; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
; RUN: opt < %s -mtriple=x86_64-unknown -basicaa -slp-vectorizer -S | FileCheck %s --check-prefix=NO-FMA
; RUN: opt < %s -mtriple=x86_64-unknown -mcpu=corei7-avx -basicaa -slp-vectorizer -S | FileCheck %s --check-prefix=NO-FMA
; RUN: opt < %s -mtriple=x86_64-unknown -mcpu=bdver1 -basicaa -slp-vectorizer -S | FileCheck %s --check-prefix=FMA --check-prefix=FMA256
; RUN: opt < %s -mtriple=x86_64-unknown -mcpu=core-avx2 -basicaa -slp-vectorizer -S | FileCheck %s --check-prefix=FMA --check-prefix=FMA256
; RUN: opt < %s -mtriple=x86_64-unknown -mcpu=skylake-avx512 -mattr=-prefer-256-bit -basicaa -slp-vectorizer -S | FileCheck %s --check-prefix=FMA --check-prefix=FMA512
; RUN: opt < %s -mtriple=x86_64-unknown -mcpu=skylake-avx512 -mattr=+prefer-256-bit -basicaa -slp-vectorizer -S | FileCheck %s --check-prefix=FMA --check-prefix=FMA256

target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"

@srcA64 = common global [8 x double] zeroinitializer, align 64
@srcB64 = common global [8 x double] zeroinitializer, align 64
@srcC64 = common global [8 x double] zeroinitializer, align 64
@srcA32 = common global [16 x float] zeroinitializer, align 64
@srcB32 = common global [16 x float] zeroinitializer, align 64
@srcC32 = common global [16 x float] zeroinitializer, align 64
@dst64 = common global [8 x double] zeroinitializer, align 64
@dst32 = common global [16 x float] zeroinitializer, align 64

declare float @llvm.fma.f32(float, float, float)
declare double @llvm.fma.f64(double, double, double)

;
; FMA
;

define void @fma_2f64() #0 {
; NO-FMA-LABEL: @fma_2f64(
; NO-FMA-NEXT:    [[A0:%.*]] = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcA64, i32 0, i64 0), align 8
; NO-FMA-NEXT:    [[A1:%.*]] = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcA64, i32 0, i64 1), align 8
; NO-FMA-NEXT:    [[B0:%.*]] = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcB64, i32 0, i64 0), align 8
; NO-FMA-NEXT:    [[B1:%.*]] = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcB64, i32 0, i64 1), align 8
; NO-FMA-NEXT:    [[C0:%.*]] = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcC64, i32 0, i64 0), align 8
; NO-FMA-NEXT:    [[C1:%.*]] = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcC64, i32 0, i64 1), align 8
; NO-FMA-NEXT:    [[FMA0:%.*]] = call double @llvm.fma.f64(double [[A0]], double [[B0]], double [[C0]])
; NO-FMA-NEXT:    [[FMA1:%.*]] = call double @llvm.fma.f64(double [[A1]], double [[B1]], double [[C1]])
; NO-FMA-NEXT:    store double [[FMA0]], double* getelementptr inbounds ([8 x double], [8 x double]* @dst64, i32 0, i64 0), align 8
; NO-FMA-NEXT:    store double [[FMA1]], double* getelementptr inbounds ([8 x double], [8 x double]* @dst64, i32 0, i64 1), align 8
; NO-FMA-NEXT:    ret void
;
; FMA-LABEL: @fma_2f64(
; FMA-NEXT:    [[TMP1:%.*]] = load <2 x double>, <2 x double>* bitcast ([8 x double]* @srcA64 to <2 x double>*), align 8
; FMA-NEXT:    [[TMP2:%.*]] = load <2 x double>, <2 x double>* bitcast ([8 x double]* @srcB64 to <2 x double>*), align 8
; FMA-NEXT:    [[TMP3:%.*]] = load <2 x double>, <2 x double>* bitcast ([8 x double]* @srcC64 to <2 x double>*), align 8
; FMA-NEXT:    [[TMP4:%.*]] = call <2 x double> @llvm.fma.v2f64(<2 x double> [[TMP1]], <2 x double> [[TMP2]], <2 x double> [[TMP3]])
; FMA-NEXT:    store <2 x double> [[TMP4]], <2 x double>* bitcast ([8 x double]* @dst64 to <2 x double>*), align 8
; FMA-NEXT:    ret void
;
  %a0 = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcA64, i32 0, i64 0), align 8
  %a1 = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcA64, i32 0, i64 1), align 8
  %b0 = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcB64, i32 0, i64 0), align 8
  %b1 = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcB64, i32 0, i64 1), align 8
  %c0 = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcC64, i32 0, i64 0), align 8
  %c1 = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcC64, i32 0, i64 1), align 8
  %fma0 = call double @llvm.fma.f64(double %a0, double %b0, double %c0)
  %fma1 = call double @llvm.fma.f64(double %a1, double %b1, double %c1)
  store double %fma0, double* getelementptr inbounds ([8 x double], [8 x double]* @dst64, i32 0, i64 0), align 8
  store double %fma1, double* getelementptr inbounds ([8 x double], [8 x double]* @dst64, i32 0, i64 1), align 8
  ret void
}

define void @fma_4f64() #0 {
; NO-FMA-LABEL: @fma_4f64(
; NO-FMA-NEXT:    [[A0:%.*]] = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcA64, i32 0, i64 0), align 8
; NO-FMA-NEXT:    [[A1:%.*]] = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcA64, i32 0, i64 1), align 8
; NO-FMA-NEXT:    [[A2:%.*]] = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcA64, i32 0, i64 2), align 8
; NO-FMA-NEXT:    [[A3:%.*]] = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcA64, i32 0, i64 3), align 8
; NO-FMA-NEXT:    [[B0:%.*]] = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcB64, i32 0, i64 0), align 8
; NO-FMA-NEXT:    [[B1:%.*]] = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcB64, i32 0, i64 1), align 8
; NO-FMA-NEXT:    [[B2:%.*]] = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcB64, i32 0, i64 2), align 8
; NO-FMA-NEXT:    [[B3:%.*]] = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcB64, i32 0, i64 3), align 8
; NO-FMA-NEXT:    [[C0:%.*]] = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcC64, i32 0, i64 0), align 8
; NO-FMA-NEXT:    [[C1:%.*]] = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcC64, i32 0, i64 1), align 8
; NO-FMA-NEXT:    [[C2:%.*]] = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcC64, i32 0, i64 2), align 8
; NO-FMA-NEXT:    [[C3:%.*]] = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcC64, i32 0, i64 3), align 8
; NO-FMA-NEXT:    [[FMA0:%.*]] = call double @llvm.fma.f64(double [[A0]], double [[B0]], double [[C0]])
; NO-FMA-NEXT:    [[FMA1:%.*]] = call double @llvm.fma.f64(double [[A1]], double [[B1]], double [[C1]])
; NO-FMA-NEXT:    [[FMA2:%.*]] = call double @llvm.fma.f64(double [[A2]], double [[B2]], double [[C2]])
; NO-FMA-NEXT:    [[FMA3:%.*]] = call double @llvm.fma.f64(double [[A3]], double [[B3]], double [[C3]])
; NO-FMA-NEXT:    store double [[FMA0]], double* getelementptr inbounds ([8 x double], [8 x double]* @dst64, i32 0, i64 0), align 8
; NO-FMA-NEXT:    store double [[FMA1]], double* getelementptr inbounds ([8 x double], [8 x double]* @dst64, i32 0, i64 1), align 8
