Handle sqrt() shrinking in SimplifyLibCalls like any other call

This patch removes a chunk of special case logic for folding
(float)sqrt((double)x) -> sqrtf(x)
in InstCombineCasts and handles it in the mainstream path of SimplifyLibCalls.

No functional change intended, but I loosened the restriction on the existing
sqrt testcases to allow for this optimization even without unsafe-fp-math because
that's the existing behavior.

I also added a missing test case for not shrinking the llvm.sqrt.f64 intrinsic
in case the result is used as a double.

Differential Revision: http://reviews.llvm.org/D5919

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@220514 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Transforms/InstCombine/InstCombineCasts.cpp b/lib/Transforms/InstCombine/InstCombineCasts.cpp
index c16992ff62600f559ab17bd5303a0b3234aef6a6..cbcc85944a75de4e425d6cf3754ce17cce518728 100644 (file)
--- a/lib/Transforms/InstCombine/InstCombineCasts.cpp
+++ b/lib/Transforms/InstCombine/InstCombineCasts.cpp
@@ -1317,42 +1317,6 @@ Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) {
     }
   }
 
-  // Fold (fptrunc (sqrt (fpext x))) -> (sqrtf x)
-  // Note that we restrict this transformation based on
-  // TLI->has(LibFunc::sqrtf), even for the sqrt intrinsic, because
-  // TLI->has(LibFunc::sqrtf) is sufficient to guarantee that the
-  // single-precision intrinsic can be expanded in the backend.
-  CallInst *Call = dyn_cast<CallInst>(CI.getOperand(0));
-  if (Call && Call->getCalledFunction() && TLI->has(LibFunc::sqrtf) &&
-      (Call->getCalledFunction()->getName() == TLI->getName(LibFunc::sqrt) ||
-       Call->getCalledFunction()->getIntrinsicID() == Intrinsic::sqrt) &&
-      Call->getNumArgOperands() == 1 &&
-      Call->hasOneUse()) {
-    CastInst *Arg = dyn_cast<CastInst>(Call->getArgOperand(0));
-    if (Arg && Arg->getOpcode() == Instruction::FPExt &&
-        CI.getType()->isFloatTy() &&
-        Call->getType()->isDoubleTy() &&
-        Arg->getType()->isDoubleTy() &&
-        Arg->getOperand(0)->getType()->isFloatTy()) {
-      Function *Callee = Call->getCalledFunction();
-      Module *M = CI.getParent()->getParent()->getParent();
-      Constant *SqrtfFunc = (Callee->getIntrinsicID() == Intrinsic::sqrt) ?
-        Intrinsic::getDeclaration(M, Intrinsic::sqrt, Builder->getFloatTy()) :
-        M->getOrInsertFunction("sqrtf", Callee->getAttributes(),
-                               Builder->getFloatTy(), Builder->getFloatTy(),
-                               NULL);
-      CallInst *ret = CallInst::Create(SqrtfFunc, Arg->getOperand(0),
-                                       "sqrtfcall");
-      ret->setAttributes(Callee->getAttributes());
-
-
-      // Remove the old Call.  With -fmath-errno, it won't get marked readnone.
-      ReplaceInstUsesWith(*Call, UndefValue::get(Call->getType()));
-      EraseInstFromFunction(*Call);
-      return ret;
-    }
-  }
-
   return nullptr;
 }
 

diff --git a/lib/Transforms/Utils/SimplifyLibCalls.cpp b/lib/Transforms/Utils/SimplifyLibCalls.cpp
index caae06c4211e1bf349fec1147f09bee7e17445ac..207d3664789e6995196f2ee8302f9517cc0b7995 100644 (file)
--- a/lib/Transforms/Utils/SimplifyLibCalls.cpp
+++ b/lib/Transforms/Utils/SimplifyLibCalls.cpp
@@ -1058,7 +1058,16 @@ Value *LibCallSimplifier::optimizeUnaryDoubleFP(CallInst *CI, IRBuilder<> &B,
 
