9626 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_stride5_vf2(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecptr2, ptr %in.vecptr3, ptr %in.vecptr4, ptr %out.vec) nounwind {19; SSE-LABEL: store_i32_stride5_vf2:20; SSE: # %bb.0:21; SSE-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero22; SSE-NEXT: movsd {{.*#+}} xmm1 = mem[0],zero23; SSE-NEXT: movsd {{.*#+}} xmm2 = mem[0],zero24; SSE-NEXT: movsd {{.*#+}} xmm3 = mem[0],zero25; SSE-NEXT: movsd {{.*#+}} xmm4 = mem[0],zero26; SSE-NEXT: movaps %xmm2, %xmm527; SSE-NEXT: unpcklps {{.*#+}} xmm5 = xmm5[0],xmm3[0],xmm5[1],xmm3[1]28; SSE-NEXT: movaps %xmm0, %xmm629; SSE-NEXT: unpcklps {{.*#+}} xmm6 = xmm6[0],xmm1[0],xmm6[1],xmm1[1]30; SSE-NEXT: movlhps {{.*#+}} xmm6 = xmm6[0],xmm5[0]31; SSE-NEXT: shufps {{.*#+}} xmm2 = xmm2[1,1],xmm1[1,1]32; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[0,1],xmm2[2,0]33; SSE-NEXT: movss {{.*#+}} xmm0 = xmm4[0],xmm0[1,2,3]34; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm4[1,1,1,1]35; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm3[1,1,1,1]36; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]37; SSE-NEXT: movq %xmm2, 32(%r9)38; SSE-NEXT: movaps %xmm6, (%r9)39; SSE-NEXT: movaps %xmm0, 16(%r9)40; SSE-NEXT: retq41;42; AVX-LABEL: store_i32_stride5_vf2:43; AVX: # %bb.0:44; AVX-NEXT: vmovsd {{.*#+}} xmm0 = mem[0],zero45; AVX-NEXT: vmovsd {{.*#+}} xmm1 = mem[0],zero46; AVX-NEXT: vmovlhps {{.*#+}} xmm2 = xmm1[0],xmm0[0]47; AVX-NEXT: vmovsd {{.*#+}} xmm3 = mem[0],zero48; AVX-NEXT: vmovsd {{.*#+}} xmm4 = mem[0],zero49; AVX-NEXT: vmovlhps {{.*#+}} xmm3 = xmm4[0],xmm3[0]50; AVX-NEXT: vmovsd {{.*#+}} xmm5 = mem[0],zero51; AVX-NEXT: vshufps {{.*#+}} xmm2 = xmm2[0,2],xmm3[0,2]52; AVX-NEXT: vshufps {{.*#+}} xmm0 = xmm4[1,1],xmm0[1,1]53; AVX-NEXT: vshufps {{.*#+}} xmm0 = xmm1[0,1],xmm0[2,0]54; AVX-NEXT: vmovss {{.*#+}} xmm0 = xmm5[0],xmm0[1,2,3]55; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm2, %ymm056; AVX-NEXT: vshufps {{.*#+}} xmm1 = xmm3[3,3,3,3]57; AVX-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0],xmm5[1],xmm1[2,3]58; AVX-NEXT: vmovlps %xmm1, 32(%r9)59; AVX-NEXT: vmovaps %ymm0, (%r9)60; AVX-NEXT: vzeroupper61; AVX-NEXT: retq62;63; AVX2-LABEL: store_i32_stride5_vf2:64; AVX2: # %bb.0:65; AVX2-NEXT: movq (%r8), %rax66; AVX2-NEXT: vmovq {{.*#+}} xmm0 = mem[0],zero67; AVX2-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero68; AVX2-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0]69; AVX2-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero70; AVX2-NEXT: vmovq {{.*#+}} xmm2 = mem[0],zero71; AVX2-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm2[0],xmm1[0]72; AVX2-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm073; AVX2-NEXT: vmovq %rax, %xmm274; AVX2-NEXT: vpmovsxbd {{.*#+}} ymm3 = [0,2,4,6,0,1,3,5]75; AVX2-NEXT: vpermd %ymm0, %ymm3, %ymm076; AVX2-NEXT: vmovd %eax, %xmm377; AVX2-NEXT: vpbroadcastd %xmm3, %ymm378; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm3[4],ymm0[5,6,7]79; AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[1,1,1,1]80; AVX2-NEXT: vpblendd {{.*#+}} xmm1 = xmm1[0],xmm2[1],xmm1[2,3]81; AVX2-NEXT: vmovq %xmm1, 32(%r9)82; AVX2-NEXT: vmovdqa %ymm0, (%r9)83; AVX2-NEXT: vzeroupper84; AVX2-NEXT: retq85;86; AVX2-FP-LABEL: store_i32_stride5_vf2:87; AVX2-FP: # %bb.0:88; AVX2-FP-NEXT: movq (%r8), %rax89; AVX2-FP-NEXT: vmovq {{.*#+}} xmm0 = mem[0],zero90; AVX2-FP-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero91; AVX2-FP-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0]92; AVX2-FP-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero93; AVX2-FP-NEXT: vmovq {{.*#+}} xmm2 = mem[0],zero94; AVX2-FP-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm2[0],xmm1[0]95; AVX2-FP-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm096; AVX2-FP-NEXT: vmovq %rax, %xmm297; AVX2-FP-NEXT: vpmovsxbd {{.*#+}} ymm3 = [0,2,4,6,0,1,3,5]98; AVX2-FP-NEXT: vpermd %ymm0, %ymm3, %ymm099; AVX2-FP-NEXT: vmovd %eax, %xmm3100; AVX2-FP-NEXT: vpbroadcastd %xmm3, %ymm3101; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm3[4],ymm0[5,6,7]102; AVX2-FP-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[1,1,1,1]103; AVX2-FP-NEXT: vpblendd {{.*#+}} xmm1 = xmm1[0],xmm2[1],xmm1[2,3]104; AVX2-FP-NEXT: vmovq %xmm1, 32(%r9)105; AVX2-FP-NEXT: vmovdqa %ymm0, (%r9)106; AVX2-FP-NEXT: vzeroupper107; AVX2-FP-NEXT: retq108;109; AVX2-FCP-LABEL: store_i32_stride5_vf2:110; AVX2-FCP: # %bb.0:111; AVX2-FCP-NEXT: movq (%r8), %rax112; AVX2-FCP-NEXT: vmovq {{.*#+}} xmm0 = mem[0],zero113; AVX2-FCP-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero114; AVX2-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0]115; AVX2-FCP-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero116; AVX2-FCP-NEXT: vmovq {{.*#+}} xmm2 = mem[0],zero117; AVX2-FCP-NEXT: vpunpcklqdq {{.*#+}} xmm2 = xmm2[0],xmm1[0]118; AVX2-FCP-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm0119; AVX2-FCP-NEXT: vmovq %rax, %xmm2120; AVX2-FCP-NEXT: vpmovsxbd {{.*#+}} ymm3 = [0,2,4,6,0,1,3,5]121; AVX2-FCP-NEXT: vpermd %ymm0, %ymm3, %ymm0122; AVX2-FCP-NEXT: vmovd %eax, %xmm3123; AVX2-FCP-NEXT: vpbroadcastd %xmm3, %ymm3124; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm3[4],ymm0[5,6,7]125; AVX2-FCP-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[1,1,1,1]126; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm1 = xmm1[0],xmm2[1],xmm1[2,3]127; AVX2-FCP-NEXT: vmovq %xmm1, 32(%r9)128; AVX2-FCP-NEXT: vmovdqa %ymm0, (%r9)129; AVX2-FCP-NEXT: vzeroupper130; AVX2-FCP-NEXT: retq131;132; AVX512-LABEL: store_i32_stride5_vf2:133; AVX512: # %bb.0:134; AVX512-NEXT: vmovsd {{.*#+}} xmm0 = mem[0],zero135; AVX512-NEXT: vmovsd {{.*#+}} xmm1 = mem[0],zero136; AVX512-NEXT: vmovlhps {{.*#+}} xmm0 = xmm1[0],xmm0[0]137; AVX512-NEXT: vmovsd {{.*#+}} xmm1 = mem[0],zero138; AVX512-NEXT: vmovsd {{.*#+}} xmm2 = mem[0],zero139; AVX512-NEXT: vmovlhps {{.*#+}} xmm1 = xmm2[0],xmm1[0]140; AVX512-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0141; AVX512-NEXT: vmovsd {{.*#+}} xmm1 = mem[0],zero142; AVX512-NEXT: vinsertf32x4 $2, %xmm1, %zmm0, %zmm0143; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm1 = [0,2,4,6,8,1,3,5,7,9,0,0,0,0,0,0]144; AVX512-NEXT: vpermps %zmm0, %zmm1, %zmm0145; AVX512-NEXT: vextractf32x4 $2, %zmm0, %xmm1146; AVX512-NEXT: vmovlps %xmm1, 32(%r9)147; AVX512-NEXT: vmovaps %ymm0, (%r9)148; AVX512-NEXT: vzeroupper149; AVX512-NEXT: retq150;151; AVX512-FCP-LABEL: store_i32_stride5_vf2:152; AVX512-FCP: # %bb.0:153; AVX512-FCP-NEXT: vmovsd {{.*#+}} xmm0 = mem[0],zero154; AVX512-FCP-NEXT: vmovsd {{.*#+}} xmm1 = mem[0],zero155; AVX512-FCP-NEXT: vmovlhps {{.*#+}} xmm0 = xmm1[0],xmm0[0]156; AVX512-FCP-NEXT: vmovsd {{.*#+}} xmm1 = mem[0],zero157; AVX512-FCP-NEXT: vmovsd {{.*#+}} xmm2 = mem[0],zero158; AVX512-FCP-NEXT: vmovlhps {{.*#+}} xmm1 = xmm2[0],xmm1[0]159; AVX512-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0160; AVX512-FCP-NEXT: vmovsd {{.*#+}} xmm1 = mem[0],zero161; AVX512-FCP-NEXT: vinsertf32x4 $2, %xmm1, %zmm0, %zmm0162; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm1 = [0,2,4,6,8,1,3,5,7,9,0,0,0,0,0,0]163; AVX512-FCP-NEXT: vpermps %zmm0, %zmm1, %zmm0164; AVX512-FCP-NEXT: vextractf32x4 $2, %zmm0, %xmm1165; AVX512-FCP-NEXT: vmovlps %xmm1, 32(%r9)166; AVX512-FCP-NEXT: vmovaps %ymm0, (%r9)167; AVX512-FCP-NEXT: vzeroupper168; AVX512-FCP-NEXT: retq169;170; AVX512DQ-LABEL: store_i32_stride5_vf2:171; AVX512DQ: # %bb.0:172; AVX512DQ-NEXT: vmovsd {{.*#+}} xmm0 = mem[0],zero173; AVX512DQ-NEXT: vmovsd {{.*#+}} xmm1 = mem[0],zero174; AVX512DQ-NEXT: vmovlhps {{.*#+}} xmm0 = xmm1[0],xmm0[0]175; AVX512DQ-NEXT: vmovsd {{.*#+}} xmm1 = mem[0],zero176; AVX512DQ-NEXT: vmovsd {{.*#+}} xmm2 = mem[0],zero177; AVX512DQ-NEXT: vmovlhps {{.*#+}} xmm1 = xmm2[0],xmm1[0]178; AVX512DQ-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0179; AVX512DQ-NEXT: vmovsd {{.*#+}} xmm1 = mem[0],zero180; AVX512DQ-NEXT: vinsertf32x4 $2, %xmm1, %zmm0, %zmm0181; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm1 = [0,2,4,6,8,1,3,5,7,9,0,0,0,0,0,0]182; AVX512DQ-NEXT: vpermps %zmm0, %zmm1, %zmm0183; AVX512DQ-NEXT: vextractf32x4 $2, %zmm0, %xmm1184; AVX512DQ-NEXT: vmovlps %xmm1, 32(%r9)185; AVX512DQ-NEXT: vmovaps %ymm0, (%r9)186; AVX512DQ-NEXT: vzeroupper187; AVX512DQ-NEXT: retq188;189; AVX512DQ-FCP-LABEL: store_i32_stride5_vf2:190; AVX512DQ-FCP: # %bb.0:191; AVX512DQ-FCP-NEXT: vmovsd {{.*#+}} xmm0 = mem[0],zero192; AVX512DQ-FCP-NEXT: vmovsd {{.*#+}} xmm1 = mem[0],zero193; AVX512DQ-FCP-NEXT: vmovlhps {{.*#+}} xmm0 = xmm1[0],xmm0[0]194; AVX512DQ-FCP-NEXT: vmovsd {{.*#+}} xmm1 = mem[0],zero195; AVX512DQ-FCP-NEXT: vmovsd {{.*#+}} xmm2 = mem[0],zero196; AVX512DQ-FCP-NEXT: vmovlhps {{.*#+}} xmm1 = xmm2[0],xmm1[0]197; AVX512DQ-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0198; AVX512DQ-FCP-NEXT: vmovsd {{.*#+}} xmm1 = mem[0],zero199; AVX512DQ-FCP-NEXT: vinsertf32x4 $2, %xmm1, %zmm0, %zmm0200; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm1 = [0,2,4,6,8,1,3,5,7,9,0,0,0,0,0,0]201; AVX512DQ-FCP-NEXT: vpermps %zmm0, %zmm1, %zmm0202; AVX512DQ-FCP-NEXT: vextractf32x4 $2, %zmm0, %xmm1203; AVX512DQ-FCP-NEXT: vmovlps %xmm1, 32(%r9)204; AVX512DQ-FCP-NEXT: vmovaps %ymm0, (%r9)205; AVX512DQ-FCP-NEXT: vzeroupper206; AVX512DQ-FCP-NEXT: retq207;208; AVX512BW-LABEL: store_i32_stride5_vf2:209; AVX512BW: # %bb.0:210; AVX512BW-NEXT: vmovsd {{.*#+}} xmm0 = mem[0],zero211; AVX512BW-NEXT: vmovsd {{.*#+}} xmm1 = mem[0],zero212; AVX512BW-NEXT: vmovlhps {{.*#+}} xmm0 = xmm1[0],xmm0[0]213; AVX512BW-NEXT: vmovsd {{.*#+}} xmm1 = mem[0],zero214; AVX512BW-NEXT: vmovsd {{.*#+}} xmm2 = mem[0],zero215; AVX512BW-NEXT: vmovlhps {{.*#+}} xmm1 = xmm2[0],xmm1[0]216; AVX512BW-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0217; AVX512BW-NEXT: vmovsd {{.*#+}} xmm1 = mem[0],zero218; AVX512BW-NEXT: vinsertf32x4 $2, %xmm1, %zmm0, %zmm0219; AVX512BW-NEXT: vpmovsxbd {{.*#+}} zmm1 = [0,2,4,6,8,1,3,5,7,9,0,0,0,0,0,0]220; AVX512BW-NEXT: vpermps %zmm0, %zmm1, %zmm0221; AVX512BW-NEXT: vextractf32x4 $2, %zmm0, %xmm1222; AVX512BW-NEXT: vmovlps %xmm1, 32(%r9)223; AVX512BW-NEXT: vmovaps %ymm0, (%r9)224; AVX512BW-NEXT: vzeroupper225; AVX512BW-NEXT: retq226;227; AVX512BW-FCP-LABEL: store_i32_stride5_vf2:228; AVX512BW-FCP: # %bb.0:229; AVX512BW-FCP-NEXT: vmovsd {{.*#+}} xmm0 = mem[0],zero230; AVX512BW-FCP-NEXT: vmovsd {{.*#+}} xmm1 = mem[0],zero231; AVX512BW-FCP-NEXT: vmovlhps {{.*#+}} xmm0 = xmm1[0],xmm0[0]232; AVX512BW-FCP-NEXT: vmovsd {{.*#+}} xmm1 = mem[0],zero233; AVX512BW-FCP-NEXT: vmovsd {{.*#+}} xmm2 = mem[0],zero234; AVX512BW-FCP-NEXT: vmovlhps {{.*#+}} xmm1 = xmm2[0],xmm1[0]235; AVX512BW-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0236; AVX512BW-FCP-NEXT: vmovsd {{.*#+}} xmm1 = mem[0],zero237; AVX512BW-FCP-NEXT: vinsertf32x4 $2, %xmm1, %zmm0, %zmm0238; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} zmm1 = [0,2,4,6,8,1,3,5,7,9,0,0,0,0,0,0]239; AVX512BW-FCP-NEXT: vpermps %zmm0, %zmm1, %zmm0240; AVX512BW-FCP-NEXT: vextractf32x4 $2, %zmm0, %xmm1241; AVX512BW-FCP-NEXT: vmovlps %xmm1, 32(%r9)242; AVX512BW-FCP-NEXT: vmovaps %ymm0, (%r9)243; AVX512BW-FCP-NEXT: vzeroupper244; AVX512BW-FCP-NEXT: retq245;246; AVX512DQ-BW-LABEL: store_i32_stride5_vf2:247; AVX512DQ-BW: # %bb.0:248; AVX512DQ-BW-NEXT: vmovsd {{.*#+}} xmm0 = mem[0],zero249; AVX512DQ-BW-NEXT: vmovsd {{.*#+}} xmm1 = mem[0],zero250; AVX512DQ-BW-NEXT: vmovlhps {{.*#+}} xmm0 = xmm1[0],xmm0[0]251; AVX512DQ-BW-NEXT: vmovsd {{.*#+}} xmm1 = mem[0],zero252; AVX512DQ-BW-NEXT: vmovsd {{.*#+}} xmm2 = mem[0],zero253; AVX512DQ-BW-NEXT: vmovlhps {{.*#+}} xmm1 = xmm2[0],xmm1[0]254; AVX512DQ-BW-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0255; AVX512DQ-BW-NEXT: vmovsd {{.*#+}} xmm1 = mem[0],zero256; AVX512DQ-BW-NEXT: vinsertf32x4 $2, %xmm1, %zmm0, %zmm0257; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} zmm1 = [0,2,4,6,8,1,3,5,7,9,0,0,0,0,0,0]258; AVX512DQ-BW-NEXT: vpermps %zmm0, %zmm1, %zmm0259; AVX512DQ-BW-NEXT: vextractf32x4 $2, %zmm0, %xmm1260; AVX512DQ-BW-NEXT: vmovlps %xmm1, 32(%r9)261; AVX512DQ-BW-NEXT: vmovaps %ymm0, (%r9)262; AVX512DQ-BW-NEXT: vzeroupper263; AVX512DQ-BW-NEXT: retq264;265; AVX512DQ-BW-FCP-LABEL: store_i32_stride5_vf2:266; AVX512DQ-BW-FCP: # %bb.0:267; AVX512DQ-BW-FCP-NEXT: vmovsd {{.*#+}} xmm0 = mem[0],zero268; AVX512DQ-BW-FCP-NEXT: vmovsd {{.*#+}} xmm1 = mem[0],zero269; AVX512DQ-BW-FCP-NEXT: vmovlhps {{.*#+}} xmm0 = xmm1[0],xmm0[0]270; AVX512DQ-BW-FCP-NEXT: vmovsd {{.*#+}} xmm1 = mem[0],zero271; AVX512DQ-BW-FCP-NEXT: vmovsd {{.*#+}} xmm2 = mem[0],zero272; AVX512DQ-BW-FCP-NEXT: vmovlhps {{.*#+}} xmm1 = xmm2[0],xmm1[0]273; AVX512DQ-BW-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0274; AVX512DQ-BW-FCP-NEXT: vmovsd {{.*#+}} xmm1 = mem[0],zero275; AVX512DQ-BW-FCP-NEXT: vinsertf32x4 $2, %xmm1, %zmm0, %zmm0276; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} zmm1 = [0,2,4,6,8,1,3,5,7,9,0,0,0,0,0,0]277; AVX512DQ-BW-FCP-NEXT: vpermps %zmm0, %zmm1, %zmm0278; AVX512DQ-BW-FCP-NEXT: vextractf32x4 $2, %zmm0, %xmm1279; AVX512DQ-BW-FCP-NEXT: vmovlps %xmm1, 32(%r9)280; AVX512DQ-BW-FCP-NEXT: vmovaps %ymm0, (%r9)281; AVX512DQ-BW-FCP-NEXT: vzeroupper282; AVX512DQ-BW-FCP-NEXT: retq283 %in.vec0 = load <2 x i32>, ptr %in.vecptr0, align 64284 %in.vec1 = load <2 x i32>, ptr %in.vecptr1, align 64285 %in.vec2 = load <2 x i32>, ptr %in.vecptr2, align 64286 %in.vec3 = load <2 x i32>, ptr %in.vecptr3, align 64287 %in.vec4 = load <2 x i32>, ptr %in.vecptr4, align 64288 %1 = shufflevector <2 x i32> %in.vec0, <2 x i32> %in.vec1, <4 x i32> <i32 0, i32 1, i32 2, i32 3>289 %2 = shufflevector <2 x i32> %in.vec2, <2 x i32> %in.vec3, <4 x i32> <i32 0, i32 1, i32 2, i32 3>290 %3 = 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>291 %4 = shufflevector <2 x i32> %in.vec4, <2 x i32> poison, <8 x i32> <i32 0, i32 1, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>292 %5 = shufflevector <8 x i32> %3, <8 x i32> %4, <10 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9>293 %interleaved.vec = shufflevector <10 x i32> %5, <10 x i32> poison, <10 x i32> <i32 0, i32 2, i32 4, i32 6, i32 8, i32 1, i32 3, i32 5, i32 7, i32 9>294 store <10 x i32> %interleaved.vec, ptr %out.vec, align 64295 ret void296}297 298define void @store_i32_stride5_vf4(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecptr2, ptr %in.vecptr3, ptr %in.vecptr4, ptr %out.vec) nounwind {299; SSE-LABEL: store_i32_stride5_vf4:300; SSE: # %bb.0:301; SSE-NEXT: movaps (%rdi), %xmm0302; SSE-NEXT: movaps (%rsi), %xmm1303; SSE-NEXT: movaps (%rdx), %xmm2304; SSE-NEXT: movaps (%rcx), %xmm3305; SSE-NEXT: movaps (%r8), %xmm4306; SSE-NEXT: movaps %xmm2, %xmm5307; SSE-NEXT: unpcklps {{.*#+}} xmm5 = xmm5[0],xmm3[0],xmm5[1],xmm3[1]308; SSE-NEXT: movaps %xmm0, %xmm6309; SSE-NEXT: unpcklps {{.*#+}} xmm6 = xmm6[0],xmm1[0],xmm6[1],xmm1[1]310; SSE-NEXT: movlhps {{.*#+}} xmm6 = xmm6[0],xmm5[0]311; SSE-NEXT: movaps %xmm3, %xmm5312; SSE-NEXT: shufps {{.*#+}} xmm5 = xmm5[3,3],xmm4[3,3]313; SSE-NEXT: movaps %xmm1, %xmm7314; SSE-NEXT: unpckhps {{.*#+}} xmm7 = xmm7[2],xmm2[2],xmm7[3],xmm2[3]315; SSE-NEXT: shufps {{.*#+}} xmm7 = xmm7[2,3],xmm5[0,2]316; SSE-NEXT: movaps %xmm0, %xmm5317; SSE-NEXT: shufps {{.*#+}} xmm5 = xmm5[3,1],xmm4[2,3]318; SSE-NEXT: movaps %xmm3, %xmm8319; SSE-NEXT: unpckhpd {{.*#+}} xmm8 = xmm8[1],xmm2[1]320; SSE-NEXT: shufps {{.*#+}} xmm8 = xmm8[2,0],xmm5[2,0]321; SSE-NEXT: movaps %xmm0, %xmm5322; SSE-NEXT: unpckhps {{.*#+}} xmm5 = xmm5[2],xmm1[2],xmm5[3],xmm1[3]323; SSE-NEXT: shufps {{.*#+}} xmm3 = xmm3[1,1],xmm4[1,1]324; SSE-NEXT: shufps {{.*#+}} xmm3 = xmm3[0,2],xmm5[0,1]325; SSE-NEXT: shufps {{.*#+}} xmm2 = xmm2[1,1],xmm1[1,1]326; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[0,1],xmm2[2,0]327; SSE-NEXT: movss {{.*#+}} xmm0 = xmm4[0],xmm0[1,2,3]328; SSE-NEXT: movaps %xmm3, 32(%r9)329; SSE-NEXT: movaps %xmm8, 48(%r9)330; SSE-NEXT: movaps %xmm7, 64(%r9)331; SSE-NEXT: movaps %xmm6, (%r9)332; SSE-NEXT: movaps %xmm0, 16(%r9)333; SSE-NEXT: retq334;335; AVX-LABEL: store_i32_stride5_vf4:336; AVX: # %bb.0:337; AVX-NEXT: vmovaps (%rdi), %xmm0338; AVX-NEXT: vmovaps (%rsi), %xmm1339; AVX-NEXT: vmovaps (%rdx), %xmm2340; AVX-NEXT: vmovaps (%rcx), %xmm3341; AVX-NEXT: vunpckhps {{.*#+}} xmm4 = xmm2[2],xmm3[2],xmm2[3],xmm3[3]342; AVX-NEXT: vblendps {{.*#+}} xmm4 = xmm4[0,1],xmm0[2,3]343; AVX-NEXT: vshufps {{.*#+}} xmm5 = xmm1[2,2,2,2]344; AVX-NEXT: vblendps {{.*#+}} xmm5 = xmm0[0,1,2],xmm5[3]345; AVX-NEXT: vinsertps {{.*#+}} xmm5 = xmm3[1],zero,xmm5[2,3]346; AVX-NEXT: vinsertf128 $1, %xmm4, %ymm5, %ymm4347; AVX-NEXT: vbroadcastf128 {{.*#+}} ymm5 = mem[0,1,0,1]348; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0],ymm5[1],ymm4[2,3,4,5],ymm5[6],ymm4[7]349; AVX-NEXT: vunpcklps {{.*#+}} xmm6 = xmm2[0],xmm3[0],xmm2[1],xmm3[1]350; AVX-NEXT: vunpcklps {{.*#+}} xmm7 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]351; AVX-NEXT: vmovlhps {{.*#+}} xmm6 = xmm7[0],xmm6[0]352; AVX-NEXT: vinsertps {{.*#+}} xmm7 = zero,xmm0[1],xmm1[1],zero353; AVX-NEXT: vinsertps {{.*#+}} xmm7 = xmm7[0,1,2],xmm2[1]354; AVX-NEXT: vinsertf128 $1, %xmm7, %ymm6, %ymm6355; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm5[4,5,6,7]356; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm6[1,2,3],ymm0[4],ymm6[5,6,7]357; AVX-NEXT: vshufps {{.*#+}} xmm2 = xmm2[3,3],xmm3[3,3]358; AVX-NEXT: vinsertps {{.*#+}} xmm1 = xmm1[3],xmm2[1,2,3]359; AVX-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1,2],xmm5[3]360; AVX-NEXT: vmovaps %xmm1, 64(%r9)361; AVX-NEXT: vmovaps %ymm0, (%r9)362; AVX-NEXT: vmovaps %ymm4, 32(%r9)363; AVX-NEXT: vzeroupper364; AVX-NEXT: retq365;366; AVX2-LABEL: store_i32_stride5_vf4:367; AVX2: # %bb.0:368; AVX2-NEXT: vmovaps (%rdi), %xmm0369; AVX2-NEXT: vmovaps (%rsi), %xmm1370; AVX2-NEXT: vmovaps (%rdx), %xmm2371; AVX2-NEXT: vmovaps (%rcx), %xmm3372; AVX2-NEXT: vmovaps (%r8), %xmm4373; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm5374; AVX2-NEXT: vinsertf128 $1, %xmm3, %ymm2, %ymm6375; AVX2-NEXT: vmovaps {{.*#+}} ymm7 = [u,u,0,4,u,u,u,1]376; AVX2-NEXT: vpermps %ymm6, %ymm7, %ymm7377; AVX2-NEXT: vmovaps {{.*#+}} ymm8 = [u,4,u,u,u,1,5,u]378; AVX2-NEXT: vpermps %ymm5, %ymm8, %ymm8379; AVX2-NEXT: vblendps {{.*#+}} ymm7 = ymm8[0,1],ymm7[2,3],ymm8[4,5,6],ymm7[7]380; AVX2-NEXT: vinsertf128 $1, %xmm4, %ymm0, %ymm0381; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm7[1,2,3],ymm0[4],ymm7[5,6,7]382; AVX2-NEXT: vmovaps {{.*#+}} ymm7 = [u,u,2,6,u,u,u,3]383; AVX2-NEXT: vpermps %ymm5, %ymm7, %ymm5384; AVX2-NEXT: vmovaps {{.*#+}} ymm7 = [5,u,u,u,2,6,u,u]385; AVX2-NEXT: vpermps %ymm6, %ymm7, %ymm6386; AVX2-NEXT: vblendps {{.*#+}} ymm5 = ymm6[0,1],ymm5[2,3],ymm6[4,5],ymm5[6,7]387; AVX2-NEXT: vpermpd {{.*#+}} ymm6 = ymm4[0,1,2,1]388; AVX2-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0],ymm6[1],ymm5[2,3,4,5],ymm6[6],ymm5[7]389; AVX2-NEXT: vshufps {{.*#+}} xmm2 = xmm2[3,3],xmm3[3,3]390; AVX2-NEXT: vinsertps {{.*#+}} xmm1 = xmm1[3],xmm2[1,2],zero391; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1,2],xmm4[3]392; AVX2-NEXT: vmovaps %xmm1, 64(%r9)393; AVX2-NEXT: vmovaps %ymm5, 32(%r9)394; AVX2-NEXT: vmovaps %ymm0, (%r9)395; AVX2-NEXT: vzeroupper396; AVX2-NEXT: retq397;398; AVX2-FP-LABEL: store_i32_stride5_vf4:399; AVX2-FP: # %bb.0:400; AVX2-FP-NEXT: vmovaps (%rdi), %xmm0401; AVX2-FP-NEXT: vmovaps (%rsi), %xmm1402; AVX2-FP-NEXT: vmovaps (%rdx), %xmm2403; AVX2-FP-NEXT: vmovaps (%rcx), %xmm3404; AVX2-FP-NEXT: vmovaps (%r8), %xmm4405; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm5406; AVX2-FP-NEXT: vinsertf128 $1, %xmm3, %ymm2, %ymm6407; AVX2-FP-NEXT: vmovaps {{.*#+}} ymm7 = [u,u,0,4,u,u,u,1]408; AVX2-FP-NEXT: vpermps %ymm6, %ymm7, %ymm7409; AVX2-FP-NEXT: vmovaps {{.*#+}} ymm8 = [u,4,u,u,u,1,5,u]410; AVX2-FP-NEXT: vpermps %ymm5, %ymm8, %ymm8411; AVX2-FP-NEXT: vblendps {{.*#+}} ymm7 = ymm8[0,1],ymm7[2,3],ymm8[4,5,6],ymm7[7]412; AVX2-FP-NEXT: vinsertf128 $1, %xmm4, %ymm0, %ymm0413; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm7[1,2,3],ymm0[4],ymm7[5,6,7]414; AVX2-FP-NEXT: vmovaps {{.*#+}} ymm7 = [u,u,2,6,u,u,u,3]415; AVX2-FP-NEXT: vpermps %ymm5, %ymm7, %ymm5416; AVX2-FP-NEXT: vmovaps {{.*#+}} ymm7 = [5,u,u,u,2,6,u,u]417; AVX2-FP-NEXT: vpermps %ymm6, %ymm7, %ymm6418; AVX2-FP-NEXT: vblendps {{.*#+}} ymm5 = ymm6[0,1],ymm5[2,3],ymm6[4,5],ymm5[6,7]419; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm6 = ymm4[0,1,2,1]420; AVX2-FP-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0],ymm6[1],ymm5[2,3,4,5],ymm6[6],ymm5[7]421; AVX2-FP-NEXT: vshufps {{.*#+}} xmm2 = xmm2[3,3],xmm3[3,3]422; AVX2-FP-NEXT: vinsertps {{.*#+}} xmm1 = xmm1[3],xmm2[1,2],zero423; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1,2],xmm4[3]424; AVX2-FP-NEXT: vmovaps %xmm1, 64(%r9)425; AVX2-FP-NEXT: vmovaps %ymm5, 32(%r9)426; AVX2-FP-NEXT: vmovaps %ymm0, (%r9)427; AVX2-FP-NEXT: vzeroupper428; AVX2-FP-NEXT: retq429;430; AVX2-FCP-LABEL: store_i32_stride5_vf4:431; AVX2-FCP: # %bb.0:432; AVX2-FCP-NEXT: vmovaps (%rdi), %xmm0433; AVX2-FCP-NEXT: vmovaps (%rsi), %xmm1434; AVX2-FCP-NEXT: vmovaps (%rdx), %xmm2435; AVX2-FCP-NEXT: vmovaps (%r8), %xmm3436; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm4437; AVX2-FCP-NEXT: vinsertf128 $1, (%rcx), %ymm2, %ymm2438; AVX2-FCP-NEXT: vmovaps {{.*#+}} ymm5 = [u,u,0,4,u,u,u,1]439; AVX2-FCP-NEXT: vpermps %ymm2, %ymm5, %ymm5440; AVX2-FCP-NEXT: vmovaps {{.*#+}} ymm6 = [u,4,u,u,u,1,5,u]441; AVX2-FCP-NEXT: vpermps %ymm4, %ymm6, %ymm6442; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm5 = ymm6[0,1],ymm5[2,3],ymm6[4,5,6],ymm5[7]443; AVX2-FCP-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm0444; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm5[1,2,3],ymm0[4],ymm5[5,6,7]445; AVX2-FCP-NEXT: vmovaps {{.*#+}} ymm5 = [u,u,2,6,u,u,u,3]446; AVX2-FCP-NEXT: vpermps %ymm4, %ymm5, %ymm4447; AVX2-FCP-NEXT: vmovaps {{.*#+}} ymm5 = [5,u,u,u,2,6,u,u]448; AVX2-FCP-NEXT: vpermps %ymm2, %ymm5, %ymm5449; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm5[0,1],ymm4[2,3],ymm5[4,5],ymm4[6,7]450; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm5 = ymm3[0,1,2,1]451; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0],ymm5[1],ymm4[2,3,4,5],ymm5[6],ymm4[7]452; AVX2-FCP-NEXT: vbroadcastsd {{.*#+}} ymm5 = [7,3,7,3,7,3,7,3]453; AVX2-FCP-NEXT: vpermps %ymm2, %ymm5, %ymm2454; AVX2-FCP-NEXT: vshufps {{.*#+}} xmm1 = xmm1[3,3,3,3]455; AVX2-FCP-NEXT: vmovss {{.*#+}} xmm1 = xmm1[0],xmm2[1,2,3]456; AVX2-FCP-NEXT: vmovaps %ymm4, 32(%r9)457; AVX2-FCP-NEXT: vmovaps %ymm0, (%r9)458; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm0 = xmm1[0,1,2],xmm3[3]459; AVX2-FCP-NEXT: vmovaps %xmm0, 64(%r9)460; AVX2-FCP-NEXT: vzeroupper461; AVX2-FCP-NEXT: retq462;463; AVX512-LABEL: store_i32_stride5_vf4:464; AVX512: # %bb.0:465; AVX512-NEXT: vmovdqa (%rdi), %xmm0466; AVX512-NEXT: vmovdqa (%rdx), %xmm1467; AVX512-NEXT: vmovdqa (%r8), %xmm2468; AVX512-NEXT: vinserti128 $1, (%rcx), %ymm1, %ymm1469; AVX512-NEXT: vinserti128 $1, (%rsi), %ymm0, %ymm0470; AVX512-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0471; AVX512-NEXT: vpmovsxbd {{.*#+}} xmm1 = [7,11,15,19]472; AVX512-NEXT: vpermi2d %zmm2, %zmm0, %zmm1473; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm3 = [0,4,8,12,16,1,5,9,13,17,2,6,10,14,18,3]474; AVX512-NEXT: vpermi2d %zmm2, %zmm0, %zmm3475; AVX512-NEXT: vmovdqa64 %zmm3, (%r9)476; AVX512-NEXT: vmovdqa %xmm1, 64(%r9)477; AVX512-NEXT: vzeroupper478; AVX512-NEXT: retq479;480; AVX512-FCP-LABEL: store_i32_stride5_vf4:481; AVX512-FCP: # %bb.0:482; AVX512-FCP-NEXT: vmovdqa (%rdi), %xmm0483; AVX512-FCP-NEXT: vmovdqa (%rdx), %xmm1484; AVX512-FCP-NEXT: vmovdqa (%r8), %xmm2485; AVX512-FCP-NEXT: vinserti128 $1, (%rcx), %ymm1, %ymm1486; AVX512-FCP-NEXT: vinserti128 $1, (%rsi), %ymm0, %ymm0487; AVX512-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0488; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} xmm1 = [7,11,15,19]489; AVX512-FCP-NEXT: vpermi2d %zmm2, %zmm0, %zmm1490; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm3 = [0,4,8,12,16,1,5,9,13,17,2,6,10,14,18,3]491; AVX512-FCP-NEXT: vpermi2d %zmm2, %zmm0, %zmm3492; AVX512-FCP-NEXT: vmovdqa64 %zmm3, (%r9)493; AVX512-FCP-NEXT: vmovdqa %xmm1, 64(%r9)494; AVX512-FCP-NEXT: vzeroupper495; AVX512-FCP-NEXT: retq496;497; AVX512DQ-LABEL: store_i32_stride5_vf4:498; AVX512DQ: # %bb.0:499; AVX512DQ-NEXT: vmovdqa (%rdi), %xmm0500; AVX512DQ-NEXT: vmovdqa (%rdx), %xmm1501; AVX512DQ-NEXT: vmovdqa (%r8), %xmm2502; AVX512DQ-NEXT: vinserti128 $1, (%rcx), %ymm1, %ymm1503; AVX512DQ-NEXT: vinserti128 $1, (%rsi), %ymm0, %ymm0504; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0505; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} xmm1 = [7,11,15,19]506; AVX512DQ-NEXT: vpermi2d %zmm2, %zmm0, %zmm1507; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm3 = [0,4,8,12,16,1,5,9,13,17,2,6,10,14,18,3]508; AVX512DQ-NEXT: vpermi2d %zmm2, %zmm0, %zmm3509; AVX512DQ-NEXT: vmovdqa64 %zmm3, (%r9)510; AVX512DQ-NEXT: vmovdqa %xmm1, 64(%r9)511; AVX512DQ-NEXT: vzeroupper512; AVX512DQ-NEXT: retq513;514; AVX512DQ-FCP-LABEL: store_i32_stride5_vf4:515; AVX512DQ-FCP: # %bb.0:516; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %xmm0517; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %xmm1518; AVX512DQ-FCP-NEXT: vmovdqa (%r8), %xmm2519; AVX512DQ-FCP-NEXT: vinserti128 $1, (%rcx), %ymm1, %ymm1520; AVX512DQ-FCP-NEXT: vinserti128 $1, (%rsi), %ymm0, %ymm0521; AVX512DQ-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0522; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} xmm1 = [7,11,15,19]523; AVX512DQ-FCP-NEXT: vpermi2d %zmm2, %zmm0, %zmm1524; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm3 = [0,4,8,12,16,1,5,9,13,17,2,6,10,14,18,3]525; AVX512DQ-FCP-NEXT: vpermi2d %zmm2, %zmm0, %zmm3526; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, (%r9)527; AVX512DQ-FCP-NEXT: vmovdqa %xmm1, 64(%r9)528; AVX512DQ-FCP-NEXT: vzeroupper529; AVX512DQ-FCP-NEXT: retq530;531; AVX512BW-LABEL: store_i32_stride5_vf4:532; AVX512BW: # %bb.0:533; AVX512BW-NEXT: vmovdqa (%rdi), %xmm0534; AVX512BW-NEXT: vmovdqa (%rdx), %xmm1535; AVX512BW-NEXT: vmovdqa (%r8), %xmm2536; AVX512BW-NEXT: vinserti128 $1, (%rcx), %ymm1, %ymm1537; AVX512BW-NEXT: vinserti128 $1, (%rsi), %ymm0, %ymm0538; AVX512BW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0539; AVX512BW-NEXT: vpmovsxbd {{.*#+}} xmm1 = [7,11,15,19]540; AVX512BW-NEXT: vpermi2d %zmm2, %zmm0, %zmm1541; AVX512BW-NEXT: vpmovsxbd {{.*#+}} zmm3 = [0,4,8,12,16,1,5,9,13,17,2,6,10,14,18,3]542; AVX512BW-NEXT: vpermi2d %zmm2, %zmm0, %zmm3543; AVX512BW-NEXT: vmovdqa64 %zmm3, (%r9)544; AVX512BW-NEXT: vmovdqa %xmm1, 64(%r9)545; AVX512BW-NEXT: vzeroupper546; AVX512BW-NEXT: retq547;548; AVX512BW-FCP-LABEL: store_i32_stride5_vf4:549; AVX512BW-FCP: # %bb.0:550; AVX512BW-FCP-NEXT: vmovdqa (%rdi), %xmm0551; AVX512BW-FCP-NEXT: vmovdqa (%rdx), %xmm1552; AVX512BW-FCP-NEXT: vmovdqa (%r8), %xmm2553; AVX512BW-FCP-NEXT: vinserti128 $1, (%rcx), %ymm1, %ymm1554; AVX512BW-FCP-NEXT: vinserti128 $1, (%rsi), %ymm0, %ymm0555; AVX512BW-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0556; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} xmm1 = [7,11,15,19]557; AVX512BW-FCP-NEXT: vpermi2d %zmm2, %zmm0, %zmm1558; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} zmm3 = [0,4,8,12,16,1,5,9,13,17,2,6,10,14,18,3]559; AVX512BW-FCP-NEXT: vpermi2d %zmm2, %zmm0, %zmm3560; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, (%r9)561; AVX512BW-FCP-NEXT: vmovdqa %xmm1, 64(%r9)562; AVX512BW-FCP-NEXT: vzeroupper563; AVX512BW-FCP-NEXT: retq564;565; AVX512DQ-BW-LABEL: store_i32_stride5_vf4:566; AVX512DQ-BW: # %bb.0:567; AVX512DQ-BW-NEXT: vmovdqa (%rdi), %xmm0568; AVX512DQ-BW-NEXT: vmovdqa (%rdx), %xmm1569; AVX512DQ-BW-NEXT: vmovdqa (%r8), %xmm2570; AVX512DQ-BW-NEXT: vinserti128 $1, (%rcx), %ymm1, %ymm1571; AVX512DQ-BW-NEXT: vinserti128 $1, (%rsi), %ymm0, %ymm0572; AVX512DQ-BW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0573; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} xmm1 = [7,11,15,19]574; AVX512DQ-BW-NEXT: vpermi2d %zmm2, %zmm0, %zmm1575; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} zmm3 = [0,4,8,12,16,1,5,9,13,17,2,6,10,14,18,3]576; AVX512DQ-BW-NEXT: vpermi2d %zmm2, %zmm0, %zmm3577; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, (%r9)578; AVX512DQ-BW-NEXT: vmovdqa %xmm1, 64(%r9)579; AVX512DQ-BW-NEXT: vzeroupper580; AVX512DQ-BW-NEXT: retq581;582; AVX512DQ-BW-FCP-LABEL: store_i32_stride5_vf4:583; AVX512DQ-BW-FCP: # %bb.0:584; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdi), %xmm0585; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdx), %xmm1586; AVX512DQ-BW-FCP-NEXT: vmovdqa (%r8), %xmm2587; AVX512DQ-BW-FCP-NEXT: vinserti128 $1, (%rcx), %ymm1, %ymm1588; AVX512DQ-BW-FCP-NEXT: vinserti128 $1, (%rsi), %ymm0, %ymm0589; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0590; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} xmm1 = [7,11,15,19]591; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm2, %zmm0, %zmm1592; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} zmm3 = [0,4,8,12,16,1,5,9,13,17,2,6,10,14,18,3]593; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm2, %zmm0, %zmm3594; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, (%r9)595; AVX512DQ-BW-FCP-NEXT: vmovdqa %xmm1, 64(%r9)596; AVX512DQ-BW-FCP-NEXT: vzeroupper597; AVX512DQ-BW-FCP-NEXT: retq598 %in.vec0 = load <4 x i32>, ptr %in.vecptr0, align 64599 %in.vec1 = load <4 x i32>, ptr %in.vecptr1, align 64600 %in.vec2 = load <4 x i32>, ptr %in.vecptr2, align 64601 %in.vec3 = load <4 x i32>, ptr %in.vecptr3, align 64602 %in.vec4 = load <4 x i32>, ptr %in.vecptr4, align 64603 %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>604 %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>605 %3 = 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>606 %4 = shufflevector <4 x i32> %in.vec4, <4 x i32> poison, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>607 %5 = shufflevector <16 x i32> %3, <16 x i32> %4, <20 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>608 %interleaved.vec = shufflevector <20 x i32> %5, <20 x i32> poison, <20 x i32> <i32 0, i32 4, i32 8, i32 12, i32 16, i32 1, i32 5, i32 9, i32 13, i32 17, i32 2, i32 6, i32 10, i32 14, i32 18, i32 3, i32 7, i32 11, i32 15, i32 19>609 store <20 x i32> %interleaved.vec, ptr %out.vec, align 64610 ret void611}612 613define void @store_i32_stride5_vf8(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecptr2, ptr %in.vecptr3, ptr %in.vecptr4, ptr %out.vec) nounwind {614; SSE-LABEL: store_i32_stride5_vf8:615; SSE: # %bb.0:616; SSE-NEXT: movaps (%rdi), %xmm1617; SSE-NEXT: movaps 16(%rdi), %xmm5618; SSE-NEXT: movdqa (%rsi), %xmm6619; SSE-NEXT: movdqa 16(%rsi), %xmm3620; SSE-NEXT: movdqa (%rdx), %xmm7621; SSE-NEXT: movdqa 16(%rdx), %xmm10622; SSE-NEXT: movaps (%rcx), %xmm4623; SSE-NEXT: movaps 16(%rcx), %xmm2624; SSE-NEXT: movaps (%r8), %xmm0625; SSE-NEXT: movaps 16(%r8), %xmm11626; SSE-NEXT: movaps %xmm2, %xmm9627; SSE-NEXT: shufps {{.*#+}} xmm9 = xmm9[3,3],xmm11[3,3]628; SSE-NEXT: pshufd {{.*#+}} xmm12 = xmm10[3,3,3,3]629; SSE-NEXT: pshufd {{.*#+}} xmm8 = xmm3[3,3,3,3]630; SSE-NEXT: punpckldq {{.*#+}} xmm8 = xmm8[0],xmm12[0],xmm8[1],xmm12[1]631; SSE-NEXT: shufps {{.*#+}} xmm8 = xmm8[0,1],xmm9[0,2]632; SSE-NEXT: movaps %xmm4, %xmm12633; SSE-NEXT: shufps {{.*#+}} xmm12 = xmm12[3,3],xmm0[3,3]634; SSE-NEXT: pshufd {{.*#+}} xmm13 = xmm6[3,3,3,3]635; SSE-NEXT: pshufd {{.*#+}} xmm9 = xmm7[2,3,2,3]636; SSE-NEXT: movss {{.*#+}} xmm9 = xmm13[0],xmm9[1,2,3]637; SSE-NEXT: shufps {{.*#+}} xmm9 = xmm9[0,1],xmm12[0,2]638; SSE-NEXT: movaps %xmm5, %xmm13639; SSE-NEXT: shufps {{.*#+}} xmm13 = xmm13[3,1],xmm11[2,3]640; SSE-NEXT: movaps %xmm2, %xmm12641; SSE-NEXT: unpckhpd {{.*#+}} xmm12 = xmm12[1],xmm10[1]642; SSE-NEXT: shufps {{.*#+}} xmm12 = xmm12[2,0],xmm13[2,0]643; SSE-NEXT: movaps %xmm5, %xmm13644; SSE-NEXT: movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill645; SSE-NEXT: unpckhps {{.*#+}} xmm13 = xmm13[2],xmm3[2],xmm13[3],xmm3[3]646; SSE-NEXT: movaps %xmm10, %xmm14647; SSE-NEXT: unpcklps {{.*#+}} xmm14 = xmm14[0],xmm2[0],xmm14[1],xmm2[1]648; SSE-NEXT: shufps {{.*#+}} xmm2 = xmm2[1,1],xmm11[1,1]649; SSE-NEXT: shufps {{.*#+}} xmm2 = xmm2[0,2],xmm13[0,1]650; SSE-NEXT: movaps %xmm5, %xmm13651; SSE-NEXT: unpcklps {{.*#+}} xmm13 = xmm13[0],xmm3[0],xmm13[1],xmm3[1]652; SSE-NEXT: movlhps {{.*#+}} xmm13 = xmm13[0],xmm14[0]653; SSE-NEXT: movaps %xmm1, %xmm14654; SSE-NEXT: shufps {{.*#+}} xmm14 = xmm14[3,0],xmm0[2,0]655; SSE-NEXT: movaps %xmm0, %xmm3656; SSE-NEXT: movdqa %xmm7, %xmm15657; SSE-NEXT: punpckhdq {{.*#+}} xmm15 = xmm15[2],xmm4[2],xmm15[3],xmm4[3]658; SSE-NEXT: shufps {{.*#+}} xmm15 = xmm15[0,1],xmm14[2,0]659; SSE-NEXT: movaps %xmm1, %xmm14660; SSE-NEXT: movdqa %xmm6, %xmm0661; SSE-NEXT: unpckhps {{.*#+}} xmm14 = xmm14[2],xmm6[2],xmm14[3],xmm6[3]662; SSE-NEXT: movdqa %xmm7, %xmm6663; SSE-NEXT: punpckldq {{.*#+}} xmm6 = xmm6[0],xmm4[0],xmm6[1],xmm4[1]664; SSE-NEXT: shufps {{.*#+}} xmm4 = xmm4[1,1],xmm3[1,1]665; SSE-NEXT: shufps {{.*#+}} xmm4 = xmm4[0,2],xmm14[0,1]666; SSE-NEXT: movaps %xmm1, %xmm14667; SSE-NEXT: unpcklps {{.*#+}} xmm14 = xmm14[0],xmm0[0],xmm14[1],xmm0[1]668; SSE-NEXT: movlhps {{.*#+}} xmm14 = xmm14[0],xmm6[0]669; SSE-NEXT: shufps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm10 # 16-byte Folded Reload670; SSE-NEXT: # xmm10 = xmm10[1,1],mem[1,1]671; SSE-NEXT: shufps {{.*#+}} xmm5 = xmm5[0,1],xmm10[2,0]672; SSE-NEXT: movss {{.*#+}} xmm5 = xmm11[0],xmm5[1,2,3]673; SSE-NEXT: shufps {{.*#+}} xmm7 = xmm7[1,1],xmm0[1,1]674; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,1],xmm7[2,0]675; SSE-NEXT: movss {{.*#+}} xmm1 = xmm3[0],xmm1[1,2,3]676; SSE-NEXT: movaps %xmm14, (%r9)677; SSE-NEXT: movaps %xmm4, 32(%r9)678; SSE-NEXT: movaps %xmm15, 48(%r9)679; SSE-NEXT: movaps %xmm13, 80(%r9)680; SSE-NEXT: movaps %xmm2, 112(%r9)681; SSE-NEXT: movaps %xmm12, 128(%r9)682; SSE-NEXT: movaps %xmm1, 16(%r9)683; SSE-NEXT: movaps %xmm9, 64(%r9)684; SSE-NEXT: movaps %xmm5, 96(%r9)685; SSE-NEXT: movaps %xmm8, 144(%r9)686; SSE-NEXT: retq687;688; AVX-LABEL: store_i32_stride5_vf8:689; AVX: # %bb.0:690; AVX-NEXT: vmovaps (%rdi), %ymm0691; AVX-NEXT: vmovaps (%rsi), %ymm2692; AVX-NEXT: vmovaps (%rdx), %ymm1693; AVX-NEXT: vmovaps (%rcx), %ymm3694; AVX-NEXT: vmovaps (%rsi), %xmm5695; AVX-NEXT: vmovaps 16(%rsi), %xmm6696; AVX-NEXT: vmovaps (%rdi), %xmm7697; AVX-NEXT: vmovaps 16(%rdi), %xmm8698; AVX-NEXT: vinsertps {{.*#+}} xmm4 = zero,xmm7[1],xmm5[1],zero699; AVX-NEXT: vunpcklps {{.*#+}} xmm9 = xmm7[0],xmm5[0],xmm7[1],xmm5[1]700; AVX-NEXT: vinsertf128 $1, %xmm4, %ymm9, %ymm4701; AVX-NEXT: vmovaps (%rcx), %xmm9702; AVX-NEXT: vmovaps 16(%rcx), %xmm10703; AVX-NEXT: vmovaps (%rdx), %xmm11704; AVX-NEXT: vmovaps 16(%rdx), %xmm12705; AVX-NEXT: vunpcklps {{.*#+}} xmm13 = xmm11[0],xmm9[0],xmm11[1],xmm9[1]706; AVX-NEXT: vinsertf128 $1, %xmm11, %ymm13, %ymm13707; AVX-NEXT: vmovddup {{.*#+}} ymm13 = ymm13[0,0,2,2]708; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1],ymm13[2,3],ymm4[4,5,6],ymm13[7]709; AVX-NEXT: vbroadcastf128 {{.*#+}} ymm13 = mem[0,1,0,1]710; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm13[4],ymm4[5,6,7]711; AVX-NEXT: vunpckhps {{.*#+}} xmm11 = xmm11[2],xmm9[2],xmm11[3],xmm9[3]712; AVX-NEXT: vshufps {{.*#+}} xmm9 = xmm9[1,1,1,1]713; AVX-NEXT: vinsertf128 $1, %xmm11, %ymm9, %ymm9714; AVX-NEXT: vshufps {{.*#+}} xmm5 = xmm5[2,2,2,2]715; AVX-NEXT: vblendps {{.*#+}} xmm5 = xmm7[0,1,2],xmm5[3]716; AVX-NEXT: vinsertf128 $1, %xmm7, %ymm5, %ymm5717; AVX-NEXT: vblendps {{.*#+}} ymm5 = ymm9[0,1],ymm5[2,3],ymm9[4,5],ymm5[6,7]718; AVX-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0],ymm13[1],ymm5[2,3,4,5],ymm13[6],ymm5[7]719; AVX-NEXT: vshufps {{.*#+}} xmm7 = xmm6[2,2,2,2]720; AVX-NEXT: vblendps {{.*#+}} xmm7 = xmm8[0,1,2],xmm7[3]721; AVX-NEXT: vinsertps {{.*#+}} xmm6 = zero,xmm8[1],xmm6[1],zero722; AVX-NEXT: vinsertf128 $1, %xmm7, %ymm6, %ymm6723; AVX-NEXT: vshufps {{.*#+}} ymm7 = ymm3[1,1],ymm1[1,1],ymm3[5,5],ymm1[5,5]724; AVX-NEXT: vperm2f128 {{.*#+}} ymm7 = ymm7[2,3,2,3]725; AVX-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0,1,2],ymm7[3,4],ymm6[5,6,7]726; AVX-NEXT: vbroadcastf128 {{.*#+}} ymm7 = mem[0,1,0,1]727; AVX-NEXT: vblendps {{.*#+}} ymm6 = ymm7[0],ymm6[1,2,3,4],ymm7[5],ymm6[6,7]728; AVX-NEXT: vshufps {{.*#+}} xmm8 = xmm12[3,3],xmm10[3,3]729; AVX-NEXT: vunpckhps {{.*#+}} xmm9 = xmm12[2],xmm10[2],xmm12[3],xmm10[3]730; AVX-NEXT: vinsertf128 $1, %xmm8, %ymm9, %ymm8731; AVX-NEXT: vshufps {{.*#+}} ymm9 = ymm2[3,3],ymm0[3,3],ymm2[7,7],ymm0[7,7]732; AVX-NEXT: vperm2f128 {{.*#+}} ymm9 = ymm9[2,3,2,3]733; AVX-NEXT: vblendps {{.*#+}} ymm8 = ymm8[0,1,2],ymm9[3,4],ymm8[5,6,7]734; AVX-NEXT: vblendps {{.*#+}} ymm7 = ymm8[0,1],ymm7[2],ymm8[3,4,5,6],ymm7[7]735; AVX-NEXT: vshufps {{.*#+}} ymm2 = ymm2[3,0,2,3,7,4,6,7]736; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4],ymm2[5,6,7]737; AVX-NEXT: vshufps {{.*#+}} ymm2 = ymm3[0,1,3,0,4,5,7,4]738; AVX-NEXT: vshufpd {{.*#+}} ymm1 = ymm1[1,0,2,2]739; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4,5,6],ymm2[7]740; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm1[1,2,3],ymm0[4,5],ymm1[6,7]741; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],mem[3],ymm0[4,5,6,7]742; AVX-NEXT: vmovaps %ymm0, 64(%r9)743; AVX-NEXT: vmovaps %ymm7, 128(%r9)744; AVX-NEXT: vmovaps %ymm6, 96(%r9)745; AVX-NEXT: vmovaps %ymm5, 32(%r9)746; AVX-NEXT: vmovaps %ymm4, (%r9)747; AVX-NEXT: vzeroupper748; AVX-NEXT: retq749;750; AVX2-LABEL: store_i32_stride5_vf8:751; AVX2: # %bb.0:752; AVX2-NEXT: vmovaps (%rdi), %ymm1753; AVX2-NEXT: vmovaps (%rsi), %ymm3754; AVX2-NEXT: vmovaps (%rdx), %ymm2755; AVX2-NEXT: vmovaps (%rcx), %ymm4756; AVX2-NEXT: vmovaps (%r8), %ymm0757; AVX2-NEXT: vmovaps (%rcx), %xmm6758; AVX2-NEXT: vmovaps (%rdx), %xmm7759; AVX2-NEXT: vunpcklps {{.*#+}} xmm5 = xmm7[0],xmm6[0],xmm7[1],xmm6[1]760; AVX2-NEXT: vshufps {{.*#+}} xmm5 = xmm5[0,1,2,2]761; AVX2-NEXT: vpermpd {{.*#+}} ymm5 = ymm5[0,0,2,1]762; AVX2-NEXT: vmovaps (%rsi), %xmm8763; AVX2-NEXT: vmovaps (%rdi), %xmm9764; AVX2-NEXT: vunpcklps {{.*#+}} xmm10 = xmm9[0],xmm8[0],xmm9[1],xmm8[1]765; AVX2-NEXT: vshufps {{.*#+}} xmm10 = xmm10[0,1,3,2]766; AVX2-NEXT: vpermpd {{.*#+}} ymm11 = ymm10[0,1,1,1]767; AVX2-NEXT: vblendps {{.*#+}} ymm5 = ymm11[0,1],ymm5[2,3],ymm11[4,5,6],ymm5[7]768; AVX2-NEXT: vinsertf128 $1, (%r8), %ymm10, %ymm10769; AVX2-NEXT: vblendps {{.*#+}} ymm5 = ymm10[0],ymm5[1,2,3],ymm10[4],ymm5[5,6,7]770; AVX2-NEXT: vunpckhps {{.*#+}} xmm8 = xmm9[2],xmm8[2],xmm9[3],xmm8[3]771; AVX2-NEXT: vshufps {{.*#+}} xmm8 = xmm8[0,1,2,2]772; AVX2-NEXT: vpermpd {{.*#+}} ymm8 = ymm8[0,0,2,1]773; AVX2-NEXT: vshufps {{.*#+}} xmm6 = xmm6[1,1,2,2]774; AVX2-NEXT: vblendps {{.*#+}} xmm6 = xmm6[0,1],xmm7[2],xmm6[3]775; AVX2-NEXT: vpermpd {{.*#+}} ymm6 = ymm6[0,1,1,3]776; AVX2-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0,1],ymm8[2,3],ymm6[4,5],ymm8[6,7]777; AVX2-NEXT: vpermpd {{.*#+}} ymm7 = ymm0[0,1,2,1]778; AVX2-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0],ymm7[1],ymm6[2,3,4,5],ymm7[6],ymm6[7]779; AVX2-NEXT: vshufps {{.*#+}} ymm7 = ymm4[1,1,1,1,5,5,5,5]780; AVX2-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0],ymm2[1],ymm7[2,3,4],ymm2[5],ymm7[6,7]781; AVX2-NEXT: vpermpd {{.*#+}} ymm7 = ymm7[2,2,2,2]782; AVX2-NEXT: vshufps {{.*#+}} ymm8 = ymm3[1,1,2,2,5,5,6,6]783; AVX2-NEXT: vblendps {{.*#+}} ymm8 = ymm8[0],ymm1[1,2],ymm8[3,4],ymm1[5,6],ymm8[7]784; AVX2-NEXT: vpermpd {{.*#+}} ymm8 = ymm8[2,2,2,3]785; AVX2-NEXT: vblendps {{.*#+}} ymm7 = ymm8[0,1,2],ymm7[3,4],ymm8[5,6,7]786; AVX2-NEXT: vbroadcastsd 16(%r8), %ymm8787; AVX2-NEXT: vblendps {{.*#+}} ymm7 = ymm8[0],ymm7[1,2,3,4],ymm8[5],ymm7[6,7]788; AVX2-NEXT: vunpckhps {{.*#+}} ymm8 = ymm3[2],ymm1[2],ymm3[3],ymm1[3],ymm3[6],ymm1[6],ymm3[7],ymm1[7]789; AVX2-NEXT: vpermpd {{.*#+}} ymm8 = ymm8[3,3,3,3]790; AVX2-NEXT: vshufps {{.*#+}} ymm9 = ymm4[0,2,3,3,4,6,7,7]791; AVX2-NEXT: vshufps {{.*#+}} ymm10 = ymm2[2,3,2,3,6,7,6,7]792; AVX2-NEXT: vblendps {{.*#+}} ymm9 = ymm10[0],ymm9[1,2],ymm10[3,4],ymm9[5,6],ymm10[7]793; AVX2-NEXT: vpermpd {{.*#+}} ymm9 = ymm9[2,1,3,3]794; AVX2-NEXT: vblendps {{.*#+}} ymm8 = ymm9[0,1,2],ymm8[3,4],ymm9[5,6,7]795; AVX2-NEXT: vbroadcastsd 24(%r8), %ymm9796; AVX2-NEXT: vblendps {{.*#+}} ymm8 = ymm8[0,1],ymm9[2],ymm8[3,4,5,6],ymm9[7]797; AVX2-NEXT: vshufps {{.*#+}} ymm3 = ymm3[3,0,2,3,7,4,6,7]798; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm3[0,1,2,3],ymm1[4],ymm3[5,6,7]799; AVX2-NEXT: vshufps {{.*#+}} ymm3 = ymm4[0,1,3,0,4,5,7,4]800; AVX2-NEXT: vshufps {{.*#+}} ymm2 = ymm2[2,3,0,1,6,7,4,5]801; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1],ymm3[2],ymm2[3,4,5,6],ymm3[7]802; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1,2,3],ymm1[4,5],ymm2[6,7]803; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3],ymm1[4,5,6,7]804; AVX2-NEXT: vmovaps %ymm0, 64(%r9)805; AVX2-NEXT: vmovaps %ymm8, 128(%r9)806; AVX2-NEXT: vmovaps %ymm7, 96(%r9)807; AVX2-NEXT: vmovaps %ymm6, 32(%r9)808; AVX2-NEXT: vmovaps %ymm5, (%r9)809; AVX2-NEXT: vzeroupper810; AVX2-NEXT: retq811;812; AVX2-FP-LABEL: store_i32_stride5_vf8:813; AVX2-FP: # %bb.0:814; AVX2-FP-NEXT: vmovaps (%rdi), %ymm1815; AVX2-FP-NEXT: vmovaps (%rsi), %ymm3816; AVX2-FP-NEXT: vmovaps (%rdx), %ymm2817; AVX2-FP-NEXT: vmovaps (%rcx), %ymm4818; AVX2-FP-NEXT: vmovaps (%r8), %ymm0819; AVX2-FP-NEXT: vmovaps (%rcx), %xmm6820; AVX2-FP-NEXT: vmovaps (%rdx), %xmm7821; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm5 = xmm7[0],xmm6[0],xmm7[1],xmm6[1]822; AVX2-FP-NEXT: vshufps {{.*#+}} xmm5 = xmm5[0,1,2,2]823; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm5 = ymm5[0,0,2,1]824; AVX2-FP-NEXT: vmovaps (%rsi), %xmm8825; AVX2-FP-NEXT: vmovaps (%rdi), %xmm9826; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm10 = xmm9[0],xmm8[0],xmm9[1],xmm8[1]827; AVX2-FP-NEXT: vshufps {{.*#+}} xmm10 = xmm10[0,1,3,2]828; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm11 = ymm10[0,1,1,1]829; AVX2-FP-NEXT: vblendps {{.*#+}} ymm5 = ymm11[0,1],ymm5[2,3],ymm11[4,5,6],ymm5[7]830; AVX2-FP-NEXT: vinsertf128 $1, (%r8), %ymm10, %ymm10831; AVX2-FP-NEXT: vblendps {{.*#+}} ymm5 = ymm10[0],ymm5[1,2,3],ymm10[4],ymm5[5,6,7]832; AVX2-FP-NEXT: vunpckhps {{.*#+}} xmm8 = xmm9[2],xmm8[2],xmm9[3],xmm8[3]833; AVX2-FP-NEXT: vshufps {{.*#+}} xmm8 = xmm8[0,1,2,2]834; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm8 = ymm8[0,0,2,1]835; AVX2-FP-NEXT: vshufps {{.*#+}} xmm6 = xmm6[1,1,2,2]836; AVX2-FP-NEXT: vblendps {{.*#+}} xmm6 = xmm6[0,1],xmm7[2],xmm6[3]837; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm6 = ymm6[0,1,1,3]838; AVX2-FP-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0,1],ymm8[2,3],ymm6[4,5],ymm8[6,7]839; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm7 = ymm0[0,1,2,1]840; AVX2-FP-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0],ymm7[1],ymm6[2,3,4,5],ymm7[6],ymm6[7]841; AVX2-FP-NEXT: vshufps {{.*#+}} ymm7 = ymm4[1,1,1,1,5,5,5,5]842; AVX2-FP-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0],ymm2[1],ymm7[2,3,4],ymm2[5],ymm7[6,7]843; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm7 = ymm7[2,2,2,2]844; AVX2-FP-NEXT: vshufps {{.*#+}} ymm8 = ymm3[1,1,2,2,5,5,6,6]845; AVX2-FP-NEXT: vblendps {{.*#+}} ymm8 = ymm8[0],ymm1[1,2],ymm8[3,4],ymm1[5,6],ymm8[7]846; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm8 = ymm8[2,2,2,3]847; AVX2-FP-NEXT: vblendps {{.*#+}} ymm7 = ymm8[0,1,2],ymm7[3,4],ymm8[5,6,7]848; AVX2-FP-NEXT: vbroadcastsd 16(%r8), %ymm8849; AVX2-FP-NEXT: vblendps {{.*#+}} ymm7 = ymm8[0],ymm7[1,2,3,4],ymm8[5],ymm7[6,7]850; AVX2-FP-NEXT: vunpckhps {{.*#+}} ymm8 = ymm3[2],ymm1[2],ymm3[3],ymm1[3],ymm3[6],ymm1[6],ymm3[7],ymm1[7]851; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm8 = ymm8[3,3,3,3]852; AVX2-FP-NEXT: vshufps {{.*#+}} ymm9 = ymm4[0,2,3,3,4,6,7,7]853; AVX2-FP-NEXT: vshufps {{.*#+}} ymm10 = ymm2[2,3,2,3,6,7,6,7]854; AVX2-FP-NEXT: vblendps {{.*#+}} ymm9 = ymm10[0],ymm9[1,2],ymm10[3,4],ymm9[5,6],ymm10[7]855; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm9 = ymm9[2,1,3,3]856; AVX2-FP-NEXT: vblendps {{.*#+}} ymm8 = ymm9[0,1,2],ymm8[3,4],ymm9[5,6,7]857; AVX2-FP-NEXT: vbroadcastsd 24(%r8), %ymm9858; AVX2-FP-NEXT: vblendps {{.*#+}} ymm8 = ymm8[0,1],ymm9[2],ymm8[3,4,5,6],ymm9[7]859; AVX2-FP-NEXT: vshufps {{.*#+}} ymm3 = ymm3[3,0,2,3,7,4,6,7]860; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm3[0,1,2,3],ymm1[4],ymm3[5,6,7]861; AVX2-FP-NEXT: vshufps {{.*#+}} ymm3 = ymm4[0,1,3,0,4,5,7,4]862; AVX2-FP-NEXT: vshufps {{.*#+}} ymm2 = ymm2[2,3,0,1,6,7,4,5]863; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1],ymm3[2],ymm2[3,4,5,6],ymm3[7]864; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1,2,3],ymm1[4,5],ymm2[6,7]865; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3],ymm1[4,5,6,7]866; AVX2-FP-NEXT: vmovaps %ymm0, 64(%r9)867; AVX2-FP-NEXT: vmovaps %ymm8, 128(%r9)868; AVX2-FP-NEXT: vmovaps %ymm7, 96(%r9)869; AVX2-FP-NEXT: vmovaps %ymm6, 32(%r9)870; AVX2-FP-NEXT: vmovaps %ymm5, (%r9)871; AVX2-FP-NEXT: vzeroupper872; AVX2-FP-NEXT: retq873;874; AVX2-FCP-LABEL: store_i32_stride5_vf8:875; AVX2-FCP: # %bb.0:876; AVX2-FCP-NEXT: vmovaps (%rdi), %ymm1877; AVX2-FCP-NEXT: vmovaps (%rsi), %ymm3878; AVX2-FCP-NEXT: vmovaps (%rdx), %ymm2879; AVX2-FCP-NEXT: vmovaps (%rcx), %ymm4880; AVX2-FCP-NEXT: vmovaps (%r8), %ymm0881; AVX2-FCP-NEXT: vmovaps (%rsi), %xmm6882; AVX2-FCP-NEXT: vmovaps (%rdi), %xmm7883; AVX2-FCP-NEXT: vunpckhps {{.*#+}} xmm5 = xmm7[2],xmm6[2],xmm7[3],xmm6[3]884; AVX2-FCP-NEXT: vmovaps {{.*#+}} ymm8 = [0,1,0,1,u,u,2,2]885; AVX2-FCP-NEXT: vpermps %ymm5, %ymm8, %ymm5886; AVX2-FCP-NEXT: vmovaps (%rdx), %xmm9887; AVX2-FCP-NEXT: vmovaps (%rcx), %xmm10888; AVX2-FCP-NEXT: vshufps {{.*#+}} xmm11 = xmm10[1,1,2,2]889; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm11 = xmm11[0,1],xmm9[2],xmm11[3]890; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm11 = ymm11[0,1,1,3]891; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm5 = ymm11[0,1],ymm5[2,3],ymm11[4,5],ymm5[6,7]892; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm11 = ymm0[0,1,2,1]893; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0],ymm11[1],ymm5[2,3,4,5],ymm11[6],ymm5[7]894; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm9 = xmm9[0],xmm10[0],xmm9[1],xmm10[1]895; AVX2-FCP-NEXT: vpermps %ymm9, %ymm8, %ymm8896; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm6 = xmm7[0],xmm6[0],xmm7[1],xmm6[1]897; AVX2-FCP-NEXT: vmovaps {{.*#+}} ymm7 = [0,1,u,u,3,2,3,u]898; AVX2-FCP-NEXT: vpermps %ymm6, %ymm7, %ymm7899; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0,1],ymm8[2,3],ymm7[4,5,6],ymm8[7]900; AVX2-FCP-NEXT: vinsertf128 $1, (%r8), %ymm6, %ymm6901; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0],ymm7[1,2,3],ymm6[4],ymm7[5,6,7]902; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm7 = ymm4[1,1,1,1,5,5,5,5]903; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0],ymm2[1],ymm7[2,3,4],ymm2[5],ymm7[6,7]904; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm7 = ymm7[2,2,2,2]905; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm8 = ymm3[1,1,2,2,5,5,6,6]906; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm8 = ymm8[0],ymm1[1,2],ymm8[3,4],ymm1[5,6],ymm8[7]907; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm8 = ymm8[2,2,2,3]908; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm7 = ymm8[0,1,2],ymm7[3,4],ymm8[5,6,7]909; AVX2-FCP-NEXT: vbroadcastsd 16(%r8), %ymm8910; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm7 = ymm8[0],ymm7[1,2,3,4],ymm8[5],ymm7[6,7]911; AVX2-FCP-NEXT: vunpckhps {{.*#+}} ymm8 = ymm3[2],ymm1[2],ymm3[3],ymm1[3],ymm3[6],ymm1[6],ymm3[7],ymm1[7]912; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm8 = ymm8[3,3,3,3]913; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm9 = ymm4[0,2,3,3,4,6,7,7]914; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm10 = ymm2[2,3,2,3,6,7,6,7]915; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm9 = ymm10[0],ymm9[1,2],ymm10[3,4],ymm9[5,6],ymm10[7]916; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm9 = ymm9[2,1,3,3]917; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm8 = ymm9[0,1,2],ymm8[3,4],ymm9[5,6,7]918; AVX2-FCP-NEXT: vbroadcastsd 24(%r8), %ymm9919; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm8 = ymm8[0,1],ymm9[2],ymm8[3,4,5,6],ymm9[7]920; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm3 = ymm3[3,0,2,3,7,4,6,7]921; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm3[0,1,2,3],ymm1[4],ymm3[5,6,7]922; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm3 = ymm4[0,1,3,0,4,5,7,4]923; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm2 = ymm2[2,3,0,1,6,7,4,5]924; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1],ymm3[2],ymm2[3,4,5,6],ymm3[7]925; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1,2,3],ymm1[4,5],ymm2[6,7]926; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3],ymm1[4,5,6,7]927; AVX2-FCP-NEXT: vmovaps %ymm0, 64(%r9)928; AVX2-FCP-NEXT: vmovaps %ymm8, 128(%r9)929; AVX2-FCP-NEXT: vmovaps %ymm7, 96(%r9)930; AVX2-FCP-NEXT: vmovaps %ymm6, (%r9)931; AVX2-FCP-NEXT: vmovaps %ymm5, 32(%r9)932; AVX2-FCP-NEXT: vzeroupper933; AVX2-FCP-NEXT: retq934;935; AVX512-LABEL: store_i32_stride5_vf8:936; AVX512: # %bb.0:937; AVX512-NEXT: vmovdqa (%rdi), %ymm0938; AVX512-NEXT: vmovdqa (%rdx), %ymm1939; AVX512-NEXT: vmovdqa (%r8), %ymm2940; AVX512-NEXT: vinserti64x4 $1, (%rsi), %zmm0, %zmm0941; AVX512-NEXT: vinserti64x4 $1, (%rcx), %zmm1, %zmm1942; AVX512-NEXT: vpmovsxbd {{.*#+}} ymm3 = [6,14,0,23,31,7,15,0]943; AVX512-NEXT: vpermi2d %zmm0, %zmm1, %zmm3944; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm4 = [11,19,27,0,4,12,20,28,0,5,13,21,29,0,6,14]945; AVX512-NEXT: vpermi2d %zmm1, %zmm0, %zmm4946; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm5 = [0,1,2,19,4,5,6,7,20,9,10,11,12,21,14,15]947; AVX512-NEXT: vpermi2d %zmm2, %zmm4, %zmm5948; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm4 = [0,8,16,24,0,1,9,17,25,0,2,10,18,26,0,3]949; AVX512-NEXT: vpermi2d %zmm1, %zmm0, %zmm4950; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm0 = [0,1,2,3,16,5,6,7,8,17,10,11,12,13,18,15]951; AVX512-NEXT: vpermi2d %zmm2, %zmm4, %zmm0952; AVX512-NEXT: vpbroadcastq 24(%r8), %ymm1953; AVX512-NEXT: vmovdqa64 %zmm0, (%r9)954; AVX512-NEXT: vmovdqa64 %zmm5, 64(%r9)955; AVX512-NEXT: vpblendd {{.*#+}} ymm0 = ymm3[0,1],ymm1[2],ymm3[3,4,5,6],ymm1[7]956; AVX512-NEXT: vmovdqa %ymm0, 128(%r9)957; AVX512-NEXT: vzeroupper958; AVX512-NEXT: retq959;960; AVX512-FCP-LABEL: store_i32_stride5_vf8:961; AVX512-FCP: # %bb.0:962; AVX512-FCP-NEXT: vmovdqa (%rdi), %ymm0963; AVX512-FCP-NEXT: vmovdqa (%rdx), %ymm1964; AVX512-FCP-NEXT: vmovdqa (%r8), %ymm2965; AVX512-FCP-NEXT: vinserti64x4 $1, (%rsi), %zmm0, %zmm0966; AVX512-FCP-NEXT: vinserti64x4 $1, (%rcx), %zmm1, %zmm1967; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} ymm3 = [6,14,0,23,31,7,15,0]968; AVX512-FCP-NEXT: vpermi2d %zmm0, %zmm1, %zmm3969; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm4 = [11,19,27,0,4,12,20,28,0,5,13,21,29,0,6,14]970; AVX512-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm4971; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm5 = [0,1,2,19,4,5,6,7,20,9,10,11,12,21,14,15]972; AVX512-FCP-NEXT: vpermi2d %zmm2, %zmm4, %zmm5973; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm4 = [0,8,16,24,0,1,9,17,25,0,2,10,18,26,0,3]974; AVX512-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm4975; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm0 = [0,1,2,3,16,5,6,7,8,17,10,11,12,13,18,15]976; AVX512-FCP-NEXT: vpermi2d %zmm2, %zmm4, %zmm0977; AVX512-FCP-NEXT: vpbroadcastq 24(%r8), %ymm1978; AVX512-FCP-NEXT: vmovdqa64 %zmm0, (%r9)979; AVX512-FCP-NEXT: vmovdqa64 %zmm5, 64(%r9)980; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm3[0,1],ymm1[2],ymm3[3,4,5,6],ymm1[7]981; AVX512-FCP-NEXT: vmovdqa %ymm0, 128(%r9)982; AVX512-FCP-NEXT: vzeroupper983; AVX512-FCP-NEXT: retq984;985; AVX512DQ-LABEL: store_i32_stride5_vf8:986; AVX512DQ: # %bb.0:987; AVX512DQ-NEXT: vmovdqa (%rdi), %ymm0988; AVX512DQ-NEXT: vmovdqa (%rdx), %ymm1989; AVX512DQ-NEXT: vmovdqa (%r8), %ymm2990; AVX512DQ-NEXT: vinserti64x4 $1, (%rsi), %zmm0, %zmm0991; AVX512DQ-NEXT: vinserti64x4 $1, (%rcx), %zmm1, %zmm1992; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} ymm3 = [6,14,0,23,31,7,15,0]993; AVX512DQ-NEXT: vpermi2d %zmm0, %zmm1, %zmm3994; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm4 = [11,19,27,0,4,12,20,28,0,5,13,21,29,0,6,14]995; AVX512DQ-NEXT: vpermi2d %zmm1, %zmm0, %zmm4996; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm5 = [0,1,2,19,4,5,6,7,20,9,10,11,12,21,14,15]997; AVX512DQ-NEXT: vpermi2d %zmm2, %zmm4, %zmm5998; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm4 = [0,8,16,24,0,1,9,17,25,0,2,10,18,26,0,3]999; AVX512DQ-NEXT: vpermi2d %zmm1, %zmm0, %zmm41000; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm0 = [0,1,2,3,16,5,6,7,8,17,10,11,12,13,18,15]1001; AVX512DQ-NEXT: vpermi2d %zmm2, %zmm4, %zmm01002; AVX512DQ-NEXT: vpbroadcastq 24(%r8), %ymm11003; AVX512DQ-NEXT: vmovdqa64 %zmm0, (%r9)1004; AVX512DQ-NEXT: vmovdqa64 %zmm5, 64(%r9)1005; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm0 = ymm3[0,1],ymm1[2],ymm3[3,4,5,6],ymm1[7]1006; AVX512DQ-NEXT: vmovdqa %ymm0, 128(%r9)1007; AVX512DQ-NEXT: vzeroupper1008; AVX512DQ-NEXT: retq1009;1010; AVX512DQ-FCP-LABEL: store_i32_stride5_vf8:1011; AVX512DQ-FCP: # %bb.0:1012; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %ymm01013; AVX512DQ-FCP-NEXT: vmovdqa (%rdx), %ymm11014; AVX512DQ-FCP-NEXT: vmovdqa (%r8), %ymm21015; AVX512DQ-FCP-NEXT: vinserti64x4 $1, (%rsi), %zmm0, %zmm01016; AVX512DQ-FCP-NEXT: vinserti64x4 $1, (%rcx), %zmm1, %zmm11017; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} ymm3 = [6,14,0,23,31,7,15,0]1018; AVX512DQ-FCP-NEXT: vpermi2d %zmm0, %zmm1, %zmm31019; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm4 = [11,19,27,0,4,12,20,28,0,5,13,21,29,0,6,14]1020; AVX512DQ-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm41021; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm5 = [0,1,2,19,4,5,6,7,20,9,10,11,12,21,14,15]1022; AVX512DQ-FCP-NEXT: vpermi2d %zmm2, %zmm4, %zmm51023; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm4 = [0,8,16,24,0,1,9,17,25,0,2,10,18,26,0,3]1024; AVX512DQ-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm41025; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm0 = [0,1,2,3,16,5,6,7,8,17,10,11,12,13,18,15]1026; AVX512DQ-FCP-NEXT: vpermi2d %zmm2, %zmm4, %zmm01027; AVX512DQ-FCP-NEXT: vpbroadcastq 24(%r8), %ymm11028; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, (%r9)1029; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm5, 64(%r9)1030; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm3[0,1],ymm1[2],ymm3[3,4,5,6],ymm1[7]1031; AVX512DQ-FCP-NEXT: vmovdqa %ymm0, 128(%r9)1032; AVX512DQ-FCP-NEXT: vzeroupper1033; AVX512DQ-FCP-NEXT: retq1034;1035; AVX512BW-LABEL: store_i32_stride5_vf8:1036; AVX512BW: # %bb.0:1037; AVX512BW-NEXT: vmovdqa (%rdi), %ymm01038; AVX512BW-NEXT: vmovdqa (%rdx), %ymm11039; AVX512BW-NEXT: vmovdqa (%r8), %ymm21040; AVX512BW-NEXT: vinserti64x4 $1, (%rsi), %zmm0, %zmm01041; AVX512BW-NEXT: vinserti64x4 $1, (%rcx), %zmm1, %zmm11042; AVX512BW-NEXT: vpmovsxbd {{.*#+}} ymm3 = [6,14,0,23,31,7,15,0]1043; AVX512BW-NEXT: vpermi2d %zmm0, %zmm1, %zmm31044; AVX512BW-NEXT: vpmovsxbd {{.*#+}} zmm4 = [11,19,27,0,4,12,20,28,0,5,13,21,29,0,6,14]1045; AVX512BW-NEXT: vpermi2d %zmm1, %zmm0, %zmm41046; AVX512BW-NEXT: vpmovsxbd {{.*#+}} zmm5 = [0,1,2,19,4,5,6,7,20,9,10,11,12,21,14,15]1047; AVX512BW-NEXT: vpermi2d %zmm2, %zmm4, %zmm51048; AVX512BW-NEXT: vpmovsxbd {{.*#+}} zmm4 = [0,8,16,24,0,1,9,17,25,0,2,10,18,26,0,3]1049; AVX512BW-NEXT: vpermi2d %zmm1, %zmm0, %zmm41050; AVX512BW-NEXT: vpmovsxbd {{.*#+}} zmm0 = [0,1,2,3,16,5,6,7,8,17,10,11,12,13,18,15]1051; AVX512BW-NEXT: vpermi2d %zmm2, %zmm4, %zmm01052; AVX512BW-NEXT: vpbroadcastq 24(%r8), %ymm11053; AVX512BW-NEXT: vmovdqa64 %zmm0, (%r9)1054; AVX512BW-NEXT: vmovdqa64 %zmm5, 64(%r9)1055; AVX512BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm3[0,1],ymm1[2],ymm3[3,4,5,6],ymm1[7]1056; AVX512BW-NEXT: vmovdqa %ymm0, 128(%r9)1057; AVX512BW-NEXT: vzeroupper1058; AVX512BW-NEXT: retq1059;1060; AVX512BW-FCP-LABEL: store_i32_stride5_vf8:1061; AVX512BW-FCP: # %bb.0:1062; AVX512BW-FCP-NEXT: vmovdqa (%rdi), %ymm01063; AVX512BW-FCP-NEXT: vmovdqa (%rdx), %ymm11064; AVX512BW-FCP-NEXT: vmovdqa (%r8), %ymm21065; AVX512BW-FCP-NEXT: vinserti64x4 $1, (%rsi), %zmm0, %zmm01066; AVX512BW-FCP-NEXT: vinserti64x4 $1, (%rcx), %zmm1, %zmm11067; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm3 = [6,14,0,23,31,7,15,0]1068; AVX512BW-FCP-NEXT: vpermi2d %zmm0, %zmm1, %zmm31069; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} zmm4 = [11,19,27,0,4,12,20,28,0,5,13,21,29,0,6,14]1070; AVX512BW-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm41071; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} zmm5 = [0,1,2,19,4,5,6,7,20,9,10,11,12,21,14,15]1072; AVX512BW-FCP-NEXT: vpermi2d %zmm2, %zmm4, %zmm51073; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} zmm4 = [0,8,16,24,0,1,9,17,25,0,2,10,18,26,0,3]1074; AVX512BW-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm41075; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} zmm0 = [0,1,2,3,16,5,6,7,8,17,10,11,12,13,18,15]1076; AVX512BW-FCP-NEXT: vpermi2d %zmm2, %zmm4, %zmm01077; AVX512BW-FCP-NEXT: vpbroadcastq 24(%r8), %ymm11078; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, (%r9)1079; AVX512BW-FCP-NEXT: vmovdqa64 %zmm5, 64(%r9)1080; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm3[0,1],ymm1[2],ymm3[3,4,5,6],ymm1[7]1081; AVX512BW-FCP-NEXT: vmovdqa %ymm0, 128(%r9)1082; AVX512BW-FCP-NEXT: vzeroupper1083; AVX512BW-FCP-NEXT: retq1084;1085; AVX512DQ-BW-LABEL: store_i32_stride5_vf8:1086; AVX512DQ-BW: # %bb.0:1087; AVX512DQ-BW-NEXT: vmovdqa (%rdi), %ymm01088; AVX512DQ-BW-NEXT: vmovdqa (%rdx), %ymm11089; AVX512DQ-BW-NEXT: vmovdqa (%r8), %ymm21090; AVX512DQ-BW-NEXT: vinserti64x4 $1, (%rsi), %zmm0, %zmm01091; AVX512DQ-BW-NEXT: vinserti64x4 $1, (%rcx), %zmm1, %zmm11092; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} ymm3 = [6,14,0,23,31,7,15,0]1093; AVX512DQ-BW-NEXT: vpermi2d %zmm0, %zmm1, %zmm31094; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} zmm4 = [11,19,27,0,4,12,20,28,0,5,13,21,29,0,6,14]1095; AVX512DQ-BW-NEXT: vpermi2d %zmm1, %zmm0, %zmm41096; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} zmm5 = [0,1,2,19,4,5,6,7,20,9,10,11,12,21,14,15]1097; AVX512DQ-BW-NEXT: vpermi2d %zmm2, %zmm4, %zmm51098; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} zmm4 = [0,8,16,24,0,1,9,17,25,0,2,10,18,26,0,3]1099; AVX512DQ-BW-NEXT: vpermi2d %zmm1, %zmm0, %zmm41100; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} zmm0 = [0,1,2,3,16,5,6,7,8,17,10,11,12,13,18,15]1101; AVX512DQ-BW-NEXT: vpermi2d %zmm2, %zmm4, %zmm01102; AVX512DQ-BW-NEXT: vpbroadcastq 24(%r8), %ymm11103; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, (%r9)1104; AVX512DQ-BW-NEXT: vmovdqa64 %zmm5, 64(%r9)1105; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm3[0,1],ymm1[2],ymm3[3,4,5,6],ymm1[7]1106; AVX512DQ-BW-NEXT: vmovdqa %ymm0, 128(%r9)1107; AVX512DQ-BW-NEXT: vzeroupper1108; AVX512DQ-BW-NEXT: retq1109;1110; AVX512DQ-BW-FCP-LABEL: store_i32_stride5_vf8:1111; AVX512DQ-BW-FCP: # %bb.0:1112; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdi), %ymm01113; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdx), %ymm11114; AVX512DQ-BW-FCP-NEXT: vmovdqa (%r8), %ymm21115; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, (%rsi), %zmm0, %zmm01116; AVX512DQ-BW-FCP-NEXT: vinserti64x4 $1, (%rcx), %zmm1, %zmm11117; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} ymm3 = [6,14,0,23,31,7,15,0]1118; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm0, %zmm1, %zmm31119; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} zmm4 = [11,19,27,0,4,12,20,28,0,5,13,21,29,0,6,14]1120; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm41121; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} zmm5 = [0,1,2,19,4,5,6,7,20,9,10,11,12,21,14,15]1122; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm2, %zmm4, %zmm51123; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} zmm4 = [0,8,16,24,0,1,9,17,25,0,2,10,18,26,0,3]1124; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm41125; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} zmm0 = [0,1,2,3,16,5,6,7,8,17,10,11,12,13,18,15]1126; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm2, %zmm4, %zmm01127; AVX512DQ-BW-FCP-NEXT: vpbroadcastq 24(%r8), %ymm11128; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, (%r9)1129; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm5, 64(%r9)1130; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm3[0,1],ymm1[2],ymm3[3,4,5,6],ymm1[7]1131; AVX512DQ-BW-FCP-NEXT: vmovdqa %ymm0, 128(%r9)1132; AVX512DQ-BW-FCP-NEXT: vzeroupper1133; AVX512DQ-BW-FCP-NEXT: retq1134 %in.vec0 = load <8 x i32>, ptr %in.vecptr0, align 641135 %in.vec1 = load <8 x i32>, ptr %in.vecptr1, align 641136 %in.vec2 = load <8 x i32>, ptr %in.vecptr2, align 641137 %in.vec3 = load <8 x i32>, ptr %in.vecptr3, align 641138 %in.vec4 = load <8 x i32>, ptr %in.vecptr4, align 641139 %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>1140 %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>1141 %3 = 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>1142 %4 = shufflevector <8 x i32> %in.vec4, <8 x i32> poison, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>1143 %5 = shufflevector <32 x i32> %3, <32 x i32> %4, <40 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>1144 %interleaved.vec = shufflevector <40 x i32> %5, <40 x i32> poison, <40 x i32> <i32 0, i32 8, i32 16, i32 24, i32 32, i32 1, i32 9, i32 17, i32 25, i32 33, i32 2, i32 10, i32 18, i32 26, i32 34, i32 3, i32 11, i32 19, i32 27, i32 35, i32 4, i32 12, i32 20, i32 28, i32 36, i32 5, i32 13, i32 21, i32 29, i32 37, i32 6, i32 14, i32 22, i32 30, i32 38, i32 7, i32 15, i32 23, i32 31, i32 39>1145 store <40 x i32> %interleaved.vec, ptr %out.vec, align 641146 ret void1147}1148 1149define void @store_i32_stride5_vf16(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecptr2, ptr %in.vecptr3, ptr %in.vecptr4, ptr %out.vec) nounwind {1150; SSE-LABEL: store_i32_stride5_vf16:1151; SSE: # %bb.0:1152; SSE-NEXT: subq $168, %rsp1153; SSE-NEXT: movdqa (%rsi), %xmm81154; SSE-NEXT: movdqa 16(%rsi), %xmm61155; SSE-NEXT: movdqa 32(%rsi), %xmm91156; SSE-NEXT: movdqa %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1157; SSE-NEXT: movdqa (%rdx), %xmm101158; SSE-NEXT: movdqa 16(%rdx), %xmm71159; SSE-NEXT: movdqa 32(%rdx), %xmm41160; SSE-NEXT: movaps (%rcx), %xmm51161; SSE-NEXT: movaps 16(%rcx), %xmm141162; SSE-NEXT: movaps 32(%rcx), %xmm121163; SSE-NEXT: movaps (%r8), %xmm31164; SSE-NEXT: movaps 16(%r8), %xmm151165; SSE-NEXT: movaps 32(%r8), %xmm111166; SSE-NEXT: movaps %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1167; SSE-NEXT: movaps %xmm5, %xmm01168; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,3],xmm3[3,3]1169; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm8[3,3,3,3]1170; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm10[2,3,2,3]1171; SSE-NEXT: movss {{.*#+}} xmm2 = xmm1[0],xmm2[1,2,3]1172; SSE-NEXT: shufps {{.*#+}} xmm2 = xmm2[0,1],xmm0[0,2]1173; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1174; SSE-NEXT: movaps %xmm14, %xmm01175; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,3],xmm15[3,3]1176; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm7[3,3,3,3]1177; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm6[3,3,3,3]1178; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]1179; SSE-NEXT: shufps {{.*#+}} xmm2 = xmm2[0,1],xmm0[0,2]1180; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1181; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm9[3,3,3,3]1182; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm4[2,3,2,3]1183; SSE-NEXT: movss {{.*#+}} xmm1 = xmm0[0],xmm1[1,2,3]1184; SSE-NEXT: movaps %xmm12, %xmm01185; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,3],xmm11[3,3]1186; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,1],xmm0[0,2]1187; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1188; SSE-NEXT: movdqa 48(%rsi), %xmm11189; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1190; SSE-NEXT: movdqa 48(%rdx), %xmm111191; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm11[3,3,3,3]1192; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[3,3,3,3]1193; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]1194; SSE-NEXT: movaps 48(%rcx), %xmm91195; SSE-NEXT: movaps 48(%r8), %xmm21196; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1197; SSE-NEXT: movaps %xmm9, %xmm01198; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,3],xmm2[3,3]1199; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,1],xmm0[0,2]1200; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1201; SSE-NEXT: movdqa %xmm10, (%rsp) # 16-byte Spill1202; SSE-NEXT: movdqa %xmm10, %xmm01203; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm5[0],xmm0[1],xmm5[1]1204; SSE-NEXT: movdqa (%rdi), %xmm131205; SSE-NEXT: movdqa %xmm13, %xmm21206; SSE-NEXT: movdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1207; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm8[0],xmm2[1],xmm8[1]1208; SSE-NEXT: punpcklqdq {{.*#+}} xmm2 = xmm2[0],xmm0[0]1209; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1210; SSE-NEXT: movdqa %xmm13, %xmm01211; SSE-NEXT: punpckhdq {{.*#+}} xmm0 = xmm0[2],xmm8[2],xmm0[3],xmm8[3]1212; SSE-NEXT: movdqa %xmm10, %xmm21213; SSE-NEXT: punpckhdq {{.*#+}} xmm2 = xmm2[2],xmm5[2],xmm2[3],xmm5[3]1214; SSE-NEXT: movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1215; SSE-NEXT: shufps {{.*#+}} xmm5 = xmm5[1,1],xmm3[1,1]1216; SSE-NEXT: shufps {{.*#+}} xmm5 = xmm5[0,2],xmm0[0,1]1217; SSE-NEXT: movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1218; SSE-NEXT: movdqa %xmm13, %xmm01219; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,0],xmm3[2,0]1220; SSE-NEXT: shufps {{.*#+}} xmm2 = xmm2[0,1],xmm0[2,0]1221; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1222; SSE-NEXT: movdqa %xmm7, %xmm11223; SSE-NEXT: movdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1224; SSE-NEXT: movdqa %xmm7, %xmm01225; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm14[0],xmm0[1],xmm14[1]1226; SSE-NEXT: movdqa 16(%rdi), %xmm71227; SSE-NEXT: movdqa %xmm7, %xmm81228; SSE-NEXT: movdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1229; SSE-NEXT: punpckldq {{.*#+}} xmm8 = xmm8[0],xmm6[0],xmm8[1],xmm6[1]1230; SSE-NEXT: punpcklqdq {{.*#+}} xmm8 = xmm8[0],xmm0[0]1231; SSE-NEXT: movdqa %xmm7, %xmm01232; SSE-NEXT: punpckhdq {{.*#+}} xmm0 = xmm0[2],xmm6[2],xmm0[3],xmm6[3]1233; SSE-NEXT: movaps %xmm14, %xmm101234; SSE-NEXT: movaps %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1235; SSE-NEXT: shufps {{.*#+}} xmm10 = xmm10[1,1],xmm15[1,1]1236; SSE-NEXT: shufps {{.*#+}} xmm10 = xmm10[0,2],xmm0[0,1]1237; SSE-NEXT: movdqa %xmm7, %xmm01238; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,1],xmm15[2,3]1239; SSE-NEXT: punpckhqdq {{.*#+}} xmm14 = xmm14[1],xmm1[1]1240; SSE-NEXT: shufps {{.*#+}} xmm14 = xmm14[2,0],xmm0[2,0]1241; SSE-NEXT: movdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1242; SSE-NEXT: movdqa %xmm4, %xmm01243; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm12[0],xmm0[1],xmm12[1]1244; SSE-NEXT: movdqa 32(%rdi), %xmm31245; SSE-NEXT: movdqa %xmm3, %xmm61246; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload1247; SSE-NEXT: punpckldq {{.*#+}} xmm6 = xmm6[0],xmm2[0],xmm6[1],xmm2[1]1248; SSE-NEXT: punpcklqdq {{.*#+}} xmm6 = xmm6[0],xmm0[0]1249; SSE-NEXT: movdqa %xmm3, %xmm01250; SSE-NEXT: punpckhdq {{.*#+}} xmm0 = xmm0[2],xmm2[2],xmm0[3],xmm2[3]1251; SSE-NEXT: movdqa %xmm4, %xmm151252; SSE-NEXT: punpckhdq {{.*#+}} xmm15 = xmm15[2],xmm12[2],xmm15[3],xmm12[3]1253; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload1254; SSE-NEXT: shufps {{.*#+}} xmm12 = xmm12[1,1],xmm1[1,1]1255; SSE-NEXT: shufps {{.*#+}} xmm12 = xmm12[0,2],xmm0[0,1]1256; SSE-NEXT: movdqa %xmm3, %xmm01257; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,0],xmm1[2,0]1258; SSE-NEXT: shufps {{.*#+}} xmm15 = xmm15[0,1],xmm0[2,0]1259; SSE-NEXT: movdqa %xmm11, %xmm41260; SSE-NEXT: punpckldq {{.*#+}} xmm4 = xmm4[0],xmm9[0],xmm4[1],xmm9[1]1261; SSE-NEXT: movdqa 48(%rdi), %xmm21262; SSE-NEXT: movdqa %xmm2, %xmm51263; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload1264; SSE-NEXT: punpckldq {{.*#+}} xmm5 = xmm5[0],xmm1[0],xmm5[1],xmm1[1]1265; SSE-NEXT: punpcklqdq {{.*#+}} xmm5 = xmm5[0],xmm4[0]1266; SSE-NEXT: movdqa %xmm2, %xmm01267; SSE-NEXT: punpckhdq {{.*#+}} xmm0 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]1268; SSE-NEXT: movaps %xmm9, %xmm41269; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload1270; SSE-NEXT: shufps {{.*#+}} xmm4 = xmm4[1,1],xmm1[1,1]1271; SSE-NEXT: shufps {{.*#+}} xmm4 = xmm4[0,2],xmm0[0,1]1272; SSE-NEXT: movdqa %xmm2, %xmm01273; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,1],xmm1[2,3]1274; SSE-NEXT: punpckhqdq {{.*#+}} xmm9 = xmm9[1],xmm11[1]1275; SSE-NEXT: shufps {{.*#+}} xmm9 = xmm9[2,0],xmm0[2,0]1276; SSE-NEXT: movaps (%rsp), %xmm0 # 16-byte Reload1277; SSE-NEXT: shufps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload1278; SSE-NEXT: # xmm0 = xmm0[1,1],mem[1,1]1279; SSE-NEXT: shufps {{.*#+}} xmm13 = xmm13[0,1],xmm0[2,0]1280; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload1281; SSE-NEXT: movss {{.*#+}} xmm13 = xmm0[0],xmm13[1,2,3]1282; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload1283; SSE-NEXT: shufps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload1284; SSE-NEXT: # xmm0 = xmm0[1,1],mem[1,1]1285; SSE-NEXT: shufps {{.*#+}} xmm7 = xmm7[0,1],xmm0[2,0]1286; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload1287; SSE-NEXT: movss {{.*#+}} xmm7 = xmm0[0],xmm7[1,2,3]1288; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload1289; SSE-NEXT: shufps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload1290; SSE-NEXT: # xmm0 = xmm0[1,1],mem[1,1]1291; SSE-NEXT: shufps {{.*#+}} xmm3 = xmm3[0,1],xmm0[2,0]1292; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload1293; SSE-NEXT: movss {{.*#+}} xmm3 = xmm0[0],xmm3[1,2,3]1294; SSE-NEXT: shufps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm11 # 16-byte Folded Reload1295; SSE-NEXT: # xmm11 = xmm11[1,1],mem[1,1]1296; SSE-NEXT: shufps {{.*#+}} xmm2 = xmm2[0,1],xmm11[2,0]1297; SSE-NEXT: movss {{.*#+}} xmm2 = xmm1[0],xmm2[1,2,3]1298; SSE-NEXT: movaps %xmm9, 288(%r9)1299; SSE-NEXT: movaps %xmm4, 272(%r9)1300; SSE-NEXT: movdqa %xmm5, 240(%r9)1301; SSE-NEXT: movaps %xmm15, 208(%r9)1302; SSE-NEXT: movaps %xmm12, 192(%r9)1303; SSE-NEXT: movdqa %xmm6, 160(%r9)1304; SSE-NEXT: movaps %xmm14, 128(%r9)1305; SSE-NEXT: movaps %xmm10, 112(%r9)1306; SSE-NEXT: movdqa %xmm8, 80(%r9)1307; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload1308; SSE-NEXT: movaps %xmm0, 48(%r9)1309; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload1310; SSE-NEXT: movaps %xmm0, 32(%r9)1311; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload1312; SSE-NEXT: movaps %xmm0, (%r9)1313; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload1314; SSE-NEXT: movaps %xmm0, 304(%r9)1315; SSE-NEXT: movaps %xmm2, 256(%r9)1316; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload1317; SSE-NEXT: movaps %xmm0, 224(%r9)1318; SSE-NEXT: movaps %xmm3, 176(%r9)1319; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload1320; SSE-NEXT: movaps %xmm0, 144(%r9)1321; SSE-NEXT: movaps %xmm7, 96(%r9)1322; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload1323; SSE-NEXT: movaps %xmm0, 64(%r9)1324; SSE-NEXT: movaps %xmm13, 16(%r9)1325; SSE-NEXT: addq $168, %rsp1326; SSE-NEXT: retq1327;1328; AVX-LABEL: store_i32_stride5_vf16:1329; AVX: # %bb.0:1330; AVX-NEXT: vmovaps (%rdi), %ymm01331; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill1332; AVX-NEXT: vmovaps 32(%rdi), %ymm21333; AVX-NEXT: vmovaps (%rsi), %ymm01334; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill1335; AVX-NEXT: vmovaps 32(%rsi), %ymm41336; AVX-NEXT: vmovaps 32(%rdx), %ymm11337; AVX-NEXT: vmovaps 32(%rcx), %ymm31338; AVX-NEXT: vmovaps (%rcx), %xmm61339; AVX-NEXT: vmovaps 16(%rcx), %xmm81340; AVX-NEXT: vmovaps 32(%rcx), %xmm111341; AVX-NEXT: vmovaps (%rdx), %xmm71342; AVX-NEXT: vmovaps 16(%rdx), %xmm91343; AVX-NEXT: vmovaps 32(%rdx), %xmm121344; AVX-NEXT: vunpckhps {{.*#+}} xmm10 = xmm7[2],xmm6[2],xmm7[3],xmm6[3]1345; AVX-NEXT: vshufps {{.*#+}} xmm13 = xmm6[1,1,1,1]1346; AVX-NEXT: vinsertf128 $1, %xmm10, %ymm13, %ymm101347; AVX-NEXT: vmovaps (%rdi), %xmm131348; AVX-NEXT: vmovaps (%rsi), %xmm141349; AVX-NEXT: vshufps {{.*#+}} xmm15 = xmm14[2,2,2,2]1350; AVX-NEXT: vblendps {{.*#+}} xmm15 = xmm13[0,1,2],xmm15[3]1351; AVX-NEXT: vinsertf128 $1, %xmm13, %ymm15, %ymm151352; AVX-NEXT: vblendps {{.*#+}} ymm10 = ymm10[0,1],ymm15[2,3],ymm10[4,5],ymm15[6,7]1353; AVX-NEXT: vinsertps {{.*#+}} xmm15 = zero,xmm13[1],xmm14[1],zero1354; AVX-NEXT: vunpcklps {{.*#+}} xmm13 = xmm13[0],xmm14[0],xmm13[1],xmm14[1]1355; AVX-NEXT: vmovaps 32(%rdi), %xmm141356; AVX-NEXT: vinsertf128 $1, %xmm15, %ymm13, %ymm131357; AVX-NEXT: vmovaps 32(%rsi), %xmm151358; AVX-NEXT: vunpcklps {{.*#+}} xmm6 = xmm7[0],xmm6[0],xmm7[1],xmm6[1]1359; AVX-NEXT: vinsertf128 $1, %xmm7, %ymm6, %ymm71360; AVX-NEXT: vbroadcastf128 {{.*#+}} ymm0 = mem[0,1,0,1]1361; AVX-NEXT: vblendps {{.*#+}} ymm5 = ymm10[0],ymm0[1],ymm10[2,3,4,5],ymm0[6],ymm10[7]1362; AVX-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill1363; AVX-NEXT: vmovddup {{.*#+}} ymm7 = ymm7[0,0,2,2]1364; AVX-NEXT: vblendps {{.*#+}} ymm7 = ymm13[0,1],ymm7[2,3],ymm13[4,5,6],ymm7[7]1365; AVX-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0,1,2,3],ymm0[4],ymm7[5,6,7]1366; AVX-NEXT: vunpckhps {{.*#+}} xmm0 = xmm12[2],xmm11[2],xmm12[3],xmm11[3]1367; AVX-NEXT: vshufps {{.*#+}} xmm10 = xmm11[1,1,1,1]1368; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm10, %ymm01369; AVX-NEXT: vshufps {{.*#+}} xmm10 = xmm15[2,2,2,2]1370; AVX-NEXT: vblendps {{.*#+}} xmm10 = xmm14[0,1,2],xmm10[3]1371; AVX-NEXT: vinsertf128 $1, %xmm14, %ymm10, %ymm101372; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm10[2,3],ymm0[4,5],ymm10[6,7]1373; AVX-NEXT: vinsertps {{.