1478 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc < %s -mtriple=x86_64-unknown -mattr=+sse2 | FileCheck %s --check-prefixes=SSE,SSE23; RUN: llc < %s -mtriple=x86_64-unknown -mattr=+sse4.2 | FileCheck %s --check-prefixes=SSE,SSE424; RUN: llc < %s -mtriple=x86_64-unknown -mattr=+avx | FileCheck %s --check-prefixes=AVX,AVX1OR2,AVX15; RUN: llc < %s -mtriple=x86_64-unknown -mattr=+avx2 | FileCheck %s --check-prefixes=AVX,AVX1OR2,AVX26; RUN: llc < %s -mtriple=x86_64-unknown -mattr=+avx512f,+avx512bw,+avx512vl | FileCheck %s --check-prefixes=AVX,AVX5127 8define i64 @test_v2f64_sext(<2 x double> %a0, <2 x double> %a1) {9; SSE-LABEL: test_v2f64_sext:10; SSE: # %bb.0:11; SSE-NEXT: cmpltpd %xmm0, %xmm112; SSE-NEXT: movmskpd %xmm1, %ecx13; SSE-NEXT: xorl %eax, %eax14; SSE-NEXT: negl %ecx15; SSE-NEXT: sbbq %rax, %rax16; SSE-NEXT: retq17;18; AVX-LABEL: test_v2f64_sext:19; AVX: # %bb.0:20; AVX-NEXT: vcmpltpd %xmm0, %xmm1, %xmm021; AVX-NEXT: xorl %eax, %eax22; AVX-NEXT: vtestpd %xmm0, %xmm023; AVX-NEXT: setne %al24; AVX-NEXT: negq %rax25; AVX-NEXT: retq26 %c = fcmp ogt <2 x double> %a0, %a127 %s = sext <2 x i1> %c to <2 x i64>28 %1 = shufflevector <2 x i64> %s, <2 x i64> undef, <2 x i32> <i32 1, i32 undef>29 %2 = or <2 x i64> %s, %130 %3 = extractelement <2 x i64> %2, i32 031 ret i64 %332}33 34define i64 @test_v4f64_sext(<4 x double> %a0, <4 x double> %a1) {35; SSE-LABEL: test_v4f64_sext:36; SSE: # %bb.0:37; SSE-NEXT: cmpltpd %xmm1, %xmm338; SSE-NEXT: cmpltpd %xmm0, %xmm239; SSE-NEXT: orpd %xmm3, %xmm240; SSE-NEXT: movmskpd %xmm2, %ecx41; SSE-NEXT: xorl %eax, %eax42; SSE-NEXT: negl %ecx43; SSE-NEXT: sbbq %rax, %rax44; SSE-NEXT: retq45;46; AVX-LABEL: test_v4f64_sext:47; AVX: # %bb.0:48; AVX-NEXT: vcmpltpd %ymm0, %ymm1, %ymm049; AVX-NEXT: xorl %eax, %eax50; AVX-NEXT: vtestpd %ymm0, %ymm051; AVX-NEXT: setne %al52; AVX-NEXT: negq %rax53; AVX-NEXT: vzeroupper54; AVX-NEXT: retq55 %c = fcmp ogt <4 x double> %a0, %a156 %s = sext <4 x i1> %c to <4 x i64>57 %1 = shufflevector <4 x i64> %s, <4 x i64> undef, <4 x i32> <i32 2, i32 3, i32 undef, i32 undef>58 %2 = or <4 x i64> %s, %159 %3 = shufflevector <4 x i64> %2, <4 x i64> undef, <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef>60 %4 = or <4 x i64> %2, %361 %5 = extractelement <4 x i64> %4, i64 062 ret i64 %563}64 65define i64 @test_v4f64_legal_sext(<4 x double> %a0, <4 x double> %a1) {66; SSE-LABEL: test_v4f64_legal_sext:67; SSE: # %bb.0:68; SSE-NEXT: cmpltpd %xmm1, %xmm369; SSE-NEXT: cmpltpd %xmm0, %xmm270; SSE-NEXT: shufps {{.*#+}} xmm2 = xmm2[0,2],xmm3[0,2]71; SSE-NEXT: movmskps %xmm2, %ecx72; SSE-NEXT: xorl %eax, %eax73; SSE-NEXT: negl %ecx74; SSE-NEXT: sbbq %rax, %rax75; SSE-NEXT: retq76;77; AVX1OR2-LABEL: test_v4f64_legal_sext:78; AVX1OR2: # %bb.0:79; AVX1OR2-NEXT: vcmpltpd %ymm0, %ymm1, %ymm080; AVX1OR2-NEXT: vextractf128 $1, %ymm0, %xmm181; AVX1OR2-NEXT: vpackssdw %xmm1, %xmm0, %xmm082; AVX1OR2-NEXT: xorl %eax, %eax83; AVX1OR2-NEXT: vtestps %xmm0, %xmm084; AVX1OR2-NEXT: setne %al85; AVX1OR2-NEXT: negq %rax86; AVX1OR2-NEXT: vzeroupper87; AVX1OR2-NEXT: retq88;89; AVX512-LABEL: test_v4f64_legal_sext:90; AVX512: # %bb.0:91; AVX512-NEXT: vcmpltpd %ymm0, %ymm1, %k192; AVX512-NEXT: vpcmpeqd %xmm0, %xmm0, %xmm093; AVX512-NEXT: vmovdqa32 %xmm0, %xmm0 {%k1} {z}94; AVX512-NEXT: xorl %eax, %eax95; AVX512-NEXT: vtestps %xmm0, %xmm096; AVX512-NEXT: setne %al97; AVX512-NEXT: negq %rax98; AVX512-NEXT: vzeroupper99; AVX512-NEXT: retq100 %c = fcmp ogt <4 x double> %a0, %a1101 %s = sext <4 x i1> %c to <4 x i32>102 %1 = shufflevector <4 x i32> %s, <4 x i32> undef, <4 x i32> <i32 2, i32 3, i32 undef, i32 undef>103 %2 = or <4 x i32> %s, %1104 %3 = shufflevector <4 x i32> %2, <4 x i32> undef, <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef>105 %4 = or <4 x i32> %2, %3106 %5 = extractelement <4 x i32> %4, i64 0107 %6 = sext i32 %5 to i64108 ret i64 %6109}110 111define i32 @test_v4f32_sext(<4 x float> %a0, <4 x float> %a1) {112; SSE-LABEL: test_v4f32_sext:113; SSE: # %bb.0:114; SSE-NEXT: cmpltps %xmm0, %xmm1115; SSE-NEXT: movmskps %xmm1, %ecx116; SSE-NEXT: xorl %eax, %eax117; SSE-NEXT: negl %ecx118; SSE-NEXT: sbbl %eax, %eax119; SSE-NEXT: retq120;121; AVX-LABEL: test_v4f32_sext:122; AVX: # %bb.0:123; AVX-NEXT: vcmpltps %xmm0, %xmm1, %xmm0124; AVX-NEXT: xorl %eax, %eax125; AVX-NEXT: vtestps %xmm0, %xmm0126; AVX-NEXT: setne %al127; AVX-NEXT: negl %eax128; AVX-NEXT: retq129 %c = fcmp ogt <4 x float> %a0, %a1130 %s = sext <4 x i1> %c to <4 x i32>131 %1 = shufflevector <4 x i32> %s, <4 x i32> undef, <4 x i32> <i32 2, i32 3, i32 undef, i32 undef>132 %2 = or <4 x i32> %s, %1133 %3 = shufflevector <4 x i32> %2, <4 x i32> undef, <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef>134 %4 = or <4 x i32> %2, %3135 %5 = extractelement <4 x i32> %4, i32 0136 ret i32 %5137}138 139define i32 @test_v8f32_sext(<8 x float> %a0, <8 x float> %a1) {140; SSE-LABEL: test_v8f32_sext:141; SSE: # %bb.0:142; SSE-NEXT: cmpltps %xmm1, %xmm3143; SSE-NEXT: cmpltps %xmm0, %xmm2144; SSE-NEXT: orps %xmm3, %xmm2145; SSE-NEXT: movmskps %xmm2, %ecx146; SSE-NEXT: xorl %eax, %eax147; SSE-NEXT: negl %ecx148; SSE-NEXT: sbbl %eax, %eax149; SSE-NEXT: retq150;151; AVX-LABEL: test_v8f32_sext:152; AVX: # %bb.0:153; AVX-NEXT: vcmpltps %ymm0, %ymm1, %ymm0154; AVX-NEXT: xorl %eax, %eax155; AVX-NEXT: vtestps %ymm0, %ymm0156; AVX-NEXT: setne %al157; AVX-NEXT: negl %eax158; AVX-NEXT: vzeroupper159; AVX-NEXT: retq160 %c = fcmp ogt <8 x float> %a0, %a1161 %s = sext <8 x i1> %c to <8 x i32>162 %1 = shufflevector <8 x i32> %s, <8 x i32> undef, <8 x i32> <i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef>163 %2 = or <8 x i32> %s, %1164 %3 = shufflevector <8 x i32> %2, <8 x i32> undef, <8 x i32> <i32 2, i32 3, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>165 %4 = or <8 x i32> %2, %3166 %5 = shufflevector <8 x i32> %4, <8 x i32> undef, <8 x i32> <i32 1, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>167 %6 = or <8 x i32> %4, %5168 %7 = extractelement <8 x i32> %6, i32 0169 ret i32 %7170}171 172define i32 @test_v8f32_legal_sext(<8 x float> %a0, <8 x float> %a1) {173; SSE-LABEL: test_v8f32_legal_sext:174; SSE: # %bb.0:175; SSE-NEXT: cmpltps %xmm1, %xmm3176; SSE-NEXT: cmpltps %xmm0, %xmm2177; SSE-NEXT: packssdw %xmm3, %xmm2178; SSE-NEXT: pmovmskb %xmm2, %ecx179; SSE-NEXT: xorl %eax, %eax180; SSE-NEXT: negl %ecx181; SSE-NEXT: sbbl %eax, %eax182; SSE-NEXT: retq183;184; AVX1OR2-LABEL: test_v8f32_legal_sext:185; AVX1OR2: # %bb.0:186; AVX1OR2-NEXT: vcmpltps %ymm0, %ymm1, %ymm0187; AVX1OR2-NEXT: vextractf128 $1, %ymm0, %xmm1188; AVX1OR2-NEXT: vpackssdw %xmm1, %xmm0, %xmm0189; AVX1OR2-NEXT: vpmovmskb %xmm0, %ecx190; AVX1OR2-NEXT: xorl %eax, %eax191; AVX1OR2-NEXT: negl %ecx192; AVX1OR2-NEXT: sbbl %eax, %eax193; AVX1OR2-NEXT: vzeroupper194; AVX1OR2-NEXT: retq195;196; AVX512-LABEL: test_v8f32_legal_sext:197; AVX512: # %bb.0:198; AVX512-NEXT: vcmpltps %ymm0, %ymm1, %k0199; AVX512-NEXT: vpmovm2w %k0, %xmm0200; AVX512-NEXT: vpmovmskb %xmm0, %ecx201; AVX512-NEXT: xorl %eax, %eax202; AVX512-NEXT: negl %ecx203; AVX512-NEXT: sbbl %eax, %eax204; AVX512-NEXT: vzeroupper205; AVX512-NEXT: retq206 %c = fcmp ogt <8 x float> %a0, %a1207 %s = sext <8 x i1> %c to <8 x i16>208 %1 = shufflevector <8 x i16> %s, <8 x i16> undef, <8 x i32> <i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef>209 %2 = or <8 x i16> %s, %1210 %3 = shufflevector <8 x i16> %2, <8 x i16> undef, <8 x i32> <i32 2, i32 3, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>211 %4 = or <8 x i16> %2, %3212 %5 = shufflevector <8 x i16> %4, <8 x i16> undef, <8 x i32> <i32 1, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>213 %6 = or <8 x i16> %4, %5214 %7 = extractelement <8 x i16> %6, i32 0215 %8 = sext i16 %7 to i32216 ret i32 %8217}218 219define i64 @test_v2i64_sext(<2 x i64> %a0, <2 x i64> %a1) {220; SSE2-LABEL: test_v2i64_sext:221; SSE2: # %bb.0:222; SSE2-NEXT: movdqa {{.*#+}} xmm2 = [2147483648,2147483648]223; SSE2-NEXT: pxor %xmm2, %xmm1224; SSE2-NEXT: pxor %xmm2, %xmm0225; SSE2-NEXT: movdqa %xmm0, %xmm2226; SSE2-NEXT: pcmpeqd %xmm1, %xmm2227; SSE2-NEXT: pcmpgtd %xmm1, %xmm0228; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[0,0,2,2]229; SSE2-NEXT: pand %xmm2, %xmm1230; SSE2-NEXT: por %xmm0, %xmm1231; SSE2-NEXT: movmskpd %xmm1, %ecx232; SSE2-NEXT: xorl %eax, %eax233; SSE2-NEXT: negl %ecx234; SSE2-NEXT: sbbq %rax, %rax235; SSE2-NEXT: retq236;237; SSE42-LABEL: test_v2i64_sext:238; SSE42: # %bb.0:239; SSE42-NEXT: pcmpgtq %xmm1, %xmm0240; SSE42-NEXT: movmskpd %xmm0, %ecx241; SSE42-NEXT: xorl %eax, %eax242; SSE42-NEXT: negl %ecx243; SSE42-NEXT: sbbq %rax, %rax244; SSE42-NEXT: retq245;246; AVX-LABEL: test_v2i64_sext:247; AVX: # %bb.0:248; AVX-NEXT: vpcmpgtq %xmm1, %xmm0, %xmm0249; AVX-NEXT: xorl %eax, %eax250; AVX-NEXT: vtestpd %xmm0, %xmm0251; AVX-NEXT: setne %al252; AVX-NEXT: negq %rax253; AVX-NEXT: retq254 %c = icmp sgt <2 x i64> %a0, %a1255 %s = sext <2 x i1> %c to <2 x i64>256 %1 = shufflevector <2 x i64> %s, <2 x i64> undef, <2 x i32> <i32 1, i32 undef>257 %2 = or <2 x i64> %s, %1258 %3 = extractelement <2 x i64> %2, i32 0259 ret i64 %3260}261 262define i64 @test_v4i64_sext(<4 x i64> %a0, <4 x i64> %a1) {263; SSE2-LABEL: test_v4i64_sext:264; SSE2: # %bb.0:265; SSE2-NEXT: movdqa {{.*#+}} xmm4 = [2147483648,2147483648]266; SSE2-NEXT: pxor %xmm4, %xmm3267; SSE2-NEXT: pxor %xmm4, %xmm1268; SSE2-NEXT: movdqa %xmm1, %xmm5269; SSE2-NEXT: pcmpeqd %xmm3, %xmm5270; SSE2-NEXT: pcmpgtd %xmm3, %xmm1271; SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm1[0,0,2,2]272; SSE2-NEXT: pand %xmm5, %xmm3273; SSE2-NEXT: por %xmm1, %xmm3274; SSE2-NEXT: pxor %xmm4, %xmm2275; SSE2-NEXT: pxor %xmm4, %xmm0276; SSE2-NEXT: movdqa %xmm0, %xmm1277; SSE2-NEXT: pcmpeqd %xmm2, %xmm1278; SSE2-NEXT: pcmpgtd %xmm2, %xmm0279; SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm0[0,0,2,2]280; SSE2-NEXT: pand %xmm1, %xmm2281; SSE2-NEXT: por %xmm0, %xmm2282; SSE2-NEXT: por %xmm3, %xmm2283; SSE2-NEXT: movmskpd %xmm2, %ecx284; SSE2-NEXT: xorl %eax, %eax285; SSE2-NEXT: negl %ecx286; SSE2-NEXT: sbbq %rax, %rax287; SSE2-NEXT: retq288;289; SSE42-LABEL: test_v4i64_sext:290; SSE42: # %bb.0:291; SSE42-NEXT: pcmpgtq %xmm3, %xmm1292; SSE42-NEXT: pcmpgtq %xmm2, %xmm0293; SSE42-NEXT: por %xmm1, %xmm0294; SSE42-NEXT: movmskpd %xmm0, %ecx295; SSE42-NEXT: xorl %eax, %eax296; SSE42-NEXT: negl %ecx297; SSE42-NEXT: sbbq %rax, %rax298; SSE42-NEXT: retq299;300; AVX1-LABEL: test_v4i64_sext:301; AVX1: # %bb.0:302; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm2303; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm3304; AVX1-NEXT: vpcmpgtq %xmm2, %xmm3, %xmm2305; AVX1-NEXT: vpcmpgtq %xmm1, %xmm0, %xmm0306; AVX1-NEXT: vpor %xmm2, %xmm0, %xmm0307; AVX1-NEXT: xorl %eax, %eax308; AVX1-NEXT: vtestpd %xmm0, %xmm0309; AVX1-NEXT: setne %al310; AVX1-NEXT: negq %rax311; AVX1-NEXT: vzeroupper312; AVX1-NEXT: retq313;314; AVX2-LABEL: test_v4i64_sext:315; AVX2: # %bb.0:316; AVX2-NEXT: vpcmpgtq %ymm1, %ymm0, %ymm0317; AVX2-NEXT: xorl %eax, %eax318; AVX2-NEXT: vtestpd %ymm0, %ymm0319; AVX2-NEXT: setne %al320; AVX2-NEXT: negq %rax321; AVX2-NEXT: vzeroupper322; AVX2-NEXT: retq323;324; AVX512-LABEL: test_v4i64_sext:325; AVX512: # %bb.0:326; AVX512-NEXT: vpcmpgtq %ymm1, %ymm0, %ymm0327; AVX512-NEXT: xorl %eax, %eax328; AVX512-NEXT: vtestpd %ymm0, %ymm0329; AVX512-NEXT: setne %al330; AVX512-NEXT: negq %rax331; AVX512-NEXT: vzeroupper332; AVX512-NEXT: retq333 %c = icmp sgt <4 x i64> %a0, %a1334 %s = sext <4 x i1> %c to <4 x i64>335 %1 = shufflevector <4 x i64> %s, <4 x i64> undef, <4 x i32> <i32 2, i32 3, i32 undef, i32 undef>336 %2 = or <4 x i64> %s, %1337 %3 = shufflevector <4 x i64> %2, <4 x i64> undef, <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef>338 %4 = or <4 x i64> %2, %3339 %5 = extractelement <4 x i64> %4, i64 0340 ret i64 %5341}342 343define i64 @test_v4i64_legal_sext(<4 x i64> %a0, <4 x i64> %a1) {344; SSE2-LABEL: test_v4i64_legal_sext:345; SSE2: # %bb.0:346; SSE2-NEXT: movdqa {{.*#+}} xmm4 = [2147483648,2147483648]347; SSE2-NEXT: pxor %xmm4, %xmm3348; SSE2-NEXT: pxor %xmm4, %xmm1349; SSE2-NEXT: movdqa %xmm1, %xmm5350; SSE2-NEXT: pcmpgtd %xmm3, %xmm5351; SSE2-NEXT: pxor %xmm4, %xmm2352; SSE2-NEXT: pxor %xmm4, %xmm0353; SSE2-NEXT: movdqa %xmm0, %xmm4354; SSE2-NEXT: pcmpgtd %xmm2, %xmm4355; SSE2-NEXT: movdqa %xmm4, %xmm6356; SSE2-NEXT: shufps {{.*#+}} xmm6 = xmm6[0,2],xmm5[0,2]357; SSE2-NEXT: pcmpeqd %xmm3, %xmm1358; SSE2-NEXT: pcmpeqd %xmm2, %xmm0359; SSE2-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,3],xmm1[1,3]360; SSE2-NEXT: andps %xmm6, %xmm0361; SSE2-NEXT: shufps {{.