SimplifyLibCalls: Replace more unary libcalls with intrinsics

[davide]

lib/Transforms/Utils/SimplifyLibCalls.cpp

2148–2158

Does your code handle double-float shrinking correctly? (just checking)

2224–2225

This can go away, right? If we lower to an intrinsic we should be able to constant fold it, no? (please add a test to verify)

2233–2234

Ditto.

[arsenm]

Update comment

lib/Transforms/Utils/SimplifyLibCalls.cpp

2148–2158

The InstCombine part of the patch adds the shrinking with the intrinsics

[efriedma]

For anyone else following along with a copy of IEEE 754, this is the roundToIntegral* family of functions.

Aren't llvm.rint() and llvm.nearbyint() precisely equivalent? We should probably get rid of llvm.rint()... (we can add it back once we actually start modeling FP exceptions).

test/Transforms/InstCombine/win-math.ll

273

Windows doesn't have round()? Weird.

[davide]

In D28484#640523, @efriedma wrote:

For anyone else following along with a copy of IEEE 754, this is the roundToIntegral* family of functions.

Aren't llvm.rint() and llvm.nearbyint() precisely equivalent? We should probably get rid of llvm.rint()... (we can add it back once we actually start modeling FP exceptions).

I think so.

test/Transforms/InstCombine/win-math.ll

273

I'm surprised too https://msdn.microsoft.com/en-us/library/dn353646.aspx
This also compiles http://rextester.com/MUCW42464 (unless I'm misunderstanding).

[arsenm]

In D28484#640523, @efriedma wrote:

For anyone else following along with a copy of IEEE 754, this is the roundToIntegral* family of functions.

Aren't llvm.rint() and llvm.nearbyint() precisely equivalent? We should probably get rid of llvm.rint()... (we can add it back once we actually start modeling FP exceptions).

Should I drop rint handling here or keep it until later when it's removed?

[davide]

In D28484#647688, @arsenm wrote:

In D28484#640523, @efriedma wrote:

For anyone else following along with a copy of IEEE 754, this is the roundToIntegral* family of functions.

Aren't llvm.rint() and llvm.nearbyint() precisely equivalent? We should probably get rid of llvm.rint()... (we can add it back once we actually start modeling FP exceptions).

Should I drop rint handling here or keep it until later when it's removed?

I'd say keep it until we agree we're going to remove it.

lib/Transforms/InstCombine/InstCombineCalls.cpp

1642

Why this newline.

[efriedma]

I would say map @rint() to @llvm.nearbyint... but otherwise leave in the @llvm.rint() handling. We can kill off @llvm.rint() in a followup.

[arsenm]

Only use nearbyint

[arsenm]

ping

[efriedma]

The intrinsics aren't used nearly as much as the function calls, so it's possible this will expose issues... but that's fine, I think.

[arsenm]

r292855

include/llvm/Transforms/Utils/SimplifyLibCalls.h

private:
   Value *optimizeMemSet(CallInst *CI, IRBuilder<> &B);
   // Wrapper for all String/Memory Library Call Optimizations
   Value *optimizeStringMemoryLibCall(CallInst *CI, IRBuilder<> &B);
 
   // Math Library Optimizations
   Value *optimizeCos(CallInst *CI, IRBuilder<> &B);
   Value *optimizePow(CallInst *CI, IRBuilder<> &B);
   Value *optimizeExp2(CallInst *CI, IRBuilder<> &B);
-  Value *optimizeFabs(CallInst *CI, IRBuilder<> &B);
   Value *optimizeFMinFMax(CallInst *CI, IRBuilder<> &B);
   Value *optimizeLog(CallInst *CI, IRBuilder<> &B);
   Value *optimizeSqrt(CallInst *CI, IRBuilder<> &B);
   Value *optimizeSinCosPi(CallInst *CI, IRBuilder<> &B);
   Value *optimizeTan(CallInst *CI, IRBuilder<> &B);
 
   // Integer Library Call Optimizations
   Value *optimizeFFS(CallInst *CI, IRBuilder<> &B);

lib/Transforms/InstCombine/InstCombineCalls.cpp

case Intrinsic::fabs: {
     Constant *LHS, *RHS;
     if (match(II->getArgOperand(0),
               m_Select(m_Value(Cond), m_Constant(LHS), m_Constant(RHS)))) {
       CallInst *Call0 = Builder->CreateCall(II->getCalledFunction(), {LHS});
       CallInst *Call1 = Builder->CreateCall(II->getCalledFunction(), {RHS});
       return SelectInst::Create(Cond, Call0, Call1);
     }
 
+    LLVM_FALLTHROUGH;
+  }
+  case Intrinsic::ceil:
+  case Intrinsic::floor:
+  case Intrinsic::round:
+  case Intrinsic::nearbyint:
+  case Intrinsic::trunc: {
     Value *ExtSrc;
     if (match(II->getArgOperand(0), m_FPExt(m_Value(ExtSrc))) &&
         II->getArgOperand(0)->hasOneUse()) {
       // fabs (fpext x) -> fpext (fabs x)
-      Value *F = Intrinsic::getDeclaration(II->getModule(), Intrinsic::fabs,
+      Value *F = Intrinsic::getDeclaration(II->getModule(), II->getIntrinsicID(),
                                            { ExtSrc->getType() });
       CallInst *NewFabs = Builder->CreateCall(F, ExtSrc);
       NewFabs->copyFastMathFlags(II);
       NewFabs->takeName(II);
       return new FPExtInst(NewFabs, II->getType());
     }
 
     break;

lib/Transforms/InstCombine/InstCombineCasts.cpp

Value *RHSTrunc = Builder->CreateFPTrunc(SI->getOperand(2),
                                              CI.getType());
     return SelectInst::Create(SI->getOperand(0), LHSTrunc, RHSTrunc);
   }
 
   IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI.getOperand(0));
   if (II) {
     switch (II->getIntrinsicID()) {
     default: break;
-    case Intrinsic::fabs: {
+    case Intrinsic::fabs:
+    case Intrinsic::ceil:
+    case Intrinsic::floor:
+    case Intrinsic::rint:
+    case Intrinsic::round:
+    case Intrinsic::nearbyint:
+    case Intrinsic::trunc: {
+      // Do unary FP operation on smaller type.
       // (fptrunc (fabs x)) -> (fabs (fptrunc x))
       Value *InnerTrunc = Builder->CreateFPTrunc(II->getArgOperand(0),
                                                  CI.getType());
       Type *IntrinsicType[] = { CI.getType() };
       Function *Overload = Intrinsic::getDeclaration(
         CI.getModule(), II->getIntrinsicID(), IntrinsicType);
 
       SmallVector<OperandBundleDef, 1> OpBundles;

lib/Transforms/Utils/SimplifyLibCalls.cpp

static Value *optimizeUnaryDoubleFP(CallInst *CI, IRBuilder<> &B,
   } else {
     // The call is a library call rather than an intrinsic.
     V = emitUnaryFloatFnCall(V, Callee->getName(), B, Callee->getAttributes());
   }
 
   return B.CreateFPExt(V, B.getDoubleTy());
 }
 
+// Replace a libcall \p CI with a call to intrinsic \p IID
+static Value *replaceUnaryCall(CallInst *CI, IRBuilder<> &B, Intrinsic::ID IID) {
+  // Propagate fast-math flags from the existing call to the new call.
+  IRBuilder<>::FastMathFlagGuard Guard(B);
+  B.setFastMathFlags(CI->getFastMathFlags());
+
+  Module *M = CI->getModule();
+  Value *V = CI->getArgOperand(0);
+  Function *F = Intrinsic::getDeclaration(M, IID, CI->getType());
+  CallInst *NewCall = B.CreateCall(F, V);
+  NewCall->takeName(CI);
+  return NewCall;
+}
+
 /// Shrink double -> float for binary functions like 'fmin/fmax'.
 static Value *optimizeBinaryDoubleFP(CallInst *CI, IRBuilder<> &B) {
   Function *Callee = CI->getCalledFunction();
   // We know this libcall has a valid prototype, but we don't know which.
   if (!CI->getType()->isDoubleTy())
     return nullptr;
 
   // If this is something like 'fmin((double)floatval1, (double)floatval2)',
if (LdExpArg) {
         CI->setCallingConv(F->getCallingConv());
 
       return CI;
     }
   }
   return Ret;
 }
 
-Value *LibCallSimplifier::optimizeFabs(CallInst *CI, IRBuilder<> &B) {
-  Function *Callee = CI->getCalledFunction();
-  IRBuilder<>::FastMathFlagGuard Guard(B);
-  B.setFastMathFlags(CI->getFastMathFlags());
-
-  // fabs/fabsf -> llvm.fabs.*
-  Value *F = Intrinsic::getDeclaration(Callee->getParent(), Intrinsic::fabs,
-                                       CI->getType());
-  Value *NewCall = B.CreateCall(F, { CI->getArgOperand(0) });
-  NewCall->takeName(CI);
-  return NewCall;
-}
-
 Value *LibCallSimplifier::optimizeFMinFMax(CallInst *CI, IRBuilder<> &B) {
   Function *Callee = CI->getCalledFunction();
   // If we can shrink the call to a float function rather than a double
   // function, do that first.
   StringRef Name = Callee->getName();
   if ((Name == "fmin" || Name == "fmax") && hasFloatVersion(Name))
     if (Value *Ret = optimizeBinaryDoubleFP(CI, B))
       return Ret;
case LibFunc::powl:
       return optimizePow(CI, Builder);
     case LibFunc::exp2l:
     case LibFunc::exp2:
     case LibFunc::exp2f:
       return optimizeExp2(CI, Builder);
     case LibFunc::fabsf:
     case LibFunc::fabs:
     case LibFunc::fabsl:
-      return optimizeFabs(CI, Builder);
+      return replaceUnaryCall(CI, Builder, Intrinsic::fabs);
     case LibFunc::sqrtf:
     case LibFunc::sqrt:
     case LibFunc::sqrtl:
       return optimizeSqrt(CI, Builder);
     case LibFunc::ffs:
     case LibFunc::ffsl:
     case LibFunc::ffsll:
       return optimizeFFS(CI, Builder);
if (TLI->getLibFunc(*Callee, Func) && TLI->has(Func)) {
     case LibFunc::perror:
       return optimizeErrorReporting(CI, Builder);
     case LibFunc::vfprintf:
     case LibFunc::fiprintf:
       return optimizeErrorReporting(CI, Builder, 0);
     case LibFunc::fputc:
       return optimizeErrorReporting(CI, Builder, 1);
     case LibFunc::ceil:
+      return replaceUnaryCall(CI, Builder, Intrinsic::ceil);
     case LibFunc::floor:
-    case LibFunc::rint:
+      return replaceUnaryCall(CI, Builder, Intrinsic::floor);
     case LibFunc::round:
+      return replaceUnaryCall(CI, Builder, Intrinsic::round);
     case LibFunc::nearbyint:
+    case LibFunc::rint:
+      return replaceUnaryCall(CI, Builder, Intrinsic::nearbyint);
     case LibFunc::trunc:
-      if (hasFloatVersion(FuncName))
-        return optimizeUnaryDoubleFP(CI, Builder, false);
-      return nullptr;
+      return replaceUnaryCall(CI, Builder, Intrinsic::trunc);
     case LibFunc::acos:
     case LibFunc::acosh:
     case LibFunc::asin:
     case LibFunc::asinh:
     case LibFunc::atan:
     case LibFunc::atanh:
     case LibFunc::cbrt:
     case LibFunc::cosh:
     case LibFunc::exp:

