1272 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=AVX512,AVX512-VL8; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX512,AVX512-FCP9; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+avx512dq | FileCheck %s --check-prefixes=AVX512,AVX512DQ10; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+avx512dq,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX512,AVX512DQ-FCP11; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+avx512bw | FileCheck %s --check-prefixes=AVX512,AVX512BW12; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+avx512bw,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX512,AVX512BW-FCP13; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+avx512dq,+avx512bw | FileCheck %s --check-prefixes=AVX512,AVX512DQ-BW14; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl,+avx512dq,+avx512bw,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX512,AVX512DQ-BW-FCP15 16; These patterns are produced by LoopVectorizer for interleaved loads.17 18define void @load_i16_stride2_vf2(ptr %in.vec, ptr %out.vec0, ptr %out.vec1) nounwind {19; SSE-LABEL: load_i16_stride2_vf2:20; SSE: # %bb.0:21; SSE-NEXT: movq {{.*#+}} xmm0 = mem[0],zero22; SSE-NEXT: pshuflw {{.*#+}} xmm1 = xmm0[0,2,2,3,4,5,6,7]23; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[1,3,2,3,4,5,6,7]24; SSE-NEXT: movd %xmm1, (%rsi)25; SSE-NEXT: movd %xmm0, (%rdx)26; SSE-NEXT: retq27;28; AVX-LABEL: load_i16_stride2_vf2:29; AVX: # %bb.0:30; AVX-NEXT: vmovq {{.*#+}} xmm0 = mem[0],zero31; AVX-NEXT: vpshuflw {{.*#+}} xmm1 = xmm0[0,2,2,3,4,5,6,7]32; AVX-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[1,3,2,3,4,5,6,7]33; AVX-NEXT: vmovd %xmm1, (%rsi)34; AVX-NEXT: vmovd %xmm0, (%rdx)35; AVX-NEXT: retq36;37; AVX2-LABEL: load_i16_stride2_vf2:38; AVX2: # %bb.0:39; AVX2-NEXT: vmovq {{.*#+}} xmm0 = mem[0],zero40; AVX2-NEXT: vpshuflw {{.*#+}} xmm1 = xmm0[0,2,2,3,4,5,6,7]41; AVX2-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[1,3,2,3,4,5,6,7]42; AVX2-NEXT: vmovd %xmm1, (%rsi)43; AVX2-NEXT: vmovd %xmm0, (%rdx)44; AVX2-NEXT: retq45;46; AVX2-FP-LABEL: load_i16_stride2_vf2:47; AVX2-FP: # %bb.0:48; AVX2-FP-NEXT: vmovq {{.*#+}} xmm0 = mem[0],zero49; AVX2-FP-NEXT: vpshuflw {{.*#+}} xmm1 = xmm0[0,2,2,3,4,5,6,7]50; AVX2-FP-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[1,3,2,3,4,5,6,7]51; AVX2-FP-NEXT: vmovd %xmm1, (%rsi)52; AVX2-FP-NEXT: vmovd %xmm0, (%rdx)53; AVX2-FP-NEXT: retq54;55; AVX2-FCP-LABEL: load_i16_stride2_vf2:56; AVX2-FCP: # %bb.0:57; AVX2-FCP-NEXT: vmovq {{.*#+}} xmm0 = mem[0],zero58; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm1 = xmm0[0,2,2,3,4,5,6,7]59; AVX2-FCP-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[1,3,2,3,4,5,6,7]60; AVX2-FCP-NEXT: vmovd %xmm1, (%rsi)61; AVX2-FCP-NEXT: vmovd %xmm0, (%rdx)62; AVX2-FCP-NEXT: retq63;64; AVX512-LABEL: load_i16_stride2_vf2:65; AVX512: # %bb.0:66; AVX512-NEXT: vmovq {{.*#+}} xmm0 = mem[0],zero67; AVX512-NEXT: vpshuflw {{.*#+}} xmm1 = xmm0[0,2,2,3,4,5,6,7]68; AVX512-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[1,3,2,3,4,5,6,7]69; AVX512-NEXT: vmovd %xmm1, (%rsi)70; AVX512-NEXT: vmovd %xmm0, (%rdx)71; AVX512-NEXT: retq72 %wide.vec = load <4 x i16>, ptr %in.vec, align 6473 %strided.vec0 = shufflevector <4 x i16> %wide.vec, <4 x i16> poison, <2 x i32> <i32 0, i32 2>74 %strided.vec1 = shufflevector <4 x i16> %wide.vec, <4 x i16> poison, <2 x i32> <i32 1, i32 3>75 store <2 x i16> %strided.vec0, ptr %out.vec0, align 6476 store <2 x i16> %strided.vec1, ptr %out.vec1, align 6477 ret void78}79 80define void @load_i16_stride2_vf4(ptr %in.vec, ptr %out.vec0, ptr %out.vec1) nounwind {81; SSE-LABEL: load_i16_stride2_vf4:82; SSE: # %bb.0:83; SSE-NEXT: movdqa (%rdi), %xmm084; SSE-NEXT: pshuflw {{.*#+}} xmm1 = xmm0[0,2,2,3,4,5,6,7]85; SSE-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,4,6,6,7]86; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,2,2,3]87; SSE-NEXT: movq %xmm1, (%rsi)88; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[3,1,2,3,4,5,6,7]89; SSE-NEXT: pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,7,5,6,7]90; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]91; SSE-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[1,0,3,2,4,5,6,7]92; SSE-NEXT: movq %xmm0, (%rdx)93; SSE-NEXT: retq94;95; AVX-LABEL: load_i16_stride2_vf4:96; AVX: # %bb.0:97; AVX-NEXT: vmovdqa (%rdi), %xmm098; AVX-NEXT: vpshufb {{.*#+}} xmm1 = xmm0[0,1,4,5,8,9,12,13,u,u,u,u,u,u,u,u]99; AVX-NEXT: vmovq %xmm1, (%rsi)100; AVX-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[2,3,6,7,10,11,14,15,u,u,u,u,u,u,u,u]101; AVX-NEXT: vmovq %xmm0, (%rdx)102; AVX-NEXT: retq103;104; AVX2-LABEL: load_i16_stride2_vf4:105; AVX2: # %bb.0:106; AVX2-NEXT: vmovdqa (%rdi), %xmm0107; AVX2-NEXT: vpshufb {{.*#+}} xmm1 = xmm0[0,1,4,5,8,9,12,13,u,u,u,u,u,u,u,u]108; AVX2-NEXT: vmovq %xmm1, (%rsi)109; AVX2-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[2,3,6,7,10,11,14,15,u,u,u,u,u,u,u,u]110; AVX2-NEXT: vmovq %xmm0, (%rdx)111; AVX2-NEXT: retq112;113; AVX2-FP-LABEL: load_i16_stride2_vf4:114; AVX2-FP: # %bb.0:115; AVX2-FP-NEXT: vmovdqa (%rdi), %xmm0116; AVX2-FP-NEXT: vpshufb {{.*#+}} xmm1 = xmm0[0,1,4,5,8,9,12,13,u,u,u,u,u,u,u,u]117; AVX2-FP-NEXT: vmovq %xmm1, (%rsi)118; AVX2-FP-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[2,3,6,7,10,11,14,15,u,u,u,u,u,u,u,u]119; AVX2-FP-NEXT: vmovq %xmm0, (%rdx)120; AVX2-FP-NEXT: retq121;122; AVX2-FCP-LABEL: load_i16_stride2_vf4:123; AVX2-FCP: # %bb.0:124; AVX2-FCP-NEXT: vmovdqa (%rdi), %xmm0125; AVX2-FCP-NEXT: vpshufb {{.*#+}} xmm1 = xmm0[0,1,4,5,8,9,12,13,u,u,u,u,u,u,u,u]126; AVX2-FCP-NEXT: vmovq %xmm1, (%rsi)127; AVX2-FCP-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[2,3,6,7,10,11,14,15,u,u,u,u,u,u,u,u]128; AVX2-FCP-NEXT: vmovq %xmm0, (%rdx)129; AVX2-FCP-NEXT: retq130;131; AVX512-LABEL: load_i16_stride2_vf4:132; AVX512: # %bb.0:133; AVX512-NEXT: vmovdqa (%rdi), %xmm0134; AVX512-NEXT: vpmovdw %xmm0, (%rsi)135; AVX512-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[2,3,6,7,10,11,14,15,u,u,u,u,u,u,u,u]136; AVX512-NEXT: vmovq %xmm0, (%rdx)137; AVX512-NEXT: retq138 %wide.vec = load <8 x i16>, ptr %in.vec, align 64139 %strided.vec0 = shufflevector <8 x i16> %wide.vec, <8 x i16> poison, <4 x i32> <i32 0, i32 2, i32 4, i32 6>140 %strided.vec1 = shufflevector <8 x i16> %wide.vec, <8 x i16> poison, <4 x i32> <i32 1, i32 3, i32 5, i32 7>141 store <4 x i16> %strided.vec0, ptr %out.vec0, align 64142 store <4 x i16> %strided.vec1, ptr %out.vec1, align 64143 ret void144}145 146define void @load_i16_stride2_vf8(ptr %in.vec, ptr %out.vec0, ptr %out.vec1) nounwind {147; SSE-LABEL: load_i16_stride2_vf8:148; SSE: # %bb.0:149; SSE-NEXT: movdqa (%rdi), %xmm0150; SSE-NEXT: movdqa 16(%rdi), %xmm1151; SSE-NEXT: movdqa %xmm1, %xmm2152; SSE-NEXT: pslld $16, %xmm2153; SSE-NEXT: psrad $16, %xmm2154; SSE-NEXT: movdqa %xmm0, %xmm3155; SSE-NEXT: pslld $16, %xmm3156; SSE-NEXT: psrad $16, %xmm3157; SSE-NEXT: packssdw %xmm2, %xmm3158; SSE-NEXT: psrad $16, %xmm1159; SSE-NEXT: psrad $16, %xmm0160; SSE-NEXT: packssdw %xmm1, %xmm0161; SSE-NEXT: movdqa %xmm3, (%rsi)162; SSE-NEXT: movdqa %xmm0, (%rdx)163; SSE-NEXT: retq164;165; AVX-LABEL: load_i16_stride2_vf8:166; AVX: # %bb.0:167; AVX-NEXT: vpxor %xmm0, %xmm0, %xmm0168; AVX-NEXT: vmovdqa (%rdi), %xmm1169; AVX-NEXT: vmovdqa 16(%rdi), %xmm2170; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm2[0],xmm0[1],xmm2[2],xmm0[3],xmm2[4],xmm0[5],xmm2[6],xmm0[7]171; AVX-NEXT: vpblendw {{.*#+}} xmm0 = xmm1[0],xmm0[1],xmm1[2],xmm0[3],xmm1[4],xmm0[5],xmm1[6],xmm0[7]172; AVX-NEXT: vpackusdw %xmm3, %xmm0, %xmm0173; AVX-NEXT: vpsrld $16, %xmm2, %xmm2174; AVX-NEXT: vpsrld $16, %xmm1, %xmm1175; AVX-NEXT: vpackusdw %xmm2, %xmm1, %xmm1176; AVX-NEXT: vmovdqa %xmm0, (%rsi)177; AVX-NEXT: vmovdqa %xmm1, (%rdx)178; AVX-NEXT: retq179;180; AVX2-LABEL: load_i16_stride2_vf8:181; AVX2: # %bb.0:182; AVX2-NEXT: vpxor %xmm0, %xmm0, %xmm0183; AVX2-NEXT: vmovdqa (%rdi), %xmm1184; AVX2-NEXT: vmovdqa 16(%rdi), %xmm2185; AVX2-NEXT: vpblendw {{.