %ymm6 ; AVX2-FCP-NEXT: vmovdqa 48(%r9), %xmm4 ; AVX2-FCP-NEXT: vmovdqa 48(%r8), %xmm5 ; AVX2-FCP-NEXT: vpunpckhbw {{.*#+}} xmm0 = 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] ; AVX2-FCP-NEXT: vinserti128 $1, %xmm0, %ymm0, %ymm7 ; AVX2-FCP-NEXT: vpshufb %ymm9, %ymm6, %ymm0 ; AVX2-FCP-NEXT: vpshufb %ymm10, %ymm7, %ymm2 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm11 = ymm2[0,1,2],ymm0[3],ymm2[4,5,6],ymm0[7],ymm2[8,9,10],ymm0[11],ymm2[12,13,14],ymm0[15] ; AVX2-FCP-NEXT: vmovdqa 48(%rsi), %xmm10 ; AVX2-FCP-NEXT: vmovdqa 48(%rdi), %xmm9 ; AVX2-FCP-NEXT: vpunpckhbw {{.*#+}} xmm0 = 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-FCP-NEXT: vpshufb %xmm13, %xmm0, %xmm1 ; AVX2-FCP-NEXT: vpmovzxwq {{.*#+}} xmm13 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero ; AVX2-FCP-NEXT: vinserti128 $1, %xmm1, %ymm13, %ymm1 ; AVX2-FCP-NEXT: vmovdqa 48(%rcx), %xmm13 ; AVX2-FCP-NEXT: vmovdqa 48(%rdx), %xmm2 ; AVX2-FCP-NEXT: vpunpckhbw {{.*#+}} xmm15 = xmm2[8],xmm13[8],xmm2[9],xmm13[9],xmm2[10],xmm13[10],xmm2[11],xmm13[11],xmm2[12],xmm13[12],xmm2[13],xmm13[13],xmm2[14],xmm13[14],xmm2[15],xmm13[15] ; AVX2-FCP-NEXT: vinserti128 $1, %xmm15, %ymm15, %ymm15 ; AVX2-FCP-NEXT: vpshufb %ymm12, %ymm15, %ymm14 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm1[0],ymm14[1],ymm1[2,3,4],ymm14[5],ymm1[6,7,8],ymm14[9],ymm1[10,11,12],ymm14[13],ymm1[14,15] ; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0],ymm11[1],ymm1[2],ymm11[3],ymm1[4],ymm11[5],ymm1[6],ymm11[7] ; AVX2-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm12 = [0,1,2,3,8,9,8,9,8,9,8,9,12,13,10,11,0,1,2,3,8,9,12,13,8,9,12,13,12,13,14,15] ; AVX2-FCP-NEXT: vpshufb %ymm12, %ymm6, %ymm1 ; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm11 = [0,1,2,3,8,9,10,11,8,9,10,11,10,11,14,15,0,1,2,3,12,13,10,11,12,13,10,11,14,15,14,15] ; AVX2-FCP-NEXT: vpshufb %ymm11, %ymm7, %ymm6 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm6[0,1,2],ymm1[3],ymm6[4,5,6],ymm1[7],ymm6[8,9,10],ymm1[11],ymm6[12,13,14],ymm1[15] ; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm7 = [8,9,8,9,4,5,6,7,12,13,10,11,12,13,10,11,8,9,12,13,4,5,6,7,12,13,14,15,12,13,14,15] ; AVX2-FCP-NEXT: vpshufb %ymm7, %ymm15, %ymm6 ; AVX2-FCP-NEXT: vinserti128 $1, %xmm0, %ymm0, %ymm0 ; AVX2-FCP-NEXT: vpmovsxwq {{.*#+}} ymm14 = [2312,2826,3340,3854] ; AVX2-FCP-NEXT: vpshufb %ymm14, %ymm0, %ymm0 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm6[1],ymm0[2,3,4],ymm6[5],ymm0[6,7,8],ymm6[9],ymm0[10,11,12],ymm6[13],ymm0[14,15] ; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2],ymm1[3],ymm0[4],ymm1[5],ymm0[6],ymm1[7] ; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FCP-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm3[0],xmm8[0],xmm3[1],xmm8[1],xmm3[2],xmm8[2],xmm3[3],xmm8[3],xmm3[4],xmm8[4],xmm3[5],xmm8[5],xmm3[6],xmm8[6],xmm3[7],xmm8[7] ; AVX2-FCP-NEXT: vpunpcklbw {{.*#+}} xmm1 = 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: vinserti128 $1, %xmm0, %ymm0, %ymm0 ; AVX2-FCP-NEXT: vinserti128 $1, %xmm1, %ymm1, %ymm1 ; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm15 = [0,1,0,1,0,1,0,1,8,9,10,11,4,5,2,3,0,1,4,5,0,1,4,5,8,9,10,11,4,5,6,7] ; AVX2-FCP-NEXT: vpshufb %ymm15, %ymm0, %ymm3 ; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm8 = [0,1,2,3,0,1,2,3,8,9,10,11,2,3,6,7,4,5,2,3,4,5,2,3,8,9,10,11,6,7,6,7] ; AVX2-FCP-NEXT: vpshufb %ymm8, %ymm1, %ymm4 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm4[0,1,2],ymm3[3],ymm4[4,5,6],ymm3[7],ymm4[8,9,10],ymm3[11],ymm4[12,13,14],ymm3[15] ; AVX2-FCP-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm2[0],xmm13[0],xmm2[1],xmm13[1],xmm2[2],xmm13[2],xmm2[3],xmm13[3],xmm2[4],xmm13[4],xmm2[5],xmm13[5],xmm2[6],xmm13[6],xmm2[7],xmm13[7] ; AVX2-FCP-NEXT: vpunpcklbw {{.*#+}} xmm4 = xmm9[0],xmm10[0],xmm9[1],xmm10[1],xmm9[2],xmm10[2],xmm9[3],xmm10[3],xmm9[4],xmm10[4],xmm9[5],xmm10[5],xmm9[6],xmm10[6],xmm9[7],xmm10[7] ; AVX2-FCP-NEXT: vpmovsxwq {{.*#+}} xmm10 = [1284,1798] ; AVX2-FCP-NEXT: vpshufb %xmm10, %xmm4, %xmm5 ; AVX2-FCP-NEXT: vpmovzxwq {{.*#+}} xmm6 = xmm4[0],zero,zero,zero,xmm4[1],zero,zero,zero ; AVX2-FCP-NEXT: vinserti128 $1, %xmm5, %ymm6, %ymm5 ; AVX2-FCP-NEXT: vinserti128 $1, %xmm2, %ymm2, %ymm2 ; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm6 = ymm2[0,1,0,1,4,5,2,3,4,5,2,3,12,13,14,15,16,17,20,21,20,21,22,23,20,21,22,23,28,29,30,31] ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm5 = ymm5[0],ymm6[1],ymm5[2,3,4],ymm6[5],ymm5[6,7,8],ymm6[9],ymm5[10,11,12],ymm6[13],ymm5[14,15] ; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm3 = ymm5[0],ymm3[1],ymm5[2],ymm3[3],ymm5[4],ymm3[5],ymm5[6],ymm3[7] ; AVX2-FCP-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FCP-NEXT: vpshufb %ymm12, %ymm0, %ymm0 ; AVX2-FCP-NEXT: vpshufb %ymm11, %ymm1, %ymm1 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3],ymm1[4,5,6],ymm0[7],ymm1[8,9,10],ymm0[11],ymm1[12,13,14],ymm0[15] ; AVX2-FCP-NEXT: vpshufb %ymm7, %ymm2, %ymm1 ; AVX2-FCP-NEXT: vinserti128 $1, %xmm4, %ymm4, %ymm2 ; AVX2-FCP-NEXT: vpshufb %ymm14, %ymm2, %ymm2 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2,3,4],ymm1[5],ymm2[6,7,8],ymm1[9],ymm2[10,11,12],ymm1[13],ymm2[14,15] ; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[2],ymm0[3],ymm1[4],ymm0[5],ymm1[6],ymm0[7] ; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FCP-NEXT: vmovdqa 32(%r10), %xmm1 ; AVX2-FCP-NEXT: vmovdqa 32(%rax), %xmm3 ; AVX2-FCP-NEXT: vpunpckhbw {{.*#+}} xmm0 = 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] ; AVX2-FCP-NEXT: vinserti128 $1, %xmm0, %ymm0, %ymm6 ; AVX2-FCP-NEXT: vmovdqa 32(%r9), %xmm4 ; AVX2-FCP-NEXT: vmovdqa 32(%r8), %xmm5 ; AVX2-FCP-NEXT: vpunpckhbw {{.*#+}} xmm0 = 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] ; AVX2-FCP-NEXT: vinserti128 $1, %xmm0, %ymm0, %ymm7 ; AVX2-FCP-NEXT: vpshufb %ymm15, %ymm6, %ymm0 ; AVX2-FCP-NEXT: vpshufb %ymm8, %ymm7, %ymm2 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm8 = ymm2[0,1,2],ymm0[3],ymm2[4,5,6],ymm0[7],ymm2[8,9,10],ymm0[11],ymm2[12,13,14],ymm0[15] ; AVX2-FCP-NEXT: vmovdqa 32(%rsi), %xmm0 ; AVX2-FCP-NEXT: vmovdqa 32(%rdi), %xmm2 ; AVX2-FCP-NEXT: vpunpckhbw {{.*#+}} xmm9 = xmm2[8],xmm0[8],xmm2[9],xmm0[9],xmm2[10],xmm0[10],xmm2[11],xmm0[11],xmm2[12],xmm0[12],xmm2[13],xmm0[13],xmm2[14],xmm0[14],xmm2[15],xmm0[15] ; AVX2-FCP-NEXT: vpshufb %xmm10, %xmm9, %xmm10 ; AVX2-FCP-NEXT: vpmovzxwq {{.*#+}} xmm11 = xmm9[0],zero,zero,zero,xmm9[1],zero,zero,zero ; AVX2-FCP-NEXT: vinserti128 $1, %xmm10, %ymm11, %ymm10 ; AVX2-FCP-NEXT: vmovdqa 32(%rcx), %xmm11 ; AVX2-FCP-NEXT: vmovdqa 32(%rdx), %xmm13 ; AVX2-FCP-NEXT: vpunpckhbw {{.*#+}} xmm14 = 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-FCP-NEXT: vinserti128 $1, %xmm14, %ymm14, %ymm14 ; AVX2-FCP-NEXT: vpshufb {{.*#+}} ymm15 = ymm14[0,1,0,1,4,5,2,3,4,5,2,3,12,13,14,15,16,17,20,21,20,21,22,23,20,21,22,23,28,29,30,31] ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm10 = ymm10[0],ymm15[1],ymm10[2,3,4],ymm15[5],ymm10[6,7,8],ymm15[9],ymm10[10,11,12],ymm15[13],ymm10[14,15] ; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm8 = ymm10[0],ymm8[1],ymm10[2],ymm8[3],ymm10[4],ymm8[5],ymm10[6],ymm8[7] ; AVX2-FCP-NEXT: vmovdqu %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FCP-NEXT: vmovdqa %ymm12, %ymm10 ; AVX2-FCP-NEXT: vpshufb %ymm12, %ymm6, %ymm6 ; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm15 = [0,1,2,3,8,9,10,11,8,9,10,11,10,11,14,15,0,1,2,3,12,13,10,11,12,13,10,11,14,15,14,15] ; AVX2-FCP-NEXT: vpshufb %ymm15, %ymm7, %ymm7 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm6 = ymm7[0,1,2],ymm6[3],ymm7[4,5,6],ymm6[7],ymm7[8,9,10],ymm6[11],ymm7[12,13,14],ymm6[15] ; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm12 = [8,9,8,9,4,5,6,7,12,13,10,11,12,13,10,11,8,9,12,13,4,5,6,7,12,13,14,15,12,13,14,15] ; AVX2-FCP-NEXT: vpshufb %ymm12, %ymm14, %ymm7 ; AVX2-FCP-NEXT: vinserti128 $1, %xmm9, %ymm9, %ymm8 ; AVX2-FCP-NEXT: vpmovsxwq {{.*#+}} ymm14 = [2312,2826,3340,3854] ; AVX2-FCP-NEXT: vpshufb %ymm14, %ymm8, %ymm8 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm7 = ymm8[0],ymm7[1],ymm8[2,3,4],ymm7[5],ymm8[6,7,8],ymm7[9],ymm8[10,11,12],ymm7[13],ymm8[14,15] ; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm7[0],ymm6[1],ymm7[2],ymm6[3],ymm7[4],ymm6[5],ymm7[6],ymm6[7] ; AVX2-FCP-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FCP-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3],xmm3[4],xmm1[4],xmm3[5],xmm1[5],xmm3[6],xmm1[6],xmm3[7],xmm1[7] ; AVX2-FCP-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] ; AVX2-FCP-NEXT: vinserti128 $1, %xmm1, %ymm1, %ymm1 ; AVX2-FCP-NEXT: vinserti128 $1, %xmm3, %ymm3, %ymm3 ; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm9 = [0,1,0,1,0,1,0,1,8,9,10,11,4,5,2,3,0,1,4,5,0,1,4,5,8,9,10,11,4,5,6,7] ; AVX2-FCP-NEXT: vpshufb %ymm9, %ymm1, %ymm4 ; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm7 = [0,1,2,3,0,1,2,3,8,9,10,11,2,3,6,7,4,5,2,3,4,5,2,3,8,9,10,11,6,7,6,7] ; AVX2-FCP-NEXT: vpshufb %ymm7, %ymm3, %ymm5 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm4 = ymm5[0,1,2],ymm4[3],ymm5[4,5,6],ymm4[7],ymm5[8,9,10],ymm4[11],ymm5[12,13,14],ymm4[15] ; AVX2-FCP-NEXT: vpunpcklbw {{.*#+}} xmm5 = 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-FCP-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm2[0],xmm0[0],xmm2[1],xmm0[1],xmm2[2],xmm0[2],xmm2[3],xmm0[3],xmm2[4],xmm0[4],xmm2[5],xmm0[5],xmm2[6],xmm0[6],xmm2[7],xmm0[7] ; AVX2-FCP-NEXT: vpmovsxwq {{.*#+}} xmm11 = [1284,1798] ; AVX2-FCP-NEXT: vpshufb %xmm11, %xmm0, %xmm2 ; AVX2-FCP-NEXT: vpmovzxwq {{.*#+}} xmm6 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero ; AVX2-FCP-NEXT: vinserti128 $1, %xmm2, %ymm6, %ymm2 ; AVX2-FCP-NEXT: vinserti128 $1, %xmm5, %ymm5, %ymm5 ; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm6 = [0,1,0,1,4,5,2,3,4,5,2,3,12,13,14,15,0,1,4,5,4,5,6,7,4,5,6,7,12,13,14,15] ; AVX2-FCP-NEXT: vpshufb %ymm6, %ymm5, %ymm6 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0],ymm6[1],ymm2[2,3,4],ymm6[5],ymm2[6,7,8],ymm6[9],ymm2[10,11,12],ymm6[13],ymm2[14,15] ; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0],ymm4[1],ymm2[2],ymm4[3],ymm2[4],ymm4[5],ymm2[6],ymm4[7] ; AVX2-FCP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FCP-NEXT: vpshufb %ymm10, %ymm1, %ymm1 ; AVX2-FCP-NEXT: vpshufb %ymm15, %ymm3, %ymm2 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0,1,2],ymm1[3],ymm2[4,5,6],ymm1[7],ymm2[8,9,10],ymm1[11],ymm2[12,13,14],ymm1[15] ; AVX2-FCP-NEXT: vpshufb %ymm12, %ymm5, %ymm2 ; AVX2-FCP-NEXT: vinserti128 $1, %xmm0, %ymm0, %ymm0 ; AVX2-FCP-NEXT: vpshufb %ymm14, %ymm0, %ymm0 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm2[1],ymm0[2,3,4],ymm2[5],ymm0[6,7,8],ymm2[9],ymm0[10,11,12],ymm2[13],ymm0[14,15] ; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2],ymm1[3],ymm0[4],ymm1[5],ymm0[6],ymm1[7] ; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FCP-NEXT: vmovdqa 16(%r10), %xmm4 ; AVX2-FCP-NEXT: vmovdqa 16(%rax), %xmm2 ; AVX2-FCP-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm2[8],xmm4[8],xmm2[9],xmm4[9],xmm2[10],xmm4[10],xmm2[11],xmm4[11],xmm2[12],xmm4[12],xmm2[13],xmm4[13],xmm2[14],xmm4[14],xmm2[15],xmm4[15] ; AVX2-FCP-NEXT: vinserti128 $1, %xmm0, %ymm0, %ymm6 ; AVX2-FCP-NEXT: vmovdqa 16(%r9), %xmm1 ; AVX2-FCP-NEXT: vmovdqa 16(%r8), %xmm0 ; AVX2-FCP-NEXT: vpunpckhbw {{.*#+}} xmm3 = 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] ; AVX2-FCP-NEXT: vinserti128 $1, %xmm3, %ymm3, %ymm8 ; AVX2-FCP-NEXT: vpshufb %ymm9, %ymm6, %ymm3 ; AVX2-FCP-NEXT: vpshufb %ymm7, %ymm8, %ymm5 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm7 = ymm5[0,1,2],ymm3[3],ymm5[4,5,6],ymm3[7],ymm5[8,9,10],ymm3[11],ymm5[12,13,14],ymm3[15] ; AVX2-FCP-NEXT: vmovdqa 16(%rsi), %xmm5 ; AVX2-FCP-NEXT: vmovdqa 16(%rdi), %xmm3 ; AVX2-FCP-NEXT: vpunpckhbw {{.*#+}} xmm9 = 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: vpshufb %xmm11, %xmm9, %xmm10 ; AVX2-FCP-NEXT: vpmovzxwq {{.*#+}} xmm13 = xmm9[0],zero,zero,zero,xmm9[1],zero,zero,zero ; AVX2-FCP-NEXT: vinserti128 $1, %xmm10, %ymm13, %ymm10 ; AVX2-FCP-NEXT: vmovdqa 16(%rcx), %xmm13 ; AVX2-FCP-NEXT: vmovdqa 16(%rdx), %xmm14 ; AVX2-FCP-NEXT: vpunpckhbw {{.*#+}} xmm15 = xmm14[8],xmm13[8],xmm14[9],xmm13[9],xmm14[10],xmm13[10],xmm14[11],xmm13[11],xmm14[12],xmm13[12],xmm14[13],xmm13[13],xmm14[14],xmm13[14],xmm14[15],xmm13[15] ; AVX2-FCP-NEXT: vinserti128 $1, %xmm15, %ymm15, %ymm15 ; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm11 = [0,1,0,1,4,5,2,3,4,5,2,3,12,13,14,15,0,1,4,5,4,5,6,7,4,5,6,7,12,13,14,15] ; AVX2-FCP-NEXT: vpshufb %ymm11, %ymm15, %ymm11 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm10 = ymm10[0],ymm11[1],ymm10[2,3,4],ymm11[5],ymm10[6,7,8],ymm11[9],ymm10[10,11,12],ymm11[13],ymm10[14,15] ; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm7 = ymm10[0],ymm7[1],ymm10[2],ymm7[3],ymm10[4],ymm7[5],ymm10[6],ymm7[7] ; AVX2-FCP-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm12 = [0,1,2,3,8,9,8,9,8,9,8,9,12,13,10,11,0,1,2,3,8,9,12,13,8,9,12,13,12,13,14,15] ; AVX2-FCP-NEXT: vpshufb %ymm12, %ymm6, %ymm6 ; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm7 = [0,1,2,3,8,9,10,11,8,9,10,11,10,11,14,15,0,1,2,3,12,13,10,11,12,13,10,11,14,15,14,15] ; AVX2-FCP-NEXT: vpshufb %ymm7, %ymm8, %ymm8 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm6 = ymm8[0,1,2],ymm6[3],ymm8[4,5,6],ymm6[7],ymm8[8,9,10],ymm6[11],ymm8[12,13,14],ymm6[15] ; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm11 = [8,9,8,9,4,5,6,7,12,13,10,11,12,13,10,11,8,9,12,13,4,5,6,7,12,13,14,15,12,13,14,15] ; AVX2-FCP-NEXT: vpshufb %ymm11, %ymm15, %ymm8 ; AVX2-FCP-NEXT: vinserti128 $1, %xmm9, %ymm9, %ymm9 ; AVX2-FCP-NEXT: vpmovsxwq {{.*#+}} ymm10 = [2312,2826,3340,3854] ; AVX2-FCP-NEXT: vpshufb %ymm10, %ymm9, %ymm9 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm8 = ymm9[0],ymm8[1],ymm9[2,3,4],ymm8[5],ymm9[6,7,8],ymm8[9],ymm9[10,11,12],ymm8[13],ymm9[14,15] ; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm8[0],ymm6[1],ymm8[2],ymm6[3],ymm8[4],ymm6[5],ymm8[6],ymm6[7] ; AVX2-FCP-NEXT: vmovdqu %ymm6, (%rsp) # 32-byte Spill ; AVX2-FCP-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3],xmm2[4],xmm4[4],xmm2[5],xmm4[5],xmm2[6],xmm4[6],xmm2[7],xmm4[7] ; AVX2-FCP-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] ; AVX2-FCP-NEXT: vinserti128 $1, %xmm2, %ymm2, %ymm1 ; AVX2-FCP-NEXT: vinserti128 $1, %xmm0, %ymm0, %ymm2 ; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm6 = [0,1,0,1,0,1,0,1,8,9,10,11,4,5,2,3,0,1,4,5,0,1,4,5,8,9,10,11,4,5,6,7] ; AVX2-FCP-NEXT: vpshufb %ymm6, %ymm1, %ymm0 ; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm9 = [0,1,2,3,0,1,2,3,8,9,10,11,2,3,6,7,4,5,2,3,4,5,2,3,8,9,10,11,6,7,6,7] ; AVX2-FCP-NEXT: vpshufb %ymm9, %ymm2, %ymm4 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm0 = ymm4[0,1,2],ymm0[3],ymm4[4,5,6],ymm0[7],ymm4[8,9,10],ymm0[11],ymm4[12,13,14],ymm0[15] ; AVX2-FCP-NEXT: vpunpcklbw {{.*#+}} xmm4 = xmm14[0],xmm13[0],xmm14[1],xmm13[1],xmm14[2],xmm13[2],xmm14[3],xmm13[3],xmm14[4],xmm13[4],xmm14[5],xmm13[5],xmm14[6],xmm13[6],xmm14[7],xmm13[7] ; AVX2-FCP-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm3[0],xmm5[0],xmm3[1],xmm5[1],xmm3[2],xmm5[2],xmm3[3],xmm5[3],xmm3[4],xmm5[4],xmm3[5],xmm5[5],xmm3[6],xmm5[6],xmm3[7],xmm5[7] ; AVX2-FCP-NEXT: vpmovsxwq {{.*#+}} xmm14 = [1284,1798] ; AVX2-FCP-NEXT: vpshufb %xmm14, %xmm3, %xmm5 ; AVX2-FCP-NEXT: vpmovzxwq {{.*#+}} xmm8 = xmm3[0],zero,zero,zero,xmm3[1],zero,zero,zero ; AVX2-FCP-NEXT: vinserti128 $1, %xmm5, %ymm8, %ymm5 ; AVX2-FCP-NEXT: vinserti128 $1, %xmm4, %ymm4, %ymm4 ; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm13 = [0,1,0,1,4,5,2,3,4,5,2,3,12,13,14,15,0,1,4,5,4,5,6,7,4,5,6,7,12,13,14,15] ; AVX2-FCP-NEXT: vpshufb %ymm13, %ymm4, %ymm8 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm5 = ymm5[0],ymm8[1],ymm5[2,3,4],ymm8[5],ymm5[6,7,8],ymm8[9],ymm5[10,11,12],ymm8[13],ymm5[14,15] ; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm5[0],ymm0[1],ymm5[2],ymm0[3],ymm5[4],ymm0[5],ymm5[6],ymm0[7] ; AVX2-FCP-NEXT: vpshufb %ymm12, %ymm1, %ymm1 ; AVX2-FCP-NEXT: vmovdqa %ymm7, %ymm15 ; AVX2-FCP-NEXT: vpshufb %ymm7, %ymm2, %ymm2 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm1 = ymm2[0,1,2],ymm1[3],ymm2[4,5,6],ymm1[7],ymm2[8,9,10],ymm1[11],ymm2[12,13,14],ymm1[15] ; AVX2-FCP-NEXT: vpshufb %ymm11, %ymm4, %ymm2 ; AVX2-FCP-NEXT: vmovdqa %ymm11, %ymm7 ; AVX2-FCP-NEXT: vinserti128 $1, %xmm3, %ymm3, %ymm3 ; AVX2-FCP-NEXT: vpshufb %ymm10, %ymm3, %ymm3 ; AVX2-FCP-NEXT: vmovdqa %ymm10, %ymm11 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm3[0],ymm2[1],ymm3[2,3,4],ymm2[5],ymm3[6,7,8],ymm2[9],ymm3[10,11,12],ymm2[13],ymm3[14,15] ; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm1 = ymm2[0],ymm1[1],ymm2[2],ymm1[3],ymm2[4],ymm1[5],ymm2[6],ymm1[7] ; AVX2-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload ; AVX2-FCP-NEXT: vpunpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm2 # 16-byte Folded Reload ; AVX2-FCP-NEXT: # xmm2 = xmm2[8],mem[8],xmm2[9],mem[9],xmm2[10],mem[10],xmm2[11],mem[11],xmm2[12],mem[12],xmm2[13],mem[13],xmm2[14],mem[14],xmm2[15],mem[15] ; AVX2-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload ; AVX2-FCP-NEXT: vpunpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm3 # 16-byte Folded Reload ; AVX2-FCP-NEXT: # xmm3 = xmm3[8],mem[8],xmm3[9],mem[9],xmm3[10],mem[10],xmm3[11],mem[11],xmm3[12],mem[12],xmm3[13],mem[13],xmm3[14],mem[14],xmm3[15],mem[15] ; AVX2-FCP-NEXT: vinserti128 $1, %xmm2, %ymm2, %ymm2 ; AVX2-FCP-NEXT: vpshufb %ymm6, %ymm2, %ymm4 ; AVX2-FCP-NEXT: vinserti128 $1, %xmm3, %ymm3, %ymm3 ; AVX2-FCP-NEXT: vpshufb %ymm9, %ymm3, %ymm5 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm4 = ymm5[0,1,2],ymm4[3],ymm5[4,5,6],ymm4[7],ymm5[8,9,10],ymm4[11],ymm5[12,13,14],ymm4[15] ; AVX2-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload ; AVX2-FCP-NEXT: vpunpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm5, %xmm5 # 16-byte Folded Reload ; AVX2-FCP-NEXT: # xmm5 = xmm5[8],mem[8],xmm5[9],mem[9],xmm5[10],mem[10],xmm5[11],mem[11],xmm5[12],mem[12],xmm5[13],mem[13],xmm5[14],mem[14],xmm5[15],mem[15] ; AVX2-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload ; AVX2-FCP-NEXT: vpunpckhbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm8, %xmm8 # 16-byte Folded Reload ; AVX2-FCP-NEXT: # xmm8 = xmm8[8],mem[8],xmm8[9],mem[9],xmm8[10],mem[10],xmm8[11],mem[11],xmm8[12],mem[12],xmm8[13],mem[13],xmm8[14],mem[14],xmm8[15],mem[15] ; AVX2-FCP-NEXT: vpshufb %xmm14, %xmm8, %xmm9 ; AVX2-FCP-NEXT: vpmovzxwq {{.*#+}} xmm10 = xmm8[0],zero,zero,zero,xmm8[1],zero,zero,zero ; AVX2-FCP-NEXT: vinserti128 $1, %xmm9, %ymm10, %ymm9 ; AVX2-FCP-NEXT: vinserti128 $1, %xmm5, %ymm5, %ymm5 ; AVX2-FCP-NEXT: vpshufb %ymm13, %ymm5, %ymm10 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm9 = ymm9[0],ymm10[1],ymm9[2,3,4],ymm10[5],ymm9[6,7,8],ymm10[9],ymm9[10,11,12],ymm10[13],ymm9[14,15] ; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm9[0],ymm4[1],ymm9[2],ymm4[3],ymm9[4],ymm4[5],ymm9[6],ymm4[7] ; AVX2-FCP-NEXT: vpshufb %ymm12, %ymm2, %ymm2 ; AVX2-FCP-NEXT: vpshufb %ymm15, %ymm3, %ymm3 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm2 = ymm3[0,1,2],ymm2[3],ymm3[4,5,6],ymm2[7],ymm3[8,9,10],ymm2[11],ymm3[12,13,14],ymm2[15] ; AVX2-FCP-NEXT: vpshufb %ymm7, %ymm5, %ymm3 ; AVX2-FCP-NEXT: vinserti128 $1, %xmm8, %ymm8, %ymm5 ; AVX2-FCP-NEXT: vpshufb %ymm11, %ymm5, %ymm5 ; AVX2-FCP-NEXT: vpblendw {{.*#+}} ymm3 = ymm5[0],ymm3[1],ymm5[2,3,4],ymm3[5],ymm5[6,7,8],ymm3[9],ymm5[10,11,12],ymm3[13],ymm5[14,15] ; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0],ymm2[1],ymm3[2],ymm2[3],ymm3[4],ymm2[5],ymm3[6],ymm2[7] ; AVX2-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX2-FCP-NEXT: vmovdqa %ymm2, 96(%rax) ; AVX2-FCP-NEXT: vmovdqa %ymm4, 64(%rax) ; AVX2-FCP-NEXT: vmovdqa %ymm1, 160(%rax) ; AVX2-FCP-NEXT: vmovdqa %ymm0, 128(%rax) ; AVX2-FCP-NEXT: vmovups (%rsp), %ymm0 # 32-byte Reload ; AVX2-FCP-NEXT: vmovaps %ymm0, 224(%rax) ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload ; AVX2-FCP-NEXT: vmovaps %ymm0, 192(%rax) ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload ; AVX2-FCP-NEXT: vmovaps %ymm0, 288(%rax) ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload ; AVX2-FCP-NEXT: vmovaps %ymm0, 256(%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, 416(%rax) ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload ; AVX2-FCP-NEXT: vmovaps %ymm0, 384(%rax) ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload ; AVX2-FCP-NEXT: vmovaps %ymm0, 480(%rax) ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload ; AVX2-FCP-NEXT: vmovaps %ymm0, 448(%rax) ; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload ; AVX2-FCP-NEXT: vmovaps %ymm0, (%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 $392, %rsp # imm = 0x188 ; AVX2-FCP-NEXT: vzeroupper ; AVX2-FCP-NEXT: retq ; ; AVX512-LABEL: store_i8_stride8_vf64: ; AVX512: # %bb.0: ; AVX512-NEXT: subq $552, %rsp # imm = 0x228 ; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX512-NEXT: vmovdqa (%rcx), %xmm3 ; AVX512-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX512-NEXT: vmovdqa 32(%rcx), %xmm12 ; AVX512-NEXT: vmovdqa 48(%rcx), %xmm0 ; AVX512-NEXT: vmovdqa (%rdx), %xmm2 ; AVX512-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX512-NEXT: vmovdqa 48(%rdx), %xmm1 ; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm2 = 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: vpshuflw {{.