 //   * exp(log(x))  -> x
 //
 // log, logf, logl:
 //   * log(exp(x))   -> x
 //   * log(exp(y))   -> y*log(e)
 //   * log(exp10(y)) -> y*log(10)
 //   * log(sqrt(x))  -> 0.5*log(x)
 //
-// lround, lroundf, lroundl:
-//   * lround(cnst) -> cnst'
-//
 // pow, powf, powl:
 //   * pow(sqrt(x),y) -> pow(x,y*0.5)
 //   * pow(pow(x,y),z)-> pow(x,y*z)
 //
-// round, roundf, roundl:
-//   * round(cnst) -> cnst'
-//
 // signbit:
 //   * signbit(cnst) -> cnst'
 //   * signbit(nncst) -> 0 (if pstv is a non-negative constant)
 //
 // sqrt, sqrtf, sqrtl:
 //   * sqrt(expN(x))  -> expN(x*0.5)
 //   * sqrt(Nroot(x)) -> pow(x,1/(2*N))
 //   * sqrt(pow(x,y)) -> pow(|x|,y*0.5)
 //
-// trunc, truncf, truncl:
-//   * trunc(cnst) -> cnst'
-//
-//
 
 //===----------------------------------------------------------------------===//
 // Fortified Library Call Optimizations
 //===----------------------------------------------------------------------===//
 
 bool FortifiedLibCallSimplifier::isFortifiedCallFoldable(CallInst *CI,
                                                          unsigned ObjSizeOp,
                                                          unsigned SizeOp,
                                                          bool isString) {

test/Transforms/InstCombine/double-float-shrink-2.ll

 ; RUN: opt < %s -instcombine -S -mtriple "i386-pc-linux" | FileCheck -check-prefix=DO-SIMPLIFY %s
 ; RUN: opt < %s -instcombine -S -mtriple "i386-pc-win32" | FileCheck -check-prefix=DONT-SIMPLIFY %s
 ; RUN: opt < %s -instcombine -S -mtriple "x86_64-pc-win32" | FileCheck -check-prefix=C89-SIMPLIFY %s
 ; RUN: opt < %s -instcombine -S -mtriple "i386-pc-mingw32" | FileCheck -check-prefix=DO-SIMPLIFY %s
 ; RUN: opt < %s -instcombine -S -mtriple "x86_64-pc-mingw32" | FileCheck -check-prefix=DO-SIMPLIFY %s
 ; RUN: opt < %s -instcombine -S -mtriple "sparc-sun-solaris" | FileCheck -check-prefix=DO-SIMPLIFY %s
 
-; DO-SIMPLIFY: call float @floorf(
-; DO-SIMPLIFY: call float @ceilf(
-; DO-SIMPLIFY: call float @roundf(
-; DO-SIMPLIFY: call float @nearbyintf(
-; DO-SIMPLIFY: call float @truncf(
+; DO-SIMPLIFY: call float @llvm.floor.f32(
+; DO-SIMPLIFY: call float @llvm.ceil.f32(
+; DO-SIMPLIFY: call float @llvm.round.f32(
+; DO-SIMPLIFY: call float @llvm.nearbyint.f32(
+; DO-SIMPLIFY: call float @llvm.trunc.f32(
 ; DO-SIMPLIFY: call float @llvm.fabs.f32(
 ; DO-SIMPLIFY: call fast float @llvm.fabs.f32(
 
-; C89-SIMPLIFY: call float @floorf(
-; C89-SIMPLIFY: call float @ceilf(
+; C89-SIMPLIFY: call float @llvm.floor.f32(
+; C89-SIMPLIFY: call float @llvm.ceil.f32(
 ; C89-SIMPLIFY: call double @round(
 ; C89-SIMPLIFY: call double @nearbyint(
 
-; DONT-SIMPLIFY: call double @floor(
-; DONT-SIMPLIFY: call double @ceil(
+; DONT-SIMPLIFY: call float @llvm.floor.f32(
+; DONT-SIMPLIFY: call float @llvm.ceil.f32(
 ; DONT-SIMPLIFY: call double @round(
 ; DONT-SIMPLIFY: call double @nearbyint(
 ; DONT-SIMPLIFY: call double @trunc(
 