*#+}} xmm3 = xmm2[0],xmm0[1],xmm2[2],xmm0[3],xmm2[4],xmm0[5],xmm2[6],xmm0[7]186; AVX2-NEXT: vpblendw {{.*#+}} xmm0 = xmm1[0],xmm0[1],xmm1[2],xmm0[3],xmm1[4],xmm0[5],xmm1[6],xmm0[7]187; AVX2-NEXT: vpackusdw %xmm3, %xmm0, %xmm0188; AVX2-NEXT: vpsrld $16, %xmm2, %xmm2189; AVX2-NEXT: vpsrld $16, %xmm1, %xmm1190; AVX2-NEXT: vpackusdw %xmm2, %xmm1, %xmm1191; AVX2-NEXT: vmovdqa %xmm0, (%rsi)192; AVX2-NEXT: vmovdqa %xmm1, (%rdx)193; AVX2-NEXT: retq194;195; AVX2-FP-LABEL: load_i16_stride2_vf8:196; AVX2-FP: # %bb.0:197; AVX2-FP-NEXT: vpxor %xmm0, %xmm0, %xmm0198; AVX2-FP-NEXT: vmovdqa (%rdi), %xmm1199; AVX2-FP-NEXT: vmovdqa 16(%rdi), %xmm2200; AVX2-FP-NEXT: vpblendw {{.*#+}} xmm3 = xmm2[0],xmm0[1],xmm2[2],xmm0[3],xmm2[4],xmm0[5],xmm2[6],xmm0[7]201; AVX2-FP-NEXT: vpblendw {{.*#+}} xmm0 = xmm1[0],xmm0[1],xmm1[2],xmm0[3],xmm1[4],xmm0[5],xmm1[6],xmm0[7]202; AVX2-FP-NEXT: vpackusdw %xmm3, %xmm0, %xmm0203; AVX2-FP-NEXT: vpsrld $16, %xmm2, %xmm2204; AVX2-FP-NEXT: vpsrld $16, %xmm1, %xmm1205; AVX2-FP-NEXT: vpackusdw %xmm2, %xmm1, %xmm1206; AVX2-FP-NEXT: vmovdqa %xmm0, (%rsi)207; AVX2-FP-NEXT: vmovdqa %xmm1, (%rdx)208; AVX2-FP-NEXT: retq209;210; AVX2-FCP-LABEL: load_i16_stride2_vf8:211; AVX2-FCP: # %bb.0:212; AVX2-FCP-NEXT: vpxor %xmm0, %xmm0, %xmm0213; AVX2-FCP-NEXT: vmovdqa (%rdi), %xmm1214; AVX2-FCP-NEXT: vmovdqa 16(%rdi), %xmm2215; AVX2-FCP-NEXT: vpblendw {{.*#+}} xmm3 = xmm2[0],xmm0[1],xmm2[2],xmm0[3],xmm2[4],xmm0[5],xmm2[6],xmm0[7]216; AVX2-FCP-NEXT: vpblendw {{.*#+}} xmm0 = xmm1[0],xmm0[1],xmm1[2],xmm0[3],xmm1[4],xmm0[5],xmm1[6],xmm0[7]217; AVX2-FCP-NEXT: vpackusdw %xmm3, %xmm0, %xmm0218; AVX2-FCP-NEXT: vpsrld $16, %xmm2, %xmm2219; AVX2-FCP-NEXT: vpsrld $16, %xmm1, %xmm1220; AVX2-FCP-NEXT: vpackusdw %xmm2, %xmm1, %xmm1221; AVX2-FCP-NEXT: vmovdqa %xmm0, (%rsi)222; AVX2-FCP-NEXT: vmovdqa %xmm1, (%rdx)223; AVX2-FCP-NEXT: retq224;225; AVX512-LABEL: load_i16_stride2_vf8:226; AVX512: # %bb.0:227; AVX512-NEXT: vmovdqa (%rdi), %ymm0228; AVX512-NEXT: vpsrld $16, %ymm0, %ymm1229; AVX512-NEXT: vpmovdw %ymm0, (%rsi)230; AVX512-NEXT: vpmovdw %ymm1, (%rdx)231; AVX512-NEXT: vzeroupper232; AVX512-NEXT: retq233 %wide.vec = load <16 x i16>, ptr %in.vec, align 64234 %strided.vec0 = shufflevector <16 x i16> %wide.vec, <16 x i16> poison, <8 x i32> <i32 0, i32 2, i32 4, i32 6, i32 8, i32 10, i32 12, i32 14>235 %strided.vec1 = shufflevector <16 x i16> %wide.vec, <16 x i16> poison, <8 x i32> <i32 1, i32 3, i32 5, i32 7, i32 9, i32 11, i32 13, i32 15>236 store <8 x i16> %strided.vec0, ptr %out.vec0, align 64237 store <8 x i16> %strided.vec1, ptr %out.vec1, align 64238 ret void239}240 241define void @load_i16_stride2_vf16(ptr %in.vec, ptr %out.vec0, ptr %out.vec1) nounwind {242; SSE-LABEL: load_i16_stride2_vf16:243; SSE: # %bb.0:244; SSE-NEXT: movdqa (%rdi), %xmm0245; SSE-NEXT: movdqa 16(%rdi), %xmm2246; SSE-NEXT: movdqa 32(%rdi), %xmm1247; SSE-NEXT: movdqa 48(%rdi), %xmm3248; SSE-NEXT: movdqa %xmm3, %xmm4249; SSE-NEXT: pslld $16, %xmm4250; SSE-NEXT: psrad $16, %xmm4251; SSE-NEXT: movdqa %xmm1, %xmm5252; SSE-NEXT: pslld $16, %xmm5253; SSE-NEXT: psrad $16, %xmm5254; SSE-NEXT: packssdw %xmm4, %xmm5255; SSE-NEXT: movdqa %xmm2, %xmm4256; SSE-NEXT: pslld $16, %xmm4257; SSE-NEXT: psrad $16, %xmm4258; SSE-NEXT: movdqa %xmm0, %xmm6259; SSE-NEXT: pslld $16, %xmm6260; SSE-NEXT: psrad $16, %xmm6261; SSE-NEXT: packssdw %xmm4, %xmm6262; SSE-NEXT: psrad $16, %xmm3263; SSE-NEXT: psrad $16, %xmm1264; SSE-NEXT: packssdw %xmm3, %xmm1265; SSE-NEXT: psrad $16, %xmm2266; SSE-NEXT: psrad $16, %xmm0267; SSE-NEXT: packssdw %xmm2, %xmm0268; SSE-NEXT: movdqa %xmm6, (%rsi)269; SSE-NEXT: movdqa %xmm5, 16(%rsi)270; SSE-NEXT: movdqa %xmm0, (%rdx)271; SSE-NEXT: movdqa %xmm1, 16(%rdx)272; SSE-NEXT: retq273;274; AVX-LABEL: load_i16_stride2_vf16:275; AVX: # %bb.0:276; AVX-NEXT: vpxor %xmm0, %xmm0, %xmm0277; AVX-NEXT: vmovdqa (%rdi), %xmm1278; AVX-NEXT: vmovdqa 16(%rdi), %xmm2279; AVX-NEXT: vmovdqa 32(%rdi), %xmm3280; AVX-NEXT: vmovdqa 48(%rdi), %xmm4281; AVX-NEXT: vpblendw {{.*#+}} xmm5 = xmm4[0],xmm0[1],xmm4[2],xmm0[3],xmm4[4],xmm0[5],xmm4[6],xmm0[7]282; AVX-NEXT: vpblendw {{.*#+}} xmm6 = xmm3[0],xmm0[1],xmm3[2],xmm0[3],xmm3[4],xmm0[5],xmm3[6],xmm0[7]283; AVX-NEXT: vpackusdw %xmm5, %xmm6, %xmm5284; AVX-NEXT: vpblendw {{.*#+}} xmm6 = xmm2[0],xmm0[1],xmm2[2],xmm0[3],xmm2[4],xmm0[5],xmm2[6],xmm0[7]285; AVX-NEXT: vpblendw {{.*#+}} xmm0 = xmm1[0],xmm0[1],xmm1[2],xmm0[3],xmm1[4],xmm0[5],xmm1[6],xmm0[7]286; AVX-NEXT: vpackusdw %xmm6, %xmm0, %xmm0287; AVX-NEXT: vpsrld $16, %xmm4, %xmm4288; AVX-NEXT: vpsrld $16, %xmm3, %xmm3289; AVX-NEXT: vpackusdw %xmm4, %xmm3, %xmm3290; AVX-NEXT: vpsrld $16, %xmm2, %xmm2291; AVX-NEXT: vpsrld $16, %xmm1, %xmm1292; AVX-NEXT: vpackusdw %xmm2, %xmm1, %xmm1293; AVX-NEXT: vmovdqa %xmm0, (%rsi)294; AVX-NEXT: vmovdqa %xmm5, 16(%rsi)295; AVX-NEXT: vmovdqa %xmm1, (%rdx)296; AVX-NEXT: vmovdqa %xmm3, 16(%rdx)297; AVX-NEXT: retq298;299; AVX2-LABEL: load_i16_stride2_vf16:300; AVX2: # %bb.0:301; AVX2-NEXT: vmovdqa (%rdi), %ymm0302; AVX2-NEXT: vmovdqa 32(%rdi), %ymm1303; AVX2-NEXT: vpshuflw {{.*#+}} ymm2 = ymm1[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]304; AVX2-NEXT: vpshufhw {{.*#+}} ymm2 = ymm2[0,1,2,3,4,6,6,7,8,9,10,11,12,14,14,15]305; AVX2-NEXT: vpshuflw {{.*#+}} ymm3 = ymm0[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]306; AVX2-NEXT: vpshufhw {{.*#+}} ymm3 = ymm3[0,1,2,3,4,6,6,7,8,9,10,11,12,14,14,15]307; AVX2-NEXT: vshufps {{.*#+}} ymm2 = ymm3[0,2],ymm2[0,2],ymm3[4,6],ymm2[4,6]308; AVX2-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,2,1,3]309; AVX2-NEXT: vmovdqa {{.*#+}} ymm3 = [2,3,6,7,10,11,14,15,2,3,6,7,10,11,14,15,18,19,22,23,26,27,30,31,18,19,22,23,26,27,30,31]310; AVX2-NEXT: vpshufb %ymm3, %ymm1, %ymm1311; AVX2-NEXT: vpshufb %ymm3, %ymm0, %ymm0312; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7]313; AVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]314; AVX2-NEXT: vmovaps %ymm2, (%rsi)315; AVX2-NEXT: vmovdqa %ymm0, (%rdx)316; AVX2-NEXT: vzeroupper317; AVX2-NEXT: retq318;319; AVX2-FP-LABEL: load_i16_stride2_vf16:320; AVX2-FP: # %bb.0:321; AVX2-FP-NEXT: vmovdqa (%rdi), %ymm0322; AVX2-FP-NEXT: vmovdqa 32(%rdi), %ymm1323; AVX2-FP-NEXT: vpmovsxdq {{.*#+}} ymm2 = [84148480,218892552,353636624,488380696]324; AVX2-FP-NEXT: vpshufb %ymm2, %ymm1, %ymm3325; AVX2-FP-NEXT: vpshufb %ymm2, %ymm0, %ymm2326; AVX2-FP-NEXT: vshufps {{.*#+}} ymm2 = ymm2[0,2],ymm3[0,2],ymm2[4,6],ymm3[4,6]327; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,2,1,3]328; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm3 = [2,3,6,7,10,11,14,15,2,3,6,7,10,11,14,15,18,19,22,23,26,27,30,31,18,19,22,23,26,27,30,31]329; AVX2-FP-NEXT: vpshufb %ymm3, %ymm1, %ymm1330; AVX2-FP-NEXT: vpshufb %ymm3, %ymm0, %ymm0331; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7]332; AVX2-FP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]333; AVX2-FP-NEXT: vmovaps %ymm2, (%rsi)334; AVX2-FP-NEXT: vmovdqa %ymm0, (%rdx)335; AVX2-FP-NEXT: vzeroupper336; AVX2-FP-NEXT: retq337;338; AVX2-FCP-LABEL: load_i16_stride2_vf16:339; AVX2-FCP: # %bb.0:340; AVX2-FCP-NEXT: vmovdqa (%rdi), %ymm0341; AVX2-FCP-NEXT: vmovdqa 32(%rdi), %ymm1342; AVX2-FCP-NEXT: vpmovsxdq {{.