Allow constant folding of round function whenever feasible

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203198 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Analysis/ConstantFolding.cpp b/lib/Analysis/ConstantFolding.cpp
index ff1124ad87887578683fb82cba632a8d42ee8199..8ff79d6f6255702a2fa975047464c4e9ae7cb584 100644 (file)
--- a/lib/Analysis/ConstantFolding.cpp
+++ b/lib/Analysis/ConstantFolding.cpp
@@ -1196,6 +1196,7 @@ bool llvm::canConstantFoldCallTo(const Function *F) {
   case Intrinsic::fma:
   case Intrinsic::fmuladd:
   case Intrinsic::copysign:
+  case Intrinsic::round:
   case Intrinsic::sadd_with_overflow:
   case Intrinsic::uadd_with_overflow:
   case Intrinsic::ssub_with_overflow:
@@ -1345,6 +1346,12 @@ static Constant *ConstantFoldScalarCall(StringRef Name, unsigned IntrinsicID,
       if (!Ty->isHalfTy() && !Ty->isFloatTy() && !Ty->isDoubleTy())
         return 0;
 
+      if (IntrinsicID == Intrinsic::round) {
+        APFloat V = Op->getValueAPF();
+        V.roundToIntegral(APFloat::rmNearestTiesToAway);
+        return ConstantFP::get(Ty->getContext(), V);
+      }
+
       /// We only fold functions with finite arguments. Folding NaN and inf is
       /// likely to be aborted with an exception anyway, and some host libms
       /// have known errors raising exceptions.

diff --git a/test/Transforms/InstCombine/round.ll b/test/Transforms/InstCombine/round.ll
new file mode 100644 (file)
index 0000000..ecc62dd
--- /dev/null
+++ b/test/Transforms/InstCombine/round.ll
@@ -0,0 +1,90 @@
+; RUN: opt -S -instcombine < %s | FileCheck %s
+
+declare float @llvm.round.f32(float) #0
+declare double @llvm.round.f64(double) #0
+
+; CHECK-LABEL: @constant_fold_round_f32_01
+; CHECK-NEXT: ret float 1.000000e+00
+define float @constant_fold_round_f32_01() #0 {
+  %x = call float @llvm.round.f32(float 1.25) #0
+  ret float %x
+}
+
+; CHECK-LABEL: @constant_fold_round_f32_02
+; CHECK-NEXT: ret float -1.000000e+00
+define float @constant_fold_round_f32_02() #0 {
+  %x = call float @llvm.round.f32(float -1.25) #0
+  ret float %x
+}
+
+; CHECK-LABEL: @constant_fold_round_f32_03
+; CHECK-NEXT: ret float 2.000000e+00
+define float @constant_fold_round_f32_03() #0 {
+  %x = call float @llvm.round.f32(float 1.5) #0
+  ret float %x
+}
+
+; CHECK-LABEL: @constant_fold_round_f32_04
+; CHECK-NEXT: ret float -2.000000e+00
+define float @constant_fold_round_f32_04() #0 {
+  %x = call float @llvm.round.f32(float -1.5) #0
+  ret float %x
+}
+
+; CHECK-LABEL: @constant_fold_round_f32_05
+; CHECK-NEXT: ret float 3.000000e+00
+define float @constant_fold_round_f32_05() #0 {
+  %x = call float @llvm.round.f32(float 2.75) #0
+  ret float %x
+}
+
+; CHECK-LABEL: @constant_fold_round_f32_06
+; CHECK-NEXT: ret float -3.000000e+00
+define float @constant_fold_round_f32_06() #0 {
+  %x = call float @llvm.round.f32(float -2.75) #0
+  ret float %x
+}
+
+; CHECK-LABEL: @constant_fold_round_f64_01
+; CHECK-NEXT: ret double 1.000000e+00
+define double @constant_fold_round_f64_01() #0 {
+  %x = call double @llvm.round.f64(double 1.3) #0
+  ret double %x
+}
+
+; CHECK-LABEL: @constant_fold_round_f64_02
+; CHECK-NEXT: ret double -1.000000e+00
+define double @constant_fold_round_f64_02() #0 {
+  %x = call double @llvm.round.f64(double -1.3) #0
+  ret double %x
+}
+
+; CHECK-LABEL: @constant_fold_round_f64_03
+; CHECK-NEXT: ret double 2.000000e+00
+define double @constant_fold_round_f64_03() #0 {
+  %x = call double @llvm.round.f64(double 1.5) #0
+  ret double %x
+}
+
+; CHECK-LABEL: @constant_fold_round_f64_04
+; CHECK-NEXT: ret double -2.000000e+00
+define double @constant_fold_round_f64_04() #0 {
+  %x = call double @llvm.round.f64(double -1.5) #0
+  ret double %x
+}
+
+; CHECK-LABEL: @constant_fold_round_f64_05
+; CHECK-NEXT: ret double 3.000000e+00
+define double @constant_fold_round_f64_05() #0 {
+  %x = call double @llvm.round.f64(double 2.7) #0
+  ret double %x
+}
+
+; CHECK-LABEL: @constant_fold_round_f64_06
+; CHECK-NEXT: ret double -3.000000e+00
+define double @constant_fold_round_f64_06() #0 {
+  %x = call double @llvm.round.f64(double -2.7) #0
+  ret double %x
+}
+
+attributes #0 = { nounwind readnone }