; NO-FMA-NEXT:    store double [[FMA2]], double* getelementptr inbounds ([8 x double], [8 x double]* @dst64, i32 0, i64 2), align 8
; NO-FMA-NEXT:    store double [[FMA3]], double* getelementptr inbounds ([8 x double], [8 x double]* @dst64, i32 0, i64 3), align 8
; NO-FMA-NEXT:    ret void
;
; FMA-LABEL: @fma_4f64(
; FMA-NEXT:    [[TMP1:%.*]] = load <4 x double>, <4 x double>* bitcast ([8 x double]* @srcA64 to <4 x double>*), align 8
; FMA-NEXT:    [[TMP2:%.*]] = load <4 x double>, <4 x double>* bitcast ([8 x double]* @srcB64 to <4 x double>*), align 8
; FMA-NEXT:    [[TMP3:%.*]] = load <4 x double>, <4 x double>* bitcast ([8 x double]* @srcC64 to <4 x double>*), align 8
; FMA-NEXT:    [[TMP4:%.*]] = call <4 x double> @llvm.fma.v4f64(<4 x double> [[TMP1]], <4 x double> [[TMP2]], <4 x double> [[TMP3]])
; FMA-NEXT:    store <4 x double> [[TMP4]], <4 x double>* bitcast ([8 x double]* @dst64 to <4 x double>*), align 8
; FMA-NEXT:    ret void
;
  %a0 = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcA64, i32 0, i64 0), align 8
  %a1 = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcA64, i32 0, i64 1), align 8
  %a2 = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcA64, i32 0, i64 2), align 8
  %a3 = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcA64, i32 0, i64 3), align 8
  %b0 = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcB64, i32 0, i64 0), align 8
  %b1 = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcB64, i32 0, i64 1), align 8
  %b2 = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcB64, i32 0, i64 2), align 8
  %b3 = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcB64, i32 0, i64 3), align 8
  %c0 = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcC64, i32 0, i64 0), align 8
  %c1 = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcC64, i32 0, i64 1), align 8
  %c2 = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcC64, i32 0, i64 2), align 8
  %c3 = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcC64, i32 0, i64 3), align 8
  %fma0 = call double @llvm.fma.f64(double %a0, double %b0, double %c0)
  %fma1 = call double @llvm.fma.f64(double %a1, double %b1, double %c1)
  %fma2 = call double @llvm.fma.f64(double %a2, double %b2, double %c2)
  %fma3 = call double @llvm.fma.f64(double %a3, double %b3, double %c3)
  store double %fma0, double* getelementptr inbounds ([8 x double], [8 x double]* @dst64, i32 0, i64 0), align 8
  store double %fma1, double* getelementptr inbounds ([8 x double], [8 x double]* @dst64, i32 0, i64 1), align 8
  store double %fma2, double* getelementptr inbounds ([8 x double], [8 x double]* @dst64, i32 0, i64 2), align 8
  store double %fma3, double* getelementptr inbounds ([8 x double], [8 x double]* @dst64, i32 0, i64 3), align 8
  ret void
}

define void @fma_8f64() #0 {
; NO-FMA-LABEL: @fma_8f64(
; NO-FMA-NEXT:    [[A0:%.*]] = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcA64, i32 0, i64 0), align 4
; NO-FMA-NEXT:    [[A1:%.*]] = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcA64, i32 0, i64 1), align 4
; NO-FMA-NEXT:    [[A2:%.*]] = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcA64, i32 0, i64 2), align 4
; NO-FMA-NEXT:    [[A3:%.*]] = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcA64, i32 0, i64 3), align 4
; NO-FMA-NEXT:    [[A4:%.*]] = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcA64, i32 0, i64 4), align 4
; NO-FMA-NEXT:    [[A5:%.*]] = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcA64, i32 0, i64 5), align 4
; NO-FMA-NEXT:    [[A6:%.*]] = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcA64, i32 0, i64 6), align 4
; NO-FMA-NEXT:    [[A7:%.*]] = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcA64, i32 0, i64 7), align 4
; NO-FMA-NEXT:    [[B0:%.*]] = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcB64, i32 0, i64 0), align 4
; NO-FMA-NEXT:    [[B1:%.*]] = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcB64, i32 0, i64 1), align 4
; NO-FMA-NEXT:    [[B2:%.*]] = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcB64, i32 0, i64 2), align 4
; NO-FMA-NEXT:    [[B3:%.*]] = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcB64, i32 0, i64 3), align 4
; NO-FMA-NEXT:    [[B4:%.*]] = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcB64, i32 0, i64 4), align 4
; NO-FMA-NEXT:    [[B5:%.*]] = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcB64, i32 0, i64 5), align 4
; NO-FMA-NEXT:    [[B6:%.*]] = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcB64, i32 0, i64 6), align 4
; NO-FMA-NEXT:    [[B7:%.*]] = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcB64, i32 0, i64 7), align 4
; NO-FMA-NEXT:    [[C0:%.*]] = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcC64, i32 0, i64 0), align 4
; NO-FMA-NEXT:    [[C1:%.*]] = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcC64, i32 0, i64 1), align 4
; NO-FMA-NEXT:    [[C2:%.*]] = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcC64, i32 0, i64 2), align 4
; NO-FMA-NEXT:    [[C3:%.*]] = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcC64, i32 0, i64 3), align 4
; NO-FMA-NEXT:    [[C4:%.*]] = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcC64, i32 0, i64 4), align 4
; NO-FMA-NEXT:    [[C5:%.*]] = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcC64, i32 0, i64 5), align 4
; NO-FMA-NEXT:    [[C6:%.*]] = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcC64, i32 0, i64 6), align 4
; NO-FMA-NEXT:    [[C7:%.*]] = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcC64, i32 0, i64 7), align 4
; NO-FMA-NEXT:    [[FMA0:%.*]] = call double @llvm.fma.f64(double [[A0]], double [[B0]], double [[C0]])
; NO-FMA-NEXT:    [[FMA1:%.*]] = call double @llvm.fma.f64(double [[A1]], double [[B1]], double [[C1]])
; NO-FMA-NEXT:    [[FMA2:%.*]] = call double @llvm.fma.f64(double [[A2]], double [[B2]], double [[C2]])
; NO-FMA-NEXT:    [[FMA3:%.*]] = call double @llvm.fma.f64(double [[A3]], double [[B3]], double [[C3]])
; NO-FMA-NEXT:    [[FMA4:%.*]] = call double @llvm.fma.f64(double [[A4]], double [[B4]], double [[C4]])
; NO-FMA-NEXT:    [[FMA5:%.*]] = call double @llvm.fma.f64(double [[A5]], double [[B5]], double [[C5]])
; NO-FMA-NEXT:    [[FMA6:%.*]] = call double @llvm.fma.f64(double [[A6]], double [[B6]], double [[C6]])
; NO-FMA-NEXT:    [[FMA7:%.*]] = call double @llvm.fma.f64(double [[A7]], double [[B7]], double [[C7]])
; NO-FMA-NEXT:    store double [[FMA0]], double* getelementptr inbounds ([8 x double], [8 x double]* @dst64, i32 0, i64 0), align 4
