brintos

brintos / llvm-project-archived public Read only

0
0
Text · 1.1 MiB · 29be2b2 Raw
16485 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc < %s -mtriple=x86_64-- -mattr=+sse2 | FileCheck %s --check-prefixes=SSE3; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx  | FileCheck %s --check-prefixes=AVX4; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx2 | FileCheck %s --check-prefixes=AVX25; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx2,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX2-FP6; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx2,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX2-FCP7; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl | FileCheck %s --check-prefixes=AVX5128; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX512-FCP9; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+avx512dq | FileCheck %s --check-prefixes=AVX512DQ10; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+avx512dq,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX512DQ-FCP11; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+avx512bw | FileCheck %s --check-prefixes=AVX512BW12; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+avx512bw,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX512BW-FCP13; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+avx512dq,+avx512bw | FileCheck %s --check-prefixes=AVX512DQ-BW14; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+avx512dq,+avx512bw,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX512DQ-BW-FCP15 16; These patterns are produced by LoopVectorizer for interleaved stores.17 18define void @store_i32_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 {19; SSE-LABEL: store_i32_stride8_vf2:20; SSE:       # %bb.0:21; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %rax22; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %r1023; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %r1124; SSE-NEXT:    movsd {{.*#+}} xmm0 = mem[0],zero25; SSE-NEXT:    movsd {{.*#+}} xmm1 = mem[0],zero26; SSE-NEXT:    unpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]27; SSE-NEXT:    movsd {{.*#+}} xmm1 = mem[0],zero28; SSE-NEXT:    movsd {{.*#+}} xmm2 = mem[0],zero29; SSE-NEXT:    unpcklps {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]30; SSE-NEXT:    movsd {{.*#+}} xmm2 = mem[0],zero31; SSE-NEXT:    movsd {{.*#+}} xmm3 = mem[0],zero32; SSE-NEXT:    unpcklps {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1]33; SSE-NEXT:    movsd {{.*#+}} xmm3 = mem[0],zero34; SSE-NEXT:    movsd {{.*#+}} xmm4 = mem[0],zero35; SSE-NEXT:    unpcklps {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1]36; SSE-NEXT:    movaps %xmm0, %xmm437; SSE-NEXT:    movlhps {{.*#+}} xmm4 = xmm4[0],xmm1[0]38; SSE-NEXT:    movaps %xmm2, %xmm539; SSE-NEXT:    movlhps {{.*#+}} xmm5 = xmm5[0],xmm3[0]40; SSE-NEXT:    movhlps {{.*#+}} xmm1 = xmm0[1],xmm1[1]41; SSE-NEXT:    movhlps {{.*#+}} xmm3 = xmm2[1],xmm3[1]42; SSE-NEXT:    movaps %xmm3, 48(%rax)43; SSE-NEXT:    movaps %xmm1, 32(%rax)44; SSE-NEXT:    movaps %xmm5, 16(%rax)45; SSE-NEXT:    movaps %xmm4, (%rax)46; SSE-NEXT:    retq47;48; AVX-LABEL: store_i32_stride8_vf2:49; AVX:       # %bb.0:50; AVX-NEXT:    movq {{[0-9]+}}(%rsp), %rax51; AVX-NEXT:    movq {{[0-9]+}}(%rsp), %r1052; AVX-NEXT:    movq {{[0-9]+}}(%rsp), %r1153; AVX-NEXT:    vmovsd {{.*#+}} xmm0 = mem[0],zero54; AVX-NEXT:    vmovsd {{.*#+}} xmm1 = mem[0],zero55; AVX-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm1[0],xmm0[0]56; AVX-NEXT:    vmovsd {{.*#+}} xmm1 = mem[0],zero57; AVX-NEXT:    vmovsd {{.*#+}} xmm2 = mem[0],zero58; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm2[0],xmm1[0]59; AVX-NEXT:    vmovsd {{.*#+}} xmm2 = mem[0],zero60; AVX-NEXT:    vmovsd {{.*#+}} xmm3 = mem[0],zero61; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm3[0],xmm2[0]62; AVX-NEXT:    vmovsd {{.*#+}} xmm3 = mem[0],zero63; AVX-NEXT:    vmovsd {{.*#+}} xmm4 = mem[0],zero64; AVX-NEXT:    vmovlhps {{.*#+}} xmm3 = xmm4[0],xmm3[0]65; AVX-NEXT:    vinsertf128 $1, %xmm3, %ymm1, %ymm166; AVX-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm067; AVX-NEXT:    vshufps {{.*#+}} ymm2 = ymm0[0,2],ymm1[0,2],ymm0[4,6],ymm1[4,6]68; AVX-NEXT:    vshufps {{.*#+}} ymm0 = ymm0[1,3],ymm1[1,3],ymm0[5,7],ymm1[5,7]69; AVX-NEXT:    vmovaps %ymm0, 32(%rax)70; AVX-NEXT:    vmovaps %ymm2, (%rax)71; AVX-NEXT:    vzeroupper72; AVX-NEXT:    retq73;74; AVX2-LABEL: store_i32_stride8_vf2:75; AVX2:       # %bb.0:76; AVX2-NEXT:    movq {{[0-9]+}}(%rsp), %rax77; AVX2-NEXT:    movq {{[0-9]+}}(%rsp), %r1078; AVX2-NEXT:    movq {{[0-9]+}}(%rsp), %r1179; AVX2-NEXT:    vmovsd {{.*#+}} xmm0 = mem[0],zero80; AVX2-NEXT:    vmovsd {{.*#+}} xmm1 = mem[0],zero81; AVX2-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm1[0],xmm0[0]82; AVX2-NEXT:    vmovsd {{.*#+}} xmm1 = mem[0],zero83; AVX2-NEXT:    vmovsd {{.*#+}} xmm2 = mem[0],zero84; AVX2-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm2[0],xmm1[0]85; AVX2-NEXT:    vmovsd {{.*#+}} xmm2 = mem[0],zero86; AVX2-NEXT:    vmovsd {{.*#+}} xmm3 = mem[0],zero87; AVX2-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm3[0],xmm2[0]88; AVX2-NEXT:    vmovsd {{.*#+}} xmm3 = mem[0],zero89; AVX2-NEXT:    vmovsd {{.*#+}} xmm4 = mem[0],zero90; AVX2-NEXT:    vmovlhps {{.*#+}} xmm3 = xmm4[0],xmm3[0]91; AVX2-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm092; AVX2-NEXT:    vinsertf128 $1, %xmm3, %ymm2, %ymm193; AVX2-NEXT:    vshufps {{.*#+}} ymm2 = ymm0[0,2],ymm1[0,2],ymm0[4,6],ymm1[4,6]94; AVX2-NEXT:    vpermpd {{.*#+}} ymm2 = ymm2[0,2,1,3]95; AVX2-NEXT:    vshufps {{.*#+}} ymm0 = ymm0[1,3],ymm1[1,3],ymm0[5,7],ymm1[5,7]96; AVX2-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,2,1,3]97; AVX2-NEXT:    vmovaps %ymm0, 32(%rax)98; AVX2-NEXT:    vmovaps %ymm2, (%rax)99; AVX2-NEXT:    vzeroupper100; AVX2-NEXT:    retq101;102; AVX2-FP-LABEL: store_i32_stride8_vf2:103; AVX2-FP:       # %bb.0:104; AVX2-FP-NEXT:    movq {{[0-9]+}}(%rsp), %rax105; AVX2-FP-NEXT:    movq {{[0-9]+}}(%rsp), %r10106; AVX2-FP-NEXT:    movq {{[0-9]+}}(%rsp), %r11107; AVX2-FP-NEXT:    vmovsd {{.*#+}} xmm0 = mem[0],zero108; AVX2-FP-NEXT:    vmovsd {{.*#+}} xmm1 = mem[0],zero109; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm1[0],xmm0[0]110; AVX2-FP-NEXT:    vmovsd {{.*#+}} xmm1 = mem[0],zero111; AVX2-FP-NEXT:    vmovsd {{.*#+}} xmm2 = mem[0],zero112; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm2[0],xmm1[0]113; AVX2-FP-NEXT:    vmovsd {{.*#+}} xmm2 = mem[0],zero114; AVX2-FP-NEXT:    vmovsd {{.*#+}} xmm3 = mem[0],zero115; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm3[0],xmm2[0]116; AVX2-FP-NEXT:    vmovsd {{.*#+}} xmm3 = mem[0],zero117; AVX2-FP-NEXT:    vmovsd {{.*#+}} xmm4 = mem[0],zero118; AVX2-FP-NEXT:    vmovlhps {{.*#+}} xmm3 = xmm4[0],xmm3[0]119; AVX2-FP-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0120; AVX2-FP-NEXT:    vinsertf128 $1, %xmm3, %ymm2, %ymm1121; AVX2-FP-NEXT:    vshufps {{.*#+}} ymm2 = ymm0[0,2],ymm1[0,2],ymm0[4,6],ymm1[4,6]122; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm2 = ymm2[0,2,1,3]123; AVX2-FP-NEXT:    vshufps {{.*#+}} ymm0 = ymm0[1,3],ymm1[1,3],ymm0[5,7],ymm1[5,7]124; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,2,1,3]125; AVX2-FP-NEXT:    vmovaps %ymm0, 32(%rax)126; AVX2-FP-NEXT:    vmovaps %ymm2, (%rax)127; AVX2-FP-NEXT:    vzeroupper128; AVX2-FP-NEXT:    retq129;130; AVX2-FCP-LABEL: store_i32_stride8_vf2:131; AVX2-FCP:       # %bb.0:132; AVX2-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax133; AVX2-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r10134; AVX2-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r11135; AVX2-FCP-NEXT:    vmovsd {{.*#+}} xmm0 = mem[0],zero136; AVX2-FCP-NEXT:    vmovsd {{.*#+}} xmm1 = mem[0],zero137; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm1[0],xmm0[0]138; AVX2-FCP-NEXT:    vmovsd {{.*#+}} xmm1 = mem[0],zero139; AVX2-FCP-NEXT:    vmovsd {{.*#+}} xmm2 = mem[0],zero140; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm2[0],xmm1[0]141; AVX2-FCP-NEXT:    vmovsd {{.*#+}} xmm2 = mem[0],zero142; AVX2-FCP-NEXT:    vmovsd {{.*#+}} xmm3 = mem[0],zero143; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm3[0],xmm2[0]144; AVX2-FCP-NEXT:    vmovsd {{.*#+}} xmm3 = mem[0],zero145; AVX2-FCP-NEXT:    vmovsd {{.*#+}} xmm4 = mem[0],zero146; AVX2-FCP-NEXT:    vmovlhps {{.*#+}} xmm3 = xmm4[0],xmm3[0]147; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0148; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm3, %ymm2, %ymm1149; AVX2-FCP-NEXT:    vshufps {{.*#+}} ymm2 = ymm0[0,2],ymm1[0,2],ymm0[4,6],ymm1[4,6]150; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm2 = ymm2[0,2,1,3]151; AVX2-FCP-NEXT:    vshufps {{.*#+}} ymm0 = ymm0[1,3],ymm1[1,3],ymm0[5,7],ymm1[5,7]152; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,2,1,3]153; AVX2-FCP-NEXT:    vmovaps %ymm0, 32(%rax)154; AVX2-FCP-NEXT:    vmovaps %ymm2, (%rax)155; AVX2-FCP-NEXT:    vzeroupper156; AVX2-FCP-NEXT:    retq157;158; AVX512-LABEL: store_i32_stride8_vf2:159; AVX512:       # %bb.0:160; AVX512-NEXT:    movq {{[0-9]+}}(%rsp), %rax161; AVX512-NEXT:    movq {{[0-9]+}}(%rsp), %r10162; AVX512-NEXT:    movq {{[0-9]+}}(%rsp), %r11163; AVX512-NEXT:    vmovq {{.*#+}} xmm0 = mem[0],zero164; AVX512-NEXT:    vmovq {{.*#+}} xmm1 = mem[0],zero165; AVX512-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0]166; AVX512-NEXT:    vmovq {{.*#+}} xmm1 = mem[0],zero167; AVX512-NEXT:    vmovq {{.*#+}} xmm2 = mem[0],zero168; AVX512-NEXT:    vpunpcklqdq {{.*#+}} xmm1 = xmm2[0],xmm1[0]169; AVX512-NEXT:    vmovq {{.*#+}} xmm2 = mem[0],zero170; AVX512-NEXT:    vmovq {{.*#+}} xmm3 = mem[0],zero171; AVX512-NEXT:    vpunpcklqdq {{.*#+}} xmm2 = xmm3[0],xmm2[0]172; AVX512-NEXT:    vmovq {{.*#+}} xmm3 = mem[0],zero173; AVX512-NEXT:    vmovq {{.*#+}} xmm4 = mem[0],zero174; AVX512-NEXT:    vpunpcklqdq {{.*#+}} xmm3 = xmm4[0],xmm3[0]175; AVX512-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0176; AVX512-NEXT:    vinserti128 $1, %xmm3, %ymm2, %ymm1177; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [0,2,4,6,16,18,20,22,1,3,5,7,17,19,21,23]178; AVX512-NEXT:    vpermi2d %zmm1, %zmm0, %zmm2179; AVX512-NEXT:    vmovdqa64 %zmm2, (%rax)180; AVX512-NEXT:    vzeroupper181; AVX512-NEXT:    retq182;183; AVX512-FCP-LABEL: store_i32_stride8_vf2:184; AVX512-FCP:       # %bb.0:185; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax186; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r10187; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r11188; AVX512-FCP-NEXT:    vmovq {{.*#+}} xmm0 = mem[0],zero189; AVX512-FCP-NEXT:    vmovq {{.*#+}} xmm1 = mem[0],zero190; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0]191; AVX512-FCP-NEXT:    vmovq {{.*#+}} xmm1 = mem[0],zero192; AVX512-FCP-NEXT:    vmovq {{.*#+}} xmm2 = mem[0],zero193; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} xmm1 = xmm2[0],xmm1[0]194; AVX512-FCP-NEXT:    vmovq {{.*#+}} xmm2 = mem[0],zero195; AVX512-FCP-NEXT:    vmovq {{.*#+}} xmm3 = mem[0],zero196; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} xmm2 = xmm3[0],xmm2[0]197; AVX512-FCP-NEXT:    vmovq {{.*#+}} xmm3 = mem[0],zero198; AVX512-FCP-NEXT:    vmovq {{.*#+}} xmm4 = mem[0],zero199; AVX512-FCP-NEXT:    vpunpcklqdq {{.*#+}} xmm3 = xmm4[0],xmm3[0]200; AVX512-FCP-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0201; AVX512-FCP-NEXT:    vinserti128 $1, %xmm3, %ymm2, %ymm1202; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [0,2,4,6,16,18,20,22,1,3,5,7,17,19,21,23]203; AVX512-FCP-NEXT:    vpermi2d %zmm1, %zmm0, %zmm2204; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, (%rax)205; AVX512-FCP-NEXT:    vzeroupper206; AVX512-FCP-NEXT:    retq207;208; AVX512DQ-LABEL: store_i32_stride8_vf2:209; AVX512DQ:       # %bb.0:210; AVX512DQ-NEXT:    movq {{[0-9]+}}(%rsp), %rax211; AVX512DQ-NEXT:    movq {{[0-9]+}}(%rsp), %r10212; AVX512DQ-NEXT:    movq {{[0-9]+}}(%rsp), %r11213; AVX512DQ-NEXT:    vmovq {{.*#+}} xmm0 = mem[0],zero214; AVX512DQ-NEXT:    vmovq {{.*#+}} xmm1 = mem[0],zero215; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0]216; AVX512DQ-NEXT:    vmovq {{.*#+}} xmm1 = mem[0],zero217; AVX512DQ-NEXT:    vmovq {{.*#+}} xmm2 = mem[0],zero218; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} xmm1 = xmm2[0],xmm1[0]219; AVX512DQ-NEXT:    vmovq {{.*#+}} xmm2 = mem[0],zero220; AVX512DQ-NEXT:    vmovq {{.*#+}} xmm3 = mem[0],zero221; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} xmm2 = xmm3[0],xmm2[0]222; AVX512DQ-NEXT:    vmovq {{.*#+}} xmm3 = mem[0],zero223; AVX512DQ-NEXT:    vmovq {{.*#+}} xmm4 = mem[0],zero224; AVX512DQ-NEXT:    vpunpcklqdq {{.*#+}} xmm3 = xmm4[0],xmm3[0]225; AVX512DQ-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0226; AVX512DQ-NEXT:    vinserti128 $1, %xmm3, %ymm2, %ymm1227; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [0,2,4,6,16,18,20,22,1,3,5,7,17,19,21,23]228; AVX512DQ-NEXT:    vpermi2d %zmm1, %zmm0, %zmm2229; AVX512DQ-NEXT:    vmovdqa64 %zmm2, (%rax)230; AVX512DQ-NEXT:    vzeroupper231; AVX512DQ-NEXT:    retq232;233; AVX512DQ-FCP-LABEL: store_i32_stride8_vf2:234; AVX512DQ-FCP:       # %bb.0:235; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax236; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r10237; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r11238; AVX512DQ-FCP-NEXT:    vmovq {{.*#+}} xmm0 = mem[0],zero239; AVX512DQ-FCP-NEXT:    vmovq {{.*#+}} xmm1 = mem[0],zero240; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0]241; AVX512DQ-FCP-NEXT:    vmovq {{.*#+}} xmm1 = mem[0],zero242; AVX512DQ-FCP-NEXT:    vmovq {{.*#+}} xmm2 = mem[0],zero243; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} xmm1 = xmm2[0],xmm1[0]244; AVX512DQ-FCP-NEXT:    vmovq {{.*#+}} xmm2 = mem[0],zero245; AVX512DQ-FCP-NEXT:    vmovq {{.*#+}} xmm3 = mem[0],zero246; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} xmm2 = xmm3[0],xmm2[0]247; AVX512DQ-FCP-NEXT:    vmovq {{.*#+}} xmm3 = mem[0],zero248; AVX512DQ-FCP-NEXT:    vmovq {{.*#+}} xmm4 = mem[0],zero249; AVX512DQ-FCP-NEXT:    vpunpcklqdq {{.*#+}} xmm3 = xmm4[0],xmm3[0]250; AVX512DQ-FCP-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0251; AVX512DQ-FCP-NEXT:    vinserti128 $1, %xmm3, %ymm2, %ymm1252; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [0,2,4,6,16,18,20,22,1,3,5,7,17,19,21,23]253; AVX512DQ-FCP-NEXT:    vpermi2d %zmm1, %zmm0, %zmm2254; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm2, (%rax)255; AVX512DQ-FCP-NEXT:    vzeroupper256; AVX512DQ-FCP-NEXT:    retq257;258; AVX512BW-LABEL: store_i32_stride8_vf2:259; AVX512BW:       # %bb.0:260; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax261; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %r10262; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %r11263; AVX512BW-NEXT:    vmovq {{.*#+}} xmm0 = mem[0],zero264; AVX512BW-NEXT:    vmovq {{.*#+}} xmm1 = mem[0],zero265; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0]266; AVX512BW-NEXT:    vmovq {{.*#+}} xmm1 = mem[0],zero267; AVX512BW-NEXT:    vmovq {{.*#+}} xmm2 = mem[0],zero268; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} xmm1 = xmm2[0],xmm1[0]269; AVX512BW-NEXT:    vmovq {{.*#+}} xmm2 = mem[0],zero270; AVX512BW-NEXT:    vmovq {{.*#+}} xmm3 = mem[0],zero271; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} xmm2 = xmm3[0],xmm2[0]272; AVX512BW-NEXT:    vmovq {{.*#+}} xmm3 = mem[0],zero273; AVX512BW-NEXT:    vmovq {{.*#+}} xmm4 = mem[0],zero274; AVX512BW-NEXT:    vpunpcklqdq {{.*#+}} xmm3 = xmm4[0],xmm3[0]275; AVX512BW-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0276; AVX512BW-NEXT:    vinserti128 $1, %xmm3, %ymm2, %ymm1277; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [0,2,4,6,16,18,20,22,1,3,5,7,17,19,21,23]278; AVX512BW-NEXT:    vpermi2d %zmm1, %zmm0, %zmm2279; AVX512BW-NEXT:    vmovdqa64 %zmm2, (%rax)280; AVX512BW-NEXT:    vzeroupper281; AVX512BW-NEXT:    retq282;283; AVX512BW-FCP-LABEL: store_i32_stride8_vf2:284; AVX512BW-FCP:       # %bb.0:285; AVX512BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax286; AVX512BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r10287; AVX512BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r11288; AVX512BW-FCP-NEXT:    vmovq {{.*#+}} xmm0 = mem[0],zero289; AVX512BW-FCP-NEXT:    vmovq {{.*#+}} xmm1 = mem[0],zero290; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0]291; AVX512BW-FCP-NEXT:    vmovq {{.*#+}} xmm1 = mem[0],zero292; AVX512BW-FCP-NEXT:    vmovq {{.*#+}} xmm2 = mem[0],zero293; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} xmm1 = xmm2[0],xmm1[0]294; AVX512BW-FCP-NEXT:    vmovq {{.*#+}} xmm2 = mem[0],zero295; AVX512BW-FCP-NEXT:    vmovq {{.*#+}} xmm3 = mem[0],zero296; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} xmm2 = xmm3[0],xmm2[0]297; AVX512BW-FCP-NEXT:    vmovq {{.*#+}} xmm3 = mem[0],zero298; AVX512BW-FCP-NEXT:    vmovq {{.*#+}} xmm4 = mem[0],zero299; AVX512BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} xmm3 = xmm4[0],xmm3[0]300; AVX512BW-FCP-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0301; AVX512BW-FCP-NEXT:    vinserti128 $1, %xmm3, %ymm2, %ymm1302; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [0,2,4,6,16,18,20,22,1,3,5,7,17,19,21,23]303; AVX512BW-FCP-NEXT:    vpermi2d %zmm1, %zmm0, %zmm2304; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm2, (%rax)305; AVX512BW-FCP-NEXT:    vzeroupper306; AVX512BW-FCP-NEXT:    retq307;308; AVX512DQ-BW-LABEL: store_i32_stride8_vf2:309; AVX512DQ-BW:       # %bb.0:310; AVX512DQ-BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax311; AVX512DQ-BW-NEXT:    movq {{[0-9]+}}(%rsp), %r10312; AVX512DQ-BW-NEXT:    movq {{[0-9]+}}(%rsp), %r11313; AVX512DQ-BW-NEXT:    vmovq {{.*#+}} xmm0 = mem[0],zero314; AVX512DQ-BW-NEXT:    vmovq {{.*#+}} xmm1 = mem[0],zero315; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0]316; AVX512DQ-BW-NEXT:    vmovq {{.*#+}} xmm1 = mem[0],zero317; AVX512DQ-BW-NEXT:    vmovq {{.*#+}} xmm2 = mem[0],zero318; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} xmm1 = xmm2[0],xmm1[0]319; AVX512DQ-BW-NEXT:    vmovq {{.*#+}} xmm2 = mem[0],zero320; AVX512DQ-BW-NEXT:    vmovq {{.*#+}} xmm3 = mem[0],zero321; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} xmm2 = xmm3[0],xmm2[0]322; AVX512DQ-BW-NEXT:    vmovq {{.*#+}} xmm3 = mem[0],zero323; AVX512DQ-BW-NEXT:    vmovq {{.*#+}} xmm4 = mem[0],zero324; AVX512DQ-BW-NEXT:    vpunpcklqdq {{.*#+}} xmm3 = xmm4[0],xmm3[0]325; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0326; AVX512DQ-BW-NEXT:    vinserti128 $1, %xmm3, %ymm2, %ymm1327; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [0,2,4,6,16,18,20,22,1,3,5,7,17,19,21,23]328; AVX512DQ-BW-NEXT:    vpermi2d %zmm1, %zmm0, %zmm2329; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm2, (%rax)330; AVX512DQ-BW-NEXT:    vzeroupper331; AVX512DQ-BW-NEXT:    retq332;333; AVX512DQ-BW-FCP-LABEL: store_i32_stride8_vf2:334; AVX512DQ-BW-FCP:       # %bb.0:335; AVX512DQ-BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax336; AVX512DQ-BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r10337; AVX512DQ-BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r11338; AVX512DQ-BW-FCP-NEXT:    vmovq {{.*#+}} xmm0 = mem[0],zero339; AVX512DQ-BW-FCP-NEXT:    vmovq {{.*#+}} xmm1 = mem[0],zero340; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0]341; AVX512DQ-BW-FCP-NEXT:    vmovq {{.*#+}} xmm1 = mem[0],zero342; AVX512DQ-BW-FCP-NEXT:    vmovq {{.*#+}} xmm2 = mem[0],zero343; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} xmm1 = xmm2[0],xmm1[0]344; AVX512DQ-BW-FCP-NEXT:    vmovq {{.*#+}} xmm2 = mem[0],zero345; AVX512DQ-BW-FCP-NEXT:    vmovq {{.*#+}} xmm3 = mem[0],zero346; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} xmm2 = xmm3[0],xmm2[0]347; AVX512DQ-BW-FCP-NEXT:    vmovq {{.*#+}} xmm3 = mem[0],zero348; AVX512DQ-BW-FCP-NEXT:    vmovq {{.*#+}} xmm4 = mem[0],zero349; AVX512DQ-BW-FCP-NEXT:    vpunpcklqdq {{.*#+}} xmm3 = xmm4[0],xmm3[0]350; AVX512DQ-BW-FCP-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0351; AVX512DQ-BW-FCP-NEXT:    vinserti128 $1, %xmm3, %ymm2, %ymm1352; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [0,2,4,6,16,18,20,22,1,3,5,7,17,19,21,23]353; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm1, %zmm0, %zmm2354; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm2, (%rax)355; AVX512DQ-BW-FCP-NEXT:    vzeroupper356; AVX512DQ-BW-FCP-NEXT:    retq357  %in.vec0 = load <2 x i32>, ptr %in.vecptr0, align 64358  %in.vec1 = load <2 x i32>, ptr %in.vecptr1, align 64359  %in.vec2 = load <2 x i32>, ptr %in.vecptr2, align 64360  %in.vec3 = load <2 x i32>, ptr %in.vecptr3, align 64361  %in.vec4 = load <2 x i32>, ptr %in.vecptr4, align 64362  %in.vec5 = load <2 x i32>, ptr %in.vecptr5, align 64363  %in.vec6 = load <2 x i32>, ptr %in.vecptr6, align 64364  %in.vec7 = load <2 x i32>, ptr %in.vecptr7, align 64365  %1 = shufflevector <2 x i32> %in.vec0, <2 x i32> %in.vec1, <4 x i32> <i32 0, i32 1, i32 2, i32 3>366  %2 = shufflevector <2 x i32> %in.vec2, <2 x i32> %in.vec3, <4 x i32> <i32 0, i32 1, i32 2, i32 3>367  %3 = shufflevector <2 x i32> %in.vec4, <2 x i32> %in.vec5, <4 x i32> <i32 0, i32 1, i32 2, i32 3>368  %4 = shufflevector <2 x i32> %in.vec6, <2 x i32> %in.vec7, <4 x i32> <i32 0, i32 1, i32 2, i32 3>369  %5 = shufflevector <4 x i32> %1, <4 x i32> %2, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>370  %6 = shufflevector <4 x i32> %3, <4 x i32> %4, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>371  %7 = shufflevector <8 x i32> %5, <8 x i32> %6, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>372  %interleaved.vec = shufflevector <16 x i32> %7, <16 x i32> poison, <16 x i32> <i32 0, i32 2, i32 4, i32 6, i32 8, i32 10, i32 12, i32 14, i32 1, i32 3, i32 5, i32 7, i32 9, i32 11, i32 13, i32 15>373  store <16 x i32> %interleaved.vec, ptr %out.vec, align 64374  ret void375}376 377define void @store_i32_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 {378; SSE-LABEL: store_i32_stride8_vf4:379; SSE:       # %bb.0:380; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %rax381; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %r10382; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %r11383; SSE-NEXT:    movaps (%rdi), %xmm0384; SSE-NEXT:    movaps (%rsi), %xmm5385; SSE-NEXT:    movaps (%rdx), %xmm1386; SSE-NEXT:    movaps (%rcx), %xmm6387; SSE-NEXT:    movaps (%r8), %xmm2388; SSE-NEXT:    movaps (%r9), %xmm7389; SSE-NEXT:    movaps (%r11), %xmm8390; SSE-NEXT:    movaps (%r10), %xmm9391; SSE-NEXT:    movaps %xmm1, %xmm10392; SSE-NEXT:    unpcklps {{.*#+}} xmm10 = xmm10[0],xmm6[0],xmm10[1],xmm6[1]393; SSE-NEXT:    movaps %xmm0, %xmm4394; SSE-NEXT:    unpcklps {{.*#+}} xmm4 = xmm4[0],xmm5[0],xmm4[1],xmm5[1]395; SSE-NEXT:    movaps %xmm4, %xmm3396; SSE-NEXT:    movlhps {{.*#+}} xmm3 = xmm3[0],xmm10[0]397; SSE-NEXT:    movaps %xmm8, %xmm11398; SSE-NEXT:    unpcklps {{.*#+}} xmm11 = xmm11[0],xmm9[0],xmm11[1],xmm9[1]399; SSE-NEXT:    movaps %xmm2, %xmm12400; SSE-NEXT:    unpcklps {{.*#+}} xmm12 = xmm12[0],xmm7[0],xmm12[1],xmm7[1]401; SSE-NEXT:    movaps %xmm12, %xmm13402; SSE-NEXT:    movlhps {{.*#+}} xmm13 = xmm13[0],xmm11[0]403; SSE-NEXT:    unpckhpd {{.*#+}} xmm12 = xmm12[1],xmm11[1]404; SSE-NEXT:    unpckhpd {{.*#+}} xmm4 = xmm4[1],xmm10[1]405; SSE-NEXT:    unpckhps {{.*#+}} xmm1 = xmm1[2],xmm6[2],xmm1[3],xmm6[3]406; SSE-NEXT:    unpckhps {{.*#+}} xmm0 = xmm0[2],xmm5[2],xmm0[3],xmm5[3]407; SSE-NEXT:    movaps %xmm0, %xmm5408; SSE-NEXT:    movlhps {{.*#+}} xmm5 = xmm5[0],xmm1[0]409; SSE-NEXT:    unpckhps {{.*#+}} xmm8 = xmm8[2],xmm9[2],xmm8[3],xmm9[3]410; SSE-NEXT:    unpckhps {{.*#+}} xmm2 = xmm2[2],xmm7[2],xmm2[3],xmm7[3]411; SSE-NEXT:    movaps %xmm2, %xmm6412; SSE-NEXT:    movlhps {{.*#+}} xmm6 = xmm6[0],xmm8[0]413; SSE-NEXT:    unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm8[1]414; SSE-NEXT:    unpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]415; SSE-NEXT:    movaps %xmm0, 96(%rax)416; SSE-NEXT:    movaps %xmm2, 112(%rax)417; SSE-NEXT:    movaps %xmm6, 80(%rax)418; SSE-NEXT:    movaps %xmm5, 64(%rax)419; SSE-NEXT:    movaps %xmm4, 32(%rax)420; SSE-NEXT:    movaps %xmm12, 48(%rax)421; SSE-NEXT:    movaps %xmm13, 16(%rax)422; SSE-NEXT:    movaps %xmm3, (%rax)423; SSE-NEXT:    retq424;425; AVX-LABEL: store_i32_stride8_vf4:426; AVX:       # %bb.0:427; AVX-NEXT:    movq {{[0-9]+}}(%rsp), %rax428; AVX-NEXT:    movq {{[0-9]+}}(%rsp), %r10429; AVX-NEXT:    movq {{[0-9]+}}(%rsp), %r11430; AVX-NEXT:    vmovaps (%rdi), %xmm2431; AVX-NEXT:    vmovaps (%rsi), %xmm3432; AVX-NEXT:    vmovaps (%rdx), %xmm0433; AVX-NEXT:    vmovaps (%rcx), %xmm1434; AVX-NEXT:    vmovaps (%r8), %xmm4435; AVX-NEXT:    vmovaps (%r9), %xmm5436; AVX-NEXT:    vmovaps (%r11), %xmm6437; AVX-NEXT:    vmovaps (%r10), %xmm7438; AVX-NEXT:    vunpcklps {{.*#+}} xmm8 = xmm6[0],xmm7[0],xmm6[1],xmm7[1]439; AVX-NEXT:    vshufps {{.*#+}} xmm9 = xmm4[1,1,1,1]440; AVX-NEXT:    vblendps {{.*#+}} xmm9 = xmm9[0],xmm5[1],xmm9[2,3]441; AVX-NEXT:    vblendps {{.*#+}} xmm9 = xmm9[0,1],xmm8[2,3]442; AVX-NEXT:    vinsertf128 $1, %xmm9, %ymm0, %ymm9443; AVX-NEXT:    vunpcklps {{.*#+}} xmm10 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]444; AVX-NEXT:    vshufps {{.*#+}} xmm11 = xmm2[1,1,1,1]445; AVX-NEXT:    vblendps {{.*#+}} xmm11 = xmm11[0],xmm3[1],xmm11[2,3]446; AVX-NEXT:    vblendps {{.*#+}} xmm11 = xmm11[0,1],xmm10[2,3]447; AVX-NEXT:    vblendps {{.*#+}} ymm9 = ymm11[0,1,2,3],ymm9[4,5,6,7]448; AVX-NEXT:    vunpckhps {{.*#+}} xmm11 = xmm4[2],xmm5[2],xmm4[3],xmm5[3]449; AVX-NEXT:    vshufps {{.*#+}} xmm12 = xmm7[2,2,2,2]450; AVX-NEXT:    vblendps {{.*#+}} xmm12 = xmm6[0,1,2],xmm12[3]451; AVX-NEXT:    vblendps {{.*#+}} xmm12 = xmm11[0,1],xmm12[2,3]452; AVX-NEXT:    vinsertf128 $1, %xmm12, %ymm0, %ymm12453; AVX-NEXT:    vunpckhps {{.*#+}} xmm13 = xmm2[2],xmm3[2],xmm2[3],xmm3[3]454; AVX-NEXT:    vshufps {{.*#+}} xmm14 = xmm1[2,2,2,2]455; AVX-NEXT:    vblendps {{.*#+}} xmm14 = xmm0[0,1,2],xmm14[3]456; AVX-NEXT:    vblendps {{.*#+}} xmm14 = xmm13[0,1],xmm14[2,3]457; AVX-NEXT:    vblendps {{.*#+}} ymm12 = ymm14[0,1,2,3],ymm12[4,5,6,7]458; AVX-NEXT:    vunpcklps {{.*#+}} xmm4 = xmm4[0],xmm5[0],xmm4[1],xmm5[1]459; AVX-NEXT:    vmovlhps {{.*#+}} xmm4 = xmm4[0],xmm8[0]460; AVX-NEXT:    vinsertf128 $1, %xmm4, %ymm0, %ymm4461; AVX-NEXT:    vunpcklps {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1]462; AVX-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm2[0],xmm10[0]463; AVX-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm4[4,5,6,7]464; AVX-NEXT:    vunpckhps {{.*#+}} xmm3 = xmm6[2],xmm7[2],xmm6[3],xmm7[3]465; AVX-NEXT:    vunpckhpd {{.*#+}} xmm3 = xmm11[1],xmm3[1]466; AVX-NEXT:    vinsertf128 $1, %xmm3, %ymm0, %ymm3467; AVX-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]468; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm13[1],xmm0[1]469; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm3[4,5,6,7]470; AVX-NEXT:    vmovaps %ymm0, 96(%rax)471; AVX-NEXT:    vmovaps %ymm2, (%rax)472; AVX-NEXT:    vmovaps %ymm12, 64(%rax)473; AVX-NEXT:    vmovaps %ymm9, 32(%rax)474; AVX-NEXT:    vzeroupper475; AVX-NEXT:    retq476;477; AVX2-LABEL: store_i32_stride8_vf4:478; AVX2:       # %bb.0:479; AVX2-NEXT:    movq {{[0-9]+}}(%rsp), %rax480; AVX2-NEXT:    movq {{[0-9]+}}(%rsp), %r10481; AVX2-NEXT:    movq {{[0-9]+}}(%rsp), %r11482; AVX2-NEXT:    vmovaps (%rdi), %xmm3483; AVX2-NEXT:    vmovaps (%rsi), %xmm4484; AVX2-NEXT:    vmovaps (%rdx), %xmm0485; AVX2-NEXT:    vmovaps (%rcx), %xmm1486; AVX2-NEXT:    vmovaps (%r8), %xmm5487; AVX2-NEXT:    vmovaps (%r11), %xmm6488; AVX2-NEXT:    vinsertf128 $1, %xmm4, %ymm3, %ymm2489; AVX2-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm7490; AVX2-NEXT:    vinsertf128 $1, (%r9), %ymm5, %ymm5491; AVX2-NEXT:    vinsertf128 $1, (%r10), %ymm6, %ymm8492; AVX2-NEXT:    vbroadcastsd {{.*#+}} ymm6 = [0,4,0,4,0,4,0,4]493; AVX2-NEXT:    vpermps %ymm8, %ymm6, %ymm9494; AVX2-NEXT:    vpermps %ymm5, %ymm6, %ymm10495; AVX2-NEXT:    vblendps {{.*#+}} ymm9 = ymm10[0,1,2,3,4,5],ymm9[6,7]496; AVX2-NEXT:    vunpcklps {{.*#+}} xmm10 = xmm3[0],xmm4[0],xmm3[1],xmm4[1]497; AVX2-NEXT:    vpermps %ymm7, %ymm6, %ymm6498; AVX2-NEXT:    vblendps {{.*#+}} xmm6 = xmm10[0,1],xmm6[2,3]499; AVX2-NEXT:    vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm9[4,5,6,7]500; AVX2-NEXT:    vbroadcastsd {{.*#+}} ymm9 = [1,5,1,5,1,5,1,5]501; AVX2-NEXT:    vpermps %ymm8, %ymm9, %ymm10502; AVX2-NEXT:    vpermps %ymm5, %ymm9, %ymm11503; AVX2-NEXT:    vblendps {{.*#+}} ymm10 = ymm11[0,1,2,3,4,5],ymm10[6,7]504; AVX2-NEXT:    vunpcklps {{.*#+}} xmm11 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]505; AVX2-NEXT:    vpermps %ymm2, %ymm9, %ymm9506; AVX2-NEXT:    vblendps {{.*#+}} xmm9 = xmm9[0,1],xmm11[2,3]507; AVX2-NEXT:    vblendps {{.*#+}} ymm9 = ymm9[0,1,2,3],ymm10[4,5,6,7]508; AVX2-NEXT:    vbroadcastsd {{.*#+}} ymm10 = [2,6,2,6,2,6,2,6]509; AVX2-NEXT:    vpermps %ymm8, %ymm10, %ymm11510; AVX2-NEXT:    vpermps %ymm5, %ymm10, %ymm12511; AVX2-NEXT:    vblendps {{.*#+}} ymm11 = ymm12[0,1,2,3,4,5],ymm11[6,7]512; AVX2-NEXT:    vunpckhps {{.*#+}} xmm3 = xmm3[2],xmm4[2],xmm3[3],xmm4[3]513; AVX2-NEXT:    vpermps %ymm7, %ymm10, %ymm4514; AVX2-NEXT:    vblendps {{.*#+}} xmm3 = xmm3[0,1],xmm4[2,3]515; AVX2-NEXT:    vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm11[4,5,6,7]516; AVX2-NEXT:    vbroadcastsd {{.*#+}} ymm4 = [3,7,3,7,3,7,3,7]517; AVX2-NEXT:    vpermps %ymm8, %ymm4, %ymm7518; AVX2-NEXT:    vpermps %ymm5, %ymm4, %ymm5519; AVX2-NEXT:    vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3,4,5],ymm7[6,7]520; AVX2-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]521; AVX2-NEXT:    vpermps %ymm2, %ymm4, %ymm1522; AVX2-NEXT:    vblendps {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3]523; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm5[4,5,6,7]524; AVX2-NEXT:    vmovaps %ymm0, 96(%rax)525; AVX2-NEXT:    vmovaps %ymm3, 64(%rax)526; AVX2-NEXT:    vmovaps %ymm9, 32(%rax)527; AVX2-NEXT:    vmovaps %ymm6, (%rax)528; AVX2-NEXT:    vzeroupper529; AVX2-NEXT:    retq530;531; AVX2-FP-LABEL: store_i32_stride8_vf4:532; AVX2-FP:       # %bb.0:533; AVX2-FP-NEXT:    movq {{[0-9]+}}(%rsp), %rax534; AVX2-FP-NEXT:    movq {{[0-9]+}}(%rsp), %r10535; AVX2-FP-NEXT:    movq {{[0-9]+}}(%rsp), %r11536; AVX2-FP-NEXT:    vmovaps (%rdi), %xmm3537; AVX2-FP-NEXT:    vmovaps (%rsi), %xmm4538; AVX2-FP-NEXT:    vmovaps (%rdx), %xmm0539; AVX2-FP-NEXT:    vmovaps (%rcx), %xmm1540; AVX2-FP-NEXT:    vmovaps (%r8), %xmm5541; AVX2-FP-NEXT:    vmovaps (%r11), %xmm6542; AVX2-FP-NEXT:    vinsertf128 $1, %xmm4, %ymm3, %ymm2543; AVX2-FP-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm7544; AVX2-FP-NEXT:    vinsertf128 $1, (%r9), %ymm5, %ymm5545; AVX2-FP-NEXT:    vinsertf128 $1, (%r10), %ymm6, %ymm8546; AVX2-FP-NEXT:    vbroadcastsd {{.*#+}} ymm6 = [0,4,0,4,0,4,0,4]547; AVX2-FP-NEXT:    vpermps %ymm8, %ymm6, %ymm9548; AVX2-FP-NEXT:    vpermps %ymm5, %ymm6, %ymm10549; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm9 = ymm10[0,1,2,3,4,5],ymm9[6,7]550; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm10 = xmm3[0],xmm4[0],xmm3[1],xmm4[1]551; AVX2-FP-NEXT:    vpermps %ymm7, %ymm6, %ymm6552; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm6 = xmm10[0,1],xmm6[2,3]553; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm9[4,5,6,7]554; AVX2-FP-NEXT:    vbroadcastsd {{.*#+}} ymm9 = [1,5,1,5,1,5,1,5]555; AVX2-FP-NEXT:    vpermps %ymm8, %ymm9, %ymm10556; AVX2-FP-NEXT:    vpermps %ymm5, %ymm9, %ymm11557; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm10 = ymm11[0,1,2,3,4,5],ymm10[6,7]558; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm11 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]559; AVX2-FP-NEXT:    vpermps %ymm2, %ymm9, %ymm9560; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm9 = xmm9[0,1],xmm11[2,3]561; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm9 = ymm9[0,1,2,3],ymm10[4,5,6,7]562; AVX2-FP-NEXT:    vbroadcastsd {{.*#+}} ymm10 = [2,6,2,6,2,6,2,6]563; AVX2-FP-NEXT:    vpermps %ymm8, %ymm10, %ymm11564; AVX2-FP-NEXT:    vpermps %ymm5, %ymm10, %ymm12565; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm11 = ymm12[0,1,2,3,4,5],ymm11[6,7]566; AVX2-FP-NEXT:    vunpckhps {{.*#+}} xmm3 = xmm3[2],xmm4[2],xmm3[3],xmm4[3]567; AVX2-FP-NEXT:    vpermps %ymm7, %ymm10, %ymm4568; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm3 = xmm3[0,1],xmm4[2,3]569; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm11[4,5,6,7]570; AVX2-FP-NEXT:    vbroadcastsd {{.*#+}} ymm4 = [3,7,3,7,3,7,3,7]571; AVX2-FP-NEXT:    vpermps %ymm8, %ymm4, %ymm7572; AVX2-FP-NEXT:    vpermps %ymm5, %ymm4, %ymm5573; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3,4,5],ymm7[6,7]574; AVX2-FP-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]575; AVX2-FP-NEXT:    vpermps %ymm2, %ymm4, %ymm1576; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3]577; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm5[4,5,6,7]578; AVX2-FP-NEXT:    vmovaps %ymm0, 96(%rax)579; AVX2-FP-NEXT:    vmovaps %ymm3, 64(%rax)580; AVX2-FP-NEXT:    vmovaps %ymm9, 32(%rax)581; AVX2-FP-NEXT:    vmovaps %ymm6, (%rax)582; AVX2-FP-NEXT:    vzeroupper583; AVX2-FP-NEXT:    retq584;585; AVX2-FCP-LABEL: store_i32_stride8_vf4:586; AVX2-FCP:       # %bb.0:587; AVX2-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax588; AVX2-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r10589; AVX2-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r11590; AVX2-FCP-NEXT:    vmovaps (%rdi), %xmm3591; AVX2-FCP-NEXT:    vmovaps (%rsi), %xmm4592; AVX2-FCP-NEXT:    vmovaps (%rdx), %xmm0593; AVX2-FCP-NEXT:    vmovaps (%rcx), %xmm1594; AVX2-FCP-NEXT:    vmovaps (%r8), %xmm5595; AVX2-FCP-NEXT:    vmovaps (%r11), %xmm6596; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm4, %ymm3, %ymm2597; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm7598; AVX2-FCP-NEXT:    vinsertf128 $1, (%r9), %ymm5, %ymm5599; AVX2-FCP-NEXT:    vinsertf128 $1, (%r10), %ymm6, %ymm8600; AVX2-FCP-NEXT:    vbroadcastsd {{.*#+}} ymm6 = [0,4,0,4,0,4,0,4]601; AVX2-FCP-NEXT:    vpermps %ymm8, %ymm6, %ymm9602; AVX2-FCP-NEXT:    vpermps %ymm5, %ymm6, %ymm10603; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm9 = ymm10[0,1,2,3,4,5],ymm9[6,7]604; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm10 = xmm3[0],xmm4[0],xmm3[1],xmm4[1]605; AVX2-FCP-NEXT:    vpermps %ymm7, %ymm6, %ymm6606; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm6 = xmm10[0,1],xmm6[2,3]607; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm9[4,5,6,7]608; AVX2-FCP-NEXT:    vbroadcastsd {{.*#+}} ymm9 = [1,5,1,5,1,5,1,5]609; AVX2-FCP-NEXT:    vpermps %ymm8, %ymm9, %ymm10610; AVX2-FCP-NEXT:    vpermps %ymm5, %ymm9, %ymm11611; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm10 = ymm11[0,1,2,3,4,5],ymm10[6,7]612; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm11 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]613; AVX2-FCP-NEXT:    vpermps %ymm2, %ymm9, %ymm9614; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm9 = xmm9[0,1],xmm11[2,3]615; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm9 = ymm9[0,1,2,3],ymm10[4,5,6,7]616; AVX2-FCP-NEXT:    vbroadcastsd {{.*#+}} ymm10 = [2,6,2,6,2,6,2,6]617; AVX2-FCP-NEXT:    vpermps %ymm8, %ymm10, %ymm11618; AVX2-FCP-NEXT:    vpermps %ymm5, %ymm10, %ymm12619; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm11 = ymm12[0,1,2,3,4,5],ymm11[6,7]620; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} xmm3 = xmm3[2],xmm4[2],xmm3[3],xmm4[3]621; AVX2-FCP-NEXT:    vpermps %ymm7, %ymm10, %ymm4622; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm3 = xmm3[0,1],xmm4[2,3]623; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm11[4,5,6,7]624; AVX2-FCP-NEXT:    vbroadcastsd {{.*#+}} ymm4 = [3,7,3,7,3,7,3,7]625; AVX2-FCP-NEXT:    vpermps %ymm8, %ymm4, %ymm7626; AVX2-FCP-NEXT:    vpermps %ymm5, %ymm4, %ymm5627; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3,4,5],ymm7[6,7]628; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]629; AVX2-FCP-NEXT:    vpermps %ymm2, %ymm4, %ymm1630; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3]631; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm5[4,5,6,7]632; AVX2-FCP-NEXT:    vmovaps %ymm0, 96(%rax)633; AVX2-FCP-NEXT:    vmovaps %ymm3, 64(%rax)634; AVX2-FCP-NEXT:    vmovaps %ymm9, 32(%rax)635; AVX2-FCP-NEXT:    vmovaps %ymm6, (%rax)636; AVX2-FCP-NEXT:    vzeroupper637; AVX2-FCP-NEXT:    retq638;639; AVX512-LABEL: store_i32_stride8_vf4:640; AVX512:       # %bb.0:641; AVX512-NEXT:    movq {{[0-9]+}}(%rsp), %rax642; AVX512-NEXT:    movq {{[0-9]+}}(%rsp), %r10643; AVX512-NEXT:    movq {{[0-9]+}}(%rsp), %r11644; AVX512-NEXT:    vmovdqa (%rdi), %xmm0645; AVX512-NEXT:    vmovdqa (%rdx), %xmm1646; AVX512-NEXT:    vmovdqa (%r8), %xmm2647; AVX512-NEXT:    vmovdqa (%r11), %xmm3648; AVX512-NEXT:    vinserti128 $1, (%rcx), %ymm1, %ymm1649; AVX512-NEXT:    vinserti128 $1, (%rsi), %ymm0, %ymm0650; AVX512-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0651; AVX512-NEXT:    vinserti128 $1, (%r10), %ymm3, %ymm1652; AVX512-NEXT:    vinserti128 $1, (%r9), %ymm2, %ymm2653; AVX512-NEXT:    vinserti64x4 $1, %ymm1, %zmm2, %zmm1654; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [0,4,8,12,16,20,24,28,1,5,9,13,17,21,25,29]655; AVX512-NEXT:    vpermi2d %zmm1, %zmm0, %zmm2656; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm3 = [2,6,10,14,18,22,26,30,3,7,11,15,19,23,27,31]657; AVX512-NEXT:    vpermi2d %zmm1, %zmm0, %zmm3658; AVX512-NEXT:    vmovdqa64 %zmm3, 64(%rax)659; AVX512-NEXT:    vmovdqa64 %zmm2, (%rax)660; AVX512-NEXT:    vzeroupper661; AVX512-NEXT:    retq662;663; AVX512-FCP-LABEL: store_i32_stride8_vf4:664; AVX512-FCP:       # %bb.0:665; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax666; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r10667; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r11668; AVX512-FCP-NEXT:    vmovdqa (%rdi), %xmm0669; AVX512-FCP-NEXT:    vmovdqa (%rdx), %xmm1670; AVX512-FCP-NEXT:    vmovdqa (%r8), %xmm2671; AVX512-FCP-NEXT:    vmovdqa (%r11), %xmm3672; AVX512-FCP-NEXT:    vinserti128 $1, (%rcx), %ymm1, %ymm1673; AVX512-FCP-NEXT:    vinserti128 $1, (%rsi), %ymm0, %ymm0674; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0675; AVX512-FCP-NEXT:    vinserti128 $1, (%r10), %ymm3, %ymm1676; AVX512-FCP-NEXT:    vinserti128 $1, (%r9), %ymm2, %ymm2677; AVX512-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm2, %zmm1678; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [0,4,8,12,16,20,24,28,1,5,9,13,17,21,25,29]679; AVX512-FCP-NEXT:    vpermi2d %zmm1, %zmm0, %zmm2680; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm3 = [2,6,10,14,18,22,26,30,3,7,11,15,19,23,27,31]681; AVX512-FCP-NEXT:    vpermi2d %zmm1, %zmm0, %zmm3682; AVX512-FCP-NEXT:    vmovdqa64 %zmm3, 64(%rax)683; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, (%rax)684; AVX512-FCP-NEXT:    vzeroupper685; AVX512-FCP-NEXT:    retq686;687; AVX512DQ-LABEL: store_i32_stride8_vf4:688; AVX512DQ:       # %bb.0:689; AVX512DQ-NEXT:    movq {{[0-9]+}}(%rsp), %rax690; AVX512DQ-NEXT:    movq {{[0-9]+}}(%rsp), %r10691; AVX512DQ-NEXT:    movq {{[0-9]+}}(%rsp), %r11692; AVX512DQ-NEXT:    vmovdqa (%rdi), %xmm0693; AVX512DQ-NEXT:    vmovdqa (%rdx), %xmm1694; AVX512DQ-NEXT:    vmovdqa (%r8), %xmm2695; AVX512DQ-NEXT:    vmovdqa (%r11), %xmm3696; AVX512DQ-NEXT:    vinserti128 $1, (%rcx), %ymm1, %ymm1697; AVX512DQ-NEXT:    vinserti128 $1, (%rsi), %ymm0, %ymm0698; AVX512DQ-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0699; AVX512DQ-NEXT:    vinserti128 $1, (%r10), %ymm3, %ymm1700; AVX512DQ-NEXT:    vinserti128 $1, (%r9), %ymm2, %ymm2701; AVX512DQ-NEXT:    vinserti64x4 $1, %ymm1, %zmm2, %zmm1702; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [0,4,8,12,16,20,24,28,1,5,9,13,17,21,25,29]703; AVX512DQ-NEXT:    vpermi2d %zmm1, %zmm0, %zmm2704; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm3 = [2,6,10,14,18,22,26,30,3,7,11,15,19,23,27,31]705; AVX512DQ-NEXT:    vpermi2d %zmm1, %zmm0, %zmm3706; AVX512DQ-NEXT:    vmovdqa64 %zmm3, 64(%rax)707; AVX512DQ-NEXT:    vmovdqa64 %zmm2, (%rax)708; AVX512DQ-NEXT:    vzeroupper709; AVX512DQ-NEXT:    retq710;711; AVX512DQ-FCP-LABEL: store_i32_stride8_vf4:712; AVX512DQ-FCP:       # %bb.0:713; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax714; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r10715; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r11716; AVX512DQ-FCP-NEXT:    vmovdqa (%rdi), %xmm0717; AVX512DQ-FCP-NEXT:    vmovdqa (%rdx), %xmm1718; AVX512DQ-FCP-NEXT:    vmovdqa (%r8), %xmm2719; AVX512DQ-FCP-NEXT:    vmovdqa (%r11), %xmm3720; AVX512DQ-FCP-NEXT:    vinserti128 $1, (%rcx), %ymm1, %ymm1721; AVX512DQ-FCP-NEXT:    vinserti128 $1, (%rsi), %ymm0, %ymm0722; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0723; AVX512DQ-FCP-NEXT:    vinserti128 $1, (%r10), %ymm3, %ymm1724; AVX512DQ-FCP-NEXT:    vinserti128 $1, (%r9), %ymm2, %ymm2725; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm2, %zmm1726; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [0,4,8,12,16,20,24,28,1,5,9,13,17,21,25,29]727; AVX512DQ-FCP-NEXT:    vpermi2d %zmm1, %zmm0, %zmm2728; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm3 = [2,6,10,14,18,22,26,30,3,7,11,15,19,23,27,31]729; AVX512DQ-FCP-NEXT:    vpermi2d %zmm1, %zmm0, %zmm3730; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm3, 64(%rax)731; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm2, (%rax)732; AVX512DQ-FCP-NEXT:    vzeroupper733; AVX512DQ-FCP-NEXT:    retq734;735; AVX512BW-LABEL: store_i32_stride8_vf4:736; AVX512BW:       # %bb.0:737; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax738; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %r10739; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %r11740; AVX512BW-NEXT:    vmovdqa (%rdi), %xmm0741; AVX512BW-NEXT:    vmovdqa (%rdx), %xmm1742; AVX512BW-NEXT:    vmovdqa (%r8), %xmm2743; AVX512BW-NEXT:    vmovdqa (%r11), %xmm3744; AVX512BW-NEXT:    vinserti128 $1, (%rcx), %ymm1, %ymm1745; AVX512BW-NEXT:    vinserti128 $1, (%rsi), %ymm0, %ymm0746; AVX512BW-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0747; AVX512BW-NEXT:    vinserti128 $1, (%r10), %ymm3, %ymm1748; AVX512BW-NEXT:    vinserti128 $1, (%r9), %ymm2, %ymm2749; AVX512BW-NEXT:    vinserti64x4 $1, %ymm1, %zmm2, %zmm1750; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [0,4,8,12,16,20,24,28,1,5,9,13,17,21,25,29]751; AVX512BW-NEXT:    vpermi2d %zmm1, %zmm0, %zmm2752; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm3 = [2,6,10,14,18,22,26,30,3,7,11,15,19,23,27,31]753; AVX512BW-NEXT:    vpermi2d %zmm1, %zmm0, %zmm3754; AVX512BW-NEXT:    vmovdqa64 %zmm3, 64(%rax)755; AVX512BW-NEXT:    vmovdqa64 %zmm2, (%rax)756; AVX512BW-NEXT:    vzeroupper757; AVX512BW-NEXT:    retq758;759; AVX512BW-FCP-LABEL: store_i32_stride8_vf4:760; AVX512BW-FCP:       # %bb.0:761; AVX512BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax762; AVX512BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r10763; AVX512BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r11764; AVX512BW-FCP-NEXT:    vmovdqa (%rdi), %xmm0765; AVX512BW-FCP-NEXT:    vmovdqa (%rdx), %xmm1766; AVX512BW-FCP-NEXT:    vmovdqa (%r8), %xmm2767; AVX512BW-FCP-NEXT:    vmovdqa (%r11), %xmm3768; AVX512BW-FCP-NEXT:    vinserti128 $1, (%rcx), %ymm1, %ymm1769; AVX512BW-FCP-NEXT:    vinserti128 $1, (%rsi), %ymm0, %ymm0770; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0771; AVX512BW-FCP-NEXT:    vinserti128 $1, (%r10), %ymm3, %ymm1772; AVX512BW-FCP-NEXT:    vinserti128 $1, (%r9), %ymm2, %ymm2773; AVX512BW-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm2, %zmm1774; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [0,4,8,12,16,20,24,28,1,5,9,13,17,21,25,29]775; AVX512BW-FCP-NEXT:    vpermi2d %zmm1, %zmm0, %zmm2776; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm3 = [2,6,10,14,18,22,26,30,3,7,11,15,19,23,27,31]777; AVX512BW-FCP-NEXT:    vpermi2d %zmm1, %zmm0, %zmm3778; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm3, 64(%rax)779; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm2, (%rax)780; AVX512BW-FCP-NEXT:    vzeroupper781; AVX512BW-FCP-NEXT:    retq782;783; AVX512DQ-BW-LABEL: store_i32_stride8_vf4:784; AVX512DQ-BW:       # %bb.0:785; AVX512DQ-BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax786; AVX512DQ-BW-NEXT:    movq {{[0-9]+}}(%rsp), %r10787; AVX512DQ-BW-NEXT:    movq {{[0-9]+}}(%rsp), %r11788; AVX512DQ-BW-NEXT:    vmovdqa (%rdi), %xmm0789; AVX512DQ-BW-NEXT:    vmovdqa (%rdx), %xmm1790; AVX512DQ-BW-NEXT:    vmovdqa (%r8), %xmm2791; AVX512DQ-BW-NEXT:    vmovdqa (%r11), %xmm3792; AVX512DQ-BW-NEXT:    vinserti128 $1, (%rcx), %ymm1, %ymm1793; AVX512DQ-BW-NEXT:    vinserti128 $1, (%rsi), %ymm0, %ymm0794; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0795; AVX512DQ-BW-NEXT:    vinserti128 $1, (%r10), %ymm3, %ymm1796; AVX512DQ-BW-NEXT:    vinserti128 $1, (%r9), %ymm2, %ymm2797; AVX512DQ-BW-NEXT:    vinserti64x4 $1, %ymm1, %zmm2, %zmm1798; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [0,4,8,12,16,20,24,28,1,5,9,13,17,21,25,29]799; AVX512DQ-BW-NEXT:    vpermi2d %zmm1, %zmm0, %zmm2800; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm3 = [2,6,10,14,18,22,26,30,3,7,11,15,19,23,27,31]801; AVX512DQ-BW-NEXT:    vpermi2d %zmm1, %zmm0, %zmm3802; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm3, 64(%rax)803; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm2, (%rax)804; AVX512DQ-BW-NEXT:    vzeroupper805; AVX512DQ-BW-NEXT:    retq806;807; AVX512DQ-BW-FCP-LABEL: store_i32_stride8_vf4:808; AVX512DQ-BW-FCP:       # %bb.0:809; AVX512DQ-BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax810; AVX512DQ-BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r10811; AVX512DQ-BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r11812; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%rdi), %xmm0813; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%rdx), %xmm1814; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%r8), %xmm2815; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%r11), %xmm3816; AVX512DQ-BW-FCP-NEXT:    vinserti128 $1, (%rcx), %ymm1, %ymm1817; AVX512DQ-BW-FCP-NEXT:    vinserti128 $1, (%rsi), %ymm0, %ymm0818; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0819; AVX512DQ-BW-FCP-NEXT:    vinserti128 $1, (%r10), %ymm3, %ymm1820; AVX512DQ-BW-FCP-NEXT:    vinserti128 $1, (%r9), %ymm2, %ymm2821; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, %ymm1, %zmm2, %zmm1822; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [0,4,8,12,16,20,24,28,1,5,9,13,17,21,25,29]823; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm1, %zmm0, %zmm2824; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm3 = [2,6,10,14,18,22,26,30,3,7,11,15,19,23,27,31]825; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm1, %zmm0, %zmm3826; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm3, 64(%rax)827; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm2, (%rax)828; AVX512DQ-BW-FCP-NEXT:    vzeroupper829; AVX512DQ-BW-FCP-NEXT:    retq830  %in.vec0 = load <4 x i32>, ptr %in.vecptr0, align 64831  %in.vec1 = load <4 x i32>, ptr %in.vecptr1, align 64832  %in.vec2 = load <4 x i32>, ptr %in.vecptr2, align 64833  %in.vec3 = load <4 x i32>, ptr %in.vecptr3, align 64834  %in.vec4 = load <4 x i32>, ptr %in.vecptr4, align 64835  %in.vec5 = load <4 x i32>, ptr %in.vecptr5, align 64836  %in.vec6 = load <4 x i32>, ptr %in.vecptr6, align 64837  %in.vec7 = load <4 x i32>, ptr %in.vecptr7, align 64838  %1 = shufflevector <4 x i32> %in.vec0, <4 x i32> %in.vec1, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>839  %2 = shufflevector <4 x i32> %in.vec2, <4 x i32> %in.vec3, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>840  %3 = shufflevector <4 x i32> %in.vec4, <4 x i32> %in.vec5, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>841  %4 = shufflevector <4 x i32> %in.vec6, <4 x i32> %in.vec7, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>842  %5 = shufflevector <8 x i32> %1, <8 x i32> %2, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>843  %6 = shufflevector <8 x i32> %3, <8 x i32> %4, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>844  %7 = shufflevector <16 x i32> %5, <16 x i32> %6, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>845  %interleaved.vec = shufflevector <32 x i32> %7, <32 x i32> poison, <32 x i32> <i32 0, i32 4, i32 8, i32 12, i32 16, i32 20, i32 24, i32 28, i32 1, i32 5, i32 9, i32 13, i32 17, i32 21, i32 25, i32 29, i32 2, i32 6, i32 10, i32 14, i32 18, i32 22, i32 26, i32 30, i32 3, i32 7, i32 11, i32 15, i32 19, i32 23, i32 27, i32 31>846  store <32 x i32> %interleaved.vec, ptr %out.vec, align 64847  ret void848}849 850define void @store_i32_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 {851; SSE-LABEL: store_i32_stride8_vf8:852; SSE:       # %bb.0:853; SSE-NEXT:    subq $72, %rsp854; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %rax855; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %r10856; SSE-NEXT:    movaps (%rdi), %xmm8857; SSE-NEXT:    movaps (%rsi), %xmm0858; SSE-NEXT:    movaps (%rdx), %xmm9859; SSE-NEXT:    movaps (%rcx), %xmm3860; SSE-NEXT:    movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill861; SSE-NEXT:    movaps (%r8), %xmm15862; SSE-NEXT:    movaps 16(%r8), %xmm10863; SSE-NEXT:    movaps (%r9), %xmm1864; SSE-NEXT:    movaps (%r10), %xmm14865; SSE-NEXT:    movaps 16(%r10), %xmm12866; SSE-NEXT:    movaps (%rax), %xmm4867; SSE-NEXT:    movaps 16(%rax), %xmm7868; SSE-NEXT:    movaps %xmm4, %xmm2869; SSE-NEXT:    movaps %xmm4, %xmm11870; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm14[0]871; SSE-NEXT:    movaps %xmm15, %xmm4872; SSE-NEXT:    unpcklps {{.*#+}} xmm4 = xmm4[0],xmm1[0],xmm4[1],xmm1[1]873; SSE-NEXT:    movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill874; SSE-NEXT:    shufps {{.*#+}} xmm4 = xmm4[0,1],xmm2[2,0]875; SSE-NEXT:    movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill876; SSE-NEXT:    movaps %xmm3, %xmm2877; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm9[0]878; SSE-NEXT:    movaps %xmm8, %xmm4879; SSE-NEXT:    unpcklps {{.*#+}} xmm4 = xmm4[0],xmm0[0],xmm4[1],xmm0[1]880; SSE-NEXT:    movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill881; SSE-NEXT:    shufps {{.*#+}} xmm4 = xmm4[0,1],xmm2[2,0]882; SSE-NEXT:    movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill883; SSE-NEXT:    movaps %xmm14, %xmm2884; SSE-NEXT:    unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm11[1]885; SSE-NEXT:    movaps %xmm11, %xmm4886; SSE-NEXT:    unpckhps {{.*#+}} xmm15 = xmm15[2],xmm1[2],xmm15[3],xmm1[3]887; SSE-NEXT:    movaps %xmm15, %xmm1888; SSE-NEXT:    shufps {{.*#+}} xmm1 = xmm1[0,1],xmm2[0,2]889; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill890; SSE-NEXT:    movaps %xmm9, %xmm1891; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm3[1]892; SSE-NEXT:    unpckhps {{.*#+}} xmm8 = xmm8[2],xmm0[2],xmm8[3],xmm0[3]893; SSE-NEXT:    movaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill894; SSE-NEXT:    movaps %xmm8, %xmm0895; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[0,1],xmm1[0,2]896; SSE-NEXT:    movaps %xmm0, (%rsp) # 16-byte Spill897; SSE-NEXT:    movaps %xmm7, %xmm0898; SSE-NEXT:    movaps %xmm7, %xmm3899; SSE-NEXT:    movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill900; SSE-NEXT:    movlhps {{.*#+}} xmm0 = xmm0[0],xmm12[0]901; SSE-NEXT:    movaps 16(%r9), %xmm7902; SSE-NEXT:    movaps %xmm10, %xmm13903; SSE-NEXT:    unpcklps {{.*#+}} xmm13 = xmm13[0],xmm7[0],xmm13[1],xmm7[1]904; SSE-NEXT:    movaps %xmm13, %xmm1905; SSE-NEXT:    shufps {{.*#+}} xmm1 = xmm1[0,1],xmm0[2,0]906; SSE-NEXT:    movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill907; SSE-NEXT:    movaps 16(%rdx), %xmm6908; SSE-NEXT:    movaps 16(%rcx), %xmm8909; SSE-NEXT:    movaps %xmm8, %xmm11910; SSE-NEXT:    movlhps {{.*#+}} xmm11 = xmm11[0],xmm6[0]911; SSE-NEXT:    movaps 16(%rdi), %xmm2912; SSE-NEXT:    movaps 16(%rsi), %xmm1913; SSE-NEXT:    movaps %xmm2, %xmm0914; SSE-NEXT:    unpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]915; SSE-NEXT:    movaps %xmm0, %xmm5916; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[0,1],xmm11[2,0]917; SSE-NEXT:    movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill918; SSE-NEXT:    unpckhps {{.*#+}} xmm10 = xmm10[2],xmm7[2],xmm10[3],xmm7[3]919; SSE-NEXT:    movaps %xmm12, %xmm7920; SSE-NEXT:    unpckhpd {{.*#+}} xmm7 = xmm7[1],xmm3[1]921; SSE-NEXT:    movaps %xmm10, %xmm3922; SSE-NEXT:    shufps {{.*#+}} xmm3 = xmm3[0,1],xmm7[0,2]923; SSE-NEXT:    movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill924; SSE-NEXT:    unpckhps {{.*#+}} xmm2 = xmm2[2],xmm1[2],xmm2[3],xmm1[3]925; SSE-NEXT:    movaps %xmm6, %xmm1926; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm8[1]927; SSE-NEXT:    movaps %xmm2, %xmm7928; SSE-NEXT:    shufps {{.*#+}} xmm7 = xmm7[0,1],xmm1[0,2]929; SSE-NEXT:    movaps %xmm4, %xmm1930; SSE-NEXT:    shufps {{.*#+}} xmm1 = xmm1[1,1],xmm14[1,1]931; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload932; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[2,3],xmm1[2,0]933; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload934; SSE-NEXT:    movaps %xmm3, %xmm1935; SSE-NEXT:    shufps {{.*#+}} xmm1 = xmm1[1,1],xmm9[1,1]936; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload937; SSE-NEXT:    shufps {{.*#+}} xmm11 = xmm11[2,3],xmm1[2,0]938; SSE-NEXT:    movaps %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill939; SSE-NEXT:    shufps {{.*#+}} xmm14 = xmm14[3,3],xmm4[3,3]940; SSE-NEXT:    shufps {{.*#+}} xmm15 = xmm15[2,3],xmm14[0,2]941; SSE-NEXT:    shufps {{.*#+}} xmm9 = xmm9[3,3],xmm3[3,3]942; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload943; SSE-NEXT:    shufps {{.*#+}} xmm14 = xmm14[2,3],xmm9[0,2]944; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm9 # 16-byte Reload945; SSE-NEXT:    movaps %xmm9, %xmm1946; SSE-NEXT:    shufps {{.*#+}} xmm1 = xmm1[1,1],xmm12[1,1]947; SSE-NEXT:    shufps {{.*#+}} xmm13 = xmm13[2,3],xmm1[2,0]948; SSE-NEXT:    movaps %xmm8, %xmm1949; SSE-NEXT:    shufps {{.*#+}} xmm1 = xmm1[1,1],xmm6[1,1]950; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[2,3],xmm1[2,0]951; SSE-NEXT:    shufps {{.*#+}} xmm12 = xmm12[3,3],xmm9[3,3]952; SSE-NEXT:    shufps {{.*#+}} xmm10 = xmm10[2,3],xmm12[0,2]953; SSE-NEXT:    shufps {{.*#+}} xmm6 = xmm6[3,3],xmm8[3,3]954; SSE-NEXT:    shufps {{.*#+}} xmm2 = xmm2[2,3],xmm6[0,2]955; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %rax956; SSE-NEXT:    movaps %xmm2, 224(%rax)957; SSE-NEXT:    movaps %xmm10, 240(%rax)958; SSE-NEXT:    movaps %xmm0, 160(%rax)959; SSE-NEXT:    movaps %xmm13, 176(%rax)960; SSE-NEXT:    movaps %xmm14, 96(%rax)961; SSE-NEXT:    movaps %xmm15, 112(%rax)962; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload963; SSE-NEXT:    movaps %xmm0, 32(%rax)964; SSE-NEXT:    movaps %xmm5, 48(%rax)965; SSE-NEXT:    movaps %xmm7, 192(%rax)966; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload967; SSE-NEXT:    movaps %xmm0, 208(%rax)968; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload969; SSE-NEXT:    movaps %xmm0, 128(%rax)970; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload971; SSE-NEXT:    movaps %xmm0, 144(%rax)972; SSE-NEXT:    movaps (%rsp), %xmm0 # 16-byte Reload973; SSE-NEXT:    movaps %xmm0, 64(%rax)974; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload975; SSE-NEXT:    movaps %xmm0, 80(%rax)976; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload977; SSE-NEXT:    movaps %xmm0, (%rax)978; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload979; SSE-NEXT:    movaps %xmm0, 16(%rax)980; SSE-NEXT:    addq $72, %rsp981; SSE-NEXT:    retq982;983; AVX-LABEL: store_i32_stride8_vf8:984; AVX:       # %bb.0:985; AVX-NEXT:    subq $120, %rsp986; AVX-NEXT:    movq {{[0-9]+}}(%rsp), %rax987; AVX-NEXT:    movq {{[0-9]+}}(%rsp), %r10988; AVX-NEXT:    vmovaps (%r8), %ymm9989; AVX-NEXT:    vmovaps (%rcx), %xmm4990; AVX-NEXT:    vmovaps (%rdx), %xmm3991; AVX-NEXT:    vunpcklps {{.*#+}} xmm2 = xmm3[0],xmm4[0],xmm3[1],xmm4[1]992; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill993; AVX-NEXT:    vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill994; AVX-NEXT:    vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill995; AVX-NEXT:    vmovaps (%rsi), %xmm11996; AVX-NEXT:    vmovaps (%rdi), %xmm12997; AVX-NEXT:    vshufps {{.*#+}} xmm0 = xmm12[1,1,1,1]998; AVX-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0],xmm11[1],xmm0[2,3]999; AVX-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm2[2,3]1000; AVX-NEXT:    vmovaps (%rax), %xmm61001; AVX-NEXT:    vmovaps (%r10), %xmm51002; AVX-NEXT:    vunpcklps {{.*#+}} xmm13 = xmm5[0],xmm6[0],xmm5[1],xmm6[1]1003; AVX-NEXT:    vmovaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1004; AVX-NEXT:    vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1005; AVX-NEXT:    vinsertf128 $1, %xmm13, %ymm0, %ymm11006; AVX-NEXT:    vmovaps (%r9), %xmm141007; AVX-NEXT:    vmovaps (%r8), %xmm151008; AVX-NEXT:    vshufps {{.*#+}} xmm2 = xmm15[1,1,1,1]1009; AVX-NEXT:    vblendps {{.*#+}} xmm2 = xmm2[0],xmm14[1],xmm2[2,3]1010; AVX-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm21011; AVX-NEXT:    vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]1012; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]1013; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill1014; AVX-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm3[2],xmm4[2],xmm3[3],xmm4[3]1015; AVX-NEXT:    vunpckhps {{.*#+}} xmm1 = xmm5[2],xmm6[2],xmm5[3],xmm6[3]1016; AVX-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm11017; AVX-NEXT:    vunpckhps {{.*#+}} xmm2 = xmm15[2],xmm14[2],xmm15[3],xmm14[3]1018; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1019; AVX-NEXT:    vshufps {{.*#+}} xmm2 = xmm2[2,3,2,3]1020; AVX-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm21021; AVX-NEXT:    vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]1022; AVX-NEXT:    vunpckhps {{.*#+}} xmm2 = xmm12[2],xmm11[2],xmm12[3],xmm11[3]1023; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1024; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm2[1],xmm0[1]1025; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]1026; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill1027; AVX-NEXT:    vmovaps (%r9), %ymm61028; AVX-NEXT:    vmovaps (%r10), %ymm51029; AVX-NEXT:    vmovaps (%rax), %ymm41030; AVX-NEXT:    vunpcklps {{.*#+}} ymm0 = ymm5[0],ymm4[0],ymm5[1],ymm4[1],ymm5[4],ymm4[4],ymm5[5],ymm4[5]1031; AVX-NEXT:    vshufps {{.*#+}} ymm1 = ymm6[1,0],ymm9[1,0],ymm6[5,4],ymm9[5,4]1032; AVX-NEXT:    vshufps {{.*#+}} ymm8 = ymm1[2,0],ymm0[2,3],ymm1[6,4],ymm0[6,7]1033; AVX-NEXT:    vmovaps (%rdx), %ymm31034; AVX-NEXT:    vmovaps (%rcx), %ymm21035; AVX-NEXT:    vunpcklps {{.*#+}} ymm0 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[4],ymm2[4],ymm3[5],ymm2[5]1036; AVX-NEXT:    vextractf128 $1, %ymm0, %xmm71037; AVX-NEXT:    vmovaps (%rdi), %ymm11038; AVX-NEXT:    vmovaps (%rsi), %ymm01039; AVX-NEXT:    vshufps {{.*#+}} ymm10 = ymm0[1,0],ymm1[1,0],ymm0[5,4],ymm1[5,4]1040; AVX-NEXT:    vextractf128 $1, %ymm10, %xmm101041; AVX-NEXT:    vshufps {{.*#+}} xmm7 = xmm10[2,0],xmm7[2,3]1042; AVX-NEXT:    vblendps {{.*#+}} ymm7 = ymm7[0,1,2,3],ymm8[4,5,6,7]1043; AVX-NEXT:    vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill1044; AVX-NEXT:    vunpckhps {{.*#+}} ymm7 = ymm5[2],ymm4[2],ymm5[3],ymm4[3],ymm5[6],ymm4[6],ymm5[7],ymm4[7]1045; AVX-NEXT:    vshufps {{.*#+}} ymm8 = ymm6[3,0],ymm9[3,0],ymm6[7,4],ymm9[7,4]1046; AVX-NEXT:    vshufps {{.*#+}} ymm7 = ymm8[2,0],ymm7[2,3],ymm8[6,4],ymm7[6,7]1047; AVX-NEXT:    vunpckhps {{.*#+}} ymm8 = ymm3[2],ymm2[2],ymm3[3],ymm2[3],ymm3[6],ymm2[6],ymm3[7],ymm2[7]1048; AVX-NEXT:    vextractf128 $1, %ymm8, %xmm81049; AVX-NEXT:    vshufps {{.*#+}} ymm10 = ymm0[3,0],ymm1[3,0],ymm0[7,4],ymm1[7,4]1050; AVX-NEXT:    vextractf128 $1, %ymm10, %xmm101051; AVX-NEXT:    vshufps {{.*#+}} xmm8 = xmm10[2,0],xmm8[2,3]1052; AVX-NEXT:    vblendps {{.*#+}} ymm7 = ymm8[0,1,2,3],ymm7[4,5,6,7]1053; AVX-NEXT:    vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill1054; AVX-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm12[0],xmm11[0],xmm12[1],xmm11[1]1055; AVX-NEXT:    vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm7, %xmm7 # 16-byte Folded Reload1056; AVX-NEXT:    # xmm7 = xmm7[0],mem[0]1057; AVX-NEXT:    vunpcklps {{.*#+}} xmm8 = xmm15[0],xmm14[0],xmm15[1],xmm14[1]1058; AVX-NEXT:    vinsertf128 $1, %xmm8, %ymm0, %ymm81059; AVX-NEXT:    vshufps {{.*#+}} xmm10 = xmm13[0,1,0,1]1060; AVX-NEXT:    vinsertf128 $1, %xmm10, %ymm0, %ymm101061; AVX-NEXT:    vblendps {{.*#+}} ymm8 = ymm8[0,1,2,3,4,5],ymm10[6,7]1062; AVX-NEXT:    vblendps {{.*#+}} ymm7 = ymm7[0,1,2,3],ymm8[4,5,6,7]1063; AVX-NEXT:    vunpckhpd {{.*#+}} ymm8 = ymm4[1],ymm5[1],ymm4[3],ymm5[3]1064; AVX-NEXT:    vunpckhps {{.*#+}} ymm10 = ymm9[2],ymm6[2],ymm9[3],ymm6[3],ymm9[6],ymm6[6],ymm9[7],ymm6[7]1065; AVX-NEXT:    vshufps {{.*#+}} ymm8 = ymm10[0,1],ymm8[2,0],ymm10[4,5],ymm8[6,4]1066; AVX-NEXT:    vunpckhpd {{.*#+}} ymm10 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]1067; AVX-NEXT:    vextractf128 $1, %ymm10, %xmm101068; AVX-NEXT:    vunpckhps {{.*#+}} ymm11 = ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[6],ymm0[6],ymm1[7],ymm0[7]1069; AVX-NEXT:    vextractf128 $1, %ymm11, %xmm111070; AVX-NEXT:    vshufps {{.*#+}} xmm10 = xmm11[0,1],xmm10[2,0]1071; AVX-NEXT:    vblendps {{.*#+}} ymm8 = ymm10[0,1,2,3],ymm8[4,5,6,7]1072; AVX-NEXT:    vunpcklpd {{.*#+}} ymm4 = ymm4[0],ymm5[0],ymm4[2],ymm5[2]1073; AVX-NEXT:    vunpcklps {{.*#+}} ymm5 = ymm9[0],ymm6[0],ymm9[1],ymm6[1],ymm9[4],ymm6[4],ymm9[5],ymm6[5]1074; AVX-NEXT:    vshufps {{.*#+}} ymm4 = ymm5[0,1],ymm4[2,0],ymm5[4,5],ymm4[6,4]1075; AVX-NEXT:    vunpcklpd {{.*#+}} ymm2 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]1076; AVX-NEXT:    vunpcklps {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[4],ymm0[4],ymm1[5],ymm0[5]1077; AVX-NEXT:    vextractf128 $1, %ymm2, %xmm11078; AVX-NEXT:    vextractf128 $1, %ymm0, %xmm01079; AVX-NEXT:    vshufps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,0]1080; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm4[4,5,6,7]1081; AVX-NEXT:    vpermilps $170, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload1082; AVX-NEXT:    # xmm1 = mem[2,2,2,2]1083; AVX-NEXT:    vblendps $7, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload1084; AVX-NEXT:    # xmm1 = mem[0,1,2],xmm1[3]1085; AVX-NEXT:    vblendps $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload1086; AVX-NEXT:    # xmm1 = mem[0,1],xmm1[2,3]1087; AVX-NEXT:    vpermilps $170, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload1088; AVX-NEXT:    # xmm2 = mem[2,2,2,2]1089; AVX-NEXT:    vblendps $7, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm2 # 16-byte Folded Reload1090; AVX-NEXT:    # xmm2 = mem[0,1,2],xmm2[3]1091; AVX-NEXT:    vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm3 # 16-byte Folded Reload1092; AVX-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm21093; AVX-NEXT:    vblendps {{.*#+}} ymm2 = ymm3[0,1,2,3,4,5],ymm2[6,7]1094; AVX-NEXT:    vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]1095; AVX-NEXT:    movq {{[0-9]+}}(%rsp), %rax1096; AVX-NEXT:    vmovaps %ymm1, 64(%rax)1097; AVX-NEXT:    vmovaps %ymm0, 128(%rax)1098; AVX-NEXT:    vmovaps %ymm8, 192(%rax)1099; AVX-NEXT:    vmovaps %ymm7, (%rax)1100; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload1101; AVX-NEXT:    vmovaps %ymm0, 224(%rax)1102; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload1103; AVX-NEXT:    vmovaps %ymm0, 160(%rax)1104; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload1105; AVX-NEXT:    vmovaps %ymm0, 96(%rax)1106; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload1107; AVX-NEXT:    vmovaps %ymm0, 32(%rax)1108; AVX-NEXT:    addq $120, %rsp1109; AVX-NEXT:    vzeroupper1110; AVX-NEXT:    retq1111;1112; AVX2-LABEL: store_i32_stride8_vf8:1113; AVX2:       # %bb.0:1114; AVX2-NEXT:    movq {{[0-9]+}}(%rsp), %rax1115; AVX2-NEXT:    movq {{[0-9]+}}(%rsp), %r101116; AVX2-NEXT:    vmovaps (%rdi), %ymm31117; AVX2-NEXT:    vmovaps (%rsi), %ymm41118; AVX2-NEXT:    vmovaps (%rdx), %ymm11119; AVX2-NEXT:    vmovaps (%rcx), %ymm21120; AVX2-NEXT:    vmovaps (%r8), %ymm51121; AVX2-NEXT:    vmovaps (%r9), %ymm61122; AVX2-NEXT:    vmovaps (%r10), %ymm91123; AVX2-NEXT:    vmovaps (%rax), %ymm101124; AVX2-NEXT:    vunpcklps {{.*#+}} ymm7 = ymm9[0],ymm10[0],ymm9[1],ymm10[1],ymm9[4],ymm10[4],ymm9[5],ymm10[5]1125; AVX2-NEXT:    vbroadcastss 20(%r8), %ymm01126; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm6[5],ymm0[6,7]1127; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm7[6,7]1128; AVX2-NEXT:    vunpcklps {{.*#+}} ymm8 = ymm1[0],ymm2[0],ymm1[1],ymm2[1],ymm1[4],ymm2[4],ymm1[5],ymm2[5]1129; AVX2-NEXT:    vextractf128 $1, %ymm8, %xmm111130; AVX2-NEXT:    vshufps {{.*#+}} ymm12 = ymm3[1,1,1,1,5,5,5,5]1131; AVX2-NEXT:    vblendps {{.*#+}} ymm12 = ymm12[0],ymm4[1],ymm12[2,3,4],ymm4[5],ymm12[6,7]1132; AVX2-NEXT:    vextractf128 $1, %ymm12, %xmm121133; AVX2-NEXT:    vblendps {{.*#+}} xmm11 = xmm12[0,1],xmm11[2,3]1134; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm11[0,1,2,3],ymm0[4,5,6,7]1135; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill1136; AVX2-NEXT:    vbroadcastss 28(%r10), %ymm111137; AVX2-NEXT:    vunpckhps {{.*#+}} ymm10 = ymm11[2],ymm10[2],ymm11[3],ymm10[3],ymm11[6],ymm10[6],ymm11[7],ymm10[7]1138; AVX2-NEXT:    vunpckhps {{.*#+}} ymm11 = ymm5[2],ymm6[2],ymm5[3],ymm6[3],ymm5[6],ymm6[6],ymm5[7],ymm6[7]1139; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm10 = ymm11[1],ymm10[1],ymm11[3],ymm10[3]1140; AVX2-NEXT:    vunpckhps {{.*#+}} ymm2 = ymm1[2],ymm2[2],ymm1[3],ymm2[3],ymm1[6],ymm2[6],ymm1[7],ymm2[7]1141; AVX2-NEXT:    vextractf128 $1, %ymm2, %xmm11142; AVX2-NEXT:    vunpckhps {{.*#+}} ymm12 = ymm3[2],ymm4[2],ymm3[3],ymm4[3],ymm3[6],ymm4[6],ymm3[7],ymm4[7]1143; AVX2-NEXT:    vshufps {{.*#+}} ymm13 = ymm12[2,3,2,3,6,7,6,7]1144; AVX2-NEXT:    vextractf128 $1, %ymm13, %xmm131145; AVX2-NEXT:    vblendps {{.*#+}} xmm1 = xmm13[0,1],xmm1[2,3]1146; AVX2-NEXT:    vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm10[4,5,6,7]1147; AVX2-NEXT:    vbroadcastss 24(%rax), %ymm101148; AVX2-NEXT:    vblendps {{.*#+}} ymm9 = ymm9[0,1,2,3,4,5,6],ymm10[7]1149; AVX2-NEXT:    vblendps {{.*#+}} ymm9 = ymm11[0,1,2,3,4,5],ymm9[6,7]1150; AVX2-NEXT:    vextractf128 $1, %ymm12, %xmm101151; AVX2-NEXT:    vpermpd {{.*#+}} ymm2 = ymm2[2,2,2,2]1152; AVX2-NEXT:    vblendps {{.*#+}} xmm2 = xmm10[0,1],xmm2[2,3]1153; AVX2-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm9[4,5,6,7]1154; AVX2-NEXT:    vunpcklps {{.*#+}} ymm5 = ymm5[0],ymm6[0],ymm5[1],ymm6[1],ymm5[4],ymm6[4],ymm5[5],ymm6[5]1155; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm5 = ymm5[0],ymm7[0],ymm5[2],ymm7[2]1156; AVX2-NEXT:    vunpcklps {{.*#+}} ymm3 = ymm3[0],ymm4[0],ymm3[1],ymm4[1],ymm3[4],ymm4[4],ymm3[5],ymm4[5]1157; AVX2-NEXT:    vextractf128 $1, %ymm3, %xmm31158; AVX2-NEXT:    vpermpd {{.*#+}} ymm4 = ymm8[2,2,2,2]1159; AVX2-NEXT:    vblendps {{.*#+}} xmm3 = xmm3[0,1],xmm4[2,3]1160; AVX2-NEXT:    vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm5[4,5,6,7]1161; AVX2-NEXT:    vmovaps (%rcx), %xmm41162; AVX2-NEXT:    vmovaps (%rdx), %xmm51163; AVX2-NEXT:    vunpcklps {{.*#+}} xmm6 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]1164; AVX2-NEXT:    vmovaps (%rsi), %xmm91165; AVX2-NEXT:    vmovaps (%rdi), %xmm101166; AVX2-NEXT:    vshufps {{.*#+}} xmm7 = xmm10[1,1,1,1]1167; AVX2-NEXT:    vblendps {{.*#+}} xmm7 = xmm7[0],xmm9[1],xmm7[2,3]1168; AVX2-NEXT:    vblendps {{.*#+}} xmm8 = xmm7[0,1],xmm6[2,3]1169; AVX2-NEXT:    vmovaps (%rax), %xmm61170; AVX2-NEXT:    vmovaps (%r10), %xmm71171; AVX2-NEXT:    vunpcklps {{.*#+}} xmm11 = xmm7[0],xmm6[0],xmm7[1],xmm6[1]1172; AVX2-NEXT:    vinsertf128 $1, %xmm11, %ymm0, %ymm111173; AVX2-NEXT:    vmovaps (%r9), %xmm121174; AVX2-NEXT:    vmovaps (%r8), %xmm131175; AVX2-NEXT:    vshufps {{.*#+}} xmm14 = xmm13[1,1,1,1]1176; AVX2-NEXT:    vblendps {{.*#+}} xmm14 = xmm14[0],xmm12[1],xmm14[2,3]1177; AVX2-NEXT:    vinsertf128 $1, %xmm14, %ymm0, %ymm141178; AVX2-NEXT:    vblendps {{.*#+}} ymm11 = ymm14[0,1,2,3,4,5],ymm11[6,7]1179; AVX2-NEXT:    vblendps {{.*#+}} ymm8 = ymm8[0,1,2,3],ymm11[4,5,6,7]1180; AVX2-NEXT:    vunpckhps {{.*#+}} xmm11 = xmm7[2],xmm6[2],xmm7[3],xmm6[3]1181; AVX2-NEXT:    vinsertf128 $1, %xmm11, %ymm0, %ymm111182; AVX2-NEXT:    vunpckhps {{.*#+}} xmm14 = xmm13[2],xmm12[2],xmm13[3],xmm12[3]1183; AVX2-NEXT:    vshufps {{.*#+}} xmm15 = xmm14[2,3,2,3]1184; AVX2-NEXT:    vinsertf128 $1, %xmm15, %ymm0, %ymm151185; AVX2-NEXT:    vblendps {{.*#+}} ymm11 = ymm15[0,1,2,3,4,5],ymm11[6,7]1186; AVX2-NEXT:    vunpckhps {{.*#+}} xmm15 = xmm5[2],xmm4[2],xmm5[3],xmm4[3]1187; AVX2-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm10[2],xmm9[2],xmm10[3],xmm9[3]1188; AVX2-NEXT:    vunpckhpd {{.*#+}} xmm15 = xmm0[1],xmm15[1]1189; AVX2-NEXT:    vblendps {{.*#+}} ymm11 = ymm15[0,1,2,3],ymm11[4,5,6,7]1190; AVX2-NEXT:    vunpcklps {{.*#+}} xmm9 = xmm10[0],xmm9[0],xmm10[1],xmm9[1]1191; AVX2-NEXT:    vbroadcastss %xmm4, %xmm101192; AVX2-NEXT:    vbroadcastss %xmm5, %xmm151193; AVX2-NEXT:    vunpcklps {{.*#+}} xmm10 = xmm15[0],xmm10[0],xmm15[1],xmm10[1]1194; AVX2-NEXT:    vblendps {{.*#+}} xmm9 = xmm9[0,1],xmm10[2,3]1195; AVX2-NEXT:    vunpcklps {{.*#+}} xmm10 = xmm13[0],xmm12[0],xmm13[1],xmm12[1]1196; AVX2-NEXT:    vbroadcastss %xmm6, %xmm121197; AVX2-NEXT:    vbroadcastss %xmm7, %xmm131198; AVX2-NEXT:    vunpcklps {{.*#+}} xmm12 = xmm13[0],xmm12[0],xmm13[1],xmm12[1]1199; AVX2-NEXT:    vinsertf128 $1, %xmm10, %ymm0, %ymm101200; AVX2-NEXT:    vinsertf128 $1, %xmm12, %ymm0, %ymm121201; AVX2-NEXT:    vblendps {{.*#+}} ymm10 = ymm10[0,1,2,3,4,5],ymm12[6,7]1202; AVX2-NEXT:    vblendps {{.*#+}} ymm9 = ymm9[0,1,2,3],ymm10[4,5,6,7]1203; AVX2-NEXT:    vshufps {{.*#+}} xmm4 = xmm4[2,2,2,2]1204; AVX2-NEXT:    vblendps {{.*#+}} xmm4 = xmm5[0,1,2],xmm4[3]1205; AVX2-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm4[2,3]1206; AVX2-NEXT:    vshufps {{.*#+}} xmm4 = xmm6[2,2,2,2]1207; AVX2-NEXT:    vblendps {{.*#+}} xmm4 = xmm7[0,1,2],xmm4[3]1208; AVX2-NEXT:    vinsertf128 $1, %xmm14, %ymm0, %ymm51209; AVX2-NEXT:    vinsertf128 $1, %xmm4, %ymm0, %ymm41210; AVX2-NEXT:    vblendps {{.*#+}} ymm4 = ymm5[0,1,2,3,4,5],ymm4[6,7]1211; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm4[4,5,6,7]1212; AVX2-NEXT:    movq {{[0-9]+}}(%rsp), %rax1213; AVX2-NEXT:    vmovaps %ymm0, 64(%rax)1214; AVX2-NEXT:    vmovaps %ymm9, (%rax)1215; AVX2-NEXT:    vmovaps %ymm11, 96(%rax)1216; AVX2-NEXT:    vmovaps %ymm8, 32(%rax)1217; AVX2-NEXT:    vmovaps %ymm3, 128(%rax)1218; AVX2-NEXT:    vmovaps %ymm2, 192(%rax)1219; AVX2-NEXT:    vmovaps %ymm1, 224(%rax)1220; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload1221; AVX2-NEXT:    vmovaps %ymm0, 160(%rax)1222; AVX2-NEXT:    vzeroupper1223; AVX2-NEXT:    retq1224;1225; AVX2-FP-LABEL: store_i32_stride8_vf8:1226; AVX2-FP:       # %bb.0:1227; AVX2-FP-NEXT:    movq {{[0-9]+}}(%rsp), %rax1228; AVX2-FP-NEXT:    movq {{[0-9]+}}(%rsp), %r101229; AVX2-FP-NEXT:    vmovaps (%rdi), %ymm31230; AVX2-FP-NEXT:    vmovaps (%rsi), %ymm41231; AVX2-FP-NEXT:    vmovaps (%rdx), %ymm11232; AVX2-FP-NEXT:    vmovaps (%rcx), %ymm21233; AVX2-FP-NEXT:    vmovaps (%r8), %ymm51234; AVX2-FP-NEXT:    vmovaps (%r9), %ymm61235; AVX2-FP-NEXT:    vmovaps (%r10), %ymm91236; AVX2-FP-NEXT:    vmovaps (%rax), %ymm101237; AVX2-FP-NEXT:    vunpcklps {{.*#+}} ymm7 = ymm9[0],ymm10[0],ymm9[1],ymm10[1],ymm9[4],ymm10[4],ymm9[5],ymm10[5]1238; AVX2-FP-NEXT:    vbroadcastss 20(%r8), %ymm01239; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm6[5],ymm0[6,7]1240; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm7[6,7]1241; AVX2-FP-NEXT:    vunpcklps {{.*#+}} ymm8 = ymm1[0],ymm2[0],ymm1[1],ymm2[1],ymm1[4],ymm2[4],ymm1[5],ymm2[5]1242; AVX2-FP-NEXT:    vextractf128 $1, %ymm8, %xmm111243; AVX2-FP-NEXT:    vshufps {{.*#+}} ymm12 = ymm3[1,1,1,1,5,5,5,5]1244; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm12 = ymm12[0],ymm4[1],ymm12[2,3,4],ymm4[5],ymm12[6,7]1245; AVX2-FP-NEXT:    vextractf128 $1, %ymm12, %xmm121246; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm11 = xmm12[0,1],xmm11[2,3]1247; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm11[0,1,2,3],ymm0[4,5,6,7]1248; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill1249; AVX2-FP-NEXT:    vbroadcastss 28(%r10), %ymm111250; AVX2-FP-NEXT:    vunpckhps {{.*#+}} ymm10 = ymm11[2],ymm10[2],ymm11[3],ymm10[3],ymm11[6],ymm10[6],ymm11[7],ymm10[7]1251; AVX2-FP-NEXT:    vunpckhps {{.*#+}} ymm11 = ymm5[2],ymm6[2],ymm5[3],ymm6[3],ymm5[6],ymm6[6],ymm5[7],ymm6[7]1252; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm10 = ymm11[1],ymm10[1],ymm11[3],ymm10[3]1253; AVX2-FP-NEXT:    vunpckhps {{.*#+}} ymm2 = ymm1[2],ymm2[2],ymm1[3],ymm2[3],ymm1[6],ymm2[6],ymm1[7],ymm2[7]1254; AVX2-FP-NEXT:    vextractf128 $1, %ymm2, %xmm11255; AVX2-FP-NEXT:    vunpckhps {{.*#+}} ymm12 = ymm3[2],ymm4[2],ymm3[3],ymm4[3],ymm3[6],ymm4[6],ymm3[7],ymm4[7]1256; AVX2-FP-NEXT:    vshufps {{.*#+}} ymm13 = ymm12[2,3,2,3,6,7,6,7]1257; AVX2-FP-NEXT:    vextractf128 $1, %ymm13, %xmm131258; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm1 = xmm13[0,1],xmm1[2,3]1259; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm10[4,5,6,7]1260; AVX2-FP-NEXT:    vbroadcastss 24(%rax), %ymm101261; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm9 = ymm9[0,1,2,3,4,5,6],ymm10[7]1262; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm9 = ymm11[0,1,2,3,4,5],ymm9[6,7]1263; AVX2-FP-NEXT:    vextractf128 $1, %ymm12, %xmm101264; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm2 = ymm2[2,2,2,2]1265; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm2 = xmm10[0,1],xmm2[2,3]1266; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm9[4,5,6,7]1267; AVX2-FP-NEXT:    vunpcklps {{.*#+}} ymm5 = ymm5[0],ymm6[0],ymm5[1],ymm6[1],ymm5[4],ymm6[4],ymm5[5],ymm6[5]1268; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm5 = ymm5[0],ymm7[0],ymm5[2],ymm7[2]1269; AVX2-FP-NEXT:    vunpcklps {{.*#+}} ymm3 = ymm3[0],ymm4[0],ymm3[1],ymm4[1],ymm3[4],ymm4[4],ymm3[5],ymm4[5]1270; AVX2-FP-NEXT:    vextractf128 $1, %ymm3, %xmm31271; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm4 = ymm8[2,2,2,2]1272; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm3 = xmm3[0,1],xmm4[2,3]1273; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm5[4,5,6,7]1274; AVX2-FP-NEXT:    vmovaps (%rcx), %xmm41275; AVX2-FP-NEXT:    vmovaps (%rdx), %xmm51276; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm6 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]1277; AVX2-FP-NEXT:    vmovaps (%rsi), %xmm91278; AVX2-FP-NEXT:    vmovaps (%rdi), %xmm101279; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm7 = xmm10[1,1,1,1]1280; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm7 = xmm7[0],xmm9[1],xmm7[2,3]1281; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm8 = xmm7[0,1],xmm6[2,3]1282; AVX2-FP-NEXT:    vmovaps (%rax), %xmm61283; AVX2-FP-NEXT:    vmovaps (%r10), %xmm71284; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm11 = xmm7[0],xmm6[0],xmm7[1],xmm6[1]1285; AVX2-FP-NEXT:    vinsertf128 $1, %xmm11, %ymm0, %ymm111286; AVX2-FP-NEXT:    vmovaps (%r9), %xmm121287; AVX2-FP-NEXT:    vmovaps (%r8), %xmm131288; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm14 = xmm13[1,1,1,1]1289; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm14 = xmm14[0],xmm12[1],xmm14[2,3]1290; AVX2-FP-NEXT:    vinsertf128 $1, %xmm14, %ymm0, %ymm141291; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm11 = ymm14[0,1,2,3,4,5],ymm11[6,7]1292; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm8 = ymm8[0,1,2,3],ymm11[4,5,6,7]1293; AVX2-FP-NEXT:    vunpckhps {{.*#+}} xmm11 = xmm7[2],xmm6[2],xmm7[3],xmm6[3]1294; AVX2-FP-NEXT:    vinsertf128 $1, %xmm11, %ymm0, %ymm111295; AVX2-FP-NEXT:    vunpckhps {{.*#+}} xmm14 = xmm13[2],xmm12[2],xmm13[3],xmm12[3]1296; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm15 = xmm14[2,3,2,3]1297; AVX2-FP-NEXT:    vinsertf128 $1, %xmm15, %ymm0, %ymm151298; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm11 = ymm15[0,1,2,3,4,5],ymm11[6,7]1299; AVX2-FP-NEXT:    vunpckhps {{.*#+}} xmm15 = xmm5[2],xmm4[2],xmm5[3],xmm4[3]1300; AVX2-FP-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm10[2],xmm9[2],xmm10[3],xmm9[3]1301; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} xmm15 = xmm0[1],xmm15[1]1302; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm11 = ymm15[0,1,2,3],ymm11[4,5,6,7]1303; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm9 = xmm10[0],xmm9[0],xmm10[1],xmm9[1]1304; AVX2-FP-NEXT:    vbroadcastss %xmm4, %xmm101305; AVX2-FP-NEXT:    vbroadcastss %xmm5, %xmm151306; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm10 = xmm15[0],xmm10[0],xmm15[1],xmm10[1]1307; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm9 = xmm9[0,1],xmm10[2,3]1308; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm10 = xmm13[0],xmm12[0],xmm13[1],xmm12[1]1309; AVX2-FP-NEXT:    vbroadcastss %xmm6, %xmm121310; AVX2-FP-NEXT:    vbroadcastss %xmm7, %xmm131311; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm12 = xmm13[0],xmm12[0],xmm13[1],xmm12[1]1312; AVX2-FP-NEXT:    vinsertf128 $1, %xmm10, %ymm0, %ymm101313; AVX2-FP-NEXT:    vinsertf128 $1, %xmm12, %ymm0, %ymm121314; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm10 = ymm10[0,1,2,3,4,5],ymm12[6,7]1315; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm9 = ymm9[0,1,2,3],ymm10[4,5,6,7]1316; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm4 = xmm4[2,2,2,2]1317; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm4 = xmm5[0,1,2],xmm4[3]1318; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm4[2,3]1319; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm4 = xmm6[2,2,2,2]1320; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm4 = xmm7[0,1,2],xmm4[3]1321; AVX2-FP-NEXT:    vinsertf128 $1, %xmm14, %ymm0, %ymm51322; AVX2-FP-NEXT:    vinsertf128 $1, %xmm4, %ymm0, %ymm41323; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm4 = ymm5[0,1,2,3,4,5],ymm4[6,7]1324; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm4[4,5,6,7]1325; AVX2-FP-NEXT:    movq {{[0-9]+}}(%rsp), %rax1326; AVX2-FP-NEXT:    vmovaps %ymm0, 64(%rax)1327; AVX2-FP-NEXT:    vmovaps %ymm9, (%rax)1328; AVX2-FP-NEXT:    vmovaps %ymm11, 96(%rax)1329; AVX2-FP-NEXT:    vmovaps %ymm8, 32(%rax)1330; AVX2-FP-NEXT:    vmovaps %ymm3, 128(%rax)1331; AVX2-FP-NEXT:    vmovaps %ymm2, 192(%rax)1332; AVX2-FP-NEXT:    vmovaps %ymm1, 224(%rax)1333; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload1334; AVX2-FP-NEXT:    vmovaps %ymm0, 160(%rax)1335; AVX2-FP-NEXT:    vzeroupper1336; AVX2-FP-NEXT:    retq1337;1338; AVX2-FCP-LABEL: store_i32_stride8_vf8:1339; AVX2-FCP:       # %bb.0:1340; AVX2-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax1341; AVX2-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r101342; AVX2-FCP-NEXT:    vmovaps (%rdi), %ymm31343; AVX2-FCP-NEXT:    vmovaps (%rsi), %ymm41344; AVX2-FCP-NEXT:    vmovaps (%rdx), %ymm11345; AVX2-FCP-NEXT:    vmovaps (%rcx), %ymm21346; AVX2-FCP-NEXT:    vmovaps (%r8), %ymm51347; AVX2-FCP-NEXT:    vmovaps (%r9), %ymm61348; AVX2-FCP-NEXT:    vmovaps (%r10), %ymm91349; AVX2-FCP-NEXT:    vmovaps (%rax), %ymm101350; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} ymm7 = ymm9[0],ymm10[0],ymm9[1],ymm10[1],ymm9[4],ymm10[4],ymm9[5],ymm10[5]1351; AVX2-FCP-NEXT:    vbroadcastss 20(%r8), %ymm01352; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm6[5],ymm0[6,7]1353; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm7[6,7]1354; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} ymm8 = ymm1[0],ymm2[0],ymm1[1],ymm2[1],ymm1[4],ymm2[4],ymm1[5],ymm2[5]1355; AVX2-FCP-NEXT:    vextractf128 $1, %ymm8, %xmm111356; AVX2-FCP-NEXT:    vshufps {{.*#+}} ymm12 = ymm3[1,1,1,1,5,5,5,5]1357; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm12 = ymm12[0],ymm4[1],ymm12[2,3,4],ymm4[5],ymm12[6,7]1358; AVX2-FCP-NEXT:    vextractf128 $1, %ymm12, %xmm121359; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm11 = xmm12[0,1],xmm11[2,3]1360; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm11[0,1,2,3],ymm0[4,5,6,7]1361; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill1362; AVX2-FCP-NEXT:    vbroadcastss 28(%r10), %ymm111363; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} ymm10 = ymm11[2],ymm10[2],ymm11[3],ymm10[3],ymm11[6],ymm10[6],ymm11[7],ymm10[7]1364; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} ymm11 = ymm5[2],ymm6[2],ymm5[3],ymm6[3],ymm5[6],ymm6[6],ymm5[7],ymm6[7]1365; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm10 = ymm11[1],ymm10[1],ymm11[3],ymm10[3]1366; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} ymm2 = ymm1[2],ymm2[2],ymm1[3],ymm2[3],ymm1[6],ymm2[6],ymm1[7],ymm2[7]1367; AVX2-FCP-NEXT:    vextractf128 $1, %ymm2, %xmm11368; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} ymm12 = ymm3[2],ymm4[2],ymm3[3],ymm4[3],ymm3[6],ymm4[6],ymm3[7],ymm4[7]1369; AVX2-FCP-NEXT:    vshufps {{.*#+}} ymm13 = ymm12[2,3,2,3,6,7,6,7]1370; AVX2-FCP-NEXT:    vextractf128 $1, %ymm13, %xmm131371; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm1 = xmm13[0,1],xmm1[2,3]1372; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm10[4,5,6,7]1373; AVX2-FCP-NEXT:    vbroadcastss 24(%rax), %ymm101374; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm9 = ymm9[0,1,2,3,4,5,6],ymm10[7]1375; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm9 = ymm11[0,1,2,3,4,5],ymm9[6,7]1376; AVX2-FCP-NEXT:    vextractf128 $1, %ymm12, %xmm101377; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm2 = ymm2[2,2,2,2]1378; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm2 = xmm10[0,1],xmm2[2,3]1379; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm9[4,5,6,7]1380; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} ymm5 = ymm5[0],ymm6[0],ymm5[1],ymm6[1],ymm5[4],ymm6[4],ymm5[5],ymm6[5]1381; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm5 = ymm5[0],ymm7[0],ymm5[2],ymm7[2]1382; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} ymm3 = ymm3[0],ymm4[0],ymm3[1],ymm4[1],ymm3[4],ymm4[4],ymm3[5],ymm4[5]1383; AVX2-FCP-NEXT:    vextractf128 $1, %ymm3, %xmm31384; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm4 = ymm8[2,2,2,2]1385; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm3 = xmm3[0,1],xmm4[2,3]1386; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm5[4,5,6,7]1387; AVX2-FCP-NEXT:    vmovaps (%rcx), %xmm41388; AVX2-FCP-NEXT:    vmovaps (%rdx), %xmm51389; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm6 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]1390; AVX2-FCP-NEXT:    vmovaps (%rsi), %xmm91391; AVX2-FCP-NEXT:    vmovaps (%rdi), %xmm101392; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm7 = xmm10[1,1,1,1]1393; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm7 = xmm7[0],xmm9[1],xmm7[2,3]1394; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm8 = xmm7[0,1],xmm6[2,3]1395; AVX2-FCP-NEXT:    vmovaps (%rax), %xmm61396; AVX2-FCP-NEXT:    vmovaps (%r10), %xmm71397; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm11 = xmm7[0],xmm6[0],xmm7[1],xmm6[1]1398; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm11, %ymm0, %ymm111399; AVX2-FCP-NEXT:    vmovaps (%r9), %xmm121400; AVX2-FCP-NEXT:    vmovaps (%r8), %xmm131401; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm14 = xmm13[1,1,1,1]1402; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm14 = xmm14[0],xmm12[1],xmm14[2,3]1403; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm14, %ymm0, %ymm141404; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm11 = ymm14[0,1,2,3,4,5],ymm11[6,7]1405; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm8 = ymm8[0,1,2,3],ymm11[4,5,6,7]1406; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} xmm11 = xmm7[2],xmm6[2],xmm7[3],xmm6[3]1407; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm11, %ymm0, %ymm111408; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} xmm14 = xmm13[2],xmm12[2],xmm13[3],xmm12[3]1409; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm15 = xmm14[2,3,2,3]1410; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm15, %ymm0, %ymm151411; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm11 = ymm15[0,1,2,3,4,5],ymm11[6,7]1412; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} xmm15 = xmm5[2],xmm4[2],xmm5[3],xmm4[3]1413; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm10[2],xmm9[2],xmm10[3],xmm9[3]1414; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} xmm15 = xmm0[1],xmm15[1]1415; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm11 = ymm15[0,1,2,3],ymm11[4,5,6,7]1416; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm9 = xmm10[0],xmm9[0],xmm10[1],xmm9[1]1417; AVX2-FCP-NEXT:    vbroadcastss %xmm4, %xmm101418; AVX2-FCP-NEXT:    vbroadcastss %xmm5, %xmm151419; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm10 = xmm15[0],xmm10[0],xmm15[1],xmm10[1]1420; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm9 = xmm9[0,1],xmm10[2,3]1421; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm10 = xmm13[0],xmm12[0],xmm13[1],xmm12[1]1422; AVX2-FCP-NEXT:    vbroadcastss %xmm6, %xmm121423; AVX2-FCP-NEXT:    vbroadcastss %xmm7, %xmm131424; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm12 = xmm13[0],xmm12[0],xmm13[1],xmm12[1]1425; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm10, %ymm0, %ymm101426; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm12, %ymm0, %ymm121427; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm10 = ymm10[0,1,2,3,4,5],ymm12[6,7]1428; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm9 = ymm9[0,1,2,3],ymm10[4,5,6,7]1429; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm4 = xmm4[2,2,2,2]1430; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm4 = xmm5[0,1,2],xmm4[3]1431; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm4[2,3]1432; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm4 = xmm6[2,2,2,2]1433; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm4 = xmm7[0,1,2],xmm4[3]1434; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm14, %ymm0, %ymm51435; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm4, %ymm0, %ymm41436; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm4 = ymm5[0,1,2,3,4,5],ymm4[6,7]1437; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm4[4,5,6,7]1438; AVX2-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax1439; AVX2-FCP-NEXT:    vmovaps %ymm0, 64(%rax)1440; AVX2-FCP-NEXT:    vmovaps %ymm9, (%rax)1441; AVX2-FCP-NEXT:    vmovaps %ymm11, 96(%rax)1442; AVX2-FCP-NEXT:    vmovaps %ymm8, 32(%rax)1443; AVX2-FCP-NEXT:    vmovaps %ymm3, 128(%rax)1444; AVX2-FCP-NEXT:    vmovaps %ymm2, 192(%rax)1445; AVX2-FCP-NEXT:    vmovaps %ymm1, 224(%rax)1446; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload1447; AVX2-FCP-NEXT:    vmovaps %ymm0, 160(%rax)1448; AVX2-FCP-NEXT:    vzeroupper1449; AVX2-FCP-NEXT:    retq1450;1451; AVX512-LABEL: store_i32_stride8_vf8:1452; AVX512:       # %bb.0:1453; AVX512-NEXT:    movq {{[0-9]+}}(%rsp), %rax1454; AVX512-NEXT:    movq {{[0-9]+}}(%rsp), %r101455; AVX512-NEXT:    movq {{[0-9]+}}(%rsp), %r111456; AVX512-NEXT:    vmovdqa (%rdi), %ymm01457; AVX512-NEXT:    vmovdqa (%rdx), %ymm11458; AVX512-NEXT:    vmovdqa (%r8), %ymm21459; AVX512-NEXT:    vmovdqa (%r11), %ymm31460; AVX512-NEXT:    vinserti64x4 $1, (%rsi), %zmm0, %zmm01461; AVX512-NEXT:    vinserti64x4 $1, (%rcx), %zmm1, %zmm11462; AVX512-NEXT:    vinserti64x4 $1, (%r9), %zmm2, %zmm21463; AVX512-NEXT:    vinserti64x4 $1, (%r10), %zmm3, %zmm31464; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [0,0,0,0,0,8,16,24,0,0,0,0,1,9,17,25]1465; AVX512-NEXT:    vpermi2d %zmm3, %zmm2, %zmm41466; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm5 = [0,8,16,24,0,0,0,0,1,9,17,25,0,0,0,0]1467; AVX512-NEXT:    vpermi2d %zmm1, %zmm0, %zmm51468; AVX512-NEXT:    movb $-52, %cl1469; AVX512-NEXT:    kmovw %ecx, %k11470; AVX512-NEXT:    vmovdqa64 %zmm4, %zmm5 {%k1}1471; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [0,0,0,0,2,10,18,26,0,0,0,0,3,11,19,27]1472; AVX512-NEXT:    vpermi2d %zmm3, %zmm2, %zmm41473; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm6 = [2,10,18,26,0,0,0,0,3,11,19,27,0,0,0,0]1474; AVX512-NEXT:    vpermi2d %zmm1, %zmm0, %zmm61475; AVX512-NEXT:    vmovdqa64 %zmm4, %zmm6 {%k1}1476; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [0,0,0,0,4,12,20,28,0,0,0,0,5,13,21,29]1477; AVX512-NEXT:    vpermi2d %zmm3, %zmm2, %zmm41478; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm7 = [4,12,20,28,0,0,0,0,5,13,21,29,0,0,0,0]1479; AVX512-NEXT:    vpermi2d %zmm1, %zmm0, %zmm71480; AVX512-NEXT:    vmovdqa64 %zmm4, %zmm7 {%k1}1481; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [0,0,0,0,6,14,22,30,0,0,0,0,7,15,23,31]1482; AVX512-NEXT:    vpermi2d %zmm3, %zmm2, %zmm41483; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [6,14,22,30,0,0,0,0,7,15,23,31,0,0,0,0]1484; AVX512-NEXT:    vpermi2d %zmm1, %zmm0, %zmm21485; AVX512-NEXT:    vmovdqa64 %zmm4, %zmm2 {%k1}1486; AVX512-NEXT:    vmovdqa64 %zmm2, 192(%rax)1487; AVX512-NEXT:    vmovdqa64 %zmm7, 128(%rax)1488; AVX512-NEXT:    vmovdqa64 %zmm6, 64(%rax)1489; AVX512-NEXT:    vmovdqa64 %zmm5, (%rax)1490; AVX512-NEXT:    vzeroupper1491; AVX512-NEXT:    retq1492;1493; AVX512-FCP-LABEL: store_i32_stride8_vf8:1494; AVX512-FCP:       # %bb.0:1495; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax1496; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r101497; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r111498; AVX512-FCP-NEXT:    vmovdqa (%rdi), %ymm01499; AVX512-FCP-NEXT:    vmovdqa (%rdx), %ymm11500; AVX512-FCP-NEXT:    vmovdqa (%r8), %ymm21501; AVX512-FCP-NEXT:    vmovdqa (%r11), %ymm31502; AVX512-FCP-NEXT:    vinserti64x4 $1, (%rsi), %zmm0, %zmm01503; AVX512-FCP-NEXT:    vinserti64x4 $1, (%rcx), %zmm1, %zmm11504; AVX512-FCP-NEXT:    vinserti64x4 $1, (%r9), %zmm2, %zmm21505; AVX512-FCP-NEXT:    vinserti64x4 $1, (%r10), %zmm3, %zmm31506; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [0,0,0,0,0,8,16,24,0,0,0,0,1,9,17,25]1507; AVX512-FCP-NEXT:    vpermi2d %zmm3, %zmm2, %zmm41508; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm5 = [0,8,16,24,0,0,0,0,1,9,17,25,0,0,0,0]1509; AVX512-FCP-NEXT:    vpermi2d %zmm1, %zmm0, %zmm51510; AVX512-FCP-NEXT:    movb $-52, %cl1511; AVX512-FCP-NEXT:    kmovw %ecx, %k11512; AVX512-FCP-NEXT:    vmovdqa64 %zmm4, %zmm5 {%k1}1513; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [0,0,0,0,2,10,18,26,0,0,0,0,3,11,19,27]1514; AVX512-FCP-NEXT:    vpermi2d %zmm3, %zmm2, %zmm41515; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm6 = [2,10,18,26,0,0,0,0,3,11,19,27,0,0,0,0]1516; AVX512-FCP-NEXT:    vpermi2d %zmm1, %zmm0, %zmm61517; AVX512-FCP-NEXT:    vmovdqa64 %zmm4, %zmm6 {%k1}1518; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [0,0,0,0,4,12,20,28,0,0,0,0,5,13,21,29]1519; AVX512-FCP-NEXT:    vpermi2d %zmm3, %zmm2, %zmm41520; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm7 = [4,12,20,28,0,0,0,0,5,13,21,29,0,0,0,0]1521; AVX512-FCP-NEXT:    vpermi2d %zmm1, %zmm0, %zmm71522; AVX512-FCP-NEXT:    vmovdqa64 %zmm4, %zmm7 {%k1}1523; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [0,0,0,0,6,14,22,30,0,0,0,0,7,15,23,31]1524; AVX512-FCP-NEXT:    vpermi2d %zmm3, %zmm2, %zmm41525; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [6,14,22,30,0,0,0,0,7,15,23,31,0,0,0,0]1526; AVX512-FCP-NEXT:    vpermi2d %zmm1, %zmm0, %zmm21527; AVX512-FCP-NEXT:    vmovdqa64 %zmm4, %zmm2 {%k1}1528; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, 192(%rax)1529; AVX512-FCP-NEXT:    vmovdqa64 %zmm7, 128(%rax)1530; AVX512-FCP-NEXT:    vmovdqa64 %zmm6, 64(%rax)1531; AVX512-FCP-NEXT:    vmovdqa64 %zmm5, (%rax)1532; AVX512-FCP-NEXT:    vzeroupper1533; AVX512-FCP-NEXT:    retq1534;1535; AVX512DQ-LABEL: store_i32_stride8_vf8:1536; AVX512DQ:       # %bb.0:1537; AVX512DQ-NEXT:    movq {{[0-9]+}}(%rsp), %rax1538; AVX512DQ-NEXT:    movq {{[0-9]+}}(%rsp), %r101539; AVX512DQ-NEXT:    movq {{[0-9]+}}(%rsp), %r111540; AVX512DQ-NEXT:    vmovdqa (%rdi), %ymm01541; AVX512DQ-NEXT:    vmovdqa (%rdx), %ymm11542; AVX512DQ-NEXT:    vmovdqa (%r8), %ymm21543; AVX512DQ-NEXT:    vmovdqa (%r11), %ymm31544; AVX512DQ-NEXT:    vinserti64x4 $1, (%rsi), %zmm0, %zmm01545; AVX512DQ-NEXT:    vinserti64x4 $1, (%rcx), %zmm1, %zmm11546; AVX512DQ-NEXT:    vinserti64x4 $1, (%r9), %zmm2, %zmm21547; AVX512DQ-NEXT:    vinserti64x4 $1, (%r10), %zmm3, %zmm31548; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [0,0,0,0,0,8,16,24,0,0,0,0,1,9,17,25]1549; AVX512DQ-NEXT:    vpermi2d %zmm3, %zmm2, %zmm41550; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm5 = [0,8,16,24,0,0,0,0,1,9,17,25,0,0,0,0]1551; AVX512DQ-NEXT:    vpermi2d %zmm1, %zmm0, %zmm51552; AVX512DQ-NEXT:    movb $-52, %cl1553; AVX512DQ-NEXT:    kmovw %ecx, %k11554; AVX512DQ-NEXT:    vmovdqa64 %zmm4, %zmm5 {%k1}1555; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [0,0,0,0,2,10,18,26,0,0,0,0,3,11,19,27]1556; AVX512DQ-NEXT:    vpermi2d %zmm3, %zmm2, %zmm41557; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm6 = [2,10,18,26,0,0,0,0,3,11,19,27,0,0,0,0]1558; AVX512DQ-NEXT:    vpermi2d %zmm1, %zmm0, %zmm61559; AVX512DQ-NEXT:    vmovdqa64 %zmm4, %zmm6 {%k1}1560; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [0,0,0,0,4,12,20,28,0,0,0,0,5,13,21,29]1561; AVX512DQ-NEXT:    vpermi2d %zmm3, %zmm2, %zmm41562; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm7 = [4,12,20,28,0,0,0,0,5,13,21,29,0,0,0,0]1563; AVX512DQ-NEXT:    vpermi2d %zmm1, %zmm0, %zmm71564; AVX512DQ-NEXT:    vmovdqa64 %zmm4, %zmm7 {%k1}1565; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [0,0,0,0,6,14,22,30,0,0,0,0,7,15,23,31]1566; AVX512DQ-NEXT:    vpermi2d %zmm3, %zmm2, %zmm41567; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [6,14,22,30,0,0,0,0,7,15,23,31,0,0,0,0]1568; AVX512DQ-NEXT:    vpermi2d %zmm1, %zmm0, %zmm21569; AVX512DQ-NEXT:    vmovdqa64 %zmm4, %zmm2 {%k1}1570; AVX512DQ-NEXT:    vmovdqa64 %zmm2, 192(%rax)1571; AVX512DQ-NEXT:    vmovdqa64 %zmm7, 128(%rax)1572; AVX512DQ-NEXT:    vmovdqa64 %zmm6, 64(%rax)1573; AVX512DQ-NEXT:    vmovdqa64 %zmm5, (%rax)1574; AVX512DQ-NEXT:    vzeroupper1575; AVX512DQ-NEXT:    retq1576;1577; AVX512DQ-FCP-LABEL: store_i32_stride8_vf8:1578; AVX512DQ-FCP:       # %bb.0:1579; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax1580; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r101581; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r111582; AVX512DQ-FCP-NEXT:    vmovdqa (%rdi), %ymm01583; AVX512DQ-FCP-NEXT:    vmovdqa (%rdx), %ymm11584; AVX512DQ-FCP-NEXT:    vmovdqa (%r8), %ymm21585; AVX512DQ-FCP-NEXT:    vmovdqa (%r11), %ymm31586; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, (%rsi), %zmm0, %zmm01587; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, (%rcx), %zmm1, %zmm11588; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, (%r9), %zmm2, %zmm21589; AVX512DQ-FCP-NEXT:    vinserti64x4 $1, (%r10), %zmm3, %zmm31590; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [0,0,0,0,0,8,16,24,0,0,0,0,1,9,17,25]1591; AVX512DQ-FCP-NEXT:    vpermi2d %zmm3, %zmm2, %zmm41592; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm5 = [0,8,16,24,0,0,0,0,1,9,17,25,0,0,0,0]1593; AVX512DQ-FCP-NEXT:    vpermi2d %zmm1, %zmm0, %zmm51594; AVX512DQ-FCP-NEXT:    movb $-52, %cl1595; AVX512DQ-FCP-NEXT:    kmovw %ecx, %k11596; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm4, %zmm5 {%k1}1597; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [0,0,0,0,2,10,18,26,0,0,0,0,3,11,19,27]1598; AVX512DQ-FCP-NEXT:    vpermi2d %zmm3, %zmm2, %zmm41599; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm6 = [2,10,18,26,0,0,0,0,3,11,19,27,0,0,0,0]1600; AVX512DQ-FCP-NEXT:    vpermi2d %zmm1, %zmm0, %zmm61601; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm4, %zmm6 {%k1}1602; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [0,0,0,0,4,12,20,28,0,0,0,0,5,13,21,29]1603; AVX512DQ-FCP-NEXT:    vpermi2d %zmm3, %zmm2, %zmm41604; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm7 = [4,12,20,28,0,0,0,0,5,13,21,29,0,0,0,0]1605; AVX512DQ-FCP-NEXT:    vpermi2d %zmm1, %zmm0, %zmm71606; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm4, %zmm7 {%k1}1607; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [0,0,0,0,6,14,22,30,0,0,0,0,7,15,23,31]1608; AVX512DQ-FCP-NEXT:    vpermi2d %zmm3, %zmm2, %zmm41609; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [6,14,22,30,0,0,0,0,7,15,23,31,0,0,0,0]1610; AVX512DQ-FCP-NEXT:    vpermi2d %zmm1, %zmm0, %zmm21611; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm4, %zmm2 {%k1}1612; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm2, 192(%rax)1613; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm7, 128(%rax)1614; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm6, 64(%rax)1615; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm5, (%rax)1616; AVX512DQ-FCP-NEXT:    vzeroupper1617; AVX512DQ-FCP-NEXT:    retq1618;1619; AVX512BW-LABEL: store_i32_stride8_vf8:1620; AVX512BW:       # %bb.0:1621; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax1622; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %r101623; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %r111624; AVX512BW-NEXT:    vmovdqa (%rdi), %ymm01625; AVX512BW-NEXT:    vmovdqa (%rdx), %ymm11626; AVX512BW-NEXT:    vmovdqa (%r8), %ymm21627; AVX512BW-NEXT:    vmovdqa (%r11), %ymm31628; AVX512BW-NEXT:    vinserti64x4 $1, (%rsi), %zmm0, %zmm01629; AVX512BW-NEXT:    vinserti64x4 $1, (%rcx), %zmm1, %zmm11630; AVX512BW-NEXT:    vinserti64x4 $1, (%r9), %zmm2, %zmm21631; AVX512BW-NEXT:    vinserti64x4 $1, (%r10), %zmm3, %zmm31632; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [0,0,0,0,0,8,16,24,0,0,0,0,1,9,17,25]1633; AVX512BW-NEXT:    vpermi2d %zmm3, %zmm2, %zmm41634; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm5 = [0,8,16,24,0,0,0,0,1,9,17,25,0,0,0,0]1635; AVX512BW-NEXT:    vpermi2d %zmm1, %zmm0, %zmm51636; AVX512BW-NEXT:    movb $-52, %cl1637; AVX512BW-NEXT:    kmovd %ecx, %k11638; AVX512BW-NEXT:    vmovdqa64 %zmm4, %zmm5 {%k1}1639; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [0,0,0,0,2,10,18,26,0,0,0,0,3,11,19,27]1640; AVX512BW-NEXT:    vpermi2d %zmm3, %zmm2, %zmm41641; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm6 = [2,10,18,26,0,0,0,0,3,11,19,27,0,0,0,0]1642; AVX512BW-NEXT:    vpermi2d %zmm1, %zmm0, %zmm61643; AVX512BW-NEXT:    vmovdqa64 %zmm4, %zmm6 {%k1}1644; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [0,0,0,0,4,12,20,28,0,0,0,0,5,13,21,29]1645; AVX512BW-NEXT:    vpermi2d %zmm3, %zmm2, %zmm41646; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm7 = [4,12,20,28,0,0,0,0,5,13,21,29,0,0,0,0]1647; AVX512BW-NEXT:    vpermi2d %zmm1, %zmm0, %zmm71648; AVX512BW-NEXT:    vmovdqa64 %zmm4, %zmm7 {%k1}1649; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [0,0,0,0,6,14,22,30,0,0,0,0,7,15,23,31]1650; AVX512BW-NEXT:    vpermi2d %zmm3, %zmm2, %zmm41651; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [6,14,22,30,0,0,0,0,7,15,23,31,0,0,0,0]1652; AVX512BW-NEXT:    vpermi2d %zmm1, %zmm0, %zmm21653; AVX512BW-NEXT:    vmovdqa64 %zmm4, %zmm2 {%k1}1654; AVX512BW-NEXT:    vmovdqa64 %zmm2, 192(%rax)1655; AVX512BW-NEXT:    vmovdqa64 %zmm7, 128(%rax)1656; AVX512BW-NEXT:    vmovdqa64 %zmm6, 64(%rax)1657; AVX512BW-NEXT:    vmovdqa64 %zmm5, (%rax)1658; AVX512BW-NEXT:    vzeroupper1659; AVX512BW-NEXT:    retq1660;1661; AVX512BW-FCP-LABEL: store_i32_stride8_vf8:1662; AVX512BW-FCP:       # %bb.0:1663; AVX512BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax1664; AVX512BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r101665; AVX512BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r111666; AVX512BW-FCP-NEXT:    vmovdqa (%rdi), %ymm01667; AVX512BW-FCP-NEXT:    vmovdqa (%rdx), %ymm11668; AVX512BW-FCP-NEXT:    vmovdqa (%r8), %ymm21669; AVX512BW-FCP-NEXT:    vmovdqa (%r11), %ymm31670; AVX512BW-FCP-NEXT:    vinserti64x4 $1, (%rsi), %zmm0, %zmm01671; AVX512BW-FCP-NEXT:    vinserti64x4 $1, (%rcx), %zmm1, %zmm11672; AVX512BW-FCP-NEXT:    vinserti64x4 $1, (%r9), %zmm2, %zmm21673; AVX512BW-FCP-NEXT:    vinserti64x4 $1, (%r10), %zmm3, %zmm31674; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [0,0,0,0,0,8,16,24,0,0,0,0,1,9,17,25]1675; AVX512BW-FCP-NEXT:    vpermi2d %zmm3, %zmm2, %zmm41676; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm5 = [0,8,16,24,0,0,0,0,1,9,17,25,0,0,0,0]1677; AVX512BW-FCP-NEXT:    vpermi2d %zmm1, %zmm0, %zmm51678; AVX512BW-FCP-NEXT:    movb $-52, %cl1679; AVX512BW-FCP-NEXT:    kmovd %ecx, %k11680; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm4, %zmm5 {%k1}1681; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [0,0,0,0,2,10,18,26,0,0,0,0,3,11,19,27]1682; AVX512BW-FCP-NEXT:    vpermi2d %zmm3, %zmm2, %zmm41683; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm6 = [2,10,18,26,0,0,0,0,3,11,19,27,0,0,0,0]1684; AVX512BW-FCP-NEXT:    vpermi2d %zmm1, %zmm0, %zmm61685; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm4, %zmm6 {%k1}1686; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [0,0,0,0,4,12,20,28,0,0,0,0,5,13,21,29]1687; AVX512BW-FCP-NEXT:    vpermi2d %zmm3, %zmm2, %zmm41688; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm7 = [4,12,20,28,0,0,0,0,5,13,21,29,0,0,0,0]1689; AVX512BW-FCP-NEXT:    vpermi2d %zmm1, %zmm0, %zmm71690; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm4, %zmm7 {%k1}1691; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [0,0,0,0,6,14,22,30,0,0,0,0,7,15,23,31]1692; AVX512BW-FCP-NEXT:    vpermi2d %zmm3, %zmm2, %zmm41693; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [6,14,22,30,0,0,0,0,7,15,23,31,0,0,0,0]1694; AVX512BW-FCP-NEXT:    vpermi2d %zmm1, %zmm0, %zmm21695; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm4, %zmm2 {%k1}1696; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm2, 192(%rax)1697; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm7, 128(%rax)1698; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm6, 64(%rax)1699; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm5, (%rax)1700; AVX512BW-FCP-NEXT:    vzeroupper1701; AVX512BW-FCP-NEXT:    retq1702;1703; AVX512DQ-BW-LABEL: store_i32_stride8_vf8:1704; AVX512DQ-BW:       # %bb.0:1705; AVX512DQ-BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax1706; AVX512DQ-BW-NEXT:    movq {{[0-9]+}}(%rsp), %r101707; AVX512DQ-BW-NEXT:    movq {{[0-9]+}}(%rsp), %r111708; AVX512DQ-BW-NEXT:    vmovdqa (%rdi), %ymm01709; AVX512DQ-BW-NEXT:    vmovdqa (%rdx), %ymm11710; AVX512DQ-BW-NEXT:    vmovdqa (%r8), %ymm21711; AVX512DQ-BW-NEXT:    vmovdqa (%r11), %ymm31712; AVX512DQ-BW-NEXT:    vinserti64x4 $1, (%rsi), %zmm0, %zmm01713; AVX512DQ-BW-NEXT:    vinserti64x4 $1, (%rcx), %zmm1, %zmm11714; AVX512DQ-BW-NEXT:    vinserti64x4 $1, (%r9), %zmm2, %zmm21715; AVX512DQ-BW-NEXT:    vinserti64x4 $1, (%r10), %zmm3, %zmm31716; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [0,0,0,0,0,8,16,24,0,0,0,0,1,9,17,25]1717; AVX512DQ-BW-NEXT:    vpermi2d %zmm3, %zmm2, %zmm41718; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm5 = [0,8,16,24,0,0,0,0,1,9,17,25,0,0,0,0]1719; AVX512DQ-BW-NEXT:    vpermi2d %zmm1, %zmm0, %zmm51720; AVX512DQ-BW-NEXT:    movb $-52, %cl1721; AVX512DQ-BW-NEXT:    kmovd %ecx, %k11722; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm4, %zmm5 {%k1}1723; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [0,0,0,0,2,10,18,26,0,0,0,0,3,11,19,27]1724; AVX512DQ-BW-NEXT:    vpermi2d %zmm3, %zmm2, %zmm41725; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm6 = [2,10,18,26,0,0,0,0,3,11,19,27,0,0,0,0]1726; AVX512DQ-BW-NEXT:    vpermi2d %zmm1, %zmm0, %zmm61727; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm4, %zmm6 {%k1}1728; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [0,0,0,0,4,12,20,28,0,0,0,0,5,13,21,29]1729; AVX512DQ-BW-NEXT:    vpermi2d %zmm3, %zmm2, %zmm41730; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm7 = [4,12,20,28,0,0,0,0,5,13,21,29,0,0,0,0]1731; AVX512DQ-BW-NEXT:    vpermi2d %zmm1, %zmm0, %zmm71732; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm4, %zmm7 {%k1}1733; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [0,0,0,0,6,14,22,30,0,0,0,0,7,15,23,31]1734; AVX512DQ-BW-NEXT:    vpermi2d %zmm3, %zmm2, %zmm41735; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [6,14,22,30,0,0,0,0,7,15,23,31,0,0,0,0]1736; AVX512DQ-BW-NEXT:    vpermi2d %zmm1, %zmm0, %zmm21737; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm4, %zmm2 {%k1}1738; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm2, 192(%rax)1739; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm7, 128(%rax)1740; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm6, 64(%rax)1741; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm5, (%rax)1742; AVX512DQ-BW-NEXT:    vzeroupper1743; AVX512DQ-BW-NEXT:    retq1744;1745; AVX512DQ-BW-FCP-LABEL: store_i32_stride8_vf8:1746; AVX512DQ-BW-FCP:       # %bb.0:1747; AVX512DQ-BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax1748; AVX512DQ-BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r101749; AVX512DQ-BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r111750; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%rdi), %ymm01751; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%rdx), %ymm11752; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%r8), %ymm21753; AVX512DQ-BW-FCP-NEXT:    vmovdqa (%r11), %ymm31754; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, (%rsi), %zmm0, %zmm01755; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, (%rcx), %zmm1, %zmm11756; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, (%r9), %zmm2, %zmm21757; AVX512DQ-BW-FCP-NEXT:    vinserti64x4 $1, (%r10), %zmm3, %zmm31758; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [0,0,0,0,0,8,16,24,0,0,0,0,1,9,17,25]1759; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm3, %zmm2, %zmm41760; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm5 = [0,8,16,24,0,0,0,0,1,9,17,25,0,0,0,0]1761; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm1, %zmm0, %zmm51762; AVX512DQ-BW-FCP-NEXT:    movb $-52, %cl1763; AVX512DQ-BW-FCP-NEXT:    kmovd %ecx, %k11764; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm4, %zmm5 {%k1}1765; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [0,0,0,0,2,10,18,26,0,0,0,0,3,11,19,27]1766; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm3, %zmm2, %zmm41767; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm6 = [2,10,18,26,0,0,0,0,3,11,19,27,0,0,0,0]1768; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm1, %zmm0, %zmm61769; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm4, %zmm6 {%k1}1770; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [0,0,0,0,4,12,20,28,0,0,0,0,5,13,21,29]1771; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm3, %zmm2, %zmm41772; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm7 = [4,12,20,28,0,0,0,0,5,13,21,29,0,0,0,0]1773; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm1, %zmm0, %zmm71774; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm4, %zmm7 {%k1}1775; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [0,0,0,0,6,14,22,30,0,0,0,0,7,15,23,31]1776; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm3, %zmm2, %zmm41777; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [6,14,22,30,0,0,0,0,7,15,23,31,0,0,0,0]1778; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm1, %zmm0, %zmm21779; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm4, %zmm2 {%k1}1780; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm2, 192(%rax)1781; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm7, 128(%rax)1782; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm6, 64(%rax)1783; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm5, (%rax)1784; AVX512DQ-BW-FCP-NEXT:    vzeroupper1785; AVX512DQ-BW-FCP-NEXT:    retq1786  %in.vec0 = load <8 x i32>, ptr %in.vecptr0, align 641787  %in.vec1 = load <8 x i32>, ptr %in.vecptr1, align 641788  %in.vec2 = load <8 x i32>, ptr %in.vecptr2, align 641789  %in.vec3 = load <8 x i32>, ptr %in.vecptr3, align 641790  %in.vec4 = load <8 x i32>, ptr %in.vecptr4, align 641791  %in.vec5 = load <8 x i32>, ptr %in.vecptr5, align 641792  %in.vec6 = load <8 x i32>, ptr %in.vecptr6, align 641793  %in.vec7 = load <8 x i32>, ptr %in.vecptr7, align 641794  %1 = shufflevector <8 x i32> %in.vec0, <8 x i32> %in.vec1, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>1795  %2 = shufflevector <8 x i32> %in.vec2, <8 x i32> %in.vec3, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>1796  %3 = shufflevector <8 x i32> %in.vec4, <8 x i32> %in.vec5, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>1797  %4 = shufflevector <8 x i32> %in.vec6, <8 x i32> %in.vec7, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>1798  %5 = shufflevector <16 x i32> %1, <16 x i32> %2, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>1799  %6 = shufflevector <16 x i32> %3, <16 x i32> %4, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>1800  %7 = shufflevector <32 x i32> %5, <32 x i32> %6, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>1801  %interleaved.vec = shufflevector <64 x i32> %7, <64 x i32> poison, <64 x i32> <i32 0, i32 8, i32 16, i32 24, i32 32, i32 40, i32 48, i32 56, i32 1, i32 9, i32 17, i32 25, i32 33, i32 41, i32 49, i32 57, i32 2, i32 10, i32 18, i32 26, i32 34, i32 42, i32 50, i32 58, i32 3, i32 11, i32 19, i32 27, i32 35, i32 43, i32 51, i32 59, i32 4, i32 12, i32 20, i32 28, i32 36, i32 44, i32 52, i32 60, i32 5, i32 13, i32 21, i32 29, i32 37, i32 45, i32 53, i32 61, i32 6, i32 14, i32 22, i32 30, i32 38, i32 46, i32 54, i32 62, i32 7, i32 15, i32 23, i32 31, i32 39, i32 47, i32 55, i32 63>1802  store <64 x i32> %interleaved.vec, ptr %out.vec, align 641803  ret void1804}1805 1806define void @store_i32_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 {1807; SSE-LABEL: store_i32_stride8_vf16:1808; SSE:       # %bb.0:1809; SSE-NEXT:    subq $216, %rsp1810; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %rax1811; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %r101812; SSE-NEXT:    movaps (%rdi), %xmm91813; SSE-NEXT:    movaps 16(%rdi), %xmm101814; SSE-NEXT:    movaps (%rsi), %xmm41815; SSE-NEXT:    movaps 16(%rsi), %xmm11816; SSE-NEXT:    movaps (%rdx), %xmm21817; SSE-NEXT:    movaps 16(%rdx), %xmm01818; SSE-NEXT:    movaps (%rcx), %xmm31819; SSE-NEXT:    movaps (%r8), %xmm111820; SSE-NEXT:    movaps (%r9), %xmm71821; SSE-NEXT:    movaps (%r10), %xmm51822; SSE-NEXT:    movaps (%rax), %xmm61823; SSE-NEXT:    movaps %xmm3, %xmm81824; SSE-NEXT:    movlhps {{.*#+}} xmm8 = xmm8[0],xmm2[0]1825; SSE-NEXT:    movaps %xmm9, %xmm131826; SSE-NEXT:    unpcklps {{.*#+}} xmm13 = xmm13[0],xmm4[0],xmm13[1],xmm4[1]1827; SSE-NEXT:    movaps %xmm13, %xmm121828; SSE-NEXT:    shufps {{.*#+}} xmm12 = xmm12[0,1],xmm8[2,0]1829; SSE-NEXT:    movaps %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1830; SSE-NEXT:    movaps %xmm6, %xmm81831; SSE-NEXT:    movlhps {{.*#+}} xmm8 = xmm8[0],xmm5[0]1832; SSE-NEXT:    movaps %xmm11, %xmm141833; SSE-NEXT:    unpcklps {{.*#+}} xmm14 = xmm14[0],xmm7[0],xmm14[1],xmm7[1]1834; SSE-NEXT:    movaps %xmm14, %xmm121835; SSE-NEXT:    shufps {{.*#+}} xmm12 = xmm12[0,1],xmm8[2,0]1836; SSE-NEXT:    movaps %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1837; SSE-NEXT:    movaps %xmm3, %xmm81838; SSE-NEXT:    shufps {{.*#+}} xmm8 = xmm8[1,1],xmm2[1,1]1839; SSE-NEXT:    shufps {{.*#+}} xmm13 = xmm13[2,3],xmm8[2,0]1840; SSE-NEXT:    movaps %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1841; SSE-NEXT:    movaps %xmm6, %xmm81842; SSE-NEXT:    shufps {{.*#+}} xmm8 = xmm8[1,1],xmm5[1,1]1843; SSE-NEXT:    shufps {{.*#+}} xmm14 = xmm14[2,3],xmm8[2,0]1844; SSE-NEXT:    movaps %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1845; SSE-NEXT:    movaps %xmm2, %xmm81846; SSE-NEXT:    unpckhpd {{.*#+}} xmm8 = xmm8[1],xmm3[1]1847; SSE-NEXT:    unpckhps {{.*#+}} xmm9 = xmm9[2],xmm4[2],xmm9[3],xmm4[3]1848; SSE-NEXT:    movaps %xmm9, %xmm41849; SSE-NEXT:    shufps {{.*#+}} xmm4 = xmm4[0,1],xmm8[0,2]1850; SSE-NEXT:    movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1851; SSE-NEXT:    movaps %xmm5, %xmm41852; SSE-NEXT:    unpckhpd {{.*#+}} xmm4 = xmm4[1],xmm6[1]1853; SSE-NEXT:    unpckhps {{.*#+}} xmm11 = xmm11[2],xmm7[2],xmm11[3],xmm7[3]1854; SSE-NEXT:    movaps %xmm11, %xmm71855; SSE-NEXT:    shufps {{.*#+}} xmm7 = xmm7[0,1],xmm4[0,2]1856; SSE-NEXT:    movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1857; SSE-NEXT:    movaps 16(%rcx), %xmm41858; SSE-NEXT:    shufps {{.*#+}} xmm2 = xmm2[3,3],xmm3[3,3]1859; SSE-NEXT:    movaps 16(%r10), %xmm31860; SSE-NEXT:    shufps {{.*#+}} xmm9 = xmm9[2,3],xmm2[0,2]1861; SSE-NEXT:    movaps %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1862; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[3,3],xmm6[3,3]1863; SSE-NEXT:    shufps {{.*#+}} xmm11 = xmm11[2,3],xmm5[0,2]1864; SSE-NEXT:    movaps %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1865; SSE-NEXT:    movaps %xmm4, %xmm21866; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]1867; SSE-NEXT:    movaps %xmm10, %xmm61868; SSE-NEXT:    unpcklps {{.*#+}} xmm6 = xmm6[0],xmm1[0],xmm6[1],xmm1[1]1869; SSE-NEXT:    movaps %xmm6, %xmm51870; SSE-NEXT:    movaps %xmm6, %xmm71871; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[0,1],xmm2[2,0]1872; SSE-NEXT:    movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1873; SSE-NEXT:    movaps 16(%rax), %xmm21874; SSE-NEXT:    movaps %xmm2, %xmm51875; SSE-NEXT:    movlhps {{.*#+}} xmm5 = xmm5[0],xmm3[0]1876; SSE-NEXT:    movaps 16(%r8), %xmm111877; SSE-NEXT:    movaps 16(%r9), %xmm61878; SSE-NEXT:    movaps %xmm11, %xmm91879; SSE-NEXT:    unpcklps {{.*#+}} xmm9 = xmm9[0],xmm6[0],xmm9[1],xmm6[1]1880; SSE-NEXT:    movaps %xmm9, %xmm81881; SSE-NEXT:    shufps {{.*#+}} xmm8 = xmm8[0,1],xmm5[2,0]1882; SSE-NEXT:    movaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1883; SSE-NEXT:    movaps %xmm4, %xmm51884; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[1,1],xmm0[1,1]1885; SSE-NEXT:    shufps {{.*#+}} xmm7 = xmm7[2,3],xmm5[2,0]1886; SSE-NEXT:    movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1887; SSE-NEXT:    movaps %xmm2, %xmm51888; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[1,1],xmm3[1,1]1889; SSE-NEXT:    shufps {{.*#+}} xmm9 = xmm9[2,3],xmm5[2,0]1890; SSE-NEXT:    movaps %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1891; SSE-NEXT:    unpckhps {{.*#+}} xmm10 = xmm10[2],xmm1[2],xmm10[3],xmm1[3]1892; SSE-NEXT:    movaps %xmm0, %xmm11893; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm4[1]1894; SSE-NEXT:    movaps %xmm10, %xmm51895; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[0,1],xmm1[0,2]1896; SSE-NEXT:    movaps %xmm5, (%rsp) # 16-byte Spill1897; SSE-NEXT:    unpckhps {{.*#+}} xmm11 = xmm11[2],xmm6[2],xmm11[3],xmm6[3]1898; SSE-NEXT:    movaps %xmm3, %xmm11899; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm2[1]1900; SSE-NEXT:    movaps %xmm11, %xmm51901; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[0,1],xmm1[0,2]1902; SSE-NEXT:    movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1903; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[3,3],xmm4[3,3]1904; SSE-NEXT:    shufps {{.*#+}} xmm10 = xmm10[2,3],xmm0[0,2]1905; SSE-NEXT:    movaps %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1906; SSE-NEXT:    shufps {{.*#+}} xmm3 = xmm3[3,3],xmm2[3,3]1907; SSE-NEXT:    shufps {{.*#+}} xmm11 = xmm11[2,3],xmm3[0,2]1908; SSE-NEXT:    movaps %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1909; SSE-NEXT:    movaps 32(%rdx), %xmm31910; SSE-NEXT:    movaps 32(%rcx), %xmm01911; SSE-NEXT:    movaps %xmm0, %xmm11912; SSE-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm3[0]1913; SSE-NEXT:    movaps 32(%rdi), %xmm141914; SSE-NEXT:    movaps 32(%rsi), %xmm41915; SSE-NEXT:    movaps %xmm14, %xmm51916; SSE-NEXT:    unpcklps {{.*#+}} xmm5 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]1917; SSE-NEXT:    movaps %xmm5, %xmm21918; SSE-NEXT:    movaps %xmm5, %xmm71919; SSE-NEXT:    shufps {{.*#+}} xmm2 = xmm2[0,1],xmm1[2,0]1920; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1921; SSE-NEXT:    movaps 32(%r10), %xmm11922; SSE-NEXT:    movaps 32(%rax), %xmm21923; SSE-NEXT:    movaps %xmm2, %xmm51924; SSE-NEXT:    movlhps {{.*#+}} xmm5 = xmm5[0],xmm1[0]1925; SSE-NEXT:    movaps 32(%r8), %xmm111926; SSE-NEXT:    movaps 32(%r9), %xmm61927; SSE-NEXT:    movaps %xmm11, %xmm131928; SSE-NEXT:    unpcklps {{.*#+}} xmm13 = xmm13[0],xmm6[0],xmm13[1],xmm6[1]1929; SSE-NEXT:    movaps %xmm13, %xmm81930; SSE-NEXT:    shufps {{.*#+}} xmm8 = xmm8[0,1],xmm5[2,0]1931; SSE-NEXT:    movaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1932; SSE-NEXT:    movaps %xmm0, %xmm51933; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[1,1],xmm3[1,1]1934; SSE-NEXT:    shufps {{.*#+}} xmm7 = xmm7[2,3],xmm5[2,0]1935; SSE-NEXT:    movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1936; SSE-NEXT:    movaps %xmm2, %xmm51937; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[1,1],xmm1[1,1]1938; SSE-NEXT:    shufps {{.*#+}} xmm13 = xmm13[2,3],xmm5[2,0]1939; SSE-NEXT:    unpckhps {{.*#+}} xmm14 = xmm14[2],xmm4[2],xmm14[3],xmm4[3]1940; SSE-NEXT:    movaps %xmm3, %xmm41941; SSE-NEXT:    unpckhpd {{.*#+}} xmm4 = xmm4[1],xmm0[1]1942; SSE-NEXT:    movaps %xmm14, %xmm51943; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[0,1],xmm4[0,2]1944; SSE-NEXT:    movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1945; SSE-NEXT:    unpckhps {{.*#+}} xmm11 = xmm11[2],xmm6[2],xmm11[3],xmm6[3]1946; SSE-NEXT:    movaps %xmm1, %xmm41947; SSE-NEXT:    unpckhpd {{.*#+}} xmm4 = xmm4[1],xmm2[1]1948; SSE-NEXT:    movaps %xmm11, %xmm51949; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[0,1],xmm4[0,2]1950; SSE-NEXT:    movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1951; SSE-NEXT:    shufps {{.*#+}} xmm3 = xmm3[3,3],xmm0[3,3]1952; SSE-NEXT:    shufps {{.*#+}} xmm14 = xmm14[2,3],xmm3[0,2]1953; SSE-NEXT:    shufps {{.*#+}} xmm1 = xmm1[3,3],xmm2[3,3]1954; SSE-NEXT:    shufps {{.*#+}} xmm11 = xmm11[2,3],xmm1[0,2]1955; SSE-NEXT:    movaps 48(%rdx), %xmm11956; SSE-NEXT:    movaps 48(%rcx), %xmm81957; SSE-NEXT:    movaps %xmm8, %xmm01958; SSE-NEXT:    movlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0]1959; SSE-NEXT:    movaps 48(%rdi), %xmm31960; SSE-NEXT:    movaps 48(%rsi), %xmm121961; SSE-NEXT:    movaps %xmm3, %xmm41962; SSE-NEXT:    unpcklps {{.*#+}} xmm4 = xmm4[0],xmm12[0],xmm4[1],xmm12[1]1963; SSE-NEXT:    movaps %xmm4, %xmm151964; SSE-NEXT:    shufps {{.*#+}} xmm15 = xmm15[0,1],xmm0[2,0]1965; SSE-NEXT:    movaps 48(%r10), %xmm01966; SSE-NEXT:    movaps 48(%rax), %xmm71967; SSE-NEXT:    movaps %xmm7, %xmm21968; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]1969; SSE-NEXT:    movaps 48(%r8), %xmm51970; SSE-NEXT:    movaps 48(%r9), %xmm91971; SSE-NEXT:    movaps %xmm5, %xmm61972; SSE-NEXT:    unpcklps {{.*#+}} xmm6 = xmm6[0],xmm9[0],xmm6[1],xmm9[1]1973; SSE-NEXT:    movaps %xmm6, %xmm101974; SSE-NEXT:    shufps {{.*#+}} xmm10 = xmm10[0,1],xmm2[2,0]1975; SSE-NEXT:    movaps %xmm8, %xmm21976; SSE-NEXT:    shufps {{.*#+}} xmm2 = xmm2[1,1],xmm1[1,1]1977; SSE-NEXT:    shufps {{.*#+}} xmm4 = xmm4[2,3],xmm2[2,0]1978; SSE-NEXT:    movaps %xmm7, %xmm21979; SSE-NEXT:    shufps {{.*#+}} xmm2 = xmm2[1,1],xmm0[1,1]1980; SSE-NEXT:    shufps {{.*#+}} xmm6 = xmm6[2,3],xmm2[2,0]1981; SSE-NEXT:    unpckhps {{.*#+}} xmm3 = xmm3[2],xmm12[2],xmm3[3],xmm12[3]1982; SSE-NEXT:    movaps %xmm1, %xmm21983; SSE-NEXT:    unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm8[1]1984; SSE-NEXT:    movaps %xmm3, %xmm121985; SSE-NEXT:    shufps {{.*#+}} xmm12 = xmm12[0,1],xmm2[0,2]1986; SSE-NEXT:    unpckhps {{.*#+}} xmm5 = xmm5[2],xmm9[2],xmm5[3],xmm9[3]1987; SSE-NEXT:    movaps %xmm0, %xmm21988; SSE-NEXT:    unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm7[1]1989; SSE-NEXT:    movaps %xmm5, %xmm91990; SSE-NEXT:    shufps {{.*#+}} xmm9 = xmm9[0,1],xmm2[0,2]1991; SSE-NEXT:    shufps {{.*#+}} xmm1 = xmm1[3,3],xmm8[3,3]1992; SSE-NEXT:    shufps {{.*#+}} xmm3 = xmm3[2,3],xmm1[0,2]1993; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[3,3],xmm7[3,3]1994; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[2,3],xmm0[0,2]1995; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %rax1996; SSE-NEXT:    movaps %xmm5, 496(%rax)1997; SSE-NEXT:    movaps %xmm3, 480(%rax)1998; SSE-NEXT:    movaps %xmm9, 464(%rax)1999; SSE-NEXT:    movaps %xmm12, 448(%rax)2000; SSE-NEXT:    movaps %xmm6, 432(%rax)2001; SSE-NEXT:    movaps %xmm4, 416(%rax)2002; SSE-NEXT:    movaps %xmm10, 400(%rax)2003; SSE-NEXT:    movaps %xmm15, 384(%rax)2004; SSE-NEXT:    movaps %xmm11, 368(%rax)2005; SSE-NEXT:    movaps %xmm14, 352(%rax)2006; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2007; SSE-NEXT:    movaps %xmm0, 336(%rax)2008; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2009; SSE-NEXT:    movaps %xmm0, 320(%rax)2010; SSE-NEXT:    movaps %xmm13, 304(%rax)2011; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2012; SSE-NEXT:    movaps %xmm0, 288(%rax)2013; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2014; SSE-NEXT:    movaps %xmm0, 272(%rax)2015; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2016; SSE-NEXT:    movaps %xmm0, 256(%rax)2017; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2018; SSE-NEXT:    movaps %xmm0, 240(%rax)2019; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2020; SSE-NEXT:    movaps %xmm0, 224(%rax)2021; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2022; SSE-NEXT:    movaps %xmm0, 208(%rax)2023; SSE-NEXT:    movaps (%rsp), %xmm0 # 16-byte Reload2024; SSE-NEXT:    movaps %xmm0, 192(%rax)2025; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2026; SSE-NEXT:    movaps %xmm0, 176(%rax)2027; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2028; SSE-NEXT:    movaps %xmm0, 160(%rax)2029; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2030; SSE-NEXT:    movaps %xmm0, 144(%rax)2031; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2032; SSE-NEXT:    movaps %xmm0, 128(%rax)2033; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2034; SSE-NEXT:    movaps %xmm0, 112(%rax)2035; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2036; SSE-NEXT:    movaps %xmm0, 96(%rax)2037; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2038; SSE-NEXT:    movaps %xmm0, 80(%rax)2039; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2040; SSE-NEXT:    movaps %xmm0, 64(%rax)2041; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2042; SSE-NEXT:    movaps %xmm0, 48(%rax)2043; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2044; SSE-NEXT:    movaps %xmm0, 32(%rax)2045; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2046; SSE-NEXT:    movaps %xmm0, 16(%rax)2047; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2048; SSE-NEXT:    movaps %xmm0, (%rax)2049; SSE-NEXT:    addq $216, %rsp2050; SSE-NEXT:    retq2051;2052; AVX-LABEL: store_i32_stride8_vf16:2053; AVX:       # %bb.0:2054; AVX-NEXT:    subq $264, %rsp # imm = 0x1082055; AVX-NEXT:    movq {{[0-9]+}}(%rsp), %rax2056; AVX-NEXT:    movq {{[0-9]+}}(%rsp), %r102057; AVX-NEXT:    vmovaps 32(%r8), %ymm02058; AVX-NEXT:    vmovaps 32(%r9), %ymm52059; AVX-NEXT:    vmovaps 32(%r10), %ymm132060; AVX-NEXT:    vmovaps 32(%rax), %ymm142061; AVX-NEXT:    vmovaps (%rcx), %xmm22062; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2063; AVX-NEXT:    vmovaps (%rdx), %xmm12064; AVX-NEXT:    vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2065; AVX-NEXT:    vunpcklps {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]2066; AVX-NEXT:    vmovaps (%rsi), %xmm62067; AVX-NEXT:    vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2068; AVX-NEXT:    vmovaps (%rdi), %xmm72069; AVX-NEXT:    vmovaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2070; AVX-NEXT:    vmovaps (%rax), %xmm32071; AVX-NEXT:    vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2072; AVX-NEXT:    vmovaps (%r10), %xmm22073; AVX-NEXT:    vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2074; AVX-NEXT:    vunpcklps {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1]2075; AVX-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm32076; AVX-NEXT:    vmovaps (%r9), %xmm152077; AVX-NEXT:    vmovaps (%r8), %xmm112078; AVX-NEXT:    vshufps {{.*#+}} xmm4 = xmm11[1,1,1,1]2079; AVX-NEXT:    vblendps {{.*#+}} xmm4 = xmm4[0],xmm15[1],xmm4[2,3]2080; AVX-NEXT:    vinsertf128 $1, %xmm4, %ymm0, %ymm42081; AVX-NEXT:    vblendps {{.*#+}} ymm3 = ymm4[0,1,2,3,4,5],ymm3[6,7]2082; AVX-NEXT:    vshufps {{.*#+}} xmm4 = xmm7[1,1,1,1]2083; AVX-NEXT:    vblendps {{.*#+}} xmm4 = xmm4[0],xmm6[1],xmm4[2,3]2084; AVX-NEXT:    vblendps {{.*#+}} xmm4 = xmm4[0,1],xmm1[2,3]2085; AVX-NEXT:    vblendps {{.*#+}} ymm3 = ymm4[0,1,2,3],ymm3[4,5,6,7]2086; AVX-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2087; AVX-NEXT:    vunpcklps {{.*#+}} xmm3 = xmm7[0],xmm6[0],xmm7[1],xmm6[1]2088; AVX-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm3[0],xmm1[0]2089; AVX-NEXT:    vunpcklps {{.*#+}} xmm3 = xmm11[0],xmm15[0],xmm11[1],xmm15[1]2090; AVX-NEXT:    vinsertf128 $1, %xmm3, %ymm0, %ymm32091; AVX-NEXT:    vshufps {{.*#+}} xmm2 = xmm2[0,1,0,1]2092; AVX-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm22093; AVX-NEXT:    vblendps {{.*#+}} ymm2 = ymm3[0,1,2,3,4,5],ymm2[6,7]2094; AVX-NEXT:    vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]2095; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2096; AVX-NEXT:    vunpckhpd {{.*#+}} ymm1 = ymm14[1],ymm13[1],ymm14[3],ymm13[3]2097; AVX-NEXT:    vunpckhps {{.*#+}} ymm2 = ymm0[2],ymm5[2],ymm0[3],ymm5[3],ymm0[6],ymm5[6],ymm0[7],ymm5[7]2098; AVX-NEXT:    vshufps {{.*#+}} ymm3 = ymm2[0,1],ymm1[2,0],ymm2[4,5],ymm1[6,4]2099; AVX-NEXT:    vmovaps 32(%rdx), %ymm12100; AVX-NEXT:    vmovaps 32(%rcx), %ymm22101; AVX-NEXT:    vunpckhpd {{.*#+}} ymm4 = ymm2[1],ymm1[1],ymm2[3],ymm1[3]2102; AVX-NEXT:    vextractf128 $1, %ymm4, %xmm42103; AVX-NEXT:    vmovaps 32(%rdi), %ymm62104; AVX-NEXT:    vmovaps 32(%rsi), %ymm72105; AVX-NEXT:    vunpckhps {{.*#+}} ymm8 = ymm6[2],ymm7[2],ymm6[3],ymm7[3],ymm6[6],ymm7[6],ymm6[7],ymm7[7]2106; AVX-NEXT:    vextractf128 $1, %ymm8, %xmm82107; AVX-NEXT:    vshufps {{.*#+}} xmm4 = xmm8[0,1],xmm4[2,0]2108; AVX-NEXT:    vblendps {{.*#+}} ymm3 = ymm4[0,1,2,3],ymm3[4,5,6,7]2109; AVX-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2110; AVX-NEXT:    vunpckhps {{.*#+}} ymm3 = ymm13[2],ymm14[2],ymm13[3],ymm14[3],ymm13[6],ymm14[6],ymm13[7],ymm14[7]2111; AVX-NEXT:    vshufps {{.*#+}} ymm4 = ymm5[3,0],ymm0[3,0],ymm5[7,4],ymm0[7,4]2112; AVX-NEXT:    vshufps {{.*#+}} ymm3 = ymm4[2,0],ymm3[2,3],ymm4[6,4],ymm3[6,7]2113; AVX-NEXT:    vunpckhps {{.*#+}} ymm4 = ymm1[2],ymm2[2],ymm1[3],ymm2[3],ymm1[6],ymm2[6],ymm1[7],ymm2[7]2114; AVX-NEXT:    vextractf128 $1, %ymm4, %xmm42115; AVX-NEXT:    vshufps {{.*#+}} ymm8 = ymm7[3,0],ymm6[3,0],ymm7[7,4],ymm6[7,4]2116; AVX-NEXT:    vextractf128 $1, %ymm8, %xmm82117; AVX-NEXT:    vshufps {{.*#+}} xmm4 = xmm8[2,0],xmm4[2,3]2118; AVX-NEXT:    vblendps {{.*#+}} ymm3 = ymm4[0,1,2,3],ymm3[4,5,6,7]2119; AVX-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2120; AVX-NEXT:    vunpcklpd {{.*#+}} ymm3 = ymm14[0],ymm13[0],ymm14[2],ymm13[2]2121; AVX-NEXT:    vunpcklps {{.*#+}} ymm4 = ymm0[0],ymm5[0],ymm0[1],ymm5[1],ymm0[4],ymm5[4],ymm0[5],ymm5[5]2122; AVX-NEXT:    vshufps {{.*#+}} ymm3 = ymm4[0,1],ymm3[2,0],ymm4[4,5],ymm3[6,4]2123; AVX-NEXT:    vunpcklpd {{.*#+}} ymm4 = ymm2[0],ymm1[0],ymm2[2],ymm1[2]2124; AVX-NEXT:    vextractf128 $1, %ymm4, %xmm42125; AVX-NEXT:    vunpcklps {{.*#+}} ymm8 = ymm6[0],ymm7[0],ymm6[1],ymm7[1],ymm6[4],ymm7[4],ymm6[5],ymm7[5]2126; AVX-NEXT:    vextractf128 $1, %ymm8, %xmm82127; AVX-NEXT:    vshufps {{.*#+}} xmm4 = xmm8[0,1],xmm4[2,0]2128; AVX-NEXT:    vblendps {{.*#+}} ymm3 = ymm4[0,1,2,3],ymm3[4,5,6,7]2129; AVX-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2130; AVX-NEXT:    vunpcklps {{.*#+}} ymm3 = ymm13[0],ymm14[0],ymm13[1],ymm14[1],ymm13[4],ymm14[4],ymm13[5],ymm14[5]2131; AVX-NEXT:    vshufps {{.*#+}} ymm0 = ymm5[1,0],ymm0[1,0],ymm5[5,4],ymm0[5,4]2132; AVX-NEXT:    vshufps {{.*#+}} ymm0 = ymm0[2,0],ymm3[2,3],ymm0[6,4],ymm3[6,7]2133; AVX-NEXT:    vunpcklps {{.*#+}} ymm1 = ymm1[0],ymm2[0],ymm1[1],ymm2[1],ymm1[4],ymm2[4],ymm1[5],ymm2[5]2134; AVX-NEXT:    vshufps {{.*#+}} ymm2 = ymm7[1,0],ymm6[1,0],ymm7[5,4],ymm6[5,4]2135; AVX-NEXT:    vextractf128 $1, %ymm1, %xmm12136; AVX-NEXT:    vextractf128 $1, %ymm2, %xmm22137; AVX-NEXT:    vshufps {{.*#+}} xmm1 = xmm2[2,0],xmm1[2,3]2138; AVX-NEXT:    vmovaps 32(%rax), %xmm132139; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]2140; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2141; AVX-NEXT:    vmovaps 32(%r10), %xmm12142; AVX-NEXT:    vshufps {{.*#+}} xmm0 = xmm13[2,2,2,2]2143; AVX-NEXT:    vblendps {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[3]2144; AVX-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm32145; AVX-NEXT:    vmovaps 32(%r9), %xmm22146; AVX-NEXT:    vmovaps 32(%r8), %xmm02147; AVX-NEXT:    vunpckhps {{.*#+}} xmm4 = xmm0[2],xmm2[2],xmm0[3],xmm2[3]2148; AVX-NEXT:    vinsertf128 $1, %xmm4, %ymm0, %ymm62149; AVX-NEXT:    vblendps {{.*#+}} ymm3 = ymm6[0,1,2,3,4,5],ymm3[6,7]2150; AVX-NEXT:    vmovaps 32(%rcx), %xmm62151; AVX-NEXT:    vmovaps 32(%rdx), %xmm72152; AVX-NEXT:    vshufps {{.*#+}} xmm8 = xmm6[2,2,2,2]2153; AVX-NEXT:    vblendps {{.*#+}} xmm8 = xmm7[0,1,2],xmm8[3]2154; AVX-NEXT:    vmovaps 32(%rsi), %xmm122155; AVX-NEXT:    vmovaps 32(%rdi), %xmm52156; AVX-NEXT:    vunpckhps {{.*#+}} xmm14 = xmm5[2],xmm12[2],xmm5[3],xmm12[3]2157; AVX-NEXT:    vblendps {{.*#+}} xmm8 = xmm14[0,1],xmm8[2,3]2158; AVX-NEXT:    vblendps {{.*#+}} ymm3 = ymm8[0,1,2,3],ymm3[4,5,6,7]2159; AVX-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2160; AVX-NEXT:    vunpckhps {{.*#+}} xmm3 = xmm7[2],xmm6[2],xmm7[3],xmm6[3]2161; AVX-NEXT:    vunpckhpd {{.*#+}} xmm3 = xmm14[1],xmm3[1]2162; AVX-NEXT:    vunpckhps {{.*#+}} xmm8 = xmm1[2],xmm13[2],xmm1[3],xmm13[3]2163; AVX-NEXT:    vinsertf128 $1, %xmm8, %ymm0, %ymm82164; AVX-NEXT:    vshufps {{.*#+}} xmm4 = xmm4[2,3,2,3]2165; AVX-NEXT:    vinsertf128 $1, %xmm4, %ymm0, %ymm42166; AVX-NEXT:    vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm8[6,7]2167; AVX-NEXT:    vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm4[4,5,6,7]2168; AVX-NEXT:    vmovups %ymm3, (%rsp) # 32-byte Spill2169; AVX-NEXT:    vunpcklps {{.*#+}} xmm3 = xmm7[0],xmm6[0],xmm7[1],xmm6[1]2170; AVX-NEXT:    vunpcklps {{.*#+}} xmm1 = xmm1[0],xmm13[0],xmm1[1],xmm13[1]2171; AVX-NEXT:    vunpcklps {{.*#+}} xmm4 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]2172; AVX-NEXT:    vinsertf128 $1, %xmm4, %ymm0, %ymm42173; AVX-NEXT:    vshufps {{.*#+}} xmm6 = xmm1[0,1,0,1]2174; AVX-NEXT:    vinsertf128 $1, %xmm6, %ymm0, %ymm62175; AVX-NEXT:    vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3,4,5],ymm6[6,7]2176; AVX-NEXT:    vunpcklps {{.*#+}} xmm6 = xmm5[0],xmm12[0],xmm5[1],xmm12[1]2177; AVX-NEXT:    vmovlhps {{.*#+}} xmm6 = xmm6[0],xmm3[0]2178; AVX-NEXT:    vblendps {{.*#+}} ymm4 = ymm6[0,1,2,3],ymm4[4,5,6,7]2179; AVX-NEXT:    vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2180; AVX-NEXT:    vshufps {{.*#+}} xmm4 = xmm5[1,1,1,1]2181; AVX-NEXT:    vblendps {{.*#+}} xmm4 = xmm4[0],xmm12[1],xmm4[2,3]2182; AVX-NEXT:    vblendps {{.*#+}} xmm3 = xmm4[0,1],xmm3[2,3]2183; AVX-NEXT:    vshufps {{.*#+}} xmm0 = xmm0[1,1,1,1]2184; AVX-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0],xmm2[1],xmm0[2,3]2185; AVX-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm12186; AVX-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm02187; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]2188; AVX-NEXT:    vblendps {{.*#+}} ymm12 = ymm3[0,1,2,3],ymm0[4,5,6,7]2189; AVX-NEXT:    vmovaps (%r8), %ymm82190; AVX-NEXT:    vmovaps (%r9), %ymm62191; AVX-NEXT:    vmovaps (%r10), %ymm52192; AVX-NEXT:    vmovaps (%rax), %ymm42193; AVX-NEXT:    vunpckhpd {{.*#+}} ymm0 = ymm4[1],ymm5[1],ymm4[3],ymm5[3]2194; AVX-NEXT:    vunpckhps {{.*#+}} ymm1 = ymm8[2],ymm6[2],ymm8[3],ymm6[3],ymm8[6],ymm6[6],ymm8[7],ymm6[7]2195; AVX-NEXT:    vshufps {{.