*#+}} ymm2 = [84148480,218892552,353636624,488380696]343; AVX2-FCP-NEXT: vpshufb %ymm2, %ymm1, %ymm3344; AVX2-FCP-NEXT: vpshufb %ymm2, %ymm0, %ymm2345; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm2 = ymm2[0,2],ymm3[0,2],ymm2[4,6],ymm3[4,6]346; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,2,1,3]347; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm3 = [2,3,6,7,10,11,14,15,2,3,6,7,10,11,14,15,18,19,22,23,26,27,30,31,18,19,22,23,26,27,30,31]348; AVX2-FCP-NEXT: vpshufb %ymm3, %ymm1, %ymm1349; AVX2-FCP-NEXT: vpshufb %ymm3, %ymm0, %ymm0350; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7]351; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]352; AVX2-FCP-NEXT: vmovaps %ymm2, (%rsi)353; AVX2-FCP-NEXT: vmovdqa %ymm0, (%rdx)354; AVX2-FCP-NEXT: vzeroupper355; AVX2-FCP-NEXT: retq356;357; AVX512-LABEL: load_i16_stride2_vf16:358; AVX512: # %bb.0:359; AVX512-NEXT: vmovdqa64 (%rdi), %zmm0360; AVX512-NEXT: vpsrld $16, %zmm0, %zmm1361; AVX512-NEXT: vpmovdw %zmm0, (%rsi)362; AVX512-NEXT: vpmovdw %zmm1, (%rdx)363; AVX512-NEXT: vzeroupper364; AVX512-NEXT: retq365 %wide.vec = load <32 x i16>, ptr %in.vec, align 64366 %strided.vec0 = shufflevector <32 x i16> %wide.vec, <32 x i16> poison, <16 x i32> <i32 0, i32 2, i32 4, i32 6, i32 8, i32 10, i32 12, i32 14, i32 16, i32 18, i32 20, i32 22, i32 24, i32 26, i32 28, i32 30>367 %strided.vec1 = shufflevector <32 x i16> %wide.vec, <32 x i16> poison, <16 x i32> <i32 1, i32 3, i32 5, i32 7, i32 9, i32 11, i32 13, i32 15, i32 17, i32 19, i32 21, i32 23, i32 25, i32 27, i32 29, i32 31>368 store <16 x i16> %strided.vec0, ptr %out.vec0, align 64369 store <16 x i16> %strided.vec1, ptr %out.vec1, align 64370 ret void371}372 373define void @load_i16_stride2_vf32(ptr %in.vec, ptr %out.vec0, ptr %out.vec1) nounwind {374; SSE-LABEL: load_i16_stride2_vf32:375; SSE: # %bb.0:376; SSE-NEXT: movdqa 64(%rdi), %xmm0377; SSE-NEXT: movdqa 80(%rdi), %xmm4378; SSE-NEXT: movdqa 96(%rdi), %xmm1379; SSE-NEXT: movdqa 112(%rdi), %xmm6380; SSE-NEXT: movdqa (%rdi), %xmm2381; SSE-NEXT: movdqa 16(%rdi), %xmm7382; SSE-NEXT: movdqa 32(%rdi), %xmm3383; SSE-NEXT: movdqa 48(%rdi), %xmm9384; SSE-NEXT: movdqa %xmm9, %xmm8385; SSE-NEXT: pslld $16, %xmm8386; SSE-NEXT: psrad $16, %xmm8387; SSE-NEXT: movdqa %xmm3, %xmm5388; SSE-NEXT: pslld $16, %xmm5389; SSE-NEXT: psrad $16, %xmm5390; SSE-NEXT: packssdw %xmm8, %xmm5391; SSE-NEXT: movdqa %xmm7, %xmm10392; SSE-NEXT: pslld $16, %xmm10393; SSE-NEXT: psrad $16, %xmm10394; SSE-NEXT: movdqa %xmm2, %xmm8395; SSE-NEXT: pslld $16, %xmm8396; SSE-NEXT: psrad $16, %xmm8397; SSE-NEXT: packssdw %xmm10, %xmm8398; SSE-NEXT: movdqa %xmm6, %xmm11399; SSE-NEXT: pslld $16, %xmm11400; SSE-NEXT: psrad $16, %xmm11401; SSE-NEXT: movdqa %xmm1, %xmm10402; SSE-NEXT: pslld $16, %xmm10403; SSE-NEXT: psrad $16, %xmm10404; SSE-NEXT: packssdw %xmm11, %xmm10405; SSE-NEXT: movdqa %xmm4, %xmm11406; SSE-NEXT: pslld $16, %xmm11407; SSE-NEXT: psrad $16, %xmm11408; SSE-NEXT: movdqa %xmm0, %xmm12409; SSE-NEXT: pslld $16, %xmm12410; SSE-NEXT: psrad $16, %xmm12411; SSE-NEXT: packssdw %xmm11, %xmm12412; SSE-NEXT: psrad $16, %xmm9413; SSE-NEXT: psrad $16, %xmm3414; SSE-NEXT: packssdw %xmm9, %xmm3415; SSE-NEXT: psrad $16, %xmm7416; SSE-NEXT: psrad $16, %xmm2417; SSE-NEXT: packssdw %xmm7, %xmm2418; SSE-NEXT: psrad $16, %xmm6419; SSE-NEXT: psrad $16, %xmm1420; SSE-NEXT: packssdw %xmm6, %xmm1421; SSE-NEXT: psrad $16, %xmm4422; SSE-NEXT: psrad $16, %xmm0423; SSE-NEXT: packssdw %xmm4, %xmm0424; SSE-NEXT: movdqa %xmm12, 32(%rsi)425; SSE-NEXT: movdqa %xmm10, 48(%rsi)426; SSE-NEXT: movdqa %xmm8, (%rsi)427; SSE-NEXT: movdqa %xmm5, 16(%rsi)428; SSE-NEXT: movdqa %xmm0, 32(%rdx)429; SSE-NEXT: movdqa %xmm1, 48(%rdx)430; SSE-NEXT: movdqa %xmm2, (%rdx)431; SSE-NEXT: movdqa %xmm3, 16(%rdx)432; SSE-NEXT: retq433;434; AVX-LABEL: load_i16_stride2_vf32:435; AVX: # %bb.0:436; AVX-NEXT: vpxor %xmm0, %xmm0, %xmm0437; AVX-NEXT: vmovdqa (%rdi), %xmm1438; AVX-NEXT: vmovdqa 16(%rdi), %xmm2439; AVX-NEXT: vmovdqa 32(%rdi), %xmm3440; AVX-NEXT: vmovdqa 48(%rdi), %xmm4441; AVX-NEXT: vpblendw {{.*#+}} xmm5 = xmm4[0],xmm0[1],xmm4[2],xmm0[3],xmm4[4],xmm0[5],xmm4[6],xmm0[7]442; AVX-NEXT: vpblendw {{.*#+}} xmm6 = xmm3[0],xmm0[1],xmm3[2],xmm0[3],xmm3[4],xmm0[5],xmm3[6],xmm0[7]443; AVX-NEXT: vpackusdw %xmm5, %xmm6, %xmm5444; AVX-NEXT: vpblendw {{.*#+}} xmm6 = xmm2[0],xmm0[1],xmm2[2],xmm0[3],xmm2[4],xmm0[5],xmm2[6],xmm0[7]445; AVX-NEXT: vpblendw {{.*#+}} xmm7 = xmm1[0],xmm0[1],xmm1[2],xmm0[3],xmm1[4],xmm0[5],xmm1[6],xmm0[7]446; AVX-NEXT: vpackusdw %xmm6, %xmm7, %xmm6447; AVX-NEXT: vmovdqa 112(%rdi), %xmm7448; AVX-NEXT: vpblendw {{.*#+}} xmm8 = xmm7[0],xmm0[1],xmm7[2],xmm0[3],xmm7[4],xmm0[5],xmm7[6],xmm0[7]449; AVX-NEXT: vmovdqa 96(%rdi), %xmm9450; AVX-NEXT: vpblendw {{.*#+}} xmm10 = xmm9[0],xmm0[1],xmm9[2],xmm0[3],xmm9[4],xmm0[5],xmm9[6],xmm0[7]451; AVX-NEXT: vpackusdw %xmm8, %xmm10, %xmm8452; AVX-NEXT: vmovdqa 80(%rdi), %xmm10453; AVX-NEXT: vpblendw {{.*#+}} xmm11 = xmm10[0],xmm0[1],xmm10[2],xmm0[3],xmm10[4],xmm0[5],xmm10[6],xmm0[7]454; AVX-NEXT: vmovdqa 64(%rdi), %xmm12455; AVX-NEXT: vpblendw {{.*#+}} xmm0 = xmm12[0],xmm0[1],xmm12[2],xmm0[3],xmm12[4],xmm0[5],xmm12[6],xmm0[7]456; AVX-NEXT: vpackusdw %xmm11, %xmm0, %xmm0457; AVX-NEXT: vpsrld $16, %xmm4, %xmm4458; AVX-NEXT: vpsrld $16, %xmm3, %xmm3459; AVX-NEXT: vpackusdw %xmm4, %xmm3, %xmm3460; AVX-NEXT: vpsrld $16, %xmm2, %xmm2461; AVX-NEXT: vpsrld $16, %xmm1, %xmm1462; AVX-NEXT: vpackusdw %xmm2, %xmm1, %xmm1463; AVX-NEXT: vpsrld $16, %xmm7, %xmm2464; AVX-NEXT: vpsrld $16, %xmm9, %xmm4465; AVX-NEXT: vpackusdw %xmm2, %xmm4, %xmm2466; AVX-NEXT: vpsrld $16, %xmm10, %xmm4467; AVX-NEXT: vpsrld $16, %xmm12, %xmm7468; AVX-NEXT: vpackusdw %xmm4, %xmm7, %xmm4469; AVX-NEXT: vmovdqa %xmm0, 32(%rsi)470; AVX-NEXT: vmovdqa %xmm8, 48(%rsi)471; AVX-NEXT: vmovdqa %xmm6, (%rsi)472; AVX-NEXT: vmovdqa %xmm5, 16(%rsi)473; AVX-NEXT: vmovdqa %xmm4, 32(%rdx)474; AVX-NEXT: vmovdqa %xmm2, 48(%rdx)475; AVX-NEXT: vmovdqa %xmm1, (%rdx)476; AVX-NEXT: vmovdqa %xmm3, 16(%rdx)477; AVX-NEXT: retq478;479; AVX2-LABEL: load_i16_stride2_vf32:480; AVX2: # %bb.0:481; AVX2-NEXT: vmovdqa (%rdi), %ymm0482; AVX2-NEXT: vmovdqa 32(%rdi), %ymm1483; AVX2-NEXT: vmovdqa 64(%rdi), %ymm2484; AVX2-NEXT: vmovdqa 96(%rdi), %ymm3485; AVX2-NEXT: vpshuflw {{.*#+}} ymm4 = ymm3[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]486; AVX2-NEXT: vpshufhw {{.*#+}} ymm4 = ymm4[0,1,2,3,4,6,6,7,8,9,10,11,12,14,14,15]487; AVX2-NEXT: vpshuflw {{.*#+}} ymm5 = ymm2[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]488; AVX2-NEXT: vpshufhw {{.*#+}} ymm5 = ymm5[0,1,2,3,4,6,6,7,8,9,10,11,12,14,14,15]489; AVX2-NEXT: vshufps {{.*#+}} ymm4 = ymm5[0,2],ymm4[0,2],ymm5[4,6],ymm4[4,6]490; AVX2-NEXT: vpermpd {{.*#+}} ymm4 = ymm4[0,2,1,3]491; AVX2-NEXT: vpshuflw {{.*#+}} ymm5 = ymm1[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]492; AVX2-NEXT: vpshufhw {{.*#+}} ymm5 = ymm5[0,1,2,3,4,6,6,7,8,9,10,11,12,14,14,15]493; AVX2-NEXT: vpshuflw {{.*#+}} ymm6 = ymm0[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]494; AVX2-NEXT: vpshufhw {{.*#+}} ymm6 = ymm6[0,1,2,3,4,6,6,7,8,9,10,11,12,14,14,15]495; AVX2-NEXT: vshufps {{.*#+}} ymm5 = ymm6[0,2],ymm5[0,2],ymm6[4,6],ymm5[4,6]496; AVX2-NEXT: vpermpd {{.*#+}} ymm5 = ymm5[0,2,1,3]497; AVX2-NEXT: vmovdqa {{.*#+}} ymm6 = [2,3,6,7,10,11,14,15,2,3,6,7,10,11,14,15,18,19,22,23,26,27,30,31,18,19,22,23,26,27,30,31]498; AVX2-NEXT: vpshufb %ymm6, %ymm3, %ymm3499; AVX2-NEXT: vpshufb %ymm6, %ymm2, %ymm2500; AVX2-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1],ymm3[2,3],ymm2[4,5],ymm3[6,7]501; AVX2-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,2,1,3]502; AVX2-NEXT: vpshufb %ymm6, %ymm1, %ymm1503; AVX2-NEXT: vpshufb %ymm6, %ymm0, %ymm0504; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7]505; AVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]506; AVX2-NEXT: vmovaps %ymm5, (%rsi)507; AVX2-NEXT: vmovaps %ymm4, 32(%rsi)508; AVX2-NEXT: vmovdqa %ymm0, (%rdx)509; AVX2-NEXT: vmovdqa %ymm2, 32(%rdx)510; AVX2-NEXT: vzeroupper511; AVX2-NEXT: retq512;513; AVX2-FP-LABEL: load_i16_stride2_vf32:514; AVX2-FP: # %bb.0:515; AVX2-FP-NEXT: vmovdqa (%rdi), %ymm0516; AVX2-FP-NEXT: vmovdqa 32(%rdi), %ymm1517; AVX2-FP-NEXT: vmovdqa 64(%rdi), %ymm2518; AVX2-FP-NEXT: vmovdqa 96(%rdi), %ymm3519; AVX2-FP-NEXT: vpmovsxdq {{.*#+}} ymm4 = [84148480,218892552,353636624,488380696]520; AVX2-FP-NEXT: vpshufb %ymm4, %ymm3, %ymm5521; AVX2-FP-NEXT: vpshufb %ymm4, %ymm2, %ymm6522; AVX2-FP-NEXT: vshufps {{.*#+}} ymm5 = ymm6[0,2],ymm5[0,2],ymm6[4,6],ymm5[4,6]523; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm5 = ymm5[0,2,1,3]524; AVX2-FP-NEXT: vpshufb %ymm4, %ymm1, %ymm6525; AVX2-FP-NEXT: vpshufb %ymm4, %ymm0, %ymm4526; AVX2-FP-NEXT: vshufps {{.