564 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py2; RUN: opt < %s -mtriple=x86_64-unknown -passes=slp-vectorizer -S | FileCheck %s --check-prefix=NO-FMA3; RUN: opt < %s -mtriple=x86_64-unknown -mcpu=corei7-avx -passes=slp-vectorizer -S | FileCheck %s --check-prefix=NO-FMA4; RUN: opt < %s -mtriple=x86_64-unknown -mcpu=bdver1 -passes=slp-vectorizer -S | FileCheck %s --check-prefix=FMA --check-prefix=FMA2565; RUN: opt < %s -mtriple=x86_64-unknown -mcpu=core-avx2 -passes=slp-vectorizer -S | FileCheck %s --check-prefix=FMA --check-prefix=FMA2566; RUN: opt < %s -mtriple=x86_64-unknown -mcpu=skylake-avx512 -mattr=-prefer-256-bit -passes=slp-vectorizer -S | FileCheck %s --check-prefix=FMA --check-prefix=FMA5127; RUN: opt < %s -mtriple=x86_64-unknown -mcpu=skylake-avx512 -mattr=+prefer-256-bit -passes=slp-vectorizer -S | FileCheck %s --check-prefix=FMA --check-prefix=FMA2568 9target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"10 11@srcA64 = common global [8 x double] zeroinitializer, align 6412@srcB64 = common global [8 x double] zeroinitializer, align 6413@srcC64 = common global [8 x double] zeroinitializer, align 6414@srcA32 = common global [16 x float] zeroinitializer, align 6415@srcB32 = common global [16 x float] zeroinitializer, align 6416@srcC32 = common global [16 x float] zeroinitializer, align 6417@dst64 = common global [8 x double] zeroinitializer, align 6418@dst32 = common global [16 x float] zeroinitializer, align 6419 20declare float @llvm.fma.f32(float, float, float)21declare double @llvm.fma.f64(double, double, double)22 23;24; FMA25;26 27define void @fma_2f64() #0 {28; NO-FMA-LABEL: @fma_2f64(29; NO-FMA-NEXT: [[A0:%.*]] = load double, ptr @srcA64, align 830; NO-FMA-NEXT: [[A1:%.*]] = load double, ptr getelementptr inbounds ([8 x double], ptr @srcA64, i32 0, i64 1), align 831; NO-FMA-NEXT: [[B0:%.*]] = load double, ptr @srcB64, align 832; NO-FMA-NEXT: [[B1:%.*]] = load double, ptr getelementptr inbounds ([8 x double], ptr @srcB64, i32 0, i64 1), align 833; NO-FMA-NEXT: [[C0:%.*]] = load double, ptr @srcC64, align 834; NO-FMA-NEXT: [[C1:%.*]] = load double, ptr getelementptr inbounds ([8 x double], ptr @srcC64, i32 0, i64 1), align 835; NO-FMA-NEXT: [[FMA0:%.*]] = call double @llvm.fma.f64(double [[A0]], double [[B0]], double [[C0]])36; NO-FMA-NEXT: [[FMA1:%.*]] = call double @llvm.fma.f64(double [[A1]], double [[B1]], double [[C1]])37; NO-FMA-NEXT: store double [[FMA0]], ptr @dst64, align 838; NO-FMA-NEXT: store double [[FMA1]], ptr getelementptr inbounds ([8 x double], ptr @dst64, i32 0, i64 1), align 839; NO-FMA-NEXT: ret void40;41; FMA-LABEL: @fma_2f64(42; FMA-NEXT: [[TMP1:%.*]] = load <2 x double>, ptr @srcA64, align 843; FMA-NEXT: [[TMP2:%.*]] = load <2 x double>, ptr @srcB64, align 844; FMA-NEXT: [[TMP3:%.*]] = load <2 x double>, ptr @srcC64, align 845; FMA-NEXT: [[TMP4:%.*]] = call <2 x double> @llvm.fma.v2f64(<2 x double> [[TMP1]], <2 x double> [[TMP2]], <2 x double> [[TMP3]])46; FMA-NEXT: store <2 x double> [[TMP4]], ptr @dst64, align 847; FMA-NEXT: ret void48;49 %a0 = load double, ptr @srcA64, align 850 %a1 = load double, ptr getelementptr inbounds ([8 x double], ptr @srcA64, i32 0, i64 1), align 851 %b0 = load double, ptr @srcB64, align 852 %b1 = load double, ptr getelementptr inbounds ([8 x double], ptr @srcB64, i32 0, i64 1), align 853 %c0 = load double, ptr @srcC64, align 854 %c1 = load double, ptr getelementptr inbounds ([8 x double], ptr @srcC64, i32 0, i64 1), align 855 %fma0 = call double @llvm.fma.f64(double %a0, double %b0, double %c0)56 %fma1 = call double @llvm.fma.f64(double %a1, double %b1, double %c1)57 store double %fma0, ptr @dst64, align 858 store double %fma1, ptr getelementptr inbounds ([8 x double], ptr @dst64, i32 0, i64 1), align 859 ret void60}61 62define void @fma_4f64() #0 {63; NO-FMA-LABEL: @fma_4f64(64; NO-FMA-NEXT: [[A0:%.*]] = load double, ptr @srcA64, align 865; NO-FMA-NEXT: [[A1:%.*]] = load double, ptr getelementptr inbounds ([8 x double], ptr @srcA64, i32 0, i64 1), align 866; NO-FMA-NEXT: [[A2:%.*]] = load double, ptr getelementptr inbounds ([8 x double], ptr @srcA64, i32 0, i64 2), align 867; NO-FMA-NEXT: [[A3:%.*]] = load double, ptr getelementptr inbounds ([8 x double], ptr @srcA64, i32 0, i64 3), align 868; NO-FMA-NEXT: [[B0:%.*]] = load double, ptr @srcB64, align 869; NO-FMA-NEXT: [[B1:%.*]] = load double, ptr getelementptr inbounds ([8 x double], ptr @srcB64, i32 0, i64 1), align 870; NO-FMA-NEXT: [[B2:%.*]] = load double, ptr getelementptr inbounds ([8 x double], ptr @srcB64, i32 0, i64 2), align 871; NO-FMA-NEXT: [[B3:%.*]] = load double, ptr getelementptr inbounds ([8 x double], ptr @srcB64, i32 0, i64 3), align 872; NO-FMA-NEXT: [[C0:%.*]] = load double, ptr @srcC64, align 873; NO-FMA-NEXT: [[C1:%.*]] = load double, ptr getelementptr inbounds ([8 x double], ptr @srcC64, i32 0, i64 1), align 874; NO-FMA-NEXT: [[C2:%.*]] = load double, ptr getelementptr inbounds ([8 x double], ptr @srcC64, i32 0, i64 2), align 875; NO-FMA-NEXT: [[C3:%.*]] = load double, ptr getelementptr inbounds ([8 x double], ptr @srcC64, i32 0, i64 3), align 876; NO-FMA-NEXT: [[FMA0:%.*]] = call double @llvm.fma.f64(double [[A0]], double [[B0]], double [[C0]])77; NO-FMA-NEXT: [[FMA1:%.*]] = call double @llvm.fma.f64(double [[A1]], double [[B1]], double [[C1]])78; NO-FMA-NEXT: [[FMA2:%.*]] = call double @llvm.fma.f64(double [[A2]], double [[B2]], double [[C2]])79; NO-FMA-NEXT: [[FMA3:%.*]] = call double @llvm.fma.f64(double [[A3]], double [[B3]], double [[C3]])80; NO-FMA-NEXT: store double [[FMA0]], ptr @dst64, align 881; NO-FMA-NEXT: store double [[FMA1]], ptr getelementptr inbounds ([8 x double], ptr @dst64, i32 0, i64 1), align 882; NO-FMA-NEXT: store double [[FMA2]], ptr getelementptr inbounds ([8 x double], ptr @dst64, i32 0, i64 2), align 883; NO-FMA-NEXT: store double [[FMA3]], ptr getelementptr inbounds ([8 x double], ptr @dst64, i32 0, i64 3), align 884; NO-FMA-NEXT: ret void85;86; FMA-LABEL: @fma_4f64(87; FMA-NEXT: [[TMP1:%.