*#+}} xmm3 = xmm2[0,0,2,1,4,5,6,7] ; AVX512-NEXT: vpshuflw {{.*#+}} xmm4 = xmm2[0,2,2,3,4,5,6,7] ; AVX512-NEXT: vinserti128 $1, %xmm4, %ymm3, %ymm10 ; AVX512-NEXT: vpshufhw {{.*#+}} xmm3 = xmm2[0,1,2,3,4,4,6,5] ; AVX512-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,6,6,7] ; AVX512-NEXT: vinserti128 $1, %xmm2, %ymm3, %ymm2 ; AVX512-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX512-NEXT: vmovdqa (%r10), %xmm5 ; AVX512-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX512-NEXT: vmovdqa 48(%r10), %xmm3 ; AVX512-NEXT: vmovdqa (%rax), %xmm2 ; AVX512-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX512-NEXT: vmovdqa 48(%rax), %xmm4 ; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm2[8],xmm5[8],xmm2[9],xmm5[9],xmm2[10],xmm5[10],xmm2[11],xmm5[11],xmm2[12],xmm5[12],xmm2[13],xmm5[13],xmm2[14],xmm5[14],xmm2[15],xmm5[15] ; AVX512-NEXT: vpshuflw {{.*#+}} xmm5 = xmm2[0,0,2,1,4,5,6,7] ; AVX512-NEXT: vpshuflw {{.*#+}} xmm6 = xmm2[0,2,2,3,4,5,6,7] ; AVX512-NEXT: vinserti128 $1, %xmm6, %ymm5, %ymm5 ; AVX512-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX512-NEXT: vpshufhw {{.*#+}} xmm5 = xmm2[0,1,2,3,4,4,6,5] ; AVX512-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,6,6,7] ; AVX512-NEXT: vinserti128 $1, %xmm2, %ymm5, %ymm2 ; AVX512-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX512-NEXT: vmovdqa (%r9), %xmm5 ; AVX512-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX512-NEXT: vmovdqa 48(%r9), %xmm6 ; AVX512-NEXT: vmovdqa (%r8), %xmm2 ; AVX512-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX512-NEXT: vmovdqa 48(%r8), %xmm8 ; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm2[8],xmm5[8],xmm2[9],xmm5[9],xmm2[10],xmm5[10],xmm2[11],xmm5[11],xmm2[12],xmm5[12],xmm2[13],xmm5[13],xmm2[14],xmm5[14],xmm2[15],xmm5[15] ; AVX512-NEXT: vpshuflw {{.*#+}} xmm5 = xmm2[0,1,1,3,4,5,6,7] ; AVX512-NEXT: vpshuflw {{.*#+}} xmm7 = xmm2[2,1,3,3,4,5,6,7] ; AVX512-NEXT: vinserti128 $1, %xmm7, %ymm5, %ymm5 ; AVX512-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX512-NEXT: vpshufhw {{.*#+}} xmm5 = xmm2[0,1,2,3,4,5,5,7] ; AVX512-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,6,5,7,7] ; AVX512-NEXT: vinserti128 $1, %xmm2, %ymm5, %ymm2 ; AVX512-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3],xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7] ; AVX512-NEXT: vpshuflw {{.*#+}} xmm5 = xmm2[0,0,2,1,4,5,6,7] ; AVX512-NEXT: vpshuflw {{.*#+}} xmm9 = xmm2[0,2,2,3,4,5,6,7] ; AVX512-NEXT: vinserti128 $1, %xmm9, %ymm5, %ymm5 ; AVX512-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX512-NEXT: vpshufhw {{.*#+}} xmm5 = xmm2[0,1,2,3,4,4,6,5] ; AVX512-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,6,6,7] ; AVX512-NEXT: vinserti128 $1, %xmm2, %ymm5, %ymm2 ; AVX512-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3],xmm4[4],xmm3[4],xmm4[5],xmm3[5],xmm4[6],xmm3[6],xmm4[7],xmm3[7] ; AVX512-NEXT: vpshuflw {{.*#+}} xmm5 = xmm2[0,0,2,1,4,5,6,7] ; AVX512-NEXT: vpshuflw {{.*#+}} xmm9 = xmm2[0,2,2,3,4,5,6,7] ; AVX512-NEXT: vinserti128 $1, %xmm9, %ymm5, %ymm5 ; AVX512-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX512-NEXT: vpshufhw {{.*#+}} xmm5 = xmm2[0,1,2,3,4,4,6,5] ; AVX512-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,6,6,7] ; AVX512-NEXT: vinserti128 $1, %xmm2, %ymm5, %ymm2 ; AVX512-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm2 = 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] ; AVX512-NEXT: vpshuflw {{.*#+}} xmm5 = xmm2[0,1,1,3,4,5,6,7] ; AVX512-NEXT: vpshuflw {{.*#+}} xmm9 = xmm2[2,1,3,3,4,5,6,7] ; AVX512-NEXT: vinserti128 $1, %xmm9, %ymm5, %ymm5 ; AVX512-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX512-NEXT: vmovdqa 32(%rdx), %xmm5 ; AVX512-NEXT: vpshufhw {{.*#+}} xmm9 = xmm2[0,1,2,3,4,5,5,7] ; AVX512-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,6,5,7,7] ; AVX512-NEXT: vinserti32x4 $1, %xmm2, %ymm9, %ymm26 ; AVX512-NEXT: vmovdqa 32(%r10), %xmm2 ; AVX512-NEXT: vpunpckhbw {{.*#+}} 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] ; AVX512-NEXT: vpshuflw {{.*#+}} xmm0 = xmm1[0,0,2,1,4,5,6,7] ; AVX512-NEXT: vpshuflw {{.*#+}} xmm9 = xmm1[0,2,2,3,4,5,6,7] ; AVX512-NEXT: vinserti32x4 $1, %xmm9, %ymm0, %ymm28 ; AVX512-NEXT: vmovdqa 32(%rax), %xmm0 ; AVX512-NEXT: vpshufhw {{.*#+}} xmm9 = xmm1[0,1,2,3,4,4,6,5] ; AVX512-NEXT: vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,6,6,7] ; AVX512-NEXT: vinserti32x4 $1, %xmm1, %ymm9, %ymm29 ; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm4[8],xmm3[8],xmm4[9],xmm3[9],xmm4[10],xmm3[10],xmm4[11],xmm3[11],xmm4[12],xmm3[12],xmm4[13],xmm3[13],xmm4[14],xmm3[14],xmm4[15],xmm3[15] ; AVX512-NEXT: vpshuflw {{.*#+}} xmm3 = xmm1[0,0,2,1,4,5,6,7] ; AVX512-NEXT: vpshuflw {{.*#+}} xmm4 = xmm1[0,2,2,3,4,5,6,7] ; AVX512-NEXT: vinserti32x4 $1, %xmm4, %ymm3, %ymm25 ; AVX512-NEXT: vpshufhw {{.*#+}} xmm3 = xmm1[0,1,2,3,4,4,6,5] ; AVX512-NEXT: vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,6,6,7] ; AVX512-NEXT: vinserti32x4 $1, %xmm1, %ymm3, %ymm23 ; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm1 = 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] ; AVX512-NEXT: vpshuflw {{.*#+}} xmm3 = xmm1[0,1,1,3,4,5,6,7] ; AVX512-NEXT: vpshuflw {{.*#+}} xmm4 = xmm1[2,1,3,3,4,5,6,7] ; AVX512-NEXT: vinserti32x4 $1, %xmm4, %ymm3, %ymm19 ; AVX512-NEXT: vpshufhw {{.*#+}} xmm3 = xmm1[0,1,2,3,4,5,5,7] ; AVX512-NEXT: vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,6,5,7,7] ; AVX512-NEXT: vinserti32x4 $1, %xmm1, %ymm3, %ymm17 ; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm5[0],xmm12[0],xmm5[1],xmm12[1],xmm5[2],xmm12[2],xmm5[3],xmm12[3],xmm5[4],xmm12[4],xmm5[5],xmm12[5],xmm5[6],xmm12[6],xmm5[7],xmm12[7] ; AVX512-NEXT: vpshuflw {{.*#+}} xmm3 = xmm1[0,0,2,1,4,5,6,7] ; AVX512-NEXT: vpshuflw {{.*#+}} xmm4 = xmm1[0,2,2,3,4,5,6,7] ; AVX512-NEXT: vinserti32x4 $1, %xmm4, %ymm3, %ymm16 ; AVX512-NEXT: vpshufhw {{.*#+}} xmm3 = xmm1[0,1,2,3,4,4,6,5] ; AVX512-NEXT: vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,6,6,7] ; AVX512-NEXT: vinserti128 $1, %xmm1, %ymm3, %ymm14 ; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],xmm2[3],xmm0[4],xmm2[4],xmm0[5],xmm2[5],xmm0[6],xmm2[6],xmm0[7],xmm2[7] ; AVX512-NEXT: vpshuflw {{.*#+}} xmm3 = xmm1[0,0,2,1,4,5,6,7] ; AVX512-NEXT: vpshuflw {{.*#+}} xmm6 = xmm1[0,2,2,3,4,5,6,7] ; AVX512-NEXT: vinserti32x4 $1, %xmm6, %ymm3, %ymm20 ; AVX512-NEXT: vpshufhw {{.*#+}} xmm3 = xmm1[0,1,2,3,4,4,6,5] ; AVX512-NEXT: vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,6,6,7] ; AVX512-NEXT: vinserti32x4 $1, %xmm1, %ymm3, %ymm18 ; AVX512-NEXT: vmovdqa 32(%r9), %xmm1 ; AVX512-NEXT: vmovdqa 32(%r8), %xmm3 ; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm6 = 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] ; AVX512-NEXT: vpshuflw {{.*#+}} xmm8 = xmm6[0,1,1,3,4,5,6,7] ; AVX512-NEXT: vpshuflw {{.*#+}} xmm9 = xmm6[2,1,3,3,4,5,6,7] ; AVX512-NEXT: vinserti128 $1, %xmm9, %ymm8, %ymm15 ; AVX512-NEXT: vpshufhw {{.*#+}} xmm8 = xmm6[0,1,2,3,4,5,5,7] ; AVX512-NEXT: vpshufhw {{.*#+}} xmm6 = xmm6[0,1,2,3,6,5,7,7] ; AVX512-NEXT: vinserti128 $1, %xmm6, %ymm8, %ymm13 ; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm5 = xmm5[8],xmm12[8],xmm5[9],xmm12[9],xmm5[10],xmm12[10],xmm5[11],xmm12[11],xmm5[12],xmm12[12],xmm5[13],xmm12[13],xmm5[14],xmm12[14],xmm5[15],xmm12[15] ; AVX512-NEXT: vpshuflw {{.*#+}} xmm6 = xmm5[0,0,2,1,4,5,6,7] ; AVX512-NEXT: vpshuflw {{.*#+}} xmm8 = xmm5[0,2,2,3,4,5,6,7] ; AVX512-NEXT: vinserti128 $1, %xmm8, %ymm6, %ymm4 ; AVX512-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX512-NEXT: vpshufhw {{.*#+}} xmm6 = xmm5[0,1,2,3,4,4,6,5] ; AVX512-NEXT: vpshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,4,6,6,7] ; AVX512-NEXT: vinserti128 $1, %xmm5, %ymm6, %ymm4 ; AVX512-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm0[8],xmm2[8],xmm0[9],xmm2[9],xmm0[10],xmm2[10],xmm0[11],xmm2[11],xmm0[12],xmm2[12],xmm0[13],xmm2[13],xmm0[14],xmm2[14],xmm0[15],xmm2[15] ; AVX512-NEXT: vpshuflw {{.*#+}} xmm2 = xmm0[0,0,2,1,4,5,6,7] ; AVX512-NEXT: vpshuflw {{.*#+}} xmm5 = xmm0[0,2,2,3,4,5,6,7] ; AVX512-NEXT: vinserti128 $1, %xmm5, %ymm2, %ymm2 ; AVX512-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX512-NEXT: vpshufhw {{.*#+}} xmm2 = xmm0[0,1,2,3,4,4,6,5] ; AVX512-NEXT: vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,6,6,7] ; AVX512-NEXT: vinserti128 $1, %xmm0, %ymm2, %ymm0 ; AVX512-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm0 = 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] ; AVX512-NEXT: vpshuflw {{.*#+}} xmm1 = xmm0[0,1,1,3,4,5,6,7] ; AVX512-NEXT: vpshuflw {{.*#+}} xmm2 = xmm0[2,1,3,3,4,5,6,7] ; AVX512-NEXT: vinserti128 $1, %xmm2, %ymm1, %ymm1 ; AVX512-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX512-NEXT: vpshufhw {{.*#+}} xmm1 = xmm0[0,1,2,3,4,5,5,7] ; AVX512-NEXT: vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,6,5,7,7] ; AVX512-NEXT: vinserti128 $1, %xmm0, %ymm1, %ymm0 ; AVX512-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX512-NEXT: vmovdqa 16(%rcx), %xmm4 ; AVX512-NEXT: vmovdqa 16(%rdx), %xmm3 ; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm0 = 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] ; AVX512-NEXT: vpshuflw {{.*#+}} xmm1 = xmm0[0,0,2,1,4,5,6,7] ; AVX512-NEXT: vpshuflw {{.*#+}} xmm5 = xmm0[0,2,2,3,4,5,6,7] ; AVX512-NEXT: vinserti128 $1, %xmm5, %ymm1, %ymm1 ; AVX512-NEXT: vmovdqu %ymm1, (%rsp) # 32-byte Spill ; AVX512-NEXT: vpshufhw {{.*#+}} xmm1 = xmm0[0,1,2,3,4,4,6,5] ; AVX512-NEXT: vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,6,6,7] ; AVX512-NEXT: vinserti128 $1, %xmm0, %ymm1, %ymm0 ; AVX512-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX512-NEXT: vmovdqa 16(%r10), %xmm9 ; AVX512-NEXT: vmovdqa 16(%rax), %xmm8 ; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm8[0],xmm9[0],xmm8[1],xmm9[1],xmm8[2],xmm9[2],xmm8[3],xmm9[3],xmm8[4],xmm9[4],xmm8[5],xmm9[5],xmm8[6],xmm9[6],xmm8[7],xmm9[7] ; AVX512-NEXT: vpshuflw {{.*#+}} xmm1 = xmm0[0,0,2,1,4,5,6,7] ; AVX512-NEXT: vpshuflw {{.*#+}} xmm5 = xmm0[0,2,2,3,4,5,6,7] ; AVX512-NEXT: vinserti32x4 $1, %xmm5, %ymm1, %ymm30 ; AVX512-NEXT: vpshufhw {{.*#+}} xmm1 = xmm0[0,1,2,3,4,4,6,5] ; AVX512-NEXT: vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,6,6,7] ; AVX512-NEXT: vinserti32x4 $1, %xmm0, %ymm1, %ymm31 ; AVX512-NEXT: vmovdqa 16(%r9), %xmm6 ; AVX512-NEXT: vmovdqa 16(%r8), %xmm5 ; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm2 = 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] ; AVX512-NEXT: vpshuflw {{.*#+}} xmm1 = xmm2[0,1,1,3,4,5,6,7] ; AVX512-NEXT: vpshuflw {{.*#+}} xmm11 = xmm2[2,1,3,3,4,5,6,7] ; AVX512-NEXT: vinserti32x4 $1, %xmm11, %ymm1, %ymm27 ; AVX512-NEXT: vmovdqa 48(%rsi), %xmm1 ; AVX512-NEXT: vmovdqa 48(%rdi), %xmm11 ; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm11[0],xmm1[0],xmm11[1],xmm1[1],xmm11[2],xmm1[2],xmm11[3],xmm1[3],xmm11[4],xmm1[4],xmm11[5],xmm1[5],xmm11[6],xmm1[6],xmm11[7],xmm1[7] ; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm11[8],xmm1[8],xmm11[9],xmm1[9],xmm11[10],xmm1[10],xmm11[11],xmm1[11],xmm11[12],xmm1[12],xmm11[13],xmm1[13],xmm11[14],xmm1[14],xmm11[15],xmm1[15] ; AVX512-NEXT: vmovdqa (%rsi), %xmm7 ; AVX512-NEXT: vmovdqa (%rdi), %xmm12 ; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm11 = xmm12[8],xmm7[8],xmm12[9],xmm7[9],xmm12[10],xmm7[10],xmm12[11],xmm7[11],xmm12[12],xmm7[12],xmm12[13],xmm7[13],xmm12[14],xmm7[14],xmm12[15],xmm7[15] ; AVX512-NEXT: vmovdqa64 %xmm7, %xmm21 ; AVX512-NEXT: vpmovzxwq {{.*#+}} zmm22 = xmm11[0],zero,zero,zero,xmm11[1],zero,zero,zero,xmm11[2],zero,zero,zero,xmm11[3],zero,zero,zero,xmm11[4],zero,zero,zero,xmm11[5],zero,zero,zero,xmm11[6],zero,zero,zero,xmm11[7],zero,zero,zero ; AVX512-NEXT: vpmovzxwq {{.*#+}} zmm24 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero,xmm0[4],zero,zero,zero,xmm0[5],zero,zero,zero,xmm0[6],zero,zero,zero,xmm0[7],zero,zero,zero ; AVX512-NEXT: vpshufd {{.*#+}} ymm0 = ymm10[0,1,1,3,4,5,5,7] ; AVX512-NEXT: vpshufd $246, {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Folded Reload ; AVX512-NEXT: # ymm11 = mem[2,1,3,3,6,5,7,7] ; AVX512-NEXT: vinserti64x4 $1, %ymm11, %zmm0, %zmm0 ; AVX512-NEXT: vpbroadcastq {{.*#+}} zmm11 = [65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535] ; AVX512-NEXT: vpternlogq {{.*#+}} zmm22 = zmm22 | (zmm0 & ~zmm11) ; AVX512-NEXT: vpshufd $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Folded Reload ; AVX512-NEXT: # ymm0 = mem[0,0,2,1,4,4,6,5] ; AVX512-NEXT: vpshufd $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Folded Reload ; AVX512-NEXT: # ymm7 = mem[0,2,2,3,4,6,6,7] ; AVX512-NEXT: vinserti64x4 $1, %ymm7, %zmm0, %zmm0 ; AVX512-NEXT: vpshufd $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Folded Reload ; AVX512-NEXT: # ymm7 = mem[0,0,2,1,4,4,6,5] ; AVX512-NEXT: vpshufd $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Folded Reload ; AVX512-NEXT: # ymm10 = mem[0,2,2,3,4,6,6,7] ; AVX512-NEXT: vinserti64x4 $1, %ymm10, %zmm7, %zmm10 ; AVX512-NEXT: vpbroadcastq {{.*#+}} zmm7 = [65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0] ; AVX512-NEXT: vpandnq %zmm0, %zmm7, %zmm0 ; AVX512-NEXT: vpandq %zmm7, %zmm10, %zmm10 ; AVX512-NEXT: movw $-21846, %ax # imm = 0xAAAA ; AVX512-NEXT: kmovw %eax, %k1 ; AVX512-NEXT: vpord %zmm0, %zmm10, %zmm22 {%k1} ; AVX512-NEXT: vpshufd $212, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Folded Reload ; AVX512-NEXT: # ymm0 = mem[0,1,1,3,4,5,5,7] ; AVX512-NEXT: vpshufd $246, {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Folded Reload ; AVX512-NEXT: # ymm10 = mem[2,1,3,3,6,5,7,7] ; AVX512-NEXT: vinserti64x4 $1, %ymm10, %zmm0, %zmm0 ; AVX512-NEXT: vpternlogq {{.*#+}} zmm24 = zmm24 | (zmm0 & ~zmm11) ; AVX512-NEXT: vpshufd $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Folded Reload ; AVX512-NEXT: # ymm0 = mem[0,0,2,1,4,4,6,5] ; AVX512-NEXT: vpshufd $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Folded Reload ; AVX512-NEXT: # ymm10 = mem[0,2,2,3,4,6,6,7] ; AVX512-NEXT: vinserti64x4 $1, %ymm10, %zmm0, %zmm0 ; AVX512-NEXT: vpshufd $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Folded Reload ; AVX512-NEXT: # ymm10 = mem[0,0,2,1,4,4,6,5] ; AVX512-NEXT: vpshufd {{.*#+}} ymm26 = ymm26[0,2,2,3,4,6,6,7] ; AVX512-NEXT: vinserti64x4 $1, %ymm26, %zmm10, %zmm10 ; AVX512-NEXT: vpandnq %zmm0, %zmm7, %zmm0 ; AVX512-NEXT: vpandq %zmm7, %zmm10, %zmm10 ; AVX512-NEXT: vpord %zmm0, %zmm10, %zmm24 {%k1} ; AVX512-NEXT: vpshufd {{.*#+}} ymm0 = ymm28[0,1,1,3,4,5,5,7] ; AVX512-NEXT: vpshufd {{.*#+}} ymm10 = ymm29[2,1,3,3,6,5,7,7] ; AVX512-NEXT: vinserti64x4 $1, %ymm10, %zmm0, %zmm0 ; AVX512-NEXT: vpmovzxwq {{.*#+}} zmm10 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero,xmm1[2],zero,zero,zero,xmm1[3],zero,zero,zero,xmm1[4],zero,zero,zero,xmm1[5],zero,zero,zero,xmm1[6],zero,zero,zero,xmm1[7],zero,zero,zero ; AVX512-NEXT: vpternlogq {{.*#+}} zmm10 = zmm10 | (zmm0 & ~zmm11) ; AVX512-NEXT: vpshufd {{.*#+}} ymm0 = ymm25[0,0,2,1,4,4,6,5] ; AVX512-NEXT: vpshufd {{.*#+}} ymm1 = ymm23[0,2,2,3,4,6,6,7] ; AVX512-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0 ; AVX512-NEXT: vpshufd {{.*#+}} ymm1 = ymm19[0,0,2,1,4,4,6,5] ; AVX512-NEXT: vpshufd {{.*#+}} ymm17 = ymm17[0,2,2,3,4,6,6,7] ; AVX512-NEXT: vinserti64x4 $1, %ymm17, %zmm1, %zmm1 ; AVX512-NEXT: vpandnq %zmm0, %zmm7, %zmm0 ; AVX512-NEXT: vpandq %zmm7, %zmm1, %zmm1 ; AVX512-NEXT: vpord %zmm0, %zmm1, %zmm10 {%k1} ; AVX512-NEXT: vpshufd {{.*#+}} ymm0 = ymm16[0,1,1,3,4,5,5,7] ; AVX512-NEXT: vpshufd {{.*#+}} ymm1 = ymm14[2,1,3,3,6,5,7,7] ; AVX512-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm1 ; AVX512-NEXT: vpshufhw {{.*#+}} xmm0 = xmm2[0,1,2,3,4,5,5,7] ; AVX512-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,6,5,7,7] ; AVX512-NEXT: vinserti32x4 $1, %xmm2, %ymm0, %ymm17 ; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm2 = 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] ; AVX512-NEXT: vpshuflw {{.*#+}} xmm0 = xmm2[0,0,2,1,4,5,6,7] ; AVX512-NEXT: vpshuflw {{.*#+}} xmm3 = xmm2[0,2,2,3,4,5,6,7] ; AVX512-NEXT: vinserti32x4 $1, %xmm3, %ymm0, %ymm19 ; AVX512-NEXT: vmovdqa 32(%rsi), %xmm14 ; AVX512-NEXT: vmovdqa 32(%rdi), %xmm0 ; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm0[0],xmm14[0],xmm0[1],xmm14[1],xmm0[2],xmm14[2],xmm0[3],xmm14[3],xmm0[4],xmm14[4],xmm0[5],xmm14[5],xmm0[6],xmm14[6],xmm0[7],xmm14[7] ; AVX512-NEXT: vpmovzxwq {{.*#+}} zmm4 = xmm3[0],zero,zero,zero,xmm3[1],zero,zero,zero,xmm3[2],zero,zero,zero,xmm3[3],zero,zero,zero,xmm3[4],zero,zero,zero,xmm3[5],zero,zero,zero,xmm3[6],zero,zero,zero,xmm3[7],zero,zero,zero ; AVX512-NEXT: vpternlogq {{.*#+}} zmm4 = zmm4 | (zmm1 & ~zmm11) ; AVX512-NEXT: vpshufd {{.*#+}} ymm1 = ymm20[0,0,2,1,4,4,6,5] ; AVX512-NEXT: vpshufd {{.*#+}} ymm3 = ymm18[0,2,2,3,4,6,6,7] ; AVX512-NEXT: vinserti64x4 $1, %ymm3, %zmm1, %zmm1 ; AVX512-NEXT: vpshufd {{.*#+}} ymm3 = ymm15[0,0,2,1,4,4,6,5] ; AVX512-NEXT: vpshufd {{.*#+}} ymm13 = ymm13[0,2,2,3,4,6,6,7] ; AVX512-NEXT: vinserti64x4 $1, %ymm13, %zmm3, %zmm3 ; AVX512-NEXT: vpandnq %zmm1, %zmm7, %zmm1 ; AVX512-NEXT: vpandq %zmm7, %zmm3, %zmm3 ; AVX512-NEXT: vpord %zmm1, %zmm3, %zmm4 {%k1} ; AVX512-NEXT: vpshufhw {{.*#+}} xmm1 = xmm2[0,1,2,3,4,4,6,5] ; AVX512-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,6,6,7] ; AVX512-NEXT: vinserti32x4 $1, %xmm2, %ymm1, %ymm20 ; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm8[8],xmm9[8],xmm8[9],xmm9[9],xmm8[10],xmm9[10],xmm8[11],xmm9[11],xmm8[12],xmm9[12],xmm8[13],xmm9[13],xmm8[14],xmm9[14],xmm8[15],xmm9[15] ; AVX512-NEXT: vpshuflw {{.*#+}} xmm3 = xmm2[0,0,2,1,4,5,6,7] ; AVX512-NEXT: vpshuflw {{.*#+}} xmm8 = xmm2[0,2,2,3,4,5,6,7] ; AVX512-NEXT: vinserti32x4 $1, %xmm8, %ymm3, %ymm23 ; AVX512-NEXT: vpshufhw {{.*#+}} xmm3 = xmm2[0,1,2,3,4,4,6,5] ; AVX512-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,6,6,7] ; AVX512-NEXT: vinserti128 $1, %xmm2, %ymm3, %ymm9 ; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm3 = 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] ; AVX512-NEXT: vpshuflw {{.*#+}} xmm2 = xmm3[0,1,1,3,4,5,6,7] ; AVX512-NEXT: vpshuflw {{.*#+}} xmm5 = xmm3[2,1,3,3,4,5,6,7] ; AVX512-NEXT: vinserti128 $1, %xmm5, %ymm2, %ymm2 ; AVX512-NEXT: vpshufhw {{.*#+}} xmm5 = xmm3[0,1,2,3,4,5,5,7] ; AVX512-NEXT: vpshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,6,5,7,7] ; AVX512-NEXT: vinserti128 $1, %xmm3, %ymm5, %ymm3 ; AVX512-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload ; AVX512-NEXT: vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm5 # 16-byte Folded Reload ; AVX512-NEXT: # xmm5 = xmm1[0],mem[0],xmm1[1],mem[1],xmm1[2],mem[2],xmm1[3],mem[3],xmm1[4],mem[4],xmm1[5],mem[5],xmm1[6],mem[6],xmm1[7],mem[7] ; AVX512-NEXT: vpshuflw {{.*#+}} xmm6 = xmm5[0,0,2,1,4,5,6,7] ; AVX512-NEXT: vpshuflw {{.*#+}} xmm13 = xmm5[0,2,2,3,4,5,6,7] ; AVX512-NEXT: vinserti128 $1, %xmm13, %ymm6, %ymm13 ; AVX512-NEXT: vpshufhw {{.*#+}} xmm6 = xmm5[0,1,2,3,4,4,6,5] ; AVX512-NEXT: vpshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,4,6,6,7] ; AVX512-NEXT: vinserti128 $1, %xmm5, %ymm6, %ymm15 ; AVX512-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload ; AVX512-NEXT: vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm5 # 16-byte Folded Reload ; AVX512-NEXT: # xmm5 = xmm1[0],mem[0],xmm1[1],mem[1],xmm1[2],mem[2],xmm1[3],mem[3],xmm1[4],mem[4],xmm1[5],mem[5],xmm1[6],mem[6],xmm1[7],mem[7] ; AVX512-NEXT: vpshuflw {{.*#+}} xmm6 = xmm5[0,0,2,1,4,5,6,7] ; AVX512-NEXT: vpshuflw {{.*#+}} xmm1 = xmm5[0,2,2,3,4,5,6,7] ; AVX512-NEXT: vinserti32x4 $1, %xmm1, %ymm6, %ymm16 ; AVX512-NEXT: vpshufhw {{.