Shrinkify libcalls: use float versions of double libm functions with fast-math (bug 17850)

[spatel]

Summary

When a call to a double-precision libm function has fast-math semantics (via function attribute for now because there is no IR-level FMF on calls), we can avoid fpext/fptrunc operations and use the float version of the call if the input and output are both float.

We already do this optimization using a command-line option; this patch just adds the ability for fast-math to use the existing functionality.

I moved the cl::opt from InstructionCombining into SimplifyLibCalls because it's only ever used internally to that class.

Is it possible to conditionalize test case code from a RUN line? If yes, I could combine all of the new test cases with the existing test file instead of creating a new file.

This patch should solve: http://llvm.org/bugs/show_bug.cgi?id=17850

[hfinkel]

LGTM, thanks!

Don't duplicate the tests... we don't need to test the command line option specifically, just remove those tests and add these.

[spatel]

Closed by commit rL220390 (authored by @spatel).

llvm/trunk/include/llvm/Transforms/Utils/SimplifyLibCalls.h

private:
   const DataLayout *DL;
   const TargetLibraryInfo *TLI;
   bool UnsafeFPShrink;
 
 protected:
   ~LibCallSimplifier() {}
 
 public:
-  LibCallSimplifier(const DataLayout *TD, const TargetLibraryInfo *TLI,
-                    bool UnsafeFPShrink);
+  LibCallSimplifier(const DataLayout *TD, const TargetLibraryInfo *TLI);
 
   /// optimizeCall - Take the given call instruction and return a more
   /// optimal value to replace the instruction with or 0 if a more
   /// optimal form can't be found.  Note that the returned value may
   /// be equal to the instruction being optimized.  In this case all
   /// other instructions that use the given instruction were modified
   /// and the given instruction is dead.
   Value *optimizeCall(CallInst *CI);

llvm/trunk/lib/Transforms/InstCombine/InstructionCombining.cpp

 STATISTIC(NumCombined , "Number of insts combined");
 STATISTIC(NumConstProp, "Number of constant folds");
 STATISTIC(NumDeadInst , "Number of dead inst eliminated");
 STATISTIC(NumSunkInst , "Number of instructions sunk");
 STATISTIC(NumExpand,    "Number of expansions");
 STATISTIC(NumFactor   , "Number of factorizations");
 STATISTIC(NumReassoc  , "Number of reassociations");
 
-static cl::opt<bool>
-    EnableUnsafeFPShrink("enable-double-float-shrink", cl::Hidden,
-                         cl::init(false),
-                         cl::desc("Enable unsafe double to float "
-                                  "shrinking for math lib calls"));
-
 // Initialization Routines
 void llvm::initializeInstCombine(PassRegistry &Registry) {
   initializeInstCombinerPass(Registry);
 }
 
 void LLVMInitializeInstCombine(LLVMPassRegistryRef R) {
   initializeInstCombine(*unwrap(R));
 }

 
 namespace {
 class InstCombinerLibCallSimplifier final : public LibCallSimplifier {
   InstCombiner *IC;
 public:
   InstCombinerLibCallSimplifier(const DataLayout *DL,
                                 const TargetLibraryInfo *TLI,
                                 InstCombiner *IC)
-    : LibCallSimplifier(DL, TLI, EnableUnsafeFPShrink) {
+    : LibCallSimplifier(DL, TLI) {
     this->IC = IC;
   }
 
   /// replaceAllUsesWith - override so that instruction replacement
   /// can be defined in terms of the instruction combiner framework.
   void replaceAllUsesWith(Instruction *I, Value *With) const override {
     IC->ReplaceInstUsesWith(*I, With);
   }

llvm/trunk/lib/Transforms/Utils/SimplifyLibCalls.cpp

 
 using namespace llvm;
 using namespace PatternMatch;
 
 static cl::opt<bool>
     ColdErrorCalls("error-reporting-is-cold", cl::init(true), cl::Hidden,
                    cl::desc("Treat error-reporting calls as cold"));
 
+static cl::opt<bool>
+    EnableUnsafeFPShrink("enable-double-float-shrink", cl::Hidden,
+                         cl::init(false),
+                         cl::desc("Enable unsafe double to float "
+                                  "shrinking for math lib calls"));
+
+
 //===----------------------------------------------------------------------===//
 // Helper Functions
 //===----------------------------------------------------------------------===//
 
 static bool ignoreCallingConv(LibFunc::Func Func) {
   switch (Func) {
   case LibFunc::abs:
   case LibFunc::labs:
if (CI->isNoBuiltin())
     return nullptr;
 