*#+}} ymm9 = ymm1[0,1],ymm0[2,0],ymm1[4,5],ymm0[6,4]2196; AVX-NEXT:    vmovaps (%rdx), %ymm32197; AVX-NEXT:    vmovaps (%rcx), %ymm22198; AVX-NEXT:    vunpckhpd {{.*#+}} ymm7 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]2199; AVX-NEXT:    vextractf128 $1, %ymm7, %xmm72200; AVX-NEXT:    vmovaps (%rdi), %ymm12201; AVX-NEXT:    vmovaps (%rsi), %ymm02202; AVX-NEXT:    vunpckhps {{.*#+}} ymm10 = ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[6],ymm0[6],ymm1[7],ymm0[7]2203; AVX-NEXT:    vextractf128 $1, %ymm10, %xmm102204; AVX-NEXT:    vshufps {{.*#+}} xmm7 = xmm10[0,1],xmm7[2,0]2205; AVX-NEXT:    vblendps {{.*#+}} ymm10 = ymm7[0,1,2,3],ymm9[4,5,6,7]2206; AVX-NEXT:    vunpckhps {{.*#+}} ymm7 = ymm5[2],ymm4[2],ymm5[3],ymm4[3],ymm5[6],ymm4[6],ymm5[7],ymm4[7]2207; AVX-NEXT:    vshufps {{.*#+}} ymm9 = ymm6[3,0],ymm8[3,0],ymm6[7,4],ymm8[7,4]2208; AVX-NEXT:    vshufps {{.*#+}} ymm7 = ymm9[2,0],ymm7[2,3],ymm9[6,4],ymm7[6,7]2209; AVX-NEXT:    vunpckhps {{.*#+}} ymm9 = ymm3[2],ymm2[2],ymm3[3],ymm2[3],ymm3[6],ymm2[6],ymm3[7],ymm2[7]2210; AVX-NEXT:    vextractf128 $1, %ymm9, %xmm92211; AVX-NEXT:    vshufps {{.*#+}} ymm14 = ymm0[3,0],ymm1[3,0],ymm0[7,4],ymm1[7,4]2212; AVX-NEXT:    vextractf128 $1, %ymm14, %xmm142213; AVX-NEXT:    vshufps {{.*#+}} xmm9 = xmm14[2,0],xmm9[2,3]2214; AVX-NEXT:    vblendps {{.*#+}} ymm9 = ymm9[0,1,2,3],ymm7[4,5,6,7]2215; AVX-NEXT:    vunpcklpd {{.*#+}} ymm7 = ymm4[0],ymm5[0],ymm4[2],ymm5[2]2216; AVX-NEXT:    vunpcklps {{.*#+}} ymm14 = ymm8[0],ymm6[0],ymm8[1],ymm6[1],ymm8[4],ymm6[4],ymm8[5],ymm6[5]2217; AVX-NEXT:    vshufps {{.*#+}} ymm7 = ymm14[0,1],ymm7[2,0],ymm14[4,5],ymm7[6,4]2218; AVX-NEXT:    vunpcklpd {{.*#+}} ymm14 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]2219; AVX-NEXT:    vextractf128 $1, %ymm14, %xmm142220; AVX-NEXT:    vunpcklps {{.*#+}} ymm13 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[4],ymm0[4],ymm1[5],ymm0[5]2221; AVX-NEXT:    vextractf128 $1, %ymm13, %xmm132222; AVX-NEXT:    vshufps {{.*#+}} xmm13 = xmm13[0,1],xmm14[2,0]2223; AVX-NEXT:    vblendps {{.*#+}} ymm7 = ymm13[0,1,2,3],ymm7[4,5,6,7]2224; AVX-NEXT:    vunpcklps {{.*#+}} ymm4 = ymm5[0],ymm4[0],ymm5[1],ymm4[1],ymm5[4],ymm4[4],ymm5[5],ymm4[5]2225; AVX-NEXT:    vshufps {{.*#+}} ymm5 = ymm6[1,0],ymm8[1,0],ymm6[5,4],ymm8[5,4]2226; AVX-NEXT:    vshufps {{.*#+}} ymm4 = ymm5[2,0],ymm4[2,3],ymm5[6,4],ymm4[6,7]2227; AVX-NEXT:    vunpcklps {{.*#+}} ymm2 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[4],ymm2[4],ymm3[5],ymm2[5]2228; AVX-NEXT:    vshufps {{.*#+}} ymm0 = ymm0[1,0],ymm1[1,0],ymm0[5,4],ymm1[5,4]2229; AVX-NEXT:    vextractf128 $1, %ymm2, %xmm12230; AVX-NEXT:    vextractf128 $1, %ymm0, %xmm02231; AVX-NEXT:    vshufps {{.*#+}} xmm0 = xmm0[2,0],xmm1[2,3]2232; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm4[4,5,6,7]2233; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload2234; AVX-NEXT:    vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload2235; AVX-NEXT:    # xmm1 = xmm1[2],mem[2],xmm1[3],mem[3]2236; AVX-NEXT:    vunpckhps {{.*#+}} xmm2 = xmm11[2],xmm15[2],xmm11[3],xmm15[3]2237; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Reload2238; AVX-NEXT:    vshufps {{.*#+}} xmm3 = xmm11[2,2,2,2]2239; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 16-byte Reload2240; AVX-NEXT:    vblendps {{.*#+}} xmm3 = xmm8[0,1,2],xmm3[3]2241; AVX-NEXT:    vinsertf128 $1, %xmm3, %ymm0, %ymm32242; AVX-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm42243; AVX-NEXT:    vblendps {{.*#+}} ymm3 = ymm4[0,1,2,3,4,5],ymm3[6,7]2244; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload2245; AVX-NEXT:    vshufps {{.*#+}} xmm4 = xmm5[2,2,2,2]2246; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload2247; AVX-NEXT:    vblendps {{.*#+}} xmm4 = xmm6[0,1,2],xmm4[3]2248; AVX-NEXT:    vblendps {{.*#+}} xmm4 = xmm1[0,1],xmm4[2,3]2249; AVX-NEXT:    vblendps {{.*#+}} ymm3 = ymm4[0,1,2,3],ymm3[4,5,6,7]2250; AVX-NEXT:    vunpckhps {{.*#+}} xmm4 = xmm6[2],xmm5[2],xmm6[3],xmm5[3]2251; AVX-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm4[1]2252; AVX-NEXT:    vunpckhps {{.*#+}} xmm4 = xmm8[2],xmm11[2],xmm8[3],xmm11[3]2253; AVX-NEXT:    vinsertf128 $1, %xmm4, %ymm0, %ymm42254; AVX-NEXT:    vshufps {{.*#+}} xmm2 = xmm2[2,3,2,3]2255; AVX-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm22256; AVX-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm4[6,7]2257; AVX-NEXT:    vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]2258; AVX-NEXT:    movq {{[0-9]+}}(%rsp), %rax2259; AVX-NEXT:    vmovaps %ymm1, 96(%rax)2260; AVX-NEXT:    vmovaps %ymm3, 64(%rax)2261; AVX-NEXT:    vmovaps %ymm0, 160(%rax)2262; AVX-NEXT:    vmovaps %ymm7, 128(%rax)2263; AVX-NEXT:    vmovaps %ymm9, 224(%rax)2264; AVX-NEXT:    vmovaps %ymm10, 192(%rax)2265; AVX-NEXT:    vmovaps %ymm12, 288(%rax)2266; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload2267; AVX-NEXT:    vmovaps %ymm0, 256(%rax)2268; AVX-NEXT:    vmovups (%rsp), %ymm0 # 32-byte Reload2269; AVX-NEXT:    vmovaps %ymm0, 352(%rax)2270; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload2271; AVX-NEXT:    vmovaps %ymm0, 320(%rax)2272; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload2273; AVX-NEXT:    vmovaps %ymm0, 416(%rax)2274; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload2275; AVX-NEXT:    vmovaps %ymm0, 384(%rax)2276; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload2277; AVX-NEXT:    vmovaps %ymm0, 480(%rax)2278; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload2279; AVX-NEXT:    vmovaps %ymm0, 448(%rax)2280; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload2281; AVX-NEXT:    vmovaps %ymm0, (%rax)2282; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload2283; AVX-NEXT:    vmovaps %ymm0, 32(%rax)2284; AVX-NEXT:    addq $264, %rsp # imm = 0x1082285; AVX-NEXT:    vzeroupper2286; AVX-NEXT:    retq2287;2288; AVX2-LABEL: store_i32_stride8_vf16:2289; AVX2:       # %bb.0:2290; AVX2-NEXT:    subq $328, %rsp # imm = 0x1482291; AVX2-NEXT:    movq {{[0-9]+}}(%rsp), %rax2292; AVX2-NEXT:    movq {{[0-9]+}}(%rsp), %r102293; AVX2-NEXT:    vmovaps (%rdi), %ymm72294; AVX2-NEXT:    vmovaps 32(%rdi), %ymm32295; AVX2-NEXT:    vmovaps (%rsi), %ymm82296; AVX2-NEXT:    vmovaps 32(%rsi), %ymm42297; AVX2-NEXT:    vmovaps (%rdx), %ymm02298; AVX2-NEXT:    vmovaps 32(%rdx), %ymm102299; AVX2-NEXT:    vmovaps 32(%rcx), %ymm112300; AVX2-NEXT:    vmovaps 32(%r8), %ymm92301; AVX2-NEXT:    vmovaps 32(%r9), %ymm62302; AVX2-NEXT:    vmovaps 32(%r10), %ymm122303; AVX2-NEXT:    vmovaps 32(%rax), %ymm132304; AVX2-NEXT:    vbroadcastss 56(%rax), %ymm12305; AVX2-NEXT:    vblendps {{.*#+}} ymm1 = ymm12[0,1,2,3,4,5,6],ymm1[7]2306; AVX2-NEXT:    vunpckhps {{.*#+}} ymm2 = ymm3[2],ymm4[2],ymm3[3],ymm4[3],ymm3[6],ymm4[6],ymm3[7],ymm4[7]2307; AVX2-NEXT:    vextractf128 $1, %ymm2, %xmm52308; AVX2-NEXT:    vunpckhps {{.*#+}} ymm14 = ymm10[2],ymm11[2],ymm10[3],ymm11[3],ymm10[6],ymm11[6],ymm10[7],ymm11[7]2309; AVX2-NEXT:    vpermpd {{.*#+}} ymm15 = ymm14[2,2,2,2]2310; AVX2-NEXT:    vblendps {{.*#+}} xmm5 = xmm5[0,1],xmm15[2,3]2311; AVX2-NEXT:    vunpckhps {{.*#+}} ymm15 = ymm9[2],ymm6[2],ymm9[3],ymm6[3],ymm9[6],ymm6[6],ymm9[7],ymm6[7]2312; AVX2-NEXT:    vblendps {{.*#+}} ymm1 = ymm15[0,1,2,3,4,5],ymm1[6,7]2313; AVX2-NEXT:    vblendps {{.*#+}} ymm1 = ymm5[0,1,2,3],ymm1[4,5,6,7]2314; AVX2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2315; AVX2-NEXT:    vbroadcastss 60(%r10), %ymm12316; AVX2-NEXT:    vunpckhps {{.*#+}} ymm1 = ymm1[2],ymm13[2],ymm1[3],ymm13[3],ymm1[6],ymm13[6],ymm1[7],ymm13[7]2317; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm15 = ymm15[1],ymm1[1],ymm15[3],ymm1[3]2318; AVX2-NEXT:    vmovaps (%rcx), %ymm12319; AVX2-NEXT:    vextractf128 $1, %ymm14, %xmm52320; AVX2-NEXT:    vshufps {{.*#+}} ymm2 = ymm2[2,3,2,3,6,7,6,7]2321; AVX2-NEXT:    vextractf128 $1, %ymm2, %xmm22322; AVX2-NEXT:    vblendps {{.*#+}} xmm2 = xmm2[0,1],xmm5[2,3]2323; AVX2-NEXT:    vmovaps (%r8), %ymm52324; AVX2-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm15[4,5,6,7]2325; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2326; AVX2-NEXT:    vmovaps (%r9), %ymm22327; AVX2-NEXT:    vunpcklps {{.*#+}} ymm12 = ymm12[0],ymm13[0],ymm12[1],ymm13[1],ymm12[4],ymm13[4],ymm12[5],ymm13[5]2328; AVX2-NEXT:    vunpcklps {{.*#+}} ymm13 = ymm3[0],ymm4[0],ymm3[1],ymm4[1],ymm3[4],ymm4[4],ymm3[5],ymm4[5]2329; AVX2-NEXT:    vextractf128 $1, %ymm13, %xmm132330; AVX2-NEXT:    vunpcklps {{.*#+}} ymm11 = ymm10[0],ymm11[0],ymm10[1],ymm11[1],ymm10[4],ymm11[4],ymm10[5],ymm11[5]2331; AVX2-NEXT:    vpermpd {{.*#+}} ymm10 = ymm11[2,2,2,2]2332; AVX2-NEXT:    vblendps {{.*#+}} xmm13 = xmm13[0,1],xmm10[2,3]2333; AVX2-NEXT:    vmovaps (%r10), %ymm102334; AVX2-NEXT:    vunpcklps {{.*#+}} ymm9 = ymm9[0],ymm6[0],ymm9[1],ymm6[1],ymm9[4],ymm6[4],ymm9[5],ymm6[5]2335; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm9 = ymm9[0],ymm12[0],ymm9[2],ymm12[2]2336; AVX2-NEXT:    vblendps {{.*#+}} ymm9 = ymm13[0,1,2,3],ymm9[4,5,6,7]2337; AVX2-NEXT:    vmovups %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2338; AVX2-NEXT:    vbroadcastss 52(%r8), %ymm92339; AVX2-NEXT:    vblendps {{.*#+}} ymm6 = ymm9[0,1,2,3,4],ymm6[5],ymm9[6,7]2340; AVX2-NEXT:    vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3,4,5],ymm12[6,7]2341; AVX2-NEXT:    vextractf128 $1, %ymm11, %xmm92342; AVX2-NEXT:    vshufps {{.*#+}} ymm3 = ymm3[1,1,1,1,5,5,5,5]2343; AVX2-NEXT:    vblendps {{.*#+}} ymm3 = ymm3[0],ymm4[1],ymm3[2,3,4],ymm4[5],ymm3[6,7]2344; AVX2-NEXT:    vextractf128 $1, %ymm3, %xmm32345; AVX2-NEXT:    vblendps {{.*#+}} xmm3 = xmm3[0,1],xmm9[2,3]2346; AVX2-NEXT:    vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm6[4,5,6,7]2347; AVX2-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2348; AVX2-NEXT:    vbroadcastss 24(%rax), %ymm32349; AVX2-NEXT:    vblendps {{.*#+}} ymm3 = ymm10[0,1,2,3,4,5,6],ymm3[7]2350; AVX2-NEXT:    vunpckhps {{.*#+}} ymm4 = ymm7[2],ymm8[2],ymm7[3],ymm8[3],ymm7[6],ymm8[6],ymm7[7],ymm8[7]2351; AVX2-NEXT:    vextractf128 $1, %ymm4, %xmm62352; AVX2-NEXT:    vunpckhps {{.*#+}} ymm9 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7]2353; AVX2-NEXT:    vpermpd {{.*#+}} ymm11 = ymm9[2,2,2,2]2354; AVX2-NEXT:    vblendps {{.*#+}} xmm6 = xmm6[0,1],xmm11[2,3]2355; AVX2-NEXT:    vunpckhps {{.*#+}} ymm11 = ymm5[2],ymm2[2],ymm5[3],ymm2[3],ymm5[6],ymm2[6],ymm5[7],ymm2[7]2356; AVX2-NEXT:    vblendps {{.*#+}} ymm3 = ymm11[0,1,2,3,4,5],ymm3[6,7]2357; AVX2-NEXT:    vblendps {{.*#+}} ymm3 = ymm6[0,1,2,3],ymm3[4,5,6,7]2358; AVX2-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2359; AVX2-NEXT:    vmovaps (%rax), %ymm32360; AVX2-NEXT:    vbroadcastss 28(%r10), %ymm62361; AVX2-NEXT:    vunpckhps {{.*#+}} ymm6 = ymm6[2],ymm3[2],ymm6[3],ymm3[3],ymm6[6],ymm3[6],ymm6[7],ymm3[7]2362; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm6 = ymm11[1],ymm6[1],ymm11[3],ymm6[3]2363; AVX2-NEXT:    vextractf128 $1, %ymm9, %xmm92364; AVX2-NEXT:    vshufps {{.*#+}} ymm4 = ymm4[2,3,2,3,6,7,6,7]2365; AVX2-NEXT:    vextractf128 $1, %ymm4, %xmm42366; AVX2-NEXT:    vblendps {{.*#+}} xmm4 = xmm4[0,1],xmm9[2,3]2367; AVX2-NEXT:    vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm6[4,5,6,7]2368; AVX2-NEXT:    vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2369; AVX2-NEXT:    vunpcklps {{.*#+}} ymm3 = ymm10[0],ymm3[0],ymm10[1],ymm3[1],ymm10[4],ymm3[4],ymm10[5],ymm3[5]2370; AVX2-NEXT:    vunpcklps {{.*#+}} ymm4 = ymm5[0],ymm2[0],ymm5[1],ymm2[1],ymm5[4],ymm2[4],ymm5[5],ymm2[5]2371; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm4 = ymm4[0],ymm3[0],ymm4[2],ymm3[2]2372; AVX2-NEXT:    vunpcklps {{.*#+}} ymm5 = ymm7[0],ymm8[0],ymm7[1],ymm8[1],ymm7[4],ymm8[4],ymm7[5],ymm8[5]2373; AVX2-NEXT:    vextractf128 $1, %ymm5, %xmm52374; AVX2-NEXT:    vunpcklps {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]2375; AVX2-NEXT:    vpermpd {{.*#+}} ymm1 = ymm0[2,2,2,2]2376; AVX2-NEXT:    vblendps {{.*#+}} xmm1 = xmm5[0,1],xmm1[2,3]2377; AVX2-NEXT:    vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm4[4,5,6,7]2378; AVX2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2379; AVX2-NEXT:    vbroadcastss 20(%r8), %ymm12380; AVX2-NEXT:    vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3,4],ymm2[5],ymm1[6,7]2381; AVX2-NEXT:    vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5],ymm3[6,7]2382; AVX2-NEXT:    vextractf128 $1, %ymm0, %xmm02383; AVX2-NEXT:    vshufps {{.*#+}} ymm2 = ymm7[1,1,1,1,5,5,5,5]2384; AVX2-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0],ymm8[1],ymm2[2,3,4],ymm8[5],ymm2[6,7]2385; AVX2-NEXT:    vextractf128 $1, %ymm2, %xmm22386; AVX2-NEXT:    vblendps {{.*#+}} xmm0 = xmm2[0,1],xmm0[2,3]2387; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]2388; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2389; AVX2-NEXT:    vmovaps (%rcx), %xmm32390; AVX2-NEXT:    vmovaps (%rdx), %xmm142391; AVX2-NEXT:    vunpcklps {{.*#+}} xmm0 = xmm14[0],xmm3[0],xmm14[1],xmm3[1]2392; AVX2-NEXT:    vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2393; AVX2-NEXT:    vmovaps (%rsi), %xmm52394; AVX2-NEXT:    vmovaps (%rdi), %xmm42395; AVX2-NEXT:    vshufps {{.*#+}} xmm1 = xmm4[1,1,1,1]2396; AVX2-NEXT:    vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2397; AVX2-NEXT:    vblendps {{.*#+}} xmm1 = xmm1[0],xmm5[1],xmm1[2,3]2398; AVX2-NEXT:    vmovaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2399; AVX2-NEXT:    vblendps {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3]2400; AVX2-NEXT:    vmovaps (%rax), %xmm112401; AVX2-NEXT:    vmovaps (%r10), %xmm122402; AVX2-NEXT:    vunpcklps {{.*#+}} xmm1 = xmm12[0],xmm11[0],xmm12[1],xmm11[1]2403; AVX2-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm12404; AVX2-NEXT:    vmovaps (%r9), %xmm62405; AVX2-NEXT:    vmovaps (%r8), %xmm132406; AVX2-NEXT:    vshufps {{.*#+}} xmm2 = xmm13[1,1,1,1]2407; AVX2-NEXT:    vblendps {{.*#+}} xmm2 = xmm2[0],xmm6[1],xmm2[2,3]2408; AVX2-NEXT:    vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2409; AVX2-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm22410; AVX2-NEXT:    vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]2411; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]2412; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2413; AVX2-NEXT:    vbroadcastss %xmm3, %xmm02414; AVX2-NEXT:    vbroadcastss %xmm14, %xmm12415; AVX2-NEXT:    vunpcklps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]2416; AVX2-NEXT:    vunpcklps {{.*#+}} xmm1 = xmm4[0],xmm5[0],xmm4[1],xmm5[1]2417; AVX2-NEXT:    vblendps {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3]2418; AVX2-NEXT:    vbroadcastss %xmm11, %xmm12419; AVX2-NEXT:    vbroadcastss %xmm12, %xmm22420; AVX2-NEXT:    vunpcklps {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]2421; AVX2-NEXT:    vunpcklps {{.*#+}} xmm2 = xmm13[0],xmm6[0],xmm13[1],xmm6[1]2422; AVX2-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm22423; AVX2-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm12424; AVX2-NEXT:    vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]2425; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]2426; AVX2-NEXT:    vmovups %ymm0, (%rsp) # 32-byte Spill2427; AVX2-NEXT:    vmovaps 32(%rax), %xmm72428; AVX2-NEXT:    vmovaps 32(%r10), %xmm62429; AVX2-NEXT:    vshufps {{.*#+}} xmm0 = xmm7[2,2,2,2]2430; AVX2-NEXT:    vblendps {{.*#+}} xmm0 = xmm6[0,1,2],xmm0[3]2431; AVX2-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm02432; AVX2-NEXT:    vmovaps 32(%r9), %xmm102433; AVX2-NEXT:    vmovaps 32(%r8), %xmm92434; AVX2-NEXT:    vunpckhps {{.*#+}} xmm15 = xmm9[2],xmm10[2],xmm9[3],xmm10[3]2435; AVX2-NEXT:    vinsertf128 $1, %xmm15, %ymm0, %ymm12436; AVX2-NEXT:    vblendps {{.*#+}} ymm8 = ymm1[0,1,2,3,4,5],ymm0[6,7]2437; AVX2-NEXT:    vmovaps 32(%rcx), %xmm32438; AVX2-NEXT:    vmovaps 32(%rdx), %xmm12439; AVX2-NEXT:    vshufps {{.*#+}} xmm2 = xmm3[2,2,2,2]2440; AVX2-NEXT:    vblendps {{.*#+}} xmm5 = xmm1[0,1,2],xmm2[3]2441; AVX2-NEXT:    vmovaps 32(%rsi), %xmm42442; AVX2-NEXT:    vmovaps 32(%rdi), %xmm22443; AVX2-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm2[2],xmm4[2],xmm2[3],xmm4[3]2444; AVX2-NEXT:    vblendps {{.*#+}} xmm5 = xmm0[0,1],xmm5[2,3]2445; AVX2-NEXT:    vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm8[4,5,6,7]2446; AVX2-NEXT:    vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2447; AVX2-NEXT:    vunpckhps {{.*#+}} xmm5 = xmm1[2],xmm3[2],xmm1[3],xmm3[3]2448; AVX2-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm5[1]2449; AVX2-NEXT:    vunpckhps {{.*#+}} xmm5 = xmm6[2],xmm7[2],xmm6[3],xmm7[3]2450; AVX2-NEXT:    vinsertf128 $1, %xmm5, %ymm0, %ymm52451; AVX2-NEXT:    vshufps {{.*#+}} xmm8 = xmm15[2,3,2,3]2452; AVX2-NEXT:    vinsertf128 $1, %xmm8, %ymm0, %ymm82453; AVX2-NEXT:    vblendps {{.*#+}} ymm5 = ymm8[0,1,2,3,4,5],ymm5[6,7]2454; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm5[4,5,6,7]2455; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2456; AVX2-NEXT:    vbroadcastss %xmm3, %xmm02457; AVX2-NEXT:    vbroadcastss %xmm1, %xmm52458; AVX2-NEXT:    vunpcklps {{.*#+}} xmm0 = xmm5[0],xmm0[0],xmm5[1],xmm0[1]2459; AVX2-NEXT:    vunpcklps {{.*#+}} xmm5 = xmm2[0],xmm4[0],xmm2[1],xmm4[1]2460; AVX2-NEXT:    vblendps {{.*#+}} xmm0 = xmm5[0,1],xmm0[2,3]2461; AVX2-NEXT:    vbroadcastss %xmm7, %xmm52462; AVX2-NEXT:    vbroadcastss %xmm6, %xmm82463; AVX2-NEXT:    vunpcklps {{.*#+}} xmm5 = xmm8[0],xmm5[0],xmm8[1],xmm5[1]2464; AVX2-NEXT:    vunpcklps {{.*#+}} xmm8 = xmm9[0],xmm10[0],xmm9[1],xmm10[1]2465; AVX2-NEXT:    vinsertf128 $1, %xmm8, %ymm0, %ymm82466; AVX2-NEXT:    vinsertf128 $1, %xmm5, %ymm0, %ymm52467; AVX2-NEXT:    vblendps {{.*#+}} ymm5 = ymm8[0,1,2,3,4,5],ymm5[6,7]2468; AVX2-NEXT:    vblendps {{.*#+}} ymm15 = ymm0[0,1,2,3],ymm5[4,5,6,7]2469; AVX2-NEXT:    vunpcklps {{.*#+}} xmm0 = xmm1[0],xmm3[0],xmm1[1],xmm3[1]2470; AVX2-NEXT:    vshufps {{.*#+}} xmm1 = xmm2[1,1,1,1]2471; AVX2-NEXT:    vblendps {{.*#+}} xmm1 = xmm1[0],xmm4[1],xmm1[2,3]2472; AVX2-NEXT:    vblendps {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3]2473; AVX2-NEXT:    vunpcklps {{.*#+}} xmm1 = xmm6[0],xmm7[0],xmm6[1],xmm7[1]2474; AVX2-NEXT:    vshufps {{.*#+}} xmm2 = xmm9[1,1,1,1]2475; AVX2-NEXT:    vblendps {{.*#+}} xmm2 = xmm2[0],xmm10[1],xmm2[2,3]2476; AVX2-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm12477; AVX2-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm22478; AVX2-NEXT:    vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]2479; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]2480; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload2481; AVX2-NEXT:    vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload2482; AVX2-NEXT:    # xmm1 = xmm1[2],mem[2],xmm1[3],mem[3]2483; AVX2-NEXT:    vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm13, %xmm2 # 16-byte Folded Reload2484; AVX2-NEXT:    # xmm2 = xmm13[2],mem[2],xmm13[3],mem[3]2485; AVX2-NEXT:    vshufps {{.*#+}} xmm3 = xmm11[2,2,2,2]2486; AVX2-NEXT:    vblendps {{.*#+}} xmm3 = xmm12[0,1,2],xmm3[3]2487; AVX2-NEXT:    vinsertf128 $1, %xmm3, %ymm0, %ymm32488; AVX2-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm42489; AVX2-NEXT:    vblendps {{.*#+}} ymm3 = ymm4[0,1,2,3,4,5],ymm3[6,7]2490; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload2491; AVX2-NEXT:    vshufps {{.*#+}} xmm4 = xmm5[2,2,2,2]2492; AVX2-NEXT:    vblendps {{.*#+}} xmm4 = xmm14[0,1,2],xmm4[3]2493; AVX2-NEXT:    vblendps {{.*#+}} xmm4 = xmm1[0,1],xmm4[2,3]2494; AVX2-NEXT:    vblendps {{.*#+}} ymm3 = ymm4[0,1,2,3],ymm3[4,5,6,7]2495; AVX2-NEXT:    vunpckhps {{.*#+}} xmm4 = xmm14[2],xmm5[2],xmm14[3],xmm5[3]2496; AVX2-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm4[1]2497; AVX2-NEXT:    vunpckhps {{.*#+}} xmm4 = xmm12[2],xmm11[2],xmm12[3],xmm11[3]2498; AVX2-NEXT:    vinsertf128 $1, %xmm4, %ymm0, %ymm42499; AVX2-NEXT:    vshufps {{.*#+}} xmm2 = xmm2[2,3,2,3]2500; AVX2-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm22501; AVX2-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm4[6,7]2502; AVX2-NEXT:    vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]2503; AVX2-NEXT:    movq {{[0-9]+}}(%rsp), %rax2504; AVX2-NEXT:    vmovaps %ymm1, 96(%rax)2505; AVX2-NEXT:    vmovaps %ymm3, 64(%rax)2506; AVX2-NEXT:    vmovaps %ymm0, 288(%rax)2507; AVX2-NEXT:    vmovaps %ymm15, 256(%rax)2508; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload2509; AVX2-NEXT:    vmovaps %ymm0, 352(%rax)2510; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload2511; AVX2-NEXT:    vmovaps %ymm0, 320(%rax)2512; AVX2-NEXT:    vmovups (%rsp), %ymm0 # 32-byte Reload2513; AVX2-NEXT:    vmovaps %ymm0, (%rax)2514; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload2515; AVX2-NEXT:    vmovaps %ymm0, 32(%rax)2516; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload2517; AVX2-NEXT:    vmovaps %ymm0, 160(%rax)2518; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload2519; AVX2-NEXT:    vmovaps %ymm0, 128(%rax)2520; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload2521; AVX2-NEXT:    vmovaps %ymm0, 224(%rax)2522; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload2523; AVX2-NEXT:    vmovaps %ymm0, 192(%rax)2524; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload2525; AVX2-NEXT:    vmovaps %ymm0, 416(%rax)2526; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload2527; AVX2-NEXT:    vmovaps %ymm0, 384(%rax)2528; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload2529; AVX2-NEXT:    vmovaps %ymm0, 480(%rax)2530; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload2531; AVX2-NEXT:    vmovaps %ymm0, 448(%rax)2532; AVX2-NEXT:    addq $328, %rsp # imm = 0x1482533; AVX2-NEXT:    vzeroupper2534; AVX2-NEXT:    retq2535;2536; AVX2-FP-LABEL: store_i32_stride8_vf16:2537; AVX2-FP:       # %bb.0:2538; AVX2-FP-NEXT:    subq $328, %rsp # imm = 0x1482539; AVX2-FP-NEXT:    movq {{[0-9]+}}(%rsp), %rax2540; AVX2-FP-NEXT:    movq {{[0-9]+}}(%rsp), %r102541; AVX2-FP-NEXT:    vmovaps (%rdi), %ymm72542; AVX2-FP-NEXT:    vmovaps 32(%rdi), %ymm32543; AVX2-FP-NEXT:    vmovaps (%rsi), %ymm82544; AVX2-FP-NEXT:    vmovaps 32(%rsi), %ymm42545; AVX2-FP-NEXT:    vmovaps (%rdx), %ymm02546; AVX2-FP-NEXT:    vmovaps 32(%rdx), %ymm102547; AVX2-FP-NEXT:    vmovaps 32(%rcx), %ymm112548; AVX2-FP-NEXT:    vmovaps 32(%r8), %ymm92549; AVX2-FP-NEXT:    vmovaps 32(%r9), %ymm62550; AVX2-FP-NEXT:    vmovaps 32(%r10), %ymm122551; AVX2-FP-NEXT:    vmovaps 32(%rax), %ymm132552; AVX2-FP-NEXT:    vbroadcastss 56(%rax), %ymm12553; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm1 = ymm12[0,1,2,3,4,5,6],ymm1[7]2554; AVX2-FP-NEXT:    vunpckhps {{.*#+}} ymm2 = ymm3[2],ymm4[2],ymm3[3],ymm4[3],ymm3[6],ymm4[6],ymm3[7],ymm4[7]2555; AVX2-FP-NEXT:    vextractf128 $1, %ymm2, %xmm52556; AVX2-FP-NEXT:    vunpckhps {{.*#+}} ymm14 = ymm10[2],ymm11[2],ymm10[3],ymm11[3],ymm10[6],ymm11[6],ymm10[7],ymm11[7]2557; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm15 = ymm14[2,2,2,2]2558; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm5 = xmm5[0,1],xmm15[2,3]2559; AVX2-FP-NEXT:    vunpckhps {{.*#+}} ymm15 = ymm9[2],ymm6[2],ymm9[3],ymm6[3],ymm9[6],ymm6[6],ymm9[7],ymm6[7]2560; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm1 = ymm15[0,1,2,3,4,5],ymm1[6,7]2561; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm1 = ymm5[0,1,2,3],ymm1[4,5,6,7]2562; AVX2-FP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2563; AVX2-FP-NEXT:    vbroadcastss 60(%r10), %ymm12564; AVX2-FP-NEXT:    vunpckhps {{.*#+}} ymm1 = ymm1[2],ymm13[2],ymm1[3],ymm13[3],ymm1[6],ymm13[6],ymm1[7],ymm13[7]2565; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm15 = ymm15[1],ymm1[1],ymm15[3],ymm1[3]2566; AVX2-FP-NEXT:    vmovaps (%rcx), %ymm12567; AVX2-FP-NEXT:    vextractf128 $1, %ymm14, %xmm52568; AVX2-FP-NEXT:    vshufps {{.*#+}} ymm2 = ymm2[2,3,2,3,6,7,6,7]2569; AVX2-FP-NEXT:    vextractf128 $1, %ymm2, %xmm22570; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm2 = xmm2[0,1],xmm5[2,3]2571; AVX2-FP-NEXT:    vmovaps (%r8), %ymm52572; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm15[4,5,6,7]2573; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2574; AVX2-FP-NEXT:    vmovaps (%r9), %ymm22575; AVX2-FP-NEXT:    vunpcklps {{.*#+}} ymm12 = ymm12[0],ymm13[0],ymm12[1],ymm13[1],ymm12[4],ymm13[4],ymm12[5],ymm13[5]2576; AVX2-FP-NEXT:    vunpcklps {{.*#+}} ymm13 = ymm3[0],ymm4[0],ymm3[1],ymm4[1],ymm3[4],ymm4[4],ymm3[5],ymm4[5]2577; AVX2-FP-NEXT:    vextractf128 $1, %ymm13, %xmm132578; AVX2-FP-NEXT:    vunpcklps {{.*#+}} ymm11 = ymm10[0],ymm11[0],ymm10[1],ymm11[1],ymm10[4],ymm11[4],ymm10[5],ymm11[5]2579; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm10 = ymm11[2,2,2,2]2580; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm13 = xmm13[0,1],xmm10[2,3]2581; AVX2-FP-NEXT:    vmovaps (%r10), %ymm102582; AVX2-FP-NEXT:    vunpcklps {{.*#+}} ymm9 = ymm9[0],ymm6[0],ymm9[1],ymm6[1],ymm9[4],ymm6[4],ymm9[5],ymm6[5]2583; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm9 = ymm9[0],ymm12[0],ymm9[2],ymm12[2]2584; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm9 = ymm13[0,1,2,3],ymm9[4,5,6,7]2585; AVX2-FP-NEXT:    vmovups %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2586; AVX2-FP-NEXT:    vbroadcastss 52(%r8), %ymm92587; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm6 = ymm9[0,1,2,3,4],ymm6[5],ymm9[6,7]2588; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3,4,5],ymm12[6,7]2589; AVX2-FP-NEXT:    vextractf128 $1, %ymm11, %xmm92590; AVX2-FP-NEXT:    vshufps {{.*#+}} ymm3 = ymm3[1,1,1,1,5,5,5,5]2591; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm3 = ymm3[0],ymm4[1],ymm3[2,3,4],ymm4[5],ymm3[6,7]2592; AVX2-FP-NEXT:    vextractf128 $1, %ymm3, %xmm32593; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm3 = xmm3[0,1],xmm9[2,3]2594; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm6[4,5,6,7]2595; AVX2-FP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2596; AVX2-FP-NEXT:    vbroadcastss 24(%rax), %ymm32597; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm3 = ymm10[0,1,2,3,4,5,6],ymm3[7]2598; AVX2-FP-NEXT:    vunpckhps {{.*#+}} ymm4 = ymm7[2],ymm8[2],ymm7[3],ymm8[3],ymm7[6],ymm8[6],ymm7[7],ymm8[7]2599; AVX2-FP-NEXT:    vextractf128 $1, %ymm4, %xmm62600; AVX2-FP-NEXT:    vunpckhps {{.*#+}} ymm9 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7]2601; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm11 = ymm9[2,2,2,2]2602; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm6 = xmm6[0,1],xmm11[2,3]2603; AVX2-FP-NEXT:    vunpckhps {{.*#+}} ymm11 = ymm5[2],ymm2[2],ymm5[3],ymm2[3],ymm5[6],ymm2[6],ymm5[7],ymm2[7]2604; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm3 = ymm11[0,1,2,3,4,5],ymm3[6,7]2605; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm3 = ymm6[0,1,2,3],ymm3[4,5,6,7]2606; AVX2-FP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2607; AVX2-FP-NEXT:    vmovaps (%rax), %ymm32608; AVX2-FP-NEXT:    vbroadcastss 28(%r10), %ymm62609; AVX2-FP-NEXT:    vunpckhps {{.*#+}} ymm6 = ymm6[2],ymm3[2],ymm6[3],ymm3[3],ymm6[6],ymm3[6],ymm6[7],ymm3[7]2610; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm6 = ymm11[1],ymm6[1],ymm11[3],ymm6[3]2611; AVX2-FP-NEXT:    vextractf128 $1, %ymm9, %xmm92612; AVX2-FP-NEXT:    vshufps {{.*#+}} ymm4 = ymm4[2,3,2,3,6,7,6,7]2613; AVX2-FP-NEXT:    vextractf128 $1, %ymm4, %xmm42614; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm4 = xmm4[0,1],xmm9[2,3]2615; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm6[4,5,6,7]2616; AVX2-FP-NEXT:    vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2617; AVX2-FP-NEXT:    vunpcklps {{.*#+}} ymm3 = ymm10[0],ymm3[0],ymm10[1],ymm3[1],ymm10[4],ymm3[4],ymm10[5],ymm3[5]2618; AVX2-FP-NEXT:    vunpcklps {{.*#+}} ymm4 = ymm5[0],ymm2[0],ymm5[1],ymm2[1],ymm5[4],ymm2[4],ymm5[5],ymm2[5]2619; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm4 = ymm4[0],ymm3[0],ymm4[2],ymm3[2]2620; AVX2-FP-NEXT:    vunpcklps {{.*#+}} ymm5 = ymm7[0],ymm8[0],ymm7[1],ymm8[1],ymm7[4],ymm8[4],ymm7[5],ymm8[5]2621; AVX2-FP-NEXT:    vextractf128 $1, %ymm5, %xmm52622; AVX2-FP-NEXT:    vunpcklps {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]2623; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm1 = ymm0[2,2,2,2]2624; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm1 = xmm5[0,1],xmm1[2,3]2625; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm4[4,5,6,7]2626; AVX2-FP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2627; AVX2-FP-NEXT:    vbroadcastss 20(%r8), %ymm12628; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3,4],ymm2[5],ymm1[6,7]2629; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5],ymm3[6,7]2630; AVX2-FP-NEXT:    vextractf128 $1, %ymm0, %xmm02631; AVX2-FP-NEXT:    vshufps {{.*#+}} ymm2 = ymm7[1,1,1,1,5,5,5,5]2632; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0],ymm8[1],ymm2[2,3,4],ymm8[5],ymm2[6,7]2633; AVX2-FP-NEXT:    vextractf128 $1, %ymm2, %xmm22634; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm0 = xmm2[0,1],xmm0[2,3]2635; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]2636; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2637; AVX2-FP-NEXT:    vmovaps (%rcx), %xmm32638; AVX2-FP-NEXT:    vmovaps (%rdx), %xmm142639; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm0 = xmm14[0],xmm3[0],xmm14[1],xmm3[1]2640; AVX2-FP-NEXT:    vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2641; AVX2-FP-NEXT:    vmovaps (%rsi), %xmm52642; AVX2-FP-NEXT:    vmovaps (%rdi), %xmm42643; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm1 = xmm4[1,1,1,1]2644; AVX2-FP-NEXT:    vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2645; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm1 = xmm1[0],xmm5[1],xmm1[2,3]2646; AVX2-FP-NEXT:    vmovaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2647; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3]2648; AVX2-FP-NEXT:    vmovaps (%rax), %xmm112649; AVX2-FP-NEXT:    vmovaps (%r10), %xmm122650; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm1 = xmm12[0],xmm11[0],xmm12[1],xmm11[1]2651; AVX2-FP-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm12652; AVX2-FP-NEXT:    vmovaps (%r9), %xmm62653; AVX2-FP-NEXT:    vmovaps (%r8), %xmm132654; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm2 = xmm13[1,1,1,1]2655; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm2 = xmm2[0],xmm6[1],xmm2[2,3]2656; AVX2-FP-NEXT:    vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2657; AVX2-FP-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm22658; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]2659; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]2660; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2661; AVX2-FP-NEXT:    vbroadcastss %xmm3, %xmm02662; AVX2-FP-NEXT:    vbroadcastss %xmm14, %xmm12663; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]2664; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm1 = xmm4[0],xmm5[0],xmm4[1],xmm5[1]2665; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3]2666; AVX2-FP-NEXT:    vbroadcastss %xmm11, %xmm12667; AVX2-FP-NEXT:    vbroadcastss %xmm12, %xmm22668; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]2669; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm2 = xmm13[0],xmm6[0],xmm13[1],xmm6[1]2670; AVX2-FP-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm22671; AVX2-FP-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm12672; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]2673; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]2674; AVX2-FP-NEXT:    vmovups %ymm0, (%rsp) # 32-byte Spill2675; AVX2-FP-NEXT:    vmovaps 32(%rax), %xmm72676; AVX2-FP-NEXT:    vmovaps 32(%r10), %xmm62677; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm0 = xmm7[2,2,2,2]2678; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm0 = xmm6[0,1,2],xmm0[3]2679; AVX2-FP-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm02680; AVX2-FP-NEXT:    vmovaps 32(%r9), %xmm102681; AVX2-FP-NEXT:    vmovaps 32(%r8), %xmm92682; AVX2-FP-NEXT:    vunpckhps {{.*#+}} xmm15 = xmm9[2],xmm10[2],xmm9[3],xmm10[3]2683; AVX2-FP-NEXT:    vinsertf128 $1, %xmm15, %ymm0, %ymm12684; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm8 = ymm1[0,1,2,3,4,5],ymm0[6,7]2685; AVX2-FP-NEXT:    vmovaps 32(%rcx), %xmm32686; AVX2-FP-NEXT:    vmovaps 32(%rdx), %xmm12687; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm2 = xmm3[2,2,2,2]2688; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm5 = xmm1[0,1,2],xmm2[3]2689; AVX2-FP-NEXT:    vmovaps 32(%rsi), %xmm42690; AVX2-FP-NEXT:    vmovaps 32(%rdi), %xmm22691; AVX2-FP-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm2[2],xmm4[2],xmm2[3],xmm4[3]2692; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm5 = xmm0[0,1],xmm5[2,3]2693; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm8[4,5,6,7]2694; AVX2-FP-NEXT:    vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2695; AVX2-FP-NEXT:    vunpckhps {{.*#+}} xmm5 = xmm1[2],xmm3[2],xmm1[3],xmm3[3]2696; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm5[1]2697; AVX2-FP-NEXT:    vunpckhps {{.*#+}} xmm5 = xmm6[2],xmm7[2],xmm6[3],xmm7[3]2698; AVX2-FP-NEXT:    vinsertf128 $1, %xmm5, %ymm0, %ymm52699; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm8 = xmm15[2,3,2,3]2700; AVX2-FP-NEXT:    vinsertf128 $1, %xmm8, %ymm0, %ymm82701; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm5 = ymm8[0,1,2,3,4,5],ymm5[6,7]2702; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm5[4,5,6,7]2703; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2704; AVX2-FP-NEXT:    vbroadcastss %xmm3, %xmm02705; AVX2-FP-NEXT:    vbroadcastss %xmm1, %xmm52706; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm0 = xmm5[0],xmm0[0],xmm5[1],xmm0[1]2707; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm5 = xmm2[0],xmm4[0],xmm2[1],xmm4[1]2708; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm0 = xmm5[0,1],xmm0[2,3]2709; AVX2-FP-NEXT:    vbroadcastss %xmm7, %xmm52710; AVX2-FP-NEXT:    vbroadcastss %xmm6, %xmm82711; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm5 = xmm8[0],xmm5[0],xmm8[1],xmm5[1]2712; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm8 = xmm9[0],xmm10[0],xmm9[1],xmm10[1]2713; AVX2-FP-NEXT:    vinsertf128 $1, %xmm8, %ymm0, %ymm82714; AVX2-FP-NEXT:    vinsertf128 $1, %xmm5, %ymm0, %ymm52715; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm5 = ymm8[0,1,2,3,4,5],ymm5[6,7]2716; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm15 = ymm0[0,1,2,3],ymm5[4,5,6,7]2717; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm0 = xmm1[0],xmm3[0],xmm1[1],xmm3[1]2718; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm1 = xmm2[1,1,1,1]2719; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm1 = xmm1[0],xmm4[1],xmm1[2,3]2720; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3]2721; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm1 = xmm6[0],xmm7[0],xmm6[1],xmm7[1]2722; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm2 = xmm9[1,1,1,1]2723; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm2 = xmm2[0],xmm10[1],xmm2[2,3]2724; AVX2-FP-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm12725; AVX2-FP-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm22726; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]2727; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]2728; AVX2-FP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload2729; AVX2-FP-NEXT:    vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload2730; AVX2-FP-NEXT:    # xmm1 = xmm1[2],mem[2],xmm1[3],mem[3]2731; AVX2-FP-NEXT:    vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm13, %xmm2 # 16-byte Folded Reload2732; AVX2-FP-NEXT:    # xmm2 = xmm13[2],mem[2],xmm13[3],mem[3]2733; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm3 = xmm11[2,2,2,2]2734; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm3 = xmm12[0,1,2],xmm3[3]2735; AVX2-FP-NEXT:    vinsertf128 $1, %xmm3, %ymm0, %ymm32736; AVX2-FP-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm42737; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm3 = ymm4[0,1,2,3,4,5],ymm3[6,7]2738; AVX2-FP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload2739; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm4 = xmm5[2,2,2,2]2740; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm4 = xmm14[0,1,2],xmm4[3]2741; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm4 = xmm1[0,1],xmm4[2,3]2742; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm3 = ymm4[0,1,2,3],ymm3[4,5,6,7]2743; AVX2-FP-NEXT:    vunpckhps {{.*#+}} xmm4 = xmm14[2],xmm5[2],xmm14[3],xmm5[3]2744; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm4[1]2745; AVX2-FP-NEXT:    vunpckhps {{.*#+}} xmm4 = xmm12[2],xmm11[2],xmm12[3],xmm11[3]2746; AVX2-FP-NEXT:    vinsertf128 $1, %xmm4, %ymm0, %ymm42747; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm2 = xmm2[2,3,2,3]2748; AVX2-FP-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm22749; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm4[6,7]2750; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]2751; AVX2-FP-NEXT:    movq {{[0-9]+}}(%rsp), %rax2752; AVX2-FP-NEXT:    vmovaps %ymm1, 96(%rax)2753; AVX2-FP-NEXT:    vmovaps %ymm3, 64(%rax)2754; AVX2-FP-NEXT:    vmovaps %ymm0, 288(%rax)2755; AVX2-FP-NEXT:    vmovaps %ymm15, 256(%rax)2756; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload2757; AVX2-FP-NEXT:    vmovaps %ymm0, 352(%rax)2758; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload2759; AVX2-FP-NEXT:    vmovaps %ymm0, 320(%rax)2760; AVX2-FP-NEXT:    vmovups (%rsp), %ymm0 # 32-byte Reload2761; AVX2-FP-NEXT:    vmovaps %ymm0, (%rax)2762; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload2763; AVX2-FP-NEXT:    vmovaps %ymm0, 32(%rax)2764; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload2765; AVX2-FP-NEXT:    vmovaps %ymm0, 160(%rax)2766; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload2767; AVX2-FP-NEXT:    vmovaps %ymm0, 128(%rax)2768; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload2769; AVX2-FP-NEXT:    vmovaps %ymm0, 224(%rax)2770; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload2771; AVX2-FP-NEXT:    vmovaps %ymm0, 192(%rax)2772; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload2773; AVX2-FP-NEXT:    vmovaps %ymm0, 416(%rax)2774; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload2775; AVX2-FP-NEXT:    vmovaps %ymm0, 384(%rax)2776; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload2777; AVX2-FP-NEXT:    vmovaps %ymm0, 480(%rax)2778; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload2779; AVX2-FP-NEXT:    vmovaps %ymm0, 448(%rax)2780; AVX2-FP-NEXT:    addq $328, %rsp # imm = 0x1482781; AVX2-FP-NEXT:    vzeroupper2782; AVX2-FP-NEXT:    retq2783;2784; AVX2-FCP-LABEL: store_i32_stride8_vf16:2785; AVX2-FCP:       # %bb.0:2786; AVX2-FCP-NEXT:    subq $328, %rsp # imm = 0x1482787; AVX2-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax2788; AVX2-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r102789; AVX2-FCP-NEXT:    vmovaps (%rdi), %ymm72790; AVX2-FCP-NEXT:    vmovaps 32(%rdi), %ymm32791; AVX2-FCP-NEXT:    vmovaps (%rsi), %ymm82792; AVX2-FCP-NEXT:    vmovaps 32(%rsi), %ymm42793; AVX2-FCP-NEXT:    vmovaps (%rdx), %ymm02794; AVX2-FCP-NEXT:    vmovaps 32(%rdx), %ymm102795; AVX2-FCP-NEXT:    vmovaps 32(%rcx), %ymm112796; AVX2-FCP-NEXT:    vmovaps 32(%r8), %ymm92797; AVX2-FCP-NEXT:    vmovaps 32(%r9), %ymm62798; AVX2-FCP-NEXT:    vmovaps 32(%r10), %ymm122799; AVX2-FCP-NEXT:    vmovaps 32(%rax), %ymm132800; AVX2-FCP-NEXT:    vbroadcastss 56(%rax), %ymm12801; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm1 = ymm12[0,1,2,3,4,5,6],ymm1[7]2802; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} ymm2 = ymm3[2],ymm4[2],ymm3[3],ymm4[3],ymm3[6],ymm4[6],ymm3[7],ymm4[7]2803; AVX2-FCP-NEXT:    vextractf128 $1, %ymm2, %xmm52804; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} ymm14 = ymm10[2],ymm11[2],ymm10[3],ymm11[3],ymm10[6],ymm11[6],ymm10[7],ymm11[7]2805; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm15 = ymm14[2,2,2,2]2806; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm5 = xmm5[0,1],xmm15[2,3]2807; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} ymm15 = ymm9[2],ymm6[2],ymm9[3],ymm6[3],ymm9[6],ymm6[6],ymm9[7],ymm6[7]2808; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm1 = ymm15[0,1,2,3,4,5],ymm1[6,7]2809; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm1 = ymm5[0,1,2,3],ymm1[4,5,6,7]2810; AVX2-FCP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2811; AVX2-FCP-NEXT:    vbroadcastss 60(%r10), %ymm12812; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} ymm1 = ymm1[2],ymm13[2],ymm1[3],ymm13[3],ymm1[6],ymm13[6],ymm1[7],ymm13[7]2813; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm15 = ymm15[1],ymm1[1],ymm15[3],ymm1[3]2814; AVX2-FCP-NEXT:    vmovaps (%rcx), %ymm12815; AVX2-FCP-NEXT:    vextractf128 $1, %ymm14, %xmm52816; AVX2-FCP-NEXT:    vshufps {{.*#+}} ymm2 = ymm2[2,3,2,3,6,7,6,7]2817; AVX2-FCP-NEXT:    vextractf128 $1, %ymm2, %xmm22818; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm2 = xmm2[0,1],xmm5[2,3]2819; AVX2-FCP-NEXT:    vmovaps (%r8), %ymm52820; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm15[4,5,6,7]2821; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2822; AVX2-FCP-NEXT:    vmovaps (%r9), %ymm22823; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} ymm12 = ymm12[0],ymm13[0],ymm12[1],ymm13[1],ymm12[4],ymm13[4],ymm12[5],ymm13[5]2824; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} ymm13 = ymm3[0],ymm4[0],ymm3[1],ymm4[1],ymm3[4],ymm4[4],ymm3[5],ymm4[5]2825; AVX2-FCP-NEXT:    vextractf128 $1, %ymm13, %xmm132826; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} ymm11 = ymm10[0],ymm11[0],ymm10[1],ymm11[1],ymm10[4],ymm11[4],ymm10[5],ymm11[5]2827; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm10 = ymm11[2,2,2,2]2828; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm13 = xmm13[0,1],xmm10[2,3]2829; AVX2-FCP-NEXT:    vmovaps (%r10), %ymm102830; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} ymm9 = ymm9[0],ymm6[0],ymm9[1],ymm6[1],ymm9[4],ymm6[4],ymm9[5],ymm6[5]2831; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm9 = ymm9[0],ymm12[0],ymm9[2],ymm12[2]2832; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm9 = ymm13[0,1,2,3],ymm9[4,5,6,7]2833; AVX2-FCP-NEXT:    vmovups %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2834; AVX2-FCP-NEXT:    vbroadcastss 52(%r8), %ymm92835; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm6 = ymm9[0,1,2,3,4],ymm6[5],ymm9[6,7]2836; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3,4,5],ymm12[6,7]2837; AVX2-FCP-NEXT:    vextractf128 $1, %ymm11, %xmm92838; AVX2-FCP-NEXT:    vshufps {{.*#+}} ymm3 = ymm3[1,1,1,1,5,5,5,5]2839; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm3 = ymm3[0],ymm4[1],ymm3[2,3,4],ymm4[5],ymm3[6,7]2840; AVX2-FCP-NEXT:    vextractf128 $1, %ymm3, %xmm32841; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm3 = xmm3[0,1],xmm9[2,3]2842; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm6[4,5,6,7]2843; AVX2-FCP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2844; AVX2-FCP-NEXT:    vbroadcastss 24(%rax), %ymm32845; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm3 = ymm10[0,1,2,3,4,5,6],ymm3[7]2846; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} ymm4 = ymm7[2],ymm8[2],ymm7[3],ymm8[3],ymm7[6],ymm8[6],ymm7[7],ymm8[7]2847; AVX2-FCP-NEXT:    vextractf128 $1, %ymm4, %xmm62848; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} ymm9 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7]2849; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm11 = ymm9[2,2,2,2]2850; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm6 = xmm6[0,1],xmm11[2,3]2851; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} ymm11 = ymm5[2],ymm2[2],ymm5[3],ymm2[3],ymm5[6],ymm2[6],ymm5[7],ymm2[7]2852; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm3 = ymm11[0,1,2,3,4,5],ymm3[6,7]2853; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm3 = ymm6[0,1,2,3],ymm3[4,5,6,7]2854; AVX2-FCP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2855; AVX2-FCP-NEXT:    vmovaps (%rax), %ymm32856; AVX2-FCP-NEXT:    vbroadcastss 28(%r10), %ymm62857; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} ymm6 = ymm6[2],ymm3[2],ymm6[3],ymm3[3],ymm6[6],ymm3[6],ymm6[7],ymm3[7]2858; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm6 = ymm11[1],ymm6[1],ymm11[3],ymm6[3]2859; AVX2-FCP-NEXT:    vextractf128 $1, %ymm9, %xmm92860; AVX2-FCP-NEXT:    vshufps {{.*#+}} ymm4 = ymm4[2,3,2,3,6,7,6,7]2861; AVX2-FCP-NEXT:    vextractf128 $1, %ymm4, %xmm42862; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm4 = xmm4[0,1],xmm9[2,3]2863; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm6[4,5,6,7]2864; AVX2-FCP-NEXT:    vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2865; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} ymm3 = ymm10[0],ymm3[0],ymm10[1],ymm3[1],ymm10[4],ymm3[4],ymm10[5],ymm3[5]2866; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} ymm4 = ymm5[0],ymm2[0],ymm5[1],ymm2[1],ymm5[4],ymm2[4],ymm5[5],ymm2[5]2867; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm4 = ymm4[0],ymm3[0],ymm4[2],ymm3[2]2868; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} ymm5 = ymm7[0],ymm8[0],ymm7[1],ymm8[1],ymm7[4],ymm8[4],ymm7[5],ymm8[5]2869; AVX2-FCP-NEXT:    vextractf128 $1, %ymm5, %xmm52870; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} ymm0 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]2871; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm1 = ymm0[2,2,2,2]2872; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm1 = xmm5[0,1],xmm1[2,3]2873; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm4[4,5,6,7]2874; AVX2-FCP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2875; AVX2-FCP-NEXT:    vbroadcastss 20(%r8), %ymm12876; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3,4],ymm2[5],ymm1[6,7]2877; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3,4,5],ymm3[6,7]2878; AVX2-FCP-NEXT:    vextractf128 $1, %ymm0, %xmm02879; AVX2-FCP-NEXT:    vshufps {{.*#+}} ymm2 = ymm7[1,1,1,1,5,5,5,5]2880; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0],ymm8[1],ymm2[2,3,4],ymm8[5],ymm2[6,7]2881; AVX2-FCP-NEXT:    vextractf128 $1, %ymm2, %xmm22882; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm0 = xmm2[0,1],xmm0[2,3]2883; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]2884; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2885; AVX2-FCP-NEXT:    vmovaps (%rcx), %xmm32886; AVX2-FCP-NEXT:    vmovaps (%rdx), %xmm142887; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm0 = xmm14[0],xmm3[0],xmm14[1],xmm3[1]2888; AVX2-FCP-NEXT:    vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2889; AVX2-FCP-NEXT:    vmovaps (%rsi), %xmm52890; AVX2-FCP-NEXT:    vmovaps (%rdi), %xmm42891; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm1 = xmm4[1,1,1,1]2892; AVX2-FCP-NEXT:    vmovaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2893; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm1 = xmm1[0],xmm5[1],xmm1[2,3]2894; AVX2-FCP-NEXT:    vmovaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2895; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3]2896; AVX2-FCP-NEXT:    vmovaps (%rax), %xmm112897; AVX2-FCP-NEXT:    vmovaps (%r10), %xmm122898; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm1 = xmm12[0],xmm11[0],xmm12[1],xmm11[1]2899; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm12900; AVX2-FCP-NEXT:    vmovaps (%r9), %xmm62901; AVX2-FCP-NEXT:    vmovaps (%r8), %xmm132902; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm2 = xmm13[1,1,1,1]2903; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm2 = xmm2[0],xmm6[1],xmm2[2,3]2904; AVX2-FCP-NEXT:    vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2905; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm22906; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]2907; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]2908; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2909; AVX2-FCP-NEXT:    vbroadcastss %xmm3, %xmm02910; AVX2-FCP-NEXT:    vbroadcastss %xmm14, %xmm12911; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]2912; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm1 = xmm4[0],xmm5[0],xmm4[1],xmm5[1]2913; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3]2914; AVX2-FCP-NEXT:    vbroadcastss %xmm11, %xmm12915; AVX2-FCP-NEXT:    vbroadcastss %xmm12, %xmm22916; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]2917; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm2 = xmm13[0],xmm6[0],xmm13[1],xmm6[1]2918; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm22919; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm12920; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]2921; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]2922; AVX2-FCP-NEXT:    vmovups %ymm0, (%rsp) # 32-byte Spill2923; AVX2-FCP-NEXT:    vmovaps 32(%rax), %xmm72924; AVX2-FCP-NEXT:    vmovaps 32(%r10), %xmm62925; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm0 = xmm7[2,2,2,2]2926; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm0 = xmm6[0,1,2],xmm0[3]2927; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm02928; AVX2-FCP-NEXT:    vmovaps 32(%r9), %xmm102929; AVX2-FCP-NEXT:    vmovaps 32(%r8), %xmm92930; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} xmm15 = xmm9[2],xmm10[2],xmm9[3],xmm10[3]2931; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm15, %ymm0, %ymm12932; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm8 = ymm1[0,1,2,3,4,5],ymm0[6,7]2933; AVX2-FCP-NEXT:    vmovaps 32(%rcx), %xmm32934; AVX2-FCP-NEXT:    vmovaps 32(%rdx), %xmm12935; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm2 = xmm3[2,2,2,2]2936; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm5 = xmm1[0,1,2],xmm2[3]2937; AVX2-FCP-NEXT:    vmovaps 32(%rsi), %xmm42938; AVX2-FCP-NEXT:    vmovaps 32(%rdi), %xmm22939; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm2[2],xmm4[2],xmm2[3],xmm4[3]2940; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm5 = xmm0[0,1],xmm5[2,3]2941; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm8[4,5,6,7]2942; AVX2-FCP-NEXT:    vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2943; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} xmm5 = xmm1[2],xmm3[2],xmm1[3],xmm3[3]2944; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm5[1]2945; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} xmm5 = xmm6[2],xmm7[2],xmm6[3],xmm7[3]2946; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm5, %ymm0, %ymm52947; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm8 = xmm15[2,3,2,3]2948; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm8, %ymm0, %ymm82949; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm5 = ymm8[0,1,2,3,4,5],ymm5[6,7]2950; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm5[4,5,6,7]2951; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2952; AVX2-FCP-NEXT:    vbroadcastss %xmm3, %xmm02953; AVX2-FCP-NEXT:    vbroadcastss %xmm1, %xmm52954; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm0 = xmm5[0],xmm0[0],xmm5[1],xmm0[1]2955; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm5 = xmm2[0],xmm4[0],xmm2[1],xmm4[1]2956; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm0 = xmm5[0,1],xmm0[2,3]2957; AVX2-FCP-NEXT:    vbroadcastss %xmm7, %xmm52958; AVX2-FCP-NEXT:    vbroadcastss %xmm6, %xmm82959; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm5 = xmm8[0],xmm5[0],xmm8[1],xmm5[1]2960; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm8 = xmm9[0],xmm10[0],xmm9[1],xmm10[1]2961; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm8, %ymm0, %ymm82962; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm5, %ymm0, %ymm52963; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm5 = ymm8[0,1,2,3,4,5],ymm5[6,7]2964; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm15 = ymm0[0,1,2,3],ymm5[4,5,6,7]2965; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm0 = xmm1[0],xmm3[0],xmm1[1],xmm3[1]2966; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm1 = xmm2[1,1,1,1]2967; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm1 = xmm1[0],xmm4[1],xmm1[2,3]2968; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,3]2969; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm1 = xmm6[0],xmm7[0],xmm6[1],xmm7[1]2970; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm2 = xmm9[1,1,1,1]2971; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm2 = xmm2[0],xmm10[1],xmm2[2,3]2972; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm12973; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm22974; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]2975; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]2976; AVX2-FCP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload2977; AVX2-FCP-NEXT:    vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload2978; AVX2-FCP-NEXT:    # xmm1 = xmm1[2],mem[2],xmm1[3],mem[3]2979; AVX2-FCP-NEXT:    vunpckhps {{[-0-9]+}}(%r{{[sb]}}p), %xmm13, %xmm2 # 16-byte Folded Reload2980; AVX2-FCP-NEXT:    # xmm2 = xmm13[2],mem[2],xmm13[3],mem[3]2981; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm3 = xmm11[2,2,2,2]2982; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm3 = xmm12[0,1,2],xmm3[3]2983; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm3, %ymm0, %ymm32984; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm42985; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm3 = ymm4[0,1,2,3,4,5],ymm3[6,7]2986; AVX2-FCP-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload2987; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm4 = xmm5[2,2,2,2]2988; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm4 = xmm14[0,1,2],xmm4[3]2989; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm4 = xmm1[0,1],xmm4[2,3]2990; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm3 = ymm4[0,1,2,3],ymm3[4,5,6,7]2991; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} xmm4 = xmm14[2],xmm5[2],xmm14[3],xmm5[3]2992; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm1[1],xmm4[1]2993; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} xmm4 = xmm12[2],xmm11[2],xmm12[3],xmm11[3]2994; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm4, %ymm0, %ymm42995; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm2 = xmm2[2,3,2,3]2996; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm22997; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm4[6,7]2998; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm2[4,5,6,7]2999; AVX2-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax3000; AVX2-FCP-NEXT:    vmovaps %ymm1, 96(%rax)3001; AVX2-FCP-NEXT:    vmovaps %ymm3, 64(%rax)3002; AVX2-FCP-NEXT:    vmovaps %ymm0, 288(%rax)3003; AVX2-FCP-NEXT:    vmovaps %ymm15, 256(%rax)3004; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3005; AVX2-FCP-NEXT:    vmovaps %ymm0, 352(%rax)3006; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3007; AVX2-FCP-NEXT:    vmovaps %ymm0, 320(%rax)3008; AVX2-FCP-NEXT:    vmovups (%rsp), %ymm0 # 32-byte Reload3009; AVX2-FCP-NEXT:    vmovaps %ymm0, (%rax)3010; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3011; AVX2-FCP-NEXT:    vmovaps %ymm0, 32(%rax)3012; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3013; AVX2-FCP-NEXT:    vmovaps %ymm0, 160(%rax)3014; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3015; AVX2-FCP-NEXT:    vmovaps %ymm0, 128(%rax)3016; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3017; AVX2-FCP-NEXT:    vmovaps %ymm0, 224(%rax)3018; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3019; AVX2-FCP-NEXT:    vmovaps %ymm0, 192(%rax)3020; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3021; AVX2-FCP-NEXT:    vmovaps %ymm0, 416(%rax)3022; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3023; AVX2-FCP-NEXT:    vmovaps %ymm0, 384(%rax)3024; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3025; AVX2-FCP-NEXT:    vmovaps %ymm0, 480(%rax)3026; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3027; AVX2-FCP-NEXT:    vmovaps %ymm0, 448(%rax)3028; AVX2-FCP-NEXT:    addq $328, %rsp # imm = 0x1483029; AVX2-FCP-NEXT:    vzeroupper3030; AVX2-FCP-NEXT:    retq3031;3032; AVX512-LABEL: store_i32_stride8_vf16:3033; AVX512:       # %bb.0:3034; AVX512-NEXT:    movq {{[0-9]+}}(%rsp), %rax3035; AVX512-NEXT:    movq {{[0-9]+}}(%rsp), %r103036; AVX512-NEXT:    movq {{[0-9]+}}(%rsp), %r113037; AVX512-NEXT:    vmovdqa64 (%rdi), %zmm03038; AVX512-NEXT:    vmovdqa64 (%rsi), %zmm13039; AVX512-NEXT:    vmovdqa64 (%rdx), %zmm23040; AVX512-NEXT:    vmovdqa64 (%rcx), %zmm33041; AVX512-NEXT:    vmovdqa64 (%r8), %zmm53042; AVX512-NEXT:    vmovdqa64 (%r9), %zmm63043; AVX512-NEXT:    vmovdqa64 (%r11), %zmm73044; AVX512-NEXT:    vmovdqa64 (%r10), %zmm83045; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [0,0,0,0,0,0,0,16,0,0,0,0,0,0,1,17]3046; AVX512-NEXT:    vpermi2d %zmm8, %zmm7, %zmm43047; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm9 = [0,0,0,0,0,16,0,0,0,0,0,0,1,17,0,0]3048; AVX512-NEXT:    vpermi2d %zmm6, %zmm5, %zmm93049; AVX512-NEXT:    movb $-120, %cl3050; AVX512-NEXT:    kmovw %ecx, %k13051; AVX512-NEXT:    vmovdqa64 %zmm4, %zmm9 {%k1}3052; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm10 = [0,0,0,16,0,0,0,0,0,0,1,17,0,0,0,0]3053; AVX512-NEXT:    vpermi2d %zmm3, %zmm2, %zmm103054; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [0,16,0,0,0,0,0,0,1,17,0,0,0,0,0,0]3055; AVX512-NEXT:    vpermi2d %zmm1, %zmm0, %zmm43056; AVX512-NEXT:    movb $34, %cl3057; AVX512-NEXT:    kmovw %ecx, %k23058; AVX512-NEXT:    vmovdqa64 %zmm10, %zmm4 {%k2}3059; AVX512-NEXT:    movb $-52, %cl3060; AVX512-NEXT:    kmovw %ecx, %k33061; AVX512-NEXT:    vmovdqa64 %zmm9, %zmm4 {%k3}3062; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm9 = [0,0,0,0,0,0,2,18,0,0,0,0,0,0,3,19]3063; AVX512-NEXT:    vpermi2d %zmm8, %zmm7, %zmm93064; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm10 = [0,0,0,0,2,18,0,0,0,0,0,0,3,19,0,0]3065; AVX512-NEXT:    vpermi2d %zmm6, %zmm5, %zmm103066; AVX512-NEXT:    vmovdqa64 %zmm9, %zmm10 {%k1}3067; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm11 = [0,0,2,18,0,0,0,0,0,0,3,19,0,0,0,0]3068; AVX512-NEXT:    vpermi2d %zmm3, %zmm2, %zmm113069; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm9 = [2,18,0,0,0,0,0,0,3,19,0,0,0,0,0,0]3070; AVX512-NEXT:    vpermi2d %zmm1, %zmm0, %zmm93071; AVX512-NEXT:    vmovdqa64 %zmm11, %zmm9 {%k2}3072; AVX512-NEXT:    vmovdqa64 %zmm10, %zmm9 {%k3}3073; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm10 = [0,0,0,0,0,0,4,20,0,0,0,0,0,0,5,21]3074; AVX512-NEXT:    vpermi2d %zmm8, %zmm7, %zmm103075; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm11 = [0,0,0,0,4,20,0,0,0,0,0,0,5,21,0,0]3076; AVX512-NEXT:    vpermi2d %zmm6, %zmm5, %zmm113077; AVX512-NEXT:    vmovdqa64 %zmm10, %zmm11 {%k1}3078; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,4,20,0,0,0,0,0,0,5,21,0,0,0,0]3079; AVX512-NEXT:    vpermi2d %zmm3, %zmm2, %zmm123080; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm10 = [4,20,0,0,0,0,0,0,5,21,0,0,0,0,0,0]3081; AVX512-NEXT:    vpermi2d %zmm1, %zmm0, %zmm103082; AVX512-NEXT:    vmovdqa64 %zmm12, %zmm10 {%k2}3083; AVX512-NEXT:    vmovdqa64 %zmm11, %zmm10 {%k3}3084; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm11 = [0,0,0,0,0,0,6,22,0,0,0,0,0,0,7,23]3085; AVX512-NEXT:    vpermi2d %zmm8, %zmm7, %zmm113086; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,0,0,6,22,0,0,0,0,0,0,7,23,0,0]3087; AVX512-NEXT:    vpermi2d %zmm6, %zmm5, %zmm123088; AVX512-NEXT:    vmovdqa64 %zmm11, %zmm12 {%k1}3089; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm13 = [0,0,6,22,0,0,0,0,0,0,7,23,0,0,0,0]3090; AVX512-NEXT:    vpermi2d %zmm3, %zmm2, %zmm133091; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm11 = [6,22,0,0,0,0,0,0,7,23,0,0,0,0,0,0]3092; AVX512-NEXT:    vpermi2d %zmm1, %zmm0, %zmm113093; AVX512-NEXT:    vmovdqa64 %zmm13, %zmm11 {%k2}3094; AVX512-NEXT:    vmovdqa64 %zmm12, %zmm11 {%k3}3095; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,0,0,0,0,8,24,0,0,0,0,0,0,9,25]3096; AVX512-NEXT:    vpermi2d %zmm8, %zmm7, %zmm123097; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm13 = [0,0,0,0,8,24,0,0,0,0,0,0,9,25,0,0]3098; AVX512-NEXT:    vpermi2d %zmm6, %zmm5, %zmm133099; AVX512-NEXT:    vmovdqa64 %zmm12, %zmm13 {%k1}3100; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,8,24,0,0,0,0,0,0,9,25,0,0,0,0]3101; AVX512-NEXT:    vpermi2d %zmm3, %zmm2, %zmm123102; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm14 = [8,24,0,0,0,0,0,0,9,25,0,0,0,0,0,0]3103; AVX512-NEXT:    vpermi2d %zmm1, %zmm0, %zmm143104; AVX512-NEXT:    vmovdqa64 %zmm12, %zmm14 {%k2}3105; AVX512-NEXT:    vmovdqa64 %zmm13, %zmm14 {%k3}3106; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,0,0,0,0,10,26,0,0,0,0,0,0,11,27]3107; AVX512-NEXT:    vpermi2d %zmm8, %zmm7, %zmm123108; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm13 = [0,0,0,0,10,26,0,0,0,0,0,0,11,27,0,0]3109; AVX512-NEXT:    vpermi2d %zmm6, %zmm5, %zmm133110; AVX512-NEXT:    vmovdqa64 %zmm12, %zmm13 {%k1}3111; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,10,26,0,0,0,0,0,0,11,27,0,0,0,0]3112; AVX512-NEXT:    vpermi2d %zmm3, %zmm2, %zmm123113; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm15 = [10,26,0,0,0,0,0,0,11,27,0,0,0,0,0,0]3114; AVX512-NEXT:    vpermi2d %zmm1, %zmm0, %zmm153115; AVX512-NEXT:    vmovdqa64 %zmm12, %zmm15 {%k2}3116; AVX512-NEXT:    vmovdqa64 %zmm13, %zmm15 {%k3}3117; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,0,0,0,0,12,28,0,0,0,0,0,0,13,29]3118; AVX512-NEXT:    vpermi2d %zmm8, %zmm7, %zmm123119; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm13 = [0,0,0,0,12,28,0,0,0,0,0,0,13,29,0,0]3120; AVX512-NEXT:    vpermi2d %zmm6, %zmm5, %zmm133121; AVX512-NEXT:    vmovdqa64 %zmm12, %zmm13 {%k1}3122; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,12,28,0,0,0,0,0,0,13,29,0,0,0,0]3123; AVX512-NEXT:    vpermi2d %zmm3, %zmm2, %zmm123124; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm16 = [12,28,0,0,0,0,0,0,13,29,0,0,0,0,0,0]3125; AVX512-NEXT:    vpermi2d %zmm1, %zmm0, %zmm163126; AVX512-NEXT:    vmovdqa64 %zmm12, %zmm16 {%k2}3127; AVX512-NEXT:    vmovdqa64 %zmm13, %zmm16 {%k3}3128; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,0,0,0,0,14,30,0,0,0,0,0,0,15,31]3129; AVX512-NEXT:    vpermi2d %zmm8, %zmm7, %zmm123130; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm7 = [0,0,0,0,14,30,0,0,0,0,0,0,15,31,0,0]3131; AVX512-NEXT:    vpermi2d %zmm6, %zmm5, %zmm73132; AVX512-NEXT:    vmovdqa64 %zmm12, %zmm7 {%k1}3133; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm5 = [0,0,14,30,0,0,0,0,0,0,15,31,0,0,0,0]3134; AVX512-NEXT:    vpermi2d %zmm3, %zmm2, %zmm53135; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [14,30,0,0,0,0,0,0,15,31,0,0,0,0,0,0]3136; AVX512-NEXT:    vpermi2d %zmm1, %zmm0, %zmm23137; AVX512-NEXT:    vmovdqa64 %zmm5, %zmm2 {%k2}3138; AVX512-NEXT:    vmovdqa64 %zmm7, %zmm2 {%k3}3139; AVX512-NEXT:    vmovdqa64 %zmm2, 448(%rax)3140; AVX512-NEXT:    vmovdqa64 %zmm16, 384(%rax)3141; AVX512-NEXT:    vmovdqa64 %zmm15, 320(%rax)3142; AVX512-NEXT:    vmovdqa64 %zmm14, 256(%rax)3143; AVX512-NEXT:    vmovdqa64 %zmm11, 192(%rax)3144; AVX512-NEXT:    vmovdqa64 %zmm10, 128(%rax)3145; AVX512-NEXT:    vmovdqa64 %zmm9, 64(%rax)3146; AVX512-NEXT:    vmovdqa64 %zmm4, (%rax)3147; AVX512-NEXT:    vzeroupper3148; AVX512-NEXT:    retq3149;3150; AVX512-FCP-LABEL: store_i32_stride8_vf16:3151; AVX512-FCP:       # %bb.0:3152; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax3153; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r103154; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r113155; AVX512-FCP-NEXT:    vmovdqa64 (%rdi), %zmm03156; AVX512-FCP-NEXT:    vmovdqa64 (%rsi), %zmm13157; AVX512-FCP-NEXT:    vmovdqa64 (%rdx), %zmm23158; AVX512-FCP-NEXT:    vmovdqa64 (%rcx), %zmm33159; AVX512-FCP-NEXT:    vmovdqa64 (%r8), %zmm53160; AVX512-FCP-NEXT:    vmovdqa64 (%r9), %zmm63161; AVX512-FCP-NEXT:    vmovdqa64 (%r11), %zmm73162; AVX512-FCP-NEXT:    vmovdqa64 (%r10), %zmm83163; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [0,0,0,0,0,0,0,16,0,0,0,0,0,0,1,17]3164; AVX512-FCP-NEXT:    vpermi2d %zmm8, %zmm7, %zmm43165; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm9 = [0,0,0,0,0,16,0,0,0,0,0,0,1,17,0,0]3166; AVX512-FCP-NEXT:    vpermi2d %zmm6, %zmm5, %zmm93167; AVX512-FCP-NEXT:    movb $-120, %cl3168; AVX512-FCP-NEXT:    kmovw %ecx, %k13169; AVX512-FCP-NEXT:    vmovdqa64 %zmm4, %zmm9 {%k1}3170; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm10 = [0,0,0,16,0,0,0,0,0,0,1,17,0,0,0,0]3171; AVX512-FCP-NEXT:    vpermi2d %zmm3, %zmm2, %zmm103172; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [0,16,0,0,0,0,0,0,1,17,0,0,0,0,0,0]3173; AVX512-FCP-NEXT:    vpermi2d %zmm1, %zmm0, %zmm43174; AVX512-FCP-NEXT:    movb $34, %cl3175; AVX512-FCP-NEXT:    kmovw %ecx, %k23176; AVX512-FCP-NEXT:    vmovdqa64 %zmm10, %zmm4 {%k2}3177; AVX512-FCP-NEXT:    movb $-52, %cl3178; AVX512-FCP-NEXT:    kmovw %ecx, %k33179; AVX512-FCP-NEXT:    vmovdqa64 %zmm9, %zmm4 {%k3}3180; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm9 = [0,0,0,0,0,0,2,18,0,0,0,0,0,0,3,19]3181; AVX512-FCP-NEXT:    vpermi2d %zmm8, %zmm7, %zmm93182; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm10 = [0,0,0,0,2,18,0,0,0,0,0,0,3,19,0,0]3183; AVX512-FCP-NEXT:    vpermi2d %zmm6, %zmm5, %zmm103184; AVX512-FCP-NEXT:    vmovdqa64 %zmm9, %zmm10 {%k1}3185; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm11 = [0,0,2,18,0,0,0,0,0,0,3,19,0,0,0,0]3186; AVX512-FCP-NEXT:    vpermi2d %zmm3, %zmm2, %zmm113187; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm9 = [2,18,0,0,0,0,0,0,3,19,0,0,0,0,0,0]3188; AVX512-FCP-NEXT:    vpermi2d %zmm1, %zmm0, %zmm93189; AVX512-FCP-NEXT:    vmovdqa64 %zmm11, %zmm9 {%k2}3190; AVX512-FCP-NEXT:    vmovdqa64 %zmm10, %zmm9 {%k3}3191; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm10 = [0,0,0,0,0,0,4,20,0,0,0,0,0,0,5,21]3192; AVX512-FCP-NEXT:    vpermi2d %zmm8, %zmm7, %zmm103193; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm11 = [0,0,0,0,4,20,0,0,0,0,0,0,5,21,0,0]3194; AVX512-FCP-NEXT:    vpermi2d %zmm6, %zmm5, %zmm113195; AVX512-FCP-NEXT:    vmovdqa64 %zmm10, %zmm11 {%k1}3196; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,4,20,0,0,0,0,0,0,5,21,0,0,0,0]3197; AVX512-FCP-NEXT:    vpermi2d %zmm3, %zmm2, %zmm123198; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm10 = [4,20,0,0,0,0,0,0,5,21,0,0,0,0,0,0]3199; AVX512-FCP-NEXT:    vpermi2d %zmm1, %zmm0, %zmm103200; AVX512-FCP-NEXT:    vmovdqa64 %zmm12, %zmm10 {%k2}3201; AVX512-FCP-NEXT:    vmovdqa64 %zmm11, %zmm10 {%k3}3202; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm11 = [0,0,0,0,0,0,6,22,0,0,0,0,0,0,7,23]3203; AVX512-FCP-NEXT:    vpermi2d %zmm8, %zmm7, %zmm113204; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,0,0,6,22,0,0,0,0,0,0,7,23,0,0]3205; AVX512-FCP-NEXT:    vpermi2d %zmm6, %zmm5, %zmm123206; AVX512-FCP-NEXT:    vmovdqa64 %zmm11, %zmm12 {%k1}3207; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm13 = [0,0,6,22,0,0,0,0,0,0,7,23,0,0,0,0]3208; AVX512-FCP-NEXT:    vpermi2d %zmm3, %zmm2, %zmm133209; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm11 = [6,22,0,0,0,0,0,0,7,23,0,0,0,0,0,0]3210; AVX512-FCP-NEXT:    vpermi2d %zmm1, %zmm0, %zmm113211; AVX512-FCP-NEXT:    vmovdqa64 %zmm13, %zmm11 {%k2}3212; AVX512-FCP-NEXT:    vmovdqa64 %zmm12, %zmm11 {%k3}3213; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,0,0,0,0,8,24,0,0,0,0,0,0,9,25]3214; AVX512-FCP-NEXT:    vpermi2d %zmm8, %zmm7, %zmm123215; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm13 = [0,0,0,0,8,24,0,0,0,0,0,0,9,25,0,0]3216; AVX512-FCP-NEXT:    vpermi2d %zmm6, %zmm5, %zmm133217; AVX512-FCP-NEXT:    vmovdqa64 %zmm12, %zmm13 {%k1}3218; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,8,24,0,0,0,0,0,0,9,25,0,0,0,0]3219; AVX512-FCP-NEXT:    vpermi2d %zmm3, %zmm2, %zmm123220; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm14 = [8,24,0,0,0,0,0,0,9,25,0,0,0,0,0,0]3221; AVX512-FCP-NEXT:    vpermi2d %zmm1, %zmm0, %zmm143222; AVX512-FCP-NEXT:    vmovdqa64 %zmm12, %zmm14 {%k2}3223; AVX512-FCP-NEXT:    vmovdqa64 %zmm13, %zmm14 {%k3}3224; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,0,0,0,0,10,26,0,0,0,0,0,0,11,27]3225; AVX512-FCP-NEXT:    vpermi2d %zmm8, %zmm7, %zmm123226; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm13 = [0,0,0,0,10,26,0,0,0,0,0,0,11,27,0,0]3227; AVX512-FCP-NEXT:    vpermi2d %zmm6, %zmm5, %zmm133228; AVX512-FCP-NEXT:    vmovdqa64 %zmm12, %zmm13 {%k1}3229; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,10,26,0,0,0,0,0,0,11,27,0,0,0,0]3230; AVX512-FCP-NEXT:    vpermi2d %zmm3, %zmm2, %zmm123231; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm15 = [10,26,0,0,0,0,0,0,11,27,0,0,0,0,0,0]3232; AVX512-FCP-NEXT:    vpermi2d %zmm1, %zmm0, %zmm153233; AVX512-FCP-NEXT:    vmovdqa64 %zmm12, %zmm15 {%k2}3234; AVX512-FCP-NEXT:    vmovdqa64 %zmm13, %zmm15 {%k3}3235; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,0,0,0,0,12,28,0,0,0,0,0,0,13,29]3236; AVX512-FCP-NEXT:    vpermi2d %zmm8, %zmm7, %zmm123237; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm13 = [0,0,0,0,12,28,0,0,0,0,0,0,13,29,0,0]3238; AVX512-FCP-NEXT:    vpermi2d %zmm6, %zmm5, %zmm133239; AVX512-FCP-NEXT:    vmovdqa64 %zmm12, %zmm13 {%k1}3240; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,12,28,0,0,0,0,0,0,13,29,0,0,0,0]3241; AVX512-FCP-NEXT:    vpermi2d %zmm3, %zmm2, %zmm123242; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm16 = [12,28,0,0,0,0,0,0,13,29,0,0,0,0,0,0]3243; AVX512-FCP-NEXT:    vpermi2d %zmm1, %zmm0, %zmm163244; AVX512-FCP-NEXT:    vmovdqa64 %zmm12, %zmm16 {%k2}3245; AVX512-FCP-NEXT:    vmovdqa64 %zmm13, %zmm16 {%k3}3246; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,0,0,0,0,14,30,0,0,0,0,0,0,15,31]3247; AVX512-FCP-NEXT:    vpermi2d %zmm8, %zmm7, %zmm123248; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm7 = [0,0,0,0,14,30,0,0,0,0,0,0,15,31,0,0]3249; AVX512-FCP-NEXT:    vpermi2d %zmm6, %zmm5, %zmm73250; AVX512-FCP-NEXT:    vmovdqa64 %zmm12, %zmm7 {%k1}3251; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm5 = [0,0,14,30,0,0,0,0,0,0,15,31,0,0,0,0]3252; AVX512-FCP-NEXT:    vpermi2d %zmm3, %zmm2, %zmm53253; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [14,30,0,0,0,0,0,0,15,31,0,0,0,0,0,0]3254; AVX512-FCP-NEXT:    vpermi2d %zmm1, %zmm0, %zmm23255; AVX512-FCP-NEXT:    vmovdqa64 %zmm5, %zmm2 {%k2}3256; AVX512-FCP-NEXT:    vmovdqa64 %zmm7, %zmm2 {%k3}3257; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, 448(%rax)3258; AVX512-FCP-NEXT:    vmovdqa64 %zmm16, 384(%rax)3259; AVX512-FCP-NEXT:    vmovdqa64 %zmm15, 320(%rax)3260; AVX512-FCP-NEXT:    vmovdqa64 %zmm14, 256(%rax)3261; AVX512-FCP-NEXT:    vmovdqa64 %zmm11, 192(%rax)3262; AVX512-FCP-NEXT:    vmovdqa64 %zmm10, 128(%rax)3263; AVX512-FCP-NEXT:    vmovdqa64 %zmm9, 64(%rax)3264; AVX512-FCP-NEXT:    vmovdqa64 %zmm4, (%rax)3265; AVX512-FCP-NEXT:    vzeroupper3266; AVX512-FCP-NEXT:    retq3267;3268; AVX512DQ-LABEL: store_i32_stride8_vf16:3269; AVX512DQ:       # %bb.0:3270; AVX512DQ-NEXT:    movq {{[0-9]+}}(%rsp), %rax3271; AVX512DQ-NEXT:    movq {{[0-9]+}}(%rsp), %r103272; AVX512DQ-NEXT:    movq {{[0-9]+}}(%rsp), %r113273; AVX512DQ-NEXT:    vmovdqa64 (%rdi), %zmm03274; AVX512DQ-NEXT:    vmovdqa64 (%rsi), %zmm13275; AVX512DQ-NEXT:    vmovdqa64 (%rdx), %zmm23276; AVX512DQ-NEXT:    vmovdqa64 (%rcx), %zmm33277; AVX512DQ-NEXT:    vmovdqa64 (%r8), %zmm53278; AVX512DQ-NEXT:    vmovdqa64 (%r9), %zmm63279; AVX512DQ-NEXT:    vmovdqa64 (%r11), %zmm73280; AVX512DQ-NEXT:    vmovdqa64 (%r10), %zmm83281; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [0,0,0,0,0,0,0,16,0,0,0,0,0,0,1,17]3282; AVX512DQ-NEXT:    vpermi2d %zmm8, %zmm7, %zmm43283; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm9 = [0,0,0,0,0,16,0,0,0,0,0,0,1,17,0,0]3284; AVX512DQ-NEXT:    vpermi2d %zmm6, %zmm5, %zmm93285; AVX512DQ-NEXT:    movb $-120, %cl3286; AVX512DQ-NEXT:    kmovw %ecx, %k13287; AVX512DQ-NEXT:    vmovdqa64 %zmm4, %zmm9 {%k1}3288; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm10 = [0,0,0,16,0,0,0,0,0,0,1,17,0,0,0,0]3289; AVX512DQ-NEXT:    vpermi2d %zmm3, %zmm2, %zmm103290; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [0,16,0,0,0,0,0,0,1,17,0,0,0,0,0,0]3291; AVX512DQ-NEXT:    vpermi2d %zmm1, %zmm0, %zmm43292; AVX512DQ-NEXT:    movb $34, %cl3293; AVX512DQ-NEXT:    kmovw %ecx, %k23294; AVX512DQ-NEXT:    vmovdqa64 %zmm10, %zmm4 {%k2}3295; AVX512DQ-NEXT:    movb $-52, %cl3296; AVX512DQ-NEXT:    kmovw %ecx, %k33297; AVX512DQ-NEXT:    vmovdqa64 %zmm9, %zmm4 {%k3}3298; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm9 = [0,0,0,0,0,0,2,18,0,0,0,0,0,0,3,19]3299; AVX512DQ-NEXT:    vpermi2d %zmm8, %zmm7, %zmm93300; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm10 = [0,0,0,0,2,18,0,0,0,0,0,0,3,19,0,0]3301; AVX512DQ-NEXT:    vpermi2d %zmm6, %zmm5, %zmm103302; AVX512DQ-NEXT:    vmovdqa64 %zmm9, %zmm10 {%k1}3303; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm11 = [0,0,2,18,0,0,0,0,0,0,3,19,0,0,0,0]3304; AVX512DQ-NEXT:    vpermi2d %zmm3, %zmm2, %zmm113305; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm9 = [2,18,0,0,0,0,0,0,3,19,0,0,0,0,0,0]3306; AVX512DQ-NEXT:    vpermi2d %zmm1, %zmm0, %zmm93307; AVX512DQ-NEXT:    vmovdqa64 %zmm11, %zmm9 {%k2}3308; AVX512DQ-NEXT:    vmovdqa64 %zmm10, %zmm9 {%k3}3309; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm10 = [0,0,0,0,0,0,4,20,0,0,0,0,0,0,5,21]3310; AVX512DQ-NEXT:    vpermi2d %zmm8, %zmm7, %zmm103311; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm11 = [0,0,0,0,4,20,0,0,0,0,0,0,5,21,0,0]3312; AVX512DQ-NEXT:    vpermi2d %zmm6, %zmm5, %zmm113313; AVX512DQ-NEXT:    vmovdqa64 %zmm10, %zmm11 {%k1}3314; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,4,20,0,0,0,0,0,0,5,21,0,0,0,0]3315; AVX512DQ-NEXT:    vpermi2d %zmm3, %zmm2, %zmm123316; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm10 = [4,20,0,0,0,0,0,0,5,21,0,0,0,0,0,0]3317; AVX512DQ-NEXT:    vpermi2d %zmm1, %zmm0, %zmm103318; AVX512DQ-NEXT:    vmovdqa64 %zmm12, %zmm10 {%k2}3319; AVX512DQ-NEXT:    vmovdqa64 %zmm11, %zmm10 {%k3}3320; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm11 = [0,0,0,0,0,0,6,22,0,0,0,0,0,0,7,23]3321; AVX512DQ-NEXT:    vpermi2d %zmm8, %zmm7, %zmm113322; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,0,0,6,22,0,0,0,0,0,0,7,23,0,0]3323; AVX512DQ-NEXT:    vpermi2d %zmm6, %zmm5, %zmm123324; AVX512DQ-NEXT:    vmovdqa64 %zmm11, %zmm12 {%k1}3325; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm13 = [0,0,6,22,0,0,0,0,0,0,7,23,0,0,0,0]3326; AVX512DQ-NEXT:    vpermi2d %zmm3, %zmm2, %zmm133327; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm11 = [6,22,0,0,0,0,0,0,7,23,0,0,0,0,0,0]3328; AVX512DQ-NEXT:    vpermi2d %zmm1, %zmm0, %zmm113329; AVX512DQ-NEXT:    vmovdqa64 %zmm13, %zmm11 {%k2}3330; AVX512DQ-NEXT:    vmovdqa64 %zmm12, %zmm11 {%k3}3331; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,0,0,0,0,8,24,0,0,0,0,0,0,9,25]3332; AVX512DQ-NEXT:    vpermi2d %zmm8, %zmm7, %zmm123333; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm13 = [0,0,0,0,8,24,0,0,0,0,0,0,9,25,0,0]3334; AVX512DQ-NEXT:    vpermi2d %zmm6, %zmm5, %zmm133335; AVX512DQ-NEXT:    vmovdqa64 %zmm12, %zmm13 {%k1}3336; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,8,24,0,0,0,0,0,0,9,25,0,0,0,0]3337; AVX512DQ-NEXT:    vpermi2d %zmm3, %zmm2, %zmm123338; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm14 = [8,24,0,0,0,0,0,0,9,25,0,0,0,0,0,0]3339; AVX512DQ-NEXT:    vpermi2d %zmm1, %zmm0, %zmm143340; AVX512DQ-NEXT:    vmovdqa64 %zmm12, %zmm14 {%k2}3341; AVX512DQ-NEXT:    vmovdqa64 %zmm13, %zmm14 {%k3}3342; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,0,0,0,0,10,26,0,0,0,0,0,0,11,27]3343; AVX512DQ-NEXT:    vpermi2d %zmm8, %zmm7, %zmm123344; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm13 = [0,0,0,0,10,26,0,0,0,0,0,0,11,27,0,0]3345; AVX512DQ-NEXT:    vpermi2d %zmm6, %zmm5, %zmm133346; AVX512DQ-NEXT:    vmovdqa64 %zmm12, %zmm13 {%k1}3347; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,10,26,0,0,0,0,0,0,11,27,0,0,0,0]3348; AVX512DQ-NEXT:    vpermi2d %zmm3, %zmm2, %zmm123349; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm15 = [10,26,0,0,0,0,0,0,11,27,0,0,0,0,0,0]3350; AVX512DQ-NEXT:    vpermi2d %zmm1, %zmm0, %zmm153351; AVX512DQ-NEXT:    vmovdqa64 %zmm12, %zmm15 {%k2}3352; AVX512DQ-NEXT:    vmovdqa64 %zmm13, %zmm15 {%k3}3353; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,0,0,0,0,12,28,0,0,0,0,0,0,13,29]3354; AVX512DQ-NEXT:    vpermi2d %zmm8, %zmm7, %zmm123355; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm13 = [0,0,0,0,12,28,0,0,0,0,0,0,13,29,0,0]3356; AVX512DQ-NEXT:    vpermi2d %zmm6, %zmm5, %zmm133357; AVX512DQ-NEXT:    vmovdqa64 %zmm12, %zmm13 {%k1}3358; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,12,28,0,0,0,0,0,0,13,29,0,0,0,0]3359; AVX512DQ-NEXT:    vpermi2d %zmm3, %zmm2, %zmm123360; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm16 = [12,28,0,0,0,0,0,0,13,29,0,0,0,0,0,0]3361; AVX512DQ-NEXT:    vpermi2d %zmm1, %zmm0, %zmm163362; AVX512DQ-NEXT:    vmovdqa64 %zmm12, %zmm16 {%k2}3363; AVX512DQ-NEXT:    vmovdqa64 %zmm13, %zmm16 {%k3}3364; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,0,0,0,0,14,30,0,0,0,0,0,0,15,31]3365; AVX512DQ-NEXT:    vpermi2d %zmm8, %zmm7, %zmm123366; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm7 = [0,0,0,0,14,30,0,0,0,0,0,0,15,31,0,0]3367; AVX512DQ-NEXT:    vpermi2d %zmm6, %zmm5, %zmm73368; AVX512DQ-NEXT:    vmovdqa64 %zmm12, %zmm7 {%k1}3369; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm5 = [0,0,14,30,0,0,0,0,0,0,15,31,0,0,0,0]3370; AVX512DQ-NEXT:    vpermi2d %zmm3, %zmm2, %zmm53371; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [14,30,0,0,0,0,0,0,15,31,0,0,0,0,0,0]3372; AVX512DQ-NEXT:    vpermi2d %zmm1, %zmm0, %zmm23373; AVX512DQ-NEXT:    vmovdqa64 %zmm5, %zmm2 {%k2}3374; AVX512DQ-NEXT:    vmovdqa64 %zmm7, %zmm2 {%k3}3375; AVX512DQ-NEXT:    vmovdqa64 %zmm2, 448(%rax)3376; AVX512DQ-NEXT:    vmovdqa64 %zmm16, 384(%rax)3377; AVX512DQ-NEXT:    vmovdqa64 %zmm15, 320(%rax)3378; AVX512DQ-NEXT:    vmovdqa64 %zmm14, 256(%rax)3379; AVX512DQ-NEXT:    vmovdqa64 %zmm11, 192(%rax)3380; AVX512DQ-NEXT:    vmovdqa64 %zmm10, 128(%rax)3381; AVX512DQ-NEXT:    vmovdqa64 %zmm9, 64(%rax)3382; AVX512DQ-NEXT:    vmovdqa64 %zmm4, (%rax)3383; AVX512DQ-NEXT:    vzeroupper3384; AVX512DQ-NEXT:    retq3385;3386; AVX512DQ-FCP-LABEL: store_i32_stride8_vf16:3387; AVX512DQ-FCP:       # %bb.0:3388; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax3389; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r103390; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r113391; AVX512DQ-FCP-NEXT:    vmovdqa64 (%rdi), %zmm03392; AVX512DQ-FCP-NEXT:    vmovdqa64 (%rsi), %zmm13393; AVX512DQ-FCP-NEXT:    vmovdqa64 (%rdx), %zmm23394; AVX512DQ-FCP-NEXT:    vmovdqa64 (%rcx), %zmm33395; AVX512DQ-FCP-NEXT:    vmovdqa64 (%r8), %zmm53396; AVX512DQ-FCP-NEXT:    vmovdqa64 (%r9), %zmm63397; AVX512DQ-FCP-NEXT:    vmovdqa64 (%r11), %zmm73398; AVX512DQ-FCP-NEXT:    vmovdqa64 (%r10), %zmm83399; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [0,0,0,0,0,0,0,16,0,0,0,0,0,0,1,17]3400; AVX512DQ-FCP-NEXT:    vpermi2d %zmm8, %zmm7, %zmm43401; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm9 = [0,0,0,0,0,16,0,0,0,0,0,0,1,17,0,0]3402; AVX512DQ-FCP-NEXT:    vpermi2d %zmm6, %zmm5, %zmm93403; AVX512DQ-FCP-NEXT:    movb $-120, %cl3404; AVX512DQ-FCP-NEXT:    kmovw %ecx, %k13405; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm4, %zmm9 {%k1}3406; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm10 = [0,0,0,16,0,0,0,0,0,0,1,17,0,0,0,0]3407; AVX512DQ-FCP-NEXT:    vpermi2d %zmm3, %zmm2, %zmm103408; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [0,16,0,0,0,0,0,0,1,17,0,0,0,0,0,0]3409; AVX512DQ-FCP-NEXT:    vpermi2d %zmm1, %zmm0, %zmm43410; AVX512DQ-FCP-NEXT:    movb $34, %cl3411; AVX512DQ-FCP-NEXT:    kmovw %ecx, %k23412; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm10, %zmm4 {%k2}3413; AVX512DQ-FCP-NEXT:    movb $-52, %cl3414; AVX512DQ-FCP-NEXT:    kmovw %ecx, %k33415; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm9, %zmm4 {%k3}3416; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm9 = [0,0,0,0,0,0,2,18,0,0,0,0,0,0,3,19]3417; AVX512DQ-FCP-NEXT:    vpermi2d %zmm8, %zmm7, %zmm93418; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm10 = [0,0,0,0,2,18,0,0,0,0,0,0,3,19,0,0]3419; AVX512DQ-FCP-NEXT:    vpermi2d %zmm6, %zmm5, %zmm103420; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm9, %zmm10 {%k1}3421; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm11 = [0,0,2,18,0,0,0,0,0,0,3,19,0,0,0,0]3422; AVX512DQ-FCP-NEXT:    vpermi2d %zmm3, %zmm2, %zmm113423; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm9 = [2,18,0,0,0,0,0,0,3,19,0,0,0,0,0,0]3424; AVX512DQ-FCP-NEXT:    vpermi2d %zmm1, %zmm0, %zmm93425; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm11, %zmm9 {%k2}3426; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm10, %zmm9 {%k3}3427; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm10 = [0,0,0,0,0,0,4,20,0,0,0,0,0,0,5,21]3428; AVX512DQ-FCP-NEXT:    vpermi2d %zmm8, %zmm7, %zmm103429; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm11 = [0,0,0,0,4,20,0,0,0,0,0,0,5,21,0,0]3430; AVX512DQ-FCP-NEXT:    vpermi2d %zmm6, %zmm5, %zmm113431; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm10, %zmm11 {%k1}3432; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,4,20,0,0,0,0,0,0,5,21,0,0,0,0]3433; AVX512DQ-FCP-NEXT:    vpermi2d %zmm3, %zmm2, %zmm123434; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm10 = [4,20,0,0,0,0,0,0,5,21,0,0,0,0,0,0]3435; AVX512DQ-FCP-NEXT:    vpermi2d %zmm1, %zmm0, %zmm103436; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm12, %zmm10 {%k2}3437; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm11, %zmm10 {%k3}3438; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm11 = [0,0,0,0,0,0,6,22,0,0,0,0,0,0,7,23]3439; AVX512DQ-FCP-NEXT:    vpermi2d %zmm8, %zmm7, %zmm113440; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,0,0,6,22,0,0,0,0,0,0,7,23,0,0]3441; AVX512DQ-FCP-NEXT:    vpermi2d %zmm6, %zmm5, %zmm123442; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm11, %zmm12 {%k1}3443; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm13 = [0,0,6,22,0,0,0,0,0,0,7,23,0,0,0,0]3444; AVX512DQ-FCP-NEXT:    vpermi2d %zmm3, %zmm2, %zmm133445; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm11 = [6,22,0,0,0,0,0,0,7,23,0,0,0,0,0,0]3446; AVX512DQ-FCP-NEXT:    vpermi2d %zmm1, %zmm0, %zmm113447; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm13, %zmm11 {%k2}3448; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm12, %zmm11 {%k3}3449; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,0,0,0,0,8,24,0,0,0,0,0,0,9,25]3450; AVX512DQ-FCP-NEXT:    vpermi2d %zmm8, %zmm7, %zmm123451; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm13 = [0,0,0,0,8,24,0,0,0,0,0,0,9,25,0,0]3452; AVX512DQ-FCP-NEXT:    vpermi2d %zmm6, %zmm5, %zmm133453; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm12, %zmm13 {%k1}3454; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,8,24,0,0,0,0,0,0,9,25,0,0,0,0]3455; AVX512DQ-FCP-NEXT:    vpermi2d %zmm3, %zmm2, %zmm123456; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm14 = [8,24,0,0,0,0,0,0,9,25,0,0,0,0,0,0]3457; AVX512DQ-FCP-NEXT:    vpermi2d %zmm1, %zmm0, %zmm143458; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm12, %zmm14 {%k2}3459; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm13, %zmm14 {%k3}3460; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,0,0,0,0,10,26,0,0,0,0,0,0,11,27]3461; AVX512DQ-FCP-NEXT:    vpermi2d %zmm8, %zmm7, %zmm123462; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm13 = [0,0,0,0,10,26,0,0,0,0,0,0,11,27,0,0]3463; AVX512DQ-FCP-NEXT:    vpermi2d %zmm6, %zmm5, %zmm133464; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm12, %zmm13 {%k1}3465; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,10,26,0,0,0,0,0,0,11,27,0,0,0,0]3466; AVX512DQ-FCP-NEXT:    vpermi2d %zmm3, %zmm2, %zmm123467; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm15 = [10,26,0,0,0,0,0,0,11,27,0,0,0,0,0,0]3468; AVX512DQ-FCP-NEXT:    vpermi2d %zmm1, %zmm0, %zmm153469; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm12, %zmm15 {%k2}3470; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm13, %zmm15 {%k3}3471; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,0,0,0,0,12,28,0,0,0,0,0,0,13,29]3472; AVX512DQ-FCP-NEXT:    vpermi2d %zmm8, %zmm7, %zmm123473; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm13 = [0,0,0,0,12,28,0,0,0,0,0,0,13,29,0,0]3474; AVX512DQ-FCP-NEXT:    vpermi2d %zmm6, %zmm5, %zmm133475; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm12, %zmm13 {%k1}3476; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,12,28,0,0,0,0,0,0,13,29,0,0,0,0]3477; AVX512DQ-FCP-NEXT:    vpermi2d %zmm3, %zmm2, %zmm123478; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm16 = [12,28,0,0,0,0,0,0,13,29,0,0,0,0,0,0]3479; AVX512DQ-FCP-NEXT:    vpermi2d %zmm1, %zmm0, %zmm163480; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm12, %zmm16 {%k2}3481; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm13, %zmm16 {%k3}3482; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,0,0,0,0,14,30,0,0,0,0,0,0,15,31]3483; AVX512DQ-FCP-NEXT:    vpermi2d %zmm8, %zmm7, %zmm123484; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm7 = [0,0,0,0,14,30,0,0,0,0,0,0,15,31,0,0]3485; AVX512DQ-FCP-NEXT:    vpermi2d %zmm6, %zmm5, %zmm73486; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm12, %zmm7 {%k1}3487; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm5 = [0,0,14,30,0,0,0,0,0,0,15,31,0,0,0,0]3488; AVX512DQ-FCP-NEXT:    vpermi2d %zmm3, %zmm2, %zmm53489; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [14,30,0,0,0,0,0,0,15,31,0,0,0,0,0,0]3490; AVX512DQ-FCP-NEXT:    vpermi2d %zmm1, %zmm0, %zmm23491; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm5, %zmm2 {%k2}3492; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm7, %zmm2 {%k3}3493; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm2, 448(%rax)3494; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm16, 384(%rax)3495; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm15, 320(%rax)3496; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm14, 256(%rax)3497; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm11, 192(%rax)3498; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm10, 128(%rax)3499; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm9, 64(%rax)3500; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm4, (%rax)3501; AVX512DQ-FCP-NEXT:    vzeroupper3502; AVX512DQ-FCP-NEXT:    retq3503;3504; AVX512BW-LABEL: store_i32_stride8_vf16:3505; AVX512BW:       # %bb.0:3506; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax3507; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %r103508; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %r113509; AVX512BW-NEXT:    vmovdqa64 (%rdi), %zmm03510; AVX512BW-NEXT:    vmovdqa64 (%rsi), %zmm13511; AVX512BW-NEXT:    vmovdqa64 (%rdx), %zmm23512; AVX512BW-NEXT:    vmovdqa64 (%rcx), %zmm33513; AVX512BW-NEXT:    vmovdqa64 (%r8), %zmm53514; AVX512BW-NEXT:    vmovdqa64 (%r9), %zmm63515; AVX512BW-NEXT:    vmovdqa64 (%r11), %zmm73516; AVX512BW-NEXT:    vmovdqa64 (%r10), %zmm83517; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [0,0,0,0,0,0,0,16,0,0,0,0,0,0,1,17]3518; AVX512BW-NEXT:    vpermi2d %zmm8, %zmm7, %zmm43519; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm9 = [0,0,0,0,0,16,0,0,0,0,0,0,1,17,0,0]3520; AVX512BW-NEXT:    vpermi2d %zmm6, %zmm5, %zmm93521; AVX512BW-NEXT:    movb $-120, %cl3522; AVX512BW-NEXT:    kmovd %ecx, %k13523; AVX512BW-NEXT:    vmovdqa64 %zmm4, %zmm9 {%k1}3524; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm10 = [0,0,0,16,0,0,0,0,0,0,1,17,0,0,0,0]3525; AVX512BW-NEXT:    vpermi2d %zmm3, %zmm2, %zmm103526; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [0,16,0,0,0,0,0,0,1,17,0,0,0,0,0,0]3527; AVX512BW-NEXT:    vpermi2d %zmm1, %zmm0, %zmm43528; AVX512BW-NEXT:    movb $34, %cl3529; AVX512BW-NEXT:    kmovd %ecx, %k23530; AVX512BW-NEXT:    vmovdqa64 %zmm10, %zmm4 {%k2}3531; AVX512BW-NEXT:    movb $-52, %cl3532; AVX512BW-NEXT:    kmovd %ecx, %k33533; AVX512BW-NEXT:    vmovdqa64 %zmm9, %zmm4 {%k3}3534; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm9 = [0,0,0,0,0,0,2,18,0,0,0,0,0,0,3,19]3535; AVX512BW-NEXT:    vpermi2d %zmm8, %zmm7, %zmm93536; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm10 = [0,0,0,0,2,18,0,0,0,0,0,0,3,19,0,0]3537; AVX512BW-NEXT:    vpermi2d %zmm6, %zmm5, %zmm103538; AVX512BW-NEXT:    vmovdqa64 %zmm9, %zmm10 {%k1}3539; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm11 = [0,0,2,18,0,0,0,0,0,0,3,19,0,0,0,0]3540; AVX512BW-NEXT:    vpermi2d %zmm3, %zmm2, %zmm113541; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm9 = [2,18,0,0,0,0,0,0,3,19,0,0,0,0,0,0]3542; AVX512BW-NEXT:    vpermi2d %zmm1, %zmm0, %zmm93543; AVX512BW-NEXT:    vmovdqa64 %zmm11, %zmm9 {%k2}3544; AVX512BW-NEXT:    vmovdqa64 %zmm10, %zmm9 {%k3}3545; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm10 = [0,0,0,0,0,0,4,20,0,0,0,0,0,0,5,21]3546; AVX512BW-NEXT:    vpermi2d %zmm8, %zmm7, %zmm103547; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm11 = [0,0,0,0,4,20,0,0,0,0,0,0,5,21,0,0]3548; AVX512BW-NEXT:    vpermi2d %zmm6, %zmm5, %zmm113549; AVX512BW-NEXT:    vmovdqa64 %zmm10, %zmm11 {%k1}3550; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,4,20,0,0,0,0,0,0,5,21,0,0,0,0]3551; AVX512BW-NEXT:    vpermi2d %zmm3, %zmm2, %zmm123552; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm10 = [4,20,0,0,0,0,0,0,5,21,0,0,0,0,0,0]3553; AVX512BW-NEXT:    vpermi2d %zmm1, %zmm0, %zmm103554; AVX512BW-NEXT:    vmovdqa64 %zmm12, %zmm10 {%k2}3555; AVX512BW-NEXT:    vmovdqa64 %zmm11, %zmm10 {%k3}3556; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm11 = [0,0,0,0,0,0,6,22,0,0,0,0,0,0,7,23]3557; AVX512BW-NEXT:    vpermi2d %zmm8, %zmm7, %zmm113558; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,0,0,6,22,0,0,0,0,0,0,7,23,0,0]3559; AVX512BW-NEXT:    vpermi2d %zmm6, %zmm5, %zmm123560; AVX512BW-NEXT:    vmovdqa64 %zmm11, %zmm12 {%k1}3561; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm13 = [0,0,6,22,0,0,0,0,0,0,7,23,0,0,0,0]3562; AVX512BW-NEXT:    vpermi2d %zmm3, %zmm2, %zmm133563; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm11 = [6,22,0,0,0,0,0,0,7,23,0,0,0,0,0,0]3564; AVX512BW-NEXT:    vpermi2d %zmm1, %zmm0, %zmm113565; AVX512BW-NEXT:    vmovdqa64 %zmm13, %zmm11 {%k2}3566; AVX512BW-NEXT:    vmovdqa64 %zmm12, %zmm11 {%k3}3567; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,0,0,0,0,8,24,0,0,0,0,0,0,9,25]3568; AVX512BW-NEXT:    vpermi2d %zmm8, %zmm7, %zmm123569; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm13 = [0,0,0,0,8,24,0,0,0,0,0,0,9,25,0,0]3570; AVX512BW-NEXT:    vpermi2d %zmm6, %zmm5, %zmm133571; AVX512BW-NEXT:    vmovdqa64 %zmm12, %zmm13 {%k1}3572; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,8,24,0,0,0,0,0,0,9,25,0,0,0,0]3573; AVX512BW-NEXT:    vpermi2d %zmm3, %zmm2, %zmm123574; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm14 = [8,24,0,0,0,0,0,0,9,25,0,0,0,0,0,0]3575; AVX512BW-NEXT:    vpermi2d %zmm1, %zmm0, %zmm143576; AVX512BW-NEXT:    vmovdqa64 %zmm12, %zmm14 {%k2}3577; AVX512BW-NEXT:    vmovdqa64 %zmm13, %zmm14 {%k3}3578; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,0,0,0,0,10,26,0,0,0,0,0,0,11,27]3579; AVX512BW-NEXT:    vpermi2d %zmm8, %zmm7, %zmm123580; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm13 = [0,0,0,0,10,26,0,0,0,0,0,0,11,27,0,0]3581; AVX512BW-NEXT:    vpermi2d %zmm6, %zmm5, %zmm133582; AVX512BW-NEXT:    vmovdqa64 %zmm12, %zmm13 {%k1}3583; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,10,26,0,0,0,0,0,0,11,27,0,0,0,0]3584; AVX512BW-NEXT:    vpermi2d %zmm3, %zmm2, %zmm123585; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm15 = [10,26,0,0,0,0,0,0,11,27,0,0,0,0,0,0]3586; AVX512BW-NEXT:    vpermi2d %zmm1, %zmm0, %zmm153587; AVX512BW-NEXT:    vmovdqa64 %zmm12, %zmm15 {%k2}3588; AVX512BW-NEXT:    vmovdqa64 %zmm13, %zmm15 {%k3}3589; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,0,0,0,0,12,28,0,0,0,0,0,0,13,29]3590; AVX512BW-NEXT:    vpermi2d %zmm8, %zmm7, %zmm123591; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm13 = [0,0,0,0,12,28,0,0,0,0,0,0,13,29,0,0]3592; AVX512BW-NEXT:    vpermi2d %zmm6, %zmm5, %zmm133593; AVX512BW-NEXT:    vmovdqa64 %zmm12, %zmm13 {%k1}3594; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,12,28,0,0,0,0,0,0,13,29,0,0,0,0]3595; AVX512BW-NEXT:    vpermi2d %zmm3, %zmm2, %zmm123596; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm16 = [12,28,0,0,0,0,0,0,13,29,0,0,0,0,0,0]3597; AVX512BW-NEXT:    vpermi2d %zmm1, %zmm0, %zmm163598; AVX512BW-NEXT:    vmovdqa64 %zmm12, %zmm16 {%k2}3599; AVX512BW-NEXT:    vmovdqa64 %zmm13, %zmm16 {%k3}3600; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,0,0,0,0,14,30,0,0,0,0,0,0,15,31]3601; AVX512BW-NEXT:    vpermi2d %zmm8, %zmm7, %zmm123602; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm7 = [0,0,0,0,14,30,0,0,0,0,0,0,15,31,0,0]3603; AVX512BW-NEXT:    vpermi2d %zmm6, %zmm5, %zmm73604; AVX512BW-NEXT:    vmovdqa64 %zmm12, %zmm7 {%k1}3605; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm5 = [0,0,14,30,0,0,0,0,0,0,15,31,0,0,0,0]3606; AVX512BW-NEXT:    vpermi2d %zmm3, %zmm2, %zmm53607; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [14,30,0,0,0,0,0,0,15,31,0,0,0,0,0,0]3608; AVX512BW-NEXT:    vpermi2d %zmm1, %zmm0, %zmm23609; AVX512BW-NEXT:    vmovdqa64 %zmm5, %zmm2 {%k2}3610; AVX512BW-NEXT:    vmovdqa64 %zmm7, %zmm2 {%k3}3611; AVX512BW-NEXT:    vmovdqa64 %zmm2, 448(%rax)3612; AVX512BW-NEXT:    vmovdqa64 %zmm16, 384(%rax)3613; AVX512BW-NEXT:    vmovdqa64 %zmm15, 320(%rax)3614; AVX512BW-NEXT:    vmovdqa64 %zmm14, 256(%rax)3615; AVX512BW-NEXT:    vmovdqa64 %zmm11, 192(%rax)3616; AVX512BW-NEXT:    vmovdqa64 %zmm10, 128(%rax)3617; AVX512BW-NEXT:    vmovdqa64 %zmm9, 64(%rax)3618; AVX512BW-NEXT:    vmovdqa64 %zmm4, (%rax)3619; AVX512BW-NEXT:    vzeroupper3620; AVX512BW-NEXT:    retq3621;3622; AVX512BW-FCP-LABEL: store_i32_stride8_vf16:3623; AVX512BW-FCP:       # %bb.0:3624; AVX512BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax3625; AVX512BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r103626; AVX512BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r113627; AVX512BW-FCP-NEXT:    vmovdqa64 (%rdi), %zmm03628; AVX512BW-FCP-NEXT:    vmovdqa64 (%rsi), %zmm13629; AVX512BW-FCP-NEXT:    vmovdqa64 (%rdx), %zmm23630; AVX512BW-FCP-NEXT:    vmovdqa64 (%rcx), %zmm33631; AVX512BW-FCP-NEXT:    vmovdqa64 (%r8), %zmm53632; AVX512BW-FCP-NEXT:    vmovdqa64 (%r9), %zmm63633; AVX512BW-FCP-NEXT:    vmovdqa64 (%r11), %zmm73634; AVX512BW-FCP-NEXT:    vmovdqa64 (%r10), %zmm83635; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [0,0,0,0,0,0,0,16,0,0,0,0,0,0,1,17]3636; AVX512BW-FCP-NEXT:    vpermi2d %zmm8, %zmm7, %zmm43637; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm9 = [0,0,0,0,0,16,0,0,0,0,0,0,1,17,0,0]3638; AVX512BW-FCP-NEXT:    vpermi2d %zmm6, %zmm5, %zmm93639; AVX512BW-FCP-NEXT:    movb $-120, %cl3640; AVX512BW-FCP-NEXT:    kmovd %ecx, %k13641; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm4, %zmm9 {%k1}3642; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm10 = [0,0,0,16,0,0,0,0,0,0,1,17,0,0,0,0]3643; AVX512BW-FCP-NEXT:    vpermi2d %zmm3, %zmm2, %zmm103644; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [0,16,0,0,0,0,0,0,1,17,0,0,0,0,0,0]3645; AVX512BW-FCP-NEXT:    vpermi2d %zmm1, %zmm0, %zmm43646; AVX512BW-FCP-NEXT:    movb $34, %cl3647; AVX512BW-FCP-NEXT:    kmovd %ecx, %k23648; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm10, %zmm4 {%k2}3649; AVX512BW-FCP-NEXT:    movb $-52, %cl3650; AVX512BW-FCP-NEXT:    kmovd %ecx, %k33651; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm9, %zmm4 {%k3}3652; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm9 = [0,0,0,0,0,0,2,18,0,0,0,0,0,0,3,19]3653; AVX512BW-FCP-NEXT:    vpermi2d %zmm8, %zmm7, %zmm93654; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm10 = [0,0,0,0,2,18,0,0,0,0,0,0,3,19,0,0]3655; AVX512BW-FCP-NEXT:    vpermi2d %zmm6, %zmm5, %zmm103656; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm9, %zmm10 {%k1}3657; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm11 = [0,0,2,18,0,0,0,0,0,0,3,19,0,0,0,0]3658; AVX512BW-FCP-NEXT:    vpermi2d %zmm3, %zmm2, %zmm113659; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm9 = [2,18,0,0,0,0,0,0,3,19,0,0,0,0,0,0]3660; AVX512BW-FCP-NEXT:    vpermi2d %zmm1, %zmm0, %zmm93661; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm11, %zmm9 {%k2}3662; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm10, %zmm9 {%k3}3663; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm10 = [0,0,0,0,0,0,4,20,0,0,0,0,0,0,5,21]3664; AVX512BW-FCP-NEXT:    vpermi2d %zmm8, %zmm7, %zmm103665; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm11 = [0,0,0,0,4,20,0,0,0,0,0,0,5,21,0,0]3666; AVX512BW-FCP-NEXT:    vpermi2d %zmm6, %zmm5, %zmm113667; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm10, %zmm11 {%k1}3668; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,4,20,0,0,0,0,0,0,5,21,0,0,0,0]3669; AVX512BW-FCP-NEXT:    vpermi2d %zmm3, %zmm2, %zmm123670; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm10 = [4,20,0,0,0,0,0,0,5,21,0,0,0,0,0,0]3671; AVX512BW-FCP-NEXT:    vpermi2d %zmm1, %zmm0, %zmm103672; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm12, %zmm10 {%k2}3673; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm11, %zmm10 {%k3}3674; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm11 = [0,0,0,0,0,0,6,22,0,0,0,0,0,0,7,23]3675; AVX512BW-FCP-NEXT:    vpermi2d %zmm8, %zmm7, %zmm113676; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,0,0,6,22,0,0,0,0,0,0,7,23,0,0]3677; AVX512BW-FCP-NEXT:    vpermi2d %zmm6, %zmm5, %zmm123678; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm11, %zmm12 {%k1}3679; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm13 = [0,0,6,22,0,0,0,0,0,0,7,23,0,0,0,0]3680; AVX512BW-FCP-NEXT:    vpermi2d %zmm3, %zmm2, %zmm133681; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm11 = [6,22,0,0,0,0,0,0,7,23,0,0,0,0,0,0]3682; AVX512BW-FCP-NEXT:    vpermi2d %zmm1, %zmm0, %zmm113683; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm13, %zmm11 {%k2}3684; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm12, %zmm11 {%k3}3685; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,0,0,0,0,8,24,0,0,0,0,0,0,9,25]3686; AVX512BW-FCP-NEXT:    vpermi2d %zmm8, %zmm7, %zmm123687; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm13 = [0,0,0,0,8,24,0,0,0,0,0,0,9,25,0,0]3688; AVX512BW-FCP-NEXT:    vpermi2d %zmm6, %zmm5, %zmm133689; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm12, %zmm13 {%k1}3690; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,8,24,0,0,0,0,0,0,9,25,0,0,0,0]3691; AVX512BW-FCP-NEXT:    vpermi2d %zmm3, %zmm2, %zmm123692; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm14 = [8,24,0,0,0,0,0,0,9,25,0,0,0,0,0,0]3693; AVX512BW-FCP-NEXT:    vpermi2d %zmm1, %zmm0, %zmm143694; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm12, %zmm14 {%k2}3695; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm13, %zmm14 {%k3}3696; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,0,0,0,0,10,26,0,0,0,0,0,0,11,27]3697; AVX512BW-FCP-NEXT:    vpermi2d %zmm8, %zmm7, %zmm123698; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm13 = [0,0,0,0,10,26,0,0,0,0,0,0,11,27,0,0]3699; AVX512BW-FCP-NEXT:    vpermi2d %zmm6, %zmm5, %zmm133700; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm12, %zmm13 {%k1}3701; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,10,26,0,0,0,0,0,0,11,27,0,0,0,0]3702; AVX512BW-FCP-NEXT:    vpermi2d %zmm3, %zmm2, %zmm123703; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm15 = [10,26,0,0,0,0,0,0,11,27,0,0,0,0,0,0]3704; AVX512BW-FCP-NEXT:    vpermi2d %zmm1, %zmm0, %zmm153705; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm12, %zmm15 {%k2}3706; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm13, %zmm15 {%k3}3707; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,0,0,0,0,12,28,0,0,0,0,0,0,13,29]3708; AVX512BW-FCP-NEXT:    vpermi2d %zmm8, %zmm7, %zmm123709; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm13 = [0,0,0,0,12,28,0,0,0,0,0,0,13,29,0,0]3710; AVX512BW-FCP-NEXT:    vpermi2d %zmm6, %zmm5, %zmm133711; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm12, %zmm13 {%k1}3712; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,12,28,0,0,0,0,0,0,13,29,0,0,0,0]3713; AVX512BW-FCP-NEXT:    vpermi2d %zmm3, %zmm2, %zmm123714; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm16 = [12,28,0,0,0,0,0,0,13,29,0,0,0,0,0,0]3715; AVX512BW-FCP-NEXT:    vpermi2d %zmm1, %zmm0, %zmm163716; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm12, %zmm16 {%k2}3717; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm13, %zmm16 {%k3}3718; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,0,0,0,0,14,30,0,0,0,0,0,0,15,31]3719; AVX512BW-FCP-NEXT:    vpermi2d %zmm8, %zmm7, %zmm123720; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm7 = [0,0,0,0,14,30,0,0,0,0,0,0,15,31,0,0]3721; AVX512BW-FCP-NEXT:    vpermi2d %zmm6, %zmm5, %zmm73722; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm12, %zmm7 {%k1}3723; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm5 = [0,0,14,30,0,0,0,0,0,0,15,31,0,0,0,0]3724; AVX512BW-FCP-NEXT:    vpermi2d %zmm3, %zmm2, %zmm53725; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [14,30,0,0,0,0,0,0,15,31,0,0,0,0,0,0]3726; AVX512BW-FCP-NEXT:    vpermi2d %zmm1, %zmm0, %zmm23727; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm5, %zmm2 {%k2}3728; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm7, %zmm2 {%k3}3729; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm2, 448(%rax)3730; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm16, 384(%rax)3731; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm15, 320(%rax)3732; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm14, 256(%rax)3733; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm11, 192(%rax)3734; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm10, 128(%rax)3735; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm9, 64(%rax)3736; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm4, (%rax)3737; AVX512BW-FCP-NEXT:    vzeroupper3738; AVX512BW-FCP-NEXT:    retq3739;3740; AVX512DQ-BW-LABEL: store_i32_stride8_vf16:3741; AVX512DQ-BW:       # %bb.0:3742; AVX512DQ-BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax3743; AVX512DQ-BW-NEXT:    movq {{[0-9]+}}(%rsp), %r103744; AVX512DQ-BW-NEXT:    movq {{[0-9]+}}(%rsp), %r113745; AVX512DQ-BW-NEXT:    vmovdqa64 (%rdi), %zmm03746; AVX512DQ-BW-NEXT:    vmovdqa64 (%rsi), %zmm13747; AVX512DQ-BW-NEXT:    vmovdqa64 (%rdx), %zmm23748; AVX512DQ-BW-NEXT:    vmovdqa64 (%rcx), %zmm33749; AVX512DQ-BW-NEXT:    vmovdqa64 (%r8), %zmm53750; AVX512DQ-BW-NEXT:    vmovdqa64 (%r9), %zmm63751; AVX512DQ-BW-NEXT:    vmovdqa64 (%r11), %zmm73752; AVX512DQ-BW-NEXT:    vmovdqa64 (%r10), %zmm83753; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [0,0,0,0,0,0,0,16,0,0,0,0,0,0,1,17]3754; AVX512DQ-BW-NEXT:    vpermi2d %zmm8, %zmm7, %zmm43755; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm9 = [0,0,0,0,0,16,0,0,0,0,0,0,1,17,0,0]3756; AVX512DQ-BW-NEXT:    vpermi2d %zmm6, %zmm5, %zmm93757; AVX512DQ-BW-NEXT:    movb $-120, %cl3758; AVX512DQ-BW-NEXT:    kmovd %ecx, %k13759; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm4, %zmm9 {%k1}3760; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm10 = [0,0,0,16,0,0,0,0,0,0,1,17,0,0,0,0]3761; AVX512DQ-BW-NEXT:    vpermi2d %zmm3, %zmm2, %zmm103762; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [0,16,0,0,0,0,0,0,1,17,0,0,0,0,0,0]3763; AVX512DQ-BW-NEXT:    vpermi2d %zmm1, %zmm0, %zmm43764; AVX512DQ-BW-NEXT:    movb $34, %cl3765; AVX512DQ-BW-NEXT:    kmovd %ecx, %k23766; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm10, %zmm4 {%k2}3767; AVX512DQ-BW-NEXT:    movb $-52, %cl3768; AVX512DQ-BW-NEXT:    kmovd %ecx, %k33769; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm9, %zmm4 {%k3}3770; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm9 = [0,0,0,0,0,0,2,18,0,0,0,0,0,0,3,19]3771; AVX512DQ-BW-NEXT:    vpermi2d %zmm8, %zmm7, %zmm93772; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm10 = [0,0,0,0,2,18,0,0,0,0,0,0,3,19,0,0]3773; AVX512DQ-BW-NEXT:    vpermi2d %zmm6, %zmm5, %zmm103774; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm9, %zmm10 {%k1}3775; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm11 = [0,0,2,18,0,0,0,0,0,0,3,19,0,0,0,0]3776; AVX512DQ-BW-NEXT:    vpermi2d %zmm3, %zmm2, %zmm113777; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm9 = [2,18,0,0,0,0,0,0,3,19,0,0,0,0,0,0]3778; AVX512DQ-BW-NEXT:    vpermi2d %zmm1, %zmm0, %zmm93779; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm11, %zmm9 {%k2}3780; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm10, %zmm9 {%k3}3781; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm10 = [0,0,0,0,0,0,4,20,0,0,0,0,0,0,5,21]3782; AVX512DQ-BW-NEXT:    vpermi2d %zmm8, %zmm7, %zmm103783; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm11 = [0,0,0,0,4,20,0,0,0,0,0,0,5,21,0,0]3784; AVX512DQ-BW-NEXT:    vpermi2d %zmm6, %zmm5, %zmm113785; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm10, %zmm11 {%k1}3786; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,4,20,0,0,0,0,0,0,5,21,0,0,0,0]3787; AVX512DQ-BW-NEXT:    vpermi2d %zmm3, %zmm2, %zmm123788; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm10 = [4,20,0,0,0,0,0,0,5,21,0,0,0,0,0,0]3789; AVX512DQ-BW-NEXT:    vpermi2d %zmm1, %zmm0, %zmm103790; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm12, %zmm10 {%k2}3791; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm11, %zmm10 {%k3}3792; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm11 = [0,0,0,0,0,0,6,22,0,0,0,0,0,0,7,23]3793; AVX512DQ-BW-NEXT:    vpermi2d %zmm8, %zmm7, %zmm113794; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,0,0,6,22,0,0,0,0,0,0,7,23,0,0]3795; AVX512DQ-BW-NEXT:    vpermi2d %zmm6, %zmm5, %zmm123796; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm11, %zmm12 {%k1}3797; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm13 = [0,0,6,22,0,0,0,0,0,0,7,23,0,0,0,0]3798; AVX512DQ-BW-NEXT:    vpermi2d %zmm3, %zmm2, %zmm133799; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm11 = [6,22,0,0,0,0,0,0,7,23,0,0,0,0,0,0]3800; AVX512DQ-BW-NEXT:    vpermi2d %zmm1, %zmm0, %zmm113801; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm13, %zmm11 {%k2}3802; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm12, %zmm11 {%k3}3803; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,0,0,0,0,8,24,0,0,0,0,0,0,9,25]3804; AVX512DQ-BW-NEXT:    vpermi2d %zmm8, %zmm7, %zmm123805; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm13 = [0,0,0,0,8,24,0,0,0,0,0,0,9,25,0,0]3806; AVX512DQ-BW-NEXT:    vpermi2d %zmm6, %zmm5, %zmm133807; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm12, %zmm13 {%k1}3808; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,8,24,0,0,0,0,0,0,9,25,0,0,0,0]3809; AVX512DQ-BW-NEXT:    vpermi2d %zmm3, %zmm2, %zmm123810; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm14 = [8,24,0,0,0,0,0,0,9,25,0,0,0,0,0,0]3811; AVX512DQ-BW-NEXT:    vpermi2d %zmm1, %zmm0, %zmm143812; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm12, %zmm14 {%k2}3813; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm13, %zmm14 {%k3}3814; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,0,0,0,0,10,26,0,0,0,0,0,0,11,27]3815; AVX512DQ-BW-NEXT:    vpermi2d %zmm8, %zmm7, %zmm123816; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm13 = [0,0,0,0,10,26,0,0,0,0,0,0,11,27,0,0]3817; AVX512DQ-BW-NEXT:    vpermi2d %zmm6, %zmm5, %zmm133818; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm12, %zmm13 {%k1}3819; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,10,26,0,0,0,0,0,0,11,27,0,0,0,0]3820; AVX512DQ-BW-NEXT:    vpermi2d %zmm3, %zmm2, %zmm123821; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm15 = [10,26,0,0,0,0,0,0,11,27,0,0,0,0,0,0]3822; AVX512DQ-BW-NEXT:    vpermi2d %zmm1, %zmm0, %zmm153823; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm12, %zmm15 {%k2}3824; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm13, %zmm15 {%k3}3825; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,0,0,0,0,12,28,0,0,0,0,0,0,13,29]3826; AVX512DQ-BW-NEXT:    vpermi2d %zmm8, %zmm7, %zmm123827; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm13 = [0,0,0,0,12,28,0,0,0,0,0,0,13,29,0,0]3828; AVX512DQ-BW-NEXT:    vpermi2d %zmm6, %zmm5, %zmm133829; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm12, %zmm13 {%k1}3830; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,12,28,0,0,0,0,0,0,13,29,0,0,0,0]3831; AVX512DQ-BW-NEXT:    vpermi2d %zmm3, %zmm2, %zmm123832; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm16 = [12,28,0,0,0,0,0,0,13,29,0,0,0,0,0,0]3833; AVX512DQ-BW-NEXT:    vpermi2d %zmm1, %zmm0, %zmm163834; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm12, %zmm16 {%k2}3835; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm13, %zmm16 {%k3}3836; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,0,0,0,0,14,30,0,0,0,0,0,0,15,31]3837; AVX512DQ-BW-NEXT:    vpermi2d %zmm8, %zmm7, %zmm123838; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm7 = [0,0,0,0,14,30,0,0,0,0,0,0,15,31,0,0]3839; AVX512DQ-BW-NEXT:    vpermi2d %zmm6, %zmm5, %zmm73840; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm12, %zmm7 {%k1}3841; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm5 = [0,0,14,30,0,0,0,0,0,0,15,31,0,0,0,0]3842; AVX512DQ-BW-NEXT:    vpermi2d %zmm3, %zmm2, %zmm53843; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [14,30,0,0,0,0,0,0,15,31,0,0,0,0,0,0]3844; AVX512DQ-BW-NEXT:    vpermi2d %zmm1, %zmm0, %zmm23845; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm5, %zmm2 {%k2}3846; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm7, %zmm2 {%k3}3847; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm2, 448(%rax)3848; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm16, 384(%rax)3849; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm15, 320(%rax)3850; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm14, 256(%rax)3851; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm11, 192(%rax)3852; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm10, 128(%rax)3853; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm9, 64(%rax)3854; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm4, (%rax)3855; AVX512DQ-BW-NEXT:    vzeroupper3856; AVX512DQ-BW-NEXT:    retq3857;3858; AVX512DQ-BW-FCP-LABEL: store_i32_stride8_vf16:3859; AVX512DQ-BW-FCP:       # %bb.0:3860; AVX512DQ-BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax3861; AVX512DQ-BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r103862; AVX512DQ-BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r113863; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 (%rdi), %zmm03864; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 (%rsi), %zmm13865; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 (%rdx), %zmm23866; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 (%rcx), %zmm33867; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 (%r8), %zmm53868; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 (%r9), %zmm63869; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 (%r11), %zmm73870; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 (%r10), %zmm83871; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [0,0,0,0,0,0,0,16,0,0,0,0,0,0,1,17]3872; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm8, %zmm7, %zmm43873; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm9 = [0,0,0,0,0,16,0,0,0,0,0,0,1,17,0,0]3874; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm6, %zmm5, %zmm93875; AVX512DQ-BW-FCP-NEXT:    movb $-120, %cl3876; AVX512DQ-BW-FCP-NEXT:    kmovd %ecx, %k13877; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm4, %zmm9 {%k1}3878; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm10 = [0,0,0,16,0,0,0,0,0,0,1,17,0,0,0,0]3879; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm3, %zmm2, %zmm103880; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [0,16,0,0,0,0,0,0,1,17,0,0,0,0,0,0]3881; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm1, %zmm0, %zmm43882; AVX512DQ-BW-FCP-NEXT:    movb $34, %cl3883; AVX512DQ-BW-FCP-NEXT:    kmovd %ecx, %k23884; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm10, %zmm4 {%k2}3885; AVX512DQ-BW-FCP-NEXT:    movb $-52, %cl3886; AVX512DQ-BW-FCP-NEXT:    kmovd %ecx, %k33887; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm9, %zmm4 {%k3}3888; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm9 = [0,0,0,0,0,0,2,18,0,0,0,0,0,0,3,19]3889; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm8, %zmm7, %zmm93890; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm10 = [0,0,0,0,2,18,0,0,0,0,0,0,3,19,0,0]3891; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm6, %zmm5, %zmm103892; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm9, %zmm10 {%k1}3893; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm11 = [0,0,2,18,0,0,0,0,0,0,3,19,0,0,0,0]3894; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm3, %zmm2, %zmm113895; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm9 = [2,18,0,0,0,0,0,0,3,19,0,0,0,0,0,0]3896; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm1, %zmm0, %zmm93897; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm11, %zmm9 {%k2}3898; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm10, %zmm9 {%k3}3899; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm10 = [0,0,0,0,0,0,4,20,0,0,0,0,0,0,5,21]3900; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm8, %zmm7, %zmm103901; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm11 = [0,0,0,0,4,20,0,0,0,0,0,0,5,21,0,0]3902; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm6, %zmm5, %zmm113903; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm10, %zmm11 {%k1}3904; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,4,20,0,0,0,0,0,0,5,21,0,0,0,0]3905; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm3, %zmm2, %zmm123906; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm10 = [4,20,0,0,0,0,0,0,5,21,0,0,0,0,0,0]3907; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm1, %zmm0, %zmm103908; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm12, %zmm10 {%k2}3909; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm11, %zmm10 {%k3}3910; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm11 = [0,0,0,0,0,0,6,22,0,0,0,0,0,0,7,23]3911; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm8, %zmm7, %zmm113912; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,0,0,6,22,0,0,0,0,0,0,7,23,0,0]3913; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm6, %zmm5, %zmm123914; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm11, %zmm12 {%k1}3915; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm13 = [0,0,6,22,0,0,0,0,0,0,7,23,0,0,0,0]3916; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm3, %zmm2, %zmm133917; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm11 = [6,22,0,0,0,0,0,0,7,23,0,0,0,0,0,0]3918; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm1, %zmm0, %zmm113919; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm13, %zmm11 {%k2}3920; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm12, %zmm11 {%k3}3921; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,0,0,0,0,8,24,0,0,0,0,0,0,9,25]3922; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm8, %zmm7, %zmm123923; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm13 = [0,0,0,0,8,24,0,0,0,0,0,0,9,25,0,0]3924; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm6, %zmm5, %zmm133925; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm12, %zmm13 {%k1}3926; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,8,24,0,0,0,0,0,0,9,25,0,0,0,0]3927; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm3, %zmm2, %zmm123928; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm14 = [8,24,0,0,0,0,0,0,9,25,0,0,0,0,0,0]3929; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm1, %zmm0, %zmm143930; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm12, %zmm14 {%k2}3931; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm13, %zmm14 {%k3}3932; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,0,0,0,0,10,26,0,0,0,0,0,0,11,27]3933; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm8, %zmm7, %zmm123934; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm13 = [0,0,0,0,10,26,0,0,0,0,0,0,11,27,0,0]3935; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm6, %zmm5, %zmm133936; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm12, %zmm13 {%k1}3937; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,10,26,0,0,0,0,0,0,11,27,0,0,0,0]3938; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm3, %zmm2, %zmm123939; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm15 = [10,26,0,0,0,0,0,0,11,27,0,0,0,0,0,0]3940; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm1, %zmm0, %zmm153941; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm12, %zmm15 {%k2}3942; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm13, %zmm15 {%k3}3943; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,0,0,0,0,12,28,0,0,0,0,0,0,13,29]3944; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm8, %zmm7, %zmm123945; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm13 = [0,0,0,0,12,28,0,0,0,0,0,0,13,29,0,0]3946; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm6, %zmm5, %zmm133947; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm12, %zmm13 {%k1}3948; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,12,28,0,0,0,0,0,0,13,29,0,0,0,0]3949; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm3, %zmm2, %zmm123950; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm16 = [12,28,0,0,0,0,0,0,13,29,0,0,0,0,0,0]3951; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm1, %zmm0, %zmm163952; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm12, %zmm16 {%k2}3953; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm13, %zmm16 {%k3}3954; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,0,0,0,0,14,30,0,0,0,0,0,0,15,31]3955; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm8, %zmm7, %zmm123956; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm7 = [0,0,0,0,14,30,0,0,0,0,0,0,15,31,0,0]3957; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm6, %zmm5, %zmm73958; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm12, %zmm7 {%k1}3959; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm5 = [0,0,14,30,0,0,0,0,0,0,15,31,0,0,0,0]3960; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm3, %zmm2, %zmm53961; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [14,30,0,0,0,0,0,0,15,31,0,0,0,0,0,0]3962; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm1, %zmm0, %zmm23963; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm5, %zmm2 {%k2}3964; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm7, %zmm2 {%k3}3965; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm2, 448(%rax)3966; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm16, 384(%rax)3967; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm15, 320(%rax)3968; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm14, 256(%rax)3969; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm11, 192(%rax)3970; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm10, 128(%rax)3971; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm9, 64(%rax)3972; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm4, (%rax)3973; AVX512DQ-BW-FCP-NEXT:    vzeroupper3974; AVX512DQ-BW-FCP-NEXT:    retq3975  %in.vec0 = load <16 x i32>, ptr %in.vecptr0, align 643976  %in.vec1 = load <16 x i32>, ptr %in.vecptr1, align 643977  %in.vec2 = load <16 x i32>, ptr %in.vecptr2, align 643978  %in.vec3 = load <16 x i32>, ptr %in.vecptr3, align 643979  %in.vec4 = load <16 x i32>, ptr %in.vecptr4, align 643980  %in.vec5 = load <16 x i32>, ptr %in.vecptr5, align 643981  %in.vec6 = load <16 x i32>, ptr %in.vecptr6, align 643982  %in.vec7 = load <16 x i32>, ptr %in.vecptr7, align 643983  %1 = shufflevector <16 x i32> %in.vec0, <16 x i32> %in.vec1, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>3984  %2 = shufflevector <16 x i32> %in.vec2, <16 x i32> %in.vec3, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>3985  %3 = shufflevector <16 x i32> %in.vec4, <16 x i32> %in.vec5, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>3986  %4 = shufflevector <16 x i32> %in.vec6, <16 x i32> %in.vec7, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>3987  %5 = shufflevector <32 x i32> %1, <32 x i32> %2, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>3988  %6 = shufflevector <32 x i32> %3, <32 x i32> %4, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>3989  %7 = shufflevector <64 x i32> %5, <64 x i32> %6, <128 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 64, i32 65, i32 66, i32 67, i32 68, i32 69, i32 70, i32 71, i32 72, i32 73, i32 74, i32 75, i32 76, i32 77, i32 78, i32 79, i32 80, i32 81, i32 82, i32 83, i32 84, i32 85, i32 86, i32 87, i32 88, i32 89, i32 90, i32 91, i32 92, i32 93, i32 94, i32 95, i32 96, i32 97, i32 98, i32 99, i32 100, i32 101, i32 102, i32 103, i32 104, i32 105, i32 106, i32 107, i32 108, i32 109, i32 110, i32 111, i32 112, i32 113, i32 114, i32 115, i32 116, i32 117, i32 118, i32 119, i32 120, i32 121, i32 122, i32 123, i32 124, i32 125, i32 126, i32 127>3990  %interleaved.vec = shufflevector <128 x i32> %7, <128 x i32> poison, <128 x i32> <i32 0, i32 16, i32 32, i32 48, i32 64, i32 80, i32 96, i32 112, i32 1, i32 17, i32 33, i32 49, i32 65, i32 81, i32 97, i32 113, i32 2, i32 18, i32 34, i32 50, i32 66, i32 82, i32 98, i32 114, i32 3, i32 19, i32 35, i32 51, i32 67, i32 83, i32 99, i32 115, i32 4, i32 20, i32 36, i32 52, i32 68, i32 84, i32 100, i32 116, i32 5, i32 21, i32 37, i32 53, i32 69, i32 85, i32 101, i32 117, i32 6, i32 22, i32 38, i32 54, i32 70, i32 86, i32 102, i32 118, i32 7, i32 23, i32 39, i32 55, i32 71, i32 87, i32 103, i32 119, i32 8, i32 24, i32 40, i32 56, i32 72, i32 88, i32 104, i32 120, i32 9, i32 25, i32 41, i32 57, i32 73, i32 89, i32 105, i32 121, i32 10, i32 26, i32 42, i32 58, i32 74, i32 90, i32 106, i32 122, i32 11, i32 27, i32 43, i32 59, i32 75, i32 91, i32 107, i32 123, i32 12, i32 28, i32 44, i32 60, i32 76, i32 92, i32 108, i32 124, i32 13, i32 29, i32 45, i32 61, i32 77, i32 93, i32 109, i32 125, i32 14, i32 30, i32 46, i32 62, i32 78, i32 94, i32 110, i32 126, i32 15, i32 31, i32 47, i32 63, i32 79, i32 95, i32 111, i32 127>3991  store <128 x i32> %interleaved.vec, ptr %out.vec, align 643992  ret void3993}3994 3995define void @store_i32_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 {3996; SSE-LABEL: store_i32_stride8_vf32:3997; SSE:       # %bb.0:3998; SSE-NEXT:    subq $728, %rsp # imm = 0x2D83999; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %rax4000; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %r104001; SSE-NEXT:    movaps (%rdi), %xmm94002; SSE-NEXT:    movaps 16(%rdi), %xmm104003; SSE-NEXT:    movaps (%rsi), %xmm34004; SSE-NEXT:    movaps 16(%rsi), %xmm14005; SSE-NEXT:    movaps (%rdx), %xmm24006; SSE-NEXT:    movaps 16(%rdx), %xmm04007; SSE-NEXT:    movaps (%rcx), %xmm44008; SSE-NEXT:    movaps (%r8), %xmm114009; SSE-NEXT:    movaps (%r9), %xmm64010; SSE-NEXT:    movaps (%r10), %xmm54011; SSE-NEXT:    movaps (%rax), %xmm74012; SSE-NEXT:    movaps %xmm4, %xmm84013; SSE-NEXT:    movlhps {{.*#+}} xmm8 = xmm8[0],xmm2[0]4014; SSE-NEXT:    movaps %xmm9, %xmm134015; SSE-NEXT:    unpcklps {{.*#+}} xmm13 = xmm13[0],xmm3[0],xmm13[1],xmm3[1]4016; SSE-NEXT:    movaps %xmm13, %xmm124017; SSE-NEXT:    shufps {{.*#+}} xmm12 = xmm12[0,1],xmm8[2,0]4018; SSE-NEXT:    movaps %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4019; SSE-NEXT:    movaps %xmm7, %xmm84020; SSE-NEXT:    movlhps {{.*#+}} xmm8 = xmm8[0],xmm5[0]4021; SSE-NEXT:    movaps %xmm11, %xmm144022; SSE-NEXT:    unpcklps {{.*#+}} xmm14 = xmm14[0],xmm6[0],xmm14[1],xmm6[1]4023; SSE-NEXT:    movaps %xmm14, %xmm124024; SSE-NEXT:    shufps {{.*#+}} xmm12 = xmm12[0,1],xmm8[2,0]4025; SSE-NEXT:    movaps %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4026; SSE-NEXT:    movaps %xmm4, %xmm84027; SSE-NEXT:    shufps {{.*#+}} xmm8 = xmm8[1,1],xmm2[1,1]4028; SSE-NEXT:    shufps {{.*#+}} xmm13 = xmm13[2,3],xmm8[2,0]4029; SSE-NEXT:    movaps %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4030; SSE-NEXT:    movaps %xmm7, %xmm84031; SSE-NEXT:    shufps {{.*#+}} xmm8 = xmm8[1,1],xmm5[1,1]4032; SSE-NEXT:    shufps {{.*#+}} xmm14 = xmm14[2,3],xmm8[2,0]4033; SSE-NEXT:    movaps %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4034; SSE-NEXT:    movaps %xmm2, %xmm84035; SSE-NEXT:    unpckhpd {{.*#+}} xmm8 = xmm8[1],xmm4[1]4036; SSE-NEXT:    unpckhps {{.*#+}} xmm9 = xmm9[2],xmm3[2],xmm9[3],xmm3[3]4037; SSE-NEXT:    movaps %xmm9, %xmm34038; SSE-NEXT:    shufps {{.*#+}} xmm3 = xmm3[0,1],xmm8[0,2]4039; SSE-NEXT:    movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4040; SSE-NEXT:    movaps %xmm5, %xmm34041; SSE-NEXT:    unpckhpd {{.*#+}} xmm3 = xmm3[1],xmm7[1]4042; SSE-NEXT:    unpckhps {{.*#+}} xmm11 = xmm11[2],xmm6[2],xmm11[3],xmm6[3]4043; SSE-NEXT:    movaps %xmm11, %xmm64044; SSE-NEXT:    shufps {{.*#+}} xmm6 = xmm6[0,1],xmm3[0,2]4045; SSE-NEXT:    movaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4046; SSE-NEXT:    movaps 16(%rcx), %xmm64047; SSE-NEXT:    shufps {{.*#+}} xmm2 = xmm2[3,3],xmm4[3,3]4048; SSE-NEXT:    movaps 16(%r10), %xmm34049; SSE-NEXT:    shufps {{.*#+}} xmm9 = xmm9[2,3],xmm2[0,2]4050; SSE-NEXT:    movaps %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4051; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[3,3],xmm7[3,3]4052; SSE-NEXT:    shufps {{.*#+}} xmm11 = xmm11[2,3],xmm5[0,2]4053; SSE-NEXT:    movaps %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4054; SSE-NEXT:    movaps %xmm6, %xmm24055; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]4056; SSE-NEXT:    movaps %xmm10, %xmm54057; SSE-NEXT:    unpcklps {{.*#+}} xmm5 = xmm5[0],xmm1[0],xmm5[1],xmm1[1]4058; SSE-NEXT:    movaps %xmm5, %xmm44059; SSE-NEXT:    movaps %xmm5, %xmm74060; SSE-NEXT:    shufps {{.*#+}} xmm4 = xmm4[0,1],xmm2[2,0]4061; SSE-NEXT:    movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4062; SSE-NEXT:    movaps 16(%rax), %xmm24063; SSE-NEXT:    movaps %xmm2, %xmm44064; SSE-NEXT:    movlhps {{.*#+}} xmm4 = xmm4[0],xmm3[0]4065; SSE-NEXT:    movaps 16(%r8), %xmm114066; SSE-NEXT:    movaps 16(%r9), %xmm54067; SSE-NEXT:    movaps %xmm11, %xmm94068; SSE-NEXT:    unpcklps {{.*#+}} xmm9 = xmm9[0],xmm5[0],xmm9[1],xmm5[1]4069; SSE-NEXT:    movaps %xmm9, %xmm84070; SSE-NEXT:    shufps {{.*#+}} xmm8 = xmm8[0,1],xmm4[2,0]4071; SSE-NEXT:    movaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4072; SSE-NEXT:    movaps %xmm6, %xmm44073; SSE-NEXT:    shufps {{.*#+}} xmm4 = xmm4[1,1],xmm0[1,1]4074; SSE-NEXT:    shufps {{.*#+}} xmm7 = xmm7[2,3],xmm4[2,0]4075; SSE-NEXT:    movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4076; SSE-NEXT:    movaps %xmm2, %xmm44077; SSE-NEXT:    shufps {{.*#+}} xmm4 = xmm4[1,1],xmm3[1,1]4078; SSE-NEXT:    shufps {{.*#+}} xmm9 = xmm9[2,3],xmm4[2,0]4079; SSE-NEXT:    movaps %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4080; SSE-NEXT:    unpckhps {{.*#+}} xmm10 = xmm10[2],xmm1[2],xmm10[3],xmm1[3]4081; SSE-NEXT:    movaps %xmm0, %xmm14082; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm6[1]4083; SSE-NEXT:    movaps %xmm10, %xmm44084; SSE-NEXT:    shufps {{.*#+}} xmm4 = xmm4[0,1],xmm1[0,2]4085; SSE-NEXT:    movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4086; SSE-NEXT:    unpckhps {{.*#+}} xmm11 = xmm11[2],xmm5[2],xmm11[3],xmm5[3]4087; SSE-NEXT:    movaps %xmm3, %xmm14088; SSE-NEXT:    unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm2[1]4089; SSE-NEXT:    movaps %xmm11, %xmm44090; SSE-NEXT:    shufps {{.*#+}} xmm4 = xmm4[0,1],xmm1[0,2]4091; SSE-NEXT:    movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4092; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[3,3],xmm6[3,3]4093; SSE-NEXT:    shufps {{.*#+}} xmm10 = xmm10[2,3],xmm0[0,2]4094; SSE-NEXT:    movaps %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4095; SSE-NEXT:    shufps {{.*#+}} xmm3 = xmm3[3,3],xmm2[3,3]4096; SSE-NEXT:    shufps {{.*#+}} xmm11 = xmm11[2,3],xmm3[0,2]4097; SSE-NEXT:    movaps %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4098; SSE-NEXT:    movaps 32(%rdx), %xmm04099; SSE-NEXT:    movaps 32(%rcx), %xmm14100; SSE-NEXT:    movaps %xmm1, %xmm24101; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]4102; SSE-NEXT:    movaps 32(%rdi), %xmm74103; SSE-NEXT:    movaps 32(%rsi), %xmm44104; SSE-NEXT:    movaps %xmm7, %xmm54105; SSE-NEXT:    unpcklps {{.*#+}} xmm5 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]4106; SSE-NEXT:    movaps %xmm5, %xmm34107; SSE-NEXT:    movaps %xmm5, %xmm84108; SSE-NEXT:    shufps {{.*#+}} xmm3 = xmm3[0,1],xmm2[2,0]4109; SSE-NEXT:    movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4110; SSE-NEXT:    movaps 32(%r10), %xmm24111; SSE-NEXT:    movaps 32(%rax), %xmm34112; SSE-NEXT:    movaps %xmm3, %xmm54113; SSE-NEXT:    movlhps {{.*#+}} xmm5 = xmm5[0],xmm2[0]4114; SSE-NEXT:    movaps 32(%r8), %xmm114115; SSE-NEXT:    movaps 32(%r9), %xmm64116; SSE-NEXT:    movaps %xmm11, %xmm104117; SSE-NEXT:    unpcklps {{.*#+}} xmm10 = xmm10[0],xmm6[0],xmm10[1],xmm6[1]4118; SSE-NEXT:    movaps %xmm10, %xmm94119; SSE-NEXT:    shufps {{.*#+}} xmm9 = xmm9[0,1],xmm5[2,0]4120; SSE-NEXT:    movaps %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4121; SSE-NEXT:    movaps %xmm1, %xmm54122; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[1,1],xmm0[1,1]4123; SSE-NEXT:    shufps {{.*#+}} xmm8 = xmm8[2,3],xmm5[2,0]4124; SSE-NEXT:    movaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4125; SSE-NEXT:    movaps %xmm3, %xmm54126; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[1,1],xmm2[1,1]4127; SSE-NEXT:    shufps {{.*#+}} xmm10 = xmm10[2,3],xmm5[2,0]4128; SSE-NEXT:    movaps %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4129; SSE-NEXT:    unpckhps {{.*#+}} xmm7 = xmm7[2],xmm4[2],xmm7[3],xmm4[3]4130; SSE-NEXT:    movaps %xmm0, %xmm44131; SSE-NEXT:    unpckhpd {{.*#+}} xmm4 = xmm4[1],xmm1[1]4132; SSE-NEXT:    movaps %xmm7, %xmm54133; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[0,1],xmm4[0,2]4134; SSE-NEXT:    movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4135; SSE-NEXT:    unpckhps {{.*#+}} xmm11 = xmm11[2],xmm6[2],xmm11[3],xmm6[3]4136; SSE-NEXT:    movaps %xmm2, %xmm44137; SSE-NEXT:    unpckhpd {{.*#+}} xmm4 = xmm4[1],xmm3[1]4138; SSE-NEXT:    movaps %xmm11, %xmm54139; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[0,1],xmm4[0,2]4140; SSE-NEXT:    movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4141; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[3,3],xmm1[3,3]4142; SSE-NEXT:    shufps {{.*#+}} xmm7 = xmm7[2,3],xmm0[0,2]4143; SSE-NEXT:    movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4144; SSE-NEXT:    shufps {{.*#+}} xmm2 = xmm2[3,3],xmm3[3,3]4145; SSE-NEXT:    shufps {{.*#+}} xmm11 = xmm11[2,3],xmm2[0,2]4146; SSE-NEXT:    movaps %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4147; SSE-NEXT:    movaps 48(%rdx), %xmm04148; SSE-NEXT:    movaps 48(%rcx), %xmm14149; SSE-NEXT:    movaps %xmm1, %xmm24150; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]4151; SSE-NEXT:    movaps 48(%rdi), %xmm74152; SSE-NEXT:    movaps 48(%rsi), %xmm44153; SSE-NEXT:    movaps %xmm7, %xmm54154; SSE-NEXT:    unpcklps {{.*#+}} xmm5 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]4155; SSE-NEXT:    movaps %xmm5, %xmm34156; SSE-NEXT:    movaps %xmm5, %xmm84157; SSE-NEXT:    shufps {{.*#+}} xmm3 = xmm3[0,1],xmm2[2,0]4158; SSE-NEXT:    movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4159; SSE-NEXT:    movaps 48(%r10), %xmm24160; SSE-NEXT:    movaps 48(%rax), %xmm34161; SSE-NEXT:    movaps %xmm3, %xmm54162; SSE-NEXT:    movlhps {{.*#+}} xmm5 = xmm5[0],xmm2[0]4163; SSE-NEXT:    movaps 48(%r8), %xmm114164; SSE-NEXT:    movaps 48(%r9), %xmm64165; SSE-NEXT:    movaps %xmm11, %xmm104166; SSE-NEXT:    unpcklps {{.*#+}} xmm10 = xmm10[0],xmm6[0],xmm10[1],xmm6[1]4167; SSE-NEXT:    movaps %xmm10, %xmm94168; SSE-NEXT:    shufps {{.