451 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+sse2 | FileCheck %s --check-prefix=SSE23; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+sse4.2 | FileCheck %s --check-prefix=SSE424; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx | FileCheck %s --check-prefix=AVX15; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx2 | FileCheck %s --check-prefix=AVX26; 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 | FileCheck %s --check-prefix=AVX512VL8; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512bw | FileCheck %s --check-prefix=AVX512BWNOVL9; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512bw,+avx512vl | FileCheck %s --check-prefix=AVX512BWVL10 11;12; General cases - packing of vector comparison to legal vector result types13;14 15define <16 x i8> @vselect_packss_v16i16(<16 x i16> %a0, <16 x i16> %a1, <16 x i8> %a2, <16 x i8> %a3) {16; SSE2-LABEL: vselect_packss_v16i16:17; SSE2: # %bb.0:18; SSE2-NEXT: pcmpeqw %xmm3, %xmm119; SSE2-NEXT: pcmpeqw %xmm2, %xmm020; SSE2-NEXT: packsswb %xmm1, %xmm021; SSE2-NEXT: pand %xmm0, %xmm422; SSE2-NEXT: pandn %xmm5, %xmm023; SSE2-NEXT: por %xmm4, %xmm024; SSE2-NEXT: retq25;26; SSE42-LABEL: vselect_packss_v16i16:27; SSE42: # %bb.0:28; SSE42-NEXT: pcmpeqw %xmm3, %xmm129; SSE42-NEXT: pcmpeqw %xmm2, %xmm030; SSE42-NEXT: packsswb %xmm1, %xmm031; SSE42-NEXT: pblendvb %xmm0, %xmm4, %xmm532; SSE42-NEXT: movdqa %xmm5, %xmm033; SSE42-NEXT: retq34;35; AVX1-LABEL: vselect_packss_v16i16:36; AVX1: # %bb.0:37; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm438; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm539; AVX1-NEXT: vpcmpeqw %xmm4, %xmm5, %xmm440; AVX1-NEXT: vpcmpeqw %xmm1, %xmm0, %xmm041; AVX1-NEXT: vpacksswb %xmm4, %xmm0, %xmm042; AVX1-NEXT: vpblendvb %xmm0, %xmm2, %xmm3, %xmm043; AVX1-NEXT: vzeroupper44; AVX1-NEXT: retq45;46; AVX2-LABEL: vselect_packss_v16i16:47; AVX2: # %bb.0:48; AVX2-NEXT: vpcmpeqw %ymm1, %ymm0, %ymm049; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm150; AVX2-NEXT: vpacksswb %xmm1, %xmm0, %xmm051; AVX2-NEXT: vpblendvb %xmm0, %xmm2, %xmm3, %xmm052; AVX2-NEXT: vzeroupper53; AVX2-NEXT: retq54;55; AVX512F-LABEL: vselect_packss_v16i16:56; AVX512F: # %bb.0:57; AVX512F-NEXT: vpcmpeqw %ymm1, %ymm0, %ymm058; AVX512F-NEXT: vpmovzxwd {{.*#+}} zmm0 = ymm0[0],zero,ymm0[1],zero,ymm0[2],zero,ymm0[3],zero,ymm0[4],zero,ymm0[5],zero,ymm0[6],zero,ymm0[7],zero,ymm0[8],zero,ymm0[9],zero,ymm0[10],zero,ymm0[11],zero,ymm0[12],zero,ymm0[13],zero,ymm0[14],zero,ymm0[15],zero59; AVX512F-NEXT: vpmovdb %zmm0, %xmm060; AVX512F-NEXT: vpblendvb %xmm0, %xmm2, %xmm3, %xmm061; AVX512F-NEXT: vzeroupper62; AVX512F-NEXT: retq63;64; AVX512VL-LABEL: vselect_packss_v16i16:65; AVX512VL: # %bb.0:66; AVX512VL-NEXT: vpcmpeqw %ymm1, %ymm0, %ymm067; AVX512VL-NEXT: vpmovzxwd {{.