14668 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 loads.17 18define void @load_i64_stride5_vf2(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr %out.vec2, ptr %out.vec3, ptr %out.vec4) nounwind {19; SSE-LABEL: load_i64_stride5_vf2:20; SSE: # %bb.0:21; SSE-NEXT: movapd 64(%rdi), %xmm022; SSE-NEXT: movapd (%rdi), %xmm123; SSE-NEXT: movapd 16(%rdi), %xmm224; SSE-NEXT: movapd 32(%rdi), %xmm325; SSE-NEXT: movapd 48(%rdi), %xmm426; SSE-NEXT: movapd %xmm3, %xmm527; SSE-NEXT: movsd {{.*#+}} xmm5 = xmm1[0],xmm5[1]28; SSE-NEXT: shufpd {{.*#+}} xmm1 = xmm1[1],xmm4[0]29; SSE-NEXT: movsd {{.*#+}} xmm4 = xmm2[0],xmm4[1]30; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm0[0]31; SSE-NEXT: movsd {{.*#+}} xmm0 = xmm3[0],xmm0[1]32; SSE-NEXT: movapd %xmm5, (%rsi)33; SSE-NEXT: movapd %xmm1, (%rdx)34; SSE-NEXT: movapd %xmm4, (%rcx)35; SSE-NEXT: movapd %xmm2, (%r8)36; SSE-NEXT: movapd %xmm0, (%r9)37; SSE-NEXT: retq38;39; AVX-LABEL: load_i64_stride5_vf2:40; AVX: # %bb.0:41; AVX-NEXT: vmovdqa (%rdi), %xmm042; AVX-NEXT: vmovdqa 16(%rdi), %xmm143; AVX-NEXT: vmovdqa 32(%rdi), %xmm244; AVX-NEXT: vmovdqa 48(%rdi), %xmm345; AVX-NEXT: vpblendw {{.*#+}} xmm4 = xmm0[0,1,2,3],xmm2[4,5,6,7]46; AVX-NEXT: vpalignr {{.*#+}} xmm0 = xmm0[8,9,10,11,12,13,14,15],xmm3[0,1,2,3,4,5,6,7]47; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm1[0,1,2,3],xmm3[4,5,6,7]48; AVX-NEXT: vmovdqa 64(%rdi), %xmm549; AVX-NEXT: vpalignr {{.*#+}} xmm1 = xmm1[8,9,10,11,12,13,14,15],xmm5[0,1,2,3,4,5,6,7]50; AVX-NEXT: vpblendw {{.*#+}} xmm2 = xmm2[0,1,2,3],xmm5[4,5,6,7]51; AVX-NEXT: vmovdqa %xmm4, (%rsi)52; AVX-NEXT: vmovdqa %xmm0, (%rdx)53; AVX-NEXT: vmovdqa %xmm3, (%rcx)54; AVX-NEXT: vmovdqa %xmm1, (%r8)55; AVX-NEXT: vmovdqa %xmm2, (%r9)56; AVX-NEXT: retq57;58; AVX2-LABEL: load_i64_stride5_vf2:59; AVX2: # %bb.0:60; AVX2-NEXT: vmovdqa (%rdi), %xmm061; AVX2-NEXT: vmovaps 16(%rdi), %xmm162; AVX2-NEXT: vmovdqa 32(%rdi), %xmm263; AVX2-NEXT: vmovdqa 48(%rdi), %xmm364; AVX2-NEXT: vpblendd {{.*#+}} xmm4 = xmm0[0,1],xmm2[2,3]65; AVX2-NEXT: vpalignr {{.*#+}} xmm0 = xmm0[8,9,10,11,12,13,14,15],xmm3[0,1,2,3,4,5,6,7]66; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]67; AVX2-NEXT: vmovdqa 64(%rdi), %xmm368; AVX2-NEXT: vpalignr {{.*#+}} xmm5 = mem[8,9,10,11,12,13,14,15],xmm3[0,1,2,3,4,5,6,7]69; AVX2-NEXT: vpblendd {{.*#+}} xmm2 = xmm2[0,1],xmm3[2,3]70; AVX2-NEXT: vmovdqa %xmm4, (%rsi)71; AVX2-NEXT: vmovdqa %xmm0, (%rdx)72; AVX2-NEXT: vmovaps %xmm1, (%rcx)73; AVX2-NEXT: vmovdqa %xmm5, (%r8)74; AVX2-NEXT: vmovdqa %xmm2, (%r9)75; AVX2-NEXT: retq76;77; AVX2-FP-LABEL: load_i64_stride5_vf2:78; AVX2-FP: # %bb.0:79; AVX2-FP-NEXT: vmovdqa (%rdi), %xmm080; AVX2-FP-NEXT: vmovaps 16(%rdi), %xmm181; AVX2-FP-NEXT: vmovdqa 32(%rdi), %xmm282; AVX2-FP-NEXT: vmovdqa 48(%rdi), %xmm383; AVX2-FP-NEXT: vpblendd {{.*#+}} xmm4 = xmm0[0,1],xmm2[2,3]84; AVX2-FP-NEXT: vpalignr {{.*#+}} xmm0 = xmm0[8,9,10,11,12,13,14,15],xmm3[0,1,2,3,4,5,6,7]85; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]86; AVX2-FP-NEXT: vmovdqa 64(%rdi), %xmm387; AVX2-FP-NEXT: vpalignr {{.*#+}} xmm5 = mem[8,9,10,11,12,13,14,15],xmm3[0,1,2,3,4,5,6,7]88; AVX2-FP-NEXT: vpblendd {{.*#+}} xmm2 = xmm2[0,1],xmm3[2,3]89; AVX2-FP-NEXT: vmovdqa %xmm4, (%rsi)90; AVX2-FP-NEXT: vmovdqa %xmm0, (%rdx)91; AVX2-FP-NEXT: vmovaps %xmm1, (%rcx)92; AVX2-FP-NEXT: vmovdqa %xmm5, (%r8)93; AVX2-FP-NEXT: vmovdqa %xmm2, (%r9)94; AVX2-FP-NEXT: retq95;96; AVX2-FCP-LABEL: load_i64_stride5_vf2:97; AVX2-FCP: # %bb.0:98; AVX2-FCP-NEXT: vmovdqa (%rdi), %xmm099; AVX2-FCP-NEXT: vmovaps 16(%rdi), %xmm1100; AVX2-FCP-NEXT: vmovdqa 32(%rdi), %xmm2101; AVX2-FCP-NEXT: vmovdqa 48(%rdi), %xmm3102; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm4 = xmm0[0,1],xmm2[2,3]103; AVX2-FCP-NEXT: vpalignr {{.*#+}} xmm0 = xmm0[8,9,10,11,12,13,14,15],xmm3[0,1,2,3,4,5,6,7]104; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]105; AVX2-FCP-NEXT: vmovdqa 64(%rdi), %xmm3106; AVX2-FCP-NEXT: vpalignr {{.*#+}} xmm5 = mem[8,9,10,11,12,13,14,15],xmm3[0,1,2,3,4,5,6,7]107; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm2 = xmm2[0,1],xmm3[2,3]108; AVX2-FCP-NEXT: vmovdqa %xmm4, (%rsi)109; AVX2-FCP-NEXT: vmovdqa %xmm0, (%rdx)110; AVX2-FCP-NEXT: vmovaps %xmm1, (%rcx)111; AVX2-FCP-NEXT: vmovdqa %xmm5, (%r8)112; AVX2-FCP-NEXT: vmovdqa %xmm2, (%r9)113; AVX2-FCP-NEXT: retq114;115; AVX512-LABEL: load_i64_stride5_vf2:116; AVX512: # %bb.0:117; AVX512-NEXT: vmovdqa (%rdi), %xmm0118; AVX512-NEXT: vmovdqa 16(%rdi), %xmm1119; AVX512-NEXT: vmovdqa 32(%rdi), %xmm2120; AVX512-NEXT: vmovdqa 48(%rdi), %xmm3121; AVX512-NEXT: vpblendd {{.*#+}} xmm4 = xmm0[0,1],xmm2[2,3]122; AVX512-NEXT: vpalignr {{.*#+}} xmm0 = xmm0[8,9,10,11,12,13,14,15],xmm3[0,1,2,3,4,5,6,7]123; AVX512-NEXT: vpblendd {{.*#+}} xmm3 = xmm1[0,1],xmm3[2,3]124; AVX512-NEXT: vmovdqa 64(%rdi), %xmm5125; AVX512-NEXT: vpalignr {{.*#+}} xmm1 = xmm1[8,9,10,11,12,13,14,15],xmm5[0,1,2,3,4,5,6,7]126; AVX512-NEXT: vpblendd {{.*#+}} xmm2 = xmm2[0,1],xmm5[2,3]127; AVX512-NEXT: vmovdqa %xmm4, (%rsi)128; AVX512-NEXT: vmovdqa %xmm0, (%rdx)129; AVX512-NEXT: vmovdqa %xmm3, (%rcx)130; AVX512-NEXT: vmovdqa %xmm1, (%r8)131; AVX512-NEXT: vmovdqa %xmm2, (%r9)132; AVX512-NEXT: retq133;134; AVX512-FCP-LABEL: load_i64_stride5_vf2:135; AVX512-FCP: # %bb.0:136; AVX512-FCP-NEXT: vmovdqa (%rdi), %xmm0137; AVX512-FCP-NEXT: vmovdqa 16(%rdi), %xmm1138; AVX512-FCP-NEXT: vmovdqa 32(%rdi), %xmm2139; AVX512-FCP-NEXT: vmovdqa 48(%rdi), %xmm3140; AVX512-FCP-NEXT: vpblendd {{.*#+}} xmm4 = xmm0[0,1],xmm2[2,3]141; AVX512-FCP-NEXT: vpalignr {{.*#+}} xmm0 = xmm0[8,9,10,11,12,13,14,15],xmm3[0,1,2,3,4,5,6,7]142; AVX512-FCP-NEXT: vpblendd {{.*#+}} xmm3 = xmm1[0,1],xmm3[2,3]143; AVX512-FCP-NEXT: vmovdqa 64(%rdi), %xmm5144; AVX512-FCP-NEXT: vpalignr {{.*#+}} xmm1 = xmm1[8,9,10,11,12,13,14,15],xmm5[0,1,2,3,4,5,6,7]145; AVX512-FCP-NEXT: vpblendd {{.*#+}} xmm2 = xmm2[0,1],xmm5[2,3]146; AVX512-FCP-NEXT: vmovdqa %xmm4, (%rsi)147; AVX512-FCP-NEXT: vmovdqa %xmm0, (%rdx)148; AVX512-FCP-NEXT: vmovdqa %xmm3, (%rcx)149; AVX512-FCP-NEXT: vmovdqa %xmm1, (%r8)150; AVX512-FCP-NEXT: vmovdqa %xmm2, (%r9)151; AVX512-FCP-NEXT: retq152;153; AVX512DQ-LABEL: load_i64_stride5_vf2:154; AVX512DQ: # %bb.0:155; AVX512DQ-NEXT: vmovdqa (%rdi), %xmm0156; AVX512DQ-NEXT: vmovdqa 16(%rdi), %xmm1157; AVX512DQ-NEXT: vmovdqa 32(%rdi), %xmm2158; AVX512DQ-NEXT: vmovdqa 48(%rdi), %xmm3159; AVX512DQ-NEXT: vpblendd {{.*#+}} xmm4 = xmm0[0,1],xmm2[2,3]160; AVX512DQ-NEXT: vpalignr {{.*#+}} xmm0 = xmm0[8,9,10,11,12,13,14,15],xmm3[0,1,2,3,4,5,6,7]161; AVX512DQ-NEXT: vpblendd {{.*#+}} xmm3 = xmm1[0,1],xmm3[2,3]162; AVX512DQ-NEXT: vmovdqa 64(%rdi), %xmm5163; AVX512DQ-NEXT: vpalignr {{.*#+}} xmm1 = xmm1[8,9,10,11,12,13,14,15],xmm5[0,1,2,3,4,5,6,7]164; AVX512DQ-NEXT: vpblendd {{.*#+}} xmm2 = xmm2[0,1],xmm5[2,3]165; AVX512DQ-NEXT: vmovdqa %xmm4, (%rsi)166; AVX512DQ-NEXT: vmovdqa %xmm0, (%rdx)167; AVX512DQ-NEXT: vmovdqa %xmm3, (%rcx)168; AVX512DQ-NEXT: vmovdqa %xmm1, (%r8)169; AVX512DQ-NEXT: vmovdqa %xmm2, (%r9)170; AVX512DQ-NEXT: retq171;172; AVX512DQ-FCP-LABEL: load_i64_stride5_vf2:173; AVX512DQ-FCP: # %bb.0:174; AVX512DQ-FCP-NEXT: vmovdqa (%rdi), %xmm0175; AVX512DQ-FCP-NEXT: vmovdqa 16(%rdi), %xmm1176; AVX512DQ-FCP-NEXT: vmovdqa 32(%rdi), %xmm2177; AVX512DQ-FCP-NEXT: vmovdqa 48(%rdi), %xmm3178; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} xmm4 = xmm0[0,1],xmm2[2,3]179; AVX512DQ-FCP-NEXT: vpalignr {{.*#+}} xmm0 = xmm0[8,9,10,11,12,13,14,15],xmm3[0,1,2,3,4,5,6,7]180; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} xmm3 = xmm1[0,1],xmm3[2,3]181; AVX512DQ-FCP-NEXT: vmovdqa 64(%rdi), %xmm5182; AVX512DQ-FCP-NEXT: vpalignr {{.*#+}} xmm1 = xmm1[8,9,10,11,12,13,14,15],xmm5[0,1,2,3,4,5,6,7]183; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} xmm2 = xmm2[0,1],xmm5[2,3]184; AVX512DQ-FCP-NEXT: vmovdqa %xmm4, (%rsi)185; AVX512DQ-FCP-NEXT: vmovdqa %xmm0, (%rdx)186; AVX512DQ-FCP-NEXT: vmovdqa %xmm3, (%rcx)187; AVX512DQ-FCP-NEXT: vmovdqa %xmm1, (%r8)188; AVX512DQ-FCP-NEXT: vmovdqa %xmm2, (%r9)189; AVX512DQ-FCP-NEXT: retq190;191; AVX512BW-LABEL: load_i64_stride5_vf2:192; AVX512BW: # %bb.0:193; AVX512BW-NEXT: vmovdqa (%rdi), %xmm0194; AVX512BW-NEXT: vmovdqa 16(%rdi), %xmm1195; AVX512BW-NEXT: vmovdqa 32(%rdi), %xmm2196; AVX512BW-NEXT: vmovdqa 48(%rdi), %xmm3197; AVX512BW-NEXT: vpblendd {{.*#+}} xmm4 = xmm0[0,1],xmm2[2,3]198; AVX512BW-NEXT: vpalignr {{.*#+}} xmm0 = xmm0[8,9,10,11,12,13,14,15],xmm3[0,1,2,3,4,5,6,7]199; AVX512BW-NEXT: vpblendd {{.*#+}} xmm3 = xmm1[0,1],xmm3[2,3]200; AVX512BW-NEXT: vmovdqa 64(%rdi), %xmm5201; AVX512BW-NEXT: vpalignr {{.*#+}} xmm1 = xmm1[8,9,10,11,12,13,14,15],xmm5[0,1,2,3,4,5,6,7]202; AVX512BW-NEXT: vpblendd {{.*#+}} xmm2 = xmm2[0,1],xmm5[2,3]203; AVX512BW-NEXT: vmovdqa %xmm4, (%rsi)204; AVX512BW-NEXT: vmovdqa %xmm0, (%rdx)205; AVX512BW-NEXT: vmovdqa %xmm3, (%rcx)206; AVX512BW-NEXT: vmovdqa %xmm1, (%r8)207; AVX512BW-NEXT: vmovdqa %xmm2, (%r9)208; AVX512BW-NEXT: retq209;210; AVX512BW-FCP-LABEL: load_i64_stride5_vf2:211; AVX512BW-FCP: # %bb.0:212; AVX512BW-FCP-NEXT: vmovdqa (%rdi), %xmm0213; AVX512BW-FCP-NEXT: vmovdqa 16(%rdi), %xmm1214; AVX512BW-FCP-NEXT: vmovdqa 32(%rdi), %xmm2215; AVX512BW-FCP-NEXT: vmovdqa 48(%rdi), %xmm3216; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} xmm4 = xmm0[0,1],xmm2[2,3]217; AVX512BW-FCP-NEXT: vpalignr {{.*#+}} xmm0 = xmm0[8,9,10,11,12,13,14,15],xmm3[0,1,2,3,4,5,6,7]218; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} xmm3 = xmm1[0,1],xmm3[2,3]219; AVX512BW-FCP-NEXT: vmovdqa 64(%rdi), %xmm5220; AVX512BW-FCP-NEXT: vpalignr {{.*#+}} xmm1 = xmm1[8,9,10,11,12,13,14,15],xmm5[0,1,2,3,4,5,6,7]221; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} xmm2 = xmm2[0,1],xmm5[2,3]222; AVX512BW-FCP-NEXT: vmovdqa %xmm4, (%rsi)223; AVX512BW-FCP-NEXT: vmovdqa %xmm0, (%rdx)224; AVX512BW-FCP-NEXT: vmovdqa %xmm3, (%rcx)225; AVX512BW-FCP-NEXT: vmovdqa %xmm1, (%r8)226; AVX512BW-FCP-NEXT: vmovdqa %xmm2, (%r9)227; AVX512BW-FCP-NEXT: retq228;229; AVX512DQ-BW-LABEL: load_i64_stride5_vf2:230; AVX512DQ-BW: # %bb.0:231; AVX512DQ-BW-NEXT: vmovdqa (%rdi), %xmm0232; AVX512DQ-BW-NEXT: vmovdqa 16(%rdi), %xmm1233; AVX512DQ-BW-NEXT: vmovdqa 32(%rdi), %xmm2234; AVX512DQ-BW-NEXT: vmovdqa 48(%rdi), %xmm3235; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm4 = xmm0[0,1],xmm2[2,3]236; AVX512DQ-BW-NEXT: vpalignr {{.*#+}} xmm0 = xmm0[8,9,10,11,12,13,14,15],xmm3[0,1,2,3,4,5,6,7]237; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm3 = xmm1[0,1],xmm3[2,3]238; AVX512DQ-BW-NEXT: vmovdqa 64(%rdi), %xmm5239; AVX512DQ-BW-NEXT: vpalignr {{.*#+}} xmm1 = xmm1[8,9,10,11,12,13,14,15],xmm5[0,1,2,3,4,5,6,7]240; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} xmm2 = xmm2[0,1],xmm5[2,3]241; AVX512DQ-BW-NEXT: vmovdqa %xmm4, (%rsi)242; AVX512DQ-BW-NEXT: vmovdqa %xmm0, (%rdx)243; AVX512DQ-BW-NEXT: vmovdqa %xmm3, (%rcx)244; AVX512DQ-BW-NEXT: vmovdqa %xmm1, (%r8)245; AVX512DQ-BW-NEXT: vmovdqa %xmm2, (%r9)246; AVX512DQ-BW-NEXT: retq247;248; AVX512DQ-BW-FCP-LABEL: load_i64_stride5_vf2:249; AVX512DQ-BW-FCP: # %bb.0:250; AVX512DQ-BW-FCP-NEXT: vmovdqa (%rdi), %xmm0251; AVX512DQ-BW-FCP-NEXT: vmovdqa 16(%rdi), %xmm1252; AVX512DQ-BW-FCP-NEXT: vmovdqa 32(%rdi), %xmm2253; AVX512DQ-BW-FCP-NEXT: vmovdqa 48(%rdi), %xmm3254; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} xmm4 = xmm0[0,1],xmm2[2,3]255; AVX512DQ-BW-FCP-NEXT: vpalignr {{.*#+}} xmm0 = xmm0[8,9,10,11,12,13,14,15],xmm3[0,1,2,3,4,5,6,7]256; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} xmm3 = xmm1[0,1],xmm3[2,3]257; AVX512DQ-BW-FCP-NEXT: vmovdqa 64(%rdi), %xmm5258; AVX512DQ-BW-FCP-NEXT: vpalignr {{.*#+}} xmm1 = xmm1[8,9,10,11,12,13,14,15],xmm5[0,1,2,3,4,5,6,7]259; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} xmm2 = xmm2[0,1],xmm5[2,3]260; AVX512DQ-BW-FCP-NEXT: vmovdqa %xmm4, (%rsi)261; AVX512DQ-BW-FCP-NEXT: vmovdqa %xmm0, (%rdx)262; AVX512DQ-BW-FCP-NEXT: vmovdqa %xmm3, (%rcx)263; AVX512DQ-BW-FCP-NEXT: vmovdqa %xmm1, (%r8)264; AVX512DQ-BW-FCP-NEXT: vmovdqa %xmm2, (%r9)265; AVX512DQ-BW-FCP-NEXT: retq266 %wide.vec = load <10 x i64>, ptr %in.vec, align 64267 %strided.vec0 = shufflevector <10 x i64> %wide.vec, <10 x i64> poison, <2 x i32> <i32 0, i32 5>268 %strided.vec1 = shufflevector <10 x i64> %wide.vec, <10 x i64> poison, <2 x i32> <i32 1, i32 6>269 %strided.vec2 = shufflevector <10 x i64> %wide.vec, <10 x i64> poison, <2 x i32> <i32 2, i32 7>270 %strided.vec3 = shufflevector <10 x i64> %wide.vec, <10 x i64> poison, <2 x i32> <i32 3, i32 8>271 %strided.vec4 = shufflevector <10 x i64> %wide.vec, <10 x i64> poison, <2 x i32> <i32 4, i32 9>272 store <2 x i64> %strided.vec0, ptr %out.vec0, align 64273 store <2 x i64> %strided.vec1, ptr %out.vec1, align 64274 store <2 x i64> %strided.vec2, ptr %out.vec2, align 64275 store <2 x i64> %strided.vec3, ptr %out.vec3, align 64276 store <2 x i64> %strided.vec4, ptr %out.vec4, align 64277 ret void278}279 280define void @load_i64_stride5_vf4(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr %out.vec2, ptr %out.vec3, ptr %out.vec4) nounwind {281; SSE-LABEL: load_i64_stride5_vf4:282; SSE: # %bb.0:283; SSE-NEXT: movapd 144(%rdi), %xmm1284; SSE-NEXT: movapd 64(%rdi), %xmm0285; SSE-NEXT: movapd 96(%rdi), %xmm2286; SSE-NEXT: movapd 128(%rdi), %xmm3287; SSE-NEXT: movapd (%rdi), %xmm4288; SSE-NEXT: movapd 16(%rdi), %xmm5289; SSE-NEXT: movapd 32(%rdi), %xmm6290; SSE-NEXT: movapd 48(%rdi), %xmm7291; SSE-NEXT: movapd 80(%rdi), %xmm8292; SSE-NEXT: movapd 112(%rdi), %xmm9293; SSE-NEXT: movapd %xmm9, %xmm10294; SSE-NEXT: movsd {{.*#+}} xmm10 = xmm8[0],xmm10[1]295; SSE-NEXT: movapd %xmm6, %xmm11296; SSE-NEXT: movsd {{.*#+}} xmm11 = xmm4[0],xmm11[1]297; SSE-NEXT: shufpd {{.*#+}} xmm8 = xmm8[1],xmm3[0]298; SSE-NEXT: shufpd {{.*#+}} xmm4 = xmm4[1],xmm7[0]299; SSE-NEXT: movsd {{.*#+}} xmm3 = xmm2[0],xmm3[1]300; SSE-NEXT: movsd {{.*#+}} xmm7 = xmm5[0],xmm7[1]301; SSE-NEXT: shufpd {{.*#+}} xmm5 = xmm5[1],xmm0[0]302; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]303; SSE-NEXT: movsd {{.*#+}} xmm0 = xmm6[0],xmm0[1]304; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm9[0],xmm1[1]305; SSE-NEXT: movapd %xmm10, 16(%rsi)306; SSE-NEXT: movapd %xmm11, (%rsi)307; SSE-NEXT: movapd %xmm8, 16(%rdx)308; SSE-NEXT: movapd %xmm4, (%rdx)309; SSE-NEXT: movapd %xmm3, 16(%rcx)310; SSE-NEXT: movapd %xmm7, (%rcx)311; SSE-NEXT: movapd %xmm2, 16(%r8)312; SSE-NEXT: movapd %xmm5, (%r8)313; SSE-NEXT: movapd %xmm1, 16(%r9)314; SSE-NEXT: movapd %xmm0, (%r9)315; SSE-NEXT: retq316;317; AVX-LABEL: load_i64_stride5_vf4:318; AVX: # %bb.0:319; AVX-NEXT: vmovapd 128(%rdi), %ymm0320; AVX-NEXT: vmovapd 96(%rdi), %ymm1321; AVX-NEXT: vmovapd 64(%rdi), %ymm2322; AVX-NEXT: vblendpd {{.*#+}} ymm3 = ymm2[0,1,2],ymm1[3]323; AVX-NEXT: vmovapd (%rdi), %xmm4324; AVX-NEXT: vmovdqa 16(%rdi), %xmm5325; AVX-NEXT: vmovapd 32(%rdi), %xmm6326; AVX-NEXT: vmovdqa 48(%rdi), %xmm7327; AVX-NEXT: vblendpd {{.*#+}} xmm8 = xmm4[0],xmm6[1]328; AVX-NEXT: vblendpd {{.*#+}} ymm3 = ymm8[0,1],ymm3[2,3]329; AVX-NEXT: vmovaps 128(%rdi), %xmm8330; AVX-NEXT: vinsertf128 $1, %xmm8, %ymm0, %ymm9331; AVX-NEXT: vshufpd {{.*#+}} ymm2 = ymm2[0],ymm9[1],ymm2[3],ymm9[2]332; AVX-NEXT: vpalignr {{.*#+}} xmm4 = xmm4[8,9,10,11,12,13,14,15],xmm7[0,1,2,3,4,5,6,7]333; AVX-NEXT: vblendpd {{.*#+}} ymm2 = ymm4[0,1],ymm2[2,3]334; AVX-NEXT: vmovaps 96(%rdi), %xmm4335; AVX-NEXT: vblendps {{.*#+}} xmm8 = xmm4[0,1],xmm8[2,3]336; AVX-NEXT: vinsertf128 $1, %xmm8, %ymm0, %ymm8337; AVX-NEXT: vpblendw {{.*#+}} xmm7 = xmm5[0,1,2,3],xmm7[4,5,6,7]338; AVX-NEXT: vblendps {{.*#+}} ymm7 = ymm7[0,1,2,3],ymm8[4,5,6,7]339; AVX-NEXT: vinsertf128 $1, %xmm4, %ymm0, %ymm4340; AVX-NEXT: vshufpd {{.*#+}} ymm4 = ymm4[0],ymm0[1],ymm4[3],ymm0[2]341; AVX-NEXT: vmovdqa 64(%rdi), %xmm8342; AVX-NEXT: vpalignr {{.*#+}} xmm5 = xmm5[8,9,10,11,12,13,14,15],xmm8[0,1,2,3,4,5,6,7]343; AVX-NEXT: vblendpd {{.*#+}} ymm4 = ymm5[0,1],ymm4[2,3]344; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm1[0,1,2],ymm0[3]345; AVX-NEXT: vblendpd {{.*#+}} xmm1 = xmm6[0],xmm8[1]346; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3]347; AVX-NEXT: vmovapd %ymm3, (%rsi)348; AVX-NEXT: vmovapd %ymm2, (%rdx)349; AVX-NEXT: vmovaps %ymm7, (%rcx)350; AVX-NEXT: vmovapd %ymm4, (%r8)351; AVX-NEXT: vmovapd %ymm0, (%r9)352; AVX-NEXT: vzeroupper353; AVX-NEXT: retq354;355; AVX2-LABEL: load_i64_stride5_vf4:356; AVX2: # %bb.0:357; AVX2-NEXT: vmovaps (%rdi), %ymm0358; AVX2-NEXT: vmovdqa 128(%rdi), %ymm1359; AVX2-NEXT: vmovdqa 96(%rdi), %ymm2360; AVX2-NEXT: vmovdqa 64(%rdi), %ymm3361; AVX2-NEXT: vpblendd {{.*#+}} ymm4 = ymm3[0,1,2,3,4,5],ymm2[6,7]362; AVX2-NEXT: vmovdqa (%rdi), %xmm5363; AVX2-NEXT: vmovdqa 32(%rdi), %xmm6364; AVX2-NEXT: vmovdqa 48(%rdi), %xmm7365; AVX2-NEXT: vmovdqa 64(%rdi), %xmm8366; AVX2-NEXT: vpblendd {{.*#+}} xmm9 = xmm5[0,1],xmm6[2,3]367; AVX2-NEXT: vpblendd {{.*#+}} ymm4 = ymm9[0,1,2,3],ymm4[4,5,6,7]368; AVX2-NEXT: vpalignr {{.*#+}} xmm5 = xmm5[8,9,10,11,12,13,14,15],xmm7[0,1,2,3,4,5,6,7]369; AVX2-NEXT: vpalignr {{.*#+}} ymm3 = ymm3[8,9,10,11,12,13,14,15],ymm1[0,1,2,3,4,5,6,7],ymm3[24,25,26,27,28,29,30,31],ymm1[16,17,18,19,20,21,22,23]370; AVX2-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,1,2,1]371; AVX2-NEXT: vpblendd {{.*#+}} ymm3 = ymm5[0,1,2,3],ymm3[4,5,6,7]372; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]373; AVX2-NEXT: vmovaps 96(%rdi), %xmm5374; AVX2-NEXT: vblendps {{.*#+}} xmm5 = xmm5[0,1],mem[2,3]375; AVX2-NEXT: vinsertf128 $1, %xmm5, %ymm0, %ymm5376; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm5[2,3]377; AVX2-NEXT: vpalignr {{.*#+}} xmm5 = mem[8,9,10,11,12,13,14,15],xmm8[0,1,2,3,4,5,6,7]378; AVX2-NEXT: vpalignr {{.*#+}} ymm7 = ymm2[8,9,10,11,12,13,14,15],ymm1[0,1,2,3,4,5,6,7],ymm2[24,25,26,27,28,29,30,31],ymm1[16,17,18,19,20,21,22,23]379; AVX2-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,1,0,3]380; AVX2-NEXT: vpblendd {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm7[4,5,6,7]381; AVX2-NEXT: vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]382; AVX2-NEXT: vpblendd {{.*#+}} xmm2 = xmm6[0,1],xmm8[2,3]383; AVX2-NEXT: vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]384; AVX2-NEXT: vmovdqa %ymm4, (%rsi)385; AVX2-NEXT: vmovdqa %ymm3, (%rdx)386; AVX2-NEXT: vmovaps %ymm0, (%rcx)387; AVX2-NEXT: vmovdqa %ymm5, (%r8)388; AVX2-NEXT: vmovdqa %ymm1, (%r9)389; AVX2-NEXT: vzeroupper390; AVX2-NEXT: retq391;392; AVX2-FP-LABEL: load_i64_stride5_vf4:393; AVX2-FP: # %bb.0:394; AVX2-FP-NEXT: vmovaps (%rdi), %ymm0395; AVX2-FP-NEXT: vmovdqa 128(%rdi), %ymm1396; AVX2-FP-NEXT: vmovdqa 96(%rdi), %ymm2397; AVX2-FP-NEXT: vmovdqa 64(%rdi), %ymm3398; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm4 = ymm3[0,1,2,3,4,5],ymm2[6,7]399; AVX2-FP-NEXT: vmovdqa (%rdi), %xmm5400; AVX2-FP-NEXT: vmovdqa 32(%rdi), %xmm6401; AVX2-FP-NEXT: vmovdqa 48(%rdi), %xmm7402; AVX2-FP-NEXT: vmovdqa 64(%rdi), %xmm8403; AVX2-FP-NEXT: vpblendd {{.*#+}} xmm9 = xmm5[0,1],xmm6[2,3]404; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm4 = ymm9[0,1,2,3],ymm4[4,5,6,7]405; AVX2-FP-NEXT: vpalignr {{.*#+}} xmm5 = xmm5[8,9,10,11,12,13,14,15],xmm7[0,1,2,3,4,5,6,7]406; AVX2-FP-NEXT: vpalignr {{.*#+}} ymm3 = ymm3[8,9,10,11,12,13,14,15],ymm1[0,1,2,3,4,5,6,7],ymm3[24,25,26,27,28,29,30,31],ymm1[16,17,18,19,20,21,22,23]407; AVX2-FP-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,1,2,1]408; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm3 = ymm5[0,1,2,3],ymm3[4,5,6,7]409; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]410; AVX2-FP-NEXT: vmovaps 96(%rdi), %xmm5411; AVX2-FP-NEXT: vblendps {{.*#+}} xmm5 = xmm5[0,1],mem[2,3]412; AVX2-FP-NEXT: vinsertf128 $1, %xmm5, %ymm0, %ymm5413; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm5[2,3]414; AVX2-FP-NEXT: vpalignr {{.*#+}} xmm5 = mem[8,9,10,11,12,13,14,15],xmm8[0,1,2,3,4,5,6,7]415; AVX2-FP-NEXT: vpalignr {{.*#+}} ymm7 = ymm2[8,9,10,11,12,13,14,15],ymm1[0,1,2,3,4,5,6,7],ymm2[24,25,26,27,28,29,30,31],ymm1[16,17,18,19,20,21,22,23]416; AVX2-FP-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,1,0,3]417; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm7[4,5,6,7]418; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]419; AVX2-FP-NEXT: vpblendd {{.*#+}} xmm2 = xmm6[0,1],xmm8[2,3]420; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]421; AVX2-FP-NEXT: vmovdqa %ymm4, (%rsi)422; AVX2-FP-NEXT: vmovdqa %ymm3, (%rdx)423; AVX2-FP-NEXT: vmovaps %ymm0, (%rcx)424; AVX2-FP-NEXT: vmovdqa %ymm5, (%r8)425; AVX2-FP-NEXT: vmovdqa %ymm1, (%r9)426; AVX2-FP-NEXT: vzeroupper427; AVX2-FP-NEXT: retq428;429; AVX2-FCP-LABEL: load_i64_stride5_vf4:430; AVX2-FCP: # %bb.0:431; AVX2-FCP-NEXT: vmovaps (%rdi), %ymm0432; AVX2-FCP-NEXT: vmovdqa 128(%rdi), %ymm1433; AVX2-FCP-NEXT: vmovdqa 96(%rdi), %ymm2434; AVX2-FCP-NEXT: vmovdqa 64(%rdi), %ymm3435; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm3[0,1,2,3,4,5],ymm2[6,7]436; AVX2-FCP-NEXT: vmovdqa (%rdi), %xmm5437; AVX2-FCP-NEXT: vmovdqa 32(%rdi), %xmm6438; AVX2-FCP-NEXT: vmovdqa 48(%rdi), %xmm7439; AVX2-FCP-NEXT: vmovdqa 64(%rdi), %xmm8440; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm9 = xmm5[0,1],xmm6[2,3]441; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm9[0,1,2,3],ymm4[4,5,6,7]442; AVX2-FCP-NEXT: vpalignr {{.*#+}} xmm5 = xmm5[8,9,10,11,12,13,14,15],xmm7[0,1,2,3,4,5,6,7]443; AVX2-FCP-NEXT: vpalignr {{.*#+}} ymm3 = ymm3[8,9,10,11,12,13,14,15],ymm1[0,1,2,3,4,5,6,7],ymm3[24,25,26,27,28,29,30,31],ymm1[16,17,18,19,20,21,22,23]444; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,1,2,1]445; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm3 = ymm5[0,1,2,3],ymm3[4,5,6,7]446; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]447; AVX2-FCP-NEXT: vmovaps 96(%rdi), %xmm5448; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm5 = xmm5[0,1],mem[2,3]449; AVX2-FCP-NEXT: vinsertf128 $1, %xmm5, %ymm0, %ymm5450; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm5[2,3]451; AVX2-FCP-NEXT: vpalignr {{.*#+}} xmm5 = mem[8,9,10,11,12,13,14,15],xmm8[0,1,2,3,4,5,6,7]452; AVX2-FCP-NEXT: vpalignr {{.*#+}} ymm7 = ymm2[8,9,10,11,12,13,14,15],ymm1[0,1,2,3,4,5,6,7],ymm2[24,25,26,27,28,29,30,31],ymm1[16,17,18,19,20,21,22,23]453; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,1,0,3]454; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm7[4,5,6,7]455; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5],ymm1[6,7]456; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm2 = xmm6[0,1],xmm8[2,3]457; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]458; AVX2-FCP-NEXT: vmovdqa %ymm4, (%rsi)459; AVX2-FCP-NEXT: vmovdqa %ymm3, (%rdx)460; AVX2-FCP-NEXT: vmovaps %ymm0, (%rcx)461; AVX2-FCP-NEXT: vmovdqa %ymm5, (%r8)462; AVX2-FCP-NEXT: vmovdqa %ymm1, (%r9)463; AVX2-FCP-NEXT: vzeroupper464; AVX2-FCP-NEXT: retq465;466; AVX512-LABEL: load_i64_stride5_vf4:467; AVX512: # %bb.0:468; AVX512-NEXT: vmovdqa64 (%rdi), %zmm0469; AVX512-NEXT: vmovdqa64 64(%rdi), %zmm1470; AVX512-NEXT: vpmovsxbq {{.*#+}} ymm2 = [0,5,10,15]471; AVX512-NEXT: vpermi2q %zmm1, %zmm0, %zmm2472; AVX512-NEXT: vpmovsxbq {{.*#+}} ymm3 = [1,6,11,0]473; AVX512-NEXT: vpermi2q %zmm1, %zmm0, %zmm3474; AVX512-NEXT: vmovdqa 128(%rdi), %xmm4475; AVX512-NEXT: vpbroadcastq %xmm4, %ymm5476; AVX512-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm5[6,7]477; AVX512-NEXT: vinserti128 $1, %xmm4, %ymm0, %ymm4478; AVX512-NEXT: vpmovsxbq {{.*#+}} ymm5 = [2,7,12,0]479; AVX512-NEXT: vpermi2q %zmm1, %zmm0, %zmm5480; AVX512-NEXT: vpblendd {{.*#+}} ymm4 = ymm5[0,1,2,3,4,5],ymm4[6,7]481; AVX512-NEXT: vpmovsxbq {{.*#+}} ymm5 = [11,0,5,0]482; AVX512-NEXT: vpermi2q %zmm0, %zmm1, %zmm5483; AVX512-NEXT: vpbroadcastq 144(%rdi), %ymm6484; AVX512-NEXT: vpblendd {{.*#+}} ymm5 = ymm5[0,1,2,3,4,5],ymm6[6,7]485; AVX512-NEXT: vpmovsxbq {{.*#+}} ymm6 = [12,1,6,0]486; AVX512-NEXT: vpermi2q %zmm0, %zmm1, %zmm6487; AVX512-NEXT: vpblendd {{.*#+}} ymm0 = ymm6[0,1,2,3,4,5],mem[6,7]488; AVX512-NEXT: vmovdqa %ymm2, (%rsi)489; AVX512-NEXT: vmovdqa %ymm3, (%rdx)490; AVX512-NEXT: vmovdqa %ymm4, (%rcx)491; AVX512-NEXT: vmovdqa %ymm5, (%r8)492; AVX512-NEXT: vmovdqa %ymm0, (%r9)493; AVX512-NEXT: vzeroupper494; AVX512-NEXT: retq495;496; AVX512-FCP-LABEL: load_i64_stride5_vf4:497; AVX512-FCP: # %bb.0:498; AVX512-FCP-NEXT: vmovdqa64 (%rdi), %zmm0499; AVX512-FCP-NEXT: vmovdqa64 64(%rdi), %zmm1500; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} ymm2 = [0,5,10,15]501; AVX512-FCP-NEXT: vpermi2q %zmm1, %zmm0, %zmm2502; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} ymm3 = [1,6,11,0]503; AVX512-FCP-NEXT: vpermi2q %zmm1, %zmm0, %zmm3504; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} ymm4 = [0,1,2,4]505; AVX512-FCP-NEXT: vmovdqa 128(%rdi), %ymm5506; AVX512-FCP-NEXT: vpermi2q %ymm5, %ymm3, %ymm4507; AVX512-FCP-NEXT: vinserti128 $1, 128(%rdi), %ymm0, %ymm3508; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} ymm6 = [2,7,12,0]509; AVX512-FCP-NEXT: vpermi2q %zmm1, %zmm0, %zmm6510; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm3 = ymm6[0,1,2,3,4,5],ymm3[6,7]511; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} ymm6 = [11,0,5,0]512; AVX512-FCP-NEXT: vpermi2q %zmm0, %zmm1, %zmm6513; AVX512-FCP-NEXT: vpbroadcastq 144(%rdi), %ymm7514; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm6[0,1,2,3,4,5],ymm7[6,7]515; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} ymm7 = [12,1,6,0]516; AVX512-FCP-NEXT: vpermi2q %zmm0, %zmm1, %zmm7517; AVX512-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm7[0,1,2,3,4,5],ymm5[6,7]518; AVX512-FCP-NEXT: vmovdqa %ymm2, (%rsi)519; AVX512-FCP-NEXT: vmovdqa %ymm4, (%rdx)520; AVX512-FCP-NEXT: vmovdqa %ymm3, (%rcx)521; AVX512-FCP-NEXT: vmovdqa %ymm6, (%r8)522; AVX512-FCP-NEXT: vmovdqa %ymm0, (%r9)523; AVX512-FCP-NEXT: vzeroupper524; AVX512-FCP-NEXT: retq525;526; AVX512DQ-LABEL: load_i64_stride5_vf4:527; AVX512DQ: # %bb.0:528; AVX512DQ-NEXT: vmovdqa64 (%rdi), %zmm0529; AVX512DQ-NEXT: vmovdqa64 64(%rdi), %zmm1530; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} ymm2 = [0,5,10,15]531; AVX512DQ-NEXT: vpermi2q %zmm1, %zmm0, %zmm2532; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} ymm3 = [1,6,11,0]533; AVX512DQ-NEXT: vpermi2q %zmm1, %zmm0, %zmm3534; AVX512DQ-NEXT: vmovdqa 128(%rdi), %xmm4535; AVX512DQ-NEXT: vpbroadcastq %xmm4, %ymm5536; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm5[6,7]537; AVX512DQ-NEXT: vinserti128 $1, %xmm4, %ymm0, %ymm4538; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} ymm5 = [2,7,12,0]539; AVX512DQ-NEXT: vpermi2q %zmm1, %zmm0, %zmm5540; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm4 = ymm5[0,1,2,3,4,5],ymm4[6,7]541; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} ymm5 = [11,0,5,0]542; AVX512DQ-NEXT: vpermi2q %zmm0, %zmm1, %zmm5543; AVX512DQ-NEXT: vpbroadcastq 144(%rdi), %ymm6544; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm5 = ymm5[0,1,2,3,4,5],ymm6[6,7]545; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} ymm6 = [12,1,6,0]546; AVX512DQ-NEXT: vpermi2q %zmm0, %zmm1, %zmm6547; AVX512DQ-NEXT: vpblendd {{.*#+}} ymm0 = ymm6[0,1,2,3,4,5],mem[6,7]548; AVX512DQ-NEXT: vmovdqa %ymm2, (%rsi)549; AVX512DQ-NEXT: vmovdqa %ymm3, (%rdx)550; AVX512DQ-NEXT: vmovdqa %ymm4, (%rcx)551; AVX512DQ-NEXT: vmovdqa %ymm5, (%r8)552; AVX512DQ-NEXT: vmovdqa %ymm0, (%r9)553; AVX512DQ-NEXT: vzeroupper554; AVX512DQ-NEXT: retq555;556; AVX512DQ-FCP-LABEL: load_i64_stride5_vf4:557; AVX512DQ-FCP: # %bb.0:558; AVX512DQ-FCP-NEXT: vmovdqa64 (%rdi), %zmm0559; AVX512DQ-FCP-NEXT: vmovdqa64 64(%rdi), %zmm1560; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} ymm2 = [0,5,10,15]561; AVX512DQ-FCP-NEXT: vpermi2q %zmm1, %zmm0, %zmm2562; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} ymm3 = [1,6,11,0]563; AVX512DQ-FCP-NEXT: vpermi2q %zmm1, %zmm0, %zmm3564; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} ymm4 = [0,1,2,4]565; AVX512DQ-FCP-NEXT: vmovdqa 128(%rdi), %ymm5566; AVX512DQ-FCP-NEXT: vpermi2q %ymm5, %ymm3, %ymm4567; AVX512DQ-FCP-NEXT: vinserti128 $1, 128(%rdi), %ymm0, %ymm3568; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} ymm6 = [2,7,12,0]569; AVX512DQ-FCP-NEXT: vpermi2q %zmm1, %zmm0, %zmm6570; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm3 = ymm6[0,1,2,3,4,5],ymm3[6,7]571; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} ymm6 = [11,0,5,0]572; AVX512DQ-FCP-NEXT: vpermi2q %zmm0, %zmm1, %zmm6573; AVX512DQ-FCP-NEXT: vpbroadcastq 144(%rdi), %ymm7574; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm6[0,1,2,3,4,5],ymm7[6,7]575; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} ymm7 = [12,1,6,0]576; AVX512DQ-FCP-NEXT: vpermi2q %zmm0, %zmm1, %zmm7577; AVX512DQ-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm7[0,1,2,3,4,5],ymm5[6,7]578; AVX512DQ-FCP-NEXT: vmovdqa %ymm2, (%rsi)579; AVX512DQ-FCP-NEXT: vmovdqa %ymm4, (%rdx)580; AVX512DQ-FCP-NEXT: vmovdqa %ymm3, (%rcx)581; AVX512DQ-FCP-NEXT: vmovdqa %ymm6, (%r8)582; AVX512DQ-FCP-NEXT: vmovdqa %ymm0, (%r9)583; AVX512DQ-FCP-NEXT: vzeroupper584; AVX512DQ-FCP-NEXT: retq585;586; AVX512BW-LABEL: load_i64_stride5_vf4:587; AVX512BW: # %bb.0:588; AVX512BW-NEXT: vmovdqa64 (%rdi), %zmm0589; AVX512BW-NEXT: vmovdqa64 64(%rdi), %zmm1590; AVX512BW-NEXT: vpmovsxbq {{.*#+}} ymm2 = [0,5,10,15]591; AVX512BW-NEXT: vpermi2q %zmm1, %zmm0, %zmm2592; AVX512BW-NEXT: vpmovsxbq {{.*#+}} ymm3 = [1,6,11,0]593; AVX512BW-NEXT: vpermi2q %zmm1, %zmm0, %zmm3594; AVX512BW-NEXT: vmovdqa 128(%rdi), %xmm4595; AVX512BW-NEXT: vpbroadcastq %xmm4, %ymm5596; AVX512BW-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm5[6,7]597; AVX512BW-NEXT: vinserti128 $1, %xmm4, %ymm0, %ymm4598; AVX512BW-NEXT: vpmovsxbq {{.*#+}} ymm5 = [2,7,12,0]599; AVX512BW-NEXT: vpermi2q %zmm1, %zmm0, %zmm5600; AVX512BW-NEXT: vpblendd {{.*#+}} ymm4 = ymm5[0,1,2,3,4,5],ymm4[6,7]601; AVX512BW-NEXT: vpmovsxbq {{.*#+}} ymm5 = [11,0,5,0]602; AVX512BW-NEXT: vpermi2q %zmm0, %zmm1, %zmm5603; AVX512BW-NEXT: vpbroadcastq 144(%rdi), %ymm6604; AVX512BW-NEXT: vpblendd {{.*#+}} ymm5 = ymm5[0,1,2,3,4,5],ymm6[6,7]605; AVX512BW-NEXT: vpmovsxbq {{.*#+}} ymm6 = [12,1,6,0]606; AVX512BW-NEXT: vpermi2q %zmm0, %zmm1, %zmm6607; AVX512BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm6[0,1,2,3,4,5],mem[6,7]608; AVX512BW-NEXT: vmovdqa %ymm2, (%rsi)609; AVX512BW-NEXT: vmovdqa %ymm3, (%rdx)610; AVX512BW-NEXT: vmovdqa %ymm4, (%rcx)611; AVX512BW-NEXT: vmovdqa %ymm5, (%r8)612; AVX512BW-NEXT: vmovdqa %ymm0, (%r9)613; AVX512BW-NEXT: vzeroupper614; AVX512BW-NEXT: retq615;616; AVX512BW-FCP-LABEL: load_i64_stride5_vf4:617; AVX512BW-FCP: # %bb.0:618; AVX512BW-FCP-NEXT: vmovdqa64 (%rdi), %zmm0619; AVX512BW-FCP-NEXT: vmovdqa64 64(%rdi), %zmm1620; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} ymm2 = [0,5,10,15]621; AVX512BW-FCP-NEXT: vpermi2q %zmm1, %zmm0, %zmm2622; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} ymm3 = [1,6,11,0]623; AVX512BW-FCP-NEXT: vpermi2q %zmm1, %zmm0, %zmm3624; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} ymm4 = [0,1,2,4]625; AVX512BW-FCP-NEXT: vmovdqa 128(%rdi), %ymm5626; AVX512BW-FCP-NEXT: vpermi2q %ymm5, %ymm3, %ymm4627; AVX512BW-FCP-NEXT: vinserti128 $1, 128(%rdi), %ymm0, %ymm3628; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} ymm6 = [2,7,12,0]629; AVX512BW-FCP-NEXT: vpermi2q %zmm1, %zmm0, %zmm6630; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm3 = ymm6[0,1,2,3,4,5],ymm3[6,7]631; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} ymm6 = [11,0,5,0]632; AVX512BW-FCP-NEXT: vpermi2q %zmm0, %zmm1, %zmm6633; AVX512BW-FCP-NEXT: vpbroadcastq 144(%rdi), %ymm7634; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm6[0,1,2,3,4,5],ymm7[6,7]635; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} ymm7 = [12,1,6,0]636; AVX512BW-FCP-NEXT: vpermi2q %zmm0, %zmm1, %zmm7637; AVX512BW-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm7[0,1,2,3,4,5],ymm5[6,7]638; AVX512BW-FCP-NEXT: vmovdqa %ymm2, (%rsi)639; AVX512BW-FCP-NEXT: vmovdqa %ymm4, (%rdx)640; AVX512BW-FCP-NEXT: vmovdqa %ymm3, (%rcx)641; AVX512BW-FCP-NEXT: vmovdqa %ymm6, (%r8)642; AVX512BW-FCP-NEXT: vmovdqa %ymm0, (%r9)643; AVX512BW-FCP-NEXT: vzeroupper644; AVX512BW-FCP-NEXT: retq645;646; AVX512DQ-BW-LABEL: load_i64_stride5_vf4:647; AVX512DQ-BW: # %bb.0:648; AVX512DQ-BW-NEXT: vmovdqa64 (%rdi), %zmm0649; AVX512DQ-BW-NEXT: vmovdqa64 64(%rdi), %zmm1650; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} ymm2 = [0,5,10,15]651; AVX512DQ-BW-NEXT: vpermi2q %zmm1, %zmm0, %zmm2652; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} ymm3 = [1,6,11,0]653; AVX512DQ-BW-NEXT: vpermi2q %zmm1, %zmm0, %zmm3654; AVX512DQ-BW-NEXT: vmovdqa 128(%rdi), %xmm4655; AVX512DQ-BW-NEXT: vpbroadcastq %xmm4, %ymm5656; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0,1,2,3,4,5],ymm5[6,7]657; AVX512DQ-BW-NEXT: vinserti128 $1, %xmm4, %ymm0, %ymm4658; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} ymm5 = [2,7,12,0]659; AVX512DQ-BW-NEXT: vpermi2q %zmm1, %zmm0, %zmm5660; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm4 = ymm5[0,1,2,3,4,5],ymm4[6,7]661; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} ymm5 = [11,0,5,0]662; AVX512DQ-BW-NEXT: vpermi2q %zmm0, %zmm1, %zmm5663; AVX512DQ-BW-NEXT: vpbroadcastq 144(%rdi), %ymm6664; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm5 = ymm5[0,1,2,3,4,5],ymm6[6,7]665; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} ymm6 = [12,1,6,0]666; AVX512DQ-BW-NEXT: vpermi2q %zmm0, %zmm1, %zmm6667; AVX512DQ-BW-NEXT: vpblendd {{.*#+}} ymm0 = ymm6[0,1,2,3,4,5],mem[6,7]668; AVX512DQ-BW-NEXT: vmovdqa %ymm2, (%rsi)669; AVX512DQ-BW-NEXT: vmovdqa %ymm3, (%rdx)670; AVX512DQ-BW-NEXT: vmovdqa %ymm4, (%rcx)671; AVX512DQ-BW-NEXT: vmovdqa %ymm5, (%r8)672; AVX512DQ-BW-NEXT: vmovdqa %ymm0, (%r9)673; AVX512DQ-BW-NEXT: vzeroupper674; AVX512DQ-BW-NEXT: retq675;676; AVX512DQ-BW-FCP-LABEL: load_i64_stride5_vf4:677; AVX512DQ-BW-FCP: # %bb.0:678; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rdi), %zmm0679; AVX512DQ-BW-FCP-NEXT: vmovdqa64 64(%rdi), %zmm1680; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} ymm2 = [0,5,10,15]681; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm1, %zmm0, %zmm2682; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} ymm3 = [1,6,11,0]683; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm1, %zmm0, %zmm3684; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} ymm4 = [0,1,2,4]685; AVX512DQ-BW-FCP-NEXT: vmovdqa 128(%rdi), %ymm5686; AVX512DQ-BW-FCP-NEXT: vpermi2q %ymm5, %ymm3, %ymm4687; AVX512DQ-BW-FCP-NEXT: vinserti128 $1, 128(%rdi), %ymm0, %ymm3688; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} ymm6 = [2,7,12,0]689; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm1, %zmm0, %zmm6690; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm3 = ymm6[0,1,2,3,4,5],ymm3[6,7]691; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} ymm6 = [11,0,5,0]692; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm0, %zmm1, %zmm6693; AVX512DQ-BW-FCP-NEXT: vpbroadcastq 144(%rdi), %ymm7694; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm6[0,1,2,3,4,5],ymm7[6,7]695; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} ymm7 = [12,1,6,0]696; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm0, %zmm1, %zmm7697; AVX512DQ-BW-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm7[0,1,2,3,4,5],ymm5[6,7]698; AVX512DQ-BW-FCP-NEXT: vmovdqa %ymm2, (%rsi)699; AVX512DQ-BW-FCP-NEXT: vmovdqa %ymm4, (%rdx)700; AVX512DQ-BW-FCP-NEXT: vmovdqa %ymm3, (%rcx)701; AVX512DQ-BW-FCP-NEXT: vmovdqa %ymm6, (%r8)702; AVX512DQ-BW-FCP-NEXT: vmovdqa %ymm0, (%r9)703; AVX512DQ-BW-FCP-NEXT: vzeroupper704; AVX512DQ-BW-FCP-NEXT: retq705 %wide.vec = load <20 x i64>, ptr %in.vec, align 64706 %strided.vec0 = shufflevector <20 x i64> %wide.vec, <20 x i64> poison, <4 x i32> <i32 0, i32 5, i32 10, i32 15>707 %strided.vec1 = shufflevector <20 x i64> %wide.vec, <20 x i64> poison, <4 x i32> <i32 1, i32 6, i32 11, i32 16>708 %strided.vec2 = shufflevector <20 x i64> %wide.vec, <20 x i64> poison, <4 x i32> <i32 2, i32 7, i32 12, i32 17>709 %strided.vec3 = shufflevector <20 x i64> %wide.vec, <20 x i64> poison, <4 x i32> <i32 3, i32 8, i32 13, i32 18>710 %strided.vec4 = shufflevector <20 x i64> %wide.vec, <20 x i64> poison, <4 x i32> <i32 4, i32 9, i32 14, i32 19>711 store <4 x i64> %strided.vec0, ptr %out.vec0, align 64712 store <4 x i64> %strided.vec1, ptr %out.vec1, align 64713 store <4 x i64> %strided.vec2, ptr %out.vec2, align 64714 store <4 x i64> %strided.vec3, ptr %out.vec3, align 64715 store <4 x i64> %strided.vec4, ptr %out.vec4, align 64716 ret void717}718 719define void @load_i64_stride5_vf8(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr %out.vec2, ptr %out.vec3, ptr %out.vec4) nounwind {720; SSE-LABEL: load_i64_stride5_vf8:721; SSE: # %bb.0:722; SSE-NEXT: movapd 304(%rdi), %xmm2723; SSE-NEXT: movapd 64(%rdi), %xmm1724; SSE-NEXT: movapd 224(%rdi), %xmm0725; SSE-NEXT: movapd 256(%rdi), %xmm4726; SSE-NEXT: movapd 176(%rdi), %xmm3727; SSE-NEXT: movapd 288(%rdi), %xmm7728; SSE-NEXT: movapd 208(%rdi), %xmm6729; SSE-NEXT: movapd (%rdi), %xmm9730; SSE-NEXT: movapd 16(%rdi), %xmm5731; SSE-NEXT: movapd 32(%rdi), %xmm14732; SSE-NEXT: movapd 48(%rdi), %xmm8733; SSE-NEXT: movapd 240(%rdi), %xmm11734; SSE-NEXT: movapd 272(%rdi), %xmm13735; SSE-NEXT: movapd 160(%rdi), %xmm10736; SSE-NEXT: movapd 192(%rdi), %xmm15737; SSE-NEXT: movapd %xmm15, %xmm12738; SSE-NEXT: movsd {{.*#+}} xmm12 = xmm10[0],xmm12[1]739; SSE-NEXT: shufpd {{.*#+}} xmm10 = xmm10[1],xmm6[0]740; SSE-NEXT: movsd {{.*#+}} xmm6 = xmm3[0],xmm6[1]741; SSE-NEXT: shufpd {{.*#+}} xmm3 = xmm3[1],xmm0[0]742; SSE-NEXT: movapd %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill743; SSE-NEXT: movsd {{.*#+}} xmm0 = xmm15[0],xmm0[1]744; SSE-NEXT: movapd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill745; SSE-NEXT: movapd %xmm14, %xmm15746; SSE-NEXT: movsd {{.*#+}} xmm15 = xmm9[0],xmm15[1]747; SSE-NEXT: shufpd {{.*#+}} xmm9 = xmm9[1],xmm8[0]748; SSE-NEXT: movsd {{.*#+}} xmm8 = xmm5[0],xmm8[1]749; SSE-NEXT: shufpd {{.*#+}} xmm5 = xmm5[1],xmm1[0]750; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm14[0],xmm1[1]751; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill752; SSE-NEXT: movapd %xmm13, %xmm14753; SSE-NEXT: movsd {{.*#+}} xmm14 = xmm11[0],xmm14[1]754; SSE-NEXT: shufpd {{.*#+}} xmm11 = xmm11[1],xmm7[0]755; SSE-NEXT: movsd {{.*#+}} xmm7 = xmm4[0],xmm7[1]756; SSE-NEXT: shufpd {{.*#+}} xmm4 = xmm4[1],xmm2[0]757; SSE-NEXT: movapd %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill758; SSE-NEXT: movsd {{.*#+}} xmm2 = xmm13[0],xmm2[1]759; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill760; SSE-NEXT: movapd 80(%rdi), %xmm13761; SSE-NEXT: movapd 112(%rdi), %xmm4762; SSE-NEXT: movapd %xmm4, %xmm2763; SSE-NEXT: movsd {{.*#+}} xmm2 = xmm13[0],xmm2[1]764; SSE-NEXT: movapd 128(%rdi), %xmm0765; SSE-NEXT: shufpd {{.*#+}} xmm13 = xmm13[1],xmm0[0]766; SSE-NEXT: movapd 96(%rdi), %xmm1767; SSE-NEXT: movsd {{.*#+}} xmm0 = xmm1[0],xmm0[1]768; SSE-NEXT: movapd 144(%rdi), %xmm3769; SSE-NEXT: shufpd {{.*#+}} xmm1 = xmm1[1],xmm3[0]770; SSE-NEXT: movsd {{.*#+}} xmm3 = xmm4[0],xmm3[1]771; SSE-NEXT: movapd %xmm2, 16(%rsi)772; SSE-NEXT: movapd %xmm14, 48(%rsi)773; SSE-NEXT: movapd %xmm15, (%rsi)774; SSE-NEXT: movapd %xmm12, 32(%rsi)775; SSE-NEXT: movapd %xmm13, 16(%rdx)776; SSE-NEXT: movapd %xmm11, 48(%rdx)777; SSE-NEXT: movapd %xmm9, (%rdx)778; SSE-NEXT: movapd %xmm10, 32(%rdx)779; SSE-NEXT: movapd %xmm0, 16(%rcx)780; SSE-NEXT: movapd %xmm7, 48(%rcx)781; SSE-NEXT: movapd %xmm8, (%rcx)782; SSE-NEXT: movapd %xmm6, 32(%rcx)783; SSE-NEXT: movapd %xmm1, 16(%r8)784; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload785; SSE-NEXT: movaps %xmm0, 48(%r8)786; SSE-NEXT: movapd %xmm5, (%r8)787; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload788; SSE-NEXT: movaps %xmm0, 32(%r8)789; SSE-NEXT: movapd %xmm3, 16(%r9)790; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload791; SSE-NEXT: movaps %xmm0, 48(%r9)792; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload793; SSE-NEXT: movaps %xmm0, (%r9)794; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload795; SSE-NEXT: movaps %xmm0, 32(%r9)796; SSE-NEXT: retq797;798; AVX-LABEL: load_i64_stride5_vf8:799; AVX: # %bb.0:800; AVX-NEXT: vmovapd 128(%rdi), %ymm1801; AVX-NEXT: vmovapd 256(%rdi), %ymm0802; AVX-NEXT: vmovapd 224(%rdi), %ymm9803; AVX-NEXT: vmovapd 96(%rdi), %ymm2804; AVX-NEXT: vmovapd 64(%rdi), %ymm7805; AVX-NEXT: vblendpd {{.*#+}} ymm3 = ymm7[0,1,2],ymm2[3]806; AVX-NEXT: vmovapd (%rdi), %xmm10807; AVX-NEXT: vmovdqa 16(%rdi), %xmm8808; AVX-NEXT: vmovapd 32(%rdi), %xmm4809; AVX-NEXT: vmovdqa 48(%rdi), %xmm11810; AVX-NEXT: vblendpd {{.*#+}} xmm5 = xmm10[0],xmm4[1]811; AVX-NEXT: vblendpd {{.*#+}} ymm3 = ymm5[0,1],ymm3[2,3]812; AVX-NEXT: vmovupd %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill813; AVX-NEXT: vblendpd {{.*#+}} ymm6 = ymm9[0,1,2],ymm0[3]814; AVX-NEXT: vmovapd %ymm0, %ymm3815; AVX-NEXT: vmovapd 192(%rdi), %xmm5816; AVX-NEXT: vmovapd 160(%rdi), %xmm12817; AVX-NEXT: vblendpd {{.*#+}} xmm13 = xmm12[0],xmm5[1]818; AVX-NEXT: vblendpd {{.*#+}} ymm6 = ymm13[0,1],ymm6[2,3]819; AVX-NEXT: vmovaps 128(%rdi), %xmm13820; AVX-NEXT: vinsertf128 $1, %xmm13, %ymm0, %ymm14821; AVX-NEXT: vshufpd {{.*#+}} ymm7 = ymm7[0],ymm14[1],ymm7[3],ymm14[2]822; AVX-NEXT: vpalignr {{.*#+}} xmm10 = xmm10[8,9,10,11,12,13,14,15],xmm11[0,1,2,3,4,5,6,7]823; AVX-NEXT: vblendpd {{.*#+}} ymm7 = ymm10[0,1],ymm7[2,3]824; AVX-NEXT: vmovaps 288(%rdi), %xmm14825; AVX-NEXT: vinsertf128 $1, %xmm14, %ymm0, %ymm10826; AVX-NEXT: vshufpd {{.*#+}} ymm9 = ymm9[0],ymm10[1],ymm9[3],ymm10[2]827; AVX-NEXT: vmovdqa 208(%rdi), %xmm15828; AVX-NEXT: vpalignr {{.*#+}} xmm10 = xmm12[8,9,10,11,12,13,14,15],xmm15[0,1,2,3,4,5,6,7]829; AVX-NEXT: vblendpd {{.*#+}} ymm9 = ymm10[0,1],ymm9[2,3]830; AVX-NEXT: vmovaps 96(%rdi), %xmm12831; AVX-NEXT: vblendps {{.*#+}} xmm10 = xmm12[0,1],xmm13[2,3]832; AVX-NEXT: vinsertf128 $1, %xmm10, %ymm0, %ymm10833; AVX-NEXT: vpblendw {{.*#+}} xmm11 = xmm8[0,1,2,3],xmm11[4,5,6,7]834; AVX-NEXT: vblendps {{.*#+}} ymm10 = ymm11[0,1,2,3],ymm10[4,5,6,7]835; AVX-NEXT: vmovaps 256(%rdi), %xmm11836; AVX-NEXT: vblendps {{.*#+}} xmm13 = xmm11[0,1],xmm14[2,3]837; AVX-NEXT: vinsertf128 $1, %xmm13, %ymm0, %ymm13838; AVX-NEXT: vmovdqa 176(%rdi), %xmm14839; AVX-NEXT: vpblendw {{.*#+}} xmm15 = xmm14[0,1,2,3],xmm15[4,5,6,7]840; AVX-NEXT: vblendps {{.*#+}} ymm13 = ymm15[0,1,2,3],ymm13[4,5,6,7]841; AVX-NEXT: vinsertf128 $1, %xmm12, %ymm0, %ymm12842; AVX-NEXT: vshufpd {{.*#+}} ymm12 = ymm12[0],ymm1[1],ymm12[3],ymm1[2]843; AVX-NEXT: vmovdqa 64(%rdi), %xmm15844; AVX-NEXT: vpalignr {{.*#+}} xmm8 = xmm8[8,9,10,11,12,13,14,15],xmm15[0,1,2,3,4,5,6,7]845; AVX-NEXT: vblendpd {{.*#+}} ymm8 = ymm8[0,1],ymm12[2,3]846; AVX-NEXT: vmovapd 288(%rdi), %ymm12847; AVX-NEXT: vinsertf128 $1, %xmm11, %ymm0, %ymm11848; AVX-NEXT: vshufpd {{.*#+}} ymm11 = ymm11[0],ymm12[1],ymm11[3],ymm12[2]849; AVX-NEXT: vmovdqa 224(%rdi), %xmm0850; AVX-NEXT: vpalignr {{.*#+}} xmm14 = xmm14[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]851; AVX-NEXT: vblendpd {{.*#+}} ymm11 = ymm14[0,1],ymm11[2,3]852; AVX-NEXT: vblendpd {{.*#+}} ymm1 = ymm2[0,1,2],ymm1[3]853; AVX-NEXT: vblendpd {{.*#+}} xmm2 = xmm4[0],xmm15[1]854; AVX-NEXT: vblendpd {{.*#+}} ymm1 = ymm2[0,1],ymm1[2,3]855; AVX-NEXT: vblendpd {{.*#+}} ymm2 = ymm3[0,1,2],ymm12[3]856; AVX-NEXT: vblendpd {{.*#+}} xmm0 = xmm5[0],xmm0[1]857; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm0[0,1],ymm2[2,3]858; AVX-NEXT: vmovapd %ymm6, 32(%rsi)859; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload860; AVX-NEXT: vmovaps %ymm2, (%rsi)861; AVX-NEXT: vmovapd %ymm9, 32(%rdx)862; AVX-NEXT: vmovapd %ymm7, (%rdx)863; AVX-NEXT: vmovaps %ymm13, 32(%rcx)864; AVX-NEXT: vmovaps %ymm10, (%rcx)865; AVX-NEXT: vmovapd %ymm11, 32(%r8)866; AVX-NEXT: vmovapd %ymm8, (%r8)867; AVX-NEXT: vmovapd %ymm0, 32(%r9)868; AVX-NEXT: vmovapd %ymm1, (%r9)869; AVX-NEXT: vzeroupper870; AVX-NEXT: retq871;872; AVX2-LABEL: load_i64_stride5_vf8:873; AVX2: # %bb.0:874; AVX2-NEXT: vmovdqa 160(%rdi), %ymm9875; AVX2-NEXT: vmovdqa (%rdi), %ymm11876; AVX2-NEXT: vmovdqa 288(%rdi), %ymm0877; AVX2-NEXT: vmovdqa 128(%rdi), %ymm2878; AVX2-NEXT: vmovdqa 256(%rdi), %ymm1879; AVX2-NEXT: vmovdqa 224(%rdi), %ymm12880; AVX2-NEXT: vmovdqa 96(%rdi), %ymm3881; AVX2-NEXT: vmovdqa 64(%rdi), %ymm8882; AVX2-NEXT: vpblendd {{.*#+}} ymm4 = ymm8[0,1,2,3,4,5],ymm3[6,7]883; AVX2-NEXT: vmovdqa (%rdi), %xmm13884; AVX2-NEXT: vmovdqa 32(%rdi), %xmm5885; AVX2-NEXT: vmovdqa 48(%rdi), %xmm14886; AVX2-NEXT: vpblendd {{.*#+}} xmm6 = xmm13[0,1],xmm5[2,3]887; AVX2-NEXT: vpblendd {{.*#+}} ymm4 = ymm6[0,1,2,3],ymm4[4,5,6,7]888; AVX2-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill889; AVX2-NEXT: vpblendd {{.*#+}} ymm7 = ymm12[0,1,2,3,4,5],ymm1[6,7]890; AVX2-NEXT: vmovdqa 192(%rdi), %xmm6891; AVX2-NEXT: vmovdqa 160(%rdi), %xmm15892; AVX2-NEXT: vpblendd {{.*#+}} xmm10 = xmm15[0,1],xmm6[2,3]893; AVX2-NEXT: vpblendd {{.*#+}} ymm7 = ymm10[0,1,2,3],ymm7[4,5,6,7]894; AVX2-NEXT: vmovdqa 64(%rdi), %xmm10895; AVX2-NEXT: vpalignr {{.*#+}} xmm13 = xmm13[8,9,10,11,12,13,14,15],xmm14[0,1,2,3,4,5,6,7]896; AVX2-NEXT: vpalignr {{.*#+}} ymm8 = ymm8[8,9,10,11,12,13,14,15],ymm2[0,1,2,3,4,5,6,7],ymm8[24,25,26,27,28,29,30,31],ymm2[16,17,18,19,20,21,22,23]897; AVX2-NEXT: vpermq {{.*#+}} ymm8 = ymm8[0,1,2,1]898; AVX2-NEXT: vpblendd {{.*#+}} ymm8 = ymm13[0,1,2,3],ymm8[4,5,6,7]899; AVX2-NEXT: vmovdqa 208(%rdi), %xmm13900; AVX2-NEXT: vpalignr {{.*#+}} xmm13 = xmm15[8,9,10,11,12,13,14,15],xmm13[0,1,2,3,4,5,6,7]901; AVX2-NEXT: vpalignr {{.*#+}} ymm12 = ymm12[8,9,10,11,12,13,14,15],ymm0[0,1,2,3,4,5,6,7],ymm12[24,25,26,27,28,29,30,31],ymm0[16,17,18,19,20,21,22,23]902; AVX2-NEXT: vpermq {{.*#+}} ymm12 = ymm12[0,1,2,1]903; AVX2-NEXT: vpblendd {{.*#+}} ymm12 = ymm13[0,1,2,3],ymm12[4,5,6,7]904; AVX2-NEXT: vpblendd {{.*#+}} ymm11 = ymm11[0,1],mem[2,3],ymm11[4,5],mem[6,7]905; AVX2-NEXT: vmovdqa 96(%rdi), %xmm13906; AVX2-NEXT: vpblendd {{.*#+}} xmm13 = xmm13[0,1],mem[2,3]907; AVX2-NEXT: vinserti128 $1, %xmm13, %ymm0, %ymm13908; AVX2-NEXT: vperm2i128 {{.*#+}} ymm11 = ymm11[2,3],ymm13[2,3]909; AVX2-NEXT: vpblendd {{.*#+}} ymm9 = ymm9[0,1],mem[2,3],ymm9[4,5],mem[6,7]910; AVX2-NEXT: vmovdqa 256(%rdi), %xmm13911; AVX2-NEXT: vpblendd {{.*#+}} xmm13 = xmm13[0,1],mem[2,3]912; AVX2-NEXT: vinserti128 $1, %xmm13, %ymm0, %ymm13913; AVX2-NEXT: vperm2i128 {{.*#+}} ymm9 = ymm9[2,3],ymm13[2,3]914; AVX2-NEXT: vpalignr {{.*#+}} xmm13 = mem[8,9,10,11,12,13,14,15],xmm10[0,1,2,3,4,5,6,7]915; AVX2-NEXT: vpalignr {{.*#+}} ymm14 = ymm3[8,9,10,11,12,13,14,15],ymm2[0,1,2,3,4,5,6,7],ymm3[24,25,26,27,28,29,30,31],ymm2[16,17,18,19,20,21,22,23]916; AVX2-NEXT: vpermq {{.*#+}} ymm14 = ymm14[0,1,0,3]917; AVX2-NEXT: vpblendd {{.*#+}} ymm13 = ymm13[0,1,2,3],ymm14[4,5,6,7]918; AVX2-NEXT: vmovdqa 224(%rdi), %xmm14919; AVX2-NEXT: vpalignr {{.*#+}} ymm15 = ymm1[8,9,10,11,12,13,14,15],ymm0[0,1,2,3,4,5,6,7],ymm1[24,25,26,27,28,29,30,31],ymm0[16,17,18,19,20,21,22,23]920; AVX2-NEXT: vpermq {{.*#+}} ymm15 = ymm15[0,1,0,3]921; AVX2-NEXT: vpalignr {{.*#+}} xmm4 = mem[8,9,10,11,12,13,14,15],xmm14[0,1,2,3,4,5,6,7]922; AVX2-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm15[4,5,6,7]923; AVX2-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4,5],ymm2[6,7]924; AVX2-NEXT: vpblendd {{.*#+}} xmm3 = xmm5[0,1],xmm10[2,3]925; AVX2-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3],ymm2[4,5,6,7]926; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]927; AVX2-NEXT: vpblendd {{.*#+}} xmm1 = xmm6[0,1],xmm14[2,3]928; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]929; AVX2-NEXT: vmovdqa %ymm7, 32(%rsi)930; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload931; AVX2-NEXT: vmovaps %ymm1, (%rsi)932; AVX2-NEXT: vmovdqa %ymm12, 32(%rdx)933; AVX2-NEXT: vmovdqa %ymm8, (%rdx)934; AVX2-NEXT: vmovdqa %ymm9, 32(%rcx)935; AVX2-NEXT: vmovdqa %ymm11, (%rcx)936; AVX2-NEXT: vmovdqa %ymm4, 32(%r8)937; AVX2-NEXT: vmovdqa %ymm13, (%r8)938; AVX2-NEXT: vmovdqa %ymm0, 32(%r9)939; AVX2-NEXT: vmovdqa %ymm2, (%r9)940; AVX2-NEXT: vzeroupper941; AVX2-NEXT: retq942;943; AVX2-FP-LABEL: load_i64_stride5_vf8:944; AVX2-FP: # %bb.0:945; AVX2-FP-NEXT: vmovdqa 160(%rdi), %ymm9946; AVX2-FP-NEXT: vmovdqa (%rdi), %ymm11947; AVX2-FP-NEXT: vmovdqa 288(%rdi), %ymm0948; AVX2-FP-NEXT: vmovdqa 128(%rdi), %ymm2949; AVX2-FP-NEXT: vmovdqa 256(%rdi), %ymm1950; AVX2-FP-NEXT: vmovdqa 224(%rdi), %ymm12951; AVX2-FP-NEXT: vmovdqa 96(%rdi), %ymm3952; AVX2-FP-NEXT: vmovdqa 64(%rdi), %ymm8953; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm4 = ymm8[0,1,2,3,4,5],ymm3[6,7]954; AVX2-FP-NEXT: vmovdqa (%rdi), %xmm13955; AVX2-FP-NEXT: vmovdqa 32(%rdi), %xmm5956; AVX2-FP-NEXT: vmovdqa 48(%rdi), %xmm14957; AVX2-FP-NEXT: vpblendd {{.*#+}} xmm6 = xmm13[0,1],xmm5[2,3]958; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm4 = ymm6[0,1,2,3],ymm4[4,5,6,7]959; AVX2-FP-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill960; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm7 = ymm12[0,1,2,3,4,5],ymm1[6,7]961; AVX2-FP-NEXT: vmovdqa 192(%rdi), %xmm6962; AVX2-FP-NEXT: vmovdqa 160(%rdi), %xmm15963; AVX2-FP-NEXT: vpblendd {{.*#+}} xmm10 = xmm15[0,1],xmm6[2,3]964; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm7 = ymm10[0,1,2,3],ymm7[4,5,6,7]965; AVX2-FP-NEXT: vmovdqa 64(%rdi), %xmm10966; AVX2-FP-NEXT: vpalignr {{.*#+}} xmm13 = xmm13[8,9,10,11,12,13,14,15],xmm14[0,1,2,3,4,5,6,7]967; AVX2-FP-NEXT: vpalignr {{.*#+}} ymm8 = ymm8[8,9,10,11,12,13,14,15],ymm2[0,1,2,3,4,5,6,7],ymm8[24,25,26,27,28,29,30,31],ymm2[16,17,18,19,20,21,22,23]968; AVX2-FP-NEXT: vpermq {{.*#+}} ymm8 = ymm8[0,1,2,1]969; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm8 = ymm13[0,1,2,3],ymm8[4,5,6,7]970; AVX2-FP-NEXT: vmovdqa 208(%rdi), %xmm13971; AVX2-FP-NEXT: vpalignr {{.*#+}} xmm13 = xmm15[8,9,10,11,12,13,14,15],xmm13[0,1,2,3,4,5,6,7]972; AVX2-FP-NEXT: vpalignr {{.*#+}} ymm12 = ymm12[8,9,10,11,12,13,14,15],ymm0[0,1,2,3,4,5,6,7],ymm12[24,25,26,27,28,29,30,31],ymm0[16,17,18,19,20,21,22,23]973; AVX2-FP-NEXT: vpermq {{.*#+}} ymm12 = ymm12[0,1,2,1]974; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm12 = ymm13[0,1,2,3],ymm12[4,5,6,7]975; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm11 = ymm11[0,1],mem[2,3],ymm11[4,5],mem[6,7]976; AVX2-FP-NEXT: vmovdqa 96(%rdi), %xmm13977; AVX2-FP-NEXT: vpblendd {{.*#+}} xmm13 = xmm13[0,1],mem[2,3]978; AVX2-FP-NEXT: vinserti128 $1, %xmm13, %ymm0, %ymm13979; AVX2-FP-NEXT: vperm2i128 {{.*#+}} ymm11 = ymm11[2,3],ymm13[2,3]980; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm9 = ymm9[0,1],mem[2,3],ymm9[4,5],mem[6,7]981; AVX2-FP-NEXT: vmovdqa 256(%rdi), %xmm13982; AVX2-FP-NEXT: vpblendd {{.*#+}} xmm13 = xmm13[0,1],mem[2,3]983; AVX2-FP-NEXT: vinserti128 $1, %xmm13, %ymm0, %ymm13984; AVX2-FP-NEXT: vperm2i128 {{.*#+}} ymm9 = ymm9[2,3],ymm13[2,3]985; AVX2-FP-NEXT: vpalignr {{.*#+}} xmm13 = mem[8,9,10,11,12,13,14,15],xmm10[0,1,2,3,4,5,6,7]986; AVX2-FP-NEXT: vpalignr {{.*#+}} ymm14 = ymm3[8,9,10,11,12,13,14,15],ymm2[0,1,2,3,4,5,6,7],ymm3[24,25,26,27,28,29,30,31],ymm2[16,17,18,19,20,21,22,23]987; AVX2-FP-NEXT: vpermq {{.*#+}} ymm14 = ymm14[0,1,0,3]988; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm13 = ymm13[0,1,2,3],ymm14[4,5,6,7]989; AVX2-FP-NEXT: vmovdqa 224(%rdi), %xmm14990; AVX2-FP-NEXT: vpalignr {{.*#+}} ymm15 = ymm1[8,9,10,11,12,13,14,15],ymm0[0,1,2,3,4,5,6,7],ymm1[24,25,26,27,28,29,30,31],ymm0[16,17,18,19,20,21,22,23]991; AVX2-FP-NEXT: vpermq {{.*#+}} ymm15 = ymm15[0,1,0,3]992; AVX2-FP-NEXT: vpalignr {{.*#+}} xmm4 = mem[8,9,10,11,12,13,14,15],xmm14[0,1,2,3,4,5,6,7]993; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm15[4,5,6,7]994; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4,5],ymm2[6,7]995; AVX2-FP-NEXT: vpblendd {{.*#+}} xmm3 = xmm5[0,1],xmm10[2,3]996; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3],ymm2[4,5,6,7]997; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]998; AVX2-FP-NEXT: vpblendd {{.*#+}} xmm1 = xmm6[0,1],xmm14[2,3]999; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]1000; AVX2-FP-NEXT: vmovdqa %ymm7, 32(%rsi)1001; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload1002; AVX2-FP-NEXT: vmovaps %ymm1, (%rsi)1003; AVX2-FP-NEXT: vmovdqa %ymm12, 32(%rdx)1004; AVX2-FP-NEXT: vmovdqa %ymm8, (%rdx)1005; AVX2-FP-NEXT: vmovdqa %ymm9, 32(%rcx)1006; AVX2-FP-NEXT: vmovdqa %ymm11, (%rcx)1007; AVX2-FP-NEXT: vmovdqa %ymm4, 32(%r8)1008; AVX2-FP-NEXT: vmovdqa %ymm13, (%r8)1009; AVX2-FP-NEXT: vmovdqa %ymm0, 32(%r9)1010; AVX2-FP-NEXT: vmovdqa %ymm2, (%r9)1011; AVX2-FP-NEXT: vzeroupper1012; AVX2-FP-NEXT: retq1013;1014; AVX2-FCP-LABEL: load_i64_stride5_vf8:1015; AVX2-FCP: # %bb.0:1016; AVX2-FCP-NEXT: vmovdqa 160(%rdi), %ymm91017; AVX2-FCP-NEXT: vmovdqa (%rdi), %ymm111018; AVX2-FCP-NEXT: vmovdqa 288(%rdi), %ymm01019; AVX2-FCP-NEXT: vmovdqa 128(%rdi), %ymm21020; AVX2-FCP-NEXT: vmovdqa 256(%rdi), %ymm11021; AVX2-FCP-NEXT: vmovdqa 224(%rdi), %ymm121022; AVX2-FCP-NEXT: vmovdqa 96(%rdi), %ymm31023; AVX2-FCP-NEXT: vmovdqa 64(%rdi), %ymm81024; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm8[0,1,2,3,4,5],ymm3[6,7]1025; AVX2-FCP-NEXT: vmovdqa (%rdi), %xmm131026; AVX2-FCP-NEXT: vmovdqa 32(%rdi), %xmm51027; AVX2-FCP-NEXT: vmovdqa 48(%rdi), %xmm141028; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm6 = xmm13[0,1],xmm5[2,3]1029; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm6[0,1,2,3],ymm4[4,5,6,7]1030; AVX2-FCP-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill1031; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm7 = ymm12[0,1,2,3,4,5],ymm1[6,7]1032; AVX2-FCP-NEXT: vmovdqa 192(%rdi), %xmm61033; AVX2-FCP-NEXT: vmovdqa 160(%rdi), %xmm151034; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm10 = xmm15[0,1],xmm6[2,3]1035; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm7 = ymm10[0,1,2,3],ymm7[4,5,6,7]1036; AVX2-FCP-NEXT: vmovdqa 64(%rdi), %xmm101037; AVX2-FCP-NEXT: vpalignr {{.*#+}} xmm13 = xmm13[8,9,10,11,12,13,14,15],xmm14[0,1,2,3,4,5,6,7]1038; AVX2-FCP-NEXT: vpalignr {{.*#+}} ymm8 = ymm8[8,9,10,11,12,13,14,15],ymm2[0,1,2,3,4,5,6,7],ymm8[24,25,26,27,28,29,30,31],ymm2[16,17,18,19,20,21,22,23]1039; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm8 = ymm8[0,1,2,1]1040; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm8 = ymm13[0,1,2,3],ymm8[4,5,6,7]1041; AVX2-FCP-NEXT: vmovdqa 208(%rdi), %xmm131042; AVX2-FCP-NEXT: vpalignr {{.*#+}} xmm13 = xmm15[8,9,10,11,12,13,14,15],xmm13[0,1,2,3,4,5,6,7]1043; AVX2-FCP-NEXT: vpalignr {{.*#+}} ymm12 = ymm12[8,9,10,11,12,13,14,15],ymm0[0,1,2,3,4,5,6,7],ymm12[24,25,26,27,28,29,30,31],ymm0[16,17,18,19,20,21,22,23]1044; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm12 = ymm12[0,1,2,1]1045; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm12 = ymm13[0,1,2,3],ymm12[4,5,6,7]1046; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm11 = ymm11[0,1],mem[2,3],ymm11[4,5],mem[6,7]1047; AVX2-FCP-NEXT: vmovdqa 96(%rdi), %xmm131048; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm13 = xmm13[0,1],mem[2,3]1049; AVX2-FCP-NEXT: vinserti128 $1, %xmm13, %ymm0, %ymm131050; AVX2-FCP-NEXT: vperm2i128 {{.*#+}} ymm11 = ymm11[2,3],ymm13[2,3]1051; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm9 = ymm9[0,1],mem[2,3],ymm9[4,5],mem[6,7]1052; AVX2-FCP-NEXT: vmovdqa 256(%rdi), %xmm131053; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm13 = xmm13[0,1],mem[2,3]1054; AVX2-FCP-NEXT: vinserti128 $1, %xmm13, %ymm0, %ymm131055; AVX2-FCP-NEXT: vperm2i128 {{.*#+}} ymm9 = ymm9[2,3],ymm13[2,3]1056; AVX2-FCP-NEXT: vpalignr {{.*#+}} xmm13 = mem[8,9,10,11,12,13,14,15],xmm10[0,1,2,3,4,5,6,7]1057; AVX2-FCP-NEXT: vpalignr {{.*#+}} ymm14 = ymm3[8,9,10,11,12,13,14,15],ymm2[0,1,2,3,4,5,6,7],ymm3[24,25,26,27,28,29,30,31],ymm2[16,17,18,19,20,21,22,23]1058; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm14 = ymm14[0,1,0,3]1059; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm13 = ymm13[0,1,2,3],ymm14[4,5,6,7]1060; AVX2-FCP-NEXT: vmovdqa 224(%rdi), %xmm141061; AVX2-FCP-NEXT: vpalignr {{.*#+}} ymm15 = ymm1[8,9,10,11,12,13,14,15],ymm0[0,1,2,3,4,5,6,7],ymm1[24,25,26,27,28,29,30,31],ymm0[16,17,18,19,20,21,22,23]1062; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm15 = ymm15[0,1,0,3]1063; AVX2-FCP-NEXT: vpalignr {{.*#+}} xmm4 = mem[8,9,10,11,12,13,14,15],xmm14[0,1,2,3,4,5,6,7]1064; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm15[4,5,6,7]1065; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3,4,5],ymm2[6,7]1066; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm3 = xmm5[0,1],xmm10[2,3]1067; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm3[0,1,2,3],ymm2[4,5,6,7]1068; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]1069; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm1 = xmm6[0,1],xmm14[2,3]1070; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]1071; AVX2-FCP-NEXT: vmovdqa %ymm7, 32(%rsi)1072; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload1073; AVX2-FCP-NEXT: vmovaps %ymm1, (%rsi)1074; AVX2-FCP-NEXT: vmovdqa %ymm12, 32(%rdx)1075; AVX2-FCP-NEXT: vmovdqa %ymm8, (%rdx)1076; AVX2-FCP-NEXT: vmovdqa %ymm9, 32(%rcx)1077; AVX2-FCP-NEXT: vmovdqa %ymm11, (%rcx)1078; AVX2-FCP-NEXT: vmovdqa %ymm4, 32(%r8)1079; AVX2-FCP-NEXT: vmovdqa %ymm13, (%r8)1080; AVX2-FCP-NEXT: vmovdqa %ymm0, 32(%r9)1081; AVX2-FCP-NEXT: vmovdqa %ymm2, (%r9)1082; AVX2-FCP-NEXT: vzeroupper1083; AVX2-FCP-NEXT: retq1084;1085; AVX512-LABEL: load_i64_stride5_vf8:1086; AVX512: # %bb.0:1087; AVX512-NEXT: vmovdqa64 256(%rdi), %zmm01088; AVX512-NEXT: vmovdqa64 (%rdi), %zmm21089; AVX512-NEXT: vmovdqa64 64(%rdi), %zmm11090; AVX512-NEXT: vmovdqa64 128(%rdi), %zmm41091; AVX512-NEXT: vmovdqa64 192(%rdi), %zmm51092; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm3 = [12,1,6,0,12,1,6,0]1093; AVX512-NEXT: # zmm3 = mem[0,1,2,3,0,1,2,3]1094; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = [5,10,15,0,5,10,15,0]1095; AVX512-NEXT: # zmm6 = mem[0,1,2,3,0,1,2,3]1096; AVX512-NEXT: vpermi2q %zmm5, %zmm4, %zmm61097; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [6,11,0,1,6,11,0,1]1098; AVX512-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]1099; AVX512-NEXT: vpermi2q %zmm5, %zmm4, %zmm71100; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [7,12,0,2,7,12,0,2]1101; AVX512-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]1102; AVX512-NEXT: vpermi2q %zmm5, %zmm4, %zmm81103; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm9 = [0,5,0,11,0,5,0,11]1104; AVX512-NEXT: # zmm9 = mem[0,1,2,3,0,1,2,3]1105; AVX512-NEXT: vpermi2q %zmm4, %zmm5, %zmm91106; AVX512-NEXT: vpermt2q %zmm4, %zmm3, %zmm51107; AVX512-NEXT: vpmovsxbq {{.*#+}} ymm4 = [0,5,10,15]1108; AVX512-NEXT: vpermi2q %zmm1, %zmm2, %zmm41109; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm4 = zmm4[0,1,2,3],zmm5[4,5,6,7]1110; AVX512-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,1,2,3,4,5,6,11]1111; AVX512-NEXT: vpermi2q %zmm0, %zmm4, %zmm51112; AVX512-NEXT: vpmovsxbq {{.*#+}} ymm4 = [1,6,11,0]1113; AVX512-NEXT: vpermi2q %zmm1, %zmm2, %zmm41114; AVX512-NEXT: movb $7, %al1115; AVX512-NEXT: kmovw %eax, %k11116; AVX512-NEXT: vmovdqa64 %zmm4, %zmm6 {%k1}1117; AVX512-NEXT: vpmovsxbq {{.*#+}} zmm4 = [0,1,2,3,4,5,6,12]1118; AVX512-NEXT: vpermi2q %zmm0, %zmm6, %zmm41119; AVX512-NEXT: vpmovsxbq {{.*#+}} ymm6 = [2,7,12,0]1120; AVX512-NEXT: vpermi2q %zmm1, %zmm2, %zmm61121; AVX512-NEXT: movb $56, %al1122; AVX512-NEXT: kmovw %eax, %k11123; AVX512-NEXT: vmovdqa64 %zmm7, %zmm6 {%k1}1124; AVX512-NEXT: vpmovsxbq {{.*#+}} zmm7 = [0,1,2,3,4,5,8,13]1125; AVX512-NEXT: vpermi2q %zmm0, %zmm6, %zmm71126; AVX512-NEXT: vpmovsxbq {{.*#+}} ymm6 = [11,0,5,0]1127; AVX512-NEXT: vpermi2q %zmm2, %zmm1, %zmm61128; AVX512-NEXT: vmovdqa64 %zmm8, %zmm6 {%k1}1129; AVX512-NEXT: vpmovsxbq {{.*#+}} zmm8 = [0,1,2,3,4,5,9,14]1130; AVX512-NEXT: vpermi2q %zmm0, %zmm6, %zmm81131; AVX512-NEXT: vpermt2q %zmm2, %zmm3, %zmm11132; AVX512-NEXT: vmovdqa64 %zmm9, %zmm1 {%k1}1133; AVX512-NEXT: vpmovsxbq {{.*#+}} zmm2 = [0,1,2,3,4,5,10,15]1134; AVX512-NEXT: vpermi2q %zmm0, %zmm1, %zmm21135; AVX512-NEXT: vmovdqa64 %zmm5, (%rsi)1136; AVX512-NEXT: vmovdqa64 %zmm4, (%rdx)1137; AVX512-NEXT: vmovdqa64 %zmm7, (%rcx)1138; AVX512-NEXT: vmovdqa64 %zmm8, (%r8)1139; AVX512-NEXT: vmovdqa64 %zmm2, (%r9)1140; AVX512-NEXT: vzeroupper1141; AVX512-NEXT: retq1142;1143; AVX512-FCP-LABEL: load_i64_stride5_vf8:1144; AVX512-FCP: # %bb.0:1145; AVX512-FCP-NEXT: vmovdqa64 256(%rdi), %zmm01146; AVX512-FCP-NEXT: vmovdqa64 (%rdi), %zmm21147; AVX512-FCP-NEXT: vmovdqa64 64(%rdi), %zmm11148; AVX512-FCP-NEXT: vmovdqa64 128(%rdi), %zmm41149; AVX512-FCP-NEXT: vmovdqa64 192(%rdi), %zmm51150; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm3 = [12,1,6,0,12,1,6,0]1151; AVX512-FCP-NEXT: # zmm3 = mem[0,1,2,3,0,1,2,3]1152; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = [5,10,15,0,5,10,15,0]1153; AVX512-FCP-NEXT: # zmm6 = mem[0,1,2,3,0,1,2,3]1154; AVX512-FCP-NEXT: vpermi2q %zmm5, %zmm4, %zmm61155; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [6,11,0,1,6,11,0,1]1156; AVX512-FCP-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]1157; AVX512-FCP-NEXT: vpermi2q %zmm5, %zmm4, %zmm71158; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [7,12,0,2,7,12,0,2]1159; AVX512-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]1160; AVX512-FCP-NEXT: vpermi2q %zmm5, %zmm4, %zmm81161; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm9 = [0,5,0,11,0,5,0,11]1162; AVX512-FCP-NEXT: # zmm9 = mem[0,1,2,3,0,1,2,3]1163; AVX512-FCP-NEXT: vpermi2q %zmm4, %zmm5, %zmm91164; AVX512-FCP-NEXT: vpermt2q %zmm4, %zmm3, %zmm51165; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} ymm4 = [0,5,10,15]1166; AVX512-FCP-NEXT: vpermi2q %zmm1, %zmm2, %zmm41167; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm4 = zmm4[0,1,2,3],zmm5[4,5,6,7]1168; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,1,2,3,4,5,6,11]1169; AVX512-FCP-NEXT: vpermi2q %zmm0, %zmm4, %zmm51170; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} ymm4 = [1,6,11,0]1171; AVX512-FCP-NEXT: vpermi2q %zmm1, %zmm2, %zmm41172; AVX512-FCP-NEXT: movb $7, %al1173; AVX512-FCP-NEXT: kmovw %eax, %k11174; AVX512-FCP-NEXT: vmovdqa64 %zmm4, %zmm6 {%k1}1175; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm4 = [0,1,2,3,4,5,6,12]1176; AVX512-FCP-NEXT: vpermi2q %zmm0, %zmm6, %zmm41177; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} ymm6 = [2,7,12,0]1178; AVX512-FCP-NEXT: vpermi2q %zmm1, %zmm2, %zmm61179; AVX512-FCP-NEXT: movb $56, %al1180; AVX512-FCP-NEXT: kmovw %eax, %k11181; AVX512-FCP-NEXT: vmovdqa64 %zmm7, %zmm6 {%k1}1182; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm7 = [0,1,2,3,4,5,8,13]1183; AVX512-FCP-NEXT: vpermi2q %zmm0, %zmm6, %zmm71184; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} ymm6 = [11,0,5,0]1185; AVX512-FCP-NEXT: vpermi2q %zmm2, %zmm1, %zmm61186; AVX512-FCP-NEXT: vmovdqa64 %zmm8, %zmm6 {%k1}1187; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm8 = [0,1,2,3,4,5,9,14]1188; AVX512-FCP-NEXT: vpermi2q %zmm0, %zmm6, %zmm81189; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm3, %zmm11190; AVX512-FCP-NEXT: vmovdqa64 %zmm9, %zmm1 {%k1}1191; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm2 = [0,1,2,3,4,5,10,15]1192; AVX512-FCP-NEXT: vpermi2q %zmm0, %zmm1, %zmm21193; AVX512-FCP-NEXT: vmovdqa64 %zmm5, (%rsi)1194; AVX512-FCP-NEXT: vmovdqa64 %zmm4, (%rdx)1195; AVX512-FCP-NEXT: vmovdqa64 %zmm7, (%rcx)1196; AVX512-FCP-NEXT: vmovdqa64 %zmm8, (%r8)1197; AVX512-FCP-NEXT: vmovdqa64 %zmm2, (%r9)1198; AVX512-FCP-NEXT: vzeroupper1199; AVX512-FCP-NEXT: retq1200;1201; AVX512DQ-LABEL: load_i64_stride5_vf8:1202; AVX512DQ: # %bb.0:1203; AVX512DQ-NEXT: vmovdqa64 256(%rdi), %zmm01204; AVX512DQ-NEXT: vmovdqa64 (%rdi), %zmm21205; AVX512DQ-NEXT: vmovdqa64 64(%rdi), %zmm11206; AVX512DQ-NEXT: vmovdqa64 128(%rdi), %zmm41207; AVX512DQ-NEXT: vmovdqa64 192(%rdi), %zmm51208; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm3 = [12,1,6,0,12,1,6,0]1209; AVX512DQ-NEXT: # zmm3 = mem[0,1,2,3,0,1,2,3]1210; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = [5,10,15,0,5,10,15,0]1211; AVX512DQ-NEXT: # zmm6 = mem[0,1,2,3,0,1,2,3]1212; AVX512DQ-NEXT: vpermi2q %zmm5, %zmm4, %zmm61213; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [6,11,0,1,6,11,0,1]1214; AVX512DQ-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]1215; AVX512DQ-NEXT: vpermi2q %zmm5, %zmm4, %zmm71216; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [7,12,0,2,7,12,0,2]1217; AVX512DQ-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]1218; AVX512DQ-NEXT: vpermi2q %zmm5, %zmm4, %zmm81219; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm9 = [0,5,0,11,0,5,0,11]1220; AVX512DQ-NEXT: # zmm9 = mem[0,1,2,3,0,1,2,3]1221; AVX512DQ-NEXT: vpermi2q %zmm4, %zmm5, %zmm91222; AVX512DQ-NEXT: vpermt2q %zmm4, %zmm3, %zmm51223; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} ymm4 = [0,5,10,15]1224; AVX512DQ-NEXT: vpermi2q %zmm1, %zmm2, %zmm41225; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm4 = zmm4[0,1,2,3],zmm5[4,5,6,7]1226; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,1,2,3,4,5,6,11]1227; AVX512DQ-NEXT: vpermi2q %zmm0, %zmm4, %zmm51228; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} ymm4 = [1,6,11,0]1229; AVX512DQ-NEXT: vpermi2q %zmm1, %zmm2, %zmm41230; AVX512DQ-NEXT: movb $7, %al1231; AVX512DQ-NEXT: kmovw %eax, %k11232; AVX512DQ-NEXT: vmovdqa64 %zmm4, %zmm6 {%k1}1233; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} zmm4 = [0,1,2,3,4,5,6,12]1234; AVX512DQ-NEXT: vpermi2q %zmm0, %zmm6, %zmm41235; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} ymm6 = [2,7,12,0]1236; AVX512DQ-NEXT: vpermi2q %zmm1, %zmm2, %zmm61237; AVX512DQ-NEXT: movb $56, %al1238; AVX512DQ-NEXT: kmovw %eax, %k11239; AVX512DQ-NEXT: vmovdqa64 %zmm7, %zmm6 {%k1}1240; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} zmm7 = [0,1,2,3,4,5,8,13]1241; AVX512DQ-NEXT: vpermi2q %zmm0, %zmm6, %zmm71242; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} ymm6 = [11,0,5,0]1243; AVX512DQ-NEXT: vpermi2q %zmm2, %zmm1, %zmm61244; AVX512DQ-NEXT: vmovdqa64 %zmm8, %zmm6 {%k1}1245; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} zmm8 = [0,1,2,3,4,5,9,14]1246; AVX512DQ-NEXT: vpermi2q %zmm0, %zmm6, %zmm81247; AVX512DQ-NEXT: vpermt2q %zmm2, %zmm3, %zmm11248; AVX512DQ-NEXT: vmovdqa64 %zmm9, %zmm1 {%k1}1249; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} zmm2 = [0,1,2,3,4,5,10,15]1250; AVX512DQ-NEXT: vpermi2q %zmm0, %zmm1, %zmm21251; AVX512DQ-NEXT: vmovdqa64 %zmm5, (%rsi)1252; AVX512DQ-NEXT: vmovdqa64 %zmm4, (%rdx)1253; AVX512DQ-NEXT: vmovdqa64 %zmm7, (%rcx)1254; AVX512DQ-NEXT: vmovdqa64 %zmm8, (%r8)1255; AVX512DQ-NEXT: vmovdqa64 %zmm2, (%r9)1256; AVX512DQ-NEXT: vzeroupper1257; AVX512DQ-NEXT: retq1258;1259; AVX512DQ-FCP-LABEL: load_i64_stride5_vf8:1260; AVX512DQ-FCP: # %bb.0:1261; AVX512DQ-FCP-NEXT: vmovdqa64 256(%rdi), %zmm01262; AVX512DQ-FCP-NEXT: vmovdqa64 (%rdi), %zmm21263; AVX512DQ-FCP-NEXT: vmovdqa64 64(%rdi), %zmm11264; AVX512DQ-FCP-NEXT: vmovdqa64 128(%rdi), %zmm41265; AVX512DQ-FCP-NEXT: vmovdqa64 192(%rdi), %zmm51266; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm3 = [12,1,6,0,12,1,6,0]1267; AVX512DQ-FCP-NEXT: # zmm3 = mem[0,1,2,3,0,1,2,3]1268; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = [5,10,15,0,5,10,15,0]1269; AVX512DQ-FCP-NEXT: # zmm6 = mem[0,1,2,3,0,1,2,3]1270; AVX512DQ-FCP-NEXT: vpermi2q %zmm5, %zmm4, %zmm61271; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [6,11,0,1,6,11,0,1]1272; AVX512DQ-FCP-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]1273; AVX512DQ-FCP-NEXT: vpermi2q %zmm5, %zmm4, %zmm71274; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [7,12,0,2,7,12,0,2]1275; AVX512DQ-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]1276; AVX512DQ-FCP-NEXT: vpermi2q %zmm5, %zmm4, %zmm81277; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm9 = [0,5,0,11,0,5,0,11]1278; AVX512DQ-FCP-NEXT: # zmm9 = mem[0,1,2,3,0,1,2,3]1279; AVX512DQ-FCP-NEXT: vpermi2q %zmm4, %zmm5, %zmm91280; AVX512DQ-FCP-NEXT: vpermt2q %zmm4, %zmm3, %zmm51281; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} ymm4 = [0,5,10,15]1282; AVX512DQ-FCP-NEXT: vpermi2q %zmm1, %zmm2, %zmm41283; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm4 = zmm4[0,1,2,3],zmm5[4,5,6,7]1284; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,1,2,3,4,5,6,11]1285; AVX512DQ-FCP-NEXT: vpermi2q %zmm0, %zmm4, %zmm51286; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} ymm4 = [1,6,11,0]1287; AVX512DQ-FCP-NEXT: vpermi2q %zmm1, %zmm2, %zmm41288; AVX512DQ-FCP-NEXT: movb $7, %al1289; AVX512DQ-FCP-NEXT: kmovw %eax, %k11290; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm4, %zmm6 {%k1}1291; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm4 = [0,1,2,3,4,5,6,12]1292; AVX512DQ-FCP-NEXT: vpermi2q %zmm0, %zmm6, %zmm41293; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} ymm6 = [2,7,12,0]1294; AVX512DQ-FCP-NEXT: vpermi2q %zmm1, %zmm2, %zmm61295; AVX512DQ-FCP-NEXT: movb $56, %al1296; AVX512DQ-FCP-NEXT: kmovw %eax, %k11297; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm7, %zmm6 {%k1}1298; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm7 = [0,1,2,3,4,5,8,13]1299; AVX512DQ-FCP-NEXT: vpermi2q %zmm0, %zmm6, %zmm71300; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} ymm6 = [11,0,5,0]1301; AVX512DQ-FCP-NEXT: vpermi2q %zmm2, %zmm1, %zmm61302; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm8, %zmm6 {%k1}1303; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm8 = [0,1,2,3,4,5,9,14]1304; AVX512DQ-FCP-NEXT: vpermi2q %zmm0, %zmm6, %zmm81305; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm3, %zmm11306; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm9, %zmm1 {%k1}1307; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm2 = [0,1,2,3,4,5,10,15]1308; AVX512DQ-FCP-NEXT: vpermi2q %zmm0, %zmm1, %zmm21309; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm5, (%rsi)1310; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm4, (%rdx)1311; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm7, (%rcx)1312; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm8, (%r8)1313; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, (%r9)1314; AVX512DQ-FCP-NEXT: vzeroupper1315; AVX512DQ-FCP-NEXT: retq1316;1317; AVX512BW-LABEL: load_i64_stride5_vf8:1318; AVX512BW: # %bb.0:1319; AVX512BW-NEXT: vmovdqa64 256(%rdi), %zmm01320; AVX512BW-NEXT: vmovdqa64 (%rdi), %zmm21321; AVX512BW-NEXT: vmovdqa64 64(%rdi), %zmm11322; AVX512BW-NEXT: vmovdqa64 128(%rdi), %zmm41323; AVX512BW-NEXT: vmovdqa64 192(%rdi), %zmm51324; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm3 = [12,1,6,0,12,1,6,0]1325; AVX512BW-NEXT: # zmm3 = mem[0,1,2,3,0,1,2,3]1326; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = [5,10,15,0,5,10,15,0]1327; AVX512BW-NEXT: # zmm6 = mem[0,1,2,3,0,1,2,3]1328; AVX512BW-NEXT: vpermi2q %zmm5, %zmm4, %zmm61329; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [6,11,0,1,6,11,0,1]1330; AVX512BW-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]1331; AVX512BW-NEXT: vpermi2q %zmm5, %zmm4, %zmm71332; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [7,12,0,2,7,12,0,2]1333; AVX512BW-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]1334; AVX512BW-NEXT: vpermi2q %zmm5, %zmm4, %zmm81335; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm9 = [0,5,0,11,0,5,0,11]1336; AVX512BW-NEXT: # zmm9 = mem[0,1,2,3,0,1,2,3]1337; AVX512BW-NEXT: vpermi2q %zmm4, %zmm5, %zmm91338; AVX512BW-NEXT: vpermt2q %zmm4, %zmm3, %zmm51339; AVX512BW-NEXT: vpmovsxbq {{.*#+}} ymm4 = [0,5,10,15]1340; AVX512BW-NEXT: vpermi2q %zmm1, %zmm2, %zmm41341; AVX512BW-NEXT: vshufi64x2 {{.*#+}} zmm4 = zmm4[0,1,2,3],zmm5[4,5,6,7]1342; AVX512BW-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,1,2,3,4,5,6,11]1343; AVX512BW-NEXT: vpermi2q %zmm0, %zmm4, %zmm51344; AVX512BW-NEXT: vpmovsxbq {{.*#+}} ymm4 = [1,6,11,0]1345; AVX512BW-NEXT: vpermi2q %zmm1, %zmm2, %zmm41346; AVX512BW-NEXT: movb $7, %al1347; AVX512BW-NEXT: kmovd %eax, %k11348; AVX512BW-NEXT: vmovdqa64 %zmm4, %zmm6 {%k1}1349; AVX512BW-NEXT: vpmovsxbq {{.*#+}} zmm4 = [0,1,2,3,4,5,6,12]1350; AVX512BW-NEXT: vpermi2q %zmm0, %zmm6, %zmm41351; AVX512BW-NEXT: vpmovsxbq {{.*#+}} ymm6 = [2,7,12,0]1352; AVX512BW-NEXT: vpermi2q %zmm1, %zmm2, %zmm61353; AVX512BW-NEXT: movb $56, %al1354; AVX512BW-NEXT: kmovd %eax, %k11355; AVX512BW-NEXT: vmovdqa64 %zmm7, %zmm6 {%k1}1356; AVX512BW-NEXT: vpmovsxbq {{.*#+}} zmm7 = [0,1,2,3,4,5,8,13]1357; AVX512BW-NEXT: vpermi2q %zmm0, %zmm6, %zmm71358; AVX512BW-NEXT: vpmovsxbq {{.*#+}} ymm6 = [11,0,5,0]1359; AVX512BW-NEXT: vpermi2q %zmm2, %zmm1, %zmm61360; AVX512BW-NEXT: vmovdqa64 %zmm8, %zmm6 {%k1}1361; AVX512BW-NEXT: vpmovsxbq {{.*#+}} zmm8 = [0,1,2,3,4,5,9,14]1362; AVX512BW-NEXT: vpermi2q %zmm0, %zmm6, %zmm81363; AVX512BW-NEXT: vpermt2q %zmm2, %zmm3, %zmm11364; AVX512BW-NEXT: vmovdqa64 %zmm9, %zmm1 {%k1}1365; AVX512BW-NEXT: vpmovsxbq {{.*#+}} zmm2 = [0,1,2,3,4,5,10,15]1366; AVX512BW-NEXT: vpermi2q %zmm0, %zmm1, %zmm21367; AVX512BW-NEXT: vmovdqa64 %zmm5, (%rsi)1368; AVX512BW-NEXT: vmovdqa64 %zmm4, (%rdx)1369; AVX512BW-NEXT: vmovdqa64 %zmm7, (%rcx)1370; AVX512BW-NEXT: vmovdqa64 %zmm8, (%r8)1371; AVX512BW-NEXT: vmovdqa64 %zmm2, (%r9)1372; AVX512BW-NEXT: vzeroupper1373; AVX512BW-NEXT: retq1374;1375; AVX512BW-FCP-LABEL: load_i64_stride5_vf8:1376; AVX512BW-FCP: # %bb.0:1377; AVX512BW-FCP-NEXT: vmovdqa64 256(%rdi), %zmm01378; AVX512BW-FCP-NEXT: vmovdqa64 (%rdi), %zmm21379; AVX512BW-FCP-NEXT: vmovdqa64 64(%rdi), %zmm11380; AVX512BW-FCP-NEXT: vmovdqa64 128(%rdi), %zmm41381; AVX512BW-FCP-NEXT: vmovdqa64 192(%rdi), %zmm51382; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm3 = [12,1,6,0,12,1,6,0]1383; AVX512BW-FCP-NEXT: # zmm3 = mem[0,1,2,3,0,1,2,3]1384; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = [5,10,15,0,5,10,15,0]1385; AVX512BW-FCP-NEXT: # zmm6 = mem[0,1,2,3,0,1,2,3]1386; AVX512BW-FCP-NEXT: vpermi2q %zmm5, %zmm4, %zmm61387; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [6,11,0,1,6,11,0,1]1388; AVX512BW-FCP-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]1389; AVX512BW-FCP-NEXT: vpermi2q %zmm5, %zmm4, %zmm71390; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [7,12,0,2,7,12,0,2]1391; AVX512BW-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]1392; AVX512BW-FCP-NEXT: vpermi2q %zmm5, %zmm4, %zmm81393; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm9 = [0,5,0,11,0,5,0,11]1394; AVX512BW-FCP-NEXT: # zmm9 = mem[0,1,2,3,0,1,2,3]1395; AVX512BW-FCP-NEXT: vpermi2q %zmm4, %zmm5, %zmm91396; AVX512BW-FCP-NEXT: vpermt2q %zmm4, %zmm3, %zmm51397; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} ymm4 = [0,5,10,15]1398; AVX512BW-FCP-NEXT: vpermi2q %zmm1, %zmm2, %zmm41399; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm4 = zmm4[0,1,2,3],zmm5[4,5,6,7]1400; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,1,2,3,4,5,6,11]1401; AVX512BW-FCP-NEXT: vpermi2q %zmm0, %zmm4, %zmm51402; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} ymm4 = [1,6,11,0]1403; AVX512BW-FCP-NEXT: vpermi2q %zmm1, %zmm2, %zmm41404; AVX512BW-FCP-NEXT: movb $7, %al1405; AVX512BW-FCP-NEXT: kmovd %eax, %k11406; AVX512BW-FCP-NEXT: vmovdqa64 %zmm4, %zmm6 {%k1}1407; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm4 = [0,1,2,3,4,5,6,12]1408; AVX512BW-FCP-NEXT: vpermi2q %zmm0, %zmm6, %zmm41409; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} ymm6 = [2,7,12,0]1410; AVX512BW-FCP-NEXT: vpermi2q %zmm1, %zmm2, %zmm61411; AVX512BW-FCP-NEXT: movb $56, %al1412; AVX512BW-FCP-NEXT: kmovd %eax, %k11413; AVX512BW-FCP-NEXT: vmovdqa64 %zmm7, %zmm6 {%k1}1414; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm7 = [0,1,2,3,4,5,8,13]1415; AVX512BW-FCP-NEXT: vpermi2q %zmm0, %zmm6, %zmm71416; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} ymm6 = [11,0,5,0]1417; AVX512BW-FCP-NEXT: vpermi2q %zmm2, %zmm1, %zmm61418; AVX512BW-FCP-NEXT: vmovdqa64 %zmm8, %zmm6 {%k1}1419; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm8 = [0,1,2,3,4,5,9,14]1420; AVX512BW-FCP-NEXT: vpermi2q %zmm0, %zmm6, %zmm81421; AVX512BW-FCP-NEXT: vpermt2q %zmm2, %zmm3, %zmm11422; AVX512BW-FCP-NEXT: vmovdqa64 %zmm9, %zmm1 {%k1}1423; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm2 = [0,1,2,3,4,5,10,15]1424; AVX512BW-FCP-NEXT: vpermi2q %zmm0, %zmm1, %zmm21425; AVX512BW-FCP-NEXT: vmovdqa64 %zmm5, (%rsi)1426; AVX512BW-FCP-NEXT: vmovdqa64 %zmm4, (%rdx)1427; AVX512BW-FCP-NEXT: vmovdqa64 %zmm7, (%rcx)1428; AVX512BW-FCP-NEXT: vmovdqa64 %zmm8, (%r8)1429; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, (%r9)1430; AVX512BW-FCP-NEXT: vzeroupper1431; AVX512BW-FCP-NEXT: retq1432;1433; AVX512DQ-BW-LABEL: load_i64_stride5_vf8:1434; AVX512DQ-BW: # %bb.0:1435; AVX512DQ-BW-NEXT: vmovdqa64 256(%rdi), %zmm01436; AVX512DQ-BW-NEXT: vmovdqa64 (%rdi), %zmm21437; AVX512DQ-BW-NEXT: vmovdqa64 64(%rdi), %zmm11438; AVX512DQ-BW-NEXT: vmovdqa64 128(%rdi), %zmm41439; AVX512DQ-BW-NEXT: vmovdqa64 192(%rdi), %zmm51440; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm3 = [12,1,6,0,12,1,6,0]1441; AVX512DQ-BW-NEXT: # zmm3 = mem[0,1,2,3,0,1,2,3]1442; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = [5,10,15,0,5,10,15,0]1443; AVX512DQ-BW-NEXT: # zmm6 = mem[0,1,2,3,0,1,2,3]1444; AVX512DQ-BW-NEXT: vpermi2q %zmm5, %zmm4, %zmm61445; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [6,11,0,1,6,11,0,1]1446; AVX512DQ-BW-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]1447; AVX512DQ-BW-NEXT: vpermi2q %zmm5, %zmm4, %zmm71448; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [7,12,0,2,7,12,0,2]1449; AVX512DQ-BW-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]1450; AVX512DQ-BW-NEXT: vpermi2q %zmm5, %zmm4, %zmm81451; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm9 = [0,5,0,11,0,5,0,11]1452; AVX512DQ-BW-NEXT: # zmm9 = mem[0,1,2,3,0,1,2,3]1453; AVX512DQ-BW-NEXT: vpermi2q %zmm4, %zmm5, %zmm91454; AVX512DQ-BW-NEXT: vpermt2q %zmm4, %zmm3, %zmm51455; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} ymm4 = [0,5,10,15]1456; AVX512DQ-BW-NEXT: vpermi2q %zmm1, %zmm2, %zmm41457; AVX512DQ-BW-NEXT: vshufi64x2 {{.*#+}} zmm4 = zmm4[0,1,2,3],zmm5[4,5,6,7]1458; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,1,2,3,4,5,6,11]1459; AVX512DQ-BW-NEXT: vpermi2q %zmm0, %zmm4, %zmm51460; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} ymm4 = [1,6,11,0]1461; AVX512DQ-BW-NEXT: vpermi2q %zmm1, %zmm2, %zmm41462; AVX512DQ-BW-NEXT: movb $7, %al1463; AVX512DQ-BW-NEXT: kmovd %eax, %k11464; AVX512DQ-BW-NEXT: vmovdqa64 %zmm4, %zmm6 {%k1}1465; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} zmm4 = [0,1,2,3,4,5,6,12]1466; AVX512DQ-BW-NEXT: vpermi2q %zmm0, %zmm6, %zmm41467; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} ymm6 = [2,7,12,0]1468; AVX512DQ-BW-NEXT: vpermi2q %zmm1, %zmm2, %zmm61469; AVX512DQ-BW-NEXT: movb $56, %al1470; AVX512DQ-BW-NEXT: kmovd %eax, %k11471; AVX512DQ-BW-NEXT: vmovdqa64 %zmm7, %zmm6 {%k1}1472; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} zmm7 = [0,1,2,3,4,5,8,13]1473; AVX512DQ-BW-NEXT: vpermi2q %zmm0, %zmm6, %zmm71474; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} ymm6 = [11,0,5,0]1475; AVX512DQ-BW-NEXT: vpermi2q %zmm2, %zmm1, %zmm61476; AVX512DQ-BW-NEXT: vmovdqa64 %zmm8, %zmm6 {%k1}1477; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} zmm8 = [0,1,2,3,4,5,9,14]1478; AVX512DQ-BW-NEXT: vpermi2q %zmm0, %zmm6, %zmm81479; AVX512DQ-BW-NEXT: vpermt2q %zmm2, %zmm3, %zmm11480; AVX512DQ-BW-NEXT: vmovdqa64 %zmm9, %zmm1 {%k1}1481; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} zmm2 = [0,1,2,3,4,5,10,15]1482; AVX512DQ-BW-NEXT: vpermi2q %zmm0, %zmm1, %zmm21483; AVX512DQ-BW-NEXT: vmovdqa64 %zmm5, (%rsi)1484; AVX512DQ-BW-NEXT: vmovdqa64 %zmm4, (%rdx)1485; AVX512DQ-BW-NEXT: vmovdqa64 %zmm7, (%rcx)1486; AVX512DQ-BW-NEXT: vmovdqa64 %zmm8, (%r8)1487; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, (%r9)1488; AVX512DQ-BW-NEXT: vzeroupper1489; AVX512DQ-BW-NEXT: retq1490;1491; AVX512DQ-BW-FCP-LABEL: load_i64_stride5_vf8:1492; AVX512DQ-BW-FCP: # %bb.0:1493; AVX512DQ-BW-FCP-NEXT: vmovdqa64 256(%rdi), %zmm01494; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rdi), %zmm21495; AVX512DQ-BW-FCP-NEXT: vmovdqa64 64(%rdi), %zmm11496; AVX512DQ-BW-FCP-NEXT: vmovdqa64 128(%rdi), %zmm41497; AVX512DQ-BW-FCP-NEXT: vmovdqa64 192(%rdi), %zmm51498; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm3 = [12,1,6,0,12,1,6,0]1499; AVX512DQ-BW-FCP-NEXT: # zmm3 = mem[0,1,2,3,0,1,2,3]1500; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = [5,10,15,0,5,10,15,0]1501; AVX512DQ-BW-FCP-NEXT: # zmm6 = mem[0,1,2,3,0,1,2,3]1502; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm5, %zmm4, %zmm61503; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [6,11,0,1,6,11,0,1]1504; AVX512DQ-BW-FCP-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]1505; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm5, %zmm4, %zmm71506; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm8 = [7,12,0,2,7,12,0,2]1507; AVX512DQ-BW-FCP-NEXT: # zmm8 = mem[0,1,2,3,0,1,2,3]1508; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm5, %zmm4, %zmm81509; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm9 = [0,5,0,11,0,5,0,11]1510; AVX512DQ-BW-FCP-NEXT: # zmm9 = mem[0,1,2,3,0,1,2,3]1511; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm4, %zmm5, %zmm91512; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm4, %zmm3, %zmm51513; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} ymm4 = [0,5,10,15]1514; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm1, %zmm2, %zmm41515; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm4 = zmm4[0,1,2,3],zmm5[4,5,6,7]1516; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm5 = [0,1,2,3,4,5,6,11]1517; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm0, %zmm4, %zmm51518; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} ymm4 = [1,6,11,0]1519; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm1, %zmm2, %zmm41520; AVX512DQ-BW-FCP-NEXT: movb $7, %al1521; AVX512DQ-BW-FCP-NEXT: kmovd %eax, %k11522; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm4, %zmm6 {%k1}1523; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm4 = [0,1,2,3,4,5,6,12]1524; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm0, %zmm6, %zmm41525; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} ymm6 = [2,7,12,0]1526; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm1, %zmm2, %zmm61527; AVX512DQ-BW-FCP-NEXT: movb $56, %al1528; AVX512DQ-BW-FCP-NEXT: kmovd %eax, %k11529; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm7, %zmm6 {%k1}1530; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm7 = [0,1,2,3,4,5,8,13]1531; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm0, %zmm6, %zmm71532; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} ymm6 = [11,0,5,0]1533; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm2, %zmm1, %zmm61534; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm8, %zmm6 {%k1}1535; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm8 = [0,1,2,3,4,5,9,14]1536; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm0, %zmm6, %zmm81537; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm2, %zmm3, %zmm11538; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm9, %zmm1 {%k1}1539; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm2 = [0,1,2,3,4,5,10,15]1540; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm0, %zmm1, %zmm21541; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm5, (%rsi)1542; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm4, (%rdx)1543; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm7, (%rcx)1544; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm8, (%r8)1545; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, (%r9)1546; AVX512DQ-BW-FCP-NEXT: vzeroupper1547; AVX512DQ-BW-FCP-NEXT: retq1548 %wide.vec = load <40 x i64>, ptr %in.vec, align 641549 %strided.vec0 = shufflevector <40 x i64> %wide.vec, <40 x i64> poison, <8 x i32> <i32 0, i32 5, i32 10, i32 15, i32 20, i32 25, i32 30, i32 35>1550 %strided.vec1 = shufflevector <40 x i64> %wide.vec, <40 x i64> poison, <8 x i32> <i32 1, i32 6, i32 11, i32 16, i32 21, i32 26, i32 31, i32 36>1551 %strided.vec2 = shufflevector <40 x i64> %wide.vec, <40 x i64> poison, <8 x i32> <i32 2, i32 7, i32 12, i32 17, i32 22, i32 27, i32 32, i32 37>1552 %strided.vec3 = shufflevector <40 x i64> %wide.vec, <40 x i64> poison, <8 x i32> <i32 3, i32 8, i32 13, i32 18, i32 23, i32 28, i32 33, i32 38>1553 %strided.vec4 = shufflevector <40 x i64> %wide.vec, <40 x i64> poison, <8 x i32> <i32 4, i32 9, i32 14, i32 19, i32 24, i32 29, i32 34, i32 39>1554 store <8 x i64> %strided.vec0, ptr %out.vec0, align 641555 store <8 x i64> %strided.vec1, ptr %out.vec1, align 641556 store <8 x i64> %strided.vec2, ptr %out.vec2, align 641557 store <8 x i64> %strided.vec3, ptr %out.vec3, align 641558 store <8 x i64> %strided.vec4, ptr %out.vec4, align 641559 ret void1560}1561 1562define void @load_i64_stride5_vf16(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr %out.vec2, ptr %out.vec3, ptr %out.vec4) nounwind {1563; SSE-LABEL: load_i64_stride5_vf16:1564; SSE: # %bb.0:1565; SSE-NEXT: subq $280, %rsp # imm = 0x1181566; SSE-NEXT: movapd 224(%rdi), %xmm31567; SSE-NEXT: movapd 144(%rdi), %xmm21568; SSE-NEXT: movapd 64(%rdi), %xmm11569; SSE-NEXT: movapd 176(%rdi), %xmm41570; SSE-NEXT: movapd 96(%rdi), %xmm51571; SSE-NEXT: movapd 208(%rdi), %xmm71572; SSE-NEXT: movapd 128(%rdi), %xmm81573; SSE-NEXT: movapd (%rdi), %xmm101574; SSE-NEXT: movapd 16(%rdi), %xmm61575; SSE-NEXT: movapd 32(%rdi), %xmm141576; SSE-NEXT: movapd 48(%rdi), %xmm91577; SSE-NEXT: movapd 160(%rdi), %xmm111578; SSE-NEXT: movapd 192(%rdi), %xmm131579; SSE-NEXT: movapd 80(%rdi), %xmm121580; SSE-NEXT: movapd 112(%rdi), %xmm01581; SSE-NEXT: movapd %xmm14, %xmm151582; SSE-NEXT: movsd {{.*#+}} xmm15 = xmm10[0],xmm15[1]1583; SSE-NEXT: movapd %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1584; SSE-NEXT: shufpd {{.*#+}} xmm10 = xmm10[1],xmm9[0]1585; SSE-NEXT: movapd %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1586; SSE-NEXT: movsd {{.*#+}} xmm9 = xmm6[0],xmm9[1]1587; SSE-NEXT: movapd %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1588; SSE-NEXT: shufpd {{.*#+}} xmm6 = xmm6[1],xmm1[0]1589; SSE-NEXT: movapd %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1590; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm14[0],xmm1[1]1591; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1592; SSE-NEXT: movapd %xmm0, %xmm11593; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm12[0],xmm1[1]1594; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1595; SSE-NEXT: shufpd {{.*#+}} xmm12 = xmm12[1],xmm8[0]1596; SSE-NEXT: movapd %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1597; SSE-NEXT: movsd {{.*#+}} xmm8 = xmm5[0],xmm8[1]1598; SSE-NEXT: movapd %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1599; SSE-NEXT: shufpd {{.*#+}} xmm5 = xmm5[1],xmm2[0]1600; SSE-NEXT: movapd %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1601; SSE-NEXT: movsd {{.*#+}} xmm2 = xmm0[0],xmm2[1]1602; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1603; SSE-NEXT: movapd %xmm13, %xmm01604; SSE-NEXT: movsd {{.*#+}} xmm0 = xmm11[0],xmm0[1]1605; SSE-NEXT: movapd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1606; SSE-NEXT: shufpd {{.*#+}} xmm11 = xmm11[1],xmm7[0]1607; SSE-NEXT: movapd %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1608; SSE-NEXT: movsd {{.*#+}} xmm7 = xmm4[0],xmm7[1]1609; SSE-NEXT: movapd %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1610; SSE-NEXT: shufpd {{.*#+}} xmm4 = xmm4[1],xmm3[0]1611; SSE-NEXT: movapd %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1612; SSE-NEXT: movsd {{.*#+}} xmm3 = xmm13[0],xmm3[1]1613; SSE-NEXT: movapd %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1614; SSE-NEXT: movapd 240(%rdi), %xmm21615; SSE-NEXT: movapd 272(%rdi), %xmm01616; SSE-NEXT: movapd %xmm0, %xmm11617; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]1618; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1619; SSE-NEXT: movapd 288(%rdi), %xmm11620; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]1621; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1622; SSE-NEXT: movapd 256(%rdi), %xmm21623; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]1624; SSE-NEXT: movapd %xmm1, (%rsp) # 16-byte Spill1625; SSE-NEXT: movapd 304(%rdi), %xmm11626; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]1627; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1628; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]1629; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1630; SSE-NEXT: movapd 320(%rdi), %xmm151631; SSE-NEXT: movapd 352(%rdi), %xmm01632; SSE-NEXT: movapd %xmm0, %xmm11633; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm15[0],xmm1[1]1634; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1635; SSE-NEXT: movapd 368(%rdi), %xmm11636; SSE-NEXT: shufpd {{.*#+}} xmm15 = xmm15[1],xmm1[0]1637; SSE-NEXT: movapd 336(%rdi), %xmm21638; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]1639; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1640; SSE-NEXT: movapd 384(%rdi), %xmm11641; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]1642; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1643; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]1644; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1645; SSE-NEXT: movapd 400(%rdi), %xmm111646; SSE-NEXT: movapd 432(%rdi), %xmm01647; SSE-NEXT: movapd %xmm0, %xmm131648; SSE-NEXT: movsd {{.*#+}} xmm13 = xmm11[0],xmm13[1]1649; SSE-NEXT: movapd 448(%rdi), %xmm121650; SSE-NEXT: shufpd {{.*#+}} xmm11 = xmm11[1],xmm12[0]1651; SSE-NEXT: movapd 416(%rdi), %xmm141652; SSE-NEXT: movsd {{.*#+}} xmm12 = xmm14[0],xmm12[1]1653; SSE-NEXT: movapd 464(%rdi), %xmm11654; SSE-NEXT: shufpd {{.*#+}} xmm14 = xmm14[1],xmm1[0]1655; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]1656; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1657; SSE-NEXT: movapd 480(%rdi), %xmm41658; SSE-NEXT: movapd 512(%rdi), %xmm01659; SSE-NEXT: movapd %xmm0, %xmm61660; SSE-NEXT: movsd {{.*#+}} xmm6 = xmm4[0],xmm6[1]1661; SSE-NEXT: movapd 528(%rdi), %xmm71662; SSE-NEXT: shufpd {{.*#+}} xmm4 = xmm4[1],xmm7[0]1663; SSE-NEXT: movapd 496(%rdi), %xmm91664; SSE-NEXT: movsd {{.*#+}} xmm7 = xmm9[0],xmm7[1]1665; SSE-NEXT: movapd 544(%rdi), %xmm101666; SSE-NEXT: shufpd {{.*#+}} xmm9 = xmm9[1],xmm10[0]1667; SSE-NEXT: movsd {{.*#+}} xmm10 = xmm0[0],xmm10[1]1668; SSE-NEXT: movapd 560(%rdi), %xmm01669; SSE-NEXT: movapd 592(%rdi), %xmm51670; SSE-NEXT: movapd %xmm5, %xmm21671; SSE-NEXT: movsd {{.*#+}} xmm2 = xmm0[0],xmm2[1]1672; SSE-NEXT: movapd 608(%rdi), %xmm11673; SSE-NEXT: shufpd {{.*#+}} xmm0 = xmm0[1],xmm1[0]1674; SSE-NEXT: movapd 576(%rdi), %xmm31675; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm3[0],xmm1[1]1676; SSE-NEXT: movapd 624(%rdi), %xmm81677; SSE-NEXT: shufpd {{.*#+}} xmm3 = xmm3[1],xmm8[0]1678; SSE-NEXT: movsd {{.*#+}} xmm8 = xmm5[0],xmm8[1]1679; SSE-NEXT: movapd %xmm6, 96(%rsi)1680; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload1681; SSE-NEXT: movaps %xmm5, 32(%rsi)1682; SSE-NEXT: movapd %xmm2, 112(%rsi)1683; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload1684; SSE-NEXT: movaps %xmm2, 48(%rsi)1685; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload1686; SSE-NEXT: movaps %xmm2, 64(%rsi)1687; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload1688; SSE-NEXT: movaps %xmm2, (%rsi)1689; SSE-NEXT: movapd %xmm13, 80(%rsi)1690; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload1691; SSE-NEXT: movaps %xmm2, 16(%rsi)1692; SSE-NEXT: movapd %xmm4, 96(%rdx)1693; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload1694; SSE-NEXT: movaps %xmm2, 32(%rdx)1695; SSE-NEXT: movapd %xmm0, 112(%rdx)1696; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload1697; SSE-NEXT: movaps %xmm0, 48(%rdx)1698; SSE-NEXT: movapd %xmm15, 64(%rdx)1699; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload1700; SSE-NEXT: movaps %xmm0, (%rdx)1701; SSE-NEXT: movapd %xmm11, 80(%rdx)1702; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload1703; SSE-NEXT: movaps %xmm0, 16(%rdx)1704; SSE-NEXT: movapd %xmm7, 96(%rcx)1705; SSE-NEXT: movapd %xmm1, 112(%rcx)1706; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload1707; SSE-NEXT: movaps %xmm0, 64(%rcx)1708; SSE-NEXT: movapd %xmm12, 80(%rcx)1709; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload1710; SSE-NEXT: movaps %xmm0, 32(%rcx)1711; SSE-NEXT: movaps (%rsp), %xmm0 # 16-byte Reload1712; SSE-NEXT: movaps %xmm0, 48(%rcx)1713; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload1714; SSE-NEXT: movaps %xmm0, (%rcx)1715; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload1716; SSE-NEXT: movaps %xmm0, 16(%rcx)1717; SSE-NEXT: movapd %xmm3, 112(%r8)1718; SSE-NEXT: movapd %xmm9, 96(%r8)1719; SSE-NEXT: movapd %xmm14, 80(%r8)1720; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload1721; SSE-NEXT: movaps %xmm0, 64(%r8)1722; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload1723; SSE-NEXT: movaps %xmm0, 48(%r8)1724; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload1725; SSE-NEXT: movaps %xmm0, 32(%r8)1726; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload1727; SSE-NEXT: movaps %xmm0, 16(%r8)1728; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload1729; SSE-NEXT: movaps %xmm0, (%r8)1730; SSE-NEXT: movapd %xmm8, 112(%r9)1731; SSE-NEXT: movapd %xmm10, 96(%r9)1732; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload1733; SSE-NEXT: movaps %xmm0, 80(%r9)1734; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload1735; SSE-NEXT: movaps %xmm0, 64(%r9)1736; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload1737; SSE-NEXT: movaps %xmm0, 48(%r9)1738; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload1739; SSE-NEXT: movaps %xmm0, 32(%r9)1740; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload1741; SSE-NEXT: movaps %xmm0, 16(%r9)1742; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload1743; SSE-NEXT: movaps %xmm0, (%r9)1744; SSE-NEXT: addq $280, %rsp # imm = 0x1181745; SSE-NEXT: retq1746;1747; AVX-LABEL: load_i64_stride5_vf16:1748; AVX: # %bb.0:1749; AVX-NEXT: subq $360, %rsp # imm = 0x1681750; AVX-NEXT: vmovapd 96(%rdi), %ymm01751; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill1752; AVX-NEXT: vmovapd 64(%rdi), %ymm51753; AVX-NEXT: vmovapd 576(%rdi), %ymm71754; AVX-NEXT: vmovupd %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill1755; AVX-NEXT: vmovapd 544(%rdi), %ymm11756; AVX-NEXT: vmovapd 256(%rdi), %ymm21757; AVX-NEXT: vmovupd %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill1758; AVX-NEXT: vmovapd 224(%rdi), %ymm31759; AVX-NEXT: vblendpd {{.*#+}} ymm2 = ymm3[0,1,2],ymm2[3]1760; AVX-NEXT: vmovapd 192(%rdi), %xmm61761; AVX-NEXT: vmovapd %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1762; AVX-NEXT: vmovapd 160(%rdi), %xmm41763; AVX-NEXT: vblendpd {{.*#+}} xmm6 = xmm4[0],xmm6[1]1764; AVX-NEXT: vblendpd {{.*#+}} ymm2 = ymm6[0,1],ymm2[2,3]1765; AVX-NEXT: vmovupd %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill1766; AVX-NEXT: vblendpd {{.*#+}} ymm2 = ymm1[0,1,2],ymm7[3]1767; AVX-NEXT: vmovapd 512(%rdi), %xmm61768; AVX-NEXT: vmovapd %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1769; AVX-NEXT: vmovapd 480(%rdi), %xmm111770; AVX-NEXT: vblendpd {{.*#+}} xmm8 = xmm11[0],xmm6[1]1771; AVX-NEXT: vblendpd {{.*#+}} ymm2 = ymm8[0,1],ymm2[2,3]1772; AVX-NEXT: vmovupd %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill1773; AVX-NEXT: vblendpd {{.*#+}} ymm9 = ymm5[0,1,2],ymm0[3]1774; AVX-NEXT: vmovapd (%rdi), %xmm21775; AVX-NEXT: vmovapd 32(%rdi), %xmm01776; AVX-NEXT: vmovapd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1777; AVX-NEXT: vblendpd {{.*#+}} xmm10 = xmm2[0],xmm0[1]1778; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm10[0,1],ymm9[2,3]1779; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill1780; AVX-NEXT: vmovapd 416(%rdi), %ymm01781; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill1782; AVX-NEXT: vmovapd 384(%rdi), %ymm141783; AVX-NEXT: vblendpd {{.*#+}} ymm15 = ymm14[0,1,2],ymm0[3]1784; AVX-NEXT: vmovapd 352(%rdi), %xmm61785; AVX-NEXT: vmovapd %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1786; AVX-NEXT: vmovapd 320(%rdi), %xmm01787; AVX-NEXT: vblendpd {{.*#+}} xmm13 = xmm0[0],xmm6[1]1788; AVX-NEXT: vblendpd {{.*#+}} ymm6 = ymm13[0,1],ymm15[2,3]1789; AVX-NEXT: vmovupd %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill1790; AVX-NEXT: vmovaps 288(%rdi), %xmm131791; AVX-NEXT: vinsertf128 $1, %xmm13, %ymm0, %ymm151792; AVX-NEXT: vshufpd {{.*#+}} ymm3 = ymm3[0],ymm15[1],ymm3[3],ymm15[2]1793; AVX-NEXT: vmovdqa 208(%rdi), %xmm151794; AVX-NEXT: vpalignr {{.*#+}} xmm4 = xmm4[8,9,10,11,12,13,14,15],xmm15[0,1,2,3,4,5,6,7]1795; AVX-NEXT: vblendpd {{.*#+}} ymm3 = ymm4[0,1],ymm3[2,3]1796; AVX-NEXT: vmovupd %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill1797; AVX-NEXT: vmovaps 608(%rdi), %xmm31798; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm41799; AVX-NEXT: vshufpd {{.*#+}} ymm1 = ymm1[0],ymm4[1],ymm1[3],ymm4[2]1800; AVX-NEXT: vmovdqa 528(%rdi), %xmm41801; AVX-NEXT: vpalignr {{.*#+}} xmm11 = xmm11[8,9,10,11,12,13,14,15],xmm4[0,1,2,3,4,5,6,7]1802; AVX-NEXT: vblendpd {{.*#+}} ymm1 = ymm11[0,1],ymm1[2,3]1803; AVX-NEXT: vmovupd %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill1804; AVX-NEXT: vmovaps 128(%rdi), %xmm61805; AVX-NEXT: vinsertf128 $1, %xmm6, %ymm0, %ymm111806; AVX-NEXT: vshufpd {{.*#+}} ymm5 = ymm5[0],ymm11[1],ymm5[3],ymm11[2]1807; AVX-NEXT: vmovdqa 48(%rdi), %xmm111808; AVX-NEXT: vpalignr {{.*#+}} xmm2 = xmm2[8,9,10,11,12,13,14,15],xmm11[0,1,2,3,4,5,6,7]1809; AVX-NEXT: vblendpd {{.*#+}} ymm1 = ymm2[0,1],ymm5[2,3]1810; AVX-NEXT: vmovupd %ymm1, (%rsp) # 32-byte Spill1811; AVX-NEXT: vmovaps 448(%rdi), %xmm21812; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm51813; AVX-NEXT: vshufpd {{.*#+}} ymm5 = ymm14[0],ymm5[1],ymm14[3],ymm5[2]1814; AVX-NEXT: vmovdqa 368(%rdi), %xmm141815; AVX-NEXT: vpalignr {{.*#+}} xmm0 = xmm0[8,9,10,11,12,13,14,15],xmm14[0,1,2,3,4,5,6,7]1816; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm0[0,1],ymm5[2,3]1817; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill1818; AVX-NEXT: vmovaps 256(%rdi), %xmm01819; AVX-NEXT: vblendps {{.*#+}} xmm13 = xmm0[0,1],xmm13[2,3]1820; AVX-NEXT: vinsertf128 $1, %xmm13, %ymm0, %ymm131821; AVX-NEXT: vmovdqa 176(%rdi), %xmm91822; AVX-NEXT: vpblendw {{.*#+}} xmm15 = xmm9[0,1,2,3],xmm15[4,5,6,7]1823; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm15[0,1,2,3],ymm13[4,5,6,7]1824; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill1825; AVX-NEXT: vmovaps 576(%rdi), %xmm51826; AVX-NEXT: vblendps {{.*#+}} xmm3 = xmm5[0,1],xmm3[2,3]1827; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm31828; AVX-NEXT: vmovdqa 496(%rdi), %xmm71829; AVX-NEXT: vpblendw {{.*#+}} xmm4 = xmm7[0,1,2,3],xmm4[4,5,6,7]1830; AVX-NEXT: vblendps {{.*#+}} ymm15 = ymm4[0,1,2,3],ymm3[4,5,6,7]1831; AVX-NEXT: vmovaps 96(%rdi), %xmm41832; AVX-NEXT: vblendps {{.*#+}} xmm3 = xmm4[0,1],xmm6[2,3]1833; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm31834; AVX-NEXT: vmovdqa 16(%rdi), %xmm101835; AVX-NEXT: vpblendw {{.*#+}} xmm11 = xmm10[0,1,2,3],xmm11[4,5,6,7]1836; AVX-NEXT: vblendps {{.*#+}} ymm13 = ymm11[0,1,2,3],ymm3[4,5,6,7]1837; AVX-NEXT: vmovaps 416(%rdi), %xmm61838; AVX-NEXT: vblendps {{.*#+}} xmm2 = xmm6[0,1],xmm2[2,3]1839; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm21840; AVX-NEXT: vmovdqa 336(%rdi), %xmm81841; AVX-NEXT: vpblendw {{.*#+}} xmm14 = xmm8[0,1,2,3],xmm14[4,5,6,7]1842; AVX-NEXT: vblendps {{.*#+}} ymm11 = ymm14[0,1,2,3],ymm2[4,5,6,7]1843; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm01844; AVX-NEXT: vmovapd 288(%rdi), %ymm141845; AVX-NEXT: vshufpd {{.*#+}} ymm0 = ymm0[0],ymm14[1],ymm0[3],ymm14[2]1846; AVX-NEXT: vmovdqa 224(%rdi), %xmm31847; AVX-NEXT: vpalignr {{.*#+}} xmm9 = xmm9[8,9,10,11,12,13,14,15],xmm3[0,1,2,3,4,5,6,7]1848; AVX-NEXT: vblendpd {{.*#+}} ymm12 = ymm9[0,1],ymm0[2,3]1849; AVX-NEXT: vinsertf128 $1, %xmm5, %ymm0, %ymm01850; AVX-NEXT: vmovapd 608(%rdi), %ymm51851; AVX-NEXT: vshufpd {{.*#+}} ymm0 = ymm0[0],ymm5[1],ymm0[3],ymm5[2]1852; AVX-NEXT: vmovdqa 544(%rdi), %xmm21853; AVX-NEXT: vpalignr {{.*#+}} xmm7 = xmm7[8,9,10,11,12,13,14,15],xmm2[0,1,2,3,4,5,6,7]1854; AVX-NEXT: vblendpd {{.*#+}} ymm9 = ymm7[0,1],ymm0[2,3]1855; AVX-NEXT: vinsertf128 $1, %xmm4, %ymm0, %ymm01856; AVX-NEXT: vmovapd 128(%rdi), %ymm41857; AVX-NEXT: vshufpd {{.*#+}} ymm0 = ymm0[0],ymm4[1],ymm0[3],ymm4[2]1858; AVX-NEXT: vmovdqa 64(%rdi), %xmm11859; AVX-NEXT: vpalignr {{.*#+}} xmm10 = xmm10[8,9,10,11,12,13,14,15],xmm1[0,1,2,3,4,5,6,7]1860; AVX-NEXT: vblendpd {{.*#+}} ymm7 = ymm10[0,1],ymm0[2,3]1861; AVX-NEXT: vinsertf128 $1, %xmm6, %ymm0, %ymm61862; AVX-NEXT: vmovapd 448(%rdi), %ymm101863; AVX-NEXT: vshufpd {{.*#+}} ymm6 = ymm6[0],ymm10[1],ymm6[3],ymm10[2]1864; AVX-NEXT: vmovdqa 384(%rdi), %xmm01865; AVX-NEXT: vpalignr {{.*#+}} xmm8 = xmm8[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]1866; AVX-NEXT: vblendpd {{.*#+}} ymm6 = ymm8[0,1],ymm6[2,3]1867; AVX-NEXT: vblendpd $7, {{[-0-9]+}}(%r{{[sb]}}p), %ymm14, %ymm8 # 32-byte Folded Reload1868; AVX-NEXT: # ymm8 = mem[0,1,2],ymm14[3]1869; AVX-NEXT: vpblendw $15, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm3 # 16-byte Folded Reload1870; AVX-NEXT: # xmm3 = mem[0,1,2,3],xmm3[4,5,6,7]1871; AVX-NEXT: vblendpd {{.*#+}} ymm3 = ymm3[0,1],ymm8[2,3]1872; AVX-NEXT: vblendpd $7, {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm5 # 32-byte Folded Reload1873; AVX-NEXT: # ymm5 = mem[0,1,2],ymm5[3]1874; AVX-NEXT: vpblendw $15, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm2 # 16-byte Folded Reload1875; AVX-NEXT: # xmm2 = mem[0,1,2,3],xmm2[4,5,6,7]1876; AVX-NEXT: vblendpd {{.*#+}} ymm2 = ymm2[0,1],ymm5[2,3]1877; AVX-NEXT: vblendpd $7, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4, %ymm4 # 32-byte Folded Reload1878; AVX-NEXT: # ymm4 = mem[0,1,2],ymm4[3]1879; AVX-NEXT: vpblendw $15, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload1880; AVX-NEXT: # xmm1 = mem[0,1,2,3],xmm1[4,5,6,7]1881; AVX-NEXT: vblendpd {{.*#+}} ymm1 = ymm1[0,1],ymm4[2,3]1882; AVX-NEXT: vblendpd $7, {{[-0-9]+}}(%r{{[sb]}}p), %ymm10, %ymm4 # 32-byte Folded Reload1883; AVX-NEXT: # ymm4 = mem[0,1,2],ymm10[3]1884; AVX-NEXT: vpblendw $15, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload1885; AVX-NEXT: # xmm0 = mem[0,1,2,3],xmm0[4,5,6,7]1886; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm0[0,1],ymm4[2,3]1887; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload1888; AVX-NEXT: vmovaps %ymm4, 64(%rsi)1889; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload1890; AVX-NEXT: vmovaps %ymm4, (%rsi)1891; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload1892; AVX-NEXT: vmovaps %ymm4, 96(%rsi)1893; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload1894; AVX-NEXT: vmovaps %ymm4, 32(%rsi)1895; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload1896; AVX-NEXT: vmovaps %ymm4, 64(%rdx)1897; AVX-NEXT: vmovups (%rsp), %ymm4 # 32-byte Reload1898; AVX-NEXT: vmovaps %ymm4, (%rdx)1899; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload1900; AVX-NEXT: vmovaps %ymm4, 96(%rdx)1901; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload1902; AVX-NEXT: vmovaps %ymm4, 32(%rdx)1903; AVX-NEXT: vmovaps %ymm11, 64(%rcx)1904; AVX-NEXT: vmovaps %ymm13, (%rcx)1905; AVX-NEXT: vmovaps %ymm15, 96(%rcx)1906; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload1907; AVX-NEXT: vmovaps %ymm4, 32(%rcx)1908; AVX-NEXT: vmovapd %ymm6, 64(%r8)1909; AVX-NEXT: vmovapd %ymm7, (%r8)1910; AVX-NEXT: vmovapd %ymm9, 96(%r8)1911; AVX-NEXT: vmovapd %ymm12, 32(%r8)1912; AVX-NEXT: vmovapd %ymm0, 64(%r9)1913; AVX-NEXT: vmovapd %ymm1, (%r9)1914; AVX-NEXT: vmovapd %ymm2, 96(%r9)1915; AVX-NEXT: vmovapd %ymm3, 32(%r9)1916; AVX-NEXT: addq $360, %rsp # imm = 0x1681917; AVX-NEXT: vzeroupper1918; AVX-NEXT: retq1919;1920; AVX2-LABEL: load_i64_stride5_vf16:1921; AVX2: # %bb.0:1922; AVX2-NEXT: subq $360, %rsp # imm = 0x1681923; AVX2-NEXT: vmovdqa 96(%rdi), %ymm141924; AVX2-NEXT: vmovdqa 64(%rdi), %ymm41925; AVX2-NEXT: vmovdqa 576(%rdi), %ymm121926; AVX2-NEXT: vmovdqa 544(%rdi), %ymm11927; AVX2-NEXT: vmovdqa 256(%rdi), %ymm111928; AVX2-NEXT: vmovdqa 224(%rdi), %ymm21929; AVX2-NEXT: vpblendd {{.*#+}} ymm3 = ymm2[0,1,2,3,4,5],ymm11[6,7]1930; AVX2-NEXT: vmovdqa 192(%rdi), %xmm01931; AVX2-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1932; AVX2-NEXT: vmovdqa 160(%rdi), %xmm51933; AVX2-NEXT: vpblendd {{.*#+}} xmm8 = xmm5[0,1],xmm0[2,3]1934; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm8[0,1,2,3],ymm3[4,5,6,7]1935; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill1936; AVX2-NEXT: vpblendd {{.*#+}} ymm3 = ymm1[0,1,2,3,4,5],ymm12[6,7]1937; AVX2-NEXT: vmovdqa 512(%rdi), %xmm01938; AVX2-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1939; AVX2-NEXT: vmovdqa 480(%rdi), %xmm81940; AVX2-NEXT: vpblendd {{.*#+}} xmm9 = xmm8[0,1],xmm0[2,3]1941; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm9[0,1,2,3],ymm3[4,5,6,7]1942; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill1943; AVX2-NEXT: vpblendd {{.*#+}} ymm3 = ymm4[0,1,2,3,4,5],ymm14[6,7]1944; AVX2-NEXT: vmovdqa (%rdi), %xmm101945; AVX2-NEXT: vmovdqa 32(%rdi), %xmm01946; AVX2-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1947; AVX2-NEXT: vpblendd {{.*#+}} xmm9 = xmm10[0,1],xmm0[2,3]1948; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm9[0,1,2,3],ymm3[4,5,6,7]1949; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill1950; AVX2-NEXT: vmovdqa 416(%rdi), %ymm91951; AVX2-NEXT: vmovdqa 384(%rdi), %ymm31952; AVX2-NEXT: vpblendd {{.*#+}} ymm13 = ymm3[0,1,2,3,4,5],ymm9[6,7]1953; AVX2-NEXT: vmovdqa 352(%rdi), %xmm61954; AVX2-NEXT: vmovdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill1955; AVX2-NEXT: vmovdqa 320(%rdi), %xmm01956; AVX2-NEXT: vpblendd {{.*#+}} xmm15 = xmm0[0,1],xmm6[2,3]1957; AVX2-NEXT: vpblendd {{.*#+}} ymm6 = ymm15[0,1,2,3],ymm13[4,5,6,7]1958; AVX2-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill1959; AVX2-NEXT: vmovdqa 208(%rdi), %xmm131960; AVX2-NEXT: vpalignr {{.*#+}} xmm5 = xmm5[8,9,10,11,12,13,14,15],xmm13[0,1,2,3,4,5,6,7]1961; AVX2-NEXT: vmovdqa 288(%rdi), %ymm151962; AVX2-NEXT: vpalignr {{.*#+}} ymm2 = ymm2[8,9,10,11,12,13,14,15],ymm15[0,1,2,3,4,5,6,7],ymm2[24,25,26,27,28,29,30,31],ymm15[16,17,18,19,20,21,22,23]1963; AVX2-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,1,2,1]1964; AVX2-NEXT: vpblendd {{.*#+}} ymm2 = ymm5[0,1,2,3],ymm2[4,5,6,7]1965; AVX2-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill1966; AVX2-NEXT: vmovdqa 528(%rdi), %xmm21967; AVX2-NEXT: vpalignr {{.*#+}} xmm5 = xmm8[8,9,10,11,12,13,14,15],xmm2[0,1,2,3,4,5,6,7]1968; AVX2-NEXT: vmovdqa 608(%rdi), %ymm81969; AVX2-NEXT: vpalignr {{.*#+}} ymm1 = ymm1[8,9,10,11,12,13,14,15],ymm8[0,1,2,3,4,5,6,7],ymm1[24,25,26,27,28,29,30,31],ymm8[16,17,18,19,20,21,22,23]1970; AVX2-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,1,2,1]1971; AVX2-NEXT: vpblendd {{.*#+}} ymm1 = ymm5[0,1,2,3],ymm1[4,5,6,7]1972; AVX2-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill1973; AVX2-NEXT: vmovdqa 48(%rdi), %xmm11974; AVX2-NEXT: vpalignr {{.*#+}} xmm5 = xmm10[8,9,10,11,12,13,14,15],xmm1[0,1,2,3,4,5,6,7]1975; AVX2-NEXT: vmovdqa 128(%rdi), %ymm71976; AVX2-NEXT: vpalignr {{.*#+}} ymm4 = ymm4[8,9,10,11,12,13,14,15],ymm7[0,1,2,3,4,5,6,7],ymm4[24,25,26,27,28,29,30,31],ymm7[16,17,18,19,20,21,22,23]1977; AVX2-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,1,2,1]1978; AVX2-NEXT: vpblendd {{.*#+}} ymm1 = ymm5[0,1,2,3],ymm4[4,5,6,7]1979; AVX2-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill1980; AVX2-NEXT: vmovdqa 368(%rdi), %xmm41981; AVX2-NEXT: vpalignr {{.*#+}} xmm4 = xmm0[8,9,10,11,12,13,14,15],xmm4[0,1,2,3,4,5,6,7]1982; AVX2-NEXT: vmovdqa 448(%rdi), %ymm61983; AVX2-NEXT: vpalignr {{.*#+}} ymm3 = ymm3[8,9,10,11,12,13,14,15],ymm6[0,1,2,3,4,5,6,7],ymm3[24,25,26,27,28,29,30,31],ymm6[16,17,18,19,20,21,22,23]1984; AVX2-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,1,2,1]1985; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm4[0,1,2,3],ymm3[4,5,6,7]1986; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill1987; AVX2-NEXT: vmovdqa 160(%rdi), %ymm31988; AVX2-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0,1],mem[2,3],ymm3[4,5],mem[6,7]1989; AVX2-NEXT: vmovdqa 256(%rdi), %xmm41990; AVX2-NEXT: vpblendd {{.*#+}} xmm4 = xmm4[0,1],mem[2,3]1991; AVX2-NEXT: vinserti128 $1, %xmm4, %ymm0, %ymm41992; AVX2-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm4[2,3]1993; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill1994; AVX2-NEXT: vmovdqa 480(%rdi), %ymm31995; AVX2-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0,1],mem[2,3],ymm3[4,5],mem[6,7]1996; AVX2-NEXT: vmovdqa 576(%rdi), %xmm41997; AVX2-NEXT: vpblendd {{.