 ; This is replaced with the intrinsic, which does the right thing on
 ; all platforms.
 ; DONT-SIMPLIFY: call float @llvm.fabs.f32(
 

test/Transforms/InstCombine/float-shrink-compare.ll

 ; RUN: opt -S -instcombine < %s | FileCheck %s
 target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64-S128"
 target triple = "x86_64-apple-macosx10.8.0"
 
 define i32 @test1(float %x, float %y) nounwind uwtable {
-  %1 = fpext float %x to double
-  %2 = call double @ceil(double %1) nounwind readnone
-  %3 = fpext float %y to double
-  %4 = fcmp oeq double %2, %3
-  %5 = zext i1 %4 to i32
-  ret i32 %5
+  %x.ext = fpext float %x to double
+  %ceil = call double @ceil(double %x.ext) nounwind readnone
+  %ext.y = fpext float %y to double
+  %cmp = fcmp oeq double %ceil, %ext.y
+  %cmp.ext = zext i1 %cmp to i32
+  ret i32 %cmp.ext
 ; CHECK-LABEL: @test1(
-; CHECK-NEXT: %ceilf = call float @ceilf(float %x)
-; CHECK-NEXT: fcmp oeq float %ceilf, %y
+; CHECK-NEXT: %ceil = call float @llvm.ceil.f32(float %x)
+; CHECK-NEXT: fcmp oeq float %ceil, %y
+}
+
+define i32 @test1_intrin(float %x, float %y) nounwind uwtable {
+  %x.ext = fpext float %x to double
+  %ceil = call double @llvm.ceil.f64(double %x.ext) nounwind readnone
+  %ext.y = fpext float %y to double
+  %cmp = fcmp oeq double %ceil, %ext.y
+  %cmp.ext = zext i1 %cmp to i32
+  ret i32 %cmp.ext
+; CHECK-LABEL: @test1_intrin(
+; CHECK-NEXT: %ceil = call float @llvm.ceil.f32(float %x)
+; CHECK-NEXT: fcmp oeq float %ceil, %y
 }
 
 define i32 @test2(float %x, float %y) nounwind uwtable {
-  %1 = fpext float %x to double
-  %2 = call double @fabs(double %1) nounwind readnone
-  %3 = fpext float %y to double
-  %4 = fcmp oeq double %2, %3
-  %5 = zext i1 %4 to i32
-  ret i32 %5
+  %x.ext = fpext float %x to double
+  %fabs = call double @fabs(double %x.ext) nounwind readnone
+  %y.ext = fpext float %y to double
+  %cmp = fcmp oeq double %fabs, %y.ext
+  %cmp.ext = zext i1 %cmp to i32
+  ret i32 %cmp.ext
 ; CHECK-LABEL: @test2(
-; CHECK-NEXT: [[FABS:%[0-9]+]] = call float @llvm.fabs.f32(float %x)
-; CHECK-NEXT: fcmp oeq float [[FABS]], %y
+; CHECK-NEXT: %fabs = call float @llvm.fabs.f32(float %x)
+; CHECK-NEXT: fcmp oeq float %fabs, %y
+}
+
+define i32 @test2_intrin(float %x, float %y) nounwind uwtable {
+  %x.ext = fpext float %x to double
+  %fabs = call double @llvm.fabs.f64(double %x.ext) nounwind readnone
+  %y.ext = fpext float %y to double
+  %cmp = fcmp oeq double %fabs, %y.ext
+  %cmp.ext = zext i1 %cmp to i32
+  ret i32 %cmp.ext
+; CHECK-LABEL: @test2_intrin(
+; CHECK-NEXT: %fabs = call float @llvm.fabs.f32(float %x)
+; CHECK-NEXT: fcmp oeq float %fabs, %y
 }
 
 define i32 @fmf_test2(float %x, float %y) nounwind uwtable {
   %1 = fpext float %x to double
   %2 = call nnan double @fabs(double %1) nounwind readnone
   %3 = fpext float %y to double
   %4 = fcmp oeq double %2, %3
   %5 = zext i1 %4 to i32
   ret i32 %5
 ; CHECK-LABEL: @fmf_test2(
 ; CHECK-NEXT: [[FABS:%[0-9]+]] = call nnan float @llvm.fabs.f32(float %x)
 ; CHECK-NEXT: fcmp oeq float [[FABS]], %y
 }
 
 define i32 @test3(float %x, float %y) nounwind uwtable {
-  %1 = fpext float %x to double
-  %2 = call double @floor(double %1) nounwind readnone
-  %3 = fpext float %y to double
-  %4 = fcmp oeq double %2, %3
-  %5 = zext i1 %4 to i32
-  ret i32 %5
+  %x.ext = fpext float %x to double
+  %floor = call double @floor(double %x.ext) nounwind readnone
+  %y.ext = fpext float %y to double
+  %cmp = fcmp oeq double %floor, %y.ext
+  %cmp.ext = zext i1 %cmp to i32
+  ret i32 %cmp.ext
 ; CHECK-LABEL: @test3(
-; CHECK-NEXT: %floorf = call float @floorf(float %x)
-; CHECK-NEXT: fcmp oeq float %floorf, %y
+; CHECK-NEXT: %floor = call float @llvm.floor.f32(float %x)
+; CHECK-NEXT: fcmp oeq float %floor, %y
+}
+
+
+define i32 @test3_intrin(float %x, float %y) nounwind uwtable {
+  %x.ext = fpext float %x to double
+  %floor = call double @llvm.floor.f64(double %x.ext) nounwind readnone
+  %y.ext = fpext float %y to double
+  %cmp = fcmp oeq double %floor, %y.ext
+  %cmp.ext = zext i1 %cmp to i32
+  ret i32 %cmp.ext
+; CHECK-LABEL: @test3_intrin(
+; CHECK-NEXT: %floor = call float @llvm.floor.f32(float %x)
+; CHECK-NEXT: fcmp oeq float %floor, %y
 }
 