*#+}} xmm9 = xmm9[0,1],xmm5[2,0]4169; SSE-NEXT:    movaps %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4170; SSE-NEXT:    movaps %xmm1, %xmm54171; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[1,1],xmm0[1,1]4172; SSE-NEXT:    shufps {{.*#+}} xmm8 = xmm8[2,3],xmm5[2,0]4173; SSE-NEXT:    movaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4174; SSE-NEXT:    movaps %xmm3, %xmm54175; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[1,1],xmm2[1,1]4176; SSE-NEXT:    shufps {{.*#+}} xmm10 = xmm10[2,3],xmm5[2,0]4177; SSE-NEXT:    movaps %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4178; SSE-NEXT:    unpckhps {{.*#+}} xmm7 = xmm7[2],xmm4[2],xmm7[3],xmm4[3]4179; SSE-NEXT:    movaps %xmm0, %xmm44180; SSE-NEXT:    unpckhpd {{.*#+}} xmm4 = xmm4[1],xmm1[1]4181; SSE-NEXT:    movaps %xmm7, %xmm54182; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[0,1],xmm4[0,2]4183; SSE-NEXT:    movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4184; SSE-NEXT:    unpckhps {{.*#+}} xmm11 = xmm11[2],xmm6[2],xmm11[3],xmm6[3]4185; SSE-NEXT:    movaps %xmm2, %xmm44186; SSE-NEXT:    unpckhpd {{.*#+}} xmm4 = xmm4[1],xmm3[1]4187; SSE-NEXT:    movaps %xmm11, %xmm54188; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[0,1],xmm4[0,2]4189; SSE-NEXT:    movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4190; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[3,3],xmm1[3,3]4191; SSE-NEXT:    shufps {{.*#+}} xmm7 = xmm7[2,3],xmm0[0,2]4192; SSE-NEXT:    movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4193; SSE-NEXT:    shufps {{.*#+}} xmm2 = xmm2[3,3],xmm3[3,3]4194; SSE-NEXT:    shufps {{.*#+}} xmm11 = xmm11[2,3],xmm2[0,2]4195; SSE-NEXT:    movaps %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4196; SSE-NEXT:    movaps 64(%rdx), %xmm04197; SSE-NEXT:    movaps 64(%rcx), %xmm14198; SSE-NEXT:    movaps %xmm1, %xmm24199; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]4200; SSE-NEXT:    movaps 64(%rdi), %xmm74201; SSE-NEXT:    movaps 64(%rsi), %xmm44202; SSE-NEXT:    movaps %xmm7, %xmm54203; SSE-NEXT:    unpcklps {{.*#+}} xmm5 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]4204; SSE-NEXT:    movaps %xmm5, %xmm34205; SSE-NEXT:    movaps %xmm5, %xmm84206; SSE-NEXT:    shufps {{.*#+}} xmm3 = xmm3[0,1],xmm2[2,0]4207; SSE-NEXT:    movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4208; SSE-NEXT:    movaps 64(%r10), %xmm24209; SSE-NEXT:    movaps 64(%rax), %xmm34210; SSE-NEXT:    movaps %xmm3, %xmm54211; SSE-NEXT:    movlhps {{.*#+}} xmm5 = xmm5[0],xmm2[0]4212; SSE-NEXT:    movaps 64(%r8), %xmm114213; SSE-NEXT:    movaps 64(%r9), %xmm64214; SSE-NEXT:    movaps %xmm11, %xmm104215; SSE-NEXT:    unpcklps {{.*#+}} xmm10 = xmm10[0],xmm6[0],xmm10[1],xmm6[1]4216; SSE-NEXT:    movaps %xmm10, %xmm94217; SSE-NEXT:    shufps {{.*#+}} xmm9 = xmm9[0,1],xmm5[2,0]4218; SSE-NEXT:    movaps %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4219; SSE-NEXT:    movaps %xmm1, %xmm54220; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[1,1],xmm0[1,1]4221; SSE-NEXT:    shufps {{.*#+}} xmm8 = xmm8[2,3],xmm5[2,0]4222; SSE-NEXT:    movaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4223; SSE-NEXT:    movaps %xmm3, %xmm54224; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[1,1],xmm2[1,1]4225; SSE-NEXT:    shufps {{.*#+}} xmm10 = xmm10[2,3],xmm5[2,0]4226; SSE-NEXT:    movaps %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4227; SSE-NEXT:    unpckhps {{.*#+}} xmm7 = xmm7[2],xmm4[2],xmm7[3],xmm4[3]4228; SSE-NEXT:    movaps %xmm0, %xmm44229; SSE-NEXT:    unpckhpd {{.*#+}} xmm4 = xmm4[1],xmm1[1]4230; SSE-NEXT:    movaps %xmm7, %xmm54231; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[0,1],xmm4[0,2]4232; SSE-NEXT:    movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4233; SSE-NEXT:    unpckhps {{.*#+}} xmm11 = xmm11[2],xmm6[2],xmm11[3],xmm6[3]4234; SSE-NEXT:    movaps %xmm2, %xmm44235; SSE-NEXT:    unpckhpd {{.*#+}} xmm4 = xmm4[1],xmm3[1]4236; SSE-NEXT:    movaps %xmm11, %xmm54237; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[0,1],xmm4[0,2]4238; SSE-NEXT:    movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4239; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[3,3],xmm1[3,3]4240; SSE-NEXT:    shufps {{.*#+}} xmm7 = xmm7[2,3],xmm0[0,2]4241; SSE-NEXT:    movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4242; SSE-NEXT:    shufps {{.*#+}} xmm2 = xmm2[3,3],xmm3[3,3]4243; SSE-NEXT:    shufps {{.*#+}} xmm11 = xmm11[2,3],xmm2[0,2]4244; SSE-NEXT:    movaps %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4245; SSE-NEXT:    movaps 80(%rdx), %xmm04246; SSE-NEXT:    movaps 80(%rcx), %xmm14247; SSE-NEXT:    movaps %xmm1, %xmm24248; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]4249; SSE-NEXT:    movaps 80(%rdi), %xmm74250; SSE-NEXT:    movaps 80(%rsi), %xmm44251; SSE-NEXT:    movaps %xmm7, %xmm54252; SSE-NEXT:    unpcklps {{.*#+}} xmm5 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]4253; SSE-NEXT:    movaps %xmm5, %xmm34254; SSE-NEXT:    movaps %xmm5, %xmm84255; SSE-NEXT:    shufps {{.*#+}} xmm3 = xmm3[0,1],xmm2[2,0]4256; SSE-NEXT:    movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4257; SSE-NEXT:    movaps 80(%r10), %xmm24258; SSE-NEXT:    movaps 80(%rax), %xmm34259; SSE-NEXT:    movaps %xmm3, %xmm54260; SSE-NEXT:    movlhps {{.*#+}} xmm5 = xmm5[0],xmm2[0]4261; SSE-NEXT:    movaps 80(%r8), %xmm114262; SSE-NEXT:    movaps 80(%r9), %xmm64263; SSE-NEXT:    movaps %xmm11, %xmm104264; SSE-NEXT:    unpcklps {{.*#+}} xmm10 = xmm10[0],xmm6[0],xmm10[1],xmm6[1]4265; SSE-NEXT:    movaps %xmm10, %xmm94266; SSE-NEXT:    shufps {{.*#+}} xmm9 = xmm9[0,1],xmm5[2,0]4267; SSE-NEXT:    movaps %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4268; SSE-NEXT:    movaps %xmm1, %xmm54269; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[1,1],xmm0[1,1]4270; SSE-NEXT:    shufps {{.*#+}} xmm8 = xmm8[2,3],xmm5[2,0]4271; SSE-NEXT:    movaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4272; SSE-NEXT:    movaps %xmm3, %xmm54273; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[1,1],xmm2[1,1]4274; SSE-NEXT:    shufps {{.*#+}} xmm10 = xmm10[2,3],xmm5[2,0]4275; SSE-NEXT:    movaps %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4276; SSE-NEXT:    unpckhps {{.*#+}} xmm7 = xmm7[2],xmm4[2],xmm7[3],xmm4[3]4277; SSE-NEXT:    movaps %xmm0, %xmm44278; SSE-NEXT:    unpckhpd {{.*#+}} xmm4 = xmm4[1],xmm1[1]4279; SSE-NEXT:    movaps %xmm7, %xmm54280; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[0,1],xmm4[0,2]4281; SSE-NEXT:    movaps %xmm5, (%rsp) # 16-byte Spill4282; SSE-NEXT:    unpckhps {{.*#+}} xmm11 = xmm11[2],xmm6[2],xmm11[3],xmm6[3]4283; SSE-NEXT:    movaps %xmm2, %xmm44284; SSE-NEXT:    unpckhpd {{.*#+}} xmm4 = xmm4[1],xmm3[1]4285; SSE-NEXT:    movaps %xmm11, %xmm54286; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[0,1],xmm4[0,2]4287; SSE-NEXT:    movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4288; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[3,3],xmm1[3,3]4289; SSE-NEXT:    shufps {{.*#+}} xmm7 = xmm7[2,3],xmm0[0,2]4290; SSE-NEXT:    movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4291; SSE-NEXT:    shufps {{.*#+}} xmm2 = xmm2[3,3],xmm3[3,3]4292; SSE-NEXT:    shufps {{.*#+}} xmm11 = xmm11[2,3],xmm2[0,2]4293; SSE-NEXT:    movaps %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4294; SSE-NEXT:    movaps 96(%rdx), %xmm24295; SSE-NEXT:    movaps 96(%rcx), %xmm04296; SSE-NEXT:    movaps %xmm0, %xmm14297; SSE-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]4298; SSE-NEXT:    movaps 96(%rdi), %xmm154299; SSE-NEXT:    movaps 96(%rsi), %xmm44300; SSE-NEXT:    movaps %xmm15, %xmm54301; SSE-NEXT:    unpcklps {{.*#+}} xmm5 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]4302; SSE-NEXT:    movaps %xmm5, %xmm34303; SSE-NEXT:    movaps %xmm5, %xmm74304; SSE-NEXT:    shufps {{.*#+}} xmm3 = xmm3[0,1],xmm1[2,0]4305; SSE-NEXT:    movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4306; SSE-NEXT:    movaps 96(%r10), %xmm14307; SSE-NEXT:    movaps 96(%rax), %xmm34308; SSE-NEXT:    movaps %xmm3, %xmm54309; SSE-NEXT:    movlhps {{.*#+}} xmm5 = xmm5[0],xmm1[0]4310; SSE-NEXT:    movaps 96(%r8), %xmm134311; SSE-NEXT:    movaps 96(%r9), %xmm64312; SSE-NEXT:    movaps %xmm13, %xmm124313; SSE-NEXT:    unpcklps {{.*#+}} xmm12 = xmm12[0],xmm6[0],xmm12[1],xmm6[1]4314; SSE-NEXT:    movaps %xmm12, %xmm84315; SSE-NEXT:    shufps {{.*#+}} xmm8 = xmm8[0,1],xmm5[2,0]4316; SSE-NEXT:    movaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4317; SSE-NEXT:    movaps %xmm0, %xmm54318; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[1,1],xmm2[1,1]4319; SSE-NEXT:    shufps {{.*#+}} xmm7 = xmm7[2,3],xmm5[2,0]4320; SSE-NEXT:    movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4321; SSE-NEXT:    movaps %xmm3, %xmm54322; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[1,1],xmm1[1,1]4323; SSE-NEXT:    shufps {{.*#+}} xmm12 = xmm12[2,3],xmm5[2,0]4324; SSE-NEXT:    unpckhps {{.*#+}} xmm15 = xmm15[2],xmm4[2],xmm15[3],xmm4[3]4325; SSE-NEXT:    movaps %xmm2, %xmm44326; SSE-NEXT:    unpckhpd {{.*#+}} xmm4 = xmm4[1],xmm0[1]4327; SSE-NEXT:    movaps %xmm15, %xmm54328; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[0,1],xmm4[0,2]4329; SSE-NEXT:    movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4330; SSE-NEXT:    unpckhps {{.*#+}} xmm13 = xmm13[2],xmm6[2],xmm13[3],xmm6[3]4331; SSE-NEXT:    movaps %xmm1, %xmm44332; SSE-NEXT:    unpckhpd {{.*#+}} xmm4 = xmm4[1],xmm3[1]4333; SSE-NEXT:    movaps %xmm13, %xmm54334; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[0,1],xmm4[0,2]4335; SSE-NEXT:    movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4336; SSE-NEXT:    shufps {{.*#+}} xmm2 = xmm2[3,3],xmm0[3,3]4337; SSE-NEXT:    shufps {{.*#+}} xmm15 = xmm15[2,3],xmm2[0,2]4338; SSE-NEXT:    shufps {{.*#+}} xmm1 = xmm1[3,3],xmm3[3,3]4339; SSE-NEXT:    shufps {{.*#+}} xmm13 = xmm13[2,3],xmm1[0,2]4340; SSE-NEXT:    movaps 112(%rdx), %xmm14341; SSE-NEXT:    movaps 112(%rcx), %xmm84342; SSE-NEXT:    movaps %xmm8, %xmm04343; SSE-NEXT:    movlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0]4344; SSE-NEXT:    movaps 112(%rdi), %xmm24345; SSE-NEXT:    movaps 112(%rsi), %xmm114346; SSE-NEXT:    movaps %xmm2, %xmm34347; SSE-NEXT:    unpcklps {{.*#+}} xmm3 = xmm3[0],xmm11[0],xmm3[1],xmm11[1]4348; SSE-NEXT:    movaps %xmm3, %xmm144349; SSE-NEXT:    shufps {{.*#+}} xmm14 = xmm14[0,1],xmm0[2,0]4350; SSE-NEXT:    movaps 112(%r10), %xmm04351; SSE-NEXT:    movaps 112(%rax), %xmm74352; SSE-NEXT:    movaps %xmm7, %xmm54353; SSE-NEXT:    movlhps {{.*#+}} xmm5 = xmm5[0],xmm0[0]4354; SSE-NEXT:    movaps 112(%r8), %xmm44355; SSE-NEXT:    movaps 112(%r9), %xmm94356; SSE-NEXT:    movaps %xmm4, %xmm64357; SSE-NEXT:    unpcklps {{.*#+}} xmm6 = xmm6[0],xmm9[0],xmm6[1],xmm9[1]4358; SSE-NEXT:    movaps %xmm6, %xmm104359; SSE-NEXT:    shufps {{.*#+}} xmm10 = xmm10[0,1],xmm5[2,0]4360; SSE-NEXT:    movaps %xmm8, %xmm54361; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[1,1],xmm1[1,1]4362; SSE-NEXT:    shufps {{.*#+}} xmm3 = xmm3[2,3],xmm5[2,0]4363; SSE-NEXT:    movaps %xmm7, %xmm54364; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[1,1],xmm0[1,1]4365; SSE-NEXT:    shufps {{.*#+}} xmm6 = xmm6[2,3],xmm5[2,0]4366; SSE-NEXT:    unpckhps {{.*#+}} xmm2 = xmm2[2],xmm11[2],xmm2[3],xmm11[3]4367; SSE-NEXT:    movaps %xmm1, %xmm54368; SSE-NEXT:    unpckhpd {{.*#+}} xmm5 = xmm5[1],xmm8[1]4369; SSE-NEXT:    movaps %xmm2, %xmm114370; SSE-NEXT:    shufps {{.*#+}} xmm11 = xmm11[0,1],xmm5[0,2]4371; SSE-NEXT:    unpckhps {{.*#+}} xmm4 = xmm4[2],xmm9[2],xmm4[3],xmm9[3]4372; SSE-NEXT:    movaps %xmm0, %xmm54373; SSE-NEXT:    unpckhpd {{.*#+}} xmm5 = xmm5[1],xmm7[1]4374; SSE-NEXT:    movaps %xmm4, %xmm94375; SSE-NEXT:    shufps {{.*#+}} xmm9 = xmm9[0,1],xmm5[0,2]4376; SSE-NEXT:    shufps {{.*#+}} xmm1 = xmm1[3,3],xmm8[3,3]4377; SSE-NEXT:    shufps {{.*#+}} xmm2 = xmm2[2,3],xmm1[0,2]4378; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[3,3],xmm7[3,3]4379; SSE-NEXT:    shufps {{.*#+}} xmm4 = xmm4[2,3],xmm0[0,2]4380; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %rax4381; SSE-NEXT:    movaps %xmm4, 1008(%rax)4382; SSE-NEXT:    movaps %xmm2, 992(%rax)4383; SSE-NEXT:    movaps %xmm9, 976(%rax)4384; SSE-NEXT:    movaps %xmm11, 960(%rax)4385; SSE-NEXT:    movaps %xmm6, 944(%rax)4386; SSE-NEXT:    movaps %xmm3, 928(%rax)4387; SSE-NEXT:    movaps %xmm10, 912(%rax)4388; SSE-NEXT:    movaps %xmm14, 896(%rax)4389; SSE-NEXT:    movaps %xmm13, 880(%rax)4390; SSE-NEXT:    movaps %xmm15, 864(%rax)4391; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4392; SSE-NEXT:    movaps %xmm0, 848(%rax)4393; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4394; SSE-NEXT:    movaps %xmm0, 832(%rax)4395; SSE-NEXT:    movaps %xmm12, 816(%rax)4396; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4397; SSE-NEXT:    movaps %xmm0, 800(%rax)4398; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4399; SSE-NEXT:    movaps %xmm0, 784(%rax)4400; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4401; SSE-NEXT:    movaps %xmm0, 768(%rax)4402; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4403; SSE-NEXT:    movaps %xmm0, 752(%rax)4404; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4405; SSE-NEXT:    movaps %xmm0, 736(%rax)4406; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4407; SSE-NEXT:    movaps %xmm0, 720(%rax)4408; SSE-NEXT:    movaps (%rsp), %xmm0 # 16-byte Reload4409; SSE-NEXT:    movaps %xmm0, 704(%rax)4410; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4411; SSE-NEXT:    movaps %xmm0, 688(%rax)4412; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4413; SSE-NEXT:    movaps %xmm0, 672(%rax)4414; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4415; SSE-NEXT:    movaps %xmm0, 656(%rax)4416; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4417; SSE-NEXT:    movaps %xmm0, 640(%rax)4418; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4419; SSE-NEXT:    movaps %xmm0, 624(%rax)4420; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4421; SSE-NEXT:    movaps %xmm0, 608(%rax)4422; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4423; SSE-NEXT:    movaps %xmm0, 592(%rax)4424; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4425; SSE-NEXT:    movaps %xmm0, 576(%rax)4426; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4427; SSE-NEXT:    movaps %xmm0, 560(%rax)4428; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4429; SSE-NEXT:    movaps %xmm0, 544(%rax)4430; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4431; SSE-NEXT:    movaps %xmm0, 528(%rax)4432; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4433; SSE-NEXT:    movaps %xmm0, 512(%rax)4434; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4435; SSE-NEXT:    movaps %xmm0, 496(%rax)4436; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4437; SSE-NEXT:    movaps %xmm0, 480(%rax)4438; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4439; SSE-NEXT:    movaps %xmm0, 464(%rax)4440; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4441; SSE-NEXT:    movaps %xmm0, 448(%rax)4442; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4443; SSE-NEXT:    movaps %xmm0, 432(%rax)4444; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4445; SSE-NEXT:    movaps %xmm0, 416(%rax)4446; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4447; SSE-NEXT:    movaps %xmm0, 400(%rax)4448; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4449; SSE-NEXT:    movaps %xmm0, 384(%rax)4450; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4451; SSE-NEXT:    movaps %xmm0, 368(%rax)4452; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4453; SSE-NEXT:    movaps %xmm0, 352(%rax)4454; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4455; SSE-NEXT:    movaps %xmm0, 336(%rax)4456; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4457; SSE-NEXT:    movaps %xmm0, 320(%rax)4458; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4459; SSE-NEXT:    movaps %xmm0, 304(%rax)4460; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4461; SSE-NEXT:    movaps %xmm0, 288(%rax)4462; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4463; SSE-NEXT:    movaps %xmm0, 272(%rax)4464; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4465; SSE-NEXT:    movaps %xmm0, 256(%rax)4466; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4467; SSE-NEXT:    movaps %xmm0, 240(%rax)4468; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4469; SSE-NEXT:    movaps %xmm0, 224(%rax)4470; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4471; SSE-NEXT:    movaps %xmm0, 208(%rax)4472; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4473; SSE-NEXT:    movaps %xmm0, 192(%rax)4474; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4475; SSE-NEXT:    movaps %xmm0, 176(%rax)4476; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4477; SSE-NEXT:    movaps %xmm0, 160(%rax)4478; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4479; SSE-NEXT:    movaps %xmm0, 144(%rax)4480; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4481; SSE-NEXT:    movaps %xmm0, 128(%rax)4482; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4483; SSE-NEXT:    movaps %xmm0, 112(%rax)4484; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4485; SSE-NEXT:    movaps %xmm0, 96(%rax)4486; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4487; SSE-NEXT:    movaps %xmm0, 80(%rax)4488; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4489; SSE-NEXT:    movaps %xmm0, 64(%rax)4490; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4491; SSE-NEXT:    movaps %xmm0, 48(%rax)4492; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4493; SSE-NEXT:    movaps %xmm0, 32(%rax)4494; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4495; SSE-NEXT:    movaps %xmm0, 16(%rax)4496; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload4497; SSE-NEXT:    movaps %xmm0, (%rax)4498; SSE-NEXT:    addq $728, %rsp # imm = 0x2D84499; SSE-NEXT:    retq4500;4501; AVX-LABEL: store_i32_stride8_vf32:4502; AVX:       # %bb.0:4503; AVX-NEXT:    subq $680, %rsp # imm = 0x2A84504; AVX-NEXT:    movq {{[0-9]+}}(%rsp), %rax4505; AVX-NEXT:    movq {{[0-9]+}}(%rsp), %r104506; AVX-NEXT:    vmovaps (%rdi), %ymm04507; AVX-NEXT:    vmovaps (%rsi), %ymm14508; AVX-NEXT:    vmovaps (%rdx), %ymm24509; AVX-NEXT:    vmovaps (%rcx), %ymm34510; AVX-NEXT:    vmovaps (%rcx), %xmm54511; AVX-NEXT:    vmovaps (%rdx), %xmm74512; AVX-NEXT:    vunpcklps {{.*#+}} xmm4 = xmm7[0],xmm5[0],xmm7[1],xmm5[1]4513; AVX-NEXT:    vmovaps (%rsi), %xmm94514; AVX-NEXT:    vmovaps (%rdi), %xmm124515; AVX-NEXT:    vmovaps (%r9), %xmm114516; AVX-NEXT:    vmovaps (%r8), %xmm134517; AVX-NEXT:    vunpcklps {{.*#+}} xmm6 = xmm13[0],xmm11[0],xmm13[1],xmm11[1]4518; AVX-NEXT:    vinsertf128 $1, %xmm6, %ymm0, %ymm64519; AVX-NEXT:    vmovaps (%rax), %xmm84520; AVX-NEXT:    vmovaps (%r10), %xmm104521; AVX-NEXT:    vunpcklps {{.*#+}} xmm14 = xmm10[0],xmm8[0],xmm10[1],xmm8[1]4522; AVX-NEXT:    vshufps {{.*#+}} xmm15 = xmm14[0,1,0,1]4523; AVX-NEXT:    vinsertf128 $1, %xmm15, %ymm0, %ymm154524; AVX-NEXT:    vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3,4,5],ymm15[6,7]4525; AVX-NEXT:    vunpcklps {{.*#+}} xmm15 = xmm12[0],xmm9[0],xmm12[1],xmm9[1]4526; AVX-NEXT:    vmovlhps {{.*#+}} xmm15 = xmm15[0],xmm4[0]4527; AVX-NEXT:    vblendps {{.*#+}} ymm6 = ymm15[0,1,2,3],ymm6[4,5,6,7]4528; AVX-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4529; AVX-NEXT:    vshufps {{.*#+}} xmm6 = xmm12[1,1,1,1]4530; AVX-NEXT:    vblendps {{.*#+}} xmm6 = xmm6[0],xmm9[1],xmm6[2,3]4531; AVX-NEXT:    vblendps {{.*#+}} xmm6 = xmm6[0,1],xmm4[2,3]4532; AVX-NEXT:    vinsertf128 $1, %xmm14, %ymm0, %ymm44533; AVX-NEXT:    vshufps {{.*#+}} xmm14 = xmm13[1,1,1,1]4534; AVX-NEXT:    vblendps {{.*#+}} xmm14 = xmm14[0],xmm11[1],xmm14[2,3]4535; AVX-NEXT:    vinsertf128 $1, %xmm14, %ymm0, %ymm144536; AVX-NEXT:    vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm4[6,7]4537; AVX-NEXT:    vmovaps (%r8), %ymm44538; AVX-NEXT:    vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm14[4,5,6,7]4539; AVX-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4540; AVX-NEXT:    vmovaps (%r9), %ymm64541; AVX-NEXT:    vunpckhps {{.*#+}} xmm12 = xmm12[2],xmm9[2],xmm12[3],xmm9[3]4542; AVX-NEXT:    vunpckhps {{.*#+}} xmm11 = xmm13[2],xmm11[2],xmm13[3],xmm11[3]4543; AVX-NEXT:    vinsertf128 $1, %xmm11, %ymm0, %ymm94544; AVX-NEXT:    vshufps {{.*#+}} xmm13 = xmm8[2,2,2,2]4545; AVX-NEXT:    vblendps {{.*#+}} xmm13 = xmm10[0,1,2],xmm13[3]4546; AVX-NEXT:    vinsertf128 $1, %xmm13, %ymm0, %ymm134547; AVX-NEXT:    vblendps {{.*#+}} ymm9 = ymm9[0,1,2,3,4,5],ymm13[6,7]4548; AVX-NEXT:    vshufps {{.*#+}} xmm13 = xmm5[2,2,2,2]4549; AVX-NEXT:    vblendps {{.*#+}} xmm13 = xmm7[0,1,2],xmm13[3]4550; AVX-NEXT:    vblendps {{.*#+}} xmm13 = xmm12[0,1],xmm13[2,3]4551; AVX-NEXT:    vblendps {{.*#+}} ymm9 = ymm13[0,1,2,3],ymm9[4,5,6,7]4552; AVX-NEXT:    vmovups %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4553; AVX-NEXT:    vmovaps (%r10), %ymm94554; AVX-NEXT:    vunpckhps {{.*#+}} xmm7 = xmm7[2],xmm5[2],xmm7[3],xmm5[3]4555; AVX-NEXT:    vmovaps (%rax), %ymm54556; AVX-NEXT:    vunpckhpd {{.*#+}} xmm7 = xmm12[1],xmm7[1]4557; AVX-NEXT:    vunpckhps {{.*#+}} xmm8 = xmm10[2],xmm8[2],xmm10[3],xmm8[3]4558; AVX-NEXT:    vinsertf128 $1, %xmm8, %ymm0, %ymm84559; AVX-NEXT:    vshufps {{.*#+}} xmm10 = xmm11[2,3,2,3]4560; AVX-NEXT:    vinsertf128 $1, %xmm10, %ymm0, %ymm104561; AVX-NEXT:    vblendps {{.*#+}} ymm8 = ymm10[0,1,2,3,4,5],ymm8[6,7]4562; AVX-NEXT:    vblendps {{.*#+}} ymm7 = ymm7[0,1,2,3],ymm8[4,5,6,7]4563; AVX-NEXT:    vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4564; AVX-NEXT:    vunpcklpd {{.*#+}} ymm7 = ymm5[0],ymm9[0],ymm5[2],ymm9[2]4565; AVX-NEXT:    vunpcklps {{.*#+}} ymm8 = ymm4[0],ymm6[0],ymm4[1],ymm6[1],ymm4[4],ymm6[4],ymm4[5],ymm6[5]4566; AVX-NEXT:    vshufps {{.*#+}} ymm7 = ymm8[0,1],ymm7[2,0],ymm8[4,5],ymm7[6,4]4567; AVX-NEXT:    vunpcklpd {{.*#+}} ymm8 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]4568; AVX-NEXT:    vextractf128 $1, %ymm8, %xmm84569; AVX-NEXT:    vunpcklps {{.*#+}} ymm10 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]4570; AVX-NEXT:    vextractf128 $1, %ymm10, %xmm104571; AVX-NEXT:    vshufps {{.*#+}} xmm8 = xmm10[0,1],xmm8[2,0]4572; AVX-NEXT:    vblendps {{.*#+}} ymm7 = ymm8[0,1,2,3],ymm7[4,5,6,7]4573; AVX-NEXT:    vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4574; AVX-NEXT:    vunpcklps {{.*#+}} ymm7 = ymm9[0],ymm5[0],ymm9[1],ymm5[1],ymm9[4],ymm5[4],ymm9[5],ymm5[5]4575; AVX-NEXT:    vshufps {{.*#+}} ymm8 = ymm6[1,0],ymm4[1,0],ymm6[5,4],ymm4[5,4]4576; AVX-NEXT:    vshufps {{.*#+}} ymm7 = ymm8[2,0],ymm7[2,3],ymm8[6,4],ymm7[6,7]4577; AVX-NEXT:    vunpcklps {{.*#+}} ymm8 = ymm2[0],ymm3[0],ymm2[1],ymm3[1],ymm2[4],ymm3[4],ymm2[5],ymm3[5]4578; AVX-NEXT:    vextractf128 $1, %ymm8, %xmm84579; AVX-NEXT:    vshufps {{.*#+}} ymm10 = ymm1[1,0],ymm0[1,0],ymm1[5,4],ymm0[5,4]4580; AVX-NEXT:    vextractf128 $1, %ymm10, %xmm104581; AVX-NEXT:    vshufps {{.*#+}} xmm8 = xmm10[2,0],xmm8[2,3]4582; AVX-NEXT:    vblendps {{.*#+}} ymm7 = ymm8[0,1,2,3],ymm7[4,5,6,7]4583; AVX-NEXT:    vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4584; AVX-NEXT:    vunpckhpd {{.*#+}} ymm7 = ymm5[1],ymm9[1],ymm5[3],ymm9[3]4585; AVX-NEXT:    vunpckhps {{.*#+}} ymm8 = ymm4[2],ymm6[2],ymm4[3],ymm6[3],ymm4[6],ymm6[6],ymm4[7],ymm6[7]4586; AVX-NEXT:    vshufps {{.*#+}} ymm7 = ymm8[0,1],ymm7[2,0],ymm8[4,5],ymm7[6,4]4587; AVX-NEXT:    vunpckhpd {{.*#+}} ymm8 = ymm3[1],ymm2[1],ymm3[3],ymm2[3]4588; AVX-NEXT:    vextractf128 $1, %ymm8, %xmm84589; AVX-NEXT:    vunpckhps {{.*#+}} ymm10 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7]4590; AVX-NEXT:    vextractf128 $1, %ymm10, %xmm104591; AVX-NEXT:    vshufps {{.*#+}} xmm8 = xmm10[0,1],xmm8[2,0]4592; AVX-NEXT:    vblendps {{.*#+}} ymm7 = ymm8[0,1,2,3],ymm7[4,5,6,7]4593; AVX-NEXT:    vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4594; AVX-NEXT:    vmovaps 32(%rcx), %xmm74595; AVX-NEXT:    vunpckhps {{.*#+}} ymm8 = ymm9[2],ymm5[2],ymm9[3],ymm5[3],ymm9[6],ymm5[6],ymm9[7],ymm5[7]4596; AVX-NEXT:    vmovaps 32(%rax), %xmm54597; AVX-NEXT:    vshufps {{.*#+}} ymm6 = ymm6[3,0],ymm4[3,0],ymm6[7,4],ymm4[7,4]4598; AVX-NEXT:    vmovaps 32(%r10), %xmm44599; AVX-NEXT:    vshufps {{.*#+}} ymm6 = ymm6[2,0],ymm8[2,3],ymm6[6,4],ymm8[6,7]4600; AVX-NEXT:    vunpckhps {{.*#+}} ymm2 = ymm2[2],ymm3[2],ymm2[3],ymm3[3],ymm2[6],ymm3[6],ymm2[7],ymm3[7]4601; AVX-NEXT:    vextractf128 $1, %ymm2, %xmm24602; AVX-NEXT:    vshufps {{.*#+}} ymm0 = ymm1[3,0],ymm0[3,0],ymm1[7,4],ymm0[7,4]4603; AVX-NEXT:    vextractf128 $1, %ymm0, %xmm04604; AVX-NEXT:    vshufps {{.*#+}} xmm0 = xmm0[2,0],xmm2[2,3]4605; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm6[4,5,6,7]4606; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4607; AVX-NEXT:    vmovaps 32(%r9), %xmm14608; AVX-NEXT:    vmovaps 32(%r8), %xmm24609; AVX-NEXT:    vunpcklps {{.*#+}} xmm0 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]4610; AVX-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm04611; AVX-NEXT:    vunpcklps {{.*#+}} xmm3 = xmm4[0],xmm5[0],xmm4[1],xmm5[1]4612; AVX-NEXT:    vshufps {{.*#+}} xmm6 = xmm3[0,1,0,1]4613; AVX-NEXT:    vinsertf128 $1, %xmm6, %ymm0, %ymm64614; AVX-NEXT:    vblendps {{.*#+}} ymm6 = ymm0[0,1,2,3,4,5],ymm6[6,7]4615; AVX-NEXT:    vmovaps 32(%rdx), %xmm04616; AVX-NEXT:    vunpcklps {{.*#+}} xmm8 = xmm0[0],xmm7[0],xmm0[1],xmm7[1]4617; AVX-NEXT:    vmovaps 32(%rsi), %xmm94618; AVX-NEXT:    vmovaps 32(%rdi), %xmm104619; AVX-NEXT:    vunpcklps {{.*#+}} xmm11 = xmm10[0],xmm9[0],xmm10[1],xmm9[1]4620; AVX-NEXT:    vmovlhps {{.*#+}} xmm11 = xmm11[0],xmm8[0]4621; AVX-NEXT:    vblendps {{.*#+}} ymm6 = ymm11[0,1,2,3],ymm6[4,5,6,7]4622; AVX-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4623; AVX-NEXT:    vshufps {{.*#+}} xmm6 = xmm10[1,1,1,1]4624; AVX-NEXT:    vblendps {{.*#+}} xmm6 = xmm6[0],xmm9[1],xmm6[2,3]4625; AVX-NEXT:    vblendps {{.*#+}} xmm6 = xmm6[0,1],xmm8[2,3]4626; AVX-NEXT:    vinsertf128 $1, %xmm3, %ymm0, %ymm34627; AVX-NEXT:    vshufps {{.*#+}} xmm8 = xmm2[1,1,1,1]4628; AVX-NEXT:    vblendps {{.*#+}} xmm8 = xmm8[0],xmm1[1],xmm8[2,3]4629; AVX-NEXT:    vinsertf128 $1, %xmm8, %ymm0, %ymm84630; AVX-NEXT:    vblendps {{.*#+}} ymm3 = ymm8[0,1,2,3,4,5],ymm3[6,7]4631; AVX-NEXT:    vblendps {{.*#+}} ymm3 = ymm6[0,1,2,3],ymm3[4,5,6,7]4632; AVX-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4633; AVX-NEXT:    vunpckhps {{.*#+}} xmm3 = xmm10[2],xmm9[2],xmm10[3],xmm9[3]4634; AVX-NEXT:    vunpckhps {{.*#+}} xmm2 = xmm2[2],xmm1[2],xmm2[3],xmm1[3]4635; AVX-NEXT:    vshufps {{.*#+}} xmm1 = xmm5[2,2,2,2]4636; AVX-NEXT:    vblendps {{.*#+}} xmm1 = xmm4[0,1,2],xmm1[3]4637; AVX-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm14638; AVX-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm64639; AVX-NEXT:    vblendps {{.*#+}} ymm1 = ymm6[0,1,2,3,4,5],ymm1[6,7]4640; AVX-NEXT:    vshufps {{.*#+}} xmm6 = xmm7[2,2,2,2]4641; AVX-NEXT:    vblendps {{.*#+}} xmm6 = xmm0[0,1,2],xmm6[3]4642; AVX-NEXT:    vblendps {{.*#+}} xmm6 = xmm3[0,1],xmm6[2,3]4643; AVX-NEXT:    vblendps {{.*#+}} ymm1 = ymm6[0,1,2,3],ymm1[4,5,6,7]4644; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4645; AVX-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm0[2],xmm7[2],xmm0[3],xmm7[3]4646; AVX-NEXT:    vunpckhpd {{.*#+}} xmm3 = xmm3[1],xmm0[1]4647; AVX-NEXT:    vmovaps 32(%r8), %ymm04648; AVX-NEXT:    vunpckhps {{.*#+}} xmm4 = xmm4[2],xmm5[2],xmm4[3],xmm5[3]4649; AVX-NEXT:    vmovaps 32(%r9), %ymm14650; AVX-NEXT:    vinsertf128 $1, %xmm4, %ymm0, %ymm44651; AVX-NEXT:    vshufps {{.*#+}} xmm2 = xmm2[2,3,2,3]4652; AVX-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm24653; AVX-NEXT:    vblendps {{.*#+}} ymm4 = ymm2[0,1,2,3,4,5],ymm4[6,7]4654; AVX-NEXT:    vmovaps 32(%r10), %ymm24655; AVX-NEXT:    vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm4[4,5,6,7]4656; AVX-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4657; AVX-NEXT:    vmovaps 32(%rax), %ymm34658; AVX-NEXT:    vunpcklpd {{.*#+}} ymm4 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]4659; AVX-NEXT:    vunpcklps {{.*#+}} ymm5 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]4660; AVX-NEXT:    vshufps {{.*#+}} ymm6 = ymm5[0,1],ymm4[2,0],ymm5[4,5],ymm4[6,4]4661; AVX-NEXT:    vmovaps 32(%rdx), %ymm44662; AVX-NEXT:    vmovaps 32(%rcx), %ymm54663; AVX-NEXT:    vunpcklpd {{.*#+}} ymm7 = ymm5[0],ymm4[0],ymm5[2],ymm4[2]4664; AVX-NEXT:    vextractf128 $1, %ymm7, %xmm74665; AVX-NEXT:    vmovaps 32(%rdi), %ymm84666; AVX-NEXT:    vmovaps 32(%rsi), %ymm94667; AVX-NEXT:    vunpcklps {{.*#+}} ymm10 = ymm8[0],ymm9[0],ymm8[1],ymm9[1],ymm8[4],ymm9[4],ymm8[5],ymm9[5]4668; AVX-NEXT:    vextractf128 $1, %ymm10, %xmm104669; AVX-NEXT:    vshufps {{.*#+}} xmm7 = xmm10[0,1],xmm7[2,0]4670; AVX-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]4671; AVX-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4672; AVX-NEXT:    vunpcklps {{.*#+}} ymm6 = ymm2[0],ymm3[0],ymm2[1],ymm3[1],ymm2[4],ymm3[4],ymm2[5],ymm3[5]4673; AVX-NEXT:    vshufps {{.*#+}} ymm7 = ymm1[1,0],ymm0[1,0],ymm1[5,4],ymm0[5,4]4674; AVX-NEXT:    vshufps {{.*#+}} ymm6 = ymm7[2,0],ymm6[2,3],ymm7[6,4],ymm6[6,7]4675; AVX-NEXT:    vunpcklps {{.*#+}} ymm7 = ymm4[0],ymm5[0],ymm4[1],ymm5[1],ymm4[4],ymm5[4],ymm4[5],ymm5[5]4676; AVX-NEXT:    vextractf128 $1, %ymm7, %xmm74677; AVX-NEXT:    vshufps {{.*#+}} ymm10 = ymm9[1,0],ymm8[1,0],ymm9[5,4],ymm8[5,4]4678; AVX-NEXT:    vextractf128 $1, %ymm10, %xmm104679; AVX-NEXT:    vshufps {{.*#+}} xmm7 = xmm10[2,0],xmm7[2,3]4680; AVX-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]4681; AVX-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4682; AVX-NEXT:    vunpckhpd {{.*#+}} ymm6 = ymm3[1],ymm2[1],ymm3[3],ymm2[3]4683; AVX-NEXT:    vunpckhps {{.*#+}} ymm7 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7]4684; AVX-NEXT:    vshufps {{.*#+}} ymm6 = ymm7[0,1],ymm6[2,0],ymm7[4,5],ymm6[6,4]4685; AVX-NEXT:    vunpckhpd {{.*#+}} ymm7 = ymm5[1],ymm4[1],ymm5[3],ymm4[3]4686; AVX-NEXT:    vextractf128 $1, %ymm7, %xmm74687; AVX-NEXT:    vunpckhps {{.*#+}} ymm10 = ymm8[2],ymm9[2],ymm8[3],ymm9[3],ymm8[6],ymm9[6],ymm8[7],ymm9[7]4688; AVX-NEXT:    vextractf128 $1, %ymm10, %xmm104689; AVX-NEXT:    vshufps {{.*#+}} xmm7 = xmm10[0,1],xmm7[2,0]4690; AVX-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]4691; AVX-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4692; AVX-NEXT:    vunpckhps {{.*#+}} ymm2 = ymm2[2],ymm3[2],ymm2[3],ymm3[3],ymm2[6],ymm3[6],ymm2[7],ymm3[7]4693; AVX-NEXT:    vshufps {{.*#+}} ymm0 = ymm1[3,0],ymm0[3,0],ymm1[7,4],ymm0[7,4]4694; AVX-NEXT:    vshufps {{.*#+}} ymm0 = ymm0[2,0],ymm2[2,3],ymm0[6,4],ymm2[6,7]4695; AVX-NEXT:    vunpckhps {{.*#+}} ymm1 = ymm4[2],ymm5[2],ymm4[3],ymm5[3],ymm4[6],ymm5[6],ymm4[7],ymm5[7]4696; AVX-NEXT:    vshufps {{.*#+}} ymm2 = ymm9[3,0],ymm8[3,0],ymm9[7,4],ymm8[7,4]4697; AVX-NEXT:    vextractf128 $1, %ymm1, %xmm14698; AVX-NEXT:    vextractf128 $1, %ymm2, %xmm24699; AVX-NEXT:    vshufps {{.*#+}} xmm1 = xmm2[2,0],xmm1[2,3]4700; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]4701; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4702; AVX-NEXT:    vmovaps 64(%r9), %xmm44703; AVX-NEXT:    vmovaps 64(%r8), %xmm54704; AVX-NEXT:    vunpcklps {{.*#+}} xmm0 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]4705; AVX-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm24706; AVX-NEXT:    vmovaps 64(%rax), %xmm04707; AVX-NEXT:    vmovaps 64(%r10), %xmm14708; AVX-NEXT:    vunpcklps {{.*#+}} xmm6 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]4709; AVX-NEXT:    vshufps {{.*#+}} xmm3 = xmm6[0,1,0,1]4710; AVX-NEXT:    vinsertf128 $1, %xmm3, %ymm0, %ymm34711; AVX-NEXT:    vblendps {{.*#+}} ymm7 = ymm2[0,1,2,3,4,5],ymm3[6,7]4712; AVX-NEXT:    vmovaps 64(%rcx), %xmm24713; AVX-NEXT:    vmovaps 64(%rdx), %xmm34714; AVX-NEXT:    vunpcklps {{.*#+}} xmm8 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]4715; AVX-NEXT:    vmovaps 64(%rsi), %xmm94716; AVX-NEXT:    vmovaps 64(%rdi), %xmm104717; AVX-NEXT:    vunpcklps {{.*#+}} xmm11 = xmm10[0],xmm9[0],xmm10[1],xmm9[1]4718; AVX-NEXT:    vmovlhps {{.*#+}} xmm11 = xmm11[0],xmm8[0]4719; AVX-NEXT:    vblendps {{.*#+}} ymm7 = ymm11[0,1,2,3],ymm7[4,5,6,7]4720; AVX-NEXT:    vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4721; AVX-NEXT:    vshufps {{.*#+}} xmm7 = xmm10[1,1,1,1]4722; AVX-NEXT:    vblendps {{.*#+}} xmm7 = xmm7[0],xmm9[1],xmm7[2,3]4723; AVX-NEXT:    vblendps {{.*#+}} xmm7 = xmm7[0,1],xmm8[2,3]4724; AVX-NEXT:    vinsertf128 $1, %xmm6, %ymm0, %ymm64725; AVX-NEXT:    vshufps {{.*#+}} xmm8 = xmm5[1,1,1,1]4726; AVX-NEXT:    vblendps {{.*#+}} xmm8 = xmm8[0],xmm4[1],xmm8[2,3]4727; AVX-NEXT:    vinsertf128 $1, %xmm8, %ymm0, %ymm84728; AVX-NEXT:    vblendps {{.*#+}} ymm6 = ymm8[0,1,2,3,4,5],ymm6[6,7]4729; AVX-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]4730; AVX-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4731; AVX-NEXT:    vunpckhps {{.*#+}} xmm6 = xmm10[2],xmm9[2],xmm10[3],xmm9[3]4732; AVX-NEXT:    vunpckhps {{.*#+}} xmm4 = xmm5[2],xmm4[2],xmm5[3],xmm4[3]4733; AVX-NEXT:    vshufps {{.*#+}} xmm5 = xmm0[2,2,2,2]4734; AVX-NEXT:    vblendps {{.*#+}} xmm5 = xmm1[0,1,2],xmm5[3]4735; AVX-NEXT:    vinsertf128 $1, %xmm5, %ymm0, %ymm54736; AVX-NEXT:    vinsertf128 $1, %xmm4, %ymm0, %ymm74737; AVX-NEXT:    vblendps {{.*#+}} ymm5 = ymm7[0,1,2,3,4,5],ymm5[6,7]4738; AVX-NEXT:    vshufps {{.*#+}} xmm7 = xmm2[2,2,2,2]4739; AVX-NEXT:    vblendps {{.*#+}} xmm7 = xmm3[0,1,2],xmm7[3]4740; AVX-NEXT:    vblendps {{.*#+}} xmm7 = xmm6[0,1],xmm7[2,3]4741; AVX-NEXT:    vblendps {{.*#+}} ymm5 = ymm7[0,1,2,3],ymm5[4,5,6,7]4742; AVX-NEXT:    vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4743; AVX-NEXT:    vunpckhps {{.*#+}} xmm2 = xmm3[2],xmm2[2],xmm3[3],xmm2[3]4744; AVX-NEXT:    vunpckhpd {{.*#+}} xmm2 = xmm6[1],xmm2[1]4745; AVX-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm1[2],xmm0[2],xmm1[3],xmm0[3]4746; AVX-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm04747; AVX-NEXT:    vshufps {{.*#+}} xmm1 = xmm4[2,3,2,3]4748; AVX-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm14749; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]4750; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]4751; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4752; AVX-NEXT:    vmovaps 64(%r8), %ymm04753; AVX-NEXT:    vmovaps 64(%r9), %ymm14754; AVX-NEXT:    vmovaps 64(%r10), %ymm24755; AVX-NEXT:    vmovaps 64(%rax), %ymm34756; AVX-NEXT:    vunpcklpd {{.*#+}} ymm4 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]4757; AVX-NEXT:    vunpcklps {{.*#+}} ymm5 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]4758; AVX-NEXT:    vshufps {{.*#+}} ymm6 = ymm5[0,1],ymm4[2,0],ymm5[4,5],ymm4[6,4]4759; AVX-NEXT:    vmovaps 64(%rdx), %ymm44760; AVX-NEXT:    vmovaps 64(%rcx), %ymm54761; AVX-NEXT:    vunpcklpd {{.*#+}} ymm7 = ymm5[0],ymm4[0],ymm5[2],ymm4[2]4762; AVX-NEXT:    vextractf128 $1, %ymm7, %xmm74763; AVX-NEXT:    vmovaps 64(%rdi), %ymm84764; AVX-NEXT:    vmovaps 64(%rsi), %ymm94765; AVX-NEXT:    vunpcklps {{.*#+}} ymm10 = ymm8[0],ymm9[0],ymm8[1],ymm9[1],ymm8[4],ymm9[4],ymm8[5],ymm9[5]4766; AVX-NEXT:    vextractf128 $1, %ymm10, %xmm104767; AVX-NEXT:    vshufps {{.*#+}} xmm7 = xmm10[0,1],xmm7[2,0]4768; AVX-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]4769; AVX-NEXT:    vmovups %ymm6, (%rsp) # 32-byte Spill4770; AVX-NEXT:    vunpcklps {{.*#+}} ymm6 = ymm2[0],ymm3[0],ymm2[1],ymm3[1],ymm2[4],ymm3[4],ymm2[5],ymm3[5]4771; AVX-NEXT:    vshufps {{.*#+}} ymm7 = ymm1[1,0],ymm0[1,0],ymm1[5,4],ymm0[5,4]4772; AVX-NEXT:    vshufps {{.*#+}} ymm6 = ymm7[2,0],ymm6[2,3],ymm7[6,4],ymm6[6,7]4773; AVX-NEXT:    vunpcklps {{.*#+}} ymm7 = ymm4[0],ymm5[0],ymm4[1],ymm5[1],ymm4[4],ymm5[4],ymm4[5],ymm5[5]4774; AVX-NEXT:    vextractf128 $1, %ymm7, %xmm74775; AVX-NEXT:    vshufps {{.*#+}} ymm10 = ymm9[1,0],ymm8[1,0],ymm9[5,4],ymm8[5,4]4776; AVX-NEXT:    vextractf128 $1, %ymm10, %xmm104777; AVX-NEXT:    vshufps {{.*#+}} xmm7 = xmm10[2,0],xmm7[2,3]4778; AVX-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]4779; AVX-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4780; AVX-NEXT:    vunpckhpd {{.*#+}} ymm6 = ymm3[1],ymm2[1],ymm3[3],ymm2[3]4781; AVX-NEXT:    vunpckhps {{.*#+}} ymm7 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7]4782; AVX-NEXT:    vshufps {{.*#+}} ymm6 = ymm7[0,1],ymm6[2,0],ymm7[4,5],ymm6[6,4]4783; AVX-NEXT:    vunpckhpd {{.*#+}} ymm7 = ymm5[1],ymm4[1],ymm5[3],ymm4[3]4784; AVX-NEXT:    vextractf128 $1, %ymm7, %xmm74785; AVX-NEXT:    vunpckhps {{.*#+}} ymm10 = ymm8[2],ymm9[2],ymm8[3],ymm9[3],ymm8[6],ymm9[6],ymm8[7],ymm9[7]4786; AVX-NEXT:    vextractf128 $1, %ymm10, %xmm104787; AVX-NEXT:    vshufps {{.*#+}} xmm7 = xmm10[0,1],xmm7[2,0]4788; AVX-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]4789; AVX-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4790; AVX-NEXT:    vunpckhps {{.*#+}} ymm2 = ymm2[2],ymm3[2],ymm2[3],ymm3[3],ymm2[6],ymm3[6],ymm2[7],ymm3[7]4791; AVX-NEXT:    vshufps {{.*#+}} ymm0 = ymm1[3,0],ymm0[3,0],ymm1[7,4],ymm0[7,4]4792; AVX-NEXT:    vshufps {{.*#+}} ymm0 = ymm0[2,0],ymm2[2,3],ymm0[6,4],ymm2[6,7]4793; AVX-NEXT:    vunpckhps {{.*#+}} ymm1 = ymm4[2],ymm5[2],ymm4[3],ymm5[3],ymm4[6],ymm5[6],ymm4[7],ymm5[7]4794; AVX-NEXT:    vshufps {{.*#+}} ymm2 = ymm9[3,0],ymm8[3,0],ymm9[7,4],ymm8[7,4]4795; AVX-NEXT:    vextractf128 $1, %ymm1, %xmm14796; AVX-NEXT:    vextractf128 $1, %ymm2, %xmm24797; AVX-NEXT:    vshufps {{.*#+}} xmm1 = xmm2[2,0],xmm1[2,3]4798; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]4799; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4800; AVX-NEXT:    vmovaps 96(%r9), %xmm44801; AVX-NEXT:    vmovaps 96(%r8), %xmm54802; AVX-NEXT:    vunpcklps {{.*#+}} xmm0 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]4803; AVX-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm24804; AVX-NEXT:    vmovaps 96(%rax), %xmm94805; AVX-NEXT:    vmovaps 96(%r10), %xmm14806; AVX-NEXT:    vunpcklps {{.*#+}} xmm6 = xmm1[0],xmm9[0],xmm1[1],xmm9[1]4807; AVX-NEXT:    vshufps {{.*#+}} xmm3 = xmm6[0,1,0,1]4808; AVX-NEXT:    vinsertf128 $1, %xmm3, %ymm0, %ymm34809; AVX-NEXT:    vblendps {{.*#+}} ymm10 = ymm2[0,1,2,3,4,5],ymm3[6,7]4810; AVX-NEXT:    vmovaps 96(%rcx), %xmm24811; AVX-NEXT:    vmovaps 96(%rdx), %xmm34812; AVX-NEXT:    vunpcklps {{.*#+}} xmm8 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]4813; AVX-NEXT:    vmovaps 96(%rsi), %xmm04814; AVX-NEXT:    vmovaps 96(%rdi), %xmm74815; AVX-NEXT:    vunpcklps {{.*#+}} xmm11 = xmm7[0],xmm0[0],xmm7[1],xmm0[1]4816; AVX-NEXT:    vmovlhps {{.*#+}} xmm11 = xmm11[0],xmm8[0]4817; AVX-NEXT:    vblendps {{.*#+}} ymm10 = ymm11[0,1,2,3],ymm10[4,5,6,7]4818; AVX-NEXT:    vmovups %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4819; AVX-NEXT:    vshufps {{.*#+}} xmm10 = xmm7[1,1,1,1]4820; AVX-NEXT:    vblendps {{.*#+}} xmm10 = xmm10[0],xmm0[1],xmm10[2,3]4821; AVX-NEXT:    vblendps {{.*#+}} xmm8 = xmm10[0,1],xmm8[2,3]4822; AVX-NEXT:    vinsertf128 $1, %xmm6, %ymm0, %ymm64823; AVX-NEXT:    vshufps {{.*#+}} xmm10 = xmm5[1,1,1,1]4824; AVX-NEXT:    vblendps {{.*#+}} xmm10 = xmm10[0],xmm4[1],xmm10[2,3]4825; AVX-NEXT:    vinsertf128 $1, %xmm10, %ymm0, %ymm104826; AVX-NEXT:    vblendps {{.*#+}} ymm6 = ymm10[0,1,2,3,4,5],ymm6[6,7]4827; AVX-NEXT:    vblendps {{.*#+}} ymm12 = ymm8[0,1,2,3],ymm6[4,5,6,7]4828; AVX-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm7[2],xmm0[2],xmm7[3],xmm0[3]4829; AVX-NEXT:    vunpckhps {{.*#+}} xmm4 = xmm5[2],xmm4[2],xmm5[3],xmm4[3]4830; AVX-NEXT:    vshufps {{.*#+}} xmm5 = xmm9[2,2,2,2]4831; AVX-NEXT:    vblendps {{.*#+}} xmm5 = xmm1[0,1,2],xmm5[3]4832; AVX-NEXT:    vinsertf128 $1, %xmm5, %ymm0, %ymm54833; AVX-NEXT:    vinsertf128 $1, %xmm4, %ymm0, %ymm64834; AVX-NEXT:    vblendps {{.*#+}} ymm5 = ymm6[0,1,2,3,4,5],ymm5[6,7]4835; AVX-NEXT:    vshufps {{.*#+}} xmm6 = xmm2[2,2,2,2]4836; AVX-NEXT:    vblendps {{.*#+}} xmm6 = xmm3[0,1,2],xmm6[3]4837; AVX-NEXT:    vblendps {{.*#+}} xmm6 = xmm0[0,1],xmm6[2,3]4838; AVX-NEXT:    vblendps {{.*#+}} ymm11 = ymm6[0,1,2,3],ymm5[4,5,6,7]4839; AVX-NEXT:    vunpckhps {{.*#+}} xmm2 = xmm3[2],xmm2[2],xmm3[3],xmm2[3]4840; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm2[1]4841; AVX-NEXT:    vunpckhps {{.*#+}} xmm1 = xmm1[2],xmm9[2],xmm1[3],xmm9[3]4842; AVX-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm14843; AVX-NEXT:    vshufps {{.*#+}} xmm2 = xmm4[2,3,2,3]4844; AVX-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm24845; AVX-NEXT:    vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]4846; AVX-NEXT:    vblendps {{.*#+}} ymm10 = ymm0[0,1,2,3],ymm1[4,5,6,7]4847; AVX-NEXT:    vmovaps 96(%r8), %ymm74848; AVX-NEXT:    vmovaps 96(%r9), %ymm64849; AVX-NEXT:    vmovaps 96(%r10), %ymm54850; AVX-NEXT:    vmovaps 96(%rax), %ymm44851; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm4[0],ymm5[0],ymm4[2],ymm5[2]4852; AVX-NEXT:    vunpcklps {{.*#+}} ymm1 = ymm7[0],ymm6[0],ymm7[1],ymm6[1],ymm7[4],ymm6[4],ymm7[5],ymm6[5]4853; AVX-NEXT:    vshufps {{.*#+}} ymm8 = ymm1[0,1],ymm0[2,0],ymm1[4,5],ymm0[6,4]4854; AVX-NEXT:    vmovaps 96(%rdx), %ymm34855; AVX-NEXT:    vmovaps 96(%rcx), %ymm24856; AVX-NEXT:    vunpcklpd {{.*#+}} ymm9 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]4857; AVX-NEXT:    vextractf128 $1, %ymm9, %xmm94858; AVX-NEXT:    vmovaps 96(%rdi), %ymm14859; AVX-NEXT:    vmovaps 96(%rsi), %ymm04860; AVX-NEXT:    vunpcklps {{.*#+}} ymm15 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[4],ymm0[4],ymm1[5],ymm0[5]4861; AVX-NEXT:    vextractf128 $1, %ymm15, %xmm154862; AVX-NEXT:    vshufps {{.*#+}} xmm9 = xmm15[0,1],xmm9[2,0]4863; AVX-NEXT:    vblendps {{.*#+}} ymm9 = ymm9[0,1,2,3],ymm8[4,5,6,7]4864; AVX-NEXT:    vunpcklps {{.*#+}} ymm8 = ymm5[0],ymm4[0],ymm5[1],ymm4[1],ymm5[4],ymm4[4],ymm5[5],ymm4[5]4865; AVX-NEXT:    vshufps {{.*#+}} ymm15 = ymm6[1,0],ymm7[1,0],ymm6[5,4],ymm7[5,4]4866; AVX-NEXT:    vshufps {{.*#+}} ymm8 = ymm15[2,0],ymm8[2,3],ymm15[6,4],ymm8[6,7]4867; AVX-NEXT:    vunpcklps {{.*#+}} ymm15 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[4],ymm2[4],ymm3[5],ymm2[5]4868; AVX-NEXT:    vextractf128 $1, %ymm15, %xmm154869; AVX-NEXT:    vshufps {{.*#+}} ymm14 = ymm0[1,0],ymm1[1,0],ymm0[5,4],ymm1[5,4]4870; AVX-NEXT:    vextractf128 $1, %ymm14, %xmm144871; AVX-NEXT:    vshufps {{.*#+}} xmm14 = xmm14[2,0],xmm15[2,3]4872; AVX-NEXT:    vblendps {{.*#+}} ymm8 = ymm14[0,1,2,3],ymm8[4,5,6,7]4873; AVX-NEXT:    vunpckhpd {{.*#+}} ymm14 = ymm4[1],ymm5[1],ymm4[3],ymm5[3]4874; AVX-NEXT:    vunpckhps {{.*#+}} ymm15 = ymm7[2],ymm6[2],ymm7[3],ymm6[3],ymm7[6],ymm6[6],ymm7[7],ymm6[7]4875; AVX-NEXT:    vshufps {{.*#+}} ymm14 = ymm15[0,1],ymm14[2,0],ymm15[4,5],ymm14[6,4]4876; AVX-NEXT:    vunpckhpd {{.*#+}} ymm15 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]4877; AVX-NEXT:    vextractf128 $1, %ymm15, %xmm154878; AVX-NEXT:    vunpckhps {{.*#+}} ymm13 = ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[6],ymm0[6],ymm1[7],ymm0[7]4879; AVX-NEXT:    vextractf128 $1, %ymm13, %xmm134880; AVX-NEXT:    vshufps {{.*#+}} xmm13 = xmm13[0,1],xmm15[2,0]4881; AVX-NEXT:    vblendps {{.*#+}} ymm13 = ymm13[0,1,2,3],ymm14[4,5,6,7]4882; AVX-NEXT:    vunpckhps {{.*#+}} ymm4 = ymm5[2],ymm4[2],ymm5[3],ymm4[3],ymm5[6],ymm4[6],ymm5[7],ymm4[7]4883; AVX-NEXT:    vshufps {{.*#+}} ymm5 = ymm6[3,0],ymm7[3,0],ymm6[7,4],ymm7[7,4]4884; AVX-NEXT:    vshufps {{.*#+}} ymm4 = ymm5[2,0],ymm4[2,3],ymm5[6,4],ymm4[6,7]4885; AVX-NEXT:    vunpckhps {{.*#+}} ymm2 = ymm3[2],ymm2[2],ymm3[3],ymm2[3],ymm3[6],ymm2[6],ymm3[7],ymm2[7]4886; AVX-NEXT:    vshufps {{.*#+}} ymm0 = ymm0[3,0],ymm1[3,0],ymm0[7,4],ymm1[7,4]4887; AVX-NEXT:    vextractf128 $1, %ymm2, %xmm14888; AVX-NEXT:    vextractf128 $1, %ymm0, %xmm04889; AVX-NEXT:    vshufps {{.*#+}} xmm0 = xmm0[2,0],xmm1[2,3]4890; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm4[4,5,6,7]4891; AVX-NEXT:    movq {{[0-9]+}}(%rsp), %rax4892; AVX-NEXT:    vmovaps %ymm0, 992(%rax)4893; AVX-NEXT:    vmovaps %ymm13, 960(%rax)4894; AVX-NEXT:    vmovaps %ymm8, 928(%rax)4895; AVX-NEXT:    vmovaps %ymm9, 896(%rax)4896; AVX-NEXT:    vmovaps %ymm10, 864(%rax)4897; AVX-NEXT:    vmovaps %ymm11, 832(%rax)4898; AVX-NEXT:    vmovaps %ymm12, 800(%rax)4899; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload4900; AVX-NEXT:    vmovaps %ymm0, 768(%rax)4901; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload4902; AVX-NEXT:    vmovaps %ymm0, 736(%rax)4903; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload4904; AVX-NEXT:    vmovaps %ymm0, 704(%rax)4905; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload4906; AVX-NEXT:    vmovaps %ymm0, 672(%rax)4907; AVX-NEXT:    vmovups (%rsp), %ymm0 # 32-byte Reload4908; AVX-NEXT:    vmovaps %ymm0, 640(%rax)4909; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload4910; AVX-NEXT:    vmovaps %ymm0, 608(%rax)4911; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload4912; AVX-NEXT:    vmovaps %ymm0, 576(%rax)4913; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload4914; AVX-NEXT:    vmovaps %ymm0, 544(%rax)4915; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload4916; AVX-NEXT:    vmovaps %ymm0, 512(%rax)4917; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload4918; AVX-NEXT:    vmovaps %ymm0, 480(%rax)4919; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload4920; AVX-NEXT:    vmovaps %ymm0, 448(%rax)4921; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload4922; AVX-NEXT:    vmovaps %ymm0, 416(%rax)4923; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload4924; AVX-NEXT:    vmovaps %ymm0, 384(%rax)4925; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload4926; AVX-NEXT:    vmovaps %ymm0, 352(%rax)4927; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload4928; AVX-NEXT:    vmovaps %ymm0, 320(%rax)4929; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload4930; AVX-NEXT:    vmovaps %ymm0, 288(%rax)4931; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload4932; AVX-NEXT:    vmovaps %ymm0, 256(%rax)4933; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload4934; AVX-NEXT:    vmovaps %ymm0, 224(%rax)4935; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload4936; AVX-NEXT:    vmovaps %ymm0, 192(%rax)4937; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload4938; AVX-NEXT:    vmovaps %ymm0, 160(%rax)4939; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload4940; AVX-NEXT:    vmovaps %ymm0, 128(%rax)4941; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload4942; AVX-NEXT:    vmovaps %ymm0, 96(%rax)4943; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload4944; AVX-NEXT:    vmovaps %ymm0, 64(%rax)4945; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload4946; AVX-NEXT:    vmovaps %ymm0, 32(%rax)4947; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload4948; AVX-NEXT:    vmovaps %ymm0, (%rax)4949; AVX-NEXT:    addq $680, %rsp # imm = 0x2A84950; AVX-NEXT:    vzeroupper4951; AVX-NEXT:    retq4952;4953; AVX2-LABEL: store_i32_stride8_vf32:4954; AVX2:       # %bb.0:4955; AVX2-NEXT:    subq $648, %rsp # imm = 0x2884956; AVX2-NEXT:    movq {{[0-9]+}}(%rsp), %r104957; AVX2-NEXT:    movq {{[0-9]+}}(%rsp), %rax4958; AVX2-NEXT:    vmovaps 32(%rdi), %ymm04959; AVX2-NEXT:    vmovaps (%rdi), %ymm44960; AVX2-NEXT:    vmovaps 32(%rsi), %ymm14961; AVX2-NEXT:    vmovaps (%rsi), %ymm54962; AVX2-NEXT:    vmovaps 32(%rdx), %ymm24963; AVX2-NEXT:    vmovaps (%rdx), %ymm74964; AVX2-NEXT:    vmovaps 32(%rcx), %ymm34965; AVX2-NEXT:    vmovaps (%rcx), %ymm94966; AVX2-NEXT:    vmovaps (%r8), %ymm104967; AVX2-NEXT:    vmovaps (%r9), %ymm124968; AVX2-NEXT:    vmovaps (%rax), %ymm84969; AVX2-NEXT:    vmovaps (%r10), %ymm114970; AVX2-NEXT:    vunpcklps {{.*#+}} ymm6 = ymm8[0],ymm11[0],ymm8[1],ymm11[1],ymm8[4],ymm11[4],ymm8[5],ymm11[5]4971; AVX2-NEXT:    vunpcklps {{.*#+}} ymm13 = ymm4[0],ymm5[0],ymm4[1],ymm5[1],ymm4[4],ymm5[4],ymm4[5],ymm5[5]4972; AVX2-NEXT:    vextractf128 $1, %ymm13, %xmm134973; AVX2-NEXT:    vunpcklps {{.*#+}} ymm14 = ymm7[0],ymm9[0],ymm7[1],ymm9[1],ymm7[4],ymm9[4],ymm7[5],ymm9[5]4974; AVX2-NEXT:    vpermpd {{.*#+}} ymm15 = ymm14[2,2,2,2]4975; AVX2-NEXT:    vblendps {{.*#+}} xmm13 = xmm13[0,1],xmm15[2,3]4976; AVX2-NEXT:    vunpcklps {{.*#+}} ymm15 = ymm10[0],ymm12[0],ymm10[1],ymm12[1],ymm10[4],ymm12[4],ymm10[5],ymm12[5]4977; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm15 = ymm15[0],ymm6[0],ymm15[2],ymm6[2]4978; AVX2-NEXT:    vblendps {{.*#+}} ymm13 = ymm13[0,1,2,3],ymm15[4,5,6,7]4979; AVX2-NEXT:    vmovups %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4980; AVX2-NEXT:    vbroadcastss 20(%r8), %ymm134981; AVX2-NEXT:    vblendps {{.*#+}} ymm13 = ymm13[0,1,2,3,4],ymm12[5],ymm13[6,7]4982; AVX2-NEXT:    vblendps {{.*#+}} ymm13 = ymm13[0,1,2,3,4,5],ymm6[6,7]4983; AVX2-NEXT:    vextractf128 $1, %ymm14, %xmm64984; AVX2-NEXT:    vshufps {{.*#+}} ymm14 = ymm4[1,1,1,1,5,5,5,5]4985; AVX2-NEXT:    vblendps {{.*#+}} ymm14 = ymm14[0],ymm5[1],ymm14[2,3,4],ymm5[5],ymm14[6,7]4986; AVX2-NEXT:    vextractf128 $1, %ymm14, %xmm144987; AVX2-NEXT:    vblendps {{.*#+}} xmm14 = xmm14[0,1],xmm6[2,3]4988; AVX2-NEXT:    vmovaps 32(%r8), %ymm64989; AVX2-NEXT:    vblendps {{.*#+}} ymm13 = ymm14[0,1,2,3],ymm13[4,5,6,7]4990; AVX2-NEXT:    vmovups %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4991; AVX2-NEXT:    vbroadcastss 24(%r10), %ymm134992; AVX2-NEXT:    vblendps {{.*#+}} ymm13 = ymm8[0,1,2,3,4,5,6],ymm13[7]4993; AVX2-NEXT:    vmovaps 32(%r9), %ymm84994; AVX2-NEXT:    vunpckhps {{.*#+}} ymm12 = ymm10[2],ymm12[2],ymm10[3],ymm12[3],ymm10[6],ymm12[6],ymm10[7],ymm12[7]4995; AVX2-NEXT:    vmovaps 32(%rax), %ymm104996; AVX2-NEXT:    vunpckhps {{.*#+}} ymm4 = ymm4[2],ymm5[2],ymm4[3],ymm5[3],ymm4[6],ymm5[6],ymm4[7],ymm5[7]4997; AVX2-NEXT:    vextractf128 $1, %ymm4, %xmm54998; AVX2-NEXT:    vunpckhps {{.*#+}} ymm9 = ymm7[2],ymm9[2],ymm7[3],ymm9[3],ymm7[6],ymm9[6],ymm7[7],ymm9[7]4999; AVX2-NEXT:    vpermpd {{.*#+}} ymm7 = ymm9[2,2,2,2]5000; AVX2-NEXT:    vblendps {{.*#+}} xmm5 = xmm5[0,1],xmm7[2,3]5001; AVX2-NEXT:    vmovaps 32(%r10), %ymm75002; AVX2-NEXT:    vblendps {{.*#+}} ymm13 = ymm12[0,1,2,3,4,5],ymm13[6,7]5003; AVX2-NEXT:    vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm13[4,5,6,7]5004; AVX2-NEXT:    vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5005; AVX2-NEXT:    vbroadcastss 28(%rax), %ymm55006; AVX2-NEXT:    vunpckhps {{.*#+}} ymm5 = ymm5[2],ymm11[2],ymm5[3],ymm11[3],ymm5[6],ymm11[6],ymm5[7],ymm11[7]5007; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm5 = ymm12[1],ymm5[1],ymm12[3],ymm5[3]5008; AVX2-NEXT:    vextractf128 $1, %ymm9, %xmm95009; AVX2-NEXT:    vshufps {{.*#+}} ymm4 = ymm4[2,3,2,3,6,7,6,7]5010; AVX2-NEXT:    vextractf128 $1, %ymm4, %xmm45011; AVX2-NEXT:    vblendps {{.*#+}} xmm4 = xmm4[0,1],xmm9[2,3]5012; AVX2-NEXT:    vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm5[4,5,6,7]5013; AVX2-NEXT:    vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5014; AVX2-NEXT:    vunpcklps {{.*#+}} ymm4 = ymm10[0],ymm7[0],ymm10[1],ymm7[1],ymm10[4],ymm7[4],ymm10[5],ymm7[5]5015; AVX2-NEXT:    vunpcklps {{.*#+}} ymm5 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]5016; AVX2-NEXT:    vextractf128 $1, %ymm5, %xmm55017; AVX2-NEXT:    vunpcklps {{.*#+}} ymm9 = ymm2[0],ymm3[0],ymm2[1],ymm3[1],ymm2[4],ymm3[4],ymm2[5],ymm3[5]5018; AVX2-NEXT:    vpermpd {{.*#+}} ymm11 = ymm9[2,2,2,2]5019; AVX2-NEXT:    vblendps {{.*#+}} xmm5 = xmm5[0,1],xmm11[2,3]5020; AVX2-NEXT:    vunpcklps {{.*#+}} ymm11 = ymm6[0],ymm8[0],ymm6[1],ymm8[1],ymm6[4],ymm8[4],ymm6[5],ymm8[5]5021; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm11 = ymm11[0],ymm4[0],ymm11[2],ymm4[2]5022; AVX2-NEXT:    vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm11[4,5,6,7]5023; AVX2-NEXT:    vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5024; AVX2-NEXT:    vbroadcastss 52(%r8), %ymm55025; AVX2-NEXT:    vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3,4],ymm8[5],ymm5[6,7]5026; AVX2-NEXT:    vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3,4,5],ymm4[6,7]5027; AVX2-NEXT:    vextractf128 $1, %ymm9, %xmm45028; AVX2-NEXT:    vshufps {{.*#+}} ymm9 = ymm0[1,1,1,1,5,5,5,5]5029; AVX2-NEXT:    vblendps {{.*#+}} ymm9 = ymm9[0],ymm1[1],ymm9[2,3,4],ymm1[5],ymm9[6,7]5030; AVX2-NEXT:    vextractf128 $1, %ymm9, %xmm95031; AVX2-NEXT:    vblendps {{.*#+}} xmm9 = xmm9[0,1],xmm4[2,3]5032; AVX2-NEXT:    vmovaps 64(%rdi), %ymm45033; AVX2-NEXT:    vblendps {{.*#+}} ymm5 = ymm9[0,1,2,3],ymm5[4,5,6,7]5034; AVX2-NEXT:    vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5035; AVX2-NEXT:    vbroadcastss 56(%r10), %ymm55036; AVX2-NEXT:    vblendps {{.*#+}} ymm9 = ymm10[0,1,2,3,4,5,6],ymm5[7]5037; AVX2-NEXT:    vmovaps 64(%rsi), %ymm55038; AVX2-NEXT:    vunpckhps {{.*#+}} ymm8 = ymm6[2],ymm8[2],ymm6[3],ymm8[3],ymm6[6],ymm8[6],ymm6[7],ymm8[7]5039; AVX2-NEXT:    vmovaps 64(%rdx), %ymm65040; AVX2-NEXT:    vunpckhps {{.*#+}} ymm10 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7]5041; AVX2-NEXT:    vextractf128 $1, %ymm10, %xmm05042; AVX2-NEXT:    vunpckhps {{.*#+}} ymm2 = ymm2[2],ymm3[2],ymm2[3],ymm3[3],ymm2[6],ymm3[6],ymm2[7],ymm3[7]5043; AVX2-NEXT:    vpermpd {{.*#+}} ymm1 = ymm2[2,2,2,2]5044; AVX2-NEXT:    vblendps {{.*#+}} xmm1 = xmm0[0,1],xmm1[2,3]5045; AVX2-NEXT:    vmovaps 64(%rcx), %ymm05046; AVX2-NEXT:    vblendps {{.*#+}} ymm3 = ymm8[0,1,2,3,4,5],ymm9[6,7]5047; AVX2-NEXT:    vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm3[4,5,6,7]5048; AVX2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5049; AVX2-NEXT:    vbroadcastss 60(%rax), %ymm15050; AVX2-NEXT:    vunpckhps {{.*#+}} ymm3 = ymm1[2],ymm7[2],ymm1[3],ymm7[3],ymm1[6],ymm7[6],ymm1[7],ymm7[7]5051; AVX2-NEXT:    vmovaps 64(%r8), %ymm15052; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm3 = ymm8[1],ymm3[1],ymm8[3],ymm3[3]5053; AVX2-NEXT:    vextractf128 $1, %ymm2, %xmm25054; AVX2-NEXT:    vshufps {{.*#+}} ymm7 = ymm10[2,3,2,3,6,7,6,7]5055; AVX2-NEXT:    vextractf128 $1, %ymm7, %xmm75056; AVX2-NEXT:    vblendps {{.*#+}} xmm2 = xmm7[0,1],xmm2[2,3]5057; AVX2-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]5058; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5059; AVX2-NEXT:    vunpcklps {{.*#+}} ymm2 = ymm4[0],ymm5[0],ymm4[1],ymm5[1],ymm4[4],ymm5[4],ymm4[5],ymm5[5]5060; AVX2-NEXT:    vextractf128 $1, %ymm2, %xmm25061; AVX2-NEXT:    vunpcklps {{.*#+}} ymm3 = ymm6[0],ymm0[0],ymm6[1],ymm0[1],ymm6[4],ymm0[4],ymm6[5],ymm0[5]5062; AVX2-NEXT:    vpermpd {{.*#+}} ymm7 = ymm3[2,2,2,2]5063; AVX2-NEXT:    vblendps {{.*#+}} xmm2 = xmm2[0,1],xmm7[2,3]5064; AVX2-NEXT:    vmovaps 64(%r9), %ymm75065; AVX2-NEXT:    vmovaps 64(%rax), %ymm85066; AVX2-NEXT:    vmovaps 64(%r10), %ymm95067; AVX2-NEXT:    vunpcklps {{.*#+}} ymm10 = ymm8[0],ymm9[0],ymm8[1],ymm9[1],ymm8[4],ymm9[4],ymm8[5],ymm9[5]5068; AVX2-NEXT:    vunpcklps {{.*#+}} ymm11 = ymm1[0],ymm7[0],ymm1[1],ymm7[1],ymm1[4],ymm7[4],ymm1[5],ymm7[5]5069; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm11 = ymm11[0],ymm10[0],ymm11[2],ymm10[2]5070; AVX2-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm11[4,5,6,7]5071; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5072; AVX2-NEXT:    vbroadcastss 84(%r8), %ymm25073; AVX2-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4],ymm7[5],ymm2[6,7]5074; AVX2-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm10[6,7]5075; AVX2-NEXT:    vextractf128 $1, %ymm3, %xmm35076; AVX2-NEXT:    vshufps {{.*#+}} ymm10 = ymm4[1,1,1,1,5,5,5,5]5077; AVX2-NEXT:    vblendps {{.*#+}} ymm10 = ymm10[0],ymm5[1],ymm10[2,3,4],ymm5[5],ymm10[6,7]5078; AVX2-NEXT:    vextractf128 $1, %ymm10, %xmm105079; AVX2-NEXT:    vblendps {{.*#+}} xmm3 = xmm10[0,1],xmm3[2,3]5080; AVX2-NEXT:    vblendps {{.*#+}} ymm2 = ymm3[0,1,2,3],ymm2[4,5,6,7]5081; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5082; AVX2-NEXT:    vbroadcastss 88(%r10), %ymm25083; AVX2-NEXT:    vblendps {{.*#+}} ymm2 = ymm8[0,1,2,3,4,5,6],ymm2[7]5084; AVX2-NEXT:    vunpckhps {{.*#+}} ymm1 = ymm1[2],ymm7[2],ymm1[3],ymm7[3],ymm1[6],ymm7[6],ymm1[7],ymm7[7]5085; AVX2-NEXT:    vunpckhps {{.*#+}} ymm3 = ymm4[2],ymm5[2],ymm4[3],ymm5[3],ymm4[6],ymm5[6],ymm4[7],ymm5[7]5086; AVX2-NEXT:    vunpckhps {{.*#+}} ymm4 = ymm6[2],ymm0[2],ymm6[3],ymm0[3],ymm6[6],ymm0[6],ymm6[7],ymm0[7]5087; AVX2-NEXT:    vextractf128 $1, %ymm3, %xmm05088; AVX2-NEXT:    vpermpd {{.*#+}} ymm5 = ymm4[2,2,2,2]5089; AVX2-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm5[2,3]5090; AVX2-NEXT:    vblendps {{.*#+}} ymm2 = ymm1[0,1,2,3,4,5],ymm2[6,7]5091; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]5092; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5093; AVX2-NEXT:    vbroadcastss 92(%rax), %ymm05094; AVX2-NEXT:    vunpckhps {{.*#+}} ymm2 = ymm0[2],ymm9[2],ymm0[3],ymm9[3],ymm0[6],ymm9[6],ymm0[7],ymm9[7]5095; AVX2-NEXT:    vmovaps 96(%rdi), %ymm05096; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm5 = ymm1[1],ymm2[1],ymm1[3],ymm2[3]5097; AVX2-NEXT:    vmovaps 96(%rsi), %ymm15098; AVX2-NEXT:    vextractf128 $1, %ymm4, %xmm25099; AVX2-NEXT:    vshufps {{.*#+}} ymm3 = ymm3[2,3,2,3,6,7,6,7]5100; AVX2-NEXT:    vextractf128 $1, %ymm3, %xmm35101; AVX2-NEXT:    vblendps {{.*#+}} xmm3 = xmm3[0,1],xmm2[2,3]5102; AVX2-NEXT:    vmovaps 96(%rdx), %ymm25103; AVX2-NEXT:    vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm5[4,5,6,7]5104; AVX2-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5105; AVX2-NEXT:    vunpcklps {{.*#+}} ymm3 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]5106; AVX2-NEXT:    vextractf128 $1, %ymm3, %xmm35107; AVX2-NEXT:    vmovaps 96(%rcx), %ymm45108; AVX2-NEXT:    vunpcklps {{.*#+}} ymm5 = ymm2[0],ymm4[0],ymm2[1],ymm4[1],ymm2[4],ymm4[4],ymm2[5],ymm4[5]5109; AVX2-NEXT:    vpermpd {{.*#+}} ymm6 = ymm5[2,2,2,2]5110; AVX2-NEXT:    vblendps {{.*#+}} xmm3 = xmm3[0,1],xmm6[2,3]5111; AVX2-NEXT:    vmovaps 96(%r8), %ymm65112; AVX2-NEXT:    vmovaps 96(%r9), %ymm75113; AVX2-NEXT:    vmovaps 96(%rax), %ymm85114; AVX2-NEXT:    vmovaps 96(%r10), %ymm95115; AVX2-NEXT:    vunpcklps {{.*#+}} ymm10 = ymm8[0],ymm9[0],ymm8[1],ymm9[1],ymm8[4],ymm9[4],ymm8[5],ymm9[5]5116; AVX2-NEXT:    vunpcklps {{.*#+}} ymm11 = ymm6[0],ymm7[0],ymm6[1],ymm7[1],ymm6[4],ymm7[4],ymm6[5],ymm7[5]5117; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm11 = ymm11[0],ymm10[0],ymm11[2],ymm10[2]5118; AVX2-NEXT:    vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm11[4,5,6,7]5119; AVX2-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5120; AVX2-NEXT:    vbroadcastss 116(%r8), %ymm35121; AVX2-NEXT:    vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4],ymm7[5],ymm3[6,7]5122; AVX2-NEXT:    vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm10[6,7]5123; AVX2-NEXT:    vextractf128 $1, %ymm5, %xmm55124; AVX2-NEXT:    vshufps {{.*#+}} ymm10 = ymm0[1,1,1,1,5,5,5,5]5125; AVX2-NEXT:    vblendps {{.*#+}} ymm10 = ymm10[0],ymm1[1],ymm10[2,3,4],ymm1[5],ymm10[6,7]5126; AVX2-NEXT:    vextractf128 $1, %ymm10, %xmm105127; AVX2-NEXT:    vblendps {{.*#+}} xmm5 = xmm10[0,1],xmm5[2,3]5128; AVX2-NEXT:    vblendps {{.*#+}} ymm3 = ymm5[0,1,2,3],ymm3[4,5,6,7]5129; AVX2-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5130; AVX2-NEXT:    vbroadcastss 120(%r10), %ymm35131; AVX2-NEXT:    vblendps {{.*#+}} ymm3 = ymm8[0,1,2,3,4,5,6],ymm3[7]5132; AVX2-NEXT:    vunpckhps {{.*#+}} ymm5 = ymm6[2],ymm7[2],ymm6[3],ymm7[3],ymm6[6],ymm7[6],ymm6[7],ymm7[7]5133; AVX2-NEXT:    vunpckhps {{.*#+}} ymm0 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7]5134; AVX2-NEXT:    vunpckhps {{.*#+}} ymm1 = ymm2[2],ymm4[2],ymm2[3],ymm4[3],ymm2[6],ymm4[6],ymm2[7],ymm4[7]5135; AVX2-NEXT:    vextractf128 $1, %ymm0, %xmm25136; AVX2-NEXT:    vpermpd {{.*#+}} ymm4 = ymm1[2,2,2,2]5137; AVX2-NEXT:    vblendps {{.*#+}} xmm2 = xmm2[0,1],xmm4[2,3]5138; AVX2-NEXT:    vblendps {{.*#+}} ymm3 = ymm5[0,1,2,3,4,5],ymm3[6,7]5139; AVX2-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]5140; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5141; AVX2-NEXT:    vbroadcastss 124(%rax), %ymm25142; AVX2-NEXT:    vunpckhps {{.*#+}} ymm2 = ymm2[2],ymm9[2],ymm2[3],ymm9[3],ymm2[6],ymm9[6],ymm2[7],ymm9[7]5143; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm5[1],ymm2[1],ymm5[3],ymm2[3]5144; AVX2-NEXT:    vextractf128 $1, %ymm1, %xmm15145; AVX2-NEXT:    vshufps {{.*#+}} ymm0 = ymm0[2,3,2,3,6,7,6,7]5146; AVX2-NEXT:    vextractf128 $1, %ymm0, %xmm05147; AVX2-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]5148; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]5149; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5150; AVX2-NEXT:    vmovaps (%rcx), %xmm05151; AVX2-NEXT:    vbroadcastss %xmm0, %xmm25152; AVX2-NEXT:    vmovaps (%rdx), %xmm15153; AVX2-NEXT:    vbroadcastss %xmm1, %xmm35154; AVX2-NEXT:    vunpcklps {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]5155; AVX2-NEXT:    vmovaps (%rsi), %xmm45156; AVX2-NEXT:    vmovaps (%rdi), %xmm55157; AVX2-NEXT:    vunpcklps {{.*#+}} xmm3 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]5158; AVX2-NEXT:    vblendps {{.*#+}} xmm6 = xmm3[0,1],xmm2[2,3]5159; AVX2-NEXT:    vmovaps (%r10), %xmm25160; AVX2-NEXT:    vbroadcastss %xmm2, %xmm75161; AVX2-NEXT:    vmovaps (%rax), %xmm35162; AVX2-NEXT:    vbroadcastss %xmm3, %xmm85163; AVX2-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm8[0],xmm7[0],xmm8[1],xmm7[1]5164; AVX2-NEXT:    vmovaps (%r9), %xmm85165; AVX2-NEXT:    vmovaps (%r8), %xmm95166; AVX2-NEXT:    vunpcklps {{.*#+}} xmm10 = xmm9[0],xmm8[0],xmm9[1],xmm8[1]5167; AVX2-NEXT:    vinsertf128 $1, %xmm10, %ymm0, %ymm105168; AVX2-NEXT:    vinsertf128 $1, %xmm7, %ymm0, %ymm75169; AVX2-NEXT:    vblendps {{.*#+}} ymm7 = ymm10[0,1,2,3,4,5],ymm7[6,7]5170; AVX2-NEXT:    vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7]5171; AVX2-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5172; AVX2-NEXT:    vshufps {{.*#+}} xmm6 = xmm5[1,1,1,1]5173; AVX2-NEXT:    vblendps {{.*#+}} xmm6 = xmm6[0],xmm4[1],xmm6[2,3]5174; AVX2-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]5175; AVX2-NEXT:    vblendps {{.*#+}} xmm6 = xmm6[0,1],xmm7[2,3]5176; AVX2-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]5177; AVX2-NEXT:    vinsertf128 $1, %xmm7, %ymm0, %ymm75178; AVX2-NEXT:    vshufps {{.*#+}} xmm10 = xmm9[1,1,1,1]5179; AVX2-NEXT:    vblendps {{.*#+}} xmm10 = xmm10[0],xmm8[1],xmm10[2,3]5180; AVX2-NEXT:    vinsertf128 $1, %xmm10, %ymm0, %ymm105181; AVX2-NEXT:    vblendps {{.*#+}} ymm7 = ymm10[0,1,2,3,4,5],ymm7[6,7]5182; AVX2-NEXT:    vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7]5183; AVX2-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5184; AVX2-NEXT:    vunpckhps {{.*#+}} xmm4 = xmm5[2],xmm4[2],xmm5[3],xmm4[3]5185; AVX2-NEXT:    vunpckhps {{.*#+}} xmm5 = xmm9[2],xmm8[2],xmm9[3],xmm8[3]5186; AVX2-NEXT:    vshufps {{.*#+}} xmm6 = xmm2[2,2,2,2]5187; AVX2-NEXT:    vblendps {{.*#+}} xmm6 = xmm3[0,1,2],xmm6[3]5188; AVX2-NEXT:    vinsertf128 $1, %xmm6, %ymm0, %ymm65189; AVX2-NEXT:    vinsertf128 $1, %xmm5, %ymm0, %ymm75190; AVX2-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3,4,5],ymm6[6,7]5191; AVX2-NEXT:    vshufps {{.*#+}} xmm7 = xmm0[2,2,2,2]5192; AVX2-NEXT:    vblendps {{.*#+}} xmm7 = xmm1[0,1,2],xmm7[3]5193; AVX2-NEXT:    vblendps {{.*#+}} xmm7 = xmm4[0,1],xmm7[2,3]5194; AVX2-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]5195; AVX2-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5196; AVX2-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm1[2],xmm0[2],xmm1[3],xmm0[3]5197; AVX2-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm4[1],xmm0[1]5198; AVX2-NEXT:    vunpckhps {{.*#+}} xmm1 = xmm3[2],xmm2[2],xmm3[3],xmm2[3]5199; AVX2-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm15200; AVX2-NEXT:    vshufps {{.*#+}} xmm2 = xmm5[2,3,2,3]5201; AVX2-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm25202; AVX2-NEXT:    vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]5203; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]5204; AVX2-NEXT:    vmovups %ymm0, (%rsp) # 32-byte Spill5205; AVX2-NEXT:    vmovaps 32(%rcx), %xmm05206; AVX2-NEXT:    vmovaps 32(%rdx), %xmm15207; AVX2-NEXT:    vbroadcastss %xmm0, %xmm25208; AVX2-NEXT:    vbroadcastss %xmm1, %xmm35209; AVX2-NEXT:    vunpcklps {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]5210; AVX2-NEXT:    vmovaps 32(%rsi), %xmm45211; AVX2-NEXT:    vmovaps 32(%rdi), %xmm55212; AVX2-NEXT:    vunpcklps {{.*#+}} xmm3 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]5213; AVX2-NEXT:    vblendps {{.*#+}} xmm6 = xmm3[0,1],xmm2[2,3]5214; AVX2-NEXT:    vmovaps 32(%r10), %xmm25215; AVX2-NEXT:    vmovaps 32(%rax), %xmm35216; AVX2-NEXT:    vbroadcastss %xmm2, %xmm75217; AVX2-NEXT:    vbroadcastss %xmm3, %xmm85218; AVX2-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm8[0],xmm7[0],xmm8[1],xmm7[1]5219; AVX2-NEXT:    vmovaps 32(%r9), %xmm85220; AVX2-NEXT:    vmovaps 32(%r8), %xmm95221; AVX2-NEXT:    vunpcklps {{.*#+}} xmm10 = xmm9[0],xmm8[0],xmm9[1],xmm8[1]5222; AVX2-NEXT:    vinsertf128 $1, %xmm10, %ymm0, %ymm105223; AVX2-NEXT:    vinsertf128 $1, %xmm7, %ymm0, %ymm75224; AVX2-NEXT:    vblendps {{.*#+}} ymm7 = ymm10[0,1,2,3,4,5],ymm7[6,7]5225; AVX2-NEXT:    vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7]5226; AVX2-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5227; AVX2-NEXT:    vshufps {{.*#+}} xmm6 = xmm5[1,1,1,1]5228; AVX2-NEXT:    vblendps {{.*#+}} xmm6 = xmm6[0],xmm4[1],xmm6[2,3]5229; AVX2-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]5230; AVX2-NEXT:    vblendps {{.*#+}} xmm6 = xmm6[0,1],xmm7[2,3]5231; AVX2-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]5232; AVX2-NEXT:    vinsertf128 $1, %xmm7, %ymm0, %ymm75233; AVX2-NEXT:    vshufps {{.*#+}} xmm10 = xmm9[1,1,1,1]5234; AVX2-NEXT:    vblendps {{.*#+}} xmm10 = xmm10[0],xmm8[1],xmm10[2,3]5235; AVX2-NEXT:    vinsertf128 $1, %xmm10, %ymm0, %ymm105236; AVX2-NEXT:    vblendps {{.*#+}} ymm7 = ymm10[0,1,2,3,4,5],ymm7[6,7]5237; AVX2-NEXT:    vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7]5238; AVX2-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5239; AVX2-NEXT:    vunpckhps {{.*#+}} xmm4 = xmm5[2],xmm4[2],xmm5[3],xmm4[3]5240; AVX2-NEXT:    vunpckhps {{.*#+}} xmm5 = xmm9[2],xmm8[2],xmm9[3],xmm8[3]5241; AVX2-NEXT:    vshufps {{.*#+}} xmm6 = xmm2[2,2,2,2]5242; AVX2-NEXT:    vblendps {{.*#+}} xmm6 = xmm3[0,1,2],xmm6[3]5243; AVX2-NEXT:    vinsertf128 $1, %xmm6, %ymm0, %ymm65244; AVX2-NEXT:    vinsertf128 $1, %xmm5, %ymm0, %ymm75245; AVX2-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3,4,5],ymm6[6,7]5246; AVX2-NEXT:    vshufps {{.*#+}} xmm7 = xmm0[2,2,2,2]5247; AVX2-NEXT:    vblendps {{.*#+}} xmm7 = xmm1[0,1,2],xmm7[3]5248; AVX2-NEXT:    vblendps {{.*#+}} xmm7 = xmm4[0,1],xmm7[2,3]5249; AVX2-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]5250; AVX2-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5251; AVX2-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm1[2],xmm0[2],xmm1[3],xmm0[3]5252; AVX2-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm4[1],xmm0[1]5253; AVX2-NEXT:    vunpckhps {{.*#+}} xmm1 = xmm3[2],xmm2[2],xmm3[3],xmm2[3]5254; AVX2-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm15255; AVX2-NEXT:    vshufps {{.*#+}} xmm2 = xmm5[2,3,2,3]5256; AVX2-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm25257; AVX2-NEXT:    vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]5258; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]5259; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5260; AVX2-NEXT:    vmovaps 64(%rcx), %xmm35261; AVX2-NEXT:    vmovaps 64(%rdx), %xmm15262; AVX2-NEXT:    vbroadcastss %xmm3, %xmm05263; AVX2-NEXT:    vbroadcastss %xmm1, %xmm25264; AVX2-NEXT:    vunpcklps {{.*#+}} xmm0 = xmm2[0],xmm0[0],xmm2[1],xmm0[1]5265; AVX2-NEXT:    vmovaps 64(%rsi), %xmm45266; AVX2-NEXT:    vmovaps 64(%rdi), %xmm55267; AVX2-NEXT:    vunpcklps {{.*#+}} xmm2 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]5268; AVX2-NEXT:    vblendps {{.*#+}} xmm9 = xmm2[0,1],xmm0[2,3]5269; AVX2-NEXT:    vmovaps 64(%r10), %xmm25270; AVX2-NEXT:    vmovaps 64(%rax), %xmm05271; AVX2-NEXT:    vbroadcastss %xmm2, %xmm75272; AVX2-NEXT:    vbroadcastss %xmm0, %xmm85273; AVX2-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm8[0],xmm7[0],xmm8[1],xmm7[1]5274; AVX2-NEXT:    vmovaps 64(%r9), %xmm85275; AVX2-NEXT:    vmovaps 64(%r8), %xmm65276; AVX2-NEXT:    vunpcklps {{.*#+}} xmm10 = xmm6[0],xmm8[0],xmm6[1],xmm8[1]5277; AVX2-NEXT:    vinsertf128 $1, %xmm10, %ymm0, %ymm105278; AVX2-NEXT:    vinsertf128 $1, %xmm7, %ymm0, %ymm75279; AVX2-NEXT:    vblendps {{.*#+}} ymm7 = ymm10[0,1,2,3,4,5],ymm7[6,7]5280; AVX2-NEXT:    vblendps {{.*#+}} ymm11 = ymm9[0,1,2,3],ymm7[4,5,6,7]5281; AVX2-NEXT:    vshufps {{.*#+}} xmm7 = xmm5[1,1,1,1]5282; AVX2-NEXT:    vblendps {{.*#+}} xmm7 = xmm7[0],xmm4[1],xmm7[2,3]5283; AVX2-NEXT:    vunpcklps {{.*#+}} xmm9 = xmm1[0],xmm3[0],xmm1[1],xmm3[1]5284; AVX2-NEXT:    vblendps {{.*#+}} xmm7 = xmm7[0,1],xmm9[2,3]5285; AVX2-NEXT:    vunpcklps {{.*#+}} xmm9 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]5286; AVX2-NEXT:    vinsertf128 $1, %xmm9, %ymm0, %ymm95287; AVX2-NEXT:    vshufps {{.*#+}} xmm15 = xmm6[1,1,1,1]5288; AVX2-NEXT:    vblendps {{.*#+}} xmm15 = xmm15[0],xmm8[1],xmm15[2,3]5289; AVX2-NEXT:    vinsertf128 $1, %xmm15, %ymm0, %ymm155290; AVX2-NEXT:    vblendps {{.*#+}} ymm9 = ymm15[0,1,2,3,4,5],ymm9[6,7]5291; AVX2-NEXT:    vblendps {{.*#+}} ymm10 = ymm7[0,1,2,3],ymm9[4,5,6,7]5292; AVX2-NEXT:    vunpckhps {{.*#+}} xmm4 = xmm5[2],xmm4[2],xmm5[3],xmm4[3]5293; AVX2-NEXT:    vunpckhps {{.*#+}} xmm5 = xmm6[2],xmm8[2],xmm6[3],xmm8[3]5294; AVX2-NEXT:    vshufps {{.*#+}} xmm6 = xmm2[2,2,2,2]5295; AVX2-NEXT:    vblendps {{.*#+}} xmm6 = xmm0[0,1,2],xmm6[3]5296; AVX2-NEXT:    vinsertf128 $1, %xmm6, %ymm0, %ymm65297; AVX2-NEXT:    vinsertf128 $1, %xmm5, %ymm0, %ymm75298; AVX2-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3,4,5],ymm6[6,7]5299; AVX2-NEXT:    vshufps {{.*#+}} xmm7 = xmm3[2,2,2,2]5300; AVX2-NEXT:    vblendps {{.*#+}} xmm7 = xmm1[0,1,2],xmm7[3]5301; AVX2-NEXT:    vblendps {{.*#+}} xmm7 = xmm4[0,1],xmm7[2,3]5302; AVX2-NEXT:    vblendps {{.*#+}} ymm9 = ymm7[0,1,2,3],ymm6[4,5,6,7]5303; AVX2-NEXT:    vunpckhps {{.*#+}} xmm1 = xmm1[2],xmm3[2],xmm1[3],xmm3[3]5304; AVX2-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm4[1],xmm1[1]5305; AVX2-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm0[2],xmm2[2],xmm0[3],xmm2[3]5306; AVX2-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm05307; AVX2-NEXT:    vshufps {{.*#+}} xmm2 = xmm5[2,3,2,3]5308; AVX2-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm25309; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3,4,5],ymm0[6,7]5310; AVX2-NEXT:    vblendps {{.*#+}} ymm8 = ymm1[0,1,2,3],ymm0[4,5,6,7]5311; AVX2-NEXT:    vmovaps 96(%rcx), %xmm65312; AVX2-NEXT:    vbroadcastss %xmm6, %xmm05313; AVX2-NEXT:    vmovaps 96(%rdx), %xmm55314; AVX2-NEXT:    vbroadcastss %xmm5, %xmm15315; AVX2-NEXT:    vunpcklps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]5316; AVX2-NEXT:    vmovaps 96(%rsi), %xmm45317; AVX2-NEXT:    vmovaps 96(%rdi), %xmm25318; AVX2-NEXT:    vunpcklps {{.*#+}} xmm3 = xmm2[0],xmm4[0],xmm2[1],xmm4[1]5319; AVX2-NEXT:    vblendps {{.*#+}} xmm7 = xmm3[0,1],xmm0[2,3]5320; AVX2-NEXT:    vmovaps 96(%r10), %xmm35321; AVX2-NEXT:    vbroadcastss %xmm3, %xmm155322; AVX2-NEXT:    vmovaps 96(%rax), %xmm15323; AVX2-NEXT:    vbroadcastss %xmm1, %xmm145324; AVX2-NEXT:    vunpcklps {{.*#+}} xmm14 = xmm14[0],xmm15[0],xmm14[1],xmm15[1]5325; AVX2-NEXT:    vmovaps 96(%r9), %xmm155326; AVX2-NEXT:    vmovaps 96(%r8), %xmm05327; AVX2-NEXT:    vunpcklps {{.*#+}} xmm13 = xmm0[0],xmm15[0],xmm0[1],xmm15[1]5328; AVX2-NEXT:    vinsertf128 $1, %xmm13, %ymm0, %ymm135329; AVX2-NEXT:    vinsertf128 $1, %xmm14, %ymm0, %ymm145330; AVX2-NEXT:    vblendps {{.*#+}} ymm13 = ymm13[0,1,2,3,4,5],ymm14[6,7]5331; AVX2-NEXT:    vblendps {{.*#+}} ymm14 = ymm7[0,1,2,3],ymm13[4,5,6,7]5332; AVX2-NEXT:    vshufps {{.*#+}} xmm7 = xmm2[1,1,1,1]5333; AVX2-NEXT:    vblendps {{.*#+}} xmm7 = xmm7[0],xmm4[1],xmm7[2,3]5334; AVX2-NEXT:    vunpcklps {{.*#+}} xmm13 = xmm5[0],xmm6[0],xmm5[1],xmm6[1]5335; AVX2-NEXT:    vblendps {{.*#+}} xmm7 = xmm7[0,1],xmm13[2,3]5336; AVX2-NEXT:    vunpcklps {{.*#+}} xmm13 = xmm1[0],xmm3[0],xmm1[1],xmm3[1]5337; AVX2-NEXT:    vinsertf128 $1, %xmm13, %ymm0, %ymm135338; AVX2-NEXT:    vshufps {{.*#+}} xmm12 = xmm0[1,1,1,1]5339; AVX2-NEXT:    vblendps {{.*#+}} xmm12 = xmm12[0],xmm15[1],xmm12[2,3]5340; AVX2-NEXT:    vinsertf128 $1, %xmm12, %ymm0, %ymm125341; AVX2-NEXT:    vblendps {{.*#+}} ymm12 = ymm12[0,1,2,3,4,5],ymm13[6,7]5342; AVX2-NEXT:    vblendps {{.*#+}} ymm7 = ymm7[0,1,2,3],ymm12[4,5,6,7]5343; AVX2-NEXT:    vunpckhps {{.*#+}} xmm2 = xmm2[2],xmm4[2],xmm2[3],xmm4[3]5344; AVX2-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm0[2],xmm15[2],xmm0[3],xmm15[3]5345; AVX2-NEXT:    vshufps {{.*#+}} xmm4 = xmm3[2,2,2,2]5346; AVX2-NEXT:    vblendps {{.*#+}} xmm4 = xmm1[0,1,2],xmm4[3]5347; AVX2-NEXT:    vinsertf128 $1, %xmm4, %ymm0, %ymm45348; AVX2-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm125349; AVX2-NEXT:    vblendps {{.*#+}} ymm4 = ymm12[0,1,2,3,4,5],ymm4[6,7]5350; AVX2-NEXT:    vshufps {{.*#+}} xmm12 = xmm6[2,2,2,2]5351; AVX2-NEXT:    vblendps {{.*#+}} xmm12 = xmm5[0,1,2],xmm12[3]5352; AVX2-NEXT:    vblendps {{.*#+}} xmm12 = xmm2[0,1],xmm12[2,3]5353; AVX2-NEXT:    vblendps {{.*#+}} ymm4 = ymm12[0,1,2,3],ymm4[4,5,6,7]5354; AVX2-NEXT:    vunpckhps {{.*#+}} xmm5 = xmm5[2],xmm6[2],xmm5[3],xmm6[3]5355; AVX2-NEXT:    vunpckhpd {{.*#+}} xmm2 = xmm2[1],xmm5[1]5356; AVX2-NEXT:    vunpckhps {{.*#+}} xmm1 = xmm1[2],xmm3[2],xmm1[3],xmm3[3]5357; AVX2-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm15358; AVX2-NEXT:    vshufps {{.*#+}} xmm0 = xmm0[2,3,2,3]5359; AVX2-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm05360; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]5361; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]5362; AVX2-NEXT:    movq {{[0-9]+}}(%rsp), %rax5363; AVX2-NEXT:    vmovaps %ymm0, 864(%rax)5364; AVX2-NEXT:    vmovaps %ymm4, 832(%rax)5365; AVX2-NEXT:    vmovaps %ymm7, 800(%rax)5366; AVX2-NEXT:    vmovaps %ymm14, 768(%rax)5367; AVX2-NEXT:    vmovaps %ymm8, 608(%rax)5368; AVX2-NEXT:    vmovaps %ymm9, 576(%rax)5369; AVX2-NEXT:    vmovaps %ymm10, 544(%rax)5370; AVX2-NEXT:    vmovaps %ymm11, 512(%rax)5371; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5372; AVX2-NEXT:    vmovaps %ymm0, 352(%rax)5373; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5374; AVX2-NEXT:    vmovaps %ymm0, 320(%rax)5375; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5376; AVX2-NEXT:    vmovaps %ymm0, 288(%rax)5377; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5378; AVX2-NEXT:    vmovaps %ymm0, 256(%rax)5379; AVX2-NEXT:    vmovups (%rsp), %ymm0 # 32-byte Reload5380; AVX2-NEXT:    vmovaps %ymm0, 96(%rax)5381; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5382; AVX2-NEXT:    vmovaps %ymm0, 64(%rax)5383; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5384; AVX2-NEXT:    vmovaps %ymm0, 32(%rax)5385; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5386; AVX2-NEXT:    vmovaps %ymm0, (%rax)5387; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5388; AVX2-NEXT:    vmovaps %ymm0, 992(%rax)5389; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5390; AVX2-NEXT:    vmovaps %ymm0, 960(%rax)5391; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5392; AVX2-NEXT:    vmovaps %ymm0, 928(%rax)5393; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5394; AVX2-NEXT:    vmovaps %ymm0, 896(%rax)5395; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5396; AVX2-NEXT:    vmovaps %ymm0, 736(%rax)5397; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5398; AVX2-NEXT:    vmovaps %ymm0, 704(%rax)5399; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5400; AVX2-NEXT:    vmovaps %ymm0, 672(%rax)5401; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5402; AVX2-NEXT:    vmovaps %ymm0, 640(%rax)5403; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5404; AVX2-NEXT:    vmovaps %ymm0, 480(%rax)5405; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5406; AVX2-NEXT:    vmovaps %ymm0, 448(%rax)5407; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5408; AVX2-NEXT:    vmovaps %ymm0, 416(%rax)5409; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5410; AVX2-NEXT:    vmovaps %ymm0, 384(%rax)5411; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5412; AVX2-NEXT:    vmovaps %ymm0, 224(%rax)5413; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5414; AVX2-NEXT:    vmovaps %ymm0, 192(%rax)5415; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5416; AVX2-NEXT:    vmovaps %ymm0, 160(%rax)5417; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5418; AVX2-NEXT:    vmovaps %ymm0, 128(%rax)5419; AVX2-NEXT:    addq $648, %rsp # imm = 0x2885420; AVX2-NEXT:    vzeroupper5421; AVX2-NEXT:    retq5422;5423; AVX2-FP-LABEL: store_i32_stride8_vf32:5424; AVX2-FP:       # %bb.0:5425; AVX2-FP-NEXT:    subq $648, %rsp # imm = 0x2885426; AVX2-FP-NEXT:    movq {{[0-9]+}}(%rsp), %r105427; AVX2-FP-NEXT:    movq {{[0-9]+}}(%rsp), %rax5428; AVX2-FP-NEXT:    vmovaps 32(%rdi), %ymm05429; AVX2-FP-NEXT:    vmovaps (%rdi), %ymm45430; AVX2-FP-NEXT:    vmovaps 32(%rsi), %ymm15431; AVX2-FP-NEXT:    vmovaps (%rsi), %ymm55432; AVX2-FP-NEXT:    vmovaps 32(%rdx), %ymm25433; AVX2-FP-NEXT:    vmovaps (%rdx), %ymm75434; AVX2-FP-NEXT:    vmovaps 32(%rcx), %ymm35435; AVX2-FP-NEXT:    vmovaps (%rcx), %ymm95436; AVX2-FP-NEXT:    vmovaps (%r8), %ymm105437; AVX2-FP-NEXT:    vmovaps (%r9), %ymm125438; AVX2-FP-NEXT:    vmovaps (%rax), %ymm85439; AVX2-FP-NEXT:    vmovaps (%r10), %ymm115440; AVX2-FP-NEXT:    vunpcklps {{.*#+}} ymm6 = ymm8[0],ymm11[0],ymm8[1],ymm11[1],ymm8[4],ymm11[4],ymm8[5],ymm11[5]5441; AVX2-FP-NEXT:    vunpcklps {{.*#+}} ymm13 = ymm4[0],ymm5[0],ymm4[1],ymm5[1],ymm4[4],ymm5[4],ymm4[5],ymm5[5]5442; AVX2-FP-NEXT:    vextractf128 $1, %ymm13, %xmm135443; AVX2-FP-NEXT:    vunpcklps {{.*#+}} ymm14 = ymm7[0],ymm9[0],ymm7[1],ymm9[1],ymm7[4],ymm9[4],ymm7[5],ymm9[5]5444; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm15 = ymm14[2,2,2,2]5445; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm13 = xmm13[0,1],xmm15[2,3]5446; AVX2-FP-NEXT:    vunpcklps {{.*#+}} ymm15 = ymm10[0],ymm12[0],ymm10[1],ymm12[1],ymm10[4],ymm12[4],ymm10[5],ymm12[5]5447; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm15 = ymm15[0],ymm6[0],ymm15[2],ymm6[2]5448; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm13 = ymm13[0,1,2,3],ymm15[4,5,6,7]5449; AVX2-FP-NEXT:    vmovups %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5450; AVX2-FP-NEXT:    vbroadcastss 20(%r8), %ymm135451; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm13 = ymm13[0,1,2,3,4],ymm12[5],ymm13[6,7]5452; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm13 = ymm13[0,1,2,3,4,5],ymm6[6,7]5453; AVX2-FP-NEXT:    vextractf128 $1, %ymm14, %xmm65454; AVX2-FP-NEXT:    vshufps {{.*#+}} ymm14 = ymm4[1,1,1,1,5,5,5,5]5455; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm14 = ymm14[0],ymm5[1],ymm14[2,3,4],ymm5[5],ymm14[6,7]5456; AVX2-FP-NEXT:    vextractf128 $1, %ymm14, %xmm145457; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm14 = xmm14[0,1],xmm6[2,3]5458; AVX2-FP-NEXT:    vmovaps 32(%r8), %ymm65459; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm13 = ymm14[0,1,2,3],ymm13[4,5,6,7]5460; AVX2-FP-NEXT:    vmovups %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5461; AVX2-FP-NEXT:    vbroadcastss 24(%r10), %ymm135462; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm13 = ymm8[0,1,2,3,4,5,6],ymm13[7]5463; AVX2-FP-NEXT:    vmovaps 32(%r9), %ymm85464; AVX2-FP-NEXT:    vunpckhps {{.*#+}} ymm12 = ymm10[2],ymm12[2],ymm10[3],ymm12[3],ymm10[6],ymm12[6],ymm10[7],ymm12[7]5465; AVX2-FP-NEXT:    vmovaps 32(%rax), %ymm105466; AVX2-FP-NEXT:    vunpckhps {{.*#+}} ymm4 = ymm4[2],ymm5[2],ymm4[3],ymm5[3],ymm4[6],ymm5[6],ymm4[7],ymm5[7]5467; AVX2-FP-NEXT:    vextractf128 $1, %ymm4, %xmm55468; AVX2-FP-NEXT:    vunpckhps {{.*#+}} ymm9 = ymm7[2],ymm9[2],ymm7[3],ymm9[3],ymm7[6],ymm9[6],ymm7[7],ymm9[7]5469; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm7 = ymm9[2,2,2,2]5470; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm5 = xmm5[0,1],xmm7[2,3]5471; AVX2-FP-NEXT:    vmovaps 32(%r10), %ymm75472; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm13 = ymm12[0,1,2,3,4,5],ymm13[6,7]5473; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm13[4,5,6,7]5474; AVX2-FP-NEXT:    vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5475; AVX2-FP-NEXT:    vbroadcastss 28(%rax), %ymm55476; AVX2-FP-NEXT:    vunpckhps {{.*#+}} ymm5 = ymm5[2],ymm11[2],ymm5[3],ymm11[3],ymm5[6],ymm11[6],ymm5[7],ymm11[7]5477; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm5 = ymm12[1],ymm5[1],ymm12[3],ymm5[3]5478; AVX2-FP-NEXT:    vextractf128 $1, %ymm9, %xmm95479; AVX2-FP-NEXT:    vshufps {{.*#+}} ymm4 = ymm4[2,3,2,3,6,7,6,7]5480; AVX2-FP-NEXT:    vextractf128 $1, %ymm4, %xmm45481; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm4 = xmm4[0,1],xmm9[2,3]5482; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm5[4,5,6,7]5483; AVX2-FP-NEXT:    vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5484; AVX2-FP-NEXT:    vunpcklps {{.*#+}} ymm4 = ymm10[0],ymm7[0],ymm10[1],ymm7[1],ymm10[4],ymm7[4],ymm10[5],ymm7[5]5485; AVX2-FP-NEXT:    vunpcklps {{.*#+}} ymm5 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]5486; AVX2-FP-NEXT:    vextractf128 $1, %ymm5, %xmm55487; AVX2-FP-NEXT:    vunpcklps {{.*#+}} ymm9 = ymm2[0],ymm3[0],ymm2[1],ymm3[1],ymm2[4],ymm3[4],ymm2[5],ymm3[5]5488; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm11 = ymm9[2,2,2,2]5489; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm5 = xmm5[0,1],xmm11[2,3]5490; AVX2-FP-NEXT:    vunpcklps {{.*#+}} ymm11 = ymm6[0],ymm8[0],ymm6[1],ymm8[1],ymm6[4],ymm8[4],ymm6[5],ymm8[5]5491; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm11 = ymm11[0],ymm4[0],ymm11[2],ymm4[2]5492; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm11[4,5,6,7]5493; AVX2-FP-NEXT:    vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5494; AVX2-FP-NEXT:    vbroadcastss 52(%r8), %ymm55495; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3,4],ymm8[5],ymm5[6,7]5496; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3,4,5],ymm4[6,7]5497; AVX2-FP-NEXT:    vextractf128 $1, %ymm9, %xmm45498; AVX2-FP-NEXT:    vshufps {{.*#+}} ymm9 = ymm0[1,1,1,1,5,5,5,5]5499; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm9 = ymm9[0],ymm1[1],ymm9[2,3,4],ymm1[5],ymm9[6,7]5500; AVX2-FP-NEXT:    vextractf128 $1, %ymm9, %xmm95501; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm9 = xmm9[0,1],xmm4[2,3]5502; AVX2-FP-NEXT:    vmovaps 64(%rdi), %ymm45503; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm5 = ymm9[0,1,2,3],ymm5[4,5,6,7]5504; AVX2-FP-NEXT:    vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5505; AVX2-FP-NEXT:    vbroadcastss 56(%r10), %ymm55506; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm9 = ymm10[0,1,2,3,4,5,6],ymm5[7]5507; AVX2-FP-NEXT:    vmovaps 64(%rsi), %ymm55508; AVX2-FP-NEXT:    vunpckhps {{.*#+}} ymm8 = ymm6[2],ymm8[2],ymm6[3],ymm8[3],ymm6[6],ymm8[6],ymm6[7],ymm8[7]5509; AVX2-FP-NEXT:    vmovaps 64(%rdx), %ymm65510; AVX2-FP-NEXT:    vunpckhps {{.*#+}} ymm10 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7]5511; AVX2-FP-NEXT:    vextractf128 $1, %ymm10, %xmm05512; AVX2-FP-NEXT:    vunpckhps {{.*#+}} ymm2 = ymm2[2],ymm3[2],ymm2[3],ymm3[3],ymm2[6],ymm3[6],ymm2[7],ymm3[7]5513; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm1 = ymm2[2,2,2,2]5514; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm1 = xmm0[0,1],xmm1[2,3]5515; AVX2-FP-NEXT:    vmovaps 64(%rcx), %ymm05516; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm3 = ymm8[0,1,2,3,4,5],ymm9[6,7]5517; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm3[4,5,6,7]5518; AVX2-FP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5519; AVX2-FP-NEXT:    vbroadcastss 60(%rax), %ymm15520; AVX2-FP-NEXT:    vunpckhps {{.*#+}} ymm3 = ymm1[2],ymm7[2],ymm1[3],ymm7[3],ymm1[6],ymm7[6],ymm1[7],ymm7[7]5521; AVX2-FP-NEXT:    vmovaps 64(%r8), %ymm15522; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm3 = ymm8[1],ymm3[1],ymm8[3],ymm3[3]5523; AVX2-FP-NEXT:    vextractf128 $1, %ymm2, %xmm25524; AVX2-FP-NEXT:    vshufps {{.*#+}} ymm7 = ymm10[2,3,2,3,6,7,6,7]5525; AVX2-FP-NEXT:    vextractf128 $1, %ymm7, %xmm75526; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm2 = xmm7[0,1],xmm2[2,3]5527; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]5528; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5529; AVX2-FP-NEXT:    vunpcklps {{.*#+}} ymm2 = ymm4[0],ymm5[0],ymm4[1],ymm5[1],ymm4[4],ymm5[4],ymm4[5],ymm5[5]5530; AVX2-FP-NEXT:    vextractf128 $1, %ymm2, %xmm25531; AVX2-FP-NEXT:    vunpcklps {{.*#+}} ymm3 = ymm6[0],ymm0[0],ymm6[1],ymm0[1],ymm6[4],ymm0[4],ymm6[5],ymm0[5]5532; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm7 = ymm3[2,2,2,2]5533; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm2 = xmm2[0,1],xmm7[2,3]5534; AVX2-FP-NEXT:    vmovaps 64(%r9), %ymm75535; AVX2-FP-NEXT:    vmovaps 64(%rax), %ymm85536; AVX2-FP-NEXT:    vmovaps 64(%r10), %ymm95537; AVX2-FP-NEXT:    vunpcklps {{.*#+}} ymm10 = ymm8[0],ymm9[0],ymm8[1],ymm9[1],ymm8[4],ymm9[4],ymm8[5],ymm9[5]5538; AVX2-FP-NEXT:    vunpcklps {{.*#+}} ymm11 = ymm1[0],ymm7[0],ymm1[1],ymm7[1],ymm1[4],ymm7[4],ymm1[5],ymm7[5]5539; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm11 = ymm11[0],ymm10[0],ymm11[2],ymm10[2]5540; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm11[4,5,6,7]5541; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5542; AVX2-FP-NEXT:    vbroadcastss 84(%r8), %ymm25543; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4],ymm7[5],ymm2[6,7]5544; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm10[6,7]5545; AVX2-FP-NEXT:    vextractf128 $1, %ymm3, %xmm35546; AVX2-FP-NEXT:    vshufps {{.*#+}} ymm10 = ymm4[1,1,1,1,5,5,5,5]5547; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm10 = ymm10[0],ymm5[1],ymm10[2,3,4],ymm5[5],ymm10[6,7]5548; AVX2-FP-NEXT:    vextractf128 $1, %ymm10, %xmm105549; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm3 = xmm10[0,1],xmm3[2,3]5550; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm2 = ymm3[0,1,2,3],ymm2[4,5,6,7]5551; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5552; AVX2-FP-NEXT:    vbroadcastss 88(%r10), %ymm25553; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm2 = ymm8[0,1,2,3,4,5,6],ymm2[7]5554; AVX2-FP-NEXT:    vunpckhps {{.*#+}} ymm1 = ymm1[2],ymm7[2],ymm1[3],ymm7[3],ymm1[6],ymm7[6],ymm1[7],ymm7[7]5555; AVX2-FP-NEXT:    vunpckhps {{.*#+}} ymm3 = ymm4[2],ymm5[2],ymm4[3],ymm5[3],ymm4[6],ymm5[6],ymm4[7],ymm5[7]5556; AVX2-FP-NEXT:    vunpckhps {{.*#+}} ymm4 = ymm6[2],ymm0[2],ymm6[3],ymm0[3],ymm6[6],ymm0[6],ymm6[7],ymm0[7]5557; AVX2-FP-NEXT:    vextractf128 $1, %ymm3, %xmm05558; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm5 = ymm4[2,2,2,2]5559; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm5[2,3]5560; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm2 = ymm1[0,1,2,3,4,5],ymm2[6,7]5561; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]5562; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5563; AVX2-FP-NEXT:    vbroadcastss 92(%rax), %ymm05564; AVX2-FP-NEXT:    vunpckhps {{.*#+}} ymm2 = ymm0[2],ymm9[2],ymm0[3],ymm9[3],ymm0[6],ymm9[6],ymm0[7],ymm9[7]5565; AVX2-FP-NEXT:    vmovaps 96(%rdi), %ymm05566; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm5 = ymm1[1],ymm2[1],ymm1[3],ymm2[3]5567; AVX2-FP-NEXT:    vmovaps 96(%rsi), %ymm15568; AVX2-FP-NEXT:    vextractf128 $1, %ymm4, %xmm25569; AVX2-FP-NEXT:    vshufps {{.*#+}} ymm3 = ymm3[2,3,2,3,6,7,6,7]5570; AVX2-FP-NEXT:    vextractf128 $1, %ymm3, %xmm35571; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm3 = xmm3[0,1],xmm2[2,3]5572; AVX2-FP-NEXT:    vmovaps 96(%rdx), %ymm25573; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm5[4,5,6,7]5574; AVX2-FP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5575; AVX2-FP-NEXT:    vunpcklps {{.*#+}} ymm3 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]5576; AVX2-FP-NEXT:    vextractf128 $1, %ymm3, %xmm35577; AVX2-FP-NEXT:    vmovaps 96(%rcx), %ymm45578; AVX2-FP-NEXT:    vunpcklps {{.*#+}} ymm5 = ymm2[0],ymm4[0],ymm2[1],ymm4[1],ymm2[4],ymm4[4],ymm2[5],ymm4[5]5579; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm6 = ymm5[2,2,2,2]5580; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm3 = xmm3[0,1],xmm6[2,3]5581; AVX2-FP-NEXT:    vmovaps 96(%r8), %ymm65582; AVX2-FP-NEXT:    vmovaps 96(%r9), %ymm75583; AVX2-FP-NEXT:    vmovaps 96(%rax), %ymm85584; AVX2-FP-NEXT:    vmovaps 96(%r10), %ymm95585; AVX2-FP-NEXT:    vunpcklps {{.*#+}} ymm10 = ymm8[0],ymm9[0],ymm8[1],ymm9[1],ymm8[4],ymm9[4],ymm8[5],ymm9[5]5586; AVX2-FP-NEXT:    vunpcklps {{.*#+}} ymm11 = ymm6[0],ymm7[0],ymm6[1],ymm7[1],ymm6[4],ymm7[4],ymm6[5],ymm7[5]5587; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm11 = ymm11[0],ymm10[0],ymm11[2],ymm10[2]5588; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm11[4,5,6,7]5589; AVX2-FP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5590; AVX2-FP-NEXT:    vbroadcastss 116(%r8), %ymm35591; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4],ymm7[5],ymm3[6,7]5592; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm10[6,7]5593; AVX2-FP-NEXT:    vextractf128 $1, %ymm5, %xmm55594; AVX2-FP-NEXT:    vshufps {{.*#+}} ymm10 = ymm0[1,1,1,1,5,5,5,5]5595; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm10 = ymm10[0],ymm1[1],ymm10[2,3,4],ymm1[5],ymm10[6,7]5596; AVX2-FP-NEXT:    vextractf128 $1, %ymm10, %xmm105597; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm5 = xmm10[0,1],xmm5[2,3]5598; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm3 = ymm5[0,1,2,3],ymm3[4,5,6,7]5599; AVX2-FP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5600; AVX2-FP-NEXT:    vbroadcastss 120(%r10), %ymm35601; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm3 = ymm8[0,1,2,3,4,5,6],ymm3[7]5602; AVX2-FP-NEXT:    vunpckhps {{.*#+}} ymm5 = ymm6[2],ymm7[2],ymm6[3],ymm7[3],ymm6[6],ymm7[6],ymm6[7],ymm7[7]5603; AVX2-FP-NEXT:    vunpckhps {{.*#+}} ymm0 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7]5604; AVX2-FP-NEXT:    vunpckhps {{.*#+}} ymm1 = ymm2[2],ymm4[2],ymm2[3],ymm4[3],ymm2[6],ymm4[6],ymm2[7],ymm4[7]5605; AVX2-FP-NEXT:    vextractf128 $1, %ymm0, %xmm25606; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm4 = ymm1[2,2,2,2]5607; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm2 = xmm2[0,1],xmm4[2,3]5608; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm3 = ymm5[0,1,2,3,4,5],ymm3[6,7]5609; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]5610; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5611; AVX2-FP-NEXT:    vbroadcastss 124(%rax), %ymm25612; AVX2-FP-NEXT:    vunpckhps {{.*#+}} ymm2 = ymm2[2],ymm9[2],ymm2[3],ymm9[3],ymm2[6],ymm9[6],ymm2[7],ymm9[7]5613; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm5[1],ymm2[1],ymm5[3],ymm2[3]5614; AVX2-FP-NEXT:    vextractf128 $1, %ymm1, %xmm15615; AVX2-FP-NEXT:    vshufps {{.*#+}} ymm0 = ymm0[2,3,2,3,6,7,6,7]5616; AVX2-FP-NEXT:    vextractf128 $1, %ymm0, %xmm05617; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]5618; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]5619; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5620; AVX2-FP-NEXT:    vmovaps (%rcx), %xmm05621; AVX2-FP-NEXT:    vbroadcastss %xmm0, %xmm25622; AVX2-FP-NEXT:    vmovaps (%rdx), %xmm15623; AVX2-FP-NEXT:    vbroadcastss %xmm1, %xmm35624; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]5625; AVX2-FP-NEXT:    vmovaps (%rsi), %xmm45626; AVX2-FP-NEXT:    vmovaps (%rdi), %xmm55627; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm3 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]5628; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm6 = xmm3[0,1],xmm2[2,3]5629; AVX2-FP-NEXT:    vmovaps (%r10), %xmm25630; AVX2-FP-NEXT:    vbroadcastss %xmm2, %xmm75631; AVX2-FP-NEXT:    vmovaps (%rax), %xmm35632; AVX2-FP-NEXT:    vbroadcastss %xmm3, %xmm85633; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm8[0],xmm7[0],xmm8[1],xmm7[1]5634; AVX2-FP-NEXT:    vmovaps (%r9), %xmm85635; AVX2-FP-NEXT:    vmovaps (%r8), %xmm95636; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm10 = xmm9[0],xmm8[0],xmm9[1],xmm8[1]5637; AVX2-FP-NEXT:    vinsertf128 $1, %xmm10, %ymm0, %ymm105638; AVX2-FP-NEXT:    vinsertf128 $1, %xmm7, %ymm0, %ymm75639; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm7 = ymm10[0,1,2,3,4,5],ymm7[6,7]5640; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7]5641; AVX2-FP-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5642; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm6 = xmm5[1,1,1,1]5643; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm6 = xmm6[0],xmm4[1],xmm6[2,3]5644; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]5645; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm6 = xmm6[0,1],xmm7[2,3]5646; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]5647; AVX2-FP-NEXT:    vinsertf128 $1, %xmm7, %ymm0, %ymm75648; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm10 = xmm9[1,1,1,1]5649; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm10 = xmm10[0],xmm8[1],xmm10[2,3]5650; AVX2-FP-NEXT:    vinsertf128 $1, %xmm10, %ymm0, %ymm105651; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm7 = ymm10[0,1,2,3,4,5],ymm7[6,7]5652; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7]5653; AVX2-FP-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5654; AVX2-FP-NEXT:    vunpckhps {{.*#+}} xmm4 = xmm5[2],xmm4[2],xmm5[3],xmm4[3]5655; AVX2-FP-NEXT:    vunpckhps {{.*#+}} xmm5 = xmm9[2],xmm8[2],xmm9[3],xmm8[3]5656; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm6 = xmm2[2,2,2,2]5657; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm6 = xmm3[0,1,2],xmm6[3]5658; AVX2-FP-NEXT:    vinsertf128 $1, %xmm6, %ymm0, %ymm65659; AVX2-FP-NEXT:    vinsertf128 $1, %xmm5, %ymm0, %ymm75660; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3,4,5],ymm6[6,7]5661; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm7 = xmm0[2,2,2,2]5662; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm7 = xmm1[0,1,2],xmm7[3]5663; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm7 = xmm4[0,1],xmm7[2,3]5664; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]5665; AVX2-FP-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5666; AVX2-FP-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm1[2],xmm0[2],xmm1[3],xmm0[3]5667; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm4[1],xmm0[1]5668; AVX2-FP-NEXT:    vunpckhps {{.*#+}} xmm1 = xmm3[2],xmm2[2],xmm3[3],xmm2[3]5669; AVX2-FP-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm15670; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm2 = xmm5[2,3,2,3]5671; AVX2-FP-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm25672; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]5673; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]5674; AVX2-FP-NEXT:    vmovups %ymm0, (%rsp) # 32-byte Spill5675; AVX2-FP-NEXT:    vmovaps 32(%rcx), %xmm05676; AVX2-FP-NEXT:    vmovaps 32(%rdx), %xmm15677; AVX2-FP-NEXT:    vbroadcastss %xmm0, %xmm25678; AVX2-FP-NEXT:    vbroadcastss %xmm1, %xmm35679; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]5680; AVX2-FP-NEXT:    vmovaps 32(%rsi), %xmm45681; AVX2-FP-NEXT:    vmovaps 32(%rdi), %xmm55682; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm3 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]5683; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm6 = xmm3[0,1],xmm2[2,3]5684; AVX2-FP-NEXT:    vmovaps 32(%r10), %xmm25685; AVX2-FP-NEXT:    vmovaps 32(%rax), %xmm35686; AVX2-FP-NEXT:    vbroadcastss %xmm2, %xmm75687; AVX2-FP-NEXT:    vbroadcastss %xmm3, %xmm85688; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm8[0],xmm7[0],xmm8[1],xmm7[1]5689; AVX2-FP-NEXT:    vmovaps 32(%r9), %xmm85690; AVX2-FP-NEXT:    vmovaps 32(%r8), %xmm95691; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm10 = xmm9[0],xmm8[0],xmm9[1],xmm8[1]5692; AVX2-FP-NEXT:    vinsertf128 $1, %xmm10, %ymm0, %ymm105693; AVX2-FP-NEXT:    vinsertf128 $1, %xmm7, %ymm0, %ymm75694; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm7 = ymm10[0,1,2,3,4,5],ymm7[6,7]5695; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7]5696; AVX2-FP-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5697; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm6 = xmm5[1,1,1,1]5698; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm6 = xmm6[0],xmm4[1],xmm6[2,3]5699; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]5700; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm6 = xmm6[0,1],xmm7[2,3]5701; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]5702; AVX2-FP-NEXT:    vinsertf128 $1, %xmm7, %ymm0, %ymm75703; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm10 = xmm9[1,1,1,1]5704; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm10 = xmm10[0],xmm8[1],xmm10[2,3]5705; AVX2-FP-NEXT:    vinsertf128 $1, %xmm10, %ymm0, %ymm105706; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm7 = ymm10[0,1,2,3,4,5],ymm7[6,7]5707; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7]5708; AVX2-FP-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5709; AVX2-FP-NEXT:    vunpckhps {{.*#+}} xmm4 = xmm5[2],xmm4[2],xmm5[3],xmm4[3]5710; AVX2-FP-NEXT:    vunpckhps {{.*#+}} xmm5 = xmm9[2],xmm8[2],xmm9[3],xmm8[3]5711; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm6 = xmm2[2,2,2,2]5712; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm6 = xmm3[0,1,2],xmm6[3]5713; AVX2-FP-NEXT:    vinsertf128 $1, %xmm6, %ymm0, %ymm65714; AVX2-FP-NEXT:    vinsertf128 $1, %xmm5, %ymm0, %ymm75715; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3,4,5],ymm6[6,7]5716; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm7 = xmm0[2,2,2,2]5717; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm7 = xmm1[0,1,2],xmm7[3]5718; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm7 = xmm4[0,1],xmm7[2,3]5719; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]5720; AVX2-FP-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5721; AVX2-FP-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm1[2],xmm0[2],xmm1[3],xmm0[3]5722; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm4[1],xmm0[1]5723; AVX2-FP-NEXT:    vunpckhps {{.*#+}} xmm1 = xmm3[2],xmm2[2],xmm3[3],xmm2[3]5724; AVX2-FP-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm15725; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm2 = xmm5[2,3,2,3]5726; AVX2-FP-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm25727; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]5728; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]5729; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5730; AVX2-FP-NEXT:    vmovaps 64(%rcx), %xmm35731; AVX2-FP-NEXT:    vmovaps 64(%rdx), %xmm15732; AVX2-FP-NEXT:    vbroadcastss %xmm3, %xmm05733; AVX2-FP-NEXT:    vbroadcastss %xmm1, %xmm25734; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm0 = xmm2[0],xmm0[0],xmm2[1],xmm0[1]5735; AVX2-FP-NEXT:    vmovaps 64(%rsi), %xmm45736; AVX2-FP-NEXT:    vmovaps 64(%rdi), %xmm55737; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm2 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]5738; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm9 = xmm2[0,1],xmm0[2,3]5739; AVX2-FP-NEXT:    vmovaps 64(%r10), %xmm25740; AVX2-FP-NEXT:    vmovaps 64(%rax), %xmm05741; AVX2-FP-NEXT:    vbroadcastss %xmm2, %xmm75742; AVX2-FP-NEXT:    vbroadcastss %xmm0, %xmm85743; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm8[0],xmm7[0],xmm8[1],xmm7[1]5744; AVX2-FP-NEXT:    vmovaps 64(%r9), %xmm85745; AVX2-FP-NEXT:    vmovaps 64(%r8), %xmm65746; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm10 = xmm6[0],xmm8[0],xmm6[1],xmm8[1]5747; AVX2-FP-NEXT:    vinsertf128 $1, %xmm10, %ymm0, %ymm105748; AVX2-FP-NEXT:    vinsertf128 $1, %xmm7, %ymm0, %ymm75749; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm7 = ymm10[0,1,2,3,4,5],ymm7[6,7]5750; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm11 = ymm9[0,1,2,3],ymm7[4,5,6,7]5751; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm7 = xmm5[1,1,1,1]5752; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm7 = xmm7[0],xmm4[1],xmm7[2,3]5753; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm9 = xmm1[0],xmm3[0],xmm1[1],xmm3[1]5754; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm7 = xmm7[0,1],xmm9[2,3]5755; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm9 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]5756; AVX2-FP-NEXT:    vinsertf128 $1, %xmm9, %ymm0, %ymm95757; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm15 = xmm6[1,1,1,1]5758; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm15 = xmm15[0],xmm8[1],xmm15[2,3]5759; AVX2-FP-NEXT:    vinsertf128 $1, %xmm15, %ymm0, %ymm155760; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm9 = ymm15[0,1,2,3,4,5],ymm9[6,7]5761; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm10 = ymm7[0,1,2,3],ymm9[4,5,6,7]5762; AVX2-FP-NEXT:    vunpckhps {{.*#+}} xmm4 = xmm5[2],xmm4[2],xmm5[3],xmm4[3]5763; AVX2-FP-NEXT:    vunpckhps {{.*#+}} xmm5 = xmm6[2],xmm8[2],xmm6[3],xmm8[3]5764; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm6 = xmm2[2,2,2,2]5765; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm6 = xmm0[0,1,2],xmm6[3]5766; AVX2-FP-NEXT:    vinsertf128 $1, %xmm6, %ymm0, %ymm65767; AVX2-FP-NEXT:    vinsertf128 $1, %xmm5, %ymm0, %ymm75768; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3,4,5],ymm6[6,7]5769; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm7 = xmm3[2,2,2,2]5770; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm7 = xmm1[0,1,2],xmm7[3]5771; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm7 = xmm4[0,1],xmm7[2,3]5772; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm9 = ymm7[0,1,2,3],ymm6[4,5,6,7]5773; AVX2-FP-NEXT:    vunpckhps {{.*#+}} xmm1 = xmm1[2],xmm3[2],xmm1[3],xmm3[3]5774; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm4[1],xmm1[1]5775; AVX2-FP-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm0[2],xmm2[2],xmm0[3],xmm2[3]5776; AVX2-FP-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm05777; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm2 = xmm5[2,3,2,3]5778; AVX2-FP-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm25779; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3,4,5],ymm0[6,7]5780; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm8 = ymm1[0,1,2,3],ymm0[4,5,6,7]5781; AVX2-FP-NEXT:    vmovaps 96(%rcx), %xmm65782; AVX2-FP-NEXT:    vbroadcastss %xmm6, %xmm05783; AVX2-FP-NEXT:    vmovaps 96(%rdx), %xmm55784; AVX2-FP-NEXT:    vbroadcastss %xmm5, %xmm15785; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]5786; AVX2-FP-NEXT:    vmovaps 96(%rsi), %xmm45787; AVX2-FP-NEXT:    vmovaps 96(%rdi), %xmm25788; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm3 = xmm2[0],xmm4[0],xmm2[1],xmm4[1]5789; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm7 = xmm3[0,1],xmm0[2,3]5790; AVX2-FP-NEXT:    vmovaps 96(%r10), %xmm35791; AVX2-FP-NEXT:    vbroadcastss %xmm3, %xmm155792; AVX2-FP-NEXT:    vmovaps 96(%rax), %xmm15793; AVX2-FP-NEXT:    vbroadcastss %xmm1, %xmm145794; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm14 = xmm14[0],xmm15[0],xmm14[1],xmm15[1]5795; AVX2-FP-NEXT:    vmovaps 96(%r9), %xmm155796; AVX2-FP-NEXT:    vmovaps 96(%r8), %xmm05797; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm13 = xmm0[0],xmm15[0],xmm0[1],xmm15[1]5798; AVX2-FP-NEXT:    vinsertf128 $1, %xmm13, %ymm0, %ymm135799; AVX2-FP-NEXT:    vinsertf128 $1, %xmm14, %ymm0, %ymm145800; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm13 = ymm13[0,1,2,3,4,5],ymm14[6,7]5801; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm14 = ymm7[0,1,2,3],ymm13[4,5,6,7]5802; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm7 = xmm2[1,1,1,1]5803; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm7 = xmm7[0],xmm4[1],xmm7[2,3]5804; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm13 = xmm5[0],xmm6[0],xmm5[1],xmm6[1]5805; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm7 = xmm7[0,1],xmm13[2,3]5806; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm13 = xmm1[0],xmm3[0],xmm1[1],xmm3[1]5807; AVX2-FP-NEXT:    vinsertf128 $1, %xmm13, %ymm0, %ymm135808; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm12 = xmm0[1,1,1,1]5809; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm12 = xmm12[0],xmm15[1],xmm12[2,3]5810; AVX2-FP-NEXT:    vinsertf128 $1, %xmm12, %ymm0, %ymm125811; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm12 = ymm12[0,1,2,3,4,5],ymm13[6,7]5812; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm7 = ymm7[0,1,2,3],ymm12[4,5,6,7]5813; AVX2-FP-NEXT:    vunpckhps {{.*#+}} xmm2 = xmm2[2],xmm4[2],xmm2[3],xmm4[3]5814; AVX2-FP-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm0[2],xmm15[2],xmm0[3],xmm15[3]5815; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm4 = xmm3[2,2,2,2]5816; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm4 = xmm1[0,1,2],xmm4[3]5817; AVX2-FP-NEXT:    vinsertf128 $1, %xmm4, %ymm0, %ymm45818; AVX2-FP-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm125819; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm4 = ymm12[0,1,2,3,4,5],ymm4[6,7]5820; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm12 = xmm6[2,2,2,2]5821; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm12 = xmm5[0,1,2],xmm12[3]5822; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm12 = xmm2[0,1],xmm12[2,3]5823; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm4 = ymm12[0,1,2,3],ymm4[4,5,6,7]5824; AVX2-FP-NEXT:    vunpckhps {{.*#+}} xmm5 = xmm5[2],xmm6[2],xmm5[3],xmm6[3]5825; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} xmm2 = xmm2[1],xmm5[1]5826; AVX2-FP-NEXT:    vunpckhps {{.*#+}} xmm1 = xmm1[2],xmm3[2],xmm1[3],xmm3[3]5827; AVX2-FP-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm15828; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm0 = xmm0[2,3,2,3]5829; AVX2-FP-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm05830; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]5831; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]5832; AVX2-FP-NEXT:    movq {{[0-9]+}}(%rsp), %rax5833; AVX2-FP-NEXT:    vmovaps %ymm0, 864(%rax)5834; AVX2-FP-NEXT:    vmovaps %ymm4, 832(%rax)5835; AVX2-FP-NEXT:    vmovaps %ymm7, 800(%rax)5836; AVX2-FP-NEXT:    vmovaps %ymm14, 768(%rax)5837; AVX2-FP-NEXT:    vmovaps %ymm8, 608(%rax)5838; AVX2-FP-NEXT:    vmovaps %ymm9, 576(%rax)5839; AVX2-FP-NEXT:    vmovaps %ymm10, 544(%rax)5840; AVX2-FP-NEXT:    vmovaps %ymm11, 512(%rax)5841; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5842; AVX2-FP-NEXT:    vmovaps %ymm0, 352(%rax)5843; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5844; AVX2-FP-NEXT:    vmovaps %ymm0, 320(%rax)5845; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5846; AVX2-FP-NEXT:    vmovaps %ymm0, 288(%rax)5847; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5848; AVX2-FP-NEXT:    vmovaps %ymm0, 256(%rax)5849; AVX2-FP-NEXT:    vmovups (%rsp), %ymm0 # 32-byte Reload5850; AVX2-FP-NEXT:    vmovaps %ymm0, 96(%rax)5851; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5852; AVX2-FP-NEXT:    vmovaps %ymm0, 64(%rax)5853; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5854; AVX2-FP-NEXT:    vmovaps %ymm0, 32(%rax)5855; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5856; AVX2-FP-NEXT:    vmovaps %ymm0, (%rax)5857; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5858; AVX2-FP-NEXT:    vmovaps %ymm0, 992(%rax)5859; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5860; AVX2-FP-NEXT:    vmovaps %ymm0, 960(%rax)5861; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5862; AVX2-FP-NEXT:    vmovaps %ymm0, 928(%rax)5863; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5864; AVX2-FP-NEXT:    vmovaps %ymm0, 896(%rax)5865; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5866; AVX2-FP-NEXT:    vmovaps %ymm0, 736(%rax)5867; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5868; AVX2-FP-NEXT:    vmovaps %ymm0, 704(%rax)5869; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5870; AVX2-FP-NEXT:    vmovaps %ymm0, 672(%rax)5871; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5872; AVX2-FP-NEXT:    vmovaps %ymm0, 640(%rax)5873; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5874; AVX2-FP-NEXT:    vmovaps %ymm0, 480(%rax)5875; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5876; AVX2-FP-NEXT:    vmovaps %ymm0, 448(%rax)5877; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5878; AVX2-FP-NEXT:    vmovaps %ymm0, 416(%rax)5879; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5880; AVX2-FP-NEXT:    vmovaps %ymm0, 384(%rax)5881; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5882; AVX2-FP-NEXT:    vmovaps %ymm0, 224(%rax)5883; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5884; AVX2-FP-NEXT:    vmovaps %ymm0, 192(%rax)5885; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5886; AVX2-FP-NEXT:    vmovaps %ymm0, 160(%rax)5887; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5888; AVX2-FP-NEXT:    vmovaps %ymm0, 128(%rax)5889; AVX2-FP-NEXT:    addq $648, %rsp # imm = 0x2885890; AVX2-FP-NEXT:    vzeroupper5891; AVX2-FP-NEXT:    retq5892;5893; AVX2-FCP-LABEL: store_i32_stride8_vf32:5894; AVX2-FCP:       # %bb.0:5895; AVX2-FCP-NEXT:    subq $648, %rsp # imm = 0x2885896; AVX2-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r105897; AVX2-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax5898; AVX2-FCP-NEXT:    vmovaps 32(%rdi), %ymm05899; AVX2-FCP-NEXT:    vmovaps (%rdi), %ymm45900; AVX2-FCP-NEXT:    vmovaps 32(%rsi), %ymm15901; AVX2-FCP-NEXT:    vmovaps (%rsi), %ymm55902; AVX2-FCP-NEXT:    vmovaps 32(%rdx), %ymm25903; AVX2-FCP-NEXT:    vmovaps (%rdx), %ymm75904; AVX2-FCP-NEXT:    vmovaps 32(%rcx), %ymm35905; AVX2-FCP-NEXT:    vmovaps (%rcx), %ymm95906; AVX2-FCP-NEXT:    vmovaps (%r8), %ymm105907; AVX2-FCP-NEXT:    vmovaps (%r9), %ymm125908; AVX2-FCP-NEXT:    vmovaps (%rax), %ymm85909; AVX2-FCP-NEXT:    vmovaps (%r10), %ymm115910; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} ymm6 = ymm8[0],ymm11[0],ymm8[1],ymm11[1],ymm8[4],ymm11[4],ymm8[5],ymm11[5]5911; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} ymm13 = ymm4[0],ymm5[0],ymm4[1],ymm5[1],ymm4[4],ymm5[4],ymm4[5],ymm5[5]5912; AVX2-FCP-NEXT:    vextractf128 $1, %ymm13, %xmm135913; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} ymm14 = ymm7[0],ymm9[0],ymm7[1],ymm9[1],ymm7[4],ymm9[4],ymm7[5],ymm9[5]5914; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm15 = ymm14[2,2,2,2]5915; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm13 = xmm13[0,1],xmm15[2,3]5916; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} ymm15 = ymm10[0],ymm12[0],ymm10[1],ymm12[1],ymm10[4],ymm12[4],ymm10[5],ymm12[5]5917; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm15 = ymm15[0],ymm6[0],ymm15[2],ymm6[2]5918; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm13 = ymm13[0,1,2,3],ymm15[4,5,6,7]5919; AVX2-FCP-NEXT:    vmovups %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5920; AVX2-FCP-NEXT:    vbroadcastss 20(%r8), %ymm135921; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm13 = ymm13[0,1,2,3,4],ymm12[5],ymm13[6,7]5922; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm13 = ymm13[0,1,2,3,4,5],ymm6[6,7]5923; AVX2-FCP-NEXT:    vextractf128 $1, %ymm14, %xmm65924; AVX2-FCP-NEXT:    vshufps {{.*#+}} ymm14 = ymm4[1,1,1,1,5,5,5,5]5925; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm14 = ymm14[0],ymm5[1],ymm14[2,3,4],ymm5[5],ymm14[6,7]5926; AVX2-FCP-NEXT:    vextractf128 $1, %ymm14, %xmm145927; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm14 = xmm14[0,1],xmm6[2,3]5928; AVX2-FCP-NEXT:    vmovaps 32(%r8), %ymm65929; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm13 = ymm14[0,1,2,3],ymm13[4,5,6,7]5930; AVX2-FCP-NEXT:    vmovups %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5931; AVX2-FCP-NEXT:    vbroadcastss 24(%r10), %ymm135932; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm13 = ymm8[0,1,2,3,4,5,6],ymm13[7]5933; AVX2-FCP-NEXT:    vmovaps 32(%r9), %ymm85934; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} ymm12 = ymm10[2],ymm12[2],ymm10[3],ymm12[3],ymm10[6],ymm12[6],ymm10[7],ymm12[7]5935; AVX2-FCP-NEXT:    vmovaps 32(%rax), %ymm105936; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} ymm4 = ymm4[2],ymm5[2],ymm4[3],ymm5[3],ymm4[6],ymm5[6],ymm4[7],ymm5[7]5937; AVX2-FCP-NEXT:    vextractf128 $1, %ymm4, %xmm55938; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} ymm9 = ymm7[2],ymm9[2],ymm7[3],ymm9[3],ymm7[6],ymm9[6],ymm7[7],ymm9[7]5939; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm7 = ymm9[2,2,2,2]5940; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm5 = xmm5[0,1],xmm7[2,3]5941; AVX2-FCP-NEXT:    vmovaps 32(%r10), %ymm75942; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm13 = ymm12[0,1,2,3,4,5],ymm13[6,7]5943; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm13[4,5,6,7]5944; AVX2-FCP-NEXT:    vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5945; AVX2-FCP-NEXT:    vbroadcastss 28(%rax), %ymm55946; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} ymm5 = ymm5[2],ymm11[2],ymm5[3],ymm11[3],ymm5[6],ymm11[6],ymm5[7],ymm11[7]5947; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm5 = ymm12[1],ymm5[1],ymm12[3],ymm5[3]5948; AVX2-FCP-NEXT:    vextractf128 $1, %ymm9, %xmm95949; AVX2-FCP-NEXT:    vshufps {{.*#+}} ymm4 = ymm4[2,3,2,3,6,7,6,7]5950; AVX2-FCP-NEXT:    vextractf128 $1, %ymm4, %xmm45951; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm4 = xmm4[0,1],xmm9[2,3]5952; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm5[4,5,6,7]5953; AVX2-FCP-NEXT:    vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5954; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} ymm4 = ymm10[0],ymm7[0],ymm10[1],ymm7[1],ymm10[4],ymm7[4],ymm10[5],ymm7[5]5955; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} ymm5 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]5956; AVX2-FCP-NEXT:    vextractf128 $1, %ymm5, %xmm55957; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} ymm9 = ymm2[0],ymm3[0],ymm2[1],ymm3[1],ymm2[4],ymm3[4],ymm2[5],ymm3[5]5958; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm11 = ymm9[2,2,2,2]5959; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm5 = xmm5[0,1],xmm11[2,3]5960; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} ymm11 = ymm6[0],ymm8[0],ymm6[1],ymm8[1],ymm6[4],ymm8[4],ymm6[5],ymm8[5]5961; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm11 = ymm11[0],ymm4[0],ymm11[2],ymm4[2]5962; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm11[4,5,6,7]5963; AVX2-FCP-NEXT:    vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5964; AVX2-FCP-NEXT:    vbroadcastss 52(%r8), %ymm55965; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3,4],ymm8[5],ymm5[6,7]5966; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3,4,5],ymm4[6,7]5967; AVX2-FCP-NEXT:    vextractf128 $1, %ymm9, %xmm45968; AVX2-FCP-NEXT:    vshufps {{.*#+}} ymm9 = ymm0[1,1,1,1,5,5,5,5]5969; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm9 = ymm9[0],ymm1[1],ymm9[2,3,4],ymm1[5],ymm9[6,7]5970; AVX2-FCP-NEXT:    vextractf128 $1, %ymm9, %xmm95971; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm9 = xmm9[0,1],xmm4[2,3]5972; AVX2-FCP-NEXT:    vmovaps 64(%rdi), %ymm45973; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm5 = ymm9[0,1,2,3],ymm5[4,5,6,7]5974; AVX2-FCP-NEXT:    vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5975; AVX2-FCP-NEXT:    vbroadcastss 56(%r10), %ymm55976; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm9 = ymm10[0,1,2,3,4,5,6],ymm5[7]5977; AVX2-FCP-NEXT:    vmovaps 64(%rsi), %ymm55978; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} ymm8 = ymm6[2],ymm8[2],ymm6[3],ymm8[3],ymm6[6],ymm8[6],ymm6[7],ymm8[7]5979; AVX2-FCP-NEXT:    vmovaps 64(%rdx), %ymm65980; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} ymm10 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7]5981; AVX2-FCP-NEXT:    vextractf128 $1, %ymm10, %xmm05982; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} ymm2 = ymm2[2],ymm3[2],ymm2[3],ymm3[3],ymm2[6],ymm3[6],ymm2[7],ymm3[7]5983; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm1 = ymm2[2,2,2,2]5984; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm1 = xmm0[0,1],xmm1[2,3]5985; AVX2-FCP-NEXT:    vmovaps 64(%rcx), %ymm05986; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm3 = ymm8[0,1,2,3,4,5],ymm9[6,7]5987; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm3[4,5,6,7]5988; AVX2-FCP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5989; AVX2-FCP-NEXT:    vbroadcastss 60(%rax), %ymm15990; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} ymm3 = ymm1[2],ymm7[2],ymm1[3],ymm7[3],ymm1[6],ymm7[6],ymm1[7],ymm7[7]5991; AVX2-FCP-NEXT:    vmovaps 64(%r8), %ymm15992; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm3 = ymm8[1],ymm3[1],ymm8[3],ymm3[3]5993; AVX2-FCP-NEXT:    vextractf128 $1, %ymm2, %xmm25994; AVX2-FCP-NEXT:    vshufps {{.*#+}} ymm7 = ymm10[2,3,2,3,6,7,6,7]5995; AVX2-FCP-NEXT:    vextractf128 $1, %ymm7, %xmm75996; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm2 = xmm7[0,1],xmm2[2,3]5997; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]5998; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5999; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} ymm2 = ymm4[0],ymm5[0],ymm4[1],ymm5[1],ymm4[4],ymm5[4],ymm4[5],ymm5[5]6000; AVX2-FCP-NEXT:    vextractf128 $1, %ymm2, %xmm26001; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} ymm3 = ymm6[0],ymm0[0],ymm6[1],ymm0[1],ymm6[4],ymm0[4],ymm6[5],ymm0[5]6002; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm7 = ymm3[2,2,2,2]6003; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm2 = xmm2[0,1],xmm7[2,3]6004; AVX2-FCP-NEXT:    vmovaps 64(%r9), %ymm76005; AVX2-FCP-NEXT:    vmovaps 64(%rax), %ymm86006; AVX2-FCP-NEXT:    vmovaps 64(%r10), %ymm96007; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} ymm10 = ymm8[0],ymm9[0],ymm8[1],ymm9[1],ymm8[4],ymm9[4],ymm8[5],ymm9[5]6008; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} ymm11 = ymm1[0],ymm7[0],ymm1[1],ymm7[1],ymm1[4],ymm7[4],ymm1[5],ymm7[5]6009; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm11 = ymm11[0],ymm10[0],ymm11[2],ymm10[2]6010; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm11[4,5,6,7]6011; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6012; AVX2-FCP-NEXT:    vbroadcastss 84(%r8), %ymm26013; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4],ymm7[5],ymm2[6,7]6014; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm10[6,7]6015; AVX2-FCP-NEXT:    vextractf128 $1, %ymm3, %xmm36016; AVX2-FCP-NEXT:    vshufps {{.*#+}} ymm10 = ymm4[1,1,1,1,5,5,5,5]6017; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm10 = ymm10[0],ymm5[1],ymm10[2,3,4],ymm5[5],ymm10[6,7]6018; AVX2-FCP-NEXT:    vextractf128 $1, %ymm10, %xmm106019; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm3 = xmm10[0,1],xmm3[2,3]6020; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm2 = ymm3[0,1,2,3],ymm2[4,5,6,7]6021; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6022; AVX2-FCP-NEXT:    vbroadcastss 88(%r10), %ymm26023; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm2 = ymm8[0,1,2,3,4,5,6],ymm2[7]6024; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} ymm1 = ymm1[2],ymm7[2],ymm1[3],ymm7[3],ymm1[6],ymm7[6],ymm1[7],ymm7[7]6025; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} ymm3 = ymm4[2],ymm5[2],ymm4[3],ymm5[3],ymm4[6],ymm5[6],ymm4[7],ymm5[7]6026; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} ymm4 = ymm6[2],ymm0[2],ymm6[3],ymm0[3],ymm6[6],ymm0[6],ymm6[7],ymm0[7]6027; AVX2-FCP-NEXT:    vextractf128 $1, %ymm3, %xmm06028; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm5 = ymm4[2,2,2,2]6029; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm5[2,3]6030; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm2 = ymm1[0,1,2,3,4,5],ymm2[6,7]6031; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]6032; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6033; AVX2-FCP-NEXT:    vbroadcastss 92(%rax), %ymm06034; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} ymm2 = ymm0[2],ymm9[2],ymm0[3],ymm9[3],ymm0[6],ymm9[6],ymm0[7],ymm9[7]6035; AVX2-FCP-NEXT:    vmovaps 96(%rdi), %ymm06036; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm5 = ymm1[1],ymm2[1],ymm1[3],ymm2[3]6037; AVX2-FCP-NEXT:    vmovaps 96(%rsi), %ymm16038; AVX2-FCP-NEXT:    vextractf128 $1, %ymm4, %xmm26039; AVX2-FCP-NEXT:    vshufps {{.*#+}} ymm3 = ymm3[2,3,2,3,6,7,6,7]6040; AVX2-FCP-NEXT:    vextractf128 $1, %ymm3, %xmm36041; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm3 = xmm3[0,1],xmm2[2,3]6042; AVX2-FCP-NEXT:    vmovaps 96(%rdx), %ymm26043; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm5[4,5,6,7]6044; AVX2-FCP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6045; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} ymm3 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]6046; AVX2-FCP-NEXT:    vextractf128 $1, %ymm3, %xmm36047; AVX2-FCP-NEXT:    vmovaps 96(%rcx), %ymm46048; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} ymm5 = ymm2[0],ymm4[0],ymm2[1],ymm4[1],ymm2[4],ymm4[4],ymm2[5],ymm4[5]6049; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm6 = ymm5[2,2,2,2]6050; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm3 = xmm3[0,1],xmm6[2,3]6051; AVX2-FCP-NEXT:    vmovaps 96(%r8), %ymm66052; AVX2-FCP-NEXT:    vmovaps 96(%r9), %ymm76053; AVX2-FCP-NEXT:    vmovaps 96(%rax), %ymm86054; AVX2-FCP-NEXT:    vmovaps 96(%r10), %ymm96055; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} ymm10 = ymm8[0],ymm9[0],ymm8[1],ymm9[1],ymm8[4],ymm9[4],ymm8[5],ymm9[5]6056; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} ymm11 = ymm6[0],ymm7[0],ymm6[1],ymm7[1],ymm6[4],ymm7[4],ymm6[5],ymm7[5]6057; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm11 = ymm11[0],ymm10[0],ymm11[2],ymm10[2]6058; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm11[4,5,6,7]6059; AVX2-FCP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6060; AVX2-FCP-NEXT:    vbroadcastss 116(%r8), %ymm36061; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4],ymm7[5],ymm3[6,7]6062; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm10[6,7]6063; AVX2-FCP-NEXT:    vextractf128 $1, %ymm5, %xmm56064; AVX2-FCP-NEXT:    vshufps {{.*#+}} ymm10 = ymm0[1,1,1,1,5,5,5,5]6065; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm10 = ymm10[0],ymm1[1],ymm10[2,3,4],ymm1[5],ymm10[6,7]6066; AVX2-FCP-NEXT:    vextractf128 $1, %ymm10, %xmm106067; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm5 = xmm10[0,1],xmm5[2,3]6068; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm3 = ymm5[0,1,2,3],ymm3[4,5,6,7]6069; AVX2-FCP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6070; AVX2-FCP-NEXT:    vbroadcastss 120(%r10), %ymm36071; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm3 = ymm8[0,1,2,3,4,5,6],ymm3[7]6072; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} ymm5 = ymm6[2],ymm7[2],ymm6[3],ymm7[3],ymm6[6],ymm7[6],ymm6[7],ymm7[7]6073; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} ymm0 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7]6074; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} ymm1 = ymm2[2],ymm4[2],ymm2[3],ymm4[3],ymm2[6],ymm4[6],ymm2[7],ymm4[7]6075; AVX2-FCP-NEXT:    vextractf128 $1, %ymm0, %xmm26076; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm4 = ymm1[2,2,2,2]6077; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm2 = xmm2[0,1],xmm4[2,3]6078; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm3 = ymm5[0,1,2,3,4,5],ymm3[6,7]6079; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]6080; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6081; AVX2-FCP-NEXT:    vbroadcastss 124(%rax), %ymm26082; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} ymm2 = ymm2[2],ymm9[2],ymm2[3],ymm9[3],ymm2[6],ymm9[6],ymm2[7],ymm9[7]6083; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm5[1],ymm2[1],ymm5[3],ymm2[3]6084; AVX2-FCP-NEXT:    vextractf128 $1, %ymm1, %xmm16085; AVX2-FCP-NEXT:    vshufps {{.