*#+}} xmm4 = xmm4[0,1],mem[2,3]1998; AVX2-NEXT: vinserti128 $1, %xmm4, %ymm0, %ymm41999; AVX2-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm4[2,3]2000; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2001; AVX2-NEXT: vmovdqa (%rdi), %ymm32002; AVX2-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0,1],mem[2,3],ymm3[4,5],mem[6,7]2003; AVX2-NEXT: vmovdqa 96(%rdi), %xmm42004; AVX2-NEXT: vpblendd {{.*#+}} xmm4 = xmm4[0,1],mem[2,3]2005; AVX2-NEXT: vinserti128 $1, %xmm4, %ymm0, %ymm42006; AVX2-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm4[2,3]2007; AVX2-NEXT: vmovdqu %ymm0, (%rsp) # 32-byte Spill2008; AVX2-NEXT: vmovdqa 320(%rdi), %ymm32009; AVX2-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0,1],mem[2,3],ymm3[4,5],mem[6,7]2010; AVX2-NEXT: vmovdqa 416(%rdi), %xmm42011; AVX2-NEXT: vpblendd {{.*#+}} xmm4 = xmm4[0,1],mem[2,3]2012; AVX2-NEXT: vinserti128 $1, %xmm4, %ymm0, %ymm42013; AVX2-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm4[2,3]2014; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2015; AVX2-NEXT: vpalignr {{.*#+}} ymm3 = ymm11[8,9,10,11,12,13,14,15],ymm15[0,1,2,3,4,5,6,7],ymm11[24,25,26,27,28,29,30,31],ymm15[16,17,18,19,20,21,22,23]2016; AVX2-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,1,0,3]2017; AVX2-NEXT: vmovdqa 224(%rdi), %xmm42018; AVX2-NEXT: vpalignr {{.*#+}} xmm5 = mem[8,9,10,11,12,13,14,15],xmm4[0,1,2,3,4,5,6,7]2019; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm5[0,1,2,3],ymm3[4,5,6,7]2020; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2021; AVX2-NEXT: vpalignr {{.*#+}} ymm3 = ymm12[8,9,10,11,12,13,14,15],ymm8[0,1,2,3,4,5,6,7],ymm12[24,25,26,27,28,29,30,31],ymm8[16,17,18,19,20,21,22,23]2022; AVX2-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,1,0,3]2023; AVX2-NEXT: vmovdqa 544(%rdi), %xmm22024; AVX2-NEXT: vpalignr {{.*#+}} xmm5 = mem[8,9,10,11,12,13,14,15],xmm2[0,1,2,3,4,5,6,7]2025; AVX2-NEXT: vpblendd {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm3[4,5,6,7]2026; AVX2-NEXT: vpalignr {{.*#+}} ymm3 = ymm14[8,9,10,11,12,13,14,15],ymm7[0,1,2,3,4,5,6,7],ymm14[24,25,26,27,28,29,30,31],ymm7[16,17,18,19,20,21,22,23]2027; AVX2-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,1,0,3]2028; AVX2-NEXT: vmovdqa 64(%rdi), %xmm12029; AVX2-NEXT: vpalignr {{.*#+}} xmm13 = mem[8,9,10,11,12,13,14,15],xmm1[0,1,2,3,4,5,6,7]2030; AVX2-NEXT: vpblendd {{.*#+}} ymm3 = ymm13[0,1,2,3],ymm3[4,5,6,7]2031; AVX2-NEXT: vpalignr {{.*#+}} ymm13 = ymm9[8,9,10,11,12,13,14,15],ymm6[0,1,2,3,4,5,6,7],ymm9[24,25,26,27,28,29,30,31],ymm6[16,17,18,19,20,21,22,23]2032; AVX2-NEXT: vpermq {{.*#+}} ymm13 = ymm13[0,1,0,3]2033; AVX2-NEXT: vmovdqa 384(%rdi), %xmm02034; AVX2-NEXT: vpalignr {{.*#+}} xmm10 = mem[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]2035; AVX2-NEXT: vpblendd {{.*#+}} ymm10 = ymm10[0,1,2,3],ymm13[4,5,6,7]2036; AVX2-NEXT: vpblendd {{.*#+}} ymm11 = ymm11[0,1,2,3,4,5],ymm15[6,7]2037; AVX2-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4, %xmm4 # 16-byte Folded Reload2038; AVX2-NEXT: # xmm4 = mem[0,1],xmm4[2,3]2039; AVX2-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm11[4,5,6,7]2040; AVX2-NEXT: vpblendd {{.*#+}} ymm8 = ymm12[0,1,2,3,4,5],ymm8[6,7]2041; AVX2-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm2 # 16-byte Folded Reload2042; AVX2-NEXT: # xmm2 = mem[0,1],xmm2[2,3]2043; AVX2-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm8[4,5,6,7]2044; AVX2-NEXT: vpblendd {{.*#+}} ymm7 = ymm14[0,1,2,3,4,5],ymm7[6,7]2045; AVX2-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload2046; AVX2-NEXT: # xmm1 = mem[0,1],xmm1[2,3]2047; AVX2-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm7[4,5,6,7]2048; AVX2-NEXT: vpblendd {{.*#+}} ymm6 = ymm9[0,1,2,3,4,5],ymm6[6,7]2049; AVX2-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload2050; AVX2-NEXT: # xmm0 = mem[0,1],xmm0[2,3]2051; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm6[4,5,6,7]2052; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload2053; AVX2-NEXT: vmovaps %ymm6, 64(%rsi)2054; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload2055; AVX2-NEXT: vmovaps %ymm6, (%rsi)2056; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload2057; AVX2-NEXT: vmovaps %ymm6, 96(%rsi)2058; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload2059; AVX2-NEXT: vmovaps %ymm6, 32(%rsi)2060; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload2061; AVX2-NEXT: vmovaps %ymm6, 64(%rdx)2062; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload2063; AVX2-NEXT: vmovaps %ymm6, (%rdx)2064; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload2065; AVX2-NEXT: vmovaps %ymm6, 96(%rdx)2066; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload2067; AVX2-NEXT: vmovaps %ymm6, 32(%rdx)2068; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload2069; AVX2-NEXT: vmovaps %ymm6, 64(%rcx)2070; AVX2-NEXT: vmovups (%rsp), %ymm6 # 32-byte Reload2071; AVX2-NEXT: vmovaps %ymm6, (%rcx)2072; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload2073; AVX2-NEXT: vmovaps %ymm6, 96(%rcx)2074; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload2075; AVX2-NEXT: vmovaps %ymm6, 32(%rcx)2076; AVX2-NEXT: vmovdqa %ymm10, 64(%r8)2077; AVX2-NEXT: vmovdqa %ymm3, (%r8)2078; AVX2-NEXT: vmovdqa %ymm5, 96(%r8)2079; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload2080; AVX2-NEXT: vmovaps %ymm3, 32(%r8)2081; AVX2-NEXT: vmovdqa %ymm0, 64(%r9)2082; AVX2-NEXT: vmovdqa %ymm1, (%r9)2083; AVX2-NEXT: vmovdqa %ymm2, 96(%r9)2084; AVX2-NEXT: vmovdqa %ymm4, 32(%r9)2085; AVX2-NEXT: addq $360, %rsp # imm = 0x1682086; AVX2-NEXT: vzeroupper2087; AVX2-NEXT: retq2088;2089; AVX2-FP-LABEL: load_i64_stride5_vf16:2090; AVX2-FP: # %bb.0:2091; AVX2-FP-NEXT: subq $360, %rsp # imm = 0x1682092; AVX2-FP-NEXT: vmovdqa 96(%rdi), %ymm142093; AVX2-FP-NEXT: vmovdqa 64(%rdi), %ymm42094; AVX2-FP-NEXT: vmovdqa 576(%rdi), %ymm122095; AVX2-FP-NEXT: vmovdqa 544(%rdi), %ymm12096; AVX2-FP-NEXT: vmovdqa 256(%rdi), %ymm112097; AVX2-FP-NEXT: vmovdqa 224(%rdi), %ymm22098; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm3 = ymm2[0,1,2,3,4,5],ymm11[6,7]2099; AVX2-FP-NEXT: vmovdqa 192(%rdi), %xmm02100; AVX2-FP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2101; AVX2-FP-NEXT: vmovdqa 160(%rdi), %xmm52102; AVX2-FP-NEXT: vpblendd {{.*#+}} xmm8 = xmm5[0,1],xmm0[2,3]2103; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm8[0,1,2,3],ymm3[4,5,6,7]2104; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2105; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm3 = ymm1[0,1,2,3,4,5],ymm12[6,7]2106; AVX2-FP-NEXT: vmovdqa 512(%rdi), %xmm02107; AVX2-FP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2108; AVX2-FP-NEXT: vmovdqa 480(%rdi), %xmm82109; AVX2-FP-NEXT: vpblendd {{.*#+}} xmm9 = xmm8[0,1],xmm0[2,3]2110; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm9[0,1,2,3],ymm3[4,5,6,7]2111; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2112; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm3 = ymm4[0,1,2,3,4,5],ymm14[6,7]2113; AVX2-FP-NEXT: vmovdqa (%rdi), %xmm102114; AVX2-FP-NEXT: vmovdqa 32(%rdi), %xmm02115; AVX2-FP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2116; AVX2-FP-NEXT: vpblendd {{.*#+}} xmm9 = xmm10[0,1],xmm0[2,3]2117; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm9[0,1,2,3],ymm3[4,5,6,7]2118; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2119; AVX2-FP-NEXT: vmovdqa 416(%rdi), %ymm92120; AVX2-FP-NEXT: vmovdqa 384(%rdi), %ymm32121; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm13 = ymm3[0,1,2,3,4,5],ymm9[6,7]2122; AVX2-FP-NEXT: vmovdqa 352(%rdi), %xmm62123; AVX2-FP-NEXT: vmovdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2124; AVX2-FP-NEXT: vmovdqa 320(%rdi), %xmm02125; AVX2-FP-NEXT: vpblendd {{.*#+}} xmm15 = xmm0[0,1],xmm6[2,3]2126; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm6 = ymm15[0,1,2,3],ymm13[4,5,6,7]2127; AVX2-FP-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2128; AVX2-FP-NEXT: vmovdqa 208(%rdi), %xmm132129; AVX2-FP-NEXT: vpalignr {{.*#+}} xmm5 = xmm5[8,9,10,11,12,13,14,15],xmm13[0,1,2,3,4,5,6,7]2130; AVX2-FP-NEXT: vmovdqa 288(%rdi), %ymm152131; AVX2-FP-NEXT: vpalignr {{.*#+}} ymm2 = ymm2[8,9,10,11,12,13,14,15],ymm15[0,1,2,3,4,5,6,7],ymm2[24,25,26,27,28,29,30,31],ymm15[16,17,18,19,20,21,22,23]2132; AVX2-FP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,1,2,1]2133; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm2 = ymm5[0,1,2,3],ymm2[4,5,6,7]2134; AVX2-FP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2135; AVX2-FP-NEXT: vmovdqa 528(%rdi), %xmm22136; AVX2-FP-NEXT: vpalignr {{.*#+}} xmm5 = xmm8[8,9,10,11,12,13,14,15],xmm2[0,1,2,3,4,5,6,7]2137; AVX2-FP-NEXT: vmovdqa 608(%rdi), %ymm82138; AVX2-FP-NEXT: vpalignr {{.*#+}} ymm1 = ymm1[8,9,10,11,12,13,14,15],ymm8[0,1,2,3,4,5,6,7],ymm1[24,25,26,27,28,29,30,31],ymm8[16,17,18,19,20,21,22,23]2139; AVX2-FP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,1,2,1]2140; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm1 = ymm5[0,1,2,3],ymm1[4,5,6,7]2141; AVX2-FP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2142; AVX2-FP-NEXT: vmovdqa 48(%rdi), %xmm12143; AVX2-FP-NEXT: vpalignr {{.*#+}} xmm5 = xmm10[8,9,10,11,12,13,14,15],xmm1[0,1,2,3,4,5,6,7]2144; AVX2-FP-NEXT: vmovdqa 128(%rdi), %ymm72145; AVX2-FP-NEXT: vpalignr {{.*#+}} ymm4 = ymm4[8,9,10,11,12,13,14,15],ymm7[0,1,2,3,4,5,6,7],ymm4[24,25,26,27,28,29,30,31],ymm7[16,17,18,19,20,21,22,23]2146; AVX2-FP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,1,2,1]2147; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm1 = ymm5[0,1,2,3],ymm4[4,5,6,7]2148; AVX2-FP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2149; AVX2-FP-NEXT: vmovdqa 368(%rdi), %xmm42150; AVX2-FP-NEXT: vpalignr {{.*#+}} xmm4 = xmm0[8,9,10,11,12,13,14,15],xmm4[0,1,2,3,4,5,6,7]2151; AVX2-FP-NEXT: vmovdqa 448(%rdi), %ymm62152; AVX2-FP-NEXT: vpalignr {{.*#+}} ymm3 = ymm3[8,9,10,11,12,13,14,15],ymm6[0,1,2,3,4,5,6,7],ymm3[24,25,26,27,28,29,30,31],ymm6[16,17,18,19,20,21,22,23]2153; AVX2-FP-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,1,2,1]2154; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm4[0,1,2,3],ymm3[4,5,6,7]2155; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2156; AVX2-FP-NEXT: vmovdqa 160(%rdi), %ymm32157; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0,1],mem[2,3],ymm3[4,5],mem[6,7]2158; AVX2-FP-NEXT: vmovdqa 256(%rdi), %xmm42159; AVX2-FP-NEXT: vpblendd {{.*#+}} xmm4 = xmm4[0,1],mem[2,3]2160; AVX2-FP-NEXT: vinserti128 $1, %xmm4, %ymm0, %ymm42161; AVX2-FP-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm4[2,3]2162; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2163; AVX2-FP-NEXT: vmovdqa 480(%rdi), %ymm32164; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0,1],mem[2,3],ymm3[4,5],mem[6,7]2165; AVX2-FP-NEXT: vmovdqa 576(%rdi), %xmm42166; AVX2-FP-NEXT: vpblendd {{.*#+}} xmm4 = xmm4[0,1],mem[2,3]2167; AVX2-FP-NEXT: vinserti128 $1, %xmm4, %ymm0, %ymm42168; AVX2-FP-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm4[2,3]2169; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2170; AVX2-FP-NEXT: vmovdqa (%rdi), %ymm32171; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0,1],mem[2,3],ymm3[4,5],mem[6,7]2172; AVX2-FP-NEXT: vmovdqa 96(%rdi), %xmm42173; AVX2-FP-NEXT: vpblendd {{.*#+}} xmm4 = xmm4[0,1],mem[2,3]2174; AVX2-FP-NEXT: vinserti128 $1, %xmm4, %ymm0, %ymm42175; AVX2-FP-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm4[2,3]2176; AVX2-FP-NEXT: vmovdqu %ymm0, (%rsp) # 32-byte Spill2177; AVX2-FP-NEXT: vmovdqa 320(%rdi), %ymm32178; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0,1],mem[2,3],ymm3[4,5],mem[6,7]2179; AVX2-FP-NEXT: vmovdqa 416(%rdi), %xmm42180; AVX2-FP-NEXT: vpblendd {{.*#+}} xmm4 = xmm4[0,1],mem[2,3]2181; AVX2-FP-NEXT: vinserti128 $1, %xmm4, %ymm0, %ymm42182; AVX2-FP-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm4[2,3]2183; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2184; AVX2-FP-NEXT: vpalignr {{.*#+}} ymm3 = ymm11[8,9,10,11,12,13,14,15],ymm15[0,1,2,3,4,5,6,7],ymm11[24,25,26,27,28,29,30,31],ymm15[16,17,18,19,20,21,22,23]2185; AVX2-FP-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,1,0,3]2186; AVX2-FP-NEXT: vmovdqa 224(%rdi), %xmm42187; AVX2-FP-NEXT: vpalignr {{.*#+}} xmm5 = mem[8,9,10,11,12,13,14,15],xmm4[0,1,2,3,4,5,6,7]2188; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm5[0,1,2,3],ymm3[4,5,6,7]2189; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2190; AVX2-FP-NEXT: vpalignr {{.*#+}} ymm3 = ymm12[8,9,10,11,12,13,14,15],ymm8[0,1,2,3,4,5,6,7],ymm12[24,25,26,27,28,29,30,31],ymm8[16,17,18,19,20,21,22,23]2191; AVX2-FP-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,1,0,3]2192; AVX2-FP-NEXT: vmovdqa 544(%rdi), %xmm22193; AVX2-FP-NEXT: vpalignr {{.*#+}} xmm5 = mem[8,9,10,11,12,13,14,15],xmm2[0,1,2,3,4,5,6,7]2194; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm3[4,5,6,7]2195; AVX2-FP-NEXT: vpalignr {{.*#+}} ymm3 = ymm14[8,9,10,11,12,13,14,15],ymm7[0,1,2,3,4,5,6,7],ymm14[24,25,26,27,28,29,30,31],ymm7[16,17,18,19,20,21,22,23]2196; AVX2-FP-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,1,0,3]2197; AVX2-FP-NEXT: vmovdqa 64(%rdi), %xmm12198; AVX2-FP-NEXT: vpalignr {{.*#+}} xmm13 = mem[8,9,10,11,12,13,14,15],xmm1[0,1,2,3,4,5,6,7]2199; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm3 = ymm13[0,1,2,3],ymm3[4,5,6,7]2200; AVX2-FP-NEXT: vpalignr {{.*#+}} ymm13 = ymm9[8,9,10,11,12,13,14,15],ymm6[0,1,2,3,4,5,6,7],ymm9[24,25,26,27,28,29,30,31],ymm6[16,17,18,19,20,21,22,23]2201; AVX2-FP-NEXT: vpermq {{.*#+}} ymm13 = ymm13[0,1,0,3]2202; AVX2-FP-NEXT: vmovdqa 384(%rdi), %xmm02203; AVX2-FP-NEXT: vpalignr {{.*#+}} xmm10 = mem[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]2204; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm10 = ymm10[0,1,2,3],ymm13[4,5,6,7]2205; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm11 = ymm11[0,1,2,3,4,5],ymm15[6,7]2206; AVX2-FP-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4, %xmm4 # 16-byte Folded Reload2207; AVX2-FP-NEXT: # xmm4 = mem[0,1],xmm4[2,3]2208; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm11[4,5,6,7]2209; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm8 = ymm12[0,1,2,3,4,5],ymm8[6,7]2210; AVX2-FP-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm2 # 16-byte Folded Reload2211; AVX2-FP-NEXT: # xmm2 = mem[0,1],xmm2[2,3]2212; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm8[4,5,6,7]2213; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm7 = ymm14[0,1,2,3,4,5],ymm7[6,7]2214; AVX2-FP-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload2215; AVX2-FP-NEXT: # xmm1 = mem[0,1],xmm1[2,3]2216; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm7[4,5,6,7]2217; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm6 = ymm9[0,1,2,3,4,5],ymm6[6,7]2218; AVX2-FP-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload2219; AVX2-FP-NEXT: # xmm0 = mem[0,1],xmm0[2,3]2220; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm6[4,5,6,7]2221; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload2222; AVX2-FP-NEXT: vmovaps %ymm6, 64(%rsi)2223; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload2224; AVX2-FP-NEXT: vmovaps %ymm6, (%rsi)2225; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload2226; AVX2-FP-NEXT: vmovaps %ymm6, 96(%rsi)2227; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload2228; AVX2-FP-NEXT: vmovaps %ymm6, 32(%rsi)2229; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload2230; AVX2-FP-NEXT: vmovaps %ymm6, 64(%rdx)2231; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload2232; AVX2-FP-NEXT: vmovaps %ymm6, (%rdx)2233; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload2234; AVX2-FP-NEXT: vmovaps %ymm6, 96(%rdx)2235; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload2236; AVX2-FP-NEXT: vmovaps %ymm6, 32(%rdx)2237; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload2238; AVX2-FP-NEXT: vmovaps %ymm6, 64(%rcx)2239; AVX2-FP-NEXT: vmovups (%rsp), %ymm6 # 32-byte Reload2240; AVX2-FP-NEXT: vmovaps %ymm6, (%rcx)2241; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload2242; AVX2-FP-NEXT: vmovaps %ymm6, 96(%rcx)2243; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload2244; AVX2-FP-NEXT: vmovaps %ymm6, 32(%rcx)2245; AVX2-FP-NEXT: vmovdqa %ymm10, 64(%r8)2246; AVX2-FP-NEXT: vmovdqa %ymm3, (%r8)2247; AVX2-FP-NEXT: vmovdqa %ymm5, 96(%r8)2248; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload2249; AVX2-FP-NEXT: vmovaps %ymm3, 32(%r8)2250; AVX2-FP-NEXT: vmovdqa %ymm0, 64(%r9)2251; AVX2-FP-NEXT: vmovdqa %ymm1, (%r9)2252; AVX2-FP-NEXT: vmovdqa %ymm2, 96(%r9)2253; AVX2-FP-NEXT: vmovdqa %ymm4, 32(%r9)2254; AVX2-FP-NEXT: addq $360, %rsp # imm = 0x1682255; AVX2-FP-NEXT: vzeroupper2256; AVX2-FP-NEXT: retq2257;2258; AVX2-FCP-LABEL: load_i64_stride5_vf16:2259; AVX2-FCP: # %bb.0:2260; AVX2-FCP-NEXT: subq $360, %rsp # imm = 0x1682261; AVX2-FCP-NEXT: vmovdqa 96(%rdi), %ymm142262; AVX2-FCP-NEXT: vmovdqa 64(%rdi), %ymm42263; AVX2-FCP-NEXT: vmovdqa 576(%rdi), %ymm122264; AVX2-FCP-NEXT: vmovdqa 544(%rdi), %ymm12265; AVX2-FCP-NEXT: vmovdqa 256(%rdi), %ymm112266; AVX2-FCP-NEXT: vmovdqa 224(%rdi), %ymm22267; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm3 = ymm2[0,1,2,3,4,5],ymm11[6,7]2268; AVX2-FCP-NEXT: vmovdqa 192(%rdi), %xmm02269; AVX2-FCP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2270; AVX2-FCP-NEXT: vmovdqa 160(%rdi), %xmm52271; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm8 = xmm5[0,1],xmm0[2,3]2272; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm8[0,1,2,3],ymm3[4,5,6,7]2273; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2274; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm3 = ymm1[0,1,2,3,4,5],ymm12[6,7]2275; AVX2-FCP-NEXT: vmovdqa 512(%rdi), %xmm02276; AVX2-FCP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2277; AVX2-FCP-NEXT: vmovdqa 480(%rdi), %xmm82278; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm9 = xmm8[0,1],xmm0[2,3]2279; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm9[0,1,2,3],ymm3[4,5,6,7]2280; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2281; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm3 = ymm4[0,1,2,3,4,5],ymm14[6,7]2282; AVX2-FCP-NEXT: vmovdqa (%rdi), %xmm102283; AVX2-FCP-NEXT: vmovdqa 32(%rdi), %xmm02284; AVX2-FCP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2285; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm9 = xmm10[0,1],xmm0[2,3]2286; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm9[0,1,2,3],ymm3[4,5,6,7]2287; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2288; AVX2-FCP-NEXT: vmovdqa 416(%rdi), %ymm92289; AVX2-FCP-NEXT: vmovdqa 384(%rdi), %ymm32290; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm13 = ymm3[0,1,2,3,4,5],ymm9[6,7]2291; AVX2-FCP-NEXT: vmovdqa 352(%rdi), %xmm62292; AVX2-FCP-NEXT: vmovdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill2293; AVX2-FCP-NEXT: vmovdqa 320(%rdi), %xmm02294; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm15 = xmm0[0,1],xmm6[2,3]2295; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm15[0,1,2,3],ymm13[4,5,6,7]2296; AVX2-FCP-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2297; AVX2-FCP-NEXT: vmovdqa 208(%rdi), %xmm132298; AVX2-FCP-NEXT: vpalignr {{.*#+}} xmm5 = xmm5[8,9,10,11,12,13,14,15],xmm13[0,1,2,3,4,5,6,7]2299; AVX2-FCP-NEXT: vmovdqa 288(%rdi), %ymm152300; AVX2-FCP-NEXT: vpalignr {{.*#+}} ymm2 = ymm2[8,9,10,11,12,13,14,15],ymm15[0,1,2,3,4,5,6,7],ymm2[24,25,26,27,28,29,30,31],ymm15[16,17,18,19,20,21,22,23]2301; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,1,2,1]2302; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm5[0,1,2,3],ymm2[4,5,6,7]2303; AVX2-FCP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2304; AVX2-FCP-NEXT: vmovdqa 528(%rdi), %xmm22305; AVX2-FCP-NEXT: vpalignr {{.*#+}} xmm5 = xmm8[8,9,10,11,12,13,14,15],xmm2[0,1,2,3,4,5,6,7]2306; AVX2-FCP-NEXT: vmovdqa 608(%rdi), %ymm82307; AVX2-FCP-NEXT: vpalignr {{.*#+}} ymm1 = ymm1[8,9,10,11,12,13,14,15],ymm8[0,1,2,3,4,5,6,7],ymm1[24,25,26,27,28,29,30,31],ymm8[16,17,18,19,20,21,22,23]2308; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,1,2,1]2309; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm1 = ymm5[0,1,2,3],ymm1[4,5,6,7]2310; AVX2-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2311; AVX2-FCP-NEXT: vmovdqa 48(%rdi), %xmm12312; AVX2-FCP-NEXT: vpalignr {{.*#+}} xmm5 = xmm10[8,9,10,11,12,13,14,15],xmm1[0,1,2,3,4,5,6,7]2313; AVX2-FCP-NEXT: vmovdqa 128(%rdi), %ymm72314; AVX2-FCP-NEXT: vpalignr {{.*#+}} ymm4 = ymm4[8,9,10,11,12,13,14,15],ymm7[0,1,2,3,4,5,6,7],ymm4[24,25,26,27,28,29,30,31],ymm7[16,17,18,19,20,21,22,23]2315; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,1,2,1]2316; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm1 = ymm5[0,1,2,3],ymm4[4,5,6,7]2317; AVX2-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2318; AVX2-FCP-NEXT: vmovdqa 368(%rdi), %xmm42319; AVX2-FCP-NEXT: vpalignr {{.*#+}} xmm4 = xmm0[8,9,10,11,12,13,14,15],xmm4[0,1,2,3,4,5,6,7]2320; AVX2-FCP-NEXT: vmovdqa 448(%rdi), %ymm62321; AVX2-FCP-NEXT: vpalignr {{.*#+}} ymm3 = ymm3[8,9,10,11,12,13,14,15],ymm6[0,1,2,3,4,5,6,7],ymm3[24,25,26,27,28,29,30,31],ymm6[16,17,18,19,20,21,22,23]2322; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,1,2,1]2323; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm4[0,1,2,3],ymm3[4,5,6,7]2324; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2325; AVX2-FCP-NEXT: vmovdqa 160(%rdi), %ymm32326; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0,1],mem[2,3],ymm3[4,5],mem[6,7]2327; AVX2-FCP-NEXT: vmovdqa 256(%rdi), %xmm42328; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm4 = xmm4[0,1],mem[2,3]2329; AVX2-FCP-NEXT: vinserti128 $1, %xmm4, %ymm0, %ymm42330; AVX2-FCP-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm4[2,3]2331; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2332; AVX2-FCP-NEXT: vmovdqa 480(%rdi), %ymm32333; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0,1],mem[2,3],ymm3[4,5],mem[6,7]2334; AVX2-FCP-NEXT: vmovdqa 576(%rdi), %xmm42335; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm4 = xmm4[0,1],mem[2,3]2336; AVX2-FCP-NEXT: vinserti128 $1, %xmm4, %ymm0, %ymm42337; AVX2-FCP-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm4[2,3]2338; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2339; AVX2-FCP-NEXT: vmovdqa (%rdi), %ymm32340; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0,1],mem[2,3],ymm3[4,5],mem[6,7]2341; AVX2-FCP-NEXT: vmovdqa 96(%rdi), %xmm42342; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm4 = xmm4[0,1],mem[2,3]2343; AVX2-FCP-NEXT: vinserti128 $1, %xmm4, %ymm0, %ymm42344; AVX2-FCP-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm4[2,3]2345; AVX2-FCP-NEXT: vmovdqu %ymm0, (%rsp) # 32-byte Spill2346; AVX2-FCP-NEXT: vmovdqa 320(%rdi), %ymm32347; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0,1],mem[2,3],ymm3[4,5],mem[6,7]2348; AVX2-FCP-NEXT: vmovdqa 416(%rdi), %xmm42349; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm4 = xmm4[0,1],mem[2,3]2350; AVX2-FCP-NEXT: vinserti128 $1, %xmm4, %ymm0, %ymm42351; AVX2-FCP-NEXT: vperm2i128 {{.*#+}} ymm0 = ymm3[2,3],ymm4[2,3]2352; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2353; AVX2-FCP-NEXT: vpalignr {{.*#+}} ymm3 = ymm11[8,9,10,11,12,13,14,15],ymm15[0,1,2,3,4,5,6,7],ymm11[24,25,26,27,28,29,30,31],ymm15[16,17,18,19,20,21,22,23]2354; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,1,0,3]2355; AVX2-FCP-NEXT: vmovdqa 224(%rdi), %xmm42356; AVX2-FCP-NEXT: vpalignr {{.*#+}} xmm5 = mem[8,9,10,11,12,13,14,15],xmm4[0,1,2,3,4,5,6,7]2357; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm5[0,1,2,3],ymm3[4,5,6,7]2358; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill2359; AVX2-FCP-NEXT: vpalignr {{.*#+}} ymm3 = ymm12[8,9,10,11,12,13,14,15],ymm8[0,1,2,3,4,5,6,7],ymm12[24,25,26,27,28,29,30,31],ymm8[16,17,18,19,20,21,22,23]2360; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,1,0,3]2361; AVX2-FCP-NEXT: vmovdqa 544(%rdi), %xmm22362; AVX2-FCP-NEXT: vpalignr {{.*#+}} xmm5 = mem[8,9,10,11,12,13,14,15],xmm2[0,1,2,3,4,5,6,7]2363; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm3[4,5,6,7]2364; AVX2-FCP-NEXT: vpalignr {{.*#+}} ymm3 = ymm14[8,9,10,11,12,13,14,15],ymm7[0,1,2,3,4,5,6,7],ymm14[24,25,26,27,28,29,30,31],ymm7[16,17,18,19,20,21,22,23]2365; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,1,0,3]2366; AVX2-FCP-NEXT: vmovdqa 64(%rdi), %xmm12367; AVX2-FCP-NEXT: vpalignr {{.*#+}} xmm13 = mem[8,9,10,11,12,13,14,15],xmm1[0,1,2,3,4,5,6,7]2368; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm3 = ymm13[0,1,2,3],ymm3[4,5,6,7]2369; AVX2-FCP-NEXT: vpalignr {{.*#+}} ymm13 = ymm9[8,9,10,11,12,13,14,15],ymm6[0,1,2,3,4,5,6,7],ymm9[24,25,26,27,28,29,30,31],ymm6[16,17,18,19,20,21,22,23]2370; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm13 = ymm13[0,1,0,3]2371; AVX2-FCP-NEXT: vmovdqa 384(%rdi), %xmm02372; AVX2-FCP-NEXT: vpalignr {{.*#+}} xmm10 = mem[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]2373; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm10 = ymm10[0,1,2,3],ymm13[4,5,6,7]2374; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm11 = ymm11[0,1,2,3,4,5],ymm15[6,7]2375; AVX2-FCP-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4, %xmm4 # 16-byte Folded Reload2376; AVX2-FCP-NEXT: # xmm4 = mem[0,1],xmm4[2,3]2377; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm11[4,5,6,7]2378; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm8 = ymm12[0,1,2,3,4,5],ymm8[6,7]2379; AVX2-FCP-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm2 # 16-byte Folded Reload2380; AVX2-FCP-NEXT: # xmm2 = mem[0,1],xmm2[2,3]2381; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1,2,3],ymm8[4,5,6,7]2382; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm7 = ymm14[0,1,2,3,4,5],ymm7[6,7]2383; AVX2-FCP-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload2384; AVX2-FCP-NEXT: # xmm1 = mem[0,1],xmm1[2,3]2385; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm1 = ymm1[0,1,2,3],ymm7[4,5,6,7]2386; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm9[0,1,2,3,4,5],ymm6[6,7]2387; AVX2-FCP-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload2388; AVX2-FCP-NEXT: # xmm0 = mem[0,1],xmm0[2,3]2389; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm6[4,5,6,7]2390; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload2391; AVX2-FCP-NEXT: vmovaps %ymm6, 64(%rsi)2392; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload2393; AVX2-FCP-NEXT: vmovaps %ymm6, (%rsi)2394; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload2395; AVX2-FCP-NEXT: vmovaps %ymm6, 96(%rsi)2396; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload2397; AVX2-FCP-NEXT: vmovaps %ymm6, 32(%rsi)2398; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload2399; AVX2-FCP-NEXT: vmovaps %ymm6, 64(%rdx)2400; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload2401; AVX2-FCP-NEXT: vmovaps %ymm6, (%rdx)2402; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload2403; AVX2-FCP-NEXT: vmovaps %ymm6, 96(%rdx)2404; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload2405; AVX2-FCP-NEXT: vmovaps %ymm6, 32(%rdx)2406; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload2407; AVX2-FCP-NEXT: vmovaps %ymm6, 64(%rcx)2408; AVX2-FCP-NEXT: vmovups (%rsp), %ymm6 # 32-byte Reload2409; AVX2-FCP-NEXT: vmovaps %ymm6, (%rcx)2410; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload2411; AVX2-FCP-NEXT: vmovaps %ymm6, 96(%rcx)2412; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload2413; AVX2-FCP-NEXT: vmovaps %ymm6, 32(%rcx)2414; AVX2-FCP-NEXT: vmovdqa %ymm10, 64(%r8)2415; AVX2-FCP-NEXT: vmovdqa %ymm3, (%r8)2416; AVX2-FCP-NEXT: vmovdqa %ymm5, 96(%r8)2417; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload2418; AVX2-FCP-NEXT: vmovaps %ymm3, 32(%r8)2419; AVX2-FCP-NEXT: vmovdqa %ymm0, 64(%r9)2420; AVX2-FCP-NEXT: vmovdqa %ymm1, (%r9)2421; AVX2-FCP-NEXT: vmovdqa %ymm2, 96(%r9)2422; AVX2-FCP-NEXT: vmovdqa %ymm4, 32(%r9)2423; AVX2-FCP-NEXT: addq $360, %rsp # imm = 0x1682424; AVX2-FCP-NEXT: vzeroupper2425; AVX2-FCP-NEXT: retq2426;2427; AVX512-LABEL: load_i64_stride5_vf16:2428; AVX512: # %bb.0:2429; AVX512-NEXT: vmovdqa64 576(%rdi), %zmm22430; AVX512-NEXT: vmovdqa64 384(%rdi), %zmm02431; AVX512-NEXT: vmovdqa64 320(%rdi), %zmm32432; AVX512-NEXT: vmovdqa64 448(%rdi), %zmm52433; AVX512-NEXT: vmovdqa64 512(%rdi), %zmm42434; AVX512-NEXT: vmovdqa64 256(%rdi), %zmm62435; AVX512-NEXT: vmovdqa64 (%rdi), %zmm92436; AVX512-NEXT: vmovdqa64 64(%rdi), %zmm12437; AVX512-NEXT: vmovdqa64 128(%rdi), %zmm112438; AVX512-NEXT: vmovdqa64 192(%rdi), %zmm102439; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [12,1,6,0,12,1,6,0]2440; AVX512-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]2441; AVX512-NEXT: vmovdqa64 %zmm10, %zmm82442; AVX512-NEXT: vpermt2q %zmm11, %zmm7, %zmm82443; AVX512-NEXT: vpmovsxbq {{.*#+}} ymm12 = [0,5,10,15]2444; AVX512-NEXT: vmovdqa64 %zmm9, %zmm132445; AVX512-NEXT: vpermt2q %zmm1, %zmm12, %zmm132446; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm8 = zmm13[0,1,2,3],zmm8[4,5,6,7]2447; AVX512-NEXT: vpmovsxbq {{.*#+}} zmm13 = [0,1,2,3,4,5,6,11]2448; AVX512-NEXT: vpermt2q %zmm6, %zmm13, %zmm82449; AVX512-NEXT: vmovdqa64 %zmm4, %zmm142450; AVX512-NEXT: vpermt2q %zmm5, %zmm7, %zmm142451; AVX512-NEXT: vpermi2q %zmm0, %zmm3, %zmm122452; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm12 = zmm12[0,1,2,3],zmm14[4,5,6,7]2453; AVX512-NEXT: vpermt2q %zmm2, %zmm13, %zmm122454; AVX512-NEXT: vpmovsxbq {{.*#+}} ymm15 = [1,6,11,0]2455; AVX512-NEXT: vmovdqa64 %zmm9, %zmm162456; AVX512-NEXT: vpermt2q %zmm1, %zmm15, %zmm162457; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm14 = [5,10,15,0,5,10,15,0]2458; AVX512-NEXT: # zmm14 = mem[0,1,2,3,0,1,2,3]2459; AVX512-NEXT: vmovdqa64 %zmm11, %zmm132460; AVX512-NEXT: vpermt2q %zmm10, %zmm14, %zmm132461; AVX512-NEXT: movb $7, %al2462; AVX512-NEXT: kmovw %eax, %k12463; AVX512-NEXT: vmovdqa64 %zmm16, %zmm13 {%k1}2464; AVX512-NEXT: vpmovsxbq {{.*#+}} zmm16 = [0,1,2,3,4,5,6,12]2465; AVX512-NEXT: vpermt2q %zmm6, %zmm16, %zmm132466; AVX512-NEXT: vpermi2q %zmm4, %zmm5, %zmm142467; AVX512-NEXT: vpermi2q %zmm0, %zmm3, %zmm152468; AVX512-NEXT: vmovdqa64 %zmm15, %zmm14 {%k1}2469; AVX512-NEXT: vpermt2q %zmm2, %zmm16, %zmm142470; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm15 = [6,11,0,1,6,11,0,1]2471; AVX512-NEXT: # zmm15 = mem[0,1,2,3,0,1,2,3]2472; AVX512-NEXT: vmovdqa64 %zmm11, %zmm162473; AVX512-NEXT: vpermt2q %zmm10, %zmm15, %zmm162474; AVX512-NEXT: vpmovsxbq {{.*#+}} ymm17 = [2,7,12,0]2475; AVX512-NEXT: vmovdqa64 %zmm9, %zmm182476; AVX512-NEXT: vpermt2q %zmm1, %zmm17, %zmm182477; AVX512-NEXT: movb $56, %al2478; AVX512-NEXT: kmovw %eax, %k12479; AVX512-NEXT: vmovdqa64 %zmm16, %zmm18 {%k1}2480; AVX512-NEXT: vpmovsxbq {{.*#+}} zmm16 = [0,1,2,3,4,5,8,13]2481; AVX512-NEXT: vpermt2q %zmm6, %zmm16, %zmm182482; AVX512-NEXT: vpermi2q %zmm4, %zmm5, %zmm152483; AVX512-NEXT: vpermi2q %zmm0, %zmm3, %zmm172484; AVX512-NEXT: vmovdqa64 %zmm15, %zmm17 {%k1}2485; AVX512-NEXT: vpermt2q %zmm2, %zmm16, %zmm172486; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm15 = [7,12,0,2,7,12,0,2]2487; AVX512-NEXT: # zmm15 = mem[0,1,2,3,0,1,2,3]2488; AVX512-NEXT: vmovdqa64 %zmm11, %zmm162489; AVX512-NEXT: vpermt2q %zmm10, %zmm15, %zmm162490; AVX512-NEXT: vpmovsxbq {{.*#+}} ymm19 = [11,0,5,0]2491; AVX512-NEXT: vmovdqa64 %zmm1, %zmm202492; AVX512-NEXT: vpermt2q %zmm9, %zmm19, %zmm202493; AVX512-NEXT: vmovdqa64 %zmm16, %zmm20 {%k1}2494; AVX512-NEXT: vpmovsxbq {{.*#+}} zmm16 = [0,1,2,3,4,5,9,14]2495; AVX512-NEXT: vpermt2q %zmm6, %zmm16, %zmm202496; AVX512-NEXT: vpermi2q %zmm4, %zmm5, %zmm152497; AVX512-NEXT: vpermi2q %zmm3, %zmm0, %zmm192498; AVX512-NEXT: vmovdqa64 %zmm15, %zmm19 {%k1}2499; AVX512-NEXT: vpermt2q %zmm2, %zmm16, %zmm192500; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm15 = [0,5,0,11,0,5,0,11]2501; AVX512-NEXT: # zmm15 = mem[0,1,2,3,0,1,2,3]2502; AVX512-NEXT: vpermt2q %zmm11, %zmm15, %zmm102503; AVX512-NEXT: vpermt2q %zmm9, %zmm7, %zmm12504; AVX512-NEXT: vmovdqa64 %zmm10, %zmm1 {%k1}2505; AVX512-NEXT: vpmovsxbq {{.*#+}} zmm9 = [0,1,2,3,4,5,10,15]2506; AVX512-NEXT: vpermt2q %zmm6, %zmm9, %zmm12507; AVX512-NEXT: vpermt2q %zmm5, %zmm15, %zmm42508; AVX512-NEXT: vpermt2q %zmm3, %zmm7, %zmm02509; AVX512-NEXT: vmovdqa64 %zmm4, %zmm0 {%k1}2510; AVX512-NEXT: vpermt2q %zmm2, %zmm9, %zmm02511; AVX512-NEXT: vmovdqa64 %zmm12, 64(%rsi)2512; AVX512-NEXT: vmovdqa64 %zmm8, (%rsi)2513; AVX512-NEXT: vmovdqa64 %zmm14, 64(%rdx)2514; AVX512-NEXT: vmovdqa64 %zmm13, (%rdx)2515; AVX512-NEXT: vmovdqa64 %zmm17, 64(%rcx)2516; AVX512-NEXT: vmovdqa64 %zmm18, (%rcx)2517; AVX512-NEXT: vmovdqa64 %zmm19, 64(%r8)2518; AVX512-NEXT: vmovdqa64 %zmm20, (%r8)2519; AVX512-NEXT: vmovdqa64 %zmm0, 64(%r9)2520; AVX512-NEXT: vmovdqa64 %zmm1, (%r9)2521; AVX512-NEXT: vzeroupper2522; AVX512-NEXT: retq2523;2524; AVX512-FCP-LABEL: load_i64_stride5_vf16:2525; AVX512-FCP: # %bb.0:2526; AVX512-FCP-NEXT: vmovdqa64 576(%rdi), %zmm22527; AVX512-FCP-NEXT: vmovdqa64 384(%rdi), %zmm02528; AVX512-FCP-NEXT: vmovdqa64 320(%rdi), %zmm32529; AVX512-FCP-NEXT: vmovdqa64 448(%rdi), %zmm52530; AVX512-FCP-NEXT: vmovdqa64 512(%rdi), %zmm42531; AVX512-FCP-NEXT: vmovdqa64 256(%rdi), %zmm62532; AVX512-FCP-NEXT: vmovdqa64 (%rdi), %zmm92533; AVX512-FCP-NEXT: vmovdqa64 64(%rdi), %zmm12534; AVX512-FCP-NEXT: vmovdqa64 128(%rdi), %zmm112535; AVX512-FCP-NEXT: vmovdqa64 192(%rdi), %zmm102536; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [12,1,6,0,12,1,6,0]2537; AVX512-FCP-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]2538; AVX512-FCP-NEXT: vmovdqa64 %zmm10, %zmm82539; AVX512-FCP-NEXT: vpermt2q %zmm11, %zmm7, %zmm82540; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} ymm12 = [0,5,10,15]2541; AVX512-FCP-NEXT: vmovdqa64 %zmm9, %zmm132542; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm12, %zmm132543; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm8 = zmm13[0,1,2,3],zmm8[4,5,6,7]2544; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm13 = [0,1,2,3,4,5,6,11]2545; AVX512-FCP-NEXT: vpermt2q %zmm6, %zmm13, %zmm82546; AVX512-FCP-NEXT: vmovdqa64 %zmm4, %zmm142547; AVX512-FCP-NEXT: vpermt2q %zmm5, %zmm7, %zmm142548; AVX512-FCP-NEXT: vpermi2q %zmm0, %zmm3, %zmm122549; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm12 = zmm12[0,1,2,3],zmm14[4,5,6,7]2550; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm13, %zmm122551; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} ymm15 = [1,6,11,0]2552; AVX512-FCP-NEXT: vmovdqa64 %zmm9, %zmm162553; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm15, %zmm162554; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm14 = [5,10,15,0,5,10,15,0]2555; AVX512-FCP-NEXT: # zmm14 = mem[0,1,2,3,0,1,2,3]2556; AVX512-FCP-NEXT: vmovdqa64 %zmm11, %zmm132557; AVX512-FCP-NEXT: vpermt2q %zmm10, %zmm14, %zmm132558; AVX512-FCP-NEXT: movb $7, %al2559; AVX512-FCP-NEXT: kmovw %eax, %k12560; AVX512-FCP-NEXT: vmovdqa64 %zmm16, %zmm13 {%k1}2561; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm16 = [0,1,2,3,4,5,6,12]2562; AVX512-FCP-NEXT: vpermt2q %zmm6, %zmm16, %zmm132563; AVX512-FCP-NEXT: vpermi2q %zmm4, %zmm5, %zmm142564; AVX512-FCP-NEXT: vpermi2q %zmm0, %zmm3, %zmm152565; AVX512-FCP-NEXT: vmovdqa64 %zmm15, %zmm14 {%k1}2566; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm16, %zmm142567; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm15 = [6,11,0,1,6,11,0,1]2568; AVX512-FCP-NEXT: # zmm15 = mem[0,1,2,3,0,1,2,3]2569; AVX512-FCP-NEXT: vmovdqa64 %zmm11, %zmm162570; AVX512-FCP-NEXT: vpermt2q %zmm10, %zmm15, %zmm162571; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} ymm17 = [2,7,12,0]2572; AVX512-FCP-NEXT: vmovdqa64 %zmm9, %zmm182573; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm17, %zmm182574; AVX512-FCP-NEXT: movb $56, %al2575; AVX512-FCP-NEXT: kmovw %eax, %k12576; AVX512-FCP-NEXT: vmovdqa64 %zmm16, %zmm18 {%k1}2577; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm16 = [0,1,2,3,4,5,8,13]2578; AVX512-FCP-NEXT: vpermt2q %zmm6, %zmm16, %zmm182579; AVX512-FCP-NEXT: vpermi2q %zmm4, %zmm5, %zmm152580; AVX512-FCP-NEXT: vpermi2q %zmm0, %zmm3, %zmm172581; AVX512-FCP-NEXT: vmovdqa64 %zmm15, %zmm17 {%k1}2582; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm16, %zmm172583; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm15 = [7,12,0,2,7,12,0,2]2584; AVX512-FCP-NEXT: # zmm15 = mem[0,1,2,3,0,1,2,3]2585; AVX512-FCP-NEXT: vmovdqa64 %zmm11, %zmm162586; AVX512-FCP-NEXT: vpermt2q %zmm10, %zmm15, %zmm162587; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} ymm19 = [11,0,5,0]2588; AVX512-FCP-NEXT: vmovdqa64 %zmm1, %zmm202589; AVX512-FCP-NEXT: vpermt2q %zmm9, %zmm19, %zmm202590; AVX512-FCP-NEXT: vmovdqa64 %zmm16, %zmm20 {%k1}2591; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm16 = [0,1,2,3,4,5,9,14]2592; AVX512-FCP-NEXT: vpermt2q %zmm6, %zmm16, %zmm202593; AVX512-FCP-NEXT: vpermi2q %zmm4, %zmm5, %zmm152594; AVX512-FCP-NEXT: vpermi2q %zmm3, %zmm0, %zmm192595; AVX512-FCP-NEXT: vmovdqa64 %zmm15, %zmm19 {%k1}2596; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm16, %zmm192597; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm15 = [0,5,0,11,0,5,0,11]2598; AVX512-FCP-NEXT: # zmm15 = mem[0,1,2,3,0,1,2,3]2599; AVX512-FCP-NEXT: vpermt2q %zmm11, %zmm15, %zmm102600; AVX512-FCP-NEXT: vpermt2q %zmm9, %zmm7, %zmm12601; AVX512-FCP-NEXT: vmovdqa64 %zmm10, %zmm1 {%k1}2602; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm9 = [0,1,2,3,4,5,10,15]2603; AVX512-FCP-NEXT: vpermt2q %zmm6, %zmm9, %zmm12604; AVX512-FCP-NEXT: vpermt2q %zmm5, %zmm15, %zmm42605; AVX512-FCP-NEXT: vpermt2q %zmm3, %zmm7, %zmm02606; AVX512-FCP-NEXT: vmovdqa64 %zmm4, %zmm0 {%k1}2607; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm9, %zmm02608; AVX512-FCP-NEXT: vmovdqa64 %zmm12, 64(%rsi)2609; AVX512-FCP-NEXT: vmovdqa64 %zmm8, (%rsi)2610; AVX512-FCP-NEXT: vmovdqa64 %zmm14, 64(%rdx)2611; AVX512-FCP-NEXT: vmovdqa64 %zmm13, (%rdx)2612; AVX512-FCP-NEXT: vmovdqa64 %zmm17, 64(%rcx)2613; AVX512-FCP-NEXT: vmovdqa64 %zmm18, (%rcx)2614; AVX512-FCP-NEXT: vmovdqa64 %zmm19, 64(%r8)2615; AVX512-FCP-NEXT: vmovdqa64 %zmm20, (%r8)2616; AVX512-FCP-NEXT: vmovdqa64 %zmm0, 64(%r9)2617; AVX512-FCP-NEXT: vmovdqa64 %zmm1, (%r9)2618; AVX512-FCP-NEXT: vzeroupper2619; AVX512-FCP-NEXT: retq2620;2621; AVX512DQ-LABEL: load_i64_stride5_vf16:2622; AVX512DQ: # %bb.0:2623; AVX512DQ-NEXT: vmovdqa64 576(%rdi), %zmm22624; AVX512DQ-NEXT: vmovdqa64 384(%rdi), %zmm02625; AVX512DQ-NEXT: vmovdqa64 320(%rdi), %zmm32626; AVX512DQ-NEXT: vmovdqa64 448(%rdi), %zmm52627; AVX512DQ-NEXT: vmovdqa64 512(%rdi), %zmm42628; AVX512DQ-NEXT: vmovdqa64 256(%rdi), %zmm62629; AVX512DQ-NEXT: vmovdqa64 (%rdi), %zmm92630; AVX512DQ-NEXT: vmovdqa64 64(%rdi), %zmm12631; AVX512DQ-NEXT: vmovdqa64 128(%rdi), %zmm112632; AVX512DQ-NEXT: vmovdqa64 192(%rdi), %zmm102633; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [12,1,6,0,12,1,6,0]2634; AVX512DQ-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]2635; AVX512DQ-NEXT: vmovdqa64 %zmm10, %zmm82636; AVX512DQ-NEXT: vpermt2q %zmm11, %zmm7, %zmm82637; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} ymm12 = [0,5,10,15]2638; AVX512DQ-NEXT: vmovdqa64 %zmm9, %zmm132639; AVX512DQ-NEXT: vpermt2q %zmm1, %zmm12, %zmm132640; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm8 = zmm13[0,1,2,3],zmm8[4,5,6,7]2641; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} zmm13 = [0,1,2,3,4,5,6,11]2642; AVX512DQ-NEXT: vpermt2q %zmm6, %zmm13, %zmm82643; AVX512DQ-NEXT: vmovdqa64 %zmm4, %zmm142644; AVX512DQ-NEXT: vpermt2q %zmm5, %zmm7, %zmm142645; AVX512DQ-NEXT: vpermi2q %zmm0, %zmm3, %zmm122646; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm12 = zmm12[0,1,2,3],zmm14[4,5,6,7]2647; AVX512DQ-NEXT: vpermt2q %zmm2, %zmm13, %zmm122648; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} ymm15 = [1,6,11,0]2649; AVX512DQ-NEXT: vmovdqa64 %zmm9, %zmm162650; AVX512DQ-NEXT: vpermt2q %zmm1, %zmm15, %zmm162651; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm14 = [5,10,15,0,5,10,15,0]2652; AVX512DQ-NEXT: # zmm14 = mem[0,1,2,3,0,1,2,3]2653; AVX512DQ-NEXT: vmovdqa64 %zmm11, %zmm132654; AVX512DQ-NEXT: vpermt2q %zmm10, %zmm14, %zmm132655; AVX512DQ-NEXT: movb $7, %al2656; AVX512DQ-NEXT: kmovw %eax, %k12657; AVX512DQ-NEXT: vmovdqa64 %zmm16, %zmm13 {%k1}2658; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} zmm16 = [0,1,2,3,4,5,6,12]2659; AVX512DQ-NEXT: vpermt2q %zmm6, %zmm16, %zmm132660; AVX512DQ-NEXT: vpermi2q %zmm4, %zmm5, %zmm142661; AVX512DQ-NEXT: vpermi2q %zmm0, %zmm3, %zmm152662; AVX512DQ-NEXT: vmovdqa64 %zmm15, %zmm14 {%k1}2663; AVX512DQ-NEXT: vpermt2q %zmm2, %zmm16, %zmm142664; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm15 = [6,11,0,1,6,11,0,1]2665; AVX512DQ-NEXT: # zmm15 = mem[0,1,2,3,0,1,2,3]2666; AVX512DQ-NEXT: vmovdqa64 %zmm11, %zmm162667; AVX512DQ-NEXT: vpermt2q %zmm10, %zmm15, %zmm162668; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} ymm17 = [2,7,12,0]2669; AVX512DQ-NEXT: vmovdqa64 %zmm9, %zmm182670; AVX512DQ-NEXT: vpermt2q %zmm1, %zmm17, %zmm182671; AVX512DQ-NEXT: movb $56, %al2672; AVX512DQ-NEXT: kmovw %eax, %k12673; AVX512DQ-NEXT: vmovdqa64 %zmm16, %zmm18 {%k1}2674; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} zmm16 = [0,1,2,3,4,5,8,13]2675; AVX512DQ-NEXT: vpermt2q %zmm6, %zmm16, %zmm182676; AVX512DQ-NEXT: vpermi2q %zmm4, %zmm5, %zmm152677; AVX512DQ-NEXT: vpermi2q %zmm0, %zmm3, %zmm172678; AVX512DQ-NEXT: vmovdqa64 %zmm15, %zmm17 {%k1}2679; AVX512DQ-NEXT: vpermt2q %zmm2, %zmm16, %zmm172680; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm15 = [7,12,0,2,7,12,0,2]2681; AVX512DQ-NEXT: # zmm15 = mem[0,1,2,3,0,1,2,3]2682; AVX512DQ-NEXT: vmovdqa64 %zmm11, %zmm162683; AVX512DQ-NEXT: vpermt2q %zmm10, %zmm15, %zmm162684; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} ymm19 = [11,0,5,0]2685; AVX512DQ-NEXT: vmovdqa64 %zmm1, %zmm202686; AVX512DQ-NEXT: vpermt2q %zmm9, %zmm19, %zmm202687; AVX512DQ-NEXT: vmovdqa64 %zmm16, %zmm20 {%k1}2688; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} zmm16 = [0,1,2,3,4,5,9,14]2689; AVX512DQ-NEXT: vpermt2q %zmm6, %zmm16, %zmm202690; AVX512DQ-NEXT: vpermi2q %zmm4, %zmm5, %zmm152691; AVX512DQ-NEXT: vpermi2q %zmm3, %zmm0, %zmm192692; AVX512DQ-NEXT: vmovdqa64 %zmm15, %zmm19 {%k1}2693; AVX512DQ-NEXT: vpermt2q %zmm2, %zmm16, %zmm192694; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm15 = [0,5,0,11,0,5,0,11]2695; AVX512DQ-NEXT: # zmm15 = mem[0,1,2,3,0,1,2,3]2696; AVX512DQ-NEXT: vpermt2q %zmm11, %zmm15, %zmm102697; AVX512DQ-NEXT: vpermt2q %zmm9, %zmm7, %zmm12698; AVX512DQ-NEXT: vmovdqa64 %zmm10, %zmm1 {%k1}2699; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} zmm9 = [0,1,2,3,4,5,10,15]2700; AVX512DQ-NEXT: vpermt2q %zmm6, %zmm9, %zmm12701; AVX512DQ-NEXT: vpermt2q %zmm5, %zmm15, %zmm42702; AVX512DQ-NEXT: vpermt2q %zmm3, %zmm7, %zmm02703; AVX512DQ-NEXT: vmovdqa64 %zmm4, %zmm0 {%k1}2704; AVX512DQ-NEXT: vpermt2q %zmm2, %zmm9, %zmm02705; AVX512DQ-NEXT: vmovdqa64 %zmm12, 64(%rsi)2706; AVX512DQ-NEXT: vmovdqa64 %zmm8, (%rsi)2707; AVX512DQ-NEXT: vmovdqa64 %zmm14, 64(%rdx)2708; AVX512DQ-NEXT: vmovdqa64 %zmm13, (%rdx)2709; AVX512DQ-NEXT: vmovdqa64 %zmm17, 64(%rcx)2710; AVX512DQ-NEXT: vmovdqa64 %zmm18, (%rcx)2711; AVX512DQ-NEXT: vmovdqa64 %zmm19, 64(%r8)2712; AVX512DQ-NEXT: vmovdqa64 %zmm20, (%r8)2713; AVX512DQ-NEXT: vmovdqa64 %zmm0, 64(%r9)2714; AVX512DQ-NEXT: vmovdqa64 %zmm1, (%r9)2715; AVX512DQ-NEXT: vzeroupper2716; AVX512DQ-NEXT: retq2717;2718; AVX512DQ-FCP-LABEL: load_i64_stride5_vf16:2719; AVX512DQ-FCP: # %bb.0:2720; AVX512DQ-FCP-NEXT: vmovdqa64 576(%rdi), %zmm22721; AVX512DQ-FCP-NEXT: vmovdqa64 384(%rdi), %zmm02722; AVX512DQ-FCP-NEXT: vmovdqa64 320(%rdi), %zmm32723; AVX512DQ-FCP-NEXT: vmovdqa64 448(%rdi), %zmm52724; AVX512DQ-FCP-NEXT: vmovdqa64 512(%rdi), %zmm42725; AVX512DQ-FCP-NEXT: vmovdqa64 256(%rdi), %zmm62726; AVX512DQ-FCP-NEXT: vmovdqa64 (%rdi), %zmm92727; AVX512DQ-FCP-NEXT: vmovdqa64 64(%rdi), %zmm12728; AVX512DQ-FCP-NEXT: vmovdqa64 128(%rdi), %zmm112729; AVX512DQ-FCP-NEXT: vmovdqa64 192(%rdi), %zmm102730; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [12,1,6,0,12,1,6,0]2731; AVX512DQ-FCP-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]2732; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm10, %zmm82733; AVX512DQ-FCP-NEXT: vpermt2q %zmm11, %zmm7, %zmm82734; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} ymm12 = [0,5,10,15]2735; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm9, %zmm132736; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm12, %zmm132737; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm8 = zmm13[0,1,2,3],zmm8[4,5,6,7]2738; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm13 = [0,1,2,3,4,5,6,11]2739; AVX512DQ-FCP-NEXT: vpermt2q %zmm6, %zmm13, %zmm82740; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm4, %zmm142741; AVX512DQ-FCP-NEXT: vpermt2q %zmm5, %zmm7, %zmm142742; AVX512DQ-FCP-NEXT: vpermi2q %zmm0, %zmm3, %zmm122743; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm12 = zmm12[0,1,2,3],zmm14[4,5,6,7]2744; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm13, %zmm122745; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} ymm15 = [1,6,11,0]2746; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm9, %zmm162747; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm15, %zmm162748; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm14 = [5,10,15,0,5,10,15,0]2749; AVX512DQ-FCP-NEXT: # zmm14 = mem[0,1,2,3,0,1,2,3]2750; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm11, %zmm132751; AVX512DQ-FCP-NEXT: vpermt2q %zmm10, %zmm14, %zmm132752; AVX512DQ-FCP-NEXT: movb $7, %al2753; AVX512DQ-FCP-NEXT: kmovw %eax, %k12754; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm16, %zmm13 {%k1}2755; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm16 = [0,1,2,3,4,5,6,12]2756; AVX512DQ-FCP-NEXT: vpermt2q %zmm6, %zmm16, %zmm132757; AVX512DQ-FCP-NEXT: vpermi2q %zmm4, %zmm5, %zmm142758; AVX512DQ-FCP-NEXT: vpermi2q %zmm0, %zmm3, %zmm152759; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm15, %zmm14 {%k1}2760; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm16, %zmm142761; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm15 = [6,11,0,1,6,11,0,1]2762; AVX512DQ-FCP-NEXT: # zmm15 = mem[0,1,2,3,0,1,2,3]2763; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm11, %zmm162764; AVX512DQ-FCP-NEXT: vpermt2q %zmm10, %zmm15, %zmm162765; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} ymm17 = [2,7,12,0]2766; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm9, %zmm182767; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm17, %zmm182768; AVX512DQ-FCP-NEXT: movb $56, %al2769; AVX512DQ-FCP-NEXT: kmovw %eax, %k12770; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm16, %zmm18 {%k1}2771; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm16 = [0,1,2,3,4,5,8,13]2772; AVX512DQ-FCP-NEXT: vpermt2q %zmm6, %zmm16, %zmm182773; AVX512DQ-FCP-NEXT: vpermi2q %zmm4, %zmm5, %zmm152774; AVX512DQ-FCP-NEXT: vpermi2q %zmm0, %zmm3, %zmm172775; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm15, %zmm17 {%k1}2776; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm16, %zmm172777; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm15 = [7,12,0,2,7,12,0,2]2778; AVX512DQ-FCP-NEXT: # zmm15 = mem[0,1,2,3,0,1,2,3]2779; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm11, %zmm162780; AVX512DQ-FCP-NEXT: vpermt2q %zmm10, %zmm15, %zmm162781; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} ymm19 = [11,0,5,0]2782; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, %zmm202783; AVX512DQ-FCP-NEXT: vpermt2q %zmm9, %zmm19, %zmm202784; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm16, %zmm20 {%k1}2785; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm16 = [0,1,2,3,4,5,9,14]2786; AVX512DQ-FCP-NEXT: vpermt2q %zmm6, %zmm16, %zmm202787; AVX512DQ-FCP-NEXT: vpermi2q %zmm4, %zmm5, %zmm152788; AVX512DQ-FCP-NEXT: vpermi2q %zmm3, %zmm0, %zmm192789; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm15, %zmm19 {%k1}2790; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm16, %zmm192791; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm15 = [0,5,0,11,0,5,0,11]2792; AVX512DQ-FCP-NEXT: # zmm15 = mem[0,1,2,3,0,1,2,3]2793; AVX512DQ-FCP-NEXT: vpermt2q %zmm11, %zmm15, %zmm102794; AVX512DQ-FCP-NEXT: vpermt2q %zmm9, %zmm7, %zmm12795; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm10, %zmm1 {%k1}2796; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm9 = [0,1,2,3,4,5,10,15]2797; AVX512DQ-FCP-NEXT: vpermt2q %zmm6, %zmm9, %zmm12798; AVX512DQ-FCP-NEXT: vpermt2q %zmm5, %zmm15, %zmm42799; AVX512DQ-FCP-NEXT: vpermt2q %zmm3, %zmm7, %zmm02800; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm4, %zmm0 {%k1}2801; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm9, %zmm02802; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm12, 64(%rsi)2803; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm8, (%rsi)2804; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm14, 64(%rdx)2805; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm13, (%rdx)2806; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm17, 64(%rcx)2807; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm18, (%rcx)2808; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm19, 64(%r8)2809; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm20, (%r8)2810; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, 64(%r9)2811; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, (%r9)2812; AVX512DQ-FCP-NEXT: vzeroupper2813; AVX512DQ-FCP-NEXT: retq2814;2815; AVX512BW-LABEL: load_i64_stride5_vf16:2816; AVX512BW: # %bb.0:2817; AVX512BW-NEXT: vmovdqa64 576(%rdi), %zmm22818; AVX512BW-NEXT: vmovdqa64 384(%rdi), %zmm02819; AVX512BW-NEXT: vmovdqa64 320(%rdi), %zmm32820; AVX512BW-NEXT: vmovdqa64 448(%rdi), %zmm52821; AVX512BW-NEXT: vmovdqa64 512(%rdi), %zmm42822; AVX512BW-NEXT: vmovdqa64 256(%rdi), %zmm62823; AVX512BW-NEXT: vmovdqa64 (%rdi), %zmm92824; AVX512BW-NEXT: vmovdqa64 64(%rdi), %zmm12825; AVX512BW-NEXT: vmovdqa64 128(%rdi), %zmm112826; AVX512BW-NEXT: vmovdqa64 192(%rdi), %zmm102827; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [12,1,6,0,12,1,6,0]2828; AVX512BW-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]2829; AVX512BW-NEXT: vmovdqa64 %zmm10, %zmm82830; AVX512BW-NEXT: vpermt2q %zmm11, %zmm7, %zmm82831; AVX512BW-NEXT: vpmovsxbq {{.*#+}} ymm12 = [0,5,10,15]2832; AVX512BW-NEXT: vmovdqa64 %zmm9, %zmm132833; AVX512BW-NEXT: vpermt2q %zmm1, %zmm12, %zmm132834; AVX512BW-NEXT: vshufi64x2 {{.*#+}} zmm8 = zmm13[0,1,2,3],zmm8[4,5,6,7]2835; AVX512BW-NEXT: vpmovsxbq {{.*#+}} zmm13 = [0,1,2,3,4,5,6,11]2836; AVX512BW-NEXT: vpermt2q %zmm6, %zmm13, %zmm82837; AVX512BW-NEXT: vmovdqa64 %zmm4, %zmm142838; AVX512BW-NEXT: vpermt2q %zmm5, %zmm7, %zmm142839; AVX512BW-NEXT: vpermi2q %zmm0, %zmm3, %zmm122840; AVX512BW-NEXT: vshufi64x2 {{.*#+}} zmm12 = zmm12[0,1,2,3],zmm14[4,5,6,7]2841; AVX512BW-NEXT: vpermt2q %zmm2, %zmm13, %zmm122842; AVX512BW-NEXT: vpmovsxbq {{.*#+}} ymm15 = [1,6,11,0]2843; AVX512BW-NEXT: vmovdqa64 %zmm9, %zmm162844; AVX512BW-NEXT: vpermt2q %zmm1, %zmm15, %zmm162845; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm14 = [5,10,15,0,5,10,15,0]2846; AVX512BW-NEXT: # zmm14 = mem[0,1,2,3,0,1,2,3]2847; AVX512BW-NEXT: vmovdqa64 %zmm11, %zmm132848; AVX512BW-NEXT: vpermt2q %zmm10, %zmm14, %zmm132849; AVX512BW-NEXT: movb $7, %al2850; AVX512BW-NEXT: kmovd %eax, %k12851; AVX512BW-NEXT: vmovdqa64 %zmm16, %zmm13 {%k1}2852; AVX512BW-NEXT: vpmovsxbq {{.*#+}} zmm16 = [0,1,2,3,4,5,6,12]2853; AVX512BW-NEXT: vpermt2q %zmm6, %zmm16, %zmm132854; AVX512BW-NEXT: vpermi2q %zmm4, %zmm5, %zmm142855; AVX512BW-NEXT: vpermi2q %zmm0, %zmm3, %zmm152856; AVX512BW-NEXT: vmovdqa64 %zmm15, %zmm14 {%k1}2857; AVX512BW-NEXT: vpermt2q %zmm2, %zmm16, %zmm142858; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm15 = [6,11,0,1,6,11,0,1]2859; AVX512BW-NEXT: # zmm15 = mem[0,1,2,3,0,1,2,3]2860; AVX512BW-NEXT: vmovdqa64 %zmm11, %zmm162861; AVX512BW-NEXT: vpermt2q %zmm10, %zmm15, %zmm162862; AVX512BW-NEXT: vpmovsxbq {{.*#+}} ymm17 = [2,7,12,0]2863; AVX512BW-NEXT: vmovdqa64 %zmm9, %zmm182864; AVX512BW-NEXT: vpermt2q %zmm1, %zmm17, %zmm182865; AVX512BW-NEXT: movb $56, %al2866; AVX512BW-NEXT: kmovd %eax, %k12867; AVX512BW-NEXT: vmovdqa64 %zmm16, %zmm18 {%k1}2868; AVX512BW-NEXT: vpmovsxbq {{.*#+}} zmm16 = [0,1,2,3,4,5,8,13]2869; AVX512BW-NEXT: vpermt2q %zmm6, %zmm16, %zmm182870; AVX512BW-NEXT: vpermi2q %zmm4, %zmm5, %zmm152871; AVX512BW-NEXT: vpermi2q %zmm0, %zmm3, %zmm172872; AVX512BW-NEXT: vmovdqa64 %zmm15, %zmm17 {%k1}2873; AVX512BW-NEXT: vpermt2q %zmm2, %zmm16, %zmm172874; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm15 = [7,12,0,2,7,12,0,2]2875; AVX512BW-NEXT: # zmm15 = mem[0,1,2,3,0,1,2,3]2876; AVX512BW-NEXT: vmovdqa64 %zmm11, %zmm162877; AVX512BW-NEXT: vpermt2q %zmm10, %zmm15, %zmm162878; AVX512BW-NEXT: vpmovsxbq {{.*#+}} ymm19 = [11,0,5,0]2879; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm202880; AVX512BW-NEXT: vpermt2q %zmm9, %zmm19, %zmm202881; AVX512BW-NEXT: vmovdqa64 %zmm16, %zmm20 {%k1}2882; AVX512BW-NEXT: vpmovsxbq {{.*#+}} zmm16 = [0,1,2,3,4,5,9,14]2883; AVX512BW-NEXT: vpermt2q %zmm6, %zmm16, %zmm202884; AVX512BW-NEXT: vpermi2q %zmm4, %zmm5, %zmm152885; AVX512BW-NEXT: vpermi2q %zmm3, %zmm0, %zmm192886; AVX512BW-NEXT: vmovdqa64 %zmm15, %zmm19 {%k1}2887; AVX512BW-NEXT: vpermt2q %zmm2, %zmm16, %zmm192888; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm15 = [0,5,0,11,0,5,0,11]2889; AVX512BW-NEXT: # zmm15 = mem[0,1,2,3,0,1,2,3]2890; AVX512BW-NEXT: vpermt2q %zmm11, %zmm15, %zmm102891; AVX512BW-NEXT: vpermt2q %zmm9, %zmm7, %zmm12892; AVX512BW-NEXT: vmovdqa64 %zmm10, %zmm1 {%k1}2893; AVX512BW-NEXT: vpmovsxbq {{.*#+}} zmm9 = [0,1,2,3,4,5,10,15]2894; AVX512BW-NEXT: vpermt2q %zmm6, %zmm9, %zmm12895; AVX512BW-NEXT: vpermt2q %zmm5, %zmm15, %zmm42896; AVX512BW-NEXT: vpermt2q %zmm3, %zmm7, %zmm02897; AVX512BW-NEXT: vmovdqa64 %zmm4, %zmm0 {%k1}2898; AVX512BW-NEXT: vpermt2q %zmm2, %zmm9, %zmm02899; AVX512BW-NEXT: vmovdqa64 %zmm12, 64(%rsi)2900; AVX512BW-NEXT: vmovdqa64 %zmm8, (%rsi)2901; AVX512BW-NEXT: vmovdqa64 %zmm14, 64(%rdx)2902; AVX512BW-NEXT: vmovdqa64 %zmm13, (%rdx)2903; AVX512BW-NEXT: vmovdqa64 %zmm17, 64(%rcx)2904; AVX512BW-NEXT: vmovdqa64 %zmm18, (%rcx)2905; AVX512BW-NEXT: vmovdqa64 %zmm19, 64(%r8)2906; AVX512BW-NEXT: vmovdqa64 %zmm20, (%r8)2907; AVX512BW-NEXT: vmovdqa64 %zmm0, 64(%r9)2908; AVX512BW-NEXT: vmovdqa64 %zmm1, (%r9)2909; AVX512BW-NEXT: vzeroupper2910; AVX512BW-NEXT: retq2911;2912; AVX512BW-FCP-LABEL: load_i64_stride5_vf16:2913; AVX512BW-FCP: # %bb.0:2914; AVX512BW-FCP-NEXT: vmovdqa64 576(%rdi), %zmm22915; AVX512BW-FCP-NEXT: vmovdqa64 384(%rdi), %zmm02916; AVX512BW-FCP-NEXT: vmovdqa64 320(%rdi), %zmm32917; AVX512BW-FCP-NEXT: vmovdqa64 448(%rdi), %zmm52918; AVX512BW-FCP-NEXT: vmovdqa64 512(%rdi), %zmm42919; AVX512BW-FCP-NEXT: vmovdqa64 256(%rdi), %zmm62920; AVX512BW-FCP-NEXT: vmovdqa64 (%rdi), %zmm92921; AVX512BW-FCP-NEXT: vmovdqa64 64(%rdi), %zmm12922; AVX512BW-FCP-NEXT: vmovdqa64 128(%rdi), %zmm112923; AVX512BW-FCP-NEXT: vmovdqa64 192(%rdi), %zmm102924; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [12,1,6,0,12,1,6,0]2925; AVX512BW-FCP-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]2926; AVX512BW-FCP-NEXT: vmovdqa64 %zmm10, %zmm82927; AVX512BW-FCP-NEXT: vpermt2q %zmm11, %zmm7, %zmm82928; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} ymm12 = [0,5,10,15]2929; AVX512BW-FCP-NEXT: vmovdqa64 %zmm9, %zmm132930; AVX512BW-FCP-NEXT: vpermt2q %zmm1, %zmm12, %zmm132931; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm8 = zmm13[0,1,2,3],zmm8[4,5,6,7]2932; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm13 = [0,1,2,3,4,5,6,11]2933; AVX512BW-FCP-NEXT: vpermt2q %zmm6, %zmm13, %zmm82934; AVX512BW-FCP-NEXT: vmovdqa64 %zmm4, %zmm142935; AVX512BW-FCP-NEXT: vpermt2q %zmm5, %zmm7, %zmm142936; AVX512BW-FCP-NEXT: vpermi2q %zmm0, %zmm3, %zmm122937; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm12 = zmm12[0,1,2,3],zmm14[4,5,6,7]2938; AVX512BW-FCP-NEXT: vpermt2q %zmm2, %zmm13, %zmm122939; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} ymm15 = [1,6,11,0]2940; AVX512BW-FCP-NEXT: vmovdqa64 %zmm9, %zmm162941; AVX512BW-FCP-NEXT: vpermt2q %zmm1, %zmm15, %zmm162942; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm14 = [5,10,15,0,5,10,15,0]2943; AVX512BW-FCP-NEXT: # zmm14 = mem[0,1,2,3,0,1,2,3]2944; AVX512BW-FCP-NEXT: vmovdqa64 %zmm11, %zmm132945; AVX512BW-FCP-NEXT: vpermt2q %zmm10, %zmm14, %zmm132946; AVX512BW-FCP-NEXT: movb $7, %al2947; AVX512BW-FCP-NEXT: kmovd %eax, %k12948; AVX512BW-FCP-NEXT: vmovdqa64 %zmm16, %zmm13 {%k1}2949; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm16 = [0,1,2,3,4,5,6,12]2950; AVX512BW-FCP-NEXT: vpermt2q %zmm6, %zmm16, %zmm132951; AVX512BW-FCP-NEXT: vpermi2q %zmm4, %zmm5, %zmm142952; AVX512BW-FCP-NEXT: vpermi2q %zmm0, %zmm3, %zmm152953; AVX512BW-FCP-NEXT: vmovdqa64 %zmm15, %zmm14 {%k1}2954; AVX512BW-FCP-NEXT: vpermt2q %zmm2, %zmm16, %zmm142955; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm15 = [6,11,0,1,6,11,0,1]2956; AVX512BW-FCP-NEXT: # zmm15 = mem[0,1,2,3,0,1,2,3]2957; AVX512BW-FCP-NEXT: vmovdqa64 %zmm11, %zmm162958; AVX512BW-FCP-NEXT: vpermt2q %zmm10, %zmm15, %zmm162959; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} ymm17 = [2,7,12,0]2960; AVX512BW-FCP-NEXT: vmovdqa64 %zmm9, %zmm182961; AVX512BW-FCP-NEXT: vpermt2q %zmm1, %zmm17, %zmm182962; AVX512BW-FCP-NEXT: movb $56, %al2963; AVX512BW-FCP-NEXT: kmovd %eax, %k12964; AVX512BW-FCP-NEXT: vmovdqa64 %zmm16, %zmm18 {%k1}2965; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm16 = [0,1,2,3,4,5,8,13]2966; AVX512BW-FCP-NEXT: vpermt2q %zmm6, %zmm16, %zmm182967; AVX512BW-FCP-NEXT: vpermi2q %zmm4, %zmm5, %zmm152968; AVX512BW-FCP-NEXT: vpermi2q %zmm0, %zmm3, %zmm172969; AVX512BW-FCP-NEXT: vmovdqa64 %zmm15, %zmm17 {%k1}2970; AVX512BW-FCP-NEXT: vpermt2q %zmm2, %zmm16, %zmm172971; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm15 = [7,12,0,2,7,12,0,2]2972; AVX512BW-FCP-NEXT: # zmm15 = mem[0,1,2,3,0,1,2,3]2973; AVX512BW-FCP-NEXT: vmovdqa64 %zmm11, %zmm162974; AVX512BW-FCP-NEXT: vpermt2q %zmm10, %zmm15, %zmm162975; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} ymm19 = [11,0,5,0]2976; AVX512BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm202977; AVX512BW-FCP-NEXT: vpermt2q %zmm9, %zmm19, %zmm202978; AVX512BW-FCP-NEXT: vmovdqa64 %zmm16, %zmm20 {%k1}2979; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm16 = [0,1,2,3,4,5,9,14]2980; AVX512BW-FCP-NEXT: vpermt2q %zmm6, %zmm16, %zmm202981; AVX512BW-FCP-NEXT: vpermi2q %zmm4, %zmm5, %zmm152982; AVX512BW-FCP-NEXT: vpermi2q %zmm3, %zmm0, %zmm192983; AVX512BW-FCP-NEXT: vmovdqa64 %zmm15, %zmm19 {%k1}2984; AVX512BW-FCP-NEXT: vpermt2q %zmm2, %zmm16, %zmm192985; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm15 = [0,5,0,11,0,5,0,11]2986; AVX512BW-FCP-NEXT: # zmm15 = mem[0,1,2,3,0,1,2,3]2987; AVX512BW-FCP-NEXT: vpermt2q %zmm11, %zmm15, %zmm102988; AVX512BW-FCP-NEXT: vpermt2q %zmm9, %zmm7, %zmm12989; AVX512BW-FCP-NEXT: vmovdqa64 %zmm10, %zmm1 {%k1}2990; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm9 = [0,1,2,3,4,5,10,15]2991; AVX512BW-FCP-NEXT: vpermt2q %zmm6, %zmm9, %zmm12992; AVX512BW-FCP-NEXT: vpermt2q %zmm5, %zmm15, %zmm42993; AVX512BW-FCP-NEXT: vpermt2q %zmm3, %zmm7, %zmm02994; AVX512BW-FCP-NEXT: vmovdqa64 %zmm4, %zmm0 {%k1}2995; AVX512BW-FCP-NEXT: vpermt2q %zmm2, %zmm9, %zmm02996; AVX512BW-FCP-NEXT: vmovdqa64 %zmm12, 64(%rsi)2997; AVX512BW-FCP-NEXT: vmovdqa64 %zmm8, (%rsi)2998; AVX512BW-FCP-NEXT: vmovdqa64 %zmm14, 64(%rdx)2999; AVX512BW-FCP-NEXT: vmovdqa64 %zmm13, (%rdx)3000; AVX512BW-FCP-NEXT: vmovdqa64 %zmm17, 64(%rcx)3001; AVX512BW-FCP-NEXT: vmovdqa64 %zmm18, (%rcx)3002; AVX512BW-FCP-NEXT: vmovdqa64 %zmm19, 64(%r8)3003; AVX512BW-FCP-NEXT: vmovdqa64 %zmm20, (%r8)3004; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, 64(%r9)3005; AVX512BW-FCP-NEXT: vmovdqa64 %zmm1, (%r9)3006; AVX512BW-FCP-NEXT: vzeroupper3007; AVX512BW-FCP-NEXT: retq3008;3009; AVX512DQ-BW-LABEL: load_i64_stride5_vf16:3010; AVX512DQ-BW: # %bb.0:3011; AVX512DQ-BW-NEXT: vmovdqa64 576(%rdi), %zmm23012; AVX512DQ-BW-NEXT: vmovdqa64 384(%rdi), %zmm03013; AVX512DQ-BW-NEXT: vmovdqa64 320(%rdi), %zmm33014; AVX512DQ-BW-NEXT: vmovdqa64 448(%rdi), %zmm53015; AVX512DQ-BW-NEXT: vmovdqa64 512(%rdi), %zmm43016; AVX512DQ-BW-NEXT: vmovdqa64 256(%rdi), %zmm63017; AVX512DQ-BW-NEXT: vmovdqa64 (%rdi), %zmm93018; AVX512DQ-BW-NEXT: vmovdqa64 64(%rdi), %zmm13019; AVX512DQ-BW-NEXT: vmovdqa64 128(%rdi), %zmm113020; AVX512DQ-BW-NEXT: vmovdqa64 192(%rdi), %zmm103021; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [12,1,6,0,12,1,6,0]3022; AVX512DQ-BW-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]3023; AVX512DQ-BW-NEXT: vmovdqa64 %zmm10, %zmm83024; AVX512DQ-BW-NEXT: vpermt2q %zmm11, %zmm7, %zmm83025; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} ymm12 = [0,5,10,15]3026; AVX512DQ-BW-NEXT: vmovdqa64 %zmm9, %zmm133027; AVX512DQ-BW-NEXT: vpermt2q %zmm1, %zmm12, %zmm133028; AVX512DQ-BW-NEXT: vshufi64x2 {{.*#+}} zmm8 = zmm13[0,1,2,3],zmm8[4,5,6,7]3029; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} zmm13 = [0,1,2,3,4,5,6,11]3030; AVX512DQ-BW-NEXT: vpermt2q %zmm6, %zmm13, %zmm83031; AVX512DQ-BW-NEXT: vmovdqa64 %zmm4, %zmm143032; AVX512DQ-BW-NEXT: vpermt2q %zmm5, %zmm7, %zmm143033; AVX512DQ-BW-NEXT: vpermi2q %zmm0, %zmm3, %zmm123034; AVX512DQ-BW-NEXT: vshufi64x2 {{.*#+}} zmm12 = zmm12[0,1,2,3],zmm14[4,5,6,7]3035; AVX512DQ-BW-NEXT: vpermt2q %zmm2, %zmm13, %zmm123036; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} ymm15 = [1,6,11,0]3037; AVX512DQ-BW-NEXT: vmovdqa64 %zmm9, %zmm163038; AVX512DQ-BW-NEXT: vpermt2q %zmm1, %zmm15, %zmm163039; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm14 = [5,10,15,0,5,10,15,0]3040; AVX512DQ-BW-NEXT: # zmm14 = mem[0,1,2,3,0,1,2,3]3041; AVX512DQ-BW-NEXT: vmovdqa64 %zmm11, %zmm133042; AVX512DQ-BW-NEXT: vpermt2q %zmm10, %zmm14, %zmm133043; AVX512DQ-BW-NEXT: movb $7, %al3044; AVX512DQ-BW-NEXT: kmovd %eax, %k13045; AVX512DQ-BW-NEXT: vmovdqa64 %zmm16, %zmm13 {%k1}3046; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} zmm16 = [0,1,2,3,4,5,6,12]3047; AVX512DQ-BW-NEXT: vpermt2q %zmm6, %zmm16, %zmm133048; AVX512DQ-BW-NEXT: vpermi2q %zmm4, %zmm5, %zmm143049; AVX512DQ-BW-NEXT: vpermi2q %zmm0, %zmm3, %zmm153050; AVX512DQ-BW-NEXT: vmovdqa64 %zmm15, %zmm14 {%k1}3051; AVX512DQ-BW-NEXT: vpermt2q %zmm2, %zmm16, %zmm143052; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm15 = [6,11,0,1,6,11,0,1]3053; AVX512DQ-BW-NEXT: # zmm15 = mem[0,1,2,3,0,1,2,3]3054; AVX512DQ-BW-NEXT: vmovdqa64 %zmm11, %zmm163055; AVX512DQ-BW-NEXT: vpermt2q %zmm10, %zmm15, %zmm163056; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} ymm17 = [2,7,12,0]3057; AVX512DQ-BW-NEXT: vmovdqa64 %zmm9, %zmm183058; AVX512DQ-BW-NEXT: vpermt2q %zmm1, %zmm17, %zmm183059; AVX512DQ-BW-NEXT: movb $56, %al3060; AVX512DQ-BW-NEXT: kmovd %eax, %k13061; AVX512DQ-BW-NEXT: vmovdqa64 %zmm16, %zmm18 {%k1}3062; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} zmm16 = [0,1,2,3,4,5,8,13]3063; AVX512DQ-BW-NEXT: vpermt2q %zmm6, %zmm16, %zmm183064; AVX512DQ-BW-NEXT: vpermi2q %zmm4, %zmm5, %zmm153065; AVX512DQ-BW-NEXT: vpermi2q %zmm0, %zmm3, %zmm173066; AVX512DQ-BW-NEXT: vmovdqa64 %zmm15, %zmm17 {%k1}3067; AVX512DQ-BW-NEXT: vpermt2q %zmm2, %zmm16, %zmm173068; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm15 = [7,12,0,2,7,12,0,2]3069; AVX512DQ-BW-NEXT: # zmm15 = mem[0,1,2,3,0,1,2,3]3070; AVX512DQ-BW-NEXT: vmovdqa64 %zmm11, %zmm163071; AVX512DQ-BW-NEXT: vpermt2q %zmm10, %zmm15, %zmm163072; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} ymm19 = [11,0,5,0]3073; AVX512DQ-BW-NEXT: vmovdqa64 %zmm1, %zmm203074; AVX512DQ-BW-NEXT: vpermt2q %zmm9, %zmm19, %zmm203075; AVX512DQ-BW-NEXT: vmovdqa64 %zmm16, %zmm20 {%k1}3076; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} zmm16 = [0,1,2,3,4,5,9,14]3077; AVX512DQ-BW-NEXT: vpermt2q %zmm6, %zmm16, %zmm203078; AVX512DQ-BW-NEXT: vpermi2q %zmm4, %zmm5, %zmm153079; AVX512DQ-BW-NEXT: vpermi2q %zmm3, %zmm0, %zmm193080; AVX512DQ-BW-NEXT: vmovdqa64 %zmm15, %zmm19 {%k1}3081; AVX512DQ-BW-NEXT: vpermt2q %zmm2, %zmm16, %zmm193082; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm15 = [0,5,0,11,0,5,0,11]3083; AVX512DQ-BW-NEXT: # zmm15 = mem[0,1,2,3,0,1,2,3]3084; AVX512DQ-BW-NEXT: vpermt2q %zmm11, %zmm15, %zmm103085; AVX512DQ-BW-NEXT: vpermt2q %zmm9, %zmm7, %zmm13086; AVX512DQ-BW-NEXT: vmovdqa64 %zmm10, %zmm1 {%k1}3087; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} zmm9 = [0,1,2,3,4,5,10,15]3088; AVX512DQ-BW-NEXT: vpermt2q %zmm6, %zmm9, %zmm13089; AVX512DQ-BW-NEXT: vpermt2q %zmm5, %zmm15, %zmm43090; AVX512DQ-BW-NEXT: vpermt2q %zmm3, %zmm7, %zmm03091; AVX512DQ-BW-NEXT: vmovdqa64 %zmm4, %zmm0 {%k1}3092; AVX512DQ-BW-NEXT: vpermt2q %zmm2, %zmm9, %zmm03093; AVX512DQ-BW-NEXT: vmovdqa64 %zmm12, 64(%rsi)3094; AVX512DQ-BW-NEXT: vmovdqa64 %zmm8, (%rsi)3095; AVX512DQ-BW-NEXT: vmovdqa64 %zmm14, 64(%rdx)3096; AVX512DQ-BW-NEXT: vmovdqa64 %zmm13, (%rdx)3097; AVX512DQ-BW-NEXT: vmovdqa64 %zmm17, 64(%rcx)3098; AVX512DQ-BW-NEXT: vmovdqa64 %zmm18, (%rcx)3099; AVX512DQ-BW-NEXT: vmovdqa64 %zmm19, 64(%r8)3100; AVX512DQ-BW-NEXT: vmovdqa64 %zmm20, (%r8)3101; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, 64(%r9)3102; AVX512DQ-BW-NEXT: vmovdqa64 %zmm1, (%r9)3103; AVX512DQ-BW-NEXT: vzeroupper3104; AVX512DQ-BW-NEXT: retq3105;3106; AVX512DQ-BW-FCP-LABEL: load_i64_stride5_vf16:3107; AVX512DQ-BW-FCP: # %bb.0:3108; AVX512DQ-BW-FCP-NEXT: vmovdqa64 576(%rdi), %zmm23109; AVX512DQ-BW-FCP-NEXT: vmovdqa64 384(%rdi), %zmm03110; AVX512DQ-BW-FCP-NEXT: vmovdqa64 320(%rdi), %zmm33111; AVX512DQ-BW-FCP-NEXT: vmovdqa64 448(%rdi), %zmm53112; AVX512DQ-BW-FCP-NEXT: vmovdqa64 512(%rdi), %zmm43113; AVX512DQ-BW-FCP-NEXT: vmovdqa64 256(%rdi), %zmm63114; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rdi), %zmm93115; AVX512DQ-BW-FCP-NEXT: vmovdqa64 64(%rdi), %zmm13116; AVX512DQ-BW-FCP-NEXT: vmovdqa64 128(%rdi), %zmm113117; AVX512DQ-BW-FCP-NEXT: vmovdqa64 192(%rdi), %zmm103118; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm7 = [12,1,6,0,12,1,6,0]3119; AVX512DQ-BW-FCP-NEXT: # zmm7 = mem[0,1,2,3,0,1,2,3]3120; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm10, %zmm83121; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm11, %zmm7, %zmm83122; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} ymm12 = [0,5,10,15]3123; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm9, %zmm133124; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm1, %zmm12, %zmm133125; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm8 = zmm13[0,1,2,3],zmm8[4,5,6,7]3126; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm13 = [0,1,2,3,4,5,6,11]3127; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm6, %zmm13, %zmm83128; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm4, %zmm143129; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm5, %zmm7, %zmm143130; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm0, %zmm3, %zmm123131; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm12 = zmm12[0,1,2,3],zmm14[4,5,6,7]3132; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm2, %zmm13, %zmm123133; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} ymm15 = [1,6,11,0]3134; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm9, %zmm163135; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm1, %zmm15, %zmm163136; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm14 = [5,10,15,0,5,10,15,0]3137; AVX512DQ-BW-FCP-NEXT: # zmm14 = mem[0,1,2,3,0,1,2,3]3138; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm11, %zmm133139; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm10, %zmm14, %zmm133140; AVX512DQ-BW-FCP-NEXT: movb $7, %al3141; AVX512DQ-BW-FCP-NEXT: kmovd %eax, %k13142; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm16, %zmm13 {%k1}3143; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm16 = [0,1,2,3,4,5,6,12]3144; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm6, %zmm16, %zmm133145; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm4, %zmm5, %zmm143146; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm0, %zmm3, %zmm153147; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm15, %zmm14 {%k1}3148; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm2, %zmm16, %zmm143149; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm15 = [6,11,0,1,6,11,0,1]3150; AVX512DQ-BW-FCP-NEXT: # zmm15 = mem[0,1,2,3,0,1,2,3]3151; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm11, %zmm163152; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm10, %zmm15, %zmm163153; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} ymm17 = [2,7,12,0]3154; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm9, %zmm183155; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm1, %zmm17, %zmm183156; AVX512DQ-BW-FCP-NEXT: movb $56, %al3157; AVX512DQ-BW-FCP-NEXT: kmovd %eax, %k13158; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm16, %zmm18 {%k1}3159; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm16 = [0,1,2,3,4,5,8,13]3160; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm6, %zmm16, %zmm183161; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm4, %zmm5, %zmm153162; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm0, %zmm3, %zmm173163; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm15, %zmm17 {%k1}3164; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm2, %zmm16, %zmm173165; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm15 = [7,12,0,2,7,12,0,2]3166; AVX512DQ-BW-FCP-NEXT: # zmm15 = mem[0,1,2,3,0,1,2,3]3167; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm11, %zmm163168; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm10, %zmm15, %zmm163169; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} ymm19 = [11,0,5,0]3170; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm203171; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm9, %zmm19, %zmm203172; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm16, %zmm20 {%k1}3173; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm16 = [0,1,2,3,4,5,9,14]3174; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm6, %zmm16, %zmm203175; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm4, %zmm5, %zmm153176; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm3, %zmm0, %zmm193177; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm15, %zmm19 {%k1}3178; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm2, %zmm16, %zmm193179; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm15 = [0,5,0,11,0,5,0,11]3180; AVX512DQ-BW-FCP-NEXT: # zmm15 = mem[0,1,2,3,0,1,2,3]3181; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm11, %zmm15, %zmm103182; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm9, %zmm7, %zmm13183; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm10, %zmm1 {%k1}3184; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm9 = [0,1,2,3,4,5,10,15]3185; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm6, %zmm9, %zmm13186; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm5, %zmm15, %zmm43187; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm3, %zmm7, %zmm03188; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm4, %zmm0 {%k1}3189; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm2, %zmm9, %zmm03190; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm12, 64(%rsi)3191; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm8, (%rsi)3192; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm14, 64(%rdx)3193; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm13, (%rdx)3194; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm17, 64(%rcx)3195; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm18, (%rcx)3196; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm19, 64(%r8)3197; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm20, (%r8)3198; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, 64(%r9)3199; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm1, (%r9)3200; AVX512DQ-BW-FCP-NEXT: vzeroupper3201; AVX512DQ-BW-FCP-NEXT: retq3202 %wide.vec = load <80 x i64>, ptr %in.vec, align 643203 %strided.vec0 = shufflevector <80 x i64> %wide.vec, <80 x i64> poison, <16 x i32> <i32 0, i32 5, i32 10, i32 15, i32 20, i32 25, i32 30, i32 35, i32 40, i32 45, i32 50, i32 55, i32 60, i32 65, i32 70, i32 75>3204 %strided.vec1 = shufflevector <80 x i64> %wide.vec, <80 x i64> poison, <16 x i32> <i32 1, i32 6, i32 11, i32 16, i32 21, i32 26, i32 31, i32 36, i32 41, i32 46, i32 51, i32 56, i32 61, i32 66, i32 71, i32 76>3205 %strided.vec2 = shufflevector <80 x i64> %wide.vec, <80 x i64> poison, <16 x i32> <i32 2, i32 7, i32 12, i32 17, i32 22, i32 27, i32 32, i32 37, i32 42, i32 47, i32 52, i32 57, i32 62, i32 67, i32 72, i32 77>3206 %strided.vec3 = shufflevector <80 x i64> %wide.vec, <80 x i64> poison, <16 x i32> <i32 3, i32 8, i32 13, i32 18, i32 23, i32 28, i32 33, i32 38, i32 43, i32 48, i32 53, i32 58, i32 63, i32 68, i32 73, i32 78>3207 %strided.vec4 = shufflevector <80 x i64> %wide.vec, <80 x i64> poison, <16 x i32> <i32 4, i32 9, i32 14, i32 19, i32 24, i32 29, i32 34, i32 39, i32 44, i32 49, i32 54, i32 59, i32 64, i32 69, i32 74, i32 79>3208 store <16 x i64> %strided.vec0, ptr %out.vec0, align 643209 store <16 x i64> %strided.vec1, ptr %out.vec1, align 643210 store <16 x i64> %strided.vec2, ptr %out.vec2, align 643211 store <16 x i64> %strided.vec3, ptr %out.vec3, align 643212 store <16 x i64> %strided.vec4, ptr %out.vec4, align 643213 ret void3214}3215 3216define void @load_i64_stride5_vf32(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr %out.vec2, ptr %out.vec3, ptr %out.vec4) nounwind {3217; SSE-LABEL: load_i64_stride5_vf32:3218; SSE: # %bb.0:3219; SSE-NEXT: subq $920, %rsp # imm = 0x3983220; SSE-NEXT: movapd 224(%rdi), %xmm53221; SSE-NEXT: movapd 144(%rdi), %xmm43222; SSE-NEXT: movapd 64(%rdi), %xmm33223; SSE-NEXT: movapd 176(%rdi), %xmm73224; SSE-NEXT: movapd 96(%rdi), %xmm63225; SSE-NEXT: movapd 208(%rdi), %xmm93226; SSE-NEXT: movapd 128(%rdi), %xmm103227; SSE-NEXT: movapd (%rdi), %xmm123228; SSE-NEXT: movapd 16(%rdi), %xmm83229; SSE-NEXT: movapd 32(%rdi), %xmm13230; SSE-NEXT: movapd 48(%rdi), %xmm113231; SSE-NEXT: movapd 160(%rdi), %xmm133232; SSE-NEXT: movapd 192(%rdi), %xmm03233; SSE-NEXT: movapd 80(%rdi), %xmm143234; SSE-NEXT: movapd 112(%rdi), %xmm23235; SSE-NEXT: movapd %xmm1, %xmm153236; SSE-NEXT: movsd {{.*#+}} xmm15 = xmm12[0],xmm15[1]3237; SSE-NEXT: movapd %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3238; SSE-NEXT: shufpd {{.*#+}} xmm12 = xmm12[1],xmm11[0]3239; SSE-NEXT: movapd %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3240; SSE-NEXT: movsd {{.*#+}} xmm11 = xmm8[0],xmm11[1]3241; SSE-NEXT: movapd %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3242; SSE-NEXT: shufpd {{.*#+}} xmm8 = xmm8[1],xmm3[0]3243; SSE-NEXT: movapd %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3244; SSE-NEXT: movsd {{.*#+}} xmm3 = xmm1[0],xmm3[1]3245; SSE-NEXT: movapd %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3246; SSE-NEXT: movapd %xmm2, %xmm13247; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm14[0],xmm1[1]3248; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3249; SSE-NEXT: shufpd {{.*#+}} xmm14 = xmm14[1],xmm10[0]3250; SSE-NEXT: movapd %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3251; SSE-NEXT: movsd {{.*#+}} xmm10 = xmm6[0],xmm10[1]3252; SSE-NEXT: movapd %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3253; SSE-NEXT: shufpd {{.*#+}} xmm6 = xmm6[1],xmm4[0]3254; SSE-NEXT: movapd %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3255; SSE-NEXT: movsd {{.*#+}} xmm4 = xmm2[0],xmm4[1]3256; SSE-NEXT: movapd %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3257; SSE-NEXT: movapd %xmm0, %xmm13258; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm13[0],xmm1[1]3259; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3260; SSE-NEXT: shufpd {{.*#+}} xmm13 = xmm13[1],xmm9[0]3261; SSE-NEXT: movapd %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3262; SSE-NEXT: movsd {{.*#+}} xmm9 = xmm7[0],xmm9[1]3263; SSE-NEXT: movapd %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3264; SSE-NEXT: shufpd {{.*#+}} xmm7 = xmm7[1],xmm5[0]3265; SSE-NEXT: movapd %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3266; SSE-NEXT: movsd {{.*#+}} xmm5 = xmm0[0],xmm5[1]3267; SSE-NEXT: movapd %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3268; SSE-NEXT: movapd 240(%rdi), %xmm23269; SSE-NEXT: movapd 272(%rdi), %xmm03270; SSE-NEXT: movapd %xmm0, %xmm13271; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]3272; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3273; SSE-NEXT: movapd 288(%rdi), %xmm13274; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]3275; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3276; SSE-NEXT: movapd 256(%rdi), %xmm23277; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]3278; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3279; SSE-NEXT: movapd 304(%rdi), %xmm13280; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]3281; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3282; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]3283; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3284; SSE-NEXT: movapd 320(%rdi), %xmm23285; SSE-NEXT: movapd 352(%rdi), %xmm03286; SSE-NEXT: movapd %xmm0, %xmm13287; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]3288; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3289; SSE-NEXT: movapd 368(%rdi), %xmm13290; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]3291; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3292; SSE-NEXT: movapd 336(%rdi), %xmm23293; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]3294; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3295; SSE-NEXT: movapd 384(%rdi), %xmm13296; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]3297; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3298; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]3299; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3300; SSE-NEXT: movapd 400(%rdi), %xmm23301; SSE-NEXT: movapd 432(%rdi), %xmm03302; SSE-NEXT: movapd %xmm0, %xmm13303; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]3304; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3305; SSE-NEXT: movapd 448(%rdi), %xmm13306; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]3307; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3308; SSE-NEXT: movapd 416(%rdi), %xmm23309; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]3310; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3311; SSE-NEXT: movapd 464(%rdi), %xmm13312; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]3313; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3314; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]3315; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3316; SSE-NEXT: movapd 480(%rdi), %xmm23317; SSE-NEXT: movapd 512(%rdi), %xmm03318; SSE-NEXT: movapd %xmm0, %xmm13319; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]3320; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3321; SSE-NEXT: movapd 528(%rdi), %xmm13322; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]3323; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3324; SSE-NEXT: movapd 496(%rdi), %xmm23325; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]3326; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3327; SSE-NEXT: movapd 544(%rdi), %xmm13328; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]3329; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3330; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]3331; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3332; SSE-NEXT: movapd 560(%rdi), %xmm23333; SSE-NEXT: movapd 592(%rdi), %xmm03334; SSE-NEXT: movapd %xmm0, %xmm13335; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]3336; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3337; SSE-NEXT: movapd 608(%rdi), %xmm13338; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]3339; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3340; SSE-NEXT: movapd 576(%rdi), %xmm23341; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]3342; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3343; SSE-NEXT: movapd 624(%rdi), %xmm13344; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]3345; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3346; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]3347; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3348; SSE-NEXT: movapd 640(%rdi), %xmm23349; SSE-NEXT: movapd 672(%rdi), %xmm03350; SSE-NEXT: movapd %xmm0, %xmm13351; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]3352; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3353; SSE-NEXT: movapd 688(%rdi), %xmm13354; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]3355; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3356; SSE-NEXT: movapd 656(%rdi), %xmm23357; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]3358; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3359; SSE-NEXT: movapd 704(%rdi), %xmm13360; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]3361; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3362; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]3363; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3364; SSE-NEXT: movapd 720(%rdi), %xmm23365; SSE-NEXT: movapd 752(%rdi), %xmm03366; SSE-NEXT: movapd %xmm0, %xmm13367; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]3368; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3369; SSE-NEXT: movapd 768(%rdi), %xmm13370; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]3371; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3372; SSE-NEXT: movapd 736(%rdi), %xmm23373; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]3374; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3375; SSE-NEXT: movapd 784(%rdi), %xmm13376; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]3377; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3378; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]3379; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3380; SSE-NEXT: movapd 800(%rdi), %xmm23381; SSE-NEXT: movapd 832(%rdi), %xmm03382; SSE-NEXT: movapd %xmm0, %xmm13383; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]3384; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3385; SSE-NEXT: movapd 848(%rdi), %xmm13386; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]3387; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3388; SSE-NEXT: movapd 816(%rdi), %xmm23389; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]3390; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3391; SSE-NEXT: movapd 864(%rdi), %xmm13392; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]3393; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3394; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]3395; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3396; SSE-NEXT: movapd 880(%rdi), %xmm23397; SSE-NEXT: movapd 912(%rdi), %xmm03398; SSE-NEXT: movapd %xmm0, %xmm13399; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]3400; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3401; SSE-NEXT: movapd 928(%rdi), %xmm13402; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]3403; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3404; SSE-NEXT: movapd 896(%rdi), %xmm23405; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]3406; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3407; SSE-NEXT: movapd 944(%rdi), %xmm13408; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]3409; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3410; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]3411; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3412; SSE-NEXT: movapd 960(%rdi), %xmm103413; SSE-NEXT: movapd 992(%rdi), %xmm03414; SSE-NEXT: movapd %xmm0, %xmm143415; SSE-NEXT: movsd {{.*#+}} xmm14 = xmm10[0],xmm14[1]3416; SSE-NEXT: movapd 1008(%rdi), %xmm153417; SSE-NEXT: shufpd {{.*#+}} xmm10 = xmm10[1],xmm15[0]3418; SSE-NEXT: movapd 976(%rdi), %xmm23419; SSE-NEXT: movsd {{.*#+}} xmm15 = xmm2[0],xmm15[1]3420; SSE-NEXT: movapd 1024(%rdi), %xmm13421; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]3422; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3423; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]3424; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3425; SSE-NEXT: movapd 1040(%rdi), %xmm83426; SSE-NEXT: movapd 1072(%rdi), %xmm03427; SSE-NEXT: movapd %xmm0, %xmm93428; SSE-NEXT: movsd {{.*#+}} xmm9 = xmm8[0],xmm9[1]3429; SSE-NEXT: movapd 1088(%rdi), %xmm113430; SSE-NEXT: shufpd {{.*#+}} xmm8 = xmm8[1],xmm11[0]3431; SSE-NEXT: movapd 1056(%rdi), %xmm23432; SSE-NEXT: movsd {{.*#+}} xmm11 = xmm2[0],xmm11[1]3433; SSE-NEXT: movapd 1104(%rdi), %xmm13434; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]3435; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3436; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]3437; SSE-NEXT: movapd %xmm1, (%rsp) # 16-byte Spill3438; SSE-NEXT: movapd 1120(%rdi), %xmm53439; SSE-NEXT: movapd 1152(%rdi), %xmm13440; SSE-NEXT: movapd %xmm1, %xmm33441; SSE-NEXT: movsd {{.*#+}} xmm3 = xmm5[0],xmm3[1]3442; SSE-NEXT: movapd 1168(%rdi), %xmm63443; SSE-NEXT: shufpd {{.*#+}} xmm5 = xmm5[1],xmm6[0]3444; SSE-NEXT: movapd 1136(%rdi), %xmm123445; SSE-NEXT: movsd {{.*#+}} xmm6 = xmm12[0],xmm6[1]3446; SSE-NEXT: movapd 1184(%rdi), %xmm03447; SSE-NEXT: shufpd {{.*#+}} xmm12 = xmm12[1],xmm0[0]3448; SSE-NEXT: movsd {{.*#+}} xmm0 = xmm1[0],xmm0[1]3449; SSE-NEXT: movapd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3450; SSE-NEXT: movapd 1200(%rdi), %xmm03451; SSE-NEXT: movapd 1232(%rdi), %xmm23452; SSE-NEXT: movapd %xmm2, %xmm13453; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]3454; SSE-NEXT: movapd 1248(%rdi), %xmm43455; SSE-NEXT: shufpd {{.*#+}} xmm0 = xmm0[1],xmm4[0]3456; SSE-NEXT: movapd 1216(%rdi), %xmm73457; SSE-NEXT: movsd {{.*#+}} xmm4 = xmm7[0],xmm4[1]3458; SSE-NEXT: movapd 1264(%rdi), %xmm133459; SSE-NEXT: shufpd {{.*#+}} xmm7 = xmm7[1],xmm13[0]3460; SSE-NEXT: movsd {{.*#+}} xmm13 = xmm2[0],xmm13[1]3461; SSE-NEXT: movapd %xmm3, 224(%rsi)3462; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload3463; SSE-NEXT: movaps %xmm2, 160(%rsi)3464; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload3465; SSE-NEXT: movaps %xmm2, 96(%rsi)3466; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload3467; SSE-NEXT: movaps %xmm2, 32(%rsi)3468; SSE-NEXT: movapd %xmm1, 240(%rsi)3469; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload3470; SSE-NEXT: movaps %xmm1, 176(%rsi)3471; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload3472; SSE-NEXT: movaps %xmm1, 112(%rsi)3473; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload3474; SSE-NEXT: movaps %xmm1, 48(%rsi)3475; SSE-NEXT: movapd %xmm14, 192(%rsi)3476; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload3477; SSE-NEXT: movaps %xmm1, 128(%rsi)3478; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload3479; SSE-NEXT: movaps %xmm1, 64(%rsi)3480; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload3481; SSE-NEXT: movaps %xmm1, (%rsi)3482; SSE-NEXT: movapd %xmm9, 208(%rsi)3483; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload3484; SSE-NEXT: movaps %xmm1, 144(%rsi)3485; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload3486; SSE-NEXT: movaps %xmm1, 80(%rsi)3487; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload3488; SSE-NEXT: movaps %xmm1, 16(%rsi)3489; SSE-NEXT: movapd %xmm5, 224(%rdx)3490; SSE-NEXT: movapd %xmm0, 240(%rdx)3491; SSE-NEXT: movapd %xmm10, 192(%rdx)3492; SSE-NEXT: movapd %xmm8, 208(%rdx)3493; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3494; SSE-NEXT: movaps %xmm0, 160(%rdx)3495; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3496; SSE-NEXT: movaps %xmm0, 176(%rdx)3497; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3498; SSE-NEXT: movaps %xmm0, 128(%rdx)3499; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3500; SSE-NEXT: movaps %xmm0, 144(%rdx)3501; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3502; SSE-NEXT: movaps %xmm0, 96(%rdx)3503; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3504; SSE-NEXT: movaps %xmm0, 112(%rdx)3505; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3506; SSE-NEXT: movaps %xmm0, 64(%rdx)3507; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3508; SSE-NEXT: movaps %xmm0, 80(%rdx)3509; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3510; SSE-NEXT: movaps %xmm0, 32(%rdx)3511; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3512; SSE-NEXT: movaps %xmm0, 48(%rdx)3513; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3514; SSE-NEXT: movaps %xmm0, (%rdx)3515; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3516; SSE-NEXT: movaps %xmm0, 16(%rdx)3517; SSE-NEXT: movapd %xmm4, 240(%rcx)3518; SSE-NEXT: movapd %xmm6, 224(%rcx)3519; SSE-NEXT: movapd %xmm11, 208(%rcx)3520; SSE-NEXT: movapd %xmm15, 192(%rcx)3521; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3522; SSE-NEXT: movaps %xmm0, 176(%rcx)3523; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3524; SSE-NEXT: movaps %xmm0, 160(%rcx)3525; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3526; SSE-NEXT: movaps %xmm0, 144(%rcx)3527; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3528; SSE-NEXT: movaps %xmm0, 128(%rcx)3529; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3530; SSE-NEXT: movaps %xmm0, 112(%rcx)3531; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3532; SSE-NEXT: movaps %xmm0, 96(%rcx)3533; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3534; SSE-NEXT: movaps %xmm0, 80(%rcx)3535; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3536; SSE-NEXT: movaps %xmm0, 64(%rcx)3537; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3538; SSE-NEXT: movaps %xmm0, 48(%rcx)3539; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3540; SSE-NEXT: movaps %xmm0, 32(%rcx)3541; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3542; SSE-NEXT: movaps %xmm0, 16(%rcx)3543; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3544; SSE-NEXT: movaps %xmm0, (%rcx)3545; SSE-NEXT: movapd %xmm7, 240(%r8)3546; SSE-NEXT: movapd %xmm12, 224(%r8)3547; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3548; SSE-NEXT: movaps %xmm0, 208(%r8)3549; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3550; SSE-NEXT: movaps %xmm0, 192(%r8)3551; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3552; SSE-NEXT: movaps %xmm0, 176(%r8)3553; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3554; SSE-NEXT: movaps %xmm0, 160(%r8)3555; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3556; SSE-NEXT: movaps %xmm0, 144(%r8)3557; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3558; SSE-NEXT: movaps %xmm0, 128(%r8)3559; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3560; SSE-NEXT: movaps %xmm0, 112(%r8)3561; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3562; SSE-NEXT: movaps %xmm0, 96(%r8)3563; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3564; SSE-NEXT: movaps %xmm0, 80(%r8)3565; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3566; SSE-NEXT: movaps %xmm0, 64(%r8)3567; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3568; SSE-NEXT: movaps %xmm0, 48(%r8)3569; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3570; SSE-NEXT: movaps %xmm0, 32(%r8)3571; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3572; SSE-NEXT: movaps %xmm0, 16(%r8)3573; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3574; SSE-NEXT: movaps %xmm0, (%r8)3575; SSE-NEXT: movapd %xmm13, 240(%r9)3576; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3577; SSE-NEXT: movaps %xmm0, 224(%r9)3578; SSE-NEXT: movaps (%rsp), %xmm0 # 16-byte Reload3579; SSE-NEXT: movaps %xmm0, 208(%r9)3580; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3581; SSE-NEXT: movaps %xmm0, 192(%r9)3582; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3583; SSE-NEXT: movaps %xmm0, 176(%r9)3584; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3585; SSE-NEXT: movaps %xmm0, 160(%r9)3586; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3587; SSE-NEXT: movaps %xmm0, 144(%r9)3588; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3589; SSE-NEXT: movaps %xmm0, 128(%r9)3590; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3591; SSE-NEXT: movaps %xmm0, 112(%r9)3592; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3593; SSE-NEXT: movaps %xmm0, 96(%r9)3594; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3595; SSE-NEXT: movaps %xmm0, 80(%r9)3596; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3597; SSE-NEXT: movaps %xmm0, 64(%r9)3598; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3599; SSE-NEXT: movaps %xmm0, 48(%r9)3600; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3601; SSE-NEXT: movaps %xmm0, 32(%r9)3602; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3603; SSE-NEXT: movaps %xmm0, 16(%r9)3604; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload3605; SSE-NEXT: movaps %xmm0, (%r9)3606; SSE-NEXT: addq $920, %rsp # imm = 0x3983607; SSE-NEXT: retq3608;3609; AVX-LABEL: load_i64_stride5_vf32:3610; AVX: # %bb.0:3611; AVX-NEXT: subq $1368, %rsp # imm = 0x5583612; AVX-NEXT: vmovapd 896(%rdi), %ymm33613; AVX-NEXT: vmovupd %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3614; AVX-NEXT: vmovapd 864(%rdi), %ymm73615; AVX-NEXT: vmovapd 576(%rdi), %ymm43616; AVX-NEXT: vmovupd %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3617; AVX-NEXT: vmovapd 544(%rdi), %ymm23618; AVX-NEXT: vmovapd 256(%rdi), %ymm03619; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3620; AVX-NEXT: vmovapd 224(%rdi), %ymm53621; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm5[0,1,2],ymm0[3]3622; AVX-NEXT: vmovapd 192(%rdi), %xmm13623; AVX-NEXT: vmovapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3624; AVX-NEXT: vmovapd 160(%rdi), %xmm93625; AVX-NEXT: vblendpd {{.*#+}} xmm1 = xmm9[0],xmm1[1]3626; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3]3627; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3628; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm2[0,1,2],ymm4[3]3629; AVX-NEXT: vmovapd 512(%rdi), %xmm13630; AVX-NEXT: vmovapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3631; AVX-NEXT: vmovapd 480(%rdi), %xmm63632; AVX-NEXT: vblendpd {{.*#+}} xmm1 = xmm6[0],xmm1[1]3633; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3]3634; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3635; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm7[0,1,2],ymm3[3]3636; AVX-NEXT: vmovapd 832(%rdi), %xmm13637; AVX-NEXT: vmovapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3638; AVX-NEXT: vmovapd 800(%rdi), %xmm103639; AVX-NEXT: vblendpd {{.*#+}} xmm1 = xmm10[0],xmm1[1]3640; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3]3641; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3642; AVX-NEXT: vmovapd 1216(%rdi), %ymm03643; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3644; AVX-NEXT: vmovapd 1184(%rdi), %ymm83645; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm8[0,1,2],ymm0[3]3646; AVX-NEXT: vmovapd 1152(%rdi), %xmm13647; AVX-NEXT: vmovapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3648; AVX-NEXT: vmovapd 1120(%rdi), %xmm113649; AVX-NEXT: vblendpd {{.*#+}} xmm1 = xmm11[0],xmm1[1]3650; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3]3651; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3652; AVX-NEXT: vmovapd 96(%rdi), %ymm03653; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3654; AVX-NEXT: vmovapd 64(%rdi), %ymm43655; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm4[0,1,2],ymm0[3]3656; AVX-NEXT: vmovapd (%rdi), %xmm133657; AVX-NEXT: vmovapd 32(%rdi), %xmm13658; AVX-NEXT: vmovapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3659; AVX-NEXT: vblendpd {{.*#+}} xmm1 = xmm13[0],xmm1[1]3660; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3]3661; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3662; AVX-NEXT: vmovaps 416(%rdi), %ymm03663; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3664; AVX-NEXT: vmovaps 384(%rdi), %ymm13665; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3666; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]3667; AVX-NEXT: vmovaps 352(%rdi), %xmm13668; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3669; AVX-NEXT: vmovaps 320(%rdi), %xmm123670; AVX-NEXT: vblendps {{.*#+}} xmm1 = xmm12[0,1],xmm1[2,3]3671; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]3672; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3673; AVX-NEXT: vmovaps 736(%rdi), %ymm03674; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3675; AVX-NEXT: vmovaps 704(%rdi), %ymm13676; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3677; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]3678; AVX-NEXT: vmovaps 672(%rdi), %xmm13679; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3680; AVX-NEXT: vmovaps 640(%rdi), %xmm143681; AVX-NEXT: vblendps {{.*#+}} xmm1 = xmm14[0,1],xmm1[2,3]3682; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]3683; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3684; AVX-NEXT: vmovaps 1056(%rdi), %ymm03685; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3686; AVX-NEXT: vmovaps 1024(%rdi), %ymm13687; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3688; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]3689; AVX-NEXT: vmovaps 992(%rdi), %xmm13690; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3691; AVX-NEXT: vmovaps 960(%rdi), %xmm153692; AVX-NEXT: vblendps {{.*#+}} xmm1 = xmm15[0,1],xmm1[2,3]3693; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]3694; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3695; AVX-NEXT: vmovaps 288(%rdi), %xmm13696; AVX-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm03697; AVX-NEXT: vshufpd {{.*#+}} ymm0 = ymm5[0],ymm0[1],ymm5[3],ymm0[2]3698; AVX-NEXT: vmovdqa 208(%rdi), %xmm53699; AVX-NEXT: vpalignr {{.*#+}} xmm3 = xmm9[8,9,10,11,12,13,14,15],xmm5[0,1,2,3,4,5,6,7]3700; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm3[0,1],ymm0[2,3]3701; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3702; AVX-NEXT: vmovaps 608(%rdi), %xmm93703; AVX-NEXT: vinsertf128 $1, %xmm9, %ymm0, %ymm03704; AVX-NEXT: vshufpd {{.*#+}} ymm0 = ymm2[0],ymm0[1],ymm2[3],ymm0[2]3705; AVX-NEXT: vmovdqa 528(%rdi), %xmm23706; AVX-NEXT: vpalignr {{.*#+}} xmm3 = xmm6[8,9,10,11,12,13,14,15],xmm2[0,1,2,3,4,5,6,7]3707; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm3[0,1],ymm0[2,3]3708; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3709; AVX-NEXT: vmovaps 928(%rdi), %xmm03710; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3711; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm03712; AVX-NEXT: vshufpd {{.*#+}} ymm0 = ymm7[0],ymm0[1],ymm7[3],ymm0[2]3713; AVX-NEXT: vmovdqa 848(%rdi), %xmm33714; AVX-NEXT: vpalignr {{.*#+}} xmm6 = xmm10[8,9,10,11,12,13,14,15],xmm3[0,1,2,3,4,5,6,7]3715; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm6[0,1],ymm0[2,3]3716; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3717; AVX-NEXT: vmovaps 1248(%rdi), %xmm03718; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3719; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm63720; AVX-NEXT: vshufpd {{.*#+}} ymm6 = ymm8[0],ymm6[1],ymm8[3],ymm6[2]3721; AVX-NEXT: vmovdqa 1168(%rdi), %xmm73722; AVX-NEXT: vpalignr {{.*#+}} xmm8 = xmm11[8,9,10,11,12,13,14,15],xmm7[0,1,2,3,4,5,6,7]3723; AVX-NEXT: vblendpd {{.*#+}} ymm6 = ymm8[0,1],ymm6[2,3]3724; AVX-NEXT: vmovupd %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3725; AVX-NEXT: vmovaps 128(%rdi), %xmm83726; AVX-NEXT: vinsertf128 $1, %xmm8, %ymm0, %ymm63727; AVX-NEXT: vshufpd {{.*#+}} ymm4 = ymm4[0],ymm6[1],ymm4[3],ymm6[2]3728; AVX-NEXT: vmovdqa 48(%rdi), %xmm03729; AVX-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3730; AVX-NEXT: vpalignr {{.*#+}} xmm6 = xmm13[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]3731; AVX-NEXT: vblendpd {{.*#+}} ymm4 = ymm6[0,1],ymm4[2,3]3732; AVX-NEXT: vmovupd %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3733; AVX-NEXT: vmovaps 448(%rdi), %xmm103734; AVX-NEXT: vinsertf128 $1, %xmm10, %ymm0, %ymm43735; AVX-NEXT: vmovupd {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload3736; AVX-NEXT: vshufpd {{.*#+}} ymm4 = ymm6[0],ymm4[1],ymm6[3],ymm4[2]3737; AVX-NEXT: vmovdqa 368(%rdi), %xmm113738; AVX-NEXT: vpalignr {{.*#+}} xmm6 = xmm12[8,9,10,11,12,13,14,15],xmm11[0,1,2,3,4,5,6,7]3739; AVX-NEXT: vblendpd {{.*#+}} ymm4 = ymm6[0,1],ymm4[2,3]3740; AVX-NEXT: vmovupd %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3741; AVX-NEXT: vmovaps 768(%rdi), %xmm123742; AVX-NEXT: vinsertf128 $1, %xmm12, %ymm0, %ymm43743; AVX-NEXT: vmovupd {{[-0-9]+}}(%r{{[sb]}}p), %ymm6 # 32-byte Reload3744; AVX-NEXT: vshufpd {{.*#+}} ymm4 = ymm6[0],ymm4[1],ymm6[3],ymm4[2]3745; AVX-NEXT: vmovdqa 688(%rdi), %xmm133746; AVX-NEXT: vpalignr {{.*#+}} xmm6 = xmm14[8,9,10,11,12,13,14,15],xmm13[0,1,2,3,4,5,6,7]3747; AVX-NEXT: vblendpd {{.*#+}} ymm4 = ymm6[0,1],ymm4[2,3]3748; AVX-NEXT: vmovupd %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3749; AVX-NEXT: vmovaps 1088(%rdi), %xmm143750; AVX-NEXT: vinsertf128 $1, %xmm14, %ymm0, %ymm43751; AVX-NEXT: vmovupd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3752; AVX-NEXT: vshufpd {{.*#+}} ymm4 = ymm0[0],ymm4[1],ymm0[3],ymm4[2]3753; AVX-NEXT: vmovdqa 1008(%rdi), %xmm03754; AVX-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3755; AVX-NEXT: vpalignr {{.*#+}} xmm6 = xmm15[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]3756; AVX-NEXT: vblendpd {{.*#+}} ymm4 = ymm6[0,1],ymm4[2,3]3757; AVX-NEXT: vmovupd %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3758; AVX-NEXT: vmovaps 256(%rdi), %xmm43759; AVX-NEXT: vblendps {{.*#+}} xmm1 = xmm4[0,1],xmm1[2,3]3760; AVX-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm13761; AVX-NEXT: vmovdqa 176(%rdi), %xmm153762; AVX-NEXT: vpblendw {{.*#+}} xmm5 = xmm15[0,1,2,3],xmm5[4,5,6,7]3763; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm5[0,1,2,3],ymm1[4,5,6,7]3764; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3765; AVX-NEXT: vmovaps 576(%rdi), %xmm53766; AVX-NEXT: vblendps {{.*#+}} xmm1 = xmm5[0,1],xmm9[2,3]3767; AVX-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm13768; AVX-NEXT: vmovdqa 496(%rdi), %xmm03769; AVX-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3770; AVX-NEXT: vpblendw {{.*#+}} xmm2 = xmm0[0,1,2,3],xmm2[4,5,6,7]3771; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]3772; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3773; AVX-NEXT: vmovaps 896(%rdi), %xmm63774; AVX-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %xmm6, %xmm1 # 16-byte Folded Reload3775; AVX-NEXT: # xmm1 = xmm6[0,1],mem[2,3]3776; AVX-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm13777; AVX-NEXT: vmovdqa 816(%rdi), %xmm03778; AVX-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3779; AVX-NEXT: vpblendw {{.*#+}} xmm2 = xmm0[0,1,2,3],xmm3[4,5,6,7]3780; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]3781; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3782; AVX-NEXT: vmovaps 1216(%rdi), %xmm33783; AVX-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm0 # 16-byte Folded Reload3784; AVX-NEXT: # xmm0 = xmm3[0,1],mem[2,3]3785; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm03786; AVX-NEXT: vmovdqa 1136(%rdi), %xmm93787; AVX-NEXT: vpblendw {{.*#+}} xmm1 = xmm9[0,1,2,3],xmm7[4,5,6,7]3788; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]3789; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3790; AVX-NEXT: vmovaps 96(%rdi), %xmm03791; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3792; AVX-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm8[2,3]3793; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm03794; AVX-NEXT: vmovaps 16(%rdi), %xmm13795; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3796; AVX-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload3797; AVX-NEXT: # xmm1 = xmm1[0,1],mem[2,3]3798; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]3799; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3800; AVX-NEXT: vmovaps 416(%rdi), %xmm03801; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3802; AVX-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm10[2,3]3803; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm03804; AVX-NEXT: vmovdqa 336(%rdi), %xmm13805; AVX-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill3806; AVX-NEXT: vpblendw {{.*#+}} xmm1 = xmm1[0,1,2,3],xmm11[4,5,6,7]3807; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]3808; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3809; AVX-NEXT: vmovaps 736(%rdi), %xmm23810; AVX-NEXT: vblendps {{.*#+}} xmm0 = xmm2[0,1],xmm12[2,3]3811; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm03812; AVX-NEXT: vmovdqa 656(%rdi), %xmm123813; AVX-NEXT: vpblendw {{.*#+}} xmm1 = xmm12[0,1,2,3],xmm13[4,5,6,7]3814; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]3815; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3816; AVX-NEXT: vmovaps 1056(%rdi), %xmm03817; AVX-NEXT: vblendps {{.*#+}} xmm1 = xmm0[0,1],xmm14[2,3]3818; AVX-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm73819; AVX-NEXT: vmovaps 976(%rdi), %xmm13820; AVX-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm8 # 16-byte Folded Reload3821; AVX-NEXT: # xmm8 = xmm1[0,1],mem[2,3]3822; AVX-NEXT: vblendps {{.*#+}} ymm7 = ymm8[0,1,2,3],ymm7[4,5,6,7]3823; AVX-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3824; AVX-NEXT: vinsertf128 $1, %xmm4, %ymm0, %ymm43825; AVX-NEXT: vmovapd 288(%rdi), %ymm133826; AVX-NEXT: vshufpd {{.*#+}} ymm4 = ymm4[0],ymm13[1],ymm4[3],ymm13[2]3827; AVX-NEXT: vmovdqa 224(%rdi), %xmm103828; AVX-NEXT: vpalignr {{.*#+}} xmm8 = xmm15[8,9,10,11,12,13,14,15],xmm10[0,1,2,3,4,5,6,7]3829; AVX-NEXT: vblendpd {{.*#+}} ymm4 = ymm8[0,1],ymm4[2,3]3830; AVX-NEXT: vmovupd %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3831; AVX-NEXT: vinsertf128 $1, %xmm5, %ymm0, %ymm43832; AVX-NEXT: vmovapd 608(%rdi), %ymm143833; AVX-NEXT: vshufpd {{.*#+}} ymm5 = ymm4[0],ymm14[1],ymm4[3],ymm14[2]3834; AVX-NEXT: vmovdqa 544(%rdi), %xmm113835; AVX-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm11, %xmm8 # 16-byte Folded Reload3836; AVX-NEXT: # xmm8 = mem[8,9,10,11,12,13,14,15],xmm11[0,1,2,3,4,5,6,7]3837; AVX-NEXT: vblendpd {{.*#+}} ymm4 = ymm8[0,1],ymm5[2,3]3838; AVX-NEXT: vmovupd %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3839; AVX-NEXT: vinsertf128 $1, %xmm6, %ymm0, %ymm53840; AVX-NEXT: vmovapd 928(%rdi), %ymm43841; AVX-NEXT: vmovupd %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3842; AVX-NEXT: vshufpd {{.*#+}} ymm5 = ymm5[0],ymm4[1],ymm5[3],ymm4[2]3843; AVX-NEXT: vmovdqa 864(%rdi), %xmm83844; AVX-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm8, %xmm6 # 16-byte Folded Reload3845; AVX-NEXT: # xmm6 = mem[8,9,10,11,12,13,14,15],xmm8[0,1,2,3,4,5,6,7]3846; AVX-NEXT: vblendpd {{.*#+}} ymm4 = ymm6[0,1],ymm5[2,3]3847; AVX-NEXT: vmovupd %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3848; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm33849; AVX-NEXT: vmovapd 1248(%rdi), %ymm43850; AVX-NEXT: vmovupd %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3851; AVX-NEXT: vshufpd {{.*#+}} ymm3 = ymm3[0],ymm4[1],ymm3[3],ymm4[2]3852; AVX-NEXT: vmovdqa 1184(%rdi), %xmm153853; AVX-NEXT: vpalignr {{.*#+}} xmm6 = xmm9[8,9,10,11,12,13,14,15],xmm15[0,1,2,3,4,5,6,7]3854; AVX-NEXT: vblendpd {{.*#+}} ymm3 = ymm6[0,1],ymm3[2,3]3855; AVX-NEXT: vmovupd %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3856; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm03857; AVX-NEXT: vmovapd 1088(%rdi), %ymm73858; AVX-NEXT: vshufpd {{.*#+}} ymm0 = ymm0[0],ymm7[1],ymm0[3],ymm7[2]3859; AVX-NEXT: vmovdqa 1024(%rdi), %xmm93860; AVX-NEXT: vpalignr {{.*#+}} xmm1 = xmm1[8,9,10,11,12,13,14,15],xmm9[0,1,2,3,4,5,6,7]3861; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3]3862; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3863; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm13864; AVX-NEXT: vmovapd 768(%rdi), %ymm53865; AVX-NEXT: vshufpd {{.*#+}} ymm1 = ymm1[0],ymm5[1],ymm1[3],ymm5[2]3866; AVX-NEXT: vmovdqa 704(%rdi), %xmm43867; AVX-NEXT: vpalignr {{.*#+}} xmm6 = xmm12[8,9,10,11,12,13,14,15],xmm4[0,1,2,3,4,5,6,7]3868; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm6[0,1],ymm1[2,3]3869; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3870; AVX-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm1 # 16-byte Folded Reload3871; AVX-NEXT: vmovapd 448(%rdi), %ymm33872; AVX-NEXT: vshufpd {{.