1378 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx | FileCheck %s --check-prefixes=AVX,AVX13; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx2 | FileCheck %s --check-prefixes=AVX,AVX2,AVX2-SLOW4; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx2,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX,AVX2,AVX2-FAST-ALL5; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx2,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX,AVX2,AVX2-FAST-PERLANE6; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f | FileCheck %s --check-prefix=AVX512F7; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512vl,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX512VL8; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512vl,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX512VL9; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512bw,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX512BW10; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512bw,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX512BW11; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512bw,+avx512vl,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX512BWVL12; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512bw,+avx512vl,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX512BWVL13; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512vbmi,+avx512vl,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX512VBMIVL14; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512vbmi,+avx512vl,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=AVX512VBMIVL15 16; PR3155117; Pairs of shufflevector:trunc functions with functional equivalence.18; Ideally, the shuffles should be lowered to code with the same quality as the truncates.19 20define void @shuffle_v32i8_to_v16i8(ptr %L, ptr %S) nounwind {21; AVX1-LABEL: shuffle_v32i8_to_v16i8:22; AVX1: # %bb.0:23; AVX1-NEXT: vbroadcastss {{.*#+}} xmm0 = [255,255,255,255,255,255,255,255]24; AVX1-NEXT: vpand 16(%rdi), %xmm0, %xmm125; AVX1-NEXT: vpand (%rdi), %xmm0, %xmm026; AVX1-NEXT: vpackuswb %xmm1, %xmm0, %xmm027; AVX1-NEXT: vmovdqa %xmm0, (%rsi)28; AVX1-NEXT: retq29;30; AVX2-LABEL: shuffle_v32i8_to_v16i8:31; AVX2: # %bb.0:32; AVX2-NEXT: vpbroadcastw {{.*#+}} xmm0 = [255,255,255,255,255,255,255,255]33; AVX2-NEXT: vpand 16(%rdi), %xmm0, %xmm134; AVX2-NEXT: vpand (%rdi), %xmm0, %xmm035; AVX2-NEXT: vpackuswb %xmm1, %xmm0, %xmm036; AVX2-NEXT: vmovdqa %xmm0, (%rsi)37; AVX2-NEXT: retq38;39; AVX512F-LABEL: shuffle_v32i8_to_v16i8:40; AVX512F: # %bb.0:41; AVX512F-NEXT: vpbroadcastw {{.*#+}} xmm0 = [255,255,255,255,255,255,255,255]42; AVX512F-NEXT: vpand 16(%rdi), %xmm0, %xmm143; AVX512F-NEXT: vpand (%rdi), %xmm0, %xmm044; AVX512F-NEXT: vpackuswb %xmm1, %xmm0, %xmm045; AVX512F-NEXT: vmovdqa %xmm0, (%rsi)46; AVX512F-NEXT: retq47;48; AVX512VL-LABEL: shuffle_v32i8_to_v16i8:49; AVX512VL: # %bb.0:50; AVX512VL-NEXT: vpbroadcastd {{.*#+}} xmm0 = [255,255,255,255,255,255,255,255]51; AVX512VL-NEXT: vpand 16(%rdi), %xmm0, %xmm152; AVX512VL-NEXT: vpand (%rdi), %xmm0, %xmm053; AVX512VL-NEXT: vpackuswb %xmm1, %xmm0, %xmm054; AVX512VL-NEXT: vmovdqa %xmm0, (%rsi)55; AVX512VL-NEXT: retq56;57; AVX512BW-LABEL: shuffle_v32i8_to_v16i8:58; AVX512BW: # %bb.0:59; AVX512BW-NEXT: vmovdqa (%rdi), %ymm060; AVX512BW-NEXT: vpmovwb %zmm0, %ymm061; AVX512BW-NEXT: vmovdqa %xmm0, (%rsi)62; AVX512BW-NEXT: vzeroupper63; AVX512BW-NEXT: retq64;65; AVX512BWVL-LABEL: shuffle_v32i8_to_v16i8:66; AVX512BWVL: # %bb.0:67; AVX512BWVL-NEXT: vmovdqa (%rdi), %ymm068; AVX512BWVL-NEXT: vpmovwb %ymm0, (%rsi)69; AVX512BWVL-NEXT: vzeroupper70; AVX512BWVL-NEXT: retq71;72; AVX512VBMIVL-LABEL: shuffle_v32i8_to_v16i8:73; AVX512VBMIVL: # %bb.0:74; AVX512VBMIVL-NEXT: vmovdqa (%rdi), %ymm075; AVX512VBMIVL-NEXT: vpmovwb %ymm0, (%rsi)76; AVX512VBMIVL-NEXT: vzeroupper77; AVX512VBMIVL-NEXT: retq78 %vec = load <32 x i8>, ptr %L79 %strided.vec = shufflevector <32 x i8> %vec, <32 x i8> undef, <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>80 store <16 x i8> %strided.vec, ptr %S81 ret void82}83 84define void @trunc_v16i16_to_v16i8(ptr %L, ptr %S) nounwind {85; AVX1-LABEL: trunc_v16i16_to_v16i8:86; AVX1: # %bb.0:87; AVX1-NEXT: vmovaps (%rdi), %ymm088; AVX1-NEXT: vandps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm089; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm190; AVX1-NEXT: vpackuswb %xmm1, %xmm0, %xmm091; AVX1-NEXT: vmovdqa %xmm0, (%rsi)92; AVX1-NEXT: vzeroupper93; AVX1-NEXT: retq94;95; AVX2-LABEL: trunc_v16i16_to_v16i8:96; AVX2: # %bb.0:97; AVX2-NEXT: vmovdqa (%rdi), %ymm098; AVX2-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm099; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1100; AVX2-NEXT: vpackuswb %xmm1, %xmm0, %xmm0101; AVX2-NEXT: vmovdqa %xmm0, (%rsi)102; AVX2-NEXT: vzeroupper103; AVX2-NEXT: retq104;105; AVX512F-LABEL: trunc_v16i16_to_v16i8:106; AVX512F: # %bb.0:107; AVX512F-NEXT: vpmovzxwd {{.*#+}} zmm0 = mem[0],zero,mem[1],zero,mem[2],zero,mem[3],zero,mem[4],zero,mem[5],zero,mem[6],zero,mem[7],zero,mem[8],zero,mem[9],zero,mem[10],zero,mem[11],zero,mem[12],zero,mem[13],zero,mem[14],zero,mem[15],zero108; AVX512F-NEXT: vpmovdb %zmm0, (%rsi)109; AVX512F-NEXT: vzeroupper110; AVX512F-NEXT: retq111;112; AVX512VL-LABEL: trunc_v16i16_to_v16i8:113; AVX512VL: # %bb.0:114; AVX512VL-NEXT: vpmovzxwd {{.*#+}} zmm0 = mem[0],zero,mem[1],zero,mem[2],zero,mem[3],zero,mem[4],zero,mem[5],zero,mem[6],zero,mem[7],zero,mem[8],zero,mem[9],zero,mem[10],zero,mem[11],zero,mem[12],zero,mem[13],zero,mem[14],zero,mem[15],zero115; AVX512VL-NEXT: vpmovdb %zmm0, (%rsi)116; AVX512VL-NEXT: vzeroupper117; AVX512VL-NEXT: retq118;119; AVX512BW-LABEL: trunc_v16i16_to_v16i8:120; AVX512BW: # %bb.0:121; AVX512BW-NEXT: vmovdqa (%rdi), %ymm0122; AVX512BW-NEXT: vpmovwb %zmm0, %ymm0123; AVX512BW-NEXT: vmovdqa %xmm0, (%rsi)124; AVX512BW-NEXT: vzeroupper125; AVX512BW-NEXT: retq126;127; AVX512BWVL-LABEL: trunc_v16i16_to_v16i8:128; AVX512BWVL: # %bb.0:129; AVX512BWVL-NEXT: vmovdqa (%rdi), %ymm0130; AVX512BWVL-NEXT: vpmovwb %ymm0, (%rsi)131; AVX512BWVL-NEXT: vzeroupper132; AVX512BWVL-NEXT: retq133;134; AVX512VBMIVL-LABEL: trunc_v16i16_to_v16i8:135; AVX512VBMIVL: # %bb.0:136; AVX512VBMIVL-NEXT: vmovdqa (%rdi), %ymm0137; AVX512VBMIVL-NEXT: vpmovwb %ymm0, (%rsi)138; AVX512VBMIVL-NEXT: vzeroupper139; AVX512VBMIVL-NEXT: retq140 %vec = load <32 x i8>, ptr %L141 %bc = bitcast <32 x i8> %vec to <16 x i16>142 %strided.vec = trunc <16 x i16> %bc to <16 x i8>143 store <16 x i8> %strided.vec, ptr %S144 ret void145}146 147define void @shuffle_v16i16_to_v8i16(ptr %L, ptr %S) nounwind {148; AVX-LABEL: shuffle_v16i16_to_v8i16:149; AVX: # %bb.0:150; AVX-NEXT: vpxor %xmm0, %xmm0, %xmm0151; AVX-NEXT: vpblendw {{.*#+}} xmm1 = mem[0],xmm0[1],mem[2],xmm0[3],mem[4],xmm0[5],mem[6],xmm0[7]152; AVX-NEXT: vpblendw {{.