*#+}} xmm4 = xmm4[1,3],xmm5[1,3]362; SSE2-NEXT: orps %xmm0, %xmm4363; SSE2-NEXT: movmskps %xmm4, %ecx364; SSE2-NEXT: xorl %eax, %eax365; SSE2-NEXT: negl %ecx366; SSE2-NEXT: sbbq %rax, %rax367; SSE2-NEXT: retq368;369; SSE42-LABEL: test_v4i64_legal_sext:370; SSE42: # %bb.0:371; SSE42-NEXT: pcmpgtq %xmm3, %xmm1372; SSE42-NEXT: pcmpgtq %xmm2, %xmm0373; SSE42-NEXT: packssdw %xmm1, %xmm0374; SSE42-NEXT: movmskps %xmm0, %ecx375; SSE42-NEXT: xorl %eax, %eax376; SSE42-NEXT: negl %ecx377; SSE42-NEXT: sbbq %rax, %rax378; SSE42-NEXT: retq379;380; AVX1-LABEL: test_v4i64_legal_sext:381; AVX1: # %bb.0:382; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm2383; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm3384; AVX1-NEXT: vpcmpgtq %xmm2, %xmm3, %xmm2385; AVX1-NEXT: vpcmpgtq %xmm1, %xmm0, %xmm0386; AVX1-NEXT: vpackssdw %xmm2, %xmm0, %xmm0387; AVX1-NEXT: xorl %eax, %eax388; AVX1-NEXT: vtestps %xmm0, %xmm0389; AVX1-NEXT: setne %al390; AVX1-NEXT: negq %rax391; AVX1-NEXT: vzeroupper392; AVX1-NEXT: retq393;394; AVX2-LABEL: test_v4i64_legal_sext:395; AVX2: # %bb.0:396; AVX2-NEXT: vpcmpgtq %ymm1, %ymm0, %ymm0397; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1398; AVX2-NEXT: vpackssdw %xmm1, %xmm0, %xmm0399; AVX2-NEXT: xorl %eax, %eax400; AVX2-NEXT: vtestps %xmm0, %xmm0401; AVX2-NEXT: setne %al402; AVX2-NEXT: negq %rax403; AVX2-NEXT: vzeroupper404; AVX2-NEXT: retq405;406; AVX512-LABEL: test_v4i64_legal_sext:407; AVX512: # %bb.0:408; AVX512-NEXT: vpcmpgtq %ymm1, %ymm0, %k1409; AVX512-NEXT: vpcmpeqd %xmm0, %xmm0, %xmm0410; AVX512-NEXT: vmovdqa32 %xmm0, %xmm0 {%k1} {z}411; AVX512-NEXT: xorl %eax, %eax412; AVX512-NEXT: vtestps %xmm0, %xmm0413; AVX512-NEXT: setne %al414; AVX512-NEXT: negq %rax415; AVX512-NEXT: vzeroupper416; AVX512-NEXT: retq417 %c = icmp sgt <4 x i64> %a0, %a1418 %s = sext <4 x i1> %c to <4 x i32>419 %1 = shufflevector <4 x i32> %s, <4 x i32> undef, <4 x i32> <i32 2, i32 3, i32 undef, i32 undef>420 %2 = or <4 x i32> %s, %1421 %3 = shufflevector <4 x i32> %2, <4 x i32> undef, <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef>422 %4 = or <4 x i32> %2, %3423 %5 = extractelement <4 x i32> %4, i64 0424 %6 = sext i32 %5 to i64425 ret i64 %6426}427 428define i32 @test_v4i32_sext(<4 x i32> %a0, <4 x i32> %a1) {429; SSE-LABEL: test_v4i32_sext:430; SSE: # %bb.0:431; SSE-NEXT: pcmpgtd %xmm1, %xmm0432; SSE-NEXT: movmskps %xmm0, %ecx433; SSE-NEXT: xorl %eax, %eax434; SSE-NEXT: negl %ecx435; SSE-NEXT: sbbl %eax, %eax436; SSE-NEXT: retq437;438; AVX-LABEL: test_v4i32_sext:439; AVX: # %bb.0:440; AVX-NEXT: vpcmpgtd %xmm1, %xmm0, %xmm0441; AVX-NEXT: xorl %eax, %eax442; AVX-NEXT: vtestps %xmm0, %xmm0443; AVX-NEXT: setne %al444; AVX-NEXT: negl %eax445; AVX-NEXT: retq446 %c = icmp sgt <4 x i32> %a0, %a1447 %s = sext <4 x i1> %c to <4 x i32>448 %1 = shufflevector <4 x i32> %s, <4 x i32> undef, <4 x i32> <i32 2, i32 3, i32 undef, i32 undef>449 %2 = or <4 x i32> %s, %1450 %3 = shufflevector <4 x i32> %2, <4 x i32> undef, <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef>451 %4 = or <4 x i32> %2, %3452 %5 = extractelement <4 x i32> %4, i32 0453 ret i32 %5454}455 456define i32 @test_v8i32_sext(<8 x i32> %a0, <8 x i32> %a1) {457; SSE-LABEL: test_v8i32_sext:458; SSE: # %bb.0:459; SSE-NEXT: pcmpgtd %xmm3, %xmm1460; SSE-NEXT: pcmpgtd %xmm2, %xmm0461; SSE-NEXT: por %xmm1, %xmm0462; SSE-NEXT: movmskps %xmm0, %ecx463; SSE-NEXT: xorl %eax, %eax464; SSE-NEXT: negl %ecx465; SSE-NEXT: sbbl %eax, %eax466; SSE-NEXT: retq467;468; AVX1-LABEL: test_v8i32_sext:469; AVX1: # %bb.0:470; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm2471; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm3472; AVX1-NEXT: vpcmpgtd %xmm2, %xmm3, %xmm2473; AVX1-NEXT: vpcmpgtd %xmm1, %xmm0, %xmm0474; AVX1-NEXT: vpor %xmm2, %xmm0, %xmm0475; AVX1-NEXT: xorl %eax, %eax476; AVX1-NEXT: vtestps %xmm0, %xmm0477; AVX1-NEXT: setne %al478; AVX1-NEXT: negl %eax479; AVX1-NEXT: vzeroupper480; AVX1-NEXT: retq481;482; AVX2-LABEL: test_v8i32_sext:483; AVX2: # %bb.0:484; AVX2-NEXT: vpcmpgtd %ymm1, %ymm0, %ymm0485; AVX2-NEXT: xorl %eax, %eax486; AVX2-NEXT: vtestps %ymm0, %ymm0487; AVX2-NEXT: setne %al488; AVX2-NEXT: negl %eax489; AVX2-NEXT: vzeroupper490; AVX2-NEXT: retq491;492; AVX512-LABEL: test_v8i32_sext:493; AVX512: # %bb.0:494; AVX512-NEXT: vpcmpgtd %ymm1, %ymm0, %ymm0495; AVX512-NEXT: xorl %eax, %eax496; AVX512-NEXT: vtestps %ymm0, %ymm0497; AVX512-NEXT: setne %al498; AVX512-NEXT: negl %eax499; AVX512-NEXT: vzeroupper500; AVX512-NEXT: retq501 %c = icmp sgt <8 x i32> %a0, %a1502 %s = sext <8 x i1> %c to <8 x i32>503 %1 = shufflevector <8 x i32> %s, <8 x i32> undef, <8 x i32> <i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef>504 %2 = or <8 x i32> %s, %1505 %3 = shufflevector <8 x i32> %2, <8 x i32> undef, <8 x i32> <i32 2, i32 3, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>506 %4 = or <8 x i32> %2, %3507 %5 = shufflevector <8 x i32> %4, <8 x i32> undef, <8 x i32> <i32 1, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>508 %6 = or <8 x i32> %4, %5509 %7 = extractelement <8 x i32> %6, i32 0510 ret i32 %7511}512 513define i32 @test_v8i32_legal_sext(<8 x i32> %a0, <8 x i32> %a1) {514; SSE-LABEL: test_v8i32_legal_sext:515; SSE: # %bb.0:516; SSE-NEXT: pcmpgtd %xmm3, %xmm1517; SSE-NEXT: pcmpgtd %xmm2, %xmm0518; SSE-NEXT: packssdw %xmm1, %xmm0519; SSE-NEXT: pmovmskb %xmm0, %ecx520; SSE-NEXT: xorl %eax, %eax521; SSE-NEXT: negl %ecx522; SSE-NEXT: sbbl %eax, %eax523; SSE-NEXT: retq524;525; AVX1-LABEL: test_v8i32_legal_sext:526; AVX1: # %bb.0:527; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm2528; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm3529; AVX1-NEXT: vpcmpgtd %xmm2, %xmm3, %xmm2530; AVX1-NEXT: vpcmpgtd %xmm1, %xmm0, %xmm0531; AVX1-NEXT: vpackssdw %xmm2, %xmm0, %xmm0532; AVX1-NEXT: vpmovmskb %xmm0, %ecx533; AVX1-NEXT: xorl %eax, %eax534; AVX1-NEXT: negl %ecx535; AVX1-NEXT: sbbl %eax, %eax536; AVX1-NEXT: vzeroupper537; AVX1-NEXT: retq538;539; AVX2-LABEL: test_v8i32_legal_sext:540; AVX2: # %bb.0:541; AVX2-NEXT: vpcmpgtd %ymm1, %ymm0, %ymm0542; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1543; AVX2-NEXT: vpackssdw %xmm1, %xmm0, %xmm0544; AVX2-NEXT: vpmovmskb %xmm0, %ecx545; AVX2-NEXT: xorl %eax, %eax546; AVX2-NEXT: negl %ecx547; AVX2-NEXT: sbbl %eax, %eax548; AVX2-NEXT: vzeroupper549; AVX2-NEXT: retq550;551; AVX512-LABEL: test_v8i32_legal_sext:552; AVX512: # %bb.0:553; AVX512-NEXT: vpcmpgtd %ymm1, %ymm0, %k0554; AVX512-NEXT: vpmovm2w %k0, %xmm0555; AVX512-NEXT: vpmovmskb %xmm0, %ecx556; AVX512-NEXT: xorl %eax, %eax557; AVX512-NEXT: negl %ecx558; AVX512-NEXT: sbbl %eax, %eax559; AVX512-NEXT: vzeroupper560; AVX512-NEXT: retq561 %c = icmp sgt <8 x i32> %a0, %a1562 %s = sext <8 x i1> %c to <8 x i16>563 %1 = shufflevector <8 x i16> %s, <8 x i16> undef, <8 x i32> <i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef>564 %2 = or <8 x i16> %s, %1565 %3 = shufflevector <8 x i16> %2, <8 x i16> undef, <8 x i32> <i32 2, i32 3, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>566 %4 = or <8 x i16> %2, %3567 %5 = shufflevector <8 x i16> %4, <8 x i16> undef, <8 x i32> <i32 1, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>568 %6 = or <8 x i16> %4, %5569 %7 = extractelement <8 x i16> %6, i32 0570 %8 = sext i16 %7 to i32571 ret i32 %8572}573 574define i16 @test_v8i16_sext(<8 x i16> %a0, <8 x i16> %a1) {575; SSE-LABEL: test_v8i16_sext:576; SSE: # %bb.0:577; SSE-NEXT: pcmpgtw %xmm1, %xmm0578; SSE-NEXT: pmovmskb %xmm0, %ecx579; SSE-NEXT: xorl %eax, %eax580; SSE-NEXT: negl %ecx581; SSE-NEXT: sbbl %eax, %eax582; SSE-NEXT: # kill: def $ax killed $ax killed $eax583; SSE-NEXT: retq584;585; AVX-LABEL: test_v8i16_sext:586; AVX: # %bb.0:587; AVX-NEXT: vpcmpgtw %xmm1, %xmm0, %xmm0588; AVX-NEXT: vpmovmskb %xmm0, %ecx589; AVX-NEXT: xorl %eax, %eax590; AVX-NEXT: negl %ecx591; AVX-NEXT: sbbl %eax, %eax592; AVX-NEXT: # kill: def $ax killed $ax killed $eax593; AVX-NEXT: retq594 %c = icmp sgt <8 x i16> %a0, %a1595 %s = sext <8 x i1> %c to <8 x i16>596 %1 = shufflevector <8 x i16> %s, <8 x i16> undef, <8 x i32> <i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef>597 %2 = or <8 x i16> %s, %1598 %3 = shufflevector <8 x i16> %2, <8 x i16> undef, <8 x i32> <i32 2, i32 3, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>599 %4 = or <8 x i16> %2, %3600 %5 = shufflevector <8 x i16> %4, <8 x i16> undef, <8 x i32> <i32 1, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>601 %6 = or <8 x i16> %4, %5602 %7 = extractelement <8 x i16> %6, i32 0603 ret i16 %7604}605 606define i16 @test_v16i16_sext(<16 x i16> %a0, <16 x i16> %a1) {607; SSE-LABEL: test_v16i16_sext:608; SSE: # %bb.0:609; SSE-NEXT: pcmpgtw %xmm3, %xmm1610; SSE-NEXT: pcmpgtw %xmm2, %xmm0611; SSE-NEXT: por %xmm1, %xmm0612; SSE-NEXT: pmovmskb %xmm0, %ecx613; SSE-NEXT: xorl %eax, %eax614; SSE-NEXT: negl %ecx615; SSE-NEXT: sbbl %eax, %eax616; SSE-NEXT: # kill: def $ax killed $ax killed $eax617; SSE-NEXT: retq618;619; AVX1-LABEL: test_v16i16_sext:620; AVX1: # %bb.0:621; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm2622; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm3623; AVX1-NEXT: vpcmpgtw %xmm2, %xmm3, %xmm2624; AVX1-NEXT: vpcmpgtw %xmm1, %xmm0, %xmm0625; AVX1-NEXT: vpor %xmm2, %xmm0, %xmm0626; AVX1-NEXT: vpmovmskb %xmm0, %ecx627; AVX1-NEXT: xorl %eax, %eax628; AVX1-NEXT: negl %ecx629; AVX1-NEXT: sbbl %eax, %eax630; AVX1-NEXT: # kill: def $ax killed $ax killed $eax631; AVX1-NEXT: vzeroupper632; AVX1-NEXT: retq633;634; AVX2-LABEL: test_v16i16_sext:635; AVX2: # %bb.0:636; AVX2-NEXT: vpcmpgtw %ymm1, %ymm0, %ymm0637; AVX2-NEXT: vpmovmskb %ymm0, %ecx638; AVX2-NEXT: xorl %eax, %eax639; AVX2-NEXT: negl %ecx640; AVX2-NEXT: sbbl %eax, %eax641; AVX2-NEXT: # kill: def $ax killed $ax killed $eax642; AVX2-NEXT: vzeroupper643; AVX2-NEXT: retq644;645; AVX512-LABEL: test_v16i16_sext:646; AVX512: # %bb.0:647; AVX512-NEXT: vpcmpgtw %ymm1, %ymm0, %ymm0648; AVX512-NEXT: vpmovmskb %ymm0, %ecx649; AVX512-NEXT: xorl %eax, %eax650; AVX512-NEXT: negl %ecx651; AVX512-NEXT: sbbl %eax, %eax652; AVX512-NEXT: # kill: def $ax killed $ax killed $eax653; AVX512-NEXT: vzeroupper654; AVX512-NEXT: retq655 %c = icmp sgt <16 x i16> %a0, %a1656 %s = sext <16 x i1> %c to <16 x i16>657 %1 = shufflevector <16 x i16> %s, <16 x i16> undef, <16 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>658 %2 = or <16 x i16> %s, %1659 %3 = shufflevector <16 x i16> %2, <16 x i16> undef, <16 x i32> <i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>660 %4 = or <16 x i16> %2, %3661 %5 = shufflevector <16 x i16> %4, <16 x i16> undef, <16 x i32> <i32 2, i32 3, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>662 %6 = or <16 x i16> %4, %5663 %7 = shufflevector <16 x i16> %6, <16 x i16> undef, <16 x i32> <i32 1, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>664 %8 = or <16 x i16> %6, %7665 %9 = extractelement <16 x i16> %8, i32 0666 ret i16 %9667}668 669define i16 @test_v16i16_legal_sext(<16 x i16> %a0, <16 x i16> %a1) {670; SSE-LABEL: test_v16i16_legal_sext:671; SSE: # %bb.0:672; SSE-NEXT: pcmpgtw %xmm3, %xmm1673; SSE-NEXT: pcmpgtw %xmm2, %xmm0674; SSE-NEXT: packsswb %xmm1, %xmm0675; SSE-NEXT: pmovmskb %xmm0, %ecx676; SSE-NEXT: xorl %eax, %eax677; SSE-NEXT: negl %ecx678; SSE-NEXT: sbbl %eax, %eax679; SSE-NEXT: # kill: def $ax killed $ax killed $eax680; SSE-NEXT: retq681;682; AVX1-LABEL: test_v16i16_legal_sext:683; AVX1: # %bb.0:684; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm2685; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm3686; AVX1-NEXT: vpcmpgtw %xmm2, %xmm3, %xmm2687; AVX1-NEXT: vpcmpgtw %xmm1, %xmm0, %xmm0688; AVX1-NEXT: vpacksswb %xmm2, %xmm0, %xmm0689; AVX1-NEXT: vpmovmskb %xmm0, %ecx690; AVX1-NEXT: xorl %eax, %eax691; AVX1-NEXT: negl %ecx692; AVX1-NEXT: sbbl %eax, %eax693; AVX1-NEXT: # kill: def $ax killed $ax killed $eax694; AVX1-NEXT: vzeroupper695; AVX1-NEXT: retq696;697; AVX2-LABEL: test_v16i16_legal_sext:698; AVX2: # %bb.0:699; AVX2-NEXT: vpcmpgtw %ymm1, %ymm0, %ymm0700; AVX2-NEXT: vpmovmskb %ymm0, %ecx701; AVX2-NEXT: xorl %eax, %eax702; AVX2-NEXT: negl %ecx703; AVX2-NEXT: sbbl %eax, %eax704; AVX2-NEXT: # kill: def $ax killed $ax killed $eax705; AVX2-NEXT: vzeroupper706; AVX2-NEXT: retq707;708; AVX512-LABEL: test_v16i16_legal_sext:709; AVX512: # %bb.0:710; AVX512-NEXT: vpcmpgtw %ymm1, %ymm0, %k0711; AVX512-NEXT: vpmovm2b %k0, %xmm0712; AVX512-NEXT: vpmovmskb %xmm0, %ecx713; AVX512-NEXT: xorl %eax, %eax714; AVX512-NEXT: negl %ecx715; AVX512-NEXT: sbbl %eax, %eax716; AVX512-NEXT: # kill: def $ax killed $ax killed $eax717; AVX512-NEXT: vzeroupper718; AVX512-NEXT: retq719 %c = icmp sgt <16 x i16> %a0, %a1720 %s = sext <16 x i1> %c to <16 x i8>721 %1 = shufflevector <16 x i8> %s, <16 x i8> undef, <16 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>722 %2 = or <16 x i8> %s, %1723 %3 = shufflevector <16 x i8> %2, <16 x i8> undef, <16 x i32> <i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>724 %4 = or <16 x i8> %2, %3725 %5 = shufflevector <16 x i8> %4, <16 x i8> undef, <16 x i32> <i32 2, i32 3, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>726 %6 = or <16 x i8> %4, %5727 %7 = shufflevector <16 x i8> %6, <16 x i8> undef, <16 x i32> <i32 1, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>728 %8 = or <16 x i8> %6, %7729 %9 = extractelement <16 x i8> %8, i32 0730 %10 = sext i8 %9 to i16731 ret i16 %10732}733 734define i8 @test_v16i8_sext(<16 x i8> %a0, <16 x i8> %a1) {735; SSE-LABEL: test_v16i8_sext:736; SSE: # %bb.0:737; SSE-NEXT: pcmpgtb %xmm1, %xmm0738; SSE-NEXT: pmovmskb %xmm0, %ecx739; SSE-NEXT: xorl %eax, %eax740; SSE-NEXT: negl %ecx741; SSE-NEXT: sbbl %eax, %eax742; SSE-NEXT: # kill: def $al killed $al killed $eax743; SSE-NEXT: retq744;745; AVX-LABEL: test_v16i8_sext:746; AVX: # %bb.0:747; AVX-NEXT: vpcmpgtb %xmm1, %xmm0, %xmm0748; AVX-NEXT: vpmovmskb %xmm0, %ecx749; AVX-NEXT: xorl %eax, %eax750; AVX-NEXT: negl %ecx751; AVX-NEXT: sbbl %eax, %eax752; AVX-NEXT: # kill: def $al killed $al killed $eax753; AVX-NEXT: retq754 %c = icmp sgt <16 x i8> %a0, %a1755 %s = sext <16 x i1> %c to <16 x i8>756 %1 = shufflevector <16 x i8> %s, <16 x i8> undef, <16 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>757 %2 = or <16 x i8> %s, %1758 %3 = shufflevector <16 x i8> %2, <16 x i8> undef, <16 x i32> <i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>759 %4 = or <16 x i8> %2, %3760 %5 = shufflevector <16 x i8> %4, <16 x i8> undef, <16 x i32> <i32 2, i32 3, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>761 %6 = or <16 x i8> %4, %5762 %7 = shufflevector <16 x i8> %6, <16 x i8> undef, <16 x i32> <i32 1, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>763 %8 = or <16 x i8> %6, %7764 %9 = extractelement <16 x i8> %8, i32 0765 ret i8 %9766}767 768define i8 @test_v32i8_sext(<32 x i8> %a0, <32 x i8> %a1) {769; SSE-LABEL: test_v32i8_sext:770; SSE: # %bb.0:771; SSE-NEXT: pcmpgtb %xmm3, %xmm1772; SSE-NEXT: pcmpgtb %xmm2, %xmm0773; SSE-NEXT: por %xmm1, %xmm0774; SSE-NEXT: pmovmskb %xmm0, %ecx775; SSE-NEXT: xorl %eax, %eax776; SSE-NEXT: negl %ecx777; SSE-NEXT: sbbl %eax, %eax778; SSE-NEXT: # kill: def $al killed $al killed $eax779; SSE-NEXT: retq780;781; AVX1-LABEL: test_v32i8_sext:782; AVX1: # %bb.0:783; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm2784; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm3785; AVX1-NEXT: vpcmpgtb %xmm2, %xmm3, %xmm2786; AVX1-NEXT: vpcmpgtb %xmm1, %xmm0, %xmm0787; AVX1-NEXT: vpor %xmm2, %xmm0, %xmm0788; AVX1-NEXT: vpmovmskb %xmm0, %ecx789; AVX1-NEXT: xorl %eax, %eax790; AVX1-NEXT: negl %ecx791; AVX1-NEXT: sbbl %eax, %eax792; AVX1-NEXT: # kill: def $al killed $al killed $eax793; AVX1-NEXT: vzeroupper794; AVX1-NEXT: retq795;796; AVX2-LABEL: test_v32i8_sext:797; AVX2: # %bb.0:798; AVX2-NEXT: vpcmpgtb %ymm1, %ymm0, %ymm0799; AVX2-NEXT: vpmovmskb %ymm0, %ecx800; AVX2-NEXT: xorl %eax, %eax801; AVX2-NEXT: negl %ecx802; AVX2-NEXT: sbbl %eax, %eax803; AVX2-NEXT: # kill: def $al killed $al killed $eax804; AVX2-NEXT: vzeroupper805; AVX2-NEXT: retq806;807; AVX512-LABEL: test_v32i8_sext:808; AVX512: # %bb.0:809; AVX512-NEXT: vpcmpgtb %ymm1, %ymm0, %ymm0810; AVX512-NEXT: vpmovmskb %ymm0, %ecx811; AVX512-NEXT: xorl %eax, %eax812; AVX512-NEXT: negl %ecx813; AVX512-NEXT: sbbl %eax, %eax814; AVX512-NEXT: # kill: def $al killed $al killed $eax815; AVX512-NEXT: vzeroupper816; AVX512-NEXT: retq817 %c = icmp sgt <32 x i8> %a0, %a1818 %s = sext <32 x i1> %c to <32 x i8>819 %1 = shufflevector <32 x i8> %s, <32 x i8> undef, <32 x i32> <i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>820 %2 = or <32 x i8> %s, %1821 %3 = shufflevector <32 x i8> %2, <32 x i8> undef, <32 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>822 %4 = or <32 x i8> %2, %3823 %5 = shufflevector <32 x i8> %4, <32 x i8> undef, <32 x i32> <i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>824 %6 = or <32 x i8> %4, %5825 %7 = shufflevector <32 x i8> %6, <32 x i8> undef, <32 x i32> <i32 2, i32 3, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>826 %8 = or <32 x i8> %6, %7827 %9 = shufflevector <32 x i8> %8, <32 x i8> undef, <32 x i32> <i32 1, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>828 %10 = or <32 x i8> %8, %9829 %11 = extractelement <32 x i8> %10, i32 0830 ret i8 %11831}832 833define i1 @bool_reduction_v2f64(<2 x double> %x, <2 x double> %y) {834; SSE-LABEL: bool_reduction_v2f64:835; SSE: # %bb.0:836; SSE-NEXT: cmpltpd %xmm0, %xmm1837; SSE-NEXT: movmskpd %xmm1, %eax838; SSE-NEXT: testl %eax, %eax839; SSE-NEXT: setne %al840; SSE-NEXT: retq841;842; AVX1OR2-LABEL: bool_reduction_v2f64:843; AVX1OR2: # %bb.0:844; AVX1OR2-NEXT: vcmpltpd %xmm0, %xmm1, %xmm0845; AVX1OR2-NEXT: vtestpd %xmm0, %xmm0846; AVX1OR2-NEXT: setne %al847; AVX1OR2-NEXT: retq848;849; AVX512-LABEL: bool_reduction_v2f64:850; AVX512: # %bb.0:851; AVX512-NEXT: vcmpltpd %xmm0, %xmm1, %k0852; AVX512-NEXT: kmovd %k0, %eax853; AVX512-NEXT: testb %al, %al854; AVX512-NEXT: setne %al855; AVX512-NEXT: retq856 %a = fcmp ogt <2 x double> %x, %y857 %b = shufflevector <2 x i1> %a, <2 x i1> undef, <2 x i32> <i32 1, i32 undef>858 %c = or <2 x i1> %a, %b859 %d = extractelement <2 x i1> %c, i32 0860 ret i1 %d861}862 863define i1 @bool_reduction_v4f32(<4 x float> %x, <4 x float> %y) {864; SSE-LABEL: bool_reduction_v4f32:865; SSE: # %bb.0:866; SSE-NEXT: cmpeqps %xmm1, %xmm0867; SSE-NEXT: movmskps %xmm0, %eax868; SSE-NEXT: testl %eax, %eax869; SSE-NEXT: setne %al870; SSE-NEXT: retq871;872; AVX1OR2-LABEL: bool_reduction_v4f32:873; AVX1OR2: # %bb.0:874; AVX1OR2-NEXT: vcmpeqps %xmm1, %xmm0, %xmm0875; AVX1OR2-NEXT: vtestps %xmm0, %xmm0876; AVX1OR2-NEXT: setne %al877; AVX1OR2-NEXT: retq878;879; AVX512-LABEL: bool_reduction_v4f32:880; AVX512: # %bb.0:881; AVX512-NEXT: vcmpeqps %xmm1, %xmm0, %k0882; AVX512-NEXT: kmovd %k0, %eax883; AVX512-NEXT: testb %al, %al884; AVX512-NEXT: setne %al885; AVX512-NEXT: retq886 %a = fcmp oeq <4 x float> %x, %y887 %s1 = shufflevector <4 x i1> %a, <4 x i1> undef, <4 x i32> <i32 2, i32 3, i32 undef, i32 undef>888 %b = or <4 x i1> %s1, %a889 %s2 = shufflevector <4 x i1> %b, <4 x i1> undef, <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef>890 %c = or <4 x i1> %s2, %b891 %d = extractelement <4 x i1> %c, i32 0892 ret i1 %d893}894 895define i1 @bool_reduction_v4f64(<4 x double> %x, <4 x double> %y) {896; SSE-LABEL: bool_reduction_v4f64:897; SSE: # %bb.0:898; SSE-NEXT: cmplepd %xmm1, %xmm3899; SSE-NEXT: cmplepd %xmm0, %xmm2900; SSE-NEXT: shufps {{.