*#+}} ymm1 = ymm1[0],ymm3[1],ymm1[3],ymm3[2]3873; AVX-NEXT: vmovdqa 384(%rdi), %xmm23874; AVX-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm6 # 16-byte Folded Reload3875; AVX-NEXT: # xmm6 = mem[8,9,10,11,12,13,14,15],xmm2[0,1,2,3,4,5,6,7]3876; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm6[0,1],ymm1[2,3]3877; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3878; AVX-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm1 # 16-byte Folded Reload3879; AVX-NEXT: vmovapd 128(%rdi), %ymm123880; AVX-NEXT: vshufpd {{.*#+}} ymm1 = ymm1[0],ymm12[1],ymm1[3],ymm12[2]3881; AVX-NEXT: vmovdqa 64(%rdi), %xmm03882; AVX-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm6 # 16-byte Folded Reload3883; AVX-NEXT: # xmm6 = mem[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]3884; AVX-NEXT: vblendpd {{.*#+}} ymm6 = ymm6[0,1],ymm1[2,3]3885; AVX-NEXT: vblendpd $7, {{[-0-9]+}}(%r{{[sb]}}p), %ymm12, %ymm1 # 32-byte Folded Reload3886; AVX-NEXT: # ymm1 = mem[0,1,2],ymm12[3]3887; AVX-NEXT: vpblendw $15, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload3888; AVX-NEXT: # xmm0 = mem[0,1,2,3],xmm0[4,5,6,7]3889; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3]3890; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill3891; AVX-NEXT: vblendpd $7, {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm0 # 32-byte Folded Reload3892; AVX-NEXT: # ymm0 = mem[0,1,2],ymm13[3]3893; AVX-NEXT: vpblendw $15, {{[-0-9]+}}(%r{{[sb]}}p), %xmm10, %xmm1 # 16-byte Folded Reload3894; AVX-NEXT: # xmm1 = mem[0,1,2,3],xmm10[4,5,6,7]3895; AVX-NEXT: vblendpd {{.*#+}} ymm12 = ymm1[0,1],ymm0[2,3]3896; AVX-NEXT: vblendpd $7, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3, %ymm0 # 32-byte Folded Reload3897; AVX-NEXT: # ymm0 = mem[0,1,2],ymm3[3]3898; AVX-NEXT: vpblendw $15, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm1 # 16-byte Folded Reload3899; AVX-NEXT: # xmm1 = mem[0,1,2,3],xmm2[4,5,6,7]3900; AVX-NEXT: vblendpd {{.*#+}} ymm10 = ymm1[0,1],ymm0[2,3]3901; AVX-NEXT: vblendpd $7, {{[-0-9]+}}(%r{{[sb]}}p), %ymm14, %ymm0 # 32-byte Folded Reload3902; AVX-NEXT: # ymm0 = mem[0,1,2],ymm14[3]3903; AVX-NEXT: vpblendw $15, {{[-0-9]+}}(%r{{[sb]}}p), %xmm11, %xmm1 # 16-byte Folded Reload3904; AVX-NEXT: # xmm1 = mem[0,1,2,3],xmm11[4,5,6,7]3905; AVX-NEXT: vblendpd {{.*#+}} ymm1 = ymm1[0,1],ymm0[2,3]3906; AVX-NEXT: vblendpd $7, {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm0 # 32-byte Folded Reload3907; AVX-NEXT: # ymm0 = mem[0,1,2],ymm5[3]3908; AVX-NEXT: vpblendw $15, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4, %xmm2 # 16-byte Folded Reload3909; AVX-NEXT: # xmm2 = mem[0,1,2,3],xmm4[4,5,6,7]3910; AVX-NEXT: vblendpd {{.*#+}} ymm3 = ymm2[0,1],ymm0[2,3]3911; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3912; AVX-NEXT: vblendps $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload3913; AVX-NEXT: # ymm0 = mem[0,1,2,3,4,5],ymm0[6,7]3914; AVX-NEXT: vpblendw $15, {{[-0-9]+}}(%r{{[sb]}}p), %xmm8, %xmm2 # 16-byte Folded Reload3915; AVX-NEXT: # xmm2 = mem[0,1,2,3],xmm8[4,5,6,7]3916; AVX-NEXT: vblendps {{.*#+}} ymm11 = ymm2[0,1,2,3],ymm0[4,5,6,7]3917; AVX-NEXT: vblendpd $7, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm0 # 32-byte Folded Reload3918; AVX-NEXT: # ymm0 = mem[0,1,2],ymm7[3]3919; AVX-NEXT: vpblendw $15, {{[-0-9]+}}(%r{{[sb]}}p), %xmm9, %xmm2 # 16-byte Folded Reload3920; AVX-NEXT: # xmm2 = mem[0,1,2,3],xmm9[4,5,6,7]3921; AVX-NEXT: vblendpd {{.*#+}} ymm2 = ymm2[0,1],ymm0[2,3]3922; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3923; AVX-NEXT: vblendps $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload3924; AVX-NEXT: # ymm0 = mem[0,1,2,3,4,5],ymm0[6,7]3925; AVX-NEXT: vpblendw $15, {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm4 # 16-byte Folded Reload3926; AVX-NEXT: # xmm4 = mem[0,1,2,3],xmm15[4,5,6,7]3927; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm4[0,1,2,3],ymm0[4,5,6,7]3928; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload3929; AVX-NEXT: vmovaps %ymm4, 192(%rsi)3930; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload3931; AVX-NEXT: vmovaps %ymm4, 128(%rsi)3932; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload3933; AVX-NEXT: vmovaps %ymm4, 64(%rsi)3934; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload3935; AVX-NEXT: vmovaps %ymm4, (%rsi)3936; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload3937; AVX-NEXT: vmovaps %ymm4, 224(%rsi)3938; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload3939; AVX-NEXT: vmovaps %ymm4, 160(%rsi)3940; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload3941; AVX-NEXT: vmovaps %ymm4, 96(%rsi)3942; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload3943; AVX-NEXT: vmovaps %ymm4, 32(%rsi)3944; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload3945; AVX-NEXT: vmovaps %ymm4, 192(%rdx)3946; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload3947; AVX-NEXT: vmovaps %ymm4, 128(%rdx)3948; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload3949; AVX-NEXT: vmovaps %ymm4, 64(%rdx)3950; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload3951; AVX-NEXT: vmovaps %ymm4, (%rdx)3952; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload3953; AVX-NEXT: vmovaps %ymm4, 224(%rdx)3954; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload3955; AVX-NEXT: vmovaps %ymm4, 160(%rdx)3956; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload3957; AVX-NEXT: vmovaps %ymm4, 96(%rdx)3958; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload3959; AVX-NEXT: vmovaps %ymm4, 32(%rdx)3960; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload3961; AVX-NEXT: vmovaps %ymm4, 192(%rcx)3962; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload3963; AVX-NEXT: vmovaps %ymm4, 128(%rcx)3964; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload3965; AVX-NEXT: vmovaps %ymm4, 64(%rcx)3966; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload3967; AVX-NEXT: vmovaps %ymm4, (%rcx)3968; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload3969; AVX-NEXT: vmovaps %ymm4, 224(%rcx)3970; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload3971; AVX-NEXT: vmovaps %ymm4, 160(%rcx)3972; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload3973; AVX-NEXT: vmovaps %ymm4, 96(%rcx)3974; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload3975; AVX-NEXT: vmovaps %ymm4, 32(%rcx)3976; AVX-NEXT: vmovapd %ymm6, (%r8)3977; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload3978; AVX-NEXT: vmovaps %ymm4, 64(%r8)3979; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload3980; AVX-NEXT: vmovaps %ymm4, 128(%r8)3981; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload3982; AVX-NEXT: vmovaps %ymm4, 192(%r8)3983; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload3984; AVX-NEXT: vmovaps %ymm4, 224(%r8)3985; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload3986; AVX-NEXT: vmovaps %ymm4, 160(%r8)3987; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload3988; AVX-NEXT: vmovaps %ymm4, 96(%r8)3989; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm4 # 32-byte Reload3990; AVX-NEXT: vmovaps %ymm4, 32(%r8)3991; AVX-NEXT: vmovaps %ymm0, 224(%r9)3992; AVX-NEXT: vmovapd %ymm2, 192(%r9)3993; AVX-NEXT: vmovaps %ymm11, 160(%r9)3994; AVX-NEXT: vmovapd %ymm3, 128(%r9)3995; AVX-NEXT: vmovapd %ymm1, 96(%r9)3996; AVX-NEXT: vmovapd %ymm10, 64(%r9)3997; AVX-NEXT: vmovapd %ymm12, 32(%r9)3998; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload3999; AVX-NEXT: vmovaps %ymm0, (%r9)4000; AVX-NEXT: addq $1368, %rsp # imm = 0x5584001; AVX-NEXT: vzeroupper4002; AVX-NEXT: retq4003;4004; AVX2-LABEL: load_i64_stride5_vf32:4005; AVX2: # %bb.0:4006; AVX2-NEXT: subq $1464, %rsp # imm = 0x5B84007; AVX2-NEXT: vmovdqa 896(%rdi), %ymm14008; AVX2-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4009; AVX2-NEXT: vmovdqa 864(%rdi), %ymm44010; AVX2-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4011; AVX2-NEXT: vmovdqa 576(%rdi), %ymm34012; AVX2-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4013; AVX2-NEXT: vmovdqa 544(%rdi), %ymm64014; AVX2-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4015; AVX2-NEXT: vmovdqa 256(%rdi), %ymm04016; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4017; AVX2-NEXT: vmovdqa 224(%rdi), %ymm54018; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm5[0,1,2,3,4,5],ymm0[6,7]4019; AVX2-NEXT: vmovdqa 192(%rdi), %xmm24020; AVX2-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4021; AVX2-NEXT: vmovdqa 160(%rdi), %xmm144022; AVX2-NEXT: vpblendd {{.*#+}} xmm2 = xmm14[0,1],xmm2[2,3]4023; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]4024; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4025; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm6[0,1,2,3,4,5],ymm3[6,7]4026; AVX2-NEXT: vmovdqa 512(%rdi), %xmm24027; AVX2-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4028; AVX2-NEXT: vmovdqa 480(%rdi), %xmm104029; AVX2-NEXT: vpblendd {{.*#+}} xmm2 = xmm10[0,1],xmm2[2,3]4030; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]4031; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4032; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm4[0,1,2,3,4,5],ymm1[6,7]4033; AVX2-NEXT: vmovdqa 832(%rdi), %xmm14034; AVX2-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4035; AVX2-NEXT: vmovdqa 800(%rdi), %xmm44036; AVX2-NEXT: vpblendd {{.*#+}} xmm2 = xmm4[0,1],xmm1[2,3]4037; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]4038; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4039; AVX2-NEXT: vmovdqa 1216(%rdi), %ymm04040; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4041; AVX2-NEXT: vmovdqa 1184(%rdi), %ymm24042; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3,4,5],ymm0[6,7]4043; AVX2-NEXT: vmovdqa 1152(%rdi), %xmm14044; AVX2-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4045; AVX2-NEXT: vmovdqa 1120(%rdi), %xmm124046; AVX2-NEXT: vpblendd {{.*#+}} xmm3 = xmm12[0,1],xmm1[2,3]4047; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm3[0,1,2,3],ymm0[4,5,6,7]4048; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4049; AVX2-NEXT: vmovdqa 96(%rdi), %ymm04050; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4051; AVX2-NEXT: vmovdqa 64(%rdi), %ymm14052; AVX2-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4053; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]4054; AVX2-NEXT: vmovdqa (%rdi), %xmm134055; AVX2-NEXT: vmovdqa 32(%rdi), %xmm14056; AVX2-NEXT: vmovdqa %xmm1, (%rsp) # 16-byte Spill4057; AVX2-NEXT: vpblendd {{.*#+}} xmm6 = xmm13[0,1],xmm1[2,3]4058; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm6[0,1,2,3],ymm0[4,5,6,7]4059; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4060; AVX2-NEXT: vmovdqa 416(%rdi), %ymm04061; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4062; AVX2-NEXT: vmovdqa 384(%rdi), %ymm64063; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm6[0,1,2,3,4,5],ymm0[6,7]4064; AVX2-NEXT: vmovdqa 352(%rdi), %xmm14065; AVX2-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4066; AVX2-NEXT: vmovdqa 320(%rdi), %xmm154067; AVX2-NEXT: vpblendd {{.*#+}} xmm7 = xmm15[0,1],xmm1[2,3]4068; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm7[0,1,2,3],ymm0[4,5,6,7]4069; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4070; AVX2-NEXT: vmovdqa 736(%rdi), %ymm04071; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4072; AVX2-NEXT: vmovdqa 704(%rdi), %ymm74073; AVX2-NEXT: vpblendd {{.*#+}} ymm8 = ymm7[0,1,2,3,4,5],ymm0[6,7]4074; AVX2-NEXT: vmovdqa 672(%rdi), %xmm04075; AVX2-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4076; AVX2-NEXT: vmovdqa 640(%rdi), %xmm34077; AVX2-NEXT: vpblendd {{.*#+}} xmm9 = xmm3[0,1],xmm0[2,3]4078; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm9[0,1,2,3],ymm8[4,5,6,7]4079; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4080; AVX2-NEXT: vmovdqa 1056(%rdi), %ymm04081; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4082; AVX2-NEXT: vmovdqa 1024(%rdi), %ymm84083; AVX2-NEXT: vpblendd {{.*#+}} ymm1 = ymm8[0,1,2,3,4,5],ymm0[6,7]4084; AVX2-NEXT: vmovdqa 992(%rdi), %xmm04085; AVX2-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4086; AVX2-NEXT: vmovdqa 960(%rdi), %xmm94087; AVX2-NEXT: vpblendd {{.*#+}} xmm11 = xmm9[0,1],xmm0[2,3]4088; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm11[0,1,2,3],ymm1[4,5,6,7]4089; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4090; AVX2-NEXT: vmovdqa 208(%rdi), %xmm04091; AVX2-NEXT: vpalignr {{.*#+}} xmm0 = xmm14[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]4092; AVX2-NEXT: vmovdqa 288(%rdi), %ymm144093; AVX2-NEXT: vpalignr {{.*#+}} ymm5 = ymm5[8,9,10,11,12,13,14,15],ymm14[0,1,2,3,4,5,6,7],ymm5[24,25,26,27,28,29,30,31],ymm14[16,17,18,19,20,21,22,23]4094; AVX2-NEXT: vmovdqu %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4095; AVX2-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,1,2,1]4096; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm5[4,5,6,7]4097; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4098; AVX2-NEXT: vmovdqa 528(%rdi), %xmm04099; AVX2-NEXT: vpalignr {{.*#+}} xmm0 = xmm10[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]4100; AVX2-NEXT: vmovdqa 608(%rdi), %ymm114101; AVX2-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm11, %ymm5 # 32-byte Folded Reload4102; AVX2-NEXT: # ymm5 = mem[8,9,10,11,12,13,14,15],ymm11[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm11[16,17,18,19,20,21,22,23]4103; AVX2-NEXT: vmovdqu %ymm11, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4104; AVX2-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,1,2,1]4105; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm5[4,5,6,7]4106; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4107; AVX2-NEXT: vmovdqa 848(%rdi), %xmm04108; AVX2-NEXT: vpalignr {{.*#+}} xmm0 = xmm4[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]4109; AVX2-NEXT: vmovdqa 928(%rdi), %ymm104110; AVX2-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm10, %ymm4 # 32-byte Folded Reload4111; AVX2-NEXT: # ymm4 = mem[8,9,10,11,12,13,14,15],ymm10[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm10[16,17,18,19,20,21,22,23]4112; AVX2-NEXT: vmovdqu %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4113; AVX2-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,1,2,1]4114; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm4[4,5,6,7]4115; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4116; AVX2-NEXT: vmovdqa 1168(%rdi), %xmm04117; AVX2-NEXT: vpalignr {{.*#+}} xmm0 = xmm12[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]4118; AVX2-NEXT: vmovdqa 1248(%rdi), %ymm54119; AVX2-NEXT: vpalignr {{.*#+}} ymm2 = ymm2[8,9,10,11,12,13,14,15],ymm5[0,1,2,3,4,5,6,7],ymm2[24,25,26,27,28,29,30,31],ymm5[16,17,18,19,20,21,22,23]4120; AVX2-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4121; AVX2-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,1,2,1]4122; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]4123; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4124; AVX2-NEXT: vmovdqa 48(%rdi), %xmm04125; AVX2-NEXT: vpalignr {{.*#+}} xmm0 = xmm13[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]4126; AVX2-NEXT: vmovdqa 128(%rdi), %ymm124127; AVX2-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm12, %ymm2 # 32-byte Folded Reload4128; AVX2-NEXT: # ymm2 = mem[8,9,10,11,12,13,14,15],ymm12[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm12[16,17,18,19,20,21,22,23]4129; AVX2-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,1,2,1]4130; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]4131; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4132; AVX2-NEXT: vmovdqa 368(%rdi), %xmm04133; AVX2-NEXT: vpalignr {{.*#+}} xmm0 = xmm15[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]4134; AVX2-NEXT: vmovdqa 448(%rdi), %ymm14135; AVX2-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4136; AVX2-NEXT: vpalignr {{.*#+}} ymm2 = ymm6[8,9,10,11,12,13,14,15],ymm1[0,1,2,3,4,5,6,7],ymm6[24,25,26,27,28,29,30,31],ymm1[16,17,18,19,20,21,22,23]4137; AVX2-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,1,2,1]4138; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]4139; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4140; AVX2-NEXT: vmovdqa 688(%rdi), %xmm04141; AVX2-NEXT: vpalignr {{.*#+}} xmm0 = xmm3[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]4142; AVX2-NEXT: vmovdqa 768(%rdi), %ymm154143; AVX2-NEXT: vpalignr {{.*#+}} ymm1 = ymm7[8,9,10,11,12,13,14,15],ymm15[0,1,2,3,4,5,6,7],ymm7[24,25,26,27,28,29,30,31],ymm15[16,17,18,19,20,21,22,23]4144; AVX2-NEXT: vmovdqu %ymm15, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4145; AVX2-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,1,2,1]4146; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]4147; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4148; AVX2-NEXT: vmovdqa 1008(%rdi), %xmm04149; AVX2-NEXT: vpalignr {{.*#+}} xmm0 = xmm9[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]4150; AVX2-NEXT: vmovdqa 1088(%rdi), %ymm24151; AVX2-NEXT: vpalignr {{.*#+}} ymm1 = ymm8[8,9,10,11,12,13,14,15],ymm2[0,1,2,3,4,5,6,7],ymm8[24,25,26,27,28,29,30,31],ymm2[16,17,18,19,20,21,22,23]4152; AVX2-NEXT: vmovdqa %ymm2, %ymm84153; AVX2-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4154; AVX2-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,1,2,1]4155; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]4156; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4157; AVX2-NEXT: vmovaps 160(%rdi), %ymm04158; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]4159; AVX2-NEXT: vmovaps 256(%rdi), %xmm14160; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]4161; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm14162; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]4163; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4164; AVX2-NEXT: vmovaps 480(%rdi), %ymm04165; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]4166; AVX2-NEXT: vmovaps 576(%rdi), %xmm14167; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]4168; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm14169; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]4170; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4171; AVX2-NEXT: vmovaps 800(%rdi), %ymm04172; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]4173; AVX2-NEXT: vmovaps 896(%rdi), %xmm14174; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]4175; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm14176; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]4177; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4178; AVX2-NEXT: vmovaps 1120(%rdi), %ymm04179; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]4180; AVX2-NEXT: vmovaps 1216(%rdi), %xmm14181; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]4182; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm14183; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]4184; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4185; AVX2-NEXT: vmovaps (%rdi), %ymm04186; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]4187; AVX2-NEXT: vmovaps 96(%rdi), %xmm14188; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]4189; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm14190; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]4191; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4192; AVX2-NEXT: vmovaps 320(%rdi), %ymm04193; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]4194; AVX2-NEXT: vmovaps 416(%rdi), %xmm14195; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]4196; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm14197; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]4198; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4199; AVX2-NEXT: vmovaps 640(%rdi), %ymm04200; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]4201; AVX2-NEXT: vmovaps 736(%rdi), %xmm14202; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]4203; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm14204; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]4205; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4206; AVX2-NEXT: vmovaps 960(%rdi), %ymm04207; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]4208; AVX2-NEXT: vmovaps 1056(%rdi), %xmm14209; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]4210; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm14211; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]4212; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4213; AVX2-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm14, %ymm0 # 32-byte Folded Reload4214; AVX2-NEXT: # ymm0 = mem[8,9,10,11,12,13,14,15],ymm14[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm14[16,17,18,19,20,21,22,23]4215; AVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,3]4216; AVX2-NEXT: vmovdqa 224(%rdi), %xmm44217; AVX2-NEXT: vpalignr {{.*#+}} xmm1 = mem[8,9,10,11,12,13,14,15],xmm4[0,1,2,3,4,5,6,7]4218; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]4219; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4220; AVX2-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm11, %ymm0 # 32-byte Folded Reload4221; AVX2-NEXT: # ymm0 = mem[8,9,10,11,12,13,14,15],ymm11[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm11[16,17,18,19,20,21,22,23]4222; AVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,3]4223; AVX2-NEXT: vmovdqa 544(%rdi), %xmm64224; AVX2-NEXT: vpalignr {{.*#+}} xmm2 = mem[8,9,10,11,12,13,14,15],xmm6[0,1,2,3,4,5,6,7]4225; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]4226; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4227; AVX2-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm10, %ymm0 # 32-byte Folded Reload4228; AVX2-NEXT: # ymm0 = mem[8,9,10,11,12,13,14,15],ymm10[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm10[16,17,18,19,20,21,22,23]4229; AVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,3]4230; AVX2-NEXT: vmovdqa 864(%rdi), %xmm14231; AVX2-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4232; AVX2-NEXT: vpalignr {{.*#+}} xmm3 = mem[8,9,10,11,12,13,14,15],xmm1[0,1,2,3,4,5,6,7]4233; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm3[0,1,2,3],ymm0[4,5,6,7]4234; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4235; AVX2-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm0 # 32-byte Folded Reload4236; AVX2-NEXT: # ymm0 = mem[8,9,10,11,12,13,14,15],ymm5[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm5[16,17,18,19,20,21,22,23]4237; AVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,3]4238; AVX2-NEXT: vmovdqa 1184(%rdi), %xmm144239; AVX2-NEXT: vpalignr {{.*#+}} xmm7 = mem[8,9,10,11,12,13,14,15],xmm14[0,1,2,3,4,5,6,7]4240; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm7[0,1,2,3],ymm0[4,5,6,7]4241; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4242; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload4243; AVX2-NEXT: vpalignr {{.*#+}} ymm0 = ymm13[8,9,10,11,12,13,14,15],ymm8[0,1,2,3,4,5,6,7],ymm13[24,25,26,27,28,29,30,31],ymm8[16,17,18,19,20,21,22,23]4244; AVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,3]4245; AVX2-NEXT: vmovdqa 1024(%rdi), %xmm114246; AVX2-NEXT: vpalignr {{.*#+}} xmm7 = mem[8,9,10,11,12,13,14,15],xmm11[0,1,2,3,4,5,6,7]4247; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm7[0,1,2,3],ymm0[4,5,6,7]4248; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4249; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload4250; AVX2-NEXT: vpalignr {{.*#+}} ymm0 = ymm10[8,9,10,11,12,13,14,15],ymm15[0,1,2,3,4,5,6,7],ymm10[24,25,26,27,28,29,30,31],ymm15[16,17,18,19,20,21,22,23]4251; AVX2-NEXT: vpermq {{.*#+}} ymm7 = ymm0[0,1,0,3]4252; AVX2-NEXT: vmovdqa 704(%rdi), %xmm94253; AVX2-NEXT: vpalignr {{.*#+}} xmm8 = mem[8,9,10,11,12,13,14,15],xmm9[0,1,2,3,4,5,6,7]4254; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm8[0,1,2,3],ymm7[4,5,6,7]4255; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4256; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload4257; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload4258; AVX2-NEXT: vpalignr {{.*#+}} ymm7 = ymm3[8,9,10,11,12,13,14,15],ymm2[0,1,2,3,4,5,6,7],ymm3[24,25,26,27,28,29,30,31],ymm2[16,17,18,19,20,21,22,23]4259; AVX2-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,1,0,3]4260; AVX2-NEXT: vmovdqa 384(%rdi), %xmm54261; AVX2-NEXT: vpalignr {{.*#+}} xmm8 = mem[8,9,10,11,12,13,14,15],xmm5[0,1,2,3,4,5,6,7]4262; AVX2-NEXT: vpblendd {{.*#+}} ymm8 = ymm8[0,1,2,3],ymm7[4,5,6,7]4263; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4264; AVX2-NEXT: vpalignr {{.*#+}} ymm7 = ymm1[8,9,10,11,12,13,14,15],ymm12[0,1,2,3,4,5,6,7],ymm1[24,25,26,27,28,29,30,31],ymm12[16,17,18,19,20,21,22,23]4265; AVX2-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,1,0,3]4266; AVX2-NEXT: vmovdqa 64(%rdi), %xmm04267; AVX2-NEXT: vpalignr {{.*#+}} xmm15 = mem[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]4268; AVX2-NEXT: vpblendd {{.*#+}} ymm7 = ymm15[0,1,2,3],ymm7[4,5,6,7]4269; AVX2-NEXT: vpblendd {{.*#+}} ymm12 = ymm1[0,1,2,3,4,5],ymm12[6,7]4270; AVX2-NEXT: vpblendd $3, (%rsp), %xmm0, %xmm0 # 16-byte Folded Reload4271; AVX2-NEXT: # xmm0 = mem[0,1],xmm0[2,3]4272; AVX2-NEXT: vpblendd {{.*#+}} ymm12 = ymm0[0,1,2,3],ymm12[4,5,6,7]4273; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload4274; AVX2-NEXT: vpblendd $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload4275; AVX2-NEXT: # ymm0 = mem[0,1,2,3,4,5],ymm0[6,7]4276; AVX2-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4, %xmm4 # 16-byte Folded Reload4277; AVX2-NEXT: # xmm4 = mem[0,1],xmm4[2,3]4278; AVX2-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm0[4,5,6,7]4279; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm3[0,1,2,3,4,5],ymm2[6,7]4280; AVX2-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm5, %xmm5 # 16-byte Folded Reload4281; AVX2-NEXT: # xmm5 = mem[0,1],xmm5[2,3]4282; AVX2-NEXT: vpblendd {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm0[4,5,6,7]4283; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload4284; AVX2-NEXT: vpblendd $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload4285; AVX2-NEXT: # ymm0 = mem[0,1,2,3,4,5],ymm0[6,7]4286; AVX2-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm6, %xmm6 # 16-byte Folded Reload4287; AVX2-NEXT: # xmm6 = mem[0,1],xmm6[2,3]4288; AVX2-NEXT: vpblendd {{.*#+}} ymm15 = ymm6[0,1,2,3],ymm0[4,5,6,7]4289; AVX2-NEXT: vpblendd $192, {{[-0-9]+}}(%r{{[sb]}}p), %ymm10, %ymm0 # 32-byte Folded Reload4290; AVX2-NEXT: # ymm0 = ymm10[0,1,2,3,4,5],mem[6,7]4291; AVX2-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm9, %xmm1 # 16-byte Folded Reload4292; AVX2-NEXT: # xmm1 = mem[0,1],xmm9[2,3]4293; AVX2-NEXT: vpblendd {{.*#+}} ymm6 = ymm1[0,1,2,3],ymm0[4,5,6,7]4294; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload4295; AVX2-NEXT: vblendps $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload4296; AVX2-NEXT: # ymm0 = mem[0,1,2,3,4,5],ymm0[6,7]4297; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload4298; AVX2-NEXT: vblendps $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload4299; AVX2-NEXT: # xmm1 = mem[0,1],xmm1[2,3]4300; AVX2-NEXT: vblendps {{.*#+}} ymm2 = ymm1[0,1,2,3],ymm0[4,5,6,7]4301; AVX2-NEXT: vpblendd $192, {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm0 # 32-byte Folded Reload4302; AVX2-NEXT: # ymm0 = ymm13[0,1,2,3,4,5],mem[6,7]4303; AVX2-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm11, %xmm1 # 16-byte Folded Reload4304; AVX2-NEXT: # xmm1 = mem[0,1],xmm11[2,3]4305; AVX2-NEXT: vpblendd {{.*#+}} ymm9 = ymm1[0,1,2,3],ymm0[4,5,6,7]4306; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload4307; AVX2-NEXT: vpblendd $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload4308; AVX2-NEXT: # ymm0 = mem[0,1,2,3,4,5],ymm0[6,7]4309; AVX2-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm1 # 16-byte Folded Reload4310; AVX2-NEXT: # xmm1 = mem[0,1],xmm14[2,3]4311; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]4312; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4313; AVX2-NEXT: vmovaps %ymm1, 192(%rsi)4314; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4315; AVX2-NEXT: vmovaps %ymm1, 128(%rsi)4316; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4317; AVX2-NEXT: vmovaps %ymm1, 64(%rsi)4318; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4319; AVX2-NEXT: vmovaps %ymm1, (%rsi)4320; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4321; AVX2-NEXT: vmovaps %ymm1, 224(%rsi)4322; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4323; AVX2-NEXT: vmovaps %ymm1, 160(%rsi)4324; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4325; AVX2-NEXT: vmovaps %ymm1, 96(%rsi)4326; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4327; AVX2-NEXT: vmovaps %ymm1, 32(%rsi)4328; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4329; AVX2-NEXT: vmovaps %ymm1, 192(%rdx)4330; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4331; AVX2-NEXT: vmovaps %ymm1, 128(%rdx)4332; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4333; AVX2-NEXT: vmovaps %ymm1, 64(%rdx)4334; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4335; AVX2-NEXT: vmovaps %ymm1, (%rdx)4336; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4337; AVX2-NEXT: vmovaps %ymm1, 224(%rdx)4338; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4339; AVX2-NEXT: vmovaps %ymm1, 160(%rdx)4340; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4341; AVX2-NEXT: vmovaps %ymm1, 96(%rdx)4342; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4343; AVX2-NEXT: vmovaps %ymm1, 32(%rdx)4344; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4345; AVX2-NEXT: vmovaps %ymm1, 192(%rcx)4346; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4347; AVX2-NEXT: vmovaps %ymm1, 128(%rcx)4348; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4349; AVX2-NEXT: vmovaps %ymm1, 64(%rcx)4350; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4351; AVX2-NEXT: vmovaps %ymm1, (%rcx)4352; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4353; AVX2-NEXT: vmovaps %ymm1, 224(%rcx)4354; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4355; AVX2-NEXT: vmovaps %ymm1, 160(%rcx)4356; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4357; AVX2-NEXT: vmovaps %ymm1, 96(%rcx)4358; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4359; AVX2-NEXT: vmovaps %ymm1, 32(%rcx)4360; AVX2-NEXT: vmovdqa %ymm7, (%r8)4361; AVX2-NEXT: vmovdqa %ymm8, 64(%r8)4362; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4363; AVX2-NEXT: vmovaps %ymm1, 128(%r8)4364; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4365; AVX2-NEXT: vmovaps %ymm1, 192(%r8)4366; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4367; AVX2-NEXT: vmovaps %ymm1, 224(%r8)4368; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4369; AVX2-NEXT: vmovaps %ymm1, 160(%r8)4370; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4371; AVX2-NEXT: vmovaps %ymm1, 96(%r8)4372; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4373; AVX2-NEXT: vmovaps %ymm1, 32(%r8)4374; AVX2-NEXT: vmovdqa %ymm0, 224(%r9)4375; AVX2-NEXT: vmovdqa %ymm9, 192(%r9)4376; AVX2-NEXT: vmovaps %ymm2, 160(%r9)4377; AVX2-NEXT: vmovdqa %ymm6, 128(%r9)4378; AVX2-NEXT: vmovdqa %ymm15, 96(%r9)4379; AVX2-NEXT: vmovdqa %ymm5, 64(%r9)4380; AVX2-NEXT: vmovdqa %ymm4, 32(%r9)4381; AVX2-NEXT: vmovdqa %ymm12, (%r9)4382; AVX2-NEXT: addq $1464, %rsp # imm = 0x5B84383; AVX2-NEXT: vzeroupper4384; AVX2-NEXT: retq4385;4386; AVX2-FP-LABEL: load_i64_stride5_vf32:4387; AVX2-FP: # %bb.0:4388; AVX2-FP-NEXT: subq $1464, %rsp # imm = 0x5B84389; AVX2-FP-NEXT: vmovdqa 896(%rdi), %ymm14390; AVX2-FP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4391; AVX2-FP-NEXT: vmovdqa 864(%rdi), %ymm44392; AVX2-FP-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4393; AVX2-FP-NEXT: vmovdqa 576(%rdi), %ymm34394; AVX2-FP-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4395; AVX2-FP-NEXT: vmovdqa 544(%rdi), %ymm64396; AVX2-FP-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4397; AVX2-FP-NEXT: vmovdqa 256(%rdi), %ymm04398; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4399; AVX2-FP-NEXT: vmovdqa 224(%rdi), %ymm54400; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm5[0,1,2,3,4,5],ymm0[6,7]4401; AVX2-FP-NEXT: vmovdqa 192(%rdi), %xmm24402; AVX2-FP-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4403; AVX2-FP-NEXT: vmovdqa 160(%rdi), %xmm144404; AVX2-FP-NEXT: vpblendd {{.*#+}} xmm2 = xmm14[0,1],xmm2[2,3]4405; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]4406; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4407; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm6[0,1,2,3,4,5],ymm3[6,7]4408; AVX2-FP-NEXT: vmovdqa 512(%rdi), %xmm24409; AVX2-FP-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4410; AVX2-FP-NEXT: vmovdqa 480(%rdi), %xmm104411; AVX2-FP-NEXT: vpblendd {{.*#+}} xmm2 = xmm10[0,1],xmm2[2,3]4412; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]4413; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4414; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm4[0,1,2,3,4,5],ymm1[6,7]4415; AVX2-FP-NEXT: vmovdqa 832(%rdi), %xmm14416; AVX2-FP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4417; AVX2-FP-NEXT: vmovdqa 800(%rdi), %xmm44418; AVX2-FP-NEXT: vpblendd {{.*#+}} xmm2 = xmm4[0,1],xmm1[2,3]4419; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]4420; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4421; AVX2-FP-NEXT: vmovdqa 1216(%rdi), %ymm04422; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4423; AVX2-FP-NEXT: vmovdqa 1184(%rdi), %ymm24424; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3,4,5],ymm0[6,7]4425; AVX2-FP-NEXT: vmovdqa 1152(%rdi), %xmm14426; AVX2-FP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4427; AVX2-FP-NEXT: vmovdqa 1120(%rdi), %xmm124428; AVX2-FP-NEXT: vpblendd {{.*#+}} xmm3 = xmm12[0,1],xmm1[2,3]4429; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm3[0,1,2,3],ymm0[4,5,6,7]4430; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4431; AVX2-FP-NEXT: vmovdqa 96(%rdi), %ymm04432; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4433; AVX2-FP-NEXT: vmovdqa 64(%rdi), %ymm14434; AVX2-FP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4435; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]4436; AVX2-FP-NEXT: vmovdqa (%rdi), %xmm134437; AVX2-FP-NEXT: vmovdqa 32(%rdi), %xmm14438; AVX2-FP-NEXT: vmovdqa %xmm1, (%rsp) # 16-byte Spill4439; AVX2-FP-NEXT: vpblendd {{.*#+}} xmm6 = xmm13[0,1],xmm1[2,3]4440; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm6[0,1,2,3],ymm0[4,5,6,7]4441; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4442; AVX2-FP-NEXT: vmovdqa 416(%rdi), %ymm04443; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4444; AVX2-FP-NEXT: vmovdqa 384(%rdi), %ymm64445; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm6[0,1,2,3,4,5],ymm0[6,7]4446; AVX2-FP-NEXT: vmovdqa 352(%rdi), %xmm14447; AVX2-FP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4448; AVX2-FP-NEXT: vmovdqa 320(%rdi), %xmm154449; AVX2-FP-NEXT: vpblendd {{.*#+}} xmm7 = xmm15[0,1],xmm1[2,3]4450; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm7[0,1,2,3],ymm0[4,5,6,7]4451; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4452; AVX2-FP-NEXT: vmovdqa 736(%rdi), %ymm04453; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4454; AVX2-FP-NEXT: vmovdqa 704(%rdi), %ymm74455; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm8 = ymm7[0,1,2,3,4,5],ymm0[6,7]4456; AVX2-FP-NEXT: vmovdqa 672(%rdi), %xmm04457; AVX2-FP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4458; AVX2-FP-NEXT: vmovdqa 640(%rdi), %xmm34459; AVX2-FP-NEXT: vpblendd {{.*#+}} xmm9 = xmm3[0,1],xmm0[2,3]4460; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm9[0,1,2,3],ymm8[4,5,6,7]4461; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4462; AVX2-FP-NEXT: vmovdqa 1056(%rdi), %ymm04463; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4464; AVX2-FP-NEXT: vmovdqa 1024(%rdi), %ymm84465; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm1 = ymm8[0,1,2,3,4,5],ymm0[6,7]4466; AVX2-FP-NEXT: vmovdqa 992(%rdi), %xmm04467; AVX2-FP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4468; AVX2-FP-NEXT: vmovdqa 960(%rdi), %xmm94469; AVX2-FP-NEXT: vpblendd {{.*#+}} xmm11 = xmm9[0,1],xmm0[2,3]4470; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm11[0,1,2,3],ymm1[4,5,6,7]4471; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4472; AVX2-FP-NEXT: vmovdqa 208(%rdi), %xmm04473; AVX2-FP-NEXT: vpalignr {{.*#+}} xmm0 = xmm14[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]4474; AVX2-FP-NEXT: vmovdqa 288(%rdi), %ymm144475; AVX2-FP-NEXT: vpalignr {{.*#+}} ymm5 = ymm5[8,9,10,11,12,13,14,15],ymm14[0,1,2,3,4,5,6,7],ymm5[24,25,26,27,28,29,30,31],ymm14[16,17,18,19,20,21,22,23]4476; AVX2-FP-NEXT: vmovdqu %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4477; AVX2-FP-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,1,2,1]4478; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm5[4,5,6,7]4479; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4480; AVX2-FP-NEXT: vmovdqa 528(%rdi), %xmm04481; AVX2-FP-NEXT: vpalignr {{.*#+}} xmm0 = xmm10[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]4482; AVX2-FP-NEXT: vmovdqa 608(%rdi), %ymm114483; AVX2-FP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm11, %ymm5 # 32-byte Folded Reload4484; AVX2-FP-NEXT: # ymm5 = mem[8,9,10,11,12,13,14,15],ymm11[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm11[16,17,18,19,20,21,22,23]4485; AVX2-FP-NEXT: vmovdqu %ymm11, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4486; AVX2-FP-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,1,2,1]4487; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm5[4,5,6,7]4488; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4489; AVX2-FP-NEXT: vmovdqa 848(%rdi), %xmm04490; AVX2-FP-NEXT: vpalignr {{.*#+}} xmm0 = xmm4[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]4491; AVX2-FP-NEXT: vmovdqa 928(%rdi), %ymm104492; AVX2-FP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm10, %ymm4 # 32-byte Folded Reload4493; AVX2-FP-NEXT: # ymm4 = mem[8,9,10,11,12,13,14,15],ymm10[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm10[16,17,18,19,20,21,22,23]4494; AVX2-FP-NEXT: vmovdqu %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4495; AVX2-FP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,1,2,1]4496; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm4[4,5,6,7]4497; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4498; AVX2-FP-NEXT: vmovdqa 1168(%rdi), %xmm04499; AVX2-FP-NEXT: vpalignr {{.*#+}} xmm0 = xmm12[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]4500; AVX2-FP-NEXT: vmovdqa 1248(%rdi), %ymm54501; AVX2-FP-NEXT: vpalignr {{.*#+}} ymm2 = ymm2[8,9,10,11,12,13,14,15],ymm5[0,1,2,3,4,5,6,7],ymm2[24,25,26,27,28,29,30,31],ymm5[16,17,18,19,20,21,22,23]4502; AVX2-FP-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4503; AVX2-FP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,1,2,1]4504; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]4505; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4506; AVX2-FP-NEXT: vmovdqa 48(%rdi), %xmm04507; AVX2-FP-NEXT: vpalignr {{.*#+}} xmm0 = xmm13[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]4508; AVX2-FP-NEXT: vmovdqa 128(%rdi), %ymm124509; AVX2-FP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm12, %ymm2 # 32-byte Folded Reload4510; AVX2-FP-NEXT: # ymm2 = mem[8,9,10,11,12,13,14,15],ymm12[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm12[16,17,18,19,20,21,22,23]4511; AVX2-FP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,1,2,1]4512; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]4513; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4514; AVX2-FP-NEXT: vmovdqa 368(%rdi), %xmm04515; AVX2-FP-NEXT: vpalignr {{.*#+}} xmm0 = xmm15[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]4516; AVX2-FP-NEXT: vmovdqa 448(%rdi), %ymm14517; AVX2-FP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4518; AVX2-FP-NEXT: vpalignr {{.*#+}} ymm2 = ymm6[8,9,10,11,12,13,14,15],ymm1[0,1,2,3,4,5,6,7],ymm6[24,25,26,27,28,29,30,31],ymm1[16,17,18,19,20,21,22,23]4519; AVX2-FP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,1,2,1]4520; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]4521; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4522; AVX2-FP-NEXT: vmovdqa 688(%rdi), %xmm04523; AVX2-FP-NEXT: vpalignr {{.*#+}} xmm0 = xmm3[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]4524; AVX2-FP-NEXT: vmovdqa 768(%rdi), %ymm154525; AVX2-FP-NEXT: vpalignr {{.*#+}} ymm1 = ymm7[8,9,10,11,12,13,14,15],ymm15[0,1,2,3,4,5,6,7],ymm7[24,25,26,27,28,29,30,31],ymm15[16,17,18,19,20,21,22,23]4526; AVX2-FP-NEXT: vmovdqu %ymm15, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4527; AVX2-FP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,1,2,1]4528; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]4529; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4530; AVX2-FP-NEXT: vmovdqa 1008(%rdi), %xmm04531; AVX2-FP-NEXT: vpalignr {{.*#+}} xmm0 = xmm9[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]4532; AVX2-FP-NEXT: vmovdqa 1088(%rdi), %ymm24533; AVX2-FP-NEXT: vpalignr {{.*#+}} ymm1 = ymm8[8,9,10,11,12,13,14,15],ymm2[0,1,2,3,4,5,6,7],ymm8[24,25,26,27,28,29,30,31],ymm2[16,17,18,19,20,21,22,23]4534; AVX2-FP-NEXT: vmovdqa %ymm2, %ymm84535; AVX2-FP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4536; AVX2-FP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,1,2,1]4537; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]4538; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4539; AVX2-FP-NEXT: vmovaps 160(%rdi), %ymm04540; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]4541; AVX2-FP-NEXT: vmovaps 256(%rdi), %xmm14542; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]4543; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm14544; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]4545; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4546; AVX2-FP-NEXT: vmovaps 480(%rdi), %ymm04547; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]4548; AVX2-FP-NEXT: vmovaps 576(%rdi), %xmm14549; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]4550; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm14551; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]4552; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4553; AVX2-FP-NEXT: vmovaps 800(%rdi), %ymm04554; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]4555; AVX2-FP-NEXT: vmovaps 896(%rdi), %xmm14556; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]4557; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm14558; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]4559; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4560; AVX2-FP-NEXT: vmovaps 1120(%rdi), %ymm04561; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]4562; AVX2-FP-NEXT: vmovaps 1216(%rdi), %xmm14563; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]4564; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm14565; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]4566; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4567; AVX2-FP-NEXT: vmovaps (%rdi), %ymm04568; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]4569; AVX2-FP-NEXT: vmovaps 96(%rdi), %xmm14570; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]4571; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm14572; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]4573; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4574; AVX2-FP-NEXT: vmovaps 320(%rdi), %ymm04575; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]4576; AVX2-FP-NEXT: vmovaps 416(%rdi), %xmm14577; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]4578; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm14579; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]4580; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4581; AVX2-FP-NEXT: vmovaps 640(%rdi), %ymm04582; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]4583; AVX2-FP-NEXT: vmovaps 736(%rdi), %xmm14584; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]4585; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm14586; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]4587; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4588; AVX2-FP-NEXT: vmovaps 960(%rdi), %ymm04589; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]4590; AVX2-FP-NEXT: vmovaps 1056(%rdi), %xmm14591; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]4592; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm14593; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]4594; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4595; AVX2-FP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm14, %ymm0 # 32-byte Folded Reload4596; AVX2-FP-NEXT: # ymm0 = mem[8,9,10,11,12,13,14,15],ymm14[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm14[16,17,18,19,20,21,22,23]4597; AVX2-FP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,3]4598; AVX2-FP-NEXT: vmovdqa 224(%rdi), %xmm44599; AVX2-FP-NEXT: vpalignr {{.*#+}} xmm1 = mem[8,9,10,11,12,13,14,15],xmm4[0,1,2,3,4,5,6,7]4600; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]4601; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4602; AVX2-FP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm11, %ymm0 # 32-byte Folded Reload4603; AVX2-FP-NEXT: # ymm0 = mem[8,9,10,11,12,13,14,15],ymm11[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm11[16,17,18,19,20,21,22,23]4604; AVX2-FP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,3]4605; AVX2-FP-NEXT: vmovdqa 544(%rdi), %xmm64606; AVX2-FP-NEXT: vpalignr {{.*#+}} xmm2 = mem[8,9,10,11,12,13,14,15],xmm6[0,1,2,3,4,5,6,7]4607; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]4608; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4609; AVX2-FP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm10, %ymm0 # 32-byte Folded Reload4610; AVX2-FP-NEXT: # ymm0 = mem[8,9,10,11,12,13,14,15],ymm10[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm10[16,17,18,19,20,21,22,23]4611; AVX2-FP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,3]4612; AVX2-FP-NEXT: vmovdqa 864(%rdi), %xmm14613; AVX2-FP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4614; AVX2-FP-NEXT: vpalignr {{.*#+}} xmm3 = mem[8,9,10,11,12,13,14,15],xmm1[0,1,2,3,4,5,6,7]4615; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm3[0,1,2,3],ymm0[4,5,6,7]4616; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4617; AVX2-FP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm0 # 32-byte Folded Reload4618; AVX2-FP-NEXT: # ymm0 = mem[8,9,10,11,12,13,14,15],ymm5[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm5[16,17,18,19,20,21,22,23]4619; AVX2-FP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,3]4620; AVX2-FP-NEXT: vmovdqa 1184(%rdi), %xmm144621; AVX2-FP-NEXT: vpalignr {{.*#+}} xmm7 = mem[8,9,10,11,12,13,14,15],xmm14[0,1,2,3,4,5,6,7]4622; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm7[0,1,2,3],ymm0[4,5,6,7]4623; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4624; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload4625; AVX2-FP-NEXT: vpalignr {{.*#+}} ymm0 = ymm13[8,9,10,11,12,13,14,15],ymm8[0,1,2,3,4,5,6,7],ymm13[24,25,26,27,28,29,30,31],ymm8[16,17,18,19,20,21,22,23]4626; AVX2-FP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,3]4627; AVX2-FP-NEXT: vmovdqa 1024(%rdi), %xmm114628; AVX2-FP-NEXT: vpalignr {{.*#+}} xmm7 = mem[8,9,10,11,12,13,14,15],xmm11[0,1,2,3,4,5,6,7]4629; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm7[0,1,2,3],ymm0[4,5,6,7]4630; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4631; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload4632; AVX2-FP-NEXT: vpalignr {{.*#+}} ymm0 = ymm10[8,9,10,11,12,13,14,15],ymm15[0,1,2,3,4,5,6,7],ymm10[24,25,26,27,28,29,30,31],ymm15[16,17,18,19,20,21,22,23]4633; AVX2-FP-NEXT: vpermq {{.*#+}} ymm7 = ymm0[0,1,0,3]4634; AVX2-FP-NEXT: vmovdqa 704(%rdi), %xmm94635; AVX2-FP-NEXT: vpalignr {{.*#+}} xmm8 = mem[8,9,10,11,12,13,14,15],xmm9[0,1,2,3,4,5,6,7]4636; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm8[0,1,2,3],ymm7[4,5,6,7]4637; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4638; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload4639; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload4640; AVX2-FP-NEXT: vpalignr {{.*#+}} ymm7 = ymm3[8,9,10,11,12,13,14,15],ymm2[0,1,2,3,4,5,6,7],ymm3[24,25,26,27,28,29,30,31],ymm2[16,17,18,19,20,21,22,23]4641; AVX2-FP-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,1,0,3]4642; AVX2-FP-NEXT: vmovdqa 384(%rdi), %xmm54643; AVX2-FP-NEXT: vpalignr {{.*#+}} xmm8 = mem[8,9,10,11,12,13,14,15],xmm5[0,1,2,3,4,5,6,7]4644; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm8 = ymm8[0,1,2,3],ymm7[4,5,6,7]4645; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4646; AVX2-FP-NEXT: vpalignr {{.*#+}} ymm7 = ymm1[8,9,10,11,12,13,14,15],ymm12[0,1,2,3,4,5,6,7],ymm1[24,25,26,27,28,29,30,31],ymm12[16,17,18,19,20,21,22,23]4647; AVX2-FP-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,1,0,3]4648; AVX2-FP-NEXT: vmovdqa 64(%rdi), %xmm04649; AVX2-FP-NEXT: vpalignr {{.*#+}} xmm15 = mem[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]4650; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm7 = ymm15[0,1,2,3],ymm7[4,5,6,7]4651; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm12 = ymm1[0,1,2,3,4,5],ymm12[6,7]4652; AVX2-FP-NEXT: vpblendd $3, (%rsp), %xmm0, %xmm0 # 16-byte Folded Reload4653; AVX2-FP-NEXT: # xmm0 = mem[0,1],xmm0[2,3]4654; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm12 = ymm0[0,1,2,3],ymm12[4,5,6,7]4655; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload4656; AVX2-FP-NEXT: vpblendd $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload4657; AVX2-FP-NEXT: # ymm0 = mem[0,1,2,3,4,5],ymm0[6,7]4658; AVX2-FP-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4, %xmm4 # 16-byte Folded Reload4659; AVX2-FP-NEXT: # xmm4 = mem[0,1],xmm4[2,3]4660; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm0[4,5,6,7]4661; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm3[0,1,2,3,4,5],ymm2[6,7]4662; AVX2-FP-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm5, %xmm5 # 16-byte Folded Reload4663; AVX2-FP-NEXT: # xmm5 = mem[0,1],xmm5[2,3]4664; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm0[4,5,6,7]4665; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload4666; AVX2-FP-NEXT: vpblendd $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload4667; AVX2-FP-NEXT: # ymm0 = mem[0,1,2,3,4,5],ymm0[6,7]4668; AVX2-FP-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm6, %xmm6 # 16-byte Folded Reload4669; AVX2-FP-NEXT: # xmm6 = mem[0,1],xmm6[2,3]4670; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm15 = ymm6[0,1,2,3],ymm0[4,5,6,7]4671; AVX2-FP-NEXT: vpblendd $192, {{[-0-9]+}}(%r{{[sb]}}p), %ymm10, %ymm0 # 32-byte Folded Reload4672; AVX2-FP-NEXT: # ymm0 = ymm10[0,1,2,3,4,5],mem[6,7]4673; AVX2-FP-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm9, %xmm1 # 16-byte Folded Reload4674; AVX2-FP-NEXT: # xmm1 = mem[0,1],xmm9[2,3]4675; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm6 = ymm1[0,1,2,3],ymm0[4,5,6,7]4676; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload4677; AVX2-FP-NEXT: vblendps $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload4678; AVX2-FP-NEXT: # ymm0 = mem[0,1,2,3,4,5],ymm0[6,7]4679; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload4680; AVX2-FP-NEXT: vblendps $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload4681; AVX2-FP-NEXT: # xmm1 = mem[0,1],xmm1[2,3]4682; AVX2-FP-NEXT: vblendps {{.*#+}} ymm2 = ymm1[0,1,2,3],ymm0[4,5,6,7]4683; AVX2-FP-NEXT: vpblendd $192, {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm0 # 32-byte Folded Reload4684; AVX2-FP-NEXT: # ymm0 = ymm13[0,1,2,3,4,5],mem[6,7]4685; AVX2-FP-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm11, %xmm1 # 16-byte Folded Reload4686; AVX2-FP-NEXT: # xmm1 = mem[0,1],xmm11[2,3]4687; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm9 = ymm1[0,1,2,3],ymm0[4,5,6,7]4688; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload4689; AVX2-FP-NEXT: vpblendd $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload4690; AVX2-FP-NEXT: # ymm0 = mem[0,1,2,3,4,5],ymm0[6,7]4691; AVX2-FP-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm1 # 16-byte Folded Reload4692; AVX2-FP-NEXT: # xmm1 = mem[0,1],xmm14[2,3]4693; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]4694; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4695; AVX2-FP-NEXT: vmovaps %ymm1, 192(%rsi)4696; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4697; AVX2-FP-NEXT: vmovaps %ymm1, 128(%rsi)4698; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4699; AVX2-FP-NEXT: vmovaps %ymm1, 64(%rsi)4700; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4701; AVX2-FP-NEXT: vmovaps %ymm1, (%rsi)4702; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4703; AVX2-FP-NEXT: vmovaps %ymm1, 224(%rsi)4704; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4705; AVX2-FP-NEXT: vmovaps %ymm1, 160(%rsi)4706; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4707; AVX2-FP-NEXT: vmovaps %ymm1, 96(%rsi)4708; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4709; AVX2-FP-NEXT: vmovaps %ymm1, 32(%rsi)4710; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4711; AVX2-FP-NEXT: vmovaps %ymm1, 192(%rdx)4712; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4713; AVX2-FP-NEXT: vmovaps %ymm1, 128(%rdx)4714; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4715; AVX2-FP-NEXT: vmovaps %ymm1, 64(%rdx)4716; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4717; AVX2-FP-NEXT: vmovaps %ymm1, (%rdx)4718; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4719; AVX2-FP-NEXT: vmovaps %ymm1, 224(%rdx)4720; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4721; AVX2-FP-NEXT: vmovaps %ymm1, 160(%rdx)4722; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4723; AVX2-FP-NEXT: vmovaps %ymm1, 96(%rdx)4724; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4725; AVX2-FP-NEXT: vmovaps %ymm1, 32(%rdx)4726; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4727; AVX2-FP-NEXT: vmovaps %ymm1, 192(%rcx)4728; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4729; AVX2-FP-NEXT: vmovaps %ymm1, 128(%rcx)4730; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4731; AVX2-FP-NEXT: vmovaps %ymm1, 64(%rcx)4732; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4733; AVX2-FP-NEXT: vmovaps %ymm1, (%rcx)4734; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4735; AVX2-FP-NEXT: vmovaps %ymm1, 224(%rcx)4736; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4737; AVX2-FP-NEXT: vmovaps %ymm1, 160(%rcx)4738; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4739; AVX2-FP-NEXT: vmovaps %ymm1, 96(%rcx)4740; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4741; AVX2-FP-NEXT: vmovaps %ymm1, 32(%rcx)4742; AVX2-FP-NEXT: vmovdqa %ymm7, (%r8)4743; AVX2-FP-NEXT: vmovdqa %ymm8, 64(%r8)4744; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4745; AVX2-FP-NEXT: vmovaps %ymm1, 128(%r8)4746; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4747; AVX2-FP-NEXT: vmovaps %ymm1, 192(%r8)4748; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4749; AVX2-FP-NEXT: vmovaps %ymm1, 224(%r8)4750; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4751; AVX2-FP-NEXT: vmovaps %ymm1, 160(%r8)4752; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4753; AVX2-FP-NEXT: vmovaps %ymm1, 96(%r8)4754; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload4755; AVX2-FP-NEXT: vmovaps %ymm1, 32(%r8)4756; AVX2-FP-NEXT: vmovdqa %ymm0, 224(%r9)4757; AVX2-FP-NEXT: vmovdqa %ymm9, 192(%r9)4758; AVX2-FP-NEXT: vmovaps %ymm2, 160(%r9)4759; AVX2-FP-NEXT: vmovdqa %ymm6, 128(%r9)4760; AVX2-FP-NEXT: vmovdqa %ymm15, 96(%r9)4761; AVX2-FP-NEXT: vmovdqa %ymm5, 64(%r9)4762; AVX2-FP-NEXT: vmovdqa %ymm4, 32(%r9)4763; AVX2-FP-NEXT: vmovdqa %ymm12, (%r9)4764; AVX2-FP-NEXT: addq $1464, %rsp # imm = 0x5B84765; AVX2-FP-NEXT: vzeroupper4766; AVX2-FP-NEXT: retq4767;4768; AVX2-FCP-LABEL: load_i64_stride5_vf32:4769; AVX2-FCP: # %bb.0:4770; AVX2-FCP-NEXT: subq $1464, %rsp # imm = 0x5B84771; AVX2-FCP-NEXT: vmovdqa 896(%rdi), %ymm14772; AVX2-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4773; AVX2-FCP-NEXT: vmovdqa 864(%rdi), %ymm44774; AVX2-FCP-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4775; AVX2-FCP-NEXT: vmovdqa 576(%rdi), %ymm34776; AVX2-FCP-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4777; AVX2-FCP-NEXT: vmovdqa 544(%rdi), %ymm64778; AVX2-FCP-NEXT: vmovdqu %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4779; AVX2-FCP-NEXT: vmovdqa 256(%rdi), %ymm04780; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4781; AVX2-FCP-NEXT: vmovdqa 224(%rdi), %ymm54782; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm5[0,1,2,3,4,5],ymm0[6,7]4783; AVX2-FCP-NEXT: vmovdqa 192(%rdi), %xmm24784; AVX2-FCP-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4785; AVX2-FCP-NEXT: vmovdqa 160(%rdi), %xmm144786; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm2 = xmm14[0,1],xmm2[2,3]4787; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]4788; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4789; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm6[0,1,2,3,4,5],ymm3[6,7]4790; AVX2-FCP-NEXT: vmovdqa 512(%rdi), %xmm24791; AVX2-FCP-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4792; AVX2-FCP-NEXT: vmovdqa 480(%rdi), %xmm104793; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm2 = xmm10[0,1],xmm2[2,3]4794; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]4795; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4796; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm4[0,1,2,3,4,5],ymm1[6,7]4797; AVX2-FCP-NEXT: vmovdqa 832(%rdi), %xmm14798; AVX2-FCP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4799; AVX2-FCP-NEXT: vmovdqa 800(%rdi), %xmm44800; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm2 = xmm4[0,1],xmm1[2,3]4801; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]4802; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4803; AVX2-FCP-NEXT: vmovdqa 1216(%rdi), %ymm04804; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4805; AVX2-FCP-NEXT: vmovdqa 1184(%rdi), %ymm24806; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3,4,5],ymm0[6,7]4807; AVX2-FCP-NEXT: vmovdqa 1152(%rdi), %xmm14808; AVX2-FCP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4809; AVX2-FCP-NEXT: vmovdqa 1120(%rdi), %xmm124810; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm3 = xmm12[0,1],xmm1[2,3]4811; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm3[0,1,2,3],ymm0[4,5,6,7]4812; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4813; AVX2-FCP-NEXT: vmovdqa 96(%rdi), %ymm04814; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4815; AVX2-FCP-NEXT: vmovdqa 64(%rdi), %ymm14816; AVX2-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4817; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]4818; AVX2-FCP-NEXT: vmovdqa (%rdi), %xmm134819; AVX2-FCP-NEXT: vmovdqa 32(%rdi), %xmm14820; AVX2-FCP-NEXT: vmovdqa %xmm1, (%rsp) # 16-byte Spill4821; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm6 = xmm13[0,1],xmm1[2,3]4822; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm6[0,1,2,3],ymm0[4,5,6,7]4823; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4824; AVX2-FCP-NEXT: vmovdqa 416(%rdi), %ymm04825; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4826; AVX2-FCP-NEXT: vmovdqa 384(%rdi), %ymm64827; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm6[0,1,2,3,4,5],ymm0[6,7]4828; AVX2-FCP-NEXT: vmovdqa 352(%rdi), %xmm14829; AVX2-FCP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4830; AVX2-FCP-NEXT: vmovdqa 320(%rdi), %xmm154831; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm7 = xmm15[0,1],xmm1[2,3]4832; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm7[0,1,2,3],ymm0[4,5,6,7]4833; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4834; AVX2-FCP-NEXT: vmovdqa 736(%rdi), %ymm04835; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4836; AVX2-FCP-NEXT: vmovdqa 704(%rdi), %ymm74837; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm8 = ymm7[0,1,2,3,4,5],ymm0[6,7]4838; AVX2-FCP-NEXT: vmovdqa 672(%rdi), %xmm04839; AVX2-FCP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4840; AVX2-FCP-NEXT: vmovdqa 640(%rdi), %xmm34841; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm9 = xmm3[0,1],xmm0[2,3]4842; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm9[0,1,2,3],ymm8[4,5,6,7]4843; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4844; AVX2-FCP-NEXT: vmovdqa 1056(%rdi), %ymm04845; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4846; AVX2-FCP-NEXT: vmovdqa 1024(%rdi), %ymm84847; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm1 = ymm8[0,1,2,3,4,5],ymm0[6,7]4848; AVX2-FCP-NEXT: vmovdqa 992(%rdi), %xmm04849; AVX2-FCP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4850; AVX2-FCP-NEXT: vmovdqa 960(%rdi), %xmm94851; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm11 = xmm9[0,1],xmm0[2,3]4852; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm11[0,1,2,3],ymm1[4,5,6,7]4853; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4854; AVX2-FCP-NEXT: vmovdqa 208(%rdi), %xmm04855; AVX2-FCP-NEXT: vpalignr {{.*#+}} xmm0 = xmm14[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]4856; AVX2-FCP-NEXT: vmovdqa 288(%rdi), %ymm144857; AVX2-FCP-NEXT: vpalignr {{.*#+}} ymm5 = ymm5[8,9,10,11,12,13,14,15],ymm14[0,1,2,3,4,5,6,7],ymm5[24,25,26,27,28,29,30,31],ymm14[16,17,18,19,20,21,22,23]4858; AVX2-FCP-NEXT: vmovdqu %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4859; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,1,2,1]4860; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm5[4,5,6,7]4861; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4862; AVX2-FCP-NEXT: vmovdqa 528(%rdi), %xmm04863; AVX2-FCP-NEXT: vpalignr {{.*#+}} xmm0 = xmm10[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]4864; AVX2-FCP-NEXT: vmovdqa 608(%rdi), %ymm114865; AVX2-FCP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm11, %ymm5 # 32-byte Folded Reload4866; AVX2-FCP-NEXT: # ymm5 = mem[8,9,10,11,12,13,14,15],ymm11[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm11[16,17,18,19,20,21,22,23]4867; AVX2-FCP-NEXT: vmovdqu %ymm11, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4868; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,1,2,1]4869; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm5[4,5,6,7]4870; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4871; AVX2-FCP-NEXT: vmovdqa 848(%rdi), %xmm04872; AVX2-FCP-NEXT: vpalignr {{.*#+}} xmm0 = xmm4[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]4873; AVX2-FCP-NEXT: vmovdqa 928(%rdi), %ymm104874; AVX2-FCP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm10, %ymm4 # 32-byte Folded Reload4875; AVX2-FCP-NEXT: # ymm4 = mem[8,9,10,11,12,13,14,15],ymm10[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm10[16,17,18,19,20,21,22,23]4876; AVX2-FCP-NEXT: vmovdqu %ymm10, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4877; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,1,2,1]4878; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm4[4,5,6,7]4879; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4880; AVX2-FCP-NEXT: vmovdqa 1168(%rdi), %xmm04881; AVX2-FCP-NEXT: vpalignr {{.*#+}} xmm0 = xmm12[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]4882; AVX2-FCP-NEXT: vmovdqa 1248(%rdi), %ymm54883; AVX2-FCP-NEXT: vpalignr {{.*#+}} ymm2 = ymm2[8,9,10,11,12,13,14,15],ymm5[0,1,2,3,4,5,6,7],ymm2[24,25,26,27,28,29,30,31],ymm5[16,17,18,19,20,21,22,23]4884; AVX2-FCP-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4885; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,1,2,1]4886; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]4887; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4888; AVX2-FCP-NEXT: vmovdqa 48(%rdi), %xmm04889; AVX2-FCP-NEXT: vpalignr {{.*#+}} xmm0 = xmm13[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]4890; AVX2-FCP-NEXT: vmovdqa 128(%rdi), %ymm124891; AVX2-FCP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm12, %ymm2 # 32-byte Folded Reload4892; AVX2-FCP-NEXT: # ymm2 = mem[8,9,10,11,12,13,14,15],ymm12[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm12[16,17,18,19,20,21,22,23]4893; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,1,2,1]4894; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]4895; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4896; AVX2-FCP-NEXT: vmovdqa 368(%rdi), %xmm04897; AVX2-FCP-NEXT: vpalignr {{.*#+}} xmm0 = xmm15[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]4898; AVX2-FCP-NEXT: vmovdqa 448(%rdi), %ymm14899; AVX2-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4900; AVX2-FCP-NEXT: vpalignr {{.*#+}} ymm2 = ymm6[8,9,10,11,12,13,14,15],ymm1[0,1,2,3,4,5,6,7],ymm6[24,25,26,27,28,29,30,31],ymm1[16,17,18,19,20,21,22,23]4901; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,1,2,1]4902; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]4903; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4904; AVX2-FCP-NEXT: vmovdqa 688(%rdi), %xmm04905; AVX2-FCP-NEXT: vpalignr {{.*#+}} xmm0 = xmm3[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]4906; AVX2-FCP-NEXT: vmovdqa 768(%rdi), %ymm154907; AVX2-FCP-NEXT: vpalignr {{.*#+}} ymm1 = ymm7[8,9,10,11,12,13,14,15],ymm15[0,1,2,3,4,5,6,7],ymm7[24,25,26,27,28,29,30,31],ymm15[16,17,18,19,20,21,22,23]4908; AVX2-FCP-NEXT: vmovdqu %ymm15, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4909; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,1,2,1]4910; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]4911; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4912; AVX2-FCP-NEXT: vmovdqa 1008(%rdi), %xmm04913; AVX2-FCP-NEXT: vpalignr {{.*#+}} xmm0 = xmm9[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]4914; AVX2-FCP-NEXT: vmovdqa 1088(%rdi), %ymm24915; AVX2-FCP-NEXT: vpalignr {{.*#+}} ymm1 = ymm8[8,9,10,11,12,13,14,15],ymm2[0,1,2,3,4,5,6,7],ymm8[24,25,26,27,28,29,30,31],ymm2[16,17,18,19,20,21,22,23]4916; AVX2-FCP-NEXT: vmovdqa %ymm2, %ymm84917; AVX2-FCP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4918; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,1,2,1]4919; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]4920; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4921; AVX2-FCP-NEXT: vmovaps 160(%rdi), %ymm04922; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]4923; AVX2-FCP-NEXT: vmovaps 256(%rdi), %xmm14924; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]4925; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm14926; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]4927; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4928; AVX2-FCP-NEXT: vmovaps 480(%rdi), %ymm04929; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]4930; AVX2-FCP-NEXT: vmovaps 576(%rdi), %xmm14931; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]4932; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm14933; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]4934; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4935; AVX2-FCP-NEXT: vmovaps 800(%rdi), %ymm04936; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]4937; AVX2-FCP-NEXT: vmovaps 896(%rdi), %xmm14938; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]4939; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm14940; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]4941; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4942; AVX2-FCP-NEXT: vmovaps 1120(%rdi), %ymm04943; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]4944; AVX2-FCP-NEXT: vmovaps 1216(%rdi), %xmm14945; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]4946; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm14947; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]4948; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4949; AVX2-FCP-NEXT: vmovaps (%rdi), %ymm04950; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]4951; AVX2-FCP-NEXT: vmovaps 96(%rdi), %xmm14952; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]4953; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm14954; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]4955; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4956; AVX2-FCP-NEXT: vmovaps 320(%rdi), %ymm04957; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]4958; AVX2-FCP-NEXT: vmovaps 416(%rdi), %xmm14959; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]4960; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm14961; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]4962; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4963; AVX2-FCP-NEXT: vmovaps 640(%rdi), %ymm04964; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]4965; AVX2-FCP-NEXT: vmovaps 736(%rdi), %xmm14966; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]4967; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm14968; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]4969; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4970; AVX2-FCP-NEXT: vmovaps 960(%rdi), %ymm04971; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]4972; AVX2-FCP-NEXT: vmovaps 1056(%rdi), %xmm14973; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]4974; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm14975; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]4976; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4977; AVX2-FCP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm14, %ymm0 # 32-byte Folded Reload4978; AVX2-FCP-NEXT: # ymm0 = mem[8,9,10,11,12,13,14,15],ymm14[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm14[16,17,18,19,20,21,22,23]4979; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,3]4980; AVX2-FCP-NEXT: vmovdqa 224(%rdi), %xmm44981; AVX2-FCP-NEXT: vpalignr {{.*#+}} xmm1 = mem[8,9,10,11,12,13,14,15],xmm4[0,1,2,3,4,5,6,7]4982; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]4983; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4984; AVX2-FCP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm11, %ymm0 # 32-byte Folded Reload4985; AVX2-FCP-NEXT: # ymm0 = mem[8,9,10,11,12,13,14,15],ymm11[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm11[16,17,18,19,20,21,22,23]4986; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,3]4987; AVX2-FCP-NEXT: vmovdqa 544(%rdi), %xmm64988; AVX2-FCP-NEXT: vpalignr {{.*#+}} xmm2 = mem[8,9,10,11,12,13,14,15],xmm6[0,1,2,3,4,5,6,7]4989; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]4990; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4991; AVX2-FCP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm10, %ymm0 # 32-byte Folded Reload4992; AVX2-FCP-NEXT: # ymm0 = mem[8,9,10,11,12,13,14,15],ymm10[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm10[16,17,18,19,20,21,22,23]4993; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,3]4994; AVX2-FCP-NEXT: vmovdqa 864(%rdi), %xmm14995; AVX2-FCP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill4996; AVX2-FCP-NEXT: vpalignr {{.*#+}} xmm3 = mem[8,9,10,11,12,13,14,15],xmm1[0,1,2,3,4,5,6,7]4997; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm3[0,1,2,3],ymm0[4,5,6,7]4998; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill4999; AVX2-FCP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm0 # 32-byte Folded Reload5000; AVX2-FCP-NEXT: # ymm0 = mem[8,9,10,11,12,13,14,15],ymm5[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm5[16,17,18,19,20,21,22,23]5001; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,3]5002; AVX2-FCP-NEXT: vmovdqa 1184(%rdi), %xmm145003; AVX2-FCP-NEXT: vpalignr {{.*#+}} xmm7 = mem[8,9,10,11,12,13,14,15],xmm14[0,1,2,3,4,5,6,7]5004; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm7[0,1,2,3],ymm0[4,5,6,7]5005; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5006; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload5007; AVX2-FCP-NEXT: vpalignr {{.*#+}} ymm0 = ymm13[8,9,10,11,12,13,14,15],ymm8[0,1,2,3,4,5,6,7],ymm13[24,25,26,27,28,29,30,31],ymm8[16,17,18,19,20,21,22,23]5008; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,3]5009; AVX2-FCP-NEXT: vmovdqa 1024(%rdi), %xmm115010; AVX2-FCP-NEXT: vpalignr {{.*#+}} xmm7 = mem[8,9,10,11,12,13,14,15],xmm11[0,1,2,3,4,5,6,7]5011; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm7[0,1,2,3],ymm0[4,5,6,7]5012; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5013; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload5014; AVX2-FCP-NEXT: vpalignr {{.*#+}} ymm0 = ymm10[8,9,10,11,12,13,14,15],ymm15[0,1,2,3,4,5,6,7],ymm10[24,25,26,27,28,29,30,31],ymm15[16,17,18,19,20,21,22,23]5015; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm7 = ymm0[0,1,0,3]5016; AVX2-FCP-NEXT: vmovdqa 704(%rdi), %xmm95017; AVX2-FCP-NEXT: vpalignr {{.*#+}} xmm8 = mem[8,9,10,11,12,13,14,15],xmm9[0,1,2,3,4,5,6,7]5018; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm8[0,1,2,3],ymm7[4,5,6,7]5019; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill5020; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload5021; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload5022; AVX2-FCP-NEXT: vpalignr {{.*#+}} ymm7 = ymm3[8,9,10,11,12,13,14,15],ymm2[0,1,2,3,4,5,6,7],ymm3[24,25,26,27,28,29,30,31],ymm2[16,17,18,19,20,21,22,23]5023; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,1,0,3]5024; AVX2-FCP-NEXT: vmovdqa 384(%rdi), %xmm55025; AVX2-FCP-NEXT: vpalignr {{.*#+}} xmm8 = mem[8,9,10,11,12,13,14,15],xmm5[0,1,2,3,4,5,6,7]5026; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm8 = ymm8[0,1,2,3],ymm7[4,5,6,7]5027; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5028; AVX2-FCP-NEXT: vpalignr {{.*#+}} ymm7 = ymm1[8,9,10,11,12,13,14,15],ymm12[0,1,2,3,4,5,6,7],ymm1[24,25,26,27,28,29,30,31],ymm12[16,17,18,19,20,21,22,23]5029; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,1,0,3]5030; AVX2-FCP-NEXT: vmovdqa 64(%rdi), %xmm05031; AVX2-FCP-NEXT: vpalignr {{.*#+}} xmm15 = mem[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]5032; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm7 = ymm15[0,1,2,3],ymm7[4,5,6,7]5033; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm12 = ymm1[0,1,2,3,4,5],ymm12[6,7]5034; AVX2-FCP-NEXT: vpblendd $3, (%rsp), %xmm0, %xmm0 # 16-byte Folded Reload5035; AVX2-FCP-NEXT: # xmm0 = mem[0,1],xmm0[2,3]5036; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm12 = ymm0[0,1,2,3],ymm12[4,5,6,7]5037; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5038; AVX2-FCP-NEXT: vpblendd $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload5039; AVX2-FCP-NEXT: # ymm0 = mem[0,1,2,3,4,5],ymm0[6,7]5040; AVX2-FCP-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4, %xmm4 # 16-byte Folded Reload5041; AVX2-FCP-NEXT: # xmm4 = mem[0,1],xmm4[2,3]5042; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm0[4,5,6,7]5043; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm3[0,1,2,3,4,5],ymm2[6,7]5044; AVX2-FCP-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm5, %xmm5 # 16-byte Folded Reload5045; AVX2-FCP-NEXT: # xmm5 = mem[0,1],xmm5[2,3]5046; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm5 = ymm5[0,1,2,3],ymm0[4,5,6,7]5047; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5048; AVX2-FCP-NEXT: vpblendd $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload5049; AVX2-FCP-NEXT: # ymm0 = mem[0,1,2,3,4,5],ymm0[6,7]5050; AVX2-FCP-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm6, %xmm6 # 16-byte Folded Reload5051; AVX2-FCP-NEXT: # xmm6 = mem[0,1],xmm6[2,3]5052; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm15 = ymm6[0,1,2,3],ymm0[4,5,6,7]5053; AVX2-FCP-NEXT: vpblendd $192, {{[-0-9]+}}(%r{{[sb]}}p), %ymm10, %ymm0 # 32-byte Folded Reload5054; AVX2-FCP-NEXT: # ymm0 = ymm10[0,1,2,3,4,5],mem[6,7]5055; AVX2-FCP-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm9, %xmm1 # 16-byte Folded Reload5056; AVX2-FCP-NEXT: # xmm1 = mem[0,1],xmm9[2,3]5057; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm1[0,1,2,3],ymm0[4,5,6,7]5058; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5059; AVX2-FCP-NEXT: vblendps $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload5060; AVX2-FCP-NEXT: # ymm0 = mem[0,1,2,3,4,5],ymm0[6,7]5061; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload5062; AVX2-FCP-NEXT: vblendps $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload5063; AVX2-FCP-NEXT: # xmm1 = mem[0,1],xmm1[2,3]5064; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm2 = ymm1[0,1,2,3],ymm0[4,5,6,7]5065; AVX2-FCP-NEXT: vpblendd $192, {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm0 # 32-byte Folded Reload5066; AVX2-FCP-NEXT: # ymm0 = ymm13[0,1,2,3,4,5],mem[6,7]5067; AVX2-FCP-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm11, %xmm1 # 16-byte Folded Reload5068; AVX2-FCP-NEXT: # xmm1 = mem[0,1],xmm11[2,3]5069; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm9 = ymm1[0,1,2,3],ymm0[4,5,6,7]5070; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload5071; AVX2-FCP-NEXT: vpblendd $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload5072; AVX2-FCP-NEXT: # ymm0 = mem[0,1,2,3,4,5],ymm0[6,7]5073; AVX2-FCP-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm14, %xmm1 # 16-byte Folded Reload5074; AVX2-FCP-NEXT: # xmm1 = mem[0,1],xmm14[2,3]5075; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]5076; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5077; AVX2-FCP-NEXT: vmovaps %ymm1, 192(%rsi)5078; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5079; AVX2-FCP-NEXT: vmovaps %ymm1, 128(%rsi)5080; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5081; AVX2-FCP-NEXT: vmovaps %ymm1, 64(%rsi)5082; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5083; AVX2-FCP-NEXT: vmovaps %ymm1, (%rsi)5084; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5085; AVX2-FCP-NEXT: vmovaps %ymm1, 224(%rsi)5086; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5087; AVX2-FCP-NEXT: vmovaps %ymm1, 160(%rsi)5088; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5089; AVX2-FCP-NEXT: vmovaps %ymm1, 96(%rsi)5090; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5091; AVX2-FCP-NEXT: vmovaps %ymm1, 32(%rsi)5092; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5093; AVX2-FCP-NEXT: vmovaps %ymm1, 192(%rdx)5094; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5095; AVX2-FCP-NEXT: vmovaps %ymm1, 128(%rdx)5096; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5097; AVX2-FCP-NEXT: vmovaps %ymm1, 64(%rdx)5098; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5099; AVX2-FCP-NEXT: vmovaps %ymm1, (%rdx)5100; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5101; AVX2-FCP-NEXT: vmovaps %ymm1, 224(%rdx)5102; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5103; AVX2-FCP-NEXT: vmovaps %ymm1, 160(%rdx)5104; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5105; AVX2-FCP-NEXT: vmovaps %ymm1, 96(%rdx)5106; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5107; AVX2-FCP-NEXT: vmovaps %ymm1, 32(%rdx)5108; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5109; AVX2-FCP-NEXT: vmovaps %ymm1, 192(%rcx)5110; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5111; AVX2-FCP-NEXT: vmovaps %ymm1, 128(%rcx)5112; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5113; AVX2-FCP-NEXT: vmovaps %ymm1, 64(%rcx)5114; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5115; AVX2-FCP-NEXT: vmovaps %ymm1, (%rcx)5116; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5117; AVX2-FCP-NEXT: vmovaps %ymm1, 224(%rcx)5118; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5119; AVX2-FCP-NEXT: vmovaps %ymm1, 160(%rcx)5120; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5121; AVX2-FCP-NEXT: vmovaps %ymm1, 96(%rcx)5122; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5123; AVX2-FCP-NEXT: vmovaps %ymm1, 32(%rcx)5124; AVX2-FCP-NEXT: vmovdqa %ymm7, (%r8)5125; AVX2-FCP-NEXT: vmovdqa %ymm8, 64(%r8)5126; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5127; AVX2-FCP-NEXT: vmovaps %ymm1, 128(%r8)5128; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5129; AVX2-FCP-NEXT: vmovaps %ymm1, 192(%r8)5130; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5131; AVX2-FCP-NEXT: vmovaps %ymm1, 224(%r8)5132; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5133; AVX2-FCP-NEXT: vmovaps %ymm1, 160(%r8)5134; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5135; AVX2-FCP-NEXT: vmovaps %ymm1, 96(%r8)5136; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload5137; AVX2-FCP-NEXT: vmovaps %ymm1, 32(%r8)5138; AVX2-FCP-NEXT: vmovdqa %ymm0, 224(%r9)5139; AVX2-FCP-NEXT: vmovdqa %ymm9, 192(%r9)5140; AVX2-FCP-NEXT: vmovaps %ymm2, 160(%r9)5141; AVX2-FCP-NEXT: vmovdqa %ymm6, 128(%r9)5142; AVX2-FCP-NEXT: vmovdqa %ymm15, 96(%r9)5143; AVX2-FCP-NEXT: vmovdqa %ymm5, 64(%r9)5144; AVX2-FCP-NEXT: vmovdqa %ymm4, 32(%r9)5145; AVX2-FCP-NEXT: vmovdqa %ymm12, (%r9)5146; AVX2-FCP-NEXT: addq $1464, %rsp # imm = 0x5B85147; AVX2-FCP-NEXT: vzeroupper5148; AVX2-FCP-NEXT: retq5149;5150; AVX512-LABEL: load_i64_stride5_vf32:5151; AVX512: # %bb.0:5152; AVX512-NEXT: subq $648, %rsp # imm = 0x2885153; AVX512-NEXT: vmovdqa64 1088(%rdi), %zmm215154; AVX512-NEXT: vmovdqa64 1152(%rdi), %zmm15155; AVX512-NEXT: vmovdqa64 768(%rdi), %zmm185156; AVX512-NEXT: vmovdqa64 832(%rdi), %zmm05157; AVX512-NEXT: vmovdqa64 384(%rdi), %zmm175158; AVX512-NEXT: vmovdqa64 320(%rdi), %zmm195159; AVX512-NEXT: vmovdqa64 448(%rdi), %zmm265160; AVX512-NEXT: vmovdqa64 512(%rdi), %zmm25161; AVX512-NEXT: vmovdqa64 (%rdi), %zmm165162; AVX512-NEXT: vmovdqa64 128(%rdi), %zmm255163; AVX512-NEXT: vmovdqa64 192(%rdi), %zmm35164; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm11 = [12,1,6,0,12,1,6,0]5165; AVX512-NEXT: # zmm11 = mem[0,1,2,3,0,1,2,3]5166; AVX512-NEXT: vmovdqa64 %zmm3, %zmm105167; AVX512-NEXT: vpermt2q %zmm25, %zmm11, %zmm105168; AVX512-NEXT: vpmovsxbq {{.*#+}} ymm24 = [0,5,10,15]5169; AVX512-NEXT: vmovdqa64 %zmm2, %zmm125170; AVX512-NEXT: vmovdqa64 %zmm2, %zmm45171; AVX512-NEXT: vpermt2q %zmm26, %zmm11, %zmm125172; AVX512-NEXT: vmovdqa64 %zmm19, %zmm25173; AVX512-NEXT: vpermt2q %zmm17, %zmm24, %zmm25174; AVX512-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill5175; AVX512-NEXT: vmovdqa64 %zmm0, %zmm55176; AVX512-NEXT: vmovdqa64 %zmm0, %zmm25177; AVX512-NEXT: vpermt2q %zmm18, %zmm11, %zmm55178; AVX512-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill5179; AVX512-NEXT: vmovdqa64 %zmm1, %zmm05180; AVX512-NEXT: vpermt2q %zmm21, %zmm11, %zmm05181; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill5182; AVX512-NEXT: vpmovsxbq {{.*#+}} ymm28 = [1,6,11,0]5183; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm15 = [5,10,15,0,5,10,15,0]5184; AVX512-NEXT: # zmm15 = mem[0,1,2,3,0,1,2,3]5185; AVX512-NEXT: vmovdqa64 %zmm25, %zmm305186; AVX512-NEXT: vpermt2q %zmm3, %zmm15, %zmm305187; AVX512-NEXT: vmovdqa64 %zmm18, %zmm275188; AVX512-NEXT: vpermt2q %zmm2, %zmm15, %zmm275189; AVX512-NEXT: vmovdqa64 %zmm26, %zmm225190; AVX512-NEXT: vpermt2q %zmm4, %zmm15, %zmm225191; AVX512-NEXT: vmovdqa64 %zmm19, %zmm05192; AVX512-NEXT: vpermt2q %zmm17, %zmm28, %zmm05193; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill5194; AVX512-NEXT: vpermi2q %zmm1, %zmm21, %zmm155195; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm23 = [6,11,0,1,6,11,0,1]5196; AVX512-NEXT: # zmm23 = mem[0,1,2,3,0,1,2,3]5197; AVX512-NEXT: vmovdqa64 %zmm25, %zmm05198; AVX512-NEXT: vpermt2q %zmm3, %zmm23, %zmm05199; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill5200; AVX512-NEXT: vpmovsxbq {{.*#+}} ymm13 = [2,7,12,0]5201; AVX512-NEXT: vmovdqa64 %zmm18, %zmm05202; AVX512-NEXT: vpermt2q %zmm2, %zmm23, %zmm05203; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill5204; AVX512-NEXT: vmovdqa64 %zmm26, %zmm05205; AVX512-NEXT: vpermt2q %zmm4, %zmm23, %zmm05206; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill5207; AVX512-NEXT: vpermi2q %zmm1, %zmm21, %zmm235208; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm29 = [7,12,0,2,7,12,0,2]5209; AVX512-NEXT: # zmm29 = mem[0,1,2,3,0,1,2,3]5210; AVX512-NEXT: vmovdqa64 %zmm26, %zmm05211; AVX512-NEXT: vpermt2q %zmm4, %zmm29, %zmm05212; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill5213; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm0 = [0,5,0,11,0,5,0,11]5214; AVX512-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3]5215; AVX512-NEXT: vpermt2q %zmm26, %zmm0, %zmm45216; AVX512-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill5217; AVX512-NEXT: vmovdqa64 %zmm25, %zmm265218; AVX512-NEXT: vpermt2q %zmm3, %zmm29, %zmm265219; AVX512-NEXT: vpermt2q %zmm25, %zmm0, %zmm35220; AVX512-NEXT: vmovdqu64 %zmm3, (%rsp) # 64-byte Spill5221; AVX512-NEXT: vmovdqa64 %zmm18, %zmm205222; AVX512-NEXT: vpermt2q %zmm2, %zmm29, %zmm205223; AVX512-NEXT: vpermi2q %zmm1, %zmm21, %zmm295224; AVX512-NEXT: vpermt2q %zmm21, %zmm0, %zmm15225; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill5226; AVX512-NEXT: vmovdqa64 %zmm19, %zmm215227; AVX512-NEXT: vpermt2q %zmm17, %zmm13, %zmm215228; AVX512-NEXT: vpermt2q %zmm18, %zmm0, %zmm25229; AVX512-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill5230; AVX512-NEXT: vpmovsxbq {{.*#+}} ymm31 = [11,0,5,0]5231; AVX512-NEXT: vmovdqa64 %zmm17, %zmm255232; AVX512-NEXT: vpermt2q %zmm19, %zmm31, %zmm255233; AVX512-NEXT: vpermt2q %zmm19, %zmm11, %zmm175234; AVX512-NEXT: vmovdqa64 64(%rdi), %zmm185235; AVX512-NEXT: vmovdqa64 %zmm16, %zmm65236; AVX512-NEXT: vpermt2q %zmm18, %zmm24, %zmm65237; AVX512-NEXT: vmovdqa64 %zmm16, %zmm95238; AVX512-NEXT: vpermt2q %zmm18, %zmm28, %zmm95239; AVX512-NEXT: vmovdqa64 %zmm16, %zmm85240; AVX512-NEXT: vpermt2q %zmm18, %zmm13, %zmm85241; AVX512-NEXT: vmovdqa64 %zmm18, %zmm195242; AVX512-NEXT: vpermt2q %zmm16, %zmm31, %zmm195243; AVX512-NEXT: vpermt2q %zmm16, %zmm11, %zmm185244; AVX512-NEXT: vmovdqa64 704(%rdi), %zmm165245; AVX512-NEXT: vmovdqa64 640(%rdi), %zmm15246; AVX512-NEXT: vmovdqa64 %zmm1, %zmm35247; AVX512-NEXT: vpermt2q %zmm16, %zmm24, %zmm35248; AVX512-NEXT: vmovdqa64 1024(%rdi), %zmm25249; AVX512-NEXT: vmovdqa64 960(%rdi), %zmm05250; AVX512-NEXT: vpermi2q %zmm2, %zmm0, %zmm245251; AVX512-NEXT: vmovdqa64 %zmm1, %zmm75252; AVX512-NEXT: vpermt2q %zmm16, %zmm28, %zmm75253; AVX512-NEXT: vpermi2q %zmm2, %zmm0, %zmm285254; AVX512-NEXT: vmovdqa64 %zmm1, %zmm45255; AVX512-NEXT: vpermt2q %zmm16, %zmm13, %zmm45256; AVX512-NEXT: vpermi2q %zmm2, %zmm0, %zmm135257; AVX512-NEXT: vmovdqa64 %zmm16, %zmm55258; AVX512-NEXT: vpermt2q %zmm1, %zmm31, %zmm55259; AVX512-NEXT: vpermi2q %zmm0, %zmm2, %zmm315260; AVX512-NEXT: vpermt2q %zmm0, %zmm11, %zmm25261; AVX512-NEXT: vpermt2q %zmm1, %zmm11, %zmm165262; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm14 = zmm6[0,1,2,3],zmm10[4,5,6,7]5263; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload5264; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm12 = zmm0[0,1,2,3],zmm12[4,5,6,7]5265; AVX512-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3, %zmm10 # 64-byte Folded Reload5266; AVX512-NEXT: # zmm10 = zmm3[0,1,2,3],mem[4,5,6,7]5267; AVX512-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm24, %zmm6 # 64-byte Folded Reload5268; AVX512-NEXT: # zmm6 = zmm24[0,1,2,3],mem[4,5,6,7]5269; AVX512-NEXT: vmovdqa64 256(%rdi), %zmm115270; AVX512-NEXT: vpmovsxbq {{.*#+}} zmm24 = [0,1,2,3,4,5,6,11]5271; AVX512-NEXT: vpermt2q %zmm11, %zmm24, %zmm145272; AVX512-NEXT: vmovdqa64 576(%rdi), %zmm05273; AVX512-NEXT: vpermt2q %zmm0, %zmm24, %zmm125274; AVX512-NEXT: vmovdqa64 896(%rdi), %zmm35275; AVX512-NEXT: vpermt2q %zmm3, %zmm24, %zmm105276; AVX512-NEXT: vmovdqa64 1216(%rdi), %zmm15277; AVX512-NEXT: vpermt2q %zmm1, %zmm24, %zmm65278; AVX512-NEXT: movb $7, %al5279; AVX512-NEXT: kmovw %eax, %k15280; AVX512-NEXT: vmovdqa64 %zmm9, %zmm30 {%k1}5281; AVX512-NEXT: vmovdqa64 %zmm7, %zmm27 {%k1}5282; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload5283; AVX512-NEXT: vmovdqa64 %zmm7, %zmm22 {%k1}5284; AVX512-NEXT: vmovdqa64 %zmm28, %zmm15 {%k1}5285; AVX512-NEXT: vpmovsxbq {{.*#+}} zmm7 = [0,1,2,3,4,5,6,12]5286; AVX512-NEXT: vpermt2q %zmm11, %zmm7, %zmm305287; AVX512-NEXT: vpermt2q %zmm3, %zmm7, %zmm275288; AVX512-NEXT: vpermt2q %zmm0, %zmm7, %zmm225289; AVX512-NEXT: vpermt2q %zmm1, %zmm7, %zmm155290; AVX512-NEXT: movb $56, %al5291; AVX512-NEXT: kmovw %eax, %k15292; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload5293; AVX512-NEXT: vmovdqa64 %zmm7, %zmm8 {%k1}5294; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload5295; AVX512-NEXT: vmovdqa64 %zmm7, %zmm4 {%k1}5296; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload5297; AVX512-NEXT: vmovdqa64 %zmm7, %zmm21 {%k1}5298; AVX512-NEXT: vmovdqa64 %zmm23, %zmm13 {%k1}5299; AVX512-NEXT: vpmovsxbq {{.*#+}} zmm7 = [0,1,2,3,4,5,8,13]5300; AVX512-NEXT: vpermt2q %zmm11, %zmm7, %zmm85301; AVX512-NEXT: vpermt2q %zmm3, %zmm7, %zmm45302; AVX512-NEXT: vpermt2q %zmm0, %zmm7, %zmm215303; AVX512-NEXT: vpermt2q %zmm1, %zmm7, %zmm135304; AVX512-NEXT: vmovdqa64 %zmm26, %zmm19 {%k1}5305; AVX512-NEXT: vmovdqa64 %zmm20, %zmm5 {%k1}5306; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload5307; AVX512-NEXT: vmovdqa64 %zmm7, %zmm25 {%k1}5308; AVX512-NEXT: vmovdqa64 %zmm29, %zmm31 {%k1}5309; AVX512-NEXT: vpmovsxbq {{.*#+}} zmm7 = [0,1,2,3,4,5,9,14]5310; AVX512-NEXT: vpermt2q %zmm11, %zmm7, %zmm195311; AVX512-NEXT: vpermt2q %zmm3, %zmm7, %zmm55312; AVX512-NEXT: vpermt2q %zmm0, %zmm7, %zmm255313; AVX512-NEXT: vpermt2q %zmm1, %zmm7, %zmm315314; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload5315; AVX512-NEXT: vmovdqa64 %zmm7, %zmm17 {%k1}5316; AVX512-NEXT: vpmovsxbq {{.*#+}} zmm7 = [0,1,2,3,4,5,10,15]5317; AVX512-NEXT: vpermt2q %zmm0, %zmm7, %zmm175318; AVX512-NEXT: vmovdqu64 (%rsp), %zmm0 # 64-byte Reload5319; AVX512-NEXT: vmovdqa64 %zmm0, %zmm18 {%k1}5320; AVX512-NEXT: vpermt2q %zmm11, %zmm7, %zmm185321; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload5322; AVX512-NEXT: vmovdqa64 %zmm0, %zmm2 {%k1}5323; AVX512-NEXT: vpermt2q %zmm1, %zmm7, %zmm25324; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload5325; AVX512-NEXT: vmovdqa64 %zmm0, %zmm16 {%k1}5326; AVX512-NEXT: vpermt2q %zmm3, %zmm7, %zmm165327; AVX512-NEXT: vmovdqa64 %zmm6, 192(%rsi)5328; AVX512-NEXT: vmovdqa64 %zmm10, 128(%rsi)5329; AVX512-NEXT: vmovdqa64 %zmm12, 64(%rsi)5330; AVX512-NEXT: vmovdqa64 %zmm14, (%rsi)5331; AVX512-NEXT: vmovdqa64 %zmm15, 192(%rdx)5332; AVX512-NEXT: vmovdqa64 %zmm30, (%rdx)5333; AVX512-NEXT: vmovdqa64 %zmm22, 64(%rdx)5334; AVX512-NEXT: vmovdqa64 %zmm27, 128(%rdx)5335; AVX512-NEXT: vmovdqa64 %zmm13, 192(%rcx)5336; AVX512-NEXT: vmovdqa64 %zmm8, (%rcx)5337; AVX512-NEXT: vmovdqa64 %zmm21, 64(%rcx)5338; AVX512-NEXT: vmovdqa64 %zmm4, 128(%rcx)5339; AVX512-NEXT: vmovdqa64 %zmm31, 192(%r8)5340; AVX512-NEXT: vmovdqa64 %zmm19, (%r8)5341; AVX512-NEXT: vmovdqa64 %zmm25, 64(%r8)5342; AVX512-NEXT: vmovdqa64 %zmm5, 128(%r8)5343; AVX512-NEXT: vmovdqa64 %zmm16, 128(%r9)5344; AVX512-NEXT: vmovdqa64 %zmm2, 192(%r9)5345; AVX512-NEXT: vmovdqa64 %zmm18, (%r9)5346; AVX512-NEXT: vmovdqa64 %zmm17, 64(%r9)5347; AVX512-NEXT: addq $648, %rsp # imm = 0x2885348; AVX512-NEXT: vzeroupper5349; AVX512-NEXT: retq5350;5351; AVX512-FCP-LABEL: load_i64_stride5_vf32:5352; AVX512-FCP: # %bb.0:5353; AVX512-FCP-NEXT: subq $648, %rsp # imm = 0x2885354; AVX512-FCP-NEXT: vmovdqa64 1088(%rdi), %zmm215355; AVX512-FCP-NEXT: vmovdqa64 1152(%rdi), %zmm15356; AVX512-FCP-NEXT: vmovdqa64 768(%rdi), %zmm185357; AVX512-FCP-NEXT: vmovdqa64 832(%rdi), %zmm05358; AVX512-FCP-NEXT: vmovdqa64 384(%rdi), %zmm175359; AVX512-FCP-NEXT: vmovdqa64 320(%rdi), %zmm195360; AVX512-FCP-NEXT: vmovdqa64 448(%rdi), %zmm265361; AVX512-FCP-NEXT: vmovdqa64 512(%rdi), %zmm25362; AVX512-FCP-NEXT: vmovdqa64 (%rdi), %zmm165363; AVX512-FCP-NEXT: vmovdqa64 128(%rdi), %zmm255364; AVX512-FCP-NEXT: vmovdqa64 192(%rdi), %zmm35365; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm11 = [12,1,6,0,12,1,6,0]5366; AVX512-FCP-NEXT: # zmm11 = mem[0,1,2,3,0,1,2,3]5367; AVX512-FCP-NEXT: vmovdqa64 %zmm3, %zmm105368; AVX512-FCP-NEXT: vpermt2q %zmm25, %zmm11, %zmm105369; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} ymm24 = [0,5,10,15]5370; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm125371; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm45372; AVX512-FCP-NEXT: vpermt2q %zmm26, %zmm11, %zmm125373; AVX512-FCP-NEXT: vmovdqa64 %zmm19, %zmm25374; AVX512-FCP-NEXT: vpermt2q %zmm17, %zmm24, %zmm25375; AVX512-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill5376; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm55377; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm25378; AVX512-FCP-NEXT: vpermt2q %zmm18, %zmm11, %zmm55379; AVX512-FCP-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill5380; AVX512-FCP-NEXT: vmovdqa64 %zmm1, %zmm05381; AVX512-FCP-NEXT: vpermt2q %zmm21, %zmm11, %zmm05382; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill5383; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} ymm28 = [1,6,11,0]5384; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm15 = [5,10,15,0,5,10,15,0]5385; AVX512-FCP-NEXT: # zmm15 = mem[0,1,2,3,0,1,2,3]5386; AVX512-FCP-NEXT: vmovdqa64 %zmm25, %zmm305387; AVX512-FCP-NEXT: vpermt2q %zmm3, %zmm15, %zmm305388; AVX512-FCP-NEXT: vmovdqa64 %zmm18, %zmm275389; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm15, %zmm275390; AVX512-FCP-NEXT: vmovdqa64 %zmm26, %zmm225391; AVX512-FCP-NEXT: vpermt2q %zmm4, %zmm15, %zmm225392; AVX512-FCP-NEXT: vmovdqa64 %zmm19, %zmm05393; AVX512-FCP-NEXT: vpermt2q %zmm17, %zmm28, %zmm05394; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill5395; AVX512-FCP-NEXT: vpermi2q %zmm1, %zmm21, %zmm155396; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm23 = [6,11,0,1,6,11,0,1]5397; AVX512-FCP-NEXT: # zmm23 = mem[0,1,2,3,0,1,2,3]5398; AVX512-FCP-NEXT: vmovdqa64 %zmm25, %zmm05399; AVX512-FCP-NEXT: vpermt2q %zmm3, %zmm23, %zmm05400; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill5401; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} ymm13 = [2,7,12,0]5402; AVX512-FCP-NEXT: vmovdqa64 %zmm18, %zmm05403; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm23, %zmm05404; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill5405; AVX512-FCP-NEXT: vmovdqa64 %zmm26, %zmm05406; AVX512-FCP-NEXT: vpermt2q %zmm4, %zmm23, %zmm05407; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill5408; AVX512-FCP-NEXT: vpermi2q %zmm1, %zmm21, %zmm235409; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm29 = [7,12,0,2,7,12,0,2]5410; AVX512-FCP-NEXT: # zmm29 = mem[0,1,2,3,0,1,2,3]5411; AVX512-FCP-NEXT: vmovdqa64 %zmm26, %zmm05412; AVX512-FCP-NEXT: vpermt2q %zmm4, %zmm29, %zmm05413; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill5414; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm0 = [0,5,0,11,0,5,0,11]5415; AVX512-FCP-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3]5416; AVX512-FCP-NEXT: vpermt2q %zmm26, %zmm0, %zmm45417; AVX512-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill5418; AVX512-FCP-NEXT: vmovdqa64 %zmm25, %zmm265419; AVX512-FCP-NEXT: vpermt2q %zmm3, %zmm29, %zmm265420; AVX512-FCP-NEXT: vpermt2q %zmm25, %zmm0, %zmm35421; AVX512-FCP-NEXT: vmovdqu64 %zmm3, (%rsp) # 64-byte Spill5422; AVX512-FCP-NEXT: vmovdqa64 %zmm18, %zmm205423; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm29, %zmm205424; AVX512-FCP-NEXT: vpermi2q %zmm1, %zmm21, %zmm295425; AVX512-FCP-NEXT: vpermt2q %zmm21, %zmm0, %zmm15426; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill5427; AVX512-FCP-NEXT: vmovdqa64 %zmm19, %zmm215428; AVX512-FCP-NEXT: vpermt2q %zmm17, %zmm13, %zmm215429; AVX512-FCP-NEXT: vpermt2q %zmm18, %zmm0, %zmm25430; AVX512-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill5431; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} ymm31 = [11,0,5,0]5432; AVX512-FCP-NEXT: vmovdqa64 %zmm17, %zmm255433; AVX512-FCP-NEXT: vpermt2q %zmm19, %zmm31, %zmm255434; AVX512-FCP-NEXT: vpermt2q %zmm19, %zmm11, %zmm175435; AVX512-FCP-NEXT: vmovdqa64 64(%rdi), %zmm185436; AVX512-FCP-NEXT: vmovdqa64 %zmm16, %zmm65437; AVX512-FCP-NEXT: vpermt2q %zmm18, %zmm24, %zmm65438; AVX512-FCP-NEXT: vmovdqa64 %zmm16, %zmm95439; AVX512-FCP-NEXT: vpermt2q %zmm18, %zmm28, %zmm95440; AVX512-FCP-NEXT: vmovdqa64 %zmm16, %zmm85441; AVX512-FCP-NEXT: vpermt2q %zmm18, %zmm13, %zmm85442; AVX512-FCP-NEXT: vmovdqa64 %zmm18, %zmm195443; AVX512-FCP-NEXT: vpermt2q %zmm16, %zmm31, %zmm195444; AVX512-FCP-NEXT: vpermt2q %zmm16, %zmm11, %zmm185445; AVX512-FCP-NEXT: vmovdqa64 704(%rdi), %zmm165446; AVX512-FCP-NEXT: vmovdqa64 640(%rdi), %zmm15447; AVX512-FCP-NEXT: vmovdqa64 %zmm1, %zmm35448; AVX512-FCP-NEXT: vpermt2q %zmm16, %zmm24, %zmm35449; AVX512-FCP-NEXT: vmovdqa64 1024(%rdi), %zmm25450; AVX512-FCP-NEXT: vmovdqa64 960(%rdi), %zmm05451; AVX512-FCP-NEXT: vpermi2q %zmm2, %zmm0, %zmm245452; AVX512-FCP-NEXT: vmovdqa64 %zmm1, %zmm75453; AVX512-FCP-NEXT: vpermt2q %zmm16, %zmm28, %zmm75454; AVX512-FCP-NEXT: vpermi2q %zmm2, %zmm0, %zmm285455; AVX512-FCP-NEXT: vmovdqa64 %zmm1, %zmm45456; AVX512-FCP-NEXT: vpermt2q %zmm16, %zmm13, %zmm45457; AVX512-FCP-NEXT: vpermi2q %zmm2, %zmm0, %zmm135458; AVX512-FCP-NEXT: vmovdqa64 %zmm16, %zmm55459; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm31, %zmm55460; AVX512-FCP-NEXT: vpermi2q %zmm0, %zmm2, %zmm315461; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm11, %zmm25462; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm11, %zmm165463; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm14 = zmm6[0,1,2,3],zmm10[4,5,6,7]5464; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload5465; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm12 = zmm0[0,1,2,3],zmm12[4,5,6,7]5466; AVX512-FCP-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3, %zmm10 # 64-byte Folded Reload5467; AVX512-FCP-NEXT: # zmm10 = zmm3[0,1,2,3],mem[4,5,6,7]5468; AVX512-FCP-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm24, %zmm6 # 64-byte Folded Reload5469; AVX512-FCP-NEXT: # zmm6 = zmm24[0,1,2,3],mem[4,5,6,7]5470; AVX512-FCP-NEXT: vmovdqa64 256(%rdi), %zmm115471; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm24 = [0,1,2,3,4,5,6,11]5472; AVX512-FCP-NEXT: vpermt2q %zmm11, %zmm24, %zmm145473; AVX512-FCP-NEXT: vmovdqa64 576(%rdi), %zmm05474; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm24, %zmm125475; AVX512-FCP-NEXT: vmovdqa64 896(%rdi), %zmm35476; AVX512-FCP-NEXT: vpermt2q %zmm3, %zmm24, %zmm105477; AVX512-FCP-NEXT: vmovdqa64 1216(%rdi), %zmm15478; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm24, %zmm65479; AVX512-FCP-NEXT: movb $7, %al5480; AVX512-FCP-NEXT: kmovw %eax, %k15481; AVX512-FCP-NEXT: vmovdqa64 %zmm9, %zmm30 {%k1}5482; AVX512-FCP-NEXT: vmovdqa64 %zmm7, %zmm27 {%k1}5483; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload5484; AVX512-FCP-NEXT: vmovdqa64 %zmm7, %zmm22 {%k1}5485; AVX512-FCP-NEXT: vmovdqa64 %zmm28, %zmm15 {%k1}5486; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm7 = [0,1,2,3,4,5,6,12]5487; AVX512-FCP-NEXT: vpermt2q %zmm11, %zmm7, %zmm305488; AVX512-FCP-NEXT: vpermt2q %zmm3, %zmm7, %zmm275489; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm7, %zmm225490; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm7, %zmm155491; AVX512-FCP-NEXT: movb $56, %al5492; AVX512-FCP-NEXT: kmovw %eax, %k15493; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload5494; AVX512-FCP-NEXT: vmovdqa64 %zmm7, %zmm8 {%k1}5495; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload5496; AVX512-FCP-NEXT: vmovdqa64 %zmm7, %zmm4 {%k1}5497; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload5498; AVX512-FCP-NEXT: vmovdqa64 %zmm7, %zmm21 {%k1}5499; AVX512-FCP-NEXT: vmovdqa64 %zmm23, %zmm13 {%k1}5500; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm7 = [0,1,2,3,4,5,8,13]5501; AVX512-FCP-NEXT: vpermt2q %zmm11, %zmm7, %zmm85502; AVX512-FCP-NEXT: vpermt2q %zmm3, %zmm7, %zmm45503; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm7, %zmm215504; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm7, %zmm135505; AVX512-FCP-NEXT: vmovdqa64 %zmm26, %zmm19 {%k1}5506; AVX512-FCP-NEXT: vmovdqa64 %zmm20, %zmm5 {%k1}5507; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload5508; AVX512-FCP-NEXT: vmovdqa64 %zmm7, %zmm25 {%k1}5509; AVX512-FCP-NEXT: vmovdqa64 %zmm29, %zmm31 {%k1}5510; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm7 = [0,1,2,3,4,5,9,14]5511; AVX512-FCP-NEXT: vpermt2q %zmm11, %zmm7, %zmm195512; AVX512-FCP-NEXT: vpermt2q %zmm3, %zmm7, %zmm55513; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm7, %zmm255514; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm7, %zmm315515; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload5516; AVX512-FCP-NEXT: vmovdqa64 %zmm7, %zmm17 {%k1}5517; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm7 = [0,1,2,3,4,5,10,15]5518; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm7, %zmm175519; AVX512-FCP-NEXT: vmovdqu64 (%rsp), %zmm0 # 64-byte Reload5520; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm18 {%k1}5521; AVX512-FCP-NEXT: vpermt2q %zmm11, %zmm7, %zmm185522; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload5523; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm2 {%k1}5524; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm7, %zmm25525; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload5526; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm16 {%k1}5527; AVX512-FCP-NEXT: vpermt2q %zmm3, %zmm7, %zmm165528; AVX512-FCP-NEXT: vmovdqa64 %zmm6, 192(%rsi)5529; AVX512-FCP-NEXT: vmovdqa64 %zmm10, 128(%rsi)5530; AVX512-FCP-NEXT: vmovdqa64 %zmm12, 64(%rsi)5531; AVX512-FCP-NEXT: vmovdqa64 %zmm14, (%rsi)5532; AVX512-FCP-NEXT: vmovdqa64 %zmm15, 192(%rdx)5533; AVX512-FCP-NEXT: vmovdqa64 %zmm30, (%rdx)5534; AVX512-FCP-NEXT: vmovdqa64 %zmm22, 64(%rdx)5535; AVX512-FCP-NEXT: vmovdqa64 %zmm27, 128(%rdx)5536; AVX512-FCP-NEXT: vmovdqa64 %zmm13, 192(%rcx)5537; AVX512-FCP-NEXT: vmovdqa64 %zmm8, (%rcx)5538; AVX512-FCP-NEXT: vmovdqa64 %zmm21, 64(%rcx)5539; AVX512-FCP-NEXT: vmovdqa64 %zmm4, 128(%rcx)5540; AVX512-FCP-NEXT: vmovdqa64 %zmm31, 192(%r8)5541; AVX512-FCP-NEXT: vmovdqa64 %zmm19, (%r8)5542; AVX512-FCP-NEXT: vmovdqa64 %zmm25, 64(%r8)5543; AVX512-FCP-NEXT: vmovdqa64 %zmm5, 128(%r8)5544; AVX512-FCP-NEXT: vmovdqa64 %zmm16, 128(%r9)5545; AVX512-FCP-NEXT: vmovdqa64 %zmm2, 192(%r9)5546; AVX512-FCP-NEXT: vmovdqa64 %zmm18, (%r9)5547; AVX512-FCP-NEXT: vmovdqa64 %zmm17, 64(%r9)5548; AVX512-FCP-NEXT: addq $648, %rsp # imm = 0x2885549; AVX512-FCP-NEXT: vzeroupper5550; AVX512-FCP-NEXT: retq5551;5552; AVX512DQ-LABEL: load_i64_stride5_vf32:5553; AVX512DQ: # %bb.0:5554; AVX512DQ-NEXT: subq $648, %rsp # imm = 0x2885555; AVX512DQ-NEXT: vmovdqa64 1088(%rdi), %zmm215556; AVX512DQ-NEXT: vmovdqa64 1152(%rdi), %zmm15557; AVX512DQ-NEXT: vmovdqa64 768(%rdi), %zmm185558; AVX512DQ-NEXT: vmovdqa64 832(%rdi), %zmm05559; AVX512DQ-NEXT: vmovdqa64 384(%rdi), %zmm175560; AVX512DQ-NEXT: vmovdqa64 320(%rdi), %zmm195561; AVX512DQ-NEXT: vmovdqa64 448(%rdi), %zmm265562; AVX512DQ-NEXT: vmovdqa64 512(%rdi), %zmm25563; AVX512DQ-NEXT: vmovdqa64 (%rdi), %zmm165564; AVX512DQ-NEXT: vmovdqa64 128(%rdi), %zmm255565; AVX512DQ-NEXT: vmovdqa64 192(%rdi), %zmm35566; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm11 = [12,1,6,0,12,1,6,0]5567; AVX512DQ-NEXT: # zmm11 = mem[0,1,2,3,0,1,2,3]5568; AVX512DQ-NEXT: vmovdqa64 %zmm3, %zmm105569; AVX512DQ-NEXT: vpermt2q %zmm25, %zmm11, %zmm105570; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} ymm24 = [0,5,10,15]5571; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm125572; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm45573; AVX512DQ-NEXT: vpermt2q %zmm26, %zmm11, %zmm125574; AVX512DQ-NEXT: vmovdqa64 %zmm19, %zmm25575; AVX512DQ-NEXT: vpermt2q %zmm17, %zmm24, %zmm25576; AVX512DQ-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill5577; AVX512DQ-NEXT: vmovdqa64 %zmm0, %zmm55578; AVX512DQ-NEXT: vmovdqa64 %zmm0, %zmm25579; AVX512DQ-NEXT: vpermt2q %zmm18, %zmm11, %zmm55580; AVX512DQ-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill5581; AVX512DQ-NEXT: vmovdqa64 %zmm1, %zmm05582; AVX512DQ-NEXT: vpermt2q %zmm21, %zmm11, %zmm05583; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill5584; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} ymm28 = [1,6,11,0]5585; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm15 = [5,10,15,0,5,10,15,0]5586; AVX512DQ-NEXT: # zmm15 = mem[0,1,2,3,0,1,2,3]5587; AVX512DQ-NEXT: vmovdqa64 %zmm25, %zmm305588; AVX512DQ-NEXT: vpermt2q %zmm3, %zmm15, %zmm305589; AVX512DQ-NEXT: vmovdqa64 %zmm18, %zmm275590; AVX512DQ-NEXT: vpermt2q %zmm2, %zmm15, %zmm275591; AVX512DQ-NEXT: vmovdqa64 %zmm26, %zmm225592; AVX512DQ-NEXT: vpermt2q %zmm4, %zmm15, %zmm225593; AVX512DQ-NEXT: vmovdqa64 %zmm19, %zmm05594; AVX512DQ-NEXT: vpermt2q %zmm17, %zmm28, %zmm05595; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill5596; AVX512DQ-NEXT: vpermi2q %zmm1, %zmm21, %zmm155597; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm23 = [6,11,0,1,6,11,0,1]5598; AVX512DQ-NEXT: # zmm23 = mem[0,1,2,3,0,1,2,3]5599; AVX512DQ-NEXT: vmovdqa64 %zmm25, %zmm05600; AVX512DQ-NEXT: vpermt2q %zmm3, %zmm23, %zmm05601; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill5602; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} ymm13 = [2,7,12,0]5603; AVX512DQ-NEXT: vmovdqa64 %zmm18, %zmm05604; AVX512DQ-NEXT: vpermt2q %zmm2, %zmm23, %zmm05605; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill5606; AVX512DQ-NEXT: vmovdqa64 %zmm26, %zmm05607; AVX512DQ-NEXT: vpermt2q %zmm4, %zmm23, %zmm05608; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill5609; AVX512DQ-NEXT: vpermi2q %zmm1, %zmm21, %zmm235610; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm29 = [7,12,0,2,7,12,0,2]5611; AVX512DQ-NEXT: # zmm29 = mem[0,1,2,3,0,1,2,3]5612; AVX512DQ-NEXT: vmovdqa64 %zmm26, %zmm05613; AVX512DQ-NEXT: vpermt2q %zmm4, %zmm29, %zmm05614; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill5615; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm0 = [0,5,0,11,0,5,0,11]5616; AVX512DQ-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3]5617; AVX512DQ-NEXT: vpermt2q %zmm26, %zmm0, %zmm45618; AVX512DQ-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill5619; AVX512DQ-NEXT: vmovdqa64 %zmm25, %zmm265620; AVX512DQ-NEXT: vpermt2q %zmm3, %zmm29, %zmm265621; AVX512DQ-NEXT: vpermt2q %zmm25, %zmm0, %zmm35622; AVX512DQ-NEXT: vmovdqu64 %zmm3, (%rsp) # 64-byte Spill5623; AVX512DQ-NEXT: vmovdqa64 %zmm18, %zmm205624; AVX512DQ-NEXT: vpermt2q %zmm2, %zmm29, %zmm205625; AVX512DQ-NEXT: vpermi2q %zmm1, %zmm21, %zmm295626; AVX512DQ-NEXT: vpermt2q %zmm21, %zmm0, %zmm15627; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill5628; AVX512DQ-NEXT: vmovdqa64 %zmm19, %zmm215629; AVX512DQ-NEXT: vpermt2q %zmm17, %zmm13, %zmm215630; AVX512DQ-NEXT: vpermt2q %zmm18, %zmm0, %zmm25631; AVX512DQ-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill5632; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} ymm31 = [11,0,5,0]5633; AVX512DQ-NEXT: vmovdqa64 %zmm17, %zmm255634; AVX512DQ-NEXT: vpermt2q %zmm19, %zmm31, %zmm255635; AVX512DQ-NEXT: vpermt2q %zmm19, %zmm11, %zmm175636; AVX512DQ-NEXT: vmovdqa64 64(%rdi), %zmm185637; AVX512DQ-NEXT: vmovdqa64 %zmm16, %zmm65638; AVX512DQ-NEXT: vpermt2q %zmm18, %zmm24, %zmm65639; AVX512DQ-NEXT: vmovdqa64 %zmm16, %zmm95640; AVX512DQ-NEXT: vpermt2q %zmm18, %zmm28, %zmm95641; AVX512DQ-NEXT: vmovdqa64 %zmm16, %zmm85642; AVX512DQ-NEXT: vpermt2q %zmm18, %zmm13, %zmm85643; AVX512DQ-NEXT: vmovdqa64 %zmm18, %zmm195644; AVX512DQ-NEXT: vpermt2q %zmm16, %zmm31, %zmm195645; AVX512DQ-NEXT: vpermt2q %zmm16, %zmm11, %zmm185646; AVX512DQ-NEXT: vmovdqa64 704(%rdi), %zmm165647; AVX512DQ-NEXT: vmovdqa64 640(%rdi), %zmm15648; AVX512DQ-NEXT: vmovdqa64 %zmm1, %zmm35649; AVX512DQ-NEXT: vpermt2q %zmm16, %zmm24, %zmm35650; AVX512DQ-NEXT: vmovdqa64 1024(%rdi), %zmm25651; AVX512DQ-NEXT: vmovdqa64 960(%rdi), %zmm05652; AVX512DQ-NEXT: vpermi2q %zmm2, %zmm0, %zmm245653; AVX512DQ-NEXT: vmovdqa64 %zmm1, %zmm75654; AVX512DQ-NEXT: vpermt2q %zmm16, %zmm28, %zmm75655; AVX512DQ-NEXT: vpermi2q %zmm2, %zmm0, %zmm285656; AVX512DQ-NEXT: vmovdqa64 %zmm1, %zmm45657; AVX512DQ-NEXT: vpermt2q %zmm16, %zmm13, %zmm45658; AVX512DQ-NEXT: vpermi2q %zmm2, %zmm0, %zmm135659; AVX512DQ-NEXT: vmovdqa64 %zmm16, %zmm55660; AVX512DQ-NEXT: vpermt2q %zmm1, %zmm31, %zmm55661; AVX512DQ-NEXT: vpermi2q %zmm0, %zmm2, %zmm315662; AVX512DQ-NEXT: vpermt2q %zmm0, %zmm11, %zmm25663; AVX512DQ-NEXT: vpermt2q %zmm1, %zmm11, %zmm165664; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm14 = zmm6[0,1,2,3],zmm10[4,5,6,7]5665; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload5666; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm12 = zmm0[0,1,2,3],zmm12[4,5,6,7]5667; AVX512DQ-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3, %zmm10 # 64-byte Folded Reload5668; AVX512DQ-NEXT: # zmm10 = zmm3[0,1,2,3],mem[4,5,6,7]5669; AVX512DQ-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm24, %zmm6 # 64-byte Folded Reload5670; AVX512DQ-NEXT: # zmm6 = zmm24[0,1,2,3],mem[4,5,6,7]5671; AVX512DQ-NEXT: vmovdqa64 256(%rdi), %zmm115672; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} zmm24 = [0,1,2,3,4,5,6,11]5673; AVX512DQ-NEXT: vpermt2q %zmm11, %zmm24, %zmm145674; AVX512DQ-NEXT: vmovdqa64 576(%rdi), %zmm05675; AVX512DQ-NEXT: vpermt2q %zmm0, %zmm24, %zmm125676; AVX512DQ-NEXT: vmovdqa64 896(%rdi), %zmm35677; AVX512DQ-NEXT: vpermt2q %zmm3, %zmm24, %zmm105678; AVX512DQ-NEXT: vmovdqa64 1216(%rdi), %zmm15679; AVX512DQ-NEXT: vpermt2q %zmm1, %zmm24, %zmm65680; AVX512DQ-NEXT: movb $7, %al5681; AVX512DQ-NEXT: kmovw %eax, %k15682; AVX512DQ-NEXT: vmovdqa64 %zmm9, %zmm30 {%k1}5683; AVX512DQ-NEXT: vmovdqa64 %zmm7, %zmm27 {%k1}5684; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload5685; AVX512DQ-NEXT: vmovdqa64 %zmm7, %zmm22 {%k1}5686; AVX512DQ-NEXT: vmovdqa64 %zmm28, %zmm15 {%k1}5687; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} zmm7 = [0,1,2,3,4,5,6,12]5688; AVX512DQ-NEXT: vpermt2q %zmm11, %zmm7, %zmm305689; AVX512DQ-NEXT: vpermt2q %zmm3, %zmm7, %zmm275690; AVX512DQ-NEXT: vpermt2q %zmm0, %zmm7, %zmm225691; AVX512DQ-NEXT: vpermt2q %zmm1, %zmm7, %zmm155692; AVX512DQ-NEXT: movb $56, %al5693; AVX512DQ-NEXT: kmovw %eax, %k15694; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload5695; AVX512DQ-NEXT: vmovdqa64 %zmm7, %zmm8 {%k1}5696; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload5697; AVX512DQ-NEXT: vmovdqa64 %zmm7, %zmm4 {%k1}5698; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload5699; AVX512DQ-NEXT: vmovdqa64 %zmm7, %zmm21 {%k1}5700; AVX512DQ-NEXT: vmovdqa64 %zmm23, %zmm13 {%k1}5701; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} zmm7 = [0,1,2,3,4,5,8,13]5702; AVX512DQ-NEXT: vpermt2q %zmm11, %zmm7, %zmm85703; AVX512DQ-NEXT: vpermt2q %zmm3, %zmm7, %zmm45704; AVX512DQ-NEXT: vpermt2q %zmm0, %zmm7, %zmm215705; AVX512DQ-NEXT: vpermt2q %zmm1, %zmm7, %zmm135706; AVX512DQ-NEXT: vmovdqa64 %zmm26, %zmm19 {%k1}5707; AVX512DQ-NEXT: vmovdqa64 %zmm20, %zmm5 {%k1}5708; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload5709; AVX512DQ-NEXT: vmovdqa64 %zmm7, %zmm25 {%k1}5710; AVX512DQ-NEXT: vmovdqa64 %zmm29, %zmm31 {%k1}5711; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} zmm7 = [0,1,2,3,4,5,9,14]5712; AVX512DQ-NEXT: vpermt2q %zmm11, %zmm7, %zmm195713; AVX512DQ-NEXT: vpermt2q %zmm3, %zmm7, %zmm55714; AVX512DQ-NEXT: vpermt2q %zmm0, %zmm7, %zmm255715; AVX512DQ-NEXT: vpermt2q %zmm1, %zmm7, %zmm315716; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload5717; AVX512DQ-NEXT: vmovdqa64 %zmm7, %zmm17 {%k1}5718; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} zmm7 = [0,1,2,3,4,5,10,15]5719; AVX512DQ-NEXT: vpermt2q %zmm0, %zmm7, %zmm175720; AVX512DQ-NEXT: vmovdqu64 (%rsp), %zmm0 # 64-byte Reload5721; AVX512DQ-NEXT: vmovdqa64 %zmm0, %zmm18 {%k1}5722; AVX512DQ-NEXT: vpermt2q %zmm11, %zmm7, %zmm185723; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload5724; AVX512DQ-NEXT: vmovdqa64 %zmm0, %zmm2 {%k1}5725; AVX512DQ-NEXT: vpermt2q %zmm1, %zmm7, %zmm25726; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload5727; AVX512DQ-NEXT: vmovdqa64 %zmm0, %zmm16 {%k1}5728; AVX512DQ-NEXT: vpermt2q %zmm3, %zmm7, %zmm165729; AVX512DQ-NEXT: vmovdqa64 %zmm6, 192(%rsi)5730; AVX512DQ-NEXT: vmovdqa64 %zmm10, 128(%rsi)5731; AVX512DQ-NEXT: vmovdqa64 %zmm12, 64(%rsi)5732; AVX512DQ-NEXT: vmovdqa64 %zmm14, (%rsi)5733; AVX512DQ-NEXT: vmovdqa64 %zmm15, 192(%rdx)5734; AVX512DQ-NEXT: vmovdqa64 %zmm30, (%rdx)5735; AVX512DQ-NEXT: vmovdqa64 %zmm22, 64(%rdx)5736; AVX512DQ-NEXT: vmovdqa64 %zmm27, 128(%rdx)5737; AVX512DQ-NEXT: vmovdqa64 %zmm13, 192(%rcx)5738; AVX512DQ-NEXT: vmovdqa64 %zmm8, (%rcx)5739; AVX512DQ-NEXT: vmovdqa64 %zmm21, 64(%rcx)5740; AVX512DQ-NEXT: vmovdqa64 %zmm4, 128(%rcx)5741; AVX512DQ-NEXT: vmovdqa64 %zmm31, 192(%r8)5742; AVX512DQ-NEXT: vmovdqa64 %zmm19, (%r8)5743; AVX512DQ-NEXT: vmovdqa64 %zmm25, 64(%r8)5744; AVX512DQ-NEXT: vmovdqa64 %zmm5, 128(%r8)5745; AVX512DQ-NEXT: vmovdqa64 %zmm16, 128(%r9)5746; AVX512DQ-NEXT: vmovdqa64 %zmm2, 192(%r9)5747; AVX512DQ-NEXT: vmovdqa64 %zmm18, (%r9)5748; AVX512DQ-NEXT: vmovdqa64 %zmm17, 64(%r9)5749; AVX512DQ-NEXT: addq $648, %rsp # imm = 0x2885750; AVX512DQ-NEXT: vzeroupper5751; AVX512DQ-NEXT: retq5752;5753; AVX512DQ-FCP-LABEL: load_i64_stride5_vf32:5754; AVX512DQ-FCP: # %bb.0:5755; AVX512DQ-FCP-NEXT: subq $648, %rsp # imm = 0x2885756; AVX512DQ-FCP-NEXT: vmovdqa64 1088(%rdi), %zmm215757; AVX512DQ-FCP-NEXT: vmovdqa64 1152(%rdi), %zmm15758; AVX512DQ-FCP-NEXT: vmovdqa64 768(%rdi), %zmm185759; AVX512DQ-FCP-NEXT: vmovdqa64 832(%rdi), %zmm05760; AVX512DQ-FCP-NEXT: vmovdqa64 384(%rdi), %zmm175761; AVX512DQ-FCP-NEXT: vmovdqa64 320(%rdi), %zmm195762; AVX512DQ-FCP-NEXT: vmovdqa64 448(%rdi), %zmm265763; AVX512DQ-FCP-NEXT: vmovdqa64 512(%rdi), %zmm25764; AVX512DQ-FCP-NEXT: vmovdqa64 (%rdi), %zmm165765; AVX512DQ-FCP-NEXT: vmovdqa64 128(%rdi), %zmm255766; AVX512DQ-FCP-NEXT: vmovdqa64 192(%rdi), %zmm35767; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm11 = [12,1,6,0,12,1,6,0]5768; AVX512DQ-FCP-NEXT: # zmm11 = mem[0,1,2,3,0,1,2,3]5769; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, %zmm105770; AVX512DQ-FCP-NEXT: vpermt2q %zmm25, %zmm11, %zmm105771; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} ymm24 = [0,5,10,15]5772; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm125773; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm45774; AVX512DQ-FCP-NEXT: vpermt2q %zmm26, %zmm11, %zmm125775; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm19, %zmm25776; AVX512DQ-FCP-NEXT: vpermt2q %zmm17, %zmm24, %zmm25777; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill5778; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm55779; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm25780; AVX512DQ-FCP-NEXT: vpermt2q %zmm18, %zmm11, %zmm55781; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill5782; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, %zmm05783; AVX512DQ-FCP-NEXT: vpermt2q %zmm21, %zmm11, %zmm05784; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill5785; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} ymm28 = [1,6,11,0]5786; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm15 = [5,10,15,0,5,10,15,0]5787; AVX512DQ-FCP-NEXT: # zmm15 = mem[0,1,2,3,0,1,2,3]5788; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm25, %zmm305789; AVX512DQ-FCP-NEXT: vpermt2q %zmm3, %zmm15, %zmm305790; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm18, %zmm275791; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm15, %zmm275792; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm26, %zmm225793; AVX512DQ-FCP-NEXT: vpermt2q %zmm4, %zmm15, %zmm225794; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm19, %zmm05795; AVX512DQ-FCP-NEXT: vpermt2q %zmm17, %zmm28, %zmm05796; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill5797; AVX512DQ-FCP-NEXT: vpermi2q %zmm1, %zmm21, %zmm155798; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm23 = [6,11,0,1,6,11,0,1]5799; AVX512DQ-FCP-NEXT: # zmm23 = mem[0,1,2,3,0,1,2,3]5800; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm25, %zmm05801; AVX512DQ-FCP-NEXT: vpermt2q %zmm3, %zmm23, %zmm05802; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill5803; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} ymm13 = [2,7,12,0]5804; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm18, %zmm05805; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm23, %zmm05806; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill5807; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm26, %zmm05808; AVX512DQ-FCP-NEXT: vpermt2q %zmm4, %zmm23, %zmm05809; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill5810; AVX512DQ-FCP-NEXT: vpermi2q %zmm1, %zmm21, %zmm235811; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm29 = [7,12,0,2,7,12,0,2]5812; AVX512DQ-FCP-NEXT: # zmm29 = mem[0,1,2,3,0,1,2,3]5813; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm26, %zmm05814; AVX512DQ-FCP-NEXT: vpermt2q %zmm4, %zmm29, %zmm05815; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill5816; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm0 = [0,5,0,11,0,5,0,11]5817; AVX512DQ-FCP-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3]5818; AVX512DQ-FCP-NEXT: vpermt2q %zmm26, %zmm0, %zmm45819; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill5820; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm25, %zmm265821; AVX512DQ-FCP-NEXT: vpermt2q %zmm3, %zmm29, %zmm265822; AVX512DQ-FCP-NEXT: vpermt2q %zmm25, %zmm0, %zmm35823; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm3, (%rsp) # 64-byte Spill5824; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm18, %zmm205825; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm29, %zmm205826; AVX512DQ-FCP-NEXT: vpermi2q %zmm1, %zmm21, %zmm295827; AVX512DQ-FCP-NEXT: vpermt2q %zmm21, %zmm0, %zmm15828; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill5829; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm19, %zmm215830; AVX512DQ-FCP-NEXT: vpermt2q %zmm17, %zmm13, %zmm215831; AVX512DQ-FCP-NEXT: vpermt2q %zmm18, %zmm0, %zmm25832; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill5833; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} ymm31 = [11,0,5,0]5834; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm17, %zmm255835; AVX512DQ-FCP-NEXT: vpermt2q %zmm19, %zmm31, %zmm255836; AVX512DQ-FCP-NEXT: vpermt2q %zmm19, %zmm11, %zmm175837; AVX512DQ-FCP-NEXT: vmovdqa64 64(%rdi), %zmm185838; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm16, %zmm65839; AVX512DQ-FCP-NEXT: vpermt2q %zmm18, %zmm24, %zmm65840; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm16, %zmm95841; AVX512DQ-FCP-NEXT: vpermt2q %zmm18, %zmm28, %zmm95842; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm16, %zmm85843; AVX512DQ-FCP-NEXT: vpermt2q %zmm18, %zmm13, %zmm85844; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm18, %zmm195845; AVX512DQ-FCP-NEXT: vpermt2q %zmm16, %zmm31, %zmm195846; AVX512DQ-FCP-NEXT: vpermt2q %zmm16, %zmm11, %zmm185847; AVX512DQ-FCP-NEXT: vmovdqa64 704(%rdi), %zmm165848; AVX512DQ-FCP-NEXT: vmovdqa64 640(%rdi), %zmm15849; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, %zmm35850; AVX512DQ-FCP-NEXT: vpermt2q %zmm16, %zmm24, %zmm35851; AVX512DQ-FCP-NEXT: vmovdqa64 1024(%rdi), %zmm25852; AVX512DQ-FCP-NEXT: vmovdqa64 960(%rdi), %zmm05853; AVX512DQ-FCP-NEXT: vpermi2q %zmm2, %zmm0, %zmm245854; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, %zmm75855; AVX512DQ-FCP-NEXT: vpermt2q %zmm16, %zmm28, %zmm75856; AVX512DQ-FCP-NEXT: vpermi2q %zmm2, %zmm0, %zmm285857; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, %zmm45858; AVX512DQ-FCP-NEXT: vpermt2q %zmm16, %zmm13, %zmm45859; AVX512DQ-FCP-NEXT: vpermi2q %zmm2, %zmm0, %zmm135860; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm16, %zmm55861; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm31, %zmm55862; AVX512DQ-FCP-NEXT: vpermi2q %zmm0, %zmm2, %zmm315863; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm11, %zmm25864; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm11, %zmm165865; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm14 = zmm6[0,1,2,3],zmm10[4,5,6,7]5866; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload5867; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm12 = zmm0[0,1,2,3],zmm12[4,5,6,7]5868; AVX512DQ-FCP-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3, %zmm10 # 64-byte Folded Reload5869; AVX512DQ-FCP-NEXT: # zmm10 = zmm3[0,1,2,3],mem[4,5,6,7]5870; AVX512DQ-FCP-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm24, %zmm6 # 64-byte Folded Reload5871; AVX512DQ-FCP-NEXT: # zmm6 = zmm24[0,1,2,3],mem[4,5,6,7]5872; AVX512DQ-FCP-NEXT: vmovdqa64 256(%rdi), %zmm115873; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm24 = [0,1,2,3,4,5,6,11]5874; AVX512DQ-FCP-NEXT: vpermt2q %zmm11, %zmm24, %zmm145875; AVX512DQ-FCP-NEXT: vmovdqa64 576(%rdi), %zmm05876; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm24, %zmm125877; AVX512DQ-FCP-NEXT: vmovdqa64 896(%rdi), %zmm35878; AVX512DQ-FCP-NEXT: vpermt2q %zmm3, %zmm24, %zmm105879; AVX512DQ-FCP-NEXT: vmovdqa64 1216(%rdi), %zmm15880; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm24, %zmm65881; AVX512DQ-FCP-NEXT: movb $7, %al5882; AVX512DQ-FCP-NEXT: kmovw %eax, %k15883; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm9, %zmm30 {%k1}5884; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm7, %zmm27 {%k1}5885; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload5886; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm7, %zmm22 {%k1}5887; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm28, %zmm15 {%k1}5888; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm7 = [0,1,2,3,4,5,6,12]5889; AVX512DQ-FCP-NEXT: vpermt2q %zmm11, %zmm7, %zmm305890; AVX512DQ-FCP-NEXT: vpermt2q %zmm3, %zmm7, %zmm275891; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm7, %zmm225892; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm7, %zmm155893; AVX512DQ-FCP-NEXT: movb $56, %al5894; AVX512DQ-FCP-NEXT: kmovw %eax, %k15895; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload5896; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm7, %zmm8 {%k1}5897; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload5898; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm7, %zmm4 {%k1}5899; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload5900; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm7, %zmm21 {%k1}5901; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm23, %zmm13 {%k1}5902; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm7 = [0,1,2,3,4,5,8,13]5903; AVX512DQ-FCP-NEXT: vpermt2q %zmm11, %zmm7, %zmm85904; AVX512DQ-FCP-NEXT: vpermt2q %zmm3, %zmm7, %zmm45905; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm7, %zmm215906; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm7, %zmm135907; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm26, %zmm19 {%k1}5908; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm20, %zmm5 {%k1}5909; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload5910; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm7, %zmm25 {%k1}5911; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm29, %zmm31 {%k1}5912; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm7 = [0,1,2,3,4,5,9,14]5913; AVX512DQ-FCP-NEXT: vpermt2q %zmm11, %zmm7, %zmm195914; AVX512DQ-FCP-NEXT: vpermt2q %zmm3, %zmm7, %zmm55915; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm7, %zmm255916; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm7, %zmm315917; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload5918; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm7, %zmm17 {%k1}5919; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm7 = [0,1,2,3,4,5,10,15]5920; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm7, %zmm175921; AVX512DQ-FCP-NEXT: vmovdqu64 (%rsp), %zmm0 # 64-byte Reload5922; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm18 {%k1}5923; AVX512DQ-FCP-NEXT: vpermt2q %zmm11, %zmm7, %zmm185924; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload5925; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm2 {%k1}5926; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm7, %zmm25927; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload5928; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm16 {%k1}5929; AVX512DQ-FCP-NEXT: vpermt2q %zmm3, %zmm7, %zmm165930; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, 192(%rsi)5931; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm10, 128(%rsi)5932; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm12, 64(%rsi)5933; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm14, (%rsi)5934; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm15, 192(%rdx)5935; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm30, (%rdx)5936; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm22, 64(%rdx)5937; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm27, 128(%rdx)5938; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm13, 192(%rcx)5939; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm8, (%rcx)5940; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm21, 64(%rcx)5941; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm4, 128(%rcx)5942; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm31, 192(%r8)5943; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm19, (%r8)5944; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm25, 64(%r8)5945; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm5, 128(%r8)5946; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm16, 128(%r9)5947; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, 192(%r9)5948; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm18, (%r9)5949; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm17, 64(%r9)5950; AVX512DQ-FCP-NEXT: addq $648, %rsp # imm = 0x2885951; AVX512DQ-FCP-NEXT: vzeroupper5952; AVX512DQ-FCP-NEXT: retq5953;5954; AVX512BW-LABEL: load_i64_stride5_vf32:5955; AVX512BW: # %bb.0:5956; AVX512BW-NEXT: subq $648, %rsp # imm = 0x2885957; AVX512BW-NEXT: vmovdqa64 1088(%rdi), %zmm215958; AVX512BW-NEXT: vmovdqa64 1152(%rdi), %zmm15959; AVX512BW-NEXT: vmovdqa64 768(%rdi), %zmm185960; AVX512BW-NEXT: vmovdqa64 832(%rdi), %zmm05961; AVX512BW-NEXT: vmovdqa64 384(%rdi), %zmm175962; AVX512BW-NEXT: vmovdqa64 320(%rdi), %zmm195963; AVX512BW-NEXT: vmovdqa64 448(%rdi), %zmm265964; AVX512BW-NEXT: vmovdqa64 512(%rdi), %zmm25965; AVX512BW-NEXT: vmovdqa64 (%rdi), %zmm165966; AVX512BW-NEXT: vmovdqa64 128(%rdi), %zmm255967; AVX512BW-NEXT: vmovdqa64 192(%rdi), %zmm35968; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm11 = [12,1,6,0,12,1,6,0]5969; AVX512BW-NEXT: # zmm11 = mem[0,1,2,3,0,1,2,3]5970; AVX512BW-NEXT: vmovdqa64 %zmm3, %zmm105971; AVX512BW-NEXT: vpermt2q %zmm25, %zmm11, %zmm105972; AVX512BW-NEXT: vpmovsxbq {{.*#+}} ymm24 = [0,5,10,15]5973; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm125974; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm45975; AVX512BW-NEXT: vpermt2q %zmm26, %zmm11, %zmm125976; AVX512BW-NEXT: vmovdqa64 %zmm19, %zmm25977; AVX512BW-NEXT: vpermt2q %zmm17, %zmm24, %zmm25978; AVX512BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill5979; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm55980; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm25981; AVX512BW-NEXT: vpermt2q %zmm18, %zmm11, %zmm55982; AVX512BW-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill5983; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm05984; AVX512BW-NEXT: vpermt2q %zmm21, %zmm11, %zmm05985; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill5986; AVX512BW-NEXT: vpmovsxbq {{.*#+}} ymm28 = [1,6,11,0]5987; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm15 = [5,10,15,0,5,10,15,0]5988; AVX512BW-NEXT: # zmm15 = mem[0,1,2,3,0,1,2,3]5989; AVX512BW-NEXT: vmovdqa64 %zmm25, %zmm305990; AVX512BW-NEXT: vpermt2q %zmm3, %zmm15, %zmm305991; AVX512BW-NEXT: vmovdqa64 %zmm18, %zmm275992; AVX512BW-NEXT: vpermt2q %zmm2, %zmm15, %zmm275993; AVX512BW-NEXT: vmovdqa64 %zmm26, %zmm225994; AVX512BW-NEXT: vpermt2q %zmm4, %zmm15, %zmm225995; AVX512BW-NEXT: vmovdqa64 %zmm19, %zmm05996; AVX512BW-NEXT: vpermt2q %zmm17, %zmm28, %zmm05997; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill5998; AVX512BW-NEXT: vpermi2q %zmm1, %zmm21, %zmm155999; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm23 = [6,11,0,1,6,11,0,1]6000; AVX512BW-NEXT: # zmm23 = mem[0,1,2,3,0,1,2,3]6001; AVX512BW-NEXT: vmovdqa64 %zmm25, %zmm06002; AVX512BW-NEXT: vpermt2q %zmm3, %zmm23, %zmm06003; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6004; AVX512BW-NEXT: vpmovsxbq {{.*#+}} ymm13 = [2,7,12,0]6005; AVX512BW-NEXT: vmovdqa64 %zmm18, %zmm06006; AVX512BW-NEXT: vpermt2q %zmm2, %zmm23, %zmm06007; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6008; AVX512BW-NEXT: vmovdqa64 %zmm26, %zmm06009; AVX512BW-NEXT: vpermt2q %zmm4, %zmm23, %zmm06010; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6011; AVX512BW-NEXT: vpermi2q %zmm1, %zmm21, %zmm236012; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm29 = [7,12,0,2,7,12,0,2]6013; AVX512BW-NEXT: # zmm29 = mem[0,1,2,3,0,1,2,3]6014; AVX512BW-NEXT: vmovdqa64 %zmm26, %zmm06015; AVX512BW-NEXT: vpermt2q %zmm4, %zmm29, %zmm06016; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6017; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm0 = [0,5,0,11,0,5,0,11]6018; AVX512BW-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3]6019; AVX512BW-NEXT: vpermt2q %zmm26, %zmm0, %zmm46020; AVX512BW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6021; AVX512BW-NEXT: vmovdqa64 %zmm25, %zmm266022; AVX512BW-NEXT: vpermt2q %zmm3, %zmm29, %zmm266023; AVX512BW-NEXT: vpermt2q %zmm25, %zmm0, %zmm36024; AVX512BW-NEXT: vmovdqu64 %zmm3, (%rsp) # 64-byte Spill6025; AVX512BW-NEXT: vmovdqa64 %zmm18, %zmm206026; AVX512BW-NEXT: vpermt2q %zmm2, %zmm29, %zmm206027; AVX512BW-NEXT: vpermi2q %zmm1, %zmm21, %zmm296028; AVX512BW-NEXT: vpermt2q %zmm21, %zmm0, %zmm16029; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6030; AVX512BW-NEXT: vmovdqa64 %zmm19, %zmm216031; AVX512BW-NEXT: vpermt2q %zmm17, %zmm13, %zmm216032; AVX512BW-NEXT: vpermt2q %zmm18, %zmm0, %zmm26033; AVX512BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6034; AVX512BW-NEXT: vpmovsxbq {{.*#+}} ymm31 = [11,0,5,0]6035; AVX512BW-NEXT: vmovdqa64 %zmm17, %zmm256036; AVX512BW-NEXT: vpermt2q %zmm19, %zmm31, %zmm256037; AVX512BW-NEXT: vpermt2q %zmm19, %zmm11, %zmm176038; AVX512BW-NEXT: vmovdqa64 64(%rdi), %zmm186039; AVX512BW-NEXT: vmovdqa64 %zmm16, %zmm66040; AVX512BW-NEXT: vpermt2q %zmm18, %zmm24, %zmm66041; AVX512BW-NEXT: vmovdqa64 %zmm16, %zmm96042; AVX512BW-NEXT: vpermt2q %zmm18, %zmm28, %zmm96043; AVX512BW-NEXT: vmovdqa64 %zmm16, %zmm86044; AVX512BW-NEXT: vpermt2q %zmm18, %zmm13, %zmm86045; AVX512BW-NEXT: vmovdqa64 %zmm18, %zmm196046; AVX512BW-NEXT: vpermt2q %zmm16, %zmm31, %zmm196047; AVX512BW-NEXT: vpermt2q %zmm16, %zmm11, %zmm186048; AVX512BW-NEXT: vmovdqa64 704(%rdi), %zmm166049; AVX512BW-NEXT: vmovdqa64 640(%rdi), %zmm16050; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm36051; AVX512BW-NEXT: vpermt2q %zmm16, %zmm24, %zmm36052; AVX512BW-NEXT: vmovdqa64 1024(%rdi), %zmm26053; AVX512BW-NEXT: vmovdqa64 960(%rdi), %zmm06054; AVX512BW-NEXT: vpermi2q %zmm2, %zmm0, %zmm246055; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm76056; AVX512BW-NEXT: vpermt2q %zmm16, %zmm28, %zmm76057; AVX512BW-NEXT: vpermi2q %zmm2, %zmm0, %zmm286058; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm46059; AVX512BW-NEXT: vpermt2q %zmm16, %zmm13, %zmm46060; AVX512BW-NEXT: vpermi2q %zmm2, %zmm0, %zmm136061; AVX512BW-NEXT: vmovdqa64 %zmm16, %zmm56062; AVX512BW-NEXT: vpermt2q %zmm1, %zmm31, %zmm56063; AVX512BW-NEXT: vpermi2q %zmm0, %zmm2, %zmm316064; AVX512BW-NEXT: vpermt2q %zmm0, %zmm11, %zmm26065; AVX512BW-NEXT: vpermt2q %zmm1, %zmm11, %zmm166066; AVX512BW-NEXT: vshufi64x2 {{.*#+}} zmm14 = zmm6[0,1,2,3],zmm10[4,5,6,7]6067; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload6068; AVX512BW-NEXT: vshufi64x2 {{.*#+}} zmm12 = zmm0[0,1,2,3],zmm12[4,5,6,7]6069; AVX512BW-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3, %zmm10 # 64-byte Folded Reload6070; AVX512BW-NEXT: # zmm10 = zmm3[0,1,2,3],mem[4,5,6,7]6071; AVX512BW-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm24, %zmm6 # 64-byte Folded Reload6072; AVX512BW-NEXT: # zmm6 = zmm24[0,1,2,3],mem[4,5,6,7]6073; AVX512BW-NEXT: vmovdqa64 256(%rdi), %zmm116074; AVX512BW-NEXT: vpmovsxbq {{.*#+}} zmm24 = [0,1,2,3,4,5,6,11]6075; AVX512BW-NEXT: vpermt2q %zmm11, %zmm24, %zmm146076; AVX512BW-NEXT: vmovdqa64 576(%rdi), %zmm06077; AVX512BW-NEXT: vpermt2q %zmm0, %zmm24, %zmm126078; AVX512BW-NEXT: vmovdqa64 896(%rdi), %zmm36079; AVX512BW-NEXT: vpermt2q %zmm3, %zmm24, %zmm106080; AVX512BW-NEXT: vmovdqa64 1216(%rdi), %zmm16081; AVX512BW-NEXT: vpermt2q %zmm1, %zmm24, %zmm66082; AVX512BW-NEXT: movb $7, %al6083; AVX512BW-NEXT: kmovd %eax, %k16084; AVX512BW-NEXT: vmovdqa64 %zmm9, %zmm30 {%k1}6085; AVX512BW-NEXT: vmovdqa64 %zmm7, %zmm27 {%k1}6086; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload6087; AVX512BW-NEXT: vmovdqa64 %zmm7, %zmm22 {%k1}6088; AVX512BW-NEXT: vmovdqa64 %zmm28, %zmm15 {%k1}6089; AVX512BW-NEXT: vpmovsxbq {{.*#+}} zmm7 = [0,1,2,3,4,5,6,12]6090; AVX512BW-NEXT: vpermt2q %zmm11, %zmm7, %zmm306091; AVX512BW-NEXT: vpermt2q %zmm3, %zmm7, %zmm276092; AVX512BW-NEXT: vpermt2q %zmm0, %zmm7, %zmm226093; AVX512BW-NEXT: vpermt2q %zmm1, %zmm7, %zmm156094; AVX512BW-NEXT: movb $56, %al6095; AVX512BW-NEXT: kmovd %eax, %k16096; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload6097; AVX512BW-NEXT: vmovdqa64 %zmm7, %zmm8 {%k1}6098; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload6099; AVX512BW-NEXT: vmovdqa64 %zmm7, %zmm4 {%k1}6100; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload6101; AVX512BW-NEXT: vmovdqa64 %zmm7, %zmm21 {%k1}6102; AVX512BW-NEXT: vmovdqa64 %zmm23, %zmm13 {%k1}6103; AVX512BW-NEXT: vpmovsxbq {{.*#+}} zmm7 = [0,1,2,3,4,5,8,13]6104; AVX512BW-NEXT: vpermt2q %zmm11, %zmm7, %zmm86105; AVX512BW-NEXT: vpermt2q %zmm3, %zmm7, %zmm46106; AVX512BW-NEXT: vpermt2q %zmm0, %zmm7, %zmm216107; AVX512BW-NEXT: vpermt2q %zmm1, %zmm7, %zmm136108; AVX512BW-NEXT: vmovdqa64 %zmm26, %zmm19 {%k1}6109; AVX512BW-NEXT: vmovdqa64 %zmm20, %zmm5 {%k1}6110; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload6111; AVX512BW-NEXT: vmovdqa64 %zmm7, %zmm25 {%k1}6112; AVX512BW-NEXT: vmovdqa64 %zmm29, %zmm31 {%k1}6113; AVX512BW-NEXT: vpmovsxbq {{.*#+}} zmm7 = [0,1,2,3,4,5,9,14]6114; AVX512BW-NEXT: vpermt2q %zmm11, %zmm7, %zmm196115; AVX512BW-NEXT: vpermt2q %zmm3, %zmm7, %zmm56116; AVX512BW-NEXT: vpermt2q %zmm0, %zmm7, %zmm256117; AVX512BW-NEXT: vpermt2q %zmm1, %zmm7, %zmm316118; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload6119; AVX512BW-NEXT: vmovdqa64 %zmm7, %zmm17 {%k1}6120; AVX512BW-NEXT: vpmovsxbq {{.*#+}} zmm7 = [0,1,2,3,4,5,10,15]6121; AVX512BW-NEXT: vpermt2q %zmm0, %zmm7, %zmm176122; AVX512BW-NEXT: vmovdqu64 (%rsp), %zmm0 # 64-byte Reload6123; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm18 {%k1}6124; AVX512BW-NEXT: vpermt2q %zmm11, %zmm7, %zmm186125; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload6126; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm2 {%k1}6127; AVX512BW-NEXT: vpermt2q %zmm1, %zmm7, %zmm26128; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload6129; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm16 {%k1}6130; AVX512BW-NEXT: vpermt2q %zmm3, %zmm7, %zmm166131; AVX512BW-NEXT: vmovdqa64 %zmm6, 192(%rsi)6132; AVX512BW-NEXT: vmovdqa64 %zmm10, 128(%rsi)6133; AVX512BW-NEXT: vmovdqa64 %zmm12, 64(%rsi)6134; AVX512BW-NEXT: vmovdqa64 %zmm14, (%rsi)6135; AVX512BW-NEXT: vmovdqa64 %zmm15, 192(%rdx)6136; AVX512BW-NEXT: vmovdqa64 %zmm30, (%rdx)6137; AVX512BW-NEXT: vmovdqa64 %zmm22, 64(%rdx)6138; AVX512BW-NEXT: vmovdqa64 %zmm27, 128(%rdx)6139; AVX512BW-NEXT: vmovdqa64 %zmm13, 192(%rcx)6140; AVX512BW-NEXT: vmovdqa64 %zmm8, (%rcx)6141; AVX512BW-NEXT: vmovdqa64 %zmm21, 64(%rcx)6142; AVX512BW-NEXT: vmovdqa64 %zmm4, 128(%rcx)6143; AVX512BW-NEXT: vmovdqa64 %zmm31, 192(%r8)6144; AVX512BW-NEXT: vmovdqa64 %zmm19, (%r8)6145; AVX512BW-NEXT: vmovdqa64 %zmm25, 64(%r8)6146; AVX512BW-NEXT: vmovdqa64 %zmm5, 128(%r8)6147; AVX512BW-NEXT: vmovdqa64 %zmm16, 128(%r9)6148; AVX512BW-NEXT: vmovdqa64 %zmm2, 192(%r9)6149; AVX512BW-NEXT: vmovdqa64 %zmm18, (%r9)6150; AVX512BW-NEXT: vmovdqa64 %zmm17, 64(%r9)6151; AVX512BW-NEXT: addq $648, %rsp # imm = 0x2886152; AVX512BW-NEXT: vzeroupper6153; AVX512BW-NEXT: retq6154;6155; AVX512BW-FCP-LABEL: load_i64_stride5_vf32:6156; AVX512BW-FCP: # %bb.0:6157; AVX512BW-FCP-NEXT: subq $648, %rsp # imm = 0x2886158; AVX512BW-FCP-NEXT: vmovdqa64 1088(%rdi), %zmm216159; AVX512BW-FCP-NEXT: vmovdqa64 1152(%rdi), %zmm16160; AVX512BW-FCP-NEXT: vmovdqa64 768(%rdi), %zmm186161; AVX512BW-FCP-NEXT: vmovdqa64 832(%rdi), %zmm06162; AVX512BW-FCP-NEXT: vmovdqa64 384(%rdi), %zmm176163; AVX512BW-FCP-NEXT: vmovdqa64 320(%rdi), %zmm196164; AVX512BW-FCP-NEXT: vmovdqa64 448(%rdi), %zmm266165; AVX512BW-FCP-NEXT: vmovdqa64 512(%rdi), %zmm26166; AVX512BW-FCP-NEXT: vmovdqa64 (%rdi), %zmm166167; AVX512BW-FCP-NEXT: vmovdqa64 128(%rdi), %zmm256168; AVX512BW-FCP-NEXT: vmovdqa64 192(%rdi), %zmm36169; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm11 = [12,1,6,0,12,1,6,0]6170; AVX512BW-FCP-NEXT: # zmm11 = mem[0,1,2,3,0,1,2,3]6171; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm106172; AVX512BW-FCP-NEXT: vpermt2q %zmm25, %zmm11, %zmm106173; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} ymm24 = [0,5,10,15]6174; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm126175; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm46176; AVX512BW-FCP-NEXT: vpermt2q %zmm26, %zmm11, %zmm126177; AVX512BW-FCP-NEXT: vmovdqa64 %zmm19, %zmm26178; AVX512BW-FCP-NEXT: vpermt2q %zmm17, %zmm24, %zmm26179; AVX512BW-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6180; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm56181; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm26182; AVX512BW-FCP-NEXT: vpermt2q %zmm18, %zmm11, %zmm56183; AVX512BW-FCP-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6184; AVX512BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm06185; AVX512BW-FCP-NEXT: vpermt2q %zmm21, %zmm11, %zmm06186; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6187; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} ymm28 = [1,6,11,0]6188; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm15 = [5,10,15,0,5,10,15,0]6189; AVX512BW-FCP-NEXT: # zmm15 = mem[0,1,2,3,0,1,2,3]6190; AVX512BW-FCP-NEXT: vmovdqa64 %zmm25, %zmm306191; AVX512BW-FCP-NEXT: vpermt2q %zmm3, %zmm15, %zmm306192; AVX512BW-FCP-NEXT: vmovdqa64 %zmm18, %zmm276193; AVX512BW-FCP-NEXT: vpermt2q %zmm2, %zmm15, %zmm276194; AVX512BW-FCP-NEXT: vmovdqa64 %zmm26, %zmm226195; AVX512BW-FCP-NEXT: vpermt2q %zmm4, %zmm15, %zmm226196; AVX512BW-FCP-NEXT: vmovdqa64 %zmm19, %zmm06197; AVX512BW-FCP-NEXT: vpermt2q %zmm17, %zmm28, %zmm06198; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6199; AVX512BW-FCP-NEXT: vpermi2q %zmm1, %zmm21, %zmm156200; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm23 = [6,11,0,1,6,11,0,1]6201; AVX512BW-FCP-NEXT: # zmm23 = mem[0,1,2,3,0,1,2,3]6202; AVX512BW-FCP-NEXT: vmovdqa64 %zmm25, %zmm06203; AVX512BW-FCP-NEXT: vpermt2q %zmm3, %zmm23, %zmm06204; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6205; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} ymm13 = [2,7,12,0]6206; AVX512BW-FCP-NEXT: vmovdqa64 %zmm18, %zmm06207; AVX512BW-FCP-NEXT: vpermt2q %zmm2, %zmm23, %zmm06208; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6209; AVX512BW-FCP-NEXT: vmovdqa64 %zmm26, %zmm06210; AVX512BW-FCP-NEXT: vpermt2q %zmm4, %zmm23, %zmm06211; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6212; AVX512BW-FCP-NEXT: vpermi2q %zmm1, %zmm21, %zmm236213; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm29 = [7,12,0,2,7,12,0,2]6214; AVX512BW-FCP-NEXT: # zmm29 = mem[0,1,2,3,0,1,2,3]6215; AVX512BW-FCP-NEXT: vmovdqa64 %zmm26, %zmm06216; AVX512BW-FCP-NEXT: vpermt2q %zmm4, %zmm29, %zmm06217; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6218; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm0 = [0,5,0,11,0,5,0,11]6219; AVX512BW-FCP-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3]6220; AVX512BW-FCP-NEXT: vpermt2q %zmm26, %zmm0, %zmm46221; AVX512BW-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6222; AVX512BW-FCP-NEXT: vmovdqa64 %zmm25, %zmm266223; AVX512BW-FCP-NEXT: vpermt2q %zmm3, %zmm29, %zmm266224; AVX512BW-FCP-NEXT: vpermt2q %zmm25, %zmm0, %zmm36225; AVX512BW-FCP-NEXT: vmovdqu64 %zmm3, (%rsp) # 64-byte Spill6226; AVX512BW-FCP-NEXT: vmovdqa64 %zmm18, %zmm206227; AVX512BW-FCP-NEXT: vpermt2q %zmm2, %zmm29, %zmm206228; AVX512BW-FCP-NEXT: vpermi2q %zmm1, %zmm21, %zmm296229; AVX512BW-FCP-NEXT: vpermt2q %zmm21, %zmm0, %zmm16230; AVX512BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6231; AVX512BW-FCP-NEXT: vmovdqa64 %zmm19, %zmm216232; AVX512BW-FCP-NEXT: vpermt2q %zmm17, %zmm13, %zmm216233; AVX512BW-FCP-NEXT: vpermt2q %zmm18, %zmm0, %zmm26234; AVX512BW-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6235; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} ymm31 = [11,0,5,0]6236; AVX512BW-FCP-NEXT: vmovdqa64 %zmm17, %zmm256237; AVX512BW-FCP-NEXT: vpermt2q %zmm19, %zmm31, %zmm256238; AVX512BW-FCP-NEXT: vpermt2q %zmm19, %zmm11, %zmm176239; AVX512BW-FCP-NEXT: vmovdqa64 64(%rdi), %zmm186240; AVX512BW-FCP-NEXT: vmovdqa64 %zmm16, %zmm66241; AVX512BW-FCP-NEXT: vpermt2q %zmm18, %zmm24, %zmm66242; AVX512BW-FCP-NEXT: vmovdqa64 %zmm16, %zmm96243; AVX512BW-FCP-NEXT: vpermt2q %zmm18, %zmm28, %zmm96244; AVX512BW-FCP-NEXT: vmovdqa64 %zmm16, %zmm86245; AVX512BW-FCP-NEXT: vpermt2q %zmm18, %zmm13, %zmm86246; AVX512BW-FCP-NEXT: vmovdqa64 %zmm18, %zmm196247; AVX512BW-FCP-NEXT: vpermt2q %zmm16, %zmm31, %zmm196248; AVX512BW-FCP-NEXT: vpermt2q %zmm16, %zmm11, %zmm186249; AVX512BW-FCP-NEXT: vmovdqa64 704(%rdi), %zmm166250; AVX512BW-FCP-NEXT: vmovdqa64 640(%rdi), %zmm16251; AVX512BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm36252; AVX512BW-FCP-NEXT: vpermt2q %zmm16, %zmm24, %zmm36253; AVX512BW-FCP-NEXT: vmovdqa64 1024(%rdi), %zmm26254; AVX512BW-FCP-NEXT: vmovdqa64 960(%rdi), %zmm06255; AVX512BW-FCP-NEXT: vpermi2q %zmm2, %zmm0, %zmm246256; AVX512BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm76257; AVX512BW-FCP-NEXT: vpermt2q %zmm16, %zmm28, %zmm76258; AVX512BW-FCP-NEXT: vpermi2q %zmm2, %zmm0, %zmm286259; AVX512BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm46260; AVX512BW-FCP-NEXT: vpermt2q %zmm16, %zmm13, %zmm46261; AVX512BW-FCP-NEXT: vpermi2q %zmm2, %zmm0, %zmm136262; AVX512BW-FCP-NEXT: vmovdqa64 %zmm16, %zmm56263; AVX512BW-FCP-NEXT: vpermt2q %zmm1, %zmm31, %zmm56264; AVX512BW-FCP-NEXT: vpermi2q %zmm0, %zmm2, %zmm316265; AVX512BW-FCP-NEXT: vpermt2q %zmm0, %zmm11, %zmm26266; AVX512BW-FCP-NEXT: vpermt2q %zmm1, %zmm11, %zmm166267; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm14 = zmm6[0,1,2,3],zmm10[4,5,6,7]6268; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload6269; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm12 = zmm0[0,1,2,3],zmm12[4,5,6,7]6270; AVX512BW-FCP-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3, %zmm10 # 64-byte Folded Reload6271; AVX512BW-FCP-NEXT: # zmm10 = zmm3[0,1,2,3],mem[4,5,6,7]6272; AVX512BW-FCP-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm24, %zmm6 # 64-byte Folded Reload6273; AVX512BW-FCP-NEXT: # zmm6 = zmm24[0,1,2,3],mem[4,5,6,7]6274; AVX512BW-FCP-NEXT: vmovdqa64 256(%rdi), %zmm116275; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm24 = [0,1,2,3,4,5,6,11]6276; AVX512BW-FCP-NEXT: vpermt2q %zmm11, %zmm24, %zmm146277; AVX512BW-FCP-NEXT: vmovdqa64 576(%rdi), %zmm06278; AVX512BW-FCP-NEXT: vpermt2q %zmm0, %zmm24, %zmm126279; AVX512BW-FCP-NEXT: vmovdqa64 896(%rdi), %zmm36280; AVX512BW-FCP-NEXT: vpermt2q %zmm3, %zmm24, %zmm106281; AVX512BW-FCP-NEXT: vmovdqa64 1216(%rdi), %zmm16282; AVX512BW-FCP-NEXT: vpermt2q %zmm1, %zmm24, %zmm66283; AVX512BW-FCP-NEXT: movb $7, %al6284; AVX512BW-FCP-NEXT: kmovd %eax, %k16285; AVX512BW-FCP-NEXT: vmovdqa64 %zmm9, %zmm30 {%k1}6286; AVX512BW-FCP-NEXT: vmovdqa64 %zmm7, %zmm27 {%k1}6287; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload6288; AVX512BW-FCP-NEXT: vmovdqa64 %zmm7, %zmm22 {%k1}6289; AVX512BW-FCP-NEXT: vmovdqa64 %zmm28, %zmm15 {%k1}6290; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm7 = [0,1,2,3,4,5,6,12]6291; AVX512BW-FCP-NEXT: vpermt2q %zmm11, %zmm7, %zmm306292; AVX512BW-FCP-NEXT: vpermt2q %zmm3, %zmm7, %zmm276293; AVX512BW-FCP-NEXT: vpermt2q %zmm0, %zmm7, %zmm226294; AVX512BW-FCP-NEXT: vpermt2q %zmm1, %zmm7, %zmm156295; AVX512BW-FCP-NEXT: movb $56, %al6296; AVX512BW-FCP-NEXT: kmovd %eax, %k16297; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload6298; AVX512BW-FCP-NEXT: vmovdqa64 %zmm7, %zmm8 {%k1}6299; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload6300; AVX512BW-FCP-NEXT: vmovdqa64 %zmm7, %zmm4 {%k1}6301; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload6302; AVX512BW-FCP-NEXT: vmovdqa64 %zmm7, %zmm21 {%k1}6303; AVX512BW-FCP-NEXT: vmovdqa64 %zmm23, %zmm13 {%k1}6304; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm7 = [0,1,2,3,4,5,8,13]6305; AVX512BW-FCP-NEXT: vpermt2q %zmm11, %zmm7, %zmm86306; AVX512BW-FCP-NEXT: vpermt2q %zmm3, %zmm7, %zmm46307; AVX512BW-FCP-NEXT: vpermt2q %zmm0, %zmm7, %zmm216308; AVX512BW-FCP-NEXT: vpermt2q %zmm1, %zmm7, %zmm136309; AVX512BW-FCP-NEXT: vmovdqa64 %zmm26, %zmm19 {%k1}6310; AVX512BW-FCP-NEXT: vmovdqa64 %zmm20, %zmm5 {%k1}6311; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload6312; AVX512BW-FCP-NEXT: vmovdqa64 %zmm7, %zmm25 {%k1}6313; AVX512BW-FCP-NEXT: vmovdqa64 %zmm29, %zmm31 {%k1}6314; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm7 = [0,1,2,3,4,5,9,14]6315; AVX512BW-FCP-NEXT: vpermt2q %zmm11, %zmm7, %zmm196316; AVX512BW-FCP-NEXT: vpermt2q %zmm3, %zmm7, %zmm56317; AVX512BW-FCP-NEXT: vpermt2q %zmm0, %zmm7, %zmm256318; AVX512BW-FCP-NEXT: vpermt2q %zmm1, %zmm7, %zmm316319; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload6320; AVX512BW-FCP-NEXT: vmovdqa64 %zmm7, %zmm17 {%k1}6321; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm7 = [0,1,2,3,4,5,10,15]6322; AVX512BW-FCP-NEXT: vpermt2q %zmm0, %zmm7, %zmm176323; AVX512BW-FCP-NEXT: vmovdqu64 (%rsp), %zmm0 # 64-byte Reload6324; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm18 {%k1}6325; AVX512BW-FCP-NEXT: vpermt2q %zmm11, %zmm7, %zmm186326; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload6327; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm2 {%k1}6328; AVX512BW-FCP-NEXT: vpermt2q %zmm1, %zmm7, %zmm26329; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload6330; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm16 {%k1}6331; AVX512BW-FCP-NEXT: vpermt2q %zmm3, %zmm7, %zmm166332; AVX512BW-FCP-NEXT: vmovdqa64 %zmm6, 192(%rsi)6333; AVX512BW-FCP-NEXT: vmovdqa64 %zmm10, 128(%rsi)6334; AVX512BW-FCP-NEXT: vmovdqa64 %zmm12, 64(%rsi)6335; AVX512BW-FCP-NEXT: vmovdqa64 %zmm14, (%rsi)6336; AVX512BW-FCP-NEXT: vmovdqa64 %zmm15, 192(%rdx)6337; AVX512BW-FCP-NEXT: vmovdqa64 %zmm30, (%rdx)6338; AVX512BW-FCP-NEXT: vmovdqa64 %zmm22, 64(%rdx)6339; AVX512BW-FCP-NEXT: vmovdqa64 %zmm27, 128(%rdx)6340; AVX512BW-FCP-NEXT: vmovdqa64 %zmm13, 192(%rcx)6341; AVX512BW-FCP-NEXT: vmovdqa64 %zmm8, (%rcx)6342; AVX512BW-FCP-NEXT: vmovdqa64 %zmm21, 64(%rcx)6343; AVX512BW-FCP-NEXT: vmovdqa64 %zmm4, 128(%rcx)6344; AVX512BW-FCP-NEXT: vmovdqa64 %zmm31, 192(%r8)6345; AVX512BW-FCP-NEXT: vmovdqa64 %zmm19, (%r8)6346; AVX512BW-FCP-NEXT: vmovdqa64 %zmm25, 64(%r8)6347; AVX512BW-FCP-NEXT: vmovdqa64 %zmm5, 128(%r8)6348; AVX512BW-FCP-NEXT: vmovdqa64 %zmm16, 128(%r9)6349; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, 192(%r9)6350; AVX512BW-FCP-NEXT: vmovdqa64 %zmm18, (%r9)6351; AVX512BW-FCP-NEXT: vmovdqa64 %zmm17, 64(%r9)6352; AVX512BW-FCP-NEXT: addq $648, %rsp # imm = 0x2886353; AVX512BW-FCP-NEXT: vzeroupper6354; AVX512BW-FCP-NEXT: retq6355;6356; AVX512DQ-BW-LABEL: load_i64_stride5_vf32:6357; AVX512DQ-BW: # %bb.0:6358; AVX512DQ-BW-NEXT: subq $648, %rsp # imm = 0x2886359; AVX512DQ-BW-NEXT: vmovdqa64 1088(%rdi), %zmm216360; AVX512DQ-BW-NEXT: vmovdqa64 1152(%rdi), %zmm16361; AVX512DQ-BW-NEXT: vmovdqa64 768(%rdi), %zmm186362; AVX512DQ-BW-NEXT: vmovdqa64 832(%rdi), %zmm06363; AVX512DQ-BW-NEXT: vmovdqa64 384(%rdi), %zmm176364; AVX512DQ-BW-NEXT: vmovdqa64 320(%rdi), %zmm196365; AVX512DQ-BW-NEXT: vmovdqa64 448(%rdi), %zmm266366; AVX512DQ-BW-NEXT: vmovdqa64 512(%rdi), %zmm26367; AVX512DQ-BW-NEXT: vmovdqa64 (%rdi), %zmm166368; AVX512DQ-BW-NEXT: vmovdqa64 128(%rdi), %zmm256369; AVX512DQ-BW-NEXT: vmovdqa64 192(%rdi), %zmm36370; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm11 = [12,1,6,0,12,1,6,0]6371; AVX512DQ-BW-NEXT: # zmm11 = mem[0,1,2,3,0,1,2,3]6372; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, %zmm106373; AVX512DQ-BW-NEXT: vpermt2q %zmm25, %zmm11, %zmm106374; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} ymm24 = [0,5,10,15]6375; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm126376; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm46377; AVX512DQ-BW-NEXT: vpermt2q %zmm26, %zmm11, %zmm126378; AVX512DQ-BW-NEXT: vmovdqa64 %zmm19, %zmm26379; AVX512DQ-BW-NEXT: vpermt2q %zmm17, %zmm24, %zmm26380; AVX512DQ-BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6381; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, %zmm56382; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, %zmm26383; AVX512DQ-BW-NEXT: vpermt2q %zmm18, %zmm11, %zmm56384; AVX512DQ-BW-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6385; AVX512DQ-BW-NEXT: vmovdqa64 %zmm1, %zmm06386; AVX512DQ-BW-NEXT: vpermt2q %zmm21, %zmm11, %zmm06387; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6388; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} ymm28 = [1,6,11,0]6389; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm15 = [5,10,15,0,5,10,15,0]6390; AVX512DQ-BW-NEXT: # zmm15 = mem[0,1,2,3,0,1,2,3]6391; AVX512DQ-BW-NEXT: vmovdqa64 %zmm25, %zmm306392; AVX512DQ-BW-NEXT: vpermt2q %zmm3, %zmm15, %zmm306393; AVX512DQ-BW-NEXT: vmovdqa64 %zmm18, %zmm276394; AVX512DQ-BW-NEXT: vpermt2q %zmm2, %zmm15, %zmm276395; AVX512DQ-BW-NEXT: vmovdqa64 %zmm26, %zmm226396; AVX512DQ-BW-NEXT: vpermt2q %zmm4, %zmm15, %zmm226397; AVX512DQ-BW-NEXT: vmovdqa64 %zmm19, %zmm06398; AVX512DQ-BW-NEXT: vpermt2q %zmm17, %zmm28, %zmm06399; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6400; AVX512DQ-BW-NEXT: vpermi2q %zmm1, %zmm21, %zmm156401; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm23 = [6,11,0,1,6,11,0,1]6402; AVX512DQ-BW-NEXT: # zmm23 = mem[0,1,2,3,0,1,2,3]6403; AVX512DQ-BW-NEXT: vmovdqa64 %zmm25, %zmm06404; AVX512DQ-BW-NEXT: vpermt2q %zmm3, %zmm23, %zmm06405; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6406; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} ymm13 = [2,7,12,0]6407; AVX512DQ-BW-NEXT: vmovdqa64 %zmm18, %zmm06408; AVX512DQ-BW-NEXT: vpermt2q %zmm2, %zmm23, %zmm06409; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6410; AVX512DQ-BW-NEXT: vmovdqa64 %zmm26, %zmm06411; AVX512DQ-BW-NEXT: vpermt2q %zmm4, %zmm23, %zmm06412; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6413; AVX512DQ-BW-NEXT: vpermi2q %zmm1, %zmm21, %zmm236414; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm29 = [7,12,0,2,7,12,0,2]6415; AVX512DQ-BW-NEXT: # zmm29 = mem[0,1,2,3,0,1,2,3]6416; AVX512DQ-BW-NEXT: vmovdqa64 %zmm26, %zmm06417; AVX512DQ-BW-NEXT: vpermt2q %zmm4, %zmm29, %zmm06418; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6419; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm0 = [0,5,0,11,0,5,0,11]6420; AVX512DQ-BW-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3]6421; AVX512DQ-BW-NEXT: vpermt2q %zmm26, %zmm0, %zmm46422; AVX512DQ-BW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6423; AVX512DQ-BW-NEXT: vmovdqa64 %zmm25, %zmm266424; AVX512DQ-BW-NEXT: vpermt2q %zmm3, %zmm29, %zmm266425; AVX512DQ-BW-NEXT: vpermt2q %zmm25, %zmm0, %zmm36426; AVX512DQ-BW-NEXT: vmovdqu64 %zmm3, (%rsp) # 64-byte Spill6427; AVX512DQ-BW-NEXT: vmovdqa64 %zmm18, %zmm206428; AVX512DQ-BW-NEXT: vpermt2q %zmm2, %zmm29, %zmm206429; AVX512DQ-BW-NEXT: vpermi2q %zmm1, %zmm21, %zmm296430; AVX512DQ-BW-NEXT: vpermt2q %zmm21, %zmm0, %zmm16431; AVX512DQ-BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6432; AVX512DQ-BW-NEXT: vmovdqa64 %zmm19, %zmm216433; AVX512DQ-BW-NEXT: vpermt2q %zmm17, %zmm13, %zmm216434; AVX512DQ-BW-NEXT: vpermt2q %zmm18, %zmm0, %zmm26435; AVX512DQ-BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6436; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} ymm31 = [11,0,5,0]6437; AVX512DQ-BW-NEXT: vmovdqa64 %zmm17, %zmm256438; AVX512DQ-BW-NEXT: vpermt2q %zmm19, %zmm31, %zmm256439; AVX512DQ-BW-NEXT: vpermt2q %zmm19, %zmm11, %zmm176440; AVX512DQ-BW-NEXT: vmovdqa64 64(%rdi), %zmm186441; AVX512DQ-BW-NEXT: vmovdqa64 %zmm16, %zmm66442; AVX512DQ-BW-NEXT: vpermt2q %zmm18, %zmm24, %zmm66443; AVX512DQ-BW-NEXT: vmovdqa64 %zmm16, %zmm96444; AVX512DQ-BW-NEXT: vpermt2q %zmm18, %zmm28, %zmm96445; AVX512DQ-BW-NEXT: vmovdqa64 %zmm16, %zmm86446; AVX512DQ-BW-NEXT: vpermt2q %zmm18, %zmm13, %zmm86447; AVX512DQ-BW-NEXT: vmovdqa64 %zmm18, %zmm196448; AVX512DQ-BW-NEXT: vpermt2q %zmm16, %zmm31, %zmm196449; AVX512DQ-BW-NEXT: vpermt2q %zmm16, %zmm11, %zmm186450; AVX512DQ-BW-NEXT: vmovdqa64 704(%rdi), %zmm166451; AVX512DQ-BW-NEXT: vmovdqa64 640(%rdi), %zmm16452; AVX512DQ-BW-NEXT: vmovdqa64 %zmm1, %zmm36453; AVX512DQ-BW-NEXT: vpermt2q %zmm16, %zmm24, %zmm36454; AVX512DQ-BW-NEXT: vmovdqa64 1024(%rdi), %zmm26455; AVX512DQ-BW-NEXT: vmovdqa64 960(%rdi), %zmm06456; AVX512DQ-BW-NEXT: vpermi2q %zmm2, %zmm0, %zmm246457; AVX512DQ-BW-NEXT: vmovdqa64 %zmm1, %zmm76458; AVX512DQ-BW-NEXT: vpermt2q %zmm16, %zmm28, %zmm76459; AVX512DQ-BW-NEXT: vpermi2q %zmm2, %zmm0, %zmm286460; AVX512DQ-BW-NEXT: vmovdqa64 %zmm1, %zmm46461; AVX512DQ-BW-NEXT: vpermt2q %zmm16, %zmm13, %zmm46462; AVX512DQ-BW-NEXT: vpermi2q %zmm2, %zmm0, %zmm136463; AVX512DQ-BW-NEXT: vmovdqa64 %zmm16, %zmm56464; AVX512DQ-BW-NEXT: vpermt2q %zmm1, %zmm31, %zmm56465; AVX512DQ-BW-NEXT: vpermi2q %zmm0, %zmm2, %zmm316466; AVX512DQ-BW-NEXT: vpermt2q %zmm0, %zmm11, %zmm26467; AVX512DQ-BW-NEXT: vpermt2q %zmm1, %zmm11, %zmm166468; AVX512DQ-BW-NEXT: vshufi64x2 {{.*#+}} zmm14 = zmm6[0,1,2,3],zmm10[4,5,6,7]6469; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload6470; AVX512DQ-BW-NEXT: vshufi64x2 {{.*#+}} zmm12 = zmm0[0,1,2,3],zmm12[4,5,6,7]6471; AVX512DQ-BW-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3, %zmm10 # 64-byte Folded Reload6472; AVX512DQ-BW-NEXT: # zmm10 = zmm3[0,1,2,3],mem[4,5,6,7]6473; AVX512DQ-BW-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm24, %zmm6 # 64-byte Folded Reload6474; AVX512DQ-BW-NEXT: # zmm6 = zmm24[0,1,2,3],mem[4,5,6,7]6475; AVX512DQ-BW-NEXT: vmovdqa64 256(%rdi), %zmm116476; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} zmm24 = [0,1,2,3,4,5,6,11]6477; AVX512DQ-BW-NEXT: vpermt2q %zmm11, %zmm24, %zmm146478; AVX512DQ-BW-NEXT: vmovdqa64 576(%rdi), %zmm06479; AVX512DQ-BW-NEXT: vpermt2q %zmm0, %zmm24, %zmm126480; AVX512DQ-BW-NEXT: vmovdqa64 896(%rdi), %zmm36481; AVX512DQ-BW-NEXT: vpermt2q %zmm3, %zmm24, %zmm106482; AVX512DQ-BW-NEXT: vmovdqa64 1216(%rdi), %zmm16483; AVX512DQ-BW-NEXT: vpermt2q %zmm1, %zmm24, %zmm66484; AVX512DQ-BW-NEXT: movb $7, %al6485; AVX512DQ-BW-NEXT: kmovd %eax, %k16486; AVX512DQ-BW-NEXT: vmovdqa64 %zmm9, %zmm30 {%k1}6487; AVX512DQ-BW-NEXT: vmovdqa64 %zmm7, %zmm27 {%k1}6488; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload6489; AVX512DQ-BW-NEXT: vmovdqa64 %zmm7, %zmm22 {%k1}6490; AVX512DQ-BW-NEXT: vmovdqa64 %zmm28, %zmm15 {%k1}6491; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} zmm7 = [0,1,2,3,4,5,6,12]6492; AVX512DQ-BW-NEXT: vpermt2q %zmm11, %zmm7, %zmm306493; AVX512DQ-BW-NEXT: vpermt2q %zmm3, %zmm7, %zmm276494; AVX512DQ-BW-NEXT: vpermt2q %zmm0, %zmm7, %zmm226495; AVX512DQ-BW-NEXT: vpermt2q %zmm1, %zmm7, %zmm156496; AVX512DQ-BW-NEXT: movb $56, %al6497; AVX512DQ-BW-NEXT: kmovd %eax, %k16498; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload6499; AVX512DQ-BW-NEXT: vmovdqa64 %zmm7, %zmm8 {%k1}6500; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload6501; AVX512DQ-BW-NEXT: vmovdqa64 %zmm7, %zmm4 {%k1}6502; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload6503; AVX512DQ-BW-NEXT: vmovdqa64 %zmm7, %zmm21 {%k1}6504; AVX512DQ-BW-NEXT: vmovdqa64 %zmm23, %zmm13 {%k1}6505; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} zmm7 = [0,1,2,3,4,5,8,13]6506; AVX512DQ-BW-NEXT: vpermt2q %zmm11, %zmm7, %zmm86507; AVX512DQ-BW-NEXT: vpermt2q %zmm3, %zmm7, %zmm46508; AVX512DQ-BW-NEXT: vpermt2q %zmm0, %zmm7, %zmm216509; AVX512DQ-BW-NEXT: vpermt2q %zmm1, %zmm7, %zmm136510; AVX512DQ-BW-NEXT: vmovdqa64 %zmm26, %zmm19 {%k1}6511; AVX512DQ-BW-NEXT: vmovdqa64 %zmm20, %zmm5 {%k1}6512; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload6513; AVX512DQ-BW-NEXT: vmovdqa64 %zmm7, %zmm25 {%k1}6514; AVX512DQ-BW-NEXT: vmovdqa64 %zmm29, %zmm31 {%k1}6515; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} zmm7 = [0,1,2,3,4,5,9,14]6516; AVX512DQ-BW-NEXT: vpermt2q %zmm11, %zmm7, %zmm196517; AVX512DQ-BW-NEXT: vpermt2q %zmm3, %zmm7, %zmm56518; AVX512DQ-BW-NEXT: vpermt2q %zmm0, %zmm7, %zmm256519; AVX512DQ-BW-NEXT: vpermt2q %zmm1, %zmm7, %zmm316520; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload6521; AVX512DQ-BW-NEXT: vmovdqa64 %zmm7, %zmm17 {%k1}6522; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} zmm7 = [0,1,2,3,4,5,10,15]6523; AVX512DQ-BW-NEXT: vpermt2q %zmm0, %zmm7, %zmm176524; AVX512DQ-BW-NEXT: vmovdqu64 (%rsp), %zmm0 # 64-byte Reload6525; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, %zmm18 {%k1}6526; AVX512DQ-BW-NEXT: vpermt2q %zmm11, %zmm7, %zmm186527; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload6528; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, %zmm2 {%k1}6529; AVX512DQ-BW-NEXT: vpermt2q %zmm1, %zmm7, %zmm26530; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload6531; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, %zmm16 {%k1}6532; AVX512DQ-BW-NEXT: vpermt2q %zmm3, %zmm7, %zmm166533; AVX512DQ-BW-NEXT: vmovdqa64 %zmm6, 192(%rsi)6534; AVX512DQ-BW-NEXT: vmovdqa64 %zmm10, 128(%rsi)6535; AVX512DQ-BW-NEXT: vmovdqa64 %zmm12, 64(%rsi)6536; AVX512DQ-BW-NEXT: vmovdqa64 %zmm14, (%rsi)6537; AVX512DQ-BW-NEXT: vmovdqa64 %zmm15, 192(%rdx)6538; AVX512DQ-BW-NEXT: vmovdqa64 %zmm30, (%rdx)6539; AVX512DQ-BW-NEXT: vmovdqa64 %zmm22, 64(%rdx)6540; AVX512DQ-BW-NEXT: vmovdqa64 %zmm27, 128(%rdx)6541; AVX512DQ-BW-NEXT: vmovdqa64 %zmm13, 192(%rcx)6542; AVX512DQ-BW-NEXT: vmovdqa64 %zmm8, (%rcx)6543; AVX512DQ-BW-NEXT: vmovdqa64 %zmm21, 64(%rcx)6544; AVX512DQ-BW-NEXT: vmovdqa64 %zmm4, 128(%rcx)6545; AVX512DQ-BW-NEXT: vmovdqa64 %zmm31, 192(%r8)6546; AVX512DQ-BW-NEXT: vmovdqa64 %zmm19, (%r8)6547; AVX512DQ-BW-NEXT: vmovdqa64 %zmm25, 64(%r8)6548; AVX512DQ-BW-NEXT: vmovdqa64 %zmm5, 128(%r8)6549; AVX512DQ-BW-NEXT: vmovdqa64 %zmm16, 128(%r9)6550; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, 192(%r9)6551; AVX512DQ-BW-NEXT: vmovdqa64 %zmm18, (%r9)6552; AVX512DQ-BW-NEXT: vmovdqa64 %zmm17, 64(%r9)6553; AVX512DQ-BW-NEXT: addq $648, %rsp # imm = 0x2886554; AVX512DQ-BW-NEXT: vzeroupper6555; AVX512DQ-BW-NEXT: retq6556;6557; AVX512DQ-BW-FCP-LABEL: load_i64_stride5_vf32:6558; AVX512DQ-BW-FCP: # %bb.0:6559; AVX512DQ-BW-FCP-NEXT: subq $648, %rsp # imm = 0x2886560; AVX512DQ-BW-FCP-NEXT: vmovdqa64 1088(%rdi), %zmm216561; AVX512DQ-BW-FCP-NEXT: vmovdqa64 1152(%rdi), %zmm16562; AVX512DQ-BW-FCP-NEXT: vmovdqa64 768(%rdi), %zmm186563; AVX512DQ-BW-FCP-NEXT: vmovdqa64 832(%rdi), %zmm06564; AVX512DQ-BW-FCP-NEXT: vmovdqa64 384(%rdi), %zmm176565; AVX512DQ-BW-FCP-NEXT: vmovdqa64 320(%rdi), %zmm196566; AVX512DQ-BW-FCP-NEXT: vmovdqa64 448(%rdi), %zmm266567; AVX512DQ-BW-FCP-NEXT: vmovdqa64 512(%rdi), %zmm26568; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rdi), %zmm166569; AVX512DQ-BW-FCP-NEXT: vmovdqa64 128(%rdi), %zmm256570; AVX512DQ-BW-FCP-NEXT: vmovdqa64 192(%rdi), %zmm36571; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm11 = [12,1,6,0,12,1,6,0]6572; AVX512DQ-BW-FCP-NEXT: # zmm11 = mem[0,1,2,3,0,1,2,3]6573; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm106574; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm25, %zmm11, %zmm106575; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} ymm24 = [0,5,10,15]6576; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm126577; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm46578; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm26, %zmm11, %zmm126579; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm19, %zmm26580; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm17, %zmm24, %zmm26581; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6582; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm56583; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm26584; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm18, %zmm11, %zmm56585; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6586; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm06587; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm21, %zmm11, %zmm06588; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6589; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} ymm28 = [1,6,11,0]6590; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm15 = [5,10,15,0,5,10,15,0]6591; AVX512DQ-BW-FCP-NEXT: # zmm15 = mem[0,1,2,3,0,1,2,3]6592; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm25, %zmm306593; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm3, %zmm15, %zmm306594; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm18, %zmm276595; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm2, %zmm15, %zmm276596; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm26, %zmm226597; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm4, %zmm15, %zmm226598; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm19, %zmm06599; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm17, %zmm28, %zmm06600; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6601; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm1, %zmm21, %zmm156602; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm23 = [6,11,0,1,6,11,0,1]6603; AVX512DQ-BW-FCP-NEXT: # zmm23 = mem[0,1,2,3,0,1,2,3]6604; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm25, %zmm06605; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm3, %zmm23, %zmm06606; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6607; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} ymm13 = [2,7,12,0]6608; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm18, %zmm06609; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm2, %zmm23, %zmm06610; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6611; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm26, %zmm06612; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm4, %zmm23, %zmm06613; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6614; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm1, %zmm21, %zmm236615; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm29 = [7,12,0,2,7,12,0,2]6616; AVX512DQ-BW-FCP-NEXT: # zmm29 = mem[0,1,2,3,0,1,2,3]6617; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm26, %zmm06618; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm4, %zmm29, %zmm06619; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6620; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm0 = [0,5,0,11,0,5,0,11]6621; AVX512DQ-BW-FCP-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3]6622; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm26, %zmm0, %zmm46623; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6624; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm25, %zmm266625; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm3, %zmm29, %zmm266626; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm25, %zmm0, %zmm36627; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm3, (%rsp) # 64-byte Spill6628; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm18, %zmm206629; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm2, %zmm29, %zmm206630; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm1, %zmm21, %zmm296631; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm21, %zmm0, %zmm16632; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6633; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm19, %zmm216634; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm17, %zmm13, %zmm216635; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm18, %zmm0, %zmm26636; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill6637; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} ymm31 = [11,0,5,0]6638; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm17, %zmm256639; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm19, %zmm31, %zmm256640; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm19, %zmm11, %zmm176641; AVX512DQ-BW-FCP-NEXT: vmovdqa64 64(%rdi), %zmm186642; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm16, %zmm66643; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm18, %zmm24, %zmm66644; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm16, %zmm96645; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm18, %zmm28, %zmm96646; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm16, %zmm86647; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm18, %zmm13, %zmm86648; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm18, %zmm196649; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm16, %zmm31, %zmm196650; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm16, %zmm11, %zmm186651; AVX512DQ-BW-FCP-NEXT: vmovdqa64 704(%rdi), %zmm166652; AVX512DQ-BW-FCP-NEXT: vmovdqa64 640(%rdi), %zmm16653; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm36654; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm16, %zmm24, %zmm36655; AVX512DQ-BW-FCP-NEXT: vmovdqa64 1024(%rdi), %zmm26656; AVX512DQ-BW-FCP-NEXT: vmovdqa64 960(%rdi), %zmm06657; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm2, %zmm0, %zmm246658; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm76659; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm16, %zmm28, %zmm76660; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm2, %zmm0, %zmm286661; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm46662; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm16, %zmm13, %zmm46663; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm2, %zmm0, %zmm136664; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm16, %zmm56665; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm1, %zmm31, %zmm56666; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm0, %zmm2, %zmm316667; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm0, %zmm11, %zmm26668; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm1, %zmm11, %zmm166669; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm14 = zmm6[0,1,2,3],zmm10[4,5,6,7]6670; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload6671; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm12 = zmm0[0,1,2,3],zmm12[4,5,6,7]6672; AVX512DQ-BW-FCP-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm3, %zmm10 # 64-byte Folded Reload6673; AVX512DQ-BW-FCP-NEXT: # zmm10 = zmm3[0,1,2,3],mem[4,5,6,7]6674; AVX512DQ-BW-FCP-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm24, %zmm6 # 64-byte Folded Reload6675; AVX512DQ-BW-FCP-NEXT: # zmm6 = zmm24[0,1,2,3],mem[4,5,6,7]6676; AVX512DQ-BW-FCP-NEXT: vmovdqa64 256(%rdi), %zmm116677; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm24 = [0,1,2,3,4,5,6,11]6678; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm11, %zmm24, %zmm146679; AVX512DQ-BW-FCP-NEXT: vmovdqa64 576(%rdi), %zmm06680; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm0, %zmm24, %zmm126681; AVX512DQ-BW-FCP-NEXT: vmovdqa64 896(%rdi), %zmm36682; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm3, %zmm24, %zmm106683; AVX512DQ-BW-FCP-NEXT: vmovdqa64 1216(%rdi), %zmm16684; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm1, %zmm24, %zmm66685; AVX512DQ-BW-FCP-NEXT: movb $7, %al6686; AVX512DQ-BW-FCP-NEXT: kmovd %eax, %k16687; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm9, %zmm30 {%k1}6688; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm7, %zmm27 {%k1}6689; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload6690; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm7, %zmm22 {%k1}6691; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm28, %zmm15 {%k1}6692; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm7 = [0,1,2,3,4,5,6,12]6693; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm11, %zmm7, %zmm306694; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm3, %zmm7, %zmm276695; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm0, %zmm7, %zmm226696; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm1, %zmm7, %zmm156697; AVX512DQ-BW-FCP-NEXT: movb $56, %al6698; AVX512DQ-BW-FCP-NEXT: kmovd %eax, %k16699; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload6700; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm7, %zmm8 {%k1}6701; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload6702; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm7, %zmm4 {%k1}6703; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload6704; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm7, %zmm21 {%k1}6705; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm23, %zmm13 {%k1}6706; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm7 = [0,1,2,3,4,5,8,13]6707; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm11, %zmm7, %zmm86708; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm3, %zmm7, %zmm46709; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm0, %zmm7, %zmm216710; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm1, %zmm7, %zmm136711; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm26, %zmm19 {%k1}6712; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm20, %zmm5 {%k1}6713; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload6714; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm7, %zmm25 {%k1}6715; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm29, %zmm31 {%k1}6716; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm7 = [0,1,2,3,4,5,9,14]6717; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm11, %zmm7, %zmm196718; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm3, %zmm7, %zmm56719; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm0, %zmm7, %zmm256720; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm1, %zmm7, %zmm316721; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm7 # 64-byte Reload6722; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm7, %zmm17 {%k1}6723; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm7 = [0,1,2,3,4,5,10,15]6724; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm0, %zmm7, %zmm176725; AVX512DQ-BW-FCP-NEXT: vmovdqu64 (%rsp), %zmm0 # 64-byte Reload6726; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm18 {%k1}6727; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm11, %zmm7, %zmm186728; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload6729; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm2 {%k1}6730; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm1, %zmm7, %zmm26731; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload6732; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm16 {%k1}6733; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm3, %zmm7, %zmm166734; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm6, 192(%rsi)6735; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm10, 128(%rsi)6736; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm12, 64(%rsi)6737; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm14, (%rsi)6738; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm15, 192(%rdx)6739; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm30, (%rdx)6740; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm22, 64(%rdx)6741; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm27, 128(%rdx)6742; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm13, 192(%rcx)6743; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm8, (%rcx)6744; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm21, 64(%rcx)6745; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm4, 128(%rcx)6746; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm31, 192(%r8)6747; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm19, (%r8)6748; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm25, 64(%r8)6749; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm5, 128(%r8)6750; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm16, 128(%r9)6751; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, 192(%r9)6752; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm18, (%r9)6753; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm17, 64(%r9)6754; AVX512DQ-BW-FCP-NEXT: addq $648, %rsp # imm = 0x2886755; AVX512DQ-BW-FCP-NEXT: vzeroupper6756; AVX512DQ-BW-FCP-NEXT: retq6757 %wide.vec = load <160 x i64>, ptr %in.vec, align 646758 %strided.vec0 = shufflevector <160 x i64> %wide.vec, <160 x i64> poison, <32 x i32> <i32 0, i32 5, i32 10, i32 15, i32 20, i32 25, i32 30, i32 35, i32 40, i32 45, i32 50, i32 55, i32 60, i32 65, i32 70, i32 75, i32 80, i32 85, i32 90, i32 95, i32 100, i32 105, i32 110, i32 115, i32 120, i32 125, i32 130, i32 135, i32 140, i32 145, i32 150, i32 155>6759 %strided.vec1 = shufflevector <160 x i64> %wide.vec, <160 x i64> poison, <32 x i32> <i32 1, i32 6, i32 11, i32 16, i32 21, i32 26, i32 31, i32 36, i32 41, i32 46, i32 51, i32 56, i32 61, i32 66, i32 71, i32 76, i32 81, i32 86, i32 91, i32 96, i32 101, i32 106, i32 111, i32 116, i32 121, i32 126, i32 131, i32 136, i32 141, i32 146, i32 151, i32 156>6760 %strided.vec2 = shufflevector <160 x i64> %wide.vec, <160 x i64> poison, <32 x i32> <i32 2, i32 7, i32 12, i32 17, i32 22, i32 27, i32 32, i32 37, i32 42, i32 47, i32 52, i32 57, i32 62, i32 67, i32 72, i32 77, i32 82, i32 87, i32 92, i32 97, i32 102, i32 107, i32 112, i32 117, i32 122, i32 127, i32 132, i32 137, i32 142, i32 147, i32 152, i32 157>6761 %strided.vec3 = shufflevector <160 x i64> %wide.vec, <160 x i64> poison, <32 x i32> <i32 3, i32 8, i32 13, i32 18, i32 23, i32 28, i32 33, i32 38, i32 43, i32 48, i32 53, i32 58, i32 63, i32 68, i32 73, i32 78, i32 83, i32 88, i32 93, i32 98, i32 103, i32 108, i32 113, i32 118, i32 123, i32 128, i32 133, i32 138, i32 143, i32 148, i32 153, i32 158>6762 %strided.vec4 = shufflevector <160 x i64> %wide.vec, <160 x i64> poison, <32 x i32> <i32 4, i32 9, i32 14, i32 19, i32 24, i32 29, i32 34, i32 39, i32 44, i32 49, i32 54, i32 59, i32 64, i32 69, i32 74, i32 79, i32 84, i32 89, i32 94, i32 99, i32 104, i32 109, i32 114, i32 119, i32 124, i32 129, i32 134, i32 139, i32 144, i32 149, i32 154, i32 159>6763 store <32 x i64> %strided.vec0, ptr %out.vec0, align 646764 store <32 x i64> %strided.vec1, ptr %out.vec1, align 646765 store <32 x i64> %strided.vec2, ptr %out.vec2, align 646766 store <32 x i64> %strided.vec3, ptr %out.vec3, align 646767 store <32 x i64> %strided.vec4, ptr %out.vec4, align 646768 ret void6769}6770 6771define void @load_i64_stride5_vf64(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr %out.vec2, ptr %out.vec3, ptr %out.vec4) nounwind {6772; SSE-LABEL: load_i64_stride5_vf64:6773; SSE: # %bb.0:6774; SSE-NEXT: subq $2216, %rsp # imm = 0x8A86775; SSE-NEXT: movapd 224(%rdi), %xmm56776; SSE-NEXT: movapd 144(%rdi), %xmm46777; SSE-NEXT: movapd 64(%rdi), %xmm36778; SSE-NEXT: movapd 176(%rdi), %xmm76779; SSE-NEXT: movapd 96(%rdi), %xmm66780; SSE-NEXT: movapd 208(%rdi), %xmm106781; SSE-NEXT: movapd 128(%rdi), %xmm96782; SSE-NEXT: movapd (%rdi), %xmm126783; SSE-NEXT: movapd 16(%rdi), %xmm86784; SSE-NEXT: movapd 32(%rdi), %xmm06785; SSE-NEXT: movapd 48(%rdi), %xmm116786; SSE-NEXT: movapd 160(%rdi), %xmm136787; SSE-NEXT: movapd 192(%rdi), %xmm16788; SSE-NEXT: movapd 80(%rdi), %xmm146789; SSE-NEXT: movapd 112(%rdi), %xmm26790; SSE-NEXT: movapd %xmm0, %xmm156791; SSE-NEXT: movsd {{.*#+}} xmm15 = xmm12[0],xmm15[1]6792; SSE-NEXT: movapd %xmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6793; SSE-NEXT: shufpd {{.*#+}} xmm12 = xmm12[1],xmm11[0]6794; SSE-NEXT: movapd %xmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6795; SSE-NEXT: movsd {{.*#+}} xmm11 = xmm8[0],xmm11[1]6796; SSE-NEXT: movapd %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6797; SSE-NEXT: shufpd {{.*#+}} xmm8 = xmm8[1],xmm3[0]6798; SSE-NEXT: movapd %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6799; SSE-NEXT: movsd {{.*#+}} xmm3 = xmm0[0],xmm3[1]6800; SSE-NEXT: movapd %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6801; SSE-NEXT: movapd %xmm2, %xmm06802; SSE-NEXT: movsd {{.*#+}} xmm0 = xmm14[0],xmm0[1]6803; SSE-NEXT: movapd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6804; SSE-NEXT: shufpd {{.*#+}} xmm14 = xmm14[1],xmm9[0]6805; SSE-NEXT: movapd %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6806; SSE-NEXT: movsd {{.*#+}} xmm9 = xmm6[0],xmm9[1]6807; SSE-NEXT: movapd %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6808; SSE-NEXT: shufpd {{.*#+}} xmm6 = xmm6[1],xmm4[0]6809; SSE-NEXT: movapd %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6810; SSE-NEXT: movsd {{.*#+}} xmm4 = xmm2[0],xmm4[1]6811; SSE-NEXT: movapd %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6812; SSE-NEXT: movapd %xmm1, %xmm06813; SSE-NEXT: movsd {{.*#+}} xmm0 = xmm13[0],xmm0[1]6814; SSE-NEXT: movapd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6815; SSE-NEXT: shufpd {{.*#+}} xmm13 = xmm13[1],xmm10[0]6816; SSE-NEXT: movapd %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6817; SSE-NEXT: movsd {{.*#+}} xmm10 = xmm7[0],xmm10[1]6818; SSE-NEXT: movapd %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6819; SSE-NEXT: shufpd {{.*#+}} xmm7 = xmm7[1],xmm5[0]6820; SSE-NEXT: movapd %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6821; SSE-NEXT: movsd {{.*#+}} xmm5 = xmm1[0],xmm5[1]6822; SSE-NEXT: movapd %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6823; SSE-NEXT: movapd 240(%rdi), %xmm26824; SSE-NEXT: movapd 272(%rdi), %xmm06825; SSE-NEXT: movapd %xmm0, %xmm16826; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]6827; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6828; SSE-NEXT: movapd 288(%rdi), %xmm16829; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]6830; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6831; SSE-NEXT: movapd 256(%rdi), %xmm26832; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]6833; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6834; SSE-NEXT: movapd 304(%rdi), %xmm16835; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]6836; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6837; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]6838; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6839; SSE-NEXT: movapd 320(%rdi), %xmm26840; SSE-NEXT: movapd 352(%rdi), %xmm06841; SSE-NEXT: movapd %xmm0, %xmm16842; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]6843; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6844; SSE-NEXT: movapd 368(%rdi), %xmm16845; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]6846; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6847; SSE-NEXT: movapd 336(%rdi), %xmm26848; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]6849; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6850; SSE-NEXT: movapd 384(%rdi), %xmm16851; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]6852; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6853; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]6854; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6855; SSE-NEXT: movapd 400(%rdi), %xmm26856; SSE-NEXT: movapd 432(%rdi), %xmm06857; SSE-NEXT: movapd %xmm0, %xmm16858; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]6859; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6860; SSE-NEXT: movapd 448(%rdi), %xmm16861; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]6862; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6863; SSE-NEXT: movapd 416(%rdi), %xmm26864; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]6865; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6866; SSE-NEXT: movapd 464(%rdi), %xmm16867; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]6868; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6869; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]6870; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6871; SSE-NEXT: movapd 480(%rdi), %xmm26872; SSE-NEXT: movapd 512(%rdi), %xmm06873; SSE-NEXT: movapd %xmm0, %xmm16874; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]6875; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6876; SSE-NEXT: movapd 528(%rdi), %xmm16877; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]6878; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6879; SSE-NEXT: movapd 496(%rdi), %xmm26880; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]6881; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6882; SSE-NEXT: movapd 544(%rdi), %xmm16883; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]6884; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6885; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]6886; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6887; SSE-NEXT: movapd 560(%rdi), %xmm26888; SSE-NEXT: movapd 592(%rdi), %xmm06889; SSE-NEXT: movapd %xmm0, %xmm16890; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]6891; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6892; SSE-NEXT: movapd 608(%rdi), %xmm16893; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]6894; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6895; SSE-NEXT: movapd 576(%rdi), %xmm26896; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]6897; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6898; SSE-NEXT: movapd 624(%rdi), %xmm16899; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]6900; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6901; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]6902; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6903; SSE-NEXT: movapd 640(%rdi), %xmm26904; SSE-NEXT: movapd 672(%rdi), %xmm06905; SSE-NEXT: movapd %xmm0, %xmm16906; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]6907; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6908; SSE-NEXT: movapd 688(%rdi), %xmm16909; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]6910; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6911; SSE-NEXT: movapd 656(%rdi), %xmm26912; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]6913; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6914; SSE-NEXT: movapd 704(%rdi), %xmm16915; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]6916; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6917; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]6918; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6919; SSE-NEXT: movapd 720(%rdi), %xmm26920; SSE-NEXT: movapd 752(%rdi), %xmm06921; SSE-NEXT: movapd %xmm0, %xmm16922; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]6923; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6924; SSE-NEXT: movapd 768(%rdi), %xmm16925; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]6926; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6927; SSE-NEXT: movapd 736(%rdi), %xmm26928; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]6929; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6930; SSE-NEXT: movapd 784(%rdi), %xmm16931; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]6932; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6933; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]6934; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6935; SSE-NEXT: movapd 800(%rdi), %xmm26936; SSE-NEXT: movapd 832(%rdi), %xmm06937; SSE-NEXT: movapd %xmm0, %xmm16938; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]6939; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6940; SSE-NEXT: movapd 848(%rdi), %xmm16941; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]6942; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6943; SSE-NEXT: movapd 816(%rdi), %xmm26944; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]6945; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6946; SSE-NEXT: movapd 864(%rdi), %xmm16947; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]6948; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6949; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]6950; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6951; SSE-NEXT: movapd 880(%rdi), %xmm26952; SSE-NEXT: movapd 912(%rdi), %xmm06953; SSE-NEXT: movapd %xmm0, %xmm16954; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]6955; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6956; SSE-NEXT: movapd 928(%rdi), %xmm16957; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]6958; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6959; SSE-NEXT: movapd 896(%rdi), %xmm26960; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]6961; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6962; SSE-NEXT: movapd 944(%rdi), %xmm16963; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]6964; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6965; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]6966; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6967; SSE-NEXT: movapd 960(%rdi), %xmm26968; SSE-NEXT: movapd 992(%rdi), %xmm06969; SSE-NEXT: movapd %xmm0, %xmm16970; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]6971; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6972; SSE-NEXT: movapd 1008(%rdi), %xmm16973; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]6974; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6975; SSE-NEXT: movapd 976(%rdi), %xmm26976; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]6977; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6978; SSE-NEXT: movapd 1024(%rdi), %xmm16979; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]6980; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6981; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]6982; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6983; SSE-NEXT: movapd 1040(%rdi), %xmm26984; SSE-NEXT: movapd 1072(%rdi), %xmm06985; SSE-NEXT: movapd %xmm0, %xmm16986; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]6987; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6988; SSE-NEXT: movapd 1088(%rdi), %xmm16989; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]6990; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6991; SSE-NEXT: movapd 1056(%rdi), %xmm26992; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]6993; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6994; SSE-NEXT: movapd 1104(%rdi), %xmm16995; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]6996; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6997; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]6998; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill6999; SSE-NEXT: movapd 1120(%rdi), %xmm27000; SSE-NEXT: movapd 1152(%rdi), %xmm07001; SSE-NEXT: movapd %xmm0, %xmm17002; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]7003; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7004; SSE-NEXT: movapd 1168(%rdi), %xmm17005; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]7006; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7007; SSE-NEXT: movapd 1136(%rdi), %xmm27008; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]7009; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7010; SSE-NEXT: movapd 1184(%rdi), %xmm17011; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]7012; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7013; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]7014; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7015; SSE-NEXT: movapd 1200(%rdi), %xmm27016; SSE-NEXT: movapd 1232(%rdi), %xmm07017; SSE-NEXT: movapd %xmm0, %xmm17018; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]7019; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7020; SSE-NEXT: movapd 1248(%rdi), %xmm17021; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]7022; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7023; SSE-NEXT: movapd 1216(%rdi), %xmm27024; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]7025; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7026; SSE-NEXT: movapd 1264(%rdi), %xmm17027; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]7028; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7029; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]7030; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7031; SSE-NEXT: movapd 1280(%rdi), %xmm27032; SSE-NEXT: movapd 1312(%rdi), %xmm07033; SSE-NEXT: movapd %xmm0, %xmm17034; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]7035; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7036; SSE-NEXT: movapd 1328(%rdi), %xmm17037; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]7038; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7039; SSE-NEXT: movapd 1296(%rdi), %xmm27040; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]7041; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7042; SSE-NEXT: movapd 1344(%rdi), %xmm17043; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]7044; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7045; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]7046; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7047; SSE-NEXT: movapd 1360(%rdi), %xmm27048; SSE-NEXT: movapd 1392(%rdi), %xmm07049; SSE-NEXT: movapd %xmm0, %xmm17050; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]7051; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7052; SSE-NEXT: movapd 1408(%rdi), %xmm17053; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]7054; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7055; SSE-NEXT: movapd 1376(%rdi), %xmm27056; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]7057; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7058; SSE-NEXT: movapd 1424(%rdi), %xmm17059; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]7060; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7061; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]7062; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7063; SSE-NEXT: movapd 1440(%rdi), %xmm27064; SSE-NEXT: movapd 1472(%rdi), %xmm07065; SSE-NEXT: movapd %xmm0, %xmm17066; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]7067; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7068; SSE-NEXT: movapd 1488(%rdi), %xmm17069; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]7070; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7071; SSE-NEXT: movapd 1456(%rdi), %xmm27072; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]7073; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7074; SSE-NEXT: movapd 1504(%rdi), %xmm17075; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]7076; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7077; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]7078; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7079; SSE-NEXT: movapd 1520(%rdi), %xmm27080; SSE-NEXT: movapd 1552(%rdi), %xmm07081; SSE-NEXT: movapd %xmm0, %xmm17082; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]7083; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7084; SSE-NEXT: movapd 1568(%rdi), %xmm17085; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]7086; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7087; SSE-NEXT: movapd 1536(%rdi), %xmm27088; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]7089; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7090; SSE-NEXT: movapd 1584(%rdi), %xmm17091; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]7092; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7093; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]7094; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7095; SSE-NEXT: movapd 1600(%rdi), %xmm27096; SSE-NEXT: movapd 1632(%rdi), %xmm07097; SSE-NEXT: movapd %xmm0, %xmm17098; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]7099; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7100; SSE-NEXT: movapd 1648(%rdi), %xmm17101; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]7102; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7103; SSE-NEXT: movapd 1616(%rdi), %xmm27104; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]7105; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7106; SSE-NEXT: movapd 1664(%rdi), %xmm17107; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]7108; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7109; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]7110; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7111; SSE-NEXT: movapd 1680(%rdi), %xmm27112; SSE-NEXT: movapd 1712(%rdi), %xmm07113; SSE-NEXT: movapd %xmm0, %xmm17114; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]7115; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7116; SSE-NEXT: movapd 1728(%rdi), %xmm17117; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]7118; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7119; SSE-NEXT: movapd 1696(%rdi), %xmm27120; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]7121; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7122; SSE-NEXT: movapd 1744(%rdi), %xmm17123; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]7124; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7125; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]7126; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7127; SSE-NEXT: movapd 1760(%rdi), %xmm27128; SSE-NEXT: movapd 1792(%rdi), %xmm07129; SSE-NEXT: movapd %xmm0, %xmm17130; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]7131; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7132; SSE-NEXT: movapd 1808(%rdi), %xmm17133; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]7134; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7135; SSE-NEXT: movapd 1776(%rdi), %xmm27136; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]7137; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7138; SSE-NEXT: movapd 1824(%rdi), %xmm17139; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]7140; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7141; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]7142; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7143; SSE-NEXT: movapd 1840(%rdi), %xmm27144; SSE-NEXT: movapd 1872(%rdi), %xmm07145; SSE-NEXT: movapd %xmm0, %xmm17146; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]7147; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7148; SSE-NEXT: movapd 1888(%rdi), %xmm17149; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]7150; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7151; SSE-NEXT: movapd 1856(%rdi), %xmm27152; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]7153; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7154; SSE-NEXT: movapd 1904(%rdi), %xmm17155; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]7156; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7157; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]7158; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7159; SSE-NEXT: movapd 1920(%rdi), %xmm27160; SSE-NEXT: movapd 1952(%rdi), %xmm07161; SSE-NEXT: movapd %xmm0, %xmm17162; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]7163; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7164; SSE-NEXT: movapd 1968(%rdi), %xmm17165; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]7166; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7167; SSE-NEXT: movapd 1936(%rdi), %xmm27168; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]7169; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7170; SSE-NEXT: movapd 1984(%rdi), %xmm17171; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]7172; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7173; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]7174; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7175; SSE-NEXT: movapd 2000(%rdi), %xmm27176; SSE-NEXT: movapd 2032(%rdi), %xmm07177; SSE-NEXT: movapd %xmm0, %xmm17178; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]7179; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7180; SSE-NEXT: movapd 2048(%rdi), %xmm17181; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]7182; SSE-NEXT: movapd %xmm2, (%rsp) # 16-byte Spill7183; SSE-NEXT: movapd 2016(%rdi), %xmm27184; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]7185; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7186; SSE-NEXT: movapd 2064(%rdi), %xmm17187; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]7188; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7189; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]7190; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7191; SSE-NEXT: movapd 2080(%rdi), %xmm27192; SSE-NEXT: movapd 2112(%rdi), %xmm07193; SSE-NEXT: movapd %xmm0, %xmm17194; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]7195; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7196; SSE-NEXT: movapd 2128(%rdi), %xmm17197; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]7198; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7199; SSE-NEXT: movapd 2096(%rdi), %xmm27200; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]7201; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7202; SSE-NEXT: movapd 2144(%rdi), %xmm17203; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]7204; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7205; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]7206; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7207; SSE-NEXT: movapd 2160(%rdi), %xmm147208; SSE-NEXT: movapd 2192(%rdi), %xmm07209; SSE-NEXT: movapd %xmm0, %xmm117210; SSE-NEXT: movsd {{.*#+}} xmm11 = xmm14[0],xmm11[1]7211; SSE-NEXT: movapd 2208(%rdi), %xmm17212; SSE-NEXT: shufpd {{.*#+}} xmm14 = xmm14[1],xmm1[0]7213; SSE-NEXT: movapd 2176(%rdi), %xmm27214; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]7215; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7216; SSE-NEXT: movapd 2224(%rdi), %xmm17217; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]7218; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7219; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]7220; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7221; SSE-NEXT: movapd 2240(%rdi), %xmm127222; SSE-NEXT: movapd 2272(%rdi), %xmm07223; SSE-NEXT: movapd %xmm0, %xmm87224; SSE-NEXT: movsd {{.*#+}} xmm8 = xmm12[0],xmm8[1]7225; SSE-NEXT: movapd 2288(%rdi), %xmm17226; SSE-NEXT: shufpd {{.*#+}} xmm12 = xmm12[1],xmm1[0]7227; SSE-NEXT: movapd 2256(%rdi), %xmm27228; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm2[0],xmm1[1]7229; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7230; SSE-NEXT: movapd 2304(%rdi), %xmm17231; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]7232; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7233; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]7234; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7235; SSE-NEXT: movapd 2320(%rdi), %xmm97236; SSE-NEXT: movapd 2352(%rdi), %xmm07237; SSE-NEXT: movapd %xmm0, %xmm67238; SSE-NEXT: movsd {{.*#+}} xmm6 = xmm9[0],xmm6[1]7239; SSE-NEXT: movapd 2368(%rdi), %xmm157240; SSE-NEXT: shufpd {{.*#+}} xmm9 = xmm9[1],xmm15[0]7241; SSE-NEXT: movapd 2336(%rdi), %xmm27242; SSE-NEXT: movsd {{.*#+}} xmm15 = xmm2[0],xmm15[1]7243; SSE-NEXT: movapd 2384(%rdi), %xmm17244; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]7245; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7246; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]7247; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7248; SSE-NEXT: movapd 2400(%rdi), %xmm77249; SSE-NEXT: movapd 2432(%rdi), %xmm07250; SSE-NEXT: movapd %xmm0, %xmm47251; SSE-NEXT: movsd {{.*#+}} xmm4 = xmm7[0],xmm4[1]7252; SSE-NEXT: movapd 2448(%rdi), %xmm137253; SSE-NEXT: shufpd {{.*#+}} xmm7 = xmm7[1],xmm13[0]7254; SSE-NEXT: movapd 2416(%rdi), %xmm27255; SSE-NEXT: movsd {{.*#+}} xmm13 = xmm2[0],xmm13[1]7256; SSE-NEXT: movapd 2464(%rdi), %xmm17257; SSE-NEXT: shufpd {{.*#+}} xmm2 = xmm2[1],xmm1[0]7258; SSE-NEXT: movapd %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7259; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm0[0],xmm1[1]7260; SSE-NEXT: movapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7261; SSE-NEXT: movapd 2480(%rdi), %xmm57262; SSE-NEXT: movapd 2512(%rdi), %xmm37263; SSE-NEXT: movapd %xmm3, %xmm27264; SSE-NEXT: movsd {{.*#+}} xmm2 = xmm5[0],xmm2[1]7265; SSE-NEXT: movapd 2528(%rdi), %xmm107266; SSE-NEXT: shufpd {{.*#+}} xmm5 = xmm5[1],xmm10[0]7267; SSE-NEXT: movapd 2496(%rdi), %xmm07268; SSE-NEXT: movsd {{.*#+}} xmm10 = xmm0[0],xmm10[1]7269; SSE-NEXT: movapd 2544(%rdi), %xmm17270; SSE-NEXT: shufpd {{.*#+}} xmm0 = xmm0[1],xmm1[0]7271; SSE-NEXT: movapd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7272; SSE-NEXT: movsd {{.*#+}} xmm1 = xmm3[0],xmm1[1]7273; SSE-NEXT: movapd %xmm2, 496(%rsi)7274; SSE-NEXT: movapd %xmm4, 480(%rsi)7275; SSE-NEXT: movapd %xmm6, 464(%rsi)7276; SSE-NEXT: movapd %xmm8, 448(%rsi)7277; SSE-NEXT: movapd %xmm11, 432(%rsi)7278; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7279; SSE-NEXT: movaps %xmm0, 416(%rsi)7280; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7281; SSE-NEXT: movaps %xmm0, 400(%rsi)7282; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7283; SSE-NEXT: movaps %xmm0, 384(%rsi)7284; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7285; SSE-NEXT: movaps %xmm0, 368(%rsi)7286; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7287; SSE-NEXT: movaps %xmm0, 352(%rsi)7288; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7289; SSE-NEXT: movaps %xmm0, 336(%rsi)7290; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7291; SSE-NEXT: movaps %xmm0, 320(%rsi)7292; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7293; SSE-NEXT: movaps %xmm0, 304(%rsi)7294; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7295; SSE-NEXT: movaps %xmm0, 288(%rsi)7296; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7297; SSE-NEXT: movaps %xmm0, 272(%rsi)7298; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7299; SSE-NEXT: movaps %xmm0, 256(%rsi)7300; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7301; SSE-NEXT: movaps %xmm0, 240(%rsi)7302; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7303; SSE-NEXT: movaps %xmm0, 224(%rsi)7304; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7305; SSE-NEXT: movaps %xmm0, 208(%rsi)7306; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7307; SSE-NEXT: movaps %xmm0, 192(%rsi)7308; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7309; SSE-NEXT: movaps %xmm0, 176(%rsi)7310; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7311; SSE-NEXT: movaps %xmm0, 160(%rsi)7312; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7313; SSE-NEXT: movaps %xmm0, 144(%rsi)7314; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7315; SSE-NEXT: movaps %xmm0, 128(%rsi)7316; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7317; SSE-NEXT: movaps %xmm0, 112(%rsi)7318; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7319; SSE-NEXT: movaps %xmm0, 96(%rsi)7320; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7321; SSE-NEXT: movaps %xmm0, 80(%rsi)7322; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7323; SSE-NEXT: movaps %xmm0, 64(%rsi)7324; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7325; SSE-NEXT: movaps %xmm0, 48(%rsi)7326; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7327; SSE-NEXT: movaps %xmm0, 32(%rsi)7328; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7329; SSE-NEXT: movaps %xmm0, 16(%rsi)7330; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7331; SSE-NEXT: movaps %xmm0, (%rsi)7332; SSE-NEXT: movapd %xmm5, 496(%rdx)7333; SSE-NEXT: movapd %xmm7, 480(%rdx)7334; SSE-NEXT: movapd %xmm9, 464(%rdx)7335; SSE-NEXT: movapd %xmm12, 448(%rdx)7336; SSE-NEXT: movapd %xmm14, 432(%rdx)7337; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7338; SSE-NEXT: movaps %xmm0, 416(%rdx)7339; SSE-NEXT: movaps (%rsp), %xmm0 # 16-byte Reload7340; SSE-NEXT: movaps %xmm0, 400(%rdx)7341; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7342; SSE-NEXT: movaps %xmm0, 384(%rdx)7343; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7344; SSE-NEXT: movaps %xmm0, 368(%rdx)7345; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7346; SSE-NEXT: movaps %xmm0, 352(%rdx)7347; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7348; SSE-NEXT: movaps %xmm0, 336(%rdx)7349; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7350; SSE-NEXT: movaps %xmm0, 320(%rdx)7351; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7352; SSE-NEXT: movaps %xmm0, 304(%rdx)7353; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7354; SSE-NEXT: movaps %xmm0, 288(%rdx)7355; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7356; SSE-NEXT: movaps %xmm0, 272(%rdx)7357; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7358; SSE-NEXT: movaps %xmm0, 256(%rdx)7359; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7360; SSE-NEXT: movaps %xmm0, 240(%rdx)7361; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7362; SSE-NEXT: movaps %xmm0, 224(%rdx)7363; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7364; SSE-NEXT: movaps %xmm0, 208(%rdx)7365; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7366; SSE-NEXT: movaps %xmm0, 192(%rdx)7367; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7368; SSE-NEXT: movaps %xmm0, 176(%rdx)7369; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7370; SSE-NEXT: movaps %xmm0, 160(%rdx)7371; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7372; SSE-NEXT: movaps %xmm0, 144(%rdx)7373; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7374; SSE-NEXT: movaps %xmm0, 128(%rdx)7375; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7376; SSE-NEXT: movaps %xmm0, 112(%rdx)7377; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7378; SSE-NEXT: movaps %xmm0, 96(%rdx)7379; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7380; SSE-NEXT: movaps %xmm0, 80(%rdx)7381; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7382; SSE-NEXT: movaps %xmm0, 64(%rdx)7383; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7384; SSE-NEXT: movaps %xmm0, 48(%rdx)7385; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7386; SSE-NEXT: movaps %xmm0, 32(%rdx)7387; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7388; SSE-NEXT: movaps %xmm0, 16(%rdx)7389; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7390; SSE-NEXT: movaps %xmm0, (%rdx)7391; SSE-NEXT: movapd %xmm10, 496(%rcx)7392; SSE-NEXT: movapd %xmm13, 480(%rcx)7393; SSE-NEXT: movapd %xmm15, 464(%rcx)7394; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7395; SSE-NEXT: movaps %xmm0, 448(%rcx)7396; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7397; SSE-NEXT: movaps %xmm0, 432(%rcx)7398; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7399; SSE-NEXT: movaps %xmm0, 416(%rcx)7400; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7401; SSE-NEXT: movaps %xmm0, 400(%rcx)7402; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7403; SSE-NEXT: movaps %xmm0, 384(%rcx)7404; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7405; SSE-NEXT: movaps %xmm0, 368(%rcx)7406; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7407; SSE-NEXT: movaps %xmm0, 352(%rcx)7408; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7409; SSE-NEXT: movaps %xmm0, 336(%rcx)7410; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7411; SSE-NEXT: movaps %xmm0, 320(%rcx)7412; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7413; SSE-NEXT: movaps %xmm0, 304(%rcx)7414; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7415; SSE-NEXT: movaps %xmm0, 288(%rcx)7416; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7417; SSE-NEXT: movaps %xmm0, 272(%rcx)7418; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7419; SSE-NEXT: movaps %xmm0, 256(%rcx)7420; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7421; SSE-NEXT: movaps %xmm0, 240(%rcx)7422; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7423; SSE-NEXT: movaps %xmm0, 224(%rcx)7424; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7425; SSE-NEXT: movaps %xmm0, 208(%rcx)7426; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7427; SSE-NEXT: movaps %xmm0, 192(%rcx)7428; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7429; SSE-NEXT: movaps %xmm0, 176(%rcx)7430; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7431; SSE-NEXT: movaps %xmm0, 160(%rcx)7432; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7433; SSE-NEXT: movaps %xmm0, 144(%rcx)7434; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7435; SSE-NEXT: movaps %xmm0, 128(%rcx)7436; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7437; SSE-NEXT: movaps %xmm0, 112(%rcx)7438; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7439; SSE-NEXT: movaps %xmm0, 96(%rcx)7440; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7441; SSE-NEXT: movaps %xmm0, 80(%rcx)7442; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7443; SSE-NEXT: movaps %xmm0, 64(%rcx)7444; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7445; SSE-NEXT: movaps %xmm0, 48(%rcx)7446; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7447; SSE-NEXT: movaps %xmm0, 32(%rcx)7448; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7449; SSE-NEXT: movaps %xmm0, 16(%rcx)7450; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7451; SSE-NEXT: movaps %xmm0, (%rcx)7452; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7453; SSE-NEXT: movaps %xmm0, 496(%r8)7454; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7455; SSE-NEXT: movaps %xmm0, 480(%r8)7456; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7457; SSE-NEXT: movaps %xmm0, 464(%r8)7458; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7459; SSE-NEXT: movaps %xmm0, 448(%r8)7460; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7461; SSE-NEXT: movaps %xmm0, 432(%r8)7462; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7463; SSE-NEXT: movaps %xmm0, 416(%r8)7464; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7465; SSE-NEXT: movaps %xmm0, 400(%r8)7466; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7467; SSE-NEXT: movaps %xmm0, 384(%r8)7468; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7469; SSE-NEXT: movaps %xmm0, 368(%r8)7470; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7471; SSE-NEXT: movaps %xmm0, 352(%r8)7472; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7473; SSE-NEXT: movaps %xmm0, 336(%r8)7474; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7475; SSE-NEXT: movaps %xmm0, 320(%r8)7476; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7477; SSE-NEXT: movaps %xmm0, 304(%r8)7478; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7479; SSE-NEXT: movaps %xmm0, 288(%r8)7480; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7481; SSE-NEXT: movaps %xmm0, 272(%r8)7482; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7483; SSE-NEXT: movaps %xmm0, 256(%r8)7484; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7485; SSE-NEXT: movaps %xmm0, 240(%r8)7486; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7487; SSE-NEXT: movaps %xmm0, 224(%r8)7488; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7489; SSE-NEXT: movaps %xmm0, 208(%r8)7490; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7491; SSE-NEXT: movaps %xmm0, 192(%r8)7492; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7493; SSE-NEXT: movaps %xmm0, 176(%r8)7494; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7495; SSE-NEXT: movaps %xmm0, 160(%r8)7496; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7497; SSE-NEXT: movaps %xmm0, 144(%r8)7498; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7499; SSE-NEXT: movaps %xmm0, 128(%r8)7500; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7501; SSE-NEXT: movaps %xmm0, 112(%r8)7502; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7503; SSE-NEXT: movaps %xmm0, 96(%r8)7504; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7505; SSE-NEXT: movaps %xmm0, 80(%r8)7506; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7507; SSE-NEXT: movaps %xmm0, 64(%r8)7508; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7509; SSE-NEXT: movaps %xmm0, 48(%r8)7510; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7511; SSE-NEXT: movaps %xmm0, 32(%r8)7512; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7513; SSE-NEXT: movaps %xmm0, 16(%r8)7514; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7515; SSE-NEXT: movaps %xmm0, (%r8)7516; SSE-NEXT: movapd %xmm1, 496(%r9)7517; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7518; SSE-NEXT: movaps %xmm0, 480(%r9)7519; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7520; SSE-NEXT: movaps %xmm0, 464(%r9)7521; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7522; SSE-NEXT: movaps %xmm0, 448(%r9)7523; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7524; SSE-NEXT: movaps %xmm0, 432(%r9)7525; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7526; SSE-NEXT: movaps %xmm0, 416(%r9)7527; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7528; SSE-NEXT: movaps %xmm0, 400(%r9)7529; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7530; SSE-NEXT: movaps %xmm0, 384(%r9)7531; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7532; SSE-NEXT: movaps %xmm0, 368(%r9)7533; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7534; SSE-NEXT: movaps %xmm0, 352(%r9)7535; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7536; SSE-NEXT: movaps %xmm0, 336(%r9)7537; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7538; SSE-NEXT: movaps %xmm0, 320(%r9)7539; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7540; SSE-NEXT: movaps %xmm0, 304(%r9)7541; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7542; SSE-NEXT: movaps %xmm0, 288(%r9)7543; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7544; SSE-NEXT: movaps %xmm0, 272(%r9)7545; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7546; SSE-NEXT: movaps %xmm0, 256(%r9)7547; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7548; SSE-NEXT: movaps %xmm0, 240(%r9)7549; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7550; SSE-NEXT: movaps %xmm0, 224(%r9)7551; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7552; SSE-NEXT: movaps %xmm0, 208(%r9)7553; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7554; SSE-NEXT: movaps %xmm0, 192(%r9)7555; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7556; SSE-NEXT: movaps %xmm0, 176(%r9)7557; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7558; SSE-NEXT: movaps %xmm0, 160(%r9)7559; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7560; SSE-NEXT: movaps %xmm0, 144(%r9)7561; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7562; SSE-NEXT: movaps %xmm0, 128(%r9)7563; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7564; SSE-NEXT: movaps %xmm0, 112(%r9)7565; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7566; SSE-NEXT: movaps %xmm0, 96(%r9)7567; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7568; SSE-NEXT: movaps %xmm0, 80(%r9)7569; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7570; SSE-NEXT: movaps %xmm0, 64(%r9)7571; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7572; SSE-NEXT: movaps %xmm0, 48(%r9)7573; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7574; SSE-NEXT: movaps %xmm0, 32(%r9)7575; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7576; SSE-NEXT: movaps %xmm0, 16(%r9)7577; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload7578; SSE-NEXT: movaps %xmm0, (%r9)7579; SSE-NEXT: addq $2216, %rsp # imm = 0x8A87580; SSE-NEXT: retq7581;7582; AVX-LABEL: load_i64_stride5_vf64:7583; AVX: # %bb.0:7584; AVX-NEXT: subq $3256, %rsp # imm = 0xCB87585; AVX-NEXT: vmovapd 896(%rdi), %ymm27586; AVX-NEXT: vmovupd %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7587; AVX-NEXT: vmovapd 864(%rdi), %ymm47588; AVX-NEXT: vmovapd 576(%rdi), %ymm37589; AVX-NEXT: vmovupd %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7590; AVX-NEXT: vmovapd 544(%rdi), %ymm57591; AVX-NEXT: vmovapd 256(%rdi), %ymm07592; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7593; AVX-NEXT: vmovapd 224(%rdi), %ymm77594; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm7[0,1,2],ymm0[3]7595; AVX-NEXT: vmovapd 192(%rdi), %xmm17596; AVX-NEXT: vmovapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7597; AVX-NEXT: vmovapd 160(%rdi), %xmm117598; AVX-NEXT: vblendpd {{.*#+}} xmm1 = xmm11[0],xmm1[1]7599; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3]7600; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7601; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm5[0,1,2],ymm3[3]7602; AVX-NEXT: vmovapd 512(%rdi), %xmm17603; AVX-NEXT: vmovapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7604; AVX-NEXT: vmovapd 480(%rdi), %xmm87605; AVX-NEXT: vblendpd {{.