*#+}} zmm0 = ymm0[0],zero,ymm0[1],zero,ymm0[2],zero,ymm0[3],zero,ymm0[4],zero,ymm0[5],zero,ymm0[6],zero,ymm0[7],zero,ymm0[8],zero,ymm0[9],zero,ymm0[10],zero,ymm0[11],zero,ymm0[12],zero,ymm0[13],zero,ymm0[14],zero,ymm0[15],zero68; AVX512VL-NEXT: vpmovdb %zmm0, %xmm069; AVX512VL-NEXT: vpternlogq {{.*#+}} xmm0 = xmm3 ^ (xmm0 & (xmm2 ^ xmm3))70; AVX512VL-NEXT: vzeroupper71; AVX512VL-NEXT: retq72;73; AVX512BWNOVL-LABEL: vselect_packss_v16i16:74; AVX512BWNOVL: # %bb.0:75; AVX512BWNOVL-NEXT: vpcmpeqw %ymm1, %ymm0, %ymm076; AVX512BWNOVL-NEXT: vpmovwb %zmm0, %ymm077; AVX512BWNOVL-NEXT: vpblendvb %xmm0, %xmm2, %xmm3, %xmm078; AVX512BWNOVL-NEXT: vzeroupper79; AVX512BWNOVL-NEXT: retq80;81; AVX512BWVL-LABEL: vselect_packss_v16i16:82; AVX512BWVL: # %bb.0:83; AVX512BWVL-NEXT: vpcmpeqw %ymm1, %ymm0, %k084; AVX512BWVL-NEXT: vpmovm2b %k0, %xmm085; AVX512BWVL-NEXT: vpternlogq {{.*#+}} xmm0 = xmm3 ^ (xmm0 & (xmm2 ^ xmm3))86; AVX512BWVL-NEXT: vzeroupper87; AVX512BWVL-NEXT: retq88 %1 = icmp eq <16 x i16> %a0, %a189 %2 = sext <16 x i1> %1 to <16 x i8>90 %3 = and <16 x i8> %2, %a291 %4 = xor <16 x i8> %2, <i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1>92 %5 = and <16 x i8> %4, %a393 %6 = or <16 x i8> %3, %594 ret <16 x i8> %695}96 97define <16 x i8> @vselect_packss_v16i32(<16 x i32> %a0, <16 x i32> %a1, <16 x i8> %a2, <16 x i8> %a3) {98; SSE2-LABEL: vselect_packss_v16i32:99; SSE2: # %bb.0:100; SSE2-NEXT: pcmpeqd %xmm7, %xmm3101; SSE2-NEXT: pcmpeqd %xmm6, %xmm2102; SSE2-NEXT: packssdw %xmm3, %xmm2103; SSE2-NEXT: pcmpeqd %xmm5, %xmm1104; SSE2-NEXT: pcmpeqd %xmm4, %xmm0105; SSE2-NEXT: packssdw %xmm1, %xmm0106; SSE2-NEXT: packsswb %xmm2, %xmm0107; SSE2-NEXT: movdqa {{[0-9]+}}(%rsp), %xmm1108; SSE2-NEXT: pand %xmm0, %xmm1109; SSE2-NEXT: pandn {{[0-9]+}}(%rsp), %xmm0110; SSE2-NEXT: por %xmm1, %xmm0111; SSE2-NEXT: retq112;113; SSE42-LABEL: vselect_packss_v16i32:114; SSE42: # %bb.0:115; SSE42-NEXT: movdqa {{[0-9]+}}(%rsp), %xmm8116; SSE42-NEXT: pcmpeqd %xmm7, %xmm3117; SSE42-NEXT: pcmpeqd %xmm6, %xmm2118; SSE42-NEXT: packssdw %xmm3, %xmm2119; SSE42-NEXT: pcmpeqd %xmm5, %xmm1120; SSE42-NEXT: pcmpeqd %xmm4, %xmm0121; SSE42-NEXT: packssdw %xmm1, %xmm0122; SSE42-NEXT: packsswb %xmm2, %xmm0123; SSE42-NEXT: pblendvb %xmm0, {{[0-9]+}}(%rsp), %xmm8124; SSE42-NEXT: movdqa %xmm8, %xmm0125; SSE42-NEXT: retq126;127; AVX1-LABEL: vselect_packss_v16i32:128; AVX1: # %bb.0:129; AVX1-NEXT: vextractf128 $1, %ymm3, %xmm6130; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm7131; AVX1-NEXT: vpcmpeqd %xmm6, %xmm7, %xmm6132; AVX1-NEXT: vpcmpeqd %xmm3, %xmm1, %xmm1133; AVX1-NEXT: vpackssdw %xmm6, %xmm1, %xmm1134; AVX1-NEXT: vextractf128 $1, %ymm2, %xmm3135; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm6136; AVX1-NEXT: vpcmpeqd %xmm3, %xmm6, %xmm3137; AVX1-NEXT: vpcmpeqd %xmm2, %xmm0, %xmm0138; AVX1-NEXT: vpackssdw %xmm3, %xmm0, %xmm0139; AVX1-NEXT: vpacksswb %xmm1, %xmm0, %xmm0140; AVX1-NEXT: vpblendvb %xmm0, %xmm4, %xmm5, %xmm0141; AVX1-NEXT: vzeroupper142; AVX1-NEXT: retq143;144; AVX2-LABEL: vselect_packss_v16i32:145; AVX2: # %bb.0:146; AVX2-NEXT: vpcmpeqd %ymm3, %ymm1, %ymm1147; AVX2-NEXT: vpcmpeqd %ymm2, %ymm0, %ymm0148; AVX2-NEXT: vpackssdw %ymm1, %ymm0, %ymm0149; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1150; AVX2-NEXT: vpacksswb %xmm1, %xmm0, %xmm0151; AVX2-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,2,1,3]152; AVX2-NEXT: vpblendvb %xmm0, %xmm4, %xmm5, %xmm0153; AVX2-NEXT: vzeroupper154; AVX2-NEXT: retq155;156; AVX512F-LABEL: vselect_packss_v16i32:157; AVX512F: # %bb.0:158; AVX512F-NEXT: vpcmpeqd %zmm1, %zmm0, %k1159; AVX512F-NEXT: vpternlogd {{.*#+}} zmm0 {%k1} {z} = -1160; AVX512F-NEXT: vpmovdb %zmm0, %xmm0161; AVX512F-NEXT: vpblendvb %xmm0, %xmm2, %xmm3, %xmm0162; AVX512F-NEXT: vzeroupper163; AVX512F-NEXT: retq164;165; AVX512VL-LABEL: vselect_packss_v16i32:166; AVX512VL: # %bb.0:167; AVX512VL-NEXT: vpcmpeqd %zmm1, %zmm0, %k1168; AVX512VL-NEXT: vpternlogd {{.*#+}} zmm0 {%k1} {z} = -1169; AVX512VL-NEXT: vpmovdb %zmm0, %xmm0170; AVX512VL-NEXT: vpternlogq {{.*#+}} xmm0 = xmm3 ^ (xmm0 & (xmm2 ^ xmm3))171; AVX512VL-NEXT: vzeroupper172; AVX512VL-NEXT: retq173;174; AVX512BWNOVL-LABEL: vselect_packss_v16i32:175; AVX512BWNOVL: # %bb.0:176; AVX512BWNOVL-NEXT: vpcmpeqd %zmm1, %zmm0, %k0177; AVX512BWNOVL-NEXT: vpmovm2b %k0, %zmm0178; AVX512BWNOVL-NEXT: vpblendvb %xmm0, %xmm2, %xmm3, %xmm0179; AVX512BWNOVL-NEXT: vzeroupper180; AVX512BWNOVL-NEXT: retq181;182; AVX512BWVL-LABEL: vselect_packss_v16i32:183; AVX512BWVL: # %bb.0:184; AVX512BWVL-NEXT: vpcmpeqd %zmm1, %zmm0, %k0185; AVX512BWVL-NEXT: vpmovm2b %k0, %xmm0186; AVX512BWVL-NEXT: vpternlogq {{.