   LibFunc::Func Func;
   Function *Callee = CI->getCalledFunction();
   StringRef FuncName = Callee->getName();
   IRBuilder<> Builder(CI);
   bool isCallingConvC = CI->getCallingConv() == llvm::CallingConv::C;
 
+  // Command-line parameter overrides function attribute.
+  if (EnableUnsafeFPShrink.getNumOccurrences() > 0)
+    UnsafeFPShrink = EnableUnsafeFPShrink;
+  else if (Callee->hasFnAttribute("unsafe-fp-math")) {
+    // FIXME: This is the same problem as described in optimizeSqrt().
+    // If calls gain access to IR-level FMF, then use that instead of a
+    // function attribute.
+
+    // Check for unsafe-fp-math = true.
+    Attribute Attr = Callee->getFnAttribute("unsafe-fp-math");
+    if (Attr.getValueAsString() == "true")
+      UnsafeFPShrink = true;
+  }
+
   // Next check for intrinsics.
   if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI)) {
     if (!isCallingConvC)
       return nullptr;
     switch (II->getIntrinsicID()) {
     case Intrinsic::pow:
       return optimizePow(CI, Builder);
     case Intrinsic::exp2:
if (FuncName.endswith("_chk")) {
     else if (FuncName == "__stpncpy_chk")
       return optimizeStrNCpyChk(CI, Builder);
   }
 
   return nullptr;
 }
 
 LibCallSimplifier::LibCallSimplifier(const DataLayout *DL,
-                                     const TargetLibraryInfo *TLI,
-                                     bool UnsafeFPShrink) :
+                                     const TargetLibraryInfo *TLI) :
                                      DL(DL),
                                      TLI(TLI),
-                                     UnsafeFPShrink(UnsafeFPShrink) {
+                                     UnsafeFPShrink(false) {
 }
 
 void LibCallSimplifier::replaceAllUsesWith(Instruction *I, Value *With) const {
   I->replaceAllUsesWith(With);
   I->eraseFromParent();
 }
 
 // TODO:

llvm/trunk/test/Transforms/InstCombine/double-float-shrink-1.ll

-; RUN: opt < %s -instcombine -enable-double-float-shrink -S | FileCheck %s
+; 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-unknown-linux-gnu"
 
-define float @acos_test(float %f) nounwind readnone {
-; CHECK: acos_test
+; Check for and against shrinkage when using the
+; unsafe-fp-math function attribute on a math lib
+; function. This optimization may be overridden by
+; the -enable-double-float-shrink option.
+; PR17850: http://llvm.org/bugs/show_bug.cgi?id=17850
+
+define float @acos_test(float %f)   {
    %conv = fpext float %f to double
    %call = call double @acos(double %conv)
    %conv1 = fptrunc double %call to float
    ret float %conv1
+; CHECK-LABEL: acos_test
 ; CHECK: call float @acosf(float %f)
 }
 
-define double @acos_test2(float %f) nounwind readnone {
-; CHECK: acos_test2
+define double @acos_test2(float %f)   {
    %conv = fpext float %f to double
    %call = call double @acos(double %conv)
    ret double %call
+; CHECK-LABEL: acos_test2
 ; CHECK: call double @acos(double %conv)
 }
 
-define float @acosh_test(float %f) nounwind readnone {
-; CHECK: acosh_test
+define float @acosh_test(float %f)   {
    %conv = fpext float %f to double
    %call = call double @acosh(double %conv)
    %conv1 = fptrunc double %call to float
    ret float %conv1
+; CHECK-LABEL: acosh_test
 ; CHECK: call float @acoshf(float %f)
 }
 
-define double @acosh_test2(float %f) nounwind readnone {
-; CHECK: acosh_test2
+define double @acosh_test2(float %f)   {
    %conv = fpext float %f to double
    %call = call double @acosh(double %conv)
    ret double %call
+; CHECK-LABEL: acosh_test2
 ; CHECK: call double @acosh(double %conv)
 }
 
-define float @asin_test(float %f) nounwind readnone {
-; CHECK: asin_test
+define float @asin_test(float %f)   {
    %conv = fpext float %f to double
    %call = call double @asin(double %conv)
    %conv1 = fptrunc double %call to float
    ret float %conv1
+; CHECK-LABEL: asin_test
 ; CHECK: call float @asinf(float %f)
 }
 
-define double @asin_test2(float %f) nounwind readnone {
-; CHECK: asin_test2
+define double @asin_test2(float %f)   {
    %conv = fpext float %f to double
    %call = call double @asin(double %conv)
    ret double %call
+; CHECK-LABEL: asin_test2
 ; CHECK: call double @asin(double %conv)
 }
 