*#+}} ymm0 = ymm0[2,3,2,3,6,7,6,7]6086; AVX2-FCP-NEXT:    vextractf128 $1, %ymm0, %xmm06087; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]6088; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]6089; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6090; AVX2-FCP-NEXT:    vmovaps (%rcx), %xmm06091; AVX2-FCP-NEXT:    vbroadcastss %xmm0, %xmm26092; AVX2-FCP-NEXT:    vmovaps (%rdx), %xmm16093; AVX2-FCP-NEXT:    vbroadcastss %xmm1, %xmm36094; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]6095; AVX2-FCP-NEXT:    vmovaps (%rsi), %xmm46096; AVX2-FCP-NEXT:    vmovaps (%rdi), %xmm56097; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm3 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]6098; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm6 = xmm3[0,1],xmm2[2,3]6099; AVX2-FCP-NEXT:    vmovaps (%r10), %xmm26100; AVX2-FCP-NEXT:    vbroadcastss %xmm2, %xmm76101; AVX2-FCP-NEXT:    vmovaps (%rax), %xmm36102; AVX2-FCP-NEXT:    vbroadcastss %xmm3, %xmm86103; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm8[0],xmm7[0],xmm8[1],xmm7[1]6104; AVX2-FCP-NEXT:    vmovaps (%r9), %xmm86105; AVX2-FCP-NEXT:    vmovaps (%r8), %xmm96106; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm10 = xmm9[0],xmm8[0],xmm9[1],xmm8[1]6107; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm10, %ymm0, %ymm106108; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm7, %ymm0, %ymm76109; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm7 = ymm10[0,1,2,3,4,5],ymm7[6,7]6110; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7]6111; AVX2-FCP-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6112; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm6 = xmm5[1,1,1,1]6113; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm6 = xmm6[0],xmm4[1],xmm6[2,3]6114; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]6115; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm6 = xmm6[0,1],xmm7[2,3]6116; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]6117; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm7, %ymm0, %ymm76118; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm10 = xmm9[1,1,1,1]6119; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm10 = xmm10[0],xmm8[1],xmm10[2,3]6120; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm10, %ymm0, %ymm106121; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm7 = ymm10[0,1,2,3,4,5],ymm7[6,7]6122; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7]6123; AVX2-FCP-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6124; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} xmm4 = xmm5[2],xmm4[2],xmm5[3],xmm4[3]6125; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} xmm5 = xmm9[2],xmm8[2],xmm9[3],xmm8[3]6126; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm6 = xmm2[2,2,2,2]6127; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm6 = xmm3[0,1,2],xmm6[3]6128; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm6, %ymm0, %ymm66129; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm5, %ymm0, %ymm76130; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3,4,5],ymm6[6,7]6131; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm7 = xmm0[2,2,2,2]6132; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm7 = xmm1[0,1,2],xmm7[3]6133; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm7 = xmm4[0,1],xmm7[2,3]6134; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]6135; AVX2-FCP-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6136; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm1[2],xmm0[2],xmm1[3],xmm0[3]6137; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm4[1],xmm0[1]6138; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} xmm1 = xmm3[2],xmm2[2],xmm3[3],xmm2[3]6139; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm16140; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm2 = xmm5[2,3,2,3]6141; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm26142; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]6143; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]6144; AVX2-FCP-NEXT:    vmovups %ymm0, (%rsp) # 32-byte Spill6145; AVX2-FCP-NEXT:    vmovaps 32(%rcx), %xmm06146; AVX2-FCP-NEXT:    vmovaps 32(%rdx), %xmm16147; AVX2-FCP-NEXT:    vbroadcastss %xmm0, %xmm26148; AVX2-FCP-NEXT:    vbroadcastss %xmm1, %xmm36149; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]6150; AVX2-FCP-NEXT:    vmovaps 32(%rsi), %xmm46151; AVX2-FCP-NEXT:    vmovaps 32(%rdi), %xmm56152; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm3 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]6153; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm6 = xmm3[0,1],xmm2[2,3]6154; AVX2-FCP-NEXT:    vmovaps 32(%r10), %xmm26155; AVX2-FCP-NEXT:    vmovaps 32(%rax), %xmm36156; AVX2-FCP-NEXT:    vbroadcastss %xmm2, %xmm76157; AVX2-FCP-NEXT:    vbroadcastss %xmm3, %xmm86158; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm8[0],xmm7[0],xmm8[1],xmm7[1]6159; AVX2-FCP-NEXT:    vmovaps 32(%r9), %xmm86160; AVX2-FCP-NEXT:    vmovaps 32(%r8), %xmm96161; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm10 = xmm9[0],xmm8[0],xmm9[1],xmm8[1]6162; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm10, %ymm0, %ymm106163; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm7, %ymm0, %ymm76164; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm7 = ymm10[0,1,2,3,4,5],ymm7[6,7]6165; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7]6166; AVX2-FCP-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6167; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm6 = xmm5[1,1,1,1]6168; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm6 = xmm6[0],xmm4[1],xmm6[2,3]6169; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]6170; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm6 = xmm6[0,1],xmm7[2,3]6171; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]6172; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm7, %ymm0, %ymm76173; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm10 = xmm9[1,1,1,1]6174; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm10 = xmm10[0],xmm8[1],xmm10[2,3]6175; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm10, %ymm0, %ymm106176; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm7 = ymm10[0,1,2,3,4,5],ymm7[6,7]6177; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7]6178; AVX2-FCP-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6179; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} xmm4 = xmm5[2],xmm4[2],xmm5[3],xmm4[3]6180; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} xmm5 = xmm9[2],xmm8[2],xmm9[3],xmm8[3]6181; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm6 = xmm2[2,2,2,2]6182; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm6 = xmm3[0,1,2],xmm6[3]6183; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm6, %ymm0, %ymm66184; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm5, %ymm0, %ymm76185; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3,4,5],ymm6[6,7]6186; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm7 = xmm0[2,2,2,2]6187; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm7 = xmm1[0,1,2],xmm7[3]6188; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm7 = xmm4[0,1],xmm7[2,3]6189; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]6190; AVX2-FCP-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6191; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm1[2],xmm0[2],xmm1[3],xmm0[3]6192; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm4[1],xmm0[1]6193; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} xmm1 = xmm3[2],xmm2[2],xmm3[3],xmm2[3]6194; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm16195; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm2 = xmm5[2,3,2,3]6196; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm26197; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]6198; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]6199; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6200; AVX2-FCP-NEXT:    vmovaps 64(%rcx), %xmm36201; AVX2-FCP-NEXT:    vmovaps 64(%rdx), %xmm16202; AVX2-FCP-NEXT:    vbroadcastss %xmm3, %xmm06203; AVX2-FCP-NEXT:    vbroadcastss %xmm1, %xmm26204; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm0 = xmm2[0],xmm0[0],xmm2[1],xmm0[1]6205; AVX2-FCP-NEXT:    vmovaps 64(%rsi), %xmm46206; AVX2-FCP-NEXT:    vmovaps 64(%rdi), %xmm56207; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm2 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]6208; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm9 = xmm2[0,1],xmm0[2,3]6209; AVX2-FCP-NEXT:    vmovaps 64(%r10), %xmm26210; AVX2-FCP-NEXT:    vmovaps 64(%rax), %xmm06211; AVX2-FCP-NEXT:    vbroadcastss %xmm2, %xmm76212; AVX2-FCP-NEXT:    vbroadcastss %xmm0, %xmm86213; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm8[0],xmm7[0],xmm8[1],xmm7[1]6214; AVX2-FCP-NEXT:    vmovaps 64(%r9), %xmm86215; AVX2-FCP-NEXT:    vmovaps 64(%r8), %xmm66216; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm10 = xmm6[0],xmm8[0],xmm6[1],xmm8[1]6217; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm10, %ymm0, %ymm106218; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm7, %ymm0, %ymm76219; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm7 = ymm10[0,1,2,3,4,5],ymm7[6,7]6220; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm11 = ymm9[0,1,2,3],ymm7[4,5,6,7]6221; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm7 = xmm5[1,1,1,1]6222; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm7 = xmm7[0],xmm4[1],xmm7[2,3]6223; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm9 = xmm1[0],xmm3[0],xmm1[1],xmm3[1]6224; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm7 = xmm7[0,1],xmm9[2,3]6225; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm9 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]6226; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm9, %ymm0, %ymm96227; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm15 = xmm6[1,1,1,1]6228; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm15 = xmm15[0],xmm8[1],xmm15[2,3]6229; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm15, %ymm0, %ymm156230; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm9 = ymm15[0,1,2,3,4,5],ymm9[6,7]6231; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm10 = ymm7[0,1,2,3],ymm9[4,5,6,7]6232; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} xmm4 = xmm5[2],xmm4[2],xmm5[3],xmm4[3]6233; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} xmm5 = xmm6[2],xmm8[2],xmm6[3],xmm8[3]6234; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm6 = xmm2[2,2,2,2]6235; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm6 = xmm0[0,1,2],xmm6[3]6236; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm6, %ymm0, %ymm66237; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm5, %ymm0, %ymm76238; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3,4,5],ymm6[6,7]6239; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm7 = xmm3[2,2,2,2]6240; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm7 = xmm1[0,1,2],xmm7[3]6241; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm7 = xmm4[0,1],xmm7[2,3]6242; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm9 = ymm7[0,1,2,3],ymm6[4,5,6,7]6243; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} xmm1 = xmm1[2],xmm3[2],xmm1[3],xmm3[3]6244; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} xmm1 = xmm4[1],xmm1[1]6245; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm0[2],xmm2[2],xmm0[3],xmm2[3]6246; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm06247; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm2 = xmm5[2,3,2,3]6248; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm26249; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3,4,5],ymm0[6,7]6250; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm8 = ymm1[0,1,2,3],ymm0[4,5,6,7]6251; AVX2-FCP-NEXT:    vmovaps 96(%rcx), %xmm66252; AVX2-FCP-NEXT:    vbroadcastss %xmm6, %xmm06253; AVX2-FCP-NEXT:    vmovaps 96(%rdx), %xmm56254; AVX2-FCP-NEXT:    vbroadcastss %xmm5, %xmm16255; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]6256; AVX2-FCP-NEXT:    vmovaps 96(%rsi), %xmm46257; AVX2-FCP-NEXT:    vmovaps 96(%rdi), %xmm26258; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm3 = xmm2[0],xmm4[0],xmm2[1],xmm4[1]6259; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm7 = xmm3[0,1],xmm0[2,3]6260; AVX2-FCP-NEXT:    vmovaps 96(%r10), %xmm36261; AVX2-FCP-NEXT:    vbroadcastss %xmm3, %xmm156262; AVX2-FCP-NEXT:    vmovaps 96(%rax), %xmm16263; AVX2-FCP-NEXT:    vbroadcastss %xmm1, %xmm146264; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm14 = xmm14[0],xmm15[0],xmm14[1],xmm15[1]6265; AVX2-FCP-NEXT:    vmovaps 96(%r9), %xmm156266; AVX2-FCP-NEXT:    vmovaps 96(%r8), %xmm06267; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm13 = xmm0[0],xmm15[0],xmm0[1],xmm15[1]6268; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm13, %ymm0, %ymm136269; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm14, %ymm0, %ymm146270; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm13 = ymm13[0,1,2,3,4,5],ymm14[6,7]6271; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm14 = ymm7[0,1,2,3],ymm13[4,5,6,7]6272; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm7 = xmm2[1,1,1,1]6273; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm7 = xmm7[0],xmm4[1],xmm7[2,3]6274; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm13 = xmm5[0],xmm6[0],xmm5[1],xmm6[1]6275; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm7 = xmm7[0,1],xmm13[2,3]6276; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm13 = xmm1[0],xmm3[0],xmm1[1],xmm3[1]6277; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm13, %ymm0, %ymm136278; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm12 = xmm0[1,1,1,1]6279; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm12 = xmm12[0],xmm15[1],xmm12[2,3]6280; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm12, %ymm0, %ymm126281; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm12 = ymm12[0,1,2,3,4,5],ymm13[6,7]6282; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm7 = ymm7[0,1,2,3],ymm12[4,5,6,7]6283; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} xmm2 = xmm2[2],xmm4[2],xmm2[3],xmm4[3]6284; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm0[2],xmm15[2],xmm0[3],xmm15[3]6285; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm4 = xmm3[2,2,2,2]6286; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm4 = xmm1[0,1,2],xmm4[3]6287; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm4, %ymm0, %ymm46288; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm126289; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm4 = ymm12[0,1,2,3,4,5],ymm4[6,7]6290; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm12 = xmm6[2,2,2,2]6291; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm12 = xmm5[0,1,2],xmm12[3]6292; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm12 = xmm2[0,1],xmm12[2,3]6293; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm4 = ymm12[0,1,2,3],ymm4[4,5,6,7]6294; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} xmm5 = xmm5[2],xmm6[2],xmm5[3],xmm6[3]6295; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} xmm2 = xmm2[1],xmm5[1]6296; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} xmm1 = xmm1[2],xmm3[2],xmm1[3],xmm3[3]6297; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm16298; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm0 = xmm0[2,3,2,3]6299; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm06300; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]6301; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]6302; AVX2-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax6303; AVX2-FCP-NEXT:    vmovaps %ymm0, 864(%rax)6304; AVX2-FCP-NEXT:    vmovaps %ymm4, 832(%rax)6305; AVX2-FCP-NEXT:    vmovaps %ymm7, 800(%rax)6306; AVX2-FCP-NEXT:    vmovaps %ymm14, 768(%rax)6307; AVX2-FCP-NEXT:    vmovaps %ymm8, 608(%rax)6308; AVX2-FCP-NEXT:    vmovaps %ymm9, 576(%rax)6309; AVX2-FCP-NEXT:    vmovaps %ymm10, 544(%rax)6310; AVX2-FCP-NEXT:    vmovaps %ymm11, 512(%rax)6311; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6312; AVX2-FCP-NEXT:    vmovaps %ymm0, 352(%rax)6313; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6314; AVX2-FCP-NEXT:    vmovaps %ymm0, 320(%rax)6315; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6316; AVX2-FCP-NEXT:    vmovaps %ymm0, 288(%rax)6317; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6318; AVX2-FCP-NEXT:    vmovaps %ymm0, 256(%rax)6319; AVX2-FCP-NEXT:    vmovups (%rsp), %ymm0 # 32-byte Reload6320; AVX2-FCP-NEXT:    vmovaps %ymm0, 96(%rax)6321; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6322; AVX2-FCP-NEXT:    vmovaps %ymm0, 64(%rax)6323; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6324; AVX2-FCP-NEXT:    vmovaps %ymm0, 32(%rax)6325; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6326; AVX2-FCP-NEXT:    vmovaps %ymm0, (%rax)6327; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6328; AVX2-FCP-NEXT:    vmovaps %ymm0, 992(%rax)6329; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6330; AVX2-FCP-NEXT:    vmovaps %ymm0, 960(%rax)6331; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6332; AVX2-FCP-NEXT:    vmovaps %ymm0, 928(%rax)6333; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6334; AVX2-FCP-NEXT:    vmovaps %ymm0, 896(%rax)6335; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6336; AVX2-FCP-NEXT:    vmovaps %ymm0, 736(%rax)6337; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6338; AVX2-FCP-NEXT:    vmovaps %ymm0, 704(%rax)6339; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6340; AVX2-FCP-NEXT:    vmovaps %ymm0, 672(%rax)6341; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6342; AVX2-FCP-NEXT:    vmovaps %ymm0, 640(%rax)6343; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6344; AVX2-FCP-NEXT:    vmovaps %ymm0, 480(%rax)6345; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6346; AVX2-FCP-NEXT:    vmovaps %ymm0, 448(%rax)6347; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6348; AVX2-FCP-NEXT:    vmovaps %ymm0, 416(%rax)6349; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6350; AVX2-FCP-NEXT:    vmovaps %ymm0, 384(%rax)6351; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6352; AVX2-FCP-NEXT:    vmovaps %ymm0, 224(%rax)6353; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6354; AVX2-FCP-NEXT:    vmovaps %ymm0, 192(%rax)6355; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6356; AVX2-FCP-NEXT:    vmovaps %ymm0, 160(%rax)6357; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6358; AVX2-FCP-NEXT:    vmovaps %ymm0, 128(%rax)6359; AVX2-FCP-NEXT:    addq $648, %rsp # imm = 0x2886360; AVX2-FCP-NEXT:    vzeroupper6361; AVX2-FCP-NEXT:    retq6362;6363; AVX512-LABEL: store_i32_stride8_vf32:6364; AVX512:       # %bb.0:6365; AVX512-NEXT:    subq $2056, %rsp # imm = 0x8086366; AVX512-NEXT:    movq {{[0-9]+}}(%rsp), %rax6367; AVX512-NEXT:    movq {{[0-9]+}}(%rsp), %r106368; AVX512-NEXT:    vmovdqa64 (%r8), %zmm16369; AVX512-NEXT:    vmovdqa64 (%r9), %zmm276370; AVX512-NEXT:    vmovdqa64 (%r10), %zmm26371; AVX512-NEXT:    vmovdqa64 64(%r10), %zmm06372; AVX512-NEXT:    vmovdqa64 (%rax), %zmm306373; AVX512-NEXT:    vmovdqa64 64(%rax), %zmm286374; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm3 = [0,0,0,0,0,0,2,18,0,0,0,0,0,0,3,19]6375; AVX512-NEXT:    vmovdqa64 %zmm2, %zmm56376; AVX512-NEXT:    vmovdqa64 %zmm2, %zmm46377; AVX512-NEXT:    vpermt2d %zmm30, %zmm3, %zmm56378; AVX512-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6379; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [0,0,0,0,2,18,0,0,0,0,0,0,3,19,0,0]6380; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm56381; AVX512-NEXT:    vpermt2d %zmm27, %zmm2, %zmm56382; AVX512-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6383; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm5 = [0,0,0,0,0,0,0,16,0,0,0,0,0,0,1,17]6384; AVX512-NEXT:    vmovdqa64 %zmm4, %zmm66385; AVX512-NEXT:    vpermt2d %zmm30, %zmm5, %zmm66386; AVX512-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6387; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm6 = [0,0,0,0,0,16,0,0,0,0,0,0,1,17,0,0]6388; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm76389; AVX512-NEXT:    vpermt2d %zmm27, %zmm6, %zmm76390; AVX512-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6391; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm7 = [0,0,0,0,0,0,6,22,0,0,0,0,0,0,7,23]6392; AVX512-NEXT:    vmovdqa64 %zmm4, %zmm86393; AVX512-NEXT:    vpermt2d %zmm30, %zmm7, %zmm86394; AVX512-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6395; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm8 = [0,0,0,0,6,22,0,0,0,0,0,0,7,23,0,0]6396; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm96397; AVX512-NEXT:    vpermt2d %zmm27, %zmm8, %zmm96398; AVX512-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6399; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm9 = [0,0,0,0,0,0,4,20,0,0,0,0,0,0,5,21]6400; AVX512-NEXT:    vmovdqa64 %zmm4, %zmm106401; AVX512-NEXT:    vpermt2d %zmm30, %zmm9, %zmm106402; AVX512-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6403; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm10 = [0,0,0,0,4,20,0,0,0,0,0,0,5,21,0,0]6404; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm116405; AVX512-NEXT:    vpermt2d %zmm27, %zmm10, %zmm116406; AVX512-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6407; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm11 = [0,0,0,0,0,0,10,26,0,0,0,0,0,0,11,27]6408; AVX512-NEXT:    vmovdqa64 %zmm4, %zmm126409; AVX512-NEXT:    vpermt2d %zmm30, %zmm11, %zmm126410; AVX512-NEXT:    vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6411; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,0,0,10,26,0,0,0,0,0,0,11,27,0,0]6412; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm136413; AVX512-NEXT:    vpermt2d %zmm27, %zmm12, %zmm136414; AVX512-NEXT:    vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6415; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm13 = [0,0,0,0,0,0,8,24,0,0,0,0,0,0,9,25]6416; AVX512-NEXT:    vmovdqa64 %zmm4, %zmm146417; AVX512-NEXT:    vpermt2d %zmm30, %zmm13, %zmm146418; AVX512-NEXT:    vmovdqu64 %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6419; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm14 = [0,0,0,0,8,24,0,0,0,0,0,0,9,25,0,0]6420; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm15 = [0,0,0,0,0,0,14,30,0,0,0,0,0,0,15,31]6421; AVX512-NEXT:    vmovdqa64 %zmm4, %zmm166422; AVX512-NEXT:    vpermt2d %zmm30, %zmm15, %zmm166423; AVX512-NEXT:    vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6424; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm31 = [0,0,0,0,0,0,12,28,0,0,0,0,0,0,13,29]6425; AVX512-NEXT:    vpermt2d %zmm30, %zmm31, %zmm46426; AVX512-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6427; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm46428; AVX512-NEXT:    vpermt2d %zmm27, %zmm14, %zmm46429; AVX512-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6430; AVX512-NEXT:    vpermi2d %zmm28, %zmm0, %zmm36431; AVX512-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6432; AVX512-NEXT:    vpermi2d %zmm28, %zmm0, %zmm56433; AVX512-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6434; AVX512-NEXT:    vpermi2d %zmm28, %zmm0, %zmm76435; AVX512-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6436; AVX512-NEXT:    vpermi2d %zmm28, %zmm0, %zmm96437; AVX512-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6438; AVX512-NEXT:    vpermi2d %zmm28, %zmm0, %zmm116439; AVX512-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6440; AVX512-NEXT:    vpermi2d %zmm28, %zmm0, %zmm136441; AVX512-NEXT:    vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6442; AVX512-NEXT:    vpermi2d %zmm28, %zmm0, %zmm156443; AVX512-NEXT:    vmovdqu64 %zmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6444; AVX512-NEXT:    vpermt2d %zmm28, %zmm31, %zmm06445; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6446; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm28 = [0,0,0,0,14,30,0,0,0,0,0,0,15,31,0,0]6447; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm316448; AVX512-NEXT:    vpermt2d %zmm27, %zmm28, %zmm316449; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm0 = [0,0,0,0,12,28,0,0,0,0,0,0,13,29,0,0]6450; AVX512-NEXT:    vpermt2d %zmm27, %zmm0, %zmm16451; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6452; AVX512-NEXT:    vmovdqa64 64(%r8), %zmm276453; AVX512-NEXT:    vmovdqa64 64(%r9), %zmm16454; AVX512-NEXT:    vpermi2d %zmm1, %zmm27, %zmm26455; AVX512-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6456; AVX512-NEXT:    vpermi2d %zmm1, %zmm27, %zmm66457; AVX512-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6458; AVX512-NEXT:    vpermi2d %zmm1, %zmm27, %zmm86459; AVX512-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6460; AVX512-NEXT:    vpermi2d %zmm1, %zmm27, %zmm106461; AVX512-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6462; AVX512-NEXT:    vpermi2d %zmm1, %zmm27, %zmm126463; AVX512-NEXT:    vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6464; AVX512-NEXT:    vpermi2d %zmm1, %zmm27, %zmm146465; AVX512-NEXT:    vmovdqu64 %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6466; AVX512-NEXT:    vpermi2d %zmm1, %zmm27, %zmm286467; AVX512-NEXT:    vpermt2d %zmm1, %zmm0, %zmm276468; AVX512-NEXT:    vmovdqa64 (%rdx), %zmm166469; AVX512-NEXT:    vmovdqa64 (%rcx), %zmm06470; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [0,0,2,18,0,0,0,0,0,0,3,19,0,0,0,0]6471; AVX512-NEXT:    vmovdqa64 %zmm16, %zmm206472; AVX512-NEXT:    vpermt2d %zmm0, %zmm2, %zmm206473; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm3 = [0,0,0,16,0,0,0,0,0,0,1,17,0,0,0,0]6474; AVX512-NEXT:    vmovdqa64 %zmm16, %zmm216475; AVX512-NEXT:    vpermt2d %zmm0, %zmm3, %zmm216476; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [0,0,6,22,0,0,0,0,0,0,7,23,0,0,0,0]6477; AVX512-NEXT:    vmovdqa64 %zmm16, %zmm226478; AVX512-NEXT:    vpermt2d %zmm0, %zmm4, %zmm226479; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm6 = [0,0,4,20,0,0,0,0,0,0,5,21,0,0,0,0]6480; AVX512-NEXT:    vmovdqa64 %zmm16, %zmm236481; AVX512-NEXT:    vpermt2d %zmm0, %zmm6, %zmm236482; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm7 = [0,0,10,26,0,0,0,0,0,0,11,27,0,0,0,0]6483; AVX512-NEXT:    vmovdqa64 %zmm16, %zmm246484; AVX512-NEXT:    vpermt2d %zmm0, %zmm7, %zmm246485; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm29 = [0,0,8,24,0,0,0,0,0,0,9,25,0,0,0,0]6486; AVX512-NEXT:    vmovdqa64 %zmm16, %zmm256487; AVX512-NEXT:    vpermt2d %zmm0, %zmm29, %zmm256488; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm30 = [0,0,14,30,0,0,0,0,0,0,15,31,0,0,0,0]6489; AVX512-NEXT:    vmovdqa64 %zmm16, %zmm266490; AVX512-NEXT:    vpermt2d %zmm0, %zmm30, %zmm266491; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,12,28,0,0,0,0,0,0,13,29,0,0,0,0]6492; AVX512-NEXT:    vpermt2d %zmm0, %zmm1, %zmm166493; AVX512-NEXT:    vmovdqa64 64(%rdx), %zmm56494; AVX512-NEXT:    vmovdqa64 64(%rcx), %zmm06495; AVX512-NEXT:    vpermi2d %zmm0, %zmm5, %zmm26496; AVX512-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6497; AVX512-NEXT:    vpermi2d %zmm0, %zmm5, %zmm36498; AVX512-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6499; AVX512-NEXT:    vpermi2d %zmm0, %zmm5, %zmm46500; AVX512-NEXT:    vmovdqu64 %zmm4, (%rsp) # 64-byte Spill6501; AVX512-NEXT:    vpermi2d %zmm0, %zmm5, %zmm66502; AVX512-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6503; AVX512-NEXT:    vpermi2d %zmm0, %zmm5, %zmm76504; AVX512-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6505; AVX512-NEXT:    vpermi2d %zmm0, %zmm5, %zmm296506; AVX512-NEXT:    vpermi2d %zmm0, %zmm5, %zmm306507; AVX512-NEXT:    vpermt2d %zmm0, %zmm1, %zmm56508; AVX512-NEXT:    vmovdqa64 (%rdi), %zmm36509; AVX512-NEXT:    vmovdqa64 (%rsi), %zmm06510; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [2,18,0,0,0,0,0,0,3,19,0,0,0,0,0,0]6511; AVX512-NEXT:    vmovdqa64 %zmm3, %zmm186512; AVX512-NEXT:    vpermt2d %zmm0, %zmm12, %zmm186513; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm10 = [0,16,0,0,0,0,0,0,1,17,0,0,0,0,0,0]6514; AVX512-NEXT:    vmovdqa64 %zmm3, %zmm176515; AVX512-NEXT:    vpermt2d %zmm0, %zmm10, %zmm176516; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm8 = [6,22,0,0,0,0,0,0,7,23,0,0,0,0,0,0]6517; AVX512-NEXT:    vmovdqa64 %zmm3, %zmm156518; AVX512-NEXT:    vpermt2d %zmm0, %zmm8, %zmm156519; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm7 = [4,20,0,0,0,0,0,0,5,21,0,0,0,0,0,0]6520; AVX512-NEXT:    vmovdqa64 %zmm3, %zmm146521; AVX512-NEXT:    vpermt2d %zmm0, %zmm7, %zmm146522; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm6 = [10,26,0,0,0,0,0,0,11,27,0,0,0,0,0,0]6523; AVX512-NEXT:    vmovdqa64 %zmm3, %zmm136524; AVX512-NEXT:    vpermt2d %zmm0, %zmm6, %zmm136525; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [8,24,0,0,0,0,0,0,9,25,0,0,0,0,0,0]6526; AVX512-NEXT:    vmovdqa64 %zmm3, %zmm116527; AVX512-NEXT:    vpermt2d %zmm0, %zmm4, %zmm116528; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [14,30,0,0,0,0,0,0,15,31,0,0,0,0,0,0]6529; AVX512-NEXT:    vmovdqa64 %zmm3, %zmm96530; AVX512-NEXT:    vpermt2d %zmm0, %zmm2, %zmm96531; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm19 = [12,28,0,0,0,0,0,0,13,29,0,0,0,0,0,0]6532; AVX512-NEXT:    vpermt2d %zmm0, %zmm19, %zmm36533; AVX512-NEXT:    vmovdqa64 64(%rdi), %zmm16534; AVX512-NEXT:    vmovdqa64 64(%rsi), %zmm06535; AVX512-NEXT:    vpermi2d %zmm0, %zmm1, %zmm126536; AVX512-NEXT:    vpermi2d %zmm0, %zmm1, %zmm106537; AVX512-NEXT:    vpermi2d %zmm0, %zmm1, %zmm86538; AVX512-NEXT:    vpermi2d %zmm0, %zmm1, %zmm76539; AVX512-NEXT:    vpermi2d %zmm0, %zmm1, %zmm66540; AVX512-NEXT:    vpermi2d %zmm0, %zmm1, %zmm46541; AVX512-NEXT:    vpermi2d %zmm0, %zmm1, %zmm26542; AVX512-NEXT:    vpermt2d %zmm0, %zmm19, %zmm16543; AVX512-NEXT:    movb $-120, %al6544; AVX512-NEXT:    kmovw %eax, %k16545; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload6546; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload6547; AVX512-NEXT:    vmovdqa64 %zmm19, %zmm0 {%k1}6548; AVX512-NEXT:    movb $34, %al6549; AVX512-NEXT:    kmovw %eax, %k26550; AVX512-NEXT:    vmovdqa64 %zmm20, %zmm18 {%k2}6551; AVX512-NEXT:    movb $-52, %al6552; AVX512-NEXT:    kmovw %eax, %k36553; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm18 {%k3}6554; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload6555; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload6556; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm19 {%k1}6557; AVX512-NEXT:    vmovdqa64 %zmm21, %zmm17 {%k2}6558; AVX512-NEXT:    vmovdqa64 %zmm19, %zmm17 {%k3}6559; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload6560; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload6561; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm19 {%k1}6562; AVX512-NEXT:    vmovdqa64 %zmm22, %zmm15 {%k2}6563; AVX512-NEXT:    vmovdqa64 %zmm19, %zmm15 {%k3}6564; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload6565; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload6566; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm19 {%k1}6567; AVX512-NEXT:    vmovdqa64 %zmm23, %zmm14 {%k2}6568; AVX512-NEXT:    vmovdqa64 %zmm19, %zmm14 {%k3}6569; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload6570; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload6571; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm19 {%k1}6572; AVX512-NEXT:    vmovdqa64 %zmm24, %zmm13 {%k2}6573; AVX512-NEXT:    vmovdqa64 %zmm19, %zmm13 {%k3}6574; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload6575; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload6576; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm19 {%k1}6577; AVX512-NEXT:    vmovdqa64 %zmm25, %zmm11 {%k2}6578; AVX512-NEXT:    vmovdqa64 %zmm19, %zmm11 {%k3}6579; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload6580; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm31 {%k1}6581; AVX512-NEXT:    vmovdqa64 %zmm26, %zmm9 {%k2}6582; AVX512-NEXT:    vmovdqa64 %zmm31, %zmm9 {%k3}6583; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload6584; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload6585; AVX512-NEXT:    vmovdqa64 %zmm19, %zmm0 {%k1}6586; AVX512-NEXT:    vmovdqa64 %zmm16, %zmm3 {%k2}6587; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm3 {%k3}6588; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload6589; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload6590; AVX512-NEXT:    vmovdqa64 %zmm16, %zmm0 {%k1}6591; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload6592; AVX512-NEXT:    vmovdqa64 %zmm16, %zmm12 {%k2}6593; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm12 {%k3}6594; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload6595; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload6596; AVX512-NEXT:    vmovdqa64 %zmm16, %zmm0 {%k1}6597; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload6598; AVX512-NEXT:    vmovdqa64 %zmm16, %zmm10 {%k2}6599; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm10 {%k3}6600; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload6601; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload6602; AVX512-NEXT:    vmovdqa64 %zmm16, %zmm0 {%k1}6603; AVX512-NEXT:    vmovdqu64 (%rsp), %zmm16 # 64-byte Reload6604; AVX512-NEXT:    vmovdqa64 %zmm16, %zmm8 {%k2}6605; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm8 {%k3}6606; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload6607; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload6608; AVX512-NEXT:    vmovdqa64 %zmm16, %zmm0 {%k1}6609; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload6610; AVX512-NEXT:    vmovdqa64 %zmm16, %zmm7 {%k2}6611; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm7 {%k3}6612; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload6613; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload6614; AVX512-NEXT:    vmovdqa64 %zmm16, %zmm0 {%k1}6615; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload6616; AVX512-NEXT:    vmovdqa64 %zmm16, %zmm6 {%k2}6617; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm6 {%k3}6618; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload6619; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload6620; AVX512-NEXT:    vmovdqa64 %zmm16, %zmm0 {%k1}6621; AVX512-NEXT:    vmovdqa64 %zmm29, %zmm4 {%k2}6622; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm4 {%k3}6623; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload6624; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm28 {%k1}6625; AVX512-NEXT:    vmovdqa64 %zmm30, %zmm2 {%k2}6626; AVX512-NEXT:    vmovdqa64 %zmm28, %zmm2 {%k3}6627; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload6628; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm27 {%k1}6629; AVX512-NEXT:    vmovdqa64 %zmm5, %zmm1 {%k2}6630; AVX512-NEXT:    vmovdqa64 %zmm27, %zmm1 {%k3}6631; AVX512-NEXT:    movq {{[0-9]+}}(%rsp), %rax6632; AVX512-NEXT:    vmovdqa64 %zmm1, 896(%rax)6633; AVX512-NEXT:    vmovdqa64 %zmm2, 960(%rax)6634; AVX512-NEXT:    vmovdqa64 %zmm4, 768(%rax)6635; AVX512-NEXT:    vmovdqa64 %zmm6, 832(%rax)6636; AVX512-NEXT:    vmovdqa64 %zmm7, 640(%rax)6637; AVX512-NEXT:    vmovdqa64 %zmm8, 704(%rax)6638; AVX512-NEXT:    vmovdqa64 %zmm10, 512(%rax)6639; AVX512-NEXT:    vmovdqa64 %zmm12, 576(%rax)6640; AVX512-NEXT:    vmovdqa64 %zmm3, 384(%rax)6641; AVX512-NEXT:    vmovdqa64 %zmm9, 448(%rax)6642; AVX512-NEXT:    vmovdqa64 %zmm11, 256(%rax)6643; AVX512-NEXT:    vmovdqa64 %zmm13, 320(%rax)6644; AVX512-NEXT:    vmovdqa64 %zmm14, 128(%rax)6645; AVX512-NEXT:    vmovdqa64 %zmm15, 192(%rax)6646; AVX512-NEXT:    vmovdqa64 %zmm17, (%rax)6647; AVX512-NEXT:    vmovdqa64 %zmm18, 64(%rax)6648; AVX512-NEXT:    addq $2056, %rsp # imm = 0x8086649; AVX512-NEXT:    vzeroupper6650; AVX512-NEXT:    retq6651;6652; AVX512-FCP-LABEL: store_i32_stride8_vf32:6653; AVX512-FCP:       # %bb.0:6654; AVX512-FCP-NEXT:    subq $2056, %rsp # imm = 0x8086655; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax6656; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r106657; AVX512-FCP-NEXT:    vmovdqa64 (%r8), %zmm16658; AVX512-FCP-NEXT:    vmovdqa64 (%r9), %zmm276659; AVX512-FCP-NEXT:    vmovdqa64 (%r10), %zmm26660; AVX512-FCP-NEXT:    vmovdqa64 64(%r10), %zmm06661; AVX512-FCP-NEXT:    vmovdqa64 (%rax), %zmm306662; AVX512-FCP-NEXT:    vmovdqa64 64(%rax), %zmm286663; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm3 = [0,0,0,0,0,0,2,18,0,0,0,0,0,0,3,19]6664; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, %zmm56665; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, %zmm46666; AVX512-FCP-NEXT:    vpermt2d %zmm30, %zmm3, %zmm56667; AVX512-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6668; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [0,0,0,0,2,18,0,0,0,0,0,0,3,19,0,0]6669; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm56670; AVX512-FCP-NEXT:    vpermt2d %zmm27, %zmm2, %zmm56671; AVX512-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6672; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm5 = [0,0,0,0,0,0,0,16,0,0,0,0,0,0,1,17]6673; AVX512-FCP-NEXT:    vmovdqa64 %zmm4, %zmm66674; AVX512-FCP-NEXT:    vpermt2d %zmm30, %zmm5, %zmm66675; AVX512-FCP-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6676; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm6 = [0,0,0,0,0,16,0,0,0,0,0,0,1,17,0,0]6677; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm76678; AVX512-FCP-NEXT:    vpermt2d %zmm27, %zmm6, %zmm76679; AVX512-FCP-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6680; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm7 = [0,0,0,0,0,0,6,22,0,0,0,0,0,0,7,23]6681; AVX512-FCP-NEXT:    vmovdqa64 %zmm4, %zmm86682; AVX512-FCP-NEXT:    vpermt2d %zmm30, %zmm7, %zmm86683; AVX512-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6684; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm8 = [0,0,0,0,6,22,0,0,0,0,0,0,7,23,0,0]6685; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm96686; AVX512-FCP-NEXT:    vpermt2d %zmm27, %zmm8, %zmm96687; AVX512-FCP-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6688; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm9 = [0,0,0,0,0,0,4,20,0,0,0,0,0,0,5,21]6689; AVX512-FCP-NEXT:    vmovdqa64 %zmm4, %zmm106690; AVX512-FCP-NEXT:    vpermt2d %zmm30, %zmm9, %zmm106691; AVX512-FCP-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6692; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm10 = [0,0,0,0,4,20,0,0,0,0,0,0,5,21,0,0]6693; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm116694; AVX512-FCP-NEXT:    vpermt2d %zmm27, %zmm10, %zmm116695; AVX512-FCP-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6696; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm11 = [0,0,0,0,0,0,10,26,0,0,0,0,0,0,11,27]6697; AVX512-FCP-NEXT:    vmovdqa64 %zmm4, %zmm126698; AVX512-FCP-NEXT:    vpermt2d %zmm30, %zmm11, %zmm126699; AVX512-FCP-NEXT:    vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6700; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,0,0,10,26,0,0,0,0,0,0,11,27,0,0]6701; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm136702; AVX512-FCP-NEXT:    vpermt2d %zmm27, %zmm12, %zmm136703; AVX512-FCP-NEXT:    vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6704; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm13 = [0,0,0,0,0,0,8,24,0,0,0,0,0,0,9,25]6705; AVX512-FCP-NEXT:    vmovdqa64 %zmm4, %zmm146706; AVX512-FCP-NEXT:    vpermt2d %zmm30, %zmm13, %zmm146707; AVX512-FCP-NEXT:    vmovdqu64 %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6708; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm14 = [0,0,0,0,8,24,0,0,0,0,0,0,9,25,0,0]6709; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm15 = [0,0,0,0,0,0,14,30,0,0,0,0,0,0,15,31]6710; AVX512-FCP-NEXT:    vmovdqa64 %zmm4, %zmm166711; AVX512-FCP-NEXT:    vpermt2d %zmm30, %zmm15, %zmm166712; AVX512-FCP-NEXT:    vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6713; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm31 = [0,0,0,0,0,0,12,28,0,0,0,0,0,0,13,29]6714; AVX512-FCP-NEXT:    vpermt2d %zmm30, %zmm31, %zmm46715; AVX512-FCP-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6716; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm46717; AVX512-FCP-NEXT:    vpermt2d %zmm27, %zmm14, %zmm46718; AVX512-FCP-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6719; AVX512-FCP-NEXT:    vpermi2d %zmm28, %zmm0, %zmm36720; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6721; AVX512-FCP-NEXT:    vpermi2d %zmm28, %zmm0, %zmm56722; AVX512-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6723; AVX512-FCP-NEXT:    vpermi2d %zmm28, %zmm0, %zmm76724; AVX512-FCP-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6725; AVX512-FCP-NEXT:    vpermi2d %zmm28, %zmm0, %zmm96726; AVX512-FCP-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6727; AVX512-FCP-NEXT:    vpermi2d %zmm28, %zmm0, %zmm116728; AVX512-FCP-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6729; AVX512-FCP-NEXT:    vpermi2d %zmm28, %zmm0, %zmm136730; AVX512-FCP-NEXT:    vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6731; AVX512-FCP-NEXT:    vpermi2d %zmm28, %zmm0, %zmm156732; AVX512-FCP-NEXT:    vmovdqu64 %zmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6733; AVX512-FCP-NEXT:    vpermt2d %zmm28, %zmm31, %zmm06734; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6735; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm28 = [0,0,0,0,14,30,0,0,0,0,0,0,15,31,0,0]6736; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm316737; AVX512-FCP-NEXT:    vpermt2d %zmm27, %zmm28, %zmm316738; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm0 = [0,0,0,0,12,28,0,0,0,0,0,0,13,29,0,0]6739; AVX512-FCP-NEXT:    vpermt2d %zmm27, %zmm0, %zmm16740; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6741; AVX512-FCP-NEXT:    vmovdqa64 64(%r8), %zmm276742; AVX512-FCP-NEXT:    vmovdqa64 64(%r9), %zmm16743; AVX512-FCP-NEXT:    vpermi2d %zmm1, %zmm27, %zmm26744; AVX512-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6745; AVX512-FCP-NEXT:    vpermi2d %zmm1, %zmm27, %zmm66746; AVX512-FCP-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6747; AVX512-FCP-NEXT:    vpermi2d %zmm1, %zmm27, %zmm86748; AVX512-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6749; AVX512-FCP-NEXT:    vpermi2d %zmm1, %zmm27, %zmm106750; AVX512-FCP-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6751; AVX512-FCP-NEXT:    vpermi2d %zmm1, %zmm27, %zmm126752; AVX512-FCP-NEXT:    vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6753; AVX512-FCP-NEXT:    vpermi2d %zmm1, %zmm27, %zmm146754; AVX512-FCP-NEXT:    vmovdqu64 %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6755; AVX512-FCP-NEXT:    vpermi2d %zmm1, %zmm27, %zmm286756; AVX512-FCP-NEXT:    vpermt2d %zmm1, %zmm0, %zmm276757; AVX512-FCP-NEXT:    vmovdqa64 (%rdx), %zmm166758; AVX512-FCP-NEXT:    vmovdqa64 (%rcx), %zmm06759; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [0,0,2,18,0,0,0,0,0,0,3,19,0,0,0,0]6760; AVX512-FCP-NEXT:    vmovdqa64 %zmm16, %zmm206761; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm2, %zmm206762; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm3 = [0,0,0,16,0,0,0,0,0,0,1,17,0,0,0,0]6763; AVX512-FCP-NEXT:    vmovdqa64 %zmm16, %zmm216764; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm3, %zmm216765; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [0,0,6,22,0,0,0,0,0,0,7,23,0,0,0,0]6766; AVX512-FCP-NEXT:    vmovdqa64 %zmm16, %zmm226767; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm4, %zmm226768; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm6 = [0,0,4,20,0,0,0,0,0,0,5,21,0,0,0,0]6769; AVX512-FCP-NEXT:    vmovdqa64 %zmm16, %zmm236770; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm6, %zmm236771; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm7 = [0,0,10,26,0,0,0,0,0,0,11,27,0,0,0,0]6772; AVX512-FCP-NEXT:    vmovdqa64 %zmm16, %zmm246773; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm7, %zmm246774; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm29 = [0,0,8,24,0,0,0,0,0,0,9,25,0,0,0,0]6775; AVX512-FCP-NEXT:    vmovdqa64 %zmm16, %zmm256776; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm29, %zmm256777; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm30 = [0,0,14,30,0,0,0,0,0,0,15,31,0,0,0,0]6778; AVX512-FCP-NEXT:    vmovdqa64 %zmm16, %zmm266779; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm30, %zmm266780; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,12,28,0,0,0,0,0,0,13,29,0,0,0,0]6781; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm1, %zmm166782; AVX512-FCP-NEXT:    vmovdqa64 64(%rdx), %zmm56783; AVX512-FCP-NEXT:    vmovdqa64 64(%rcx), %zmm06784; AVX512-FCP-NEXT:    vpermi2d %zmm0, %zmm5, %zmm26785; AVX512-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6786; AVX512-FCP-NEXT:    vpermi2d %zmm0, %zmm5, %zmm36787; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6788; AVX512-FCP-NEXT:    vpermi2d %zmm0, %zmm5, %zmm46789; AVX512-FCP-NEXT:    vmovdqu64 %zmm4, (%rsp) # 64-byte Spill6790; AVX512-FCP-NEXT:    vpermi2d %zmm0, %zmm5, %zmm66791; AVX512-FCP-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6792; AVX512-FCP-NEXT:    vpermi2d %zmm0, %zmm5, %zmm76793; AVX512-FCP-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6794; AVX512-FCP-NEXT:    vpermi2d %zmm0, %zmm5, %zmm296795; AVX512-FCP-NEXT:    vpermi2d %zmm0, %zmm5, %zmm306796; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm1, %zmm56797; AVX512-FCP-NEXT:    vmovdqa64 (%rdi), %zmm36798; AVX512-FCP-NEXT:    vmovdqa64 (%rsi), %zmm06799; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [2,18,0,0,0,0,0,0,3,19,0,0,0,0,0,0]6800; AVX512-FCP-NEXT:    vmovdqa64 %zmm3, %zmm186801; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm12, %zmm186802; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm10 = [0,16,0,0,0,0,0,0,1,17,0,0,0,0,0,0]6803; AVX512-FCP-NEXT:    vmovdqa64 %zmm3, %zmm176804; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm10, %zmm176805; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm8 = [6,22,0,0,0,0,0,0,7,23,0,0,0,0,0,0]6806; AVX512-FCP-NEXT:    vmovdqa64 %zmm3, %zmm156807; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm8, %zmm156808; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm7 = [4,20,0,0,0,0,0,0,5,21,0,0,0,0,0,0]6809; AVX512-FCP-NEXT:    vmovdqa64 %zmm3, %zmm146810; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm7, %zmm146811; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm6 = [10,26,0,0,0,0,0,0,11,27,0,0,0,0,0,0]6812; AVX512-FCP-NEXT:    vmovdqa64 %zmm3, %zmm136813; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm6, %zmm136814; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [8,24,0,0,0,0,0,0,9,25,0,0,0,0,0,0]6815; AVX512-FCP-NEXT:    vmovdqa64 %zmm3, %zmm116816; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm4, %zmm116817; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [14,30,0,0,0,0,0,0,15,31,0,0,0,0,0,0]6818; AVX512-FCP-NEXT:    vmovdqa64 %zmm3, %zmm96819; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm2, %zmm96820; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm19 = [12,28,0,0,0,0,0,0,13,29,0,0,0,0,0,0]6821; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm19, %zmm36822; AVX512-FCP-NEXT:    vmovdqa64 64(%rdi), %zmm16823; AVX512-FCP-NEXT:    vmovdqa64 64(%rsi), %zmm06824; AVX512-FCP-NEXT:    vpermi2d %zmm0, %zmm1, %zmm126825; AVX512-FCP-NEXT:    vpermi2d %zmm0, %zmm1, %zmm106826; AVX512-FCP-NEXT:    vpermi2d %zmm0, %zmm1, %zmm86827; AVX512-FCP-NEXT:    vpermi2d %zmm0, %zmm1, %zmm76828; AVX512-FCP-NEXT:    vpermi2d %zmm0, %zmm1, %zmm66829; AVX512-FCP-NEXT:    vpermi2d %zmm0, %zmm1, %zmm46830; AVX512-FCP-NEXT:    vpermi2d %zmm0, %zmm1, %zmm26831; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm19, %zmm16832; AVX512-FCP-NEXT:    movb $-120, %al6833; AVX512-FCP-NEXT:    kmovw %eax, %k16834; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload6835; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload6836; AVX512-FCP-NEXT:    vmovdqa64 %zmm19, %zmm0 {%k1}6837; AVX512-FCP-NEXT:    movb $34, %al6838; AVX512-FCP-NEXT:    kmovw %eax, %k26839; AVX512-FCP-NEXT:    vmovdqa64 %zmm20, %zmm18 {%k2}6840; AVX512-FCP-NEXT:    movb $-52, %al6841; AVX512-FCP-NEXT:    kmovw %eax, %k36842; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm18 {%k3}6843; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload6844; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload6845; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm19 {%k1}6846; AVX512-FCP-NEXT:    vmovdqa64 %zmm21, %zmm17 {%k2}6847; AVX512-FCP-NEXT:    vmovdqa64 %zmm19, %zmm17 {%k3}6848; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload6849; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload6850; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm19 {%k1}6851; AVX512-FCP-NEXT:    vmovdqa64 %zmm22, %zmm15 {%k2}6852; AVX512-FCP-NEXT:    vmovdqa64 %zmm19, %zmm15 {%k3}6853; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload6854; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload6855; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm19 {%k1}6856; AVX512-FCP-NEXT:    vmovdqa64 %zmm23, %zmm14 {%k2}6857; AVX512-FCP-NEXT:    vmovdqa64 %zmm19, %zmm14 {%k3}6858; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload6859; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload6860; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm19 {%k1}6861; AVX512-FCP-NEXT:    vmovdqa64 %zmm24, %zmm13 {%k2}6862; AVX512-FCP-NEXT:    vmovdqa64 %zmm19, %zmm13 {%k3}6863; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload6864; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload6865; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm19 {%k1}6866; AVX512-FCP-NEXT:    vmovdqa64 %zmm25, %zmm11 {%k2}6867; AVX512-FCP-NEXT:    vmovdqa64 %zmm19, %zmm11 {%k3}6868; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload6869; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm31 {%k1}6870; AVX512-FCP-NEXT:    vmovdqa64 %zmm26, %zmm9 {%k2}6871; AVX512-FCP-NEXT:    vmovdqa64 %zmm31, %zmm9 {%k3}6872; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload6873; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload6874; AVX512-FCP-NEXT:    vmovdqa64 %zmm19, %zmm0 {%k1}6875; AVX512-FCP-NEXT:    vmovdqa64 %zmm16, %zmm3 {%k2}6876; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm3 {%k3}6877; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload6878; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload6879; AVX512-FCP-NEXT:    vmovdqa64 %zmm16, %zmm0 {%k1}6880; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload6881; AVX512-FCP-NEXT:    vmovdqa64 %zmm16, %zmm12 {%k2}6882; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm12 {%k3}6883; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload6884; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload6885; AVX512-FCP-NEXT:    vmovdqa64 %zmm16, %zmm0 {%k1}6886; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload6887; AVX512-FCP-NEXT:    vmovdqa64 %zmm16, %zmm10 {%k2}6888; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm10 {%k3}6889; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload6890; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload6891; AVX512-FCP-NEXT:    vmovdqa64 %zmm16, %zmm0 {%k1}6892; AVX512-FCP-NEXT:    vmovdqu64 (%rsp), %zmm16 # 64-byte Reload6893; AVX512-FCP-NEXT:    vmovdqa64 %zmm16, %zmm8 {%k2}6894; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm8 {%k3}6895; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload6896; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload6897; AVX512-FCP-NEXT:    vmovdqa64 %zmm16, %zmm0 {%k1}6898; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload6899; AVX512-FCP-NEXT:    vmovdqa64 %zmm16, %zmm7 {%k2}6900; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm7 {%k3}6901; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload6902; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload6903; AVX512-FCP-NEXT:    vmovdqa64 %zmm16, %zmm0 {%k1}6904; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload6905; AVX512-FCP-NEXT:    vmovdqa64 %zmm16, %zmm6 {%k2}6906; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm6 {%k3}6907; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload6908; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload6909; AVX512-FCP-NEXT:    vmovdqa64 %zmm16, %zmm0 {%k1}6910; AVX512-FCP-NEXT:    vmovdqa64 %zmm29, %zmm4 {%k2}6911; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm4 {%k3}6912; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload6913; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm28 {%k1}6914; AVX512-FCP-NEXT:    vmovdqa64 %zmm30, %zmm2 {%k2}6915; AVX512-FCP-NEXT:    vmovdqa64 %zmm28, %zmm2 {%k3}6916; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload6917; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm27 {%k1}6918; AVX512-FCP-NEXT:    vmovdqa64 %zmm5, %zmm1 {%k2}6919; AVX512-FCP-NEXT:    vmovdqa64 %zmm27, %zmm1 {%k3}6920; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax6921; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, 896(%rax)6922; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, 960(%rax)6923; AVX512-FCP-NEXT:    vmovdqa64 %zmm4, 768(%rax)6924; AVX512-FCP-NEXT:    vmovdqa64 %zmm6, 832(%rax)6925; AVX512-FCP-NEXT:    vmovdqa64 %zmm7, 640(%rax)6926; AVX512-FCP-NEXT:    vmovdqa64 %zmm8, 704(%rax)6927; AVX512-FCP-NEXT:    vmovdqa64 %zmm10, 512(%rax)6928; AVX512-FCP-NEXT:    vmovdqa64 %zmm12, 576(%rax)6929; AVX512-FCP-NEXT:    vmovdqa64 %zmm3, 384(%rax)6930; AVX512-FCP-NEXT:    vmovdqa64 %zmm9, 448(%rax)6931; AVX512-FCP-NEXT:    vmovdqa64 %zmm11, 256(%rax)6932; AVX512-FCP-NEXT:    vmovdqa64 %zmm13, 320(%rax)6933; AVX512-FCP-NEXT:    vmovdqa64 %zmm14, 128(%rax)6934; AVX512-FCP-NEXT:    vmovdqa64 %zmm15, 192(%rax)6935; AVX512-FCP-NEXT:    vmovdqa64 %zmm17, (%rax)6936; AVX512-FCP-NEXT:    vmovdqa64 %zmm18, 64(%rax)6937; AVX512-FCP-NEXT:    addq $2056, %rsp # imm = 0x8086938; AVX512-FCP-NEXT:    vzeroupper6939; AVX512-FCP-NEXT:    retq6940;6941; AVX512DQ-LABEL: store_i32_stride8_vf32:6942; AVX512DQ:       # %bb.0:6943; AVX512DQ-NEXT:    subq $2056, %rsp # imm = 0x8086944; AVX512DQ-NEXT:    movq {{[0-9]+}}(%rsp), %rax6945; AVX512DQ-NEXT:    movq {{[0-9]+}}(%rsp), %r106946; AVX512DQ-NEXT:    vmovdqa64 (%r8), %zmm16947; AVX512DQ-NEXT:    vmovdqa64 (%r9), %zmm276948; AVX512DQ-NEXT:    vmovdqa64 (%r10), %zmm26949; AVX512DQ-NEXT:    vmovdqa64 64(%r10), %zmm06950; AVX512DQ-NEXT:    vmovdqa64 (%rax), %zmm306951; AVX512DQ-NEXT:    vmovdqa64 64(%rax), %zmm286952; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm3 = [0,0,0,0,0,0,2,18,0,0,0,0,0,0,3,19]6953; AVX512DQ-NEXT:    vmovdqa64 %zmm2, %zmm56954; AVX512DQ-NEXT:    vmovdqa64 %zmm2, %zmm46955; AVX512DQ-NEXT:    vpermt2d %zmm30, %zmm3, %zmm56956; AVX512DQ-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6957; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [0,0,0,0,2,18,0,0,0,0,0,0,3,19,0,0]6958; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm56959; AVX512DQ-NEXT:    vpermt2d %zmm27, %zmm2, %zmm56960; AVX512DQ-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6961; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm5 = [0,0,0,0,0,0,0,16,0,0,0,0,0,0,1,17]6962; AVX512DQ-NEXT:    vmovdqa64 %zmm4, %zmm66963; AVX512DQ-NEXT:    vpermt2d %zmm30, %zmm5, %zmm66964; AVX512DQ-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6965; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm6 = [0,0,0,0,0,16,0,0,0,0,0,0,1,17,0,0]6966; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm76967; AVX512DQ-NEXT:    vpermt2d %zmm27, %zmm6, %zmm76968; AVX512DQ-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6969; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm7 = [0,0,0,0,0,0,6,22,0,0,0,0,0,0,7,23]6970; AVX512DQ-NEXT:    vmovdqa64 %zmm4, %zmm86971; AVX512DQ-NEXT:    vpermt2d %zmm30, %zmm7, %zmm86972; AVX512DQ-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6973; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm8 = [0,0,0,0,6,22,0,0,0,0,0,0,7,23,0,0]6974; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm96975; AVX512DQ-NEXT:    vpermt2d %zmm27, %zmm8, %zmm96976; AVX512DQ-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6977; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm9 = [0,0,0,0,0,0,4,20,0,0,0,0,0,0,5,21]6978; AVX512DQ-NEXT:    vmovdqa64 %zmm4, %zmm106979; AVX512DQ-NEXT:    vpermt2d %zmm30, %zmm9, %zmm106980; AVX512DQ-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6981; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm10 = [0,0,0,0,4,20,0,0,0,0,0,0,5,21,0,0]6982; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm116983; AVX512DQ-NEXT:    vpermt2d %zmm27, %zmm10, %zmm116984; AVX512DQ-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6985; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm11 = [0,0,0,0,0,0,10,26,0,0,0,0,0,0,11,27]6986; AVX512DQ-NEXT:    vmovdqa64 %zmm4, %zmm126987; AVX512DQ-NEXT:    vpermt2d %zmm30, %zmm11, %zmm126988; AVX512DQ-NEXT:    vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6989; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,0,0,10,26,0,0,0,0,0,0,11,27,0,0]6990; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm136991; AVX512DQ-NEXT:    vpermt2d %zmm27, %zmm12, %zmm136992; AVX512DQ-NEXT:    vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6993; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm13 = [0,0,0,0,0,0,8,24,0,0,0,0,0,0,9,25]6994; AVX512DQ-NEXT:    vmovdqa64 %zmm4, %zmm146995; AVX512DQ-NEXT:    vpermt2d %zmm30, %zmm13, %zmm146996; AVX512DQ-NEXT:    vmovdqu64 %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6997; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm14 = [0,0,0,0,8,24,0,0,0,0,0,0,9,25,0,0]6998; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm15 = [0,0,0,0,0,0,14,30,0,0,0,0,0,0,15,31]6999; AVX512DQ-NEXT:    vmovdqa64 %zmm4, %zmm167000; AVX512DQ-NEXT:    vpermt2d %zmm30, %zmm15, %zmm167001; AVX512DQ-NEXT:    vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7002; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm31 = [0,0,0,0,0,0,12,28,0,0,0,0,0,0,13,29]7003; AVX512DQ-NEXT:    vpermt2d %zmm30, %zmm31, %zmm47004; AVX512DQ-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7005; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm47006; AVX512DQ-NEXT:    vpermt2d %zmm27, %zmm14, %zmm47007; AVX512DQ-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7008; AVX512DQ-NEXT:    vpermi2d %zmm28, %zmm0, %zmm37009; AVX512DQ-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7010; AVX512DQ-NEXT:    vpermi2d %zmm28, %zmm0, %zmm57011; AVX512DQ-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7012; AVX512DQ-NEXT:    vpermi2d %zmm28, %zmm0, %zmm77013; AVX512DQ-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7014; AVX512DQ-NEXT:    vpermi2d %zmm28, %zmm0, %zmm97015; AVX512DQ-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7016; AVX512DQ-NEXT:    vpermi2d %zmm28, %zmm0, %zmm117017; AVX512DQ-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7018; AVX512DQ-NEXT:    vpermi2d %zmm28, %zmm0, %zmm137019; AVX512DQ-NEXT:    vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7020; AVX512DQ-NEXT:    vpermi2d %zmm28, %zmm0, %zmm157021; AVX512DQ-NEXT:    vmovdqu64 %zmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7022; AVX512DQ-NEXT:    vpermt2d %zmm28, %zmm31, %zmm07023; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7024; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm28 = [0,0,0,0,14,30,0,0,0,0,0,0,15,31,0,0]7025; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm317026; AVX512DQ-NEXT:    vpermt2d %zmm27, %zmm28, %zmm317027; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm0 = [0,0,0,0,12,28,0,0,0,0,0,0,13,29,0,0]7028; AVX512DQ-NEXT:    vpermt2d %zmm27, %zmm0, %zmm17029; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7030; AVX512DQ-NEXT:    vmovdqa64 64(%r8), %zmm277031; AVX512DQ-NEXT:    vmovdqa64 64(%r9), %zmm17032; AVX512DQ-NEXT:    vpermi2d %zmm1, %zmm27, %zmm27033; AVX512DQ-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7034; AVX512DQ-NEXT:    vpermi2d %zmm1, %zmm27, %zmm67035; AVX512DQ-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7036; AVX512DQ-NEXT:    vpermi2d %zmm1, %zmm27, %zmm87037; AVX512DQ-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7038; AVX512DQ-NEXT:    vpermi2d %zmm1, %zmm27, %zmm107039; AVX512DQ-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7040; AVX512DQ-NEXT:    vpermi2d %zmm1, %zmm27, %zmm127041; AVX512DQ-NEXT:    vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7042; AVX512DQ-NEXT:    vpermi2d %zmm1, %zmm27, %zmm147043; AVX512DQ-NEXT:    vmovdqu64 %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7044; AVX512DQ-NEXT:    vpermi2d %zmm1, %zmm27, %zmm287045; AVX512DQ-NEXT:    vpermt2d %zmm1, %zmm0, %zmm277046; AVX512DQ-NEXT:    vmovdqa64 (%rdx), %zmm167047; AVX512DQ-NEXT:    vmovdqa64 (%rcx), %zmm07048; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [0,0,2,18,0,0,0,0,0,0,3,19,0,0,0,0]7049; AVX512DQ-NEXT:    vmovdqa64 %zmm16, %zmm207050; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm2, %zmm207051; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm3 = [0,0,0,16,0,0,0,0,0,0,1,17,0,0,0,0]7052; AVX512DQ-NEXT:    vmovdqa64 %zmm16, %zmm217053; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm3, %zmm217054; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [0,0,6,22,0,0,0,0,0,0,7,23,0,0,0,0]7055; AVX512DQ-NEXT:    vmovdqa64 %zmm16, %zmm227056; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm4, %zmm227057; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm6 = [0,0,4,20,0,0,0,0,0,0,5,21,0,0,0,0]7058; AVX512DQ-NEXT:    vmovdqa64 %zmm16, %zmm237059; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm6, %zmm237060; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm7 = [0,0,10,26,0,0,0,0,0,0,11,27,0,0,0,0]7061; AVX512DQ-NEXT:    vmovdqa64 %zmm16, %zmm247062; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm7, %zmm247063; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm29 = [0,0,8,24,0,0,0,0,0,0,9,25,0,0,0,0]7064; AVX512DQ-NEXT:    vmovdqa64 %zmm16, %zmm257065; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm29, %zmm257066; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm30 = [0,0,14,30,0,0,0,0,0,0,15,31,0,0,0,0]7067; AVX512DQ-NEXT:    vmovdqa64 %zmm16, %zmm267068; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm30, %zmm267069; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,12,28,0,0,0,0,0,0,13,29,0,0,0,0]7070; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm1, %zmm167071; AVX512DQ-NEXT:    vmovdqa64 64(%rdx), %zmm57072; AVX512DQ-NEXT:    vmovdqa64 64(%rcx), %zmm07073; AVX512DQ-NEXT:    vpermi2d %zmm0, %zmm5, %zmm27074; AVX512DQ-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7075; AVX512DQ-NEXT:    vpermi2d %zmm0, %zmm5, %zmm37076; AVX512DQ-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7077; AVX512DQ-NEXT:    vpermi2d %zmm0, %zmm5, %zmm47078; AVX512DQ-NEXT:    vmovdqu64 %zmm4, (%rsp) # 64-byte Spill7079; AVX512DQ-NEXT:    vpermi2d %zmm0, %zmm5, %zmm67080; AVX512DQ-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7081; AVX512DQ-NEXT:    vpermi2d %zmm0, %zmm5, %zmm77082; AVX512DQ-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7083; AVX512DQ-NEXT:    vpermi2d %zmm0, %zmm5, %zmm297084; AVX512DQ-NEXT:    vpermi2d %zmm0, %zmm5, %zmm307085; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm1, %zmm57086; AVX512DQ-NEXT:    vmovdqa64 (%rdi), %zmm37087; AVX512DQ-NEXT:    vmovdqa64 (%rsi), %zmm07088; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [2,18,0,0,0,0,0,0,3,19,0,0,0,0,0,0]7089; AVX512DQ-NEXT:    vmovdqa64 %zmm3, %zmm187090; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm12, %zmm187091; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm10 = [0,16,0,0,0,0,0,0,1,17,0,0,0,0,0,0]7092; AVX512DQ-NEXT:    vmovdqa64 %zmm3, %zmm177093; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm10, %zmm177094; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm8 = [6,22,0,0,0,0,0,0,7,23,0,0,0,0,0,0]7095; AVX512DQ-NEXT:    vmovdqa64 %zmm3, %zmm157096; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm8, %zmm157097; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm7 = [4,20,0,0,0,0,0,0,5,21,0,0,0,0,0,0]7098; AVX512DQ-NEXT:    vmovdqa64 %zmm3, %zmm147099; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm7, %zmm147100; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm6 = [10,26,0,0,0,0,0,0,11,27,0,0,0,0,0,0]7101; AVX512DQ-NEXT:    vmovdqa64 %zmm3, %zmm137102; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm6, %zmm137103; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [8,24,0,0,0,0,0,0,9,25,0,0,0,0,0,0]7104; AVX512DQ-NEXT:    vmovdqa64 %zmm3, %zmm117105; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm4, %zmm117106; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [14,30,0,0,0,0,0,0,15,31,0,0,0,0,0,0]7107; AVX512DQ-NEXT:    vmovdqa64 %zmm3, %zmm97108; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm2, %zmm97109; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm19 = [12,28,0,0,0,0,0,0,13,29,0,0,0,0,0,0]7110; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm19, %zmm37111; AVX512DQ-NEXT:    vmovdqa64 64(%rdi), %zmm17112; AVX512DQ-NEXT:    vmovdqa64 64(%rsi), %zmm07113; AVX512DQ-NEXT:    vpermi2d %zmm0, %zmm1, %zmm127114; AVX512DQ-NEXT:    vpermi2d %zmm0, %zmm1, %zmm107115; AVX512DQ-NEXT:    vpermi2d %zmm0, %zmm1, %zmm87116; AVX512DQ-NEXT:    vpermi2d %zmm0, %zmm1, %zmm77117; AVX512DQ-NEXT:    vpermi2d %zmm0, %zmm1, %zmm67118; AVX512DQ-NEXT:    vpermi2d %zmm0, %zmm1, %zmm47119; AVX512DQ-NEXT:    vpermi2d %zmm0, %zmm1, %zmm27120; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm19, %zmm17121; AVX512DQ-NEXT:    movb $-120, %al7122; AVX512DQ-NEXT:    kmovw %eax, %k17123; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7124; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload7125; AVX512DQ-NEXT:    vmovdqa64 %zmm19, %zmm0 {%k1}7126; AVX512DQ-NEXT:    movb $34, %al7127; AVX512DQ-NEXT:    kmovw %eax, %k27128; AVX512DQ-NEXT:    vmovdqa64 %zmm20, %zmm18 {%k2}7129; AVX512DQ-NEXT:    movb $-52, %al7130; AVX512DQ-NEXT:    kmovw %eax, %k37131; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm18 {%k3}7132; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7133; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload7134; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm19 {%k1}7135; AVX512DQ-NEXT:    vmovdqa64 %zmm21, %zmm17 {%k2}7136; AVX512DQ-NEXT:    vmovdqa64 %zmm19, %zmm17 {%k3}7137; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7138; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload7139; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm19 {%k1}7140; AVX512DQ-NEXT:    vmovdqa64 %zmm22, %zmm15 {%k2}7141; AVX512DQ-NEXT:    vmovdqa64 %zmm19, %zmm15 {%k3}7142; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7143; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload7144; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm19 {%k1}7145; AVX512DQ-NEXT:    vmovdqa64 %zmm23, %zmm14 {%k2}7146; AVX512DQ-NEXT:    vmovdqa64 %zmm19, %zmm14 {%k3}7147; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7148; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload7149; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm19 {%k1}7150; AVX512DQ-NEXT:    vmovdqa64 %zmm24, %zmm13 {%k2}7151; AVX512DQ-NEXT:    vmovdqa64 %zmm19, %zmm13 {%k3}7152; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7153; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload7154; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm19 {%k1}7155; AVX512DQ-NEXT:    vmovdqa64 %zmm25, %zmm11 {%k2}7156; AVX512DQ-NEXT:    vmovdqa64 %zmm19, %zmm11 {%k3}7157; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7158; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm31 {%k1}7159; AVX512DQ-NEXT:    vmovdqa64 %zmm26, %zmm9 {%k2}7160; AVX512DQ-NEXT:    vmovdqa64 %zmm31, %zmm9 {%k3}7161; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7162; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload7163; AVX512DQ-NEXT:    vmovdqa64 %zmm19, %zmm0 {%k1}7164; AVX512DQ-NEXT:    vmovdqa64 %zmm16, %zmm3 {%k2}7165; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm3 {%k3}7166; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7167; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload7168; AVX512DQ-NEXT:    vmovdqa64 %zmm16, %zmm0 {%k1}7169; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload7170; AVX512DQ-NEXT:    vmovdqa64 %zmm16, %zmm12 {%k2}7171; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm12 {%k3}7172; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7173; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload7174; AVX512DQ-NEXT:    vmovdqa64 %zmm16, %zmm0 {%k1}7175; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload7176; AVX512DQ-NEXT:    vmovdqa64 %zmm16, %zmm10 {%k2}7177; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm10 {%k3}7178; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7179; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload7180; AVX512DQ-NEXT:    vmovdqa64 %zmm16, %zmm0 {%k1}7181; AVX512DQ-NEXT:    vmovdqu64 (%rsp), %zmm16 # 64-byte Reload7182; AVX512DQ-NEXT:    vmovdqa64 %zmm16, %zmm8 {%k2}7183; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm8 {%k3}7184; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7185; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload7186; AVX512DQ-NEXT:    vmovdqa64 %zmm16, %zmm0 {%k1}7187; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload7188; AVX512DQ-NEXT:    vmovdqa64 %zmm16, %zmm7 {%k2}7189; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm7 {%k3}7190; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7191; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload7192; AVX512DQ-NEXT:    vmovdqa64 %zmm16, %zmm0 {%k1}7193; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload7194; AVX512DQ-NEXT:    vmovdqa64 %zmm16, %zmm6 {%k2}7195; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm6 {%k3}7196; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7197; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload7198; AVX512DQ-NEXT:    vmovdqa64 %zmm16, %zmm0 {%k1}7199; AVX512DQ-NEXT:    vmovdqa64 %zmm29, %zmm4 {%k2}7200; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm4 {%k3}7201; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7202; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm28 {%k1}7203; AVX512DQ-NEXT:    vmovdqa64 %zmm30, %zmm2 {%k2}7204; AVX512DQ-NEXT:    vmovdqa64 %zmm28, %zmm2 {%k3}7205; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7206; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm27 {%k1}7207; AVX512DQ-NEXT:    vmovdqa64 %zmm5, %zmm1 {%k2}7208; AVX512DQ-NEXT:    vmovdqa64 %zmm27, %zmm1 {%k3}7209; AVX512DQ-NEXT:    movq {{[0-9]+}}(%rsp), %rax7210; AVX512DQ-NEXT:    vmovdqa64 %zmm1, 896(%rax)7211; AVX512DQ-NEXT:    vmovdqa64 %zmm2, 960(%rax)7212; AVX512DQ-NEXT:    vmovdqa64 %zmm4, 768(%rax)7213; AVX512DQ-NEXT:    vmovdqa64 %zmm6, 832(%rax)7214; AVX512DQ-NEXT:    vmovdqa64 %zmm7, 640(%rax)7215; AVX512DQ-NEXT:    vmovdqa64 %zmm8, 704(%rax)7216; AVX512DQ-NEXT:    vmovdqa64 %zmm10, 512(%rax)7217; AVX512DQ-NEXT:    vmovdqa64 %zmm12, 576(%rax)7218; AVX512DQ-NEXT:    vmovdqa64 %zmm3, 384(%rax)7219; AVX512DQ-NEXT:    vmovdqa64 %zmm9, 448(%rax)7220; AVX512DQ-NEXT:    vmovdqa64 %zmm11, 256(%rax)7221; AVX512DQ-NEXT:    vmovdqa64 %zmm13, 320(%rax)7222; AVX512DQ-NEXT:    vmovdqa64 %zmm14, 128(%rax)7223; AVX512DQ-NEXT:    vmovdqa64 %zmm15, 192(%rax)7224; AVX512DQ-NEXT:    vmovdqa64 %zmm17, (%rax)7225; AVX512DQ-NEXT:    vmovdqa64 %zmm18, 64(%rax)7226; AVX512DQ-NEXT:    addq $2056, %rsp # imm = 0x8087227; AVX512DQ-NEXT:    vzeroupper7228; AVX512DQ-NEXT:    retq7229;7230; AVX512DQ-FCP-LABEL: store_i32_stride8_vf32:7231; AVX512DQ-FCP:       # %bb.0:7232; AVX512DQ-FCP-NEXT:    subq $2056, %rsp # imm = 0x8087233; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax7234; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r107235; AVX512DQ-FCP-NEXT:    vmovdqa64 (%r8), %zmm17236; AVX512DQ-FCP-NEXT:    vmovdqa64 (%r9), %zmm277237; AVX512DQ-FCP-NEXT:    vmovdqa64 (%r10), %zmm27238; AVX512DQ-FCP-NEXT:    vmovdqa64 64(%r10), %zmm07239; AVX512DQ-FCP-NEXT:    vmovdqa64 (%rax), %zmm307240; AVX512DQ-FCP-NEXT:    vmovdqa64 64(%rax), %zmm287241; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm3 = [0,0,0,0,0,0,2,18,0,0,0,0,0,0,3,19]7242; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm2, %zmm57243; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm2, %zmm47244; AVX512DQ-FCP-NEXT:    vpermt2d %zmm30, %zmm3, %zmm57245; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7246; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [0,0,0,0,2,18,0,0,0,0,0,0,3,19,0,0]7247; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, %zmm57248; AVX512DQ-FCP-NEXT:    vpermt2d %zmm27, %zmm2, %zmm57249; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7250; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm5 = [0,0,0,0,0,0,0,16,0,0,0,0,0,0,1,17]7251; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm4, %zmm67252; AVX512DQ-FCP-NEXT:    vpermt2d %zmm30, %zmm5, %zmm67253; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7254; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm6 = [0,0,0,0,0,16,0,0,0,0,0,0,1,17,0,0]7255; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, %zmm77256; AVX512DQ-FCP-NEXT:    vpermt2d %zmm27, %zmm6, %zmm77257; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7258; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm7 = [0,0,0,0,0,0,6,22,0,0,0,0,0,0,7,23]7259; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm4, %zmm87260; AVX512DQ-FCP-NEXT:    vpermt2d %zmm30, %zmm7, %zmm87261; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7262; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm8 = [0,0,0,0,6,22,0,0,0,0,0,0,7,23,0,0]7263; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, %zmm97264; AVX512DQ-FCP-NEXT:    vpermt2d %zmm27, %zmm8, %zmm97265; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7266; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm9 = [0,0,0,0,0,0,4,20,0,0,0,0,0,0,5,21]7267; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm4, %zmm107268; AVX512DQ-FCP-NEXT:    vpermt2d %zmm30, %zmm9, %zmm107269; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7270; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm10 = [0,0,0,0,4,20,0,0,0,0,0,0,5,21,0,0]7271; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, %zmm117272; AVX512DQ-FCP-NEXT:    vpermt2d %zmm27, %zmm10, %zmm117273; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7274; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm11 = [0,0,0,0,0,0,10,26,0,0,0,0,0,0,11,27]7275; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm4, %zmm127276; AVX512DQ-FCP-NEXT:    vpermt2d %zmm30, %zmm11, %zmm127277; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7278; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,0,0,10,26,0,0,0,0,0,0,11,27,0,0]7279; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, %zmm137280; AVX512DQ-FCP-NEXT:    vpermt2d %zmm27, %zmm12, %zmm137281; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7282; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm13 = [0,0,0,0,0,0,8,24,0,0,0,0,0,0,9,25]7283; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm4, %zmm147284; AVX512DQ-FCP-NEXT:    vpermt2d %zmm30, %zmm13, %zmm147285; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7286; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm14 = [0,0,0,0,8,24,0,0,0,0,0,0,9,25,0,0]7287; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm15 = [0,0,0,0,0,0,14,30,0,0,0,0,0,0,15,31]7288; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm4, %zmm167289; AVX512DQ-FCP-NEXT:    vpermt2d %zmm30, %zmm15, %zmm167290; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7291; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm31 = [0,0,0,0,0,0,12,28,0,0,0,0,0,0,13,29]7292; AVX512DQ-FCP-NEXT:    vpermt2d %zmm30, %zmm31, %zmm47293; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7294; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, %zmm47295; AVX512DQ-FCP-NEXT:    vpermt2d %zmm27, %zmm14, %zmm47296; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7297; AVX512DQ-FCP-NEXT:    vpermi2d %zmm28, %zmm0, %zmm37298; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7299; AVX512DQ-FCP-NEXT:    vpermi2d %zmm28, %zmm0, %zmm57300; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7301; AVX512DQ-FCP-NEXT:    vpermi2d %zmm28, %zmm0, %zmm77302; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7303; AVX512DQ-FCP-NEXT:    vpermi2d %zmm28, %zmm0, %zmm97304; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7305; AVX512DQ-FCP-NEXT:    vpermi2d %zmm28, %zmm0, %zmm117306; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7307; AVX512DQ-FCP-NEXT:    vpermi2d %zmm28, %zmm0, %zmm137308; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7309; AVX512DQ-FCP-NEXT:    vpermi2d %zmm28, %zmm0, %zmm157310; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7311; AVX512DQ-FCP-NEXT:    vpermt2d %zmm28, %zmm31, %zmm07312; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7313; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm28 = [0,0,0,0,14,30,0,0,0,0,0,0,15,31,0,0]7314; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, %zmm317315; AVX512DQ-FCP-NEXT:    vpermt2d %zmm27, %zmm28, %zmm317316; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm0 = [0,0,0,0,12,28,0,0,0,0,0,0,13,29,0,0]7317; AVX512DQ-FCP-NEXT:    vpermt2d %zmm27, %zmm0, %zmm17318; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7319; AVX512DQ-FCP-NEXT:    vmovdqa64 64(%r8), %zmm277320; AVX512DQ-FCP-NEXT:    vmovdqa64 64(%r9), %zmm17321; AVX512DQ-FCP-NEXT:    vpermi2d %zmm1, %zmm27, %zmm27322; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7323; AVX512DQ-FCP-NEXT:    vpermi2d %zmm1, %zmm27, %zmm67324; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7325; AVX512DQ-FCP-NEXT:    vpermi2d %zmm1, %zmm27, %zmm87326; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7327; AVX512DQ-FCP-NEXT:    vpermi2d %zmm1, %zmm27, %zmm107328; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7329; AVX512DQ-FCP-NEXT:    vpermi2d %zmm1, %zmm27, %zmm127330; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7331; AVX512DQ-FCP-NEXT:    vpermi2d %zmm1, %zmm27, %zmm147332; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7333; AVX512DQ-FCP-NEXT:    vpermi2d %zmm1, %zmm27, %zmm287334; AVX512DQ-FCP-NEXT:    vpermt2d %zmm1, %zmm0, %zmm277335; AVX512DQ-FCP-NEXT:    vmovdqa64 (%rdx), %zmm167336; AVX512DQ-FCP-NEXT:    vmovdqa64 (%rcx), %zmm07337; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [0,0,2,18,0,0,0,0,0,0,3,19,0,0,0,0]7338; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm16, %zmm207339; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm2, %zmm207340; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm3 = [0,0,0,16,0,0,0,0,0,0,1,17,0,0,0,0]7341; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm16, %zmm217342; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm3, %zmm217343; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [0,0,6,22,0,0,0,0,0,0,7,23,0,0,0,0]7344; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm16, %zmm227345; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm4, %zmm227346; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm6 = [0,0,4,20,0,0,0,0,0,0,5,21,0,0,0,0]7347; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm16, %zmm237348; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm6, %zmm237349; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm7 = [0,0,10,26,0,0,0,0,0,0,11,27,0,0,0,0]7350; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm16, %zmm247351; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm7, %zmm247352; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm29 = [0,0,8,24,0,0,0,0,0,0,9,25,0,0,0,0]7353; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm16, %zmm257354; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm29, %zmm257355; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm30 = [0,0,14,30,0,0,0,0,0,0,15,31,0,0,0,0]7356; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm16, %zmm267357; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm30, %zmm267358; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,12,28,0,0,0,0,0,0,13,29,0,0,0,0]7359; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm1, %zmm167360; AVX512DQ-FCP-NEXT:    vmovdqa64 64(%rdx), %zmm57361; AVX512DQ-FCP-NEXT:    vmovdqa64 64(%rcx), %zmm07362; AVX512DQ-FCP-NEXT:    vpermi2d %zmm0, %zmm5, %zmm27363; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7364; AVX512DQ-FCP-NEXT:    vpermi2d %zmm0, %zmm5, %zmm37365; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7366; AVX512DQ-FCP-NEXT:    vpermi2d %zmm0, %zmm5, %zmm47367; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm4, (%rsp) # 64-byte Spill7368; AVX512DQ-FCP-NEXT:    vpermi2d %zmm0, %zmm5, %zmm67369; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7370; AVX512DQ-FCP-NEXT:    vpermi2d %zmm0, %zmm5, %zmm77371; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7372; AVX512DQ-FCP-NEXT:    vpermi2d %zmm0, %zmm5, %zmm297373; AVX512DQ-FCP-NEXT:    vpermi2d %zmm0, %zmm5, %zmm307374; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm1, %zmm57375; AVX512DQ-FCP-NEXT:    vmovdqa64 (%rdi), %zmm37376; AVX512DQ-FCP-NEXT:    vmovdqa64 (%rsi), %zmm07377; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [2,18,0,0,0,0,0,0,3,19,0,0,0,0,0,0]7378; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm3, %zmm187379; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm12, %zmm187380; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm10 = [0,16,0,0,0,0,0,0,1,17,0,0,0,0,0,0]7381; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm3, %zmm177382; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm10, %zmm177383; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm8 = [6,22,0,0,0,0,0,0,7,23,0,0,0,0,0,0]7384; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm3, %zmm157385; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm8, %zmm157386; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm7 = [4,20,0,0,0,0,0,0,5,21,0,0,0,0,0,0]7387; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm3, %zmm147388; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm7, %zmm147389; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm6 = [10,26,0,0,0,0,0,0,11,27,0,0,0,0,0,0]7390; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm3, %zmm137391; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm6, %zmm137392; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [8,24,0,0,0,0,0,0,9,25,0,0,0,0,0,0]7393; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm3, %zmm117394; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm4, %zmm117395; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [14,30,0,0,0,0,0,0,15,31,0,0,0,0,0,0]7396; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm3, %zmm97397; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm2, %zmm97398; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm19 = [12,28,0,0,0,0,0,0,13,29,0,0,0,0,0,0]7399; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm19, %zmm37400; AVX512DQ-FCP-NEXT:    vmovdqa64 64(%rdi), %zmm17401; AVX512DQ-FCP-NEXT:    vmovdqa64 64(%rsi), %zmm07402; AVX512DQ-FCP-NEXT:    vpermi2d %zmm0, %zmm1, %zmm127403; AVX512DQ-FCP-NEXT:    vpermi2d %zmm0, %zmm1, %zmm107404; AVX512DQ-FCP-NEXT:    vpermi2d %zmm0, %zmm1, %zmm87405; AVX512DQ-FCP-NEXT:    vpermi2d %zmm0, %zmm1, %zmm77406; AVX512DQ-FCP-NEXT:    vpermi2d %zmm0, %zmm1, %zmm67407; AVX512DQ-FCP-NEXT:    vpermi2d %zmm0, %zmm1, %zmm47408; AVX512DQ-FCP-NEXT:    vpermi2d %zmm0, %zmm1, %zmm27409; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm19, %zmm17410; AVX512DQ-FCP-NEXT:    movb $-120, %al7411; AVX512DQ-FCP-NEXT:    kmovw %eax, %k17412; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7413; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload7414; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm19, %zmm0 {%k1}7415; AVX512DQ-FCP-NEXT:    movb $34, %al7416; AVX512DQ-FCP-NEXT:    kmovw %eax, %k27417; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm20, %zmm18 {%k2}7418; AVX512DQ-FCP-NEXT:    movb $-52, %al7419; AVX512DQ-FCP-NEXT:    kmovw %eax, %k37420; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm18 {%k3}7421; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7422; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload7423; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm19 {%k1}7424; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm21, %zmm17 {%k2}7425; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm19, %zmm17 {%k3}7426; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7427; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload7428; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm19 {%k1}7429; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm22, %zmm15 {%k2}7430; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm19, %zmm15 {%k3}7431; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7432; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload7433; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm19 {%k1}7434; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm23, %zmm14 {%k2}7435; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm19, %zmm14 {%k3}7436; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7437; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload7438; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm19 {%k1}7439; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm24, %zmm13 {%k2}7440; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm19, %zmm13 {%k3}7441; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7442; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload7443; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm19 {%k1}7444; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm25, %zmm11 {%k2}7445; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm19, %zmm11 {%k3}7446; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7447; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm31 {%k1}7448; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm26, %zmm9 {%k2}7449; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm31, %zmm9 {%k3}7450; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7451; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload7452; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm19, %zmm0 {%k1}7453; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm16, %zmm3 {%k2}7454; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm3 {%k3}7455; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7456; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload7457; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm16, %zmm0 {%k1}7458; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload7459; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm16, %zmm12 {%k2}7460; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm12 {%k3}7461; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7462; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload7463; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm16, %zmm0 {%k1}7464; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload7465; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm16, %zmm10 {%k2}7466; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm10 {%k3}7467; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7468; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload7469; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm16, %zmm0 {%k1}7470; AVX512DQ-FCP-NEXT:    vmovdqu64 (%rsp), %zmm16 # 64-byte Reload7471; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm16, %zmm8 {%k2}7472; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm8 {%k3}7473; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7474; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload7475; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm16, %zmm0 {%k1}7476; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload7477; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm16, %zmm7 {%k2}7478; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm7 {%k3}7479; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7480; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload7481; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm16, %zmm0 {%k1}7482; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload7483; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm16, %zmm6 {%k2}7484; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm6 {%k3}7485; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7486; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload7487; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm16, %zmm0 {%k1}7488; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm29, %zmm4 {%k2}7489; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm4 {%k3}7490; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7491; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm28 {%k1}7492; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm30, %zmm2 {%k2}7493; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm28, %zmm2 {%k3}7494; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7495; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm27 {%k1}7496; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm5, %zmm1 {%k2}7497; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm27, %zmm1 {%k3}7498; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax7499; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, 896(%rax)7500; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm2, 960(%rax)7501; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm4, 768(%rax)7502; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm6, 832(%rax)7503; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm7, 640(%rax)7504; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm8, 704(%rax)7505; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm10, 512(%rax)7506; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm12, 576(%rax)7507; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm3, 384(%rax)7508; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm9, 448(%rax)7509; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm11, 256(%rax)7510; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm13, 320(%rax)7511; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm14, 128(%rax)7512; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm15, 192(%rax)7513; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm17, (%rax)7514; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm18, 64(%rax)7515; AVX512DQ-FCP-NEXT:    addq $2056, %rsp # imm = 0x8087516; AVX512DQ-FCP-NEXT:    vzeroupper7517; AVX512DQ-FCP-NEXT:    retq7518;7519; AVX512BW-LABEL: store_i32_stride8_vf32:7520; AVX512BW:       # %bb.0:7521; AVX512BW-NEXT:    subq $2056, %rsp # imm = 0x8087522; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax7523; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %r107524; AVX512BW-NEXT:    vmovdqa64 (%r8), %zmm17525; AVX512BW-NEXT:    vmovdqa64 (%r9), %zmm277526; AVX512BW-NEXT:    vmovdqa64 (%r10), %zmm27527; AVX512BW-NEXT:    vmovdqa64 64(%r10), %zmm07528; AVX512BW-NEXT:    vmovdqa64 (%rax), %zmm307529; AVX512BW-NEXT:    vmovdqa64 64(%rax), %zmm287530; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm3 = [0,0,0,0,0,0,2,18,0,0,0,0,0,0,3,19]7531; AVX512BW-NEXT:    vmovdqa64 %zmm2, %zmm57532; AVX512BW-NEXT:    vmovdqa64 %zmm2, %zmm47533; AVX512BW-NEXT:    vpermt2d %zmm30, %zmm3, %zmm57534; AVX512BW-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7535; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [0,0,0,0,2,18,0,0,0,0,0,0,3,19,0,0]7536; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm57537; AVX512BW-NEXT:    vpermt2d %zmm27, %zmm2, %zmm57538; AVX512BW-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7539; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm5 = [0,0,0,0,0,0,0,16,0,0,0,0,0,0,1,17]7540; AVX512BW-NEXT:    vmovdqa64 %zmm4, %zmm67541; AVX512BW-NEXT:    vpermt2d %zmm30, %zmm5, %zmm67542; AVX512BW-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7543; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm6 = [0,0,0,0,0,16,0,0,0,0,0,0,1,17,0,0]7544; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm77545; AVX512BW-NEXT:    vpermt2d %zmm27, %zmm6, %zmm77546; AVX512BW-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7547; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm7 = [0,0,0,0,0,0,6,22,0,0,0,0,0,0,7,23]7548; AVX512BW-NEXT:    vmovdqa64 %zmm4, %zmm87549; AVX512BW-NEXT:    vpermt2d %zmm30, %zmm7, %zmm87550; AVX512BW-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7551; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm8 = [0,0,0,0,6,22,0,0,0,0,0,0,7,23,0,0]7552; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm97553; AVX512BW-NEXT:    vpermt2d %zmm27, %zmm8, %zmm97554; AVX512BW-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7555; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm9 = [0,0,0,0,0,0,4,20,0,0,0,0,0,0,5,21]7556; AVX512BW-NEXT:    vmovdqa64 %zmm4, %zmm107557; AVX512BW-NEXT:    vpermt2d %zmm30, %zmm9, %zmm107558; AVX512BW-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7559; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm10 = [0,0,0,0,4,20,0,0,0,0,0,0,5,21,0,0]7560; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm117561; AVX512BW-NEXT:    vpermt2d %zmm27, %zmm10, %zmm117562; AVX512BW-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7563; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm11 = [0,0,0,0,0,0,10,26,0,0,0,0,0,0,11,27]7564; AVX512BW-NEXT:    vmovdqa64 %zmm4, %zmm127565; AVX512BW-NEXT:    vpermt2d %zmm30, %zmm11, %zmm127566; AVX512BW-NEXT:    vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7567; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,0,0,10,26,0,0,0,0,0,0,11,27,0,0]7568; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm137569; AVX512BW-NEXT:    vpermt2d %zmm27, %zmm12, %zmm137570; AVX512BW-NEXT:    vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7571; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm13 = [0,0,0,0,0,0,8,24,0,0,0,0,0,0,9,25]7572; AVX512BW-NEXT:    vmovdqa64 %zmm4, %zmm147573; AVX512BW-NEXT:    vpermt2d %zmm30, %zmm13, %zmm147574; AVX512BW-NEXT:    vmovdqu64 %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7575; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm14 = [0,0,0,0,8,24,0,0,0,0,0,0,9,25,0,0]7576; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm15 = [0,0,0,0,0,0,14,30,0,0,0,0,0,0,15,31]7577; AVX512BW-NEXT:    vmovdqa64 %zmm4, %zmm167578; AVX512BW-NEXT:    vpermt2d %zmm30, %zmm15, %zmm167579; AVX512BW-NEXT:    vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7580; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm31 = [0,0,0,0,0,0,12,28,0,0,0,0,0,0,13,29]7581; AVX512BW-NEXT:    vpermt2d %zmm30, %zmm31, %zmm47582; AVX512BW-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7583; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm47584; AVX512BW-NEXT:    vpermt2d %zmm27, %zmm14, %zmm47585; AVX512BW-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7586; AVX512BW-NEXT:    vpermi2d %zmm28, %zmm0, %zmm37587; AVX512BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7588; AVX512BW-NEXT:    vpermi2d %zmm28, %zmm0, %zmm57589; AVX512BW-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7590; AVX512BW-NEXT:    vpermi2d %zmm28, %zmm0, %zmm77591; AVX512BW-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7592; AVX512BW-NEXT:    vpermi2d %zmm28, %zmm0, %zmm97593; AVX512BW-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7594; AVX512BW-NEXT:    vpermi2d %zmm28, %zmm0, %zmm117595; AVX512BW-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7596; AVX512BW-NEXT:    vpermi2d %zmm28, %zmm0, %zmm137597; AVX512BW-NEXT:    vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7598; AVX512BW-NEXT:    vpermi2d %zmm28, %zmm0, %zmm157599; AVX512BW-NEXT:    vmovdqu64 %zmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7600; AVX512BW-NEXT:    vpermt2d %zmm28, %zmm31, %zmm07601; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7602; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm28 = [0,0,0,0,14,30,0,0,0,0,0,0,15,31,0,0]7603; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm317604; AVX512BW-NEXT:    vpermt2d %zmm27, %zmm28, %zmm317605; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm0 = [0,0,0,0,12,28,0,0,0,0,0,0,13,29,0,0]7606; AVX512BW-NEXT:    vpermt2d %zmm27, %zmm0, %zmm17607; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7608; AVX512BW-NEXT:    vmovdqa64 64(%r8), %zmm277609; AVX512BW-NEXT:    vmovdqa64 64(%r9), %zmm17610; AVX512BW-NEXT:    vpermi2d %zmm1, %zmm27, %zmm27611; AVX512BW-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7612; AVX512BW-NEXT:    vpermi2d %zmm1, %zmm27, %zmm67613; AVX512BW-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7614; AVX512BW-NEXT:    vpermi2d %zmm1, %zmm27, %zmm87615; AVX512BW-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7616; AVX512BW-NEXT:    vpermi2d %zmm1, %zmm27, %zmm107617; AVX512BW-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7618; AVX512BW-NEXT:    vpermi2d %zmm1, %zmm27, %zmm127619; AVX512BW-NEXT:    vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7620; AVX512BW-NEXT:    vpermi2d %zmm1, %zmm27, %zmm147621; AVX512BW-NEXT:    vmovdqu64 %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7622; AVX512BW-NEXT:    vpermi2d %zmm1, %zmm27, %zmm287623; AVX512BW-NEXT:    vpermt2d %zmm1, %zmm0, %zmm277624; AVX512BW-NEXT:    vmovdqa64 (%rdx), %zmm167625; AVX512BW-NEXT:    vmovdqa64 (%rcx), %zmm07626; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [0,0,2,18,0,0,0,0,0,0,3,19,0,0,0,0]7627; AVX512BW-NEXT:    vmovdqa64 %zmm16, %zmm207628; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm2, %zmm207629; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm3 = [0,0,0,16,0,0,0,0,0,0,1,17,0,0,0,0]7630; AVX512BW-NEXT:    vmovdqa64 %zmm16, %zmm217631; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm3, %zmm217632; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [0,0,6,22,0,0,0,0,0,0,7,23,0,0,0,0]7633; AVX512BW-NEXT:    vmovdqa64 %zmm16, %zmm227634; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm4, %zmm227635; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm6 = [0,0,4,20,0,0,0,0,0,0,5,21,0,0,0,0]7636; AVX512BW-NEXT:    vmovdqa64 %zmm16, %zmm237637; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm6, %zmm237638; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm7 = [0,0,10,26,0,0,0,0,0,0,11,27,0,0,0,0]7639; AVX512BW-NEXT:    vmovdqa64 %zmm16, %zmm247640; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm7, %zmm247641; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm29 = [0,0,8,24,0,0,0,0,0,0,9,25,0,0,0,0]7642; AVX512BW-NEXT:    vmovdqa64 %zmm16, %zmm257643; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm29, %zmm257644; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm30 = [0,0,14,30,0,0,0,0,0,0,15,31,0,0,0,0]7645; AVX512BW-NEXT:    vmovdqa64 %zmm16, %zmm267646; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm30, %zmm267647; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,12,28,0,0,0,0,0,0,13,29,0,0,0,0]7648; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm1, %zmm167649; AVX512BW-NEXT:    vmovdqa64 64(%rdx), %zmm57650; AVX512BW-NEXT:    vmovdqa64 64(%rcx), %zmm07651; AVX512BW-NEXT:    vpermi2d %zmm0, %zmm5, %zmm27652; AVX512BW-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7653; AVX512BW-NEXT:    vpermi2d %zmm0, %zmm5, %zmm37654; AVX512BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7655; AVX512BW-NEXT:    vpermi2d %zmm0, %zmm5, %zmm47656; AVX512BW-NEXT:    vmovdqu64 %zmm4, (%rsp) # 64-byte Spill7657; AVX512BW-NEXT:    vpermi2d %zmm0, %zmm5, %zmm67658; AVX512BW-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7659; AVX512BW-NEXT:    vpermi2d %zmm0, %zmm5, %zmm77660; AVX512BW-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7661; AVX512BW-NEXT:    vpermi2d %zmm0, %zmm5, %zmm297662; AVX512BW-NEXT:    vpermi2d %zmm0, %zmm5, %zmm307663; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm1, %zmm57664; AVX512BW-NEXT:    vmovdqa64 (%rdi), %zmm37665; AVX512BW-NEXT:    vmovdqa64 (%rsi), %zmm07666; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [2,18,0,0,0,0,0,0,3,19,0,0,0,0,0,0]7667; AVX512BW-NEXT:    vmovdqa64 %zmm3, %zmm187668; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm12, %zmm187669; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm10 = [0,16,0,0,0,0,0,0,1,17,0,0,0,0,0,0]7670; AVX512BW-NEXT:    vmovdqa64 %zmm3, %zmm177671; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm10, %zmm177672; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm8 = [6,22,0,0,0,0,0,0,7,23,0,0,0,0,0,0]7673; AVX512BW-NEXT:    vmovdqa64 %zmm3, %zmm157674; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm8, %zmm157675; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm7 = [4,20,0,0,0,0,0,0,5,21,0,0,0,0,0,0]7676; AVX512BW-NEXT:    vmovdqa64 %zmm3, %zmm147677; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm7, %zmm147678; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm6 = [10,26,0,0,0,0,0,0,11,27,0,0,0,0,0,0]7679; AVX512BW-NEXT:    vmovdqa64 %zmm3, %zmm137680; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm6, %zmm137681; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [8,24,0,0,0,0,0,0,9,25,0,0,0,0,0,0]7682; AVX512BW-NEXT:    vmovdqa64 %zmm3, %zmm117683; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm4, %zmm117684; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [14,30,0,0,0,0,0,0,15,31,0,0,0,0,0,0]7685; AVX512BW-NEXT:    vmovdqa64 %zmm3, %zmm97686; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm2, %zmm97687; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm19 = [12,28,0,0,0,0,0,0,13,29,0,0,0,0,0,0]7688; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm19, %zmm37689; AVX512BW-NEXT:    vmovdqa64 64(%rdi), %zmm17690; AVX512BW-NEXT:    vmovdqa64 64(%rsi), %zmm07691; AVX512BW-NEXT:    vpermi2d %zmm0, %zmm1, %zmm127692; AVX512BW-NEXT:    vpermi2d %zmm0, %zmm1, %zmm107693; AVX512BW-NEXT:    vpermi2d %zmm0, %zmm1, %zmm87694; AVX512BW-NEXT:    vpermi2d %zmm0, %zmm1, %zmm77695; AVX512BW-NEXT:    vpermi2d %zmm0, %zmm1, %zmm67696; AVX512BW-NEXT:    vpermi2d %zmm0, %zmm1, %zmm47697; AVX512BW-NEXT:    vpermi2d %zmm0, %zmm1, %zmm27698; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm19, %zmm17699; AVX512BW-NEXT:    movb $-120, %al7700; AVX512BW-NEXT:    kmovd %eax, %k17701; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7702; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload7703; AVX512BW-NEXT:    vmovdqa64 %zmm19, %zmm0 {%k1}7704; AVX512BW-NEXT:    movb $34, %al7705; AVX512BW-NEXT:    kmovd %eax, %k27706; AVX512BW-NEXT:    vmovdqa64 %zmm20, %zmm18 {%k2}7707; AVX512BW-NEXT:    movb $-52, %al7708; AVX512BW-NEXT:    kmovd %eax, %k37709; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm18 {%k3}7710; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7711; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload7712; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm19 {%k1}7713; AVX512BW-NEXT:    vmovdqa64 %zmm21, %zmm17 {%k2}7714; AVX512BW-NEXT:    vmovdqa64 %zmm19, %zmm17 {%k3}7715; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7716; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload7717; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm19 {%k1}7718; AVX512BW-NEXT:    vmovdqa64 %zmm22, %zmm15 {%k2}7719; AVX512BW-NEXT:    vmovdqa64 %zmm19, %zmm15 {%k3}7720; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7721; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload7722; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm19 {%k1}7723; AVX512BW-NEXT:    vmovdqa64 %zmm23, %zmm14 {%k2}7724; AVX512BW-NEXT:    vmovdqa64 %zmm19, %zmm14 {%k3}7725; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7726; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload7727; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm19 {%k1}7728; AVX512BW-NEXT:    vmovdqa64 %zmm24, %zmm13 {%k2}7729; AVX512BW-NEXT:    vmovdqa64 %zmm19, %zmm13 {%k3}7730; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7731; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload7732; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm19 {%k1}7733; AVX512BW-NEXT:    vmovdqa64 %zmm25, %zmm11 {%k2}7734; AVX512BW-NEXT:    vmovdqa64 %zmm19, %zmm11 {%k3}7735; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7736; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm31 {%k1}7737; AVX512BW-NEXT:    vmovdqa64 %zmm26, %zmm9 {%k2}7738; AVX512BW-NEXT:    vmovdqa64 %zmm31, %zmm9 {%k3}7739; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7740; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload7741; AVX512BW-NEXT:    vmovdqa64 %zmm19, %zmm0 {%k1}7742; AVX512BW-NEXT:    vmovdqa64 %zmm16, %zmm3 {%k2}7743; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm3 {%k3}7744; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7745; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload7746; AVX512BW-NEXT:    vmovdqa64 %zmm16, %zmm0 {%k1}7747; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload7748; AVX512BW-NEXT:    vmovdqa64 %zmm16, %zmm12 {%k2}7749; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm12 {%k3}7750; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7751; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload7752; AVX512BW-NEXT:    vmovdqa64 %zmm16, %zmm0 {%k1}7753; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload7754; AVX512BW-NEXT:    vmovdqa64 %zmm16, %zmm10 {%k2}7755; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm10 {%k3}7756; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7757; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload7758; AVX512BW-NEXT:    vmovdqa64 %zmm16, %zmm0 {%k1}7759; AVX512BW-NEXT:    vmovdqu64 (%rsp), %zmm16 # 64-byte Reload7760; AVX512BW-NEXT:    vmovdqa64 %zmm16, %zmm8 {%k2}7761; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm8 {%k3}7762; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7763; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload7764; AVX512BW-NEXT:    vmovdqa64 %zmm16, %zmm0 {%k1}7765; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload7766; AVX512BW-NEXT:    vmovdqa64 %zmm16, %zmm7 {%k2}7767; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm7 {%k3}7768; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7769; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload7770; AVX512BW-NEXT:    vmovdqa64 %zmm16, %zmm0 {%k1}7771; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload7772; AVX512BW-NEXT:    vmovdqa64 %zmm16, %zmm6 {%k2}7773; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm6 {%k3}7774; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7775; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload7776; AVX512BW-NEXT:    vmovdqa64 %zmm16, %zmm0 {%k1}7777; AVX512BW-NEXT:    vmovdqa64 %zmm29, %zmm4 {%k2}7778; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm4 {%k3}7779; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7780; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm28 {%k1}7781; AVX512BW-NEXT:    vmovdqa64 %zmm30, %zmm2 {%k2}7782; AVX512BW-NEXT:    vmovdqa64 %zmm28, %zmm2 {%k3}7783; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7784; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm27 {%k1}7785; AVX512BW-NEXT:    vmovdqa64 %zmm5, %zmm1 {%k2}7786; AVX512BW-NEXT:    vmovdqa64 %zmm27, %zmm1 {%k3}7787; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax7788; AVX512BW-NEXT:    vmovdqa64 %zmm1, 896(%rax)7789; AVX512BW-NEXT:    vmovdqa64 %zmm2, 960(%rax)7790; AVX512BW-NEXT:    vmovdqa64 %zmm4, 768(%rax)7791; AVX512BW-NEXT:    vmovdqa64 %zmm6, 832(%rax)7792; AVX512BW-NEXT:    vmovdqa64 %zmm7, 640(%rax)7793; AVX512BW-NEXT:    vmovdqa64 %zmm8, 704(%rax)7794; AVX512BW-NEXT:    vmovdqa64 %zmm10, 512(%rax)7795; AVX512BW-NEXT:    vmovdqa64 %zmm12, 576(%rax)7796; AVX512BW-NEXT:    vmovdqa64 %zmm3, 384(%rax)7797; AVX512BW-NEXT:    vmovdqa64 %zmm9, 448(%rax)7798; AVX512BW-NEXT:    vmovdqa64 %zmm11, 256(%rax)7799; AVX512BW-NEXT:    vmovdqa64 %zmm13, 320(%rax)7800; AVX512BW-NEXT:    vmovdqa64 %zmm14, 128(%rax)7801; AVX512BW-NEXT:    vmovdqa64 %zmm15, 192(%rax)7802; AVX512BW-NEXT:    vmovdqa64 %zmm17, (%rax)7803; AVX512BW-NEXT:    vmovdqa64 %zmm18, 64(%rax)7804; AVX512BW-NEXT:    addq $2056, %rsp # imm = 0x8087805; AVX512BW-NEXT:    vzeroupper7806; AVX512BW-NEXT:    retq7807;7808; AVX512BW-FCP-LABEL: store_i32_stride8_vf32:7809; AVX512BW-FCP:       # %bb.0:7810; AVX512BW-FCP-NEXT:    subq $2056, %rsp # imm = 0x8087811; AVX512BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax7812; AVX512BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r107813; AVX512BW-FCP-NEXT:    vmovdqa64 (%r8), %zmm17814; AVX512BW-FCP-NEXT:    vmovdqa64 (%r9), %zmm277815; AVX512BW-FCP-NEXT:    vmovdqa64 (%r10), %zmm27816; AVX512BW-FCP-NEXT:    vmovdqa64 64(%r10), %zmm07817; AVX512BW-FCP-NEXT:    vmovdqa64 (%rax), %zmm307818; AVX512BW-FCP-NEXT:    vmovdqa64 64(%rax), %zmm287819; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm3 = [0,0,0,0,0,0,2,18,0,0,0,0,0,0,3,19]7820; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm2, %zmm57821; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm2, %zmm47822; AVX512BW-FCP-NEXT:    vpermt2d %zmm30, %zmm3, %zmm57823; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7824; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [0,0,0,0,2,18,0,0,0,0,0,0,3,19,0,0]7825; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm1, %zmm57826; AVX512BW-FCP-NEXT:    vpermt2d %zmm27, %zmm2, %zmm57827; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7828; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm5 = [0,0,0,0,0,0,0,16,0,0,0,0,0,0,1,17]7829; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm4, %zmm67830; AVX512BW-FCP-NEXT:    vpermt2d %zmm30, %zmm5, %zmm67831; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7832; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm6 = [0,0,0,0,0,16,0,0,0,0,0,0,1,17,0,0]7833; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm1, %zmm77834; AVX512BW-FCP-NEXT:    vpermt2d %zmm27, %zmm6, %zmm77835; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7836; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm7 = [0,0,0,0,0,0,6,22,0,0,0,0,0,0,7,23]7837; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm4, %zmm87838; AVX512BW-FCP-NEXT:    vpermt2d %zmm30, %zmm7, %zmm87839; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7840; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm8 = [0,0,0,0,6,22,0,0,0,0,0,0,7,23,0,0]7841; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm1, %zmm97842; AVX512BW-FCP-NEXT:    vpermt2d %zmm27, %zmm8, %zmm97843; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7844; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm9 = [0,0,0,0,0,0,4,20,0,0,0,0,0,0,5,21]7845; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm4, %zmm107846; AVX512BW-FCP-NEXT:    vpermt2d %zmm30, %zmm9, %zmm107847; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7848; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm10 = [0,0,0,0,4,20,0,0,0,0,0,0,5,21,0,0]7849; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm1, %zmm117850; AVX512BW-FCP-NEXT:    vpermt2d %zmm27, %zmm10, %zmm117851; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7852; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm11 = [0,0,0,0,0,0,10,26,0,0,0,0,0,0,11,27]7853; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm4, %zmm127854; AVX512BW-FCP-NEXT:    vpermt2d %zmm30, %zmm11, %zmm127855; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7856; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,0,0,10,26,0,0,0,0,0,0,11,27,0,0]7857; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm1, %zmm137858; AVX512BW-FCP-NEXT:    vpermt2d %zmm27, %zmm12, %zmm137859; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7860; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm13 = [0,0,0,0,0,0,8,24,0,0,0,0,0,0,9,25]7861; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm4, %zmm147862; AVX512BW-FCP-NEXT:    vpermt2d %zmm30, %zmm13, %zmm147863; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7864; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm14 = [0,0,0,0,8,24,0,0,0,0,0,0,9,25,0,0]7865; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm15 = [0,0,0,0,0,0,14,30,0,0,0,0,0,0,15,31]7866; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm4, %zmm167867; AVX512BW-FCP-NEXT:    vpermt2d %zmm30, %zmm15, %zmm167868; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7869; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm31 = [0,0,0,0,0,0,12,28,0,0,0,0,0,0,13,29]7870; AVX512BW-FCP-NEXT:    vpermt2d %zmm30, %zmm31, %zmm47871; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7872; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm1, %zmm47873; AVX512BW-FCP-NEXT:    vpermt2d %zmm27, %zmm14, %zmm47874; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7875; AVX512BW-FCP-NEXT:    vpermi2d %zmm28, %zmm0, %zmm37876; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7877; AVX512BW-FCP-NEXT:    vpermi2d %zmm28, %zmm0, %zmm57878; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7879; AVX512BW-FCP-NEXT:    vpermi2d %zmm28, %zmm0, %zmm77880; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7881; AVX512BW-FCP-NEXT:    vpermi2d %zmm28, %zmm0, %zmm97882; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7883; AVX512BW-FCP-NEXT:    vpermi2d %zmm28, %zmm0, %zmm117884; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7885; AVX512BW-FCP-NEXT:    vpermi2d %zmm28, %zmm0, %zmm137886; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7887; AVX512BW-FCP-NEXT:    vpermi2d %zmm28, %zmm0, %zmm157888; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7889; AVX512BW-FCP-NEXT:    vpermt2d %zmm28, %zmm31, %zmm07890; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7891; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm28 = [0,0,0,0,14,30,0,0,0,0,0,0,15,31,0,0]7892; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm1, %zmm317893; AVX512BW-FCP-NEXT:    vpermt2d %zmm27, %zmm28, %zmm317894; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm0 = [0,0,0,0,12,28,0,0,0,0,0,0,13,29,0,0]7895; AVX512BW-FCP-NEXT:    vpermt2d %zmm27, %zmm0, %zmm17896; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7897; AVX512BW-FCP-NEXT:    vmovdqa64 64(%r8), %zmm277898; AVX512BW-FCP-NEXT:    vmovdqa64 64(%r9), %zmm17899; AVX512BW-FCP-NEXT:    vpermi2d %zmm1, %zmm27, %zmm27900; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7901; AVX512BW-FCP-NEXT:    vpermi2d %zmm1, %zmm27, %zmm67902; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7903; AVX512BW-FCP-NEXT:    vpermi2d %zmm1, %zmm27, %zmm87904; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7905; AVX512BW-FCP-NEXT:    vpermi2d %zmm1, %zmm27, %zmm107906; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7907; AVX512BW-FCP-NEXT:    vpermi2d %zmm1, %zmm27, %zmm127908; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7909; AVX512BW-FCP-NEXT:    vpermi2d %zmm1, %zmm27, %zmm147910; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7911; AVX512BW-FCP-NEXT:    vpermi2d %zmm1, %zmm27, %zmm287912; AVX512BW-FCP-NEXT:    vpermt2d %zmm1, %zmm0, %zmm277913; AVX512BW-FCP-NEXT:    vmovdqa64 (%rdx), %zmm167914; AVX512BW-FCP-NEXT:    vmovdqa64 (%rcx), %zmm07915; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [0,0,2,18,0,0,0,0,0,0,3,19,0,0,0,0]7916; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm16, %zmm207917; AVX512BW-FCP-NEXT:    vpermt2d %zmm0, %zmm2, %zmm207918; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm3 = [0,0,0,16,0,0,0,0,0,0,1,17,0,0,0,0]7919; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm16, %zmm217920; AVX512BW-FCP-NEXT:    vpermt2d %zmm0, %zmm3, %zmm217921; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [0,0,6,22,0,0,0,0,0,0,7,23,0,0,0,0]7922; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm16, %zmm227923; AVX512BW-FCP-NEXT:    vpermt2d %zmm0, %zmm4, %zmm227924; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm6 = [0,0,4,20,0,0,0,0,0,0,5,21,0,0,0,0]7925; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm16, %zmm237926; AVX512BW-FCP-NEXT:    vpermt2d %zmm0, %zmm6, %zmm237927; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm7 = [0,0,10,26,0,0,0,0,0,0,11,27,0,0,0,0]7928; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm16, %zmm247929; AVX512BW-FCP-NEXT:    vpermt2d %zmm0, %zmm7, %zmm247930; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm29 = [0,0,8,24,0,0,0,0,0,0,9,25,0,0,0,0]7931; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm16, %zmm257932; AVX512BW-FCP-NEXT:    vpermt2d %zmm0, %zmm29, %zmm257933; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm30 = [0,0,14,30,0,0,0,0,0,0,15,31,0,0,0,0]7934; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm16, %zmm267935; AVX512BW-FCP-NEXT:    vpermt2d %zmm0, %zmm30, %zmm267936; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,12,28,0,0,0,0,0,0,13,29,0,0,0,0]7937; AVX512BW-FCP-NEXT:    vpermt2d %zmm0, %zmm1, %zmm167938; AVX512BW-FCP-NEXT:    vmovdqa64 64(%rdx), %zmm57939; AVX512BW-FCP-NEXT:    vmovdqa64 64(%rcx), %zmm07940; AVX512BW-FCP-NEXT:    vpermi2d %zmm0, %zmm5, %zmm27941; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7942; AVX512BW-FCP-NEXT:    vpermi2d %zmm0, %zmm5, %zmm37943; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7944; AVX512BW-FCP-NEXT:    vpermi2d %zmm0, %zmm5, %zmm47945; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm4, (%rsp) # 64-byte Spill7946; AVX512BW-FCP-NEXT:    vpermi2d %zmm0, %zmm5, %zmm67947; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7948; AVX512BW-FCP-NEXT:    vpermi2d %zmm0, %zmm5, %zmm77949; AVX512BW-FCP-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7950; AVX512BW-FCP-NEXT:    vpermi2d %zmm0, %zmm5, %zmm297951; AVX512BW-FCP-NEXT:    vpermi2d %zmm0, %zmm5, %zmm307952; AVX512BW-FCP-NEXT:    vpermt2d %zmm0, %zmm1, %zmm57953; AVX512BW-FCP-NEXT:    vmovdqa64 (%rdi), %zmm37954; AVX512BW-FCP-NEXT:    vmovdqa64 (%rsi), %zmm07955; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [2,18,0,0,0,0,0,0,3,19,0,0,0,0,0,0]7956; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm3, %zmm187957; AVX512BW-FCP-NEXT:    vpermt2d %zmm0, %zmm12, %zmm187958; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm10 = [0,16,0,0,0,0,0,0,1,17,0,0,0,0,0,0]7959; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm3, %zmm177960; AVX512BW-FCP-NEXT:    vpermt2d %zmm0, %zmm10, %zmm177961; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm8 = [6,22,0,0,0,0,0,0,7,23,0,0,0,0,0,0]7962; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm3, %zmm157963; AVX512BW-FCP-NEXT:    vpermt2d %zmm0, %zmm8, %zmm157964; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm7 = [4,20,0,0,0,0,0,0,5,21,0,0,0,0,0,0]7965; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm3, %zmm147966; AVX512BW-FCP-NEXT:    vpermt2d %zmm0, %zmm7, %zmm147967; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm6 = [10,26,0,0,0,0,0,0,11,27,0,0,0,0,0,0]7968; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm3, %zmm137969; AVX512BW-FCP-NEXT:    vpermt2d %zmm0, %zmm6, %zmm137970; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [8,24,0,0,0,0,0,0,9,25,0,0,0,0,0,0]7971; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm3, %zmm117972; AVX512BW-FCP-NEXT:    vpermt2d %zmm0, %zmm4, %zmm117973; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [14,30,0,0,0,0,0,0,15,31,0,0,0,0,0,0]7974; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm3, %zmm97975; AVX512BW-FCP-NEXT:    vpermt2d %zmm0, %zmm2, %zmm97976; AVX512BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm19 = [12,28,0,0,0,0,0,0,13,29,0,0,0,0,0,0]7977; AVX512BW-FCP-NEXT:    vpermt2d %zmm0, %zmm19, %zmm37978; AVX512BW-FCP-NEXT:    vmovdqa64 64(%rdi), %zmm17979; AVX512BW-FCP-NEXT:    vmovdqa64 64(%rsi), %zmm07980; AVX512BW-FCP-NEXT:    vpermi2d %zmm0, %zmm1, %zmm127981; AVX512BW-FCP-NEXT:    vpermi2d %zmm0, %zmm1, %zmm107982; AVX512BW-FCP-NEXT:    vpermi2d %zmm0, %zmm1, %zmm87983; AVX512BW-FCP-NEXT:    vpermi2d %zmm0, %zmm1, %zmm77984; AVX512BW-FCP-NEXT:    vpermi2d %zmm0, %zmm1, %zmm67985; AVX512BW-FCP-NEXT:    vpermi2d %zmm0, %zmm1, %zmm47986; AVX512BW-FCP-NEXT:    vpermi2d %zmm0, %zmm1, %zmm27987; AVX512BW-FCP-NEXT:    vpermt2d %zmm0, %zmm19, %zmm17988; AVX512BW-FCP-NEXT:    movb $-120, %al7989; AVX512BW-FCP-NEXT:    kmovd %eax, %k17990; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload7991; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload7992; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm19, %zmm0 {%k1}7993; AVX512BW-FCP-NEXT:    movb $34, %al7994; AVX512BW-FCP-NEXT:    kmovd %eax, %k27995; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm20, %zmm18 {%k2}7996; AVX512BW-FCP-NEXT:    movb $-52, %al7997; AVX512BW-FCP-NEXT:    kmovd %eax, %k37998; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm18 {%k3}7999; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8000; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload8001; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm19 {%k1}8002; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm21, %zmm17 {%k2}8003; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm19, %zmm17 {%k3}8004; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8005; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload8006; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm19 {%k1}8007; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm22, %zmm15 {%k2}8008; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm19, %zmm15 {%k3}8009; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8010; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload8011; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm19 {%k1}8012; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm23, %zmm14 {%k2}8013; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm19, %zmm14 {%k3}8014; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8015; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload8016; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm19 {%k1}8017; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm24, %zmm13 {%k2}8018; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm19, %zmm13 {%k3}8019; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8020; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload8021; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm19 {%k1}8022; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm25, %zmm11 {%k2}8023; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm19, %zmm11 {%k3}8024; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8025; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm31 {%k1}8026; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm26, %zmm9 {%k2}8027; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm31, %zmm9 {%k3}8028; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8029; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload8030; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm19, %zmm0 {%k1}8031; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm16, %zmm3 {%k2}8032; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm3 {%k3}8033; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8034; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload8035; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm16, %zmm0 {%k1}8036; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload8037; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm16, %zmm12 {%k2}8038; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm12 {%k3}8039; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8040; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload8041; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm16, %zmm0 {%k1}8042; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload8043; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm16, %zmm10 {%k2}8044; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm10 {%k3}8045; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8046; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload8047; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm16, %zmm0 {%k1}8048; AVX512BW-FCP-NEXT:    vmovdqu64 (%rsp), %zmm16 # 64-byte Reload8049; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm16, %zmm8 {%k2}8050; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm8 {%k3}8051; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8052; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload8053; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm16, %zmm0 {%k1}8054; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload8055; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm16, %zmm7 {%k2}8056; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm7 {%k3}8057; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8058; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload8059; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm16, %zmm0 {%k1}8060; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload8061; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm16, %zmm6 {%k2}8062; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm6 {%k3}8063; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8064; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload8065; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm16, %zmm0 {%k1}8066; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm29, %zmm4 {%k2}8067; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm4 {%k3}8068; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8069; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm28 {%k1}8070; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm30, %zmm2 {%k2}8071; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm28, %zmm2 {%k3}8072; AVX512BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8073; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm27 {%k1}8074; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm5, %zmm1 {%k2}8075; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm27, %zmm1 {%k3}8076; AVX512BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax8077; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm1, 896(%rax)8078; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm2, 960(%rax)8079; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm4, 768(%rax)8080; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm6, 832(%rax)8081; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm7, 640(%rax)8082; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm8, 704(%rax)8083; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm10, 512(%rax)8084; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm12, 576(%rax)8085; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm3, 384(%rax)8086; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm9, 448(%rax)8087; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm11, 256(%rax)8088; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm13, 320(%rax)8089; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm14, 128(%rax)8090; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm15, 192(%rax)8091; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm17, (%rax)8092; AVX512BW-FCP-NEXT:    vmovdqa64 %zmm18, 64(%rax)8093; AVX512BW-FCP-NEXT:    addq $2056, %rsp # imm = 0x8088094; AVX512BW-FCP-NEXT:    vzeroupper8095; AVX512BW-FCP-NEXT:    retq8096;8097; AVX512DQ-BW-LABEL: store_i32_stride8_vf32:8098; AVX512DQ-BW:       # %bb.0:8099; AVX512DQ-BW-NEXT:    subq $2056, %rsp # imm = 0x8088100; AVX512DQ-BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax8101; AVX512DQ-BW-NEXT:    movq {{[0-9]+}}(%rsp), %r108102; AVX512DQ-BW-NEXT:    vmovdqa64 (%r8), %zmm18103; AVX512DQ-BW-NEXT:    vmovdqa64 (%r9), %zmm278104; AVX512DQ-BW-NEXT:    vmovdqa64 (%r10), %zmm28105; AVX512DQ-BW-NEXT:    vmovdqa64 64(%r10), %zmm08106; AVX512DQ-BW-NEXT:    vmovdqa64 (%rax), %zmm308107; AVX512DQ-BW-NEXT:    vmovdqa64 64(%rax), %zmm288108; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm3 = [0,0,0,0,0,0,2,18,0,0,0,0,0,0,3,19]8109; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm2, %zmm58110; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm2, %zmm48111; AVX512DQ-BW-NEXT:    vpermt2d %zmm30, %zmm3, %zmm58112; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8113; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [0,0,0,0,2,18,0,0,0,0,0,0,3,19,0,0]8114; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm1, %zmm58115; AVX512DQ-BW-NEXT:    vpermt2d %zmm27, %zmm2, %zmm58116; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8117; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm5 = [0,0,0,0,0,0,0,16,0,0,0,0,0,0,1,17]8118; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm4, %zmm68119; AVX512DQ-BW-NEXT:    vpermt2d %zmm30, %zmm5, %zmm68120; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8121; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm6 = [0,0,0,0,0,16,0,0,0,0,0,0,1,17,0,0]8122; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm1, %zmm78123; AVX512DQ-BW-NEXT:    vpermt2d %zmm27, %zmm6, %zmm78124; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8125; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm7 = [0,0,0,0,0,0,6,22,0,0,0,0,0,0,7,23]8126; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm4, %zmm88127; AVX512DQ-BW-NEXT:    vpermt2d %zmm30, %zmm7, %zmm88128; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8129; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm8 = [0,0,0,0,6,22,0,0,0,0,0,0,7,23,0,0]8130; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm1, %zmm98131; AVX512DQ-BW-NEXT:    vpermt2d %zmm27, %zmm8, %zmm98132; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8133; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm9 = [0,0,0,0,0,0,4,20,0,0,0,0,0,0,5,21]8134; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm4, %zmm108135; AVX512DQ-BW-NEXT:    vpermt2d %zmm30, %zmm9, %zmm108136; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8137; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm10 = [0,0,0,0,4,20,0,0,0,0,0,0,5,21,0,0]8138; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm1, %zmm118139; AVX512DQ-BW-NEXT:    vpermt2d %zmm27, %zmm10, %zmm118140; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8141; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm11 = [0,0,0,0,0,0,10,26,0,0,0,0,0,0,11,27]8142; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm4, %zmm128143; AVX512DQ-BW-NEXT:    vpermt2d %zmm30, %zmm11, %zmm128144; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8145; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,0,0,10,26,0,0,0,0,0,0,11,27,0,0]8146; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm1, %zmm138147; AVX512DQ-BW-NEXT:    vpermt2d %zmm27, %zmm12, %zmm138148; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8149; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm13 = [0,0,0,0,0,0,8,24,0,0,0,0,0,0,9,25]8150; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm4, %zmm148151; AVX512DQ-BW-NEXT:    vpermt2d %zmm30, %zmm13, %zmm148152; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8153; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm14 = [0,0,0,0,8,24,0,0,0,0,0,0,9,25,0,0]8154; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm15 = [0,0,0,0,0,0,14,30,0,0,0,0,0,0,15,31]8155; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm4, %zmm168156; AVX512DQ-BW-NEXT:    vpermt2d %zmm30, %zmm15, %zmm168157; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8158; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm31 = [0,0,0,0,0,0,12,28,0,0,0,0,0,0,13,29]8159; AVX512DQ-BW-NEXT:    vpermt2d %zmm30, %zmm31, %zmm48160; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8161; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm1, %zmm48162; AVX512DQ-BW-NEXT:    vpermt2d %zmm27, %zmm14, %zmm48163; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8164; AVX512DQ-BW-NEXT:    vpermi2d %zmm28, %zmm0, %zmm38165; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8166; AVX512DQ-BW-NEXT:    vpermi2d %zmm28, %zmm0, %zmm58167; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8168; AVX512DQ-BW-NEXT:    vpermi2d %zmm28, %zmm0, %zmm78169; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8170; AVX512DQ-BW-NEXT:    vpermi2d %zmm28, %zmm0, %zmm98171; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8172; AVX512DQ-BW-NEXT:    vpermi2d %zmm28, %zmm0, %zmm118173; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8174; AVX512DQ-BW-NEXT:    vpermi2d %zmm28, %zmm0, %zmm138175; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8176; AVX512DQ-BW-NEXT:    vpermi2d %zmm28, %zmm0, %zmm158177; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8178; AVX512DQ-BW-NEXT:    vpermt2d %zmm28, %zmm31, %zmm08179; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8180; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm28 = [0,0,0,0,14,30,0,0,0,0,0,0,15,31,0,0]8181; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm1, %zmm318182; AVX512DQ-BW-NEXT:    vpermt2d %zmm27, %zmm28, %zmm318183; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm0 = [0,0,0,0,12,28,0,0,0,0,0,0,13,29,0,0]8184; AVX512DQ-BW-NEXT:    vpermt2d %zmm27, %zmm0, %zmm18185; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8186; AVX512DQ-BW-NEXT:    vmovdqa64 64(%r8), %zmm278187; AVX512DQ-BW-NEXT:    vmovdqa64 64(%r9), %zmm18188; AVX512DQ-BW-NEXT:    vpermi2d %zmm1, %zmm27, %zmm28189; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8190; AVX512DQ-BW-NEXT:    vpermi2d %zmm1, %zmm27, %zmm68191; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8192; AVX512DQ-BW-NEXT:    vpermi2d %zmm1, %zmm27, %zmm88193; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8194; AVX512DQ-BW-NEXT:    vpermi2d %zmm1, %zmm27, %zmm108195; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8196; AVX512DQ-BW-NEXT:    vpermi2d %zmm1, %zmm27, %zmm128197; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8198; AVX512DQ-BW-NEXT:    vpermi2d %zmm1, %zmm27, %zmm148199; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8200; AVX512DQ-BW-NEXT:    vpermi2d %zmm1, %zmm27, %zmm288201; AVX512DQ-BW-NEXT:    vpermt2d %zmm1, %zmm0, %zmm278202; AVX512DQ-BW-NEXT:    vmovdqa64 (%rdx), %zmm168203; AVX512DQ-BW-NEXT:    vmovdqa64 (%rcx), %zmm08204; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [0,0,2,18,0,0,0,0,0,0,3,19,0,0,0,0]8205; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm16, %zmm208206; AVX512DQ-BW-NEXT:    vpermt2d %zmm0, %zmm2, %zmm208207; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm3 = [0,0,0,16,0,0,0,0,0,0,1,17,0,0,0,0]8208; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm16, %zmm218209; AVX512DQ-BW-NEXT:    vpermt2d %zmm0, %zmm3, %zmm218210; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [0,0,6,22,0,0,0,0,0,0,7,23,0,0,0,0]8211; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm16, %zmm228212; AVX512DQ-BW-NEXT:    vpermt2d %zmm0, %zmm4, %zmm228213; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm6 = [0,0,4,20,0,0,0,0,0,0,5,21,0,0,0,0]8214; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm16, %zmm238215; AVX512DQ-BW-NEXT:    vpermt2d %zmm0, %zmm6, %zmm238216; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm7 = [0,0,10,26,0,0,0,0,0,0,11,27,0,0,0,0]8217; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm16, %zmm248218; AVX512DQ-BW-NEXT:    vpermt2d %zmm0, %zmm7, %zmm248219; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm29 = [0,0,8,24,0,0,0,0,0,0,9,25,0,0,0,0]8220; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm16, %zmm258221; AVX512DQ-BW-NEXT:    vpermt2d %zmm0, %zmm29, %zmm258222; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm30 = [0,0,14,30,0,0,0,0,0,0,15,31,0,0,0,0]8223; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm16, %zmm268224; AVX512DQ-BW-NEXT:    vpermt2d %zmm0, %zmm30, %zmm268225; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,12,28,0,0,0,0,0,0,13,29,0,0,0,0]8226; AVX512DQ-BW-NEXT:    vpermt2d %zmm0, %zmm1, %zmm168227; AVX512DQ-BW-NEXT:    vmovdqa64 64(%rdx), %zmm58228; AVX512DQ-BW-NEXT:    vmovdqa64 64(%rcx), %zmm08229; AVX512DQ-BW-NEXT:    vpermi2d %zmm0, %zmm5, %zmm28230; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8231; AVX512DQ-BW-NEXT:    vpermi2d %zmm0, %zmm5, %zmm38232; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8233; AVX512DQ-BW-NEXT:    vpermi2d %zmm0, %zmm5, %zmm48234; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm4, (%rsp) # 64-byte Spill8235; AVX512DQ-BW-NEXT:    vpermi2d %zmm0, %zmm5, %zmm68236; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8237; AVX512DQ-BW-NEXT:    vpermi2d %zmm0, %zmm5, %zmm78238; AVX512DQ-BW-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8239; AVX512DQ-BW-NEXT:    vpermi2d %zmm0, %zmm5, %zmm298240; AVX512DQ-BW-NEXT:    vpermi2d %zmm0, %zmm5, %zmm308241; AVX512DQ-BW-NEXT:    vpermt2d %zmm0, %zmm1, %zmm58242; AVX512DQ-BW-NEXT:    vmovdqa64 (%rdi), %zmm38243; AVX512DQ-BW-NEXT:    vmovdqa64 (%rsi), %zmm08244; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [2,18,0,0,0,0,0,0,3,19,0,0,0,0,0,0]8245; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm3, %zmm188246; AVX512DQ-BW-NEXT:    vpermt2d %zmm0, %zmm12, %zmm188247; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm10 = [0,16,0,0,0,0,0,0,1,17,0,0,0,0,0,0]8248; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm3, %zmm178249; AVX512DQ-BW-NEXT:    vpermt2d %zmm0, %zmm10, %zmm178250; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm8 = [6,22,0,0,0,0,0,0,7,23,0,0,0,0,0,0]8251; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm3, %zmm158252; AVX512DQ-BW-NEXT:    vpermt2d %zmm0, %zmm8, %zmm158253; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm7 = [4,20,0,0,0,0,0,0,5,21,0,0,0,0,0,0]8254; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm3, %zmm148255; AVX512DQ-BW-NEXT:    vpermt2d %zmm0, %zmm7, %zmm148256; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm6 = [10,26,0,0,0,0,0,0,11,27,0,0,0,0,0,0]8257; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm3, %zmm138258; AVX512DQ-BW-NEXT:    vpermt2d %zmm0, %zmm6, %zmm138259; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [8,24,0,0,0,0,0,0,9,25,0,0,0,0,0,0]8260; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm3, %zmm118261; AVX512DQ-BW-NEXT:    vpermt2d %zmm0, %zmm4, %zmm118262; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [14,30,0,0,0,0,0,0,15,31,0,0,0,0,0,0]8263; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm3, %zmm98264; AVX512DQ-BW-NEXT:    vpermt2d %zmm0, %zmm2, %zmm98265; AVX512DQ-BW-NEXT:    vpmovsxbd {{.