*#+}} xmm0 = mem[0],xmm0[1],mem[2],xmm0[3],mem[4],xmm0[5],mem[6],xmm0[7]153; AVX-NEXT: vpackusdw %xmm1, %xmm0, %xmm0154; AVX-NEXT: vmovdqa %xmm0, (%rsi)155; AVX-NEXT: retq156;157; AVX512F-LABEL: shuffle_v16i16_to_v8i16:158; AVX512F: # %bb.0:159; AVX512F-NEXT: vmovdqa (%rdi), %ymm0160; AVX512F-NEXT: vpmovdw %zmm0, %ymm0161; AVX512F-NEXT: vmovdqa %xmm0, (%rsi)162; AVX512F-NEXT: vzeroupper163; AVX512F-NEXT: retq164;165; AVX512VL-LABEL: shuffle_v16i16_to_v8i16:166; AVX512VL: # %bb.0:167; AVX512VL-NEXT: vmovdqa (%rdi), %ymm0168; AVX512VL-NEXT: vpmovdw %ymm0, (%rsi)169; AVX512VL-NEXT: vzeroupper170; AVX512VL-NEXT: retq171;172; AVX512BW-LABEL: shuffle_v16i16_to_v8i16:173; AVX512BW: # %bb.0:174; AVX512BW-NEXT: vmovdqa (%rdi), %ymm0175; AVX512BW-NEXT: vpmovdw %zmm0, %ymm0176; AVX512BW-NEXT: vmovdqa %xmm0, (%rsi)177; AVX512BW-NEXT: vzeroupper178; AVX512BW-NEXT: retq179;180; AVX512BWVL-LABEL: shuffle_v16i16_to_v8i16:181; AVX512BWVL: # %bb.0:182; AVX512BWVL-NEXT: vmovdqa (%rdi), %ymm0183; AVX512BWVL-NEXT: vpmovdw %ymm0, (%rsi)184; AVX512BWVL-NEXT: vzeroupper185; AVX512BWVL-NEXT: retq186;187; AVX512VBMIVL-LABEL: shuffle_v16i16_to_v8i16:188; AVX512VBMIVL: # %bb.0:189; AVX512VBMIVL-NEXT: vmovdqa (%rdi), %ymm0190; AVX512VBMIVL-NEXT: vpmovdw %ymm0, (%rsi)191; AVX512VBMIVL-NEXT: vzeroupper192; AVX512VBMIVL-NEXT: retq193 %vec = load <16 x i16>, ptr %L194 %strided.vec = shufflevector <16 x i16> %vec, <16 x i16> undef, <8 x i32> <i32 0, i32 2, i32 4, i32 6, i32 8, i32 10, i32 12, i32 14>195 store <8 x i16> %strided.vec, ptr %S196 ret void197}198 199define void @trunc_v8i32_to_v8i16(ptr %L, ptr %S) nounwind {200; AVX1-LABEL: trunc_v8i32_to_v8i16:201; AVX1: # %bb.0:202; AVX1-NEXT: vmovaps (%rdi), %ymm0203; AVX1-NEXT: vandps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm0204; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1205; AVX1-NEXT: vpackusdw %xmm1, %xmm0, %xmm0206; AVX1-NEXT: vmovdqa %xmm0, (%rsi)207; AVX1-NEXT: vzeroupper208; AVX1-NEXT: retq209;210; AVX2-LABEL: trunc_v8i32_to_v8i16:211; AVX2: # %bb.0:212; AVX2-NEXT: vmovdqa (%rdi), %ymm0213; AVX2-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,1,4,5,8,9,12,13,u,u,u,u,u,u,u,u,16,17,20,21,24,25,28,29,u,u,u,u,u,u,u,u]214; AVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,2,2,3]215; AVX2-NEXT: vmovdqa %xmm0, (%rsi)216; AVX2-NEXT: vzeroupper217; AVX2-NEXT: retq218;219; AVX512F-LABEL: trunc_v8i32_to_v8i16:220; AVX512F: # %bb.0:221; AVX512F-NEXT: vmovdqa (%rdi), %ymm0222; AVX512F-NEXT: vpmovdw %zmm0, %ymm0223; AVX512F-NEXT: vmovdqa %xmm0, (%rsi)224; AVX512F-NEXT: vzeroupper225; AVX512F-NEXT: retq226;227; AVX512VL-LABEL: trunc_v8i32_to_v8i16:228; AVX512VL: # %bb.0:229; AVX512VL-NEXT: vmovdqa (%rdi), %ymm0230; AVX512VL-NEXT: vpmovdw %ymm0, (%rsi)231; AVX512VL-NEXT: vzeroupper232; AVX512VL-NEXT: retq233;234; AVX512BW-LABEL: trunc_v8i32_to_v8i16:235; AVX512BW: # %bb.0:236; AVX512BW-NEXT: vmovdqa (%rdi), %ymm0237; AVX512BW-NEXT: vpmovdw %zmm0, %ymm0238; AVX512BW-NEXT: vmovdqa %xmm0, (%rsi)239; AVX512BW-NEXT: vzeroupper240; AVX512BW-NEXT: retq241;242; AVX512BWVL-LABEL: trunc_v8i32_to_v8i16:243; AVX512BWVL: # %bb.0:244; AVX512BWVL-NEXT: vmovdqa (%rdi), %ymm0245; AVX512BWVL-NEXT: vpmovdw %ymm0, (%rsi)246; AVX512BWVL-NEXT: vzeroupper247; AVX512BWVL-NEXT: retq248;249; AVX512VBMIVL-LABEL: trunc_v8i32_to_v8i16:250; AVX512VBMIVL: # %bb.0:251; AVX512VBMIVL-NEXT: vmovdqa (%rdi), %ymm0252; AVX512VBMIVL-NEXT: vpmovdw %ymm0, (%rsi)253; AVX512VBMIVL-NEXT: vzeroupper254; AVX512VBMIVL-NEXT: retq255 %vec = load <16 x i16>, ptr %L256 %bc = bitcast <16 x i16> %vec to <8 x i32>257 %strided.vec = trunc <8 x i32> %bc to <8 x i16>258 store <8 x i16> %strided.vec, ptr %S259 ret void260}261 262define void @shuffle_v8i32_to_v4i32(ptr %L, ptr %S) nounwind {263; AVX-LABEL: shuffle_v8i32_to_v4i32:264; AVX: # %bb.0:265; AVX-NEXT: vmovaps (%rdi), %xmm0266; AVX-NEXT: vshufps {{.*#+}} xmm0 = xmm0[0,2],mem[0,2]267; AVX-NEXT: vmovaps %xmm0, (%rsi)268; AVX-NEXT: retq269;270; AVX512F-LABEL: shuffle_v8i32_to_v4i32:271; AVX512F: # %bb.0:272; AVX512F-NEXT: vmovaps (%rdi), %xmm0273; AVX512F-NEXT: vshufps {{.*#+}} xmm0 = xmm0[0,2],mem[0,2]274; AVX512F-NEXT: vmovaps %xmm0, (%rsi)275; AVX512F-NEXT: retq276;277; AVX512VL-LABEL: shuffle_v8i32_to_v4i32:278; AVX512VL: # %bb.0:279; AVX512VL-NEXT: vmovdqa (%rdi), %ymm0280; AVX512VL-NEXT: vpmovqd %ymm0, (%rsi)281; AVX512VL-NEXT: vzeroupper282; AVX512VL-NEXT: retq283;284; AVX512BW-LABEL: shuffle_v8i32_to_v4i32:285; AVX512BW: # %bb.0:286; AVX512BW-NEXT: vmovaps (%rdi), %xmm0287; AVX512BW-NEXT: vshufps {{.*#+}} xmm0 = xmm0[0,2],mem[0,2]288; AVX512BW-NEXT: vmovaps %xmm0, (%rsi)289; AVX512BW-NEXT: retq290;291; AVX512BWVL-LABEL: shuffle_v8i32_to_v4i32:292; AVX512BWVL: # %bb.0:293; AVX512BWVL-NEXT: vmovdqa (%rdi), %ymm0294; AVX512BWVL-NEXT: vpmovqd %ymm0, (%rsi)295; AVX512BWVL-NEXT: vzeroupper296; AVX512BWVL-NEXT: retq297;298; AVX512VBMIVL-LABEL: shuffle_v8i32_to_v4i32:299; AVX512VBMIVL: # %bb.0:300; AVX512VBMIVL-NEXT: vmovdqa (%rdi), %ymm0301; AVX512VBMIVL-NEXT: vpmovqd %ymm0, (%rsi)302; AVX512VBMIVL-NEXT: vzeroupper303; AVX512VBMIVL-NEXT: retq304 %vec = load <8 x i32>, ptr %L305 %strided.vec = shufflevector <8 x i32> %vec, <8 x i32> undef, <4 x i32> <i32 0, i32 2, i32 4, i32 6>306 store <4 x i32> %strided.vec, ptr %S307 ret void308}309 310define void @trunc_v4i64_to_v4i32(ptr %L, ptr %S) nounwind {311; AVX1-LABEL: trunc_v4i64_to_v4i32:312; AVX1: # %bb.0:313; AVX1-NEXT: vmovaps (%rdi), %xmm0314; AVX1-NEXT: vshufps {{.*#+}} xmm0 = xmm0[0,2],mem[0,2]315; AVX1-NEXT: vmovaps %xmm0, (%rsi)316; AVX1-NEXT: retq317;318; AVX2-SLOW-LABEL: trunc_v4i64_to_v4i32:319; AVX2-SLOW: # %bb.0:320; AVX2-SLOW-NEXT: vmovaps (%rdi), %xmm0321; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm0 = xmm0[0,2],mem[0,2]322; AVX2-SLOW-NEXT: vmovaps %xmm0, (%rsi)323; AVX2-SLOW-NEXT: retq324;325; AVX2-FAST-ALL-LABEL: trunc_v4i64_to_v4i32:326; AVX2-FAST-ALL: # %bb.0:327; AVX2-FAST-ALL-NEXT: vbroadcastf128 {{.