*#+}} xmm13 = zero,xmm14[1],xmm15[1],zero1374; AVX-NEXT: vunpcklps {{.*#+}} xmm14 = xmm14[0],xmm15[0],xmm14[1],xmm15[1]1375; AVX-NEXT: vbroadcastf128 {{.*#+}} ymm15 = mem[0,1,0,1]1376; AVX-NEXT: vblendps {{.*#+}} ymm10 = ymm0[0],ymm15[1],ymm0[2,3,4,5],ymm15[6],ymm0[7]1377; AVX-NEXT: vinsertf128 $1, %xmm13, %ymm14, %ymm01378; AVX-NEXT: vunpcklps {{.*#+}} xmm11 = xmm12[0],xmm11[0],xmm12[1],xmm11[1]1379; AVX-NEXT: vinsertf128 $1, %xmm12, %ymm11, %ymm111380; AVX-NEXT: vmovddup {{.*#+}} ymm11 = ymm11[0,0,2,2]1381; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm11[2,3],ymm0[4,5,6],ymm11[7]1382; AVX-NEXT: vblendps {{.*#+}} ymm11 = ymm0[0,1,2,3],ymm15[4],ymm0[5,6,7]1383; AVX-NEXT: vmovaps 48(%rdi), %xmm01384; AVX-NEXT: vmovaps 48(%rsi), %xmm121385; AVX-NEXT: vshufps {{.*#+}} xmm13 = xmm12[2,2,2,2]1386; AVX-NEXT: vblendps {{.*#+}} xmm13 = xmm0[0,1,2],xmm13[3]1387; AVX-NEXT: vinsertps {{.*#+}} xmm0 = zero,xmm0[1],xmm12[1],zero1388; AVX-NEXT: vinsertf128 $1, %xmm13, %ymm0, %ymm01389; AVX-NEXT: vshufps {{.*#+}} ymm12 = ymm3[1,1],ymm1[1,1],ymm3[5,5],ymm1[5,5]1390; AVX-NEXT: vmovaps %ymm1, %ymm61391; AVX-NEXT: vperm2f128 {{.*#+}} ymm12 = ymm12[2,3,2,3]1392; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm12[3,4],ymm0[5,6,7]1393; AVX-NEXT: vmovaps 48(%rcx), %xmm121394; AVX-NEXT: vmovaps 48(%rdx), %xmm131395; AVX-NEXT: vshufps {{.*#+}} xmm14 = xmm13[3,3],xmm12[3,3]1396; AVX-NEXT: vunpckhps {{.*#+}} xmm12 = xmm13[2],xmm12[2],xmm13[3],xmm12[3]1397; AVX-NEXT: vmovaps 16(%rdi), %xmm151398; AVX-NEXT: vinsertf128 $1, %xmm14, %ymm12, %ymm121399; AVX-NEXT: vshufps {{.*#+}} ymm13 = ymm4[3,3],ymm2[3,3],ymm4[7,7],ymm2[7,7]1400; AVX-NEXT: vperm2f128 {{.*#+}} ymm13 = ymm13[2,3,2,3]1401; AVX-NEXT: vblendps {{.*#+}} ymm13 = ymm12[0,1,2],ymm13[3,4],ymm12[5,6,7]1402; AVX-NEXT: vbroadcastf128 {{.*#+}} ymm14 = mem[0,1,0,1]1403; AVX-NEXT: vblendps {{.*#+}} ymm12 = ymm14[0],ymm0[1,2,3,4],ymm14[5],ymm0[6,7]1404; AVX-NEXT: vblendps {{.*#+}} ymm13 = ymm13[0,1],ymm14[2],ymm13[3,4,5,6],ymm14[7]1405; AVX-NEXT: vshufps {{.*#+}} xmm0 = xmm9[3,3],xmm8[3,3]1406; AVX-NEXT: vunpckhps {{.*#+}} xmm8 = xmm9[2],xmm8[2],xmm9[3],xmm8[3]1407; AVX-NEXT: vmovaps 16(%rsi), %xmm91408; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm8, %ymm01409; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload1410; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload1411; AVX-NEXT: vshufps {{.*#+}} ymm8 = ymm1[3,3],ymm5[3,3],ymm1[7,7],ymm5[7,7]1412; AVX-NEXT: vperm2f128 {{.*#+}} ymm8 = ymm8[2,3,2,3]1413; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm8[3,4],ymm0[5,6,7]1414; AVX-NEXT: vshufps {{.*#+}} xmm8 = xmm9[2,2,2,2]1415; AVX-NEXT: vblendps {{.*#+}} xmm8 = xmm15[0,1,2],xmm8[3]1416; AVX-NEXT: vinsertps {{.*#+}} xmm9 = zero,xmm15[1],xmm9[1],zero1417; AVX-NEXT: vmovaps (%rdx), %ymm141418; AVX-NEXT: vinsertf128 $1, %xmm8, %ymm9, %ymm81419; AVX-NEXT: vmovaps (%rcx), %ymm91420; AVX-NEXT: vshufps {{.*#+}} ymm15 = ymm9[1,1],ymm14[1,1],ymm9[5,5],ymm14[5,5]1421; AVX-NEXT: vperm2f128 {{.*#+}} ymm15 = ymm15[2,3,2,3]1422; AVX-NEXT: vblendps {{.*#+}} ymm8 = ymm8[0,1,2],ymm15[3,4],ymm8[5,6,7]1423; AVX-NEXT: vbroadcastf128 {{.*#+}} ymm15 = mem[0,1,0,1]1424; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm15[2],ymm0[3,4,5,6],ymm15[7]1425; AVX-NEXT: vblendps {{.*#+}} ymm8 = ymm15[0],ymm8[1,2,3,4],ymm15[5],ymm8[6,7]1426; AVX-NEXT: vshufps {{.*#+}} ymm4 = ymm4[3,0,2,3,7,4,6,7]1427; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm4[0,1,2,3],ymm2[4],ymm4[5,6,7]1428; AVX-NEXT: vshufps {{.*#+}} ymm4 = ymm3[0,1,3,0,4,5,7,4]1429; AVX-NEXT: vshufpd {{.*#+}} ymm3 = ymm6[1,0,2,2]1430; AVX-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1],ymm4[2],ymm3[3,4,5,6],ymm4[7]1431; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0],ymm3[1,2,3],ymm2[4,5],ymm3[6,7]1432; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm1[3,0,2,3,7,4,6,7]1433; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm5[4],ymm1[5,6,7]1434; AVX-NEXT: vshufps {{.*#+}} ymm3 = ymm9[0,1,3,0,4,5,7,4]1435; AVX-NEXT: vshufpd {{.*#+}} ymm4 = ymm14[1,0,2,2]1436; AVX-NEXT: vblendps {{.*#+}} ymm3 = ymm4[0,1],ymm3[2],ymm4[3,4,5,6],ymm3[7]1437; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm3[1,2,3],ymm1[4,5],ymm3[6,7]1438; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2],mem[3],ymm2[4,5,6,7]1439; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2],mem[3],ymm1[4,5,6,7]1440; AVX-NEXT: vmovaps %ymm1, 64(%r9)1441; AVX-NEXT: vmovaps %ymm2, 224(%r9)1442; AVX-NEXT: vmovaps %ymm8, 96(%r9)1443; AVX-NEXT: vmovaps %ymm0, 128(%r9)1444; AVX-NEXT: vmovaps %ymm13, 288(%r9)1445; AVX-NEXT: vmovaps %ymm12, 256(%r9)1446; AVX-NEXT: vmovaps %ymm11, 160(%r9)1447; AVX-NEXT: vmovaps %ymm10, 192(%r9)1448; AVX-NEXT: vmovaps %ymm7, (%r9)1449; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload1450; AVX-NEXT: vmovaps %ymm0, 32(%r9)1451; AVX-NEXT: vzeroupper1452; AVX-NEXT: retq1453;1454; AVX2-LABEL: store_i32_stride5_vf16:1455; AVX2: # %bb.0:1456; AVX2-NEXT: vmovaps 32(%rdi), %ymm31457; AVX2-NEXT: vmovaps 32(%rsi), %ymm51458; AVX2-NEXT: vmovaps (%rdx), %ymm11459; AVX2-NEXT: vmovaps 32(%rdx), %ymm41460; AVX2-NEXT: vmovaps (%rcx), %ymm21461; AVX2-NEXT: vmovaps 32(%rcx), %ymm71462; AVX2-NEXT: vmovaps (%r8), %ymm01463; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill1464; AVX2-NEXT: vmovaps 32(%r8), %ymm61465; AVX2-NEXT: vmovaps (%rsi), %xmm91466; AVX2-NEXT: vmovaps 32(%rsi), %xmm111467; AVX2-NEXT: vmovaps (%rdi), %xmm101468; AVX2-NEXT: vmovaps 32(%rdi), %xmm121469; AVX2-NEXT: vunpckhps {{.*#+}} xmm8 = xmm10[2],xmm9[2],xmm10[3],xmm9[3]1470; AVX2-NEXT: vshufps {{.*#+}} xmm8 = xmm8[0,1,2,2]1471; AVX2-NEXT: vpermpd {{.*#+}} ymm8 = ymm8[0,0,2,1]1472; AVX2-NEXT: vmovaps (%rdx), %xmm131473; AVX2-NEXT: vmovaps (%rcx), %xmm141474; AVX2-NEXT: vshufps {{.*#+}} xmm15 = xmm14[1,1,2,2]1475; AVX2-NEXT: vblendps {{.*#+}} xmm15 = xmm15[0,1],xmm13[2],xmm15[3]1476; AVX2-NEXT: vpermpd {{.*#+}} ymm15 = ymm15[0,1,1,3]1477; AVX2-NEXT: vblendps {{.*#+}} ymm8 = ymm15[0,1],ymm8[2,3],ymm15[4,5],ymm8[6,7]1478; AVX2-NEXT: vpermpd {{.*#+}} ymm15 = ymm0[0,1,2,1]1479; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm8[0],ymm15[1],ymm8[2,3,4,5],ymm15[6],ymm8[7]1480; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill1481; AVX2-NEXT: vmovaps 32(%rdx), %xmm151482; AVX2-NEXT: vunpcklps {{.*#+}} xmm13 = xmm13[0],xmm14[0],xmm13[1],xmm14[1]1483; AVX2-NEXT: vmovaps 32(%rcx), %xmm141484; AVX2-NEXT: vshufps {{.*#+}} xmm13 = xmm13[0,1,2,2]1485; AVX2-NEXT: vpermpd {{.*#+}} ymm13 = ymm13[0,0,2,1]1486; AVX2-NEXT: vunpcklps {{.*#+}} xmm9 = xmm10[0],xmm9[0],xmm10[1],xmm9[1]1487; AVX2-NEXT: vshufps {{.*#+}} xmm9 = xmm9[0,1,3,2]1488; AVX2-NEXT: vpermpd {{.*#+}} ymm10 = ymm9[0,1,1,1]1489; AVX2-NEXT: vblendps {{.*#+}} ymm10 = ymm10[0,1],ymm13[2,3],ymm10[4,5,6],ymm13[7]1490; AVX2-NEXT: vinsertf128 $1, (%r8), %ymm9, %ymm91491; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm9[0],ymm10[1,2,3],ymm9[4],ymm10[5,6,7]1492; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill1493; AVX2-NEXT: vunpckhps {{.*#+}} xmm10 = xmm12[2],xmm11[2],xmm12[3],xmm11[3]1494; AVX2-NEXT: vshufps {{.*#+}} xmm10 = xmm10[0,1,2,2]1495; AVX2-NEXT: vpermpd {{.*#+}} ymm10 = ymm10[0,0,2,1]1496; AVX2-NEXT: vshufps {{.*#+}} xmm13 = xmm14[1,1,2,2]1497; AVX2-NEXT: vblendps {{.*#+}} xmm13 = xmm13[0,1],xmm15[2],xmm13[3]1498; AVX2-NEXT: vpermpd {{.*#+}} ymm13 = ymm13[0,1,1,3]1499; AVX2-NEXT: vblendps {{.*#+}} ymm10 = ymm13[0,1],ymm10[2,3],ymm13[4,5],ymm10[6,7]1500; AVX2-NEXT: vpermpd {{.*#+}} ymm13 = ymm6[0,1,2,1]1501; AVX2-NEXT: vblendps {{.*#+}} ymm10 = ymm10[0],ymm13[1],ymm10[2,3,4,5],ymm13[6],ymm10[7]1502; AVX2-NEXT: vunpcklps {{.*#+}} xmm13 = xmm15[0],xmm14[0],xmm15[1],xmm14[1]1503; AVX2-NEXT: vshufps {{.*#+}} xmm13 = xmm13[0,1,2,2]1504; AVX2-NEXT: vpermpd {{.*#+}} ymm13 = ymm13[0,0,2,1]1505; AVX2-NEXT: vunpcklps {{.*#+}} xmm11 = xmm12[0],xmm11[0],xmm12[1],xmm11[1]1506; AVX2-NEXT: vshufps {{.*#+}} xmm11 = xmm11[0,1,3,2]1507; AVX2-NEXT: vpermpd {{.*#+}} ymm12 = ymm11[0,1,1,1]1508; AVX2-NEXT: vblendps {{.*#+}} ymm12 = ymm12[0,1],ymm13[2,3],ymm12[4,5,6],ymm13[7]1509; AVX2-NEXT: vinsertf128 $1, 32(%r8), %ymm11, %ymm111510; AVX2-NEXT: vblendps {{.*#+}} ymm11 = ymm11[0],ymm12[1,2,3],ymm11[4],ymm12[5,6,7]1511; AVX2-NEXT: vshufps {{.*#+}} ymm12 = ymm7[1,1,1,1,5,5,5,5]1512; AVX2-NEXT: vblendps {{.*#+}} ymm12 = ymm12[0],ymm4[1],ymm12[2,3,4],ymm4[5],ymm12[6,7]1513; AVX2-NEXT: vpermpd {{.*#+}} ymm12 = ymm12[2,2,2,2]1514; AVX2-NEXT: vshufps {{.*#+}} ymm13 = ymm5[1,1,2,2,5,5,6,6]1515; AVX2-NEXT: vblendps {{.*#+}} ymm13 = ymm13[0],ymm3[1,2],ymm13[3,4],ymm3[5,6],ymm13[7]1516; AVX2-NEXT: vpermpd {{.*#+}} ymm13 = ymm13[2,2,2,3]1517; AVX2-NEXT: vblendps {{.*#+}} ymm12 = ymm13[0,1,2],ymm12[3,4],ymm13[5,6,7]1518; AVX2-NEXT: vbroadcastsd 48(%r8), %ymm131519; AVX2-NEXT: vblendps {{.*#+}} ymm12 = ymm13[0],ymm12[1,2,3,4],ymm13[5],ymm12[6,7]1520; AVX2-NEXT: vunpckhps {{.*#+}} ymm13 = ymm5[2],ymm3[2],ymm5[3],ymm3[3],ymm5[6],ymm3[6],ymm5[7],ymm3[7]1521; AVX2-NEXT: vpermpd {{.*#+}} ymm13 = ymm13[3,3,3,3]1522; AVX2-NEXT: vshufps {{.*#+}} ymm14 = ymm7[0,2,3,3,4,6,7,7]1523; AVX2-NEXT: vshufps {{.*#+}} ymm15 = ymm4[2,3,2,3,6,7,6,7]1524; AVX2-NEXT: vblendps {{.*#+}} ymm14 = ymm15[0],ymm14[1,2],ymm15[3,4],ymm14[5,6],ymm15[7]1525; AVX2-NEXT: vpermpd {{.*#+}} ymm14 = ymm14[2,1,3,3]1526; AVX2-NEXT: vblendps {{.*#+}} ymm13 = ymm14[0,1,2],ymm13[3,4],ymm14[5,6,7]1527; AVX2-NEXT: vbroadcastsd 56(%r8), %ymm141528; AVX2-NEXT: vblendps {{.*#+}} ymm13 = ymm13[0,1],ymm14[2],ymm13[3,4,5,6],ymm14[7]1529; AVX2-NEXT: vshufps {{.*#+}} ymm14 = ymm2[0,2,3,3,4,6,7,7]1530; AVX2-NEXT: vshufps {{.*#+}} ymm15 = ymm1[2,3,2,3,6,7,6,7]1531; AVX2-NEXT: vblendps {{.*#+}} ymm14 = ymm15[0],ymm14[1,2],ymm15[3,4],ymm14[5,6],ymm15[7]1532; AVX2-NEXT: vmovaps (%rdi), %ymm151533; AVX2-NEXT: vmovaps (%rsi), %ymm01534; AVX2-NEXT: vunpckhps {{.*#+}} ymm8 = ymm0[2],ymm15[2],ymm0[3],ymm15[3],ymm0[6],ymm15[6],ymm0[7],ymm15[7]1535; AVX2-NEXT: vpermpd {{.*#+}} ymm8 = ymm8[3,3,3,3]1536; AVX2-NEXT: vpermpd {{.*#+}} ymm14 = ymm14[2,1,3,3]1537; AVX2-NEXT: vblendps {{.*#+}} ymm8 = ymm14[0,1,2],ymm8[3,4],ymm14[5,6,7]1538; AVX2-NEXT: vbroadcastsd 24(%r8), %ymm141539; AVX2-NEXT: vblendps {{.*#+}} ymm8 = ymm8[0,1],ymm14[2],ymm8[3,4,5,6],ymm14[7]1540; AVX2-NEXT: vshufps {{.*#+}} ymm14 = ymm2[1,1,1,1,5,5,5,5]1541; AVX2-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0],ymm1[1],ymm14[2,3,4],ymm1[5],ymm14[6,7]1542; AVX2-NEXT: vpermpd {{.*#+}} ymm14 = ymm14[2,2,2,2]1543; AVX2-NEXT: vshufps {{.*#+}} ymm9 = ymm0[1,1,2,2,5,5,6,6]1544; AVX2-NEXT: vblendps {{.*#+}} ymm9 = ymm9[0],ymm15[1,2],ymm9[3,4],ymm15[5,6],ymm9[7]1545; AVX2-NEXT: vpermpd {{.*#+}} ymm9 = ymm9[2,2,2,3]1546; AVX2-NEXT: vblendps {{.*#+}} ymm9 = ymm9[0,1,2],ymm14[3,4],ymm9[5,6,7]1547; AVX2-NEXT: vbroadcastsd 16(%r8), %ymm141548; AVX2-NEXT: vblendps {{.*#+}} ymm9 = ymm14[0],ymm9[1,2,3,4],ymm14[5],ymm9[6,7]1549; AVX2-NEXT: vshufps {{.*#+}} ymm5 = ymm5[3,0,2,3,7,4,6,7]1550; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm5[0,1,2,3],ymm3[4],ymm5[5,6,7]1551; AVX2-NEXT: vshufps {{.*#+}} ymm5 = ymm7[0,1,3,0,4,5,7,4]1552; AVX2-NEXT: vshufps {{.*#+}} ymm4 = ymm4[2,3,0,1,6,7,4,5]1553; AVX2-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1],ymm5[2],ymm4[3,4,5,6],ymm5[7]1554; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0],ymm4[1,2,3],ymm3[4,5],ymm4[6,7]1555; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2],ymm6[3],ymm3[4,5,6,7]1556; AVX2-NEXT: vshufps {{.*#+}} ymm0 = ymm0[3,0,2,3,7,4,6,7]1557; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm15[4],ymm0[5,6,7]1558; AVX2-NEXT: vshufps {{.*#+}} ymm2 = ymm2[0,1,3,0,4,5,7,4]1559; AVX2-NEXT: vshufps {{.*#+}} ymm1 = ymm1[2,3,0,1,6,7,4,5]1560; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4,5,6],ymm2[7]1561; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm1[1,2,3],ymm0[4,5],ymm1[6,7]1562; AVX2-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload1563; AVX2-NEXT: # ymm0 = ymm0[0,1,2],mem[3],ymm0[4,5,6,7]1564; AVX2-NEXT: vmovaps %ymm0, 64(%r9)1565; AVX2-NEXT: vmovaps %ymm3, 224(%r9)1566; AVX2-NEXT: vmovaps %ymm9, 96(%r9)1567; AVX2-NEXT: vmovaps %ymm8, 128(%r9)1568; AVX2-NEXT: vmovaps %ymm13, 288(%r9)1569; AVX2-NEXT: vmovaps %ymm12, 256(%r9)1570; AVX2-NEXT: vmovaps %ymm11, 160(%r9)1571; AVX2-NEXT: vmovaps %ymm10, 192(%r9)1572; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload1573; AVX2-NEXT: vmovaps %ymm0, (%r9)1574; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload1575; AVX2-NEXT: vmovaps %ymm0, 32(%r9)1576; AVX2-NEXT: vzeroupper1577; AVX2-NEXT: retq1578;1579; AVX2-FP-LABEL: store_i32_stride5_vf16:1580; AVX2-FP: # %bb.0:1581; AVX2-FP-NEXT: vmovaps 32(%rdi), %ymm31582; AVX2-FP-NEXT: vmovaps 32(%rsi), %ymm51583; AVX2-FP-NEXT: vmovaps (%rdx), %ymm11584; AVX2-FP-NEXT: vmovaps 32(%rdx), %ymm41585; AVX2-FP-NEXT: vmovaps (%rcx), %ymm21586; AVX2-FP-NEXT: vmovaps 32(%rcx), %ymm71587; AVX2-FP-NEXT: vmovaps (%r8), %ymm01588; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill1589; AVX2-FP-NEXT: vmovaps 32(%r8), %ymm61590; AVX2-FP-NEXT: vmovaps (%rsi), %xmm91591; AVX2-FP-NEXT: vmovaps 32(%rsi), %xmm111592; AVX2-FP-NEXT: vmovaps (%rdi), %xmm101593; AVX2-FP-NEXT: vmovaps 32(%rdi), %xmm121594; AVX2-FP-NEXT: vunpckhps {{.*#+}} xmm8 = xmm10[2],xmm9[2],xmm10[3],xmm9[3]1595; AVX2-FP-NEXT: vshufps {{.*#+}} xmm8 = xmm8[0,1,2,2]1596; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm8 = ymm8[0,0,2,1]1597; AVX2-FP-NEXT: vmovaps (%rdx), %xmm131598; AVX2-FP-NEXT: vmovaps (%rcx), %xmm141599; AVX2-FP-NEXT: vshufps {{.*#+}} xmm15 = xmm14[1,1,2,2]1600; AVX2-FP-NEXT: vblendps {{.*#+}} xmm15 = xmm15[0,1],xmm13[2],xmm15[3]1601; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm15 = ymm15[0,1,1,3]1602; AVX2-FP-NEXT: vblendps {{.*#+}} ymm8 = ymm15[0,1],ymm8[2,3],ymm15[4,5],ymm8[6,7]1603; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm15 = ymm0[0,1,2,1]1604; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm8[0],ymm15[1],ymm8[2,3,4,5],ymm15[6],ymm8[7]1605; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill1606; AVX2-FP-NEXT: vmovaps 32(%rdx), %xmm151607; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm13 = xmm13[0],xmm14[0],xmm13[1],xmm14[1]1608; AVX2-FP-NEXT: vmovaps 32(%rcx), %xmm141609; AVX2-FP-NEXT: vshufps {{.*#+}} xmm13 = xmm13[0,1,2,2]1610; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm13 = ymm13[0,0,2,1]1611; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm9 = xmm10[0],xmm9[0],xmm10[1],xmm9[1]1612; AVX2-FP-NEXT: vshufps {{.*#+}} xmm9 = xmm9[0,1,3,2]1613; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm10 = ymm9[0,1,1,1]1614; AVX2-FP-NEXT: vblendps {{.*#+}} ymm10 = ymm10[0,1],ymm13[2,3],ymm10[4,5,6],ymm13[7]1615; AVX2-FP-NEXT: vinsertf128 $1, (%r8), %ymm9, %ymm91616; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm9[0],ymm10[1,2,3],ymm9[4],ymm10[5,6,7]1617; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill1618; AVX2-FP-NEXT: vunpckhps {{.*#+}} xmm10 = xmm12[2],xmm11[2],xmm12[3],xmm11[3]1619; AVX2-FP-NEXT: vshufps {{.*#+}} xmm10 = xmm10[0,1,2,2]1620; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm10 = ymm10[0,0,2,1]1621; AVX2-FP-NEXT: vshufps {{.*#+}} xmm13 = xmm14[1,1,2,2]1622; AVX2-FP-NEXT: vblendps {{.*#+}} xmm13 = xmm13[0,1],xmm15[2],xmm13[3]1623; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm13 = ymm13[0,1,1,3]1624; AVX2-FP-NEXT: vblendps {{.*#+}} ymm10 = ymm13[0,1],ymm10[2,3],ymm13[4,5],ymm10[6,7]1625; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm13 = ymm6[0,1,2,1]1626; AVX2-FP-NEXT: vblendps {{.*#+}} ymm10 = ymm10[0],ymm13[1],ymm10[2,3,4,5],ymm13[6],ymm10[7]1627; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm13 = xmm15[0],xmm14[0],xmm15[1],xmm14[1]1628; AVX2-FP-NEXT: vshufps {{.*#+}} xmm13 = xmm13[0,1,2,2]1629; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm13 = ymm13[0,0,2,1]1630; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm11 = xmm12[0],xmm11[0],xmm12[1],xmm11[1]1631; AVX2-FP-NEXT: vshufps {{.*#+}} xmm11 = xmm11[0,1,3,2]1632; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm12 = ymm11[0,1,1,1]1633; AVX2-FP-NEXT: vblendps {{.*#+}} ymm12 = ymm12[0,1],ymm13[2,3],ymm12[4,5,6],ymm13[7]1634; AVX2-FP-NEXT: vinsertf128 $1, 32(%r8), %ymm11, %ymm111635; AVX2-FP-NEXT: vblendps {{.*#+}} ymm11 = ymm11[0],ymm12[1,2,3],ymm11[4],ymm12[5,6,7]1636; AVX2-FP-NEXT: vshufps {{.*#+}} ymm12 = ymm7[1,1,1,1,5,5,5,5]1637; AVX2-FP-NEXT: vblendps {{.*#+}} ymm12 = ymm12[0],ymm4[1],ymm12[2,3,4],ymm4[5],ymm12[6,7]1638; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm12 = ymm12[2,2,2,2]1639; AVX2-FP-NEXT: vshufps {{.*#+}} ymm13 = ymm5[1,1,2,2,5,5,6,6]1640; AVX2-FP-NEXT: vblendps {{.*#+}} ymm13 = ymm13[0],ymm3[1,2],ymm13[3,4],ymm3[5,6],ymm13[7]1641; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm13 = ymm13[2,2,2,3]1642; AVX2-FP-NEXT: vblendps {{.*#+}} ymm12 = ymm13[0,1,2],ymm12[3,4],ymm13[5,6,7]1643; AVX2-FP-NEXT: vbroadcastsd 48(%r8), %ymm131644; AVX2-FP-NEXT: vblendps {{.*#+}} ymm12 = ymm13[0],ymm12[1,2,3,4],ymm13[5],ymm12[6,7]1645; AVX2-FP-NEXT: vunpckhps {{.*#+}} ymm13 = ymm5[2],ymm3[2],ymm5[3],ymm3[3],ymm5[6],ymm3[6],ymm5[7],ymm3[7]1646; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm13 = ymm13[3,3,3,3]1647; AVX2-FP-NEXT: vshufps {{.*#+}} ymm14 = ymm7[0,2,3,3,4,6,7,7]1648; AVX2-FP-NEXT: vshufps {{.*#+}} ymm15 = ymm4[2,3,2,3,6,7,6,7]1649; AVX2-FP-NEXT: vblendps {{.*#+}} ymm14 = ymm15[0],ymm14[1,2],ymm15[3,4],ymm14[5,6],ymm15[7]1650; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm14 = ymm14[2,1,3,3]1651; AVX2-FP-NEXT: vblendps {{.*#+}} ymm13 = ymm14[0,1,2],ymm13[3,4],ymm14[5,6,7]1652; AVX2-FP-NEXT: vbroadcastsd 56(%r8), %ymm141653; AVX2-FP-NEXT: vblendps {{.*#+}} ymm13 = ymm13[0,1],ymm14[2],ymm13[3,4,5,6],ymm14[7]1654; AVX2-FP-NEXT: vshufps {{.*#+}} ymm14 = ymm2[0,2,3,3,4,6,7,7]1655; AVX2-FP-NEXT: vshufps {{.*#+}} ymm15 = ymm1[2,3,2,3,6,7,6,7]1656; AVX2-FP-NEXT: vblendps {{.*#+}} ymm14 = ymm15[0],ymm14[1,2],ymm15[3,4],ymm14[5,6],ymm15[7]1657; AVX2-FP-NEXT: vmovaps (%rdi), %ymm151658; AVX2-FP-NEXT: vmovaps (%rsi), %ymm01659; AVX2-FP-NEXT: vunpckhps {{.*#+}} ymm8 = ymm0[2],ymm15[2],ymm0[3],ymm15[3],ymm0[6],ymm15[6],ymm0[7],ymm15[7]1660; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm8 = ymm8[3,3,3,3]1661; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm14 = ymm14[2,1,3,3]1662; AVX2-FP-NEXT: vblendps {{.*#+}} ymm8 = ymm14[0,1,2],ymm8[3,4],ymm14[5,6,7]1663; AVX2-FP-NEXT: vbroadcastsd 24(%r8), %ymm141664; AVX2-FP-NEXT: vblendps {{.*#+}} ymm8 = ymm8[0,1],ymm14[2],ymm8[3,4,5,6],ymm14[7]1665; AVX2-FP-NEXT: vshufps {{.*#+}} ymm14 = ymm2[1,1,1,1,5,5,5,5]1666; AVX2-FP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0],ymm1[1],ymm14[2,3,4],ymm1[5],ymm14[6,7]1667; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm14 = ymm14[2,2,2,2]1668; AVX2-FP-NEXT: vshufps {{.*#+}} ymm9 = ymm0[1,1,2,2,5,5,6,6]1669; AVX2-FP-NEXT: vblendps {{.*#+}} ymm9 = ymm9[0],ymm15[1,2],ymm9[3,4],ymm15[5,6],ymm9[7]1670; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm9 = ymm9[2,2,2,3]1671; AVX2-FP-NEXT: vblendps {{.*#+}} ymm9 = ymm9[0,1,2],ymm14[3,4],ymm9[5,6,7]1672; AVX2-FP-NEXT: vbroadcastsd 16(%r8), %ymm141673; AVX2-FP-NEXT: vblendps {{.*#+}} ymm9 = ymm14[0],ymm9[1,2,3,4],ymm14[5],ymm9[6,7]1674; AVX2-FP-NEXT: vshufps {{.*#+}} ymm5 = ymm5[3,0,2,3,7,4,6,7]1675; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm5[0,1,2,3],ymm3[4],ymm5[5,6,7]1676; AVX2-FP-NEXT: vshufps {{.*#+}} ymm5 = ymm7[0,1,3,0,4,5,7,4]1677; AVX2-FP-NEXT: vshufps {{.*#+}} ymm4 = ymm4[2,3,0,1,6,7,4,5]1678; AVX2-FP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1],ymm5[2],ymm4[3,4,5,6],ymm5[7]1679; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0],ymm4[1,2,3],ymm3[4,5],ymm4[6,7]1680; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2],ymm6[3],ymm3[4,5,6,7]1681; AVX2-FP-NEXT: vshufps {{.*#+}} ymm0 = ymm0[3,0,2,3,7,4,6,7]1682; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm15[4],ymm0[5,6,7]1683; AVX2-FP-NEXT: vshufps {{.*#+}} ymm2 = ymm2[0,1,3,0,4,5,7,4]1684; AVX2-FP-NEXT: vshufps {{.*#+}} ymm1 = ymm1[2,3,0,1,6,7,4,5]1685; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1],ymm2[2],ymm1[3,4,5,6],ymm2[7]1686; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm1[1,2,3],ymm0[4,5],ymm1[6,7]1687; AVX2-FP-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload1688; AVX2-FP-NEXT: # ymm0 = ymm0[0,1,2],mem[3],ymm0[4,5,6,7]1689; AVX2-FP-NEXT: vmovaps %ymm0, 64(%r9)1690; AVX2-FP-NEXT: vmovaps %ymm3, 224(%r9)1691; AVX2-FP-NEXT: vmovaps %ymm9, 96(%r9)1692; AVX2-FP-NEXT: vmovaps %ymm8, 128(%r9)1693; AVX2-FP-NEXT: vmovaps %ymm13, 288(%r9)1694; AVX2-FP-NEXT: vmovaps %ymm12, 256(%r9)1695; AVX2-FP-NEXT: vmovaps %ymm11, 160(%r9)1696; AVX2-FP-NEXT: vmovaps %ymm10, 192(%r9)1697; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload1698; AVX2-FP-NEXT: vmovaps %ymm0, (%r9)1699; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload1700; AVX2-FP-NEXT: vmovaps %ymm0, 32(%r9)1701; AVX2-FP-NEXT: vzeroupper1702; AVX2-FP-NEXT: retq1703;1704; AVX2-FCP-LABEL: store_i32_stride5_vf16:1705; AVX2-FCP: # %bb.0:1706; AVX2-FCP-NEXT: vmovaps 32(%rdi), %ymm11707; AVX2-FCP-NEXT: vmovaps 32(%rsi), %ymm21708; AVX2-FCP-NEXT: vmovaps (%r8), %ymm41709; AVX2-FCP-NEXT: vmovaps 32(%r8), %ymm31710; AVX2-FCP-NEXT: vmovaps (%rsi), %xmm61711; AVX2-FCP-NEXT: vmovaps 32(%rsi), %xmm91712; AVX2-FCP-NEXT: vmovaps (%rdi), %xmm01713; AVX2-FCP-NEXT: vmovaps 32(%rdi), %xmm101714; AVX2-FCP-NEXT: vunpckhps {{.*#+}} xmm5 = xmm0[2],xmm6[2],xmm0[3],xmm6[3]1715; AVX2-FCP-NEXT: vmovaps {{.*#+}} ymm12 = [0,1,0,1,u,u,2,2]1716; AVX2-FCP-NEXT: vpermps %ymm5, %ymm12, %ymm51717; AVX2-FCP-NEXT: vmovaps (%rdx), %xmm81718; AVX2-FCP-NEXT: vmovaps 32(%rdx), %xmm141719; AVX2-FCP-NEXT: vmovaps (%rcx), %xmm111720; AVX2-FCP-NEXT: vmovaps 32(%rcx), %xmm151721; AVX2-FCP-NEXT: vshufps {{.*#+}} xmm7 = xmm11[1,1,2,2]1722; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm7 = xmm7[0,1],xmm8[2],xmm7[3]1723; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm7 = ymm7[0,1,1,3]1724; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm5 = ymm7[0,1],ymm5[2,3],ymm7[4,5],ymm5[6,7]1725; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm7 = ymm4[0,1,2,1]1726; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0],ymm7[1],ymm5[2,3,4,5],ymm7[6],ymm5[7]1727; AVX2-FCP-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill1728; AVX2-FCP-NEXT: vunpckhps {{.*#+}} xmm5 = xmm10[2],xmm9[2],xmm10[3],xmm9[3]1729; AVX2-FCP-NEXT: vpermps %ymm5, %ymm12, %ymm51730; AVX2-FCP-NEXT: vshufps {{.*#+}} xmm7 = xmm15[1,1,2,2]1731; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm7 = xmm7[0,1],xmm14[2],xmm7[3]1732; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm7 = ymm7[0,1,1,3]1733; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm5 = ymm7[0,1],ymm5[2,3],ymm7[4,5],ymm5[6,7]1734; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm7 = ymm3[0,1,2,1]1735; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0],ymm7[1],ymm5[2,3,4,5],ymm7[6],ymm5[7]1736; AVX2-FCP-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill1737; AVX2-FCP-NEXT: vmovaps 32(%rdx), %ymm71738; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm8 = xmm8[0],xmm11[0],xmm8[1],xmm11[1]1739; AVX2-FCP-NEXT: vpermps %ymm8, %ymm12, %ymm81740; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm6 = xmm0[0],xmm6[0],xmm0[1],xmm6[1]1741; AVX2-FCP-NEXT: vmovaps {{.*#+}} ymm0 = [0,1,3,2,3,2,3,2]1742; AVX2-FCP-NEXT: vpermps %ymm6, %ymm0, %ymm111743; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm8 = ymm11[0,1],ymm8[2,3],ymm11[4,5,6],ymm8[7]1744; AVX2-FCP-NEXT: vmovaps 32(%rcx), %ymm111745; AVX2-FCP-NEXT: vinsertf128 $1, (%r8), %ymm6, %ymm61746; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm5 = ymm6[0],ymm8[1,2,3],ymm6[4],ymm8[5,6,7]1747; AVX2-FCP-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill1748; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm8 = ymm11[1,1,1,1,5,5,5,5]1749; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm8 = ymm8[0],ymm7[1],ymm8[2,3,4],ymm7[5],ymm8[6,7]1750; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm8 = ymm8[2,2,2,2]1751; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm13 = ymm2[1,1,2,2,5,5,6,6]1752; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm13 = ymm13[0],ymm1[1,2],ymm13[3,4],ymm1[5,6],ymm13[7]1753; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm13 = ymm13[2,2,2,3]1754; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm8 = ymm13[0,1,2],ymm8[3,4],ymm13[5,6,7]1755; AVX2-FCP-NEXT: vbroadcastsd 48(%r8), %ymm131756; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm8 = ymm13[0],ymm8[1,2,3,4],ymm13[5],ymm8[6,7]1757; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm13 = ymm11[0,2,3,3,4,6,7,7]1758; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm5 = ymm7[2,3,2,3,6,7,6,7]1759; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0],ymm13[1,2],ymm5[3,4],ymm13[5,6],ymm5[7]1760; AVX2-FCP-NEXT: vunpckhps {{.*#+}} ymm13 = ymm2[2],ymm1[2],ymm2[3],ymm1[3],ymm2[6],ymm1[6],ymm2[7],ymm1[7]1761; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm13 = ymm13[3,3,3,3]1762; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm5 = ymm5[2,1,3,3]1763; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2],ymm13[3,4],ymm5[5,6,7]1764; AVX2-FCP-NEXT: vbroadcastsd 56(%r8), %ymm131765; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm13 = ymm5[0,1],ymm13[2],ymm5[3,4,5,6],ymm13[7]1766; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm5 = xmm14[0],xmm15[0],xmm14[1],xmm15[1]1767; AVX2-FCP-NEXT: vpermps %ymm5, %ymm12, %ymm51768; AVX2-FCP-NEXT: vmovaps (%rdx), %ymm121769; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm9 = xmm10[0],xmm9[0],xmm10[1],xmm9[1]1770; AVX2-FCP-NEXT: vmovaps (%rcx), %ymm101771; AVX2-FCP-NEXT: vpermps %ymm9, %ymm0, %ymm01772; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm5[2,3],ymm0[4,5,6],ymm5[7]1773; AVX2-FCP-NEXT: vinsertf128 $1, 32(%r8), %ymm9, %ymm51774; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm9 = ymm5[0],ymm0[1,2,3],ymm5[4],ymm0[5,6,7]1775; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm0 = ymm10[0,2,3,3,4,6,7,7]1776; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm5 = ymm12[2,3,2,3,6,7,6,7]1777; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm5[0],ymm0[1,2],ymm5[3,4],ymm0[5,6],ymm5[7]1778; AVX2-FCP-NEXT: vmovaps (%rdi), %ymm51779; AVX2-FCP-NEXT: vmovaps (%rsi), %ymm141780; AVX2-FCP-NEXT: vunpckhps {{.*#+}} ymm15 = ymm14[2],ymm5[2],ymm14[3],ymm5[3],ymm14[6],ymm5[6],ymm14[7],ymm5[7]1781; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm15 = ymm15[3,3,3,3]1782; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,1,3,3]1783; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm15[3,4],ymm0[5,6,7]1784; AVX2-FCP-NEXT: vbroadcastsd 24(%r8), %ymm151785; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm15[2],ymm0[3,4,5,6],ymm15[7]1786; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm15 = ymm10[1,1,1,1,5,5,5,5]1787; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm15 = ymm15[0],ymm12[1],ymm15[2,3,4],ymm12[5],ymm15[6,7]1788; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm15 = ymm15[2,2,2,2]1789; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm6 = ymm14[1,1,2,2,5,5,6,6]1790; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0],ymm5[1,2],ymm6[3,4],ymm5[5,6],ymm6[7]1791; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm6 = ymm6[2,2,2,3]1792; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0,1,2],ymm15[3,4],ymm6[5,6,7]1793; AVX2-FCP-NEXT: vbroadcastsd 16(%r8), %ymm151794; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm6 = ymm15[0],ymm6[1,2,3,4],ymm15[5],ymm6[6,7]1795; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm2 = ymm2[3,0,2,3,7,4,6,7]1796; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4],ymm2[5,6,7]1797; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm2 = ymm11[0,1,3,0,4,5,7,4]1798; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm7 = ymm7[2,3,0,1,6,7,4,5]1799; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm7[0,1],ymm2[2],ymm7[3,4,5,6],ymm2[7]1800; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1,2,3],ymm1[4,5],ymm2[6,7]1801; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2],ymm3[3],ymm1[4,5,6,7]1802; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm2 = ymm14[3,0,2,3,7,4,6,7]1803; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm5[4],ymm2[5,6,7]1804; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm3 = ymm10[0,1,3,0,4,5,7,4]1805; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm5 = ymm12[2,3,0,1,6,7,4,5]1806; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm3 = ymm5[0,1],ymm3[2],ymm5[3,4,5,6],ymm3[7]1807; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0],ymm3[1,2,3],ymm2[4,5],ymm3[6,7]1808; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2],ymm4[3],ymm2[4,5,6,7]1809; AVX2-FCP-NEXT: vmovaps %ymm2, 64(%r9)1810; AVX2-FCP-NEXT: vmovaps %ymm1, 224(%r9)1811; AVX2-FCP-NEXT: vmovaps %ymm6, 96(%r9)1812; AVX2-FCP-NEXT: vmovaps %ymm0, 128(%r9)1813; AVX2-FCP-NEXT: vmovaps %ymm9, 160(%r9)1814; AVX2-FCP-NEXT: vmovaps %ymm13, 288(%r9)1815; AVX2-FCP-NEXT: vmovaps %ymm8, 256(%r9)1816; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload1817; AVX2-FCP-NEXT: vmovaps %ymm0, (%r9)1818; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload1819; AVX2-FCP-NEXT: vmovaps %ymm0, 192(%r9)1820; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload1821; AVX2-FCP-NEXT: vmovaps %ymm0, 32(%r9)1822; AVX2-FCP-NEXT: vzeroupper1823; AVX2-FCP-NEXT: retq1824;1825; AVX512-LABEL: store_i32_stride5_vf16:1826; AVX512: # %bb.0:1827; AVX512-NEXT: vmovdqa64 (%rdi), %zmm01828; AVX512-NEXT: vmovdqa64 (%rsi), %zmm11829; AVX512-NEXT: vmovdqa64 (%rdx), %zmm21830; AVX512-NEXT: vmovdqa64 (%rcx), %zmm31831; AVX512-NEXT: vmovdqa64 (%r8), %zmm41832; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm5 = [17,0,0,16,2,18,0,1,17,0,0,16,2,18,0,1]1833; AVX512-NEXT: # zmm5 = mem[0,1,2,3,0,1,2,3]1834; AVX512-NEXT: vpermi2d %zmm3, %zmm2, %zmm51835; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = [0,16,2,18,0,1,17,3,0,16,2,18,0,1,17,3]1836; AVX512-NEXT: # zmm6 = mem[0,1,2,3,0,1,2,3]1837; AVX512-NEXT: vpermi2d %zmm1, %zmm0, %zmm61838; AVX512-NEXT: movw $12684, %ax # imm = 0x318C1839; AVX512-NEXT: kmovw %eax, %k11840; AVX512-NEXT: vmovdqa32 %zmm5, %zmm6 {%k1}1841; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm5 = [0,1,2,3,16,5,6,7,8,17,10,11,12,13,18,15]1842; AVX512-NEXT: vpermi2d %zmm4, %zmm6, %zmm51843; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = [0,3,19,5,21,0,4,20,0,3,19,5,21,0,4,20]1844; AVX512-NEXT: # zmm6 = mem[0,1,2,3,0,1,2,3]1845; AVX512-NEXT: vpermi2d %zmm3, %zmm2, %zmm61846; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [3,21,5,0,20,4,22,6,3,21,5,0,20,4,22,6]1847; AVX512-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]1848; AVX512-NEXT: vpermi2d %zmm0, %zmm1, %zmm71849; AVX512-NEXT: movw $6342, %ax # imm = 0x18C61850; AVX512-NEXT: kmovw %eax, %k21851; AVX512-NEXT: vmovdqa32 %zmm6, %zmm7 {%k2}1852; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm6 = [0,1,2,19,4,5,6,7,20,9,10,11,12,21,14,15]1853; AVX512-NEXT: vpermi2d %zmm4, %zmm7, %zmm61854; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [8,24,0,7,23,9,25,0,8,24,0,7,23,9,25,0]1855; AVX512-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]1856; AVX512-NEXT: vpermi2d %zmm1, %zmm0, %zmm71857; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [6,22,8,24,0,7,23,9,6,22,8,24,0,7,23,9]1858; AVX512-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]1859; AVX512-NEXT: vpermi2d %zmm3, %zmm2, %zmm81860; AVX512-NEXT: movw $25368, %ax # imm = 0x63181861; AVX512-NEXT: kmovw %eax, %k21862; AVX512-NEXT: vmovdqa32 %zmm7, %zmm8 {%k2}1863; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm7 = [0,1,22,3,4,5,6,23,8,9,10,11,24,13,14,15]1864; AVX512-NEXT: vpermi2d %zmm4, %zmm8, %zmm71865; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [27,0,10,26,12,28,0,11,27,0,10,26,12,28,0,11]1866; AVX512-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]1867; AVX512-NEXT: vpermi2d %zmm1, %zmm0, %zmm81868; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm9 = [9,27,11,0,26,10,28,12,9,27,11,0,26,10,28,12]1869; AVX512-NEXT: # zmm9 = mem[0,1,2,3,0,1,2,3]1870; AVX512-NEXT: vpermi2d %zmm2, %zmm3, %zmm91871; AVX512-NEXT: vmovdqa32 %zmm8, %zmm9 {%k1}1872; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm8 = [0,25,2,3,4,5,26,7,8,9,10,27,12,13,14,15]1873; AVX512-NEXT: vpermi2d %zmm4, %zmm9, %zmm81874; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm9 = [14,30,0,13,29,15,31,0,14,30,0,13,29,15,31,0]1875; AVX512-NEXT: # zmm9 = mem[0,1,2,3,0,1,2,3]1876; AVX512-NEXT: vpermi2d %zmm3, %zmm2, %zmm91877; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm2 = [0,13,29,15,31,0,14,30,0,13,29,15,31,0,14,30]1878; AVX512-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3]1879; AVX512-NEXT: vpermi2d %zmm1, %zmm0, %zmm21880; AVX512-NEXT: vmovdqa32 %zmm9, %zmm2 {%k2}1881; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm0 = [28,1,2,3,4,29,6,7,8,9,30,11,12,13,14,31]1882; AVX512-NEXT: vpermi2d %zmm4, %zmm2, %zmm01883; AVX512-NEXT: vmovdqa64 %zmm0, 256(%r9)1884; AVX512-NEXT: vmovdqa64 %zmm8, 192(%r9)1885; AVX512-NEXT: vmovdqa64 %zmm7, 128(%r9)1886; AVX512-NEXT: vmovdqa64 %zmm6, 64(%r9)1887; AVX512-NEXT: vmovdqa64 %zmm5, (%r9)1888; AVX512-NEXT: vzeroupper1889; AVX512-NEXT: retq1890;1891; AVX512-FCP-LABEL: store_i32_stride5_vf16:1892; AVX512-FCP: # %bb.0:1893; AVX512-FCP-NEXT: vmovdqa64 (%rdi), %zmm01894; AVX512-FCP-NEXT: vmovdqa64 (%rsi), %zmm11895; AVX512-FCP-NEXT: vmovdqa64 (%rdx), %zmm21896; AVX512-FCP-NEXT: vmovdqa64 (%rcx), %zmm31897; AVX512-FCP-NEXT: vmovdqa64 (%r8), %zmm41898; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm5 = [17,0,0,16,2,18,0,1,17,0,0,16,2,18,0,1]1899; AVX512-FCP-NEXT: # zmm5 = mem[0,1,2,3,0,1,2,3]1900; AVX512-FCP-NEXT: vpermi2d %zmm3, %zmm2, %zmm51901; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = [0,16,2,18,0,1,17,3,0,16,2,18,0,1,17,3]1902; AVX512-FCP-NEXT: # zmm6 = mem[0,1,2,3,0,1,2,3]1903; AVX512-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm61904; AVX512-FCP-NEXT: movw $12684, %ax # imm = 0x318C1905; AVX512-FCP-NEXT: kmovw %eax, %k11906; AVX512-FCP-NEXT: vmovdqa32 %zmm5, %zmm6 {%k1}1907; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm5 = [0,1,2,3,16,5,6,7,8,17,10,11,12,13,18,15]1908; AVX512-FCP-NEXT: vpermi2d %zmm4, %zmm6, %zmm51909; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = [0,3,19,5,21,0,4,20,0,3,19,5,21,0,4,20]1910; AVX512-FCP-NEXT: # zmm6 = mem[0,1,2,3,0,1,2,3]1911; AVX512-FCP-NEXT: vpermi2d %zmm3, %zmm2, %zmm61912; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [3,21,5,0,20,4,22,6,3,21,5,0,20,4,22,6]1913; AVX512-FCP-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]1914; AVX512-FCP-NEXT: vpermi2d %zmm0, %zmm1, %zmm71915; AVX512-FCP-NEXT: movw $6342, %ax # imm = 0x18C61916; AVX512-FCP-NEXT: kmovw %eax, %k21917; AVX512-FCP-NEXT: vmovdqa32 %zmm6, %zmm7 {%k2}1918; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm6 = [0,1,2,19,4,5,6,7,20,9,10,11,12,21,14,15]1919; AVX512-FCP-NEXT: vpermi2d %zmm4, %zmm7, %zmm61920; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [8,24,0,7,23,9,25,0,8,24,0,7,23,9,25,0]1921; AVX512-FCP-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]1922; AVX512-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm71923; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [6,22,8,24,0,7,23,9,6,22,8,24,0,7,23,9]1924; AVX512-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]1925; AVX512-FCP-NEXT: vpermi2d %zmm3, %zmm2, %zmm81926; AVX512-FCP-NEXT: movw $25368, %ax # imm = 0x63181927; AVX512-FCP-NEXT: kmovw %eax, %k21928; AVX512-FCP-NEXT: vmovdqa32 %zmm7, %zmm8 {%k2}1929; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm7 = [0,1,22,3,4,5,6,23,8,9,10,11,24,13,14,15]1930; AVX512-FCP-NEXT: vpermi2d %zmm4, %zmm8, %zmm71931; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [27,0,10,26,12,28,0,11,27,0,10,26,12,28,0,11]1932; AVX512-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]1933; AVX512-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm81934; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm9 = [9,27,11,0,26,10,28,12,9,27,11,0,26,10,28,12]1935; AVX512-FCP-NEXT: # zmm9 = mem[0,1,2,3,0,1,2,3]1936; AVX512-FCP-NEXT: vpermi2d %zmm2, %zmm3, %zmm91937; AVX512-FCP-NEXT: vmovdqa32 %zmm8, %zmm9 {%k1}1938; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm8 = [0,25,2,3,4,5,26,7,8,9,10,27,12,13,14,15]1939; AVX512-FCP-NEXT: vpermi2d %zmm4, %zmm9, %zmm81940; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm9 = [14,30,0,13,29,15,31,0,14,30,0,13,29,15,31,0]1941; AVX512-FCP-NEXT: # zmm9 = mem[0,1,2,3,0,1,2,3]1942; AVX512-FCP-NEXT: vpermi2d %zmm3, %zmm2, %zmm91943; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm2 = [0,13,29,15,31,0,14,30,0,13,29,15,31,0,14,30]1944; AVX512-FCP-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3]1945; AVX512-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm21946; AVX512-FCP-NEXT: vmovdqa32 %zmm9, %zmm2 {%k2}1947; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm0 = [28,1,2,3,4,29,6,7,8,9,30,11,12,13,14,31]1948; AVX512-FCP-NEXT: vpermi2d %zmm4, %zmm2, %zmm01949; AVX512-FCP-NEXT: vmovdqa64 %zmm0, 256(%r9)1950; AVX512-FCP-NEXT: vmovdqa64 %zmm8, 192(%r9)1951; AVX512-FCP-NEXT: vmovdqa64 %zmm7, 128(%r9)1952; AVX512-FCP-NEXT: vmovdqa64 %zmm6, 64(%r9)1953; AVX512-FCP-NEXT: vmovdqa64 %zmm5, (%r9)1954; AVX512-FCP-NEXT: vzeroupper1955; AVX512-FCP-NEXT: retq1956;1957; AVX512DQ-LABEL: store_i32_stride5_vf16:1958; AVX512DQ: # %bb.0:1959; AVX512DQ-NEXT: vmovdqa64 (%rdi), %zmm01960; AVX512DQ-NEXT: vmovdqa64 (%rsi), %zmm11961; AVX512DQ-NEXT: vmovdqa64 (%rdx), %zmm21962; AVX512DQ-NEXT: vmovdqa64 (%rcx), %zmm31963; AVX512DQ-NEXT: vmovdqa64 (%r8), %zmm41964; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm5 = [17,0,0,16,2,18,0,1,17,0,0,16,2,18,0,1]1965; AVX512DQ-NEXT: # zmm5 = mem[0,1,2,3,0,1,2,3]1966; AVX512DQ-NEXT: vpermi2d %zmm3, %zmm2, %zmm51967; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = [0,16,2,18,0,1,17,3,0,16,2,18,0,1,17,3]1968; AVX512DQ-NEXT: # zmm6 = mem[0,1,2,3,0,1,2,3]1969; AVX512DQ-NEXT: vpermi2d %zmm1, %zmm0, %zmm61970; AVX512DQ-NEXT: movw $12684, %ax # imm = 0x318C1971; AVX512DQ-NEXT: kmovw %eax, %k11972; AVX512DQ-NEXT: vmovdqa32 %zmm5, %zmm6 {%k1}1973; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm5 = [0,1,2,3,16,5,6,7,8,17,10,11,12,13,18,15]1974; AVX512DQ-NEXT: vpermi2d %zmm4, %zmm6, %zmm51975; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = [0,3,19,5,21,0,4,20,0,3,19,5,21,0,4,20]1976; AVX512DQ-NEXT: # zmm6 = mem[0,1,2,3,0,1,2,3]1977; AVX512DQ-NEXT: vpermi2d %zmm3, %zmm2, %zmm61978; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [3,21,5,0,20,4,22,6,3,21,5,0,20,4,22,6]1979; AVX512DQ-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]1980; AVX512DQ-NEXT: vpermi2d %zmm0, %zmm1, %zmm71981; AVX512DQ-NEXT: movw $6342, %ax # imm = 0x18C61982; AVX512DQ-NEXT: kmovw %eax, %k21983; AVX512DQ-NEXT: vmovdqa32 %zmm6, %zmm7 {%k2}1984; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm6 = [0,1,2,19,4,5,6,7,20,9,10,11,12,21,14,15]1985; AVX512DQ-NEXT: vpermi2d %zmm4, %zmm7, %zmm61986; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [8,24,0,7,23,9,25,0,8,24,0,7,23,9,25,0]1987; AVX512DQ-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]1988; AVX512DQ-NEXT: vpermi2d %zmm1, %zmm0, %zmm71989; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [6,22,8,24,0,7,23,9,6,22,8,24,0,7,23,9]1990; AVX512DQ-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]1991; AVX512DQ-NEXT: vpermi2d %zmm3, %zmm2, %zmm81992; AVX512DQ-NEXT: movw $25368, %ax # imm = 0x63181993; AVX512DQ-NEXT: kmovw %eax, %k21994; AVX512DQ-NEXT: vmovdqa32 %zmm7, %zmm8 {%k2}1995; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm7 = [0,1,22,3,4,5,6,23,8,9,10,11,24,13,14,15]1996; AVX512DQ-NEXT: vpermi2d %zmm4, %zmm8, %zmm71997; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [27,0,10,26,12,28,0,11,27,0,10,26,12,28,0,11]1998; AVX512DQ-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]1999; AVX512DQ-NEXT: vpermi2d %zmm1, %zmm0, %zmm82000; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm9 = [9,27,11,0,26,10,28,12,9,27,11,0,26,10,28,12]2001; AVX512DQ-NEXT: # zmm9 = mem[0,1,2,3,0,1,2,3]2002; AVX512DQ-NEXT: vpermi2d %zmm2, %zmm3, %zmm92003; AVX512DQ-NEXT: vmovdqa32 %zmm8, %zmm9 {%k1}2004; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm8 = [0,25,2,3,4,5,26,7,8,9,10,27,12,13,14,15]2005; AVX512DQ-NEXT: vpermi2d %zmm4, %zmm9, %zmm82006; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm9 = [14,30,0,13,29,15,31,0,14,30,0,13,29,15,31,0]2007; AVX512DQ-NEXT: # zmm9 = mem[0,1,2,3,0,1,2,3]2008; AVX512DQ-NEXT: vpermi2d %zmm3, %zmm2, %zmm92009; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm2 = [0,13,29,15,31,0,14,30,0,13,29,15,31,0,14,30]2010; AVX512DQ-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3]2011; AVX512DQ-NEXT: vpermi2d %zmm1, %zmm0, %zmm22012; AVX512DQ-NEXT: vmovdqa32 %zmm9, %zmm2 {%k2}2013; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm0 = [28,1,2,3,4,29,6,7,8,9,30,11,12,13,14,31]2014; AVX512DQ-NEXT: vpermi2d %zmm4, %zmm2, %zmm02015; AVX512DQ-NEXT: vmovdqa64 %zmm0, 256(%r9)2016; AVX512DQ-NEXT: vmovdqa64 %zmm8, 192(%r9)2017; AVX512DQ-NEXT: vmovdqa64 %zmm7, 128(%r9)2018; AVX512DQ-NEXT: vmovdqa64 %zmm6, 64(%r9)2019; AVX512DQ-NEXT: vmovdqa64 %zmm5, (%r9)2020; AVX512DQ-NEXT: vzeroupper2021; AVX512DQ-NEXT: retq2022;2023; AVX512DQ-FCP-LABEL: store_i32_stride5_vf16:2024; AVX512DQ-FCP: # %bb.0:2025; AVX512DQ-FCP-NEXT: vmovdqa64 (%rdi), %zmm02026; AVX512DQ-FCP-NEXT: vmovdqa64 (%rsi), %zmm12027; AVX512DQ-FCP-NEXT: vmovdqa64 (%rdx), %zmm22028; AVX512DQ-FCP-NEXT: vmovdqa64 (%rcx), %zmm32029; AVX512DQ-FCP-NEXT: vmovdqa64 (%r8), %zmm42030; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm5 = [17,0,0,16,2,18,0,1,17,0,0,16,2,18,0,1]2031; AVX512DQ-FCP-NEXT: # zmm5 = mem[0,1,2,3,0,1,2,3]2032; AVX512DQ-FCP-NEXT: vpermi2d %zmm3, %zmm2, %zmm52033; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = [0,16,2,18,0,1,17,3,0,16,2,18,0,1,17,3]2034; AVX512DQ-FCP-NEXT: # zmm6 = mem[0,1,2,3,0,1,2,3]2035; AVX512DQ-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm62036; AVX512DQ-FCP-NEXT: movw $12684, %ax # imm = 0x318C2037; AVX512DQ-FCP-NEXT: kmovw %eax, %k12038; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm5, %zmm6 {%k1}2039; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm5 = [0,1,2,3,16,5,6,7,8,17,10,11,12,13,18,15]2040; AVX512DQ-FCP-NEXT: vpermi2d %zmm4, %zmm6, %zmm52041; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = [0,3,19,5,21,0,4,20,0,3,19,5,21,0,4,20]2042; AVX512DQ-FCP-NEXT: # zmm6 = mem[0,1,2,3,0,1,2,3]2043; AVX512DQ-FCP-NEXT: vpermi2d %zmm3, %zmm2, %zmm62044; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [3,21,5,0,20,4,22,6,3,21,5,0,20,4,22,6]2045; AVX512DQ-FCP-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]2046; AVX512DQ-FCP-NEXT: vpermi2d %zmm0, %zmm1, %zmm72047; AVX512DQ-FCP-NEXT: movw $6342, %ax # imm = 0x18C62048; AVX512DQ-FCP-NEXT: kmovw %eax, %k22049; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm6, %zmm7 {%k2}2050; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm6 = [0,1,2,19,4,5,6,7,20,9,10,11,12,21,14,15]2051; AVX512DQ-FCP-NEXT: vpermi2d %zmm4, %zmm7, %zmm62052; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [8,24,0,7,23,9,25,0,8,24,0,7,23,9,25,0]2053; AVX512DQ-FCP-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]2054; AVX512DQ-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm72055; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [6,22,8,24,0,7,23,9,6,22,8,24,0,7,23,9]2056; AVX512DQ-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]2057; AVX512DQ-FCP-NEXT: vpermi2d %zmm3, %zmm2, %zmm82058; AVX512DQ-FCP-NEXT: movw $25368, %ax # imm = 0x63182059; AVX512DQ-FCP-NEXT: kmovw %eax, %k22060; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm7, %zmm8 {%k2}2061; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm7 = [0,1,22,3,4,5,6,23,8,9,10,11,24,13,14,15]2062; AVX512DQ-FCP-NEXT: vpermi2d %zmm4, %zmm8, %zmm72063; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [27,0,10,26,12,28,0,11,27,0,10,26,12,28,0,11]2064; AVX512DQ-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]2065; AVX512DQ-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm82066; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm9 = [9,27,11,0,26,10,28,12,9,27,11,0,26,10,28,12]2067; AVX512DQ-FCP-NEXT: # zmm9 = mem[0,1,2,3,0,1,2,3]2068; AVX512DQ-FCP-NEXT: vpermi2d %zmm2, %zmm3, %zmm92069; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm8, %zmm9 {%k1}2070; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm8 = [0,25,2,3,4,5,26,7,8,9,10,27,12,13,14,15]2071; AVX512DQ-FCP-NEXT: vpermi2d %zmm4, %zmm9, %zmm82072; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm9 = [14,30,0,13,29,15,31,0,14,30,0,13,29,15,31,0]2073; AVX512DQ-FCP-NEXT: # zmm9 = mem[0,1,2,3,0,1,2,3]2074; AVX512DQ-FCP-NEXT: vpermi2d %zmm3, %zmm2, %zmm92075; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm2 = [0,13,29,15,31,0,14,30,0,13,29,15,31,0,14,30]2076; AVX512DQ-FCP-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3]2077; AVX512DQ-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm22078; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm9, %zmm2 {%k2}2079; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm0 = [28,1,2,3,4,29,6,7,8,9,30,11,12,13,14,31]2080; AVX512DQ-FCP-NEXT: vpermi2d %zmm4, %zmm2, %zmm02081; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, 256(%r9)2082; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm8, 192(%r9)2083; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm7, 128(%r9)2084; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, 64(%r9)2085; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm5, (%r9)2086; AVX512DQ-FCP-NEXT: vzeroupper2087; AVX512DQ-FCP-NEXT: retq2088;2089; AVX512BW-LABEL: store_i32_stride5_vf16:2090; AVX512BW: # %bb.0:2091; AVX512BW-NEXT: vmovdqa64 (%rdi), %zmm02092; AVX512BW-NEXT: vmovdqa64 (%rsi), %zmm12093; AVX512BW-NEXT: vmovdqa64 (%rdx), %zmm22094; AVX512BW-NEXT: vmovdqa64 (%rcx), %zmm32095; AVX512BW-NEXT: vmovdqa64 (%r8), %zmm42096; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm5 = [17,0,0,16,2,18,0,1,17,0,0,16,2,18,0,1]2097; AVX512BW-NEXT: # zmm5 = mem[0,1,2,3,0,1,2,3]2098; AVX512BW-NEXT: vpermi2d %zmm3, %zmm2, %zmm52099; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = [0,16,2,18,0,1,17,3,0,16,2,18,0,1,17,3]2100; AVX512BW-NEXT: # zmm6 = mem[0,1,2,3,0,1,2,3]2101; AVX512BW-NEXT: vpermi2d %zmm1, %zmm0, %zmm62102; AVX512BW-NEXT: movw $12684, %ax # imm = 0x318C2103; AVX512BW-NEXT: kmovd %eax, %k12104; AVX512BW-NEXT: vmovdqa32 %zmm5, %zmm6 {%k1}2105; AVX512BW-NEXT: vpmovsxbd {{.*#+}} zmm5 = [0,1,2,3,16,5,6,7,8,17,10,11,12,13,18,15]2106; AVX512BW-NEXT: vpermi2d %zmm4, %zmm6, %zmm52107; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = [0,3,19,5,21,0,4,20,0,3,19,5,21,0,4,20]2108; AVX512BW-NEXT: # zmm6 = mem[0,1,2,3,0,1,2,3]2109; AVX512BW-NEXT: vpermi2d %zmm3, %zmm2, %zmm62110; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [3,21,5,0,20,4,22,6,3,21,5,0,20,4,22,6]2111; AVX512BW-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]2112; AVX512BW-NEXT: vpermi2d %zmm0, %zmm1, %zmm72113; AVX512BW-NEXT: movw $6342, %ax # imm = 0x18C62114; AVX512BW-NEXT: kmovd %eax, %k22115; AVX512BW-NEXT: vmovdqa32 %zmm6, %zmm7 {%k2}2116; AVX512BW-NEXT: vpmovsxbd {{.*#+}} zmm6 = [0,1,2,19,4,5,6,7,20,9,10,11,12,21,14,15]2117; AVX512BW-NEXT: vpermi2d %zmm4, %zmm7, %zmm62118; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [8,24,0,7,23,9,25,0,8,24,0,7,23,9,25,0]2119; AVX512BW-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]2120; AVX512BW-NEXT: vpermi2d %zmm1, %zmm0, %zmm72121; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [6,22,8,24,0,7,23,9,6,22,8,24,0,7,23,9]2122; AVX512BW-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]2123; AVX512BW-NEXT: vpermi2d %zmm3, %zmm2, %zmm82124; AVX512BW-NEXT: movw $25368, %ax # imm = 0x63182125; AVX512BW-NEXT: kmovd %eax, %k22126; AVX512BW-NEXT: vmovdqa32 %zmm7, %zmm8 {%k2}2127; AVX512BW-NEXT: vpmovsxbd {{.*#+}} zmm7 = [0,1,22,3,4,5,6,23,8,9,10,11,24,13,14,15]2128; AVX512BW-NEXT: vpermi2d %zmm4, %zmm8, %zmm72129; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [27,0,10,26,12,28,0,11,27,0,10,26,12,28,0,11]2130; AVX512BW-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]2131; AVX512BW-NEXT: vpermi2d %zmm1, %zmm0, %zmm82132; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm9 = [9,27,11,0,26,10,28,12,9,27,11,0,26,10,28,12]2133; AVX512BW-NEXT: # zmm9 = mem[0,1,2,3,0,1,2,3]2134; AVX512BW-NEXT: vpermi2d %zmm2, %zmm3, %zmm92135; AVX512BW-NEXT: vmovdqa32 %zmm8, %zmm9 {%k1}2136; AVX512BW-NEXT: vpmovsxbd {{.*#+}} zmm8 = [0,25,2,3,4,5,26,7,8,9,10,27,12,13,14,15]2137; AVX512BW-NEXT: vpermi2d %zmm4, %zmm9, %zmm82138; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm9 = [14,30,0,13,29,15,31,0,14,30,0,13,29,15,31,0]2139; AVX512BW-NEXT: # zmm9 = mem[0,1,2,3,0,1,2,3]2140; AVX512BW-NEXT: vpermi2d %zmm3, %zmm2, %zmm92141; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm2 = [0,13,29,15,31,0,14,30,0,13,29,15,31,0,14,30]2142; AVX512BW-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3]2143; AVX512BW-NEXT: vpermi2d %zmm1, %zmm0, %zmm22144; AVX512BW-NEXT: vmovdqa32 %zmm9, %zmm2 {%k2}2145; AVX512BW-NEXT: vpmovsxbd {{.*#+}} zmm0 = [28,1,2,3,4,29,6,7,8,9,30,11,12,13,14,31]2146; AVX512BW-NEXT: vpermi2d %zmm4, %zmm2, %zmm02147; AVX512BW-NEXT: vmovdqa64 %zmm0, 256(%r9)2148; AVX512BW-NEXT: vmovdqa64 %zmm8, 192(%r9)2149; AVX512BW-NEXT: vmovdqa64 %zmm7, 128(%r9)2150; AVX512BW-NEXT: vmovdqa64 %zmm6, 64(%r9)2151; AVX512BW-NEXT: vmovdqa64 %zmm5, (%r9)2152; AVX512BW-NEXT: vzeroupper2153; AVX512BW-NEXT: retq2154;2155; AVX512BW-FCP-LABEL: store_i32_stride5_vf16:2156; AVX512BW-FCP: # %bb.0:2157; AVX512BW-FCP-NEXT: vmovdqa64 (%rdi), %zmm02158; AVX512BW-FCP-NEXT: vmovdqa64 (%rsi), %zmm12159; AVX512BW-FCP-NEXT: vmovdqa64 (%rdx), %zmm22160; AVX512BW-FCP-NEXT: vmovdqa64 (%rcx), %zmm32161; AVX512BW-FCP-NEXT: vmovdqa64 (%r8), %zmm42162; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm5 = [17,0,0,16,2,18,0,1,17,0,0,16,2,18,0,1]2163; AVX512BW-FCP-NEXT: # zmm5 = mem[0,1,2,3,0,1,2,3]2164; AVX512BW-FCP-NEXT: vpermi2d %zmm3, %zmm2, %zmm52165; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = [0,16,2,18,0,1,17,3,0,16,2,18,0,1,17,3]2166; AVX512BW-FCP-NEXT: # zmm6 = mem[0,1,2,3,0,1,2,3]2167; AVX512BW-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm62168; AVX512BW-FCP-NEXT: movw $12684, %ax # imm = 0x318C2169; AVX512BW-FCP-NEXT: kmovd %eax, %k12170; AVX512BW-FCP-NEXT: vmovdqa32 %zmm5, %zmm6 {%k1}2171; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} zmm5 = [0,1,2,3,16,5,6,7,8,17,10,11,12,13,18,15]2172; AVX512BW-FCP-NEXT: vpermi2d %zmm4, %zmm6, %zmm52173; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = [0,3,19,5,21,0,4,20,0,3,19,5,21,0,4,20]2174; AVX512BW-FCP-NEXT: # zmm6 = mem[0,1,2,3,0,1,2,3]2175; AVX512BW-FCP-NEXT: vpermi2d %zmm3, %zmm2, %zmm62176; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [3,21,5,0,20,4,22,6,3,21,5,0,20,4,22,6]2177; AVX512BW-FCP-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]2178; AVX512BW-FCP-NEXT: vpermi2d %zmm0, %zmm1, %zmm72179; AVX512BW-FCP-NEXT: movw $6342, %ax # imm = 0x18C62180; AVX512BW-FCP-NEXT: kmovd %eax, %k22181; AVX512BW-FCP-NEXT: vmovdqa32 %zmm6, %zmm7 {%k2}2182; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} zmm6 = [0,1,2,19,4,5,6,7,20,9,10,11,12,21,14,15]2183; AVX512BW-FCP-NEXT: vpermi2d %zmm4, %zmm7, %zmm62184; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [8,24,0,7,23,9,25,0,8,24,0,7,23,9,25,0]2185; AVX512BW-FCP-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]2186; AVX512BW-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm72187; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [6,22,8,24,0,7,23,9,6,22,8,24,0,7,23,9]2188; AVX512BW-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]2189; AVX512BW-FCP-NEXT: vpermi2d %zmm3, %zmm2, %zmm82190; AVX512BW-FCP-NEXT: movw $25368, %ax # imm = 0x63182191; AVX512BW-FCP-NEXT: kmovd %eax, %k22192; AVX512BW-FCP-NEXT: vmovdqa32 %zmm7, %zmm8 {%k2}2193; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} zmm7 = [0,1,22,3,4,5,6,23,8,9,10,11,24,13,14,15]2194; AVX512BW-FCP-NEXT: vpermi2d %zmm4, %zmm8, %zmm72195; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [27,0,10,26,12,28,0,11,27,0,10,26,12,28,0,11]2196; AVX512BW-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]2197; AVX512BW-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm82198; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm9 = [9,27,11,0,26,10,28,12,9,27,11,0,26,10,28,12]2199; AVX512BW-FCP-NEXT: # zmm9 = mem[0,1,2,3,0,1,2,3]2200; AVX512BW-FCP-NEXT: vpermi2d %zmm2, %zmm3, %zmm92201; AVX512BW-FCP-NEXT: vmovdqa32 %zmm8, %zmm9 {%k1}2202; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} zmm8 = [0,25,2,3,4,5,26,7,8,9,10,27,12,13,14,15]2203; AVX512BW-FCP-NEXT: vpermi2d %zmm4, %zmm9, %zmm82204; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm9 = [14,30,0,13,29,15,31,0,14,30,0,13,29,15,31,0]2205; AVX512BW-FCP-NEXT: # zmm9 = mem[0,1,2,3,0,1,2,3]2206; AVX512BW-FCP-NEXT: vpermi2d %zmm3, %zmm2, %zmm92207; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm2 = [0,13,29,15,31,0,14,30,0,13,29,15,31,0,14,30]2208; AVX512BW-FCP-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3]2209; AVX512BW-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm22210; AVX512BW-FCP-NEXT: vmovdqa32 %zmm9, %zmm2 {%k2}2211; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} zmm0 = [28,1,2,3,4,29,6,7,8,9,30,11,12,13,14,31]2212; AVX512BW-FCP-NEXT: vpermi2d %zmm4, %zmm2, %zmm02213; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, 256(%r9)2214; AVX512BW-FCP-NEXT: vmovdqa64 %zmm8, 192(%r9)2215; AVX512BW-FCP-NEXT: vmovdqa64 %zmm7, 128(%r9)2216; AVX512BW-FCP-NEXT: vmovdqa64 %zmm6, 64(%r9)2217; AVX512BW-FCP-NEXT: vmovdqa64 %zmm5, (%r9)2218; AVX512BW-FCP-NEXT: vzeroupper2219; AVX512BW-FCP-NEXT: retq2220;2221; AVX512DQ-BW-LABEL: store_i32_stride5_vf16:2222; AVX512DQ-BW: # %bb.0:2223; AVX512DQ-BW-NEXT: vmovdqa64 (%rdi), %zmm02224; AVX512DQ-BW-NEXT: vmovdqa64 (%rsi), %zmm12225; AVX512DQ-BW-NEXT: vmovdqa64 (%rdx), %zmm22226; AVX512DQ-BW-NEXT: vmovdqa64 (%rcx), %zmm32227; AVX512DQ-BW-NEXT: vmovdqa64 (%r8), %zmm42228; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm5 = [17,0,0,16,2,18,0,1,17,0,0,16,2,18,0,1]2229; AVX512DQ-BW-NEXT: # zmm5 = mem[0,1,2,3,0,1,2,3]2230; AVX512DQ-BW-NEXT: vpermi2d %zmm3, %zmm2, %zmm52231; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = [0,16,2,18,0,1,17,3,0,16,2,18,0,1,17,3]2232; AVX512DQ-BW-NEXT: # zmm6 = mem[0,1,2,3,0,1,2,3]2233; AVX512DQ-BW-NEXT: vpermi2d %zmm1, %zmm0, %zmm62234; AVX512DQ-BW-NEXT: movw $12684, %ax # imm = 0x318C2235; AVX512DQ-BW-NEXT: kmovd %eax, %k12236; AVX512DQ-BW-NEXT: vmovdqa32 %zmm5, %zmm6 {%k1}2237; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} zmm5 = [0,1,2,3,16,5,6,7,8,17,10,11,12,13,18,15]2238; AVX512DQ-BW-NEXT: vpermi2d %zmm4, %zmm6, %zmm52239; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = [0,3,19,5,21,0,4,20,0,3,19,5,21,0,4,20]2240; AVX512DQ-BW-NEXT: # zmm6 = mem[0,1,2,3,0,1,2,3]2241; AVX512DQ-BW-NEXT: vpermi2d %zmm3, %zmm2, %zmm62242; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [3,21,5,0,20,4,22,6,3,21,5,0,20,4,22,6]2243; AVX512DQ-BW-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]2244; AVX512DQ-BW-NEXT: vpermi2d %zmm0, %zmm1, %zmm72245; AVX512DQ-BW-NEXT: movw $6342, %ax # imm = 0x18C62246; AVX512DQ-BW-NEXT: kmovd %eax, %k22247; AVX512DQ-BW-NEXT: vmovdqa32 %zmm6, %zmm7 {%k2}2248; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} zmm6 = [0,1,2,19,4,5,6,7,20,9,10,11,12,21,14,15]2249; AVX512DQ-BW-NEXT: vpermi2d %zmm4, %zmm7, %zmm62250; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [8,24,0,7,23,9,25,0,8,24,0,7,23,9,25,0]2251; AVX512DQ-BW-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]2252; AVX512DQ-BW-NEXT: vpermi2d %zmm1, %zmm0, %zmm72253; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [6,22,8,24,0,7,23,9,6,22,8,24,0,7,23,9]2254; AVX512DQ-BW-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]2255; AVX512DQ-BW-NEXT: vpermi2d %zmm3, %zmm2, %zmm82256; AVX512DQ-BW-NEXT: movw $25368, %ax # imm = 0x63182257; AVX512DQ-BW-NEXT: kmovd %eax, %k22258; AVX512DQ-BW-NEXT: vmovdqa32 %zmm7, %zmm8 {%k2}2259; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} zmm7 = [0,1,22,3,4,5,6,23,8,9,10,11,24,13,14,15]2260; AVX512DQ-BW-NEXT: vpermi2d %zmm4, %zmm8, %zmm72261; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [27,0,10,26,12,28,0,11,27,0,10,26,12,28,0,11]2262; AVX512DQ-BW-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]2263; AVX512DQ-BW-NEXT: vpermi2d %zmm1, %zmm0, %zmm82264; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm9 = [9,27,11,0,26,10,28,12,9,27,11,0,26,10,28,12]2265; AVX512DQ-BW-NEXT: # zmm9 = mem[0,1,2,3,0,1,2,3]2266; AVX512DQ-BW-NEXT: vpermi2d %zmm2, %zmm3, %zmm92267; AVX512DQ-BW-NEXT: vmovdqa32 %zmm8, %zmm9 {%k1}2268; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} zmm8 = [0,25,2,3,4,5,26,7,8,9,10,27,12,13,14,15]2269; AVX512DQ-BW-NEXT: vpermi2d %zmm4, %zmm9, %zmm82270; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm9 = [14,30,0,13,29,15,31,0,14,30,0,13,29,15,31,0]2271; AVX512DQ-BW-NEXT: # zmm9 = mem[0,1,2,3,0,1,2,3]2272; AVX512DQ-BW-NEXT: vpermi2d %zmm3, %zmm2, %zmm92273; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm2 = [0,13,29,15,31,0,14,30,0,13,29,15,31,0,14,30]2274; AVX512DQ-BW-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3]2275; AVX512DQ-BW-NEXT: vpermi2d %zmm1, %zmm0, %zmm22276; AVX512DQ-BW-NEXT: vmovdqa32 %zmm9, %zmm2 {%k2}2277; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} zmm0 = [28,1,2,3,4,29,6,7,8,9,30,11,12,13,14,31]2278; AVX512DQ-BW-NEXT: vpermi2d %zmm4, %zmm2, %zmm02279; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, 256(%r9)2280; AVX512DQ-BW-NEXT: vmovdqa64 %zmm8, 192(%r9)2281; AVX512DQ-BW-NEXT: vmovdqa64 %zmm7, 128(%r9)2282; AVX512DQ-BW-NEXT: vmovdqa64 %zmm6, 64(%r9)2283; AVX512DQ-BW-NEXT: vmovdqa64 %zmm5, (%r9)2284; AVX512DQ-BW-NEXT: vzeroupper2285; AVX512DQ-BW-NEXT: retq2286;2287; AVX512DQ-BW-FCP-LABEL: store_i32_stride5_vf16:2288; AVX512DQ-BW-FCP: # %bb.0:2289; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rdi), %zmm02290; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rsi), %zmm12291; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rdx), %zmm22292; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rcx), %zmm32293; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%r8), %zmm42294; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm5 = [17,0,0,16,2,18,0,1,17,0,0,16,2,18,0,1]2295; AVX512DQ-BW-FCP-NEXT: # zmm5 = mem[0,1,2,3,0,1,2,3]2296; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm3, %zmm2, %zmm52297; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = [0,16,2,18,0,1,17,3,0,16,2,18,0,1,17,3]2298; AVX512DQ-BW-FCP-NEXT: # zmm6 = mem[0,1,2,3,0,1,2,3]2299; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm62300; AVX512DQ-BW-FCP-NEXT: movw $12684, %ax # imm = 0x318C2301; AVX512DQ-BW-FCP-NEXT: kmovd %eax, %k12302; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm5, %zmm6 {%k1}2303; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} zmm5 = [0,1,2,3,16,5,6,7,8,17,10,11,12,13,18,15]2304; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm4, %zmm6, %zmm52305; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = [0,3,19,5,21,0,4,20,0,3,19,5,21,0,4,20]2306; AVX512DQ-BW-FCP-NEXT: # zmm6 = mem[0,1,2,3,0,1,2,3]2307; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm3, %zmm2, %zmm62308; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [3,21,5,0,20,4,22,6,3,21,5,0,20,4,22,6]2309; AVX512DQ-BW-FCP-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]2310; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm0, %zmm1, %zmm72311; AVX512DQ-BW-FCP-NEXT: movw $6342, %ax # imm = 0x18C62312; AVX512DQ-BW-FCP-NEXT: kmovd %eax, %k22313; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm6, %zmm7 {%k2}2314; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} zmm6 = [0,1,2,19,4,5,6,7,20,9,10,11,12,21,14,15]2315; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm4, %zmm7, %zmm62316; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [8,24,0,7,23,9,25,0,8,24,0,7,23,9,25,0]2317; AVX512DQ-BW-FCP-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]2318; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm72319; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [6,22,8,24,0,7,23,9,6,22,8,24,0,7,23,9]2320; AVX512DQ-BW-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]2321; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm3, %zmm2, %zmm82322; AVX512DQ-BW-FCP-NEXT: movw $25368, %ax # imm = 0x63182323; AVX512DQ-BW-FCP-NEXT: kmovd %eax, %k22324; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm7, %zmm8 {%k2}2325; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} zmm7 = [0,1,22,3,4,5,6,23,8,9,10,11,24,13,14,15]2326; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm4, %zmm8, %zmm72327; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [27,0,10,26,12,28,0,11,27,0,10,26,12,28,0,11]2328; AVX512DQ-BW-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]2329; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm82330; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm9 = [9,27,11,0,26,10,28,12,9,27,11,0,26,10,28,12]2331; AVX512DQ-BW-FCP-NEXT: # zmm9 = mem[0,1,2,3,0,1,2,3]2332; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm2, %zmm3, %zmm92333; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm8, %zmm9 {%k1}2334; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} zmm8 = [0,25,2,3,4,5,26,7,8,9,10,27,12,13,14,15]2335; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm4, %zmm9, %zmm82336; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm9 = [14,30,0,13,29,15,31,0,14,30,0,13,29,15,31,0]2337; AVX512DQ-BW-FCP-NEXT: # zmm9 = mem[0,1,2,3,0,1,2,3]2338; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm3, %zmm2, %zmm92339; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm2 = [0,13,29,15,31,0,14,30,0,13,29,15,31,0,14,30]2340; AVX512DQ-BW-FCP-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3]2341; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm1, %zmm0, %zmm22342; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm9, %zmm2 {%k2}2343; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} zmm0 = [28,1,2,3,4,29,6,7,8,9,30,11,12,13,14,31]2344; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm4, %zmm2, %zmm02345; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, 256(%r9)2346; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm8, 192(%r9)2347; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm7, 128(%r9)2348; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm6, 64(%r9)2349; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm5, (%r9)2350; AVX512DQ-BW-FCP-NEXT: vzeroupper2351; AVX512DQ-BW-FCP-NEXT: retq2352 %in.vec0 = load <16 x i32>, ptr %in.vecptr0, align 642353 %in.vec1 = load <16 x i32>, ptr %in.vecptr1, align 642354 %in.vec2 = load <16 x i32>, ptr %in.vecptr2, align 642355 %in.vec3 = load <16 x i32>, ptr %in.vecptr3, align 642356 %in.vec4 = load <16 x i32>, ptr %in.vecptr4, align 642357 %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>2358 %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>2359 %3 = 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>2360 %4 = shufflevector <16 x i32> %in.vec4, <16 x i32> poison, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>2361 %5 = shufflevector <64 x i32> %3, <64 x i32> %4, <80 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>2362 %interleaved.vec = shufflevector <80 x i32> %5, <80 x i32> poison, <80 x i32> <i32 0, i32 16, i32 32, i32 48, i32 64, i32 1, i32 17, i32 33, i32 49, i32 65, i32 2, i32 18, i32 34, i32 50, i32 66, i32 3, i32 19, i32 35, i32 51, i32 67, i32 4, i32 20, i32 36, i32 52, i32 68, i32 5, i32 21, i32 37, i32 53, i32 69, i32 6, i32 22, i32 38, i32 54, i32 70, i32 7, i32 23, i32 39, i32 55, i32 71, i32 8, i32 24, i32 40, i32 56, i32 72, i32 9, i32 25, i32 41, i32 57, i32 73, i32 10, i32 26, i32 42, i32 58, i32 74, i32 11, i32 27, i32 43, i32 59, i32 75, i32 12, i32 28, i32 44, i32 60, i32 76, i32 13, i32 29, i32 45, i32 61, i32 77, i32 14, i32 30, i32 46, i32 62, i32 78, i32 15, i32 31, i32 47, i32 63, i32 79>2363 store <80 x i32> %interleaved.vec, ptr %out.vec, align 642364 ret void2365}2366 2367define void @store_i32_stride5_vf32(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecptr2, ptr %in.vecptr3, ptr %in.vecptr4, ptr %out.vec) nounwind {2368; SSE-LABEL: store_i32_stride5_vf32:2369; SSE: # %bb.0:2370; SSE-NEXT: subq $712, %rsp # imm = 0x2C82371; SSE-NEXT: movdqa (%rsi), %xmm92372; SSE-NEXT: movdqa 16(%rsi), %xmm72373; SSE-NEXT: movdqa 32(%rsi), %xmm82374; SSE-NEXT: movdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2375; SSE-NEXT: movdqa (%rdx), %xmm112376; SSE-NEXT: movdqa 16(%rdx), %xmm102377; SSE-NEXT: movdqa 32(%rdx), %xmm142378; SSE-NEXT: movdqa %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2379; SSE-NEXT: movaps (%rcx), %xmm32380; SSE-NEXT: movaps 16(%rcx), %xmm52381; SSE-NEXT: movaps 32(%rcx), %xmm62382; SSE-NEXT: movaps (%r8), %xmm42383; SSE-NEXT: movaps 16(%r8), %xmm132384; SSE-NEXT: movaps 32(%r8), %xmm122385; SSE-NEXT: movaps %xmm3, %xmm02386; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,3],xmm4[3,3]2387; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm9[3,3,3,3]2388; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm11[2,3,2,3]2389; SSE-NEXT: movss {{.*#+}} xmm2 = xmm1[0],xmm2[1,2,3]2390; SSE-NEXT: shufps {{.*#+}} xmm2 = xmm2[0,1],xmm0[0,2]2391; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2392; SSE-NEXT: movaps %xmm5, %xmm02393; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,3],xmm13[3,3]2394; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm10[3,3,3,3]2395; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm7[3,3,3,3]2396; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]2397; SSE-NEXT: shufps {{.*#+}} xmm2 = xmm2[0,1],xmm0[0,2]2398; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2399; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm8[3,3,3,3]2400; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm14[2,3,2,3]2401; SSE-NEXT: movss {{.*#+}} xmm1 = xmm0[0],xmm1[1,2,3]2402; SSE-NEXT: movaps %xmm6, %xmm02403; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,3],xmm12[3,3]2404; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,1],xmm0[0,2]2405; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2406; SSE-NEXT: movdqa 48(%rsi), %xmm82407; SSE-NEXT: movdqa 48(%rdx), %xmm02408; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2409; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[3,3,3,3]2410; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm8[3,3,3,3]2411; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]2412; SSE-NEXT: movaps 48(%rcx), %xmm22413; SSE-NEXT: movaps 48(%r8), %xmm142414; SSE-NEXT: movaps %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2415; SSE-NEXT: movaps %xmm2, %xmm02416; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,3],xmm14[3,3]2417; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,1],xmm0[0,2]2418; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2419; SSE-NEXT: movdqa 64(%rsi), %xmm02420; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2421; SSE-NEXT: movdqa 64(%rdx), %xmm12422; SSE-NEXT: movdqa %xmm1, (%rsp) # 16-byte Spill2423; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[3,3,3,3]2424; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[2,3,2,3]2425; SSE-NEXT: movss {{.*#+}} xmm1 = xmm0[0],xmm1[1,2,3]2426; SSE-NEXT: movaps 64(%rcx), %xmm02427; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2428; SSE-NEXT: movaps 64(%r8), %xmm142429; SSE-NEXT: movaps %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2430; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,3],xmm14[3,3]2431; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,1],xmm0[0,2]2432; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2433; SSE-NEXT: movdqa 80(%rsi), %xmm142434; SSE-NEXT: movdqa %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2435; SSE-NEXT: movdqa 80(%rdx), %xmm02436; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2437; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[3,3,3,3]2438; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm14[3,3,3,3]2439; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]2440; SSE-NEXT: movaps 80(%rcx), %xmm02441; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2442; SSE-NEXT: movaps 80(%r8), %xmm142443; SSE-NEXT: movaps %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2444; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,3],xmm14[3,3]2445; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,1],xmm0[0,2]2446; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2447; SSE-NEXT: movdqa 96(%rsi), %xmm02448; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2449; SSE-NEXT: movdqa 96(%rdx), %xmm12450; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2451; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[3,3,3,3]2452; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[2,3,2,3]2453; SSE-NEXT: movss {{.*#+}} xmm1 = xmm0[0],xmm1[1,2,3]2454; SSE-NEXT: movaps 96(%rcx), %xmm02455; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2456; SSE-NEXT: movaps 96(%r8), %xmm142457; SSE-NEXT: movaps %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2458; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,3],xmm14[3,3]2459; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,1],xmm0[0,2]2460; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2461; SSE-NEXT: movdqa 112(%rsi), %xmm142462; SSE-NEXT: movdqa %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2463; SSE-NEXT: movdqa 112(%rdx), %xmm02464; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2465; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[3,3,3,3]2466; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm14[3,3,3,3]2467; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]2468; SSE-NEXT: movaps 112(%rcx), %xmm152469; SSE-NEXT: movaps 112(%r8), %xmm142470; SSE-NEXT: movaps %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2471; SSE-NEXT: movaps %xmm15, %xmm02472; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,3],xmm14[3,3]2473; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,1],xmm0[0,2]2474; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2475; SSE-NEXT: movdqa %xmm11, %xmm12476; SSE-NEXT: movdqa %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2477; SSE-NEXT: movdqa %xmm11, %xmm02478; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1]2479; SSE-NEXT: movdqa (%rdi), %xmm142480; SSE-NEXT: movdqa %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2481; SSE-NEXT: movdqa %xmm14, %xmm112482; SSE-NEXT: movdqa %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2483; SSE-NEXT: punpckldq {{.*#+}} xmm11 = xmm11[0],xmm9[0],xmm11[1],xmm9[1]2484; SSE-NEXT: punpcklqdq {{.*#+}} xmm11 = xmm11[0],xmm0[0]2485; SSE-NEXT: movdqa %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2486; SSE-NEXT: movdqa %xmm14, %xmm02487; SSE-NEXT: punpckhdq {{.*#+}} xmm0 = xmm0[2],xmm9[2],xmm0[3],xmm9[3]2488; SSE-NEXT: movdqa %xmm1, %xmm92489; SSE-NEXT: punpckhdq {{.*#+}} xmm9 = xmm9[2],xmm3[2],xmm9[3],xmm3[3]2490; SSE-NEXT: movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2491; SSE-NEXT: shufps {{.*#+}} xmm3 = xmm3[1,1],xmm4[1,1]2492; SSE-NEXT: shufps {{.*#+}} xmm3 = xmm3[0,2],xmm0[0,1]2493; SSE-NEXT: movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2494; SSE-NEXT: movdqa %xmm14, %xmm02495; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,0],xmm4[2,0]2496; SSE-NEXT: shufps {{.*#+}} xmm9 = xmm9[0,1],xmm0[2,0]2497; SSE-NEXT: movaps %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2498; SSE-NEXT: movdqa %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2499; SSE-NEXT: movdqa %xmm10, %xmm02500; SSE-NEXT: movaps %xmm5, %xmm42501; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm5[0],xmm0[1],xmm5[1]2502; SSE-NEXT: movdqa 16(%rdi), %xmm32503; SSE-NEXT: movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2504; SSE-NEXT: movdqa %xmm3, %xmm52505; SSE-NEXT: movdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2506; SSE-NEXT: punpckldq {{.*#+}} xmm5 = xmm5[0],xmm7[0],xmm5[1],xmm7[1]2507; SSE-NEXT: punpcklqdq {{.*#+}} xmm5 = xmm5[0],xmm0[0]2508; SSE-NEXT: movdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2509; SSE-NEXT: movdqa %xmm3, %xmm02510; SSE-NEXT: punpckhdq {{.*#+}} xmm0 = xmm0[2],xmm7[2],xmm0[3],xmm7[3]2511; SSE-NEXT: movaps %xmm4, %xmm52512; SSE-NEXT: movaps %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2513; SSE-NEXT: shufps {{.*#+}} xmm5 = xmm5[1,1],xmm13[1,1]2514; SSE-NEXT: shufps {{.*#+}} xmm5 = xmm5[0,2],xmm0[0,1]2515; SSE-NEXT: movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2516; SSE-NEXT: movdqa %xmm3, %xmm02517; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,1],xmm13[2,3]2518; SSE-NEXT: unpckhpd {{.*#+}} xmm4 = xmm4[1],xmm10[1]2519; SSE-NEXT: shufps {{.*#+}} xmm4 = xmm4[2,0],xmm0[2,0]2520; SSE-NEXT: movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2521; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload2522; SSE-NEXT: movaps %xmm1, %xmm02523; SSE-NEXT: movaps %xmm6, %xmm42524; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm6[0],xmm0[1],xmm6[1]2525; SSE-NEXT: movaps 32(%rdi), %xmm52526; SSE-NEXT: movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2527; SSE-NEXT: movaps %xmm5, %xmm32528; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm6 # 16-byte Reload2529; SSE-NEXT: unpcklps {{.*#+}} xmm3 = xmm3[0],xmm6[0],xmm3[1],xmm6[1]2530; SSE-NEXT: movlhps {{.