*#+}} ymm4 = ymm4[0,2],ymm6[0,2],ymm4[4,6],ymm6[4,6]527; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm4 = ymm4[0,2,1,3]528; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm6 = [2,3,6,7,10,11,14,15,2,3,6,7,10,11,14,15,18,19,22,23,26,27,30,31,18,19,22,23,26,27,30,31]529; AVX2-FP-NEXT: vpshufb %ymm6, %ymm3, %ymm3530; AVX2-FP-NEXT: vpshufb %ymm6, %ymm2, %ymm2531; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1],ymm3[2,3],ymm2[4,5],ymm3[6,7]532; AVX2-FP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,2,1,3]533; AVX2-FP-NEXT: vpshufb %ymm6, %ymm1, %ymm1534; AVX2-FP-NEXT: vpshufb %ymm6, %ymm0, %ymm0535; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7]536; AVX2-FP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]537; AVX2-FP-NEXT: vmovaps %ymm4, (%rsi)538; AVX2-FP-NEXT: vmovaps %ymm5, 32(%rsi)539; AVX2-FP-NEXT: vmovdqa %ymm0, (%rdx)540; AVX2-FP-NEXT: vmovdqa %ymm2, 32(%rdx)541; AVX2-FP-NEXT: vzeroupper542; AVX2-FP-NEXT: retq543;544; AVX2-FCP-LABEL: load_i16_stride2_vf32:545; AVX2-FCP: # %bb.0:546; AVX2-FCP-NEXT: vmovdqa (%rdi), %ymm0547; AVX2-FCP-NEXT: vmovdqa 32(%rdi), %ymm1548; AVX2-FCP-NEXT: vmovdqa 64(%rdi), %ymm2549; AVX2-FCP-NEXT: vmovdqa 96(%rdi), %ymm3550; AVX2-FCP-NEXT: vpmovsxdq {{.*#+}} ymm4 = [84148480,218892552,353636624,488380696]551; AVX2-FCP-NEXT: vpshufb %ymm4, %ymm3, %ymm5552; AVX2-FCP-NEXT: vpshufb %ymm4, %ymm2, %ymm6553; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm5 = ymm6[0,2],ymm5[0,2],ymm6[4,6],ymm5[4,6]554; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm5 = ymm5[0,2,1,3]555; AVX2-FCP-NEXT: vpshufb %ymm4, %ymm1, %ymm6556; AVX2-FCP-NEXT: vpshufb %ymm4, %ymm0, %ymm4557; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm4 = ymm4[0,2],ymm6[0,2],ymm4[4,6],ymm6[4,6]558; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm4 = ymm4[0,2,1,3]559; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm6 = [2,3,6,7,10,11,14,15,2,3,6,7,10,11,14,15,18,19,22,23,26,27,30,31,18,19,22,23,26,27,30,31]560; AVX2-FCP-NEXT: vpshufb %ymm6, %ymm3, %ymm3561; AVX2-FCP-NEXT: vpshufb %ymm6, %ymm2, %ymm2562; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm2 = ymm2[0,1],ymm3[2,3],ymm2[4,5],ymm3[6,7]563; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,2,1,3]564; AVX2-FCP-NEXT: vpshufb %ymm6, %ymm1, %ymm1565; AVX2-FCP-NEXT: vpshufb %ymm6, %ymm0, %ymm0566; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7]567; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]568; AVX2-FCP-NEXT: vmovaps %ymm4, (%rsi)569; AVX2-FCP-NEXT: vmovaps %ymm5, 32(%rsi)570; AVX2-FCP-NEXT: vmovdqa %ymm0, (%rdx)571; AVX2-FCP-NEXT: vmovdqa %ymm2, 32(%rdx)572; AVX2-FCP-NEXT: vzeroupper573; AVX2-FCP-NEXT: retq574;575; AVX512-VL-LABEL: load_i16_stride2_vf32:576; AVX512-VL: # %bb.0:577; AVX512-VL-NEXT: vmovdqa64 (%rdi), %zmm0578; AVX512-VL-NEXT: vmovdqa64 64(%rdi), %zmm1579; AVX512-VL-NEXT: vpsrld $16, %zmm0, %zmm2580; AVX512-VL-NEXT: vpsrld $16, %zmm1, %zmm3581; AVX512-VL-NEXT: vpmovdw %zmm1, 32(%rsi)582; AVX512-VL-NEXT: vpmovdw %zmm0, (%rsi)583; AVX512-VL-NEXT: vpmovdw %zmm3, 32(%rdx)584; AVX512-VL-NEXT: vpmovdw %zmm2, (%rdx)585; AVX512-VL-NEXT: vzeroupper586; AVX512-VL-NEXT: retq587;588; AVX512-FCP-LABEL: load_i16_stride2_vf32:589; AVX512-FCP: # %bb.0:590; AVX512-FCP-NEXT: vmovdqa64 (%rdi), %zmm0591; AVX512-FCP-NEXT: vmovdqa64 64(%rdi), %zmm1592; AVX512-FCP-NEXT: vpsrld $16, %zmm0, %zmm2593; AVX512-FCP-NEXT: vpsrld $16, %zmm1, %zmm3594; AVX512-FCP-NEXT: vpmovdw %zmm1, 32(%rsi)595; AVX512-FCP-NEXT: vpmovdw %zmm0, (%rsi)596; AVX512-FCP-NEXT: vpmovdw %zmm3, 32(%rdx)597; AVX512-FCP-NEXT: vpmovdw %zmm2, (%rdx)598; AVX512-FCP-NEXT: vzeroupper599; AVX512-FCP-NEXT: retq600;601; AVX512DQ-LABEL: load_i16_stride2_vf32:602; AVX512DQ: # %bb.0:603; AVX512DQ-NEXT: vmovdqa64 (%rdi), %zmm0604; AVX512DQ-NEXT: vmovdqa64 64(%rdi), %zmm1605; AVX512DQ-NEXT: vpsrld $16, %zmm0, %zmm2606; AVX512DQ-NEXT: vpsrld $16, %zmm1, %zmm3607; AVX512DQ-NEXT: vpmovdw %zmm1, 32(%rsi)608; AVX512DQ-NEXT: vpmovdw %zmm0, (%rsi)609; AVX512DQ-NEXT: vpmovdw %zmm3, 32(%rdx)610; AVX512DQ-NEXT: vpmovdw %zmm2, (%rdx)611; AVX512DQ-NEXT: vzeroupper612; AVX512DQ-NEXT: retq613;614; AVX512DQ-FCP-LABEL: load_i16_stride2_vf32:615; AVX512DQ-FCP: # %bb.0:616; AVX512DQ-FCP-NEXT: vmovdqa64 (%rdi), %zmm0617; AVX512DQ-FCP-NEXT: vmovdqa64 64(%rdi), %zmm1618; AVX512DQ-FCP-NEXT: vpsrld $16, %zmm0, %zmm2619; AVX512DQ-FCP-NEXT: vpsrld $16, %zmm1, %zmm3620; AVX512DQ-FCP-NEXT: vpmovdw %zmm1, 32(%rsi)621; AVX512DQ-FCP-NEXT: vpmovdw %zmm0, (%rsi)622; AVX512DQ-FCP-NEXT: vpmovdw %zmm3, 32(%rdx)623; AVX512DQ-FCP-NEXT: vpmovdw %zmm2, (%rdx)624; AVX512DQ-FCP-NEXT: vzeroupper625; AVX512DQ-FCP-NEXT: retq626;627; AVX512BW-LABEL: load_i16_stride2_vf32:628; AVX512BW: # %bb.0:629; AVX512BW-NEXT: vmovdqa64 (%rdi), %zmm0630; AVX512BW-NEXT: vmovdqa64 64(%rdi), %zmm1631; AVX512BW-NEXT: vpmovsxbw {{.*#+}} zmm2 = [0,2,4,6,8,10,12,14,16,18,20,22,24,26,28,30,32,34,36,38,40,42,44,46,48,50,52,54,56,58,60,62]632; AVX512BW-NEXT: vpermi2w %zmm1, %zmm0, %zmm2633; AVX512BW-NEXT: vpmovsxbw {{.*#+}} zmm3 = [1,3,5,7,9,11,13,15,17,19,21,23,25,27,29,31,33,35,37,39,41,43,45,47,49,51,53,55,57,59,61,63]634; AVX512BW-NEXT: vpermi2w %zmm1, %zmm0, %zmm3635; AVX512BW-NEXT: vmovdqa64 %zmm2, (%rsi)636; AVX512BW-NEXT: vmovdqa64 %zmm3, (%rdx)637; AVX512BW-NEXT: vzeroupper638; AVX512BW-NEXT: retq639;640; AVX512BW-FCP-LABEL: load_i16_stride2_vf32:641; AVX512BW-FCP: # %bb.0:642; AVX512BW-FCP-NEXT: vmovdqa64 (%rdi), %zmm0643; AVX512BW-FCP-NEXT: vmovdqa64 64(%rdi), %zmm1644; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm2 = [0,2,4,6,8,10,12,14,16,18,20,22,24,26,28,30,32,34,36,38,40,42,44,46,48,50,52,54,56,58,60,62]645; AVX512BW-FCP-NEXT: vpermi2w %zmm1, %zmm0, %zmm2646; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm3 = [1,3,5,7,9,11,13,15,17,19,21,23,25,27,29,31,33,35,37,39,41,43,45,47,49,51,53,55,57,59,61,63]647; AVX512BW-FCP-NEXT: vpermi2w %zmm1, %zmm0, %zmm3648; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, (%rsi)649; AVX512BW-FCP-NEXT: vmovdqa64 %zmm3, (%rdx)650; AVX512BW-FCP-NEXT: vzeroupper651; AVX512BW-FCP-NEXT: retq652;653; AVX512DQ-BW-LABEL: load_i16_stride2_vf32:654; AVX512DQ-BW: # %bb.0:655; AVX512DQ-BW-NEXT: vmovdqa64 (%rdi), %zmm0656; AVX512DQ-BW-NEXT: vmovdqa64 64(%rdi), %zmm1657; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} zmm2 = [0,2,4,6,8,10,12,14,16,18,20,22,24,26,28,30,32,34,36,38,40,42,44,46,48,50,52,54,56,58,60,62]658; AVX512DQ-BW-NEXT: vpermi2w %zmm1, %zmm0, %zmm2659; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} zmm3 = [1,3,5,7,9,11,13,15,17,19,21,23,25,27,29,31,33,35,37,39,41,43,45,47,49,51,53,55,57,59,61,63]660; AVX512DQ-BW-NEXT: vpermi2w %zmm1, %zmm0, %zmm3661; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, (%rsi)662; AVX512DQ-BW-NEXT: vmovdqa64 %zmm3, (%rdx)663; AVX512DQ-BW-NEXT: vzeroupper664; AVX512DQ-BW-NEXT: retq665;666; AVX512DQ-BW-FCP-LABEL: load_i16_stride2_vf32:667; AVX512DQ-BW-FCP: # %bb.0:668; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rdi), %zmm0669; AVX512DQ-BW-FCP-NEXT: vmovdqa64 64(%rdi), %zmm1670; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm2 = [0,2,4,6,8,10,12,14,16,18,20,22,24,26,28,30,32,34,36,38,40,42,44,46,48,50,52,54,56,58,60,62]671; AVX512DQ-BW-FCP-NEXT: vpermi2w %zmm1, %zmm0, %zmm2672; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm3 = [1,3,5,7,9,11,13,15,17,19,21,23,25,27,29,31,33,35,37,39,41,43,45,47,49,51,53,55,57,59,61,63]673; AVX512DQ-BW-FCP-NEXT: vpermi2w %zmm1, %zmm0, %zmm3674; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, (%rsi)675; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm3, (%rdx)676; AVX512DQ-BW-FCP-NEXT: vzeroupper677; AVX512DQ-BW-FCP-NEXT: retq678 %wide.vec = load <64 x i16>, ptr %in.vec, align 64679 %strided.vec0 = shufflevector <64 x i16> %wide.vec, <64 x i16> poison, <32 x i32> <i32 0, i32 2, i32 4, i32 6, i32 