*]] = load <4 x double>, ptr @srcA64, align 888; FMA-NEXT: [[TMP2:%.*]] = load <4 x double>, ptr @srcB64, align 889; FMA-NEXT: [[TMP3:%.*]] = load <4 x double>, ptr @srcC64, align 890; FMA-NEXT: [[TMP4:%.*]] = call <4 x double> @llvm.fma.v4f64(<4 x double> [[TMP1]], <4 x double> [[TMP2]], <4 x double> [[TMP3]])91; FMA-NEXT: store <4 x double> [[TMP4]], ptr @dst64, align 892; FMA-NEXT: ret void93;94 %a0 = load double, ptr @srcA64, align 895 %a1 = load double, ptr getelementptr inbounds ([8 x double], ptr @srcA64, i32 0, i64 1), align 896 %a2 = load double, ptr getelementptr inbounds ([8 x double], ptr @srcA64, i32 0, i64 2), align 897 %a3 = load double, ptr getelementptr inbounds ([8 x double], ptr @srcA64, i32 0, i64 3), align 898 %b0 = load double, ptr @srcB64, align 899 %b1 = load double, ptr getelementptr inbounds ([8 x double], ptr @srcB64, i32 0, i64 1), align 8100 %b2 = load double, ptr getelementptr inbounds ([8 x double], ptr @srcB64, i32 0, i64 2), align 8101 %b3 = load double, ptr getelementptr inbounds ([8 x double], ptr @srcB64, i32 0, i64 3), align 8102 %c0 = load double, ptr @srcC64, align 8103 %c1 = load double, ptr getelementptr inbounds ([8 x double], ptr @srcC64, i32 0, i64 1), align 8104 %c2 = load double, ptr getelementptr inbounds ([8 x double], ptr @srcC64, i32 0, i64 2), align 8105 %c3 = load double, ptr getelementptr inbounds ([8 x double], ptr @srcC64, i32 0, i64 3), align 8106 %fma0 = call double @llvm.fma.f64(double %a0, double %b0, double %c0)107 %fma1 = call double @llvm.fma.f64(double %a1, double %b1, double %c1)108 %fma2 = call double @llvm.fma.f64(double %a2, double %b2, double %c2)109 %fma3 = call double @llvm.fma.f64(double %a3, double %b3, double %c3)110 store double %fma0, ptr @dst64, align 8111 store double %fma1, ptr getelementptr inbounds ([8 x double], ptr @dst64, i32 0, i64 1), align 8112 store double %fma2, ptr getelementptr inbounds ([8 x double], ptr @dst64, i32 0, i64 2), align 8113 store double %fma3, ptr getelementptr inbounds ([8 x double], ptr @dst64, i32 0, i64 3), align 8114 ret void115}116 117define void @fma_8f64() #0 {118; NO-FMA-LABEL: @fma_8f64(119; NO-FMA-NEXT: [[A0:%.*]] = load double, ptr @srcA64, align 4120; NO-FMA-NEXT: [[A1:%.*]] = load double, ptr getelementptr inbounds ([8 x double], ptr @srcA64, i32 0, i64 1), align 4121; NO-FMA-NEXT: [[A2:%.*]] = load double, ptr getelementptr inbounds ([8 x double], ptr @srcA64, i32 0, i64 2), align 4122; NO-FMA-NEXT: [[A3:%.*]] = load double, ptr getelementptr inbounds ([8 x double], ptr @srcA64, i32 0, i64 3), align 4123; NO-FMA-NEXT: [[A4:%.*]] = load double, ptr getelementptr inbounds ([8 x double], ptr @srcA64, i32 0, i64 4), align 4124; NO-FMA-NEXT: [[A5:%.*]] = load double, ptr getelementptr inbounds ([8 x double], ptr @srcA64, i32 0, i64 5), align 4125; NO-FMA-NEXT: [[A6:%.*]] = load double, ptr getelementptr inbounds ([8 x double], ptr @srcA64, i32 0, i64 6), align 4126; NO-FMA-NEXT: [[A7:%.*]] = load double, ptr getelementptr inbounds ([8 x double], ptr @srcA64, i32 0, i64 7), align 4127; NO-FMA-NEXT: [[B0:%.*]] = load double, ptr @srcB64, align 4128; NO-FMA-NEXT: [[B1:%.*]] = load double, ptr getelementptr inbounds ([8 x double], ptr @srcB64, i32 0, i64 1), align 4129; NO-FMA-NEXT: [[B2:%.*]] = load double, ptr getelementptr inbounds ([8 x double], ptr @srcB64, i32 0, i64 2), align 4130; NO-FMA-NEXT: [[B3:%.*]] = load double, ptr getelementptr inbounds ([8 x double], ptr @srcB64, i32 0, i64 3), align 4131; NO-FMA-NEXT: [[B4:%.*]] = load double, ptr getelementptr inbounds ([8 x double], ptr @srcB64, i32 0, i64 4), align 4132; NO-FMA-NEXT: [[B5:%.*]] = load double, ptr getelementptr inbounds ([8 x double], ptr @srcB64, i32 0, i64 5), align 4133; NO-FMA-NEXT: [[B6:%.*]] = load double, ptr getelementptr inbounds ([8 x double], ptr @srcB64, i32 0, i64 6), align 4134; NO-FMA-NEXT: [[B7:%.*]] = load double, ptr getelementptr inbounds ([8 x double], ptr @srcB64, i32 0, i64 7), align 4135; NO-FMA-NEXT: [[C0:%.*]] = load double, ptr @srcC64, align 4136; NO-FMA-NEXT: [[C1:%.*]] = load double, ptr getelementptr inbounds ([8 x double], ptr @srcC64, i32 0, i64 1), align 4137; NO-FMA-NEXT: [[C2:%.*]] = load double, ptr getelementptr inbounds ([8 x double], ptr @srcC64, i32 0, i64 2), align 4138; NO-FMA-NEXT: [[C3:%.*]] = load double, ptr getelementptr inbounds ([8 x double], ptr @srcC64, i32 0, i64 3), align 4139; NO-FMA-NEXT: [[C4:%.*]] = load double, ptr getelementptr inbounds ([8 x double], ptr @srcC64, i32 0, i64 4), align 4140; NO-FMA-NEXT: [[C5:%.*]] = load double, ptr getelementptr inbounds ([8 x double], ptr @srcC64, i32 0, i64 5), align 4141; NO-FMA-NEXT: [[C6:%.*]] = load double, ptr getelementptr inbounds ([8 x double], ptr @srcC64, i32 0, i64 6), align 4142; NO-FMA-NEXT: [[C7:%.*]] = load double, ptr getelementptr inbounds ([8 x double], ptr @srcC64, i32 0, i64 7), align 4143; NO-FMA-NEXT: [[FMA0:%.*]] = call double @llvm.fma.f64(double [[A0]], double [[B0]], double [[C0]])144; NO-FMA-NEXT: [[FMA1:%.*]] = call double @llvm.fma.f64(double [[A1]], double [[B1]], double [[C1]])145; NO-FMA-NEXT: [[FMA2:%.*]] = call double @llvm.fma.f64(double [[A2]], double [[B2]], double [[C2]])146; NO-FMA-NEXT: [[FMA3:%.*]] = call double @llvm.fma.f64(double [[A3]], double [[B3]], double [[C3]])147; NO-FMA-NEXT: [[FMA4:%.*]] = call double @llvm.fma.f64(double [[A4]], double [[B4]], double [[C4]])148; NO-FMA-NEXT: [[FMA5:%.*]] = call double @llvm.fma.f64(double [[A5]], double [[B5]], double [[C5]])149; NO-FMA-NEXT: [[FMA6:%.*]] = call double @llvm.fma.f64(double [[A6]], double [[B6]], double [[C6]])150; NO-FMA-NEXT: [[FMA7:%.