*#+}} xmm1 = xmm8[0],xmm1[1]7606; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3]7607; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7608; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm4[0,1,2],ymm2[3]7609; AVX-NEXT: vmovapd 832(%rdi), %xmm17610; AVX-NEXT: vmovapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7611; AVX-NEXT: vmovapd 800(%rdi), %xmm37612; AVX-NEXT: vblendpd {{.*#+}} xmm1 = xmm3[0],xmm1[1]7613; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3]7614; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7615; AVX-NEXT: vmovaps 1216(%rdi), %ymm07616; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7617; AVX-NEXT: vmovaps 1184(%rdi), %ymm17618; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7619; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]7620; AVX-NEXT: vmovaps 1152(%rdi), %xmm17621; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7622; AVX-NEXT: vmovaps 1120(%rdi), %xmm97623; AVX-NEXT: vblendps {{.*#+}} xmm1 = xmm9[0,1],xmm1[2,3]7624; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]7625; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7626; AVX-NEXT: vmovapd 1536(%rdi), %ymm07627; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7628; AVX-NEXT: vmovapd 1504(%rdi), %ymm67629; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm6[0,1,2],ymm0[3]7630; AVX-NEXT: vmovapd 1472(%rdi), %xmm17631; AVX-NEXT: vmovapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7632; AVX-NEXT: vmovapd 1440(%rdi), %xmm127633; AVX-NEXT: vblendpd {{.*#+}} xmm1 = xmm12[0],xmm1[1]7634; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3]7635; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7636; AVX-NEXT: vmovapd 1856(%rdi), %ymm07637; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7638; AVX-NEXT: vmovapd 1824(%rdi), %ymm107639; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm10[0,1,2],ymm0[3]7640; AVX-NEXT: vmovapd 1792(%rdi), %xmm17641; AVX-NEXT: vmovapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7642; AVX-NEXT: vmovapd 1760(%rdi), %xmm137643; AVX-NEXT: vblendpd {{.*#+}} xmm1 = xmm13[0],xmm1[1]7644; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3]7645; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7646; AVX-NEXT: vmovapd 2176(%rdi), %ymm07647; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7648; AVX-NEXT: vmovapd 2144(%rdi), %ymm147649; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm14[0,1,2],ymm0[3]7650; AVX-NEXT: vmovapd 2112(%rdi), %xmm17651; AVX-NEXT: vmovapd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7652; AVX-NEXT: vmovapd 2080(%rdi), %xmm157653; AVX-NEXT: vblendpd {{.*#+}} xmm1 = xmm15[0],xmm1[1]7654; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3]7655; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7656; AVX-NEXT: vmovaps 2496(%rdi), %ymm07657; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7658; AVX-NEXT: vmovaps 2464(%rdi), %ymm17659; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7660; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm1[0,1,2,3,4,5],ymm0[6,7]7661; AVX-NEXT: vmovaps 2432(%rdi), %xmm17662; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7663; AVX-NEXT: vmovaps 2400(%rdi), %xmm07664; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7665; AVX-NEXT: vblendps {{.*#+}} xmm1 = xmm0[0,1],xmm1[2,3]7666; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm2[4,5,6,7]7667; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7668; AVX-NEXT: vmovaps 96(%rdi), %ymm07669; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7670; AVX-NEXT: vmovaps 64(%rdi), %ymm17671; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7672; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]7673; AVX-NEXT: vmovaps (%rdi), %xmm27674; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7675; AVX-NEXT: vmovaps 32(%rdi), %xmm17676; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7677; AVX-NEXT: vblendps {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3]7678; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]7679; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7680; AVX-NEXT: vmovaps 416(%rdi), %ymm07681; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7682; AVX-NEXT: vmovaps 384(%rdi), %ymm17683; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7684; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm1[0,1,2,3,4,5],ymm0[6,7]7685; AVX-NEXT: vmovaps 352(%rdi), %xmm17686; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7687; AVX-NEXT: vmovaps 320(%rdi), %xmm07688; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7689; AVX-NEXT: vblendps {{.*#+}} xmm1 = xmm0[0,1],xmm1[2,3]7690; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm2[4,5,6,7]7691; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7692; AVX-NEXT: vmovaps 736(%rdi), %ymm07693; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7694; AVX-NEXT: vmovaps 704(%rdi), %ymm17695; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7696; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]7697; AVX-NEXT: vmovaps 672(%rdi), %xmm17698; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7699; AVX-NEXT: vmovaps 640(%rdi), %xmm27700; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7701; AVX-NEXT: vblendps {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3]7702; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]7703; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7704; AVX-NEXT: vmovaps 1056(%rdi), %ymm07705; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7706; AVX-NEXT: vmovaps 1024(%rdi), %ymm17707; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7708; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm1[0,1,2,3,4,5],ymm0[6,7]7709; AVX-NEXT: vmovaps 992(%rdi), %xmm17710; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7711; AVX-NEXT: vmovaps 960(%rdi), %xmm07712; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7713; AVX-NEXT: vblendps {{.*#+}} xmm1 = xmm0[0,1],xmm1[2,3]7714; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm2[4,5,6,7]7715; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7716; AVX-NEXT: vmovaps 1376(%rdi), %ymm07717; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7718; AVX-NEXT: vmovaps 1344(%rdi), %ymm17719; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7720; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]7721; AVX-NEXT: vmovaps 1312(%rdi), %xmm17722; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7723; AVX-NEXT: vmovaps 1280(%rdi), %xmm27724; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7725; AVX-NEXT: vblendps {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3]7726; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]7727; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7728; AVX-NEXT: vmovaps 1696(%rdi), %ymm07729; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7730; AVX-NEXT: vmovaps 1664(%rdi), %ymm17731; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7732; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm1[0,1,2,3,4,5],ymm0[6,7]7733; AVX-NEXT: vmovaps 1632(%rdi), %xmm17734; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7735; AVX-NEXT: vmovaps 1600(%rdi), %xmm07736; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7737; AVX-NEXT: vblendps {{.*#+}} xmm1 = xmm0[0,1],xmm1[2,3]7738; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm2[4,5,6,7]7739; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7740; AVX-NEXT: vmovaps 2016(%rdi), %ymm07741; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7742; AVX-NEXT: vmovaps 1984(%rdi), %ymm17743; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7744; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]7745; AVX-NEXT: vmovaps 1952(%rdi), %xmm17746; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7747; AVX-NEXT: vmovaps 1920(%rdi), %xmm27748; AVX-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7749; AVX-NEXT: vblendps {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3]7750; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]7751; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7752; AVX-NEXT: vmovaps 2336(%rdi), %ymm07753; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7754; AVX-NEXT: vmovaps 2304(%rdi), %ymm17755; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7756; AVX-NEXT: vblendps {{.*#+}} ymm2 = ymm1[0,1,2,3,4,5],ymm0[6,7]7757; AVX-NEXT: vmovaps 2272(%rdi), %xmm17758; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7759; AVX-NEXT: vmovaps 2240(%rdi), %xmm07760; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7761; AVX-NEXT: vblendps {{.*#+}} xmm1 = xmm0[0,1],xmm1[2,3]7762; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm2[4,5,6,7]7763; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7764; AVX-NEXT: vmovaps 288(%rdi), %xmm07765; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7766; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm07767; AVX-NEXT: vshufpd {{.*#+}} ymm0 = ymm7[0],ymm0[1],ymm7[3],ymm0[2]7768; AVX-NEXT: vmovdqa 208(%rdi), %xmm17769; AVX-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7770; AVX-NEXT: vpalignr {{.*#+}} xmm1 = xmm11[8,9,10,11,12,13,14,15],xmm1[0,1,2,3,4,5,6,7]7771; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3]7772; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7773; AVX-NEXT: vmovaps 608(%rdi), %xmm07774; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7775; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm07776; AVX-NEXT: vshufpd {{.*#+}} ymm0 = ymm5[0],ymm0[1],ymm5[3],ymm0[2]7777; AVX-NEXT: vmovdqa 528(%rdi), %xmm17778; AVX-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7779; AVX-NEXT: vpalignr {{.*#+}} xmm1 = xmm8[8,9,10,11,12,13,14,15],xmm1[0,1,2,3,4,5,6,7]7780; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3]7781; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7782; AVX-NEXT: vmovaps 928(%rdi), %xmm07783; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7784; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm07785; AVX-NEXT: vshufpd {{.*#+}} ymm0 = ymm4[0],ymm0[1],ymm4[3],ymm0[2]7786; AVX-NEXT: vmovdqa 848(%rdi), %xmm17787; AVX-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7788; AVX-NEXT: vpalignr {{.*#+}} xmm1 = xmm3[8,9,10,11,12,13,14,15],xmm1[0,1,2,3,4,5,6,7]7789; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3]7790; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7791; AVX-NEXT: vmovaps 1248(%rdi), %xmm07792; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7793; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm07794; AVX-NEXT: vmovupd {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload7795; AVX-NEXT: vshufpd {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[3],ymm0[2]7796; AVX-NEXT: vmovdqa 1168(%rdi), %xmm17797; AVX-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7798; AVX-NEXT: vpalignr {{.*#+}} xmm1 = xmm9[8,9,10,11,12,13,14,15],xmm1[0,1,2,3,4,5,6,7]7799; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3]7800; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7801; AVX-NEXT: vmovaps 1568(%rdi), %xmm07802; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7803; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm07804; AVX-NEXT: vshufpd {{.*#+}} ymm0 = ymm6[0],ymm0[1],ymm6[3],ymm0[2]7805; AVX-NEXT: vmovdqa 1488(%rdi), %xmm17806; AVX-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7807; AVX-NEXT: vpalignr {{.*#+}} xmm1 = xmm12[8,9,10,11,12,13,14,15],xmm1[0,1,2,3,4,5,6,7]7808; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3]7809; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7810; AVX-NEXT: vmovaps 1888(%rdi), %xmm117811; AVX-NEXT: vinsertf128 $1, %xmm11, %ymm0, %ymm07812; AVX-NEXT: vshufpd {{.*#+}} ymm0 = ymm10[0],ymm0[1],ymm10[3],ymm0[2]7813; AVX-NEXT: vmovdqa 1808(%rdi), %xmm127814; AVX-NEXT: vpalignr {{.*#+}} xmm1 = xmm13[8,9,10,11,12,13,14,15],xmm12[0,1,2,3,4,5,6,7]7815; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3]7816; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7817; AVX-NEXT: vmovaps 2208(%rdi), %xmm137818; AVX-NEXT: vinsertf128 $1, %xmm13, %ymm0, %ymm07819; AVX-NEXT: vshufpd {{.*#+}} ymm0 = ymm14[0],ymm0[1],ymm14[3],ymm0[2]7820; AVX-NEXT: vmovdqa 2128(%rdi), %xmm147821; AVX-NEXT: vpalignr {{.*#+}} xmm1 = xmm15[8,9,10,11,12,13,14,15],xmm14[0,1,2,3,4,5,6,7]7822; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3]7823; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7824; AVX-NEXT: vmovaps 2528(%rdi), %xmm157825; AVX-NEXT: vinsertf128 $1, %xmm15, %ymm0, %ymm07826; AVX-NEXT: vmovupd {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload7827; AVX-NEXT: vshufpd {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[3],ymm0[2]7828; AVX-NEXT: vmovdqa 2448(%rdi), %xmm17829; AVX-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7830; AVX-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload7831; AVX-NEXT: # xmm1 = mem[8,9,10,11,12,13,14,15],xmm1[0,1,2,3,4,5,6,7]7832; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3]7833; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7834; AVX-NEXT: vmovaps 128(%rdi), %xmm07835; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7836; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm07837; AVX-NEXT: vmovupd {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload7838; AVX-NEXT: vshufpd {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[3],ymm0[2]7839; AVX-NEXT: vmovdqa 48(%rdi), %xmm17840; AVX-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7841; AVX-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload7842; AVX-NEXT: # xmm1 = mem[8,9,10,11,12,13,14,15],xmm1[0,1,2,3,4,5,6,7]7843; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3]7844; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7845; AVX-NEXT: vmovaps 448(%rdi), %xmm07846; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7847; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm07848; AVX-NEXT: vmovupd {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload7849; AVX-NEXT: vshufpd {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[3],ymm0[2]7850; AVX-NEXT: vmovdqa 368(%rdi), %xmm17851; AVX-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7852; AVX-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload7853; AVX-NEXT: # xmm1 = mem[8,9,10,11,12,13,14,15],xmm1[0,1,2,3,4,5,6,7]7854; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3]7855; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7856; AVX-NEXT: vmovaps 768(%rdi), %xmm07857; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7858; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm07859; AVX-NEXT: vmovupd {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload7860; AVX-NEXT: vshufpd {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[3],ymm0[2]7861; AVX-NEXT: vmovdqa 688(%rdi), %xmm17862; AVX-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7863; AVX-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload7864; AVX-NEXT: # xmm1 = mem[8,9,10,11,12,13,14,15],xmm1[0,1,2,3,4,5,6,7]7865; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3]7866; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7867; AVX-NEXT: vmovaps 1088(%rdi), %xmm07868; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7869; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm07870; AVX-NEXT: vmovupd {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload7871; AVX-NEXT: vshufpd {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[3],ymm0[2]7872; AVX-NEXT: vmovdqa 1008(%rdi), %xmm17873; AVX-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7874; AVX-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload7875; AVX-NEXT: # xmm1 = mem[8,9,10,11,12,13,14,15],xmm1[0,1,2,3,4,5,6,7]7876; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3]7877; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7878; AVX-NEXT: vmovaps 1408(%rdi), %xmm87879; AVX-NEXT: vinsertf128 $1, %xmm8, %ymm0, %ymm07880; AVX-NEXT: vmovupd {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload7881; AVX-NEXT: vshufpd {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[3],ymm0[2]7882; AVX-NEXT: vmovdqa 1328(%rdi), %xmm107883; AVX-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm10, %xmm1 # 16-byte Folded Reload7884; AVX-NEXT: # xmm1 = mem[8,9,10,11,12,13,14,15],xmm10[0,1,2,3,4,5,6,7]7885; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3]7886; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7887; AVX-NEXT: vmovaps 1728(%rdi), %xmm67888; AVX-NEXT: vinsertf128 $1, %xmm6, %ymm0, %ymm07889; AVX-NEXT: vmovupd {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload7890; AVX-NEXT: vshufpd {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[3],ymm0[2]7891; AVX-NEXT: vmovdqa 1648(%rdi), %xmm77892; AVX-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm7, %xmm1 # 16-byte Folded Reload7893; AVX-NEXT: # xmm1 = mem[8,9,10,11,12,13,14,15],xmm7[0,1,2,3,4,5,6,7]7894; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3]7895; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7896; AVX-NEXT: vmovaps 2048(%rdi), %xmm47897; AVX-NEXT: vinsertf128 $1, %xmm4, %ymm0, %ymm07898; AVX-NEXT: vmovupd {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload7899; AVX-NEXT: vshufpd {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[3],ymm0[2]7900; AVX-NEXT: vmovdqa 1968(%rdi), %xmm57901; AVX-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm5, %xmm1 # 16-byte Folded Reload7902; AVX-NEXT: # xmm1 = mem[8,9,10,11,12,13,14,15],xmm5[0,1,2,3,4,5,6,7]7903; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3]7904; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7905; AVX-NEXT: vmovaps 2368(%rdi), %xmm27906; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm17907; AVX-NEXT: vmovupd {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload7908; AVX-NEXT: vshufpd {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[3],ymm1[2]7909; AVX-NEXT: vmovdqa 2288(%rdi), %xmm37910; AVX-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm9 # 16-byte Folded Reload7911; AVX-NEXT: # xmm9 = mem[8,9,10,11,12,13,14,15],xmm3[0,1,2,3,4,5,6,7]7912; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm9[0,1],ymm0[2,3]7913; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7914; AVX-NEXT: vmovaps 256(%rdi), %xmm07915; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7916; AVX-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload7917; AVX-NEXT: # xmm0 = xmm0[0,1],mem[2,3]7918; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm07919; AVX-NEXT: vmovaps 176(%rdi), %xmm17920; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7921; AVX-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm9 # 16-byte Folded Reload7922; AVX-NEXT: # xmm9 = xmm1[0,1],mem[2,3]7923; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm9[0,1,2,3],ymm0[4,5,6,7]7924; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7925; AVX-NEXT: vmovaps 576(%rdi), %xmm07926; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7927; AVX-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload7928; AVX-NEXT: # xmm0 = xmm0[0,1],mem[2,3]7929; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm07930; AVX-NEXT: vmovaps 496(%rdi), %xmm17931; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7932; AVX-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm9 # 16-byte Folded Reload7933; AVX-NEXT: # xmm9 = xmm1[0,1],mem[2,3]7934; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm9[0,1,2,3],ymm0[4,5,6,7]7935; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7936; AVX-NEXT: vmovaps 896(%rdi), %xmm07937; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7938; AVX-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload7939; AVX-NEXT: # xmm0 = xmm0[0,1],mem[2,3]7940; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm07941; AVX-NEXT: vmovaps 816(%rdi), %xmm17942; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7943; AVX-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm9 # 16-byte Folded Reload7944; AVX-NEXT: # xmm9 = xmm1[0,1],mem[2,3]7945; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm9[0,1,2,3],ymm0[4,5,6,7]7946; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7947; AVX-NEXT: vmovaps 1216(%rdi), %xmm07948; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7949; AVX-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload7950; AVX-NEXT: # xmm0 = xmm0[0,1],mem[2,3]7951; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm07952; AVX-NEXT: vmovaps 1136(%rdi), %xmm17953; AVX-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7954; AVX-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm9 # 16-byte Folded Reload7955; AVX-NEXT: # xmm9 = xmm1[0,1],mem[2,3]7956; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm9[0,1,2,3],ymm0[4,5,6,7]7957; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7958; AVX-NEXT: vmovaps 1536(%rdi), %xmm07959; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7960; AVX-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload7961; AVX-NEXT: # xmm0 = xmm0[0,1],mem[2,3]7962; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm17963; AVX-NEXT: vmovaps 1456(%rdi), %xmm07964; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7965; AVX-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm9 # 16-byte Folded Reload7966; AVX-NEXT: # xmm9 = xmm0[0,1],mem[2,3]7967; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm9[0,1,2,3],ymm1[4,5,6,7]7968; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7969; AVX-NEXT: vmovaps 1856(%rdi), %xmm07970; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7971; AVX-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm11[2,3]7972; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm07973; AVX-NEXT: vmovdqa 1776(%rdi), %xmm97974; AVX-NEXT: vmovdqa %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7975; AVX-NEXT: vpblendw {{.*#+}} xmm9 = xmm9[0,1,2,3],xmm12[4,5,6,7]7976; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm9[0,1,2,3],ymm0[4,5,6,7]7977; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7978; AVX-NEXT: vmovaps 2176(%rdi), %xmm07979; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7980; AVX-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm13[2,3]7981; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm07982; AVX-NEXT: vmovdqa 2096(%rdi), %xmm97983; AVX-NEXT: vmovdqa %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7984; AVX-NEXT: vpblendw {{.*#+}} xmm9 = xmm9[0,1,2,3],xmm14[4,5,6,7]7985; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm9[0,1,2,3],ymm0[4,5,6,7]7986; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7987; AVX-NEXT: vmovaps 2496(%rdi), %xmm07988; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7989; AVX-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm15[2,3]7990; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm07991; AVX-NEXT: vmovaps 2416(%rdi), %xmm157992; AVX-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm9 # 16-byte Folded Reload7993; AVX-NEXT: # xmm9 = xmm15[0,1],mem[2,3]7994; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm9[0,1,2,3],ymm0[4,5,6,7]7995; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill7996; AVX-NEXT: vmovaps 2336(%rdi), %xmm07997; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill7998; AVX-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm2[2,3]7999; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm08000; AVX-NEXT: vmovdqa 2256(%rdi), %xmm28001; AVX-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8002; AVX-NEXT: vpblendw {{.*#+}} xmm1 = xmm2[0,1,2,3],xmm3[4,5,6,7]8003; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]8004; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8005; AVX-NEXT: vmovaps 2016(%rdi), %xmm08006; AVX-NEXT: vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8007; AVX-NEXT: vblendps {{.*#+}} xmm0 = xmm0[0,1],xmm4[2,3]8008; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm08009; AVX-NEXT: vmovdqa 1936(%rdi), %xmm18010; AVX-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8011; AVX-NEXT: vpblendw {{.*#+}} xmm1 = xmm1[0,1,2,3],xmm5[4,5,6,7]8012; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]8013; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8014; AVX-NEXT: vmovaps 1696(%rdi), %xmm128015; AVX-NEXT: vblendps {{.*#+}} xmm0 = xmm12[0,1],xmm6[2,3]8016; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm08017; AVX-NEXT: vmovdqa 1616(%rdi), %xmm18018; AVX-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8019; AVX-NEXT: vpblendw {{.*#+}} xmm1 = xmm1[0,1,2,3],xmm7[4,5,6,7]8020; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]8021; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8022; AVX-NEXT: vmovaps 1376(%rdi), %xmm28023; AVX-NEXT: vblendps {{.*#+}} xmm0 = xmm2[0,1],xmm8[2,3]8024; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm18025; AVX-NEXT: vmovdqa 1296(%rdi), %xmm08026; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm0[0,1,2,3],xmm10[4,5,6,7]8027; AVX-NEXT: vblendps {{.*#+}} ymm1 = ymm3[0,1,2,3],ymm1[4,5,6,7]8028; AVX-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8029; AVX-NEXT: vmovaps 1056(%rdi), %xmm38030; AVX-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm1 # 16-byte Folded Reload8031; AVX-NEXT: # xmm1 = xmm3[0,1],mem[2,3]8032; AVX-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm48033; AVX-NEXT: vmovaps 976(%rdi), %xmm18034; AVX-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm5 # 16-byte Folded Reload8035; AVX-NEXT: # xmm5 = xmm1[0,1],mem[2,3]8036; AVX-NEXT: vblendps {{.*#+}} ymm4 = ymm5[0,1,2,3],ymm4[4,5,6,7]8037; AVX-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8038; AVX-NEXT: vmovaps 736(%rdi), %xmm58039; AVX-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %xmm5, %xmm4 # 16-byte Folded Reload8040; AVX-NEXT: # xmm4 = xmm5[0,1],mem[2,3]8041; AVX-NEXT: vinsertf128 $1, %xmm4, %ymm0, %ymm68042; AVX-NEXT: vmovaps 656(%rdi), %xmm48043; AVX-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4, %xmm7 # 16-byte Folded Reload8044; AVX-NEXT: # xmm7 = xmm4[0,1],mem[2,3]8045; AVX-NEXT: vblendps {{.*#+}} ymm6 = ymm7[0,1,2,3],ymm6[4,5,6,7]8046; AVX-NEXT: vmovups %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8047; AVX-NEXT: vmovaps 416(%rdi), %xmm68048; AVX-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %xmm6, %xmm7 # 16-byte Folded Reload8049; AVX-NEXT: # xmm7 = xmm6[0,1],mem[2,3]8050; AVX-NEXT: vinsertf128 $1, %xmm7, %ymm0, %ymm78051; AVX-NEXT: vmovaps 336(%rdi), %xmm88052; AVX-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %xmm8, %xmm9 # 16-byte Folded Reload8053; AVX-NEXT: # xmm9 = xmm8[0,1],mem[2,3]8054; AVX-NEXT: vblendps {{.*#+}} ymm7 = ymm9[0,1,2,3],ymm7[4,5,6,7]8055; AVX-NEXT: vmovups %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8056; AVX-NEXT: vmovaps 96(%rdi), %xmm78057; AVX-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %xmm7, %xmm9 # 16-byte Folded Reload8058; AVX-NEXT: # xmm9 = xmm7[0,1],mem[2,3]8059; AVX-NEXT: vinsertf128 $1, %xmm9, %ymm0, %ymm98060; AVX-NEXT: vmovaps 16(%rdi), %xmm108061; AVX-NEXT: vblendps $12, {{[-0-9]+}}(%r{{[sb]}}p), %xmm10, %xmm11 # 16-byte Folded Reload8062; AVX-NEXT: # xmm11 = xmm10[0,1],mem[2,3]8063; AVX-NEXT: vblendps {{.*#+}} ymm9 = ymm11[0,1,2,3],ymm9[4,5,6,7]8064; AVX-NEXT: vmovups %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8065; AVX-NEXT: vinsertf128 $1, %xmm7, %ymm0, %ymm78066; AVX-NEXT: vmovapd 128(%rdi), %ymm98067; AVX-NEXT: vmovupd %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8068; AVX-NEXT: vshufpd {{.*#+}} ymm7 = ymm7[0],ymm9[1],ymm7[3],ymm9[2]8069; AVX-NEXT: vmovdqa 64(%rdi), %xmm98070; AVX-NEXT: vmovdqa %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8071; AVX-NEXT: vpalignr {{.*#+}} xmm9 = xmm10[8,9,10,11,12,13,14,15],xmm9[0,1,2,3,4,5,6,7]8072; AVX-NEXT: vblendpd {{.*#+}} ymm7 = ymm9[0,1],ymm7[2,3]8073; AVX-NEXT: vmovupd %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8074; AVX-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm7 # 16-byte Folded Reload8075; AVX-NEXT: vmovapd 288(%rdi), %ymm98076; AVX-NEXT: vmovupd %ymm9, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8077; AVX-NEXT: vshufpd {{.*#+}} ymm7 = ymm7[0],ymm9[1],ymm7[3],ymm9[2]8078; AVX-NEXT: vmovdqa 224(%rdi), %xmm98079; AVX-NEXT: vmovdqa %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8080; AVX-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm9, %xmm9 # 16-byte Folded Reload8081; AVX-NEXT: # xmm9 = mem[8,9,10,11,12,13,14,15],xmm9[0,1,2,3,4,5,6,7]8082; AVX-NEXT: vblendpd {{.*#+}} ymm7 = ymm9[0,1],ymm7[2,3]8083; AVX-NEXT: vmovupd %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8084; AVX-NEXT: vinsertf128 $1, %xmm6, %ymm0, %ymm68085; AVX-NEXT: vmovapd 448(%rdi), %ymm78086; AVX-NEXT: vmovupd %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8087; AVX-NEXT: vshufpd {{.*#+}} ymm6 = ymm6[0],ymm7[1],ymm6[3],ymm7[2]8088; AVX-NEXT: vmovdqa 384(%rdi), %xmm78089; AVX-NEXT: vmovdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8090; AVX-NEXT: vpalignr {{.*#+}} xmm7 = xmm8[8,9,10,11,12,13,14,15],xmm7[0,1,2,3,4,5,6,7]8091; AVX-NEXT: vblendpd {{.*#+}} ymm6 = ymm7[0,1],ymm6[2,3]8092; AVX-NEXT: vmovupd %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8093; AVX-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm6 # 16-byte Folded Reload8094; AVX-NEXT: vmovapd 608(%rdi), %ymm78095; AVX-NEXT: vmovupd %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8096; AVX-NEXT: vshufpd {{.*#+}} ymm6 = ymm6[0],ymm7[1],ymm6[3],ymm7[2]8097; AVX-NEXT: vmovdqa 544(%rdi), %xmm78098; AVX-NEXT: vmovdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8099; AVX-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm7, %xmm7 # 16-byte Folded Reload8100; AVX-NEXT: # xmm7 = mem[8,9,10,11,12,13,14,15],xmm7[0,1,2,3,4,5,6,7]8101; AVX-NEXT: vblendpd {{.*#+}} ymm6 = ymm7[0,1],ymm6[2,3]8102; AVX-NEXT: vmovupd %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8103; AVX-NEXT: vinsertf128 $1, %xmm5, %ymm0, %ymm58104; AVX-NEXT: vmovapd 768(%rdi), %ymm68105; AVX-NEXT: vmovupd %ymm6, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8106; AVX-NEXT: vshufpd {{.*#+}} ymm5 = ymm5[0],ymm6[1],ymm5[3],ymm6[2]8107; AVX-NEXT: vmovdqa 704(%rdi), %xmm68108; AVX-NEXT: vmovdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8109; AVX-NEXT: vpalignr {{.*#+}} xmm4 = xmm4[8,9,10,11,12,13,14,15],xmm6[0,1,2,3,4,5,6,7]8110; AVX-NEXT: vblendpd {{.*#+}} ymm4 = ymm4[0,1],ymm5[2,3]8111; AVX-NEXT: vmovupd %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8112; AVX-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm4 # 16-byte Folded Reload8113; AVX-NEXT: vmovapd 928(%rdi), %ymm58114; AVX-NEXT: vmovupd %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8115; AVX-NEXT: vshufpd {{.*#+}} ymm4 = ymm4[0],ymm5[1],ymm4[3],ymm5[2]8116; AVX-NEXT: vmovdqa 864(%rdi), %xmm58117; AVX-NEXT: vmovdqa %xmm5, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8118; AVX-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm5, %xmm5 # 16-byte Folded Reload8119; AVX-NEXT: # xmm5 = mem[8,9,10,11,12,13,14,15],xmm5[0,1,2,3,4,5,6,7]8120; AVX-NEXT: vblendpd {{.*#+}} ymm4 = ymm5[0,1],ymm4[2,3]8121; AVX-NEXT: vmovupd %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8122; AVX-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm38123; AVX-NEXT: vmovapd 1088(%rdi), %ymm48124; AVX-NEXT: vmovupd %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8125; AVX-NEXT: vshufpd {{.*#+}} ymm3 = ymm3[0],ymm4[1],ymm3[3],ymm4[2]8126; AVX-NEXT: vmovdqa 1024(%rdi), %xmm48127; AVX-NEXT: vmovdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8128; AVX-NEXT: vpalignr {{.*#+}} xmm1 = xmm1[8,9,10,11,12,13,14,15],xmm4[0,1,2,3,4,5,6,7]8129; AVX-NEXT: vblendpd {{.*#+}} ymm1 = ymm1[0,1],ymm3[2,3]8130; AVX-NEXT: vmovupd %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8131; AVX-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm1 # 16-byte Folded Reload8132; AVX-NEXT: vmovapd 1248(%rdi), %ymm38133; AVX-NEXT: vmovupd %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8134; AVX-NEXT: vshufpd {{.*#+}} ymm1 = ymm1[0],ymm3[1],ymm1[3],ymm3[2]8135; AVX-NEXT: vmovdqa 1184(%rdi), %xmm38136; AVX-NEXT: vmovdqa %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8137; AVX-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm3 # 16-byte Folded Reload8138; AVX-NEXT: # xmm3 = mem[8,9,10,11,12,13,14,15],xmm3[0,1,2,3,4,5,6,7]8139; AVX-NEXT: vblendpd {{.*#+}} ymm1 = ymm3[0,1],ymm1[2,3]8140; AVX-NEXT: vmovupd %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8141; AVX-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm18142; AVX-NEXT: vmovapd 1408(%rdi), %ymm28143; AVX-NEXT: vmovupd %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8144; AVX-NEXT: vshufpd {{.*#+}} ymm1 = ymm1[0],ymm2[1],ymm1[3],ymm2[2]8145; AVX-NEXT: vmovdqa 1344(%rdi), %xmm28146; AVX-NEXT: vmovdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8147; AVX-NEXT: vpalignr {{.*#+}} xmm0 = xmm0[8,9,10,11,12,13,14,15],xmm2[0,1,2,3,4,5,6,7]8148; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3]8149; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8150; AVX-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload8151; AVX-NEXT: vmovapd 1568(%rdi), %ymm148152; AVX-NEXT: vshufpd {{.*#+}} ymm0 = ymm0[0],ymm14[1],ymm0[3],ymm14[2]8153; AVX-NEXT: vmovdqa 1504(%rdi), %xmm138154; AVX-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm13, %xmm1 # 16-byte Folded Reload8155; AVX-NEXT: # xmm1 = mem[8,9,10,11,12,13,14,15],xmm13[0,1,2,3,4,5,6,7]8156; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3]8157; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8158; AVX-NEXT: vinsertf128 $1, %xmm12, %ymm0, %ymm08159; AVX-NEXT: vmovapd 1728(%rdi), %ymm128160; AVX-NEXT: vshufpd {{.*#+}} ymm0 = ymm0[0],ymm12[1],ymm0[3],ymm12[2]8161; AVX-NEXT: vmovdqa 1664(%rdi), %xmm118162; AVX-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm11, %xmm1 # 16-byte Folded Reload8163; AVX-NEXT: # xmm1 = mem[8,9,10,11,12,13,14,15],xmm11[0,1,2,3,4,5,6,7]8164; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3]8165; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8166; AVX-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload8167; AVX-NEXT: vmovapd 1888(%rdi), %ymm108168; AVX-NEXT: vshufpd {{.*#+}} ymm0 = ymm0[0],ymm10[1],ymm0[3],ymm10[2]8169; AVX-NEXT: vmovdqa 1824(%rdi), %xmm98170; AVX-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm9, %xmm1 # 16-byte Folded Reload8171; AVX-NEXT: # xmm1 = mem[8,9,10,11,12,13,14,15],xmm9[0,1,2,3,4,5,6,7]8172; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3]8173; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8174; AVX-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload8175; AVX-NEXT: vmovapd 2048(%rdi), %ymm88176; AVX-NEXT: vshufpd {{.*#+}} ymm0 = ymm0[0],ymm8[1],ymm0[3],ymm8[2]8177; AVX-NEXT: vmovdqa 1984(%rdi), %xmm78178; AVX-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm7, %xmm1 # 16-byte Folded Reload8179; AVX-NEXT: # xmm1 = mem[8,9,10,11,12,13,14,15],xmm7[0,1,2,3,4,5,6,7]8180; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3]8181; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8182; AVX-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload8183; AVX-NEXT: vmovapd 2208(%rdi), %ymm68184; AVX-NEXT: vshufpd {{.*#+}} ymm0 = ymm0[0],ymm6[1],ymm0[3],ymm6[2]8185; AVX-NEXT: vmovdqa 2144(%rdi), %xmm58186; AVX-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm5, %xmm1 # 16-byte Folded Reload8187; AVX-NEXT: # xmm1 = mem[8,9,10,11,12,13,14,15],xmm5[0,1,2,3,4,5,6,7]8188; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3]8189; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8190; AVX-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload8191; AVX-NEXT: vmovapd 2368(%rdi), %ymm48192; AVX-NEXT: vshufpd {{.*#+}} ymm0 = ymm0[0],ymm4[1],ymm0[3],ymm4[2]8193; AVX-NEXT: vmovdqa 2304(%rdi), %xmm38194; AVX-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm1 # 16-byte Folded Reload8195; AVX-NEXT: # xmm1 = mem[8,9,10,11,12,13,14,15],xmm3[0,1,2,3,4,5,6,7]8196; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3]8197; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8198; AVX-NEXT: vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload8199; AVX-NEXT: vmovapd 2528(%rdi), %ymm28200; AVX-NEXT: vshufpd {{.*#+}} ymm0 = ymm0[0],ymm2[1],ymm0[3],ymm2[2]8201; AVX-NEXT: vmovdqa 2464(%rdi), %xmm18202; AVX-NEXT: vpalignr {{.*#+}} xmm15 = xmm15[8,9,10,11,12,13,14,15],xmm1[0,1,2,3,4,5,6,7]8203; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm15[0,1],ymm0[2,3]8204; AVX-NEXT: vmovupd %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8205; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload8206; AVX-NEXT: vblendps $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload8207; AVX-NEXT: # ymm0 = mem[0,1,2,3,4,5],ymm0[6,7]8208; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload8209; AVX-NEXT: vblendps $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload8210; AVX-NEXT: # xmm15 = mem[0,1],xmm15[2,3]8211; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]8212; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8213; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload8214; AVX-NEXT: vblendps $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload8215; AVX-NEXT: # ymm0 = mem[0,1,2,3,4,5],ymm0[6,7]8216; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload8217; AVX-NEXT: vblendps $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload8218; AVX-NEXT: # xmm15 = mem[0,1],xmm15[2,3]8219; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]8220; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8221; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload8222; AVX-NEXT: vblendps $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload8223; AVX-NEXT: # ymm0 = mem[0,1,2,3,4,5],ymm0[6,7]8224; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload8225; AVX-NEXT: vblendps $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload8226; AVX-NEXT: # xmm15 = mem[0,1],xmm15[2,3]8227; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]8228; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8229; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload8230; AVX-NEXT: vblendps $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload8231; AVX-NEXT: # ymm0 = mem[0,1,2,3,4,5],ymm0[6,7]8232; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload8233; AVX-NEXT: vblendps $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload8234; AVX-NEXT: # xmm15 = mem[0,1],xmm15[2,3]8235; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]8236; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8237; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload8238; AVX-NEXT: vblendps $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload8239; AVX-NEXT: # ymm0 = mem[0,1,2,3,4,5],ymm0[6,7]8240; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload8241; AVX-NEXT: vblendps $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload8242; AVX-NEXT: # xmm15 = mem[0,1],xmm15[2,3]8243; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]8244; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8245; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload8246; AVX-NEXT: vblendps $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload8247; AVX-NEXT: # ymm0 = mem[0,1,2,3,4,5],ymm0[6,7]8248; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload8249; AVX-NEXT: vblendps $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload8250; AVX-NEXT: # xmm15 = mem[0,1],xmm15[2,3]8251; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]8252; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8253; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload8254; AVX-NEXT: vblendps $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload8255; AVX-NEXT: # ymm0 = mem[0,1,2,3,4,5],ymm0[6,7]8256; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload8257; AVX-NEXT: vblendps $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload8258; AVX-NEXT: # xmm15 = mem[0,1],xmm15[2,3]8259; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]8260; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8261; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload8262; AVX-NEXT: vblendps $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload8263; AVX-NEXT: # ymm0 = mem[0,1,2,3,4,5],ymm0[6,7]8264; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload8265; AVX-NEXT: vblendps $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload8266; AVX-NEXT: # xmm15 = mem[0,1],xmm15[2,3]8267; AVX-NEXT: vblendps {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]8268; AVX-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8269; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload8270; AVX-NEXT: vblendps $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload8271; AVX-NEXT: # ymm0 = mem[0,1,2,3,4,5],ymm0[6,7]8272; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload8273; AVX-NEXT: vblendps $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload8274; AVX-NEXT: # xmm15 = mem[0,1],xmm15[2,3]8275; AVX-NEXT: vblendps {{.*#+}} ymm15 = ymm15[0,1,2,3],ymm0[4,5,6,7]8276; AVX-NEXT: vblendpd $7, {{[-0-9]+}}(%r{{[sb]}}p), %ymm14, %ymm0 # 32-byte Folded Reload8277; AVX-NEXT: # ymm0 = mem[0,1,2],ymm14[3]8278; AVX-NEXT: vpblendw $15, {{[-0-9]+}}(%r{{[sb]}}p), %xmm13, %xmm13 # 16-byte Folded Reload8279; AVX-NEXT: # xmm13 = mem[0,1,2,3],xmm13[4,5,6,7]8280; AVX-NEXT: vblendpd {{.*#+}} ymm13 = ymm13[0,1],ymm0[2,3]8281; AVX-NEXT: vblendpd $7, {{[-0-9]+}}(%r{{[sb]}}p), %ymm12, %ymm0 # 32-byte Folded Reload8282; AVX-NEXT: # ymm0 = mem[0,1,2],ymm12[3]8283; AVX-NEXT: vpblendw $15, {{[-0-9]+}}(%r{{[sb]}}p), %xmm11, %xmm11 # 16-byte Folded Reload8284; AVX-NEXT: # xmm11 = mem[0,1,2,3],xmm11[4,5,6,7]8285; AVX-NEXT: vblendpd {{.*#+}} ymm11 = ymm11[0,1],ymm0[2,3]8286; AVX-NEXT: vblendpd $7, {{[-0-9]+}}(%r{{[sb]}}p), %ymm10, %ymm0 # 32-byte Folded Reload8287; AVX-NEXT: # ymm0 = mem[0,1,2],ymm10[3]8288; AVX-NEXT: vpblendw $15, {{[-0-9]+}}(%r{{[sb]}}p), %xmm9, %xmm9 # 16-byte Folded Reload8289; AVX-NEXT: # xmm9 = mem[0,1,2,3],xmm9[4,5,6,7]8290; AVX-NEXT: vblendpd {{.*#+}} ymm9 = ymm9[0,1],ymm0[2,3]8291; AVX-NEXT: vblendpd $7, {{[-0-9]+}}(%r{{[sb]}}p), %ymm8, %ymm0 # 32-byte Folded Reload8292; AVX-NEXT: # ymm0 = mem[0,1,2],ymm8[3]8293; AVX-NEXT: vpblendw $15, {{[-0-9]+}}(%r{{[sb]}}p), %xmm7, %xmm7 # 16-byte Folded Reload8294; AVX-NEXT: # xmm7 = mem[0,1,2,3],xmm7[4,5,6,7]8295; AVX-NEXT: vblendpd {{.*#+}} ymm7 = ymm7[0,1],ymm0[2,3]8296; AVX-NEXT: vblendpd $7, {{[-0-9]+}}(%r{{[sb]}}p), %ymm6, %ymm0 # 32-byte Folded Reload8297; AVX-NEXT: # ymm0 = mem[0,1,2],ymm6[3]8298; AVX-NEXT: vpblendw $15, {{[-0-9]+}}(%r{{[sb]}}p), %xmm5, %xmm5 # 16-byte Folded Reload8299; AVX-NEXT: # xmm5 = mem[0,1,2,3],xmm5[4,5,6,7]8300; AVX-NEXT: vblendpd {{.*#+}} ymm5 = ymm5[0,1],ymm0[2,3]8301; AVX-NEXT: vblendpd $7, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4, %ymm0 # 32-byte Folded Reload8302; AVX-NEXT: # ymm0 = mem[0,1,2],ymm4[3]8303; AVX-NEXT: vpblendw $15, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm3 # 16-byte Folded Reload8304; AVX-NEXT: # xmm3 = mem[0,1,2,3],xmm3[4,5,6,7]8305; AVX-NEXT: vblendpd {{.*#+}} ymm3 = ymm3[0,1],ymm0[2,3]8306; AVX-NEXT: vblendpd $7, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm0 # 32-byte Folded Reload8307; AVX-NEXT: # ymm0 = mem[0,1,2],ymm2[3]8308; AVX-NEXT: vpblendw $15, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload8309; AVX-NEXT: # xmm1 = mem[0,1,2,3],xmm1[4,5,6,7]8310; AVX-NEXT: vblendpd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3]8311; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8312; AVX-NEXT: vmovaps %ymm1, 448(%rsi)8313; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8314; AVX-NEXT: vmovaps %ymm1, 384(%rsi)8315; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8316; AVX-NEXT: vmovaps %ymm1, 320(%rsi)8317; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8318; AVX-NEXT: vmovaps %ymm1, 256(%rsi)8319; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8320; AVX-NEXT: vmovaps %ymm1, 192(%rsi)8321; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8322; AVX-NEXT: vmovaps %ymm1, 128(%rsi)8323; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8324; AVX-NEXT: vmovaps %ymm1, 64(%rsi)8325; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8326; AVX-NEXT: vmovaps %ymm1, (%rsi)8327; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8328; AVX-NEXT: vmovaps %ymm1, 480(%rsi)8329; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8330; AVX-NEXT: vmovaps %ymm1, 416(%rsi)8331; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8332; AVX-NEXT: vmovaps %ymm1, 352(%rsi)8333; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8334; AVX-NEXT: vmovaps %ymm1, 288(%rsi)8335; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8336; AVX-NEXT: vmovaps %ymm1, 224(%rsi)8337; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8338; AVX-NEXT: vmovaps %ymm1, 160(%rsi)8339; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8340; AVX-NEXT: vmovaps %ymm1, 96(%rsi)8341; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8342; AVX-NEXT: vmovaps %ymm1, 32(%rsi)8343; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8344; AVX-NEXT: vmovaps %ymm1, 448(%rdx)8345; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8346; AVX-NEXT: vmovaps %ymm1, 384(%rdx)8347; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8348; AVX-NEXT: vmovaps %ymm1, 320(%rdx)8349; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8350; AVX-NEXT: vmovaps %ymm1, 256(%rdx)8351; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8352; AVX-NEXT: vmovaps %ymm1, 192(%rdx)8353; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8354; AVX-NEXT: vmovaps %ymm1, 128(%rdx)8355; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8356; AVX-NEXT: vmovaps %ymm1, 64(%rdx)8357; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8358; AVX-NEXT: vmovaps %ymm1, (%rdx)8359; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8360; AVX-NEXT: vmovaps %ymm1, 480(%rdx)8361; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8362; AVX-NEXT: vmovaps %ymm1, 416(%rdx)8363; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8364; AVX-NEXT: vmovaps %ymm1, 352(%rdx)8365; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8366; AVX-NEXT: vmovaps %ymm1, 288(%rdx)8367; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8368; AVX-NEXT: vmovaps %ymm1, 224(%rdx)8369; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8370; AVX-NEXT: vmovaps %ymm1, 160(%rdx)8371; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8372; AVX-NEXT: vmovaps %ymm1, 96(%rdx)8373; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8374; AVX-NEXT: vmovaps %ymm1, 32(%rdx)8375; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8376; AVX-NEXT: vmovaps %ymm1, (%rcx)8377; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8378; AVX-NEXT: vmovaps %ymm1, 64(%rcx)8379; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8380; AVX-NEXT: vmovaps %ymm1, 128(%rcx)8381; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8382; AVX-NEXT: vmovaps %ymm1, 192(%rcx)8383; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8384; AVX-NEXT: vmovaps %ymm1, 256(%rcx)8385; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8386; AVX-NEXT: vmovaps %ymm1, 320(%rcx)8387; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8388; AVX-NEXT: vmovaps %ymm1, 384(%rcx)8389; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8390; AVX-NEXT: vmovaps %ymm1, 448(%rcx)8391; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8392; AVX-NEXT: vmovaps %ymm1, 480(%rcx)8393; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8394; AVX-NEXT: vmovaps %ymm1, 416(%rcx)8395; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8396; AVX-NEXT: vmovaps %ymm1, 352(%rcx)8397; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8398; AVX-NEXT: vmovaps %ymm1, 288(%rcx)8399; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8400; AVX-NEXT: vmovaps %ymm1, 224(%rcx)8401; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8402; AVX-NEXT: vmovaps %ymm1, 160(%rcx)8403; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8404; AVX-NEXT: vmovaps %ymm1, 96(%rcx)8405; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8406; AVX-NEXT: vmovaps %ymm1, 32(%rcx)8407; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8408; AVX-NEXT: vmovaps %ymm1, 480(%r8)8409; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8410; AVX-NEXT: vmovaps %ymm1, 448(%r8)8411; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8412; AVX-NEXT: vmovaps %ymm1, 416(%r8)8413; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8414; AVX-NEXT: vmovaps %ymm1, 384(%r8)8415; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8416; AVX-NEXT: vmovaps %ymm1, 352(%r8)8417; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8418; AVX-NEXT: vmovaps %ymm1, 320(%r8)8419; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8420; AVX-NEXT: vmovaps %ymm1, 288(%r8)8421; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8422; AVX-NEXT: vmovaps %ymm1, 256(%r8)8423; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8424; AVX-NEXT: vmovaps %ymm1, 224(%r8)8425; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8426; AVX-NEXT: vmovaps %ymm1, 192(%r8)8427; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8428; AVX-NEXT: vmovaps %ymm1, 160(%r8)8429; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8430; AVX-NEXT: vmovaps %ymm1, 128(%r8)8431; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8432; AVX-NEXT: vmovaps %ymm1, 96(%r8)8433; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8434; AVX-NEXT: vmovaps %ymm1, 64(%r8)8435; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8436; AVX-NEXT: vmovaps %ymm1, 32(%r8)8437; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload8438; AVX-NEXT: vmovaps %ymm1, (%r8)8439; AVX-NEXT: vmovapd %ymm0, 480(%r9)8440; AVX-NEXT: vmovapd %ymm3, 448(%r9)8441; AVX-NEXT: vmovapd %ymm5, 416(%r9)8442; AVX-NEXT: vmovapd %ymm7, 384(%r9)8443; AVX-NEXT: vmovapd %ymm9, 352(%r9)8444; AVX-NEXT: vmovapd %ymm11, 320(%r9)8445; AVX-NEXT: vmovapd %ymm13, 288(%r9)8446; AVX-NEXT: vmovaps %ymm15, 256(%r9)8447; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload8448; AVX-NEXT: vmovaps %ymm0, 224(%r9)8449; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload8450; AVX-NEXT: vmovaps %ymm0, 192(%r9)8451; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload8452; AVX-NEXT: vmovaps %ymm0, 160(%r9)8453; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload8454; AVX-NEXT: vmovaps %ymm0, 128(%r9)8455; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload8456; AVX-NEXT: vmovaps %ymm0, 96(%r9)8457; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload8458; AVX-NEXT: vmovaps %ymm0, 64(%r9)8459; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload8460; AVX-NEXT: vmovaps %ymm0, 32(%r9)8461; AVX-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload8462; AVX-NEXT: vmovaps %ymm0, (%r9)8463; AVX-NEXT: addq $3256, %rsp # imm = 0xCB88464; AVX-NEXT: vzeroupper8465; AVX-NEXT: retq8466;8467; AVX2-LABEL: load_i64_stride5_vf64:8468; AVX2: # %bb.0:8469; AVX2-NEXT: subq $3240, %rsp # imm = 0xCA88470; AVX2-NEXT: vmovaps 896(%rdi), %ymm28471; AVX2-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8472; AVX2-NEXT: vmovaps 864(%rdi), %ymm48473; AVX2-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8474; AVX2-NEXT: vmovaps 576(%rdi), %ymm38475; AVX2-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8476; AVX2-NEXT: vmovaps 544(%rdi), %ymm58477; AVX2-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8478; AVX2-NEXT: vmovaps 256(%rdi), %ymm08479; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8480; AVX2-NEXT: vmovaps 224(%rdi), %ymm18481; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8482; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]8483; AVX2-NEXT: vmovaps 192(%rdi), %xmm18484; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8485; AVX2-NEXT: vmovaps 160(%rdi), %xmm68486; AVX2-NEXT: vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8487; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm6[0,1],xmm1[2,3]8488; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]8489; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8490; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm5[0,1,2,3,4,5],ymm3[6,7]8491; AVX2-NEXT: vmovaps 512(%rdi), %xmm18492; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8493; AVX2-NEXT: vmovaps 480(%rdi), %xmm38494; AVX2-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8495; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm3[0,1],xmm1[2,3]8496; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]8497; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8498; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm4[0,1,2,3,4,5],ymm2[6,7]8499; AVX2-NEXT: vmovaps 832(%rdi), %xmm18500; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8501; AVX2-NEXT: vmovaps 800(%rdi), %xmm28502; AVX2-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8503; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3]8504; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]8505; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8506; AVX2-NEXT: vmovaps 1216(%rdi), %ymm08507; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8508; AVX2-NEXT: vmovaps 1184(%rdi), %ymm18509; AVX2-NEXT: vmovups %ymm1, (%rsp) # 32-byte Spill8510; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]8511; AVX2-NEXT: vmovaps 1152(%rdi), %xmm18512; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8513; AVX2-NEXT: vmovaps 1120(%rdi), %xmm28514; AVX2-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8515; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3]8516; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]8517; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8518; AVX2-NEXT: vmovaps 1536(%rdi), %ymm08519; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8520; AVX2-NEXT: vmovaps 1504(%rdi), %ymm18521; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8522; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]8523; AVX2-NEXT: vmovaps 1472(%rdi), %xmm18524; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8525; AVX2-NEXT: vmovaps 1440(%rdi), %xmm28526; AVX2-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8527; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3]8528; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]8529; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8530; AVX2-NEXT: vmovaps 1856(%rdi), %ymm08531; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8532; AVX2-NEXT: vmovaps 1824(%rdi), %ymm18533; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8534; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]8535; AVX2-NEXT: vmovaps 1792(%rdi), %xmm18536; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8537; AVX2-NEXT: vmovaps 1760(%rdi), %xmm28538; AVX2-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8539; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3]8540; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]8541; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8542; AVX2-NEXT: vmovaps 2176(%rdi), %ymm08543; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8544; AVX2-NEXT: vmovaps 2144(%rdi), %ymm18545; AVX2-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8546; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]8547; AVX2-NEXT: vmovaps 2112(%rdi), %xmm18548; AVX2-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8549; AVX2-NEXT: vmovaps 2080(%rdi), %xmm28550; AVX2-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8551; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3]8552; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]8553; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8554; AVX2-NEXT: vmovdqa 2496(%rdi), %ymm08555; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8556; AVX2-NEXT: vmovdqa 2464(%rdi), %ymm18557; AVX2-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8558; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]8559; AVX2-NEXT: vmovdqa 2432(%rdi), %xmm18560; AVX2-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8561; AVX2-NEXT: vmovdqa 2400(%rdi), %xmm138562; AVX2-NEXT: vpblendd {{.*#+}} xmm1 = xmm13[0,1],xmm1[2,3]8563; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]8564; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8565; AVX2-NEXT: vmovdqa 96(%rdi), %ymm08566; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8567; AVX2-NEXT: vmovdqa 64(%rdi), %ymm18568; AVX2-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8569; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]8570; AVX2-NEXT: vmovdqa (%rdi), %xmm148571; AVX2-NEXT: vmovdqa 32(%rdi), %xmm18572; AVX2-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8573; AVX2-NEXT: vpblendd {{.*#+}} xmm1 = xmm14[0,1],xmm1[2,3]8574; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]8575; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8576; AVX2-NEXT: vmovdqa 416(%rdi), %ymm08577; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8578; AVX2-NEXT: vmovdqa 384(%rdi), %ymm18579; AVX2-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8580; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]8581; AVX2-NEXT: vmovdqa 352(%rdi), %xmm18582; AVX2-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8583; AVX2-NEXT: vmovdqa 320(%rdi), %xmm88584; AVX2-NEXT: vpblendd {{.*#+}} xmm1 = xmm8[0,1],xmm1[2,3]8585; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]8586; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8587; AVX2-NEXT: vmovdqa 736(%rdi), %ymm08588; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8589; AVX2-NEXT: vmovdqa 704(%rdi), %ymm18590; AVX2-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8591; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]8592; AVX2-NEXT: vmovdqa 672(%rdi), %xmm18593; AVX2-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8594; AVX2-NEXT: vmovdqa 640(%rdi), %xmm78595; AVX2-NEXT: vpblendd {{.*#+}} xmm1 = xmm7[0,1],xmm1[2,3]8596; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]8597; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8598; AVX2-NEXT: vmovdqa 1056(%rdi), %ymm08599; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8600; AVX2-NEXT: vmovdqa 1024(%rdi), %ymm18601; AVX2-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8602; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]8603; AVX2-NEXT: vmovdqa 992(%rdi), %xmm18604; AVX2-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8605; AVX2-NEXT: vmovdqa 960(%rdi), %xmm58606; AVX2-NEXT: vpblendd {{.*#+}} xmm1 = xmm5[0,1],xmm1[2,3]8607; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]8608; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8609; AVX2-NEXT: vmovdqa 1376(%rdi), %ymm08610; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8611; AVX2-NEXT: vmovdqa 1344(%rdi), %ymm128612; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm12[0,1,2,3,4,5],ymm0[6,7]8613; AVX2-NEXT: vmovdqa 1312(%rdi), %xmm18614; AVX2-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8615; AVX2-NEXT: vmovdqa 1280(%rdi), %xmm48616; AVX2-NEXT: vpblendd {{.*#+}} xmm1 = xmm4[0,1],xmm1[2,3]8617; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]8618; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8619; AVX2-NEXT: vmovdqa 1696(%rdi), %ymm08620; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8621; AVX2-NEXT: vmovdqa 1664(%rdi), %ymm118622; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm11[0,1,2,3,4,5],ymm0[6,7]8623; AVX2-NEXT: vmovdqa 1632(%rdi), %xmm18624; AVX2-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8625; AVX2-NEXT: vmovdqa 1600(%rdi), %xmm38626; AVX2-NEXT: vpblendd {{.*#+}} xmm1 = xmm3[0,1],xmm1[2,3]8627; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]8628; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8629; AVX2-NEXT: vmovdqa 2016(%rdi), %ymm08630; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8631; AVX2-NEXT: vmovdqa 1984(%rdi), %ymm108632; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm10[0,1,2,3,4,5],ymm0[6,7]8633; AVX2-NEXT: vmovdqa 1952(%rdi), %xmm18634; AVX2-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8635; AVX2-NEXT: vmovdqa 1920(%rdi), %xmm98636; AVX2-NEXT: vpblendd {{.*#+}} xmm6 = xmm9[0,1],xmm1[2,3]8637; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm6[0,1,2,3],ymm0[4,5,6,7]8638; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8639; AVX2-NEXT: vmovdqa 2336(%rdi), %ymm08640; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8641; AVX2-NEXT: vmovdqa 2304(%rdi), %ymm68642; AVX2-NEXT: vpblendd {{.*#+}} ymm2 = ymm6[0,1,2,3,4,5],ymm0[6,7]8643; AVX2-NEXT: vmovdqa 2272(%rdi), %xmm08644; AVX2-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8645; AVX2-NEXT: vmovdqa 2240(%rdi), %xmm18646; AVX2-NEXT: vpblendd {{.*#+}} xmm15 = xmm1[0,1],xmm0[2,3]8647; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm2[4,5,6,7]8648; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8649; AVX2-NEXT: vmovdqa 208(%rdi), %xmm08650; AVX2-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload8651; AVX2-NEXT: # xmm0 = mem[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]8652; AVX2-NEXT: vmovdqa 288(%rdi), %ymm28653; AVX2-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8654; AVX2-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm15 # 32-byte Folded Reload8655; AVX2-NEXT: # ymm15 = mem[8,9,10,11,12,13,14,15],ymm2[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm2[16,17,18,19,20,21,22,23]8656; AVX2-NEXT: vpermq {{.*#+}} ymm15 = ymm15[0,1,2,1]8657; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm15[4,5,6,7]8658; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8659; AVX2-NEXT: vmovdqa 528(%rdi), %xmm08660; AVX2-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload8661; AVX2-NEXT: # xmm0 = mem[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]8662; AVX2-NEXT: vmovdqa 608(%rdi), %ymm28663; AVX2-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8664; AVX2-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm15 # 32-byte Folded Reload8665; AVX2-NEXT: # ymm15 = mem[8,9,10,11,12,13,14,15],ymm2[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm2[16,17,18,19,20,21,22,23]8666; AVX2-NEXT: vpermq {{.*#+}} ymm15 = ymm15[0,1,2,1]8667; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm15[4,5,6,7]8668; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8669; AVX2-NEXT: vmovdqa 848(%rdi), %xmm08670; AVX2-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload8671; AVX2-NEXT: # xmm0 = mem[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]8672; AVX2-NEXT: vmovdqa 928(%rdi), %ymm28673; AVX2-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8674; AVX2-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm15 # 32-byte Folded Reload8675; AVX2-NEXT: # ymm15 = mem[8,9,10,11,12,13,14,15],ymm2[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm2[16,17,18,19,20,21,22,23]8676; AVX2-NEXT: vpermq {{.*#+}} ymm15 = ymm15[0,1,2,1]8677; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm15[4,5,6,7]8678; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8679; AVX2-NEXT: vmovdqa 1168(%rdi), %xmm08680; AVX2-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload8681; AVX2-NEXT: # xmm0 = mem[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]8682; AVX2-NEXT: vmovdqa 1248(%rdi), %ymm28683; AVX2-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8684; AVX2-NEXT: vpalignr $8, (%rsp), %ymm2, %ymm15 # 32-byte Folded Reload8685; AVX2-NEXT: # ymm15 = mem[8,9,10,11,12,13,14,15],ymm2[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm2[16,17,18,19,20,21,22,23]8686; AVX2-NEXT: vpermq {{.*#+}} ymm15 = ymm15[0,1,2,1]8687; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm15[4,5,6,7]8688; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8689; AVX2-NEXT: vmovdqa 1488(%rdi), %xmm08690; AVX2-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload8691; AVX2-NEXT: # xmm0 = mem[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]8692; AVX2-NEXT: vmovdqa 1568(%rdi), %ymm28693; AVX2-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8694; AVX2-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm15 # 32-byte Folded Reload8695; AVX2-NEXT: # ymm15 = mem[8,9,10,11,12,13,14,15],ymm2[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm2[16,17,18,19,20,21,22,23]8696; AVX2-NEXT: vpermq {{.*#+}} ymm15 = ymm15[0,1,2,1]8697; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm15[4,5,6,7]8698; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8699; AVX2-NEXT: vmovdqa 1808(%rdi), %xmm08700; AVX2-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload8701; AVX2-NEXT: # xmm0 = mem[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]8702; AVX2-NEXT: vmovdqa 1888(%rdi), %ymm28703; AVX2-NEXT: vmovdqu %ymm2, (%rsp) # 32-byte Spill8704; AVX2-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm15 # 32-byte Folded Reload8705; AVX2-NEXT: # ymm15 = mem[8,9,10,11,12,13,14,15],ymm2[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm2[16,17,18,19,20,21,22,23]8706; AVX2-NEXT: vpermq {{.*#+}} ymm15 = ymm15[0,1,2,1]8707; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm15[4,5,6,7]8708; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8709; AVX2-NEXT: vmovdqa 2128(%rdi), %xmm08710; AVX2-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload8711; AVX2-NEXT: # xmm0 = mem[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]8712; AVX2-NEXT: vmovdqa 2208(%rdi), %ymm28713; AVX2-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8714; AVX2-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm15 # 32-byte Folded Reload8715; AVX2-NEXT: # ymm15 = mem[8,9,10,11,12,13,14,15],ymm2[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm2[16,17,18,19,20,21,22,23]8716; AVX2-NEXT: vpermq {{.*#+}} ymm15 = ymm15[0,1,2,1]8717; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm15[4,5,6,7]8718; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8719; AVX2-NEXT: vmovdqa 2448(%rdi), %xmm08720; AVX2-NEXT: vpalignr {{.*#+}} xmm0 = xmm13[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]8721; AVX2-NEXT: vmovdqa 2528(%rdi), %ymm28722; AVX2-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8723; AVX2-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm13 # 32-byte Folded Reload8724; AVX2-NEXT: # ymm13 = mem[8,9,10,11,12,13,14,15],ymm2[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm2[16,17,18,19,20,21,22,23]8725; AVX2-NEXT: vpermq {{.*#+}} ymm13 = ymm13[0,1,2,1]8726; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm13[4,5,6,7]8727; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8728; AVX2-NEXT: vmovdqa 48(%rdi), %xmm08729; AVX2-NEXT: vpalignr {{.*#+}} xmm0 = xmm14[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]8730; AVX2-NEXT: vmovdqa 128(%rdi), %ymm158731; AVX2-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm13 # 32-byte Folded Reload8732; AVX2-NEXT: # ymm13 = mem[8,9,10,11,12,13,14,15],ymm15[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm15[16,17,18,19,20,21,22,23]8733; AVX2-NEXT: vmovdqu %ymm15, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8734; AVX2-NEXT: vpermq {{.*#+}} ymm13 = ymm13[0,1,2,1]8735; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm13[4,5,6,7]8736; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8737; AVX2-NEXT: vmovdqa 368(%rdi), %xmm08738; AVX2-NEXT: vpalignr {{.*#+}} xmm0 = xmm8[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]8739; AVX2-NEXT: vmovdqa 448(%rdi), %ymm138740; AVX2-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm8 # 32-byte Folded Reload8741; AVX2-NEXT: # ymm8 = mem[8,9,10,11,12,13,14,15],ymm13[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm13[16,17,18,19,20,21,22,23]8742; AVX2-NEXT: vmovdqu %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8743; AVX2-NEXT: vpermq {{.*#+}} ymm8 = ymm8[0,1,2,1]8744; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm8[4,5,6,7]8745; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8746; AVX2-NEXT: vmovdqa 688(%rdi), %xmm08747; AVX2-NEXT: vpalignr {{.*#+}} xmm0 = xmm7[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]8748; AVX2-NEXT: vmovdqa 768(%rdi), %ymm88749; AVX2-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm8, %ymm7 # 32-byte Folded Reload8750; AVX2-NEXT: # ymm7 = mem[8,9,10,11,12,13,14,15],ymm8[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm8[16,17,18,19,20,21,22,23]8751; AVX2-NEXT: vmovdqu %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8752; AVX2-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,1,2,1]8753; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm7[4,5,6,7]8754; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8755; AVX2-NEXT: vmovdqa 1008(%rdi), %xmm08756; AVX2-NEXT: vpalignr {{.*#+}} xmm0 = xmm5[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]8757; AVX2-NEXT: vmovdqa 1088(%rdi), %ymm78758; AVX2-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm5 # 32-byte Folded Reload8759; AVX2-NEXT: # ymm5 = mem[8,9,10,11,12,13,14,15],ymm7[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm7[16,17,18,19,20,21,22,23]8760; AVX2-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8761; AVX2-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,1,2,1]8762; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm5[4,5,6,7]8763; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8764; AVX2-NEXT: vmovdqa 1328(%rdi), %xmm08765; AVX2-NEXT: vpalignr {{.*#+}} xmm0 = xmm4[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]8766; AVX2-NEXT: vmovdqa 1408(%rdi), %ymm58767; AVX2-NEXT: vpalignr {{.*#+}} ymm4 = ymm12[8,9,10,11,12,13,14,15],ymm5[0,1,2,3,4,5,6,7],ymm12[24,25,26,27,28,29,30,31],ymm5[16,17,18,19,20,21,22,23]8768; AVX2-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8769; AVX2-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,1,2,1]8770; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm4[4,5,6,7]8771; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8772; AVX2-NEXT: vmovdqa 1648(%rdi), %xmm08773; AVX2-NEXT: vpalignr {{.*#+}} xmm0 = xmm3[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]8774; AVX2-NEXT: vmovdqa 1728(%rdi), %ymm48775; AVX2-NEXT: vpalignr {{.*#+}} ymm3 = ymm11[8,9,10,11,12,13,14,15],ymm4[0,1,2,3,4,5,6,7],ymm11[24,25,26,27,28,29,30,31],ymm4[16,17,18,19,20,21,22,23]8776; AVX2-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8777; AVX2-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,1,2,1]8778; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm3[4,5,6,7]8779; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8780; AVX2-NEXT: vmovdqa 1968(%rdi), %xmm08781; AVX2-NEXT: vpalignr {{.*#+}} xmm0 = xmm9[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]8782; AVX2-NEXT: vmovdqa 2048(%rdi), %ymm38783; AVX2-NEXT: vpalignr {{.*#+}} ymm2 = ymm10[8,9,10,11,12,13,14,15],ymm3[0,1,2,3,4,5,6,7],ymm10[24,25,26,27,28,29,30,31],ymm3[16,17,18,19,20,21,22,23]8784; AVX2-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8785; AVX2-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,1,2,1]8786; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]8787; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8788; AVX2-NEXT: vmovdqa 2288(%rdi), %xmm08789; AVX2-NEXT: vpalignr {{.*#+}} xmm0 = xmm1[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]8790; AVX2-NEXT: vmovdqa 2368(%rdi), %ymm148791; AVX2-NEXT: vpalignr {{.*#+}} ymm1 = ymm6[8,9,10,11,12,13,14,15],ymm14[0,1,2,3,4,5,6,7],ymm6[24,25,26,27,28,29,30,31],ymm14[16,17,18,19,20,21,22,23]8792; AVX2-NEXT: vmovdqu %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8793; AVX2-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,1,2,1]8794; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]8795; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8796; AVX2-NEXT: vmovaps 160(%rdi), %ymm08797; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]8798; AVX2-NEXT: vmovaps 256(%rdi), %xmm18799; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]8800; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm18801; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]8802; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8803; AVX2-NEXT: vmovaps 480(%rdi), %ymm08804; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]8805; AVX2-NEXT: vmovaps 576(%rdi), %xmm18806; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]8807; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm18808; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]8809; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8810; AVX2-NEXT: vmovaps 800(%rdi), %ymm08811; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]8812; AVX2-NEXT: vmovaps 896(%rdi), %xmm18813; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]8814; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm18815; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]8816; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8817; AVX2-NEXT: vmovaps 1120(%rdi), %ymm08818; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]8819; AVX2-NEXT: vmovaps 1216(%rdi), %xmm18820; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]8821; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm18822; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]8823; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8824; AVX2-NEXT: vmovaps 1440(%rdi), %ymm08825; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]8826; AVX2-NEXT: vmovaps 1536(%rdi), %xmm18827; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]8828; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm18829; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]8830; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8831; AVX2-NEXT: vmovaps 1760(%rdi), %ymm08832; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]8833; AVX2-NEXT: vmovaps 1856(%rdi), %xmm18834; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]8835; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm18836; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]8837; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8838; AVX2-NEXT: vmovaps 2080(%rdi), %ymm08839; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]8840; AVX2-NEXT: vmovaps 2176(%rdi), %xmm18841; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]8842; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm18843; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]8844; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8845; AVX2-NEXT: vmovaps 2400(%rdi), %ymm08846; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]8847; AVX2-NEXT: vmovaps 2496(%rdi), %xmm18848; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]8849; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm18850; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]8851; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8852; AVX2-NEXT: vmovaps 2240(%rdi), %ymm08853; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]8854; AVX2-NEXT: vmovaps 2336(%rdi), %xmm18855; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]8856; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm18857; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]8858; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8859; AVX2-NEXT: vmovaps 1920(%rdi), %ymm08860; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]8861; AVX2-NEXT: vmovaps 2016(%rdi), %xmm18862; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]8863; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm18864; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]8865; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8866; AVX2-NEXT: vmovaps 1600(%rdi), %ymm08867; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]8868; AVX2-NEXT: vmovaps 1696(%rdi), %xmm18869; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]8870; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm18871; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]8872; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8873; AVX2-NEXT: vmovaps 1280(%rdi), %ymm08874; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]8875; AVX2-NEXT: vmovaps 1376(%rdi), %xmm18876; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]8877; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm18878; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]8879; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8880; AVX2-NEXT: vmovaps 960(%rdi), %ymm08881; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]8882; AVX2-NEXT: vmovaps 1056(%rdi), %xmm18883; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]8884; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm18885; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]8886; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8887; AVX2-NEXT: vmovaps 640(%rdi), %ymm08888; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]8889; AVX2-NEXT: vmovaps 736(%rdi), %xmm18890; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]8891; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm18892; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]8893; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8894; AVX2-NEXT: vmovaps 320(%rdi), %ymm08895; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]8896; AVX2-NEXT: vmovaps 416(%rdi), %xmm18897; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]8898; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm18899; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]8900; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8901; AVX2-NEXT: vmovaps (%rdi), %ymm08902; AVX2-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]8903; AVX2-NEXT: vmovaps 96(%rdi), %xmm18904; AVX2-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]8905; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm18906; AVX2-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]8907; AVX2-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8908; AVX2-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm0 # 32-byte Folded Reload8909; AVX2-NEXT: # ymm0 = mem[8,9,10,11,12,13,14,15],ymm15[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm15[16,17,18,19,20,21,22,23]8910; AVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,3]8911; AVX2-NEXT: vmovdqa 64(%rdi), %xmm18912; AVX2-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8913; AVX2-NEXT: vpalignr {{.*#+}} xmm1 = mem[8,9,10,11,12,13,14,15],xmm1[0,1,2,3,4,5,6,7]8914; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]8915; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8916; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload8917; AVX2-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload8918; AVX2-NEXT: # ymm0 = mem[8,9,10,11,12,13,14,15],ymm0[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm0[16,17,18,19,20,21,22,23]8919; AVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,3]8920; AVX2-NEXT: vmovdqa 224(%rdi), %xmm18921; AVX2-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill8922; AVX2-NEXT: vpalignr {{.*#+}} xmm1 = mem[8,9,10,11,12,13,14,15],xmm1[0,1,2,3,4,5,6,7]8923; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]8924; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8925; AVX2-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm0 # 32-byte Folded Reload8926; AVX2-NEXT: # ymm0 = mem[8,9,10,11,12,13,14,15],ymm13[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm13[16,17,18,19,20,21,22,23]8927; AVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,3]8928; AVX2-NEXT: vmovdqa 384(%rdi), %xmm138929; AVX2-NEXT: vpalignr {{.*#+}} xmm1 = mem[8,9,10,11,12,13,14,15],xmm13[0,1,2,3,4,5,6,7]8930; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]8931; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8932; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload8933; AVX2-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload8934; AVX2-NEXT: # ymm0 = mem[8,9,10,11,12,13,14,15],ymm0[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm0[16,17,18,19,20,21,22,23]8935; AVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,3]8936; AVX2-NEXT: vmovdqa 544(%rdi), %xmm118937; AVX2-NEXT: vpalignr {{.*#+}} xmm1 = mem[8,9,10,11,12,13,14,15],xmm11[0,1,2,3,4,5,6,7]8938; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]8939; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8940; AVX2-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm8, %ymm0 # 32-byte Folded Reload8941; AVX2-NEXT: # ymm0 = mem[8,9,10,11,12,13,14,15],ymm8[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm8[16,17,18,19,20,21,22,23]8942; AVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,3]8943; AVX2-NEXT: vmovdqa 704(%rdi), %xmm128944; AVX2-NEXT: vpalignr {{.*#+}} xmm1 = mem[8,9,10,11,12,13,14,15],xmm12[0,1,2,3,4,5,6,7]8945; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]8946; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8947; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload8948; AVX2-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload8949; AVX2-NEXT: # ymm0 = mem[8,9,10,11,12,13,14,15],ymm0[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm0[16,17,18,19,20,21,22,23]8950; AVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,3]8951; AVX2-NEXT: vmovdqa 864(%rdi), %xmm108952; AVX2-NEXT: vpalignr {{.*#+}} xmm1 = mem[8,9,10,11,12,13,14,15],xmm10[0,1,2,3,4,5,6,7]8953; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]8954; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8955; AVX2-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm0 # 32-byte Folded Reload8956; AVX2-NEXT: # ymm0 = mem[8,9,10,11,12,13,14,15],ymm7[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm7[16,17,18,19,20,21,22,23]8957; AVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,3]8958; AVX2-NEXT: vmovdqa 1024(%rdi), %xmm98959; AVX2-NEXT: vpalignr {{.*#+}} xmm1 = mem[8,9,10,11,12,13,14,15],xmm9[0,1,2,3,4,5,6,7]8960; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]8961; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8962; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload8963; AVX2-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload8964; AVX2-NEXT: # ymm0 = mem[8,9,10,11,12,13,14,15],ymm0[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm0[16,17,18,19,20,21,22,23]8965; AVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,3]8966; AVX2-NEXT: vmovdqa 1184(%rdi), %xmm88967; AVX2-NEXT: vpalignr {{.*#+}} xmm1 = mem[8,9,10,11,12,13,14,15],xmm8[0,1,2,3,4,5,6,7]8968; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]8969; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8970; AVX2-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm0 # 32-byte Folded Reload8971; AVX2-NEXT: # ymm0 = mem[8,9,10,11,12,13,14,15],ymm5[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm5[16,17,18,19,20,21,22,23]8972; AVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,3]8973; AVX2-NEXT: vmovdqa 1344(%rdi), %xmm78974; AVX2-NEXT: vpalignr {{.*#+}} xmm1 = mem[8,9,10,11,12,13,14,15],xmm7[0,1,2,3,4,5,6,7]8975; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]8976; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8977; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload8978; AVX2-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload8979; AVX2-NEXT: # ymm0 = mem[8,9,10,11,12,13,14,15],ymm0[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm0[16,17,18,19,20,21,22,23]8980; AVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,3]8981; AVX2-NEXT: vmovdqa 1504(%rdi), %xmm68982; AVX2-NEXT: vpalignr {{.*#+}} xmm1 = mem[8,9,10,11,12,13,14,15],xmm6[0,1,2,3,4,5,6,7]8983; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]8984; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8985; AVX2-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4, %ymm0 # 32-byte Folded Reload8986; AVX2-NEXT: # ymm0 = mem[8,9,10,11,12,13,14,15],ymm4[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm4[16,17,18,19,20,21,22,23]8987; AVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,3]8988; AVX2-NEXT: vmovdqa 1664(%rdi), %xmm58989; AVX2-NEXT: vpalignr {{.*#+}} xmm1 = mem[8,9,10,11,12,13,14,15],xmm5[0,1,2,3,4,5,6,7]8990; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]8991; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill8992; AVX2-NEXT: vmovdqu (%rsp), %ymm0 # 32-byte Reload8993; AVX2-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload8994; AVX2-NEXT: # ymm0 = mem[8,9,10,11,12,13,14,15],ymm0[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm0[16,17,18,19,20,21,22,23]8995; AVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,3]8996; AVX2-NEXT: vmovdqa 1824(%rdi), %xmm48997; AVX2-NEXT: vpalignr {{.*#+}} xmm1 = mem[8,9,10,11,12,13,14,15],xmm4[0,1,2,3,4,5,6,7]8998; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]8999; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9000; AVX2-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3, %ymm0 # 32-byte Folded Reload9001; AVX2-NEXT: # ymm0 = mem[8,9,10,11,12,13,14,15],ymm3[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm3[16,17,18,19,20,21,22,23]9002; AVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,3]9003; AVX2-NEXT: vmovdqa 1984(%rdi), %xmm39004; AVX2-NEXT: vpalignr {{.*#+}} xmm1 = mem[8,9,10,11,12,13,14,15],xmm3[0,1,2,3,4,5,6,7]9005; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]9006; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9007; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload9008; AVX2-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload9009; AVX2-NEXT: # ymm0 = mem[8,9,10,11,12,13,14,15],ymm0[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm0[16,17,18,19,20,21,22,23]9010; AVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,3]9011; AVX2-NEXT: vmovdqa 2144(%rdi), %xmm29012; AVX2-NEXT: vpalignr {{.*#+}} xmm1 = mem[8,9,10,11,12,13,14,15],xmm2[0,1,2,3,4,5,6,7]9013; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]9014; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9015; AVX2-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm14, %ymm0 # 32-byte Folded Reload9016; AVX2-NEXT: # ymm0 = mem[8,9,10,11,12,13,14,15],ymm14[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm14[16,17,18,19,20,21,22,23]9017; AVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,3]9018; AVX2-NEXT: vmovdqa 2304(%rdi), %xmm19019; AVX2-NEXT: vpalignr {{.*#+}} xmm15 = mem[8,9,10,11,12,13,14,15],xmm1[0,1,2,3,4,5,6,7]9020; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]9021; AVX2-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9022; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload9023; AVX2-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload9024; AVX2-NEXT: # ymm0 = mem[8,9,10,11,12,13,14,15],ymm0[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm0[16,17,18,19,20,21,22,23]9025; AVX2-NEXT: vpermq {{.*#+}} ymm15 = ymm0[0,1,0,3]9026; AVX2-NEXT: vmovdqa 2464(%rdi), %xmm09027; AVX2-NEXT: vpalignr {{.*#+}} xmm14 = mem[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]9028; AVX2-NEXT: vpblendd {{.*#+}} ymm14 = ymm14[0,1,2,3],ymm15[4,5,6,7]9029; AVX2-NEXT: vmovdqu %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9030; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Reload9031; AVX2-NEXT: vblendps $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm14, %ymm14 # 32-byte Folded Reload9032; AVX2-NEXT: # ymm14 = mem[0,1,2,3,4,5],ymm14[6,7]9033; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload9034; AVX2-NEXT: vblendps $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload9035; AVX2-NEXT: # xmm15 = mem[0,1],xmm15[2,3]9036; AVX2-NEXT: vblendps {{.*#+}} ymm14 = ymm15[0,1,2,3],ymm14[4,5,6,7]9037; AVX2-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9038; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Reload9039; AVX2-NEXT: vblendps $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm14, %ymm14 # 32-byte Folded Reload9040; AVX2-NEXT: # ymm14 = mem[0,1,2,3,4,5],ymm14[6,7]9041; AVX2-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload9042; AVX2-NEXT: vblendps $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload9043; AVX2-NEXT: # xmm15 = mem[0,1],xmm15[2,3]9044; AVX2-NEXT: vblendps {{.*#+}} ymm15 = ymm15[0,1,2,3],ymm14[4,5,6,7]9045; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Reload9046; AVX2-NEXT: vpblendd $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm14, %ymm14 # 32-byte Folded Reload9047; AVX2-NEXT: # ymm14 = mem[0,1,2,3,4,5],ymm14[6,7]9048; AVX2-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm13, %xmm13 # 16-byte Folded Reload9049; AVX2-NEXT: # xmm13 = mem[0,1],xmm13[2,3]9050; AVX2-NEXT: vpblendd {{.*#+}} ymm13 = ymm13[0,1,2,3],ymm14[4,5,6,7]9051; AVX2-NEXT: vmovdqu %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9052; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload9053; AVX2-NEXT: vpblendd $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm13 # 32-byte Folded Reload9054; AVX2-NEXT: # ymm13 = mem[0,1,2,3,4,5],ymm13[6,7]9055; AVX2-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm11, %xmm11 # 16-byte Folded Reload9056; AVX2-NEXT: # xmm11 = mem[0,1],xmm11[2,3]9057; AVX2-NEXT: vpblendd {{.*#+}} ymm11 = ymm11[0,1,2,3],ymm13[4,5,6,7]9058; AVX2-NEXT: vmovdqu %ymm11, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9059; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload9060; AVX2-NEXT: vpblendd $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm11, %ymm11 # 32-byte Folded Reload9061; AVX2-NEXT: # ymm11 = mem[0,1,2,3,4,5],ymm11[6,7]9062; AVX2-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm12, %xmm12 # 16-byte Folded Reload9063; AVX2-NEXT: # xmm12 = mem[0,1],xmm12[2,3]9064; AVX2-NEXT: vpblendd {{.*#+}} ymm12 = ymm12[0,1,2,3],ymm11[4,5,6,7]9065; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload9066; AVX2-NEXT: vpblendd $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm11, %ymm11 # 32-byte Folded Reload9067; AVX2-NEXT: # ymm11 = mem[0,1,2,3,4,5],ymm11[6,7]9068; AVX2-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm10, %xmm10 # 16-byte Folded Reload9069; AVX2-NEXT: # xmm10 = mem[0,1],xmm10[2,3]9070; AVX2-NEXT: vpblendd {{.*#+}} ymm11 = ymm10[0,1,2,3],ymm11[4,5,6,7]9071; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload9072; AVX2-NEXT: vpblendd $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm10, %ymm10 # 32-byte Folded Reload9073; AVX2-NEXT: # ymm10 = mem[0,1,2,3,4,5],ymm10[6,7]9074; AVX2-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm9, %xmm9 # 16-byte Folded Reload9075; AVX2-NEXT: # xmm9 = mem[0,1],xmm9[2,3]9076; AVX2-NEXT: vpblendd {{.*#+}} ymm9 = ymm9[0,1,2,3],ymm10[4,5,6,7]9077; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload9078; AVX2-NEXT: vpblendd $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm10, %ymm10 # 32-byte Folded Reload9079; AVX2-NEXT: # ymm10 = mem[0,1,2,3,4,5],ymm10[6,7]9080; AVX2-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm8, %xmm8 # 16-byte Folded Reload9081; AVX2-NEXT: # xmm8 = mem[0,1],xmm8[2,3]9082; AVX2-NEXT: vpblendd {{.*#+}} ymm14 = ymm8[0,1,2,3],ymm10[4,5,6,7]9083; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload9084; AVX2-NEXT: vpblendd $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm8, %ymm8 # 32-byte Folded Reload9085; AVX2-NEXT: # ymm8 = mem[0,1,2,3,4,5],ymm8[6,7]9086; AVX2-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm7, %xmm7 # 16-byte Folded Reload9087; AVX2-NEXT: # xmm7 = mem[0,1],xmm7[2,3]9088; AVX2-NEXT: vpblendd {{.*#+}} ymm10 = ymm7[0,1,2,3],ymm8[4,5,6,7]9089; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload9090; AVX2-NEXT: vpblendd $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm7 # 32-byte Folded Reload9091; AVX2-NEXT: # ymm7 = mem[0,1,2,3,4,5],ymm7[6,7]9092; AVX2-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm6, %xmm6 # 16-byte Folded Reload9093; AVX2-NEXT: # xmm6 = mem[0,1],xmm6[2,3]9094; AVX2-NEXT: vpblendd {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7]9095; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload9096; AVX2-NEXT: vpblendd $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm7 # 32-byte Folded Reload9097; AVX2-NEXT: # ymm7 = mem[0,1,2,3,4,5],ymm7[6,7]9098; AVX2-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm5, %xmm5 # 16-byte Folded Reload9099; AVX2-NEXT: # xmm5 = mem[0,1],xmm5[2,3]9100; AVX2-NEXT: vpblendd {{.*#+}} ymm13 = ymm5[0,1,2,3],ymm7[4,5,6,7]9101; AVX2-NEXT: vmovdqu (%rsp), %ymm5 # 32-byte Reload9102; AVX2-NEXT: vpblendd $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm5 # 32-byte Folded Reload9103; AVX2-NEXT: # ymm5 = mem[0,1,2,3,4,5],ymm5[6,7]9104; AVX2-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4, %xmm4 # 16-byte Folded Reload9105; AVX2-NEXT: # xmm4 = mem[0,1],xmm4[2,3]9106; AVX2-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm5[4,5,6,7]9107; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload9108; AVX2-NEXT: vpblendd $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm5 # 32-byte Folded Reload9109; AVX2-NEXT: # ymm5 = mem[0,1,2,3,4,5],ymm5[6,7]9110; AVX2-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm3 # 16-byte Folded Reload9111; AVX2-NEXT: # xmm3 = mem[0,1],xmm3[2,3]9112; AVX2-NEXT: vpblendd {{.*#+}} ymm7 = ymm3[0,1,2,3],ymm5[4,5,6,7]9113; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload9114; AVX2-NEXT: vpblendd $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3, %ymm3 # 32-byte Folded Reload9115; AVX2-NEXT: # ymm3 = mem[0,1,2,3,4,5],ymm3[6,7]9116; AVX2-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm2 # 16-byte Folded Reload9117; AVX2-NEXT: # xmm2 = mem[0,1],xmm2[2,3]9118; AVX2-NEXT: vpblendd {{.*#+}} ymm5 = ymm2[0,1,2,3],ymm3[4,5,6,7]9119; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload9120; AVX2-NEXT: vpblendd $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload9121; AVX2-NEXT: # ymm2 = mem[0,1,2,3,4,5],ymm2[6,7]9122; AVX2-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload9123; AVX2-NEXT: # xmm1 = mem[0,1],xmm1[2,3]9124; AVX2-NEXT: vpblendd {{.*#+}} ymm2 = ymm1[0,1,2,3],ymm2[4,5,6,7]9125; AVX2-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9126; AVX2-NEXT: vpblendd $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload9127; AVX2-NEXT: # ymm1 = mem[0,1,2,3,4,5],ymm1[6,7]9128; AVX2-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload9129; AVX2-NEXT: # xmm0 = mem[0,1],xmm0[2,3]9130; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]9131; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9132; AVX2-NEXT: vmovaps %ymm1, 448(%rsi)9133; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9134; AVX2-NEXT: vmovaps %ymm1, 384(%rsi)9135; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9136; AVX2-NEXT: vmovaps %ymm1, 320(%rsi)9137; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9138; AVX2-NEXT: vmovaps %ymm1, 256(%rsi)9139; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9140; AVX2-NEXT: vmovaps %ymm1, 192(%rsi)9141; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9142; AVX2-NEXT: vmovaps %ymm1, 128(%rsi)9143; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9144; AVX2-NEXT: vmovaps %ymm1, 64(%rsi)9145; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9146; AVX2-NEXT: vmovaps %ymm1, (%rsi)9147; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9148; AVX2-NEXT: vmovaps %ymm1, 480(%rsi)9149; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9150; AVX2-NEXT: vmovaps %ymm1, 416(%rsi)9151; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9152; AVX2-NEXT: vmovaps %ymm1, 352(%rsi)9153; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9154; AVX2-NEXT: vmovaps %ymm1, 288(%rsi)9155; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9156; AVX2-NEXT: vmovaps %ymm1, 224(%rsi)9157; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9158; AVX2-NEXT: vmovaps %ymm1, 160(%rsi)9159; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9160; AVX2-NEXT: vmovaps %ymm1, 96(%rsi)9161; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9162; AVX2-NEXT: vmovaps %ymm1, 32(%rsi)9163; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9164; AVX2-NEXT: vmovaps %ymm1, 448(%rdx)9165; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9166; AVX2-NEXT: vmovaps %ymm1, 384(%rdx)9167; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9168; AVX2-NEXT: vmovaps %ymm1, 320(%rdx)9169; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9170; AVX2-NEXT: vmovaps %ymm1, 256(%rdx)9171; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9172; AVX2-NEXT: vmovaps %ymm1, 192(%rdx)9173; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9174; AVX2-NEXT: vmovaps %ymm1, 128(%rdx)9175; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9176; AVX2-NEXT: vmovaps %ymm1, 64(%rdx)9177; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9178; AVX2-NEXT: vmovaps %ymm1, (%rdx)9179; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9180; AVX2-NEXT: vmovaps %ymm1, 480(%rdx)9181; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9182; AVX2-NEXT: vmovaps %ymm1, 416(%rdx)9183; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9184; AVX2-NEXT: vmovaps %ymm1, 352(%rdx)9185; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9186; AVX2-NEXT: vmovaps %ymm1, 288(%rdx)9187; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9188; AVX2-NEXT: vmovaps %ymm1, 224(%rdx)9189; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9190; AVX2-NEXT: vmovaps %ymm1, 160(%rdx)9191; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9192; AVX2-NEXT: vmovaps %ymm1, 96(%rdx)9193; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9194; AVX2-NEXT: vmovaps %ymm1, 32(%rdx)9195; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9196; AVX2-NEXT: vmovaps %ymm1, (%rcx)9197; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9198; AVX2-NEXT: vmovaps %ymm1, 64(%rcx)9199; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9200; AVX2-NEXT: vmovaps %ymm1, 128(%rcx)9201; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9202; AVX2-NEXT: vmovaps %ymm1, 192(%rcx)9203; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9204; AVX2-NEXT: vmovaps %ymm1, 256(%rcx)9205; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9206; AVX2-NEXT: vmovaps %ymm1, 320(%rcx)9207; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9208; AVX2-NEXT: vmovaps %ymm1, 384(%rcx)9209; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9210; AVX2-NEXT: vmovaps %ymm1, 448(%rcx)9211; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9212; AVX2-NEXT: vmovaps %ymm1, 480(%rcx)9213; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9214; AVX2-NEXT: vmovaps %ymm1, 416(%rcx)9215; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9216; AVX2-NEXT: vmovaps %ymm1, 352(%rcx)9217; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9218; AVX2-NEXT: vmovaps %ymm1, 288(%rcx)9219; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9220; AVX2-NEXT: vmovaps %ymm1, 224(%rcx)9221; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9222; AVX2-NEXT: vmovaps %ymm1, 160(%rcx)9223; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9224; AVX2-NEXT: vmovaps %ymm1, 96(%rcx)9225; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9226; AVX2-NEXT: vmovaps %ymm1, 32(%rcx)9227; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9228; AVX2-NEXT: vmovaps %ymm1, 480(%r8)9229; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9230; AVX2-NEXT: vmovaps %ymm1, 448(%r8)9231; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9232; AVX2-NEXT: vmovaps %ymm1, 416(%r8)9233; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9234; AVX2-NEXT: vmovaps %ymm1, 384(%r8)9235; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9236; AVX2-NEXT: vmovaps %ymm1, 352(%r8)9237; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9238; AVX2-NEXT: vmovaps %ymm1, 320(%r8)9239; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9240; AVX2-NEXT: vmovaps %ymm1, 288(%r8)9241; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9242; AVX2-NEXT: vmovaps %ymm1, 256(%r8)9243; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9244; AVX2-NEXT: vmovaps %ymm1, 224(%r8)9245; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9246; AVX2-NEXT: vmovaps %ymm1, 192(%r8)9247; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9248; AVX2-NEXT: vmovaps %ymm1, 160(%r8)9249; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9250; AVX2-NEXT: vmovaps %ymm1, 128(%r8)9251; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9252; AVX2-NEXT: vmovaps %ymm1, 96(%r8)9253; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9254; AVX2-NEXT: vmovaps %ymm1, 64(%r8)9255; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9256; AVX2-NEXT: vmovaps %ymm1, 32(%r8)9257; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9258; AVX2-NEXT: vmovaps %ymm1, (%r8)9259; AVX2-NEXT: vmovdqa %ymm0, 480(%r9)9260; AVX2-NEXT: vmovdqa %ymm2, 448(%r9)9261; AVX2-NEXT: vmovdqa %ymm5, 416(%r9)9262; AVX2-NEXT: vmovdqa %ymm7, 384(%r9)9263; AVX2-NEXT: vmovdqa %ymm4, 352(%r9)9264; AVX2-NEXT: vmovdqa %ymm13, 320(%r9)9265; AVX2-NEXT: vmovdqa %ymm6, 288(%r9)9266; AVX2-NEXT: vmovdqa %ymm10, 256(%r9)9267; AVX2-NEXT: vmovdqa %ymm14, 224(%r9)9268; AVX2-NEXT: vmovdqa %ymm9, 192(%r9)9269; AVX2-NEXT: vmovdqa %ymm11, 160(%r9)9270; AVX2-NEXT: vmovdqa %ymm12, 128(%r9)9271; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload9272; AVX2-NEXT: vmovaps %ymm0, 96(%r9)9273; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload9274; AVX2-NEXT: vmovaps %ymm0, 64(%r9)9275; AVX2-NEXT: vmovaps %ymm15, 32(%r9)9276; AVX2-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload9277; AVX2-NEXT: vmovaps %ymm0, (%r9)9278; AVX2-NEXT: addq $3240, %rsp # imm = 0xCA89279; AVX2-NEXT: vzeroupper9280; AVX2-NEXT: retq9281;9282; AVX2-FP-LABEL: load_i64_stride5_vf64:9283; AVX2-FP: # %bb.0:9284; AVX2-FP-NEXT: subq $3240, %rsp # imm = 0xCA89285; AVX2-FP-NEXT: vmovaps 896(%rdi), %ymm29286; AVX2-FP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9287; AVX2-FP-NEXT: vmovaps 864(%rdi), %ymm49288; AVX2-FP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9289; AVX2-FP-NEXT: vmovaps 576(%rdi), %ymm39290; AVX2-FP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9291; AVX2-FP-NEXT: vmovaps 544(%rdi), %ymm59292; AVX2-FP-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9293; AVX2-FP-NEXT: vmovaps 256(%rdi), %ymm09294; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9295; AVX2-FP-NEXT: vmovaps 224(%rdi), %ymm19296; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9297; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]9298; AVX2-FP-NEXT: vmovaps 192(%rdi), %xmm19299; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9300; AVX2-FP-NEXT: vmovaps 160(%rdi), %xmm69301; AVX2-FP-NEXT: vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9302; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm6[0,1],xmm1[2,3]9303; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]9304; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9305; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm5[0,1,2,3,4,5],ymm3[6,7]9306; AVX2-FP-NEXT: vmovaps 512(%rdi), %xmm19307; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9308; AVX2-FP-NEXT: vmovaps 480(%rdi), %xmm39309; AVX2-FP-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9310; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm3[0,1],xmm1[2,3]9311; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]9312; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9313; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm4[0,1,2,3,4,5],ymm2[6,7]9314; AVX2-FP-NEXT: vmovaps 832(%rdi), %xmm19315; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9316; AVX2-FP-NEXT: vmovaps 800(%rdi), %xmm29317; AVX2-FP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9318; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3]9319; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]9320; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9321; AVX2-FP-NEXT: vmovaps 1216(%rdi), %ymm09322; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9323; AVX2-FP-NEXT: vmovaps 1184(%rdi), %ymm19324; AVX2-FP-NEXT: vmovups %ymm1, (%rsp) # 32-byte Spill9325; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]9326; AVX2-FP-NEXT: vmovaps 1152(%rdi), %xmm19327; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9328; AVX2-FP-NEXT: vmovaps 1120(%rdi), %xmm29329; AVX2-FP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9330; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3]9331; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]9332; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9333; AVX2-FP-NEXT: vmovaps 1536(%rdi), %ymm09334; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9335; AVX2-FP-NEXT: vmovaps 1504(%rdi), %ymm19336; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9337; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]9338; AVX2-FP-NEXT: vmovaps 1472(%rdi), %xmm19339; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9340; AVX2-FP-NEXT: vmovaps 1440(%rdi), %xmm29341; AVX2-FP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9342; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3]9343; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]9344; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9345; AVX2-FP-NEXT: vmovaps 1856(%rdi), %ymm09346; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9347; AVX2-FP-NEXT: vmovaps 1824(%rdi), %ymm19348; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9349; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]9350; AVX2-FP-NEXT: vmovaps 1792(%rdi), %xmm19351; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9352; AVX2-FP-NEXT: vmovaps 1760(%rdi), %xmm29353; AVX2-FP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9354; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3]9355; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]9356; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9357; AVX2-FP-NEXT: vmovaps 2176(%rdi), %ymm09358; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9359; AVX2-FP-NEXT: vmovaps 2144(%rdi), %ymm19360; AVX2-FP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9361; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]9362; AVX2-FP-NEXT: vmovaps 2112(%rdi), %xmm19363; AVX2-FP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9364; AVX2-FP-NEXT: vmovaps 2080(%rdi), %xmm29365; AVX2-FP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9366; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3]9367; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]9368; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9369; AVX2-FP-NEXT: vmovdqa 2496(%rdi), %ymm09370; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9371; AVX2-FP-NEXT: vmovdqa 2464(%rdi), %ymm19372; AVX2-FP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9373; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]9374; AVX2-FP-NEXT: vmovdqa 2432(%rdi), %xmm19375; AVX2-FP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9376; AVX2-FP-NEXT: vmovdqa 2400(%rdi), %xmm139377; AVX2-FP-NEXT: vpblendd {{.*#+}} xmm1 = xmm13[0,1],xmm1[2,3]9378; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]9379; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9380; AVX2-FP-NEXT: vmovdqa 96(%rdi), %ymm09381; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9382; AVX2-FP-NEXT: vmovdqa 64(%rdi), %ymm19383; AVX2-FP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9384; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]9385; AVX2-FP-NEXT: vmovdqa (%rdi), %xmm149386; AVX2-FP-NEXT: vmovdqa 32(%rdi), %xmm19387; AVX2-FP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9388; AVX2-FP-NEXT: vpblendd {{.*#+}} xmm1 = xmm14[0,1],xmm1[2,3]9389; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]9390; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9391; AVX2-FP-NEXT: vmovdqa 416(%rdi), %ymm09392; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9393; AVX2-FP-NEXT: vmovdqa 384(%rdi), %ymm19394; AVX2-FP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9395; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]9396; AVX2-FP-NEXT: vmovdqa 352(%rdi), %xmm19397; AVX2-FP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9398; AVX2-FP-NEXT: vmovdqa 320(%rdi), %xmm89399; AVX2-FP-NEXT: vpblendd {{.