8, i32 10, i32 12, i32 14, i32 16, i32 18, i32 20, i32 22, i32 24, i32 26, i32 28, i32 30, i32 32, i32 34, i32 36, i32 38, i32 40, i32 42, i32 44, i32 46, i32 48, i32 50, i32 52, i32 54, i32 56, i32 58, i32 60, i32 62>680 %strided.vec1 = shufflevector <64 x i16> %wide.vec, <64 x i16> poison, <32 x i32> <i32 1, i32 3, i32 5, i32 7, i32 9, i32 11, i32 13, i32 15, i32 17, i32 19, i32 21, i32 23, i32 25, i32 27, i32 29, i32 31, i32 33, i32 35, i32 37, i32 39, i32 41, i32 43, i32 45, i32 47, i32 49, i32 51, i32 53, i32 55, i32 57, i32 59, i32 61, i32 63>681 store <32 x i16> %strided.vec0, ptr %out.vec0, align 64682 store <32 x i16> %strided.vec1, ptr %out.vec1, align 64683 ret void684}685 686define void @load_i16_stride2_vf64(ptr %in.vec, ptr %out.vec0, ptr %out.vec1) nounwind {687; SSE-LABEL: load_i16_stride2_vf64:688; SSE: # %bb.0:689; SSE-NEXT: subq $40, %rsp690; SSE-NEXT: movdqa 160(%rdi), %xmm14691; SSE-NEXT: movdqa %xmm14, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill692; SSE-NEXT: movdqa 176(%rdi), %xmm2693; SSE-NEXT: movdqa %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill694; SSE-NEXT: movdqa 64(%rdi), %xmm11695; SSE-NEXT: movdqa %xmm11, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill696; SSE-NEXT: movdqa 80(%rdi), %xmm1697; SSE-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill698; SSE-NEXT: movdqa 96(%rdi), %xmm9699; SSE-NEXT: movdqa %xmm9, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill700; SSE-NEXT: movdqa 112(%rdi), %xmm4701; SSE-NEXT: movdqa %xmm4, (%rsp) # 16-byte Spill702; SSE-NEXT: movdqa (%rdi), %xmm10703; SSE-NEXT: movdqa %xmm10, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill704; SSE-NEXT: movdqa 16(%rdi), %xmm7705; SSE-NEXT: movdqa %xmm7, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill706; SSE-NEXT: movdqa 32(%rdi), %xmm13707; SSE-NEXT: movdqa %xmm13, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill708; SSE-NEXT: movdqa 48(%rdi), %xmm0709; SSE-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill710; SSE-NEXT: pslld $16, %xmm0711; SSE-NEXT: psrad $16, %xmm0712; SSE-NEXT: pslld $16, %xmm13713; SSE-NEXT: psrad $16, %xmm13714; SSE-NEXT: packssdw %xmm0, %xmm13715; SSE-NEXT: movdqa %xmm7, %xmm0716; SSE-NEXT: pslld $16, %xmm0717; SSE-NEXT: psrad $16, %xmm0718; SSE-NEXT: pslld $16, %xmm10719; SSE-NEXT: psrad $16, %xmm10720; SSE-NEXT: packssdw %xmm0, %xmm10721; SSE-NEXT: movdqa %xmm4, %xmm0722; SSE-NEXT: pslld $16, %xmm0723; SSE-NEXT: psrad $16, %xmm0724; SSE-NEXT: pslld $16, %xmm9725; SSE-NEXT: psrad $16, %xmm9726; SSE-NEXT: packssdw %xmm0, %xmm9727; SSE-NEXT: movdqa %xmm1, %xmm0728; SSE-NEXT: pslld $16, %xmm0729; SSE-NEXT: psrad $16, %xmm0730; SSE-NEXT: pslld $16, %xmm11731; SSE-NEXT: psrad $16, %xmm11732; SSE-NEXT: packssdw %xmm0, %xmm11733; SSE-NEXT: movdqa %xmm2, %xmm0734; SSE-NEXT: pslld $16, %xmm0735; SSE-NEXT: psrad $16, %xmm0736; SSE-NEXT: pslld $16, %xmm14737; SSE-NEXT: psrad $16, %xmm14738; SSE-NEXT: packssdw %xmm0, %xmm14739; SSE-NEXT: movdqa 144(%rdi), %xmm7740; SSE-NEXT: movdqa %xmm7, %xmm0741; SSE-NEXT: pslld $16, %xmm0742; SSE-NEXT: psrad $16, %xmm0743; SSE-NEXT: movdqa 128(%rdi), %xmm8744; SSE-NEXT: movdqa %xmm8, %xmm15745; SSE-NEXT: pslld $16, %xmm15746; SSE-NEXT: psrad $16, %xmm15747; SSE-NEXT: packssdw %xmm0, %xmm15748; SSE-NEXT: movdqa 240(%rdi), %xmm12749; SSE-NEXT: movdqa %xmm12, %xmm1750; SSE-NEXT: pslld $16, %xmm1751; SSE-NEXT: psrad $16, %xmm1752; SSE-NEXT: movdqa 224(%rdi), %xmm5753; SSE-NEXT: movdqa %xmm5, %xmm3754; SSE-NEXT: pslld $16, %xmm3755; SSE-NEXT: psrad $16, %xmm3756; SSE-NEXT: packssdw %xmm1, %xmm3757; SSE-NEXT: movdqa 208(%rdi), %xmm6758; SSE-NEXT: movdqa %xmm6, %xmm4759; SSE-NEXT: pslld $16, %xmm4760; SSE-NEXT: psrad $16, %xmm4761; SSE-NEXT: movdqa 192(%rdi), %xmm2762; SSE-NEXT: movdqa %xmm2, %xmm1763; SSE-NEXT: pslld $16, %xmm1764; SSE-NEXT: psrad $16, %xmm1765; SSE-NEXT: packssdw %xmm4, %xmm1766; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload767; SSE-NEXT: psrad $16, %xmm0768; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload769; SSE-NEXT: psrad $16, %xmm4770; SSE-NEXT: packssdw %xmm0, %xmm4771; SSE-NEXT: movdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill772; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload773; SSE-NEXT: psrad $16, %xmm0774; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload775; SSE-NEXT: psrad $16, %xmm4776; SSE-NEXT: packssdw %xmm0, %xmm4777; SSE-NEXT: movdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill778; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload779; SSE-NEXT: psrad $16, %xmm0780; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload781; SSE-NEXT: psrad $16, %xmm4782; SSE-NEXT: packssdw %xmm0, %xmm4783; SSE-NEXT: movdqa %xmm4, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill784; SSE-NEXT: movdqa (%rsp), %xmm0 # 16-byte Reload785; SSE-NEXT: psrad $16, %xmm0786; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm4 # 16-byte Reload787; SSE-NEXT: psrad $16, %xmm4788; SSE-NEXT: packssdw %xmm0, %xmm4789; SSE-NEXT: psrad $16, %xmm7790; SSE-NEXT: psrad $16, %xmm8791; SSE-NEXT: packssdw %xmm7, %xmm8792; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload793; SSE-NEXT: psrad $16, %xmm0794; SSE-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload795; SSE-NEXT: psrad $16, %xmm7796; SSE-NEXT: packssdw %xmm0, %xmm7797; SSE-NEXT: psrad $16, %xmm6798; SSE-NEXT: psrad $16, %xmm2799; SSE-NEXT: packssdw %xmm6, %xmm2800; SSE-NEXT: psrad $16, %xmm12801; SSE-NEXT: psrad $16, %xmm5802; SSE-NEXT: packssdw %xmm12, %xmm5803; SSE-NEXT: movdqa %xmm1, 96(%rsi)804; SSE-NEXT: movdqa %xmm3, 112(%rsi)805; SSE-NEXT: movdqa %xmm15, 64(%rsi)806; SSE-NEXT: movdqa %xmm14, 80(%rsi)807; SSE-NEXT: movdqa %xmm11, 32(%rsi)808; SSE-NEXT: movdqa %xmm9, 48(%rsi)809; SSE-NEXT: movdqa %xmm10, (%rsi)810; SSE-NEXT: movdqa %xmm13, 16(%rsi)811; SSE-NEXT: movdqa %xmm5, 112(%rdx)812; SSE-NEXT: movdqa %xmm2, 96(%rdx)813; SSE-NEXT: movdqa %xmm7, 80(%rdx)814; SSE-NEXT: movdqa %xmm8, 64(%rdx)815; SSE-NEXT: movdqa %xmm4, 48(%rdx)816; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload817; SSE-NEXT: movaps %xmm0, 32(%rdx)818; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload819; SSE-NEXT: movaps %xmm0, 16(%rdx)820; SSE-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload821; SSE-NEXT: movaps %xmm0, (%rdx)822; SSE-NEXT: addq $40, %rsp823; SSE-NEXT: retq824;825; AVX-LABEL: load_i16_stride2_vf64:826; AVX: # %bb.0:827; AVX-NEXT: subq $24, %rsp828; AVX-NEXT: vpxor %xmm11, %xmm11, %xmm11829; AVX-NEXT: vmovdqa 176(%rdi), %xmm0830; AVX-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill831; AVX-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0],xmm11[1],xmm0[2],xmm11[3],xmm0[4],xmm11[5],xmm0[6],xmm11[7]832; AVX-NEXT: vmovdqa 160(%rdi), %xmm9833; AVX-NEXT: vpblendw {{.*#+}} xmm1 = xmm9[0],xmm11[1],xmm9[2],xmm11[3],xmm9[4],xmm11[5],xmm9[6],xmm11[7]834; AVX-NEXT: vpackusdw %xmm0, %xmm1, %xmm0835; AVX-NEXT: vmovdqa %xmm0, (%rsp) # 16-byte Spill836; AVX-NEXT: vmovdqa (%rdi), %xmm6837; AVX-NEXT: vmovdqa %xmm6, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill838; AVX-NEXT: vmovdqa 16(%rdi), %xmm8839; AVX-NEXT: vmovdqa %xmm8, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill840; AVX-NEXT: vmovdqa 32(%rdi), %xmm7841; AVX-NEXT: vmovdqa 48(%rdi), %xmm10842; AVX-NEXT: vpblendw {{.*#+}} xmm0 = xmm10[0],xmm11[1],xmm10[2],xmm11[3],xmm10[4],xmm11[5],xmm10[6],xmm11[7]843; AVX-NEXT: vpblendw {{.*#+}} xmm1 = xmm7[0],xmm11[1],xmm7[2],xmm11[3],xmm7[4],xmm11[5],xmm7[6],xmm11[7]844; AVX-NEXT: vpackusdw %xmm0, %xmm1, %xmm0845; AVX-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill846; AVX-NEXT: vmovdqa 144(%rdi), %xmm5847; AVX-NEXT: vpblendw {{.*#+}} xmm2 = xmm5[0],xmm11[1],xmm5[2],xmm11[3],xmm5[4],xmm11[5],xmm5[6],xmm11[7]848; AVX-NEXT: vmovdqa 128(%rdi), %xmm4849; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm4[0],xmm11[1],xmm4[2],xmm11[3],xmm4[4],xmm11[5],xmm4[6],xmm11[7]850; AVX-NEXT: vpackusdw %xmm2, %xmm3, %xmm0851; AVX-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill852; AVX-NEXT: vpblendw {{.