*#+}} ymm0 = [0,2,4,6,0,2,4,6]328; AVX2-FAST-ALL-NEXT: # ymm0 = mem[0,1,0,1]329; AVX2-FAST-ALL-NEXT: vpermps (%rdi), %ymm0, %ymm0330; AVX2-FAST-ALL-NEXT: vmovaps %xmm0, (%rsi)331; AVX2-FAST-ALL-NEXT: vzeroupper332; AVX2-FAST-ALL-NEXT: retq333;334; AVX2-FAST-PERLANE-LABEL: trunc_v4i64_to_v4i32:335; AVX2-FAST-PERLANE: # %bb.0:336; AVX2-FAST-PERLANE-NEXT: vmovaps (%rdi), %xmm0337; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm0 = xmm0[0,2],mem[0,2]338; AVX2-FAST-PERLANE-NEXT: vmovaps %xmm0, (%rsi)339; AVX2-FAST-PERLANE-NEXT: retq340;341; AVX512F-LABEL: trunc_v4i64_to_v4i32:342; AVX512F: # %bb.0:343; AVX512F-NEXT: vmovdqa (%rdi), %ymm0344; AVX512F-NEXT: vpmovqd %zmm0, %ymm0345; AVX512F-NEXT: vmovdqa %xmm0, (%rsi)346; AVX512F-NEXT: vzeroupper347; AVX512F-NEXT: retq348;349; AVX512VL-LABEL: trunc_v4i64_to_v4i32:350; AVX512VL: # %bb.0:351; AVX512VL-NEXT: vmovdqa (%rdi), %ymm0352; AVX512VL-NEXT: vpmovqd %ymm0, (%rsi)353; AVX512VL-NEXT: vzeroupper354; AVX512VL-NEXT: retq355;356; AVX512BW-LABEL: trunc_v4i64_to_v4i32:357; AVX512BW: # %bb.0:358; AVX512BW-NEXT: vmovdqa (%rdi), %ymm0359; AVX512BW-NEXT: vpmovqd %zmm0, %ymm0360; AVX512BW-NEXT: vmovdqa %xmm0, (%rsi)361; AVX512BW-NEXT: vzeroupper362; AVX512BW-NEXT: retq363;364; AVX512BWVL-LABEL: trunc_v4i64_to_v4i32:365; AVX512BWVL: # %bb.0:366; AVX512BWVL-NEXT: vmovdqa (%rdi), %ymm0367; AVX512BWVL-NEXT: vpmovqd %ymm0, (%rsi)368; AVX512BWVL-NEXT: vzeroupper369; AVX512BWVL-NEXT: retq370;371; AVX512VBMIVL-LABEL: trunc_v4i64_to_v4i32:372; AVX512VBMIVL: # %bb.0:373; AVX512VBMIVL-NEXT: vmovdqa (%rdi), %ymm0374; AVX512VBMIVL-NEXT: vpmovqd %ymm0, (%rsi)375; AVX512VBMIVL-NEXT: vzeroupper376; AVX512VBMIVL-NEXT: retq377 %vec = load <8 x i32>, ptr %L378 %bc = bitcast <8 x i32> %vec to <4 x i64>379 %strided.vec = trunc <4 x i64> %bc to <4 x i32>380 store <4 x i32> %strided.vec, ptr %S381 ret void382}383 384define void @shuffle_v32i8_to_v8i8(ptr %L, ptr %S) nounwind {385; AVX-LABEL: shuffle_v32i8_to_v8i8:386; AVX: # %bb.0:387; AVX-NEXT: vmovdqa (%rdi), %xmm0388; AVX-NEXT: vmovdqa 16(%rdi), %xmm1389; AVX-NEXT: vmovd {{.*#+}} xmm2 = [0,4,8,12,0,0,0,0,0,0,0,0,0,0,0,0]390; AVX-NEXT: vpshufb %xmm2, %xmm1, %xmm1391; AVX-NEXT: vpshufb %xmm2, %xmm0, %xmm0392; AVX-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]393; AVX-NEXT: vmovq %xmm0, (%rsi)394; AVX-NEXT: retq395;396; AVX512F-LABEL: shuffle_v32i8_to_v8i8:397; AVX512F: # %bb.0:398; AVX512F-NEXT: vmovdqa (%rdi), %ymm0399; AVX512F-NEXT: vpmovdb %zmm0, %xmm0400; AVX512F-NEXT: vmovq %xmm0, (%rsi)401; AVX512F-NEXT: vzeroupper402; AVX512F-NEXT: retq403;404; AVX512VL-LABEL: shuffle_v32i8_to_v8i8:405; AVX512VL: # %bb.0:406; AVX512VL-NEXT: vmovdqa (%rdi), %ymm0407; AVX512VL-NEXT: vpmovdb %ymm0, (%rsi)408; AVX512VL-NEXT: vzeroupper409; AVX512VL-NEXT: retq410;411; AVX512BW-LABEL: shuffle_v32i8_to_v8i8:412; AVX512BW: # %bb.0:413; AVX512BW-NEXT: vmovdqa (%rdi), %ymm0414; AVX512BW-NEXT: vpmovdb %zmm0, %xmm0415; AVX512BW-NEXT: vmovq %xmm0, (%rsi)416; AVX512BW-NEXT: vzeroupper417; AVX512BW-NEXT: retq418;419; AVX512BWVL-LABEL: shuffle_v32i8_to_v8i8:420; AVX512BWVL: # %bb.0:421; AVX512BWVL-NEXT: vmovdqa (%rdi), %ymm0422; AVX512BWVL-NEXT: vpmovdb %ymm0, (%rsi)423; AVX512BWVL-NEXT: vzeroupper424; AVX512BWVL-NEXT: retq425;426; AVX512VBMIVL-LABEL: shuffle_v32i8_to_v8i8:427; AVX512VBMIVL: # %bb.0:428; AVX512VBMIVL-NEXT: vmovdqa (%rdi), %ymm0429; AVX512VBMIVL-NEXT: vpmovdb %ymm0, (%rsi)430; AVX512VBMIVL-NEXT: vzeroupper431; AVX512VBMIVL-NEXT: retq432 %vec = load <32 x i8>, ptr %L433 %strided.vec = shufflevector <32 x i8> %vec, <32 x i8> undef, <8 x i32> <i32 0, i32 4, i32 8, i32 12, i32 16, i32 20, i32 24, i32 28>434 store <8 x i8> %strided.vec, ptr %S435 ret void436}437 438define void @trunc_v8i32_to_v8i8(ptr %L, ptr %S) nounwind {439; AVX-LABEL: trunc_v8i32_to_v8i8:440; AVX: # %bb.0:441; AVX-NEXT: vmovdqa (%rdi), %xmm0442; AVX-NEXT: vmovdqa 16(%rdi), %xmm1443; AVX-NEXT: vmovd {{.*#+}} xmm2 = [0,4,8,12,0,0,0,0,0,0,0,0,0,0,0,0]444; AVX-NEXT: vpshufb %xmm2, %xmm1, %xmm1445; AVX-NEXT: vpshufb %xmm2, %xmm0, %xmm0446; AVX-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]447; AVX-NEXT: vmovq %xmm0, (%rsi)448; AVX-NEXT: retq449;450; AVX512F-LABEL: trunc_v8i32_to_v8i8:451; AVX512F: # %bb.0:452; AVX512F-NEXT: vmovdqa (%rdi), %ymm0453; AVX512F-NEXT: vpmovdb %zmm0, %xmm0454; AVX512F-NEXT: vmovq %xmm0, (%rsi)455; AVX512F-NEXT: vzeroupper456; AVX512F-NEXT: retq457;458; AVX512VL-LABEL: trunc_v8i32_to_v8i8:459; AVX512VL: # %bb.0:460; AVX512VL-NEXT: vmovdqa (%rdi), %ymm0461; AVX512VL-NEXT: vpmovdb %ymm0, (%rsi)462; AVX512VL-NEXT: vzeroupper463; AVX512VL-NEXT: retq464;465; AVX512BW-LABEL: trunc_v8i32_to_v8i8:466; AVX512BW: # %bb.0:467; AVX512BW-NEXT: vmovdqa (%rdi), %ymm0468; AVX512BW-NEXT: vpmovdb %zmm0, %xmm0469; AVX512BW-NEXT: vmovq %xmm0, (%rsi)470; AVX512BW-NEXT: vzeroupper471; AVX512BW-NEXT: retq472;473; AVX512BWVL-LABEL: trunc_v8i32_to_v8i8:474; AVX512BWVL: # %bb.0:475; AVX512BWVL-NEXT: vmovdqa (%rdi), %ymm0476; AVX512BWVL-NEXT: vpmovdb %ymm0, (%rsi)477; AVX512BWVL-NEXT: vzeroupper478; AVX512BWVL-NEXT: retq479;480; AVX512VBMIVL-LABEL: trunc_v8i32_to_v8i8:481; AVX512VBMIVL: # %bb.0:482; AVX512VBMIVL-NEXT: vmovdqa (%rdi), %ymm0483; AVX512VBMIVL-NEXT: vpmovdb %ymm0, (%rsi)484; AVX512VBMIVL-NEXT: vzeroupper485; AVX512VBMIVL-NEXT: retq486 %vec = load <32 x i8>, ptr %L487 %bc = bitcast <32 x i8> %vec to <8 x i32>488 %strided.vec = trunc <8 x i32> %bc to <8 x i8>489 store <8 x i8> %strided.vec, ptr %S490 ret void491}492 493; IR generated from:494; return (__m128i) {(long long)__builtin_convertvector((__v8si)__A, __v8qi), 0};495 496define <2 x i64> @trunc_v8i32_to_v8i8_return_v2i64(<8 x i32> %vec) nounwind {497; AVX1-LABEL: trunc_v8i32_to_v8i8_return_v2i64:498; AVX1: # %bb.0:499; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1500; AVX1-NEXT: vmovd {{.*#+}} xmm2 = [0,4,8,12,0,0,0,0,0,0,0,0,0,0,0,0]501; AVX1-NEXT: vpshufb %xmm2, %xmm1, %xmm1502; AVX1-NEXT: vpshufb %xmm2, %xmm0, %xmm0503; AVX1-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]504; AVX1-NEXT: vmovq {{.*#+}} xmm0 = xmm0[0],zero505; AVX1-NEXT: vzeroupper506; AVX1-NEXT: retq507;508; AVX2-LABEL: trunc_v8i32_to_v8i8_return_v2i64:509; AVX2: # %bb.0:510; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1511; AVX2-NEXT: vmovd {{.*#+}} xmm2 = [0,4,8,12,0,0,0,0,0,0,0,0,0,0,0,0]512; AVX2-NEXT: vpshufb %xmm2, %xmm1, %xmm1513; AVX2-NEXT: vpshufb %xmm2, %xmm0, %xmm0514; AVX2-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]515; AVX2-NEXT: vmovq {{.*#+}} xmm0 = xmm0[0],zero516; AVX2-NEXT: vzeroupper517; AVX2-NEXT: retq518;519; AVX512F-LABEL: trunc_v8i32_to_v8i8_return_v2i64:520; AVX512F: # %bb.0:521; AVX512F-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0522; AVX512F-NEXT: vpmovdb %zmm0, %xmm0523; AVX512F-NEXT: vmovq {{.*#+}} xmm0 = xmm0[0],zero524; AVX512F-NEXT: vzeroupper525; AVX512F-NEXT: retq526;527; AVX512VL-LABEL: trunc_v8i32_to_v8i8_return_v2i64:528; AVX512VL: # %bb.0:529; AVX512VL-NEXT: vpmovdb %ymm0, %xmm0530; AVX512VL-NEXT: vzeroupper531; AVX512VL-NEXT: retq532;533; AVX512BW-LABEL: trunc_v8i32_to_v8i8_return_v2i64:534; AVX512BW: # %bb.0:535; AVX512BW-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0536; AVX512BW-NEXT: vpmovdb %zmm0, %xmm0537; AVX512BW-NEXT: vmovq {{.