-define float @asinh_test(float %f) nounwind readnone {
-; CHECK: asinh_test
+define float @asinh_test(float %f)   {
    %conv = fpext float %f to double
    %call = call double @asinh(double %conv)
    %conv1 = fptrunc double %call to float
    ret float %conv1
+; CHECK-LABEL: asinh_test
 ; CHECK: call float @asinhf(float %f)
 }
 
-define double @asinh_test2(float %f) nounwind readnone {
-; CHECK: asinh_test2
+define double @asinh_test2(float %f)   {
    %conv = fpext float %f to double
    %call = call double @asinh(double %conv)
    ret double %call
+; CHECK-LABEL: asinh_test2
 ; CHECK: call double @asinh(double %conv)
 }
 
-define float @atan_test(float %f) nounwind readnone {
-; CHECK: atan_test
+define float @atan_test(float %f)   {
    %conv = fpext float %f to double
    %call = call double @atan(double %conv)
    %conv1 = fptrunc double %call to float
    ret float %conv1
+; CHECK-LABEL: atan_test
 ; CHECK: call float @atanf(float %f)
 }
 
-define double @atan_test2(float %f) nounwind readnone {
-; CHECK: atan_test2
+define double @atan_test2(float %f)   {
    %conv = fpext float %f to double
    %call = call double @atan(double %conv)
    ret double %call
+; CHECK-LABEL: atan_test2
 ; CHECK: call double @atan(double %conv)
 }
-define float @atanh_test(float %f) nounwind readnone {
-; CHECK: atanh_test
+define float @atanh_test(float %f)   {
    %conv = fpext float %f to double
    %call = call double @atanh(double %conv)
    %conv1 = fptrunc double %call to float
    ret float %conv1
+; CHECK-LABEL: atanh_test
 ; CHECK: call float @atanhf(float %f)
 }
 
-define double @atanh_test2(float %f) nounwind readnone {
-; CHECK: atanh_test2
+define double @atanh_test2(float %f)   {
     %conv = fpext float %f to double
     %call = call double @atanh(double %conv)
     ret double %call
+; CHECK-LABEL: atanh_test2
 ; CHECK: call double @atanh(double %conv)
 }
-define float @cbrt_test(float %f) nounwind readnone {
-; CHECK: cbrt_test
+define float @cbrt_test(float %f)   {
    %conv = fpext float %f to double
    %call = call double @cbrt(double %conv)
    %conv1 = fptrunc double %call to float
    ret float %conv1
+; CHECK-LABEL: cbrt_test
 ; CHECK: call float @cbrtf(float %f)
 }
 
-define double @cbrt_test2(float %f) nounwind readnone {
-; CHECK: cbrt_test2
+define double @cbrt_test2(float %f)   {
    %conv = fpext float %f to double
    %call = call double @cbrt(double %conv)
    ret double %call
+; CHECK-LABEL: cbrt_test2
 ; CHECK: call double @cbrt(double %conv)
 }
-define float @exp_test(float %f) nounwind readnone {
-; CHECK: exp_test
+define float @exp_test(float %f)   {
    %conv = fpext float %f to double
    %call = call double @exp(double %conv)
    %conv1 = fptrunc double %call to float
    ret float %conv1
+; CHECK-LABEL: exp_test
 ; CHECK: call float @expf(float %f)
 }
 
-define double @exp_test2(float %f) nounwind readnone {
-; CHECK: exp_test2
+define double @exp_test2(float %f)   {
    %conv = fpext float %f to double
    %call = call double @exp(double %conv)
    ret double %call
+; CHECK-LABEL: exp_test2
 ; CHECK: call double @exp(double %conv)
 }
-define float @expm1_test(float %f) nounwind readnone {
-; CHECK: expm1_test
+define float @expm1_test(float %f)   {
    %conv = fpext float %f to double
    %call = call double @expm1(double %conv)
    %conv1 = fptrunc double %call to float
    ret float %conv1
+; CHECK-LABEL: expm1_test
 ; CHECK: call float @expm1f(float %f)
 }
 
-define double @expm1_test2(float %f) nounwind readnone {
-; CHECK: expm1_test2
+define double @expm1_test2(float %f)   {
    %conv = fpext float %f to double
    %call = call double @expm1(double %conv)
    ret double %call
+; CHECK-LABEL: expm1_test2
 ; CHECK: call double @expm1(double %conv)
 }
-define float @exp10_test(float %f) nounwind readnone {
-; CHECK: exp10_test
+define float @exp10_test(float %f)   {
    %conv = fpext float %f to double
    %call = call double @exp10(double %conv)
    %conv1 = fptrunc double %call to float
    ret float %conv1
-; FIXME: Re-enable this when Linux allows transforming this again, or when we
-; can use builtin attributes to test the transform regardless of OS.
-; DISABLED-CHECK: call float @exp10f(float %f)
+; CHECK-LABEL: exp10_test
 ; CHECK: call double @exp10(double %conv)
 }
 
