DCE math library calls with a constant operand

[efriedma]

Summary

On platforms which use -fmath-errno, math libcalls without any uses require some extra checks to figure out if they are actually dead.

Fixes https://llvm.org/bugs/show_bug.cgi?id=30464 .

[majnemer]

LGTM, you could use an ImmutableCallSite if you wish.

[davidxl]

lib/Analysis/ConstantFolding.cpp

1971 ↗

(On Diff #75798)

There are lots of duplicated logic here with LibcallShinkWrap pass. I think the right way to do is let the later passes to clean up the dead calls. This can be done by pushing libcall shrinkwrap later in the pipeline.

[efriedma]

lib/Analysis/ConstantFolding.cpp

1971 ↗

(On Diff #75798)

The case which I originally ran into involves pow(), and LibCallsShrinkWrap can't really deal with that effectively, as far as I can tell.

It's probably possible to refactor the logic for some of the other libcalls... but we want to perform the check in isInstructionTriviallyDead to get rid of dead calls as soon as possible, so they don't block other optimizations.

[davidxl]

lib/Analysis/ConstantFolding.cpp

1971 ↗

(On Diff #75798)

Refactoring is ideal, but it may not be easy to do. Please add a comment here to point to the refactoring opportunity in the future.

llvm/trunk/include/llvm/Analysis/ConstantFolding.h

 //===----------------------------------------------------------------------===//
 
 #ifndef LLVM_ANALYSIS_CONSTANTFOLDING_H
 #define LLVM_ANALYSIS_CONSTANTFOLDING_H
 
 namespace llvm {
 class APInt;
 template <typename T> class ArrayRef;
+class CallSite;
 class Constant;
 class ConstantExpr;
 class ConstantVector;
 class DataLayout;
 class Function;
 class GlobalValue;
 class Instruction;
 class TargetLibraryInfo;

 /// canConstantFoldCallTo - Return true if its even possible to fold a call to
 /// the specified function.
 bool canConstantFoldCallTo(const Function *F);
 
 /// ConstantFoldCall - Attempt to constant fold a call to the specified function
 /// with the specified arguments, returning null if unsuccessful.
 Constant *ConstantFoldCall(Function *F, ArrayRef<Constant *> Operands,
                            const TargetLibraryInfo *TLI = nullptr);
+
+/// \brief Check whether the given call has no side-effects.
+/// Specifically checks for math routimes which sometimes set errno.
+bool isMathLibCallNoop(CallSite CS, const TargetLibraryInfo *TLI);
 }
 
 #endif

llvm/trunk/lib/Analysis/ConstantFolding.cpp

llvm::ConstantFoldCall(Function *F, ArrayRef<Constant *> Operands,
   Type *Ty = F->getReturnType();
 
   if (auto *VTy = dyn_cast<VectorType>(Ty))
     return ConstantFoldVectorCall(Name, F->getIntrinsicID(), VTy, Operands,
                                   F->getParent()->getDataLayout(), TLI);
 