; NO-FMA-NEXT:    store double [[FMA1]], double* getelementptr inbounds ([8 x double], [8 x double]* @dst64, i32 0, i64 1), align 4
; NO-FMA-NEXT:    store double [[FMA2]], double* getelementptr inbounds ([8 x double], [8 x double]* @dst64, i32 0, i64 2), align 4
; NO-FMA-NEXT:    store double [[FMA3]], double* getelementptr inbounds ([8 x double], [8 x double]* @dst64, i32 0, i64 3), align 4
; NO-FMA-NEXT:    store double [[FMA4]], double* getelementptr inbounds ([8 x double], [8 x double]* @dst64, i32 0, i64 4), align 4
; NO-FMA-NEXT:    store double [[FMA5]], double* getelementptr inbounds ([8 x double], [8 x double]* @dst64, i32 0, i64 5), align 4
; NO-FMA-NEXT:    store double [[FMA6]], double* getelementptr inbounds ([8 x double], [8 x double]* @dst64, i32 0, i64 6), align 4
; NO-FMA-NEXT:    store double [[FMA7]], double* getelementptr inbounds ([8 x double], [8 x double]* @dst64, i32 0, i64 7), align 4
; NO-FMA-NEXT:    ret void
;
; FMA256-LABEL: @fma_8f64(
; FMA256-NEXT:    [[TMP1:%.*]] = load <4 x double>, <4 x double>* bitcast ([8 x double]* @srcA64 to <4 x double>*), align 4
; FMA256-NEXT:    [[TMP2:%.*]] = load <4 x double>, <4 x double>* bitcast (double* getelementptr inbounds ([8 x double], [8 x double]* @srcA64, i32 0, i64 4) to <4 x double>*), align 4
; FMA256-NEXT:    [[TMP3:%.*]] = load <4 x double>, <4 x double>* bitcast ([8 x double]* @srcB64 to <4 x double>*), align 4
; FMA256-NEXT:    [[TMP4:%.*]] = load <4 x double>, <4 x double>* bitcast (double* getelementptr inbounds ([8 x double], [8 x double]* @srcB64, i32 0, i64 4) to <4 x double>*), align 4
; FMA256-NEXT:    [[TMP5:%.*]] = load <4 x double>, <4 x double>* bitcast ([8 x double]* @srcC64 to <4 x double>*), align 4
; FMA256-NEXT:    [[TMP6:%.*]] = load <4 x double>, <4 x double>* bitcast (double* getelementptr inbounds ([8 x double], [8 x double]* @srcC64, i32 0, i64 4) to <4 x double>*), align 4
; FMA256-NEXT:    [[TMP7:%.*]] = call <4 x double> @llvm.fma.v4f64(<4 x double> [[TMP1]], <4 x double> [[TMP3]], <4 x double> [[TMP5]])
; FMA256-NEXT:    [[TMP8:%.*]] = call <4 x double> @llvm.fma.v4f64(<4 x double> [[TMP2]], <4 x double> [[TMP4]], <4 x double> [[TMP6]])
; FMA256-NEXT:    store <4 x double> [[TMP7]], <4 x double>* bitcast ([8 x double]* @dst64 to <4 x double>*), align 4
; FMA256-NEXT:    store <4 x double> [[TMP8]], <4 x double>* bitcast (double* getelementptr inbounds ([8 x double], [8 x double]* @dst64, i32 0, i64 4) to <4 x double>*), align 4
; FMA256-NEXT:    ret void
;
; FMA512-LABEL: @fma_8f64(
; FMA512-NEXT:    [[TMP1:%.*]] = load <8 x double>, <8 x double>* bitcast ([8 x double]* @srcA64 to <8 x double>*), align 4
; FMA512-NEXT:    [[TMP2:%.*]] = load <8 x double>, <8 x double>* bitcast ([8 x double]* @srcB64 to <8 x double>*), align 4
; FMA512-NEXT:    [[TMP3:%.*]] = load <8 x double>, <8 x double>* bitcast ([8 x double]* @srcC64 to <8 x double>*), align 4
; FMA512-NEXT:    [[TMP4:%.*]] = call <8 x double> @llvm.fma.v8f64(<8 x double> [[TMP1]], <8 x double> [[TMP2]], <8 x double> [[TMP3]])
; FMA512-NEXT:    store <8 x double> [[TMP4]], <8 x double>* bitcast ([8 x double]* @dst64 to <8 x double>*), align 4
; FMA512-NEXT:    ret void
;
  %a0 = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcA64, i32 0, i64 0), align 4
  %a1 = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcA64, i32 0, i64 1), align 4
  %a2 = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcA64, i32 0, i64 2), align 4
  %a3 = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcA64, i32 0, i64 3), align 4
  %a4 = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcA64, i32 0, i64 4), align 4
  %a5 = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcA64, i32 0, i64 5), align 4
  %a6 = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcA64, i32 0, i64 6), align 4
  %a7 = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcA64, i32 0, i64 7), align 4
  %b0 = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcB64, i32 0, i64 0), align 4
  %b1 = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcB64, i32 0, i64 1), align 4
  %b2 = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcB64, i32 0, i64 2), align 4
  %b3 = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcB64, i32 0, i64 3), align 4
  %b4 = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcB64, i32 0, i64 4), align 4
  %b5 = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcB64, i32 0, i64 5), align 4
  %b6 = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcB64, i32 0, i64 6), align 4
  %b7 = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcB64, i32 0, i64 7), align 4
  %c0 = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcC64, i32 0, i64 0), align 4
  %c1 = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcC64, i32 0, i64 1), align 4
  %c2 = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcC64, i32 0, i64 2), align 4
  %c3 = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcC64, i32 0, i64 3), align 4
  %c4 = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcC64, i32 0, i64 4), align 4
  %c5 = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcC64, i32 0, i64 5), align 4
  %c6 = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcC64, i32 0, i64 6), align 4
  %c7 = load double, double* getelementptr inbounds ([8 x double], [8 x double]* @srcC64, i32 0, i64 7), align 4
  %fma0 = call double @llvm.fma.f64(double %a0, double %b0, double %c0)
  %fma1 = call double @llvm.fma.f64(double %a1, double %b1, double %c1)
  %fma2 = call double @llvm.fma.f64(double %a2, double %b2, double %c2)
  %fma3 = call double @llvm.fma.f64(double %a3, double %b3, double %c3)
  %fma4 = call double @llvm.fma.f64(double %a4, double %b4, double %c4)
  %fma5 = call double @llvm.fma.f64(double %a5, double %b5, double %c5)
  %fma6 = call double @llvm.fma.f64(double %a6, double %b6, double %c6)
  %fma7 = call double @llvm.fma.f64(double %a7, double %b7, double %c7)
  store double %fma0, double* getelementptr inbounds ([8 x double], [8 x double]* @dst64, i32 0, i64 0), align 4
  store double %fma1, double* getelementptr inbounds ([8 x double], [8 x double]* @dst64, i32 0, i64 1), align 4
  store double %fma2, double* getelementptr inbounds ([8 x double], [8 x double]* @dst64, i32 0, i64 2), align 4
  store double %fma3, double* getelementptr inbounds ([8 x double], [8 x double]* @dst64, i32 0, i64 3), align 4