*#+}} xmm3 = xmm3[0],xmm0[0]2531; SSE-NEXT: movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2532; SSE-NEXT: movaps %xmm5, %xmm02533; SSE-NEXT: unpckhps {{.*#+}} xmm0 = xmm0[2],xmm6[2],xmm0[3],xmm6[3]2534; SSE-NEXT: movaps %xmm1, %xmm32535; SSE-NEXT: unpckhps {{.*#+}} xmm3 = xmm3[2],xmm4[2],xmm3[3],xmm4[3]2536; SSE-NEXT: movaps %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2537; SSE-NEXT: shufps {{.*#+}} xmm4 = xmm4[1,1],xmm12[1,1]2538; SSE-NEXT: shufps {{.*#+}} xmm4 = xmm4[0,2],xmm0[0,1]2539; SSE-NEXT: movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2540; SSE-NEXT: movaps %xmm5, %xmm02541; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,0],xmm12[2,0]2542; SSE-NEXT: shufps {{.*#+}} xmm3 = xmm3[0,1],xmm0[2,0]2543; SSE-NEXT: movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2544; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload2545; SSE-NEXT: movaps %xmm3, %xmm02546; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]2547; SSE-NEXT: movdqa 48(%rdi), %xmm102548; SSE-NEXT: movdqa %xmm10, %xmm12549; SSE-NEXT: movdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2550; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm8[0],xmm1[1],xmm8[1]2551; SSE-NEXT: punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm0[0]2552; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2553; SSE-NEXT: movdqa %xmm10, %xmm02554; SSE-NEXT: punpckhdq {{.*#+}} xmm0 = xmm0[2],xmm8[2],xmm0[3],xmm8[3]2555; SSE-NEXT: movaps %xmm2, %xmm42556; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload2557; SSE-NEXT: shufps {{.*#+}} xmm4 = xmm4[1,1],xmm1[1,1]2558; SSE-NEXT: shufps {{.*#+}} xmm4 = xmm4[0,2],xmm0[0,1]2559; SSE-NEXT: movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2560; SSE-NEXT: movdqa %xmm10, %xmm02561; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,1],xmm1[2,3]2562; SSE-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm3[1]2563; SSE-NEXT: shufps {{.*#+}} xmm2 = xmm2[2,0],xmm0[2,0]2564; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2565; SSE-NEXT: movaps (%rsp), %xmm2 # 16-byte Reload2566; SSE-NEXT: movaps %xmm2, %xmm02567; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload2568; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]2569; SSE-NEXT: movaps 64(%rdi), %xmm92570; SSE-NEXT: movaps %xmm9, %xmm42571; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload2572; SSE-NEXT: unpcklps {{.*#+}} xmm4 = xmm4[0],xmm3[0],xmm4[1],xmm3[1]2573; SSE-NEXT: movlhps {{.*#+}} xmm4 = xmm4[0],xmm0[0]2574; SSE-NEXT: movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2575; SSE-NEXT: movaps %xmm9, %xmm02576; SSE-NEXT: unpckhps {{.*#+}} xmm0 = xmm0[2],xmm3[2],xmm0[3],xmm3[3]2577; SSE-NEXT: movaps %xmm2, %xmm32578; SSE-NEXT: unpckhps {{.*#+}} xmm3 = xmm3[2],xmm1[2],xmm3[3],xmm1[3]2579; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload2580; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[1,1],xmm2[1,1]2581; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,2],xmm0[0,1]2582; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2583; SSE-NEXT: movaps %xmm9, %xmm02584; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,0],xmm2[2,0]2585; SSE-NEXT: shufps {{.*#+}} xmm3 = xmm3[0,1],xmm0[2,0]2586; SSE-NEXT: movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2587; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload2588; SSE-NEXT: movaps %xmm3, %xmm02589; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload2590; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]2591; SSE-NEXT: movaps 80(%rdi), %xmm72592; SSE-NEXT: movaps %xmm7, %xmm132593; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload2594; SSE-NEXT: unpcklps {{.*#+}} xmm13 = xmm13[0],xmm2[0],xmm13[1],xmm2[1]2595; SSE-NEXT: movlhps {{.*#+}} xmm13 = xmm13[0],xmm0[0]2596; SSE-NEXT: movaps %xmm7, %xmm02597; SSE-NEXT: unpckhps {{.*#+}} xmm0 = xmm0[2],xmm2[2],xmm0[3],xmm2[3]2598; SSE-NEXT: movaps %xmm1, %xmm122599; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload2600; SSE-NEXT: shufps {{.*#+}} xmm12 = xmm12[1,1],xmm2[1,1]2601; SSE-NEXT: shufps {{.*#+}} xmm12 = xmm12[0,2],xmm0[0,1]2602; SSE-NEXT: movaps %xmm7, %xmm02603; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,1],xmm2[2,3]2604; SSE-NEXT: unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm3[1]2605; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[2,0],xmm0[2,0]2606; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2607; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Reload2608; SSE-NEXT: movaps %xmm14, %xmm02609; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload2610; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]2611; SSE-NEXT: movaps 96(%rdi), %xmm52612; SSE-NEXT: movaps %xmm5, %xmm82613; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload2614; SSE-NEXT: unpcklps {{.*#+}} xmm8 = xmm8[0],xmm2[0],xmm8[1],xmm2[1]2615; SSE-NEXT: movlhps {{.*#+}} xmm8 = xmm8[0],xmm0[0]2616; SSE-NEXT: movaps %xmm5, %xmm02617; SSE-NEXT: unpckhps {{.*#+}} xmm0 = xmm0[2],xmm2[2],xmm0[3],xmm2[3]2618; SSE-NEXT: movaps %xmm14, %xmm62619; SSE-NEXT: unpckhps {{.*#+}} xmm6 = xmm6[2],xmm1[2],xmm6[3],xmm1[3]2620; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload2621; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[1,1],xmm2[1,1]2622; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,2],xmm0[0,1]2623; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2624; SSE-NEXT: movaps %xmm5, %xmm02625; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,0],xmm2[2,0]2626; SSE-NEXT: shufps {{.*#+}} xmm6 = xmm6[0,1],xmm0[2,0]2627; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload2628; SSE-NEXT: movaps %xmm2, %xmm112629; SSE-NEXT: movaps %xmm15, %xmm12630; SSE-NEXT: unpcklps {{.*#+}} xmm11 = xmm11[0],xmm15[0],xmm11[1],xmm15[1]2631; SSE-NEXT: movaps 112(%rdi), %xmm32632; SSE-NEXT: movaps %xmm3, %xmm42633; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload2634; SSE-NEXT: unpcklps {{.*#+}} xmm4 = xmm4[0],xmm15[0],xmm4[1],xmm15[1]2635; SSE-NEXT: movlhps {{.*#+}} xmm4 = xmm4[0],xmm11[0]2636; SSE-NEXT: movaps %xmm3, %xmm02637; SSE-NEXT: unpckhps {{.*#+}} xmm0 = xmm0[2],xmm15[2],xmm0[3],xmm15[3]2638; SSE-NEXT: movaps %xmm1, %xmm112639; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload2640; SSE-NEXT: shufps {{.*#+}} xmm11 = xmm11[1,1],xmm15[1,1]2641; SSE-NEXT: shufps {{.*#+}} xmm11 = xmm11[0,2],xmm0[0,1]2642; SSE-NEXT: movaps %xmm3, %xmm02643; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,1],xmm15[2,3]2644; SSE-NEXT: unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm2[1]2645; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[2,0],xmm0[2,0]2646; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload2647; SSE-NEXT: shufps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload2648; SSE-NEXT: # xmm2 = xmm2[1,1],mem[1,1]2649; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2650; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[0,1],xmm2[2,0]2651; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload2652; SSE-NEXT: movss {{.*#+}} xmm0 = xmm2[0],xmm0[1,2,3]2653; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2654; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2655; SSE-NEXT: shufps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload2656; SSE-NEXT: # xmm0 = xmm0[1,1],mem[1,1]2657; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload2658; SSE-NEXT: shufps {{.*#+}} xmm15 = xmm15[0,1],xmm0[2,0]2659; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2660; SSE-NEXT: movss {{.*#+}} xmm15 = xmm0[0],xmm15[1,2,3]2661; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2662; SSE-NEXT: shufps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload2663; SSE-NEXT: # xmm0 = xmm0[1,1],mem[1,1]2664; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload2665; SSE-NEXT: shufps {{.*#+}} xmm2 = xmm2[0,1],xmm0[2,0]2666; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2667; SSE-NEXT: movss {{.*#+}} xmm2 = xmm0[0],xmm2[1,2,3]2668; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2669; SSE-NEXT: shufps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload2670; SSE-NEXT: # xmm0 = xmm0[1,1],mem[1,1]2671; SSE-NEXT: shufps {{.*#+}} xmm10 = xmm10[0,1],xmm0[2,0]2672; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2673; SSE-NEXT: movss {{.*#+}} xmm10 = xmm0[0],xmm10[1,2,3]2674; SSE-NEXT: movaps (%rsp), %xmm0 # 16-byte Reload2675; SSE-NEXT: shufps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload2676; SSE-NEXT: # xmm0 = xmm0[1,1],mem[1,1]2677; SSE-NEXT: shufps {{.*#+}} xmm9 = xmm9[0,1],xmm0[2,0]2678; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2679; SSE-NEXT: movss {{.*#+}} xmm9 = xmm0[0],xmm9[1,2,3]2680; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2681; SSE-NEXT: shufps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload2682; SSE-NEXT: # xmm0 = xmm0[1,1],mem[1,1]2683; SSE-NEXT: shufps {{.*#+}} xmm7 = xmm7[0,1],xmm0[2,0]2684; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2685; SSE-NEXT: movss {{.*#+}} xmm7 = xmm0[0],xmm7[1,2,3]2686; SSE-NEXT: shufps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm14 # 16-byte Folded Reload2687; SSE-NEXT: # xmm14 = xmm14[1,1],mem[1,1]2688; SSE-NEXT: shufps {{.*#+}} xmm5 = xmm5[0,1],xmm14[2,0]2689; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2690; SSE-NEXT: movss {{.*#+}} xmm5 = xmm0[0],xmm5[1,2,3]2691; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2692; SSE-NEXT: shufps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload2693; SSE-NEXT: # xmm0 = xmm0[1,1],mem[1,1]2694; SSE-NEXT: shufps {{.*#+}} xmm3 = xmm3[0,1],xmm0[2,0]2695; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2696; SSE-NEXT: movss {{.*#+}} xmm3 = xmm0[0],xmm3[1,2,3]2697; SSE-NEXT: movaps %xmm1, 608(%r9)2698; SSE-NEXT: movaps %xmm11, 592(%r9)2699; SSE-NEXT: movaps %xmm4, 560(%r9)2700; SSE-NEXT: movaps %xmm6, 528(%r9)2701; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2702; SSE-NEXT: movaps %xmm0, 512(%r9)2703; SSE-NEXT: movaps %xmm8, 480(%r9)2704; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2705; SSE-NEXT: movaps %xmm0, 448(%r9)2706; SSE-NEXT: movaps %xmm12, 432(%r9)2707; SSE-NEXT: movaps %xmm13, 400(%r9)2708; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2709; SSE-NEXT: movaps %xmm0, 368(%r9)2710; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2711; SSE-NEXT: movaps %xmm0, 352(%r9)2712; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2713; SSE-NEXT: movaps %xmm0, 320(%r9)2714; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2715; SSE-NEXT: movaps %xmm0, 288(%r9)2716; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2717; SSE-NEXT: movaps %xmm0, 272(%r9)2718; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2719; SSE-NEXT: movaps %xmm0, 240(%r9)2720; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2721; SSE-NEXT: movaps %xmm0, 208(%r9)2722; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2723; SSE-NEXT: movaps %xmm0, 192(%r9)2724; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2725; SSE-NEXT: movaps %xmm0, 160(%r9)2726; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2727; SSE-NEXT: movaps %xmm0, 128(%r9)2728; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2729; SSE-NEXT: movaps %xmm0, 112(%r9)2730; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2731; SSE-NEXT: movaps %xmm0, 80(%r9)2732; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2733; SSE-NEXT: movaps %xmm0, 48(%r9)2734; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2735; SSE-NEXT: movaps %xmm0, 32(%r9)2736; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2737; SSE-NEXT: movaps %xmm0, (%r9)2738; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2739; SSE-NEXT: movaps %xmm0, 624(%r9)2740; SSE-NEXT: movaps %xmm3, 576(%r9)2741; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2742; SSE-NEXT: movaps %xmm0, 544(%r9)2743; SSE-NEXT: movaps %xmm5, 496(%r9)2744; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2745; SSE-NEXT: movaps %xmm0, 464(%r9)2746; SSE-NEXT: movaps %xmm7, 416(%r9)2747; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2748; SSE-NEXT: movaps %xmm0, 384(%r9)2749; SSE-NEXT: movaps %xmm9, 336(%r9)2750; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2751; SSE-NEXT: movaps %xmm0, 304(%r9)2752; SSE-NEXT: movaps %xmm10, 256(%r9)2753; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2754; SSE-NEXT: movaps %xmm0, 224(%r9)2755; SSE-NEXT: movaps %xmm2, 176(%r9)2756; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2757; SSE-NEXT: movaps %xmm0, 144(%r9)2758; SSE-NEXT: movaps %xmm15, 96(%r9)2759; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2760; SSE-NEXT: movaps %xmm0, 64(%r9)2761; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload2762; SSE-NEXT: movaps %xmm0, 16(%r9)2763; SSE-NEXT: addq $712, %rsp # imm = 0x2C82764; SSE-NEXT: retq2765;2766; AVX-LABEL: store_i32_stride5_vf32:2767; AVX: # %bb.0:2768; AVX-NEXT: subq $456, %rsp # imm = 0x1C82769; AVX-NEXT: vmovaps (%rdi), %ymm112770; AVX-NEXT: vmovups %ymm11, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2771; AVX-NEXT: vmovaps (%rsi), %ymm142772; AVX-NEXT: vmovaps (%rdx), %ymm122773; AVX-NEXT: vmovups %ymm12, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2774; AVX-NEXT: vmovaps (%rcx), %ymm132775; AVX-NEXT: vmovups %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2776; AVX-NEXT: vmovaps (%rsi), %xmm32777; AVX-NEXT: vmovaps 32(%rsi), %xmm02778; AVX-NEXT: vmovaps (%rdi), %xmm42779; AVX-NEXT: vmovaps 32(%rdi), %xmm12780; AVX-NEXT: vinsertps {{.*#+}} xmm2 = zero,xmm4[1],xmm3[1],zero2781; AVX-NEXT: vunpcklps {{.*#+}} xmm5 = xmm4[0],xmm3[0],xmm4[1],xmm3[1]2782; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm5, %ymm52783; AVX-NEXT: vmovaps (%rcx), %xmm62784; AVX-NEXT: vmovaps 32(%rcx), %xmm22785; AVX-NEXT: vmovaps (%rdx), %xmm72786; AVX-NEXT: vmovaps 32(%rdx), %xmm82787; AVX-NEXT: vunpcklps {{.*#+}} xmm9 = xmm7[0],xmm6[0],xmm7[1],xmm6[1]2788; AVX-NEXT: vinsertf128 $1, %xmm7, %ymm9, %ymm92789; AVX-NEXT: vmovddup {{.*#+}} ymm9 = ymm9[0,0,2,2]2790; AVX-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1],ymm9[2,3],ymm5[4,5,6],ymm9[7]2791; AVX-NEXT: vbroadcastf128 {{.*#+}} ymm9 = mem[0,1,0,1]2792; AVX-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm9[4],ymm5[5,6,7]2793; AVX-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2794; AVX-NEXT: vunpckhps {{.*#+}} xmm5 = xmm7[2],xmm6[2],xmm7[3],xmm6[3]2795; AVX-NEXT: vshufps {{.*#+}} xmm6 = xmm6[1,1,1,1]2796; AVX-NEXT: vinsertf128 $1, %xmm5, %ymm6, %ymm52797; AVX-NEXT: vshufps {{.*#+}} xmm3 = xmm3[2,2,2,2]2798; AVX-NEXT: vblendps {{.*#+}} xmm3 = xmm4[0,1,2],xmm3[3]2799; AVX-NEXT: vinsertf128 $1, %xmm4, %ymm3, %ymm32800; AVX-NEXT: vblendps {{.*#+}} ymm3 = ymm5[0,1],ymm3[2,3],ymm5[4,5],ymm3[6,7]2801; AVX-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0],ymm9[1],ymm3[2,3,4,5],ymm9[6],ymm3[7]2802; AVX-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2803; AVX-NEXT: vinsertps {{.*#+}} xmm3 = zero,xmm1[1],xmm0[1],zero2804; AVX-NEXT: vunpcklps {{.*#+}} xmm4 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]2805; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm4, %ymm32806; AVX-NEXT: vunpcklps {{.*#+}} xmm4 = xmm8[0],xmm2[0],xmm8[1],xmm2[1]2807; AVX-NEXT: vinsertf128 $1, %xmm8, %ymm4, %ymm42808; AVX-NEXT: vmovddup {{.*#+}} ymm4 = ymm4[0,0,2,2]2809; AVX-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1],ymm4[2,3],ymm3[4,5,6],ymm4[7]2810; AVX-NEXT: vbroadcastf128 {{.*#+}} ymm4 = mem[0,1,0,1]2811; AVX-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm4[4],ymm3[5,6,7]2812; AVX-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2813; AVX-NEXT: vunpckhps {{.*#+}} xmm3 = xmm8[2],xmm2[2],xmm8[3],xmm2[3]2814; AVX-NEXT: vshufps {{.*#+}} xmm2 = xmm2[1,1,1,1]2815; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm2, %ymm22816; AVX-NEXT: vshufps {{.*#+}} xmm0 = xmm0[2,2,2,2]2817; AVX-NEXT: vblendps {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[3]2818; AVX-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm02819; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1],ymm0[2,3],ymm2[4,5],ymm0[6,7]2820; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm4[1],ymm0[2,3,4,5],ymm4[6],ymm0[7]2821; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2822; AVX-NEXT: vmovaps 64(%rsi), %xmm02823; AVX-NEXT: vmovaps 64(%rdi), %xmm12824; AVX-NEXT: vinsertps {{.*#+}} xmm2 = zero,xmm1[1],xmm0[1],zero2825; AVX-NEXT: vunpcklps {{.*#+}} xmm3 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]2826; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm3, %ymm22827; AVX-NEXT: vmovaps 64(%rcx), %xmm32828; AVX-NEXT: vmovaps 64(%rdx), %xmm42829; AVX-NEXT: vunpcklps {{.*#+}} xmm5 = xmm4[0],xmm3[0],xmm4[1],xmm3[1]2830; AVX-NEXT: vinsertf128 $1, %xmm4, %ymm5, %ymm52831; AVX-NEXT: vmovddup {{.*#+}} ymm5 = ymm5[0,0,2,2]2832; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1],ymm5[2,3],ymm2[4,5,6],ymm5[7]2833; AVX-NEXT: vbroadcastf128 {{.*#+}} ymm5 = mem[0,1,0,1]2834; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm5[4],ymm2[5,6,7]2835; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2836; AVX-NEXT: vunpckhps {{.*#+}} xmm2 = xmm4[2],xmm3[2],xmm4[3],xmm3[3]2837; AVX-NEXT: vshufps {{.*#+}} xmm3 = xmm3[1,1,1,1]2838; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm3, %ymm22839; AVX-NEXT: vshufps {{.*#+}} xmm0 = xmm0[2,2,2,2]2840; AVX-NEXT: vblendps {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[3]2841; AVX-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm02842; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1],ymm0[2,3],ymm2[4,5],ymm0[6,7]2843; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm5[1],ymm0[2,3,4,5],ymm5[6],ymm0[7]2844; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2845; AVX-NEXT: vmovaps 96(%rsi), %xmm02846; AVX-NEXT: vmovaps 96(%rdi), %xmm12847; AVX-NEXT: vinsertps {{.*#+}} xmm2 = zero,xmm1[1],xmm0[1],zero2848; AVX-NEXT: vunpcklps {{.*#+}} xmm3 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]2849; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm3, %ymm22850; AVX-NEXT: vmovaps 96(%rcx), %xmm32851; AVX-NEXT: vmovaps 96(%rdx), %xmm42852; AVX-NEXT: vunpcklps {{.*#+}} xmm5 = xmm4[0],xmm3[0],xmm4[1],xmm3[1]2853; AVX-NEXT: vinsertf128 $1, %xmm4, %ymm5, %ymm52854; AVX-NEXT: vmovddup {{.*#+}} ymm5 = ymm5[0,0,2,2]2855; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1],ymm5[2,3],ymm2[4,5,6],ymm5[7]2856; AVX-NEXT: vmovaps 16(%rsi), %xmm52857; AVX-NEXT: vunpckhps {{.*#+}} xmm4 = xmm4[2],xmm3[2],xmm4[3],xmm3[3]2858; AVX-NEXT: vshufps {{.*#+}} xmm3 = xmm3[1,1,1,1]2859; AVX-NEXT: vinsertf128 $1, %xmm4, %ymm3, %ymm32860; AVX-NEXT: vmovaps 16(%rdi), %xmm42861; AVX-NEXT: vshufps {{.*#+}} xmm0 = xmm0[2,2,2,2]2862; AVX-NEXT: vblendps {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[3]2863; AVX-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm02864; AVX-NEXT: vmovaps 16(%rcx), %xmm12865; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm3[0,1],ymm0[2,3],ymm3[4,5],ymm0[6,7]2866; AVX-NEXT: vbroadcastf128 {{.*#+}} ymm3 = mem[0,1,0,1]2867; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4],ymm2[5,6,7]2868; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2869; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm3[1],ymm0[2,3,4,5],ymm3[6],ymm0[7]2870; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2871; AVX-NEXT: vshufps {{.*#+}} xmm0 = xmm5[2,2,2,2]2872; AVX-NEXT: vblendps {{.*#+}} xmm0 = xmm4[0,1,2],xmm0[3]2873; AVX-NEXT: vinsertps {{.*#+}} xmm2 = zero,xmm4[1],xmm5[1],zero2874; AVX-NEXT: vmovaps 16(%rdx), %xmm32875; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm2, %ymm02876; AVX-NEXT: vshufps {{.*#+}} ymm2 = ymm13[1,1],ymm12[1,1],ymm13[5,5],ymm12[5,5]2877; AVX-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3,2,3]2878; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm2[3,4],ymm0[5,6,7]2879; AVX-NEXT: vshufps {{.*#+}} xmm2 = xmm3[3,3],xmm1[3,3]2880; AVX-NEXT: vunpckhps {{.*#+}} xmm1 = xmm3[2],xmm1[2],xmm3[3],xmm1[3]2881; AVX-NEXT: vbroadcastf128 {{.*#+}} ymm3 = mem[0,1,0,1]2882; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm3[0],ymm0[1,2,3,4],ymm3[5],ymm0[6,7]2883; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2884; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm1, %ymm02885; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm14[3,3],ymm11[3,3],ymm14[7,7],ymm11[7,7]2886; AVX-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm1[2,3,2,3]2887; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3,4],ymm0[5,6,7]2888; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm3[2],ymm0[3,4,5,6],ymm3[7]2889; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2890; AVX-NEXT: vmovaps 48(%rdi), %xmm02891; AVX-NEXT: vmovaps 48(%rsi), %xmm12892; AVX-NEXT: vshufps {{.*#+}} xmm2 = xmm1[2,2,2,2]2893; AVX-NEXT: vblendps {{.*#+}} xmm2 = xmm0[0,1,2],xmm2[3]2894; AVX-NEXT: vinsertps {{.*#+}} xmm0 = zero,xmm0[1],xmm1[1],zero2895; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm02896; AVX-NEXT: vmovaps 32(%rdx), %ymm12897; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2898; AVX-NEXT: vmovaps 32(%rcx), %ymm132899; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm13[1,1],ymm1[1,1],ymm13[5,5],ymm1[5,5]2900; AVX-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm1[2,3,2,3]2901; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3,4],ymm0[5,6,7]2902; AVX-NEXT: vmovaps 48(%rcx), %xmm12903; AVX-NEXT: vmovaps 48(%rdx), %xmm22904; AVX-NEXT: vshufps {{.*#+}} xmm3 = xmm2[3,3],xmm1[3,3]2905; AVX-NEXT: vunpckhps {{.*#+}} xmm1 = xmm2[2],xmm1[2],xmm2[3],xmm1[3]2906; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm1, %ymm12907; AVX-NEXT: vmovaps 32(%rdi), %ymm122908; AVX-NEXT: vmovaps 32(%rsi), %ymm112909; AVX-NEXT: vshufps {{.*#+}} ymm2 = ymm11[3,3],ymm12[3,3],ymm11[7,7],ymm12[7,7]2910; AVX-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3,2,3]2911; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2],ymm2[3,4],ymm1[5,6,7]2912; AVX-NEXT: vbroadcastf128 {{.*#+}} ymm2 = mem[0,1,0,1]2913; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0],ymm0[1,2,3,4],ymm2[5],ymm0[6,7]2914; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2915; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm2[2],ymm1[3,4,5,6],ymm2[7]2916; AVX-NEXT: vmovups %ymm0, (%rsp) # 32-byte Spill2917; AVX-NEXT: vmovaps 80(%rdi), %xmm02918; AVX-NEXT: vmovaps 80(%rsi), %xmm12919; AVX-NEXT: vshufps {{.*#+}} xmm2 = xmm1[2,2,2,2]2920; AVX-NEXT: vblendps {{.*#+}} xmm2 = xmm0[0,1,2],xmm2[3]2921; AVX-NEXT: vinsertps {{.*#+}} xmm0 = zero,xmm0[1],xmm1[1],zero2922; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm02923; AVX-NEXT: vmovaps 64(%rdx), %ymm92924; AVX-NEXT: vmovaps 64(%rcx), %ymm82925; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm8[1,1],ymm9[1,1],ymm8[5,5],ymm9[5,5]2926; AVX-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm1[2,3,2,3]2927; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3,4],ymm0[5,6,7]2928; AVX-NEXT: vmovaps 80(%rcx), %xmm12929; AVX-NEXT: vmovaps 80(%rdx), %xmm22930; AVX-NEXT: vshufps {{.*#+}} xmm3 = xmm2[3,3],xmm1[3,3]2931; AVX-NEXT: vunpckhps {{.*#+}} xmm1 = xmm2[2],xmm1[2],xmm2[3],xmm1[3]2932; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm1, %ymm12933; AVX-NEXT: vmovaps 64(%rdi), %ymm52934; AVX-NEXT: vmovaps 64(%rsi), %ymm42935; AVX-NEXT: vshufps {{.*#+}} ymm6 = ymm4[3,3],ymm5[3,3],ymm4[7,7],ymm5[7,7]2936; AVX-NEXT: vperm2f128 {{.*#+}} ymm6 = ymm6[2,3,2,3]2937; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2],ymm6[3,4],ymm1[5,6,7]2938; AVX-NEXT: vbroadcastf128 {{.*#+}} ymm6 = mem[0,1,0,1]2939; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm6[0],ymm0[1,2,3,4],ymm6[5],ymm0[6,7]2940; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2941; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm6[2],ymm1[3,4,5,6],ymm6[7]2942; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2943; AVX-NEXT: vmovaps 112(%rdi), %xmm02944; AVX-NEXT: vmovaps 112(%rsi), %xmm12945; AVX-NEXT: vshufps {{.*#+}} xmm6 = xmm1[2,2,2,2]2946; AVX-NEXT: vblendps {{.*#+}} xmm6 = xmm0[0,1,2],xmm6[3]2947; AVX-NEXT: vinsertps {{.*#+}} xmm0 = zero,xmm0[1],xmm1[1],zero2948; AVX-NEXT: vinsertf128 $1, %xmm6, %ymm0, %ymm62949; AVX-NEXT: vmovaps 96(%rdx), %ymm32950; AVX-NEXT: vmovaps 96(%rcx), %ymm22951; AVX-NEXT: vshufps {{.*#+}} ymm7 = ymm2[1,1],ymm3[1,1],ymm2[5,5],ymm3[5,5]2952; AVX-NEXT: vperm2f128 {{.*#+}} ymm7 = ymm7[2,3,2,3]2953; AVX-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0,1,2],ymm7[3,4],ymm6[5,6,7]2954; AVX-NEXT: vmovaps 112(%rcx), %xmm72955; AVX-NEXT: vmovaps 112(%rdx), %xmm02956; AVX-NEXT: vshufps {{.*#+}} xmm1 = xmm0[3,3],xmm7[3,3]2957; AVX-NEXT: vunpckhps {{.*#+}} xmm0 = xmm0[2],xmm7[2],xmm0[3],xmm7[3]2958; AVX-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm02959; AVX-NEXT: vmovaps 96(%rdi), %ymm12960; AVX-NEXT: vmovaps 96(%rsi), %ymm72961; AVX-NEXT: vshufps {{.*#+}} ymm15 = ymm7[3,3],ymm1[3,3],ymm7[7,7],ymm1[7,7]2962; AVX-NEXT: vperm2f128 {{.*#+}} ymm15 = ymm15[2,3,2,3]2963; AVX-NEXT: vblendps {{.*#+}} ymm15 = ymm0[0,1,2],ymm15[3,4],ymm0[5,6,7]2964; AVX-NEXT: vbroadcastf128 {{.*#+}} ymm0 = mem[0,1,0,1]2965; AVX-NEXT: vblendps {{.*#+}} ymm10 = ymm0[0],ymm6[1,2,3,4],ymm0[5],ymm6[6,7]2966; AVX-NEXT: vblendps {{.*#+}} ymm6 = ymm15[0,1],ymm0[2],ymm15[3,4,5,6],ymm0[7]2967; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm14[3,0,2,3,7,4,6,7]2968; AVX-NEXT: vblendps $16, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload2969; AVX-NEXT: # ymm0 = ymm0[0,1,2,3],mem[4],ymm0[5,6,7]2970; AVX-NEXT: vpermilps $52, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Folded Reload2971; AVX-NEXT: # ymm15 = mem[0,1,3,0,4,5,7,4]2972; AVX-NEXT: vpermilpd $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Folded Reload2973; AVX-NEXT: # ymm14 = mem[1,0,2,2]2974; AVX-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1],ymm15[2],ymm14[3,4,5,6],ymm15[7]2975; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm14[1,2,3],ymm0[4,5],ymm14[6,7]2976; AVX-NEXT: vshufps {{.*#+}} ymm11 = ymm11[3,0,2,3,7,4,6,7]2977; AVX-NEXT: vblendps {{.*#+}} ymm11 = ymm11[0,1,2,3],ymm12[4],ymm11[5,6,7]2978; AVX-NEXT: vshufps {{.*#+}} ymm12 = ymm13[0,1,3,0,4,5,7,4]2979; AVX-NEXT: vpermilpd $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Folded Reload2980; AVX-NEXT: # ymm13 = mem[1,0,2,2]2981; AVX-NEXT: vblendps {{.*#+}} ymm12 = ymm13[0,1],ymm12[2],ymm13[3,4,5,6],ymm12[7]2982; AVX-NEXT: vblendps {{.*#+}} ymm11 = ymm11[0],ymm12[1,2,3],ymm11[4,5],ymm12[6,7]2983; AVX-NEXT: vshufps {{.*#+}} ymm4 = ymm4[3,0,2,3,7,4,6,7]2984; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm5[4],ymm4[5,6,7]2985; AVX-NEXT: vshufps {{.*#+}} ymm5 = ymm8[0,1,3,0,4,5,7,4]2986; AVX-NEXT: vshufpd {{.*#+}} ymm8 = ymm9[1,0,2,2]2987; AVX-NEXT: vblendps {{.*#+}} ymm5 = ymm8[0,1],ymm5[2],ymm8[3,4,5,6],ymm5[7]2988; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0],ymm5[1,2,3],ymm4[4,5],ymm5[6,7]2989; AVX-NEXT: vshufps {{.*#+}} ymm5 = ymm7[3,0,2,3,7,4,6,7]2990; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm5[0,1,2,3],ymm1[4],ymm5[5,6,7]2991; AVX-NEXT: vshufps {{.*#+}} ymm2 = ymm2[0,1,3,0,4,5,7,4]2992; AVX-NEXT: vshufpd {{.*#+}} ymm3 = ymm3[1,0,2,2]2993; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm3[0,1],ymm2[2],ymm3[3,4,5,6],ymm2[7]2994; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1,2,3],ymm1[4,5],ymm2[6,7]2995; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],mem[3],ymm0[4,5,6,7]2996; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm11[0,1,2],mem[3],ymm11[4,5,6,7]2997; AVX-NEXT: vblendps {{.*#+}} ymm3 = ymm4[0,1,2],mem[3],ymm4[4,5,6,7]2998; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2],mem[3],ymm1[4,5,6,7]2999; AVX-NEXT: vmovaps %ymm1, 544(%r9)3000; AVX-NEXT: vmovaps %ymm3, 384(%r9)3001; AVX-NEXT: vmovaps %ymm2, 224(%r9)3002; AVX-NEXT: vmovaps %ymm0, 64(%r9)3003; AVX-NEXT: vmovaps %ymm6, 608(%r9)3004; AVX-NEXT: vmovaps %ymm10, 576(%r9)3005; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3006; AVX-NEXT: vmovaps %ymm0, 448(%r9)3007; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3008; AVX-NEXT: vmovaps %ymm0, 416(%r9)3009; AVX-NEXT: vmovups (%rsp), %ymm0 # 32-byte Reload3010; AVX-NEXT: vmovaps %ymm0, 288(%r9)3011; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3012; AVX-NEXT: vmovaps %ymm0, 256(%r9)3013; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3014; AVX-NEXT: vmovaps %ymm0, 128(%r9)3015; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3016; AVX-NEXT: vmovaps %ymm0, 96(%r9)3017; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3018; AVX-NEXT: vmovaps %ymm0, 512(%r9)3019; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3020; AVX-NEXT: vmovaps %ymm0, 480(%r9)3021; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3022; AVX-NEXT: vmovaps %ymm0, 352(%r9)3023; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3024; AVX-NEXT: vmovaps %ymm0, 320(%r9)3025; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3026; AVX-NEXT: vmovaps %ymm0, 192(%r9)3027; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3028; AVX-NEXT: vmovaps %ymm0, 160(%r9)3029; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3030; AVX-NEXT: vmovaps %ymm0, 32(%r9)3031; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3032; AVX-NEXT: vmovaps %ymm0, (%r9)3033; AVX-NEXT: addq $456, %rsp # imm = 0x1C83034; AVX-NEXT: vzeroupper3035; AVX-NEXT: retq3036;3037; AVX2-LABEL: store_i32_stride5_vf32:3038; AVX2: # %bb.0:3039; AVX2-NEXT: subq $552, %rsp # imm = 0x2283040; AVX2-NEXT: vmovaps (%r8), %ymm153041; AVX2-NEXT: vmovups %ymm15, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3042; AVX2-NEXT: vmovaps 32(%r8), %ymm143043; AVX2-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3044; AVX2-NEXT: vmovaps 64(%r8), %ymm133045; AVX2-NEXT: vmovups %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3046; AVX2-NEXT: vmovaps (%rcx), %xmm73047; AVX2-NEXT: vmovaps 32(%rcx), %xmm33048; AVX2-NEXT: vmovaps 64(%rcx), %xmm03049; AVX2-NEXT: vmovaps (%rdx), %xmm83050; AVX2-NEXT: vmovaps 32(%rdx), %xmm43051; AVX2-NEXT: vmovaps 64(%rdx), %xmm13052; AVX2-NEXT: vunpcklps {{.*#+}} xmm2 = xmm8[0],xmm7[0],xmm8[1],xmm7[1]3053; AVX2-NEXT: vshufps {{.*#+}} xmm2 = xmm2[0,1,2,2]3054; AVX2-NEXT: vpermpd {{.*#+}} ymm5 = ymm2[0,0,2,1]3055; AVX2-NEXT: vmovaps (%rsi), %xmm93056; AVX2-NEXT: vmovaps 32(%rsi), %xmm63057; AVX2-NEXT: vmovaps 64(%rsi), %xmm23058; AVX2-NEXT: vmovaps (%rdi), %xmm103059; AVX2-NEXT: vunpcklps {{.*#+}} xmm11 = xmm10[0],xmm9[0],xmm10[1],xmm9[1]3060; AVX2-NEXT: vshufps {{.*#+}} xmm11 = xmm11[0,1,3,2]3061; AVX2-NEXT: vpermpd {{.*#+}} ymm12 = ymm11[0,1,1,1]3062; AVX2-NEXT: vblendps {{.*#+}} ymm5 = ymm12[0,1],ymm5[2,3],ymm12[4,5,6],ymm5[7]3063; AVX2-NEXT: vmovaps 32(%rdi), %xmm123064; AVX2-NEXT: vinsertf128 $1, (%r8), %ymm11, %ymm113065; AVX2-NEXT: vblendps {{.*#+}} ymm5 = ymm11[0],ymm5[1,2,3],ymm11[4],ymm5[5,6,7]3066; AVX2-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3067; AVX2-NEXT: vmovaps 64(%rdi), %xmm53068; AVX2-NEXT: vunpckhps {{.*#+}} xmm9 = xmm10[2],xmm9[2],xmm10[3],xmm9[3]3069; AVX2-NEXT: vshufps {{.*#+}} xmm9 = xmm9[0,1,2,2]3070; AVX2-NEXT: vpermpd {{.*#+}} ymm9 = ymm9[0,0,2,1]3071; AVX2-NEXT: vshufps {{.*#+}} xmm7 = xmm7[1,1,2,2]3072; AVX2-NEXT: vblendps {{.*#+}} xmm7 = xmm7[0,1],xmm8[2],xmm7[3]3073; AVX2-NEXT: vpermpd {{.*#+}} ymm7 = ymm7[0,1,1,3]3074; AVX2-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0,1],ymm9[2,3],ymm7[4,5],ymm9[6,7]3075; AVX2-NEXT: vpermpd {{.*#+}} ymm8 = ymm15[0,1,2,1]3076; AVX2-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0],ymm8[1],ymm7[2,3,4,5],ymm8[6],ymm7[7]3077; AVX2-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3078; AVX2-NEXT: vunpcklps {{.*#+}} xmm7 = xmm4[0],xmm3[0],xmm4[1],xmm3[1]3079; AVX2-NEXT: vshufps {{.*#+}} xmm7 = xmm7[0,1,2,2]3080; AVX2-NEXT: vpermpd {{.*#+}} ymm7 = ymm7[0,0,2,1]3081; AVX2-NEXT: vunpcklps {{.*#+}} xmm8 = xmm12[0],xmm6[0],xmm12[1],xmm6[1]3082; AVX2-NEXT: vshufps {{.*#+}} xmm8 = xmm8[0,1,3,2]3083; AVX2-NEXT: vpermpd {{.*#+}} ymm9 = ymm8[0,1,1,1]3084; AVX2-NEXT: vblendps {{.*#+}} ymm7 = ymm9[0,1],ymm7[2,3],ymm9[4,5,6],ymm7[7]3085; AVX2-NEXT: vinsertf128 $1, 32(%r8), %ymm8, %ymm83086; AVX2-NEXT: vblendps {{.*#+}} ymm7 = ymm8[0],ymm7[1,2,3],ymm8[4],ymm7[5,6,7]3087; AVX2-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3088; AVX2-NEXT: vunpckhps {{.*#+}} xmm6 = xmm12[2],xmm6[2],xmm12[3],xmm6[3]3089; AVX2-NEXT: vshufps {{.*#+}} xmm6 = xmm6[0,1,2,2]3090; AVX2-NEXT: vpermpd {{.*#+}} ymm6 = ymm6[0,0,2,1]3091; AVX2-NEXT: vshufps {{.*#+}} xmm3 = xmm3[1,1,2,2]3092; AVX2-NEXT: vblendps {{.*#+}} xmm3 = xmm3[0,1],xmm4[2],xmm3[3]3093; AVX2-NEXT: vpermpd {{.*#+}} ymm3 = ymm3[0,1,1,3]3094; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1],ymm6[2,3],ymm3[4,5],ymm6[6,7]3095; AVX2-NEXT: vpermpd {{.*#+}} ymm4 = ymm14[0,1,2,1]3096; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0],ymm4[1],ymm3[2,3,4,5],ymm4[6],ymm3[7]3097; AVX2-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3098; AVX2-NEXT: vunpcklps {{.*#+}} xmm3 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]3099; AVX2-NEXT: vshufps {{.*#+}} xmm3 = xmm3[0,1,2,2]3100; AVX2-NEXT: vpermpd {{.*#+}} ymm3 = ymm3[0,0,2,1]3101; AVX2-NEXT: vunpcklps {{.*#+}} xmm4 = xmm5[0],xmm2[0],xmm5[1],xmm2[1]3102; AVX2-NEXT: vshufps {{.*#+}} xmm4 = xmm4[0,1,3,2]3103; AVX2-NEXT: vpermpd {{.*#+}} ymm6 = ymm4[0,1,1,1]3104; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm6[0,1],ymm3[2,3],ymm6[4,5,6],ymm3[7]3105; AVX2-NEXT: vinsertf128 $1, 64(%r8), %ymm4, %ymm43106; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm4[0],ymm3[1,2,3],ymm4[4],ymm3[5,6,7]3107; AVX2-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3108; AVX2-NEXT: vunpckhps {{.*#+}} xmm2 = xmm5[2],xmm2[2],xmm5[3],xmm2[3]3109; AVX2-NEXT: vshufps {{.*#+}} xmm2 = xmm2[0,1,2,2]3110; AVX2-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,0,2,1]3111; AVX2-NEXT: vshufps {{.*#+}} xmm0 = xmm0[1,1,2,2]3112; AVX2-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2],xmm0[3]3113; AVX2-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,1,1,3]3114; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3],ymm0[4,5],ymm2[6,7]3115; AVX2-NEXT: vpermpd {{.*#+}} ymm1 = ymm13[0,1,2,1]3116; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2,3,4,5],ymm1[6],ymm0[7]3117; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3118; AVX2-NEXT: vmovaps 96(%rcx), %xmm03119; AVX2-NEXT: vmovaps 96(%rdx), %xmm13120; AVX2-NEXT: vunpcklps {{.*#+}} xmm2 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]3121; AVX2-NEXT: vshufps {{.*#+}} xmm2 = xmm2[0,1,2,2]3122; AVX2-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,0,2,1]3123; AVX2-NEXT: vmovaps 96(%rsi), %xmm33124; AVX2-NEXT: vmovaps 96(%rdi), %xmm43125; AVX2-NEXT: vunpcklps {{.*#+}} xmm5 = xmm4[0],xmm3[0],xmm4[1],xmm3[1]3126; AVX2-NEXT: vshufps {{.*#+}} xmm5 = xmm5[0,1,3,2]3127; AVX2-NEXT: vpermpd {{.*#+}} ymm6 = ymm5[0,1,1,1]3128; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm6[0,1],ymm2[2,3],ymm6[4,5,6],ymm2[7]3129; AVX2-NEXT: vmovaps (%rdi), %ymm63130; AVX2-NEXT: vinsertf128 $1, 96(%r8), %ymm5, %ymm53131; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm5[0],ymm2[1,2,3],ymm5[4],ymm2[5,6,7]3132; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3133; AVX2-NEXT: vmovaps (%rdx), %ymm53134; AVX2-NEXT: vunpckhps {{.*#+}} xmm2 = xmm4[2],xmm3[2],xmm4[3],xmm3[3]3135; AVX2-NEXT: vmovaps (%rcx), %ymm143136; AVX2-NEXT: vshufps {{.*#+}} xmm0 = xmm0[1,1,2,2]3137; AVX2-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2],xmm0[3]3138; AVX2-NEXT: vmovaps 96(%r8), %ymm33139; AVX2-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3140; AVX2-NEXT: vshufps {{.*#+}} xmm1 = xmm2[0,1,2,2]3141; AVX2-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,0,2,1]3142; AVX2-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,1,1,3]3143; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7]3144; AVX2-NEXT: vpermpd {{.*#+}} ymm1 = ymm3[0,1,2,1]3145; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2,3,4,5],ymm1[6],ymm0[7]3146; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3147; AVX2-NEXT: vshufps {{.*#+}} ymm0 = ymm14[1,1,1,1,5,5,5,5]3148; AVX2-NEXT: vmovups %ymm5, (%rsp) # 32-byte Spill3149; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm5[1],ymm0[2,3,4],ymm5[5],ymm0[6,7]3150; AVX2-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2]3151; AVX2-NEXT: vmovaps (%rsi), %ymm123152; AVX2-NEXT: vshufps {{.*#+}} ymm1 = ymm12[1,1,2,2,5,5,6,6]3153; AVX2-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3154; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm6[1,2],ymm1[3,4],ymm6[5,6],ymm1[7]3155; AVX2-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,2,2,3]3156; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4],ymm1[5,6,7]3157; AVX2-NEXT: vbroadcastsd 16(%r8), %ymm13158; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2,3,4],ymm1[5],ymm0[6,7]3159; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3160; AVX2-NEXT: vshufps {{.*#+}} ymm0 = ymm14[0,2,3,3,4,6,7,7]3161; AVX2-NEXT: vshufps {{.*#+}} ymm1 = ymm5[2,3,2,3,6,7,6,7]3162; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2],ymm1[3,4],ymm0[5,6],ymm1[7]3163; AVX2-NEXT: vunpckhps {{.*#+}} ymm1 = ymm12[2],ymm6[2],ymm12[3],ymm6[3],ymm12[6],ymm6[6],ymm12[7],ymm6[7]3164; AVX2-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[3,3,3,3]3165; AVX2-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,1,3,3]3166; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3,4],ymm0[5,6,7]3167; AVX2-NEXT: vbroadcastsd 24(%r8), %ymm13168; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4,5,6],ymm1[7]3169; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3170; AVX2-NEXT: vmovaps 32(%rdx), %ymm113171; AVX2-NEXT: vmovaps 32(%rcx), %ymm103172; AVX2-NEXT: vshufps {{.*#+}} ymm0 = ymm10[1,1,1,1,5,5,5,5]3173; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm11[1],ymm0[2,3,4],ymm11[5],ymm0[6,7]3174; AVX2-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2]3175; AVX2-NEXT: vmovaps 32(%rdi), %ymm93176; AVX2-NEXT: vmovaps 32(%rsi), %ymm83177; AVX2-NEXT: vshufps {{.*#+}} ymm1 = ymm8[1,1,2,2,5,5,6,6]3178; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm9[1,2],ymm1[3,4],ymm9[5,6],ymm1[7]3179; AVX2-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,2,2,3]3180; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4],ymm1[5,6,7]3181; AVX2-NEXT: vbroadcastsd 48(%r8), %ymm13182; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2,3,4],ymm1[5],ymm0[6,7]3183; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3184; AVX2-NEXT: vshufps {{.*#+}} ymm0 = ymm10[0,2,3,3,4,6,7,7]3185; AVX2-NEXT: vshufps {{.*#+}} ymm1 = ymm11[2,3,2,3,6,7,6,7]3186; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2],ymm1[3,4],ymm0[5,6],ymm1[7]3187; AVX2-NEXT: vunpckhps {{.*#+}} ymm1 = ymm8[2],ymm9[2],ymm8[3],ymm9[3],ymm8[6],ymm9[6],ymm8[7],ymm9[7]3188; AVX2-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[3,3,3,3]3189; AVX2-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,1,3,3]3190; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3,4],ymm0[5,6,7]3191; AVX2-NEXT: vbroadcastsd 56(%r8), %ymm13192; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4,5,6],ymm1[7]3193; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3194; AVX2-NEXT: vmovaps 64(%rdx), %ymm73195; AVX2-NEXT: vmovaps 64(%rcx), %ymm63196; AVX2-NEXT: vshufps {{.*#+}} ymm0 = ymm6[1,1,1,1,5,5,5,5]3197; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm7[1],ymm0[2,3,4],ymm7[5],ymm0[6,7]3198; AVX2-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2]3199; AVX2-NEXT: vmovaps 64(%rdi), %ymm53200; AVX2-NEXT: vmovaps 64(%rsi), %ymm43201; AVX2-NEXT: vshufps {{.*#+}} ymm1 = ymm4[1,1,2,2,5,5,6,6]3202; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm5[1,2],ymm1[3,4],ymm5[5,6],ymm1[7]3203; AVX2-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,2,2,3]3204; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4],ymm1[5,6,7]3205; AVX2-NEXT: vbroadcastsd 80(%r8), %ymm13206; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2,3,4],ymm1[5],ymm0[6,7]3207; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3208; AVX2-NEXT: vshufps {{.*#+}} ymm0 = ymm6[0,2,3,3,4,6,7,7]3209; AVX2-NEXT: vshufps {{.*#+}} ymm1 = ymm7[2,3,2,3,6,7,6,7]3210; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2],ymm1[3,4],ymm0[5,6],ymm1[7]3211; AVX2-NEXT: vunpckhps {{.*#+}} ymm1 = ymm4[2],ymm5[2],ymm4[3],ymm5[3],ymm4[6],ymm5[6],ymm4[7],ymm5[7]3212; AVX2-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[3,3,3,3]3213; AVX2-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,1,3,3]3214; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3,4],ymm0[5,6,7]3215; AVX2-NEXT: vbroadcastsd 88(%r8), %ymm13216; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4,5,6],ymm1[7]3217; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3218; AVX2-NEXT: vmovaps 96(%rdx), %ymm33219; AVX2-NEXT: vmovaps 96(%rcx), %ymm23220; AVX2-NEXT: vshufps {{.*#+}} ymm0 = ymm2[1,1,1,1,5,5,5,5]3221; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm3[1],ymm0[2,3,4],ymm3[5],ymm0[6,7]3222; AVX2-NEXT: vpermpd {{.*#+}} ymm13 = ymm0[2,2,2,2]3223; AVX2-NEXT: vmovaps 96(%rdi), %ymm13224; AVX2-NEXT: vmovaps 96(%rsi), %ymm03225; AVX2-NEXT: vshufps {{.*#+}} ymm15 = ymm0[1,1,2,2,5,5,6,6]3226; AVX2-NEXT: vblendps {{.*#+}} ymm15 = ymm15[0],ymm1[1,2],ymm15[3,4],ymm1[5,6],ymm15[7]3227; AVX2-NEXT: vpermpd {{.*#+}} ymm15 = ymm15[2,2,2,3]3228; AVX2-NEXT: vblendps {{.*#+}} ymm13 = ymm15[0,1,2],ymm13[3,4],ymm15[5,6,7]3229; AVX2-NEXT: vbroadcastsd 112(%r8), %ymm153230; AVX2-NEXT: vblendps {{.*#+}} ymm13 = ymm15[0],ymm13[1,2,3,4],ymm15[5],ymm13[6,7]3231; AVX2-NEXT: vmovups %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3232; AVX2-NEXT: vshufps {{.*#+}} ymm13 = ymm2[0,2,3,3,4,6,7,7]3233; AVX2-NEXT: vshufps {{.*#+}} ymm15 = ymm3[2,3,2,3,6,7,6,7]3234; AVX2-NEXT: vblendps {{.*#+}} ymm13 = ymm15[0],ymm13[1,2],ymm15[3,4],ymm13[5,6],ymm15[7]3235; AVX2-NEXT: vunpckhps {{.*#+}} ymm15 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7]3236; AVX2-NEXT: vpermpd {{.*#+}} ymm15 = ymm15[3,3,3,3]3237; AVX2-NEXT: vpermpd {{.*#+}} ymm13 = ymm13[2,1,3,3]3238; AVX2-NEXT: vblendps {{.*#+}} ymm13 = ymm13[0,1,2],ymm15[3,4],ymm13[5,6,7]3239; AVX2-NEXT: vbroadcastsd 120(%r8), %ymm153240; AVX2-NEXT: vblendps {{.*#+}} ymm13 = ymm13[0,1],ymm15[2],ymm13[3,4,5,6],ymm15[7]3241; AVX2-NEXT: vshufps {{.*#+}} ymm12 = ymm12[3,0,2,3,7,4,6,7]3242; AVX2-NEXT: vblendps $16, {{[-0-9]+}}(%r{{[sb]}}p), %ymm12, %ymm12 # 32-byte Folded Reload3243; AVX2-NEXT: # ymm12 = ymm12[0,1,2,3],mem[4],ymm12[5,6,7]3244; AVX2-NEXT: vshufps {{.*#+}} ymm14 = ymm14[0,1,3,0,4,5,7,4]3245; AVX2-NEXT: vpermilps $78, (%rsp), %ymm15 # 32-byte Folded Reload3246; AVX2-NEXT: # ymm15 = mem[2,3,0,1,6,7,4,5]3247; AVX2-NEXT: vblendps {{.*#+}} ymm14 = ymm15[0,1],ymm14[2],ymm15[3,4,5,6],ymm14[7]3248; AVX2-NEXT: vblendps {{.*#+}} ymm12 = ymm12[0],ymm14[1,2,3],ymm12[4,5],ymm14[6,7]3249; AVX2-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm12, %ymm12 # 32-byte Folded Reload3250; AVX2-NEXT: # ymm12 = ymm12[0,1,2],mem[3],ymm12[4,5,6,7]3251; AVX2-NEXT: vshufps {{.*#+}} ymm8 = ymm8[3,0,2,3,7,4,6,7]3252; AVX2-NEXT: vblendps {{.*#+}} ymm8 = ymm8[0,1,2,3],ymm9[4],ymm8[5,6,7]3253; AVX2-NEXT: vshufps {{.*#+}} ymm9 = ymm10[0,1,3,0,4,5,7,4]3254; AVX2-NEXT: vshufps {{.*#+}} ymm10 = ymm11[2,3,0,1,6,7,4,5]3255; AVX2-NEXT: vblendps {{.*#+}} ymm9 = ymm10[0,1],ymm9[2],ymm10[3,4,5,6],ymm9[7]3256; AVX2-NEXT: vblendps {{.*#+}} ymm8 = ymm8[0],ymm9[1,2,3],ymm8[4,5],ymm9[6,7]3257; AVX2-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm8, %ymm8 # 32-byte Folded Reload3258; AVX2-NEXT: # ymm8 = ymm8[0,1,2],mem[3],ymm8[4,5,6,7]3259; AVX2-NEXT: vshufps {{.*#+}} ymm4 = ymm4[3,0,2,3,7,4,6,7]3260; AVX2-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm5[4],ymm4[5,6,7]3261; AVX2-NEXT: vshufps {{.*#+}} ymm5 = ymm6[0,1,3,0,4,5,7,4]3262; AVX2-NEXT: vshufps {{.*#+}} ymm6 = ymm7[2,3,0,1,6,7,4,5]3263; AVX2-NEXT: vblendps {{.*#+}} ymm5 = ymm6[0,1],ymm5[2],ymm6[3,4,5,6],ymm5[7]3264; AVX2-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0],ymm5[1,2,3],ymm4[4,5],ymm5[6,7]3265; AVX2-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4, %ymm4 # 32-byte Folded Reload3266; AVX2-NEXT: # ymm4 = ymm4[0,1,2],mem[3],ymm4[4,5,6,7]3267; AVX2-NEXT: vshufps {{.*#+}} ymm0 = ymm0[3,0,2,3,7,4,6,7]3268; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4],ymm0[5,6,7]3269; AVX2-NEXT: vshufps {{.*#+}} ymm1 = ymm2[0,1,3,0,4,5,7,4]3270; AVX2-NEXT: vshufps {{.*#+}} ymm2 = ymm3[2,3,0,1,6,7,4,5]3271; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1],ymm1[2],ymm2[3,4,5,6],ymm1[7]3272; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm1[1,2,3],ymm0[4,5],ymm1[6,7]3273; AVX2-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload3274; AVX2-NEXT: # ymm0 = ymm0[0,1,2],mem[3],ymm0[4,5,6,7]3275; AVX2-NEXT: vmovaps %ymm0, 544(%r9)3276; AVX2-NEXT: vmovaps %ymm4, 384(%r9)3277; AVX2-NEXT: vmovaps %ymm8, 224(%r9)3278; AVX2-NEXT: vmovaps %ymm12, 64(%r9)3279; AVX2-NEXT: vmovaps %ymm13, 608(%r9)3280; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3281; AVX2-NEXT: vmovaps %ymm0, 576(%r9)3282; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3283; AVX2-NEXT: vmovaps %ymm0, 448(%r9)3284; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3285; AVX2-NEXT: vmovaps %ymm0, 416(%r9)3286; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3287; AVX2-NEXT: vmovaps %ymm0, 288(%r9)3288; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3289; AVX2-NEXT: vmovaps %ymm0, 256(%r9)3290; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3291; AVX2-NEXT: vmovaps %ymm0, 128(%r9)3292; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3293; AVX2-NEXT: vmovaps %ymm0, 96(%r9)3294; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3295; AVX2-NEXT: vmovaps %ymm0, 512(%r9)3296; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3297; AVX2-NEXT: vmovaps %ymm0, 480(%r9)3298; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3299; AVX2-NEXT: vmovaps %ymm0, 352(%r9)3300; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3301; AVX2-NEXT: vmovaps %ymm0, 320(%r9)3302; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3303; AVX2-NEXT: vmovaps %ymm0, 192(%r9)3304; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3305; AVX2-NEXT: vmovaps %ymm0, 160(%r9)3306; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3307; AVX2-NEXT: vmovaps %ymm0, 32(%r9)3308; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3309; AVX2-NEXT: vmovaps %ymm0, (%r9)3310; AVX2-NEXT: addq $552, %rsp # imm = 0x2283311; AVX2-NEXT: vzeroupper3312; AVX2-NEXT: retq3313;3314; AVX2-FP-LABEL: store_i32_stride5_vf32:3315; AVX2-FP: # %bb.0:3316; AVX2-FP-NEXT: subq $552, %rsp # imm = 0x2283317; AVX2-FP-NEXT: vmovaps (%r8), %ymm153318; AVX2-FP-NEXT: vmovups %ymm15, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3319; AVX2-FP-NEXT: vmovaps 32(%r8), %ymm143320; AVX2-FP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3321; AVX2-FP-NEXT: vmovaps 64(%r8), %ymm133322; AVX2-FP-NEXT: vmovups %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3323; AVX2-FP-NEXT: vmovaps (%rcx), %xmm73324; AVX2-FP-NEXT: vmovaps 32(%rcx), %xmm33325; AVX2-FP-NEXT: vmovaps 64(%rcx), %xmm03326; AVX2-FP-NEXT: vmovaps (%rdx), %xmm83327; AVX2-FP-NEXT: vmovaps 32(%rdx), %xmm43328; AVX2-FP-NEXT: vmovaps 64(%rdx), %xmm13329; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm2 = xmm8[0],xmm7[0],xmm8[1],xmm7[1]3330; AVX2-FP-NEXT: vshufps {{.*#+}} xmm2 = xmm2[0,1,2,2]3331; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm5 = ymm2[0,0,2,1]3332; AVX2-FP-NEXT: vmovaps (%rsi), %xmm93333; AVX2-FP-NEXT: vmovaps 32(%rsi), %xmm63334; AVX2-FP-NEXT: vmovaps 64(%rsi), %xmm23335; AVX2-FP-NEXT: vmovaps (%rdi), %xmm103336; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm11 = xmm10[0],xmm9[0],xmm10[1],xmm9[1]3337; AVX2-FP-NEXT: vshufps {{.*#+}} xmm11 = xmm11[0,1,3,2]3338; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm12 = ymm11[0,1,1,1]3339; AVX2-FP-NEXT: vblendps {{.*#+}} ymm5 = ymm12[0,1],ymm5[2,3],ymm12[4,5,6],ymm5[7]3340; AVX2-FP-NEXT: vmovaps 32(%rdi), %xmm123341; AVX2-FP-NEXT: vinsertf128 $1, (%r8), %ymm11, %ymm113342; AVX2-FP-NEXT: vblendps {{.*#+}} ymm5 = ymm11[0],ymm5[1,2,3],ymm11[4],ymm5[5,6,7]3343; AVX2-FP-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3344; AVX2-FP-NEXT: vmovaps 64(%rdi), %xmm53345; AVX2-FP-NEXT: vunpckhps {{.*#+}} xmm9 = xmm10[2],xmm9[2],xmm10[3],xmm9[3]3346; AVX2-FP-NEXT: vshufps {{.*#+}} xmm9 = xmm9[0,1,2,2]3347; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm9 = ymm9[0,0,2,1]3348; AVX2-FP-NEXT: vshufps {{.*#+}} xmm7 = xmm7[1,1,2,2]3349; AVX2-FP-NEXT: vblendps {{.*#+}} xmm7 = xmm7[0,1],xmm8[2],xmm7[3]3350; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm7 = ymm7[0,1,1,3]3351; AVX2-FP-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0,1],ymm9[2,3],ymm7[4,5],ymm9[6,7]3352; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm8 = ymm15[0,1,2,1]3353; AVX2-FP-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0],ymm8[1],ymm7[2,3,4,5],ymm8[6],ymm7[7]3354; AVX2-FP-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3355; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm7 = xmm4[0],xmm3[0],xmm4[1],xmm3[1]3356; AVX2-FP-NEXT: vshufps {{.*#+}} xmm7 = xmm7[0,1,2,2]3357; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm7 = ymm7[0,0,2,1]3358; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm8 = xmm12[0],xmm6[0],xmm12[1],xmm6[1]3359; AVX2-FP-NEXT: vshufps {{.*#+}} xmm8 = xmm8[0,1,3,2]3360; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm9 = ymm8[0,1,1,1]3361; AVX2-FP-NEXT: vblendps {{.*#+}} ymm7 = ymm9[0,1],ymm7[2,3],ymm9[4,5,6],ymm7[7]3362; AVX2-FP-NEXT: vinsertf128 $1, 32(%r8), %ymm8, %ymm83363; AVX2-FP-NEXT: vblendps {{.*#+}} ymm7 = ymm8[0],ymm7[1,2,3],ymm8[4],ymm7[5,6,7]3364; AVX2-FP-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3365; AVX2-FP-NEXT: vunpckhps {{.*#+}} xmm6 = xmm12[2],xmm6[2],xmm12[3],xmm6[3]3366; AVX2-FP-NEXT: vshufps {{.*#+}} xmm6 = xmm6[0,1,2,2]3367; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm6 = ymm6[0,0,2,1]3368; AVX2-FP-NEXT: vshufps {{.*#+}} xmm3 = xmm3[1,1,2,2]3369; AVX2-FP-NEXT: vblendps {{.*#+}} xmm3 = xmm3[0,1],xmm4[2],xmm3[3]3370; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm3 = ymm3[0,1,1,3]3371; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1],ymm6[2,3],ymm3[4,5],ymm6[6,7]3372; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm4 = ymm14[0,1,2,1]3373; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0],ymm4[1],ymm3[2,3,4,5],ymm4[6],ymm3[7]3374; AVX2-FP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3375; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm3 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]3376; AVX2-FP-NEXT: vshufps {{.*#+}} xmm3 = xmm3[0,1,2,2]3377; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm3 = ymm3[0,0,2,1]3378; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm4 = xmm5[0],xmm2[0],xmm5[1],xmm2[1]3379; AVX2-FP-NEXT: vshufps {{.*#+}} xmm4 = xmm4[0,1,3,2]3380; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm6 = ymm4[0,1,1,1]3381; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm6[0,1],ymm3[2,3],ymm6[4,5,6],ymm3[7]3382; AVX2-FP-NEXT: vinsertf128 $1, 64(%r8), %ymm4, %ymm43383; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm4[0],ymm3[1,2,3],ymm4[4],ymm3[5,6,7]3384; AVX2-FP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3385; AVX2-FP-NEXT: vunpckhps {{.*#+}} xmm2 = xmm5[2],xmm2[2],xmm5[3],xmm2[3]3386; AVX2-FP-NEXT: vshufps {{.*#+}} xmm2 = xmm2[0,1,2,2]3387; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,0,2,1]3388; AVX2-FP-NEXT: vshufps {{.*#+}} xmm0 = xmm0[1,1,2,2]3389; AVX2-FP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2],xmm0[3]3390; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,1,1,3]3391; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3],ymm0[4,5],ymm2[6,7]3392; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm1 = ymm13[0,1,2,1]3393; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2,3,4,5],ymm1[6],ymm0[7]3394; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3395; AVX2-FP-NEXT: vmovaps 96(%rcx), %xmm03396; AVX2-FP-NEXT: vmovaps 96(%rdx), %xmm13397; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm2 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]3398; AVX2-FP-NEXT: vshufps {{.*#+}} xmm2 = xmm2[0,1,2,2]3399; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,0,2,1]3400; AVX2-FP-NEXT: vmovaps 96(%rsi), %xmm33401; AVX2-FP-NEXT: vmovaps 96(%rdi), %xmm43402; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm5 = xmm4[0],xmm3[0],xmm4[1],xmm3[1]3403; AVX2-FP-NEXT: vshufps {{.*#+}} xmm5 = xmm5[0,1,3,2]3404; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm6 = ymm5[0,1,1,1]3405; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm6[0,1],ymm2[2,3],ymm6[4,5,6],ymm2[7]3406; AVX2-FP-NEXT: vmovaps (%rdi), %ymm63407; AVX2-FP-NEXT: vinsertf128 $1, 96(%r8), %ymm5, %ymm53408; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm5[0],ymm2[1,2,3],ymm5[4],ymm2[5,6,7]3409; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3410; AVX2-FP-NEXT: vmovaps (%rdx), %ymm53411; AVX2-FP-NEXT: vunpckhps {{.*#+}} xmm2 = xmm4[2],xmm3[2],xmm4[3],xmm3[3]3412; AVX2-FP-NEXT: vmovaps (%rcx), %ymm143413; AVX2-FP-NEXT: vshufps {{.*#+}} xmm0 = xmm0[1,1,2,2]3414; AVX2-FP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2],xmm0[3]3415; AVX2-FP-NEXT: vmovaps 96(%r8), %ymm33416; AVX2-FP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3417; AVX2-FP-NEXT: vshufps {{.*#+}} xmm1 = xmm2[0,1,2,2]3418; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,0,2,1]3419; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,1,1,3]3420; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7]3421; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm1 = ymm3[0,1,2,1]3422; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2,3,4,5],ymm1[6],ymm0[7]3423; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3424; AVX2-FP-NEXT: vshufps {{.*#+}} ymm0 = ymm14[1,1,1,1,5,5,5,5]3425; AVX2-FP-NEXT: vmovups %ymm5, (%rsp) # 32-byte Spill3426; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm5[1],ymm0[2,3,4],ymm5[5],ymm0[6,7]3427; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2]3428; AVX2-FP-NEXT: vmovaps (%rsi), %ymm123429; AVX2-FP-NEXT: vshufps {{.*#+}} ymm1 = ymm12[1,1,2,2,5,5,6,6]3430; AVX2-FP-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3431; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm6[1,2],ymm1[3,4],ymm6[5,6],ymm1[7]3432; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,2,2,3]3433; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4],ymm1[5,6,7]3434; AVX2-FP-NEXT: vbroadcastsd 16(%r8), %ymm13435; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2,3,4],ymm1[5],ymm0[6,7]3436; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3437; AVX2-FP-NEXT: vshufps {{.*#+}} ymm0 = ymm14[0,2,3,3,4,6,7,7]3438; AVX2-FP-NEXT: vshufps {{.*#+}} ymm1 = ymm5[2,3,2,3,6,7,6,7]3439; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2],ymm1[3,4],ymm0[5,6],ymm1[7]3440; AVX2-FP-NEXT: vunpckhps {{.*#+}} ymm1 = ymm12[2],ymm6[2],ymm12[3],ymm6[3],ymm12[6],ymm6[6],ymm12[7],ymm6[7]3441; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[3,3,3,3]3442; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,1,3,3]3443; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3,4],ymm0[5,6,7]3444; AVX2-FP-NEXT: vbroadcastsd 24(%r8), %ymm13445; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4,5,6],ymm1[7]3446; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3447; AVX2-FP-NEXT: vmovaps 32(%rdx), %ymm113448; AVX2-FP-NEXT: vmovaps 32(%rcx), %ymm103449; AVX2-FP-NEXT: vshufps {{.*#+}} ymm0 = ymm10[1,1,1,1,5,5,5,5]3450; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm11[1],ymm0[2,3,4],ymm11[5],ymm0[6,7]3451; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2]3452; AVX2-FP-NEXT: vmovaps 32(%rdi), %ymm93453; AVX2-FP-NEXT: vmovaps 32(%rsi), %ymm83454; AVX2-FP-NEXT: vshufps {{.*#+}} ymm1 = ymm8[1,1,2,2,5,5,6,6]3455; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm9[1,2],ymm1[3,4],ymm9[5,6],ymm1[7]3456; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,2,2,3]3457; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4],ymm1[5,6,7]3458; AVX2-FP-NEXT: vbroadcastsd 48(%r8), %ymm13459; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2,3,4],ymm1[5],ymm0[6,7]3460; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3461; AVX2-FP-NEXT: vshufps {{.*#+}} ymm0 = ymm10[0,2,3,3,4,6,7,7]3462; AVX2-FP-NEXT: vshufps {{.*#+}} ymm1 = ymm11[2,3,2,3,6,7,6,7]3463; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2],ymm1[3,4],ymm0[5,6],ymm1[7]3464; AVX2-FP-NEXT: vunpckhps {{.*#+}} ymm1 = ymm8[2],ymm9[2],ymm8[3],ymm9[3],ymm8[6],ymm9[6],ymm8[7],ymm9[7]3465; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[3,3,3,3]3466; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,1,3,3]3467; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3,4],ymm0[5,6,7]3468; AVX2-FP-NEXT: vbroadcastsd 56(%r8), %ymm13469; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4,5,6],ymm1[7]3470; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3471; AVX2-FP-NEXT: vmovaps 64(%rdx), %ymm73472; AVX2-FP-NEXT: vmovaps 64(%rcx), %ymm63473; AVX2-FP-NEXT: vshufps {{.*#+}} ymm0 = ymm6[1,1,1,1,5,5,5,5]3474; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm7[1],ymm0[2,3,4],ymm7[5],ymm0[6,7]3475; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2]3476; AVX2-FP-NEXT: vmovaps 64(%rdi), %ymm53477; AVX2-FP-NEXT: vmovaps 64(%rsi), %ymm43478; AVX2-FP-NEXT: vshufps {{.*#+}} ymm1 = ymm4[1,1,2,2,5,5,6,6]3479; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm5[1,2],ymm1[3,4],ymm5[5,6],ymm1[7]3480; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,2,2,3]3481; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4],ymm1[5,6,7]3482; AVX2-FP-NEXT: vbroadcastsd 80(%r8), %ymm13483; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2,3,4],ymm1[5],ymm0[6,7]3484; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3485; AVX2-FP-NEXT: vshufps {{.*#+}} ymm0 = ymm6[0,2,3,3,4,6,7,7]3486; AVX2-FP-NEXT: vshufps {{.*#+}} ymm1 = ymm7[2,3,2,3,6,7,6,7]3487; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2],ymm1[3,4],ymm0[5,6],ymm1[7]3488; AVX2-FP-NEXT: vunpckhps {{.*#+}} ymm1 = ymm4[2],ymm5[2],ymm4[3],ymm5[3],ymm4[6],ymm5[6],ymm4[7],ymm5[7]3489; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[3,3,3,3]3490; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,1,3,3]3491; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3,4],ymm0[5,6,7]3492; AVX2-FP-NEXT: vbroadcastsd 88(%r8), %ymm13493; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4,5,6],ymm1[7]3494; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3495; AVX2-FP-NEXT: vmovaps 96(%rdx), %ymm33496; AVX2-FP-NEXT: vmovaps 96(%rcx), %ymm23497; AVX2-FP-NEXT: vshufps {{.*#+}} ymm0 = ymm2[1,1,1,1,5,5,5,5]3498; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm3[1],ymm0[2,3,4],ymm3[5],ymm0[6,7]3499; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm13 = ymm0[2,2,2,2]3500; AVX2-FP-NEXT: vmovaps 96(%rdi), %ymm13501; AVX2-FP-NEXT: vmovaps 96(%rsi), %ymm03502; AVX2-FP-NEXT: vshufps {{.*#+}} ymm15 = ymm0[1,1,2,2,5,5,6,6]3503; AVX2-FP-NEXT: vblendps {{.*#+}} ymm15 = ymm15[0],ymm1[1,2],ymm15[3,4],ymm1[5,6],ymm15[7]3504; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm15 = ymm15[2,2,2,3]3505; AVX2-FP-NEXT: vblendps {{.*#+}} ymm13 = ymm15[0,1,2],ymm13[3,4],ymm15[5,6,7]3506; AVX2-FP-NEXT: vbroadcastsd 112(%r8), %ymm153507; AVX2-FP-NEXT: vblendps {{.*#+}} ymm13 = ymm15[0],ymm13[1,2,3,4],ymm15[5],ymm13[6,7]3508; AVX2-FP-NEXT: vmovups %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3509; AVX2-FP-NEXT: vshufps {{.*#+}} ymm13 = ymm2[0,2,3,3,4,6,7,7]3510; AVX2-FP-NEXT: vshufps {{.*#+}} ymm15 = ymm3[2,3,2,3,6,7,6,7]3511; AVX2-FP-NEXT: vblendps {{.*#+}} ymm13 = ymm15[0],ymm13[1,2],ymm15[3,4],ymm13[5,6],ymm15[7]3512; AVX2-FP-NEXT: vunpckhps {{.*#+}} ymm15 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7]3513; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm15 = ymm15[3,3,3,3]3514; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm13 = ymm13[2,1,3,3]3515; AVX2-FP-NEXT: vblendps {{.*#+}} ymm13 = ymm13[0,1,2],ymm15[3,4],ymm13[5,6,7]3516; AVX2-FP-NEXT: vbroadcastsd 120(%r8), %ymm153517; AVX2-FP-NEXT: vblendps {{.*#+}} ymm13 = ymm13[0,1],ymm15[2],ymm13[3,4,5,6],ymm15[7]3518; AVX2-FP-NEXT: vshufps {{.*#+}} ymm12 = ymm12[3,0,2,3,7,4,6,7]3519; AVX2-FP-NEXT: vblendps $16, {{[-0-9]+}}(%r{{[sb]}}p), %ymm12, %ymm12 # 32-byte Folded Reload3520; AVX2-FP-NEXT: # ymm12 = ymm12[0,1,2,3],mem[4],ymm12[5,6,7]3521; AVX2-FP-NEXT: vshufps {{.*#+}} ymm14 = ymm14[0,1,3,0,4,5,7,4]3522; AVX2-FP-NEXT: vpermilps $78, (%rsp), %ymm15 # 32-byte Folded Reload3523; AVX2-FP-NEXT: # ymm15 = mem[2,3,0,1,6,7,4,5]3524; AVX2-FP-NEXT: vblendps {{.*#+}} ymm14 = ymm15[0,1],ymm14[2],ymm15[3,4,5,6],ymm14[7]3525; AVX2-FP-NEXT: vblendps {{.*#+}} ymm12 = ymm12[0],ymm14[1,2,3],ymm12[4,5],ymm14[6,7]3526; AVX2-FP-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm12, %ymm12 # 32-byte Folded Reload3527; AVX2-FP-NEXT: # ymm12 = ymm12[0,1,2],mem[3],ymm12[4,5,6,7]3528; AVX2-FP-NEXT: vshufps {{.*#+}} ymm8 = ymm8[3,0,2,3,7,4,6,7]3529; AVX2-FP-NEXT: vblendps {{.*#+}} ymm8 = ymm8[0,1,2,3],ymm9[4],ymm8[5,6,7]3530; AVX2-FP-NEXT: vshufps {{.*#+}} ymm9 = ymm10[0,1,3,0,4,5,7,4]3531; AVX2-FP-NEXT: vshufps {{.*#+}} ymm10 = ymm11[2,3,0,1,6,7,4,5]3532; AVX2-FP-NEXT: vblendps {{.*#+}} ymm9 = ymm10[0,1],ymm9[2],ymm10[3,4,5,6],ymm9[7]3533; AVX2-FP-NEXT: vblendps {{.*#+}} ymm8 = ymm8[0],ymm9[1,2,3],ymm8[4,5],ymm9[6,7]3534; AVX2-FP-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm8, %ymm8 # 32-byte Folded Reload3535; AVX2-FP-NEXT: # ymm8 = ymm8[0,1,2],mem[3],ymm8[4,5,6,7]3536; AVX2-FP-NEXT: vshufps {{.*#+}} ymm4 = ymm4[3,0,2,3,7,4,6,7]3537; AVX2-FP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm5[4],ymm4[5,6,7]3538; AVX2-FP-NEXT: vshufps {{.*#+}} ymm5 = ymm6[0,1,3,0,4,5,7,4]3539; AVX2-FP-NEXT: vshufps {{.*#+}} ymm6 = ymm7[2,3,0,1,6,7,4,5]3540; AVX2-FP-NEXT: vblendps {{.*#+}} ymm5 = ymm6[0,1],ymm5[2],ymm6[3,4,5,6],ymm5[7]3541; AVX2-FP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0],ymm5[1,2,3],ymm4[4,5],ymm5[6,7]3542; AVX2-FP-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4, %ymm4 # 32-byte Folded Reload3543; AVX2-FP-NEXT: # ymm4 = ymm4[0,1,2],mem[3],ymm4[4,5,6,7]3544; AVX2-FP-NEXT: vshufps {{.*#+}} ymm0 = ymm0[3,0,2,3,7,4,6,7]3545; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4],ymm0[5,6,7]3546; AVX2-FP-NEXT: vshufps {{.*#+}} ymm1 = ymm2[0,1,3,0,4,5,7,4]3547; AVX2-FP-NEXT: vshufps {{.*#+}} ymm2 = ymm3[2,3,0,1,6,7,4,5]3548; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1],ymm1[2],ymm2[3,4,5,6],ymm1[7]3549; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm1[1,2,3],ymm0[4,5],ymm1[6,7]3550; AVX2-FP-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload3551; AVX2-FP-NEXT: # ymm0 = ymm0[0,1,2],mem[3],ymm0[4,5,6,7]3552; AVX2-FP-NEXT: vmovaps %ymm0, 544(%r9)3553; AVX2-FP-NEXT: vmovaps %ymm4, 384(%r9)3554; AVX2-FP-NEXT: vmovaps %ymm8, 224(%r9)3555; AVX2-FP-NEXT: vmovaps %ymm12, 64(%r9)3556; AVX2-FP-NEXT: vmovaps %ymm13, 608(%r9)3557; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3558; AVX2-FP-NEXT: vmovaps %ymm0, 576(%r9)3559; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3560; AVX2-FP-NEXT: vmovaps %ymm0, 448(%r9)3561; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3562; AVX2-FP-NEXT: vmovaps %ymm0, 416(%r9)3563; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3564; AVX2-FP-NEXT: vmovaps %ymm0, 288(%r9)3565; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3566; AVX2-FP-NEXT: vmovaps %ymm0, 256(%r9)3567; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3568; AVX2-FP-NEXT: vmovaps %ymm0, 128(%r9)3569; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3570; AVX2-FP-NEXT: vmovaps %ymm0, 96(%r9)3571; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3572; AVX2-FP-NEXT: vmovaps %ymm0, 512(%r9)3573; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3574; AVX2-FP-NEXT: vmovaps %ymm0, 480(%r9)3575; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3576; AVX2-FP-NEXT: vmovaps %ymm0, 352(%r9)3577; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3578; AVX2-FP-NEXT: vmovaps %ymm0, 320(%r9)3579; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3580; AVX2-FP-NEXT: vmovaps %ymm0, 192(%r9)3581; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3582; AVX2-FP-NEXT: vmovaps %ymm0, 160(%r9)3583; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3584; AVX2-FP-NEXT: vmovaps %ymm0, 32(%r9)3585; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3586; AVX2-FP-NEXT: vmovaps %ymm0, (%r9)3587; AVX2-FP-NEXT: addq $552, %rsp # imm = 0x2283588; AVX2-FP-NEXT: vzeroupper3589; AVX2-FP-NEXT: retq3590;3591; AVX2-FCP-LABEL: store_i32_stride5_vf32:3592; AVX2-FCP: # %bb.0:3593; AVX2-FCP-NEXT: subq $600, %rsp # imm = 0x2583594; AVX2-FCP-NEXT: vmovaps (%r8), %ymm113595; AVX2-FCP-NEXT: vmovups %ymm11, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3596; AVX2-FCP-NEXT: vmovaps 32(%r8), %ymm103597; AVX2-FCP-NEXT: vmovups %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3598; AVX2-FCP-NEXT: vmovaps (%rsi), %xmm63599; AVX2-FCP-NEXT: vmovaps 32(%rsi), %xmm43600; AVX2-FCP-NEXT: vmovaps 64(%rsi), %xmm153601; AVX2-FCP-NEXT: vmovaps (%rdi), %xmm73602; AVX2-FCP-NEXT: vmovaps 32(%rdi), %xmm53603; AVX2-FCP-NEXT: vmovaps 64(%rdi), %xmm133604; AVX2-FCP-NEXT: vmovaps %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3605; AVX2-FCP-NEXT: vunpckhps {{.*#+}} xmm0 = xmm7[2],xmm6[2],xmm7[3],xmm6[3]3606; AVX2-FCP-NEXT: vmovaps {{.*#+}} ymm12 = [0,1,0,1,u,u,2,2]3607; AVX2-FCP-NEXT: vpermps %ymm0, %ymm12, %ymm03608; AVX2-FCP-NEXT: vmovaps (%rdx), %xmm83609; AVX2-FCP-NEXT: vmovaps 32(%rdx), %xmm23610; AVX2-FCP-NEXT: vmovaps (%rcx), %xmm93611; AVX2-FCP-NEXT: vmovaps 32(%rcx), %xmm33612; AVX2-FCP-NEXT: vshufps {{.*#+}} xmm1 = xmm9[1,1,2,2]3613; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],xmm8[2],xmm1[3]3614; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,1,1,3]3615; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3],ymm1[4,5],ymm0[6,7]3616; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm1 = ymm11[0,1,2,1]3617; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2,3,4,5],ymm1[6],ymm0[7]3618; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3619; AVX2-FCP-NEXT: vunpckhps {{.*#+}} xmm0 = xmm5[2],xmm4[2],xmm5[3],xmm4[3]3620; AVX2-FCP-NEXT: vpermps %ymm0, %ymm12, %ymm03621; AVX2-FCP-NEXT: vshufps {{.*#+}} xmm1 = xmm3[1,1,2,2]3622; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],xmm2[2],xmm1[3]3623; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,1,1,3]3624; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3],ymm1[4,5],ymm0[6,7]3625; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm1 = ymm10[0,1,2,1]3626; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2,3,4,5],ymm1[6],ymm0[7]3627; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3628; AVX2-FCP-NEXT: vunpckhps {{.*#+}} xmm0 = xmm13[2],xmm15[2],xmm13[3],xmm15[3]3629; AVX2-FCP-NEXT: vpermps %ymm0, %ymm12, %ymm103630; AVX2-FCP-NEXT: vmovaps 64(%rdx), %xmm13631; AVX2-FCP-NEXT: vmovaps 64(%rcx), %xmm03632; AVX2-FCP-NEXT: vshufps {{.*#+}} xmm11 = xmm0[1,1,2,2]3633; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm11 = xmm11[0,1],xmm1[2],xmm11[3]3634; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm11 = ymm11[0,1,1,3]3635; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm10 = ymm11[0,1],ymm10[2,3],ymm11[4,5],ymm10[6,7]3636; AVX2-FCP-NEXT: vmovaps 64(%r8), %ymm113637; AVX2-FCP-NEXT: vmovups %ymm11, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3638; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm11 = ymm11[0,1,2,1]3639; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm10 = ymm10[0],ymm11[1],ymm10[2,3,4,5],ymm11[6],ymm10[7]3640; AVX2-FCP-NEXT: vmovups %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3641; AVX2-FCP-NEXT: vmovaps 96(%rsi), %xmm113642; AVX2-FCP-NEXT: vmovaps %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3643; AVX2-FCP-NEXT: vmovaps 96(%rdi), %xmm103644; AVX2-FCP-NEXT: vmovaps %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3645; AVX2-FCP-NEXT: vunpckhps {{.*#+}} xmm10 = xmm10[2],xmm11[2],xmm10[3],xmm11[3]3646; AVX2-FCP-NEXT: vpermps %ymm10, %ymm12, %ymm103647; AVX2-FCP-NEXT: vmovaps 96(%rdx), %xmm143648; AVX2-FCP-NEXT: vmovaps %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3649; AVX2-FCP-NEXT: vmovaps 96(%rcx), %xmm133650; AVX2-FCP-NEXT: vshufps {{.*#+}} xmm11 = xmm13[1,1,2,2]3651; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm11 = xmm11[0,1],xmm14[2],xmm11[3]3652; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm11 = ymm11[0,1,1,3]3653; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm10 = ymm11[0,1],ymm10[2,3],ymm11[4,5],ymm10[6,7]3654; AVX2-FCP-NEXT: vmovaps 96(%r8), %ymm113655; AVX2-FCP-NEXT: vmovups %ymm11, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3656; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm11 = ymm11[0,1,2,1]3657; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm10 = ymm10[0],ymm11[1],ymm10[2,3,4,5],ymm11[6],ymm10[7]3658; AVX2-FCP-NEXT: vmovups %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3659; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm8 = xmm8[0],xmm9[0],xmm8[1],xmm9[1]3660; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm6 = xmm7[0],xmm6[0],xmm7[1],xmm6[1]3661; AVX2-FCP-NEXT: vpermps %ymm8, %ymm12, %ymm73662; AVX2-FCP-NEXT: vmovaps {{.*#+}} ymm11 = [0,1,3,2,3,2,3,2]3663; AVX2-FCP-NEXT: vpermps %ymm6, %ymm11, %ymm83664; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm7 = ymm8[0,1],ymm7[2,3],ymm8[4,5,6],ymm7[7]3665; AVX2-FCP-NEXT: vinsertf128 $1, (%r8), %ymm6, %ymm63666; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0],ymm7[1,2,3],ymm6[4],ymm7[5,6,7]3667; AVX2-FCP-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3668; AVX2-FCP-NEXT: vmovaps (%rdx), %ymm73669; AVX2-FCP-NEXT: vmovaps (%rcx), %ymm93670; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm6 = ymm9[1,1,1,1,5,5,5,5]3671; AVX2-FCP-NEXT: vmovups %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3672; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0],ymm7[1],ymm6[2,3,4],ymm7[5],ymm6[6,7]3673; AVX2-FCP-NEXT: vmovaps %ymm7, %ymm83674; AVX2-FCP-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3675; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm6 = ymm6[2,2,2,2]3676; AVX2-FCP-NEXT: vmovaps (%rdi), %ymm103677; AVX2-FCP-NEXT: vmovups %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3678; AVX2-FCP-NEXT: vmovaps (%rsi), %ymm143679; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm7 = ymm14[1,1,2,2,5,5,6,6]3680; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0],ymm10[1,2],ymm7[3,4],ymm10[5,6],ymm7[7]3681; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm7 = ymm7[2,2,2,3]3682; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm6 = ymm7[0,1,2],ymm6[3,4],ymm7[5,6,7]3683; AVX2-FCP-NEXT: vbroadcastsd 16(%r8), %ymm73684; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm6 = ymm7[0],ymm6[1,2,3,4],ymm7[5],ymm6[6,7]3685; AVX2-FCP-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3686; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm6 = ymm9[0,2,3,3,4,6,7,7]3687; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm7 = ymm8[2,3,2,3,6,7,6,7]3688; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm6 = ymm7[0],ymm6[1,2],ymm7[3,4],ymm6[5,6],ymm7[7]3689; AVX2-FCP-NEXT: vunpckhps {{.*#+}} ymm7 = ymm14[2],ymm10[2],ymm14[3],ymm10[3],ymm14[6],ymm10[6],ymm14[7],ymm10[7]3690; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm7 = ymm7[3,3,3,3]3691; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm6 = ymm6[2,1,3,3]3692; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0,1,2],ymm7[3,4],ymm6[5,6,7]3693; AVX2-FCP-NEXT: vbroadcastsd 24(%r8), %ymm73694; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0,1],ymm7[2],ymm6[3,4,5,6],ymm7[7]3695; AVX2-FCP-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3696; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1]3697; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm3 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]3698; AVX2-FCP-NEXT: vpermps %ymm2, %ymm12, %ymm23699; AVX2-FCP-NEXT: vpermps %ymm3, %ymm11, %ymm43700; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm4[0,1],ymm2[2,3],ymm4[4,5,6],ymm2[7]3701; AVX2-FCP-NEXT: vinsertf128 $1, 32(%r8), %ymm3, %ymm33702; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm3[0],ymm2[1,2,3],ymm3[4],ymm2[5,6,7]3703; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3704; AVX2-FCP-NEXT: vmovaps 32(%rdx), %ymm103705; AVX2-FCP-NEXT: vmovaps 32(%rcx), %ymm93706; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm2 = ymm9[1,1,1,1,5,5,5,5]3707; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0],ymm10[1],ymm2[2,3,4],ymm10[5],ymm2[6,7]3708; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[2,2,2,2]3709; AVX2-FCP-NEXT: vmovaps 32(%rdi), %ymm83710; AVX2-FCP-NEXT: vmovaps 32(%rsi), %ymm73711; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm3 = ymm7[1,1,2,2,5,5,6,6]3712; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0],ymm8[1,2],ymm3[3,4],ymm8[5,6],ymm3[7]3713; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm3 = ymm3[2,2,2,3]3714; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm3[0,1,2],ymm2[3,4],ymm3[5,6,7]3715; AVX2-FCP-NEXT: vbroadcastsd 48(%r8), %ymm33716; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm3[0],ymm2[1,2,3,4],ymm3[5],ymm2[6,7]3717; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3718; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm2 = ymm9[0,2,3,3,4,6,7,7]3719; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm3 = ymm10[2,3,2,3,6,7,6,7]3720; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm3[0],ymm2[1,2],ymm3[3,4],ymm2[5,6],ymm3[7]3721; AVX2-FCP-NEXT: vunpckhps {{.