*#+}} xmm0 = xmm0[0],zero538; AVX512BW-NEXT: vzeroupper539; AVX512BW-NEXT: retq540;541; AVX512BWVL-LABEL: trunc_v8i32_to_v8i8_return_v2i64:542; AVX512BWVL: # %bb.0:543; AVX512BWVL-NEXT: vpmovdb %ymm0, %xmm0544; AVX512BWVL-NEXT: vzeroupper545; AVX512BWVL-NEXT: retq546;547; AVX512VBMIVL-LABEL: trunc_v8i32_to_v8i8_return_v2i64:548; AVX512VBMIVL: # %bb.0:549; AVX512VBMIVL-NEXT: vpmovdb %ymm0, %xmm0550; AVX512VBMIVL-NEXT: vzeroupper551; AVX512VBMIVL-NEXT: retq552 %truncated.vec = trunc <8 x i32> %vec to <8 x i8>553 %bc = bitcast <8 x i8> %truncated.vec to i64554 %result = insertelement <2 x i64> zeroinitializer, i64 %bc, i32 0555 ret <2 x i64> %result556}557 558define <16 x i8> @trunc_v8i32_to_v8i8_with_zext_return_v16i8(<8 x i32> %vec) nounwind {559; AVX1-LABEL: trunc_v8i32_to_v8i8_with_zext_return_v16i8:560; AVX1: # %bb.0:561; AVX1-NEXT: vandps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm0562; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1563; AVX1-NEXT: vpackusdw %xmm1, %xmm0, %xmm0564; AVX1-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,2,4,6,8,10,12,14],zero,zero,zero,zero,zero,zero,zero,zero565; AVX1-NEXT: vzeroupper566; AVX1-NEXT: retq567;568; AVX2-LABEL: trunc_v8i32_to_v8i8_with_zext_return_v16i8:569; AVX2: # %bb.0:570; AVX2-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,1,4,5,8,9,12,13,u,u,u,u,u,u,u,u,16,17,20,21,24,25,28,29,u,u,u,u,u,u,u,u]571; AVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,2,2,3]572; AVX2-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,2,4,6,8,10,12,14],zero,zero,zero,zero,zero,zero,zero,zero573; AVX2-NEXT: vzeroupper574; AVX2-NEXT: retq575;576; AVX512F-LABEL: trunc_v8i32_to_v8i8_with_zext_return_v16i8:577; AVX512F: # %bb.0:578; AVX512F-NEXT: vmovdqa %ymm0, %ymm0579; AVX512F-NEXT: vpmovdb %zmm0, %xmm0580; AVX512F-NEXT: vzeroupper581; AVX512F-NEXT: retq582;583; AVX512VL-LABEL: trunc_v8i32_to_v8i8_with_zext_return_v16i8:584; AVX512VL: # %bb.0:585; AVX512VL-NEXT: vpmovdb %ymm0, %xmm0586; AVX512VL-NEXT: vzeroupper587; AVX512VL-NEXT: retq588;589; AVX512BW-LABEL: trunc_v8i32_to_v8i8_with_zext_return_v16i8:590; AVX512BW: # %bb.0:591; AVX512BW-NEXT: vmovdqa %ymm0, %ymm0592; AVX512BW-NEXT: vpmovdb %zmm0, %xmm0593; AVX512BW-NEXT: vzeroupper594; AVX512BW-NEXT: retq595;596; AVX512BWVL-LABEL: trunc_v8i32_to_v8i8_with_zext_return_v16i8:597; AVX512BWVL: # %bb.0:598; AVX512BWVL-NEXT: vpmovdb %ymm0, %xmm0599; AVX512BWVL-NEXT: vzeroupper600; AVX512BWVL-NEXT: retq601;602; AVX512VBMIVL-LABEL: trunc_v8i32_to_v8i8_with_zext_return_v16i8:603; AVX512VBMIVL: # %bb.0:604; AVX512VBMIVL-NEXT: vpmovdb %ymm0, %xmm0605; AVX512VBMIVL-NEXT: vzeroupper606; AVX512VBMIVL-NEXT: retq607 %truncated = trunc <8 x i32> %vec to <8 x i8>608 %truncated.ext = zext <8 x i8> %truncated to <8 x i16>609 %bc = bitcast <8 x i16> %truncated.ext to <16 x i8>610 %result = shufflevector <16 x i8> %bc, <16 x i8> zeroinitializer, <16 x i32> <i32 0, i32 2, i32 4, i32 6, i32 8, i32 10, i32 12, i32 14, i32 1, i32 3, i32 5, i32 7, i32 9, i32 11, i32 13, i32 15>611 ret <16 x i8> %result612}613 614define <16 x i8> @trunc_v8i32_to_v8i8_via_v8i16_return_v16i8(<8 x i32> %vec) nounwind {615; AVX1-LABEL: trunc_v8i32_to_v8i8_via_v8i16_return_v16i8:616; AVX1: # %bb.0:617; AVX1-NEXT: vandps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm0618; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1619; AVX1-NEXT: vpackusdw %xmm1, %xmm0, %xmm0620; AVX1-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,2,4,6,8,10,12,14],zero,zero,zero,zero,zero,zero,zero,zero621; AVX1-NEXT: vzeroupper622; AVX1-NEXT: retq623;624; AVX2-LABEL: trunc_v8i32_to_v8i8_via_v8i16_return_v16i8:625; AVX2: # %bb.0:626; AVX2-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,1,4,5,8,9,12,13,u,u,u,u,u,u,u,u,16,17,20,21,24,25,28,29,u,u,u,u,u,u,u,u]627; AVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,2,2,3]628; AVX2-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,2,4,6,8,10,12,14],zero,zero,zero,zero,zero,zero,zero,zero629; AVX2-NEXT: vzeroupper630; AVX2-NEXT: retq631;632; AVX512F-LABEL: trunc_v8i32_to_v8i8_via_v8i16_return_v16i8:633; AVX512F: # %bb.0:634; AVX512F-NEXT: vmovdqa %ymm0, %ymm0635; AVX512F-NEXT: vpmovdb %zmm0, %xmm0636; AVX512F-NEXT: vzeroupper637; AVX512F-NEXT: retq638;639; AVX512VL-LABEL: trunc_v8i32_to_v8i8_via_v8i16_return_v16i8:640; AVX512VL: # %bb.0:641; AVX512VL-NEXT: vpmovdb %ymm0, %xmm0642; AVX512VL-NEXT: vzeroupper643; AVX512VL-NEXT: retq644;645; AVX512BW-LABEL: trunc_v8i32_to_v8i8_via_v8i16_return_v16i8:646; AVX512BW: # %bb.0:647; AVX512BW-NEXT: vmovdqa %ymm0, %ymm0648; AVX512BW-NEXT: vpmovdb %zmm0, %xmm0649; AVX512BW-NEXT: vzeroupper650; AVX512BW-NEXT: retq651;652; AVX512BWVL-LABEL: trunc_v8i32_to_v8i8_via_v8i16_return_v16i8:653; AVX512BWVL: # %bb.0:654; AVX512BWVL-NEXT: vpmovdb %ymm0, %xmm0655; AVX512BWVL-NEXT: vzeroupper656; AVX512BWVL-NEXT: retq657;658; AVX512VBMIVL-LABEL: trunc_v8i32_to_v8i8_via_v8i16_return_v16i8:659; AVX512VBMIVL: # %bb.0:660; AVX512VBMIVL-NEXT: vpmovdb %ymm0, %xmm0661; AVX512VBMIVL-NEXT: vzeroupper662; AVX512VBMIVL-NEXT: retq663 %truncated = trunc <8 x i32> %vec to <8 x i16>664 %bc = bitcast <8 x i16> %truncated to <16 x i8>665 %result = shufflevector <16 x i8> %bc, <16 x i8> zeroinitializer, <16 x i32> <i32 0, i32 2, i32 4, i32 6, i32 8, i32 10, i32 12, i32 14, i32 17, i32 20, i32 24, i32 22, i32 31, i32 28, i32 28, i32 29>666 ret <16 x i8> %result667}668 669define <16 x i8> @trunc_v8i32_to_v8i8_return_v16i8(<8 x i32> %vec) nounwind {670; AVX1-LABEL: trunc_v8i32_to_v8i8_return_v16i8:671; AVX1: # %bb.0:672; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1673; AVX1-NEXT: vmovd {{.*#+}} xmm2 = [0,4,8,12,0,0,0,0,0,0,0,0,0,0,0,0]674; AVX1-NEXT: vpshufb %xmm2, %xmm1, %xmm1675; AVX1-NEXT: vpshufb %xmm2, %xmm0, %xmm0676; AVX1-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]677; AVX1-NEXT: vmovq {{.*#+}} xmm0 = xmm0[0],zero678; AVX1-NEXT: vzeroupper679; AVX1-NEXT: retq680;681; AVX2-LABEL: trunc_v8i32_to_v8i8_return_v16i8:682; AVX2: # %bb.0:683; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1684; AVX2-NEXT: vmovd {{.*#+}} xmm2 = [0,4,8,12,0,0,0,0,0,0,0,0,0,0,0,0]685; AVX2-NEXT: vpshufb %xmm2, %xmm1, %xmm1686; AVX2-NEXT: vpshufb %xmm2, %xmm0, %xmm0687; AVX2-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]688; AVX2-NEXT: vmovq {{.*#+}} xmm0 = xmm0[0],zero689; AVX2-NEXT: vzeroupper690; AVX2-NEXT: retq691;692; AVX512F-LABEL: trunc_v8i32_to_v8i8_return_v16i8:693; AVX512F: # %bb.0:694; AVX512F-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0695; AVX512F-NEXT: vpmovdb %zmm0, %xmm0696; AVX512F-NEXT: vmovq {{.*#+}} xmm0 = xmm0[0],zero697; AVX512F-NEXT: vzeroupper698; AVX512F-NEXT: retq699;700; AVX512VL-LABEL: trunc_v8i32_to_v8i8_return_v16i8:701; AVX512VL: # %bb.0:702; AVX512VL-NEXT: vpmovdb %ymm0, %xmm0703; AVX512VL-NEXT: vzeroupper704; AVX512VL-NEXT: retq705;706; AVX512BW-LABEL: trunc_v8i32_to_v8i8_return_v16i8:707; AVX512BW: # %bb.0:708; AVX512BW-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0709; AVX512BW-NEXT: vpmovdb %zmm0, %xmm0710; AVX512BW-NEXT: vmovq {{.