  store double %fma4, double* getelementptr inbounds ([8 x double], [8 x double]* @dst64, i32 0, i64 4), align 4
  store double %fma5, double* getelementptr inbounds ([8 x double], [8 x double]* @dst64, i32 0, i64 5), align 4
  store double %fma6, double* getelementptr inbounds ([8 x double], [8 x double]* @dst64, i32 0, i64 6), align 4
  store double %fma7, double* getelementptr inbounds ([8 x double], [8 x double]* @dst64, i32 0, i64 7), align 4
  ret void
}

define void @fma_4f32() #0 {
; NO-FMA-LABEL: @fma_4f32(
; NO-FMA-NEXT:    [[A0:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcA32, i32 0, i64 0), align 4
; NO-FMA-NEXT:    [[A1:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcA32, i32 0, i64 1), align 4
; NO-FMA-NEXT:    [[A2:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcA32, i32 0, i64 2), align 4
; NO-FMA-NEXT:    [[A3:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcA32, i32 0, i64 3), align 4
; NO-FMA-NEXT:    [[B0:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcB32, i32 0, i64 0), align 4
; NO-FMA-NEXT:    [[B1:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcB32, i32 0, i64 1), align 4
; NO-FMA-NEXT:    [[B2:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcB32, i32 0, i64 2), align 4
; NO-FMA-NEXT:    [[B3:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcB32, i32 0, i64 3), align 4
; NO-FMA-NEXT:    [[C0:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcC32, i32 0, i64 0), align 4
; NO-FMA-NEXT:    [[C1:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcC32, i32 0, i64 1), align 4
; NO-FMA-NEXT:    [[C2:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcC32, i32 0, i64 2), align 4
; NO-FMA-NEXT:    [[C3:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcC32, i32 0, i64 3), align 4
; NO-FMA-NEXT:    [[FMA0:%.*]] = call float @llvm.fma.f32(float [[A0]], float [[B0]], float [[C0]])
; NO-FMA-NEXT:    [[FMA1:%.*]] = call float @llvm.fma.f32(float [[A1]], float [[B1]], float [[C1]])
; NO-FMA-NEXT:    [[FMA2:%.*]] = call float @llvm.fma.f32(float [[A2]], float [[B2]], float [[C2]])
; NO-FMA-NEXT:    [[FMA3:%.*]] = call float @llvm.fma.f32(float [[A3]], float [[B3]], float [[C3]])
; NO-FMA-NEXT:    store float [[FMA0]], float* getelementptr inbounds ([16 x float], [16 x float]* @dst32, i32 0, i64 0), align 4
; NO-FMA-NEXT:    store float [[FMA1]], float* getelementptr inbounds ([16 x float], [16 x float]* @dst32, i32 0, i64 1), align 4
; NO-FMA-NEXT:    store float [[FMA2]], float* getelementptr inbounds ([16 x float], [16 x float]* @dst32, i32 0, i64 2), align 4
; NO-FMA-NEXT:    store float [[FMA3]], float* getelementptr inbounds ([16 x float], [16 x float]* @dst32, i32 0, i64 3), align 4
; NO-FMA-NEXT:    ret void
;
; FMA-LABEL: @fma_4f32(
; FMA-NEXT:    [[TMP1:%.*]] = load <4 x float>, <4 x float>* bitcast ([16 x float]* @srcA32 to <4 x float>*), align 4
; FMA-NEXT:    [[TMP2:%.*]] = load <4 x float>, <4 x float>* bitcast ([16 x float]* @srcB32 to <4 x float>*), align 4
; FMA-NEXT:    [[TMP3:%.*]] = load <4 x float>, <4 x float>* bitcast ([16 x float]* @srcC32 to <4 x float>*), align 4
; FMA-NEXT:    [[TMP4:%.*]] = call <4 x float> @llvm.fma.v4f32(<4 x float> [[TMP1]], <4 x float> [[TMP2]], <4 x float> [[TMP3]])
; FMA-NEXT:    store <4 x float> [[TMP4]], <4 x float>* bitcast ([16 x float]* @dst32 to <4 x float>*), align 4
; FMA-NEXT:    ret void
;
  %a0 = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcA32, i32 0, i64 0), align 4
  %a1 = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcA32, i32 0, i64 1), align 4
  %a2 = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcA32, i32 0, i64 2), align 4
  %a3 = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcA32, i32 0, i64 3), align 4
  %b0 = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcB32, i32 0, i64 0), align 4
  %b1 = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcB32, i32 0, i64 1), align 4
  %b2 = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcB32, i32 0, i64 2), align 4
  %b3 = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcB32, i32 0, i64 3), align 4
  %c0 = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcC32, i32 0, i64 0), align 4
  %c1 = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcC32, i32 0, i64 1), align 4
  %c2 = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcC32, i32 0, i64 2), align 4
  %c3 = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcC32, i32 0, i64 3), align 4
  %fma0 = call float @llvm.fma.f32(float %a0, float %b0, float %c0)
  %fma1 = call float @llvm.fma.f32(float %a1, float %b1, float %c1)
  %fma2 = call float @llvm.fma.f32(float %a2, float %b2, float %c2)
  %fma3 = call float @llvm.fma.f32(float %a3, float %b3, float %c3)
  store float %fma0, float* getelementptr inbounds ([16 x float], [16 x float]* @dst32, i32 0, i64 0), align 4
  store float %fma1, float* getelementptr inbounds ([16 x float], [16 x float]* @dst32, i32 0, i64 1), align 4
  store float %fma2, float* getelementptr inbounds ([16 x float], [16 x float]* @dst32, i32 0, i64 2), align 4
  store float %fma3, float* getelementptr inbounds ([16 x float], [16 x float]* @dst32, i32 0, i64 3), align 4
  ret void
}

define void @fma_8f32() #0 {
; NO-FMA-LABEL: @fma_8f32(
; NO-FMA-NEXT:    [[A0:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcA32, i32 0, i64 0), align 4
; NO-FMA-NEXT:    [[A1:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcA32, i32 0, i64 1), align 4
; NO-FMA-NEXT:    [[A2:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcA32, i32 0, i64 2), align 4
; NO-FMA-NEXT:    [[A3:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcA32, i32 0, i64 3), align 4
; NO-FMA-NEXT:    [[A4:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcA32, i32 0, i64 4), align 4
; NO-FMA-NEXT:    [[A5:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcA32, i32 0, i64 5), align 4
; NO-FMA-NEXT:    [[A6:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcA32, i32 0, i64 6), align 4
; NO-FMA-NEXT:    [[A7:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcA32, i32 0, i64 7), align 4
; NO-FMA-NEXT:    [[B0:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcB32, i32 0, i64 0), align 4
; NO-FMA-NEXT:    [[B1:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcB32, i32 0, i64 1), align 4
; NO-FMA-NEXT:    [[B2:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcB32, i32 0, i64 2), align 4
; NO-FMA-NEXT:    [[B3:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcB32, i32 0, i64 3), align 4