 define i32 @test4(float %x, float %y) nounwind uwtable {
-  %1 = fpext float %x to double
-  %2 = call double @nearbyint(double %1) nounwind
-  %3 = fpext float %y to double
-  %4 = fcmp oeq double %2, %3
-  %5 = zext i1 %4 to i32
-  ret i32 %5
+  %x.ext = fpext float %x to double
+  %nearbyint = call double @nearbyint(double %x.ext) nounwind
+  %y.ext = fpext float %y to double
+  %cmp = fcmp oeq double %nearbyint, %y.ext
+  %cmp.ext = zext i1 %cmp to i32
+  ret i32 %cmp.ext
 ; CHECK-LABEL: @test4(
-; CHECK-NEXT: %nearbyintf = call float @nearbyintf(float %x)
-; CHECK-NEXT: fcmp oeq float %nearbyintf, %y
+; CHECK-NEXT: %nearbyint = call float @llvm.nearbyint.f32(float %x)
+; CHECK-NEXT: fcmp oeq float %nearbyint, %y
+}
+
+define i32 @shrink_nearbyint_intrin(float %x, float %y) nounwind uwtable {
+  %x.ext = fpext float %x to double
+  %nearbyint = call double @llvm.nearbyint.f64(double %x.ext) nounwind
+  %y.ext = fpext float %y to double
+  %cmp = fcmp oeq double %nearbyint, %y.ext
+  %cmp.ext = zext i1 %cmp to i32
+  ret i32 %cmp.ext
+; CHECK-LABEL: @shrink_nearbyint_intrin(
+; CHECK-NEXT: %nearbyint = call float @llvm.nearbyint.f32(float %x)
+; CHECK-NEXT: fcmp oeq float %nearbyint, %y
 }
 
 define i32 @test5(float %x, float %y) nounwind uwtable {
-  %1 = fpext float %x to double
-  %2 = call double @rint(double %1) nounwind
-  %3 = fpext float %y to double
-  %4 = fcmp oeq double %2, %3
-  %5 = zext i1 %4 to i32
-  ret i32 %5
+  %x.ext = fpext float %x to double
+  %rint = call double @rint(double %x.ext) nounwind
+  %y.ext = fpext float %y to double
+  %cmp = fcmp oeq double %rint, %y.ext
+  %cmp.ext = zext i1 %cmp to i32
+  ret i32 %cmp.ext
 ; CHECK-LABEL: @test5(
-; CHECK-NEXT: %rintf = call float @rintf(float %x)
-; CHECK-NEXT: fcmp oeq float %rintf, %y
+; CHECK-NEXT: %rint = call float @llvm.nearbyint.f32(float %x)
+; CHECK-NEXT: fcmp oeq float %rint, %y
 }
 
 define i32 @test6(float %x, float %y) nounwind uwtable {
-  %1 = fpext float %x to double
-  %2 = call double @round(double %1) nounwind readnone
-  %3 = fpext float %y to double
-  %4 = fcmp oeq double %2, %3
-  %5 = zext i1 %4 to i32
-  ret i32 %5
+  %x.ext = fpext float %x to double
+  %round = call double @round(double %x.ext) nounwind readnone
+  %y.ext = fpext float %y to double
+  %cmp = fcmp oeq double %round, %y.ext
+  %cmp.ext = zext i1 %cmp to i32
+  ret i32 %cmp.ext
 ; CHECK-LABEL: @test6(
-; CHECK-NEXT: %roundf = call float @roundf(float %x)
-; CHECK-NEXT: fcmp oeq float %roundf, %y
+; CHECK-NEXT: %round = call float @llvm.round.f32(float %x)
+; CHECK-NEXT: fcmp oeq float %round, %y
+}
+
+define i32 @test6_intrin(float %x, float %y) nounwind uwtable {
+  %x.ext = fpext float %x to double
+  %round = call double @llvm.round.f64(double %x.ext) nounwind readnone
+  %y.ext = fpext float %y to double
+  %cmp = fcmp oeq double %round, %y.ext
+  %cmp.ext = zext i1 %cmp to i32
+  ret i32 %cmp.ext
+; CHECK-LABEL: @test6_intrin(
+; CHECK-NEXT: %round = call float @llvm.round.f32(float %x)
+; CHECK-NEXT: fcmp oeq float %round, %y
 }
 