*]] = call double @llvm.fma.f64(double [[A7]], double [[B7]], double [[C7]])151; NO-FMA-NEXT: store double [[FMA0]], ptr @dst64, align 4152; NO-FMA-NEXT: store double [[FMA1]], ptr getelementptr inbounds ([8 x double], ptr @dst64, i32 0, i64 1), align 4153; NO-FMA-NEXT: store double [[FMA2]], ptr getelementptr inbounds ([8 x double], ptr @dst64, i32 0, i64 2), align 4154; NO-FMA-NEXT: store double [[FMA3]], ptr getelementptr inbounds ([8 x double], ptr @dst64, i32 0, i64 3), align 4155; NO-FMA-NEXT: store double [[FMA4]], ptr getelementptr inbounds ([8 x double], ptr @dst64, i32 0, i64 4), align 4156; NO-FMA-NEXT: store double [[FMA5]], ptr getelementptr inbounds ([8 x double], ptr @dst64, i32 0, i64 5), align 4157; NO-FMA-NEXT: store double [[FMA6]], ptr getelementptr inbounds ([8 x double], ptr @dst64, i32 0, i64 6), align 4158; NO-FMA-NEXT: store double [[FMA7]], ptr getelementptr inbounds ([8 x double], ptr @dst64, i32 0, i64 7), align 4159; NO-FMA-NEXT: ret void160;161; FMA256-LABEL: @fma_8f64(162; FMA256-NEXT: [[TMP1:%.*]] = load <4 x double>, ptr @srcA64, align 4163; FMA256-NEXT: [[TMP2:%.*]] = load <4 x double>, ptr @srcB64, align 4164; FMA256-NEXT: [[TMP3:%.*]] = load <4 x double>, ptr @srcC64, align 4165; FMA256-NEXT: [[TMP4:%.*]] = call <4 x double> @llvm.fma.v4f64(<4 x double> [[TMP1]], <4 x double> [[TMP2]], <4 x double> [[TMP3]])166; FMA256-NEXT: store <4 x double> [[TMP4]], ptr @dst64, align 4167; FMA256-NEXT: [[TMP5:%.*]] = load <4 x double>, ptr getelementptr inbounds ([8 x double], ptr @srcA64, i32 0, i64 4), align 4168; FMA256-NEXT: [[TMP6:%.*]] = load <4 x double>, ptr getelementptr inbounds ([8 x double], ptr @srcB64, i32 0, i64 4), align 4169; FMA256-NEXT: [[TMP7:%.*]] = load <4 x double>, ptr getelementptr inbounds ([8 x double], ptr @srcC64, i32 0, i64 4), align 4170; FMA256-NEXT: [[TMP8:%.*]] = call <4 x double> @llvm.fma.v4f64(<4 x double> [[TMP5]], <4 x double> [[TMP6]], <4 x double> [[TMP7]])171; FMA256-NEXT: store <4 x double> [[TMP8]], ptr getelementptr inbounds ([8 x double], ptr @dst64, i32 0, i64 4), align 4172; FMA256-NEXT: ret void173;174; FMA512-LABEL: @fma_8f64(175; FMA512-NEXT: [[TMP1:%.*]] = load <8 x double>, ptr @srcA64, align 4176; FMA512-NEXT: [[TMP2:%.*]] = load <8 x double>, ptr @srcB64, align 4177; FMA512-NEXT: [[TMP3:%.*]] = load <8 x double>, ptr @srcC64, align 4178; FMA512-NEXT: [[TMP4:%.*]] = call <8 x double> @llvm.fma.v8f64(<8 x double> [[TMP1]], <8 x double> [[TMP2]], <8 x double> [[TMP3]])179; FMA512-NEXT: store <8 x double> [[TMP4]], ptr @dst64, align 4180; FMA512-NEXT: ret void181;182 %a0 = load double, ptr @srcA64, align 4183 %a1 = load double, ptr getelementptr inbounds ([8 x double], ptr @srcA64, i32 0, i64 1), align 4184 %a2 = load double, ptr getelementptr inbounds ([8 x double], ptr @srcA64, i32 0, i64 2), align 4185 %a3 = load double, ptr getelementptr inbounds ([8 x double], ptr @srcA64, i32 0, i64 3), align 4186 %a4 = load double, ptr getelementptr inbounds ([8 x double], ptr @srcA64, i32 0, i64 4), align 4187 %a5 = load double, ptr getelementptr inbounds ([8 x double], ptr @srcA64, i32 0, i64 5), align 4188 %a6 = load double, ptr getelementptr inbounds ([8 x double], ptr @srcA64, i32 0, i64 6), align 4189 %a7 = load double, ptr getelementptr inbounds ([8 x double], ptr @srcA64, i32 0, i64 7), align 4190 %b0 = load double, ptr @srcB64, align 4191 %b1 = load double, ptr getelementptr inbounds ([8 x double], ptr @srcB64, i32 0, i64 1), align 4192 %b2 = load double, ptr getelementptr inbounds ([8 x double], ptr @srcB64, i32 0, i64 2), align 4193 %b3 = load double, ptr getelementptr inbounds ([8 x double], ptr @srcB64, i32 0, i64 3), align 4194 %b4 = load double, ptr getelementptr inbounds ([8 x double], ptr @srcB64, i32 0, i64 4), align 4195 %b5 = load double, ptr getelementptr inbounds ([8 x double], ptr @srcB64, i32 0, i64 5), align 4196 %b6 = load double, ptr getelementptr inbounds ([8 x double], ptr @srcB64, i32 0, i64 6), align 4197 %b7 = load double, ptr getelementptr inbounds ([8 x double], ptr @srcB64, i32 0, i64 7), align 4198 %c0 = load double, ptr @srcC64, align 4199 %c1 = load double, ptr getelementptr inbounds ([8 x double], ptr @srcC64, i32 0, i64 1), align 4200 %c2 = load double, ptr getelementptr inbounds ([8 x double], ptr @srcC64, i32 0, i64 2), align 4201 %c3 = load double, ptr getelementptr inbounds ([8 x double], ptr @srcC64, i32 0, i64 3), align 4202 %c4 = load double, ptr getelementptr inbounds ([8 x double], ptr @srcC64, i32 0, i64 4), align 4203 %c5 = load double, ptr getelementptr inbounds ([8 x double], ptr @srcC64, i32 0, i64 5), align 4204 %c6 = load double, ptr getelementptr inbounds ([8 x double], ptr @srcC64, i32 0, i64 6), align 4205 %c7 = load double, ptr getelementptr inbounds ([8 x double], ptr @srcC64, i32 0, i64 7), align 4206 %fma0 = call double @llvm.fma.f64(double %a0, double %b0, double %c0)207 %fma1 = call double @llvm.fma.f64(double %a1, double %b1, double %c1)208 %fma2 = call double @llvm.fma.f64(double %a2, double %b2, double %c2)209 %fma3 = call double @llvm.fma.f64(double %a3, double %b3, double %c3)210 %fma4 = call double @llvm.fma.f64(double %a4, double %b4, double %c4)211 %fma5 = call double @llvm.fma.f64(double %a5, double %b5, double %c5)212 %fma6 = call double @llvm.fma.f64(double %a6, double %b6, double %c6)213 %fma7 = call double @llvm.fma.f64(double %a7, double %b7, double %c7)214 store double %fma0, ptr @dst64, align 4215 store double %fma1, ptr getelementptr inbounds ([8 x double], ptr @dst64, i32 0, i64 1), align 4216 store double %fma2, ptr getelementptr inbounds ([8 x double], ptr @dst64, i32 0, i64 2), align 4217 store double %fma3, ptr getelementptr inbounds ([8 x double], ptr @dst64, i32 0, i64 3), align 4218 store double %fma4, ptr getelementptr inbounds ([8 x double], ptr @dst64, i32 0, i64 4), align 4219 store double %fma5, ptr getelementptr inbounds ([8 x double], ptr @dst64, i32 0, i64 5), align 4220 store double %fma6, ptr getelementptr inbounds ([8 x double], ptr @dst64, i32 0, i64 6), align 4221 store double %fma7, ptr getelementptr inbounds ([8 x double], ptr @dst64, i32 0, i64 7), align 4222 ret void223}224 225define void @fma_4f32() #0 {226; NO-FMA-LABEL: @fma_4f32(227; NO-FMA-NEXT: [[A0:%.