; NO-FMA-NEXT:    [[B4:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcB32, i32 0, i64 4), align 4
; NO-FMA-NEXT:    [[B5:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcB32, i32 0, i64 5), align 4
; NO-FMA-NEXT:    [[B6:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcB32, i32 0, i64 6), align 4
; NO-FMA-NEXT:    [[B7:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcB32, i32 0, i64 7), align 4
; NO-FMA-NEXT:    [[C0:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcC32, i32 0, i64 0), align 4
; NO-FMA-NEXT:    [[C1:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcC32, i32 0, i64 1), align 4
; NO-FMA-NEXT:    [[C2:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcC32, i32 0, i64 2), align 4
; NO-FMA-NEXT:    [[C3:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcC32, i32 0, i64 3), align 4
; NO-FMA-NEXT:    [[C4:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcC32, i32 0, i64 4), align 4
; NO-FMA-NEXT:    [[C5:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcC32, i32 0, i64 5), align 4
; NO-FMA-NEXT:    [[C6:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcC32, i32 0, i64 6), align 4
; NO-FMA-NEXT:    [[C7:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcC32, i32 0, i64 7), align 4
; NO-FMA-NEXT:    [[FMA0:%.*]] = call float @llvm.fma.f32(float [[A0]], float [[B0]], float [[C0]])
; NO-FMA-NEXT:    [[FMA1:%.*]] = call float @llvm.fma.f32(float [[A1]], float [[B1]], float [[C1]])
; NO-FMA-NEXT:    [[FMA2:%.*]] = call float @llvm.fma.f32(float [[A2]], float [[B2]], float [[C2]])
; NO-FMA-NEXT:    [[FMA3:%.*]] = call float @llvm.fma.f32(float [[A3]], float [[B3]], float [[C3]])
; NO-FMA-NEXT:    [[FMA4:%.*]] = call float @llvm.fma.f32(float [[A4]], float [[B4]], float [[C4]])
; NO-FMA-NEXT:    [[FMA5:%.*]] = call float @llvm.fma.f32(float [[A5]], float [[B5]], float [[C5]])
; NO-FMA-NEXT:    [[FMA6:%.*]] = call float @llvm.fma.f32(float [[A6]], float [[B6]], float [[C6]])
; NO-FMA-NEXT:    [[FMA7:%.*]] = call float @llvm.fma.f32(float [[A7]], float [[B7]], float [[C7]])
; NO-FMA-NEXT:    store float [[FMA0]], float* getelementptr inbounds ([16 x float], [16 x float]* @dst32, i32 0, i64 0), align 4
; NO-FMA-NEXT:    store float [[FMA1]], float* getelementptr inbounds ([16 x float], [16 x float]* @dst32, i32 0, i64 1), align 4
; NO-FMA-NEXT:    store float [[FMA2]], float* getelementptr inbounds ([16 x float], [16 x float]* @dst32, i32 0, i64 2), align 4
; NO-FMA-NEXT:    store float [[FMA3]], float* getelementptr inbounds ([16 x float], [16 x float]* @dst32, i32 0, i64 3), align 4
; NO-FMA-NEXT:    store float [[FMA4]], float* getelementptr inbounds ([16 x float], [16 x float]* @dst32, i32 0, i64 4), align 4
; NO-FMA-NEXT:    store float [[FMA5]], float* getelementptr inbounds ([16 x float], [16 x float]* @dst32, i32 0, i64 5), align 4
; NO-FMA-NEXT:    store float [[FMA6]], float* getelementptr inbounds ([16 x float], [16 x float]* @dst32, i32 0, i64 6), align 4
; NO-FMA-NEXT:    store float [[FMA7]], float* getelementptr inbounds ([16 x float], [16 x float]* @dst32, i32 0, i64 7), align 4
; NO-FMA-NEXT:    ret void
;
; FMA-LABEL: @fma_8f32(
; FMA-NEXT:    [[TMP1:%.*]] = load <8 x float>, <8 x float>* bitcast ([16 x float]* @srcA32 to <8 x float>*), align 4
; FMA-NEXT:    [[TMP2:%.*]] = load <8 x float>, <8 x float>* bitcast ([16 x float]* @srcB32 to <8 x float>*), align 4
; FMA-NEXT:    [[TMP3:%.*]] = load <8 x float>, <8 x float>* bitcast ([16 x float]* @srcC32 to <8 x float>*), align 4
; FMA-NEXT:    [[TMP4:%.*]] = call <8 x float> @llvm.fma.v8f32(<8 x float> [[TMP1]], <8 x float> [[TMP2]], <8 x float> [[TMP3]])
; FMA-NEXT:    store <8 x float> [[TMP4]], <8 x float>* bitcast ([16 x float]* @dst32 to <8 x float>*), align 4
; FMA-NEXT:    ret void
;
  %a0 = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcA32, i32 0, i64 0), align 4
  %a1 = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcA32, i32 0, i64 1), align 4
  %a2 = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcA32, i32 0, i64 2), align 4
  %a3 = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcA32, i32 0, i64 3), align 4
  %a4 = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcA32, i32 0, i64 4), align 4
  %a5 = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcA32, i32 0, i64 5), align 4
  %a6 = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcA32, i32 0, i64 6), align 4
  %a7 = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcA32, i32 0, i64 7), align 4
  %b0 = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcB32, i32 0, i64 0), align 4
  %b1 = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcB32, i32 0, i64 1), align 4
  %b2 = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcB32, i32 0, i64 2), align 4
  %b3 = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcB32, i32 0, i64 3), align 4
  %b4 = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcB32, i32 0, i64 4), align 4
  %b5 = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcB32, i32 0, i64 5), align 4
  %b6 = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcB32, i32 0, i64 6), align 4
  %b7 = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcB32, i32 0, i64 7), align 4
  %c0 = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcC32, i32 0, i64 0), align 4
  %c1 = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcC32, i32 0, i64 1), align 4
  %c2 = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcC32, i32 0, i64 2), align 4
  %c3 = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcC32, i32 0, i64 3), align 4
  %c4 = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcC32, i32 0, i64 4), align 4
  %c5 = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcC32, i32 0, i64 5), align 4
  %c6 = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcC32, i32 0, i64 6), align 4
  %c7 = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcC32, i32 0, i64 7), align 4
  %fma0 = call float @llvm.fma.f32(float %a0, float %b0, float %c0)
  %fma1 = call float @llvm.fma.f32(float %a1, float %b1, float %c1)
  %fma2 = call float @llvm.fma.f32(float %a2, float %b2, float %c2)
  %fma3 = call float @llvm.fma.f32(float %a3, float %b3, float %c3)
  %fma4 = call float @llvm.fma.f32(float %a4, float %b4, float %c4)
  %fma5 = call float @llvm.fma.f32(float %a5, float %b5, float %c5)
  %fma6 = call float @llvm.fma.f32(float %a6, float %b6, float %c6)
  %fma7 = call float @llvm.fma.f32(float %a7, float %b7, float %c7)
