diff options
Diffstat (limited to 'llvm/lib/Analysis/ConstantFolding.cpp')
| -rw-r--r-- | llvm/lib/Analysis/ConstantFolding.cpp | 20 |
1 files changed, 12 insertions, 8 deletions
diff --git a/llvm/lib/Analysis/ConstantFolding.cpp b/llvm/lib/Analysis/ConstantFolding.cpp index 759c553..dd98b62 100644 --- a/llvm/lib/Analysis/ConstantFolding.cpp +++ b/llvm/lib/Analysis/ConstantFolding.cpp @@ -1373,7 +1373,7 @@ Constant *llvm::FlushFPConstant(Constant *Operand, const Instruction *Inst, if (ConstantFP *CFP = dyn_cast<ConstantFP>(Operand)) return flushDenormalConstantFP(CFP, Inst, IsOutput); - if (isa<ConstantAggregateZero, UndefValue, ConstantExpr>(Operand)) + if (isa<ConstantAggregateZero, UndefValue>(Operand)) return Operand; Type *Ty = Operand->getType(); @@ -1389,6 +1389,9 @@ Constant *llvm::FlushFPConstant(Constant *Operand, const Instruction *Inst, Ty = VecTy->getElementType(); } + if (isa<ConstantExpr>(Operand)) + return Operand; + if (const auto *CV = dyn_cast<ConstantVector>(Operand)) { SmallVector<Constant *, 16> NewElts; for (unsigned i = 0, e = CV->getNumOperands(); i != e; ++i) { @@ -2628,14 +2631,14 @@ static Constant *ConstantFoldScalarCall1(StringRef Name, case Intrinsic::nvvm_ceil_d: return ConstantFoldFP( ceil, APF, Ty, - nvvm::GetNVVMDenromMode( + nvvm::GetNVVMDenormMode( nvvm::UnaryMathIntrinsicShouldFTZ(IntrinsicID))); case Intrinsic::nvvm_fabs_ftz: case Intrinsic::nvvm_fabs: return ConstantFoldFP( fabs, APF, Ty, - nvvm::GetNVVMDenromMode( + nvvm::GetNVVMDenormMode( nvvm::UnaryMathIntrinsicShouldFTZ(IntrinsicID))); case Intrinsic::nvvm_floor_ftz_f: @@ -2643,7 +2646,7 @@ static Constant *ConstantFoldScalarCall1(StringRef Name, case Intrinsic::nvvm_floor_d: return ConstantFoldFP( floor, APF, Ty, - nvvm::GetNVVMDenromMode( + nvvm::GetNVVMDenormMode( nvvm::UnaryMathIntrinsicShouldFTZ(IntrinsicID))); case Intrinsic::nvvm_rcp_rm_ftz_f: @@ -2676,11 +2679,12 @@ 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. + // nvvm_round is lowered to PTX cvt.rni, which will round to nearest + // integer, choosing even integer if source is equidistant between two + // integers, so the semantics are closer to "rint" rather than "round". 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); } @@ -2705,7 +2709,7 @@ static Constant *ConstantFoldScalarCall1(StringRef Name, return nullptr; return ConstantFoldFP( sqrt, APF, Ty, - nvvm::GetNVVMDenromMode( + nvvm::GetNVVMDenormMode( nvvm::UnaryMathIntrinsicShouldFTZ(IntrinsicID))); // AMDGCN Intrinsics: |