-define double @exp10_test2(float %f) nounwind readnone {
-; CHECK: exp10_test2
+define double @exp10_test2(float %f)   {
    %conv = fpext float %f to double
    %call = call double @exp10(double %conv)
    ret double %call
+; CHECK-LABEL: exp10_test2
 ; CHECK: call double @exp10(double %conv)
 }
-define float @log_test(float %f) nounwind readnone {
-; CHECK: log_test
+define float @log_test(float %f)   {
    %conv = fpext float %f to double
    %call = call double @log(double %conv)
    %conv1 = fptrunc double %call to float
    ret float %conv1
+; CHECK-LABEL: log_test
 ; CHECK: call float @logf(float %f)
 }
 
-define double @log_test2(float %f) nounwind readnone {
-; CHECK: log_test2
+define double @log_test2(float %f)   {
    %conv = fpext float %f to double
    %call = call double @log(double %conv)
    ret double %call
+; CHECK-LABEL: log_test2
 ; CHECK: call double @log(double %conv)
 }
-define float @log10_test(float %f) nounwind readnone {
-; CHECK: log10_test
+define float @log10_test(float %f)   {
    %conv = fpext float %f to double
    %call = call double @log10(double %conv)
    %conv1 = fptrunc double %call to float
    ret float %conv1
+; CHECK-LABEL: log10_test
 ; CHECK: call float @log10f(float %f)
 }
 
-define double @log10_test2(float %f) nounwind readnone {
-; CHECK: log10_test2
+define double @log10_test2(float %f) {
    %conv = fpext float %f to double
    %call = call double @log10(double %conv)
    ret double %call
+; CHECK-LABEL: log10_test2
 ; CHECK: call double @log10(double %conv)
 }
-define float @log1p_test(float %f) nounwind readnone {
-; CHECK: log1p_test
+define float @log1p_test(float %f)   {
    %conv = fpext float %f to double
    %call = call double @log1p(double %conv)
    %conv1 = fptrunc double %call to float
    ret float %conv1
+; CHECK-LABEL: log1p_test
 ; CHECK: call float @log1pf(float %f)
 }
 
-define double @log1p_test2(float %f) nounwind readnone {
-; CHECK: log1p_test2
+define double @log1p_test2(float %f)   {
    %conv = fpext float %f to double
    %call = call double @log1p(double %conv)
    ret double %call
+; CHECK-LABEL: log1p_test2
 ; CHECK: call double @log1p(double %conv)
 }
-define float @log2_test(float %f) nounwind readnone {
-; CHECK: log2_test
+define float @log2_test(float %f)   {
    %conv = fpext float %f to double
    %call = call double @log2(double %conv)
    %conv1 = fptrunc double %call to float
    ret float %conv1
+; CHECK-LABEL: log2_test
 ; CHECK: call float @log2f(float %f)
 }
 
-define double @log2_test2(float %f) nounwind readnone {
-; CHECK: log2_test2
+define double @log2_test2(float %f)   {
    %conv = fpext float %f to double
    %call = call double @log2(double %conv)
    ret double %call
+; CHECK-LABEL: log2_test2
 ; CHECK: call double @log2(double %conv)
 }
-define float @logb_test(float %f) nounwind readnone {
-; CHECK: logb_test
+define float @logb_test(float %f)   {
    %conv = fpext float %f to double
    %call = call double @logb(double %conv)
    %conv1 = fptrunc double %call to float
    ret float %conv1
+; CHECK-LABEL: logb_test
 ; CHECK: call float @logbf(float %f)
 }
 
-define double @logb_test2(float %f) nounwind readnone {
-; CHECK: logb_test2
+define double @logb_test2(float %f)   {
    %conv = fpext float %f to double
    %call = call double @logb(double %conv)
    ret double %call
+; CHECK-LABEL: logb_test2
 ; CHECK: call double @logb(double %conv)
 }
-define float @sin_test(float %f) nounwind readnone {
-; CHECK: sin_test
+define float @sin_test(float %f)   {
    %conv = fpext float %f to double
    %call = call double @sin(double %conv)
    %conv1 = fptrunc double %call to float
    ret float %conv1
+; CHECK-LABEL: sin_test
 ; CHECK: call float @sinf(float %f)
 }
 