  store float %fma0, float* getelementptr inbounds ([16 x float], [16 x float]* @dst32, i32 0, i64 0), align 4
  store float %fma1, float* getelementptr inbounds ([16 x float], [16 x float]* @dst32, i32 0, i64 1), align 4
  store float %fma2, float* getelementptr inbounds ([16 x float], [16 x float]* @dst32, i32 0, i64 2), align 4
  store float %fma3, float* getelementptr inbounds ([16 x float], [16 x float]* @dst32, i32 0, i64 3), align 4
  store float %fma4, float* getelementptr inbounds ([16 x float], [16 x float]* @dst32, i32 0, i64 4), align 4
  store float %fma5, float* getelementptr inbounds ([16 x float], [16 x float]* @dst32, i32 0, i64 5), align 4
  store float %fma6, float* getelementptr inbounds ([16 x float], [16 x float]* @dst32, i32 0, i64 6), align 4
  store float %fma7, float* getelementptr inbounds ([16 x float], [16 x float]* @dst32, i32 0, i64 7), align 4
  ret void
}

define void @fma_16f32() #0 {
; NO-FMA-LABEL: @fma_16f32(
; NO-FMA-NEXT:    [[A0:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcA32, i32 0, i64 0), align 4
; NO-FMA-NEXT:    [[A1:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcA32, i32 0, i64 1), align 4
; NO-FMA-NEXT:    [[A2:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcA32, i32 0, i64 2), align 4
; NO-FMA-NEXT:    [[A3:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcA32, i32 0, i64 3), align 4
; NO-FMA-NEXT:    [[A4:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcA32, i32 0, i64 4), align 4
; NO-FMA-NEXT:    [[A5:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcA32, i32 0, i64 5), align 4
; NO-FMA-NEXT:    [[A6:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcA32, i32 0, i64 6), align 4
; NO-FMA-NEXT:    [[A7:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcA32, i32 0, i64 7), align 4
; NO-FMA-NEXT:    [[A8:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcA32, i32 0, i64 8), align 4
; NO-FMA-NEXT:    [[A9:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcA32, i32 0, i64 9), align 4
; NO-FMA-NEXT:    [[A10:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcA32, i32 0, i64 10), align 4
; NO-FMA-NEXT:    [[A11:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcA32, i32 0, i64 11), align 4
; NO-FMA-NEXT:    [[A12:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcA32, i32 0, i64 12), align 4
; NO-FMA-NEXT:    [[A13:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcA32, i32 0, i64 13), align 4
; NO-FMA-NEXT:    [[A14:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcA32, i32 0, i64 14), align 4
; NO-FMA-NEXT:    [[A15:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcA32, i32 0, i64 15), align 4
; NO-FMA-NEXT:    [[B0:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcB32, i32 0, i64 0), align 4
; NO-FMA-NEXT:    [[B1:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcB32, i32 0, i64 1), align 4
; NO-FMA-NEXT:    [[B2:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcB32, i32 0, i64 2), align 4
; NO-FMA-NEXT:    [[B3:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcB32, i32 0, i64 3), align 4
; NO-FMA-NEXT:    [[B4:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcB32, i32 0, i64 4), align 4
; NO-FMA-NEXT:    [[B5:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcB32, i32 0, i64 5), align 4
; NO-FMA-NEXT:    [[B6:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcB32, i32 0, i64 6), align 4
; NO-FMA-NEXT:    [[B7:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcB32, i32 0, i64 7), align 4
; NO-FMA-NEXT:    [[B8:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcB32, i32 0, i64 8), align 4
; NO-FMA-NEXT:    [[B9:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcB32, i32 0, i64 9), align 4
; NO-FMA-NEXT:    [[B10:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcB32, i32 0, i64 10), align 4
; NO-FMA-NEXT:    [[B11:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcB32, i32 0, i64 11), align 4
; NO-FMA-NEXT:    [[B12:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcB32, i32 0, i64 12), align 4
; NO-FMA-NEXT:    [[B13:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcB32, i32 0, i64 13), align 4
; NO-FMA-NEXT:    [[B14:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcB32, i32 0, i64 14), align 4
; NO-FMA-NEXT:    [[B15:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcB32, i32 0, i64 15), align 4
; NO-FMA-NEXT:    [[C0:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcC32, i32 0, i64 0), align 4
; NO-FMA-NEXT:    [[C1:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcC32, i32 0, i64 1), align 4
; NO-FMA-NEXT:    [[C2:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcC32, i32 0, i64 2), align 4
; NO-FMA-NEXT:    [[C3:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcC32, i32 0, i64 3), align 4
; NO-FMA-NEXT:    [[C4:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcC32, i32 0, i64 4), align 4
; NO-FMA-NEXT:    [[C5:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcC32, i32 0, i64 5), align 4
; NO-FMA-NEXT:    [[C6:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcC32, i32 0, i64 6), align 4
; NO-FMA-NEXT:    [[C7:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcC32, i32 0, i64 7), align 4
; NO-FMA-NEXT:    [[C8:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcC32, i32 0, i64 8), align 4
; NO-FMA-NEXT:    [[C9:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcC32, i32 0, i64 9), align 4
; NO-FMA-NEXT:    [[C10:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcC32, i32 0, i64 10), align 4
; NO-FMA-NEXT:    [[C11:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcC32, i32 0, i64 11), align 4
; NO-FMA-NEXT:    [[C12:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcC32, i32 0, i64 12), align 4
; NO-FMA-NEXT:    [[C13:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcC32, i32 0, i64 13), align 4
; NO-FMA-NEXT:    [[C14:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcC32, i32 0, i64 14), align 4
; NO-FMA-NEXT:    [[C15:%.*]] = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcC32, i32 0, i64 15), align 4
; NO-FMA-NEXT:    [[FMA0:%.*]] = call float @llvm.fma.f32(float [[A0]], float [[B0]], float [[C0]])
; NO-FMA-NEXT:    [[FMA1:%.*]] = call float @llvm.fma.f32(float [[A1]], float [[B1]], float [[C1]])
; NO-FMA-NEXT:    [[FMA2:%.*]] = call float @llvm.fma.f32(float [[A2]], float [[B2]], float [[C2]])
; NO-FMA-NEXT:    [[FMA3:%.*]] = call float @llvm.fma.f32(float [[A3]], float [[B3]], float [[C3]])