*#+}} ymm3 = ymm7[2],ymm8[2],ymm7[3],ymm8[3],ymm7[6],ymm8[6],ymm7[7],ymm8[7]3722; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm3 = ymm3[3,3,3,3]3723; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[2,1,3,3]3724; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2],ymm3[3,4],ymm2[5,6,7]3725; AVX2-FCP-NEXT: vbroadcastsd 56(%r8), %ymm33726; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1],ymm3[2],ymm2[3,4,5,6],ymm3[7]3727; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3728; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]3729; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload3730; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm1[0],xmm15[0],xmm1[1],xmm15[1]3731; AVX2-FCP-NEXT: vpermps %ymm0, %ymm12, %ymm03732; AVX2-FCP-NEXT: vpermps %ymm1, %ymm11, %ymm23733; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1],ymm0[2,3],ymm2[4,5,6],ymm0[7]3734; AVX2-FCP-NEXT: vinsertf128 $1, 64(%r8), %ymm1, %ymm13735; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2,3],ymm1[4],ymm0[5,6,7]3736; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3737; AVX2-FCP-NEXT: vmovaps 64(%rdx), %ymm53738; AVX2-FCP-NEXT: vmovaps 64(%rcx), %ymm43739; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm0 = ymm4[1,1,1,1,5,5,5,5]3740; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm5[1],ymm0[2,3,4],ymm5[5],ymm0[6,7]3741; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2]3742; AVX2-FCP-NEXT: vmovaps 64(%rdi), %ymm33743; AVX2-FCP-NEXT: vmovaps 64(%rsi), %ymm23744; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm1 = ymm2[1,1,2,2,5,5,6,6]3745; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm3[1,2],ymm1[3,4],ymm3[5,6],ymm1[7]3746; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,2,2,3]3747; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4],ymm1[5,6,7]3748; AVX2-FCP-NEXT: vbroadcastsd 80(%r8), %ymm13749; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2,3,4],ymm1[5],ymm0[6,7]3750; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3751; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm0 = ymm4[0,2,3,3,4,6,7,7]3752; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm1 = ymm5[2,3,2,3,6,7,6,7]3753; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2],ymm1[3,4],ymm0[5,6],ymm1[7]3754; AVX2-FCP-NEXT: vunpckhps {{.*#+}} ymm1 = ymm2[2],ymm3[2],ymm2[3],ymm3[3],ymm2[6],ymm3[6],ymm2[7],ymm3[7]3755; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[3,3,3,3]3756; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,1,3,3]3757; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3,4],ymm0[5,6,7]3758; AVX2-FCP-NEXT: vbroadcastsd 88(%r8), %ymm13759; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4,5,6],ymm1[7]3760; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3761; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3762; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm0 = xmm0[0],xmm13[0],xmm0[1],xmm13[1]3763; AVX2-FCP-NEXT: vpermps %ymm0, %ymm12, %ymm03764; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload3765; AVX2-FCP-NEXT: vunpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload3766; AVX2-FCP-NEXT: # xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]3767; AVX2-FCP-NEXT: vpermps %ymm1, %ymm11, %ymm63768; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm6[0,1],ymm0[2,3],ymm6[4,5,6],ymm0[7]3769; AVX2-FCP-NEXT: vinsertf128 $1, 96(%r8), %ymm1, %ymm13770; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2,3],ymm1[4],ymm0[5,6,7]3771; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3772; AVX2-FCP-NEXT: vmovaps 96(%rdx), %ymm123773; AVX2-FCP-NEXT: vmovaps 96(%rcx), %ymm133774; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm0 = ymm13[1,1,1,1,5,5,5,5]3775; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm12[1],ymm0[2,3,4],ymm12[5],ymm0[6,7]3776; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm6 = ymm0[2,2,2,2]3777; AVX2-FCP-NEXT: vmovaps 96(%rdi), %ymm13778; AVX2-FCP-NEXT: vmovaps 96(%rsi), %ymm03779; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm11 = ymm0[1,1,2,2,5,5,6,6]3780; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm11 = ymm11[0],ymm1[1,2],ymm11[3,4],ymm1[5,6],ymm11[7]3781; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm11 = ymm11[2,2,2,3]3782; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm6 = ymm11[0,1,2],ymm6[3,4],ymm11[5,6,7]3783; AVX2-FCP-NEXT: vbroadcastsd 112(%r8), %ymm113784; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm6 = ymm11[0],ymm6[1,2,3,4],ymm11[5],ymm6[6,7]3785; AVX2-FCP-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3786; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm6 = ymm13[0,2,3,3,4,6,7,7]3787; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm15 = ymm12[2,3,2,3,6,7,6,7]3788; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm6 = ymm15[0],ymm6[1,2],ymm15[3,4],ymm6[5,6],ymm15[7]3789; AVX2-FCP-NEXT: vunpckhps {{.*#+}} ymm15 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7]3790; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm15 = ymm15[3,3,3,3]3791; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm6 = ymm6[2,1,3,3]3792; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0,1,2],ymm15[3,4],ymm6[5,6,7]3793; AVX2-FCP-NEXT: vbroadcastsd 120(%r8), %ymm153794; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0,1],ymm15[2],ymm6[3,4,5,6],ymm15[7]3795; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm14 = ymm14[3,0,2,3,7,4,6,7]3796; AVX2-FCP-NEXT: vblendps $16, {{[-0-9]+}}(%r{{[sb]}}p), %ymm14, %ymm14 # 32-byte Folded Reload3797; AVX2-FCP-NEXT: # ymm14 = ymm14[0,1,2,3],mem[4],ymm14[5,6,7]3798; AVX2-FCP-NEXT: vpermilps $52, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Folded Reload3799; AVX2-FCP-NEXT: # ymm15 = mem[0,1,3,0,4,5,7,4]3800; AVX2-FCP-NEXT: vpermilps $78, {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Folded Reload3801; AVX2-FCP-NEXT: # ymm11 = mem[2,3,0,1,6,7,4,5]3802; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm11 = ymm11[0,1],ymm15[2],ymm11[3,4,5,6],ymm15[7]3803; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm11 = ymm14[0],ymm11[1,2,3],ymm14[4,5],ymm11[6,7]3804; AVX2-FCP-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm11, %ymm11 # 32-byte Folded Reload3805; AVX2-FCP-NEXT: # ymm11 = ymm11[0,1,2],mem[3],ymm11[4,5,6,7]3806; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm7 = ymm7[3,0,2,3,7,4,6,7]3807; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0,1,2,3],ymm8[4],ymm7[5,6,7]3808; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm8 = ymm9[0,1,3,0,4,5,7,4]3809; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm9 = ymm10[2,3,0,1,6,7,4,5]3810; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm8 = ymm9[0,1],ymm8[2],ymm9[3,4,5,6],ymm8[7]3811; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0],ymm8[1,2,3],ymm7[4,5],ymm8[6,7]3812; AVX2-FCP-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm7 # 32-byte Folded Reload3813; AVX2-FCP-NEXT: # ymm7 = ymm7[0,1,2],mem[3],ymm7[4,5,6,7]3814; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm2 = ymm2[3,0,2,3,7,4,6,7]3815; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm3[4],ymm2[5,6,7]3816; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm3 = ymm4[0,1,3,0,4,5,7,4]3817; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm4 = ymm5[2,3,0,1,6,7,4,5]3818; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm3 = ymm4[0,1],ymm3[2],ymm4[3,4,5,6],ymm3[7]3819; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0],ymm3[1,2,3],ymm2[4,5],ymm3[6,7]3820; AVX2-FCP-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload3821; AVX2-FCP-NEXT: # ymm2 = ymm2[0,1,2],mem[3],ymm2[4,5,6,7]3822; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm0 = ymm0[3,0,2,3,7,4,6,7]3823; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4],ymm0[5,6,7]3824; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm1 = ymm13[0,1,3,0,4,5,7,4]3825; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm3 = ymm12[2,3,0,1,6,7,4,5]3826; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm3[0,1],ymm1[2],ymm3[3,4,5,6],ymm1[7]3827; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm1[1,2,3],ymm0[4,5],ymm1[6,7]3828; AVX2-FCP-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload3829; AVX2-FCP-NEXT: # ymm0 = ymm0[0,1,2],mem[3],ymm0[4,5,6,7]3830; AVX2-FCP-NEXT: vmovaps %ymm0, 544(%r9)3831; AVX2-FCP-NEXT: vmovaps %ymm2, 384(%r9)3832; AVX2-FCP-NEXT: vmovaps %ymm7, 224(%r9)3833; AVX2-FCP-NEXT: vmovaps %ymm11, 64(%r9)3834; AVX2-FCP-NEXT: vmovaps %ymm6, 608(%r9)3835; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3836; AVX2-FCP-NEXT: vmovaps %ymm0, 576(%r9)3837; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3838; AVX2-FCP-NEXT: vmovaps %ymm0, 480(%r9)3839; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3840; AVX2-FCP-NEXT: vmovaps %ymm0, 448(%r9)3841; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3842; AVX2-FCP-NEXT: vmovaps %ymm0, 416(%r9)3843; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3844; AVX2-FCP-NEXT: vmovaps %ymm0, 320(%r9)3845; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3846; AVX2-FCP-NEXT: vmovaps %ymm0, 288(%r9)3847; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3848; AVX2-FCP-NEXT: vmovaps %ymm0, 256(%r9)3849; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3850; AVX2-FCP-NEXT: vmovaps %ymm0, 160(%r9)3851; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3852; AVX2-FCP-NEXT: vmovaps %ymm0, 128(%r9)3853; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3854; AVX2-FCP-NEXT: vmovaps %ymm0, 96(%r9)3855; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3856; AVX2-FCP-NEXT: vmovaps %ymm0, (%r9)3857; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3858; AVX2-FCP-NEXT: vmovaps %ymm0, 512(%r9)3859; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3860; AVX2-FCP-NEXT: vmovaps %ymm0, 352(%r9)3861; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3862; AVX2-FCP-NEXT: vmovaps %ymm0, 192(%r9)3863; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3864; AVX2-FCP-NEXT: vmovaps %ymm0, 32(%r9)3865; AVX2-FCP-NEXT: addq $600, %rsp # imm = 0x2583866; AVX2-FCP-NEXT: vzeroupper3867; AVX2-FCP-NEXT: retq3868;3869; AVX512-LABEL: store_i32_stride5_vf32:3870; AVX512: # %bb.0:3871; AVX512-NEXT: vmovdqa64 (%rdi), %zmm23872; AVX512-NEXT: vmovdqa64 64(%rdi), %zmm03873; AVX512-NEXT: vmovdqa64 (%rsi), %zmm13874; AVX512-NEXT: vmovdqa64 64(%rsi), %zmm113875; AVX512-NEXT: vmovdqa64 (%rdx), %zmm43876; AVX512-NEXT: vmovdqa64 64(%rdx), %zmm123877; AVX512-NEXT: vmovdqa64 (%rcx), %zmm63878; AVX512-NEXT: vmovdqa64 64(%rcx), %zmm143879; AVX512-NEXT: vmovdqa64 (%r8), %zmm53880; AVX512-NEXT: vmovdqa64 64(%r8), %zmm133881; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm16 = [17,0,0,16,2,18,0,1,17,0,0,16,2,18,0,1]3882; AVX512-NEXT: # zmm16 = mem[0,1,2,3,0,1,2,3]3883; AVX512-NEXT: vmovdqa64 %zmm4, %zmm73884; AVX512-NEXT: vpermt2d %zmm6, %zmm16, %zmm73885; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm17 = [0,16,2,18,0,1,17,3,0,16,2,18,0,1,17,3]3886; AVX512-NEXT: # zmm17 = mem[0,1,2,3,0,1,2,3]3887; AVX512-NEXT: vmovdqa64 %zmm2, %zmm33888; AVX512-NEXT: vpermt2d %zmm1, %zmm17, %zmm33889; AVX512-NEXT: movw $12684, %ax # imm = 0x318C3890; AVX512-NEXT: kmovw %eax, %k13891; AVX512-NEXT: vmovdqa32 %zmm7, %zmm3 {%k1}3892; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm18 = [0,1,2,3,16,5,6,7,8,17,10,11,12,13,18,15]3893; AVX512-NEXT: vpermt2d %zmm5, %zmm18, %zmm33894; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm19 = [14,30,0,13,29,15,31,0,14,30,0,13,29,15,31,0]3895; AVX512-NEXT: # zmm19 = mem[0,1,2,3,0,1,2,3]3896; AVX512-NEXT: vmovdqa64 %zmm12, %zmm93897; AVX512-NEXT: vpermt2d %zmm14, %zmm19, %zmm93898; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [0,13,29,15,31,0,14,30,0,13,29,15,31,0,14,30]3899; AVX512-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]3900; AVX512-NEXT: vmovdqa64 %zmm0, %zmm73901; AVX512-NEXT: vpermt2d %zmm11, %zmm8, %zmm73902; AVX512-NEXT: movw $25368, %ax # imm = 0x63183903; AVX512-NEXT: kmovw %eax, %k23904; AVX512-NEXT: vmovdqa32 %zmm9, %zmm7 {%k2}3905; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm20 = [28,1,2,3,4,29,6,7,8,9,30,11,12,13,14,31]3906; AVX512-NEXT: vpermt2d %zmm13, %zmm20, %zmm73907; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm21 = [27,0,10,26,12,28,0,11,27,0,10,26,12,28,0,11]3908; AVX512-NEXT: # zmm21 = mem[0,1,2,3,0,1,2,3]3909; AVX512-NEXT: vmovdqa64 %zmm0, %zmm153910; AVX512-NEXT: vpermt2d %zmm11, %zmm21, %zmm153911; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm10 = [9,27,11,0,26,10,28,12,9,27,11,0,26,10,28,12]3912; AVX512-NEXT: # zmm10 = mem[0,1,2,3,0,1,2,3]3913; AVX512-NEXT: vmovdqa64 %zmm14, %zmm93914; AVX512-NEXT: vpermt2d %zmm12, %zmm10, %zmm93915; AVX512-NEXT: vmovdqa32 %zmm15, %zmm9 {%k1}3916; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm22 = [0,25,2,3,4,5,26,7,8,9,10,27,12,13,14,15]3917; AVX512-NEXT: vpermt2d %zmm13, %zmm22, %zmm93918; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm23 = [8,24,0,7,23,9,25,0,8,24,0,7,23,9,25,0]3919; AVX512-NEXT: # zmm23 = mem[0,1,2,3,0,1,2,3]3920; AVX512-NEXT: vmovdqa64 %zmm0, %zmm243921; AVX512-NEXT: vpermt2d %zmm11, %zmm23, %zmm243922; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm25 = [6,22,8,24,0,7,23,9,6,22,8,24,0,7,23,9]3923; AVX512-NEXT: # zmm25 = mem[0,1,2,3,0,1,2,3]3924; AVX512-NEXT: vmovdqa64 %zmm12, %zmm153925; AVX512-NEXT: vpermt2d %zmm14, %zmm25, %zmm153926; AVX512-NEXT: vmovdqa32 %zmm24, %zmm15 {%k2}3927; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm24 = [0,1,22,3,4,5,6,23,8,9,10,11,24,13,14,15]3928; AVX512-NEXT: vpermt2d %zmm13, %zmm24, %zmm153929; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm26 = [0,3,19,5,21,0,4,20,0,3,19,5,21,0,4,20]3930; AVX512-NEXT: # zmm26 = mem[0,1,2,3,0,1,2,3]3931; AVX512-NEXT: vmovdqa64 %zmm12, %zmm273932; AVX512-NEXT: vpermt2d %zmm14, %zmm26, %zmm273933; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm28 = [3,21,5,0,20,4,22,6,3,21,5,0,20,4,22,6]3934; AVX512-NEXT: # zmm28 = mem[0,1,2,3,0,1,2,3]3935; AVX512-NEXT: vmovdqa64 %zmm11, %zmm293936; AVX512-NEXT: vpermt2d %zmm0, %zmm28, %zmm293937; AVX512-NEXT: movw $6342, %ax # imm = 0x18C63938; AVX512-NEXT: kmovw %eax, %k33939; AVX512-NEXT: vmovdqa32 %zmm27, %zmm29 {%k3}3940; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm27 = [0,1,2,19,4,5,6,7,20,9,10,11,12,21,14,15]3941; AVX512-NEXT: vpermt2d %zmm13, %zmm27, %zmm293942; AVX512-NEXT: vpermt2d %zmm14, %zmm16, %zmm123943; AVX512-NEXT: vpermt2d %zmm11, %zmm17, %zmm03944; AVX512-NEXT: vmovdqa32 %zmm12, %zmm0 {%k1}3945; AVX512-NEXT: vpermt2d %zmm13, %zmm18, %zmm03946; AVX512-NEXT: vpermi2d %zmm6, %zmm4, %zmm193947; AVX512-NEXT: vpermi2d %zmm1, %zmm2, %zmm83948; AVX512-NEXT: vmovdqa32 %zmm19, %zmm8 {%k2}3949; AVX512-NEXT: vpermt2d %zmm5, %zmm20, %zmm83950; AVX512-NEXT: vpermi2d %zmm1, %zmm2, %zmm213951; AVX512-NEXT: vpermi2d %zmm4, %zmm6, %zmm103952; AVX512-NEXT: vmovdqa32 %zmm21, %zmm10 {%k1}3953; AVX512-NEXT: vpermt2d %zmm5, %zmm22, %zmm103954; AVX512-NEXT: vpermi2d %zmm1, %zmm2, %zmm233955; AVX512-NEXT: vpermi2d %zmm6, %zmm4, %zmm253956; AVX512-NEXT: vmovdqa32 %zmm23, %zmm25 {%k2}3957; AVX512-NEXT: vpermt2d %zmm5, %zmm24, %zmm253958; AVX512-NEXT: vpermt2d %zmm6, %zmm26, %zmm43959; AVX512-NEXT: vpermt2d %zmm2, %zmm28, %zmm13960; AVX512-NEXT: vmovdqa32 %zmm4, %zmm1 {%k3}3961; AVX512-NEXT: vpermt2d %zmm5, %zmm27, %zmm13962; AVX512-NEXT: vmovdqa64 %zmm1, 64(%r9)3963; AVX512-NEXT: vmovdqa64 %zmm25, 128(%r9)3964; AVX512-NEXT: vmovdqa64 %zmm10, 192(%r9)3965; AVX512-NEXT: vmovdqa64 %zmm8, 256(%r9)3966; AVX512-NEXT: vmovdqa64 %zmm0, 320(%r9)3967; AVX512-NEXT: vmovdqa64 %zmm29, 384(%r9)3968; AVX512-NEXT: vmovdqa64 %zmm15, 448(%r9)3969; AVX512-NEXT: vmovdqa64 %zmm9, 512(%r9)3970; AVX512-NEXT: vmovdqa64 %zmm7, 576(%r9)3971; AVX512-NEXT: vmovdqa64 %zmm3, (%r9)3972; AVX512-NEXT: vzeroupper3973; AVX512-NEXT: retq3974;3975; AVX512-FCP-LABEL: store_i32_stride5_vf32:3976; AVX512-FCP: # %bb.0:3977; AVX512-FCP-NEXT: vmovdqa64 (%rdi), %zmm23978; AVX512-FCP-NEXT: vmovdqa64 64(%rdi), %zmm03979; AVX512-FCP-NEXT: vmovdqa64 (%rsi), %zmm13980; AVX512-FCP-NEXT: vmovdqa64 64(%rsi), %zmm113981; AVX512-FCP-NEXT: vmovdqa64 (%rdx), %zmm43982; AVX512-FCP-NEXT: vmovdqa64 64(%rdx), %zmm123983; AVX512-FCP-NEXT: vmovdqa64 (%rcx), %zmm63984; AVX512-FCP-NEXT: vmovdqa64 64(%rcx), %zmm143985; AVX512-FCP-NEXT: vmovdqa64 (%r8), %zmm53986; AVX512-FCP-NEXT: vmovdqa64 64(%r8), %zmm133987; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm16 = [17,0,0,16,2,18,0,1,17,0,0,16,2,18,0,1]3988; AVX512-FCP-NEXT: # zmm16 = mem[0,1,2,3,0,1,2,3]3989; AVX512-FCP-NEXT: vmovdqa64 %zmm4, %zmm73990; AVX512-FCP-NEXT: vpermt2d %zmm6, %zmm16, %zmm73991; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm17 = [0,16,2,18,0,1,17,3,0,16,2,18,0,1,17,3]3992; AVX512-FCP-NEXT: # zmm17 = mem[0,1,2,3,0,1,2,3]3993; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm33994; AVX512-FCP-NEXT: vpermt2d %zmm1, %zmm17, %zmm33995; AVX512-FCP-NEXT: movw $12684, %ax # imm = 0x318C3996; AVX512-FCP-NEXT: kmovw %eax, %k13997; AVX512-FCP-NEXT: vmovdqa32 %zmm7, %zmm3 {%k1}3998; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm18 = [0,1,2,3,16,5,6,7,8,17,10,11,12,13,18,15]3999; AVX512-FCP-NEXT: vpermt2d %zmm5, %zmm18, %zmm34000; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm19 = [14,30,0,13,29,15,31,0,14,30,0,13,29,15,31,0]4001; AVX512-FCP-NEXT: # zmm19 = mem[0,1,2,3,0,1,2,3]4002; AVX512-FCP-NEXT: vmovdqa64 %zmm12, %zmm94003; AVX512-FCP-NEXT: vpermt2d %zmm14, %zmm19, %zmm94004; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [0,13,29,15,31,0,14,30,0,13,29,15,31,0,14,30]4005; AVX512-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]4006; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm74007; AVX512-FCP-NEXT: vpermt2d %zmm11, %zmm8, %zmm74008; AVX512-FCP-NEXT: movw $25368, %ax # imm = 0x63184009; AVX512-FCP-NEXT: kmovw %eax, %k24010; AVX512-FCP-NEXT: vmovdqa32 %zmm9, %zmm7 {%k2}4011; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm20 = [28,1,2,3,4,29,6,7,8,9,30,11,12,13,14,31]4012; AVX512-FCP-NEXT: vpermt2d %zmm13, %zmm20, %zmm74013; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm21 = [27,0,10,26,12,28,0,11,27,0,10,26,12,28,0,11]4014; AVX512-FCP-NEXT: # zmm21 = mem[0,1,2,3,0,1,2,3]4015; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm154016; AVX512-FCP-NEXT: vpermt2d %zmm11, %zmm21, %zmm154017; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm10 = [9,27,11,0,26,10,28,12,9,27,11,0,26,10,28,12]4018; AVX512-FCP-NEXT: # zmm10 = mem[0,1,2,3,0,1,2,3]4019; AVX512-FCP-NEXT: vmovdqa64 %zmm14, %zmm94020; AVX512-FCP-NEXT: vpermt2d %zmm12, %zmm10, %zmm94021; AVX512-FCP-NEXT: vmovdqa32 %zmm15, %zmm9 {%k1}4022; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm22 = [0,25,2,3,4,5,26,7,8,9,10,27,12,13,14,15]4023; AVX512-FCP-NEXT: vpermt2d %zmm13, %zmm22, %zmm94024; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm23 = [8,24,0,7,23,9,25,0,8,24,0,7,23,9,25,0]4025; AVX512-FCP-NEXT: # zmm23 = mem[0,1,2,3,0,1,2,3]4026; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm244027; AVX512-FCP-NEXT: vpermt2d %zmm11, %zmm23, %zmm244028; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm25 = [6,22,8,24,0,7,23,9,6,22,8,24,0,7,23,9]4029; AVX512-FCP-NEXT: # zmm25 = mem[0,1,2,3,0,1,2,3]4030; AVX512-FCP-NEXT: vmovdqa64 %zmm12, %zmm154031; AVX512-FCP-NEXT: vpermt2d %zmm14, %zmm25, %zmm154032; AVX512-FCP-NEXT: vmovdqa32 %zmm24, %zmm15 {%k2}4033; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm24 = [0,1,22,3,4,5,6,23,8,9,10,11,24,13,14,15]4034; AVX512-FCP-NEXT: vpermt2d %zmm13, %zmm24, %zmm154035; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm26 = [0,3,19,5,21,0,4,20,0,3,19,5,21,0,4,20]4036; AVX512-FCP-NEXT: # zmm26 = mem[0,1,2,3,0,1,2,3]4037; AVX512-FCP-NEXT: vmovdqa64 %zmm12, %zmm274038; AVX512-FCP-NEXT: vpermt2d %zmm14, %zmm26, %zmm274039; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm28 = [3,21,5,0,20,4,22,6,3,21,5,0,20,4,22,6]4040; AVX512-FCP-NEXT: # zmm28 = mem[0,1,2,3,0,1,2,3]4041; AVX512-FCP-NEXT: vmovdqa64 %zmm11, %zmm294042; AVX512-FCP-NEXT: vpermt2d %zmm0, %zmm28, %zmm294043; AVX512-FCP-NEXT: movw $6342, %ax # imm = 0x18C64044; AVX512-FCP-NEXT: kmovw %eax, %k34045; AVX512-FCP-NEXT: vmovdqa32 %zmm27, %zmm29 {%k3}4046; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm27 = [0,1,2,19,4,5,6,7,20,9,10,11,12,21,14,15]4047; AVX512-FCP-NEXT: vpermt2d %zmm13, %zmm27, %zmm294048; AVX512-FCP-NEXT: vpermt2d %zmm14, %zmm16, %zmm124049; AVX512-FCP-NEXT: vpermt2d %zmm11, %zmm17, %zmm04050; AVX512-FCP-NEXT: vmovdqa32 %zmm12, %zmm0 {%k1}4051; AVX512-FCP-NEXT: vpermt2d %zmm13, %zmm18, %zmm04052; AVX512-FCP-NEXT: vpermi2d %zmm6, %zmm4, %zmm194053; AVX512-FCP-NEXT: vpermi2d %zmm1, %zmm2, %zmm84054; AVX512-FCP-NEXT: vmovdqa32 %zmm19, %zmm8 {%k2}4055; AVX512-FCP-NEXT: vpermt2d %zmm5, %zmm20, %zmm84056; AVX512-FCP-NEXT: vpermi2d %zmm1, %zmm2, %zmm214057; AVX512-FCP-NEXT: vpermi2d %zmm4, %zmm6, %zmm104058; AVX512-FCP-NEXT: vmovdqa32 %zmm21, %zmm10 {%k1}4059; AVX512-FCP-NEXT: vpermt2d %zmm5, %zmm22, %zmm104060; AVX512-FCP-NEXT: vpermi2d %zmm1, %zmm2, %zmm234061; AVX512-FCP-NEXT: vpermi2d %zmm6, %zmm4, %zmm254062; AVX512-FCP-NEXT: vmovdqa32 %zmm23, %zmm25 {%k2}4063; AVX512-FCP-NEXT: vpermt2d %zmm5, %zmm24, %zmm254064; AVX512-FCP-NEXT: vpermt2d %zmm6, %zmm26, %zmm44065; AVX512-FCP-NEXT: vpermt2d %zmm2, %zmm28, %zmm14066; AVX512-FCP-NEXT: vmovdqa32 %zmm4, %zmm1 {%k3}4067; AVX512-FCP-NEXT: vpermt2d %zmm5, %zmm27, %zmm14068; AVX512-FCP-NEXT: vmovdqa64 %zmm1, 64(%r9)4069; AVX512-FCP-NEXT: vmovdqa64 %zmm25, 128(%r9)4070; AVX512-FCP-NEXT: vmovdqa64 %zmm10, 192(%r9)4071; AVX512-FCP-NEXT: vmovdqa64 %zmm8, 256(%r9)4072; AVX512-FCP-NEXT: vmovdqa64 %zmm0, 320(%r9)4073; AVX512-FCP-NEXT: vmovdqa64 %zmm29, 384(%r9)4074; AVX512-FCP-NEXT: vmovdqa64 %zmm15, 448(%r9)4075; AVX512-FCP-NEXT: vmovdqa64 %zmm9, 512(%r9)4076; AVX512-FCP-NEXT: vmovdqa64 %zmm7, 576(%r9)4077; AVX512-FCP-NEXT: vmovdqa64 %zmm3, (%r9)4078; AVX512-FCP-NEXT: vzeroupper4079; AVX512-FCP-NEXT: retq4080;4081; AVX512DQ-LABEL: store_i32_stride5_vf32:4082; AVX512DQ: # %bb.0:4083; AVX512DQ-NEXT: vmovdqa64 (%rdi), %zmm24084; AVX512DQ-NEXT: vmovdqa64 64(%rdi), %zmm04085; AVX512DQ-NEXT: vmovdqa64 (%rsi), %zmm14086; AVX512DQ-NEXT: vmovdqa64 64(%rsi), %zmm114087; AVX512DQ-NEXT: vmovdqa64 (%rdx), %zmm44088; AVX512DQ-NEXT: vmovdqa64 64(%rdx), %zmm124089; AVX512DQ-NEXT: vmovdqa64 (%rcx), %zmm64090; AVX512DQ-NEXT: vmovdqa64 64(%rcx), %zmm144091; AVX512DQ-NEXT: vmovdqa64 (%r8), %zmm54092; AVX512DQ-NEXT: vmovdqa64 64(%r8), %zmm134093; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm16 = [17,0,0,16,2,18,0,1,17,0,0,16,2,18,0,1]4094; AVX512DQ-NEXT: # zmm16 = mem[0,1,2,3,0,1,2,3]4095; AVX512DQ-NEXT: vmovdqa64 %zmm4, %zmm74096; AVX512DQ-NEXT: vpermt2d %zmm6, %zmm16, %zmm74097; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm17 = [0,16,2,18,0,1,17,3,0,16,2,18,0,1,17,3]4098; AVX512DQ-NEXT: # zmm17 = mem[0,1,2,3,0,1,2,3]4099; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm34100; AVX512DQ-NEXT: vpermt2d %zmm1, %zmm17, %zmm34101; AVX512DQ-NEXT: movw $12684, %ax # imm = 0x318C4102; AVX512DQ-NEXT: kmovw %eax, %k14103; AVX512DQ-NEXT: vmovdqa32 %zmm7, %zmm3 {%k1}4104; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm18 = [0,1,2,3,16,5,6,7,8,17,10,11,12,13,18,15]4105; AVX512DQ-NEXT: vpermt2d %zmm5, %zmm18, %zmm34106; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm19 = [14,30,0,13,29,15,31,0,14,30,0,13,29,15,31,0]4107; AVX512DQ-NEXT: # zmm19 = mem[0,1,2,3,0,1,2,3]4108; AVX512DQ-NEXT: vmovdqa64 %zmm12, %zmm94109; AVX512DQ-NEXT: vpermt2d %zmm14, %zmm19, %zmm94110; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [0,13,29,15,31,0,14,30,0,13,29,15,31,0,14,30]4111; AVX512DQ-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]4112; AVX512DQ-NEXT: vmovdqa64 %zmm0, %zmm74113; AVX512DQ-NEXT: vpermt2d %zmm11, %zmm8, %zmm74114; AVX512DQ-NEXT: movw $25368, %ax # imm = 0x63184115; AVX512DQ-NEXT: kmovw %eax, %k24116; AVX512DQ-NEXT: vmovdqa32 %zmm9, %zmm7 {%k2}4117; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm20 = [28,1,2,3,4,29,6,7,8,9,30,11,12,13,14,31]4118; AVX512DQ-NEXT: vpermt2d %zmm13, %zmm20, %zmm74119; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm21 = [27,0,10,26,12,28,0,11,27,0,10,26,12,28,0,11]4120; AVX512DQ-NEXT: # zmm21 = mem[0,1,2,3,0,1,2,3]4121; AVX512DQ-NEXT: vmovdqa64 %zmm0, %zmm154122; AVX512DQ-NEXT: vpermt2d %zmm11, %zmm21, %zmm154123; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm10 = [9,27,11,0,26,10,28,12,9,27,11,0,26,10,28,12]4124; AVX512DQ-NEXT: # zmm10 = mem[0,1,2,3,0,1,2,3]4125; AVX512DQ-NEXT: vmovdqa64 %zmm14, %zmm94126; AVX512DQ-NEXT: vpermt2d %zmm12, %zmm10, %zmm94127; AVX512DQ-NEXT: vmovdqa32 %zmm15, %zmm9 {%k1}4128; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm22 = [0,25,2,3,4,5,26,7,8,9,10,27,12,13,14,15]4129; AVX512DQ-NEXT: vpermt2d %zmm13, %zmm22, %zmm94130; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm23 = [8,24,0,7,23,9,25,0,8,24,0,7,23,9,25,0]4131; AVX512DQ-NEXT: # zmm23 = mem[0,1,2,3,0,1,2,3]4132; AVX512DQ-NEXT: vmovdqa64 %zmm0, %zmm244133; AVX512DQ-NEXT: vpermt2d %zmm11, %zmm23, %zmm244134; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm25 = [6,22,8,24,0,7,23,9,6,22,8,24,0,7,23,9]4135; AVX512DQ-NEXT: # zmm25 = mem[0,1,2,3,0,1,2,3]4136; AVX512DQ-NEXT: vmovdqa64 %zmm12, %zmm154137; AVX512DQ-NEXT: vpermt2d %zmm14, %zmm25, %zmm154138; AVX512DQ-NEXT: vmovdqa32 %zmm24, %zmm15 {%k2}4139; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm24 = [0,1,22,3,4,5,6,23,8,9,10,11,24,13,14,15]4140; AVX512DQ-NEXT: vpermt2d %zmm13, %zmm24, %zmm154141; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm26 = [0,3,19,5,21,0,4,20,0,3,19,5,21,0,4,20]4142; AVX512DQ-NEXT: # zmm26 = mem[0,1,2,3,0,1,2,3]4143; AVX512DQ-NEXT: vmovdqa64 %zmm12, %zmm274144; AVX512DQ-NEXT: vpermt2d %zmm14, %zmm26, %zmm274145; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm28 = [3,21,5,0,20,4,22,6,3,21,5,0,20,4,22,6]4146; AVX512DQ-NEXT: # zmm28 = mem[0,1,2,3,0,1,2,3]4147; AVX512DQ-NEXT: vmovdqa64 %zmm11, %zmm294148; AVX512DQ-NEXT: vpermt2d %zmm0, %zmm28, %zmm294149; AVX512DQ-NEXT: movw $6342, %ax # imm = 0x18C64150; AVX512DQ-NEXT: kmovw %eax, %k34151; AVX512DQ-NEXT: vmovdqa32 %zmm27, %zmm29 {%k3}4152; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm27 = [0,1,2,19,4,5,6,7,20,9,10,11,12,21,14,15]4153; AVX512DQ-NEXT: vpermt2d %zmm13, %zmm27, %zmm294154; AVX512DQ-NEXT: vpermt2d %zmm14, %zmm16, %zmm124155; AVX512DQ-NEXT: vpermt2d %zmm11, %zmm17, %zmm04156; AVX512DQ-NEXT: vmovdqa32 %zmm12, %zmm0 {%k1}4157; AVX512DQ-NEXT: vpermt2d %zmm13, %zmm18, %zmm04158; AVX512DQ-NEXT: vpermi2d %zmm6, %zmm4, %zmm194159; AVX512DQ-NEXT: vpermi2d %zmm1, %zmm2, %zmm84160; AVX512DQ-NEXT: vmovdqa32 %zmm19, %zmm8 {%k2}4161; AVX512DQ-NEXT: vpermt2d %zmm5, %zmm20, %zmm84162; AVX512DQ-NEXT: vpermi2d %zmm1, %zmm2, %zmm214163; AVX512DQ-NEXT: vpermi2d %zmm4, %zmm6, %zmm104164; AVX512DQ-NEXT: vmovdqa32 %zmm21, %zmm10 {%k1}4165; AVX512DQ-NEXT: vpermt2d %zmm5, %zmm22, %zmm104166; AVX512DQ-NEXT: vpermi2d %zmm1, %zmm2, %zmm234167; AVX512DQ-NEXT: vpermi2d %zmm6, %zmm4, %zmm254168; AVX512DQ-NEXT: vmovdqa32 %zmm23, %zmm25 {%k2}4169; AVX512DQ-NEXT: vpermt2d %zmm5, %zmm24, %zmm254170; AVX512DQ-NEXT: vpermt2d %zmm6, %zmm26, %zmm44171; AVX512DQ-NEXT: vpermt2d %zmm2, %zmm28, %zmm14172; AVX512DQ-NEXT: vmovdqa32 %zmm4, %zmm1 {%k3}4173; AVX512DQ-NEXT: vpermt2d %zmm5, %zmm27, %zmm14174; AVX512DQ-NEXT: vmovdqa64 %zmm1, 64(%r9)4175; AVX512DQ-NEXT: vmovdqa64 %zmm25, 128(%r9)4176; AVX512DQ-NEXT: vmovdqa64 %zmm10, 192(%r9)4177; AVX512DQ-NEXT: vmovdqa64 %zmm8, 256(%r9)4178; AVX512DQ-NEXT: vmovdqa64 %zmm0, 320(%r9)4179; AVX512DQ-NEXT: vmovdqa64 %zmm29, 384(%r9)4180; AVX512DQ-NEXT: vmovdqa64 %zmm15, 448(%r9)4181; AVX512DQ-NEXT: vmovdqa64 %zmm9, 512(%r9)4182; AVX512DQ-NEXT: vmovdqa64 %zmm7, 576(%r9)4183; AVX512DQ-NEXT: vmovdqa64 %zmm3, (%r9)4184; AVX512DQ-NEXT: vzeroupper4185; AVX512DQ-NEXT: retq4186;4187; AVX512DQ-FCP-LABEL: store_i32_stride5_vf32:4188; AVX512DQ-FCP: # %bb.0:4189; AVX512DQ-FCP-NEXT: vmovdqa64 (%rdi), %zmm24190; AVX512DQ-FCP-NEXT: vmovdqa64 64(%rdi), %zmm04191; AVX512DQ-FCP-NEXT: vmovdqa64 (%rsi), %zmm14192; AVX512DQ-FCP-NEXT: vmovdqa64 64(%rsi), %zmm114193; AVX512DQ-FCP-NEXT: vmovdqa64 (%rdx), %zmm44194; AVX512DQ-FCP-NEXT: vmovdqa64 64(%rdx), %zmm124195; AVX512DQ-FCP-NEXT: vmovdqa64 (%rcx), %zmm64196; AVX512DQ-FCP-NEXT: vmovdqa64 64(%rcx), %zmm144197; AVX512DQ-FCP-NEXT: vmovdqa64 (%r8), %zmm54198; AVX512DQ-FCP-NEXT: vmovdqa64 64(%r8), %zmm134199; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm16 = [17,0,0,16,2,18,0,1,17,0,0,16,2,18,0,1]4200; AVX512DQ-FCP-NEXT: # zmm16 = mem[0,1,2,3,0,1,2,3]4201; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm4, %zmm74202; AVX512DQ-FCP-NEXT: vpermt2d %zmm6, %zmm16, %zmm74203; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm17 = [0,16,2,18,0,1,17,3,0,16,2,18,0,1,17,3]4204; AVX512DQ-FCP-NEXT: # zmm17 = mem[0,1,2,3,0,1,2,3]4205; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm34206; AVX512DQ-FCP-NEXT: vpermt2d %zmm1, %zmm17, %zmm34207; AVX512DQ-FCP-NEXT: movw $12684, %ax # imm = 0x318C4208; AVX512DQ-FCP-NEXT: kmovw %eax, %k14209; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm7, %zmm3 {%k1}4210; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm18 = [0,1,2,3,16,5,6,7,8,17,10,11,12,13,18,15]4211; AVX512DQ-FCP-NEXT: vpermt2d %zmm5, %zmm18, %zmm34212; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm19 = [14,30,0,13,29,15,31,0,14,30,0,13,29,15,31,0]4213; AVX512DQ-FCP-NEXT: # zmm19 = mem[0,1,2,3,0,1,2,3]4214; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm12, %zmm94215; AVX512DQ-FCP-NEXT: vpermt2d %zmm14, %zmm19, %zmm94216; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [0,13,29,15,31,0,14,30,0,13,29,15,31,0,14,30]4217; AVX512DQ-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]4218; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm74219; AVX512DQ-FCP-NEXT: vpermt2d %zmm11, %zmm8, %zmm74220; AVX512DQ-FCP-NEXT: movw $25368, %ax # imm = 0x63184221; AVX512DQ-FCP-NEXT: kmovw %eax, %k24222; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm9, %zmm7 {%k2}4223; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm20 = [28,1,2,3,4,29,6,7,8,9,30,11,12,13,14,31]4224; AVX512DQ-FCP-NEXT: vpermt2d %zmm13, %zmm20, %zmm74225; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm21 = [27,0,10,26,12,28,0,11,27,0,10,26,12,28,0,11]4226; AVX512DQ-FCP-NEXT: # zmm21 = mem[0,1,2,3,0,1,2,3]4227; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm154228; AVX512DQ-FCP-NEXT: vpermt2d %zmm11, %zmm21, %zmm154229; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm10 = [9,27,11,0,26,10,28,12,9,27,11,0,26,10,28,12]4230; AVX512DQ-FCP-NEXT: # zmm10 = mem[0,1,2,3,0,1,2,3]4231; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm14, %zmm94232; AVX512DQ-FCP-NEXT: vpermt2d %zmm12, %zmm10, %zmm94233; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm15, %zmm9 {%k1}4234; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm22 = [0,25,2,3,4,5,26,7,8,9,10,27,12,13,14,15]4235; AVX512DQ-FCP-NEXT: vpermt2d %zmm13, %zmm22, %zmm94236; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm23 = [8,24,0,7,23,9,25,0,8,24,0,7,23,9,25,0]4237; AVX512DQ-FCP-NEXT: # zmm23 = mem[0,1,2,3,0,1,2,3]4238; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm244239; AVX512DQ-FCP-NEXT: vpermt2d %zmm11, %zmm23, %zmm244240; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm25 = [6,22,8,24,0,7,23,9,6,22,8,24,0,7,23,9]4241; AVX512DQ-FCP-NEXT: # zmm25 = mem[0,1,2,3,0,1,2,3]4242; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm12, %zmm154243; AVX512DQ-FCP-NEXT: vpermt2d %zmm14, %zmm25, %zmm154244; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm24, %zmm15 {%k2}4245; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm24 = [0,1,22,3,4,5,6,23,8,9,10,11,24,13,14,15]4246; AVX512DQ-FCP-NEXT: vpermt2d %zmm13, %zmm24, %zmm154247; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm26 = [0,3,19,5,21,0,4,20,0,3,19,5,21,0,4,20]4248; AVX512DQ-FCP-NEXT: # zmm26 = mem[0,1,2,3,0,1,2,3]4249; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm12, %zmm274250; AVX512DQ-FCP-NEXT: vpermt2d %zmm14, %zmm26, %zmm274251; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm28 = [3,21,5,0,20,4,22,6,3,21,5,0,20,4,22,6]4252; AVX512DQ-FCP-NEXT: # zmm28 = mem[0,1,2,3,0,1,2,3]4253; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm11, %zmm294254; AVX512DQ-FCP-NEXT: vpermt2d %zmm0, %zmm28, %zmm294255; AVX512DQ-FCP-NEXT: movw $6342, %ax # imm = 0x18C64256; AVX512DQ-FCP-NEXT: kmovw %eax, %k34257; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm27, %zmm29 {%k3}4258; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm27 = [0,1,2,19,4,5,6,7,20,9,10,11,12,21,14,15]4259; AVX512DQ-FCP-NEXT: vpermt2d %zmm13, %zmm27, %zmm294260; AVX512DQ-FCP-NEXT: vpermt2d %zmm14, %zmm16, %zmm124261; AVX512DQ-FCP-NEXT: vpermt2d %zmm11, %zmm17, %zmm04262; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm12, %zmm0 {%k1}4263; AVX512DQ-FCP-NEXT: vpermt2d %zmm13, %zmm18, %zmm04264; AVX512DQ-FCP-NEXT: vpermi2d %zmm6, %zmm4, %zmm194265; AVX512DQ-FCP-NEXT: vpermi2d %zmm1, %zmm2, %zmm84266; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm19, %zmm8 {%k2}4267; AVX512DQ-FCP-NEXT: vpermt2d %zmm5, %zmm20, %zmm84268; AVX512DQ-FCP-NEXT: vpermi2d %zmm1, %zmm2, %zmm214269; AVX512DQ-FCP-NEXT: vpermi2d %zmm4, %zmm6, %zmm104270; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm21, %zmm10 {%k1}4271; AVX512DQ-FCP-NEXT: vpermt2d %zmm5, %zmm22, %zmm104272; AVX512DQ-FCP-NEXT: vpermi2d %zmm1, %zmm2, %zmm234273; AVX512DQ-FCP-NEXT: vpermi2d %zmm6, %zmm4, %zmm254274; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm23, %zmm25 {%k2}4275; AVX512DQ-FCP-NEXT: vpermt2d %zmm5, %zmm24, %zmm254276; AVX512DQ-FCP-NEXT: vpermt2d %zmm6, %zmm26, %zmm44277; AVX512DQ-FCP-NEXT: vpermt2d %zmm2, %zmm28, %zmm14278; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm4, %zmm1 {%k3}4279; AVX512DQ-FCP-NEXT: vpermt2d %zmm5, %zmm27, %zmm14280; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, 64(%r9)4281; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm25, 128(%r9)4282; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm10, 192(%r9)4283; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm8, 256(%r9)4284; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, 320(%r9)4285; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm29, 384(%r9)4286; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm15, 448(%r9)4287; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm9, 512(%r9)4288; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm7, 576(%r9)4289; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, (%r9)4290; AVX512DQ-FCP-NEXT: vzeroupper4291; AVX512DQ-FCP-NEXT: retq4292;4293; AVX512BW-LABEL: store_i32_stride5_vf32:4294; AVX512BW: # %bb.0:4295; AVX512BW-NEXT: vmovdqa64 (%rdi), %zmm24296; AVX512BW-NEXT: vmovdqa64 64(%rdi), %zmm04297; AVX512BW-NEXT: vmovdqa64 (%rsi), %zmm14298; AVX512BW-NEXT: vmovdqa64 64(%rsi), %zmm114299; AVX512BW-NEXT: vmovdqa64 (%rdx), %zmm44300; AVX512BW-NEXT: vmovdqa64 64(%rdx), %zmm124301; AVX512BW-NEXT: vmovdqa64 (%rcx), %zmm64302; AVX512BW-NEXT: vmovdqa64 64(%rcx), %zmm144303; AVX512BW-NEXT: vmovdqa64 (%r8), %zmm54304; AVX512BW-NEXT: vmovdqa64 64(%r8), %zmm134305; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm16 = [17,0,0,16,2,18,0,1,17,0,0,16,2,18,0,1]4306; AVX512BW-NEXT: # zmm16 = mem[0,1,2,3,0,1,2,3]4307; AVX512BW-NEXT: vmovdqa64 %zmm4, %zmm74308; AVX512BW-NEXT: vpermt2d %zmm6, %zmm16, %zmm74309; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm17 = [0,16,2,18,0,1,17,3,0,16,2,18,0,1,17,3]4310; AVX512BW-NEXT: # zmm17 = mem[0,1,2,3,0,1,2,3]4311; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm34312; AVX512BW-NEXT: vpermt2d %zmm1, %zmm17, %zmm34313; AVX512BW-NEXT: movw $12684, %ax # imm = 0x318C4314; AVX512BW-NEXT: kmovd %eax, %k14315; AVX512BW-NEXT: vmovdqa32 %zmm7, %zmm3 {%k1}4316; AVX512BW-NEXT: vpmovsxbd {{.*#+}} zmm18 = [0,1,2,3,16,5,6,7,8,17,10,11,12,13,18,15]4317; AVX512BW-NEXT: vpermt2d %zmm5, %zmm18, %zmm34318; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm19 = [14,30,0,13,29,15,31,0,14,30,0,13,29,15,31,0]4319; AVX512BW-NEXT: # zmm19 = mem[0,1,2,3,0,1,2,3]4320; AVX512BW-NEXT: vmovdqa64 %zmm12, %zmm94321; AVX512BW-NEXT: vpermt2d %zmm14, %zmm19, %zmm94322; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [0,13,29,15,31,0,14,30,0,13,29,15,31,0,14,30]4323; AVX512BW-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]4324; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm74325; AVX512BW-NEXT: vpermt2d %zmm11, %zmm8, %zmm74326; AVX512BW-NEXT: movw $25368, %ax # imm = 0x63184327; AVX512BW-NEXT: kmovd %eax, %k24328; AVX512BW-NEXT: vmovdqa32 %zmm9, %zmm7 {%k2}4329; AVX512BW-NEXT: vpmovsxbd {{.*#+}} zmm20 = [28,1,2,3,4,29,6,7,8,9,30,11,12,13,14,31]4330; AVX512BW-NEXT: vpermt2d %zmm13, %zmm20, %zmm74331; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm21 = [27,0,10,26,12,28,0,11,27,0,10,26,12,28,0,11]4332; AVX512BW-NEXT: # zmm21 = mem[0,1,2,3,0,1,2,3]4333; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm154334; AVX512BW-NEXT: vpermt2d %zmm11, %zmm21, %zmm154335; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm10 = [9,27,11,0,26,10,28,12,9,27,11,0,26,10,28,12]4336; AVX512BW-NEXT: # zmm10 = mem[0,1,2,3,0,1,2,3]4337; AVX512BW-NEXT: vmovdqa64 %zmm14, %zmm94338; AVX512BW-NEXT: vpermt2d %zmm12, %zmm10, %zmm94339; AVX512BW-NEXT: vmovdqa32 %zmm15, %zmm9 {%k1}4340; AVX512BW-NEXT: vpmovsxbd {{.*#+}} zmm22 = [0,25,2,3,4,5,26,7,8,9,10,27,12,13,14,15]4341; AVX512BW-NEXT: vpermt2d %zmm13, %zmm22, %zmm94342; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm23 = [8,24,0,7,23,9,25,0,8,24,0,7,23,9,25,0]4343; AVX512BW-NEXT: # zmm23 = mem[0,1,2,3,0,1,2,3]4344; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm244345; AVX512BW-NEXT: vpermt2d %zmm11, %zmm23, %zmm244346; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm25 = [6,22,8,24,0,7,23,9,6,22,8,24,0,7,23,9]4347; AVX512BW-NEXT: # zmm25 = mem[0,1,2,3,0,1,2,3]4348; AVX512BW-NEXT: vmovdqa64 %zmm12, %zmm154349; AVX512BW-NEXT: vpermt2d %zmm14, %zmm25, %zmm154350; AVX512BW-NEXT: vmovdqa32 %zmm24, %zmm15 {%k2}4351; AVX512BW-NEXT: vpmovsxbd {{.*#+}} zmm24 = [0,1,22,3,4,5,6,23,8,9,10,11,24,13,14,15]4352; AVX512BW-NEXT: vpermt2d %zmm13, %zmm24, %zmm154353; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm26 = [0,3,19,5,21,0,4,20,0,3,19,5,21,0,4,20]4354; AVX512BW-NEXT: # zmm26 = mem[0,1,2,3,0,1,2,3]4355; AVX512BW-NEXT: vmovdqa64 %zmm12, %zmm274356; AVX512BW-NEXT: vpermt2d %zmm14, %zmm26, %zmm274357; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm28 = [3,21,5,0,20,4,22,6,3,21,5,0,20,4,22,6]4358; AVX512BW-NEXT: # zmm28 = mem[0,1,2,3,0,1,2,3]4359; AVX512BW-NEXT: vmovdqa64 %zmm11, %zmm294360; AVX512BW-NEXT: vpermt2d %zmm0, %zmm28, %zmm294361; AVX512BW-NEXT: movw $6342, %ax # imm = 0x18C64362; AVX512BW-NEXT: kmovd %eax, %k34363; AVX512BW-NEXT: vmovdqa32 %zmm27, %zmm29 {%k3}4364; AVX512BW-NEXT: vpmovsxbd {{.*#+}} zmm27 = [0,1,2,19,4,5,6,7,20,9,10,11,12,21,14,15]4365; AVX512BW-NEXT: vpermt2d %zmm13, %zmm27, %zmm294366; AVX512BW-NEXT: vpermt2d %zmm14, %zmm16, %zmm124367; AVX512BW-NEXT: vpermt2d %zmm11, %zmm17, %zmm04368; AVX512BW-NEXT: vmovdqa32 %zmm12, %zmm0 {%k1}4369; AVX512BW-NEXT: vpermt2d %zmm13, %zmm18, %zmm04370; AVX512BW-NEXT: vpermi2d %zmm6, %zmm4, %zmm194371; AVX512BW-NEXT: vpermi2d %zmm1, %zmm2, %zmm84372; AVX512BW-NEXT: vmovdqa32 %zmm19, %zmm8 {%k2}4373; AVX512BW-NEXT: vpermt2d %zmm5, %zmm20, %zmm84374; AVX512BW-NEXT: vpermi2d %zmm1, %zmm2, %zmm214375; AVX512BW-NEXT: vpermi2d %zmm4, %zmm6, %zmm104376; AVX512BW-NEXT: vmovdqa32 %zmm21, %zmm10 {%k1}4377; AVX512BW-NEXT: vpermt2d %zmm5, %zmm22, %zmm104378; AVX512BW-NEXT: vpermi2d %zmm1, %zmm2, %zmm234379; AVX512BW-NEXT: vpermi2d %zmm6, %zmm4, %zmm254380; AVX512BW-NEXT: vmovdqa32 %zmm23, %zmm25 {%k2}4381; AVX512BW-NEXT: vpermt2d %zmm5, %zmm24, %zmm254382; AVX512BW-NEXT: vpermt2d %zmm6, %zmm26, %zmm44383; AVX512BW-NEXT: vpermt2d %zmm2, %zmm28, %zmm14384; AVX512BW-NEXT: vmovdqa32 %zmm4, %zmm1 {%k3}4385; AVX512BW-NEXT: vpermt2d %zmm5, %zmm27, %zmm14386; AVX512BW-NEXT: vmovdqa64 %zmm1, 64(%r9)4387; AVX512BW-NEXT: vmovdqa64 %zmm25, 128(%r9)4388; AVX512BW-NEXT: vmovdqa64 %zmm10, 192(%r9)4389; AVX512BW-NEXT: vmovdqa64 %zmm8, 256(%r9)4390; AVX512BW-NEXT: vmovdqa64 %zmm0, 320(%r9)4391; AVX512BW-NEXT: vmovdqa64 %zmm29, 384(%r9)4392; AVX512BW-NEXT: vmovdqa64 %zmm15, 448(%r9)4393; AVX512BW-NEXT: vmovdqa64 %zmm9, 512(%r9)4394; AVX512BW-NEXT: vmovdqa64 %zmm7, 576(%r9)4395; AVX512BW-NEXT: vmovdqa64 %zmm3, (%r9)4396; AVX512BW-NEXT: vzeroupper4397; AVX512BW-NEXT: retq4398;4399; AVX512BW-FCP-LABEL: store_i32_stride5_vf32:4400; AVX512BW-FCP: # %bb.0:4401; AVX512BW-FCP-NEXT: vmovdqa64 (%rdi), %zmm24402; AVX512BW-FCP-NEXT: vmovdqa64 64(%rdi), %zmm04403; AVX512BW-FCP-NEXT: vmovdqa64 (%rsi), %zmm14404; AVX512BW-FCP-NEXT: vmovdqa64 64(%rsi), %zmm114405; AVX512BW-FCP-NEXT: vmovdqa64 (%rdx), %zmm44406; AVX512BW-FCP-NEXT: vmovdqa64 64(%rdx), %zmm124407; AVX512BW-FCP-NEXT: vmovdqa64 (%rcx), %zmm64408; AVX512BW-FCP-NEXT: vmovdqa64 64(%rcx), %zmm144409; AVX512BW-FCP-NEXT: vmovdqa64 (%r8), %zmm54410; AVX512BW-FCP-NEXT: vmovdqa64 64(%r8), %zmm134411; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm16 = [17,0,0,16,2,18,0,1,17,0,0,16,2,18,0,1]4412; AVX512BW-FCP-NEXT: # zmm16 = mem[0,1,2,3,0,1,2,3]4413; AVX512BW-FCP-NEXT: vmovdqa64 %zmm4, %zmm74414; AVX512BW-FCP-NEXT: vpermt2d %zmm6, %zmm16, %zmm74415; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm17 = [0,16,2,18,0,1,17,3,0,16,2,18,0,1,17,3]4416; AVX512BW-FCP-NEXT: # zmm17 = mem[0,1,2,3,0,1,2,3]4417; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm34418; AVX512BW-FCP-NEXT: vpermt2d %zmm1, %zmm17, %zmm34419; AVX512BW-FCP-NEXT: movw $12684, %ax # imm = 0x318C4420; AVX512BW-FCP-NEXT: kmovd %eax, %k14421; AVX512BW-FCP-NEXT: vmovdqa32 %zmm7, %zmm3 {%k1}4422; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} zmm18 = [0,1,2,3,16,5,6,7,8,17,10,11,12,13,18,15]4423; AVX512BW-FCP-NEXT: vpermt2d %zmm5, %zmm18, %zmm34424; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm19 = [14,30,0,13,29,15,31,0,14,30,0,13,29,15,31,0]4425; AVX512BW-FCP-NEXT: # zmm19 = mem[0,1,2,3,0,1,2,3]4426; AVX512BW-FCP-NEXT: vmovdqa64 %zmm12, %zmm94427; AVX512BW-FCP-NEXT: vpermt2d %zmm14, %zmm19, %zmm94428; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [0,13,29,15,31,0,14,30,0,13,29,15,31,0,14,30]4429; AVX512BW-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]4430; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm74431; AVX512BW-FCP-NEXT: vpermt2d %zmm11, %zmm8, %zmm74432; AVX512BW-FCP-NEXT: movw $25368, %ax # imm = 0x63184433; AVX512BW-FCP-NEXT: kmovd %eax, %k24434; AVX512BW-FCP-NEXT: vmovdqa32 %zmm9, %zmm7 {%k2}4435; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} zmm20 = [28,1,2,3,4,29,6,7,8,9,30,11,12,13,14,31]4436; AVX512BW-FCP-NEXT: vpermt2d %zmm13, %zmm20, %zmm74437; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm21 = [27,0,10,26,12,28,0,11,27,0,10,26,12,28,0,11]4438; AVX512BW-FCP-NEXT: # zmm21 = mem[0,1,2,3,0,1,2,3]4439; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm154440; AVX512BW-FCP-NEXT: vpermt2d %zmm11, %zmm21, %zmm154441; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm10 = [9,27,11,0,26,10,28,12,9,27,11,0,26,10,28,12]4442; AVX512BW-FCP-NEXT: # zmm10 = mem[0,1,2,3,0,1,2,3]4443; AVX512BW-FCP-NEXT: vmovdqa64 %zmm14, %zmm94444; AVX512BW-FCP-NEXT: vpermt2d %zmm12, %zmm10, %zmm94445; AVX512BW-FCP-NEXT: vmovdqa32 %zmm15, %zmm9 {%k1}4446; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} zmm22 = [0,25,2,3,4,5,26,7,8,9,10,27,12,13,14,15]4447; AVX512BW-FCP-NEXT: vpermt2d %zmm13, %zmm22, %zmm94448; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm23 = [8,24,0,7,23,9,25,0,8,24,0,7,23,9,25,0]4449; AVX512BW-FCP-NEXT: # zmm23 = mem[0,1,2,3,0,1,2,3]4450; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm244451; AVX512BW-FCP-NEXT: vpermt2d %zmm11, %zmm23, %zmm244452; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm25 = [6,22,8,24,0,7,23,9,6,22,8,24,0,7,23,9]4453; AVX512BW-FCP-NEXT: # zmm25 = mem[0,1,2,3,0,1,2,3]4454; AVX512BW-FCP-NEXT: vmovdqa64 %zmm12, %zmm154455; AVX512BW-FCP-NEXT: vpermt2d %zmm14, %zmm25, %zmm154456; AVX512BW-FCP-NEXT: vmovdqa32 %zmm24, %zmm15 {%k2}4457; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} zmm24 = [0,1,22,3,4,5,6,23,8,9,10,11,24,13,14,15]4458; AVX512BW-FCP-NEXT: vpermt2d %zmm13, %zmm24, %zmm154459; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm26 = [0,3,19,5,21,0,4,20,0,3,19,5,21,0,4,20]4460; AVX512BW-FCP-NEXT: # zmm26 = mem[0,1,2,3,0,1,2,3]4461; AVX512BW-FCP-NEXT: vmovdqa64 %zmm12, %zmm274462; AVX512BW-FCP-NEXT: vpermt2d %zmm14, %zmm26, %zmm274463; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm28 = [3,21,5,0,20,4,22,6,3,21,5,0,20,4,22,6]4464; AVX512BW-FCP-NEXT: # zmm28 = mem[0,1,2,3,0,1,2,3]4465; AVX512BW-FCP-NEXT: vmovdqa64 %zmm11, %zmm294466; AVX512BW-FCP-NEXT: vpermt2d %zmm0, %zmm28, %zmm294467; AVX512BW-FCP-NEXT: movw $6342, %ax # imm = 0x18C64468; AVX512BW-FCP-NEXT: kmovd %eax, %k34469; AVX512BW-FCP-NEXT: vmovdqa32 %zmm27, %zmm29 {%k3}4470; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} zmm27 = [0,1,2,19,4,5,6,7,20,9,10,11,12,21,14,15]4471; AVX512BW-FCP-NEXT: vpermt2d %zmm13, %zmm27, %zmm294472; AVX512BW-FCP-NEXT: vpermt2d %zmm14, %zmm16, %zmm124473; AVX512BW-FCP-NEXT: vpermt2d %zmm11, %zmm17, %zmm04474; AVX512BW-FCP-NEXT: vmovdqa32 %zmm12, %zmm0 {%k1}4475; AVX512BW-FCP-NEXT: vpermt2d %zmm13, %zmm18, %zmm04476; AVX512BW-FCP-NEXT: vpermi2d %zmm6, %zmm4, %zmm194477; AVX512BW-FCP-NEXT: vpermi2d %zmm1, %zmm2, %zmm84478; AVX512BW-FCP-NEXT: vmovdqa32 %zmm19, %zmm8 {%k2}4479; AVX512BW-FCP-NEXT: vpermt2d %zmm5, %zmm20, %zmm84480; AVX512BW-FCP-NEXT: vpermi2d %zmm1, %zmm2, %zmm214481; AVX512BW-FCP-NEXT: vpermi2d %zmm4, %zmm6, %zmm104482; AVX512BW-FCP-NEXT: vmovdqa32 %zmm21, %zmm10 {%k1}4483; AVX512BW-FCP-NEXT: vpermt2d %zmm5, %zmm22, %zmm104484; AVX512BW-FCP-NEXT: vpermi2d %zmm1, %zmm2, %zmm234485; AVX512BW-FCP-NEXT: vpermi2d %zmm6, %zmm4, %zmm254486; AVX512BW-FCP-NEXT: vmovdqa32 %zmm23, %zmm25 {%k2}4487; AVX512BW-FCP-NEXT: vpermt2d %zmm5, %zmm24, %zmm254488; AVX512BW-FCP-NEXT: vpermt2d %zmm6, %zmm26, %zmm44489; AVX512BW-FCP-NEXT: vpermt2d %zmm2, %zmm28, %zmm14490; AVX512BW-FCP-NEXT: vmovdqa32 %zmm4, %zmm1 {%k3}4491; AVX512BW-FCP-NEXT: vpermt2d %zmm5, %zmm27, %zmm14492; AVX512BW-FCP-NEXT: vmovdqa64 %zmm1, 64(%r9)4493; AVX512BW-FCP-NEXT: vmovdqa64 %zmm25, 128(%r9)4494; AVX512BW-FCP-NEXT: vmovdqa64 %zmm10, 192(%r9)4495; AVX512BW-FCP-NEXT: vmovdqa64 %zmm8, 256(%r9)4496; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, 320(%r9)4497; AVX512BW-FCP-NEXT: vmovdqa64 %zmm29, 384(%r9)4498; AVX512BW-FCP-NEXT: vmovdqa64 %zmm15, 448(%r9)4499; AVX512BW-FCP-NEXT: vmovdqa64 %zmm9, 512(%r9)4500; AVX512BW-FCP-NEXT: vmovdqa64 %zmm7, 576(%r9)4501; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, (%r9)4502; AVX512BW-FCP-NEXT: vzeroupper4503; AVX512BW-FCP-NEXT: retq4504;4505; AVX512DQ-BW-LABEL: store_i32_stride5_vf32:4506; AVX512DQ-BW: # %bb.0:4507; AVX512DQ-BW-NEXT: vmovdqa64 (%rdi), %zmm24508; AVX512DQ-BW-NEXT: vmovdqa64 64(%rdi), %zmm04509; AVX512DQ-BW-NEXT: vmovdqa64 (%rsi), %zmm14510; AVX512DQ-BW-NEXT: vmovdqa64 64(%rsi), %zmm114511; AVX512DQ-BW-NEXT: vmovdqa64 (%rdx), %zmm44512; AVX512DQ-BW-NEXT: vmovdqa64 64(%rdx), %zmm124513; AVX512DQ-BW-NEXT: vmovdqa64 (%rcx), %zmm64514; AVX512DQ-BW-NEXT: vmovdqa64 64(%rcx), %zmm144515; AVX512DQ-BW-NEXT: vmovdqa64 (%r8), %zmm54516; AVX512DQ-BW-NEXT: vmovdqa64 64(%r8), %zmm134517; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm16 = [17,0,0,16,2,18,0,1,17,0,0,16,2,18,0,1]4518; AVX512DQ-BW-NEXT: # zmm16 = mem[0,1,2,3,0,1,2,3]4519; AVX512DQ-BW-NEXT: vmovdqa64 %zmm4, %zmm74520; AVX512DQ-BW-NEXT: vpermt2d %zmm6, %zmm16, %zmm74521; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm17 = [0,16,2,18,0,1,17,3,0,16,2,18,0,1,17,3]4522; AVX512DQ-BW-NEXT: # zmm17 = mem[0,1,2,3,0,1,2,3]4523; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm34524; AVX512DQ-BW-NEXT: vpermt2d %zmm1, %zmm17, %zmm34525; AVX512DQ-BW-NEXT: movw $12684, %ax # imm = 0x318C4526; AVX512DQ-BW-NEXT: kmovd %eax, %k14527; AVX512DQ-BW-NEXT: vmovdqa32 %zmm7, %zmm3 {%k1}4528; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} zmm18 = [0,1,2,3,16,5,6,7,8,17,10,11,12,13,18,15]4529; AVX512DQ-BW-NEXT: vpermt2d %zmm5, %zmm18, %zmm34530; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm19 = [14,30,0,13,29,15,31,0,14,30,0,13,29,15,31,0]4531; AVX512DQ-BW-NEXT: # zmm19 = mem[0,1,2,3,0,1,2,3]4532; AVX512DQ-BW-NEXT: vmovdqa64 %zmm12, %zmm94533; AVX512DQ-BW-NEXT: vpermt2d %zmm14, %zmm19, %zmm94534; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [0,13,29,15,31,0,14,30,0,13,29,15,31,0,14,30]4535; AVX512DQ-BW-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]4536; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, %zmm74537; AVX512DQ-BW-NEXT: vpermt2d %zmm11, %zmm8, %zmm74538; AVX512DQ-BW-NEXT: movw $25368, %ax # imm = 0x63184539; AVX512DQ-BW-NEXT: kmovd %eax, %k24540; AVX512DQ-BW-NEXT: vmovdqa32 %zmm9, %zmm7 {%k2}4541; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} zmm20 = [28,1,2,3,4,29,6,7,8,9,30,11,12,13,14,31]4542; AVX512DQ-BW-NEXT: vpermt2d %zmm13, %zmm20, %zmm74543; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm21 = [27,0,10,26,12,28,0,11,27,0,10,26,12,28,0,11]4544; AVX512DQ-BW-NEXT: # zmm21 = mem[0,1,2,3,0,1,2,3]4545; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, %zmm154546; AVX512DQ-BW-NEXT: vpermt2d %zmm11, %zmm21, %zmm154547; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm10 = [9,27,11,0,26,10,28,12,9,27,11,0,26,10,28,12]4548; AVX512DQ-BW-NEXT: # zmm10 = mem[0,1,2,3,0,1,2,3]4549; AVX512DQ-BW-NEXT: vmovdqa64 %zmm14, %zmm94550; AVX512DQ-BW-NEXT: vpermt2d %zmm12, %zmm10, %zmm94551; AVX512DQ-BW-NEXT: vmovdqa32 %zmm15, %zmm9 {%k1}4552; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} zmm22 = [0,25,2,3,4,5,26,7,8,9,10,27,12,13,14,15]4553; AVX512DQ-BW-NEXT: vpermt2d %zmm13, %zmm22, %zmm94554; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm23 = [8,24,0,7,23,9,25,0,8,24,0,7,23,9,25,0]4555; AVX512DQ-BW-NEXT: # zmm23 = mem[0,1,2,3,0,1,2,3]4556; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, %zmm244557; AVX512DQ-BW-NEXT: vpermt2d %zmm11, %zmm23, %zmm244558; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm25 = [6,22,8,24,0,7,23,9,6,22,8,24,0,7,23,9]4559; AVX512DQ-BW-NEXT: # zmm25 = mem[0,1,2,3,0,1,2,3]4560; AVX512DQ-BW-NEXT: vmovdqa64 %zmm12, %zmm154561; AVX512DQ-BW-NEXT: vpermt2d %zmm14, %zmm25, %zmm154562; AVX512DQ-BW-NEXT: vmovdqa32 %zmm24, %zmm15 {%k2}4563; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} zmm24 = [0,1,22,3,4,5,6,23,8,9,10,11,24,13,14,15]4564; AVX512DQ-BW-NEXT: vpermt2d %zmm13, %zmm24, %zmm154565; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm26 = [0,3,19,5,21,0,4,20,0,3,19,5,21,0,4,20]4566; AVX512DQ-BW-NEXT: # zmm26 = mem[0,1,2,3,0,1,2,3]4567; AVX512DQ-BW-NEXT: vmovdqa64 %zmm12, %zmm274568; AVX512DQ-BW-NEXT: vpermt2d %zmm14, %zmm26, %zmm274569; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm28 = [3,21,5,0,20,4,22,6,3,21,5,0,20,4,22,6]4570; AVX512DQ-BW-NEXT: # zmm28 = mem[0,1,2,3,0,1,2,3]4571; AVX512DQ-BW-NEXT: vmovdqa64 %zmm11, %zmm294572; AVX512DQ-BW-NEXT: vpermt2d %zmm0, %zmm28, %zmm294573; AVX512DQ-BW-NEXT: movw $6342, %ax # imm = 0x18C64574; AVX512DQ-BW-NEXT: kmovd %eax, %k34575; AVX512DQ-BW-NEXT: vmovdqa32 %zmm27, %zmm29 {%k3}4576; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} zmm27 = [0,1,2,19,4,5,6,7,20,9,10,11,12,21,14,15]4577; AVX512DQ-BW-NEXT: vpermt2d %zmm13, %zmm27, %zmm294578; AVX512DQ-BW-NEXT: vpermt2d %zmm14, %zmm16, %zmm124579; AVX512DQ-BW-NEXT: vpermt2d %zmm11, %zmm17, %zmm04580; AVX512DQ-BW-NEXT: vmovdqa32 %zmm12, %zmm0 {%k1}4581; AVX512DQ-BW-NEXT: vpermt2d %zmm13, %zmm18, %zmm04582; AVX512DQ-BW-NEXT: vpermi2d %zmm6, %zmm4, %zmm194583; AVX512DQ-BW-NEXT: vpermi2d %zmm1, %zmm2, %zmm84584; AVX512DQ-BW-NEXT: vmovdqa32 %zmm19, %zmm8 {%k2}4585; AVX512DQ-BW-NEXT: vpermt2d %zmm5, %zmm20, %zmm84586; AVX512DQ-BW-NEXT: vpermi2d %zmm1, %zmm2, %zmm214587; AVX512DQ-BW-NEXT: vpermi2d %zmm4, %zmm6, %zmm104588; AVX512DQ-BW-NEXT: vmovdqa32 %zmm21, %zmm10 {%k1}4589; AVX512DQ-BW-NEXT: vpermt2d %zmm5, %zmm22, %zmm104590; AVX512DQ-BW-NEXT: vpermi2d %zmm1, %zmm2, %zmm234591; AVX512DQ-BW-NEXT: vpermi2d %zmm6, %zmm4, %zmm254592; AVX512DQ-BW-NEXT: vmovdqa32 %zmm23, %zmm25 {%k2}4593; AVX512DQ-BW-NEXT: vpermt2d %zmm5, %zmm24, %zmm254594; AVX512DQ-BW-NEXT: vpermt2d %zmm6, %zmm26, %zmm44595; AVX512DQ-BW-NEXT: vpermt2d %zmm2, %zmm28, %zmm14596; AVX512DQ-BW-NEXT: vmovdqa32 %zmm4, %zmm1 {%k3}4597; AVX512DQ-BW-NEXT: vpermt2d %zmm5, %zmm27, %zmm14598; AVX512DQ-BW-NEXT: vmovdqa64 %zmm1, 64(%r9)4599; AVX512DQ-BW-NEXT: vmovdqa64 %zmm25, 128(%r9)4600; AVX512DQ-BW-NEXT: vmovdqa64 %zmm10, 192(%r9)4601; AVX512DQ-BW-NEXT: vmovdqa64 %zmm8, 256(%r9)4602; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, 320(%r9)4603; AVX512DQ-BW-NEXT: vmovdqa64 %zmm29, 384(%r9)4604; AVX512DQ-BW-NEXT: vmovdqa64 %zmm15, 448(%r9)4605; AVX512DQ-BW-NEXT: vmovdqa64 %zmm9, 512(%r9)4606; AVX512DQ-BW-NEXT: vmovdqa64 %zmm7, 576(%r9)4607; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, (%r9)4608; AVX512DQ-BW-NEXT: vzeroupper4609; AVX512DQ-BW-NEXT: retq4610;4611; AVX512DQ-BW-FCP-LABEL: store_i32_stride5_vf32:4612; AVX512DQ-BW-FCP: # %bb.0:4613; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rdi), %zmm24614; AVX512DQ-BW-FCP-NEXT: vmovdqa64 64(%rdi), %zmm04615; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rsi), %zmm14616; AVX512DQ-BW-FCP-NEXT: vmovdqa64 64(%rsi), %zmm114617; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rdx), %zmm44618; AVX512DQ-BW-FCP-NEXT: vmovdqa64 64(%rdx), %zmm124619; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rcx), %zmm64620; AVX512DQ-BW-FCP-NEXT: vmovdqa64 64(%rcx), %zmm144621; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%r8), %zmm54622; AVX512DQ-BW-FCP-NEXT: vmovdqa64 64(%r8), %zmm134623; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm16 = [17,0,0,16,2,18,0,1,17,0,0,16,2,18,0,1]4624; AVX512DQ-BW-FCP-NEXT: # zmm16 = mem[0,1,2,3,0,1,2,3]4625; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm4, %zmm74626; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm6, %zmm16, %zmm74627; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm17 = [0,16,2,18,0,1,17,3,0,16,2,18,0,1,17,3]4628; AVX512DQ-BW-FCP-NEXT: # zmm17 = mem[0,1,2,3,0,1,2,3]4629; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm34630; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm1, %zmm17, %zmm34631; AVX512DQ-BW-FCP-NEXT: movw $12684, %ax # imm = 0x318C4632; AVX512DQ-BW-FCP-NEXT: kmovd %eax, %k14633; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm7, %zmm3 {%k1}4634; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} zmm18 = [0,1,2,3,16,5,6,7,8,17,10,11,12,13,18,15]4635; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm5, %zmm18, %zmm34636; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm19 = [14,30,0,13,29,15,31,0,14,30,0,13,29,15,31,0]4637; AVX512DQ-BW-FCP-NEXT: # zmm19 = mem[0,1,2,3,0,1,2,3]4638; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm12, %zmm94639; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm14, %zmm19, %zmm94640; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [0,13,29,15,31,0,14,30,0,13,29,15,31,0,14,30]4641; AVX512DQ-BW-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]4642; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm74643; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm11, %zmm8, %zmm74644; AVX512DQ-BW-FCP-NEXT: movw $25368, %ax # imm = 0x63184645; AVX512DQ-BW-FCP-NEXT: kmovd %eax, %k24646; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm9, %zmm7 {%k2}4647; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} zmm20 = [28,1,2,3,4,29,6,7,8,9,30,11,12,13,14,31]4648; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm13, %zmm20, %zmm74649; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm21 = [27,0,10,26,12,28,0,11,27,0,10,26,12,28,0,11]4650; AVX512DQ-BW-FCP-NEXT: # zmm21 = mem[0,1,2,3,0,1,2,3]4651; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm154652; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm11, %zmm21, %zmm154653; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm10 = [9,27,11,0,26,10,28,12,9,27,11,0,26,10,28,12]4654; AVX512DQ-BW-FCP-NEXT: # zmm10 = mem[0,1,2,3,0,1,2,3]4655; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm14, %zmm94656; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm12, %zmm10, %zmm94657; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm15, %zmm9 {%k1}4658; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} zmm22 = [0,25,2,3,4,5,26,7,8,9,10,27,12,13,14,15]4659; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm13, %zmm22, %zmm94660; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm23 = [8,24,0,7,23,9,25,0,8,24,0,7,23,9,25,0]4661; AVX512DQ-BW-FCP-NEXT: # zmm23 = mem[0,1,2,3,0,1,2,3]4662; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm244663; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm11, %zmm23, %zmm244664; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm25 = [6,22,8,24,0,7,23,9,6,22,8,24,0,7,23,9]4665; AVX512DQ-BW-FCP-NEXT: # zmm25 = mem[0,1,2,3,0,1,2,3]4666; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm12, %zmm154667; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm14, %zmm25, %zmm154668; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm24, %zmm15 {%k2}4669; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} zmm24 = [0,1,22,3,4,5,6,23,8,9,10,11,24,13,14,15]4670; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm13, %zmm24, %zmm154671; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm26 = [0,3,19,5,21,0,4,20,0,3,19,5,21,0,4,20]4672; AVX512DQ-BW-FCP-NEXT: # zmm26 = mem[0,1,2,3,0,1,2,3]4673; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm12, %zmm274674; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm14, %zmm26, %zmm274675; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm28 = [3,21,5,0,20,4,22,6,3,21,5,0,20,4,22,6]4676; AVX512DQ-BW-FCP-NEXT: # zmm28 = mem[0,1,2,3,0,1,2,3]4677; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm11, %zmm294678; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm0, %zmm28, %zmm294679; AVX512DQ-BW-FCP-NEXT: movw $6342, %ax # imm = 0x18C64680; AVX512DQ-BW-FCP-NEXT: kmovd %eax, %k34681; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm27, %zmm29 {%k3}4682; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} zmm27 = [0,1,2,19,4,5,6,7,20,9,10,11,12,21,14,15]4683; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm13, %zmm27, %zmm294684; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm14, %zmm16, %zmm124685; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm11, %zmm17, %zmm04686; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm12, %zmm0 {%k1}4687; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm13, %zmm18, %zmm04688; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm6, %zmm4, %zmm194689; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm1, %zmm2, %zmm84690; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm19, %zmm8 {%k2}4691; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm5, %zmm20, %zmm84692; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm1, %zmm2, %zmm214693; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm4, %zmm6, %zmm104694; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm21, %zmm10 {%k1}4695; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm5, %zmm22, %zmm104696; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm1, %zmm2, %zmm234697; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm6, %zmm4, %zmm254698; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm23, %zmm25 {%k2}4699; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm5, %zmm24, %zmm254700; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm6, %zmm26, %zmm44701; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm2, %zmm28, %zmm14702; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm4, %zmm1 {%k3}4703; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm5, %zmm27, %zmm14704; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm1, 64(%r9)4705; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm25, 128(%r9)4706; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm10, 192(%r9)4707; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm8, 256(%r9)4708; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, 320(%r9)4709; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm29, 384(%r9)4710; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm15, 448(%r9)4711; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm9, 512(%r9)4712; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm7, 576(%r9)4713; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, (%r9)4714; AVX512DQ-BW-FCP-NEXT: vzeroupper4715; AVX512DQ-BW-FCP-NEXT: retq4716 %in.vec0 = load <32 x i32>, ptr %in.vecptr0, align 644717 %in.vec1 = load <32 x i32>, ptr %in.vecptr1, align 644718 %in.vec2 = load <32 x i32>, ptr %in.vecptr2, align 644719 %in.vec3 = load <32 x i32>, ptr %in.vecptr3, align 644720 %in.vec4 = load <32 x i32>, ptr %in.vecptr4, align 644721 %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>4722 %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>4723 %3 = 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>4724 %4 = shufflevector <32 x i32> %in.vec4, <32 x i32> poison, <128 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>4725 %5 = shufflevector <128 x i32> %3, <128 x i32> %4, <160 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>4726 %interleaved.vec = shufflevector <160 x i32> %5, <160 x i32> poison, <160 x i32> <i32 0, i32 32, i32 64, i32 96, i32 128, i32 1, i32 33, i32 65, i32 97, i32 129, i32 2, i32 34, i32 66, i32 98, i32 130, i32 3, i32 35, i32 67, i32 99, i32 131, i32 4, i32 36, i32 68, i32 100, i32 132, i32 5, i32 37, i32 69, i32 101, i32 133, i32 6, i32 38, i32 70, i32 102, i32 134, i32 7, i32 39, i32 71, i32 103, i32 135, i32 8, i32 40, i32 72, i32 104, i32 136, i32 9, i32 41, i32 73, i32 105, i32 137, i32 10, i32 42, i32 74, i32 106, i32 138, i32 11, i32 43, i32 75, i32 107, i32 139, i32 12, i32 44, i32 76, i32 108, i32 140, i32 13, i32 45, i32 77, i32 109, i32 141, i32 14, i32 46, i32 78, i32 110, i32 142, i32 15, i32 47, i32 79, i32 111, i32 143, i32 16, i32 48, i32 80, i32 112, i32 144, i32 17, i32 49, i32 81, i32 113, i32 145, i32 18, i32 50, i32 82, i32 114, i32 146, i32 19, i32 51, i32 83, i32 115, i32 147, i32 20, i32 52, i32 84, i32 116, i32 148, i32 21, i32 53, i32 85, i32 117, i32 149, i32 22, i32 54, i32 86, i32 118, i32 150, i32 23, i32 55, i32 87, i32 119, i32 151, i32 24, i32 56, i32 88, i32 120, i32 152, i32 25, i32 57, i32 89, i32 121, i32 153, i32 26, i32 58, i32 90, i32 122, i32 154, i32 27, i32 59, i32 91, i32 123, i32 155, i32 28, i32 60, i32 92, i32 124, i32 156, i32 29, i32 61, i32 93, i32 125, i32 157, i32 30, i32 62, i32 94, i32 126, i32 158, i32 31, i32 63, i32 95, i32 127, i32 159>4727 store <160 x i32> %interleaved.vec, ptr %out.vec, align 644728 ret void4729}4730 4731define void @store_i32_stride5_vf64(ptr %in.vecptr0, ptr %in.vecptr1, ptr %in.vecptr2, ptr %in.vecptr3, ptr %in.vecptr4, ptr %out.vec) nounwind {4732; SSE-LABEL: store_i32_stride5_vf64:4733; SSE: # %bb.0:4734; SSE-NEXT: subq $1736, %rsp # imm = 0x6C84735; SSE-NEXT: movdqa (%rsi), %xmm104736; SSE-NEXT: movdqa 16(%rsi), %xmm34737; SSE-NEXT: movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4738; SSE-NEXT: movdqa 32(%rsi), %xmm64739; SSE-NEXT: movdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4740; SSE-NEXT: movdqa (%rdx), %xmm84741; SSE-NEXT: movdqa 16(%rdx), %xmm134742; SSE-NEXT: movdqa %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4743; SSE-NEXT: movdqa 32(%rdx), %xmm74744; SSE-NEXT: movdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4745; SSE-NEXT: movaps (%rcx), %xmm54746; SSE-NEXT: movaps 16(%rcx), %xmm94747; SSE-NEXT: movaps 32(%rcx), %xmm124748; SSE-NEXT: movaps %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4749; SSE-NEXT: movaps (%r8), %xmm44750; SSE-NEXT: movaps 16(%r8), %xmm114751; SSE-NEXT: movaps 32(%r8), %xmm144752; SSE-NEXT: movaps %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4753; SSE-NEXT: movaps %xmm5, %xmm04754; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,3],xmm4[3,3]4755; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm10[3,3,3,3]4756; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm8[2,3,2,3]4757; SSE-NEXT: movss {{.*#+}} xmm2 = xmm1[0],xmm2[1,2,3]4758; SSE-NEXT: shufps {{.*#+}} xmm2 = xmm2[0,1],xmm0[0,2]4759; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4760; SSE-NEXT: movaps %xmm9, %xmm04761; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,3],xmm11[3,3]4762; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm13[3,3,3,3]4763; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm3[3,3,3,3]4764; SSE-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]4765; SSE-NEXT: shufps {{.*#+}} xmm2 = xmm2[0,1],xmm0[0,2]4766; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4767; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm6[3,3,3,3]4768; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm7[2,3,2,3]4769; SSE-NEXT: movss {{.*#+}} xmm1 = xmm0[0],xmm1[1,2,3]4770; SSE-NEXT: movaps %xmm12, %xmm04771; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,3],xmm14[3,3]4772; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,1],xmm0[0,2]4773; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4774; SSE-NEXT: movdqa 48(%rsi), %xmm124775; SSE-NEXT: movdqa 48(%rdx), %xmm04776; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4777; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[3,3,3,3]4778; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm12[3,3,3,3]4779; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]4780; SSE-NEXT: movaps 48(%rcx), %xmm64781; SSE-NEXT: movaps 48(%r8), %xmm74782; SSE-NEXT: movaps %xmm6, %xmm04783; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,3],xmm7[3,3]4784; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,1],xmm0[0,2]4785; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4786; SSE-NEXT: movdqa 64(%rsi), %xmm04787; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4788; SSE-NEXT: movdqa 64(%rdx), %xmm154789; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[3,3,3,3]4790; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm15[2,3,2,3]4791; SSE-NEXT: movss {{.*#+}} xmm1 = xmm0[0],xmm1[1,2,3]4792; SSE-NEXT: movaps 64(%rcx), %xmm134793; SSE-NEXT: movaps 64(%r8), %xmm24794; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4795; SSE-NEXT: movaps %xmm13, %xmm04796; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,3],xmm2[3,3]4797; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,1],xmm0[0,2]4798; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4799; SSE-NEXT: movdqa 80(%rsi), %xmm24800; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4801; SSE-NEXT: movdqa 80(%rdx), %xmm04802; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4803; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[3,3,3,3]4804; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm2[3,3,3,3]4805; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]4806; SSE-NEXT: movaps 80(%rcx), %xmm144807; SSE-NEXT: movaps 80(%r8), %xmm24808; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4809; SSE-NEXT: movaps %xmm14, %xmm04810; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,3],xmm2[3,3]4811; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,1],xmm0[0,2]4812; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4813; SSE-NEXT: movdqa 96(%rsi), %xmm04814; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4815; SSE-NEXT: movdqa 96(%rdx), %xmm14816; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4817; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[3,3,3,3]4818; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[2,3,2,3]4819; SSE-NEXT: movss {{.*#+}} xmm1 = xmm0[0],xmm1[1,2,3]4820; SSE-NEXT: movaps 96(%rcx), %xmm04821; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4822; SSE-NEXT: movaps 96(%r8), %xmm24823; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4824; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,3],xmm2[3,3]4825; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,1],xmm0[0,2]4826; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4827; SSE-NEXT: movdqa 112(%rsi), %xmm24828; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4829; SSE-NEXT: movdqa 112(%rdx), %xmm04830; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4831; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[3,3,3,3]4832; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm2[3,3,3,3]4833; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]4834; SSE-NEXT: movaps 112(%rcx), %xmm04835; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4836; SSE-NEXT: movaps 112(%r8), %xmm24837; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4838; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,3],xmm2[3,3]4839; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,1],xmm0[0,2]4840; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4841; SSE-NEXT: movdqa 128(%rsi), %xmm04842; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4843; SSE-NEXT: movdqa 128(%rdx), %xmm14844; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4845; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[3,3,3,3]4846; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[2,3,2,3]4847; SSE-NEXT: movss {{.*#+}} xmm1 = xmm0[0],xmm1[1,2,3]4848; SSE-NEXT: movaps 128(%rcx), %xmm04849; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4850; SSE-NEXT: movaps 128(%r8), %xmm24851; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4852; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,3],xmm2[3,3]4853; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,1],xmm0[0,2]4854; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4855; SSE-NEXT: movdqa 144(%rsi), %xmm24856; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4857; SSE-NEXT: movdqa 144(%rdx), %xmm04858; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4859; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[3,3,3,3]4860; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm2[3,3,3,3]4861; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]4862; SSE-NEXT: movaps 144(%rcx), %xmm04863; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4864; SSE-NEXT: movaps 144(%r8), %xmm24865; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4866; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,3],xmm2[3,3]4867; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,1],xmm0[0,2]4868; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4869; SSE-NEXT: movdqa 160(%rsi), %xmm04870; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4871; SSE-NEXT: movdqa 160(%rdx), %xmm14872; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4873; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[3,3,3,3]4874; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[2,3,2,3]4875; SSE-NEXT: movss {{.*#+}} xmm1 = xmm0[0],xmm1[1,2,3]4876; SSE-NEXT: movaps 160(%rcx), %xmm04877; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4878; SSE-NEXT: movaps 160(%r8), %xmm24879; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4880; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,3],xmm2[3,3]4881; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,1],xmm0[0,2]4882; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4883; SSE-NEXT: movdqa 176(%rsi), %xmm24884; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4885; SSE-NEXT: movdqa 176(%rdx), %xmm04886; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4887; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[3,3,3,3]4888; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm2[3,3,3,3]4889; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]4890; SSE-NEXT: movaps 176(%rcx), %xmm04891; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4892; SSE-NEXT: movaps 176(%r8), %xmm24893; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4894; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,3],xmm2[3,3]4895; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,1],xmm0[0,2]4896; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4897; SSE-NEXT: movdqa 192(%rsi), %xmm04898; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4899; SSE-NEXT: movdqa 192(%rdx), %xmm14900; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4901; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[3,3,3,3]4902; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[2,3,2,3]4903; SSE-NEXT: movss {{.*#+}} xmm1 = xmm0[0],xmm1[1,2,3]4904; SSE-NEXT: movaps 192(%rcx), %xmm04905; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4906; SSE-NEXT: movaps 192(%r8), %xmm24907; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4908; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,3],xmm2[3,3]4909; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,1],xmm0[0,2]4910; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4911; SSE-NEXT: movdqa 208(%rsi), %xmm24912; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4913; SSE-NEXT: movdqa 208(%rdx), %xmm04914; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4915; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[3,3,3,3]4916; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm2[3,3,3,3]4917; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]4918; SSE-NEXT: movaps 208(%rcx), %xmm04919; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4920; SSE-NEXT: movaps 208(%r8), %xmm24921; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4922; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,3],xmm2[3,3]4923; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,1],xmm0[0,2]4924; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4925; SSE-NEXT: movdqa 224(%rsi), %xmm04926; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4927; SSE-NEXT: movdqa 224(%rdx), %xmm14928; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4929; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[3,3,3,3]4930; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[2,3,2,3]4931; SSE-NEXT: movss {{.