*]] = load float, ptr @srcA32, align 4228; NO-FMA-NEXT: [[A1:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcA32, i32 0, i64 1), align 4229; NO-FMA-NEXT: [[A2:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcA32, i32 0, i64 2), align 4230; NO-FMA-NEXT: [[A3:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcA32, i32 0, i64 3), align 4231; NO-FMA-NEXT: [[B0:%.*]] = load float, ptr @srcB32, align 4232; NO-FMA-NEXT: [[B1:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcB32, i32 0, i64 1), align 4233; NO-FMA-NEXT: [[B2:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcB32, i32 0, i64 2), align 4234; NO-FMA-NEXT: [[B3:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcB32, i32 0, i64 3), align 4235; NO-FMA-NEXT: [[C0:%.*]] = load float, ptr @srcC32, align 4236; NO-FMA-NEXT: [[C1:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcC32, i32 0, i64 1), align 4237; NO-FMA-NEXT: [[C2:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcC32, i32 0, i64 2), align 4238; NO-FMA-NEXT: [[C3:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcC32, i32 0, i64 3), align 4239; NO-FMA-NEXT: [[FMA0:%.*]] = call float @llvm.fma.f32(float [[A0]], float [[B0]], float [[C0]])240; NO-FMA-NEXT: [[FMA1:%.*]] = call float @llvm.fma.f32(float [[A1]], float [[B1]], float [[C1]])241; NO-FMA-NEXT: [[FMA2:%.*]] = call float @llvm.fma.f32(float [[A2]], float [[B2]], float [[C2]])242; NO-FMA-NEXT: [[FMA3:%.*]] = call float @llvm.fma.f32(float [[A3]], float [[B3]], float [[C3]])243; NO-FMA-NEXT: store float [[FMA0]], ptr @dst32, align 4244; NO-FMA-NEXT: store float [[FMA1]], ptr getelementptr inbounds ([16 x float], ptr @dst32, i32 0, i64 1), align 4245; NO-FMA-NEXT: store float [[FMA2]], ptr getelementptr inbounds ([16 x float], ptr @dst32, i32 0, i64 2), align 4246; NO-FMA-NEXT: store float [[FMA3]], ptr getelementptr inbounds ([16 x float], ptr @dst32, i32 0, i64 3), align 4247; NO-FMA-NEXT: ret void248;249; FMA-LABEL: @fma_4f32(250; FMA-NEXT: [[TMP1:%.*]] = load <4 x float>, ptr @srcA32, align 4251; FMA-NEXT: [[TMP2:%.*]] = load <4 x float>, ptr @srcB32, align 4252; FMA-NEXT: [[TMP3:%.*]] = load <4 x float>, ptr @srcC32, align 4253; FMA-NEXT: [[TMP4:%.*]] = call <4 x float> @llvm.fma.v4f32(<4 x float> [[TMP1]], <4 x float> [[TMP2]], <4 x float> [[TMP3]])254; FMA-NEXT: store <4 x float> [[TMP4]], ptr @dst32, align 4255; FMA-NEXT: ret void256;257 %a0 = load float, ptr @srcA32, align 4258 %a1 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcA32, i32 0, i64 1), align 4259 %a2 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcA32, i32 0, i64 2), align 4260 %a3 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcA32, i32 0, i64 3), align 4261 %b0 = load float, ptr @srcB32, align 4262 %b1 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcB32, i32 0, i64 1), align 4263 %b2 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcB32, i32 0, i64 2), align 4264 %b3 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcB32, i32 0, i64 3), align 4265 %c0 = load float, ptr @srcC32, align 4266 %c1 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcC32, i32 0, i64 1), align 4267 %c2 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcC32, i32 0, i64 2), align 4268 %c3 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcC32, i32 0, i64 3), align 4269 %fma0 = call float @llvm.fma.f32(float %a0, float %b0, float %c0)270 %fma1 = call float @llvm.fma.f32(float %a1, float %b1, float %c1)271 %fma2 = call float @llvm.fma.f32(float %a2, float %b2, float %c2)272 %fma3 = call float @llvm.fma.f32(float %a3, float %b3, float %c3)273 store float %fma0, ptr @dst32, align 4274 store float %fma1, ptr getelementptr inbounds ([16 x float], ptr @dst32, i32 0, i64 1), align 4275 store float %fma2, ptr getelementptr inbounds ([16 x float], ptr @dst32, i32 0, i64 2), align 4276 store float %fma3, ptr getelementptr inbounds ([16 x float], ptr @dst32, i32 0, i64 3), align 4277 ret void278}279 280define void @fma_8f32() #0 {281; NO-FMA-LABEL: @fma_8f32(282; NO-FMA-NEXT: [[A0:%.*]] = load float, ptr @srcA32, align 4283; NO-FMA-NEXT: [[A1:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcA32, i32 0, i64 1), align 4284; NO-FMA-NEXT: [[A2:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcA32, i32 0, i64 2), align 4285; NO-FMA-NEXT: [[A3:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcA32, i32 0, i64 3), align 4286; NO-FMA-NEXT: [[A4:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcA32, i32 0, i64 4), align 4287; NO-FMA-NEXT: [[A5:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcA32, i32 0, i64 5), align 4288; NO-FMA-NEXT: [[A6:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcA32, i32 0, i64 6), align 4289; NO-FMA-NEXT: [[A7:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcA32, i32 0, i64 7), align 4290; NO-FMA-NEXT: [[B0:%.*]] = load float, ptr @srcB32, align 4291; NO-FMA-NEXT: [[B1:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcB32, i32 0, i64 1), align 4292; NO-FMA-NEXT: [[B2:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcB32, i32 0, i64 2), align 4293; NO-FMA-NEXT: [[B3:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcB32, i32 0, i64 3), align 4294; NO-FMA-NEXT: [[B4:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcB32, i32 0, i64 4), align 4295; NO-FMA-NEXT: [[B5:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcB32, i32 0, i64 5), align 4296; NO-FMA-NEXT: [[B6:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcB32, i32 0, i64 6), align 4297; NO-FMA-NEXT: [[B7:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcB32, i32 0, i64 7), align 4298; NO-FMA-NEXT: [[C0:%.