; NO-FMA-NEXT:    [[FMA4:%.*]] = call float @llvm.fma.f32(float [[A4]], float [[B4]], float [[C4]])
; NO-FMA-NEXT:    [[FMA5:%.*]] = call float @llvm.fma.f32(float [[A5]], float [[B5]], float [[C5]])
; NO-FMA-NEXT:    [[FMA6:%.*]] = call float @llvm.fma.f32(float [[A6]], float [[B6]], float [[C6]])
; NO-FMA-NEXT:    [[FMA7:%.*]] = call float @llvm.fma.f32(float [[A7]], float [[B7]], float [[C7]])
; NO-FMA-NEXT:    [[FMA8:%.*]] = call float @llvm.fma.f32(float [[A8]], float [[B8]], float [[C8]])
; NO-FMA-NEXT:    [[FMA9:%.*]] = call float @llvm.fma.f32(float [[A9]], float [[B9]], float [[C9]])
; NO-FMA-NEXT:    [[FMA10:%.*]] = call float @llvm.fma.f32(float [[A10]], float [[B10]], float [[C10]])
; NO-FMA-NEXT:    [[FMA11:%.*]] = call float @llvm.fma.f32(float [[A11]], float [[B11]], float [[C11]])
; NO-FMA-NEXT:    [[FMA12:%.*]] = call float @llvm.fma.f32(float [[A12]], float [[B12]], float [[C12]])
; NO-FMA-NEXT:    [[FMA13:%.*]] = call float @llvm.fma.f32(float [[A13]], float [[B13]], float [[C13]])
; NO-FMA-NEXT:    [[FMA14:%.*]] = call float @llvm.fma.f32(float [[A14]], float [[B14]], float [[C14]])
; NO-FMA-NEXT:    [[FMA15:%.*]] = call float @llvm.fma.f32(float [[A15]], float [[B15]], float [[C15]])
; NO-FMA-NEXT:    store float [[FMA0]], float* getelementptr inbounds ([16 x float], [16 x float]* @dst32, i32 0, i64 0), align 4
; NO-FMA-NEXT:    store float [[FMA1]], float* getelementptr inbounds ([16 x float], [16 x float]* @dst32, i32 0, i64 1), align 4
; NO-FMA-NEXT:    store float [[FMA2]], float* getelementptr inbounds ([16 x float], [16 x float]* @dst32, i32 0, i64 2), align 4
; NO-FMA-NEXT:    store float [[FMA3]], float* getelementptr inbounds ([16 x float], [16 x float]* @dst32, i32 0, i64 3), align 4
; NO-FMA-NEXT:    store float [[FMA4]], float* getelementptr inbounds ([16 x float], [16 x float]* @dst32, i32 0, i64 4), align 4
; NO-FMA-NEXT:    store float [[FMA5]], float* getelementptr inbounds ([16 x float], [16 x float]* @dst32, i32 0, i64 5), align 4
; NO-FMA-NEXT:    store float [[FMA6]], float* getelementptr inbounds ([16 x float], [16 x float]* @dst32, i32 0, i64 6), align 4
; NO-FMA-NEXT:    store float [[FMA7]], float* getelementptr inbounds ([16 x float], [16 x float]* @dst32, i32 0, i64 7), align 4
; NO-FMA-NEXT:    store float [[FMA8]], float* getelementptr inbounds ([16 x float], [16 x float]* @dst32, i32 0, i64 8), align 4
; NO-FMA-NEXT:    store float [[FMA9]], float* getelementptr inbounds ([16 x float], [16 x float]* @dst32, i32 0, i64 9), align 4
; NO-FMA-NEXT:    store float [[FMA10]], float* getelementptr inbounds ([16 x float], [16 x float]* @dst32, i32 0, i64 10), align 4
; NO-FMA-NEXT:    store float [[FMA11]], float* getelementptr inbounds ([16 x float], [16 x float]* @dst32, i32 0, i64 11), align 4
; NO-FMA-NEXT:    store float [[FMA12]], float* getelementptr inbounds ([16 x float], [16 x float]* @dst32, i32 0, i64 12), align 4
; NO-FMA-NEXT:    store float [[FMA13]], float* getelementptr inbounds ([16 x float], [16 x float]* @dst32, i32 0, i64 13), align 4
; NO-FMA-NEXT:    store float [[FMA14]], float* getelementptr inbounds ([16 x float], [16 x float]* @dst32, i32 0, i64 14), align 4
; NO-FMA-NEXT:    store float [[FMA15]], float* getelementptr inbounds ([16 x float], [16 x float]* @dst32, i32 0, i64 15), align 4
; NO-FMA-NEXT:    ret void
;
; FMA256-LABEL: @fma_16f32(
; FMA256-NEXT:    [[TMP1:%.*]] = load <8 x float>, <8 x float>* bitcast ([16 x float]* @srcA32 to <8 x float>*), align 4
; FMA256-NEXT:    [[TMP2:%.*]] = load <8 x float>, <8 x float>* bitcast (float* getelementptr inbounds ([16 x float], [16 x float]* @srcA32, i32 0, i64 8) to <8 x float>*), align 4
; FMA256-NEXT:    [[TMP3:%.*]] = load <8 x float>, <8 x float>* bitcast ([16 x float]* @srcB32 to <8 x float>*), align 4
; FMA256-NEXT:    [[TMP4:%.*]] = load <8 x float>, <8 x float>* bitcast (float* getelementptr inbounds ([16 x float], [16 x float]* @srcB32, i32 0, i64 8) to <8 x float>*), align 4
; FMA256-NEXT:    [[TMP5:%.*]] = load <8 x float>, <8 x float>* bitcast ([16 x float]* @srcC32 to <8 x float>*), align 4
; FMA256-NEXT:    [[TMP6:%.*]] = load <8 x float>, <8 x float>* bitcast (float* getelementptr inbounds ([16 x float], [16 x float]* @srcC32, i32 0, i64 8) to <8 x float>*), align 4
; FMA256-NEXT:    [[TMP7:%.*]] = call <8 x float> @llvm.fma.v8f32(<8 x float> [[TMP1]], <8 x float> [[TMP3]], <8 x float> [[TMP5]])
; FMA256-NEXT:    [[TMP8:%.*]] = call <8 x float> @llvm.fma.v8f32(<8 x float> [[TMP2]], <8 x float> [[TMP4]], <8 x float> [[TMP6]])
; FMA256-NEXT:    store <8 x float> [[TMP7]], <8 x float>* bitcast ([16 x float]* @dst32 to <8 x float>*), align 4
; FMA256-NEXT:    store <8 x float> [[TMP8]], <8 x float>* bitcast (float* getelementptr inbounds ([16 x float], [16 x float]* @dst32, i32 0, i64 8) to <8 x float>*), align 4
; FMA256-NEXT:    ret void
;
; FMA512-LABEL: @fma_16f32(
; FMA512-NEXT:    [[TMP1:%.*]] = load <16 x float>, <16 x float>* bitcast ([16 x float]* @srcA32 to <16 x float>*), align 4
; FMA512-NEXT:    [[TMP2:%.*]] = load <16 x float>, <16 x float>* bitcast ([16 x float]* @srcB32 to <16 x float>*), align 4
; FMA512-NEXT:    [[TMP3:%.*]] = load <16 x float>, <16 x float>* bitcast ([16 x float]* @srcC32 to <16 x float>*), align 4
; FMA512-NEXT:    [[TMP4:%.*]] = call <16 x float> @llvm.fma.v16f32(<16 x float> [[TMP1]], <16 x float> [[TMP2]], <16 x float> [[TMP3]])
; FMA512-NEXT:    store <16 x float> [[TMP4]], <16 x float>* bitcast ([16 x float]* @dst32 to <16 x float>*), align 4
; FMA512-NEXT:    ret void
;
  %a0  = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcA32, i32 0, i64  0), align 4
  %a1  = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcA32, i32 0, i64  1), align 4
  %a2  = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcA32, i32 0, i64  2), align 4
  %a3  = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcA32, i32 0, i64  3), align 4
  %a4  = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcA32, i32 0, i64  4), align 4
  %a5  = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcA32, i32 0, i64  5), align 4
  %a6  = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcA32, i32 0, i64  6), align 4
  %a7  = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcA32, i32 0, i64  7), align 4
  %a8  = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcA32, i32 0, i64  8), align 4
  %a9  = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcA32, i32 0, i64  9), align 4