*#+}} xmm0 = xmm0[0],zero711; AVX512BW-NEXT: vzeroupper712; AVX512BW-NEXT: retq713;714; AVX512BWVL-LABEL: trunc_v8i32_to_v8i8_return_v16i8:715; AVX512BWVL: # %bb.0:716; AVX512BWVL-NEXT: vpmovdb %ymm0, %xmm0717; AVX512BWVL-NEXT: vzeroupper718; AVX512BWVL-NEXT: retq719;720; AVX512VBMIVL-LABEL: trunc_v8i32_to_v8i8_return_v16i8:721; AVX512VBMIVL: # %bb.0:722; AVX512VBMIVL-NEXT: vpmovdb %ymm0, %xmm0723; AVX512VBMIVL-NEXT: vzeroupper724; AVX512VBMIVL-NEXT: retq725 %truncated = trunc <8 x i32> %vec to <8 x i8>726 %result = shufflevector <8 x i8> %truncated, <8 x i8> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>727 ret <16 x i8> %result728}729 730define <2 x i64> @trunc_v4i64_to_v4i16_return_v2i64(<4 x i64> %vec) nounwind {731; IR generated from:732; return (__m128i) {(long long)__builtin_convertvector((__v4di)x, __v4hi), 0};733; AVX1-LABEL: trunc_v4i64_to_v4i16_return_v2i64:734; AVX1: # %bb.0:735; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1736; AVX1-NEXT: vpxor %xmm2, %xmm2, %xmm2737; AVX1-NEXT: vpblendw {{.*#+}} xmm1 = xmm1[0],xmm2[1,2,3],xmm1[4],xmm2[5,6,7]738; AVX1-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0],xmm2[1,2,3],xmm0[4],xmm2[5,6,7]739; AVX1-NEXT: vpackusdw %xmm1, %xmm0, %xmm0740; AVX1-NEXT: vpackusdw %xmm2, %xmm0, %xmm0741; AVX1-NEXT: vzeroupper742; AVX1-NEXT: retq743;744; AVX2-LABEL: trunc_v4i64_to_v4i16_return_v2i64:745; AVX2: # %bb.0:746; AVX2-NEXT: vpxor %xmm1, %xmm1, %xmm1747; AVX2-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm1[1,2,3],ymm0[4],ymm1[5,6,7],ymm0[8],ymm1[9,10,11],ymm0[12],ymm1[13,14,15]748; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1749; AVX2-NEXT: vpackusdw %xmm1, %xmm0, %xmm0750; AVX2-NEXT: vpxor %xmm1, %xmm1, %xmm1751; AVX2-NEXT: vpackusdw %xmm1, %xmm0, %xmm0752; AVX2-NEXT: vzeroupper753; AVX2-NEXT: retq754;755; AVX512F-LABEL: trunc_v4i64_to_v4i16_return_v2i64:756; AVX512F: # %bb.0:757; AVX512F-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0758; AVX512F-NEXT: vpmovqw %zmm0, %xmm0759; AVX512F-NEXT: vmovq {{.*#+}} xmm0 = xmm0[0],zero760; AVX512F-NEXT: vzeroupper761; AVX512F-NEXT: retq762;763; AVX512VL-LABEL: trunc_v4i64_to_v4i16_return_v2i64:764; AVX512VL: # %bb.0:765; AVX512VL-NEXT: vpmovqw %ymm0, %xmm0766; AVX512VL-NEXT: vzeroupper767; AVX512VL-NEXT: retq768;769; AVX512BW-LABEL: trunc_v4i64_to_v4i16_return_v2i64:770; AVX512BW: # %bb.0:771; AVX512BW-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0772; AVX512BW-NEXT: vpmovqw %zmm0, %xmm0773; AVX512BW-NEXT: vmovq {{.*#+}} xmm0 = xmm0[0],zero774; AVX512BW-NEXT: vzeroupper775; AVX512BW-NEXT: retq776;777; AVX512BWVL-LABEL: trunc_v4i64_to_v4i16_return_v2i64:778; AVX512BWVL: # %bb.0:779; AVX512BWVL-NEXT: vpmovqw %ymm0, %xmm0780; AVX512BWVL-NEXT: vzeroupper781; AVX512BWVL-NEXT: retq782;783; AVX512VBMIVL-LABEL: trunc_v4i64_to_v4i16_return_v2i64:784; AVX512VBMIVL: # %bb.0:785; AVX512VBMIVL-NEXT: vpmovqw %ymm0, %xmm0786; AVX512VBMIVL-NEXT: vzeroupper787; AVX512VBMIVL-NEXT: retq788 %truncated = trunc <4 x i64> %vec to <4 x i16>789 %bc = bitcast <4 x i16> %truncated to i64790 %result = insertelement <2 x i64> zeroinitializer, i64 %bc, i32 0791 ret <2 x i64> %result792}793 794define <8 x i16> @trunc_v4i64_to_v4i16_with_zext_return_v8i16(<4 x i64> %vec) nounwind {795; AVX1-LABEL: trunc_v4i64_to_v4i16_with_zext_return_v8i16:796; AVX1: # %bb.0:797; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1798; AVX1-NEXT: vshufps {{.*#+}} xmm0 = xmm0[0,2],xmm1[0,2]799; AVX1-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,1,4,5,8,9,12,13],zero,zero,zero,zero,zero,zero,zero,zero800; AVX1-NEXT: vzeroupper801; AVX1-NEXT: retq802;803; AVX2-SLOW-LABEL: trunc_v4i64_to_v4i16_with_zext_return_v8i16:804; AVX2-SLOW: # %bb.0:805; AVX2-SLOW-NEXT: vextractf128 $1, %ymm0, %xmm1806; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm0 = xmm0[0,2],xmm1[0,2]807; AVX2-SLOW-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,1,4,5,8,9,12,13],zero,zero,zero,zero,zero,zero,zero,zero808; AVX2-SLOW-NEXT: vzeroupper809; AVX2-SLOW-NEXT: retq810;811; AVX2-FAST-ALL-LABEL: trunc_v4i64_to_v4i16_with_zext_return_v8i16:812; AVX2-FAST-ALL: # %bb.0:813; AVX2-FAST-ALL-NEXT: vpmovsxbd {{.*#+}} ymm1 = [0,2,4,6,0,0,0,0]814; AVX2-FAST-ALL-NEXT: vpermd %ymm0, %ymm1, %ymm0815; AVX2-FAST-ALL-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,1,4,5,8,9,12,13],zero,zero,zero,zero,zero,zero,zero,zero816; AVX2-FAST-ALL-NEXT: vzeroupper817; AVX2-FAST-ALL-NEXT: retq818;819; AVX2-FAST-PERLANE-LABEL: trunc_v4i64_to_v4i16_with_zext_return_v8i16:820; AVX2-FAST-PERLANE: # %bb.0:821; AVX2-FAST-PERLANE-NEXT: vextractf128 $1, %ymm0, %xmm1822; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm0 = xmm0[0,2],xmm1[0,2]823; AVX2-FAST-PERLANE-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,1,4,5,8,9,12,13],zero,zero,zero,zero,zero,zero,zero,zero824; AVX2-FAST-PERLANE-NEXT: vzeroupper825; AVX2-FAST-PERLANE-NEXT: retq826;827; AVX512F-LABEL: trunc_v4i64_to_v4i16_with_zext_return_v8i16:828; AVX512F: # %bb.0:829; AVX512F-NEXT: vmovdqa %ymm0, %ymm0830; AVX512F-NEXT: vpmovqw %zmm0, %xmm0831; AVX512F-NEXT: vzeroupper832; AVX512F-NEXT: retq833;834; AVX512VL-LABEL: trunc_v4i64_to_v4i16_with_zext_return_v8i16:835; AVX512VL: # %bb.0:836; AVX512VL-NEXT: vpmovqw %ymm0, %xmm0837; AVX512VL-NEXT: vzeroupper838; AVX512VL-NEXT: retq839;840; AVX512BW-LABEL: trunc_v4i64_to_v4i16_with_zext_return_v8i16:841; AVX512BW: # %bb.0:842; AVX512BW-NEXT: vmovdqa %ymm0, %ymm0843; AVX512BW-NEXT: vpmovqw %zmm0, %xmm0844; AVX512BW-NEXT: vzeroupper845; AVX512BW-NEXT: retq846;847; AVX512BWVL-LABEL: trunc_v4i64_to_v4i16_with_zext_return_v8i16:848; AVX512BWVL: # %bb.0:849; AVX512BWVL-NEXT: vpmovqw %ymm0, %xmm0850; AVX512BWVL-NEXT: vzeroupper851; AVX512BWVL-NEXT: retq852;853; AVX512VBMIVL-LABEL: trunc_v4i64_to_v4i16_with_zext_return_v8i16:854; AVX512VBMIVL: # %bb.0:855; AVX512VBMIVL-NEXT: vpmovqw %ymm0, %xmm0856; AVX512VBMIVL-NEXT: vzeroupper857; AVX512VBMIVL-NEXT: retq858 %truncated = trunc <4 x i64> %vec to <4 x i16>859 %truncated.ext = zext <4 x i16> %truncated to <4 x i32>860 %bc = bitcast <4 x i32> %truncated.ext to <8 x i16>861 %result = shufflevector <8 x i16> %bc, <8 x i16> zeroinitializer, <8 x i32> <i32 0, i32 2, i32 4, i32 6, i32 1, i32 3, i32 5, i32 7>862 ret <8 x i16> %result863}864 865define <8 x i16> @trunc_v4i64_to_v4i16_via_v4i32_return_v8i16(<4 x i64> %vec) nounwind {866; AVX1-LABEL: trunc_v4i64_to_v4i16_via_v4i32_return_v8i16:867; AVX1: # %bb.0:868; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1869; AVX1-NEXT: vshufps {{.*#+}} xmm0 = xmm0[0,2],xmm1[0,2]870; AVX1-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,1,4,5,8,9,12,13],zero,zero,zero,zero,zero,zero,zero,zero871; AVX1-NEXT: vzeroupper872; AVX1-NEXT: retq873;874; AVX2-SLOW-LABEL: trunc_v4i64_to_v4i16_via_v4i32_return_v8i16:875; AVX2-SLOW: # %bb.0:876; AVX2-SLOW-NEXT: vextractf128 $1, %ymm0, %xmm1877; AVX2-SLOW-NEXT: vshufps {{.