*#+}} xmm1 = xmm8[0,1],xmm1[2,3]9400; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]9401; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9402; AVX2-FP-NEXT: vmovdqa 736(%rdi), %ymm09403; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9404; AVX2-FP-NEXT: vmovdqa 704(%rdi), %ymm19405; AVX2-FP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9406; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]9407; AVX2-FP-NEXT: vmovdqa 672(%rdi), %xmm19408; AVX2-FP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9409; AVX2-FP-NEXT: vmovdqa 640(%rdi), %xmm79410; AVX2-FP-NEXT: vpblendd {{.*#+}} xmm1 = xmm7[0,1],xmm1[2,3]9411; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]9412; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9413; AVX2-FP-NEXT: vmovdqa 1056(%rdi), %ymm09414; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9415; AVX2-FP-NEXT: vmovdqa 1024(%rdi), %ymm19416; AVX2-FP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9417; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]9418; AVX2-FP-NEXT: vmovdqa 992(%rdi), %xmm19419; AVX2-FP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9420; AVX2-FP-NEXT: vmovdqa 960(%rdi), %xmm59421; AVX2-FP-NEXT: vpblendd {{.*#+}} xmm1 = xmm5[0,1],xmm1[2,3]9422; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]9423; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9424; AVX2-FP-NEXT: vmovdqa 1376(%rdi), %ymm09425; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9426; AVX2-FP-NEXT: vmovdqa 1344(%rdi), %ymm129427; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm12[0,1,2,3,4,5],ymm0[6,7]9428; AVX2-FP-NEXT: vmovdqa 1312(%rdi), %xmm19429; AVX2-FP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9430; AVX2-FP-NEXT: vmovdqa 1280(%rdi), %xmm49431; AVX2-FP-NEXT: vpblendd {{.*#+}} xmm1 = xmm4[0,1],xmm1[2,3]9432; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]9433; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9434; AVX2-FP-NEXT: vmovdqa 1696(%rdi), %ymm09435; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9436; AVX2-FP-NEXT: vmovdqa 1664(%rdi), %ymm119437; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm11[0,1,2,3,4,5],ymm0[6,7]9438; AVX2-FP-NEXT: vmovdqa 1632(%rdi), %xmm19439; AVX2-FP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9440; AVX2-FP-NEXT: vmovdqa 1600(%rdi), %xmm39441; AVX2-FP-NEXT: vpblendd {{.*#+}} xmm1 = xmm3[0,1],xmm1[2,3]9442; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]9443; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9444; AVX2-FP-NEXT: vmovdqa 2016(%rdi), %ymm09445; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9446; AVX2-FP-NEXT: vmovdqa 1984(%rdi), %ymm109447; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm10[0,1,2,3,4,5],ymm0[6,7]9448; AVX2-FP-NEXT: vmovdqa 1952(%rdi), %xmm19449; AVX2-FP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9450; AVX2-FP-NEXT: vmovdqa 1920(%rdi), %xmm99451; AVX2-FP-NEXT: vpblendd {{.*#+}} xmm6 = xmm9[0,1],xmm1[2,3]9452; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm6[0,1,2,3],ymm0[4,5,6,7]9453; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9454; AVX2-FP-NEXT: vmovdqa 2336(%rdi), %ymm09455; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9456; AVX2-FP-NEXT: vmovdqa 2304(%rdi), %ymm69457; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm2 = ymm6[0,1,2,3,4,5],ymm0[6,7]9458; AVX2-FP-NEXT: vmovdqa 2272(%rdi), %xmm09459; AVX2-FP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9460; AVX2-FP-NEXT: vmovdqa 2240(%rdi), %xmm19461; AVX2-FP-NEXT: vpblendd {{.*#+}} xmm15 = xmm1[0,1],xmm0[2,3]9462; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm2[4,5,6,7]9463; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9464; AVX2-FP-NEXT: vmovdqa 208(%rdi), %xmm09465; AVX2-FP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload9466; AVX2-FP-NEXT: # xmm0 = mem[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]9467; AVX2-FP-NEXT: vmovdqa 288(%rdi), %ymm29468; AVX2-FP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9469; AVX2-FP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm15 # 32-byte Folded Reload9470; AVX2-FP-NEXT: # ymm15 = mem[8,9,10,11,12,13,14,15],ymm2[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm2[16,17,18,19,20,21,22,23]9471; AVX2-FP-NEXT: vpermq {{.*#+}} ymm15 = ymm15[0,1,2,1]9472; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm15[4,5,6,7]9473; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9474; AVX2-FP-NEXT: vmovdqa 528(%rdi), %xmm09475; AVX2-FP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload9476; AVX2-FP-NEXT: # xmm0 = mem[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]9477; AVX2-FP-NEXT: vmovdqa 608(%rdi), %ymm29478; AVX2-FP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9479; AVX2-FP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm15 # 32-byte Folded Reload9480; AVX2-FP-NEXT: # ymm15 = mem[8,9,10,11,12,13,14,15],ymm2[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm2[16,17,18,19,20,21,22,23]9481; AVX2-FP-NEXT: vpermq {{.*#+}} ymm15 = ymm15[0,1,2,1]9482; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm15[4,5,6,7]9483; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9484; AVX2-FP-NEXT: vmovdqa 848(%rdi), %xmm09485; AVX2-FP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload9486; AVX2-FP-NEXT: # xmm0 = mem[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]9487; AVX2-FP-NEXT: vmovdqa 928(%rdi), %ymm29488; AVX2-FP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9489; AVX2-FP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm15 # 32-byte Folded Reload9490; AVX2-FP-NEXT: # ymm15 = mem[8,9,10,11,12,13,14,15],ymm2[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm2[16,17,18,19,20,21,22,23]9491; AVX2-FP-NEXT: vpermq {{.*#+}} ymm15 = ymm15[0,1,2,1]9492; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm15[4,5,6,7]9493; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9494; AVX2-FP-NEXT: vmovdqa 1168(%rdi), %xmm09495; AVX2-FP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload9496; AVX2-FP-NEXT: # xmm0 = mem[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]9497; AVX2-FP-NEXT: vmovdqa 1248(%rdi), %ymm29498; AVX2-FP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9499; AVX2-FP-NEXT: vpalignr $8, (%rsp), %ymm2, %ymm15 # 32-byte Folded Reload9500; AVX2-FP-NEXT: # ymm15 = mem[8,9,10,11,12,13,14,15],ymm2[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm2[16,17,18,19,20,21,22,23]9501; AVX2-FP-NEXT: vpermq {{.*#+}} ymm15 = ymm15[0,1,2,1]9502; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm15[4,5,6,7]9503; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9504; AVX2-FP-NEXT: vmovdqa 1488(%rdi), %xmm09505; AVX2-FP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload9506; AVX2-FP-NEXT: # xmm0 = mem[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]9507; AVX2-FP-NEXT: vmovdqa 1568(%rdi), %ymm29508; AVX2-FP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9509; AVX2-FP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm15 # 32-byte Folded Reload9510; AVX2-FP-NEXT: # ymm15 = mem[8,9,10,11,12,13,14,15],ymm2[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm2[16,17,18,19,20,21,22,23]9511; AVX2-FP-NEXT: vpermq {{.*#+}} ymm15 = ymm15[0,1,2,1]9512; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm15[4,5,6,7]9513; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9514; AVX2-FP-NEXT: vmovdqa 1808(%rdi), %xmm09515; AVX2-FP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload9516; AVX2-FP-NEXT: # xmm0 = mem[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]9517; AVX2-FP-NEXT: vmovdqa 1888(%rdi), %ymm29518; AVX2-FP-NEXT: vmovdqu %ymm2, (%rsp) # 32-byte Spill9519; AVX2-FP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm15 # 32-byte Folded Reload9520; AVX2-FP-NEXT: # ymm15 = mem[8,9,10,11,12,13,14,15],ymm2[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm2[16,17,18,19,20,21,22,23]9521; AVX2-FP-NEXT: vpermq {{.*#+}} ymm15 = ymm15[0,1,2,1]9522; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm15[4,5,6,7]9523; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9524; AVX2-FP-NEXT: vmovdqa 2128(%rdi), %xmm09525; AVX2-FP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload9526; AVX2-FP-NEXT: # xmm0 = mem[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]9527; AVX2-FP-NEXT: vmovdqa 2208(%rdi), %ymm29528; AVX2-FP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9529; AVX2-FP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm15 # 32-byte Folded Reload9530; AVX2-FP-NEXT: # ymm15 = mem[8,9,10,11,12,13,14,15],ymm2[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm2[16,17,18,19,20,21,22,23]9531; AVX2-FP-NEXT: vpermq {{.*#+}} ymm15 = ymm15[0,1,2,1]9532; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm15[4,5,6,7]9533; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9534; AVX2-FP-NEXT: vmovdqa 2448(%rdi), %xmm09535; AVX2-FP-NEXT: vpalignr {{.*#+}} xmm0 = xmm13[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]9536; AVX2-FP-NEXT: vmovdqa 2528(%rdi), %ymm29537; AVX2-FP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9538; AVX2-FP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm13 # 32-byte Folded Reload9539; AVX2-FP-NEXT: # ymm13 = mem[8,9,10,11,12,13,14,15],ymm2[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm2[16,17,18,19,20,21,22,23]9540; AVX2-FP-NEXT: vpermq {{.*#+}} ymm13 = ymm13[0,1,2,1]9541; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm13[4,5,6,7]9542; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9543; AVX2-FP-NEXT: vmovdqa 48(%rdi), %xmm09544; AVX2-FP-NEXT: vpalignr {{.*#+}} xmm0 = xmm14[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]9545; AVX2-FP-NEXT: vmovdqa 128(%rdi), %ymm159546; AVX2-FP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm13 # 32-byte Folded Reload9547; AVX2-FP-NEXT: # ymm13 = mem[8,9,10,11,12,13,14,15],ymm15[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm15[16,17,18,19,20,21,22,23]9548; AVX2-FP-NEXT: vmovdqu %ymm15, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9549; AVX2-FP-NEXT: vpermq {{.*#+}} ymm13 = ymm13[0,1,2,1]9550; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm13[4,5,6,7]9551; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9552; AVX2-FP-NEXT: vmovdqa 368(%rdi), %xmm09553; AVX2-FP-NEXT: vpalignr {{.*#+}} xmm0 = xmm8[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]9554; AVX2-FP-NEXT: vmovdqa 448(%rdi), %ymm139555; AVX2-FP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm8 # 32-byte Folded Reload9556; AVX2-FP-NEXT: # ymm8 = mem[8,9,10,11,12,13,14,15],ymm13[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm13[16,17,18,19,20,21,22,23]9557; AVX2-FP-NEXT: vmovdqu %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9558; AVX2-FP-NEXT: vpermq {{.*#+}} ymm8 = ymm8[0,1,2,1]9559; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm8[4,5,6,7]9560; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9561; AVX2-FP-NEXT: vmovdqa 688(%rdi), %xmm09562; AVX2-FP-NEXT: vpalignr {{.*#+}} xmm0 = xmm7[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]9563; AVX2-FP-NEXT: vmovdqa 768(%rdi), %ymm89564; AVX2-FP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm8, %ymm7 # 32-byte Folded Reload9565; AVX2-FP-NEXT: # ymm7 = mem[8,9,10,11,12,13,14,15],ymm8[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm8[16,17,18,19,20,21,22,23]9566; AVX2-FP-NEXT: vmovdqu %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9567; AVX2-FP-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,1,2,1]9568; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm7[4,5,6,7]9569; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9570; AVX2-FP-NEXT: vmovdqa 1008(%rdi), %xmm09571; AVX2-FP-NEXT: vpalignr {{.*#+}} xmm0 = xmm5[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]9572; AVX2-FP-NEXT: vmovdqa 1088(%rdi), %ymm79573; AVX2-FP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm5 # 32-byte Folded Reload9574; AVX2-FP-NEXT: # ymm5 = mem[8,9,10,11,12,13,14,15],ymm7[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm7[16,17,18,19,20,21,22,23]9575; AVX2-FP-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9576; AVX2-FP-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,1,2,1]9577; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm5[4,5,6,7]9578; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9579; AVX2-FP-NEXT: vmovdqa 1328(%rdi), %xmm09580; AVX2-FP-NEXT: vpalignr {{.*#+}} xmm0 = xmm4[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]9581; AVX2-FP-NEXT: vmovdqa 1408(%rdi), %ymm59582; AVX2-FP-NEXT: vpalignr {{.*#+}} ymm4 = ymm12[8,9,10,11,12,13,14,15],ymm5[0,1,2,3,4,5,6,7],ymm12[24,25,26,27,28,29,30,31],ymm5[16,17,18,19,20,21,22,23]9583; AVX2-FP-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9584; AVX2-FP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,1,2,1]9585; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm4[4,5,6,7]9586; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9587; AVX2-FP-NEXT: vmovdqa 1648(%rdi), %xmm09588; AVX2-FP-NEXT: vpalignr {{.*#+}} xmm0 = xmm3[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]9589; AVX2-FP-NEXT: vmovdqa 1728(%rdi), %ymm49590; AVX2-FP-NEXT: vpalignr {{.*#+}} ymm3 = ymm11[8,9,10,11,12,13,14,15],ymm4[0,1,2,3,4,5,6,7],ymm11[24,25,26,27,28,29,30,31],ymm4[16,17,18,19,20,21,22,23]9591; AVX2-FP-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9592; AVX2-FP-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,1,2,1]9593; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm3[4,5,6,7]9594; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9595; AVX2-FP-NEXT: vmovdqa 1968(%rdi), %xmm09596; AVX2-FP-NEXT: vpalignr {{.*#+}} xmm0 = xmm9[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]9597; AVX2-FP-NEXT: vmovdqa 2048(%rdi), %ymm39598; AVX2-FP-NEXT: vpalignr {{.*#+}} ymm2 = ymm10[8,9,10,11,12,13,14,15],ymm3[0,1,2,3,4,5,6,7],ymm10[24,25,26,27,28,29,30,31],ymm3[16,17,18,19,20,21,22,23]9599; AVX2-FP-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9600; AVX2-FP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,1,2,1]9601; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]9602; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9603; AVX2-FP-NEXT: vmovdqa 2288(%rdi), %xmm09604; AVX2-FP-NEXT: vpalignr {{.*#+}} xmm0 = xmm1[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]9605; AVX2-FP-NEXT: vmovdqa 2368(%rdi), %ymm149606; AVX2-FP-NEXT: vpalignr {{.*#+}} ymm1 = ymm6[8,9,10,11,12,13,14,15],ymm14[0,1,2,3,4,5,6,7],ymm6[24,25,26,27,28,29,30,31],ymm14[16,17,18,19,20,21,22,23]9607; AVX2-FP-NEXT: vmovdqu %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9608; AVX2-FP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,1,2,1]9609; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]9610; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9611; AVX2-FP-NEXT: vmovaps 160(%rdi), %ymm09612; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]9613; AVX2-FP-NEXT: vmovaps 256(%rdi), %xmm19614; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]9615; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm19616; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]9617; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9618; AVX2-FP-NEXT: vmovaps 480(%rdi), %ymm09619; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]9620; AVX2-FP-NEXT: vmovaps 576(%rdi), %xmm19621; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]9622; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm19623; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]9624; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9625; AVX2-FP-NEXT: vmovaps 800(%rdi), %ymm09626; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]9627; AVX2-FP-NEXT: vmovaps 896(%rdi), %xmm19628; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]9629; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm19630; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]9631; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9632; AVX2-FP-NEXT: vmovaps 1120(%rdi), %ymm09633; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]9634; AVX2-FP-NEXT: vmovaps 1216(%rdi), %xmm19635; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]9636; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm19637; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]9638; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9639; AVX2-FP-NEXT: vmovaps 1440(%rdi), %ymm09640; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]9641; AVX2-FP-NEXT: vmovaps 1536(%rdi), %xmm19642; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]9643; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm19644; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]9645; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9646; AVX2-FP-NEXT: vmovaps 1760(%rdi), %ymm09647; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]9648; AVX2-FP-NEXT: vmovaps 1856(%rdi), %xmm19649; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]9650; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm19651; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]9652; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9653; AVX2-FP-NEXT: vmovaps 2080(%rdi), %ymm09654; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]9655; AVX2-FP-NEXT: vmovaps 2176(%rdi), %xmm19656; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]9657; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm19658; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]9659; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9660; AVX2-FP-NEXT: vmovaps 2400(%rdi), %ymm09661; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]9662; AVX2-FP-NEXT: vmovaps 2496(%rdi), %xmm19663; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]9664; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm19665; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]9666; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9667; AVX2-FP-NEXT: vmovaps 2240(%rdi), %ymm09668; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]9669; AVX2-FP-NEXT: vmovaps 2336(%rdi), %xmm19670; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]9671; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm19672; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]9673; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9674; AVX2-FP-NEXT: vmovaps 1920(%rdi), %ymm09675; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]9676; AVX2-FP-NEXT: vmovaps 2016(%rdi), %xmm19677; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]9678; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm19679; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]9680; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9681; AVX2-FP-NEXT: vmovaps 1600(%rdi), %ymm09682; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]9683; AVX2-FP-NEXT: vmovaps 1696(%rdi), %xmm19684; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]9685; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm19686; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]9687; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9688; AVX2-FP-NEXT: vmovaps 1280(%rdi), %ymm09689; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]9690; AVX2-FP-NEXT: vmovaps 1376(%rdi), %xmm19691; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]9692; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm19693; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]9694; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9695; AVX2-FP-NEXT: vmovaps 960(%rdi), %ymm09696; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]9697; AVX2-FP-NEXT: vmovaps 1056(%rdi), %xmm19698; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]9699; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm19700; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]9701; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9702; AVX2-FP-NEXT: vmovaps 640(%rdi), %ymm09703; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]9704; AVX2-FP-NEXT: vmovaps 736(%rdi), %xmm19705; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]9706; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm19707; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]9708; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9709; AVX2-FP-NEXT: vmovaps 320(%rdi), %ymm09710; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]9711; AVX2-FP-NEXT: vmovaps 416(%rdi), %xmm19712; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]9713; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm19714; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]9715; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9716; AVX2-FP-NEXT: vmovaps (%rdi), %ymm09717; AVX2-FP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]9718; AVX2-FP-NEXT: vmovaps 96(%rdi), %xmm19719; AVX2-FP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]9720; AVX2-FP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm19721; AVX2-FP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]9722; AVX2-FP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9723; AVX2-FP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm0 # 32-byte Folded Reload9724; AVX2-FP-NEXT: # ymm0 = mem[8,9,10,11,12,13,14,15],ymm15[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm15[16,17,18,19,20,21,22,23]9725; AVX2-FP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,3]9726; AVX2-FP-NEXT: vmovdqa 64(%rdi), %xmm19727; AVX2-FP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9728; AVX2-FP-NEXT: vpalignr {{.*#+}} xmm1 = mem[8,9,10,11,12,13,14,15],xmm1[0,1,2,3,4,5,6,7]9729; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]9730; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9731; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload9732; AVX2-FP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload9733; AVX2-FP-NEXT: # ymm0 = mem[8,9,10,11,12,13,14,15],ymm0[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm0[16,17,18,19,20,21,22,23]9734; AVX2-FP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,3]9735; AVX2-FP-NEXT: vmovdqa 224(%rdi), %xmm19736; AVX2-FP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill9737; AVX2-FP-NEXT: vpalignr {{.*#+}} xmm1 = mem[8,9,10,11,12,13,14,15],xmm1[0,1,2,3,4,5,6,7]9738; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]9739; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9740; AVX2-FP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm0 # 32-byte Folded Reload9741; AVX2-FP-NEXT: # ymm0 = mem[8,9,10,11,12,13,14,15],ymm13[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm13[16,17,18,19,20,21,22,23]9742; AVX2-FP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,3]9743; AVX2-FP-NEXT: vmovdqa 384(%rdi), %xmm139744; AVX2-FP-NEXT: vpalignr {{.*#+}} xmm1 = mem[8,9,10,11,12,13,14,15],xmm13[0,1,2,3,4,5,6,7]9745; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]9746; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9747; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload9748; AVX2-FP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload9749; AVX2-FP-NEXT: # ymm0 = mem[8,9,10,11,12,13,14,15],ymm0[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm0[16,17,18,19,20,21,22,23]9750; AVX2-FP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,3]9751; AVX2-FP-NEXT: vmovdqa 544(%rdi), %xmm119752; AVX2-FP-NEXT: vpalignr {{.*#+}} xmm1 = mem[8,9,10,11,12,13,14,15],xmm11[0,1,2,3,4,5,6,7]9753; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]9754; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9755; AVX2-FP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm8, %ymm0 # 32-byte Folded Reload9756; AVX2-FP-NEXT: # ymm0 = mem[8,9,10,11,12,13,14,15],ymm8[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm8[16,17,18,19,20,21,22,23]9757; AVX2-FP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,3]9758; AVX2-FP-NEXT: vmovdqa 704(%rdi), %xmm129759; AVX2-FP-NEXT: vpalignr {{.*#+}} xmm1 = mem[8,9,10,11,12,13,14,15],xmm12[0,1,2,3,4,5,6,7]9760; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]9761; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9762; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload9763; AVX2-FP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload9764; AVX2-FP-NEXT: # ymm0 = mem[8,9,10,11,12,13,14,15],ymm0[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm0[16,17,18,19,20,21,22,23]9765; AVX2-FP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,3]9766; AVX2-FP-NEXT: vmovdqa 864(%rdi), %xmm109767; AVX2-FP-NEXT: vpalignr {{.*#+}} xmm1 = mem[8,9,10,11,12,13,14,15],xmm10[0,1,2,3,4,5,6,7]9768; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]9769; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9770; AVX2-FP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm0 # 32-byte Folded Reload9771; AVX2-FP-NEXT: # ymm0 = mem[8,9,10,11,12,13,14,15],ymm7[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm7[16,17,18,19,20,21,22,23]9772; AVX2-FP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,3]9773; AVX2-FP-NEXT: vmovdqa 1024(%rdi), %xmm99774; AVX2-FP-NEXT: vpalignr {{.*#+}} xmm1 = mem[8,9,10,11,12,13,14,15],xmm9[0,1,2,3,4,5,6,7]9775; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]9776; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9777; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload9778; AVX2-FP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload9779; AVX2-FP-NEXT: # ymm0 = mem[8,9,10,11,12,13,14,15],ymm0[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm0[16,17,18,19,20,21,22,23]9780; AVX2-FP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,3]9781; AVX2-FP-NEXT: vmovdqa 1184(%rdi), %xmm89782; AVX2-FP-NEXT: vpalignr {{.*#+}} xmm1 = mem[8,9,10,11,12,13,14,15],xmm8[0,1,2,3,4,5,6,7]9783; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]9784; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9785; AVX2-FP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm0 # 32-byte Folded Reload9786; AVX2-FP-NEXT: # ymm0 = mem[8,9,10,11,12,13,14,15],ymm5[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm5[16,17,18,19,20,21,22,23]9787; AVX2-FP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,3]9788; AVX2-FP-NEXT: vmovdqa 1344(%rdi), %xmm79789; AVX2-FP-NEXT: vpalignr {{.*#+}} xmm1 = mem[8,9,10,11,12,13,14,15],xmm7[0,1,2,3,4,5,6,7]9790; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]9791; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9792; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload9793; AVX2-FP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload9794; AVX2-FP-NEXT: # ymm0 = mem[8,9,10,11,12,13,14,15],ymm0[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm0[16,17,18,19,20,21,22,23]9795; AVX2-FP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,3]9796; AVX2-FP-NEXT: vmovdqa 1504(%rdi), %xmm69797; AVX2-FP-NEXT: vpalignr {{.*#+}} xmm1 = mem[8,9,10,11,12,13,14,15],xmm6[0,1,2,3,4,5,6,7]9798; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]9799; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9800; AVX2-FP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4, %ymm0 # 32-byte Folded Reload9801; AVX2-FP-NEXT: # ymm0 = mem[8,9,10,11,12,13,14,15],ymm4[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm4[16,17,18,19,20,21,22,23]9802; AVX2-FP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,3]9803; AVX2-FP-NEXT: vmovdqa 1664(%rdi), %xmm59804; AVX2-FP-NEXT: vpalignr {{.*#+}} xmm1 = mem[8,9,10,11,12,13,14,15],xmm5[0,1,2,3,4,5,6,7]9805; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]9806; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9807; AVX2-FP-NEXT: vmovdqu (%rsp), %ymm0 # 32-byte Reload9808; AVX2-FP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload9809; AVX2-FP-NEXT: # ymm0 = mem[8,9,10,11,12,13,14,15],ymm0[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm0[16,17,18,19,20,21,22,23]9810; AVX2-FP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,3]9811; AVX2-FP-NEXT: vmovdqa 1824(%rdi), %xmm49812; AVX2-FP-NEXT: vpalignr {{.*#+}} xmm1 = mem[8,9,10,11,12,13,14,15],xmm4[0,1,2,3,4,5,6,7]9813; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]9814; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9815; AVX2-FP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3, %ymm0 # 32-byte Folded Reload9816; AVX2-FP-NEXT: # ymm0 = mem[8,9,10,11,12,13,14,15],ymm3[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm3[16,17,18,19,20,21,22,23]9817; AVX2-FP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,3]9818; AVX2-FP-NEXT: vmovdqa 1984(%rdi), %xmm39819; AVX2-FP-NEXT: vpalignr {{.*#+}} xmm1 = mem[8,9,10,11,12,13,14,15],xmm3[0,1,2,3,4,5,6,7]9820; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]9821; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9822; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload9823; AVX2-FP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload9824; AVX2-FP-NEXT: # ymm0 = mem[8,9,10,11,12,13,14,15],ymm0[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm0[16,17,18,19,20,21,22,23]9825; AVX2-FP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,3]9826; AVX2-FP-NEXT: vmovdqa 2144(%rdi), %xmm29827; AVX2-FP-NEXT: vpalignr {{.*#+}} xmm1 = mem[8,9,10,11,12,13,14,15],xmm2[0,1,2,3,4,5,6,7]9828; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]9829; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9830; AVX2-FP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm14, %ymm0 # 32-byte Folded Reload9831; AVX2-FP-NEXT: # ymm0 = mem[8,9,10,11,12,13,14,15],ymm14[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm14[16,17,18,19,20,21,22,23]9832; AVX2-FP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,3]9833; AVX2-FP-NEXT: vmovdqa 2304(%rdi), %xmm19834; AVX2-FP-NEXT: vpalignr {{.*#+}} xmm15 = mem[8,9,10,11,12,13,14,15],xmm1[0,1,2,3,4,5,6,7]9835; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]9836; AVX2-FP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9837; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload9838; AVX2-FP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload9839; AVX2-FP-NEXT: # ymm0 = mem[8,9,10,11,12,13,14,15],ymm0[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm0[16,17,18,19,20,21,22,23]9840; AVX2-FP-NEXT: vpermq {{.*#+}} ymm15 = ymm0[0,1,0,3]9841; AVX2-FP-NEXT: vmovdqa 2464(%rdi), %xmm09842; AVX2-FP-NEXT: vpalignr {{.*#+}} xmm14 = mem[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]9843; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm14 = ymm14[0,1,2,3],ymm15[4,5,6,7]9844; AVX2-FP-NEXT: vmovdqu %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9845; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Reload9846; AVX2-FP-NEXT: vblendps $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm14, %ymm14 # 32-byte Folded Reload9847; AVX2-FP-NEXT: # ymm14 = mem[0,1,2,3,4,5],ymm14[6,7]9848; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload9849; AVX2-FP-NEXT: vblendps $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload9850; AVX2-FP-NEXT: # xmm15 = mem[0,1],xmm15[2,3]9851; AVX2-FP-NEXT: vblendps {{.*#+}} ymm14 = ymm15[0,1,2,3],ymm14[4,5,6,7]9852; AVX2-FP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9853; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Reload9854; AVX2-FP-NEXT: vblendps $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm14, %ymm14 # 32-byte Folded Reload9855; AVX2-FP-NEXT: # ymm14 = mem[0,1,2,3,4,5],ymm14[6,7]9856; AVX2-FP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload9857; AVX2-FP-NEXT: vblendps $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload9858; AVX2-FP-NEXT: # xmm15 = mem[0,1],xmm15[2,3]9859; AVX2-FP-NEXT: vblendps {{.*#+}} ymm15 = ymm15[0,1,2,3],ymm14[4,5,6,7]9860; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Reload9861; AVX2-FP-NEXT: vpblendd $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm14, %ymm14 # 32-byte Folded Reload9862; AVX2-FP-NEXT: # ymm14 = mem[0,1,2,3,4,5],ymm14[6,7]9863; AVX2-FP-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm13, %xmm13 # 16-byte Folded Reload9864; AVX2-FP-NEXT: # xmm13 = mem[0,1],xmm13[2,3]9865; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm13 = ymm13[0,1,2,3],ymm14[4,5,6,7]9866; AVX2-FP-NEXT: vmovdqu %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9867; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload9868; AVX2-FP-NEXT: vpblendd $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm13 # 32-byte Folded Reload9869; AVX2-FP-NEXT: # ymm13 = mem[0,1,2,3,4,5],ymm13[6,7]9870; AVX2-FP-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm11, %xmm11 # 16-byte Folded Reload9871; AVX2-FP-NEXT: # xmm11 = mem[0,1],xmm11[2,3]9872; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm11 = ymm11[0,1,2,3],ymm13[4,5,6,7]9873; AVX2-FP-NEXT: vmovdqu %ymm11, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill9874; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload9875; AVX2-FP-NEXT: vpblendd $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm11, %ymm11 # 32-byte Folded Reload9876; AVX2-FP-NEXT: # ymm11 = mem[0,1,2,3,4,5],ymm11[6,7]9877; AVX2-FP-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm12, %xmm12 # 16-byte Folded Reload9878; AVX2-FP-NEXT: # xmm12 = mem[0,1],xmm12[2,3]9879; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm12 = ymm12[0,1,2,3],ymm11[4,5,6,7]9880; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload9881; AVX2-FP-NEXT: vpblendd $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm11, %ymm11 # 32-byte Folded Reload9882; AVX2-FP-NEXT: # ymm11 = mem[0,1,2,3,4,5],ymm11[6,7]9883; AVX2-FP-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm10, %xmm10 # 16-byte Folded Reload9884; AVX2-FP-NEXT: # xmm10 = mem[0,1],xmm10[2,3]9885; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm11 = ymm10[0,1,2,3],ymm11[4,5,6,7]9886; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload9887; AVX2-FP-NEXT: vpblendd $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm10, %ymm10 # 32-byte Folded Reload9888; AVX2-FP-NEXT: # ymm10 = mem[0,1,2,3,4,5],ymm10[6,7]9889; AVX2-FP-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm9, %xmm9 # 16-byte Folded Reload9890; AVX2-FP-NEXT: # xmm9 = mem[0,1],xmm9[2,3]9891; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm9 = ymm9[0,1,2,3],ymm10[4,5,6,7]9892; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload9893; AVX2-FP-NEXT: vpblendd $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm10, %ymm10 # 32-byte Folded Reload9894; AVX2-FP-NEXT: # ymm10 = mem[0,1,2,3,4,5],ymm10[6,7]9895; AVX2-FP-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm8, %xmm8 # 16-byte Folded Reload9896; AVX2-FP-NEXT: # xmm8 = mem[0,1],xmm8[2,3]9897; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm14 = ymm8[0,1,2,3],ymm10[4,5,6,7]9898; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload9899; AVX2-FP-NEXT: vpblendd $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm8, %ymm8 # 32-byte Folded Reload9900; AVX2-FP-NEXT: # ymm8 = mem[0,1,2,3,4,5],ymm8[6,7]9901; AVX2-FP-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm7, %xmm7 # 16-byte Folded Reload9902; AVX2-FP-NEXT: # xmm7 = mem[0,1],xmm7[2,3]9903; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm10 = ymm7[0,1,2,3],ymm8[4,5,6,7]9904; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload9905; AVX2-FP-NEXT: vpblendd $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm7 # 32-byte Folded Reload9906; AVX2-FP-NEXT: # ymm7 = mem[0,1,2,3,4,5],ymm7[6,7]9907; AVX2-FP-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm6, %xmm6 # 16-byte Folded Reload9908; AVX2-FP-NEXT: # xmm6 = mem[0,1],xmm6[2,3]9909; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7]9910; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload9911; AVX2-FP-NEXT: vpblendd $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm7 # 32-byte Folded Reload9912; AVX2-FP-NEXT: # ymm7 = mem[0,1,2,3,4,5],ymm7[6,7]9913; AVX2-FP-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm5, %xmm5 # 16-byte Folded Reload9914; AVX2-FP-NEXT: # xmm5 = mem[0,1],xmm5[2,3]9915; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm13 = ymm5[0,1,2,3],ymm7[4,5,6,7]9916; AVX2-FP-NEXT: vmovdqu (%rsp), %ymm5 # 32-byte Reload9917; AVX2-FP-NEXT: vpblendd $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm5 # 32-byte Folded Reload9918; AVX2-FP-NEXT: # ymm5 = mem[0,1,2,3,4,5],ymm5[6,7]9919; AVX2-FP-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4, %xmm4 # 16-byte Folded Reload9920; AVX2-FP-NEXT: # xmm4 = mem[0,1],xmm4[2,3]9921; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm5[4,5,6,7]9922; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload9923; AVX2-FP-NEXT: vpblendd $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm5 # 32-byte Folded Reload9924; AVX2-FP-NEXT: # ymm5 = mem[0,1,2,3,4,5],ymm5[6,7]9925; AVX2-FP-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm3 # 16-byte Folded Reload9926; AVX2-FP-NEXT: # xmm3 = mem[0,1],xmm3[2,3]9927; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm7 = ymm3[0,1,2,3],ymm5[4,5,6,7]9928; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload9929; AVX2-FP-NEXT: vpblendd $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3, %ymm3 # 32-byte Folded Reload9930; AVX2-FP-NEXT: # ymm3 = mem[0,1,2,3,4,5],ymm3[6,7]9931; AVX2-FP-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm2 # 16-byte Folded Reload9932; AVX2-FP-NEXT: # xmm2 = mem[0,1],xmm2[2,3]9933; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm5 = ymm2[0,1,2,3],ymm3[4,5,6,7]9934; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload9935; AVX2-FP-NEXT: vpblendd $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload9936; AVX2-FP-NEXT: # ymm2 = mem[0,1,2,3,4,5],ymm2[6,7]9937; AVX2-FP-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload9938; AVX2-FP-NEXT: # xmm1 = mem[0,1],xmm1[2,3]9939; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm2 = ymm1[0,1,2,3],ymm2[4,5,6,7]9940; AVX2-FP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9941; AVX2-FP-NEXT: vpblendd $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload9942; AVX2-FP-NEXT: # ymm1 = mem[0,1,2,3,4,5],ymm1[6,7]9943; AVX2-FP-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload9944; AVX2-FP-NEXT: # xmm0 = mem[0,1],xmm0[2,3]9945; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]9946; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9947; AVX2-FP-NEXT: vmovaps %ymm1, 448(%rsi)9948; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9949; AVX2-FP-NEXT: vmovaps %ymm1, 384(%rsi)9950; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9951; AVX2-FP-NEXT: vmovaps %ymm1, 320(%rsi)9952; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9953; AVX2-FP-NEXT: vmovaps %ymm1, 256(%rsi)9954; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9955; AVX2-FP-NEXT: vmovaps %ymm1, 192(%rsi)9956; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9957; AVX2-FP-NEXT: vmovaps %ymm1, 128(%rsi)9958; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9959; AVX2-FP-NEXT: vmovaps %ymm1, 64(%rsi)9960; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9961; AVX2-FP-NEXT: vmovaps %ymm1, (%rsi)9962; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9963; AVX2-FP-NEXT: vmovaps %ymm1, 480(%rsi)9964; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9965; AVX2-FP-NEXT: vmovaps %ymm1, 416(%rsi)9966; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9967; AVX2-FP-NEXT: vmovaps %ymm1, 352(%rsi)9968; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9969; AVX2-FP-NEXT: vmovaps %ymm1, 288(%rsi)9970; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9971; AVX2-FP-NEXT: vmovaps %ymm1, 224(%rsi)9972; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9973; AVX2-FP-NEXT: vmovaps %ymm1, 160(%rsi)9974; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9975; AVX2-FP-NEXT: vmovaps %ymm1, 96(%rsi)9976; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9977; AVX2-FP-NEXT: vmovaps %ymm1, 32(%rsi)9978; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9979; AVX2-FP-NEXT: vmovaps %ymm1, 448(%rdx)9980; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9981; AVX2-FP-NEXT: vmovaps %ymm1, 384(%rdx)9982; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9983; AVX2-FP-NEXT: vmovaps %ymm1, 320(%rdx)9984; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9985; AVX2-FP-NEXT: vmovaps %ymm1, 256(%rdx)9986; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9987; AVX2-FP-NEXT: vmovaps %ymm1, 192(%rdx)9988; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9989; AVX2-FP-NEXT: vmovaps %ymm1, 128(%rdx)9990; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9991; AVX2-FP-NEXT: vmovaps %ymm1, 64(%rdx)9992; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9993; AVX2-FP-NEXT: vmovaps %ymm1, (%rdx)9994; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9995; AVX2-FP-NEXT: vmovaps %ymm1, 480(%rdx)9996; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9997; AVX2-FP-NEXT: vmovaps %ymm1, 416(%rdx)9998; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload9999; AVX2-FP-NEXT: vmovaps %ymm1, 352(%rdx)10000; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10001; AVX2-FP-NEXT: vmovaps %ymm1, 288(%rdx)10002; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10003; AVX2-FP-NEXT: vmovaps %ymm1, 224(%rdx)10004; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10005; AVX2-FP-NEXT: vmovaps %ymm1, 160(%rdx)10006; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10007; AVX2-FP-NEXT: vmovaps %ymm1, 96(%rdx)10008; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10009; AVX2-FP-NEXT: vmovaps %ymm1, 32(%rdx)10010; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10011; AVX2-FP-NEXT: vmovaps %ymm1, (%rcx)10012; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10013; AVX2-FP-NEXT: vmovaps %ymm1, 64(%rcx)10014; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10015; AVX2-FP-NEXT: vmovaps %ymm1, 128(%rcx)10016; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10017; AVX2-FP-NEXT: vmovaps %ymm1, 192(%rcx)10018; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10019; AVX2-FP-NEXT: vmovaps %ymm1, 256(%rcx)10020; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10021; AVX2-FP-NEXT: vmovaps %ymm1, 320(%rcx)10022; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10023; AVX2-FP-NEXT: vmovaps %ymm1, 384(%rcx)10024; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10025; AVX2-FP-NEXT: vmovaps %ymm1, 448(%rcx)10026; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10027; AVX2-FP-NEXT: vmovaps %ymm1, 480(%rcx)10028; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10029; AVX2-FP-NEXT: vmovaps %ymm1, 416(%rcx)10030; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10031; AVX2-FP-NEXT: vmovaps %ymm1, 352(%rcx)10032; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10033; AVX2-FP-NEXT: vmovaps %ymm1, 288(%rcx)10034; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10035; AVX2-FP-NEXT: vmovaps %ymm1, 224(%rcx)10036; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10037; AVX2-FP-NEXT: vmovaps %ymm1, 160(%rcx)10038; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10039; AVX2-FP-NEXT: vmovaps %ymm1, 96(%rcx)10040; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10041; AVX2-FP-NEXT: vmovaps %ymm1, 32(%rcx)10042; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10043; AVX2-FP-NEXT: vmovaps %ymm1, 480(%r8)10044; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10045; AVX2-FP-NEXT: vmovaps %ymm1, 448(%r8)10046; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10047; AVX2-FP-NEXT: vmovaps %ymm1, 416(%r8)10048; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10049; AVX2-FP-NEXT: vmovaps %ymm1, 384(%r8)10050; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10051; AVX2-FP-NEXT: vmovaps %ymm1, 352(%r8)10052; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10053; AVX2-FP-NEXT: vmovaps %ymm1, 320(%r8)10054; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10055; AVX2-FP-NEXT: vmovaps %ymm1, 288(%r8)10056; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10057; AVX2-FP-NEXT: vmovaps %ymm1, 256(%r8)10058; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10059; AVX2-FP-NEXT: vmovaps %ymm1, 224(%r8)10060; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10061; AVX2-FP-NEXT: vmovaps %ymm1, 192(%r8)10062; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10063; AVX2-FP-NEXT: vmovaps %ymm1, 160(%r8)10064; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10065; AVX2-FP-NEXT: vmovaps %ymm1, 128(%r8)10066; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10067; AVX2-FP-NEXT: vmovaps %ymm1, 96(%r8)10068; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10069; AVX2-FP-NEXT: vmovaps %ymm1, 64(%r8)10070; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10071; AVX2-FP-NEXT: vmovaps %ymm1, 32(%r8)10072; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10073; AVX2-FP-NEXT: vmovaps %ymm1, (%r8)10074; AVX2-FP-NEXT: vmovdqa %ymm0, 480(%r9)10075; AVX2-FP-NEXT: vmovdqa %ymm2, 448(%r9)10076; AVX2-FP-NEXT: vmovdqa %ymm5, 416(%r9)10077; AVX2-FP-NEXT: vmovdqa %ymm7, 384(%r9)10078; AVX2-FP-NEXT: vmovdqa %ymm4, 352(%r9)10079; AVX2-FP-NEXT: vmovdqa %ymm13, 320(%r9)10080; AVX2-FP-NEXT: vmovdqa %ymm6, 288(%r9)10081; AVX2-FP-NEXT: vmovdqa %ymm10, 256(%r9)10082; AVX2-FP-NEXT: vmovdqa %ymm14, 224(%r9)10083; AVX2-FP-NEXT: vmovdqa %ymm9, 192(%r9)10084; AVX2-FP-NEXT: vmovdqa %ymm11, 160(%r9)10085; AVX2-FP-NEXT: vmovdqa %ymm12, 128(%r9)10086; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10087; AVX2-FP-NEXT: vmovaps %ymm0, 96(%r9)10088; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10089; AVX2-FP-NEXT: vmovaps %ymm0, 64(%r9)10090; AVX2-FP-NEXT: vmovaps %ymm15, 32(%r9)10091; AVX2-FP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10092; AVX2-FP-NEXT: vmovaps %ymm0, (%r9)10093; AVX2-FP-NEXT: addq $3240, %rsp # imm = 0xCA810094; AVX2-FP-NEXT: vzeroupper10095; AVX2-FP-NEXT: retq10096;10097; AVX2-FCP-LABEL: load_i64_stride5_vf64:10098; AVX2-FCP: # %bb.0:10099; AVX2-FCP-NEXT: subq $3240, %rsp # imm = 0xCA810100; AVX2-FCP-NEXT: vmovaps 896(%rdi), %ymm210101; AVX2-FCP-NEXT: vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10102; AVX2-FCP-NEXT: vmovaps 864(%rdi), %ymm410103; AVX2-FCP-NEXT: vmovups %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10104; AVX2-FCP-NEXT: vmovaps 576(%rdi), %ymm310105; AVX2-FCP-NEXT: vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10106; AVX2-FCP-NEXT: vmovaps 544(%rdi), %ymm510107; AVX2-FCP-NEXT: vmovups %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10108; AVX2-FCP-NEXT: vmovaps 256(%rdi), %ymm010109; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10110; AVX2-FCP-NEXT: vmovaps 224(%rdi), %ymm110111; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10112; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]10113; AVX2-FCP-NEXT: vmovaps 192(%rdi), %xmm110114; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10115; AVX2-FCP-NEXT: vmovaps 160(%rdi), %xmm610116; AVX2-FCP-NEXT: vmovaps %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10117; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm6[0,1],xmm1[2,3]10118; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]10119; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10120; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm5[0,1,2,3,4,5],ymm3[6,7]10121; AVX2-FCP-NEXT: vmovaps 512(%rdi), %xmm110122; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10123; AVX2-FCP-NEXT: vmovaps 480(%rdi), %xmm310124; AVX2-FCP-NEXT: vmovaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10125; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm3[0,1],xmm1[2,3]10126; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]10127; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10128; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm4[0,1,2,3,4,5],ymm2[6,7]10129; AVX2-FCP-NEXT: vmovaps 832(%rdi), %xmm110130; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10131; AVX2-FCP-NEXT: vmovaps 800(%rdi), %xmm210132; AVX2-FCP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10133; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3]10134; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]10135; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10136; AVX2-FCP-NEXT: vmovaps 1216(%rdi), %ymm010137; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10138; AVX2-FCP-NEXT: vmovaps 1184(%rdi), %ymm110139; AVX2-FCP-NEXT: vmovups %ymm1, (%rsp) # 32-byte Spill10140; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]10141; AVX2-FCP-NEXT: vmovaps 1152(%rdi), %xmm110142; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10143; AVX2-FCP-NEXT: vmovaps 1120(%rdi), %xmm210144; AVX2-FCP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10145; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3]10146; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]10147; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10148; AVX2-FCP-NEXT: vmovaps 1536(%rdi), %ymm010149; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10150; AVX2-FCP-NEXT: vmovaps 1504(%rdi), %ymm110151; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10152; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]10153; AVX2-FCP-NEXT: vmovaps 1472(%rdi), %xmm110154; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10155; AVX2-FCP-NEXT: vmovaps 1440(%rdi), %xmm210156; AVX2-FCP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10157; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3]10158; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]10159; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10160; AVX2-FCP-NEXT: vmovaps 1856(%rdi), %ymm010161; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10162; AVX2-FCP-NEXT: vmovaps 1824(%rdi), %ymm110163; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10164; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]10165; AVX2-FCP-NEXT: vmovaps 1792(%rdi), %xmm110166; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10167; AVX2-FCP-NEXT: vmovaps 1760(%rdi), %xmm210168; AVX2-FCP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10169; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3]10170; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]10171; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10172; AVX2-FCP-NEXT: vmovaps 2176(%rdi), %ymm010173; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10174; AVX2-FCP-NEXT: vmovaps 2144(%rdi), %ymm110175; AVX2-FCP-NEXT: vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10176; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]10177; AVX2-FCP-NEXT: vmovaps 2112(%rdi), %xmm110178; AVX2-FCP-NEXT: vmovaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10179; AVX2-FCP-NEXT: vmovaps 2080(%rdi), %xmm210180; AVX2-FCP-NEXT: vmovaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10181; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm2[0,1],xmm1[2,3]10182; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]10183; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10184; AVX2-FCP-NEXT: vmovdqa 2496(%rdi), %ymm010185; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10186; AVX2-FCP-NEXT: vmovdqa 2464(%rdi), %ymm110187; AVX2-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10188; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]10189; AVX2-FCP-NEXT: vmovdqa 2432(%rdi), %xmm110190; AVX2-FCP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10191; AVX2-FCP-NEXT: vmovdqa 2400(%rdi), %xmm1310192; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm1 = xmm13[0,1],xmm1[2,3]10193; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]10194; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10195; AVX2-FCP-NEXT: vmovdqa 96(%rdi), %ymm010196; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10197; AVX2-FCP-NEXT: vmovdqa 64(%rdi), %ymm110198; AVX2-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10199; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]10200; AVX2-FCP-NEXT: vmovdqa (%rdi), %xmm1410201; AVX2-FCP-NEXT: vmovdqa 32(%rdi), %xmm110202; AVX2-FCP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10203; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm1 = xmm14[0,1],xmm1[2,3]10204; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]10205; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10206; AVX2-FCP-NEXT: vmovdqa 416(%rdi), %ymm010207; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10208; AVX2-FCP-NEXT: vmovdqa 384(%rdi), %ymm110209; AVX2-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10210; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]10211; AVX2-FCP-NEXT: vmovdqa 352(%rdi), %xmm110212; AVX2-FCP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10213; AVX2-FCP-NEXT: vmovdqa 320(%rdi), %xmm810214; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm1 = xmm8[0,1],xmm1[2,3]10215; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]10216; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10217; AVX2-FCP-NEXT: vmovdqa 736(%rdi), %ymm010218; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10219; AVX2-FCP-NEXT: vmovdqa 704(%rdi), %ymm110220; AVX2-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10221; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]10222; AVX2-FCP-NEXT: vmovdqa 672(%rdi), %xmm110223; AVX2-FCP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10224; AVX2-FCP-NEXT: vmovdqa 640(%rdi), %xmm710225; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm1 = xmm7[0,1],xmm1[2,3]10226; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]10227; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10228; AVX2-FCP-NEXT: vmovdqa 1056(%rdi), %ymm010229; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10230; AVX2-FCP-NEXT: vmovdqa 1024(%rdi), %ymm110231; AVX2-FCP-NEXT: vmovdqu %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10232; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]10233; AVX2-FCP-NEXT: vmovdqa 992(%rdi), %xmm110234; AVX2-FCP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10235; AVX2-FCP-NEXT: vmovdqa 960(%rdi), %xmm510236; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm1 = xmm5[0,1],xmm1[2,3]10237; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]10238; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10239; AVX2-FCP-NEXT: vmovdqa 1376(%rdi), %ymm010240; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10241; AVX2-FCP-NEXT: vmovdqa 1344(%rdi), %ymm1210242; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm12[0,1,2,3,4,5],ymm0[6,7]10243; AVX2-FCP-NEXT: vmovdqa 1312(%rdi), %xmm110244; AVX2-FCP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10245; AVX2-FCP-NEXT: vmovdqa 1280(%rdi), %xmm410246; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm1 = xmm4[0,1],xmm1[2,3]10247; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]10248; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10249; AVX2-FCP-NEXT: vmovdqa 1696(%rdi), %ymm010250; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10251; AVX2-FCP-NEXT: vmovdqa 1664(%rdi), %ymm1110252; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm11[0,1,2,3,4,5],ymm0[6,7]10253; AVX2-FCP-NEXT: vmovdqa 1632(%rdi), %xmm110254; AVX2-FCP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10255; AVX2-FCP-NEXT: vmovdqa 1600(%rdi), %xmm310256; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm1 = xmm3[0,1],xmm1[2,3]10257; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]10258; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10259; AVX2-FCP-NEXT: vmovdqa 2016(%rdi), %ymm010260; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10261; AVX2-FCP-NEXT: vmovdqa 1984(%rdi), %ymm1010262; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm10[0,1,2,3,4,5],ymm0[6,7]10263; AVX2-FCP-NEXT: vmovdqa 1952(%rdi), %xmm110264; AVX2-FCP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10265; AVX2-FCP-NEXT: vmovdqa 1920(%rdi), %xmm910266; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm6 = xmm9[0,1],xmm1[2,3]10267; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm6[0,1,2,3],ymm0[4,5,6,7]10268; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10269; AVX2-FCP-NEXT: vmovdqa 2336(%rdi), %ymm010270; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10271; AVX2-FCP-NEXT: vmovdqa 2304(%rdi), %ymm610272; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm6[0,1,2,3,4,5],ymm0[6,7]10273; AVX2-FCP-NEXT: vmovdqa 2272(%rdi), %xmm010274; AVX2-FCP-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10275; AVX2-FCP-NEXT: vmovdqa 2240(%rdi), %xmm110276; AVX2-FCP-NEXT: vpblendd {{.*#+}} xmm15 = xmm1[0,1],xmm0[2,3]10277; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm2[4,5,6,7]10278; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10279; AVX2-FCP-NEXT: vmovdqa 208(%rdi), %xmm010280; AVX2-FCP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload10281; AVX2-FCP-NEXT: # xmm0 = mem[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]10282; AVX2-FCP-NEXT: vmovdqa 288(%rdi), %ymm210283; AVX2-FCP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10284; AVX2-FCP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm15 # 32-byte Folded Reload10285; AVX2-FCP-NEXT: # ymm15 = mem[8,9,10,11,12,13,14,15],ymm2[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm2[16,17,18,19,20,21,22,23]10286; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm15 = ymm15[0,1,2,1]10287; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm15[4,5,6,7]10288; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10289; AVX2-FCP-NEXT: vmovdqa 528(%rdi), %xmm010290; AVX2-FCP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload10291; AVX2-FCP-NEXT: # xmm0 = mem[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]10292; AVX2-FCP-NEXT: vmovdqa 608(%rdi), %ymm210293; AVX2-FCP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10294; AVX2-FCP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm15 # 32-byte Folded Reload10295; AVX2-FCP-NEXT: # ymm15 = mem[8,9,10,11,12,13,14,15],ymm2[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm2[16,17,18,19,20,21,22,23]10296; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm15 = ymm15[0,1,2,1]10297; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm15[4,5,6,7]10298; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10299; AVX2-FCP-NEXT: vmovdqa 848(%rdi), %xmm010300; AVX2-FCP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload10301; AVX2-FCP-NEXT: # xmm0 = mem[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]10302; AVX2-FCP-NEXT: vmovdqa 928(%rdi), %ymm210303; AVX2-FCP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10304; AVX2-FCP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm15 # 32-byte Folded Reload10305; AVX2-FCP-NEXT: # ymm15 = mem[8,9,10,11,12,13,14,15],ymm2[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm2[16,17,18,19,20,21,22,23]10306; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm15 = ymm15[0,1,2,1]10307; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm15[4,5,6,7]10308; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10309; AVX2-FCP-NEXT: vmovdqa 1168(%rdi), %xmm010310; AVX2-FCP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload10311; AVX2-FCP-NEXT: # xmm0 = mem[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]10312; AVX2-FCP-NEXT: vmovdqa 1248(%rdi), %ymm210313; AVX2-FCP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10314; AVX2-FCP-NEXT: vpalignr $8, (%rsp), %ymm2, %ymm15 # 32-byte Folded Reload10315; AVX2-FCP-NEXT: # ymm15 = mem[8,9,10,11,12,13,14,15],ymm2[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm2[16,17,18,19,20,21,22,23]10316; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm15 = ymm15[0,1,2,1]10317; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm15[4,5,6,7]10318; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10319; AVX2-FCP-NEXT: vmovdqa 1488(%rdi), %xmm010320; AVX2-FCP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload10321; AVX2-FCP-NEXT: # xmm0 = mem[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]10322; AVX2-FCP-NEXT: vmovdqa 1568(%rdi), %ymm210323; AVX2-FCP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10324; AVX2-FCP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm15 # 32-byte Folded Reload10325; AVX2-FCP-NEXT: # ymm15 = mem[8,9,10,11,12,13,14,15],ymm2[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm2[16,17,18,19,20,21,22,23]10326; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm15 = ymm15[0,1,2,1]10327; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm15[4,5,6,7]10328; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10329; AVX2-FCP-NEXT: vmovdqa 1808(%rdi), %xmm010330; AVX2-FCP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload10331; AVX2-FCP-NEXT: # xmm0 = mem[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]10332; AVX2-FCP-NEXT: vmovdqa 1888(%rdi), %ymm210333; AVX2-FCP-NEXT: vmovdqu %ymm2, (%rsp) # 32-byte Spill10334; AVX2-FCP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm15 # 32-byte Folded Reload10335; AVX2-FCP-NEXT: # ymm15 = mem[8,9,10,11,12,13,14,15],ymm2[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm2[16,17,18,19,20,21,22,23]10336; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm15 = ymm15[0,1,2,1]10337; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm15[4,5,6,7]10338; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10339; AVX2-FCP-NEXT: vmovdqa 2128(%rdi), %xmm010340; AVX2-FCP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload10341; AVX2-FCP-NEXT: # xmm0 = mem[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]10342; AVX2-FCP-NEXT: vmovdqa 2208(%rdi), %ymm210343; AVX2-FCP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10344; AVX2-FCP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm15 # 32-byte Folded Reload10345; AVX2-FCP-NEXT: # ymm15 = mem[8,9,10,11,12,13,14,15],ymm2[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm2[16,17,18,19,20,21,22,23]10346; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm15 = ymm15[0,1,2,1]10347; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm15[4,5,6,7]10348; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10349; AVX2-FCP-NEXT: vmovdqa 2448(%rdi), %xmm010350; AVX2-FCP-NEXT: vpalignr {{.*#+}} xmm0 = xmm13[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]10351; AVX2-FCP-NEXT: vmovdqa 2528(%rdi), %ymm210352; AVX2-FCP-NEXT: vmovdqu %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10353; AVX2-FCP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm13 # 32-byte Folded Reload10354; AVX2-FCP-NEXT: # ymm13 = mem[8,9,10,11,12,13,14,15],ymm2[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm2[16,17,18,19,20,21,22,23]10355; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm13 = ymm13[0,1,2,1]10356; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm13[4,5,6,7]10357; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10358; AVX2-FCP-NEXT: vmovdqa 48(%rdi), %xmm010359; AVX2-FCP-NEXT: vpalignr {{.*#+}} xmm0 = xmm14[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]10360; AVX2-FCP-NEXT: vmovdqa 128(%rdi), %ymm1510361; AVX2-FCP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm13 # 32-byte Folded Reload10362; AVX2-FCP-NEXT: # ymm13 = mem[8,9,10,11,12,13,14,15],ymm15[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm15[16,17,18,19,20,21,22,23]10363; AVX2-FCP-NEXT: vmovdqu %ymm15, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10364; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm13 = ymm13[0,1,2,1]10365; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm13[4,5,6,7]10366; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10367; AVX2-FCP-NEXT: vmovdqa 368(%rdi), %xmm010368; AVX2-FCP-NEXT: vpalignr {{.*#+}} xmm0 = xmm8[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]10369; AVX2-FCP-NEXT: vmovdqa 448(%rdi), %ymm1310370; AVX2-FCP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm8 # 32-byte Folded Reload10371; AVX2-FCP-NEXT: # ymm8 = mem[8,9,10,11,12,13,14,15],ymm13[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm13[16,17,18,19,20,21,22,23]10372; AVX2-FCP-NEXT: vmovdqu %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10373; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm8 = ymm8[0,1,2,1]10374; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm8[4,5,6,7]10375; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10376; AVX2-FCP-NEXT: vmovdqa 688(%rdi), %xmm010377; AVX2-FCP-NEXT: vpalignr {{.*#+}} xmm0 = xmm7[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]10378; AVX2-FCP-NEXT: vmovdqa 768(%rdi), %ymm810379; AVX2-FCP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm8, %ymm7 # 32-byte Folded Reload10380; AVX2-FCP-NEXT: # ymm7 = mem[8,9,10,11,12,13,14,15],ymm8[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm8[16,17,18,19,20,21,22,23]10381; AVX2-FCP-NEXT: vmovdqu %ymm8, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10382; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,1,2,1]10383; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm7[4,5,6,7]10384; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10385; AVX2-FCP-NEXT: vmovdqa 1008(%rdi), %xmm010386; AVX2-FCP-NEXT: vpalignr {{.*#+}} xmm0 = xmm5[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]10387; AVX2-FCP-NEXT: vmovdqa 1088(%rdi), %ymm710388; AVX2-FCP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm5 # 32-byte Folded Reload10389; AVX2-FCP-NEXT: # ymm5 = mem[8,9,10,11,12,13,14,15],ymm7[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm7[16,17,18,19,20,21,22,23]10390; AVX2-FCP-NEXT: vmovdqu %ymm7, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10391; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm5 = ymm5[0,1,2,1]10392; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm5[4,5,6,7]10393; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10394; AVX2-FCP-NEXT: vmovdqa 1328(%rdi), %xmm010395; AVX2-FCP-NEXT: vpalignr {{.*#+}} xmm0 = xmm4[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]10396; AVX2-FCP-NEXT: vmovdqa 1408(%rdi), %ymm510397; AVX2-FCP-NEXT: vpalignr {{.*#+}} ymm4 = ymm12[8,9,10,11,12,13,14,15],ymm5[0,1,2,3,4,5,6,7],ymm12[24,25,26,27,28,29,30,31],ymm5[16,17,18,19,20,21,22,23]10398; AVX2-FCP-NEXT: vmovdqu %ymm5, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10399; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,1,2,1]10400; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm4[4,5,6,7]10401; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10402; AVX2-FCP-NEXT: vmovdqa 1648(%rdi), %xmm010403; AVX2-FCP-NEXT: vpalignr {{.*#+}} xmm0 = xmm3[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]10404; AVX2-FCP-NEXT: vmovdqa 1728(%rdi), %ymm410405; AVX2-FCP-NEXT: vpalignr {{.*#+}} ymm3 = ymm11[8,9,10,11,12,13,14,15],ymm4[0,1,2,3,4,5,6,7],ymm11[24,25,26,27,28,29,30,31],ymm4[16,17,18,19,20,21,22,23]10406; AVX2-FCP-NEXT: vmovdqu %ymm4, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10407; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,1,2,1]10408; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm3[4,5,6,7]10409; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10410; AVX2-FCP-NEXT: vmovdqa 1968(%rdi), %xmm010411; AVX2-FCP-NEXT: vpalignr {{.*#+}} xmm0 = xmm9[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]10412; AVX2-FCP-NEXT: vmovdqa 2048(%rdi), %ymm310413; AVX2-FCP-NEXT: vpalignr {{.*#+}} ymm2 = ymm10[8,9,10,11,12,13,14,15],ymm3[0,1,2,3,4,5,6,7],ymm10[24,25,26,27,28,29,30,31],ymm3[16,17,18,19,20,21,22,23]10414; AVX2-FCP-NEXT: vmovdqu %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10415; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,1,2,1]10416; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm2[4,5,6,7]10417; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10418; AVX2-FCP-NEXT: vmovdqa 2288(%rdi), %xmm010419; AVX2-FCP-NEXT: vpalignr {{.*#+}} xmm0 = xmm1[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]10420; AVX2-FCP-NEXT: vmovdqa 2368(%rdi), %ymm1410421; AVX2-FCP-NEXT: vpalignr {{.*#+}} ymm1 = ymm6[8,9,10,11,12,13,14,15],ymm14[0,1,2,3,4,5,6,7],ymm6[24,25,26,27,28,29,30,31],ymm14[16,17,18,19,20,21,22,23]10422; AVX2-FCP-NEXT: vmovdqu %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10423; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm1 = ymm1[0,1,2,1]10424; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]10425; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10426; AVX2-FCP-NEXT: vmovaps 160(%rdi), %ymm010427; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]10428; AVX2-FCP-NEXT: vmovaps 256(%rdi), %xmm110429; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]10430; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm110431; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]10432; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10433; AVX2-FCP-NEXT: vmovaps 480(%rdi), %ymm010434; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]10435; AVX2-FCP-NEXT: vmovaps 576(%rdi), %xmm110436; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]10437; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm110438; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]10439; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10440; AVX2-FCP-NEXT: vmovaps 800(%rdi), %ymm010441; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]10442; AVX2-FCP-NEXT: vmovaps 896(%rdi), %xmm110443; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]10444; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm110445; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]10446; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10447; AVX2-FCP-NEXT: vmovaps 1120(%rdi), %ymm010448; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]10449; AVX2-FCP-NEXT: vmovaps 1216(%rdi), %xmm110450; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]10451; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm110452; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]10453; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10454; AVX2-FCP-NEXT: vmovaps 1440(%rdi), %ymm010455; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]10456; AVX2-FCP-NEXT: vmovaps 1536(%rdi), %xmm110457; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]10458; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm110459; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]10460; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10461; AVX2-FCP-NEXT: vmovaps 1760(%rdi), %ymm010462; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]10463; AVX2-FCP-NEXT: vmovaps 1856(%rdi), %xmm110464; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]10465; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm110466; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]10467; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10468; AVX2-FCP-NEXT: vmovaps 2080(%rdi), %ymm010469; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]10470; AVX2-FCP-NEXT: vmovaps 2176(%rdi), %xmm110471; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]10472; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm110473; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]10474; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10475; AVX2-FCP-NEXT: vmovaps 2400(%rdi), %ymm010476; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]10477; AVX2-FCP-NEXT: vmovaps 2496(%rdi), %xmm110478; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]10479; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm110480; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]10481; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10482; AVX2-FCP-NEXT: vmovaps 2240(%rdi), %ymm010483; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]10484; AVX2-FCP-NEXT: vmovaps 2336(%rdi), %xmm110485; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]10486; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm110487; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]10488; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10489; AVX2-FCP-NEXT: vmovaps 1920(%rdi), %ymm010490; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]10491; AVX2-FCP-NEXT: vmovaps 2016(%rdi), %xmm110492; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]10493; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm110494; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]10495; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10496; AVX2-FCP-NEXT: vmovaps 1600(%rdi), %ymm010497; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]10498; AVX2-FCP-NEXT: vmovaps 1696(%rdi), %xmm110499; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]10500; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm110501; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]10502; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10503; AVX2-FCP-NEXT: vmovaps 1280(%rdi), %ymm010504; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]10505; AVX2-FCP-NEXT: vmovaps 1376(%rdi), %xmm110506; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]10507; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm110508; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]10509; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10510; AVX2-FCP-NEXT: vmovaps 960(%rdi), %ymm010511; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]10512; AVX2-FCP-NEXT: vmovaps 1056(%rdi), %xmm110513; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]10514; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm110515; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]10516; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10517; AVX2-FCP-NEXT: vmovaps 640(%rdi), %ymm010518; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]10519; AVX2-FCP-NEXT: vmovaps 736(%rdi), %xmm110520; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]10521; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm110522; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]10523; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10524; AVX2-FCP-NEXT: vmovaps 320(%rdi), %ymm010525; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]10526; AVX2-FCP-NEXT: vmovaps 416(%rdi), %xmm110527; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]10528; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm110529; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]10530; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10531; AVX2-FCP-NEXT: vmovaps (%rdi), %ymm010532; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm0 = ymm0[0,1],mem[2,3],ymm0[4,5],mem[6,7]10533; AVX2-FCP-NEXT: vmovaps 96(%rdi), %xmm110534; AVX2-FCP-NEXT: vblendps {{.*#+}} xmm1 = xmm1[0,1],mem[2,3]10535; AVX2-FCP-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm110536; AVX2-FCP-NEXT: vperm2f128 {{.*#+}} ymm0 = ymm0[2,3],ymm1[2,3]10537; AVX2-FCP-NEXT: vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10538; AVX2-FCP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm15, %ymm0 # 32-byte Folded Reload10539; AVX2-FCP-NEXT: # ymm0 = mem[8,9,10,11,12,13,14,15],ymm15[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm15[16,17,18,19,20,21,22,23]10540; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,3]10541; AVX2-FCP-NEXT: vmovdqa 64(%rdi), %xmm110542; AVX2-FCP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10543; AVX2-FCP-NEXT: vpalignr {{.*#+}} xmm1 = mem[8,9,10,11,12,13,14,15],xmm1[0,1,2,3,4,5,6,7]10544; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]10545; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10546; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10547; AVX2-FCP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload10548; AVX2-FCP-NEXT: # ymm0 = mem[8,9,10,11,12,13,14,15],ymm0[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm0[16,17,18,19,20,21,22,23]10549; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,3]10550; AVX2-FCP-NEXT: vmovdqa 224(%rdi), %xmm110551; AVX2-FCP-NEXT: vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill10552; AVX2-FCP-NEXT: vpalignr {{.*#+}} xmm1 = mem[8,9,10,11,12,13,14,15],xmm1[0,1,2,3,4,5,6,7]10553; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]10554; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10555; AVX2-FCP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm0 # 32-byte Folded Reload10556; AVX2-FCP-NEXT: # ymm0 = mem[8,9,10,11,12,13,14,15],ymm13[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm13[16,17,18,19,20,21,22,23]10557; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,3]10558; AVX2-FCP-NEXT: vmovdqa 384(%rdi), %xmm1310559; AVX2-FCP-NEXT: vpalignr {{.*#+}} xmm1 = mem[8,9,10,11,12,13,14,15],xmm13[0,1,2,3,4,5,6,7]10560; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]10561; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10562; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10563; AVX2-FCP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload10564; AVX2-FCP-NEXT: # ymm0 = mem[8,9,10,11,12,13,14,15],ymm0[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm0[16,17,18,19,20,21,22,23]10565; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,3]10566; AVX2-FCP-NEXT: vmovdqa 544(%rdi), %xmm1110567; AVX2-FCP-NEXT: vpalignr {{.*#+}} xmm1 = mem[8,9,10,11,12,13,14,15],xmm11[0,1,2,3,4,5,6,7]10568; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]10569; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10570; AVX2-FCP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm8, %ymm0 # 32-byte Folded Reload10571; AVX2-FCP-NEXT: # ymm0 = mem[8,9,10,11,12,13,14,15],ymm8[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm8[16,17,18,19,20,21,22,23]10572; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,3]10573; AVX2-FCP-NEXT: vmovdqa 704(%rdi), %xmm1210574; AVX2-FCP-NEXT: vpalignr {{.*#+}} xmm1 = mem[8,9,10,11,12,13,14,15],xmm12[0,1,2,3,4,5,6,7]10575; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]10576; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10577; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10578; AVX2-FCP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload10579; AVX2-FCP-NEXT: # ymm0 = mem[8,9,10,11,12,13,14,15],ymm0[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm0[16,17,18,19,20,21,22,23]10580; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,3]10581; AVX2-FCP-NEXT: vmovdqa 864(%rdi), %xmm1010582; AVX2-FCP-NEXT: vpalignr {{.*#+}} xmm1 = mem[8,9,10,11,12,13,14,15],xmm10[0,1,2,3,4,5,6,7]10583; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]10584; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10585; AVX2-FCP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm0 # 32-byte Folded Reload10586; AVX2-FCP-NEXT: # ymm0 = mem[8,9,10,11,12,13,14,15],ymm7[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm7[16,17,18,19,20,21,22,23]10587; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,3]10588; AVX2-FCP-NEXT: vmovdqa 1024(%rdi), %xmm910589; AVX2-FCP-NEXT: vpalignr {{.*#+}} xmm1 = mem[8,9,10,11,12,13,14,15],xmm9[0,1,2,3,4,5,6,7]10590; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]10591; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10592; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10593; AVX2-FCP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload10594; AVX2-FCP-NEXT: # ymm0 = mem[8,9,10,11,12,13,14,15],ymm0[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm0[16,17,18,19,20,21,22,23]10595; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,3]10596; AVX2-FCP-NEXT: vmovdqa 1184(%rdi), %xmm810597; AVX2-FCP-NEXT: vpalignr {{.*#+}} xmm1 = mem[8,9,10,11,12,13,14,15],xmm8[0,1,2,3,4,5,6,7]10598; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]10599; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10600; AVX2-FCP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm0 # 32-byte Folded Reload10601; AVX2-FCP-NEXT: # ymm0 = mem[8,9,10,11,12,13,14,15],ymm5[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm5[16,17,18,19,20,21,22,23]10602; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,3]10603; AVX2-FCP-NEXT: vmovdqa 1344(%rdi), %xmm710604; AVX2-FCP-NEXT: vpalignr {{.*#+}} xmm1 = mem[8,9,10,11,12,13,14,15],xmm7[0,1,2,3,4,5,6,7]10605; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]10606; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10607; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10608; AVX2-FCP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload10609; AVX2-FCP-NEXT: # ymm0 = mem[8,9,10,11,12,13,14,15],ymm0[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm0[16,17,18,19,20,21,22,23]10610; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,3]10611; AVX2-FCP-NEXT: vmovdqa 1504(%rdi), %xmm610612; AVX2-FCP-NEXT: vpalignr {{.*#+}} xmm1 = mem[8,9,10,11,12,13,14,15],xmm6[0,1,2,3,4,5,6,7]10613; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]10614; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10615; AVX2-FCP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm4, %ymm0 # 32-byte Folded Reload10616; AVX2-FCP-NEXT: # ymm0 = mem[8,9,10,11,12,13,14,15],ymm4[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm4[16,17,18,19,20,21,22,23]10617; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,3]10618; AVX2-FCP-NEXT: vmovdqa 1664(%rdi), %xmm510619; AVX2-FCP-NEXT: vpalignr {{.*#+}} xmm1 = mem[8,9,10,11,12,13,14,15],xmm5[0,1,2,3,4,5,6,7]10620; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]10621; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10622; AVX2-FCP-NEXT: vmovdqu (%rsp), %ymm0 # 32-byte Reload10623; AVX2-FCP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload10624; AVX2-FCP-NEXT: # ymm0 = mem[8,9,10,11,12,13,14,15],ymm0[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm0[16,17,18,19,20,21,22,23]10625; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,3]10626; AVX2-FCP-NEXT: vmovdqa 1824(%rdi), %xmm410627; AVX2-FCP-NEXT: vpalignr {{.*#+}} xmm1 = mem[8,9,10,11,12,13,14,15],xmm4[0,1,2,3,4,5,6,7]10628; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]10629; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10630; AVX2-FCP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3, %ymm0 # 32-byte Folded Reload10631; AVX2-FCP-NEXT: # ymm0 = mem[8,9,10,11,12,13,14,15],ymm3[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm3[16,17,18,19,20,21,22,23]10632; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,3]10633; AVX2-FCP-NEXT: vmovdqa 1984(%rdi), %xmm310634; AVX2-FCP-NEXT: vpalignr {{.*#+}} xmm1 = mem[8,9,10,11,12,13,14,15],xmm3[0,1,2,3,4,5,6,7]10635; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]10636; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10637; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10638; AVX2-FCP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload10639; AVX2-FCP-NEXT: # ymm0 = mem[8,9,10,11,12,13,14,15],ymm0[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm0[16,17,18,19,20,21,22,23]10640; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,3]10641; AVX2-FCP-NEXT: vmovdqa 2144(%rdi), %xmm210642; AVX2-FCP-NEXT: vpalignr {{.*#+}} xmm1 = mem[8,9,10,11,12,13,14,15],xmm2[0,1,2,3,4,5,6,7]10643; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]10644; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10645; AVX2-FCP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm14, %ymm0 # 32-byte Folded Reload10646; AVX2-FCP-NEXT: # ymm0 = mem[8,9,10,11,12,13,14,15],ymm14[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm14[16,17,18,19,20,21,22,23]10647; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,1,0,3]10648; AVX2-FCP-NEXT: vmovdqa 2304(%rdi), %xmm110649; AVX2-FCP-NEXT: vpalignr {{.*#+}} xmm15 = mem[8,9,10,11,12,13,14,15],xmm1[0,1,2,3,4,5,6,7]10650; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm15[0,1,2,3],ymm0[4,5,6,7]10651; AVX2-FCP-NEXT: vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10652; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10653; AVX2-FCP-NEXT: vpalignr $8, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 32-byte Folded Reload10654; AVX2-FCP-NEXT: # ymm0 = mem[8,9,10,11,12,13,14,15],ymm0[0,1,2,3,4,5,6,7],mem[24,25,26,27,28,29,30,31],ymm0[16,17,18,19,20,21,22,23]10655; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm15 = ymm0[0,1,0,3]10656; AVX2-FCP-NEXT: vmovdqa 2464(%rdi), %xmm010657; AVX2-FCP-NEXT: vpalignr {{.*#+}} xmm14 = mem[8,9,10,11,12,13,14,15],xmm0[0,1,2,3,4,5,6,7]10658; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm14 = ymm14[0,1,2,3],ymm15[4,5,6,7]10659; AVX2-FCP-NEXT: vmovdqu %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10660; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Reload10661; AVX2-FCP-NEXT: vblendps $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm14, %ymm14 # 32-byte Folded Reload10662; AVX2-FCP-NEXT: # ymm14 = mem[0,1,2,3,4,5],ymm14[6,7]10663; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload10664; AVX2-FCP-NEXT: vblendps $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload10665; AVX2-FCP-NEXT: # xmm15 = mem[0,1],xmm15[2,3]10666; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm14 = ymm15[0,1,2,3],ymm14[4,5,6,7]10667; AVX2-FCP-NEXT: vmovups %ymm14, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10668; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Reload10669; AVX2-FCP-NEXT: vblendps $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm14, %ymm14 # 32-byte Folded Reload10670; AVX2-FCP-NEXT: # ymm14 = mem[0,1,2,3,4,5],ymm14[6,7]10671; AVX2-FCP-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 16-byte Reload10672; AVX2-FCP-NEXT: vblendps $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm15, %xmm15 # 16-byte Folded Reload10673; AVX2-FCP-NEXT: # xmm15 = mem[0,1],xmm15[2,3]10674; AVX2-FCP-NEXT: vblendps {{.*#+}} ymm15 = ymm15[0,1,2,3],ymm14[4,5,6,7]10675; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm14 # 32-byte Reload10676; AVX2-FCP-NEXT: vpblendd $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm14, %ymm14 # 32-byte Folded Reload10677; AVX2-FCP-NEXT: # ymm14 = mem[0,1,2,3,4,5],ymm14[6,7]10678; AVX2-FCP-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm13, %xmm13 # 16-byte Folded Reload10679; AVX2-FCP-NEXT: # xmm13 = mem[0,1],xmm13[2,3]10680; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm13 = ymm13[0,1,2,3],ymm14[4,5,6,7]10681; AVX2-FCP-NEXT: vmovdqu %ymm13, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10682; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm13 # 32-byte Reload10683; AVX2-FCP-NEXT: vpblendd $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm13, %ymm13 # 32-byte Folded Reload10684; AVX2-FCP-NEXT: # ymm13 = mem[0,1,2,3,4,5],ymm13[6,7]10685; AVX2-FCP-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm11, %xmm11 # 16-byte Folded Reload10686; AVX2-FCP-NEXT: # xmm11 = mem[0,1],xmm11[2,3]10687; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm11 = ymm11[0,1,2,3],ymm13[4,5,6,7]10688; AVX2-FCP-NEXT: vmovdqu %ymm11, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill10689; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload10690; AVX2-FCP-NEXT: vpblendd $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm11, %ymm11 # 32-byte Folded Reload10691; AVX2-FCP-NEXT: # ymm11 = mem[0,1,2,3,4,5],ymm11[6,7]10692; AVX2-FCP-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm12, %xmm12 # 16-byte Folded Reload10693; AVX2-FCP-NEXT: # xmm12 = mem[0,1],xmm12[2,3]10694; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm12 = ymm12[0,1,2,3],ymm11[4,5,6,7]10695; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm11 # 32-byte Reload10696; AVX2-FCP-NEXT: vpblendd $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm11, %ymm11 # 32-byte Folded Reload10697; AVX2-FCP-NEXT: # ymm11 = mem[0,1,2,3,4,5],ymm11[6,7]10698; AVX2-FCP-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm10, %xmm10 # 16-byte Folded Reload10699; AVX2-FCP-NEXT: # xmm10 = mem[0,1],xmm10[2,3]10700; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm11 = ymm10[0,1,2,3],ymm11[4,5,6,7]10701; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload10702; AVX2-FCP-NEXT: vpblendd $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm10, %ymm10 # 32-byte Folded Reload10703; AVX2-FCP-NEXT: # ymm10 = mem[0,1,2,3,4,5],ymm10[6,7]10704; AVX2-FCP-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm9, %xmm9 # 16-byte Folded Reload10705; AVX2-FCP-NEXT: # xmm9 = mem[0,1],xmm9[2,3]10706; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm9 = ymm9[0,1,2,3],ymm10[4,5,6,7]10707; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm10 # 32-byte Reload10708; AVX2-FCP-NEXT: vpblendd $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm10, %ymm10 # 32-byte Folded Reload10709; AVX2-FCP-NEXT: # ymm10 = mem[0,1,2,3,4,5],ymm10[6,7]10710; AVX2-FCP-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm8, %xmm8 # 16-byte Folded Reload10711; AVX2-FCP-NEXT: # xmm8 = mem[0,1],xmm8[2,3]10712; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm14 = ymm8[0,1,2,3],ymm10[4,5,6,7]10713; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm8 # 32-byte Reload10714; AVX2-FCP-NEXT: vpblendd $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm8, %ymm8 # 32-byte Folded Reload10715; AVX2-FCP-NEXT: # ymm8 = mem[0,1,2,3,4,5],ymm8[6,7]10716; AVX2-FCP-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm7, %xmm7 # 16-byte Folded Reload10717; AVX2-FCP-NEXT: # xmm7 = mem[0,1],xmm7[2,3]10718; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm10 = ymm7[0,1,2,3],ymm8[4,5,6,7]10719; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload10720; AVX2-FCP-NEXT: vpblendd $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm7 # 32-byte Folded Reload10721; AVX2-FCP-NEXT: # ymm7 = mem[0,1,2,3,4,5],ymm7[6,7]10722; AVX2-FCP-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm6, %xmm6 # 16-byte Folded Reload10723; AVX2-FCP-NEXT: # xmm6 = mem[0,1],xmm6[2,3]10724; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm6 = ymm6[0,1,2,3],ymm7[4,5,6,7]10725; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm7 # 32-byte Reload10726; AVX2-FCP-NEXT: vpblendd $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm7, %ymm7 # 32-byte Folded Reload10727; AVX2-FCP-NEXT: # ymm7 = mem[0,1,2,3,4,5],ymm7[6,7]10728; AVX2-FCP-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm5, %xmm5 # 16-byte Folded Reload10729; AVX2-FCP-NEXT: # xmm5 = mem[0,1],xmm5[2,3]10730; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm13 = ymm5[0,1,2,3],ymm7[4,5,6,7]10731; AVX2-FCP-NEXT: vmovdqu (%rsp), %ymm5 # 32-byte Reload10732; AVX2-FCP-NEXT: vpblendd $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm5 # 32-byte Folded Reload10733; AVX2-FCP-NEXT: # ymm5 = mem[0,1,2,3,4,5],ymm5[6,7]10734; AVX2-FCP-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm4, %xmm4 # 16-byte Folded Reload10735; AVX2-FCP-NEXT: # xmm4 = mem[0,1],xmm4[2,3]10736; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1,2,3],ymm5[4,5,6,7]10737; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm5 # 32-byte Reload10738; AVX2-FCP-NEXT: vpblendd $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm5, %ymm5 # 32-byte Folded Reload10739; AVX2-FCP-NEXT: # ymm5 = mem[0,1,2,3,4,5],ymm5[6,7]10740; AVX2-FCP-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm3, %xmm3 # 16-byte Folded Reload10741; AVX2-FCP-NEXT: # xmm3 = mem[0,1],xmm3[2,3]10742; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm7 = ymm3[0,1,2,3],ymm5[4,5,6,7]10743; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm3 # 32-byte Reload10744; AVX2-FCP-NEXT: vpblendd $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm3, %ymm3 # 32-byte Folded Reload10745; AVX2-FCP-NEXT: # ymm3 = mem[0,1,2,3,4,5],ymm3[6,7]10746; AVX2-FCP-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm2, %xmm2 # 16-byte Folded Reload10747; AVX2-FCP-NEXT: # xmm2 = mem[0,1],xmm2[2,3]10748; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm5 = ymm2[0,1,2,3],ymm3[4,5,6,7]10749; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm2 # 32-byte Reload10750; AVX2-FCP-NEXT: vpblendd $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm2, %ymm2 # 32-byte Folded Reload10751; AVX2-FCP-NEXT: # ymm2 = mem[0,1,2,3,4,5],ymm2[6,7]10752; AVX2-FCP-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload10753; AVX2-FCP-NEXT: # xmm1 = mem[0,1],xmm1[2,3]10754; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm1[0,1,2,3],ymm2[4,5,6,7]10755; AVX2-FCP-NEXT: vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10756; AVX2-FCP-NEXT: vpblendd $63, {{[-0-9]+}}(%r{{[sb]}}p), %ymm1, %ymm1 # 32-byte Folded Reload10757; AVX2-FCP-NEXT: # ymm1 = mem[0,1,2,3,4,5],ymm1[6,7]10758; AVX2-FCP-NEXT: vpblendd $3, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload10759; AVX2-FCP-NEXT: # xmm0 = mem[0,1],xmm0[2,3]10760; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]10761; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10762; AVX2-FCP-NEXT: vmovaps %ymm1, 448(%rsi)10763; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10764; AVX2-FCP-NEXT: vmovaps %ymm1, 384(%rsi)10765; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10766; AVX2-FCP-NEXT: vmovaps %ymm1, 320(%rsi)10767; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10768; AVX2-FCP-NEXT: vmovaps %ymm1, 256(%rsi)10769; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10770; AVX2-FCP-NEXT: vmovaps %ymm1, 192(%rsi)10771; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10772; AVX2-FCP-NEXT: vmovaps %ymm1, 128(%rsi)10773; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10774; AVX2-FCP-NEXT: vmovaps %ymm1, 64(%rsi)10775; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10776; AVX2-FCP-NEXT: vmovaps %ymm1, (%rsi)10777; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10778; AVX2-FCP-NEXT: vmovaps %ymm1, 480(%rsi)10779; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10780; AVX2-FCP-NEXT: vmovaps %ymm1, 416(%rsi)10781; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10782; AVX2-FCP-NEXT: vmovaps %ymm1, 352(%rsi)10783; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10784; AVX2-FCP-NEXT: vmovaps %ymm1, 288(%rsi)10785; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10786; AVX2-FCP-NEXT: vmovaps %ymm1, 224(%rsi)10787; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10788; AVX2-FCP-NEXT: vmovaps %ymm1, 160(%rsi)10789; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10790; AVX2-FCP-NEXT: vmovaps %ymm1, 96(%rsi)10791; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10792; AVX2-FCP-NEXT: vmovaps %ymm1, 32(%rsi)10793; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10794; AVX2-FCP-NEXT: vmovaps %ymm1, 448(%rdx)10795; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10796; AVX2-FCP-NEXT: vmovaps %ymm1, 384(%rdx)10797; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10798; AVX2-FCP-NEXT: vmovaps %ymm1, 320(%rdx)10799; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10800; AVX2-FCP-NEXT: vmovaps %ymm1, 256(%rdx)10801; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10802; AVX2-FCP-NEXT: vmovaps %ymm1, 192(%rdx)10803; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10804; AVX2-FCP-NEXT: vmovaps %ymm1, 128(%rdx)10805; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10806; AVX2-FCP-NEXT: vmovaps %ymm1, 64(%rdx)10807; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10808; AVX2-FCP-NEXT: vmovaps %ymm1, (%rdx)10809; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10810; AVX2-FCP-NEXT: vmovaps %ymm1, 480(%rdx)10811; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10812; AVX2-FCP-NEXT: vmovaps %ymm1, 416(%rdx)10813; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10814; AVX2-FCP-NEXT: vmovaps %ymm1, 352(%rdx)10815; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10816; AVX2-FCP-NEXT: vmovaps %ymm1, 288(%rdx)10817; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10818; AVX2-FCP-NEXT: vmovaps %ymm1, 224(%rdx)10819; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10820; AVX2-FCP-NEXT: vmovaps %ymm1, 160(%rdx)10821; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10822; AVX2-FCP-NEXT: vmovaps %ymm1, 96(%rdx)10823; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10824; AVX2-FCP-NEXT: vmovaps %ymm1, 32(%rdx)10825; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10826; AVX2-FCP-NEXT: vmovaps %ymm1, (%rcx)10827; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10828; AVX2-FCP-NEXT: vmovaps %ymm1, 64(%rcx)10829; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10830; AVX2-FCP-NEXT: vmovaps %ymm1, 128(%rcx)10831; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10832; AVX2-FCP-NEXT: vmovaps %ymm1, 192(%rcx)10833; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10834; AVX2-FCP-NEXT: vmovaps %ymm1, 256(%rcx)10835; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10836; AVX2-FCP-NEXT: vmovaps %ymm1, 320(%rcx)10837; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10838; AVX2-FCP-NEXT: vmovaps %ymm1, 384(%rcx)10839; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10840; AVX2-FCP-NEXT: vmovaps %ymm1, 448(%rcx)10841; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10842; AVX2-FCP-NEXT: vmovaps %ymm1, 480(%rcx)10843; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10844; AVX2-FCP-NEXT: vmovaps %ymm1, 416(%rcx)10845; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10846; AVX2-FCP-NEXT: vmovaps %ymm1, 352(%rcx)10847; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10848; AVX2-FCP-NEXT: vmovaps %ymm1, 288(%rcx)10849; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10850; AVX2-FCP-NEXT: vmovaps %ymm1, 224(%rcx)10851; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10852; AVX2-FCP-NEXT: vmovaps %ymm1, 160(%rcx)10853; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10854; AVX2-FCP-NEXT: vmovaps %ymm1, 96(%rcx)10855; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10856; AVX2-FCP-NEXT: vmovaps %ymm1, 32(%rcx)10857; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10858; AVX2-FCP-NEXT: vmovaps %ymm1, 480(%r8)10859; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10860; AVX2-FCP-NEXT: vmovaps %ymm1, 448(%r8)10861; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10862; AVX2-FCP-NEXT: vmovaps %ymm1, 416(%r8)10863; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10864; AVX2-FCP-NEXT: vmovaps %ymm1, 384(%r8)10865; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10866; AVX2-FCP-NEXT: vmovaps %ymm1, 352(%r8)10867; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10868; AVX2-FCP-NEXT: vmovaps %ymm1, 320(%r8)10869; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10870; AVX2-FCP-NEXT: vmovaps %ymm1, 288(%r8)10871; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10872; AVX2-FCP-NEXT: vmovaps %ymm1, 256(%r8)10873; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10874; AVX2-FCP-NEXT: vmovaps %ymm1, 224(%r8)10875; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10876; AVX2-FCP-NEXT: vmovaps %ymm1, 192(%r8)10877; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10878; AVX2-FCP-NEXT: vmovaps %ymm1, 160(%r8)10879; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10880; AVX2-FCP-NEXT: vmovaps %ymm1, 128(%r8)10881; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10882; AVX2-FCP-NEXT: vmovaps %ymm1, 96(%r8)10883; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10884; AVX2-FCP-NEXT: vmovaps %ymm1, 64(%r8)10885; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10886; AVX2-FCP-NEXT: vmovaps %ymm1, 32(%r8)10887; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm1 # 32-byte Reload10888; AVX2-FCP-NEXT: vmovaps %ymm1, (%r8)10889; AVX2-FCP-NEXT: vmovdqa %ymm0, 480(%r9)10890; AVX2-FCP-NEXT: vmovdqa %ymm2, 448(%r9)10891; AVX2-FCP-NEXT: vmovdqa %ymm5, 416(%r9)10892; AVX2-FCP-NEXT: vmovdqa %ymm7, 384(%r9)10893; AVX2-FCP-NEXT: vmovdqa %ymm4, 352(%r9)10894; AVX2-FCP-NEXT: vmovdqa %ymm13, 320(%r9)10895; AVX2-FCP-NEXT: vmovdqa %ymm6, 288(%r9)10896; AVX2-FCP-NEXT: vmovdqa %ymm10, 256(%r9)10897; AVX2-FCP-NEXT: vmovdqa %ymm14, 224(%r9)10898; AVX2-FCP-NEXT: vmovdqa %ymm9, 192(%r9)10899; AVX2-FCP-NEXT: vmovdqa %ymm11, 160(%r9)10900; AVX2-FCP-NEXT: vmovdqa %ymm12, 128(%r9)10901; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10902; AVX2-FCP-NEXT: vmovaps %ymm0, 96(%r9)10903; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10904; AVX2-FCP-NEXT: vmovaps %ymm0, 64(%r9)10905; AVX2-FCP-NEXT: vmovaps %ymm15, 32(%r9)10906; AVX2-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload10907; AVX2-FCP-NEXT: vmovaps %ymm0, (%r9)10908; AVX2-FCP-NEXT: addq $3240, %rsp # imm = 0xCA810909; AVX2-FCP-NEXT: vzeroupper10910; AVX2-FCP-NEXT: retq10911;10912; AVX512-LABEL: load_i64_stride5_vf64:10913; AVX512: # %bb.0:10914; AVX512-NEXT: subq $3336, %rsp # imm = 0xD0810915; AVX512-NEXT: vmovdqa64 1728(%rdi), %zmm310916; AVX512-NEXT: vmovdqa64 1792(%rdi), %zmm710917; AVX512-NEXT: vmovdqa64 1408(%rdi), %zmm010918; AVX512-NEXT: vmovdqa64 1088(%rdi), %zmm110919; AVX512-NEXT: vmovdqa64 1152(%rdi), %zmm610920; AVX512-NEXT: vmovdqa64 768(%rdi), %zmm510921; AVX512-NEXT: vmovdqa64 832(%rdi), %zmm810922; AVX512-NEXT: vmovdqa64 448(%rdi), %zmm410923; AVX512-NEXT: vmovdqa64 512(%rdi), %zmm910924; AVX512-NEXT: vmovdqa64 128(%rdi), %zmm210925; AVX512-NEXT: vmovdqa64 192(%rdi), %zmm1010926; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm16 = [12,1,6,0,12,1,6,0]10927; AVX512-NEXT: # zmm16 = mem[0,1,2,3,0,1,2,3]10928; AVX512-NEXT: vmovdqa64 %zmm10, %zmm1110929; AVX512-NEXT: vpermt2q %zmm2, %zmm16, %zmm1110930; AVX512-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill10931; AVX512-NEXT: vmovdqa64 %zmm9, %zmm1110932; AVX512-NEXT: vpermt2q %zmm4, %zmm16, %zmm1110933; AVX512-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill10934; AVX512-NEXT: vmovdqa64 %zmm8, %zmm1110935; AVX512-NEXT: vpermt2q %zmm5, %zmm16, %zmm1110936; AVX512-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill10937; AVX512-NEXT: vmovdqa64 %zmm6, %zmm1210938; AVX512-NEXT: vmovdqa64 %zmm6, %zmm1110939; AVX512-NEXT: vpermt2q %zmm1, %zmm16, %zmm1210940; AVX512-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill10941; AVX512-NEXT: vmovdqa64 %zmm7, %zmm610942; AVX512-NEXT: vpermt2q %zmm3, %zmm16, %zmm610943; AVX512-NEXT: vmovdqu64 %zmm6, (%rsp) # 64-byte Spill10944; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm12 = [5,10,15,0,5,10,15,0]10945; AVX512-NEXT: # zmm12 = mem[0,1,2,3,0,1,2,3]10946; AVX512-NEXT: vmovdqa64 %zmm2, %zmm610947; AVX512-NEXT: vpermt2q %zmm10, %zmm12, %zmm610948; AVX512-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill10949; AVX512-NEXT: vmovdqa64 %zmm4, %zmm610950; AVX512-NEXT: vpermt2q %zmm9, %zmm12, %zmm610951; AVX512-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill10952; AVX512-NEXT: vmovdqa64 %zmm1, %zmm610953; AVX512-NEXT: vpermt2q %zmm11, %zmm12, %zmm610954; AVX512-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill10955; AVX512-NEXT: vmovdqa64 %zmm5, %zmm610956; AVX512-NEXT: vpermt2q %zmm8, %zmm12, %zmm610957; AVX512-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill10958; AVX512-NEXT: vmovdqa64 %zmm3, %zmm610959; AVX512-NEXT: vpermt2q %zmm7, %zmm12, %zmm610960; AVX512-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill10961; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = [6,11,0,1,6,11,0,1]10962; AVX512-NEXT: # zmm6 = mem[0,1,2,3,0,1,2,3]10963; AVX512-NEXT: vmovdqa64 %zmm2, %zmm1310964; AVX512-NEXT: vpermt2q %zmm10, %zmm6, %zmm1310965; AVX512-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill10966; AVX512-NEXT: vmovdqa64 %zmm6, %zmm1310967; AVX512-NEXT: vmovdqa64 %zmm4, %zmm610968; AVX512-NEXT: vpermt2q %zmm9, %zmm13, %zmm610969; AVX512-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill10970; AVX512-NEXT: vmovdqa64 %zmm1, %zmm610971; AVX512-NEXT: vpermt2q %zmm11, %zmm13, %zmm610972; AVX512-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill10973; AVX512-NEXT: vmovdqa64 %zmm5, %zmm610974; AVX512-NEXT: vpermt2q %zmm8, %zmm13, %zmm610975; AVX512-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill10976; AVX512-NEXT: vmovdqa64 %zmm3, %zmm610977; AVX512-NEXT: vpermt2q %zmm7, %zmm13, %zmm610978; AVX512-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill10979; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm14 = [7,12,0,2,7,12,0,2]10980; AVX512-NEXT: # zmm14 = mem[0,1,2,3,0,1,2,3]10981; AVX512-NEXT: vmovdqa64 %zmm4, %zmm610982; AVX512-NEXT: vpermt2q %zmm9, %zmm14, %zmm610983; AVX512-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill10984; AVX512-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = [0,5,0,11,0,5,0,11]10985; AVX512-NEXT: # zmm6 = mem[0,1,2,3,0,1,2,3]10986; AVX512-NEXT: vpermt2q %zmm4, %zmm6, %zmm910987; AVX512-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill10988; AVX512-NEXT: vmovdqa64 %zmm2, %zmm410989; AVX512-NEXT: vpermt2q %zmm10, %zmm14, %zmm410990; AVX512-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill10991; AVX512-NEXT: vpermt2q %zmm2, %zmm6, %zmm1010992; AVX512-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill10993; AVX512-NEXT: vmovdqa64 %zmm1, %zmm210994; AVX512-NEXT: vpermt2q %zmm11, %zmm14, %zmm210995; AVX512-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill10996; AVX512-NEXT: vpermt2q %zmm1, %zmm6, %zmm1110997; AVX512-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill10998; AVX512-NEXT: vmovdqa64 %zmm5, %zmm110999; AVX512-NEXT: vpermt2q %zmm8, %zmm14, %zmm111000; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11001; AVX512-NEXT: vpermt2q %zmm5, %zmm6, %zmm811002; AVX512-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11003; AVX512-NEXT: vmovdqa64 %zmm3, %zmm111004; AVX512-NEXT: vpermt2q %zmm7, %zmm14, %zmm111005; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11006; AVX512-NEXT: vpermt2q %zmm3, %zmm6, %zmm711007; AVX512-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11008; AVX512-NEXT: vmovdqa64 1472(%rdi), %zmm211009; AVX512-NEXT: vmovdqa64 %zmm2, %zmm1911010; AVX512-NEXT: vpermt2q %zmm0, %zmm16, %zmm1911011; AVX512-NEXT: vmovdqa64 %zmm0, %zmm111012; AVX512-NEXT: vpermt2q %zmm2, %zmm12, %zmm111013; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11014; AVX512-NEXT: vmovdqa64 %zmm0, %zmm111015; AVX512-NEXT: vpermt2q %zmm2, %zmm13, %zmm111016; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11017; AVX512-NEXT: vmovdqa64 %zmm0, %zmm111018; AVX512-NEXT: vpermt2q %zmm2, %zmm14, %zmm111019; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11020; AVX512-NEXT: vpermt2q %zmm0, %zmm6, %zmm211021; AVX512-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11022; AVX512-NEXT: vmovdqa64 2048(%rdi), %zmm011023; AVX512-NEXT: vmovdqa64 2112(%rdi), %zmm211024; AVX512-NEXT: vmovdqa64 %zmm2, %zmm1811025; AVX512-NEXT: vpermt2q %zmm0, %zmm16, %zmm1811026; AVX512-NEXT: vmovdqa64 %zmm0, %zmm111027; AVX512-NEXT: vpermt2q %zmm2, %zmm12, %zmm111028; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11029; AVX512-NEXT: vmovdqa64 %zmm0, %zmm111030; AVX512-NEXT: vpermt2q %zmm2, %zmm13, %zmm111031; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11032; AVX512-NEXT: vmovdqa64 %zmm0, %zmm111033; AVX512-NEXT: vpermt2q %zmm2, %zmm14, %zmm111034; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11035; AVX512-NEXT: vpermt2q %zmm0, %zmm6, %zmm211036; AVX512-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11037; AVX512-NEXT: vmovdqa64 2368(%rdi), %zmm011038; AVX512-NEXT: vmovdqa64 2432(%rdi), %zmm111039; AVX512-NEXT: vpermi2q %zmm1, %zmm0, %zmm1211040; AVX512-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11041; AVX512-NEXT: vpermi2q %zmm1, %zmm0, %zmm1311042; AVX512-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11043; AVX512-NEXT: vpermi2q %zmm1, %zmm0, %zmm1411044; AVX512-NEXT: vmovdqu64 %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11045; AVX512-NEXT: vmovdqa64 %zmm1, %zmm1711046; AVX512-NEXT: vpermt2q %zmm0, %zmm6, %zmm111047; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11048; AVX512-NEXT: vpermt2q %zmm0, %zmm16, %zmm1711049; AVX512-NEXT: vmovdqa64 384(%rdi), %zmm3111050; AVX512-NEXT: vmovdqa64 320(%rdi), %zmm211051; AVX512-NEXT: vpmovsxbq {{.*#+}} ymm3 = [0,5,10,15]11052; AVX512-NEXT: vmovdqa64 %zmm2, %zmm111053; AVX512-NEXT: vpermt2q %zmm31, %zmm3, %zmm111054; AVX512-NEXT: vpmovsxbq {{.*#+}} ymm29 = [1,6,11,0]11055; AVX512-NEXT: vmovdqa64 %zmm2, %zmm011056; AVX512-NEXT: vpermt2q %zmm31, %zmm29, %zmm011057; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11058; AVX512-NEXT: vpmovsxbq {{.*#+}} ymm6 = [2,7,12,0]11059; AVX512-NEXT: vmovdqa64 %zmm2, %zmm011060; AVX512-NEXT: vpermt2q %zmm31, %zmm6, %zmm011061; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11062; AVX512-NEXT: vpmovsxbq {{.