  %a10 = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcA32, i32 0, i64 10), align 4
  %a11 = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcA32, i32 0, i64 11), align 4
  %a12 = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcA32, i32 0, i64 12), align 4
  %a13 = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcA32, i32 0, i64 13), align 4
  %a14 = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcA32, i32 0, i64 14), align 4
  %a15 = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcA32, i32 0, i64 15), align 4
  %b0  = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcB32, i32 0, i64  0), align 4
  %b1  = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcB32, i32 0, i64  1), align 4
  %b2  = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcB32, i32 0, i64  2), align 4
  %b3  = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcB32, i32 0, i64  3), align 4
  %b4  = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcB32, i32 0, i64  4), align 4
  %b5  = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcB32, i32 0, i64  5), align 4
  %b6  = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcB32, i32 0, i64  6), align 4
  %b7  = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcB32, i32 0, i64  7), align 4
  %b8  = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcB32, i32 0, i64  8), align 4
  %b9  = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcB32, i32 0, i64  9), align 4
  %b10 = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcB32, i32 0, i64 10), align 4
  %b11 = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcB32, i32 0, i64 11), align 4
  %b12 = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcB32, i32 0, i64 12), align 4
  %b13 = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcB32, i32 0, i64 13), align 4
  %b14 = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcB32, i32 0, i64 14), align 4
  %b15 = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcB32, i32 0, i64 15), align 4
  %c0  = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcC32, i32 0, i64  0), align 4
  %c1  = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcC32, i32 0, i64  1), align 4
  %c2  = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcC32, i32 0, i64  2), align 4
  %c3  = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcC32, i32 0, i64  3), align 4
  %c4  = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcC32, i32 0, i64  4), align 4
  %c5  = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcC32, i32 0, i64  5), align 4
  %c6  = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcC32, i32 0, i64  6), align 4
  %c7  = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcC32, i32 0, i64  7), align 4
  %c8  = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcC32, i32 0, i64  8), align 4
  %c9  = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcC32, i32 0, i64  9), align 4
  %c10 = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcC32, i32 0, i64 10), align 4
  %c11 = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcC32, i32 0, i64 11), align 4
  %c12 = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcC32, i32 0, i64 12), align 4
  %c13 = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcC32, i32 0, i64 13), align 4
  %c14 = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcC32, i32 0, i64 14), align 4
  %c15 = load float, float* getelementptr inbounds ([16 x float], [16 x float]* @srcC32, i32 0, i64 15), align 4
  %fma0  = call float @llvm.fma.f32(float %a0 , float %b0 , float %c0 )
  %fma1  = call float @llvm.fma.f32(float %a1 , float %b1 , float %c1 )
  %fma2  = call float @llvm.fma.f32(float %a2 , float %b2 , float %c2 )
  %fma3  = call float @llvm.fma.f32(float %a3 , float %b3 , float %c3 )
  %fma4  = call float @llvm.fma.f32(float %a4 , float %b4 , float %c4 )
  %fma5  = call float @llvm.fma.f32(float %a5 , float %b5 , float %c5 )
  %fma6  = call float @llvm.fma.f32(float %a6 , float %b6 , float %c6 )
  %fma7  = call float @llvm.fma.f32(float %a7 , float %b7 , float %c7 )
  %fma8  = call float @llvm.fma.f32(float %a8 , float %b8 , float %c8 )
  %fma9  = call float @llvm.fma.f32(float %a9 , float %b9 , float %c9 )
  %fma10 = call float @llvm.fma.f32(float %a10, float %b10, float %c10)
  %fma11 = call float @llvm.fma.f32(float %a11, float %b11, float %c11)
  %fma12 = call float @llvm.fma.f32(float %a12, float %b12, float %c12)
  %fma13 = call float @llvm.fma.f32(float %a13, float %b13, float %c13)
  %fma14 = call float @llvm.fma.f32(float %a14, float %b14, float %c14)
  %fma15 = call float @llvm.fma.f32(float %a15, float %b15, float %c15)
  store float %fma0 , float* getelementptr inbounds ([16 x float], [16 x float]* @dst32, i32 0, i64  0), align 4
  store float %fma1 , float* getelementptr inbounds ([16 x float], [16 x float]* @dst32, i32 0, i64  1), align 4
  store float %fma2 , float* getelementptr inbounds ([16 x float], [16 x float]* @dst32, i32 0, i64  2), align 4
  store float %fma3 , float* getelementptr inbounds ([16 x float], [16 x float]* @dst32, i32 0, i64  3), align 4
  store float %fma4 , float* getelementptr inbounds ([16 x float], [16 x float]* @dst32, i32 0, i64  4), align 4
  store float %fma5 , float* getelementptr inbounds ([16 x float], [16 x float]* @dst32, i32 0, i64  5), align 4
  store float %fma6 , float* getelementptr inbounds ([16 x float], [16 x float]* @dst32, i32 0, i64  6), align 4
  store float %fma7 , float* getelementptr inbounds ([16 x float], [16 x float]* @dst32, i32 0, i64  7), align 4
  store float %fma8 , float* getelementptr inbounds ([16 x float], [16 x float]* @dst32, i32 0, i64  8), align 4
  store float %fma9 , float* getelementptr inbounds ([16 x float], [16 x float]* @dst32, i32 0, i64  9), align 4
  store float %fma10, float* getelementptr inbounds ([16 x float], [16 x float]* @dst32, i32 0, i64 10), align 4
  store float %fma11, float* getelementptr inbounds ([16 x float], [16 x float]* @dst32, i32 0, i64 11), align 4
  store float %fma12, float* getelementptr inbounds ([16 x float], [16 x float]* @dst32, i32 0, i64 12), align 4
  store float %fma13, float* getelementptr inbounds ([16 x float], [16 x float]* @dst32, i32 0, i64 13), align 4
  store float %fma14, float* getelementptr inbounds ([16 x float], [16 x float]* @dst32, i32 0, i64 14), align 4
  store float %fma15, float* getelementptr inbounds ([16 x float], [16 x float]* @dst32, i32 0, i64 15), align 4
  ret void
}

attributes #0 = { nounwind }