*#+}} xmm2 = xmm8[0],xmm11[1],xmm8[2],xmm11[3],xmm8[4],xmm11[5],xmm8[6],xmm11[7]853; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm6[0],xmm11[1],xmm6[2],xmm11[3],xmm6[4],xmm11[5],xmm6[6],xmm11[7]854; AVX-NEXT: vpackusdw %xmm2, %xmm3, %xmm0855; AVX-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill856; AVX-NEXT: vmovdqa 112(%rdi), %xmm12857; AVX-NEXT: vpblendw {{.*#+}} xmm2 = xmm12[0],xmm11[1],xmm12[2],xmm11[3],xmm12[4],xmm11[5],xmm12[6],xmm11[7]858; AVX-NEXT: vmovdqa 96(%rdi), %xmm13859; AVX-NEXT: vpblendw {{.*#+}} xmm8 = xmm13[0],xmm11[1],xmm13[2],xmm11[3],xmm13[4],xmm11[5],xmm13[6],xmm11[7]860; AVX-NEXT: vpackusdw %xmm2, %xmm8, %xmm0861; AVX-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill862; AVX-NEXT: vmovdqa 80(%rdi), %xmm15863; AVX-NEXT: vpblendw {{.*#+}} xmm14 = xmm15[0],xmm11[1],xmm15[2],xmm11[3],xmm15[4],xmm11[5],xmm15[6],xmm11[7]864; AVX-NEXT: vmovdqa 64(%rdi), %xmm6865; AVX-NEXT: vpblendw {{.*#+}} xmm3 = xmm6[0],xmm11[1],xmm6[2],xmm11[3],xmm6[4],xmm11[5],xmm6[6],xmm11[7]866; AVX-NEXT: vpackusdw %xmm14, %xmm3, %xmm0867; AVX-NEXT: vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill868; AVX-NEXT: vmovdqa 240(%rdi), %xmm3869; AVX-NEXT: vpblendw {{.*#+}} xmm1 = xmm3[0],xmm11[1],xmm3[2],xmm11[3],xmm3[4],xmm11[5],xmm3[6],xmm11[7]870; AVX-NEXT: vmovdqa 224(%rdi), %xmm0871; AVX-NEXT: vpblendw {{.*#+}} xmm8 = xmm0[0],xmm11[1],xmm0[2],xmm11[3],xmm0[4],xmm11[5],xmm0[6],xmm11[7]872; AVX-NEXT: vpackusdw %xmm1, %xmm8, %xmm14873; AVX-NEXT: vmovdqa 208(%rdi), %xmm8874; AVX-NEXT: vpblendw {{.*#+}} xmm1 = xmm8[0],xmm11[1],xmm8[2],xmm11[3],xmm8[4],xmm11[5],xmm8[6],xmm11[7]875; AVX-NEXT: vmovdqa 192(%rdi), %xmm2876; AVX-NEXT: vpblendw {{.*#+}} xmm11 = xmm2[0],xmm11[1],xmm2[2],xmm11[3],xmm2[4],xmm11[5],xmm2[6],xmm11[7]877; AVX-NEXT: vpackusdw %xmm1, %xmm11, %xmm1878; AVX-NEXT: vpsrld $16, %xmm3, %xmm3879; AVX-NEXT: vpsrld $16, %xmm0, %xmm0880; AVX-NEXT: vpackusdw %xmm3, %xmm0, %xmm11881; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload882; AVX-NEXT: vpsrld $16, %xmm0, %xmm0883; AVX-NEXT: vpsrld $16, %xmm9, %xmm3884; AVX-NEXT: vpackusdw %xmm0, %xmm3, %xmm0885; AVX-NEXT: vpsrld $16, %xmm5, %xmm3886; AVX-NEXT: vpsrld $16, %xmm4, %xmm4887; AVX-NEXT: vpackusdw %xmm3, %xmm4, %xmm3888; AVX-NEXT: vpsrld $16, %xmm10, %xmm4889; AVX-NEXT: vpsrld $16, %xmm7, %xmm5890; AVX-NEXT: vpackusdw %xmm4, %xmm5, %xmm4891; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm5 # 16-byte Reload892; AVX-NEXT: vpsrld $16, %xmm5, %xmm5893; AVX-NEXT: vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm7 # 16-byte Reload894; AVX-NEXT: vpsrld $16, %xmm7, %xmm7895; AVX-NEXT: vpackusdw %xmm5, %xmm7, %xmm5896; AVX-NEXT: vpsrld $16, %xmm8, %xmm7897; AVX-NEXT: vpsrld $16, %xmm2, %xmm2898; AVX-NEXT: vpackusdw %xmm7, %xmm2, %xmm2899; AVX-NEXT: vpsrld $16, %xmm12, %xmm7900; AVX-NEXT: vpsrld $16, %xmm13, %xmm8901; AVX-NEXT: vpackusdw %xmm7, %xmm8, %xmm7902; AVX-NEXT: vpsrld $16, %xmm15, %xmm8903; AVX-NEXT: vpsrld $16, %xmm6, %xmm6904; AVX-NEXT: vpackusdw %xmm8, %xmm6, %xmm6905; AVX-NEXT: vmovdqa %xmm1, 96(%rsi)906; AVX-NEXT: vmovdqa %xmm14, 112(%rsi)907; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload908; AVX-NEXT: vmovaps %xmm1, 32(%rsi)909; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload910; AVX-NEXT: vmovaps %xmm1, 48(%rsi)911; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload912; AVX-NEXT: vmovaps %xmm1, (%rsi)913; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload914; AVX-NEXT: vmovaps %xmm1, 64(%rsi)915; AVX-NEXT: vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload916; AVX-NEXT: vmovaps %xmm1, 16(%rsi)917; AVX-NEXT: vmovaps (%rsp), %xmm1 # 16-byte Reload918; AVX-NEXT: vmovaps %xmm1, 80(%rsi)919; AVX-NEXT: vmovdqa %xmm6, 32(%rdx)920; AVX-NEXT: vmovdqa %xmm7, 48(%rdx)921; AVX-NEXT: vmovdqa %xmm2, 96(%rdx)922; AVX-NEXT: vmovdqa %xmm5, (%rdx)923; AVX-NEXT: vmovdqa %xmm4, 16(%rdx)924; AVX-NEXT: vmovdqa %xmm3, 64(%rdx)925; AVX-NEXT: vmovdqa %xmm0, 80(%rdx)926; AVX-NEXT: vmovdqa %xmm11, 112(%rdx)927; AVX-NEXT: addq $24, %rsp928; AVX-NEXT: retq929;930; AVX2-LABEL: load_i16_stride2_vf64:931; AVX2: # %bb.0:932; AVX2-NEXT: vmovdqa 128(%rdi), %ymm0933; AVX2-NEXT: vmovdqa 160(%rdi), %ymm1934; AVX2-NEXT: vmovdqa 192(%rdi), %ymm4935; AVX2-NEXT: vmovdqa 224(%rdi), %ymm6936; AVX2-NEXT: vmovdqa (%rdi), %ymm3937; AVX2-NEXT: vmovdqa 32(%rdi), %ymm5938; AVX2-NEXT: vmovdqa 64(%rdi), %ymm7939; AVX2-NEXT: vmovdqa 96(%rdi), %ymm8940; AVX2-NEXT: vpshuflw {{.*#+}} ymm2 = ymm8[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]941; AVX2-NEXT: vpshufhw {{.*#+}} ymm2 = ymm2[0,1,2,3,4,6,6,7,8,9,10,11,12,14,14,15]942; AVX2-NEXT: vpshuflw {{.*#+}} ymm9 = ymm7[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]943; AVX2-NEXT: vpshufhw {{.*#+}} ymm9 = ymm9[0,1,2,3,4,6,6,7,8,9,10,11,12,14,14,15]944; AVX2-NEXT: vshufps {{.*#+}} ymm2 = ymm9[0,2],ymm2[0,2],ymm9[4,6],ymm2[4,6]945; AVX2-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,2,1,3]946; AVX2-NEXT: vpshuflw {{.*#+}} ymm9 = ymm6[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]947; AVX2-NEXT: vpshufhw {{.*#+}} ymm9 = ymm9[0,1,2,3,4,6,6,7,8,9,10,11,12,14,14,15]948; AVX2-NEXT: vpshuflw {{.*#+}} ymm10 = ymm4[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]949; AVX2-NEXT: vpshufhw {{.*#+}} ymm10 = ymm10[0,1,2,3,4,6,6,7,8,9,10,11,12,14,14,15]950; AVX2-NEXT: vshufps {{.*#+}} ymm9 = ymm10[0,2],ymm9[0,2],ymm10[4,6],ymm9[4,6]951; AVX2-NEXT: vpermpd {{.*#+}} ymm9 = ymm9[0,2,1,3]952; AVX2-NEXT: vpshuflw {{.*#+}} ymm10 = ymm5[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]953; AVX2-NEXT: vpshufhw {{.*#+}} ymm10 = ymm10[0,1,2,3,4,6,6,7,8,9,10,11,12,14,14,15]954; AVX2-NEXT: vpshuflw {{.*#+}} ymm11 = ymm3[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]955; AVX2-NEXT: vpshufhw {{.*#+}} ymm11 = ymm11[0,1,2,3,4,6,6,7,8,9,10,11,12,14,14,15]956; AVX2-NEXT: vshufps {{.*#+}} ymm10 = ymm11[0,2],ymm10[0,2],ymm11[4,6],ymm10[4,6]957; AVX2-NEXT: vpermpd {{.*#+}} ymm10 = ymm10[0,2,1,3]958; AVX2-NEXT: vpshuflw {{.*#+}} ymm11 = ymm1[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]959; AVX2-NEXT: vpshufhw {{.*#+}} ymm11 = ymm11[0,1,2,3,4,6,6,7,8,9,10,11,12,14,14,15]960; AVX2-NEXT: vpshuflw {{.*#+}} ymm12 = ymm0[0,2,2,3,4,5,6,7,8,10,10,11,12,13,14,15]961; AVX2-NEXT: vpshufhw {{.*#+}} ymm12 = ymm12[0,1,2,3,4,6,6,7,8,9,10,11,12,14,14,15]962; AVX2-NEXT: vshufps {{.*#+}} ymm11 = ymm12[0,2],ymm11[0,2],ymm12[4,6],ymm11[4,6]963; AVX2-NEXT: vpermpd {{.*#+}} ymm11 = ymm11[0,2,1,3]964; AVX2-NEXT: vmovdqa {{.*#+}} ymm12 = [2,3,6,7,10,11,14,15,2,3,6,7,10,11,14,15,18,19,22,23,26,27,30,31,18,19,22,23,26,27,30,31]965; AVX2-NEXT: vpshufb %ymm12, %ymm8, %ymm8966; AVX2-NEXT: vpshufb %ymm12, %ymm7, %ymm7967; AVX2-NEXT: vpblendd {{.*#+}} ymm7 = ymm7[0,1],ymm8[2,3],ymm7[4,5],ymm8[6,7]968; AVX2-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,2,1,3]969; AVX2-NEXT: vpshufb %ymm12, %ymm6, %ymm6970; AVX2-NEXT: vpshufb %ymm12, %ymm4, %ymm4971; AVX2-NEXT: vpblendd {{.*#+}} ymm4 = ymm4[0,1],ymm6[2,3],ymm4[4,5],ymm6[6,7]972; AVX2-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,2,1,3]973; AVX2-NEXT: vpshufb %ymm12, %ymm5, %ymm5974; AVX2-NEXT: vpshufb %ymm12, %ymm3, %ymm3975; AVX2-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0,1],ymm5[2,3],ymm3[4,5],ymm5[6,7]976; AVX2-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,2,1,3]977; AVX2-NEXT: vpshufb %ymm12, %ymm1, %ymm1978; AVX2-NEXT: vpshufb %ymm12, %ymm0, %ymm0979; AVX2-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7]980; AVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]981; AVX2-NEXT: vmovaps %ymm11, 64(%rsi)982; AVX2-NEXT: vmovaps %ymm10, (%rsi)983; AVX2-NEXT: vmovaps %ymm9, 96(%rsi)984; AVX2-NEXT: vmovaps %ymm2, 32(%rsi)985; AVX2-NEXT: vmovdqa %ymm0, 64(%rdx)986; AVX2-NEXT: vmovdqa %ymm3, (%rdx)987; AVX2-NEXT: vmovdqa %ymm4, 96(%rdx)988; AVX2-NEXT: vmovdqa %ymm7, 32(%rdx)989; AVX2-NEXT: vzeroupper990; AVX2-NEXT: retq991;992; AVX2-FP-LABEL: load_i16_stride2_vf64:993; AVX2-FP: # %bb.0:994; AVX2-FP-NEXT: vmovdqa 128(%rdi), %ymm0995; AVX2-FP-NEXT: vmovdqa 160(%rdi), %ymm1996; AVX2-FP-NEXT: vmovdqa 192(%rdi), %ymm3997; AVX2-FP-NEXT: vmovdqa 224(%rdi), %ymm4998; AVX2-FP-NEXT: vmovdqa (%rdi), %ymm5999; AVX2-FP-NEXT: vmovdqa 32(%rdi), %ymm61000; AVX2-FP-NEXT: vmovdqa 64(%rdi), %ymm71001; AVX2-FP-NEXT: vmovdqa 96(%rdi), %ymm81002; AVX2-FP-NEXT: vpmovsxdq {{.