123 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py2; RUN: opt < %s -passes=instcombine -S | FileCheck %s3target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"4 5define <4 x float> @test_vfmadd_ss(<4 x float> %a, <4 x float> %b, <4 x float> %c) {6;7; CHECK-LABEL: @test_vfmadd_ss(8; CHECK-NEXT: [[TMP1:%.*]] = extractelement <4 x float> [[A:%.*]], i64 09; CHECK-NEXT: [[TMP2:%.*]] = extractelement <4 x float> [[B:%.*]], i64 010; CHECK-NEXT: [[TMP3:%.*]] = extractelement <4 x float> [[C:%.*]], i64 011; CHECK-NEXT: [[TMP4:%.*]] = call float @llvm.fma.f32(float [[TMP1]], float [[TMP2]], float [[TMP3]])12; CHECK-NEXT: [[TMP5:%.*]] = insertelement <4 x float> [[A]], float [[TMP4]], i64 013; CHECK-NEXT: ret <4 x float> [[TMP5]]14;15 %1 = insertelement <4 x float> %b, float 1.000000e+00, i32 116 %2 = insertelement <4 x float> %1, float 2.000000e+00, i32 217 %3 = insertelement <4 x float> %2, float 3.000000e+00, i32 318 %4 = insertelement <4 x float> %c, float 4.000000e+00, i32 119 %5 = insertelement <4 x float> %4, float 5.000000e+00, i32 220 %6 = insertelement <4 x float> %5, float 6.000000e+00, i32 321 %7 = extractelement <4 x float> %a, i64 022 %8 = extractelement <4 x float> %3, i64 023 %9 = extractelement <4 x float> %6, i64 024 %10 = call float @llvm.fma.f32(float %7, float %8, float %9)25 %11 = insertelement <4 x float> %a, float %10, i64 026 ret <4 x float> %1127}28 29define float @test_vfmadd_ss_0(<4 x float> %a, <4 x float> %b, <4 x float> %c) {30;31; CHECK-LABEL: @test_vfmadd_ss_0(32; CHECK-NEXT: [[TMP1:%.*]] = extractelement <4 x float> [[A:%.*]], i64 033; CHECK-NEXT: [[TMP2:%.*]] = extractelement <4 x float> [[B:%.*]], i64 034; CHECK-NEXT: [[TMP3:%.*]] = extractelement <4 x float> [[C:%.*]], i64 035; CHECK-NEXT: [[TMP4:%.*]] = call float @llvm.fma.f32(float [[TMP1]], float [[TMP2]], float [[TMP3]])36; CHECK-NEXT: ret float [[TMP4]]37;38 %1 = insertelement <4 x float> %a, float 1.000000e+00, i32 139 %2 = insertelement <4 x float> %1, float 2.000000e+00, i32 240 %3 = insertelement <4 x float> %2, float 3.000000e+00, i32 341 %4 = extractelement <4 x float> %3, i64 042 %5 = extractelement <4 x float> %b, i64 043 %6 = extractelement <4 x float> %c, i64 044 %7 = call float @llvm.fma.f32(float %4, float %5, float %6)45 %8 = insertelement <4 x float> %3, float %7, i64 046 %9 = extractelement <4 x float> %8, i32 047 ret float %948}49 50define float @test_vfmadd_ss_1(<4 x float> %a, <4 x float> %b, <4 x float> %c) {51;52; CHECK-LABEL: @test_vfmadd_ss_1(53; CHECK-NEXT: ret float 1.000000e+0054;55 %1 = insertelement <4 x float> %a, float 1.000000e+00, i32 156 %2 = insertelement <4 x float> %1, float 2.000000e+00, i32 257 %3 = insertelement <4 x float> %2, float 3.000000e+00, i32 358 %4 = extractelement <4 x float> %3, i64 059 %5 = extractelement <4 x float> %b, i64 060 %6 = extractelement <4 x float> %c, i64 061 %7 = call float @llvm.fma.f32(float %4, float %5, float %6)62 %8 = insertelement <4 x float> %3, float %7, i64 063 %9 = extractelement <4 x float> %8, i32 164 ret float %965}66 67define <2 x double> @test_vfmadd_sd(<2 x double> %a, <2 x double> %b, <2 x double> %c) {68;69; CHECK-LABEL: @test_vfmadd_sd(70; CHECK-NEXT: [[TMP1:%.*]] = extractelement <2 x double> [[A:%.*]], i64 071; CHECK-NEXT: [[TMP2:%.*]] = extractelement <2 x double> [[B:%.*]], i64 072; CHECK-NEXT: [[TMP3:%.*]] = extractelement <2 x double> [[C:%.*]], i64 073; CHECK-NEXT: [[TMP4:%.*]] = call double @llvm.fma.f64(double [[TMP1]], double [[TMP2]], double [[TMP3]])74; CHECK-NEXT: [[TMP5:%.*]] = insertelement <2 x double> [[A]], double [[TMP4]], i64 075; CHECK-NEXT: ret <2 x double> [[TMP5]]76;77 %1 = insertelement <2 x double> %b, double 1.000000e+00, i32 178 %2 = insertelement <2 x double> %c, double 2.000000e+00, i32 179 %3 = extractelement <2 x double> %a, i64 080 %4 = extractelement <2 x double> %1, i64 081 %5 = extractelement <2 x double> %2, i64 082 %6 = call double @llvm.fma.f64(double %3, double %4, double %5)83 %7 = insertelement <2 x double> %a, double %6, i64 084 ret <2 x double> %785}86 87define double @test_vfmadd_sd_0(<2 x double> %a, <2 x double> %b, <2 x double> %c) {88;89; CHECK-LABEL: @test_vfmadd_sd_0(90; CHECK-NEXT: [[TMP1:%.*]] = extractelement <2 x double> [[A:%.*]], i64 091; CHECK-NEXT: [[TMP2:%.*]] = extractelement <2 x double> [[B:%.*]], i64 092; CHECK-NEXT: [[TMP3:%.*]] = extractelement <2 x double> [[C:%.*]], i64 093; CHECK-NEXT: [[TMP4:%.*]] = call double @llvm.fma.f64(double [[TMP1]], double [[TMP2]], double [[TMP3]])94; CHECK-NEXT: ret double [[TMP4]]95;96 %1 = insertelement <2 x double> %a, double 1.000000e+00, i32 197 %2 = extractelement <2 x double> %1, i64 098 %3 = extractelement <2 x double> %b, i64 099 %4 = extractelement <2 x double> %c, i64 0100 %5 = call double @llvm.fma.f64(double %2, double %3, double %4)101 %6 = insertelement <2 x double> %1, double %5, i64 0102 %7 = extractelement <2 x double> %6, i32 0103 ret double %7104}105 106define double @test_vfmadd_sd_1(<2 x double> %a, <2 x double> %b, <2 x double> %c) {107;108; CHECK-LABEL: @test_vfmadd_sd_1(109; CHECK-NEXT: ret double 1.000000e+00110;111 %1 = insertelement <2 x double> %a, double 1.000000e+00, i32 1112 %2 = extractelement <2 x double> %1, i64 0113 %3 = extractelement <2 x double> %b, i64 0114 %4 = extractelement <2 x double> %c, i64 0115 %5 = call double @llvm.fma.f64(double %2, double %3, double %4)116 %6 = insertelement <2 x double> %1, double %5, i64 0117 %7 = extractelement <2 x double> %6, i32 1118 ret double %7119}120 121declare float @llvm.fma.f32(float, float, float)122declare double @llvm.fma.f64(double, double, double)123