*#+}} xmm1 = xmm0[0],xmm1[1,2,3]4932; SSE-NEXT: movaps 224(%rcx), %xmm04933; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4934; SSE-NEXT: movaps 224(%r8), %xmm24935; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4936; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,3],xmm2[3,3]4937; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,1],xmm0[0,2]4938; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4939; SSE-NEXT: movdqa 240(%rsi), %xmm24940; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4941; SSE-NEXT: movdqa 240(%rdx), %xmm04942; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4943; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[3,3,3,3]4944; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm2[3,3,3,3]4945; SSE-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]4946; SSE-NEXT: movaps 240(%rcx), %xmm04947; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4948; SSE-NEXT: movaps 240(%r8), %xmm34949; SSE-NEXT: movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4950; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,3],xmm3[3,3]4951; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,1],xmm0[0,2]4952; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4953; SSE-NEXT: movdqa %xmm8, %xmm24954; SSE-NEXT: movdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4955; SSE-NEXT: movdqa %xmm8, %xmm04956; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm5[0],xmm0[1],xmm5[1]4957; SSE-NEXT: movdqa (%rdi), %xmm14958; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4959; SSE-NEXT: movdqa %xmm1, %xmm84960; SSE-NEXT: movdqa %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4961; SSE-NEXT: punpckldq {{.*#+}} xmm8 = xmm8[0],xmm10[0],xmm8[1],xmm10[1]4962; SSE-NEXT: punpcklqdq {{.*#+}} xmm8 = xmm8[0],xmm0[0]4963; SSE-NEXT: movdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4964; SSE-NEXT: movdqa %xmm1, %xmm04965; SSE-NEXT: punpckhdq {{.*#+}} xmm0 = xmm0[2],xmm10[2],xmm0[3],xmm10[3]4966; SSE-NEXT: punpckhdq {{.*#+}} xmm2 = xmm2[2],xmm5[2],xmm2[3],xmm5[3]4967; SSE-NEXT: movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4968; SSE-NEXT: shufps {{.*#+}} xmm5 = xmm5[1,1],xmm4[1,1]4969; SSE-NEXT: shufps {{.*#+}} xmm5 = xmm5[0,2],xmm0[0,1]4970; SSE-NEXT: movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4971; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[3,0],xmm4[2,0]4972; SSE-NEXT: shufps {{.*#+}} xmm2 = xmm2[0,1],xmm1[2,0]4973; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4974; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload4975; SSE-NEXT: movaps %xmm1, %xmm04976; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm9[0],xmm0[1],xmm9[1]4977; SSE-NEXT: movaps 16(%rdi), %xmm34978; SSE-NEXT: movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4979; SSE-NEXT: movaps %xmm3, %xmm84980; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload4981; SSE-NEXT: unpcklps {{.*#+}} xmm8 = xmm8[0],xmm2[0],xmm8[1],xmm2[1]4982; SSE-NEXT: movlhps {{.*#+}} xmm8 = xmm8[0],xmm0[0]4983; SSE-NEXT: movaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4984; SSE-NEXT: movaps %xmm3, %xmm04985; SSE-NEXT: unpckhps {{.*#+}} xmm0 = xmm0[2],xmm2[2],xmm0[3],xmm2[3]4986; SSE-NEXT: movaps %xmm9, %xmm84987; SSE-NEXT: movaps %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4988; SSE-NEXT: shufps {{.*#+}} xmm8 = xmm8[1,1],xmm11[1,1]4989; SSE-NEXT: shufps {{.*#+}} xmm8 = xmm8[0,2],xmm0[0,1]4990; SSE-NEXT: movaps %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4991; SSE-NEXT: movaps %xmm3, %xmm04992; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,1],xmm11[2,3]4993; SSE-NEXT: unpckhpd {{.*#+}} xmm9 = xmm9[1],xmm1[1]4994; SSE-NEXT: shufps {{.*#+}} xmm9 = xmm9[2,0],xmm0[2,0]4995; SSE-NEXT: movaps %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4996; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload4997; SSE-NEXT: movaps %xmm1, %xmm04998; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload4999; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm4[0],xmm0[1],xmm4[1]5000; SSE-NEXT: movaps 32(%rdi), %xmm25001; SSE-NEXT: movaps %xmm2, %xmm35002; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload5003; SSE-NEXT: unpcklps {{.*#+}} xmm3 = xmm3[0],xmm5[0],xmm3[1],xmm5[1]5004; SSE-NEXT: movlhps {{.*#+}} xmm3 = xmm3[0],xmm0[0]5005; SSE-NEXT: movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5006; SSE-NEXT: movaps %xmm2, %xmm05007; SSE-NEXT: movaps %xmm2, (%rsp) # 16-byte Spill5008; SSE-NEXT: unpckhps {{.*#+}} xmm0 = xmm0[2],xmm5[2],xmm0[3],xmm5[3]5009; SSE-NEXT: movaps %xmm1, %xmm35010; SSE-NEXT: unpckhps {{.*#+}} xmm3 = xmm3[2],xmm4[2],xmm3[3],xmm4[3]5011; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload5012; SSE-NEXT: shufps {{.*#+}} xmm4 = xmm4[1,1],xmm1[1,1]5013; SSE-NEXT: shufps {{.*#+}} xmm4 = xmm4[0,2],xmm0[0,1]5014; SSE-NEXT: movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5015; SSE-NEXT: movaps %xmm2, %xmm05016; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,0],xmm1[2,0]5017; SSE-NEXT: shufps {{.*#+}} xmm3 = xmm3[0,1],xmm0[2,0]5018; SSE-NEXT: movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5019; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload5020; SSE-NEXT: movaps %xmm1, %xmm05021; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm6[0],xmm0[1],xmm6[1]5022; SSE-NEXT: movdqa 48(%rdi), %xmm35023; SSE-NEXT: movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5024; SSE-NEXT: movdqa %xmm3, %xmm45025; SSE-NEXT: movdqa %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5026; SSE-NEXT: punpckldq {{.*#+}} xmm4 = xmm4[0],xmm12[0],xmm4[1],xmm12[1]5027; SSE-NEXT: punpcklqdq {{.*#+}} xmm4 = xmm4[0],xmm0[0]5028; SSE-NEXT: movdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5029; SSE-NEXT: movdqa %xmm3, %xmm05030; SSE-NEXT: punpckhdq {{.*#+}} xmm0 = xmm0[2],xmm12[2],xmm0[3],xmm12[3]5031; SSE-NEXT: movaps %xmm6, %xmm45032; SSE-NEXT: movaps %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5033; SSE-NEXT: shufps {{.*#+}} xmm4 = xmm4[1,1],xmm7[1,1]5034; SSE-NEXT: shufps {{.*#+}} xmm4 = xmm4[0,2],xmm0[0,1]5035; SSE-NEXT: movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5036; SSE-NEXT: movdqa %xmm3, %xmm05037; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,1],xmm7[2,3]5038; SSE-NEXT: unpckhpd {{.*#+}} xmm6 = xmm6[1],xmm1[1]5039; SSE-NEXT: shufps {{.*#+}} xmm6 = xmm6[2,0],xmm0[2,0]5040; SSE-NEXT: movaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5041; SSE-NEXT: movdqa %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5042; SSE-NEXT: movdqa %xmm15, %xmm05043; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm13[0],xmm0[1],xmm13[1]5044; SSE-NEXT: movdqa 64(%rdi), %xmm45045; SSE-NEXT: movdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5046; SSE-NEXT: movdqa %xmm4, %xmm35047; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload5048; SSE-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]5049; SSE-NEXT: punpcklqdq {{.*#+}} xmm3 = xmm3[0],xmm0[0]5050; SSE-NEXT: movdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5051; SSE-NEXT: movdqa %xmm4, %xmm05052; SSE-NEXT: punpckhdq {{.*#+}} xmm0 = xmm0[2],xmm2[2],xmm0[3],xmm2[3]5053; SSE-NEXT: punpckhdq {{.*#+}} xmm15 = xmm15[2],xmm13[2],xmm15[3],xmm13[3]5054; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload5055; SSE-NEXT: shufps {{.*#+}} xmm13 = xmm13[1,1],xmm1[1,1]5056; SSE-NEXT: shufps {{.*#+}} xmm13 = xmm13[0,2],xmm0[0,1]5057; SSE-NEXT: movaps %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5058; SSE-NEXT: movdqa %xmm4, %xmm05059; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,0],xmm1[2,0]5060; SSE-NEXT: shufps {{.*#+}} xmm15 = xmm15[0,1],xmm0[2,0]5061; SSE-NEXT: movaps %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5062; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload5063; SSE-NEXT: movaps %xmm2, %xmm05064; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm14[0],xmm0[1],xmm14[1]5065; SSE-NEXT: movaps 80(%rdi), %xmm35066; SSE-NEXT: movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5067; SSE-NEXT: movaps %xmm3, %xmm45068; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload5069; SSE-NEXT: unpcklps {{.*#+}} xmm4 = xmm4[0],xmm1[0],xmm4[1],xmm1[1]5070; SSE-NEXT: movlhps {{.*#+}} xmm4 = xmm4[0],xmm0[0]5071; SSE-NEXT: movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5072; SSE-NEXT: movaps %xmm3, %xmm05073; SSE-NEXT: unpckhps {{.*#+}} xmm0 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]5074; SSE-NEXT: movaps %xmm14, %xmm45075; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload5076; SSE-NEXT: shufps {{.*#+}} xmm4 = xmm4[1,1],xmm1[1,1]5077; SSE-NEXT: shufps {{.*#+}} xmm4 = xmm4[0,2],xmm0[0,1]5078; SSE-NEXT: movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5079; SSE-NEXT: movaps %xmm3, %xmm05080; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,1],xmm1[2,3]5081; SSE-NEXT: unpckhpd {{.*#+}} xmm14 = xmm14[1],xmm2[1]5082; SSE-NEXT: shufps {{.*#+}} xmm14 = xmm14[2,0],xmm0[2,0]5083; SSE-NEXT: movaps %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5084; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload5085; SSE-NEXT: movaps %xmm1, %xmm05086; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload5087; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm4[0],xmm0[1],xmm4[1]5088; SSE-NEXT: movaps 96(%rdi), %xmm55089; SSE-NEXT: movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5090; SSE-NEXT: movaps %xmm5, %xmm35091; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload5092; SSE-NEXT: unpcklps {{.*#+}} xmm3 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]5093; SSE-NEXT: movlhps {{.*#+}} xmm3 = xmm3[0],xmm0[0]5094; SSE-NEXT: movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5095; SSE-NEXT: movaps %xmm5, %xmm05096; SSE-NEXT: unpckhps {{.*#+}} xmm0 = xmm0[2],xmm2[2],xmm0[3],xmm2[3]5097; SSE-NEXT: movaps %xmm1, %xmm25098; SSE-NEXT: unpckhps {{.*#+}} xmm2 = xmm2[2],xmm4[2],xmm2[3],xmm4[3]5099; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload5100; SSE-NEXT: shufps {{.*#+}} xmm4 = xmm4[1,1],xmm1[1,1]5101; SSE-NEXT: shufps {{.*#+}} xmm4 = xmm4[0,2],xmm0[0,1]5102; SSE-NEXT: movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5103; SSE-NEXT: movaps %xmm5, %xmm05104; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,0],xmm1[2,0]5105; SSE-NEXT: shufps {{.*#+}} xmm2 = xmm2[0,1],xmm0[2,0]5106; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5107; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload5108; SSE-NEXT: movaps %xmm2, %xmm05109; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload5110; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm5[0],xmm0[1],xmm5[1]5111; SSE-NEXT: movaps 112(%rdi), %xmm35112; SSE-NEXT: movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5113; SSE-NEXT: movaps %xmm3, %xmm45114; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload5115; SSE-NEXT: unpcklps {{.*#+}} xmm4 = xmm4[0],xmm1[0],xmm4[1],xmm1[1]5116; SSE-NEXT: movlhps {{.*#+}} xmm4 = xmm4[0],xmm0[0]5117; SSE-NEXT: movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5118; SSE-NEXT: movaps %xmm3, %xmm05119; SSE-NEXT: unpckhps {{.*#+}} xmm0 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]5120; SSE-NEXT: movaps %xmm5, %xmm45121; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload5122; SSE-NEXT: shufps {{.*#+}} xmm4 = xmm4[1,1],xmm1[1,1]5123; SSE-NEXT: shufps {{.*#+}} xmm4 = xmm4[0,2],xmm0[0,1]5124; SSE-NEXT: movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5125; SSE-NEXT: movaps %xmm3, %xmm05126; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,1],xmm1[2,3]5127; SSE-NEXT: unpckhpd {{.*#+}} xmm5 = xmm5[1],xmm2[1]5128; SSE-NEXT: shufps {{.*#+}} xmm5 = xmm5[2,0],xmm0[2,0]5129; SSE-NEXT: movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5130; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload5131; SSE-NEXT: movaps %xmm1, %xmm05132; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload5133; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm4[0],xmm0[1],xmm4[1]5134; SSE-NEXT: movaps 128(%rdi), %xmm55135; SSE-NEXT: movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5136; SSE-NEXT: movaps %xmm5, %xmm35137; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload5138; SSE-NEXT: unpcklps {{.*#+}} xmm3 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]5139; SSE-NEXT: movlhps {{.*#+}} xmm3 = xmm3[0],xmm0[0]5140; SSE-NEXT: movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5141; SSE-NEXT: movaps %xmm5, %xmm05142; SSE-NEXT: unpckhps {{.*#+}} xmm0 = xmm0[2],xmm2[2],xmm0[3],xmm2[3]5143; SSE-NEXT: movaps %xmm1, %xmm25144; SSE-NEXT: unpckhps {{.*#+}} xmm2 = xmm2[2],xmm4[2],xmm2[3],xmm4[3]5145; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload5146; SSE-NEXT: shufps {{.*#+}} xmm4 = xmm4[1,1],xmm1[1,1]5147; SSE-NEXT: shufps {{.*#+}} xmm4 = xmm4[0,2],xmm0[0,1]5148; SSE-NEXT: movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5149; SSE-NEXT: movaps %xmm5, %xmm05150; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,0],xmm1[2,0]5151; SSE-NEXT: shufps {{.*#+}} xmm2 = xmm2[0,1],xmm0[2,0]5152; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5153; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload5154; SSE-NEXT: movaps %xmm3, %xmm05155; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload5156; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]5157; SSE-NEXT: movaps 144(%rdi), %xmm45158; SSE-NEXT: movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5159; SSE-NEXT: movaps %xmm4, %xmm55160; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload5161; SSE-NEXT: unpcklps {{.*#+}} xmm5 = xmm5[0],xmm2[0],xmm5[1],xmm2[1]5162; SSE-NEXT: movlhps {{.*#+}} xmm5 = xmm5[0],xmm0[0]5163; SSE-NEXT: movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5164; SSE-NEXT: movaps %xmm4, %xmm05165; SSE-NEXT: unpckhps {{.*#+}} xmm0 = xmm0[2],xmm2[2],xmm0[3],xmm2[3]5166; SSE-NEXT: movaps %xmm1, %xmm55167; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload5168; SSE-NEXT: shufps {{.*#+}} xmm5 = xmm5[1,1],xmm2[1,1]5169; SSE-NEXT: shufps {{.*#+}} xmm5 = xmm5[0,2],xmm0[0,1]5170; SSE-NEXT: movaps %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5171; SSE-NEXT: movaps %xmm4, %xmm05172; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,1],xmm2[2,3]5173; SSE-NEXT: unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm3[1]5174; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[2,0],xmm0[2,0]5175; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5176; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload5177; SSE-NEXT: movaps %xmm2, %xmm05178; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload5179; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]5180; SSE-NEXT: movaps 160(%rdi), %xmm145181; SSE-NEXT: movaps %xmm14, %xmm45182; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload5183; SSE-NEXT: unpcklps {{.*#+}} xmm4 = xmm4[0],xmm3[0],xmm4[1],xmm3[1]5184; SSE-NEXT: movlhps {{.*#+}} xmm4 = xmm4[0],xmm0[0]5185; SSE-NEXT: movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5186; SSE-NEXT: movaps %xmm14, %xmm05187; SSE-NEXT: unpckhps {{.*#+}} xmm0 = xmm0[2],xmm3[2],xmm0[3],xmm3[3]5188; SSE-NEXT: movaps %xmm2, %xmm35189; SSE-NEXT: unpckhps {{.*#+}} xmm3 = xmm3[2],xmm1[2],xmm3[3],xmm1[3]5190; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload5191; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[1,1],xmm2[1,1]5192; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,2],xmm0[0,1]5193; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5194; SSE-NEXT: movaps %xmm14, %xmm05195; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,0],xmm2[2,0]5196; SSE-NEXT: shufps {{.*#+}} xmm3 = xmm3[0,1],xmm0[2,0]5197; SSE-NEXT: movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5198; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload5199; SSE-NEXT: movaps %xmm3, %xmm05200; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload5201; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]5202; SSE-NEXT: movaps 176(%rdi), %xmm115203; SSE-NEXT: movaps %xmm11, %xmm45204; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload5205; SSE-NEXT: unpcklps {{.*#+}} xmm4 = xmm4[0],xmm2[0],xmm4[1],xmm2[1]5206; SSE-NEXT: movlhps {{.*#+}} xmm4 = xmm4[0],xmm0[0]5207; SSE-NEXT: movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5208; SSE-NEXT: movaps %xmm11, %xmm05209; SSE-NEXT: unpckhps {{.*#+}} xmm0 = xmm0[2],xmm2[2],xmm0[3],xmm2[3]5210; SSE-NEXT: movaps %xmm1, %xmm45211; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload5212; SSE-NEXT: shufps {{.*#+}} xmm4 = xmm4[1,1],xmm2[1,1]5213; SSE-NEXT: shufps {{.*#+}} xmm4 = xmm4[0,2],xmm0[0,1]5214; SSE-NEXT: movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5215; SSE-NEXT: movaps %xmm11, %xmm05216; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,1],xmm2[2,3]5217; SSE-NEXT: unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm3[1]5218; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[2,0],xmm0[2,0]5219; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5220; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload5221; SSE-NEXT: movaps %xmm2, %xmm05222; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload5223; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]5224; SSE-NEXT: movaps 192(%rdi), %xmm105225; SSE-NEXT: movaps %xmm10, %xmm45226; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload5227; SSE-NEXT: unpcklps {{.*#+}} xmm4 = xmm4[0],xmm3[0],xmm4[1],xmm3[1]5228; SSE-NEXT: movlhps {{.*#+}} xmm4 = xmm4[0],xmm0[0]5229; SSE-NEXT: movaps %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5230; SSE-NEXT: movaps %xmm10, %xmm05231; SSE-NEXT: unpckhps {{.*#+}} xmm0 = xmm0[2],xmm3[2],xmm0[3],xmm3[3]5232; SSE-NEXT: movaps %xmm2, %xmm35233; SSE-NEXT: unpckhps {{.*#+}} xmm3 = xmm3[2],xmm1[2],xmm3[3],xmm1[3]5234; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload5235; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[1,1],xmm2[1,1]5236; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,2],xmm0[0,1]5237; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5238; SSE-NEXT: movaps %xmm10, %xmm05239; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,0],xmm2[2,0]5240; SSE-NEXT: shufps {{.*#+}} xmm3 = xmm3[0,1],xmm0[2,0]5241; SSE-NEXT: movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5242; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload5243; SSE-NEXT: movaps %xmm3, %xmm05244; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload5245; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]5246; SSE-NEXT: movaps 208(%rdi), %xmm125247; SSE-NEXT: movaps %xmm12, %xmm135248; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload5249; SSE-NEXT: unpcklps {{.*#+}} xmm13 = xmm13[0],xmm2[0],xmm13[1],xmm2[1]5250; SSE-NEXT: movlhps {{.*#+}} xmm13 = xmm13[0],xmm0[0]5251; SSE-NEXT: movaps %xmm12, %xmm05252; SSE-NEXT: unpckhps {{.*#+}} xmm0 = xmm0[2],xmm2[2],xmm0[3],xmm2[3]5253; SSE-NEXT: movaps %xmm1, %xmm95254; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload5255; SSE-NEXT: shufps {{.*#+}} xmm9 = xmm9[1,1],xmm2[1,1]5256; SSE-NEXT: shufps {{.*#+}} xmm9 = xmm9[0,2],xmm0[0,1]5257; SSE-NEXT: movaps %xmm12, %xmm05258; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,1],xmm2[2,3]5259; SSE-NEXT: unpckhpd {{.*#+}} xmm1 = xmm1[1],xmm3[1]5260; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[2,0],xmm0[2,0]5261; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5262; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload5263; SSE-NEXT: movaps %xmm7, %xmm05264; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload5265; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]5266; SSE-NEXT: movaps 224(%rdi), %xmm45267; SSE-NEXT: movaps %xmm4, %xmm85268; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload5269; SSE-NEXT: unpcklps {{.*#+}} xmm8 = xmm8[0],xmm3[0],xmm8[1],xmm3[1]5270; SSE-NEXT: movlhps {{.*#+}} xmm8 = xmm8[0],xmm0[0]5271; SSE-NEXT: movaps %xmm4, %xmm05272; SSE-NEXT: unpckhps {{.*#+}} xmm0 = xmm0[2],xmm3[2],xmm0[3],xmm3[3]5273; SSE-NEXT: unpckhps {{.*#+}} xmm7 = xmm7[2],xmm1[2],xmm7[3],xmm1[3]5274; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload5275; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[1,1],xmm2[1,1]5276; SSE-NEXT: shufps {{.*#+}} xmm1 = xmm1[0,2],xmm0[0,1]5277; SSE-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5278; SSE-NEXT: movaps %xmm4, %xmm05279; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,0],xmm2[2,0]5280; SSE-NEXT: shufps {{.*#+}} xmm7 = xmm7[0,1],xmm0[2,0]5281; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload5282; SSE-NEXT: movaps %xmm2, %xmm15283; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Reload5284; SSE-NEXT: unpcklps {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1]5285; SSE-NEXT: movaps 240(%rdi), %xmm55286; SSE-NEXT: movaps %xmm5, %xmm65287; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload5288; SSE-NEXT: unpcklps {{.*#+}} xmm6 = xmm6[0],xmm15[0],xmm6[1],xmm15[1]5289; SSE-NEXT: movlhps {{.*#+}} xmm6 = xmm6[0],xmm1[0]5290; SSE-NEXT: movaps %xmm5, %xmm05291; SSE-NEXT: unpckhps {{.*#+}} xmm0 = xmm0[2],xmm15[2],xmm0[3],xmm15[3]5292; SSE-NEXT: movaps %xmm3, %xmm155293; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload5294; SSE-NEXT: shufps {{.*#+}} xmm3 = xmm3[1,1],xmm1[1,1]5295; SSE-NEXT: shufps {{.*#+}} xmm3 = xmm3[0,2],xmm0[0,1]5296; SSE-NEXT: movaps %xmm5, %xmm05297; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,1],xmm1[2,3]5298; SSE-NEXT: unpckhpd {{.*#+}} xmm15 = xmm15[1],xmm2[1]5299; SSE-NEXT: movaps %xmm2, %xmm15300; SSE-NEXT: shufps {{.*#+}} xmm15 = xmm15[2,0],xmm0[2,0]5301; SSE-NEXT: movaps %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5302; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload5303; SSE-NEXT: shufps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Folded Reload5304; SSE-NEXT: # xmm2 = xmm2[1,1],mem[1,1]5305; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5306; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[0,1],xmm2[2,0]5307; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload5308; SSE-NEXT: movss {{.*#+}} xmm0 = xmm2[0],xmm0[1,2,3]5309; SSE-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5310; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5311; SSE-NEXT: shufps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload5312; SSE-NEXT: # xmm0 = xmm0[1,1],mem[1,1]5313; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload5314; SSE-NEXT: shufps {{.*#+}} xmm2 = xmm2[0,1],xmm0[2,0]5315; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5316; SSE-NEXT: movss {{.*#+}} xmm2 = xmm0[0],xmm2[1,2,3]5317; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5318; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5319; SSE-NEXT: shufps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload5320; SSE-NEXT: # xmm0 = xmm0[1,1],mem[1,1]5321; SSE-NEXT: movaps (%rsp), %xmm2 # 16-byte Reload5322; SSE-NEXT: shufps {{.*#+}} xmm2 = xmm2[0,1],xmm0[2,0]5323; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5324; SSE-NEXT: movss {{.*#+}} xmm2 = xmm0[0],xmm2[1,2,3]5325; SSE-NEXT: movaps %xmm2, (%rsp) # 16-byte Spill5326; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5327; SSE-NEXT: shufps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload5328; SSE-NEXT: # xmm0 = xmm0[1,1],mem[1,1]5329; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload5330; SSE-NEXT: shufps {{.*#+}} xmm2 = xmm2[0,1],xmm0[2,0]5331; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5332; SSE-NEXT: movss {{.*#+}} xmm2 = xmm0[0],xmm2[1,2,3]5333; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5334; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5335; SSE-NEXT: shufps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload5336; SSE-NEXT: # xmm0 = xmm0[1,1],mem[1,1]5337; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload5338; SSE-NEXT: shufps {{.*#+}} xmm2 = xmm2[0,1],xmm0[2,0]5339; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5340; SSE-NEXT: movss {{.*#+}} xmm2 = xmm0[0],xmm2[1,2,3]5341; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5342; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5343; SSE-NEXT: shufps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload5344; SSE-NEXT: # xmm0 = xmm0[1,1],mem[1,1]5345; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload5346; SSE-NEXT: shufps {{.*#+}} xmm2 = xmm2[0,1],xmm0[2,0]5347; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5348; SSE-NEXT: movss {{.*#+}} xmm2 = xmm0[0],xmm2[1,2,3]5349; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5350; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5351; SSE-NEXT: shufps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload5352; SSE-NEXT: # xmm0 = xmm0[1,1],mem[1,1]5353; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload5354; SSE-NEXT: shufps {{.*#+}} xmm2 = xmm2[0,1],xmm0[2,0]5355; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5356; SSE-NEXT: movss {{.*#+}} xmm2 = xmm0[0],xmm2[1,2,3]5357; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5358; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5359; SSE-NEXT: shufps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload5360; SSE-NEXT: # xmm0 = xmm0[1,1],mem[1,1]5361; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload5362; SSE-NEXT: shufps {{.*#+}} xmm2 = xmm2[0,1],xmm0[2,0]5363; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5364; SSE-NEXT: movss {{.*#+}} xmm2 = xmm0[0],xmm2[1,2,3]5365; SSE-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill5366; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5367; SSE-NEXT: shufps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload5368; SSE-NEXT: # xmm0 = xmm0[1,1],mem[1,1]5369; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload5370; SSE-NEXT: shufps {{.*#+}} xmm15 = xmm15[0,1],xmm0[2,0]5371; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5372; SSE-NEXT: movss {{.*#+}} xmm15 = xmm0[0],xmm15[1,2,3]5373; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5374; SSE-NEXT: shufps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload5375; SSE-NEXT: # xmm0 = xmm0[1,1],mem[1,1]5376; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload5377; SSE-NEXT: shufps {{.*#+}} xmm2 = xmm2[0,1],xmm0[2,0]5378; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5379; SSE-NEXT: movss {{.*#+}} xmm2 = xmm0[0],xmm2[1,2,3]5380; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5381; SSE-NEXT: shufps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload5382; SSE-NEXT: # xmm0 = xmm0[1,1],mem[1,1]5383; SSE-NEXT: shufps {{.*#+}} xmm14 = xmm14[0,1],xmm0[2,0]5384; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5385; SSE-NEXT: movss {{.*#+}} xmm14 = xmm0[0],xmm14[1,2,3]5386; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5387; SSE-NEXT: shufps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload5388; SSE-NEXT: # xmm0 = xmm0[1,1],mem[1,1]5389; SSE-NEXT: shufps {{.*#+}} xmm11 = xmm11[0,1],xmm0[2,0]5390; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5391; SSE-NEXT: movss {{.*#+}} xmm11 = xmm0[0],xmm11[1,2,3]5392; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5393; SSE-NEXT: shufps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload5394; SSE-NEXT: # xmm0 = xmm0[1,1],mem[1,1]5395; SSE-NEXT: shufps {{.*#+}} xmm10 = xmm10[0,1],xmm0[2,0]5396; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5397; SSE-NEXT: movss {{.*#+}} xmm10 = xmm0[0],xmm10[1,2,3]5398; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5399; SSE-NEXT: shufps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload5400; SSE-NEXT: # xmm0 = xmm0[1,1],mem[1,1]5401; SSE-NEXT: shufps {{.*#+}} xmm12 = xmm12[0,1],xmm0[2,0]5402; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5403; SSE-NEXT: movss {{.*#+}} xmm12 = xmm0[0],xmm12[1,2,3]5404; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5405; SSE-NEXT: shufps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload5406; SSE-NEXT: # xmm0 = xmm0[1,1],mem[1,1]5407; SSE-NEXT: shufps {{.*#+}} xmm4 = xmm4[0,1],xmm0[2,0]5408; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5409; SSE-NEXT: movss {{.*#+}} xmm4 = xmm0[0],xmm4[1,2,3]5410; SSE-NEXT: movaps %xmm1, %xmm05411; SSE-NEXT: shufps $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload5412; SSE-NEXT: # xmm0 = xmm0[1,1],mem[1,1]5413; SSE-NEXT: shufps {{.*#+}} xmm5 = xmm5[0,1],xmm0[2,0]5414; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5415; SSE-NEXT: movss {{.*#+}} xmm5 = xmm0[0],xmm5[1,2,3]5416; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5417; SSE-NEXT: movaps %xmm0, 1248(%r9)5418; SSE-NEXT: movaps %xmm3, 1232(%r9)5419; SSE-NEXT: movaps %xmm6, 1200(%r9)5420; SSE-NEXT: movaps %xmm7, 1168(%r9)5421; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5422; SSE-NEXT: movaps %xmm0, 1152(%r9)5423; SSE-NEXT: movaps %xmm8, 1120(%r9)5424; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5425; SSE-NEXT: movaps %xmm0, 1088(%r9)5426; SSE-NEXT: movaps %xmm9, 1072(%r9)5427; SSE-NEXT: movaps %xmm13, 1040(%r9)5428; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5429; SSE-NEXT: movaps %xmm0, 1008(%r9)5430; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5431; SSE-NEXT: movaps %xmm0, 992(%r9)5432; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5433; SSE-NEXT: movaps %xmm0, 960(%r9)5434; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5435; SSE-NEXT: movaps %xmm0, 928(%r9)5436; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5437; SSE-NEXT: movaps %xmm0, 912(%r9)5438; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5439; SSE-NEXT: movaps %xmm0, 880(%r9)5440; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5441; SSE-NEXT: movaps %xmm0, 848(%r9)5442; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5443; SSE-NEXT: movaps %xmm0, 832(%r9)5444; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5445; SSE-NEXT: movaps %xmm0, 800(%r9)5446; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5447; SSE-NEXT: movaps %xmm0, 768(%r9)5448; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5449; SSE-NEXT: movaps %xmm0, 752(%r9)5450; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5451; SSE-NEXT: movaps %xmm0, 720(%r9)5452; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5453; SSE-NEXT: movaps %xmm0, 688(%r9)5454; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5455; SSE-NEXT: movaps %xmm0, 672(%r9)5456; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5457; SSE-NEXT: movaps %xmm0, 640(%r9)5458; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5459; SSE-NEXT: movaps %xmm0, 608(%r9)5460; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5461; SSE-NEXT: movaps %xmm0, 592(%r9)5462; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5463; SSE-NEXT: movaps %xmm0, 560(%r9)5464; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5465; SSE-NEXT: movaps %xmm0, 528(%r9)5466; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5467; SSE-NEXT: movaps %xmm0, 512(%r9)5468; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5469; SSE-NEXT: movaps %xmm0, 480(%r9)5470; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5471; SSE-NEXT: movaps %xmm0, 448(%r9)5472; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5473; SSE-NEXT: movaps %xmm0, 432(%r9)5474; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5475; SSE-NEXT: movaps %xmm0, 400(%r9)5476; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5477; SSE-NEXT: movaps %xmm0, 368(%r9)5478; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5479; SSE-NEXT: movaps %xmm0, 352(%r9)5480; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5481; SSE-NEXT: movaps %xmm0, 320(%r9)5482; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5483; SSE-NEXT: movaps %xmm0, 288(%r9)5484; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5485; SSE-NEXT: movaps %xmm0, 272(%r9)5486; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5487; SSE-NEXT: movaps %xmm0, 240(%r9)5488; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5489; SSE-NEXT: movaps %xmm0, 208(%r9)5490; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5491; SSE-NEXT: movaps %xmm0, 192(%r9)5492; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5493; SSE-NEXT: movaps %xmm0, 160(%r9)5494; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5495; SSE-NEXT: movaps %xmm0, 128(%r9)5496; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5497; SSE-NEXT: movaps %xmm0, 112(%r9)5498; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5499; SSE-NEXT: movaps %xmm0, 80(%r9)5500; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5501; SSE-NEXT: movaps %xmm0, 48(%r9)5502; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5503; SSE-NEXT: movaps %xmm0, 32(%r9)5504; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5505; SSE-NEXT: movaps %xmm0, (%r9)5506; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5507; SSE-NEXT: movaps %xmm0, 1264(%r9)5508; SSE-NEXT: movaps %xmm5, 1216(%r9)5509; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5510; SSE-NEXT: movaps %xmm0, 1184(%r9)5511; SSE-NEXT: movaps %xmm4, 1136(%r9)5512; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5513; SSE-NEXT: movaps %xmm0, 1104(%r9)5514; SSE-NEXT: movaps %xmm12, 1056(%r9)5515; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5516; SSE-NEXT: movaps %xmm0, 1024(%r9)5517; SSE-NEXT: movaps %xmm10, 976(%r9)5518; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5519; SSE-NEXT: movaps %xmm0, 944(%r9)5520; SSE-NEXT: movaps %xmm11, 896(%r9)5521; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5522; SSE-NEXT: movaps %xmm0, 864(%r9)5523; SSE-NEXT: movaps %xmm14, 816(%r9)5524; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5525; SSE-NEXT: movaps %xmm0, 784(%r9)5526; SSE-NEXT: movaps %xmm2, 736(%r9)5527; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5528; SSE-NEXT: movaps %xmm0, 704(%r9)5529; SSE-NEXT: movaps %xmm15, 656(%r9)5530; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5531; SSE-NEXT: movaps %xmm0, 624(%r9)5532; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5533; SSE-NEXT: movaps %xmm0, 576(%r9)5534; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5535; SSE-NEXT: movaps %xmm0, 544(%r9)5536; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5537; SSE-NEXT: movaps %xmm0, 496(%r9)5538; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5539; SSE-NEXT: movaps %xmm0, 464(%r9)5540; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5541; SSE-NEXT: movaps %xmm0, 416(%r9)5542; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5543; SSE-NEXT: movaps %xmm0, 384(%r9)5544; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5545; SSE-NEXT: movaps %xmm0, 336(%r9)5546; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5547; SSE-NEXT: movaps %xmm0, 304(%r9)5548; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5549; SSE-NEXT: movaps %xmm0, 256(%r9)5550; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5551; SSE-NEXT: movaps %xmm0, 224(%r9)5552; SSE-NEXT: movaps (%rsp), %xmm0 # 16-byte Reload5553; SSE-NEXT: movaps %xmm0, 176(%r9)5554; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5555; SSE-NEXT: movaps %xmm0, 144(%r9)5556; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5557; SSE-NEXT: movaps %xmm0, 96(%r9)5558; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5559; SSE-NEXT: movaps %xmm0, 64(%r9)5560; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload5561; SSE-NEXT: movaps %xmm0, 16(%r9)5562; SSE-NEXT: addq $1736, %rsp # imm = 0x6C85563; SSE-NEXT: retq5564;5565; AVX-LABEL: store_i32_stride5_vf64:5566; AVX: # %bb.0:5567; AVX-NEXT: subq $1512, %rsp # imm = 0x5E85568; AVX-NEXT: vmovaps (%rsi), %xmm35569; AVX-NEXT: vmovaps 32(%rsi), %xmm05570; AVX-NEXT: vmovaps (%rdi), %xmm45571; AVX-NEXT: vmovaps 32(%rdi), %xmm15572; AVX-NEXT: vinsertps {{.*#+}} xmm2 = zero,xmm4[1],xmm3[1],zero5573; AVX-NEXT: vunpcklps {{.*#+}} xmm5 = xmm4[0],xmm3[0],xmm4[1],xmm3[1]5574; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm5, %ymm55575; AVX-NEXT: vmovaps (%rcx), %xmm65576; AVX-NEXT: vmovaps 32(%rcx), %xmm25577; AVX-NEXT: vmovaps (%rdx), %xmm75578; AVX-NEXT: vmovaps 32(%rdx), %xmm85579; AVX-NEXT: vunpcklps {{.*#+}} xmm9 = xmm7[0],xmm6[0],xmm7[1],xmm6[1]5580; AVX-NEXT: vinsertf128 $1, %xmm7, %ymm9, %ymm95581; AVX-NEXT: vmovddup {{.*#+}} ymm9 = ymm9[0,0,2,2]5582; AVX-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1],ymm9[2,3],ymm5[4,5,6],ymm9[7]5583; AVX-NEXT: vbroadcastf128 {{.*#+}} ymm9 = mem[0,1,0,1]5584; AVX-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm9[4],ymm5[5,6,7]5585; AVX-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5586; AVX-NEXT: vunpckhps {{.*#+}} xmm5 = xmm7[2],xmm6[2],xmm7[3],xmm6[3]5587; AVX-NEXT: vshufps {{.*#+}} xmm6 = xmm6[1,1,1,1]5588; AVX-NEXT: vinsertf128 $1, %xmm5, %ymm6, %ymm55589; AVX-NEXT: vshufps {{.*#+}} xmm3 = xmm3[2,2,2,2]5590; AVX-NEXT: vblendps {{.*#+}} xmm3 = xmm4[0,1,2],xmm3[3]5591; AVX-NEXT: vinsertf128 $1, %xmm4, %ymm3, %ymm35592; AVX-NEXT: vblendps {{.*#+}} ymm3 = ymm5[0,1],ymm3[2,3],ymm5[4,5],ymm3[6,7]5593; AVX-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0],ymm9[1],ymm3[2,3,4,5],ymm9[6],ymm3[7]5594; AVX-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5595; AVX-NEXT: vinsertps {{.*#+}} xmm3 = zero,xmm1[1],xmm0[1],zero5596; AVX-NEXT: vunpcklps {{.*#+}} xmm4 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]5597; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm4, %ymm35598; AVX-NEXT: vunpcklps {{.*#+}} xmm4 = xmm8[0],xmm2[0],xmm8[1],xmm2[1]5599; AVX-NEXT: vinsertf128 $1, %xmm8, %ymm4, %ymm45600; AVX-NEXT: vmovddup {{.*#+}} ymm4 = ymm4[0,0,2,2]5601; AVX-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1],ymm4[2,3],ymm3[4,5,6],ymm4[7]5602; AVX-NEXT: vbroadcastf128 {{.*#+}} ymm4 = mem[0,1,0,1]5603; AVX-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1,2,3],ymm4[4],ymm3[5,6,7]5604; AVX-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5605; AVX-NEXT: vunpckhps {{.*#+}} xmm3 = xmm8[2],xmm2[2],xmm8[3],xmm2[3]5606; AVX-NEXT: vshufps {{.*#+}} xmm2 = xmm2[1,1,1,1]5607; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm2, %ymm25608; AVX-NEXT: vshufps {{.*#+}} xmm0 = xmm0[2,2,2,2]5609; AVX-NEXT: vblendps {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[3]5610; AVX-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm05611; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1],ymm0[2,3],ymm2[4,5],ymm0[6,7]5612; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm4[1],ymm0[2,3,4,5],ymm4[6],ymm0[7]5613; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5614; AVX-NEXT: vmovaps 64(%rsi), %xmm05615; AVX-NEXT: vmovaps 64(%rdi), %xmm15616; AVX-NEXT: vinsertps {{.*#+}} xmm2 = zero,xmm1[1],xmm0[1],zero5617; AVX-NEXT: vunpcklps {{.*#+}} xmm3 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]5618; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm3, %ymm25619; AVX-NEXT: vmovaps 64(%rcx), %xmm35620; AVX-NEXT: vmovaps 64(%rdx), %xmm45621; AVX-NEXT: vunpcklps {{.*#+}} xmm5 = xmm4[0],xmm3[0],xmm4[1],xmm3[1]5622; AVX-NEXT: vinsertf128 $1, %xmm4, %ymm5, %ymm55623; AVX-NEXT: vmovddup {{.*#+}} ymm5 = ymm5[0,0,2,2]5624; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1],ymm5[2,3],ymm2[4,5,6],ymm5[7]5625; AVX-NEXT: vbroadcastf128 {{.*#+}} ymm5 = mem[0,1,0,1]5626; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm5[4],ymm2[5,6,7]5627; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5628; AVX-NEXT: vunpckhps {{.*#+}} xmm2 = xmm4[2],xmm3[2],xmm4[3],xmm3[3]5629; AVX-NEXT: vshufps {{.*#+}} xmm3 = xmm3[1,1,1,1]5630; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm3, %ymm25631; AVX-NEXT: vshufps {{.*#+}} xmm0 = xmm0[2,2,2,2]5632; AVX-NEXT: vblendps {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[3]5633; AVX-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm05634; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1],ymm0[2,3],ymm2[4,5],ymm0[6,7]5635; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm5[1],ymm0[2,3,4,5],ymm5[6],ymm0[7]5636; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5637; AVX-NEXT: vmovaps 96(%rsi), %xmm05638; AVX-NEXT: vmovaps 96(%rdi), %xmm15639; AVX-NEXT: vinsertps {{.*#+}} xmm2 = zero,xmm1[1],xmm0[1],zero5640; AVX-NEXT: vunpcklps {{.*#+}} xmm3 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]5641; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm3, %ymm25642; AVX-NEXT: vmovaps 96(%rcx), %xmm35643; AVX-NEXT: vmovaps 96(%rdx), %xmm45644; AVX-NEXT: vunpcklps {{.*#+}} xmm5 = xmm4[0],xmm3[0],xmm4[1],xmm3[1]5645; AVX-NEXT: vinsertf128 $1, %xmm4, %ymm5, %ymm55646; AVX-NEXT: vmovddup {{.*#+}} ymm5 = ymm5[0,0,2,2]5647; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1],ymm5[2,3],ymm2[4,5,6],ymm5[7]5648; AVX-NEXT: vbroadcastf128 {{.*#+}} ymm5 = mem[0,1,0,1]5649; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm5[4],ymm2[5,6,7]5650; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5651; AVX-NEXT: vunpckhps {{.*#+}} xmm2 = xmm4[2],xmm3[2],xmm4[3],xmm3[3]5652; AVX-NEXT: vshufps {{.*#+}} xmm3 = xmm3[1,1,1,1]5653; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm3, %ymm25654; AVX-NEXT: vshufps {{.*#+}} xmm0 = xmm0[2,2,2,2]5655; AVX-NEXT: vblendps {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[3]5656; AVX-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm05657; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1],ymm0[2,3],ymm2[4,5],ymm0[6,7]5658; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm5[1],ymm0[2,3,4,5],ymm5[6],ymm0[7]5659; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5660; AVX-NEXT: vmovaps 128(%rsi), %xmm05661; AVX-NEXT: vmovaps 128(%rdi), %xmm15662; AVX-NEXT: vinsertps {{.*#+}} xmm2 = zero,xmm1[1],xmm0[1],zero5663; AVX-NEXT: vunpcklps {{.*#+}} xmm3 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]5664; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm3, %ymm25665; AVX-NEXT: vmovaps 128(%rcx), %xmm35666; AVX-NEXT: vmovaps 128(%rdx), %xmm45667; AVX-NEXT: vunpcklps {{.*#+}} xmm5 = xmm4[0],xmm3[0],xmm4[1],xmm3[1]5668; AVX-NEXT: vinsertf128 $1, %xmm4, %ymm5, %ymm55669; AVX-NEXT: vmovddup {{.*#+}} ymm5 = ymm5[0,0,2,2]5670; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1],ymm5[2,3],ymm2[4,5,6],ymm5[7]5671; AVX-NEXT: vbroadcastf128 {{.*#+}} ymm5 = mem[0,1,0,1]5672; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm5[4],ymm2[5,6,7]5673; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5674; AVX-NEXT: vunpckhps {{.*#+}} xmm2 = xmm4[2],xmm3[2],xmm4[3],xmm3[3]5675; AVX-NEXT: vshufps {{.*#+}} xmm3 = xmm3[1,1,1,1]5676; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm3, %ymm25677; AVX-NEXT: vshufps {{.*#+}} xmm0 = xmm0[2,2,2,2]5678; AVX-NEXT: vblendps {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[3]5679; AVX-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm05680; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1],ymm0[2,3],ymm2[4,5],ymm0[6,7]5681; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm5[1],ymm0[2,3,4,5],ymm5[6],ymm0[7]5682; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5683; AVX-NEXT: vmovaps 160(%rsi), %xmm05684; AVX-NEXT: vmovaps 160(%rdi), %xmm15685; AVX-NEXT: vinsertps {{.*#+}} xmm2 = zero,xmm1[1],xmm0[1],zero5686; AVX-NEXT: vunpcklps {{.*#+}} xmm3 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]5687; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm3, %ymm25688; AVX-NEXT: vmovaps 160(%rcx), %xmm35689; AVX-NEXT: vmovaps 160(%rdx), %xmm45690; AVX-NEXT: vunpcklps {{.*#+}} xmm5 = xmm4[0],xmm3[0],xmm4[1],xmm3[1]5691; AVX-NEXT: vinsertf128 $1, %xmm4, %ymm5, %ymm55692; AVX-NEXT: vmovddup {{.*#+}} ymm5 = ymm5[0,0,2,2]5693; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1],ymm5[2,3],ymm2[4,5,6],ymm5[7]5694; AVX-NEXT: vbroadcastf128 {{.*#+}} ymm5 = mem[0,1,0,1]5695; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm5[4],ymm2[5,6,7]5696; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5697; AVX-NEXT: vunpckhps {{.*#+}} xmm2 = xmm4[2],xmm3[2],xmm4[3],xmm3[3]5698; AVX-NEXT: vshufps {{.*#+}} xmm3 = xmm3[1,1,1,1]5699; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm3, %ymm25700; AVX-NEXT: vshufps {{.*#+}} xmm0 = xmm0[2,2,2,2]5701; AVX-NEXT: vblendps {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[3]5702; AVX-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm05703; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1],ymm0[2,3],ymm2[4,5],ymm0[6,7]5704; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm5[1],ymm0[2,3,4,5],ymm5[6],ymm0[7]5705; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5706; AVX-NEXT: vmovaps 192(%rsi), %xmm05707; AVX-NEXT: vmovaps 192(%rdi), %xmm15708; AVX-NEXT: vinsertps {{.*#+}} xmm2 = zero,xmm1[1],xmm0[1],zero5709; AVX-NEXT: vunpcklps {{.*#+}} xmm3 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]5710; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm3, %ymm25711; AVX-NEXT: vmovaps 192(%rcx), %xmm35712; AVX-NEXT: vmovaps 192(%rdx), %xmm45713; AVX-NEXT: vunpcklps {{.*#+}} xmm5 = xmm4[0],xmm3[0],xmm4[1],xmm3[1]5714; AVX-NEXT: vinsertf128 $1, %xmm4, %ymm5, %ymm55715; AVX-NEXT: vmovddup {{.*#+}} ymm5 = ymm5[0,0,2,2]5716; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1],ymm5[2,3],ymm2[4,5,6],ymm5[7]5717; AVX-NEXT: vunpckhps {{.*#+}} xmm4 = xmm4[2],xmm3[2],xmm4[3],xmm3[3]5718; AVX-NEXT: vshufps {{.*#+}} xmm3 = xmm3[1,1,1,1]5719; AVX-NEXT: vinsertf128 $1, %xmm4, %ymm3, %ymm35720; AVX-NEXT: vshufps {{.*#+}} xmm0 = xmm0[2,2,2,2]5721; AVX-NEXT: vblendps {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[3]5722; AVX-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm05723; AVX-NEXT: vbroadcastf128 {{.*#+}} ymm1 = mem[0,1,0,1]5724; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm1[4],ymm2[5,6,7]5725; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5726; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm3[0,1],ymm0[2,3],ymm3[4,5],ymm0[6,7]5727; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2,3,4,5],ymm1[6],ymm0[7]5728; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5729; AVX-NEXT: vmovaps 224(%rsi), %xmm05730; AVX-NEXT: vmovaps 224(%rdi), %xmm15731; AVX-NEXT: vinsertps {{.*#+}} xmm2 = zero,xmm1[1],xmm0[1],zero5732; AVX-NEXT: vunpcklps {{.*#+}} xmm3 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]5733; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm3, %ymm25734; AVX-NEXT: vmovaps 224(%rcx), %xmm35735; AVX-NEXT: vmovaps 224(%rdx), %xmm45736; AVX-NEXT: vunpcklps {{.*#+}} xmm5 = xmm4[0],xmm3[0],xmm4[1],xmm3[1]5737; AVX-NEXT: vinsertf128 $1, %xmm4, %ymm5, %ymm55738; AVX-NEXT: vmovddup {{.*#+}} ymm5 = ymm5[0,0,2,2]5739; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1],ymm5[2,3],ymm2[4,5,6],ymm5[7]5740; AVX-NEXT: vunpckhps {{.*#+}} xmm4 = xmm4[2],xmm3[2],xmm4[3],xmm3[3]5741; AVX-NEXT: vshufps {{.*#+}} xmm3 = xmm3[1,1,1,1]5742; AVX-NEXT: vinsertf128 $1, %xmm4, %ymm3, %ymm35743; AVX-NEXT: vshufps {{.*#+}} xmm0 = xmm0[2,2,2,2]5744; AVX-NEXT: vblendps {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[3]5745; AVX-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm05746; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm3[0,1],ymm0[2,3],ymm3[4,5],ymm0[6,7]5747; AVX-NEXT: vbroadcastf128 {{.*#+}} ymm1 = mem[0,1,0,1]5748; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm1[4],ymm2[5,6,7]5749; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5750; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2,3,4,5],ymm1[6],ymm0[7]5751; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5752; AVX-NEXT: vmovaps 16(%rsi), %xmm05753; AVX-NEXT: vmovaps 16(%rdi), %xmm15754; AVX-NEXT: vshufps {{.*#+}} xmm2 = xmm0[2,2,2,2]5755; AVX-NEXT: vblendps {{.*#+}} xmm2 = xmm1[0,1,2],xmm2[3]5756; AVX-NEXT: vinsertps {{.*#+}} xmm0 = zero,xmm1[1],xmm0[1],zero5757; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm05758; AVX-NEXT: vmovaps (%rdx), %ymm15759; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5760; AVX-NEXT: vmovaps (%rcx), %ymm135761; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm13[1,1],ymm1[1,1],ymm13[5,5],ymm1[5,5]5762; AVX-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm1[2,3,2,3]5763; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3,4],ymm0[5,6,7]5764; AVX-NEXT: vmovaps 16(%rcx), %xmm15765; AVX-NEXT: vmovaps 16(%rdx), %xmm25766; AVX-NEXT: vshufps {{.*#+}} xmm3 = xmm2[3,3],xmm1[3,3]5767; AVX-NEXT: vunpckhps {{.*#+}} xmm1 = xmm2[2],xmm1[2],xmm2[3],xmm1[3]5768; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm1, %ymm15769; AVX-NEXT: vmovaps (%rdi), %ymm145770; AVX-NEXT: vmovaps (%rsi), %ymm155771; AVX-NEXT: vshufps {{.*#+}} ymm2 = ymm15[3,3],ymm14[3,3],ymm15[7,7],ymm14[7,7]5772; AVX-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3,2,3]5773; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2],ymm2[3,4],ymm1[5,6,7]5774; AVX-NEXT: vbroadcastf128 {{.*#+}} ymm2 = mem[0,1,0,1]5775; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0],ymm0[1,2,3,4],ymm2[5],ymm0[6,7]5776; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5777; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm2[2],ymm1[3,4,5,6],ymm2[7]5778; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5779; AVX-NEXT: vmovaps 48(%rdi), %xmm05780; AVX-NEXT: vmovaps 48(%rsi), %xmm15781; AVX-NEXT: vshufps {{.*#+}} xmm2 = xmm1[2,2,2,2]5782; AVX-NEXT: vblendps {{.*#+}} xmm2 = xmm0[0,1,2],xmm2[3]5783; AVX-NEXT: vinsertps {{.*#+}} xmm0 = zero,xmm0[1],xmm1[1],zero5784; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm05785; AVX-NEXT: vmovaps 32(%rdx), %ymm15786; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5787; AVX-NEXT: vmovaps 32(%rcx), %ymm25788; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5789; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm2[1,1],ymm1[1,1],ymm2[5,5],ymm1[5,5]5790; AVX-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm1[2,3,2,3]5791; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3,4],ymm0[5,6,7]5792; AVX-NEXT: vmovaps 48(%rcx), %xmm15793; AVX-NEXT: vmovaps 48(%rdx), %xmm25794; AVX-NEXT: vshufps {{.*#+}} xmm3 = xmm2[3,3],xmm1[3,3]5795; AVX-NEXT: vunpckhps {{.*#+}} xmm1 = xmm2[2],xmm1[2],xmm2[3],xmm1[3]5796; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm1, %ymm15797; AVX-NEXT: vmovaps 32(%rdi), %ymm25798; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5799; AVX-NEXT: vmovaps 32(%rsi), %ymm35800; AVX-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5801; AVX-NEXT: vshufps {{.*#+}} ymm2 = ymm3[3,3],ymm2[3,3],ymm3[7,7],ymm2[7,7]5802; AVX-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3,2,3]5803; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2],ymm2[3,4],ymm1[5,6,7]5804; AVX-NEXT: vbroadcastf128 {{.*#+}} ymm2 = mem[0,1,0,1]5805; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0],ymm0[1,2,3,4],ymm2[5],ymm0[6,7]5806; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5807; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm2[2],ymm1[3,4,5,6],ymm2[7]5808; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5809; AVX-NEXT: vmovaps 80(%rdi), %xmm05810; AVX-NEXT: vmovaps 80(%rsi), %xmm15811; AVX-NEXT: vshufps {{.*#+}} xmm2 = xmm1[2,2,2,2]5812; AVX-NEXT: vblendps {{.*#+}} xmm2 = xmm0[0,1,2],xmm2[3]5813; AVX-NEXT: vinsertps {{.*#+}} xmm0 = zero,xmm0[1],xmm1[1],zero5814; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm05815; AVX-NEXT: vmovaps 64(%rdx), %ymm15816; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5817; AVX-NEXT: vmovaps 64(%rcx), %ymm25818; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5819; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm2[1,1],ymm1[1,1],ymm2[5,5],ymm1[5,5]5820; AVX-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm1[2,3,2,3]5821; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3,4],ymm0[5,6,7]5822; AVX-NEXT: vmovaps 80(%rcx), %xmm15823; AVX-NEXT: vmovaps 80(%rdx), %xmm25824; AVX-NEXT: vshufps {{.*#+}} xmm3 = xmm2[3,3],xmm1[3,3]5825; AVX-NEXT: vunpckhps {{.*#+}} xmm1 = xmm2[2],xmm1[2],xmm2[3],xmm1[3]5826; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm1, %ymm15827; AVX-NEXT: vmovaps 64(%rdi), %ymm25828; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5829; AVX-NEXT: vmovaps 64(%rsi), %ymm35830; AVX-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5831; AVX-NEXT: vshufps {{.*#+}} ymm2 = ymm3[3,3],ymm2[3,3],ymm3[7,7],ymm2[7,7]5832; AVX-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3,2,3]5833; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2],ymm2[3,4],ymm1[5,6,7]5834; AVX-NEXT: vbroadcastf128 {{.*#+}} ymm2 = mem[0,1,0,1]5835; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0],ymm0[1,2,3,4],ymm2[5],ymm0[6,7]5836; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5837; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm2[2],ymm1[3,4,5,6],ymm2[7]5838; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5839; AVX-NEXT: vmovaps 112(%rdi), %xmm05840; AVX-NEXT: vmovaps 112(%rsi), %xmm15841; AVX-NEXT: vshufps {{.*#+}} xmm2 = xmm1[2,2,2,2]5842; AVX-NEXT: vblendps {{.*#+}} xmm2 = xmm0[0,1,2],xmm2[3]5843; AVX-NEXT: vinsertps {{.*#+}} xmm0 = zero,xmm0[1],xmm1[1],zero5844; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm05845; AVX-NEXT: vmovaps 96(%rdx), %ymm15846; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5847; AVX-NEXT: vmovaps 96(%rcx), %ymm25848; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5849; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm2[1,1],ymm1[1,1],ymm2[5,5],ymm1[5,5]5850; AVX-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm1[2,3,2,3]5851; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3,4],ymm0[5,6,7]5852; AVX-NEXT: vmovaps 112(%rcx), %xmm15853; AVX-NEXT: vmovaps 112(%rdx), %xmm25854; AVX-NEXT: vshufps {{.*#+}} xmm3 = xmm2[3,3],xmm1[3,3]5855; AVX-NEXT: vunpckhps {{.*#+}} xmm1 = xmm2[2],xmm1[2],xmm2[3],xmm1[3]5856; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm1, %ymm15857; AVX-NEXT: vmovaps 96(%rdi), %ymm35858; AVX-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5859; AVX-NEXT: vmovaps 96(%rsi), %ymm25860; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5861; AVX-NEXT: vshufps {{.*#+}} ymm2 = ymm2[3,3],ymm3[3,3],ymm2[7,7],ymm3[7,7]5862; AVX-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3,2,3]5863; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2],ymm2[3,4],ymm1[5,6,7]5864; AVX-NEXT: vbroadcastf128 {{.*#+}} ymm2 = mem[0,1,0,1]5865; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0],ymm0[1,2,3,4],ymm2[5],ymm0[6,7]5866; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5867; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm2[2],ymm1[3,4,5,6],ymm2[7]5868; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5869; AVX-NEXT: vmovaps 144(%rdi), %xmm05870; AVX-NEXT: vmovaps 144(%rsi), %xmm15871; AVX-NEXT: vshufps {{.*#+}} xmm2 = xmm1[2,2,2,2]5872; AVX-NEXT: vblendps {{.*#+}} xmm2 = xmm0[0,1,2],xmm2[3]5873; AVX-NEXT: vinsertps {{.*#+}} xmm0 = zero,xmm0[1],xmm1[1],zero5874; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm05875; AVX-NEXT: vmovaps 128(%rdx), %ymm25876; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5877; AVX-NEXT: vmovaps 128(%rcx), %ymm15878; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5879; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm1[1,1],ymm2[1,1],ymm1[5,5],ymm2[5,5]5880; AVX-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm1[2,3,2,3]5881; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3,4],ymm0[5,6,7]5882; AVX-NEXT: vmovaps 144(%rcx), %xmm15883; AVX-NEXT: vmovaps 144(%rdx), %xmm25884; AVX-NEXT: vshufps {{.*#+}} xmm3 = xmm2[3,3],xmm1[3,3]5885; AVX-NEXT: vunpckhps {{.*#+}} xmm1 = xmm2[2],xmm1[2],xmm2[3],xmm1[3]5886; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm1, %ymm15887; AVX-NEXT: vmovaps 128(%rdi), %ymm125888; AVX-NEXT: vmovaps 128(%rsi), %ymm25889; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5890; AVX-NEXT: vshufps {{.*#+}} ymm2 = ymm2[3,3],ymm12[3,3],ymm2[7,7],ymm12[7,7]5891; AVX-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3,2,3]5892; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2],ymm2[3,4],ymm1[5,6,7]5893; AVX-NEXT: vbroadcastf128 {{.*#+}} ymm2 = mem[0,1,0,1]5894; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0],ymm0[1,2,3,4],ymm2[5],ymm0[6,7]5895; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5896; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm2[2],ymm1[3,4,5,6],ymm2[7]5897; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5898; AVX-NEXT: vmovaps 176(%rdi), %xmm05899; AVX-NEXT: vmovaps 176(%rsi), %xmm15900; AVX-NEXT: vshufps {{.*#+}} xmm2 = xmm1[2,2,2,2]5901; AVX-NEXT: vblendps {{.*#+}} xmm2 = xmm0[0,1,2],xmm2[3]5902; AVX-NEXT: vinsertps {{.*#+}} xmm0 = zero,xmm0[1],xmm1[1],zero5903; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm05904; AVX-NEXT: vmovaps 160(%rdx), %ymm15905; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5906; AVX-NEXT: vmovaps 160(%rcx), %ymm25907; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5908; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm2[1,1],ymm1[1,1],ymm2[5,5],ymm1[5,5]5909; AVX-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm1[2,3,2,3]5910; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3,4],ymm0[5,6,7]5911; AVX-NEXT: vmovaps 176(%rcx), %xmm15912; AVX-NEXT: vmovaps 176(%rdx), %xmm25913; AVX-NEXT: vshufps {{.*#+}} xmm3 = xmm2[3,3],xmm1[3,3]5914; AVX-NEXT: vunpckhps {{.*#+}} xmm1 = xmm2[2],xmm1[2],xmm2[3],xmm1[3]5915; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm1, %ymm15916; AVX-NEXT: vmovaps 160(%rdi), %ymm25917; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5918; AVX-NEXT: vmovaps 160(%rsi), %ymm115919; AVX-NEXT: vshufps {{.*#+}} ymm2 = ymm11[3,3],ymm2[3,3],ymm11[7,7],ymm2[7,7]5920; AVX-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3,2,3]5921; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2],ymm2[3,4],ymm1[5,6,7]5922; AVX-NEXT: vbroadcastf128 {{.*#+}} ymm2 = mem[0,1,0,1]5923; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0],ymm0[1,2,3,4],ymm2[5],ymm0[6,7]5924; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5925; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm2[2],ymm1[3,4,5,6],ymm2[7]5926; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5927; AVX-NEXT: vmovaps 208(%rdi), %xmm05928; AVX-NEXT: vmovaps 208(%rsi), %xmm15929; AVX-NEXT: vshufps {{.*#+}} xmm2 = xmm1[2,2,2,2]5930; AVX-NEXT: vblendps {{.*#+}} xmm2 = xmm0[0,1,2],xmm2[3]5931; AVX-NEXT: vinsertps {{.*#+}} xmm0 = zero,xmm0[1],xmm1[1],zero5932; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm05933; AVX-NEXT: vmovaps 192(%rdx), %ymm105934; AVX-NEXT: vmovaps 192(%rcx), %ymm95935; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm9[1,1],ymm10[1,1],ymm9[5,5],ymm10[5,5]5936; AVX-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm1[2,3,2,3]5937; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3,4],ymm0[5,6,7]5938; AVX-NEXT: vmovaps 208(%rcx), %xmm15939; AVX-NEXT: vmovaps 208(%rdx), %xmm25940; AVX-NEXT: vshufps {{.*#+}} xmm3 = xmm2[3,3],xmm1[3,3]5941; AVX-NEXT: vunpckhps {{.*#+}} xmm1 = xmm2[2],xmm1[2],xmm2[3],xmm1[3]5942; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm1, %ymm15943; AVX-NEXT: vmovaps 192(%rdi), %ymm75944; AVX-NEXT: vmovaps 192(%rsi), %ymm55945; AVX-NEXT: vshufps {{.*#+}} ymm2 = ymm5[3,3],ymm7[3,3],ymm5[7,7],ymm7[7,7]5946; AVX-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm2[2,3,2,3]5947; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0,1,2],ymm2[3,4],ymm1[5,6,7]5948; AVX-NEXT: vbroadcastf128 {{.*#+}} ymm2 = mem[0,1,0,1]5949; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0],ymm0[1,2,3,4],ymm2[5],ymm0[6,7]5950; AVX-NEXT: vmovups %ymm0, (%rsp) # 32-byte Spill5951; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm2[2],ymm1[3,4,5,6],ymm2[7]5952; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5953; AVX-NEXT: vmovaps 240(%rdi), %xmm05954; AVX-NEXT: vmovaps 240(%rsi), %xmm15955; AVX-NEXT: vshufps {{.*#+}} xmm2 = xmm1[2,2,2,2]5956; AVX-NEXT: vblendps {{.*#+}} xmm2 = xmm0[0,1,2],xmm2[3]5957; AVX-NEXT: vinsertps {{.*#+}} xmm0 = zero,xmm0[1],xmm1[1],zero5958; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm05959; AVX-NEXT: vmovaps 224(%rdx), %ymm45960; AVX-NEXT: vmovaps 224(%rcx), %ymm35961; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm3[1,1],ymm4[1,1],ymm3[5,5],ymm4[5,5]5962; AVX-NEXT: vperm2f128 {{.*#+}} ymm1 = ymm1[2,3,2,3]5963; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm0[0,1,2],ymm1[3,4],ymm0[5,6,7]5964; AVX-NEXT: vmovaps 240(%rcx), %xmm05965; AVX-NEXT: vmovaps 240(%rdx), %xmm15966; AVX-NEXT: vshufps {{.*#+}} xmm6 = xmm1[3,3],xmm0[3,3]5967; AVX-NEXT: vunpckhps {{.*#+}} xmm0 = xmm1[2],xmm0[2],xmm1[3],xmm0[3]5968; AVX-NEXT: vinsertf128 $1, %xmm6, %ymm0, %ymm65969; AVX-NEXT: vmovaps 224(%rdi), %ymm15970; AVX-NEXT: vmovaps 224(%rsi), %ymm05971; AVX-NEXT: vshufps {{.*#+}} ymm8 = ymm0[3,3],ymm1[3,3],ymm0[7,7],ymm1[7,7]5972; AVX-NEXT: vperm2f128 {{.*#+}} ymm8 = ymm8[2,3,2,3]5973; AVX-NEXT: vblendps {{.*#+}} ymm6 = ymm6[0,1,2],ymm8[3,4],ymm6[5,6,7]5974; AVX-NEXT: vbroadcastf128 {{.*#+}} ymm8 = mem[0,1,0,1]5975; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm8[0],ymm2[1,2,3,4],ymm8[5],ymm2[6,7]5976; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5977; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm6[0,1],ymm8[2],ymm6[3,4,5,6],ymm8[7]5978; AVX-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5979; AVX-NEXT: vshufps {{.*#+}} ymm2 = ymm15[3,0,2,3,7,4,6,7]5980; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm14[4],ymm2[5,6,7]5981; AVX-NEXT: vshufps {{.*#+}} ymm6 = ymm13[0,1,3,0,4,5,7,4]5982; AVX-NEXT: vpermilpd $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Folded Reload5983; AVX-NEXT: # ymm8 = mem[1,0,2,2]5984; AVX-NEXT: vblendps {{.*#+}} ymm6 = ymm8[0,1],ymm6[2],ymm8[3,4,5,6],ymm6[7]5985; AVX-NEXT: vblendps {{.*#+}} ymm8 = ymm2[0],ymm6[1,2,3],ymm2[4,5],ymm6[6,7]5986; AVX-NEXT: vpermilps $227, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Folded Reload5987; AVX-NEXT: # ymm2 = mem[3,0,2,3,7,4,6,7]5988; AVX-NEXT: vblendps $16, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload5989; AVX-NEXT: # ymm2 = ymm2[0,1,2,3],mem[4],ymm2[5,6,7]5990; AVX-NEXT: vpermilps $52, {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Folded Reload5991; AVX-NEXT: # ymm6 = mem[0,1,3,0,4,5,7,4]5992; AVX-NEXT: vpermilpd $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Folded Reload5993; AVX-NEXT: # ymm15 = mem[1,0,2,2]5994; AVX-NEXT: vblendps {{.*#+}} ymm6 = ymm15[0,1],ymm6[2],ymm15[3,4,5,6],ymm6[7]5995; AVX-NEXT: vblendps {{.*#+}} ymm6 = ymm2[0],ymm6[1,2,3],ymm2[4,5],ymm6[6,7]5996; AVX-NEXT: vpermilps $227, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Folded Reload5997; AVX-NEXT: # ymm2 = mem[3,0,2,3,7,4,6,7]5998; AVX-NEXT: vblendps $16, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload5999; AVX-NEXT: # ymm2 = ymm2[0,1,2,3],mem[4],ymm2[5,6,7]6000; AVX-NEXT: vpermilps $52, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Folded Reload6001; AVX-NEXT: # ymm15 = mem[0,1,3,0,4,5,7,4]6002; AVX-NEXT: vpermilpd $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Folded Reload6003; AVX-NEXT: # ymm14 = mem[1,0,2,2]6004; AVX-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1],ymm15[2],ymm14[3,4,5,6],ymm15[7]6005; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0],ymm14[1,2,3],ymm2[4,5],ymm14[6,7]6006; AVX-NEXT: vpermilps $227, {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Folded Reload6007; AVX-NEXT: # ymm14 = mem[3,0,2,3,7,4,6,7]6008; AVX-NEXT: vblendps $16, {{[-0-9]+}}(%r{{[sb]}}p), %ymm14, %ymm14 # 32-byte Folded Reload6009; AVX-NEXT: # ymm14 = ymm14[0,1,2,3],mem[4],ymm14[5,6,7]6010; AVX-NEXT: vpermilps $52, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Folded Reload6011; AVX-NEXT: # ymm15 = mem[0,1,3,0,4,5,7,4]6012; AVX-NEXT: vpermilpd $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Folded Reload6013; AVX-NEXT: # ymm13 = mem[1,0,2,2]6014; AVX-NEXT: vblendps {{.*#+}} ymm13 = ymm13[0,1],ymm15[2],ymm13[3,4,5,6],ymm15[7]6015; AVX-NEXT: vblendps {{.*#+}} ymm13 = ymm14[0],ymm13[1,2,3],ymm14[4,5],ymm13[6,7]6016; AVX-NEXT: vpermilps $227, {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Folded Reload6017; AVX-NEXT: # ymm14 = mem[3,0,2,3,7,4,6,7]6018; AVX-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1,2,3],ymm12[4],ymm14[5,6,7]6019; AVX-NEXT: vpermilps $52, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Folded Reload6020; AVX-NEXT: # ymm15 = mem[0,1,3,0,4,5,7,4]6021; AVX-NEXT: vpermilpd $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Folded Reload6022; AVX-NEXT: # ymm12 = mem[1,0,2,2]6023; AVX-NEXT: vblendps {{.*#+}} ymm12 = ymm12[0,1],ymm15[2],ymm12[3,4,5,6],ymm15[7]6024; AVX-NEXT: vblendps {{.*#+}} ymm12 = ymm14[0],ymm12[1,2,3],ymm14[4,5],ymm12[6,7]6025; AVX-NEXT: vshufps {{.*#+}} ymm11 = ymm11[3,0,2,3,7,4,6,7]6026; AVX-NEXT: vblendps $16, {{[-0-9]+}}(%r{{[sb]}}p), %ymm11, %ymm11 # 32-byte Folded Reload6027; AVX-NEXT: # ymm11 = ymm11[0,1,2,3],mem[4],ymm11[5,6,7]6028; AVX-NEXT: vpermilps $52, {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Folded Reload6029; AVX-NEXT: # ymm14 = mem[0,1,3,0,4,5,7,4]6030; AVX-NEXT: vpermilpd $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Folded Reload6031; AVX-NEXT: # ymm15 = mem[1,0,2,2]6032; AVX-NEXT: vblendps {{.*#+}} ymm14 = ymm15[0,1],ymm14[2],ymm15[3,4,5,6],ymm14[7]6033; AVX-NEXT: vblendps {{.*#+}} ymm11 = ymm11[0],ymm14[1,2,3],ymm11[4,5],ymm14[6,7]6034; AVX-NEXT: vshufps {{.*#+}} ymm5 = ymm5[3,0,2,3,7,4,6,7]6035; AVX-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm7[4],ymm5[5,6,7]6036; AVX-NEXT: vshufps {{.*#+}} ymm7 = ymm9[0,1,3,0,4,5,7,4]6037; AVX-NEXT: vshufpd {{.*#+}} ymm9 = ymm10[1,0,2,2]6038; AVX-NEXT: vblendps {{.*#+}} ymm7 = ymm9[0,1],ymm7[2],ymm9[3,4,5,6],ymm7[7]6039; AVX-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0],ymm7[1,2,3],ymm5[4,5],ymm7[6,7]6040; AVX-NEXT: vshufps {{.*#+}} ymm0 = ymm0[3,0,2,3,7,4,6,7]6041; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4],ymm0[5,6,7]6042; AVX-NEXT: vshufps {{.*#+}} ymm1 = ymm3[0,1,3,0,4,5,7,4]6043; AVX-NEXT: vshufpd {{.*#+}} ymm3 = ymm4[1,0,2,2]6044; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm3[0,1],ymm1[2],ymm3[3,4,5,6],ymm1[7]6045; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm1[1,2,3],ymm0[4,5],ymm1[6,7]6046; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm8[0,1,2],mem[3],ymm8[4,5,6,7]6047; AVX-NEXT: vblendps {{.*#+}} ymm3 = ymm6[0,1,2],mem[3],ymm6[4,5,6,7]6048; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm2[0,1,2],mem[3],ymm2[4,5,6,7]6049; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm13[0,1,2],mem[3],ymm13[4,5,6,7]6050; AVX-NEXT: vblendps {{.*#+}} ymm6 = ymm12[0,1,2],mem[3],ymm12[4,5,6,7]6051; AVX-NEXT: vblendps {{.*#+}} ymm7 = ymm11[0,1,2],mem[3],ymm11[4,5,6,7]6052; AVX-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2],mem[3],ymm5[4,5,6,7]6053; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],mem[3],ymm0[4,5,6,7]6054; AVX-NEXT: vmovaps %ymm0, 1184(%r9)6055; AVX-NEXT: vmovaps %ymm5, 1024(%r9)6056; AVX-NEXT: vmovaps %ymm7, 864(%r9)6057; AVX-NEXT: vmovaps %ymm6, 704(%r9)6058; AVX-NEXT: vmovaps %ymm4, 544(%r9)6059; AVX-NEXT: vmovaps %ymm2, 384(%r9)6060; AVX-NEXT: vmovaps %ymm3, 224(%r9)6061; AVX-NEXT: vmovaps %ymm1, 64(%r9)6062; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6063; AVX-NEXT: vmovaps %ymm0, 1248(%r9)6064; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6065; AVX-NEXT: vmovaps %ymm0, 1216(%r9)6066; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6067; AVX-NEXT: vmovaps %ymm0, 1088(%r9)6068; AVX-NEXT: vmovups (%rsp), %ymm0 # 32-byte Reload6069; AVX-NEXT: vmovaps %ymm0, 1056(%r9)6070; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6071; AVX-NEXT: vmovaps %ymm0, 928(%r9)6072; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6073; AVX-NEXT: vmovaps %ymm0, 896(%r9)6074; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6075; AVX-NEXT: vmovaps %ymm0, 768(%r9)6076; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6077; AVX-NEXT: vmovaps %ymm0, 736(%r9)6078; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6079; AVX-NEXT: vmovaps %ymm0, 608(%r9)6080; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6081; AVX-NEXT: vmovaps %ymm0, 576(%r9)6082; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6083; AVX-NEXT: vmovaps %ymm0, 448(%r9)6084; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6085; AVX-NEXT: vmovaps %ymm0, 416(%r9)6086; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6087; AVX-NEXT: vmovaps %ymm0, 288(%r9)6088; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6089; AVX-NEXT: vmovaps %ymm0, 256(%r9)6090; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6091; AVX-NEXT: vmovaps %ymm0, 128(%r9)6092; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6093; AVX-NEXT: vmovaps %ymm0, 96(%r9)6094; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6095; AVX-NEXT: vmovaps %ymm0, 1152(%r9)6096; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6097; AVX-NEXT: vmovaps %ymm0, 1120(%r9)6098; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6099; AVX-NEXT: vmovaps %ymm0, 992(%r9)6100; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6101; AVX-NEXT: vmovaps %ymm0, 960(%r9)6102; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6103; AVX-NEXT: vmovaps %ymm0, 832(%r9)6104; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6105; AVX-NEXT: vmovaps %ymm0, 800(%r9)6106; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6107; AVX-NEXT: vmovaps %ymm0, 672(%r9)6108; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6109; AVX-NEXT: vmovaps %ymm0, 640(%r9)6110; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6111; AVX-NEXT: vmovaps %ymm0, 512(%r9)6112; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6113; AVX-NEXT: vmovaps %ymm0, 480(%r9)6114; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6115; AVX-NEXT: vmovaps %ymm0, 352(%r9)6116; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6117; AVX-NEXT: vmovaps %ymm0, 320(%r9)6118; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6119; AVX-NEXT: vmovaps %ymm0, 192(%r9)6120; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6121; AVX-NEXT: vmovaps %ymm0, 160(%r9)6122; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6123; AVX-NEXT: vmovaps %ymm0, 32(%r9)6124; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6125; AVX-NEXT: vmovaps %ymm0, (%r9)6126; AVX-NEXT: addq $1512, %rsp # imm = 0x5E86127; AVX-NEXT: vzeroupper6128; AVX-NEXT: retq6129;6130; AVX2-LABEL: store_i32_stride5_vf64:6131; AVX2: # %bb.0:6132; AVX2-NEXT: subq $1736, %rsp # imm = 0x6C86133; AVX2-NEXT: vmovaps (%r8), %ymm156134; AVX2-NEXT: vmovups %ymm15, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6135; AVX2-NEXT: vmovaps 32(%r8), %ymm146136; AVX2-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6137; AVX2-NEXT: vmovaps 64(%r8), %ymm136138; AVX2-NEXT: vmovups %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6139; AVX2-NEXT: vmovaps (%rcx), %xmm76140; AVX2-NEXT: vmovaps 32(%rcx), %xmm36141; AVX2-NEXT: vmovaps 64(%rcx), %xmm06142; AVX2-NEXT: vmovaps (%rdx), %xmm86143; AVX2-NEXT: vmovaps 32(%rdx), %xmm46144; AVX2-NEXT: vmovaps 64(%rdx), %xmm16145; AVX2-NEXT: vunpcklps {{.*#+}} xmm2 = xmm8[0],xmm7[0],xmm8[1],xmm7[1]6146; AVX2-NEXT: vshufps {{.*#+}} xmm2 = xmm2[0,1,2,2]6147; AVX2-NEXT: vpermpd {{.*#+}} ymm5 = ymm2[0,0,2,1]6148; AVX2-NEXT: vmovaps (%rsi), %xmm96149; AVX2-NEXT: vmovaps 32(%rsi), %xmm66150; AVX2-NEXT: vmovaps 64(%rsi), %xmm26151; AVX2-NEXT: vmovaps (%rdi), %xmm106152; AVX2-NEXT: vunpcklps {{.*#+}} xmm11 = xmm10[0],xmm9[0],xmm10[1],xmm9[1]6153; AVX2-NEXT: vshufps {{.*#+}} xmm11 = xmm11[0,1,3,2]6154; AVX2-NEXT: vpermpd {{.*#+}} ymm12 = ymm11[0,1,1,1]6155; AVX2-NEXT: vblendps {{.*#+}} ymm5 = ymm12[0,1],ymm5[2,3],ymm12[4,5,6],ymm5[7]6156; AVX2-NEXT: vmovaps 32(%rdi), %xmm126157; AVX2-NEXT: vinsertf128 $1, (%r8), %ymm11, %ymm116158; AVX2-NEXT: vblendps {{.*#+}} ymm5 = ymm11[0],ymm5[1,2,3],ymm11[4],ymm5[5,6,7]6159; AVX2-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6160; AVX2-NEXT: vmovaps 64(%rdi), %xmm56161; AVX2-NEXT: vunpckhps {{.*#+}} xmm9 = xmm10[2],xmm9[2],xmm10[3],xmm9[3]6162; AVX2-NEXT: vshufps {{.*#+}} xmm9 = xmm9[0,1,2,2]6163; AVX2-NEXT: vpermpd {{.*#+}} ymm9 = ymm9[0,0,2,1]6164; AVX2-NEXT: vshufps {{.*#+}} xmm7 = xmm7[1,1,2,2]6165; AVX2-NEXT: vblendps {{.*#+}} xmm7 = xmm7[0,1],xmm8[2],xmm7[3]6166; AVX2-NEXT: vpermpd {{.*#+}} ymm7 = ymm7[0,1,1,3]6167; AVX2-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0,1],ymm9[2,3],ymm7[4,5],ymm9[6,7]6168; AVX2-NEXT: vpermpd {{.*#+}} ymm8 = ymm15[0,1,2,1]6169; AVX2-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0],ymm8[1],ymm7[2,3,4,5],ymm8[6],ymm7[7]6170; AVX2-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6171; AVX2-NEXT: vunpcklps {{.*#+}} xmm7 = xmm4[0],xmm3[0],xmm4[1],xmm3[1]6172; AVX2-NEXT: vshufps {{.*#+}} xmm7 = xmm7[0,1,2,2]6173; AVX2-NEXT: vpermpd {{.*#+}} ymm7 = ymm7[0,0,2,1]6174; AVX2-NEXT: vunpcklps {{.*#+}} xmm8 = xmm12[0],xmm6[0],xmm12[1],xmm6[1]6175; AVX2-NEXT: vshufps {{.*#+}} xmm8 = xmm8[0,1,3,2]6176; AVX2-NEXT: vpermpd {{.*#+}} ymm9 = ymm8[0,1,1,1]6177; AVX2-NEXT: vblendps {{.*#+}} ymm7 = ymm9[0,1],ymm7[2,3],ymm9[4,5,6],ymm7[7]6178; AVX2-NEXT: vinsertf128 $1, 32(%r8), %ymm8, %ymm86179; AVX2-NEXT: vblendps {{.*#+}} ymm7 = ymm8[0],ymm7[1,2,3],ymm8[4],ymm7[5,6,7]6180; AVX2-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6181; AVX2-NEXT: vunpckhps {{.*#+}} xmm6 = xmm12[2],xmm6[2],xmm12[3],xmm6[3]6182; AVX2-NEXT: vshufps {{.*#+}} xmm6 = xmm6[0,1,2,2]6183; AVX2-NEXT: vpermpd {{.*#+}} ymm6 = ymm6[0,0,2,1]6184; AVX2-NEXT: vshufps {{.*#+}} xmm3 = xmm3[1,1,2,2]6185; AVX2-NEXT: vblendps {{.*#+}} xmm3 = xmm3[0,1],xmm4[2],xmm3[3]6186; AVX2-NEXT: vpermpd {{.*#+}} ymm3 = ymm3[0,1,1,3]6187; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1],ymm6[2,3],ymm3[4,5],ymm6[6,7]6188; AVX2-NEXT: vpermpd {{.*#+}} ymm4 = ymm14[0,1,2,1]6189; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0],ymm4[1],ymm3[2,3,4,5],ymm4[6],ymm3[7]6190; AVX2-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6191; AVX2-NEXT: vunpcklps {{.