*#+}} xmm2 = xmm2[0,2],xmm3[0,2]901; SSE-NEXT: movmskps %xmm2, %eax902; SSE-NEXT: testl %eax, %eax903; SSE-NEXT: setne %al904; SSE-NEXT: retq905;906; AVX1OR2-LABEL: bool_reduction_v4f64:907; AVX1OR2: # %bb.0:908; AVX1OR2-NEXT: vcmplepd %ymm0, %ymm1, %ymm0909; AVX1OR2-NEXT: vtestpd %ymm0, %ymm0910; AVX1OR2-NEXT: setne %al911; AVX1OR2-NEXT: vzeroupper912; AVX1OR2-NEXT: retq913;914; AVX512-LABEL: bool_reduction_v4f64:915; AVX512: # %bb.0:916; AVX512-NEXT: vcmplepd %ymm0, %ymm1, %k0917; AVX512-NEXT: kmovd %k0, %eax918; AVX512-NEXT: testb %al, %al919; AVX512-NEXT: setne %al920; AVX512-NEXT: vzeroupper921; AVX512-NEXT: retq922 %a = fcmp oge <4 x double> %x, %y923 %s1 = shufflevector <4 x i1> %a, <4 x i1> undef, <4 x i32> <i32 2, i32 3, i32 undef, i32 undef>924 %b = or <4 x i1> %s1, %a925 %s2 = shufflevector <4 x i1> %b, <4 x i1> undef, <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef>926 %c = or <4 x i1> %s2, %b927 %d = extractelement <4 x i1> %c, i32 0928 ret i1 %d929}930 931define i1 @bool_reduction_v8f32(<8 x float> %x, <8 x float> %y) {932; SSE-LABEL: bool_reduction_v8f32:933; SSE: # %bb.0:934; SSE-NEXT: cmpneqps %xmm3, %xmm1935; SSE-NEXT: cmpneqps %xmm2, %xmm0936; SSE-NEXT: packssdw %xmm1, %xmm0937; SSE-NEXT: pmovmskb %xmm0, %eax938; SSE-NEXT: testl %eax, %eax939; SSE-NEXT: setne %al940; SSE-NEXT: retq941;942; AVX1OR2-LABEL: bool_reduction_v8f32:943; AVX1OR2: # %bb.0:944; AVX1OR2-NEXT: vcmpneqps %ymm1, %ymm0, %ymm0945; AVX1OR2-NEXT: vtestps %ymm0, %ymm0946; AVX1OR2-NEXT: setne %al947; AVX1OR2-NEXT: vzeroupper948; AVX1OR2-NEXT: retq949;950; AVX512-LABEL: bool_reduction_v8f32:951; AVX512: # %bb.0:952; AVX512-NEXT: vcmpneqps %ymm1, %ymm0, %k0953; AVX512-NEXT: kmovd %k0, %eax954; AVX512-NEXT: testb %al, %al955; AVX512-NEXT: setne %al956; AVX512-NEXT: vzeroupper957; AVX512-NEXT: retq958 %a = fcmp une <8 x float> %x, %y959 %s1 = shufflevector <8 x i1> %a, <8 x i1> undef, <8 x i32> <i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef>960 %b = or <8 x i1> %s1, %a961 %s2 = shufflevector <8 x i1> %b, <8 x i1> undef, <8 x i32> <i32 2, i32 3, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>962 %c = or <8 x i1> %s2, %b963 %s3 = shufflevector <8 x i1> %c, <8 x i1> undef, <8 x i32> <i32 1, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>964 %d = or <8 x i1> %s3, %c965 %e = extractelement <8 x i1> %d, i32 0966 ret i1 %e967}968 969define i1 @bool_reduction_v2i64(<2 x i64> %x, <2 x i64> %y) {970; SSE2-LABEL: bool_reduction_v2i64:971; SSE2: # %bb.0:972; SSE2-NEXT: pcmpeqd %xmm1, %xmm0973; SSE2-NEXT: movmskps %xmm0, %eax974; SSE2-NEXT: xorl $15, %eax975; SSE2-NEXT: setne %al976; SSE2-NEXT: retq977;978; SSE42-LABEL: bool_reduction_v2i64:979; SSE42: # %bb.0:980; SSE42-NEXT: pxor %xmm1, %xmm0981; SSE42-NEXT: ptest %xmm0, %xmm0982; SSE42-NEXT: setne %al983; SSE42-NEXT: retq984;985; AVX-LABEL: bool_reduction_v2i64:986; AVX: # %bb.0:987; AVX-NEXT: vpxor %xmm1, %xmm0, %xmm0988; AVX-NEXT: vptest %xmm0, %xmm0989; AVX-NEXT: setne %al990; AVX-NEXT: retq991 %a = icmp ne <2 x i64> %x, %y992 %b = shufflevector <2 x i1> %a, <2 x i1> undef, <2 x i32> <i32 1, i32 undef>993 %c = or <2 x i1> %a, %b994 %d = extractelement <2 x i1> %c, i32 0995 ret i1 %d996}997 998define i1 @bool_reduction_v4i32(<4 x i32> %x, <4 x i32> %y) {999; SSE2-LABEL: bool_reduction_v4i32:1000; SSE2: # %bb.0:1001; SSE2-NEXT: movdqa {{.*#+}} xmm2 = [2147483648,2147483648,2147483648,2147483648]1002; SSE2-NEXT: pxor %xmm2, %xmm11003; SSE2-NEXT: pxor %xmm2, %xmm01004; SSE2-NEXT: pcmpgtd %xmm1, %xmm01005; SSE2-NEXT: movmskps %xmm0, %eax1006; SSE2-NEXT: testl %eax, %eax1007; SSE2-NEXT: setne %al1008; SSE2-NEXT: retq1009;1010; SSE42-LABEL: bool_reduction_v4i32:1011; SSE42: # %bb.0:1012; SSE42-NEXT: pminud %xmm0, %xmm11013; SSE42-NEXT: pxor %xmm0, %xmm11014; SSE42-NEXT: ptest %xmm1, %xmm11015; SSE42-NEXT: setne %al1016; SSE42-NEXT: retq1017;1018; AVX1OR2-LABEL: bool_reduction_v4i32:1019; AVX1OR2: # %bb.0:1020; AVX1OR2-NEXT: vpminud %xmm1, %xmm0, %xmm11021; AVX1OR2-NEXT: vpxor %xmm1, %xmm0, %xmm01022; AVX1OR2-NEXT: vptest %xmm0, %xmm01023; AVX1OR2-NEXT: setne %al1024; AVX1OR2-NEXT: retq1025;1026; AVX512-LABEL: bool_reduction_v4i32:1027; AVX512: # %bb.0:1028; AVX512-NEXT: vpcmpnleud %xmm1, %xmm0, %k01029; AVX512-NEXT: kmovd %k0, %eax1030; AVX512-NEXT: testb %al, %al1031; AVX512-NEXT: setne %al1032; AVX512-NEXT: retq1033 %a = icmp ugt <4 x i32> %x, %y1034 %s1 = shufflevector <4 x i1> %a, <4 x i1> undef, <4 x i32> <i32 2, i32 3, i32 undef, i32 undef>1035 %b = or <4 x i1> %s1, %a1036 %s2 = shufflevector <4 x i1> %b, <4 x i1> undef, <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef>1037 %c = or <4 x i1> %s2, %b1038 %d = extractelement <4 x i1> %c, i32 01039 ret i1 %d1040}1041 1042define i1 @bool_reduction_v8i16(<8 x i16> %x, <8 x i16> %y) {1043; SSE-LABEL: bool_reduction_v8i16:1044; SSE: # %bb.0:1045; SSE-NEXT: pcmpgtw %xmm0, %xmm11046; SSE-NEXT: pmovmskb %xmm1, %eax1047; SSE-NEXT: testl %eax, %eax1048; SSE-NEXT: setne %al1049; SSE-NEXT: retq1050;1051; AVX1OR2-LABEL: bool_reduction_v8i16:1052; AVX1OR2: # %bb.0:1053; AVX1OR2-NEXT: vpcmpgtw %xmm0, %xmm1, %xmm01054; AVX1OR2-NEXT: vpmovmskb %xmm0, %eax1055; AVX1OR2-NEXT: testl %eax, %eax1056; AVX1OR2-NEXT: setne %al1057; AVX1OR2-NEXT: retq1058;1059; AVX512-LABEL: bool_reduction_v8i16:1060; AVX512: # %bb.0:1061; AVX512-NEXT: vpcmpgtw %xmm0, %xmm1, %k01062; AVX512-NEXT: kmovd %k0, %eax1063; AVX512-NEXT: testb %al, %al1064; AVX512-NEXT: setne %al1065; AVX512-NEXT: retq1066 %a = icmp slt <8 x i16> %x, %y1067 %s1 = shufflevector <8 x i1> %a, <8 x i1> undef, <8 x i32> <i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef>1068 %b = or <8 x i1> %s1, %a1069 %s2 = shufflevector <8 x i1> %b, <8 x i1> undef, <8 x i32> <i32 2, i32 3, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>1070 %c = or <8 x i1> %s2, %b1071 %s3 = shufflevector <8 x i1> %c, <8 x i1> undef, <8 x i32> <i32 1, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>1072 %d = or <8 x i1> %s3, %c1073 %e = extractelement <8 x i1> %d, i32 01074 ret i1 %e1075}1076 1077define i1 @bool_reduction_v16i8(<16 x i8> %x, <16 x i8> %y) {1078; SSE-LABEL: bool_reduction_v16i8:1079; SSE: # %bb.0:1080; SSE-NEXT: pcmpgtb %xmm1, %xmm01081; SSE-NEXT: pmovmskb %xmm0, %eax1082; SSE-NEXT: testl %eax, %eax1083; SSE-NEXT: setne %al1084; SSE-NEXT: retq1085;1086; AVX1OR2-LABEL: bool_reduction_v16i8:1087; AVX1OR2: # %bb.0:1088; AVX1OR2-NEXT: vpcmpgtb %xmm1, %xmm0, %xmm01089; AVX1OR2-NEXT: vpmovmskb %xmm0, %eax1090; AVX1OR2-NEXT: testl %eax, %eax1091; AVX1OR2-NEXT: setne %al1092; AVX1OR2-NEXT: retq1093;1094; AVX512-LABEL: bool_reduction_v16i8:1095; AVX512: # %bb.0:1096; AVX512-NEXT: vpcmpgtb %xmm1, %xmm0, %k01097; AVX512-NEXT: kortestw %k0, %k01098; AVX512-NEXT: setne %al1099; AVX512-NEXT: retq1100 %a = icmp sgt <16 x i8> %x, %y1101 %s1 = shufflevector <16 x i1> %a, <16 x i1> undef, <16 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>1102 %b = or <16 x i1> %s1, %a1103 %s2 = shufflevector <16 x i1> %b, <16 x i1> undef, <16 x i32> <i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>1104 %c = or <16 x i1> %s2, %b1105 %s3 = shufflevector <16 x i1> %c, <16 x i1> undef, <16 x i32> <i32 2, i32 3, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>1106 %d = or <16 x i1> %s3, %c1107 %s4 = shufflevector <16 x i1> %d, <16 x i1> undef, <16 x i32> <i32 1, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>1108 %e = or <16 x i1> %s4, %d1109 %f = extractelement <16 x i1> %e, i32 01110 ret i1 %f1111}1112 1113define i1 @bool_reduction_v4i64(<4 x i64> %x, <4 x i64> %y) {1114; SSE2-LABEL: bool_reduction_v4i64:1115; SSE2: # %bb.0:1116; SSE2-NEXT: movdqa {{.