*#+}} ymm7 = [11,0,5,0]11063; AVX512-NEXT: vmovdqa64 %zmm31, %zmm011064; AVX512-NEXT: vpermt2q %zmm2, %zmm7, %zmm011065; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11066; AVX512-NEXT: vpermt2q %zmm2, %zmm16, %zmm3111067; AVX512-NEXT: vmovdqa64 (%rdi), %zmm211068; AVX512-NEXT: vmovdqa64 64(%rdi), %zmm2311069; AVX512-NEXT: vmovdqa64 %zmm2, %zmm411070; AVX512-NEXT: vpermt2q %zmm23, %zmm3, %zmm411071; AVX512-NEXT: vmovdqa64 %zmm2, %zmm011072; AVX512-NEXT: vpermt2q %zmm23, %zmm29, %zmm011073; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11074; AVX512-NEXT: vmovdqa64 %zmm2, %zmm011075; AVX512-NEXT: vpermt2q %zmm23, %zmm6, %zmm011076; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11077; AVX512-NEXT: vmovdqa64 %zmm23, %zmm011078; AVX512-NEXT: vpermt2q %zmm2, %zmm7, %zmm011079; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11080; AVX512-NEXT: vpermt2q %zmm2, %zmm16, %zmm2311081; AVX512-NEXT: vmovdqa64 1024(%rdi), %zmm2111082; AVX512-NEXT: vmovdqa64 960(%rdi), %zmm211083; AVX512-NEXT: vmovdqa64 %zmm2, %zmm1011084; AVX512-NEXT: vpermt2q %zmm21, %zmm3, %zmm1011085; AVX512-NEXT: vmovdqa64 %zmm2, %zmm011086; AVX512-NEXT: vpermt2q %zmm21, %zmm29, %zmm011087; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11088; AVX512-NEXT: vmovdqa64 %zmm2, %zmm011089; AVX512-NEXT: vpermt2q %zmm21, %zmm6, %zmm011090; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11091; AVX512-NEXT: vmovdqa64 %zmm21, %zmm011092; AVX512-NEXT: vpermt2q %zmm2, %zmm7, %zmm011093; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11094; AVX512-NEXT: vpermt2q %zmm2, %zmm16, %zmm2111095; AVX512-NEXT: vmovdqa64 704(%rdi), %zmm2011096; AVX512-NEXT: vmovdqa64 640(%rdi), %zmm211097; AVX512-NEXT: vmovdqa64 %zmm2, %zmm1111098; AVX512-NEXT: vpermt2q %zmm20, %zmm3, %zmm1111099; AVX512-NEXT: vmovdqa64 %zmm2, %zmm011100; AVX512-NEXT: vpermt2q %zmm20, %zmm29, %zmm011101; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11102; AVX512-NEXT: vmovdqa64 %zmm2, %zmm011103; AVX512-NEXT: vpermt2q %zmm20, %zmm6, %zmm011104; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11105; AVX512-NEXT: vmovdqa64 %zmm20, %zmm011106; AVX512-NEXT: vpermt2q %zmm2, %zmm7, %zmm011107; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11108; AVX512-NEXT: vpermt2q %zmm2, %zmm16, %zmm2011109; AVX512-NEXT: vmovdqa64 1664(%rdi), %zmm2511110; AVX512-NEXT: vmovdqa64 1600(%rdi), %zmm211111; AVX512-NEXT: vmovdqa64 %zmm2, %zmm1311112; AVX512-NEXT: vpermt2q %zmm25, %zmm3, %zmm1311113; AVX512-NEXT: vmovdqa64 %zmm2, %zmm011114; AVX512-NEXT: vpermt2q %zmm25, %zmm29, %zmm011115; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11116; AVX512-NEXT: vmovdqa64 %zmm2, %zmm3011117; AVX512-NEXT: vpermt2q %zmm25, %zmm6, %zmm3011118; AVX512-NEXT: vmovdqa64 %zmm25, %zmm011119; AVX512-NEXT: vpermt2q %zmm2, %zmm7, %zmm011120; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11121; AVX512-NEXT: vpermt2q %zmm2, %zmm16, %zmm2511122; AVX512-NEXT: vmovdqa64 1344(%rdi), %zmm1211123; AVX512-NEXT: vmovdqa64 1280(%rdi), %zmm211124; AVX512-NEXT: vmovdqa64 %zmm2, %zmm1411125; AVX512-NEXT: vpermt2q %zmm12, %zmm3, %zmm1411126; AVX512-NEXT: vmovdqa64 %zmm2, %zmm2611127; AVX512-NEXT: vpermt2q %zmm12, %zmm29, %zmm2611128; AVX512-NEXT: vmovdqa64 %zmm2, %zmm2711129; AVX512-NEXT: vpermt2q %zmm12, %zmm6, %zmm2711130; AVX512-NEXT: vmovdqa64 %zmm12, %zmm2811131; AVX512-NEXT: vpermt2q %zmm2, %zmm7, %zmm2811132; AVX512-NEXT: vpermt2q %zmm2, %zmm16, %zmm1211133; AVX512-NEXT: vmovdqa64 1984(%rdi), %zmm911134; AVX512-NEXT: vmovdqa64 1920(%rdi), %zmm1511135; AVX512-NEXT: vmovdqa64 %zmm15, %zmm211136; AVX512-NEXT: vpermt2q %zmm9, %zmm3, %zmm211137; AVX512-NEXT: vmovdqa64 2304(%rdi), %zmm511138; AVX512-NEXT: vmovdqa64 2240(%rdi), %zmm011139; AVX512-NEXT: vpermi2q %zmm5, %zmm0, %zmm311140; AVX512-NEXT: vmovdqa64 %zmm15, %zmm811141; AVX512-NEXT: vpermt2q %zmm9, %zmm29, %zmm811142; AVX512-NEXT: vpermi2q %zmm5, %zmm0, %zmm2911143; AVX512-NEXT: vmovdqa64 %zmm15, %zmm2211144; AVX512-NEXT: vpermt2q %zmm9, %zmm6, %zmm2211145; AVX512-NEXT: vpermi2q %zmm5, %zmm0, %zmm611146; AVX512-NEXT: vmovdqa64 %zmm9, %zmm2411147; AVX512-NEXT: vpermt2q %zmm15, %zmm7, %zmm2411148; AVX512-NEXT: vpermi2q %zmm0, %zmm5, %zmm711149; AVX512-NEXT: vpermt2q %zmm0, %zmm16, %zmm511150; AVX512-NEXT: vpermt2q %zmm15, %zmm16, %zmm911151; AVX512-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm0 # 64-byte Folded Reload11152; AVX512-NEXT: # zmm0 = zmm4[0,1,2,3],mem[4,5,6,7]11153; AVX512-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm4 # 64-byte Folded Reload11154; AVX512-NEXT: # zmm4 = zmm1[0,1,2,3],mem[4,5,6,7]11155; AVX512-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm11, %zmm11 # 64-byte Folded Reload11156; AVX512-NEXT: # zmm11 = zmm11[0,1,2,3],mem[4,5,6,7]11157; AVX512-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm10, %zmm10 # 64-byte Folded Reload11158; AVX512-NEXT: # zmm10 = zmm10[0,1,2,3],mem[4,5,6,7]11159; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm19 = zmm14[0,1,2,3],zmm19[4,5,6,7]11160; AVX512-NEXT: vshufi64x2 $228, (%rsp), %zmm13, %zmm1 # 64-byte Folded Reload11161; AVX512-NEXT: # zmm1 = zmm13[0,1,2,3],mem[4,5,6,7]11162; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm2[0,1,2,3],zmm18[4,5,6,7]11163; AVX512-NEXT: vshufi64x2 {{.*#+}} zmm3 = zmm3[0,1,2,3],zmm17[4,5,6,7]11164; AVX512-NEXT: vmovdqa64 256(%rdi), %zmm1811165; AVX512-NEXT: vpmovsxbq {{.*#+}} zmm15 = [0,1,2,3,4,5,6,11]11166; AVX512-NEXT: vpermt2q %zmm18, %zmm15, %zmm011167; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11168; AVX512-NEXT: vmovdqa64 576(%rdi), %zmm1411169; AVX512-NEXT: vpermt2q %zmm14, %zmm15, %zmm411170; AVX512-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11171; AVX512-NEXT: vmovdqa64 896(%rdi), %zmm1711172; AVX512-NEXT: vpermt2q %zmm17, %zmm15, %zmm1111173; AVX512-NEXT: vmovdqu64 %zmm11, (%rsp) # 64-byte Spill11174; AVX512-NEXT: vmovdqa64 1216(%rdi), %zmm1311175; AVX512-NEXT: vpermt2q %zmm13, %zmm15, %zmm1011176; AVX512-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11177; AVX512-NEXT: vmovdqa64 1536(%rdi), %zmm1611178; AVX512-NEXT: vpermt2q %zmm16, %zmm15, %zmm1911179; AVX512-NEXT: vmovdqu64 %zmm19, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11180; AVX512-NEXT: vmovdqa64 1856(%rdi), %zmm1111181; AVX512-NEXT: vpermt2q %zmm11, %zmm15, %zmm111182; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11183; AVX512-NEXT: vmovdqa64 2176(%rdi), %zmm1911184; AVX512-NEXT: vpermt2q %zmm19, %zmm15, %zmm211185; AVX512-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11186; AVX512-NEXT: vmovdqa64 2496(%rdi), %zmm1011187; AVX512-NEXT: vpermt2q %zmm10, %zmm15, %zmm311188; AVX512-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11189; AVX512-NEXT: movb $7, %al11190; AVX512-NEXT: kmovw %eax, %k111191; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload11192; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload11193; AVX512-NEXT: vmovdqa64 %zmm0, %zmm15 {%k1}11194; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload11195; AVX512-NEXT: vmovdqa64 %zmm8, %zmm4 {%k1}11196; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload11197; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload11198; AVX512-NEXT: vmovdqa64 %zmm0, %zmm8 {%k1}11199; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload11200; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload11201; AVX512-NEXT: vmovdqa64 %zmm1, %zmm0 {%k1}11202; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload11203; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11204; AVX512-NEXT: vmovdqa64 %zmm2, %zmm1 {%k1}11205; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11206; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload11207; AVX512-NEXT: vmovdqa64 %zmm3, %zmm2 {%k1}11208; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload11209; AVX512-NEXT: vmovdqa64 %zmm26, %zmm3 {%k1}11210; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload11211; AVX512-NEXT: vmovdqa64 %zmm29, %zmm26 {%k1}11212; AVX512-NEXT: vpmovsxbq {{.*#+}} zmm29 = [0,1,2,3,4,5,6,12]11213; AVX512-NEXT: vpermt2q %zmm18, %zmm29, %zmm1511214; AVX512-NEXT: vmovdqu64 %zmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11215; AVX512-NEXT: vpermt2q %zmm19, %zmm29, %zmm411216; AVX512-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11217; AVX512-NEXT: vpermt2q %zmm14, %zmm29, %zmm811218; AVX512-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11219; AVX512-NEXT: vpermt2q %zmm13, %zmm29, %zmm011220; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11221; AVX512-NEXT: vpermt2q %zmm17, %zmm29, %zmm111222; AVX512-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11223; AVX512-NEXT: vpermt2q %zmm11, %zmm29, %zmm211224; AVX512-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11225; AVX512-NEXT: vpermt2q %zmm16, %zmm29, %zmm311226; AVX512-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11227; AVX512-NEXT: vpermt2q %zmm10, %zmm29, %zmm2611228; AVX512-NEXT: vmovdqu64 %zmm26, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11229; AVX512-NEXT: movb $56, %al11230; AVX512-NEXT: kmovw %eax, %k111231; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11232; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload11233; AVX512-NEXT: vmovdqa64 %zmm2, %zmm0 {%k1}11234; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload11235; AVX512-NEXT: vmovdqa64 %zmm1, %zmm22 {%k1}11236; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11237; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload11238; AVX512-NEXT: vmovdqa64 %zmm2, %zmm8 {%k1}11239; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11240; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload11241; AVX512-NEXT: vmovdqa64 %zmm2, %zmm26 {%k1}11242; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11243; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload11244; AVX512-NEXT: vmovdqa64 %zmm2, %zmm15 {%k1}11245; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11246; AVX512-NEXT: vmovdqa64 %zmm2, %zmm30 {%k1}11247; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload11248; AVX512-NEXT: vmovdqa64 %zmm1, %zmm27 {%k1}11249; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload11250; AVX512-NEXT: vmovdqa64 %zmm1, %zmm6 {%k1}11251; AVX512-NEXT: vpmovsxbq {{.*#+}} zmm1 = [0,1,2,3,4,5,8,13]11252; AVX512-NEXT: vpermt2q %zmm18, %zmm1, %zmm011253; AVX512-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11254; AVX512-NEXT: vpermt2q %zmm19, %zmm1, %zmm2211255; AVX512-NEXT: vpermt2q %zmm14, %zmm1, %zmm811256; AVX512-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11257; AVX512-NEXT: vpermt2q %zmm13, %zmm1, %zmm2611258; AVX512-NEXT: vpermt2q %zmm17, %zmm1, %zmm1511259; AVX512-NEXT: vpermt2q %zmm11, %zmm1, %zmm3011260; AVX512-NEXT: vpermt2q %zmm16, %zmm1, %zmm2711261; AVX512-NEXT: vpermt2q %zmm10, %zmm1, %zmm611262; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11263; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload11264; AVX512-NEXT: vmovdqa64 %zmm2, %zmm8 {%k1}11265; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload11266; AVX512-NEXT: vmovdqa64 %zmm1, %zmm24 {%k1}11267; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11268; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 # 64-byte Reload11269; AVX512-NEXT: vmovdqa64 %zmm2, %zmm29 {%k1}11270; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11271; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload11272; AVX512-NEXT: vmovdqa64 %zmm2, %zmm1 {%k1}11273; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11274; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload11275; AVX512-NEXT: vmovdqa64 %zmm2, %zmm3 {%k1}11276; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11277; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload11278; AVX512-NEXT: vmovdqa64 %zmm2, %zmm4 {%k1}11279; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11280; AVX512-NEXT: vmovdqa64 %zmm2, %zmm28 {%k1}11281; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11282; AVX512-NEXT: vmovdqa64 %zmm2, %zmm7 {%k1}11283; AVX512-NEXT: vpmovsxbq {{.*#+}} zmm2 = [0,1,2,3,4,5,9,14]11284; AVX512-NEXT: vpermt2q %zmm18, %zmm2, %zmm811285; AVX512-NEXT: vpermt2q %zmm19, %zmm2, %zmm2411286; AVX512-NEXT: vpermt2q %zmm14, %zmm2, %zmm2911287; AVX512-NEXT: vpermt2q %zmm13, %zmm2, %zmm111288; AVX512-NEXT: vpermt2q %zmm17, %zmm2, %zmm311289; AVX512-NEXT: vmovdqa64 %zmm3, %zmm011290; AVX512-NEXT: vpermt2q %zmm11, %zmm2, %zmm411291; AVX512-NEXT: vpermt2q %zmm16, %zmm2, %zmm2811292; AVX512-NEXT: vpermt2q %zmm10, %zmm2, %zmm711293; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11294; AVX512-NEXT: vmovdqa64 %zmm2, %zmm31 {%k1}11295; AVX512-NEXT: vpmovsxbq {{.*#+}} zmm3 = [0,1,2,3,4,5,10,15]11296; AVX512-NEXT: vpermt2q %zmm14, %zmm3, %zmm3111297; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11298; AVX512-NEXT: vmovdqa64 %zmm2, %zmm23 {%k1}11299; AVX512-NEXT: vpermt2q %zmm18, %zmm3, %zmm2311300; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11301; AVX512-NEXT: vmovdqa64 %zmm2, %zmm21 {%k1}11302; AVX512-NEXT: vpermt2q %zmm13, %zmm3, %zmm2111303; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11304; AVX512-NEXT: vmovdqa64 %zmm2, %zmm20 {%k1}11305; AVX512-NEXT: vpermt2q %zmm17, %zmm3, %zmm2011306; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11307; AVX512-NEXT: vmovdqa64 %zmm2, %zmm25 {%k1}11308; AVX512-NEXT: vpermt2q %zmm11, %zmm3, %zmm2511309; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11310; AVX512-NEXT: vmovdqa64 %zmm2, %zmm12 {%k1}11311; AVX512-NEXT: vpermt2q %zmm16, %zmm3, %zmm1211312; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11313; AVX512-NEXT: vmovdqa64 %zmm2, %zmm5 {%k1}11314; AVX512-NEXT: vpermt2q %zmm10, %zmm3, %zmm511315; AVX512-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11316; AVX512-NEXT: vmovdqa64 %zmm2, %zmm9 {%k1}11317; AVX512-NEXT: vpermt2q %zmm19, %zmm3, %zmm911318; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11319; AVX512-NEXT: vmovaps %zmm2, 448(%rsi)11320; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11321; AVX512-NEXT: vmovaps %zmm2, 384(%rsi)11322; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11323; AVX512-NEXT: vmovaps %zmm2, 320(%rsi)11324; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11325; AVX512-NEXT: vmovaps %zmm2, 256(%rsi)11326; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11327; AVX512-NEXT: vmovaps %zmm2, 192(%rsi)11328; AVX512-NEXT: vmovups (%rsp), %zmm2 # 64-byte Reload11329; AVX512-NEXT: vmovaps %zmm2, 128(%rsi)11330; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11331; AVX512-NEXT: vmovaps %zmm2, 64(%rsi)11332; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11333; AVX512-NEXT: vmovaps %zmm2, (%rsi)11334; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11335; AVX512-NEXT: vmovaps %zmm2, 448(%rdx)11336; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11337; AVX512-NEXT: vmovaps %zmm2, 256(%rdx)11338; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11339; AVX512-NEXT: vmovaps %zmm2, 320(%rdx)11340; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11341; AVX512-NEXT: vmovaps %zmm2, 128(%rdx)11342; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11343; AVX512-NEXT: vmovaps %zmm2, 192(%rdx)11344; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11345; AVX512-NEXT: vmovaps %zmm2, (%rdx)11346; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11347; AVX512-NEXT: vmovaps %zmm2, 64(%rdx)11348; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11349; AVX512-NEXT: vmovaps %zmm2, 384(%rdx)11350; AVX512-NEXT: vmovdqa64 %zmm6, 448(%rcx)11351; AVX512-NEXT: vmovdqa64 %zmm27, 256(%rcx)11352; AVX512-NEXT: vmovdqa64 %zmm30, 320(%rcx)11353; AVX512-NEXT: vmovdqa64 %zmm15, 128(%rcx)11354; AVX512-NEXT: vmovdqa64 %zmm26, 192(%rcx)11355; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11356; AVX512-NEXT: vmovaps %zmm2, (%rcx)11357; AVX512-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11358; AVX512-NEXT: vmovaps %zmm2, 64(%rcx)11359; AVX512-NEXT: vmovdqa64 %zmm22, 384(%rcx)11360; AVX512-NEXT: vmovdqa64 %zmm7, 448(%r8)11361; AVX512-NEXT: vmovdqa64 %zmm28, 256(%r8)11362; AVX512-NEXT: vmovdqa64 %zmm4, 320(%r8)11363; AVX512-NEXT: vmovdqa64 %zmm0, 128(%r8)11364; AVX512-NEXT: vmovdqa64 %zmm1, 192(%r8)11365; AVX512-NEXT: vmovdqa64 %zmm8, (%r8)11366; AVX512-NEXT: vmovdqa64 %zmm29, 64(%r8)11367; AVX512-NEXT: vmovdqa64 %zmm24, 384(%r8)11368; AVX512-NEXT: vmovdqa64 %zmm9, 384(%r9)11369; AVX512-NEXT: vmovdqa64 %zmm5, 448(%r9)11370; AVX512-NEXT: vmovdqa64 %zmm12, 256(%r9)11371; AVX512-NEXT: vmovdqa64 %zmm25, 320(%r9)11372; AVX512-NEXT: vmovdqa64 %zmm20, 128(%r9)11373; AVX512-NEXT: vmovdqa64 %zmm21, 192(%r9)11374; AVX512-NEXT: vmovdqa64 %zmm23, (%r9)11375; AVX512-NEXT: vmovdqa64 %zmm31, 64(%r9)11376; AVX512-NEXT: addq $3336, %rsp # imm = 0xD0811377; AVX512-NEXT: vzeroupper11378; AVX512-NEXT: retq11379;11380; AVX512-FCP-LABEL: load_i64_stride5_vf64:11381; AVX512-FCP: # %bb.0:11382; AVX512-FCP-NEXT: subq $3336, %rsp # imm = 0xD0811383; AVX512-FCP-NEXT: vmovdqa64 1728(%rdi), %zmm311384; AVX512-FCP-NEXT: vmovdqa64 1792(%rdi), %zmm711385; AVX512-FCP-NEXT: vmovdqa64 1408(%rdi), %zmm011386; AVX512-FCP-NEXT: vmovdqa64 1088(%rdi), %zmm111387; AVX512-FCP-NEXT: vmovdqa64 1152(%rdi), %zmm611388; AVX512-FCP-NEXT: vmovdqa64 768(%rdi), %zmm511389; AVX512-FCP-NEXT: vmovdqa64 832(%rdi), %zmm811390; AVX512-FCP-NEXT: vmovdqa64 448(%rdi), %zmm411391; AVX512-FCP-NEXT: vmovdqa64 512(%rdi), %zmm911392; AVX512-FCP-NEXT: vmovdqa64 128(%rdi), %zmm211393; AVX512-FCP-NEXT: vmovdqa64 192(%rdi), %zmm1011394; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm16 = [12,1,6,0,12,1,6,0]11395; AVX512-FCP-NEXT: # zmm16 = mem[0,1,2,3,0,1,2,3]11396; AVX512-FCP-NEXT: vmovdqa64 %zmm10, %zmm1111397; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm16, %zmm1111398; AVX512-FCP-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11399; AVX512-FCP-NEXT: vmovdqa64 %zmm9, %zmm1111400; AVX512-FCP-NEXT: vpermt2q %zmm4, %zmm16, %zmm1111401; AVX512-FCP-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11402; AVX512-FCP-NEXT: vmovdqa64 %zmm8, %zmm1111403; AVX512-FCP-NEXT: vpermt2q %zmm5, %zmm16, %zmm1111404; AVX512-FCP-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11405; AVX512-FCP-NEXT: vmovdqa64 %zmm6, %zmm1211406; AVX512-FCP-NEXT: vmovdqa64 %zmm6, %zmm1111407; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm16, %zmm1211408; AVX512-FCP-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11409; AVX512-FCP-NEXT: vmovdqa64 %zmm7, %zmm611410; AVX512-FCP-NEXT: vpermt2q %zmm3, %zmm16, %zmm611411; AVX512-FCP-NEXT: vmovdqu64 %zmm6, (%rsp) # 64-byte Spill11412; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm12 = [5,10,15,0,5,10,15,0]11413; AVX512-FCP-NEXT: # zmm12 = mem[0,1,2,3,0,1,2,3]11414; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm611415; AVX512-FCP-NEXT: vpermt2q %zmm10, %zmm12, %zmm611416; AVX512-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11417; AVX512-FCP-NEXT: vmovdqa64 %zmm4, %zmm611418; AVX512-FCP-NEXT: vpermt2q %zmm9, %zmm12, %zmm611419; AVX512-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11420; AVX512-FCP-NEXT: vmovdqa64 %zmm1, %zmm611421; AVX512-FCP-NEXT: vpermt2q %zmm11, %zmm12, %zmm611422; AVX512-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11423; AVX512-FCP-NEXT: vmovdqa64 %zmm5, %zmm611424; AVX512-FCP-NEXT: vpermt2q %zmm8, %zmm12, %zmm611425; AVX512-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11426; AVX512-FCP-NEXT: vmovdqa64 %zmm3, %zmm611427; AVX512-FCP-NEXT: vpermt2q %zmm7, %zmm12, %zmm611428; AVX512-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11429; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = [6,11,0,1,6,11,0,1]11430; AVX512-FCP-NEXT: # zmm6 = mem[0,1,2,3,0,1,2,3]11431; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm1311432; AVX512-FCP-NEXT: vpermt2q %zmm10, %zmm6, %zmm1311433; AVX512-FCP-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11434; AVX512-FCP-NEXT: vmovdqa64 %zmm6, %zmm1311435; AVX512-FCP-NEXT: vmovdqa64 %zmm4, %zmm611436; AVX512-FCP-NEXT: vpermt2q %zmm9, %zmm13, %zmm611437; AVX512-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11438; AVX512-FCP-NEXT: vmovdqa64 %zmm1, %zmm611439; AVX512-FCP-NEXT: vpermt2q %zmm11, %zmm13, %zmm611440; AVX512-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11441; AVX512-FCP-NEXT: vmovdqa64 %zmm5, %zmm611442; AVX512-FCP-NEXT: vpermt2q %zmm8, %zmm13, %zmm611443; AVX512-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11444; AVX512-FCP-NEXT: vmovdqa64 %zmm3, %zmm611445; AVX512-FCP-NEXT: vpermt2q %zmm7, %zmm13, %zmm611446; AVX512-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11447; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm14 = [7,12,0,2,7,12,0,2]11448; AVX512-FCP-NEXT: # zmm14 = mem[0,1,2,3,0,1,2,3]11449; AVX512-FCP-NEXT: vmovdqa64 %zmm4, %zmm611450; AVX512-FCP-NEXT: vpermt2q %zmm9, %zmm14, %zmm611451; AVX512-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11452; AVX512-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = [0,5,0,11,0,5,0,11]11453; AVX512-FCP-NEXT: # zmm6 = mem[0,1,2,3,0,1,2,3]11454; AVX512-FCP-NEXT: vpermt2q %zmm4, %zmm6, %zmm911455; AVX512-FCP-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11456; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm411457; AVX512-FCP-NEXT: vpermt2q %zmm10, %zmm14, %zmm411458; AVX512-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11459; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm6, %zmm1011460; AVX512-FCP-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11461; AVX512-FCP-NEXT: vmovdqa64 %zmm1, %zmm211462; AVX512-FCP-NEXT: vpermt2q %zmm11, %zmm14, %zmm211463; AVX512-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11464; AVX512-FCP-NEXT: vpermt2q %zmm1, %zmm6, %zmm1111465; AVX512-FCP-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11466; AVX512-FCP-NEXT: vmovdqa64 %zmm5, %zmm111467; AVX512-FCP-NEXT: vpermt2q %zmm8, %zmm14, %zmm111468; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11469; AVX512-FCP-NEXT: vpermt2q %zmm5, %zmm6, %zmm811470; AVX512-FCP-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11471; AVX512-FCP-NEXT: vmovdqa64 %zmm3, %zmm111472; AVX512-FCP-NEXT: vpermt2q %zmm7, %zmm14, %zmm111473; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11474; AVX512-FCP-NEXT: vpermt2q %zmm3, %zmm6, %zmm711475; AVX512-FCP-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11476; AVX512-FCP-NEXT: vmovdqa64 1472(%rdi), %zmm211477; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm1911478; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm16, %zmm1911479; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm111480; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm12, %zmm111481; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11482; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm111483; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm13, %zmm111484; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11485; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm111486; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm14, %zmm111487; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11488; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm6, %zmm211489; AVX512-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11490; AVX512-FCP-NEXT: vmovdqa64 2048(%rdi), %zmm011491; AVX512-FCP-NEXT: vmovdqa64 2112(%rdi), %zmm211492; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm1811493; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm16, %zmm1811494; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm111495; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm12, %zmm111496; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11497; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm111498; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm13, %zmm111499; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11500; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm111501; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm14, %zmm111502; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11503; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm6, %zmm211504; AVX512-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11505; AVX512-FCP-NEXT: vmovdqa64 2368(%rdi), %zmm011506; AVX512-FCP-NEXT: vmovdqa64 2432(%rdi), %zmm111507; AVX512-FCP-NEXT: vpermi2q %zmm1, %zmm0, %zmm1211508; AVX512-FCP-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11509; AVX512-FCP-NEXT: vpermi2q %zmm1, %zmm0, %zmm1311510; AVX512-FCP-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11511; AVX512-FCP-NEXT: vpermi2q %zmm1, %zmm0, %zmm1411512; AVX512-FCP-NEXT: vmovdqu64 %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11513; AVX512-FCP-NEXT: vmovdqa64 %zmm1, %zmm1711514; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm6, %zmm111515; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11516; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm16, %zmm1711517; AVX512-FCP-NEXT: vmovdqa64 384(%rdi), %zmm3111518; AVX512-FCP-NEXT: vmovdqa64 320(%rdi), %zmm211519; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} ymm3 = [0,5,10,15]11520; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm111521; AVX512-FCP-NEXT: vpermt2q %zmm31, %zmm3, %zmm111522; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} ymm29 = [1,6,11,0]11523; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm011524; AVX512-FCP-NEXT: vpermt2q %zmm31, %zmm29, %zmm011525; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11526; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} ymm6 = [2,7,12,0]11527; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm011528; AVX512-FCP-NEXT: vpermt2q %zmm31, %zmm6, %zmm011529; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11530; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} ymm7 = [11,0,5,0]11531; AVX512-FCP-NEXT: vmovdqa64 %zmm31, %zmm011532; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm7, %zmm011533; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11534; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm16, %zmm3111535; AVX512-FCP-NEXT: vmovdqa64 (%rdi), %zmm211536; AVX512-FCP-NEXT: vmovdqa64 64(%rdi), %zmm2311537; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm411538; AVX512-FCP-NEXT: vpermt2q %zmm23, %zmm3, %zmm411539; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm011540; AVX512-FCP-NEXT: vpermt2q %zmm23, %zmm29, %zmm011541; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11542; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm011543; AVX512-FCP-NEXT: vpermt2q %zmm23, %zmm6, %zmm011544; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11545; AVX512-FCP-NEXT: vmovdqa64 %zmm23, %zmm011546; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm7, %zmm011547; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11548; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm16, %zmm2311549; AVX512-FCP-NEXT: vmovdqa64 1024(%rdi), %zmm2111550; AVX512-FCP-NEXT: vmovdqa64 960(%rdi), %zmm211551; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm1011552; AVX512-FCP-NEXT: vpermt2q %zmm21, %zmm3, %zmm1011553; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm011554; AVX512-FCP-NEXT: vpermt2q %zmm21, %zmm29, %zmm011555; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11556; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm011557; AVX512-FCP-NEXT: vpermt2q %zmm21, %zmm6, %zmm011558; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11559; AVX512-FCP-NEXT: vmovdqa64 %zmm21, %zmm011560; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm7, %zmm011561; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11562; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm16, %zmm2111563; AVX512-FCP-NEXT: vmovdqa64 704(%rdi), %zmm2011564; AVX512-FCP-NEXT: vmovdqa64 640(%rdi), %zmm211565; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm1111566; AVX512-FCP-NEXT: vpermt2q %zmm20, %zmm3, %zmm1111567; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm011568; AVX512-FCP-NEXT: vpermt2q %zmm20, %zmm29, %zmm011569; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11570; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm011571; AVX512-FCP-NEXT: vpermt2q %zmm20, %zmm6, %zmm011572; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11573; AVX512-FCP-NEXT: vmovdqa64 %zmm20, %zmm011574; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm7, %zmm011575; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11576; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm16, %zmm2011577; AVX512-FCP-NEXT: vmovdqa64 1664(%rdi), %zmm2511578; AVX512-FCP-NEXT: vmovdqa64 1600(%rdi), %zmm211579; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm1311580; AVX512-FCP-NEXT: vpermt2q %zmm25, %zmm3, %zmm1311581; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm011582; AVX512-FCP-NEXT: vpermt2q %zmm25, %zmm29, %zmm011583; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11584; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm3011585; AVX512-FCP-NEXT: vpermt2q %zmm25, %zmm6, %zmm3011586; AVX512-FCP-NEXT: vmovdqa64 %zmm25, %zmm011587; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm7, %zmm011588; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11589; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm16, %zmm2511590; AVX512-FCP-NEXT: vmovdqa64 1344(%rdi), %zmm1211591; AVX512-FCP-NEXT: vmovdqa64 1280(%rdi), %zmm211592; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm1411593; AVX512-FCP-NEXT: vpermt2q %zmm12, %zmm3, %zmm1411594; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm2611595; AVX512-FCP-NEXT: vpermt2q %zmm12, %zmm29, %zmm2611596; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm2711597; AVX512-FCP-NEXT: vpermt2q %zmm12, %zmm6, %zmm2711598; AVX512-FCP-NEXT: vmovdqa64 %zmm12, %zmm2811599; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm7, %zmm2811600; AVX512-FCP-NEXT: vpermt2q %zmm2, %zmm16, %zmm1211601; AVX512-FCP-NEXT: vmovdqa64 1984(%rdi), %zmm911602; AVX512-FCP-NEXT: vmovdqa64 1920(%rdi), %zmm1511603; AVX512-FCP-NEXT: vmovdqa64 %zmm15, %zmm211604; AVX512-FCP-NEXT: vpermt2q %zmm9, %zmm3, %zmm211605; AVX512-FCP-NEXT: vmovdqa64 2304(%rdi), %zmm511606; AVX512-FCP-NEXT: vmovdqa64 2240(%rdi), %zmm011607; AVX512-FCP-NEXT: vpermi2q %zmm5, %zmm0, %zmm311608; AVX512-FCP-NEXT: vmovdqa64 %zmm15, %zmm811609; AVX512-FCP-NEXT: vpermt2q %zmm9, %zmm29, %zmm811610; AVX512-FCP-NEXT: vpermi2q %zmm5, %zmm0, %zmm2911611; AVX512-FCP-NEXT: vmovdqa64 %zmm15, %zmm2211612; AVX512-FCP-NEXT: vpermt2q %zmm9, %zmm6, %zmm2211613; AVX512-FCP-NEXT: vpermi2q %zmm5, %zmm0, %zmm611614; AVX512-FCP-NEXT: vmovdqa64 %zmm9, %zmm2411615; AVX512-FCP-NEXT: vpermt2q %zmm15, %zmm7, %zmm2411616; AVX512-FCP-NEXT: vpermi2q %zmm0, %zmm5, %zmm711617; AVX512-FCP-NEXT: vpermt2q %zmm0, %zmm16, %zmm511618; AVX512-FCP-NEXT: vpermt2q %zmm15, %zmm16, %zmm911619; AVX512-FCP-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm0 # 64-byte Folded Reload11620; AVX512-FCP-NEXT: # zmm0 = zmm4[0,1,2,3],mem[4,5,6,7]11621; AVX512-FCP-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm4 # 64-byte Folded Reload11622; AVX512-FCP-NEXT: # zmm4 = zmm1[0,1,2,3],mem[4,5,6,7]11623; AVX512-FCP-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm11, %zmm11 # 64-byte Folded Reload11624; AVX512-FCP-NEXT: # zmm11 = zmm11[0,1,2,3],mem[4,5,6,7]11625; AVX512-FCP-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm10, %zmm10 # 64-byte Folded Reload11626; AVX512-FCP-NEXT: # zmm10 = zmm10[0,1,2,3],mem[4,5,6,7]11627; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm19 = zmm14[0,1,2,3],zmm19[4,5,6,7]11628; AVX512-FCP-NEXT: vshufi64x2 $228, (%rsp), %zmm13, %zmm1 # 64-byte Folded Reload11629; AVX512-FCP-NEXT: # zmm1 = zmm13[0,1,2,3],mem[4,5,6,7]11630; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm2[0,1,2,3],zmm18[4,5,6,7]11631; AVX512-FCP-NEXT: vshufi64x2 {{.*#+}} zmm3 = zmm3[0,1,2,3],zmm17[4,5,6,7]11632; AVX512-FCP-NEXT: vmovdqa64 256(%rdi), %zmm1811633; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm15 = [0,1,2,3,4,5,6,11]11634; AVX512-FCP-NEXT: vpermt2q %zmm18, %zmm15, %zmm011635; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11636; AVX512-FCP-NEXT: vmovdqa64 576(%rdi), %zmm1411637; AVX512-FCP-NEXT: vpermt2q %zmm14, %zmm15, %zmm411638; AVX512-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11639; AVX512-FCP-NEXT: vmovdqa64 896(%rdi), %zmm1711640; AVX512-FCP-NEXT: vpermt2q %zmm17, %zmm15, %zmm1111641; AVX512-FCP-NEXT: vmovdqu64 %zmm11, (%rsp) # 64-byte Spill11642; AVX512-FCP-NEXT: vmovdqa64 1216(%rdi), %zmm1311643; AVX512-FCP-NEXT: vpermt2q %zmm13, %zmm15, %zmm1011644; AVX512-FCP-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11645; AVX512-FCP-NEXT: vmovdqa64 1536(%rdi), %zmm1611646; AVX512-FCP-NEXT: vpermt2q %zmm16, %zmm15, %zmm1911647; AVX512-FCP-NEXT: vmovdqu64 %zmm19, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11648; AVX512-FCP-NEXT: vmovdqa64 1856(%rdi), %zmm1111649; AVX512-FCP-NEXT: vpermt2q %zmm11, %zmm15, %zmm111650; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11651; AVX512-FCP-NEXT: vmovdqa64 2176(%rdi), %zmm1911652; AVX512-FCP-NEXT: vpermt2q %zmm19, %zmm15, %zmm211653; AVX512-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11654; AVX512-FCP-NEXT: vmovdqa64 2496(%rdi), %zmm1011655; AVX512-FCP-NEXT: vpermt2q %zmm10, %zmm15, %zmm311656; AVX512-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11657; AVX512-FCP-NEXT: movb $7, %al11658; AVX512-FCP-NEXT: kmovw %eax, %k111659; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload11660; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload11661; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm15 {%k1}11662; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload11663; AVX512-FCP-NEXT: vmovdqa64 %zmm8, %zmm4 {%k1}11664; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload11665; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload11666; AVX512-FCP-NEXT: vmovdqa64 %zmm0, %zmm8 {%k1}11667; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload11668; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload11669; AVX512-FCP-NEXT: vmovdqa64 %zmm1, %zmm0 {%k1}11670; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload11671; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11672; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm1 {%k1}11673; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11674; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload11675; AVX512-FCP-NEXT: vmovdqa64 %zmm3, %zmm2 {%k1}11676; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload11677; AVX512-FCP-NEXT: vmovdqa64 %zmm26, %zmm3 {%k1}11678; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload11679; AVX512-FCP-NEXT: vmovdqa64 %zmm29, %zmm26 {%k1}11680; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm29 = [0,1,2,3,4,5,6,12]11681; AVX512-FCP-NEXT: vpermt2q %zmm18, %zmm29, %zmm1511682; AVX512-FCP-NEXT: vmovdqu64 %zmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11683; AVX512-FCP-NEXT: vpermt2q %zmm19, %zmm29, %zmm411684; AVX512-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11685; AVX512-FCP-NEXT: vpermt2q %zmm14, %zmm29, %zmm811686; AVX512-FCP-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11687; AVX512-FCP-NEXT: vpermt2q %zmm13, %zmm29, %zmm011688; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11689; AVX512-FCP-NEXT: vpermt2q %zmm17, %zmm29, %zmm111690; AVX512-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11691; AVX512-FCP-NEXT: vpermt2q %zmm11, %zmm29, %zmm211692; AVX512-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11693; AVX512-FCP-NEXT: vpermt2q %zmm16, %zmm29, %zmm311694; AVX512-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11695; AVX512-FCP-NEXT: vpermt2q %zmm10, %zmm29, %zmm2611696; AVX512-FCP-NEXT: vmovdqu64 %zmm26, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11697; AVX512-FCP-NEXT: movb $56, %al11698; AVX512-FCP-NEXT: kmovw %eax, %k111699; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11700; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload11701; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm0 {%k1}11702; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload11703; AVX512-FCP-NEXT: vmovdqa64 %zmm1, %zmm22 {%k1}11704; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11705; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload11706; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm8 {%k1}11707; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11708; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload11709; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm26 {%k1}11710; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11711; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload11712; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm15 {%k1}11713; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11714; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm30 {%k1}11715; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload11716; AVX512-FCP-NEXT: vmovdqa64 %zmm1, %zmm27 {%k1}11717; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload11718; AVX512-FCP-NEXT: vmovdqa64 %zmm1, %zmm6 {%k1}11719; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm1 = [0,1,2,3,4,5,8,13]11720; AVX512-FCP-NEXT: vpermt2q %zmm18, %zmm1, %zmm011721; AVX512-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11722; AVX512-FCP-NEXT: vpermt2q %zmm19, %zmm1, %zmm2211723; AVX512-FCP-NEXT: vpermt2q %zmm14, %zmm1, %zmm811724; AVX512-FCP-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11725; AVX512-FCP-NEXT: vpermt2q %zmm13, %zmm1, %zmm2611726; AVX512-FCP-NEXT: vpermt2q %zmm17, %zmm1, %zmm1511727; AVX512-FCP-NEXT: vpermt2q %zmm11, %zmm1, %zmm3011728; AVX512-FCP-NEXT: vpermt2q %zmm16, %zmm1, %zmm2711729; AVX512-FCP-NEXT: vpermt2q %zmm10, %zmm1, %zmm611730; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11731; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload11732; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm8 {%k1}11733; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload11734; AVX512-FCP-NEXT: vmovdqa64 %zmm1, %zmm24 {%k1}11735; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11736; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 # 64-byte Reload11737; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm29 {%k1}11738; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11739; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload11740; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm1 {%k1}11741; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11742; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload11743; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm3 {%k1}11744; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11745; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload11746; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm4 {%k1}11747; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11748; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm28 {%k1}11749; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11750; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm7 {%k1}11751; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm2 = [0,1,2,3,4,5,9,14]11752; AVX512-FCP-NEXT: vpermt2q %zmm18, %zmm2, %zmm811753; AVX512-FCP-NEXT: vpermt2q %zmm19, %zmm2, %zmm2411754; AVX512-FCP-NEXT: vpermt2q %zmm14, %zmm2, %zmm2911755; AVX512-FCP-NEXT: vpermt2q %zmm13, %zmm2, %zmm111756; AVX512-FCP-NEXT: vpermt2q %zmm17, %zmm2, %zmm311757; AVX512-FCP-NEXT: vmovdqa64 %zmm3, %zmm011758; AVX512-FCP-NEXT: vpermt2q %zmm11, %zmm2, %zmm411759; AVX512-FCP-NEXT: vpermt2q %zmm16, %zmm2, %zmm2811760; AVX512-FCP-NEXT: vpermt2q %zmm10, %zmm2, %zmm711761; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11762; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm31 {%k1}11763; AVX512-FCP-NEXT: vpmovsxbq {{.*#+}} zmm3 = [0,1,2,3,4,5,10,15]11764; AVX512-FCP-NEXT: vpermt2q %zmm14, %zmm3, %zmm3111765; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11766; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm23 {%k1}11767; AVX512-FCP-NEXT: vpermt2q %zmm18, %zmm3, %zmm2311768; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11769; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm21 {%k1}11770; AVX512-FCP-NEXT: vpermt2q %zmm13, %zmm3, %zmm2111771; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11772; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm20 {%k1}11773; AVX512-FCP-NEXT: vpermt2q %zmm17, %zmm3, %zmm2011774; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11775; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm25 {%k1}11776; AVX512-FCP-NEXT: vpermt2q %zmm11, %zmm3, %zmm2511777; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11778; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm12 {%k1}11779; AVX512-FCP-NEXT: vpermt2q %zmm16, %zmm3, %zmm1211780; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11781; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm5 {%k1}11782; AVX512-FCP-NEXT: vpermt2q %zmm10, %zmm3, %zmm511783; AVX512-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11784; AVX512-FCP-NEXT: vmovdqa64 %zmm2, %zmm9 {%k1}11785; AVX512-FCP-NEXT: vpermt2q %zmm19, %zmm3, %zmm911786; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11787; AVX512-FCP-NEXT: vmovaps %zmm2, 448(%rsi)11788; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11789; AVX512-FCP-NEXT: vmovaps %zmm2, 384(%rsi)11790; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11791; AVX512-FCP-NEXT: vmovaps %zmm2, 320(%rsi)11792; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11793; AVX512-FCP-NEXT: vmovaps %zmm2, 256(%rsi)11794; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11795; AVX512-FCP-NEXT: vmovaps %zmm2, 192(%rsi)11796; AVX512-FCP-NEXT: vmovups (%rsp), %zmm2 # 64-byte Reload11797; AVX512-FCP-NEXT: vmovaps %zmm2, 128(%rsi)11798; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11799; AVX512-FCP-NEXT: vmovaps %zmm2, 64(%rsi)11800; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11801; AVX512-FCP-NEXT: vmovaps %zmm2, (%rsi)11802; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11803; AVX512-FCP-NEXT: vmovaps %zmm2, 448(%rdx)11804; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11805; AVX512-FCP-NEXT: vmovaps %zmm2, 256(%rdx)11806; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11807; AVX512-FCP-NEXT: vmovaps %zmm2, 320(%rdx)11808; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11809; AVX512-FCP-NEXT: vmovaps %zmm2, 128(%rdx)11810; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11811; AVX512-FCP-NEXT: vmovaps %zmm2, 192(%rdx)11812; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11813; AVX512-FCP-NEXT: vmovaps %zmm2, (%rdx)11814; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11815; AVX512-FCP-NEXT: vmovaps %zmm2, 64(%rdx)11816; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11817; AVX512-FCP-NEXT: vmovaps %zmm2, 384(%rdx)11818; AVX512-FCP-NEXT: vmovdqa64 %zmm6, 448(%rcx)11819; AVX512-FCP-NEXT: vmovdqa64 %zmm27, 256(%rcx)11820; AVX512-FCP-NEXT: vmovdqa64 %zmm30, 320(%rcx)11821; AVX512-FCP-NEXT: vmovdqa64 %zmm15, 128(%rcx)11822; AVX512-FCP-NEXT: vmovdqa64 %zmm26, 192(%rcx)11823; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11824; AVX512-FCP-NEXT: vmovaps %zmm2, (%rcx)11825; AVX512-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload11826; AVX512-FCP-NEXT: vmovaps %zmm2, 64(%rcx)11827; AVX512-FCP-NEXT: vmovdqa64 %zmm22, 384(%rcx)11828; AVX512-FCP-NEXT: vmovdqa64 %zmm7, 448(%r8)11829; AVX512-FCP-NEXT: vmovdqa64 %zmm28, 256(%r8)11830; AVX512-FCP-NEXT: vmovdqa64 %zmm4, 320(%r8)11831; AVX512-FCP-NEXT: vmovdqa64 %zmm0, 128(%r8)11832; AVX512-FCP-NEXT: vmovdqa64 %zmm1, 192(%r8)11833; AVX512-FCP-NEXT: vmovdqa64 %zmm8, (%r8)11834; AVX512-FCP-NEXT: vmovdqa64 %zmm29, 64(%r8)11835; AVX512-FCP-NEXT: vmovdqa64 %zmm24, 384(%r8)11836; AVX512-FCP-NEXT: vmovdqa64 %zmm9, 384(%r9)11837; AVX512-FCP-NEXT: vmovdqa64 %zmm5, 448(%r9)11838; AVX512-FCP-NEXT: vmovdqa64 %zmm12, 256(%r9)11839; AVX512-FCP-NEXT: vmovdqa64 %zmm25, 320(%r9)11840; AVX512-FCP-NEXT: vmovdqa64 %zmm20, 128(%r9)11841; AVX512-FCP-NEXT: vmovdqa64 %zmm21, 192(%r9)11842; AVX512-FCP-NEXT: vmovdqa64 %zmm23, (%r9)11843; AVX512-FCP-NEXT: vmovdqa64 %zmm31, 64(%r9)11844; AVX512-FCP-NEXT: addq $3336, %rsp # imm = 0xD0811845; AVX512-FCP-NEXT: vzeroupper11846; AVX512-FCP-NEXT: retq11847;11848; AVX512DQ-LABEL: load_i64_stride5_vf64:11849; AVX512DQ: # %bb.0:11850; AVX512DQ-NEXT: subq $3336, %rsp # imm = 0xD0811851; AVX512DQ-NEXT: vmovdqa64 1728(%rdi), %zmm311852; AVX512DQ-NEXT: vmovdqa64 1792(%rdi), %zmm711853; AVX512DQ-NEXT: vmovdqa64 1408(%rdi), %zmm011854; AVX512DQ-NEXT: vmovdqa64 1088(%rdi), %zmm111855; AVX512DQ-NEXT: vmovdqa64 1152(%rdi), %zmm611856; AVX512DQ-NEXT: vmovdqa64 768(%rdi), %zmm511857; AVX512DQ-NEXT: vmovdqa64 832(%rdi), %zmm811858; AVX512DQ-NEXT: vmovdqa64 448(%rdi), %zmm411859; AVX512DQ-NEXT: vmovdqa64 512(%rdi), %zmm911860; AVX512DQ-NEXT: vmovdqa64 128(%rdi), %zmm211861; AVX512DQ-NEXT: vmovdqa64 192(%rdi), %zmm1011862; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm16 = [12,1,6,0,12,1,6,0]11863; AVX512DQ-NEXT: # zmm16 = mem[0,1,2,3,0,1,2,3]11864; AVX512DQ-NEXT: vmovdqa64 %zmm10, %zmm1111865; AVX512DQ-NEXT: vpermt2q %zmm2, %zmm16, %zmm1111866; AVX512DQ-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11867; AVX512DQ-NEXT: vmovdqa64 %zmm9, %zmm1111868; AVX512DQ-NEXT: vpermt2q %zmm4, %zmm16, %zmm1111869; AVX512DQ-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11870; AVX512DQ-NEXT: vmovdqa64 %zmm8, %zmm1111871; AVX512DQ-NEXT: vpermt2q %zmm5, %zmm16, %zmm1111872; AVX512DQ-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11873; AVX512DQ-NEXT: vmovdqa64 %zmm6, %zmm1211874; AVX512DQ-NEXT: vmovdqa64 %zmm6, %zmm1111875; AVX512DQ-NEXT: vpermt2q %zmm1, %zmm16, %zmm1211876; AVX512DQ-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11877; AVX512DQ-NEXT: vmovdqa64 %zmm7, %zmm611878; AVX512DQ-NEXT: vpermt2q %zmm3, %zmm16, %zmm611879; AVX512DQ-NEXT: vmovdqu64 %zmm6, (%rsp) # 64-byte Spill11880; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm12 = [5,10,15,0,5,10,15,0]11881; AVX512DQ-NEXT: # zmm12 = mem[0,1,2,3,0,1,2,3]11882; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm611883; AVX512DQ-NEXT: vpermt2q %zmm10, %zmm12, %zmm611884; AVX512DQ-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11885; AVX512DQ-NEXT: vmovdqa64 %zmm4, %zmm611886; AVX512DQ-NEXT: vpermt2q %zmm9, %zmm12, %zmm611887; AVX512DQ-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11888; AVX512DQ-NEXT: vmovdqa64 %zmm1, %zmm611889; AVX512DQ-NEXT: vpermt2q %zmm11, %zmm12, %zmm611890; AVX512DQ-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11891; AVX512DQ-NEXT: vmovdqa64 %zmm5, %zmm611892; AVX512DQ-NEXT: vpermt2q %zmm8, %zmm12, %zmm611893; AVX512DQ-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11894; AVX512DQ-NEXT: vmovdqa64 %zmm3, %zmm611895; AVX512DQ-NEXT: vpermt2q %zmm7, %zmm12, %zmm611896; AVX512DQ-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11897; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = [6,11,0,1,6,11,0,1]11898; AVX512DQ-NEXT: # zmm6 = mem[0,1,2,3,0,1,2,3]11899; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm1311900; AVX512DQ-NEXT: vpermt2q %zmm10, %zmm6, %zmm1311901; AVX512DQ-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11902; AVX512DQ-NEXT: vmovdqa64 %zmm6, %zmm1311903; AVX512DQ-NEXT: vmovdqa64 %zmm4, %zmm611904; AVX512DQ-NEXT: vpermt2q %zmm9, %zmm13, %zmm611905; AVX512DQ-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11906; AVX512DQ-NEXT: vmovdqa64 %zmm1, %zmm611907; AVX512DQ-NEXT: vpermt2q %zmm11, %zmm13, %zmm611908; AVX512DQ-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11909; AVX512DQ-NEXT: vmovdqa64 %zmm5, %zmm611910; AVX512DQ-NEXT: vpermt2q %zmm8, %zmm13, %zmm611911; AVX512DQ-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11912; AVX512DQ-NEXT: vmovdqa64 %zmm3, %zmm611913; AVX512DQ-NEXT: vpermt2q %zmm7, %zmm13, %zmm611914; AVX512DQ-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11915; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm14 = [7,12,0,2,7,12,0,2]11916; AVX512DQ-NEXT: # zmm14 = mem[0,1,2,3,0,1,2,3]11917; AVX512DQ-NEXT: vmovdqa64 %zmm4, %zmm611918; AVX512DQ-NEXT: vpermt2q %zmm9, %zmm14, %zmm611919; AVX512DQ-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11920; AVX512DQ-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = [0,5,0,11,0,5,0,11]11921; AVX512DQ-NEXT: # zmm6 = mem[0,1,2,3,0,1,2,3]11922; AVX512DQ-NEXT: vpermt2q %zmm4, %zmm6, %zmm911923; AVX512DQ-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11924; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm411925; AVX512DQ-NEXT: vpermt2q %zmm10, %zmm14, %zmm411926; AVX512DQ-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11927; AVX512DQ-NEXT: vpermt2q %zmm2, %zmm6, %zmm1011928; AVX512DQ-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11929; AVX512DQ-NEXT: vmovdqa64 %zmm1, %zmm211930; AVX512DQ-NEXT: vpermt2q %zmm11, %zmm14, %zmm211931; AVX512DQ-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11932; AVX512DQ-NEXT: vpermt2q %zmm1, %zmm6, %zmm1111933; AVX512DQ-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11934; AVX512DQ-NEXT: vmovdqa64 %zmm5, %zmm111935; AVX512DQ-NEXT: vpermt2q %zmm8, %zmm14, %zmm111936; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11937; AVX512DQ-NEXT: vpermt2q %zmm5, %zmm6, %zmm811938; AVX512DQ-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11939; AVX512DQ-NEXT: vmovdqa64 %zmm3, %zmm111940; AVX512DQ-NEXT: vpermt2q %zmm7, %zmm14, %zmm111941; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11942; AVX512DQ-NEXT: vpermt2q %zmm3, %zmm6, %zmm711943; AVX512DQ-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11944; AVX512DQ-NEXT: vmovdqa64 1472(%rdi), %zmm211945; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm1911946; AVX512DQ-NEXT: vpermt2q %zmm0, %zmm16, %zmm1911947; AVX512DQ-NEXT: vmovdqa64 %zmm0, %zmm111948; AVX512DQ-NEXT: vpermt2q %zmm2, %zmm12, %zmm111949; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11950; AVX512DQ-NEXT: vmovdqa64 %zmm0, %zmm111951; AVX512DQ-NEXT: vpermt2q %zmm2, %zmm13, %zmm111952; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11953; AVX512DQ-NEXT: vmovdqa64 %zmm0, %zmm111954; AVX512DQ-NEXT: vpermt2q %zmm2, %zmm14, %zmm111955; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11956; AVX512DQ-NEXT: vpermt2q %zmm0, %zmm6, %zmm211957; AVX512DQ-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11958; AVX512DQ-NEXT: vmovdqa64 2048(%rdi), %zmm011959; AVX512DQ-NEXT: vmovdqa64 2112(%rdi), %zmm211960; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm1811961; AVX512DQ-NEXT: vpermt2q %zmm0, %zmm16, %zmm1811962; AVX512DQ-NEXT: vmovdqa64 %zmm0, %zmm111963; AVX512DQ-NEXT: vpermt2q %zmm2, %zmm12, %zmm111964; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11965; AVX512DQ-NEXT: vmovdqa64 %zmm0, %zmm111966; AVX512DQ-NEXT: vpermt2q %zmm2, %zmm13, %zmm111967; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11968; AVX512DQ-NEXT: vmovdqa64 %zmm0, %zmm111969; AVX512DQ-NEXT: vpermt2q %zmm2, %zmm14, %zmm111970; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11971; AVX512DQ-NEXT: vpermt2q %zmm0, %zmm6, %zmm211972; AVX512DQ-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11973; AVX512DQ-NEXT: vmovdqa64 2368(%rdi), %zmm011974; AVX512DQ-NEXT: vmovdqa64 2432(%rdi), %zmm111975; AVX512DQ-NEXT: vpermi2q %zmm1, %zmm0, %zmm1211976; AVX512DQ-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11977; AVX512DQ-NEXT: vpermi2q %zmm1, %zmm0, %zmm1311978; AVX512DQ-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11979; AVX512DQ-NEXT: vpermi2q %zmm1, %zmm0, %zmm1411980; AVX512DQ-NEXT: vmovdqu64 %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11981; AVX512DQ-NEXT: vmovdqa64 %zmm1, %zmm1711982; AVX512DQ-NEXT: vpermt2q %zmm0, %zmm6, %zmm111983; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11984; AVX512DQ-NEXT: vpermt2q %zmm0, %zmm16, %zmm1711985; AVX512DQ-NEXT: vmovdqa64 384(%rdi), %zmm3111986; AVX512DQ-NEXT: vmovdqa64 320(%rdi), %zmm211987; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} ymm3 = [0,5,10,15]11988; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm111989; AVX512DQ-NEXT: vpermt2q %zmm31, %zmm3, %zmm111990; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} ymm29 = [1,6,11,0]11991; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm011992; AVX512DQ-NEXT: vpermt2q %zmm31, %zmm29, %zmm011993; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11994; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} ymm6 = [2,7,12,0]11995; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm011996; AVX512DQ-NEXT: vpermt2q %zmm31, %zmm6, %zmm011997; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill11998; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} ymm7 = [11,0,5,0]11999; AVX512DQ-NEXT: vmovdqa64 %zmm31, %zmm012000; AVX512DQ-NEXT: vpermt2q %zmm2, %zmm7, %zmm012001; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12002; AVX512DQ-NEXT: vpermt2q %zmm2, %zmm16, %zmm3112003; AVX512DQ-NEXT: vmovdqa64 (%rdi), %zmm212004; AVX512DQ-NEXT: vmovdqa64 64(%rdi), %zmm2312005; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm412006; AVX512DQ-NEXT: vpermt2q %zmm23, %zmm3, %zmm412007; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm012008; AVX512DQ-NEXT: vpermt2q %zmm23, %zmm29, %zmm012009; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12010; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm012011; AVX512DQ-NEXT: vpermt2q %zmm23, %zmm6, %zmm012012; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12013; AVX512DQ-NEXT: vmovdqa64 %zmm23, %zmm012014; AVX512DQ-NEXT: vpermt2q %zmm2, %zmm7, %zmm012015; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12016; AVX512DQ-NEXT: vpermt2q %zmm2, %zmm16, %zmm2312017; AVX512DQ-NEXT: vmovdqa64 1024(%rdi), %zmm2112018; AVX512DQ-NEXT: vmovdqa64 960(%rdi), %zmm212019; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm1012020; AVX512DQ-NEXT: vpermt2q %zmm21, %zmm3, %zmm1012021; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm012022; AVX512DQ-NEXT: vpermt2q %zmm21, %zmm29, %zmm012023; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12024; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm012025; AVX512DQ-NEXT: vpermt2q %zmm21, %zmm6, %zmm012026; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12027; AVX512DQ-NEXT: vmovdqa64 %zmm21, %zmm012028; AVX512DQ-NEXT: vpermt2q %zmm2, %zmm7, %zmm012029; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12030; AVX512DQ-NEXT: vpermt2q %zmm2, %zmm16, %zmm2112031; AVX512DQ-NEXT: vmovdqa64 704(%rdi), %zmm2012032; AVX512DQ-NEXT: vmovdqa64 640(%rdi), %zmm212033; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm1112034; AVX512DQ-NEXT: vpermt2q %zmm20, %zmm3, %zmm1112035; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm012036; AVX512DQ-NEXT: vpermt2q %zmm20, %zmm29, %zmm012037; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12038; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm012039; AVX512DQ-NEXT: vpermt2q %zmm20, %zmm6, %zmm012040; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12041; AVX512DQ-NEXT: vmovdqa64 %zmm20, %zmm012042; AVX512DQ-NEXT: vpermt2q %zmm2, %zmm7, %zmm012043; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12044; AVX512DQ-NEXT: vpermt2q %zmm2, %zmm16, %zmm2012045; AVX512DQ-NEXT: vmovdqa64 1664(%rdi), %zmm2512046; AVX512DQ-NEXT: vmovdqa64 1600(%rdi), %zmm212047; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm1312048; AVX512DQ-NEXT: vpermt2q %zmm25, %zmm3, %zmm1312049; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm012050; AVX512DQ-NEXT: vpermt2q %zmm25, %zmm29, %zmm012051; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12052; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm3012053; AVX512DQ-NEXT: vpermt2q %zmm25, %zmm6, %zmm3012054; AVX512DQ-NEXT: vmovdqa64 %zmm25, %zmm012055; AVX512DQ-NEXT: vpermt2q %zmm2, %zmm7, %zmm012056; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12057; AVX512DQ-NEXT: vpermt2q %zmm2, %zmm16, %zmm2512058; AVX512DQ-NEXT: vmovdqa64 1344(%rdi), %zmm1212059; AVX512DQ-NEXT: vmovdqa64 1280(%rdi), %zmm212060; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm1412061; AVX512DQ-NEXT: vpermt2q %zmm12, %zmm3, %zmm1412062; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm2612063; AVX512DQ-NEXT: vpermt2q %zmm12, %zmm29, %zmm2612064; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm2712065; AVX512DQ-NEXT: vpermt2q %zmm12, %zmm6, %zmm2712066; AVX512DQ-NEXT: vmovdqa64 %zmm12, %zmm2812067; AVX512DQ-NEXT: vpermt2q %zmm2, %zmm7, %zmm2812068; AVX512DQ-NEXT: vpermt2q %zmm2, %zmm16, %zmm1212069; AVX512DQ-NEXT: vmovdqa64 1984(%rdi), %zmm912070; AVX512DQ-NEXT: vmovdqa64 1920(%rdi), %zmm1512071; AVX512DQ-NEXT: vmovdqa64 %zmm15, %zmm212072; AVX512DQ-NEXT: vpermt2q %zmm9, %zmm3, %zmm212073; AVX512DQ-NEXT: vmovdqa64 2304(%rdi), %zmm512074; AVX512DQ-NEXT: vmovdqa64 2240(%rdi), %zmm012075; AVX512DQ-NEXT: vpermi2q %zmm5, %zmm0, %zmm312076; AVX512DQ-NEXT: vmovdqa64 %zmm15, %zmm812077; AVX512DQ-NEXT: vpermt2q %zmm9, %zmm29, %zmm812078; AVX512DQ-NEXT: vpermi2q %zmm5, %zmm0, %zmm2912079; AVX512DQ-NEXT: vmovdqa64 %zmm15, %zmm2212080; AVX512DQ-NEXT: vpermt2q %zmm9, %zmm6, %zmm2212081; AVX512DQ-NEXT: vpermi2q %zmm5, %zmm0, %zmm612082; AVX512DQ-NEXT: vmovdqa64 %zmm9, %zmm2412083; AVX512DQ-NEXT: vpermt2q %zmm15, %zmm7, %zmm2412084; AVX512DQ-NEXT: vpermi2q %zmm0, %zmm5, %zmm712085; AVX512DQ-NEXT: vpermt2q %zmm0, %zmm16, %zmm512086; AVX512DQ-NEXT: vpermt2q %zmm15, %zmm16, %zmm912087; AVX512DQ-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm0 # 64-byte Folded Reload12088; AVX512DQ-NEXT: # zmm0 = zmm4[0,1,2,3],mem[4,5,6,7]12089; AVX512DQ-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm4 # 64-byte Folded Reload12090; AVX512DQ-NEXT: # zmm4 = zmm1[0,1,2,3],mem[4,5,6,7]12091; AVX512DQ-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm11, %zmm11 # 64-byte Folded Reload12092; AVX512DQ-NEXT: # zmm11 = zmm11[0,1,2,3],mem[4,5,6,7]12093; AVX512DQ-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm10, %zmm10 # 64-byte Folded Reload12094; AVX512DQ-NEXT: # zmm10 = zmm10[0,1,2,3],mem[4,5,6,7]12095; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm19 = zmm14[0,1,2,3],zmm19[4,5,6,7]12096; AVX512DQ-NEXT: vshufi64x2 $228, (%rsp), %zmm13, %zmm1 # 64-byte Folded Reload12097; AVX512DQ-NEXT: # zmm1 = zmm13[0,1,2,3],mem[4,5,6,7]12098; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm2[0,1,2,3],zmm18[4,5,6,7]12099; AVX512DQ-NEXT: vshufi64x2 {{.*#+}} zmm3 = zmm3[0,1,2,3],zmm17[4,5,6,7]12100; AVX512DQ-NEXT: vmovdqa64 256(%rdi), %zmm1812101; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} zmm15 = [0,1,2,3,4,5,6,11]12102; AVX512DQ-NEXT: vpermt2q %zmm18, %zmm15, %zmm012103; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12104; AVX512DQ-NEXT: vmovdqa64 576(%rdi), %zmm1412105; AVX512DQ-NEXT: vpermt2q %zmm14, %zmm15, %zmm412106; AVX512DQ-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12107; AVX512DQ-NEXT: vmovdqa64 896(%rdi), %zmm1712108; AVX512DQ-NEXT: vpermt2q %zmm17, %zmm15, %zmm1112109; AVX512DQ-NEXT: vmovdqu64 %zmm11, (%rsp) # 64-byte Spill12110; AVX512DQ-NEXT: vmovdqa64 1216(%rdi), %zmm1312111; AVX512DQ-NEXT: vpermt2q %zmm13, %zmm15, %zmm1012112; AVX512DQ-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12113; AVX512DQ-NEXT: vmovdqa64 1536(%rdi), %zmm1612114; AVX512DQ-NEXT: vpermt2q %zmm16, %zmm15, %zmm1912115; AVX512DQ-NEXT: vmovdqu64 %zmm19, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12116; AVX512DQ-NEXT: vmovdqa64 1856(%rdi), %zmm1112117; AVX512DQ-NEXT: vpermt2q %zmm11, %zmm15, %zmm112118; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12119; AVX512DQ-NEXT: vmovdqa64 2176(%rdi), %zmm1912120; AVX512DQ-NEXT: vpermt2q %zmm19, %zmm15, %zmm212121; AVX512DQ-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12122; AVX512DQ-NEXT: vmovdqa64 2496(%rdi), %zmm1012123; AVX512DQ-NEXT: vpermt2q %zmm10, %zmm15, %zmm312124; AVX512DQ-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12125; AVX512DQ-NEXT: movb $7, %al12126; AVX512DQ-NEXT: kmovw %eax, %k112127; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload12128; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12129; AVX512DQ-NEXT: vmovdqa64 %zmm0, %zmm15 {%k1}12130; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload12131; AVX512DQ-NEXT: vmovdqa64 %zmm8, %zmm4 {%k1}12132; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload12133; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12134; AVX512DQ-NEXT: vmovdqa64 %zmm0, %zmm8 {%k1}12135; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12136; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload12137; AVX512DQ-NEXT: vmovdqa64 %zmm1, %zmm0 {%k1}12138; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload12139; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12140; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm1 {%k1}12141; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12142; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload12143; AVX512DQ-NEXT: vmovdqa64 %zmm3, %zmm2 {%k1}12144; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload12145; AVX512DQ-NEXT: vmovdqa64 %zmm26, %zmm3 {%k1}12146; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload12147; AVX512DQ-NEXT: vmovdqa64 %zmm29, %zmm26 {%k1}12148; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} zmm29 = [0,1,2,3,4,5,6,12]12149; AVX512DQ-NEXT: vpermt2q %zmm18, %zmm29, %zmm1512150; AVX512DQ-NEXT: vmovdqu64 %zmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12151; AVX512DQ-NEXT: vpermt2q %zmm19, %zmm29, %zmm412152; AVX512DQ-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12153; AVX512DQ-NEXT: vpermt2q %zmm14, %zmm29, %zmm812154; AVX512DQ-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12155; AVX512DQ-NEXT: vpermt2q %zmm13, %zmm29, %zmm012156; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12157; AVX512DQ-NEXT: vpermt2q %zmm17, %zmm29, %zmm112158; AVX512DQ-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12159; AVX512DQ-NEXT: vpermt2q %zmm11, %zmm29, %zmm212160; AVX512DQ-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12161; AVX512DQ-NEXT: vpermt2q %zmm16, %zmm29, %zmm312162; AVX512DQ-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12163; AVX512DQ-NEXT: vpermt2q %zmm10, %zmm29, %zmm2612164; AVX512DQ-NEXT: vmovdqu64 %zmm26, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12165; AVX512DQ-NEXT: movb $56, %al12166; AVX512DQ-NEXT: kmovw %eax, %k112167; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12168; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12169; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm0 {%k1}12170; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload12171; AVX512DQ-NEXT: vmovdqa64 %zmm1, %zmm22 {%k1}12172; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12173; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload12174; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm8 {%k1}12175; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12176; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload12177; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm26 {%k1}12178; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12179; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload12180; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm15 {%k1}12181; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12182; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm30 {%k1}12183; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload12184; AVX512DQ-NEXT: vmovdqa64 %zmm1, %zmm27 {%k1}12185; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload12186; AVX512DQ-NEXT: vmovdqa64 %zmm1, %zmm6 {%k1}12187; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} zmm1 = [0,1,2,3,4,5,8,13]12188; AVX512DQ-NEXT: vpermt2q %zmm18, %zmm1, %zmm012189; AVX512DQ-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12190; AVX512DQ-NEXT: vpermt2q %zmm19, %zmm1, %zmm2212191; AVX512DQ-NEXT: vpermt2q %zmm14, %zmm1, %zmm812192; AVX512DQ-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12193; AVX512DQ-NEXT: vpermt2q %zmm13, %zmm1, %zmm2612194; AVX512DQ-NEXT: vpermt2q %zmm17, %zmm1, %zmm1512195; AVX512DQ-NEXT: vpermt2q %zmm11, %zmm1, %zmm3012196; AVX512DQ-NEXT: vpermt2q %zmm16, %zmm1, %zmm2712197; AVX512DQ-NEXT: vpermt2q %zmm10, %zmm1, %zmm612198; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12199; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload12200; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm8 {%k1}12201; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload12202; AVX512DQ-NEXT: vmovdqa64 %zmm1, %zmm24 {%k1}12203; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12204; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 # 64-byte Reload12205; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm29 {%k1}12206; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12207; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload12208; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm1 {%k1}12209; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12210; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload12211; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm3 {%k1}12212; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12213; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload12214; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm4 {%k1}12215; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12216; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm28 {%k1}12217; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12218; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm7 {%k1}12219; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} zmm2 = [0,1,2,3,4,5,9,14]12220; AVX512DQ-NEXT: vpermt2q %zmm18, %zmm2, %zmm812221; AVX512DQ-NEXT: vpermt2q %zmm19, %zmm2, %zmm2412222; AVX512DQ-NEXT: vpermt2q %zmm14, %zmm2, %zmm2912223; AVX512DQ-NEXT: vpermt2q %zmm13, %zmm2, %zmm112224; AVX512DQ-NEXT: vpermt2q %zmm17, %zmm2, %zmm312225; AVX512DQ-NEXT: vmovdqa64 %zmm3, %zmm012226; AVX512DQ-NEXT: vpermt2q %zmm11, %zmm2, %zmm412227; AVX512DQ-NEXT: vpermt2q %zmm16, %zmm2, %zmm2812228; AVX512DQ-NEXT: vpermt2q %zmm10, %zmm2, %zmm712229; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12230; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm31 {%k1}12231; AVX512DQ-NEXT: vpmovsxbq {{.*#+}} zmm3 = [0,1,2,3,4,5,10,15]12232; AVX512DQ-NEXT: vpermt2q %zmm14, %zmm3, %zmm3112233; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12234; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm23 {%k1}12235; AVX512DQ-NEXT: vpermt2q %zmm18, %zmm3, %zmm2312236; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12237; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm21 {%k1}12238; AVX512DQ-NEXT: vpermt2q %zmm13, %zmm3, %zmm2112239; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12240; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm20 {%k1}12241; AVX512DQ-NEXT: vpermt2q %zmm17, %zmm3, %zmm2012242; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12243; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm25 {%k1}12244; AVX512DQ-NEXT: vpermt2q %zmm11, %zmm3, %zmm2512245; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12246; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm12 {%k1}12247; AVX512DQ-NEXT: vpermt2q %zmm16, %zmm3, %zmm1212248; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12249; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm5 {%k1}12250; AVX512DQ-NEXT: vpermt2q %zmm10, %zmm3, %zmm512251; AVX512DQ-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12252; AVX512DQ-NEXT: vmovdqa64 %zmm2, %zmm9 {%k1}12253; AVX512DQ-NEXT: vpermt2q %zmm19, %zmm3, %zmm912254; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12255; AVX512DQ-NEXT: vmovaps %zmm2, 448(%rsi)12256; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12257; AVX512DQ-NEXT: vmovaps %zmm2, 384(%rsi)12258; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12259; AVX512DQ-NEXT: vmovaps %zmm2, 320(%rsi)12260; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12261; AVX512DQ-NEXT: vmovaps %zmm2, 256(%rsi)12262; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12263; AVX512DQ-NEXT: vmovaps %zmm2, 192(%rsi)12264; AVX512DQ-NEXT: vmovups (%rsp), %zmm2 # 64-byte Reload12265; AVX512DQ-NEXT: vmovaps %zmm2, 128(%rsi)12266; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12267; AVX512DQ-NEXT: vmovaps %zmm2, 64(%rsi)12268; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12269; AVX512DQ-NEXT: vmovaps %zmm2, (%rsi)12270; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12271; AVX512DQ-NEXT: vmovaps %zmm2, 448(%rdx)12272; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12273; AVX512DQ-NEXT: vmovaps %zmm2, 256(%rdx)12274; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12275; AVX512DQ-NEXT: vmovaps %zmm2, 320(%rdx)12276; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12277; AVX512DQ-NEXT: vmovaps %zmm2, 128(%rdx)12278; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12279; AVX512DQ-NEXT: vmovaps %zmm2, 192(%rdx)12280; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12281; AVX512DQ-NEXT: vmovaps %zmm2, (%rdx)12282; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12283; AVX512DQ-NEXT: vmovaps %zmm2, 64(%rdx)12284; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12285; AVX512DQ-NEXT: vmovaps %zmm2, 384(%rdx)12286; AVX512DQ-NEXT: vmovdqa64 %zmm6, 448(%rcx)12287; AVX512DQ-NEXT: vmovdqa64 %zmm27, 256(%rcx)12288; AVX512DQ-NEXT: vmovdqa64 %zmm30, 320(%rcx)12289; AVX512DQ-NEXT: vmovdqa64 %zmm15, 128(%rcx)12290; AVX512DQ-NEXT: vmovdqa64 %zmm26, 192(%rcx)12291; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12292; AVX512DQ-NEXT: vmovaps %zmm2, (%rcx)12293; AVX512DQ-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12294; AVX512DQ-NEXT: vmovaps %zmm2, 64(%rcx)12295; AVX512DQ-NEXT: vmovdqa64 %zmm22, 384(%rcx)12296; AVX512DQ-NEXT: vmovdqa64 %zmm7, 448(%r8)12297; AVX512DQ-NEXT: vmovdqa64 %zmm28, 256(%r8)12298; AVX512DQ-NEXT: vmovdqa64 %zmm4, 320(%r8)12299; AVX512DQ-NEXT: vmovdqa64 %zmm0, 128(%r8)12300; AVX512DQ-NEXT: vmovdqa64 %zmm1, 192(%r8)12301; AVX512DQ-NEXT: vmovdqa64 %zmm8, (%r8)12302; AVX512DQ-NEXT: vmovdqa64 %zmm29, 64(%r8)12303; AVX512DQ-NEXT: vmovdqa64 %zmm24, 384(%r8)12304; AVX512DQ-NEXT: vmovdqa64 %zmm9, 384(%r9)12305; AVX512DQ-NEXT: vmovdqa64 %zmm5, 448(%r9)12306; AVX512DQ-NEXT: vmovdqa64 %zmm12, 256(%r9)12307; AVX512DQ-NEXT: vmovdqa64 %zmm25, 320(%r9)12308; AVX512DQ-NEXT: vmovdqa64 %zmm20, 128(%r9)12309; AVX512DQ-NEXT: vmovdqa64 %zmm21, 192(%r9)12310; AVX512DQ-NEXT: vmovdqa64 %zmm23, (%r9)12311; AVX512DQ-NEXT: vmovdqa64 %zmm31, 64(%r9)12312; AVX512DQ-NEXT: addq $3336, %rsp # imm = 0xD0812313; AVX512DQ-NEXT: vzeroupper12314; AVX512DQ-NEXT: retq12315;12316; AVX512DQ-FCP-LABEL: load_i64_stride5_vf64:12317; AVX512DQ-FCP: # %bb.0:12318; AVX512DQ-FCP-NEXT: subq $3336, %rsp # imm = 0xD0812319; AVX512DQ-FCP-NEXT: vmovdqa64 1728(%rdi), %zmm312320; AVX512DQ-FCP-NEXT: vmovdqa64 1792(%rdi), %zmm712321; AVX512DQ-FCP-NEXT: vmovdqa64 1408(%rdi), %zmm012322; AVX512DQ-FCP-NEXT: vmovdqa64 1088(%rdi), %zmm112323; AVX512DQ-FCP-NEXT: vmovdqa64 1152(%rdi), %zmm612324; AVX512DQ-FCP-NEXT: vmovdqa64 768(%rdi), %zmm512325; AVX512DQ-FCP-NEXT: vmovdqa64 832(%rdi), %zmm812326; AVX512DQ-FCP-NEXT: vmovdqa64 448(%rdi), %zmm412327; AVX512DQ-FCP-NEXT: vmovdqa64 512(%rdi), %zmm912328; AVX512DQ-FCP-NEXT: vmovdqa64 128(%rdi), %zmm212329; AVX512DQ-FCP-NEXT: vmovdqa64 192(%rdi), %zmm1012330; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm16 = [12,1,6,0,12,1,6,0]12331; AVX512DQ-FCP-NEXT: # zmm16 = mem[0,1,2,3,0,1,2,3]12332; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm10, %zmm1112333; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm16, %zmm1112334; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12335; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm9, %zmm1112336; AVX512DQ-FCP-NEXT: vpermt2q %zmm4, %zmm16, %zmm1112337; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12338; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm8, %zmm1112339; AVX512DQ-FCP-NEXT: vpermt2q %zmm5, %zmm16, %zmm1112340; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12341; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, %zmm1212342; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, %zmm1112343; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm16, %zmm1212344; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12345; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm7, %zmm612346; AVX512DQ-FCP-NEXT: vpermt2q %zmm3, %zmm16, %zmm612347; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm6, (%rsp) # 64-byte Spill12348; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm12 = [5,10,15,0,5,10,15,0]12349; AVX512DQ-FCP-NEXT: # zmm12 = mem[0,1,2,3,0,1,2,3]12350; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm612351; AVX512DQ-FCP-NEXT: vpermt2q %zmm10, %zmm12, %zmm612352; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12353; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm4, %zmm612354; AVX512DQ-FCP-NEXT: vpermt2q %zmm9, %zmm12, %zmm612355; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12356; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, %zmm612357; AVX512DQ-FCP-NEXT: vpermt2q %zmm11, %zmm12, %zmm612358; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12359; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm5, %zmm612360; AVX512DQ-FCP-NEXT: vpermt2q %zmm8, %zmm12, %zmm612361; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12362; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, %zmm612363; AVX512DQ-FCP-NEXT: vpermt2q %zmm7, %zmm12, %zmm612364; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12365; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = [6,11,0,1,6,11,0,1]12366; AVX512DQ-FCP-NEXT: # zmm6 = mem[0,1,2,3,0,1,2,3]12367; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm1312368; AVX512DQ-FCP-NEXT: vpermt2q %zmm10, %zmm6, %zmm1312369; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12370; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, %zmm1312371; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm4, %zmm612372; AVX512DQ-FCP-NEXT: vpermt2q %zmm9, %zmm13, %zmm612373; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12374; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, %zmm612375; AVX512DQ-FCP-NEXT: vpermt2q %zmm11, %zmm13, %zmm612376; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12377; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm5, %zmm612378; AVX512DQ-FCP-NEXT: vpermt2q %zmm8, %zmm13, %zmm612379; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12380; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, %zmm612381; AVX512DQ-FCP-NEXT: vpermt2q %zmm7, %zmm13, %zmm612382; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12383; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm14 = [7,12,0,2,7,12,0,2]12384; AVX512DQ-FCP-NEXT: # zmm14 = mem[0,1,2,3,0,1,2,3]12385; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm4, %zmm612386; AVX512DQ-FCP-NEXT: vpermt2q %zmm9, %zmm14, %zmm612387; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12388; AVX512DQ-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = [0,5,0,11,0,5,0,11]12389; AVX512DQ-FCP-NEXT: # zmm6 = mem[0,1,2,3,0,1,2,3]12390; AVX512DQ-FCP-NEXT: vpermt2q %zmm4, %zmm6, %zmm912391; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12392; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm412393; AVX512DQ-FCP-NEXT: vpermt2q %zmm10, %zmm14, %zmm412394; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12395; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm6, %zmm1012396; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12397; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, %zmm212398; AVX512DQ-FCP-NEXT: vpermt2q %zmm11, %zmm14, %zmm212399; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12400; AVX512DQ-FCP-NEXT: vpermt2q %zmm1, %zmm6, %zmm1112401; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12402; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm5, %zmm112403; AVX512DQ-FCP-NEXT: vpermt2q %zmm8, %zmm14, %zmm112404; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12405; AVX512DQ-FCP-NEXT: vpermt2q %zmm5, %zmm6, %zmm812406; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12407; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, %zmm112408; AVX512DQ-FCP-NEXT: vpermt2q %zmm7, %zmm14, %zmm112409; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12410; AVX512DQ-FCP-NEXT: vpermt2q %zmm3, %zmm6, %zmm712411; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12412; AVX512DQ-FCP-NEXT: vmovdqa64 1472(%rdi), %zmm212413; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm1912414; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm16, %zmm1912415; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm112416; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm12, %zmm112417; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12418; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm112419; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm13, %zmm112420; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12421; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm112422; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm14, %zmm112423; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12424; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm6, %zmm212425; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12426; AVX512DQ-FCP-NEXT: vmovdqa64 2048(%rdi), %zmm012427; AVX512DQ-FCP-NEXT: vmovdqa64 2112(%rdi), %zmm212428; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm1812429; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm16, %zmm1812430; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm112431; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm12, %zmm112432; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12433; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm112434; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm13, %zmm112435; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12436; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm112437; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm14, %zmm112438; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12439; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm6, %zmm212440; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12441; AVX512DQ-FCP-NEXT: vmovdqa64 2368(%rdi), %zmm012442; AVX512DQ-FCP-NEXT: vmovdqa64 2432(%rdi), %zmm112443; AVX512DQ-FCP-NEXT: vpermi2q %zmm1, %zmm0, %zmm1212444; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12445; AVX512DQ-FCP-NEXT: vpermi2q %zmm1, %zmm0, %zmm1312446; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12447; AVX512DQ-FCP-NEXT: vpermi2q %zmm1, %zmm0, %zmm1412448; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12449; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, %zmm1712450; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm6, %zmm112451; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12452; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm16, %zmm1712453; AVX512DQ-FCP-NEXT: vmovdqa64 384(%rdi), %zmm3112454; AVX512DQ-FCP-NEXT: vmovdqa64 320(%rdi), %zmm212455; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} ymm3 = [0,5,10,15]12456; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm112457; AVX512DQ-FCP-NEXT: vpermt2q %zmm31, %zmm3, %zmm112458; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} ymm29 = [1,6,11,0]12459; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm012460; AVX512DQ-FCP-NEXT: vpermt2q %zmm31, %zmm29, %zmm012461; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12462; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} ymm6 = [2,7,12,0]12463; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm012464; AVX512DQ-FCP-NEXT: vpermt2q %zmm31, %zmm6, %zmm012465; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12466; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} ymm7 = [11,0,5,0]12467; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm31, %zmm012468; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm7, %zmm012469; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12470; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm16, %zmm3112471; AVX512DQ-FCP-NEXT: vmovdqa64 (%rdi), %zmm212472; AVX512DQ-FCP-NEXT: vmovdqa64 64(%rdi), %zmm2312473; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm412474; AVX512DQ-FCP-NEXT: vpermt2q %zmm23, %zmm3, %zmm412475; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm012476; AVX512DQ-FCP-NEXT: vpermt2q %zmm23, %zmm29, %zmm012477; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12478; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm012479; AVX512DQ-FCP-NEXT: vpermt2q %zmm23, %zmm6, %zmm012480; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12481; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm23, %zmm012482; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm7, %zmm012483; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12484; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm16, %zmm2312485; AVX512DQ-FCP-NEXT: vmovdqa64 1024(%rdi), %zmm2112486; AVX512DQ-FCP-NEXT: vmovdqa64 960(%rdi), %zmm212487; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm1012488; AVX512DQ-FCP-NEXT: vpermt2q %zmm21, %zmm3, %zmm1012489; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm012490; AVX512DQ-FCP-NEXT: vpermt2q %zmm21, %zmm29, %zmm012491; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12492; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm012493; AVX512DQ-FCP-NEXT: vpermt2q %zmm21, %zmm6, %zmm012494; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12495; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm21, %zmm012496; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm7, %zmm012497; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12498; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm16, %zmm2112499; AVX512DQ-FCP-NEXT: vmovdqa64 704(%rdi), %zmm2012500; AVX512DQ-FCP-NEXT: vmovdqa64 640(%rdi), %zmm212501; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm1112502; AVX512DQ-FCP-NEXT: vpermt2q %zmm20, %zmm3, %zmm1112503; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm012504; AVX512DQ-FCP-NEXT: vpermt2q %zmm20, %zmm29, %zmm012505; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12506; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm012507; AVX512DQ-FCP-NEXT: vpermt2q %zmm20, %zmm6, %zmm012508; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12509; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm20, %zmm012510; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm7, %zmm012511; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12512; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm16, %zmm2012513; AVX512DQ-FCP-NEXT: vmovdqa64 1664(%rdi), %zmm2512514; AVX512DQ-FCP-NEXT: vmovdqa64 1600(%rdi), %zmm212515; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm1312516; AVX512DQ-FCP-NEXT: vpermt2q %zmm25, %zmm3, %zmm1312517; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm012518; AVX512DQ-FCP-NEXT: vpermt2q %zmm25, %zmm29, %zmm012519; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12520; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm3012521; AVX512DQ-FCP-NEXT: vpermt2q %zmm25, %zmm6, %zmm3012522; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm25, %zmm012523; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm7, %zmm012524; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12525; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm16, %zmm2512526; AVX512DQ-FCP-NEXT: vmovdqa64 1344(%rdi), %zmm1212527; AVX512DQ-FCP-NEXT: vmovdqa64 1280(%rdi), %zmm212528; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm1412529; AVX512DQ-FCP-NEXT: vpermt2q %zmm12, %zmm3, %zmm1412530; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm2612531; AVX512DQ-FCP-NEXT: vpermt2q %zmm12, %zmm29, %zmm2612532; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm2712533; AVX512DQ-FCP-NEXT: vpermt2q %zmm12, %zmm6, %zmm2712534; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm12, %zmm2812535; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm7, %zmm2812536; AVX512DQ-FCP-NEXT: vpermt2q %zmm2, %zmm16, %zmm1212537; AVX512DQ-FCP-NEXT: vmovdqa64 1984(%rdi), %zmm912538; AVX512DQ-FCP-NEXT: vmovdqa64 1920(%rdi), %zmm1512539; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm15, %zmm212540; AVX512DQ-FCP-NEXT: vpermt2q %zmm9, %zmm3, %zmm212541; AVX512DQ-FCP-NEXT: vmovdqa64 2304(%rdi), %zmm512542; AVX512DQ-FCP-NEXT: vmovdqa64 2240(%rdi), %zmm012543; AVX512DQ-FCP-NEXT: vpermi2q %zmm5, %zmm0, %zmm312544; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm15, %zmm812545; AVX512DQ-FCP-NEXT: vpermt2q %zmm9, %zmm29, %zmm812546; AVX512DQ-FCP-NEXT: vpermi2q %zmm5, %zmm0, %zmm2912547; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm15, %zmm2212548; AVX512DQ-FCP-NEXT: vpermt2q %zmm9, %zmm6, %zmm2212549; AVX512DQ-FCP-NEXT: vpermi2q %zmm5, %zmm0, %zmm612550; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm9, %zmm2412551; AVX512DQ-FCP-NEXT: vpermt2q %zmm15, %zmm7, %zmm2412552; AVX512DQ-FCP-NEXT: vpermi2q %zmm0, %zmm5, %zmm712553; AVX512DQ-FCP-NEXT: vpermt2q %zmm0, %zmm16, %zmm512554; AVX512DQ-FCP-NEXT: vpermt2q %zmm15, %zmm16, %zmm912555; AVX512DQ-FCP-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm0 # 64-byte Folded Reload12556; AVX512DQ-FCP-NEXT: # zmm0 = zmm4[0,1,2,3],mem[4,5,6,7]12557; AVX512DQ-FCP-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm4 # 64-byte Folded Reload12558; AVX512DQ-FCP-NEXT: # zmm4 = zmm1[0,1,2,3],mem[4,5,6,7]12559; AVX512DQ-FCP-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm11, %zmm11 # 64-byte Folded Reload12560; AVX512DQ-FCP-NEXT: # zmm11 = zmm11[0,1,2,3],mem[4,5,6,7]12561; AVX512DQ-FCP-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm10, %zmm10 # 64-byte Folded Reload12562; AVX512DQ-FCP-NEXT: # zmm10 = zmm10[0,1,2,3],mem[4,5,6,7]12563; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm19 = zmm14[0,1,2,3],zmm19[4,5,6,7]12564; AVX512DQ-FCP-NEXT: vshufi64x2 $228, (%rsp), %zmm13, %zmm1 # 64-byte Folded Reload12565; AVX512DQ-FCP-NEXT: # zmm1 = zmm13[0,1,2,3],mem[4,5,6,7]12566; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm2[0,1,2,3],zmm18[4,5,6,7]12567; AVX512DQ-FCP-NEXT: vshufi64x2 {{.*#+}} zmm3 = zmm3[0,1,2,3],zmm17[4,5,6,7]12568; AVX512DQ-FCP-NEXT: vmovdqa64 256(%rdi), %zmm1812569; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm15 = [0,1,2,3,4,5,6,11]12570; AVX512DQ-FCP-NEXT: vpermt2q %zmm18, %zmm15, %zmm012571; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12572; AVX512DQ-FCP-NEXT: vmovdqa64 576(%rdi), %zmm1412573; AVX512DQ-FCP-NEXT: vpermt2q %zmm14, %zmm15, %zmm412574; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12575; AVX512DQ-FCP-NEXT: vmovdqa64 896(%rdi), %zmm1712576; AVX512DQ-FCP-NEXT: vpermt2q %zmm17, %zmm15, %zmm1112577; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm11, (%rsp) # 64-byte Spill12578; AVX512DQ-FCP-NEXT: vmovdqa64 1216(%rdi), %zmm1312579; AVX512DQ-FCP-NEXT: vpermt2q %zmm13, %zmm15, %zmm1012580; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12581; AVX512DQ-FCP-NEXT: vmovdqa64 1536(%rdi), %zmm1612582; AVX512DQ-FCP-NEXT: vpermt2q %zmm16, %zmm15, %zmm1912583; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm19, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12584; AVX512DQ-FCP-NEXT: vmovdqa64 1856(%rdi), %zmm1112585; AVX512DQ-FCP-NEXT: vpermt2q %zmm11, %zmm15, %zmm112586; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12587; AVX512DQ-FCP-NEXT: vmovdqa64 2176(%rdi), %zmm1912588; AVX512DQ-FCP-NEXT: vpermt2q %zmm19, %zmm15, %zmm212589; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12590; AVX512DQ-FCP-NEXT: vmovdqa64 2496(%rdi), %zmm1012591; AVX512DQ-FCP-NEXT: vpermt2q %zmm10, %zmm15, %zmm312592; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12593; AVX512DQ-FCP-NEXT: movb $7, %al12594; AVX512DQ-FCP-NEXT: kmovw %eax, %k112595; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload12596; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12597; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm15 {%k1}12598; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload12599; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm8, %zmm4 {%k1}12600; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload12601; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12602; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, %zmm8 {%k1}12603; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12604; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload12605; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, %zmm0 {%k1}12606; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload12607; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12608; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm1 {%k1}12609; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12610; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload12611; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, %zmm2 {%k1}12612; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload12613; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm26, %zmm3 {%k1}12614; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload12615; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm29, %zmm26 {%k1}12616; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm29 = [0,1,2,3,4,5,6,12]12617; AVX512DQ-FCP-NEXT: vpermt2q %zmm18, %zmm29, %zmm1512618; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12619; AVX512DQ-FCP-NEXT: vpermt2q %zmm19, %zmm29, %zmm412620; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12621; AVX512DQ-FCP-NEXT: vpermt2q %zmm14, %zmm29, %zmm812622; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12623; AVX512DQ-FCP-NEXT: vpermt2q %zmm13, %zmm29, %zmm012624; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12625; AVX512DQ-FCP-NEXT: vpermt2q %zmm17, %zmm29, %zmm112626; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12627; AVX512DQ-FCP-NEXT: vpermt2q %zmm11, %zmm29, %zmm212628; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12629; AVX512DQ-FCP-NEXT: vpermt2q %zmm16, %zmm29, %zmm312630; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12631; AVX512DQ-FCP-NEXT: vpermt2q %zmm10, %zmm29, %zmm2612632; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm26, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12633; AVX512DQ-FCP-NEXT: movb $56, %al12634; AVX512DQ-FCP-NEXT: kmovw %eax, %k112635; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12636; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload12637; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm0 {%k1}12638; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload12639; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, %zmm22 {%k1}12640; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12641; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload12642; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm8 {%k1}12643; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12644; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload12645; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm26 {%k1}12646; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12647; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload12648; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm15 {%k1}12649; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12650; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm30 {%k1}12651; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload12652; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, %zmm27 {%k1}12653; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload12654; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, %zmm6 {%k1}12655; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm1 = [0,1,2,3,4,5,8,13]12656; AVX512DQ-FCP-NEXT: vpermt2q %zmm18, %zmm1, %zmm012657; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12658; AVX512DQ-FCP-NEXT: vpermt2q %zmm19, %zmm1, %zmm2212659; AVX512DQ-FCP-NEXT: vpermt2q %zmm14, %zmm1, %zmm812660; AVX512DQ-FCP-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12661; AVX512DQ-FCP-NEXT: vpermt2q %zmm13, %zmm1, %zmm2612662; AVX512DQ-FCP-NEXT: vpermt2q %zmm17, %zmm1, %zmm1512663; AVX512DQ-FCP-NEXT: vpermt2q %zmm11, %zmm1, %zmm3012664; AVX512DQ-FCP-NEXT: vpermt2q %zmm16, %zmm1, %zmm2712665; AVX512DQ-FCP-NEXT: vpermt2q %zmm10, %zmm1, %zmm612666; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12667; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload12668; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm8 {%k1}12669; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload12670; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, %zmm24 {%k1}12671; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12672; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 # 64-byte Reload12673; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm29 {%k1}12674; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12675; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload12676; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm1 {%k1}12677; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12678; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload12679; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm3 {%k1}12680; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12681; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload12682; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm4 {%k1}12683; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12684; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm28 {%k1}12685; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12686; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm7 {%k1}12687; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm2 = [0,1,2,3,4,5,9,14]12688; AVX512DQ-FCP-NEXT: vpermt2q %zmm18, %zmm2, %zmm812689; AVX512DQ-FCP-NEXT: vpermt2q %zmm19, %zmm2, %zmm2412690; AVX512DQ-FCP-NEXT: vpermt2q %zmm14, %zmm2, %zmm2912691; AVX512DQ-FCP-NEXT: vpermt2q %zmm13, %zmm2, %zmm112692; AVX512DQ-FCP-NEXT: vpermt2q %zmm17, %zmm2, %zmm312693; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm3, %zmm012694; AVX512DQ-FCP-NEXT: vpermt2q %zmm11, %zmm2, %zmm412695; AVX512DQ-FCP-NEXT: vpermt2q %zmm16, %zmm2, %zmm2812696; AVX512DQ-FCP-NEXT: vpermt2q %zmm10, %zmm2, %zmm712697; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12698; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm31 {%k1}12699; AVX512DQ-FCP-NEXT: vpmovsxbq {{.*#+}} zmm3 = [0,1,2,3,4,5,10,15]12700; AVX512DQ-FCP-NEXT: vpermt2q %zmm14, %zmm3, %zmm3112701; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12702; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm23 {%k1}12703; AVX512DQ-FCP-NEXT: vpermt2q %zmm18, %zmm3, %zmm2312704; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12705; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm21 {%k1}12706; AVX512DQ-FCP-NEXT: vpermt2q %zmm13, %zmm3, %zmm2112707; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12708; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm20 {%k1}12709; AVX512DQ-FCP-NEXT: vpermt2q %zmm17, %zmm3, %zmm2012710; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12711; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm25 {%k1}12712; AVX512DQ-FCP-NEXT: vpermt2q %zmm11, %zmm3, %zmm2512713; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12714; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm12 {%k1}12715; AVX512DQ-FCP-NEXT: vpermt2q %zmm16, %zmm3, %zmm1212716; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12717; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm5 {%k1}12718; AVX512DQ-FCP-NEXT: vpermt2q %zmm10, %zmm3, %zmm512719; AVX512DQ-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12720; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm2, %zmm9 {%k1}12721; AVX512DQ-FCP-NEXT: vpermt2q %zmm19, %zmm3, %zmm912722; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12723; AVX512DQ-FCP-NEXT: vmovaps %zmm2, 448(%rsi)12724; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12725; AVX512DQ-FCP-NEXT: vmovaps %zmm2, 384(%rsi)12726; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12727; AVX512DQ-FCP-NEXT: vmovaps %zmm2, 320(%rsi)12728; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12729; AVX512DQ-FCP-NEXT: vmovaps %zmm2, 256(%rsi)12730; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12731; AVX512DQ-FCP-NEXT: vmovaps %zmm2, 192(%rsi)12732; AVX512DQ-FCP-NEXT: vmovups (%rsp), %zmm2 # 64-byte Reload12733; AVX512DQ-FCP-NEXT: vmovaps %zmm2, 128(%rsi)12734; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12735; AVX512DQ-FCP-NEXT: vmovaps %zmm2, 64(%rsi)12736; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12737; AVX512DQ-FCP-NEXT: vmovaps %zmm2, (%rsi)12738; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12739; AVX512DQ-FCP-NEXT: vmovaps %zmm2, 448(%rdx)12740; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12741; AVX512DQ-FCP-NEXT: vmovaps %zmm2, 256(%rdx)12742; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12743; AVX512DQ-FCP-NEXT: vmovaps %zmm2, 320(%rdx)12744; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12745; AVX512DQ-FCP-NEXT: vmovaps %zmm2, 128(%rdx)12746; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12747; AVX512DQ-FCP-NEXT: vmovaps %zmm2, 192(%rdx)12748; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12749; AVX512DQ-FCP-NEXT: vmovaps %zmm2, (%rdx)12750; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12751; AVX512DQ-FCP-NEXT: vmovaps %zmm2, 64(%rdx)12752; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12753; AVX512DQ-FCP-NEXT: vmovaps %zmm2, 384(%rdx)12754; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm6, 448(%rcx)12755; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm27, 256(%rcx)12756; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm30, 320(%rcx)12757; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm15, 128(%rcx)12758; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm26, 192(%rcx)12759; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12760; AVX512DQ-FCP-NEXT: vmovaps %zmm2, (%rcx)12761; AVX512DQ-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload12762; AVX512DQ-FCP-NEXT: vmovaps %zmm2, 64(%rcx)12763; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm22, 384(%rcx)12764; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm7, 448(%r8)12765; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm28, 256(%r8)12766; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm4, 320(%r8)12767; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm0, 128(%r8)12768; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm1, 192(%r8)12769; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm8, (%r8)12770; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm29, 64(%r8)12771; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm24, 384(%r8)12772; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm9, 384(%r9)12773; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm5, 448(%r9)12774; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm12, 256(%r9)12775; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm25, 320(%r9)12776; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm20, 128(%r9)12777; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm21, 192(%r9)12778; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm23, (%r9)12779; AVX512DQ-FCP-NEXT: vmovdqa64 %zmm31, 64(%r9)12780; AVX512DQ-FCP-NEXT: addq $3336, %rsp # imm = 0xD0812781; AVX512DQ-FCP-NEXT: vzeroupper12782; AVX512DQ-FCP-NEXT: retq12783;12784; AVX512BW-LABEL: load_i64_stride5_vf64:12785; AVX512BW: # %bb.0:12786; AVX512BW-NEXT: subq $3336, %rsp # imm = 0xD0812787; AVX512BW-NEXT: vmovdqa64 1728(%rdi), %zmm312788; AVX512BW-NEXT: vmovdqa64 1792(%rdi), %zmm712789; AVX512BW-NEXT: vmovdqa64 1408(%rdi), %zmm012790; AVX512BW-NEXT: vmovdqa64 1088(%rdi), %zmm112791; AVX512BW-NEXT: vmovdqa64 1152(%rdi), %zmm612792; AVX512BW-NEXT: vmovdqa64 768(%rdi), %zmm512793; AVX512BW-NEXT: vmovdqa64 832(%rdi), %zmm812794; AVX512BW-NEXT: vmovdqa64 448(%rdi), %zmm412795; AVX512BW-NEXT: vmovdqa64 512(%rdi), %zmm912796; AVX512BW-NEXT: vmovdqa64 128(%rdi), %zmm212797; AVX512BW-NEXT: vmovdqa64 192(%rdi), %zmm1012798; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm16 = [12,1,6,0,12,1,6,0]12799; AVX512BW-NEXT: # zmm16 = mem[0,1,2,3,0,1,2,3]12800; AVX512BW-NEXT: vmovdqa64 %zmm10, %zmm1112801; AVX512BW-NEXT: vpermt2q %zmm2, %zmm16, %zmm1112802; AVX512BW-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12803; AVX512BW-NEXT: vmovdqa64 %zmm9, %zmm1112804; AVX512BW-NEXT: vpermt2q %zmm4, %zmm16, %zmm1112805; AVX512BW-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12806; AVX512BW-NEXT: vmovdqa64 %zmm8, %zmm1112807; AVX512BW-NEXT: vpermt2q %zmm5, %zmm16, %zmm1112808; AVX512BW-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12809; AVX512BW-NEXT: vmovdqa64 %zmm6, %zmm1212810; AVX512BW-NEXT: vmovdqa64 %zmm6, %zmm1112811; AVX512BW-NEXT: vpermt2q %zmm1, %zmm16, %zmm1212812; AVX512BW-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12813; AVX512BW-NEXT: vmovdqa64 %zmm7, %zmm612814; AVX512BW-NEXT: vpermt2q %zmm3, %zmm16, %zmm612815; AVX512BW-NEXT: vmovdqu64 %zmm6, (%rsp) # 64-byte Spill12816; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm12 = [5,10,15,0,5,10,15,0]12817; AVX512BW-NEXT: # zmm12 = mem[0,1,2,3,0,1,2,3]12818; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm612819; AVX512BW-NEXT: vpermt2q %zmm10, %zmm12, %zmm612820; AVX512BW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12821; AVX512BW-NEXT: vmovdqa64 %zmm4, %zmm612822; AVX512BW-NEXT: vpermt2q %zmm9, %zmm12, %zmm612823; AVX512BW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12824; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm612825; AVX512BW-NEXT: vpermt2q %zmm11, %zmm12, %zmm612826; AVX512BW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12827; AVX512BW-NEXT: vmovdqa64 %zmm5, %zmm612828; AVX512BW-NEXT: vpermt2q %zmm8, %zmm12, %zmm612829; AVX512BW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12830; AVX512BW-NEXT: vmovdqa64 %zmm3, %zmm612831; AVX512BW-NEXT: vpermt2q %zmm7, %zmm12, %zmm612832; AVX512BW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12833; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = [6,11,0,1,6,11,0,1]12834; AVX512BW-NEXT: # zmm6 = mem[0,1,2,3,0,1,2,3]12835; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm1312836; AVX512BW-NEXT: vpermt2q %zmm10, %zmm6, %zmm1312837; AVX512BW-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12838; AVX512BW-NEXT: vmovdqa64 %zmm6, %zmm1312839; AVX512BW-NEXT: vmovdqa64 %zmm4, %zmm612840; AVX512BW-NEXT: vpermt2q %zmm9, %zmm13, %zmm612841; AVX512BW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12842; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm612843; AVX512BW-NEXT: vpermt2q %zmm11, %zmm13, %zmm612844; AVX512BW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12845; AVX512BW-NEXT: vmovdqa64 %zmm5, %zmm612846; AVX512BW-NEXT: vpermt2q %zmm8, %zmm13, %zmm612847; AVX512BW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12848; AVX512BW-NEXT: vmovdqa64 %zmm3, %zmm612849; AVX512BW-NEXT: vpermt2q %zmm7, %zmm13, %zmm612850; AVX512BW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12851; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm14 = [7,12,0,2,7,12,0,2]12852; AVX512BW-NEXT: # zmm14 = mem[0,1,2,3,0,1,2,3]12853; AVX512BW-NEXT: vmovdqa64 %zmm4, %zmm612854; AVX512BW-NEXT: vpermt2q %zmm9, %zmm14, %zmm612855; AVX512BW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12856; AVX512BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = [0,5,0,11,0,5,0,11]12857; AVX512BW-NEXT: # zmm6 = mem[0,1,2,3,0,1,2,3]12858; AVX512BW-NEXT: vpermt2q %zmm4, %zmm6, %zmm912859; AVX512BW-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12860; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm412861; AVX512BW-NEXT: vpermt2q %zmm10, %zmm14, %zmm412862; AVX512BW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12863; AVX512BW-NEXT: vpermt2q %zmm2, %zmm6, %zmm1012864; AVX512BW-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12865; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm212866; AVX512BW-NEXT: vpermt2q %zmm11, %zmm14, %zmm212867; AVX512BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12868; AVX512BW-NEXT: vpermt2q %zmm1, %zmm6, %zmm1112869; AVX512BW-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12870; AVX512BW-NEXT: vmovdqa64 %zmm5, %zmm112871; AVX512BW-NEXT: vpermt2q %zmm8, %zmm14, %zmm112872; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12873; AVX512BW-NEXT: vpermt2q %zmm5, %zmm6, %zmm812874; AVX512BW-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12875; AVX512BW-NEXT: vmovdqa64 %zmm3, %zmm112876; AVX512BW-NEXT: vpermt2q %zmm7, %zmm14, %zmm112877; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12878; AVX512BW-NEXT: vpermt2q %zmm3, %zmm6, %zmm712879; AVX512BW-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12880; AVX512BW-NEXT: vmovdqa64 1472(%rdi), %zmm212881; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm1912882; AVX512BW-NEXT: vpermt2q %zmm0, %zmm16, %zmm1912883; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm112884; AVX512BW-NEXT: vpermt2q %zmm2, %zmm12, %zmm112885; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12886; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm112887; AVX512BW-NEXT: vpermt2q %zmm2, %zmm13, %zmm112888; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12889; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm112890; AVX512BW-NEXT: vpermt2q %zmm2, %zmm14, %zmm112891; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12892; AVX512BW-NEXT: vpermt2q %zmm0, %zmm6, %zmm212893; AVX512BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12894; AVX512BW-NEXT: vmovdqa64 2048(%rdi), %zmm012895; AVX512BW-NEXT: vmovdqa64 2112(%rdi), %zmm212896; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm1812897; AVX512BW-NEXT: vpermt2q %zmm0, %zmm16, %zmm1812898; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm112899; AVX512BW-NEXT: vpermt2q %zmm2, %zmm12, %zmm112900; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12901; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm112902; AVX512BW-NEXT: vpermt2q %zmm2, %zmm13, %zmm112903; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12904; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm112905; AVX512BW-NEXT: vpermt2q %zmm2, %zmm14, %zmm112906; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12907; AVX512BW-NEXT: vpermt2q %zmm0, %zmm6, %zmm212908; AVX512BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12909; AVX512BW-NEXT: vmovdqa64 2368(%rdi), %zmm012910; AVX512BW-NEXT: vmovdqa64 2432(%rdi), %zmm112911; AVX512BW-NEXT: vpermi2q %zmm1, %zmm0, %zmm1212912; AVX512BW-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12913; AVX512BW-NEXT: vpermi2q %zmm1, %zmm0, %zmm1312914; AVX512BW-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12915; AVX512BW-NEXT: vpermi2q %zmm1, %zmm0, %zmm1412916; AVX512BW-NEXT: vmovdqu64 %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12917; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm1712918; AVX512BW-NEXT: vpermt2q %zmm0, %zmm6, %zmm112919; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12920; AVX512BW-NEXT: vpermt2q %zmm0, %zmm16, %zmm1712921; AVX512BW-NEXT: vmovdqa64 384(%rdi), %zmm3112922; AVX512BW-NEXT: vmovdqa64 320(%rdi), %zmm212923; AVX512BW-NEXT: vpmovsxbq {{.*#+}} ymm3 = [0,5,10,15]12924; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm112925; AVX512BW-NEXT: vpermt2q %zmm31, %zmm3, %zmm112926; AVX512BW-NEXT: vpmovsxbq {{.*#+}} ymm29 = [1,6,11,0]12927; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm012928; AVX512BW-NEXT: vpermt2q %zmm31, %zmm29, %zmm012929; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12930; AVX512BW-NEXT: vpmovsxbq {{.*#+}} ymm6 = [2,7,12,0]12931; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm012932; AVX512BW-NEXT: vpermt2q %zmm31, %zmm6, %zmm012933; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12934; AVX512BW-NEXT: vpmovsxbq {{.