*#+}} xmm0 = xmm0[0,2],xmm1[0,2]878; AVX2-SLOW-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,1,4,5,8,9,12,13],zero,zero,zero,zero,zero,zero,zero,zero879; AVX2-SLOW-NEXT: vzeroupper880; AVX2-SLOW-NEXT: retq881;882; AVX2-FAST-ALL-LABEL: trunc_v4i64_to_v4i16_via_v4i32_return_v8i16:883; AVX2-FAST-ALL: # %bb.0:884; AVX2-FAST-ALL-NEXT: vpmovsxbd {{.*#+}} ymm1 = [0,2,4,6,0,0,0,0]885; AVX2-FAST-ALL-NEXT: vpermd %ymm0, %ymm1, %ymm0886; AVX2-FAST-ALL-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,1,4,5,8,9,12,13],zero,zero,zero,zero,zero,zero,zero,zero887; AVX2-FAST-ALL-NEXT: vzeroupper888; AVX2-FAST-ALL-NEXT: retq889;890; AVX2-FAST-PERLANE-LABEL: trunc_v4i64_to_v4i16_via_v4i32_return_v8i16:891; AVX2-FAST-PERLANE: # %bb.0:892; AVX2-FAST-PERLANE-NEXT: vextractf128 $1, %ymm0, %xmm1893; AVX2-FAST-PERLANE-NEXT: vshufps {{.*#+}} xmm0 = xmm0[0,2],xmm1[0,2]894; AVX2-FAST-PERLANE-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,1,4,5,8,9,12,13],zero,zero,zero,zero,zero,zero,zero,zero895; AVX2-FAST-PERLANE-NEXT: vzeroupper896; AVX2-FAST-PERLANE-NEXT: retq897;898; AVX512F-LABEL: trunc_v4i64_to_v4i16_via_v4i32_return_v8i16:899; AVX512F: # %bb.0:900; AVX512F-NEXT: vmovdqa %ymm0, %ymm0901; AVX512F-NEXT: vpmovqw %zmm0, %xmm0902; AVX512F-NEXT: vzeroupper903; AVX512F-NEXT: retq904;905; AVX512VL-LABEL: trunc_v4i64_to_v4i16_via_v4i32_return_v8i16:906; AVX512VL: # %bb.0:907; AVX512VL-NEXT: vpmovqw %ymm0, %xmm0908; AVX512VL-NEXT: vzeroupper909; AVX512VL-NEXT: retq910;911; AVX512BW-LABEL: trunc_v4i64_to_v4i16_via_v4i32_return_v8i16:912; AVX512BW: # %bb.0:913; AVX512BW-NEXT: vmovdqa %ymm0, %ymm0914; AVX512BW-NEXT: vpmovqw %zmm0, %xmm0915; AVX512BW-NEXT: vzeroupper916; AVX512BW-NEXT: retq917;918; AVX512BWVL-LABEL: trunc_v4i64_to_v4i16_via_v4i32_return_v8i16:919; AVX512BWVL: # %bb.0:920; AVX512BWVL-NEXT: vpmovqw %ymm0, %xmm0921; AVX512BWVL-NEXT: vzeroupper922; AVX512BWVL-NEXT: retq923;924; AVX512VBMIVL-LABEL: trunc_v4i64_to_v4i16_via_v4i32_return_v8i16:925; AVX512VBMIVL: # %bb.0:926; AVX512VBMIVL-NEXT: vpmovqw %ymm0, %xmm0927; AVX512VBMIVL-NEXT: vzeroupper928; AVX512VBMIVL-NEXT: retq929 %truncated = trunc <4 x i64> %vec to <4 x i32>930 %bc = bitcast <4 x i32> %truncated to <8 x i16>931 %result = shufflevector <8 x i16> %bc, <8 x i16> zeroinitializer, <8 x i32> <i32 0, i32 2, i32 4, i32 6, i32 8, i32 8, i32 undef, i32 13>932 ret <8 x i16> %result933}934 935define <8 x i16> @trunc_v4i64_to_v4i16_return_v8i16(<4 x i64> %vec) nounwind {936; AVX1-LABEL: trunc_v4i64_to_v4i16_return_v8i16:937; AVX1: # %bb.0:938; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1939; AVX1-NEXT: vpxor %xmm2, %xmm2, %xmm2940; AVX1-NEXT: vpblendw {{.*#+}} xmm1 = xmm1[0],xmm2[1,2,3],xmm1[4],xmm2[5,6,7]941; AVX1-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0],xmm2[1,2,3],xmm0[4],xmm2[5,6,7]942; AVX1-NEXT: vpackusdw %xmm1, %xmm0, %xmm0943; AVX1-NEXT: vpackusdw %xmm2, %xmm0, %xmm0944; AVX1-NEXT: vzeroupper945; AVX1-NEXT: retq946;947; AVX2-LABEL: trunc_v4i64_to_v4i16_return_v8i16:948; AVX2: # %bb.0:949; AVX2-NEXT: vpxor %xmm1, %xmm1, %xmm1950; AVX2-NEXT: vpblendw {{.*#+}} ymm0 = ymm0[0],ymm1[1,2,3],ymm0[4],ymm1[5,6,7],ymm0[8],ymm1[9,10,11],ymm0[12],ymm1[13,14,15]951; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1952; AVX2-NEXT: vpackusdw %xmm1, %xmm0, %xmm0953; AVX2-NEXT: vpxor %xmm1, %xmm1, %xmm1954; AVX2-NEXT: vpackusdw %xmm1, %xmm0, %xmm0955; AVX2-NEXT: vzeroupper956; AVX2-NEXT: retq957;958; AVX512F-LABEL: trunc_v4i64_to_v4i16_return_v8i16:959; AVX512F: # %bb.0:960; AVX512F-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0961; AVX512F-NEXT: vpmovqw %zmm0, %xmm0962; AVX512F-NEXT: vmovq {{.*#+}} xmm0 = xmm0[0],zero963; AVX512F-NEXT: vzeroupper964; AVX512F-NEXT: retq965;966; AVX512VL-LABEL: trunc_v4i64_to_v4i16_return_v8i16:967; AVX512VL: # %bb.0:968; AVX512VL-NEXT: vpmovqw %ymm0, %xmm0969; AVX512VL-NEXT: vzeroupper970; AVX512VL-NEXT: retq971;972; AVX512BW-LABEL: trunc_v4i64_to_v4i16_return_v8i16:973; AVX512BW: # %bb.0:974; AVX512BW-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0975; AVX512BW-NEXT: vpmovqw %zmm0, %xmm0976; AVX512BW-NEXT: vmovq {{.*#+}} xmm0 = xmm0[0],zero977; AVX512BW-NEXT: vzeroupper978; AVX512BW-NEXT: retq979;980; AVX512BWVL-LABEL: trunc_v4i64_to_v4i16_return_v8i16:981; AVX512BWVL: # %bb.0:982; AVX512BWVL-NEXT: vpmovqw %ymm0, %xmm0983; AVX512BWVL-NEXT: vzeroupper984; AVX512BWVL-NEXT: retq985;986; AVX512VBMIVL-LABEL: trunc_v4i64_to_v4i16_return_v8i16:987; AVX512VBMIVL: # %bb.0:988; AVX512VBMIVL-NEXT: vpmovqw %ymm0, %xmm0989; AVX512VBMIVL-NEXT: vzeroupper990; AVX512VBMIVL-NEXT: retq991 %truncated = trunc <4 x i64> %vec to <4 x i16>992 %result = shufflevector <4 x i16> %truncated, <4 x i16> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>993 ret <8 x i16> %result994}995 996define <16 x i8> @trunc_v4i64_to_v4i8_return_v16i8(<4 x i64> %vec) nounwind {997; AVX1-LABEL: trunc_v4i64_to_v4i8_return_v16i8:998; AVX1: # %bb.0:999; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm11000; AVX1-NEXT: vmovd {{.*#+}} xmm2 = [0,8,0,0,0,0,0,0,0,0,0,0,0,0,0,0]1001; AVX1-NEXT: vpshufb %xmm2, %xmm1, %xmm11002; AVX1-NEXT: vpshufb %xmm2, %xmm0, %xmm01003; AVX1-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]1004; AVX1-NEXT: vpxor %xmm1, %xmm1, %xmm11005; AVX1-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3,4,5,6,7]1006; AVX1-NEXT: vzeroupper1007; AVX1-NEXT: retq1008;1009; AVX2-LABEL: trunc_v4i64_to_v4i8_return_v16i8:1010; AVX2: # %bb.0:1011; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm11012; AVX2-NEXT: vpbroadcastw {{.*#+}} xmm2 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]1013; AVX2-NEXT: vpshufb %xmm2, %xmm1, %xmm11014; AVX2-NEXT: vpshufb %xmm2, %xmm0, %xmm01015; AVX2-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]1016; AVX2-NEXT: vpxor %xmm1, %xmm1, %xmm11017; AVX2-NEXT: vpblendd {{.*#+}} xmm0 = xmm0[0],xmm1[1,2,3]1018; AVX2-NEXT: vzeroupper1019; AVX2-NEXT: retq1020;1021; AVX512F-LABEL: trunc_v4i64_to_v4i8_return_v16i8:1022; AVX512F: # %bb.0:1023; AVX512F-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm01024; AVX512F-NEXT: vpmovqb %zmm0, %xmm01025; AVX512F-NEXT: vpxor %xmm1, %xmm1, %xmm11026; AVX512F-NEXT: vpblendd {{.*#+}} xmm0 = xmm0[0],xmm1[1,2,3]1027; AVX512F-NEXT: vzeroupper1028; AVX512F-NEXT: retq1029;1030; AVX512VL-LABEL: trunc_v4i64_to_v4i8_return_v16i8:1031; AVX512VL: # %bb.0:1032; AVX512VL-NEXT: vpmovqb %ymm0, %xmm01033; AVX512VL-NEXT: vzeroupper1034; AVX512VL-NEXT: retq1035;1036; AVX512BW-LABEL: trunc_v4i64_to_v4i8_return_v16i8:1037; AVX512BW: # %bb.0:1038; AVX512BW-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm01039; AVX512BW-NEXT: vpmovqb %zmm0, %xmm01040; AVX512BW-NEXT: vpxor %xmm1, %xmm1, %xmm11041; AVX512BW-NEXT: vpblendd {{.