[ConstantFolding] Fix incorrect nvvm_round folding

[LewisCrawford]

The nvvm_round intrinsic should round to the nearest even number in the case of ties. It lowers to PTX cvt.rni, which will "round to nearest integer, choosing even integer if source is equidistant between two integers", so it matches the semantics of rint (and not round as the name suggests).

[llvmbot]

@llvm/pr-subscribers-llvm-transforms
@llvm/pr-subscribers-llvm-analysis

@llvm/pr-subscribers-backend-nvptx

Author: Lewis Crawford (LewisCrawford)

Changes

The nvvm_round intrinsic should round to the nearest even number in the case of ties. It matches the semantics of rint (not round as the name suggests).

---

Full diff: https://github.com/llvm/llvm-project/pull/151563.diff

2 Files Affected:

• (modified) llvm/lib/Analysis/ConstantFolding.cpp (+1-3)
• (modified) llvm/test/Transforms/InstSimplify/const-fold-nvvm-unary-arithmetic.ll (+48)

diff --git a/llvm/lib/Analysis/ConstantFolding.cpp b/llvm/lib/Analysis/ConstantFolding.cpp
index 2d52f3440938a..5e8d24b056245 100644
--- a/llvm/lib/Analysis/ConstantFolding.cpp
+++ b/llvm/lib/Analysis/ConstantFolding.cpp
@@ -2679,11 +2679,9 @@ static Constant *ConstantFoldScalarCall1(StringRef Name,
       case Intrinsic::nvvm_round_ftz_f:
       case Intrinsic::nvvm_round_f:
       case Intrinsic::nvvm_round_d: {
-        // Use APFloat implementation instead of native libm call, as some
-        // implementations (e.g. on PPC) do not preserve the sign of negative 0.
         bool IsFTZ = nvvm::UnaryMathIntrinsicShouldFTZ(IntrinsicID);
         auto V = IsFTZ ? FTZPreserveSign(APF) : APF;
-        V.roundToIntegral(APFloat::rmNearestTiesToAway);
+        V.roundToIntegral(APFloat::rmNearestTiesToEven);
         return ConstantFP::get(Ty->getContext(), V);
       }
 
diff --git a/llvm/test/Transforms/InstSimplify/const-fold-nvvm-unary-arithmetic.ll b/llvm/test/Transforms/InstSimplify/const-fold-nvvm-unary-arithmetic.ll
index 75b850978b75a..6eed7f8f1a2ae 100644
--- a/llvm/test/Transforms/InstSimplify/const-fold-nvvm-unary-arithmetic.ll
+++ b/llvm/test/Transforms/InstSimplify/const-fold-nvvm-unary-arithmetic.ll
@@ -416,6 +416,54 @@ define float @test_round_ftz_f_neg_1_5() {
   ret float %res
 }
 
+define double @test_round_d_2_5() {
+; CHECK-LABEL: define double @test_round_d_2_5() {
+; CHECK-NEXT:    ret double 2.000000e+00
+;
+  %res = call double @llvm.nvvm.round.d(double 2.5)
+  ret double %res
+}
+
+define float @test_round_f_2_5() {
+; CHECK-LABEL: define float @test_round_f_2_5() {
+; CHECK-NEXT:    ret float 2.000000e+00
+;
+  %res = call float @llvm.nvvm.round.f(float 2.5)
+  ret float %res
+}
+
+define float @test_round_ftz_f_2_5() {
+; CHECK-LABEL: define float @test_round_ftz_f_2_5() {
+; CHECK-NEXT:    ret float 2.000000e+00
+;
+  %res = call float @llvm.nvvm.round.ftz.f(float 2.5)
+  ret float %res
+}
+
+define double @test_round_d_neg_2_5() {
+; CHECK-LABEL: define double @test_round_d_neg_2_5() {
+; CHECK-NEXT:    ret double -2.000000e+00
+;
+  %res = call double @llvm.nvvm.round.d(double -2.5)
+  ret double %res
+}
+
+define float @test_round_f_neg_2_5() {
+; CHECK-LABEL: define float @test_round_f_neg_2_5() {
+; CHECK-NEXT:    ret float -2.000000e+00
+;
+  %res = call float @llvm.nvvm.round.f(float -2.5)
+  ret float %res
+}
+
+define float @test_round_ftz_f_neg_2_5() {
+; CHECK-LABEL: define float @test_round_ftz_f_neg_2_5() {
+; CHECK-NEXT:    ret float -2.000000e+00
+;
+  %res = call float @llvm.nvvm.round.ftz.f(float -2.5)
+  ret float %res
+}
+
 define double @test_round_d_neg_subnorm() {
 ; CHECK-LABEL: define double @test_round_d_neg_subnorm() {
 ; CHECK-NEXT:    ret double -0.000000e+00