   return ConstantFoldScalarCall(Name, F->getIntrinsicID(), Ty, Operands, TLI);
 }
+
+bool llvm::isMathLibCallNoop(CallSite CS, const TargetLibraryInfo *TLI) {
+  // FIXME: Refactor this code; this duplicates logic in LibCallsShrinkWrap
+  // (and to some extent ConstantFoldScalarCall).
+  Function *F = CS.getCalledFunction();
+  if (!F)
+    return false;
+
+  LibFunc::Func Func;
+  if (!TLI || !TLI->getLibFunc(*F, Func))
+    return false;
+
+  if (CS.getNumArgOperands() == 1) {
+    if (ConstantFP *OpC = dyn_cast<ConstantFP>(CS.getArgOperand(0))) {
+      const APFloat &Op = OpC->getValueAPF();
+      switch (Func) {
+      case LibFunc::logl:
+      case LibFunc::log:
+      case LibFunc::logf:
+      case LibFunc::log2l:
+      case LibFunc::log2:
+      case LibFunc::log2f:
+      case LibFunc::log10l:
+      case LibFunc::log10:
+      case LibFunc::log10f:
+        return Op.isNaN() || (!Op.isZero() && !Op.isNegative());
+
+      case LibFunc::expl:
+      case LibFunc::exp:
+      case LibFunc::expf:
+        // FIXME: These boundaries are slightly conservative.
+        if (OpC->getType()->isDoubleTy())
+          return Op.compare(APFloat(-745.0)) != APFloat::cmpLessThan &&
+                 Op.compare(APFloat(709.0)) != APFloat::cmpGreaterThan;
+        if (OpC->getType()->isFloatTy())
+          return Op.compare(APFloat(-103.0f)) != APFloat::cmpLessThan &&
+                 Op.compare(APFloat(88.0f)) != APFloat::cmpGreaterThan;
+        break;
+
+      case LibFunc::exp2l:
+      case LibFunc::exp2:
+      case LibFunc::exp2f:
+        // FIXME: These boundaries are slightly conservative.
+        if (OpC->getType()->isDoubleTy())
+          return Op.compare(APFloat(-1074.0)) != APFloat::cmpLessThan &&
+                 Op.compare(APFloat(1023.0)) != APFloat::cmpGreaterThan;
+        if (OpC->getType()->isFloatTy())
+          return Op.compare(APFloat(-149.0f)) != APFloat::cmpLessThan &&
+                 Op.compare(APFloat(127.0f)) != APFloat::cmpGreaterThan;
+        break;
+
+      case LibFunc::sinl:
+      case LibFunc::sin:
+      case LibFunc::sinf:
+      case LibFunc::cosl:
+      case LibFunc::cos:
+      case LibFunc::cosf:
+        return !Op.isInfinity();
+
+      case LibFunc::tanl:
+      case LibFunc::tan:
+      case LibFunc::tanf: {
+        // FIXME: Stop using the host math library.
+        // FIXME: The computation isn't done in the right precision.
+        Type *Ty = OpC->getType();
+        if (Ty->isDoubleTy() || Ty->isFloatTy() || Ty->isHalfTy()) {
+          double OpV = getValueAsDouble(OpC);
+          return ConstantFoldFP(tan, OpV, Ty) != nullptr;
+        }
+        break;
+      }
+
+      case LibFunc::asinl:
+      case LibFunc::asin:
+      case LibFunc::asinf:
+      case LibFunc::acosl:
+      case LibFunc::acos:
+      case LibFunc::acosf:
+        return Op.compare(APFloat(Op.getSemantics(), "-1")) !=
+                   APFloat::cmpLessThan &&
+               Op.compare(APFloat(Op.getSemantics(), "1")) !=
+                   APFloat::cmpGreaterThan;
+
+      case LibFunc::sinh:
+      case LibFunc::cosh:
+      case LibFunc::sinhf:
+      case LibFunc::coshf:
+      case LibFunc::sinhl:
+      case LibFunc::coshl:
+        // FIXME: These boundaries are slightly conservative.
+        if (OpC->getType()->isDoubleTy())
+          return Op.compare(APFloat(-710.0)) != APFloat::cmpLessThan &&
+                 Op.compare(APFloat(710.0)) != APFloat::cmpGreaterThan;
+        if (OpC->getType()->isFloatTy())
+          return Op.compare(APFloat(-89.0f)) != APFloat::cmpLessThan &&
+                 Op.compare(APFloat(89.0f)) != APFloat::cmpGreaterThan;
+        break;
+
+      case LibFunc::sqrtl:
+      case LibFunc::sqrt:
+      case LibFunc::sqrtf:
+        return Op.isNaN() || Op.isZero() || !Op.isNegative();
+
+      // FIXME: Add more functions: sqrt_finite, atanh, expm1, log1p,
+      // maybe others?
+      default:
+        break;
+      }
+    }
+  }
+
+  if (CS.getNumArgOperands() == 2) {
+    ConstantFP *Op0C = dyn_cast<ConstantFP>(CS.getArgOperand(0));
+    ConstantFP *Op1C = dyn_cast<ConstantFP>(CS.getArgOperand(1));
+    if (Op0C && Op1C) {
+      const APFloat &Op0 = Op0C->getValueAPF();
+      const APFloat &Op1 = Op1C->getValueAPF();
+
+      switch (Func) {
+      case LibFunc::powl:
+      case LibFunc::pow:
+      case LibFunc::powf: {
+        // FIXME: Stop using the host math library.
+        // FIXME: The computation isn't done in the right precision.
+        Type *Ty = Op0C->getType();
+        if (Ty->isDoubleTy() || Ty->isFloatTy() || Ty->isHalfTy()) {
+          if (Ty == Op1C->getType()) {
+            double Op0V = getValueAsDouble(Op0C);
+            double Op1V = getValueAsDouble(Op1C);
+            return ConstantFoldBinaryFP(pow, Op0V, Op1V, Ty) != nullptr;
+          }
+        }
+        break;
+      }
+
+      case LibFunc::fmodl:
+      case LibFunc::fmod:
+      case LibFunc::fmodf:
+        return Op0.isNaN() || Op1.isNaN() ||
+               (!Op0.isInfinity() && !Op1.isZero());
+
+      default:
+        break;
+      }
+    }
+  }
+
+  return false;
+}