*#+}} xmm1 = xmm5[0,1,2,3,4,4,6,5] ; AVX512-NEXT: vpshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,4,6,6,7] ; AVX512-NEXT: vinserti32x4 $1, %xmm5, %ymm1, %ymm18 ; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm0[8],xmm14[8],xmm0[9],xmm14[9],xmm0[10],xmm14[10],xmm0[11],xmm14[11],xmm0[12],xmm14[12],xmm0[13],xmm14[13],xmm0[14],xmm14[14],xmm0[15],xmm14[15] ; AVX512-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload ; AVX512-NEXT: vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload ; AVX512-NEXT: # xmm1 = xmm1[0],mem[0],xmm1[1],mem[1],xmm1[2],mem[2],xmm1[3],mem[3],xmm1[4],mem[4],xmm1[5],mem[5],xmm1[6],mem[6],xmm1[7],mem[7] ; AVX512-NEXT: vpshuflw {{.*#+}} xmm5 = xmm1[0,1,1,3,4,5,6,7] ; AVX512-NEXT: vpshuflw {{.*#+}} xmm6 = xmm1[2,1,3,3,4,5,6,7] ; AVX512-NEXT: vinserti128 $1, %xmm6, %ymm5, %ymm5 ; AVX512-NEXT: vpshufhw {{.*#+}} xmm6 = xmm1[0,1,2,3,4,5,5,7] ; AVX512-NEXT: vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,6,5,7,7] ; AVX512-NEXT: vinserti128 $1, %xmm1, %ymm6, %ymm1 ; AVX512-NEXT: vmovdqa 16(%rsi), %xmm6 ; AVX512-NEXT: vmovdqa 16(%rdi), %xmm14 ; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm8 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3],xmm14[4],xmm6[4],xmm14[5],xmm6[5],xmm14[6],xmm6[6],xmm14[7],xmm6[7] ; AVX512-NEXT: vpunpckhbw {{.*#+}} xmm6 = xmm14[8],xmm6[8],xmm14[9],xmm6[9],xmm14[10],xmm6[10],xmm14[11],xmm6[11],xmm14[12],xmm6[12],xmm14[13],xmm6[13],xmm14[14],xmm6[14],xmm14[15],xmm6[15] ; AVX512-NEXT: vmovdqa64 %xmm21, %xmm14 ; AVX512-NEXT: vpunpcklbw {{.*#+}} xmm12 = xmm12[0],xmm14[0],xmm12[1],xmm14[1],xmm12[2],xmm14[2],xmm12[3],xmm14[3],xmm12[4],xmm14[4],xmm12[5],xmm14[5],xmm12[6],xmm14[6],xmm12[7],xmm14[7] ; AVX512-NEXT: vpshufd $212, {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Folded Reload ; AVX512-NEXT: # ymm14 = mem[0,1,1,3,4,5,5,7] ; AVX512-NEXT: vpshufd $246, {{[-0-9]+}}(%r{{[sb]}}p), %ymm25 # 32-byte Folded Reload ; AVX512-NEXT: # ymm25 = mem[2,1,3,3,6,5,7,7] ; AVX512-NEXT: vinserti64x4 $1, %ymm25, %zmm14, %zmm14 ; AVX512-NEXT: vpmovzxwq {{.*#+}} zmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero,xmm0[4],zero,zero,zero,xmm0[5],zero,zero,zero,xmm0[6],zero,zero,zero,xmm0[7],zero,zero,zero ; AVX512-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 | (zmm14 & ~zmm11) ; AVX512-NEXT: vpmovzxwq {{.*#+}} zmm8 = xmm8[0],zero,zero,zero,xmm8[1],zero,zero,zero,xmm8[2],zero,zero,zero,xmm8[3],zero,zero,zero,xmm8[4],zero,zero,zero,xmm8[5],zero,zero,zero,xmm8[6],zero,zero,zero,xmm8[7],zero,zero,zero ; AVX512-NEXT: vpmovzxwq {{.*#+}} zmm6 = xmm6[0],zero,zero,zero,xmm6[1],zero,zero,zero,xmm6[2],zero,zero,zero,xmm6[3],zero,zero,zero,xmm6[4],zero,zero,zero,xmm6[5],zero,zero,zero,xmm6[6],zero,zero,zero,xmm6[7],zero,zero,zero ; AVX512-NEXT: vpmovzxwq {{.*#+}} zmm12 = xmm12[0],zero,zero,zero,xmm12[1],zero,zero,zero,xmm12[2],zero,zero,zero,xmm12[3],zero,zero,zero,xmm12[4],zero,zero,zero,xmm12[5],zero,zero,zero,xmm12[6],zero,zero,zero,xmm12[7],zero,zero,zero ; AVX512-NEXT: vpshufd $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Folded Reload ; AVX512-NEXT: # ymm14 = mem[0,0,2,1,4,4,6,5] ; AVX512-NEXT: vpshufd $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm25 # 32-byte Folded Reload ; AVX512-NEXT: # ymm25 = mem[0,2,2,3,4,6,6,7] ; AVX512-NEXT: vinserti64x4 $1, %ymm25, %zmm14, %zmm14 ; AVX512-NEXT: vpshufd $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm25 # 32-byte Folded Reload ; AVX512-NEXT: # ymm25 = mem[0,0,2,1,4,4,6,5] ; AVX512-NEXT: vpshufd $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm26 # 32-byte Folded Reload ; AVX512-NEXT: # ymm26 = mem[0,2,2,3,4,6,6,7] ; AVX512-NEXT: vinserti64x4 $1, %ymm26, %zmm25, %zmm25 ; AVX512-NEXT: vpandnq %zmm14, %zmm7, %zmm14 ; AVX512-NEXT: vpandq %zmm7, %zmm25, %zmm25 ; AVX512-NEXT: vpord %zmm14, %zmm25, %zmm0 {%k1} ; AVX512-NEXT: vpshufd $212, (%rsp), %ymm14 # 32-byte Folded Reload ; AVX512-NEXT: # ymm14 = mem[0,1,1,3,4,5,5,7] ; AVX512-NEXT: vpshufd $246, {{[-0-9]+}}(%r{{[sb]}}p), %ymm25 # 32-byte Folded Reload ; AVX512-NEXT: # ymm25 = mem[2,1,3,3,6,5,7,7] ; AVX512-NEXT: vinserti64x4 $1, %ymm25, %zmm14, %zmm14 ; AVX512-NEXT: vpternlogq {{.*#+}} zmm8 = zmm8 | (zmm14 & ~zmm11) ; AVX512-NEXT: vpshufd {{.*#+}} ymm14 = ymm30[0,0,2,1,4,4,6,5] ; AVX512-NEXT: vpshufd {{.*#+}} ymm25 = ymm31[0,2,2,3,4,6,6,7] ; AVX512-NEXT: vinserti64x4 $1, %ymm25, %zmm14, %zmm14 ; AVX512-NEXT: vpshufd {{.*#+}} ymm21 = ymm27[0,0,2,1,4,4,6,5] ; AVX512-NEXT: vpshufd {{.*#+}} ymm17 = ymm17[0,2,2,3,4,6,6,7] ; AVX512-NEXT: vinserti64x4 $1, %ymm17, %zmm21, %zmm17 ; AVX512-NEXT: vpandnq %zmm14, %zmm7, %zmm14 ; AVX512-NEXT: vpandq %zmm7, %zmm17, %zmm17 ; AVX512-NEXT: vpord %zmm14, %zmm17, %zmm8 {%k1} ; AVX512-NEXT: vpshufd {{.*#+}} ymm14 = ymm19[0,1,1,3,4,5,5,7] ; AVX512-NEXT: vpshufd {{.*#+}} ymm17 = ymm20[2,1,3,3,6,5,7,7] ; AVX512-NEXT: vinserti64x4 $1, %ymm17, %zmm14, %zmm14 ; AVX512-NEXT: vpternlogq {{.*#+}} zmm6 = zmm6 | (zmm14 & ~zmm11) ; AVX512-NEXT: vpshufd {{.*#+}} ymm14 = ymm23[0,0,2,1,4,4,6,5] ; AVX512-NEXT: vpshufd {{.*#+}} ymm9 = ymm9[0,2,2,3,4,6,6,7] ; AVX512-NEXT: vinserti64x4 $1, %ymm9, %zmm14, %zmm9 ; AVX512-NEXT: vpshufd {{.*#+}} ymm2 = ymm2[0,0,2,1,4,4,6,5] ; AVX512-NEXT: vpshufd {{.*#+}} ymm3 = ymm3[0,2,2,3,4,6,6,7] ; AVX512-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm2 ; AVX512-NEXT: vpandnq %zmm9, %zmm7, %zmm3 ; AVX512-NEXT: vpandq %zmm7, %zmm2, %zmm2 ; AVX512-NEXT: vpord %zmm3, %zmm2, %zmm6 {%k1} ; AVX512-NEXT: vpshufd {{.*#+}} ymm2 = ymm13[0,1,1,3,4,5,5,7] ; AVX512-NEXT: vpshufd {{.*#+}} ymm3 = ymm15[2,1,3,3,6,5,7,7] ; AVX512-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm2 ; AVX512-NEXT: vpternlogq {{.*#+}} zmm12 = zmm12 | (zmm2 & ~zmm11) ; AVX512-NEXT: vpshufd {{.*#+}} ymm2 = ymm16[0,0,2,1,4,4,6,5] ; AVX512-NEXT: vpshufd {{.*#+}} ymm3 = ymm18[0,2,2,3,4,6,6,7] ; AVX512-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm2 ; AVX512-NEXT: vpshufd {{.*#+}} ymm3 = ymm5[0,0,2,1,4,4,6,5] ; AVX512-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[0,2,2,3,4,6,6,7] ; AVX512-NEXT: vinserti64x4 $1, %ymm1, %zmm3, %zmm1 ; AVX512-NEXT: vpandnq %zmm2, %zmm7, %zmm2 ; AVX512-NEXT: vpandq %zmm7, %zmm1, %zmm1 ; AVX512-NEXT: vpord %zmm2, %zmm1, %zmm12 {%k1} ; AVX512-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512-NEXT: vmovdqa64 %zmm12, (%rax) ; AVX512-NEXT: vmovdqa64 %zmm6, 192(%rax) ; AVX512-NEXT: vmovdqa64 %zmm8, 128(%rax) ; AVX512-NEXT: vmovdqa64 %zmm0, 320(%rax) ; AVX512-NEXT: vmovdqa64 %zmm4, 256(%rax) ; AVX512-NEXT: vmovdqa64 %zmm10, 448(%rax) ; AVX512-NEXT: vmovdqa64 %zmm24, 384(%rax) ; AVX512-NEXT: vmovdqa64 %zmm22, 64(%rax) ; AVX512-NEXT: addq $552, %rsp # imm = 0x228 ; AVX512-NEXT: vzeroupper ; AVX512-NEXT: retq ; ; AVX512-FCP-LABEL: store_i8_stride8_vf64: ; AVX512-FCP: # %bb.0: ; AVX512-FCP-NEXT: subq $392, %rsp # imm = 0x188 ; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX512-FCP-NEXT: vmovdqa (%rcx), %xmm2 ; AVX512-FCP-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX512-FCP-NEXT: vmovdqa 48(%rcx), %xmm0 ; AVX512-FCP-NEXT: vmovdqa (%rdx), %xmm15 ; AVX512-FCP-NEXT: vmovdqa 48(%rdx), %xmm1 ; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm15[8],xmm2[8],xmm15[9],xmm2[9],xmm15[10],xmm2[10],xmm15[11],xmm2[11],xmm15[12],xmm2[12],xmm15[13],xmm2[13],xmm15[14],xmm2[14],xmm15[15],xmm2[15] ; AVX512-FCP-NEXT: vinserti128 $1, %xmm2, %ymm2, %ymm2 ; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm6 = [8,9,8,9,4,5,6,7,12,13,10,11,12,13,10,11,8,9,12,13,4,5,6,7,12,13,14,15,12,13,14,15] ; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm2, %ymm3 ; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm8 = [0,1,0,1,4,5,2,3,4,5,2,3,12,13,14,15,0,1,4,5,4,5,6,7,4,5,6,7,12,13,14,15] ; AVX512-FCP-NEXT: vpshufb %ymm8, %ymm2, %ymm2 ; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm2 ; AVX512-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill ; AVX512-FCP-NEXT: vmovdqa (%r10), %xmm2 ; AVX512-FCP-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX512-FCP-NEXT: vmovdqa (%rax), %xmm14 ; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm14[8],xmm2[8],xmm14[9],xmm2[9],xmm14[10],xmm2[10],xmm14[11],xmm2[11],xmm14[12],xmm2[12],xmm14[13],xmm2[13],xmm14[14],xmm2[14],xmm14[15],xmm2[15] ; AVX512-FCP-NEXT: vinserti128 $1, %xmm2, %ymm2, %ymm2 ; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm11 = [0,1,2,3,8,9,8,9,8,9,8,9,12,13,10,11,0,1,2,3,8,9,12,13,8,9,12,13,12,13,14,15] ; AVX512-FCP-NEXT: vpshufb %ymm11, %ymm2, %ymm3 ; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm10 = [0,1,0,1,0,1,0,1,8,9,10,11,4,5,2,3,0,1,4,5,0,1,4,5,8,9,10,11,4,5,6,7] ; AVX512-FCP-NEXT: vpshufb %ymm10, %ymm2, %ymm2 ; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm19 ; AVX512-FCP-NEXT: vmovdqa (%r9), %xmm3 ; AVX512-FCP-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX512-FCP-NEXT: vmovdqa (%r8), %xmm2 ; AVX512-FCP-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm2 = 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-FCP-NEXT: vinserti128 $1, %xmm2, %ymm2, %ymm2 ; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm9 = [0,1,2,3,8,9,10,11,8,9,10,11,10,11,14,15,0,1,2,3,12,13,10,11,12,13,10,11,14,15,14,15] ; AVX512-FCP-NEXT: vpshufb %ymm9, %ymm2, %ymm3 ; AVX512-FCP-NEXT: vmovdqa {{.*#+}} ymm12 = [0,1,2,3,0,1,2,3,8,9,10,11,2,3,6,7,4,5,2,3,4,5,2,3,8,9,10,11,6,7,6,7] ; AVX512-FCP-NEXT: vpshufb %ymm12, %ymm2, %ymm2 ; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm21 ; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3],xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7] ; AVX512-FCP-NEXT: vinserti128 $1, %xmm2, %ymm2, %ymm2 ; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm2, %ymm3 ; AVX512-FCP-NEXT: vpshufb %ymm8, %ymm2, %ymm2 ; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm2 ; AVX512-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill ; AVX512-FCP-NEXT: vmovdqa 48(%r10), %xmm2 ; AVX512-FCP-NEXT: vmovdqa 48(%rax), %xmm3 ; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm4 = 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] ; AVX512-FCP-NEXT: vinserti128 $1, %xmm4, %ymm4, %ymm4 ; AVX512-FCP-NEXT: vpshufb %ymm11, %ymm4, %ymm5 ; AVX512-FCP-NEXT: vpshufb %ymm10, %ymm4, %ymm4 ; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm5, %zmm4, %zmm4 ; AVX512-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill ; AVX512-FCP-NEXT: vmovdqa 48(%r9), %xmm4 ; AVX512-FCP-NEXT: vmovdqa 48(%r8), %xmm5 ; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm7 = 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] ; AVX512-FCP-NEXT: vinserti128 $1, %xmm7, %ymm7, %ymm7 ; AVX512-FCP-NEXT: vpshufb %ymm9, %ymm7, %ymm13 ; AVX512-FCP-NEXT: vpshufb %ymm12, %ymm7, %ymm7 ; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm13, %zmm7, %zmm7 ; AVX512-FCP-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill ; AVX512-FCP-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-FCP-NEXT: vinserti128 $1, %xmm0, %ymm0, %ymm0 ; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm0, %ymm1 ; AVX512-FCP-NEXT: vpshufb %ymm8, %ymm0, %ymm0 ; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill ; 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: vinserti128 $1, %xmm0, %ymm0, %ymm0 ; AVX512-FCP-NEXT: vpshufb %ymm11, %ymm0, %ymm1 ; AVX512-FCP-NEXT: vpshufb %ymm10, %ymm0, %ymm0 ; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0 ; AVX512-FCP-NEXT: vmovdqu64 %zmm0, (%rsp) # 64-byte Spill ; AVX512-FCP-NEXT: vmovdqa 32(%rcx), %xmm0 ; AVX512-FCP-NEXT: vmovdqa 32(%rdx), %xmm1 ; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm5[8],xmm4[8],xmm5[9],xmm4[9],xmm5[10],xmm4[10],xmm5[11],xmm4[11],xmm5[12],xmm4[12],xmm5[13],xmm4[13],xmm5[14],xmm4[14],xmm5[15],xmm4[15] ; AVX512-FCP-NEXT: vinserti128 $1, %xmm2, %ymm2, %ymm2 ; AVX512-FCP-NEXT: vpshufb %ymm9, %ymm2, %ymm3 ; AVX512-FCP-NEXT: vpshufb %ymm12, %ymm2, %ymm2 ; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm2 ; AVX512-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill ; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3],xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7] ; AVX512-FCP-NEXT: vinserti128 $1, %xmm2, %ymm2, %ymm2 ; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm2, %ymm3 ; AVX512-FCP-NEXT: vpshufb %ymm8, %ymm2, %ymm2 ; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm20 ; AVX512-FCP-NEXT: vmovdqa 32(%r10), %xmm2 ; AVX512-FCP-NEXT: vmovdqa 32(%rax), %xmm3 ; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm4 = 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] ; AVX512-FCP-NEXT: vinserti128 $1, %xmm4, %ymm4, %ymm4 ; AVX512-FCP-NEXT: vpshufb %ymm11, %ymm4, %ymm5 ; AVX512-FCP-NEXT: vpshufb %ymm10, %ymm4, %ymm4 ; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm5, %zmm4, %zmm22 ; AVX512-FCP-NEXT: vmovdqa 32(%r9), %xmm4 ; AVX512-FCP-NEXT: vmovdqa 32(%r8), %xmm5 ; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm7 = 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] ; AVX512-FCP-NEXT: vinserti128 $1, %xmm7, %ymm7, %ymm7 ; AVX512-FCP-NEXT: vpshufb %ymm9, %ymm7, %ymm13 ; AVX512-FCP-NEXT: vpshufb %ymm12, %ymm7, %ymm7 ; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm13, %zmm7, %zmm23 ; AVX512-FCP-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-FCP-NEXT: vinserti128 $1, %xmm0, %ymm0, %ymm0 ; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm0, %ymm1 ; AVX512-FCP-NEXT: vpshufb %ymm8, %ymm0, %ymm0 ; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm24 ; 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: vinserti128 $1, %xmm0, %ymm0, %ymm0 ; AVX512-FCP-NEXT: vpshufb %ymm11, %ymm0, %ymm1 ; AVX512-FCP-NEXT: vpshufb %ymm10, %ymm0, %ymm0 ; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm25 ; AVX512-FCP-NEXT: vmovdqa 16(%rcx), %xmm0 ; AVX512-FCP-NEXT: vmovdqa 16(%rdx), %xmm1 ; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm5[8],xmm4[8],xmm5[9],xmm4[9],xmm5[10],xmm4[10],xmm5[11],xmm4[11],xmm5[12],xmm4[12],xmm5[13],xmm4[13],xmm5[14],xmm4[14],xmm5[15],xmm4[15] ; AVX512-FCP-NEXT: vinserti128 $1, %xmm2, %ymm2, %ymm2 ; AVX512-FCP-NEXT: vpshufb %ymm9, %ymm2, %ymm3 ; AVX512-FCP-NEXT: vpshufb %ymm12, %ymm2, %ymm2 ; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm26 ; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3],xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7] ; AVX512-FCP-NEXT: vinserti128 $1, %xmm2, %ymm2, %ymm2 ; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm2, %ymm3 ; AVX512-FCP-NEXT: vpshufb %ymm8, %ymm2, %ymm2 ; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm27 ; AVX512-FCP-NEXT: vmovdqa 16(%r10), %xmm2 ; AVX512-FCP-NEXT: vmovdqa 16(%rax), %xmm3 ; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm4 = 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] ; AVX512-FCP-NEXT: vinserti128 $1, %xmm4, %ymm4, %ymm4 ; AVX512-FCP-NEXT: vpshufb %ymm11, %ymm4, %ymm5 ; AVX512-FCP-NEXT: vpshufb %ymm10, %ymm4, %ymm4 ; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm5, %zmm4, %zmm28 ; AVX512-FCP-NEXT: vmovdqa 16(%r9), %xmm4 ; AVX512-FCP-NEXT: vmovdqa 16(%r8), %xmm5 ; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm7 = 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] ; AVX512-FCP-NEXT: vinserti128 $1, %xmm7, %ymm7, %ymm7 ; AVX512-FCP-NEXT: vpshufb %ymm9, %ymm7, %ymm13 ; AVX512-FCP-NEXT: vpshufb %ymm12, %ymm7, %ymm7 ; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm13, %zmm7, %zmm29 ; AVX512-FCP-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-FCP-NEXT: vinserti128 $1, %xmm0, %ymm0, %ymm0 ; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm0, %ymm1 ; AVX512-FCP-NEXT: vpshufb %ymm8, %ymm0, %ymm0 ; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm30 ; 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: vinserti128 $1, %xmm0, %ymm0, %ymm0 ; AVX512-FCP-NEXT: vpshufb %ymm11, %ymm0, %ymm1 ; AVX512-FCP-NEXT: vpshufb %ymm10, %ymm0, %ymm0 ; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm31 ; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm0 = 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-FCP-NEXT: vinserti128 $1, %xmm0, %ymm0, %ymm0 ; AVX512-FCP-NEXT: vpshufb %ymm9, %ymm0, %ymm1 ; AVX512-FCP-NEXT: vmovdqa64 %ymm9, %ymm17 ; AVX512-FCP-NEXT: vpshufb %ymm12, %ymm0, %ymm0 ; AVX512-FCP-NEXT: vmovdqa64 %ymm12, %ymm18 ; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm16 ; AVX512-FCP-NEXT: vmovdqa (%rsi), %xmm12 ; AVX512-FCP-NEXT: vmovdqa (%rdi), %xmm9 ; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm9[8],xmm12[8],xmm9[9],xmm12[9],xmm9[10],xmm12[10],xmm9[11],xmm12[11],xmm9[12],xmm12[12],xmm9[13],xmm12[13],xmm9[14],xmm12[14],xmm9[15],xmm12[15] ; AVX512-FCP-NEXT: vpmovsxwq {{.*#+}} xmm5 = [1284,1798] ; AVX512-FCP-NEXT: vpmovzxwq {{.*#+}} xmm1 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero ; AVX512-FCP-NEXT: vpshufb %xmm5, %xmm0, %xmm2 ; AVX512-FCP-NEXT: vinserti128 $1, %xmm2, %ymm1, %ymm1 ; AVX512-FCP-NEXT: vinserti128 $1, %xmm0, %ymm0, %ymm0 ; AVX512-FCP-NEXT: vpmovsxwq {{.*#+}} ymm4 = [2312,2826,3340,3854] ; AVX512-FCP-NEXT: vpshufb %ymm4, %ymm0, %ymm0 ; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm3 ; AVX512-FCP-NEXT: vpbroadcastq {{.*#+}} zmm7 = [65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535] ; AVX512-FCP-NEXT: vpternlogq $226, {{[-0-9]+}}(%r{{[sb]}}p), %zmm7, %zmm3 # 64-byte Folded Reload ; AVX512-FCP-NEXT: # zmm3 = mem ^ (zmm7 & (zmm3 ^ mem)) ; AVX512-FCP-NEXT: vpbroadcastq {{.*#+}} zmm2 = [65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0] ; AVX512-FCP-NEXT: vpandnq %zmm19, %zmm2, %zmm19 ; AVX512-FCP-NEXT: vpandq %zmm2, %zmm21, %zmm21 ; AVX512-FCP-NEXT: movw $-21846, %ax # imm = 0xAAAA ; AVX512-FCP-NEXT: kmovw %eax, %k1 ; AVX512-FCP-NEXT: vpord %zmm19, %zmm21, %zmm3 {%k1} ; AVX512-FCP-NEXT: vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload ; AVX512-FCP-NEXT: # xmm15 = xmm15[0],mem[0],xmm15[1],mem[1],xmm15[2],mem[2],xmm15[3],mem[3],xmm15[4],mem[4],xmm15[5],mem[5],xmm15[6],mem[6],xmm15[7],mem[7] ; AVX512-FCP-NEXT: vinserti128 $1, %xmm15, %ymm15, %ymm15 ; AVX512-FCP-NEXT: vpshufb %ymm6, %ymm15, %ymm6 ; AVX512-FCP-NEXT: vpshufb %ymm8, %ymm15, %ymm8 ; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm6, %zmm8, %zmm19 ; AVX512-FCP-NEXT: vmovdqa 48(%rsi), %xmm15 ; AVX512-FCP-NEXT: vmovdqa 48(%rdi), %xmm8 ; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm6 = xmm8[0],xmm15[0],xmm8[1],xmm15[1],xmm8[2],xmm15[2],xmm8[3],xmm15[3],xmm8[4],xmm15[4],xmm8[5],xmm15[5],xmm8[6],xmm15[6],xmm8[7],xmm15[7] ; AVX512-FCP-NEXT: vpshufb %xmm5, %xmm6, %xmm0 ; AVX512-FCP-NEXT: vpmovzxwq {{.*#+}} xmm21 = xmm6[0],zero,zero,zero,xmm6[1],zero,zero,zero ; AVX512-FCP-NEXT: vinserti32x4 $1, %xmm0, %ymm21, %ymm0 ; AVX512-FCP-NEXT: vinserti128 $1, %xmm6, %ymm6, %ymm6 ; AVX512-FCP-NEXT: vpshufb %ymm4, %ymm6, %ymm6 ; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm6, %zmm0, %zmm6 ; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm8[8],xmm15[8],xmm8[9],xmm15[9],xmm8[10],xmm15[10],xmm8[11],xmm15[11],xmm8[12],xmm15[12],xmm8[13],xmm15[13],xmm8[14],xmm15[14],xmm8[15],xmm15[15] ; AVX512-FCP-NEXT: vpshufb %xmm5, %xmm0, %xmm8 ; AVX512-FCP-NEXT: vpmovzxwq {{.*#+}} xmm15 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero ; AVX512-FCP-NEXT: vinserti128 $1, %xmm8, %ymm15, %ymm8 ; AVX512-FCP-NEXT: vinserti128 $1, %xmm0, %ymm0, %ymm0 ; AVX512-FCP-NEXT: vpshufb %ymm4, %ymm0, %ymm0 ; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm8, %zmm15 ; AVX512-FCP-NEXT: vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm0 # 16-byte Folded Reload ; AVX512-FCP-NEXT: # xmm0 = xmm14[0],mem[0],xmm14[1],mem[1],xmm14[2],mem[2],xmm14[3],mem[3],xmm14[4],mem[4],xmm14[5],mem[5],xmm14[6],mem[6],xmm14[7],mem[7] ; AVX512-FCP-NEXT: vinserti128 $1, %xmm0, %ymm0, %ymm0 ; AVX512-FCP-NEXT: vpshufb %ymm11, %ymm0, %ymm8 ; AVX512-FCP-NEXT: vpshufb %ymm10, %ymm0, %ymm0 ; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm8, %zmm0, %zmm13 ; AVX512-FCP-NEXT: vmovdqa 32(%rsi), %xmm0 ; AVX512-FCP-NEXT: vmovdqa 32(%rdi), %xmm8 ; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm11 = 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] ; AVX512-FCP-NEXT: vpshufb %xmm5, %xmm11, %xmm14 ; AVX512-FCP-NEXT: vpmovzxwq {{.*#+}} xmm21 = xmm11[0],zero,zero,zero,xmm11[1],zero,zero,zero ; AVX512-FCP-NEXT: vinserti32x4 $1, %xmm14, %ymm21, %ymm14 ; AVX512-FCP-NEXT: vinserti128 $1, %xmm11, %ymm11, %ymm11 ; AVX512-FCP-NEXT: vpshufb %ymm4, %ymm11, %ymm11 ; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm11, %zmm14, %zmm11 ; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm8[8],xmm0[8],xmm8[9],xmm0[9],xmm8[10],xmm0[10],xmm8[11],xmm0[11],xmm8[12],xmm0[12],xmm8[13],xmm0[13],xmm8[14],xmm0[14],xmm8[15],xmm0[15] ; AVX512-FCP-NEXT: vpshufb %xmm5, %xmm0, %xmm8 ; AVX512-FCP-NEXT: vpmovzxwq {{.