*#+}} xmm4 = [2147483648,2147483648]1117; SSE2-NEXT: pxor %xmm4, %xmm11118; SSE2-NEXT: pxor %xmm4, %xmm31119; SSE2-NEXT: movdqa %xmm3, %xmm51120; SSE2-NEXT: pcmpgtd %xmm1, %xmm51121; SSE2-NEXT: pxor %xmm4, %xmm01122; SSE2-NEXT: pxor %xmm4, %xmm21123; SSE2-NEXT: movdqa %xmm2, %xmm41124; SSE2-NEXT: pcmpgtd %xmm0, %xmm41125; SSE2-NEXT: movdqa %xmm4, %xmm61126; SSE2-NEXT: shufps {{.*#+}} xmm6 = xmm6[0,2],xmm5[0,2]1127; SSE2-NEXT: pcmpeqd %xmm1, %xmm31128; SSE2-NEXT: pcmpeqd %xmm0, %xmm21129; SSE2-NEXT: shufps {{.*#+}} xmm2 = xmm2[1,3],xmm3[1,3]1130; SSE2-NEXT: andps %xmm6, %xmm21131; SSE2-NEXT: shufps {{.*#+}} xmm4 = xmm4[1,3],xmm5[1,3]1132; SSE2-NEXT: orps %xmm2, %xmm41133; SSE2-NEXT: movmskps %xmm4, %eax1134; SSE2-NEXT: testl %eax, %eax1135; SSE2-NEXT: setne %al1136; SSE2-NEXT: retq1137;1138; SSE42-LABEL: bool_reduction_v4i64:1139; SSE42: # %bb.0:1140; SSE42-NEXT: pcmpgtq %xmm1, %xmm31141; SSE42-NEXT: pcmpgtq %xmm0, %xmm21142; SSE42-NEXT: packssdw %xmm3, %xmm21143; SSE42-NEXT: movmskps %xmm2, %eax1144; SSE42-NEXT: testl %eax, %eax1145; SSE42-NEXT: setne %al1146; SSE42-NEXT: retq1147;1148; AVX1-LABEL: bool_reduction_v4i64:1149; AVX1: # %bb.0:1150; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm21151; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm31152; AVX1-NEXT: vpcmpgtq %xmm2, %xmm3, %xmm21153; AVX1-NEXT: vpcmpgtq %xmm0, %xmm1, %xmm01154; AVX1-NEXT: vpor %xmm2, %xmm0, %xmm01155; AVX1-NEXT: vtestpd %xmm0, %xmm01156; AVX1-NEXT: setne %al1157; AVX1-NEXT: vzeroupper1158; AVX1-NEXT: retq1159;1160; AVX2-LABEL: bool_reduction_v4i64:1161; AVX2: # %bb.0:1162; AVX2-NEXT: vpcmpgtq %ymm0, %ymm1, %ymm01163; AVX2-NEXT: vtestpd %ymm0, %ymm01164; AVX2-NEXT: setne %al1165; AVX2-NEXT: vzeroupper1166; AVX2-NEXT: retq1167;1168; AVX512-LABEL: bool_reduction_v4i64:1169; AVX512: # %bb.0:1170; AVX512-NEXT: vpcmpgtq %ymm0, %ymm1, %k01171; AVX512-NEXT: kmovd %k0, %eax1172; AVX512-NEXT: testb %al, %al1173; AVX512-NEXT: setne %al1174; AVX512-NEXT: vzeroupper1175; AVX512-NEXT: retq1176 %a = icmp slt <4 x i64> %x, %y1177 %s1 = shufflevector <4 x i1> %a, <4 x i1> undef, <4 x i32> <i32 2, i32 3, i32 undef, i32 undef>1178 %b = or <4 x i1> %s1, %a1179 %s2 = shufflevector <4 x i1> %b, <4 x i1> undef, <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef>1180 %c = or <4 x i1> %s2, %b1181 %d = extractelement <4 x i1> %c, i32 01182 ret i1 %d1183}1184 1185define i1 @bool_reduction_v8i32(<8 x i32> %x, <8 x i32> %y) {1186; SSE2-LABEL: bool_reduction_v8i32:1187; SSE2: # %bb.0:1188; SSE2-NEXT: movdqa {{.*#+}} xmm4 = [2147483648,2147483648,2147483648,2147483648]1189; SSE2-NEXT: pxor %xmm4, %xmm31190; SSE2-NEXT: pxor %xmm4, %xmm11191; SSE2-NEXT: pcmpgtd %xmm3, %xmm11192; SSE2-NEXT: pxor %xmm4, %xmm21193; SSE2-NEXT: pxor %xmm4, %xmm01194; SSE2-NEXT: pcmpgtd %xmm2, %xmm01195; SSE2-NEXT: packssdw %xmm1, %xmm01196; SSE2-NEXT: packsswb %xmm0, %xmm01197; SSE2-NEXT: pmovmskb %xmm0, %eax1198; SSE2-NEXT: xorb $-1, %al1199; SSE2-NEXT: setne %al1200; SSE2-NEXT: retq1201;1202; SSE42-LABEL: bool_reduction_v8i32:1203; SSE42: # %bb.0:1204; SSE42-NEXT: pminud %xmm1, %xmm31205; SSE42-NEXT: pcmpeqd %xmm1, %xmm31206; SSE42-NEXT: pminud %xmm0, %xmm21207; SSE42-NEXT: pcmpeqd %xmm0, %xmm21208; SSE42-NEXT: packssdw %xmm3, %xmm21209; SSE42-NEXT: pmovmskb %xmm2, %eax1210; SSE42-NEXT: testl %eax, %eax1211; SSE42-NEXT: setne %al1212; SSE42-NEXT: retq1213;1214; AVX1-LABEL: bool_reduction_v8i32:1215; AVX1: # %bb.0:1216; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm21217; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm31218; AVX1-NEXT: vpminud %xmm2, %xmm3, %xmm21219; AVX1-NEXT: vpcmpeqd %xmm2, %xmm3, %xmm21220; AVX1-NEXT: vpminud %xmm1, %xmm0, %xmm11221; AVX1-NEXT: vpcmpeqd %xmm1, %xmm0, %xmm01222; AVX1-NEXT: vpor %xmm2, %xmm0, %xmm01223; AVX1-NEXT: vtestps %xmm0, %xmm01224; AVX1-NEXT: setne %al1225; AVX1-NEXT: vzeroupper1226; AVX1-NEXT: retq1227;1228; AVX2-LABEL: bool_reduction_v8i32:1229; AVX2: # %bb.0:1230; AVX2-NEXT: vpminud %ymm1, %ymm0, %ymm11231; AVX2-NEXT: vpcmpeqd %ymm1, %ymm0, %ymm01232; AVX2-NEXT: vtestps %ymm0, %ymm01233; AVX2-NEXT: setne %al1234; AVX2-NEXT: vzeroupper1235; AVX2-NEXT: retq1236;1237; AVX512-LABEL: bool_reduction_v8i32:1238; AVX512: # %bb.0:1239; AVX512-NEXT: vpcmpleud %ymm1, %ymm0, %k01240; AVX512-NEXT: kmovd %k0, %eax1241; AVX512-NEXT: testb %al, %al1242; AVX512-NEXT: setne %al1243; AVX512-NEXT: vzeroupper1244; AVX512-NEXT: retq1245 %a = icmp ule <8 x i32> %x, %y1246 %s1 = shufflevector <8 x i1> %a, <8 x i1> undef, <8 x i32> <i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef>1247 %b = or <8 x i1> %s1, %a1248 %s2 = shufflevector <8 x i1> %b, <8 x i1> undef, <8 x i32> <i32 2, i32 3, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>1249 %c = or <8 x i1> %s2, %b1250 %s3 = shufflevector <8 x i1> %c, <8 x i1> undef, <8 x i32> <i32 1, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>1251 %d = or <8 x i1> %s3, %c1252 %e = extractelement <8 x i1> %d, i32 01253 ret i1 %e1254}1255 1256define i1 @bool_reduction_v16i16(<16 x i16> %x, <16 x i16> %y) {1257; SSE-LABEL: bool_reduction_v16i16:1258; SSE: # %bb.0:1259; SSE-NEXT: pcmpeqw %xmm3, %xmm11260; SSE-NEXT: pcmpeqw %xmm2, %xmm01261; SSE-NEXT: packsswb %xmm1, %xmm01262; SSE-NEXT: pmovmskb %xmm0, %eax1263; SSE-NEXT: testl %eax, %eax1264; SSE-NEXT: setne %al1265; SSE-NEXT: retq1266;1267; AVX1-LABEL: bool_reduction_v16i16:1268; AVX1: # %bb.0:1269; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm21270; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm31271; AVX1-NEXT: vpcmpeqw %xmm2, %xmm3, %xmm21272; AVX1-NEXT: vpcmpeqw %xmm1, %xmm0, %xmm01273; AVX1-NEXT: vpacksswb %xmm2, %xmm0, %xmm01274; AVX1-NEXT: vpmovmskb %xmm0, %eax1275; AVX1-NEXT: testl %eax, %eax1276; AVX1-NEXT: setne %al1277; AVX1-NEXT: vzeroupper1278; AVX1-NEXT: retq1279;1280; AVX2-LABEL: bool_reduction_v16i16:1281; AVX2: # %bb.0:1282; AVX2-NEXT: vpcmpeqw %ymm1, %ymm0, %ymm01283; AVX2-NEXT: vpmovmskb %ymm0, %eax1284; AVX2-NEXT: testl %eax, %eax1285; AVX2-NEXT: setne %al1286; AVX2-NEXT: vzeroupper1287; AVX2-NEXT: retq1288;1289; AVX512-LABEL: bool_reduction_v16i16:1290; AVX512: # %bb.0:1291; AVX512-NEXT: vpcmpeqw %ymm1, %ymm0, %k01292; AVX512-NEXT: kortestw %k0, %k01293; AVX512-NEXT: setne %al1294; AVX512-NEXT: vzeroupper1295; AVX512-NEXT: retq1296 %a = icmp eq <16 x i16> %x, %y1297 %s1 = shufflevector <16 x i1> %a, <16 x i1> undef, <16 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>1298 %b = or <16 x i1> %s1, %a1299 %s2 = shufflevector <16 x i1> %b, <16 x i1> undef, <16 x i32> <i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>1300 %c = or <16 x i1> %s2, %b1301 %s3 = shufflevector <16 x i1> %c, <16 x i1> undef, <16 x i32> <i32 2, i32 3, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>1302 %d = or <16 x i1> %s3, %c1303 %s4 = shufflevector <16 x i1> %d, <16 x i1> undef, <16 x i32> <i32 1, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>1304 %e = or <16 x i1> %s4, %d1305 %f = extractelement <16 x i1> %e, i32 01306 ret i1 %f1307}1308 1309define i1 @bool_reduction_v32i8(<32 x i8> %x, <32 x i8> %y) {1310; SSE-LABEL: bool_reduction_v32i8:1311; SSE: # %bb.0:1312; SSE-NEXT: pcmpeqb %xmm3, %xmm11313; SSE-NEXT: pcmpeqb %xmm2, %xmm01314; SSE-NEXT: por %xmm1, %xmm01315; SSE-NEXT: pmovmskb %xmm0, %eax1316; SSE-NEXT: testl %eax, %eax1317; SSE-NEXT: setne %al1318; SSE-NEXT: retq1319;1320; AVX1-LABEL: bool_reduction_v32i8:1321; AVX1: # %bb.0:1322; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm21323; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm31324; AVX1-NEXT: vpcmpeqb %xmm2, %xmm3, %xmm21325; AVX1-NEXT: vpcmpeqb %xmm1, %xmm0, %xmm01326; AVX1-NEXT: vpor %xmm2, %xmm0, %xmm01327; AVX1-NEXT: vpmovmskb %xmm0, %eax1328; AVX1-NEXT: testl %eax, %eax1329; AVX1-NEXT: setne %al1330; AVX1-NEXT: vzeroupper1331; AVX1-NEXT: retq1332;1333; AVX2-LABEL: bool_reduction_v32i8:1334; AVX2: # %bb.0:1335; AVX2-NEXT: vpcmpeqb %ymm1, %ymm0, %ymm01336; AVX2-NEXT: vpmovmskb %ymm0, %eax1337; AVX2-NEXT: testl %eax, %eax1338; AVX2-NEXT: setne %al1339; AVX2-NEXT: vzeroupper1340; AVX2-NEXT: retq1341;1342; AVX512-LABEL: bool_reduction_v32i8:1343; AVX512: # %bb.0:1344; AVX512-NEXT: vpcmpeqb %ymm1, %ymm0, %k01345; AVX512-NEXT: kortestd %k0, %k01346; AVX512-NEXT: setne %al1347; AVX512-NEXT: vzeroupper1348; AVX512-NEXT: retq1349 %a = icmp eq <32 x i8> %x, %y1350 %s1 = shufflevector <32 x i1> %a, <32 x i1> undef, <32 x i32> <i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>1351 %b = or <32 x i1> %s1, %a1352 %s2 = shufflevector <32 x i1> %b, <32 x i1> undef, <32 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>1353 %c = or <32 x i1> %s2, %b1354 %s3 = shufflevector <32 x i1> %c, <32 x i1> undef, <32 x i32> <i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>1355 %d = or <32 x i1> %s3, %c1356 %s4 = shufflevector <32 x i1> %d, <32 x i1> undef, <32 x i32> <i32 2, i32 3, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>1357 %e = or <32 x i1> %s4, %d1358 %s5 = shufflevector <32 x i1> %e, <32 x i1> undef, <32 x i32> <i32 1, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>1359 %f = or <32 x i1> %s5, %e1360 %g = extractelement <32 x i1> %f, i32 01361 ret i1 %g1362}1363 1364define {i32, i1} @test_v16i8_muti_uses(<16 x i8> %x, <16 x i8>%y, <16 x i8> %z) {1365; SSE-LABEL: test_v16i8_muti_uses:1366; SSE: # %bb.0:1367; SSE-NEXT: pcmpeqb %xmm1, %xmm01368; SSE-NEXT: pcmpeqb %xmm1, %xmm21369; SSE-NEXT: pmovmskb %xmm0, %ecx1370; SSE-NEXT: pmovmskb %xmm2, %eax1371; SSE-NEXT: shll $16, %eax1372; SSE-NEXT: orl %ecx, %eax1373; SSE-NEXT: sete %dl1374; SSE-NEXT: retq1375;1376; AVX1-LABEL: test_v16i8_muti_uses:1377; AVX1: # %bb.0:1378; AVX1-NEXT: vpcmpeqb %xmm1, %xmm0, %xmm01379; AVX1-NEXT: vpcmpeqb %xmm1, %xmm2, %xmm11380; AVX1-NEXT: vpmovmskb %xmm0, %ecx1381; AVX1-NEXT: vpmovmskb %xmm1, %eax1382; AVX1-NEXT: shll $16, %eax1383; AVX1-NEXT: orl %ecx, %eax1384; AVX1-NEXT: sete %dl1385; AVX1-NEXT: retq1386;1387; AVX2-LABEL: test_v16i8_muti_uses:1388; AVX2: # %bb.0:1389; AVX2-NEXT: vpcmpeqb %xmm1, %xmm0, %xmm01390; AVX2-NEXT: vpcmpeqb %xmm1, %xmm2, %xmm11391; AVX2-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm01392; AVX2-NEXT: vpmovmskb %ymm0, %eax1393; AVX2-NEXT: testl %eax, %eax1394; AVX2-NEXT: sete %dl1395; AVX2-NEXT: vzeroupper1396; AVX2-NEXT: retq1397;1398; AVX512-LABEL: test_v16i8_muti_uses:1399; AVX512: # %bb.0:1400; AVX512-NEXT: vpcmpeqb %xmm1, %xmm0, %k01401; AVX512-NEXT: vpcmpeqb %xmm1, %xmm2, %k11402; AVX512-NEXT: kunpckwd %k0, %k1, %k01403; AVX512-NEXT: kortestd %k0, %k01404; AVX512-NEXT: kmovd %k0, %eax1405; AVX512-NEXT: sete %dl1406; AVX512-NEXT: retq1407 %t1 = icmp eq <16 x i8> %x, %y1408 %t2 = icmp eq <16 x i8> %z, %y1409 %a = shufflevector <16 x i1> %t1, <16 x i1> %t2, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>1410 %b = bitcast <32 x i1> %a to i321411 %c = icmp eq i32 %b, 01412 %r1 = insertvalue {i32, i1} poison, i32 %b, 01413 %r2 = insertvalue {i32, i1} %r1, i1 %c, 11414 ret {i32, i1} %r21415}1416 1417; PR598671418define i1 @select_v2i8(ptr %s0, ptr %s1) {1419; SSE2-LABEL: select_v2i8:1420; SSE2: # %bb.0:1421; SSE2-NEXT: movzwl (%rdi), %eax1422; SSE2-NEXT: movd %eax, %xmm01423; SSE2-NEXT: movzwl (%rsi), %eax1424; SSE2-NEXT: movd %eax, %xmm11425; SSE2-NEXT: pcmpeqb %xmm0, %xmm11426; SSE2-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]1427; SSE2-NEXT: pshuflw {{.*#+}} xmm0 = xmm1[0,0,1,1,4,5,6,7]1428; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,0,1,1]1429; SSE2-NEXT: movmskpd %xmm0, %eax1430; SSE2-NEXT: testl %eax, %eax1431; SSE2-NEXT: setne %al1432; SSE2-NEXT: retq1433;1434; SSE42-LABEL: select_v2i8:1435; SSE42: # %bb.0:1436; SSE42-NEXT: movzwl (%rdi), %eax1437; SSE42-NEXT: movd %eax, %xmm01438; SSE42-NEXT: movzwl (%rsi), %eax1439; SSE42-NEXT: movd %eax, %xmm11440; SSE42-NEXT: pcmpeqb %xmm0, %xmm11441; SSE42-NEXT: pmovsxbq %xmm1, %xmm01442; SSE42-NEXT: movmskpd %xmm0, %eax1443; SSE42-NEXT: testl %eax, %eax1444; SSE42-NEXT: setne %al1445; SSE42-NEXT: retq1446;1447; AVX1OR2-LABEL: select_v2i8:1448; AVX1OR2: # %bb.0:1449; AVX1OR2-NEXT: movzwl (%rdi), %eax1450; AVX1OR2-NEXT: vmovd %eax, %xmm01451; AVX1OR2-NEXT: movzwl (%rsi), %eax1452; AVX1OR2-NEXT: vmovd %eax, %xmm11453; AVX1OR2-NEXT: vpcmpeqb %xmm1, %xmm0, %xmm01454; AVX1OR2-NEXT: vpmovsxbq %xmm0, %xmm01455; AVX1OR2-NEXT: vtestpd %xmm0, %xmm01456; AVX1OR2-NEXT: setne %al1457; AVX1OR2-NEXT: retq1458;1459; AVX512-LABEL: select_v2i8:1460; AVX512: # %bb.0:1461; AVX512-NEXT: movzwl (%rdi), %eax1462; AVX512-NEXT: vmovd %eax, %xmm01463; AVX512-NEXT: movzwl (%rsi), %eax1464; AVX512-NEXT: vmovd %eax, %xmm11465; AVX512-NEXT: vpcmpeqb %xmm1, %xmm0, %k01466; AVX512-NEXT: kmovd %k0, %eax1467; AVX512-NEXT: testb $3, %al1468; AVX512-NEXT: setne %al1469; AVX512-NEXT: retq1470 %v0 = load <2 x i8>, ptr %s0, align 11471 %v1 = load <2 x i8>, ptr %s1, align 11472 %cmp = icmp eq <2 x i8> %v0, %v11473 %cmp0 = extractelement <2 x i1> %cmp, i32 01474 %cmp1 = extractelement <2 x i1> %cmp, i32 11475 %res = select i1 %cmp0, i1 true, i1 %cmp11476 ret i1 %res1477}1478