*#+}} xmm0 = xmm3 ^ (xmm0 & (xmm2 ^ xmm3))187; AVX512BWVL-NEXT: vzeroupper188; AVX512BWVL-NEXT: retq189 %1 = icmp eq <16 x i32> %a0, %a1190 %2 = sext <16 x i1> %1 to <16 x i8>191 %3 = and <16 x i8> %2, %a2192 %4 = xor <16 x i8> %2, <i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1>193 %5 = and <16 x i8> %4, %a3194 %6 = or <16 x i8> %3, %5195 ret <16 x i8> %6196}197 198define <16 x i8> @vselect_packss_v16i64(<16 x i64> %a0, <16 x i64> %a1, <16 x i8> %a2, <16 x i8> %a3) {199; SSE2-LABEL: vselect_packss_v16i64:200; SSE2: # %bb.0:201; SSE2-NEXT: pcmpeqd {{[0-9]+}}(%rsp), %xmm7202; SSE2-NEXT: pshufd {{.*#+}} xmm8 = xmm7[1,0,3,2]203; SSE2-NEXT: pand %xmm7, %xmm8204; SSE2-NEXT: pcmpeqd {{[0-9]+}}(%rsp), %xmm6205; SSE2-NEXT: pshufd {{.*#+}} xmm7 = xmm6[1,0,3,2]206; SSE2-NEXT: pand %xmm6, %xmm7207; SSE2-NEXT: packssdw %xmm8, %xmm7208; SSE2-NEXT: pcmpeqd {{[0-9]+}}(%rsp), %xmm5209; SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm5[1,0,3,2]210; SSE2-NEXT: pand %xmm5, %xmm6211; SSE2-NEXT: pcmpeqd {{[0-9]+}}(%rsp), %xmm4212; SSE2-NEXT: pshufd {{.*#+}} xmm5 = xmm4[1,0,3,2]213; SSE2-NEXT: pand %xmm4, %xmm5214; SSE2-NEXT: packssdw %xmm6, %xmm5215; SSE2-NEXT: packssdw %xmm7, %xmm5216; SSE2-NEXT: pcmpeqd {{[0-9]+}}(%rsp), %xmm3217; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm3[1,0,3,2]218; SSE2-NEXT: pand %xmm3, %xmm4219; SSE2-NEXT: pcmpeqd {{[0-9]+}}(%rsp), %xmm2220; SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm2[1,0,3,2]221; SSE2-NEXT: pand %xmm2, %xmm3222; SSE2-NEXT: packssdw %xmm4, %xmm3223; SSE2-NEXT: pcmpeqd {{[0-9]+}}(%rsp), %xmm1224; SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm1[1,0,3,2]225; SSE2-NEXT: pand %xmm1, %xmm2226; SSE2-NEXT: pcmpeqd {{[0-9]+}}(%rsp), %xmm0227; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,0,3,2]228; SSE2-NEXT: pand %xmm1, %xmm0229; SSE2-NEXT: packssdw %xmm2, %xmm0230; SSE2-NEXT: packssdw %xmm3, %xmm0231; SSE2-NEXT: packsswb %xmm5, %xmm0232; SSE2-NEXT: movdqa {{[0-9]+}}(%rsp), %xmm1233; SSE2-NEXT: pand %xmm0, %xmm1234; SSE2-NEXT: pandn {{[0-9]+}}(%rsp), %xmm0235; SSE2-NEXT: por %xmm1, %xmm0236; SSE2-NEXT: retq237;238; SSE42-LABEL: vselect_packss_v16i64:239; SSE42: # %bb.0:240; SSE42-NEXT: movdqa {{[0-9]+}}(%rsp), %xmm8241; SSE42-NEXT: pcmpeqq {{[0-9]+}}(%rsp), %xmm7242; SSE42-NEXT: pcmpeqq {{[0-9]+}}(%rsp), %xmm6243; SSE42-NEXT: packssdw %xmm7, %xmm6244; SSE42-NEXT: pcmpeqq {{[0-9]+}}(%rsp), %xmm5245; SSE42-NEXT: pcmpeqq {{[0-9]+}}(%rsp), %xmm4246; SSE42-NEXT: packssdw %xmm5, %xmm4247; SSE42-NEXT: packssdw %xmm6, %xmm4248; SSE42-NEXT: pcmpeqq {{[0-9]+}}(%rsp), %xmm3249; SSE42-NEXT: pcmpeqq {{[0-9]+}}(%rsp), %xmm2250; SSE42-NEXT: packssdw %xmm3, %xmm2251; SSE42-NEXT: pcmpeqq {{[0-9]+}}(%rsp), %xmm1252; SSE42-NEXT: pcmpeqq {{[0-9]+}}(%rsp), %xmm0253; SSE42-NEXT: packssdw %xmm1, %xmm0254; SSE42-NEXT: packssdw %xmm2, %xmm0255; SSE42-NEXT: packsswb %xmm4, %xmm0256; SSE42-NEXT: pblendvb %xmm0, {{[0-9]+}}(%rsp), %xmm8257; SSE42-NEXT: movdqa %xmm8, %xmm0258; SSE42-NEXT: retq259;260; AVX1-LABEL: vselect_packss_v16i64:261; AVX1: # %bb.0:262; AVX1-NEXT: vmovdqa {{[0-9]+}}(%rsp), %xmm8263; AVX1-NEXT: vextractf128 $1, %ymm7, %xmm9264; AVX1-NEXT: vextractf128 $1, %ymm3, %xmm10265; AVX1-NEXT: vpcmpeqq %xmm9, %xmm10, %xmm9266; AVX1-NEXT: vpcmpeqq %xmm7, %xmm3, %xmm3267; AVX1-NEXT: vpackssdw %xmm9, %xmm3, %xmm3268; AVX1-NEXT: vextractf128 $1, %ymm6, %xmm7269; AVX1-NEXT: vextractf128 $1, %ymm2, %xmm9270; AVX1-NEXT: vpcmpeqq %xmm7, %xmm9, %xmm7271; AVX1-NEXT: vpcmpeqq %xmm6, %xmm2, %xmm2272; AVX1-NEXT: vpackssdw %xmm7, %xmm2, %xmm2273; AVX1-NEXT: vpackssdw %xmm3, %xmm2, %xmm2274; AVX1-NEXT: vextractf128 $1, %ymm5, %xmm3275; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm6276; AVX1-NEXT: vpcmpeqq %xmm3, %xmm6, %xmm3277; AVX1-NEXT: vpcmpeqq %xmm5, %xmm1, %xmm1278; AVX1-NEXT: vpackssdw %xmm3, %xmm1, %xmm1279; AVX1-NEXT: vextractf128 $1, %ymm4, %xmm3280; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm5281; AVX1-NEXT: vpcmpeqq %xmm3, %xmm5, %xmm3282; AVX1-NEXT: vpcmpeqq %xmm4, %xmm0, %xmm0283; AVX1-NEXT: vpackssdw %xmm3, %xmm0, %xmm0284; AVX1-NEXT: vpackssdw %xmm1, %xmm0, %xmm0285; AVX1-NEXT: vpacksswb %xmm2, %xmm0, %xmm0286; AVX1-NEXT: vpblendvb %xmm0, {{[0-9]+}}(%rsp), %xmm8, %xmm0287; AVX1-NEXT: vzeroupper288; AVX1-NEXT: retq289;290; AVX2-LABEL: vselect_packss_v16i64:291; AVX2: # %bb.0:292; AVX2-NEXT: vpcmpeqq %ymm7, %ymm3, %ymm3293; AVX2-NEXT: vpcmpeqq %ymm6, %ymm2, %ymm2294; AVX2-NEXT: vpackssdw %ymm3, %ymm2, %ymm2295; AVX2-NEXT: vpermq {{.*#+}} ymm2 = ymm2[0,2,1,3]296; AVX2-NEXT: vpcmpeqq %ymm5, %ymm1, %ymm1297; AVX2-NEXT: vpcmpeqq %ymm4, %ymm0, %ymm0298; AVX2-NEXT: vpackssdw %ymm1, %ymm0, %ymm0299; AVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]300; AVX2-NEXT: vpackssdw %ymm2, %ymm0, %ymm0301; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1302; AVX2-NEXT: vpacksswb %xmm1, %xmm0, %xmm0303; AVX2-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,2,1,3]304; AVX2-NEXT: vpand {{[0-9]+}}(%rsp), %xmm0, %xmm1305; AVX2-NEXT: vpandn {{[0-9]+}}(%rsp), %xmm0, %xmm0306; AVX2-NEXT: vpor %xmm0, %xmm1, %xmm0307; AVX2-NEXT: vzeroupper308; AVX2-NEXT: retq309;310; AVX512F-LABEL: vselect_packss_v16i64:311; AVX512F: # %bb.0:312; AVX512F-NEXT: vpcmpeqq %zmm2, %zmm0, %k0313; AVX512F-NEXT: vpcmpeqq %zmm3, %zmm1, %k1314; AVX512F-NEXT: kunpckbw %k0, %k1, %k1315; AVX512F-NEXT: vpternlogd {{.