-define double @sin_test2(float %f) nounwind readnone {
-; CHECK: sin_test2
+define double @sin_test2(float %f) {
    %conv = fpext float %f to double
    %call = call double @sin(double %conv)
    ret double %call
+; CHECK-LABEL: sin_test2
 ; CHECK: call double @sin(double %conv)
 }
 
-define float @sqrt_test(float %f) nounwind readnone {
-; CHECK: sqrt_test
+define float @sqrt_test(float %f) {
    %conv = fpext float %f to double
    %call = call double @sqrt(double %conv)
    %conv1 = fptrunc double %call to float
    ret float %conv1
+; CHECK-LABEL: sqrt_test
 ; CHECK: call float @sqrtf(float %f)
 }
 
-define float @sqrt_int_test(float %f) nounwind readnone {
-; CHECK: sqrt_int_test
+define float @sqrt_int_test(float %f) {
    %conv = fpext float %f to double
    %call = call double @llvm.sqrt.f64(double %conv)
    %conv1 = fptrunc double %call to float
    ret float %conv1
+; CHECK-LABEL: sqrt_int_test
 ; CHECK: call float @llvm.sqrt.f32(float %f)
 }
 
-define double @sqrt_test2(float %f) nounwind readnone {
-; CHECK: sqrt_test2
+define double @sqrt_test2(float %f) {
    %conv = fpext float %f to double
    %call = call double @sqrt(double %conv)
    ret double %call
+; CHECK-LABEL: sqrt_test2
 ; CHECK: call double @sqrt(double %conv)
 }
-define float @tan_test(float %f) nounwind readnone {
-; CHECK: tan_test
+define float @tan_test(float %f) {
    %conv = fpext float %f to double
    %call = call double @tan(double %conv)
    %conv1 = fptrunc double %call to float
    ret float %conv1
+; CHECK-LABEL: tan_test
 ; CHECK: call float @tanf(float %f)
 }
 
-define double @tan_test2(float %f) nounwind readnone {
-; CHECK: tan_test2
+define double @tan_test2(float %f) {
    %conv = fpext float %f to double
    %call = call double @tan(double %conv)
    ret double %call
+; CHECK-LABEL: tan_test2
 ; CHECK: call double @tan(double %conv)
 }
-define float @tanh_test(float %f) nounwind readnone {
-; CHECK: tanh_test
+define float @tanh_test(float %f) {
    %conv = fpext float %f to double
    %call = call double @tanh(double %conv)
    %conv1 = fptrunc double %call to float
    ret float %conv1
+; CHECK-LABEL: tanh_test
 ; CHECK: call float @tanhf(float %f)
 }
 
-define double @tanh_test2(float %f) nounwind readnone {
-; CHECK: tanh_test2
+define double @tanh_test2(float %f) {
    %conv = fpext float %f to double
    %call = call double @tanh(double %conv)
    ret double %call
+; CHECK-LABEL: tanh_test2
 ; CHECK: call double @tanh(double %conv)
 }
 
-declare double @tanh(double) nounwind readnone
-declare double @tan(double) nounwind readnone
-declare double @sqrt(double) nounwind readnone
-declare double @sin(double) nounwind readnone
-declare double @log2(double) nounwind readnone
-declare double @log1p(double) nounwind readnone
-declare double @log10(double) nounwind readnone
-declare double @log(double) nounwind readnone
-declare double @logb(double) nounwind readnone
-declare double @exp10(double) nounwind readnone
-declare double @expm1(double) nounwind readnone
-declare double @exp(double) nounwind readnone
-declare double @cbrt(double) nounwind readnone
-declare double @atanh(double) nounwind readnone
-declare double @atan(double) nounwind readnone
-declare double @acos(double) nounwind readnone
-declare double @acosh(double) nounwind readnone
-declare double @asin(double) nounwind readnone
-declare double @asinh(double) nounwind readnone
+declare double @tanh(double) #1
+declare double @tan(double) #1
+declare double @sqrt(double) #1
+declare double @sin(double) #1
+declare double @log2(double) #1
+declare double @log1p(double) #1
+declare double @log10(double) #1
+declare double @log(double) #1
+declare double @logb(double) #1
+declare double @exp10(double) #1
+declare double @expm1(double) #1
+declare double @exp(double) #1
+declare double @cbrt(double) #1
+declare double @atanh(double) #1
+declare double @atan(double) #1
+declare double @acos(double) #1
+declare double @acosh(double) #1
+declare double @asin(double) #1
+declare double @asinh(double) #1
 
-declare double @llvm.sqrt.f64(double) nounwind readnone
+declare double @llvm.sqrt.f64(double) #1
+attributes #1 = { "unsafe-fp-math"="true" }