*#+}} xmm3 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]6192; AVX2-NEXT: vshufps {{.*#+}} xmm3 = xmm3[0,1,2,2]6193; AVX2-NEXT: vpermpd {{.*#+}} ymm3 = ymm3[0,0,2,1]6194; AVX2-NEXT: vunpcklps {{.*#+}} xmm4 = xmm5[0],xmm2[0],xmm5[1],xmm2[1]6195; AVX2-NEXT: vshufps {{.*#+}} xmm4 = xmm4[0,1,3,2]6196; AVX2-NEXT: vpermpd {{.*#+}} ymm6 = ymm4[0,1,1,1]6197; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm6[0,1],ymm3[2,3],ymm6[4,5,6],ymm3[7]6198; AVX2-NEXT: vinsertf128 $1, 64(%r8), %ymm4, %ymm46199; AVX2-NEXT: vblendps {{.*#+}} ymm3 = ymm4[0],ymm3[1,2,3],ymm4[4],ymm3[5,6,7]6200; AVX2-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6201; AVX2-NEXT: vunpckhps {{.*#+}} xmm2 = xmm5[2],xmm2[2],xmm5[3],xmm2[3]6202; AVX2-NEXT: vshufps {{.*#+}} xmm2 = xmm2[0,1,2,2]6203; AVX2-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,0,2,1]6204; AVX2-NEXT: vshufps {{.*#+}} xmm0 = xmm0[1,1,2,2]6205; AVX2-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2],xmm0[3]6206; AVX2-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,1,1,3]6207; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3],ymm0[4,5],ymm2[6,7]6208; AVX2-NEXT: vpermpd {{.*#+}} ymm1 = ymm13[0,1,2,1]6209; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2,3,4,5],ymm1[6],ymm0[7]6210; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6211; AVX2-NEXT: vmovaps 96(%rcx), %xmm06212; AVX2-NEXT: vmovaps 96(%rdx), %xmm16213; AVX2-NEXT: vunpcklps {{.*#+}} xmm2 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]6214; AVX2-NEXT: vshufps {{.*#+}} xmm2 = xmm2[0,1,2,2]6215; AVX2-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,0,2,1]6216; AVX2-NEXT: vmovaps 96(%rsi), %xmm36217; AVX2-NEXT: vmovaps 96(%rdi), %xmm46218; AVX2-NEXT: vunpcklps {{.*#+}} xmm5 = xmm4[0],xmm3[0],xmm4[1],xmm3[1]6219; AVX2-NEXT: vshufps {{.*#+}} xmm5 = xmm5[0,1,3,2]6220; AVX2-NEXT: vpermpd {{.*#+}} ymm6 = ymm5[0,1,1,1]6221; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm6[0,1],ymm2[2,3],ymm6[4,5,6],ymm2[7]6222; AVX2-NEXT: vmovaps 96(%r8), %ymm66223; AVX2-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6224; AVX2-NEXT: vinsertf128 $1, 96(%r8), %ymm5, %ymm56225; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm5[0],ymm2[1,2,3],ymm5[4],ymm2[5,6,7]6226; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6227; AVX2-NEXT: vunpckhps {{.*#+}} xmm2 = xmm4[2],xmm3[2],xmm4[3],xmm3[3]6228; AVX2-NEXT: vshufps {{.*#+}} xmm2 = xmm2[0,1,2,2]6229; AVX2-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,0,2,1]6230; AVX2-NEXT: vshufps {{.*#+}} xmm0 = xmm0[1,1,2,2]6231; AVX2-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2],xmm0[3]6232; AVX2-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,1,1,3]6233; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3],ymm0[4,5],ymm2[6,7]6234; AVX2-NEXT: vpermpd {{.*#+}} ymm1 = ymm6[0,1,2,1]6235; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2,3,4,5],ymm1[6],ymm0[7]6236; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6237; AVX2-NEXT: vmovaps 128(%rcx), %xmm06238; AVX2-NEXT: vmovaps 128(%rdx), %xmm16239; AVX2-NEXT: vunpcklps {{.*#+}} xmm2 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]6240; AVX2-NEXT: vshufps {{.*#+}} xmm2 = xmm2[0,1,2,2]6241; AVX2-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,0,2,1]6242; AVX2-NEXT: vmovaps 128(%rsi), %xmm36243; AVX2-NEXT: vmovaps 128(%rdi), %xmm46244; AVX2-NEXT: vunpcklps {{.*#+}} xmm5 = xmm4[0],xmm3[0],xmm4[1],xmm3[1]6245; AVX2-NEXT: vshufps {{.*#+}} xmm5 = xmm5[0,1,3,2]6246; AVX2-NEXT: vpermpd {{.*#+}} ymm6 = ymm5[0,1,1,1]6247; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm6[0,1],ymm2[2,3],ymm6[4,5,6],ymm2[7]6248; AVX2-NEXT: vinsertf128 $1, 128(%r8), %ymm5, %ymm56249; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm5[0],ymm2[1,2,3],ymm5[4],ymm2[5,6,7]6250; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6251; AVX2-NEXT: vunpckhps {{.*#+}} xmm2 = xmm4[2],xmm3[2],xmm4[3],xmm3[3]6252; AVX2-NEXT: vshufps {{.*#+}} xmm0 = xmm0[1,1,2,2]6253; AVX2-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2],xmm0[3]6254; AVX2-NEXT: vmovaps 128(%r8), %ymm36255; AVX2-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6256; AVX2-NEXT: vshufps {{.*#+}} xmm1 = xmm2[0,1,2,2]6257; AVX2-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,0,2,1]6258; AVX2-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,1,1,3]6259; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7]6260; AVX2-NEXT: vpermpd {{.*#+}} ymm1 = ymm3[0,1,2,1]6261; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2,3,4,5],ymm1[6],ymm0[7]6262; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6263; AVX2-NEXT: vmovaps 160(%rcx), %xmm06264; AVX2-NEXT: vmovaps 160(%rdx), %xmm16265; AVX2-NEXT: vunpcklps {{.*#+}} xmm2 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]6266; AVX2-NEXT: vshufps {{.*#+}} xmm2 = xmm2[0,1,2,2]6267; AVX2-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,0,2,1]6268; AVX2-NEXT: vmovaps 160(%rsi), %xmm36269; AVX2-NEXT: vmovaps 160(%rdi), %xmm46270; AVX2-NEXT: vunpcklps {{.*#+}} xmm5 = xmm4[0],xmm3[0],xmm4[1],xmm3[1]6271; AVX2-NEXT: vshufps {{.*#+}} xmm5 = xmm5[0,1,3,2]6272; AVX2-NEXT: vpermpd {{.*#+}} ymm6 = ymm5[0,1,1,1]6273; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm6[0,1],ymm2[2,3],ymm6[4,5,6],ymm2[7]6274; AVX2-NEXT: vinsertf128 $1, 160(%r8), %ymm5, %ymm56275; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm5[0],ymm2[1,2,3],ymm5[4],ymm2[5,6,7]6276; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6277; AVX2-NEXT: vunpckhps {{.*#+}} xmm2 = xmm4[2],xmm3[2],xmm4[3],xmm3[3]6278; AVX2-NEXT: vshufps {{.*#+}} xmm0 = xmm0[1,1,2,2]6279; AVX2-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2],xmm0[3]6280; AVX2-NEXT: vshufps {{.*#+}} xmm1 = xmm2[0,1,2,2]6281; AVX2-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,0,2,1]6282; AVX2-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,1,1,3]6283; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7]6284; AVX2-NEXT: vmovaps 160(%r8), %ymm16285; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6286; AVX2-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]6287; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2,3,4,5],ymm1[6],ymm0[7]6288; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6289; AVX2-NEXT: vmovaps 192(%rcx), %xmm06290; AVX2-NEXT: vmovaps 192(%rdx), %xmm16291; AVX2-NEXT: vunpcklps {{.*#+}} xmm2 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]6292; AVX2-NEXT: vshufps {{.*#+}} xmm2 = xmm2[0,1,2,2]6293; AVX2-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,0,2,1]6294; AVX2-NEXT: vmovaps 192(%rsi), %xmm36295; AVX2-NEXT: vmovaps 192(%rdi), %xmm46296; AVX2-NEXT: vunpcklps {{.*#+}} xmm5 = xmm4[0],xmm3[0],xmm4[1],xmm3[1]6297; AVX2-NEXT: vshufps {{.*#+}} xmm5 = xmm5[0,1,3,2]6298; AVX2-NEXT: vpermpd {{.*#+}} ymm6 = ymm5[0,1,1,1]6299; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm6[0,1],ymm2[2,3],ymm6[4,5,6],ymm2[7]6300; AVX2-NEXT: vinsertf128 $1, 192(%r8), %ymm5, %ymm56301; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm5[0],ymm2[1,2,3],ymm5[4],ymm2[5,6,7]6302; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6303; AVX2-NEXT: vunpckhps {{.*#+}} xmm2 = xmm4[2],xmm3[2],xmm4[3],xmm3[3]6304; AVX2-NEXT: vshufps {{.*#+}} xmm0 = xmm0[1,1,2,2]6305; AVX2-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2],xmm0[3]6306; AVX2-NEXT: vshufps {{.*#+}} xmm1 = xmm2[0,1,2,2]6307; AVX2-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,0,2,1]6308; AVX2-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,1,1,3]6309; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7]6310; AVX2-NEXT: vmovaps 192(%r8), %ymm16311; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6312; AVX2-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]6313; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2,3,4,5],ymm1[6],ymm0[7]6314; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6315; AVX2-NEXT: vmovaps 224(%rcx), %xmm06316; AVX2-NEXT: vmovaps 224(%rdx), %xmm16317; AVX2-NEXT: vunpcklps {{.*#+}} xmm2 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]6318; AVX2-NEXT: vshufps {{.*#+}} xmm2 = xmm2[0,1,2,2]6319; AVX2-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,0,2,1]6320; AVX2-NEXT: vmovaps 224(%rsi), %xmm36321; AVX2-NEXT: vmovaps 224(%rdi), %xmm46322; AVX2-NEXT: vunpcklps {{.*#+}} xmm5 = xmm4[0],xmm3[0],xmm4[1],xmm3[1]6323; AVX2-NEXT: vshufps {{.*#+}} xmm5 = xmm5[0,1,3,2]6324; AVX2-NEXT: vpermpd {{.*#+}} ymm6 = ymm5[0,1,1,1]6325; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm6[0,1],ymm2[2,3],ymm6[4,5,6],ymm2[7]6326; AVX2-NEXT: vinsertf128 $1, 224(%r8), %ymm5, %ymm56327; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm5[0],ymm2[1,2,3],ymm5[4],ymm2[5,6,7]6328; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6329; AVX2-NEXT: vunpckhps {{.*#+}} xmm2 = xmm4[2],xmm3[2],xmm4[3],xmm3[3]6330; AVX2-NEXT: vshufps {{.*#+}} xmm0 = xmm0[1,1,2,2]6331; AVX2-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2],xmm0[3]6332; AVX2-NEXT: vshufps {{.*#+}} xmm1 = xmm2[0,1,2,2]6333; AVX2-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,0,2,1]6334; AVX2-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,1,1,3]6335; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7]6336; AVX2-NEXT: vmovaps 224(%r8), %ymm16337; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6338; AVX2-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]6339; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2,3,4,5],ymm1[6],ymm0[7]6340; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6341; AVX2-NEXT: vmovaps (%rdx), %ymm156342; AVX2-NEXT: vmovaps (%rcx), %ymm26343; AVX2-NEXT: vshufps {{.*#+}} ymm0 = ymm2[1,1,1,1,5,5,5,5]6344; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6345; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm15[1],ymm0[2,3,4],ymm15[5],ymm0[6,7]6346; AVX2-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2]6347; AVX2-NEXT: vmovaps (%rdi), %ymm36348; AVX2-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6349; AVX2-NEXT: vmovaps (%rsi), %ymm136350; AVX2-NEXT: vshufps {{.*#+}} ymm1 = ymm13[1,1,2,2,5,5,6,6]6351; AVX2-NEXT: vmovups %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6352; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm3[1,2],ymm1[3,4],ymm3[5,6],ymm1[7]6353; AVX2-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,2,2,3]6354; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4],ymm1[5,6,7]6355; AVX2-NEXT: vbroadcastsd 16(%r8), %ymm16356; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2,3,4],ymm1[5],ymm0[6,7]6357; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6358; AVX2-NEXT: vshufps {{.*#+}} ymm0 = ymm2[0,2,3,3,4,6,7,7]6359; AVX2-NEXT: vshufps {{.*#+}} ymm1 = ymm15[2,3,2,3,6,7,6,7]6360; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2],ymm1[3,4],ymm0[5,6],ymm1[7]6361; AVX2-NEXT: vunpckhps {{.*#+}} ymm1 = ymm13[2],ymm3[2],ymm13[3],ymm3[3],ymm13[6],ymm3[6],ymm13[7],ymm3[7]6362; AVX2-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[3,3,3,3]6363; AVX2-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,1,3,3]6364; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3,4],ymm0[5,6,7]6365; AVX2-NEXT: vbroadcastsd 24(%r8), %ymm16366; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4,5,6],ymm1[7]6367; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6368; AVX2-NEXT: vmovaps 32(%rdx), %ymm36369; AVX2-NEXT: vmovaps 32(%rcx), %ymm46370; AVX2-NEXT: vshufps {{.*#+}} ymm0 = ymm4[1,1,1,1,5,5,5,5]6371; AVX2-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6372; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm3[1],ymm0[2,3,4],ymm3[5],ymm0[6,7]6373; AVX2-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6374; AVX2-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2]6375; AVX2-NEXT: vmovaps 32(%rdi), %ymm26376; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6377; AVX2-NEXT: vmovaps 32(%rsi), %ymm136378; AVX2-NEXT: vshufps {{.*#+}} ymm1 = ymm13[1,1,2,2,5,5,6,6]6379; AVX2-NEXT: vmovups %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6380; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1,2],ymm1[3,4],ymm2[5,6],ymm1[7]6381; AVX2-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,2,2,3]6382; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4],ymm1[5,6,7]6383; AVX2-NEXT: vbroadcastsd 48(%r8), %ymm16384; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2,3,4],ymm1[5],ymm0[6,7]6385; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6386; AVX2-NEXT: vshufps {{.*#+}} ymm0 = ymm4[0,2,3,3,4,6,7,7]6387; AVX2-NEXT: vshufps {{.*#+}} ymm1 = ymm3[2,3,2,3,6,7,6,7]6388; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2],ymm1[3,4],ymm0[5,6],ymm1[7]6389; AVX2-NEXT: vunpckhps {{.*#+}} ymm1 = ymm13[2],ymm2[2],ymm13[3],ymm2[3],ymm13[6],ymm2[6],ymm13[7],ymm2[7]6390; AVX2-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[3,3,3,3]6391; AVX2-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,1,3,3]6392; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3,4],ymm0[5,6,7]6393; AVX2-NEXT: vbroadcastsd 56(%r8), %ymm16394; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4,5,6],ymm1[7]6395; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6396; AVX2-NEXT: vmovaps 64(%rdx), %ymm36397; AVX2-NEXT: vmovaps 64(%rcx), %ymm46398; AVX2-NEXT: vshufps {{.*#+}} ymm0 = ymm4[1,1,1,1,5,5,5,5]6399; AVX2-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6400; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm3[1],ymm0[2,3,4],ymm3[5],ymm0[6,7]6401; AVX2-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6402; AVX2-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2]6403; AVX2-NEXT: vmovaps 64(%rdi), %ymm26404; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6405; AVX2-NEXT: vmovaps 64(%rsi), %ymm136406; AVX2-NEXT: vshufps {{.*#+}} ymm1 = ymm13[1,1,2,2,5,5,6,6]6407; AVX2-NEXT: vmovups %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6408; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1,2],ymm1[3,4],ymm2[5,6],ymm1[7]6409; AVX2-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,2,2,3]6410; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4],ymm1[5,6,7]6411; AVX2-NEXT: vbroadcastsd 80(%r8), %ymm16412; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2,3,4],ymm1[5],ymm0[6,7]6413; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6414; AVX2-NEXT: vshufps {{.*#+}} ymm0 = ymm4[0,2,3,3,4,6,7,7]6415; AVX2-NEXT: vshufps {{.*#+}} ymm1 = ymm3[2,3,2,3,6,7,6,7]6416; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2],ymm1[3,4],ymm0[5,6],ymm1[7]6417; AVX2-NEXT: vunpckhps {{.*#+}} ymm1 = ymm13[2],ymm2[2],ymm13[3],ymm2[3],ymm13[6],ymm2[6],ymm13[7],ymm2[7]6418; AVX2-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[3,3,3,3]6419; AVX2-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,1,3,3]6420; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3,4],ymm0[5,6,7]6421; AVX2-NEXT: vbroadcastsd 88(%r8), %ymm16422; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4,5,6],ymm1[7]6423; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6424; AVX2-NEXT: vmovaps 96(%rdx), %ymm36425; AVX2-NEXT: vmovaps 96(%rcx), %ymm46426; AVX2-NEXT: vshufps {{.*#+}} ymm0 = ymm4[1,1,1,1,5,5,5,5]6427; AVX2-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6428; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm3[1],ymm0[2,3,4],ymm3[5],ymm0[6,7]6429; AVX2-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6430; AVX2-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2]6431; AVX2-NEXT: vmovaps 96(%rdi), %ymm26432; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6433; AVX2-NEXT: vmovaps 96(%rsi), %ymm146434; AVX2-NEXT: vshufps {{.*#+}} ymm1 = ymm14[1,1,2,2,5,5,6,6]6435; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1,2],ymm1[3,4],ymm2[5,6],ymm1[7]6436; AVX2-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,2,2,3]6437; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4],ymm1[5,6,7]6438; AVX2-NEXT: vbroadcastsd 112(%r8), %ymm16439; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2,3,4],ymm1[5],ymm0[6,7]6440; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6441; AVX2-NEXT: vshufps {{.*#+}} ymm0 = ymm4[0,2,3,3,4,6,7,7]6442; AVX2-NEXT: vshufps {{.*#+}} ymm1 = ymm3[2,3,2,3,6,7,6,7]6443; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2],ymm1[3,4],ymm0[5,6],ymm1[7]6444; AVX2-NEXT: vunpckhps {{.*#+}} ymm1 = ymm14[2],ymm2[2],ymm14[3],ymm2[3],ymm14[6],ymm2[6],ymm14[7],ymm2[7]6445; AVX2-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[3,3,3,3]6446; AVX2-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,1,3,3]6447; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3,4],ymm0[5,6,7]6448; AVX2-NEXT: vbroadcastsd 120(%r8), %ymm16449; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4,5,6],ymm1[7]6450; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6451; AVX2-NEXT: vmovaps 128(%rdx), %ymm46452; AVX2-NEXT: vmovaps 128(%rcx), %ymm26453; AVX2-NEXT: vshufps {{.*#+}} ymm0 = ymm2[1,1,1,1,5,5,5,5]6454; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6455; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm4[1],ymm0[2,3,4],ymm4[5],ymm0[6,7]6456; AVX2-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6457; AVX2-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2]6458; AVX2-NEXT: vmovaps 128(%rdi), %ymm36459; AVX2-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6460; AVX2-NEXT: vmovaps 128(%rsi), %ymm136461; AVX2-NEXT: vshufps {{.*#+}} ymm1 = ymm13[1,1,2,2,5,5,6,6]6462; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm3[1,2],ymm1[3,4],ymm3[5,6],ymm1[7]6463; AVX2-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,2,2,3]6464; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4],ymm1[5,6,7]6465; AVX2-NEXT: vbroadcastsd 144(%r8), %ymm16466; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2,3,4],ymm1[5],ymm0[6,7]6467; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6468; AVX2-NEXT: vshufps {{.*#+}} ymm0 = ymm2[0,2,3,3,4,6,7,7]6469; AVX2-NEXT: vshufps {{.*#+}} ymm1 = ymm4[2,3,2,3,6,7,6,7]6470; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2],ymm1[3,4],ymm0[5,6],ymm1[7]6471; AVX2-NEXT: vunpckhps {{.*#+}} ymm1 = ymm13[2],ymm3[2],ymm13[3],ymm3[3],ymm13[6],ymm3[6],ymm13[7],ymm3[7]6472; AVX2-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[3,3,3,3]6473; AVX2-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,1,3,3]6474; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3,4],ymm0[5,6,7]6475; AVX2-NEXT: vbroadcastsd 152(%r8), %ymm16476; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4,5,6],ymm1[7]6477; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6478; AVX2-NEXT: vmovaps 160(%rdx), %ymm26479; AVX2-NEXT: vmovaps 160(%rcx), %ymm126480; AVX2-NEXT: vshufps {{.*#+}} ymm0 = ymm12[1,1,1,1,5,5,5,5]6481; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm2[1],ymm0[2,3,4],ymm2[5],ymm0[6,7]6482; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6483; AVX2-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2]6484; AVX2-NEXT: vmovaps 160(%rdi), %ymm106485; AVX2-NEXT: vmovaps 160(%rsi), %ymm96486; AVX2-NEXT: vshufps {{.*#+}} ymm1 = ymm9[1,1,2,2,5,5,6,6]6487; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm10[1,2],ymm1[3,4],ymm10[5,6],ymm1[7]6488; AVX2-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,2,2,3]6489; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4],ymm1[5,6,7]6490; AVX2-NEXT: vbroadcastsd 176(%r8), %ymm16491; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2,3,4],ymm1[5],ymm0[6,7]6492; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6493; AVX2-NEXT: vshufps {{.*#+}} ymm0 = ymm12[0,2,3,3,4,6,7,7]6494; AVX2-NEXT: vshufps {{.*#+}} ymm1 = ymm2[2,3,2,3,6,7,6,7]6495; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2],ymm1[3,4],ymm0[5,6],ymm1[7]6496; AVX2-NEXT: vunpckhps {{.*#+}} ymm1 = ymm9[2],ymm10[2],ymm9[3],ymm10[3],ymm9[6],ymm10[6],ymm9[7],ymm10[7]6497; AVX2-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[3,3,3,3]6498; AVX2-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,1,3,3]6499; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3,4],ymm0[5,6,7]6500; AVX2-NEXT: vbroadcastsd 184(%r8), %ymm16501; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4,5,6],ymm1[7]6502; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6503; AVX2-NEXT: vmovaps 192(%rdx), %ymm86504; AVX2-NEXT: vmovaps 192(%rcx), %ymm66505; AVX2-NEXT: vshufps {{.*#+}} ymm0 = ymm6[1,1,1,1,5,5,5,5]6506; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm8[1],ymm0[2,3,4],ymm8[5],ymm0[6,7]6507; AVX2-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2]6508; AVX2-NEXT: vmovaps 192(%rdi), %ymm56509; AVX2-NEXT: vmovaps 192(%rsi), %ymm46510; AVX2-NEXT: vshufps {{.*#+}} ymm1 = ymm4[1,1,2,2,5,5,6,6]6511; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm5[1,2],ymm1[3,4],ymm5[5,6],ymm1[7]6512; AVX2-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,2,2,3]6513; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4],ymm1[5,6,7]6514; AVX2-NEXT: vbroadcastsd 208(%r8), %ymm16515; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2,3,4],ymm1[5],ymm0[6,7]6516; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6517; AVX2-NEXT: vshufps {{.*#+}} ymm0 = ymm6[0,2,3,3,4,6,7,7]6518; AVX2-NEXT: vshufps {{.*#+}} ymm1 = ymm8[2,3,2,3,6,7,6,7]6519; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2],ymm1[3,4],ymm0[5,6],ymm1[7]6520; AVX2-NEXT: vunpckhps {{.*#+}} ymm1 = ymm4[2],ymm5[2],ymm4[3],ymm5[3],ymm4[6],ymm5[6],ymm4[7],ymm5[7]6521; AVX2-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[3,3,3,3]6522; AVX2-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,1,3,3]6523; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3,4],ymm0[5,6,7]6524; AVX2-NEXT: vbroadcastsd 216(%r8), %ymm16525; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4,5,6],ymm1[7]6526; AVX2-NEXT: vmovups %ymm0, (%rsp) # 32-byte Spill6527; AVX2-NEXT: vmovaps 224(%rdx), %ymm36528; AVX2-NEXT: vmovaps 224(%rcx), %ymm26529; AVX2-NEXT: vshufps {{.*#+}} ymm0 = ymm2[1,1,1,1,5,5,5,5]6530; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm3[1],ymm0[2,3,4],ymm3[5],ymm0[6,7]6531; AVX2-NEXT: vpermpd {{.*#+}} ymm7 = ymm0[2,2,2,2]6532; AVX2-NEXT: vmovaps 224(%rdi), %ymm16533; AVX2-NEXT: vmovaps 224(%rsi), %ymm06534; AVX2-NEXT: vshufps {{.*#+}} ymm11 = ymm0[1,1,2,2,5,5,6,6]6535; AVX2-NEXT: vblendps {{.*#+}} ymm11 = ymm11[0],ymm1[1,2],ymm11[3,4],ymm1[5,6],ymm11[7]6536; AVX2-NEXT: vpermpd {{.*#+}} ymm11 = ymm11[2,2,2,3]6537; AVX2-NEXT: vblendps {{.*#+}} ymm7 = ymm11[0,1,2],ymm7[3,4],ymm11[5,6,7]6538; AVX2-NEXT: vbroadcastsd 240(%r8), %ymm116539; AVX2-NEXT: vblendps {{.*#+}} ymm7 = ymm11[0],ymm7[1,2,3,4],ymm11[5],ymm7[6,7]6540; AVX2-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6541; AVX2-NEXT: vshufps {{.*#+}} ymm7 = ymm2[0,2,3,3,4,6,7,7]6542; AVX2-NEXT: vshufps {{.*#+}} ymm11 = ymm3[2,3,2,3,6,7,6,7]6543; AVX2-NEXT: vblendps {{.*#+}} ymm7 = ymm11[0],ymm7[1,2],ymm11[3,4],ymm7[5,6],ymm11[7]6544; AVX2-NEXT: vunpckhps {{.*#+}} ymm11 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7]6545; AVX2-NEXT: vpermpd {{.*#+}} ymm11 = ymm11[3,3,3,3]6546; AVX2-NEXT: vpermpd {{.*#+}} ymm7 = ymm7[2,1,3,3]6547; AVX2-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0,1,2],ymm11[3,4],ymm7[5,6,7]6548; AVX2-NEXT: vbroadcastsd 248(%r8), %ymm116549; AVX2-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0,1],ymm11[2],ymm7[3,4,5,6],ymm11[7]6550; AVX2-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6551; AVX2-NEXT: vpermilps $227, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Folded Reload6552; AVX2-NEXT: # ymm7 = mem[3,0,2,3,7,4,6,7]6553; AVX2-NEXT: vblendps $16, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm7 # 32-byte Folded Reload6554; AVX2-NEXT: # ymm7 = ymm7[0,1,2,3],mem[4],ymm7[5,6,7]6555; AVX2-NEXT: vpermilps $52, {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Folded Reload6556; AVX2-NEXT: # ymm11 = mem[0,1,3,0,4,5,7,4]6557; AVX2-NEXT: vshufps {{.*#+}} ymm15 = ymm15[2,3,0,1,6,7,4,5]6558; AVX2-NEXT: vblendps {{.*#+}} ymm11 = ymm15[0,1],ymm11[2],ymm15[3,4,5,6],ymm11[7]6559; AVX2-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0],ymm11[1,2,3],ymm7[4,5],ymm11[6,7]6560; AVX2-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm7 # 32-byte Folded Reload6561; AVX2-NEXT: # ymm7 = ymm7[0,1,2],mem[3],ymm7[4,5,6,7]6562; AVX2-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6563; AVX2-NEXT: vpermilps $227, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Folded Reload6564; AVX2-NEXT: # ymm7 = mem[3,0,2,3,7,4,6,7]6565; AVX2-NEXT: vblendps $16, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm7 # 32-byte Folded Reload6566; AVX2-NEXT: # ymm7 = ymm7[0,1,2,3],mem[4],ymm7[5,6,7]6567; AVX2-NEXT: vpermilps $52, {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Folded Reload6568; AVX2-NEXT: # ymm11 = mem[0,1,3,0,4,5,7,4]6569; AVX2-NEXT: vpermilps $78, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Folded Reload6570; AVX2-NEXT: # ymm15 = mem[2,3,0,1,6,7,4,5]6571; AVX2-NEXT: vblendps {{.*#+}} ymm11 = ymm15[0,1],ymm11[2],ymm15[3,4,5,6],ymm11[7]6572; AVX2-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0],ymm11[1,2,3],ymm7[4,5],ymm11[6,7]6573; AVX2-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm7 # 32-byte Folded Reload6574; AVX2-NEXT: # ymm7 = ymm7[0,1,2],mem[3],ymm7[4,5,6,7]6575; AVX2-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6576; AVX2-NEXT: vpermilps $227, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Folded Reload6577; AVX2-NEXT: # ymm7 = mem[3,0,2,3,7,4,6,7]6578; AVX2-NEXT: vblendps $16, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm7 # 32-byte Folded Reload6579; AVX2-NEXT: # ymm7 = ymm7[0,1,2,3],mem[4],ymm7[5,6,7]6580; AVX2-NEXT: vpermilps $52, {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Folded Reload6581; AVX2-NEXT: # ymm11 = mem[0,1,3,0,4,5,7,4]6582; AVX2-NEXT: vpermilps $78, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Folded Reload6583; AVX2-NEXT: # ymm15 = mem[2,3,0,1,6,7,4,5]6584; AVX2-NEXT: vblendps {{.*#+}} ymm11 = ymm15[0,1],ymm11[2],ymm15[3,4,5,6],ymm11[7]6585; AVX2-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0],ymm11[1,2,3],ymm7[4,5],ymm11[6,7]6586; AVX2-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm11 # 32-byte Folded Reload6587; AVX2-NEXT: # ymm11 = ymm7[0,1,2],mem[3],ymm7[4,5,6,7]6588; AVX2-NEXT: vshufps {{.*#+}} ymm7 = ymm14[3,0,2,3,7,4,6,7]6589; AVX2-NEXT: vblendps $16, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm7 # 32-byte Folded Reload6590; AVX2-NEXT: # ymm7 = ymm7[0,1,2,3],mem[4],ymm7[5,6,7]6591; AVX2-NEXT: vpermilps $52, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Folded Reload6592; AVX2-NEXT: # ymm15 = mem[0,1,3,0,4,5,7,4]6593; AVX2-NEXT: vpermilps $78, {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Folded Reload6594; AVX2-NEXT: # ymm14 = mem[2,3,0,1,6,7,4,5]6595; AVX2-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1],ymm15[2],ymm14[3,4,5,6],ymm15[7]6596; AVX2-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0],ymm14[1,2,3],ymm7[4,5],ymm14[6,7]6597; AVX2-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm7 # 32-byte Folded Reload6598; AVX2-NEXT: # ymm7 = ymm7[0,1,2],mem[3],ymm7[4,5,6,7]6599; AVX2-NEXT: vshufps {{.*#+}} ymm14 = ymm13[3,0,2,3,7,4,6,7]6600; AVX2-NEXT: vblendps $16, {{[-0-9]+}}(%r{{[sb]}}p), %ymm14, %ymm14 # 32-byte Folded Reload6601; AVX2-NEXT: # ymm14 = ymm14[0,1,2,3],mem[4],ymm14[5,6,7]6602; AVX2-NEXT: vpermilps $52, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Folded Reload6603; AVX2-NEXT: # ymm15 = mem[0,1,3,0,4,5,7,4]6604; AVX2-NEXT: vpermilps $78, {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Folded Reload6605; AVX2-NEXT: # ymm13 = mem[2,3,0,1,6,7,4,5]6606; AVX2-NEXT: vblendps {{.*#+}} ymm13 = ymm13[0,1],ymm15[2],ymm13[3,4,5,6],ymm15[7]6607; AVX2-NEXT: vblendps {{.*#+}} ymm13 = ymm14[0],ymm13[1,2,3],ymm14[4,5],ymm13[6,7]6608; AVX2-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm13 # 32-byte Folded Reload6609; AVX2-NEXT: # ymm13 = ymm13[0,1,2],mem[3],ymm13[4,5,6,7]6610; AVX2-NEXT: vshufps {{.*#+}} ymm9 = ymm9[3,0,2,3,7,4,6,7]6611; AVX2-NEXT: vblendps {{.*#+}} ymm9 = ymm9[0,1,2,3],ymm10[4],ymm9[5,6,7]6612; AVX2-NEXT: vshufps {{.*#+}} ymm10 = ymm12[0,1,3,0,4,5,7,4]6613; AVX2-NEXT: vpermilps $78, {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Folded Reload6614; AVX2-NEXT: # ymm12 = mem[2,3,0,1,6,7,4,5]6615; AVX2-NEXT: vblendps {{.*#+}} ymm10 = ymm12[0,1],ymm10[2],ymm12[3,4,5,6],ymm10[7]6616; AVX2-NEXT: vblendps {{.*#+}} ymm9 = ymm9[0],ymm10[1,2,3],ymm9[4,5],ymm10[6,7]6617; AVX2-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm9, %ymm9 # 32-byte Folded Reload6618; AVX2-NEXT: # ymm9 = ymm9[0,1,2],mem[3],ymm9[4,5,6,7]6619; AVX2-NEXT: vshufps {{.*#+}} ymm4 = ymm4[3,0,2,3,7,4,6,7]6620; AVX2-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm5[4],ymm4[5,6,7]6621; AVX2-NEXT: vshufps {{.*#+}} ymm5 = ymm6[0,1,3,0,4,5,7,4]6622; AVX2-NEXT: vshufps {{.*#+}} ymm6 = ymm8[2,3,0,1,6,7,4,5]6623; AVX2-NEXT: vblendps {{.*#+}} ymm5 = ymm6[0,1],ymm5[2],ymm6[3,4,5,6],ymm5[7]6624; AVX2-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0],ymm5[1,2,3],ymm4[4,5],ymm5[6,7]6625; AVX2-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4, %ymm4 # 32-byte Folded Reload6626; AVX2-NEXT: # ymm4 = ymm4[0,1,2],mem[3],ymm4[4,5,6,7]6627; AVX2-NEXT: vshufps {{.*#+}} ymm0 = ymm0[3,0,2,3,7,4,6,7]6628; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4],ymm0[5,6,7]6629; AVX2-NEXT: vshufps {{.*#+}} ymm1 = ymm2[0,1,3,0,4,5,7,4]6630; AVX2-NEXT: vshufps {{.*#+}} ymm2 = ymm3[2,3,0,1,6,7,4,5]6631; AVX2-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1],ymm1[2],ymm2[3,4,5,6],ymm1[7]6632; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm1[1,2,3],ymm0[4,5],ymm1[6,7]6633; AVX2-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload6634; AVX2-NEXT: # ymm0 = ymm0[0,1,2],mem[3],ymm0[4,5,6,7]6635; AVX2-NEXT: vmovaps %ymm0, 1184(%r9)6636; AVX2-NEXT: vmovaps %ymm4, 1024(%r9)6637; AVX2-NEXT: vmovaps %ymm9, 864(%r9)6638; AVX2-NEXT: vmovaps %ymm13, 704(%r9)6639; AVX2-NEXT: vmovaps %ymm7, 544(%r9)6640; AVX2-NEXT: vmovaps %ymm11, 384(%r9)6641; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6642; AVX2-NEXT: vmovaps %ymm0, 224(%r9)6643; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6644; AVX2-NEXT: vmovaps %ymm0, 64(%r9)6645; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6646; AVX2-NEXT: vmovaps %ymm0, 1248(%r9)6647; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6648; AVX2-NEXT: vmovaps %ymm0, 1216(%r9)6649; AVX2-NEXT: vmovups (%rsp), %ymm0 # 32-byte Reload6650; AVX2-NEXT: vmovaps %ymm0, 1088(%r9)6651; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6652; AVX2-NEXT: vmovaps %ymm0, 1056(%r9)6653; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6654; AVX2-NEXT: vmovaps %ymm0, 928(%r9)6655; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6656; AVX2-NEXT: vmovaps %ymm0, 896(%r9)6657; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6658; AVX2-NEXT: vmovaps %ymm0, 768(%r9)6659; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6660; AVX2-NEXT: vmovaps %ymm0, 736(%r9)6661; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6662; AVX2-NEXT: vmovaps %ymm0, 608(%r9)6663; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6664; AVX2-NEXT: vmovaps %ymm0, 576(%r9)6665; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6666; AVX2-NEXT: vmovaps %ymm0, 448(%r9)6667; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6668; AVX2-NEXT: vmovaps %ymm0, 416(%r9)6669; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6670; AVX2-NEXT: vmovaps %ymm0, 288(%r9)6671; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6672; AVX2-NEXT: vmovaps %ymm0, 256(%r9)6673; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6674; AVX2-NEXT: vmovaps %ymm0, 128(%r9)6675; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6676; AVX2-NEXT: vmovaps %ymm0, 96(%r9)6677; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6678; AVX2-NEXT: vmovaps %ymm0, 1152(%r9)6679; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6680; AVX2-NEXT: vmovaps %ymm0, 1120(%r9)6681; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6682; AVX2-NEXT: vmovaps %ymm0, 992(%r9)6683; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6684; AVX2-NEXT: vmovaps %ymm0, 960(%r9)6685; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6686; AVX2-NEXT: vmovaps %ymm0, 832(%r9)6687; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6688; AVX2-NEXT: vmovaps %ymm0, 800(%r9)6689; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6690; AVX2-NEXT: vmovaps %ymm0, 672(%r9)6691; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6692; AVX2-NEXT: vmovaps %ymm0, 640(%r9)6693; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6694; AVX2-NEXT: vmovaps %ymm0, 512(%r9)6695; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6696; AVX2-NEXT: vmovaps %ymm0, 480(%r9)6697; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6698; AVX2-NEXT: vmovaps %ymm0, 352(%r9)6699; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6700; AVX2-NEXT: vmovaps %ymm0, 320(%r9)6701; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6702; AVX2-NEXT: vmovaps %ymm0, 192(%r9)6703; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6704; AVX2-NEXT: vmovaps %ymm0, 160(%r9)6705; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6706; AVX2-NEXT: vmovaps %ymm0, 32(%r9)6707; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload6708; AVX2-NEXT: vmovaps %ymm0, (%r9)6709; AVX2-NEXT: addq $1736, %rsp # imm = 0x6C86710; AVX2-NEXT: vzeroupper6711; AVX2-NEXT: retq6712;6713; AVX2-FP-LABEL: store_i32_stride5_vf64:6714; AVX2-FP: # %bb.0:6715; AVX2-FP-NEXT: subq $1736, %rsp # imm = 0x6C86716; AVX2-FP-NEXT: vmovaps (%r8), %ymm156717; AVX2-FP-NEXT: vmovups %ymm15, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6718; AVX2-FP-NEXT: vmovaps 32(%r8), %ymm146719; AVX2-FP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6720; AVX2-FP-NEXT: vmovaps 64(%r8), %ymm136721; AVX2-FP-NEXT: vmovups %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6722; AVX2-FP-NEXT: vmovaps (%rcx), %xmm76723; AVX2-FP-NEXT: vmovaps 32(%rcx), %xmm36724; AVX2-FP-NEXT: vmovaps 64(%rcx), %xmm06725; AVX2-FP-NEXT: vmovaps (%rdx), %xmm86726; AVX2-FP-NEXT: vmovaps 32(%rdx), %xmm46727; AVX2-FP-NEXT: vmovaps 64(%rdx), %xmm16728; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm2 = xmm8[0],xmm7[0],xmm8[1],xmm7[1]6729; AVX2-FP-NEXT: vshufps {{.*#+}} xmm2 = xmm2[0,1,2,2]6730; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm5 = ymm2[0,0,2,1]6731; AVX2-FP-NEXT: vmovaps (%rsi), %xmm96732; AVX2-FP-NEXT: vmovaps 32(%rsi), %xmm66733; AVX2-FP-NEXT: vmovaps 64(%rsi), %xmm26734; AVX2-FP-NEXT: vmovaps (%rdi), %xmm106735; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm11 = xmm10[0],xmm9[0],xmm10[1],xmm9[1]6736; AVX2-FP-NEXT: vshufps {{.*#+}} xmm11 = xmm11[0,1,3,2]6737; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm12 = ymm11[0,1,1,1]6738; AVX2-FP-NEXT: vblendps {{.*#+}} ymm5 = ymm12[0,1],ymm5[2,3],ymm12[4,5,6],ymm5[7]6739; AVX2-FP-NEXT: vmovaps 32(%rdi), %xmm126740; AVX2-FP-NEXT: vinsertf128 $1, (%r8), %ymm11, %ymm116741; AVX2-FP-NEXT: vblendps {{.*#+}} ymm5 = ymm11[0],ymm5[1,2,3],ymm11[4],ymm5[5,6,7]6742; AVX2-FP-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6743; AVX2-FP-NEXT: vmovaps 64(%rdi), %xmm56744; AVX2-FP-NEXT: vunpckhps {{.*#+}} xmm9 = xmm10[2],xmm9[2],xmm10[3],xmm9[3]6745; AVX2-FP-NEXT: vshufps {{.*#+}} xmm9 = xmm9[0,1,2,2]6746; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm9 = ymm9[0,0,2,1]6747; AVX2-FP-NEXT: vshufps {{.*#+}} xmm7 = xmm7[1,1,2,2]6748; AVX2-FP-NEXT: vblendps {{.*#+}} xmm7 = xmm7[0,1],xmm8[2],xmm7[3]6749; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm7 = ymm7[0,1,1,3]6750; AVX2-FP-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0,1],ymm9[2,3],ymm7[4,5],ymm9[6,7]6751; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm8 = ymm15[0,1,2,1]6752; AVX2-FP-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0],ymm8[1],ymm7[2,3,4,5],ymm8[6],ymm7[7]6753; AVX2-FP-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6754; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm7 = xmm4[0],xmm3[0],xmm4[1],xmm3[1]6755; AVX2-FP-NEXT: vshufps {{.*#+}} xmm7 = xmm7[0,1,2,2]6756; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm7 = ymm7[0,0,2,1]6757; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm8 = xmm12[0],xmm6[0],xmm12[1],xmm6[1]6758; AVX2-FP-NEXT: vshufps {{.*#+}} xmm8 = xmm8[0,1,3,2]6759; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm9 = ymm8[0,1,1,1]6760; AVX2-FP-NEXT: vblendps {{.*#+}} ymm7 = ymm9[0,1],ymm7[2,3],ymm9[4,5,6],ymm7[7]6761; AVX2-FP-NEXT: vinsertf128 $1, 32(%r8), %ymm8, %ymm86762; AVX2-FP-NEXT: vblendps {{.*#+}} ymm7 = ymm8[0],ymm7[1,2,3],ymm8[4],ymm7[5,6,7]6763; AVX2-FP-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6764; AVX2-FP-NEXT: vunpckhps {{.*#+}} xmm6 = xmm12[2],xmm6[2],xmm12[3],xmm6[3]6765; AVX2-FP-NEXT: vshufps {{.*#+}} xmm6 = xmm6[0,1,2,2]6766; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm6 = ymm6[0,0,2,1]6767; AVX2-FP-NEXT: vshufps {{.*#+}} xmm3 = xmm3[1,1,2,2]6768; AVX2-FP-NEXT: vblendps {{.*#+}} xmm3 = xmm3[0,1],xmm4[2],xmm3[3]6769; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm3 = ymm3[0,1,1,3]6770; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0,1],ymm6[2,3],ymm3[4,5],ymm6[6,7]6771; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm4 = ymm14[0,1,2,1]6772; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm3[0],ymm4[1],ymm3[2,3,4,5],ymm4[6],ymm3[7]6773; AVX2-FP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6774; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm3 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]6775; AVX2-FP-NEXT: vshufps {{.*#+}} xmm3 = xmm3[0,1,2,2]6776; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm3 = ymm3[0,0,2,1]6777; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm4 = xmm5[0],xmm2[0],xmm5[1],xmm2[1]6778; AVX2-FP-NEXT: vshufps {{.*#+}} xmm4 = xmm4[0,1,3,2]6779; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm6 = ymm4[0,1,1,1]6780; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm6[0,1],ymm3[2,3],ymm6[4,5,6],ymm3[7]6781; AVX2-FP-NEXT: vinsertf128 $1, 64(%r8), %ymm4, %ymm46782; AVX2-FP-NEXT: vblendps {{.*#+}} ymm3 = ymm4[0],ymm3[1,2,3],ymm4[4],ymm3[5,6,7]6783; AVX2-FP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6784; AVX2-FP-NEXT: vunpckhps {{.*#+}} xmm2 = xmm5[2],xmm2[2],xmm5[3],xmm2[3]6785; AVX2-FP-NEXT: vshufps {{.*#+}} xmm2 = xmm2[0,1,2,2]6786; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,0,2,1]6787; AVX2-FP-NEXT: vshufps {{.*#+}} xmm0 = xmm0[1,1,2,2]6788; AVX2-FP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2],xmm0[3]6789; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,1,1,3]6790; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3],ymm0[4,5],ymm2[6,7]6791; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm1 = ymm13[0,1,2,1]6792; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2,3,4,5],ymm1[6],ymm0[7]6793; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6794; AVX2-FP-NEXT: vmovaps 96(%rcx), %xmm06795; AVX2-FP-NEXT: vmovaps 96(%rdx), %xmm16796; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm2 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]6797; AVX2-FP-NEXT: vshufps {{.*#+}} xmm2 = xmm2[0,1,2,2]6798; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,0,2,1]6799; AVX2-FP-NEXT: vmovaps 96(%rsi), %xmm36800; AVX2-FP-NEXT: vmovaps 96(%rdi), %xmm46801; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm5 = xmm4[0],xmm3[0],xmm4[1],xmm3[1]6802; AVX2-FP-NEXT: vshufps {{.*#+}} xmm5 = xmm5[0,1,3,2]6803; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm6 = ymm5[0,1,1,1]6804; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm6[0,1],ymm2[2,3],ymm6[4,5,6],ymm2[7]6805; AVX2-FP-NEXT: vmovaps 96(%r8), %ymm66806; AVX2-FP-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6807; AVX2-FP-NEXT: vinsertf128 $1, 96(%r8), %ymm5, %ymm56808; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm5[0],ymm2[1,2,3],ymm5[4],ymm2[5,6,7]6809; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6810; AVX2-FP-NEXT: vunpckhps {{.*#+}} xmm2 = xmm4[2],xmm3[2],xmm4[3],xmm3[3]6811; AVX2-FP-NEXT: vshufps {{.*#+}} xmm2 = xmm2[0,1,2,2]6812; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,0,2,1]6813; AVX2-FP-NEXT: vshufps {{.*#+}} xmm0 = xmm0[1,1,2,2]6814; AVX2-FP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2],xmm0[3]6815; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,1,1,3]6816; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3],ymm0[4,5],ymm2[6,7]6817; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm1 = ymm6[0,1,2,1]6818; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2,3,4,5],ymm1[6],ymm0[7]6819; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6820; AVX2-FP-NEXT: vmovaps 128(%rcx), %xmm06821; AVX2-FP-NEXT: vmovaps 128(%rdx), %xmm16822; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm2 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]6823; AVX2-FP-NEXT: vshufps {{.*#+}} xmm2 = xmm2[0,1,2,2]6824; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,0,2,1]6825; AVX2-FP-NEXT: vmovaps 128(%rsi), %xmm36826; AVX2-FP-NEXT: vmovaps 128(%rdi), %xmm46827; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm5 = xmm4[0],xmm3[0],xmm4[1],xmm3[1]6828; AVX2-FP-NEXT: vshufps {{.*#+}} xmm5 = xmm5[0,1,3,2]6829; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm6 = ymm5[0,1,1,1]6830; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm6[0,1],ymm2[2,3],ymm6[4,5,6],ymm2[7]6831; AVX2-FP-NEXT: vinsertf128 $1, 128(%r8), %ymm5, %ymm56832; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm5[0],ymm2[1,2,3],ymm5[4],ymm2[5,6,7]6833; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6834; AVX2-FP-NEXT: vunpckhps {{.*#+}} xmm2 = xmm4[2],xmm3[2],xmm4[3],xmm3[3]6835; AVX2-FP-NEXT: vshufps {{.*#+}} xmm0 = xmm0[1,1,2,2]6836; AVX2-FP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2],xmm0[3]6837; AVX2-FP-NEXT: vmovaps 128(%r8), %ymm36838; AVX2-FP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6839; AVX2-FP-NEXT: vshufps {{.*#+}} xmm1 = xmm2[0,1,2,2]6840; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,0,2,1]6841; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,1,1,3]6842; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7]6843; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm1 = ymm3[0,1,2,1]6844; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2,3,4,5],ymm1[6],ymm0[7]6845; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6846; AVX2-FP-NEXT: vmovaps 160(%rcx), %xmm06847; AVX2-FP-NEXT: vmovaps 160(%rdx), %xmm16848; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm2 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]6849; AVX2-FP-NEXT: vshufps {{.*#+}} xmm2 = xmm2[0,1,2,2]6850; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,0,2,1]6851; AVX2-FP-NEXT: vmovaps 160(%rsi), %xmm36852; AVX2-FP-NEXT: vmovaps 160(%rdi), %xmm46853; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm5 = xmm4[0],xmm3[0],xmm4[1],xmm3[1]6854; AVX2-FP-NEXT: vshufps {{.*#+}} xmm5 = xmm5[0,1,3,2]6855; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm6 = ymm5[0,1,1,1]6856; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm6[0,1],ymm2[2,3],ymm6[4,5,6],ymm2[7]6857; AVX2-FP-NEXT: vinsertf128 $1, 160(%r8), %ymm5, %ymm56858; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm5[0],ymm2[1,2,3],ymm5[4],ymm2[5,6,7]6859; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6860; AVX2-FP-NEXT: vunpckhps {{.*#+}} xmm2 = xmm4[2],xmm3[2],xmm4[3],xmm3[3]6861; AVX2-FP-NEXT: vshufps {{.*#+}} xmm0 = xmm0[1,1,2,2]6862; AVX2-FP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2],xmm0[3]6863; AVX2-FP-NEXT: vshufps {{.*#+}} xmm1 = xmm2[0,1,2,2]6864; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,0,2,1]6865; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,1,1,3]6866; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7]6867; AVX2-FP-NEXT: vmovaps 160(%r8), %ymm16868; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6869; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]6870; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2,3,4,5],ymm1[6],ymm0[7]6871; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6872; AVX2-FP-NEXT: vmovaps 192(%rcx), %xmm06873; AVX2-FP-NEXT: vmovaps 192(%rdx), %xmm16874; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm2 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]6875; AVX2-FP-NEXT: vshufps {{.*#+}} xmm2 = xmm2[0,1,2,2]6876; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,0,2,1]6877; AVX2-FP-NEXT: vmovaps 192(%rsi), %xmm36878; AVX2-FP-NEXT: vmovaps 192(%rdi), %xmm46879; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm5 = xmm4[0],xmm3[0],xmm4[1],xmm3[1]6880; AVX2-FP-NEXT: vshufps {{.*#+}} xmm5 = xmm5[0,1,3,2]6881; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm6 = ymm5[0,1,1,1]6882; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm6[0,1],ymm2[2,3],ymm6[4,5,6],ymm2[7]6883; AVX2-FP-NEXT: vinsertf128 $1, 192(%r8), %ymm5, %ymm56884; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm5[0],ymm2[1,2,3],ymm5[4],ymm2[5,6,7]6885; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6886; AVX2-FP-NEXT: vunpckhps {{.*#+}} xmm2 = xmm4[2],xmm3[2],xmm4[3],xmm3[3]6887; AVX2-FP-NEXT: vshufps {{.*#+}} xmm0 = xmm0[1,1,2,2]6888; AVX2-FP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2],xmm0[3]6889; AVX2-FP-NEXT: vshufps {{.*#+}} xmm1 = xmm2[0,1,2,2]6890; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,0,2,1]6891; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,1,1,3]6892; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7]6893; AVX2-FP-NEXT: vmovaps 192(%r8), %ymm16894; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6895; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]6896; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2,3,4,5],ymm1[6],ymm0[7]6897; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6898; AVX2-FP-NEXT: vmovaps 224(%rcx), %xmm06899; AVX2-FP-NEXT: vmovaps 224(%rdx), %xmm16900; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm2 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]6901; AVX2-FP-NEXT: vshufps {{.*#+}} xmm2 = xmm2[0,1,2,2]6902; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,0,2,1]6903; AVX2-FP-NEXT: vmovaps 224(%rsi), %xmm36904; AVX2-FP-NEXT: vmovaps 224(%rdi), %xmm46905; AVX2-FP-NEXT: vunpcklps {{.*#+}} xmm5 = xmm4[0],xmm3[0],xmm4[1],xmm3[1]6906; AVX2-FP-NEXT: vshufps {{.*#+}} xmm5 = xmm5[0,1,3,2]6907; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm6 = ymm5[0,1,1,1]6908; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm6[0,1],ymm2[2,3],ymm6[4,5,6],ymm2[7]6909; AVX2-FP-NEXT: vinsertf128 $1, 224(%r8), %ymm5, %ymm56910; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm5[0],ymm2[1,2,3],ymm5[4],ymm2[5,6,7]6911; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6912; AVX2-FP-NEXT: vunpckhps {{.*#+}} xmm2 = xmm4[2],xmm3[2],xmm4[3],xmm3[3]6913; AVX2-FP-NEXT: vshufps {{.*#+}} xmm0 = xmm0[1,1,2,2]6914; AVX2-FP-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm1[2],xmm0[3]6915; AVX2-FP-NEXT: vshufps {{.*#+}} xmm1 = xmm2[0,1,2,2]6916; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,0,2,1]6917; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,1,1,3]6918; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7]6919; AVX2-FP-NEXT: vmovaps 224(%r8), %ymm16920; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6921; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[0,1,2,1]6922; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2,3,4,5],ymm1[6],ymm0[7]6923; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6924; AVX2-FP-NEXT: vmovaps (%rdx), %ymm156925; AVX2-FP-NEXT: vmovaps (%rcx), %ymm26926; AVX2-FP-NEXT: vshufps {{.*#+}} ymm0 = ymm2[1,1,1,1,5,5,5,5]6927; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6928; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm15[1],ymm0[2,3,4],ymm15[5],ymm0[6,7]6929; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2]6930; AVX2-FP-NEXT: vmovaps (%rdi), %ymm36931; AVX2-FP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6932; AVX2-FP-NEXT: vmovaps (%rsi), %ymm136933; AVX2-FP-NEXT: vshufps {{.*#+}} ymm1 = ymm13[1,1,2,2,5,5,6,6]6934; AVX2-FP-NEXT: vmovups %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6935; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm3[1,2],ymm1[3,4],ymm3[5,6],ymm1[7]6936; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,2,2,3]6937; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4],ymm1[5,6,7]6938; AVX2-FP-NEXT: vbroadcastsd 16(%r8), %ymm16939; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2,3,4],ymm1[5],ymm0[6,7]6940; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6941; AVX2-FP-NEXT: vshufps {{.*#+}} ymm0 = ymm2[0,2,3,3,4,6,7,7]6942; AVX2-FP-NEXT: vshufps {{.*#+}} ymm1 = ymm15[2,3,2,3,6,7,6,7]6943; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2],ymm1[3,4],ymm0[5,6],ymm1[7]6944; AVX2-FP-NEXT: vunpckhps {{.*#+}} ymm1 = ymm13[2],ymm3[2],ymm13[3],ymm3[3],ymm13[6],ymm3[6],ymm13[7],ymm3[7]6945; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[3,3,3,3]6946; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,1,3,3]6947; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3,4],ymm0[5,6,7]6948; AVX2-FP-NEXT: vbroadcastsd 24(%r8), %ymm16949; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4,5,6],ymm1[7]6950; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6951; AVX2-FP-NEXT: vmovaps 32(%rdx), %ymm36952; AVX2-FP-NEXT: vmovaps 32(%rcx), %ymm46953; AVX2-FP-NEXT: vshufps {{.*#+}} ymm0 = ymm4[1,1,1,1,5,5,5,5]6954; AVX2-FP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6955; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm3[1],ymm0[2,3,4],ymm3[5],ymm0[6,7]6956; AVX2-FP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6957; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2]6958; AVX2-FP-NEXT: vmovaps 32(%rdi), %ymm26959; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6960; AVX2-FP-NEXT: vmovaps 32(%rsi), %ymm136961; AVX2-FP-NEXT: vshufps {{.*#+}} ymm1 = ymm13[1,1,2,2,5,5,6,6]6962; AVX2-FP-NEXT: vmovups %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6963; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1,2],ymm1[3,4],ymm2[5,6],ymm1[7]6964; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,2,2,3]6965; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4],ymm1[5,6,7]6966; AVX2-FP-NEXT: vbroadcastsd 48(%r8), %ymm16967; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2,3,4],ymm1[5],ymm0[6,7]6968; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6969; AVX2-FP-NEXT: vshufps {{.*#+}} ymm0 = ymm4[0,2,3,3,4,6,7,7]6970; AVX2-FP-NEXT: vshufps {{.*#+}} ymm1 = ymm3[2,3,2,3,6,7,6,7]6971; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2],ymm1[3,4],ymm0[5,6],ymm1[7]6972; AVX2-FP-NEXT: vunpckhps {{.*#+}} ymm1 = ymm13[2],ymm2[2],ymm13[3],ymm2[3],ymm13[6],ymm2[6],ymm13[7],ymm2[7]6973; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[3,3,3,3]6974; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,1,3,3]6975; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3,4],ymm0[5,6,7]6976; AVX2-FP-NEXT: vbroadcastsd 56(%r8), %ymm16977; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4,5,6],ymm1[7]6978; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6979; AVX2-FP-NEXT: vmovaps 64(%rdx), %ymm36980; AVX2-FP-NEXT: vmovaps 64(%rcx), %ymm46981; AVX2-FP-NEXT: vshufps {{.*#+}} ymm0 = ymm4[1,1,1,1,5,5,5,5]6982; AVX2-FP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6983; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm3[1],ymm0[2,3,4],ymm3[5],ymm0[6,7]6984; AVX2-FP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6985; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2]6986; AVX2-FP-NEXT: vmovaps 64(%rdi), %ymm26987; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6988; AVX2-FP-NEXT: vmovaps 64(%rsi), %ymm136989; AVX2-FP-NEXT: vshufps {{.*#+}} ymm1 = ymm13[1,1,2,2,5,5,6,6]6990; AVX2-FP-NEXT: vmovups %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6991; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1,2],ymm1[3,4],ymm2[5,6],ymm1[7]6992; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,2,2,3]6993; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4],ymm1[5,6,7]6994; AVX2-FP-NEXT: vbroadcastsd 80(%r8), %ymm16995; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2,3,4],ymm1[5],ymm0[6,7]6996; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill6997; AVX2-FP-NEXT: vshufps {{.*#+}} ymm0 = ymm4[0,2,3,3,4,6,7,7]6998; AVX2-FP-NEXT: vshufps {{.*#+}} ymm1 = ymm3[2,3,2,3,6,7,6,7]6999; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2],ymm1[3,4],ymm0[5,6],ymm1[7]7000; AVX2-FP-NEXT: vunpckhps {{.*#+}} ymm1 = ymm13[2],ymm2[2],ymm13[3],ymm2[3],ymm13[6],ymm2[6],ymm13[7],ymm2[7]7001; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[3,3,3,3]7002; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,1,3,3]7003; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3,4],ymm0[5,6,7]7004; AVX2-FP-NEXT: vbroadcastsd 88(%r8), %ymm17005; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4,5,6],ymm1[7]7006; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7007; AVX2-FP-NEXT: vmovaps 96(%rdx), %ymm37008; AVX2-FP-NEXT: vmovaps 96(%rcx), %ymm47009; AVX2-FP-NEXT: vshufps {{.*#+}} ymm0 = ymm4[1,1,1,1,5,5,5,5]7010; AVX2-FP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7011; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm3[1],ymm0[2,3,4],ymm3[5],ymm0[6,7]7012; AVX2-FP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7013; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2]7014; AVX2-FP-NEXT: vmovaps 96(%rdi), %ymm27015; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7016; AVX2-FP-NEXT: vmovaps 96(%rsi), %ymm147017; AVX2-FP-NEXT: vshufps {{.*#+}} ymm1 = ymm14[1,1,2,2,5,5,6,6]7018; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1,2],ymm1[3,4],ymm2[5,6],ymm1[7]7019; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,2,2,3]7020; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4],ymm1[5,6,7]7021; AVX2-FP-NEXT: vbroadcastsd 112(%r8), %ymm17022; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2,3,4],ymm1[5],ymm0[6,7]7023; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7024; AVX2-FP-NEXT: vshufps {{.*#+}} ymm0 = ymm4[0,2,3,3,4,6,7,7]7025; AVX2-FP-NEXT: vshufps {{.*#+}} ymm1 = ymm3[2,3,2,3,6,7,6,7]7026; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2],ymm1[3,4],ymm0[5,6],ymm1[7]7027; AVX2-FP-NEXT: vunpckhps {{.*#+}} ymm1 = ymm14[2],ymm2[2],ymm14[3],ymm2[3],ymm14[6],ymm2[6],ymm14[7],ymm2[7]7028; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[3,3,3,3]7029; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,1,3,3]7030; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3,4],ymm0[5,6,7]7031; AVX2-FP-NEXT: vbroadcastsd 120(%r8), %ymm17032; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4,5,6],ymm1[7]7033; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7034; AVX2-FP-NEXT: vmovaps 128(%rdx), %ymm47035; AVX2-FP-NEXT: vmovaps 128(%rcx), %ymm27036; AVX2-FP-NEXT: vshufps {{.*#+}} ymm0 = ymm2[1,1,1,1,5,5,5,5]7037; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7038; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm4[1],ymm0[2,3,4],ymm4[5],ymm0[6,7]7039; AVX2-FP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7040; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2]7041; AVX2-FP-NEXT: vmovaps 128(%rdi), %ymm37042; AVX2-FP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7043; AVX2-FP-NEXT: vmovaps 128(%rsi), %ymm137044; AVX2-FP-NEXT: vshufps {{.*#+}} ymm1 = ymm13[1,1,2,2,5,5,6,6]7045; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm3[1,2],ymm1[3,4],ymm3[5,6],ymm1[7]7046; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,2,2,3]7047; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4],ymm1[5,6,7]7048; AVX2-FP-NEXT: vbroadcastsd 144(%r8), %ymm17049; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2,3,4],ymm1[5],ymm0[6,7]7050; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7051; AVX2-FP-NEXT: vshufps {{.*#+}} ymm0 = ymm2[0,2,3,3,4,6,7,7]7052; AVX2-FP-NEXT: vshufps {{.*#+}} ymm1 = ymm4[2,3,2,3,6,7,6,7]7053; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2],ymm1[3,4],ymm0[5,6],ymm1[7]7054; AVX2-FP-NEXT: vunpckhps {{.*#+}} ymm1 = ymm13[2],ymm3[2],ymm13[3],ymm3[3],ymm13[6],ymm3[6],ymm13[7],ymm3[7]7055; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[3,3,3,3]7056; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,1,3,3]7057; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3,4],ymm0[5,6,7]7058; AVX2-FP-NEXT: vbroadcastsd 152(%r8), %ymm17059; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4,5,6],ymm1[7]7060; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7061; AVX2-FP-NEXT: vmovaps 160(%rdx), %ymm27062; AVX2-FP-NEXT: vmovaps 160(%rcx), %ymm127063; AVX2-FP-NEXT: vshufps {{.*#+}} ymm0 = ymm12[1,1,1,1,5,5,5,5]7064; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm2[1],ymm0[2,3,4],ymm2[5],ymm0[6,7]7065; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7066; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2]7067; AVX2-FP-NEXT: vmovaps 160(%rdi), %ymm107068; AVX2-FP-NEXT: vmovaps 160(%rsi), %ymm97069; AVX2-FP-NEXT: vshufps {{.*#+}} ymm1 = ymm9[1,1,2,2,5,5,6,6]7070; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm10[1,2],ymm1[3,4],ymm10[5,6],ymm1[7]7071; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,2,2,3]7072; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4],ymm1[5,6,7]7073; AVX2-FP-NEXT: vbroadcastsd 176(%r8), %ymm17074; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2,3,4],ymm1[5],ymm0[6,7]7075; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7076; AVX2-FP-NEXT: vshufps {{.*#+}} ymm0 = ymm12[0,2,3,3,4,6,7,7]7077; AVX2-FP-NEXT: vshufps {{.*#+}} ymm1 = ymm2[2,3,2,3,6,7,6,7]7078; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2],ymm1[3,4],ymm0[5,6],ymm1[7]7079; AVX2-FP-NEXT: vunpckhps {{.*#+}} ymm1 = ymm9[2],ymm10[2],ymm9[3],ymm10[3],ymm9[6],ymm10[6],ymm9[7],ymm10[7]7080; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[3,3,3,3]7081; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,1,3,3]7082; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3,4],ymm0[5,6,7]7083; AVX2-FP-NEXT: vbroadcastsd 184(%r8), %ymm17084; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4,5,6],ymm1[7]7085; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7086; AVX2-FP-NEXT: vmovaps 192(%rdx), %ymm87087; AVX2-FP-NEXT: vmovaps 192(%rcx), %ymm67088; AVX2-FP-NEXT: vshufps {{.*#+}} ymm0 = ymm6[1,1,1,1,5,5,5,5]7089; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm8[1],ymm0[2,3,4],ymm8[5],ymm0[6,7]7090; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2]7091; AVX2-FP-NEXT: vmovaps 192(%rdi), %ymm57092; AVX2-FP-NEXT: vmovaps 192(%rsi), %ymm47093; AVX2-FP-NEXT: vshufps {{.*#+}} ymm1 = ymm4[1,1,2,2,5,5,6,6]7094; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm5[1,2],ymm1[3,4],ymm5[5,6],ymm1[7]7095; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,2,2,3]7096; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4],ymm1[5,6,7]7097; AVX2-FP-NEXT: vbroadcastsd 208(%r8), %ymm17098; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2,3,4],ymm1[5],ymm0[6,7]7099; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7100; AVX2-FP-NEXT: vshufps {{.*#+}} ymm0 = ymm6[0,2,3,3,4,6,7,7]7101; AVX2-FP-NEXT: vshufps {{.*#+}} ymm1 = ymm8[2,3,2,3,6,7,6,7]7102; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2],ymm1[3,4],ymm0[5,6],ymm1[7]7103; AVX2-FP-NEXT: vunpckhps {{.*#+}} ymm1 = ymm4[2],ymm5[2],ymm4[3],ymm5[3],ymm4[6],ymm5[6],ymm4[7],ymm5[7]7104; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[3,3,3,3]7105; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,1,3,3]7106; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3,4],ymm0[5,6,7]7107; AVX2-FP-NEXT: vbroadcastsd 216(%r8), %ymm17108; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4,5,6],ymm1[7]7109; AVX2-FP-NEXT: vmovups %ymm0, (%rsp) # 32-byte Spill7110; AVX2-FP-NEXT: vmovaps 224(%rdx), %ymm37111; AVX2-FP-NEXT: vmovaps 224(%rcx), %ymm27112; AVX2-FP-NEXT: vshufps {{.*#+}} ymm0 = ymm2[1,1,1,1,5,5,5,5]7113; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm3[1],ymm0[2,3,4],ymm3[5],ymm0[6,7]7114; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm7 = ymm0[2,2,2,2]7115; AVX2-FP-NEXT: vmovaps 224(%rdi), %ymm17116; AVX2-FP-NEXT: vmovaps 224(%rsi), %ymm07117; AVX2-FP-NEXT: vshufps {{.*#+}} ymm11 = ymm0[1,1,2,2,5,5,6,6]7118; AVX2-FP-NEXT: vblendps {{.*#+}} ymm11 = ymm11[0],ymm1[1,2],ymm11[3,4],ymm1[5,6],ymm11[7]7119; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm11 = ymm11[2,2,2,3]7120; AVX2-FP-NEXT: vblendps {{.*#+}} ymm7 = ymm11[0,1,2],ymm7[3,4],ymm11[5,6,7]7121; AVX2-FP-NEXT: vbroadcastsd 240(%r8), %ymm117122; AVX2-FP-NEXT: vblendps {{.*#+}} ymm7 = ymm11[0],ymm7[1,2,3,4],ymm11[5],ymm7[6,7]7123; AVX2-FP-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7124; AVX2-FP-NEXT: vshufps {{.*#+}} ymm7 = ymm2[0,2,3,3,4,6,7,7]7125; AVX2-FP-NEXT: vshufps {{.*#+}} ymm11 = ymm3[2,3,2,3,6,7,6,7]7126; AVX2-FP-NEXT: vblendps {{.*#+}} ymm7 = ymm11[0],ymm7[1,2],ymm11[3,4],ymm7[5,6],ymm11[7]7127; AVX2-FP-NEXT: vunpckhps {{.*#+}} ymm11 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7]7128; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm11 = ymm11[3,3,3,3]7129; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm7 = ymm7[2,1,3,3]7130; AVX2-FP-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0,1,2],ymm11[3,4],ymm7[5,6,7]7131; AVX2-FP-NEXT: vbroadcastsd 248(%r8), %ymm117132; AVX2-FP-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0,1],ymm11[2],ymm7[3,4,5,6],ymm11[7]7133; AVX2-FP-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7134; AVX2-FP-NEXT: vpermilps $227, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Folded Reload7135; AVX2-FP-NEXT: # ymm7 = mem[3,0,2,3,7,4,6,7]7136; AVX2-FP-NEXT: vblendps $16, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm7 # 32-byte Folded Reload7137; AVX2-FP-NEXT: # ymm7 = ymm7[0,1,2,3],mem[4],ymm7[5,6,7]7138; AVX2-FP-NEXT: vpermilps $52, {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Folded Reload7139; AVX2-FP-NEXT: # ymm11 = mem[0,1,3,0,4,5,7,4]7140; AVX2-FP-NEXT: vshufps {{.*#+}} ymm15 = ymm15[2,3,0,1,6,7,4,5]7141; AVX2-FP-NEXT: vblendps {{.*#+}} ymm11 = ymm15[0,1],ymm11[2],ymm15[3,4,5,6],ymm11[7]7142; AVX2-FP-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0],ymm11[1,2,3],ymm7[4,5],ymm11[6,7]7143; AVX2-FP-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm7 # 32-byte Folded Reload7144; AVX2-FP-NEXT: # ymm7 = ymm7[0,1,2],mem[3],ymm7[4,5,6,7]7145; AVX2-FP-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7146; AVX2-FP-NEXT: vpermilps $227, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Folded Reload7147; AVX2-FP-NEXT: # ymm7 = mem[3,0,2,3,7,4,6,7]7148; AVX2-FP-NEXT: vblendps $16, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm7 # 32-byte Folded Reload7149; AVX2-FP-NEXT: # ymm7 = ymm7[0,1,2,3],mem[4],ymm7[5,6,7]7150; AVX2-FP-NEXT: vpermilps $52, {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Folded Reload7151; AVX2-FP-NEXT: # ymm11 = mem[0,1,3,0,4,5,7,4]7152; AVX2-FP-NEXT: vpermilps $78, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Folded Reload7153; AVX2-FP-NEXT: # ymm15 = mem[2,3,0,1,6,7,4,5]7154; AVX2-FP-NEXT: vblendps {{.*#+}} ymm11 = ymm15[0,1],ymm11[2],ymm15[3,4,5,6],ymm11[7]7155; AVX2-FP-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0],ymm11[1,2,3],ymm7[4,5],ymm11[6,7]7156; AVX2-FP-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm7 # 32-byte Folded Reload7157; AVX2-FP-NEXT: # ymm7 = ymm7[0,1,2],mem[3],ymm7[4,5,6,7]7158; AVX2-FP-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7159; AVX2-FP-NEXT: vpermilps $227, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Folded Reload7160; AVX2-FP-NEXT: # ymm7 = mem[3,0,2,3,7,4,6,7]7161; AVX2-FP-NEXT: vblendps $16, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm7 # 32-byte Folded Reload7162; AVX2-FP-NEXT: # ymm7 = ymm7[0,1,2,3],mem[4],ymm7[5,6,7]7163; AVX2-FP-NEXT: vpermilps $52, {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Folded Reload7164; AVX2-FP-NEXT: # ymm11 = mem[0,1,3,0,4,5,7,4]7165; AVX2-FP-NEXT: vpermilps $78, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Folded Reload7166; AVX2-FP-NEXT: # ymm15 = mem[2,3,0,1,6,7,4,5]7167; AVX2-FP-NEXT: vblendps {{.*#+}} ymm11 = ymm15[0,1],ymm11[2],ymm15[3,4,5,6],ymm11[7]7168; AVX2-FP-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0],ymm11[1,2,3],ymm7[4,5],ymm11[6,7]7169; AVX2-FP-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm11 # 32-byte Folded Reload7170; AVX2-FP-NEXT: # ymm11 = ymm7[0,1,2],mem[3],ymm7[4,5,6,7]7171; AVX2-FP-NEXT: vshufps {{.*#+}} ymm7 = ymm14[3,0,2,3,7,4,6,7]7172; AVX2-FP-NEXT: vblendps $16, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm7 # 32-byte Folded Reload7173; AVX2-FP-NEXT: # ymm7 = ymm7[0,1,2,3],mem[4],ymm7[5,6,7]7174; AVX2-FP-NEXT: vpermilps $52, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Folded Reload7175; AVX2-FP-NEXT: # ymm15 = mem[0,1,3,0,4,5,7,4]7176; AVX2-FP-NEXT: vpermilps $78, {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Folded Reload7177; AVX2-FP-NEXT: # ymm14 = mem[2,3,0,1,6,7,4,5]7178; AVX2-FP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1],ymm15[2],ymm14[3,4,5,6],ymm15[7]7179; AVX2-FP-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0],ymm14[1,2,3],ymm7[4,5],ymm14[6,7]7180; AVX2-FP-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm7 # 32-byte Folded Reload7181; AVX2-FP-NEXT: # ymm7 = ymm7[0,1,2],mem[3],ymm7[4,5,6,7]7182; AVX2-FP-NEXT: vshufps {{.*#+}} ymm14 = ymm13[3,0,2,3,7,4,6,7]7183; AVX2-FP-NEXT: vblendps $16, {{[-0-9]+}}(%r{{[sb]}}p), %ymm14, %ymm14 # 32-byte Folded Reload7184; AVX2-FP-NEXT: # ymm14 = ymm14[0,1,2,3],mem[4],ymm14[5,6,7]7185; AVX2-FP-NEXT: vpermilps $52, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Folded Reload7186; AVX2-FP-NEXT: # ymm15 = mem[0,1,3,0,4,5,7,4]7187; AVX2-FP-NEXT: vpermilps $78, {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Folded Reload7188; AVX2-FP-NEXT: # ymm13 = mem[2,3,0,1,6,7,4,5]7189; AVX2-FP-NEXT: vblendps {{.*#+}} ymm13 = ymm13[0,1],ymm15[2],ymm13[3,4,5,6],ymm15[7]7190; AVX2-FP-NEXT: vblendps {{.*#+}} ymm13 = ymm14[0],ymm13[1,2,3],ymm14[4,5],ymm13[6,7]7191; AVX2-FP-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm13 # 32-byte Folded Reload7192; AVX2-FP-NEXT: # ymm13 = ymm13[0,1,2],mem[3],ymm13[4,5,6,7]7193; AVX2-FP-NEXT: vshufps {{.*#+}} ymm9 = ymm9[3,0,2,3,7,4,6,7]7194; AVX2-FP-NEXT: vblendps {{.*#+}} ymm9 = ymm9[0,1,2,3],ymm10[4],ymm9[5,6,7]7195; AVX2-FP-NEXT: vshufps {{.*#+}} ymm10 = ymm12[0,1,3,0,4,5,7,4]7196; AVX2-FP-NEXT: vpermilps $78, {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Folded Reload7197; AVX2-FP-NEXT: # ymm12 = mem[2,3,0,1,6,7,4,5]7198; AVX2-FP-NEXT: vblendps {{.*#+}} ymm10 = ymm12[0,1],ymm10[2],ymm12[3,4,5,6],ymm10[7]7199; AVX2-FP-NEXT: vblendps {{.*#+}} ymm9 = ymm9[0],ymm10[1,2,3],ymm9[4,5],ymm10[6,7]7200; AVX2-FP-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm9, %ymm9 # 32-byte Folded Reload7201; AVX2-FP-NEXT: # ymm9 = ymm9[0,1,2],mem[3],ymm9[4,5,6,7]7202; AVX2-FP-NEXT: vshufps {{.*#+}} ymm4 = ymm4[3,0,2,3,7,4,6,7]7203; AVX2-FP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm5[4],ymm4[5,6,7]7204; AVX2-FP-NEXT: vshufps {{.*#+}} ymm5 = ymm6[0,1,3,0,4,5,7,4]7205; AVX2-FP-NEXT: vshufps {{.*#+}} ymm6 = ymm8[2,3,0,1,6,7,4,5]7206; AVX2-FP-NEXT: vblendps {{.*#+}} ymm5 = ymm6[0,1],ymm5[2],ymm6[3,4,5,6],ymm5[7]7207; AVX2-FP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0],ymm5[1,2,3],ymm4[4,5],ymm5[6,7]7208; AVX2-FP-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4, %ymm4 # 32-byte Folded Reload7209; AVX2-FP-NEXT: # ymm4 = ymm4[0,1,2],mem[3],ymm4[4,5,6,7]7210; AVX2-FP-NEXT: vshufps {{.*#+}} ymm0 = ymm0[3,0,2,3,7,4,6,7]7211; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4],ymm0[5,6,7]7212; AVX2-FP-NEXT: vshufps {{.*#+}} ymm1 = ymm2[0,1,3,0,4,5,7,4]7213; AVX2-FP-NEXT: vshufps {{.*#+}} ymm2 = ymm3[2,3,0,1,6,7,4,5]7214; AVX2-FP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1],ymm1[2],ymm2[3,4,5,6],ymm1[7]7215; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm1[1,2,3],ymm0[4,5],ymm1[6,7]7216; AVX2-FP-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload7217; AVX2-FP-NEXT: # ymm0 = ymm0[0,1,2],mem[3],ymm0[4,5,6,7]7218; AVX2-FP-NEXT: vmovaps %ymm0, 1184(%r9)7219; AVX2-FP-NEXT: vmovaps %ymm4, 1024(%r9)7220; AVX2-FP-NEXT: vmovaps %ymm9, 864(%r9)7221; AVX2-FP-NEXT: vmovaps %ymm13, 704(%r9)7222; AVX2-FP-NEXT: vmovaps %ymm7, 544(%r9)7223; AVX2-FP-NEXT: vmovaps %ymm11, 384(%r9)7224; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload7225; AVX2-FP-NEXT: vmovaps %ymm0, 224(%r9)7226; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload7227; AVX2-FP-NEXT: vmovaps %ymm0, 64(%r9)7228; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload7229; AVX2-FP-NEXT: vmovaps %ymm0, 1248(%r9)7230; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload7231; AVX2-FP-NEXT: vmovaps %ymm0, 1216(%r9)7232; AVX2-FP-NEXT: vmovups (%rsp), %ymm0 # 32-byte Reload7233; AVX2-FP-NEXT: vmovaps %ymm0, 1088(%r9)7234; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload7235; AVX2-FP-NEXT: vmovaps %ymm0, 1056(%r9)7236; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload7237; AVX2-FP-NEXT: vmovaps %ymm0, 928(%r9)7238; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload7239; AVX2-FP-NEXT: vmovaps %ymm0, 896(%r9)7240; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload7241; AVX2-FP-NEXT: vmovaps %ymm0, 768(%r9)7242; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload7243; AVX2-FP-NEXT: vmovaps %ymm0, 736(%r9)7244; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload7245; AVX2-FP-NEXT: vmovaps %ymm0, 608(%r9)7246; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload7247; AVX2-FP-NEXT: vmovaps %ymm0, 576(%r9)7248; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload7249; AVX2-FP-NEXT: vmovaps %ymm0, 448(%r9)7250; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload7251; AVX2-FP-NEXT: vmovaps %ymm0, 416(%r9)7252; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload7253; AVX2-FP-NEXT: vmovaps %ymm0, 288(%r9)7254; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload7255; AVX2-FP-NEXT: vmovaps %ymm0, 256(%r9)7256; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload7257; AVX2-FP-NEXT: vmovaps %ymm0, 128(%r9)7258; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload7259; AVX2-FP-NEXT: vmovaps %ymm0, 96(%r9)7260; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload7261; AVX2-FP-NEXT: vmovaps %ymm0, 1152(%r9)7262; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload7263; AVX2-FP-NEXT: vmovaps %ymm0, 1120(%r9)7264; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload7265; AVX2-FP-NEXT: vmovaps %ymm0, 992(%r9)7266; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload7267; AVX2-FP-NEXT: vmovaps %ymm0, 960(%r9)7268; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload7269; AVX2-FP-NEXT: vmovaps %ymm0, 832(%r9)7270; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload7271; AVX2-FP-NEXT: vmovaps %ymm0, 800(%r9)7272; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload7273; AVX2-FP-NEXT: vmovaps %ymm0, 672(%r9)7274; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload7275; AVX2-FP-NEXT: vmovaps %ymm0, 640(%r9)7276; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload7277; AVX2-FP-NEXT: vmovaps %ymm0, 512(%r9)7278; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload7279; AVX2-FP-NEXT: vmovaps %ymm0, 480(%r9)7280; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload7281; AVX2-FP-NEXT: vmovaps %ymm0, 352(%r9)7282; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload7283; AVX2-FP-NEXT: vmovaps %ymm0, 320(%r9)7284; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload7285; AVX2-FP-NEXT: vmovaps %ymm0, 192(%r9)7286; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload7287; AVX2-FP-NEXT: vmovaps %ymm0, 160(%r9)7288; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload7289; AVX2-FP-NEXT: vmovaps %ymm0, 32(%r9)7290; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload7291; AVX2-FP-NEXT: vmovaps %ymm0, (%r9)7292; AVX2-FP-NEXT: addq $1736, %rsp # imm = 0x6C87293; AVX2-FP-NEXT: vzeroupper7294; AVX2-FP-NEXT: retq7295;7296; AVX2-FCP-LABEL: store_i32_stride5_vf64:7297; AVX2-FCP: # %bb.0:7298; AVX2-FCP-NEXT: subq $1800, %rsp # imm = 0x7087299; AVX2-FCP-NEXT: vmovaps (%r8), %ymm87300; AVX2-FCP-NEXT: vmovups %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7301; AVX2-FCP-NEXT: vmovaps 32(%r8), %ymm57302; AVX2-FCP-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7303; AVX2-FCP-NEXT: vmovaps (%rsi), %xmm117304; AVX2-FCP-NEXT: vmovaps 32(%rsi), %xmm137305; AVX2-FCP-NEXT: vmovaps 64(%rsi), %xmm17306; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7307; AVX2-FCP-NEXT: vmovaps (%rdi), %xmm147308; AVX2-FCP-NEXT: vmovaps 32(%rdi), %xmm77309; AVX2-FCP-NEXT: vmovaps 64(%rdi), %xmm67310; AVX2-FCP-NEXT: vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7311; AVX2-FCP-NEXT: vunpckhps {{.*#+}} xmm0 = xmm14[2],xmm11[2],xmm14[3],xmm11[3]7312; AVX2-FCP-NEXT: vmovaps {{.*#+}} ymm9 = [0,1,0,1,u,u,2,2]7313; AVX2-FCP-NEXT: vpermps %ymm0, %ymm9, %ymm27314; AVX2-FCP-NEXT: vmovaps (%rdx), %xmm157315; AVX2-FCP-NEXT: vmovaps 32(%rdx), %xmm107316; AVX2-FCP-NEXT: vmovaps (%rcx), %xmm47317; AVX2-FCP-NEXT: vmovaps 32(%rcx), %xmm127318; AVX2-FCP-NEXT: vshufps {{.*#+}} xmm3 = xmm4[1,1,2,2]7319; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm3 = xmm3[0,1],xmm15[2],xmm3[3]7320; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm3 = ymm3[0,1,1,3]7321; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm3[0,1],ymm2[2,3],ymm3[4,5],ymm2[6,7]7322; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm3 = ymm8[0,1,2,1]7323; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0],ymm3[1],ymm2[2,3,4,5],ymm3[6],ymm2[7]7324; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7325; AVX2-FCP-NEXT: vunpckhps {{.