*]] = load float, ptr @srcC32, align 4299; NO-FMA-NEXT: [[C1:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcC32, i32 0, i64 1), align 4300; NO-FMA-NEXT: [[C2:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcC32, i32 0, i64 2), align 4301; NO-FMA-NEXT: [[C3:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcC32, i32 0, i64 3), align 4302; NO-FMA-NEXT: [[C4:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcC32, i32 0, i64 4), align 4303; NO-FMA-NEXT: [[C5:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcC32, i32 0, i64 5), align 4304; NO-FMA-NEXT: [[C6:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcC32, i32 0, i64 6), align 4305; NO-FMA-NEXT: [[C7:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcC32, i32 0, i64 7), align 4306; NO-FMA-NEXT: [[FMA0:%.*]] = call float @llvm.fma.f32(float [[A0]], float [[B0]], float [[C0]])307; NO-FMA-NEXT: [[FMA1:%.*]] = call float @llvm.fma.f32(float [[A1]], float [[B1]], float [[C1]])308; NO-FMA-NEXT: [[FMA2:%.*]] = call float @llvm.fma.f32(float [[A2]], float [[B2]], float [[C2]])309; NO-FMA-NEXT: [[FMA3:%.*]] = call float @llvm.fma.f32(float [[A3]], float [[B3]], float [[C3]])310; NO-FMA-NEXT: [[FMA4:%.*]] = call float @llvm.fma.f32(float [[A4]], float [[B4]], float [[C4]])311; NO-FMA-NEXT: [[FMA5:%.*]] = call float @llvm.fma.f32(float [[A5]], float [[B5]], float [[C5]])312; NO-FMA-NEXT: [[FMA6:%.*]] = call float @llvm.fma.f32(float [[A6]], float [[B6]], float [[C6]])313; NO-FMA-NEXT: [[FMA7:%.*]] = call float @llvm.fma.f32(float [[A7]], float [[B7]], float [[C7]])314; NO-FMA-NEXT: store float [[FMA0]], ptr @dst32, align 4315; NO-FMA-NEXT: store float [[FMA1]], ptr getelementptr inbounds ([16 x float], ptr @dst32, i32 0, i64 1), align 4316; NO-FMA-NEXT: store float [[FMA2]], ptr getelementptr inbounds ([16 x float], ptr @dst32, i32 0, i64 2), align 4317; NO-FMA-NEXT: store float [[FMA3]], ptr getelementptr inbounds ([16 x float], ptr @dst32, i32 0, i64 3), align 4318; NO-FMA-NEXT: store float [[FMA4]], ptr getelementptr inbounds ([16 x float], ptr @dst32, i32 0, i64 4), align 4319; NO-FMA-NEXT: store float [[FMA5]], ptr getelementptr inbounds ([16 x float], ptr @dst32, i32 0, i64 5), align 4320; NO-FMA-NEXT: store float [[FMA6]], ptr getelementptr inbounds ([16 x float], ptr @dst32, i32 0, i64 6), align 4321; NO-FMA-NEXT: store float [[FMA7]], ptr getelementptr inbounds ([16 x float], ptr @dst32, i32 0, i64 7), align 4322; NO-FMA-NEXT: ret void323;324; FMA-LABEL: @fma_8f32(325; FMA-NEXT: [[TMP1:%.*]] = load <8 x float>, ptr @srcA32, align 4326; FMA-NEXT: [[TMP2:%.*]] = load <8 x float>, ptr @srcB32, align 4327; FMA-NEXT: [[TMP3:%.*]] = load <8 x float>, ptr @srcC32, align 4328; FMA-NEXT: [[TMP4:%.*]] = call <8 x float> @llvm.fma.v8f32(<8 x float> [[TMP1]], <8 x float> [[TMP2]], <8 x float> [[TMP3]])329; FMA-NEXT: store <8 x float> [[TMP4]], ptr @dst32, align 4330; FMA-NEXT: ret void331;332 %a0 = load float, ptr @srcA32, align 4333 %a1 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcA32, i32 0, i64 1), align 4334 %a2 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcA32, i32 0, i64 2), align 4335 %a3 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcA32, i32 0, i64 3), align 4336 %a4 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcA32, i32 0, i64 4), align 4337 %a5 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcA32, i32 0, i64 5), align 4338 %a6 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcA32, i32 0, i64 6), align 4339 %a7 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcA32, i32 0, i64 7), align 4340 %b0 = load float, ptr @srcB32, align 4341 %b1 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcB32, i32 0, i64 1), align 4342 %b2 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcB32, i32 0, i64 2), align 4343 %b3 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcB32, i32 0, i64 3), align 4344 %b4 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcB32, i32 0, i64 4), align 4345 %b5 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcB32, i32 0, i64 5), align 4346 %b6 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcB32, i32 0, i64 6), align 4347 %b7 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcB32, i32 0, i64 7), align 4348 %c0 = load float, ptr @srcC32, align 4349 %c1 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcC32, i32 0, i64 1), align 4350 %c2 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcC32, i32 0, i64 2), align 4351 %c3 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcC32, i32 0, i64 3), align 4352 %c4 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcC32, i32 0, i64 4), align 4353 %c5 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcC32, i32 0, i64 5), align 4354 %c6 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcC32, i32 0, i64 6), align 4355 %c7 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcC32, i32 0, i64 7), align 4356 %fma0 = call float @llvm.fma.f32(float %a0, float %b0, float %c0)357 %fma1 = call float @llvm.fma.f32(float %a1, float %b1, float %c1)358 %fma2 = call float @llvm.fma.f32(float %a2, float %b2, float %c2)359 %fma3 = call float @llvm.fma.f32(float %a3, float %b3, float %c3)360 %fma4 = call float @llvm.fma.f32(float %a4, float %b4, float %c4)361 %fma5 = call float @llvm.fma.f32(float %a5, float %b5, float %c5)362 %fma6 = call float @llvm.fma.f32(float %a6, float %b6, float %c6)363 %fma7 = call float @llvm.fma.f32(float %a7, float %b7, float %c7)364 store float %fma0, ptr @dst32, align 4365 store float %fma1, ptr getelementptr inbounds ([16 x float], ptr @dst32, i32 0, i64 1), align 4366 store float %fma2, ptr getelementptr inbounds ([16 x float], ptr @dst32, i32 0, i64 2), align 4367 store float %fma3, ptr getelementptr inbounds ([16 x float], ptr @dst32, i32 0, i64 3), align 4368 store float %fma4, ptr getelementptr inbounds ([16 x float], ptr @dst32, i32 0, i64 4), align 4369 store float %fma5, ptr getelementptr inbounds ([16 x float], ptr @dst32, i32 0, i64 5), align 4370 store float %fma6, ptr getelementptr inbounds ([16 x float], ptr @dst32, i32 0, i64 6), align 4371 store float %fma7, ptr getelementptr inbounds ([16 x float], ptr @dst32, i32 0, i64 7), align 4372 ret void373}374 375define void @fma_16f32() #0 {376; NO-FMA-LABEL: @fma_16f32(377; NO-FMA-NEXT: [[A0:%.