*#+}} zmm0 {%k1} {z} = -1316; AVX512F-NEXT: vpmovdb %zmm0, %xmm0317; AVX512F-NEXT: vpblendvb %xmm0, %xmm4, %xmm5, %xmm0318; AVX512F-NEXT: vzeroupper319; AVX512F-NEXT: retq320;321; AVX512VL-LABEL: vselect_packss_v16i64:322; AVX512VL: # %bb.0:323; AVX512VL-NEXT: vpcmpeqq %zmm2, %zmm0, %k0324; AVX512VL-NEXT: vpcmpeqq %zmm3, %zmm1, %k1325; AVX512VL-NEXT: kunpckbw %k0, %k1, %k1326; AVX512VL-NEXT: vpternlogd {{.*#+}} zmm0 {%k1} {z} = -1327; AVX512VL-NEXT: vpmovdb %zmm0, %xmm0328; AVX512VL-NEXT: vpternlogq {{.*#+}} xmm0 = xmm5 ^ (xmm0 & (xmm4 ^ xmm5))329; AVX512VL-NEXT: vzeroupper330; AVX512VL-NEXT: retq331;332; AVX512BWNOVL-LABEL: vselect_packss_v16i64:333; AVX512BWNOVL: # %bb.0:334; AVX512BWNOVL-NEXT: vpcmpeqq %zmm2, %zmm0, %k0335; AVX512BWNOVL-NEXT: vpcmpeqq %zmm3, %zmm1, %k1336; AVX512BWNOVL-NEXT: kunpckbw %k0, %k1, %k0337; AVX512BWNOVL-NEXT: vpmovm2b %k0, %zmm0338; AVX512BWNOVL-NEXT: vpblendvb %xmm0, %xmm4, %xmm5, %xmm0339; AVX512BWNOVL-NEXT: vzeroupper340; AVX512BWNOVL-NEXT: retq341;342; AVX512BWVL-LABEL: vselect_packss_v16i64:343; AVX512BWVL: # %bb.0:344; AVX512BWVL-NEXT: vpcmpeqq %zmm2, %zmm0, %k0345; AVX512BWVL-NEXT: vpcmpeqq %zmm3, %zmm1, %k1346; AVX512BWVL-NEXT: kunpckbw %k0, %k1, %k0347; AVX512BWVL-NEXT: vpmovm2b %k0, %xmm0348; AVX512BWVL-NEXT: vpternlogq {{.*#+}} xmm0 = xmm5 ^ (xmm0 & (xmm4 ^ xmm5))349; AVX512BWVL-NEXT: vzeroupper350; AVX512BWVL-NEXT: retq351 %1 = icmp eq <16 x i64> %a0, %a1352 %2 = sext <16 x i1> %1 to <16 x i8>353 %3 = and <16 x i8> %2, %a2354 %4 = xor <16 x i8> %2, <i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1>355 %5 = and <16 x i8> %4, %a3356 %6 = or <16 x i8> %3, %5357 ret <16 x i8> %6358}359 360;361; PACKSS case362;363 364define <16 x i8> @vselect_packss(<16 x i16> %a0, <16 x i16> %a1, <16 x i8> %a2, <16 x i8> %a3) {365; SSE2-LABEL: vselect_packss:366; SSE2: # %bb.0:367; SSE2-NEXT: pcmpeqw %xmm3, %xmm1368; SSE2-NEXT: pcmpeqw %xmm2, %xmm0369; SSE2-NEXT: packsswb %xmm1, %xmm0370; SSE2-NEXT: pand %xmm0, %xmm4371; SSE2-NEXT: pandn %xmm5, %xmm0372; SSE2-NEXT: por %xmm4, %xmm0373; SSE2-NEXT: retq374;375; SSE42-LABEL: vselect_packss:376; SSE42: # %bb.0:377; SSE42-NEXT: pcmpeqw %xmm3, %xmm1378; SSE42-NEXT: pcmpeqw %xmm2, %xmm0379; SSE42-NEXT: packsswb %xmm1, %xmm0380; SSE42-NEXT: pblendvb %xmm0, %xmm4, %xmm5381; SSE42-NEXT: movdqa %xmm5, %xmm0382; SSE42-NEXT: retq383;384; AVX1-LABEL: vselect_packss:385; AVX1: # %bb.0:386; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm4387; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm5388; AVX1-NEXT: vpcmpeqw %xmm4, %xmm5, %xmm4389; AVX1-NEXT: vpcmpeqw %xmm1, %xmm0, %xmm0390; AVX1-NEXT: vpacksswb %xmm4, %xmm0, %xmm0391; AVX1-NEXT: vpblendvb %xmm0, %xmm2, %xmm3, %xmm0392; AVX1-NEXT: vzeroupper393; AVX1-NEXT: retq394;395; AVX2-LABEL: vselect_packss:396; AVX2: # %bb.0:397; AVX2-NEXT: vpcmpeqw %ymm1, %ymm0, %ymm0398; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1399; AVX2-NEXT: vpacksswb %xmm1, %xmm0, %xmm0400; AVX2-NEXT: vpblendvb %xmm0, %xmm2, %xmm3, %xmm0401; AVX2-NEXT: vzeroupper402; AVX2-NEXT: retq403;404; AVX512F-LABEL: vselect_packss:405; AVX512F: # %bb.0:406; AVX512F-NEXT: vpcmpeqw %ymm1, %ymm0, %ymm0407; AVX512F-NEXT: vextracti128 $1, %ymm0, %xmm1408; AVX512F-NEXT: vpacksswb %xmm1, %xmm0, %xmm0409; AVX512F-NEXT: vpblendvb %xmm0, %xmm2, %xmm3, %xmm0410; AVX512F-NEXT: vzeroupper411; AVX512F-NEXT: retq412;413; AVX512VL-LABEL: vselect_packss:414; AVX512VL: # %bb.0:415; AVX512VL-NEXT: vpcmpeqw %ymm1, %ymm0, %ymm0416; AVX512VL-NEXT: vextracti128 $1, %ymm0, %xmm1417; AVX512VL-NEXT: vpacksswb %xmm1, %xmm0, %xmm0418; AVX512VL-NEXT: vpternlogq {{.*#+}} xmm0 = xmm3 ^ (xmm0 & (xmm2 ^ xmm3))419; AVX512VL-NEXT: vzeroupper420; AVX512VL-NEXT: retq421;422; AVX512BWNOVL-LABEL: vselect_packss:423; AVX512BWNOVL: # %bb.0:424; AVX512BWNOVL-NEXT: vpcmpeqw %ymm1, %ymm0, %ymm0425; AVX512BWNOVL-NEXT: vextracti128 $1, %ymm0, %xmm1426; AVX512BWNOVL-NEXT: vpacksswb %xmm1, %xmm0, %xmm0427; AVX512BWNOVL-NEXT: vpblendvb %xmm0, %xmm2, %xmm3, %xmm0428; AVX512BWNOVL-NEXT: vzeroupper429; AVX512BWNOVL-NEXT: retq430;431; AVX512BWVL-LABEL: vselect_packss:432; AVX512BWVL: # %bb.0:433; AVX512BWVL-NEXT: vpcmpeqw %ymm1, %ymm0, %ymm0434; AVX512BWVL-NEXT: vextracti128 $1, %ymm0, %xmm1435; AVX512BWVL-NEXT: vpacksswb %xmm1, %xmm0, %xmm0436; AVX512BWVL-NEXT: vpternlogq {{.*#+}} xmm0 = xmm3 ^ (xmm0 & (xmm2 ^ xmm3))437; AVX512BWVL-NEXT: vzeroupper438; AVX512BWVL-NEXT: retq439 %1 = icmp eq <16 x i16> %a0, %a1440 %2 = sext <16 x i1> %1 to <16 x i16>441 %3 = shufflevector <16 x i16> %2, <16 x i16> undef, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>442 %4 = shufflevector <16 x i16> %2, <16 x i16> undef, <8 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>443 %5 = tail call <16 x i8> @llvm.x86.sse2.packsswb.128(<8 x i16> %3, <8 x i16> %4)444 %6 = and <16 x i8> %5, %a2445 %7 = xor <16 x i8> %5, <i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1>446 %8 = and <16 x i8> %7, %a3447 %9 = or <16 x i8> %6, %8448 ret <16 x i8> %9449}450declare <16 x i8> @llvm.x86.sse2.packsswb.128(<8 x i16>, <8 x i16>)451