*#+}} xmm14 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero ; AVX512-FCP-NEXT: vinserti128 $1, %xmm8, %ymm14, %ymm8 ; AVX512-FCP-NEXT: vinserti128 $1, %xmm0, %ymm0, %ymm0 ; AVX512-FCP-NEXT: vpshufb %ymm4, %ymm0, %ymm0 ; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm8, %zmm14 ; AVX512-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload ; AVX512-FCP-NEXT: vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload ; AVX512-FCP-NEXT: # 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: vinserti128 $1, %xmm0, %ymm0, %ymm0 ; AVX512-FCP-NEXT: vmovdqa64 %ymm17, %ymm1 ; AVX512-FCP-NEXT: vpshufb %ymm1, %ymm0, %ymm8 ; AVX512-FCP-NEXT: vmovdqa64 %ymm18, %ymm1 ; AVX512-FCP-NEXT: vpshufb %ymm1, %ymm0, %ymm0 ; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm8, %zmm0, %zmm21 ; AVX512-FCP-NEXT: vmovdqa 16(%rsi), %xmm8 ; AVX512-FCP-NEXT: vmovdqa 16(%rdi), %xmm10 ; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm10[0],xmm8[0],xmm10[1],xmm8[1],xmm10[2],xmm8[2],xmm10[3],xmm8[3],xmm10[4],xmm8[4],xmm10[5],xmm8[5],xmm10[6],xmm8[6],xmm10[7],xmm8[7] ; AVX512-FCP-NEXT: vpshufb %xmm5, %xmm0, %xmm1 ; AVX512-FCP-NEXT: vpmovzxwq {{.*#+}} xmm17 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero ; AVX512-FCP-NEXT: vinserti32x4 $1, %xmm1, %ymm17, %ymm1 ; AVX512-FCP-NEXT: vinserti128 $1, %xmm0, %ymm0, %ymm0 ; AVX512-FCP-NEXT: vpshufb %ymm4, %ymm0, %ymm0 ; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0 ; AVX512-FCP-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm10[8],xmm8[8],xmm10[9],xmm8[9],xmm10[10],xmm8[10],xmm10[11],xmm8[11],xmm10[12],xmm8[12],xmm10[13],xmm8[13],xmm10[14],xmm8[14],xmm10[15],xmm8[15] ; AVX512-FCP-NEXT: vpshufb %xmm5, %xmm1, %xmm8 ; AVX512-FCP-NEXT: vpmovzxwq {{.*#+}} xmm10 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero ; AVX512-FCP-NEXT: vinserti128 $1, %xmm8, %ymm10, %ymm8 ; AVX512-FCP-NEXT: vinserti128 $1, %xmm1, %ymm1, %ymm1 ; AVX512-FCP-NEXT: vpshufb %ymm4, %ymm1, %ymm1 ; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm8, %zmm1 ; AVX512-FCP-NEXT: vpunpcklbw {{.*#+}} xmm8 = xmm9[0],xmm12[0],xmm9[1],xmm12[1],xmm9[2],xmm12[2],xmm9[3],xmm12[3],xmm9[4],xmm12[4],xmm9[5],xmm12[5],xmm9[6],xmm12[6],xmm9[7],xmm12[7] ; AVX512-FCP-NEXT: vinserti128 $1, %xmm8, %ymm8, %ymm9 ; AVX512-FCP-NEXT: vpshufb %ymm4, %ymm9, %ymm4 ; AVX512-FCP-NEXT: vpshufb %xmm5, %xmm8, %xmm5 ; AVX512-FCP-NEXT: vpmovzxwq {{.*#+}} xmm8 = xmm8[0],zero,zero,zero,xmm8[1],zero,zero,zero ; AVX512-FCP-NEXT: vinserti128 $1, %xmm5, %ymm8, %ymm5 ; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm5, %zmm4 ; AVX512-FCP-NEXT: vpternlogq $226, {{[-0-9]+}}(%r{{[sb]}}p), %zmm7, %zmm6 # 64-byte Folded Reload ; AVX512-FCP-NEXT: # zmm6 = mem ^ (zmm7 & (zmm6 ^ mem)) ; AVX512-FCP-NEXT: vpandnq {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm5 # 64-byte Folded Reload ; AVX512-FCP-NEXT: vpandq {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm8 # 64-byte Folded Reload ; AVX512-FCP-NEXT: vpord %zmm5, %zmm8, %zmm6 {%k1} ; AVX512-FCP-NEXT: vpternlogq $226, {{[-0-9]+}}(%r{{[sb]}}p), %zmm7, %zmm15 # 64-byte Folded Reload ; AVX512-FCP-NEXT: # zmm15 = mem ^ (zmm7 & (zmm15 ^ mem)) ; AVX512-FCP-NEXT: vpandnq (%rsp), %zmm2, %zmm5 # 64-byte Folded Reload ; AVX512-FCP-NEXT: vpandq {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm8 # 64-byte Folded Reload ; AVX512-FCP-NEXT: vpord %zmm5, %zmm8, %zmm15 {%k1} ; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm11 = zmm20 ^ (zmm7 & (zmm11 ^ zmm20)) ; AVX512-FCP-NEXT: vpandnq %zmm22, %zmm2, %zmm5 ; AVX512-FCP-NEXT: vpandq %zmm2, %zmm23, %zmm8 ; AVX512-FCP-NEXT: vpord %zmm5, %zmm8, %zmm11 {%k1} ; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm14 = zmm24 ^ (zmm7 & (zmm14 ^ zmm24)) ; AVX512-FCP-NEXT: vpandnq %zmm25, %zmm2, %zmm5 ; AVX512-FCP-NEXT: vpandq %zmm2, %zmm26, %zmm8 ; AVX512-FCP-NEXT: vpord %zmm5, %zmm8, %zmm14 {%k1} ; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm0 = zmm27 ^ (zmm7 & (zmm0 ^ zmm27)) ; AVX512-FCP-NEXT: vpandnq %zmm28, %zmm2, %zmm5 ; AVX512-FCP-NEXT: vpandq %zmm2, %zmm29, %zmm8 ; AVX512-FCP-NEXT: vpord %zmm5, %zmm8, %zmm0 {%k1} ; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm1 = zmm30 ^ (zmm7 & (zmm1 ^ zmm30)) ; AVX512-FCP-NEXT: vpandnq %zmm31, %zmm2, %zmm5 ; AVX512-FCP-NEXT: vpandq %zmm2, %zmm16, %zmm8 ; AVX512-FCP-NEXT: vpord %zmm5, %zmm8, %zmm1 {%k1} ; AVX512-FCP-NEXT: vpternlogq {{.*#+}} zmm4 = zmm19 ^ (zmm7 & (zmm4 ^ zmm19)) ; AVX512-FCP-NEXT: vpandnq %zmm13, %zmm2, %zmm5 ; AVX512-FCP-NEXT: vpandq %zmm2, %zmm21, %zmm2 ; AVX512-FCP-NEXT: vpord %zmm5, %zmm2, %zmm4 {%k1} ; AVX512-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512-FCP-NEXT: vmovdqa64 %zmm4, (%rax) ; AVX512-FCP-NEXT: vmovdqa64 %zmm1, 192(%rax) ; AVX512-FCP-NEXT: vmovdqa64 %zmm0, 128(%rax) ; AVX512-FCP-NEXT: vmovdqa64 %zmm14, 320(%rax) ; AVX512-FCP-NEXT: vmovdqa64 %zmm11, 256(%rax) ; AVX512-FCP-NEXT: vmovdqa64 %zmm15, 448(%rax) ; AVX512-FCP-NEXT: vmovdqa64 %zmm6, 384(%rax) ; AVX512-FCP-NEXT: vmovdqa64 %zmm3, 64(%rax) ; AVX512-FCP-NEXT: addq $392, %rsp # imm = 0x188 ; AVX512-FCP-NEXT: vzeroupper ; AVX512-FCP-NEXT: retq ; ; AVX512DQ-LABEL: store_i8_stride8_vf64: ; AVX512DQ: # %bb.0: ; AVX512DQ-NEXT: subq $552, %rsp # imm = 0x228 ; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX512DQ-NEXT: vmovdqa (%rcx), %xmm3 ; AVX512DQ-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX512DQ-NEXT: vmovdqa 32(%rcx), %xmm12 ; AVX512DQ-NEXT: vmovdqa 48(%rcx), %xmm0 ; AVX512DQ-NEXT: vmovdqa (%rdx), %xmm2 ; AVX512DQ-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX512DQ-NEXT: vmovdqa 48(%rdx), %xmm1 ; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm2 = 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: vpshuflw {{.*#+}} xmm3 = xmm2[0,0,2,1,4,5,6,7] ; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm4 = xmm2[0,2,2,3,4,5,6,7] ; AVX512DQ-NEXT: vinserti128 $1, %xmm4, %ymm3, %ymm10 ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm3 = xmm2[0,1,2,3,4,4,6,5] ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,6,6,7] ; AVX512DQ-NEXT: vinserti128 $1, %xmm2, %ymm3, %ymm2 ; AVX512DQ-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX512DQ-NEXT: vmovdqa (%r10), %xmm5 ; AVX512DQ-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX512DQ-NEXT: vmovdqa 48(%r10), %xmm3 ; AVX512DQ-NEXT: vmovdqa (%rax), %xmm2 ; AVX512DQ-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX512DQ-NEXT: vmovdqa 48(%rax), %xmm4 ; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm2[8],xmm5[8],xmm2[9],xmm5[9],xmm2[10],xmm5[10],xmm2[11],xmm5[11],xmm2[12],xmm5[12],xmm2[13],xmm5[13],xmm2[14],xmm5[14],xmm2[15],xmm5[15] ; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm5 = xmm2[0,0,2,1,4,5,6,7] ; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm6 = xmm2[0,2,2,3,4,5,6,7] ; AVX512DQ-NEXT: vinserti128 $1, %xmm6, %ymm5, %ymm5 ; AVX512DQ-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm5 = xmm2[0,1,2,3,4,4,6,5] ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,6,6,7] ; AVX512DQ-NEXT: vinserti128 $1, %xmm2, %ymm5, %ymm2 ; AVX512DQ-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX512DQ-NEXT: vmovdqa (%r9), %xmm5 ; AVX512DQ-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX512DQ-NEXT: vmovdqa 48(%r9), %xmm6 ; AVX512DQ-NEXT: vmovdqa (%r8), %xmm2 ; AVX512DQ-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX512DQ-NEXT: vmovdqa 48(%r8), %xmm8 ; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm2[8],xmm5[8],xmm2[9],xmm5[9],xmm2[10],xmm5[10],xmm2[11],xmm5[11],xmm2[12],xmm5[12],xmm2[13],xmm5[13],xmm2[14],xmm5[14],xmm2[15],xmm5[15] ; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm5 = xmm2[0,1,1,3,4,5,6,7] ; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm7 = xmm2[2,1,3,3,4,5,6,7] ; AVX512DQ-NEXT: vinserti128 $1, %xmm7, %ymm5, %ymm5 ; AVX512DQ-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm5 = xmm2[0,1,2,3,4,5,5,7] ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,6,5,7,7] ; AVX512DQ-NEXT: vinserti128 $1, %xmm2, %ymm5, %ymm2 ; AVX512DQ-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3],xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7] ; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm5 = xmm2[0,0,2,1,4,5,6,7] ; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm9 = xmm2[0,2,2,3,4,5,6,7] ; AVX512DQ-NEXT: vinserti128 $1, %xmm9, %ymm5, %ymm5 ; AVX512DQ-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm5 = xmm2[0,1,2,3,4,4,6,5] ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,6,6,7] ; AVX512DQ-NEXT: vinserti128 $1, %xmm2, %ymm5, %ymm2 ; AVX512DQ-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3],xmm4[4],xmm3[4],xmm4[5],xmm3[5],xmm4[6],xmm3[6],xmm4[7],xmm3[7] ; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm5 = xmm2[0,0,2,1,4,5,6,7] ; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm9 = xmm2[0,2,2,3,4,5,6,7] ; AVX512DQ-NEXT: vinserti128 $1, %xmm9, %ymm5, %ymm5 ; AVX512DQ-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm5 = xmm2[0,1,2,3,4,4,6,5] ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,6,6,7] ; AVX512DQ-NEXT: vinserti128 $1, %xmm2, %ymm5, %ymm2 ; AVX512DQ-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm2 = 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-NEXT: vpshuflw {{.*#+}} xmm5 = xmm2[0,1,1,3,4,5,6,7] ; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm9 = xmm2[2,1,3,3,4,5,6,7] ; AVX512DQ-NEXT: vinserti128 $1, %xmm9, %ymm5, %ymm5 ; AVX512DQ-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX512DQ-NEXT: vmovdqa 32(%rdx), %xmm5 ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm9 = xmm2[0,1,2,3,4,5,5,7] ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,6,5,7,7] ; AVX512DQ-NEXT: vinserti32x4 $1, %xmm2, %ymm9, %ymm26 ; AVX512DQ-NEXT: vmovdqa 32(%r10), %xmm2 ; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} 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] ; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm0 = xmm1[0,0,2,1,4,5,6,7] ; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm9 = xmm1[0,2,2,3,4,5,6,7] ; AVX512DQ-NEXT: vinserti32x4 $1, %xmm9, %ymm0, %ymm28 ; AVX512DQ-NEXT: vmovdqa 32(%rax), %xmm0 ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm9 = xmm1[0,1,2,3,4,4,6,5] ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,6,6,7] ; AVX512DQ-NEXT: vinserti32x4 $1, %xmm1, %ymm9, %ymm29 ; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm4[8],xmm3[8],xmm4[9],xmm3[9],xmm4[10],xmm3[10],xmm4[11],xmm3[11],xmm4[12],xmm3[12],xmm4[13],xmm3[13],xmm4[14],xmm3[14],xmm4[15],xmm3[15] ; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm3 = xmm1[0,0,2,1,4,5,6,7] ; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm4 = xmm1[0,2,2,3,4,5,6,7] ; AVX512DQ-NEXT: vinserti32x4 $1, %xmm4, %ymm3, %ymm25 ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm3 = xmm1[0,1,2,3,4,4,6,5] ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,6,6,7] ; AVX512DQ-NEXT: vinserti32x4 $1, %xmm1, %ymm3, %ymm23 ; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm1 = 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-NEXT: vpshuflw {{.*#+}} xmm3 = xmm1[0,1,1,3,4,5,6,7] ; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm4 = xmm1[2,1,3,3,4,5,6,7] ; AVX512DQ-NEXT: vinserti32x4 $1, %xmm4, %ymm3, %ymm19 ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm3 = xmm1[0,1,2,3,4,5,5,7] ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,6,5,7,7] ; AVX512DQ-NEXT: vinserti32x4 $1, %xmm1, %ymm3, %ymm17 ; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm5[0],xmm12[0],xmm5[1],xmm12[1],xmm5[2],xmm12[2],xmm5[3],xmm12[3],xmm5[4],xmm12[4],xmm5[5],xmm12[5],xmm5[6],xmm12[6],xmm5[7],xmm12[7] ; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm3 = xmm1[0,0,2,1,4,5,6,7] ; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm4 = xmm1[0,2,2,3,4,5,6,7] ; AVX512DQ-NEXT: vinserti32x4 $1, %xmm4, %ymm3, %ymm16 ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm3 = xmm1[0,1,2,3,4,4,6,5] ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,6,6,7] ; AVX512DQ-NEXT: vinserti128 $1, %xmm1, %ymm3, %ymm14 ; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],xmm2[3],xmm0[4],xmm2[4],xmm0[5],xmm2[5],xmm0[6],xmm2[6],xmm0[7],xmm2[7] ; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm3 = xmm1[0,0,2,1,4,5,6,7] ; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm6 = xmm1[0,2,2,3,4,5,6,7] ; AVX512DQ-NEXT: vinserti32x4 $1, %xmm6, %ymm3, %ymm20 ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm3 = xmm1[0,1,2,3,4,4,6,5] ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,6,6,7] ; AVX512DQ-NEXT: vinserti32x4 $1, %xmm1, %ymm3, %ymm18 ; AVX512DQ-NEXT: vmovdqa 32(%r9), %xmm1 ; AVX512DQ-NEXT: vmovdqa 32(%r8), %xmm3 ; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm6 = 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] ; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm8 = xmm6[0,1,1,3,4,5,6,7] ; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm9 = xmm6[2,1,3,3,4,5,6,7] ; AVX512DQ-NEXT: vinserti128 $1, %xmm9, %ymm8, %ymm15 ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm8 = xmm6[0,1,2,3,4,5,5,7] ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm6 = xmm6[0,1,2,3,6,5,7,7] ; AVX512DQ-NEXT: vinserti128 $1, %xmm6, %ymm8, %ymm13 ; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm5 = xmm5[8],xmm12[8],xmm5[9],xmm12[9],xmm5[10],xmm12[10],xmm5[11],xmm12[11],xmm5[12],xmm12[12],xmm5[13],xmm12[13],xmm5[14],xmm12[14],xmm5[15],xmm12[15] ; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm6 = xmm5[0,0,2,1,4,5,6,7] ; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm8 = xmm5[0,2,2,3,4,5,6,7] ; AVX512DQ-NEXT: vinserti128 $1, %xmm8, %ymm6, %ymm4 ; AVX512DQ-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm6 = xmm5[0,1,2,3,4,4,6,5] ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,4,6,6,7] ; AVX512DQ-NEXT: vinserti128 $1, %xmm5, %ymm6, %ymm4 ; AVX512DQ-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm0[8],xmm2[8],xmm0[9],xmm2[9],xmm0[10],xmm2[10],xmm0[11],xmm2[11],xmm0[12],xmm2[12],xmm0[13],xmm2[13],xmm0[14],xmm2[14],xmm0[15],xmm2[15] ; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm2 = xmm0[0,0,2,1,4,5,6,7] ; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm5 = xmm0[0,2,2,3,4,5,6,7] ; AVX512DQ-NEXT: vinserti128 $1, %xmm5, %ymm2, %ymm2 ; AVX512DQ-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm2 = xmm0[0,1,2,3,4,4,6,5] ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,6,6,7] ; AVX512DQ-NEXT: vinserti128 $1, %xmm0, %ymm2, %ymm0 ; AVX512DQ-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm0 = 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] ; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm1 = xmm0[0,1,1,3,4,5,6,7] ; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm2 = xmm0[2,1,3,3,4,5,6,7] ; AVX512DQ-NEXT: vinserti128 $1, %xmm2, %ymm1, %ymm1 ; AVX512DQ-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm1 = xmm0[0,1,2,3,4,5,5,7] ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,6,5,7,7] ; AVX512DQ-NEXT: vinserti128 $1, %xmm0, %ymm1, %ymm0 ; AVX512DQ-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX512DQ-NEXT: vmovdqa 16(%rcx), %xmm4 ; AVX512DQ-NEXT: vmovdqa 16(%rdx), %xmm3 ; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm0 = 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] ; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm1 = xmm0[0,0,2,1,4,5,6,7] ; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm5 = xmm0[0,2,2,3,4,5,6,7] ; AVX512DQ-NEXT: vinserti128 $1, %xmm5, %ymm1, %ymm1 ; AVX512DQ-NEXT: vmovdqu %ymm1, (%rsp) # 32-byte Spill ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm1 = xmm0[0,1,2,3,4,4,6,5] ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,6,6,7] ; AVX512DQ-NEXT: vinserti128 $1, %xmm0, %ymm1, %ymm0 ; AVX512DQ-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill ; AVX512DQ-NEXT: vmovdqa 16(%r10), %xmm9 ; AVX512DQ-NEXT: vmovdqa 16(%rax), %xmm8 ; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm8[0],xmm9[0],xmm8[1],xmm9[1],xmm8[2],xmm9[2],xmm8[3],xmm9[3],xmm8[4],xmm9[4],xmm8[5],xmm9[5],xmm8[6],xmm9[6],xmm8[7],xmm9[7] ; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm1 = xmm0[0,0,2,1,4,5,6,7] ; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm5 = xmm0[0,2,2,3,4,5,6,7] ; AVX512DQ-NEXT: vinserti32x4 $1, %xmm5, %ymm1, %ymm30 ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm1 = xmm0[0,1,2,3,4,4,6,5] ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,6,6,7] ; AVX512DQ-NEXT: vinserti32x4 $1, %xmm0, %ymm1, %ymm31 ; AVX512DQ-NEXT: vmovdqa 16(%r9), %xmm6 ; AVX512DQ-NEXT: vmovdqa 16(%r8), %xmm5 ; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm2 = 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] ; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm1 = xmm2[0,1,1,3,4,5,6,7] ; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm11 = xmm2[2,1,3,3,4,5,6,7] ; AVX512DQ-NEXT: vinserti32x4 $1, %xmm11, %ymm1, %ymm27 ; AVX512DQ-NEXT: vmovdqa 48(%rsi), %xmm1 ; AVX512DQ-NEXT: vmovdqa 48(%rdi), %xmm11 ; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm11[0],xmm1[0],xmm11[1],xmm1[1],xmm11[2],xmm1[2],xmm11[3],xmm1[3],xmm11[4],xmm1[4],xmm11[5],xmm1[5],xmm11[6],xmm1[6],xmm11[7],xmm1[7] ; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm11[8],xmm1[8],xmm11[9],xmm1[9],xmm11[10],xmm1[10],xmm11[11],xmm1[11],xmm11[12],xmm1[12],xmm11[13],xmm1[13],xmm11[14],xmm1[14],xmm11[15],xmm1[15] ; AVX512DQ-NEXT: vmovdqa (%rsi), %xmm7 ; AVX512DQ-NEXT: vmovdqa (%rdi), %xmm12 ; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm11 = xmm12[8],xmm7[8],xmm12[9],xmm7[9],xmm12[10],xmm7[10],xmm12[11],xmm7[11],xmm12[12],xmm7[12],xmm12[13],xmm7[13],xmm12[14],xmm7[14],xmm12[15],xmm7[15] ; AVX512DQ-NEXT: vmovdqa64 %xmm7, %xmm21 ; AVX512DQ-NEXT: vpmovzxwq {{.*#+}} zmm22 = xmm11[0],zero,zero,zero,xmm11[1],zero,zero,zero,xmm11[2],zero,zero,zero,xmm11[3],zero,zero,zero,xmm11[4],zero,zero,zero,xmm11[5],zero,zero,zero,xmm11[6],zero,zero,zero,xmm11[7],zero,zero,zero ; AVX512DQ-NEXT: vpmovzxwq {{.*#+}} zmm24 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero,xmm0[4],zero,zero,zero,xmm0[5],zero,zero,zero,xmm0[6],zero,zero,zero,xmm0[7],zero,zero,zero ; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm0 = ymm10[0,1,1,3,4,5,5,7] ; AVX512DQ-NEXT: vpshufd $246, {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Folded Reload ; AVX512DQ-NEXT: # ymm11 = mem[2,1,3,3,6,5,7,7] ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm11, %zmm0, %zmm0 ; AVX512DQ-NEXT: vpbroadcastq {{.*#+}} zmm11 = [65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535] ; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm22 = zmm22 | (zmm0 & ~zmm11) ; AVX512DQ-NEXT: vpshufd $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Folded Reload ; AVX512DQ-NEXT: # ymm0 = mem[0,0,2,1,4,4,6,5] ; AVX512DQ-NEXT: vpshufd $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Folded Reload ; AVX512DQ-NEXT: # ymm7 = mem[0,2,2,3,4,6,6,7] ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm7, %zmm0, %zmm0 ; AVX512DQ-NEXT: vpshufd $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Folded Reload ; AVX512DQ-NEXT: # ymm7 = mem[0,0,2,1,4,4,6,5] ; AVX512DQ-NEXT: vpshufd $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Folded Reload ; AVX512DQ-NEXT: # ymm10 = mem[0,2,2,3,4,6,6,7] ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm10, %zmm7, %zmm10 ; AVX512DQ-NEXT: vpbroadcastq {{.*#+}} zmm7 = [65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0] ; AVX512DQ-NEXT: vpandnq %zmm0, %zmm7, %zmm0 ; AVX512DQ-NEXT: vpandq %zmm7, %zmm10, %zmm10 ; AVX512DQ-NEXT: movw $-21846, %ax # imm = 0xAAAA ; AVX512DQ-NEXT: kmovw %eax, %k1 ; AVX512DQ-NEXT: vpord %zmm0, %zmm10, %zmm22 {%k1} ; AVX512DQ-NEXT: vpshufd $212, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Folded Reload ; AVX512DQ-NEXT: # ymm0 = mem[0,1,1,3,4,5,5,7] ; AVX512DQ-NEXT: vpshufd $246, {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Folded Reload ; AVX512DQ-NEXT: # ymm10 = mem[2,1,3,3,6,5,7,7] ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm10, %zmm0, %zmm0 ; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm24 = zmm24 | (zmm0 & ~zmm11) ; AVX512DQ-NEXT: vpshufd $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Folded Reload ; AVX512DQ-NEXT: # ymm0 = mem[0,0,2,1,4,4,6,5] ; AVX512DQ-NEXT: vpshufd $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Folded Reload ; AVX512DQ-NEXT: # ymm10 = mem[0,2,2,3,4,6,6,7] ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm10, %zmm0, %zmm0 ; AVX512DQ-NEXT: vpshufd $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Folded Reload ; AVX512DQ-NEXT: # ymm10 = mem[0,0,2,1,4,4,6,5] ; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm26 = ymm26[0,2,2,3,4,6,6,7] ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm26, %zmm10, %zmm10 ; AVX512DQ-NEXT: vpandnq %zmm0, %zmm7, %zmm0 ; AVX512DQ-NEXT: vpandq %zmm7, %zmm10, %zmm10 ; AVX512DQ-NEXT: vpord %zmm0, %zmm10, %zmm24 {%k1} ; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm0 = ymm28[0,1,1,3,4,5,5,7] ; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm10 = ymm29[2,1,3,3,6,5,7,7] ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm10, %zmm0, %zmm0 ; AVX512DQ-NEXT: vpmovzxwq {{.