*]] = load float, ptr @srcA32, align 4378; NO-FMA-NEXT: [[A1:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcA32, i32 0, i64 1), align 4379; NO-FMA-NEXT: [[A2:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcA32, i32 0, i64 2), align 4380; NO-FMA-NEXT: [[A3:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcA32, i32 0, i64 3), align 4381; NO-FMA-NEXT: [[A4:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcA32, i32 0, i64 4), align 4382; NO-FMA-NEXT: [[A5:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcA32, i32 0, i64 5), align 4383; NO-FMA-NEXT: [[A6:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcA32, i32 0, i64 6), align 4384; NO-FMA-NEXT: [[A7:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcA32, i32 0, i64 7), align 4385; NO-FMA-NEXT: [[A8:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcA32, i32 0, i64 8), align 4386; NO-FMA-NEXT: [[A9:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcA32, i32 0, i64 9), align 4387; NO-FMA-NEXT: [[A10:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcA32, i32 0, i64 10), align 4388; NO-FMA-NEXT: [[A11:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcA32, i32 0, i64 11), align 4389; NO-FMA-NEXT: [[A12:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcA32, i32 0, i64 12), align 4390; NO-FMA-NEXT: [[A13:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcA32, i32 0, i64 13), align 4391; NO-FMA-NEXT: [[A14:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcA32, i32 0, i64 14), align 4392; NO-FMA-NEXT: [[A15:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcA32, i32 0, i64 15), align 4393; NO-FMA-NEXT: [[B0:%.*]] = load float, ptr @srcB32, align 4394; NO-FMA-NEXT: [[B1:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcB32, i32 0, i64 1), align 4395; NO-FMA-NEXT: [[B2:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcB32, i32 0, i64 2), align 4396; NO-FMA-NEXT: [[B3:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcB32, i32 0, i64 3), align 4397; NO-FMA-NEXT: [[B4:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcB32, i32 0, i64 4), align 4398; NO-FMA-NEXT: [[B5:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcB32, i32 0, i64 5), align 4399; NO-FMA-NEXT: [[B6:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcB32, i32 0, i64 6), align 4400; NO-FMA-NEXT: [[B7:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcB32, i32 0, i64 7), align 4401; NO-FMA-NEXT: [[B8:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcB32, i32 0, i64 8), align 4402; NO-FMA-NEXT: [[B9:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcB32, i32 0, i64 9), align 4403; NO-FMA-NEXT: [[B10:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcB32, i32 0, i64 10), align 4404; NO-FMA-NEXT: [[B11:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcB32, i32 0, i64 11), align 4405; NO-FMA-NEXT: [[B12:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcB32, i32 0, i64 12), align 4406; NO-FMA-NEXT: [[B13:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcB32, i32 0, i64 13), align 4407; NO-FMA-NEXT: [[B14:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcB32, i32 0, i64 14), align 4408; NO-FMA-NEXT: [[B15:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcB32, i32 0, i64 15), align 4409; NO-FMA-NEXT: [[C0:%.*]] = load float, ptr @srcC32, align 4410; NO-FMA-NEXT: [[C1:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcC32, i32 0, i64 1), align 4411; NO-FMA-NEXT: [[C2:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcC32, i32 0, i64 2), align 4412; NO-FMA-NEXT: [[C3:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcC32, i32 0, i64 3), align 4413; NO-FMA-NEXT: [[C4:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcC32, i32 0, i64 4), align 4414; NO-FMA-NEXT: [[C5:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcC32, i32 0, i64 5), align 4415; NO-FMA-NEXT: [[C6:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcC32, i32 0, i64 6), align 4416; NO-FMA-NEXT: [[C7:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcC32, i32 0, i64 7), align 4417; NO-FMA-NEXT: [[C8:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcC32, i32 0, i64 8), align 4418; NO-FMA-NEXT: [[C9:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcC32, i32 0, i64 9), align 4419; NO-FMA-NEXT: [[C10:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcC32, i32 0, i64 10), align 4420; NO-FMA-NEXT: [[C11:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcC32, i32 0, i64 11), align 4421; NO-FMA-NEXT: [[C12:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcC32, i32 0, i64 12), align 4422; NO-FMA-NEXT: [[C13:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcC32, i32 0, i64 13), align 4423; NO-FMA-NEXT: [[C14:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcC32, i32 0, i64 14), align 4424; NO-FMA-NEXT: [[C15:%.*]] = load float, ptr getelementptr inbounds ([16 x float], ptr @srcC32, i32 0, i64 15), align 4425; NO-FMA-NEXT: [[FMA0:%.*]] = call float @llvm.fma.f32(float [[A0]], float [[B0]], float [[C0]])426; NO-FMA-NEXT: [[FMA1:%.*]] = call float @llvm.fma.f32(float [[A1]], float [[B1]], float [[C1]])427; NO-FMA-NEXT: [[FMA2:%.*]] = call float @llvm.fma.f32(float [[A2]], float [[B2]], float [[C2]])428; NO-FMA-NEXT: [[FMA3:%.