 define i32 @test7(float %x, float %y) nounwind uwtable {
-  %1 = fpext float %x to double
-  %2 = call double @trunc(double %1) nounwind
-  %3 = fpext float %y to double
-  %4 = fcmp oeq double %2, %3
-  %5 = zext i1 %4 to i32
-  ret i32 %5
+  %x.ext = fpext float %x to double
+  %trunc = call double @trunc(double %x.ext) nounwind
+  %y.ext = fpext float %y to double
+  %cmp = fcmp oeq double %trunc, %y.ext
+  %cmp.ext = zext i1 %cmp to i32
+  ret i32 %cmp.ext
 ; CHECK-LABEL: @test7(
-; CHECK-NEXT: %truncf = call float @truncf(float %x)
-; CHECK-NEXT: fcmp oeq float %truncf, %y
+; CHECK-NEXT: %trunc = call float @llvm.trunc.f32(float %x)
+; CHECK-NEXT: fcmp oeq float %trunc, %y
+}
+
+define i32 @test7_intrin(float %x, float %y) nounwind uwtable {
+  %x.ext = fpext float %x to double
+  %trunc = call double @llvm.trunc.f64(double %x.ext) nounwind
+  %y.ext = fpext float %y to double
+  %cmp = fcmp oeq double %trunc, %y.ext
+  %cmp.ext = zext i1 %cmp to i32
+  ret i32 %cmp.ext
+; CHECK-LABEL: @test7_intrin(
+; CHECK-NEXT: %trunc = call float @llvm.trunc.f32(float %x)
+; CHECK-NEXT: fcmp oeq float %trunc, %y
 }
 
 define i32 @test8(float %x, float %y) nounwind uwtable {
-  %1 = fpext float %y to double
-  %2 = fpext float %x to double
-  %3 = call double @ceil(double %2) nounwind readnone
-  %4 = fcmp oeq double %1, %3
-  %5 = zext i1 %4 to i32
-  ret i32 %5
+  %x.ext = fpext float %x to double
+  %y.ext = fpext float %y to double
+  %ceil = call double @ceil(double %x.ext) nounwind readnone
+  %cmp = fcmp oeq double %y.ext, %ceil
+  %cmp.ext = zext i1 %cmp to i32
+  ret i32 %cmp.ext
 ; CHECK-LABEL: @test8(
-; CHECK-NEXT: %ceilf = call float @ceilf(float %x)
-; CHECK-NEXT: fcmp oeq float %ceilf, %y
+; CHECK-NEXT: %ceil = call float @llvm.ceil.f32(float %x)
+; CHECK-NEXT: fcmp oeq float %ceil, %y
+}
+
+define i32 @test8_intrin(float %x, float %y) nounwind uwtable {
+  %x.ext = fpext float %x to double
+  %y.ext = fpext float %y to double
+  %ceil = call double @llvm.ceil.f64(double %x.ext) nounwind readnone
+  %cmp = fcmp oeq double %y.ext, %ceil
+  %cmp.ext = zext i1 %cmp to i32
+  ret i32 %cmp.ext
+; CHECK-LABEL: @test8_intrin(
+; CHECK-NEXT: %ceil = call float @llvm.ceil.f32(float %x)
+; CHECK-NEXT: fcmp oeq float %ceil, %y
 }
 
 define i32 @test9(float %x, float %y) nounwind uwtable {
   %x.ext = fpext float %x to double
   %y.ext = fpext float %y to double
   %fabs = call double @fabs(double %x.ext) nounwind readnone
   %cmp = fcmp oeq double %y.ext, %fabs
   %cmp.ext = zext i1 %cmp to i32
   ret i32 %cmp.ext
 ; CHECK-LABEL: @test9(
 ; CHECK-NEXT: %fabs = call float @llvm.fabs.f32(float %x)
 ; CHECK-NEXT: fcmp oeq float %fabs, %y
 }
 
+define i32 @test9_intrin(float %x, float %y) nounwind uwtable {
+  %x.ext = fpext float %x to double
+  %y.ext = fpext float %y to double
+  %fabs = call double @llvm.fabs.f64(double %x.ext) nounwind readnone
+  %cmp = fcmp oeq double %y.ext, %fabs
+  %cmp.ext = zext i1 %cmp to i32
+  ret i32 %cmp.ext
+; CHECK-LABEL: @test9_intrin(
+; CHECK-NEXT: %fabs = call float @llvm.fabs.f32(float %x)
+; CHECK-NEXT: fcmp oeq float %fabs, %y
+}
+
 define i32 @test10(float %x, float %y) nounwind uwtable {
-  %1 = fpext float %y to double
-  %2 = fpext float %x to double
-  %3 = call double @floor(double %2) nounwind readnone
-  %4 = fcmp oeq double %1, %3
-  %5 = zext i1 %4 to i32
-  ret i32 %5
+  %x.ext = fpext float %x to double
+  %y.ext = fpext float %y to double
+  %floor = call double @floor(double %x.ext) nounwind readnone
+  %cmp = fcmp oeq double %floor, %y.ext
+  %cmp.ext = zext i1 %cmp to i32
+  ret i32 %cmp.ext
 ; CHECK-LABEL: @test10(
-; CHECK-NEXT: %floorf = call float @floorf(float %x)
-; CHECK-NEXT: fcmp oeq float %floorf, %y
+; CHECK-NEXT: %floor = call float @llvm.floor.f32(float %x)
+; CHECK-NEXT: fcmp oeq float %floor, %y
+}
+
+define i32 @test10_intrin(float %x, float %y) nounwind uwtable {
+  %x.ext = fpext float %x to double
+  %y.ext = fpext float %y to double
+  %floor = call double @llvm.floor.f64(double %x.ext) nounwind readnone
+  %cmp = fcmp oeq double %floor, %y.ext
+  %cmp.ext = zext i1 %cmp to i32
+  ret i32 %cmp.ext
+; CHECK-LABEL: @test10_intrin(
+; CHECK-NEXT: %floor = call float @llvm.floor.f32(float %x)
+; CHECK-NEXT: fcmp oeq float %floor, %y
 }
 