*#+}} zmm19 = [12,28,0,0,0,0,0,0,13,29,0,0,0,0,0,0]8266; AVX512DQ-BW-NEXT:    vpermt2d %zmm0, %zmm19, %zmm38267; AVX512DQ-BW-NEXT:    vmovdqa64 64(%rdi), %zmm18268; AVX512DQ-BW-NEXT:    vmovdqa64 64(%rsi), %zmm08269; AVX512DQ-BW-NEXT:    vpermi2d %zmm0, %zmm1, %zmm128270; AVX512DQ-BW-NEXT:    vpermi2d %zmm0, %zmm1, %zmm108271; AVX512DQ-BW-NEXT:    vpermi2d %zmm0, %zmm1, %zmm88272; AVX512DQ-BW-NEXT:    vpermi2d %zmm0, %zmm1, %zmm78273; AVX512DQ-BW-NEXT:    vpermi2d %zmm0, %zmm1, %zmm68274; AVX512DQ-BW-NEXT:    vpermi2d %zmm0, %zmm1, %zmm48275; AVX512DQ-BW-NEXT:    vpermi2d %zmm0, %zmm1, %zmm28276; AVX512DQ-BW-NEXT:    vpermt2d %zmm0, %zmm19, %zmm18277; AVX512DQ-BW-NEXT:    movb $-120, %al8278; AVX512DQ-BW-NEXT:    kmovd %eax, %k18279; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8280; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload8281; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm19, %zmm0 {%k1}8282; AVX512DQ-BW-NEXT:    movb $34, %al8283; AVX512DQ-BW-NEXT:    kmovd %eax, %k28284; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm20, %zmm18 {%k2}8285; AVX512DQ-BW-NEXT:    movb $-52, %al8286; AVX512DQ-BW-NEXT:    kmovd %eax, %k38287; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm0, %zmm18 {%k3}8288; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8289; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload8290; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm0, %zmm19 {%k1}8291; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm21, %zmm17 {%k2}8292; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm19, %zmm17 {%k3}8293; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8294; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload8295; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm0, %zmm19 {%k1}8296; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm22, %zmm15 {%k2}8297; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm19, %zmm15 {%k3}8298; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8299; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload8300; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm0, %zmm19 {%k1}8301; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm23, %zmm14 {%k2}8302; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm19, %zmm14 {%k3}8303; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8304; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload8305; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm0, %zmm19 {%k1}8306; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm24, %zmm13 {%k2}8307; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm19, %zmm13 {%k3}8308; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8309; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload8310; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm0, %zmm19 {%k1}8311; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm25, %zmm11 {%k2}8312; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm19, %zmm11 {%k3}8313; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8314; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm0, %zmm31 {%k1}8315; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm26, %zmm9 {%k2}8316; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm31, %zmm9 {%k3}8317; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8318; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload8319; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm19, %zmm0 {%k1}8320; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm16, %zmm3 {%k2}8321; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm0, %zmm3 {%k3}8322; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8323; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload8324; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm16, %zmm0 {%k1}8325; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload8326; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm16, %zmm12 {%k2}8327; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm0, %zmm12 {%k3}8328; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8329; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload8330; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm16, %zmm0 {%k1}8331; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload8332; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm16, %zmm10 {%k2}8333; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm0, %zmm10 {%k3}8334; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8335; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload8336; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm16, %zmm0 {%k1}8337; AVX512DQ-BW-NEXT:    vmovdqu64 (%rsp), %zmm16 # 64-byte Reload8338; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm16, %zmm8 {%k2}8339; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm0, %zmm8 {%k3}8340; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8341; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload8342; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm16, %zmm0 {%k1}8343; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload8344; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm16, %zmm7 {%k2}8345; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm0, %zmm7 {%k3}8346; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8347; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload8348; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm16, %zmm0 {%k1}8349; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload8350; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm16, %zmm6 {%k2}8351; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm0, %zmm6 {%k3}8352; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8353; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload8354; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm16, %zmm0 {%k1}8355; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm29, %zmm4 {%k2}8356; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm0, %zmm4 {%k3}8357; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8358; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm0, %zmm28 {%k1}8359; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm30, %zmm2 {%k2}8360; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm28, %zmm2 {%k3}8361; AVX512DQ-BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8362; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm0, %zmm27 {%k1}8363; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm5, %zmm1 {%k2}8364; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm27, %zmm1 {%k3}8365; AVX512DQ-BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax8366; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm1, 896(%rax)8367; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm2, 960(%rax)8368; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm4, 768(%rax)8369; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm6, 832(%rax)8370; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm7, 640(%rax)8371; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm8, 704(%rax)8372; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm10, 512(%rax)8373; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm12, 576(%rax)8374; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm3, 384(%rax)8375; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm9, 448(%rax)8376; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm11, 256(%rax)8377; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm13, 320(%rax)8378; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm14, 128(%rax)8379; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm15, 192(%rax)8380; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm17, (%rax)8381; AVX512DQ-BW-NEXT:    vmovdqa64 %zmm18, 64(%rax)8382; AVX512DQ-BW-NEXT:    addq $2056, %rsp # imm = 0x8088383; AVX512DQ-BW-NEXT:    vzeroupper8384; AVX512DQ-BW-NEXT:    retq8385;8386; AVX512DQ-BW-FCP-LABEL: store_i32_stride8_vf32:8387; AVX512DQ-BW-FCP:       # %bb.0:8388; AVX512DQ-BW-FCP-NEXT:    subq $2056, %rsp # imm = 0x8088389; AVX512DQ-BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax8390; AVX512DQ-BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r108391; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 (%r8), %zmm18392; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 (%r9), %zmm278393; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 (%r10), %zmm28394; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 64(%r10), %zmm08395; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 (%rax), %zmm308396; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 64(%rax), %zmm288397; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm3 = [0,0,0,0,0,0,2,18,0,0,0,0,0,0,3,19]8398; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm2, %zmm58399; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm2, %zmm48400; AVX512DQ-BW-FCP-NEXT:    vpermt2d %zmm30, %zmm3, %zmm58401; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8402; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [0,0,0,0,2,18,0,0,0,0,0,0,3,19,0,0]8403; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm1, %zmm58404; AVX512DQ-BW-FCP-NEXT:    vpermt2d %zmm27, %zmm2, %zmm58405; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8406; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm5 = [0,0,0,0,0,0,0,16,0,0,0,0,0,0,1,17]8407; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm4, %zmm68408; AVX512DQ-BW-FCP-NEXT:    vpermt2d %zmm30, %zmm5, %zmm68409; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8410; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm6 = [0,0,0,0,0,16,0,0,0,0,0,0,1,17,0,0]8411; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm1, %zmm78412; AVX512DQ-BW-FCP-NEXT:    vpermt2d %zmm27, %zmm6, %zmm78413; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8414; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm7 = [0,0,0,0,0,0,6,22,0,0,0,0,0,0,7,23]8415; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm4, %zmm88416; AVX512DQ-BW-FCP-NEXT:    vpermt2d %zmm30, %zmm7, %zmm88417; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8418; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm8 = [0,0,0,0,6,22,0,0,0,0,0,0,7,23,0,0]8419; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm1, %zmm98420; AVX512DQ-BW-FCP-NEXT:    vpermt2d %zmm27, %zmm8, %zmm98421; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8422; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm9 = [0,0,0,0,0,0,4,20,0,0,0,0,0,0,5,21]8423; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm4, %zmm108424; AVX512DQ-BW-FCP-NEXT:    vpermt2d %zmm30, %zmm9, %zmm108425; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8426; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm10 = [0,0,0,0,4,20,0,0,0,0,0,0,5,21,0,0]8427; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm1, %zmm118428; AVX512DQ-BW-FCP-NEXT:    vpermt2d %zmm27, %zmm10, %zmm118429; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8430; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm11 = [0,0,0,0,0,0,10,26,0,0,0,0,0,0,11,27]8431; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm4, %zmm128432; AVX512DQ-BW-FCP-NEXT:    vpermt2d %zmm30, %zmm11, %zmm128433; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8434; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [0,0,0,0,10,26,0,0,0,0,0,0,11,27,0,0]8435; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm1, %zmm138436; AVX512DQ-BW-FCP-NEXT:    vpermt2d %zmm27, %zmm12, %zmm138437; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8438; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm13 = [0,0,0,0,0,0,8,24,0,0,0,0,0,0,9,25]8439; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm4, %zmm148440; AVX512DQ-BW-FCP-NEXT:    vpermt2d %zmm30, %zmm13, %zmm148441; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8442; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm14 = [0,0,0,0,8,24,0,0,0,0,0,0,9,25,0,0]8443; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm15 = [0,0,0,0,0,0,14,30,0,0,0,0,0,0,15,31]8444; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm4, %zmm168445; AVX512DQ-BW-FCP-NEXT:    vpermt2d %zmm30, %zmm15, %zmm168446; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm16, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8447; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm31 = [0,0,0,0,0,0,12,28,0,0,0,0,0,0,13,29]8448; AVX512DQ-BW-FCP-NEXT:    vpermt2d %zmm30, %zmm31, %zmm48449; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8450; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm1, %zmm48451; AVX512DQ-BW-FCP-NEXT:    vpermt2d %zmm27, %zmm14, %zmm48452; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8453; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm28, %zmm0, %zmm38454; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8455; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm28, %zmm0, %zmm58456; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8457; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm28, %zmm0, %zmm78458; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8459; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm28, %zmm0, %zmm98460; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8461; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm28, %zmm0, %zmm118462; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8463; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm28, %zmm0, %zmm138464; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8465; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm28, %zmm0, %zmm158466; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8467; AVX512DQ-BW-FCP-NEXT:    vpermt2d %zmm28, %zmm31, %zmm08468; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8469; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm28 = [0,0,0,0,14,30,0,0,0,0,0,0,15,31,0,0]8470; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm1, %zmm318471; AVX512DQ-BW-FCP-NEXT:    vpermt2d %zmm27, %zmm28, %zmm318472; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm0 = [0,0,0,0,12,28,0,0,0,0,0,0,13,29,0,0]8473; AVX512DQ-BW-FCP-NEXT:    vpermt2d %zmm27, %zmm0, %zmm18474; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8475; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 64(%r8), %zmm278476; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 64(%r9), %zmm18477; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm1, %zmm27, %zmm28478; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8479; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm1, %zmm27, %zmm68480; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8481; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm1, %zmm27, %zmm88482; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8483; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm1, %zmm27, %zmm108484; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8485; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm1, %zmm27, %zmm128486; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8487; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm1, %zmm27, %zmm148488; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8489; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm1, %zmm27, %zmm288490; AVX512DQ-BW-FCP-NEXT:    vpermt2d %zmm1, %zmm0, %zmm278491; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 (%rdx), %zmm168492; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 (%rcx), %zmm08493; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [0,0,2,18,0,0,0,0,0,0,3,19,0,0,0,0]8494; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm16, %zmm208495; AVX512DQ-BW-FCP-NEXT:    vpermt2d %zmm0, %zmm2, %zmm208496; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm3 = [0,0,0,16,0,0,0,0,0,0,1,17,0,0,0,0]8497; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm16, %zmm218498; AVX512DQ-BW-FCP-NEXT:    vpermt2d %zmm0, %zmm3, %zmm218499; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [0,0,6,22,0,0,0,0,0,0,7,23,0,0,0,0]8500; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm16, %zmm228501; AVX512DQ-BW-FCP-NEXT:    vpermt2d %zmm0, %zmm4, %zmm228502; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm6 = [0,0,4,20,0,0,0,0,0,0,5,21,0,0,0,0]8503; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm16, %zmm238504; AVX512DQ-BW-FCP-NEXT:    vpermt2d %zmm0, %zmm6, %zmm238505; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm7 = [0,0,10,26,0,0,0,0,0,0,11,27,0,0,0,0]8506; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm16, %zmm248507; AVX512DQ-BW-FCP-NEXT:    vpermt2d %zmm0, %zmm7, %zmm248508; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm29 = [0,0,8,24,0,0,0,0,0,0,9,25,0,0,0,0]8509; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm16, %zmm258510; AVX512DQ-BW-FCP-NEXT:    vpermt2d %zmm0, %zmm29, %zmm258511; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm30 = [0,0,14,30,0,0,0,0,0,0,15,31,0,0,0,0]8512; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm16, %zmm268513; AVX512DQ-BW-FCP-NEXT:    vpermt2d %zmm0, %zmm30, %zmm268514; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,12,28,0,0,0,0,0,0,13,29,0,0,0,0]8515; AVX512DQ-BW-FCP-NEXT:    vpermt2d %zmm0, %zmm1, %zmm168516; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 64(%rdx), %zmm58517; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 64(%rcx), %zmm08518; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm0, %zmm5, %zmm28519; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8520; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm0, %zmm5, %zmm38521; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8522; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm0, %zmm5, %zmm48523; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm4, (%rsp) # 64-byte Spill8524; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm0, %zmm5, %zmm68525; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8526; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm0, %zmm5, %zmm78527; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8528; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm0, %zmm5, %zmm298529; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm0, %zmm5, %zmm308530; AVX512DQ-BW-FCP-NEXT:    vpermt2d %zmm0, %zmm1, %zmm58531; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 (%rdi), %zmm38532; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 (%rsi), %zmm08533; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm12 = [2,18,0,0,0,0,0,0,3,19,0,0,0,0,0,0]8534; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm3, %zmm188535; AVX512DQ-BW-FCP-NEXT:    vpermt2d %zmm0, %zmm12, %zmm188536; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm10 = [0,16,0,0,0,0,0,0,1,17,0,0,0,0,0,0]8537; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm3, %zmm178538; AVX512DQ-BW-FCP-NEXT:    vpermt2d %zmm0, %zmm10, %zmm178539; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm8 = [6,22,0,0,0,0,0,0,7,23,0,0,0,0,0,0]8540; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm3, %zmm158541; AVX512DQ-BW-FCP-NEXT:    vpermt2d %zmm0, %zmm8, %zmm158542; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm7 = [4,20,0,0,0,0,0,0,5,21,0,0,0,0,0,0]8543; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm3, %zmm148544; AVX512DQ-BW-FCP-NEXT:    vpermt2d %zmm0, %zmm7, %zmm148545; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm6 = [10,26,0,0,0,0,0,0,11,27,0,0,0,0,0,0]8546; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm3, %zmm138547; AVX512DQ-BW-FCP-NEXT:    vpermt2d %zmm0, %zmm6, %zmm138548; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [8,24,0,0,0,0,0,0,9,25,0,0,0,0,0,0]8549; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm3, %zmm118550; AVX512DQ-BW-FCP-NEXT:    vpermt2d %zmm0, %zmm4, %zmm118551; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [14,30,0,0,0,0,0,0,15,31,0,0,0,0,0,0]8552; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm3, %zmm98553; AVX512DQ-BW-FCP-NEXT:    vpermt2d %zmm0, %zmm2, %zmm98554; AVX512DQ-BW-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm19 = [12,28,0,0,0,0,0,0,13,29,0,0,0,0,0,0]8555; AVX512DQ-BW-FCP-NEXT:    vpermt2d %zmm0, %zmm19, %zmm38556; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 64(%rdi), %zmm18557; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 64(%rsi), %zmm08558; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm0, %zmm1, %zmm128559; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm0, %zmm1, %zmm108560; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm0, %zmm1, %zmm88561; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm0, %zmm1, %zmm78562; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm0, %zmm1, %zmm68563; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm0, %zmm1, %zmm48564; AVX512DQ-BW-FCP-NEXT:    vpermi2d %zmm0, %zmm1, %zmm28565; AVX512DQ-BW-FCP-NEXT:    vpermt2d %zmm0, %zmm19, %zmm18566; AVX512DQ-BW-FCP-NEXT:    movb $-120, %al8567; AVX512DQ-BW-FCP-NEXT:    kmovd %eax, %k18568; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8569; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload8570; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm19, %zmm0 {%k1}8571; AVX512DQ-BW-FCP-NEXT:    movb $34, %al8572; AVX512DQ-BW-FCP-NEXT:    kmovd %eax, %k28573; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm20, %zmm18 {%k2}8574; AVX512DQ-BW-FCP-NEXT:    movb $-52, %al8575; AVX512DQ-BW-FCP-NEXT:    kmovd %eax, %k38576; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm18 {%k3}8577; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8578; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload8579; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm19 {%k1}8580; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm21, %zmm17 {%k2}8581; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm19, %zmm17 {%k3}8582; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8583; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload8584; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm19 {%k1}8585; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm22, %zmm15 {%k2}8586; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm19, %zmm15 {%k3}8587; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8588; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload8589; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm19 {%k1}8590; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm23, %zmm14 {%k2}8591; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm19, %zmm14 {%k3}8592; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8593; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload8594; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm19 {%k1}8595; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm24, %zmm13 {%k2}8596; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm19, %zmm13 {%k3}8597; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8598; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload8599; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm19 {%k1}8600; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm25, %zmm11 {%k2}8601; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm19, %zmm11 {%k3}8602; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8603; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm31 {%k1}8604; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm26, %zmm9 {%k2}8605; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm31, %zmm9 {%k3}8606; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8607; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm19 # 64-byte Reload8608; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm19, %zmm0 {%k1}8609; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm16, %zmm3 {%k2}8610; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm3 {%k3}8611; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8612; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload8613; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm16, %zmm0 {%k1}8614; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload8615; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm16, %zmm12 {%k2}8616; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm12 {%k3}8617; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8618; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload8619; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm16, %zmm0 {%k1}8620; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload8621; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm16, %zmm10 {%k2}8622; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm10 {%k3}8623; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8624; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload8625; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm16, %zmm0 {%k1}8626; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 (%rsp), %zmm16 # 64-byte Reload8627; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm16, %zmm8 {%k2}8628; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm8 {%k3}8629; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8630; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload8631; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm16, %zmm0 {%k1}8632; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload8633; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm16, %zmm7 {%k2}8634; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm7 {%k3}8635; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8636; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload8637; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm16, %zmm0 {%k1}8638; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload8639; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm16, %zmm6 {%k2}8640; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm6 {%k3}8641; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8642; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm16 # 64-byte Reload8643; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm16, %zmm0 {%k1}8644; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm29, %zmm4 {%k2}8645; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm4 {%k3}8646; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8647; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm28 {%k1}8648; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm30, %zmm2 {%k2}8649; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm28, %zmm2 {%k3}8650; AVX512DQ-BW-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8651; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm0, %zmm27 {%k1}8652; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm5, %zmm1 {%k2}8653; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm27, %zmm1 {%k3}8654; AVX512DQ-BW-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax8655; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm1, 896(%rax)8656; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm2, 960(%rax)8657; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm4, 768(%rax)8658; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm6, 832(%rax)8659; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm7, 640(%rax)8660; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm8, 704(%rax)8661; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm10, 512(%rax)8662; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm12, 576(%rax)8663; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm3, 384(%rax)8664; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm9, 448(%rax)8665; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm11, 256(%rax)8666; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm13, 320(%rax)8667; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm14, 128(%rax)8668; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm15, 192(%rax)8669; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm17, (%rax)8670; AVX512DQ-BW-FCP-NEXT:    vmovdqa64 %zmm18, 64(%rax)8671; AVX512DQ-BW-FCP-NEXT:    addq $2056, %rsp # imm = 0x8088672; AVX512DQ-BW-FCP-NEXT:    vzeroupper8673; AVX512DQ-BW-FCP-NEXT:    retq8674  %in.vec0 = load <32 x i32>, ptr %in.vecptr0, align 648675  %in.vec1 = load <32 x i32>, ptr %in.vecptr1, align 648676  %in.vec2 = load <32 x i32>, ptr %in.vecptr2, align 648677  %in.vec3 = load <32 x i32>, ptr %in.vecptr3, align 648678  %in.vec4 = load <32 x i32>, ptr %in.vecptr4, align 648679  %in.vec5 = load <32 x i32>, ptr %in.vecptr5, align 648680  %in.vec6 = load <32 x i32>, ptr %in.vecptr6, align 648681  %in.vec7 = load <32 x i32>, ptr %in.vecptr7, align 648682  %1 = shufflevector <32 x i32> %in.vec0, <32 x i32> %in.vec1, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>8683  %2 = shufflevector <32 x i32> %in.vec2, <32 x i32> %in.vec3, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>8684  %3 = shufflevector <32 x i32> %in.vec4, <32 x i32> %in.vec5, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>8685  %4 = shufflevector <32 x i32> %in.vec6, <32 x i32> %in.vec7, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>8686  %5 = shufflevector <64 x i32> %1, <64 x i32> %2, <128 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 64, i32 65, i32 66, i32 67, i32 68, i32 69, i32 70, i32 71, i32 72, i32 73, i32 74, i32 75, i32 76, i32 77, i32 78, i32 79, i32 80, i32 81, i32 82, i32 83, i32 84, i32 85, i32 86, i32 87, i32 88, i32 89, i32 90, i32 91, i32 92, i32 93, i32 94, i32 95, i32 96, i32 97, i32 98, i32 99, i32 100, i32 101, i32 102, i32 103, i32 104, i32 105, i32 106, i32 107, i32 108, i32 109, i32 110, i32 111, i32 112, i32 113, i32 114, i32 115, i32 116, i32 117, i32 118, i32 119, i32 120, i32 121, i32 122, i32 123, i32 124, i32 125, i32 126, i32 127>8687  %6 = shufflevector <64 x i32> %3, <64 x i32> %4, <128 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 64, i32 65, i32 66, i32 67, i32 68, i32 69, i32 70, i32 71, i32 72, i32 73, i32 74, i32 75, i32 76, i32 77, i32 78, i32 79, i32 80, i32 81, i32 82, i32 83, i32 84, i32 85, i32 86, i32 87, i32 88, i32 89, i32 90, i32 91, i32 92, i32 93, i32 94, i32 95, i32 96, i32 97, i32 98, i32 99, i32 100, i32 101, i32 102, i32 103, i32 104, i32 105, i32 106, i32 107, i32 108, i32 109, i32 110, i32 111, i32 112, i32 113, i32 114, i32 115, i32 116, i32 117, i32 118, i32 119, i32 120, i32 121, i32 122, i32 123, i32 124, i32 125, i32 126, i32 127>8688  %7 = shufflevector <128 x i32> %5, <128 x i32> %6, <256 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 64, i32 65, i32 66, i32 67, i32 68, i32 69, i32 70, i32 71, i32 72, i32 73, i32 74, i32 75, i32 76, i32 77, i32 78, i32 79, i32 80, i32 81, i32 82, i32 83, i32 84, i32 85, i32 86, i32 87, i32 88, i32 89, i32 90, i32 91, i32 92, i32 93, i32 94, i32 95, i32 96, i32 97, i32 98, i32 99, i32 100, i32 101, i32 102, i32 103, i32 104, i32 105, i32 106, i32 107, i32 108, i32 109, i32 110, i32 111, i32 112, i32 113, i32 114, i32 115, i32 116, i32 117, i32 118, i32 119, i32 120, i32 121, i32 122, i32 123, i32 124, i32 125, i32 126, i32 127, i32 128, i32 129, i32 130, i32 131, i32 132, i32 133, i32 134, i32 135, i32 136, i32 137, i32 138, i32 139, i32 140, i32 141, i32 142, i32 143, i32 144, i32 145, i32 146, i32 147, i32 148, i32 149, i32 150, i32 151, i32 152, i32 153, i32 154, i32 155, i32 156, i32 157, i32 158, i32 159, i32 160, i32 161, i32 162, i32 163, i32 164, i32 165, i32 166, i32 167, i32 168, i32 169, i32 170, i32 171, i32 172, i32 173, i32 174, i32 175, i32 176, i32 177, i32 178, i32 179, i32 180, i32 181, i32 182, i32 183, i32 184, i32 185, i32 186, i32 187, i32 188, i32 189, i32 190, i32 191, i32 192, i32 193, i32 194, i32 195, i32 196, i32 197, i32 198, i32 199, i32 200, i32 201, i32 202, i32 203, i32 204, i32 205, i32 206, i32 207, i32 208, i32 209, i32 210, i32 211, i32 212, i32 213, i32 214, i32 215, i32 216, i32 217, i32 218, i32 219, i32 220, i32 221, i32 222, i32 223, i32 224, i32 225, i32 226, i32 227, i32 228, i32 229, i32 230, i32 231, i32 232, i32 233, i32 234, i32 235, i32 236, i32 237, i32 238, i32 239, i32 240, i32 241, i32 242, i32 243, i32 244, i32 245, i32 246, i32 247, i32 248, i32 249, i32 250, i32 251, i32 252, i32 253, i32 254, i32 255>8689  %interleaved.vec = shufflevector <256 x i32> %7, <256 x i32> poison, <256 x i32> <i32 0, i32 32, i32 64, i32 96, i32 128, i32 160, i32 192, i32 224, i32 1, i32 33, i32 65, i32 97, i32 129, i32 161, i32 193, i32 225, i32 2, i32 34, i32 66, i32 98, i32 130, i32 162, i32 194, i32 226, i32 3, i32 35, i32 67, i32 99, i32 131, i32 163, i32 195, i32 227, i32 4, i32 36, i32 68, i32 100, i32 132, i32 164, i32 196, i32 228, i32 5, i32 37, i32 69, i32 101, i32 133, i32 165, i32 197, i32 229, i32 6, i32 38, i32 70, i32 102, i32 134, i32 166, i32 198, i32 230, i32 7, i32 39, i32 71, i32 103, i32 135, i32 167, i32 199, i32 231, i32 8, i32 40, i32 72, i32 104, i32 136, i32 168, i32 200, i32 232, i32 9, i32 41, i32 73, i32 105, i32 137, i32 169, i32 201, i32 233, i32 10, i32 42, i32 74, i32 106, i32 138, i32 170, i32 202, i32 234, i32 11, i32 43, i32 75, i32 107, i32 139, i32 171, i32 203, i32 235, i32 12, i32 44, i32 76, i32 108, i32 140, i32 172, i32 204, i32 236, i32 13, i32 45, i32 77, i32 109, i32 141, i32 173, i32 205, i32 237, i32 14, i32 46, i32 78, i32 110, i32 142, i32 174, i32 206, i32 238, i32 15, i32 47, i32 79, i32 111, i32 143, i32 175, i32 207, i32 239, i32 16, i32 48, i32 80, i32 112, i32 144, i32 176, i32 208, i32 240, i32 17, i32 49, i32 81, i32 113, i32 145, i32 177, i32 209, i32 241, i32 18, i32 50, i32 82, i32 114, i32 146, i32 178, i32 210, i32 242, i32 19, i32 51, i32 83, i32 115, i32 147, i32 179, i32 211, i32 243, i32 20, i32 52, i32 84, i32 116, i32 148, i32 180, i32 212, i32 244, i32 21, i32 53, i32 85, i32 117, i32 149, i32 181, i32 213, i32 245, i32 22, i32 54, i32 86, i32 118, i32 150, i32 182, i32 214, i32 246, i32 23, i32 55, i32 87, i32 119, i32 151, i32 183, i32 215, i32 247, i32 24, i32 56, i32 88, i32 120, i32 152, i32 184, i32 216, i32 248, i32 25, i32 57, i32 89, i32 121, i32 153, i32 185, i32 217, i32 249, i32 26, i32 58, i32 90, i32 122, i32 154, i32 186, i32 218, i32 250, i32 27, i32 59, i32 91, i32 123, i32 155, i32 187, i32 219, i32 251, i32 28, i32 60, i32 92, i32 124, i32 156, i32 188, i32 220, i32 252, i32 29, i32 61, i32 93, i32 125, i32 157, i32 189, i32 221, i32 253, i32 30, i32 62, i32 94, i32 126, i32 158, i32 190, i32 222, i32 254, i32 31, i32 63, i32 95, i32 127, i32 159, i32 191, i32 223, i32 255>8690  store <256 x i32> %interleaved.vec, ptr %out.vec, align 648691  ret void8692}8693 8694define void @store_i32_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 {8695; SSE-LABEL: store_i32_stride8_vf64:8696; SSE:       # %bb.0:8697; SSE-NEXT:    subq $1752, %rsp # imm = 0x6D88698; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %rax8699; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %r108700; SSE-NEXT:    movaps (%rdi), %xmm98701; SSE-NEXT:    movaps 16(%rdi), %xmm108702; SSE-NEXT:    movaps (%rsi), %xmm28703; SSE-NEXT:    movaps 16(%rsi), %xmm08704; SSE-NEXT:    movaps (%rdx), %xmm38705; SSE-NEXT:    movaps 16(%rdx), %xmm18706; SSE-NEXT:    movaps (%rcx), %xmm48707; SSE-NEXT:    movaps (%r8), %xmm118708; SSE-NEXT:    movaps (%r9), %xmm58709; SSE-NEXT:    movaps (%r10), %xmm68710; SSE-NEXT:    movaps (%rax), %xmm78711; SSE-NEXT:    movaps %xmm4, %xmm88712; SSE-NEXT:    movlhps {{.*#+}} xmm8 = xmm8[0],xmm3[0]8713; SSE-NEXT:    movaps %xmm9, %xmm138714; SSE-NEXT:    unpcklps {{.*#+}} xmm13 = xmm13[0],xmm2[0],xmm13[1],xmm2[1]8715; SSE-NEXT:    movaps %xmm13, %xmm128716; SSE-NEXT:    shufps {{.*#+}} xmm12 = xmm12[0,1],xmm8[2,0]8717; SSE-NEXT:    movaps %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8718; SSE-NEXT:    movaps %xmm7, %xmm88719; SSE-NEXT:    movlhps {{.*#+}} xmm8 = xmm8[0],xmm6[0]8720; SSE-NEXT:    movaps %xmm11, %xmm148721; SSE-NEXT:    unpcklps {{.*#+}} xmm14 = xmm14[0],xmm5[0],xmm14[1],xmm5[1]8722; SSE-NEXT:    movaps %xmm14, %xmm128723; SSE-NEXT:    shufps {{.*#+}} xmm12 = xmm12[0,1],xmm8[2,0]8724; SSE-NEXT:    movaps %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8725; SSE-NEXT:    movaps %xmm4, %xmm88726; SSE-NEXT:    shufps {{.*#+}} xmm8 = xmm8[1,1],xmm3[1,1]8727; SSE-NEXT:    shufps {{.*#+}} xmm13 = xmm13[2,3],xmm8[2,0]8728; SSE-NEXT:    movaps %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8729; SSE-NEXT:    movaps %xmm7, %xmm88730; SSE-NEXT:    shufps {{.*#+}} xmm8 = xmm8[1,1],xmm6[1,1]8731; SSE-NEXT:    shufps {{.*#+}} xmm14 = xmm14[2,3],xmm8[2,0]8732; SSE-NEXT:    movaps %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8733; SSE-NEXT:    movaps %xmm3, %xmm88734; SSE-NEXT:    unpckhpd {{.*#+}} xmm8 = xmm8[1],xmm4[1]8735; SSE-NEXT:    unpckhps {{.*#+}} xmm9 = xmm9[2],xmm2[2],xmm9[3],xmm2[3]8736; SSE-NEXT:    movaps %xmm9, %xmm28737; SSE-NEXT:    shufps {{.*#+}} xmm2 = xmm2[0,1],xmm8[0,2]8738; SSE-NEXT:    movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8739; SSE-NEXT:    movaps %xmm6, %xmm28740; SSE-NEXT:    unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm7[1]8741; SSE-NEXT:    unpckhps {{.*#+}} xmm11 = xmm11[2],xmm5[2],xmm11[3],xmm5[3]8742; SSE-NEXT:    movaps %xmm11, %xmm58743; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[0,1],xmm2[0,2]8744; SSE-NEXT:    movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8745; SSE-NEXT:    movaps 16(%rcx), %xmm58746; SSE-NEXT:    shufps {{.*#+}} xmm3 = xmm3[3,3],xmm4[3,3]8747; SSE-NEXT:    movaps 16(%r10), %xmm28748; SSE-NEXT:    shufps {{.*#+}} xmm9 = xmm9[2,3],xmm3[0,2]8749; SSE-NEXT:    movaps %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8750; SSE-NEXT:    shufps {{.*#+}} xmm6 = xmm6[3,3],xmm7[3,3]8751; SSE-NEXT:    shufps {{.*#+}} xmm11 = xmm11[2,3],xmm6[0,2]8752; SSE-NEXT:    movaps %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8753; SSE-NEXT:    movaps %xmm5, %xmm38754; SSE-NEXT:    movlhps {{.*#+}} xmm3 = xmm3[0],xmm1[0]8755; SSE-NEXT:    movaps %xmm10, %xmm68756; SSE-NEXT:    unpcklps {{.*#+}} xmm6 = xmm6[0],xmm0[0],xmm6[1],xmm0[1]8757; SSE-NEXT:    movaps %xmm6, %xmm48758; SSE-NEXT:    movaps %xmm6, %xmm78759; SSE-NEXT:    shufps {{.*#+}} xmm4 = xmm4[0,1],xmm3[2,0]8760; SSE-NEXT:    movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8761; SSE-NEXT:    movaps 16(%rax), %xmm38762; SSE-NEXT:    movaps %xmm3, %xmm48763; SSE-NEXT:    movlhps {{.*#+}} xmm4 = xmm4[0],xmm2[0]8764; SSE-NEXT:    movaps 16(%r8), %xmm118765; SSE-NEXT:    movaps 16(%r9), %xmm68766; SSE-NEXT:    movaps %xmm11, %xmm98767; SSE-NEXT:    unpcklps {{.*#+}} xmm9 = xmm9[0],xmm6[0],xmm9[1],xmm6[1]8768; SSE-NEXT:    movaps %xmm9, %xmm88769; SSE-NEXT:    shufps {{.*#+}} xmm8 = xmm8[0,1],xmm4[2,0]8770; SSE-NEXT:    movaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8771; SSE-NEXT:    movaps %xmm5, %xmm48772; SSE-NEXT:    shufps {{.*#+}} xmm4 = xmm4[1,1],xmm1[1,1]8773; SSE-NEXT:    shufps {{.*#+}} xmm7 = xmm7[2,3],xmm4[2,0]8774; SSE-NEXT:    movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8775; SSE-NEXT:    movaps %xmm3, %xmm48776; SSE-NEXT:    shufps {{.*#+}} xmm4 = xmm4[1,1],xmm2[1,1]8777; SSE-NEXT:    shufps {{.*#+}} xmm9 = xmm9[2,3],xmm4[2,0]8778; SSE-NEXT:    movaps %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8779; SSE-NEXT:    unpckhps {{.*#+}} xmm10 = xmm10[2],xmm0[2],xmm10[3],xmm0[3]8780; SSE-NEXT:    movaps %xmm1, %xmm08781; SSE-NEXT:    unpckhpd {{.*#+}} xmm0 = xmm0[1],xmm5[1]8782; SSE-NEXT:    movaps %xmm10, %xmm48783; SSE-NEXT:    shufps {{.*#+}} xmm4 = xmm4[0,1],xmm0[0,2]8784; SSE-NEXT:    movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8785; SSE-NEXT:    unpckhps {{.*#+}} xmm11 = xmm11[2],xmm6[2],xmm11[3],xmm6[3]8786; SSE-NEXT:    movaps %xmm2, %xmm08787; SSE-NEXT:    unpckhpd {{.*#+}} xmm0 = xmm0[1],xmm3[1]8788; SSE-NEXT:    movaps %xmm11, %xmm48789; SSE-NEXT:    shufps {{.*#+}} xmm4 = xmm4[0,1],xmm0[0,2]8790; SSE-NEXT:    movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8791; SSE-NEXT:    shufps {{.*#+}} xmm1 = xmm1[3,3],xmm5[3,3]8792; SSE-NEXT:    shufps {{.*#+}} xmm10 = xmm10[2,3],xmm1[0,2]8793; SSE-NEXT:    movaps %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8794; SSE-NEXT:    shufps {{.*#+}} xmm2 = xmm2[3,3],xmm3[3,3]8795; SSE-NEXT:    shufps {{.*#+}} xmm11 = xmm11[2,3],xmm2[0,2]8796; SSE-NEXT:    movaps %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8797; SSE-NEXT:    movaps 32(%rdx), %xmm08798; SSE-NEXT:    movaps 32(%rcx), %xmm18799; SSE-NEXT:    movaps %xmm1, %xmm28800; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]8801; SSE-NEXT:    movaps 32(%rdi), %xmm78802; SSE-NEXT:    movaps 32(%rsi), %xmm48803; SSE-NEXT:    movaps %xmm7, %xmm58804; SSE-NEXT:    unpcklps {{.*#+}} xmm5 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]8805; SSE-NEXT:    movaps %xmm5, %xmm38806; SSE-NEXT:    movaps %xmm5, %xmm88807; SSE-NEXT:    shufps {{.*#+}} xmm3 = xmm3[0,1],xmm2[2,0]8808; SSE-NEXT:    movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8809; SSE-NEXT:    movaps 32(%r10), %xmm28810; SSE-NEXT:    movaps 32(%rax), %xmm38811; SSE-NEXT:    movaps %xmm3, %xmm58812; SSE-NEXT:    movlhps {{.*#+}} xmm5 = xmm5[0],xmm2[0]8813; SSE-NEXT:    movaps 32(%r8), %xmm118814; SSE-NEXT:    movaps 32(%r9), %xmm68815; SSE-NEXT:    movaps %xmm11, %xmm108816; SSE-NEXT:    unpcklps {{.*#+}} xmm10 = xmm10[0],xmm6[0],xmm10[1],xmm6[1]8817; SSE-NEXT:    movaps %xmm10, %xmm98818; SSE-NEXT:    shufps {{.*#+}} xmm9 = xmm9[0,1],xmm5[2,0]8819; SSE-NEXT:    movaps %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8820; SSE-NEXT:    movaps %xmm1, %xmm58821; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[1,1],xmm0[1,1]8822; SSE-NEXT:    shufps {{.*#+}} xmm8 = xmm8[2,3],xmm5[2,0]8823; SSE-NEXT:    movaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8824; SSE-NEXT:    movaps %xmm3, %xmm58825; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[1,1],xmm2[1,1]8826; SSE-NEXT:    shufps {{.*#+}} xmm10 = xmm10[2,3],xmm5[2,0]8827; SSE-NEXT:    movaps %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8828; SSE-NEXT:    unpckhps {{.*#+}} xmm7 = xmm7[2],xmm4[2],xmm7[3],xmm4[3]8829; SSE-NEXT:    movaps %xmm0, %xmm48830; SSE-NEXT:    unpckhpd {{.*#+}} xmm4 = xmm4[1],xmm1[1]8831; SSE-NEXT:    movaps %xmm7, %xmm58832; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[0,1],xmm4[0,2]8833; SSE-NEXT:    movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8834; SSE-NEXT:    unpckhps {{.*#+}} xmm11 = xmm11[2],xmm6[2],xmm11[3],xmm6[3]8835; SSE-NEXT:    movaps %xmm2, %xmm48836; SSE-NEXT:    unpckhpd {{.*#+}} xmm4 = xmm4[1],xmm3[1]8837; SSE-NEXT:    movaps %xmm11, %xmm58838; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[0,1],xmm4[0,2]8839; SSE-NEXT:    movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8840; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[3,3],xmm1[3,3]8841; SSE-NEXT:    shufps {{.*#+}} xmm7 = xmm7[2,3],xmm0[0,2]8842; SSE-NEXT:    movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8843; SSE-NEXT:    shufps {{.*#+}} xmm2 = xmm2[3,3],xmm3[3,3]8844; SSE-NEXT:    shufps {{.*#+}} xmm11 = xmm11[2,3],xmm2[0,2]8845; SSE-NEXT:    movaps %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8846; SSE-NEXT:    movaps 48(%rdx), %xmm08847; SSE-NEXT:    movaps 48(%rcx), %xmm18848; SSE-NEXT:    movaps %xmm1, %xmm28849; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]8850; SSE-NEXT:    movaps 48(%rdi), %xmm78851; SSE-NEXT:    movaps 48(%rsi), %xmm48852; SSE-NEXT:    movaps %xmm7, %xmm58853; SSE-NEXT:    unpcklps {{.*#+}} xmm5 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]8854; SSE-NEXT:    movaps %xmm5, %xmm38855; SSE-NEXT:    movaps %xmm5, %xmm88856; SSE-NEXT:    shufps {{.*#+}} xmm3 = xmm3[0,1],xmm2[2,0]8857; SSE-NEXT:    movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8858; SSE-NEXT:    movaps 48(%r10), %xmm28859; SSE-NEXT:    movaps 48(%rax), %xmm38860; SSE-NEXT:    movaps %xmm3, %xmm58861; SSE-NEXT:    movlhps {{.*#+}} xmm5 = xmm5[0],xmm2[0]8862; SSE-NEXT:    movaps 48(%r8), %xmm118863; SSE-NEXT:    movaps 48(%r9), %xmm68864; SSE-NEXT:    movaps %xmm11, %xmm108865; SSE-NEXT:    unpcklps {{.*#+}} xmm10 = xmm10[0],xmm6[0],xmm10[1],xmm6[1]8866; SSE-NEXT:    movaps %xmm10, %xmm98867; SSE-NEXT:    shufps {{.*#+}} xmm9 = xmm9[0,1],xmm5[2,0]8868; SSE-NEXT:    movaps %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8869; SSE-NEXT:    movaps %xmm1, %xmm58870; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[1,1],xmm0[1,1]8871; SSE-NEXT:    shufps {{.*#+}} xmm8 = xmm8[2,3],xmm5[2,0]8872; SSE-NEXT:    movaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8873; SSE-NEXT:    movaps %xmm3, %xmm58874; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[1,1],xmm2[1,1]8875; SSE-NEXT:    shufps {{.*#+}} xmm10 = xmm10[2,3],xmm5[2,0]8876; SSE-NEXT:    movaps %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8877; SSE-NEXT:    unpckhps {{.*#+}} xmm7 = xmm7[2],xmm4[2],xmm7[3],xmm4[3]8878; SSE-NEXT:    movaps %xmm0, %xmm48879; SSE-NEXT:    unpckhpd {{.*#+}} xmm4 = xmm4[1],xmm1[1]8880; SSE-NEXT:    movaps %xmm7, %xmm58881; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[0,1],xmm4[0,2]8882; SSE-NEXT:    movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8883; SSE-NEXT:    unpckhps {{.*#+}} xmm11 = xmm11[2],xmm6[2],xmm11[3],xmm6[3]8884; SSE-NEXT:    movaps %xmm2, %xmm48885; SSE-NEXT:    unpckhpd {{.*#+}} xmm4 = xmm4[1],xmm3[1]8886; SSE-NEXT:    movaps %xmm11, %xmm58887; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[0,1],xmm4[0,2]8888; SSE-NEXT:    movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8889; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[3,3],xmm1[3,3]8890; SSE-NEXT:    shufps {{.*#+}} xmm7 = xmm7[2,3],xmm0[0,2]8891; SSE-NEXT:    movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8892; SSE-NEXT:    shufps {{.*#+}} xmm2 = xmm2[3,3],xmm3[3,3]8893; SSE-NEXT:    shufps {{.*#+}} xmm11 = xmm11[2,3],xmm2[0,2]8894; SSE-NEXT:    movaps %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8895; SSE-NEXT:    movaps 64(%rdx), %xmm08896; SSE-NEXT:    movaps 64(%rcx), %xmm18897; SSE-NEXT:    movaps %xmm1, %xmm28898; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]8899; SSE-NEXT:    movaps 64(%rdi), %xmm78900; SSE-NEXT:    movaps 64(%rsi), %xmm48901; SSE-NEXT:    movaps %xmm7, %xmm58902; SSE-NEXT:    unpcklps {{.*#+}} xmm5 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]8903; SSE-NEXT:    movaps %xmm5, %xmm38904; SSE-NEXT:    movaps %xmm5, %xmm88905; SSE-NEXT:    shufps {{.*#+}} xmm3 = xmm3[0,1],xmm2[2,0]8906; SSE-NEXT:    movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8907; SSE-NEXT:    movaps 64(%r10), %xmm28908; SSE-NEXT:    movaps 64(%rax), %xmm38909; SSE-NEXT:    movaps %xmm3, %xmm58910; SSE-NEXT:    movlhps {{.*#+}} xmm5 = xmm5[0],xmm2[0]8911; SSE-NEXT:    movaps 64(%r8), %xmm118912; SSE-NEXT:    movaps 64(%r9), %xmm68913; SSE-NEXT:    movaps %xmm11, %xmm108914; SSE-NEXT:    unpcklps {{.*#+}} xmm10 = xmm10[0],xmm6[0],xmm10[1],xmm6[1]8915; SSE-NEXT:    movaps %xmm10, %xmm98916; SSE-NEXT:    shufps {{.*#+}} xmm9 = xmm9[0,1],xmm5[2,0]8917; SSE-NEXT:    movaps %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8918; SSE-NEXT:    movaps %xmm1, %xmm58919; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[1,1],xmm0[1,1]8920; SSE-NEXT:    shufps {{.*#+}} xmm8 = xmm8[2,3],xmm5[2,0]8921; SSE-NEXT:    movaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8922; SSE-NEXT:    movaps %xmm3, %xmm58923; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[1,1],xmm2[1,1]8924; SSE-NEXT:    shufps {{.*#+}} xmm10 = xmm10[2,3],xmm5[2,0]8925; SSE-NEXT:    movaps %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8926; SSE-NEXT:    unpckhps {{.*#+}} xmm7 = xmm7[2],xmm4[2],xmm7[3],xmm4[3]8927; SSE-NEXT:    movaps %xmm0, %xmm48928; SSE-NEXT:    unpckhpd {{.*#+}} xmm4 = xmm4[1],xmm1[1]8929; SSE-NEXT:    movaps %xmm7, %xmm58930; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[0,1],xmm4[0,2]8931; SSE-NEXT:    movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8932; SSE-NEXT:    unpckhps {{.*#+}} xmm11 = xmm11[2],xmm6[2],xmm11[3],xmm6[3]8933; SSE-NEXT:    movaps %xmm2, %xmm48934; SSE-NEXT:    unpckhpd {{.*#+}} xmm4 = xmm4[1],xmm3[1]8935; SSE-NEXT:    movaps %xmm11, %xmm58936; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[0,1],xmm4[0,2]8937; SSE-NEXT:    movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8938; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[3,3],xmm1[3,3]8939; SSE-NEXT:    shufps {{.*#+}} xmm7 = xmm7[2,3],xmm0[0,2]8940; SSE-NEXT:    movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8941; SSE-NEXT:    shufps {{.*#+}} xmm2 = xmm2[3,3],xmm3[3,3]8942; SSE-NEXT:    shufps {{.*#+}} xmm11 = xmm11[2,3],xmm2[0,2]8943; SSE-NEXT:    movaps %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8944; SSE-NEXT:    movaps 80(%rdx), %xmm08945; SSE-NEXT:    movaps 80(%rcx), %xmm18946; SSE-NEXT:    movaps %xmm1, %xmm28947; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]8948; SSE-NEXT:    movaps 80(%rdi), %xmm78949; SSE-NEXT:    movaps 80(%rsi), %xmm48950; SSE-NEXT:    movaps %xmm7, %xmm58951; SSE-NEXT:    unpcklps {{.*#+}} xmm5 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]8952; SSE-NEXT:    movaps %xmm5, %xmm38953; SSE-NEXT:    movaps %xmm5, %xmm88954; SSE-NEXT:    shufps {{.*#+}} xmm3 = xmm3[0,1],xmm2[2,0]8955; SSE-NEXT:    movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8956; SSE-NEXT:    movaps 80(%r10), %xmm28957; SSE-NEXT:    movaps 80(%rax), %xmm38958; SSE-NEXT:    movaps %xmm3, %xmm58959; SSE-NEXT:    movlhps {{.*#+}} xmm5 = xmm5[0],xmm2[0]8960; SSE-NEXT:    movaps 80(%r8), %xmm118961; SSE-NEXT:    movaps 80(%r9), %xmm68962; SSE-NEXT:    movaps %xmm11, %xmm108963; SSE-NEXT:    unpcklps {{.*#+}} xmm10 = xmm10[0],xmm6[0],xmm10[1],xmm6[1]8964; SSE-NEXT:    movaps %xmm10, %xmm98965; SSE-NEXT:    shufps {{.*#+}} xmm9 = xmm9[0,1],xmm5[2,0]8966; SSE-NEXT:    movaps %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8967; SSE-NEXT:    movaps %xmm1, %xmm58968; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[1,1],xmm0[1,1]8969; SSE-NEXT:    shufps {{.*#+}} xmm8 = xmm8[2,3],xmm5[2,0]8970; SSE-NEXT:    movaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8971; SSE-NEXT:    movaps %xmm3, %xmm58972; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[1,1],xmm2[1,1]8973; SSE-NEXT:    shufps {{.*#+}} xmm10 = xmm10[2,3],xmm5[2,0]8974; SSE-NEXT:    movaps %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8975; SSE-NEXT:    unpckhps {{.*#+}} xmm7 = xmm7[2],xmm4[2],xmm7[3],xmm4[3]8976; SSE-NEXT:    movaps %xmm0, %xmm48977; SSE-NEXT:    unpckhpd {{.*#+}} xmm4 = xmm4[1],xmm1[1]8978; SSE-NEXT:    movaps %xmm7, %xmm58979; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[0,1],xmm4[0,2]8980; SSE-NEXT:    movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8981; SSE-NEXT:    unpckhps {{.*#+}} xmm11 = xmm11[2],xmm6[2],xmm11[3],xmm6[3]8982; SSE-NEXT:    movaps %xmm2, %xmm48983; SSE-NEXT:    unpckhpd {{.*#+}} xmm4 = xmm4[1],xmm3[1]8984; SSE-NEXT:    movaps %xmm11, %xmm58985; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[0,1],xmm4[0,2]8986; SSE-NEXT:    movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8987; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[3,3],xmm1[3,3]8988; SSE-NEXT:    shufps {{.*#+}} xmm7 = xmm7[2,3],xmm0[0,2]8989; SSE-NEXT:    movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8990; SSE-NEXT:    shufps {{.*#+}} xmm2 = xmm2[3,3],xmm3[3,3]8991; SSE-NEXT:    shufps {{.*#+}} xmm11 = xmm11[2,3],xmm2[0,2]8992; SSE-NEXT:    movaps %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8993; SSE-NEXT:    movaps 96(%rdx), %xmm08994; SSE-NEXT:    movaps 96(%rcx), %xmm18995; SSE-NEXT:    movaps %xmm1, %xmm28996; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]8997; SSE-NEXT:    movaps 96(%rdi), %xmm78998; SSE-NEXT:    movaps 96(%rsi), %xmm48999; SSE-NEXT:    movaps %xmm7, %xmm59000; SSE-NEXT:    unpcklps {{.*#+}} xmm5 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]9001; SSE-NEXT:    movaps %xmm5, %xmm39002; SSE-NEXT:    movaps %xmm5, %xmm89003; SSE-NEXT:    shufps {{.*#+}} xmm3 = xmm3[0,1],xmm2[2,0]9004; SSE-NEXT:    movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9005; SSE-NEXT:    movaps 96(%r10), %xmm29006; SSE-NEXT:    movaps 96(%rax), %xmm39007; SSE-NEXT:    movaps %xmm3, %xmm59008; SSE-NEXT:    movlhps {{.*#+}} xmm5 = xmm5[0],xmm2[0]9009; SSE-NEXT:    movaps 96(%r8), %xmm119010; SSE-NEXT:    movaps 96(%r9), %xmm69011; SSE-NEXT:    movaps %xmm11, %xmm109012; SSE-NEXT:    unpcklps {{.*#+}} xmm10 = xmm10[0],xmm6[0],xmm10[1],xmm6[1]9013; SSE-NEXT:    movaps %xmm10, %xmm99014; SSE-NEXT:    shufps {{.*#+}} xmm9 = xmm9[0,1],xmm5[2,0]9015; SSE-NEXT:    movaps %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9016; SSE-NEXT:    movaps %xmm1, %xmm59017; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[1,1],xmm0[1,1]9018; SSE-NEXT:    shufps {{.*#+}} xmm8 = xmm8[2,3],xmm5[2,0]9019; SSE-NEXT:    movaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9020; SSE-NEXT:    movaps %xmm3, %xmm59021; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[1,1],xmm2[1,1]9022; SSE-NEXT:    shufps {{.*#+}} xmm10 = xmm10[2,3],xmm5[2,0]9023; SSE-NEXT:    movaps %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9024; SSE-NEXT:    unpckhps {{.*#+}} xmm7 = xmm7[2],xmm4[2],xmm7[3],xmm4[3]9025; SSE-NEXT:    movaps %xmm0, %xmm49026; SSE-NEXT:    unpckhpd {{.*#+}} xmm4 = xmm4[1],xmm1[1]9027; SSE-NEXT:    movaps %xmm7, %xmm59028; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[0,1],xmm4[0,2]9029; SSE-NEXT:    movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9030; SSE-NEXT:    unpckhps {{.*#+}} xmm11 = xmm11[2],xmm6[2],xmm11[3],xmm6[3]9031; SSE-NEXT:    movaps %xmm2, %xmm49032; SSE-NEXT:    unpckhpd {{.*#+}} xmm4 = xmm4[1],xmm3[1]9033; SSE-NEXT:    movaps %xmm11, %xmm59034; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[0,1],xmm4[0,2]9035; SSE-NEXT:    movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9036; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[3,3],xmm1[3,3]9037; SSE-NEXT:    shufps {{.*#+}} xmm7 = xmm7[2,3],xmm0[0,2]9038; SSE-NEXT:    movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9039; SSE-NEXT:    shufps {{.*#+}} xmm2 = xmm2[3,3],xmm3[3,3]9040; SSE-NEXT:    shufps {{.*#+}} xmm11 = xmm11[2,3],xmm2[0,2]9041; SSE-NEXT:    movaps %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9042; SSE-NEXT:    movaps 112(%rdx), %xmm09043; SSE-NEXT:    movaps 112(%rcx), %xmm19044; SSE-NEXT:    movaps %xmm1, %xmm29045; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9046; SSE-NEXT:    movaps 112(%rdi), %xmm79047; SSE-NEXT:    movaps 112(%rsi), %xmm49048; SSE-NEXT:    movaps %xmm7, %xmm59049; SSE-NEXT:    unpcklps {{.*#+}} xmm5 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]9050; SSE-NEXT:    movaps %xmm5, %xmm39051; SSE-NEXT:    movaps %xmm5, %xmm89052; SSE-NEXT:    shufps {{.*#+}} xmm3 = xmm3[0,1],xmm2[2,0]9053; SSE-NEXT:    movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9054; SSE-NEXT:    movaps 112(%r10), %xmm29055; SSE-NEXT:    movaps 112(%rax), %xmm39056; SSE-NEXT:    movaps %xmm3, %xmm59057; SSE-NEXT:    movlhps {{.*#+}} xmm5 = xmm5[0],xmm2[0]9058; SSE-NEXT:    movaps 112(%r8), %xmm119059; SSE-NEXT:    movaps 112(%r9), %xmm69060; SSE-NEXT:    movaps %xmm11, %xmm109061; SSE-NEXT:    unpcklps {{.*#+}} xmm10 = xmm10[0],xmm6[0],xmm10[1],xmm6[1]9062; SSE-NEXT:    movaps %xmm10, %xmm99063; SSE-NEXT:    shufps {{.*#+}} xmm9 = xmm9[0,1],xmm5[2,0]9064; SSE-NEXT:    movaps %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9065; SSE-NEXT:    movaps %xmm1, %xmm59066; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[1,1],xmm0[1,1]9067; SSE-NEXT:    shufps {{.*#+}} xmm8 = xmm8[2,3],xmm5[2,0]9068; SSE-NEXT:    movaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9069; SSE-NEXT:    movaps %xmm3, %xmm59070; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[1,1],xmm2[1,1]9071; SSE-NEXT:    shufps {{.*#+}} xmm10 = xmm10[2,3],xmm5[2,0]9072; SSE-NEXT:    movaps %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9073; SSE-NEXT:    unpckhps {{.*#+}} xmm7 = xmm7[2],xmm4[2],xmm7[3],xmm4[3]9074; SSE-NEXT:    movaps %xmm0, %xmm49075; SSE-NEXT:    unpckhpd {{.*#+}} xmm4 = xmm4[1],xmm1[1]9076; SSE-NEXT:    movaps %xmm7, %xmm59077; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[0,1],xmm4[0,2]9078; SSE-NEXT:    movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9079; SSE-NEXT:    unpckhps {{.*#+}} xmm11 = xmm11[2],xmm6[2],xmm11[3],xmm6[3]9080; SSE-NEXT:    movaps %xmm2, %xmm49081; SSE-NEXT:    unpckhpd {{.*#+}} xmm4 = xmm4[1],xmm3[1]9082; SSE-NEXT:    movaps %xmm11, %xmm59083; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[0,1],xmm4[0,2]9084; SSE-NEXT:    movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9085; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[3,3],xmm1[3,3]9086; SSE-NEXT:    shufps {{.*#+}} xmm7 = xmm7[2,3],xmm0[0,2]9087; SSE-NEXT:    movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9088; SSE-NEXT:    shufps {{.*#+}} xmm2 = xmm2[3,3],xmm3[3,3]9089; SSE-NEXT:    shufps {{.*#+}} xmm11 = xmm11[2,3],xmm2[0,2]9090; SSE-NEXT:    movaps %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9091; SSE-NEXT:    movaps 128(%rdx), %xmm09092; SSE-NEXT:    movaps 128(%rcx), %xmm19093; SSE-NEXT:    movaps %xmm1, %xmm29094; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9095; SSE-NEXT:    movaps 128(%rdi), %xmm79096; SSE-NEXT:    movaps 128(%rsi), %xmm49097; SSE-NEXT:    movaps %xmm7, %xmm59098; SSE-NEXT:    unpcklps {{.*#+}} xmm5 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]9099; SSE-NEXT:    movaps %xmm5, %xmm39100; SSE-NEXT:    movaps %xmm5, %xmm89101; SSE-NEXT:    shufps {{.*#+}} xmm3 = xmm3[0,1],xmm2[2,0]9102; SSE-NEXT:    movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9103; SSE-NEXT:    movaps 128(%r10), %xmm29104; SSE-NEXT:    movaps 128(%rax), %xmm39105; SSE-NEXT:    movaps %xmm3, %xmm59106; SSE-NEXT:    movlhps {{.*#+}} xmm5 = xmm5[0],xmm2[0]9107; SSE-NEXT:    movaps 128(%r8), %xmm119108; SSE-NEXT:    movaps 128(%r9), %xmm69109; SSE-NEXT:    movaps %xmm11, %xmm109110; SSE-NEXT:    unpcklps {{.*#+}} xmm10 = xmm10[0],xmm6[0],xmm10[1],xmm6[1]9111; SSE-NEXT:    movaps %xmm10, %xmm99112; SSE-NEXT:    shufps {{.*#+}} xmm9 = xmm9[0,1],xmm5[2,0]9113; SSE-NEXT:    movaps %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9114; SSE-NEXT:    movaps %xmm1, %xmm59115; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[1,1],xmm0[1,1]9116; SSE-NEXT:    shufps {{.*#+}} xmm8 = xmm8[2,3],xmm5[2,0]9117; SSE-NEXT:    movaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9118; SSE-NEXT:    movaps %xmm3, %xmm59119; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[1,1],xmm2[1,1]9120; SSE-NEXT:    shufps {{.*#+}} xmm10 = xmm10[2,3],xmm5[2,0]9121; SSE-NEXT:    movaps %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9122; SSE-NEXT:    unpckhps {{.*#+}} xmm7 = xmm7[2],xmm4[2],xmm7[3],xmm4[3]9123; SSE-NEXT:    movaps %xmm0, %xmm49124; SSE-NEXT:    unpckhpd {{.*#+}} xmm4 = xmm4[1],xmm1[1]9125; SSE-NEXT:    movaps %xmm7, %xmm59126; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[0,1],xmm4[0,2]9127; SSE-NEXT:    movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9128; SSE-NEXT:    unpckhps {{.*#+}} xmm11 = xmm11[2],xmm6[2],xmm11[3],xmm6[3]9129; SSE-NEXT:    movaps %xmm2, %xmm49130; SSE-NEXT:    unpckhpd {{.*#+}} xmm4 = xmm4[1],xmm3[1]9131; SSE-NEXT:    movaps %xmm11, %xmm59132; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[0,1],xmm4[0,2]9133; SSE-NEXT:    movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9134; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[3,3],xmm1[3,3]9135; SSE-NEXT:    shufps {{.*#+}} xmm7 = xmm7[2,3],xmm0[0,2]9136; SSE-NEXT:    movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9137; SSE-NEXT:    shufps {{.*#+}} xmm2 = xmm2[3,3],xmm3[3,3]9138; SSE-NEXT:    shufps {{.*#+}} xmm11 = xmm11[2,3],xmm2[0,2]9139; SSE-NEXT:    movaps %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9140; SSE-NEXT:    movaps 144(%rdx), %xmm09141; SSE-NEXT:    movaps 144(%rcx), %xmm19142; SSE-NEXT:    movaps %xmm1, %xmm29143; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9144; SSE-NEXT:    movaps 144(%rdi), %xmm79145; SSE-NEXT:    movaps 144(%rsi), %xmm49146; SSE-NEXT:    movaps %xmm7, %xmm59147; SSE-NEXT:    unpcklps {{.*#+}} xmm5 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]9148; SSE-NEXT:    movaps %xmm5, %xmm39149; SSE-NEXT:    movaps %xmm5, %xmm89150; SSE-NEXT:    shufps {{.*#+}} xmm3 = xmm3[0,1],xmm2[2,0]9151; SSE-NEXT:    movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9152; SSE-NEXT:    movaps 144(%r10), %xmm29153; SSE-NEXT:    movaps 144(%rax), %xmm39154; SSE-NEXT:    movaps %xmm3, %xmm59155; SSE-NEXT:    movlhps {{.*#+}} xmm5 = xmm5[0],xmm2[0]9156; SSE-NEXT:    movaps 144(%r8), %xmm119157; SSE-NEXT:    movaps 144(%r9), %xmm69158; SSE-NEXT:    movaps %xmm11, %xmm109159; SSE-NEXT:    unpcklps {{.*#+}} xmm10 = xmm10[0],xmm6[0],xmm10[1],xmm6[1]9160; SSE-NEXT:    movaps %xmm10, %xmm99161; SSE-NEXT:    shufps {{.*#+}} xmm9 = xmm9[0,1],xmm5[2,0]9162; SSE-NEXT:    movaps %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9163; SSE-NEXT:    movaps %xmm1, %xmm59164; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[1,1],xmm0[1,1]9165; SSE-NEXT:    shufps {{.*#+}} xmm8 = xmm8[2,3],xmm5[2,0]9166; SSE-NEXT:    movaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9167; SSE-NEXT:    movaps %xmm3, %xmm59168; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[1,1],xmm2[1,1]9169; SSE-NEXT:    shufps {{.*#+}} xmm10 = xmm10[2,3],xmm5[2,0]9170; SSE-NEXT:    movaps %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9171; SSE-NEXT:    unpckhps {{.*#+}} xmm7 = xmm7[2],xmm4[2],xmm7[3],xmm4[3]9172; SSE-NEXT:    movaps %xmm0, %xmm49173; SSE-NEXT:    unpckhpd {{.*#+}} xmm4 = xmm4[1],xmm1[1]9174; SSE-NEXT:    movaps %xmm7, %xmm59175; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[0,1],xmm4[0,2]9176; SSE-NEXT:    movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9177; SSE-NEXT:    unpckhps {{.*#+}} xmm11 = xmm11[2],xmm6[2],xmm11[3],xmm6[3]9178; SSE-NEXT:    movaps %xmm2, %xmm49179; SSE-NEXT:    unpckhpd {{.*#+}} xmm4 = xmm4[1],xmm3[1]9180; SSE-NEXT:    movaps %xmm11, %xmm59181; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[0,1],xmm4[0,2]9182; SSE-NEXT:    movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9183; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[3,3],xmm1[3,3]9184; SSE-NEXT:    shufps {{.*#+}} xmm7 = xmm7[2,3],xmm0[0,2]9185; SSE-NEXT:    movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9186; SSE-NEXT:    shufps {{.*#+}} xmm2 = xmm2[3,3],xmm3[3,3]9187; SSE-NEXT:    shufps {{.*#+}} xmm11 = xmm11[2,3],xmm2[0,2]9188; SSE-NEXT:    movaps %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9189; SSE-NEXT:    movaps 160(%rdx), %xmm09190; SSE-NEXT:    movaps 160(%rcx), %xmm19191; SSE-NEXT:    movaps %xmm1, %xmm29192; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9193; SSE-NEXT:    movaps 160(%rdi), %xmm79194; SSE-NEXT:    movaps 160(%rsi), %xmm49195; SSE-NEXT:    movaps %xmm7, %xmm59196; SSE-NEXT:    unpcklps {{.*#+}} xmm5 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]9197; SSE-NEXT:    movaps %xmm5, %xmm39198; SSE-NEXT:    movaps %xmm5, %xmm89199; SSE-NEXT:    shufps {{.*#+}} xmm3 = xmm3[0,1],xmm2[2,0]9200; SSE-NEXT:    movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9201; SSE-NEXT:    movaps 160(%r10), %xmm29202; SSE-NEXT:    movaps 160(%rax), %xmm39203; SSE-NEXT:    movaps %xmm3, %xmm59204; SSE-NEXT:    movlhps {{.*#+}} xmm5 = xmm5[0],xmm2[0]9205; SSE-NEXT:    movaps 160(%r8), %xmm119206; SSE-NEXT:    movaps 160(%r9), %xmm69207; SSE-NEXT:    movaps %xmm11, %xmm109208; SSE-NEXT:    unpcklps {{.*#+}} xmm10 = xmm10[0],xmm6[0],xmm10[1],xmm6[1]9209; SSE-NEXT:    movaps %xmm10, %xmm99210; SSE-NEXT:    shufps {{.*#+}} xmm9 = xmm9[0,1],xmm5[2,0]9211; SSE-NEXT:    movaps %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9212; SSE-NEXT:    movaps %xmm1, %xmm59213; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[1,1],xmm0[1,1]9214; SSE-NEXT:    shufps {{.*#+}} xmm8 = xmm8[2,3],xmm5[2,0]9215; SSE-NEXT:    movaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9216; SSE-NEXT:    movaps %xmm3, %xmm59217; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[1,1],xmm2[1,1]9218; SSE-NEXT:    shufps {{.*#+}} xmm10 = xmm10[2,3],xmm5[2,0]9219; SSE-NEXT:    movaps %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9220; SSE-NEXT:    unpckhps {{.*#+}} xmm7 = xmm7[2],xmm4[2],xmm7[3],xmm4[3]9221; SSE-NEXT:    movaps %xmm0, %xmm49222; SSE-NEXT:    unpckhpd {{.*#+}} xmm4 = xmm4[1],xmm1[1]9223; SSE-NEXT:    movaps %xmm7, %xmm59224; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[0,1],xmm4[0,2]9225; SSE-NEXT:    movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9226; SSE-NEXT:    unpckhps {{.*#+}} xmm11 = xmm11[2],xmm6[2],xmm11[3],xmm6[3]9227; SSE-NEXT:    movaps %xmm2, %xmm49228; SSE-NEXT:    unpckhpd {{.*#+}} xmm4 = xmm4[1],xmm3[1]9229; SSE-NEXT:    movaps %xmm11, %xmm59230; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[0,1],xmm4[0,2]9231; SSE-NEXT:    movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9232; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[3,3],xmm1[3,3]9233; SSE-NEXT:    shufps {{.*#+}} xmm7 = xmm7[2,3],xmm0[0,2]9234; SSE-NEXT:    movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9235; SSE-NEXT:    shufps {{.*#+}} xmm2 = xmm2[3,3],xmm3[3,3]9236; SSE-NEXT:    shufps {{.*#+}} xmm11 = xmm11[2,3],xmm2[0,2]9237; SSE-NEXT:    movaps %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9238; SSE-NEXT:    movaps 176(%rdx), %xmm09239; SSE-NEXT:    movaps 176(%rcx), %xmm19240; SSE-NEXT:    movaps %xmm1, %xmm29241; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9242; SSE-NEXT:    movaps 176(%rdi), %xmm79243; SSE-NEXT:    movaps 176(%rsi), %xmm49244; SSE-NEXT:    movaps %xmm7, %xmm59245; SSE-NEXT:    unpcklps {{.*#+}} xmm5 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]9246; SSE-NEXT:    movaps %xmm5, %xmm39247; SSE-NEXT:    movaps %xmm5, %xmm89248; SSE-NEXT:    shufps {{.*#+}} xmm3 = xmm3[0,1],xmm2[2,0]9249; SSE-NEXT:    movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9250; SSE-NEXT:    movaps 176(%r10), %xmm29251; SSE-NEXT:    movaps 176(%rax), %xmm39252; SSE-NEXT:    movaps %xmm3, %xmm59253; SSE-NEXT:    movlhps {{.*#+}} xmm5 = xmm5[0],xmm2[0]9254; SSE-NEXT:    movaps 176(%r8), %xmm119255; SSE-NEXT:    movaps 176(%r9), %xmm69256; SSE-NEXT:    movaps %xmm11, %xmm109257; SSE-NEXT:    unpcklps {{.*#+}} xmm10 = xmm10[0],xmm6[0],xmm10[1],xmm6[1]9258; SSE-NEXT:    movaps %xmm10, %xmm99259; SSE-NEXT:    shufps {{.*#+}} xmm9 = xmm9[0,1],xmm5[2,0]9260; SSE-NEXT:    movaps %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9261; SSE-NEXT:    movaps %xmm1, %xmm59262; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[1,1],xmm0[1,1]9263; SSE-NEXT:    shufps {{.*#+}} xmm8 = xmm8[2,3],xmm5[2,0]9264; SSE-NEXT:    movaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9265; SSE-NEXT:    movaps %xmm3, %xmm59266; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[1,1],xmm2[1,1]9267; SSE-NEXT:    shufps {{.*#+}} xmm10 = xmm10[2,3],xmm5[2,0]9268; SSE-NEXT:    movaps %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9269; SSE-NEXT:    unpckhps {{.*#+}} xmm7 = xmm7[2],xmm4[2],xmm7[3],xmm4[3]9270; SSE-NEXT:    movaps %xmm0, %xmm49271; SSE-NEXT:    unpckhpd {{.*#+}} xmm4 = xmm4[1],xmm1[1]9272; SSE-NEXT:    movaps %xmm7, %xmm59273; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[0,1],xmm4[0,2]9274; SSE-NEXT:    movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9275; SSE-NEXT:    unpckhps {{.*#+}} xmm11 = xmm11[2],xmm6[2],xmm11[3],xmm6[3]9276; SSE-NEXT:    movaps %xmm2, %xmm49277; SSE-NEXT:    unpckhpd {{.*#+}} xmm4 = xmm4[1],xmm3[1]9278; SSE-NEXT:    movaps %xmm11, %xmm59279; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[0,1],xmm4[0,2]9280; SSE-NEXT:    movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9281; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[3,3],xmm1[3,3]9282; SSE-NEXT:    shufps {{.*#+}} xmm7 = xmm7[2,3],xmm0[0,2]9283; SSE-NEXT:    movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9284; SSE-NEXT:    shufps {{.*#+}} xmm2 = xmm2[3,3],xmm3[3,3]9285; SSE-NEXT:    shufps {{.*#+}} xmm11 = xmm11[2,3],xmm2[0,2]9286; SSE-NEXT:    movaps %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9287; SSE-NEXT:    movaps 192(%rdx), %xmm09288; SSE-NEXT:    movaps 192(%rcx), %xmm19289; SSE-NEXT:    movaps %xmm1, %xmm29290; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9291; SSE-NEXT:    movaps 192(%rdi), %xmm79292; SSE-NEXT:    movaps 192(%rsi), %xmm49293; SSE-NEXT:    movaps %xmm7, %xmm59294; SSE-NEXT:    unpcklps {{.*#+}} xmm5 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]9295; SSE-NEXT:    movaps %xmm5, %xmm39296; SSE-NEXT:    movaps %xmm5, %xmm89297; SSE-NEXT:    shufps {{.*#+}} xmm3 = xmm3[0,1],xmm2[2,0]9298; SSE-NEXT:    movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9299; SSE-NEXT:    movaps 192(%r10), %xmm29300; SSE-NEXT:    movaps 192(%rax), %xmm39301; SSE-NEXT:    movaps %xmm3, %xmm59302; SSE-NEXT:    movlhps {{.*#+}} xmm5 = xmm5[0],xmm2[0]9303; SSE-NEXT:    movaps 192(%r8), %xmm119304; SSE-NEXT:    movaps 192(%r9), %xmm69305; SSE-NEXT:    movaps %xmm11, %xmm109306; SSE-NEXT:    unpcklps {{.*#+}} xmm10 = xmm10[0],xmm6[0],xmm10[1],xmm6[1]9307; SSE-NEXT:    movaps %xmm10, %xmm99308; SSE-NEXT:    shufps {{.*#+}} xmm9 = xmm9[0,1],xmm5[2,0]9309; SSE-NEXT:    movaps %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9310; SSE-NEXT:    movaps %xmm1, %xmm59311; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[1,1],xmm0[1,1]9312; SSE-NEXT:    shufps {{.*#+}} xmm8 = xmm8[2,3],xmm5[2,0]9313; SSE-NEXT:    movaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9314; SSE-NEXT:    movaps %xmm3, %xmm59315; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[1,1],xmm2[1,1]9316; SSE-NEXT:    shufps {{.*#+}} xmm10 = xmm10[2,3],xmm5[2,0]9317; SSE-NEXT:    movaps %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9318; SSE-NEXT:    unpckhps {{.*#+}} xmm7 = xmm7[2],xmm4[2],xmm7[3],xmm4[3]9319; SSE-NEXT:    movaps %xmm0, %xmm49320; SSE-NEXT:    unpckhpd {{.*#+}} xmm4 = xmm4[1],xmm1[1]9321; SSE-NEXT:    movaps %xmm7, %xmm59322; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[0,1],xmm4[0,2]9323; SSE-NEXT:    movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9324; SSE-NEXT:    unpckhps {{.*#+}} xmm11 = xmm11[2],xmm6[2],xmm11[3],xmm6[3]9325; SSE-NEXT:    movaps %xmm2, %xmm49326; SSE-NEXT:    unpckhpd {{.*#+}} xmm4 = xmm4[1],xmm3[1]9327; SSE-NEXT:    movaps %xmm11, %xmm59328; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[0,1],xmm4[0,2]9329; SSE-NEXT:    movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9330; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[3,3],xmm1[3,3]9331; SSE-NEXT:    shufps {{.*#+}} xmm7 = xmm7[2,3],xmm0[0,2]9332; SSE-NEXT:    movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9333; SSE-NEXT:    shufps {{.*#+}} xmm2 = xmm2[3,3],xmm3[3,3]9334; SSE-NEXT:    shufps {{.*#+}} xmm11 = xmm11[2,3],xmm2[0,2]9335; SSE-NEXT:    movaps %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9336; SSE-NEXT:    movaps 208(%rdx), %xmm09337; SSE-NEXT:    movaps 208(%rcx), %xmm19338; SSE-NEXT:    movaps %xmm1, %xmm29339; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm0[0]9340; SSE-NEXT:    movaps 208(%rdi), %xmm79341; SSE-NEXT:    movaps 208(%rsi), %xmm49342; SSE-NEXT:    movaps %xmm7, %xmm59343; SSE-NEXT:    unpcklps {{.*#+}} xmm5 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]9344; SSE-NEXT:    movaps %xmm5, %xmm39345; SSE-NEXT:    movaps %xmm5, %xmm89346; SSE-NEXT:    shufps {{.*#+}} xmm3 = xmm3[0,1],xmm2[2,0]9347; SSE-NEXT:    movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9348; SSE-NEXT:    movaps 208(%r10), %xmm29349; SSE-NEXT:    movaps 208(%rax), %xmm39350; SSE-NEXT:    movaps %xmm3, %xmm59351; SSE-NEXT:    movlhps {{.*#+}} xmm5 = xmm5[0],xmm2[0]9352; SSE-NEXT:    movaps 208(%r8), %xmm119353; SSE-NEXT:    movaps 208(%r9), %xmm69354; SSE-NEXT:    movaps %xmm11, %xmm109355; SSE-NEXT:    unpcklps {{.*#+}} xmm10 = xmm10[0],xmm6[0],xmm10[1],xmm6[1]9356; SSE-NEXT:    movaps %xmm10, %xmm99357; SSE-NEXT:    shufps {{.*#+}} xmm9 = xmm9[0,1],xmm5[2,0]9358; SSE-NEXT:    movaps %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9359; SSE-NEXT:    movaps %xmm1, %xmm59360; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[1,1],xmm0[1,1]9361; SSE-NEXT:    shufps {{.*#+}} xmm8 = xmm8[2,3],xmm5[2,0]9362; SSE-NEXT:    movaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9363; SSE-NEXT:    movaps %xmm3, %xmm59364; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[1,1],xmm2[1,1]9365; SSE-NEXT:    shufps {{.*#+}} xmm10 = xmm10[2,3],xmm5[2,0]9366; SSE-NEXT:    movaps %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9367; SSE-NEXT:    unpckhps {{.*#+}} xmm7 = xmm7[2],xmm4[2],xmm7[3],xmm4[3]9368; SSE-NEXT:    movaps %xmm0, %xmm49369; SSE-NEXT:    unpckhpd {{.*#+}} xmm4 = xmm4[1],xmm1[1]9370; SSE-NEXT:    movaps %xmm7, %xmm59371; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[0,1],xmm4[0,2]9372; SSE-NEXT:    movaps %xmm5, (%rsp) # 16-byte Spill9373; SSE-NEXT:    unpckhps {{.*#+}} xmm11 = xmm11[2],xmm6[2],xmm11[3],xmm6[3]9374; SSE-NEXT:    movaps %xmm2, %xmm49375; SSE-NEXT:    unpckhpd {{.*#+}} xmm4 = xmm4[1],xmm3[1]9376; SSE-NEXT:    movaps %xmm11, %xmm59377; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[0,1],xmm4[0,2]9378; SSE-NEXT:    movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9379; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[3,3],xmm1[3,3]9380; SSE-NEXT:    shufps {{.*#+}} xmm7 = xmm7[2,3],xmm0[0,2]9381; SSE-NEXT:    movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9382; SSE-NEXT:    shufps {{.*#+}} xmm2 = xmm2[3,3],xmm3[3,3]9383; SSE-NEXT:    shufps {{.*#+}} xmm11 = xmm11[2,3],xmm2[0,2]9384; SSE-NEXT:    movaps %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9385; SSE-NEXT:    movaps 224(%rdx), %xmm19386; SSE-NEXT:    movaps 224(%rcx), %xmm09387; SSE-NEXT:    movaps %xmm0, %xmm29388; SSE-NEXT:    movlhps {{.*#+}} xmm2 = xmm2[0],xmm1[0]9389; SSE-NEXT:    movaps 224(%rdi), %xmm129390; SSE-NEXT:    movaps 224(%rsi), %xmm49391; SSE-NEXT:    movaps %xmm12, %xmm59392; SSE-NEXT:    unpcklps {{.*#+}} xmm5 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]9393; SSE-NEXT:    movaps %xmm5, %xmm39394; SSE-NEXT:    movaps %xmm5, %xmm79395; SSE-NEXT:    shufps {{.*#+}} xmm3 = xmm3[0,1],xmm2[2,0]9396; SSE-NEXT:    movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9397; SSE-NEXT:    movaps 224(%r10), %xmm29398; SSE-NEXT:    movaps 224(%rax), %xmm39399; SSE-NEXT:    movaps %xmm3, %xmm59400; SSE-NEXT:    movlhps {{.*#+}} xmm5 = xmm5[0],xmm2[0]9401; SSE-NEXT:    movaps 224(%r8), %xmm159402; SSE-NEXT:    movaps 224(%r9), %xmm69403; SSE-NEXT:    movaps %xmm15, %xmm139404; SSE-NEXT:    unpcklps {{.*#+}} xmm13 = xmm13[0],xmm6[0],xmm13[1],xmm6[1]9405; SSE-NEXT:    movaps %xmm13, %xmm89406; SSE-NEXT:    shufps {{.*#+}} xmm8 = xmm8[0,1],xmm5[2,0]9407; SSE-NEXT:    movaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9408; SSE-NEXT:    movaps %xmm0, %xmm59409; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[1,1],xmm1[1,1]9410; SSE-NEXT:    shufps {{.*#+}} xmm7 = xmm7[2,3],xmm5[2,0]9411; SSE-NEXT:    movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9412; SSE-NEXT:    movaps %xmm3, %xmm59413; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[1,1],xmm2[1,1]9414; SSE-NEXT:    shufps {{.*#+}} xmm13 = xmm13[2,3],xmm5[2,0]9415; SSE-NEXT:    unpckhps {{.*#+}} xmm12 = xmm12[2],xmm4[2],xmm12[3],xmm4[3]9416; SSE-NEXT:    movaps %xmm1, %xmm49417; SSE-NEXT:    unpckhpd {{.*#+}} xmm4 = xmm4[1],xmm0[1]9418; SSE-NEXT:    movaps %xmm12, %xmm59419; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[0,1],xmm4[0,2]9420; SSE-NEXT:    movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9421; SSE-NEXT:    unpckhps {{.*#+}} xmm15 = xmm15[2],xmm6[2],xmm15[3],xmm6[3]9422; SSE-NEXT:    movaps %xmm2, %xmm49423; SSE-NEXT:    unpckhpd {{.*#+}} xmm4 = xmm4[1],xmm3[1]9424; SSE-NEXT:    movaps %xmm15, %xmm59425; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[0,1],xmm4[0,2]9426; SSE-NEXT:    movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9427; SSE-NEXT:    shufps {{.*#+}} xmm1 = xmm1[3,3],xmm0[3,3]9428; SSE-NEXT:    shufps {{.*#+}} xmm12 = xmm12[2,3],xmm1[0,2]9429; SSE-NEXT:    shufps {{.*#+}} xmm2 = xmm2[3,3],xmm3[3,3]9430; SSE-NEXT:    shufps {{.*#+}} xmm15 = xmm15[2,3],xmm2[0,2]9431; SSE-NEXT:    movaps 240(%rdx), %xmm29432; SSE-NEXT:    movaps 240(%rcx), %xmm89433; SSE-NEXT:    movaps %xmm8, %xmm09434; SSE-NEXT:    movlhps {{.*#+}} xmm0 = xmm0[0],xmm2[0]9435; SSE-NEXT:    movaps 240(%rdi), %xmm19436; SSE-NEXT:    movaps 240(%rsi), %xmm119437; SSE-NEXT:    movaps %xmm1, %xmm49438; SSE-NEXT:    unpcklps {{.*#+}} xmm4 = xmm4[0],xmm11[0],xmm4[1],xmm11[1]9439; SSE-NEXT:    movaps %xmm4, %xmm149440; SSE-NEXT:    shufps {{.*#+}} xmm14 = xmm14[0,1],xmm0[2,0]9441; SSE-NEXT:    movaps 240(%r10), %xmm09442; SSE-NEXT:    movaps 240(%rax), %xmm79443; SSE-NEXT:    movaps %xmm7, %xmm39444; SSE-NEXT:    movlhps {{.*#+}} xmm3 = xmm3[0],xmm0[0]9445; SSE-NEXT:    movaps 240(%r8), %xmm59446; SSE-NEXT:    movaps 240(%r9), %xmm99447; SSE-NEXT:    movaps %xmm5, %xmm69448; SSE-NEXT:    unpcklps {{.*#+}} xmm6 = xmm6[0],xmm9[0],xmm6[1],xmm9[1]9449; SSE-NEXT:    movaps %xmm6, %xmm109450; SSE-NEXT:    shufps {{.*#+}} xmm10 = xmm10[0,1],xmm3[2,0]9451; SSE-NEXT:    movaps %xmm8, %xmm39452; SSE-NEXT:    shufps {{.*#+}} xmm3 = xmm3[1,1],xmm2[1,1]9453; SSE-NEXT:    shufps {{.*#+}} xmm4 = xmm4[2,3],xmm3[2,0]9454; SSE-NEXT:    movaps %xmm7, %xmm39455; SSE-NEXT:    shufps {{.*#+}} xmm3 = xmm3[1,1],xmm0[1,1]9456; SSE-NEXT:    shufps {{.*#+}} xmm6 = xmm6[2,3],xmm3[2,0]9457; SSE-NEXT:    unpckhps {{.*#+}} xmm1 = xmm1[2],xmm11[2],xmm1[3],xmm11[3]9458; SSE-NEXT:    movaps %xmm2, %xmm39459; SSE-NEXT:    unpckhpd {{.*#+}} xmm3 = xmm3[1],xmm8[1]9460; SSE-NEXT:    movaps %xmm1, %xmm119461; SSE-NEXT:    shufps {{.*#+}} xmm11 = xmm11[0,1],xmm3[0,2]9462; SSE-NEXT:    unpckhps {{.*#+}} xmm5 = xmm5[2],xmm9[2],xmm5[3],xmm9[3]9463; SSE-NEXT:    movaps %xmm0, %xmm39464; SSE-NEXT:    unpckhpd {{.*#+}} xmm3 = xmm3[1],xmm7[1]9465; SSE-NEXT:    movaps %xmm5, %xmm99466; SSE-NEXT:    shufps {{.*#+}} xmm9 = xmm9[0,1],xmm3[0,2]9467; SSE-NEXT:    shufps {{.*#+}} xmm2 = xmm2[3,3],xmm8[3,3]9468; SSE-NEXT:    shufps {{.*#+}} xmm1 = xmm1[2,3],xmm2[0,2]9469; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[3,3],xmm7[3,3]9470; SSE-NEXT:    shufps {{.*#+}} xmm5 = xmm5[2,3],xmm0[0,2]9471; SSE-NEXT:    movq {{[0-9]+}}(%rsp), %rax9472; SSE-NEXT:    movaps %xmm5, 2032(%rax)9473; SSE-NEXT:    movaps %xmm1, 2016(%rax)9474; SSE-NEXT:    movaps %xmm9, 2000(%rax)9475; SSE-NEXT:    movaps %xmm11, 1984(%rax)9476; SSE-NEXT:    movaps %xmm6, 1968(%rax)9477; SSE-NEXT:    movaps %xmm4, 1952(%rax)9478; SSE-NEXT:    movaps %xmm10, 1936(%rax)9479; SSE-NEXT:    movaps %xmm14, 1920(%rax)9480; SSE-NEXT:    movaps %xmm15, 1904(%rax)9481; SSE-NEXT:    movaps %xmm12, 1888(%rax)9482; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9483; SSE-NEXT:    movaps %xmm0, 1872(%rax)9484; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9485; SSE-NEXT:    movaps %xmm0, 1856(%rax)9486; SSE-NEXT:    movaps %xmm13, 1840(%rax)9487; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9488; SSE-NEXT:    movaps %xmm0, 1824(%rax)9489; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9490; SSE-NEXT:    movaps %xmm0, 1808(%rax)9491; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9492; SSE-NEXT:    movaps %xmm0, 1792(%rax)9493; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9494; SSE-NEXT:    movaps %xmm0, 1776(%rax)9495; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9496; SSE-NEXT:    movaps %xmm0, 1760(%rax)9497; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9498; SSE-NEXT:    movaps %xmm0, 1744(%rax)9499; SSE-NEXT:    movaps (%rsp), %xmm0 # 16-byte Reload9500; SSE-NEXT:    movaps %xmm0, 1728(%rax)9501; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9502; SSE-NEXT:    movaps %xmm0, 1712(%rax)9503; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9504; SSE-NEXT:    movaps %xmm0, 1696(%rax)9505; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9506; SSE-NEXT:    movaps %xmm0, 1680(%rax)9507; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9508; SSE-NEXT:    movaps %xmm0, 1664(%rax)9509; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9510; SSE-NEXT:    movaps %xmm0, 1648(%rax)9511; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9512; SSE-NEXT:    movaps %xmm0, 1632(%rax)9513; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9514; SSE-NEXT:    movaps %xmm0, 1616(%rax)9515; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9516; SSE-NEXT:    movaps %xmm0, 1600(%rax)9517; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9518; SSE-NEXT:    movaps %xmm0, 1584(%rax)9519; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9520; SSE-NEXT:    movaps %xmm0, 1568(%rax)9521; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9522; SSE-NEXT:    movaps %xmm0, 1552(%rax)9523; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9524; SSE-NEXT:    movaps %xmm0, 1536(%rax)9525; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9526; SSE-NEXT:    movaps %xmm0, 1520(%rax)9527; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9528; SSE-NEXT:    movaps %xmm0, 1504(%rax)9529; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9530; SSE-NEXT:    movaps %xmm0, 1488(%rax)9531; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9532; SSE-NEXT:    movaps %xmm0, 1472(%rax)9533; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9534; SSE-NEXT:    movaps %xmm0, 1456(%rax)9535; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9536; SSE-NEXT:    movaps %xmm0, 1440(%rax)9537; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9538; SSE-NEXT:    movaps %xmm0, 1424(%rax)9539; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9540; SSE-NEXT:    movaps %xmm0, 1408(%rax)9541; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9542; SSE-NEXT:    movaps %xmm0, 1392(%rax)9543; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9544; SSE-NEXT:    movaps %xmm0, 1376(%rax)9545; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9546; SSE-NEXT:    movaps %xmm0, 1360(%rax)9547; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9548; SSE-NEXT:    movaps %xmm0, 1344(%rax)9549; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9550; SSE-NEXT:    movaps %xmm0, 1328(%rax)9551; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9552; SSE-NEXT:    movaps %xmm0, 1312(%rax)9553; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9554; SSE-NEXT:    movaps %xmm0, 1296(%rax)9555; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9556; SSE-NEXT:    movaps %xmm0, 1280(%rax)9557; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9558; SSE-NEXT:    movaps %xmm0, 1264(%rax)9559; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9560; SSE-NEXT:    movaps %xmm0, 1248(%rax)9561; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9562; SSE-NEXT:    movaps %xmm0, 1232(%rax)9563; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9564; SSE-NEXT:    movaps %xmm0, 1216(%rax)9565; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9566; SSE-NEXT:    movaps %xmm0, 1200(%rax)9567; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9568; SSE-NEXT:    movaps %xmm0, 1184(%rax)9569; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9570; SSE-NEXT:    movaps %xmm0, 1168(%rax)9571; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9572; SSE-NEXT:    movaps %xmm0, 1152(%rax)9573; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9574; SSE-NEXT:    movaps %xmm0, 1136(%rax)9575; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9576; SSE-NEXT:    movaps %xmm0, 1120(%rax)9577; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9578; SSE-NEXT:    movaps %xmm0, 1104(%rax)9579; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9580; SSE-NEXT:    movaps %xmm0, 1088(%rax)9581; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9582; SSE-NEXT:    movaps %xmm0, 1072(%rax)9583; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9584; SSE-NEXT:    movaps %xmm0, 1056(%rax)9585; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9586; SSE-NEXT:    movaps %xmm0, 1040(%rax)9587; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9588; SSE-NEXT:    movaps %xmm0, 1024(%rax)9589; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9590; SSE-NEXT:    movaps %xmm0, 1008(%rax)9591; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9592; SSE-NEXT:    movaps %xmm0, 992(%rax)9593; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9594; SSE-NEXT:    movaps %xmm0, 976(%rax)9595; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9596; SSE-NEXT:    movaps %xmm0, 960(%rax)9597; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9598; SSE-NEXT:    movaps %xmm0, 944(%rax)9599; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9600; SSE-NEXT:    movaps %xmm0, 928(%rax)9601; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9602; SSE-NEXT:    movaps %xmm0, 912(%rax)9603; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9604; SSE-NEXT:    movaps %xmm0, 896(%rax)9605; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9606; SSE-NEXT:    movaps %xmm0, 880(%rax)9607; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9608; SSE-NEXT:    movaps %xmm0, 864(%rax)9609; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9610; SSE-NEXT:    movaps %xmm0, 848(%rax)9611; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9612; SSE-NEXT:    movaps %xmm0, 832(%rax)9613; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9614; SSE-NEXT:    movaps %xmm0, 816(%rax)9615; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9616; SSE-NEXT:    movaps %xmm0, 800(%rax)9617; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9618; SSE-NEXT:    movaps %xmm0, 784(%rax)9619; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9620; SSE-NEXT:    movaps %xmm0, 768(%rax)9621; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9622; SSE-NEXT:    movaps %xmm0, 752(%rax)9623; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9624; SSE-NEXT:    movaps %xmm0, 736(%rax)9625; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9626; SSE-NEXT:    movaps %xmm0, 720(%rax)9627; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9628; SSE-NEXT:    movaps %xmm0, 704(%rax)9629; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9630; SSE-NEXT:    movaps %xmm0, 688(%rax)9631; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9632; SSE-NEXT:    movaps %xmm0, 672(%rax)9633; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9634; SSE-NEXT:    movaps %xmm0, 656(%rax)9635; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9636; SSE-NEXT:    movaps %xmm0, 640(%rax)9637; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9638; SSE-NEXT:    movaps %xmm0, 624(%rax)9639; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9640; SSE-NEXT:    movaps %xmm0, 608(%rax)9641; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9642; SSE-NEXT:    movaps %xmm0, 592(%rax)9643; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9644; SSE-NEXT:    movaps %xmm0, 576(%rax)9645; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9646; SSE-NEXT:    movaps %xmm0, 560(%rax)9647; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9648; SSE-NEXT:    movaps %xmm0, 544(%rax)9649; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9650; SSE-NEXT:    movaps %xmm0, 528(%rax)9651; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9652; SSE-NEXT:    movaps %xmm0, 512(%rax)9653; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9654; SSE-NEXT:    movaps %xmm0, 496(%rax)9655; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9656; SSE-NEXT:    movaps %xmm0, 480(%rax)9657; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9658; SSE-NEXT:    movaps %xmm0, 464(%rax)9659; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9660; SSE-NEXT:    movaps %xmm0, 448(%rax)9661; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9662; SSE-NEXT:    movaps %xmm0, 432(%rax)9663; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9664; SSE-NEXT:    movaps %xmm0, 416(%rax)9665; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9666; SSE-NEXT:    movaps %xmm0, 400(%rax)9667; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9668; SSE-NEXT:    movaps %xmm0, 384(%rax)9669; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9670; SSE-NEXT:    movaps %xmm0, 368(%rax)9671; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9672; SSE-NEXT:    movaps %xmm0, 352(%rax)9673; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9674; SSE-NEXT:    movaps %xmm0, 336(%rax)9675; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9676; SSE-NEXT:    movaps %xmm0, 320(%rax)9677; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9678; SSE-NEXT:    movaps %xmm0, 304(%rax)9679; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9680; SSE-NEXT:    movaps %xmm0, 288(%rax)9681; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9682; SSE-NEXT:    movaps %xmm0, 272(%rax)9683; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9684; SSE-NEXT:    movaps %xmm0, 256(%rax)9685; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9686; SSE-NEXT:    movaps %xmm0, 240(%rax)9687; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9688; SSE-NEXT:    movaps %xmm0, 224(%rax)9689; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9690; SSE-NEXT:    movaps %xmm0, 208(%rax)9691; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9692; SSE-NEXT:    movaps %xmm0, 192(%rax)9693; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9694; SSE-NEXT:    movaps %xmm0, 176(%rax)9695; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9696; SSE-NEXT:    movaps %xmm0, 160(%rax)9697; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9698; SSE-NEXT:    movaps %xmm0, 144(%rax)9699; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9700; SSE-NEXT:    movaps %xmm0, 128(%rax)9701; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9702; SSE-NEXT:    movaps %xmm0, 112(%rax)9703; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9704; SSE-NEXT:    movaps %xmm0, 96(%rax)9705; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9706; SSE-NEXT:    movaps %xmm0, 80(%rax)9707; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9708; SSE-NEXT:    movaps %xmm0, 64(%rax)9709; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9710; SSE-NEXT:    movaps %xmm0, 48(%rax)9711; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9712; SSE-NEXT:    movaps %xmm0, 32(%rax)9713; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9714; SSE-NEXT:    movaps %xmm0, 16(%rax)9715; SSE-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload9716; SSE-NEXT:    movaps %xmm0, (%rax)9717; SSE-NEXT:    addq $1752, %rsp # imm = 0x6D89718; SSE-NEXT:    retq9719;9720; AVX-LABEL: store_i32_stride8_vf64:9721; AVX:       # %bb.0:9722; AVX-NEXT:    subq $1704, %rsp # imm = 0x6A89723; AVX-NEXT:    movq {{[0-9]+}}(%rsp), %rax9724; AVX-NEXT:    movq {{[0-9]+}}(%rsp), %r109725; AVX-NEXT:    vmovaps (%rdi), %ymm09726; AVX-NEXT:    vmovaps (%rsi), %ymm19727; AVX-NEXT:    vmovaps (%rdx), %ymm29728; AVX-NEXT:    vmovaps (%rcx), %ymm39729; AVX-NEXT:    vmovaps (%rcx), %xmm59730; AVX-NEXT:    vmovaps (%rdx), %xmm79731; AVX-NEXT:    vunpcklps {{.*#+}} xmm4 = xmm7[0],xmm5[0],xmm7[1],xmm5[1]9732; AVX-NEXT:    vmovaps (%rsi), %xmm109733; AVX-NEXT:    vmovaps (%rdi), %xmm129734; AVX-NEXT:    vmovaps (%r9), %xmm119735; AVX-NEXT:    vmovaps (%r8), %xmm139736; AVX-NEXT:    vunpcklps {{.*#+}} xmm6 = xmm13[0],xmm11[0],xmm13[1],xmm11[1]9737; AVX-NEXT:    vinsertf128 $1, %xmm6, %ymm0, %ymm69738; AVX-NEXT:    vmovaps (%rax), %xmm89739; AVX-NEXT:    vmovaps (%r10), %xmm99740; AVX-NEXT:    vunpcklps {{.*#+}} xmm14 = xmm9[0],xmm8[0],xmm9[1],xmm8[1]9741; AVX-NEXT:    vshufps {{.*#+}} xmm15 = xmm14[0,1,0,1]9742; AVX-NEXT:    vinsertf128 $1, %xmm15, %ymm0, %ymm159743; AVX-NEXT:    vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3,4,5],ymm15[6,7]9744; AVX-NEXT:    vunpcklps {{.*#+}} xmm15 = xmm12[0],xmm10[0],xmm12[1],xmm10[1]9745; AVX-NEXT:    vmovlhps {{.*#+}} xmm15 = xmm15[0],xmm4[0]9746; AVX-NEXT:    vblendps {{.*#+}} ymm6 = ymm15[0,1,2,3],ymm6[4,5,6,7]9747; AVX-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9748; AVX-NEXT:    vshufps {{.*#+}} xmm6 = xmm12[1,1,1,1]9749; AVX-NEXT:    vblendps {{.*#+}} xmm6 = xmm6[0],xmm10[1],xmm6[2,3]9750; AVX-NEXT:    vblendps {{.*#+}} xmm6 = xmm6[0,1],xmm4[2,3]9751; AVX-NEXT:    vinsertf128 $1, %xmm14, %ymm0, %ymm49752; AVX-NEXT:    vshufps {{.*#+}} xmm14 = xmm13[1,1,1,1]9753; AVX-NEXT:    vblendps {{.*#+}} xmm14 = xmm14[0],xmm11[1],xmm14[2,3]9754; AVX-NEXT:    vinsertf128 $1, %xmm14, %ymm0, %ymm149755; AVX-NEXT:    vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3,4,5],ymm4[6,7]9756; AVX-NEXT:    vmovaps (%r8), %ymm49757; AVX-NEXT:    vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm14[4,5,6,7]9758; AVX-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9759; AVX-NEXT:    vmovaps (%r9), %ymm69760; AVX-NEXT:    vunpckhps {{.*#+}} xmm12 = xmm12[2],xmm10[2],xmm12[3],xmm10[3]9761; AVX-NEXT:    vunpckhps {{.*#+}} xmm11 = xmm13[2],xmm11[2],xmm13[3],xmm11[3]9762; AVX-NEXT:    vinsertf128 $1, %xmm11, %ymm0, %ymm109763; AVX-NEXT:    vshufps {{.*#+}} xmm13 = xmm8[2,2,2,2]9764; AVX-NEXT:    vblendps {{.*#+}} xmm13 = xmm9[0,1,2],xmm13[3]9765; AVX-NEXT:    vinsertf128 $1, %xmm13, %ymm0, %ymm139766; AVX-NEXT:    vblendps {{.*#+}} ymm10 = ymm10[0,1,2,3,4,5],ymm13[6,7]9767; AVX-NEXT:    vshufps {{.*#+}} xmm13 = xmm5[2,2,2,2]9768; AVX-NEXT:    vblendps {{.*#+}} xmm13 = xmm7[0,1,2],xmm13[3]9769; AVX-NEXT:    vblendps {{.*#+}} xmm13 = xmm12[0,1],xmm13[2,3]9770; AVX-NEXT:    vblendps {{.*#+}} ymm10 = ymm13[0,1,2,3],ymm10[4,5,6,7]9771; AVX-NEXT:    vmovups %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9772; AVX-NEXT:    vmovaps (%r10), %ymm109773; AVX-NEXT:    vunpckhps {{.*#+}} xmm7 = xmm7[2],xmm5[2],xmm7[3],xmm5[3]9774; AVX-NEXT:    vmovaps (%rax), %ymm59775; AVX-NEXT:    vunpckhpd {{.*#+}} xmm7 = xmm12[1],xmm7[1]9776; AVX-NEXT:    vunpckhps {{.*#+}} xmm8 = xmm9[2],xmm8[2],xmm9[3],xmm8[3]9777; AVX-NEXT:    vinsertf128 $1, %xmm8, %ymm0, %ymm89778; AVX-NEXT:    vshufps {{.*#+}} xmm9 = xmm11[2,3,2,3]9779; AVX-NEXT:    vinsertf128 $1, %xmm9, %ymm0, %ymm99780; AVX-NEXT:    vblendps {{.*#+}} ymm8 = ymm9[0,1,2,3,4,5],ymm8[6,7]9781; AVX-NEXT:    vblendps {{.*#+}} ymm7 = ymm7[0,1,2,3],ymm8[4,5,6,7]9782; AVX-NEXT:    vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9783; AVX-NEXT:    vunpcklpd {{.*#+}} ymm7 = ymm5[0],ymm10[0],ymm5[2],ymm10[2]9784; AVX-NEXT:    vunpcklps {{.*#+}} ymm8 = ymm4[0],ymm6[0],ymm4[1],ymm6[1],ymm4[4],ymm6[4],ymm4[5],ymm6[5]9785; AVX-NEXT:    vshufps {{.*#+}} ymm7 = ymm8[0,1],ymm7[2,0],ymm8[4,5],ymm7[6,4]9786; AVX-NEXT:    vunpcklpd {{.*#+}} ymm8 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]9787; AVX-NEXT:    vextractf128 $1, %ymm8, %xmm89788; AVX-NEXT:    vunpcklps {{.*#+}} ymm9 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]9789; AVX-NEXT:    vextractf128 $1, %ymm9, %xmm99790; AVX-NEXT:    vshufps {{.*#+}} xmm8 = xmm9[0,1],xmm8[2,0]9791; AVX-NEXT:    vblendps {{.*#+}} ymm7 = ymm8[0,1,2,3],ymm7[4,5,6,7]9792; AVX-NEXT:    vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9793; AVX-NEXT:    vunpcklps {{.*#+}} ymm7 = ymm10[0],ymm5[0],ymm10[1],ymm5[1],ymm10[4],ymm5[4],ymm10[5],ymm5[5]9794; AVX-NEXT:    vshufps {{.*#+}} ymm8 = ymm6[1,0],ymm4[1,0],ymm6[5,4],ymm4[5,4]9795; AVX-NEXT:    vshufps {{.*#+}} ymm7 = ymm8[2,0],ymm7[2,3],ymm8[6,4],ymm7[6,7]9796; AVX-NEXT:    vunpcklps {{.*#+}} ymm8 = ymm2[0],ymm3[0],ymm2[1],ymm3[1],ymm2[4],ymm3[4],ymm2[5],ymm3[5]9797; AVX-NEXT:    vextractf128 $1, %ymm8, %xmm89798; AVX-NEXT:    vshufps {{.*#+}} ymm9 = ymm1[1,0],ymm0[1,0],ymm1[5,4],ymm0[5,4]9799; AVX-NEXT:    vextractf128 $1, %ymm9, %xmm99800; AVX-NEXT:    vshufps {{.*#+}} xmm8 = xmm9[2,0],xmm8[2,3]9801; AVX-NEXT:    vblendps {{.*#+}} ymm7 = ymm8[0,1,2,3],ymm7[4,5,6,7]9802; AVX-NEXT:    vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9803; AVX-NEXT:    vunpckhpd {{.*#+}} ymm7 = ymm5[1],ymm10[1],ymm5[3],ymm10[3]9804; AVX-NEXT:    vunpckhps {{.*#+}} ymm8 = ymm4[2],ymm6[2],ymm4[3],ymm6[3],ymm4[6],ymm6[6],ymm4[7],ymm6[7]9805; AVX-NEXT:    vshufps {{.*#+}} ymm7 = ymm8[0,1],ymm7[2,0],ymm8[4,5],ymm7[6,4]9806; AVX-NEXT:    vunpckhpd {{.*#+}} ymm8 = ymm3[1],ymm2[1],ymm3[3],ymm2[3]9807; AVX-NEXT:    vextractf128 $1, %ymm8, %xmm89808; AVX-NEXT:    vunpckhps {{.*#+}} ymm9 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7]9809; AVX-NEXT:    vextractf128 $1, %ymm9, %xmm99810; AVX-NEXT:    vshufps {{.*#+}} xmm8 = xmm9[0,1],xmm8[2,0]9811; AVX-NEXT:    vblendps {{.*#+}} ymm7 = ymm8[0,1,2,3],ymm7[4,5,6,7]9812; AVX-NEXT:    vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9813; AVX-NEXT:    vmovaps 32(%rcx), %xmm79814; AVX-NEXT:    vunpckhps {{.*#+}} ymm8 = ymm10[2],ymm5[2],ymm10[3],ymm5[3],ymm10[6],ymm5[6],ymm10[7],ymm5[7]9815; AVX-NEXT:    vmovaps 32(%rax), %xmm59816; AVX-NEXT:    vshufps {{.*#+}} ymm6 = ymm6[3,0],ymm4[3,0],ymm6[7,4],ymm4[7,4]9817; AVX-NEXT:    vmovaps 32(%r10), %xmm49818; AVX-NEXT:    vshufps {{.*#+}} ymm6 = ymm6[2,0],ymm8[2,3],ymm6[6,4],ymm8[6,7]9819; AVX-NEXT:    vunpckhps {{.*#+}} ymm2 = ymm2[2],ymm3[2],ymm2[3],ymm3[3],ymm2[6],ymm3[6],ymm2[7],ymm3[7]9820; AVX-NEXT:    vextractf128 $1, %ymm2, %xmm29821; AVX-NEXT:    vshufps {{.*#+}} ymm0 = ymm1[3,0],ymm0[3,0],ymm1[7,4],ymm0[7,4]9822; AVX-NEXT:    vextractf128 $1, %ymm0, %xmm09823; AVX-NEXT:    vshufps {{.*#+}} xmm0 = xmm0[2,0],xmm2[2,3]9824; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm6[4,5,6,7]9825; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9826; AVX-NEXT:    vmovaps 32(%r9), %xmm19827; AVX-NEXT:    vmovaps 32(%r8), %xmm29828; AVX-NEXT:    vunpcklps {{.*#+}} xmm0 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]9829; AVX-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm09830; AVX-NEXT:    vunpcklps {{.*#+}} xmm3 = xmm4[0],xmm5[0],xmm4[1],xmm5[1]9831; AVX-NEXT:    vshufps {{.*#+}} xmm6 = xmm3[0,1,0,1]9832; AVX-NEXT:    vinsertf128 $1, %xmm6, %ymm0, %ymm69833; AVX-NEXT:    vblendps {{.*#+}} ymm6 = ymm0[0,1,2,3,4,5],ymm6[6,7]9834; AVX-NEXT:    vmovaps 32(%rdx), %xmm09835; AVX-NEXT:    vunpcklps {{.*#+}} xmm8 = xmm0[0],xmm7[0],xmm0[1],xmm7[1]9836; AVX-NEXT:    vmovaps 32(%rsi), %xmm99837; AVX-NEXT:    vmovaps 32(%rdi), %xmm109838; AVX-NEXT:    vunpcklps {{.*#+}} xmm11 = xmm10[0],xmm9[0],xmm10[1],xmm9[1]9839; AVX-NEXT:    vmovlhps {{.*#+}} xmm11 = xmm11[0],xmm8[0]9840; AVX-NEXT:    vblendps {{.*#+}} ymm6 = ymm11[0,1,2,3],ymm6[4,5,6,7]9841; AVX-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9842; AVX-NEXT:    vshufps {{.*#+}} xmm6 = xmm10[1,1,1,1]9843; AVX-NEXT:    vblendps {{.*#+}} xmm6 = xmm6[0],xmm9[1],xmm6[2,3]9844; AVX-NEXT:    vblendps {{.*#+}} xmm6 = xmm6[0,1],xmm8[2,3]9845; AVX-NEXT:    vinsertf128 $1, %xmm3, %ymm0, %ymm39846; AVX-NEXT:    vshufps {{.*#+}} xmm8 = xmm2[1,1,1,1]9847; AVX-NEXT:    vblendps {{.*#+}} xmm8 = xmm8[0],xmm1[1],xmm8[2,3]9848; AVX-NEXT:    vinsertf128 $1, %xmm8, %ymm0, %ymm89849; AVX-NEXT:    vblendps {{.*#+}} ymm3 = ymm8[0,1,2,3,4,5],ymm3[6,7]9850; AVX-NEXT:    vblendps {{.*#+}} ymm3 = ymm6[0,1,2,3],ymm3[4,5,6,7]9851; AVX-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9852; AVX-NEXT:    vunpckhps {{.*#+}} xmm3 = xmm10[2],xmm9[2],xmm10[3],xmm9[3]9853; AVX-NEXT:    vunpckhps {{.*#+}} xmm2 = xmm2[2],xmm1[2],xmm2[3],xmm1[3]9854; AVX-NEXT:    vshufps {{.*#+}} xmm1 = xmm5[2,2,2,2]9855; AVX-NEXT:    vblendps {{.*#+}} xmm1 = xmm4[0,1,2],xmm1[3]9856; AVX-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm19857; AVX-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm69858; AVX-NEXT:    vblendps {{.*#+}} ymm1 = ymm6[0,1,2,3,4,5],ymm1[6,7]9859; AVX-NEXT:    vshufps {{.*#+}} xmm6 = xmm7[2,2,2,2]9860; AVX-NEXT:    vblendps {{.*#+}} xmm6 = xmm0[0,1,2],xmm6[3]9861; AVX-NEXT:    vblendps {{.*#+}} xmm6 = xmm3[0,1],xmm6[2,3]9862; AVX-NEXT:    vblendps {{.*#+}} ymm1 = ymm6[0,1,2,3],ymm1[4,5,6,7]9863; AVX-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9864; AVX-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm0[2],xmm7[2],xmm0[3],xmm7[3]9865; AVX-NEXT:    vunpckhpd {{.*#+}} xmm3 = xmm3[1],xmm0[1]9866; AVX-NEXT:    vmovaps 32(%r8), %ymm09867; AVX-NEXT:    vunpckhps {{.*#+}} xmm4 = xmm4[2],xmm5[2],xmm4[3],xmm5[3]9868; AVX-NEXT:    vmovaps 32(%r9), %ymm19869; AVX-NEXT:    vinsertf128 $1, %xmm4, %ymm0, %ymm49870; AVX-NEXT:    vshufps {{.*#+}} xmm2 = xmm2[2,3,2,3]9871; AVX-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm29872; AVX-NEXT:    vblendps {{.*#+}} ymm4 = ymm2[0,1,2,3,4,5],ymm4[6,7]9873; AVX-NEXT:    vmovaps 32(%r10), %ymm29874; AVX-NEXT:    vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm4[4,5,6,7]9875; AVX-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9876; AVX-NEXT:    vmovaps 32(%rax), %ymm39877; AVX-NEXT:    vunpcklpd {{.*#+}} ymm4 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]9878; AVX-NEXT:    vunpcklps {{.*#+}} ymm5 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]9879; AVX-NEXT:    vshufps {{.*#+}} ymm6 = ymm5[0,1],ymm4[2,0],ymm5[4,5],ymm4[6,4]9880; AVX-NEXT:    vmovaps 32(%rdx), %ymm49881; AVX-NEXT:    vmovaps 32(%rcx), %ymm59882; AVX-NEXT:    vunpcklpd {{.*#+}} ymm7 = ymm5[0],ymm4[0],ymm5[2],ymm4[2]9883; AVX-NEXT:    vextractf128 $1, %ymm7, %xmm79884; AVX-NEXT:    vmovaps 32(%rdi), %ymm89885; AVX-NEXT:    vmovaps 32(%rsi), %ymm99886; AVX-NEXT:    vunpcklps {{.*#+}} ymm10 = ymm8[0],ymm9[0],ymm8[1],ymm9[1],ymm8[4],ymm9[4],ymm8[5],ymm9[5]9887; AVX-NEXT:    vextractf128 $1, %ymm10, %xmm109888; AVX-NEXT:    vshufps {{.*#+}} xmm7 = xmm10[0,1],xmm7[2,0]9889; AVX-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]9890; AVX-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9891; AVX-NEXT:    vunpcklps {{.*#+}} ymm6 = ymm2[0],ymm3[0],ymm2[1],ymm3[1],ymm2[4],ymm3[4],ymm2[5],ymm3[5]9892; AVX-NEXT:    vshufps {{.*#+}} ymm7 = ymm1[1,0],ymm0[1,0],ymm1[5,4],ymm0[5,4]9893; AVX-NEXT:    vshufps {{.*#+}} ymm6 = ymm7[2,0],ymm6[2,3],ymm7[6,4],ymm6[6,7]9894; AVX-NEXT:    vunpcklps {{.*#+}} ymm7 = ymm4[0],ymm5[0],ymm4[1],ymm5[1],ymm4[4],ymm5[4],ymm4[5],ymm5[5]9895; AVX-NEXT:    vextractf128 $1, %ymm7, %xmm79896; AVX-NEXT:    vshufps {{.*#+}} ymm10 = ymm9[1,0],ymm8[1,0],ymm9[5,4],ymm8[5,4]9897; AVX-NEXT:    vextractf128 $1, %ymm10, %xmm109898; AVX-NEXT:    vshufps {{.*#+}} xmm7 = xmm10[2,0],xmm7[2,3]9899; AVX-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]9900; AVX-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9901; AVX-NEXT:    vunpckhpd {{.*#+}} ymm6 = ymm3[1],ymm2[1],ymm3[3],ymm2[3]9902; AVX-NEXT:    vunpckhps {{.*#+}} ymm7 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7]9903; AVX-NEXT:    vshufps {{.*#+}} ymm6 = ymm7[0,1],ymm6[2,0],ymm7[4,5],ymm6[6,4]9904; AVX-NEXT:    vunpckhpd {{.*#+}} ymm7 = ymm5[1],ymm4[1],ymm5[3],ymm4[3]9905; AVX-NEXT:    vextractf128 $1, %ymm7, %xmm79906; AVX-NEXT:    vunpckhps {{.*#+}} ymm10 = ymm8[2],ymm9[2],ymm8[3],ymm9[3],ymm8[6],ymm9[6],ymm8[7],ymm9[7]9907; AVX-NEXT:    vextractf128 $1, %ymm10, %xmm109908; AVX-NEXT:    vshufps {{.*#+}} xmm7 = xmm10[0,1],xmm7[2,0]9909; AVX-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]9910; AVX-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9911; AVX-NEXT:    vunpckhps {{.*#+}} ymm2 = ymm2[2],ymm3[2],ymm2[3],ymm3[3],ymm2[6],ymm3[6],ymm2[7],ymm3[7]9912; AVX-NEXT:    vshufps {{.*#+}} ymm0 = ymm1[3,0],ymm0[3,0],ymm1[7,4],ymm0[7,4]9913; AVX-NEXT:    vshufps {{.*#+}} ymm0 = ymm0[2,0],ymm2[2,3],ymm0[6,4],ymm2[6,7]9914; AVX-NEXT:    vunpckhps {{.*#+}} ymm1 = ymm4[2],ymm5[2],ymm4[3],ymm5[3],ymm4[6],ymm5[6],ymm4[7],ymm5[7]9915; AVX-NEXT:    vshufps {{.*#+}} ymm2 = ymm9[3,0],ymm8[3,0],ymm9[7,4],ymm8[7,4]9916; AVX-NEXT:    vextractf128 $1, %ymm1, %xmm19917; AVX-NEXT:    vextractf128 $1, %ymm2, %xmm29918; AVX-NEXT:    vshufps {{.*#+}} xmm1 = xmm2[2,0],xmm1[2,3]9919; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]9920; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9921; AVX-NEXT:    vmovaps 64(%r9), %xmm49922; AVX-NEXT:    vmovaps 64(%r8), %xmm59923; AVX-NEXT:    vunpcklps {{.*#+}} xmm0 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]9924; AVX-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm29925; AVX-NEXT:    vmovaps 64(%rax), %xmm09926; AVX-NEXT:    vmovaps 64(%r10), %xmm19927; AVX-NEXT:    vunpcklps {{.*#+}} xmm6 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]9928; AVX-NEXT:    vshufps {{.*#+}} xmm3 = xmm6[0,1,0,1]9929; AVX-NEXT:    vinsertf128 $1, %xmm3, %ymm0, %ymm39930; AVX-NEXT:    vblendps {{.*#+}} ymm7 = ymm2[0,1,2,3,4,5],ymm3[6,7]9931; AVX-NEXT:    vmovaps 64(%rcx), %xmm29932; AVX-NEXT:    vmovaps 64(%rdx), %xmm39933; AVX-NEXT:    vunpcklps {{.*#+}} xmm8 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]9934; AVX-NEXT:    vmovaps 64(%rsi), %xmm99935; AVX-NEXT:    vmovaps 64(%rdi), %xmm109936; AVX-NEXT:    vunpcklps {{.*#+}} xmm11 = xmm10[0],xmm9[0],xmm10[1],xmm9[1]9937; AVX-NEXT:    vmovlhps {{.*#+}} xmm11 = xmm11[0],xmm8[0]9938; AVX-NEXT:    vblendps {{.*#+}} ymm7 = ymm11[0,1,2,3],ymm7[4,5,6,7]9939; AVX-NEXT:    vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9940; AVX-NEXT:    vshufps {{.*#+}} xmm7 = xmm10[1,1,1,1]9941; AVX-NEXT:    vblendps {{.*#+}} xmm7 = xmm7[0],xmm9[1],xmm7[2,3]9942; AVX-NEXT:    vblendps {{.*#+}} xmm7 = xmm7[0,1],xmm8[2,3]9943; AVX-NEXT:    vinsertf128 $1, %xmm6, %ymm0, %ymm69944; AVX-NEXT:    vshufps {{.*#+}} xmm8 = xmm5[1,1,1,1]9945; AVX-NEXT:    vblendps {{.*#+}} xmm8 = xmm8[0],xmm4[1],xmm8[2,3]9946; AVX-NEXT:    vinsertf128 $1, %xmm8, %ymm0, %ymm89947; AVX-NEXT:    vblendps {{.*#+}} ymm6 = ymm8[0,1,2,3,4,5],ymm6[6,7]9948; AVX-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]9949; AVX-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9950; AVX-NEXT:    vunpckhps {{.*#+}} xmm6 = xmm10[2],xmm9[2],xmm10[3],xmm9[3]9951; AVX-NEXT:    vunpckhps {{.*#+}} xmm4 = xmm5[2],xmm4[2],xmm5[3],xmm4[3]9952; AVX-NEXT:    vshufps {{.*#+}} xmm5 = xmm0[2,2,2,2]9953; AVX-NEXT:    vblendps {{.*#+}} xmm5 = xmm1[0,1,2],xmm5[3]9954; AVX-NEXT:    vinsertf128 $1, %xmm5, %ymm0, %ymm59955; AVX-NEXT:    vinsertf128 $1, %xmm4, %ymm0, %ymm79956; AVX-NEXT:    vblendps {{.*#+}} ymm5 = ymm7[0,1,2,3,4,5],ymm5[6,7]9957; AVX-NEXT:    vshufps {{.*#+}} xmm7 = xmm2[2,2,2,2]9958; AVX-NEXT:    vblendps {{.*#+}} xmm7 = xmm3[0,1,2],xmm7[3]9959; AVX-NEXT:    vblendps {{.*#+}} xmm7 = xmm6[0,1],xmm7[2,3]9960; AVX-NEXT:    vblendps {{.*#+}} ymm5 = ymm7[0,1,2,3],ymm5[4,5,6,7]9961; AVX-NEXT:    vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9962; AVX-NEXT:    vunpckhps {{.*#+}} xmm2 = xmm3[2],xmm2[2],xmm3[3],xmm2[3]9963; AVX-NEXT:    vunpckhpd {{.*#+}} xmm2 = xmm6[1],xmm2[1]9964; AVX-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm1[2],xmm0[2],xmm1[3],xmm0[3]9965; AVX-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm09966; AVX-NEXT:    vshufps {{.*#+}} xmm1 = xmm4[2,3,2,3]9967; AVX-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm19968; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]9969; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]9970; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9971; AVX-NEXT:    vmovaps 64(%r8), %ymm09972; AVX-NEXT:    vmovaps 64(%r9), %ymm19973; AVX-NEXT:    vmovaps 64(%r10), %ymm29974; AVX-NEXT:    vmovaps 64(%rax), %ymm39975; AVX-NEXT:    vunpcklpd {{.*#+}} ymm4 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]9976; AVX-NEXT:    vunpcklps {{.*#+}} ymm5 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]9977; AVX-NEXT:    vshufps {{.*#+}} ymm6 = ymm5[0,1],ymm4[2,0],ymm5[4,5],ymm4[6,4]9978; AVX-NEXT:    vmovaps 64(%rdx), %ymm49979; AVX-NEXT:    vmovaps 64(%rcx), %ymm59980; AVX-NEXT:    vunpcklpd {{.*#+}} ymm7 = ymm5[0],ymm4[0],ymm5[2],ymm4[2]9981; AVX-NEXT:    vextractf128 $1, %ymm7, %xmm79982; AVX-NEXT:    vmovaps 64(%rdi), %ymm89983; AVX-NEXT:    vmovaps 64(%rsi), %ymm99984; AVX-NEXT:    vunpcklps {{.*#+}} ymm10 = ymm8[0],ymm9[0],ymm8[1],ymm9[1],ymm8[4],ymm9[4],ymm8[5],ymm9[5]9985; AVX-NEXT:    vextractf128 $1, %ymm10, %xmm109986; AVX-NEXT:    vshufps {{.*#+}} xmm7 = xmm10[0,1],xmm7[2,0]9987; AVX-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]9988; AVX-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9989; AVX-NEXT:    vunpcklps {{.*#+}} ymm6 = ymm2[0],ymm3[0],ymm2[1],ymm3[1],ymm2[4],ymm3[4],ymm2[5],ymm3[5]9990; AVX-NEXT:    vshufps {{.*#+}} ymm7 = ymm1[1,0],ymm0[1,0],ymm1[5,4],ymm0[5,4]9991; AVX-NEXT:    vshufps {{.*#+}} ymm6 = ymm7[2,0],ymm6[2,3],ymm7[6,4],ymm6[6,7]9992; AVX-NEXT:    vunpcklps {{.*#+}} ymm7 = ymm4[0],ymm5[0],ymm4[1],ymm5[1],ymm4[4],ymm5[4],ymm4[5],ymm5[5]9993; AVX-NEXT:    vextractf128 $1, %ymm7, %xmm79994; AVX-NEXT:    vshufps {{.*#+}} ymm10 = ymm9[1,0],ymm8[1,0],ymm9[5,4],ymm8[5,4]9995; AVX-NEXT:    vextractf128 $1, %ymm10, %xmm109996; AVX-NEXT:    vshufps {{.*#+}} xmm7 = xmm10[2,0],xmm7[2,3]9997; AVX-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]9998; AVX-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9999; AVX-NEXT:    vunpckhpd {{.*#+}} ymm6 = ymm3[1],ymm2[1],ymm3[3],ymm2[3]10000; AVX-NEXT:    vunpckhps {{.*#+}} ymm7 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7]10001; AVX-NEXT:    vshufps {{.*#+}} ymm6 = ymm7[0,1],ymm6[2,0],ymm7[4,5],ymm6[6,4]10002; AVX-NEXT:    vunpckhpd {{.*#+}} ymm7 = ymm5[1],ymm4[1],ymm5[3],ymm4[3]10003; AVX-NEXT:    vextractf128 $1, %ymm7, %xmm710004; AVX-NEXT:    vunpckhps {{.*#+}} ymm10 = ymm8[2],ymm9[2],ymm8[3],ymm9[3],ymm8[6],ymm9[6],ymm8[7],ymm9[7]10005; AVX-NEXT:    vextractf128 $1, %ymm10, %xmm1010006; AVX-NEXT:    vshufps {{.*#+}} xmm7 = xmm10[0,1],xmm7[2,0]10007; AVX-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]10008; AVX-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10009; AVX-NEXT:    vunpckhps {{.*#+}} ymm2 = ymm2[2],ymm3[2],ymm2[3],ymm3[3],ymm2[6],ymm3[6],ymm2[7],ymm3[7]10010; AVX-NEXT:    vshufps {{.*#+}} ymm0 = ymm1[3,0],ymm0[3,0],ymm1[7,4],ymm0[7,4]10011; AVX-NEXT:    vshufps {{.*#+}} ymm0 = ymm0[2,0],ymm2[2,3],ymm0[6,4],ymm2[6,7]10012; AVX-NEXT:    vunpckhps {{.*#+}} ymm1 = ymm4[2],ymm5[2],ymm4[3],ymm5[3],ymm4[6],ymm5[6],ymm4[7],ymm5[7]10013; AVX-NEXT:    vshufps {{.*#+}} ymm2 = ymm9[3,0],ymm8[3,0],ymm9[7,4],ymm8[7,4]10014; AVX-NEXT:    vextractf128 $1, %ymm1, %xmm110015; AVX-NEXT:    vextractf128 $1, %ymm2, %xmm210016; AVX-NEXT:    vshufps {{.*#+}} xmm1 = xmm2[2,0],xmm1[2,3]10017; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]10018; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10019; AVX-NEXT:    vmovaps 96(%r9), %xmm410020; AVX-NEXT:    vmovaps 96(%r8), %xmm510021; AVX-NEXT:    vunpcklps {{.*#+}} xmm0 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]10022; AVX-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm210023; AVX-NEXT:    vmovaps 96(%rax), %xmm010024; AVX-NEXT:    vmovaps 96(%r10), %xmm110025; AVX-NEXT:    vunpcklps {{.*#+}} xmm6 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]10026; AVX-NEXT:    vshufps {{.*#+}} xmm3 = xmm6[0,1,0,1]10027; AVX-NEXT:    vinsertf128 $1, %xmm3, %ymm0, %ymm310028; AVX-NEXT:    vblendps {{.*#+}} ymm7 = ymm2[0,1,2,3,4,5],ymm3[6,7]10029; AVX-NEXT:    vmovaps 96(%rcx), %xmm210030; AVX-NEXT:    vmovaps 96(%rdx), %xmm310031; AVX-NEXT:    vunpcklps {{.*#+}} xmm8 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]10032; AVX-NEXT:    vmovaps 96(%rsi), %xmm910033; AVX-NEXT:    vmovaps 96(%rdi), %xmm1010034; AVX-NEXT:    vunpcklps {{.*#+}} xmm11 = xmm10[0],xmm9[0],xmm10[1],xmm9[1]10035; AVX-NEXT:    vmovlhps {{.*#+}} xmm11 = xmm11[0],xmm8[0]10036; AVX-NEXT:    vblendps {{.*#+}} ymm7 = ymm11[0,1,2,3],ymm7[4,5,6,7]10037; AVX-NEXT:    vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10038; AVX-NEXT:    vshufps {{.*#+}} xmm7 = xmm10[1,1,1,1]10039; AVX-NEXT:    vblendps {{.*#+}} xmm7 = xmm7[0],xmm9[1],xmm7[2,3]10040; AVX-NEXT:    vblendps {{.*#+}} xmm7 = xmm7[0,1],xmm8[2,3]10041; AVX-NEXT:    vinsertf128 $1, %xmm6, %ymm0, %ymm610042; AVX-NEXT:    vshufps {{.*#+}} xmm8 = xmm5[1,1,1,1]10043; AVX-NEXT:    vblendps {{.*#+}} xmm8 = xmm8[0],xmm4[1],xmm8[2,3]10044; AVX-NEXT:    vinsertf128 $1, %xmm8, %ymm0, %ymm810045; AVX-NEXT:    vblendps {{.*#+}} ymm6 = ymm8[0,1,2,3,4,5],ymm6[6,7]10046; AVX-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]10047; AVX-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10048; AVX-NEXT:    vunpckhps {{.*#+}} xmm6 = xmm10[2],xmm9[2],xmm10[3],xmm9[3]10049; AVX-NEXT:    vunpckhps {{.*#+}} xmm4 = xmm5[2],xmm4[2],xmm5[3],xmm4[3]10050; AVX-NEXT:    vshufps {{.*#+}} xmm5 = xmm0[2,2,2,2]10051; AVX-NEXT:    vblendps {{.*#+}} xmm5 = xmm1[0,1,2],xmm5[3]10052; AVX-NEXT:    vinsertf128 $1, %xmm5, %ymm0, %ymm510053; AVX-NEXT:    vinsertf128 $1, %xmm4, %ymm0, %ymm710054; AVX-NEXT:    vblendps {{.*#+}} ymm5 = ymm7[0,1,2,3,4,5],ymm5[6,7]10055; AVX-NEXT:    vshufps {{.*#+}} xmm7 = xmm2[2,2,2,2]10056; AVX-NEXT:    vblendps {{.*#+}} xmm7 = xmm3[0,1,2],xmm7[3]10057; AVX-NEXT:    vblendps {{.*#+}} xmm7 = xmm6[0,1],xmm7[2,3]10058; AVX-NEXT:    vblendps {{.*#+}} ymm5 = ymm7[0,1,2,3],ymm5[4,5,6,7]10059; AVX-NEXT:    vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10060; AVX-NEXT:    vunpckhps {{.*#+}} xmm2 = xmm3[2],xmm2[2],xmm3[3],xmm2[3]10061; AVX-NEXT:    vunpckhpd {{.*#+}} xmm2 = xmm6[1],xmm2[1]10062; AVX-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm1[2],xmm0[2],xmm1[3],xmm0[3]10063; AVX-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm010064; AVX-NEXT:    vshufps {{.*#+}} xmm1 = xmm4[2,3,2,3]10065; AVX-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm110066; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]10067; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]10068; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10069; AVX-NEXT:    vmovaps 96(%r8), %ymm010070; AVX-NEXT:    vmovaps 96(%r9), %ymm110071; AVX-NEXT:    vmovaps 96(%r10), %ymm210072; AVX-NEXT:    vmovaps 96(%rax), %ymm310073; AVX-NEXT:    vunpcklpd {{.*#+}} ymm4 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]10074; AVX-NEXT:    vunpcklps {{.*#+}} ymm5 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]10075; AVX-NEXT:    vshufps {{.*#+}} ymm6 = ymm5[0,1],ymm4[2,0],ymm5[4,5],ymm4[6,4]10076; AVX-NEXT:    vmovaps 96(%rdx), %ymm410077; AVX-NEXT:    vmovaps 96(%rcx), %ymm510078; AVX-NEXT:    vunpcklpd {{.*#+}} ymm7 = ymm5[0],ymm4[0],ymm5[2],ymm4[2]10079; AVX-NEXT:    vextractf128 $1, %ymm7, %xmm710080; AVX-NEXT:    vmovaps 96(%rdi), %ymm810081; AVX-NEXT:    vmovaps 96(%rsi), %ymm910082; AVX-NEXT:    vunpcklps {{.*#+}} ymm10 = ymm8[0],ymm9[0],ymm8[1],ymm9[1],ymm8[4],ymm9[4],ymm8[5],ymm9[5]10083; AVX-NEXT:    vextractf128 $1, %ymm10, %xmm1010084; AVX-NEXT:    vshufps {{.*#+}} xmm7 = xmm10[0,1],xmm7[2,0]10085; AVX-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]10086; AVX-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10087; AVX-NEXT:    vunpcklps {{.*#+}} ymm6 = ymm2[0],ymm3[0],ymm2[1],ymm3[1],ymm2[4],ymm3[4],ymm2[5],ymm3[5]10088; AVX-NEXT:    vshufps {{.*#+}} ymm7 = ymm1[1,0],ymm0[1,0],ymm1[5,4],ymm0[5,4]10089; AVX-NEXT:    vshufps {{.*#+}} ymm6 = ymm7[2,0],ymm6[2,3],ymm7[6,4],ymm6[6,7]10090; AVX-NEXT:    vunpcklps {{.*#+}} ymm7 = ymm4[0],ymm5[0],ymm4[1],ymm5[1],ymm4[4],ymm5[4],ymm4[5],ymm5[5]10091; AVX-NEXT:    vextractf128 $1, %ymm7, %xmm710092; AVX-NEXT:    vshufps {{.*#+}} ymm10 = ymm9[1,0],ymm8[1,0],ymm9[5,4],ymm8[5,4]10093; AVX-NEXT:    vextractf128 $1, %ymm10, %xmm1010094; AVX-NEXT:    vshufps {{.*#+}} xmm7 = xmm10[2,0],xmm7[2,3]10095; AVX-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]10096; AVX-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10097; AVX-NEXT:    vunpckhpd {{.*#+}} ymm6 = ymm3[1],ymm2[1],ymm3[3],ymm2[3]10098; AVX-NEXT:    vunpckhps {{.*#+}} ymm7 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7]10099; AVX-NEXT:    vshufps {{.*#+}} ymm6 = ymm7[0,1],ymm6[2,0],ymm7[4,5],ymm6[6,4]10100; AVX-NEXT:    vunpckhpd {{.*#+}} ymm7 = ymm5[1],ymm4[1],ymm5[3],ymm4[3]10101; AVX-NEXT:    vextractf128 $1, %ymm7, %xmm710102; AVX-NEXT:    vunpckhps {{.*#+}} ymm10 = ymm8[2],ymm9[2],ymm8[3],ymm9[3],ymm8[6],ymm9[6],ymm8[7],ymm9[7]10103; AVX-NEXT:    vextractf128 $1, %ymm10, %xmm1010104; AVX-NEXT:    vshufps {{.*#+}} xmm7 = xmm10[0,1],xmm7[2,0]10105; AVX-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]10106; AVX-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10107; AVX-NEXT:    vunpckhps {{.*#+}} ymm2 = ymm2[2],ymm3[2],ymm2[3],ymm3[3],ymm2[6],ymm3[6],ymm2[7],ymm3[7]10108; AVX-NEXT:    vshufps {{.*#+}} ymm0 = ymm1[3,0],ymm0[3,0],ymm1[7,4],ymm0[7,4]10109; AVX-NEXT:    vshufps {{.*#+}} ymm0 = ymm0[2,0],ymm2[2,3],ymm0[6,4],ymm2[6,7]10110; AVX-NEXT:    vunpckhps {{.*#+}} ymm1 = ymm4[2],ymm5[2],ymm4[3],ymm5[3],ymm4[6],ymm5[6],ymm4[7],ymm5[7]10111; AVX-NEXT:    vshufps {{.*#+}} ymm2 = ymm9[3,0],ymm8[3,0],ymm9[7,4],ymm8[7,4]10112; AVX-NEXT:    vextractf128 $1, %ymm1, %xmm110113; AVX-NEXT:    vextractf128 $1, %ymm2, %xmm210114; AVX-NEXT:    vshufps {{.*#+}} xmm1 = xmm2[2,0],xmm1[2,3]10115; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]10116; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10117; AVX-NEXT:    vmovaps 128(%r9), %xmm410118; AVX-NEXT:    vmovaps 128(%r8), %xmm510119; AVX-NEXT:    vunpcklps {{.*#+}} xmm0 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]10120; AVX-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm210121; AVX-NEXT:    vmovaps 128(%rax), %xmm010122; AVX-NEXT:    vmovaps 128(%r10), %xmm110123; AVX-NEXT:    vunpcklps {{.*#+}} xmm6 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]10124; AVX-NEXT:    vshufps {{.*#+}} xmm3 = xmm6[0,1,0,1]10125; AVX-NEXT:    vinsertf128 $1, %xmm3, %ymm0, %ymm310126; AVX-NEXT:    vblendps {{.*#+}} ymm7 = ymm2[0,1,2,3,4,5],ymm3[6,7]10127; AVX-NEXT:    vmovaps 128(%rcx), %xmm210128; AVX-NEXT:    vmovaps 128(%rdx), %xmm310129; AVX-NEXT:    vunpcklps {{.*#+}} xmm8 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]10130; AVX-NEXT:    vmovaps 128(%rsi), %xmm910131; AVX-NEXT:    vmovaps 128(%rdi), %xmm1010132; AVX-NEXT:    vunpcklps {{.*#+}} xmm11 = xmm10[0],xmm9[0],xmm10[1],xmm9[1]10133; AVX-NEXT:    vmovlhps {{.*#+}} xmm11 = xmm11[0],xmm8[0]10134; AVX-NEXT:    vblendps {{.*#+}} ymm7 = ymm11[0,1,2,3],ymm7[4,5,6,7]10135; AVX-NEXT:    vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10136; AVX-NEXT:    vshufps {{.*#+}} xmm7 = xmm10[1,1,1,1]10137; AVX-NEXT:    vblendps {{.*#+}} xmm7 = xmm7[0],xmm9[1],xmm7[2,3]10138; AVX-NEXT:    vblendps {{.*#+}} xmm7 = xmm7[0,1],xmm8[2,3]10139; AVX-NEXT:    vinsertf128 $1, %xmm6, %ymm0, %ymm610140; AVX-NEXT:    vshufps {{.*#+}} xmm8 = xmm5[1,1,1,1]10141; AVX-NEXT:    vblendps {{.*#+}} xmm8 = xmm8[0],xmm4[1],xmm8[2,3]10142; AVX-NEXT:    vinsertf128 $1, %xmm8, %ymm0, %ymm810143; AVX-NEXT:    vblendps {{.*#+}} ymm6 = ymm8[0,1,2,3,4,5],ymm6[6,7]10144; AVX-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]10145; AVX-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10146; AVX-NEXT:    vunpckhps {{.*#+}} xmm6 = xmm10[2],xmm9[2],xmm10[3],xmm9[3]10147; AVX-NEXT:    vunpckhps {{.*#+}} xmm4 = xmm5[2],xmm4[2],xmm5[3],xmm4[3]10148; AVX-NEXT:    vshufps {{.*#+}} xmm5 = xmm0[2,2,2,2]10149; AVX-NEXT:    vblendps {{.*#+}} xmm5 = xmm1[0,1,2],xmm5[3]10150; AVX-NEXT:    vinsertf128 $1, %xmm5, %ymm0, %ymm510151; AVX-NEXT:    vinsertf128 $1, %xmm4, %ymm0, %ymm710152; AVX-NEXT:    vblendps {{.*#+}} ymm5 = ymm7[0,1,2,3,4,5],ymm5[6,7]10153; AVX-NEXT:    vshufps {{.*#+}} xmm7 = xmm2[2,2,2,2]10154; AVX-NEXT:    vblendps {{.*#+}} xmm7 = xmm3[0,1,2],xmm7[3]10155; AVX-NEXT:    vblendps {{.*#+}} xmm7 = xmm6[0,1],xmm7[2,3]10156; AVX-NEXT:    vblendps {{.*#+}} ymm5 = ymm7[0,1,2,3],ymm5[4,5,6,7]10157; AVX-NEXT:    vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10158; AVX-NEXT:    vunpckhps {{.*#+}} xmm2 = xmm3[2],xmm2[2],xmm3[3],xmm2[3]10159; AVX-NEXT:    vunpckhpd {{.*#+}} xmm2 = xmm6[1],xmm2[1]10160; AVX-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm1[2],xmm0[2],xmm1[3],xmm0[3]10161; AVX-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm010162; AVX-NEXT:    vshufps {{.*#+}} xmm1 = xmm4[2,3,2,3]10163; AVX-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm110164; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]10165; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]10166; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10167; AVX-NEXT:    vmovaps 128(%r8), %ymm010168; AVX-NEXT:    vmovaps 128(%r9), %ymm110169; AVX-NEXT:    vmovaps 128(%r10), %ymm210170; AVX-NEXT:    vmovaps 128(%rax), %ymm310171; AVX-NEXT:    vunpcklpd {{.*#+}} ymm4 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]10172; AVX-NEXT:    vunpcklps {{.*#+}} ymm5 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]10173; AVX-NEXT:    vshufps {{.*#+}} ymm6 = ymm5[0,1],ymm4[2,0],ymm5[4,5],ymm4[6,4]10174; AVX-NEXT:    vmovaps 128(%rdx), %ymm410175; AVX-NEXT:    vmovaps 128(%rcx), %ymm510176; AVX-NEXT:    vunpcklpd {{.*#+}} ymm7 = ymm5[0],ymm4[0],ymm5[2],ymm4[2]10177; AVX-NEXT:    vextractf128 $1, %ymm7, %xmm710178; AVX-NEXT:    vmovaps 128(%rdi), %ymm810179; AVX-NEXT:    vmovaps 128(%rsi), %ymm910180; AVX-NEXT:    vunpcklps {{.*#+}} ymm10 = ymm8[0],ymm9[0],ymm8[1],ymm9[1],ymm8[4],ymm9[4],ymm8[5],ymm9[5]10181; AVX-NEXT:    vextractf128 $1, %ymm10, %xmm1010182; AVX-NEXT:    vshufps {{.*#+}} xmm7 = xmm10[0,1],xmm7[2,0]10183; AVX-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]10184; AVX-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10185; AVX-NEXT:    vunpcklps {{.*#+}} ymm6 = ymm2[0],ymm3[0],ymm2[1],ymm3[1],ymm2[4],ymm3[4],ymm2[5],ymm3[5]10186; AVX-NEXT:    vshufps {{.*#+}} ymm7 = ymm1[1,0],ymm0[1,0],ymm1[5,4],ymm0[5,4]10187; AVX-NEXT:    vshufps {{.*#+}} ymm6 = ymm7[2,0],ymm6[2,3],ymm7[6,4],ymm6[6,7]10188; AVX-NEXT:    vunpcklps {{.*#+}} ymm7 = ymm4[0],ymm5[0],ymm4[1],ymm5[1],ymm4[4],ymm5[4],ymm4[5],ymm5[5]10189; AVX-NEXT:    vextractf128 $1, %ymm7, %xmm710190; AVX-NEXT:    vshufps {{.*#+}} ymm10 = ymm9[1,0],ymm8[1,0],ymm9[5,4],ymm8[5,4]10191; AVX-NEXT:    vextractf128 $1, %ymm10, %xmm1010192; AVX-NEXT:    vshufps {{.*#+}} xmm7 = xmm10[2,0],xmm7[2,3]10193; AVX-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]10194; AVX-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10195; AVX-NEXT:    vunpckhpd {{.*#+}} ymm6 = ymm3[1],ymm2[1],ymm3[3],ymm2[3]10196; AVX-NEXT:    vunpckhps {{.*#+}} ymm7 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7]10197; AVX-NEXT:    vshufps {{.*#+}} ymm6 = ymm7[0,1],ymm6[2,0],ymm7[4,5],ymm6[6,4]10198; AVX-NEXT:    vunpckhpd {{.*#+}} ymm7 = ymm5[1],ymm4[1],ymm5[3],ymm4[3]10199; AVX-NEXT:    vextractf128 $1, %ymm7, %xmm710200; AVX-NEXT:    vunpckhps {{.*#+}} ymm10 = ymm8[2],ymm9[2],ymm8[3],ymm9[3],ymm8[6],ymm9[6],ymm8[7],ymm9[7]10201; AVX-NEXT:    vextractf128 $1, %ymm10, %xmm1010202; AVX-NEXT:    vshufps {{.*#+}} xmm7 = xmm10[0,1],xmm7[2,0]10203; AVX-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]10204; AVX-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10205; AVX-NEXT:    vunpckhps {{.*#+}} ymm2 = ymm2[2],ymm3[2],ymm2[3],ymm3[3],ymm2[6],ymm3[6],ymm2[7],ymm3[7]10206; AVX-NEXT:    vshufps {{.*#+}} ymm0 = ymm1[3,0],ymm0[3,0],ymm1[7,4],ymm0[7,4]10207; AVX-NEXT:    vshufps {{.*#+}} ymm0 = ymm0[2,0],ymm2[2,3],ymm0[6,4],ymm2[6,7]10208; AVX-NEXT:    vunpckhps {{.*#+}} ymm1 = ymm4[2],ymm5[2],ymm4[3],ymm5[3],ymm4[6],ymm5[6],ymm4[7],ymm5[7]10209; AVX-NEXT:    vshufps {{.*#+}} ymm2 = ymm9[3,0],ymm8[3,0],ymm9[7,4],ymm8[7,4]10210; AVX-NEXT:    vextractf128 $1, %ymm1, %xmm110211; AVX-NEXT:    vextractf128 $1, %ymm2, %xmm210212; AVX-NEXT:    vshufps {{.*#+}} xmm1 = xmm2[2,0],xmm1[2,3]10213; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]10214; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10215; AVX-NEXT:    vmovaps 160(%r9), %xmm410216; AVX-NEXT:    vmovaps 160(%r8), %xmm510217; AVX-NEXT:    vunpcklps {{.*#+}} xmm0 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]10218; AVX-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm210219; AVX-NEXT:    vmovaps 160(%rax), %xmm010220; AVX-NEXT:    vmovaps 160(%r10), %xmm110221; AVX-NEXT:    vunpcklps {{.*#+}} xmm6 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]10222; AVX-NEXT:    vshufps {{.*#+}} xmm3 = xmm6[0,1,0,1]10223; AVX-NEXT:    vinsertf128 $1, %xmm3, %ymm0, %ymm310224; AVX-NEXT:    vblendps {{.*#+}} ymm7 = ymm2[0,1,2,3,4,5],ymm3[6,7]10225; AVX-NEXT:    vmovaps 160(%rcx), %xmm210226; AVX-NEXT:    vmovaps 160(%rdx), %xmm310227; AVX-NEXT:    vunpcklps {{.*#+}} xmm8 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]10228; AVX-NEXT:    vmovaps 160(%rsi), %xmm910229; AVX-NEXT:    vmovaps 160(%rdi), %xmm1010230; AVX-NEXT:    vunpcklps {{.*#+}} xmm11 = xmm10[0],xmm9[0],xmm10[1],xmm9[1]10231; AVX-NEXT:    vmovlhps {{.*#+}} xmm11 = xmm11[0],xmm8[0]10232; AVX-NEXT:    vblendps {{.*#+}} ymm7 = ymm11[0,1,2,3],ymm7[4,5,6,7]10233; AVX-NEXT:    vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10234; AVX-NEXT:    vshufps {{.*#+}} xmm7 = xmm10[1,1,1,1]10235; AVX-NEXT:    vblendps {{.*#+}} xmm7 = xmm7[0],xmm9[1],xmm7[2,3]10236; AVX-NEXT:    vblendps {{.*#+}} xmm7 = xmm7[0,1],xmm8[2,3]10237; AVX-NEXT:    vinsertf128 $1, %xmm6, %ymm0, %ymm610238; AVX-NEXT:    vshufps {{.*#+}} xmm8 = xmm5[1,1,1,1]10239; AVX-NEXT:    vblendps {{.*#+}} xmm8 = xmm8[0],xmm4[1],xmm8[2,3]10240; AVX-NEXT:    vinsertf128 $1, %xmm8, %ymm0, %ymm810241; AVX-NEXT:    vblendps {{.*#+}} ymm6 = ymm8[0,1,2,3,4,5],ymm6[6,7]10242; AVX-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]10243; AVX-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10244; AVX-NEXT:    vunpckhps {{.*#+}} xmm6 = xmm10[2],xmm9[2],xmm10[3],xmm9[3]10245; AVX-NEXT:    vunpckhps {{.*#+}} xmm4 = xmm5[2],xmm4[2],xmm5[3],xmm4[3]10246; AVX-NEXT:    vshufps {{.*#+}} xmm5 = xmm0[2,2,2,2]10247; AVX-NEXT:    vblendps {{.*#+}} xmm5 = xmm1[0,1,2],xmm5[3]10248; AVX-NEXT:    vinsertf128 $1, %xmm5, %ymm0, %ymm510249; AVX-NEXT:    vinsertf128 $1, %xmm4, %ymm0, %ymm710250; AVX-NEXT:    vblendps {{.*#+}} ymm5 = ymm7[0,1,2,3,4,5],ymm5[6,7]10251; AVX-NEXT:    vshufps {{.*#+}} xmm7 = xmm2[2,2,2,2]10252; AVX-NEXT:    vblendps {{.*#+}} xmm7 = xmm3[0,1,2],xmm7[3]10253; AVX-NEXT:    vblendps {{.*#+}} xmm7 = xmm6[0,1],xmm7[2,3]10254; AVX-NEXT:    vblendps {{.*#+}} ymm5 = ymm7[0,1,2,3],ymm5[4,5,6,7]10255; AVX-NEXT:    vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10256; AVX-NEXT:    vunpckhps {{.*#+}} xmm2 = xmm3[2],xmm2[2],xmm3[3],xmm2[3]10257; AVX-NEXT:    vunpckhpd {{.*#+}} xmm2 = xmm6[1],xmm2[1]10258; AVX-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm1[2],xmm0[2],xmm1[3],xmm0[3]10259; AVX-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm010260; AVX-NEXT:    vshufps {{.*#+}} xmm1 = xmm4[2,3,2,3]10261; AVX-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm110262; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]10263; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]10264; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10265; AVX-NEXT:    vmovaps 160(%r8), %ymm010266; AVX-NEXT:    vmovaps 160(%r9), %ymm110267; AVX-NEXT:    vmovaps 160(%r10), %ymm210268; AVX-NEXT:    vmovaps 160(%rax), %ymm310269; AVX-NEXT:    vunpcklpd {{.*#+}} ymm4 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]10270; AVX-NEXT:    vunpcklps {{.*#+}} ymm5 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]10271; AVX-NEXT:    vshufps {{.*#+}} ymm6 = ymm5[0,1],ymm4[2,0],ymm5[4,5],ymm4[6,4]10272; AVX-NEXT:    vmovaps 160(%rdx), %ymm410273; AVX-NEXT:    vmovaps 160(%rcx), %ymm510274; AVX-NEXT:    vunpcklpd {{.*#+}} ymm7 = ymm5[0],ymm4[0],ymm5[2],ymm4[2]10275; AVX-NEXT:    vextractf128 $1, %ymm7, %xmm710276; AVX-NEXT:    vmovaps 160(%rdi), %ymm810277; AVX-NEXT:    vmovaps 160(%rsi), %ymm910278; AVX-NEXT:    vunpcklps {{.*#+}} ymm10 = ymm8[0],ymm9[0],ymm8[1],ymm9[1],ymm8[4],ymm9[4],ymm8[5],ymm9[5]10279; AVX-NEXT:    vextractf128 $1, %ymm10, %xmm1010280; AVX-NEXT:    vshufps {{.*#+}} xmm7 = xmm10[0,1],xmm7[2,0]10281; AVX-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]10282; AVX-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10283; AVX-NEXT:    vunpcklps {{.*#+}} ymm6 = ymm2[0],ymm3[0],ymm2[1],ymm3[1],ymm2[4],ymm3[4],ymm2[5],ymm3[5]10284; AVX-NEXT:    vshufps {{.*#+}} ymm7 = ymm1[1,0],ymm0[1,0],ymm1[5,4],ymm0[5,4]10285; AVX-NEXT:    vshufps {{.*#+}} ymm6 = ymm7[2,0],ymm6[2,3],ymm7[6,4],ymm6[6,7]10286; AVX-NEXT:    vunpcklps {{.*#+}} ymm7 = ymm4[0],ymm5[0],ymm4[1],ymm5[1],ymm4[4],ymm5[4],ymm4[5],ymm5[5]10287; AVX-NEXT:    vextractf128 $1, %ymm7, %xmm710288; AVX-NEXT:    vshufps {{.*#+}} ymm10 = ymm9[1,0],ymm8[1,0],ymm9[5,4],ymm8[5,4]10289; AVX-NEXT:    vextractf128 $1, %ymm10, %xmm1010290; AVX-NEXT:    vshufps {{.*#+}} xmm7 = xmm10[2,0],xmm7[2,3]10291; AVX-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]10292; AVX-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10293; AVX-NEXT:    vunpckhpd {{.*#+}} ymm6 = ymm3[1],ymm2[1],ymm3[3],ymm2[3]10294; AVX-NEXT:    vunpckhps {{.*#+}} ymm7 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7]10295; AVX-NEXT:    vshufps {{.*#+}} ymm6 = ymm7[0,1],ymm6[2,0],ymm7[4,5],ymm6[6,4]10296; AVX-NEXT:    vunpckhpd {{.*#+}} ymm7 = ymm5[1],ymm4[1],ymm5[3],ymm4[3]10297; AVX-NEXT:    vextractf128 $1, %ymm7, %xmm710298; AVX-NEXT:    vunpckhps {{.*#+}} ymm10 = ymm8[2],ymm9[2],ymm8[3],ymm9[3],ymm8[6],ymm9[6],ymm8[7],ymm9[7]10299; AVX-NEXT:    vextractf128 $1, %ymm10, %xmm1010300; AVX-NEXT:    vshufps {{.*#+}} xmm7 = xmm10[0,1],xmm7[2,0]10301; AVX-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]10302; AVX-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10303; AVX-NEXT:    vunpckhps {{.*#+}} ymm2 = ymm2[2],ymm3[2],ymm2[3],ymm3[3],ymm2[6],ymm3[6],ymm2[7],ymm3[7]10304; AVX-NEXT:    vshufps {{.*#+}} ymm0 = ymm1[3,0],ymm0[3,0],ymm1[7,4],ymm0[7,4]10305; AVX-NEXT:    vshufps {{.*#+}} ymm0 = ymm0[2,0],ymm2[2,3],ymm0[6,4],ymm2[6,7]10306; AVX-NEXT:    vunpckhps {{.*#+}} ymm1 = ymm4[2],ymm5[2],ymm4[3],ymm5[3],ymm4[6],ymm5[6],ymm4[7],ymm5[7]10307; AVX-NEXT:    vshufps {{.*#+}} ymm2 = ymm9[3,0],ymm8[3,0],ymm9[7,4],ymm8[7,4]10308; AVX-NEXT:    vextractf128 $1, %ymm1, %xmm110309; AVX-NEXT:    vextractf128 $1, %ymm2, %xmm210310; AVX-NEXT:    vshufps {{.*#+}} xmm1 = xmm2[2,0],xmm1[2,3]10311; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]10312; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10313; AVX-NEXT:    vmovaps 192(%r9), %xmm410314; AVX-NEXT:    vmovaps 192(%r8), %xmm510315; AVX-NEXT:    vunpcklps {{.*#+}} xmm0 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]10316; AVX-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm210317; AVX-NEXT:    vmovaps 192(%rax), %xmm010318; AVX-NEXT:    vmovaps 192(%r10), %xmm110319; AVX-NEXT:    vunpcklps {{.*#+}} xmm6 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]10320; AVX-NEXT:    vshufps {{.*#+}} xmm3 = xmm6[0,1,0,1]10321; AVX-NEXT:    vinsertf128 $1, %xmm3, %ymm0, %ymm310322; AVX-NEXT:    vblendps {{.*#+}} ymm7 = ymm2[0,1,2,3,4,5],ymm3[6,7]10323; AVX-NEXT:    vmovaps 192(%rcx), %xmm210324; AVX-NEXT:    vmovaps 192(%rdx), %xmm310325; AVX-NEXT:    vunpcklps {{.*#+}} xmm8 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]10326; AVX-NEXT:    vmovaps 192(%rsi), %xmm910327; AVX-NEXT:    vmovaps 192(%rdi), %xmm1010328; AVX-NEXT:    vunpcklps {{.*#+}} xmm11 = xmm10[0],xmm9[0],xmm10[1],xmm9[1]10329; AVX-NEXT:    vmovlhps {{.*#+}} xmm11 = xmm11[0],xmm8[0]10330; AVX-NEXT:    vblendps {{.*#+}} ymm7 = ymm11[0,1,2,3],ymm7[4,5,6,7]10331; AVX-NEXT:    vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10332; AVX-NEXT:    vshufps {{.*#+}} xmm7 = xmm10[1,1,1,1]10333; AVX-NEXT:    vblendps {{.*#+}} xmm7 = xmm7[0],xmm9[1],xmm7[2,3]10334; AVX-NEXT:    vblendps {{.*#+}} xmm7 = xmm7[0,1],xmm8[2,3]10335; AVX-NEXT:    vinsertf128 $1, %xmm6, %ymm0, %ymm610336; AVX-NEXT:    vshufps {{.*#+}} xmm8 = xmm5[1,1,1,1]10337; AVX-NEXT:    vblendps {{.*#+}} xmm8 = xmm8[0],xmm4[1],xmm8[2,3]10338; AVX-NEXT:    vinsertf128 $1, %xmm8, %ymm0, %ymm810339; AVX-NEXT:    vblendps {{.*#+}} ymm6 = ymm8[0,1,2,3,4,5],ymm6[6,7]10340; AVX-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]10341; AVX-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10342; AVX-NEXT:    vunpckhps {{.*#+}} xmm6 = xmm10[2],xmm9[2],xmm10[3],xmm9[3]10343; AVX-NEXT:    vunpckhps {{.*#+}} xmm4 = xmm5[2],xmm4[2],xmm5[3],xmm4[3]10344; AVX-NEXT:    vshufps {{.*#+}} xmm5 = xmm0[2,2,2,2]10345; AVX-NEXT:    vblendps {{.*#+}} xmm5 = xmm1[0,1,2],xmm5[3]10346; AVX-NEXT:    vinsertf128 $1, %xmm5, %ymm0, %ymm510347; AVX-NEXT:    vinsertf128 $1, %xmm4, %ymm0, %ymm710348; AVX-NEXT:    vblendps {{.*#+}} ymm5 = ymm7[0,1,2,3,4,5],ymm5[6,7]10349; AVX-NEXT:    vshufps {{.*#+}} xmm7 = xmm2[2,2,2,2]10350; AVX-NEXT:    vblendps {{.*#+}} xmm7 = xmm3[0,1,2],xmm7[3]10351; AVX-NEXT:    vblendps {{.*#+}} xmm7 = xmm6[0,1],xmm7[2,3]10352; AVX-NEXT:    vblendps {{.*#+}} ymm5 = ymm7[0,1,2,3],ymm5[4,5,6,7]10353; AVX-NEXT:    vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10354; AVX-NEXT:    vunpckhps {{.*#+}} xmm2 = xmm3[2],xmm2[2],xmm3[3],xmm2[3]10355; AVX-NEXT:    vunpckhpd {{.*#+}} xmm2 = xmm6[1],xmm2[1]10356; AVX-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm1[2],xmm0[2],xmm1[3],xmm0[3]10357; AVX-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm010358; AVX-NEXT:    vshufps {{.*#+}} xmm1 = xmm4[2,3,2,3]10359; AVX-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm110360; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]10361; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]10362; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10363; AVX-NEXT:    vmovaps 192(%r8), %ymm010364; AVX-NEXT:    vmovaps 192(%r9), %ymm110365; AVX-NEXT:    vmovaps 192(%r10), %ymm210366; AVX-NEXT:    vmovaps 192(%rax), %ymm310367; AVX-NEXT:    vunpcklpd {{.