*]] = call float @llvm.fma.f32(float [[A3]], float [[B3]], float [[C3]])429; NO-FMA-NEXT: [[FMA4:%.*]] = call float @llvm.fma.f32(float [[A4]], float [[B4]], float [[C4]])430; NO-FMA-NEXT: [[FMA5:%.*]] = call float @llvm.fma.f32(float [[A5]], float [[B5]], float [[C5]])431; NO-FMA-NEXT: [[FMA6:%.*]] = call float @llvm.fma.f32(float [[A6]], float [[B6]], float [[C6]])432; NO-FMA-NEXT: [[FMA7:%.*]] = call float @llvm.fma.f32(float [[A7]], float [[B7]], float [[C7]])433; NO-FMA-NEXT: [[FMA8:%.*]] = call float @llvm.fma.f32(float [[A8]], float [[B8]], float [[C8]])434; NO-FMA-NEXT: [[FMA9:%.*]] = call float @llvm.fma.f32(float [[A9]], float [[B9]], float [[C9]])435; NO-FMA-NEXT: [[FMA10:%.*]] = call float @llvm.fma.f32(float [[A10]], float [[B10]], float [[C10]])436; NO-FMA-NEXT: [[FMA11:%.*]] = call float @llvm.fma.f32(float [[A11]], float [[B11]], float [[C11]])437; NO-FMA-NEXT: [[FMA12:%.*]] = call float @llvm.fma.f32(float [[A12]], float [[B12]], float [[C12]])438; NO-FMA-NEXT: [[FMA13:%.*]] = call float @llvm.fma.f32(float [[A13]], float [[B13]], float [[C13]])439; NO-FMA-NEXT: [[FMA14:%.*]] = call float @llvm.fma.f32(float [[A14]], float [[B14]], float [[C14]])440; NO-FMA-NEXT: [[FMA15:%.*]] = call float @llvm.fma.f32(float [[A15]], float [[B15]], float [[C15]])441; NO-FMA-NEXT: store float [[FMA0]], ptr @dst32, align 4442; NO-FMA-NEXT: store float [[FMA1]], ptr getelementptr inbounds ([16 x float], ptr @dst32, i32 0, i64 1), align 4443; NO-FMA-NEXT: store float [[FMA2]], ptr getelementptr inbounds ([16 x float], ptr @dst32, i32 0, i64 2), align 4444; NO-FMA-NEXT: store float [[FMA3]], ptr getelementptr inbounds ([16 x float], ptr @dst32, i32 0, i64 3), align 4445; NO-FMA-NEXT: store float [[FMA4]], ptr getelementptr inbounds ([16 x float], ptr @dst32, i32 0, i64 4), align 4446; NO-FMA-NEXT: store float [[FMA5]], ptr getelementptr inbounds ([16 x float], ptr @dst32, i32 0, i64 5), align 4447; NO-FMA-NEXT: store float [[FMA6]], ptr getelementptr inbounds ([16 x float], ptr @dst32, i32 0, i64 6), align 4448; NO-FMA-NEXT: store float [[FMA7]], ptr getelementptr inbounds ([16 x float], ptr @dst32, i32 0, i64 7), align 4449; NO-FMA-NEXT: store float [[FMA8]], ptr getelementptr inbounds ([16 x float], ptr @dst32, i32 0, i64 8), align 4450; NO-FMA-NEXT: store float [[FMA9]], ptr getelementptr inbounds ([16 x float], ptr @dst32, i32 0, i64 9), align 4451; NO-FMA-NEXT: store float [[FMA10]], ptr getelementptr inbounds ([16 x float], ptr @dst32, i32 0, i64 10), align 4452; NO-FMA-NEXT: store float [[FMA11]], ptr getelementptr inbounds ([16 x float], ptr @dst32, i32 0, i64 11), align 4453; NO-FMA-NEXT: store float [[FMA12]], ptr getelementptr inbounds ([16 x float], ptr @dst32, i32 0, i64 12), align 4454; NO-FMA-NEXT: store float [[FMA13]], ptr getelementptr inbounds ([16 x float], ptr @dst32, i32 0, i64 13), align 4455; NO-FMA-NEXT: store float [[FMA14]], ptr getelementptr inbounds ([16 x float], ptr @dst32, i32 0, i64 14), align 4456; NO-FMA-NEXT: store float [[FMA15]], ptr getelementptr inbounds ([16 x float], ptr @dst32, i32 0, i64 15), align 4457; NO-FMA-NEXT: ret void458;459; FMA256-LABEL: @fma_16f32(460; FMA256-NEXT: [[TMP1:%.*]] = load <8 x float>, ptr @srcA32, align 4461; FMA256-NEXT: [[TMP2:%.*]] = load <8 x float>, ptr @srcB32, align 4462; FMA256-NEXT: [[TMP3:%.*]] = load <8 x float>, ptr @srcC32, align 4463; FMA256-NEXT: [[TMP4:%.*]] = call <8 x float> @llvm.fma.v8f32(<8 x float> [[TMP1]], <8 x float> [[TMP2]], <8 x float> [[TMP3]])464; FMA256-NEXT: store <8 x float> [[TMP4]], ptr @dst32, align 4465; FMA256-NEXT: [[TMP5:%.*]] = load <8 x float>, ptr getelementptr inbounds ([16 x float], ptr @srcA32, i32 0, i64 8), align 4466; FMA256-NEXT: [[TMP6:%.*]] = load <8 x float>, ptr getelementptr inbounds ([16 x float], ptr @srcB32, i32 0, i64 8), align 4467; FMA256-NEXT: [[TMP7:%.*]] = load <8 x float>, ptr getelementptr inbounds ([16 x float], ptr @srcC32, i32 0, i64 8), align 4468; FMA256-NEXT: [[TMP8:%.*]] = call <8 x float> @llvm.fma.v8f32(<8 x float> [[TMP5]], <8 x float> [[TMP6]], <8 x float> [[TMP7]])469; FMA256-NEXT: store <8 x float> [[TMP8]], ptr getelementptr inbounds ([16 x float], ptr @dst32, i32 0, i64 8), align 4470; FMA256-NEXT: ret void471;472; FMA512-LABEL: @fma_16f32(473; FMA512-NEXT: [[TMP1:%.*]] = load <16 x float>, ptr @srcA32, align 4474; FMA512-NEXT: [[TMP2:%.*]] = load <16 x float>, ptr @srcB32, align 4475; FMA512-NEXT: [[TMP3:%.*]] = load <16 x float>, ptr @srcC32, align 4476; FMA512-NEXT: [[TMP4:%.*]] = call <16 x float> @llvm.fma.v16f32(<16 x float> [[TMP1]], <16 x float> [[TMP2]], <16 x float> [[TMP3]])477; FMA512-NEXT: store <16 x float> [[TMP4]], ptr @dst32, align 4478; FMA512-NEXT: ret void479;480 %a0 = load float, ptr @srcA32, align 4481 %a1 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcA32, i32 0, i64 1), align 4482 %a2 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcA32, i32 0, i64 2), align 4483 %a3 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcA32, i32 0, i64 3), align 4484 %a4 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcA32, i32 0, i64 4), align 4485 %a5 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcA32, i32 0, i64 5), align 4486 %a6 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcA32, i32 0, i64 6), align 4487 %a7 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcA32, i32 0, i64 7), align 4488 %a8 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcA32, i32 0, i64 8), align 4489 %a9 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcA32, i32 0, i64 9), align 4490 %a10 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcA32, i32 0, i64 10), align 4491 %a11 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcA32, i32 0, i64 11), align 4492 %a12 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcA32, i32 0, i64 12), align 