 define i32 @test11(float %x, float %y) nounwind uwtable {
-  %1 = fpext float %y to double
-  %2 = fpext float %x to double
-  %3 = call double @nearbyint(double %2) nounwind
-  %4 = fcmp oeq double %1, %3
-  %5 = zext i1 %4 to i32
-  ret i32 %5
+  %x.ext = fpext float %x to double
+  %y.ext = fpext float %y to double
+  %nearbyint = call double @nearbyint(double %x.ext) nounwind
+  %cmp = fcmp oeq double %nearbyint, %y.ext
+  %cmp.ext = zext i1 %cmp to i32
+  ret i32 %cmp.ext
 ; CHECK-LABEL: @test11(
-; CHECK-NEXT: %nearbyintf = call float @nearbyintf(float %x)
-; CHECK-NEXT: fcmp oeq float %nearbyintf, %y
+; CHECK-NEXT: %nearbyint = call float @llvm.nearbyint.f32(float %x)
+; CHECK-NEXT: fcmp oeq float %nearbyint, %y
+}
+
+
+define i32 @test11_intrin(float %x, float %y) nounwind uwtable {
+  %x.ext = fpext float %x to double
+  %y.ext = fpext float %y to double
+  %nearbyint = call double @llvm.nearbyint.f64(double %x.ext) nounwind
+  %cmp = fcmp oeq double %nearbyint, %y.ext
+  %cmp.ext = zext i1 %cmp to i32
+  ret i32 %cmp.ext
+; CHECK-LABEL: @test11_intrin(
+; CHECK-NEXT: %nearbyint = call float @llvm.nearbyint.f32(float %x)
+; CHECK-NEXT: fcmp oeq float %nearbyint, %y
 }
 
 define i32 @test12(float %x, float %y) nounwind uwtable {
-  %1 = fpext float %y to double
-  %2 = fpext float %x to double
-  %3 = call double @rint(double %2) nounwind
-  %4 = fcmp oeq double %1, %3
-  %5 = zext i1 %4 to i32
-  ret i32 %5
+  %x.ext = fpext float %x to double
+  %y.ext = fpext float %y to double
+  %rint = call double @rint(double %x.ext) nounwind
+  %cmp = fcmp oeq double %y.ext, %rint
+  %cmp.ext = zext i1 %cmp to i32
+  ret i32 %cmp.ext
 ; CHECK-LABEL: @test12(
-; CHECK-NEXT: %rintf = call float @rintf(float %x)
-; CHECK-NEXT: fcmp oeq float %rintf, %y
+; CHECK-NEXT: %rint = call float @llvm.nearbyint.f32(float %x)
+; CHECK-NEXT: fcmp oeq float %rint, %y
 }
 
 define i32 @test13(float %x, float %y) nounwind uwtable {
-  %1 = fpext float %y to double
-  %2 = fpext float %x to double
-  %3 = call double @round(double %2) nounwind readnone
-  %4 = fcmp oeq double %1, %3
-  %5 = zext i1 %4 to i32
-  ret i32 %5
+  %x.ext = fpext float %x to double
+  %y.ext = fpext float %y to double
+  %round = call double @round(double %x.ext) nounwind readnone
+  %cmp = fcmp oeq double %y.ext, %round
+  %cmp.ext = zext i1 %cmp to i32
+  ret i32 %cmp.ext
 ; CHECK-LABEL: @test13(
-; CHECK-NEXT: %roundf = call float @roundf(float %x)
-; CHECK-NEXT: fcmp oeq float %roundf, %y
+; CHECK-NEXT: %round = call float @llvm.round.f32(float %x)
+; CHECK-NEXT: fcmp oeq float %round, %y
+}
+
+define i32 @test13_intrin(float %x, float %y) nounwind uwtable {
+  %x.ext = fpext float %x to double
+  %y.ext = fpext float %y to double
+  %round = call double @llvm.round.f64(double %x.ext) nounwind readnone
+  %cmp = fcmp oeq double %y.ext, %round
+  %cmp.ext = zext i1 %cmp to i32
+  ret i32 %cmp.ext
+; CHECK-LABEL: @test13_intrin(
+; CHECK-NEXT: %round = call float @llvm.round.f32(float %x)
+; CHECK-NEXT: fcmp oeq float %round, %y
 }
 