*#+}} ymm4 = ymm3[0],ymm2[0],ymm3[2],ymm2[2]10368; AVX-NEXT:    vunpcklps {{.*#+}} ymm5 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]10369; AVX-NEXT:    vshufps {{.*#+}} ymm6 = ymm5[0,1],ymm4[2,0],ymm5[4,5],ymm4[6,4]10370; AVX-NEXT:    vmovaps 192(%rdx), %ymm410371; AVX-NEXT:    vmovaps 192(%rcx), %ymm510372; AVX-NEXT:    vunpcklpd {{.*#+}} ymm7 = ymm5[0],ymm4[0],ymm5[2],ymm4[2]10373; AVX-NEXT:    vextractf128 $1, %ymm7, %xmm710374; AVX-NEXT:    vmovaps 192(%rdi), %ymm810375; AVX-NEXT:    vmovaps 192(%rsi), %ymm910376; AVX-NEXT:    vunpcklps {{.*#+}} ymm10 = ymm8[0],ymm9[0],ymm8[1],ymm9[1],ymm8[4],ymm9[4],ymm8[5],ymm9[5]10377; AVX-NEXT:    vextractf128 $1, %ymm10, %xmm1010378; AVX-NEXT:    vshufps {{.*#+}} xmm7 = xmm10[0,1],xmm7[2,0]10379; AVX-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]10380; AVX-NEXT:    vmovups %ymm6, (%rsp) # 32-byte Spill10381; AVX-NEXT:    vunpcklps {{.*#+}} ymm6 = ymm2[0],ymm3[0],ymm2[1],ymm3[1],ymm2[4],ymm3[4],ymm2[5],ymm3[5]10382; AVX-NEXT:    vshufps {{.*#+}} ymm7 = ymm1[1,0],ymm0[1,0],ymm1[5,4],ymm0[5,4]10383; AVX-NEXT:    vshufps {{.*#+}} ymm6 = ymm7[2,0],ymm6[2,3],ymm7[6,4],ymm6[6,7]10384; AVX-NEXT:    vunpcklps {{.*#+}} ymm7 = ymm4[0],ymm5[0],ymm4[1],ymm5[1],ymm4[4],ymm5[4],ymm4[5],ymm5[5]10385; AVX-NEXT:    vextractf128 $1, %ymm7, %xmm710386; AVX-NEXT:    vshufps {{.*#+}} ymm10 = ymm9[1,0],ymm8[1,0],ymm9[5,4],ymm8[5,4]10387; AVX-NEXT:    vextractf128 $1, %ymm10, %xmm1010388; AVX-NEXT:    vshufps {{.*#+}} xmm7 = xmm10[2,0],xmm7[2,3]10389; AVX-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]10390; AVX-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10391; AVX-NEXT:    vunpckhpd {{.*#+}} ymm6 = ymm3[1],ymm2[1],ymm3[3],ymm2[3]10392; AVX-NEXT:    vunpckhps {{.*#+}} ymm7 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7]10393; AVX-NEXT:    vshufps {{.*#+}} ymm6 = ymm7[0,1],ymm6[2,0],ymm7[4,5],ymm6[6,4]10394; AVX-NEXT:    vunpckhpd {{.*#+}} ymm7 = ymm5[1],ymm4[1],ymm5[3],ymm4[3]10395; AVX-NEXT:    vextractf128 $1, %ymm7, %xmm710396; AVX-NEXT:    vunpckhps {{.*#+}} ymm10 = ymm8[2],ymm9[2],ymm8[3],ymm9[3],ymm8[6],ymm9[6],ymm8[7],ymm9[7]10397; AVX-NEXT:    vextractf128 $1, %ymm10, %xmm1010398; AVX-NEXT:    vshufps {{.*#+}} xmm7 = xmm10[0,1],xmm7[2,0]10399; AVX-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]10400; AVX-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10401; AVX-NEXT:    vunpckhps {{.*#+}} ymm2 = ymm2[2],ymm3[2],ymm2[3],ymm3[3],ymm2[6],ymm3[6],ymm2[7],ymm3[7]10402; AVX-NEXT:    vshufps {{.*#+}} ymm0 = ymm1[3,0],ymm0[3,0],ymm1[7,4],ymm0[7,4]10403; AVX-NEXT:    vshufps {{.*#+}} ymm0 = ymm0[2,0],ymm2[2,3],ymm0[6,4],ymm2[6,7]10404; AVX-NEXT:    vunpckhps {{.*#+}} ymm1 = ymm4[2],ymm5[2],ymm4[3],ymm5[3],ymm4[6],ymm5[6],ymm4[7],ymm5[7]10405; AVX-NEXT:    vshufps {{.*#+}} ymm2 = ymm9[3,0],ymm8[3,0],ymm9[7,4],ymm8[7,4]10406; AVX-NEXT:    vextractf128 $1, %ymm1, %xmm110407; AVX-NEXT:    vextractf128 $1, %ymm2, %xmm210408; AVX-NEXT:    vshufps {{.*#+}} xmm1 = xmm2[2,0],xmm1[2,3]10409; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]10410; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10411; AVX-NEXT:    vmovaps 224(%r9), %xmm410412; AVX-NEXT:    vmovaps 224(%r8), %xmm510413; AVX-NEXT:    vunpcklps {{.*#+}} xmm0 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]10414; AVX-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm210415; AVX-NEXT:    vmovaps 224(%rax), %xmm910416; AVX-NEXT:    vmovaps 224(%r10), %xmm110417; AVX-NEXT:    vunpcklps {{.*#+}} xmm6 = xmm1[0],xmm9[0],xmm1[1],xmm9[1]10418; AVX-NEXT:    vshufps {{.*#+}} xmm3 = xmm6[0,1,0,1]10419; AVX-NEXT:    vinsertf128 $1, %xmm3, %ymm0, %ymm310420; AVX-NEXT:    vblendps {{.*#+}} ymm10 = ymm2[0,1,2,3,4,5],ymm3[6,7]10421; AVX-NEXT:    vmovaps 224(%rcx), %xmm210422; AVX-NEXT:    vmovaps 224(%rdx), %xmm310423; AVX-NEXT:    vunpcklps {{.*#+}} xmm8 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]10424; AVX-NEXT:    vmovaps 224(%rsi), %xmm010425; AVX-NEXT:    vmovaps 224(%rdi), %xmm710426; AVX-NEXT:    vunpcklps {{.*#+}} xmm11 = xmm7[0],xmm0[0],xmm7[1],xmm0[1]10427; AVX-NEXT:    vmovlhps {{.*#+}} xmm11 = xmm11[0],xmm8[0]10428; AVX-NEXT:    vblendps {{.*#+}} ymm10 = ymm11[0,1,2,3],ymm10[4,5,6,7]10429; AVX-NEXT:    vmovups %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10430; AVX-NEXT:    vshufps {{.*#+}} xmm10 = xmm7[1,1,1,1]10431; AVX-NEXT:    vblendps {{.*#+}} xmm10 = xmm10[0],xmm0[1],xmm10[2,3]10432; AVX-NEXT:    vblendps {{.*#+}} xmm8 = xmm10[0,1],xmm8[2,3]10433; AVX-NEXT:    vinsertf128 $1, %xmm6, %ymm0, %ymm610434; AVX-NEXT:    vshufps {{.*#+}} xmm10 = xmm5[1,1,1,1]10435; AVX-NEXT:    vblendps {{.*#+}} xmm10 = xmm10[0],xmm4[1],xmm10[2,3]10436; AVX-NEXT:    vinsertf128 $1, %xmm10, %ymm0, %ymm1010437; AVX-NEXT:    vblendps {{.*#+}} ymm6 = ymm10[0,1,2,3,4,5],ymm6[6,7]10438; AVX-NEXT:    vblendps {{.*#+}} ymm12 = ymm8[0,1,2,3],ymm6[4,5,6,7]10439; AVX-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm7[2],xmm0[2],xmm7[3],xmm0[3]10440; AVX-NEXT:    vunpckhps {{.*#+}} xmm4 = xmm5[2],xmm4[2],xmm5[3],xmm4[3]10441; AVX-NEXT:    vshufps {{.*#+}} xmm5 = xmm9[2,2,2,2]10442; AVX-NEXT:    vblendps {{.*#+}} xmm5 = xmm1[0,1,2],xmm5[3]10443; AVX-NEXT:    vinsertf128 $1, %xmm5, %ymm0, %ymm510444; AVX-NEXT:    vinsertf128 $1, %xmm4, %ymm0, %ymm610445; AVX-NEXT:    vblendps {{.*#+}} ymm5 = ymm6[0,1,2,3,4,5],ymm5[6,7]10446; AVX-NEXT:    vshufps {{.*#+}} xmm6 = xmm2[2,2,2,2]10447; AVX-NEXT:    vblendps {{.*#+}} xmm6 = xmm3[0,1,2],xmm6[3]10448; AVX-NEXT:    vblendps {{.*#+}} xmm6 = xmm0[0,1],xmm6[2,3]10449; AVX-NEXT:    vblendps {{.*#+}} ymm11 = ymm6[0,1,2,3],ymm5[4,5,6,7]10450; AVX-NEXT:    vunpckhps {{.*#+}} xmm2 = xmm3[2],xmm2[2],xmm3[3],xmm2[3]10451; AVX-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm2[1]10452; AVX-NEXT:    vunpckhps {{.*#+}} xmm1 = xmm1[2],xmm9[2],xmm1[3],xmm9[3]10453; AVX-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm110454; AVX-NEXT:    vshufps {{.*#+}} xmm2 = xmm4[2,3,2,3]10455; AVX-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm210456; AVX-NEXT:    vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]10457; AVX-NEXT:    vblendps {{.*#+}} ymm10 = ymm0[0,1,2,3],ymm1[4,5,6,7]10458; AVX-NEXT:    vmovaps 224(%r8), %ymm710459; AVX-NEXT:    vmovaps 224(%r9), %ymm610460; AVX-NEXT:    vmovaps 224(%r10), %ymm510461; AVX-NEXT:    vmovaps 224(%rax), %ymm410462; AVX-NEXT:    vunpcklpd {{.*#+}} ymm0 = ymm4[0],ymm5[0],ymm4[2],ymm5[2]10463; AVX-NEXT:    vunpcklps {{.*#+}} ymm1 = ymm7[0],ymm6[0],ymm7[1],ymm6[1],ymm7[4],ymm6[4],ymm7[5],ymm6[5]10464; AVX-NEXT:    vshufps {{.*#+}} ymm8 = ymm1[0,1],ymm0[2,0],ymm1[4,5],ymm0[6,4]10465; AVX-NEXT:    vmovaps 224(%rdx), %ymm310466; AVX-NEXT:    vmovaps 224(%rcx), %ymm210467; AVX-NEXT:    vunpcklpd {{.*#+}} ymm9 = ymm2[0],ymm3[0],ymm2[2],ymm3[2]10468; AVX-NEXT:    vextractf128 $1, %ymm9, %xmm910469; AVX-NEXT:    vmovaps 224(%rdi), %ymm110470; AVX-NEXT:    vmovaps 224(%rsi), %ymm010471; AVX-NEXT:    vunpcklps {{.*#+}} ymm15 = ymm1[0],ymm0[0],ymm1[1],ymm0[1],ymm1[4],ymm0[4],ymm1[5],ymm0[5]10472; AVX-NEXT:    vextractf128 $1, %ymm15, %xmm1510473; AVX-NEXT:    vshufps {{.*#+}} xmm9 = xmm15[0,1],xmm9[2,0]10474; AVX-NEXT:    vblendps {{.*#+}} ymm9 = ymm9[0,1,2,3],ymm8[4,5,6,7]10475; AVX-NEXT:    vunpcklps {{.*#+}} ymm8 = ymm5[0],ymm4[0],ymm5[1],ymm4[1],ymm5[4],ymm4[4],ymm5[5],ymm4[5]10476; AVX-NEXT:    vshufps {{.*#+}} ymm15 = ymm6[1,0],ymm7[1,0],ymm6[5,4],ymm7[5,4]10477; AVX-NEXT:    vshufps {{.*#+}} ymm8 = ymm15[2,0],ymm8[2,3],ymm15[6,4],ymm8[6,7]10478; AVX-NEXT:    vunpcklps {{.*#+}} ymm15 = ymm3[0],ymm2[0],ymm3[1],ymm2[1],ymm3[4],ymm2[4],ymm3[5],ymm2[5]10479; AVX-NEXT:    vextractf128 $1, %ymm15, %xmm1510480; AVX-NEXT:    vshufps {{.*#+}} ymm14 = ymm0[1,0],ymm1[1,0],ymm0[5,4],ymm1[5,4]10481; AVX-NEXT:    vextractf128 $1, %ymm14, %xmm1410482; AVX-NEXT:    vshufps {{.*#+}} xmm14 = xmm14[2,0],xmm15[2,3]10483; AVX-NEXT:    vblendps {{.*#+}} ymm8 = ymm14[0,1,2,3],ymm8[4,5,6,7]10484; AVX-NEXT:    vunpckhpd {{.*#+}} ymm14 = ymm4[1],ymm5[1],ymm4[3],ymm5[3]10485; AVX-NEXT:    vunpckhps {{.*#+}} ymm15 = ymm7[2],ymm6[2],ymm7[3],ymm6[3],ymm7[6],ymm6[6],ymm7[7],ymm6[7]10486; AVX-NEXT:    vshufps {{.*#+}} ymm14 = ymm15[0,1],ymm14[2,0],ymm15[4,5],ymm14[6,4]10487; AVX-NEXT:    vunpckhpd {{.*#+}} ymm15 = ymm2[1],ymm3[1],ymm2[3],ymm3[3]10488; AVX-NEXT:    vextractf128 $1, %ymm15, %xmm1510489; AVX-NEXT:    vunpckhps {{.*#+}} ymm13 = ymm1[2],ymm0[2],ymm1[3],ymm0[3],ymm1[6],ymm0[6],ymm1[7],ymm0[7]10490; AVX-NEXT:    vextractf128 $1, %ymm13, %xmm1310491; AVX-NEXT:    vshufps {{.*#+}} xmm13 = xmm13[0,1],xmm15[2,0]10492; AVX-NEXT:    vblendps {{.*#+}} ymm13 = ymm13[0,1,2,3],ymm14[4,5,6,7]10493; AVX-NEXT:    vunpckhps {{.*#+}} ymm4 = ymm5[2],ymm4[2],ymm5[3],ymm4[3],ymm5[6],ymm4[6],ymm5[7],ymm4[7]10494; AVX-NEXT:    vshufps {{.*#+}} ymm5 = ymm6[3,0],ymm7[3,0],ymm6[7,4],ymm7[7,4]10495; AVX-NEXT:    vshufps {{.*#+}} ymm4 = ymm5[2,0],ymm4[2,3],ymm5[6,4],ymm4[6,7]10496; AVX-NEXT:    vunpckhps {{.*#+}} ymm2 = ymm3[2],ymm2[2],ymm3[3],ymm2[3],ymm3[6],ymm2[6],ymm3[7],ymm2[7]10497; AVX-NEXT:    vshufps {{.*#+}} ymm0 = ymm0[3,0],ymm1[3,0],ymm0[7,4],ymm1[7,4]10498; AVX-NEXT:    vextractf128 $1, %ymm2, %xmm110499; AVX-NEXT:    vextractf128 $1, %ymm0, %xmm010500; AVX-NEXT:    vshufps {{.*#+}} xmm0 = xmm0[2,0],xmm1[2,3]10501; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm4[4,5,6,7]10502; AVX-NEXT:    movq {{[0-9]+}}(%rsp), %rax10503; AVX-NEXT:    vmovaps %ymm0, 2016(%rax)10504; AVX-NEXT:    vmovaps %ymm13, 1984(%rax)10505; AVX-NEXT:    vmovaps %ymm8, 1952(%rax)10506; AVX-NEXT:    vmovaps %ymm9, 1920(%rax)10507; AVX-NEXT:    vmovaps %ymm10, 1888(%rax)10508; AVX-NEXT:    vmovaps %ymm11, 1856(%rax)10509; AVX-NEXT:    vmovaps %ymm12, 1824(%rax)10510; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10511; AVX-NEXT:    vmovaps %ymm0, 1792(%rax)10512; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10513; AVX-NEXT:    vmovaps %ymm0, 1760(%rax)10514; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10515; AVX-NEXT:    vmovaps %ymm0, 1728(%rax)10516; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10517; AVX-NEXT:    vmovaps %ymm0, 1696(%rax)10518; AVX-NEXT:    vmovups (%rsp), %ymm0 # 32-byte Reload10519; AVX-NEXT:    vmovaps %ymm0, 1664(%rax)10520; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10521; AVX-NEXT:    vmovaps %ymm0, 1632(%rax)10522; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10523; AVX-NEXT:    vmovaps %ymm0, 1600(%rax)10524; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10525; AVX-NEXT:    vmovaps %ymm0, 1568(%rax)10526; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10527; AVX-NEXT:    vmovaps %ymm0, 1536(%rax)10528; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10529; AVX-NEXT:    vmovaps %ymm0, 1504(%rax)10530; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10531; AVX-NEXT:    vmovaps %ymm0, 1472(%rax)10532; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10533; AVX-NEXT:    vmovaps %ymm0, 1440(%rax)10534; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10535; AVX-NEXT:    vmovaps %ymm0, 1408(%rax)10536; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10537; AVX-NEXT:    vmovaps %ymm0, 1376(%rax)10538; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10539; AVX-NEXT:    vmovaps %ymm0, 1344(%rax)10540; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10541; AVX-NEXT:    vmovaps %ymm0, 1312(%rax)10542; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10543; AVX-NEXT:    vmovaps %ymm0, 1280(%rax)10544; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10545; AVX-NEXT:    vmovaps %ymm0, 1248(%rax)10546; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10547; AVX-NEXT:    vmovaps %ymm0, 1216(%rax)10548; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10549; AVX-NEXT:    vmovaps %ymm0, 1184(%rax)10550; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10551; AVX-NEXT:    vmovaps %ymm0, 1152(%rax)10552; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10553; AVX-NEXT:    vmovaps %ymm0, 1120(%rax)10554; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10555; AVX-NEXT:    vmovaps %ymm0, 1088(%rax)10556; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10557; AVX-NEXT:    vmovaps %ymm0, 1056(%rax)10558; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10559; AVX-NEXT:    vmovaps %ymm0, 1024(%rax)10560; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10561; AVX-NEXT:    vmovaps %ymm0, 992(%rax)10562; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10563; AVX-NEXT:    vmovaps %ymm0, 960(%rax)10564; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10565; AVX-NEXT:    vmovaps %ymm0, 928(%rax)10566; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10567; AVX-NEXT:    vmovaps %ymm0, 896(%rax)10568; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10569; AVX-NEXT:    vmovaps %ymm0, 864(%rax)10570; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10571; AVX-NEXT:    vmovaps %ymm0, 832(%rax)10572; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10573; AVX-NEXT:    vmovaps %ymm0, 800(%rax)10574; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10575; AVX-NEXT:    vmovaps %ymm0, 768(%rax)10576; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10577; AVX-NEXT:    vmovaps %ymm0, 736(%rax)10578; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10579; AVX-NEXT:    vmovaps %ymm0, 704(%rax)10580; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10581; AVX-NEXT:    vmovaps %ymm0, 672(%rax)10582; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10583; AVX-NEXT:    vmovaps %ymm0, 640(%rax)10584; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10585; AVX-NEXT:    vmovaps %ymm0, 608(%rax)10586; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10587; AVX-NEXT:    vmovaps %ymm0, 576(%rax)10588; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10589; AVX-NEXT:    vmovaps %ymm0, 544(%rax)10590; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10591; AVX-NEXT:    vmovaps %ymm0, 512(%rax)10592; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10593; AVX-NEXT:    vmovaps %ymm0, 480(%rax)10594; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10595; AVX-NEXT:    vmovaps %ymm0, 448(%rax)10596; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10597; AVX-NEXT:    vmovaps %ymm0, 416(%rax)10598; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10599; AVX-NEXT:    vmovaps %ymm0, 384(%rax)10600; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10601; AVX-NEXT:    vmovaps %ymm0, 352(%rax)10602; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10603; AVX-NEXT:    vmovaps %ymm0, 320(%rax)10604; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10605; AVX-NEXT:    vmovaps %ymm0, 288(%rax)10606; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10607; AVX-NEXT:    vmovaps %ymm0, 256(%rax)10608; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10609; AVX-NEXT:    vmovaps %ymm0, 224(%rax)10610; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10611; AVX-NEXT:    vmovaps %ymm0, 192(%rax)10612; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10613; AVX-NEXT:    vmovaps %ymm0, 160(%rax)10614; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10615; AVX-NEXT:    vmovaps %ymm0, 128(%rax)10616; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10617; AVX-NEXT:    vmovaps %ymm0, 96(%rax)10618; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10619; AVX-NEXT:    vmovaps %ymm0, 64(%rax)10620; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10621; AVX-NEXT:    vmovaps %ymm0, 32(%rax)10622; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10623; AVX-NEXT:    vmovaps %ymm0, (%rax)10624; AVX-NEXT:    addq $1704, %rsp # imm = 0x6A810625; AVX-NEXT:    vzeroupper10626; AVX-NEXT:    retq10627;10628; AVX2-LABEL: store_i32_stride8_vf64:10629; AVX2:       # %bb.0:10630; AVX2-NEXT:    subq $1672, %rsp # imm = 0x68810631; AVX2-NEXT:    movq {{[0-9]+}}(%rsp), %rax10632; AVX2-NEXT:    movq {{[0-9]+}}(%rsp), %r1010633; AVX2-NEXT:    vmovaps 32(%rdi), %ymm010634; AVX2-NEXT:    vmovaps (%rdi), %ymm410635; AVX2-NEXT:    vmovaps 32(%rsi), %ymm110636; AVX2-NEXT:    vmovaps (%rsi), %ymm510637; AVX2-NEXT:    vmovaps 32(%rdx), %ymm210638; AVX2-NEXT:    vmovaps (%rdx), %ymm610639; AVX2-NEXT:    vmovaps 32(%rcx), %ymm310640; AVX2-NEXT:    vmovaps (%rcx), %ymm710641; AVX2-NEXT:    vmovaps (%r8), %ymm910642; AVX2-NEXT:    vmovaps (%r9), %ymm1110643; AVX2-NEXT:    vmovaps (%r10), %ymm1010644; AVX2-NEXT:    vmovaps (%rax), %ymm1210645; AVX2-NEXT:    vunpcklps {{.*#+}} ymm8 = ymm10[0],ymm12[0],ymm10[1],ymm12[1],ymm10[4],ymm12[4],ymm10[5],ymm12[5]10646; AVX2-NEXT:    vunpcklps {{.*#+}} ymm13 = ymm4[0],ymm5[0],ymm4[1],ymm5[1],ymm4[4],ymm5[4],ymm4[5],ymm5[5]10647; AVX2-NEXT:    vextractf128 $1, %ymm13, %xmm1310648; AVX2-NEXT:    vunpcklps {{.*#+}} ymm14 = ymm6[0],ymm7[0],ymm6[1],ymm7[1],ymm6[4],ymm7[4],ymm6[5],ymm7[5]10649; AVX2-NEXT:    vpermpd {{.*#+}} ymm15 = ymm14[2,2,2,2]10650; AVX2-NEXT:    vblendps {{.*#+}} xmm13 = xmm13[0,1],xmm15[2,3]10651; AVX2-NEXT:    vunpcklps {{.*#+}} ymm15 = ymm9[0],ymm11[0],ymm9[1],ymm11[1],ymm9[4],ymm11[4],ymm9[5],ymm11[5]10652; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm15 = ymm15[0],ymm8[0],ymm15[2],ymm8[2]10653; AVX2-NEXT:    vblendps {{.*#+}} ymm13 = ymm13[0,1,2,3],ymm15[4,5,6,7]10654; AVX2-NEXT:    vmovups %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10655; AVX2-NEXT:    vbroadcastss 20(%r8), %ymm1310656; AVX2-NEXT:    vblendps {{.*#+}} ymm13 = ymm13[0,1,2,3,4],ymm11[5],ymm13[6,7]10657; AVX2-NEXT:    vblendps {{.*#+}} ymm13 = ymm13[0,1,2,3,4,5],ymm8[6,7]10658; AVX2-NEXT:    vextractf128 $1, %ymm14, %xmm810659; AVX2-NEXT:    vshufps {{.*#+}} ymm14 = ymm4[1,1,1,1,5,5,5,5]10660; AVX2-NEXT:    vblendps {{.*#+}} ymm14 = ymm14[0],ymm5[1],ymm14[2,3,4],ymm5[5],ymm14[6,7]10661; AVX2-NEXT:    vextractf128 $1, %ymm14, %xmm1410662; AVX2-NEXT:    vblendps {{.*#+}} xmm14 = xmm14[0,1],xmm8[2,3]10663; AVX2-NEXT:    vmovaps 32(%r8), %ymm810664; AVX2-NEXT:    vblendps {{.*#+}} ymm13 = ymm14[0,1,2,3],ymm13[4,5,6,7]10665; AVX2-NEXT:    vmovups %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10666; AVX2-NEXT:    vbroadcastss 24(%rax), %ymm1310667; AVX2-NEXT:    vblendps {{.*#+}} ymm13 = ymm10[0,1,2,3,4,5,6],ymm13[7]10668; AVX2-NEXT:    vmovaps 32(%r9), %ymm1010669; AVX2-NEXT:    vunpckhps {{.*#+}} ymm11 = ymm9[2],ymm11[2],ymm9[3],ymm11[3],ymm9[6],ymm11[6],ymm9[7],ymm11[7]10670; AVX2-NEXT:    vmovaps 32(%r10), %ymm910671; AVX2-NEXT:    vunpckhps {{.*#+}} ymm4 = ymm4[2],ymm5[2],ymm4[3],ymm5[3],ymm4[6],ymm5[6],ymm4[7],ymm5[7]10672; AVX2-NEXT:    vextractf128 $1, %ymm4, %xmm510673; AVX2-NEXT:    vunpckhps {{.*#+}} ymm6 = ymm6[2],ymm7[2],ymm6[3],ymm7[3],ymm6[6],ymm7[6],ymm6[7],ymm7[7]10674; AVX2-NEXT:    vpermpd {{.*#+}} ymm7 = ymm6[2,2,2,2]10675; AVX2-NEXT:    vblendps {{.*#+}} xmm5 = xmm5[0,1],xmm7[2,3]10676; AVX2-NEXT:    vmovaps 32(%rax), %ymm710677; AVX2-NEXT:    vblendps {{.*#+}} ymm13 = ymm11[0,1,2,3,4,5],ymm13[6,7]10678; AVX2-NEXT:    vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm13[4,5,6,7]10679; AVX2-NEXT:    vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10680; AVX2-NEXT:    vbroadcastss 28(%r10), %ymm510681; AVX2-NEXT:    vunpckhps {{.*#+}} ymm5 = ymm5[2],ymm12[2],ymm5[3],ymm12[3],ymm5[6],ymm12[6],ymm5[7],ymm12[7]10682; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm5 = ymm11[1],ymm5[1],ymm11[3],ymm5[3]10683; AVX2-NEXT:    vextractf128 $1, %ymm6, %xmm610684; AVX2-NEXT:    vshufps {{.*#+}} ymm4 = ymm4[2,3,2,3,6,7,6,7]10685; AVX2-NEXT:    vextractf128 $1, %ymm4, %xmm410686; AVX2-NEXT:    vblendps {{.*#+}} xmm4 = xmm4[0,1],xmm6[2,3]10687; AVX2-NEXT:    vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm5[4,5,6,7]10688; AVX2-NEXT:    vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10689; AVX2-NEXT:    vunpcklps {{.*#+}} ymm4 = ymm9[0],ymm7[0],ymm9[1],ymm7[1],ymm9[4],ymm7[4],ymm9[5],ymm7[5]10690; AVX2-NEXT:    vunpcklps {{.*#+}} ymm5 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]10691; AVX2-NEXT:    vextractf128 $1, %ymm5, %xmm510692; AVX2-NEXT:    vunpcklps {{.*#+}} ymm6 = ymm2[0],ymm3[0],ymm2[1],ymm3[1],ymm2[4],ymm3[4],ymm2[5],ymm3[5]10693; AVX2-NEXT:    vpermpd {{.*#+}} ymm11 = ymm6[2,2,2,2]10694; AVX2-NEXT:    vblendps {{.*#+}} xmm5 = xmm5[0,1],xmm11[2,3]10695; AVX2-NEXT:    vunpcklps {{.*#+}} ymm11 = ymm8[0],ymm10[0],ymm8[1],ymm10[1],ymm8[4],ymm10[4],ymm8[5],ymm10[5]10696; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm11 = ymm11[0],ymm4[0],ymm11[2],ymm4[2]10697; AVX2-NEXT:    vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm11[4,5,6,7]10698; AVX2-NEXT:    vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10699; AVX2-NEXT:    vbroadcastss 52(%r8), %ymm510700; AVX2-NEXT:    vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3,4],ymm10[5],ymm5[6,7]10701; AVX2-NEXT:    vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3,4,5],ymm4[6,7]10702; AVX2-NEXT:    vextractf128 $1, %ymm6, %xmm410703; AVX2-NEXT:    vshufps {{.*#+}} ymm6 = ymm0[1,1,1,1,5,5,5,5]10704; AVX2-NEXT:    vblendps {{.*#+}} ymm6 = ymm6[0],ymm1[1],ymm6[2,3,4],ymm1[5],ymm6[6,7]10705; AVX2-NEXT:    vextractf128 $1, %ymm6, %xmm610706; AVX2-NEXT:    vblendps {{.*#+}} xmm6 = xmm6[0,1],xmm4[2,3]10707; AVX2-NEXT:    vmovaps 64(%rdi), %ymm410708; AVX2-NEXT:    vblendps {{.*#+}} ymm5 = ymm6[0,1,2,3],ymm5[4,5,6,7]10709; AVX2-NEXT:    vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10710; AVX2-NEXT:    vbroadcastss 56(%rax), %ymm510711; AVX2-NEXT:    vblendps {{.*#+}} ymm9 = ymm9[0,1,2,3,4,5,6],ymm5[7]10712; AVX2-NEXT:    vmovaps 64(%rsi), %ymm510713; AVX2-NEXT:    vunpckhps {{.*#+}} ymm8 = ymm8[2],ymm10[2],ymm8[3],ymm10[3],ymm8[6],ymm10[6],ymm8[7],ymm10[7]10714; AVX2-NEXT:    vmovaps 64(%rdx), %ymm610715; AVX2-NEXT:    vunpckhps {{.*#+}} ymm10 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7]10716; AVX2-NEXT:    vextractf128 $1, %ymm10, %xmm010717; AVX2-NEXT:    vunpckhps {{.*#+}} ymm2 = ymm2[2],ymm3[2],ymm2[3],ymm3[3],ymm2[6],ymm3[6],ymm2[7],ymm3[7]10718; AVX2-NEXT:    vpermpd {{.*#+}} ymm1 = ymm2[2,2,2,2]10719; AVX2-NEXT:    vblendps {{.*#+}} xmm1 = xmm0[0,1],xmm1[2,3]10720; AVX2-NEXT:    vmovaps 64(%rcx), %ymm010721; AVX2-NEXT:    vblendps {{.*#+}} ymm3 = ymm8[0,1,2,3,4,5],ymm9[6,7]10722; AVX2-NEXT:    vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm3[4,5,6,7]10723; AVX2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10724; AVX2-NEXT:    vbroadcastss 60(%r10), %ymm110725; AVX2-NEXT:    vunpckhps {{.*#+}} ymm3 = ymm1[2],ymm7[2],ymm1[3],ymm7[3],ymm1[6],ymm7[6],ymm1[7],ymm7[7]10726; AVX2-NEXT:    vmovaps 64(%r8), %ymm110727; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm3 = ymm8[1],ymm3[1],ymm8[3],ymm3[3]10728; AVX2-NEXT:    vextractf128 $1, %ymm2, %xmm210729; AVX2-NEXT:    vshufps {{.*#+}} ymm7 = ymm10[2,3,2,3,6,7,6,7]10730; AVX2-NEXT:    vextractf128 $1, %ymm7, %xmm710731; AVX2-NEXT:    vblendps {{.*#+}} xmm2 = xmm7[0,1],xmm2[2,3]10732; AVX2-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]10733; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10734; AVX2-NEXT:    vunpcklps {{.*#+}} ymm2 = ymm4[0],ymm5[0],ymm4[1],ymm5[1],ymm4[4],ymm5[4],ymm4[5],ymm5[5]10735; AVX2-NEXT:    vextractf128 $1, %ymm2, %xmm210736; AVX2-NEXT:    vunpcklps {{.*#+}} ymm3 = ymm6[0],ymm0[0],ymm6[1],ymm0[1],ymm6[4],ymm0[4],ymm6[5],ymm0[5]10737; AVX2-NEXT:    vpermpd {{.*#+}} ymm7 = ymm3[2,2,2,2]10738; AVX2-NEXT:    vblendps {{.*#+}} xmm2 = xmm2[0,1],xmm7[2,3]10739; AVX2-NEXT:    vmovaps 64(%r9), %ymm710740; AVX2-NEXT:    vmovaps 64(%r10), %ymm810741; AVX2-NEXT:    vmovaps 64(%rax), %ymm910742; AVX2-NEXT:    vunpcklps {{.*#+}} ymm10 = ymm8[0],ymm9[0],ymm8[1],ymm9[1],ymm8[4],ymm9[4],ymm8[5],ymm9[5]10743; AVX2-NEXT:    vunpcklps {{.*#+}} ymm11 = ymm1[0],ymm7[0],ymm1[1],ymm7[1],ymm1[4],ymm7[4],ymm1[5],ymm7[5]10744; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm11 = ymm11[0],ymm10[0],ymm11[2],ymm10[2]10745; AVX2-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm11[4,5,6,7]10746; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10747; AVX2-NEXT:    vbroadcastss 84(%r8), %ymm210748; AVX2-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4],ymm7[5],ymm2[6,7]10749; AVX2-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm10[6,7]10750; AVX2-NEXT:    vextractf128 $1, %ymm3, %xmm310751; AVX2-NEXT:    vshufps {{.*#+}} ymm10 = ymm4[1,1,1,1,5,5,5,5]10752; AVX2-NEXT:    vblendps {{.*#+}} ymm10 = ymm10[0],ymm5[1],ymm10[2,3,4],ymm5[5],ymm10[6,7]10753; AVX2-NEXT:    vextractf128 $1, %ymm10, %xmm1010754; AVX2-NEXT:    vblendps {{.*#+}} xmm3 = xmm10[0,1],xmm3[2,3]10755; AVX2-NEXT:    vblendps {{.*#+}} ymm2 = ymm3[0,1,2,3],ymm2[4,5,6,7]10756; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10757; AVX2-NEXT:    vbroadcastss 88(%rax), %ymm210758; AVX2-NEXT:    vblendps {{.*#+}} ymm2 = ymm8[0,1,2,3,4,5,6],ymm2[7]10759; AVX2-NEXT:    vunpckhps {{.*#+}} ymm1 = ymm1[2],ymm7[2],ymm1[3],ymm7[3],ymm1[6],ymm7[6],ymm1[7],ymm7[7]10760; AVX2-NEXT:    vunpckhps {{.*#+}} ymm3 = ymm4[2],ymm5[2],ymm4[3],ymm5[3],ymm4[6],ymm5[6],ymm4[7],ymm5[7]10761; AVX2-NEXT:    vunpckhps {{.*#+}} ymm4 = ymm6[2],ymm0[2],ymm6[3],ymm0[3],ymm6[6],ymm0[6],ymm6[7],ymm0[7]10762; AVX2-NEXT:    vextractf128 $1, %ymm3, %xmm010763; AVX2-NEXT:    vpermpd {{.*#+}} ymm5 = ymm4[2,2,2,2]10764; AVX2-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm5[2,3]10765; AVX2-NEXT:    vblendps {{.*#+}} ymm2 = ymm1[0,1,2,3,4,5],ymm2[6,7]10766; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]10767; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10768; AVX2-NEXT:    vbroadcastss 92(%r10), %ymm010769; AVX2-NEXT:    vunpckhps {{.*#+}} ymm2 = ymm0[2],ymm9[2],ymm0[3],ymm9[3],ymm0[6],ymm9[6],ymm0[7],ymm9[7]10770; AVX2-NEXT:    vmovaps 96(%rdi), %ymm010771; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm5 = ymm1[1],ymm2[1],ymm1[3],ymm2[3]10772; AVX2-NEXT:    vmovaps 96(%rsi), %ymm110773; AVX2-NEXT:    vextractf128 $1, %ymm4, %xmm210774; AVX2-NEXT:    vshufps {{.*#+}} ymm3 = ymm3[2,3,2,3,6,7,6,7]10775; AVX2-NEXT:    vextractf128 $1, %ymm3, %xmm310776; AVX2-NEXT:    vblendps {{.*#+}} xmm3 = xmm3[0,1],xmm2[2,3]10777; AVX2-NEXT:    vmovaps 96(%rdx), %ymm210778; AVX2-NEXT:    vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm5[4,5,6,7]10779; AVX2-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10780; AVX2-NEXT:    vunpcklps {{.*#+}} ymm3 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]10781; AVX2-NEXT:    vextractf128 $1, %ymm3, %xmm310782; AVX2-NEXT:    vmovaps 96(%rcx), %ymm410783; AVX2-NEXT:    vunpcklps {{.*#+}} ymm5 = ymm2[0],ymm4[0],ymm2[1],ymm4[1],ymm2[4],ymm4[4],ymm2[5],ymm4[5]10784; AVX2-NEXT:    vpermpd {{.*#+}} ymm6 = ymm5[2,2,2,2]10785; AVX2-NEXT:    vblendps {{.*#+}} xmm3 = xmm3[0,1],xmm6[2,3]10786; AVX2-NEXT:    vmovaps 96(%r8), %ymm610787; AVX2-NEXT:    vmovaps 96(%r9), %ymm710788; AVX2-NEXT:    vmovaps 96(%r10), %ymm810789; AVX2-NEXT:    vmovaps 96(%rax), %ymm910790; AVX2-NEXT:    vunpcklps {{.*#+}} ymm10 = ymm8[0],ymm9[0],ymm8[1],ymm9[1],ymm8[4],ymm9[4],ymm8[5],ymm9[5]10791; AVX2-NEXT:    vunpcklps {{.*#+}} ymm11 = ymm6[0],ymm7[0],ymm6[1],ymm7[1],ymm6[4],ymm7[4],ymm6[5],ymm7[5]10792; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm11 = ymm11[0],ymm10[0],ymm11[2],ymm10[2]10793; AVX2-NEXT:    vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm11[4,5,6,7]10794; AVX2-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10795; AVX2-NEXT:    vbroadcastss 116(%r8), %ymm310796; AVX2-NEXT:    vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4],ymm7[5],ymm3[6,7]10797; AVX2-NEXT:    vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm10[6,7]10798; AVX2-NEXT:    vextractf128 $1, %ymm5, %xmm510799; AVX2-NEXT:    vshufps {{.*#+}} ymm10 = ymm0[1,1,1,1,5,5,5,5]10800; AVX2-NEXT:    vblendps {{.*#+}} ymm10 = ymm10[0],ymm1[1],ymm10[2,3,4],ymm1[5],ymm10[6,7]10801; AVX2-NEXT:    vextractf128 $1, %ymm10, %xmm1010802; AVX2-NEXT:    vblendps {{.*#+}} xmm5 = xmm10[0,1],xmm5[2,3]10803; AVX2-NEXT:    vblendps {{.*#+}} ymm3 = ymm5[0,1,2,3],ymm3[4,5,6,7]10804; AVX2-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10805; AVX2-NEXT:    vbroadcastss 120(%rax), %ymm310806; AVX2-NEXT:    vblendps {{.*#+}} ymm3 = ymm8[0,1,2,3,4,5,6],ymm3[7]10807; AVX2-NEXT:    vunpckhps {{.*#+}} ymm5 = ymm6[2],ymm7[2],ymm6[3],ymm7[3],ymm6[6],ymm7[6],ymm6[7],ymm7[7]10808; AVX2-NEXT:    vunpckhps {{.*#+}} ymm0 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7]10809; AVX2-NEXT:    vunpckhps {{.*#+}} ymm1 = ymm2[2],ymm4[2],ymm2[3],ymm4[3],ymm2[6],ymm4[6],ymm2[7],ymm4[7]10810; AVX2-NEXT:    vextractf128 $1, %ymm0, %xmm210811; AVX2-NEXT:    vpermpd {{.*#+}} ymm4 = ymm1[2,2,2,2]10812; AVX2-NEXT:    vblendps {{.*#+}} xmm2 = xmm2[0,1],xmm4[2,3]10813; AVX2-NEXT:    vblendps {{.*#+}} ymm3 = ymm5[0,1,2,3,4,5],ymm3[6,7]10814; AVX2-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]10815; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10816; AVX2-NEXT:    vbroadcastss 124(%r10), %ymm210817; AVX2-NEXT:    vunpckhps {{.*#+}} ymm2 = ymm2[2],ymm9[2],ymm2[3],ymm9[3],ymm2[6],ymm9[6],ymm2[7],ymm9[7]10818; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm5[1],ymm2[1],ymm5[3],ymm2[3]10819; AVX2-NEXT:    vextractf128 $1, %ymm1, %xmm110820; AVX2-NEXT:    vshufps {{.*#+}} ymm0 = ymm0[2,3,2,3,6,7,6,7]10821; AVX2-NEXT:    vextractf128 $1, %ymm0, %xmm010822; AVX2-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]10823; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]10824; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10825; AVX2-NEXT:    vmovaps 128(%rdi), %ymm010826; AVX2-NEXT:    vmovaps 128(%rsi), %ymm110827; AVX2-NEXT:    vunpcklps {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]10828; AVX2-NEXT:    vextractf128 $1, %ymm2, %xmm210829; AVX2-NEXT:    vmovaps 128(%rdx), %ymm310830; AVX2-NEXT:    vmovaps 128(%rcx), %ymm410831; AVX2-NEXT:    vunpcklps {{.*#+}} ymm5 = ymm3[0],ymm4[0],ymm3[1],ymm4[1],ymm3[4],ymm4[4],ymm3[5],ymm4[5]10832; AVX2-NEXT:    vpermpd {{.*#+}} ymm6 = ymm5[2,2,2,2]10833; AVX2-NEXT:    vblendps {{.*#+}} xmm2 = xmm2[0,1],xmm6[2,3]10834; AVX2-NEXT:    vmovaps 128(%r8), %ymm610835; AVX2-NEXT:    vmovaps 128(%r9), %ymm710836; AVX2-NEXT:    vmovaps 128(%r10), %ymm810837; AVX2-NEXT:    vmovaps 128(%rax), %ymm910838; AVX2-NEXT:    vunpcklps {{.*#+}} ymm10 = ymm8[0],ymm9[0],ymm8[1],ymm9[1],ymm8[4],ymm9[4],ymm8[5],ymm9[5]10839; AVX2-NEXT:    vunpcklps {{.*#+}} ymm11 = ymm6[0],ymm7[0],ymm6[1],ymm7[1],ymm6[4],ymm7[4],ymm6[5],ymm7[5]10840; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm11 = ymm11[0],ymm10[0],ymm11[2],ymm10[2]10841; AVX2-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm11[4,5,6,7]10842; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10843; AVX2-NEXT:    vbroadcastss 148(%r8), %ymm210844; AVX2-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4],ymm7[5],ymm2[6,7]10845; AVX2-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm10[6,7]10846; AVX2-NEXT:    vextractf128 $1, %ymm5, %xmm510847; AVX2-NEXT:    vshufps {{.*#+}} ymm10 = ymm0[1,1,1,1,5,5,5,5]10848; AVX2-NEXT:    vblendps {{.*#+}} ymm10 = ymm10[0],ymm1[1],ymm10[2,3,4],ymm1[5],ymm10[6,7]10849; AVX2-NEXT:    vextractf128 $1, %ymm10, %xmm1010850; AVX2-NEXT:    vblendps {{.*#+}} xmm5 = xmm10[0,1],xmm5[2,3]10851; AVX2-NEXT:    vblendps {{.*#+}} ymm2 = ymm5[0,1,2,3],ymm2[4,5,6,7]10852; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10853; AVX2-NEXT:    vbroadcastss 152(%rax), %ymm210854; AVX2-NEXT:    vblendps {{.*#+}} ymm2 = ymm8[0,1,2,3,4,5,6],ymm2[7]10855; AVX2-NEXT:    vunpckhps {{.*#+}} ymm5 = ymm6[2],ymm7[2],ymm6[3],ymm7[3],ymm6[6],ymm7[6],ymm6[7],ymm7[7]10856; AVX2-NEXT:    vunpckhps {{.*#+}} ymm0 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7]10857; AVX2-NEXT:    vunpckhps {{.*#+}} ymm1 = ymm3[2],ymm4[2],ymm3[3],ymm4[3],ymm3[6],ymm4[6],ymm3[7],ymm4[7]10858; AVX2-NEXT:    vextractf128 $1, %ymm0, %xmm310859; AVX2-NEXT:    vpermpd {{.*#+}} ymm4 = ymm1[2,2,2,2]10860; AVX2-NEXT:    vblendps {{.*#+}} xmm3 = xmm3[0,1],xmm4[2,3]10861; AVX2-NEXT:    vblendps {{.*#+}} ymm2 = ymm5[0,1,2,3,4,5],ymm2[6,7]10862; AVX2-NEXT:    vblendps {{.*#+}} ymm2 = ymm3[0,1,2,3],ymm2[4,5,6,7]10863; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10864; AVX2-NEXT:    vbroadcastss 156(%r10), %ymm210865; AVX2-NEXT:    vunpckhps {{.*#+}} ymm2 = ymm2[2],ymm9[2],ymm2[3],ymm9[3],ymm2[6],ymm9[6],ymm2[7],ymm9[7]10866; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm5[1],ymm2[1],ymm5[3],ymm2[3]10867; AVX2-NEXT:    vextractf128 $1, %ymm1, %xmm110868; AVX2-NEXT:    vshufps {{.*#+}} ymm0 = ymm0[2,3,2,3,6,7,6,7]10869; AVX2-NEXT:    vextractf128 $1, %ymm0, %xmm010870; AVX2-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]10871; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]10872; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10873; AVX2-NEXT:    vmovaps 160(%rdi), %ymm010874; AVX2-NEXT:    vmovaps 160(%rsi), %ymm110875; AVX2-NEXT:    vunpcklps {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]10876; AVX2-NEXT:    vextractf128 $1, %ymm2, %xmm210877; AVX2-NEXT:    vmovaps 160(%rdx), %ymm310878; AVX2-NEXT:    vmovaps 160(%rcx), %ymm410879; AVX2-NEXT:    vunpcklps {{.*#+}} ymm5 = ymm3[0],ymm4[0],ymm3[1],ymm4[1],ymm3[4],ymm4[4],ymm3[5],ymm4[5]10880; AVX2-NEXT:    vpermpd {{.*#+}} ymm6 = ymm5[2,2,2,2]10881; AVX2-NEXT:    vblendps {{.*#+}} xmm2 = xmm2[0,1],xmm6[2,3]10882; AVX2-NEXT:    vmovaps 160(%r8), %ymm610883; AVX2-NEXT:    vmovaps 160(%r9), %ymm710884; AVX2-NEXT:    vmovaps 160(%r10), %ymm810885; AVX2-NEXT:    vmovaps 160(%rax), %ymm910886; AVX2-NEXT:    vunpcklps {{.*#+}} ymm10 = ymm8[0],ymm9[0],ymm8[1],ymm9[1],ymm8[4],ymm9[4],ymm8[5],ymm9[5]10887; AVX2-NEXT:    vunpcklps {{.*#+}} ymm11 = ymm6[0],ymm7[0],ymm6[1],ymm7[1],ymm6[4],ymm7[4],ymm6[5],ymm7[5]10888; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm11 = ymm11[0],ymm10[0],ymm11[2],ymm10[2]10889; AVX2-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm11[4,5,6,7]10890; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10891; AVX2-NEXT:    vbroadcastss 180(%r8), %ymm210892; AVX2-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4],ymm7[5],ymm2[6,7]10893; AVX2-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm10[6,7]10894; AVX2-NEXT:    vextractf128 $1, %ymm5, %xmm510895; AVX2-NEXT:    vshufps {{.*#+}} ymm10 = ymm0[1,1,1,1,5,5,5,5]10896; AVX2-NEXT:    vblendps {{.*#+}} ymm10 = ymm10[0],ymm1[1],ymm10[2,3,4],ymm1[5],ymm10[6,7]10897; AVX2-NEXT:    vextractf128 $1, %ymm10, %xmm1010898; AVX2-NEXT:    vblendps {{.*#+}} xmm5 = xmm10[0,1],xmm5[2,3]10899; AVX2-NEXT:    vblendps {{.*#+}} ymm2 = ymm5[0,1,2,3],ymm2[4,5,6,7]10900; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10901; AVX2-NEXT:    vbroadcastss 184(%rax), %ymm210902; AVX2-NEXT:    vblendps {{.*#+}} ymm2 = ymm8[0,1,2,3,4,5,6],ymm2[7]10903; AVX2-NEXT:    vunpckhps {{.*#+}} ymm5 = ymm6[2],ymm7[2],ymm6[3],ymm7[3],ymm6[6],ymm7[6],ymm6[7],ymm7[7]10904; AVX2-NEXT:    vunpckhps {{.*#+}} ymm0 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7]10905; AVX2-NEXT:    vunpckhps {{.*#+}} ymm1 = ymm3[2],ymm4[2],ymm3[3],ymm4[3],ymm3[6],ymm4[6],ymm3[7],ymm4[7]10906; AVX2-NEXT:    vextractf128 $1, %ymm0, %xmm310907; AVX2-NEXT:    vpermpd {{.*#+}} ymm4 = ymm1[2,2,2,2]10908; AVX2-NEXT:    vblendps {{.*#+}} xmm3 = xmm3[0,1],xmm4[2,3]10909; AVX2-NEXT:    vblendps {{.*#+}} ymm2 = ymm5[0,1,2,3,4,5],ymm2[6,7]10910; AVX2-NEXT:    vblendps {{.*#+}} ymm2 = ymm3[0,1,2,3],ymm2[4,5,6,7]10911; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10912; AVX2-NEXT:    vbroadcastss 188(%r10), %ymm210913; AVX2-NEXT:    vunpckhps {{.*#+}} ymm2 = ymm2[2],ymm9[2],ymm2[3],ymm9[3],ymm2[6],ymm9[6],ymm2[7],ymm9[7]10914; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm5[1],ymm2[1],ymm5[3],ymm2[3]10915; AVX2-NEXT:    vextractf128 $1, %ymm1, %xmm110916; AVX2-NEXT:    vshufps {{.*#+}} ymm0 = ymm0[2,3,2,3,6,7,6,7]10917; AVX2-NEXT:    vextractf128 $1, %ymm0, %xmm010918; AVX2-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]10919; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]10920; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10921; AVX2-NEXT:    vmovaps 192(%rdi), %ymm010922; AVX2-NEXT:    vmovaps 192(%rsi), %ymm110923; AVX2-NEXT:    vunpcklps {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]10924; AVX2-NEXT:    vextractf128 $1, %ymm2, %xmm210925; AVX2-NEXT:    vmovaps 192(%rdx), %ymm310926; AVX2-NEXT:    vmovaps 192(%rcx), %ymm410927; AVX2-NEXT:    vunpcklps {{.*#+}} ymm5 = ymm3[0],ymm4[0],ymm3[1],ymm4[1],ymm3[4],ymm4[4],ymm3[5],ymm4[5]10928; AVX2-NEXT:    vpermpd {{.*#+}} ymm6 = ymm5[2,2,2,2]10929; AVX2-NEXT:    vblendps {{.*#+}} xmm2 = xmm2[0,1],xmm6[2,3]10930; AVX2-NEXT:    vmovaps 192(%r8), %ymm610931; AVX2-NEXT:    vmovaps 192(%r9), %ymm710932; AVX2-NEXT:    vmovaps 192(%r10), %ymm810933; AVX2-NEXT:    vmovaps 192(%rax), %ymm910934; AVX2-NEXT:    vunpcklps {{.*#+}} ymm10 = ymm8[0],ymm9[0],ymm8[1],ymm9[1],ymm8[4],ymm9[4],ymm8[5],ymm9[5]10935; AVX2-NEXT:    vunpcklps {{.*#+}} ymm11 = ymm6[0],ymm7[0],ymm6[1],ymm7[1],ymm6[4],ymm7[4],ymm6[5],ymm7[5]10936; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm11 = ymm11[0],ymm10[0],ymm11[2],ymm10[2]10937; AVX2-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm11[4,5,6,7]10938; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10939; AVX2-NEXT:    vbroadcastss 212(%r8), %ymm210940; AVX2-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4],ymm7[5],ymm2[6,7]10941; AVX2-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm10[6,7]10942; AVX2-NEXT:    vextractf128 $1, %ymm5, %xmm510943; AVX2-NEXT:    vshufps {{.*#+}} ymm10 = ymm0[1,1,1,1,5,5,5,5]10944; AVX2-NEXT:    vblendps {{.*#+}} ymm10 = ymm10[0],ymm1[1],ymm10[2,3,4],ymm1[5],ymm10[6,7]10945; AVX2-NEXT:    vextractf128 $1, %ymm10, %xmm1010946; AVX2-NEXT:    vblendps {{.*#+}} xmm5 = xmm10[0,1],xmm5[2,3]10947; AVX2-NEXT:    vblendps {{.*#+}} ymm2 = ymm5[0,1,2,3],ymm2[4,5,6,7]10948; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10949; AVX2-NEXT:    vbroadcastss 216(%rax), %ymm210950; AVX2-NEXT:    vblendps {{.*#+}} ymm2 = ymm8[0,1,2,3,4,5,6],ymm2[7]10951; AVX2-NEXT:    vunpckhps {{.*#+}} ymm5 = ymm6[2],ymm7[2],ymm6[3],ymm7[3],ymm6[6],ymm7[6],ymm6[7],ymm7[7]10952; AVX2-NEXT:    vunpckhps {{.*#+}} ymm0 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7]10953; AVX2-NEXT:    vunpckhps {{.*#+}} ymm1 = ymm3[2],ymm4[2],ymm3[3],ymm4[3],ymm3[6],ymm4[6],ymm3[7],ymm4[7]10954; AVX2-NEXT:    vextractf128 $1, %ymm0, %xmm310955; AVX2-NEXT:    vpermpd {{.*#+}} ymm4 = ymm1[2,2,2,2]10956; AVX2-NEXT:    vblendps {{.*#+}} xmm3 = xmm3[0,1],xmm4[2,3]10957; AVX2-NEXT:    vblendps {{.*#+}} ymm2 = ymm5[0,1,2,3,4,5],ymm2[6,7]10958; AVX2-NEXT:    vblendps {{.*#+}} ymm2 = ymm3[0,1,2,3],ymm2[4,5,6,7]10959; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10960; AVX2-NEXT:    vbroadcastss 220(%r10), %ymm210961; AVX2-NEXT:    vunpckhps {{.*#+}} ymm2 = ymm2[2],ymm9[2],ymm2[3],ymm9[3],ymm2[6],ymm9[6],ymm2[7],ymm9[7]10962; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm5[1],ymm2[1],ymm5[3],ymm2[3]10963; AVX2-NEXT:    vextractf128 $1, %ymm1, %xmm110964; AVX2-NEXT:    vshufps {{.*#+}} ymm0 = ymm0[2,3,2,3,6,7,6,7]10965; AVX2-NEXT:    vextractf128 $1, %ymm0, %xmm010966; AVX2-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]10967; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]10968; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10969; AVX2-NEXT:    vmovaps 224(%rdi), %ymm010970; AVX2-NEXT:    vmovaps 224(%rsi), %ymm110971; AVX2-NEXT:    vunpcklps {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]10972; AVX2-NEXT:    vextractf128 $1, %ymm2, %xmm210973; AVX2-NEXT:    vmovaps 224(%rdx), %ymm310974; AVX2-NEXT:    vmovaps 224(%rcx), %ymm410975; AVX2-NEXT:    vunpcklps {{.*#+}} ymm5 = ymm3[0],ymm4[0],ymm3[1],ymm4[1],ymm3[4],ymm4[4],ymm3[5],ymm4[5]10976; AVX2-NEXT:    vpermpd {{.*#+}} ymm6 = ymm5[2,2,2,2]10977; AVX2-NEXT:    vblendps {{.*#+}} xmm2 = xmm2[0,1],xmm6[2,3]10978; AVX2-NEXT:    vmovaps 224(%r8), %ymm610979; AVX2-NEXT:    vmovaps 224(%r9), %ymm710980; AVX2-NEXT:    vmovaps 224(%r10), %ymm810981; AVX2-NEXT:    vmovaps 224(%rax), %ymm910982; AVX2-NEXT:    vunpcklps {{.*#+}} ymm10 = ymm8[0],ymm9[0],ymm8[1],ymm9[1],ymm8[4],ymm9[4],ymm8[5],ymm9[5]10983; AVX2-NEXT:    vunpcklps {{.*#+}} ymm11 = ymm6[0],ymm7[0],ymm6[1],ymm7[1],ymm6[4],ymm7[4],ymm6[5],ymm7[5]10984; AVX2-NEXT:    vunpcklpd {{.*#+}} ymm11 = ymm11[0],ymm10[0],ymm11[2],ymm10[2]10985; AVX2-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm11[4,5,6,7]10986; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10987; AVX2-NEXT:    vbroadcastss 244(%r8), %ymm210988; AVX2-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4],ymm7[5],ymm2[6,7]10989; AVX2-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm10[6,7]10990; AVX2-NEXT:    vextractf128 $1, %ymm5, %xmm510991; AVX2-NEXT:    vshufps {{.*#+}} ymm10 = ymm0[1,1,1,1,5,5,5,5]10992; AVX2-NEXT:    vblendps {{.*#+}} ymm10 = ymm10[0],ymm1[1],ymm10[2,3,4],ymm1[5],ymm10[6,7]10993; AVX2-NEXT:    vextractf128 $1, %ymm10, %xmm1010994; AVX2-NEXT:    vblendps {{.*#+}} xmm5 = xmm10[0,1],xmm5[2,3]10995; AVX2-NEXT:    vblendps {{.*#+}} ymm2 = ymm5[0,1,2,3],ymm2[4,5,6,7]10996; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10997; AVX2-NEXT:    vbroadcastss 248(%rax), %ymm210998; AVX2-NEXT:    vblendps {{.*#+}} ymm2 = ymm8[0,1,2,3,4,5,6],ymm2[7]10999; AVX2-NEXT:    vunpckhps {{.*#+}} ymm5 = ymm6[2],ymm7[2],ymm6[3],ymm7[3],ymm6[6],ymm7[6],ymm6[7],ymm7[7]11000; AVX2-NEXT:    vunpckhps {{.*#+}} ymm0 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7]11001; AVX2-NEXT:    vunpckhps {{.*#+}} ymm1 = ymm3[2],ymm4[2],ymm3[3],ymm4[3],ymm3[6],ymm4[6],ymm3[7],ymm4[7]11002; AVX2-NEXT:    vextractf128 $1, %ymm0, %xmm311003; AVX2-NEXT:    vpermpd {{.*#+}} ymm4 = ymm1[2,2,2,2]11004; AVX2-NEXT:    vblendps {{.*#+}} xmm3 = xmm3[0,1],xmm4[2,3]11005; AVX2-NEXT:    vblendps {{.*#+}} ymm2 = ymm5[0,1,2,3,4,5],ymm2[6,7]11006; AVX2-NEXT:    vblendps {{.*#+}} ymm2 = ymm3[0,1,2,3],ymm2[4,5,6,7]11007; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11008; AVX2-NEXT:    vbroadcastss 252(%r10), %ymm211009; AVX2-NEXT:    vunpckhps {{.*#+}} ymm2 = ymm2[2],ymm9[2],ymm2[3],ymm9[3],ymm2[6],ymm9[6],ymm2[7],ymm9[7]11010; AVX2-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm5[1],ymm2[1],ymm5[3],ymm2[3]11011; AVX2-NEXT:    vextractf128 $1, %ymm1, %xmm111012; AVX2-NEXT:    vshufps {{.*#+}} ymm0 = ymm0[2,3,2,3,6,7,6,7]11013; AVX2-NEXT:    vextractf128 $1, %ymm0, %xmm011014; AVX2-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]11015; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]11016; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11017; AVX2-NEXT:    vmovaps (%rcx), %xmm011018; AVX2-NEXT:    vbroadcastss %xmm0, %xmm211019; AVX2-NEXT:    vmovaps (%rdx), %xmm111020; AVX2-NEXT:    vbroadcastss %xmm1, %xmm311021; AVX2-NEXT:    vunpcklps {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]11022; AVX2-NEXT:    vmovaps (%rsi), %xmm411023; AVX2-NEXT:    vmovaps (%rdi), %xmm511024; AVX2-NEXT:    vunpcklps {{.*#+}} xmm3 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]11025; AVX2-NEXT:    vblendps {{.*#+}} xmm6 = xmm3[0,1],xmm2[2,3]11026; AVX2-NEXT:    vmovaps (%rax), %xmm211027; AVX2-NEXT:    vbroadcastss %xmm2, %xmm711028; AVX2-NEXT:    vmovaps (%r10), %xmm311029; AVX2-NEXT:    vbroadcastss %xmm3, %xmm811030; AVX2-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm8[0],xmm7[0],xmm8[1],xmm7[1]11031; AVX2-NEXT:    vmovaps (%r9), %xmm811032; AVX2-NEXT:    vmovaps (%r8), %xmm911033; AVX2-NEXT:    vunpcklps {{.*#+}} xmm10 = xmm9[0],xmm8[0],xmm9[1],xmm8[1]11034; AVX2-NEXT:    vinsertf128 $1, %xmm10, %ymm0, %ymm1011035; AVX2-NEXT:    vinsertf128 $1, %xmm7, %ymm0, %ymm711036; AVX2-NEXT:    vblendps {{.*#+}} ymm7 = ymm10[0,1,2,3,4,5],ymm7[6,7]11037; AVX2-NEXT:    vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7]11038; AVX2-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11039; AVX2-NEXT:    vshufps {{.*#+}} xmm6 = xmm5[1,1,1,1]11040; AVX2-NEXT:    vblendps {{.*#+}} xmm6 = xmm6[0],xmm4[1],xmm6[2,3]11041; AVX2-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]11042; AVX2-NEXT:    vblendps {{.*#+}} xmm6 = xmm6[0,1],xmm7[2,3]11043; AVX2-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]11044; AVX2-NEXT:    vinsertf128 $1, %xmm7, %ymm0, %ymm711045; AVX2-NEXT:    vshufps {{.*#+}} xmm10 = xmm9[1,1,1,1]11046; AVX2-NEXT:    vblendps {{.*#+}} xmm10 = xmm10[0],xmm8[1],xmm10[2,3]11047; AVX2-NEXT:    vinsertf128 $1, %xmm10, %ymm0, %ymm1011048; AVX2-NEXT:    vblendps {{.*#+}} ymm7 = ymm10[0,1,2,3,4,5],ymm7[6,7]11049; AVX2-NEXT:    vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7]11050; AVX2-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11051; AVX2-NEXT:    vunpckhps {{.*#+}} xmm4 = xmm5[2],xmm4[2],xmm5[3],xmm4[3]11052; AVX2-NEXT:    vunpckhps {{.*#+}} xmm5 = xmm9[2],xmm8[2],xmm9[3],xmm8[3]11053; AVX2-NEXT:    vshufps {{.*#+}} xmm6 = xmm2[2,2,2,2]11054; AVX2-NEXT:    vblendps {{.*#+}} xmm6 = xmm3[0,1,2],xmm6[3]11055; AVX2-NEXT:    vinsertf128 $1, %xmm6, %ymm0, %ymm611056; AVX2-NEXT:    vinsertf128 $1, %xmm5, %ymm0, %ymm711057; AVX2-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3,4,5],ymm6[6,7]11058; AVX2-NEXT:    vshufps {{.*#+}} xmm7 = xmm0[2,2,2,2]11059; AVX2-NEXT:    vblendps {{.*#+}} xmm7 = xmm1[0,1,2],xmm7[3]11060; AVX2-NEXT:    vblendps {{.*#+}} xmm7 = xmm4[0,1],xmm7[2,3]11061; AVX2-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]11062; AVX2-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11063; AVX2-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm1[2],xmm0[2],xmm1[3],xmm0[3]11064; AVX2-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm4[1],xmm0[1]11065; AVX2-NEXT:    vunpckhps {{.*#+}} xmm1 = xmm3[2],xmm2[2],xmm3[3],xmm2[3]11066; AVX2-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm111067; AVX2-NEXT:    vshufps {{.*#+}} xmm2 = xmm5[2,3,2,3]11068; AVX2-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm211069; AVX2-NEXT:    vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]11070; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]11071; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11072; AVX2-NEXT:    vmovaps 32(%rcx), %xmm011073; AVX2-NEXT:    vmovaps 32(%rdx), %xmm111074; AVX2-NEXT:    vbroadcastss %xmm0, %xmm211075; AVX2-NEXT:    vbroadcastss %xmm1, %xmm311076; AVX2-NEXT:    vunpcklps {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]11077; AVX2-NEXT:    vmovaps 32(%rsi), %xmm411078; AVX2-NEXT:    vmovaps 32(%rdi), %xmm511079; AVX2-NEXT:    vunpcklps {{.*#+}} xmm3 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]11080; AVX2-NEXT:    vblendps {{.*#+}} xmm6 = xmm3[0,1],xmm2[2,3]11081; AVX2-NEXT:    vmovaps 32(%r10), %xmm211082; AVX2-NEXT:    vmovaps 32(%rax), %xmm311083; AVX2-NEXT:    vbroadcastss %xmm3, %xmm711084; AVX2-NEXT:    vbroadcastss %xmm2, %xmm811085; AVX2-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm8[0],xmm7[0],xmm8[1],xmm7[1]11086; AVX2-NEXT:    vmovaps 32(%r9), %xmm811087; AVX2-NEXT:    vmovaps 32(%r8), %xmm911088; AVX2-NEXT:    vunpcklps {{.*#+}} xmm10 = xmm9[0],xmm8[0],xmm9[1],xmm8[1]11089; AVX2-NEXT:    vinsertf128 $1, %xmm10, %ymm0, %ymm1011090; AVX2-NEXT:    vinsertf128 $1, %xmm7, %ymm0, %ymm711091; AVX2-NEXT:    vblendps {{.*#+}} ymm7 = ymm10[0,1,2,3,4,5],ymm7[6,7]11092; AVX2-NEXT:    vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7]11093; AVX2-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11094; AVX2-NEXT:    vshufps {{.*#+}} xmm6 = xmm5[1,1,1,1]11095; AVX2-NEXT:    vblendps {{.*#+}} xmm6 = xmm6[0],xmm4[1],xmm6[2,3]11096; AVX2-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]11097; AVX2-NEXT:    vblendps {{.*#+}} xmm6 = xmm6[0,1],xmm7[2,3]11098; AVX2-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm2[0],xmm3[0],xmm2[1],xmm3[1]11099; AVX2-NEXT:    vinsertf128 $1, %xmm7, %ymm0, %ymm711100; AVX2-NEXT:    vshufps {{.*#+}} xmm10 = xmm9[1,1,1,1]11101; AVX2-NEXT:    vblendps {{.*#+}} xmm10 = xmm10[0],xmm8[1],xmm10[2,3]11102; AVX2-NEXT:    vinsertf128 $1, %xmm10, %ymm0, %ymm1011103; AVX2-NEXT:    vblendps {{.*#+}} ymm7 = ymm10[0,1,2,3,4,5],ymm7[6,7]11104; AVX2-NEXT:    vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7]11105; AVX2-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11106; AVX2-NEXT:    vunpckhps {{.*#+}} xmm4 = xmm5[2],xmm4[2],xmm5[3],xmm4[3]11107; AVX2-NEXT:    vunpckhps {{.*#+}} xmm5 = xmm9[2],xmm8[2],xmm9[3],xmm8[3]11108; AVX2-NEXT:    vshufps {{.*#+}} xmm6 = xmm3[2,2,2,2]11109; AVX2-NEXT:    vblendps {{.*#+}} xmm6 = xmm2[0,1,2],xmm6[3]11110; AVX2-NEXT:    vinsertf128 $1, %xmm6, %ymm0, %ymm611111; AVX2-NEXT:    vinsertf128 $1, %xmm5, %ymm0, %ymm711112; AVX2-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3,4,5],ymm6[6,7]11113; AVX2-NEXT:    vshufps {{.*#+}} xmm7 = xmm0[2,2,2,2]11114; AVX2-NEXT:    vblendps {{.*#+}} xmm7 = xmm1[0,1,2],xmm7[3]11115; AVX2-NEXT:    vblendps {{.*#+}} xmm7 = xmm4[0,1],xmm7[2,3]11116; AVX2-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]11117; AVX2-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11118; AVX2-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm1[2],xmm0[2],xmm1[3],xmm0[3]11119; AVX2-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm4[1],xmm0[1]11120; AVX2-NEXT:    vunpckhps {{.*#+}} xmm1 = xmm2[2],xmm3[2],xmm2[3],xmm3[3]11121; AVX2-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm111122; AVX2-NEXT:    vshufps {{.*#+}} xmm2 = xmm5[2,3,2,3]11123; AVX2-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm211124; AVX2-NEXT:    vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]11125; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]11126; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11127; AVX2-NEXT:    vmovaps 64(%rcx), %xmm011128; AVX2-NEXT:    vmovaps 64(%rdx), %xmm111129; AVX2-NEXT:    vbroadcastss %xmm0, %xmm211130; AVX2-NEXT:    vbroadcastss %xmm1, %xmm311131; AVX2-NEXT:    vunpcklps {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]11132; AVX2-NEXT:    vmovaps 64(%rsi), %xmm411133; AVX2-NEXT:    vmovaps 64(%rdi), %xmm511134; AVX2-NEXT:    vunpcklps {{.*#+}} xmm3 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]11135; AVX2-NEXT:    vblendps {{.*#+}} xmm6 = xmm3[0,1],xmm2[2,3]11136; AVX2-NEXT:    vmovaps 64(%r10), %xmm211137; AVX2-NEXT:    vmovaps 64(%rax), %xmm311138; AVX2-NEXT:    vbroadcastss %xmm3, %xmm711139; AVX2-NEXT:    vbroadcastss %xmm2, %xmm811140; AVX2-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm8[0],xmm7[0],xmm8[1],xmm7[1]11141; AVX2-NEXT:    vmovaps 64(%r9), %xmm811142; AVX2-NEXT:    vmovaps 64(%r8), %xmm911143; AVX2-NEXT:    vunpcklps {{.*#+}} xmm10 = xmm9[0],xmm8[0],xmm9[1],xmm8[1]11144; AVX2-NEXT:    vinsertf128 $1, %xmm10, %ymm0, %ymm1011145; AVX2-NEXT:    vinsertf128 $1, %xmm7, %ymm0, %ymm711146; AVX2-NEXT:    vblendps {{.*#+}} ymm7 = ymm10[0,1,2,3,4,5],ymm7[6,7]11147; AVX2-NEXT:    vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7]11148; AVX2-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11149; AVX2-NEXT:    vshufps {{.*#+}} xmm6 = xmm5[1,1,1,1]11150; AVX2-NEXT:    vblendps {{.*#+}} xmm6 = xmm6[0],xmm4[1],xmm6[2,3]11151; AVX2-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]11152; AVX2-NEXT:    vblendps {{.*#+}} xmm6 = xmm6[0,1],xmm7[2,3]11153; AVX2-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm2[0],xmm3[0],xmm2[1],xmm3[1]11154; AVX2-NEXT:    vinsertf128 $1, %xmm7, %ymm0, %ymm711155; AVX2-NEXT:    vshufps {{.*#+}} xmm10 = xmm9[1,1,1,1]11156; AVX2-NEXT:    vblendps {{.*#+}} xmm10 = xmm10[0],xmm8[1],xmm10[2,3]11157; AVX2-NEXT:    vinsertf128 $1, %xmm10, %ymm0, %ymm1011158; AVX2-NEXT:    vblendps {{.*#+}} ymm7 = ymm10[0,1,2,3,4,5],ymm7[6,7]11159; AVX2-NEXT:    vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7]11160; AVX2-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11161; AVX2-NEXT:    vunpckhps {{.*#+}} xmm4 = xmm5[2],xmm4[2],xmm5[3],xmm4[3]11162; AVX2-NEXT:    vunpckhps {{.*#+}} xmm5 = xmm9[2],xmm8[2],xmm9[3],xmm8[3]11163; AVX2-NEXT:    vshufps {{.*#+}} xmm6 = xmm3[2,2,2,2]11164; AVX2-NEXT:    vblendps {{.*#+}} xmm6 = xmm2[0,1,2],xmm6[3]11165; AVX2-NEXT:    vinsertf128 $1, %xmm6, %ymm0, %ymm611166; AVX2-NEXT:    vinsertf128 $1, %xmm5, %ymm0, %ymm711167; AVX2-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3,4,5],ymm6[6,7]11168; AVX2-NEXT:    vshufps {{.*#+}} xmm7 = xmm0[2,2,2,2]11169; AVX2-NEXT:    vblendps {{.*#+}} xmm7 = xmm1[0,1,2],xmm7[3]11170; AVX2-NEXT:    vblendps {{.*#+}} xmm7 = xmm4[0,1],xmm7[2,3]11171; AVX2-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]11172; AVX2-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11173; AVX2-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm1[2],xmm0[2],xmm1[3],xmm0[3]11174; AVX2-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm4[1],xmm0[1]11175; AVX2-NEXT:    vunpckhps {{.*#+}} xmm1 = xmm2[2],xmm3[2],xmm2[3],xmm3[3]11176; AVX2-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm111177; AVX2-NEXT:    vshufps {{.*#+}} xmm2 = xmm5[2,3,2,3]11178; AVX2-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm211179; AVX2-NEXT:    vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]11180; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]11181; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11182; AVX2-NEXT:    vmovaps 96(%rcx), %xmm011183; AVX2-NEXT:    vbroadcastss %xmm0, %xmm211184; AVX2-NEXT:    vmovaps 96(%rdx), %xmm111185; AVX2-NEXT:    vbroadcastss %xmm1, %xmm311186; AVX2-NEXT:    vunpcklps {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]11187; AVX2-NEXT:    vmovaps 96(%rsi), %xmm411188; AVX2-NEXT:    vmovaps 96(%rdi), %xmm511189; AVX2-NEXT:    vunpcklps {{.*#+}} xmm3 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]11190; AVX2-NEXT:    vblendps {{.*#+}} xmm6 = xmm3[0,1],xmm2[2,3]11191; AVX2-NEXT:    vmovaps 96(%rax), %xmm211192; AVX2-NEXT:    vbroadcastss %xmm2, %xmm711193; AVX2-NEXT:    vmovaps 96(%r10), %xmm311194; AVX2-NEXT:    vbroadcastss %xmm3, %xmm811195; AVX2-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm8[0],xmm7[0],xmm8[1],xmm7[1]11196; AVX2-NEXT:    vmovaps 96(%r9), %xmm811197; AVX2-NEXT:    vmovaps 96(%r8), %xmm911198; AVX2-NEXT:    vunpcklps {{.*#+}} xmm10 = xmm9[0],xmm8[0],xmm9[1],xmm8[1]11199; AVX2-NEXT:    vinsertf128 $1, %xmm10, %ymm0, %ymm1011200; AVX2-NEXT:    vinsertf128 $1, %xmm7, %ymm0, %ymm711201; AVX2-NEXT:    vblendps {{.*#+}} ymm7 = ymm10[0,1,2,3,4,5],ymm7[6,7]11202; AVX2-NEXT:    vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7]11203; AVX2-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11204; AVX2-NEXT:    vshufps {{.*#+}} xmm6 = xmm5[1,1,1,1]11205; AVX2-NEXT:    vblendps {{.*#+}} xmm6 = xmm6[0],xmm4[1],xmm6[2,3]11206; AVX2-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]11207; AVX2-NEXT:    vblendps {{.*#+}} xmm6 = xmm6[0,1],xmm7[2,3]11208; AVX2-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]11209; AVX2-NEXT:    vinsertf128 $1, %xmm7, %ymm0, %ymm711210; AVX2-NEXT:    vshufps {{.*#+}} xmm10 = xmm9[1,1,1,1]11211; AVX2-NEXT:    vblendps {{.*#+}} xmm10 = xmm10[0],xmm8[1],xmm10[2,3]11212; AVX2-NEXT:    vinsertf128 $1, %xmm10, %ymm0, %ymm1011213; AVX2-NEXT:    vblendps {{.*#+}} ymm7 = ymm10[0,1,2,3,4,5],ymm7[6,7]11214; AVX2-NEXT:    vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7]11215; AVX2-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11216; AVX2-NEXT:    vunpckhps {{.*#+}} xmm4 = xmm5[2],xmm4[2],xmm5[3],xmm4[3]11217; AVX2-NEXT:    vunpckhps {{.*#+}} xmm5 = xmm9[2],xmm8[2],xmm9[3],xmm8[3]11218; AVX2-NEXT:    vshufps {{.*#+}} xmm6 = xmm2[2,2,2,2]11219; AVX2-NEXT:    vblendps {{.*#+}} xmm6 = xmm3[0,1,2],xmm6[3]11220; AVX2-NEXT:    vinsertf128 $1, %xmm6, %ymm0, %ymm611221; AVX2-NEXT:    vinsertf128 $1, %xmm5, %ymm0, %ymm711222; AVX2-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3,4,5],ymm6[6,7]11223; AVX2-NEXT:    vshufps {{.*#+}} xmm7 = xmm0[2,2,2,2]11224; AVX2-NEXT:    vblendps {{.*#+}} xmm7 = xmm1[0,1,2],xmm7[3]11225; AVX2-NEXT:    vblendps {{.*#+}} xmm7 = xmm4[0,1],xmm7[2,3]11226; AVX2-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]11227; AVX2-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11228; AVX2-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm1[2],xmm0[2],xmm1[3],xmm0[3]11229; AVX2-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm4[1],xmm0[1]11230; AVX2-NEXT:    vunpckhps {{.*#+}} xmm1 = xmm3[2],xmm2[2],xmm3[3],xmm2[3]11231; AVX2-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm111232; AVX2-NEXT:    vshufps {{.*#+}} xmm2 = xmm5[2,3,2,3]11233; AVX2-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm211234; AVX2-NEXT:    vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]11235; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]11236; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11237; AVX2-NEXT:    vmovaps 128(%rcx), %xmm011238; AVX2-NEXT:    vbroadcastss %xmm0, %xmm211239; AVX2-NEXT:    vmovaps 128(%rdx), %xmm111240; AVX2-NEXT:    vbroadcastss %xmm1, %xmm311241; AVX2-NEXT:    vunpcklps {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]11242; AVX2-NEXT:    vmovaps 128(%rsi), %xmm411243; AVX2-NEXT:    vmovaps 128(%rdi), %xmm511244; AVX2-NEXT:    vunpcklps {{.*#+}} xmm3 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]11245; AVX2-NEXT:    vblendps {{.*#+}} xmm6 = xmm3[0,1],xmm2[2,3]11246; AVX2-NEXT:    vmovaps 128(%rax), %xmm211247; AVX2-NEXT:    vbroadcastss %xmm2, %xmm711248; AVX2-NEXT:    vmovaps 128(%r10), %xmm311249; AVX2-NEXT:    vbroadcastss %xmm3, %xmm811250; AVX2-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm8[0],xmm7[0],xmm8[1],xmm7[1]11251; AVX2-NEXT:    vmovaps 128(%r9), %xmm811252; AVX2-NEXT:    vmovaps 128(%r8), %xmm911253; AVX2-NEXT:    vunpcklps {{.*#+}} xmm10 = xmm9[0],xmm8[0],xmm9[1],xmm8[1]11254; AVX2-NEXT:    vinsertf128 $1, %xmm10, %ymm0, %ymm1011255; AVX2-NEXT:    vinsertf128 $1, %xmm7, %ymm0, %ymm711256; AVX2-NEXT:    vblendps {{.*#+}} ymm7 = ymm10[0,1,2,3,4,5],ymm7[6,7]11257; AVX2-NEXT:    vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7]11258; AVX2-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11259; AVX2-NEXT:    vshufps {{.*#+}} xmm6 = xmm5[1,1,1,1]11260; AVX2-NEXT:    vblendps {{.*#+}} xmm6 = xmm6[0],xmm4[1],xmm6[2,3]11261; AVX2-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]11262; AVX2-NEXT:    vblendps {{.*#+}} xmm6 = xmm6[0,1],xmm7[2,3]11263; AVX2-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]11264; AVX2-NEXT:    vinsertf128 $1, %xmm7, %ymm0, %ymm711265; AVX2-NEXT:    vshufps {{.*#+}} xmm10 = xmm9[1,1,1,1]11266; AVX2-NEXT:    vblendps {{.*#+}} xmm10 = xmm10[0],xmm8[1],xmm10[2,3]11267; AVX2-NEXT:    vinsertf128 $1, %xmm10, %ymm0, %ymm1011268; AVX2-NEXT:    vblendps {{.*#+}} ymm7 = ymm10[0,1,2,3,4,5],ymm7[6,7]11269; AVX2-NEXT:    vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7]11270; AVX2-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11271; AVX2-NEXT:    vunpckhps {{.*#+}} xmm4 = xmm5[2],xmm4[2],xmm5[3],xmm4[3]11272; AVX2-NEXT:    vunpckhps {{.*#+}} xmm5 = xmm9[2],xmm8[2],xmm9[3],xmm8[3]11273; AVX2-NEXT:    vshufps {{.*#+}} xmm6 = xmm2[2,2,2,2]11274; AVX2-NEXT:    vblendps {{.*#+}} xmm6 = xmm3[0,1,2],xmm6[3]11275; AVX2-NEXT:    vinsertf128 $1, %xmm6, %ymm0, %ymm611276; AVX2-NEXT:    vinsertf128 $1, %xmm5, %ymm0, %ymm711277; AVX2-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3,4,5],ymm6[6,7]11278; AVX2-NEXT:    vshufps {{.*#+}} xmm7 = xmm0[2,2,2,2]11279; AVX2-NEXT:    vblendps {{.*#+}} xmm7 = xmm1[0,1,2],xmm7[3]11280; AVX2-NEXT:    vblendps {{.*#+}} xmm7 = xmm4[0,1],xmm7[2,3]11281; AVX2-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]11282; AVX2-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11283; AVX2-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm1[2],xmm0[2],xmm1[3],xmm0[3]11284; AVX2-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm4[1],xmm0[1]11285; AVX2-NEXT:    vunpckhps {{.*#+}} xmm1 = xmm3[2],xmm2[2],xmm3[3],xmm2[3]11286; AVX2-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm111287; AVX2-NEXT:    vshufps {{.*#+}} xmm2 = xmm5[2,3,2,3]11288; AVX2-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm211289; AVX2-NEXT:    vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]11290; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]11291; AVX2-NEXT:    vmovups %ymm0, (%rsp) # 32-byte Spill11292; AVX2-NEXT:    vmovaps 160(%rcx), %xmm011293; AVX2-NEXT:    vbroadcastss %xmm0, %xmm211294; AVX2-NEXT:    vmovaps 160(%rdx), %xmm111295; AVX2-NEXT:    vbroadcastss %xmm1, %xmm311296; AVX2-NEXT:    vunpcklps {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]11297; AVX2-NEXT:    vmovaps 160(%rsi), %xmm411298; AVX2-NEXT:    vmovaps 160(%rdi), %xmm511299; AVX2-NEXT:    vunpcklps {{.*#+}} xmm3 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]11300; AVX2-NEXT:    vblendps {{.*#+}} xmm6 = xmm3[0,1],xmm2[2,3]11301; AVX2-NEXT:    vmovaps 160(%rax), %xmm211302; AVX2-NEXT:    vbroadcastss %xmm2, %xmm711303; AVX2-NEXT:    vmovaps 160(%r10), %xmm311304; AVX2-NEXT:    vbroadcastss %xmm3, %xmm811305; AVX2-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm8[0],xmm7[0],xmm8[1],xmm7[1]11306; AVX2-NEXT:    vmovaps 160(%r9), %xmm811307; AVX2-NEXT:    vmovaps 160(%r8), %xmm911308; AVX2-NEXT:    vunpcklps {{.*#+}} xmm10 = xmm9[0],xmm8[0],xmm9[1],xmm8[1]11309; AVX2-NEXT:    vinsertf128 $1, %xmm10, %ymm0, %ymm1011310; AVX2-NEXT:    vinsertf128 $1, %xmm7, %ymm0, %ymm711311; AVX2-NEXT:    vblendps {{.*#+}} ymm7 = ymm10[0,1,2,3,4,5],ymm7[6,7]11312; AVX2-NEXT:    vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7]11313; AVX2-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11314; AVX2-NEXT:    vshufps {{.*#+}} xmm6 = xmm5[1,1,1,1]11315; AVX2-NEXT:    vblendps {{.*#+}} xmm6 = xmm6[0],xmm4[1],xmm6[2,3]11316; AVX2-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]11317; AVX2-NEXT:    vblendps {{.*#+}} xmm6 = xmm6[0,1],xmm7[2,3]11318; AVX2-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]11319; AVX2-NEXT:    vinsertf128 $1, %xmm7, %ymm0, %ymm711320; AVX2-NEXT:    vshufps {{.*#+}} xmm10 = xmm9[1,1,1,1]11321; AVX2-NEXT:    vblendps {{.*#+}} xmm10 = xmm10[0],xmm8[1],xmm10[2,3]11322; AVX2-NEXT:    vinsertf128 $1, %xmm10, %ymm0, %ymm1011323; AVX2-NEXT:    vblendps {{.*#+}} ymm7 = ymm10[0,1,2,3,4,5],ymm7[6,7]11324; AVX2-NEXT:    vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7]11325; AVX2-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11326; AVX2-NEXT:    vunpckhps {{.*#+}} xmm4 = xmm5[2],xmm4[2],xmm5[3],xmm4[3]11327; AVX2-NEXT:    vunpckhps {{.*#+}} xmm5 = xmm9[2],xmm8[2],xmm9[3],xmm8[3]11328; AVX2-NEXT:    vshufps {{.*#+}} xmm6 = xmm2[2,2,2,2]11329; AVX2-NEXT:    vblendps {{.*#+}} xmm6 = xmm3[0,1,2],xmm6[3]11330; AVX2-NEXT:    vinsertf128 $1, %xmm6, %ymm0, %ymm611331; AVX2-NEXT:    vinsertf128 $1, %xmm5, %ymm0, %ymm711332; AVX2-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3,4,5],ymm6[6,7]11333; AVX2-NEXT:    vshufps {{.*#+}} xmm7 = xmm0[2,2,2,2]11334; AVX2-NEXT:    vblendps {{.*#+}} xmm7 = xmm1[0,1,2],xmm7[3]11335; AVX2-NEXT:    vblendps {{.*#+}} xmm7 = xmm4[0,1],xmm7[2,3]11336; AVX2-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]11337; AVX2-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11338; AVX2-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm1[2],xmm0[2],xmm1[3],xmm0[3]11339; AVX2-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm4[1],xmm0[1]11340; AVX2-NEXT:    vunpckhps {{.*#+}} xmm1 = xmm3[2],xmm2[2],xmm3[3],xmm2[3]11341; AVX2-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm111342; AVX2-NEXT:    vshufps {{.*#+}} xmm2 = xmm5[2,3,2,3]11343; AVX2-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm211344; AVX2-NEXT:    vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]11345; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]11346; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11347; AVX2-NEXT:    vmovaps 192(%rcx), %xmm011348; AVX2-NEXT:    vbroadcastss %xmm0, %xmm211349; AVX2-NEXT:    vmovaps 192(%rdx), %xmm111350; AVX2-NEXT:    vbroadcastss %xmm1, %xmm311351; AVX2-NEXT:    vunpcklps {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]11352; AVX2-NEXT:    vmovaps 192(%rsi), %xmm411353; AVX2-NEXT:    vmovaps 192(%rdi), %xmm511354; AVX2-NEXT:    vunpcklps {{.*#+}} xmm3 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]11355; AVX2-NEXT:    vblendps {{.*#+}} xmm9 = xmm3[0,1],xmm2[2,3]11356; AVX2-NEXT:    vmovaps 192(%rax), %xmm211357; AVX2-NEXT:    vbroadcastss %xmm2, %xmm711358; AVX2-NEXT:    vmovaps 192(%r10), %xmm311359; AVX2-NEXT:    vbroadcastss %xmm3, %xmm811360; AVX2-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm8[0],xmm7[0],xmm8[1],xmm7[1]11361; AVX2-NEXT:    vmovaps 192(%r9), %xmm811362; AVX2-NEXT:    vmovaps 192(%r8), %xmm611363; AVX2-NEXT:    vunpcklps {{.*#+}} xmm10 = xmm6[0],xmm8[0],xmm6[1],xmm8[1]11364; AVX2-NEXT:    vinsertf128 $1, %xmm10, %ymm0, %ymm1011365; AVX2-NEXT:    vinsertf128 $1, %xmm7, %ymm0, %ymm711366; AVX2-NEXT:    vblendps {{.*#+}} ymm7 = ymm10[0,1,2,3,4,5],ymm7[6,7]11367; AVX2-NEXT:    vblendps {{.*#+}} ymm11 = ymm9[0,1,2,3],ymm7[4,5,6,7]11368; AVX2-NEXT:    vshufps {{.*#+}} xmm7 = xmm5[1,1,1,1]11369; AVX2-NEXT:    vblendps {{.*#+}} xmm7 = xmm7[0],xmm4[1],xmm7[2,3]11370; AVX2-NEXT:    vunpcklps {{.*#+}} xmm9 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]11371; AVX2-NEXT:    vblendps {{.*#+}} xmm7 = xmm7[0,1],xmm9[2,3]11372; AVX2-NEXT:    vunpcklps {{.*#+}} xmm9 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]11373; AVX2-NEXT:    vinsertf128 $1, %xmm9, %ymm0, %ymm911374; AVX2-NEXT:    vshufps {{.*#+}} xmm15 = xmm6[1,1,1,1]11375; AVX2-NEXT:    vblendps {{.*#+}} xmm15 = xmm15[0],xmm8[1],xmm15[2,3]11376; AVX2-NEXT:    vinsertf128 $1, %xmm15, %ymm0, %ymm1511377; AVX2-NEXT:    vblendps {{.*#+}} ymm9 = ymm15[0,1,2,3,4,5],ymm9[6,7]11378; AVX2-NEXT:    vblendps {{.*#+}} ymm10 = ymm7[0,1,2,3],ymm9[4,5,6,7]11379; AVX2-NEXT:    vunpckhps {{.*#+}} xmm4 = xmm5[2],xmm4[2],xmm5[3],xmm4[3]11380; AVX2-NEXT:    vunpckhps {{.*#+}} xmm5 = xmm6[2],xmm8[2],xmm6[3],xmm8[3]11381; AVX2-NEXT:    vshufps {{.*#+}} xmm6 = xmm2[2,2,2,2]11382; AVX2-NEXT:    vblendps {{.*#+}} xmm6 = xmm3[0,1,2],xmm6[3]11383; AVX2-NEXT:    vinsertf128 $1, %xmm6, %ymm0, %ymm611384; AVX2-NEXT:    vinsertf128 $1, %xmm5, %ymm0, %ymm711385; AVX2-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3,4,5],ymm6[6,7]11386; AVX2-NEXT:    vshufps {{.*#+}} xmm7 = xmm0[2,2,2,2]11387; AVX2-NEXT:    vblendps {{.*#+}} xmm7 = xmm1[0,1,2],xmm7[3]11388; AVX2-NEXT:    vblendps {{.*#+}} xmm7 = xmm4[0,1],xmm7[2,3]11389; AVX2-NEXT:    vblendps {{.*#+}} ymm9 = ymm7[0,1,2,3],ymm6[4,5,6,7]11390; AVX2-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm1[2],xmm0[2],xmm1[3],xmm0[3]11391; AVX2-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm4[1],xmm0[1]11392; AVX2-NEXT:    vunpckhps {{.*#+}} xmm1 = xmm3[2],xmm2[2],xmm3[3],xmm2[3]11393; AVX2-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm111394; AVX2-NEXT:    vshufps {{.*#+}} xmm2 = xmm5[2,3,2,3]11395; AVX2-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm211396; AVX2-NEXT:    vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]11397; AVX2-NEXT:    vblendps {{.*#+}} ymm8 = ymm0[0,1,2,3],ymm1[4,5,6,7]11398; AVX2-NEXT:    vmovaps 224(%rcx), %xmm611399; AVX2-NEXT:    vbroadcastss %xmm6, %xmm011400; AVX2-NEXT:    vmovaps 224(%rdx), %xmm511401; AVX2-NEXT:    vbroadcastss %xmm5, %xmm111402; AVX2-NEXT:    vunpcklps {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]11403; AVX2-NEXT:    vmovaps 224(%rsi), %xmm411404; AVX2-NEXT:    vmovaps 224(%rdi), %xmm211405; AVX2-NEXT:    vunpcklps {{.*#+}} xmm3 = xmm2[0],xmm4[0],xmm2[1],xmm4[1]11406; AVX2-NEXT:    vblendps {{.*#+}} xmm7 = xmm3[0,1],xmm1[2,3]11407; AVX2-NEXT:    vmovaps 224(%rax), %xmm311408; AVX2-NEXT:    vmovaps 224(%r10), %xmm111409; AVX2-NEXT:    vbroadcastss %xmm3, %xmm1511410; AVX2-NEXT:    vbroadcastss %xmm1, %xmm1411411; AVX2-NEXT:    vunpcklps {{.*#+}} xmm14 = xmm14[0],xmm15[0],xmm14[1],xmm15[1]11412; AVX2-NEXT:    vmovaps 224(%r9), %xmm1511413; AVX2-NEXT:    vmovaps 224(%r8), %xmm011414; AVX2-NEXT:    vunpcklps {{.*#+}} xmm13 = xmm0[0],xmm15[0],xmm0[1],xmm15[1]11415; AVX2-NEXT:    vinsertf128 $1, %xmm13, %ymm0, %ymm1311416; AVX2-NEXT:    vinsertf128 $1, %xmm14, %ymm0, %ymm1411417; AVX2-NEXT:    vblendps {{.*#+}} ymm13 = ymm13[0,1,2,3,4,5],ymm14[6,7]11418; AVX2-NEXT:    vblendps {{.*#+}} ymm14 = ymm7[0,1,2,3],ymm13[4,5,6,7]11419; AVX2-NEXT:    vshufps {{.*#+}} xmm7 = xmm2[1,1,1,1]11420; AVX2-NEXT:    vblendps {{.*#+}} xmm7 = xmm7[0],xmm4[1],xmm7[2,3]11421; AVX2-NEXT:    vunpcklps {{.*#+}} xmm13 = xmm5[0],xmm6[0],xmm5[1],xmm6[1]11422; AVX2-NEXT:    vblendps {{.*#+}} xmm7 = xmm7[0,1],xmm13[2,3]11423; AVX2-NEXT:    vunpcklps {{.*#+}} xmm13 = xmm1[0],xmm3[0],xmm1[1],xmm3[1]11424; AVX2-NEXT:    vinsertf128 $1, %xmm13, %ymm0, %ymm1311425; AVX2-NEXT:    vshufps {{.*#+}} xmm12 = xmm0[1,1,1,1]11426; AVX2-NEXT:    vblendps {{.*#+}} xmm12 = xmm12[0],xmm15[1],xmm12[2,3]11427; AVX2-NEXT:    vinsertf128 $1, %xmm12, %ymm0, %ymm1211428; AVX2-NEXT:    vblendps {{.*#+}} ymm12 = ymm12[0,1,2,3,4,5],ymm13[6,7]11429; AVX2-NEXT:    vblendps {{.*#+}} ymm7 = ymm7[0,1,2,3],ymm12[4,5,6,7]11430; AVX2-NEXT:    vunpckhps {{.*#+}} xmm2 = xmm2[2],xmm4[2],xmm2[3],xmm4[3]11431; AVX2-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm0[2],xmm15[2],xmm0[3],xmm15[3]11432; AVX2-NEXT:    vshufps {{.*#+}} xmm4 = xmm3[2,2,2,2]11433; AVX2-NEXT:    vblendps {{.*#+}} xmm4 = xmm1[0,1,2],xmm4[3]11434; AVX2-NEXT:    vinsertf128 $1, %xmm4, %ymm0, %ymm411435; AVX2-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm1211436; AVX2-NEXT:    vblendps {{.*#+}} ymm4 = ymm12[0,1,2,3,4,5],ymm4[6,7]11437; AVX2-NEXT:    vshufps {{.*#+}} xmm12 = xmm6[2,2,2,2]11438; AVX2-NEXT:    vblendps {{.*#+}} xmm12 = xmm5[0,1,2],xmm12[3]11439; AVX2-NEXT:    vblendps {{.*#+}} xmm12 = xmm2[0,1],xmm12[2,3]11440; AVX2-NEXT:    vblendps {{.*#+}} ymm4 = ymm12[0,1,2,3],ymm4[4,5,6,7]11441; AVX2-NEXT:    vunpckhps {{.*#+}} xmm5 = xmm5[2],xmm6[2],xmm5[3],xmm6[3]11442; AVX2-NEXT:    vunpckhpd {{.*#+}} xmm2 = xmm2[1],xmm5[1]11443; AVX2-NEXT:    vunpckhps {{.*#+}} xmm1 = xmm1[2],xmm3[2],xmm1[3],xmm3[3]11444; AVX2-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm111445; AVX2-NEXT:    vshufps {{.*#+}} xmm0 = xmm0[2,3,2,3]11446; AVX2-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm011447; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]11448; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]11449; AVX2-NEXT:    movq {{[0-9]+}}(%rsp), %rax11450; AVX2-NEXT:    vmovaps %ymm0, 1888(%rax)11451; AVX2-NEXT:    vmovaps %ymm4, 1856(%rax)11452; AVX2-NEXT:    vmovaps %ymm7, 1824(%rax)11453; AVX2-NEXT:    vmovaps %ymm14, 1792(%rax)11454; AVX2-NEXT:    vmovaps %ymm8, 1632(%rax)11455; AVX2-NEXT:    vmovaps %ymm9, 1600(%rax)11456; AVX2-NEXT:    vmovaps %ymm10, 1568(%rax)11457; AVX2-NEXT:    vmovaps %ymm11, 1536(%rax)11458; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11459; AVX2-NEXT:    vmovaps %ymm0, 1376(%rax)11460; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11461; AVX2-NEXT:    vmovaps %ymm0, 1344(%rax)11462; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11463; AVX2-NEXT:    vmovaps %ymm0, 1312(%rax)11464; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11465; AVX2-NEXT:    vmovaps %ymm0, 1280(%rax)11466; AVX2-NEXT:    vmovups (%rsp), %ymm0 # 32-byte Reload11467; AVX2-NEXT:    vmovaps %ymm0, 1120(%rax)11468; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11469; AVX2-NEXT:    vmovaps %ymm0, 1088(%rax)11470; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11471; AVX2-NEXT:    vmovaps %ymm0, 1056(%rax)11472; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11473; AVX2-NEXT:    vmovaps %ymm0, 1024(%rax)11474; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11475; AVX2-NEXT:    vmovaps %ymm0, 864(%rax)11476; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11477; AVX2-NEXT:    vmovaps %ymm0, 832(%rax)11478; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11479; AVX2-NEXT:    vmovaps %ymm0, 800(%rax)11480; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11481; AVX2-NEXT:    vmovaps %ymm0, 768(%rax)11482; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11483; AVX2-NEXT:    vmovaps %ymm0, 608(%rax)11484; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11485; AVX2-NEXT:    vmovaps %ymm0, 576(%rax)11486; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11487; AVX2-NEXT:    vmovaps %ymm0, 544(%rax)11488; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11489; AVX2-NEXT:    vmovaps %ymm0, 512(%rax)11490; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11491; AVX2-NEXT:    vmovaps %ymm0, 352(%rax)11492; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11493; AVX2-NEXT:    vmovaps %ymm0, 320(%rax)11494; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11495; AVX2-NEXT:    vmovaps %ymm0, 288(%rax)11496; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11497; AVX2-NEXT:    vmovaps %ymm0, 256(%rax)11498; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11499; AVX2-NEXT:    vmovaps %ymm0, 96(%rax)11500; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11501; AVX2-NEXT:    vmovaps %ymm0, 64(%rax)11502; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11503; AVX2-NEXT:    vmovaps %ymm0, 32(%rax)11504; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11505; AVX2-NEXT:    vmovaps %ymm0, (%rax)11506; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11507; AVX2-NEXT:    vmovaps %ymm0, 2016(%rax)11508; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11509; AVX2-NEXT:    vmovaps %ymm0, 1984(%rax)11510; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11511; AVX2-NEXT:    vmovaps %ymm0, 1952(%rax)11512; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11513; AVX2-NEXT:    vmovaps %ymm0, 1920(%rax)11514; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11515; AVX2-NEXT:    vmovaps %ymm0, 1760(%rax)11516; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11517; AVX2-NEXT:    vmovaps %ymm0, 1728(%rax)11518; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11519; AVX2-NEXT:    vmovaps %ymm0, 1696(%rax)11520; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11521; AVX2-NEXT:    vmovaps %ymm0, 1664(%rax)11522; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11523; AVX2-NEXT:    vmovaps %ymm0, 1504(%rax)11524; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11525; AVX2-NEXT:    vmovaps %ymm0, 1472(%rax)11526; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11527; AVX2-NEXT:    vmovaps %ymm0, 1440(%rax)11528; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11529; AVX2-NEXT:    vmovaps %ymm0, 1408(%rax)11530; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11531; AVX2-NEXT:    vmovaps %ymm0, 1248(%rax)11532; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11533; AVX2-NEXT:    vmovaps %ymm0, 1216(%rax)11534; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11535; AVX2-NEXT:    vmovaps %ymm0, 1184(%rax)11536; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11537; AVX2-NEXT:    vmovaps %ymm0, 1152(%rax)11538; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11539; AVX2-NEXT:    vmovaps %ymm0, 992(%rax)11540; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11541; AVX2-NEXT:    vmovaps %ymm0, 960(%rax)11542; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11543; AVX2-NEXT:    vmovaps %ymm0, 928(%rax)11544; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11545; AVX2-NEXT:    vmovaps %ymm0, 896(%rax)11546; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11547; AVX2-NEXT:    vmovaps %ymm0, 736(%rax)11548; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11549; AVX2-NEXT:    vmovaps %ymm0, 704(%rax)11550; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11551; AVX2-NEXT:    vmovaps %ymm0, 672(%rax)11552; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11553; AVX2-NEXT:    vmovaps %ymm0, 640(%rax)11554; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11555; AVX2-NEXT:    vmovaps %ymm0, 480(%rax)11556; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11557; AVX2-NEXT:    vmovaps %ymm0, 448(%rax)11558; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11559; AVX2-NEXT:    vmovaps %ymm0, 416(%rax)11560; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11561; AVX2-NEXT:    vmovaps %ymm0, 384(%rax)11562; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11563; AVX2-NEXT:    vmovaps %ymm0, 224(%rax)11564; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11565; AVX2-NEXT:    vmovaps %ymm0, 192(%rax)11566; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11567; AVX2-NEXT:    vmovaps %ymm0, 160(%rax)11568; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload11569; AVX2-NEXT:    vmovaps %ymm0, 128(%rax)11570; AVX2-NEXT:    addq $1672, %rsp # imm = 0x68811571; AVX2-NEXT:    vzeroupper11572; AVX2-NEXT:    retq11573;11574; AVX2-FP-LABEL: store_i32_stride8_vf64:11575; AVX2-FP:       # %bb.0:11576; AVX2-FP-NEXT:    subq $1672, %rsp # imm = 0x68811577; AVX2-FP-NEXT:    movq {{[0-9]+}}(%rsp), %rax11578; AVX2-FP-NEXT:    movq {{[0-9]+}}(%rsp), %r1011579; AVX2-FP-NEXT:    vmovaps 32(%rdi), %ymm011580; AVX2-FP-NEXT:    vmovaps (%rdi), %ymm411581; AVX2-FP-NEXT:    vmovaps 32(%rsi), %ymm111582; AVX2-FP-NEXT:    vmovaps (%rsi), %ymm511583; AVX2-FP-NEXT:    vmovaps 32(%rdx), %ymm211584; AVX2-FP-NEXT:    vmovaps (%rdx), %ymm611585; AVX2-FP-NEXT:    vmovaps 32(%rcx), %ymm311586; AVX2-FP-NEXT:    vmovaps (%rcx), %ymm711587; AVX2-FP-NEXT:    vmovaps (%r8), %ymm911588; AVX2-FP-NEXT:    vmovaps (%r9), %ymm1111589; AVX2-FP-NEXT:    vmovaps (%r10), %ymm1011590; AVX2-FP-NEXT:    vmovaps (%rax), %ymm1211591; AVX2-FP-NEXT:    vunpcklps {{.*#+}} ymm8 = ymm10[0],ymm12[0],ymm10[1],ymm12[1],ymm10[4],ymm12[4],ymm10[5],ymm12[5]11592; AVX2-FP-NEXT:    vunpcklps {{.*#+}} ymm13 = ymm4[0],ymm5[0],ymm4[1],ymm5[1],ymm4[4],ymm5[4],ymm4[5],ymm5[5]11593; AVX2-FP-NEXT:    vextractf128 $1, %ymm13, %xmm1311594; AVX2-FP-NEXT:    vunpcklps {{.*#+}} ymm14 = ymm6[0],ymm7[0],ymm6[1],ymm7[1],ymm6[4],ymm7[4],ymm6[5],ymm7[5]11595; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm15 = ymm14[2,2,2,2]11596; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm13 = xmm13[0,1],xmm15[2,3]11597; AVX2-FP-NEXT:    vunpcklps {{.*#+}} ymm15 = ymm9[0],ymm11[0],ymm9[1],ymm11[1],ymm9[4],ymm11[4],ymm9[5],ymm11[5]11598; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm15 = ymm15[0],ymm8[0],ymm15[2],ymm8[2]11599; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm13 = ymm13[0,1,2,3],ymm15[4,5,6,7]11600; AVX2-FP-NEXT:    vmovups %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11601; AVX2-FP-NEXT:    vbroadcastss 20(%r8), %ymm1311602; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm13 = ymm13[0,1,2,3,4],ymm11[5],ymm13[6,7]11603; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm13 = ymm13[0,1,2,3,4,5],ymm8[6,7]11604; AVX2-FP-NEXT:    vextractf128 $1, %ymm14, %xmm811605; AVX2-FP-NEXT:    vshufps {{.*#+}} ymm14 = ymm4[1,1,1,1,5,5,5,5]11606; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm14 = ymm14[0],ymm5[1],ymm14[2,3,4],ymm5[5],ymm14[6,7]11607; AVX2-FP-NEXT:    vextractf128 $1, %ymm14, %xmm1411608; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm14 = xmm14[0,1],xmm8[2,3]11609; AVX2-FP-NEXT:    vmovaps 32(%r8), %ymm811610; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm13 = ymm14[0,1,2,3],ymm13[4,5,6,7]11611; AVX2-FP-NEXT:    vmovups %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11612; AVX2-FP-NEXT:    vbroadcastss 24(%rax), %ymm1311613; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm13 = ymm10[0,1,2,3,4,5,6],ymm13[7]11614; AVX2-FP-NEXT:    vmovaps 32(%r9), %ymm1011615; AVX2-FP-NEXT:    vunpckhps {{.*#+}} ymm11 = ymm9[2],ymm11[2],ymm9[3],ymm11[3],ymm9[6],ymm11[6],ymm9[7],ymm11[7]11616; AVX2-FP-NEXT:    vmovaps 32(%r10), %ymm911617; AVX2-FP-NEXT:    vunpckhps {{.*#+}} ymm4 = ymm4[2],ymm5[2],ymm4[3],ymm5[3],ymm4[6],ymm5[6],ymm4[7],ymm5[7]11618; AVX2-FP-NEXT:    vextractf128 $1, %ymm4, %xmm511619; AVX2-FP-NEXT:    vunpckhps {{.*#+}} ymm6 = ymm6[2],ymm7[2],ymm6[3],ymm7[3],ymm6[6],ymm7[6],ymm6[7],ymm7[7]11620; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm7 = ymm6[2,2,2,2]11621; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm5 = xmm5[0,1],xmm7[2,3]11622; AVX2-FP-NEXT:    vmovaps 32(%rax), %ymm711623; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm13 = ymm11[0,1,2,3,4,5],ymm13[6,7]11624; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm13[4,5,6,7]11625; AVX2-FP-NEXT:    vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11626; AVX2-FP-NEXT:    vbroadcastss 28(%r10), %ymm511627; AVX2-FP-NEXT:    vunpckhps {{.*#+}} ymm5 = ymm5[2],ymm12[2],ymm5[3],ymm12[3],ymm5[6],ymm12[6],ymm5[7],ymm12[7]11628; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm5 = ymm11[1],ymm5[1],ymm11[3],ymm5[3]11629; AVX2-FP-NEXT:    vextractf128 $1, %ymm6, %xmm611630; AVX2-FP-NEXT:    vshufps {{.*#+}} ymm4 = ymm4[2,3,2,3,6,7,6,7]11631; AVX2-FP-NEXT:    vextractf128 $1, %ymm4, %xmm411632; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm4 = xmm4[0,1],xmm6[2,3]11633; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm5[4,5,6,7]11634; AVX2-FP-NEXT:    vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11635; AVX2-FP-NEXT:    vunpcklps {{.*#+}} ymm4 = ymm9[0],ymm7[0],ymm9[1],ymm7[1],ymm9[4],ymm7[4],ymm9[5],ymm7[5]11636; AVX2-FP-NEXT:    vunpcklps {{.*#+}} ymm5 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]11637; AVX2-FP-NEXT:    vextractf128 $1, %ymm5, %xmm511638; AVX2-FP-NEXT:    vunpcklps {{.*#+}} ymm6 = ymm2[0],ymm3[0],ymm2[1],ymm3[1],ymm2[4],ymm3[4],ymm2[5],ymm3[5]11639; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm11 = ymm6[2,2,2,2]11640; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm5 = xmm5[0,1],xmm11[2,3]11641; AVX2-FP-NEXT:    vunpcklps {{.*#+}} ymm11 = ymm8[0],ymm10[0],ymm8[1],ymm10[1],ymm8[4],ymm10[4],ymm8[5],ymm10[5]11642; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm11 = ymm11[0],ymm4[0],ymm11[2],ymm4[2]11643; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm11[4,5,6,7]11644; AVX2-FP-NEXT:    vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11645; AVX2-FP-NEXT:    vbroadcastss 52(%r8), %ymm511646; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3,4],ymm10[5],ymm5[6,7]11647; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3,4,5],ymm4[6,7]11648; AVX2-FP-NEXT:    vextractf128 $1, %ymm6, %xmm411649; AVX2-FP-NEXT:    vshufps {{.*#+}} ymm6 = ymm0[1,1,1,1,5,5,5,5]11650; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm6 = ymm6[0],ymm1[1],ymm6[2,3,4],ymm1[5],ymm6[6,7]11651; AVX2-FP-NEXT:    vextractf128 $1, %ymm6, %xmm611652; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm6 = xmm6[0,1],xmm4[2,3]11653; AVX2-FP-NEXT:    vmovaps 64(%rdi), %ymm411654; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm5 = ymm6[0,1,2,3],ymm5[4,5,6,7]11655; AVX2-FP-NEXT:    vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11656; AVX2-FP-NEXT:    vbroadcastss 56(%rax), %ymm511657; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm9 = ymm9[0,1,2,3,4,5,6],ymm5[7]11658; AVX2-FP-NEXT:    vmovaps 64(%rsi), %ymm511659; AVX2-FP-NEXT:    vunpckhps {{.*#+}} ymm8 = ymm8[2],ymm10[2],ymm8[3],ymm10[3],ymm8[6],ymm10[6],ymm8[7],ymm10[7]11660; AVX2-FP-NEXT:    vmovaps 64(%rdx), %ymm611661; AVX2-FP-NEXT:    vunpckhps {{.*#+}} ymm10 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7]11662; AVX2-FP-NEXT:    vextractf128 $1, %ymm10, %xmm011663; AVX2-FP-NEXT:    vunpckhps {{.*#+}} ymm2 = ymm2[2],ymm3[2],ymm2[3],ymm3[3],ymm2[6],ymm3[6],ymm2[7],ymm3[7]11664; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm1 = ymm2[2,2,2,2]11665; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm1 = xmm0[0,1],xmm1[2,3]11666; AVX2-FP-NEXT:    vmovaps 64(%rcx), %ymm011667; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm3 = ymm8[0,1,2,3,4,5],ymm9[6,7]11668; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm3[4,5,6,7]11669; AVX2-FP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11670; AVX2-FP-NEXT:    vbroadcastss 60(%r10), %ymm111671; AVX2-FP-NEXT:    vunpckhps {{.*#+}} ymm3 = ymm1[2],ymm7[2],ymm1[3],ymm7[3],ymm1[6],ymm7[6],ymm1[7],ymm7[7]11672; AVX2-FP-NEXT:    vmovaps 64(%r8), %ymm111673; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm3 = ymm8[1],ymm3[1],ymm8[3],ymm3[3]11674; AVX2-FP-NEXT:    vextractf128 $1, %ymm2, %xmm211675; AVX2-FP-NEXT:    vshufps {{.*#+}} ymm7 = ymm10[2,3,2,3,6,7,6,7]11676; AVX2-FP-NEXT:    vextractf128 $1, %ymm7, %xmm711677; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm2 = xmm7[0,1],xmm2[2,3]11678; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]11679; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11680; AVX2-FP-NEXT:    vunpcklps {{.*#+}} ymm2 = ymm4[0],ymm5[0],ymm4[1],ymm5[1],ymm4[4],ymm5[4],ymm4[5],ymm5[5]11681; AVX2-FP-NEXT:    vextractf128 $1, %ymm2, %xmm211682; AVX2-FP-NEXT:    vunpcklps {{.*#+}} ymm3 = ymm6[0],ymm0[0],ymm6[1],ymm0[1],ymm6[4],ymm0[4],ymm6[5],ymm0[5]11683; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm7 = ymm3[2,2,2,2]11684; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm2 = xmm2[0,1],xmm7[2,3]11685; AVX2-FP-NEXT:    vmovaps 64(%r9), %ymm711686; AVX2-FP-NEXT:    vmovaps 64(%r10), %ymm811687; AVX2-FP-NEXT:    vmovaps 64(%rax), %ymm911688; AVX2-FP-NEXT:    vunpcklps {{.*#+}} ymm10 = ymm8[0],ymm9[0],ymm8[1],ymm9[1],ymm8[4],ymm9[4],ymm8[5],ymm9[5]11689; AVX2-FP-NEXT:    vunpcklps {{.*#+}} ymm11 = ymm1[0],ymm7[0],ymm1[1],ymm7[1],ymm1[4],ymm7[4],ymm1[5],ymm7[5]11690; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm11 = ymm11[0],ymm10[0],ymm11[2],ymm10[2]11691; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm11[4,5,6,7]11692; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11693; AVX2-FP-NEXT:    vbroadcastss 84(%r8), %ymm211694; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4],ymm7[5],ymm2[6,7]11695; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm10[6,7]11696; AVX2-FP-NEXT:    vextractf128 $1, %ymm3, %xmm311697; AVX2-FP-NEXT:    vshufps {{.*#+}} ymm10 = ymm4[1,1,1,1,5,5,5,5]11698; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm10 = ymm10[0],ymm5[1],ymm10[2,3,4],ymm5[5],ymm10[6,7]11699; AVX2-FP-NEXT:    vextractf128 $1, %ymm10, %xmm1011700; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm3 = xmm10[0,1],xmm3[2,3]11701; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm2 = ymm3[0,1,2,3],ymm2[4,5,6,7]11702; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11703; AVX2-FP-NEXT:    vbroadcastss 88(%rax), %ymm211704; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm2 = ymm8[0,1,2,3,4,5,6],ymm2[7]11705; AVX2-FP-NEXT:    vunpckhps {{.*#+}} ymm1 = ymm1[2],ymm7[2],ymm1[3],ymm7[3],ymm1[6],ymm7[6],ymm1[7],ymm7[7]11706; AVX2-FP-NEXT:    vunpckhps {{.*#+}} ymm3 = ymm4[2],ymm5[2],ymm4[3],ymm5[3],ymm4[6],ymm5[6],ymm4[7],ymm5[7]11707; AVX2-FP-NEXT:    vunpckhps {{.*#+}} ymm4 = ymm6[2],ymm0[2],ymm6[3],ymm0[3],ymm6[6],ymm0[6],ymm6[7],ymm0[7]11708; AVX2-FP-NEXT:    vextractf128 $1, %ymm3, %xmm011709; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm5 = ymm4[2,2,2,2]11710; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm5[2,3]11711; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm2 = ymm1[0,1,2,3,4,5],ymm2[6,7]11712; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]11713; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11714; AVX2-FP-NEXT:    vbroadcastss 92(%r10), %ymm011715; AVX2-FP-NEXT:    vunpckhps {{.*#+}} ymm2 = ymm0[2],ymm9[2],ymm0[3],ymm9[3],ymm0[6],ymm9[6],ymm0[7],ymm9[7]11716; AVX2-FP-NEXT:    vmovaps 96(%rdi), %ymm011717; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm5 = ymm1[1],ymm2[1],ymm1[3],ymm2[3]11718; AVX2-FP-NEXT:    vmovaps 96(%rsi), %ymm111719; AVX2-FP-NEXT:    vextractf128 $1, %ymm4, %xmm211720; AVX2-FP-NEXT:    vshufps {{.*#+}} ymm3 = ymm3[2,3,2,3,6,7,6,7]11721; AVX2-FP-NEXT:    vextractf128 $1, %ymm3, %xmm311722; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm3 = xmm3[0,1],xmm2[2,3]11723; AVX2-FP-NEXT:    vmovaps 96(%rdx), %ymm211724; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm5[4,5,6,7]11725; AVX2-FP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11726; AVX2-FP-NEXT:    vunpcklps {{.*#+}} ymm3 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]11727; AVX2-FP-NEXT:    vextractf128 $1, %ymm3, %xmm311728; AVX2-FP-NEXT:    vmovaps 96(%rcx), %ymm411729; AVX2-FP-NEXT:    vunpcklps {{.*#+}} ymm5 = ymm2[0],ymm4[0],ymm2[1],ymm4[1],ymm2[4],ymm4[4],ymm2[5],ymm4[5]11730; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm6 = ymm5[2,2,2,2]11731; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm3 = xmm3[0,1],xmm6[2,3]11732; AVX2-FP-NEXT:    vmovaps 96(%r8), %ymm611733; AVX2-FP-NEXT:    vmovaps 96(%r9), %ymm711734; AVX2-FP-NEXT:    vmovaps 96(%r10), %ymm811735; AVX2-FP-NEXT:    vmovaps 96(%rax), %ymm911736; AVX2-FP-NEXT:    vunpcklps {{.*#+}} ymm10 = ymm8[0],ymm9[0],ymm8[1],ymm9[1],ymm8[4],ymm9[4],ymm8[5],ymm9[5]11737; AVX2-FP-NEXT:    vunpcklps {{.*#+}} ymm11 = ymm6[0],ymm7[0],ymm6[1],ymm7[1],ymm6[4],ymm7[4],ymm6[5],ymm7[5]11738; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm11 = ymm11[0],ymm10[0],ymm11[2],ymm10[2]11739; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm11[4,5,6,7]11740; AVX2-FP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11741; AVX2-FP-NEXT:    vbroadcastss 116(%r8), %ymm311742; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4],ymm7[5],ymm3[6,7]11743; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm10[6,7]11744; AVX2-FP-NEXT:    vextractf128 $1, %ymm5, %xmm511745; AVX2-FP-NEXT:    vshufps {{.*#+}} ymm10 = ymm0[1,1,1,1,5,5,5,5]11746; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm10 = ymm10[0],ymm1[1],ymm10[2,3,4],ymm1[5],ymm10[6,7]11747; AVX2-FP-NEXT:    vextractf128 $1, %ymm10, %xmm1011748; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm5 = xmm10[0,1],xmm5[2,3]11749; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm3 = ymm5[0,1,2,3],ymm3[4,5,6,7]11750; AVX2-FP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11751; AVX2-FP-NEXT:    vbroadcastss 120(%rax), %ymm311752; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm3 = ymm8[0,1,2,3,4,5,6],ymm3[7]11753; AVX2-FP-NEXT:    vunpckhps {{.*#+}} ymm5 = ymm6[2],ymm7[2],ymm6[3],ymm7[3],ymm6[6],ymm7[6],ymm6[7],ymm7[7]11754; AVX2-FP-NEXT:    vunpckhps {{.*#+}} ymm0 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7]11755; AVX2-FP-NEXT:    vunpckhps {{.*#+}} ymm1 = ymm2[2],ymm4[2],ymm2[3],ymm4[3],ymm2[6],ymm4[6],ymm2[7],ymm4[7]11756; AVX2-FP-NEXT:    vextractf128 $1, %ymm0, %xmm211757; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm4 = ymm1[2,2,2,2]11758; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm2 = xmm2[0,1],xmm4[2,3]11759; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm3 = ymm5[0,1,2,3,4,5],ymm3[6,7]11760; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]11761; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11762; AVX2-FP-NEXT:    vbroadcastss 124(%r10), %ymm211763; AVX2-FP-NEXT:    vunpckhps {{.*#+}} ymm2 = ymm2[2],ymm9[2],ymm2[3],ymm9[3],ymm2[6],ymm9[6],ymm2[7],ymm9[7]11764; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm5[1],ymm2[1],ymm5[3],ymm2[3]11765; AVX2-FP-NEXT:    vextractf128 $1, %ymm1, %xmm111766; AVX2-FP-NEXT:    vshufps {{.*#+}} ymm0 = ymm0[2,3,2,3,6,7,6,7]11767; AVX2-FP-NEXT:    vextractf128 $1, %ymm0, %xmm011768; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]11769; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]11770; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11771; AVX2-FP-NEXT:    vmovaps 128(%rdi), %ymm011772; AVX2-FP-NEXT:    vmovaps 128(%rsi), %ymm111773; AVX2-FP-NEXT:    vunpcklps {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]11774; AVX2-FP-NEXT:    vextractf128 $1, %ymm2, %xmm211775; AVX2-FP-NEXT:    vmovaps 128(%rdx), %ymm311776; AVX2-FP-NEXT:    vmovaps 128(%rcx), %ymm411777; AVX2-FP-NEXT:    vunpcklps {{.*#+}} ymm5 = ymm3[0],ymm4[0],ymm3[1],ymm4[1],ymm3[4],ymm4[4],ymm3[5],ymm4[5]11778; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm6 = ymm5[2,2,2,2]11779; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm2 = xmm2[0,1],xmm6[2,3]11780; AVX2-FP-NEXT:    vmovaps 128(%r8), %ymm611781; AVX2-FP-NEXT:    vmovaps 128(%r9), %ymm711782; AVX2-FP-NEXT:    vmovaps 128(%r10), %ymm811783; AVX2-FP-NEXT:    vmovaps 128(%rax), %ymm911784; AVX2-FP-NEXT:    vunpcklps {{.*#+}} ymm10 = ymm8[0],ymm9[0],ymm8[1],ymm9[1],ymm8[4],ymm9[4],ymm8[5],ymm9[5]11785; AVX2-FP-NEXT:    vunpcklps {{.*#+}} ymm11 = ymm6[0],ymm7[0],ymm6[1],ymm7[1],ymm6[4],ymm7[4],ymm6[5],ymm7[5]11786; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm11 = ymm11[0],ymm10[0],ymm11[2],ymm10[2]11787; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm11[4,5,6,7]11788; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11789; AVX2-FP-NEXT:    vbroadcastss 148(%r8), %ymm211790; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4],ymm7[5],ymm2[6,7]11791; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm10[6,7]11792; AVX2-FP-NEXT:    vextractf128 $1, %ymm5, %xmm511793; AVX2-FP-NEXT:    vshufps {{.*#+}} ymm10 = ymm0[1,1,1,1,5,5,5,5]11794; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm10 = ymm10[0],ymm1[1],ymm10[2,3,4],ymm1[5],ymm10[6,7]11795; AVX2-FP-NEXT:    vextractf128 $1, %ymm10, %xmm1011796; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm5 = xmm10[0,1],xmm5[2,3]11797; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm2 = ymm5[0,1,2,3],ymm2[4,5,6,7]11798; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11799; AVX2-FP-NEXT:    vbroadcastss 152(%rax), %ymm211800; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm2 = ymm8[0,1,2,3,4,5,6],ymm2[7]11801; AVX2-FP-NEXT:    vunpckhps {{.*#+}} ymm5 = ymm6[2],ymm7[2],ymm6[3],ymm7[3],ymm6[6],ymm7[6],ymm6[7],ymm7[7]11802; AVX2-FP-NEXT:    vunpckhps {{.*#+}} ymm0 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7]11803; AVX2-FP-NEXT:    vunpckhps {{.*#+}} ymm1 = ymm3[2],ymm4[2],ymm3[3],ymm4[3],ymm3[6],ymm4[6],ymm3[7],ymm4[7]11804; AVX2-FP-NEXT:    vextractf128 $1, %ymm0, %xmm311805; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm4 = ymm1[2,2,2,2]11806; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm3 = xmm3[0,1],xmm4[2,3]11807; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm2 = ymm5[0,1,2,3,4,5],ymm2[6,7]11808; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm2 = ymm3[0,1,2,3],ymm2[4,5,6,7]11809; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11810; AVX2-FP-NEXT:    vbroadcastss 156(%r10), %ymm211811; AVX2-FP-NEXT:    vunpckhps {{.*#+}} ymm2 = ymm2[2],ymm9[2],ymm2[3],ymm9[3],ymm2[6],ymm9[6],ymm2[7],ymm9[7]11812; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm5[1],ymm2[1],ymm5[3],ymm2[3]11813; AVX2-FP-NEXT:    vextractf128 $1, %ymm1, %xmm111814; AVX2-FP-NEXT:    vshufps {{.*#+}} ymm0 = ymm0[2,3,2,3,6,7,6,7]11815; AVX2-FP-NEXT:    vextractf128 $1, %ymm0, %xmm011816; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]11817; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]11818; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11819; AVX2-FP-NEXT:    vmovaps 160(%rdi), %ymm011820; AVX2-FP-NEXT:    vmovaps 160(%rsi), %ymm111821; AVX2-FP-NEXT:    vunpcklps {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]11822; AVX2-FP-NEXT:    vextractf128 $1, %ymm2, %xmm211823; AVX2-FP-NEXT:    vmovaps 160(%rdx), %ymm311824; AVX2-FP-NEXT:    vmovaps 160(%rcx), %ymm411825; AVX2-FP-NEXT:    vunpcklps {{.*#+}} ymm5 = ymm3[0],ymm4[0],ymm3[1],ymm4[1],ymm3[4],ymm4[4],ymm3[5],ymm4[5]11826; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm6 = ymm5[2,2,2,2]11827; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm2 = xmm2[0,1],xmm6[2,3]11828; AVX2-FP-NEXT:    vmovaps 160(%r8), %ymm611829; AVX2-FP-NEXT:    vmovaps 160(%r9), %ymm711830; AVX2-FP-NEXT:    vmovaps 160(%r10), %ymm811831; AVX2-FP-NEXT:    vmovaps 160(%rax), %ymm911832; AVX2-FP-NEXT:    vunpcklps {{.*#+}} ymm10 = ymm8[0],ymm9[0],ymm8[1],ymm9[1],ymm8[4],ymm9[4],ymm8[5],ymm9[5]11833; AVX2-FP-NEXT:    vunpcklps {{.*#+}} ymm11 = ymm6[0],ymm7[0],ymm6[1],ymm7[1],ymm6[4],ymm7[4],ymm6[5],ymm7[5]11834; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm11 = ymm11[0],ymm10[0],ymm11[2],ymm10[2]11835; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm11[4,5,6,7]11836; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11837; AVX2-FP-NEXT:    vbroadcastss 180(%r8), %ymm211838; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4],ymm7[5],ymm2[6,7]11839; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm10[6,7]11840; AVX2-FP-NEXT:    vextractf128 $1, %ymm5, %xmm511841; AVX2-FP-NEXT:    vshufps {{.*#+}} ymm10 = ymm0[1,1,1,1,5,5,5,5]11842; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm10 = ymm10[0],ymm1[1],ymm10[2,3,4],ymm1[5],ymm10[6,7]11843; AVX2-FP-NEXT:    vextractf128 $1, %ymm10, %xmm1011844; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm5 = xmm10[0,1],xmm5[2,3]11845; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm2 = ymm5[0,1,2,3],ymm2[4,5,6,7]11846; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11847; AVX2-FP-NEXT:    vbroadcastss 184(%rax), %ymm211848; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm2 = ymm8[0,1,2,3,4,5,6],ymm2[7]11849; AVX2-FP-NEXT:    vunpckhps {{.*#+}} ymm5 = ymm6[2],ymm7[2],ymm6[3],ymm7[3],ymm6[6],ymm7[6],ymm6[7],ymm7[7]11850; AVX2-FP-NEXT:    vunpckhps {{.*#+}} ymm0 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7]11851; AVX2-FP-NEXT:    vunpckhps {{.*#+}} ymm1 = ymm3[2],ymm4[2],ymm3[3],ymm4[3],ymm3[6],ymm4[6],ymm3[7],ymm4[7]11852; AVX2-FP-NEXT:    vextractf128 $1, %ymm0, %xmm311853; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm4 = ymm1[2,2,2,2]11854; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm3 = xmm3[0,1],xmm4[2,3]11855; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm2 = ymm5[0,1,2,3,4,5],ymm2[6,7]11856; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm2 = ymm3[0,1,2,3],ymm2[4,5,6,7]11857; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11858; AVX2-FP-NEXT:    vbroadcastss 188(%r10), %ymm211859; AVX2-FP-NEXT:    vunpckhps {{.*#+}} ymm2 = ymm2[2],ymm9[2],ymm2[3],ymm9[3],ymm2[6],ymm9[6],ymm2[7],ymm9[7]11860; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm5[1],ymm2[1],ymm5[3],ymm2[3]11861; AVX2-FP-NEXT:    vextractf128 $1, %ymm1, %xmm111862; AVX2-FP-NEXT:    vshufps {{.*#+}} ymm0 = ymm0[2,3,2,3,6,7,6,7]11863; AVX2-FP-NEXT:    vextractf128 $1, %ymm0, %xmm011864; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]11865; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]11866; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11867; AVX2-FP-NEXT:    vmovaps 192(%rdi), %ymm011868; AVX2-FP-NEXT:    vmovaps 192(%rsi), %ymm111869; AVX2-FP-NEXT:    vunpcklps {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]11870; AVX2-FP-NEXT:    vextractf128 $1, %ymm2, %xmm211871; AVX2-FP-NEXT:    vmovaps 192(%rdx), %ymm311872; AVX2-FP-NEXT:    vmovaps 192(%rcx), %ymm411873; AVX2-FP-NEXT:    vunpcklps {{.*#+}} ymm5 = ymm3[0],ymm4[0],ymm3[1],ymm4[1],ymm3[4],ymm4[4],ymm3[5],ymm4[5]11874; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm6 = ymm5[2,2,2,2]11875; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm2 = xmm2[0,1],xmm6[2,3]11876; AVX2-FP-NEXT:    vmovaps 192(%r8), %ymm611877; AVX2-FP-NEXT:    vmovaps 192(%r9), %ymm711878; AVX2-FP-NEXT:    vmovaps 192(%r10), %ymm811879; AVX2-FP-NEXT:    vmovaps 192(%rax), %ymm911880; AVX2-FP-NEXT:    vunpcklps {{.*#+}} ymm10 = ymm8[0],ymm9[0],ymm8[1],ymm9[1],ymm8[4],ymm9[4],ymm8[5],ymm9[5]11881; AVX2-FP-NEXT:    vunpcklps {{.*#+}} ymm11 = ymm6[0],ymm7[0],ymm6[1],ymm7[1],ymm6[4],ymm7[4],ymm6[5],ymm7[5]11882; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm11 = ymm11[0],ymm10[0],ymm11[2],ymm10[2]11883; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm11[4,5,6,7]11884; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11885; AVX2-FP-NEXT:    vbroadcastss 212(%r8), %ymm211886; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4],ymm7[5],ymm2[6,7]11887; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm10[6,7]11888; AVX2-FP-NEXT:    vextractf128 $1, %ymm5, %xmm511889; AVX2-FP-NEXT:    vshufps {{.*#+}} ymm10 = ymm0[1,1,1,1,5,5,5,5]11890; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm10 = ymm10[0],ymm1[1],ymm10[2,3,4],ymm1[5],ymm10[6,7]11891; AVX2-FP-NEXT:    vextractf128 $1, %ymm10, %xmm1011892; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm5 = xmm10[0,1],xmm5[2,3]11893; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm2 = ymm5[0,1,2,3],ymm2[4,5,6,7]11894; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11895; AVX2-FP-NEXT:    vbroadcastss 216(%rax), %ymm211896; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm2 = ymm8[0,1,2,3,4,5,6],ymm2[7]11897; AVX2-FP-NEXT:    vunpckhps {{.*#+}} ymm5 = ymm6[2],ymm7[2],ymm6[3],ymm7[3],ymm6[6],ymm7[6],ymm6[7],ymm7[7]11898; AVX2-FP-NEXT:    vunpckhps {{.*#+}} ymm0 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7]11899; AVX2-FP-NEXT:    vunpckhps {{.*#+}} ymm1 = ymm3[2],ymm4[2],ymm3[3],ymm4[3],ymm3[6],ymm4[6],ymm3[7],ymm4[7]11900; AVX2-FP-NEXT:    vextractf128 $1, %ymm0, %xmm311901; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm4 = ymm1[2,2,2,2]11902; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm3 = xmm3[0,1],xmm4[2,3]11903; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm2 = ymm5[0,1,2,3,4,5],ymm2[6,7]11904; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm2 = ymm3[0,1,2,3],ymm2[4,5,6,7]11905; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11906; AVX2-FP-NEXT:    vbroadcastss 220(%r10), %ymm211907; AVX2-FP-NEXT:    vunpckhps {{.*#+}} ymm2 = ymm2[2],ymm9[2],ymm2[3],ymm9[3],ymm2[6],ymm9[6],ymm2[7],ymm9[7]11908; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm5[1],ymm2[1],ymm5[3],ymm2[3]11909; AVX2-FP-NEXT:    vextractf128 $1, %ymm1, %xmm111910; AVX2-FP-NEXT:    vshufps {{.*#+}} ymm0 = ymm0[2,3,2,3,6,7,6,7]11911; AVX2-FP-NEXT:    vextractf128 $1, %ymm0, %xmm011912; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]11913; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]11914; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11915; AVX2-FP-NEXT:    vmovaps 224(%rdi), %ymm011916; AVX2-FP-NEXT:    vmovaps 224(%rsi), %ymm111917; AVX2-FP-NEXT:    vunpcklps {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]11918; AVX2-FP-NEXT:    vextractf128 $1, %ymm2, %xmm211919; AVX2-FP-NEXT:    vmovaps 224(%rdx), %ymm311920; AVX2-FP-NEXT:    vmovaps 224(%rcx), %ymm411921; AVX2-FP-NEXT:    vunpcklps {{.*#+}} ymm5 = ymm3[0],ymm4[0],ymm3[1],ymm4[1],ymm3[4],ymm4[4],ymm3[5],ymm4[5]11922; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm6 = ymm5[2,2,2,2]11923; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm2 = xmm2[0,1],xmm6[2,3]11924; AVX2-FP-NEXT:    vmovaps 224(%r8), %ymm611925; AVX2-FP-NEXT:    vmovaps 224(%r9), %ymm711926; AVX2-FP-NEXT:    vmovaps 224(%r10), %ymm811927; AVX2-FP-NEXT:    vmovaps 224(%rax), %ymm911928; AVX2-FP-NEXT:    vunpcklps {{.*#+}} ymm10 = ymm8[0],ymm9[0],ymm8[1],ymm9[1],ymm8[4],ymm9[4],ymm8[5],ymm9[5]11929; AVX2-FP-NEXT:    vunpcklps {{.*#+}} ymm11 = ymm6[0],ymm7[0],ymm6[1],ymm7[1],ymm6[4],ymm7[4],ymm6[5],ymm7[5]11930; AVX2-FP-NEXT:    vunpcklpd {{.*#+}} ymm11 = ymm11[0],ymm10[0],ymm11[2],ymm10[2]11931; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm11[4,5,6,7]11932; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11933; AVX2-FP-NEXT:    vbroadcastss 244(%r8), %ymm211934; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4],ymm7[5],ymm2[6,7]11935; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm10[6,7]11936; AVX2-FP-NEXT:    vextractf128 $1, %ymm5, %xmm511937; AVX2-FP-NEXT:    vshufps {{.*#+}} ymm10 = ymm0[1,1,1,1,5,5,5,5]11938; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm10 = ymm10[0],ymm1[1],ymm10[2,3,4],ymm1[5],ymm10[6,7]11939; AVX2-FP-NEXT:    vextractf128 $1, %ymm10, %xmm1011940; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm5 = xmm10[0,1],xmm5[2,3]11941; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm2 = ymm5[0,1,2,3],ymm2[4,5,6,7]11942; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11943; AVX2-FP-NEXT:    vbroadcastss 248(%rax), %ymm211944; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm2 = ymm8[0,1,2,3,4,5,6],ymm2[7]11945; AVX2-FP-NEXT:    vunpckhps {{.*#+}} ymm5 = ymm6[2],ymm7[2],ymm6[3],ymm7[3],ymm6[6],ymm7[6],ymm6[7],ymm7[7]11946; AVX2-FP-NEXT:    vunpckhps {{.*#+}} ymm0 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7]11947; AVX2-FP-NEXT:    vunpckhps {{.*#+}} ymm1 = ymm3[2],ymm4[2],ymm3[3],ymm4[3],ymm3[6],ymm4[6],ymm3[7],ymm4[7]11948; AVX2-FP-NEXT:    vextractf128 $1, %ymm0, %xmm311949; AVX2-FP-NEXT:    vpermpd {{.*#+}} ymm4 = ymm1[2,2,2,2]11950; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm3 = xmm3[0,1],xmm4[2,3]11951; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm2 = ymm5[0,1,2,3,4,5],ymm2[6,7]11952; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm2 = ymm3[0,1,2,3],ymm2[4,5,6,7]11953; AVX2-FP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11954; AVX2-FP-NEXT:    vbroadcastss 252(%r10), %ymm211955; AVX2-FP-NEXT:    vunpckhps {{.*#+}} ymm2 = ymm2[2],ymm9[2],ymm2[3],ymm9[3],ymm2[6],ymm9[6],ymm2[7],ymm9[7]11956; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm5[1],ymm2[1],ymm5[3],ymm2[3]11957; AVX2-FP-NEXT:    vextractf128 $1, %ymm1, %xmm111958; AVX2-FP-NEXT:    vshufps {{.*#+}} ymm0 = ymm0[2,3,2,3,6,7,6,7]11959; AVX2-FP-NEXT:    vextractf128 $1, %ymm0, %xmm011960; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]11961; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]11962; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11963; AVX2-FP-NEXT:    vmovaps (%rcx), %xmm011964; AVX2-FP-NEXT:    vbroadcastss %xmm0, %xmm211965; AVX2-FP-NEXT:    vmovaps (%rdx), %xmm111966; AVX2-FP-NEXT:    vbroadcastss %xmm1, %xmm311967; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]11968; AVX2-FP-NEXT:    vmovaps (%rsi), %xmm411969; AVX2-FP-NEXT:    vmovaps (%rdi), %xmm511970; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm3 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]11971; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm6 = xmm3[0,1],xmm2[2,3]11972; AVX2-FP-NEXT:    vmovaps (%rax), %xmm211973; AVX2-FP-NEXT:    vbroadcastss %xmm2, %xmm711974; AVX2-FP-NEXT:    vmovaps (%r10), %xmm311975; AVX2-FP-NEXT:    vbroadcastss %xmm3, %xmm811976; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm8[0],xmm7[0],xmm8[1],xmm7[1]11977; AVX2-FP-NEXT:    vmovaps (%r9), %xmm811978; AVX2-FP-NEXT:    vmovaps (%r8), %xmm911979; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm10 = xmm9[0],xmm8[0],xmm9[1],xmm8[1]11980; AVX2-FP-NEXT:    vinsertf128 $1, %xmm10, %ymm0, %ymm1011981; AVX2-FP-NEXT:    vinsertf128 $1, %xmm7, %ymm0, %ymm711982; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm7 = ymm10[0,1,2,3,4,5],ymm7[6,7]11983; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7]11984; AVX2-FP-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11985; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm6 = xmm5[1,1,1,1]11986; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm6 = xmm6[0],xmm4[1],xmm6[2,3]11987; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]11988; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm6 = xmm6[0,1],xmm7[2,3]11989; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]11990; AVX2-FP-NEXT:    vinsertf128 $1, %xmm7, %ymm0, %ymm711991; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm10 = xmm9[1,1,1,1]11992; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm10 = xmm10[0],xmm8[1],xmm10[2,3]11993; AVX2-FP-NEXT:    vinsertf128 $1, %xmm10, %ymm0, %ymm1011994; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm7 = ymm10[0,1,2,3,4,5],ymm7[6,7]11995; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7]11996; AVX2-FP-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill11997; AVX2-FP-NEXT:    vunpckhps {{.*#+}} xmm4 = xmm5[2],xmm4[2],xmm5[3],xmm4[3]11998; AVX2-FP-NEXT:    vunpckhps {{.*#+}} xmm5 = xmm9[2],xmm8[2],xmm9[3],xmm8[3]11999; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm6 = xmm2[2,2,2,2]12000; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm6 = xmm3[0,1,2],xmm6[3]12001; AVX2-FP-NEXT:    vinsertf128 $1, %xmm6, %ymm0, %ymm612002; AVX2-FP-NEXT:    vinsertf128 $1, %xmm5, %ymm0, %ymm712003; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3,4,5],ymm6[6,7]12004; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm7 = xmm0[2,2,2,2]12005; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm7 = xmm1[0,1,2],xmm7[3]12006; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm7 = xmm4[0,1],xmm7[2,3]12007; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]12008; AVX2-FP-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12009; AVX2-FP-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm1[2],xmm0[2],xmm1[3],xmm0[3]12010; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm4[1],xmm0[1]12011; AVX2-FP-NEXT:    vunpckhps {{.*#+}} xmm1 = xmm3[2],xmm2[2],xmm3[3],xmm2[3]12012; AVX2-FP-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm112013; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm2 = xmm5[2,3,2,3]12014; AVX2-FP-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm212015; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]12016; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]12017; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12018; AVX2-FP-NEXT:    vmovaps 32(%rcx), %xmm012019; AVX2-FP-NEXT:    vmovaps 32(%rdx), %xmm112020; AVX2-FP-NEXT:    vbroadcastss %xmm0, %xmm212021; AVX2-FP-NEXT:    vbroadcastss %xmm1, %xmm312022; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]12023; AVX2-FP-NEXT:    vmovaps 32(%rsi), %xmm412024; AVX2-FP-NEXT:    vmovaps 32(%rdi), %xmm512025; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm3 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]12026; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm6 = xmm3[0,1],xmm2[2,3]12027; AVX2-FP-NEXT:    vmovaps 32(%r10), %xmm212028; AVX2-FP-NEXT:    vmovaps 32(%rax), %xmm312029; AVX2-FP-NEXT:    vbroadcastss %xmm3, %xmm712030; AVX2-FP-NEXT:    vbroadcastss %xmm2, %xmm812031; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm8[0],xmm7[0],xmm8[1],xmm7[1]12032; AVX2-FP-NEXT:    vmovaps 32(%r9), %xmm812033; AVX2-FP-NEXT:    vmovaps 32(%r8), %xmm912034; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm10 = xmm9[0],xmm8[0],xmm9[1],xmm8[1]12035; AVX2-FP-NEXT:    vinsertf128 $1, %xmm10, %ymm0, %ymm1012036; AVX2-FP-NEXT:    vinsertf128 $1, %xmm7, %ymm0, %ymm712037; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm7 = ymm10[0,1,2,3,4,5],ymm7[6,7]12038; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7]12039; AVX2-FP-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12040; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm6 = xmm5[1,1,1,1]12041; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm6 = xmm6[0],xmm4[1],xmm6[2,3]12042; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]12043; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm6 = xmm6[0,1],xmm7[2,3]12044; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm2[0],xmm3[0],xmm2[1],xmm3[1]12045; AVX2-FP-NEXT:    vinsertf128 $1, %xmm7, %ymm0, %ymm712046; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm10 = xmm9[1,1,1,1]12047; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm10 = xmm10[0],xmm8[1],xmm10[2,3]12048; AVX2-FP-NEXT:    vinsertf128 $1, %xmm10, %ymm0, %ymm1012049; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm7 = ymm10[0,1,2,3,4,5],ymm7[6,7]12050; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7]12051; AVX2-FP-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12052; AVX2-FP-NEXT:    vunpckhps {{.*#+}} xmm4 = xmm5[2],xmm4[2],xmm5[3],xmm4[3]12053; AVX2-FP-NEXT:    vunpckhps {{.*#+}} xmm5 = xmm9[2],xmm8[2],xmm9[3],xmm8[3]12054; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm6 = xmm3[2,2,2,2]12055; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm6 = xmm2[0,1,2],xmm6[3]12056; AVX2-FP-NEXT:    vinsertf128 $1, %xmm6, %ymm0, %ymm612057; AVX2-FP-NEXT:    vinsertf128 $1, %xmm5, %ymm0, %ymm712058; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3,4,5],ymm6[6,7]12059; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm7 = xmm0[2,2,2,2]12060; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm7 = xmm1[0,1,2],xmm7[3]12061; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm7 = xmm4[0,1],xmm7[2,3]12062; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]12063; AVX2-FP-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12064; AVX2-FP-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm1[2],xmm0[2],xmm1[3],xmm0[3]12065; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm4[1],xmm0[1]12066; AVX2-FP-NEXT:    vunpckhps {{.*#+}} xmm1 = xmm2[2],xmm3[2],xmm2[3],xmm3[3]12067; AVX2-FP-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm112068; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm2 = xmm5[2,3,2,3]12069; AVX2-FP-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm212070; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]12071; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]12072; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12073; AVX2-FP-NEXT:    vmovaps 64(%rcx), %xmm012074; AVX2-FP-NEXT:    vmovaps 64(%rdx), %xmm112075; AVX2-FP-NEXT:    vbroadcastss %xmm0, %xmm212076; AVX2-FP-NEXT:    vbroadcastss %xmm1, %xmm312077; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]12078; AVX2-FP-NEXT:    vmovaps 64(%rsi), %xmm412079; AVX2-FP-NEXT:    vmovaps 64(%rdi), %xmm512080; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm3 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]12081; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm6 = xmm3[0,1],xmm2[2,3]12082; AVX2-FP-NEXT:    vmovaps 64(%r10), %xmm212083; AVX2-FP-NEXT:    vmovaps 64(%rax), %xmm312084; AVX2-FP-NEXT:    vbroadcastss %xmm3, %xmm712085; AVX2-FP-NEXT:    vbroadcastss %xmm2, %xmm812086; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm8[0],xmm7[0],xmm8[1],xmm7[1]12087; AVX2-FP-NEXT:    vmovaps 64(%r9), %xmm812088; AVX2-FP-NEXT:    vmovaps 64(%r8), %xmm912089; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm10 = xmm9[0],xmm8[0],xmm9[1],xmm8[1]12090; AVX2-FP-NEXT:    vinsertf128 $1, %xmm10, %ymm0, %ymm1012091; AVX2-FP-NEXT:    vinsertf128 $1, %xmm7, %ymm0, %ymm712092; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm7 = ymm10[0,1,2,3,4,5],ymm7[6,7]12093; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7]12094; AVX2-FP-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12095; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm6 = xmm5[1,1,1,1]12096; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm6 = xmm6[0],xmm4[1],xmm6[2,3]12097; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]12098; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm6 = xmm6[0,1],xmm7[2,3]12099; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm2[0],xmm3[0],xmm2[1],xmm3[1]12100; AVX2-FP-NEXT:    vinsertf128 $1, %xmm7, %ymm0, %ymm712101; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm10 = xmm9[1,1,1,1]12102; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm10 = xmm10[0],xmm8[1],xmm10[2,3]12103; AVX2-FP-NEXT:    vinsertf128 $1, %xmm10, %ymm0, %ymm1012104; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm7 = ymm10[0,1,2,3,4,5],ymm7[6,7]12105; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7]12106; AVX2-FP-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12107; AVX2-FP-NEXT:    vunpckhps {{.*#+}} xmm4 = xmm5[2],xmm4[2],xmm5[3],xmm4[3]12108; AVX2-FP-NEXT:    vunpckhps {{.*#+}} xmm5 = xmm9[2],xmm8[2],xmm9[3],xmm8[3]12109; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm6 = xmm3[2,2,2,2]12110; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm6 = xmm2[0,1,2],xmm6[3]12111; AVX2-FP-NEXT:    vinsertf128 $1, %xmm6, %ymm0, %ymm612112; AVX2-FP-NEXT:    vinsertf128 $1, %xmm5, %ymm0, %ymm712113; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3,4,5],ymm6[6,7]12114; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm7 = xmm0[2,2,2,2]12115; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm7 = xmm1[0,1,2],xmm7[3]12116; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm7 = xmm4[0,1],xmm7[2,3]12117; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]12118; AVX2-FP-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12119; AVX2-FP-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm1[2],xmm0[2],xmm1[3],xmm0[3]12120; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm4[1],xmm0[1]12121; AVX2-FP-NEXT:    vunpckhps {{.*#+}} xmm1 = xmm2[2],xmm3[2],xmm2[3],xmm3[3]12122; AVX2-FP-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm112123; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm2 = xmm5[2,3,2,3]12124; AVX2-FP-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm212125; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]12126; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]12127; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12128; AVX2-FP-NEXT:    vmovaps 96(%rcx), %xmm012129; AVX2-FP-NEXT:    vbroadcastss %xmm0, %xmm212130; AVX2-FP-NEXT:    vmovaps 96(%rdx), %xmm112131; AVX2-FP-NEXT:    vbroadcastss %xmm1, %xmm312132; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]12133; AVX2-FP-NEXT:    vmovaps 96(%rsi), %xmm412134; AVX2-FP-NEXT:    vmovaps 96(%rdi), %xmm512135; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm3 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]12136; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm6 = xmm3[0,1],xmm2[2,3]12137; AVX2-FP-NEXT:    vmovaps 96(%rax), %xmm212138; AVX2-FP-NEXT:    vbroadcastss %xmm2, %xmm712139; AVX2-FP-NEXT:    vmovaps 96(%r10), %xmm312140; AVX2-FP-NEXT:    vbroadcastss %xmm3, %xmm812141; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm8[0],xmm7[0],xmm8[1],xmm7[1]12142; AVX2-FP-NEXT:    vmovaps 96(%r9), %xmm812143; AVX2-FP-NEXT:    vmovaps 96(%r8), %xmm912144; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm10 = xmm9[0],xmm8[0],xmm9[1],xmm8[1]12145; AVX2-FP-NEXT:    vinsertf128 $1, %xmm10, %ymm0, %ymm1012146; AVX2-FP-NEXT:    vinsertf128 $1, %xmm7, %ymm0, %ymm712147; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm7 = ymm10[0,1,2,3,4,5],ymm7[6,7]12148; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7]12149; AVX2-FP-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12150; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm6 = xmm5[1,1,1,1]12151; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm6 = xmm6[0],xmm4[1],xmm6[2,3]12152; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]12153; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm6 = xmm6[0,1],xmm7[2,3]12154; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]12155; AVX2-FP-NEXT:    vinsertf128 $1, %xmm7, %ymm0, %ymm712156; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm10 = xmm9[1,1,1,1]12157; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm10 = xmm10[0],xmm8[1],xmm10[2,3]12158; AVX2-FP-NEXT:    vinsertf128 $1, %xmm10, %ymm0, %ymm1012159; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm7 = ymm10[0,1,2,3,4,5],ymm7[6,7]12160; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7]12161; AVX2-FP-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12162; AVX2-FP-NEXT:    vunpckhps {{.*#+}} xmm4 = xmm5[2],xmm4[2],xmm5[3],xmm4[3]12163; AVX2-FP-NEXT:    vunpckhps {{.*#+}} xmm5 = xmm9[2],xmm8[2],xmm9[3],xmm8[3]12164; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm6 = xmm2[2,2,2,2]12165; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm6 = xmm3[0,1,2],xmm6[3]12166; AVX2-FP-NEXT:    vinsertf128 $1, %xmm6, %ymm0, %ymm612167; AVX2-FP-NEXT:    vinsertf128 $1, %xmm5, %ymm0, %ymm712168; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3,4,5],ymm6[6,7]12169; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm7 = xmm0[2,2,2,2]12170; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm7 = xmm1[0,1,2],xmm7[3]12171; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm7 = xmm4[0,1],xmm7[2,3]12172; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]12173; AVX2-FP-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12174; AVX2-FP-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm1[2],xmm0[2],xmm1[3],xmm0[3]12175; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm4[1],xmm0[1]12176; AVX2-FP-NEXT:    vunpckhps {{.*#+}} xmm1 = xmm3[2],xmm2[2],xmm3[3],xmm2[3]12177; AVX2-FP-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm112178; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm2 = xmm5[2,3,2,3]12179; AVX2-FP-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm212180; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]12181; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]12182; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12183; AVX2-FP-NEXT:    vmovaps 128(%rcx), %xmm012184; AVX2-FP-NEXT:    vbroadcastss %xmm0, %xmm212185; AVX2-FP-NEXT:    vmovaps 128(%rdx), %xmm112186; AVX2-FP-NEXT:    vbroadcastss %xmm1, %xmm312187; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]12188; AVX2-FP-NEXT:    vmovaps 128(%rsi), %xmm412189; AVX2-FP-NEXT:    vmovaps 128(%rdi), %xmm512190; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm3 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]12191; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm6 = xmm3[0,1],xmm2[2,3]12192; AVX2-FP-NEXT:    vmovaps 128(%rax), %xmm212193; AVX2-FP-NEXT:    vbroadcastss %xmm2, %xmm712194; AVX2-FP-NEXT:    vmovaps 128(%r10), %xmm312195; AVX2-FP-NEXT:    vbroadcastss %xmm3, %xmm812196; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm8[0],xmm7[0],xmm8[1],xmm7[1]12197; AVX2-FP-NEXT:    vmovaps 128(%r9), %xmm812198; AVX2-FP-NEXT:    vmovaps 128(%r8), %xmm912199; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm10 = xmm9[0],xmm8[0],xmm9[1],xmm8[1]12200; AVX2-FP-NEXT:    vinsertf128 $1, %xmm10, %ymm0, %ymm1012201; AVX2-FP-NEXT:    vinsertf128 $1, %xmm7, %ymm0, %ymm712202; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm7 = ymm10[0,1,2,3,4,5],ymm7[6,7]12203; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7]12204; AVX2-FP-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12205; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm6 = xmm5[1,1,1,1]12206; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm6 = xmm6[0],xmm4[1],xmm6[2,3]12207; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]12208; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm6 = xmm6[0,1],xmm7[2,3]12209; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]12210; AVX2-FP-NEXT:    vinsertf128 $1, %xmm7, %ymm0, %ymm712211; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm10 = xmm9[1,1,1,1]12212; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm10 = xmm10[0],xmm8[1],xmm10[2,3]12213; AVX2-FP-NEXT:    vinsertf128 $1, %xmm10, %ymm0, %ymm1012214; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm7 = ymm10[0,1,2,3,4,5],ymm7[6,7]12215; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7]12216; AVX2-FP-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12217; AVX2-FP-NEXT:    vunpckhps {{.*#+}} xmm4 = xmm5[2],xmm4[2],xmm5[3],xmm4[3]12218; AVX2-FP-NEXT:    vunpckhps {{.*#+}} xmm5 = xmm9[2],xmm8[2],xmm9[3],xmm8[3]12219; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm6 = xmm2[2,2,2,2]12220; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm6 = xmm3[0,1,2],xmm6[3]12221; AVX2-FP-NEXT:    vinsertf128 $1, %xmm6, %ymm0, %ymm612222; AVX2-FP-NEXT:    vinsertf128 $1, %xmm5, %ymm0, %ymm712223; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3,4,5],ymm6[6,7]12224; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm7 = xmm0[2,2,2,2]12225; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm7 = xmm1[0,1,2],xmm7[3]12226; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm7 = xmm4[0,1],xmm7[2,3]12227; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]12228; AVX2-FP-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12229; AVX2-FP-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm1[2],xmm0[2],xmm1[3],xmm0[3]12230; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm4[1],xmm0[1]12231; AVX2-FP-NEXT:    vunpckhps {{.*#+}} xmm1 = xmm3[2],xmm2[2],xmm3[3],xmm2[3]12232; AVX2-FP-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm112233; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm2 = xmm5[2,3,2,3]12234; AVX2-FP-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm212235; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]12236; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]12237; AVX2-FP-NEXT:    vmovups %ymm0, (%rsp) # 32-byte Spill12238; AVX2-FP-NEXT:    vmovaps 160(%rcx), %xmm012239; AVX2-FP-NEXT:    vbroadcastss %xmm0, %xmm212240; AVX2-FP-NEXT:    vmovaps 160(%rdx), %xmm112241; AVX2-FP-NEXT:    vbroadcastss %xmm1, %xmm312242; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]12243; AVX2-FP-NEXT:    vmovaps 160(%rsi), %xmm412244; AVX2-FP-NEXT:    vmovaps 160(%rdi), %xmm512245; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm3 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]12246; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm6 = xmm3[0,1],xmm2[2,3]12247; AVX2-FP-NEXT:    vmovaps 160(%rax), %xmm212248; AVX2-FP-NEXT:    vbroadcastss %xmm2, %xmm712249; AVX2-FP-NEXT:    vmovaps 160(%r10), %xmm312250; AVX2-FP-NEXT:    vbroadcastss %xmm3, %xmm812251; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm8[0],xmm7[0],xmm8[1],xmm7[1]12252; AVX2-FP-NEXT:    vmovaps 160(%r9), %xmm812253; AVX2-FP-NEXT:    vmovaps 160(%r8), %xmm912254; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm10 = xmm9[0],xmm8[0],xmm9[1],xmm8[1]12255; AVX2-FP-NEXT:    vinsertf128 $1, %xmm10, %ymm0, %ymm1012256; AVX2-FP-NEXT:    vinsertf128 $1, %xmm7, %ymm0, %ymm712257; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm7 = ymm10[0,1,2,3,4,5],ymm7[6,7]12258; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7]12259; AVX2-FP-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12260; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm6 = xmm5[1,1,1,1]12261; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm6 = xmm6[0],xmm4[1],xmm6[2,3]12262; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]12263; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm6 = xmm6[0,1],xmm7[2,3]12264; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]12265; AVX2-FP-NEXT:    vinsertf128 $1, %xmm7, %ymm0, %ymm712266; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm10 = xmm9[1,1,1,1]12267; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm10 = xmm10[0],xmm8[1],xmm10[2,3]12268; AVX2-FP-NEXT:    vinsertf128 $1, %xmm10, %ymm0, %ymm1012269; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm7 = ymm10[0,1,2,3,4,5],ymm7[6,7]12270; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7]12271; AVX2-FP-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12272; AVX2-FP-NEXT:    vunpckhps {{.*#+}} xmm4 = xmm5[2],xmm4[2],xmm5[3],xmm4[3]12273; AVX2-FP-NEXT:    vunpckhps {{.*#+}} xmm5 = xmm9[2],xmm8[2],xmm9[3],xmm8[3]12274; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm6 = xmm2[2,2,2,2]12275; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm6 = xmm3[0,1,2],xmm6[3]12276; AVX2-FP-NEXT:    vinsertf128 $1, %xmm6, %ymm0, %ymm612277; AVX2-FP-NEXT:    vinsertf128 $1, %xmm5, %ymm0, %ymm712278; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3,4,5],ymm6[6,7]12279; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm7 = xmm0[2,2,2,2]12280; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm7 = xmm1[0,1,2],xmm7[3]12281; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm7 = xmm4[0,1],xmm7[2,3]12282; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]12283; AVX2-FP-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12284; AVX2-FP-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm1[2],xmm0[2],xmm1[3],xmm0[3]12285; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm4[1],xmm0[1]12286; AVX2-FP-NEXT:    vunpckhps {{.*#+}} xmm1 = xmm3[2],xmm2[2],xmm3[3],xmm2[3]12287; AVX2-FP-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm112288; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm2 = xmm5[2,3,2,3]12289; AVX2-FP-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm212290; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]12291; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]12292; AVX2-FP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12293; AVX2-FP-NEXT:    vmovaps 192(%rcx), %xmm012294; AVX2-FP-NEXT:    vbroadcastss %xmm0, %xmm212295; AVX2-FP-NEXT:    vmovaps 192(%rdx), %xmm112296; AVX2-FP-NEXT:    vbroadcastss %xmm1, %xmm312297; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]12298; AVX2-FP-NEXT:    vmovaps 192(%rsi), %xmm412299; AVX2-FP-NEXT:    vmovaps 192(%rdi), %xmm512300; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm3 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]12301; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm9 = xmm3[0,1],xmm2[2,3]12302; AVX2-FP-NEXT:    vmovaps 192(%rax), %xmm212303; AVX2-FP-NEXT:    vbroadcastss %xmm2, %xmm712304; AVX2-FP-NEXT:    vmovaps 192(%r10), %xmm312305; AVX2-FP-NEXT:    vbroadcastss %xmm3, %xmm812306; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm8[0],xmm7[0],xmm8[1],xmm7[1]12307; AVX2-FP-NEXT:    vmovaps 192(%r9), %xmm812308; AVX2-FP-NEXT:    vmovaps 192(%r8), %xmm612309; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm10 = xmm6[0],xmm8[0],xmm6[1],xmm8[1]12310; AVX2-FP-NEXT:    vinsertf128 $1, %xmm10, %ymm0, %ymm1012311; AVX2-FP-NEXT:    vinsertf128 $1, %xmm7, %ymm0, %ymm712312; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm7 = ymm10[0,1,2,3,4,5],ymm7[6,7]12313; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm11 = ymm9[0,1,2,3],ymm7[4,5,6,7]12314; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm7 = xmm5[1,1,1,1]12315; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm7 = xmm7[0],xmm4[1],xmm7[2,3]12316; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm9 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]12317; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm7 = xmm7[0,1],xmm9[2,3]12318; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm9 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]12319; AVX2-FP-NEXT:    vinsertf128 $1, %xmm9, %ymm0, %ymm912320; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm15 = xmm6[1,1,1,1]12321; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm15 = xmm15[0],xmm8[1],xmm15[2,3]12322; AVX2-FP-NEXT:    vinsertf128 $1, %xmm15, %ymm0, %ymm1512323; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm9 = ymm15[0,1,2,3,4,5],ymm9[6,7]12324; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm10 = ymm7[0,1,2,3],ymm9[4,5,6,7]12325; AVX2-FP-NEXT:    vunpckhps {{.*#+}} xmm4 = xmm5[2],xmm4[2],xmm5[3],xmm4[3]12326; AVX2-FP-NEXT:    vunpckhps {{.*#+}} xmm5 = xmm6[2],xmm8[2],xmm6[3],xmm8[3]12327; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm6 = xmm2[2,2,2,2]12328; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm6 = xmm3[0,1,2],xmm6[3]12329; AVX2-FP-NEXT:    vinsertf128 $1, %xmm6, %ymm0, %ymm612330; AVX2-FP-NEXT:    vinsertf128 $1, %xmm5, %ymm0, %ymm712331; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3,4,5],ymm6[6,7]12332; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm7 = xmm0[2,2,2,2]12333; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm7 = xmm1[0,1,2],xmm7[3]12334; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm7 = xmm4[0,1],xmm7[2,3]12335; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm9 = ymm7[0,1,2,3],ymm6[4,5,6,7]12336; AVX2-FP-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm1[2],xmm0[2],xmm1[3],xmm0[3]12337; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm4[1],xmm0[1]12338; AVX2-FP-NEXT:    vunpckhps {{.