*#+}} xmm0 = xmm0[0],xmm1[1,2,3]1042; AVX512BW-NEXT: vzeroupper1043; AVX512BW-NEXT: retq1044;1045; AVX512BWVL-LABEL: trunc_v4i64_to_v4i8_return_v16i8:1046; AVX512BWVL: # %bb.0:1047; AVX512BWVL-NEXT: vpmovqb %ymm0, %xmm01048; AVX512BWVL-NEXT: vzeroupper1049; AVX512BWVL-NEXT: retq1050;1051; AVX512VBMIVL-LABEL: trunc_v4i64_to_v4i8_return_v16i8:1052; AVX512VBMIVL: # %bb.0:1053; AVX512VBMIVL-NEXT: vpmovqb %ymm0, %xmm01054; AVX512VBMIVL-NEXT: vzeroupper1055; AVX512VBMIVL-NEXT: retq1056 %truncated = trunc <4 x i64> %vec to <4 x i8>1057 %result = shufflevector <4 x i8> %truncated, <4 x i8> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 5, i32 5, i32 undef, i32 7>1058 ret <16 x i8> %result1059}1060 1061define void @shuffle_v16i16_to_v4i16(ptr %L, ptr %S) nounwind {1062; AVX-LABEL: shuffle_v16i16_to_v4i16:1063; AVX: # %bb.0:1064; AVX-NEXT: vpxor %xmm0, %xmm0, %xmm01065; AVX-NEXT: vpblendw {{.*#+}} xmm1 = mem[0],xmm0[1,2,3],mem[4],xmm0[5,6,7]1066; AVX-NEXT: vpblendw {{.*#+}} xmm0 = mem[0],xmm0[1,2,3],mem[4],xmm0[5,6,7]1067; AVX-NEXT: vpackusdw %xmm1, %xmm0, %xmm01068; AVX-NEXT: vpackusdw %xmm0, %xmm0, %xmm01069; AVX-NEXT: vmovq %xmm0, (%rsi)1070; AVX-NEXT: retq1071;1072; AVX512F-LABEL: shuffle_v16i16_to_v4i16:1073; AVX512F: # %bb.0:1074; AVX512F-NEXT: vmovdqa (%rdi), %ymm01075; AVX512F-NEXT: vpmovqw %zmm0, %xmm01076; AVX512F-NEXT: vmovq %xmm0, (%rsi)1077; AVX512F-NEXT: vzeroupper1078; AVX512F-NEXT: retq1079;1080; AVX512VL-LABEL: shuffle_v16i16_to_v4i16:1081; AVX512VL: # %bb.0:1082; AVX512VL-NEXT: vmovdqa (%rdi), %ymm01083; AVX512VL-NEXT: vpmovqw %ymm0, (%rsi)1084; AVX512VL-NEXT: vzeroupper1085; AVX512VL-NEXT: retq1086;1087; AVX512BW-LABEL: shuffle_v16i16_to_v4i16:1088; AVX512BW: # %bb.0:1089; AVX512BW-NEXT: vmovdqa (%rdi), %ymm01090; AVX512BW-NEXT: vpmovqw %zmm0, %xmm01091; AVX512BW-NEXT: vmovq %xmm0, (%rsi)1092; AVX512BW-NEXT: vzeroupper1093; AVX512BW-NEXT: retq1094;1095; AVX512BWVL-LABEL: shuffle_v16i16_to_v4i16:1096; AVX512BWVL: # %bb.0:1097; AVX512BWVL-NEXT: vmovdqa (%rdi), %ymm01098; AVX512BWVL-NEXT: vpmovqw %ymm0, (%rsi)1099; AVX512BWVL-NEXT: vzeroupper1100; AVX512BWVL-NEXT: retq1101;1102; AVX512VBMIVL-LABEL: shuffle_v16i16_to_v4i16:1103; AVX512VBMIVL: # %bb.0:1104; AVX512VBMIVL-NEXT: vmovdqa (%rdi), %ymm01105; AVX512VBMIVL-NEXT: vpmovqw %ymm0, (%rsi)1106; AVX512VBMIVL-NEXT: vzeroupper1107; AVX512VBMIVL-NEXT: retq1108 %vec = load <16 x i16>, ptr %L1109 %strided.vec = shufflevector <16 x i16> %vec, <16 x i16> undef, <4 x i32> <i32 0, i32 4, i32 8, i32 12>1110 store <4 x i16> %strided.vec, ptr %S1111 ret void1112}1113 1114define void @trunc_v4i64_to_v4i16(ptr %L, ptr %S) nounwind {1115; AVX-LABEL: trunc_v4i64_to_v4i16:1116; AVX: # %bb.0:1117; AVX-NEXT: vpxor %xmm0, %xmm0, %xmm01118; AVX-NEXT: vpblendw {{.*#+}} xmm1 = mem[0],xmm0[1,2,3],mem[4],xmm0[5,6,7]1119; AVX-NEXT: vpblendw {{.*#+}} xmm0 = mem[0],xmm0[1,2,3],mem[4],xmm0[5,6,7]1120; AVX-NEXT: vpackusdw %xmm1, %xmm0, %xmm01121; AVX-NEXT: vpackusdw %xmm0, %xmm0, %xmm01122; AVX-NEXT: vmovq %xmm0, (%rsi)1123; AVX-NEXT: retq1124;1125; AVX512F-LABEL: trunc_v4i64_to_v4i16:1126; AVX512F: # %bb.0:1127; AVX512F-NEXT: vmovdqa (%rdi), %ymm01128; AVX512F-NEXT: vpmovqw %zmm0, %xmm01129; AVX512F-NEXT: vmovq %xmm0, (%rsi)1130; AVX512F-NEXT: vzeroupper1131; AVX512F-NEXT: retq1132;1133; AVX512VL-LABEL: trunc_v4i64_to_v4i16:1134; AVX512VL: # %bb.0:1135; AVX512VL-NEXT: vmovdqa (%rdi), %ymm01136; AVX512VL-NEXT: vpmovqw %ymm0, (%rsi)1137; AVX512VL-NEXT: vzeroupper1138; AVX512VL-NEXT: retq1139;1140; AVX512BW-LABEL: trunc_v4i64_to_v4i16:1141; AVX512BW: # %bb.0:1142; AVX512BW-NEXT: vmovdqa (%rdi), %ymm01143; AVX512BW-NEXT: vpmovqw %zmm0, %xmm01144; AVX512BW-NEXT: vmovq %xmm0, (%rsi)1145; AVX512BW-NEXT: vzeroupper1146; AVX512BW-NEXT: retq1147;1148; AVX512BWVL-LABEL: trunc_v4i64_to_v4i16:1149; AVX512BWVL: # %bb.0:1150; AVX512BWVL-NEXT: vmovdqa (%rdi), %ymm01151; AVX512BWVL-NEXT: vpmovqw %ymm0, (%rsi)1152; AVX512BWVL-NEXT: vzeroupper1153; AVX512BWVL-NEXT: retq1154;1155; AVX512VBMIVL-LABEL: trunc_v4i64_to_v4i16:1156; AVX512VBMIVL: # %bb.0:1157; AVX512VBMIVL-NEXT: vmovdqa (%rdi), %ymm01158; AVX512VBMIVL-NEXT: vpmovqw %ymm0, (%rsi)1159; AVX512VBMIVL-NEXT: vzeroupper1160; AVX512VBMIVL-NEXT: retq1161 %vec = load <16 x i16>, ptr %L1162 %bc = bitcast <16 x i16> %vec to <4 x i64>1163 %strided.vec = trunc <4 x i64> %bc to <4 x i16>1164 store <4 x i16> %strided.vec, ptr %S1165 ret void1166}1167 1168define void @shuffle_v32i8_to_v4i8(ptr %L, ptr %S) nounwind {1169; AVX1-LABEL: shuffle_v32i8_to_v4i8:1170; AVX1: # %bb.0:1171; AVX1-NEXT: vmovdqa (%rdi), %xmm01172; AVX1-NEXT: vmovdqa 16(%rdi), %xmm11173; AVX1-NEXT: vmovd {{.*#+}} xmm2 = [0,8,0,0,0,0,0,0,0,0,0,0,0,0,0,0]1174; AVX1-NEXT: vpshufb %xmm2, %xmm1, %xmm11175; AVX1-NEXT: vpshufb %xmm2, %xmm0, %xmm01176; AVX1-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]1177; AVX1-NEXT: vmovd %xmm0, (%rsi)1178; AVX1-NEXT: retq1179;1180; AVX2-LABEL: shuffle_v32i8_to_v4i8:1181; AVX2: # %bb.0:1182; AVX2-NEXT: vmovdqa (%rdi), %xmm01183; AVX2-NEXT: vmovdqa 16(%rdi), %xmm11184; AVX2-NEXT: vpbroadcastw {{.*#+}} xmm2 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]1185; AVX2-NEXT: vpshufb %xmm2, %xmm1, %xmm11186; AVX2-NEXT: vpshufb %xmm2, %xmm0, %xmm01187; AVX2-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]1188; AVX2-NEXT: vmovd %xmm0, (%rsi)1189; AVX2-NEXT: retq1190;1191; AVX512F-LABEL: shuffle_v32i8_to_v4i8:1192; AVX512F: # %bb.0:1193; AVX512F-NEXT: vmovdqa (%rdi), %ymm01194; AVX512F-NEXT: vpmovqb %zmm0, %xmm01195; AVX512F-NEXT: vmovd %xmm0, (%rsi)1196; AVX512F-NEXT: vzeroupper1197; AVX512F-NEXT: retq1198;1199; AVX512VL-LABEL: shuffle_v32i8_to_v4i8:1200; AVX512VL: # %bb.0:1201; AVX512VL-NEXT: vmovdqa (%rdi), %ymm01202; AVX512VL-NEXT: vpmovqb %ymm0, (%rsi)1203; AVX512VL-NEXT: vzeroupper1204; AVX512VL-NEXT: retq1205;1206; AVX512BW-LABEL: shuffle_v32i8_to_v4i8:1207; AVX512BW: # %bb.0:1208; AVX512BW-NEXT: vmovdqa (%rdi), %ymm01209; AVX512BW-NEXT: vpmovqb %zmm0, %xmm01210; AVX512BW-NEXT: vmovd %xmm0, (%rsi)1211; AVX512BW-NEXT: vzeroupper1212; AVX512BW-NEXT: retq1213;1214; AVX512BWVL-LABEL: shuffle_v32i8_to_v4i8:1215; AVX512BWVL: # %bb.0:1216; AVX512BWVL-NEXT: vmovdqa (%rdi), %ymm01217; AVX512BWVL-NEXT: vpmovqb %ymm0, (%rsi)1218; AVX512BWVL-NEXT: vzeroupper1219; AVX512BWVL-NEXT: retq1220;1221; AVX512VBMIVL-LABEL: shuffle_v32i8_to_v4i8:1222; AVX512VBMIVL: # %bb.0:1223; AVX512VBMIVL-NEXT: vmovdqa (%rdi), %ymm01224; AVX512VBMIVL-NEXT: vpmovqb %ymm0, (%rsi)1225; AVX512VBMIVL-NEXT: vzeroupper1226; AVX512VBMIVL-NEXT: retq1227 %vec = load <32 x i8>, ptr %L1228 %strided.vec = shufflevector <32 x i8> %vec, <32 x i8> undef, <4 x i32> <i32 0, i32 8, i32 16, i32 24>1229 store <4 x i8> %strided.vec, ptr %S1230 ret void1231}1232 1233define void @trunc_v4i64_to_v4i8(ptr %L, ptr %S) nounwind {1234; AVX1-LABEL: trunc_v4i64_to_v4i8:1235; AVX1: # %bb.0:1236; AVX1-NEXT: vmovdqa (%rdi), %xmm01237; AVX1-NEXT: vmovdqa 16(%rdi), %xmm11238; AVX1-NEXT: vmovd {{.