*#+}} xmm2 = xmm7[2],xmm13[2],xmm7[3],xmm13[3]7326; AVX2-FCP-NEXT: vpermps %ymm2, %ymm9, %ymm27327; AVX2-FCP-NEXT: vshufps {{.*#+}} xmm3 = xmm12[1,1,2,2]7328; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm3 = xmm3[0,1],xmm10[2],xmm3[3]7329; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm3 = ymm3[0,1,1,3]7330; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm3[0,1],ymm2[2,3],ymm3[4,5],ymm2[6,7]7331; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm3 = ymm5[0,1,2,1]7332; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0],ymm3[1],ymm2[2,3,4,5],ymm3[6],ymm2[7]7333; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7334; AVX2-FCP-NEXT: vunpckhps {{.*#+}} xmm2 = xmm6[2],xmm1[2],xmm6[3],xmm1[3]7335; AVX2-FCP-NEXT: vpermps %ymm2, %ymm9, %ymm27336; AVX2-FCP-NEXT: vmovaps 64(%rdx), %xmm87337; AVX2-FCP-NEXT: vmovaps 64(%rcx), %xmm67338; AVX2-FCP-NEXT: vshufps {{.*#+}} xmm3 = xmm6[1,1,2,2]7339; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm3 = xmm3[0,1],xmm8[2],xmm3[3]7340; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm3 = ymm3[0,1,1,3]7341; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm3[0,1],ymm2[2,3],ymm3[4,5],ymm2[6,7]7342; AVX2-FCP-NEXT: vmovaps 64(%r8), %ymm07343; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7344; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm3 = ymm0[0,1,2,1]7345; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0],ymm3[1],ymm2[2,3,4,5],ymm3[6],ymm2[7]7346; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7347; AVX2-FCP-NEXT: vmovaps 96(%rsi), %xmm07348; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7349; AVX2-FCP-NEXT: vmovaps 96(%rdi), %xmm17350; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7351; AVX2-FCP-NEXT: vunpckhps {{.*#+}} xmm2 = xmm1[2],xmm0[2],xmm1[3],xmm0[3]7352; AVX2-FCP-NEXT: vpermps %ymm2, %ymm9, %ymm07353; AVX2-FCP-NEXT: vmovaps 96(%rdx), %xmm17354; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7355; AVX2-FCP-NEXT: vmovaps 96(%rcx), %xmm37356; AVX2-FCP-NEXT: vshufps {{.*#+}} xmm5 = xmm3[1,1,2,2]7357; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm5 = xmm5[0,1],xmm1[2],xmm5[3]7358; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm5 = ymm5[0,1,1,3]7359; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm5[0,1],ymm0[2,3],ymm5[4,5],ymm0[6,7]7360; AVX2-FCP-NEXT: vmovaps 96(%r8), %ymm07361; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7362; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm5 = ymm0[0,1,2,1]7363; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm5[1],ymm1[2,3,4,5],ymm5[6],ymm1[7]7364; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7365; AVX2-FCP-NEXT: vmovaps 128(%rsi), %xmm17366; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7367; AVX2-FCP-NEXT: vmovaps 128(%rdi), %xmm07368; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7369; AVX2-FCP-NEXT: vunpckhps {{.*#+}} xmm0 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]7370; AVX2-FCP-NEXT: vpermps %ymm0, %ymm9, %ymm07371; AVX2-FCP-NEXT: vmovaps 128(%rdx), %xmm17372; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7373; AVX2-FCP-NEXT: vmovaps 128(%rcx), %xmm27374; AVX2-FCP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7375; AVX2-FCP-NEXT: vshufps {{.*#+}} xmm5 = xmm2[1,1,2,2]7376; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm5 = xmm5[0,1],xmm1[2],xmm5[3]7377; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm5 = ymm5[0,1,1,3]7378; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm5[0,1],ymm0[2,3],ymm5[4,5],ymm0[6,7]7379; AVX2-FCP-NEXT: vmovaps 128(%r8), %ymm17380; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7381; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm5 = ymm1[0,1,2,1]7382; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm5[1],ymm0[2,3,4,5],ymm5[6],ymm0[7]7383; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7384; AVX2-FCP-NEXT: vmovaps 160(%rsi), %xmm17385; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7386; AVX2-FCP-NEXT: vmovaps 160(%rdi), %xmm07387; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7388; AVX2-FCP-NEXT: vunpckhps {{.*#+}} xmm0 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]7389; AVX2-FCP-NEXT: vpermps %ymm0, %ymm9, %ymm07390; AVX2-FCP-NEXT: vmovaps 160(%rdx), %xmm17391; AVX2-FCP-NEXT: vmovaps %xmm1, (%rsp) # 16-byte Spill7392; AVX2-FCP-NEXT: vmovaps 160(%rcx), %xmm27393; AVX2-FCP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7394; AVX2-FCP-NEXT: vshufps {{.*#+}} xmm5 = xmm2[1,1,2,2]7395; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm5 = xmm5[0,1],xmm1[2],xmm5[3]7396; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm5 = ymm5[0,1,1,3]7397; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm5[0,1],ymm0[2,3],ymm5[4,5],ymm0[6,7]7398; AVX2-FCP-NEXT: vmovaps 160(%r8), %ymm17399; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7400; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm5 = ymm1[0,1,2,1]7401; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm5[1],ymm0[2,3,4,5],ymm5[6],ymm0[7]7402; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7403; AVX2-FCP-NEXT: vmovaps 192(%rsi), %xmm17404; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7405; AVX2-FCP-NEXT: vmovaps 192(%rdi), %xmm07406; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7407; AVX2-FCP-NEXT: vunpckhps {{.*#+}} xmm0 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]7408; AVX2-FCP-NEXT: vpermps %ymm0, %ymm9, %ymm07409; AVX2-FCP-NEXT: vmovaps 192(%rdx), %xmm17410; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7411; AVX2-FCP-NEXT: vmovaps 192(%rcx), %xmm27412; AVX2-FCP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7413; AVX2-FCP-NEXT: vshufps {{.*#+}} xmm5 = xmm2[1,1,2,2]7414; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm5 = xmm5[0,1],xmm1[2],xmm5[3]7415; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm5 = ymm5[0,1,1,3]7416; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm5[0,1],ymm0[2,3],ymm5[4,5],ymm0[6,7]7417; AVX2-FCP-NEXT: vmovaps 192(%r8), %ymm17418; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7419; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm5 = ymm1[0,1,2,1]7420; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm5[1],ymm0[2,3,4,5],ymm5[6],ymm0[7]7421; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7422; AVX2-FCP-NEXT: vmovaps 224(%rsi), %xmm17423; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7424; AVX2-FCP-NEXT: vmovaps 224(%rdi), %xmm07425; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7426; AVX2-FCP-NEXT: vunpckhps {{.*#+}} xmm0 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]7427; AVX2-FCP-NEXT: vpermps %ymm0, %ymm9, %ymm27428; AVX2-FCP-NEXT: vmovaps 224(%rdx), %xmm17429; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7430; AVX2-FCP-NEXT: vmovaps 224(%rcx), %xmm07431; AVX2-FCP-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7432; AVX2-FCP-NEXT: vshufps {{.*#+}} xmm5 = xmm0[1,1,2,2]7433; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm5 = xmm5[0,1],xmm1[2],xmm5[3]7434; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm5 = ymm5[0,1,1,3]7435; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm5[0,1],ymm2[2,3],ymm5[4,5],ymm2[6,7]7436; AVX2-FCP-NEXT: vmovaps 224(%r8), %ymm17437; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7438; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm5 = ymm1[0,1,2,1]7439; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm5[1],ymm0[2,3,4,5],ymm5[6],ymm0[7]7440; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7441; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm0 = xmm15[0],xmm4[0],xmm15[1],xmm4[1]7442; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm14[0],xmm11[0],xmm14[1],xmm11[1]7443; AVX2-FCP-NEXT: vpermps %ymm0, %ymm9, %ymm07444; AVX2-FCP-NEXT: vmovaps {{.*#+}} ymm4 = [0,1,3,2,3,2,3,2]7445; AVX2-FCP-NEXT: vpermps %ymm1, %ymm4, %ymm57446; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm5[0,1],ymm0[2,3],ymm5[4,5,6],ymm0[7]7447; AVX2-FCP-NEXT: vinsertf128 $1, (%r8), %ymm1, %ymm17448; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2,3],ymm1[4],ymm0[5,6,7]7449; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7450; AVX2-FCP-NEXT: vmovaps (%rdx), %ymm117451; AVX2-FCP-NEXT: vmovaps (%rcx), %ymm157452; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm0 = ymm15[1,1,1,1,5,5,5,5]7453; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm11[1],ymm0[2,3,4],ymm11[5],ymm0[6,7]7454; AVX2-FCP-NEXT: vmovups %ymm11, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7455; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2]7456; AVX2-FCP-NEXT: vmovaps (%rdi), %ymm27457; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7458; AVX2-FCP-NEXT: vmovaps (%rsi), %ymm57459; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm1 = ymm5[1,1,2,2,5,5,6,6]7460; AVX2-FCP-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7461; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm2[1,2],ymm1[3,4],ymm2[5,6],ymm1[7]7462; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,2,2,3]7463; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4],ymm1[5,6,7]7464; AVX2-FCP-NEXT: vbroadcastsd 16(%r8), %ymm17465; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2,3,4],ymm1[5],ymm0[6,7]7466; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7467; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm0 = ymm15[0,2,3,3,4,6,7,7]7468; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm1 = ymm11[2,3,2,3,6,7,6,7]7469; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2],ymm1[3,4],ymm0[5,6],ymm1[7]7470; AVX2-FCP-NEXT: vunpckhps {{.*#+}} ymm1 = ymm5[2],ymm2[2],ymm5[3],ymm2[3],ymm5[6],ymm2[6],ymm5[7],ymm2[7]7471; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[3,3,3,3]7472; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,1,3,3]7473; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3,4],ymm0[5,6,7]7474; AVX2-FCP-NEXT: vbroadcastsd 24(%r8), %ymm17475; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4,5,6],ymm1[7]7476; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7477; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm0 = xmm10[0],xmm12[0],xmm10[1],xmm12[1]7478; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm1 = xmm7[0],xmm13[0],xmm7[1],xmm13[1]7479; AVX2-FCP-NEXT: vpermps %ymm0, %ymm9, %ymm07480; AVX2-FCP-NEXT: vpermps %ymm1, %ymm4, %ymm57481; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm5[0,1],ymm0[2,3],ymm5[4,5,6],ymm0[7]7482; AVX2-FCP-NEXT: vinsertf128 $1, 32(%r8), %ymm1, %ymm17483; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2,3],ymm1[4],ymm0[5,6,7]7484; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7485; AVX2-FCP-NEXT: vmovaps 32(%rdx), %ymm107486; AVX2-FCP-NEXT: vmovaps 32(%rcx), %ymm27487; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm0 = ymm2[1,1,1,1,5,5,5,5]7488; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7489; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm10[1],ymm0[2,3,4],ymm10[5],ymm0[6,7]7490; AVX2-FCP-NEXT: vmovups %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7491; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2]7492; AVX2-FCP-NEXT: vmovaps 32(%rdi), %ymm57493; AVX2-FCP-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7494; AVX2-FCP-NEXT: vmovaps 32(%rsi), %ymm77495; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm1 = ymm7[1,1,2,2,5,5,6,6]7496; AVX2-FCP-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7497; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm5[1,2],ymm1[3,4],ymm5[5,6],ymm1[7]7498; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,2,2,3]7499; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4],ymm1[5,6,7]7500; AVX2-FCP-NEXT: vbroadcastsd 48(%r8), %ymm17501; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2,3,4],ymm1[5],ymm0[6,7]7502; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7503; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm0 = ymm2[0,2,3,3,4,6,7,7]7504; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm1 = ymm10[2,3,2,3,6,7,6,7]7505; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2],ymm1[3,4],ymm0[5,6],ymm1[7]7506; AVX2-FCP-NEXT: vunpckhps {{.*#+}} ymm1 = ymm7[2],ymm5[2],ymm7[3],ymm5[3],ymm7[6],ymm5[6],ymm7[7],ymm5[7]7507; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[3,3,3,3]7508; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,1,3,3]7509; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3,4],ymm0[5,6,7]7510; AVX2-FCP-NEXT: vbroadcastsd 56(%r8), %ymm17511; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4,5,6],ymm1[7]7512; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7513; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm0 = xmm8[0],xmm6[0],xmm8[1],xmm6[1]7514; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload7515; AVX2-FCP-NEXT: vunpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload7516; AVX2-FCP-NEXT: # xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]7517; AVX2-FCP-NEXT: vpermps %ymm0, %ymm9, %ymm07518; AVX2-FCP-NEXT: vpermps %ymm1, %ymm4, %ymm57519; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm5[0,1],ymm0[2,3],ymm5[4,5,6],ymm0[7]7520; AVX2-FCP-NEXT: vinsertf128 $1, 64(%r8), %ymm1, %ymm17521; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2,3],ymm1[4],ymm0[5,6,7]7522; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7523; AVX2-FCP-NEXT: vmovaps 64(%rdx), %ymm17524; AVX2-FCP-NEXT: vmovaps 64(%rcx), %ymm57525; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm0 = ymm5[1,1,1,1,5,5,5,5]7526; AVX2-FCP-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7527; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2,3,4],ymm1[5],ymm0[6,7]7528; AVX2-FCP-NEXT: vmovaps %ymm1, %ymm27529; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7530; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2]7531; AVX2-FCP-NEXT: vmovaps 64(%rdi), %ymm67532; AVX2-FCP-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7533; AVX2-FCP-NEXT: vmovaps 64(%rsi), %ymm77534; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm1 = ymm7[1,1,2,2,5,5,6,6]7535; AVX2-FCP-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7536; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm6[1,2],ymm1[3,4],ymm6[5,6],ymm1[7]7537; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,2,2,3]7538; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4],ymm1[5,6,7]7539; AVX2-FCP-NEXT: vbroadcastsd 80(%r8), %ymm17540; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2,3,4],ymm1[5],ymm0[6,7]7541; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7542; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm0 = ymm5[0,2,3,3,4,6,7,7]7543; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm1 = ymm2[2,3,2,3,6,7,6,7]7544; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2],ymm1[3,4],ymm0[5,6],ymm1[7]7545; AVX2-FCP-NEXT: vunpckhps {{.*#+}} ymm1 = ymm7[2],ymm6[2],ymm7[3],ymm6[3],ymm7[6],ymm6[6],ymm7[7],ymm6[7]7546; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[3,3,3,3]7547; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,1,3,3]7548; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3,4],ymm0[5,6,7]7549; AVX2-FCP-NEXT: vbroadcastsd 88(%r8), %ymm17550; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4,5,6],ymm1[7]7551; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7552; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7553; AVX2-FCP-NEXT: vunpcklps {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1]7554; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload7555; AVX2-FCP-NEXT: vunpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload7556; AVX2-FCP-NEXT: # xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]7557; AVX2-FCP-NEXT: vpermps %ymm0, %ymm9, %ymm07558; AVX2-FCP-NEXT: vpermps %ymm1, %ymm4, %ymm27559; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1],ymm0[2,3],ymm2[4,5,6],ymm0[7]7560; AVX2-FCP-NEXT: vinsertf128 $1, 96(%r8), %ymm1, %ymm17561; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2,3],ymm1[4],ymm0[5,6,7]7562; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7563; AVX2-FCP-NEXT: vmovaps 96(%rdx), %ymm57564; AVX2-FCP-NEXT: vmovaps 96(%rcx), %ymm27565; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm0 = ymm2[1,1,1,1,5,5,5,5]7566; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7567; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm5[1],ymm0[2,3,4],ymm5[5],ymm0[6,7]7568; AVX2-FCP-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7569; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2]7570; AVX2-FCP-NEXT: vmovaps 96(%rdi), %ymm37571; AVX2-FCP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7572; AVX2-FCP-NEXT: vmovaps 96(%rsi), %ymm147573; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm1 = ymm14[1,1,2,2,5,5,6,6]7574; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm3[1,2],ymm1[3,4],ymm3[5,6],ymm1[7]7575; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,2,2,3]7576; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4],ymm1[5,6,7]7577; AVX2-FCP-NEXT: vbroadcastsd 112(%r8), %ymm17578; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2,3,4],ymm1[5],ymm0[6,7]7579; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7580; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm0 = ymm2[0,2,3,3,4,6,7,7]7581; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm1 = ymm5[2,3,2,3,6,7,6,7]7582; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2],ymm1[3,4],ymm0[5,6],ymm1[7]7583; AVX2-FCP-NEXT: vunpckhps {{.*#+}} ymm1 = ymm14[2],ymm3[2],ymm14[3],ymm3[3],ymm14[6],ymm3[6],ymm14[7],ymm3[7]7584; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[3,3,3,3]7585; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,1,3,3]7586; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3,4],ymm0[5,6,7]7587; AVX2-FCP-NEXT: vbroadcastsd 120(%r8), %ymm17588; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4,5,6],ymm1[7]7589; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7590; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7591; AVX2-FCP-NEXT: vunpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload7592; AVX2-FCP-NEXT: # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]7593; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload7594; AVX2-FCP-NEXT: vunpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload7595; AVX2-FCP-NEXT: # xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]7596; AVX2-FCP-NEXT: vpermps %ymm0, %ymm9, %ymm07597; AVX2-FCP-NEXT: vpermps %ymm1, %ymm4, %ymm27598; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1],ymm0[2,3],ymm2[4,5,6],ymm0[7]7599; AVX2-FCP-NEXT: vinsertf128 $1, 128(%r8), %ymm1, %ymm17600; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2,3],ymm1[4],ymm0[5,6,7]7601; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7602; AVX2-FCP-NEXT: vmovaps 128(%rdx), %ymm57603; AVX2-FCP-NEXT: vmovaps 128(%rcx), %ymm27604; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm0 = ymm2[1,1,1,1,5,5,5,5]7605; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7606; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm5[1],ymm0[2,3,4],ymm5[5],ymm0[6,7]7607; AVX2-FCP-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7608; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2]7609; AVX2-FCP-NEXT: vmovaps 128(%rdi), %ymm37610; AVX2-FCP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7611; AVX2-FCP-NEXT: vmovaps 128(%rsi), %ymm137612; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm1 = ymm13[1,1,2,2,5,5,6,6]7613; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm3[1,2],ymm1[3,4],ymm3[5,6],ymm1[7]7614; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,2,2,3]7615; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4],ymm1[5,6,7]7616; AVX2-FCP-NEXT: vbroadcastsd 144(%r8), %ymm17617; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2,3,4],ymm1[5],ymm0[6,7]7618; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7619; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm0 = ymm2[0,2,3,3,4,6,7,7]7620; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm1 = ymm5[2,3,2,3,6,7,6,7]7621; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2],ymm1[3,4],ymm0[5,6],ymm1[7]7622; AVX2-FCP-NEXT: vunpckhps {{.*#+}} ymm1 = ymm13[2],ymm3[2],ymm13[3],ymm3[3],ymm13[6],ymm3[6],ymm13[7],ymm3[7]7623; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[3,3,3,3]7624; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,1,3,3]7625; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3,4],ymm0[5,6,7]7626; AVX2-FCP-NEXT: vbroadcastsd 152(%r8), %ymm17627; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4,5,6],ymm1[7]7628; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7629; AVX2-FCP-NEXT: vmovaps (%rsp), %xmm0 # 16-byte Reload7630; AVX2-FCP-NEXT: vunpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload7631; AVX2-FCP-NEXT: # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]7632; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload7633; AVX2-FCP-NEXT: vunpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload7634; AVX2-FCP-NEXT: # xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]7635; AVX2-FCP-NEXT: vpermps %ymm0, %ymm9, %ymm07636; AVX2-FCP-NEXT: vpermps %ymm1, %ymm4, %ymm27637; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1],ymm0[2,3],ymm2[4,5,6],ymm0[7]7638; AVX2-FCP-NEXT: vinsertf128 $1, 160(%r8), %ymm1, %ymm17639; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2,3],ymm1[4],ymm0[5,6,7]7640; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7641; AVX2-FCP-NEXT: vmovaps 160(%rdx), %ymm17642; AVX2-FCP-NEXT: vmovaps 160(%rcx), %ymm37643; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm0 = ymm3[1,1,1,1,5,5,5,5]7644; AVX2-FCP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7645; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2,3,4],ymm1[5],ymm0[6,7]7646; AVX2-FCP-NEXT: vmovaps %ymm1, %ymm27647; AVX2-FCP-NEXT: vmovups %ymm1, (%rsp) # 32-byte Spill7648; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2]7649; AVX2-FCP-NEXT: vmovaps 160(%rdi), %ymm117650; AVX2-FCP-NEXT: vmovaps 160(%rsi), %ymm107651; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm1 = ymm10[1,1,2,2,5,5,6,6]7652; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm11[1,2],ymm1[3,4],ymm11[5,6],ymm1[7]7653; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,2,2,3]7654; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4],ymm1[5,6,7]7655; AVX2-FCP-NEXT: vbroadcastsd 176(%r8), %ymm17656; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2,3,4],ymm1[5],ymm0[6,7]7657; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7658; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm0 = ymm3[0,2,3,3,4,6,7,7]7659; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm1 = ymm2[2,3,2,3,6,7,6,7]7660; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2],ymm1[3,4],ymm0[5,6],ymm1[7]7661; AVX2-FCP-NEXT: vunpckhps {{.*#+}} ymm1 = ymm10[2],ymm11[2],ymm10[3],ymm11[3],ymm10[6],ymm11[6],ymm10[7],ymm11[7]7662; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[3,3,3,3]7663; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,1,3,3]7664; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3,4],ymm0[5,6,7]7665; AVX2-FCP-NEXT: vbroadcastsd 184(%r8), %ymm17666; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4,5,6],ymm1[7]7667; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7668; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7669; AVX2-FCP-NEXT: vunpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload7670; AVX2-FCP-NEXT: # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]7671; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload7672; AVX2-FCP-NEXT: vunpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload7673; AVX2-FCP-NEXT: # xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]7674; AVX2-FCP-NEXT: vpermps %ymm0, %ymm9, %ymm07675; AVX2-FCP-NEXT: vpermps %ymm1, %ymm4, %ymm27676; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1],ymm0[2,3],ymm2[4,5,6],ymm0[7]7677; AVX2-FCP-NEXT: vinsertf128 $1, 192(%r8), %ymm1, %ymm17678; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2,3],ymm1[4],ymm0[5,6,7]7679; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7680; AVX2-FCP-NEXT: vmovaps 192(%rdx), %ymm87681; AVX2-FCP-NEXT: vmovaps 192(%rcx), %ymm77682; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm0 = ymm7[1,1,1,1,5,5,5,5]7683; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm8[1],ymm0[2,3,4],ymm8[5],ymm0[6,7]7684; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,2,2,2]7685; AVX2-FCP-NEXT: vmovaps 192(%rdi), %ymm67686; AVX2-FCP-NEXT: vmovaps 192(%rsi), %ymm57687; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm1 = ymm5[1,1,2,2,5,5,6,6]7688; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm1[0],ymm6[1,2],ymm1[3,4],ymm6[5,6],ymm1[7]7689; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[2,2,2,3]7690; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3,4],ymm1[5,6,7]7691; AVX2-FCP-NEXT: vbroadcastsd 208(%r8), %ymm17692; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2,3,4],ymm1[5],ymm0[6,7]7693; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7694; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm0 = ymm7[0,2,3,3,4,6,7,7]7695; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm1 = ymm8[2,3,2,3,6,7,6,7]7696; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2],ymm1[3,4],ymm0[5,6],ymm1[7]7697; AVX2-FCP-NEXT: vunpckhps {{.*#+}} ymm1 = ymm5[2],ymm6[2],ymm5[3],ymm6[3],ymm5[6],ymm6[6],ymm5[7],ymm6[7]7698; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm1 = ymm1[3,3,3,3]7699; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[2,1,3,3]7700; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2],ymm1[3,4],ymm0[5,6,7]7701; AVX2-FCP-NEXT: vbroadcastsd 216(%r8), %ymm17702; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4,5,6],ymm1[7]7703; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7704; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7705; AVX2-FCP-NEXT: vunpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload7706; AVX2-FCP-NEXT: # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]7707; AVX2-FCP-NEXT: vpermps %ymm0, %ymm9, %ymm07708; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload7709; AVX2-FCP-NEXT: vunpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload7710; AVX2-FCP-NEXT: # xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]7711; AVX2-FCP-NEXT: vpermps %ymm1, %ymm4, %ymm27712; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm2[0,1],ymm0[2,3],ymm2[4,5,6],ymm0[7]7713; AVX2-FCP-NEXT: vinsertf128 $1, 224(%r8), %ymm1, %ymm17714; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2,3],ymm1[4],ymm0[5,6,7]7715; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7716; AVX2-FCP-NEXT: vmovaps 224(%rdx), %ymm37717; AVX2-FCP-NEXT: vmovaps 224(%rcx), %ymm27718; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm0 = ymm2[1,1,1,1,5,5,5,5]7719; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm3[1],ymm0[2,3,4],ymm3[5],ymm0[6,7]7720; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm4 = ymm0[2,2,2,2]7721; AVX2-FCP-NEXT: vmovaps 224(%rdi), %ymm17722; AVX2-FCP-NEXT: vmovaps 224(%rsi), %ymm07723; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm9 = ymm0[1,1,2,2,5,5,6,6]7724; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm9 = ymm9[0],ymm1[1,2],ymm9[3,4],ymm1[5,6],ymm9[7]7725; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm9 = ymm9[2,2,2,3]7726; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm9[0,1,2],ymm4[3,4],ymm9[5,6,7]7727; AVX2-FCP-NEXT: vbroadcastsd 240(%r8), %ymm97728; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm9[0],ymm4[1,2,3,4],ymm9[5],ymm4[6,7]7729; AVX2-FCP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7730; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm4 = ymm2[0,2,3,3,4,6,7,7]7731; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm9 = ymm3[2,3,2,3,6,7,6,7]7732; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm9[0],ymm4[1,2],ymm9[3,4],ymm4[5,6],ymm9[7]7733; AVX2-FCP-NEXT: vunpckhps {{.*#+}} ymm9 = ymm0[2],ymm1[2],ymm0[3],ymm1[3],ymm0[6],ymm1[6],ymm0[7],ymm1[7]7734; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm9 = ymm9[3,3,3,3]7735; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm4 = ymm4[2,1,3,3]7736; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1,2],ymm9[3,4],ymm4[5,6,7]7737; AVX2-FCP-NEXT: vbroadcastsd 248(%r8), %ymm97738; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0,1],ymm9[2],ymm4[3,4,5,6],ymm9[7]7739; AVX2-FCP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7740; AVX2-FCP-NEXT: vpermilps $227, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Folded Reload7741; AVX2-FCP-NEXT: # ymm4 = mem[3,0,2,3,7,4,6,7]7742; AVX2-FCP-NEXT: vblendps $16, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4, %ymm4 # 32-byte Folded Reload7743; AVX2-FCP-NEXT: # ymm4 = ymm4[0,1,2,3],mem[4],ymm4[5,6,7]7744; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm9 = ymm15[0,1,3,0,4,5,7,4]7745; AVX2-FCP-NEXT: vpermilps $78, {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Folded Reload7746; AVX2-FCP-NEXT: # ymm12 = mem[2,3,0,1,6,7,4,5]7747; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm9 = ymm12[0,1],ymm9[2],ymm12[3,4,5,6],ymm9[7]7748; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0],ymm9[1,2,3],ymm4[4,5],ymm9[6,7]7749; AVX2-FCP-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4, %ymm4 # 32-byte Folded Reload7750; AVX2-FCP-NEXT: # ymm4 = ymm4[0,1,2],mem[3],ymm4[4,5,6,7]7751; AVX2-FCP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7752; AVX2-FCP-NEXT: vpermilps $227, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Folded Reload7753; AVX2-FCP-NEXT: # ymm4 = mem[3,0,2,3,7,4,6,7]7754; AVX2-FCP-NEXT: vblendps $16, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4, %ymm4 # 32-byte Folded Reload7755; AVX2-FCP-NEXT: # ymm4 = ymm4[0,1,2,3],mem[4],ymm4[5,6,7]7756; AVX2-FCP-NEXT: vpermilps $52, {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Folded Reload7757; AVX2-FCP-NEXT: # ymm9 = mem[0,1,3,0,4,5,7,4]7758; AVX2-FCP-NEXT: vpermilps $78, {{[-0-9]+}}(%r{{[sb]}}p), %ymm12 # 32-byte Folded Reload7759; AVX2-FCP-NEXT: # ymm12 = mem[2,3,0,1,6,7,4,5]7760; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm9 = ymm12[0,1],ymm9[2],ymm12[3,4,5,6],ymm9[7]7761; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0],ymm9[1,2,3],ymm4[4,5],ymm9[6,7]7762; AVX2-FCP-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4, %ymm12 # 32-byte Folded Reload7763; AVX2-FCP-NEXT: # ymm12 = ymm4[0,1,2],mem[3],ymm4[4,5,6,7]7764; AVX2-FCP-NEXT: vpermilps $227, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Folded Reload7765; AVX2-FCP-NEXT: # ymm4 = mem[3,0,2,3,7,4,6,7]7766; AVX2-FCP-NEXT: vblendps $16, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4, %ymm4 # 32-byte Folded Reload7767; AVX2-FCP-NEXT: # ymm4 = ymm4[0,1,2,3],mem[4],ymm4[5,6,7]7768; AVX2-FCP-NEXT: vpermilps $52, {{[-0-9]+}}(%r{{[sb]}}p), %ymm9 # 32-byte Folded Reload7769; AVX2-FCP-NEXT: # ymm9 = mem[0,1,3,0,4,5,7,4]7770; AVX2-FCP-NEXT: vpermilps $78, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Folded Reload7771; AVX2-FCP-NEXT: # ymm15 = mem[2,3,0,1,6,7,4,5]7772; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm9 = ymm15[0,1],ymm9[2],ymm15[3,4,5,6],ymm9[7]7773; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0],ymm9[1,2,3],ymm4[4,5],ymm9[6,7]7774; AVX2-FCP-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4, %ymm9 # 32-byte Folded Reload7775; AVX2-FCP-NEXT: # ymm9 = ymm4[0,1,2],mem[3],ymm4[4,5,6,7]7776; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm4 = ymm14[3,0,2,3,7,4,6,7]7777; AVX2-FCP-NEXT: vblendps $16, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4, %ymm4 # 32-byte Folded Reload7778; AVX2-FCP-NEXT: # ymm4 = ymm4[0,1,2,3],mem[4],ymm4[5,6,7]7779; AVX2-FCP-NEXT: vpermilps $52, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Folded Reload7780; AVX2-FCP-NEXT: # ymm15 = mem[0,1,3,0,4,5,7,4]7781; AVX2-FCP-NEXT: vpermilps $78, {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Folded Reload7782; AVX2-FCP-NEXT: # ymm14 = mem[2,3,0,1,6,7,4,5]7783; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm14 = ymm14[0,1],ymm15[2],ymm14[3,4,5,6],ymm15[7]7784; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm4 = ymm4[0],ymm14[1,2,3],ymm4[4,5],ymm14[6,7]7785; AVX2-FCP-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4, %ymm4 # 32-byte Folded Reload7786; AVX2-FCP-NEXT: # ymm4 = ymm4[0,1,2],mem[3],ymm4[4,5,6,7]7787; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm14 = ymm13[3,0,2,3,7,4,6,7]7788; AVX2-FCP-NEXT: vblendps $16, {{[-0-9]+}}(%r{{[sb]}}p), %ymm14, %ymm14 # 32-byte Folded Reload7789; AVX2-FCP-NEXT: # ymm14 = ymm14[0,1,2,3],mem[4],ymm14[5,6,7]7790; AVX2-FCP-NEXT: vpermilps $52, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15 # 32-byte Folded Reload7791; AVX2-FCP-NEXT: # ymm15 = mem[0,1,3,0,4,5,7,4]7792; AVX2-FCP-NEXT: vpermilps $78, {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Folded Reload7793; AVX2-FCP-NEXT: # ymm13 = mem[2,3,0,1,6,7,4,5]7794; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm13 = ymm13[0,1],ymm15[2],ymm13[3,4,5,6],ymm15[7]7795; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm13 = ymm14[0],ymm13[1,2,3],ymm14[4,5],ymm13[6,7]7796; AVX2-FCP-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm13 # 32-byte Folded Reload7797; AVX2-FCP-NEXT: # ymm13 = ymm13[0,1,2],mem[3],ymm13[4,5,6,7]7798; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm10 = ymm10[3,0,2,3,7,4,6,7]7799; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm10 = ymm10[0,1,2,3],ymm11[4],ymm10[5,6,7]7800; AVX2-FCP-NEXT: vpermilps $52, {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Folded Reload7801; AVX2-FCP-NEXT: # ymm11 = mem[0,1,3,0,4,5,7,4]7802; AVX2-FCP-NEXT: vpermilps $78, (%rsp), %ymm14 # 32-byte Folded Reload7803; AVX2-FCP-NEXT: # ymm14 = mem[2,3,0,1,6,7,4,5]7804; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm11 = ymm14[0,1],ymm11[2],ymm14[3,4,5,6],ymm11[7]7805; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm10 = ymm10[0],ymm11[1,2,3],ymm10[4,5],ymm11[6,7]7806; AVX2-FCP-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm10, %ymm10 # 32-byte Folded Reload7807; AVX2-FCP-NEXT: # ymm10 = ymm10[0,1,2],mem[3],ymm10[4,5,6,7]7808; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm5 = ymm5[3,0,2,3,7,4,6,7]7809; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm6[4],ymm5[5,6,7]7810; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm6 = ymm7[0,1,3,0,4,5,7,4]7811; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm7 = ymm8[2,3,0,1,6,7,4,5]7812; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm6 = ymm7[0,1],ymm6[2],ymm7[3,4,5,6],ymm6[7]7813; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm5 = ymm5[0],ymm6[1,2,3],ymm5[4,5],ymm6[6,7]7814; AVX2-FCP-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm5 # 32-byte Folded Reload7815; AVX2-FCP-NEXT: # ymm5 = ymm5[0,1,2],mem[3],ymm5[4,5,6,7]7816; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm0 = ymm0[3,0,2,3,7,4,6,7]7817; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4],ymm0[5,6,7]7818; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm1 = ymm2[0,1,3,0,4,5,7,4]7819; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm2 = ymm3[2,3,0,1,6,7,4,5]7820; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1],ymm1[2],ymm2[3,4,5,6],ymm1[7]7821; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0],ymm1[1,2,3],ymm0[4,5],ymm1[6,7]7822; AVX2-FCP-NEXT: vblendps $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload7823; AVX2-FCP-NEXT: # ymm0 = ymm0[0,1,2],mem[3],ymm0[4,5,6,7]7824; AVX2-FCP-NEXT: vmovaps %ymm0, 1184(%r9)7825; AVX2-FCP-NEXT: vmovaps %ymm5, 1024(%r9)7826; AVX2-FCP-NEXT: vmovaps %ymm10, 864(%r9)7827; AVX2-FCP-NEXT: vmovaps %ymm13, 704(%r9)7828; AVX2-FCP-NEXT: vmovaps %ymm4, 544(%r9)7829; AVX2-FCP-NEXT: vmovaps %ymm9, 384(%r9)7830; AVX2-FCP-NEXT: vmovaps %ymm12, 224(%r9)7831; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload7832; AVX2-FCP-NEXT: vmovaps %ymm0, 64(%r9)7833; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload7834; AVX2-FCP-NEXT: vmovaps %ymm0, 1248(%r9)7835; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload7836; AVX2-FCP-NEXT: vmovaps %ymm0, 1216(%r9)7837; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload7838; AVX2-FCP-NEXT: vmovaps %ymm0, 1120(%r9)7839; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload7840; AVX2-FCP-NEXT: vmovaps %ymm0, 1088(%r9)7841; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload7842; AVX2-FCP-NEXT: vmovaps %ymm0, 1056(%r9)7843; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload7844; AVX2-FCP-NEXT: vmovaps %ymm0, 960(%r9)7845; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload7846; AVX2-FCP-NEXT: vmovaps %ymm0, 928(%r9)7847; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload7848; AVX2-FCP-NEXT: vmovaps %ymm0, 896(%r9)7849; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload7850; AVX2-FCP-NEXT: vmovaps %ymm0, 800(%r9)7851; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload7852; AVX2-FCP-NEXT: vmovaps %ymm0, 768(%r9)7853; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload7854; AVX2-FCP-NEXT: vmovaps %ymm0, 736(%r9)7855; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload7856; AVX2-FCP-NEXT: vmovaps %ymm0, 640(%r9)7857; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload7858; AVX2-FCP-NEXT: vmovaps %ymm0, 608(%r9)7859; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload7860; AVX2-FCP-NEXT: vmovaps %ymm0, 576(%r9)7861; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload7862; AVX2-FCP-NEXT: vmovaps %ymm0, 480(%r9)7863; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload7864; AVX2-FCP-NEXT: vmovaps %ymm0, 448(%r9)7865; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload7866; AVX2-FCP-NEXT: vmovaps %ymm0, 416(%r9)7867; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload7868; AVX2-FCP-NEXT: vmovaps %ymm0, 320(%r9)7869; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload7870; AVX2-FCP-NEXT: vmovaps %ymm0, 288(%r9)7871; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload7872; AVX2-FCP-NEXT: vmovaps %ymm0, 256(%r9)7873; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload7874; AVX2-FCP-NEXT: vmovaps %ymm0, 160(%r9)7875; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload7876; AVX2-FCP-NEXT: vmovaps %ymm0, 128(%r9)7877; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload7878; AVX2-FCP-NEXT: vmovaps %ymm0, 96(%r9)7879; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload7880; AVX2-FCP-NEXT: vmovaps %ymm0, (%r9)7881; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload7882; AVX2-FCP-NEXT: vmovaps %ymm0, 1152(%r9)7883; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload7884; AVX2-FCP-NEXT: vmovaps %ymm0, 992(%r9)7885; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload7886; AVX2-FCP-NEXT: vmovaps %ymm0, 832(%r9)7887; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload7888; AVX2-FCP-NEXT: vmovaps %ymm0, 672(%r9)7889; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload7890; AVX2-FCP-NEXT: vmovaps %ymm0, 512(%r9)7891; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload7892; AVX2-FCP-NEXT: vmovaps %ymm0, 352(%r9)7893; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload7894; AVX2-FCP-NEXT: vmovaps %ymm0, 192(%r9)7895; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload7896; AVX2-FCP-NEXT: vmovaps %ymm0, 32(%r9)7897; AVX2-FCP-NEXT: addq $1800, %rsp # imm = 0x7087898; AVX2-FCP-NEXT: vzeroupper7899; AVX2-FCP-NEXT: retq7900;7901; AVX512-LABEL: store_i32_stride5_vf64:7902; AVX512: # %bb.0:7903; AVX512-NEXT: subq $712, %rsp # imm = 0x2C87904; AVX512-NEXT: vmovdqa64 128(%rdi), %zmm197905; AVX512-NEXT: vmovdqa64 192(%rdi), %zmm237906; AVX512-NEXT: vmovdqa64 192(%rsi), %zmm187907; AVX512-NEXT: vmovdqa64 (%rdx), %zmm17908; AVX512-NEXT: vmovdqa64 64(%rdx), %zmm37909; AVX512-NEXT: vmovdqa64 128(%rdx), %zmm297910; AVX512-NEXT: vmovdqa64 192(%rdx), %zmm157911; AVX512-NEXT: vmovdqa64 (%rcx), %zmm217912; AVX512-NEXT: vmovdqa64 64(%rcx), %zmm257913; AVX512-NEXT: vmovdqa64 128(%rcx), %zmm247914; AVX512-NEXT: vmovdqa64 192(%rcx), %zmm07915; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm28 = [17,0,0,16,2,18,0,1,17,0,0,16,2,18,0,1]7916; AVX512-NEXT: # zmm28 = mem[0,1,2,3,0,1,2,3]7917; AVX512-NEXT: vmovdqa64 %zmm1, %zmm27918; AVX512-NEXT: vpermt2d %zmm21, %zmm28, %zmm27919; AVX512-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7920; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [0,16,2,18,0,1,17,3,0,16,2,18,0,1,17,3]7921; AVX512-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]7922; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm2 = [14,30,0,13,29,15,31,0,14,30,0,13,29,15,31,0]7923; AVX512-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3]7924; AVX512-NEXT: vmovdqa64 %zmm15, %zmm47925; AVX512-NEXT: vpermt2d %zmm0, %zmm2, %zmm47926; AVX512-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7927; AVX512-NEXT: vmovdqa64 %zmm2, %zmm47928; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm17 = [0,13,29,15,31,0,14,30,0,13,29,15,31,0,14,30]7929; AVX512-NEXT: # zmm17 = mem[0,1,2,3,0,1,2,3]7930; AVX512-NEXT: vmovdqa64 %zmm23, %zmm27931; AVX512-NEXT: vpermt2d %zmm18, %zmm17, %zmm27932; AVX512-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7933; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm2 = [27,0,10,26,12,28,0,11,27,0,10,26,12,28,0,11]7934; AVX512-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3]7935; AVX512-NEXT: vmovdqa64 %zmm23, %zmm57936; AVX512-NEXT: vpermt2d %zmm18, %zmm2, %zmm57937; AVX512-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7938; AVX512-NEXT: vmovdqa64 %zmm2, %zmm57939; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm14 = [9,27,11,0,26,10,28,12,9,27,11,0,26,10,28,12]7940; AVX512-NEXT: # zmm14 = mem[0,1,2,3,0,1,2,3]7941; AVX512-NEXT: vmovdqa64 %zmm0, %zmm317942; AVX512-NEXT: vpermt2d %zmm15, %zmm14, %zmm317943; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm13 = [6,22,8,24,0,7,23,9,6,22,8,24,0,7,23,9]7944; AVX512-NEXT: # zmm13 = mem[0,1,2,3,0,1,2,3]7945; AVX512-NEXT: vmovdqa64 %zmm15, %zmm227946; AVX512-NEXT: vpermt2d %zmm0, %zmm13, %zmm227947; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm30 = [0,3,19,5,21,0,4,20,0,3,19,5,21,0,4,20]7948; AVX512-NEXT: # zmm30 = mem[0,1,2,3,0,1,2,3]7949; AVX512-NEXT: vmovdqa64 %zmm15, %zmm27950; AVX512-NEXT: vpermt2d %zmm0, %zmm30, %zmm27951; AVX512-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7952; AVX512-NEXT: vpermt2d %zmm0, %zmm28, %zmm157953; AVX512-NEXT: vmovdqa64 %zmm29, %zmm07954; AVX512-NEXT: vpermt2d %zmm24, %zmm4, %zmm07955; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7956; AVX512-NEXT: vmovdqa64 %zmm24, %zmm207957; AVX512-NEXT: vpermt2d %zmm29, %zmm14, %zmm207958; AVX512-NEXT: vmovdqa64 %zmm29, %zmm277959; AVX512-NEXT: vpermt2d %zmm24, %zmm13, %zmm277960; AVX512-NEXT: vmovdqa64 %zmm29, %zmm07961; AVX512-NEXT: vpermt2d %zmm24, %zmm30, %zmm07962; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7963; AVX512-NEXT: vpermt2d %zmm24, %zmm28, %zmm297964; AVX512-NEXT: vmovdqa64 %zmm3, %zmm07965; AVX512-NEXT: vmovdqa64 %zmm25, %zmm247966; AVX512-NEXT: vpermt2d %zmm3, %zmm14, %zmm247967; AVX512-NEXT: vmovdqa64 %zmm3, %zmm267968; AVX512-NEXT: vmovdqa64 %zmm3, %zmm27969; AVX512-NEXT: vpermt2d %zmm25, %zmm28, %zmm37970; AVX512-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7971; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm28 = [8,24,0,7,23,9,25,0,8,24,0,7,23,9,25,0]7972; AVX512-NEXT: # zmm28 = mem[0,1,2,3,0,1,2,3]7973; AVX512-NEXT: vpermt2d %zmm25, %zmm4, %zmm07974; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7975; AVX512-NEXT: vpermt2d %zmm25, %zmm13, %zmm267976; AVX512-NEXT: vpermt2d %zmm25, %zmm30, %zmm27977; AVX512-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7978; AVX512-NEXT: vmovdqa64 %zmm23, %zmm167979; AVX512-NEXT: vpermt2d %zmm18, %zmm28, %zmm167980; AVX512-NEXT: vpermi2d %zmm21, %zmm1, %zmm47981; AVX512-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill7982; AVX512-NEXT: vpermi2d %zmm1, %zmm21, %zmm147983; AVX512-NEXT: vpermi2d %zmm21, %zmm1, %zmm137984; AVX512-NEXT: vpermt2d %zmm21, %zmm30, %zmm17985; AVX512-NEXT: vmovdqu64 %zmm1, (%rsp) # 64-byte Spill7986; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm2 = [3,21,5,0,20,4,22,6,3,21,5,0,20,4,22,6]7987; AVX512-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3]7988; AVX512-NEXT: vmovdqa64 %zmm18, %zmm217989; AVX512-NEXT: vpermt2d %zmm23, %zmm2, %zmm217990; AVX512-NEXT: vpermt2d %zmm18, %zmm7, %zmm237991; AVX512-NEXT: vmovdqa64 128(%rsi), %zmm07992; AVX512-NEXT: vmovdqa64 %zmm19, %zmm187993; AVX512-NEXT: vpermt2d %zmm0, %zmm17, %zmm187994; AVX512-NEXT: vmovdqa64 %zmm19, %zmm117995; AVX512-NEXT: vpermt2d %zmm0, %zmm5, %zmm117996; AVX512-NEXT: vmovdqa64 %zmm19, %zmm127997; AVX512-NEXT: vpermt2d %zmm0, %zmm28, %zmm127998; AVX512-NEXT: vmovdqa64 %zmm0, %zmm307999; AVX512-NEXT: vpermt2d %zmm19, %zmm2, %zmm308000; AVX512-NEXT: vpermt2d %zmm0, %zmm7, %zmm198001; AVX512-NEXT: vmovdqa64 (%rdi), %zmm258002; AVX512-NEXT: vmovdqa64 (%rsi), %zmm18003; AVX512-NEXT: vmovdqa64 %zmm25, %zmm88004; AVX512-NEXT: vpermt2d %zmm1, %zmm7, %zmm88005; AVX512-NEXT: vmovdqa64 64(%rdi), %zmm08006; AVX512-NEXT: vmovdqa64 64(%rsi), %zmm38007; AVX512-NEXT: vmovdqa64 %zmm3, %zmm48008; AVX512-NEXT: vpermt2d %zmm0, %zmm2, %zmm48009; AVX512-NEXT: vmovdqa64 %zmm0, %zmm68010; AVX512-NEXT: vmovdqa64 %zmm0, %zmm108011; AVX512-NEXT: vmovdqa64 %zmm0, %zmm98012; AVX512-NEXT: vpermt2d %zmm3, %zmm7, %zmm08013; AVX512-NEXT: vpermt2d %zmm3, %zmm17, %zmm68014; AVX512-NEXT: vpermt2d %zmm3, %zmm5, %zmm108015; AVX512-NEXT: vpermt2d %zmm3, %zmm28, %zmm98016; AVX512-NEXT: vpermi2d %zmm1, %zmm25, %zmm178017; AVX512-NEXT: vpermi2d %zmm1, %zmm25, %zmm58018; AVX512-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8019; AVX512-NEXT: vpermi2d %zmm1, %zmm25, %zmm288020; AVX512-NEXT: vpermt2d %zmm25, %zmm2, %zmm18021; AVX512-NEXT: movw $12684, %ax # imm = 0x318C8022; AVX512-NEXT: kmovw %eax, %k18023; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload8024; AVX512-NEXT: vmovdqa32 %zmm2, %zmm8 {%k1}8025; AVX512-NEXT: movw $25368, %ax # imm = 0x63188026; AVX512-NEXT: kmovw %eax, %k28027; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload8028; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload8029; AVX512-NEXT: vmovdqa32 %zmm2, %zmm3 {%k2}8030; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload8031; AVX512-NEXT: vmovdqa32 %zmm2, %zmm31 {%k1}8032; AVX512-NEXT: vmovdqa32 %zmm16, %zmm22 {%k2}8033; AVX512-NEXT: movw $6342, %ax # imm = 0x18C68034; AVX512-NEXT: kmovw %eax, %k38035; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload8036; AVX512-NEXT: vmovdqa32 %zmm2, %zmm21 {%k3}8037; AVX512-NEXT: vmovdqa32 %zmm15, %zmm23 {%k1}8038; AVX512-NEXT: vmovdqa64 192(%r8), %zmm28039; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm5 = [28,1,2,3,4,29,6,7,8,9,30,11,12,13,14,31]8040; AVX512-NEXT: vpermt2d %zmm2, %zmm5, %zmm38041; AVX512-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8042; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm15 = [0,25,2,3,4,5,26,7,8,9,10,27,12,13,14,15]8043; AVX512-NEXT: vpermt2d %zmm2, %zmm15, %zmm318044; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm16 = [0,1,22,3,4,5,6,23,8,9,10,11,24,13,14,15]8045; AVX512-NEXT: vpermt2d %zmm2, %zmm16, %zmm228046; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm25 = [0,1,2,19,4,5,6,7,20,9,10,11,12,21,14,15]8047; AVX512-NEXT: vpermt2d %zmm2, %zmm25, %zmm218048; AVX512-NEXT: vpmovsxbd {{.*#+}} zmm3 = [0,1,2,3,16,5,6,7,8,17,10,11,12,13,18,15]8049; AVX512-NEXT: vpermt2d %zmm2, %zmm3, %zmm238050; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload8051; AVX512-NEXT: vmovdqa32 %zmm2, %zmm18 {%k2}8052; AVX512-NEXT: vmovdqa32 %zmm11, %zmm20 {%k1}8053; AVX512-NEXT: vmovdqa32 %zmm12, %zmm27 {%k2}8054; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload8055; AVX512-NEXT: vmovdqa32 %zmm2, %zmm30 {%k3}8056; AVX512-NEXT: vmovdqa32 %zmm29, %zmm19 {%k1}8057; AVX512-NEXT: vmovdqa64 128(%r8), %zmm28058; AVX512-NEXT: vpermt2d %zmm2, %zmm5, %zmm188059; AVX512-NEXT: vpermt2d %zmm2, %zmm15, %zmm208060; AVX512-NEXT: vpermt2d %zmm2, %zmm16, %zmm278061; AVX512-NEXT: vpermt2d %zmm2, %zmm25, %zmm308062; AVX512-NEXT: vpermt2d %zmm2, %zmm3, %zmm198063; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload8064; AVX512-NEXT: vmovdqa32 %zmm2, %zmm6 {%k2}8065; AVX512-NEXT: vmovdqa32 %zmm10, %zmm24 {%k1}8066; AVX512-NEXT: vmovdqa32 %zmm9, %zmm26 {%k2}8067; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload8068; AVX512-NEXT: vmovdqa32 %zmm2, %zmm4 {%k3}8069; AVX512-NEXT: vmovdqa64 (%r8), %zmm28070; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload8071; AVX512-NEXT: vmovdqa32 %zmm7, %zmm0 {%k1}8072; AVX512-NEXT: vmovdqa64 64(%r8), %zmm78073; AVX512-NEXT: vpermt2d %zmm2, %zmm3, %zmm88074; AVX512-NEXT: vpermt2d %zmm7, %zmm3, %zmm08075; AVX512-NEXT: vpermt2d %zmm7, %zmm5, %zmm68076; AVX512-NEXT: vpermt2d %zmm7, %zmm15, %zmm248077; AVX512-NEXT: vpermt2d %zmm7, %zmm16, %zmm268078; AVX512-NEXT: vpermt2d %zmm7, %zmm25, %zmm48079; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload8080; AVX512-NEXT: vmovdqa32 %zmm3, %zmm17 {%k2}8081; AVX512-NEXT: vpermt2d %zmm2, %zmm5, %zmm178082; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload8083; AVX512-NEXT: vmovdqa32 %zmm3, %zmm14 {%k1}8084; AVX512-NEXT: vpermt2d %zmm2, %zmm15, %zmm148085; AVX512-NEXT: vmovdqa32 %zmm28, %zmm13 {%k2}8086; AVX512-NEXT: vpermt2d %zmm2, %zmm16, %zmm138087; AVX512-NEXT: vmovdqu64 (%rsp), %zmm3 # 64-byte Reload8088; AVX512-NEXT: vmovdqa32 %zmm3, %zmm1 {%k3}8089; AVX512-NEXT: vpermt2d %zmm2, %zmm25, %zmm18090; AVX512-NEXT: vmovdqa64 %zmm1, 64(%r9)8091; AVX512-NEXT: vmovdqa64 %zmm13, 128(%r9)8092; AVX512-NEXT: vmovdqa64 %zmm14, 192(%r9)8093; AVX512-NEXT: vmovdqa64 %zmm17, 256(%r9)8094; AVX512-NEXT: vmovdqa64 %zmm0, 320(%r9)8095; AVX512-NEXT: vmovdqa64 %zmm4, 384(%r9)8096; AVX512-NEXT: vmovdqa64 %zmm26, 448(%r9)8097; AVX512-NEXT: vmovdqa64 %zmm24, 512(%r9)8098; AVX512-NEXT: vmovdqa64 %zmm6, 576(%r9)8099; AVX512-NEXT: vmovdqa64 %zmm19, 640(%r9)8100; AVX512-NEXT: vmovdqa64 %zmm30, 704(%r9)8101; AVX512-NEXT: vmovdqa64 %zmm27, 768(%r9)8102; AVX512-NEXT: vmovdqa64 %zmm20, 832(%r9)8103; AVX512-NEXT: vmovdqa64 %zmm18, 896(%r9)8104; AVX512-NEXT: vmovdqa64 %zmm23, 960(%r9)8105; AVX512-NEXT: vmovdqa64 %zmm21, 1024(%r9)8106; AVX512-NEXT: vmovdqa64 %zmm22, 1088(%r9)8107; AVX512-NEXT: vmovdqa64 %zmm31, 1152(%r9)8108; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8109; AVX512-NEXT: vmovaps %zmm0, 1216(%r9)8110; AVX512-NEXT: vmovdqa64 %zmm8, (%r9)8111; AVX512-NEXT: addq $712, %rsp # imm = 0x2C88112; AVX512-NEXT: vzeroupper8113; AVX512-NEXT: retq8114;8115; AVX512-FCP-LABEL: store_i32_stride5_vf64:8116; AVX512-FCP: # %bb.0:8117; AVX512-FCP-NEXT: subq $712, %rsp # imm = 0x2C88118; AVX512-FCP-NEXT: vmovdqa64 128(%rdi), %zmm198119; AVX512-FCP-NEXT: vmovdqa64 192(%rdi), %zmm238120; AVX512-FCP-NEXT: vmovdqa64 192(%rsi), %zmm188121; AVX512-FCP-NEXT: vmovdqa64 (%rdx), %zmm18122; AVX512-FCP-NEXT: vmovdqa64 64(%rdx), %zmm38123; AVX512-FCP-NEXT: vmovdqa64 128(%rdx), %zmm298124; AVX512-FCP-NEXT: vmovdqa64 192(%rdx), %zmm158125; AVX512-FCP-NEXT: vmovdqa64 (%rcx), %zmm218126; AVX512-FCP-NEXT: vmovdqa64 64(%rcx), %zmm258127; AVX512-FCP-NEXT: vmovdqa64 128(%rcx), %zmm248128; AVX512-FCP-NEXT: vmovdqa64 192(%rcx), %zmm08129; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm28 = [17,0,0,16,2,18,0,1,17,0,0,16,2,18,0,1]8130; AVX512-FCP-NEXT: # zmm28 = mem[0,1,2,3,0,1,2,3]8131; AVX512-FCP-NEXT: vmovdqa64 %zmm1, %zmm28132; AVX512-FCP-NEXT: vpermt2d %zmm21, %zmm28, %zmm28133; AVX512-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8134; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [0,16,2,18,0,1,17,3,0,16,2,18,0,1,17,3]8135; AVX512-FCP-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]8136; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm2 = [14,30,0,13,29,15,31,0,14,30,0,13,29,15,31,0]8137; AVX512-FCP-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3]8138; AVX512-FCP-NEXT: vmovdqa64 %zmm15, %zmm48139; AVX512-FCP-NEXT: vpermt2d %zmm0, %zmm2, %zmm48140; AVX512-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8141; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm48142; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm17 = [0,13,29,15,31,0,14,30,0,13,29,15,31,0,14,30]8143; AVX512-FCP-NEXT: # zmm17 = mem[0,1,2,3,0,1,2,3]8144; AVX512-FCP-NEXT: vmovdqa64 %zmm23, %zmm28145; AVX512-FCP-NEXT: vpermt2d %zmm18, %zmm17, %zmm28146; AVX512-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8147; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm2 = [27,0,10,26,12,28,0,11,27,0,10,26,12,28,0,11]8148; AVX512-FCP-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3]8149; AVX512-FCP-NEXT: vmovdqa64 %zmm23, %zmm58150; AVX512-FCP-NEXT: vpermt2d %zmm18, %zmm2, %zmm58151; AVX512-FCP-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8152; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm58153; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm14 = [9,27,11,0,26,10,28,12,9,27,11,0,26,10,28,12]8154; AVX512-FCP-NEXT: # zmm14 = mem[0,1,2,3,0,1,2,3]8155; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm318156; AVX512-FCP-NEXT: vpermt2d %zmm15, %zmm14, %zmm318157; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm13 = [6,22,8,24,0,7,23,9,6,22,8,24,0,7,23,9]8158; AVX512-FCP-NEXT: # zmm13 = mem[0,1,2,3,0,1,2,3]8159; AVX512-FCP-NEXT: vmovdqa64 %zmm15, %zmm228160; AVX512-FCP-NEXT: vpermt2d %zmm0, %zmm13, %zmm228161; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm30 = [0,3,19,5,21,0,4,20,0,3,19,5,21,0,4,20]8162; AVX512-FCP-NEXT: # zmm30 = mem[0,1,2,3,0,1,2,3]8163; AVX512-FCP-NEXT: vmovdqa64 %zmm15, %zmm28164; AVX512-FCP-NEXT: vpermt2d %zmm0, %zmm30, %zmm28165; AVX512-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8166; AVX512-FCP-NEXT: vpermt2d %zmm0, %zmm28, %zmm158167; AVX512-FCP-NEXT: vmovdqa64 %zmm29, %zmm08168; AVX512-FCP-NEXT: vpermt2d %zmm24, %zmm4, %zmm08169; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8170; AVX512-FCP-NEXT: vmovdqa64 %zmm24, %zmm208171; AVX512-FCP-NEXT: vpermt2d %zmm29, %zmm14, %zmm208172; AVX512-FCP-NEXT: vmovdqa64 %zmm29, %zmm278173; AVX512-FCP-NEXT: vpermt2d %zmm24, %zmm13, %zmm278174; AVX512-FCP-NEXT: vmovdqa64 %zmm29, %zmm08175; AVX512-FCP-NEXT: vpermt2d %zmm24, %zmm30, %zmm08176; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8177; AVX512-FCP-NEXT: vpermt2d %zmm24, %zmm28, %zmm298178; AVX512-FCP-NEXT: vmovdqa64 %zmm3, %zmm08179; AVX512-FCP-NEXT: vmovdqa64 %zmm25, %zmm248180; AVX512-FCP-NEXT: vpermt2d %zmm3, %zmm14, %zmm248181; AVX512-FCP-NEXT: vmovdqa64 %zmm3, %zmm268182; AVX512-FCP-NEXT: vmovdqa64 %zmm3, %zmm28183; AVX512-FCP-NEXT: vpermt2d %zmm25, %zmm28, %zmm38184; AVX512-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8185; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm28 = [8,24,0,7,23,9,25,0,8,24,0,7,23,9,25,0]8186; AVX512-FCP-NEXT: # zmm28 = mem[0,1,2,3,0,1,2,3]8187; AVX512-FCP-NEXT: vpermt2d %zmm25, %zmm4, %zmm08188; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8189; AVX512-FCP-NEXT: vpermt2d %zmm25, %zmm13, %zmm268190; AVX512-FCP-NEXT: vpermt2d %zmm25, %zmm30, %zmm28191; AVX512-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8192; AVX512-FCP-NEXT: vmovdqa64 %zmm23, %zmm168193; AVX512-FCP-NEXT: vpermt2d %zmm18, %zmm28, %zmm168194; AVX512-FCP-NEXT: vpermi2d %zmm21, %zmm1, %zmm48195; AVX512-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8196; AVX512-FCP-NEXT: vpermi2d %zmm1, %zmm21, %zmm148197; AVX512-FCP-NEXT: vpermi2d %zmm21, %zmm1, %zmm138198; AVX512-FCP-NEXT: vpermt2d %zmm21, %zmm30, %zmm18199; AVX512-FCP-NEXT: vmovdqu64 %zmm1, (%rsp) # 64-byte Spill8200; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm2 = [3,21,5,0,20,4,22,6,3,21,5,0,20,4,22,6]8201; AVX512-FCP-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3]8202; AVX512-FCP-NEXT: vmovdqa64 %zmm18, %zmm218203; AVX512-FCP-NEXT: vpermt2d %zmm23, %zmm2, %zmm218204; AVX512-FCP-NEXT: vpermt2d %zmm18, %zmm7, %zmm238205; AVX512-FCP-NEXT: vmovdqa64 128(%rsi), %zmm08206; AVX512-FCP-NEXT: vmovdqa64 %zmm19, %zmm188207; AVX512-FCP-NEXT: vpermt2d %zmm0, %zmm17, %zmm188208; AVX512-FCP-NEXT: vmovdqa64 %zmm19, %zmm118209; AVX512-FCP-NEXT: vpermt2d %zmm0, %zmm5, %zmm118210; AVX512-FCP-NEXT: vmovdqa64 %zmm19, %zmm128211; AVX512-FCP-NEXT: vpermt2d %zmm0, %zmm28, %zmm128212; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm308213; AVX512-FCP-NEXT: vpermt2d %zmm19, %zmm2, %zmm308214; AVX512-FCP-NEXT: vpermt2d %zmm0, %zmm7, %zmm198215; AVX512-FCP-NEXT: vmovdqa64 (%rdi), %zmm258216; AVX512-FCP-NEXT: vmovdqa64 (%rsi), %zmm18217; AVX512-FCP-NEXT: vmovdqa64 %zmm25, %zmm88218; AVX512-FCP-NEXT: vpermt2d %zmm1, %zmm7, %zmm88219; AVX512-FCP-NEXT: vmovdqa64 64(%rdi), %zmm08220; AVX512-FCP-NEXT: vmovdqa64 64(%rsi), %zmm38221; AVX512-FCP-NEXT: vmovdqa64 %zmm3, %zmm48222; AVX512-FCP-NEXT: vpermt2d %zmm0, %zmm2, %zmm48223; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm68224; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm108225; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm98226; AVX512-FCP-NEXT: vpermt2d %zmm3, %zmm7, %zmm08227; AVX512-FCP-NEXT: vpermt2d %zmm3, %zmm17, %zmm68228; AVX512-FCP-NEXT: vpermt2d %zmm3, %zmm5, %zmm108229; AVX512-FCP-NEXT: vpermt2d %zmm3, %zmm28, %zmm98230; AVX512-FCP-NEXT: vpermi2d %zmm1, %zmm25, %zmm178231; AVX512-FCP-NEXT: vpermi2d %zmm1, %zmm25, %zmm58232; AVX512-FCP-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8233; AVX512-FCP-NEXT: vpermi2d %zmm1, %zmm25, %zmm288234; AVX512-FCP-NEXT: vpermt2d %zmm25, %zmm2, %zmm18235; AVX512-FCP-NEXT: movw $12684, %ax # imm = 0x318C8236; AVX512-FCP-NEXT: kmovw %eax, %k18237; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload8238; AVX512-FCP-NEXT: vmovdqa32 %zmm2, %zmm8 {%k1}8239; AVX512-FCP-NEXT: movw $25368, %ax # imm = 0x63188240; AVX512-FCP-NEXT: kmovw %eax, %k28241; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload8242; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload8243; AVX512-FCP-NEXT: vmovdqa32 %zmm2, %zmm3 {%k2}8244; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload8245; AVX512-FCP-NEXT: vmovdqa32 %zmm2, %zmm31 {%k1}8246; AVX512-FCP-NEXT: vmovdqa32 %zmm16, %zmm22 {%k2}8247; AVX512-FCP-NEXT: movw $6342, %ax # imm = 0x18C68248; AVX512-FCP-NEXT: kmovw %eax, %k38249; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload8250; AVX512-FCP-NEXT: vmovdqa32 %zmm2, %zmm21 {%k3}8251; AVX512-FCP-NEXT: vmovdqa32 %zmm15, %zmm23 {%k1}8252; AVX512-FCP-NEXT: vmovdqa64 192(%r8), %zmm28253; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm5 = [28,1,2,3,4,29,6,7,8,9,30,11,12,13,14,31]8254; AVX512-FCP-NEXT: vpermt2d %zmm2, %zmm5, %zmm38255; AVX512-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8256; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm15 = [0,25,2,3,4,5,26,7,8,9,10,27,12,13,14,15]8257; AVX512-FCP-NEXT: vpermt2d %zmm2, %zmm15, %zmm318258; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm16 = [0,1,22,3,4,5,6,23,8,9,10,11,24,13,14,15]8259; AVX512-FCP-NEXT: vpermt2d %zmm2, %zmm16, %zmm228260; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm25 = [0,1,2,19,4,5,6,7,20,9,10,11,12,21,14,15]8261; AVX512-FCP-NEXT: vpermt2d %zmm2, %zmm25, %zmm218262; AVX512-FCP-NEXT: vpmovsxbd {{.*#+}} zmm3 = [0,1,2,3,16,5,6,7,8,17,10,11,12,13,18,15]8263; AVX512-FCP-NEXT: vpermt2d %zmm2, %zmm3, %zmm238264; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload8265; AVX512-FCP-NEXT: vmovdqa32 %zmm2, %zmm18 {%k2}8266; AVX512-FCP-NEXT: vmovdqa32 %zmm11, %zmm20 {%k1}8267; AVX512-FCP-NEXT: vmovdqa32 %zmm12, %zmm27 {%k2}8268; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload8269; AVX512-FCP-NEXT: vmovdqa32 %zmm2, %zmm30 {%k3}8270; AVX512-FCP-NEXT: vmovdqa32 %zmm29, %zmm19 {%k1}8271; AVX512-FCP-NEXT: vmovdqa64 128(%r8), %zmm28272; AVX512-FCP-NEXT: vpermt2d %zmm2, %zmm5, %zmm188273; AVX512-FCP-NEXT: vpermt2d %zmm2, %zmm15, %zmm208274; AVX512-FCP-NEXT: vpermt2d %zmm2, %zmm16, %zmm278275; AVX512-FCP-NEXT: vpermt2d %zmm2, %zmm25, %zmm308276; AVX512-FCP-NEXT: vpermt2d %zmm2, %zmm3, %zmm198277; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload8278; AVX512-FCP-NEXT: vmovdqa32 %zmm2, %zmm6 {%k2}8279; AVX512-FCP-NEXT: vmovdqa32 %zmm10, %zmm24 {%k1}8280; AVX512-FCP-NEXT: vmovdqa32 %zmm9, %zmm26 {%k2}8281; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload8282; AVX512-FCP-NEXT: vmovdqa32 %zmm2, %zmm4 {%k3}8283; AVX512-FCP-NEXT: vmovdqa64 (%r8), %zmm28284; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload8285; AVX512-FCP-NEXT: vmovdqa32 %zmm7, %zmm0 {%k1}8286; AVX512-FCP-NEXT: vmovdqa64 64(%r8), %zmm78287; AVX512-FCP-NEXT: vpermt2d %zmm2, %zmm3, %zmm88288; AVX512-FCP-NEXT: vpermt2d %zmm7, %zmm3, %zmm08289; AVX512-FCP-NEXT: vpermt2d %zmm7, %zmm5, %zmm68290; AVX512-FCP-NEXT: vpermt2d %zmm7, %zmm15, %zmm248291; AVX512-FCP-NEXT: vpermt2d %zmm7, %zmm16, %zmm268292; AVX512-FCP-NEXT: vpermt2d %zmm7, %zmm25, %zmm48293; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload8294; AVX512-FCP-NEXT: vmovdqa32 %zmm3, %zmm17 {%k2}8295; AVX512-FCP-NEXT: vpermt2d %zmm2, %zmm5, %zmm178296; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload8297; AVX512-FCP-NEXT: vmovdqa32 %zmm3, %zmm14 {%k1}8298; AVX512-FCP-NEXT: vpermt2d %zmm2, %zmm15, %zmm148299; AVX512-FCP-NEXT: vmovdqa32 %zmm28, %zmm13 {%k2}8300; AVX512-FCP-NEXT: vpermt2d %zmm2, %zmm16, %zmm138301; AVX512-FCP-NEXT: vmovdqu64 (%rsp), %zmm3 # 64-byte Reload8302; AVX512-FCP-NEXT: vmovdqa32 %zmm3, %zmm1 {%k3}8303; AVX512-FCP-NEXT: vpermt2d %zmm2, %zmm25, %zmm18304; AVX512-FCP-NEXT: vmovdqa64 %zmm1, 64(%r9)8305; AVX512-FCP-NEXT: vmovdqa64 %zmm13, 128(%r9)8306; AVX512-FCP-NEXT: vmovdqa64 %zmm14, 192(%r9)8307; AVX512-FCP-NEXT: vmovdqa64 %zmm17, 256(%r9)8308; AVX512-FCP-NEXT: vmovdqa64 %zmm0, 320(%r9)8309; AVX512-FCP-NEXT: vmovdqa64 %zmm4, 384(%r9)8310; AVX512-FCP-NEXT: vmovdqa64 %zmm26, 448(%r9)8311; AVX512-FCP-NEXT: vmovdqa64 %zmm24, 512(%r9)8312; AVX512-FCP-NEXT: vmovdqa64 %zmm6, 576(%r9)8313; AVX512-FCP-NEXT: vmovdqa64 %zmm19, 640(%r9)8314; AVX512-FCP-NEXT: vmovdqa64 %zmm30, 704(%r9)8315; AVX512-FCP-NEXT: vmovdqa64 %zmm27, 768(%r9)8316; AVX512-FCP-NEXT: vmovdqa64 %zmm20, 832(%r9)8317; AVX512-FCP-NEXT: vmovdqa64 %zmm18, 896(%r9)8318; AVX512-FCP-NEXT: vmovdqa64 %zmm23, 960(%r9)8319; AVX512-FCP-NEXT: vmovdqa64 %zmm21, 1024(%r9)8320; AVX512-FCP-NEXT: vmovdqa64 %zmm22, 1088(%r9)8321; AVX512-FCP-NEXT: vmovdqa64 %zmm31, 1152(%r9)8322; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8323; AVX512-FCP-NEXT: vmovaps %zmm0, 1216(%r9)8324; AVX512-FCP-NEXT: vmovdqa64 %zmm8, (%r9)8325; AVX512-FCP-NEXT: addq $712, %rsp # imm = 0x2C88326; AVX512-FCP-NEXT: vzeroupper8327; AVX512-FCP-NEXT: retq8328;8329; AVX512DQ-LABEL: store_i32_stride5_vf64:8330; AVX512DQ: # %bb.0:8331; AVX512DQ-NEXT: subq $712, %rsp # imm = 0x2C88332; AVX512DQ-NEXT: vmovdqa64 128(%rdi), %zmm198333; AVX512DQ-NEXT: vmovdqa64 192(%rdi), %zmm238334; AVX512DQ-NEXT: vmovdqa64 192(%rsi), %zmm188335; AVX512DQ-NEXT: vmovdqa64 (%rdx), %zmm18336; AVX512DQ-NEXT: vmovdqa64 64(%rdx), %zmm38337; AVX512DQ-NEXT: vmovdqa64 128(%rdx), %zmm298338; AVX512DQ-NEXT: vmovdqa64 192(%rdx), %zmm158339; AVX512DQ-NEXT: vmovdqa64 (%rcx), %zmm218340; AVX512DQ-NEXT: vmovdqa64 64(%rcx), %zmm258341; AVX512DQ-NEXT: vmovdqa64 128(%rcx), %zmm248342; AVX512DQ-NEXT: vmovdqa64 192(%rcx), %zmm08343; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm28 = [17,0,0,16,2,18,0,1,17,0,0,16,2,18,0,1]8344; AVX512DQ-NEXT: # zmm28 = mem[0,1,2,3,0,1,2,3]8345; AVX512DQ-NEXT: vmovdqa64 %zmm1, %zmm28346; AVX512DQ-NEXT: vpermt2d %zmm21, %zmm28, %zmm28347; AVX512DQ-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8348; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [0,16,2,18,0,1,17,3,0,16,2,18,0,1,17,3]8349; AVX512DQ-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]8350; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm2 = [14,30,0,13,29,15,31,0,14,30,0,13,29,15,31,0]8351; AVX512DQ-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3]8352; AVX512DQ-NEXT: vmovdqa64 %zmm15, %zmm48353; AVX512DQ-NEXT: vpermt2d %zmm0, %zmm2, %zmm48354; AVX512DQ-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8355; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm48356; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm17 = [0,13,29,15,31,0,14,30,0,13,29,15,31,0,14,30]8357; AVX512DQ-NEXT: # zmm17 = mem[0,1,2,3,0,1,2,3]8358; AVX512DQ-NEXT: vmovdqa64 %zmm23, %zmm28359; AVX512DQ-NEXT: vpermt2d %zmm18, %zmm17, %zmm28360; AVX512DQ-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8361; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm2 = [27,0,10,26,12,28,0,11,27,0,10,26,12,28,0,11]8362; AVX512DQ-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3]8363; AVX512DQ-NEXT: vmovdqa64 %zmm23, %zmm58364; AVX512DQ-NEXT: vpermt2d %zmm18, %zmm2, %zmm58365; AVX512DQ-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8366; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm58367; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm14 = [9,27,11,0,26,10,28,12,9,27,11,0,26,10,28,12]8368; AVX512DQ-NEXT: # zmm14 = mem[0,1,2,3,0,1,2,3]8369; AVX512DQ-NEXT: vmovdqa64 %zmm0, %zmm318370; AVX512DQ-NEXT: vpermt2d %zmm15, %zmm14, %zmm318371; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm13 = [6,22,8,24,0,7,23,9,6,22,8,24,0,7,23,9]8372; AVX512DQ-NEXT: # zmm13 = mem[0,1,2,3,0,1,2,3]8373; AVX512DQ-NEXT: vmovdqa64 %zmm15, %zmm228374; AVX512DQ-NEXT: vpermt2d %zmm0, %zmm13, %zmm228375; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm30 = [0,3,19,5,21,0,4,20,0,3,19,5,21,0,4,20]8376; AVX512DQ-NEXT: # zmm30 = mem[0,1,2,3,0,1,2,3]8377; AVX512DQ-NEXT: vmovdqa64 %zmm15, %zmm28378; AVX512DQ-NEXT: vpermt2d %zmm0, %zmm30, %zmm28379; AVX512DQ-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8380; AVX512DQ-NEXT: vpermt2d %zmm0, %zmm28, %zmm158381; AVX512DQ-NEXT: vmovdqa64 %zmm29, %zmm08382; AVX512DQ-NEXT: vpermt2d %zmm24, %zmm4, %zmm08383; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8384; AVX512DQ-NEXT: vmovdqa64 %zmm24, %zmm208385; AVX512DQ-NEXT: vpermt2d %zmm29, %zmm14, %zmm208386; AVX512DQ-NEXT: vmovdqa64 %zmm29, %zmm278387; AVX512DQ-NEXT: vpermt2d %zmm24, %zmm13, %zmm278388; AVX512DQ-NEXT: vmovdqa64 %zmm29, %zmm08389; AVX512DQ-NEXT: vpermt2d %zmm24, %zmm30, %zmm08390; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8391; AVX512DQ-NEXT: vpermt2d %zmm24, %zmm28, %zmm298392; AVX512DQ-NEXT: vmovdqa64 %zmm3, %zmm08393; AVX512DQ-NEXT: vmovdqa64 %zmm25, %zmm248394; AVX512DQ-NEXT: vpermt2d %zmm3, %zmm14, %zmm248395; AVX512DQ-NEXT: vmovdqa64 %zmm3, %zmm268396; AVX512DQ-NEXT: vmovdqa64 %zmm3, %zmm28397; AVX512DQ-NEXT: vpermt2d %zmm25, %zmm28, %zmm38398; AVX512DQ-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8399; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm28 = [8,24,0,7,23,9,25,0,8,24,0,7,23,9,25,0]8400; AVX512DQ-NEXT: # zmm28 = mem[0,1,2,3,0,1,2,3]8401; AVX512DQ-NEXT: vpermt2d %zmm25, %zmm4, %zmm08402; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8403; AVX512DQ-NEXT: vpermt2d %zmm25, %zmm13, %zmm268404; AVX512DQ-NEXT: vpermt2d %zmm25, %zmm30, %zmm28405; AVX512DQ-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8406; AVX512DQ-NEXT: vmovdqa64 %zmm23, %zmm168407; AVX512DQ-NEXT: vpermt2d %zmm18, %zmm28, %zmm168408; AVX512DQ-NEXT: vpermi2d %zmm21, %zmm1, %zmm48409; AVX512DQ-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8410; AVX512DQ-NEXT: vpermi2d %zmm1, %zmm21, %zmm148411; AVX512DQ-NEXT: vpermi2d %zmm21, %zmm1, %zmm138412; AVX512DQ-NEXT: vpermt2d %zmm21, %zmm30, %zmm18413; AVX512DQ-NEXT: vmovdqu64 %zmm1, (%rsp) # 64-byte Spill8414; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm2 = [3,21,5,0,20,4,22,6,3,21,5,0,20,4,22,6]8415; AVX512DQ-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3]8416; AVX512DQ-NEXT: vmovdqa64 %zmm18, %zmm218417; AVX512DQ-NEXT: vpermt2d %zmm23, %zmm2, %zmm218418; AVX512DQ-NEXT: vpermt2d %zmm18, %zmm7, %zmm238419; AVX512DQ-NEXT: vmovdqa64 128(%rsi), %zmm08420; AVX512DQ-NEXT: vmovdqa64 %zmm19, %zmm188421; AVX512DQ-NEXT: vpermt2d %zmm0, %zmm17, %zmm188422; AVX512DQ-NEXT: vmovdqa64 %zmm19, %zmm118423; AVX512DQ-NEXT: vpermt2d %zmm0, %zmm5, %zmm118424; AVX512DQ-NEXT: vmovdqa64 %zmm19, %zmm128425; AVX512DQ-NEXT: vpermt2d %zmm0, %zmm28, %zmm128426; AVX512DQ-NEXT: vmovdqa64 %zmm0, %zmm308427; AVX512DQ-NEXT: vpermt2d %zmm19, %zmm2, %zmm308428; AVX512DQ-NEXT: vpermt2d %zmm0, %zmm7, %zmm198429; AVX512DQ-NEXT: vmovdqa64 (%rdi), %zmm258430; AVX512DQ-NEXT: vmovdqa64 (%rsi), %zmm18431; AVX512DQ-NEXT: vmovdqa64 %zmm25, %zmm88432; AVX512DQ-NEXT: vpermt2d %zmm1, %zmm7, %zmm88433; AVX512DQ-NEXT: vmovdqa64 64(%rdi), %zmm08434; AVX512DQ-NEXT: vmovdqa64 64(%rsi), %zmm38435; AVX512DQ-NEXT: vmovdqa64 %zmm3, %zmm48436; AVX512DQ-NEXT: vpermt2d %zmm0, %zmm2, %zmm48437; AVX512DQ-NEXT: vmovdqa64 %zmm0, %zmm68438; AVX512DQ-NEXT: vmovdqa64 %zmm0, %zmm108439; AVX512DQ-NEXT: vmovdqa64 %zmm0, %zmm98440; AVX512DQ-NEXT: vpermt2d %zmm3, %zmm7, %zmm08441; AVX512DQ-NEXT: vpermt2d %zmm3, %zmm17, %zmm68442; AVX512DQ-NEXT: vpermt2d %zmm3, %zmm5, %zmm108443; AVX512DQ-NEXT: vpermt2d %zmm3, %zmm28, %zmm98444; AVX512DQ-NEXT: vpermi2d %zmm1, %zmm25, %zmm178445; AVX512DQ-NEXT: vpermi2d %zmm1, %zmm25, %zmm58446; AVX512DQ-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8447; AVX512DQ-NEXT: vpermi2d %zmm1, %zmm25, %zmm288448; AVX512DQ-NEXT: vpermt2d %zmm25, %zmm2, %zmm18449; AVX512DQ-NEXT: movw $12684, %ax # imm = 0x318C8450; AVX512DQ-NEXT: kmovw %eax, %k18451; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload8452; AVX512DQ-NEXT: vmovdqa32 %zmm2, %zmm8 {%k1}8453; AVX512DQ-NEXT: movw $25368, %ax # imm = 0x63188454; AVX512DQ-NEXT: kmovw %eax, %k28455; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload8456; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload8457; AVX512DQ-NEXT: vmovdqa32 %zmm2, %zmm3 {%k2}8458; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload8459; AVX512DQ-NEXT: vmovdqa32 %zmm2, %zmm31 {%k1}8460; AVX512DQ-NEXT: vmovdqa32 %zmm16, %zmm22 {%k2}8461; AVX512DQ-NEXT: movw $6342, %ax # imm = 0x18C68462; AVX512DQ-NEXT: kmovw %eax, %k38463; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload8464; AVX512DQ-NEXT: vmovdqa32 %zmm2, %zmm21 {%k3}8465; AVX512DQ-NEXT: vmovdqa32 %zmm15, %zmm23 {%k1}8466; AVX512DQ-NEXT: vmovdqa64 192(%r8), %zmm28467; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm5 = [28,1,2,3,4,29,6,7,8,9,30,11,12,13,14,31]8468; AVX512DQ-NEXT: vpermt2d %zmm2, %zmm5, %zmm38469; AVX512DQ-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8470; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm15 = [0,25,2,3,4,5,26,7,8,9,10,27,12,13,14,15]8471; AVX512DQ-NEXT: vpermt2d %zmm2, %zmm15, %zmm318472; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm16 = [0,1,22,3,4,5,6,23,8,9,10,11,24,13,14,15]8473; AVX512DQ-NEXT: vpermt2d %zmm2, %zmm16, %zmm228474; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm25 = [0,1,2,19,4,5,6,7,20,9,10,11,12,21,14,15]8475; AVX512DQ-NEXT: vpermt2d %zmm2, %zmm25, %zmm218476; AVX512DQ-NEXT: vpmovsxbd {{.