239 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc < %s -disable-peephole -mtriple=x86_64-unknown-unknown -mattr=+avx | FileCheck %s --check-prefix=CHECK --check-prefix=AVX3; RUN: llc < %s -disable-peephole -mtriple=x86_64-unknown-unknown -mattr=+avx512f | FileCheck %s --check-prefix=CHECK --check-prefix=AVX5124 5define <8 x float> @sitofp00(<8 x i32> %a) nounwind {6; CHECK-LABEL: sitofp00:7; CHECK: # %bb.0:8; CHECK-NEXT: vcvtdq2ps %ymm0, %ymm09; CHECK-NEXT: retq10 %b = sitofp <8 x i32> %a to <8 x float>11 ret <8 x float> %b12}13 14define <8 x i32> @fptosi00(<8 x float> %a) nounwind {15; CHECK-LABEL: fptosi00:16; CHECK: # %bb.0:17; CHECK-NEXT: vcvttps2dq %ymm0, %ymm018; CHECK-NEXT: retq19 %b = fptosi <8 x float> %a to <8 x i32>20 ret <8 x i32> %b21}22 23define <4 x double> @sitofp01(<4 x i32> %a) {24; CHECK-LABEL: sitofp01:25; CHECK: # %bb.0:26; CHECK-NEXT: vcvtdq2pd %xmm0, %ymm027; CHECK-NEXT: retq28 %b = sitofp <4 x i32> %a to <4 x double>29 ret <4 x double> %b30}31 32define <8 x float> @sitofp02(<8 x i16> %a) {33; AVX-LABEL: sitofp02:34; AVX: # %bb.0:35; AVX-NEXT: vpmovsxwd %xmm0, %xmm136; AVX-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[2,3,2,3]37; AVX-NEXT: vpmovsxwd %xmm0, %xmm038; AVX-NEXT: vinsertf128 $1, %xmm0, %ymm1, %ymm039; AVX-NEXT: vcvtdq2ps %ymm0, %ymm040; AVX-NEXT: retq41;42; AVX512-LABEL: sitofp02:43; AVX512: # %bb.0:44; AVX512-NEXT: vpmovsxwd %xmm0, %ymm045; AVX512-NEXT: vcvtdq2ps %ymm0, %ymm046; AVX512-NEXT: retq47 %b = sitofp <8 x i16> %a to <8 x float>48 ret <8 x float> %b49}50 51define <4 x i32> @fptosi01(<4 x double> %a) {52; CHECK-LABEL: fptosi01:53; CHECK: # %bb.0:54; CHECK-NEXT: vcvttpd2dq %ymm0, %xmm055; CHECK-NEXT: vzeroupper56; CHECK-NEXT: retq57 %b = fptosi <4 x double> %a to <4 x i32>58 ret <4 x i32> %b59}60 61define <8 x float> @fptrunc00(<8 x double> %b) nounwind {62; AVX-LABEL: fptrunc00:63; AVX: # %bb.0:64; AVX-NEXT: vcvtpd2ps %ymm0, %xmm065; AVX-NEXT: vcvtpd2ps %ymm1, %xmm166; AVX-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm067; AVX-NEXT: retq68;69; AVX512-LABEL: fptrunc00:70; AVX512: # %bb.0:71; AVX512-NEXT: vcvtpd2ps %zmm0, %ymm072; AVX512-NEXT: retq73 %a = fptrunc <8 x double> %b to <8 x float>74 ret <8 x float> %a75}76 77define <4 x float> @fptrunc01(<2 x double> %a0, <4 x float> %a1) nounwind {78; CHECK-LABEL: fptrunc01:79; CHECK: # %bb.0:80; CHECK-NEXT: vcvtsd2ss %xmm0, %xmm1, %xmm081; CHECK-NEXT: retq82 %ext = extractelement <2 x double> %a0, i32 083 %cvt = fptrunc double %ext to float84 %res = insertelement <4 x float> %a1, float %cvt, i32 085 ret <4 x float> %res86}87 88define <4 x double> @fpext00(<4 x float> %b) nounwind {89; CHECK-LABEL: fpext00:90; CHECK: # %bb.0:91; CHECK-NEXT: vcvtps2pd %xmm0, %ymm092; CHECK-NEXT: retq93 %a = fpext <4 x float> %b to <4 x double>94 ret <4 x double> %a95}96 97define <2 x double> @fpext01(<2 x double> %a0, <4 x float> %a1) nounwind {98; CHECK-LABEL: fpext01:99; CHECK: # %bb.0:100; CHECK-NEXT: vcvtss2sd %xmm1, %xmm0, %xmm0101; CHECK-NEXT: retq102 %ext = extractelement <4 x float> %a1, i32 0103 %cvt = fpext float %ext to double104 %res = insertelement <2 x double> %a0, double %cvt, i32 0105 ret <2 x double> %res106}107 108define double @funcA(ptr nocapture %e) nounwind uwtable readonly ssp {109; CHECK-LABEL: funcA:110; CHECK: # %bb.0:111; CHECK-NEXT: vcvtsi2sdq (%rdi), %xmm15, %xmm0112; CHECK-NEXT: retq113 %tmp1 = load i64, ptr %e, align 8114 %conv = sitofp i64 %tmp1 to double115 ret double %conv116}117 118define double @funcB(ptr nocapture %e) nounwind uwtable readonly ssp {119; CHECK-LABEL: funcB:120; CHECK: # %bb.0:121; CHECK-NEXT: vcvtsi2sdl (%rdi), %xmm15, %xmm0122; CHECK-NEXT: retq123 %tmp1 = load i32, ptr %e, align 4124 %conv = sitofp i32 %tmp1 to double125 ret double %conv126}127 128define float @funcC(ptr nocapture %e) nounwind uwtable readonly ssp {129; CHECK-LABEL: funcC:130; CHECK: # %bb.0:131; CHECK-NEXT: vcvtsi2ssl (%rdi), %xmm15, %xmm0132; CHECK-NEXT: retq133 %tmp1 = load i32, ptr %e, align 4134 %conv = sitofp i32 %tmp1 to float135 ret float %conv136}137 138define float @funcD(ptr nocapture %e) nounwind uwtable readonly ssp {139; CHECK-LABEL: funcD:140; CHECK: # %bb.0:141; CHECK-NEXT: vcvtsi2ssq (%rdi), %xmm15, %xmm0142; CHECK-NEXT: retq143 %tmp1 = load i64, ptr %e, align 8144 %conv = sitofp i64 %tmp1 to float145 ret float %conv146}147 148define void @fpext() nounwind uwtable {149; CHECK-LABEL: fpext:150; CHECK: # %bb.0:151; CHECK-NEXT: vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero152; CHECK-NEXT: vcvtss2sd %xmm0, %xmm0, %xmm0153; CHECK-NEXT: vmovsd %xmm0, -{{[0-9]+}}(%rsp)154; CHECK-NEXT: retq155 %f = alloca float, align 4156 %d = alloca double, align 8157 %tmp = load float, ptr %f, align 4158 %conv = fpext float %tmp to double159 store double %conv, ptr %d, align 8160 ret void161}162 163define double @nearbyint_f64(double %a) {164; CHECK-LABEL: nearbyint_f64:165; CHECK: # %bb.0:166; CHECK-NEXT: vroundsd $12, %xmm0, %xmm0, %xmm0167; CHECK-NEXT: retq168 %res = call double @llvm.nearbyint.f64(double %a)169 ret double %res170}171declare double @llvm.nearbyint.f64(double %p)172 173define float @floor_f32(float %a) {174; CHECK-LABEL: floor_f32:175; CHECK: # %bb.0:176; CHECK-NEXT: vroundss $9, %xmm0, %xmm0, %xmm0177; CHECK-NEXT: retq178 %res = call float @llvm.floor.f32(float %a)179 ret float %res180}181declare float @llvm.floor.f32(float %p)182 183define float @floor_f32_load(ptr %aptr) optsize {184; CHECK-LABEL: floor_f32_load:185; CHECK: # %bb.0:186; CHECK-NEXT: vroundss $9, (%rdi), %xmm15, %xmm0187; CHECK-NEXT: retq188 %a = load float, ptr %aptr189 %res = call float @llvm.floor.f32(float %a)190 ret float %res191}192 193define float @floor_f32_load_pgso(ptr %aptr) !prof !14 {194; CHECK-LABEL: floor_f32_load_pgso:195; CHECK: # %bb.0:196; CHECK-NEXT: vroundss $9, (%rdi), %xmm15, %xmm0197; CHECK-NEXT: retq198 %a = load float, ptr %aptr199 %res = call float @llvm.floor.f32(float %a)200 ret float %res201}202 203define double @nearbyint_f64_load(ptr %aptr) optsize {204; CHECK-LABEL: nearbyint_f64_load:205; CHECK: # %bb.0:206; CHECK-NEXT: vroundsd $12, (%rdi), %xmm15, %xmm0207; CHECK-NEXT: retq208 %a = load double, ptr %aptr209 %res = call double @llvm.nearbyint.f64(double %a)210 ret double %res211}212 213define double @nearbyint_f64_load_pgso(ptr %aptr) !prof !14 {214; CHECK-LABEL: nearbyint_f64_load_pgso:215; CHECK: # %bb.0:216; CHECK-NEXT: vroundsd $12, (%rdi), %xmm15, %xmm0217; CHECK-NEXT: retq218 %a = load double, ptr %aptr219 %res = call double @llvm.nearbyint.f64(double %a)220 ret double %res221}222 223!llvm.module.flags = !{!0}224!0 = !{i32 1, !"ProfileSummary", !1}225!1 = !{!2, !3, !4, !5, !6, !7, !8, !9}226!2 = !{!"ProfileFormat", !"InstrProf"}227!3 = !{!"TotalCount", i64 10000}228!4 = !{!"MaxCount", i64 10}229!5 = !{!"MaxInternalCount", i64 1}230!6 = !{!"MaxFunctionCount", i64 1000}231!7 = !{!"NumCounts", i64 3}232!8 = !{!"NumFunctions", i64 3}233!9 = !{!"DetailedSummary", !10}234!10 = !{!11, !12, !13}235!11 = !{i32 10000, i64 100, i32 1}236!12 = !{i32 999000, i64 100, i32 1}237!13 = !{i32 999999, i64 1, i32 2}238!14 = !{!"function_entry_count", i64 0}239