*#+}} xmm1 = xmm3[2],xmm2[2],xmm3[3],xmm2[3]12339; AVX2-FP-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm112340; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm2 = xmm5[2,3,2,3]12341; AVX2-FP-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm212342; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]12343; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm8 = ymm0[0,1,2,3],ymm1[4,5,6,7]12344; AVX2-FP-NEXT:    vmovaps 224(%rcx), %xmm612345; AVX2-FP-NEXT:    vbroadcastss %xmm6, %xmm012346; AVX2-FP-NEXT:    vmovaps 224(%rdx), %xmm512347; AVX2-FP-NEXT:    vbroadcastss %xmm5, %xmm112348; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]12349; AVX2-FP-NEXT:    vmovaps 224(%rsi), %xmm412350; AVX2-FP-NEXT:    vmovaps 224(%rdi), %xmm212351; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm3 = xmm2[0],xmm4[0],xmm2[1],xmm4[1]12352; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm7 = xmm3[0,1],xmm1[2,3]12353; AVX2-FP-NEXT:    vmovaps 224(%rax), %xmm312354; AVX2-FP-NEXT:    vmovaps 224(%r10), %xmm112355; AVX2-FP-NEXT:    vbroadcastss %xmm3, %xmm1512356; AVX2-FP-NEXT:    vbroadcastss %xmm1, %xmm1412357; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm14 = xmm14[0],xmm15[0],xmm14[1],xmm15[1]12358; AVX2-FP-NEXT:    vmovaps 224(%r9), %xmm1512359; AVX2-FP-NEXT:    vmovaps 224(%r8), %xmm012360; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm13 = xmm0[0],xmm15[0],xmm0[1],xmm15[1]12361; AVX2-FP-NEXT:    vinsertf128 $1, %xmm13, %ymm0, %ymm1312362; AVX2-FP-NEXT:    vinsertf128 $1, %xmm14, %ymm0, %ymm1412363; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm13 = ymm13[0,1,2,3,4,5],ymm14[6,7]12364; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm14 = ymm7[0,1,2,3],ymm13[4,5,6,7]12365; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm7 = xmm2[1,1,1,1]12366; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm7 = xmm7[0],xmm4[1],xmm7[2,3]12367; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm13 = xmm5[0],xmm6[0],xmm5[1],xmm6[1]12368; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm7 = xmm7[0,1],xmm13[2,3]12369; AVX2-FP-NEXT:    vunpcklps {{.*#+}} xmm13 = xmm1[0],xmm3[0],xmm1[1],xmm3[1]12370; AVX2-FP-NEXT:    vinsertf128 $1, %xmm13, %ymm0, %ymm1312371; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm12 = xmm0[1,1,1,1]12372; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm12 = xmm12[0],xmm15[1],xmm12[2,3]12373; AVX2-FP-NEXT:    vinsertf128 $1, %xmm12, %ymm0, %ymm1212374; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm12 = ymm12[0,1,2,3,4,5],ymm13[6,7]12375; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm7 = ymm7[0,1,2,3],ymm12[4,5,6,7]12376; AVX2-FP-NEXT:    vunpckhps {{.*#+}} xmm2 = xmm2[2],xmm4[2],xmm2[3],xmm4[3]12377; AVX2-FP-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm0[2],xmm15[2],xmm0[3],xmm15[3]12378; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm4 = xmm3[2,2,2,2]12379; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm4 = xmm1[0,1,2],xmm4[3]12380; AVX2-FP-NEXT:    vinsertf128 $1, %xmm4, %ymm0, %ymm412381; AVX2-FP-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm1212382; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm4 = ymm12[0,1,2,3,4,5],ymm4[6,7]12383; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm12 = xmm6[2,2,2,2]12384; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm12 = xmm5[0,1,2],xmm12[3]12385; AVX2-FP-NEXT:    vblendps {{.*#+}} xmm12 = xmm2[0,1],xmm12[2,3]12386; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm4 = ymm12[0,1,2,3],ymm4[4,5,6,7]12387; AVX2-FP-NEXT:    vunpckhps {{.*#+}} xmm5 = xmm5[2],xmm6[2],xmm5[3],xmm6[3]12388; AVX2-FP-NEXT:    vunpckhpd {{.*#+}} xmm2 = xmm2[1],xmm5[1]12389; AVX2-FP-NEXT:    vunpckhps {{.*#+}} xmm1 = xmm1[2],xmm3[2],xmm1[3],xmm3[3]12390; AVX2-FP-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm112391; AVX2-FP-NEXT:    vshufps {{.*#+}} xmm0 = xmm0[2,3,2,3]12392; AVX2-FP-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm012393; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]12394; AVX2-FP-NEXT:    vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]12395; AVX2-FP-NEXT:    movq {{[0-9]+}}(%rsp), %rax12396; AVX2-FP-NEXT:    vmovaps %ymm0, 1888(%rax)12397; AVX2-FP-NEXT:    vmovaps %ymm4, 1856(%rax)12398; AVX2-FP-NEXT:    vmovaps %ymm7, 1824(%rax)12399; AVX2-FP-NEXT:    vmovaps %ymm14, 1792(%rax)12400; AVX2-FP-NEXT:    vmovaps %ymm8, 1632(%rax)12401; AVX2-FP-NEXT:    vmovaps %ymm9, 1600(%rax)12402; AVX2-FP-NEXT:    vmovaps %ymm10, 1568(%rax)12403; AVX2-FP-NEXT:    vmovaps %ymm11, 1536(%rax)12404; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12405; AVX2-FP-NEXT:    vmovaps %ymm0, 1376(%rax)12406; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12407; AVX2-FP-NEXT:    vmovaps %ymm0, 1344(%rax)12408; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12409; AVX2-FP-NEXT:    vmovaps %ymm0, 1312(%rax)12410; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12411; AVX2-FP-NEXT:    vmovaps %ymm0, 1280(%rax)12412; AVX2-FP-NEXT:    vmovups (%rsp), %ymm0 # 32-byte Reload12413; AVX2-FP-NEXT:    vmovaps %ymm0, 1120(%rax)12414; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12415; AVX2-FP-NEXT:    vmovaps %ymm0, 1088(%rax)12416; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12417; AVX2-FP-NEXT:    vmovaps %ymm0, 1056(%rax)12418; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12419; AVX2-FP-NEXT:    vmovaps %ymm0, 1024(%rax)12420; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12421; AVX2-FP-NEXT:    vmovaps %ymm0, 864(%rax)12422; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12423; AVX2-FP-NEXT:    vmovaps %ymm0, 832(%rax)12424; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12425; AVX2-FP-NEXT:    vmovaps %ymm0, 800(%rax)12426; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12427; AVX2-FP-NEXT:    vmovaps %ymm0, 768(%rax)12428; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12429; AVX2-FP-NEXT:    vmovaps %ymm0, 608(%rax)12430; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12431; AVX2-FP-NEXT:    vmovaps %ymm0, 576(%rax)12432; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12433; AVX2-FP-NEXT:    vmovaps %ymm0, 544(%rax)12434; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12435; AVX2-FP-NEXT:    vmovaps %ymm0, 512(%rax)12436; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12437; AVX2-FP-NEXT:    vmovaps %ymm0, 352(%rax)12438; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12439; AVX2-FP-NEXT:    vmovaps %ymm0, 320(%rax)12440; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12441; AVX2-FP-NEXT:    vmovaps %ymm0, 288(%rax)12442; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12443; AVX2-FP-NEXT:    vmovaps %ymm0, 256(%rax)12444; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12445; AVX2-FP-NEXT:    vmovaps %ymm0, 96(%rax)12446; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12447; AVX2-FP-NEXT:    vmovaps %ymm0, 64(%rax)12448; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12449; AVX2-FP-NEXT:    vmovaps %ymm0, 32(%rax)12450; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12451; AVX2-FP-NEXT:    vmovaps %ymm0, (%rax)12452; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12453; AVX2-FP-NEXT:    vmovaps %ymm0, 2016(%rax)12454; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12455; AVX2-FP-NEXT:    vmovaps %ymm0, 1984(%rax)12456; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12457; AVX2-FP-NEXT:    vmovaps %ymm0, 1952(%rax)12458; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12459; AVX2-FP-NEXT:    vmovaps %ymm0, 1920(%rax)12460; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12461; AVX2-FP-NEXT:    vmovaps %ymm0, 1760(%rax)12462; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12463; AVX2-FP-NEXT:    vmovaps %ymm0, 1728(%rax)12464; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12465; AVX2-FP-NEXT:    vmovaps %ymm0, 1696(%rax)12466; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12467; AVX2-FP-NEXT:    vmovaps %ymm0, 1664(%rax)12468; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12469; AVX2-FP-NEXT:    vmovaps %ymm0, 1504(%rax)12470; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12471; AVX2-FP-NEXT:    vmovaps %ymm0, 1472(%rax)12472; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12473; AVX2-FP-NEXT:    vmovaps %ymm0, 1440(%rax)12474; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12475; AVX2-FP-NEXT:    vmovaps %ymm0, 1408(%rax)12476; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12477; AVX2-FP-NEXT:    vmovaps %ymm0, 1248(%rax)12478; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12479; AVX2-FP-NEXT:    vmovaps %ymm0, 1216(%rax)12480; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12481; AVX2-FP-NEXT:    vmovaps %ymm0, 1184(%rax)12482; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12483; AVX2-FP-NEXT:    vmovaps %ymm0, 1152(%rax)12484; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12485; AVX2-FP-NEXT:    vmovaps %ymm0, 992(%rax)12486; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12487; AVX2-FP-NEXT:    vmovaps %ymm0, 960(%rax)12488; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12489; AVX2-FP-NEXT:    vmovaps %ymm0, 928(%rax)12490; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12491; AVX2-FP-NEXT:    vmovaps %ymm0, 896(%rax)12492; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12493; AVX2-FP-NEXT:    vmovaps %ymm0, 736(%rax)12494; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12495; AVX2-FP-NEXT:    vmovaps %ymm0, 704(%rax)12496; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12497; AVX2-FP-NEXT:    vmovaps %ymm0, 672(%rax)12498; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12499; AVX2-FP-NEXT:    vmovaps %ymm0, 640(%rax)12500; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12501; AVX2-FP-NEXT:    vmovaps %ymm0, 480(%rax)12502; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12503; AVX2-FP-NEXT:    vmovaps %ymm0, 448(%rax)12504; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12505; AVX2-FP-NEXT:    vmovaps %ymm0, 416(%rax)12506; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12507; AVX2-FP-NEXT:    vmovaps %ymm0, 384(%rax)12508; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12509; AVX2-FP-NEXT:    vmovaps %ymm0, 224(%rax)12510; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12511; AVX2-FP-NEXT:    vmovaps %ymm0, 192(%rax)12512; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12513; AVX2-FP-NEXT:    vmovaps %ymm0, 160(%rax)12514; AVX2-FP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload12515; AVX2-FP-NEXT:    vmovaps %ymm0, 128(%rax)12516; AVX2-FP-NEXT:    addq $1672, %rsp # imm = 0x68812517; AVX2-FP-NEXT:    vzeroupper12518; AVX2-FP-NEXT:    retq12519;12520; AVX2-FCP-LABEL: store_i32_stride8_vf64:12521; AVX2-FCP:       # %bb.0:12522; AVX2-FCP-NEXT:    subq $1672, %rsp # imm = 0x68812523; AVX2-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax12524; AVX2-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r1012525; AVX2-FCP-NEXT:    vmovaps 32(%rdi), %ymm012526; AVX2-FCP-NEXT:    vmovaps (%rdi), %ymm412527; AVX2-FCP-NEXT:    vmovaps 32(%rsi), %ymm112528; AVX2-FCP-NEXT:    vmovaps (%rsi), %ymm512529; AVX2-FCP-NEXT:    vmovaps 32(%rdx), %ymm212530; AVX2-FCP-NEXT:    vmovaps (%rdx), %ymm612531; AVX2-FCP-NEXT:    vmovaps 32(%rcx), %ymm312532; AVX2-FCP-NEXT:    vmovaps (%rcx), %ymm712533; AVX2-FCP-NEXT:    vmovaps (%r8), %ymm912534; AVX2-FCP-NEXT:    vmovaps (%r9), %ymm1112535; AVX2-FCP-NEXT:    vmovaps (%r10), %ymm1012536; AVX2-FCP-NEXT:    vmovaps (%rax), %ymm1212537; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} ymm8 = ymm10[0],ymm12[0],ymm10[1],ymm12[1],ymm10[4],ymm12[4],ymm10[5],ymm12[5]12538; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} ymm13 = ymm4[0],ymm5[0],ymm4[1],ymm5[1],ymm4[4],ymm5[4],ymm4[5],ymm5[5]12539; AVX2-FCP-NEXT:    vextractf128 $1, %ymm13, %xmm1312540; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} ymm14 = ymm6[0],ymm7[0],ymm6[1],ymm7[1],ymm6[4],ymm7[4],ymm6[5],ymm7[5]12541; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm15 = ymm14[2,2,2,2]12542; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm13 = xmm13[0,1],xmm15[2,3]12543; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} ymm15 = ymm9[0],ymm11[0],ymm9[1],ymm11[1],ymm9[4],ymm11[4],ymm9[5],ymm11[5]12544; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm15 = ymm15[0],ymm8[0],ymm15[2],ymm8[2]12545; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm13 = ymm13[0,1,2,3],ymm15[4,5,6,7]12546; AVX2-FCP-NEXT:    vmovups %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12547; AVX2-FCP-NEXT:    vbroadcastss 20(%r8), %ymm1312548; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm13 = ymm13[0,1,2,3,4],ymm11[5],ymm13[6,7]12549; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm13 = ymm13[0,1,2,3,4,5],ymm8[6,7]12550; AVX2-FCP-NEXT:    vextractf128 $1, %ymm14, %xmm812551; AVX2-FCP-NEXT:    vshufps {{.*#+}} ymm14 = ymm4[1,1,1,1,5,5,5,5]12552; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm14 = ymm14[0],ymm5[1],ymm14[2,3,4],ymm5[5],ymm14[6,7]12553; AVX2-FCP-NEXT:    vextractf128 $1, %ymm14, %xmm1412554; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm14 = xmm14[0,1],xmm8[2,3]12555; AVX2-FCP-NEXT:    vmovaps 32(%r8), %ymm812556; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm13 = ymm14[0,1,2,3],ymm13[4,5,6,7]12557; AVX2-FCP-NEXT:    vmovups %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12558; AVX2-FCP-NEXT:    vbroadcastss 24(%rax), %ymm1312559; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm13 = ymm10[0,1,2,3,4,5,6],ymm13[7]12560; AVX2-FCP-NEXT:    vmovaps 32(%r9), %ymm1012561; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} ymm11 = ymm9[2],ymm11[2],ymm9[3],ymm11[3],ymm9[6],ymm11[6],ymm9[7],ymm11[7]12562; AVX2-FCP-NEXT:    vmovaps 32(%r10), %ymm912563; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} ymm4 = ymm4[2],ymm5[2],ymm4[3],ymm5[3],ymm4[6],ymm5[6],ymm4[7],ymm5[7]12564; AVX2-FCP-NEXT:    vextractf128 $1, %ymm4, %xmm512565; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} ymm6 = ymm6[2],ymm7[2],ymm6[3],ymm7[3],ymm6[6],ymm7[6],ymm6[7],ymm7[7]12566; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm7 = ymm6[2,2,2,2]12567; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm5 = xmm5[0,1],xmm7[2,3]12568; AVX2-FCP-NEXT:    vmovaps 32(%rax), %ymm712569; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm13 = ymm11[0,1,2,3,4,5],ymm13[6,7]12570; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm13[4,5,6,7]12571; AVX2-FCP-NEXT:    vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12572; AVX2-FCP-NEXT:    vbroadcastss 28(%r10), %ymm512573; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} ymm5 = ymm5[2],ymm12[2],ymm5[3],ymm12[3],ymm5[6],ymm12[6],ymm5[7],ymm12[7]12574; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm5 = ymm11[1],ymm5[1],ymm11[3],ymm5[3]12575; AVX2-FCP-NEXT:    vextractf128 $1, %ymm6, %xmm612576; AVX2-FCP-NEXT:    vshufps {{.*#+}} ymm4 = ymm4[2,3,2,3,6,7,6,7]12577; AVX2-FCP-NEXT:    vextractf128 $1, %ymm4, %xmm412578; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm4 = xmm4[0,1],xmm6[2,3]12579; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm5[4,5,6,7]12580; AVX2-FCP-NEXT:    vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12581; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} ymm4 = ymm9[0],ymm7[0],ymm9[1],ymm7[1],ymm9[4],ymm7[4],ymm9[5],ymm7[5]12582; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} ymm5 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]12583; AVX2-FCP-NEXT:    vextractf128 $1, %ymm5, %xmm512584; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} ymm6 = ymm2[0],ymm3[0],ymm2[1],ymm3[1],ymm2[4],ymm3[4],ymm2[5],ymm3[5]12585; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm11 = ymm6[2,2,2,2]12586; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm5 = xmm5[0,1],xmm11[2,3]12587; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} ymm11 = ymm8[0],ymm10[0],ymm8[1],ymm10[1],ymm8[4],ymm10[4],ymm8[5],ymm10[5]12588; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm11 = ymm11[0],ymm4[0],ymm11[2],ymm4[2]12589; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm11[4,5,6,7]12590; AVX2-FCP-NEXT:    vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12591; AVX2-FCP-NEXT:    vbroadcastss 52(%r8), %ymm512592; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3,4],ymm10[5],ymm5[6,7]12593; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3,4,5],ymm4[6,7]12594; AVX2-FCP-NEXT:    vextractf128 $1, %ymm6, %xmm412595; AVX2-FCP-NEXT:    vshufps {{.*#+}} ymm6 = ymm0[1,1,1,1,5,5,5,5]12596; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm6 = ymm6[0],ymm1[1],ymm6[2,3,4],ymm1[5],ymm6[6,7]12597; AVX2-FCP-NEXT:    vextractf128 $1, %ymm6, %xmm612598; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm6 = xmm6[0,1],xmm4[2,3]12599; AVX2-FCP-NEXT:    vmovaps 64(%rdi), %ymm412600; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm5 = ymm6[0,1,2,3],ymm5[4,5,6,7]12601; AVX2-FCP-NEXT:    vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12602; AVX2-FCP-NEXT:    vbroadcastss 56(%rax), %ymm512603; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm9 = ymm9[0,1,2,3,4,5,6],ymm5[7]12604; AVX2-FCP-NEXT:    vmovaps 64(%rsi), %ymm512605; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} ymm8 = ymm8[2],ymm10[2],ymm8[3],ymm10[3],ymm8[6],ymm10[6],ymm8[7],ymm10[7]12606; AVX2-FCP-NEXT:    vmovaps 64(%rdx), %ymm612607; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} ymm10 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7]12608; AVX2-FCP-NEXT:    vextractf128 $1, %ymm10, %xmm012609; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} ymm2 = ymm2[2],ymm3[2],ymm2[3],ymm3[3],ymm2[6],ymm3[6],ymm2[7],ymm3[7]12610; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm1 = ymm2[2,2,2,2]12611; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm1 = xmm0[0,1],xmm1[2,3]12612; AVX2-FCP-NEXT:    vmovaps 64(%rcx), %ymm012613; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm3 = ymm8[0,1,2,3,4,5],ymm9[6,7]12614; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm3[4,5,6,7]12615; AVX2-FCP-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12616; AVX2-FCP-NEXT:    vbroadcastss 60(%r10), %ymm112617; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} ymm3 = ymm1[2],ymm7[2],ymm1[3],ymm7[3],ymm1[6],ymm7[6],ymm1[7],ymm7[7]12618; AVX2-FCP-NEXT:    vmovaps 64(%r8), %ymm112619; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm3 = ymm8[1],ymm3[1],ymm8[3],ymm3[3]12620; AVX2-FCP-NEXT:    vextractf128 $1, %ymm2, %xmm212621; AVX2-FCP-NEXT:    vshufps {{.*#+}} ymm7 = ymm10[2,3,2,3,6,7,6,7]12622; AVX2-FCP-NEXT:    vextractf128 $1, %ymm7, %xmm712623; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm2 = xmm7[0,1],xmm2[2,3]12624; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]12625; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12626; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} ymm2 = ymm4[0],ymm5[0],ymm4[1],ymm5[1],ymm4[4],ymm5[4],ymm4[5],ymm5[5]12627; AVX2-FCP-NEXT:    vextractf128 $1, %ymm2, %xmm212628; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} ymm3 = ymm6[0],ymm0[0],ymm6[1],ymm0[1],ymm6[4],ymm0[4],ymm6[5],ymm0[5]12629; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm7 = ymm3[2,2,2,2]12630; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm2 = xmm2[0,1],xmm7[2,3]12631; AVX2-FCP-NEXT:    vmovaps 64(%r9), %ymm712632; AVX2-FCP-NEXT:    vmovaps 64(%r10), %ymm812633; AVX2-FCP-NEXT:    vmovaps 64(%rax), %ymm912634; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} ymm10 = ymm8[0],ymm9[0],ymm8[1],ymm9[1],ymm8[4],ymm9[4],ymm8[5],ymm9[5]12635; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} ymm11 = ymm1[0],ymm7[0],ymm1[1],ymm7[1],ymm1[4],ymm7[4],ymm1[5],ymm7[5]12636; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm11 = ymm11[0],ymm10[0],ymm11[2],ymm10[2]12637; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm11[4,5,6,7]12638; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12639; AVX2-FCP-NEXT:    vbroadcastss 84(%r8), %ymm212640; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4],ymm7[5],ymm2[6,7]12641; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm10[6,7]12642; AVX2-FCP-NEXT:    vextractf128 $1, %ymm3, %xmm312643; AVX2-FCP-NEXT:    vshufps {{.*#+}} ymm10 = ymm4[1,1,1,1,5,5,5,5]12644; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm10 = ymm10[0],ymm5[1],ymm10[2,3,4],ymm5[5],ymm10[6,7]12645; AVX2-FCP-NEXT:    vextractf128 $1, %ymm10, %xmm1012646; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm3 = xmm10[0,1],xmm3[2,3]12647; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm2 = ymm3[0,1,2,3],ymm2[4,5,6,7]12648; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12649; AVX2-FCP-NEXT:    vbroadcastss 88(%rax), %ymm212650; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm2 = ymm8[0,1,2,3,4,5,6],ymm2[7]12651; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} ymm1 = ymm1[2],ymm7[2],ymm1[3],ymm7[3],ymm1[6],ymm7[6],ymm1[7],ymm7[7]12652; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} ymm3 = ymm4[2],ymm5[2],ymm4[3],ymm5[3],ymm4[6],ymm5[6],ymm4[7],ymm5[7]12653; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} ymm4 = ymm6[2],ymm0[2],ymm6[3],ymm0[3],ymm6[6],ymm0[6],ymm6[7],ymm0[7]12654; AVX2-FCP-NEXT:    vextractf128 $1, %ymm3, %xmm012655; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm5 = ymm4[2,2,2,2]12656; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm5[2,3]12657; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm2 = ymm1[0,1,2,3,4,5],ymm2[6,7]12658; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]12659; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12660; AVX2-FCP-NEXT:    vbroadcastss 92(%r10), %ymm012661; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} ymm2 = ymm0[2],ymm9[2],ymm0[3],ymm9[3],ymm0[6],ymm9[6],ymm0[7],ymm9[7]12662; AVX2-FCP-NEXT:    vmovaps 96(%rdi), %ymm012663; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm5 = ymm1[1],ymm2[1],ymm1[3],ymm2[3]12664; AVX2-FCP-NEXT:    vmovaps 96(%rsi), %ymm112665; AVX2-FCP-NEXT:    vextractf128 $1, %ymm4, %xmm212666; AVX2-FCP-NEXT:    vshufps {{.*#+}} ymm3 = ymm3[2,3,2,3,6,7,6,7]12667; AVX2-FCP-NEXT:    vextractf128 $1, %ymm3, %xmm312668; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm3 = xmm3[0,1],xmm2[2,3]12669; AVX2-FCP-NEXT:    vmovaps 96(%rdx), %ymm212670; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm5[4,5,6,7]12671; AVX2-FCP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12672; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} ymm3 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]12673; AVX2-FCP-NEXT:    vextractf128 $1, %ymm3, %xmm312674; AVX2-FCP-NEXT:    vmovaps 96(%rcx), %ymm412675; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} ymm5 = ymm2[0],ymm4[0],ymm2[1],ymm4[1],ymm2[4],ymm4[4],ymm2[5],ymm4[5]12676; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm6 = ymm5[2,2,2,2]12677; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm3 = xmm3[0,1],xmm6[2,3]12678; AVX2-FCP-NEXT:    vmovaps 96(%r8), %ymm612679; AVX2-FCP-NEXT:    vmovaps 96(%r9), %ymm712680; AVX2-FCP-NEXT:    vmovaps 96(%r10), %ymm812681; AVX2-FCP-NEXT:    vmovaps 96(%rax), %ymm912682; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} ymm10 = ymm8[0],ymm9[0],ymm8[1],ymm9[1],ymm8[4],ymm9[4],ymm8[5],ymm9[5]12683; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} ymm11 = ymm6[0],ymm7[0],ymm6[1],ymm7[1],ymm6[4],ymm7[4],ymm6[5],ymm7[5]12684; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm11 = ymm11[0],ymm10[0],ymm11[2],ymm10[2]12685; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm11[4,5,6,7]12686; AVX2-FCP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12687; AVX2-FCP-NEXT:    vbroadcastss 116(%r8), %ymm312688; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4],ymm7[5],ymm3[6,7]12689; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm10[6,7]12690; AVX2-FCP-NEXT:    vextractf128 $1, %ymm5, %xmm512691; AVX2-FCP-NEXT:    vshufps {{.*#+}} ymm10 = ymm0[1,1,1,1,5,5,5,5]12692; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm10 = ymm10[0],ymm1[1],ymm10[2,3,4],ymm1[5],ymm10[6,7]12693; AVX2-FCP-NEXT:    vextractf128 $1, %ymm10, %xmm1012694; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm5 = xmm10[0,1],xmm5[2,3]12695; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm3 = ymm5[0,1,2,3],ymm3[4,5,6,7]12696; AVX2-FCP-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12697; AVX2-FCP-NEXT:    vbroadcastss 120(%rax), %ymm312698; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm3 = ymm8[0,1,2,3,4,5,6],ymm3[7]12699; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} ymm5 = ymm6[2],ymm7[2],ymm6[3],ymm7[3],ymm6[6],ymm7[6],ymm6[7],ymm7[7]12700; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} ymm0 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7]12701; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} ymm1 = ymm2[2],ymm4[2],ymm2[3],ymm4[3],ymm2[6],ymm4[6],ymm2[7],ymm4[7]12702; AVX2-FCP-NEXT:    vextractf128 $1, %ymm0, %xmm212703; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm4 = ymm1[2,2,2,2]12704; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm2 = xmm2[0,1],xmm4[2,3]12705; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm3 = ymm5[0,1,2,3,4,5],ymm3[6,7]12706; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4,5,6,7]12707; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12708; AVX2-FCP-NEXT:    vbroadcastss 124(%r10), %ymm212709; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} ymm2 = ymm2[2],ymm9[2],ymm2[3],ymm9[3],ymm2[6],ymm9[6],ymm2[7],ymm9[7]12710; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm5[1],ymm2[1],ymm5[3],ymm2[3]12711; AVX2-FCP-NEXT:    vextractf128 $1, %ymm1, %xmm112712; AVX2-FCP-NEXT:    vshufps {{.*#+}} ymm0 = ymm0[2,3,2,3,6,7,6,7]12713; AVX2-FCP-NEXT:    vextractf128 $1, %ymm0, %xmm012714; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]12715; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]12716; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12717; AVX2-FCP-NEXT:    vmovaps 128(%rdi), %ymm012718; AVX2-FCP-NEXT:    vmovaps 128(%rsi), %ymm112719; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]12720; AVX2-FCP-NEXT:    vextractf128 $1, %ymm2, %xmm212721; AVX2-FCP-NEXT:    vmovaps 128(%rdx), %ymm312722; AVX2-FCP-NEXT:    vmovaps 128(%rcx), %ymm412723; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} ymm5 = ymm3[0],ymm4[0],ymm3[1],ymm4[1],ymm3[4],ymm4[4],ymm3[5],ymm4[5]12724; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm6 = ymm5[2,2,2,2]12725; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm2 = xmm2[0,1],xmm6[2,3]12726; AVX2-FCP-NEXT:    vmovaps 128(%r8), %ymm612727; AVX2-FCP-NEXT:    vmovaps 128(%r9), %ymm712728; AVX2-FCP-NEXT:    vmovaps 128(%r10), %ymm812729; AVX2-FCP-NEXT:    vmovaps 128(%rax), %ymm912730; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} ymm10 = ymm8[0],ymm9[0],ymm8[1],ymm9[1],ymm8[4],ymm9[4],ymm8[5],ymm9[5]12731; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} ymm11 = ymm6[0],ymm7[0],ymm6[1],ymm7[1],ymm6[4],ymm7[4],ymm6[5],ymm7[5]12732; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm11 = ymm11[0],ymm10[0],ymm11[2],ymm10[2]12733; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm11[4,5,6,7]12734; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12735; AVX2-FCP-NEXT:    vbroadcastss 148(%r8), %ymm212736; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4],ymm7[5],ymm2[6,7]12737; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm10[6,7]12738; AVX2-FCP-NEXT:    vextractf128 $1, %ymm5, %xmm512739; AVX2-FCP-NEXT:    vshufps {{.*#+}} ymm10 = ymm0[1,1,1,1,5,5,5,5]12740; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm10 = ymm10[0],ymm1[1],ymm10[2,3,4],ymm1[5],ymm10[6,7]12741; AVX2-FCP-NEXT:    vextractf128 $1, %ymm10, %xmm1012742; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm5 = xmm10[0,1],xmm5[2,3]12743; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm2 = ymm5[0,1,2,3],ymm2[4,5,6,7]12744; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12745; AVX2-FCP-NEXT:    vbroadcastss 152(%rax), %ymm212746; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm2 = ymm8[0,1,2,3,4,5,6],ymm2[7]12747; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} ymm5 = ymm6[2],ymm7[2],ymm6[3],ymm7[3],ymm6[6],ymm7[6],ymm6[7],ymm7[7]12748; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} ymm0 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7]12749; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} ymm1 = ymm3[2],ymm4[2],ymm3[3],ymm4[3],ymm3[6],ymm4[6],ymm3[7],ymm4[7]12750; AVX2-FCP-NEXT:    vextractf128 $1, %ymm0, %xmm312751; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm4 = ymm1[2,2,2,2]12752; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm3 = xmm3[0,1],xmm4[2,3]12753; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm2 = ymm5[0,1,2,3,4,5],ymm2[6,7]12754; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm2 = ymm3[0,1,2,3],ymm2[4,5,6,7]12755; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12756; AVX2-FCP-NEXT:    vbroadcastss 156(%r10), %ymm212757; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} ymm2 = ymm2[2],ymm9[2],ymm2[3],ymm9[3],ymm2[6],ymm9[6],ymm2[7],ymm9[7]12758; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm5[1],ymm2[1],ymm5[3],ymm2[3]12759; AVX2-FCP-NEXT:    vextractf128 $1, %ymm1, %xmm112760; AVX2-FCP-NEXT:    vshufps {{.*#+}} ymm0 = ymm0[2,3,2,3,6,7,6,7]12761; AVX2-FCP-NEXT:    vextractf128 $1, %ymm0, %xmm012762; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]12763; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]12764; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12765; AVX2-FCP-NEXT:    vmovaps 160(%rdi), %ymm012766; AVX2-FCP-NEXT:    vmovaps 160(%rsi), %ymm112767; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]12768; AVX2-FCP-NEXT:    vextractf128 $1, %ymm2, %xmm212769; AVX2-FCP-NEXT:    vmovaps 160(%rdx), %ymm312770; AVX2-FCP-NEXT:    vmovaps 160(%rcx), %ymm412771; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} ymm5 = ymm3[0],ymm4[0],ymm3[1],ymm4[1],ymm3[4],ymm4[4],ymm3[5],ymm4[5]12772; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm6 = ymm5[2,2,2,2]12773; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm2 = xmm2[0,1],xmm6[2,3]12774; AVX2-FCP-NEXT:    vmovaps 160(%r8), %ymm612775; AVX2-FCP-NEXT:    vmovaps 160(%r9), %ymm712776; AVX2-FCP-NEXT:    vmovaps 160(%r10), %ymm812777; AVX2-FCP-NEXT:    vmovaps 160(%rax), %ymm912778; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} ymm10 = ymm8[0],ymm9[0],ymm8[1],ymm9[1],ymm8[4],ymm9[4],ymm8[5],ymm9[5]12779; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} ymm11 = ymm6[0],ymm7[0],ymm6[1],ymm7[1],ymm6[4],ymm7[4],ymm6[5],ymm7[5]12780; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm11 = ymm11[0],ymm10[0],ymm11[2],ymm10[2]12781; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm11[4,5,6,7]12782; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12783; AVX2-FCP-NEXT:    vbroadcastss 180(%r8), %ymm212784; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4],ymm7[5],ymm2[6,7]12785; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm10[6,7]12786; AVX2-FCP-NEXT:    vextractf128 $1, %ymm5, %xmm512787; AVX2-FCP-NEXT:    vshufps {{.*#+}} ymm10 = ymm0[1,1,1,1,5,5,5,5]12788; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm10 = ymm10[0],ymm1[1],ymm10[2,3,4],ymm1[5],ymm10[6,7]12789; AVX2-FCP-NEXT:    vextractf128 $1, %ymm10, %xmm1012790; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm5 = xmm10[0,1],xmm5[2,3]12791; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm2 = ymm5[0,1,2,3],ymm2[4,5,6,7]12792; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12793; AVX2-FCP-NEXT:    vbroadcastss 184(%rax), %ymm212794; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm2 = ymm8[0,1,2,3,4,5,6],ymm2[7]12795; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} ymm5 = ymm6[2],ymm7[2],ymm6[3],ymm7[3],ymm6[6],ymm7[6],ymm6[7],ymm7[7]12796; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} ymm0 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7]12797; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} ymm1 = ymm3[2],ymm4[2],ymm3[3],ymm4[3],ymm3[6],ymm4[6],ymm3[7],ymm4[7]12798; AVX2-FCP-NEXT:    vextractf128 $1, %ymm0, %xmm312799; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm4 = ymm1[2,2,2,2]12800; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm3 = xmm3[0,1],xmm4[2,3]12801; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm2 = ymm5[0,1,2,3,4,5],ymm2[6,7]12802; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm2 = ymm3[0,1,2,3],ymm2[4,5,6,7]12803; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12804; AVX2-FCP-NEXT:    vbroadcastss 188(%r10), %ymm212805; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} ymm2 = ymm2[2],ymm9[2],ymm2[3],ymm9[3],ymm2[6],ymm9[6],ymm2[7],ymm9[7]12806; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm5[1],ymm2[1],ymm5[3],ymm2[3]12807; AVX2-FCP-NEXT:    vextractf128 $1, %ymm1, %xmm112808; AVX2-FCP-NEXT:    vshufps {{.*#+}} ymm0 = ymm0[2,3,2,3,6,7,6,7]12809; AVX2-FCP-NEXT:    vextractf128 $1, %ymm0, %xmm012810; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]12811; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]12812; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12813; AVX2-FCP-NEXT:    vmovaps 192(%rdi), %ymm012814; AVX2-FCP-NEXT:    vmovaps 192(%rsi), %ymm112815; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]12816; AVX2-FCP-NEXT:    vextractf128 $1, %ymm2, %xmm212817; AVX2-FCP-NEXT:    vmovaps 192(%rdx), %ymm312818; AVX2-FCP-NEXT:    vmovaps 192(%rcx), %ymm412819; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} ymm5 = ymm3[0],ymm4[0],ymm3[1],ymm4[1],ymm3[4],ymm4[4],ymm3[5],ymm4[5]12820; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm6 = ymm5[2,2,2,2]12821; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm2 = xmm2[0,1],xmm6[2,3]12822; AVX2-FCP-NEXT:    vmovaps 192(%r8), %ymm612823; AVX2-FCP-NEXT:    vmovaps 192(%r9), %ymm712824; AVX2-FCP-NEXT:    vmovaps 192(%r10), %ymm812825; AVX2-FCP-NEXT:    vmovaps 192(%rax), %ymm912826; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} ymm10 = ymm8[0],ymm9[0],ymm8[1],ymm9[1],ymm8[4],ymm9[4],ymm8[5],ymm9[5]12827; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} ymm11 = ymm6[0],ymm7[0],ymm6[1],ymm7[1],ymm6[4],ymm7[4],ymm6[5],ymm7[5]12828; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm11 = ymm11[0],ymm10[0],ymm11[2],ymm10[2]12829; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm11[4,5,6,7]12830; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12831; AVX2-FCP-NEXT:    vbroadcastss 212(%r8), %ymm212832; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4],ymm7[5],ymm2[6,7]12833; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm10[6,7]12834; AVX2-FCP-NEXT:    vextractf128 $1, %ymm5, %xmm512835; AVX2-FCP-NEXT:    vshufps {{.*#+}} ymm10 = ymm0[1,1,1,1,5,5,5,5]12836; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm10 = ymm10[0],ymm1[1],ymm10[2,3,4],ymm1[5],ymm10[6,7]12837; AVX2-FCP-NEXT:    vextractf128 $1, %ymm10, %xmm1012838; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm5 = xmm10[0,1],xmm5[2,3]12839; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm2 = ymm5[0,1,2,3],ymm2[4,5,6,7]12840; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12841; AVX2-FCP-NEXT:    vbroadcastss 216(%rax), %ymm212842; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm2 = ymm8[0,1,2,3,4,5,6],ymm2[7]12843; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} ymm5 = ymm6[2],ymm7[2],ymm6[3],ymm7[3],ymm6[6],ymm7[6],ymm6[7],ymm7[7]12844; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} ymm0 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7]12845; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} ymm1 = ymm3[2],ymm4[2],ymm3[3],ymm4[3],ymm3[6],ymm4[6],ymm3[7],ymm4[7]12846; AVX2-FCP-NEXT:    vextractf128 $1, %ymm0, %xmm312847; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm4 = ymm1[2,2,2,2]12848; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm3 = xmm3[0,1],xmm4[2,3]12849; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm2 = ymm5[0,1,2,3,4,5],ymm2[6,7]12850; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm2 = ymm3[0,1,2,3],ymm2[4,5,6,7]12851; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12852; AVX2-FCP-NEXT:    vbroadcastss 220(%r10), %ymm212853; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} ymm2 = ymm2[2],ymm9[2],ymm2[3],ymm9[3],ymm2[6],ymm9[6],ymm2[7],ymm9[7]12854; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm5[1],ymm2[1],ymm5[3],ymm2[3]12855; AVX2-FCP-NEXT:    vextractf128 $1, %ymm1, %xmm112856; AVX2-FCP-NEXT:    vshufps {{.*#+}} ymm0 = ymm0[2,3,2,3,6,7,6,7]12857; AVX2-FCP-NEXT:    vextractf128 $1, %ymm0, %xmm012858; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]12859; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]12860; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12861; AVX2-FCP-NEXT:    vmovaps 224(%rdi), %ymm012862; AVX2-FCP-NEXT:    vmovaps 224(%rsi), %ymm112863; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} ymm2 = ymm0[0],ymm1[0],ymm0[1],ymm1[1],ymm0[4],ymm1[4],ymm0[5],ymm1[5]12864; AVX2-FCP-NEXT:    vextractf128 $1, %ymm2, %xmm212865; AVX2-FCP-NEXT:    vmovaps 224(%rdx), %ymm312866; AVX2-FCP-NEXT:    vmovaps 224(%rcx), %ymm412867; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} ymm5 = ymm3[0],ymm4[0],ymm3[1],ymm4[1],ymm3[4],ymm4[4],ymm3[5],ymm4[5]12868; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm6 = ymm5[2,2,2,2]12869; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm2 = xmm2[0,1],xmm6[2,3]12870; AVX2-FCP-NEXT:    vmovaps 224(%r8), %ymm612871; AVX2-FCP-NEXT:    vmovaps 224(%r9), %ymm712872; AVX2-FCP-NEXT:    vmovaps 224(%r10), %ymm812873; AVX2-FCP-NEXT:    vmovaps 224(%rax), %ymm912874; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} ymm10 = ymm8[0],ymm9[0],ymm8[1],ymm9[1],ymm8[4],ymm9[4],ymm8[5],ymm9[5]12875; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} ymm11 = ymm6[0],ymm7[0],ymm6[1],ymm7[1],ymm6[4],ymm7[4],ymm6[5],ymm7[5]12876; AVX2-FCP-NEXT:    vunpcklpd {{.*#+}} ymm11 = ymm11[0],ymm10[0],ymm11[2],ymm10[2]12877; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm11[4,5,6,7]12878; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12879; AVX2-FCP-NEXT:    vbroadcastss 244(%r8), %ymm212880; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4],ymm7[5],ymm2[6,7]12881; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3,4,5],ymm10[6,7]12882; AVX2-FCP-NEXT:    vextractf128 $1, %ymm5, %xmm512883; AVX2-FCP-NEXT:    vshufps {{.*#+}} ymm10 = ymm0[1,1,1,1,5,5,5,5]12884; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm10 = ymm10[0],ymm1[1],ymm10[2,3,4],ymm1[5],ymm10[6,7]12885; AVX2-FCP-NEXT:    vextractf128 $1, %ymm10, %xmm1012886; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm5 = xmm10[0,1],xmm5[2,3]12887; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm2 = ymm5[0,1,2,3],ymm2[4,5,6,7]12888; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12889; AVX2-FCP-NEXT:    vbroadcastss 248(%rax), %ymm212890; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm2 = ymm8[0,1,2,3,4,5,6],ymm2[7]12891; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} ymm5 = ymm6[2],ymm7[2],ymm6[3],ymm7[3],ymm6[6],ymm7[6],ymm6[7],ymm7[7]12892; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} ymm0 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7]12893; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} ymm1 = ymm3[2],ymm4[2],ymm3[3],ymm4[3],ymm3[6],ymm4[6],ymm3[7],ymm4[7]12894; AVX2-FCP-NEXT:    vextractf128 $1, %ymm0, %xmm312895; AVX2-FCP-NEXT:    vpermpd {{.*#+}} ymm4 = ymm1[2,2,2,2]12896; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm3 = xmm3[0,1],xmm4[2,3]12897; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm2 = ymm5[0,1,2,3,4,5],ymm2[6,7]12898; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm2 = ymm3[0,1,2,3],ymm2[4,5,6,7]12899; AVX2-FCP-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12900; AVX2-FCP-NEXT:    vbroadcastss 252(%r10), %ymm212901; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} ymm2 = ymm2[2],ymm9[2],ymm2[3],ymm9[3],ymm2[6],ymm9[6],ymm2[7],ymm9[7]12902; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} ymm2 = ymm5[1],ymm2[1],ymm5[3],ymm2[3]12903; AVX2-FCP-NEXT:    vextractf128 $1, %ymm1, %xmm112904; AVX2-FCP-NEXT:    vshufps {{.*#+}} ymm0 = ymm0[2,3,2,3,6,7,6,7]12905; AVX2-FCP-NEXT:    vextractf128 $1, %ymm0, %xmm012906; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3]12907; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]12908; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12909; AVX2-FCP-NEXT:    vmovaps (%rcx), %xmm012910; AVX2-FCP-NEXT:    vbroadcastss %xmm0, %xmm212911; AVX2-FCP-NEXT:    vmovaps (%rdx), %xmm112912; AVX2-FCP-NEXT:    vbroadcastss %xmm1, %xmm312913; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]12914; AVX2-FCP-NEXT:    vmovaps (%rsi), %xmm412915; AVX2-FCP-NEXT:    vmovaps (%rdi), %xmm512916; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm3 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]12917; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm6 = xmm3[0,1],xmm2[2,3]12918; AVX2-FCP-NEXT:    vmovaps (%rax), %xmm212919; AVX2-FCP-NEXT:    vbroadcastss %xmm2, %xmm712920; AVX2-FCP-NEXT:    vmovaps (%r10), %xmm312921; AVX2-FCP-NEXT:    vbroadcastss %xmm3, %xmm812922; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm8[0],xmm7[0],xmm8[1],xmm7[1]12923; AVX2-FCP-NEXT:    vmovaps (%r9), %xmm812924; AVX2-FCP-NEXT:    vmovaps (%r8), %xmm912925; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm10 = xmm9[0],xmm8[0],xmm9[1],xmm8[1]12926; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm10, %ymm0, %ymm1012927; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm7, %ymm0, %ymm712928; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm7 = ymm10[0,1,2,3,4,5],ymm7[6,7]12929; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7]12930; AVX2-FCP-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12931; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm6 = xmm5[1,1,1,1]12932; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm6 = xmm6[0],xmm4[1],xmm6[2,3]12933; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]12934; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm6 = xmm6[0,1],xmm7[2,3]12935; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]12936; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm7, %ymm0, %ymm712937; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm10 = xmm9[1,1,1,1]12938; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm10 = xmm10[0],xmm8[1],xmm10[2,3]12939; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm10, %ymm0, %ymm1012940; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm7 = ymm10[0,1,2,3,4,5],ymm7[6,7]12941; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7]12942; AVX2-FCP-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12943; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} xmm4 = xmm5[2],xmm4[2],xmm5[3],xmm4[3]12944; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} xmm5 = xmm9[2],xmm8[2],xmm9[3],xmm8[3]12945; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm6 = xmm2[2,2,2,2]12946; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm6 = xmm3[0,1,2],xmm6[3]12947; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm6, %ymm0, %ymm612948; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm5, %ymm0, %ymm712949; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3,4,5],ymm6[6,7]12950; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm7 = xmm0[2,2,2,2]12951; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm7 = xmm1[0,1,2],xmm7[3]12952; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm7 = xmm4[0,1],xmm7[2,3]12953; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]12954; AVX2-FCP-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12955; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm1[2],xmm0[2],xmm1[3],xmm0[3]12956; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm4[1],xmm0[1]12957; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} xmm1 = xmm3[2],xmm2[2],xmm3[3],xmm2[3]12958; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm112959; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm2 = xmm5[2,3,2,3]12960; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm212961; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]12962; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]12963; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12964; AVX2-FCP-NEXT:    vmovaps 32(%rcx), %xmm012965; AVX2-FCP-NEXT:    vmovaps 32(%rdx), %xmm112966; AVX2-FCP-NEXT:    vbroadcastss %xmm0, %xmm212967; AVX2-FCP-NEXT:    vbroadcastss %xmm1, %xmm312968; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]12969; AVX2-FCP-NEXT:    vmovaps 32(%rsi), %xmm412970; AVX2-FCP-NEXT:    vmovaps 32(%rdi), %xmm512971; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm3 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]12972; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm6 = xmm3[0,1],xmm2[2,3]12973; AVX2-FCP-NEXT:    vmovaps 32(%r10), %xmm212974; AVX2-FCP-NEXT:    vmovaps 32(%rax), %xmm312975; AVX2-FCP-NEXT:    vbroadcastss %xmm3, %xmm712976; AVX2-FCP-NEXT:    vbroadcastss %xmm2, %xmm812977; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm8[0],xmm7[0],xmm8[1],xmm7[1]12978; AVX2-FCP-NEXT:    vmovaps 32(%r9), %xmm812979; AVX2-FCP-NEXT:    vmovaps 32(%r8), %xmm912980; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm10 = xmm9[0],xmm8[0],xmm9[1],xmm8[1]12981; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm10, %ymm0, %ymm1012982; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm7, %ymm0, %ymm712983; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm7 = ymm10[0,1,2,3,4,5],ymm7[6,7]12984; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7]12985; AVX2-FCP-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12986; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm6 = xmm5[1,1,1,1]12987; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm6 = xmm6[0],xmm4[1],xmm6[2,3]12988; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]12989; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm6 = xmm6[0,1],xmm7[2,3]12990; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm2[0],xmm3[0],xmm2[1],xmm3[1]12991; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm7, %ymm0, %ymm712992; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm10 = xmm9[1,1,1,1]12993; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm10 = xmm10[0],xmm8[1],xmm10[2,3]12994; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm10, %ymm0, %ymm1012995; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm7 = ymm10[0,1,2,3,4,5],ymm7[6,7]12996; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7]12997; AVX2-FCP-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill12998; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} xmm4 = xmm5[2],xmm4[2],xmm5[3],xmm4[3]12999; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} xmm5 = xmm9[2],xmm8[2],xmm9[3],xmm8[3]13000; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm6 = xmm3[2,2,2,2]13001; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm6 = xmm2[0,1,2],xmm6[3]13002; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm6, %ymm0, %ymm613003; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm5, %ymm0, %ymm713004; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3,4,5],ymm6[6,7]13005; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm7 = xmm0[2,2,2,2]13006; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm7 = xmm1[0,1,2],xmm7[3]13007; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm7 = xmm4[0,1],xmm7[2,3]13008; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]13009; AVX2-FCP-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13010; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm1[2],xmm0[2],xmm1[3],xmm0[3]13011; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm4[1],xmm0[1]13012; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} xmm1 = xmm2[2],xmm3[2],xmm2[3],xmm3[3]13013; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm113014; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm2 = xmm5[2,3,2,3]13015; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm213016; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]13017; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]13018; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13019; AVX2-FCP-NEXT:    vmovaps 64(%rcx), %xmm013020; AVX2-FCP-NEXT:    vmovaps 64(%rdx), %xmm113021; AVX2-FCP-NEXT:    vbroadcastss %xmm0, %xmm213022; AVX2-FCP-NEXT:    vbroadcastss %xmm1, %xmm313023; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]13024; AVX2-FCP-NEXT:    vmovaps 64(%rsi), %xmm413025; AVX2-FCP-NEXT:    vmovaps 64(%rdi), %xmm513026; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm3 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]13027; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm6 = xmm3[0,1],xmm2[2,3]13028; AVX2-FCP-NEXT:    vmovaps 64(%r10), %xmm213029; AVX2-FCP-NEXT:    vmovaps 64(%rax), %xmm313030; AVX2-FCP-NEXT:    vbroadcastss %xmm3, %xmm713031; AVX2-FCP-NEXT:    vbroadcastss %xmm2, %xmm813032; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm8[0],xmm7[0],xmm8[1],xmm7[1]13033; AVX2-FCP-NEXT:    vmovaps 64(%r9), %xmm813034; AVX2-FCP-NEXT:    vmovaps 64(%r8), %xmm913035; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm10 = xmm9[0],xmm8[0],xmm9[1],xmm8[1]13036; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm10, %ymm0, %ymm1013037; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm7, %ymm0, %ymm713038; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm7 = ymm10[0,1,2,3,4,5],ymm7[6,7]13039; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7]13040; AVX2-FCP-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13041; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm6 = xmm5[1,1,1,1]13042; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm6 = xmm6[0],xmm4[1],xmm6[2,3]13043; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]13044; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm6 = xmm6[0,1],xmm7[2,3]13045; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm2[0],xmm3[0],xmm2[1],xmm3[1]13046; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm7, %ymm0, %ymm713047; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm10 = xmm9[1,1,1,1]13048; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm10 = xmm10[0],xmm8[1],xmm10[2,3]13049; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm10, %ymm0, %ymm1013050; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm7 = ymm10[0,1,2,3,4,5],ymm7[6,7]13051; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7]13052; AVX2-FCP-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13053; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} xmm4 = xmm5[2],xmm4[2],xmm5[3],xmm4[3]13054; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} xmm5 = xmm9[2],xmm8[2],xmm9[3],xmm8[3]13055; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm6 = xmm3[2,2,2,2]13056; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm6 = xmm2[0,1,2],xmm6[3]13057; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm6, %ymm0, %ymm613058; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm5, %ymm0, %ymm713059; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3,4,5],ymm6[6,7]13060; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm7 = xmm0[2,2,2,2]13061; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm7 = xmm1[0,1,2],xmm7[3]13062; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm7 = xmm4[0,1],xmm7[2,3]13063; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]13064; AVX2-FCP-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13065; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm1[2],xmm0[2],xmm1[3],xmm0[3]13066; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm4[1],xmm0[1]13067; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} xmm1 = xmm2[2],xmm3[2],xmm2[3],xmm3[3]13068; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm113069; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm2 = xmm5[2,3,2,3]13070; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm213071; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]13072; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]13073; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13074; AVX2-FCP-NEXT:    vmovaps 96(%rcx), %xmm013075; AVX2-FCP-NEXT:    vbroadcastss %xmm0, %xmm213076; AVX2-FCP-NEXT:    vmovaps 96(%rdx), %xmm113077; AVX2-FCP-NEXT:    vbroadcastss %xmm1, %xmm313078; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]13079; AVX2-FCP-NEXT:    vmovaps 96(%rsi), %xmm413080; AVX2-FCP-NEXT:    vmovaps 96(%rdi), %xmm513081; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm3 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]13082; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm6 = xmm3[0,1],xmm2[2,3]13083; AVX2-FCP-NEXT:    vmovaps 96(%rax), %xmm213084; AVX2-FCP-NEXT:    vbroadcastss %xmm2, %xmm713085; AVX2-FCP-NEXT:    vmovaps 96(%r10), %xmm313086; AVX2-FCP-NEXT:    vbroadcastss %xmm3, %xmm813087; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm8[0],xmm7[0],xmm8[1],xmm7[1]13088; AVX2-FCP-NEXT:    vmovaps 96(%r9), %xmm813089; AVX2-FCP-NEXT:    vmovaps 96(%r8), %xmm913090; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm10 = xmm9[0],xmm8[0],xmm9[1],xmm8[1]13091; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm10, %ymm0, %ymm1013092; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm7, %ymm0, %ymm713093; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm7 = ymm10[0,1,2,3,4,5],ymm7[6,7]13094; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7]13095; AVX2-FCP-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13096; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm6 = xmm5[1,1,1,1]13097; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm6 = xmm6[0],xmm4[1],xmm6[2,3]13098; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]13099; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm6 = xmm6[0,1],xmm7[2,3]13100; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]13101; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm7, %ymm0, %ymm713102; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm10 = xmm9[1,1,1,1]13103; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm10 = xmm10[0],xmm8[1],xmm10[2,3]13104; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm10, %ymm0, %ymm1013105; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm7 = ymm10[0,1,2,3,4,5],ymm7[6,7]13106; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7]13107; AVX2-FCP-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13108; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} xmm4 = xmm5[2],xmm4[2],xmm5[3],xmm4[3]13109; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} xmm5 = xmm9[2],xmm8[2],xmm9[3],xmm8[3]13110; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm6 = xmm2[2,2,2,2]13111; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm6 = xmm3[0,1,2],xmm6[3]13112; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm6, %ymm0, %ymm613113; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm5, %ymm0, %ymm713114; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3,4,5],ymm6[6,7]13115; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm7 = xmm0[2,2,2,2]13116; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm7 = xmm1[0,1,2],xmm7[3]13117; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm7 = xmm4[0,1],xmm7[2,3]13118; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]13119; AVX2-FCP-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13120; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm1[2],xmm0[2],xmm1[3],xmm0[3]13121; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm4[1],xmm0[1]13122; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} xmm1 = xmm3[2],xmm2[2],xmm3[3],xmm2[3]13123; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm113124; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm2 = xmm5[2,3,2,3]13125; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm213126; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]13127; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]13128; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13129; AVX2-FCP-NEXT:    vmovaps 128(%rcx), %xmm013130; AVX2-FCP-NEXT:    vbroadcastss %xmm0, %xmm213131; AVX2-FCP-NEXT:    vmovaps 128(%rdx), %xmm113132; AVX2-FCP-NEXT:    vbroadcastss %xmm1, %xmm313133; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]13134; AVX2-FCP-NEXT:    vmovaps 128(%rsi), %xmm413135; AVX2-FCP-NEXT:    vmovaps 128(%rdi), %xmm513136; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm3 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]13137; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm6 = xmm3[0,1],xmm2[2,3]13138; AVX2-FCP-NEXT:    vmovaps 128(%rax), %xmm213139; AVX2-FCP-NEXT:    vbroadcastss %xmm2, %xmm713140; AVX2-FCP-NEXT:    vmovaps 128(%r10), %xmm313141; AVX2-FCP-NEXT:    vbroadcastss %xmm3, %xmm813142; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm8[0],xmm7[0],xmm8[1],xmm7[1]13143; AVX2-FCP-NEXT:    vmovaps 128(%r9), %xmm813144; AVX2-FCP-NEXT:    vmovaps 128(%r8), %xmm913145; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm10 = xmm9[0],xmm8[0],xmm9[1],xmm8[1]13146; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm10, %ymm0, %ymm1013147; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm7, %ymm0, %ymm713148; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm7 = ymm10[0,1,2,3,4,5],ymm7[6,7]13149; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7]13150; AVX2-FCP-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13151; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm6 = xmm5[1,1,1,1]13152; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm6 = xmm6[0],xmm4[1],xmm6[2,3]13153; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]13154; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm6 = xmm6[0,1],xmm7[2,3]13155; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]13156; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm7, %ymm0, %ymm713157; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm10 = xmm9[1,1,1,1]13158; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm10 = xmm10[0],xmm8[1],xmm10[2,3]13159; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm10, %ymm0, %ymm1013160; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm7 = ymm10[0,1,2,3,4,5],ymm7[6,7]13161; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7]13162; AVX2-FCP-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13163; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} xmm4 = xmm5[2],xmm4[2],xmm5[3],xmm4[3]13164; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} xmm5 = xmm9[2],xmm8[2],xmm9[3],xmm8[3]13165; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm6 = xmm2[2,2,2,2]13166; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm6 = xmm3[0,1,2],xmm6[3]13167; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm6, %ymm0, %ymm613168; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm5, %ymm0, %ymm713169; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3,4,5],ymm6[6,7]13170; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm7 = xmm0[2,2,2,2]13171; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm7 = xmm1[0,1,2],xmm7[3]13172; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm7 = xmm4[0,1],xmm7[2,3]13173; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]13174; AVX2-FCP-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13175; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm1[2],xmm0[2],xmm1[3],xmm0[3]13176; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm4[1],xmm0[1]13177; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} xmm1 = xmm3[2],xmm2[2],xmm3[3],xmm2[3]13178; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm113179; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm2 = xmm5[2,3,2,3]13180; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm213181; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]13182; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]13183; AVX2-FCP-NEXT:    vmovups %ymm0, (%rsp) # 32-byte Spill13184; AVX2-FCP-NEXT:    vmovaps 160(%rcx), %xmm013185; AVX2-FCP-NEXT:    vbroadcastss %xmm0, %xmm213186; AVX2-FCP-NEXT:    vmovaps 160(%rdx), %xmm113187; AVX2-FCP-NEXT:    vbroadcastss %xmm1, %xmm313188; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]13189; AVX2-FCP-NEXT:    vmovaps 160(%rsi), %xmm413190; AVX2-FCP-NEXT:    vmovaps 160(%rdi), %xmm513191; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm3 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]13192; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm6 = xmm3[0,1],xmm2[2,3]13193; AVX2-FCP-NEXT:    vmovaps 160(%rax), %xmm213194; AVX2-FCP-NEXT:    vbroadcastss %xmm2, %xmm713195; AVX2-FCP-NEXT:    vmovaps 160(%r10), %xmm313196; AVX2-FCP-NEXT:    vbroadcastss %xmm3, %xmm813197; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm8[0],xmm7[0],xmm8[1],xmm7[1]13198; AVX2-FCP-NEXT:    vmovaps 160(%r9), %xmm813199; AVX2-FCP-NEXT:    vmovaps 160(%r8), %xmm913200; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm10 = xmm9[0],xmm8[0],xmm9[1],xmm8[1]13201; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm10, %ymm0, %ymm1013202; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm7, %ymm0, %ymm713203; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm7 = ymm10[0,1,2,3,4,5],ymm7[6,7]13204; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7]13205; AVX2-FCP-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13206; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm6 = xmm5[1,1,1,1]13207; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm6 = xmm6[0],xmm4[1],xmm6[2,3]13208; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]13209; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm6 = xmm6[0,1],xmm7[2,3]13210; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]13211; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm7, %ymm0, %ymm713212; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm10 = xmm9[1,1,1,1]13213; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm10 = xmm10[0],xmm8[1],xmm10[2,3]13214; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm10, %ymm0, %ymm1013215; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm7 = ymm10[0,1,2,3,4,5],ymm7[6,7]13216; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7]13217; AVX2-FCP-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13218; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} xmm4 = xmm5[2],xmm4[2],xmm5[3],xmm4[3]13219; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} xmm5 = xmm9[2],xmm8[2],xmm9[3],xmm8[3]13220; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm6 = xmm2[2,2,2,2]13221; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm6 = xmm3[0,1,2],xmm6[3]13222; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm6, %ymm0, %ymm613223; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm5, %ymm0, %ymm713224; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3,4,5],ymm6[6,7]13225; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm7 = xmm0[2,2,2,2]13226; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm7 = xmm1[0,1,2],xmm7[3]13227; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm7 = xmm4[0,1],xmm7[2,3]13228; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]13229; AVX2-FCP-NEXT:    vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13230; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm1[2],xmm0[2],xmm1[3],xmm0[3]13231; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm4[1],xmm0[1]13232; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} xmm1 = xmm3[2],xmm2[2],xmm3[3],xmm2[3]13233; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm113234; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm2 = xmm5[2,3,2,3]13235; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm213236; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]13237; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]13238; AVX2-FCP-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill13239; AVX2-FCP-NEXT:    vmovaps 192(%rcx), %xmm013240; AVX2-FCP-NEXT:    vbroadcastss %xmm0, %xmm213241; AVX2-FCP-NEXT:    vmovaps 192(%rdx), %xmm113242; AVX2-FCP-NEXT:    vbroadcastss %xmm1, %xmm313243; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]13244; AVX2-FCP-NEXT:    vmovaps 192(%rsi), %xmm413245; AVX2-FCP-NEXT:    vmovaps 192(%rdi), %xmm513246; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm3 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]13247; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm9 = xmm3[0,1],xmm2[2,3]13248; AVX2-FCP-NEXT:    vmovaps 192(%rax), %xmm213249; AVX2-FCP-NEXT:    vbroadcastss %xmm2, %xmm713250; AVX2-FCP-NEXT:    vmovaps 192(%r10), %xmm313251; AVX2-FCP-NEXT:    vbroadcastss %xmm3, %xmm813252; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm8[0],xmm7[0],xmm8[1],xmm7[1]13253; AVX2-FCP-NEXT:    vmovaps 192(%r9), %xmm813254; AVX2-FCP-NEXT:    vmovaps 192(%r8), %xmm613255; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm10 = xmm6[0],xmm8[0],xmm6[1],xmm8[1]13256; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm10, %ymm0, %ymm1013257; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm7, %ymm0, %ymm713258; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm7 = ymm10[0,1,2,3,4,5],ymm7[6,7]13259; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm11 = ymm9[0,1,2,3],ymm7[4,5,6,7]13260; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm7 = xmm5[1,1,1,1]13261; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm7 = xmm7[0],xmm4[1],xmm7[2,3]13262; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm9 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]13263; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm7 = xmm7[0,1],xmm9[2,3]13264; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm9 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]13265; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm9, %ymm0, %ymm913266; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm15 = xmm6[1,1,1,1]13267; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm15 = xmm15[0],xmm8[1],xmm15[2,3]13268; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm15, %ymm0, %ymm1513269; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm9 = ymm15[0,1,2,3,4,5],ymm9[6,7]13270; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm10 = ymm7[0,1,2,3],ymm9[4,5,6,7]13271; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} xmm4 = xmm5[2],xmm4[2],xmm5[3],xmm4[3]13272; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} xmm5 = xmm6[2],xmm8[2],xmm6[3],xmm8[3]13273; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm6 = xmm2[2,2,2,2]13274; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm6 = xmm3[0,1,2],xmm6[3]13275; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm6, %ymm0, %ymm613276; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm5, %ymm0, %ymm713277; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3,4,5],ymm6[6,7]13278; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm7 = xmm0[2,2,2,2]13279; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm7 = xmm1[0,1,2],xmm7[3]13280; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm7 = xmm4[0,1],xmm7[2,3]13281; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm9 = ymm7[0,1,2,3],ymm6[4,5,6,7]13282; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm1[2],xmm0[2],xmm1[3],xmm0[3]13283; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm4[1],xmm0[1]13284; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} xmm1 = xmm3[2],xmm2[2],xmm3[3],xmm2[3]13285; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm113286; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm2 = xmm5[2,3,2,3]13287; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm213288; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]13289; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm8 = ymm0[0,1,2,3],ymm1[4,5,6,7]13290; AVX2-FCP-NEXT:    vmovaps 224(%rcx), %xmm613291; AVX2-FCP-NEXT:    vbroadcastss %xmm6, %xmm013292; AVX2-FCP-NEXT:    vmovaps 224(%rdx), %xmm513293; AVX2-FCP-NEXT:    vbroadcastss %xmm5, %xmm113294; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]13295; AVX2-FCP-NEXT:    vmovaps 224(%rsi), %xmm413296; AVX2-FCP-NEXT:    vmovaps 224(%rdi), %xmm213297; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm3 = xmm2[0],xmm4[0],xmm2[1],xmm4[1]13298; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm7 = xmm3[0,1],xmm1[2,3]13299; AVX2-FCP-NEXT:    vmovaps 224(%rax), %xmm313300; AVX2-FCP-NEXT:    vmovaps 224(%r10), %xmm113301; AVX2-FCP-NEXT:    vbroadcastss %xmm3, %xmm1513302; AVX2-FCP-NEXT:    vbroadcastss %xmm1, %xmm1413303; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm14 = xmm14[0],xmm15[0],xmm14[1],xmm15[1]13304; AVX2-FCP-NEXT:    vmovaps 224(%r9), %xmm1513305; AVX2-FCP-NEXT:    vmovaps 224(%r8), %xmm013306; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm13 = xmm0[0],xmm15[0],xmm0[1],xmm15[1]13307; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm13, %ymm0, %ymm1313308; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm14, %ymm0, %ymm1413309; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm13 = ymm13[0,1,2,3,4,5],ymm14[6,7]13310; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm14 = ymm7[0,1,2,3],ymm13[4,5,6,7]13311; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm7 = xmm2[1,1,1,1]13312; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm7 = xmm7[0],xmm4[1],xmm7[2,3]13313; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm13 = xmm5[0],xmm6[0],xmm5[1],xmm6[1]13314; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm7 = xmm7[0,1],xmm13[2,3]13315; AVX2-FCP-NEXT:    vunpcklps {{.*#+}} xmm13 = xmm1[0],xmm3[0],xmm1[1],xmm3[1]13316; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm13, %ymm0, %ymm1313317; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm12 = xmm0[1,1,1,1]13318; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm12 = xmm12[0],xmm15[1],xmm12[2,3]13319; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm12, %ymm0, %ymm1213320; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm12 = ymm12[0,1,2,3,4,5],ymm13[6,7]13321; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm7 = ymm7[0,1,2,3],ymm12[4,5,6,7]13322; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} xmm2 = xmm2[2],xmm4[2],xmm2[3],xmm4[3]13323; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm0[2],xmm15[2],xmm0[3],xmm15[3]13324; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm4 = xmm3[2,2,2,2]13325; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm4 = xmm1[0,1,2],xmm4[3]13326; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm4, %ymm0, %ymm413327; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm1213328; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm4 = ymm12[0,1,2,3,4,5],ymm4[6,7]13329; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm12 = xmm6[2,2,2,2]13330; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm12 = xmm5[0,1,2],xmm12[3]13331; AVX2-FCP-NEXT:    vblendps {{.*#+}} xmm12 = xmm2[0,1],xmm12[2,3]13332; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm4 = ymm12[0,1,2,3],ymm4[4,5,6,7]13333; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} xmm5 = xmm5[2],xmm6[2],xmm5[3],xmm6[3]13334; AVX2-FCP-NEXT:    vunpckhpd {{.*#+}} xmm2 = xmm2[1],xmm5[1]13335; AVX2-FCP-NEXT:    vunpckhps {{.*#+}} xmm1 = xmm1[2],xmm3[2],xmm1[3],xmm3[3]13336; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm113337; AVX2-FCP-NEXT:    vshufps {{.*#+}} xmm0 = xmm0[2,3,2,3]13338; AVX2-FCP-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm013339; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]13340; AVX2-FCP-NEXT:    vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]13341; AVX2-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax13342; AVX2-FCP-NEXT:    vmovaps %ymm0, 1888(%rax)13343; AVX2-FCP-NEXT:    vmovaps %ymm4, 1856(%rax)13344; AVX2-FCP-NEXT:    vmovaps %ymm7, 1824(%rax)13345; AVX2-FCP-NEXT:    vmovaps %ymm14, 1792(%rax)13346; AVX2-FCP-NEXT:    vmovaps %ymm8, 1632(%rax)13347; AVX2-FCP-NEXT:    vmovaps %ymm9, 1600(%rax)13348; AVX2-FCP-NEXT:    vmovaps %ymm10, 1568(%rax)13349; AVX2-FCP-NEXT:    vmovaps %ymm11, 1536(%rax)13350; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13351; AVX2-FCP-NEXT:    vmovaps %ymm0, 1376(%rax)13352; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13353; AVX2-FCP-NEXT:    vmovaps %ymm0, 1344(%rax)13354; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13355; AVX2-FCP-NEXT:    vmovaps %ymm0, 1312(%rax)13356; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13357; AVX2-FCP-NEXT:    vmovaps %ymm0, 1280(%rax)13358; AVX2-FCP-NEXT:    vmovups (%rsp), %ymm0 # 32-byte Reload13359; AVX2-FCP-NEXT:    vmovaps %ymm0, 1120(%rax)13360; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13361; AVX2-FCP-NEXT:    vmovaps %ymm0, 1088(%rax)13362; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13363; AVX2-FCP-NEXT:    vmovaps %ymm0, 1056(%rax)13364; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13365; AVX2-FCP-NEXT:    vmovaps %ymm0, 1024(%rax)13366; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13367; AVX2-FCP-NEXT:    vmovaps %ymm0, 864(%rax)13368; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13369; AVX2-FCP-NEXT:    vmovaps %ymm0, 832(%rax)13370; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13371; AVX2-FCP-NEXT:    vmovaps %ymm0, 800(%rax)13372; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13373; AVX2-FCP-NEXT:    vmovaps %ymm0, 768(%rax)13374; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13375; AVX2-FCP-NEXT:    vmovaps %ymm0, 608(%rax)13376; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13377; AVX2-FCP-NEXT:    vmovaps %ymm0, 576(%rax)13378; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13379; AVX2-FCP-NEXT:    vmovaps %ymm0, 544(%rax)13380; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13381; AVX2-FCP-NEXT:    vmovaps %ymm0, 512(%rax)13382; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13383; AVX2-FCP-NEXT:    vmovaps %ymm0, 352(%rax)13384; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13385; AVX2-FCP-NEXT:    vmovaps %ymm0, 320(%rax)13386; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13387; AVX2-FCP-NEXT:    vmovaps %ymm0, 288(%rax)13388; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13389; AVX2-FCP-NEXT:    vmovaps %ymm0, 256(%rax)13390; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13391; AVX2-FCP-NEXT:    vmovaps %ymm0, 96(%rax)13392; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13393; AVX2-FCP-NEXT:    vmovaps %ymm0, 64(%rax)13394; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13395; AVX2-FCP-NEXT:    vmovaps %ymm0, 32(%rax)13396; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13397; AVX2-FCP-NEXT:    vmovaps %ymm0, (%rax)13398; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13399; AVX2-FCP-NEXT:    vmovaps %ymm0, 2016(%rax)13400; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13401; AVX2-FCP-NEXT:    vmovaps %ymm0, 1984(%rax)13402; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13403; AVX2-FCP-NEXT:    vmovaps %ymm0, 1952(%rax)13404; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13405; AVX2-FCP-NEXT:    vmovaps %ymm0, 1920(%rax)13406; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13407; AVX2-FCP-NEXT:    vmovaps %ymm0, 1760(%rax)13408; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13409; AVX2-FCP-NEXT:    vmovaps %ymm0, 1728(%rax)13410; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13411; AVX2-FCP-NEXT:    vmovaps %ymm0, 1696(%rax)13412; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13413; AVX2-FCP-NEXT:    vmovaps %ymm0, 1664(%rax)13414; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13415; AVX2-FCP-NEXT:    vmovaps %ymm0, 1504(%rax)13416; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13417; AVX2-FCP-NEXT:    vmovaps %ymm0, 1472(%rax)13418; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13419; AVX2-FCP-NEXT:    vmovaps %ymm0, 1440(%rax)13420; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13421; AVX2-FCP-NEXT:    vmovaps %ymm0, 1408(%rax)13422; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13423; AVX2-FCP-NEXT:    vmovaps %ymm0, 1248(%rax)13424; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13425; AVX2-FCP-NEXT:    vmovaps %ymm0, 1216(%rax)13426; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13427; AVX2-FCP-NEXT:    vmovaps %ymm0, 1184(%rax)13428; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13429; AVX2-FCP-NEXT:    vmovaps %ymm0, 1152(%rax)13430; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13431; AVX2-FCP-NEXT:    vmovaps %ymm0, 992(%rax)13432; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13433; AVX2-FCP-NEXT:    vmovaps %ymm0, 960(%rax)13434; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13435; AVX2-FCP-NEXT:    vmovaps %ymm0, 928(%rax)13436; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13437; AVX2-FCP-NEXT:    vmovaps %ymm0, 896(%rax)13438; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13439; AVX2-FCP-NEXT:    vmovaps %ymm0, 736(%rax)13440; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13441; AVX2-FCP-NEXT:    vmovaps %ymm0, 704(%rax)13442; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13443; AVX2-FCP-NEXT:    vmovaps %ymm0, 672(%rax)13444; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13445; AVX2-FCP-NEXT:    vmovaps %ymm0, 640(%rax)13446; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13447; AVX2-FCP-NEXT:    vmovaps %ymm0, 480(%rax)13448; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13449; AVX2-FCP-NEXT:    vmovaps %ymm0, 448(%rax)13450; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13451; AVX2-FCP-NEXT:    vmovaps %ymm0, 416(%rax)13452; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13453; AVX2-FCP-NEXT:    vmovaps %ymm0, 384(%rax)13454; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13455; AVX2-FCP-NEXT:    vmovaps %ymm0, 224(%rax)13456; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13457; AVX2-FCP-NEXT:    vmovaps %ymm0, 192(%rax)13458; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13459; AVX2-FCP-NEXT:    vmovaps %ymm0, 160(%rax)13460; AVX2-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload13461; AVX2-FCP-NEXT:    vmovaps %ymm0, 128(%rax)13462; AVX2-FCP-NEXT:    addq $1672, %rsp # imm = 0x68813463; AVX2-FCP-NEXT:    vzeroupper13464; AVX2-FCP-NEXT:    retq13465;13466; AVX512-LABEL: store_i32_stride8_vf64:13467; AVX512:       # %bb.0:13468; AVX512-NEXT:    subq $6216, %rsp # imm = 0x184813469; AVX512-NEXT:    movq {{[0-9]+}}(%rsp), %rax13470; AVX512-NEXT:    movq {{[0-9]+}}(%rsp), %r1013471; AVX512-NEXT:    vmovdqa64 (%r10), %zmm513472; AVX512-NEXT:    vmovdqa64 64(%r10), %zmm413473; AVX512-NEXT:    vmovdqa64 128(%r10), %zmm213474; AVX512-NEXT:    vmovdqa64 (%rax), %zmm113475; AVX512-NEXT:    vmovdqa64 64(%rax), %zmm013476; AVX512-NEXT:    vmovdqa64 128(%rax), %zmm3013477; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm3 = [0,0,0,0,0,0,0,16,0,0,0,0,0,0,1,17]13478; AVX512-NEXT:    vmovdqa64 %zmm5, %zmm713479; AVX512-NEXT:    vmovdqa64 %zmm5, %zmm613480; AVX512-NEXT:    vpermt2d %zmm1, %zmm3, %zmm713481; AVX512-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13482; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm5 = [0,0,0,0,0,0,2,18,0,0,0,0,0,0,3,19]13483; AVX512-NEXT:    vmovdqa64 %zmm6, %zmm713484; AVX512-NEXT:    vpermt2d %zmm1, %zmm5, %zmm613485; AVX512-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13486; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm6 = [0,0,0,0,0,0,4,20,0,0,0,0,0,0,5,21]13487; AVX512-NEXT:    vmovdqa64 %zmm7, %zmm813488; AVX512-NEXT:    vpermt2d %zmm1, %zmm6, %zmm813489; AVX512-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13490; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm8 = [0,0,0,0,0,0,6,22,0,0,0,0,0,0,7,23]13491; AVX512-NEXT:    vmovdqa64 %zmm7, %zmm913492; AVX512-NEXT:    vpermt2d %zmm1, %zmm8, %zmm913493; AVX512-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13494; AVX512-NEXT:    vmovdqa64 %zmm8, %zmm913495; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm8 = [0,0,0,0,0,0,8,24,0,0,0,0,0,0,9,25]13496; AVX512-NEXT:    vmovdqa64 %zmm7, %zmm1013497; AVX512-NEXT:    vpermt2d %zmm1, %zmm8, %zmm1013498; AVX512-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13499; AVX512-NEXT:    vmovdqa64 %zmm8, %zmm1013500; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm8 = [0,0,0,0,0,0,10,26,0,0,0,0,0,0,11,27]13501; AVX512-NEXT:    vmovdqa64 %zmm7, %zmm1113502; AVX512-NEXT:    vpermt2d %zmm1, %zmm8, %zmm1113503; AVX512-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13504; AVX512-NEXT:    vmovdqa64 %zmm8, %zmm1113505; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm8 = [0,0,0,0,0,0,12,28,0,0,0,0,0,0,13,29]13506; AVX512-NEXT:    vmovdqa64 %zmm7, %zmm1313507; AVX512-NEXT:    vmovdqa64 %zmm7, %zmm1213508; AVX512-NEXT:    vpermt2d %zmm1, %zmm8, %zmm1313509; AVX512-NEXT:    vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13510; AVX512-NEXT:    vmovdqa64 %zmm8, %zmm713511; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm8 = [0,0,0,0,0,0,14,30,0,0,0,0,0,0,15,31]13512; AVX512-NEXT:    vpermt2d %zmm1, %zmm8, %zmm1213513; AVX512-NEXT:    vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13514; AVX512-NEXT:    vmovdqa64 %zmm4, %zmm113515; AVX512-NEXT:    vpermt2d %zmm0, %zmm3, %zmm113516; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13517; AVX512-NEXT:    vmovdqa64 %zmm4, %zmm113518; AVX512-NEXT:    vpermt2d %zmm0, %zmm5, %zmm113519; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13520; AVX512-NEXT:    vmovdqa64 %zmm4, %zmm113521; AVX512-NEXT:    vpermt2d %zmm0, %zmm6, %zmm113522; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13523; AVX512-NEXT:    vmovdqa64 %zmm4, %zmm113524; AVX512-NEXT:    vpermt2d %zmm0, %zmm9, %zmm113525; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13526; AVX512-NEXT:    vmovdqa64 %zmm4, %zmm113527; AVX512-NEXT:    vpermt2d %zmm0, %zmm10, %zmm113528; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13529; AVX512-NEXT:    vmovdqa64 %zmm4, %zmm113530; AVX512-NEXT:    vpermt2d %zmm0, %zmm11, %zmm113531; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13532; AVX512-NEXT:    vmovdqa64 %zmm4, %zmm113533; AVX512-NEXT:    vpermt2d %zmm0, %zmm7, %zmm113534; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13535; AVX512-NEXT:    vpermt2d %zmm0, %zmm8, %zmm413536; AVX512-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13537; AVX512-NEXT:    vmovdqa64 %zmm2, %zmm013538; AVX512-NEXT:    vpermt2d %zmm30, %zmm3, %zmm013539; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13540; AVX512-NEXT:    vmovdqa64 %zmm2, %zmm013541; AVX512-NEXT:    vpermt2d %zmm30, %zmm5, %zmm013542; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13543; AVX512-NEXT:    vmovdqa64 %zmm2, %zmm013544; AVX512-NEXT:    vpermt2d %zmm30, %zmm6, %zmm013545; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13546; AVX512-NEXT:    vmovdqa64 %zmm2, %zmm013547; AVX512-NEXT:    vpermt2d %zmm30, %zmm9, %zmm013548; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13549; AVX512-NEXT:    vmovdqa64 %zmm2, %zmm013550; AVX512-NEXT:    vpermt2d %zmm30, %zmm10, %zmm013551; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13552; AVX512-NEXT:    vmovdqa64 %zmm2, %zmm013553; AVX512-NEXT:    vpermt2d %zmm30, %zmm11, %zmm013554; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13555; AVX512-NEXT:    vmovdqa64 %zmm2, %zmm013556; AVX512-NEXT:    vpermt2d %zmm30, %zmm7, %zmm013557; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13558; AVX512-NEXT:    vpermt2d %zmm30, %zmm8, %zmm213559; AVX512-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13560; AVX512-NEXT:    vmovdqa64 192(%r10), %zmm113561; AVX512-NEXT:    vmovdqa64 192(%rax), %zmm013562; AVX512-NEXT:    vpermi2d %zmm0, %zmm1, %zmm313563; AVX512-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13564; AVX512-NEXT:    vpermi2d %zmm0, %zmm1, %zmm513565; AVX512-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13566; AVX512-NEXT:    vpermi2d %zmm0, %zmm1, %zmm613567; AVX512-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13568; AVX512-NEXT:    vpermi2d %zmm0, %zmm1, %zmm913569; AVX512-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13570; AVX512-NEXT:    vpermi2d %zmm0, %zmm1, %zmm1013571; AVX512-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13572; AVX512-NEXT:    vpermi2d %zmm0, %zmm1, %zmm1113573; AVX512-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13574; AVX512-NEXT:    vpermi2d %zmm0, %zmm1, %zmm713575; AVX512-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13576; AVX512-NEXT:    vpermt2d %zmm0, %zmm8, %zmm113577; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13578; AVX512-NEXT:    vmovdqa64 (%r8), %zmm2613579; AVX512-NEXT:    vmovdqa64 (%r9), %zmm013580; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,0,0,0,16,0,0,0,0,0,0,1,17,0,0]13581; AVX512-NEXT:    vmovdqa64 %zmm26, %zmm213582; AVX512-NEXT:    vpermt2d %zmm0, %zmm1, %zmm213583; AVX512-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13584; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm213585; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,0,0,2,18,0,0,0,0,0,0,3,19,0,0]13586; AVX512-NEXT:    vmovdqa64 %zmm26, %zmm313587; AVX512-NEXT:    vpermt2d %zmm0, %zmm1, %zmm313588; AVX512-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13589; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm313590; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,0,0,4,20,0,0,0,0,0,0,5,21,0,0]13591; AVX512-NEXT:    vmovdqa64 %zmm26, %zmm413592; AVX512-NEXT:    vpermt2d %zmm0, %zmm1, %zmm413593; AVX512-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13594; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm413595; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,0,0,6,22,0,0,0,0,0,0,7,23,0,0]13596; AVX512-NEXT:    vmovdqa64 %zmm26, %zmm513597; AVX512-NEXT:    vpermt2d %zmm0, %zmm1, %zmm513598; AVX512-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13599; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm513600; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,0,0,8,24,0,0,0,0,0,0,9,25,0,0]13601; AVX512-NEXT:    vmovdqa64 %zmm26, %zmm613602; AVX512-NEXT:    vpermt2d %zmm0, %zmm1, %zmm613603; AVX512-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13604; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm613605; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,0,0,10,26,0,0,0,0,0,0,11,27,0,0]13606; AVX512-NEXT:    vmovdqa64 %zmm26, %zmm713607; AVX512-NEXT:    vpermt2d %zmm0, %zmm1, %zmm713608; AVX512-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13609; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm713610; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,0,0,12,28,0,0,0,0,0,0,13,29,0,0]13611; AVX512-NEXT:    vmovdqa64 %zmm26, %zmm813612; AVX512-NEXT:    vpermt2d %zmm0, %zmm1, %zmm813613; AVX512-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13614; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm813615; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,0,0,14,30,0,0,0,0,0,0,15,31,0,0]13616; AVX512-NEXT:    vpermt2d %zmm0, %zmm1, %zmm2613617; AVX512-NEXT:    vmovdqa64 64(%r8), %zmm2513618; AVX512-NEXT:    vmovdqa64 64(%r9), %zmm013619; AVX512-NEXT:    vmovdqa64 %zmm25, %zmm913620; AVX512-NEXT:    vpermt2d %zmm0, %zmm2, %zmm913621; AVX512-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13622; AVX512-NEXT:    vmovdqa64 %zmm25, %zmm913623; AVX512-NEXT:    vpermt2d %zmm0, %zmm3, %zmm913624; AVX512-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13625; AVX512-NEXT:    vmovdqa64 %zmm25, %zmm913626; AVX512-NEXT:    vpermt2d %zmm0, %zmm4, %zmm913627; AVX512-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13628; AVX512-NEXT:    vmovdqa64 %zmm25, %zmm913629; AVX512-NEXT:    vpermt2d %zmm0, %zmm5, %zmm913630; AVX512-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13631; AVX512-NEXT:    vmovdqa64 %zmm25, %zmm913632; AVX512-NEXT:    vpermt2d %zmm0, %zmm6, %zmm913633; AVX512-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13634; AVX512-NEXT:    vmovdqa64 %zmm25, %zmm913635; AVX512-NEXT:    vpermt2d %zmm0, %zmm7, %zmm913636; AVX512-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13637; AVX512-NEXT:    vmovdqa64 %zmm25, %zmm913638; AVX512-NEXT:    vpermt2d %zmm0, %zmm8, %zmm913639; AVX512-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13640; AVX512-NEXT:    vpermt2d %zmm0, %zmm1, %zmm2513641; AVX512-NEXT:    vmovdqa64 128(%r8), %zmm2213642; AVX512-NEXT:    vmovdqa64 128(%r9), %zmm013643; AVX512-NEXT:    vmovdqa64 %zmm22, %zmm913644; AVX512-NEXT:    vpermt2d %zmm0, %zmm2, %zmm913645; AVX512-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13646; AVX512-NEXT:    vmovdqa64 %zmm22, %zmm913647; AVX512-NEXT:    vpermt2d %zmm0, %zmm3, %zmm913648; AVX512-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13649; AVX512-NEXT:    vmovdqa64 %zmm22, %zmm913650; AVX512-NEXT:    vpermt2d %zmm0, %zmm4, %zmm913651; AVX512-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13652; AVX512-NEXT:    vmovdqa64 %zmm22, %zmm913653; AVX512-NEXT:    vpermt2d %zmm0, %zmm5, %zmm913654; AVX512-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13655; AVX512-NEXT:    vmovdqa64 %zmm22, %zmm913656; AVX512-NEXT:    vpermt2d %zmm0, %zmm6, %zmm913657; AVX512-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13658; AVX512-NEXT:    vmovdqa64 %zmm22, %zmm913659; AVX512-NEXT:    vpermt2d %zmm0, %zmm7, %zmm913660; AVX512-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13661; AVX512-NEXT:    vmovdqa64 %zmm22, %zmm913662; AVX512-NEXT:    vpermt2d %zmm0, %zmm8, %zmm913663; AVX512-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13664; AVX512-NEXT:    vpermt2d %zmm0, %zmm1, %zmm2213665; AVX512-NEXT:    vmovdqa64 192(%r8), %zmm2713666; AVX512-NEXT:    vmovdqa64 192(%r9), %zmm013667; AVX512-NEXT:    vpermi2d %zmm0, %zmm27, %zmm213668; AVX512-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13669; AVX512-NEXT:    vpermi2d %zmm0, %zmm27, %zmm313670; AVX512-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13671; AVX512-NEXT:    vpermi2d %zmm0, %zmm27, %zmm413672; AVX512-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13673; AVX512-NEXT:    vpermi2d %zmm0, %zmm27, %zmm513674; AVX512-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13675; AVX512-NEXT:    vpermi2d %zmm0, %zmm27, %zmm613676; AVX512-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13677; AVX512-NEXT:    vpermi2d %zmm0, %zmm27, %zmm713678; AVX512-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13679; AVX512-NEXT:    vpermi2d %zmm0, %zmm27, %zmm813680; AVX512-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13681; AVX512-NEXT:    vpermt2d %zmm0, %zmm1, %zmm2713682; AVX512-NEXT:    vmovdqa64 (%rdx), %zmm813683; AVX512-NEXT:    vmovdqa64 (%rcx), %zmm013684; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,0,16,0,0,0,0,0,0,1,17,0,0,0,0]13685; AVX512-NEXT:    vmovdqa64 %zmm8, %zmm213686; AVX512-NEXT:    vpermt2d %zmm0, %zmm1, %zmm213687; AVX512-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13688; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm213689; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,2,18,0,0,0,0,0,0,3,19,0,0,0,0]13690; AVX512-NEXT:    vmovdqa64 %zmm8, %zmm313691; AVX512-NEXT:    vpermt2d %zmm0, %zmm1, %zmm313692; AVX512-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13693; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm313694; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,4,20,0,0,0,0,0,0,5,21,0,0,0,0]13695; AVX512-NEXT:    vmovdqa64 %zmm8, %zmm413696; AVX512-NEXT:    vpermt2d %zmm0, %zmm1, %zmm413697; AVX512-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13698; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm413699; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,6,22,0,0,0,0,0,0,7,23,0,0,0,0]13700; AVX512-NEXT:    vmovdqa64 %zmm8, %zmm513701; AVX512-NEXT:    vpermt2d %zmm0, %zmm1, %zmm513702; AVX512-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13703; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm513704; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm7 = [0,0,8,24,0,0,0,0,0,0,9,25,0,0,0,0]13705; AVX512-NEXT:    vmovdqa64 %zmm8, %zmm613706; AVX512-NEXT:    vpermt2d %zmm0, %zmm7, %zmm613707; AVX512-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13708; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm9 = [0,0,10,26,0,0,0,0,0,0,11,27,0,0,0,0]13709; AVX512-NEXT:    vmovdqa64 %zmm8, %zmm613710; AVX512-NEXT:    vpermt2d %zmm0, %zmm9, %zmm613711; AVX512-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13712; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm10 = [0,0,12,28,0,0,0,0,0,0,13,29,0,0,0,0]13713; AVX512-NEXT:    vmovdqa64 %zmm8, %zmm613714; AVX512-NEXT:    vpermt2d %zmm0, %zmm10, %zmm613715; AVX512-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13716; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,14,30,0,0,0,0,0,0,15,31,0,0,0,0]13717; AVX512-NEXT:    vpermt2d %zmm0, %zmm1, %zmm813718; AVX512-NEXT:    vmovdqa64 64(%rdx), %zmm613719; AVX512-NEXT:    vmovdqa64 64(%rcx), %zmm013720; AVX512-NEXT:    vmovdqa64 %zmm6, %zmm1113721; AVX512-NEXT:    vpermt2d %zmm0, %zmm2, %zmm1113722; AVX512-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13723; AVX512-NEXT:    vmovdqa64 %zmm6, %zmm1113724; AVX512-NEXT:    vpermt2d %zmm0, %zmm3, %zmm1113725; AVX512-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13726; AVX512-NEXT:    vmovdqa64 %zmm6, %zmm1113727; AVX512-NEXT:    vpermt2d %zmm0, %zmm4, %zmm1113728; AVX512-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13729; AVX512-NEXT:    vmovdqa64 %zmm6, %zmm1113730; AVX512-NEXT:    vpermt2d %zmm0, %zmm5, %zmm1113731; AVX512-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13732; AVX512-NEXT:    vmovdqa64 %zmm6, %zmm1113733; AVX512-NEXT:    vpermt2d %zmm0, %zmm7, %zmm1113734; AVX512-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13735; AVX512-NEXT:    vmovdqa64 %zmm6, %zmm1113736; AVX512-NEXT:    vpermt2d %zmm0, %zmm9, %zmm1113737; AVX512-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13738; AVX512-NEXT:    vmovdqa64 %zmm6, %zmm1113739; AVX512-NEXT:    vpermt2d %zmm0, %zmm10, %zmm1113740; AVX512-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13741; AVX512-NEXT:    vpermt2d %zmm0, %zmm1, %zmm613742; AVX512-NEXT:    vmovdqa64 128(%rdx), %zmm1313743; AVX512-NEXT:    vmovdqa64 128(%rcx), %zmm013744; AVX512-NEXT:    vmovdqa64 %zmm13, %zmm1113745; AVX512-NEXT:    vpermt2d %zmm0, %zmm2, %zmm1113746; AVX512-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13747; AVX512-NEXT:    vmovdqa64 %zmm13, %zmm1113748; AVX512-NEXT:    vpermt2d %zmm0, %zmm3, %zmm1113749; AVX512-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13750; AVX512-NEXT:    vmovdqa64 %zmm13, %zmm1113751; AVX512-NEXT:    vpermt2d %zmm0, %zmm4, %zmm1113752; AVX512-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13753; AVX512-NEXT:    vmovdqa64 %zmm13, %zmm1113754; AVX512-NEXT:    vpermt2d %zmm0, %zmm5, %zmm1113755; AVX512-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13756; AVX512-NEXT:    vmovdqa64 %zmm13, %zmm1113757; AVX512-NEXT:    vpermt2d %zmm0, %zmm7, %zmm1113758; AVX512-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13759; AVX512-NEXT:    vmovdqa64 %zmm13, %zmm1113760; AVX512-NEXT:    vpermt2d %zmm0, %zmm9, %zmm1113761; AVX512-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13762; AVX512-NEXT:    vmovdqa64 %zmm13, %zmm1113763; AVX512-NEXT:    vpermt2d %zmm0, %zmm10, %zmm1113764; AVX512-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13765; AVX512-NEXT:    vpermt2d %zmm0, %zmm1, %zmm1313766; AVX512-NEXT:    vmovdqa64 192(%rdx), %zmm1713767; AVX512-NEXT:    vmovdqa64 192(%rcx), %zmm013768; AVX512-NEXT:    vpermi2d %zmm0, %zmm17, %zmm213769; AVX512-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13770; AVX512-NEXT:    vpermi2d %zmm0, %zmm17, %zmm313771; AVX512-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13772; AVX512-NEXT:    vpermi2d %zmm0, %zmm17, %zmm413773; AVX512-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13774; AVX512-NEXT:    vpermi2d %zmm0, %zmm17, %zmm513775; AVX512-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13776; AVX512-NEXT:    vpermi2d %zmm0, %zmm17, %zmm713777; AVX512-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13778; AVX512-NEXT:    vpermi2d %zmm0, %zmm17, %zmm913779; AVX512-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13780; AVX512-NEXT:    vpermi2d %zmm0, %zmm17, %zmm1013781; AVX512-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13782; AVX512-NEXT:    vpermt2d %zmm0, %zmm1, %zmm1713783; AVX512-NEXT:    vmovdqa64 (%rdi), %zmm513784; AVX512-NEXT:    vmovdqa64 (%rsi), %zmm013785; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm14 = [0,16,0,0,0,0,0,0,1,17,0,0,0,0,0,0]13786; AVX512-NEXT:    vmovdqa64 %zmm5, %zmm113787; AVX512-NEXT:    vpermt2d %zmm0, %zmm14, %zmm113788; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13789; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm11 = [2,18,0,0,0,0,0,0,3,19,0,0,0,0,0,0]13790; AVX512-NEXT:    vmovdqa64 %zmm5, %zmm113791; AVX512-NEXT:    vpermt2d %zmm0, %zmm11, %zmm113792; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13793; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm10 = [4,20,0,0,0,0,0,0,5,21,0,0,0,0,0,0]13794; AVX512-NEXT:    vmovdqa64 %zmm5, %zmm113795; AVX512-NEXT:    vpermt2d %zmm0, %zmm10, %zmm113796; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13797; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm9 = [6,22,0,0,0,0,0,0,7,23,0,0,0,0,0,0]13798; AVX512-NEXT:    vmovdqa64 %zmm5, %zmm113799; AVX512-NEXT:    vpermt2d %zmm0, %zmm9, %zmm113800; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13801; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm7 = [8,24,0,0,0,0,0,0,9,25,0,0,0,0,0,0]13802; AVX512-NEXT:    vmovdqa64 %zmm5, %zmm113803; AVX512-NEXT:    vpermt2d %zmm0, %zmm7, %zmm113804; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13805; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [10,26,0,0,0,0,0,0,11,27,0,0,0,0,0,0]13806; AVX512-NEXT:    vmovdqa64 %zmm5, %zmm113807; AVX512-NEXT:    vpermt2d %zmm0, %zmm4, %zmm113808; AVX512-NEXT:    vmovdqu64 %zmm1, (%rsp) # 64-byte Spill13809; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm3 = [12,28,0,0,0,0,0,0,13,29,0,0,0,0,0,0]13810; AVX512-NEXT:    vmovdqa64 %zmm5, %zmm113811; AVX512-NEXT:    vpermt2d %zmm0, %zmm3, %zmm113812; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13813; AVX512-NEXT:    vpmovsxbd {{.*#+}} zmm31 = [14,30,0,0,0,0,0,0,15,31,0,0,0,0,0,0]13814; AVX512-NEXT:    vpermt2d %zmm0, %zmm31, %zmm513815; AVX512-NEXT:    vmovdqa64 64(%rdi), %zmm1213816; AVX512-NEXT:    vmovdqa64 64(%rsi), %zmm013817; AVX512-NEXT:    vmovdqa64 %zmm12, %zmm113818; AVX512-NEXT:    vpermt2d %zmm0, %zmm14, %zmm113819; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13820; AVX512-NEXT:    vmovdqa64 %zmm12, %zmm113821; AVX512-NEXT:    vpermt2d %zmm0, %zmm11, %zmm113822; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13823; AVX512-NEXT:    vmovdqa64 %zmm12, %zmm113824; AVX512-NEXT:    vpermt2d %zmm0, %zmm10, %zmm113825; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13826; AVX512-NEXT:    vmovdqa64 %zmm12, %zmm113827; AVX512-NEXT:    vpermt2d %zmm0, %zmm9, %zmm113828; AVX512-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13829; AVX512-NEXT:    vmovdqa64 %zmm12, %zmm3013830; AVX512-NEXT:    vpermt2d %zmm0, %zmm7, %zmm3013831; AVX512-NEXT:    vmovdqa64 %zmm12, %zmm2913832; AVX512-NEXT:    vpermt2d %zmm0, %zmm4, %zmm2913833; AVX512-NEXT:    vmovdqa64 %zmm12, %zmm2813834; AVX512-NEXT:    vpermt2d %zmm0, %zmm3, %zmm2813835; AVX512-NEXT:    vpermt2d %zmm0, %zmm31, %zmm1213836; AVX512-NEXT:    vmovdqa64 128(%rdi), %zmm1513837; AVX512-NEXT:    vmovdqa64 128(%rsi), %zmm013838; AVX512-NEXT:    vmovdqa64 %zmm15, %zmm2413839; AVX512-NEXT:    vpermt2d %zmm0, %zmm14, %zmm2413840; AVX512-NEXT:    vmovdqa64 %zmm15, %zmm2313841; AVX512-NEXT:    vpermt2d %zmm0, %zmm11, %zmm2313842; AVX512-NEXT:    vmovdqa64 %zmm15, %zmm2113843; AVX512-NEXT:    vpermt2d %zmm0, %zmm10, %zmm2113844; AVX512-NEXT:    vmovdqa64 %zmm15, %zmm2013845; AVX512-NEXT:    vpermt2d %zmm0, %zmm9, %zmm2013846; AVX512-NEXT:    vmovdqa64 %zmm15, %zmm1913847; AVX512-NEXT:    vpermt2d %zmm0, %zmm7, %zmm1913848; AVX512-NEXT:    vmovdqa64 %zmm15, %zmm1813849; AVX512-NEXT:    vpermt2d %zmm0, %zmm4, %zmm1813850; AVX512-NEXT:    vmovdqa64 %zmm15, %zmm1613851; AVX512-NEXT:    vpermt2d %zmm0, %zmm3, %zmm1613852; AVX512-NEXT:    vpermt2d %zmm0, %zmm31, %zmm1513853; AVX512-NEXT:    vmovdqa64 192(%rdi), %zmm213854; AVX512-NEXT:    vmovdqa64 192(%rsi), %zmm013855; AVX512-NEXT:    vpermi2d %zmm0, %zmm2, %zmm1413856; AVX512-NEXT:    vpermi2d %zmm0, %zmm2, %zmm1113857; AVX512-NEXT:    vpermi2d %zmm0, %zmm2, %zmm1013858; AVX512-NEXT:    vpermi2d %zmm0, %zmm2, %zmm913859; AVX512-NEXT:    vpermi2d %zmm0, %zmm2, %zmm713860; AVX512-NEXT:    vpermi2d %zmm0, %zmm2, %zmm413861; AVX512-NEXT:    vpermi2d %zmm0, %zmm2, %zmm313862; AVX512-NEXT:    vpermt2d %zmm0, %zmm31, %zmm213863; AVX512-NEXT:    movb $-120, %al13864; AVX512-NEXT:    kmovw %eax, %k113865; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13866; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload13867; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}13868; AVX512-NEXT:    movb $34, %al13869; AVX512-NEXT:    kmovw %eax, %k213870; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13871; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload13872; AVX512-NEXT:    vmovdqa64 %zmm31, %zmm0 {%k2}13873; AVX512-NEXT:    movb $-52, %al13874; AVX512-NEXT:    kmovw %eax, %k313875; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm0 {%k3}13876; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13877; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13878; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload13879; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}13880; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13881; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload13882; AVX512-NEXT:    vmovdqa64 %zmm31, %zmm0 {%k2}13883; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm0 {%k3}13884; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13885; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13886; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload13887; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}13888; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13889; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload13890; AVX512-NEXT:    vmovdqa64 %zmm31, %zmm0 {%k2}13891; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm0 {%k3}13892; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13893; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13894; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload13895; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}13896; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13897; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload13898; AVX512-NEXT:    vmovdqa64 %zmm31, %zmm0 {%k2}13899; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm0 {%k3}13900; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13901; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13902; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload13903; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}13904; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13905; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload13906; AVX512-NEXT:    vmovdqa64 %zmm31, %zmm0 {%k2}13907; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm0 {%k3}13908; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13909; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13910; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload13911; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}13912; AVX512-NEXT:    vmovdqu64 (%rsp), %zmm0 # 64-byte Reload13913; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload13914; AVX512-NEXT:    vmovdqa64 %zmm31, %zmm0 {%k2}13915; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm0 {%k3}13916; AVX512-NEXT:    vmovdqu64 %zmm0, (%rsp) # 64-byte Spill13917; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13918; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload13919; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}13920; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13921; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload13922; AVX512-NEXT:    vmovdqa64 %zmm31, %zmm0 {%k2}13923; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm0 {%k3}13924; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13925; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13926; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm26 {%k1}13927; AVX512-NEXT:    vmovdqa64 %zmm8, %zmm5 {%k2}13928; AVX512-NEXT:    vmovdqa64 %zmm26, %zmm5 {%k3}13929; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13930; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload13931; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}13932; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13933; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload13934; AVX512-NEXT:    vmovdqa64 %zmm8, %zmm0 {%k2}13935; AVX512-NEXT:    vmovdqa64 %zmm1, %zmm0 {%k3}13936; AVX512-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13937; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13938; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload13939; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm8 {%k1}13940; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload13941; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13942; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm26 {%k2}13943; AVX512-NEXT:    vmovdqa64 %zmm8, %zmm26 {%k3}13944; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13945; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload13946; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm8 {%k1}13947; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload13948; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13949; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k2}13950; AVX512-NEXT:    vmovdqa64 %zmm8, %zmm1 {%k3}13951; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13952; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload13953; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm8 {%k1}13954; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13955; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload13956; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm31 {%k2}13957; AVX512-NEXT:    vmovdqa64 %zmm8, %zmm31 {%k3}13958; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13959; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload13960; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm8 {%k1}13961; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13962; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm30 {%k2}13963; AVX512-NEXT:    vmovdqa64 %zmm8, %zmm30 {%k3}13964; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13965; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload13966; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm8 {%k1}13967; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13968; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm29 {%k2}13969; AVX512-NEXT:    vmovdqa64 %zmm8, %zmm29 {%k3}13970; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13971; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload13972; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm8 {%k1}13973; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13974; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm28 {%k2}13975; AVX512-NEXT:    vmovdqa64 %zmm8, %zmm28 {%k3}13976; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13977; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm25 {%k1}13978; AVX512-NEXT:    vmovdqa64 %zmm6, %zmm12 {%k2}13979; AVX512-NEXT:    vmovdqa64 %zmm25, %zmm12 {%k3}13980; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13981; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload13982; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm6 {%k1}13983; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13984; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm24 {%k2}13985; AVX512-NEXT:    vmovdqa64 %zmm6, %zmm24 {%k3}13986; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13987; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload13988; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm6 {%k1}13989; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13990; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm23 {%k2}13991; AVX512-NEXT:    vmovdqa64 %zmm6, %zmm23 {%k3}13992; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13993; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload13994; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm6 {%k1}13995; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13996; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm21 {%k2}13997; AVX512-NEXT:    vmovdqa64 %zmm6, %zmm21 {%k3}13998; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13999; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload14000; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm6 {%k1}14001; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14002; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm20 {%k2}14003; AVX512-NEXT:    vmovdqa64 %zmm6, %zmm20 {%k3}14004; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14005; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload14006; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm6 {%k1}14007; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14008; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm19 {%k2}14009; AVX512-NEXT:    vmovdqa64 %zmm6, %zmm19 {%k3}14010; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14011; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload14012; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm6 {%k1}14013; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14014; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm18 {%k2}14015; AVX512-NEXT:    vmovdqa64 %zmm6, %zmm18 {%k3}14016; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14017; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload14018; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm6 {%k1}14019; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14020; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm16 {%k2}14021; AVX512-NEXT:    vmovdqa64 %zmm6, %zmm16 {%k3}14022; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14023; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm22 {%k1}14024; AVX512-NEXT:    vmovdqa64 %zmm13, %zmm15 {%k2}14025; AVX512-NEXT:    vmovdqa64 %zmm22, %zmm15 {%k3}14026; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14027; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload14028; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm6 {%k1}14029; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14030; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm14 {%k2}14031; AVX512-NEXT:    vmovdqa64 %zmm6, %zmm14 {%k3}14032; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14033; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload14034; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm6 {%k1}14035; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14036; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm11 {%k2}14037; AVX512-NEXT:    vmovdqa64 %zmm6, %zmm11 {%k3}14038; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14039; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload14040; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm6 {%k1}14041; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14042; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm10 {%k2}14043; AVX512-NEXT:    vmovdqa64 %zmm6, %zmm10 {%k3}14044; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14045; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload14046; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm6 {%k1}14047; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14048; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm9 {%k2}14049; AVX512-NEXT:    vmovdqa64 %zmm6, %zmm9 {%k3}14050; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14051; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload14052; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm6 {%k1}14053; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14054; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm7 {%k2}14055; AVX512-NEXT:    vmovdqa64 %zmm6, %zmm7 {%k3}14056; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14057; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload14058; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm6 {%k1}14059; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14060; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm4 {%k2}14061; AVX512-NEXT:    vmovdqa64 %zmm6, %zmm4 {%k3}14062; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14063; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload14064; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm6 {%k1}14065; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14066; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm3 {%k2}14067; AVX512-NEXT:    vmovdqa64 %zmm6, %zmm3 {%k3}14068; AVX512-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14069; AVX512-NEXT:    vmovdqa64 %zmm0, %zmm27 {%k1}14070; AVX512-NEXT:    vmovdqa64 %zmm17, %zmm2 {%k2}14071; AVX512-NEXT:    vmovdqa64 %zmm27, %zmm2 {%k3}14072; AVX512-NEXT:    movq {{[0-9]+}}(%rsp), %rax14073; AVX512-NEXT:    vmovdqa64 %zmm2, 1984(%rax)14074; AVX512-NEXT:    vmovdqa64 %zmm3, 1920(%rax)14075; AVX512-NEXT:    vmovdqa64 %zmm4, 1856(%rax)14076; AVX512-NEXT:    vmovdqa64 %zmm7, 1792(%rax)14077; AVX512-NEXT:    vmovdqa64 %zmm9, 1728(%rax)14078; AVX512-NEXT:    vmovdqa64 %zmm10, 1664(%rax)14079; AVX512-NEXT:    vmovdqa64 %zmm11, 1600(%rax)14080; AVX512-NEXT:    vmovdqa64 %zmm14, 1536(%rax)14081; AVX512-NEXT:    vmovdqa64 %zmm15, 1472(%rax)14082; AVX512-NEXT:    vmovdqa64 %zmm16, 1408(%rax)14083; AVX512-NEXT:    vmovdqa64 %zmm18, 1344(%rax)14084; AVX512-NEXT:    vmovdqa64 %zmm19, 1280(%rax)14085; AVX512-NEXT:    vmovdqa64 %zmm20, 1216(%rax)14086; AVX512-NEXT:    vmovdqa64 %zmm21, 1152(%rax)14087; AVX512-NEXT:    vmovdqa64 %zmm23, 1088(%rax)14088; AVX512-NEXT:    vmovdqa64 %zmm24, 1024(%rax)14089; AVX512-NEXT:    vmovdqa64 %zmm12, 960(%rax)14090; AVX512-NEXT:    vmovdqa64 %zmm28, 896(%rax)14091; AVX512-NEXT:    vmovdqa64 %zmm29, 832(%rax)14092; AVX512-NEXT:    vmovdqa64 %zmm30, 768(%rax)14093; AVX512-NEXT:    vmovdqa64 %zmm31, 704(%rax)14094; AVX512-NEXT:    vmovdqa64 %zmm1, 640(%rax)14095; AVX512-NEXT:    vmovdqa64 %zmm26, 576(%rax)14096; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14097; AVX512-NEXT:    vmovaps %zmm0, 512(%rax)14098; AVX512-NEXT:    vmovdqa64 %zmm5, 448(%rax)14099; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14100; AVX512-NEXT:    vmovaps %zmm0, 384(%rax)14101; AVX512-NEXT:    vmovups (%rsp), %zmm0 # 64-byte Reload14102; AVX512-NEXT:    vmovaps %zmm0, 320(%rax)14103; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14104; AVX512-NEXT:    vmovaps %zmm0, 256(%rax)14105; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14106; AVX512-NEXT:    vmovaps %zmm0, 192(%rax)14107; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14108; AVX512-NEXT:    vmovaps %zmm0, 128(%rax)14109; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14110; AVX512-NEXT:    vmovaps %zmm0, 64(%rax)14111; AVX512-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14112; AVX512-NEXT:    vmovaps %zmm0, (%rax)14113; AVX512-NEXT:    addq $6216, %rsp # imm = 0x184814114; AVX512-NEXT:    vzeroupper14115; AVX512-NEXT:    retq14116;14117; AVX512-FCP-LABEL: store_i32_stride8_vf64:14118; AVX512-FCP:       # %bb.0:14119; AVX512-FCP-NEXT:    subq $6216, %rsp # imm = 0x184814120; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax14121; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r1014122; AVX512-FCP-NEXT:    vmovdqa64 (%r10), %zmm514123; AVX512-FCP-NEXT:    vmovdqa64 64(%r10), %zmm414124; AVX512-FCP-NEXT:    vmovdqa64 128(%r10), %zmm214125; AVX512-FCP-NEXT:    vmovdqa64 (%rax), %zmm114126; AVX512-FCP-NEXT:    vmovdqa64 64(%rax), %zmm014127; AVX512-FCP-NEXT:    vmovdqa64 128(%rax), %zmm3014128; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm3 = [0,0,0,0,0,0,0,16,0,0,0,0,0,0,1,17]14129; AVX512-FCP-NEXT:    vmovdqa64 %zmm5, %zmm714130; AVX512-FCP-NEXT:    vmovdqa64 %zmm5, %zmm614131; AVX512-FCP-NEXT:    vpermt2d %zmm1, %zmm3, %zmm714132; AVX512-FCP-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14133; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm5 = [0,0,0,0,0,0,2,18,0,0,0,0,0,0,3,19]14134; AVX512-FCP-NEXT:    vmovdqa64 %zmm6, %zmm714135; AVX512-FCP-NEXT:    vpermt2d %zmm1, %zmm5, %zmm614136; AVX512-FCP-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14137; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm6 = [0,0,0,0,0,0,4,20,0,0,0,0,0,0,5,21]14138; AVX512-FCP-NEXT:    vmovdqa64 %zmm7, %zmm814139; AVX512-FCP-NEXT:    vpermt2d %zmm1, %zmm6, %zmm814140; AVX512-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14141; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm8 = [0,0,0,0,0,0,6,22,0,0,0,0,0,0,7,23]14142; AVX512-FCP-NEXT:    vmovdqa64 %zmm7, %zmm914143; AVX512-FCP-NEXT:    vpermt2d %zmm1, %zmm8, %zmm914144; AVX512-FCP-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14145; AVX512-FCP-NEXT:    vmovdqa64 %zmm8, %zmm914146; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm8 = [0,0,0,0,0,0,8,24,0,0,0,0,0,0,9,25]14147; AVX512-FCP-NEXT:    vmovdqa64 %zmm7, %zmm1014148; AVX512-FCP-NEXT:    vpermt2d %zmm1, %zmm8, %zmm1014149; AVX512-FCP-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14150; AVX512-FCP-NEXT:    vmovdqa64 %zmm8, %zmm1014151; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm8 = [0,0,0,0,0,0,10,26,0,0,0,0,0,0,11,27]14152; AVX512-FCP-NEXT:    vmovdqa64 %zmm7, %zmm1114153; AVX512-FCP-NEXT:    vpermt2d %zmm1, %zmm8, %zmm1114154; AVX512-FCP-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14155; AVX512-FCP-NEXT:    vmovdqa64 %zmm8, %zmm1114156; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm8 = [0,0,0,0,0,0,12,28,0,0,0,0,0,0,13,29]14157; AVX512-FCP-NEXT:    vmovdqa64 %zmm7, %zmm1314158; AVX512-FCP-NEXT:    vmovdqa64 %zmm7, %zmm1214159; AVX512-FCP-NEXT:    vpermt2d %zmm1, %zmm8, %zmm1314160; AVX512-FCP-NEXT:    vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14161; AVX512-FCP-NEXT:    vmovdqa64 %zmm8, %zmm714162; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm8 = [0,0,0,0,0,0,14,30,0,0,0,0,0,0,15,31]14163; AVX512-FCP-NEXT:    vpermt2d %zmm1, %zmm8, %zmm1214164; AVX512-FCP-NEXT:    vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14165; AVX512-FCP-NEXT:    vmovdqa64 %zmm4, %zmm114166; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm3, %zmm114167; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14168; AVX512-FCP-NEXT:    vmovdqa64 %zmm4, %zmm114169; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm5, %zmm114170; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14171; AVX512-FCP-NEXT:    vmovdqa64 %zmm4, %zmm114172; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm6, %zmm114173; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14174; AVX512-FCP-NEXT:    vmovdqa64 %zmm4, %zmm114175; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm9, %zmm114176; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14177; AVX512-FCP-NEXT:    vmovdqa64 %zmm4, %zmm114178; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm10, %zmm114179; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14180; AVX512-FCP-NEXT:    vmovdqa64 %zmm4, %zmm114181; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm11, %zmm114182; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14183; AVX512-FCP-NEXT:    vmovdqa64 %zmm4, %zmm114184; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm7, %zmm114185; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14186; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm8, %zmm414187; AVX512-FCP-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14188; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, %zmm014189; AVX512-FCP-NEXT:    vpermt2d %zmm30, %zmm3, %zmm014190; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14191; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, %zmm014192; AVX512-FCP-NEXT:    vpermt2d %zmm30, %zmm5, %zmm014193; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14194; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, %zmm014195; AVX512-FCP-NEXT:    vpermt2d %zmm30, %zmm6, %zmm014196; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14197; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, %zmm014198; AVX512-FCP-NEXT:    vpermt2d %zmm30, %zmm9, %zmm014199; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14200; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, %zmm014201; AVX512-FCP-NEXT:    vpermt2d %zmm30, %zmm10, %zmm014202; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14203; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, %zmm014204; AVX512-FCP-NEXT:    vpermt2d %zmm30, %zmm11, %zmm014205; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14206; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, %zmm014207; AVX512-FCP-NEXT:    vpermt2d %zmm30, %zmm7, %zmm014208; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14209; AVX512-FCP-NEXT:    vpermt2d %zmm30, %zmm8, %zmm214210; AVX512-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14211; AVX512-FCP-NEXT:    vmovdqa64 192(%r10), %zmm114212; AVX512-FCP-NEXT:    vmovdqa64 192(%rax), %zmm014213; AVX512-FCP-NEXT:    vpermi2d %zmm0, %zmm1, %zmm314214; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14215; AVX512-FCP-NEXT:    vpermi2d %zmm0, %zmm1, %zmm514216; AVX512-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14217; AVX512-FCP-NEXT:    vpermi2d %zmm0, %zmm1, %zmm614218; AVX512-FCP-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14219; AVX512-FCP-NEXT:    vpermi2d %zmm0, %zmm1, %zmm914220; AVX512-FCP-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14221; AVX512-FCP-NEXT:    vpermi2d %zmm0, %zmm1, %zmm1014222; AVX512-FCP-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14223; AVX512-FCP-NEXT:    vpermi2d %zmm0, %zmm1, %zmm1114224; AVX512-FCP-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14225; AVX512-FCP-NEXT:    vpermi2d %zmm0, %zmm1, %zmm714226; AVX512-FCP-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14227; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm8, %zmm114228; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14229; AVX512-FCP-NEXT:    vmovdqa64 (%r8), %zmm2614230; AVX512-FCP-NEXT:    vmovdqa64 (%r9), %zmm014231; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,0,0,0,16,0,0,0,0,0,0,1,17,0,0]14232; AVX512-FCP-NEXT:    vmovdqa64 %zmm26, %zmm214233; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm1, %zmm214234; AVX512-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14235; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm214236; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,0,0,2,18,0,0,0,0,0,0,3,19,0,0]14237; AVX512-FCP-NEXT:    vmovdqa64 %zmm26, %zmm314238; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm1, %zmm314239; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14240; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm314241; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,0,0,4,20,0,0,0,0,0,0,5,21,0,0]14242; AVX512-FCP-NEXT:    vmovdqa64 %zmm26, %zmm414243; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm1, %zmm414244; AVX512-FCP-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14245; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm414246; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,0,0,6,22,0,0,0,0,0,0,7,23,0,0]14247; AVX512-FCP-NEXT:    vmovdqa64 %zmm26, %zmm514248; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm1, %zmm514249; AVX512-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14250; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm514251; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,0,0,8,24,0,0,0,0,0,0,9,25,0,0]14252; AVX512-FCP-NEXT:    vmovdqa64 %zmm26, %zmm614253; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm1, %zmm614254; AVX512-FCP-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14255; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm614256; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,0,0,10,26,0,0,0,0,0,0,11,27,0,0]14257; AVX512-FCP-NEXT:    vmovdqa64 %zmm26, %zmm714258; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm1, %zmm714259; AVX512-FCP-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14260; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm714261; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,0,0,12,28,0,0,0,0,0,0,13,29,0,0]14262; AVX512-FCP-NEXT:    vmovdqa64 %zmm26, %zmm814263; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm1, %zmm814264; AVX512-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14265; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm814266; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,0,0,14,30,0,0,0,0,0,0,15,31,0,0]14267; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm1, %zmm2614268; AVX512-FCP-NEXT:    vmovdqa64 64(%r8), %zmm2514269; AVX512-FCP-NEXT:    vmovdqa64 64(%r9), %zmm014270; AVX512-FCP-NEXT:    vmovdqa64 %zmm25, %zmm914271; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm2, %zmm914272; AVX512-FCP-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14273; AVX512-FCP-NEXT:    vmovdqa64 %zmm25, %zmm914274; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm3, %zmm914275; AVX512-FCP-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14276; AVX512-FCP-NEXT:    vmovdqa64 %zmm25, %zmm914277; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm4, %zmm914278; AVX512-FCP-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14279; AVX512-FCP-NEXT:    vmovdqa64 %zmm25, %zmm914280; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm5, %zmm914281; AVX512-FCP-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14282; AVX512-FCP-NEXT:    vmovdqa64 %zmm25, %zmm914283; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm6, %zmm914284; AVX512-FCP-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14285; AVX512-FCP-NEXT:    vmovdqa64 %zmm25, %zmm914286; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm7, %zmm914287; AVX512-FCP-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14288; AVX512-FCP-NEXT:    vmovdqa64 %zmm25, %zmm914289; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm8, %zmm914290; AVX512-FCP-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14291; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm1, %zmm2514292; AVX512-FCP-NEXT:    vmovdqa64 128(%r8), %zmm2214293; AVX512-FCP-NEXT:    vmovdqa64 128(%r9), %zmm014294; AVX512-FCP-NEXT:    vmovdqa64 %zmm22, %zmm914295; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm2, %zmm914296; AVX512-FCP-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14297; AVX512-FCP-NEXT:    vmovdqa64 %zmm22, %zmm914298; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm3, %zmm914299; AVX512-FCP-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14300; AVX512-FCP-NEXT:    vmovdqa64 %zmm22, %zmm914301; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm4, %zmm914302; AVX512-FCP-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14303; AVX512-FCP-NEXT:    vmovdqa64 %zmm22, %zmm914304; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm5, %zmm914305; AVX512-FCP-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14306; AVX512-FCP-NEXT:    vmovdqa64 %zmm22, %zmm914307; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm6, %zmm914308; AVX512-FCP-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14309; AVX512-FCP-NEXT:    vmovdqa64 %zmm22, %zmm914310; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm7, %zmm914311; AVX512-FCP-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14312; AVX512-FCP-NEXT:    vmovdqa64 %zmm22, %zmm914313; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm8, %zmm914314; AVX512-FCP-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14315; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm1, %zmm2214316; AVX512-FCP-NEXT:    vmovdqa64 192(%r8), %zmm2714317; AVX512-FCP-NEXT:    vmovdqa64 192(%r9), %zmm014318; AVX512-FCP-NEXT:    vpermi2d %zmm0, %zmm27, %zmm214319; AVX512-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14320; AVX512-FCP-NEXT:    vpermi2d %zmm0, %zmm27, %zmm314321; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14322; AVX512-FCP-NEXT:    vpermi2d %zmm0, %zmm27, %zmm414323; AVX512-FCP-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14324; AVX512-FCP-NEXT:    vpermi2d %zmm0, %zmm27, %zmm514325; AVX512-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14326; AVX512-FCP-NEXT:    vpermi2d %zmm0, %zmm27, %zmm614327; AVX512-FCP-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14328; AVX512-FCP-NEXT:    vpermi2d %zmm0, %zmm27, %zmm714329; AVX512-FCP-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14330; AVX512-FCP-NEXT:    vpermi2d %zmm0, %zmm27, %zmm814331; AVX512-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14332; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm1, %zmm2714333; AVX512-FCP-NEXT:    vmovdqa64 (%rdx), %zmm814334; AVX512-FCP-NEXT:    vmovdqa64 (%rcx), %zmm014335; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,0,16,0,0,0,0,0,0,1,17,0,0,0,0]14336; AVX512-FCP-NEXT:    vmovdqa64 %zmm8, %zmm214337; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm1, %zmm214338; AVX512-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14339; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm214340; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,2,18,0,0,0,0,0,0,3,19,0,0,0,0]14341; AVX512-FCP-NEXT:    vmovdqa64 %zmm8, %zmm314342; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm1, %zmm314343; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14344; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm314345; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,4,20,0,0,0,0,0,0,5,21,0,0,0,0]14346; AVX512-FCP-NEXT:    vmovdqa64 %zmm8, %zmm414347; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm1, %zmm414348; AVX512-FCP-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14349; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm414350; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,6,22,0,0,0,0,0,0,7,23,0,0,0,0]14351; AVX512-FCP-NEXT:    vmovdqa64 %zmm8, %zmm514352; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm1, %zmm514353; AVX512-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14354; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm514355; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm7 = [0,0,8,24,0,0,0,0,0,0,9,25,0,0,0,0]14356; AVX512-FCP-NEXT:    vmovdqa64 %zmm8, %zmm614357; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm7, %zmm614358; AVX512-FCP-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14359; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm9 = [0,0,10,26,0,0,0,0,0,0,11,27,0,0,0,0]14360; AVX512-FCP-NEXT:    vmovdqa64 %zmm8, %zmm614361; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm9, %zmm614362; AVX512-FCP-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14363; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm10 = [0,0,12,28,0,0,0,0,0,0,13,29,0,0,0,0]14364; AVX512-FCP-NEXT:    vmovdqa64 %zmm8, %zmm614365; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm10, %zmm614366; AVX512-FCP-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14367; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,14,30,0,0,0,0,0,0,15,31,0,0,0,0]14368; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm1, %zmm814369; AVX512-FCP-NEXT:    vmovdqa64 64(%rdx), %zmm614370; AVX512-FCP-NEXT:    vmovdqa64 64(%rcx), %zmm014371; AVX512-FCP-NEXT:    vmovdqa64 %zmm6, %zmm1114372; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm2, %zmm1114373; AVX512-FCP-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14374; AVX512-FCP-NEXT:    vmovdqa64 %zmm6, %zmm1114375; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm3, %zmm1114376; AVX512-FCP-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14377; AVX512-FCP-NEXT:    