*#+}} zmm10 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero,xmm1[2],zero,zero,zero,xmm1[3],zero,zero,zero,xmm1[4],zero,zero,zero,xmm1[5],zero,zero,zero,xmm1[6],zero,zero,zero,xmm1[7],zero,zero,zero ; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm10 = zmm10 | (zmm0 & ~zmm11) ; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm0 = ymm25[0,0,2,1,4,4,6,5] ; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm1 = ymm23[0,2,2,3,4,6,6,7] ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0 ; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm1 = ymm19[0,0,2,1,4,4,6,5] ; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm17 = ymm17[0,2,2,3,4,6,6,7] ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm17, %zmm1, %zmm1 ; AVX512DQ-NEXT: vpandnq %zmm0, %zmm7, %zmm0 ; AVX512DQ-NEXT: vpandq %zmm7, %zmm1, %zmm1 ; AVX512DQ-NEXT: vpord %zmm0, %zmm1, %zmm10 {%k1} ; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm0 = ymm16[0,1,1,3,4,5,5,7] ; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm1 = ymm14[2,1,3,3,6,5,7,7] ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm1 ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm0 = xmm2[0,1,2,3,4,5,5,7] ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,6,5,7,7] ; AVX512DQ-NEXT: vinserti32x4 $1, %xmm2, %ymm0, %ymm17 ; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm2 = 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] ; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm0 = xmm2[0,0,2,1,4,5,6,7] ; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm3 = xmm2[0,2,2,3,4,5,6,7] ; AVX512DQ-NEXT: vinserti32x4 $1, %xmm3, %ymm0, %ymm19 ; AVX512DQ-NEXT: vmovdqa 32(%rsi), %xmm14 ; AVX512DQ-NEXT: vmovdqa 32(%rdi), %xmm0 ; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm0[0],xmm14[0],xmm0[1],xmm14[1],xmm0[2],xmm14[2],xmm0[3],xmm14[3],xmm0[4],xmm14[4],xmm0[5],xmm14[5],xmm0[6],xmm14[6],xmm0[7],xmm14[7] ; AVX512DQ-NEXT: vpmovzxwq {{.*#+}} zmm4 = xmm3[0],zero,zero,zero,xmm3[1],zero,zero,zero,xmm3[2],zero,zero,zero,xmm3[3],zero,zero,zero,xmm3[4],zero,zero,zero,xmm3[5],zero,zero,zero,xmm3[6],zero,zero,zero,xmm3[7],zero,zero,zero ; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm4 = zmm4 | (zmm1 & ~zmm11) ; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm1 = ymm20[0,0,2,1,4,4,6,5] ; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm3 = ymm18[0,2,2,3,4,6,6,7] ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm3, %zmm1, %zmm1 ; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm3 = ymm15[0,0,2,1,4,4,6,5] ; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm13 = ymm13[0,2,2,3,4,6,6,7] ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm13, %zmm3, %zmm3 ; AVX512DQ-NEXT: vpandnq %zmm1, %zmm7, %zmm1 ; AVX512DQ-NEXT: vpandq %zmm7, %zmm3, %zmm3 ; AVX512DQ-NEXT: vpord %zmm1, %zmm3, %zmm4 {%k1} ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm1 = xmm2[0,1,2,3,4,4,6,5] ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,6,6,7] ; AVX512DQ-NEXT: vinserti32x4 $1, %xmm2, %ymm1, %ymm20 ; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm8[8],xmm9[8],xmm8[9],xmm9[9],xmm8[10],xmm9[10],xmm8[11],xmm9[11],xmm8[12],xmm9[12],xmm8[13],xmm9[13],xmm8[14],xmm9[14],xmm8[15],xmm9[15] ; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm3 = xmm2[0,0,2,1,4,5,6,7] ; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm8 = xmm2[0,2,2,3,4,5,6,7] ; AVX512DQ-NEXT: vinserti32x4 $1, %xmm8, %ymm3, %ymm23 ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm3 = xmm2[0,1,2,3,4,4,6,5] ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,4,6,6,7] ; AVX512DQ-NEXT: vinserti128 $1, %xmm2, %ymm3, %ymm9 ; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm3 = 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] ; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm2 = xmm3[0,1,1,3,4,5,6,7] ; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm5 = xmm3[2,1,3,3,4,5,6,7] ; AVX512DQ-NEXT: vinserti128 $1, %xmm5, %ymm2, %ymm2 ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm5 = xmm3[0,1,2,3,4,5,5,7] ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,6,5,7,7] ; AVX512DQ-NEXT: vinserti128 $1, %xmm3, %ymm5, %ymm3 ; AVX512DQ-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload ; AVX512DQ-NEXT: vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm5 # 16-byte Folded Reload ; AVX512DQ-NEXT: # xmm5 = xmm1[0],mem[0],xmm1[1],mem[1],xmm1[2],mem[2],xmm1[3],mem[3],xmm1[4],mem[4],xmm1[5],mem[5],xmm1[6],mem[6],xmm1[7],mem[7] ; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm6 = xmm5[0,0,2,1,4,5,6,7] ; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm13 = xmm5[0,2,2,3,4,5,6,7] ; AVX512DQ-NEXT: vinserti128 $1, %xmm13, %ymm6, %ymm13 ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm6 = xmm5[0,1,2,3,4,4,6,5] ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,4,6,6,7] ; AVX512DQ-NEXT: vinserti128 $1, %xmm5, %ymm6, %ymm15 ; AVX512DQ-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload ; AVX512DQ-NEXT: vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm5 # 16-byte Folded Reload ; AVX512DQ-NEXT: # xmm5 = xmm1[0],mem[0],xmm1[1],mem[1],xmm1[2],mem[2],xmm1[3],mem[3],xmm1[4],mem[4],xmm1[5],mem[5],xmm1[6],mem[6],xmm1[7],mem[7] ; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm6 = xmm5[0,0,2,1,4,5,6,7] ; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm1 = xmm5[0,2,2,3,4,5,6,7] ; AVX512DQ-NEXT: vinserti32x4 $1, %xmm1, %ymm6, %ymm16 ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm1 = xmm5[0,1,2,3,4,4,6,5] ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,4,6,6,7] ; AVX512DQ-NEXT: vinserti32x4 $1, %xmm5, %ymm1, %ymm18 ; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm0[8],xmm14[8],xmm0[9],xmm14[9],xmm0[10],xmm14[10],xmm0[11],xmm14[11],xmm0[12],xmm14[12],xmm0[13],xmm14[13],xmm0[14],xmm14[14],xmm0[15],xmm14[15] ; AVX512DQ-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload ; AVX512DQ-NEXT: vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload ; AVX512DQ-NEXT: # xmm1 = xmm1[0],mem[0],xmm1[1],mem[1],xmm1[2],mem[2],xmm1[3],mem[3],xmm1[4],mem[4],xmm1[5],mem[5],xmm1[6],mem[6],xmm1[7],mem[7] ; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm5 = xmm1[0,1,1,3,4,5,6,7] ; AVX512DQ-NEXT: vpshuflw {{.*#+}} xmm6 = xmm1[2,1,3,3,4,5,6,7] ; AVX512DQ-NEXT: vinserti128 $1, %xmm6, %ymm5, %ymm5 ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm6 = xmm1[0,1,2,3,4,5,5,7] ; AVX512DQ-NEXT: vpshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,6,5,7,7] ; AVX512DQ-NEXT: vinserti128 $1, %xmm1, %ymm6, %ymm1 ; AVX512DQ-NEXT: vmovdqa 16(%rsi), %xmm6 ; AVX512DQ-NEXT: vmovdqa 16(%rdi), %xmm14 ; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm8 = xmm14[0],xmm6[0],xmm14[1],xmm6[1],xmm14[2],xmm6[2],xmm14[3],xmm6[3],xmm14[4],xmm6[4],xmm14[5],xmm6[5],xmm14[6],xmm6[6],xmm14[7],xmm6[7] ; AVX512DQ-NEXT: vpunpckhbw {{.*#+}} xmm6 = xmm14[8],xmm6[8],xmm14[9],xmm6[9],xmm14[10],xmm6[10],xmm14[11],xmm6[11],xmm14[12],xmm6[12],xmm14[13],xmm6[13],xmm14[14],xmm6[14],xmm14[15],xmm6[15] ; AVX512DQ-NEXT: vmovdqa64 %xmm21, %xmm14 ; AVX512DQ-NEXT: vpunpcklbw {{.*#+}} xmm12 = xmm12[0],xmm14[0],xmm12[1],xmm14[1],xmm12[2],xmm14[2],xmm12[3],xmm14[3],xmm12[4],xmm14[4],xmm12[5],xmm14[5],xmm12[6],xmm14[6],xmm12[7],xmm14[7] ; AVX512DQ-NEXT: vpshufd $212, {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Folded Reload ; AVX512DQ-NEXT: # ymm14 = mem[0,1,1,3,4,5,5,7] ; AVX512DQ-NEXT: vpshufd $246, {{[-0-9]+}}(%r{{[sb]}}p), %ymm25 # 32-byte Folded Reload ; AVX512DQ-NEXT: # ymm25 = mem[2,1,3,3,6,5,7,7] ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm25, %zmm14, %zmm14 ; AVX512DQ-NEXT: vpmovzxwq {{.*#+}} zmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero,xmm0[4],zero,zero,zero,xmm0[5],zero,zero,zero,xmm0[6],zero,zero,zero,xmm0[7],zero,zero,zero ; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 | (zmm14 & ~zmm11) ; AVX512DQ-NEXT: vpmovzxwq {{.*#+}} zmm8 = xmm8[0],zero,zero,zero,xmm8[1],zero,zero,zero,xmm8[2],zero,zero,zero,xmm8[3],zero,zero,zero,xmm8[4],zero,zero,zero,xmm8[5],zero,zero,zero,xmm8[6],zero,zero,zero,xmm8[7],zero,zero,zero ; AVX512DQ-NEXT: vpmovzxwq {{.*#+}} zmm6 = xmm6[0],zero,zero,zero,xmm6[1],zero,zero,zero,xmm6[2],zero,zero,zero,xmm6[3],zero,zero,zero,xmm6[4],zero,zero,zero,xmm6[5],zero,zero,zero,xmm6[6],zero,zero,zero,xmm6[7],zero,zero,zero ; AVX512DQ-NEXT: vpmovzxwq {{.*#+}} zmm12 = xmm12[0],zero,zero,zero,xmm12[1],zero,zero,zero,xmm12[2],zero,zero,zero,xmm12[3],zero,zero,zero,xmm12[4],zero,zero,zero,xmm12[5],zero,zero,zero,xmm12[6],zero,zero,zero,xmm12[7],zero,zero,zero ; AVX512DQ-NEXT: vpshufd $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Folded Reload ; AVX512DQ-NEXT: # ymm14 = mem[0,0,2,1,4,4,6,5] ; AVX512DQ-NEXT: vpshufd $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm25 # 32-byte Folded Reload ; AVX512DQ-NEXT: # ymm25 = mem[0,2,2,3,4,6,6,7] ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm25, %zmm14, %zmm14 ; AVX512DQ-NEXT: vpshufd $96, {{[-0-9]+}}(%r{{[sb]}}p), %ymm25 # 32-byte Folded Reload ; AVX512DQ-NEXT: # ymm25 = mem[0,0,2,1,4,4,6,5] ; AVX512DQ-NEXT: vpshufd $232, {{[-0-9]+}}(%r{{[sb]}}p), %ymm26 # 32-byte Folded Reload ; AVX512DQ-NEXT: # ymm26 = mem[0,2,2,3,4,6,6,7] ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm26, %zmm25, %zmm25 ; AVX512DQ-NEXT: vpandnq %zmm14, %zmm7, %zmm14 ; AVX512DQ-NEXT: vpandq %zmm7, %zmm25, %zmm25 ; AVX512DQ-NEXT: vpord %zmm14, %zmm25, %zmm0 {%k1} ; AVX512DQ-NEXT: vpshufd $212, (%rsp), %ymm14 # 32-byte Folded Reload ; AVX512DQ-NEXT: # ymm14 = mem[0,1,1,3,4,5,5,7] ; AVX512DQ-NEXT: vpshufd $246, {{[-0-9]+}}(%r{{[sb]}}p), %ymm25 # 32-byte Folded Reload ; AVX512DQ-NEXT: # ymm25 = mem[2,1,3,3,6,5,7,7] ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm25, %zmm14, %zmm14 ; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm8 = zmm8 | (zmm14 & ~zmm11) ; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm14 = ymm30[0,0,2,1,4,4,6,5] ; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm25 = ymm31[0,2,2,3,4,6,6,7] ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm25, %zmm14, %zmm14 ; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm21 = ymm27[0,0,2,1,4,4,6,5] ; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm17 = ymm17[0,2,2,3,4,6,6,7] ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm17, %zmm21, %zmm17 ; AVX512DQ-NEXT: vpandnq %zmm14, %zmm7, %zmm14 ; AVX512DQ-NEXT: vpandq %zmm7, %zmm17, %zmm17 ; AVX512DQ-NEXT: vpord %zmm14, %zmm17, %zmm8 {%k1} ; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm14 = ymm19[0,1,1,3,4,5,5,7] ; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm17 = ymm20[2,1,3,3,6,5,7,7] ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm17, %zmm14, %zmm14 ; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm6 = zmm6 | (zmm14 & ~zmm11) ; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm14 = ymm23[0,0,2,1,4,4,6,5] ; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm9 = ymm9[0,2,2,3,4,6,6,7] ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm9, %zmm14, %zmm9 ; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm2 = ymm2[0,0,2,1,4,4,6,5] ; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm3 = ymm3[0,2,2,3,4,6,6,7] ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm2 ; AVX512DQ-NEXT: vpandnq %zmm9, %zmm7, %zmm3 ; AVX512DQ-NEXT: vpandq %zmm7, %zmm2, %zmm2 ; AVX512DQ-NEXT: vpord %zmm3, %zmm2, %zmm6 {%k1} ; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm2 = ymm13[0,1,1,3,4,5,5,7] ; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm3 = ymm15[2,1,3,3,6,5,7,7] ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm2 ; AVX512DQ-NEXT: vpternlogq {{.*#+}} zmm12 = zmm12 | (zmm2 & ~zmm11) ; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm2 = ymm16[0,0,2,1,4,4,6,5] ; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm3 = ymm18[0,2,2,3,4,6,6,7] ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm2 ; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm3 = ymm5[0,0,2,1,4,4,6,5] ; AVX512DQ-NEXT: vpshufd {{.*#+}} ymm1 = ymm1[0,2,2,3,4,6,6,7] ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm3, %zmm1 ; AVX512DQ-NEXT: vpandnq %zmm2, %zmm7, %zmm2 ; AVX512DQ-NEXT: vpandq %zmm7, %zmm1, %zmm1 ; AVX512DQ-NEXT: vpord %zmm2, %zmm1, %zmm12 {%k1} ; AVX512DQ-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512DQ-NEXT: vmovdqa64 %zmm12, (%rax) ; AVX512DQ-NEXT: vmovdqa64 %zmm6, 192(%rax) ; AVX512DQ-NEXT: vmovdqa64 %zmm8, 128(%rax) ; AVX512DQ-NEXT: vmovdqa64 %zmm0, 320(%rax) ; AVX512DQ-NEXT: vmovdqa64 %zmm4, 256(%rax) ; AVX512DQ-NEXT: vmovdqa64 %zmm10, 448(%rax) ; AVX512DQ-NEXT: vmovdqa64 %zmm24, 384(%rax) ; AVX512DQ-NEXT: vmovdqa64 %zmm22, 64(%rax) ; AVX512DQ-NEXT: addq $552, %rsp # imm = 0x228 ; AVX512DQ-NEXT: vzeroupper ; AVX512DQ-NEXT: retq ; ; AVX512DQ-FCP-LABEL: store_i8_stride8_vf64: ; AVX512DQ-FCP: # %bb.0: ; AVX512DQ-FCP-NEXT: subq $392, %rsp # imm = 0x188 ; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX512DQ-FCP-NEXT: vmovdqa (%rcx), %xmm2 ; AVX512DQ-FCP-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX512DQ-FCP-NEXT: vmovdqa 48(%rcx), %xmm0 ; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %xmm15 ; AVX512DQ-FCP-NEXT: vmovdqa 48(%rdx), %xmm1 ; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm15[8],xmm2[8],xmm15[9],xmm2[9],xmm15[10],xmm2[10],xmm15[11],xmm2[11],xmm15[12],xmm2[12],xmm15[13],xmm2[13],xmm15[14],xmm2[14],xmm15[15],xmm2[15] ; AVX512DQ-FCP-NEXT: vinserti128 $1, %xmm2, %ymm2, %ymm2 ; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm6 = [8,9,8,9,4,5,6,7,12,13,10,11,12,13,10,11,8,9,12,13,4,5,6,7,12,13,14,15,12,13,14,15] ; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm2, %ymm3 ; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm8 = [0,1,0,1,4,5,2,3,4,5,2,3,12,13,14,15,0,1,4,5,4,5,6,7,4,5,6,7,12,13,14,15] ; AVX512DQ-FCP-NEXT: vpshufb %ymm8, %ymm2, %ymm2 ; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm2 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill ; AVX512DQ-FCP-NEXT: vmovdqa (%r10), %xmm2 ; AVX512DQ-FCP-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX512DQ-FCP-NEXT: vmovdqa (%rax), %xmm14 ; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm14[8],xmm2[8],xmm14[9],xmm2[9],xmm14[10],xmm2[10],xmm14[11],xmm2[11],xmm14[12],xmm2[12],xmm14[13],xmm2[13],xmm14[14],xmm2[14],xmm14[15],xmm2[15] ; AVX512DQ-FCP-NEXT: vinserti128 $1, %xmm2, %ymm2, %ymm2 ; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm11 = [0,1,2,3,8,9,8,9,8,9,8,9,12,13,10,11,0,1,2,3,8,9,12,13,8,9,12,13,12,13,14,15] ; AVX512DQ-FCP-NEXT: vpshufb %ymm11, %ymm2, %ymm3 ; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm10 = [0,1,0,1,0,1,0,1,8,9,10,11,4,5,2,3,0,1,4,5,0,1,4,5,8,9,10,11,4,5,6,7] ; AVX512DQ-FCP-NEXT: vpshufb %ymm10, %ymm2, %ymm2 ; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm19 ; AVX512DQ-FCP-NEXT: vmovdqa (%r9), %xmm3 ; AVX512DQ-FCP-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX512DQ-FCP-NEXT: vmovdqa (%r8), %xmm2 ; AVX512DQ-FCP-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill ; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm2 = 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-FCP-NEXT: vinserti128 $1, %xmm2, %ymm2, %ymm2 ; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm9 = [0,1,2,3,8,9,10,11,8,9,10,11,10,11,14,15,0,1,2,3,12,13,10,11,12,13,10,11,14,15,14,15] ; AVX512DQ-FCP-NEXT: vpshufb %ymm9, %ymm2, %ymm3 ; AVX512DQ-FCP-NEXT: vmovdqa {{.*#+}} ymm12 = [0,1,2,3,0,1,2,3,8,9,10,11,2,3,6,7,4,5,2,3,4,5,2,3,8,9,10,11,6,7,6,7] ; AVX512DQ-FCP-NEXT: vpshufb %ymm12, %ymm2, %ymm2 ; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm21 ; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3],xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7] ; AVX512DQ-FCP-NEXT: vinserti128 $1, %xmm2, %ymm2, %ymm2 ; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm2, %ymm3 ; AVX512DQ-FCP-NEXT: vpshufb %ymm8, %ymm2, %ymm2 ; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm2 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill ; AVX512DQ-FCP-NEXT: vmovdqa 48(%r10), %xmm2 ; AVX512DQ-FCP-NEXT: vmovdqa 48(%rax), %xmm3 ; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm4 = 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] ; AVX512DQ-FCP-NEXT: vinserti128 $1, %xmm4, %ymm4, %ymm4 ; AVX512DQ-FCP-NEXT: vpshufb %ymm11, %ymm4, %ymm5 ; AVX512DQ-FCP-NEXT: vpshufb %ymm10, %ymm4, %ymm4 ; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm5, %zmm4, %zmm4 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill ; AVX512DQ-FCP-NEXT: vmovdqa 48(%r9), %xmm4 ; AVX512DQ-FCP-NEXT: vmovdqa 48(%r8), %xmm5 ; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm7 = 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] ; AVX512DQ-FCP-NEXT: vinserti128 $1, %xmm7, %ymm7, %ymm7 ; AVX512DQ-FCP-NEXT: vpshufb %ymm9, %ymm7, %ymm13 ; AVX512DQ-FCP-NEXT: vpshufb %ymm12, %ymm7, %ymm7 ; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm13, %zmm7, %zmm7 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill ; AVX512DQ-FCP-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-FCP-NEXT: vinserti128 $1, %xmm0, %ymm0, %ymm0 ; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm0, %ymm1 ; AVX512DQ-FCP-NEXT: vpshufb %ymm8, %ymm0, %ymm0 ; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill ; 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: vinserti128 $1, %xmm0, %ymm0, %ymm0 ; AVX512DQ-FCP-NEXT: vpshufb %ymm11, %ymm0, %ymm1 ; AVX512DQ-FCP-NEXT: vpshufb %ymm10, %ymm0, %ymm0 ; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, (%rsp) # 64-byte Spill ; AVX512DQ-FCP-NEXT: vmovdqa 32(%rcx), %xmm0 ; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdx), %xmm1 ; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm5[8],xmm4[8],xmm5[9],xmm4[9],xmm5[10],xmm4[10],xmm5[11],xmm4[11],xmm5[12],xmm4[12],xmm5[13],xmm4[13],xmm5[14],xmm4[14],xmm5[15],xmm4[15] ; AVX512DQ-FCP-NEXT: vinserti128 $1, %xmm2, %ymm2, %ymm2 ; AVX512DQ-FCP-NEXT: vpshufb %ymm9, %ymm2, %ymm3 ; AVX512DQ-FCP-NEXT: vpshufb %ymm12, %ymm2, %ymm2 ; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm2 ; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill ; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3],xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7] ; AVX512DQ-FCP-NEXT: vinserti128 $1, %xmm2, %ymm2, %ymm2 ; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm2, %ymm3 ; AVX512DQ-FCP-NEXT: vpshufb %ymm8, %ymm2, %ymm2 ; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm20 ; AVX512DQ-FCP-NEXT: vmovdqa 32(%r10), %xmm2 ; AVX512DQ-FCP-NEXT: vmovdqa 32(%rax), %xmm3 ; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm4 = 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] ; AVX512DQ-FCP-NEXT: vinserti128 $1, %xmm4, %ymm4, %ymm4 ; AVX512DQ-FCP-NEXT: vpshufb %ymm11, %ymm4, %ymm5 ; AVX512DQ-FCP-NEXT: vpshufb %ymm10, %ymm4, %ymm4 ; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm5, %zmm4, %zmm22 ; AVX512DQ-FCP-NEXT: vmovdqa 32(%r9), %xmm4 ; AVX512DQ-FCP-NEXT: vmovdqa 32(%r8), %xmm5 ; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm7 = 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] ; AVX512DQ-FCP-NEXT: vinserti128 $1, %xmm7, %ymm7, %ymm7 ; AVX512DQ-FCP-NEXT: vpshufb %ymm9, %ymm7, %ymm13 ; AVX512DQ-FCP-NEXT: vpshufb %ymm12, %ymm7, %ymm7 ; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm13, %zmm7, %zmm23 ; AVX512DQ-FCP-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-FCP-NEXT: vinserti128 $1, %xmm0, %ymm0, %ymm0 ; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm0, %ymm1 ; AVX512DQ-FCP-NEXT: vpshufb %ymm8, %ymm0, %ymm0 ; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm24 ; 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: vinserti128 $1, %xmm0, %ymm0, %ymm0 ; AVX512DQ-FCP-NEXT: vpshufb %ymm11, %ymm0, %ymm1 ; AVX512DQ-FCP-NEXT: vpshufb %ymm10, %ymm0, %ymm0 ; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm25 ; AVX512DQ-FCP-NEXT: vmovdqa 16(%rcx), %xmm0 ; AVX512DQ-FCP-NEXT: vmovdqa 16(%rdx), %xmm1 ; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm5[8],xmm4[8],xmm5[9],xmm4[9],xmm5[10],xmm4[10],xmm5[11],xmm4[11],xmm5[12],xmm4[12],xmm5[13],xmm4[13],xmm5[14],xmm4[14],xmm5[15],xmm4[15] ; AVX512DQ-FCP-NEXT: vinserti128 $1, %xmm2, %ymm2, %ymm2 ; AVX512DQ-FCP-NEXT: vpshufb %ymm9, %ymm2, %ymm3 ; AVX512DQ-FCP-NEXT: vpshufb %ymm12, %ymm2, %ymm2 ; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm26 ; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm2 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3],xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7] ; AVX512DQ-FCP-NEXT: vinserti128 $1, %xmm2, %ymm2, %ymm2 ; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm2, %ymm3 ; AVX512DQ-FCP-NEXT: vpshufb %ymm8, %ymm2, %ymm2 ; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm3, %zmm2, %zmm27 ; AVX512DQ-FCP-NEXT: vmovdqa 16(%r10), %xmm2 ; AVX512DQ-FCP-NEXT: vmovdqa 16(%rax), %xmm3 ; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm4 = 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] ; AVX512DQ-FCP-NEXT: vinserti128 $1, %xmm4, %ymm4, %ymm4 ; AVX512DQ-FCP-NEXT: vpshufb %ymm11, %ymm4, %ymm5 ; AVX512DQ-FCP-NEXT: vpshufb %ymm10, %ymm4, %ymm4 ; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm5, %zmm4, %zmm28 ; AVX512DQ-FCP-NEXT: vmovdqa 16(%r9), %xmm4 ; AVX512DQ-FCP-NEXT: vmovdqa 16(%r8), %xmm5 ; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm7 = 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] ; AVX512DQ-FCP-NEXT: vinserti128 $1, %xmm7, %ymm7, %ymm7 ; AVX512DQ-FCP-NEXT: vpshufb %ymm9, %ymm7, %ymm13 ; AVX512DQ-FCP-NEXT: vpshufb %ymm12, %ymm7, %ymm7 ; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm13, %zmm7, %zmm29 ; AVX512DQ-FCP-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-FCP-NEXT: vinserti128 $1, %xmm0, %ymm0, %ymm0 ; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm0, %ymm1 ; AVX512DQ-FCP-NEXT: vpshufb %ymm8, %ymm0, %ymm0 ; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm30 ; 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: vinserti128 $1, %xmm0, %ymm0, %ymm0 ; AVX512DQ-FCP-NEXT: vpshufb %ymm11, %ymm0, %ymm1 ; AVX512DQ-FCP-NEXT: vpshufb %ymm10, %ymm0, %ymm0 ; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm31 ; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm0 = 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-FCP-NEXT: vinserti128 $1, %xmm0, %ymm0, %ymm0 ; AVX512DQ-FCP-NEXT: vpshufb %ymm9, %ymm0, %ymm1 ; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm9, %ymm17 ; AVX512DQ-FCP-NEXT: vpshufb %ymm12, %ymm0, %ymm0 ; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm12, %ymm18 ; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm16 ; AVX512DQ-FCP-NEXT: vmovdqa (%rsi), %xmm12 ; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %xmm9 ; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm9[8],xmm12[8],xmm9[9],xmm12[9],xmm9[10],xmm12[10],xmm9[11],xmm12[11],xmm9[12],xmm12[12],xmm9[13],xmm12[13],xmm9[14],xmm12[14],xmm9[15],xmm12[15] ; AVX512DQ-FCP-NEXT: vpmovsxwq {{.