*#+}} ymm7 = [11,0,5,0]12935; AVX512BW-NEXT: vmovdqa64 %zmm31, %zmm012936; AVX512BW-NEXT: vpermt2q %zmm2, %zmm7, %zmm012937; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12938; AVX512BW-NEXT: vpermt2q %zmm2, %zmm16, %zmm3112939; AVX512BW-NEXT: vmovdqa64 (%rdi), %zmm212940; AVX512BW-NEXT: vmovdqa64 64(%rdi), %zmm2312941; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm412942; AVX512BW-NEXT: vpermt2q %zmm23, %zmm3, %zmm412943; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm012944; AVX512BW-NEXT: vpermt2q %zmm23, %zmm29, %zmm012945; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12946; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm012947; AVX512BW-NEXT: vpermt2q %zmm23, %zmm6, %zmm012948; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12949; AVX512BW-NEXT: vmovdqa64 %zmm23, %zmm012950; AVX512BW-NEXT: vpermt2q %zmm2, %zmm7, %zmm012951; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12952; AVX512BW-NEXT: vpermt2q %zmm2, %zmm16, %zmm2312953; AVX512BW-NEXT: vmovdqa64 1024(%rdi), %zmm2112954; AVX512BW-NEXT: vmovdqa64 960(%rdi), %zmm212955; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm1012956; AVX512BW-NEXT: vpermt2q %zmm21, %zmm3, %zmm1012957; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm012958; AVX512BW-NEXT: vpermt2q %zmm21, %zmm29, %zmm012959; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12960; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm012961; AVX512BW-NEXT: vpermt2q %zmm21, %zmm6, %zmm012962; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12963; AVX512BW-NEXT: vmovdqa64 %zmm21, %zmm012964; AVX512BW-NEXT: vpermt2q %zmm2, %zmm7, %zmm012965; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12966; AVX512BW-NEXT: vpermt2q %zmm2, %zmm16, %zmm2112967; AVX512BW-NEXT: vmovdqa64 704(%rdi), %zmm2012968; AVX512BW-NEXT: vmovdqa64 640(%rdi), %zmm212969; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm1112970; AVX512BW-NEXT: vpermt2q %zmm20, %zmm3, %zmm1112971; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm012972; AVX512BW-NEXT: vpermt2q %zmm20, %zmm29, %zmm012973; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12974; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm012975; AVX512BW-NEXT: vpermt2q %zmm20, %zmm6, %zmm012976; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12977; AVX512BW-NEXT: vmovdqa64 %zmm20, %zmm012978; AVX512BW-NEXT: vpermt2q %zmm2, %zmm7, %zmm012979; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12980; AVX512BW-NEXT: vpermt2q %zmm2, %zmm16, %zmm2012981; AVX512BW-NEXT: vmovdqa64 1664(%rdi), %zmm2512982; AVX512BW-NEXT: vmovdqa64 1600(%rdi), %zmm212983; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm1312984; AVX512BW-NEXT: vpermt2q %zmm25, %zmm3, %zmm1312985; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm012986; AVX512BW-NEXT: vpermt2q %zmm25, %zmm29, %zmm012987; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12988; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm3012989; AVX512BW-NEXT: vpermt2q %zmm25, %zmm6, %zmm3012990; AVX512BW-NEXT: vmovdqa64 %zmm25, %zmm012991; AVX512BW-NEXT: vpermt2q %zmm2, %zmm7, %zmm012992; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill12993; AVX512BW-NEXT: vpermt2q %zmm2, %zmm16, %zmm2512994; AVX512BW-NEXT: vmovdqa64 1344(%rdi), %zmm1212995; AVX512BW-NEXT: vmovdqa64 1280(%rdi), %zmm212996; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm1412997; AVX512BW-NEXT: vpermt2q %zmm12, %zmm3, %zmm1412998; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm2612999; AVX512BW-NEXT: vpermt2q %zmm12, %zmm29, %zmm2613000; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm2713001; AVX512BW-NEXT: vpermt2q %zmm12, %zmm6, %zmm2713002; AVX512BW-NEXT: vmovdqa64 %zmm12, %zmm2813003; AVX512BW-NEXT: vpermt2q %zmm2, %zmm7, %zmm2813004; AVX512BW-NEXT: vpermt2q %zmm2, %zmm16, %zmm1213005; AVX512BW-NEXT: vmovdqa64 1984(%rdi), %zmm913006; AVX512BW-NEXT: vmovdqa64 1920(%rdi), %zmm1513007; AVX512BW-NEXT: vmovdqa64 %zmm15, %zmm213008; AVX512BW-NEXT: vpermt2q %zmm9, %zmm3, %zmm213009; AVX512BW-NEXT: vmovdqa64 2304(%rdi), %zmm513010; AVX512BW-NEXT: vmovdqa64 2240(%rdi), %zmm013011; AVX512BW-NEXT: vpermi2q %zmm5, %zmm0, %zmm313012; AVX512BW-NEXT: vmovdqa64 %zmm15, %zmm813013; AVX512BW-NEXT: vpermt2q %zmm9, %zmm29, %zmm813014; AVX512BW-NEXT: vpermi2q %zmm5, %zmm0, %zmm2913015; AVX512BW-NEXT: vmovdqa64 %zmm15, %zmm2213016; AVX512BW-NEXT: vpermt2q %zmm9, %zmm6, %zmm2213017; AVX512BW-NEXT: vpermi2q %zmm5, %zmm0, %zmm613018; AVX512BW-NEXT: vmovdqa64 %zmm9, %zmm2413019; AVX512BW-NEXT: vpermt2q %zmm15, %zmm7, %zmm2413020; AVX512BW-NEXT: vpermi2q %zmm0, %zmm5, %zmm713021; AVX512BW-NEXT: vpermt2q %zmm0, %zmm16, %zmm513022; AVX512BW-NEXT: vpermt2q %zmm15, %zmm16, %zmm913023; AVX512BW-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm0 # 64-byte Folded Reload13024; AVX512BW-NEXT: # zmm0 = zmm4[0,1,2,3],mem[4,5,6,7]13025; AVX512BW-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm4 # 64-byte Folded Reload13026; AVX512BW-NEXT: # zmm4 = zmm1[0,1,2,3],mem[4,5,6,7]13027; AVX512BW-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm11, %zmm11 # 64-byte Folded Reload13028; AVX512BW-NEXT: # zmm11 = zmm11[0,1,2,3],mem[4,5,6,7]13029; AVX512BW-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm10, %zmm10 # 64-byte Folded Reload13030; AVX512BW-NEXT: # zmm10 = zmm10[0,1,2,3],mem[4,5,6,7]13031; AVX512BW-NEXT: vshufi64x2 {{.*#+}} zmm19 = zmm14[0,1,2,3],zmm19[4,5,6,7]13032; AVX512BW-NEXT: vshufi64x2 $228, (%rsp), %zmm13, %zmm1 # 64-byte Folded Reload13033; AVX512BW-NEXT: # zmm1 = zmm13[0,1,2,3],mem[4,5,6,7]13034; AVX512BW-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm2[0,1,2,3],zmm18[4,5,6,7]13035; AVX512BW-NEXT: vshufi64x2 {{.*#+}} zmm3 = zmm3[0,1,2,3],zmm17[4,5,6,7]13036; AVX512BW-NEXT: vmovdqa64 256(%rdi), %zmm1813037; AVX512BW-NEXT: vpmovsxbq {{.*#+}} zmm15 = [0,1,2,3,4,5,6,11]13038; AVX512BW-NEXT: vpermt2q %zmm18, %zmm15, %zmm013039; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13040; AVX512BW-NEXT: vmovdqa64 576(%rdi), %zmm1413041; AVX512BW-NEXT: vpermt2q %zmm14, %zmm15, %zmm413042; AVX512BW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13043; AVX512BW-NEXT: vmovdqa64 896(%rdi), %zmm1713044; AVX512BW-NEXT: vpermt2q %zmm17, %zmm15, %zmm1113045; AVX512BW-NEXT: vmovdqu64 %zmm11, (%rsp) # 64-byte Spill13046; AVX512BW-NEXT: vmovdqa64 1216(%rdi), %zmm1313047; AVX512BW-NEXT: vpermt2q %zmm13, %zmm15, %zmm1013048; AVX512BW-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13049; AVX512BW-NEXT: vmovdqa64 1536(%rdi), %zmm1613050; AVX512BW-NEXT: vpermt2q %zmm16, %zmm15, %zmm1913051; AVX512BW-NEXT: vmovdqu64 %zmm19, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13052; AVX512BW-NEXT: vmovdqa64 1856(%rdi), %zmm1113053; AVX512BW-NEXT: vpermt2q %zmm11, %zmm15, %zmm113054; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13055; AVX512BW-NEXT: vmovdqa64 2176(%rdi), %zmm1913056; AVX512BW-NEXT: vpermt2q %zmm19, %zmm15, %zmm213057; AVX512BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13058; AVX512BW-NEXT: vmovdqa64 2496(%rdi), %zmm1013059; AVX512BW-NEXT: vpermt2q %zmm10, %zmm15, %zmm313060; AVX512BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13061; AVX512BW-NEXT: movb $7, %al13062; AVX512BW-NEXT: kmovd %eax, %k113063; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload13064; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13065; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm15 {%k1}13066; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload13067; AVX512BW-NEXT: vmovdqa64 %zmm8, %zmm4 {%k1}13068; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload13069; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13070; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm8 {%k1}13071; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13072; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload13073; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm0 {%k1}13074; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload13075; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13076; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm1 {%k1}13077; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13078; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload13079; AVX512BW-NEXT: vmovdqa64 %zmm3, %zmm2 {%k1}13080; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload13081; AVX512BW-NEXT: vmovdqa64 %zmm26, %zmm3 {%k1}13082; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload13083; AVX512BW-NEXT: vmovdqa64 %zmm29, %zmm26 {%k1}13084; AVX512BW-NEXT: vpmovsxbq {{.*#+}} zmm29 = [0,1,2,3,4,5,6,12]13085; AVX512BW-NEXT: vpermt2q %zmm18, %zmm29, %zmm1513086; AVX512BW-NEXT: vmovdqu64 %zmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13087; AVX512BW-NEXT: vpermt2q %zmm19, %zmm29, %zmm413088; AVX512BW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13089; AVX512BW-NEXT: vpermt2q %zmm14, %zmm29, %zmm813090; AVX512BW-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13091; AVX512BW-NEXT: vpermt2q %zmm13, %zmm29, %zmm013092; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13093; AVX512BW-NEXT: vpermt2q %zmm17, %zmm29, %zmm113094; AVX512BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13095; AVX512BW-NEXT: vpermt2q %zmm11, %zmm29, %zmm213096; AVX512BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13097; AVX512BW-NEXT: vpermt2q %zmm16, %zmm29, %zmm313098; AVX512BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13099; AVX512BW-NEXT: vpermt2q %zmm10, %zmm29, %zmm2613100; AVX512BW-NEXT: vmovdqu64 %zmm26, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13101; AVX512BW-NEXT: movb $56, %al13102; AVX512BW-NEXT: kmovd %eax, %k113103; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13104; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13105; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm0 {%k1}13106; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload13107; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm22 {%k1}13108; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13109; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload13110; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm8 {%k1}13111; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13112; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload13113; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm26 {%k1}13114; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13115; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload13116; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm15 {%k1}13117; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13118; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm30 {%k1}13119; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload13120; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm27 {%k1}13121; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload13122; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm6 {%k1}13123; AVX512BW-NEXT: vpmovsxbq {{.*#+}} zmm1 = [0,1,2,3,4,5,8,13]13124; AVX512BW-NEXT: vpermt2q %zmm18, %zmm1, %zmm013125; AVX512BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13126; AVX512BW-NEXT: vpermt2q %zmm19, %zmm1, %zmm2213127; AVX512BW-NEXT: vpermt2q %zmm14, %zmm1, %zmm813128; AVX512BW-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13129; AVX512BW-NEXT: vpermt2q %zmm13, %zmm1, %zmm2613130; AVX512BW-NEXT: vpermt2q %zmm17, %zmm1, %zmm1513131; AVX512BW-NEXT: vpermt2q %zmm11, %zmm1, %zmm3013132; AVX512BW-NEXT: vpermt2q %zmm16, %zmm1, %zmm2713133; AVX512BW-NEXT: vpermt2q %zmm10, %zmm1, %zmm613134; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13135; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload13136; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm8 {%k1}13137; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload13138; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm24 {%k1}13139; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13140; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 # 64-byte Reload13141; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm29 {%k1}13142; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13143; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload13144; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm1 {%k1}13145; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13146; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload13147; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm3 {%k1}13148; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13149; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload13150; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm4 {%k1}13151; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13152; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm28 {%k1}13153; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13154; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm7 {%k1}13155; AVX512BW-NEXT: vpmovsxbq {{.*#+}} zmm2 = [0,1,2,3,4,5,9,14]13156; AVX512BW-NEXT: vpermt2q %zmm18, %zmm2, %zmm813157; AVX512BW-NEXT: vpermt2q %zmm19, %zmm2, %zmm2413158; AVX512BW-NEXT: vpermt2q %zmm14, %zmm2, %zmm2913159; AVX512BW-NEXT: vpermt2q %zmm13, %zmm2, %zmm113160; AVX512BW-NEXT: vpermt2q %zmm17, %zmm2, %zmm313161; AVX512BW-NEXT: vmovdqa64 %zmm3, %zmm013162; AVX512BW-NEXT: vpermt2q %zmm11, %zmm2, %zmm413163; AVX512BW-NEXT: vpermt2q %zmm16, %zmm2, %zmm2813164; AVX512BW-NEXT: vpermt2q %zmm10, %zmm2, %zmm713165; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13166; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm31 {%k1}13167; AVX512BW-NEXT: vpmovsxbq {{.*#+}} zmm3 = [0,1,2,3,4,5,10,15]13168; AVX512BW-NEXT: vpermt2q %zmm14, %zmm3, %zmm3113169; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13170; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm23 {%k1}13171; AVX512BW-NEXT: vpermt2q %zmm18, %zmm3, %zmm2313172; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13173; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm21 {%k1}13174; AVX512BW-NEXT: vpermt2q %zmm13, %zmm3, %zmm2113175; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13176; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm20 {%k1}13177; AVX512BW-NEXT: vpermt2q %zmm17, %zmm3, %zmm2013178; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13179; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm25 {%k1}13180; AVX512BW-NEXT: vpermt2q %zmm11, %zmm3, %zmm2513181; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13182; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm12 {%k1}13183; AVX512BW-NEXT: vpermt2q %zmm16, %zmm3, %zmm1213184; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13185; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm5 {%k1}13186; AVX512BW-NEXT: vpermt2q %zmm10, %zmm3, %zmm513187; AVX512BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13188; AVX512BW-NEXT: vmovdqa64 %zmm2, %zmm9 {%k1}13189; AVX512BW-NEXT: vpermt2q %zmm19, %zmm3, %zmm913190; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13191; AVX512BW-NEXT: vmovaps %zmm2, 448(%rsi)13192; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13193; AVX512BW-NEXT: vmovaps %zmm2, 384(%rsi)13194; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13195; AVX512BW-NEXT: vmovaps %zmm2, 320(%rsi)13196; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13197; AVX512BW-NEXT: vmovaps %zmm2, 256(%rsi)13198; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13199; AVX512BW-NEXT: vmovaps %zmm2, 192(%rsi)13200; AVX512BW-NEXT: vmovups (%rsp), %zmm2 # 64-byte Reload13201; AVX512BW-NEXT: vmovaps %zmm2, 128(%rsi)13202; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13203; AVX512BW-NEXT: vmovaps %zmm2, 64(%rsi)13204; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13205; AVX512BW-NEXT: vmovaps %zmm2, (%rsi)13206; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13207; AVX512BW-NEXT: vmovaps %zmm2, 448(%rdx)13208; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13209; AVX512BW-NEXT: vmovaps %zmm2, 256(%rdx)13210; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13211; AVX512BW-NEXT: vmovaps %zmm2, 320(%rdx)13212; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13213; AVX512BW-NEXT: vmovaps %zmm2, 128(%rdx)13214; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13215; AVX512BW-NEXT: vmovaps %zmm2, 192(%rdx)13216; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13217; AVX512BW-NEXT: vmovaps %zmm2, (%rdx)13218; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13219; AVX512BW-NEXT: vmovaps %zmm2, 64(%rdx)13220; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13221; AVX512BW-NEXT: vmovaps %zmm2, 384(%rdx)13222; AVX512BW-NEXT: vmovdqa64 %zmm6, 448(%rcx)13223; AVX512BW-NEXT: vmovdqa64 %zmm27, 256(%rcx)13224; AVX512BW-NEXT: vmovdqa64 %zmm30, 320(%rcx)13225; AVX512BW-NEXT: vmovdqa64 %zmm15, 128(%rcx)13226; AVX512BW-NEXT: vmovdqa64 %zmm26, 192(%rcx)13227; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13228; AVX512BW-NEXT: vmovaps %zmm2, (%rcx)13229; AVX512BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13230; AVX512BW-NEXT: vmovaps %zmm2, 64(%rcx)13231; AVX512BW-NEXT: vmovdqa64 %zmm22, 384(%rcx)13232; AVX512BW-NEXT: vmovdqa64 %zmm7, 448(%r8)13233; AVX512BW-NEXT: vmovdqa64 %zmm28, 256(%r8)13234; AVX512BW-NEXT: vmovdqa64 %zmm4, 320(%r8)13235; AVX512BW-NEXT: vmovdqa64 %zmm0, 128(%r8)13236; AVX512BW-NEXT: vmovdqa64 %zmm1, 192(%r8)13237; AVX512BW-NEXT: vmovdqa64 %zmm8, (%r8)13238; AVX512BW-NEXT: vmovdqa64 %zmm29, 64(%r8)13239; AVX512BW-NEXT: vmovdqa64 %zmm24, 384(%r8)13240; AVX512BW-NEXT: vmovdqa64 %zmm9, 384(%r9)13241; AVX512BW-NEXT: vmovdqa64 %zmm5, 448(%r9)13242; AVX512BW-NEXT: vmovdqa64 %zmm12, 256(%r9)13243; AVX512BW-NEXT: vmovdqa64 %zmm25, 320(%r9)13244; AVX512BW-NEXT: vmovdqa64 %zmm20, 128(%r9)13245; AVX512BW-NEXT: vmovdqa64 %zmm21, 192(%r9)13246; AVX512BW-NEXT: vmovdqa64 %zmm23, (%r9)13247; AVX512BW-NEXT: vmovdqa64 %zmm31, 64(%r9)13248; AVX512BW-NEXT: addq $3336, %rsp # imm = 0xD0813249; AVX512BW-NEXT: vzeroupper13250; AVX512BW-NEXT: retq13251;13252; AVX512BW-FCP-LABEL: load_i64_stride5_vf64:13253; AVX512BW-FCP: # %bb.0:13254; AVX512BW-FCP-NEXT: subq $3336, %rsp # imm = 0xD0813255; AVX512BW-FCP-NEXT: vmovdqa64 1728(%rdi), %zmm313256; AVX512BW-FCP-NEXT: vmovdqa64 1792(%rdi), %zmm713257; AVX512BW-FCP-NEXT: vmovdqa64 1408(%rdi), %zmm013258; AVX512BW-FCP-NEXT: vmovdqa64 1088(%rdi), %zmm113259; AVX512BW-FCP-NEXT: vmovdqa64 1152(%rdi), %zmm613260; AVX512BW-FCP-NEXT: vmovdqa64 768(%rdi), %zmm513261; AVX512BW-FCP-NEXT: vmovdqa64 832(%rdi), %zmm813262; AVX512BW-FCP-NEXT: vmovdqa64 448(%rdi), %zmm413263; AVX512BW-FCP-NEXT: vmovdqa64 512(%rdi), %zmm913264; AVX512BW-FCP-NEXT: vmovdqa64 128(%rdi), %zmm213265; AVX512BW-FCP-NEXT: vmovdqa64 192(%rdi), %zmm1013266; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm16 = [12,1,6,0,12,1,6,0]13267; AVX512BW-FCP-NEXT: # zmm16 = mem[0,1,2,3,0,1,2,3]13268; AVX512BW-FCP-NEXT: vmovdqa64 %zmm10, %zmm1113269; AVX512BW-FCP-NEXT: vpermt2q %zmm2, %zmm16, %zmm1113270; AVX512BW-FCP-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13271; AVX512BW-FCP-NEXT: vmovdqa64 %zmm9, %zmm1113272; AVX512BW-FCP-NEXT: vpermt2q %zmm4, %zmm16, %zmm1113273; AVX512BW-FCP-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13274; AVX512BW-FCP-NEXT: vmovdqa64 %zmm8, %zmm1113275; AVX512BW-FCP-NEXT: vpermt2q %zmm5, %zmm16, %zmm1113276; AVX512BW-FCP-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13277; AVX512BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm1213278; AVX512BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm1113279; AVX512BW-FCP-NEXT: vpermt2q %zmm1, %zmm16, %zmm1213280; AVX512BW-FCP-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13281; AVX512BW-FCP-NEXT: vmovdqa64 %zmm7, %zmm613282; AVX512BW-FCP-NEXT: vpermt2q %zmm3, %zmm16, %zmm613283; AVX512BW-FCP-NEXT: vmovdqu64 %zmm6, (%rsp) # 64-byte Spill13284; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm12 = [5,10,15,0,5,10,15,0]13285; AVX512BW-FCP-NEXT: # zmm12 = mem[0,1,2,3,0,1,2,3]13286; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm613287; AVX512BW-FCP-NEXT: vpermt2q %zmm10, %zmm12, %zmm613288; AVX512BW-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13289; AVX512BW-FCP-NEXT: vmovdqa64 %zmm4, %zmm613290; AVX512BW-FCP-NEXT: vpermt2q %zmm9, %zmm12, %zmm613291; AVX512BW-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13292; AVX512BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm613293; AVX512BW-FCP-NEXT: vpermt2q %zmm11, %zmm12, %zmm613294; AVX512BW-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13295; AVX512BW-FCP-NEXT: vmovdqa64 %zmm5, %zmm613296; AVX512BW-FCP-NEXT: vpermt2q %zmm8, %zmm12, %zmm613297; AVX512BW-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13298; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm613299; AVX512BW-FCP-NEXT: vpermt2q %zmm7, %zmm12, %zmm613300; AVX512BW-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13301; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = [6,11,0,1,6,11,0,1]13302; AVX512BW-FCP-NEXT: # zmm6 = mem[0,1,2,3,0,1,2,3]13303; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm1313304; AVX512BW-FCP-NEXT: vpermt2q %zmm10, %zmm6, %zmm1313305; AVX512BW-FCP-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13306; AVX512BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm1313307; AVX512BW-FCP-NEXT: vmovdqa64 %zmm4, %zmm613308; AVX512BW-FCP-NEXT: vpermt2q %zmm9, %zmm13, %zmm613309; AVX512BW-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13310; AVX512BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm613311; AVX512BW-FCP-NEXT: vpermt2q %zmm11, %zmm13, %zmm613312; AVX512BW-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13313; AVX512BW-FCP-NEXT: vmovdqa64 %zmm5, %zmm613314; AVX512BW-FCP-NEXT: vpermt2q %zmm8, %zmm13, %zmm613315; AVX512BW-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13316; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm613317; AVX512BW-FCP-NEXT: vpermt2q %zmm7, %zmm13, %zmm613318; AVX512BW-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13319; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm14 = [7,12,0,2,7,12,0,2]13320; AVX512BW-FCP-NEXT: # zmm14 = mem[0,1,2,3,0,1,2,3]13321; AVX512BW-FCP-NEXT: vmovdqa64 %zmm4, %zmm613322; AVX512BW-FCP-NEXT: vpermt2q %zmm9, %zmm14, %zmm613323; AVX512BW-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13324; AVX512BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = [0,5,0,11,0,5,0,11]13325; AVX512BW-FCP-NEXT: # zmm6 = mem[0,1,2,3,0,1,2,3]13326; AVX512BW-FCP-NEXT: vpermt2q %zmm4, %zmm6, %zmm913327; AVX512BW-FCP-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13328; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm413329; AVX512BW-FCP-NEXT: vpermt2q %zmm10, %zmm14, %zmm413330; AVX512BW-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13331; AVX512BW-FCP-NEXT: vpermt2q %zmm2, %zmm6, %zmm1013332; AVX512BW-FCP-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13333; AVX512BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm213334; AVX512BW-FCP-NEXT: vpermt2q %zmm11, %zmm14, %zmm213335; AVX512BW-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13336; AVX512BW-FCP-NEXT: vpermt2q %zmm1, %zmm6, %zmm1113337; AVX512BW-FCP-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13338; AVX512BW-FCP-NEXT: vmovdqa64 %zmm5, %zmm113339; AVX512BW-FCP-NEXT: vpermt2q %zmm8, %zmm14, %zmm113340; AVX512BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13341; AVX512BW-FCP-NEXT: vpermt2q %zmm5, %zmm6, %zmm813342; AVX512BW-FCP-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13343; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm113344; AVX512BW-FCP-NEXT: vpermt2q %zmm7, %zmm14, %zmm113345; AVX512BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13346; AVX512BW-FCP-NEXT: vpermt2q %zmm3, %zmm6, %zmm713347; AVX512BW-FCP-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13348; AVX512BW-FCP-NEXT: vmovdqa64 1472(%rdi), %zmm213349; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm1913350; AVX512BW-FCP-NEXT: vpermt2q %zmm0, %zmm16, %zmm1913351; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm113352; AVX512BW-FCP-NEXT: vpermt2q %zmm2, %zmm12, %zmm113353; AVX512BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13354; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm113355; AVX512BW-FCP-NEXT: vpermt2q %zmm2, %zmm13, %zmm113356; AVX512BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13357; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm113358; AVX512BW-FCP-NEXT: vpermt2q %zmm2, %zmm14, %zmm113359; AVX512BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13360; AVX512BW-FCP-NEXT: vpermt2q %zmm0, %zmm6, %zmm213361; AVX512BW-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13362; AVX512BW-FCP-NEXT: vmovdqa64 2048(%rdi), %zmm013363; AVX512BW-FCP-NEXT: vmovdqa64 2112(%rdi), %zmm213364; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm1813365; AVX512BW-FCP-NEXT: vpermt2q %zmm0, %zmm16, %zmm1813366; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm113367; AVX512BW-FCP-NEXT: vpermt2q %zmm2, %zmm12, %zmm113368; AVX512BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13369; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm113370; AVX512BW-FCP-NEXT: vpermt2q %zmm2, %zmm13, %zmm113371; AVX512BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13372; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm113373; AVX512BW-FCP-NEXT: vpermt2q %zmm2, %zmm14, %zmm113374; AVX512BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13375; AVX512BW-FCP-NEXT: vpermt2q %zmm0, %zmm6, %zmm213376; AVX512BW-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13377; AVX512BW-FCP-NEXT: vmovdqa64 2368(%rdi), %zmm013378; AVX512BW-FCP-NEXT: vmovdqa64 2432(%rdi), %zmm113379; AVX512BW-FCP-NEXT: vpermi2q %zmm1, %zmm0, %zmm1213380; AVX512BW-FCP-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13381; AVX512BW-FCP-NEXT: vpermi2q %zmm1, %zmm0, %zmm1313382; AVX512BW-FCP-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13383; AVX512BW-FCP-NEXT: vpermi2q %zmm1, %zmm0, %zmm1413384; AVX512BW-FCP-NEXT: vmovdqu64 %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13385; AVX512BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm1713386; AVX512BW-FCP-NEXT: vpermt2q %zmm0, %zmm6, %zmm113387; AVX512BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13388; AVX512BW-FCP-NEXT: vpermt2q %zmm0, %zmm16, %zmm1713389; AVX512BW-FCP-NEXT: vmovdqa64 384(%rdi), %zmm3113390; AVX512BW-FCP-NEXT: vmovdqa64 320(%rdi), %zmm213391; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} ymm3 = [0,5,10,15]13392; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm113393; AVX512BW-FCP-NEXT: vpermt2q %zmm31, %zmm3, %zmm113394; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} ymm29 = [1,6,11,0]13395; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm013396; AVX512BW-FCP-NEXT: vpermt2q %zmm31, %zmm29, %zmm013397; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13398; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} ymm6 = [2,7,12,0]13399; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm013400; AVX512BW-FCP-NEXT: vpermt2q %zmm31, %zmm6, %zmm013401; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13402; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} ymm7 = [11,0,5,0]13403; AVX512BW-FCP-NEXT: vmovdqa64 %zmm31, %zmm013404; AVX512BW-FCP-NEXT: vpermt2q %zmm2, %zmm7, %zmm013405; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13406; AVX512BW-FCP-NEXT: vpermt2q %zmm2, %zmm16, %zmm3113407; AVX512BW-FCP-NEXT: vmovdqa64 (%rdi), %zmm213408; AVX512BW-FCP-NEXT: vmovdqa64 64(%rdi), %zmm2313409; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm413410; AVX512BW-FCP-NEXT: vpermt2q %zmm23, %zmm3, %zmm413411; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm013412; AVX512BW-FCP-NEXT: vpermt2q %zmm23, %zmm29, %zmm013413; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13414; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm013415; AVX512BW-FCP-NEXT: vpermt2q %zmm23, %zmm6, %zmm013416; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13417; AVX512BW-FCP-NEXT: vmovdqa64 %zmm23, %zmm013418; AVX512BW-FCP-NEXT: vpermt2q %zmm2, %zmm7, %zmm013419; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13420; AVX512BW-FCP-NEXT: vpermt2q %zmm2, %zmm16, %zmm2313421; AVX512BW-FCP-NEXT: vmovdqa64 1024(%rdi), %zmm2113422; AVX512BW-FCP-NEXT: vmovdqa64 960(%rdi), %zmm213423; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm1013424; AVX512BW-FCP-NEXT: vpermt2q %zmm21, %zmm3, %zmm1013425; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm013426; AVX512BW-FCP-NEXT: vpermt2q %zmm21, %zmm29, %zmm013427; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13428; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm013429; AVX512BW-FCP-NEXT: vpermt2q %zmm21, %zmm6, %zmm013430; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13431; AVX512BW-FCP-NEXT: vmovdqa64 %zmm21, %zmm013432; AVX512BW-FCP-NEXT: vpermt2q %zmm2, %zmm7, %zmm013433; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13434; AVX512BW-FCP-NEXT: vpermt2q %zmm2, %zmm16, %zmm2113435; AVX512BW-FCP-NEXT: vmovdqa64 704(%rdi), %zmm2013436; AVX512BW-FCP-NEXT: vmovdqa64 640(%rdi), %zmm213437; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm1113438; AVX512BW-FCP-NEXT: vpermt2q %zmm20, %zmm3, %zmm1113439; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm013440; AVX512BW-FCP-NEXT: vpermt2q %zmm20, %zmm29, %zmm013441; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13442; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm013443; AVX512BW-FCP-NEXT: vpermt2q %zmm20, %zmm6, %zmm013444; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13445; AVX512BW-FCP-NEXT: vmovdqa64 %zmm20, %zmm013446; AVX512BW-FCP-NEXT: vpermt2q %zmm2, %zmm7, %zmm013447; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13448; AVX512BW-FCP-NEXT: vpermt2q %zmm2, %zmm16, %zmm2013449; AVX512BW-FCP-NEXT: vmovdqa64 1664(%rdi), %zmm2513450; AVX512BW-FCP-NEXT: vmovdqa64 1600(%rdi), %zmm213451; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm1313452; AVX512BW-FCP-NEXT: vpermt2q %zmm25, %zmm3, %zmm1313453; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm013454; AVX512BW-FCP-NEXT: vpermt2q %zmm25, %zmm29, %zmm013455; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13456; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm3013457; AVX512BW-FCP-NEXT: vpermt2q %zmm25, %zmm6, %zmm3013458; AVX512BW-FCP-NEXT: vmovdqa64 %zmm25, %zmm013459; AVX512BW-FCP-NEXT: vpermt2q %zmm2, %zmm7, %zmm013460; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13461; AVX512BW-FCP-NEXT: vpermt2q %zmm2, %zmm16, %zmm2513462; AVX512BW-FCP-NEXT: vmovdqa64 1344(%rdi), %zmm1213463; AVX512BW-FCP-NEXT: vmovdqa64 1280(%rdi), %zmm213464; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm1413465; AVX512BW-FCP-NEXT: vpermt2q %zmm12, %zmm3, %zmm1413466; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm2613467; AVX512BW-FCP-NEXT: vpermt2q %zmm12, %zmm29, %zmm2613468; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm2713469; AVX512BW-FCP-NEXT: vpermt2q %zmm12, %zmm6, %zmm2713470; AVX512BW-FCP-NEXT: vmovdqa64 %zmm12, %zmm2813471; AVX512BW-FCP-NEXT: vpermt2q %zmm2, %zmm7, %zmm2813472; AVX512BW-FCP-NEXT: vpermt2q %zmm2, %zmm16, %zmm1213473; AVX512BW-FCP-NEXT: vmovdqa64 1984(%rdi), %zmm913474; AVX512BW-FCP-NEXT: vmovdqa64 1920(%rdi), %zmm1513475; AVX512BW-FCP-NEXT: vmovdqa64 %zmm15, %zmm213476; AVX512BW-FCP-NEXT: vpermt2q %zmm9, %zmm3, %zmm213477; AVX512BW-FCP-NEXT: vmovdqa64 2304(%rdi), %zmm513478; AVX512BW-FCP-NEXT: vmovdqa64 2240(%rdi), %zmm013479; AVX512BW-FCP-NEXT: vpermi2q %zmm5, %zmm0, %zmm313480; AVX512BW-FCP-NEXT: vmovdqa64 %zmm15, %zmm813481; AVX512BW-FCP-NEXT: vpermt2q %zmm9, %zmm29, %zmm813482; AVX512BW-FCP-NEXT: vpermi2q %zmm5, %zmm0, %zmm2913483; AVX512BW-FCP-NEXT: vmovdqa64 %zmm15, %zmm2213484; AVX512BW-FCP-NEXT: vpermt2q %zmm9, %zmm6, %zmm2213485; AVX512BW-FCP-NEXT: vpermi2q %zmm5, %zmm0, %zmm613486; AVX512BW-FCP-NEXT: vmovdqa64 %zmm9, %zmm2413487; AVX512BW-FCP-NEXT: vpermt2q %zmm15, %zmm7, %zmm2413488; AVX512BW-FCP-NEXT: vpermi2q %zmm0, %zmm5, %zmm713489; AVX512BW-FCP-NEXT: vpermt2q %zmm0, %zmm16, %zmm513490; AVX512BW-FCP-NEXT: vpermt2q %zmm15, %zmm16, %zmm913491; AVX512BW-FCP-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm0 # 64-byte Folded Reload13492; AVX512BW-FCP-NEXT: # zmm0 = zmm4[0,1,2,3],mem[4,5,6,7]13493; AVX512BW-FCP-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm4 # 64-byte Folded Reload13494; AVX512BW-FCP-NEXT: # zmm4 = zmm1[0,1,2,3],mem[4,5,6,7]13495; AVX512BW-FCP-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm11, %zmm11 # 64-byte Folded Reload13496; AVX512BW-FCP-NEXT: # zmm11 = zmm11[0,1,2,3],mem[4,5,6,7]13497; AVX512BW-FCP-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm10, %zmm10 # 64-byte Folded Reload13498; AVX512BW-FCP-NEXT: # zmm10 = zmm10[0,1,2,3],mem[4,5,6,7]13499; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm19 = zmm14[0,1,2,3],zmm19[4,5,6,7]13500; AVX512BW-FCP-NEXT: vshufi64x2 $228, (%rsp), %zmm13, %zmm1 # 64-byte Folded Reload13501; AVX512BW-FCP-NEXT: # zmm1 = zmm13[0,1,2,3],mem[4,5,6,7]13502; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm2[0,1,2,3],zmm18[4,5,6,7]13503; AVX512BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm3 = zmm3[0,1,2,3],zmm17[4,5,6,7]13504; AVX512BW-FCP-NEXT: vmovdqa64 256(%rdi), %zmm1813505; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm15 = [0,1,2,3,4,5,6,11]13506; AVX512BW-FCP-NEXT: vpermt2q %zmm18, %zmm15, %zmm013507; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13508; AVX512BW-FCP-NEXT: vmovdqa64 576(%rdi), %zmm1413509; AVX512BW-FCP-NEXT: vpermt2q %zmm14, %zmm15, %zmm413510; AVX512BW-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13511; AVX512BW-FCP-NEXT: vmovdqa64 896(%rdi), %zmm1713512; AVX512BW-FCP-NEXT: vpermt2q %zmm17, %zmm15, %zmm1113513; AVX512BW-FCP-NEXT: vmovdqu64 %zmm11, (%rsp) # 64-byte Spill13514; AVX512BW-FCP-NEXT: vmovdqa64 1216(%rdi), %zmm1313515; AVX512BW-FCP-NEXT: vpermt2q %zmm13, %zmm15, %zmm1013516; AVX512BW-FCP-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13517; AVX512BW-FCP-NEXT: vmovdqa64 1536(%rdi), %zmm1613518; AVX512BW-FCP-NEXT: vpermt2q %zmm16, %zmm15, %zmm1913519; AVX512BW-FCP-NEXT: vmovdqu64 %zmm19, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13520; AVX512BW-FCP-NEXT: vmovdqa64 1856(%rdi), %zmm1113521; AVX512BW-FCP-NEXT: vpermt2q %zmm11, %zmm15, %zmm113522; AVX512BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13523; AVX512BW-FCP-NEXT: vmovdqa64 2176(%rdi), %zmm1913524; AVX512BW-FCP-NEXT: vpermt2q %zmm19, %zmm15, %zmm213525; AVX512BW-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13526; AVX512BW-FCP-NEXT: vmovdqa64 2496(%rdi), %zmm1013527; AVX512BW-FCP-NEXT: vpermt2q %zmm10, %zmm15, %zmm313528; AVX512BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13529; AVX512BW-FCP-NEXT: movb $7, %al13530; AVX512BW-FCP-NEXT: kmovd %eax, %k113531; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload13532; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13533; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm15 {%k1}13534; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload13535; AVX512BW-FCP-NEXT: vmovdqa64 %zmm8, %zmm4 {%k1}13536; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload13537; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13538; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm8 {%k1}13539; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13540; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload13541; AVX512BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm0 {%k1}13542; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload13543; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13544; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm1 {%k1}13545; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13546; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload13547; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm2 {%k1}13548; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload13549; AVX512BW-FCP-NEXT: vmovdqa64 %zmm26, %zmm3 {%k1}13550; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload13551; AVX512BW-FCP-NEXT: vmovdqa64 %zmm29, %zmm26 {%k1}13552; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm29 = [0,1,2,3,4,5,6,12]13553; AVX512BW-FCP-NEXT: vpermt2q %zmm18, %zmm29, %zmm1513554; AVX512BW-FCP-NEXT: vmovdqu64 %zmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13555; AVX512BW-FCP-NEXT: vpermt2q %zmm19, %zmm29, %zmm413556; AVX512BW-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13557; AVX512BW-FCP-NEXT: vpermt2q %zmm14, %zmm29, %zmm813558; AVX512BW-FCP-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13559; AVX512BW-FCP-NEXT: vpermt2q %zmm13, %zmm29, %zmm013560; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13561; AVX512BW-FCP-NEXT: vpermt2q %zmm17, %zmm29, %zmm113562; AVX512BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13563; AVX512BW-FCP-NEXT: vpermt2q %zmm11, %zmm29, %zmm213564; AVX512BW-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13565; AVX512BW-FCP-NEXT: vpermt2q %zmm16, %zmm29, %zmm313566; AVX512BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13567; AVX512BW-FCP-NEXT: vpermt2q %zmm10, %zmm29, %zmm2613568; AVX512BW-FCP-NEXT: vmovdqu64 %zmm26, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13569; AVX512BW-FCP-NEXT: movb $56, %al13570; AVX512BW-FCP-NEXT: kmovd %eax, %k113571; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13572; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload13573; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm0 {%k1}13574; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload13575; AVX512BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm22 {%k1}13576; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13577; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload13578; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm8 {%k1}13579; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13580; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload13581; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm26 {%k1}13582; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13583; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload13584; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm15 {%k1}13585; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13586; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm30 {%k1}13587; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload13588; AVX512BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm27 {%k1}13589; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload13590; AVX512BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm6 {%k1}13591; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm1 = [0,1,2,3,4,5,8,13]13592; AVX512BW-FCP-NEXT: vpermt2q %zmm18, %zmm1, %zmm013593; AVX512BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13594; AVX512BW-FCP-NEXT: vpermt2q %zmm19, %zmm1, %zmm2213595; AVX512BW-FCP-NEXT: vpermt2q %zmm14, %zmm1, %zmm813596; AVX512BW-FCP-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13597; AVX512BW-FCP-NEXT: vpermt2q %zmm13, %zmm1, %zmm2613598; AVX512BW-FCP-NEXT: vpermt2q %zmm17, %zmm1, %zmm1513599; AVX512BW-FCP-NEXT: vpermt2q %zmm11, %zmm1, %zmm3013600; AVX512BW-FCP-NEXT: vpermt2q %zmm16, %zmm1, %zmm2713601; AVX512BW-FCP-NEXT: vpermt2q %zmm10, %zmm1, %zmm613602; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13603; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload13604; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm8 {%k1}13605; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload13606; AVX512BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm24 {%k1}13607; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13608; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 # 64-byte Reload13609; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm29 {%k1}13610; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13611; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload13612; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm1 {%k1}13613; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13614; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload13615; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm3 {%k1}13616; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13617; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload13618; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm4 {%k1}13619; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13620; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm28 {%k1}13621; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13622; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm7 {%k1}13623; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm2 = [0,1,2,3,4,5,9,14]13624; AVX512BW-FCP-NEXT: vpermt2q %zmm18, %zmm2, %zmm813625; AVX512BW-FCP-NEXT: vpermt2q %zmm19, %zmm2, %zmm2413626; AVX512BW-FCP-NEXT: vpermt2q %zmm14, %zmm2, %zmm2913627; AVX512BW-FCP-NEXT: vpermt2q %zmm13, %zmm2, %zmm113628; AVX512BW-FCP-NEXT: vpermt2q %zmm17, %zmm2, %zmm313629; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm013630; AVX512BW-FCP-NEXT: vpermt2q %zmm11, %zmm2, %zmm413631; AVX512BW-FCP-NEXT: vpermt2q %zmm16, %zmm2, %zmm2813632; AVX512BW-FCP-NEXT: vpermt2q %zmm10, %zmm2, %zmm713633; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13634; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm31 {%k1}13635; AVX512BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm3 = [0,1,2,3,4,5,10,15]13636; AVX512BW-FCP-NEXT: vpermt2q %zmm14, %zmm3, %zmm3113637; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13638; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm23 {%k1}13639; AVX512BW-FCP-NEXT: vpermt2q %zmm18, %zmm3, %zmm2313640; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13641; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm21 {%k1}13642; AVX512BW-FCP-NEXT: vpermt2q %zmm13, %zmm3, %zmm2113643; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13644; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm20 {%k1}13645; AVX512BW-FCP-NEXT: vpermt2q %zmm17, %zmm3, %zmm2013646; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13647; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm25 {%k1}13648; AVX512BW-FCP-NEXT: vpermt2q %zmm11, %zmm3, %zmm2513649; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13650; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm12 {%k1}13651; AVX512BW-FCP-NEXT: vpermt2q %zmm16, %zmm3, %zmm1213652; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13653; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm5 {%k1}13654; AVX512BW-FCP-NEXT: vpermt2q %zmm10, %zmm3, %zmm513655; AVX512BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13656; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm9 {%k1}13657; AVX512BW-FCP-NEXT: vpermt2q %zmm19, %zmm3, %zmm913658; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13659; AVX512BW-FCP-NEXT: vmovaps %zmm2, 448(%rsi)13660; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13661; AVX512BW-FCP-NEXT: vmovaps %zmm2, 384(%rsi)13662; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13663; AVX512BW-FCP-NEXT: vmovaps %zmm2, 320(%rsi)13664; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13665; AVX512BW-FCP-NEXT: vmovaps %zmm2, 256(%rsi)13666; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13667; AVX512BW-FCP-NEXT: vmovaps %zmm2, 192(%rsi)13668; AVX512BW-FCP-NEXT: vmovups (%rsp), %zmm2 # 64-byte Reload13669; AVX512BW-FCP-NEXT: vmovaps %zmm2, 128(%rsi)13670; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13671; AVX512BW-FCP-NEXT: vmovaps %zmm2, 64(%rsi)13672; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13673; AVX512BW-FCP-NEXT: vmovaps %zmm2, (%rsi)13674; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13675; AVX512BW-FCP-NEXT: vmovaps %zmm2, 448(%rdx)13676; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13677; AVX512BW-FCP-NEXT: vmovaps %zmm2, 256(%rdx)13678; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13679; AVX512BW-FCP-NEXT: vmovaps %zmm2, 320(%rdx)13680; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13681; AVX512BW-FCP-NEXT: vmovaps %zmm2, 128(%rdx)13682; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13683; AVX512BW-FCP-NEXT: vmovaps %zmm2, 192(%rdx)13684; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13685; AVX512BW-FCP-NEXT: vmovaps %zmm2, (%rdx)13686; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13687; AVX512BW-FCP-NEXT: vmovaps %zmm2, 64(%rdx)13688; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13689; AVX512BW-FCP-NEXT: vmovaps %zmm2, 384(%rdx)13690; AVX512BW-FCP-NEXT: vmovdqa64 %zmm6, 448(%rcx)13691; AVX512BW-FCP-NEXT: vmovdqa64 %zmm27, 256(%rcx)13692; AVX512BW-FCP-NEXT: vmovdqa64 %zmm30, 320(%rcx)13693; AVX512BW-FCP-NEXT: vmovdqa64 %zmm15, 128(%rcx)13694; AVX512BW-FCP-NEXT: vmovdqa64 %zmm26, 192(%rcx)13695; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13696; AVX512BW-FCP-NEXT: vmovaps %zmm2, (%rcx)13697; AVX512BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload13698; AVX512BW-FCP-NEXT: vmovaps %zmm2, 64(%rcx)13699; AVX512BW-FCP-NEXT: vmovdqa64 %zmm22, 384(%rcx)13700; AVX512BW-FCP-NEXT: vmovdqa64 %zmm7, 448(%r8)13701; AVX512BW-FCP-NEXT: vmovdqa64 %zmm28, 256(%r8)13702; AVX512BW-FCP-NEXT: vmovdqa64 %zmm4, 320(%r8)13703; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, 128(%r8)13704; AVX512BW-FCP-NEXT: vmovdqa64 %zmm1, 192(%r8)13705; AVX512BW-FCP-NEXT: vmovdqa64 %zmm8, (%r8)13706; AVX512BW-FCP-NEXT: vmovdqa64 %zmm29, 64(%r8)13707; AVX512BW-FCP-NEXT: vmovdqa64 %zmm24, 384(%r8)13708; AVX512BW-FCP-NEXT: vmovdqa64 %zmm9, 384(%r9)13709; AVX512BW-FCP-NEXT: vmovdqa64 %zmm5, 448(%r9)13710; AVX512BW-FCP-NEXT: vmovdqa64 %zmm12, 256(%r9)13711; AVX512BW-FCP-NEXT: vmovdqa64 %zmm25, 320(%r9)13712; AVX512BW-FCP-NEXT: vmovdqa64 %zmm20, 128(%r9)13713; AVX512BW-FCP-NEXT: vmovdqa64 %zmm21, 192(%r9)13714; AVX512BW-FCP-NEXT: vmovdqa64 %zmm23, (%r9)13715; AVX512BW-FCP-NEXT: vmovdqa64 %zmm31, 64(%r9)13716; AVX512BW-FCP-NEXT: addq $3336, %rsp # imm = 0xD0813717; AVX512BW-FCP-NEXT: vzeroupper13718; AVX512BW-FCP-NEXT: retq13719;13720; AVX512DQ-BW-LABEL: load_i64_stride5_vf64:13721; AVX512DQ-BW: # %bb.0:13722; AVX512DQ-BW-NEXT: subq $3336, %rsp # imm = 0xD0813723; AVX512DQ-BW-NEXT: vmovdqa64 1728(%rdi), %zmm313724; AVX512DQ-BW-NEXT: vmovdqa64 1792(%rdi), %zmm713725; AVX512DQ-BW-NEXT: vmovdqa64 1408(%rdi), %zmm013726; AVX512DQ-BW-NEXT: vmovdqa64 1088(%rdi), %zmm113727; AVX512DQ-BW-NEXT: vmovdqa64 1152(%rdi), %zmm613728; AVX512DQ-BW-NEXT: vmovdqa64 768(%rdi), %zmm513729; AVX512DQ-BW-NEXT: vmovdqa64 832(%rdi), %zmm813730; AVX512DQ-BW-NEXT: vmovdqa64 448(%rdi), %zmm413731; AVX512DQ-BW-NEXT: vmovdqa64 512(%rdi), %zmm913732; AVX512DQ-BW-NEXT: vmovdqa64 128(%rdi), %zmm213733; AVX512DQ-BW-NEXT: vmovdqa64 192(%rdi), %zmm1013734; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm16 = [12,1,6,0,12,1,6,0]13735; AVX512DQ-BW-NEXT: # zmm16 = mem[0,1,2,3,0,1,2,3]13736; AVX512DQ-BW-NEXT: vmovdqa64 %zmm10, %zmm1113737; AVX512DQ-BW-NEXT: vpermt2q %zmm2, %zmm16, %zmm1113738; AVX512DQ-BW-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13739; AVX512DQ-BW-NEXT: vmovdqa64 %zmm9, %zmm1113740; AVX512DQ-BW-NEXT: vpermt2q %zmm4, %zmm16, %zmm1113741; AVX512DQ-BW-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13742; AVX512DQ-BW-NEXT: vmovdqa64 %zmm8, %zmm1113743; AVX512DQ-BW-NEXT: vpermt2q %zmm5, %zmm16, %zmm1113744; AVX512DQ-BW-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13745; AVX512DQ-BW-NEXT: vmovdqa64 %zmm6, %zmm1213746; AVX512DQ-BW-NEXT: vmovdqa64 %zmm6, %zmm1113747; AVX512DQ-BW-NEXT: vpermt2q %zmm1, %zmm16, %zmm1213748; AVX512DQ-BW-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13749; AVX512DQ-BW-NEXT: vmovdqa64 %zmm7, %zmm613750; AVX512DQ-BW-NEXT: vpermt2q %zmm3, %zmm16, %zmm613751; AVX512DQ-BW-NEXT: vmovdqu64 %zmm6, (%rsp) # 64-byte Spill13752; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm12 = [5,10,15,0,5,10,15,0]13753; AVX512DQ-BW-NEXT: # zmm12 = mem[0,1,2,3,0,1,2,3]13754; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm613755; AVX512DQ-BW-NEXT: vpermt2q %zmm10, %zmm12, %zmm613756; AVX512DQ-BW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13757; AVX512DQ-BW-NEXT: vmovdqa64 %zmm4, %zmm613758; AVX512DQ-BW-NEXT: vpermt2q %zmm9, %zmm12, %zmm613759; AVX512DQ-BW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13760; AVX512DQ-BW-NEXT: vmovdqa64 %zmm1, %zmm613761; AVX512DQ-BW-NEXT: vpermt2q %zmm11, %zmm12, %zmm613762; AVX512DQ-BW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13763; AVX512DQ-BW-NEXT: vmovdqa64 %zmm5, %zmm613764; AVX512DQ-BW-NEXT: vpermt2q %zmm8, %zmm12, %zmm613765; AVX512DQ-BW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13766; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, %zmm613767; AVX512DQ-BW-NEXT: vpermt2q %zmm7, %zmm12, %zmm613768; AVX512DQ-BW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13769; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = [6,11,0,1,6,11,0,1]13770; AVX512DQ-BW-NEXT: # zmm6 = mem[0,1,2,3,0,1,2,3]13771; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm1313772; AVX512DQ-BW-NEXT: vpermt2q %zmm10, %zmm6, %zmm1313773; AVX512DQ-BW-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13774; AVX512DQ-BW-NEXT: vmovdqa64 %zmm6, %zmm1313775; AVX512DQ-BW-NEXT: vmovdqa64 %zmm4, %zmm613776; AVX512DQ-BW-NEXT: vpermt2q %zmm9, %zmm13, %zmm613777; AVX512DQ-BW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13778; AVX512DQ-BW-NEXT: vmovdqa64 %zmm1, %zmm613779; AVX512DQ-BW-NEXT: vpermt2q %zmm11, %zmm13, %zmm613780; AVX512DQ-BW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13781; AVX512DQ-BW-NEXT: vmovdqa64 %zmm5, %zmm613782; AVX512DQ-BW-NEXT: vpermt2q %zmm8, %zmm13, %zmm613783; AVX512DQ-BW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13784; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, %zmm613785; AVX512DQ-BW-NEXT: vpermt2q %zmm7, %zmm13, %zmm613786; AVX512DQ-BW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13787; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm14 = [7,12,0,2,7,12,0,2]13788; AVX512DQ-BW-NEXT: # zmm14 = mem[0,1,2,3,0,1,2,3]13789; AVX512DQ-BW-NEXT: vmovdqa64 %zmm4, %zmm613790; AVX512DQ-BW-NEXT: vpermt2q %zmm9, %zmm14, %zmm613791; AVX512DQ-BW-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13792; AVX512DQ-BW-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = [0,5,0,11,0,5,0,11]13793; AVX512DQ-BW-NEXT: # zmm6 = mem[0,1,2,3,0,1,2,3]13794; AVX512DQ-BW-NEXT: vpermt2q %zmm4, %zmm6, %zmm913795; AVX512DQ-BW-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13796; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm413797; AVX512DQ-BW-NEXT: vpermt2q %zmm10, %zmm14, %zmm413798; AVX512DQ-BW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13799; AVX512DQ-BW-NEXT: vpermt2q %zmm2, %zmm6, %zmm1013800; AVX512DQ-BW-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13801; AVX512DQ-BW-NEXT: vmovdqa64 %zmm1, %zmm213802; AVX512DQ-BW-NEXT: vpermt2q %zmm11, %zmm14, %zmm213803; AVX512DQ-BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13804; AVX512DQ-BW-NEXT: vpermt2q %zmm1, %zmm6, %zmm1113805; AVX512DQ-BW-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13806; AVX512DQ-BW-NEXT: vmovdqa64 %zmm5, %zmm113807; AVX512DQ-BW-NEXT: vpermt2q %zmm8, %zmm14, %zmm113808; AVX512DQ-BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13809; AVX512DQ-BW-NEXT: vpermt2q %zmm5, %zmm6, %zmm813810; AVX512DQ-BW-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13811; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, %zmm113812; AVX512DQ-BW-NEXT: vpermt2q %zmm7, %zmm14, %zmm113813; AVX512DQ-BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13814; AVX512DQ-BW-NEXT: vpermt2q %zmm3, %zmm6, %zmm713815; AVX512DQ-BW-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13816; AVX512DQ-BW-NEXT: vmovdqa64 1472(%rdi), %zmm213817; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm1913818; AVX512DQ-BW-NEXT: vpermt2q %zmm0, %zmm16, %zmm1913819; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, %zmm113820; AVX512DQ-BW-NEXT: vpermt2q %zmm2, %zmm12, %zmm113821; AVX512DQ-BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13822; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, %zmm113823; AVX512DQ-BW-NEXT: vpermt2q %zmm2, %zmm13, %zmm113824; AVX512DQ-BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13825; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, %zmm113826; AVX512DQ-BW-NEXT: vpermt2q %zmm2, %zmm14, %zmm113827; AVX512DQ-BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13828; AVX512DQ-BW-NEXT: vpermt2q %zmm0, %zmm6, %zmm213829; AVX512DQ-BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13830; AVX512DQ-BW-NEXT: vmovdqa64 2048(%rdi), %zmm013831; AVX512DQ-BW-NEXT: vmovdqa64 2112(%rdi), %zmm213832; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm1813833; AVX512DQ-BW-NEXT: vpermt2q %zmm0, %zmm16, %zmm1813834; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, %zmm113835; AVX512DQ-BW-NEXT: vpermt2q %zmm2, %zmm12, %zmm113836; AVX512DQ-BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13837; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, %zmm113838; AVX512DQ-BW-NEXT: vpermt2q %zmm2, %zmm13, %zmm113839; AVX512DQ-BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13840; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, %zmm113841; AVX512DQ-BW-NEXT: vpermt2q %zmm2, %zmm14, %zmm113842; AVX512DQ-BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13843; AVX512DQ-BW-NEXT: vpermt2q %zmm0, %zmm6, %zmm213844; AVX512DQ-BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13845; AVX512DQ-BW-NEXT: vmovdqa64 2368(%rdi), %zmm013846; AVX512DQ-BW-NEXT: vmovdqa64 2432(%rdi), %zmm113847; AVX512DQ-BW-NEXT: vpermi2q %zmm1, %zmm0, %zmm1213848; AVX512DQ-BW-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13849; AVX512DQ-BW-NEXT: vpermi2q %zmm1, %zmm0, %zmm1313850; AVX512DQ-BW-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13851; AVX512DQ-BW-NEXT: vpermi2q %zmm1, %zmm0, %zmm1413852; AVX512DQ-BW-NEXT: vmovdqu64 %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13853; AVX512DQ-BW-NEXT: vmovdqa64 %zmm1, %zmm1713854; AVX512DQ-BW-NEXT: vpermt2q %zmm0, %zmm6, %zmm113855; AVX512DQ-BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13856; AVX512DQ-BW-NEXT: vpermt2q %zmm0, %zmm16, %zmm1713857; AVX512DQ-BW-NEXT: vmovdqa64 384(%rdi), %zmm3113858; AVX512DQ-BW-NEXT: vmovdqa64 320(%rdi), %zmm213859; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} ymm3 = [0,5,10,15]13860; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm113861; AVX512DQ-BW-NEXT: vpermt2q %zmm31, %zmm3, %zmm113862; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} ymm29 = [1,6,11,0]13863; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm013864; AVX512DQ-BW-NEXT: vpermt2q %zmm31, %zmm29, %zmm013865; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13866; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} ymm6 = [2,7,12,0]13867; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm013868; AVX512DQ-BW-NEXT: vpermt2q %zmm31, %zmm6, %zmm013869; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13870; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} ymm7 = [11,0,5,0]13871; AVX512DQ-BW-NEXT: vmovdqa64 %zmm31, %zmm013872; AVX512DQ-BW-NEXT: vpermt2q %zmm2, %zmm7, %zmm013873; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13874; AVX512DQ-BW-NEXT: vpermt2q %zmm2, %zmm16, %zmm3113875; AVX512DQ-BW-NEXT: vmovdqa64 (%rdi), %zmm213876; AVX512DQ-BW-NEXT: vmovdqa64 64(%rdi), %zmm2313877; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm413878; AVX512DQ-BW-NEXT: vpermt2q %zmm23, %zmm3, %zmm413879; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm013880; AVX512DQ-BW-NEXT: vpermt2q %zmm23, %zmm29, %zmm013881; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13882; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm013883; AVX512DQ-BW-NEXT: vpermt2q %zmm23, %zmm6, %zmm013884; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13885; AVX512DQ-BW-NEXT: vmovdqa64 %zmm23, %zmm013886; AVX512DQ-BW-NEXT: vpermt2q %zmm2, %zmm7, %zmm013887; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13888; AVX512DQ-BW-NEXT: vpermt2q %zmm2, %zmm16, %zmm2313889; AVX512DQ-BW-NEXT: vmovdqa64 1024(%rdi), %zmm2113890; AVX512DQ-BW-NEXT: vmovdqa64 960(%rdi), %zmm213891; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm1013892; AVX512DQ-BW-NEXT: vpermt2q %zmm21, %zmm3, %zmm1013893; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm013894; AVX512DQ-BW-NEXT: vpermt2q %zmm21, %zmm29, %zmm013895; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13896; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm013897; AVX512DQ-BW-NEXT: vpermt2q %zmm21, %zmm6, %zmm013898; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13899; AVX512DQ-BW-NEXT: vmovdqa64 %zmm21, %zmm013900; AVX512DQ-BW-NEXT: vpermt2q %zmm2, %zmm7, %zmm013901; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13902; AVX512DQ-BW-NEXT: vpermt2q %zmm2, %zmm16, %zmm2113903; AVX512DQ-BW-NEXT: vmovdqa64 704(%rdi), %zmm2013904; AVX512DQ-BW-NEXT: vmovdqa64 640(%rdi), %zmm213905; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm1113906; AVX512DQ-BW-NEXT: vpermt2q %zmm20, %zmm3, %zmm1113907; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm013908; AVX512DQ-BW-NEXT: vpermt2q %zmm20, %zmm29, %zmm013909; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13910; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm013911; AVX512DQ-BW-NEXT: vpermt2q %zmm20, %zmm6, %zmm013912; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13913; AVX512DQ-BW-NEXT: vmovdqa64 %zmm20, %zmm013914; AVX512DQ-BW-NEXT: vpermt2q %zmm2, %zmm7, %zmm013915; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13916; AVX512DQ-BW-NEXT: vpermt2q %zmm2, %zmm16, %zmm2013917; AVX512DQ-BW-NEXT: vmovdqa64 1664(%rdi), %zmm2513918; AVX512DQ-BW-NEXT: vmovdqa64 1600(%rdi), %zmm213919; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm1313920; AVX512DQ-BW-NEXT: vpermt2q %zmm25, %zmm3, %zmm1313921; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm013922; AVX512DQ-BW-NEXT: vpermt2q %zmm25, %zmm29, %zmm013923; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13924; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm3013925; AVX512DQ-BW-NEXT: vpermt2q %zmm25, %zmm6, %zmm3013926; AVX512DQ-BW-NEXT: vmovdqa64 %zmm25, %zmm013927; AVX512DQ-BW-NEXT: vpermt2q %zmm2, %zmm7, %zmm013928; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13929; AVX512DQ-BW-NEXT: vpermt2q %zmm2, %zmm16, %zmm2513930; AVX512DQ-BW-NEXT: vmovdqa64 1344(%rdi), %zmm1213931; AVX512DQ-BW-NEXT: vmovdqa64 1280(%rdi), %zmm213932; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm1413933; AVX512DQ-BW-NEXT: vpermt2q %zmm12, %zmm3, %zmm1413934; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm2613935; AVX512DQ-BW-NEXT: vpermt2q %zmm12, %zmm29, %zmm2613936; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm2713937; AVX512DQ-BW-NEXT: vpermt2q %zmm12, %zmm6, %zmm2713938; AVX512DQ-BW-NEXT: vmovdqa64 %zmm12, %zmm2813939; AVX512DQ-BW-NEXT: vpermt2q %zmm2, %zmm7, %zmm2813940; AVX512DQ-BW-NEXT: vpermt2q %zmm2, %zmm16, %zmm1213941; AVX512DQ-BW-NEXT: vmovdqa64 1984(%rdi), %zmm913942; AVX512DQ-BW-NEXT: vmovdqa64 1920(%rdi), %zmm1513943; AVX512DQ-BW-NEXT: vmovdqa64 %zmm15, %zmm213944; AVX512DQ-BW-NEXT: vpermt2q %zmm9, %zmm3, %zmm213945; AVX512DQ-BW-NEXT: vmovdqa64 2304(%rdi), %zmm513946; AVX512DQ-BW-NEXT: vmovdqa64 2240(%rdi), %zmm013947; AVX512DQ-BW-NEXT: vpermi2q %zmm5, %zmm0, %zmm313948; AVX512DQ-BW-NEXT: vmovdqa64 %zmm15, %zmm813949; AVX512DQ-BW-NEXT: vpermt2q %zmm9, %zmm29, %zmm813950; AVX512DQ-BW-NEXT: vpermi2q %zmm5, %zmm0, %zmm2913951; AVX512DQ-BW-NEXT: vmovdqa64 %zmm15, %zmm2213952; AVX512DQ-BW-NEXT: vpermt2q %zmm9, %zmm6, %zmm2213953; AVX512DQ-BW-NEXT: vpermi2q %zmm5, %zmm0, %zmm613954; AVX512DQ-BW-NEXT: vmovdqa64 %zmm9, %zmm2413955; AVX512DQ-BW-NEXT: vpermt2q %zmm15, %zmm7, %zmm2413956; AVX512DQ-BW-NEXT: vpermi2q %zmm0, %zmm5, %zmm713957; AVX512DQ-BW-NEXT: vpermt2q %zmm0, %zmm16, %zmm513958; AVX512DQ-BW-NEXT: vpermt2q %zmm15, %zmm16, %zmm913959; AVX512DQ-BW-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm0 # 64-byte Folded Reload13960; AVX512DQ-BW-NEXT: # zmm0 = zmm4[0,1,2,3],mem[4,5,6,7]13961; AVX512DQ-BW-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm4 # 64-byte Folded Reload13962; AVX512DQ-BW-NEXT: # zmm4 = zmm1[0,1,2,3],mem[4,5,6,7]13963; AVX512DQ-BW-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm11, %zmm11 # 64-byte Folded Reload13964; AVX512DQ-BW-NEXT: # zmm11 = zmm11[0,1,2,3],mem[4,5,6,7]13965; AVX512DQ-BW-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm10, %zmm10 # 64-byte Folded Reload13966; AVX512DQ-BW-NEXT: # zmm10 = zmm10[0,1,2,3],mem[4,5,6,7]13967; AVX512DQ-BW-NEXT: vshufi64x2 {{.*#+}} zmm19 = zmm14[0,1,2,3],zmm19[4,5,6,7]13968; AVX512DQ-BW-NEXT: vshufi64x2 $228, (%rsp), %zmm13, %zmm1 # 64-byte Folded Reload13969; AVX512DQ-BW-NEXT: # zmm1 = zmm13[0,1,2,3],mem[4,5,6,7]13970; AVX512DQ-BW-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm2[0,1,2,3],zmm18[4,5,6,7]13971; AVX512DQ-BW-NEXT: vshufi64x2 {{.*#+}} zmm3 = zmm3[0,1,2,3],zmm17[4,5,6,7]13972; AVX512DQ-BW-NEXT: vmovdqa64 256(%rdi), %zmm1813973; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} zmm15 = [0,1,2,3,4,5,6,11]13974; AVX512DQ-BW-NEXT: vpermt2q %zmm18, %zmm15, %zmm013975; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13976; AVX512DQ-BW-NEXT: vmovdqa64 576(%rdi), %zmm1413977; AVX512DQ-BW-NEXT: vpermt2q %zmm14, %zmm15, %zmm413978; AVX512DQ-BW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13979; AVX512DQ-BW-NEXT: vmovdqa64 896(%rdi), %zmm1713980; AVX512DQ-BW-NEXT: vpermt2q %zmm17, %zmm15, %zmm1113981; AVX512DQ-BW-NEXT: vmovdqu64 %zmm11, (%rsp) # 64-byte Spill13982; AVX512DQ-BW-NEXT: vmovdqa64 1216(%rdi), %zmm1313983; AVX512DQ-BW-NEXT: vpermt2q %zmm13, %zmm15, %zmm1013984; AVX512DQ-BW-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13985; AVX512DQ-BW-NEXT: vmovdqa64 1536(%rdi), %zmm1613986; AVX512DQ-BW-NEXT: vpermt2q %zmm16, %zmm15, %zmm1913987; AVX512DQ-BW-NEXT: vmovdqu64 %zmm19, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13988; AVX512DQ-BW-NEXT: vmovdqa64 1856(%rdi), %zmm1113989; AVX512DQ-BW-NEXT: vpermt2q %zmm11, %zmm15, %zmm113990; AVX512DQ-BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13991; AVX512DQ-BW-NEXT: vmovdqa64 2176(%rdi), %zmm1913992; AVX512DQ-BW-NEXT: vpermt2q %zmm19, %zmm15, %zmm213993; AVX512DQ-BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13994; AVX512DQ-BW-NEXT: vmovdqa64 2496(%rdi), %zmm1013995; AVX512DQ-BW-NEXT: vpermt2q %zmm10, %zmm15, %zmm313996; AVX512DQ-BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill13997; AVX512DQ-BW-NEXT: movb $7, %al13998; AVX512DQ-BW-NEXT: kmovd %eax, %k113999; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload14000; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14001; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, %zmm15 {%k1}14002; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload14003; AVX512DQ-BW-NEXT: vmovdqa64 %zmm8, %zmm4 {%k1}14004; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload14005; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14006; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, %zmm8 {%k1}14007; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14008; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload14009; AVX512DQ-BW-NEXT: vmovdqa64 %zmm1, %zmm0 {%k1}14010; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload14011; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14012; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm1 {%k1}14013; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14014; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload14015; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, %zmm2 {%k1}14016; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload14017; AVX512DQ-BW-NEXT: vmovdqa64 %zmm26, %zmm3 {%k1}14018; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload14019; AVX512DQ-BW-NEXT: vmovdqa64 %zmm29, %zmm26 {%k1}14020; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} zmm29 = [0,1,2,3,4,5,6,12]14021; AVX512DQ-BW-NEXT: vpermt2q %zmm18, %zmm29, %zmm1514022; AVX512DQ-BW-NEXT: vmovdqu64 %zmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14023; AVX512DQ-BW-NEXT: vpermt2q %zmm19, %zmm29, %zmm414024; AVX512DQ-BW-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14025; AVX512DQ-BW-NEXT: vpermt2q %zmm14, %zmm29, %zmm814026; AVX512DQ-BW-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14027; AVX512DQ-BW-NEXT: vpermt2q %zmm13, %zmm29, %zmm014028; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14029; AVX512DQ-BW-NEXT: vpermt2q %zmm17, %zmm29, %zmm114030; AVX512DQ-BW-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14031; AVX512DQ-BW-NEXT: vpermt2q %zmm11, %zmm29, %zmm214032; AVX512DQ-BW-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14033; AVX512DQ-BW-NEXT: vpermt2q %zmm16, %zmm29, %zmm314034; AVX512DQ-BW-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14035; AVX512DQ-BW-NEXT: vpermt2q %zmm10, %zmm29, %zmm2614036; AVX512DQ-BW-NEXT: vmovdqu64 %zmm26, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14037; AVX512DQ-BW-NEXT: movb $56, %al14038; AVX512DQ-BW-NEXT: kmovd %eax, %k114039; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14040; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14041; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm0 {%k1}14042; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload14043; AVX512DQ-BW-NEXT: vmovdqa64 %zmm1, %zmm22 {%k1}14044; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14045; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload14046; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm8 {%k1}14047; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14048; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload14049; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm26 {%k1}14050; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14051; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload14052; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm15 {%k1}14053; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14054; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm30 {%k1}14055; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload14056; AVX512DQ-BW-NEXT: vmovdqa64 %zmm1, %zmm27 {%k1}14057; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload14058; AVX512DQ-BW-NEXT: vmovdqa64 %zmm1, %zmm6 {%k1}14059; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} zmm1 = [0,1,2,3,4,5,8,13]14060; AVX512DQ-BW-NEXT: vpermt2q %zmm18, %zmm1, %zmm014061; AVX512DQ-BW-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14062; AVX512DQ-BW-NEXT: vpermt2q %zmm19, %zmm1, %zmm2214063; AVX512DQ-BW-NEXT: vpermt2q %zmm14, %zmm1, %zmm814064; AVX512DQ-BW-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14065; AVX512DQ-BW-NEXT: vpermt2q %zmm13, %zmm1, %zmm2614066; AVX512DQ-BW-NEXT: vpermt2q %zmm17, %zmm1, %zmm1514067; AVX512DQ-BW-NEXT: vpermt2q %zmm11, %zmm1, %zmm3014068; AVX512DQ-BW-NEXT: vpermt2q %zmm16, %zmm1, %zmm2714069; AVX512DQ-BW-NEXT: vpermt2q %zmm10, %zmm1, %zmm614070; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14071; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload14072; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm8 {%k1}14073; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload14074; AVX512DQ-BW-NEXT: vmovdqa64 %zmm1, %zmm24 {%k1}14075; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14076; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 # 64-byte Reload14077; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm29 {%k1}14078; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14079; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload14080; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm1 {%k1}14081; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14082; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload14083; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm3 {%k1}14084; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14085; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload14086; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm4 {%k1}14087; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14088; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm28 {%k1}14089; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14090; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm7 {%k1}14091; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} zmm2 = [0,1,2,3,4,5,9,14]14092; AVX512DQ-BW-NEXT: vpermt2q %zmm18, %zmm2, %zmm814093; AVX512DQ-BW-NEXT: vpermt2q %zmm19, %zmm2, %zmm2414094; AVX512DQ-BW-NEXT: vpermt2q %zmm14, %zmm2, %zmm2914095; AVX512DQ-BW-NEXT: vpermt2q %zmm13, %zmm2, %zmm114096; AVX512DQ-BW-NEXT: vpermt2q %zmm17, %zmm2, %zmm314097; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, %zmm014098; AVX512DQ-BW-NEXT: vpermt2q %zmm11, %zmm2, %zmm414099; AVX512DQ-BW-NEXT: vpermt2q %zmm16, %zmm2, %zmm2814100; AVX512DQ-BW-NEXT: vpermt2q %zmm10, %zmm2, %zmm714101; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14102; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm31 {%k1}14103; AVX512DQ-BW-NEXT: vpmovsxbq {{.*#+}} zmm3 = [0,1,2,3,4,5,10,15]14104; AVX512DQ-BW-NEXT: vpermt2q %zmm14, %zmm3, %zmm3114105; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14106; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm23 {%k1}14107; AVX512DQ-BW-NEXT: vpermt2q %zmm18, %zmm3, %zmm2314108; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14109; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm21 {%k1}14110; AVX512DQ-BW-NEXT: vpermt2q %zmm13, %zmm3, %zmm2114111; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14112; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm20 {%k1}14113; AVX512DQ-BW-NEXT: vpermt2q %zmm17, %zmm3, %zmm2014114; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14115; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm25 {%k1}14116; AVX512DQ-BW-NEXT: vpermt2q %zmm11, %zmm3, %zmm2514117; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14118; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm12 {%k1}14119; AVX512DQ-BW-NEXT: vpermt2q %zmm16, %zmm3, %zmm1214120; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14121; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm5 {%k1}14122; AVX512DQ-BW-NEXT: vpermt2q %zmm10, %zmm3, %zmm514123; AVX512DQ-BW-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14124; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, %zmm9 {%k1}14125; AVX512DQ-BW-NEXT: vpermt2q %zmm19, %zmm3, %zmm914126; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14127; AVX512DQ-BW-NEXT: vmovaps %zmm2, 448(%rsi)14128; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14129; AVX512DQ-BW-NEXT: vmovaps %zmm2, 384(%rsi)14130; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14131; AVX512DQ-BW-NEXT: vmovaps %zmm2, 320(%rsi)14132; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14133; AVX512DQ-BW-NEXT: vmovaps %zmm2, 256(%rsi)14134; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14135; AVX512DQ-BW-NEXT: vmovaps %zmm2, 192(%rsi)14136; AVX512DQ-BW-NEXT: vmovups (%rsp), %zmm2 # 64-byte Reload14137; AVX512DQ-BW-NEXT: vmovaps %zmm2, 128(%rsi)14138; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14139; AVX512DQ-BW-NEXT: vmovaps %zmm2, 64(%rsi)14140; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14141; AVX512DQ-BW-NEXT: vmovaps %zmm2, (%rsi)14142; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14143; AVX512DQ-BW-NEXT: vmovaps %zmm2, 448(%rdx)14144; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14145; AVX512DQ-BW-NEXT: vmovaps %zmm2, 256(%rdx)14146; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14147; AVX512DQ-BW-NEXT: vmovaps %zmm2, 320(%rdx)14148; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14149; AVX512DQ-BW-NEXT: vmovaps %zmm2, 128(%rdx)14150; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14151; AVX512DQ-BW-NEXT: vmovaps %zmm2, 192(%rdx)14152; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14153; AVX512DQ-BW-NEXT: vmovaps %zmm2, (%rdx)14154; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14155; AVX512DQ-BW-NEXT: vmovaps %zmm2, 64(%rdx)14156; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14157; AVX512DQ-BW-NEXT: vmovaps %zmm2, 384(%rdx)14158; AVX512DQ-BW-NEXT: vmovdqa64 %zmm6, 448(%rcx)14159; AVX512DQ-BW-NEXT: vmovdqa64 %zmm27, 256(%rcx)14160; AVX512DQ-BW-NEXT: vmovdqa64 %zmm30, 320(%rcx)14161; AVX512DQ-BW-NEXT: vmovdqa64 %zmm15, 128(%rcx)14162; AVX512DQ-BW-NEXT: vmovdqa64 %zmm26, 192(%rcx)14163; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14164; AVX512DQ-BW-NEXT: vmovaps %zmm2, (%rcx)14165; AVX512DQ-BW-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14166; AVX512DQ-BW-NEXT: vmovaps %zmm2, 64(%rcx)14167; AVX512DQ-BW-NEXT: vmovdqa64 %zmm22, 384(%rcx)14168; AVX512DQ-BW-NEXT: vmovdqa64 %zmm7, 448(%r8)14169; AVX512DQ-BW-NEXT: vmovdqa64 %zmm28, 256(%r8)14170; AVX512DQ-BW-NEXT: vmovdqa64 %zmm4, 320(%r8)14171; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, 128(%r8)14172; AVX512DQ-BW-NEXT: vmovdqa64 %zmm1, 192(%r8)14173; AVX512DQ-BW-NEXT: vmovdqa64 %zmm8, (%r8)14174; AVX512DQ-BW-NEXT: vmovdqa64 %zmm29, 64(%r8)14175; AVX512DQ-BW-NEXT: vmovdqa64 %zmm24, 384(%r8)14176; AVX512DQ-BW-NEXT: vmovdqa64 %zmm9, 384(%r9)14177; AVX512DQ-BW-NEXT: vmovdqa64 %zmm5, 448(%r9)14178; AVX512DQ-BW-NEXT: vmovdqa64 %zmm12, 256(%r9)14179; AVX512DQ-BW-NEXT: vmovdqa64 %zmm25, 320(%r9)14180; AVX512DQ-BW-NEXT: vmovdqa64 %zmm20, 128(%r9)14181; AVX512DQ-BW-NEXT: vmovdqa64 %zmm21, 192(%r9)14182; AVX512DQ-BW-NEXT: vmovdqa64 %zmm23, (%r9)14183; AVX512DQ-BW-NEXT: vmovdqa64 %zmm31, 64(%r9)14184; AVX512DQ-BW-NEXT: addq $3336, %rsp # imm = 0xD0814185; AVX512DQ-BW-NEXT: vzeroupper14186; AVX512DQ-BW-NEXT: retq14187;14188; AVX512DQ-BW-FCP-LABEL: load_i64_stride5_vf64:14189; AVX512DQ-BW-FCP: # %bb.0:14190; AVX512DQ-BW-FCP-NEXT: subq $3336, %rsp # imm = 0xD0814191; AVX512DQ-BW-FCP-NEXT: vmovdqa64 1728(%rdi), %zmm314192; AVX512DQ-BW-FCP-NEXT: vmovdqa64 1792(%rdi), %zmm714193; AVX512DQ-BW-FCP-NEXT: vmovdqa64 1408(%rdi), %zmm014194; AVX512DQ-BW-FCP-NEXT: vmovdqa64 1088(%rdi), %zmm114195; AVX512DQ-BW-FCP-NEXT: vmovdqa64 1152(%rdi), %zmm614196; AVX512DQ-BW-FCP-NEXT: vmovdqa64 768(%rdi), %zmm514197; AVX512DQ-BW-FCP-NEXT: vmovdqa64 832(%rdi), %zmm814198; AVX512DQ-BW-FCP-NEXT: vmovdqa64 448(%rdi), %zmm414199; AVX512DQ-BW-FCP-NEXT: vmovdqa64 512(%rdi), %zmm914200; AVX512DQ-BW-FCP-NEXT: vmovdqa64 128(%rdi), %zmm214201; AVX512DQ-BW-FCP-NEXT: vmovdqa64 192(%rdi), %zmm1014202; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm16 = [12,1,6,0,12,1,6,0]14203; AVX512DQ-BW-FCP-NEXT: # zmm16 = mem[0,1,2,3,0,1,2,3]14204; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm10, %zmm1114205; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm2, %zmm16, %zmm1114206; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14207; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm9, %zmm1114208; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm4, %zmm16, %zmm1114209; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14210; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm8, %zmm1114211; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm5, %zmm16, %zmm1114212; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14213; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm1214214; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm1114215; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm1, %zmm16, %zmm1214216; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14217; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm7, %zmm614218; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm3, %zmm16, %zmm614219; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm6, (%rsp) # 64-byte Spill14220; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm12 = [5,10,15,0,5,10,15,0]14221; AVX512DQ-BW-FCP-NEXT: # zmm12 = mem[0,1,2,3,0,1,2,3]14222; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm614223; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm10, %zmm12, %zmm614224; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14225; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm4, %zmm614226; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm9, %zmm12, %zmm614227; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14228; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm614229; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm11, %zmm12, %zmm614230; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14231; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm5, %zmm614232; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm8, %zmm12, %zmm614233; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14234; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm614235; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm7, %zmm12, %zmm614236; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14237; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = [6,11,0,1,6,11,0,1]14238; AVX512DQ-BW-FCP-NEXT: # zmm6 = mem[0,1,2,3,0,1,2,3]14239; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm1314240; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm10, %zmm6, %zmm1314241; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14242; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm6, %zmm1314243; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm4, %zmm614244; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm9, %zmm13, %zmm614245; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14246; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm614247; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm11, %zmm13, %zmm614248; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14249; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm5, %zmm614250; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm8, %zmm13, %zmm614251; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14252; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm614253; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm7, %zmm13, %zmm614254; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14255; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm14 = [7,12,0,2,7,12,0,2]14256; AVX512DQ-BW-FCP-NEXT: # zmm14 = mem[0,1,2,3,0,1,2,3]14257; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm4, %zmm614258; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm9, %zmm14, %zmm614259; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14260; AVX512DQ-BW-FCP-NEXT: vbroadcasti64x4 {{.*#+}} zmm6 = [0,5,0,11,0,5,0,11]14261; AVX512DQ-BW-FCP-NEXT: # zmm6 = mem[0,1,2,3,0,1,2,3]14262; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm4, %zmm6, %zmm914263; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14264; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm414265; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm10, %zmm14, %zmm414266; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14267; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm2, %zmm6, %zmm1014268; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14269; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm214270; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm11, %zmm14, %zmm214271; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14272; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm1, %zmm6, %zmm1114273; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14274; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm5, %zmm114275; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm8, %zmm14, %zmm114276; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14277; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm5, %zmm6, %zmm814278; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14279; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm114280; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm7, %zmm14, %zmm114281; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14282; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm3, %zmm6, %zmm714283; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14284; AVX512DQ-BW-FCP-NEXT: vmovdqa64 1472(%rdi), %zmm214285; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm1914286; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm0, %zmm16, %zmm1914287; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm114288; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm2, %zmm12, %zmm114289; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14290; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm114291; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm2, %zmm13, %zmm114292; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14293; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm114294; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm2, %zmm14, %zmm114295; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14296; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm0, %zmm6, %zmm214297; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14298; AVX512DQ-BW-FCP-NEXT: vmovdqa64 2048(%rdi), %zmm014299; AVX512DQ-BW-FCP-NEXT: vmovdqa64 2112(%rdi), %zmm214300; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm1814301; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm0, %zmm16, %zmm1814302; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm114303; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm2, %zmm12, %zmm114304; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14305; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm114306; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm2, %zmm13, %zmm114307; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14308; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm114309; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm2, %zmm14, %zmm114310; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14311; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm0, %zmm6, %zmm214312; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14313; AVX512DQ-BW-FCP-NEXT: vmovdqa64 2368(%rdi), %zmm014314; AVX512DQ-BW-FCP-NEXT: vmovdqa64 2432(%rdi), %zmm114315; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm1, %zmm0, %zmm1214316; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm12, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14317; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm1, %zmm0, %zmm1314318; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14319; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm1, %zmm0, %zmm1414320; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14321; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm1714322; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm0, %zmm6, %zmm114323; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14324; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm0, %zmm16, %zmm1714325; AVX512DQ-BW-FCP-NEXT: vmovdqa64 384(%rdi), %zmm3114326; AVX512DQ-BW-FCP-NEXT: vmovdqa64 320(%rdi), %zmm214327; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} ymm3 = [0,5,10,15]14328; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm114329; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm31, %zmm3, %zmm114330; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} ymm29 = [1,6,11,0]14331; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm014332; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm31, %zmm29, %zmm014333; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14334; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} ymm6 = [2,7,12,0]14335; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm014336; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm31, %zmm6, %zmm014337; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14338; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} ymm7 = [11,0,5,0]14339; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm31, %zmm014340; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm2, %zmm7, %zmm014341; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14342; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm2, %zmm16, %zmm3114343; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rdi), %zmm214344; AVX512DQ-BW-FCP-NEXT: vmovdqa64 64(%rdi), %zmm2314345; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm414346; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm23, %zmm3, %zmm414347; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm014348; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm23, %zmm29, %zmm014349; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14350; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm014351; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm23, %zmm6, %zmm014352; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14353; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm23, %zmm014354; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm2, %zmm7, %zmm014355; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14356; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm2, %zmm16, %zmm2314357; AVX512DQ-BW-FCP-NEXT: vmovdqa64 1024(%rdi), %zmm2114358; AVX512DQ-BW-FCP-NEXT: vmovdqa64 960(%rdi), %zmm214359; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm1014360; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm21, %zmm3, %zmm1014361; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm014362; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm21, %zmm29, %zmm014363; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14364; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm014365; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm21, %zmm6, %zmm014366; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14367; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm21, %zmm014368; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm2, %zmm7, %zmm014369; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14370; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm2, %zmm16, %zmm2114371; AVX512DQ-BW-FCP-NEXT: vmovdqa64 704(%rdi), %zmm2014372; AVX512DQ-BW-FCP-NEXT: vmovdqa64 640(%rdi), %zmm214373; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm1114374; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm20, %zmm3, %zmm1114375; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm014376; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm20, %zmm29, %zmm014377; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14378; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm014379; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm20, %zmm6, %zmm014380; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14381; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm20, %zmm014382; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm2, %zmm7, %zmm014383; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14384; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm2, %zmm16, %zmm2014385; AVX512DQ-BW-FCP-NEXT: vmovdqa64 1664(%rdi), %zmm2514386; AVX512DQ-BW-FCP-NEXT: vmovdqa64 1600(%rdi), %zmm214387; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm1314388; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm25, %zmm3, %zmm1314389; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm014390; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm25, %zmm29, %zmm014391; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14392; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm3014393; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm25, %zmm6, %zmm3014394; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm25, %zmm014395; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm2, %zmm7, %zmm014396; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14397; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm2, %zmm16, %zmm2514398; AVX512DQ-BW-FCP-NEXT: vmovdqa64 1344(%rdi), %zmm1214399; AVX512DQ-BW-FCP-NEXT: vmovdqa64 1280(%rdi), %zmm214400; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm1414401; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm12, %zmm3, %zmm1414402; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm2614403; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm12, %zmm29, %zmm2614404; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm2714405; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm12, %zmm6, %zmm2714406; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm12, %zmm2814407; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm2, %zmm7, %zmm2814408; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm2, %zmm16, %zmm1214409; AVX512DQ-BW-FCP-NEXT: vmovdqa64 1984(%rdi), %zmm914410; AVX512DQ-BW-FCP-NEXT: vmovdqa64 1920(%rdi), %zmm1514411; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm15, %zmm214412; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm9, %zmm3, %zmm214413; AVX512DQ-BW-FCP-NEXT: vmovdqa64 2304(%rdi), %zmm514414; AVX512DQ-BW-FCP-NEXT: vmovdqa64 2240(%rdi), %zmm014415; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm5, %zmm0, %zmm314416; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm15, %zmm814417; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm9, %zmm29, %zmm814418; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm5, %zmm0, %zmm2914419; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm15, %zmm2214420; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm9, %zmm6, %zmm2214421; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm5, %zmm0, %zmm614422; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm9, %zmm2414423; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm15, %zmm7, %zmm2414424; AVX512DQ-BW-FCP-NEXT: vpermi2q %zmm0, %zmm5, %zmm714425; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm0, %zmm16, %zmm514426; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm15, %zmm16, %zmm914427; AVX512DQ-BW-FCP-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm4, %zmm0 # 64-byte Folded Reload14428; AVX512DQ-BW-FCP-NEXT: # zmm0 = zmm4[0,1,2,3],mem[4,5,6,7]14429; AVX512DQ-BW-FCP-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm1, %zmm4 # 64-byte Folded Reload14430; AVX512DQ-BW-FCP-NEXT: # zmm4 = zmm1[0,1,2,3],mem[4,5,6,7]14431; AVX512DQ-BW-FCP-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm11, %zmm11 # 64-byte Folded Reload14432; AVX512DQ-BW-FCP-NEXT: # zmm11 = zmm11[0,1,2,3],mem[4,5,6,7]14433; AVX512DQ-BW-FCP-NEXT: vshufi64x2 $228, {{[-0-9]+}}(%r{{[sb]}}p), %zmm10, %zmm10 # 64-byte Folded Reload14434; AVX512DQ-BW-FCP-NEXT: # zmm10 = zmm10[0,1,2,3],mem[4,5,6,7]14435; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm19 = zmm14[0,1,2,3],zmm19[4,5,6,7]14436; AVX512DQ-BW-FCP-NEXT: vshufi64x2 $228, (%rsp), %zmm13, %zmm1 # 64-byte Folded Reload14437; AVX512DQ-BW-FCP-NEXT: # zmm1 = zmm13[0,1,2,3],mem[4,5,6,7]14438; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm2 = zmm2[0,1,2,3],zmm18[4,5,6,7]14439; AVX512DQ-BW-FCP-NEXT: vshufi64x2 {{.*#+}} zmm3 = zmm3[0,1,2,3],zmm17[4,5,6,7]14440; AVX512DQ-BW-FCP-NEXT: vmovdqa64 256(%rdi), %zmm1814441; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm15 = [0,1,2,3,4,5,6,11]14442; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm18, %zmm15, %zmm014443; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14444; AVX512DQ-BW-FCP-NEXT: vmovdqa64 576(%rdi), %zmm1414445; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm14, %zmm15, %zmm414446; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14447; AVX512DQ-BW-FCP-NEXT: vmovdqa64 896(%rdi), %zmm1714448; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm17, %zmm15, %zmm1114449; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm11, (%rsp) # 64-byte Spill14450; AVX512DQ-BW-FCP-NEXT: vmovdqa64 1216(%rdi), %zmm1314451; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm13, %zmm15, %zmm1014452; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14453; AVX512DQ-BW-FCP-NEXT: vmovdqa64 1536(%rdi), %zmm1614454; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm16, %zmm15, %zmm1914455; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm19, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14456; AVX512DQ-BW-FCP-NEXT: vmovdqa64 1856(%rdi), %zmm1114457; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm11, %zmm15, %zmm114458; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14459; AVX512DQ-BW-FCP-NEXT: vmovdqa64 2176(%rdi), %zmm1914460; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm19, %zmm15, %zmm214461; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14462; AVX512DQ-BW-FCP-NEXT: vmovdqa64 2496(%rdi), %zmm1014463; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm10, %zmm15, %zmm314464; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14465; AVX512DQ-BW-FCP-NEXT: movb $7, %al14466; AVX512DQ-BW-FCP-NEXT: kmovd %eax, %k114467; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload14468; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14469; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm15 {%k1}14470; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload14471; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm8, %zmm4 {%k1}14472; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload14473; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14474; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm8 {%k1}14475; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14476; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload14477; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm0 {%k1}14478; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload14479; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14480; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm1 {%k1}14481; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14482; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload14483; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm2 {%k1}14484; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload14485; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm26, %zmm3 {%k1}14486; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload14487; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm29, %zmm26 {%k1}14488; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm29 = [0,1,2,3,4,5,6,12]14489; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm18, %zmm29, %zmm1514490; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm15, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14491; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm19, %zmm29, %zmm414492; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14493; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm14, %zmm29, %zmm814494; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14495; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm13, %zmm29, %zmm014496; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14497; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm17, %zmm29, %zmm114498; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14499; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm11, %zmm29, %zmm214500; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14501; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm16, %zmm29, %zmm314502; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14503; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm10, %zmm29, %zmm2614504; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm26, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14505; AVX512DQ-BW-FCP-NEXT: movb $56, %al14506; AVX512DQ-BW-FCP-NEXT: kmovd %eax, %k114507; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14508; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm0 # 64-byte Reload14509; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm0 {%k1}14510; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload14511; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm22 {%k1}14512; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14513; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload14514; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm8 {%k1}14515; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14516; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm26 # 64-byte Reload14517; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm26 {%k1}14518; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14519; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm15 # 64-byte Reload14520; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm15 {%k1}14521; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14522; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm30 {%k1}14523; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload14524; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm27 {%k1}14525; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload14526; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm6 {%k1}14527; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm1 = [0,1,2,3,4,5,8,13]14528; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm18, %zmm1, %zmm014529; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14530; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm19, %zmm1, %zmm2214531; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm14, %zmm1, %zmm814532; AVX512DQ-BW-FCP-NEXT: vmovdqu64 %zmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill14533; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm13, %zmm1, %zmm2614534; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm17, %zmm1, %zmm1514535; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm11, %zmm1, %zmm3014536; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm16, %zmm1, %zmm2714537; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm10, %zmm1, %zmm614538; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14539; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm8 # 64-byte Reload14540; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm8 {%k1}14541; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload14542; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm1, %zmm24 {%k1}14543; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14544; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm29 # 64-byte Reload14545; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm29 {%k1}14546; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14547; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm1 # 64-byte Reload14548; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm1 {%k1}14549; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14550; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm3 # 64-byte Reload14551; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm3 {%k1}14552; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14553; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm4 # 64-byte Reload14554; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm4 {%k1}14555; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14556; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm28 {%k1}14557; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14558; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm7 {%k1}14559; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm2 = [0,1,2,3,4,5,9,14]14560; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm18, %zmm2, %zmm814561; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm19, %zmm2, %zmm2414562; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm14, %zmm2, %zmm2914563; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm13, %zmm2, %zmm114564; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm17, %zmm2, %zmm314565; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, %zmm014566; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm11, %zmm2, %zmm414567; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm16, %zmm2, %zmm2814568; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm10, %zmm2, %zmm714569; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14570; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm31 {%k1}14571; AVX512DQ-BW-FCP-NEXT: vpmovsxbq {{.*#+}} zmm3 = [0,1,2,3,4,5,10,15]14572; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm14, %zmm3, %zmm3114573; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14574; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm23 {%k1}14575; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm18, %zmm3, %zmm2314576; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14577; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm21 {%k1}14578; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm13, %zmm3, %zmm2114579; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14580; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm20 {%k1}14581; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm17, %zmm3, %zmm2014582; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14583; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm25 {%k1}14584; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm11, %zmm3, %zmm2514585; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14586; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm12 {%k1}14587; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm16, %zmm3, %zmm1214588; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14589; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm5 {%k1}14590; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm10, %zmm3, %zmm514591; AVX512DQ-BW-FCP-NEXT: vmovdqu64 {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14592; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, %zmm9 {%k1}14593; AVX512DQ-BW-FCP-NEXT: vpermt2q %zmm19, %zmm3, %zmm914594; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14595; AVX512DQ-BW-FCP-NEXT: vmovaps %zmm2, 448(%rsi)14596; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14597; AVX512DQ-BW-FCP-NEXT: vmovaps %zmm2, 384(%rsi)14598; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14599; AVX512DQ-BW-FCP-NEXT: vmovaps %zmm2, 320(%rsi)14600; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14601; AVX512DQ-BW-FCP-NEXT: vmovaps %zmm2, 256(%rsi)14602; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14603; AVX512DQ-BW-FCP-NEXT: vmovaps %zmm2, 192(%rsi)14604; AVX512DQ-BW-FCP-NEXT: vmovups (%rsp), %zmm2 # 64-byte Reload14605; AVX512DQ-BW-FCP-NEXT: vmovaps %zmm2, 128(%rsi)14606; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14607; AVX512DQ-BW-FCP-NEXT: vmovaps %zmm2, 64(%rsi)14608; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14609; AVX512DQ-BW-FCP-NEXT: vmovaps %zmm2, (%rsi)14610; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14611; AVX512DQ-BW-FCP-NEXT: vmovaps %zmm2, 448(%rdx)14612; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14613; AVX512DQ-BW-FCP-NEXT: vmovaps %zmm2, 256(%rdx)14614; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14615; AVX512DQ-BW-FCP-NEXT: vmovaps %zmm2, 320(%rdx)14616; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14617; AVX512DQ-BW-FCP-NEXT: vmovaps %zmm2, 128(%rdx)14618; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14619; AVX512DQ-BW-FCP-NEXT: vmovaps %zmm2, 192(%rdx)14620; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14621; AVX512DQ-BW-FCP-NEXT: vmovaps %zmm2, (%rdx)14622; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14623; AVX512DQ-BW-FCP-NEXT: vmovaps %zmm2, 64(%rdx)14624; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14625; AVX512DQ-BW-FCP-NEXT: vmovaps %zmm2, 384(%rdx)14626; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm6, 448(%rcx)14627; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm27, 256(%rcx)14628; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm30, 320(%rcx)14629; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm15, 128(%rcx)14630; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm26, 192(%rcx)14631; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14632; AVX512DQ-BW-FCP-NEXT: vmovaps %zmm2, (%rcx)14633; AVX512DQ-BW-FCP-NEXT: vmovups {{[-0-9]+}}(%r{{[sb]}}p), %zmm2 # 64-byte Reload14634; AVX512DQ-BW-FCP-NEXT: vmovaps %zmm2, 64(%rcx)14635; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm22, 384(%rcx)14636; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm7, 448(%r8)14637; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm28, 256(%r8)14638; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm4, 320(%r8)14639; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, 128(%r8)14640; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm1, 192(%r8)14641; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm8, (%r8)14642; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm29, 64(%r8)14643; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm24, 384(%r8)14644; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm9, 384(%r9)14645; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm5, 448(%r9)14646; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm12, 256(%r9)14647; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm25, 320(%r9)14648; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm20, 128(%r9)14649; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm21, 192(%r9)14650; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm23, (%r9)14651; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm31, 64(%r9)14652; AVX512DQ-BW-FCP-NEXT: addq $3336, %rsp # imm = 0xD0814653; AVX512DQ-BW-FCP-NEXT: vzeroupper14654; AVX512DQ-BW-FCP-NEXT: retq14655 %wide.vec = load <320 x i64>, ptr %in.vec, align 6414656 %strided.vec0 = shufflevector <320 x i64> %wide.vec, <320 x i64> poison, <64 x i32> <i32 0, i32 5, i32 10, i32 15, i32 20, i32 25, i32 30, i32 35, i32 40, i32 45, i32 50, i32 55, i32 60, i32 65, i32 70, i32 75, i32 80, i32 85, i32 90, i32 95, i32 100, i32 105, i32 110, i32 115, i32 120, i32 125, i32 130, i32 135, i32 140, i32 145, i32 150, i32 155, i32 160, i32 165, i32 170, i32 175, i32 180, i32 185, i32 190, i32 195, i32 200, i32 205, i32 210, i32 215, i32 220, i32 225, i32 230, i32 235, i32 240, i32 245, i32 250, i32 255, i32 260, i32 265, i32 270, i32 275, i32 280, i32 285, i32 290, i32 295, i32 300, i32 305, i32 310, i32 315>14657 %strided.vec1 = shufflevector <320 x i64> %wide.vec, <320 x i64> poison, <64 x i32> <i32 1, i32 6, i32 11, i32 16, i32 21, i32 26, i32 31, i32 36, i32 41, i32 46, i32 51, i32 56, i32 61, i32 66, i32 71, i32 76, i32 81, i32 86, i32 91, i32 96, i32 101, i32 106, i32 111, i32 116, i32 121, i32 126, i32 131, i32 136, i32 141, i32 146, i32 151, i32 156, i32 161, i32 166, i32 171, i32 176, i32 181, i32 186, i32 191, i32 196, i32 201, i32 206, i32 211, i32 216, i32 221, i32 226, i32 231, i32 236, i32 241, i32 246, i32 251, i32 256, i32 261, i32 266, i32 271, i32 276, i32 281, i32 286, i32 291, i32 296, i32 301, i32 306, i32 311, i32 316>14658 %strided.vec2 = shufflevector <320 x i64> %wide.vec, <320 x i64> poison, <64 x i32> <i32 2, i32 7, i32 12, i32 17, i32 22, i32 27, i32 32, i32 37, i32 42, i32 47, i32 52, i32 57, i32 62, i32 67, i32 72, i32 77, i32 82, i32 87, i32 92, i32 97, i32 102, i32 107, i32 112, i32 117, i32 122, i32 127, i32 132, i32 137, i32 142, i32 147, i32 152, i32 157, i32 162, i32 167, i32 172, i32 177, i32 182, i32 187, i32 192, i32 197, i32 202, i32 207, i32 212, i32 217, i32 222, i32 227, i32 232, i32 237, i32 242, i32 247, i32 252, i32 257, i32 262, i32 267, i32 272, i32 277, i32 282, i32 287, i32 292, i32 297, i32 302, i32 307, i32 312, i32 317>14659 %strided.vec3 = shufflevector <320 x i64> %wide.vec, <320 x i64> poison, <64 x i32> <i32 3, i32 8, i32 13, i32 18, i32 23, i32 28, i32 33, i32 38, i32 43, i32 48, i32 53, i32 58, i32 63, i32 68, i32 73, i32 78, i32 83, i32 88, i32 93, i32 98, i32 103, i32 108, i32 113, i32 118, i32 123, i32 128, i32 133, i32 138, i32 143, i32 148, i32 153, i32 158, i32 163, i32 168, i32 173, i32 178, i32 183, i32 188, i32 193, i32 198, i32 203, i32 208, i32 213, i32 218, i32 223, i32 228, i32 233, i32 238, i32 243, i32 248, i32 253, i32 258, i32 263, i32 268, i32 273, i32 278, i32 283, i32 288, i32 293, i32 298, i32 303, i32 308, i32 313, i32 318>14660 %strided.vec4 = shufflevector <320 x i64> %wide.vec, <320 x i64> poison, <64 x i32> <i32 4, i32 9, i32 14, i32 19, i32 24, i32 29, i32 34, i32 39, i32 44, i32 49, i32 54, i32 59, i32 64, i32 69, i32 74, i32 79, i32 84, i32 89, i32 94, i32 99, i32 104, i32 109, i32 114, i32 119, i32 124, i32 129, i32 134, i32 139, i32 144, i32 149, i32 154, i32 159, i32 164, i32 169, i32 174, i32 179, i32 184, i32 189, i32 194, i32 199, i32 204, i32 209, i32 214, i32 219, i32 224, i32 229, i32 234, i32 239, i32 244, i32 249, i32 254, i32 259, i32 264, i32 269, i32 274, i32 279, i32 284, i32 289, i32 294, i32 299, i32 304, i32 309, i32 314, i32 319>14661 store <64 x i64> %strided.vec0, ptr %out.vec0, align 6414662 store <64 x i64> %strided.vec1, ptr %out.vec1, align 6414663 store <64 x i64> %strided.vec2, ptr %out.vec2, align 6414664 store <64 x i64> %strided.vec3, ptr %out.vec3, align 6414665 store <64 x i64> %strided.vec4, ptr %out.vec4, align 6414666 ret void14667}14668