4493 %a13 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcA32, i32 0, i64 13), align 4494 %a14 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcA32, i32 0, i64 14), align 4495 %a15 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcA32, i32 0, i64 15), align 4496 %b0 = load float, ptr @srcB32, align 4497 %b1 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcB32, i32 0, i64 1), align 4498 %b2 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcB32, i32 0, i64 2), align 4499 %b3 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcB32, i32 0, i64 3), align 4500 %b4 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcB32, i32 0, i64 4), align 4501 %b5 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcB32, i32 0, i64 5), align 4502 %b6 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcB32, i32 0, i64 6), align 4503 %b7 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcB32, i32 0, i64 7), align 4504 %b8 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcB32, i32 0, i64 8), align 4505 %b9 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcB32, i32 0, i64 9), align 4506 %b10 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcB32, i32 0, i64 10), align 4507 %b11 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcB32, i32 0, i64 11), align 4508 %b12 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcB32, i32 0, i64 12), align 4509 %b13 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcB32, i32 0, i64 13), align 4510 %b14 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcB32, i32 0, i64 14), align 4511 %b15 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcB32, i32 0, i64 15), align 4512 %c0 = load float, ptr @srcC32, align 4513 %c1 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcC32, i32 0, i64 1), align 4514 %c2 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcC32, i32 0, i64 2), align 4515 %c3 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcC32, i32 0, i64 3), align 4516 %c4 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcC32, i32 0, i64 4), align 4517 %c5 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcC32, i32 0, i64 5), align 4518 %c6 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcC32, i32 0, i64 6), align 4519 %c7 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcC32, i32 0, i64 7), align 4520 %c8 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcC32, i32 0, i64 8), align 4521 %c9 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcC32, i32 0, i64 9), align 4522 %c10 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcC32, i32 0, i64 10), align 4523 %c11 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcC32, i32 0, i64 11), align 4524 %c12 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcC32, i32 0, i64 12), align 4525 %c13 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcC32, i32 0, i64 13), align 4526 %c14 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcC32, i32 0, i64 14), align 4527 %c15 = load float, ptr getelementptr inbounds ([16 x float], ptr @srcC32, i32 0, i64 15), align 4528 %fma0 = call float @llvm.fma.f32(float %a0 , float %b0 , float %c0 )529 %fma1 = call float @llvm.fma.f32(float %a1 , float %b1 , float %c1 )530 %fma2 = call float @llvm.fma.f32(float %a2 , float %b2 , float %c2 )531 %fma3 = call float @llvm.fma.f32(float %a3 , float %b3 , float %c3 )532 %fma4 = call float @llvm.fma.f32(float %a4 , float %b4 , float %c4 )533 %fma5 = call float @llvm.fma.f32(float %a5 , float %b5 , float %c5 )534 %fma6 = call float @llvm.fma.f32(float %a6 , float %b6 , float %c6 )535 %fma7 = call float @llvm.fma.f32(float %a7 , float %b7 , float %c7 )536 %fma8 = call float @llvm.fma.f32(float %a8 , float %b8 , float %c8 )537 %fma9 = call float @llvm.fma.f32(float %a9 , float %b9 , float %c9 )538 %fma10 = call float @llvm.fma.f32(float %a10, float %b10, float %c10)539 %fma11 = call float @llvm.fma.f32(float %a11, float %b11, float %c11)540 %fma12 = call float @llvm.fma.f32(float %a12, float %b12, float %c12)541 %fma13 = call float @llvm.fma.f32(float %a13, float %b13, float %c13)542 %fma14 = call float @llvm.fma.f32(float %a14, float %b14, float %c14)543 %fma15 = call float @llvm.fma.f32(float %a15, float %b15, float %c15)544 store float %fma0 , ptr @dst32, align 4545 store float %fma1 , ptr getelementptr inbounds ([16 x float], ptr @dst32, i32 0, i64 1), align 4546 store float %fma2 , ptr getelementptr inbounds ([16 x float], ptr @dst32, i32 0, i64 2), align 4547 store float %fma3 , ptr getelementptr inbounds ([16 x float], ptr @dst32, i32 0, i64 3), align 4548 store float %fma4 , ptr getelementptr inbounds ([16 x float], ptr @dst32, i32 0, i64 4), align 4549 store float %fma5 , ptr getelementptr inbounds ([16 x float], ptr @dst32, i32 0, i64 5), align 4550 store float %fma6 , ptr getelementptr inbounds ([16 x float], ptr @dst32, i32 0, i64 6), align 4551 store float %fma7 , ptr getelementptr inbounds ([16 x float], ptr @dst32, i32 0, i64 7), align 4552 store float %fma8 , ptr getelementptr inbounds ([16 x float], ptr @dst32, i32 0, i64 8), align 4553 store float %fma9 , ptr getelementptr inbounds ([16 x float], ptr @dst32, i32 0, i64 9), align 4554 store float %fma10, ptr getelementptr inbounds ([16 x float], ptr @dst32, i32 0, i64 10), align 4555 store float %fma11, ptr getelementptr inbounds ([16 x float], ptr @dst32, i32 0, i64 11), align 4556 store float %fma12, ptr getelementptr inbounds ([16 x float], ptr @dst32, i32 0, i64 12), align 4557 store float %fma13, ptr getelementptr inbounds ([16 x float], ptr @dst32, i32 0, i64 13), align 4558 store float %fma14, ptr getelementptr inbounds ([16 x float], ptr @dst32, i32 0, i64 14), align 4559 store float %fma15, ptr getelementptr inbounds ([16 x float], ptr @dst32, i32 0, i64 15), align 4560 ret void561}562 563attributes #0 = { nounwind }564