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

bool llvm::isInstructionTriviallyDead(Instruction *I,
   }
 
   if (isAllocLikeFn(I, TLI)) return true;
 
   if (CallInst *CI = isFreeCall(I, TLI))
     if (Constant *C = dyn_cast<Constant>(CI->getArgOperand(0)))
       return C->isNullValue() || isa<UndefValue>(C);
 
+  if (CallSite CS = CallSite(I))
+    if (isMathLibCallNoop(CS, TLI))
+      return true;
+
   return false;
 }
 
 /// RecursivelyDeleteTriviallyDeadInstructions - If the specified value is a
 /// trivially dead instruction, delete it.  If that makes any of its operands
 /// trivially dead, delete them too, recursively.  Return true if any
 /// instructions were deleted.
 bool

llvm/trunk/test/Transforms/DCE/calls-errno.ll

+; RUN: opt < %s -dce -S | FileCheck %s
+
+declare double @acos(double) nounwind
+declare double @asin(double) nounwind
+declare double @atan(double) nounwind
+declare double @atan2(double, double) nounwind
+declare double @ceil(double) nounwind
+declare double @cos(double) nounwind
+declare double @cosh(double) nounwind
+declare double @exp(double) nounwind
+declare double @exp2(double) nounwind
+declare double @fabs(double) nounwind
+declare double @floor(double) nounwind
+declare double @fmod(double, double) nounwind
+declare double @log(double) nounwind
+declare double @log10(double) nounwind
+declare double @pow(double, double) nounwind
+declare double @sin(double) nounwind
+declare double @sinh(double) nounwind
+declare double @sqrt(double) nounwind
+declare double @tan(double) nounwind
+declare double @tanh(double) nounwind
+
+declare float @acosf(float) nounwind
+declare float @asinf(float) nounwind
+declare float @atanf(float) nounwind
+declare float @atan2f(float, float) nounwind
+declare float @ceilf(float) nounwind
+declare float @cosf(float) nounwind
+declare float @coshf(float) nounwind
+declare float @expf(float) nounwind
+declare float @exp2f(float) nounwind
+declare float @fabsf(float) nounwind
+declare float @floorf(float) nounwind
+declare float @fmodf(float, float) nounwind
+declare float @logf(float) nounwind
+declare float @log10f(float) nounwind
+declare float @powf(float, float) nounwind
+declare float @sinf(float) nounwind
+declare float @sinhf(float) nounwind
+declare float @sqrtf(float) nounwind
+declare float @tanf(float) nounwind
+declare float @tanhf(float) nounwind
+
+define void @T() {
+entry:
+; CHECK-LABEL: @T(
+; CHECK-NEXT: entry:
+
+; log(0) produces a pole error
+; CHECK-NEXT: %log1 = call double @log(double 0.000000e+00)
+  %log1 = call double @log(double 0.000000e+00)
+
+; log(-1) produces a domain error
+; CHECK-NEXT: %log2 = call double @log(double -1.000000e+00)
+  %log2 = call double @log(double -1.000000e+00)
+
+; log(1) is 0
+  %log3 = call double @log(double 1.000000e+00)
+
+; exp(100) is roughly 2.6e+43
+  %exp1 = call double @exp(double 1.000000e+02)
+
+; exp(1000) is a range error
+; CHECK-NEXT: %exp2 = call double @exp(double 1.000000e+03)
+  %exp2 = call double @exp(double 1.000000e+03)
+
+; cos(0) is 1
+  %cos1 = call double @cos(double 0.000000e+00)
+
+; cos(inf) is a domain error
+; CHECK-NEXT: %cos2 = call double @cos(double 0x7FF0000000000000)
+  %cos2 = call double @cos(double 0x7FF0000000000000)
+
+; pow(0, 1) is 0
+  %pow1 = call double @pow(double 0x7FF0000000000000, double 1.000000e+00)
+
+; pow(0, -1) is a pole error
+; CHECK-NEXT: %pow2 = call double @pow(double 0.000000e+00, double -1.000000e+00)
+  %pow2 = call double @pow(double 0.000000e+00, double -1.000000e+00)
+
+; fmod(inf, nan) is nan
+  %fmod1 = call double @fmod(double 0x7FF0000000000000, double 0x7FF0000000000001)
+
+; fmod(inf, 1) is a domain error
+; CHECK-NEXT: %fmod2 = call double @fmod(double 0x7FF0000000000000, double 1.000000e+00)
+  %fmod2 = call double @fmod(double 0x7FF0000000000000, double 1.000000e+00)
+
+; CHECK-NEXT: ret void
+  ret void
+}