*#+}} ymm9 = [84148480,218892552,353636624,488380696]1003; AVX2-FP-NEXT: vpshufb %ymm9, %ymm8, %ymm21004; AVX2-FP-NEXT: vpshufb %ymm9, %ymm7, %ymm101005; AVX2-FP-NEXT: vshufps {{.*#+}} ymm2 = ymm10[0,2],ymm2[0,2],ymm10[4,6],ymm2[4,6]1006; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,2,1,3]1007; AVX2-FP-NEXT: vpshufb %ymm9, %ymm4, %ymm101008; AVX2-FP-NEXT: vpshufb %ymm9, %ymm3, %ymm111009; AVX2-FP-NEXT: vshufps {{.*#+}} ymm10 = ymm11[0,2],ymm10[0,2],ymm11[4,6],ymm10[4,6]1010; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm10 = ymm10[0,2,1,3]1011; AVX2-FP-NEXT: vpshufb %ymm9, %ymm6, %ymm111012; AVX2-FP-NEXT: vpshufb %ymm9, %ymm5, %ymm121013; AVX2-FP-NEXT: vshufps {{.*#+}} ymm11 = ymm12[0,2],ymm11[0,2],ymm12[4,6],ymm11[4,6]1014; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm11 = ymm11[0,2,1,3]1015; AVX2-FP-NEXT: vpshufb %ymm9, %ymm1, %ymm121016; AVX2-FP-NEXT: vpshufb %ymm9, %ymm0, %ymm91017; AVX2-FP-NEXT: vshufps {{.*#+}} ymm9 = ymm9[0,2],ymm12[0,2],ymm9[4,6],ymm12[4,6]1018; AVX2-FP-NEXT: vpermpd {{.*#+}} ymm9 = ymm9[0,2,1,3]1019; AVX2-FP-NEXT: vmovdqa {{.*#+}} ymm12 = [2,3,6,7,10,11,14,15,2,3,6,7,10,11,14,15,18,19,22,23,26,27,30,31,18,19,22,23,26,27,30,31]1020; AVX2-FP-NEXT: vpshufb %ymm12, %ymm8, %ymm81021; AVX2-FP-NEXT: vpshufb %ymm12, %ymm7, %ymm71022; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm7 = ymm7[0,1],ymm8[2,3],ymm7[4,5],ymm8[6,7]1023; AVX2-FP-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,2,1,3]1024; AVX2-FP-NEXT: vpshufb %ymm12, %ymm4, %ymm41025; AVX2-FP-NEXT: vpshufb %ymm12, %ymm3, %ymm31026; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0,1],ymm4[2,3],ymm3[4,5],ymm4[6,7]1027; AVX2-FP-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,2,1,3]1028; AVX2-FP-NEXT: vpshufb %ymm12, %ymm6, %ymm41029; AVX2-FP-NEXT: vpshufb %ymm12, %ymm5, %ymm51030; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm4 = ymm5[0,1],ymm4[2,3],ymm5[4,5],ymm4[6,7]1031; AVX2-FP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,2,1,3]1032; AVX2-FP-NEXT: vpshufb %ymm12, %ymm1, %ymm11033; AVX2-FP-NEXT: vpshufb %ymm12, %ymm0, %ymm01034; AVX2-FP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7]1035; AVX2-FP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]1036; AVX2-FP-NEXT: vmovaps %ymm9, 64(%rsi)1037; AVX2-FP-NEXT: vmovaps %ymm11, (%rsi)1038; AVX2-FP-NEXT: vmovaps %ymm10, 96(%rsi)1039; AVX2-FP-NEXT: vmovaps %ymm2, 32(%rsi)1040; AVX2-FP-NEXT: vmovdqa %ymm0, 64(%rdx)1041; AVX2-FP-NEXT: vmovdqa %ymm4, (%rdx)1042; AVX2-FP-NEXT: vmovdqa %ymm3, 96(%rdx)1043; AVX2-FP-NEXT: vmovdqa %ymm7, 32(%rdx)1044; AVX2-FP-NEXT: vzeroupper1045; AVX2-FP-NEXT: retq1046;1047; AVX2-FCP-LABEL: load_i16_stride2_vf64:1048; AVX2-FCP: # %bb.0:1049; AVX2-FCP-NEXT: vmovdqa 128(%rdi), %ymm01050; AVX2-FCP-NEXT: vmovdqa 160(%rdi), %ymm11051; AVX2-FCP-NEXT: vmovdqa 192(%rdi), %ymm31052; AVX2-FCP-NEXT: vmovdqa 224(%rdi), %ymm41053; AVX2-FCP-NEXT: vmovdqa (%rdi), %ymm51054; AVX2-FCP-NEXT: vmovdqa 32(%rdi), %ymm61055; AVX2-FCP-NEXT: vmovdqa 64(%rdi), %ymm71056; AVX2-FCP-NEXT: vmovdqa 96(%rdi), %ymm81057; AVX2-FCP-NEXT: vpmovsxdq {{.*#+}} ymm9 = [84148480,218892552,353636624,488380696]1058; AVX2-FCP-NEXT: vpshufb %ymm9, %ymm8, %ymm21059; AVX2-FCP-NEXT: vpshufb %ymm9, %ymm7, %ymm101060; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm2 = ymm10[0,2],ymm2[0,2],ymm10[4,6],ymm2[4,6]1061; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,2,1,3]1062; AVX2-FCP-NEXT: vpshufb %ymm9, %ymm4, %ymm101063; AVX2-FCP-NEXT: vpshufb %ymm9, %ymm3, %ymm111064; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm10 = ymm11[0,2],ymm10[0,2],ymm11[4,6],ymm10[4,6]1065; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm10 = ymm10[0,2,1,3]1066; AVX2-FCP-NEXT: vpshufb %ymm9, %ymm6, %ymm111067; AVX2-FCP-NEXT: vpshufb %ymm9, %ymm5, %ymm121068; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm11 = ymm12[0,2],ymm11[0,2],ymm12[4,6],ymm11[4,6]1069; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm11 = ymm11[0,2,1,3]1070; AVX2-FCP-NEXT: vpshufb %ymm9, %ymm1, %ymm121071; AVX2-FCP-NEXT: vpshufb %ymm9, %ymm0, %ymm91072; AVX2-FCP-NEXT: vshufps {{.*#+}} ymm9 = ymm9[0,2],ymm12[0,2],ymm9[4,6],ymm12[4,6]1073; AVX2-FCP-NEXT: vpermpd {{.*#+}} ymm9 = ymm9[0,2,1,3]1074; AVX2-FCP-NEXT: vmovdqa {{.*#+}} ymm12 = [2,3,6,7,10,11,14,15,2,3,6,7,10,11,14,15,18,19,22,23,26,27,30,31,18,19,22,23,26,27,30,31]1075; AVX2-FCP-NEXT: vpshufb %ymm12, %ymm8, %ymm81076; AVX2-FCP-NEXT: vpshufb %ymm12, %ymm7, %ymm71077; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm7 = ymm7[0,1],ymm8[2,3],ymm7[4,5],ymm8[6,7]1078; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm7 = ymm7[0,2,1,3]1079; AVX2-FCP-NEXT: vpshufb %ymm12, %ymm4, %ymm41080; AVX2-FCP-NEXT: vpshufb %ymm12, %ymm3, %ymm31081; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm3 = ymm3[0,1],ymm4[2,3],ymm3[4,5],ymm4[6,7]1082; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm3 = ymm3[0,2,1,3]1083; AVX2-FCP-NEXT: vpshufb %ymm12, %ymm6, %ymm41084; AVX2-FCP-NEXT: vpshufb %ymm12, %ymm5, %ymm51085; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm4 = ymm5[0,1],ymm4[2,3],ymm5[4,5],ymm4[6,7]1086; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm4 = ymm4[0,2,1,3]1087; AVX2-FCP-NEXT: vpshufb %ymm12, %ymm1, %ymm11088; AVX2-FCP-NEXT: vpshufb %ymm12, %ymm0, %ymm01089; AVX2-FCP-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7]1090; AVX2-FCP-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]1091; AVX2-FCP-NEXT: vmovaps %ymm9, 64(%rsi)1092; AVX2-FCP-NEXT: vmovaps %ymm11, (%rsi)1093; AVX2-FCP-NEXT: vmovaps %ymm10, 96(%rsi)1094; AVX2-FCP-NEXT: vmovaps %ymm2, 32(%rsi)1095; AVX2-FCP-NEXT: vmovdqa %ymm0, 64(%rdx)1096; AVX2-FCP-NEXT: vmovdqa %ymm4, (%rdx)1097; AVX2-FCP-NEXT: vmovdqa %ymm3, 96(%rdx)1098; AVX2-FCP-NEXT: vmovdqa %ymm7, 32(%rdx)1099; AVX2-FCP-NEXT: vzeroupper1100; AVX2-FCP-NEXT: retq1101;1102; AVX512-VL-LABEL: load_i16_stride2_vf64:1103; AVX512-VL: # %bb.0:1104; AVX512-VL-NEXT: vmovdqa64 (%rdi), %zmm01105; AVX512-VL-NEXT: vmovdqa64 64(%rdi), %zmm11106; AVX512-VL-NEXT: vmovdqa64 128(%rdi), %zmm21107; AVX512-VL-NEXT: vmovdqa64 192(%rdi), %zmm31108; AVX512-VL-NEXT: vpsrld $16, %zmm0, %zmm41109; AVX512-VL-NEXT: vpsrld $16, %zmm1, %zmm51110; AVX512-VL-NEXT: vpsrld $16, %zmm2, %zmm61111; AVX512-VL-NEXT: vpsrld $16, %zmm3, %zmm71112; AVX512-VL-NEXT: vpmovdw %zmm1, 32(%rsi)1113; AVX512-VL-NEXT: vpmovdw %zmm0, (%rsi)1114; AVX512-VL-NEXT: vpmovdw %zmm3, 96(%rsi)1115; AVX512-VL-NEXT: vpmovdw %zmm2, 64(%rsi)1116; AVX512-VL-NEXT: vpmovdw %zmm7, 96(%rdx)1117; AVX512-VL-NEXT: vpmovdw %zmm6, 64(%rdx)1118; AVX512-VL-NEXT: vpmovdw %zmm5, 32(%rdx)1119; AVX512-VL-NEXT: vpmovdw %zmm4, (%rdx)1120; AVX512-VL-NEXT: vzeroupper1121; AVX512-VL-NEXT: retq1122;1123; AVX512-FCP-LABEL: load_i16_stride2_vf64:1124; AVX512-FCP: # %bb.0:1125; AVX512-FCP-NEXT: vmovdqa64 (%rdi), %zmm01126; AVX512-FCP-NEXT: vmovdqa64 64(%rdi), %zmm11127; AVX512-FCP-NEXT: vmovdqa64 128(%rdi), %zmm21128; AVX512-FCP-NEXT: vmovdqa64 192(%rdi), %zmm31129; AVX512-FCP-NEXT: vpsrld $16, %zmm0, %zmm41130; AVX512-FCP-NEXT: vpsrld $16, %zmm1, %zmm51131; AVX512-FCP-NEXT: vpsrld $16, %zmm2, %zmm61132; AVX512-FCP-NEXT: vpsrld $16, %zmm3, %zmm71133; AVX512-FCP-NEXT: vpmovdw %zmm1, 32(%rsi)1134; AVX512-FCP-NEXT: vpmovdw %zmm0, (%rsi)1135; AVX512-FCP-NEXT: vpmovdw %zmm3, 96(%rsi)1136; AVX512-FCP-NEXT: vpmovdw %zmm2, 64(%rsi)1137; AVX512-FCP-NEXT: vpmovdw %zmm7, 96(%rdx)1138; AVX512-FCP-NEXT: vpmovdw %zmm6, 64(%rdx)1139; AVX512-FCP-NEXT: vpmovdw %zmm5, 32(%rdx)1140; AVX512-FCP-NEXT: vpmovdw %zmm4, (%rdx)1141; AVX512-FCP-NEXT: vzeroupper1142; AVX512-FCP-NEXT: retq1143;1144; AVX512DQ-LABEL: load_i16_stride2_vf64:1145; AVX512DQ: # %bb.0:1146; AVX512DQ-NEXT: vmovdqa64 (%rdi), %zmm01147; AVX512DQ-NEXT: vmovdqa64 64(%rdi), %zmm11148; AVX512DQ-NEXT: vmovdqa64 128(%rdi), %zmm21149; AVX512DQ-NEXT: vmovdqa64 192(%rdi), %zmm31150; AVX512DQ-NEXT: vpsrld $16, %zmm0, %zmm41151; AVX512DQ-NEXT: vpsrld $16, %zmm1, %zmm51152; AVX512DQ-NEXT: vpsrld $16, %zmm2, %zmm61153; AVX512DQ-NEXT: vpsrld $16, %zmm3, %zmm71154; AVX512DQ-NEXT: vpmovdw %zmm1, 32(%rsi)1155; AVX512DQ-NEXT: vpmovdw %zmm0, (%rsi)1156; AVX512DQ-NEXT: vpmovdw %zmm3, 96(%rsi)1157; AVX512DQ-NEXT: vpmovdw %zmm2, 64(%rsi)1158; AVX512DQ-NEXT: vpmovdw %zmm7, 96(%rdx)1159; AVX512DQ-NEXT: vpmovdw %zmm6, 64(%rdx)1160; AVX512DQ-NEXT: vpmovdw %zmm5, 32(%rdx)1161; AVX512DQ-NEXT: vpmovdw %zmm4, (%rdx)1162; AVX512DQ-NEXT: vzeroupper1163; AVX512DQ-NEXT: retq1164;1165; AVX512DQ-FCP-LABEL: load_i16_stride2_vf64:1166; AVX512DQ-FCP: # %bb.0:1167; AVX512DQ-FCP-NEXT: vmovdqa64 (%rdi), %zmm01168; AVX512DQ-FCP-NEXT: vmovdqa64 64(%rdi), %zmm11169; AVX512DQ-FCP-NEXT: vmovdqa64 128(%rdi), %zmm21170; AVX512DQ-FCP-NEXT: vmovdqa64 192(%rdi), %zmm31171; AVX512DQ-FCP-NEXT: vpsrld $16, %zmm0, %zmm41172; AVX512DQ-FCP-NEXT: vpsrld $16, %zmm1, %zmm51173; AVX512DQ-FCP-NEXT: vpsrld $16, %zmm2, %zmm61174; AVX512DQ-FCP-NEXT: vpsrld $16, %zmm3, %zmm71175; AVX512DQ-FCP-NEXT: vpmovdw %zmm1, 32(%rsi)1176; AVX512DQ-FCP-NEXT: vpmovdw %zmm0, (%rsi)1177; AVX512DQ-FCP-NEXT: vpmovdw %zmm3, 96(%rsi)1178; AVX512DQ-FCP-NEXT: vpmovdw %zmm2, 64(%rsi)1179; AVX512DQ-FCP-NEXT: vpmovdw %zmm7, 96(%rdx)1180; AVX512DQ-FCP-NEXT: vpmovdw %zmm6, 64(%rdx)1181; AVX512DQ-FCP-NEXT: vpmovdw %zmm5, 32(%rdx)1182; AVX512DQ-FCP-NEXT: vpmovdw %zmm4, (%rdx)1183; AVX512DQ-FCP-NEXT: vzeroupper1184; AVX512DQ-FCP-NEXT: retq1185;1186; AVX512BW-LABEL: load_i16_stride2_vf64:1187; AVX512BW: # %bb.0:1188; AVX512BW-NEXT: vmovdqa64 (%rdi), %zmm01189; AVX512BW-NEXT: vmovdqa64 64(%rdi), %zmm11190; AVX512BW-NEXT: vmovdqa64 128(%rdi), %zmm21191; AVX512BW-NEXT: vmovdqa64 192(%rdi), %zmm31192; AVX512BW-NEXT: vpmovsxbw {{.*#+}} zmm4 = [0,2,4,6,8,10,12,14,16,18,20,22,24,26,28,30,32,34,36,38,40,42,44,46,48,50,52,54,56,58,60,62]1193; AVX512BW-NEXT: vmovdqa64 %zmm0, %zmm51194; AVX512BW-NEXT: vpermt2w %zmm1, %zmm4, %zmm51195; AVX512BW-NEXT: vpermi2w %zmm3, %zmm2, %zmm41196; AVX512BW-NEXT: vpmovsxbw {{.*#+}} zmm6 = [1,3,5,7,9,11,13,15,17,19,21,23,25,27,29,31,33,35,37,39,41,43,45,47,49,51,53,55,57,59,61,63]1197; AVX512BW-NEXT: vpermt2w %zmm1, %zmm6, %zmm01198; AVX512BW-NEXT: vpermt2w %zmm3, %zmm6, %zmm21199; AVX512BW-NEXT: vmovdqa64 %zmm4, 64(%rsi)1200; AVX512BW-NEXT: vmovdqa64 %zmm5, (%rsi)1201; AVX512BW-NEXT: vmovdqa64 %zmm2, 64(%rdx)1202; AVX512BW-NEXT: vmovdqa64 %zmm0, (%rdx)1203; AVX512BW-NEXT: vzeroupper1204; AVX512BW-NEXT: retq1205;1206; AVX512BW-FCP-LABEL: load_i16_stride2_vf64:1207; AVX512BW-FCP: # %bb.0:1208; AVX512BW-FCP-NEXT: vmovdqa64 (%rdi), %zmm01209; AVX512BW-FCP-NEXT: vmovdqa64 64(%rdi), %zmm11210; AVX512BW-FCP-NEXT: vmovdqa64 128(%rdi), %zmm21211; AVX512BW-FCP-NEXT: vmovdqa64 192(%rdi), %zmm31212; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm4 = [0,2,4,6,8,10,12,14,16,18,20,22,24,26,28,30,32,34,36,38,40,42,44,46,48,50,52,54,56,58,60,62]1213; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm51214; AVX512BW-FCP-NEXT: vpermt2w %zmm1, %zmm4, %zmm51215; AVX512BW-FCP-NEXT: vpermi2w %zmm3, %zmm2, %zmm41216; AVX512BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm6 = [1,3,5,7,9,11,13,15,17,19,21,23,25,27,29,31,33,35,37,39,41,43,45,47,49,51,53,55,57,59,61,63]1217; AVX512BW-FCP-NEXT: vpermt2w %zmm1, %zmm6, %zmm01218; AVX512BW-FCP-NEXT: vpermt2w %zmm3, %zmm6, %zmm21219; AVX512BW-FCP-NEXT: vmovdqa64 %zmm4, 64(%rsi)1220; AVX512BW-FCP-NEXT: vmovdqa64 %zmm5, (%rsi)1221; AVX512BW-FCP-NEXT: vmovdqa64 %zmm2, 64(%rdx)1222; AVX512BW-FCP-NEXT: vmovdqa64 %zmm0, (%rdx)1223; AVX512BW-FCP-NEXT: vzeroupper1224; AVX512BW-FCP-NEXT: retq1225;1226; AVX512DQ-BW-LABEL: load_i16_stride2_vf64:1227; AVX512DQ-BW: # %bb.0:1228; AVX512DQ-BW-NEXT: vmovdqa64 (%rdi), %zmm01229; AVX512DQ-BW-NEXT: vmovdqa64 64(%rdi), %zmm11230; AVX512DQ-BW-NEXT: vmovdqa64 128(%rdi), %zmm21231; AVX512DQ-BW-NEXT: vmovdqa64 192(%rdi), %zmm31232; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} zmm4 = [0,2,4,6,8,10,12,14,16,18,20,22,24,26,28,30,32,34,36,38,40,42,44,46,48,50,52,54,56,58,60,62]1233; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, %zmm51234; AVX512DQ-BW-NEXT: vpermt2w %zmm1, %zmm4, %zmm51235; AVX512DQ-BW-NEXT: vpermi2w %zmm3, %zmm2, %zmm41236; AVX512DQ-BW-NEXT: vpmovsxbw {{.*#+}} zmm6 = [1,3,5,7,9,11,13,15,17,19,21,23,25,27,29,31,33,35,37,39,41,43,45,47,49,51,53,55,57,59,61,63]1237; AVX512DQ-BW-NEXT: vpermt2w %zmm1, %zmm6, %zmm01238; AVX512DQ-BW-NEXT: vpermt2w %zmm3, %zmm6, %zmm21239; AVX512DQ-BW-NEXT: vmovdqa64 %zmm4, 64(%rsi)1240; AVX512DQ-BW-NEXT: vmovdqa64 %zmm5, (%rsi)1241; AVX512DQ-BW-NEXT: vmovdqa64 %zmm2, 64(%rdx)1242; AVX512DQ-BW-NEXT: vmovdqa64 %zmm0, (%rdx)1243; AVX512DQ-BW-NEXT: vzeroupper1244; AVX512DQ-BW-NEXT: retq1245;1246; AVX512DQ-BW-FCP-LABEL: load_i16_stride2_vf64:1247; AVX512DQ-BW-FCP: # %bb.0:1248; AVX512DQ-BW-FCP-NEXT: vmovdqa64 (%rdi), %zmm01249; AVX512DQ-BW-FCP-NEXT: vmovdqa64 64(%rdi), %zmm11250; AVX512DQ-BW-FCP-NEXT: vmovdqa64 128(%rdi), %zmm21251; AVX512DQ-BW-FCP-NEXT: vmovdqa64 192(%rdi), %zmm31252; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm4 = [0,2,4,6,8,10,12,14,16,18,20,22,24,26,28,30,32,34,36,38,40,42,44,46,48,50,52,54,56,58,60,62]1253; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, %zmm51254; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm1, %zmm4, %zmm51255; AVX512DQ-BW-FCP-NEXT: vpermi2w %zmm3, %zmm2, %zmm41256; AVX512DQ-BW-FCP-NEXT: vpmovsxbw {{.*#+}} zmm6 = [1,3,5,7,9,11,13,15,17,19,21,23,25,27,29,31,33,35,37,39,41,43,45,47,49,51,53,55,57,59,61,63]1257; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm1, %zmm6, %zmm01258; AVX512DQ-BW-FCP-NEXT: vpermt2w %zmm3, %zmm6, %zmm21259; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm4, 64(%rsi)1260; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm5, (%rsi)1261; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm2, 64(%rdx)1262; AVX512DQ-BW-FCP-NEXT: vmovdqa64 %zmm0, (%rdx)1263; AVX512DQ-BW-FCP-NEXT: vzeroupper1264; AVX512DQ-BW-FCP-NEXT: retq1265 %wide.vec = load <128 x i16>, ptr %in.vec, align 641266 %strided.vec0 = shufflevector <128 x i16> %wide.vec, <128 x i16> poison, <64 x i32> <i32 0, i32 2, i32 4, i32 6, i32 8, i32 10, i32 12, i32 14, i32 16, i32 18, i32 20, i32 22, i32 24, i32 26, i32 28, i32 30, i32 32, i32 34, i32 36, i32 38, i32 40, i32 42, i32 44, i32 46, i32 48, i32 50, i32 52, i32 54, i32 56, i32 58, i32 60, i32 62, i32 64, i32 66, i32 68, i32 70, i32 72, i32 74, i32 76, i32 78, i32 80, i32 82, i32 84, i32 86, i32 88, i32 90, i32 92, i32 94, i32 96, i32 98, i32 100, i32 102, i32 104, i32 106, i32 108, i32 110, i32 112, i32 114, i32 116, i32 118, i32 120, i32 122, i32 124, i32 126>1267 %strided.vec1 = shufflevector <128 x i16> %wide.vec, <128 x i16> poison, <64 x i32> <i32 1, i32 3, i32 5, i32 7, i32 9, i32 11, i32 13, i32 15, i32 17, i32 19, i32 21, i32 23, i32 25, i32 27, i32 29, i32 31, i32 33, i32 35, i32 37, i32 39, i32 41, i32 43, i32 45, i32 47, i32 49, i32 51, i32 53, i32 55, i32 57, i32 59, i32 61, i32 63, i32 65, i32 67, i32 69, i32 71, i32 73, i32 75, i32 77, i32 79, i32 81, i32 83, i32 85, i32 87, i32 89, i32 91, i32 93, i32 95, i32 97, i32 99, i32 101, i32 103, i32 105, i32 107, i32 109, i32 111, i32 113, i32 115, i32 117, i32 119, i32 121, i32 123, i32 125, i32 127>1268 store <64 x i16> %strided.vec0, ptr %out.vec0, align 641269 store <64 x i16> %strided.vec1, ptr %out.vec1, align 641270 ret void1271}1272