*#+}} xmm5 = [1284,1798] ; AVX512DQ-FCP-NEXT: vpmovzxwq {{.*#+}} xmm1 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero ; AVX512DQ-FCP-NEXT: vpshufb %xmm5, %xmm0, %xmm2 ; AVX512DQ-FCP-NEXT: vinserti128 $1, %xmm2, %ymm1, %ymm1 ; AVX512DQ-FCP-NEXT: vinserti128 $1, %xmm0, %ymm0, %ymm0 ; AVX512DQ-FCP-NEXT: vpmovsxwq {{.*#+}} ymm4 = [2312,2826,3340,3854] ; AVX512DQ-FCP-NEXT: vpshufb %ymm4, %ymm0, %ymm0 ; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm3 ; AVX512DQ-FCP-NEXT: vpbroadcastq {{.*#+}} zmm7 = [65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535] ; AVX512DQ-FCP-NEXT: vpternlogq $226, {{[-0-9]+}}(%r{{[sb]}}p), %zmm7, %zmm3 # 64-byte Folded Reload ; AVX512DQ-FCP-NEXT: # zmm3 = mem ^ (zmm7 & (zmm3 ^ mem)) ; AVX512DQ-FCP-NEXT: vpbroadcastq {{.*#+}} zmm2 = [65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0,65535,65535,65535,0] ; AVX512DQ-FCP-NEXT: vpandnq %zmm19, %zmm2, %zmm19 ; AVX512DQ-FCP-NEXT: vpandq %zmm2, %zmm21, %zmm21 ; AVX512DQ-FCP-NEXT: movw $-21846, %ax # imm = 0xAAAA ; AVX512DQ-FCP-NEXT: kmovw %eax, %k1 ; AVX512DQ-FCP-NEXT: vpord %zmm19, %zmm21, %zmm3 {%k1} ; AVX512DQ-FCP-NEXT: vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload ; AVX512DQ-FCP-NEXT: # xmm15 = xmm15[0],mem[0],xmm15[1],mem[1],xmm15[2],mem[2],xmm15[3],mem[3],xmm15[4],mem[4],xmm15[5],mem[5],xmm15[6],mem[6],xmm15[7],mem[7] ; AVX512DQ-FCP-NEXT: vinserti128 $1, %xmm15, %ymm15, %ymm15 ; AVX512DQ-FCP-NEXT: vpshufb %ymm6, %ymm15, %ymm6 ; AVX512DQ-FCP-NEXT: vpshufb %ymm8, %ymm15, %ymm8 ; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm6, %zmm8, %zmm19 ; AVX512DQ-FCP-NEXT: vmovdqa 48(%rsi), %xmm15 ; AVX512DQ-FCP-NEXT: vmovdqa 48(%rdi), %xmm8 ; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm6 = xmm8[0],xmm15[0],xmm8[1],xmm15[1],xmm8[2],xmm15[2],xmm8[3],xmm15[3],xmm8[4],xmm15[4],xmm8[5],xmm15[5],xmm8[6],xmm15[6],xmm8[7],xmm15[7] ; AVX512DQ-FCP-NEXT: vpshufb %xmm5, %xmm6, %xmm0 ; AVX512DQ-FCP-NEXT: vpmovzxwq {{.*#+}} xmm21 = xmm6[0],zero,zero,zero,xmm6[1],zero,zero,zero ; AVX512DQ-FCP-NEXT: vinserti32x4 $1, %xmm0, %ymm21, %ymm0 ; AVX512DQ-FCP-NEXT: vinserti128 $1, %xmm6, %ymm6, %ymm6 ; AVX512DQ-FCP-NEXT: vpshufb %ymm4, %ymm6, %ymm6 ; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm6, %zmm0, %zmm6 ; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm8[8],xmm15[8],xmm8[9],xmm15[9],xmm8[10],xmm15[10],xmm8[11],xmm15[11],xmm8[12],xmm15[12],xmm8[13],xmm15[13],xmm8[14],xmm15[14],xmm8[15],xmm15[15] ; AVX512DQ-FCP-NEXT: vpshufb %xmm5, %xmm0, %xmm8 ; AVX512DQ-FCP-NEXT: vpmovzxwq {{.*#+}} xmm15 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero ; AVX512DQ-FCP-NEXT: vinserti128 $1, %xmm8, %ymm15, %ymm8 ; AVX512DQ-FCP-NEXT: vinserti128 $1, %xmm0, %ymm0, %ymm0 ; AVX512DQ-FCP-NEXT: vpshufb %ymm4, %ymm0, %ymm0 ; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm8, %zmm15 ; AVX512DQ-FCP-NEXT: vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm0 # 16-byte Folded Reload ; AVX512DQ-FCP-NEXT: # xmm0 = xmm14[0],mem[0],xmm14[1],mem[1],xmm14[2],mem[2],xmm14[3],mem[3],xmm14[4],mem[4],xmm14[5],mem[5],xmm14[6],mem[6],xmm14[7],mem[7] ; AVX512DQ-FCP-NEXT: vinserti128 $1, %xmm0, %ymm0, %ymm0 ; AVX512DQ-FCP-NEXT: vpshufb %ymm11, %ymm0, %ymm8 ; AVX512DQ-FCP-NEXT: vpshufb %ymm10, %ymm0, %ymm0 ; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm8, %zmm0, %zmm13 ; AVX512DQ-FCP-NEXT: vmovdqa 32(%rsi), %xmm0 ; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdi), %xmm8 ; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm11 = 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] ; AVX512DQ-FCP-NEXT: vpshufb %xmm5, %xmm11, %xmm14 ; AVX512DQ-FCP-NEXT: vpmovzxwq {{.*#+}} xmm21 = xmm11[0],zero,zero,zero,xmm11[1],zero,zero,zero ; AVX512DQ-FCP-NEXT: vinserti32x4 $1, %xmm14, %ymm21, %ymm14 ; AVX512DQ-FCP-NEXT: vinserti128 $1, %xmm11, %ymm11, %ymm11 ; AVX512DQ-FCP-NEXT: vpshufb %ymm4, %ymm11, %ymm11 ; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm11, %zmm14, %zmm11 ; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm8[8],xmm0[8],xmm8[9],xmm0[9],xmm8[10],xmm0[10],xmm8[11],xmm0[11],xmm8[12],xmm0[12],xmm8[13],xmm0[13],xmm8[14],xmm0[14],xmm8[15],xmm0[15] ; AVX512DQ-FCP-NEXT: vpshufb %xmm5, %xmm0, %xmm8 ; AVX512DQ-FCP-NEXT: vpmovzxwq {{.*#+}} xmm14 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero ; AVX512DQ-FCP-NEXT: vinserti128 $1, %xmm8, %ymm14, %ymm8 ; AVX512DQ-FCP-NEXT: vinserti128 $1, %xmm0, %ymm0, %ymm0 ; AVX512DQ-FCP-NEXT: vpshufb %ymm4, %ymm0, %ymm0 ; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm8, %zmm14 ; AVX512DQ-FCP-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload ; AVX512DQ-FCP-NEXT: vpunpcklbw {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload ; AVX512DQ-FCP-NEXT: # 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: vinserti128 $1, %xmm0, %ymm0, %ymm0 ; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm17, %ymm1 ; AVX512DQ-FCP-NEXT: vpshufb %ymm1, %ymm0, %ymm8 ; AVX512DQ-FCP-NEXT: vmovdqa64 %ymm18, %ymm1 ; AVX512DQ-FCP-NEXT: vpshufb %ymm1, %ymm0, %ymm0 ; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm8, %zmm0, %zmm21 ; AVX512DQ-FCP-NEXT: vmovdqa 16(%rsi), %xmm8 ; AVX512DQ-FCP-NEXT: vmovdqa 16(%rdi), %xmm10 ; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm0 = xmm10[0],xmm8[0],xmm10[1],xmm8[1],xmm10[2],xmm8[2],xmm10[3],xmm8[3],xmm10[4],xmm8[4],xmm10[5],xmm8[5],xmm10[6],xmm8[6],xmm10[7],xmm8[7] ; AVX512DQ-FCP-NEXT: vpshufb %xmm5, %xmm0, %xmm1 ; AVX512DQ-FCP-NEXT: vpmovzxwq {{.*#+}} xmm17 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero ; AVX512DQ-FCP-NEXT: vinserti32x4 $1, %xmm1, %ymm17, %ymm1 ; AVX512DQ-FCP-NEXT: vinserti128 $1, %xmm0, %ymm0, %ymm0 ; AVX512DQ-FCP-NEXT: vpshufb %ymm4, %ymm0, %ymm0 ; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0 ; AVX512DQ-FCP-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm10[8],xmm8[8],xmm10[9],xmm8[9],xmm10[10],xmm8[10],xmm10[11],xmm8[11],xmm10[12],xmm8[12],xmm10[13],xmm8[13],xmm10[14],xmm8[14],xmm10[15],xmm8[15] ; AVX512DQ-FCP-NEXT: vpshufb %xmm5, %xmm1, %xmm8 ; AVX512DQ-FCP-NEXT: vpmovzxwq {{.*#+}} xmm10 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero ; AVX512DQ-FCP-NEXT: vinserti128 $1, %xmm8, %ymm10, %ymm8 ; AVX512DQ-FCP-NEXT: vinserti128 $1, %xmm1, %ymm1, %ymm1 ; AVX512DQ-FCP-NEXT: vpshufb %ymm4, %ymm1, %ymm1 ; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm8, %zmm1 ; AVX512DQ-FCP-NEXT: vpunpcklbw {{.*#+}} xmm8 = xmm9[0],xmm12[0],xmm9[1],xmm12[1],xmm9[2],xmm12[2],xmm9[3],xmm12[3],xmm9[4],xmm12[4],xmm9[5],xmm12[5],xmm9[6],xmm12[6],xmm9[7],xmm12[7] ; AVX512DQ-FCP-NEXT: vinserti128 $1, %xmm8, %ymm8, %ymm9 ; AVX512DQ-FCP-NEXT: vpshufb %ymm4, %ymm9, %ymm4 ; AVX512DQ-FCP-NEXT: vpshufb %xmm5, %xmm8, %xmm5 ; AVX512DQ-FCP-NEXT: vpmovzxwq {{.*#+}} xmm8 = xmm8[0],zero,zero,zero,xmm8[1],zero,zero,zero ; AVX512DQ-FCP-NEXT: vinserti128 $1, %xmm5, %ymm8, %ymm5 ; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm4, %zmm5, %zmm4 ; AVX512DQ-FCP-NEXT: vpternlogq $226, {{[-0-9]+}}(%r{{[sb]}}p), %zmm7, %zmm6 # 64-byte Folded Reload ; AVX512DQ-FCP-NEXT: # zmm6 = mem ^ (zmm7 & (zmm6 ^ mem)) ; AVX512DQ-FCP-NEXT: vpandnq {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm5 # 64-byte Folded Reload ; AVX512DQ-FCP-NEXT: vpandq {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm8 # 64-byte Folded Reload ; AVX512DQ-FCP-NEXT: vpord %zmm5, %zmm8, %zmm6 {%k1} ; AVX512DQ-FCP-NEXT: vpternlogq $226, {{[-0-9]+}}(%r{{[sb]}}p), %zmm7, %zmm15 # 64-byte Folded Reload ; AVX512DQ-FCP-NEXT: # zmm15 = mem ^ (zmm7 & (zmm15 ^ mem)) ; AVX512DQ-FCP-NEXT: vpandnq (%rsp), %zmm2, %zmm5 # 64-byte Folded Reload ; AVX512DQ-FCP-NEXT: vpandq {{[-0-9]+}}(%r{{[sb]}}p), %zmm2, %zmm8 # 64-byte Folded Reload ; AVX512DQ-FCP-NEXT: vpord %zmm5, %zmm8, %zmm15 {%k1} ; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm11 = zmm20 ^ (zmm7 & (zmm11 ^ zmm20)) ; AVX512DQ-FCP-NEXT: vpandnq %zmm22, %zmm2, %zmm5 ; AVX512DQ-FCP-NEXT: vpandq %zmm2, %zmm23, %zmm8 ; AVX512DQ-FCP-NEXT: vpord %zmm5, %zmm8, %zmm11 {%k1} ; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm14 = zmm24 ^ (zmm7 & (zmm14 ^ zmm24)) ; AVX512DQ-FCP-NEXT: vpandnq %zmm25, %zmm2, %zmm5 ; AVX512DQ-FCP-NEXT: vpandq %zmm2, %zmm26, %zmm8 ; AVX512DQ-FCP-NEXT: vpord %zmm5, %zmm8, %zmm14 {%k1} ; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm0 = zmm27 ^ (zmm7 & (zmm0 ^ zmm27)) ; AVX512DQ-FCP-NEXT: vpandnq %zmm28, %zmm2, %zmm5 ; AVX512DQ-FCP-NEXT: vpandq %zmm2, %zmm29, %zmm8 ; AVX512DQ-FCP-NEXT: vpord %zmm5, %zmm8, %zmm0 {%k1} ; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm1 = zmm30 ^ (zmm7 & (zmm1 ^ zmm30)) ; AVX512DQ-FCP-NEXT: vpandnq %zmm31, %zmm2, %zmm5 ; AVX512DQ-FCP-NEXT: vpandq %zmm2, %zmm16, %zmm8 ; AVX512DQ-FCP-NEXT: vpord %zmm5, %zmm8, %zmm1 {%k1} ; AVX512DQ-FCP-NEXT: vpternlogq {{.*#+}} zmm4 = zmm19 ^ (zmm7 & (zmm4 ^ zmm19)) ; AVX512DQ-FCP-NEXT: vpandnq %zmm13, %zmm2, %zmm5 ; AVX512DQ-FCP-NEXT: vpandq %zmm2, %zmm21, %zmm2 ; AVX512DQ-FCP-NEXT: vpord %zmm5, %zmm2, %zmm4 {%k1} ; AVX512DQ-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm4, (%rax) ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, 192(%rax) ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, 128(%rax) ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm14, 320(%rax) ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm11, 256(%rax) ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm15, 448(%rax) ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, 384(%rax) ; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, 64(%rax) ; AVX512DQ-FCP-NEXT: addq $392, %rsp # imm = 0x188 ; AVX512DQ-FCP-NEXT: vzeroupper ; AVX512DQ-FCP-NEXT: retq ; ; AVX512BW-LABEL: store_i8_stride8_vf64: ; AVX512BW: # %bb.0: ; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX512BW-NEXT: vmovdqa (%r10), %xmm1 ; AVX512BW-NEXT: vmovdqa64 32(%r10), %xmm16 ; AVX512BW-NEXT: vmovdqa 48(%r10), %xmm14 ; AVX512BW-NEXT: vmovdqa (%rax), %xmm3 ; AVX512BW-NEXT: vmovdqa64 32(%rax), %xmm17 ; AVX512BW-NEXT: vmovdqa 48(%rax), %xmm15 ; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm3[8],xmm1[8],xmm3[9],xmm1[9],xmm3[10],xmm1[10],xmm3[11],xmm1[11],xmm3[12],xmm1[12],xmm3[13],xmm1[13],xmm3[14],xmm1[14],xmm3[15],xmm1[15] ; AVX512BW-NEXT: vmovdqa (%r9), %xmm4 ; AVX512BW-NEXT: vmovdqa64 32(%r9), %xmm19 ; AVX512BW-NEXT: vmovdqa64 48(%r9), %xmm18 ; AVX512BW-NEXT: vmovdqa (%r8), %xmm6 ; AVX512BW-NEXT: vmovdqa64 32(%r8), %xmm21 ; AVX512BW-NEXT: vmovdqa64 48(%r8), %xmm20 ; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm0 = 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: vpmovsxbw {{.*#+}} zmm7 = [0,1,0,32,4,5,1,33,2,1,2,34,4,5,3,35,0,1,4,36,4,5,5,37,0,1,6,38,6,5,7,39] ; AVX512BW-NEXT: vpermt2w %zmm2, %zmm7, %zmm0 ; AVX512BW-NEXT: vmovdqa (%rcx), %xmm5 ; AVX512BW-NEXT: vmovdqa64 48(%rcx), %xmm22 ; AVX512BW-NEXT: vmovdqa (%rdx), %xmm8 ; AVX512BW-NEXT: vmovdqa64 48(%rdx), %xmm23 ; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm11 = xmm8[8],xmm5[8],xmm8[9],xmm5[9],xmm8[10],xmm5[10],xmm8[11],xmm5[11],xmm8[12],xmm5[12],xmm8[13],xmm5[13],xmm8[14],xmm5[14],xmm8[15],xmm5[15] ; AVX512BW-NEXT: vmovdqa (%rsi), %xmm9 ; AVX512BW-NEXT: vmovdqa64 48(%rsi), %xmm24 ; AVX512BW-NEXT: vmovdqa (%rdi), %xmm10 ; AVX512BW-NEXT: vmovdqa64 48(%rdi), %xmm25 ; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm2 = 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-NEXT: vpmovsxbw {{.*#+}} zmm12 = [0,32,0,0,1,33,0,0,2,34,0,0,3,35,0,0,4,36,0,0,5,37,0,0,6,38,0,0,7,39,0,0] ; AVX512BW-NEXT: vpermt2w %zmm11, %zmm12, %zmm2 ; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm11 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3],xmm15[4],xmm14[4],xmm15[5],xmm14[5],xmm15[6],xmm14[6],xmm15[7],xmm14[7] ; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm13 = xmm20[0],xmm18[0],xmm20[1],xmm18[1],xmm20[2],xmm18[2],xmm20[3],xmm18[3],xmm20[4],xmm18[4],xmm20[5],xmm18[5],xmm20[6],xmm18[6],xmm20[7],xmm18[7] ; AVX512BW-NEXT: vpermt2w %zmm11, %zmm7, %zmm13 ; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm26 = xmm23[0],xmm22[0],xmm23[1],xmm22[1],xmm23[2],xmm22[2],xmm23[3],xmm22[3],xmm23[4],xmm22[4],xmm23[5],xmm22[5],xmm23[6],xmm22[6],xmm23[7],xmm22[7] ; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm11 = xmm25[0],xmm24[0],xmm25[1],xmm24[1],xmm25[2],xmm24[2],xmm25[3],xmm24[3],xmm25[4],xmm24[4],xmm25[5],xmm24[5],xmm25[6],xmm24[6],xmm25[7],xmm24[7] ; AVX512BW-NEXT: vpermt2w %zmm26, %zmm12, %zmm11 ; AVX512BW-NEXT: vmovdqa64 32(%rcx), %xmm26 ; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm14 = xmm15[8],xmm14[8],xmm15[9],xmm14[9],xmm15[10],xmm14[10],xmm15[11],xmm14[11],xmm15[12],xmm14[12],xmm15[13],xmm14[13],xmm15[14],xmm14[14],xmm15[15],xmm14[15] ; AVX512BW-NEXT: vmovdqa64 32(%rdx), %xmm27 ; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm15 = xmm20[8],xmm18[8],xmm20[9],xmm18[9],xmm20[10],xmm18[10],xmm20[11],xmm18[11],xmm20[12],xmm18[12],xmm20[13],xmm18[13],xmm20[14],xmm18[14],xmm20[15],xmm18[15] ; AVX512BW-NEXT: vmovdqa64 32(%rsi), %xmm28 ; AVX512BW-NEXT: vpermt2w %zmm14, %zmm7, %zmm15 ; AVX512BW-NEXT: vmovdqa64 32(%rdi), %xmm29 ; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm18 = xmm23[8],xmm22[8],xmm23[9],xmm22[9],xmm23[10],xmm22[10],xmm23[11],xmm22[11],xmm23[12],xmm22[12],xmm23[13],xmm22[13],xmm23[14],xmm22[14],xmm23[15],xmm22[15] ; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm14 = xmm25[8],xmm24[8],xmm25[9],xmm24[9],xmm25[10],xmm24[10],xmm25[11],xmm24[11],xmm25[12],xmm24[12],xmm25[13],xmm24[13],xmm25[14],xmm24[14],xmm25[15],xmm24[15] ; AVX512BW-NEXT: vpermt2w %zmm18, %zmm12, %zmm14 ; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm18 = 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: vpunpcklbw {{.*#+}} xmm20 = xmm21[0],xmm19[0],xmm21[1],xmm19[1],xmm21[2],xmm19[2],xmm21[3],xmm19[3],xmm21[4],xmm19[4],xmm21[5],xmm19[5],xmm21[6],xmm19[6],xmm21[7],xmm19[7] ; AVX512BW-NEXT: vpermt2w %zmm18, %zmm7, %zmm20 ; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm22 = xmm27[0],xmm26[0],xmm27[1],xmm26[1],xmm27[2],xmm26[2],xmm27[3],xmm26[3],xmm27[4],xmm26[4],xmm27[5],xmm26[5],xmm27[6],xmm26[6],xmm27[7],xmm26[7] ; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm18 = xmm29[0],xmm28[0],xmm29[1],xmm28[1],xmm29[2],xmm28[2],xmm29[3],xmm28[3],xmm29[4],xmm28[4],xmm29[5],xmm28[5],xmm29[6],xmm28[6],xmm29[7],xmm28[7] ; AVX512BW-NEXT: vpermt2w %zmm22, %zmm12, %zmm18 ; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm17 = 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: vmovdqa64 16(%r10), %xmm22 ; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm16 = xmm21[8],xmm19[8],xmm21[9],xmm19[9],xmm21[10],xmm19[10],xmm21[11],xmm19[11],xmm21[12],xmm19[12],xmm21[13],xmm19[13],xmm21[14],xmm19[14],xmm21[15],xmm19[15] ; AVX512BW-NEXT: vmovdqa64 16(%rax), %xmm19 ; AVX512BW-NEXT: vpermt2w %zmm17, %zmm7, %zmm16 ; AVX512BW-NEXT: vmovdqa64 16(%r9), %xmm21 ; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm23 = xmm27[8],xmm26[8],xmm27[9],xmm26[9],xmm27[10],xmm26[10],xmm27[11],xmm26[11],xmm27[12],xmm26[12],xmm27[13],xmm26[13],xmm27[14],xmm26[14],xmm27[15],xmm26[15] ; AVX512BW-NEXT: vmovdqa64 16(%r8), %xmm24 ; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm17 = xmm29[8],xmm28[8],xmm29[9],xmm28[9],xmm29[10],xmm28[10],xmm29[11],xmm28[11],xmm29[12],xmm28[12],xmm29[13],xmm28[13],xmm29[14],xmm28[14],xmm29[15],xmm28[15] ; AVX512BW-NEXT: vpermt2w %zmm23, %zmm12, %zmm17 ; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm23 = xmm19[0],xmm22[0],xmm19[1],xmm22[1],xmm19[2],xmm22[2],xmm19[3],xmm22[3],xmm19[4],xmm22[4],xmm19[5],xmm22[5],xmm19[6],xmm22[6],xmm19[7],xmm22[7] ; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm25 = xmm24[0],xmm21[0],xmm24[1],xmm21[1],xmm24[2],xmm21[2],xmm24[3],xmm21[3],xmm24[4],xmm21[4],xmm24[5],xmm21[5],xmm24[6],xmm21[6],xmm24[7],xmm21[7] ; AVX512BW-NEXT: vpermt2w %zmm23, %zmm7, %zmm25 ; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm19 = xmm19[8],xmm22[8],xmm19[9],xmm22[9],xmm19[10],xmm22[10],xmm19[11],xmm22[11],xmm19[12],xmm22[12],xmm19[13],xmm22[13],xmm19[14],xmm22[14],xmm19[15],xmm22[15] ; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm21 = xmm24[8],xmm21[8],xmm24[9],xmm21[9],xmm24[10],xmm21[10],xmm24[11],xmm21[11],xmm24[12],xmm21[12],xmm24[13],xmm21[13],xmm24[14],xmm21[14],xmm24[15],xmm21[15] ; AVX512BW-NEXT: vmovdqa64 16(%rcx), %xmm22 ; AVX512BW-NEXT: vpermt2w %zmm19, %zmm7, %zmm21 ; AVX512BW-NEXT: vmovdqa64 16(%rdx), %xmm19 ; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3],xmm3[4],xmm1[4],xmm3[5],xmm1[5],xmm3[6],xmm1[6],xmm3[7],xmm1[7] ; AVX512BW-NEXT: vmovdqa 16(%rsi), %xmm3 ; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm4 = 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: vmovdqa 16(%rdi), %xmm6 ; AVX512BW-NEXT: vpermt2w %zmm1, %zmm7, %zmm4 ; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm19[0],xmm22[0],xmm19[1],xmm22[1],xmm19[2],xmm22[2],xmm19[3],xmm22[3],xmm19[4],xmm22[4],xmm19[5],xmm22[5],xmm19[6],xmm22[6],xmm19[7],xmm22[7] ; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm7 = xmm6[0],xmm3[0],xmm6[1],xmm3[1],xmm6[2],xmm3[2],xmm6[3],xmm3[3],xmm6[4],xmm3[4],xmm6[5],xmm3[5],xmm6[6],xmm3[6],xmm6[7],xmm3[7] ; AVX512BW-NEXT: vpermt2w %zmm1, %zmm12, %zmm7 ; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm19[8],xmm22[8],xmm19[9],xmm22[9],xmm19[10],xmm22[10],xmm19[11],xmm22[11],xmm19[12],xmm22[12],xmm19[13],xmm22[13],xmm19[14],xmm22[14],xmm19[15],xmm22[15] ; AVX512BW-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm6[8],xmm3[8],xmm6[9],xmm3[9],xmm6[10],xmm3[10],xmm6[11],xmm3[11],xmm6[12],xmm3[12],xmm6[13],xmm3[13],xmm6[14],xmm3[14],xmm6[15],xmm3[15] ; AVX512BW-NEXT: vpermt2w %zmm1, %zmm12, %zmm3 ; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm1 = 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] ; AVX512BW-NEXT: vpunpcklbw {{.