   // floor((double)floatval) -> (double)floorf(floatval)
   Value *V = Cast->getOperand(0);
-  V = EmitUnaryFloatFnCall(V, Callee->getName(), B, Callee->getAttributes());
+  if (Callee->isIntrinsic()) {
+    Module *M = CI->getParent()->getParent()->getParent();
+    Intrinsic::ID IID = (Intrinsic::ID) Callee->getIntrinsicID();
+    Function *F = Intrinsic::getDeclaration(M, IID, B.getFloatTy());
+    V = B.CreateCall(F, V);
+  } 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());
 }
 
@@ -1086,6 +1095,7 @@ Value *LibCallSimplifier::optimizeBinaryDoubleFP(CallInst *CI, IRBuilder<> &B) {
   Value *V = nullptr;
   Value *V1 = Cast1->getOperand(0);
   Value *V2 = Cast2->getOperand(0);
+  // TODO: Handle intrinsics in the same way as in optimizeUnaryDoubleFP().
   V = EmitBinaryFloatFnCall(V1, V2, Callee->getName(), B,
                             Callee->getAttributes());
   return B.CreateFPExt(V, B.getDoubleTy());
@@ -1267,10 +1277,9 @@ Value *LibCallSimplifier::optimizeSqrt(CallInst *CI, IRBuilder<> &B) {
   Function *Callee = CI->getCalledFunction();
   
   Value *Ret = nullptr;
-  if (UnsafeFPShrink && Callee->getName() == "sqrt" &&
-      TLI->has(LibFunc::sqrtf)) {
+  if (TLI->has(LibFunc::sqrtf) && (Callee->getName() == "sqrt" ||
+                                   Callee->getIntrinsicID() == Intrinsic::sqrt))
     Ret = optimizeUnaryDoubleFP(CI, B, true);
-  }
 
   // FIXME: For finer-grain optimization, we need intrinsics to have the same
   // fast-math flag decorations that are applied to FP instructions. For now,
@@ -2010,7 +2019,7 @@ Value *LibCallSimplifier::optimizeCall(CallInst *CI) {
       UnsafeFPShrink = true;
   }
 
-  // Next check for intrinsics.
+  // First, check for intrinsics.
   if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI)) {
     if (!isCallingConvC)
       return nullptr;

diff --git a/test/Transforms/InstCombine/double-float-shrink-1.ll b/test/Transforms/InstCombine/double-float-shrink-1.ll
index 9b422b7729644a3d4dfafe4f82986e938decda02..63a02bbd8572b045d2bfae57f87610ac854ef776 100644 (file)
--- a/test/Transforms/InstCombine/double-float-shrink-1.ll
+++ b/test/Transforms/InstCombine/double-float-shrink-1.ll
@@ -279,6 +279,14 @@ define float @sqrt_test(float %f) {
 ; CHECK: call float @sqrtf(float %f)
 }
 
+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 @sqrt_int_test(float %f) {
    %conv = fpext float %f to double
    %call = call double @llvm.sqrt.f64(double %conv)
@@ -288,13 +296,14 @@ define float @sqrt_int_test(float %f) {
 ; CHECK: call float @llvm.sqrt.f32(float %f)
 }
 
-define double @sqrt_test2(float %f) {
+define double @sqrt_int_test2(float %f) {
    %conv = fpext float %f to double
-   %call = call double @sqrt(double %conv)
+   %call = call double @llvm.sqrt.f64(double %conv)
    ret double %call
-; CHECK-LABEL: sqrt_test2
-; CHECK: call double @sqrt(double %conv)
+; CHECK-LABEL: sqrt_int_test2
+; CHECK: call double @llvm.sqrt.f64(double %conv)
 }
+
 define float @tan_test(float %f) {
    %conv = fpext float %f to double
    %call = call double @tan(double %conv)
@@ -330,7 +339,12 @@ define double @tanh_test2(float %f) {
 
 declare double @tanh(double) #1
 declare double @tan(double) #1
-declare double @sqrt(double) #1
+
+; sqrt is a special case: the shrinking optimization 
+; is valid even without unsafe-fp-math.
+declare double @sqrt(double) 
+declare double @llvm.sqrt.f64(double) 
+
 declare double @sin(double) #1
 declare double @log2(double) #1
 declare double @log1p(double) #1
@@ -348,6 +362,5 @@ declare double @acosh(double) #1
 declare double @asin(double) #1
 declare double @asinh(double) #1
 
-declare double @llvm.sqrt.f64(double) #1
 attributes #1 = { "unsafe-fp-math"="true" }