*#+}} zmm3 = [0,1,2,3,16,5,6,7,8,17,10,11,12,13,18,15]8477; AVX512DQ-NEXT: vpermt2d %zmm2, %zmm3, %zmm238478; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload8479; AVX512DQ-NEXT: vmovdqa32 %zmm2, %zmm18 {%k2}8480; AVX512DQ-NEXT: vmovdqa32 %zmm11, %zmm20 {%k1}8481; AVX512DQ-NEXT: vmovdqa32 %zmm12, %zmm27 {%k2}8482; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload8483; AVX512DQ-NEXT: vmovdqa32 %zmm2, %zmm30 {%k3}8484; AVX512DQ-NEXT: vmovdqa32 %zmm29, %zmm19 {%k1}8485; AVX512DQ-NEXT: vmovdqa64 128(%r8), %zmm28486; AVX512DQ-NEXT: vpermt2d %zmm2, %zmm5, %zmm188487; AVX512DQ-NEXT: vpermt2d %zmm2, %zmm15, %zmm208488; AVX512DQ-NEXT: vpermt2d %zmm2, %zmm16, %zmm278489; AVX512DQ-NEXT: vpermt2d %zmm2, %zmm25, %zmm308490; AVX512DQ-NEXT: vpermt2d %zmm2, %zmm3, %zmm198491; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload8492; AVX512DQ-NEXT: vmovdqa32 %zmm2, %zmm6 {%k2}8493; AVX512DQ-NEXT: vmovdqa32 %zmm10, %zmm24 {%k1}8494; AVX512DQ-NEXT: vmovdqa32 %zmm9, %zmm26 {%k2}8495; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload8496; AVX512DQ-NEXT: vmovdqa32 %zmm2, %zmm4 {%k3}8497; AVX512DQ-NEXT: vmovdqa64 (%r8), %zmm28498; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload8499; AVX512DQ-NEXT: vmovdqa32 %zmm7, %zmm0 {%k1}8500; AVX512DQ-NEXT: vmovdqa64 64(%r8), %zmm78501; AVX512DQ-NEXT: vpermt2d %zmm2, %zmm3, %zmm88502; AVX512DQ-NEXT: vpermt2d %zmm7, %zmm3, %zmm08503; AVX512DQ-NEXT: vpermt2d %zmm7, %zmm5, %zmm68504; AVX512DQ-NEXT: vpermt2d %zmm7, %zmm15, %zmm248505; AVX512DQ-NEXT: vpermt2d %zmm7, %zmm16, %zmm268506; AVX512DQ-NEXT: vpermt2d %zmm7, %zmm25, %zmm48507; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload8508; AVX512DQ-NEXT: vmovdqa32 %zmm3, %zmm17 {%k2}8509; AVX512DQ-NEXT: vpermt2d %zmm2, %zmm5, %zmm178510; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload8511; AVX512DQ-NEXT: vmovdqa32 %zmm3, %zmm14 {%k1}8512; AVX512DQ-NEXT: vpermt2d %zmm2, %zmm15, %zmm148513; AVX512DQ-NEXT: vmovdqa32 %zmm28, %zmm13 {%k2}8514; AVX512DQ-NEXT: vpermt2d %zmm2, %zmm16, %zmm138515; AVX512DQ-NEXT: vmovdqu64 (%rsp), %zmm3 # 64-byte Reload8516; AVX512DQ-NEXT: vmovdqa32 %zmm3, %zmm1 {%k3}8517; AVX512DQ-NEXT: vpermt2d %zmm2, %zmm25, %zmm18518; AVX512DQ-NEXT: vmovdqa64 %zmm1, 64(%r9)8519; AVX512DQ-NEXT: vmovdqa64 %zmm13, 128(%r9)8520; AVX512DQ-NEXT: vmovdqa64 %zmm14, 192(%r9)8521; AVX512DQ-NEXT: vmovdqa64 %zmm17, 256(%r9)8522; AVX512DQ-NEXT: vmovdqa64 %zmm0, 320(%r9)8523; AVX512DQ-NEXT: vmovdqa64 %zmm4, 384(%r9)8524; AVX512DQ-NEXT: vmovdqa64 %zmm26, 448(%r9)8525; AVX512DQ-NEXT: vmovdqa64 %zmm24, 512(%r9)8526; AVX512DQ-NEXT: vmovdqa64 %zmm6, 576(%r9)8527; AVX512DQ-NEXT: vmovdqa64 %zmm19, 640(%r9)8528; AVX512DQ-NEXT: vmovdqa64 %zmm30, 704(%r9)8529; AVX512DQ-NEXT: vmovdqa64 %zmm27, 768(%r9)8530; AVX512DQ-NEXT: vmovdqa64 %zmm20, 832(%r9)8531; AVX512DQ-NEXT: vmovdqa64 %zmm18, 896(%r9)8532; AVX512DQ-NEXT: vmovdqa64 %zmm23, 960(%r9)8533; AVX512DQ-NEXT: vmovdqa64 %zmm21, 1024(%r9)8534; AVX512DQ-NEXT: vmovdqa64 %zmm22, 1088(%r9)8535; AVX512DQ-NEXT: vmovdqa64 %zmm31, 1152(%r9)8536; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8537; AVX512DQ-NEXT: vmovaps %zmm0, 1216(%r9)8538; AVX512DQ-NEXT: vmovdqa64 %zmm8, (%r9)8539; AVX512DQ-NEXT: addq $712, %rsp # imm = 0x2C88540; AVX512DQ-NEXT: vzeroupper8541; AVX512DQ-NEXT: retq8542;8543; AVX512DQ-FCP-LABEL: store_i32_stride5_vf64:8544; AVX512DQ-FCP: # %bb.0:8545; AVX512DQ-FCP-NEXT: subq $712, %rsp # imm = 0x2C88546; AVX512DQ-FCP-NEXT: vmovdqa64 128(%rdi), %zmm198547; AVX512DQ-FCP-NEXT: vmovdqa64 192(%rdi), %zmm238548; AVX512DQ-FCP-NEXT: vmovdqa64 192(%rsi), %zmm188549; AVX512DQ-FCP-NEXT: vmovdqa64 (%rdx), %zmm18550; AVX512DQ-FCP-NEXT: vmovdqa64 64(%rdx), %zmm38551; AVX512DQ-FCP-NEXT: vmovdqa64 128(%rdx), %zmm298552; AVX512DQ-FCP-NEXT: vmovdqa64 192(%rdx), %zmm158553; AVX512DQ-FCP-NEXT: vmovdqa64 (%rcx), %zmm218554; AVX512DQ-FCP-NEXT: vmovdqa64 64(%rcx), %zmm258555; AVX512DQ-FCP-NEXT: vmovdqa64 128(%rcx), %zmm248556; AVX512DQ-FCP-NEXT: vmovdqa64 192(%rcx), %zmm08557; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm28 = [17,0,0,16,2,18,0,1,17,0,0,16,2,18,0,1]8558; AVX512DQ-FCP-NEXT: # zmm28 = mem[0,1,2,3,0,1,2,3]8559; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, %zmm28560; AVX512DQ-FCP-NEXT: vpermt2d %zmm21, %zmm28, %zmm28561; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8562; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [0,16,2,18,0,1,17,3,0,16,2,18,0,1,17,3]8563; AVX512DQ-FCP-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]8564; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm2 = [14,30,0,13,29,15,31,0,14,30,0,13,29,15,31,0]8565; AVX512DQ-FCP-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3]8566; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm15, %zmm48567; AVX512DQ-FCP-NEXT: vpermt2d %zmm0, %zmm2, %zmm48568; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8569; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm48570; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm17 = [0,13,29,15,31,0,14,30,0,13,29,15,31,0,14,30]8571; AVX512DQ-FCP-NEXT: # zmm17 = mem[0,1,2,3,0,1,2,3]8572; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm23, %zmm28573; AVX512DQ-FCP-NEXT: vpermt2d %zmm18, %zmm17, %zmm28574; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8575; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm2 = [27,0,10,26,12,28,0,11,27,0,10,26,12,28,0,11]8576; AVX512DQ-FCP-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3]8577; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm23, %zmm58578; AVX512DQ-FCP-NEXT: vpermt2d %zmm18, %zmm2, %zmm58579; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8580; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm58581; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm14 = [9,27,11,0,26,10,28,12,9,27,11,0,26,10,28,12]8582; AVX512DQ-FCP-NEXT: # zmm14 = mem[0,1,2,3,0,1,2,3]8583; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm318584; AVX512DQ-FCP-NEXT: vpermt2d %zmm15, %zmm14, %zmm318585; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm13 = [6,22,8,24,0,7,23,9,6,22,8,24,0,7,23,9]8586; AVX512DQ-FCP-NEXT: # zmm13 = mem[0,1,2,3,0,1,2,3]8587; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm15, %zmm228588; AVX512DQ-FCP-NEXT: vpermt2d %zmm0, %zmm13, %zmm228589; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm30 = [0,3,19,5,21,0,4,20,0,3,19,5,21,0,4,20]8590; AVX512DQ-FCP-NEXT: # zmm30 = mem[0,1,2,3,0,1,2,3]8591; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm15, %zmm28592; AVX512DQ-FCP-NEXT: vpermt2d %zmm0, %zmm30, %zmm28593; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8594; AVX512DQ-FCP-NEXT: vpermt2d %zmm0, %zmm28, %zmm158595; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm29, %zmm08596; AVX512DQ-FCP-NEXT: vpermt2d %zmm24, %zmm4, %zmm08597; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8598; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm24, %zmm208599; AVX512DQ-FCP-NEXT: vpermt2d %zmm29, %zmm14, %zmm208600; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm29, %zmm278601; AVX512DQ-FCP-NEXT: vpermt2d %zmm24, %zmm13, %zmm278602; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm29, %zmm08603; AVX512DQ-FCP-NEXT: vpermt2d %zmm24, %zmm30, %zmm08604; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8605; AVX512DQ-FCP-NEXT: vpermt2d %zmm24, %zmm28, %zmm298606; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, %zmm08607; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm25, %zmm248608; AVX512DQ-FCP-NEXT: vpermt2d %zmm3, %zmm14, %zmm248609; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, %zmm268610; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, %zmm28611; AVX512DQ-FCP-NEXT: vpermt2d %zmm25, %zmm28, %zmm38612; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8613; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm28 = [8,24,0,7,23,9,25,0,8,24,0,7,23,9,25,0]8614; AVX512DQ-FCP-NEXT: # zmm28 = mem[0,1,2,3,0,1,2,3]8615; AVX512DQ-FCP-NEXT: vpermt2d %zmm25, %zmm4, %zmm08616; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8617; AVX512DQ-FCP-NEXT: vpermt2d %zmm25, %zmm13, %zmm268618; AVX512DQ-FCP-NEXT: vpermt2d %zmm25, %zmm30, %zmm28619; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8620; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm23, %zmm168621; AVX512DQ-FCP-NEXT: vpermt2d %zmm18, %zmm28, %zmm168622; AVX512DQ-FCP-NEXT: vpermi2d %zmm21, %zmm1, %zmm48623; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8624; AVX512DQ-FCP-NEXT: vpermi2d %zmm1, %zmm21, %zmm148625; AVX512DQ-FCP-NEXT: vpermi2d %zmm21, %zmm1, %zmm138626; AVX512DQ-FCP-NEXT: vpermt2d %zmm21, %zmm30, %zmm18627; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, (%rsp) # 64-byte Spill8628; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm2 = [3,21,5,0,20,4,22,6,3,21,5,0,20,4,22,6]8629; AVX512DQ-FCP-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3]8630; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm18, %zmm218631; AVX512DQ-FCP-NEXT: vpermt2d %zmm23, %zmm2, %zmm218632; AVX512DQ-FCP-NEXT: vpermt2d %zmm18, %zmm7, %zmm238633; AVX512DQ-FCP-NEXT: vmovdqa64 128(%rsi), %zmm08634; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm19, %zmm188635; AVX512DQ-FCP-NEXT: vpermt2d %zmm0, %zmm17, %zmm188636; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm19, %zmm118637; AVX512DQ-FCP-NEXT: vpermt2d %zmm0, %zmm5, %zmm118638; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm19, %zmm128639; AVX512DQ-FCP-NEXT: vpermt2d %zmm0, %zmm28, %zmm128640; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm308641; AVX512DQ-FCP-NEXT: vpermt2d %zmm19, %zmm2, %zmm308642; AVX512DQ-FCP-NEXT: vpermt2d %zmm0, %zmm7, %zmm198643; AVX512DQ-FCP-NEXT: vmovdqa64 (%rdi), %zmm258644; AVX512DQ-FCP-NEXT: vmovdqa64 (%rsi), %zmm18645; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm25, %zmm88646; AVX512DQ-FCP-NEXT: vpermt2d %zmm1, %zmm7, %zmm88647; AVX512DQ-FCP-NEXT: vmovdqa64 64(%rdi), %zmm08648; AVX512DQ-FCP-NEXT: vmovdqa64 64(%rsi), %zmm38649; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, %zmm48650; AVX512DQ-FCP-NEXT: vpermt2d %zmm0, %zmm2, %zmm48651; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm68652; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm108653; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm98654; AVX512DQ-FCP-NEXT: vpermt2d %zmm3, %zmm7, %zmm08655; AVX512DQ-FCP-NEXT: vpermt2d %zmm3, %zmm17, %zmm68656; AVX512DQ-FCP-NEXT: vpermt2d %zmm3, %zmm5, %zmm108657; AVX512DQ-FCP-NEXT: vpermt2d %zmm3, %zmm28, %zmm98658; AVX512DQ-FCP-NEXT: vpermi2d %zmm1, %zmm25, %zmm178659; AVX512DQ-FCP-NEXT: vpermi2d %zmm1, %zmm25, %zmm58660; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8661; AVX512DQ-FCP-NEXT: vpermi2d %zmm1, %zmm25, %zmm288662; AVX512DQ-FCP-NEXT: vpermt2d %zmm25, %zmm2, %zmm18663; AVX512DQ-FCP-NEXT: movw $12684, %ax # imm = 0x318C8664; AVX512DQ-FCP-NEXT: kmovw %eax, %k18665; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload8666; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm2, %zmm8 {%k1}8667; AVX512DQ-FCP-NEXT: movw $25368, %ax # imm = 0x63188668; AVX512DQ-FCP-NEXT: kmovw %eax, %k28669; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload8670; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload8671; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm2, %zmm3 {%k2}8672; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload8673; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm2, %zmm31 {%k1}8674; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm16, %zmm22 {%k2}8675; AVX512DQ-FCP-NEXT: movw $6342, %ax # imm = 0x18C68676; AVX512DQ-FCP-NEXT: kmovw %eax, %k38677; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload8678; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm2, %zmm21 {%k3}8679; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm15, %zmm23 {%k1}8680; AVX512DQ-FCP-NEXT: vmovdqa64 192(%r8), %zmm28681; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm5 = [28,1,2,3,4,29,6,7,8,9,30,11,12,13,14,31]8682; AVX512DQ-FCP-NEXT: vpermt2d %zmm2, %zmm5, %zmm38683; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8684; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm15 = [0,25,2,3,4,5,26,7,8,9,10,27,12,13,14,15]8685; AVX512DQ-FCP-NEXT: vpermt2d %zmm2, %zmm15, %zmm318686; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm16 = [0,1,22,3,4,5,6,23,8,9,10,11,24,13,14,15]8687; AVX512DQ-FCP-NEXT: vpermt2d %zmm2, %zmm16, %zmm228688; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm25 = [0,1,2,19,4,5,6,7,20,9,10,11,12,21,14,15]8689; AVX512DQ-FCP-NEXT: vpermt2d %zmm2, %zmm25, %zmm218690; AVX512DQ-FCP-NEXT: vpmovsxbd {{.*#+}} zmm3 = [0,1,2,3,16,5,6,7,8,17,10,11,12,13,18,15]8691; AVX512DQ-FCP-NEXT: vpermt2d %zmm2, %zmm3, %zmm238692; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload8693; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm2, %zmm18 {%k2}8694; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm11, %zmm20 {%k1}8695; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm12, %zmm27 {%k2}8696; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload8697; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm2, %zmm30 {%k3}8698; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm29, %zmm19 {%k1}8699; AVX512DQ-FCP-NEXT: vmovdqa64 128(%r8), %zmm28700; AVX512DQ-FCP-NEXT: vpermt2d %zmm2, %zmm5, %zmm188701; AVX512DQ-FCP-NEXT: vpermt2d %zmm2, %zmm15, %zmm208702; AVX512DQ-FCP-NEXT: vpermt2d %zmm2, %zmm16, %zmm278703; AVX512DQ-FCP-NEXT: vpermt2d %zmm2, %zmm25, %zmm308704; AVX512DQ-FCP-NEXT: vpermt2d %zmm2, %zmm3, %zmm198705; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload8706; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm2, %zmm6 {%k2}8707; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm10, %zmm24 {%k1}8708; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm9, %zmm26 {%k2}8709; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload8710; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm2, %zmm4 {%k3}8711; AVX512DQ-FCP-NEXT: vmovdqa64 (%r8), %zmm28712; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload8713; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm7, %zmm0 {%k1}8714; AVX512DQ-FCP-NEXT: vmovdqa64 64(%r8), %zmm78715; AVX512DQ-FCP-NEXT: vpermt2d %zmm2, %zmm3, %zmm88716; AVX512DQ-FCP-NEXT: vpermt2d %zmm7, %zmm3, %zmm08717; AVX512DQ-FCP-NEXT: vpermt2d %zmm7, %zmm5, %zmm68718; AVX512DQ-FCP-NEXT: vpermt2d %zmm7, %zmm15, %zmm248719; AVX512DQ-FCP-NEXT: vpermt2d %zmm7, %zmm16, %zmm268720; AVX512DQ-FCP-NEXT: vpermt2d %zmm7, %zmm25, %zmm48721; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload8722; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm3, %zmm17 {%k2}8723; AVX512DQ-FCP-NEXT: vpermt2d %zmm2, %zmm5, %zmm178724; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload8725; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm3, %zmm14 {%k1}8726; AVX512DQ-FCP-NEXT: vpermt2d %zmm2, %zmm15, %zmm148727; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm28, %zmm13 {%k2}8728; AVX512DQ-FCP-NEXT: vpermt2d %zmm2, %zmm16, %zmm138729; AVX512DQ-FCP-NEXT: vmovdqu64 (%rsp), %zmm3 # 64-byte Reload8730; AVX512DQ-FCP-NEXT: vmovdqa32 %zmm3, %zmm1 {%k3}8731; AVX512DQ-FCP-NEXT: vpermt2d %zmm2, %zmm25, %zmm18732; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, 64(%r9)8733; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm13, 128(%r9)8734; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm14, 192(%r9)8735; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm17, 256(%r9)8736; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, 320(%r9)8737; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm4, 384(%r9)8738; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm26, 448(%r9)8739; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm24, 512(%r9)8740; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, 576(%r9)8741; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm19, 640(%r9)8742; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm30, 704(%r9)8743; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm27, 768(%r9)8744; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm20, 832(%r9)8745; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm18, 896(%r9)8746; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm23, 960(%r9)8747; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm21, 1024(%r9)8748; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm22, 1088(%r9)8749; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm31, 1152(%r9)8750; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8751; AVX512DQ-FCP-NEXT: vmovaps %zmm0, 1216(%r9)8752; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm8, (%r9)8753; AVX512DQ-FCP-NEXT: addq $712, %rsp # imm = 0x2C88754; AVX512DQ-FCP-NEXT: vzeroupper8755; AVX512DQ-FCP-NEXT: retq8756;8757; AVX512BW-LABEL: store_i32_stride5_vf64:8758; AVX512BW: # %bb.0:8759; AVX512BW-NEXT: subq $712, %rsp # imm = 0x2C88760; AVX512BW-NEXT: vmovdqa64 128(%rdi), %zmm198761; AVX512BW-NEXT: vmovdqa64 192(%rdi), %zmm238762; AVX512BW-NEXT: vmovdqa64 192(%rsi), %zmm188763; AVX512BW-NEXT: vmovdqa64 (%rdx), %zmm18764; AVX512BW-NEXT: vmovdqa64 64(%rdx), %zmm38765; AVX512BW-NEXT: vmovdqa64 128(%rdx), %zmm298766; AVX512BW-NEXT: vmovdqa64 192(%rdx), %zmm158767; AVX512BW-NEXT: vmovdqa64 (%rcx), %zmm218768; AVX512BW-NEXT: vmovdqa64 64(%rcx), %zmm258769; AVX512BW-NEXT: vmovdqa64 128(%rcx), %zmm248770; AVX512BW-NEXT: vmovdqa64 192(%rcx), %zmm08771; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm28 = [17,0,0,16,2,18,0,1,17,0,0,16,2,18,0,1]8772; AVX512BW-NEXT: # zmm28 = mem[0,1,2,3,0,1,2,3]8773; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm28774; AVX512BW-NEXT: vpermt2d %zmm21, %zmm28, %zmm28775; AVX512BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8776; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [0,16,2,18,0,1,17,3,0,16,2,18,0,1,17,3]8777; AVX512BW-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]8778; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm2 = [14,30,0,13,29,15,31,0,14,30,0,13,29,15,31,0]8779; AVX512BW-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3]8780; AVX512BW-NEXT: vmovdqa64 %zmm15, %zmm48781; AVX512BW-NEXT: vpermt2d %zmm0, %zmm2, %zmm48782; AVX512BW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8783; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm48784; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm17 = [0,13,29,15,31,0,14,30,0,13,29,15,31,0,14,30]8785; AVX512BW-NEXT: # zmm17 = mem[0,1,2,3,0,1,2,3]8786; AVX512BW-NEXT: vmovdqa64 %zmm23, %zmm28787; AVX512BW-NEXT: vpermt2d %zmm18, %zmm17, %zmm28788; AVX512BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8789; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm2 = [27,0,10,26,12,28,0,11,27,0,10,26,12,28,0,11]8790; AVX512BW-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3]8791; AVX512BW-NEXT: vmovdqa64 %zmm23, %zmm58792; AVX512BW-NEXT: vpermt2d %zmm18, %zmm2, %zmm58793; AVX512BW-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8794; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm58795; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm14 = [9,27,11,0,26,10,28,12,9,27,11,0,26,10,28,12]8796; AVX512BW-NEXT: # zmm14 = mem[0,1,2,3,0,1,2,3]8797; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm318798; AVX512BW-NEXT: vpermt2d %zmm15, %zmm14, %zmm318799; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm13 = [6,22,8,24,0,7,23,9,6,22,8,24,0,7,23,9]8800; AVX512BW-NEXT: # zmm13 = mem[0,1,2,3,0,1,2,3]8801; AVX512BW-NEXT: vmovdqa64 %zmm15, %zmm228802; AVX512BW-NEXT: vpermt2d %zmm0, %zmm13, %zmm228803; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm30 = [0,3,19,5,21,0,4,20,0,3,19,5,21,0,4,20]8804; AVX512BW-NEXT: # zmm30 = mem[0,1,2,3,0,1,2,3]8805; AVX512BW-NEXT: vmovdqa64 %zmm15, %zmm28806; AVX512BW-NEXT: vpermt2d %zmm0, %zmm30, %zmm28807; AVX512BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8808; AVX512BW-NEXT: vpermt2d %zmm0, %zmm28, %zmm158809; AVX512BW-NEXT: vmovdqa64 %zmm29, %zmm08810; AVX512BW-NEXT: vpermt2d %zmm24, %zmm4, %zmm08811; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8812; AVX512BW-NEXT: vmovdqa64 %zmm24, %zmm208813; AVX512BW-NEXT: vpermt2d %zmm29, %zmm14, %zmm208814; AVX512BW-NEXT: vmovdqa64 %zmm29, %zmm278815; AVX512BW-NEXT: vpermt2d %zmm24, %zmm13, %zmm278816; AVX512BW-NEXT: vmovdqa64 %zmm29, %zmm08817; AVX512BW-NEXT: vpermt2d %zmm24, %zmm30, %zmm08818; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8819; AVX512BW-NEXT: vpermt2d %zmm24, %zmm28, %zmm298820; AVX512BW-NEXT: vmovdqa64 %zmm3, %zmm08821; AVX512BW-NEXT: vmovdqa64 %zmm25, %zmm248822; AVX512BW-NEXT: vpermt2d %zmm3, %zmm14, %zmm248823; AVX512BW-NEXT: vmovdqa64 %zmm3, %zmm268824; AVX512BW-NEXT: vmovdqa64 %zmm3, %zmm28825; AVX512BW-NEXT: vpermt2d %zmm25, %zmm28, %zmm38826; AVX512BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8827; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm28 = [8,24,0,7,23,9,25,0,8,24,0,7,23,9,25,0]8828; AVX512BW-NEXT: # zmm28 = mem[0,1,2,3,0,1,2,3]8829; AVX512BW-NEXT: vpermt2d %zmm25, %zmm4, %zmm08830; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8831; AVX512BW-NEXT: vpermt2d %zmm25, %zmm13, %zmm268832; AVX512BW-NEXT: vpermt2d %zmm25, %zmm30, %zmm28833; AVX512BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8834; AVX512BW-NEXT: vmovdqa64 %zmm23, %zmm168835; AVX512BW-NEXT: vpermt2d %zmm18, %zmm28, %zmm168836; AVX512BW-NEXT: vpermi2d %zmm21, %zmm1, %zmm48837; AVX512BW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8838; AVX512BW-NEXT: vpermi2d %zmm1, %zmm21, %zmm148839; AVX512BW-NEXT: vpermi2d %zmm21, %zmm1, %zmm138840; AVX512BW-NEXT: vpermt2d %zmm21, %zmm30, %zmm18841; AVX512BW-NEXT: vmovdqu64 %zmm1, (%rsp) # 64-byte Spill8842; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm2 = [3,21,5,0,20,4,22,6,3,21,5,0,20,4,22,6]8843; AVX512BW-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3]8844; AVX512BW-NEXT: vmovdqa64 %zmm18, %zmm218845; AVX512BW-NEXT: vpermt2d %zmm23, %zmm2, %zmm218846; AVX512BW-NEXT: vpermt2d %zmm18, %zmm7, %zmm238847; AVX512BW-NEXT: vmovdqa64 128(%rsi), %zmm08848; AVX512BW-NEXT: vmovdqa64 %zmm19, %zmm188849; AVX512BW-NEXT: vpermt2d %zmm0, %zmm17, %zmm188850; AVX512BW-NEXT: vmovdqa64 %zmm19, %zmm118851; AVX512BW-NEXT: vpermt2d %zmm0, %zmm5, %zmm118852; AVX512BW-NEXT: vmovdqa64 %zmm19, %zmm128853; AVX512BW-NEXT: vpermt2d %zmm0, %zmm28, %zmm128854; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm308855; AVX512BW-NEXT: vpermt2d %zmm19, %zmm2, %zmm308856; AVX512BW-NEXT: vpermt2d %zmm0, %zmm7, %zmm198857; AVX512BW-NEXT: vmovdqa64 (%rdi), %zmm258858; AVX512BW-NEXT: vmovdqa64 (%rsi), %zmm18859; AVX512BW-NEXT: vmovdqa64 %zmm25, %zmm88860; AVX512BW-NEXT: vpermt2d %zmm1, %zmm7, %zmm88861; AVX512BW-NEXT: vmovdqa64 64(%rdi), %zmm08862; AVX512BW-NEXT: vmovdqa64 64(%rsi), %zmm38863; AVX512BW-NEXT: vmovdqa64 %zmm3, %zmm48864; AVX512BW-NEXT: vpermt2d %zmm0, %zmm2, %zmm48865; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm68866; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm108867; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm98868; AVX512BW-NEXT: vpermt2d %zmm3, %zmm7, %zmm08869; AVX512BW-NEXT: vpermt2d %zmm3, %zmm17, %zmm68870; AVX512BW-NEXT: vpermt2d %zmm3, %zmm5, %zmm108871; AVX512BW-NEXT: vpermt2d %zmm3, %zmm28, %zmm98872; AVX512BW-NEXT: vpermi2d %zmm1, %zmm25, %zmm178873; AVX512BW-NEXT: vpermi2d %zmm1, %zmm25, %zmm58874; AVX512BW-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8875; AVX512BW-NEXT: vpermi2d %zmm1, %zmm25, %zmm288876; AVX512BW-NEXT: vpermt2d %zmm25, %zmm2, %zmm18877; AVX512BW-NEXT: movw $12684, %ax # imm = 0x318C8878; AVX512BW-NEXT: kmovd %eax, %k18879; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload8880; AVX512BW-NEXT: vmovdqa32 %zmm2, %zmm8 {%k1}8881; AVX512BW-NEXT: movw $25368, %ax # imm = 0x63188882; AVX512BW-NEXT: kmovd %eax, %k28883; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload8884; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload8885; AVX512BW-NEXT: vmovdqa32 %zmm2, %zmm3 {%k2}8886; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload8887; AVX512BW-NEXT: vmovdqa32 %zmm2, %zmm31 {%k1}8888; AVX512BW-NEXT: vmovdqa32 %zmm16, %zmm22 {%k2}8889; AVX512BW-NEXT: movw $6342, %ax # imm = 0x18C68890; AVX512BW-NEXT: kmovd %eax, %k38891; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload8892; AVX512BW-NEXT: vmovdqa32 %zmm2, %zmm21 {%k3}8893; AVX512BW-NEXT: vmovdqa32 %zmm15, %zmm23 {%k1}8894; AVX512BW-NEXT: vmovdqa64 192(%r8), %zmm28895; AVX512BW-NEXT: vpmovsxbd {{.*#+}} zmm5 = [28,1,2,3,4,29,6,7,8,9,30,11,12,13,14,31]8896; AVX512BW-NEXT: vpermt2d %zmm2, %zmm5, %zmm38897; AVX512BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8898; AVX512BW-NEXT: vpmovsxbd {{.*#+}} zmm15 = [0,25,2,3,4,5,26,7,8,9,10,27,12,13,14,15]8899; AVX512BW-NEXT: vpermt2d %zmm2, %zmm15, %zmm318900; AVX512BW-NEXT: vpmovsxbd {{.*#+}} zmm16 = [0,1,22,3,4,5,6,23,8,9,10,11,24,13,14,15]8901; AVX512BW-NEXT: vpermt2d %zmm2, %zmm16, %zmm228902; AVX512BW-NEXT: vpmovsxbd {{.*#+}} zmm25 = [0,1,2,19,4,5,6,7,20,9,10,11,12,21,14,15]8903; AVX512BW-NEXT: vpermt2d %zmm2, %zmm25, %zmm218904; AVX512BW-NEXT: vpmovsxbd {{.*#+}} zmm3 = [0,1,2,3,16,5,6,7,8,17,10,11,12,13,18,15]8905; AVX512BW-NEXT: vpermt2d %zmm2, %zmm3, %zmm238906; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload8907; AVX512BW-NEXT: vmovdqa32 %zmm2, %zmm18 {%k2}8908; AVX512BW-NEXT: vmovdqa32 %zmm11, %zmm20 {%k1}8909; AVX512BW-NEXT: vmovdqa32 %zmm12, %zmm27 {%k2}8910; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload8911; AVX512BW-NEXT: vmovdqa32 %zmm2, %zmm30 {%k3}8912; AVX512BW-NEXT: vmovdqa32 %zmm29, %zmm19 {%k1}8913; AVX512BW-NEXT: vmovdqa64 128(%r8), %zmm28914; AVX512BW-NEXT: vpermt2d %zmm2, %zmm5, %zmm188915; AVX512BW-NEXT: vpermt2d %zmm2, %zmm15, %zmm208916; AVX512BW-NEXT: vpermt2d %zmm2, %zmm16, %zmm278917; AVX512BW-NEXT: vpermt2d %zmm2, %zmm25, %zmm308918; AVX512BW-NEXT: vpermt2d %zmm2, %zmm3, %zmm198919; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload8920; AVX512BW-NEXT: vmovdqa32 %zmm2, %zmm6 {%k2}8921; AVX512BW-NEXT: vmovdqa32 %zmm10, %zmm24 {%k1}8922; AVX512BW-NEXT: vmovdqa32 %zmm9, %zmm26 {%k2}8923; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload8924; AVX512BW-NEXT: vmovdqa32 %zmm2, %zmm4 {%k3}8925; AVX512BW-NEXT: vmovdqa64 (%r8), %zmm28926; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload8927; AVX512BW-NEXT: vmovdqa32 %zmm7, %zmm0 {%k1}8928; AVX512BW-NEXT: vmovdqa64 64(%r8), %zmm78929; AVX512BW-NEXT: vpermt2d %zmm2, %zmm3, %zmm88930; AVX512BW-NEXT: vpermt2d %zmm7, %zmm3, %zmm08931; AVX512BW-NEXT: vpermt2d %zmm7, %zmm5, %zmm68932; AVX512BW-NEXT: vpermt2d %zmm7, %zmm15, %zmm248933; AVX512BW-NEXT: vpermt2d %zmm7, %zmm16, %zmm268934; AVX512BW-NEXT: vpermt2d %zmm7, %zmm25, %zmm48935; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload8936; AVX512BW-NEXT: vmovdqa32 %zmm3, %zmm17 {%k2}8937; AVX512BW-NEXT: vpermt2d %zmm2, %zmm5, %zmm178938; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload8939; AVX512BW-NEXT: vmovdqa32 %zmm3, %zmm14 {%k1}8940; AVX512BW-NEXT: vpermt2d %zmm2, %zmm15, %zmm148941; AVX512BW-NEXT: vmovdqa32 %zmm28, %zmm13 {%k2}8942; AVX512BW-NEXT: vpermt2d %zmm2, %zmm16, %zmm138943; AVX512BW-NEXT: vmovdqu64 (%rsp), %zmm3 # 64-byte Reload8944; AVX512BW-NEXT: vmovdqa32 %zmm3, %zmm1 {%k3}8945; AVX512BW-NEXT: vpermt2d %zmm2, %zmm25, %zmm18946; AVX512BW-NEXT: vmovdqa64 %zmm1, 64(%r9)8947; AVX512BW-NEXT: vmovdqa64 %zmm13, 128(%r9)8948; AVX512BW-NEXT: vmovdqa64 %zmm14, 192(%r9)8949; AVX512BW-NEXT: vmovdqa64 %zmm17, 256(%r9)8950; AVX512BW-NEXT: vmovdqa64 %zmm0, 320(%r9)8951; AVX512BW-NEXT: vmovdqa64 %zmm4, 384(%r9)8952; AVX512BW-NEXT: vmovdqa64 %zmm26, 448(%r9)8953; AVX512BW-NEXT: vmovdqa64 %zmm24, 512(%r9)8954; AVX512BW-NEXT: vmovdqa64 %zmm6, 576(%r9)8955; AVX512BW-NEXT: vmovdqa64 %zmm19, 640(%r9)8956; AVX512BW-NEXT: vmovdqa64 %zmm30, 704(%r9)8957; AVX512BW-NEXT: vmovdqa64 %zmm27, 768(%r9)8958; AVX512BW-NEXT: vmovdqa64 %zmm20, 832(%r9)8959; AVX512BW-NEXT: vmovdqa64 %zmm18, 896(%r9)8960; AVX512BW-NEXT: vmovdqa64 %zmm23, 960(%r9)8961; AVX512BW-NEXT: vmovdqa64 %zmm21, 1024(%r9)8962; AVX512BW-NEXT: vmovdqa64 %zmm22, 1088(%r9)8963; AVX512BW-NEXT: vmovdqa64 %zmm31, 1152(%r9)8964; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload8965; AVX512BW-NEXT: vmovaps %zmm0, 1216(%r9)8966; AVX512BW-NEXT: vmovdqa64 %zmm8, (%r9)8967; AVX512BW-NEXT: addq $712, %rsp # imm = 0x2C88968; AVX512BW-NEXT: vzeroupper8969; AVX512BW-NEXT: retq8970;8971; AVX512BW-FCP-LABEL: store_i32_stride5_vf64:8972; AVX512BW-FCP: # %bb.0:8973; AVX512BW-FCP-NEXT: subq $712, %rsp # imm = 0x2C88974; AVX512BW-FCP-NEXT: vmovdqa64 128(%rdi), %zmm198975; AVX512BW-FCP-NEXT: vmovdqa64 192(%rdi), %zmm238976; AVX512BW-FCP-NEXT: vmovdqa64 192(%rsi), %zmm188977; AVX512BW-FCP-NEXT: vmovdqa64 (%rdx), %zmm18978; AVX512BW-FCP-NEXT: vmovdqa64 64(%rdx), %zmm38979; AVX512BW-FCP-NEXT: vmovdqa64 128(%rdx), %zmm298980; AVX512BW-FCP-NEXT: vmovdqa64 192(%rdx), %zmm158981; AVX512BW-FCP-NEXT: vmovdqa64 (%rcx), %zmm218982; AVX512BW-FCP-NEXT: vmovdqa64 64(%rcx), %zmm258983; AVX512BW-FCP-NEXT: vmovdqa64 128(%rcx), %zmm248984; AVX512BW-FCP-NEXT: vmovdqa64 192(%rcx), %zmm08985; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm28 = [17,0,0,16,2,18,0,1,17,0,0,16,2,18,0,1]8986; AVX512BW-FCP-NEXT: # zmm28 = mem[0,1,2,3,0,1,2,3]8987; AVX512BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm28988; AVX512BW-FCP-NEXT: vpermt2d %zmm21, %zmm28, %zmm28989; AVX512BW-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8990; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [0,16,2,18,0,1,17,3,0,16,2,18,0,1,17,3]8991; AVX512BW-FCP-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]8992; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm2 = [14,30,0,13,29,15,31,0,14,30,0,13,29,15,31,0]8993; AVX512BW-FCP-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3]8994; AVX512BW-FCP-NEXT: vmovdqa64 %zmm15, %zmm48995; AVX512BW-FCP-NEXT: vpermt2d %zmm0, %zmm2, %zmm48996; AVX512BW-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill8997; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm48998; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm17 = [0,13,29,15,31,0,14,30,0,13,29,15,31,0,14,30]8999; AVX512BW-FCP-NEXT: # zmm17 = mem[0,1,2,3,0,1,2,3]9000; AVX512BW-FCP-NEXT: vmovdqa64 %zmm23, %zmm29001; AVX512BW-FCP-NEXT: vpermt2d %zmm18, %zmm17, %zmm29002; AVX512BW-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9003; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm2 = [27,0,10,26,12,28,0,11,27,0,10,26,12,28,0,11]9004; AVX512BW-FCP-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3]9005; AVX512BW-FCP-NEXT: vmovdqa64 %zmm23, %zmm59006; AVX512BW-FCP-NEXT: vpermt2d %zmm18, %zmm2, %zmm59007; AVX512BW-FCP-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9008; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm59009; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm14 = [9,27,11,0,26,10,28,12,9,27,11,0,26,10,28,12]9010; AVX512BW-FCP-NEXT: # zmm14 = mem[0,1,2,3,0,1,2,3]9011; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm319012; AVX512BW-FCP-NEXT: vpermt2d %zmm15, %zmm14, %zmm319013; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm13 = [6,22,8,24,0,7,23,9,6,22,8,24,0,7,23,9]9014; AVX512BW-FCP-NEXT: # zmm13 = mem[0,1,2,3,0,1,2,3]9015; AVX512BW-FCP-NEXT: vmovdqa64 %zmm15, %zmm229016; AVX512BW-FCP-NEXT: vpermt2d %zmm0, %zmm13, %zmm229017; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm30 = [0,3,19,5,21,0,4,20,0,3,19,5,21,0,4,20]9018; AVX512BW-FCP-NEXT: # zmm30 = mem[0,1,2,3,0,1,2,3]9019; AVX512BW-FCP-NEXT: vmovdqa64 %zmm15, %zmm29020; AVX512BW-FCP-NEXT: vpermt2d %zmm0, %zmm30, %zmm29021; AVX512BW-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9022; AVX512BW-FCP-NEXT: vpermt2d %zmm0, %zmm28, %zmm159023; AVX512BW-FCP-NEXT: vmovdqa64 %zmm29, %zmm09024; AVX512BW-FCP-NEXT: vpermt2d %zmm24, %zmm4, %zmm09025; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9026; AVX512BW-FCP-NEXT: vmovdqa64 %zmm24, %zmm209027; AVX512BW-FCP-NEXT: vpermt2d %zmm29, %zmm14, %zmm209028; AVX512BW-FCP-NEXT: vmovdqa64 %zmm29, %zmm279029; AVX512BW-FCP-NEXT: vpermt2d %zmm24, %zmm13, %zmm279030; AVX512BW-FCP-NEXT: vmovdqa64 %zmm29, %zmm09031; AVX512BW-FCP-NEXT: vpermt2d %zmm24, %zmm30, %zmm09032; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9033; AVX512BW-FCP-NEXT: vpermt2d %zmm24, %zmm28, %zmm299034; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm09035; AVX512BW-FCP-NEXT: vmovdqa64 %zmm25, %zmm249036; AVX512BW-FCP-NEXT: vpermt2d %zmm3, %zmm14, %zmm249037; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm269038; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm29039; AVX512BW-FCP-NEXT: vpermt2d %zmm25, %zmm28, %zmm39040; AVX512BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9041; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm28 = [8,24,0,7,23,9,25,0,8,24,0,7,23,9,25,0]9042; AVX512BW-FCP-NEXT: # zmm28 = mem[0,1,2,3,0,1,2,3]9043; AVX512BW-FCP-NEXT: vpermt2d %zmm25, %zmm4, %zmm09044; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9045; AVX512BW-FCP-NEXT: vpermt2d %zmm25, %zmm13, %zmm269046; AVX512BW-FCP-NEXT: vpermt2d %zmm25, %zmm30, %zmm29047; AVX512BW-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9048; AVX512BW-FCP-NEXT: vmovdqa64 %zmm23, %zmm169049; AVX512BW-FCP-NEXT: vpermt2d %zmm18, %zmm28, %zmm169050; AVX512BW-FCP-NEXT: vpermi2d %zmm21, %zmm1, %zmm49051; AVX512BW-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9052; AVX512BW-FCP-NEXT: vpermi2d %zmm1, %zmm21, %zmm149053; AVX512BW-FCP-NEXT: vpermi2d %zmm21, %zmm1, %zmm139054; AVX512BW-FCP-NEXT: vpermt2d %zmm21, %zmm30, %zmm19055; AVX512BW-FCP-NEXT: vmovdqu64 %zmm1, (%rsp) # 64-byte Spill9056; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm2 = [3,21,5,0,20,4,22,6,3,21,5,0,20,4,22,6]9057; AVX512BW-FCP-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3]9058; AVX512BW-FCP-NEXT: vmovdqa64 %zmm18, %zmm219059; AVX512BW-FCP-NEXT: vpermt2d %zmm23, %zmm2, %zmm219060; AVX512BW-FCP-NEXT: vpermt2d %zmm18, %zmm7, %zmm239061; AVX512BW-FCP-NEXT: vmovdqa64 128(%rsi), %zmm09062; AVX512BW-FCP-NEXT: vmovdqa64 %zmm19, %zmm189063; AVX512BW-FCP-NEXT: vpermt2d %zmm0, %zmm17, %zmm189064; AVX512BW-FCP-NEXT: vmovdqa64 %zmm19, %zmm119065; AVX512BW-FCP-NEXT: vpermt2d %zmm0, %zmm5, %zmm119066; AVX512BW-FCP-NEXT: vmovdqa64 %zmm19, %zmm129067; AVX512BW-FCP-NEXT: vpermt2d %zmm0, %zmm28, %zmm129068; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm309069; AVX512BW-FCP-NEXT: vpermt2d %zmm19, %zmm2, %zmm309070; AVX512BW-FCP-NEXT: vpermt2d %zmm0, %zmm7, %zmm199071; AVX512BW-FCP-NEXT: vmovdqa64 (%rdi), %zmm259072; AVX512BW-FCP-NEXT: vmovdqa64 (%rsi), %zmm19073; AVX512BW-FCP-NEXT: vmovdqa64 %zmm25, %zmm89074; AVX512BW-FCP-NEXT: vpermt2d %zmm1, %zmm7, %zmm89075; AVX512BW-FCP-NEXT: vmovdqa64 64(%rdi), %zmm09076; AVX512BW-FCP-NEXT: vmovdqa64 64(%rsi), %zmm39077; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm49078; AVX512BW-FCP-NEXT: vpermt2d %zmm0, %zmm2, %zmm49079; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm69080; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm109081; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm99082; AVX512BW-FCP-NEXT: vpermt2d %zmm3, %zmm7, %zmm09083; AVX512BW-FCP-NEXT: vpermt2d %zmm3, %zmm17, %zmm69084; AVX512BW-FCP-NEXT: vpermt2d %zmm3, %zmm5, %zmm109085; AVX512BW-FCP-NEXT: vpermt2d %zmm3, %zmm28, %zmm99086; AVX512BW-FCP-NEXT: vpermi2d %zmm1, %zmm25, %zmm179087; AVX512BW-FCP-NEXT: vpermi2d %zmm1, %zmm25, %zmm59088; AVX512BW-FCP-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9089; AVX512BW-FCP-NEXT: vpermi2d %zmm1, %zmm25, %zmm289090; AVX512BW-FCP-NEXT: vpermt2d %zmm25, %zmm2, %zmm19091; AVX512BW-FCP-NEXT: movw $12684, %ax # imm = 0x318C9092; AVX512BW-FCP-NEXT: kmovd %eax, %k19093; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload9094; AVX512BW-FCP-NEXT: vmovdqa32 %zmm2, %zmm8 {%k1}9095; AVX512BW-FCP-NEXT: movw $25368, %ax # imm = 0x63189096; AVX512BW-FCP-NEXT: kmovd %eax, %k29097; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload9098; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload9099; AVX512BW-FCP-NEXT: vmovdqa32 %zmm2, %zmm3 {%k2}9100; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload9101; AVX512BW-FCP-NEXT: vmovdqa32 %zmm2, %zmm31 {%k1}9102; AVX512BW-FCP-NEXT: vmovdqa32 %zmm16, %zmm22 {%k2}9103; AVX512BW-FCP-NEXT: movw $6342, %ax # imm = 0x18C69104; AVX512BW-FCP-NEXT: kmovd %eax, %k39105; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload9106; AVX512BW-FCP-NEXT: vmovdqa32 %zmm2, %zmm21 {%k3}9107; AVX512BW-FCP-NEXT: vmovdqa32 %zmm15, %zmm23 {%k1}9108; AVX512BW-FCP-NEXT: vmovdqa64 192(%r8), %zmm29109; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} zmm5 = [28,1,2,3,4,29,6,7,8,9,30,11,12,13,14,31]9110; AVX512BW-FCP-NEXT: vpermt2d %zmm2, %zmm5, %zmm39111; AVX512BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9112; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} zmm15 = [0,25,2,3,4,5,26,7,8,9,10,27,12,13,14,15]9113; AVX512BW-FCP-NEXT: vpermt2d %zmm2, %zmm15, %zmm319114; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} zmm16 = [0,1,22,3,4,5,6,23,8,9,10,11,24,13,14,15]9115; AVX512BW-FCP-NEXT: vpermt2d %zmm2, %zmm16, %zmm229116; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} zmm25 = [0,1,2,19,4,5,6,7,20,9,10,11,12,21,14,15]9117; AVX512BW-FCP-NEXT: vpermt2d %zmm2, %zmm25, %zmm219118; AVX512BW-FCP-NEXT: vpmovsxbd {{.*#+}} zmm3 = [0,1,2,3,16,5,6,7,8,17,10,11,12,13,18,15]9119; AVX512BW-FCP-NEXT: vpermt2d %zmm2, %zmm3, %zmm239120; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload9121; AVX512BW-FCP-NEXT: vmovdqa32 %zmm2, %zmm18 {%k2}9122; AVX512BW-FCP-NEXT: vmovdqa32 %zmm11, %zmm20 {%k1}9123; AVX512BW-FCP-NEXT: vmovdqa32 %zmm12, %zmm27 {%k2}9124; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload9125; AVX512BW-FCP-NEXT: vmovdqa32 %zmm2, %zmm30 {%k3}9126; AVX512BW-FCP-NEXT: vmovdqa32 %zmm29, %zmm19 {%k1}9127; AVX512BW-FCP-NEXT: vmovdqa64 128(%r8), %zmm29128; AVX512BW-FCP-NEXT: vpermt2d %zmm2, %zmm5, %zmm189129; AVX512BW-FCP-NEXT: vpermt2d %zmm2, %zmm15, %zmm209130; AVX512BW-FCP-NEXT: vpermt2d %zmm2, %zmm16, %zmm279131; AVX512BW-FCP-NEXT: vpermt2d %zmm2, %zmm25, %zmm309132; AVX512BW-FCP-NEXT: vpermt2d %zmm2, %zmm3, %zmm199133; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload9134; AVX512BW-FCP-NEXT: vmovdqa32 %zmm2, %zmm6 {%k2}9135; AVX512BW-FCP-NEXT: vmovdqa32 %zmm10, %zmm24 {%k1}9136; AVX512BW-FCP-NEXT: vmovdqa32 %zmm9, %zmm26 {%k2}9137; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload9138; AVX512BW-FCP-NEXT: vmovdqa32 %zmm2, %zmm4 {%k3}9139; AVX512BW-FCP-NEXT: vmovdqa64 (%r8), %zmm29140; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload9141; AVX512BW-FCP-NEXT: vmovdqa32 %zmm7, %zmm0 {%k1}9142; AVX512BW-FCP-NEXT: vmovdqa64 64(%r8), %zmm79143; AVX512BW-FCP-NEXT: vpermt2d %zmm2, %zmm3, %zmm89144; AVX512BW-FCP-NEXT: vpermt2d %zmm7, %zmm3, %zmm09145; AVX512BW-FCP-NEXT: vpermt2d %zmm7, %zmm5, %zmm69146; AVX512BW-FCP-NEXT: vpermt2d %zmm7, %zmm15, %zmm249147; AVX512BW-FCP-NEXT: vpermt2d %zmm7, %zmm16, %zmm269148; AVX512BW-FCP-NEXT: vpermt2d %zmm7, %zmm25, %zmm49149; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload9150; AVX512BW-FCP-NEXT: vmovdqa32 %zmm3, %zmm17 {%k2}9151; AVX512BW-FCP-NEXT: vpermt2d %zmm2, %zmm5, %zmm179152; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload9153; AVX512BW-FCP-NEXT: vmovdqa32 %zmm3, %zmm14 {%k1}9154; AVX512BW-FCP-NEXT: vpermt2d %zmm2, %zmm15, %zmm149155; AVX512BW-FCP-NEXT: vmovdqa32 %zmm28, %zmm13 {%k2}9156; AVX512BW-FCP-NEXT: vpermt2d %zmm2, %zmm16, %zmm139157; AVX512BW-FCP-NEXT: vmovdqu64 (%rsp), %zmm3 # 64-byte Reload9158; AVX512BW-FCP-NEXT: vmovdqa32 %zmm3, %zmm1 {%k3}9159; AVX512BW-FCP-NEXT: vpermt2d %zmm2, %zmm25, %zmm19160; AVX512BW-FCP-NEXT: vmovdqa64 %zmm1, 64(%r9)9161; AVX512BW-FCP-NEXT: vmovdqa64 %zmm13, 128(%r9)9162; AVX512BW-FCP-NEXT: vmovdqa64 %zmm14, 192(%r9)9163; AVX512BW-FCP-NEXT: vmovdqa64 %zmm17, 256(%r9)9164; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, 320(%r9)9165; AVX512BW-FCP-NEXT: vmovdqa64 %zmm4, 384(%r9)9166; AVX512BW-FCP-NEXT: vmovdqa64 %zmm26, 448(%r9)9167; AVX512BW-FCP-NEXT: vmovdqa64 %zmm24, 512(%r9)9168; AVX512BW-FCP-NEXT: vmovdqa64 %zmm6, 576(%r9)9169; AVX512BW-FCP-NEXT: vmovdqa64 %zmm19, 640(%r9)9170; AVX512BW-FCP-NEXT: vmovdqa64 %zmm30, 704(%r9)9171; AVX512BW-FCP-NEXT: vmovdqa64 %zmm27, 768(%r9)9172; AVX512BW-FCP-NEXT: vmovdqa64 %zmm20, 832(%r9)9173; AVX512BW-FCP-NEXT: vmovdqa64 %zmm18, 896(%r9)9174; AVX512BW-FCP-NEXT: vmovdqa64 %zmm23, 960(%r9)9175; AVX512BW-FCP-NEXT: vmovdqa64 %zmm21, 1024(%r9)9176; AVX512BW-FCP-NEXT: vmovdqa64 %zmm22, 1088(%r9)9177; AVX512BW-FCP-NEXT: vmovdqa64 %zmm31, 1152(%r9)9178; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload9179; AVX512BW-FCP-NEXT: vmovaps %zmm0, 1216(%r9)9180; AVX512BW-FCP-NEXT: vmovdqa64 %zmm8, (%r9)9181; AVX512BW-FCP-NEXT: addq $712, %rsp # imm = 0x2C89182; AVX512BW-FCP-NEXT: vzeroupper9183; AVX512BW-FCP-NEXT: retq9184;9185; AVX512DQ-BW-LABEL: store_i32_stride5_vf64:9186; AVX512DQ-BW: # %bb.0:9187; AVX512DQ-BW-NEXT: subq $712, %rsp # imm = 0x2C89188; AVX512DQ-BW-NEXT: vmovdqa64 128(%rdi), %zmm199189; AVX512DQ-BW-NEXT: vmovdqa64 192(%rdi), %zmm239190; AVX512DQ-BW-NEXT: vmovdqa64 192(%rsi), %zmm189191; AVX512DQ-BW-NEXT: vmovdqa64 (%rdx), %zmm19192; AVX512DQ-BW-NEXT: vmovdqa64 64(%rdx), %zmm39193; AVX512DQ-BW-NEXT: vmovdqa64 128(%rdx), %zmm299194; AVX512DQ-BW-NEXT: vmovdqa64 192(%rdx), %zmm159195; AVX512DQ-BW-NEXT: vmovdqa64 (%rcx), %zmm219196; AVX512DQ-BW-NEXT: vmovdqa64 64(%rcx), %zmm259197; AVX512DQ-BW-NEXT: vmovdqa64 128(%rcx), %zmm249198; AVX512DQ-BW-NEXT: vmovdqa64 192(%rcx), %zmm09199; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm28 = [17,0,0,16,2,18,0,1,17,0,0,16,2,18,0,1]9200; AVX512DQ-BW-NEXT: # zmm28 = mem[0,1,2,3,0,1,2,3]9201; AVX512DQ-BW-NEXT: vmovdqa64 %zmm1, %zmm29202; AVX512DQ-BW-NEXT: vpermt2d %zmm21, %zmm28, %zmm29203; AVX512DQ-BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9204; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [0,16,2,18,0,1,17,3,0,16,2,18,0,1,17,3]9205; AVX512DQ-BW-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]9206; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm2 = [14,30,0,13,29,15,31,0,14,30,0,13,29,15,31,0]9207; AVX512DQ-BW-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3]9208; AVX512DQ-BW-NEXT: vmovdqa64 %zmm15, %zmm49209; AVX512DQ-BW-NEXT: vpermt2d %zmm0, %zmm2, %zmm49210; AVX512DQ-BW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9211; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm49212; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm17 = [0,13,29,15,31,0,14,30,0,13,29,15,31,0,14,30]9213; AVX512DQ-BW-NEXT: # zmm17 = mem[0,1,2,3,0,1,2,3]9214; AVX512DQ-BW-NEXT: vmovdqa64 %zmm23, %zmm29215; AVX512DQ-BW-NEXT: vpermt2d %zmm18, %zmm17, %zmm29216; AVX512DQ-BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9217; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm2 = [27,0,10,26,12,28,0,11,27,0,10,26,12,28,0,11]9218; AVX512DQ-BW-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3]9219; AVX512DQ-BW-NEXT: vmovdqa64 %zmm23, %zmm59220; AVX512DQ-BW-NEXT: vpermt2d %zmm18, %zmm2, %zmm59221; AVX512DQ-BW-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9222; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm59223; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm14 = [9,27,11,0,26,10,28,12,9,27,11,0,26,10,28,12]9224; AVX512DQ-BW-NEXT: # zmm14 = mem[0,1,2,3,0,1,2,3]9225; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, %zmm319226; AVX512DQ-BW-NEXT: vpermt2d %zmm15, %zmm14, %zmm319227; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm13 = [6,22,8,24,0,7,23,9,6,22,8,24,0,7,23,9]9228; AVX512DQ-BW-NEXT: # zmm13 = mem[0,1,2,3,0,1,2,3]9229; AVX512DQ-BW-NEXT: vmovdqa64 %zmm15, %zmm229230; AVX512DQ-BW-NEXT: vpermt2d %zmm0, %zmm13, %zmm229231; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm30 = [0,3,19,5,21,0,4,20,0,3,19,5,21,0,4,20]9232; AVX512DQ-BW-NEXT: # zmm30 = mem[0,1,2,3,0,1,2,3]9233; AVX512DQ-BW-NEXT: vmovdqa64 %zmm15, %zmm29234; AVX512DQ-BW-NEXT: vpermt2d %zmm0, %zmm30, %zmm29235; AVX512DQ-BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9236; AVX512DQ-BW-NEXT: vpermt2d %zmm0, %zmm28, %zmm159237; AVX512DQ-BW-NEXT: vmovdqa64 %zmm29, %zmm09238; AVX512DQ-BW-NEXT: vpermt2d %zmm24, %zmm4, %zmm09239; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9240; AVX512DQ-BW-NEXT: vmovdqa64 %zmm24, %zmm209241; AVX512DQ-BW-NEXT: vpermt2d %zmm29, %zmm14, %zmm209242; AVX512DQ-BW-NEXT: vmovdqa64 %zmm29, %zmm279243; AVX512DQ-BW-NEXT: vpermt2d %zmm24, %zmm13, %zmm279244; AVX512DQ-BW-NEXT: vmovdqa64 %zmm29, %zmm09245; AVX512DQ-BW-NEXT: vpermt2d %zmm24, %zmm30, %zmm09246; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9247; AVX512DQ-BW-NEXT: vpermt2d %zmm24, %zmm28, %zmm299248; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, %zmm09249; AVX512DQ-BW-NEXT: vmovdqa64 %zmm25, %zmm249250; AVX512DQ-BW-NEXT: vpermt2d %zmm3, %zmm14, %zmm249251; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, %zmm269252; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, %zmm29253; AVX512DQ-BW-NEXT: vpermt2d %zmm25, %zmm28, %zmm39254; AVX512DQ-BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9255; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm28 = [8,24,0,7,23,9,25,0,8,24,0,7,23,9,25,0]9256; AVX512DQ-BW-NEXT: # zmm28 = mem[0,1,2,3,0,1,2,3]9257; AVX512DQ-BW-NEXT: vpermt2d %zmm25, %zmm4, %zmm09258; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9259; AVX512DQ-BW-NEXT: vpermt2d %zmm25, %zmm13, %zmm269260; AVX512DQ-BW-NEXT: vpermt2d %zmm25, %zmm30, %zmm29261; AVX512DQ-BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9262; AVX512DQ-BW-NEXT: vmovdqa64 %zmm23, %zmm169263; AVX512DQ-BW-NEXT: vpermt2d %zmm18, %zmm28, %zmm169264; AVX512DQ-BW-NEXT: vpermi2d %zmm21, %zmm1, %zmm49265; AVX512DQ-BW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9266; AVX512DQ-BW-NEXT: vpermi2d %zmm1, %zmm21, %zmm149267; AVX512DQ-BW-NEXT: vpermi2d %zmm21, %zmm1, %zmm139268; AVX512DQ-BW-NEXT: vpermt2d %zmm21, %zmm30, %zmm19269; AVX512DQ-BW-NEXT: vmovdqu64 %zmm1, (%rsp) # 64-byte Spill9270; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm2 = [3,21,5,0,20,4,22,6,3,21,5,0,20,4,22,6]9271; AVX512DQ-BW-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3]9272; AVX512DQ-BW-NEXT: vmovdqa64 %zmm18, %zmm219273; AVX512DQ-BW-NEXT: vpermt2d %zmm23, %zmm2, %zmm219274; AVX512DQ-BW-NEXT: vpermt2d %zmm18, %zmm7, %zmm239275; AVX512DQ-BW-NEXT: vmovdqa64 128(%rsi), %zmm09276; AVX512DQ-BW-NEXT: vmovdqa64 %zmm19, %zmm189277; AVX512DQ-BW-NEXT: vpermt2d %zmm0, %zmm17, %zmm189278; AVX512DQ-BW-NEXT: vmovdqa64 %zmm19, %zmm119279; AVX512DQ-BW-NEXT: vpermt2d %zmm0, %zmm5, %zmm119280; AVX512DQ-BW-NEXT: vmovdqa64 %zmm19, %zmm129281; AVX512DQ-BW-NEXT: vpermt2d %zmm0, %zmm28, %zmm129282; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, %zmm309283; AVX512DQ-BW-NEXT: vpermt2d %zmm19, %zmm2, %zmm309284; AVX512DQ-BW-NEXT: vpermt2d %zmm0, %zmm7, %zmm199285; AVX512DQ-BW-NEXT: vmovdqa64 (%rdi), %zmm259286; AVX512DQ-BW-NEXT: vmovdqa64 (%rsi), %zmm19287; AVX512DQ-BW-NEXT: vmovdqa64 %zmm25, %zmm89288; AVX512DQ-BW-NEXT: vpermt2d %zmm1, %zmm7, %zmm89289; AVX512DQ-BW-NEXT: vmovdqa64 64(%rdi), %zmm09290; AVX512DQ-BW-NEXT: vmovdqa64 64(%rsi), %zmm39291; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, %zmm49292; AVX512DQ-BW-NEXT: vpermt2d %zmm0, %zmm2, %zmm49293; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, %zmm69294; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, %zmm109295; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, %zmm99296; AVX512DQ-BW-NEXT: vpermt2d %zmm3, %zmm7, %zmm09297; AVX512DQ-BW-NEXT: vpermt2d %zmm3, %zmm17, %zmm69298; AVX512DQ-BW-NEXT: vpermt2d %zmm3, %zmm5, %zmm109299; AVX512DQ-BW-NEXT: vpermt2d %zmm3, %zmm28, %zmm99300; AVX512DQ-BW-NEXT: vpermi2d %zmm1, %zmm25, %zmm179301; AVX512DQ-BW-NEXT: vpermi2d %zmm1, %zmm25, %zmm59302; AVX512DQ-BW-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9303; AVX512DQ-BW-NEXT: vpermi2d %zmm1, %zmm25, %zmm289304; AVX512DQ-BW-NEXT: vpermt2d %zmm25, %zmm2, %zmm19305; AVX512DQ-BW-NEXT: movw $12684, %ax # imm = 0x318C9306; AVX512DQ-BW-NEXT: kmovd %eax, %k19307; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload9308; AVX512DQ-BW-NEXT: vmovdqa32 %zmm2, %zmm8 {%k1}9309; AVX512DQ-BW-NEXT: movw $25368, %ax # imm = 0x63189310; AVX512DQ-BW-NEXT: kmovd %eax, %k29311; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload9312; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload9313; AVX512DQ-BW-NEXT: vmovdqa32 %zmm2, %zmm3 {%k2}9314; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload9315; AVX512DQ-BW-NEXT: vmovdqa32 %zmm2, %zmm31 {%k1}9316; AVX512DQ-BW-NEXT: vmovdqa32 %zmm16, %zmm22 {%k2}9317; AVX512DQ-BW-NEXT: movw $6342, %ax # imm = 0x18C69318; AVX512DQ-BW-NEXT: kmovd %eax, %k39319; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload9320; AVX512DQ-BW-NEXT: vmovdqa32 %zmm2, %zmm21 {%k3}9321; AVX512DQ-BW-NEXT: vmovdqa32 %zmm15, %zmm23 {%k1}9322; AVX512DQ-BW-NEXT: vmovdqa64 192(%r8), %zmm29323; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} zmm5 = [28,1,2,3,4,29,6,7,8,9,30,11,12,13,14,31]9324; AVX512DQ-BW-NEXT: vpermt2d %zmm2, %zmm5, %zmm39325; AVX512DQ-BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9326; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} zmm15 = [0,25,2,3,4,5,26,7,8,9,10,27,12,13,14,15]9327; AVX512DQ-BW-NEXT: vpermt2d %zmm2, %zmm15, %zmm319328; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} zmm16 = [0,1,22,3,4,5,6,23,8,9,10,11,24,13,14,15]9329; AVX512DQ-BW-NEXT: vpermt2d %zmm2, %zmm16, %zmm229330; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} zmm25 = [0,1,2,19,4,5,6,7,20,9,10,11,12,21,14,15]9331; AVX512DQ-BW-NEXT: vpermt2d %zmm2, %zmm25, %zmm219332; AVX512DQ-BW-NEXT: vpmovsxbd {{.*#+}} zmm3 = [0,1,2,3,16,5,6,7,8,17,10,11,12,13,18,15]9333; AVX512DQ-BW-NEXT: vpermt2d %zmm2, %zmm3, %zmm239334; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload9335; AVX512DQ-BW-NEXT: vmovdqa32 %zmm2, %zmm18 {%k2}9336; AVX512DQ-BW-NEXT: vmovdqa32 %zmm11, %zmm20 {%k1}9337; AVX512DQ-BW-NEXT: vmovdqa32 %zmm12, %zmm27 {%k2}9338; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload9339; AVX512DQ-BW-NEXT: vmovdqa32 %zmm2, %zmm30 {%k3}9340; AVX512DQ-BW-NEXT: vmovdqa32 %zmm29, %zmm19 {%k1}9341; AVX512DQ-BW-NEXT: vmovdqa64 128(%r8), %zmm29342; AVX512DQ-BW-NEXT: vpermt2d %zmm2, %zmm5, %zmm189343; AVX512DQ-BW-NEXT: vpermt2d %zmm2, %zmm15, %zmm209344; AVX512DQ-BW-NEXT: vpermt2d %zmm2, %zmm16, %zmm279345; AVX512DQ-BW-NEXT: vpermt2d %zmm2, %zmm25, %zmm309346; AVX512DQ-BW-NEXT: vpermt2d %zmm2, %zmm3, %zmm199347; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload9348; AVX512DQ-BW-NEXT: vmovdqa32 %zmm2, %zmm6 {%k2}9349; AVX512DQ-BW-NEXT: vmovdqa32 %zmm10, %zmm24 {%k1}9350; AVX512DQ-BW-NEXT: vmovdqa32 %zmm9, %zmm26 {%k2}9351; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload9352; AVX512DQ-BW-NEXT: vmovdqa32 %zmm2, %zmm4 {%k3}9353; AVX512DQ-BW-NEXT: vmovdqa64 (%r8), %zmm29354; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload9355; AVX512DQ-BW-NEXT: vmovdqa32 %zmm7, %zmm0 {%k1}9356; AVX512DQ-BW-NEXT: vmovdqa64 64(%r8), %zmm79357; AVX512DQ-BW-NEXT: vpermt2d %zmm2, %zmm3, %zmm89358; AVX512DQ-BW-NEXT: vpermt2d %zmm7, %zmm3, %zmm09359; AVX512DQ-BW-NEXT: vpermt2d %zmm7, %zmm5, %zmm69360; AVX512DQ-BW-NEXT: vpermt2d %zmm7, %zmm15, %zmm249361; AVX512DQ-BW-NEXT: vpermt2d %zmm7, %zmm16, %zmm269362; AVX512DQ-BW-NEXT: vpermt2d %zmm7, %zmm25, %zmm49363; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload9364; AVX512DQ-BW-NEXT: vmovdqa32 %zmm3, %zmm17 {%k2}9365; AVX512DQ-BW-NEXT: vpermt2d %zmm2, %zmm5, %zmm179366; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload9367; AVX512DQ-BW-NEXT: vmovdqa32 %zmm3, %zmm14 {%k1}9368; AVX512DQ-BW-NEXT: vpermt2d %zmm2, %zmm15, %zmm149369; AVX512DQ-BW-NEXT: vmovdqa32 %zmm28, %zmm13 {%k2}9370; AVX512DQ-BW-NEXT: vpermt2d %zmm2, %zmm16, %zmm139371; AVX512DQ-BW-NEXT: vmovdqu64 (%rsp), %zmm3 # 64-byte Reload9372; AVX512DQ-BW-NEXT: vmovdqa32 %zmm3, %zmm1 {%k3}9373; AVX512DQ-BW-NEXT: vpermt2d %zmm2, %zmm25, %zmm19374; AVX512DQ-BW-NEXT: vmovdqa64 %zmm1, 64(%r9)9375; AVX512DQ-BW-NEXT: vmovdqa64 %zmm13, 128(%r9)9376; AVX512DQ-BW-NEXT: vmovdqa64 %zmm14, 192(%r9)9377; AVX512DQ-BW-NEXT: vmovdqa64 %zmm17, 256(%r9)9378; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, 320(%r9)9379; AVX512DQ-BW-NEXT: vmovdqa64 %zmm4, 384(%r9)9380; AVX512DQ-BW-NEXT: vmovdqa64 %zmm26, 448(%r9)9381; AVX512DQ-BW-NEXT: vmovdqa64 %zmm24, 512(%r9)9382; AVX512DQ-BW-NEXT: vmovdqa64 %zmm6, 576(%r9)9383; AVX512DQ-BW-NEXT: vmovdqa64 %zmm19, 640(%r9)9384; AVX512DQ-BW-NEXT: vmovdqa64 %zmm30, 704(%r9)9385; AVX512DQ-BW-NEXT: vmovdqa64 %zmm27, 768(%r9)9386; AVX512DQ-BW-NEXT: vmovdqa64 %zmm20, 832(%r9)9387; AVX512DQ-BW-NEXT: vmovdqa64 %zmm18, 896(%r9)9388; AVX512DQ-BW-NEXT: vmovdqa64 %zmm23, 960(%r9)9389; AVX512DQ-BW-NEXT: vmovdqa64 %zmm21, 1024(%r9)9390; AVX512DQ-BW-NEXT: vmovdqa64 %zmm22, 1088(%r9)9391; AVX512DQ-BW-NEXT: vmovdqa64 %zmm31, 1152(%r9)9392; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload9393; AVX512DQ-BW-NEXT: vmovaps %zmm0, 1216(%r9)9394; AVX512DQ-BW-NEXT: vmovdqa64 %zmm8, (%r9)9395; AVX512DQ-BW-NEXT: addq $712, %rsp # imm = 0x2C89396; AVX512DQ-BW-NEXT: vzeroupper9397; AVX512DQ-BW-NEXT: retq9398;9399; AVX512DQ-BW-FCP-LABEL: store_i32_stride5_vf64:9400; AVX512DQ-BW-FCP: # %bb.0:9401; AVX512DQ-BW-FCP-NEXT: subq $712, %rsp # imm = 0x2C89402; AVX512DQ-BW-FCP-NEXT: vmovdqa64 128(%rdi), %zmm199403; AVX512DQ-BW-FCP-NEXT: vmovdqa64 192(%rdi), %zmm239404; AVX512DQ-BW-FCP-NEXT: vmovdqa64 192(%rsi), %zmm189405; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rdx), %zmm19406; AVX512DQ-BW-FCP-NEXT: vmovdqa64 64(%rdx), %zmm39407; AVX512DQ-BW-FCP-NEXT: vmovdqa64 128(%rdx), %zmm299408; AVX512DQ-BW-FCP-NEXT: vmovdqa64 192(%rdx), %zmm159409; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rcx), %zmm219410; AVX512DQ-BW-FCP-NEXT: vmovdqa64 64(%rcx), %zmm259411; AVX512DQ-BW-FCP-NEXT: vmovdqa64 128(%rcx), %zmm249412; AVX512DQ-BW-FCP-NEXT: vmovdqa64 192(%rcx), %zmm09413; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm28 = [17,0,0,16,2,18,0,1,17,0,0,16,2,18,0,1]9414; AVX512DQ-BW-FCP-NEXT: # zmm28 = mem[0,1,2,3,0,1,2,3]9415; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm29416; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm21, %zmm28, %zmm29417; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9418; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [0,16,2,18,0,1,17,3,0,16,2,18,0,1,17,3]9419; AVX512DQ-BW-FCP-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]9420; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm2 = [14,30,0,13,29,15,31,0,14,30,0,13,29,15,31,0]9421; AVX512DQ-BW-FCP-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3]9422; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm15, %zmm49423; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm0, %zmm2, %zmm49424; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9425; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm49426; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm17 = [0,13,29,15,31,0,14,30,0,13,29,15,31,0,14,30]9427; AVX512DQ-BW-FCP-NEXT: # zmm17 = mem[0,1,2,3,0,1,2,3]9428; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm23, %zmm29429; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm18, %zmm17, %zmm29430; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9431; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm2 = [27,0,10,26,12,28,0,11,27,0,10,26,12,28,0,11]9432; AVX512DQ-BW-FCP-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3]9433; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm23, %zmm59434; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm18, %zmm2, %zmm59435; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9436; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm59437; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm14 = [9,27,11,0,26,10,28,12,9,27,11,0,26,10,28,12]9438; AVX512DQ-BW-FCP-NEXT: # zmm14 = mem[0,1,2,3,0,1,2,3]9439; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm319440; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm15, %zmm14, %zmm319441; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm13 = [6,22,8,24,0,7,23,9,6,22,8,24,0,7,23,9]9442; AVX512DQ-BW-FCP-NEXT: # zmm13 = mem[0,1,2,3,0,1,2,3]9443; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm15, %zmm229444; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm0, %zmm13, %zmm229445; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm30 = [0,3,19,5,21,0,4,20,0,3,19,5,21,0,4,20]9446; AVX512DQ-BW-FCP-NEXT: # zmm30 = mem[0,1,2,3,0,1,2,3]9447; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm15, %zmm29448; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm0, %zmm30, %zmm29449; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9450; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm0, %zmm28, %zmm159451; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm29, %zmm09452; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm24, %zmm4, %zmm09453; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9454; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm24, %zmm209455; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm29, %zmm14, %zmm209456; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm29, %zmm279457; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm24, %zmm13, %zmm279458; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm29, %zmm09459; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm24, %zmm30, %zmm09460; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9461; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm24, %zmm28, %zmm299462; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm09463; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm25, %zmm249464; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm3, %zmm14, %zmm249465; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm269466; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm29467; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm25, %zmm28, %zmm39468; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9469; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm28 = [8,24,0,7,23,9,25,0,8,24,0,7,23,9,25,0]9470; AVX512DQ-BW-FCP-NEXT: # zmm28 = mem[0,1,2,3,0,1,2,3]9471; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm25, %zmm4, %zmm09472; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9473; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm25, %zmm13, %zmm269474; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm25, %zmm30, %zmm29475; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9476; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm23, %zmm169477; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm18, %zmm28, %zmm169478; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm21, %zmm1, %zmm49479; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9480; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm1, %zmm21, %zmm149481; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm21, %zmm1, %zmm139482; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm21, %zmm30, %zmm19483; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm1, (%rsp) # 64-byte Spill9484; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm2 = [3,21,5,0,20,4,22,6,3,21,5,0,20,4,22,6]9485; AVX512DQ-BW-FCP-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3]9486; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm18, %zmm219487; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm23, %zmm2, %zmm219488; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm18, %zmm7, %zmm239489; AVX512DQ-BW-FCP-NEXT: vmovdqa64 128(%rsi), %zmm09490; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm19, %zmm189491; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm0, %zmm17, %zmm189492; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm19, %zmm119493; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm0, %zmm5, %zmm119494; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm19, %zmm129495; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm0, %zmm28, %zmm129496; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm309497; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm19, %zmm2, %zmm309498; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm0, %zmm7, %zmm199499; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rdi), %zmm259500; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rsi), %zmm19501; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm25, %zmm89502; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm1, %zmm7, %zmm89503; AVX512DQ-BW-FCP-NEXT: vmovdqa64 64(%rdi), %zmm09504; AVX512DQ-BW-FCP-NEXT: vmovdqa64 64(%rsi), %zmm39505; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm49506; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm0, %zmm2, %zmm49507; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm69508; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm109509; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm99510; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm3, %zmm7, %zmm09511; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm3, %zmm17, %zmm69512; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm3, %zmm5, %zmm109513; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm3, %zmm28, %zmm99514; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm1, %zmm25, %zmm179515; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm1, %zmm25, %zmm59516; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9517; AVX512DQ-BW-FCP-NEXT: vpermi2d %zmm1, %zmm25, %zmm289518; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm25, %zmm2, %zmm19519; AVX512DQ-BW-FCP-NEXT: movw $12684, %ax # imm = 0x318C9520; AVX512DQ-BW-FCP-NEXT: kmovd %eax, %k19521; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload9522; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm2, %zmm8 {%k1}9523; AVX512DQ-BW-FCP-NEXT: movw $25368, %ax # imm = 0x63189524; AVX512DQ-BW-FCP-NEXT: kmovd %eax, %k29525; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload9526; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload9527; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm2, %zmm3 {%k2}9528; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload9529; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm2, %zmm31 {%k1}9530; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm16, %zmm22 {%k2}9531; AVX512DQ-BW-FCP-NEXT: movw $6342, %ax # imm = 0x18C69532; AVX512DQ-BW-FCP-NEXT: kmovd %eax, %k39533; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload9534; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm2, %zmm21 {%k3}9535; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm15, %zmm23 {%k1}9536; AVX512DQ-BW-FCP-NEXT: vmovdqa64 192(%r8), %zmm29537; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} zmm5 = [28,1,2,3,4,29,6,7,8,9,30,11,12,13,14,31]9538; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm2, %zmm5, %zmm39539; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill9540; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} zmm15 = [0,25,2,3,4,5,26,7,8,9,10,27,12,13,14,15]9541; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm2, %zmm15, %zmm319542; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} zmm16 = [0,1,22,3,4,5,6,23,8,9,10,11,24,13,14,15]9543; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm2, %zmm16, %zmm229544; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} zmm25 = [0,1,2,19,4,5,6,7,20,9,10,11,12,21,14,15]9545; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm2, %zmm25, %zmm219546; AVX512DQ-BW-FCP-NEXT: vpmovsxbd {{.*#+}} zmm3 = [0,1,2,3,16,5,6,7,8,17,10,11,12,13,18,15]9547; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm2, %zmm3, %zmm239548; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload9549; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm2, %zmm18 {%k2}9550; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm11, %zmm20 {%k1}9551; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm12, %zmm27 {%k2}9552; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload9553; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm2, %zmm30 {%k3}9554; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm29, %zmm19 {%k1}9555; AVX512DQ-BW-FCP-NEXT: vmovdqa64 128(%r8), %zmm29556; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm2, %zmm5, %zmm189557; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm2, %zmm15, %zmm209558; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm2, %zmm16, %zmm279559; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm2, %zmm25, %zmm309560; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm2, %zmm3, %zmm199561; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload9562; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm2, %zmm6 {%k2}9563; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm10, %zmm24 {%k1}9564; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm9, %zmm26 {%k2}9565; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload9566; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm2, %zmm4 {%k3}9567; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%r8), %zmm29568; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload9569; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm7, %zmm0 {%k1}9570; AVX512DQ-BW-FCP-NEXT: vmovdqa64 64(%r8), %zmm79571; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm2, %zmm3, %zmm89572; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm7, %zmm3, %zmm09573; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm7, %zmm5, %zmm69574; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm7, %zmm15, %zmm249575; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm7, %zmm16, %zmm269576; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm7, %zmm25, %zmm49577; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload9578; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm3, %zmm17 {%k2}9579; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm2, %zmm5, %zmm179580; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload9581; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm3, %zmm14 {%k1}9582; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm2, %zmm15, %zmm149583; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm28, %zmm13 {%k2}9584; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm2, %zmm16, %zmm139585; AVX512DQ-BW-FCP-NEXT: vmovdqu64 (%rsp), %zmm3 # 64-byte Reload9586; AVX512DQ-BW-FCP-NEXT: vmovdqa32 %zmm3, %zmm1 {%k3}9587; AVX512DQ-BW-FCP-NEXT: vpermt2d %zmm2, %zmm25, %zmm19588; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm1, 64(%r9)9589; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm13, 128(%r9)9590; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm14, 192(%r9)9591; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm17, 256(%r9)9592; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, 320(%r9)9593; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm4, 384(%r9)9594; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm26, 448(%r9)9595; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm24, 512(%r9)9596; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm6, 576(%r9)9597; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm19, 640(%r9)9598; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm30, 704(%r9)9599; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm27, 768(%r9)9600; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm20, 832(%r9)9601; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm18, 896(%r9)9602; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm23, 960(%r9)9603; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm21, 1024(%r9)9604; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm22, 1088(%r9)9605; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm31, 1152(%r9)9606; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload9607; AVX512DQ-BW-FCP-NEXT: vmovaps %zmm0, 1216(%r9)9608; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm8, (%r9)9609; AVX512DQ-BW-FCP-NEXT: addq $712, %rsp # imm = 0x2C89610; AVX512DQ-BW-FCP-NEXT: vzeroupper9611; AVX512DQ-BW-FCP-NEXT: retq9612 %in.vec0 = load <64 x i32>, ptr %in.vecptr0, align 649613 %in.vec1 = load <64 x i32>, ptr %in.vecptr1, align 649614 %in.vec2 = load <64 x i32>, ptr %in.vecptr2, align 649615 %in.vec3 = load <64 x i32>, ptr %in.vecptr3, align 649616 %in.vec4 = load <64 x i32>, ptr %in.vecptr4, align 649617 %1 = shufflevector <64 x i32> %in.vec0, <64 x i32> %in.vec1, <128 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 64, i32 65, i32 66, i32 67, i32 68, i32 69, i32 70, i32 71, i32 72, i32 73, i32 74, i32 75, i32 76, i32 77, i32 78, i32 79, i32 80, i32 81, i32 82, i32 83, i32 84, i32 85, i32 86, i32 87, i32 88, i32 89, i32 90, i32 91, i32 92, i32 93, i32 94, i32 95, i32 96, i32 97, i32 98, i32 99, i32 100, i32 101, i32 102, i32 103, i32 104, i32 105, i32 106, i32 107, i32 108, i32 109, i32 110, i32 111, i32 112, i32 113, i32 114, i32 115, i32 116, i32 117, i32 118, i32 119, i32 120, i32 121, i32 122, i32 123, i32 124, i32 125, i32 126, i32 127>9618 %2 = shufflevector <64 x i32> %in.vec2, <64 x i32> %in.vec3, <128 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 64, i32 65, i32 66, i32 67, i32 68, i32 69, i32 70, i32 71, i32 72, i32 73, i32 74, i32 75, i32 76, i32 77, i32 78, i32 79, i32 80, i32 81, i32 82, i32 83, i32 84, i32 85, i32 86, i32 87, i32 88, i32 89, i32 90, i32 91, i32 92, i32 93, i32 94, i32 95, i32 96, i32 97, i32 98, i32 99, i32 100, i32 101, i32 102, i32 103, i32 104, i32 105, i32 106, i32 107, i32 108, i32 109, i32 110, i32 111, i32 112, i32 113, i32 114, i32 115, i32 116, i32 117, i32 118, i32 119, i32 120, i32 121, i32 122, i32 123, i32 124, i32 125, i32 126, i32 127>9619 %3 = shufflevector <128 x i32> %1, <128 x i32> %2, <256 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 64, i32 65, i32 66, i32 67, i32 68, i32 69, i32 70, i32 71, i32 72, i32 73, i32 74, i32 75, i32 76, i32 77, i32 78, i32 79, i32 80, i32 81, i32 82, i32 83, i32 84, i32 85, i32 86, i32 87, i32 88, i32 89, i32 90, i32 91, i32 92, i32 93, i32 94, i32 95, i32 96, i32 97, i32 98, i32 99, i32 100, i32 101, i32 102, i32 103, i32 104, i32 105, i32 106, i32 107, i32 108, i32 109, i32 110, i32 111, i32 112, i32 113, i32 114, i32 115, i32 116, i32 117, i32 118, i32 119, i32 120, i32 121, i32 122, i32 123, i32 124, i32 125, i32 126, i32 127, i32 128, i32 129, i32 130, i32 131, i32 132, i32 133, i32 134, i32 135, i32 136, i32 137, i32 138, i32 139, i32 140, i32 141, i32 142, i32 143, i32 144, i32 145, i32 146, i32 147, i32 148, i32 149, i32 150, i32 151, i32 152, i32 153, i32 154, i32 155, i32 156, i32 157, i32 158, i32 159, i32 160, i32 161, i32 162, i32 163, i32 164, i32 165, i32 166, i32 167, i32 168, i32 169, i32 170, i32 171, i32 172, i32 173, i32 174, i32 175, i32 176, i32 177, i32 178, i32 179, i32 180, i32 181, i32 182, i32 183, i32 184, i32 185, i32 186, i32 187, i32 188, i32 189, i32 190, i32 191, i32 192, i32 193, i32 194, i32 195, i32 196, i32 197, i32 198, i32 199, i32 200, i32 201, i32 202, i32 203, i32 204, i32 205, i32 206, i32 207, i32 208, i32 209, i32 210, i32 211, i32 212, i32 213, i32 214, i32 215, i32 216, i32 217, i32 218, i32 219, i32 220, i32 221, i32 222, i32 223, i32 224, i32 225, i32 226, i32 227, i32 228, i32 229, i32 230, i32 231, i32 232, i32 233, i32 234, i32 235, i32 236, i32 237, i32 238, i32 239, i32 240, i32 241, i32 242, i32 243, i32 244, i32 245, i32 246, i32 247, i32 248, i32 249, i32 250, i32 251, i32 252, i32 253, i32 254, i32 255>9620 %4 = shufflevector <64 x i32> %in.vec4, <64 x i32> poison, <256 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>9621 %5 = shufflevector <256 x i32> %3, <256 x i32> %4, <320 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 64, i32 65, i32 66, i32 67, i32 68, i32 69, i32 70, i32 71, i32 72, i32 73, i32 74, i32 75, i32 76, i32 77, i32 78, i32 79, i32 80, i32 81, i32 82, i32 83, i32 84, i32 85, i32 86, i32 87, i32 88, i32 89, i32 90, i32 91, i32 92, i32 93, i32 94, i32 95, i32 96, i32 97, i32 98, i32 99, i32 100, i32 101, i32 102, i32 103, i32 104, i32 105, i32 106, i32 107, i32 108, i32 109, i32 110, i32 111, i32 112, i32 113, i32 114, i32 115, i32 116, i32 117, i32 118, i32 119, i32 120, i32 121, i32 122, i32 123, i32 124, i32 125, i32 126, i32 127, i32 128, i32 129, i32 130, i32 131, i32 132, i32 133, i32 134, i32 135, i32 136, i32 137, i32 138, i32 139, i32 140, i32 141, i32 142, i32 143, i32 144, i32 145, i32 146, i32 147, i32 148, i32 149, i32 150, i32 151, i32 152, i32 153, i32 154, i32 155, i32 156, i32 157, i32 158, i32 159, i32 160, i32 161, i32 162, i32 163, i32 164, i32 165, i32 166, i32 167, i32 168, i32 169, i32 170, i32 171, i32 172, i32 173, i32 174, i32 175, i32 176, i32 177, i32 178, i32 179, i32 180, i32 181, i32 182, i32 183, i32 184, i32 185, i32 186, i32 187, i32 188, i32 189, i32 190, i32 191, i32 192, i32 193, i32 194, i32 195, i32 196, i32 197, i32 198, i32 199, i32 200, i32 201, i32 202, i32 203, i32 204, i32 205, i32 206, i32 207, i32 208, i32 209, i32 210, i32 211, i32 212, i32 213, i32 214, i32 215, i32 216, i32 217, i32 218, i32 219, i32 220, i32 221, i32 222, i32 223, i32 224, i32 225, i32 226, i32 227, i32 228, i32 229, i32 230, i32 231, i32 232, i32 233, i32 234, i32 235, i32 236, i32 237, i32 238, i32 239, i32 240, i32 241, i32 242, i32 243, i32 244, i32 245, i32 246, i32 247, i32 248, i32 249, i32 250, i32 251, i32 252, i32 253, i32 254, i32 255, i32 256, i32 257, i32 258, i32 259, i32 260, i32 261, i32 262, i32 263, i32 264, i32 265, i32 266, i32 267, i32 268, i32 269, i32 270, i32 271, i32 272, i32 273, i32 274, i32 275, i32 276, i32 277, i32 278, i32 279, i32 280, i32 281, i32 282, i32 283, i32 284, i32 285, i32 286, i32 287, i32 288, i32 289, i32 290, i32 291, i32 292, i32 293, i32 294, i32 295, i32 296, i32 297, i32 298, i32 299, i32 300, i32 301, i32 302, i32 303, i32 304, i32 305, i32 306, i32 307, i32 308, i32 309, i32 310, i32 311, i32 312, i32 313, i32 314, i32 315, i32 316, i32 317, i32 318, i32 319>9622 %interleaved.vec = shufflevector <320 x i32> %5, <320 x i32> poison, <320 x i32> <i32 0, i32 64, i32 128, i32 192, i32 256, i32 1, i32 65, i32 129, i32 193, i32 257, i32 2, i32 66, i32 130, i32 194, i32 258, i32 3, i32 67, i32 131, i32 195, i32 259, i32 4, i32 68, i32 132, i32 196, i32 260, i32 5, i32 69, i32 133, i32 197, i32 261, i32 6, i32 70, i32 134, i32 198, i32 262, i32 7, i32 71, i32 135, i32 199, i32 263, i32 8, i32 72, i32 136, i32 200, i32 264, i32 9, i32 73, i32 137, i32 201, i32 265, i32 10, i32 74, i32 138, i32 202, i32 266, i32 11, i32 75, i32 139, i32 203, i32 267, i32 12, i32 76, i32 140, i32 204, i32 268, i32 13, i32 77, i32 141, i32 205, i32 269, i32 14, i32 78, i32 142, i32 206, i32 270, i32 15, i32 79, i32 143, i32 207, i32 271, i32 16, i32 80, i32 144, i32 208, i32 272, i32 17, i32 81, i32 145, i32 209, i32 273, i32 18, i32 82, i32 146, i32 210, i32 274, i32 19, i32 83, i32 147, i32 211, i32 275, i32 20, i32 84, i32 148, i32 212, i32 276, i32 21, i32 85, i32 149, i32 213, i32 277, i32 22, i32 86, i32 150, i32 214, i32 278, i32 23, i32 87, i32 151, i32 215, i32 279, i32 24, i32 88, i32 152, i32 216, i32 280, i32 25, i32 89, i32 153, i32 217, i32 281, i32 26, i32 90, i32 154, i32 218, i32 282, i32 27, i32 91, i32 155, i32 219, i32 283, i32 28, i32 92, i32 156, i32 220, i32 284, i32 29, i32 93, i32 157, i32 221, i32 285, i32 30, i32 94, i32 158, i32 222, i32 286, i32 31, i32 95, i32 159, i32 223, i32 287, i32 32, i32 96, i32 160, i32 224, i32 288, i32 33, i32 97, i32 161, i32 225, i32 289, i32 34, i32 98, i32 162, i32 226, i32 290, i32 35, i32 99, i32 163, i32 227, i32 291, i32 36, i32 100, i32 164, i32 228, i32 292, i32 37, i32 101, i32 165, i32 229, i32 293, i32 38, i32 102, i32 166, i32 230, i32 294, i32 39, i32 103, i32 167, i32 231, i32 295, i32 40, i32 104, i32 168, i32 232, i32 296, i32 41, i32 105, i32 169, i32 233, i32 297, i32 42, i32 106, i32 170, i32 234, i32 298, i32 43, i32 107, i32 171, i32 235, i32 299, i32 44, i32 108, i32 172, i32 236, i32 300, i32 45, i32 109, i32 173, i32 237, i32 301, i32 46, i32 110, i32 174, i32 238, i32 302, i32 47, i32 111, i32 175, i32 239, i32 303, i32 48, i32 112, i32 176, i32 240, i32 304, i32 49, i32 113, i32 177, i32 241, i32 305, i32 50, i32 114, i32 178, i32 242, i32 306, i32 51, i32 115, i32 179, i32 243, i32 307, i32 52, i32 116, i32 180, i32 244, i32 308, i32 53, i32 117, i32 181, i32 245, i32 309, i32 54, i32 118, i32 182, i32 246, i32 310, i32 55, i32 119, i32 183, i32 247, i32 311, i32 56, i32 120, i32 184, i32 248, i32 312, i32 57, i32 121, i32 185, i32 249, i32 313, i32 58, i32 122, i32 186, i32 250, i32 314, i32 59, i32 123, i32 187, i32 251, i32 315, i32 60, i32 124, i32 188, i32 252, i32 316, i32 61, i32 125, i32 189, i32 253, i32 317, i32 62, i32 126, i32 190, i32 254, i32 318, i32 63, i32 127, i32 191, i32 255, i32 319>9623 store <320 x i32> %interleaved.vec, ptr %out.vec, align 649624 ret void9625}9626