*#+}} xmm5 = 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-NEXT: vpermt2w %zmm1, %zmm12, %zmm5 ; AVX512BW-NEXT: movw $-21846, %ax # imm = 0xAAAA ; AVX512BW-NEXT: kmovd %eax, %k1 ; AVX512BW-NEXT: vmovdqa32 %zmm0, %zmm2 {%k1} ; AVX512BW-NEXT: vmovdqa32 %zmm13, %zmm11 {%k1} ; AVX512BW-NEXT: vmovdqa32 %zmm15, %zmm14 {%k1} ; AVX512BW-NEXT: vmovdqa32 %zmm20, %zmm18 {%k1} ; AVX512BW-NEXT: vmovdqa32 %zmm16, %zmm17 {%k1} ; AVX512BW-NEXT: vmovdqa32 %zmm25, %zmm7 {%k1} ; AVX512BW-NEXT: vmovdqa32 %zmm21, %zmm3 {%k1} ; AVX512BW-NEXT: vmovdqa32 %zmm4, %zmm5 {%k1} ; AVX512BW-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512BW-NEXT: vmovdqa64 %zmm5, (%rax) ; AVX512BW-NEXT: vmovdqa64 %zmm3, 192(%rax) ; AVX512BW-NEXT: vmovdqa64 %zmm7, 128(%rax) ; AVX512BW-NEXT: vmovdqa64 %zmm17, 320(%rax) ; AVX512BW-NEXT: vmovdqa64 %zmm18, 256(%rax) ; AVX512BW-NEXT: vmovdqa64 %zmm14, 448(%rax) ; AVX512BW-NEXT: vmovdqa64 %zmm11, 384(%rax) ; AVX512BW-NEXT: vmovdqa64 %zmm2, 64(%rax) ; AVX512BW-NEXT: vzeroupper ; AVX512BW-NEXT: retq ; ; AVX512BW-FCP-LABEL: store_i8_stride8_vf64: ; AVX512BW-FCP: # %bb.0: ; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX512BW-FCP-NEXT: vmovdqa (%rsi), %xmm1 ; AVX512BW-FCP-NEXT: vmovdqa64 32(%rsi), %xmm19 ; AVX512BW-FCP-NEXT: vmovdqa64 48(%rsi), %xmm16 ; AVX512BW-FCP-NEXT: vmovdqa (%rdi), %xmm2 ; AVX512BW-FCP-NEXT: vmovdqa64 32(%rdi), %xmm20 ; AVX512BW-FCP-NEXT: vmovdqa64 48(%rdi), %xmm17 ; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm2[8],xmm1[8],xmm2[9],xmm1[9],xmm2[10],xmm1[10],xmm2[11],xmm1[11],xmm2[12],xmm1[12],xmm2[13],xmm1[13],xmm2[14],xmm1[14],xmm2[15],xmm1[15] ; AVX512BW-FCP-NEXT: vinserti128 $1, %xmm0, %ymm0, %ymm3 ; AVX512BW-FCP-NEXT: vpmovsxwq {{.*#+}} ymm5 = [2312,2826,3340,3854] ; AVX512BW-FCP-NEXT: vpshufb %ymm5, %ymm3, %ymm3 ; AVX512BW-FCP-NEXT: vpmovsxwq {{.*#+}} xmm6 = [1284,1798] ; AVX512BW-FCP-NEXT: vpshufb %xmm6, %xmm0, %xmm4 ; AVX512BW-FCP-NEXT: vpmovzxwq {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero ; AVX512BW-FCP-NEXT: vinserti128 $1, %xmm4, %ymm0, %ymm0 ; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm3, %zmm0, %zmm0 ; AVX512BW-FCP-NEXT: vmovdqa (%rcx), %xmm3 ; AVX512BW-FCP-NEXT: vmovdqa64 32(%rcx), %xmm21 ; AVX512BW-FCP-NEXT: vmovdqa64 48(%rcx), %xmm18 ; AVX512BW-FCP-NEXT: vmovdqa (%rdx), %xmm7 ; AVX512BW-FCP-NEXT: vmovdqa64 32(%rdx), %xmm22 ; AVX512BW-FCP-NEXT: vmovdqa64 48(%rdx), %xmm23 ; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm4 = 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] ; AVX512BW-FCP-NEXT: vinserti128 $1, %xmm4, %ymm4, %ymm4 ; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm9 = [0,32,2,3,4,33,6,7,8,42,10,11,12,43,14,15,16,36,18,19,20,37,22,23,24,46,26,27,28,47,30,31] ; AVX512BW-FCP-NEXT: vpermt2w %zmm4, %zmm9, %zmm0 ; AVX512BW-FCP-NEXT: vmovdqa (%r10), %xmm4 ; AVX512BW-FCP-NEXT: vmovdqa64 48(%r10), %xmm24 ; AVX512BW-FCP-NEXT: vmovdqa (%rax), %xmm8 ; AVX512BW-FCP-NEXT: vmovdqa64 48(%rax), %xmm26 ; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm10 = xmm8[8],xmm4[8],xmm8[9],xmm4[9],xmm8[10],xmm4[10],xmm8[11],xmm4[11],xmm8[12],xmm4[12],xmm8[13],xmm4[13],xmm8[14],xmm4[14],xmm8[15],xmm4[15] ; AVX512BW-FCP-NEXT: vinserti128 $1, %xmm10, %ymm10, %ymm13 ; AVX512BW-FCP-NEXT: vmovdqa (%r9), %xmm10 ; AVX512BW-FCP-NEXT: vmovdqa64 48(%r9), %xmm28 ; AVX512BW-FCP-NEXT: vmovdqa (%r8), %xmm12 ; AVX512BW-FCP-NEXT: vmovdqa64 48(%r8), %xmm30 ; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm11 = xmm12[8],xmm10[8],xmm12[9],xmm10[9],xmm12[10],xmm10[10],xmm12[11],xmm10[11],xmm12[12],xmm10[12],xmm12[13],xmm10[13],xmm12[14],xmm10[14],xmm12[15],xmm10[15] ; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm14 = [0,1,0,32,4,5,1,33,2,1,2,42,4,5,3,43,0,1,4,36,4,5,5,37,0,1,6,46,6,5,7,47] ; AVX512BW-FCP-NEXT: vpermt2w %zmm13, %zmm14, %zmm11 ; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm13 = 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-FCP-NEXT: vinserti128 $1, %xmm13, %ymm13, %ymm15 ; AVX512BW-FCP-NEXT: vpshufb %ymm5, %ymm15, %ymm15 ; AVX512BW-FCP-NEXT: vpshufb %xmm6, %xmm13, %xmm25 ; AVX512BW-FCP-NEXT: vpmovzxwq {{.*#+}} xmm13 = xmm13[0],zero,zero,zero,xmm13[1],zero,zero,zero ; AVX512BW-FCP-NEXT: vinserti32x4 $1, %xmm25, %ymm13, %ymm13 ; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm15, %zmm13, %zmm13 ; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm15 = xmm23[0],xmm18[0],xmm23[1],xmm18[1],xmm23[2],xmm18[2],xmm23[3],xmm18[3],xmm23[4],xmm18[4],xmm23[5],xmm18[5],xmm23[6],xmm18[6],xmm23[7],xmm18[7] ; AVX512BW-FCP-NEXT: vinserti128 $1, %xmm15, %ymm15, %ymm15 ; AVX512BW-FCP-NEXT: vpermt2w %zmm15, %zmm9, %zmm13 ; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm15 = xmm26[0],xmm24[0],xmm26[1],xmm24[1],xmm26[2],xmm24[2],xmm26[3],xmm24[3],xmm26[4],xmm24[4],xmm26[5],xmm24[5],xmm26[6],xmm24[6],xmm26[7],xmm24[7] ; AVX512BW-FCP-NEXT: vinserti32x4 $1, %xmm15, %ymm15, %ymm25 ; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm15 = xmm30[0],xmm28[0],xmm30[1],xmm28[1],xmm30[2],xmm28[2],xmm30[3],xmm28[3],xmm30[4],xmm28[4],xmm30[5],xmm28[5],xmm30[6],xmm28[6],xmm30[7],xmm28[7] ; AVX512BW-FCP-NEXT: vpermt2w %zmm25, %zmm14, %zmm15 ; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm16 = 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-FCP-NEXT: vpshufb %xmm6, %xmm16, %xmm17 ; AVX512BW-FCP-NEXT: vpmovzxwq {{.*#+}} xmm25 = xmm16[0],zero,zero,zero,xmm16[1],zero,zero,zero ; AVX512BW-FCP-NEXT: vinserti32x4 $1, %xmm17, %ymm25, %ymm17 ; AVX512BW-FCP-NEXT: vmovdqa64 32(%r10), %xmm25 ; AVX512BW-FCP-NEXT: vinserti32x4 $1, %xmm16, %ymm16, %ymm16 ; AVX512BW-FCP-NEXT: vpshufb %ymm5, %ymm16, %ymm16 ; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm16, %zmm17, %zmm16 ; AVX512BW-FCP-NEXT: vmovdqa64 32(%rax), %xmm29 ; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm17 = xmm23[8],xmm18[8],xmm23[9],xmm18[9],xmm23[10],xmm18[10],xmm23[11],xmm18[11],xmm23[12],xmm18[12],xmm23[13],xmm18[13],xmm23[14],xmm18[14],xmm23[15],xmm18[15] ; AVX512BW-FCP-NEXT: vmovdqa64 32(%r9), %xmm27 ; AVX512BW-FCP-NEXT: vinserti32x4 $1, %xmm17, %ymm17, %ymm17 ; AVX512BW-FCP-NEXT: vpermt2w %zmm17, %zmm9, %zmm16 ; AVX512BW-FCP-NEXT: vmovdqa64 32(%r8), %xmm31 ; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm17 = xmm26[8],xmm24[8],xmm26[9],xmm24[9],xmm26[10],xmm24[10],xmm26[11],xmm24[11],xmm26[12],xmm24[12],xmm26[13],xmm24[13],xmm26[14],xmm24[14],xmm26[15],xmm24[15] ; AVX512BW-FCP-NEXT: vinserti32x4 $1, %xmm17, %ymm17, %ymm18 ; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm17 = xmm30[8],xmm28[8],xmm30[9],xmm28[9],xmm30[10],xmm28[10],xmm30[11],xmm28[11],xmm30[12],xmm28[12],xmm30[13],xmm28[13],xmm30[14],xmm28[14],xmm30[15],xmm28[15] ; AVX512BW-FCP-NEXT: vpermt2w %zmm18, %zmm14, %zmm17 ; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm18 = xmm20[0],xmm19[0],xmm20[1],xmm19[1],xmm20[2],xmm19[2],xmm20[3],xmm19[3],xmm20[4],xmm19[4],xmm20[5],xmm19[5],xmm20[6],xmm19[6],xmm20[7],xmm19[7] ; AVX512BW-FCP-NEXT: vinserti32x4 $1, %xmm18, %ymm18, %ymm23 ; AVX512BW-FCP-NEXT: vpshufb %ymm5, %ymm23, %ymm23 ; AVX512BW-FCP-NEXT: vpshufb %xmm6, %xmm18, %xmm24 ; AVX512BW-FCP-NEXT: vpmovzxwq {{.*#+}} xmm18 = xmm18[0],zero,zero,zero,xmm18[1],zero,zero,zero ; AVX512BW-FCP-NEXT: vinserti32x4 $1, %xmm24, %ymm18, %ymm18 ; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm23, %zmm18, %zmm18 ; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm23 = xmm22[0],xmm21[0],xmm22[1],xmm21[1],xmm22[2],xmm21[2],xmm22[3],xmm21[3],xmm22[4],xmm21[4],xmm22[5],xmm21[5],xmm22[6],xmm21[6],xmm22[7],xmm21[7] ; AVX512BW-FCP-NEXT: vinserti32x4 $1, %xmm23, %ymm23, %ymm23 ; AVX512BW-FCP-NEXT: vpermt2w %zmm23, %zmm9, %zmm18 ; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm23 = xmm29[0],xmm25[0],xmm29[1],xmm25[1],xmm29[2],xmm25[2],xmm29[3],xmm25[3],xmm29[4],xmm25[4],xmm29[5],xmm25[5],xmm29[6],xmm25[6],xmm29[7],xmm25[7] ; AVX512BW-FCP-NEXT: vinserti32x4 $1, %xmm23, %ymm23, %ymm24 ; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm23 = xmm31[0],xmm27[0],xmm31[1],xmm27[1],xmm31[2],xmm27[2],xmm31[3],xmm27[3],xmm31[4],xmm27[4],xmm31[5],xmm27[5],xmm31[6],xmm27[6],xmm31[7],xmm27[7] ; AVX512BW-FCP-NEXT: vpermt2w %zmm24, %zmm14, %zmm23 ; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm19 = xmm20[8],xmm19[8],xmm20[9],xmm19[9],xmm20[10],xmm19[10],xmm20[11],xmm19[11],xmm20[12],xmm19[12],xmm20[13],xmm19[13],xmm20[14],xmm19[14],xmm20[15],xmm19[15] ; AVX512BW-FCP-NEXT: vpshufb %xmm6, %xmm19, %xmm20 ; AVX512BW-FCP-NEXT: vpmovzxwq {{.*#+}} xmm24 = xmm19[0],zero,zero,zero,xmm19[1],zero,zero,zero ; AVX512BW-FCP-NEXT: vinserti32x4 $1, %xmm20, %ymm24, %ymm20 ; AVX512BW-FCP-NEXT: vinserti32x4 $1, %xmm19, %ymm19, %ymm19 ; AVX512BW-FCP-NEXT: vpshufb %ymm5, %ymm19, %ymm19 ; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm19, %zmm20, %zmm19 ; AVX512BW-FCP-NEXT: vmovdqa64 16(%rsi), %xmm24 ; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm20 = xmm22[8],xmm21[8],xmm22[9],xmm21[9],xmm22[10],xmm21[10],xmm22[11],xmm21[11],xmm22[12],xmm21[12],xmm22[13],xmm21[13],xmm22[14],xmm21[14],xmm22[15],xmm21[15] ; AVX512BW-FCP-NEXT: vmovdqa64 16(%rdi), %xmm22 ; AVX512BW-FCP-NEXT: vinserti32x4 $1, %xmm20, %ymm20, %ymm20 ; AVX512BW-FCP-NEXT: vpermt2w %zmm20, %zmm9, %zmm19 ; AVX512BW-FCP-NEXT: vmovdqa64 16(%rcx), %xmm26 ; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm20 = xmm29[8],xmm25[8],xmm29[9],xmm25[9],xmm29[10],xmm25[10],xmm29[11],xmm25[11],xmm29[12],xmm25[12],xmm29[13],xmm25[13],xmm29[14],xmm25[14],xmm29[15],xmm25[15] ; AVX512BW-FCP-NEXT: vinserti32x4 $1, %xmm20, %ymm20, %ymm21 ; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm20 = xmm31[8],xmm27[8],xmm31[9],xmm27[9],xmm31[10],xmm27[10],xmm31[11],xmm27[11],xmm31[12],xmm27[12],xmm31[13],xmm27[13],xmm31[14],xmm27[14],xmm31[15],xmm27[15] ; AVX512BW-FCP-NEXT: vpermt2w %zmm21, %zmm14, %zmm20 ; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm21 = xmm22[0],xmm24[0],xmm22[1],xmm24[1],xmm22[2],xmm24[2],xmm22[3],xmm24[3],xmm22[4],xmm24[4],xmm22[5],xmm24[5],xmm22[6],xmm24[6],xmm22[7],xmm24[7] ; AVX512BW-FCP-NEXT: vpshufb %xmm6, %xmm21, %xmm25 ; AVX512BW-FCP-NEXT: vpmovzxwq {{.*#+}} xmm27 = xmm21[0],zero,zero,zero,xmm21[1],zero,zero,zero ; AVX512BW-FCP-NEXT: vinserti32x4 $1, %xmm25, %ymm27, %ymm25 ; AVX512BW-FCP-NEXT: vmovdqa64 16(%rdx), %xmm27 ; AVX512BW-FCP-NEXT: vinserti32x4 $1, %xmm21, %ymm21, %ymm21 ; AVX512BW-FCP-NEXT: vpshufb %ymm5, %ymm21, %ymm21 ; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm21, %zmm25, %zmm21 ; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm25 = xmm27[0],xmm26[0],xmm27[1],xmm26[1],xmm27[2],xmm26[2],xmm27[3],xmm26[3],xmm27[4],xmm26[4],xmm27[5],xmm26[5],xmm27[6],xmm26[6],xmm27[7],xmm26[7] ; AVX512BW-FCP-NEXT: vinserti32x4 $1, %xmm25, %ymm25, %ymm25 ; AVX512BW-FCP-NEXT: vpermt2w %zmm25, %zmm9, %zmm21 ; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm22 = xmm22[8],xmm24[8],xmm22[9],xmm24[9],xmm22[10],xmm24[10],xmm22[11],xmm24[11],xmm22[12],xmm24[12],xmm22[13],xmm24[13],xmm22[14],xmm24[14],xmm22[15],xmm24[15] ; AVX512BW-FCP-NEXT: vpshufb %xmm6, %xmm22, %xmm24 ; AVX512BW-FCP-NEXT: vpmovzxwq {{.*#+}} xmm25 = xmm22[0],zero,zero,zero,xmm22[1],zero,zero,zero ; AVX512BW-FCP-NEXT: vinserti32x4 $1, %xmm24, %ymm25, %ymm24 ; AVX512BW-FCP-NEXT: vinserti32x4 $1, %xmm22, %ymm22, %ymm22 ; AVX512BW-FCP-NEXT: vpshufb %ymm5, %ymm22, %ymm22 ; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm22, %zmm24, %zmm22 ; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm24 = xmm27[8],xmm26[8],xmm27[9],xmm26[9],xmm27[10],xmm26[10],xmm27[11],xmm26[11],xmm27[12],xmm26[12],xmm27[13],xmm26[13],xmm27[14],xmm26[14],xmm27[15],xmm26[15] ; AVX512BW-FCP-NEXT: vinserti32x4 $1, %xmm24, %ymm24, %ymm24 ; AVX512BW-FCP-NEXT: vpermt2w %zmm24, %zmm9, %zmm22 ; AVX512BW-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] ; AVX512BW-FCP-NEXT: vinserti128 $1, %xmm1, %ymm1, %ymm2 ; AVX512BW-FCP-NEXT: vpshufb %ymm5, %ymm2, %ymm2 ; AVX512BW-FCP-NEXT: vmovdqa 16(%r10), %xmm5 ; AVX512BW-FCP-NEXT: vpshufb %xmm6, %xmm1, %xmm6 ; AVX512BW-FCP-NEXT: vpmovzxwq {{.*#+}} xmm1 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero ; AVX512BW-FCP-NEXT: vinserti128 $1, %xmm6, %ymm1, %ymm1 ; AVX512BW-FCP-NEXT: vmovdqa 16(%rax), %xmm6 ; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm1 ; AVX512BW-FCP-NEXT: vmovdqa 16(%r9), %xmm2 ; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm7[0],xmm3[0],xmm7[1],xmm3[1],xmm7[2],xmm3[2],xmm7[3],xmm3[3],xmm7[4],xmm3[4],xmm7[5],xmm3[5],xmm7[6],xmm3[6],xmm7[7],xmm3[7] ; AVX512BW-FCP-NEXT: vmovdqa 16(%r8), %xmm7 ; AVX512BW-FCP-NEXT: vinserti128 $1, %xmm3, %ymm3, %ymm3 ; AVX512BW-FCP-NEXT: vpermt2w %zmm3, %zmm9, %zmm1 ; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3],xmm6[4],xmm5[4],xmm6[5],xmm5[5],xmm6[6],xmm5[6],xmm6[7],xmm5[7] ; AVX512BW-FCP-NEXT: vinserti128 $1, %xmm3, %ymm3, %ymm3 ; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm9 = 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] ; AVX512BW-FCP-NEXT: vpermt2w %zmm3, %zmm14, %zmm9 ; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm6[8],xmm5[8],xmm6[9],xmm5[9],xmm6[10],xmm5[10],xmm6[11],xmm5[11],xmm6[12],xmm5[12],xmm6[13],xmm5[13],xmm6[14],xmm5[14],xmm6[15],xmm5[15] ; AVX512BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm7[8],xmm2[8],xmm7[9],xmm2[9],xmm7[10],xmm2[10],xmm7[11],xmm2[11],xmm7[12],xmm2[12],xmm7[13],xmm2[13],xmm7[14],xmm2[14],xmm7[15],xmm2[15] ; AVX512BW-FCP-NEXT: vinserti128 $1, %xmm3, %ymm3, %ymm3 ; AVX512BW-FCP-NEXT: vpermt2w %zmm3, %zmm14, %zmm2 ; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm8[0],xmm4[0],xmm8[1],xmm4[1],xmm8[2],xmm4[2],xmm8[3],xmm4[3],xmm8[4],xmm4[4],xmm8[5],xmm4[5],xmm8[6],xmm4[6],xmm8[7],xmm4[7] ; AVX512BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm4 = xmm12[0],xmm10[0],xmm12[1],xmm10[1],xmm12[2],xmm10[2],xmm12[3],xmm10[3],xmm12[4],xmm10[4],xmm12[5],xmm10[5],xmm12[6],xmm10[6],xmm12[7],xmm10[7] ; AVX512BW-FCP-NEXT: vinserti128 $1, %xmm3, %ymm3, %ymm3 ; AVX512BW-FCP-NEXT: vpermt2w %zmm3, %zmm14, %zmm4 ; AVX512BW-FCP-NEXT: movw $-21846, %ax # imm = 0xAAAA ; AVX512BW-FCP-NEXT: kmovd %eax, %k1 ; AVX512BW-FCP-NEXT: vmovdqa32 %zmm11, %zmm0 {%k1} ; AVX512BW-FCP-NEXT: vmovdqa32 %zmm15, %zmm13 {%k1} ; AVX512BW-FCP-NEXT: vmovdqa32 %zmm17, %zmm16 {%k1} ; AVX512BW-FCP-NEXT: vmovdqa32 %zmm23, %zmm18 {%k1} ; AVX512BW-FCP-NEXT: vmovdqa32 %zmm20, %zmm19 {%k1} ; AVX512BW-FCP-NEXT: vmovdqa32 %zmm9, %zmm21 {%k1} ; AVX512BW-FCP-NEXT: vmovdqa32 %zmm2, %zmm22 {%k1} ; AVX512BW-FCP-NEXT: vmovdqa32 %zmm4, %zmm1 {%k1} ; AVX512BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm1, (%rax) ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm22, 192(%rax) ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm21, 128(%rax) ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm19, 320(%rax) ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm18, 256(%rax) ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm16, 448(%rax) ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm13, 384(%rax) ; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, 64(%rax) ; AVX512BW-FCP-NEXT: vzeroupper ; AVX512BW-FCP-NEXT: retq ; ; AVX512DQ-BW-LABEL: store_i8_stride8_vf64: ; AVX512DQ-BW: # %bb.0: ; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX512DQ-BW-NEXT: vmovdqa (%r10), %xmm1 ; AVX512DQ-BW-NEXT: vmovdqa64 32(%r10), %xmm16 ; AVX512DQ-BW-NEXT: vmovdqa 48(%r10), %xmm14 ; AVX512DQ-BW-NEXT: vmovdqa (%rax), %xmm3 ; AVX512DQ-BW-NEXT: vmovdqa64 32(%rax), %xmm17 ; AVX512DQ-BW-NEXT: vmovdqa 48(%rax), %xmm15 ; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm3[8],xmm1[8],xmm3[9],xmm1[9],xmm3[10],xmm1[10],xmm3[11],xmm1[11],xmm3[12],xmm1[12],xmm3[13],xmm1[13],xmm3[14],xmm1[14],xmm3[15],xmm1[15] ; AVX512DQ-BW-NEXT: vmovdqa (%r9), %xmm4 ; AVX512DQ-BW-NEXT: vmovdqa64 32(%r9), %xmm19 ; AVX512DQ-BW-NEXT: vmovdqa64 48(%r9), %xmm18 ; AVX512DQ-BW-NEXT: vmovdqa (%r8), %xmm6 ; AVX512DQ-BW-NEXT: vmovdqa64 32(%r8), %xmm21 ; AVX512DQ-BW-NEXT: vmovdqa64 48(%r8), %xmm20 ; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm0 = 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: vpmovsxbw {{.*#+}} zmm7 = [0,1,0,32,4,5,1,33,2,1,2,34,4,5,3,35,0,1,4,36,4,5,5,37,0,1,6,38,6,5,7,39] ; AVX512DQ-BW-NEXT: vpermt2w %zmm2, %zmm7, %zmm0 ; AVX512DQ-BW-NEXT: vmovdqa (%rcx), %xmm5 ; AVX512DQ-BW-NEXT: vmovdqa64 48(%rcx), %xmm22 ; AVX512DQ-BW-NEXT: vmovdqa (%rdx), %xmm8 ; AVX512DQ-BW-NEXT: vmovdqa64 48(%rdx), %xmm23 ; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm11 = xmm8[8],xmm5[8],xmm8[9],xmm5[9],xmm8[10],xmm5[10],xmm8[11],xmm5[11],xmm8[12],xmm5[12],xmm8[13],xmm5[13],xmm8[14],xmm5[14],xmm8[15],xmm5[15] ; AVX512DQ-BW-NEXT: vmovdqa (%rsi), %xmm9 ; AVX512DQ-BW-NEXT: vmovdqa64 48(%rsi), %xmm24 ; AVX512DQ-BW-NEXT: vmovdqa (%rdi), %xmm10 ; AVX512DQ-BW-NEXT: vmovdqa64 48(%rdi), %xmm25 ; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm2 = 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-NEXT: vpmovsxbw {{.*#+}} zmm12 = [0,32,0,0,1,33,0,0,2,34,0,0,3,35,0,0,4,36,0,0,5,37,0,0,6,38,0,0,7,39,0,0] ; AVX512DQ-BW-NEXT: vpermt2w %zmm11, %zmm12, %zmm2 ; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm11 = xmm15[0],xmm14[0],xmm15[1],xmm14[1],xmm15[2],xmm14[2],xmm15[3],xmm14[3],xmm15[4],xmm14[4],xmm15[5],xmm14[5],xmm15[6],xmm14[6],xmm15[7],xmm14[7] ; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm13 = xmm20[0],xmm18[0],xmm20[1],xmm18[1],xmm20[2],xmm18[2],xmm20[3],xmm18[3],xmm20[4],xmm18[4],xmm20[5],xmm18[5],xmm20[6],xmm18[6],xmm20[7],xmm18[7] ; AVX512DQ-BW-NEXT: vpermt2w %zmm11, %zmm7, %zmm13 ; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm26 = xmm23[0],xmm22[0],xmm23[1],xmm22[1],xmm23[2],xmm22[2],xmm23[3],xmm22[3],xmm23[4],xmm22[4],xmm23[5],xmm22[5],xmm23[6],xmm22[6],xmm23[7],xmm22[7] ; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm11 = xmm25[0],xmm24[0],xmm25[1],xmm24[1],xmm25[2],xmm24[2],xmm25[3],xmm24[3],xmm25[4],xmm24[4],xmm25[5],xmm24[5],xmm25[6],xmm24[6],xmm25[7],xmm24[7] ; AVX512DQ-BW-NEXT: vpermt2w %zmm26, %zmm12, %zmm11 ; AVX512DQ-BW-NEXT: vmovdqa64 32(%rcx), %xmm26 ; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm14 = xmm15[8],xmm14[8],xmm15[9],xmm14[9],xmm15[10],xmm14[10],xmm15[11],xmm14[11],xmm15[12],xmm14[12],xmm15[13],xmm14[13],xmm15[14],xmm14[14],xmm15[15],xmm14[15] ; AVX512DQ-BW-NEXT: vmovdqa64 32(%rdx), %xmm27 ; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm15 = xmm20[8],xmm18[8],xmm20[9],xmm18[9],xmm20[10],xmm18[10],xmm20[11],xmm18[11],xmm20[12],xmm18[12],xmm20[13],xmm18[13],xmm20[14],xmm18[14],xmm20[15],xmm18[15] ; AVX512DQ-BW-NEXT: vmovdqa64 32(%rsi), %xmm28 ; AVX512DQ-BW-NEXT: vpermt2w %zmm14, %zmm7, %zmm15 ; AVX512DQ-BW-NEXT: vmovdqa64 32(%rdi), %xmm29 ; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm18 = xmm23[8],xmm22[8],xmm23[9],xmm22[9],xmm23[10],xmm22[10],xmm23[11],xmm22[11],xmm23[12],xmm22[12],xmm23[13],xmm22[13],xmm23[14],xmm22[14],xmm23[15],xmm22[15] ; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm14 = xmm25[8],xmm24[8],xmm25[9],xmm24[9],xmm25[10],xmm24[10],xmm25[11],xmm24[11],xmm25[12],xmm24[12],xmm25[13],xmm24[13],xmm25[14],xmm24[14],xmm25[15],xmm24[15] ; AVX512DQ-BW-NEXT: vpermt2w %zmm18, %zmm12, %zmm14 ; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm18 = 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: vpunpcklbw {{.