 define i32 @test14(float %x, float %y) nounwind uwtable {
-  %1 = fpext float %y to double
-  %2 = fpext float %x to double
-  %3 = call double @trunc(double %2) nounwind
-  %4 = fcmp oeq double %1, %3
-  %5 = zext i1 %4 to i32
-  ret i32 %5
+  %x.ext = fpext float %x to double
+  %y.ext = fpext float %y to double
+  %trunc = call double @trunc(double %x.ext) nounwind
+  %cmp = fcmp oeq double %y.ext, %trunc
+  %cmp.ext = zext i1 %cmp to i32
+  ret i32 %cmp.ext
 ; CHECK-LABEL: @test14(
-; CHECK-NEXT: %truncf = call float @truncf(float %x)
-; CHECK-NEXT: fcmp oeq float %truncf, %y
+; CHECK-NEXT: %trunc = call float @llvm.trunc.f32(float %x)
+; CHECK-NEXT: fcmp oeq float %trunc, %y
+}
+
+define i32 @test14_intrin(float %x, float %y) nounwind uwtable {
+  %x.ext = fpext float %x to double
+  %y.ext = fpext float %y to double
+  %trunc = call double @llvm.trunc.f64(double %x.ext) nounwind
+  %cmp = fcmp oeq double %y.ext, %trunc
+  %cmp.ext = zext i1 %cmp to i32
+  ret i32 %cmp.ext
+; CHECK-LABEL: @test14_intrin(
+; CHECK-NEXT: %trunc = call float @llvm.trunc.f32(float %x)
+; CHECK-NEXT: fcmp oeq float %trunc, %y
 }
 
 define i32 @test15(float %x, float %y, float %z) nounwind uwtable {
   %1 = fpext float %x to double
   %2 = fpext float %y to double
   %3 = call double @fmin(double %1, double %2) nounwind
   %4 = fpext float %z to double
   %5 = fcmp oeq double %3, %4

 declare double @copysign(double, double) nounwind readnone
 declare double @floor(double) nounwind readnone
 declare double @nearbyint(double) nounwind readnone
 declare double @rint(double) nounwind readnone
 declare double @round(double) nounwind readnone
 declare double @trunc(double) nounwind readnone
 declare double @fmin(double, double) nounwind readnone
 declare double @fmax(double, double) nounwind readnone
+
+declare double @llvm.fabs.f64(double) nounwind readnone
+declare double @llvm.ceil.f64(double) nounwind readnone
+declare double @llvm.floor.f64(double) nounwind readnone
+declare double @llvm.nearbyint.f64(double) nounwind readnone
+declare double @llvm.round.f64(double) nounwind readnone
+declare double @llvm.trunc.f64(double) nounwind readnone

test/Transforms/InstCombine/win-math.ll

; WIN32: double @atan2
     %4 = fptrunc double %3 to float
     ret float %4
 }
 
 declare double @ceil(double %x)
 define float @float_ceil(float %x) nounwind readnone {
 ; WIN32-LABEL: @float_ceil(
 ; WIN32-NOT: float @ceilf
-; WIN32: double @ceil
+; WIN32: float @llvm.ceil.f32
 ; WIN64-LABEL: @float_ceil(
-; WIN64: float @ceilf
+; WIN64: float @llvm.ceil.f32
 ; WIN64-NOT: double @ceil
 ; MINGW32-LABEL: @float_ceil(
-; MINGW32: float @ceilf
+; MINGW32: float @llvm.ceil.f32
 ; MINGW32-NOT: double @ceil
 ; MINGW64-LABEL: @float_ceil(
-; MINGW64: float @ceilf
+; MINGW64: float @llvm.ceil.f32
 ; MINGW64-NOT: double @ceil
     %1 = fpext float %x to double
     %2 = call double @ceil(double %1)
     %3 = fptrunc double %2 to float
     ret float %3
 }
 
 declare double @_copysign(double %x)
; WIN64: double @fabs
     %4 = fptrunc double %3 to float
     ret float %4
 }
 
 declare double @floor(double %x)
 define float @float_floor(float %x) nounwind readnone {
 ; WIN32-LABEL: @float_floor(
 ; WIN32-NOT: float @floorf
-; WIN32: double @floor
+; WIN32: float @llvm.floor.f32
 ; WIN64-LABEL: @float_floor(
-; WIN64: float @floorf
+; WIN64: float @llvm.floor.f32
 ; WIN64-NOT: double @floor
 ; MINGW32-LABEL: @float_floor(
-; MINGW32: float @floorf
+; MINGW32: float @llvm.floor.f32
 ; MINGW32-NOT: double @floor
 ; MINGW64-LABEL: @float_floor(
-; MINGW64: float @floorf
+; MINGW64: float @llvm.floor.f32
 ; MINGW64-NOT: double @floor
     %1 = fpext float %x to double
     %2 = call double @floor(double %1)
     %3 = fptrunc double %2 to float
     ret float %3
 }
 
 declare double @fmod(double %x, double %y)

 define float @float_round(float %x) nounwind readnone {
 ; WIN32-LABEL: @float_round(
 ; WIN32-NOT: float @roundf
 ; WIN32: double @round
 ; WIN64-LABEL: @float_round(
 ; WIN64-NOT: float @roundf
 ; WIN64: double @round
 ; MINGW32-LABEL: @float_round(
-; MINGW32: float @roundf
+; MINGW32: float @llvm.round.f32
 ; MINGW32-NOT: double @round
 ; MINGW64-LABEL: @float_round(
-; MINGW64: float @roundf
+; MINGW64: float @llvm.round.f32
 ; MINGW64-NOT: double @round
     %1 = fpext float %x to double
     %2 = call double @round(double %1)
     %3 = fptrunc double %2 to float
     ret float %3
 }
 
 declare float @powf(float, float)
 
 ; win32 lacks sqrtf&fabsf, win64 lacks fabsf, but
 ; calls to the intrinsics can be emitted instead.
 define float @float_powsqrt(float %x) nounwind readnone {
 ; WIN32-LABEL: @float_powsqrt(