*#+}} xmm2 = [0,8,0,0,0,0,0,0,0,0,0,0,0,0,0,0]1239; AVX1-NEXT: vpshufb %xmm2, %xmm1, %xmm11240; AVX1-NEXT: vpshufb %xmm2, %xmm0, %xmm01241; AVX1-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]1242; AVX1-NEXT: vmovd %xmm0, (%rsi)1243; AVX1-NEXT: retq1244;1245; AVX2-LABEL: trunc_v4i64_to_v4i8:1246; AVX2: # %bb.0:1247; AVX2-NEXT: vmovdqa (%rdi), %xmm01248; AVX2-NEXT: vmovdqa 16(%rdi), %xmm11249; AVX2-NEXT: vpbroadcastw {{.*#+}} xmm2 = [0,8,0,8,0,8,0,8,0,8,0,8,0,8,0,8]1250; AVX2-NEXT: vpshufb %xmm2, %xmm1, %xmm11251; AVX2-NEXT: vpshufb %xmm2, %xmm0, %xmm01252; AVX2-NEXT: vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]1253; AVX2-NEXT: vmovd %xmm0, (%rsi)1254; AVX2-NEXT: retq1255;1256; AVX512F-LABEL: trunc_v4i64_to_v4i8:1257; AVX512F: # %bb.0:1258; AVX512F-NEXT: vmovdqa (%rdi), %ymm01259; AVX512F-NEXT: vpmovqb %zmm0, %xmm01260; AVX512F-NEXT: vmovd %xmm0, (%rsi)1261; AVX512F-NEXT: vzeroupper1262; AVX512F-NEXT: retq1263;1264; AVX512VL-LABEL: trunc_v4i64_to_v4i8:1265; AVX512VL: # %bb.0:1266; AVX512VL-NEXT: vmovdqa (%rdi), %ymm01267; AVX512VL-NEXT: vpmovqb %ymm0, (%rsi)1268; AVX512VL-NEXT: vzeroupper1269; AVX512VL-NEXT: retq1270;1271; AVX512BW-LABEL: trunc_v4i64_to_v4i8:1272; AVX512BW: # %bb.0:1273; AVX512BW-NEXT: vmovdqa (%rdi), %ymm01274; AVX512BW-NEXT: vpmovqb %zmm0, %xmm01275; AVX512BW-NEXT: vmovd %xmm0, (%rsi)1276; AVX512BW-NEXT: vzeroupper1277; AVX512BW-NEXT: retq1278;1279; AVX512BWVL-LABEL: trunc_v4i64_to_v4i8:1280; AVX512BWVL: # %bb.0:1281; AVX512BWVL-NEXT: vmovdqa (%rdi), %ymm01282; AVX512BWVL-NEXT: vpmovqb %ymm0, (%rsi)1283; AVX512BWVL-NEXT: vzeroupper1284; AVX512BWVL-NEXT: retq1285;1286; AVX512VBMIVL-LABEL: trunc_v4i64_to_v4i8:1287; AVX512VBMIVL: # %bb.0:1288; AVX512VBMIVL-NEXT: vmovdqa (%rdi), %ymm01289; AVX512VBMIVL-NEXT: vpmovqb %ymm0, (%rsi)1290; AVX512VBMIVL-NEXT: vzeroupper1291; AVX512VBMIVL-NEXT: retq1292 %vec = load <32 x i8>, ptr %L1293 %bc = bitcast <32 x i8> %vec to <4 x i64>1294 %strided.vec = trunc <4 x i64> %bc to <4 x i8>1295 store <4 x i8> %strided.vec, ptr %S1296 ret void1297}1298 1299; In this case not all elements are collected from the same source vector, so1300; the resulting BUILD_VECTOR should not be combined to a truncate.1301define <16 x i8> @negative(<32 x i8> %v, <32 x i8> %w) nounwind {1302; AVX1-LABEL: negative:1303; AVX1: # %bb.0:1304; AVX1-NEXT: vpshufb {{.*#+}} xmm2 = xmm0[u,2,4,6,8,10,12,14],zero,zero,zero,zero,zero,zero,zero,zero1305; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm01306; AVX1-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[u],zero,zero,zero,zero,zero,zero,zero,xmm0[0,2,4,6,8,10,12,14]1307; AVX1-NEXT: vpor %xmm2, %xmm0, %xmm01308; AVX1-NEXT: vpmovsxwq {{.*#+}} xmm2 = [18446744073709551360,18446744073709551615]1309; AVX1-NEXT: vpblendvb %xmm2, %xmm0, %xmm1, %xmm01310; AVX1-NEXT: vzeroupper1311; AVX1-NEXT: retq1312;1313; AVX2-LABEL: negative:1314; AVX2: # %bb.0:1315; AVX2-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[u,2,4,6,8,10,12,14,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,16,18,20,22,24,26,28,30]1316; AVX2-NEXT: vpmovsxwq {{.*#+}} ymm2 = [18446744073709551360,18446744073709551615,18446744073709551360,18446744073709551615]1317; AVX2-NEXT: vpblendvb %ymm2, %ymm0, %ymm1, %ymm01318; AVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,3,2,3]1319; AVX2-NEXT: # kill: def $xmm0 killed $xmm0 killed $ymm01320; AVX2-NEXT: vzeroupper1321; AVX2-NEXT: retq1322;1323; AVX512F-LABEL: negative:1324; AVX512F: # %bb.0:1325; AVX512F-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[u,2,4,6,8,10,12,14,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,16,18,20,22,24,26,28,30]1326; AVX512F-NEXT: vpmovsxwq {{.*#+}} ymm2 = [18446744073709551360,18446744073709551615,18446744073709551360,18446744073709551615]1327; AVX512F-NEXT: vpblendvb %ymm2, %ymm0, %ymm1, %ymm01328; AVX512F-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,3,2,3]1329; AVX512F-NEXT: # kill: def $xmm0 killed $xmm0 killed $ymm01330; AVX512F-NEXT: vzeroupper1331; AVX512F-NEXT: retq1332;1333; AVX512VL-LABEL: negative:1334; AVX512VL: # %bb.0:1335; AVX512VL-NEXT: vpshufb {{.*#+}} ymm0 = zero,ymm0[2,4,6,8,10,12,14,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,16,18,20,22,24,26,28,30]1336; AVX512VL-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [0,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,0,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255]1337; AVX512VL-NEXT: # ymm2 = mem[0,1,0,1]1338; AVX512VL-NEXT: vpternlogq {{.*#+}} ymm2 = (ymm1 & ~ymm2) | ymm01339; AVX512VL-NEXT: vpermq {{.*#+}} ymm0 = ymm2[0,3,2,3]1340; AVX512VL-NEXT: # kill: def $xmm0 killed $xmm0 killed $ymm01341; AVX512VL-NEXT: vzeroupper1342; AVX512VL-NEXT: retq1343;1344; AVX512BW-LABEL: negative:1345; AVX512BW: # %bb.0:1346; AVX512BW-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[u,2,4,6,8,10,12,14,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,16,18,20,22,24,26,28,30]1347; AVX512BW-NEXT: vpmovsxwq {{.*#+}} ymm2 = [18446744073709551360,18446744073709551615,18446744073709551360,18446744073709551615]1348; AVX512BW-NEXT: vpblendvb %ymm2, %ymm0, %ymm1, %ymm01349; AVX512BW-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,3,2,3]1350; AVX512BW-NEXT: # kill: def $xmm0 killed $xmm0 killed $ymm01351; AVX512BW-NEXT: vzeroupper1352; AVX512BW-NEXT: retq1353;1354; AVX512BWVL-LABEL: negative:1355; AVX512BWVL: # %bb.0:1356; AVX512BWVL-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[u,2,4,6,8,10,12,14,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,u,16,18,20,22,24,26,28,30]1357; AVX512BWVL-NEXT: movl $65537, %eax # imm = 0x100011358; AVX512BWVL-NEXT: kmovd %eax, %k11359; AVX512BWVL-NEXT: vmovdqu8 %ymm1, %ymm0 {%k1}1360; AVX512BWVL-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,3,2,3]1361; AVX512BWVL-NEXT: # kill: def $xmm0 killed $xmm0 killed $ymm01362; AVX512BWVL-NEXT: vzeroupper1363; AVX512BWVL-NEXT: retq1364;1365; AVX512VBMIVL-LABEL: negative:1366; AVX512VBMIVL: # %bb.0:1367; AVX512VBMIVL-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [32,2,4,6,8,10,12,14,16,18,20,22,24,26,28,30,32,2,4,6,8,10,12,14,16,18,20,22,24,26,28,30]1368; AVX512VBMIVL-NEXT: # ymm2 = mem[0,1,0,1]1369; AVX512VBMIVL-NEXT: vpermt2b %ymm1, %ymm2, %ymm01370; AVX512VBMIVL-NEXT: # kill: def $xmm0 killed $xmm0 killed $ymm01371; AVX512VBMIVL-NEXT: vzeroupper1372; AVX512VBMIVL-NEXT: retq1373 %strided.vec = shufflevector <32 x i8> %v, <32 x i8> undef, <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>1374 %w0 = extractelement <32 x i8> %w, i32 01375 %merged = insertelement <16 x i8> %strided.vec, i8 %w0, i32 01376 ret <16 x i8> %merged1377}1378