*#+}} xmm20 = xmm21[0],xmm19[0],xmm21[1],xmm19[1],xmm21[2],xmm19[2],xmm21[3],xmm19[3],xmm21[4],xmm19[4],xmm21[5],xmm19[5],xmm21[6],xmm19[6],xmm21[7],xmm19[7] ; AVX512DQ-BW-NEXT: vpermt2w %zmm18, %zmm7, %zmm20 ; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm22 = xmm27[0],xmm26[0],xmm27[1],xmm26[1],xmm27[2],xmm26[2],xmm27[3],xmm26[3],xmm27[4],xmm26[4],xmm27[5],xmm26[5],xmm27[6],xmm26[6],xmm27[7],xmm26[7] ; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm18 = xmm29[0],xmm28[0],xmm29[1],xmm28[1],xmm29[2],xmm28[2],xmm29[3],xmm28[3],xmm29[4],xmm28[4],xmm29[5],xmm28[5],xmm29[6],xmm28[6],xmm29[7],xmm28[7] ; AVX512DQ-BW-NEXT: vpermt2w %zmm22, %zmm12, %zmm18 ; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm17 = 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: vmovdqa64 16(%r10), %xmm22 ; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm16 = xmm21[8],xmm19[8],xmm21[9],xmm19[9],xmm21[10],xmm19[10],xmm21[11],xmm19[11],xmm21[12],xmm19[12],xmm21[13],xmm19[13],xmm21[14],xmm19[14],xmm21[15],xmm19[15] ; AVX512DQ-BW-NEXT: vmovdqa64 16(%rax), %xmm19 ; AVX512DQ-BW-NEXT: vpermt2w %zmm17, %zmm7, %zmm16 ; AVX512DQ-BW-NEXT: vmovdqa64 16(%r9), %xmm21 ; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm23 = xmm27[8],xmm26[8],xmm27[9],xmm26[9],xmm27[10],xmm26[10],xmm27[11],xmm26[11],xmm27[12],xmm26[12],xmm27[13],xmm26[13],xmm27[14],xmm26[14],xmm27[15],xmm26[15] ; AVX512DQ-BW-NEXT: vmovdqa64 16(%r8), %xmm24 ; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm17 = xmm29[8],xmm28[8],xmm29[9],xmm28[9],xmm29[10],xmm28[10],xmm29[11],xmm28[11],xmm29[12],xmm28[12],xmm29[13],xmm28[13],xmm29[14],xmm28[14],xmm29[15],xmm28[15] ; AVX512DQ-BW-NEXT: vpermt2w %zmm23, %zmm12, %zmm17 ; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm23 = xmm19[0],xmm22[0],xmm19[1],xmm22[1],xmm19[2],xmm22[2],xmm19[3],xmm22[3],xmm19[4],xmm22[4],xmm19[5],xmm22[5],xmm19[6],xmm22[6],xmm19[7],xmm22[7] ; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm25 = xmm24[0],xmm21[0],xmm24[1],xmm21[1],xmm24[2],xmm21[2],xmm24[3],xmm21[3],xmm24[4],xmm21[4],xmm24[5],xmm21[5],xmm24[6],xmm21[6],xmm24[7],xmm21[7] ; AVX512DQ-BW-NEXT: vpermt2w %zmm23, %zmm7, %zmm25 ; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm19 = xmm19[8],xmm22[8],xmm19[9],xmm22[9],xmm19[10],xmm22[10],xmm19[11],xmm22[11],xmm19[12],xmm22[12],xmm19[13],xmm22[13],xmm19[14],xmm22[14],xmm19[15],xmm22[15] ; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm21 = xmm24[8],xmm21[8],xmm24[9],xmm21[9],xmm24[10],xmm21[10],xmm24[11],xmm21[11],xmm24[12],xmm21[12],xmm24[13],xmm21[13],xmm24[14],xmm21[14],xmm24[15],xmm21[15] ; AVX512DQ-BW-NEXT: vmovdqa64 16(%rcx), %xmm22 ; AVX512DQ-BW-NEXT: vpermt2w %zmm19, %zmm7, %zmm21 ; AVX512DQ-BW-NEXT: vmovdqa64 16(%rdx), %xmm19 ; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm3[0],xmm1[0],xmm3[1],xmm1[1],xmm3[2],xmm1[2],xmm3[3],xmm1[3],xmm3[4],xmm1[4],xmm3[5],xmm1[5],xmm3[6],xmm1[6],xmm3[7],xmm1[7] ; AVX512DQ-BW-NEXT: vmovdqa 16(%rsi), %xmm3 ; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm4 = 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: vmovdqa 16(%rdi), %xmm6 ; AVX512DQ-BW-NEXT: vpermt2w %zmm1, %zmm7, %zmm4 ; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm1 = xmm19[0],xmm22[0],xmm19[1],xmm22[1],xmm19[2],xmm22[2],xmm19[3],xmm22[3],xmm19[4],xmm22[4],xmm19[5],xmm22[5],xmm19[6],xmm22[6],xmm19[7],xmm22[7] ; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm7 = xmm6[0],xmm3[0],xmm6[1],xmm3[1],xmm6[2],xmm3[2],xmm6[3],xmm3[3],xmm6[4],xmm3[4],xmm6[5],xmm3[5],xmm6[6],xmm3[6],xmm6[7],xmm3[7] ; AVX512DQ-BW-NEXT: vpermt2w %zmm1, %zmm12, %zmm7 ; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm19[8],xmm22[8],xmm19[9],xmm22[9],xmm19[10],xmm22[10],xmm19[11],xmm22[11],xmm19[12],xmm22[12],xmm19[13],xmm22[13],xmm19[14],xmm22[14],xmm19[15],xmm22[15] ; AVX512DQ-BW-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm6[8],xmm3[8],xmm6[9],xmm3[9],xmm6[10],xmm3[10],xmm6[11],xmm3[11],xmm6[12],xmm3[12],xmm6[13],xmm3[13],xmm6[14],xmm3[14],xmm6[15],xmm3[15] ; AVX512DQ-BW-NEXT: vpermt2w %zmm1, %zmm12, %zmm3 ; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm1 = 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] ; AVX512DQ-BW-NEXT: vpunpcklbw {{.*#+}} xmm5 = 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-NEXT: vpermt2w %zmm1, %zmm12, %zmm5 ; AVX512DQ-BW-NEXT: movw $-21846, %ax # imm = 0xAAAA ; AVX512DQ-BW-NEXT: kmovd %eax, %k1 ; AVX512DQ-BW-NEXT: vmovdqa32 %zmm0, %zmm2 {%k1} ; AVX512DQ-BW-NEXT: vmovdqa32 %zmm13, %zmm11 {%k1} ; AVX512DQ-BW-NEXT: vmovdqa32 %zmm15, %zmm14 {%k1} ; AVX512DQ-BW-NEXT: vmovdqa32 %zmm20, %zmm18 {%k1} ; AVX512DQ-BW-NEXT: vmovdqa32 %zmm16, %zmm17 {%k1} ; AVX512DQ-BW-NEXT: vmovdqa32 %zmm25, %zmm7 {%k1} ; AVX512DQ-BW-NEXT: vmovdqa32 %zmm21, %zmm3 {%k1} ; AVX512DQ-BW-NEXT: vmovdqa32 %zmm4, %zmm5 {%k1} ; AVX512DQ-BW-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm5, (%rax) ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, 192(%rax) ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm7, 128(%rax) ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm17, 320(%rax) ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm18, 256(%rax) ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm14, 448(%rax) ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm11, 384(%rax) ; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, 64(%rax) ; AVX512DQ-BW-NEXT: vzeroupper ; AVX512DQ-BW-NEXT: retq ; ; AVX512DQ-BW-FCP-LABEL: store_i8_stride8_vf64: ; AVX512DQ-BW-FCP: # %bb.0: ; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %r10 ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rsi), %xmm1 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 32(%rsi), %xmm19 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 48(%rsi), %xmm16 ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdi), %xmm2 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 32(%rdi), %xmm20 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 48(%rdi), %xmm17 ; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm0 = xmm2[8],xmm1[8],xmm2[9],xmm1[9],xmm2[10],xmm1[10],xmm2[11],xmm1[11],xmm2[12],xmm1[12],xmm2[13],xmm1[13],xmm2[14],xmm1[14],xmm2[15],xmm1[15] ; AVX512DQ-BW-FCP-NEXT: vinserti128 $1, %xmm0, %ymm0, %ymm3 ; AVX512DQ-BW-FCP-NEXT: vpmovsxwq {{.*#+}} ymm5 = [2312,2826,3340,3854] ; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm5, %ymm3, %ymm3 ; AVX512DQ-BW-FCP-NEXT: vpmovsxwq {{.*#+}} xmm6 = [1284,1798] ; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm6, %xmm0, %xmm4 ; AVX512DQ-BW-FCP-NEXT: vpmovzxwq {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero ; AVX512DQ-BW-FCP-NEXT: vinserti128 $1, %xmm4, %ymm0, %ymm0 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm3, %zmm0, %zmm0 ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rcx), %xmm3 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 32(%rcx), %xmm21 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 48(%rcx), %xmm18 ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdx), %xmm7 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 32(%rdx), %xmm22 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 48(%rdx), %xmm23 ; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm4 = 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-BW-FCP-NEXT: vinserti128 $1, %xmm4, %ymm4, %ymm4 ; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm9 = [0,32,2,3,4,33,6,7,8,42,10,11,12,43,14,15,16,36,18,19,20,37,22,23,24,46,26,27,28,47,30,31] ; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm4, %zmm9, %zmm0 ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%r10), %xmm4 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 48(%r10), %xmm24 ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rax), %xmm8 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 48(%rax), %xmm26 ; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm10 = xmm8[8],xmm4[8],xmm8[9],xmm4[9],xmm8[10],xmm4[10],xmm8[11],xmm4[11],xmm8[12],xmm4[12],xmm8[13],xmm4[13],xmm8[14],xmm4[14],xmm8[15],xmm4[15] ; AVX512DQ-BW-FCP-NEXT: vinserti128 $1, %xmm10, %ymm10, %ymm13 ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%r9), %xmm10 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 48(%r9), %xmm28 ; AVX512DQ-BW-FCP-NEXT: vmovdqa (%r8), %xmm12 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 48(%r8), %xmm30 ; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm11 = xmm12[8],xmm10[8],xmm12[9],xmm10[9],xmm12[10],xmm10[10],xmm12[11],xmm10[11],xmm12[12],xmm10[12],xmm12[13],xmm10[13],xmm12[14],xmm10[14],xmm12[15],xmm10[15] ; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm14 = [0,1,0,32,4,5,1,33,2,1,2,42,4,5,3,43,0,1,4,36,4,5,5,37,0,1,6,46,6,5,7,47] ; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm13, %zmm14, %zmm11 ; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm13 = 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-FCP-NEXT: vinserti128 $1, %xmm13, %ymm13, %ymm15 ; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm5, %ymm15, %ymm15 ; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm6, %xmm13, %xmm25 ; AVX512DQ-BW-FCP-NEXT: vpmovzxwq {{.*#+}} xmm13 = xmm13[0],zero,zero,zero,xmm13[1],zero,zero,zero ; AVX512DQ-BW-FCP-NEXT: vinserti32x4 $1, %xmm25, %ymm13, %ymm13 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm15, %zmm13, %zmm13 ; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm15 = xmm23[0],xmm18[0],xmm23[1],xmm18[1],xmm23[2],xmm18[2],xmm23[3],xmm18[3],xmm23[4],xmm18[4],xmm23[5],xmm18[5],xmm23[6],xmm18[6],xmm23[7],xmm18[7] ; AVX512DQ-BW-FCP-NEXT: vinserti128 $1, %xmm15, %ymm15, %ymm15 ; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm15, %zmm9, %zmm13 ; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm15 = xmm26[0],xmm24[0],xmm26[1],xmm24[1],xmm26[2],xmm24[2],xmm26[3],xmm24[3],xmm26[4],xmm24[4],xmm26[5],xmm24[5],xmm26[6],xmm24[6],xmm26[7],xmm24[7] ; AVX512DQ-BW-FCP-NEXT: vinserti32x4 $1, %xmm15, %ymm15, %ymm25 ; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm15 = xmm30[0],xmm28[0],xmm30[1],xmm28[1],xmm30[2],xmm28[2],xmm30[3],xmm28[3],xmm30[4],xmm28[4],xmm30[5],xmm28[5],xmm30[6],xmm28[6],xmm30[7],xmm28[7] ; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm25, %zmm14, %zmm15 ; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm16 = 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-FCP-NEXT: vpshufb %xmm6, %xmm16, %xmm17 ; AVX512DQ-BW-FCP-NEXT: vpmovzxwq {{.*#+}} xmm25 = xmm16[0],zero,zero,zero,xmm16[1],zero,zero,zero ; AVX512DQ-BW-FCP-NEXT: vinserti32x4 $1, %xmm17, %ymm25, %ymm17 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 32(%r10), %xmm25 ; AVX512DQ-BW-FCP-NEXT: vinserti32x4 $1, %xmm16, %ymm16, %ymm16 ; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm5, %ymm16, %ymm16 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm16, %zmm17, %zmm16 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 32(%rax), %xmm29 ; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm17 = xmm23[8],xmm18[8],xmm23[9],xmm18[9],xmm23[10],xmm18[10],xmm23[11],xmm18[11],xmm23[12],xmm18[12],xmm23[13],xmm18[13],xmm23[14],xmm18[14],xmm23[15],xmm18[15] ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 32(%r9), %xmm27 ; AVX512DQ-BW-FCP-NEXT: vinserti32x4 $1, %xmm17, %ymm17, %ymm17 ; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm17, %zmm9, %zmm16 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 32(%r8), %xmm31 ; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm17 = xmm26[8],xmm24[8],xmm26[9],xmm24[9],xmm26[10],xmm24[10],xmm26[11],xmm24[11],xmm26[12],xmm24[12],xmm26[13],xmm24[13],xmm26[14],xmm24[14],xmm26[15],xmm24[15] ; AVX512DQ-BW-FCP-NEXT: vinserti32x4 $1, %xmm17, %ymm17, %ymm18 ; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm17 = xmm30[8],xmm28[8],xmm30[9],xmm28[9],xmm30[10],xmm28[10],xmm30[11],xmm28[11],xmm30[12],xmm28[12],xmm30[13],xmm28[13],xmm30[14],xmm28[14],xmm30[15],xmm28[15] ; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm18, %zmm14, %zmm17 ; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm18 = xmm20[0],xmm19[0],xmm20[1],xmm19[1],xmm20[2],xmm19[2],xmm20[3],xmm19[3],xmm20[4],xmm19[4],xmm20[5],xmm19[5],xmm20[6],xmm19[6],xmm20[7],xmm19[7] ; AVX512DQ-BW-FCP-NEXT: vinserti32x4 $1, %xmm18, %ymm18, %ymm23 ; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm5, %ymm23, %ymm23 ; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm6, %xmm18, %xmm24 ; AVX512DQ-BW-FCP-NEXT: vpmovzxwq {{.*#+}} xmm18 = xmm18[0],zero,zero,zero,xmm18[1],zero,zero,zero ; AVX512DQ-BW-FCP-NEXT: vinserti32x4 $1, %xmm24, %ymm18, %ymm18 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm23, %zmm18, %zmm18 ; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm23 = xmm22[0],xmm21[0],xmm22[1],xmm21[1],xmm22[2],xmm21[2],xmm22[3],xmm21[3],xmm22[4],xmm21[4],xmm22[5],xmm21[5],xmm22[6],xmm21[6],xmm22[7],xmm21[7] ; AVX512DQ-BW-FCP-NEXT: vinserti32x4 $1, %xmm23, %ymm23, %ymm23 ; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm23, %zmm9, %zmm18 ; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm23 = xmm29[0],xmm25[0],xmm29[1],xmm25[1],xmm29[2],xmm25[2],xmm29[3],xmm25[3],xmm29[4],xmm25[4],xmm29[5],xmm25[5],xmm29[6],xmm25[6],xmm29[7],xmm25[7] ; AVX512DQ-BW-FCP-NEXT: vinserti32x4 $1, %xmm23, %ymm23, %ymm24 ; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm23 = xmm31[0],xmm27[0],xmm31[1],xmm27[1],xmm31[2],xmm27[2],xmm31[3],xmm27[3],xmm31[4],xmm27[4],xmm31[5],xmm27[5],xmm31[6],xmm27[6],xmm31[7],xmm27[7] ; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm24, %zmm14, %zmm23 ; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm19 = xmm20[8],xmm19[8],xmm20[9],xmm19[9],xmm20[10],xmm19[10],xmm20[11],xmm19[11],xmm20[12],xmm19[12],xmm20[13],xmm19[13],xmm20[14],xmm19[14],xmm20[15],xmm19[15] ; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm6, %xmm19, %xmm20 ; AVX512DQ-BW-FCP-NEXT: vpmovzxwq {{.*#+}} xmm24 = xmm19[0],zero,zero,zero,xmm19[1],zero,zero,zero ; AVX512DQ-BW-FCP-NEXT: vinserti32x4 $1, %xmm20, %ymm24, %ymm20 ; AVX512DQ-BW-FCP-NEXT: vinserti32x4 $1, %xmm19, %ymm19, %ymm19 ; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm5, %ymm19, %ymm19 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm19, %zmm20, %zmm19 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 16(%rsi), %xmm24 ; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm20 = xmm22[8],xmm21[8],xmm22[9],xmm21[9],xmm22[10],xmm21[10],xmm22[11],xmm21[11],xmm22[12],xmm21[12],xmm22[13],xmm21[13],xmm22[14],xmm21[14],xmm22[15],xmm21[15] ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 16(%rdi), %xmm22 ; AVX512DQ-BW-FCP-NEXT: vinserti32x4 $1, %xmm20, %ymm20, %ymm20 ; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm20, %zmm9, %zmm19 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 16(%rcx), %xmm26 ; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm20 = xmm29[8],xmm25[8],xmm29[9],xmm25[9],xmm29[10],xmm25[10],xmm29[11],xmm25[11],xmm29[12],xmm25[12],xmm29[13],xmm25[13],xmm29[14],xmm25[14],xmm29[15],xmm25[15] ; AVX512DQ-BW-FCP-NEXT: vinserti32x4 $1, %xmm20, %ymm20, %ymm21 ; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm20 = xmm31[8],xmm27[8],xmm31[9],xmm27[9],xmm31[10],xmm27[10],xmm31[11],xmm27[11],xmm31[12],xmm27[12],xmm31[13],xmm27[13],xmm31[14],xmm27[14],xmm31[15],xmm27[15] ; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm21, %zmm14, %zmm20 ; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm21 = xmm22[0],xmm24[0],xmm22[1],xmm24[1],xmm22[2],xmm24[2],xmm22[3],xmm24[3],xmm22[4],xmm24[4],xmm22[5],xmm24[5],xmm22[6],xmm24[6],xmm22[7],xmm24[7] ; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm6, %xmm21, %xmm25 ; AVX512DQ-BW-FCP-NEXT: vpmovzxwq {{.*#+}} xmm27 = xmm21[0],zero,zero,zero,xmm21[1],zero,zero,zero ; AVX512DQ-BW-FCP-NEXT: vinserti32x4 $1, %xmm25, %ymm27, %ymm25 ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 16(%rdx), %xmm27 ; AVX512DQ-BW-FCP-NEXT: vinserti32x4 $1, %xmm21, %ymm21, %ymm21 ; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm5, %ymm21, %ymm21 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm21, %zmm25, %zmm21 ; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm25 = xmm27[0],xmm26[0],xmm27[1],xmm26[1],xmm27[2],xmm26[2],xmm27[3],xmm26[3],xmm27[4],xmm26[4],xmm27[5],xmm26[5],xmm27[6],xmm26[6],xmm27[7],xmm26[7] ; AVX512DQ-BW-FCP-NEXT: vinserti32x4 $1, %xmm25, %ymm25, %ymm25 ; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm25, %zmm9, %zmm21 ; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm22 = xmm22[8],xmm24[8],xmm22[9],xmm24[9],xmm22[10],xmm24[10],xmm22[11],xmm24[11],xmm22[12],xmm24[12],xmm22[13],xmm24[13],xmm22[14],xmm24[14],xmm22[15],xmm24[15] ; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm6, %xmm22, %xmm24 ; AVX512DQ-BW-FCP-NEXT: vpmovzxwq {{.*#+}} xmm25 = xmm22[0],zero,zero,zero,xmm22[1],zero,zero,zero ; AVX512DQ-BW-FCP-NEXT: vinserti32x4 $1, %xmm24, %ymm25, %ymm24 ; AVX512DQ-BW-FCP-NEXT: vinserti32x4 $1, %xmm22, %ymm22, %ymm22 ; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm5, %ymm22, %ymm22 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm22, %zmm24, %zmm22 ; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm24 = xmm27[8],xmm26[8],xmm27[9],xmm26[9],xmm27[10],xmm26[10],xmm27[11],xmm26[11],xmm27[12],xmm26[12],xmm27[13],xmm26[13],xmm27[14],xmm26[14],xmm27[15],xmm26[15] ; AVX512DQ-BW-FCP-NEXT: vinserti32x4 $1, %xmm24, %ymm24, %ymm24 ; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm24, %zmm9, %zmm22 ; AVX512DQ-BW-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] ; AVX512DQ-BW-FCP-NEXT: vinserti128 $1, %xmm1, %ymm1, %ymm2 ; AVX512DQ-BW-FCP-NEXT: vpshufb %ymm5, %ymm2, %ymm2 ; AVX512DQ-BW-FCP-NEXT: vmovdqa 16(%r10), %xmm5 ; AVX512DQ-BW-FCP-NEXT: vpshufb %xmm6, %xmm1, %xmm6 ; AVX512DQ-BW-FCP-NEXT: vpmovzxwq {{.*#+}} xmm1 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero ; AVX512DQ-BW-FCP-NEXT: vinserti128 $1, %xmm6, %ymm1, %ymm1 ; AVX512DQ-BW-FCP-NEXT: vmovdqa 16(%rax), %xmm6 ; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm2, %zmm1, %zmm1 ; AVX512DQ-BW-FCP-NEXT: vmovdqa 16(%r9), %xmm2 ; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm7[0],xmm3[0],xmm7[1],xmm3[1],xmm7[2],xmm3[2],xmm7[3],xmm3[3],xmm7[4],xmm3[4],xmm7[5],xmm3[5],xmm7[6],xmm3[6],xmm7[7],xmm3[7] ; AVX512DQ-BW-FCP-NEXT: vmovdqa 16(%r8), %xmm7 ; AVX512DQ-BW-FCP-NEXT: vinserti128 $1, %xmm3, %ymm3, %ymm3 ; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm3, %zmm9, %zmm1 ; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3],xmm6[4],xmm5[4],xmm6[5],xmm5[5],xmm6[6],xmm5[6],xmm6[7],xmm5[7] ; AVX512DQ-BW-FCP-NEXT: vinserti128 $1, %xmm3, %ymm3, %ymm3 ; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm9 = 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] ; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm3, %zmm14, %zmm9 ; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm6[8],xmm5[8],xmm6[9],xmm5[9],xmm6[10],xmm5[10],xmm6[11],xmm5[11],xmm6[12],xmm5[12],xmm6[13],xmm5[13],xmm6[14],xmm5[14],xmm6[15],xmm5[15] ; AVX512DQ-BW-FCP-NEXT: vpunpckhbw {{.*#+}} xmm2 = xmm7[8],xmm2[8],xmm7[9],xmm2[9],xmm7[10],xmm2[10],xmm7[11],xmm2[11],xmm7[12],xmm2[12],xmm7[13],xmm2[13],xmm7[14],xmm2[14],xmm7[15],xmm2[15] ; AVX512DQ-BW-FCP-NEXT: vinserti128 $1, %xmm3, %ymm3, %ymm3 ; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm3, %zmm14, %zmm2 ; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm3 = xmm8[0],xmm4[0],xmm8[1],xmm4[1],xmm8[2],xmm4[2],xmm8[3],xmm4[3],xmm8[4],xmm4[4],xmm8[5],xmm4[5],xmm8[6],xmm4[6],xmm8[7],xmm4[7] ; AVX512DQ-BW-FCP-NEXT: vpunpcklbw {{.*#+}} xmm4 = xmm12[0],xmm10[0],xmm12[1],xmm10[1],xmm12[2],xmm10[2],xmm12[3],xmm10[3],xmm12[4],xmm10[4],xmm12[5],xmm10[5],xmm12[6],xmm10[6],xmm12[7],xmm10[7] ; AVX512DQ-BW-FCP-NEXT: vinserti128 $1, %xmm3, %ymm3, %ymm3 ; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm3, %zmm14, %zmm4 ; AVX512DQ-BW-FCP-NEXT: movw $-21846, %ax # imm = 0xAAAA ; AVX512DQ-BW-FCP-NEXT: kmovd %eax, %k1 ; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm11, %zmm0 {%k1} ; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm15, %zmm13 {%k1} ; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm17, %zmm16 {%k1} ; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm23, %zmm18 {%k1} ; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm20, %zmm19 {%k1} ; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm9, %zmm21 {%k1} ; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm2, %zmm22 {%k1} ; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm4, %zmm1 {%k1} ; AVX512DQ-BW-FCP-NEXT: movq {{[0-9]+}}(%rsp), %rax ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm1, (%rax) ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm22, 192(%rax) ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm21, 128(%rax) ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm19, 320(%rax) ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm18, 256(%rax) ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm16, 448(%rax) ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm13, 384(%rax) ; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, 64(%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 %in.vec6 = load <64 x i8>, ptr %in.vecptr6, align 64 %in.vec7 = load <64 x i8>, ptr %in.vecptr7, 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 <64 x i8> %in.vec6, <64 x i8> %in.vec7, <128 x i32> %5 = shufflevector <128 x i8> %1, <128 x i8> %2, <256 x i32> %6 = shufflevector <128 x i8> %3, <128 x i8> %4, <256 x i32> %7 = shufflevector <256 x i8> %5, <256 x i8> %6, <512 x i32> %interleaved.vec = shufflevector <512 x i8> %7, <512 x i8> poison, <512 x i32> store <512 x i8> %interleaved.vec, ptr %out.vec, align 64 ret void }