vmovdqa64 %zmm6, %zmm1114378; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm4, %zmm1114379; AVX512-FCP-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14380; AVX512-FCP-NEXT:    vmovdqa64 %zmm6, %zmm1114381; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm5, %zmm1114382; AVX512-FCP-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14383; AVX512-FCP-NEXT:    vmovdqa64 %zmm6, %zmm1114384; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm7, %zmm1114385; AVX512-FCP-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14386; AVX512-FCP-NEXT:    vmovdqa64 %zmm6, %zmm1114387; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm9, %zmm1114388; AVX512-FCP-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14389; AVX512-FCP-NEXT:    vmovdqa64 %zmm6, %zmm1114390; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm10, %zmm1114391; AVX512-FCP-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14392; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm1, %zmm614393; AVX512-FCP-NEXT:    vmovdqa64 128(%rdx), %zmm1314394; AVX512-FCP-NEXT:    vmovdqa64 128(%rcx), %zmm014395; AVX512-FCP-NEXT:    vmovdqa64 %zmm13, %zmm1114396; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm2, %zmm1114397; AVX512-FCP-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14398; AVX512-FCP-NEXT:    vmovdqa64 %zmm13, %zmm1114399; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm3, %zmm1114400; AVX512-FCP-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14401; AVX512-FCP-NEXT:    vmovdqa64 %zmm13, %zmm1114402; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm4, %zmm1114403; AVX512-FCP-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14404; AVX512-FCP-NEXT:    vmovdqa64 %zmm13, %zmm1114405; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm5, %zmm1114406; AVX512-FCP-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14407; AVX512-FCP-NEXT:    vmovdqa64 %zmm13, %zmm1114408; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm7, %zmm1114409; AVX512-FCP-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14410; AVX512-FCP-NEXT:    vmovdqa64 %zmm13, %zmm1114411; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm9, %zmm1114412; AVX512-FCP-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14413; AVX512-FCP-NEXT:    vmovdqa64 %zmm13, %zmm1114414; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm10, %zmm1114415; AVX512-FCP-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14416; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm1, %zmm1314417; AVX512-FCP-NEXT:    vmovdqa64 192(%rdx), %zmm1714418; AVX512-FCP-NEXT:    vmovdqa64 192(%rcx), %zmm014419; AVX512-FCP-NEXT:    vpermi2d %zmm0, %zmm17, %zmm214420; AVX512-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14421; AVX512-FCP-NEXT:    vpermi2d %zmm0, %zmm17, %zmm314422; AVX512-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14423; AVX512-FCP-NEXT:    vpermi2d %zmm0, %zmm17, %zmm414424; AVX512-FCP-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14425; AVX512-FCP-NEXT:    vpermi2d %zmm0, %zmm17, %zmm514426; AVX512-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14427; AVX512-FCP-NEXT:    vpermi2d %zmm0, %zmm17, %zmm714428; AVX512-FCP-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14429; AVX512-FCP-NEXT:    vpermi2d %zmm0, %zmm17, %zmm914430; AVX512-FCP-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14431; AVX512-FCP-NEXT:    vpermi2d %zmm0, %zmm17, %zmm1014432; AVX512-FCP-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14433; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm1, %zmm1714434; AVX512-FCP-NEXT:    vmovdqa64 (%rdi), %zmm514435; AVX512-FCP-NEXT:    vmovdqa64 (%rsi), %zmm014436; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm14 = [0,16,0,0,0,0,0,0,1,17,0,0,0,0,0,0]14437; AVX512-FCP-NEXT:    vmovdqa64 %zmm5, %zmm114438; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm14, %zmm114439; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14440; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm11 = [2,18,0,0,0,0,0,0,3,19,0,0,0,0,0,0]14441; AVX512-FCP-NEXT:    vmovdqa64 %zmm5, %zmm114442; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm11, %zmm114443; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14444; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm10 = [4,20,0,0,0,0,0,0,5,21,0,0,0,0,0,0]14445; AVX512-FCP-NEXT:    vmovdqa64 %zmm5, %zmm114446; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm10, %zmm114447; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14448; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm9 = [6,22,0,0,0,0,0,0,7,23,0,0,0,0,0,0]14449; AVX512-FCP-NEXT:    vmovdqa64 %zmm5, %zmm114450; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm9, %zmm114451; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14452; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm7 = [8,24,0,0,0,0,0,0,9,25,0,0,0,0,0,0]14453; AVX512-FCP-NEXT:    vmovdqa64 %zmm5, %zmm114454; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm7, %zmm114455; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14456; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [10,26,0,0,0,0,0,0,11,27,0,0,0,0,0,0]14457; AVX512-FCP-NEXT:    vmovdqa64 %zmm5, %zmm114458; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm4, %zmm114459; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, (%rsp) # 64-byte Spill14460; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm3 = [12,28,0,0,0,0,0,0,13,29,0,0,0,0,0,0]14461; AVX512-FCP-NEXT:    vmovdqa64 %zmm5, %zmm114462; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm3, %zmm114463; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14464; AVX512-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm31 = [14,30,0,0,0,0,0,0,15,31,0,0,0,0,0,0]14465; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm31, %zmm514466; AVX512-FCP-NEXT:    vmovdqa64 64(%rdi), %zmm1214467; AVX512-FCP-NEXT:    vmovdqa64 64(%rsi), %zmm014468; AVX512-FCP-NEXT:    vmovdqa64 %zmm12, %zmm114469; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm14, %zmm114470; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14471; AVX512-FCP-NEXT:    vmovdqa64 %zmm12, %zmm114472; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm11, %zmm114473; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14474; AVX512-FCP-NEXT:    vmovdqa64 %zmm12, %zmm114475; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm10, %zmm114476; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14477; AVX512-FCP-NEXT:    vmovdqa64 %zmm12, %zmm114478; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm9, %zmm114479; AVX512-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14480; AVX512-FCP-NEXT:    vmovdqa64 %zmm12, %zmm3014481; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm7, %zmm3014482; AVX512-FCP-NEXT:    vmovdqa64 %zmm12, %zmm2914483; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm4, %zmm2914484; AVX512-FCP-NEXT:    vmovdqa64 %zmm12, %zmm2814485; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm3, %zmm2814486; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm31, %zmm1214487; AVX512-FCP-NEXT:    vmovdqa64 128(%rdi), %zmm1514488; AVX512-FCP-NEXT:    vmovdqa64 128(%rsi), %zmm014489; AVX512-FCP-NEXT:    vmovdqa64 %zmm15, %zmm2414490; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm14, %zmm2414491; AVX512-FCP-NEXT:    vmovdqa64 %zmm15, %zmm2314492; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm11, %zmm2314493; AVX512-FCP-NEXT:    vmovdqa64 %zmm15, %zmm2114494; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm10, %zmm2114495; AVX512-FCP-NEXT:    vmovdqa64 %zmm15, %zmm2014496; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm9, %zmm2014497; AVX512-FCP-NEXT:    vmovdqa64 %zmm15, %zmm1914498; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm7, %zmm1914499; AVX512-FCP-NEXT:    vmovdqa64 %zmm15, %zmm1814500; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm4, %zmm1814501; AVX512-FCP-NEXT:    vmovdqa64 %zmm15, %zmm1614502; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm3, %zmm1614503; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm31, %zmm1514504; AVX512-FCP-NEXT:    vmovdqa64 192(%rdi), %zmm214505; AVX512-FCP-NEXT:    vmovdqa64 192(%rsi), %zmm014506; AVX512-FCP-NEXT:    vpermi2d %zmm0, %zmm2, %zmm1414507; AVX512-FCP-NEXT:    vpermi2d %zmm0, %zmm2, %zmm1114508; AVX512-FCP-NEXT:    vpermi2d %zmm0, %zmm2, %zmm1014509; AVX512-FCP-NEXT:    vpermi2d %zmm0, %zmm2, %zmm914510; AVX512-FCP-NEXT:    vpermi2d %zmm0, %zmm2, %zmm714511; AVX512-FCP-NEXT:    vpermi2d %zmm0, %zmm2, %zmm414512; AVX512-FCP-NEXT:    vpermi2d %zmm0, %zmm2, %zmm314513; AVX512-FCP-NEXT:    vpermt2d %zmm0, %zmm31, %zmm214514; AVX512-FCP-NEXT:    movb $-120, %al14515; AVX512-FCP-NEXT:    kmovw %eax, %k114516; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14517; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload14518; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}14519; AVX512-FCP-NEXT:    movb $34, %al14520; AVX512-FCP-NEXT:    kmovw %eax, %k214521; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14522; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload14523; AVX512-FCP-NEXT:    vmovdqa64 %zmm31, %zmm0 {%k2}14524; AVX512-FCP-NEXT:    movb $-52, %al14525; AVX512-FCP-NEXT:    kmovw %eax, %k314526; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm0 {%k3}14527; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14528; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14529; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload14530; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}14531; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14532; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload14533; AVX512-FCP-NEXT:    vmovdqa64 %zmm31, %zmm0 {%k2}14534; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm0 {%k3}14535; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14536; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14537; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload14538; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}14539; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14540; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload14541; AVX512-FCP-NEXT:    vmovdqa64 %zmm31, %zmm0 {%k2}14542; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm0 {%k3}14543; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14544; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14545; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload14546; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}14547; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14548; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload14549; AVX512-FCP-NEXT:    vmovdqa64 %zmm31, %zmm0 {%k2}14550; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm0 {%k3}14551; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14552; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14553; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload14554; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}14555; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14556; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload14557; AVX512-FCP-NEXT:    vmovdqa64 %zmm31, %zmm0 {%k2}14558; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm0 {%k3}14559; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14560; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14561; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload14562; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}14563; AVX512-FCP-NEXT:    vmovdqu64 (%rsp), %zmm0 # 64-byte Reload14564; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload14565; AVX512-FCP-NEXT:    vmovdqa64 %zmm31, %zmm0 {%k2}14566; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm0 {%k3}14567; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, (%rsp) # 64-byte Spill14568; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14569; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload14570; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}14571; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14572; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload14573; AVX512-FCP-NEXT:    vmovdqa64 %zmm31, %zmm0 {%k2}14574; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm0 {%k3}14575; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14576; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14577; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm26 {%k1}14578; AVX512-FCP-NEXT:    vmovdqa64 %zmm8, %zmm5 {%k2}14579; AVX512-FCP-NEXT:    vmovdqa64 %zmm26, %zmm5 {%k3}14580; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14581; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload14582; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}14583; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14584; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload14585; AVX512-FCP-NEXT:    vmovdqa64 %zmm8, %zmm0 {%k2}14586; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, %zmm0 {%k3}14587; AVX512-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14588; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14589; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload14590; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm8 {%k1}14591; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload14592; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14593; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm26 {%k2}14594; AVX512-FCP-NEXT:    vmovdqa64 %zmm8, %zmm26 {%k3}14595; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14596; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload14597; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm8 {%k1}14598; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload14599; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14600; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k2}14601; AVX512-FCP-NEXT:    vmovdqa64 %zmm8, %zmm1 {%k3}14602; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14603; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload14604; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm8 {%k1}14605; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14606; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload14607; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm31 {%k2}14608; AVX512-FCP-NEXT:    vmovdqa64 %zmm8, %zmm31 {%k3}14609; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14610; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload14611; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm8 {%k1}14612; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14613; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm30 {%k2}14614; AVX512-FCP-NEXT:    vmovdqa64 %zmm8, %zmm30 {%k3}14615; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14616; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload14617; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm8 {%k1}14618; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14619; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm29 {%k2}14620; AVX512-FCP-NEXT:    vmovdqa64 %zmm8, %zmm29 {%k3}14621; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14622; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload14623; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm8 {%k1}14624; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14625; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm28 {%k2}14626; AVX512-FCP-NEXT:    vmovdqa64 %zmm8, %zmm28 {%k3}14627; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14628; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm25 {%k1}14629; AVX512-FCP-NEXT:    vmovdqa64 %zmm6, %zmm12 {%k2}14630; AVX512-FCP-NEXT:    vmovdqa64 %zmm25, %zmm12 {%k3}14631; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14632; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload14633; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm6 {%k1}14634; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14635; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm24 {%k2}14636; AVX512-FCP-NEXT:    vmovdqa64 %zmm6, %zmm24 {%k3}14637; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14638; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload14639; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm6 {%k1}14640; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14641; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm23 {%k2}14642; AVX512-FCP-NEXT:    vmovdqa64 %zmm6, %zmm23 {%k3}14643; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14644; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload14645; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm6 {%k1}14646; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14647; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm21 {%k2}14648; AVX512-FCP-NEXT:    vmovdqa64 %zmm6, %zmm21 {%k3}14649; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14650; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload14651; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm6 {%k1}14652; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14653; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm20 {%k2}14654; AVX512-FCP-NEXT:    vmovdqa64 %zmm6, %zmm20 {%k3}14655; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14656; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload14657; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm6 {%k1}14658; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14659; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm19 {%k2}14660; AVX512-FCP-NEXT:    vmovdqa64 %zmm6, %zmm19 {%k3}14661; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14662; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload14663; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm6 {%k1}14664; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14665; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm18 {%k2}14666; AVX512-FCP-NEXT:    vmovdqa64 %zmm6, %zmm18 {%k3}14667; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14668; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload14669; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm6 {%k1}14670; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14671; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm16 {%k2}14672; AVX512-FCP-NEXT:    vmovdqa64 %zmm6, %zmm16 {%k3}14673; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14674; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm22 {%k1}14675; AVX512-FCP-NEXT:    vmovdqa64 %zmm13, %zmm15 {%k2}14676; AVX512-FCP-NEXT:    vmovdqa64 %zmm22, %zmm15 {%k3}14677; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14678; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload14679; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm6 {%k1}14680; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14681; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm14 {%k2}14682; AVX512-FCP-NEXT:    vmovdqa64 %zmm6, %zmm14 {%k3}14683; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14684; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload14685; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm6 {%k1}14686; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14687; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm11 {%k2}14688; AVX512-FCP-NEXT:    vmovdqa64 %zmm6, %zmm11 {%k3}14689; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14690; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload14691; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm6 {%k1}14692; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14693; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm10 {%k2}14694; AVX512-FCP-NEXT:    vmovdqa64 %zmm6, %zmm10 {%k3}14695; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14696; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload14697; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm6 {%k1}14698; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14699; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm9 {%k2}14700; AVX512-FCP-NEXT:    vmovdqa64 %zmm6, %zmm9 {%k3}14701; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14702; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload14703; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm6 {%k1}14704; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14705; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm7 {%k2}14706; AVX512-FCP-NEXT:    vmovdqa64 %zmm6, %zmm7 {%k3}14707; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14708; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload14709; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm6 {%k1}14710; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14711; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm4 {%k2}14712; AVX512-FCP-NEXT:    vmovdqa64 %zmm6, %zmm4 {%k3}14713; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14714; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload14715; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm6 {%k1}14716; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14717; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm3 {%k2}14718; AVX512-FCP-NEXT:    vmovdqa64 %zmm6, %zmm3 {%k3}14719; AVX512-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14720; AVX512-FCP-NEXT:    vmovdqa64 %zmm0, %zmm27 {%k1}14721; AVX512-FCP-NEXT:    vmovdqa64 %zmm17, %zmm2 {%k2}14722; AVX512-FCP-NEXT:    vmovdqa64 %zmm27, %zmm2 {%k3}14723; AVX512-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax14724; AVX512-FCP-NEXT:    vmovdqa64 %zmm2, 1984(%rax)14725; AVX512-FCP-NEXT:    vmovdqa64 %zmm3, 1920(%rax)14726; AVX512-FCP-NEXT:    vmovdqa64 %zmm4, 1856(%rax)14727; AVX512-FCP-NEXT:    vmovdqa64 %zmm7, 1792(%rax)14728; AVX512-FCP-NEXT:    vmovdqa64 %zmm9, 1728(%rax)14729; AVX512-FCP-NEXT:    vmovdqa64 %zmm10, 1664(%rax)14730; AVX512-FCP-NEXT:    vmovdqa64 %zmm11, 1600(%rax)14731; AVX512-FCP-NEXT:    vmovdqa64 %zmm14, 1536(%rax)14732; AVX512-FCP-NEXT:    vmovdqa64 %zmm15, 1472(%rax)14733; AVX512-FCP-NEXT:    vmovdqa64 %zmm16, 1408(%rax)14734; AVX512-FCP-NEXT:    vmovdqa64 %zmm18, 1344(%rax)14735; AVX512-FCP-NEXT:    vmovdqa64 %zmm19, 1280(%rax)14736; AVX512-FCP-NEXT:    vmovdqa64 %zmm20, 1216(%rax)14737; AVX512-FCP-NEXT:    vmovdqa64 %zmm21, 1152(%rax)14738; AVX512-FCP-NEXT:    vmovdqa64 %zmm23, 1088(%rax)14739; AVX512-FCP-NEXT:    vmovdqa64 %zmm24, 1024(%rax)14740; AVX512-FCP-NEXT:    vmovdqa64 %zmm12, 960(%rax)14741; AVX512-FCP-NEXT:    vmovdqa64 %zmm28, 896(%rax)14742; AVX512-FCP-NEXT:    vmovdqa64 %zmm29, 832(%rax)14743; AVX512-FCP-NEXT:    vmovdqa64 %zmm30, 768(%rax)14744; AVX512-FCP-NEXT:    vmovdqa64 %zmm31, 704(%rax)14745; AVX512-FCP-NEXT:    vmovdqa64 %zmm1, 640(%rax)14746; AVX512-FCP-NEXT:    vmovdqa64 %zmm26, 576(%rax)14747; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14748; AVX512-FCP-NEXT:    vmovaps %zmm0, 512(%rax)14749; AVX512-FCP-NEXT:    vmovdqa64 %zmm5, 448(%rax)14750; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14751; AVX512-FCP-NEXT:    vmovaps %zmm0, 384(%rax)14752; AVX512-FCP-NEXT:    vmovups (%rsp), %zmm0 # 64-byte Reload14753; AVX512-FCP-NEXT:    vmovaps %zmm0, 320(%rax)14754; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14755; AVX512-FCP-NEXT:    vmovaps %zmm0, 256(%rax)14756; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14757; AVX512-FCP-NEXT:    vmovaps %zmm0, 192(%rax)14758; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14759; AVX512-FCP-NEXT:    vmovaps %zmm0, 128(%rax)14760; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14761; AVX512-FCP-NEXT:    vmovaps %zmm0, 64(%rax)14762; AVX512-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14763; AVX512-FCP-NEXT:    vmovaps %zmm0, (%rax)14764; AVX512-FCP-NEXT:    addq $6216, %rsp # imm = 0x184814765; AVX512-FCP-NEXT:    vzeroupper14766; AVX512-FCP-NEXT:    retq14767;14768; AVX512DQ-LABEL: store_i32_stride8_vf64:14769; AVX512DQ:       # %bb.0:14770; AVX512DQ-NEXT:    subq $6216, %rsp # imm = 0x184814771; AVX512DQ-NEXT:    movq {{[0-9]+}}(%rsp), %rax14772; AVX512DQ-NEXT:    movq {{[0-9]+}}(%rsp), %r1014773; AVX512DQ-NEXT:    vmovdqa64 (%r10), %zmm514774; AVX512DQ-NEXT:    vmovdqa64 64(%r10), %zmm414775; AVX512DQ-NEXT:    vmovdqa64 128(%r10), %zmm214776; AVX512DQ-NEXT:    vmovdqa64 (%rax), %zmm114777; AVX512DQ-NEXT:    vmovdqa64 64(%rax), %zmm014778; AVX512DQ-NEXT:    vmovdqa64 128(%rax), %zmm3014779; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm3 = [0,0,0,0,0,0,0,16,0,0,0,0,0,0,1,17]14780; AVX512DQ-NEXT:    vmovdqa64 %zmm5, %zmm714781; AVX512DQ-NEXT:    vmovdqa64 %zmm5, %zmm614782; AVX512DQ-NEXT:    vpermt2d %zmm1, %zmm3, %zmm714783; AVX512DQ-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14784; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm5 = [0,0,0,0,0,0,2,18,0,0,0,0,0,0,3,19]14785; AVX512DQ-NEXT:    vmovdqa64 %zmm6, %zmm714786; AVX512DQ-NEXT:    vpermt2d %zmm1, %zmm5, %zmm614787; AVX512DQ-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14788; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm6 = [0,0,0,0,0,0,4,20,0,0,0,0,0,0,5,21]14789; AVX512DQ-NEXT:    vmovdqa64 %zmm7, %zmm814790; AVX512DQ-NEXT:    vpermt2d %zmm1, %zmm6, %zmm814791; AVX512DQ-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14792; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm8 = [0,0,0,0,0,0,6,22,0,0,0,0,0,0,7,23]14793; AVX512DQ-NEXT:    vmovdqa64 %zmm7, %zmm914794; AVX512DQ-NEXT:    vpermt2d %zmm1, %zmm8, %zmm914795; AVX512DQ-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14796; AVX512DQ-NEXT:    vmovdqa64 %zmm8, %zmm914797; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm8 = [0,0,0,0,0,0,8,24,0,0,0,0,0,0,9,25]14798; AVX512DQ-NEXT:    vmovdqa64 %zmm7, %zmm1014799; AVX512DQ-NEXT:    vpermt2d %zmm1, %zmm8, %zmm1014800; AVX512DQ-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14801; AVX512DQ-NEXT:    vmovdqa64 %zmm8, %zmm1014802; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm8 = [0,0,0,0,0,0,10,26,0,0,0,0,0,0,11,27]14803; AVX512DQ-NEXT:    vmovdqa64 %zmm7, %zmm1114804; AVX512DQ-NEXT:    vpermt2d %zmm1, %zmm8, %zmm1114805; AVX512DQ-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14806; AVX512DQ-NEXT:    vmovdqa64 %zmm8, %zmm1114807; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm8 = [0,0,0,0,0,0,12,28,0,0,0,0,0,0,13,29]14808; AVX512DQ-NEXT:    vmovdqa64 %zmm7, %zmm1314809; AVX512DQ-NEXT:    vmovdqa64 %zmm7, %zmm1214810; AVX512DQ-NEXT:    vpermt2d %zmm1, %zmm8, %zmm1314811; AVX512DQ-NEXT:    vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14812; AVX512DQ-NEXT:    vmovdqa64 %zmm8, %zmm714813; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm8 = [0,0,0,0,0,0,14,30,0,0,0,0,0,0,15,31]14814; AVX512DQ-NEXT:    vpermt2d %zmm1, %zmm8, %zmm1214815; AVX512DQ-NEXT:    vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14816; AVX512DQ-NEXT:    vmovdqa64 %zmm4, %zmm114817; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm3, %zmm114818; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14819; AVX512DQ-NEXT:    vmovdqa64 %zmm4, %zmm114820; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm5, %zmm114821; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14822; AVX512DQ-NEXT:    vmovdqa64 %zmm4, %zmm114823; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm6, %zmm114824; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14825; AVX512DQ-NEXT:    vmovdqa64 %zmm4, %zmm114826; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm9, %zmm114827; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14828; AVX512DQ-NEXT:    vmovdqa64 %zmm4, %zmm114829; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm10, %zmm114830; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14831; AVX512DQ-NEXT:    vmovdqa64 %zmm4, %zmm114832; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm11, %zmm114833; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14834; AVX512DQ-NEXT:    vmovdqa64 %zmm4, %zmm114835; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm7, %zmm114836; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14837; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm8, %zmm414838; AVX512DQ-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14839; AVX512DQ-NEXT:    vmovdqa64 %zmm2, %zmm014840; AVX512DQ-NEXT:    vpermt2d %zmm30, %zmm3, %zmm014841; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14842; AVX512DQ-NEXT:    vmovdqa64 %zmm2, %zmm014843; AVX512DQ-NEXT:    vpermt2d %zmm30, %zmm5, %zmm014844; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14845; AVX512DQ-NEXT:    vmovdqa64 %zmm2, %zmm014846; AVX512DQ-NEXT:    vpermt2d %zmm30, %zmm6, %zmm014847; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14848; AVX512DQ-NEXT:    vmovdqa64 %zmm2, %zmm014849; AVX512DQ-NEXT:    vpermt2d %zmm30, %zmm9, %zmm014850; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14851; AVX512DQ-NEXT:    vmovdqa64 %zmm2, %zmm014852; AVX512DQ-NEXT:    vpermt2d %zmm30, %zmm10, %zmm014853; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14854; AVX512DQ-NEXT:    vmovdqa64 %zmm2, %zmm014855; AVX512DQ-NEXT:    vpermt2d %zmm30, %zmm11, %zmm014856; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14857; AVX512DQ-NEXT:    vmovdqa64 %zmm2, %zmm014858; AVX512DQ-NEXT:    vpermt2d %zmm30, %zmm7, %zmm014859; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14860; AVX512DQ-NEXT:    vpermt2d %zmm30, %zmm8, %zmm214861; AVX512DQ-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14862; AVX512DQ-NEXT:    vmovdqa64 192(%r10), %zmm114863; AVX512DQ-NEXT:    vmovdqa64 192(%rax), %zmm014864; AVX512DQ-NEXT:    vpermi2d %zmm0, %zmm1, %zmm314865; AVX512DQ-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14866; AVX512DQ-NEXT:    vpermi2d %zmm0, %zmm1, %zmm514867; AVX512DQ-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14868; AVX512DQ-NEXT:    vpermi2d %zmm0, %zmm1, %zmm614869; AVX512DQ-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14870; AVX512DQ-NEXT:    vpermi2d %zmm0, %zmm1, %zmm914871; AVX512DQ-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14872; AVX512DQ-NEXT:    vpermi2d %zmm0, %zmm1, %zmm1014873; AVX512DQ-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14874; AVX512DQ-NEXT:    vpermi2d %zmm0, %zmm1, %zmm1114875; AVX512DQ-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14876; AVX512DQ-NEXT:    vpermi2d %zmm0, %zmm1, %zmm714877; AVX512DQ-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14878; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm8, %zmm114879; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14880; AVX512DQ-NEXT:    vmovdqa64 (%r8), %zmm2614881; AVX512DQ-NEXT:    vmovdqa64 (%r9), %zmm014882; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,0,0,0,16,0,0,0,0,0,0,1,17,0,0]14883; AVX512DQ-NEXT:    vmovdqa64 %zmm26, %zmm214884; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm1, %zmm214885; AVX512DQ-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14886; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm214887; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,0,0,2,18,0,0,0,0,0,0,3,19,0,0]14888; AVX512DQ-NEXT:    vmovdqa64 %zmm26, %zmm314889; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm1, %zmm314890; AVX512DQ-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14891; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm314892; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,0,0,4,20,0,0,0,0,0,0,5,21,0,0]14893; AVX512DQ-NEXT:    vmovdqa64 %zmm26, %zmm414894; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm1, %zmm414895; AVX512DQ-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14896; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm414897; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,0,0,6,22,0,0,0,0,0,0,7,23,0,0]14898; AVX512DQ-NEXT:    vmovdqa64 %zmm26, %zmm514899; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm1, %zmm514900; AVX512DQ-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14901; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm514902; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,0,0,8,24,0,0,0,0,0,0,9,25,0,0]14903; AVX512DQ-NEXT:    vmovdqa64 %zmm26, %zmm614904; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm1, %zmm614905; AVX512DQ-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14906; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm614907; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,0,0,10,26,0,0,0,0,0,0,11,27,0,0]14908; AVX512DQ-NEXT:    vmovdqa64 %zmm26, %zmm714909; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm1, %zmm714910; AVX512DQ-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14911; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm714912; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,0,0,12,28,0,0,0,0,0,0,13,29,0,0]14913; AVX512DQ-NEXT:    vmovdqa64 %zmm26, %zmm814914; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm1, %zmm814915; AVX512DQ-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14916; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm814917; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,0,0,14,30,0,0,0,0,0,0,15,31,0,0]14918; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm1, %zmm2614919; AVX512DQ-NEXT:    vmovdqa64 64(%r8), %zmm2514920; AVX512DQ-NEXT:    vmovdqa64 64(%r9), %zmm014921; AVX512DQ-NEXT:    vmovdqa64 %zmm25, %zmm914922; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm2, %zmm914923; AVX512DQ-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14924; AVX512DQ-NEXT:    vmovdqa64 %zmm25, %zmm914925; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm3, %zmm914926; AVX512DQ-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14927; AVX512DQ-NEXT:    vmovdqa64 %zmm25, %zmm914928; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm4, %zmm914929; AVX512DQ-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14930; AVX512DQ-NEXT:    vmovdqa64 %zmm25, %zmm914931; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm5, %zmm914932; AVX512DQ-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14933; AVX512DQ-NEXT:    vmovdqa64 %zmm25, %zmm914934; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm6, %zmm914935; AVX512DQ-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14936; AVX512DQ-NEXT:    vmovdqa64 %zmm25, %zmm914937; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm7, %zmm914938; AVX512DQ-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14939; AVX512DQ-NEXT:    vmovdqa64 %zmm25, %zmm914940; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm8, %zmm914941; AVX512DQ-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14942; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm1, %zmm2514943; AVX512DQ-NEXT:    vmovdqa64 128(%r8), %zmm2214944; AVX512DQ-NEXT:    vmovdqa64 128(%r9), %zmm014945; AVX512DQ-NEXT:    vmovdqa64 %zmm22, %zmm914946; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm2, %zmm914947; AVX512DQ-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14948; AVX512DQ-NEXT:    vmovdqa64 %zmm22, %zmm914949; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm3, %zmm914950; AVX512DQ-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14951; AVX512DQ-NEXT:    vmovdqa64 %zmm22, %zmm914952; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm4, %zmm914953; AVX512DQ-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14954; AVX512DQ-NEXT:    vmovdqa64 %zmm22, %zmm914955; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm5, %zmm914956; AVX512DQ-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14957; AVX512DQ-NEXT:    vmovdqa64 %zmm22, %zmm914958; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm6, %zmm914959; AVX512DQ-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14960; AVX512DQ-NEXT:    vmovdqa64 %zmm22, %zmm914961; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm7, %zmm914962; AVX512DQ-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14963; AVX512DQ-NEXT:    vmovdqa64 %zmm22, %zmm914964; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm8, %zmm914965; AVX512DQ-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14966; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm1, %zmm2214967; AVX512DQ-NEXT:    vmovdqa64 192(%r8), %zmm2714968; AVX512DQ-NEXT:    vmovdqa64 192(%r9), %zmm014969; AVX512DQ-NEXT:    vpermi2d %zmm0, %zmm27, %zmm214970; AVX512DQ-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14971; AVX512DQ-NEXT:    vpermi2d %zmm0, %zmm27, %zmm314972; AVX512DQ-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14973; AVX512DQ-NEXT:    vpermi2d %zmm0, %zmm27, %zmm414974; AVX512DQ-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14975; AVX512DQ-NEXT:    vpermi2d %zmm0, %zmm27, %zmm514976; AVX512DQ-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14977; AVX512DQ-NEXT:    vpermi2d %zmm0, %zmm27, %zmm614978; AVX512DQ-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14979; AVX512DQ-NEXT:    vpermi2d %zmm0, %zmm27, %zmm714980; AVX512DQ-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14981; AVX512DQ-NEXT:    vpermi2d %zmm0, %zmm27, %zmm814982; AVX512DQ-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14983; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm1, %zmm2714984; AVX512DQ-NEXT:    vmovdqa64 (%rdx), %zmm814985; AVX512DQ-NEXT:    vmovdqa64 (%rcx), %zmm014986; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,0,16,0,0,0,0,0,0,1,17,0,0,0,0]14987; AVX512DQ-NEXT:    vmovdqa64 %zmm8, %zmm214988; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm1, %zmm214989; AVX512DQ-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14990; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm214991; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,2,18,0,0,0,0,0,0,3,19,0,0,0,0]14992; AVX512DQ-NEXT:    vmovdqa64 %zmm8, %zmm314993; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm1, %zmm314994; AVX512DQ-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14995; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm314996; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,4,20,0,0,0,0,0,0,5,21,0,0,0,0]14997; AVX512DQ-NEXT:    vmovdqa64 %zmm8, %zmm414998; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm1, %zmm414999; AVX512DQ-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15000; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm415001; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,6,22,0,0,0,0,0,0,7,23,0,0,0,0]15002; AVX512DQ-NEXT:    vmovdqa64 %zmm8, %zmm515003; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm1, %zmm515004; AVX512DQ-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15005; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm515006; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm7 = [0,0,8,24,0,0,0,0,0,0,9,25,0,0,0,0]15007; AVX512DQ-NEXT:    vmovdqa64 %zmm8, %zmm615008; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm7, %zmm615009; AVX512DQ-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15010; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm9 = [0,0,10,26,0,0,0,0,0,0,11,27,0,0,0,0]15011; AVX512DQ-NEXT:    vmovdqa64 %zmm8, %zmm615012; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm9, %zmm615013; AVX512DQ-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15014; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm10 = [0,0,12,28,0,0,0,0,0,0,13,29,0,0,0,0]15015; AVX512DQ-NEXT:    vmovdqa64 %zmm8, %zmm615016; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm10, %zmm615017; AVX512DQ-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15018; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,14,30,0,0,0,0,0,0,15,31,0,0,0,0]15019; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm1, %zmm815020; AVX512DQ-NEXT:    vmovdqa64 64(%rdx), %zmm615021; AVX512DQ-NEXT:    vmovdqa64 64(%rcx), %zmm015022; AVX512DQ-NEXT:    vmovdqa64 %zmm6, %zmm1115023; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm2, %zmm1115024; AVX512DQ-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15025; AVX512DQ-NEXT:    vmovdqa64 %zmm6, %zmm1115026; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm3, %zmm1115027; AVX512DQ-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15028; AVX512DQ-NEXT:    vmovdqa64 %zmm6, %zmm1115029; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm4, %zmm1115030; AVX512DQ-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15031; AVX512DQ-NEXT:    vmovdqa64 %zmm6, %zmm1115032; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm5, %zmm1115033; AVX512DQ-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15034; AVX512DQ-NEXT:    vmovdqa64 %zmm6, %zmm1115035; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm7, %zmm1115036; AVX512DQ-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15037; AVX512DQ-NEXT:    vmovdqa64 %zmm6, %zmm1115038; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm9, %zmm1115039; AVX512DQ-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15040; AVX512DQ-NEXT:    vmovdqa64 %zmm6, %zmm1115041; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm10, %zmm1115042; AVX512DQ-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15043; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm1, %zmm615044; AVX512DQ-NEXT:    vmovdqa64 128(%rdx), %zmm1315045; AVX512DQ-NEXT:    vmovdqa64 128(%rcx), %zmm015046; AVX512DQ-NEXT:    vmovdqa64 %zmm13, %zmm1115047; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm2, %zmm1115048; AVX512DQ-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15049; AVX512DQ-NEXT:    vmovdqa64 %zmm13, %zmm1115050; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm3, %zmm1115051; AVX512DQ-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15052; AVX512DQ-NEXT:    vmovdqa64 %zmm13, %zmm1115053; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm4, %zmm1115054; AVX512DQ-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15055; AVX512DQ-NEXT:    vmovdqa64 %zmm13, %zmm1115056; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm5, %zmm1115057; AVX512DQ-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15058; AVX512DQ-NEXT:    vmovdqa64 %zmm13, %zmm1115059; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm7, %zmm1115060; AVX512DQ-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15061; AVX512DQ-NEXT:    vmovdqa64 %zmm13, %zmm1115062; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm9, %zmm1115063; AVX512DQ-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15064; AVX512DQ-NEXT:    vmovdqa64 %zmm13, %zmm1115065; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm10, %zmm1115066; AVX512DQ-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15067; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm1, %zmm1315068; AVX512DQ-NEXT:    vmovdqa64 192(%rdx), %zmm1715069; AVX512DQ-NEXT:    vmovdqa64 192(%rcx), %zmm015070; AVX512DQ-NEXT:    vpermi2d %zmm0, %zmm17, %zmm215071; AVX512DQ-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15072; AVX512DQ-NEXT:    vpermi2d %zmm0, %zmm17, %zmm315073; AVX512DQ-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15074; AVX512DQ-NEXT:    vpermi2d %zmm0, %zmm17, %zmm415075; AVX512DQ-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15076; AVX512DQ-NEXT:    vpermi2d %zmm0, %zmm17, %zmm515077; AVX512DQ-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15078; AVX512DQ-NEXT:    vpermi2d %zmm0, %zmm17, %zmm715079; AVX512DQ-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15080; AVX512DQ-NEXT:    vpermi2d %zmm0, %zmm17, %zmm915081; AVX512DQ-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15082; AVX512DQ-NEXT:    vpermi2d %zmm0, %zmm17, %zmm1015083; AVX512DQ-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15084; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm1, %zmm1715085; AVX512DQ-NEXT:    vmovdqa64 (%rdi), %zmm515086; AVX512DQ-NEXT:    vmovdqa64 (%rsi), %zmm015087; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm14 = [0,16,0,0,0,0,0,0,1,17,0,0,0,0,0,0]15088; AVX512DQ-NEXT:    vmovdqa64 %zmm5, %zmm115089; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm14, %zmm115090; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15091; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm11 = [2,18,0,0,0,0,0,0,3,19,0,0,0,0,0,0]15092; AVX512DQ-NEXT:    vmovdqa64 %zmm5, %zmm115093; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm11, %zmm115094; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15095; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm10 = [4,20,0,0,0,0,0,0,5,21,0,0,0,0,0,0]15096; AVX512DQ-NEXT:    vmovdqa64 %zmm5, %zmm115097; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm10, %zmm115098; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15099; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm9 = [6,22,0,0,0,0,0,0,7,23,0,0,0,0,0,0]15100; AVX512DQ-NEXT:    vmovdqa64 %zmm5, %zmm115101; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm9, %zmm115102; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15103; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm7 = [8,24,0,0,0,0,0,0,9,25,0,0,0,0,0,0]15104; AVX512DQ-NEXT:    vmovdqa64 %zmm5, %zmm115105; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm7, %zmm115106; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15107; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [10,26,0,0,0,0,0,0,11,27,0,0,0,0,0,0]15108; AVX512DQ-NEXT:    vmovdqa64 %zmm5, %zmm115109; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm4, %zmm115110; AVX512DQ-NEXT:    vmovdqu64 %zmm1, (%rsp) # 64-byte Spill15111; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm3 = [12,28,0,0,0,0,0,0,13,29,0,0,0,0,0,0]15112; AVX512DQ-NEXT:    vmovdqa64 %zmm5, %zmm115113; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm3, %zmm115114; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15115; AVX512DQ-NEXT:    vpmovsxbd {{.*#+}} zmm31 = [14,30,0,0,0,0,0,0,15,31,0,0,0,0,0,0]15116; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm31, %zmm515117; AVX512DQ-NEXT:    vmovdqa64 64(%rdi), %zmm1215118; AVX512DQ-NEXT:    vmovdqa64 64(%rsi), %zmm015119; AVX512DQ-NEXT:    vmovdqa64 %zmm12, %zmm115120; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm14, %zmm115121; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15122; AVX512DQ-NEXT:    vmovdqa64 %zmm12, %zmm115123; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm11, %zmm115124; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15125; AVX512DQ-NEXT:    vmovdqa64 %zmm12, %zmm115126; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm10, %zmm115127; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15128; AVX512DQ-NEXT:    vmovdqa64 %zmm12, %zmm115129; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm9, %zmm115130; AVX512DQ-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15131; AVX512DQ-NEXT:    vmovdqa64 %zmm12, %zmm3015132; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm7, %zmm3015133; AVX512DQ-NEXT:    vmovdqa64 %zmm12, %zmm2915134; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm4, %zmm2915135; AVX512DQ-NEXT:    vmovdqa64 %zmm12, %zmm2815136; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm3, %zmm2815137; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm31, %zmm1215138; AVX512DQ-NEXT:    vmovdqa64 128(%rdi), %zmm1515139; AVX512DQ-NEXT:    vmovdqa64 128(%rsi), %zmm015140; AVX512DQ-NEXT:    vmovdqa64 %zmm15, %zmm2415141; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm14, %zmm2415142; AVX512DQ-NEXT:    vmovdqa64 %zmm15, %zmm2315143; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm11, %zmm2315144; AVX512DQ-NEXT:    vmovdqa64 %zmm15, %zmm2115145; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm10, %zmm2115146; AVX512DQ-NEXT:    vmovdqa64 %zmm15, %zmm2015147; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm9, %zmm2015148; AVX512DQ-NEXT:    vmovdqa64 %zmm15, %zmm1915149; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm7, %zmm1915150; AVX512DQ-NEXT:    vmovdqa64 %zmm15, %zmm1815151; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm4, %zmm1815152; AVX512DQ-NEXT:    vmovdqa64 %zmm15, %zmm1615153; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm3, %zmm1615154; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm31, %zmm1515155; AVX512DQ-NEXT:    vmovdqa64 192(%rdi), %zmm215156; AVX512DQ-NEXT:    vmovdqa64 192(%rsi), %zmm015157; AVX512DQ-NEXT:    vpermi2d %zmm0, %zmm2, %zmm1415158; AVX512DQ-NEXT:    vpermi2d %zmm0, %zmm2, %zmm1115159; AVX512DQ-NEXT:    vpermi2d %zmm0, %zmm2, %zmm1015160; AVX512DQ-NEXT:    vpermi2d %zmm0, %zmm2, %zmm915161; AVX512DQ-NEXT:    vpermi2d %zmm0, %zmm2, %zmm715162; AVX512DQ-NEXT:    vpermi2d %zmm0, %zmm2, %zmm415163; AVX512DQ-NEXT:    vpermi2d %zmm0, %zmm2, %zmm315164; AVX512DQ-NEXT:    vpermt2d %zmm0, %zmm31, %zmm215165; AVX512DQ-NEXT:    movb $-120, %al15166; AVX512DQ-NEXT:    kmovw %eax, %k115167; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15168; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload15169; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}15170; AVX512DQ-NEXT:    movb $34, %al15171; AVX512DQ-NEXT:    kmovw %eax, %k215172; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15173; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload15174; AVX512DQ-NEXT:    vmovdqa64 %zmm31, %zmm0 {%k2}15175; AVX512DQ-NEXT:    movb $-52, %al15176; AVX512DQ-NEXT:    kmovw %eax, %k315177; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm0 {%k3}15178; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15179; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15180; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload15181; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}15182; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15183; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload15184; AVX512DQ-NEXT:    vmovdqa64 %zmm31, %zmm0 {%k2}15185; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm0 {%k3}15186; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15187; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15188; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload15189; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}15190; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15191; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload15192; AVX512DQ-NEXT:    vmovdqa64 %zmm31, %zmm0 {%k2}15193; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm0 {%k3}15194; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15195; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15196; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload15197; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}15198; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15199; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload15200; AVX512DQ-NEXT:    vmovdqa64 %zmm31, %zmm0 {%k2}15201; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm0 {%k3}15202; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15203; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15204; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload15205; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}15206; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15207; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload15208; AVX512DQ-NEXT:    vmovdqa64 %zmm31, %zmm0 {%k2}15209; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm0 {%k3}15210; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15211; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15212; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload15213; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}15214; AVX512DQ-NEXT:    vmovdqu64 (%rsp), %zmm0 # 64-byte Reload15215; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload15216; AVX512DQ-NEXT:    vmovdqa64 %zmm31, %zmm0 {%k2}15217; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm0 {%k3}15218; AVX512DQ-NEXT:    vmovdqu64 %zmm0, (%rsp) # 64-byte Spill15219; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15220; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload15221; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}15222; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15223; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload15224; AVX512DQ-NEXT:    vmovdqa64 %zmm31, %zmm0 {%k2}15225; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm0 {%k3}15226; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15227; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15228; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm26 {%k1}15229; AVX512DQ-NEXT:    vmovdqa64 %zmm8, %zmm5 {%k2}15230; AVX512DQ-NEXT:    vmovdqa64 %zmm26, %zmm5 {%k3}15231; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15232; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload15233; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}15234; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15235; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload15236; AVX512DQ-NEXT:    vmovdqa64 %zmm8, %zmm0 {%k2}15237; AVX512DQ-NEXT:    vmovdqa64 %zmm1, %zmm0 {%k3}15238; AVX512DQ-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15239; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15240; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload15241; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm8 {%k1}15242; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload15243; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15244; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm26 {%k2}15245; AVX512DQ-NEXT:    vmovdqa64 %zmm8, %zmm26 {%k3}15246; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15247; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload15248; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm8 {%k1}15249; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload15250; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15251; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k2}15252; AVX512DQ-NEXT:    vmovdqa64 %zmm8, %zmm1 {%k3}15253; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15254; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload15255; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm8 {%k1}15256; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15257; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload15258; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm31 {%k2}15259; AVX512DQ-NEXT:    vmovdqa64 %zmm8, %zmm31 {%k3}15260; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15261; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload15262; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm8 {%k1}15263; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15264; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm30 {%k2}15265; AVX512DQ-NEXT:    vmovdqa64 %zmm8, %zmm30 {%k3}15266; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15267; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload15268; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm8 {%k1}15269; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15270; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm29 {%k2}15271; AVX512DQ-NEXT:    vmovdqa64 %zmm8, %zmm29 {%k3}15272; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15273; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload15274; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm8 {%k1}15275; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15276; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm28 {%k2}15277; AVX512DQ-NEXT:    vmovdqa64 %zmm8, %zmm28 {%k3}15278; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15279; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm25 {%k1}15280; AVX512DQ-NEXT:    vmovdqa64 %zmm6, %zmm12 {%k2}15281; AVX512DQ-NEXT:    vmovdqa64 %zmm25, %zmm12 {%k3}15282; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15283; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload15284; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm6 {%k1}15285; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15286; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm24 {%k2}15287; AVX512DQ-NEXT:    vmovdqa64 %zmm6, %zmm24 {%k3}15288; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15289; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload15290; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm6 {%k1}15291; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15292; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm23 {%k2}15293; AVX512DQ-NEXT:    vmovdqa64 %zmm6, %zmm23 {%k3}15294; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15295; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload15296; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm6 {%k1}15297; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15298; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm21 {%k2}15299; AVX512DQ-NEXT:    vmovdqa64 %zmm6, %zmm21 {%k3}15300; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15301; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload15302; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm6 {%k1}15303; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15304; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm20 {%k2}15305; AVX512DQ-NEXT:    vmovdqa64 %zmm6, %zmm20 {%k3}15306; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15307; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload15308; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm6 {%k1}15309; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15310; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm19 {%k2}15311; AVX512DQ-NEXT:    vmovdqa64 %zmm6, %zmm19 {%k3}15312; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15313; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload15314; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm6 {%k1}15315; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15316; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm18 {%k2}15317; AVX512DQ-NEXT:    vmovdqa64 %zmm6, %zmm18 {%k3}15318; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15319; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload15320; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm6 {%k1}15321; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15322; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm16 {%k2}15323; AVX512DQ-NEXT:    vmovdqa64 %zmm6, %zmm16 {%k3}15324; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15325; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm22 {%k1}15326; AVX512DQ-NEXT:    vmovdqa64 %zmm13, %zmm15 {%k2}15327; AVX512DQ-NEXT:    vmovdqa64 %zmm22, %zmm15 {%k3}15328; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15329; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload15330; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm6 {%k1}15331; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15332; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm14 {%k2}15333; AVX512DQ-NEXT:    vmovdqa64 %zmm6, %zmm14 {%k3}15334; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15335; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload15336; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm6 {%k1}15337; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15338; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm11 {%k2}15339; AVX512DQ-NEXT:    vmovdqa64 %zmm6, %zmm11 {%k3}15340; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15341; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload15342; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm6 {%k1}15343; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15344; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm10 {%k2}15345; AVX512DQ-NEXT:    vmovdqa64 %zmm6, %zmm10 {%k3}15346; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15347; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload15348; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm6 {%k1}15349; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15350; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm9 {%k2}15351; AVX512DQ-NEXT:    vmovdqa64 %zmm6, %zmm9 {%k3}15352; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15353; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload15354; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm6 {%k1}15355; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15356; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm7 {%k2}15357; AVX512DQ-NEXT:    vmovdqa64 %zmm6, %zmm7 {%k3}15358; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15359; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload15360; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm6 {%k1}15361; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15362; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm4 {%k2}15363; AVX512DQ-NEXT:    vmovdqa64 %zmm6, %zmm4 {%k3}15364; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15365; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload15366; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm6 {%k1}15367; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15368; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm3 {%k2}15369; AVX512DQ-NEXT:    vmovdqa64 %zmm6, %zmm3 {%k3}15370; AVX512DQ-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15371; AVX512DQ-NEXT:    vmovdqa64 %zmm0, %zmm27 {%k1}15372; AVX512DQ-NEXT:    vmovdqa64 %zmm17, %zmm2 {%k2}15373; AVX512DQ-NEXT:    vmovdqa64 %zmm27, %zmm2 {%k3}15374; AVX512DQ-NEXT:    movq {{[0-9]+}}(%rsp), %rax15375; AVX512DQ-NEXT:    vmovdqa64 %zmm2, 1984(%rax)15376; AVX512DQ-NEXT:    vmovdqa64 %zmm3, 1920(%rax)15377; AVX512DQ-NEXT:    vmovdqa64 %zmm4, 1856(%rax)15378; AVX512DQ-NEXT:    vmovdqa64 %zmm7, 1792(%rax)15379; AVX512DQ-NEXT:    vmovdqa64 %zmm9, 1728(%rax)15380; AVX512DQ-NEXT:    vmovdqa64 %zmm10, 1664(%rax)15381; AVX512DQ-NEXT:    vmovdqa64 %zmm11, 1600(%rax)15382; AVX512DQ-NEXT:    vmovdqa64 %zmm14, 1536(%rax)15383; AVX512DQ-NEXT:    vmovdqa64 %zmm15, 1472(%rax)15384; AVX512DQ-NEXT:    vmovdqa64 %zmm16, 1408(%rax)15385; AVX512DQ-NEXT:    vmovdqa64 %zmm18, 1344(%rax)15386; AVX512DQ-NEXT:    vmovdqa64 %zmm19, 1280(%rax)15387; AVX512DQ-NEXT:    vmovdqa64 %zmm20, 1216(%rax)15388; AVX512DQ-NEXT:    vmovdqa64 %zmm21, 1152(%rax)15389; AVX512DQ-NEXT:    vmovdqa64 %zmm23, 1088(%rax)15390; AVX512DQ-NEXT:    vmovdqa64 %zmm24, 1024(%rax)15391; AVX512DQ-NEXT:    vmovdqa64 %zmm12, 960(%rax)15392; AVX512DQ-NEXT:    vmovdqa64 %zmm28, 896(%rax)15393; AVX512DQ-NEXT:    vmovdqa64 %zmm29, 832(%rax)15394; AVX512DQ-NEXT:    vmovdqa64 %zmm30, 768(%rax)15395; AVX512DQ-NEXT:    vmovdqa64 %zmm31, 704(%rax)15396; AVX512DQ-NEXT:    vmovdqa64 %zmm1, 640(%rax)15397; AVX512DQ-NEXT:    vmovdqa64 %zmm26, 576(%rax)15398; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15399; AVX512DQ-NEXT:    vmovaps %zmm0, 512(%rax)15400; AVX512DQ-NEXT:    vmovdqa64 %zmm5, 448(%rax)15401; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15402; AVX512DQ-NEXT:    vmovaps %zmm0, 384(%rax)15403; AVX512DQ-NEXT:    vmovups (%rsp), %zmm0 # 64-byte Reload15404; AVX512DQ-NEXT:    vmovaps %zmm0, 320(%rax)15405; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15406; AVX512DQ-NEXT:    vmovaps %zmm0, 256(%rax)15407; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15408; AVX512DQ-NEXT:    vmovaps %zmm0, 192(%rax)15409; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15410; AVX512DQ-NEXT:    vmovaps %zmm0, 128(%rax)15411; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15412; AVX512DQ-NEXT:    vmovaps %zmm0, 64(%rax)15413; AVX512DQ-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15414; AVX512DQ-NEXT:    vmovaps %zmm0, (%rax)15415; AVX512DQ-NEXT:    addq $6216, %rsp # imm = 0x184815416; AVX512DQ-NEXT:    vzeroupper15417; AVX512DQ-NEXT:    retq15418;15419; AVX512DQ-FCP-LABEL: store_i32_stride8_vf64:15420; AVX512DQ-FCP:       # %bb.0:15421; AVX512DQ-FCP-NEXT:    subq $6216, %rsp # imm = 0x184815422; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax15423; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %r1015424; AVX512DQ-FCP-NEXT:    vmovdqa64 (%r10), %zmm515425; AVX512DQ-FCP-NEXT:    vmovdqa64 64(%r10), %zmm415426; AVX512DQ-FCP-NEXT:    vmovdqa64 128(%r10), %zmm215427; AVX512DQ-FCP-NEXT:    vmovdqa64 (%rax), %zmm115428; AVX512DQ-FCP-NEXT:    vmovdqa64 64(%rax), %zmm015429; AVX512DQ-FCP-NEXT:    vmovdqa64 128(%rax), %zmm3015430; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm3 = [0,0,0,0,0,0,0,16,0,0,0,0,0,0,1,17]15431; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm5, %zmm715432; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm5, %zmm615433; AVX512DQ-FCP-NEXT:    vpermt2d %zmm1, %zmm3, %zmm715434; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15435; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm5 = [0,0,0,0,0,0,2,18,0,0,0,0,0,0,3,19]15436; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm6, %zmm715437; AVX512DQ-FCP-NEXT:    vpermt2d %zmm1, %zmm5, %zmm615438; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15439; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm6 = [0,0,0,0,0,0,4,20,0,0,0,0,0,0,5,21]15440; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm7, %zmm815441; AVX512DQ-FCP-NEXT:    vpermt2d %zmm1, %zmm6, %zmm815442; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15443; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm8 = [0,0,0,0,0,0,6,22,0,0,0,0,0,0,7,23]15444; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm7, %zmm915445; AVX512DQ-FCP-NEXT:    vpermt2d %zmm1, %zmm8, %zmm915446; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15447; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm8, %zmm915448; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm8 = [0,0,0,0,0,0,8,24,0,0,0,0,0,0,9,25]15449; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm7, %zmm1015450; AVX512DQ-FCP-NEXT:    vpermt2d %zmm1, %zmm8, %zmm1015451; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15452; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm8, %zmm1015453; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm8 = [0,0,0,0,0,0,10,26,0,0,0,0,0,0,11,27]15454; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm7, %zmm1115455; AVX512DQ-FCP-NEXT:    vpermt2d %zmm1, %zmm8, %zmm1115456; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15457; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm8, %zmm1115458; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm8 = [0,0,0,0,0,0,12,28,0,0,0,0,0,0,13,29]15459; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm7, %zmm1315460; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm7, %zmm1215461; AVX512DQ-FCP-NEXT:    vpermt2d %zmm1, %zmm8, %zmm1315462; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15463; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm8, %zmm715464; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm8 = [0,0,0,0,0,0,14,30,0,0,0,0,0,0,15,31]15465; AVX512DQ-FCP-NEXT:    vpermt2d %zmm1, %zmm8, %zmm1215466; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15467; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm4, %zmm115468; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm3, %zmm115469; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15470; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm4, %zmm115471; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm5, %zmm115472; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15473; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm4, %zmm115474; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm6, %zmm115475; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15476; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm4, %zmm115477; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm9, %zmm115478; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15479; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm4, %zmm115480; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm10, %zmm115481; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15482; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm4, %zmm115483; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm11, %zmm115484; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15485; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm4, %zmm115486; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm7, %zmm115487; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15488; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm8, %zmm415489; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15490; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm2, %zmm015491; AVX512DQ-FCP-NEXT:    vpermt2d %zmm30, %zmm3, %zmm015492; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15493; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm2, %zmm015494; AVX512DQ-FCP-NEXT:    vpermt2d %zmm30, %zmm5, %zmm015495; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15496; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm2, %zmm015497; AVX512DQ-FCP-NEXT:    vpermt2d %zmm30, %zmm6, %zmm015498; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15499; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm2, %zmm015500; AVX512DQ-FCP-NEXT:    vpermt2d %zmm30, %zmm9, %zmm015501; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15502; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm2, %zmm015503; AVX512DQ-FCP-NEXT:    vpermt2d %zmm30, %zmm10, %zmm015504; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15505; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm2, %zmm015506; AVX512DQ-FCP-NEXT:    vpermt2d %zmm30, %zmm11, %zmm015507; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15508; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm2, %zmm015509; AVX512DQ-FCP-NEXT:    vpermt2d %zmm30, %zmm7, %zmm015510; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15511; AVX512DQ-FCP-NEXT:    vpermt2d %zmm30, %zmm8, %zmm215512; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15513; AVX512DQ-FCP-NEXT:    vmovdqa64 192(%r10), %zmm115514; AVX512DQ-FCP-NEXT:    vmovdqa64 192(%rax), %zmm015515; AVX512DQ-FCP-NEXT:    vpermi2d %zmm0, %zmm1, %zmm315516; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15517; AVX512DQ-FCP-NEXT:    vpermi2d %zmm0, %zmm1, %zmm515518; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15519; AVX512DQ-FCP-NEXT:    vpermi2d %zmm0, %zmm1, %zmm615520; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15521; AVX512DQ-FCP-NEXT:    vpermi2d %zmm0, %zmm1, %zmm915522; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15523; AVX512DQ-FCP-NEXT:    vpermi2d %zmm0, %zmm1, %zmm1015524; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15525; AVX512DQ-FCP-NEXT:    vpermi2d %zmm0, %zmm1, %zmm1115526; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15527; AVX512DQ-FCP-NEXT:    vpermi2d %zmm0, %zmm1, %zmm715528; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15529; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm8, %zmm115530; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15531; AVX512DQ-FCP-NEXT:    vmovdqa64 (%r8), %zmm2615532; AVX512DQ-FCP-NEXT:    vmovdqa64 (%r9), %zmm015533; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,0,0,0,16,0,0,0,0,0,0,1,17,0,0]15534; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm26, %zmm215535; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm1, %zmm215536; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15537; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, %zmm215538; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,0,0,2,18,0,0,0,0,0,0,3,19,0,0]15539; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm26, %zmm315540; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm1, %zmm315541; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15542; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, %zmm315543; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,0,0,4,20,0,0,0,0,0,0,5,21,0,0]15544; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm26, %zmm415545; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm1, %zmm415546; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15547; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, %zmm415548; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,0,0,6,22,0,0,0,0,0,0,7,23,0,0]15549; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm26, %zmm515550; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm1, %zmm515551; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15552; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, %zmm515553; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,0,0,8,24,0,0,0,0,0,0,9,25,0,0]15554; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm26, %zmm615555; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm1, %zmm615556; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15557; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, %zmm615558; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,0,0,10,26,0,0,0,0,0,0,11,27,0,0]15559; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm26, %zmm715560; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm1, %zmm715561; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15562; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, %zmm715563; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,0,0,12,28,0,0,0,0,0,0,13,29,0,0]15564; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm26, %zmm815565; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm1, %zmm815566; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15567; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, %zmm815568; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,0,0,14,30,0,0,0,0,0,0,15,31,0,0]15569; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm1, %zmm2615570; AVX512DQ-FCP-NEXT:    vmovdqa64 64(%r8), %zmm2515571; AVX512DQ-FCP-NEXT:    vmovdqa64 64(%r9), %zmm015572; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm25, %zmm915573; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm2, %zmm915574; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15575; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm25, %zmm915576; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm3, %zmm915577; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15578; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm25, %zmm915579; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm4, %zmm915580; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15581; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm25, %zmm915582; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm5, %zmm915583; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15584; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm25, %zmm915585; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm6, %zmm915586; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15587; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm25, %zmm915588; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm7, %zmm915589; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15590; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm25, %zmm915591; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm8, %zmm915592; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15593; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm1, %zmm2515594; AVX512DQ-FCP-NEXT:    vmovdqa64 128(%r8), %zmm2215595; AVX512DQ-FCP-NEXT:    vmovdqa64 128(%r9), %zmm015596; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm22, %zmm915597; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm2, %zmm915598; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15599; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm22, %zmm915600; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm3, %zmm915601; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15602; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm22, %zmm915603; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm4, %zmm915604; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15605; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm22, %zmm915606; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm5, %zmm915607; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15608; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm22, %zmm915609; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm6, %zmm915610; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15611; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm22, %zmm915612; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm7, %zmm915613; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15614; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm22, %zmm915615; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm8, %zmm915616; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15617; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm1, %zmm2215618; AVX512DQ-FCP-NEXT:    vmovdqa64 192(%r8), %zmm2715619; AVX512DQ-FCP-NEXT:    vmovdqa64 192(%r9), %zmm015620; AVX512DQ-FCP-NEXT:    vpermi2d %zmm0, %zmm27, %zmm215621; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15622; AVX512DQ-FCP-NEXT:    vpermi2d %zmm0, %zmm27, %zmm315623; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15624; AVX512DQ-FCP-NEXT:    vpermi2d %zmm0, %zmm27, %zmm415625; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15626; AVX512DQ-FCP-NEXT:    vpermi2d %zmm0, %zmm27, %zmm515627; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15628; AVX512DQ-FCP-NEXT:    vpermi2d %zmm0, %zmm27, %zmm615629; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15630; AVX512DQ-FCP-NEXT:    vpermi2d %zmm0, %zmm27, %zmm715631; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15632; AVX512DQ-FCP-NEXT:    vpermi2d %zmm0, %zmm27, %zmm815633; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15634; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm1, %zmm2715635; AVX512DQ-FCP-NEXT:    vmovdqa64 (%rdx), %zmm815636; AVX512DQ-FCP-NEXT:    vmovdqa64 (%rcx), %zmm015637; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,0,16,0,0,0,0,0,0,1,17,0,0,0,0]15638; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm8, %zmm215639; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm1, %zmm215640; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15641; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, %zmm215642; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,2,18,0,0,0,0,0,0,3,19,0,0,0,0]15643; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm8, %zmm315644; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm1, %zmm315645; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15646; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, %zmm315647; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,4,20,0,0,0,0,0,0,5,21,0,0,0,0]15648; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm8, %zmm415649; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm1, %zmm415650; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15651; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, %zmm415652; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,6,22,0,0,0,0,0,0,7,23,0,0,0,0]15653; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm8, %zmm515654; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm1, %zmm515655; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15656; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, %zmm515657; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm7 = [0,0,8,24,0,0,0,0,0,0,9,25,0,0,0,0]15658; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm8, %zmm615659; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm7, %zmm615660; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15661; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm9 = [0,0,10,26,0,0,0,0,0,0,11,27,0,0,0,0]15662; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm8, %zmm615663; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm9, %zmm615664; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15665; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm10 = [0,0,12,28,0,0,0,0,0,0,13,29,0,0,0,0]15666; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm8, %zmm615667; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm10, %zmm615668; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15669; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,14,30,0,0,0,0,0,0,15,31,0,0,0,0]15670; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm1, %zmm815671; AVX512DQ-FCP-NEXT:    vmovdqa64 64(%rdx), %zmm615672; AVX512DQ-FCP-NEXT:    vmovdqa64 64(%rcx), %zmm015673; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm6, %zmm1115674; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm2, %zmm1115675; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15676; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm6, %zmm1115677; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm3, %zmm1115678; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15679; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm6, %zmm1115680; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm4, %zmm1115681; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15682; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm6, %zmm1115683; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm5, %zmm1115684; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15685; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm6, %zmm1115686; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm7, %zmm1115687; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15688; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm6, %zmm1115689; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm9, %zmm1115690; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15691; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm6, %zmm1115692; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm10, %zmm1115693; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15694; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm1, %zmm615695; AVX512DQ-FCP-NEXT:    vmovdqa64 128(%rdx), %zmm1315696; AVX512DQ-FCP-NEXT:    vmovdqa64 128(%rcx), %zmm015697; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm13, %zmm1115698; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm2, %zmm1115699; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15700; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm13, %zmm1115701; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm3, %zmm1115702; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15703; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm13, %zmm1115704; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm4, %zmm1115705; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15706; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm13, %zmm1115707; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm5, %zmm1115708; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15709; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm13, %zmm1115710; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm7, %zmm1115711; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15712; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm13, %zmm1115713; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm9, %zmm1115714; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15715; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm13, %zmm1115716; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm10, %zmm1115717; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15718; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm1, %zmm1315719; AVX512DQ-FCP-NEXT:    vmovdqa64 192(%rdx), %zmm1715720; AVX512DQ-FCP-NEXT:    vmovdqa64 192(%rcx), %zmm015721; AVX512DQ-FCP-NEXT:    vpermi2d %zmm0, %zmm17, %zmm215722; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15723; AVX512DQ-FCP-NEXT:    vpermi2d %zmm0, %zmm17, %zmm315724; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15725; AVX512DQ-FCP-NEXT:    vpermi2d %zmm0, %zmm17, %zmm415726; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15727; AVX512DQ-FCP-NEXT:    vpermi2d %zmm0, %zmm17, %zmm515728; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15729; AVX512DQ-FCP-NEXT:    vpermi2d %zmm0, %zmm17, %zmm715730; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15731; AVX512DQ-FCP-NEXT:    vpermi2d %zmm0, %zmm17, %zmm915732; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15733; AVX512DQ-FCP-NEXT:    vpermi2d %zmm0, %zmm17, %zmm1015734; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15735; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm1, %zmm1715736; AVX512DQ-FCP-NEXT:    vmovdqa64 (%rdi), %zmm515737; AVX512DQ-FCP-NEXT:    vmovdqa64 (%rsi), %zmm015738; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm14 = [0,16,0,0,0,0,0,0,1,17,0,0,0,0,0,0]15739; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm5, %zmm115740; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm14, %zmm115741; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15742; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm11 = [2,18,0,0,0,0,0,0,3,19,0,0,0,0,0,0]15743; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm5, %zmm115744; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm11, %zmm115745; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15746; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm10 = [4,20,0,0,0,0,0,0,5,21,0,0,0,0,0,0]15747; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm5, %zmm115748; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm10, %zmm115749; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15750; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm9 = [6,22,0,0,0,0,0,0,7,23,0,0,0,0,0,0]15751; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm5, %zmm115752; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm9, %zmm115753; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15754; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm7 = [8,24,0,0,0,0,0,0,9,25,0,0,0,0,0,0]15755; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm5, %zmm115756; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm7, %zmm115757; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15758; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [10,26,0,0,0,0,0,0,11,27,0,0,0,0,0,0]15759; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm5, %zmm115760; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm4, %zmm115761; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, (%rsp) # 64-byte Spill15762; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm3 = [12,28,0,0,0,0,0,0,13,29,0,0,0,0,0,0]15763; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm5, %zmm115764; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm3, %zmm115765; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15766; AVX512DQ-FCP-NEXT:    vpmovsxbd {{.*#+}} zmm31 = [14,30,0,0,0,0,0,0,15,31,0,0,0,0,0,0]15767; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm31, %zmm515768; AVX512DQ-FCP-NEXT:    vmovdqa64 64(%rdi), %zmm1215769; AVX512DQ-FCP-NEXT:    vmovdqa64 64(%rsi), %zmm015770; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm12, %zmm115771; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm14, %zmm115772; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15773; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm12, %zmm115774; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm11, %zmm115775; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15776; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm12, %zmm115777; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm10, %zmm115778; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15779; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm12, %zmm115780; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm9, %zmm115781; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15782; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm12, %zmm3015783; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm7, %zmm3015784; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm12, %zmm2915785; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm4, %zmm2915786; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm12, %zmm2815787; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm3, %zmm2815788; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm31, %zmm1215789; AVX512DQ-FCP-NEXT:    vmovdqa64 128(%rdi), %zmm1515790; AVX512DQ-FCP-NEXT:    vmovdqa64 128(%rsi), %zmm015791; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm15, %zmm2415792; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm14, %zmm2415793; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm15, %zmm2315794; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm11, %zmm2315795; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm15, %zmm2115796; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm10, %zmm2115797; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm15, %zmm2015798; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm9, %zmm2015799; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm15, %zmm1915800; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm7, %zmm1915801; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm15, %zmm1815802; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm4, %zmm1815803; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm15, %zmm1615804; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm3, %zmm1615805; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm31, %zmm1515806; AVX512DQ-FCP-NEXT:    vmovdqa64 192(%rdi), %zmm215807; AVX512DQ-FCP-NEXT:    vmovdqa64 192(%rsi), %zmm015808; AVX512DQ-FCP-NEXT:    vpermi2d %zmm0, %zmm2, %zmm1415809; AVX512DQ-FCP-NEXT:    vpermi2d %zmm0, %zmm2, %zmm1115810; AVX512DQ-FCP-NEXT:    vpermi2d %zmm0, %zmm2, %zmm1015811; AVX512DQ-FCP-NEXT:    vpermi2d %zmm0, %zmm2, %zmm915812; AVX512DQ-FCP-NEXT:    vpermi2d %zmm0, %zmm2, %zmm715813; AVX512DQ-FCP-NEXT:    vpermi2d %zmm0, %zmm2, %zmm415814; AVX512DQ-FCP-NEXT:    vpermi2d %zmm0, %zmm2, %zmm315815; AVX512DQ-FCP-NEXT:    vpermt2d %zmm0, %zmm31, %zmm215816; AVX512DQ-FCP-NEXT:    movb $-120, %al15817; AVX512DQ-FCP-NEXT:    kmovw %eax, %k115818; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15819; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload15820; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}15821; AVX512DQ-FCP-NEXT:    movb $34, %al15822; AVX512DQ-FCP-NEXT:    kmovw %eax, %k215823; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15824; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload15825; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm31, %zmm0 {%k2}15826; AVX512DQ-FCP-NEXT:    movb $-52, %al15827; AVX512DQ-FCP-NEXT:    kmovw %eax, %k315828; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, %zmm0 {%k3}15829; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15830; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15831; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload15832; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}15833; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15834; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload15835; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm31, %zmm0 {%k2}15836; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, %zmm0 {%k3}15837; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15838; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15839; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload15840; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}15841; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15842; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload15843; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm31, %zmm0 {%k2}15844; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, %zmm0 {%k3}15845; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15846; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15847; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload15848; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}15849; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15850; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload15851; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm31, %zmm0 {%k2}15852; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, %zmm0 {%k3}15853; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15854; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15855; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload15856; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}15857; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15858; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload15859; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm31, %zmm0 {%k2}15860; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, %zmm0 {%k3}15861; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15862; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15863; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload15864; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}15865; AVX512DQ-FCP-NEXT:    vmovdqu64 (%rsp), %zmm0 # 64-byte Reload15866; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload15867; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm31, %zmm0 {%k2}15868; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, %zmm0 {%k3}15869; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, (%rsp) # 64-byte Spill15870; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15871; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload15872; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}15873; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15874; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload15875; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm31, %zmm0 {%k2}15876; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, %zmm0 {%k3}15877; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15878; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15879; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm26 {%k1}15880; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm8, %zmm5 {%k2}15881; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm26, %zmm5 {%k3}15882; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15883; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload15884; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}15885; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15886; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload15887; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm8, %zmm0 {%k2}15888; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, %zmm0 {%k3}15889; AVX512DQ-FCP-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill15890; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15891; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload15892; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm8 {%k1}15893; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload15894; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15895; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm26 {%k2}15896; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm8, %zmm26 {%k3}15897; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15898; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload15899; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm8 {%k1}15900; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload15901; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15902; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k2}15903; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm8, %zmm1 {%k3}15904; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15905; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload15906; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm8 {%k1}15907; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15908; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload15909; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm31 {%k2}15910; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm8, %zmm31 {%k3}15911; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15912; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload15913; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm8 {%k1}15914; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15915; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm30 {%k2}15916; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm8, %zmm30 {%k3}15917; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15918; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload15919; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm8 {%k1}15920; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15921; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm29 {%k2}15922; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm8, %zmm29 {%k3}15923; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15924; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload15925; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm8 {%k1}15926; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15927; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm28 {%k2}15928; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm8, %zmm28 {%k3}15929; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15930; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm25 {%k1}15931; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm6, %zmm12 {%k2}15932; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm25, %zmm12 {%k3}15933; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15934; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload15935; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm6 {%k1}15936; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15937; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm24 {%k2}15938; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm6, %zmm24 {%k3}15939; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15940; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload15941; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm6 {%k1}15942; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15943; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm23 {%k2}15944; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm6, %zmm23 {%k3}15945; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15946; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload15947; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm6 {%k1}15948; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15949; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm21 {%k2}15950; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm6, %zmm21 {%k3}15951; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15952; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload15953; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm6 {%k1}15954; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15955; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm20 {%k2}15956; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm6, %zmm20 {%k3}15957; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15958; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload15959; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm6 {%k1}15960; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15961; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm19 {%k2}15962; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm6, %zmm19 {%k3}15963; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15964; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload15965; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm6 {%k1}15966; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15967; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm18 {%k2}15968; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm6, %zmm18 {%k3}15969; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15970; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload15971; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm6 {%k1}15972; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15973; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm16 {%k2}15974; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm6, %zmm16 {%k3}15975; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15976; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm22 {%k1}15977; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm13, %zmm15 {%k2}15978; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm22, %zmm15 {%k3}15979; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15980; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload15981; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm6 {%k1}15982; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15983; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm14 {%k2}15984; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm6, %zmm14 {%k3}15985; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15986; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload15987; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm6 {%k1}15988; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15989; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm11 {%k2}15990; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm6, %zmm11 {%k3}15991; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15992; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload15993; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm6 {%k1}15994; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15995; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm10 {%k2}15996; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm6, %zmm10 {%k3}15997; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload15998; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload15999; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm6 {%k1}16000; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16001; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm9 {%k2}16002; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm6, %zmm9 {%k3}16003; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16004; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload16005; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm6 {%k1}16006; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16007; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm7 {%k2}16008; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm6, %zmm7 {%k3}16009; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16010; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload16011; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm6 {%k1}16012; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16013; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm4 {%k2}16014; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm6, %zmm4 {%k3}16015; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16016; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm6 # 64-byte Reload16017; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm6 {%k1}16018; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16019; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm3 {%k2}16020; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm6, %zmm3 {%k3}16021; AVX512DQ-FCP-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16022; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm0, %zmm27 {%k1}16023; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm17, %zmm2 {%k2}16024; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm27, %zmm2 {%k3}16025; AVX512DQ-FCP-NEXT:    movq {{[0-9]+}}(%rsp), %rax16026; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm2, 1984(%rax)16027; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm3, 1920(%rax)16028; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm4, 1856(%rax)16029; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm7, 1792(%rax)16030; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm9, 1728(%rax)16031; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm10, 1664(%rax)16032; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm11, 1600(%rax)16033; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm14, 1536(%rax)16034; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm15, 1472(%rax)16035; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm16, 1408(%rax)16036; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm18, 1344(%rax)16037; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm19, 1280(%rax)16038; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm20, 1216(%rax)16039; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm21, 1152(%rax)16040; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm23, 1088(%rax)16041; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm24, 1024(%rax)16042; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm12, 960(%rax)16043; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm28, 896(%rax)16044; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm29, 832(%rax)16045; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm30, 768(%rax)16046; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm31, 704(%rax)16047; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm1, 640(%rax)16048; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm26, 576(%rax)16049; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16050; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, 512(%rax)16051; AVX512DQ-FCP-NEXT:    vmovdqa64 %zmm5, 448(%rax)16052; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16053; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, 384(%rax)16054; AVX512DQ-FCP-NEXT:    vmovups (%rsp), %zmm0 # 64-byte Reload16055; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, 320(%rax)16056; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16057; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, 256(%rax)16058; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16059; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, 192(%rax)16060; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16061; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, 128(%rax)16062; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16063; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, 64(%rax)16064; AVX512DQ-FCP-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16065; AVX512DQ-FCP-NEXT:    vmovaps %zmm0, (%rax)16066; AVX512DQ-FCP-NEXT:    addq $6216, %rsp # imm = 0x184816067; AVX512DQ-FCP-NEXT:    vzeroupper16068; AVX512DQ-FCP-NEXT:    retq16069;16070; AVX512BW-LABEL: store_i32_stride8_vf64:16071; AVX512BW:       # %bb.0:16072; AVX512BW-NEXT:    subq $6216, %rsp # imm = 0x184816073; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %rax16074; AVX512BW-NEXT:    movq {{[0-9]+}}(%rsp), %r1016075; AVX512BW-NEXT:    vmovdqa64 (%r10), %zmm516076; AVX512BW-NEXT:    vmovdqa64 64(%r10), %zmm416077; AVX512BW-NEXT:    vmovdqa64 128(%r10), %zmm216078; AVX512BW-NEXT:    vmovdqa64 (%rax), %zmm116079; AVX512BW-NEXT:    vmovdqa64 64(%rax), %zmm016080; AVX512BW-NEXT:    vmovdqa64 128(%rax), %zmm3016081; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm3 = [0,0,0,0,0,0,0,16,0,0,0,0,0,0,1,17]16082; AVX512BW-NEXT:    vmovdqa64 %zmm5, %zmm716083; AVX512BW-NEXT:    vmovdqa64 %zmm5, %zmm616084; AVX512BW-NEXT:    vpermt2d %zmm1, %zmm3, %zmm716085; AVX512BW-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16086; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm5 = [0,0,0,0,0,0,2,18,0,0,0,0,0,0,3,19]16087; AVX512BW-NEXT:    vmovdqa64 %zmm6, %zmm716088; AVX512BW-NEXT:    vpermt2d %zmm1, %zmm5, %zmm616089; AVX512BW-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16090; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm6 = [0,0,0,0,0,0,4,20,0,0,0,0,0,0,5,21]16091; AVX512BW-NEXT:    vmovdqa64 %zmm7, %zmm816092; AVX512BW-NEXT:    vpermt2d %zmm1, %zmm6, %zmm816093; AVX512BW-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16094; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm8 = [0,0,0,0,0,0,6,22,0,0,0,0,0,0,7,23]16095; AVX512BW-NEXT:    vmovdqa64 %zmm7, %zmm916096; AVX512BW-NEXT:    vpermt2d %zmm1, %zmm8, %zmm916097; AVX512BW-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16098; AVX512BW-NEXT:    vmovdqa64 %zmm8, %zmm916099; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm8 = [0,0,0,0,0,0,8,24,0,0,0,0,0,0,9,25]16100; AVX512BW-NEXT:    vmovdqa64 %zmm7, %zmm1016101; AVX512BW-NEXT:    vpermt2d %zmm1, %zmm8, %zmm1016102; AVX512BW-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16103; AVX512BW-NEXT:    vmovdqa64 %zmm8, %zmm1016104; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm8 = [0,0,0,0,0,0,10,26,0,0,0,0,0,0,11,27]16105; AVX512BW-NEXT:    vmovdqa64 %zmm7, %zmm1116106; AVX512BW-NEXT:    vpermt2d %zmm1, %zmm8, %zmm1116107; AVX512BW-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16108; AVX512BW-NEXT:    vmovdqa64 %zmm8, %zmm1116109; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm8 = [0,0,0,0,0,0,12,28,0,0,0,0,0,0,13,29]16110; AVX512BW-NEXT:    vmovdqa64 %zmm7, %zmm1316111; AVX512BW-NEXT:    vmovdqa64 %zmm7, %zmm1216112; AVX512BW-NEXT:    vpermt2d %zmm1, %zmm8, %zmm1316113; AVX512BW-NEXT:    vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16114; AVX512BW-NEXT:    vmovdqa64 %zmm8, %zmm716115; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm8 = [0,0,0,0,0,0,14,30,0,0,0,0,0,0,15,31]16116; AVX512BW-NEXT:    vpermt2d %zmm1, %zmm8, %zmm1216117; AVX512BW-NEXT:    vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16118; AVX512BW-NEXT:    vmovdqa64 %zmm4, %zmm116119; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm3, %zmm116120; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16121; AVX512BW-NEXT:    vmovdqa64 %zmm4, %zmm116122; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm5, %zmm116123; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16124; AVX512BW-NEXT:    vmovdqa64 %zmm4, %zmm116125; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm6, %zmm116126; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16127; AVX512BW-NEXT:    vmovdqa64 %zmm4, %zmm116128; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm9, %zmm116129; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16130; AVX512BW-NEXT:    vmovdqa64 %zmm4, %zmm116131; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm10, %zmm116132; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16133; AVX512BW-NEXT:    vmovdqa64 %zmm4, %zmm116134; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm11, %zmm116135; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16136; AVX512BW-NEXT:    vmovdqa64 %zmm4, %zmm116137; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm7, %zmm116138; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16139; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm8, %zmm416140; AVX512BW-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16141; AVX512BW-NEXT:    vmovdqa64 %zmm2, %zmm016142; AVX512BW-NEXT:    vpermt2d %zmm30, %zmm3, %zmm016143; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16144; AVX512BW-NEXT:    vmovdqa64 %zmm2, %zmm016145; AVX512BW-NEXT:    vpermt2d %zmm30, %zmm5, %zmm016146; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16147; AVX512BW-NEXT:    vmovdqa64 %zmm2, %zmm016148; AVX512BW-NEXT:    vpermt2d %zmm30, %zmm6, %zmm016149; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16150; AVX512BW-NEXT:    vmovdqa64 %zmm2, %zmm016151; AVX512BW-NEXT:    vpermt2d %zmm30, %zmm9, %zmm016152; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16153; AVX512BW-NEXT:    vmovdqa64 %zmm2, %zmm016154; AVX512BW-NEXT:    vpermt2d %zmm30, %zmm10, %zmm016155; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16156; AVX512BW-NEXT:    vmovdqa64 %zmm2, %zmm016157; AVX512BW-NEXT:    vpermt2d %zmm30, %zmm11, %zmm016158; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16159; AVX512BW-NEXT:    vmovdqa64 %zmm2, %zmm016160; AVX512BW-NEXT:    vpermt2d %zmm30, %zmm7, %zmm016161; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16162; AVX512BW-NEXT:    vpermt2d %zmm30, %zmm8, %zmm216163; AVX512BW-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16164; AVX512BW-NEXT:    vmovdqa64 192(%r10), %zmm116165; AVX512BW-NEXT:    vmovdqa64 192(%rax), %zmm016166; AVX512BW-NEXT:    vpermi2d %zmm0, %zmm1, %zmm316167; AVX512BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16168; AVX512BW-NEXT:    vpermi2d %zmm0, %zmm1, %zmm516169; AVX512BW-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16170; AVX512BW-NEXT:    vpermi2d %zmm0, %zmm1, %zmm616171; AVX512BW-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16172; AVX512BW-NEXT:    vpermi2d %zmm0, %zmm1, %zmm916173; AVX512BW-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16174; AVX512BW-NEXT:    vpermi2d %zmm0, %zmm1, %zmm1016175; AVX512BW-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16176; AVX512BW-NEXT:    vpermi2d %zmm0, %zmm1, %zmm1116177; AVX512BW-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16178; AVX512BW-NEXT:    vpermi2d %zmm0, %zmm1, %zmm716179; AVX512BW-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16180; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm8, %zmm116181; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16182; AVX512BW-NEXT:    vmovdqa64 (%r8), %zmm2616183; AVX512BW-NEXT:    vmovdqa64 (%r9), %zmm016184; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,0,0,0,16,0,0,0,0,0,0,1,17,0,0]16185; AVX512BW-NEXT:    vmovdqa64 %zmm26, %zmm216186; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm1, %zmm216187; AVX512BW-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16188; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm216189; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,0,0,2,18,0,0,0,0,0,0,3,19,0,0]16190; AVX512BW-NEXT:    vmovdqa64 %zmm26, %zmm316191; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm1, %zmm316192; AVX512BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16193; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm316194; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,0,0,4,20,0,0,0,0,0,0,5,21,0,0]16195; AVX512BW-NEXT:    vmovdqa64 %zmm26, %zmm416196; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm1, %zmm416197; AVX512BW-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16198; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm416199; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,0,0,6,22,0,0,0,0,0,0,7,23,0,0]16200; AVX512BW-NEXT:    vmovdqa64 %zmm26, %zmm516201; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm1, %zmm516202; AVX512BW-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16203; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm516204; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,0,0,8,24,0,0,0,0,0,0,9,25,0,0]16205; AVX512BW-NEXT:    vmovdqa64 %zmm26, %zmm616206; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm1, %zmm616207; AVX512BW-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16208; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm616209; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,0,0,10,26,0,0,0,0,0,0,11,27,0,0]16210; AVX512BW-NEXT:    vmovdqa64 %zmm26, %zmm716211; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm1, %zmm716212; AVX512BW-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16213; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm716214; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,0,0,12,28,0,0,0,0,0,0,13,29,0,0]16215; AVX512BW-NEXT:    vmovdqa64 %zmm26, %zmm816216; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm1, %zmm816217; AVX512BW-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16218; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm816219; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,0,0,14,30,0,0,0,0,0,0,15,31,0,0]16220; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm1, %zmm2616221; AVX512BW-NEXT:    vmovdqa64 64(%r8), %zmm2516222; AVX512BW-NEXT:    vmovdqa64 64(%r9), %zmm016223; AVX512BW-NEXT:    vmovdqa64 %zmm25, %zmm916224; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm2, %zmm916225; AVX512BW-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16226; AVX512BW-NEXT:    vmovdqa64 %zmm25, %zmm916227; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm3, %zmm916228; AVX512BW-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16229; AVX512BW-NEXT:    vmovdqa64 %zmm25, %zmm916230; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm4, %zmm916231; AVX512BW-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16232; AVX512BW-NEXT:    vmovdqa64 %zmm25, %zmm916233; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm5, %zmm916234; AVX512BW-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16235; AVX512BW-NEXT:    vmovdqa64 %zmm25, %zmm916236; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm6, %zmm916237; AVX512BW-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16238; AVX512BW-NEXT:    vmovdqa64 %zmm25, %zmm916239; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm7, %zmm916240; AVX512BW-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16241; AVX512BW-NEXT:    vmovdqa64 %zmm25, %zmm916242; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm8, %zmm916243; AVX512BW-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16244; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm1, %zmm2516245; AVX512BW-NEXT:    vmovdqa64 128(%r8), %zmm2216246; AVX512BW-NEXT:    vmovdqa64 128(%r9), %zmm016247; AVX512BW-NEXT:    vmovdqa64 %zmm22, %zmm916248; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm2, %zmm916249; AVX512BW-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16250; AVX512BW-NEXT:    vmovdqa64 %zmm22, %zmm916251; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm3, %zmm916252; AVX512BW-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16253; AVX512BW-NEXT:    vmovdqa64 %zmm22, %zmm916254; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm4, %zmm916255; AVX512BW-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16256; AVX512BW-NEXT:    vmovdqa64 %zmm22, %zmm916257; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm5, %zmm916258; AVX512BW-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16259; AVX512BW-NEXT:    vmovdqa64 %zmm22, %zmm916260; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm6, %zmm916261; AVX512BW-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16262; AVX512BW-NEXT:    vmovdqa64 %zmm22, %zmm916263; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm7, %zmm916264; AVX512BW-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16265; AVX512BW-NEXT:    vmovdqa64 %zmm22, %zmm916266; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm8, %zmm916267; AVX512BW-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16268; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm1, %zmm2216269; AVX512BW-NEXT:    vmovdqa64 192(%r8), %zmm2716270; AVX512BW-NEXT:    vmovdqa64 192(%r9), %zmm016271; AVX512BW-NEXT:    vpermi2d %zmm0, %zmm27, %zmm216272; AVX512BW-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16273; AVX512BW-NEXT:    vpermi2d %zmm0, %zmm27, %zmm316274; AVX512BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16275; AVX512BW-NEXT:    vpermi2d %zmm0, %zmm27, %zmm416276; AVX512BW-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16277; AVX512BW-NEXT:    vpermi2d %zmm0, %zmm27, %zmm516278; AVX512BW-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16279; AVX512BW-NEXT:    vpermi2d %zmm0, %zmm27, %zmm616280; AVX512BW-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16281; AVX512BW-NEXT:    vpermi2d %zmm0, %zmm27, %zmm716282; AVX512BW-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16283; AVX512BW-NEXT:    vpermi2d %zmm0, %zmm27, %zmm816284; AVX512BW-NEXT:    vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16285; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm1, %zmm2716286; AVX512BW-NEXT:    vmovdqa64 (%rdx), %zmm816287; AVX512BW-NEXT:    vmovdqa64 (%rcx), %zmm016288; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,0,16,0,0,0,0,0,0,1,17,0,0,0,0]16289; AVX512BW-NEXT:    vmovdqa64 %zmm8, %zmm216290; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm1, %zmm216291; AVX512BW-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16292; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm216293; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,2,18,0,0,0,0,0,0,3,19,0,0,0,0]16294; AVX512BW-NEXT:    vmovdqa64 %zmm8, %zmm316295; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm1, %zmm316296; AVX512BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16297; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm316298; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,4,20,0,0,0,0,0,0,5,21,0,0,0,0]16299; AVX512BW-NEXT:    vmovdqa64 %zmm8, %zmm416300; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm1, %zmm416301; AVX512BW-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16302; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm416303; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,6,22,0,0,0,0,0,0,7,23,0,0,0,0]16304; AVX512BW-NEXT:    vmovdqa64 %zmm8, %zmm516305; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm1, %zmm516306; AVX512BW-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16307; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm516308; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm7 = [0,0,8,24,0,0,0,0,0,0,9,25,0,0,0,0]16309; AVX512BW-NEXT:    vmovdqa64 %zmm8, %zmm616310; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm7, %zmm616311; AVX512BW-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16312; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm9 = [0,0,10,26,0,0,0,0,0,0,11,27,0,0,0,0]16313; AVX512BW-NEXT:    vmovdqa64 %zmm8, %zmm616314; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm9, %zmm616315; AVX512BW-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16316; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm10 = [0,0,12,28,0,0,0,0,0,0,13,29,0,0,0,0]16317; AVX512BW-NEXT:    vmovdqa64 %zmm8, %zmm616318; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm10, %zmm616319; AVX512BW-NEXT:    vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16320; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm1 = [0,0,14,30,0,0,0,0,0,0,15,31,0,0,0,0]16321; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm1, %zmm816322; AVX512BW-NEXT:    vmovdqa64 64(%rdx), %zmm616323; AVX512BW-NEXT:    vmovdqa64 64(%rcx), %zmm016324; AVX512BW-NEXT:    vmovdqa64 %zmm6, %zmm1116325; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm2, %zmm1116326; AVX512BW-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16327; AVX512BW-NEXT:    vmovdqa64 %zmm6, %zmm1116328; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm3, %zmm1116329; AVX512BW-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16330; AVX512BW-NEXT:    vmovdqa64 %zmm6, %zmm1116331; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm4, %zmm1116332; AVX512BW-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16333; AVX512BW-NEXT:    vmovdqa64 %zmm6, %zmm1116334; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm5, %zmm1116335; AVX512BW-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16336; AVX512BW-NEXT:    vmovdqa64 %zmm6, %zmm1116337; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm7, %zmm1116338; AVX512BW-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16339; AVX512BW-NEXT:    vmovdqa64 %zmm6, %zmm1116340; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm9, %zmm1116341; AVX512BW-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16342; AVX512BW-NEXT:    vmovdqa64 %zmm6, %zmm1116343; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm10, %zmm1116344; AVX512BW-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16345; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm1, %zmm616346; AVX512BW-NEXT:    vmovdqa64 128(%rdx), %zmm1316347; AVX512BW-NEXT:    vmovdqa64 128(%rcx), %zmm016348; AVX512BW-NEXT:    vmovdqa64 %zmm13, %zmm1116349; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm2, %zmm1116350; AVX512BW-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16351; AVX512BW-NEXT:    vmovdqa64 %zmm13, %zmm1116352; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm3, %zmm1116353; AVX512BW-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16354; AVX512BW-NEXT:    vmovdqa64 %zmm13, %zmm1116355; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm4, %zmm1116356; AVX512BW-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16357; AVX512BW-NEXT:    vmovdqa64 %zmm13, %zmm1116358; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm5, %zmm1116359; AVX512BW-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16360; AVX512BW-NEXT:    vmovdqa64 %zmm13, %zmm1116361; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm7, %zmm1116362; AVX512BW-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16363; AVX512BW-NEXT:    vmovdqa64 %zmm13, %zmm1116364; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm9, %zmm1116365; AVX512BW-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16366; AVX512BW-NEXT:    vmovdqa64 %zmm13, %zmm1116367; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm10, %zmm1116368; AVX512BW-NEXT:    vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16369; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm1, %zmm1316370; AVX512BW-NEXT:    vmovdqa64 192(%rdx), %zmm1716371; AVX512BW-NEXT:    vmovdqa64 192(%rcx), %zmm016372; AVX512BW-NEXT:    vpermi2d %zmm0, %zmm17, %zmm216373; AVX512BW-NEXT:    vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16374; AVX512BW-NEXT:    vpermi2d %zmm0, %zmm17, %zmm316375; AVX512BW-NEXT:    vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16376; AVX512BW-NEXT:    vpermi2d %zmm0, %zmm17, %zmm416377; AVX512BW-NEXT:    vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16378; AVX512BW-NEXT:    vpermi2d %zmm0, %zmm17, %zmm516379; AVX512BW-NEXT:    vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16380; AVX512BW-NEXT:    vpermi2d %zmm0, %zmm17, %zmm716381; AVX512BW-NEXT:    vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16382; AVX512BW-NEXT:    vpermi2d %zmm0, %zmm17, %zmm916383; AVX512BW-NEXT:    vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16384; AVX512BW-NEXT:    vpermi2d %zmm0, %zmm17, %zmm1016385; AVX512BW-NEXT:    vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16386; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm1, %zmm1716387; AVX512BW-NEXT:    vmovdqa64 (%rdi), %zmm516388; AVX512BW-NEXT:    vmovdqa64 (%rsi), %zmm016389; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm14 = [0,16,0,0,0,0,0,0,1,17,0,0,0,0,0,0]16390; AVX512BW-NEXT:    vmovdqa64 %zmm5, %zmm116391; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm14, %zmm116392; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16393; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm11 = [2,18,0,0,0,0,0,0,3,19,0,0,0,0,0,0]16394; AVX512BW-NEXT:    vmovdqa64 %zmm5, %zmm116395; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm11, %zmm116396; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16397; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm10 = [4,20,0,0,0,0,0,0,5,21,0,0,0,0,0,0]16398; AVX512BW-NEXT:    vmovdqa64 %zmm5, %zmm116399; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm10, %zmm116400; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16401; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm9 = [6,22,0,0,0,0,0,0,7,23,0,0,0,0,0,0]16402; AVX512BW-NEXT:    vmovdqa64 %zmm5, %zmm116403; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm9, %zmm116404; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16405; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm7 = [8,24,0,0,0,0,0,0,9,25,0,0,0,0,0,0]16406; AVX512BW-NEXT:    vmovdqa64 %zmm5, %zmm116407; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm7, %zmm116408; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16409; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm4 = [10,26,0,0,0,0,0,0,11,27,0,0,0,0,0,0]16410; AVX512BW-NEXT:    vmovdqa64 %zmm5, %zmm116411; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm4, %zmm116412; AVX512BW-NEXT:    vmovdqu64 %zmm1, (%rsp) # 64-byte Spill16413; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm3 = [12,28,0,0,0,0,0,0,13,29,0,0,0,0,0,0]16414; AVX512BW-NEXT:    vmovdqa64 %zmm5, %zmm116415; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm3, %zmm116416; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16417; AVX512BW-NEXT:    vpmovsxbd {{.*#+}} zmm31 = [14,30,0,0,0,0,0,0,15,31,0,0,0,0,0,0]16418; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm31, %zmm516419; AVX512BW-NEXT:    vmovdqa64 64(%rdi), %zmm1216420; AVX512BW-NEXT:    vmovdqa64 64(%rsi), %zmm016421; AVX512BW-NEXT:    vmovdqa64 %zmm12, %zmm116422; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm14, %zmm116423; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16424; AVX512BW-NEXT:    vmovdqa64 %zmm12, %zmm116425; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm11, %zmm116426; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16427; AVX512BW-NEXT:    vmovdqa64 %zmm12, %zmm116428; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm10, %zmm116429; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16430; AVX512BW-NEXT:    vmovdqa64 %zmm12, %zmm116431; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm9, %zmm116432; AVX512BW-NEXT:    vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16433; AVX512BW-NEXT:    vmovdqa64 %zmm12, %zmm3016434; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm7, %zmm3016435; AVX512BW-NEXT:    vmovdqa64 %zmm12, %zmm2916436; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm4, %zmm2916437; AVX512BW-NEXT:    vmovdqa64 %zmm12, %zmm2816438; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm3, %zmm2816439; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm31, %zmm1216440; AVX512BW-NEXT:    vmovdqa64 128(%rdi), %zmm1516441; AVX512BW-NEXT:    vmovdqa64 128(%rsi), %zmm016442; AVX512BW-NEXT:    vmovdqa64 %zmm15, %zmm2416443; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm14, %zmm2416444; AVX512BW-NEXT:    vmovdqa64 %zmm15, %zmm2316445; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm11, %zmm2316446; AVX512BW-NEXT:    vmovdqa64 %zmm15, %zmm2116447; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm10, %zmm2116448; AVX512BW-NEXT:    vmovdqa64 %zmm15, %zmm2016449; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm9, %zmm2016450; AVX512BW-NEXT:    vmovdqa64 %zmm15, %zmm1916451; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm7, %zmm1916452; AVX512BW-NEXT:    vmovdqa64 %zmm15, %zmm1816453; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm4, %zmm1816454; AVX512BW-NEXT:    vmovdqa64 %zmm15, %zmm1616455; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm3, %zmm1616456; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm31, %zmm1516457; AVX512BW-NEXT:    vmovdqa64 192(%rdi), %zmm216458; AVX512BW-NEXT:    vmovdqa64 192(%rsi), %zmm016459; AVX512BW-NEXT:    vpermi2d %zmm0, %zmm2, %zmm1416460; AVX512BW-NEXT:    vpermi2d %zmm0, %zmm2, %zmm1116461; AVX512BW-NEXT:    vpermi2d %zmm0, %zmm2, %zmm1016462; AVX512BW-NEXT:    vpermi2d %zmm0, %zmm2, %zmm916463; AVX512BW-NEXT:    vpermi2d %zmm0, %zmm2, %zmm716464; AVX512BW-NEXT:    vpermi2d %zmm0, %zmm2, %zmm416465; AVX512BW-NEXT:    vpermi2d %zmm0, %zmm2, %zmm316466; AVX512BW-NEXT:    vpermt2d %zmm0, %zmm31, %zmm216467; AVX512BW-NEXT:    movb $-120, %al16468; AVX512BW-NEXT:    kmovd %eax, %k116469; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16470; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload16471; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}16472; AVX512BW-NEXT:    movb $34, %al16473; AVX512BW-NEXT:    kmovd %eax, %k216474; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16475; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm31 # 64-byte Reload16476; AVX512BW-NEXT:    vmovdqa64 %zmm31, %zmm0 {%k2}16477; AVX512BW-NEXT:    movb $-52, %al16478; AVX512BW-NEXT:    kmovd %eax, %k316479; AVX512BW-NEXT:    vmovdqa64 %zmm1, %zmm0 {%k3}16480; AVX512BW-NEXT:    vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill16481; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16482; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload16483; AVX512BW-NEXT:    vmovdqa64 %zmm0, %zmm1 {%k1}16484; AVX512BW-NEXT:    vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload16485; AVX512BW-NEXT:    vm

File truncated to the first 1 MiB.