From d9c4c312d81b4a2059d0ca2bb454c3452e52042e Mon Sep 17 00:00:00 2001 From: Zahira Ammarguellat Date: Wed, 20 Mar 2024 12:14:09 -0700 Subject: [CLANG] Full support of complex multiplication and division. (#81514) In clang there are two options `-f[no]cx-limited-range` and `-f[no]cx-fortran-rules` that control the range of complex multiplication and division. However, it is unclear how these options interact with one another. For instance, what should happen when the users compile with `-fcx-fortran-rules -fno-cx-limited-range` or `-fcx-limited-range -fno-cx-fortran-rules`? In this patch we are introducing a new option to solve the issue and give a greater flexibility to the user to control the behavior of the compiler when performing multiplication and division of complex floating-point values. `-fcomplex-arihmetic=[full|improved|promoted|basic]` `full`: Implementation of complex division and multiplication using a call to runtime library functions (generally the case, but the BE might sometimes replace the library call if it knows enough about the potential range of the inputs). Overflow and non-finite values are handled by the library implementation. For the case of multiplication overflow will occur in accordance with normal floating-point rules. This is the default value. `improved`: Implementation of complex division using the Smith algorithm at source precision. Smith's algorithm for complex division. See SMITH, R. L. Algorithm 116: Complex division. Commun. ACM 5, 8 (1962). This value offers improved handling for overflow in intermediate calculations, but overflow may occur. NaN and infinite values are not handled in some cases. `promoted`: Implementation of complex division using algebraic formulas at higher precision. Overflow is handled. Non-finite values are handled in some cases. If the target does not have native support for a higher precision data type, the implementation for the complex operation using the Smith algorithm will be used. Overflow may still occur in some cases. NaN and infinite values are not handled. `basic`: Implementation of complex division and multiplication using algebraic formulas at source precision. No special handling to avoid overflow. NaN and infinite values are not handled. `fcx-limited-range` will alias `-fcomplex-arithmetic=basic` `-fcx-fortran-rules` will alias `-fcomplex-arithmetic=improved` `-fno-cx-limited-range` and `-fno-cx-fortran-rules` will alias `-fcomplex-arithmetic=full` The complex division and multiplication will be implemented as follows depending on the option used. -fcomplex-arithmetic | div | mul -- | -- | -- basic | algebraic form | algebraic form improved | smith's algorithm | algebraic form full | libcall | libcall + nan processing promoted | algebraic form + higher precision | algebraic form --- clang/lib/CodeGen/CGExprComplex.cpp | 83 ++++++++++++++++++++++++++++++------- 1 file changed, 67 insertions(+), 16 deletions(-) (limited to 'clang/lib/CodeGen/CGExprComplex.cpp') diff --git a/clang/lib/CodeGen/CGExprComplex.cpp b/clang/lib/CodeGen/CGExprComplex.cpp index 0266ba934da6..27ddaacc28f5 100644 --- a/clang/lib/CodeGen/CGExprComplex.cpp +++ b/clang/lib/CodeGen/CGExprComplex.cpp @@ -51,11 +51,12 @@ class ComplexExprEmitter CGBuilderTy &Builder; bool IgnoreReal; bool IgnoreImag; -public: - ComplexExprEmitter(CodeGenFunction &cgf, bool ir=false, bool ii=false) - : CGF(cgf), Builder(CGF.Builder), IgnoreReal(ir), IgnoreImag(ii) { - } + bool FPHasBeenPromoted; +public: + ComplexExprEmitter(CodeGenFunction &cgf, bool ir = false, bool ii = false) + : CGF(cgf), Builder(CGF.Builder), IgnoreReal(ir), IgnoreImag(ii), + FPHasBeenPromoted(false) {} //===--------------------------------------------------------------------===// // Utilities @@ -287,9 +288,54 @@ public: ComplexPairTy EmitComplexBinOpLibCall(StringRef LibCallName, const BinOpInfo &Op); - QualType getPromotionType(QualType Ty) { + QualType GetHigherPrecisionFPType(QualType ElementType) { + const auto *CurrentBT = dyn_cast(ElementType); + switch (CurrentBT->getKind()) { + case BuiltinType::Kind::Float16: + return CGF.getContext().FloatTy; + case BuiltinType::Kind::Float: + case BuiltinType::Kind::BFloat16: + return CGF.getContext().DoubleTy; + case BuiltinType::Kind::Double: + return CGF.getContext().LongDoubleTy; + default: + return ElementType; + } + } + + QualType HigherPrecisionTypeForComplexArithmetic(QualType ElementType, + bool IsDivOpCode) { + QualType HigherElementType = GetHigherPrecisionFPType(ElementType); + const llvm::fltSemantics &ElementTypeSemantics = + CGF.getContext().getFloatTypeSemantics(ElementType); + const llvm::fltSemantics &HigherElementTypeSemantics = + CGF.getContext().getFloatTypeSemantics(HigherElementType); + // Check that the promoted type can handle the intermediate values without + // overflowing. This can be interpreted as: + // (SmallerType.LargestFiniteVal * SmallerType.LargestFiniteVal) * 2 <= + // LargerType.LargestFiniteVal. + // In terms of exponent it gives this formula: + // (SmallerType.LargestFiniteVal * SmallerType.LargestFiniteVal + // doubles the exponent of SmallerType.LargestFiniteVal) + if (llvm::APFloat::semanticsMaxExponent(ElementTypeSemantics) * 2 + 1 <= + llvm::APFloat::semanticsMaxExponent(HigherElementTypeSemantics)) { + return CGF.getContext().getComplexType(HigherElementType); + } else { + FPHasBeenPromoted = true; + DiagnosticsEngine &Diags = CGF.CGM.getDiags(); + Diags.Report(diag::warn_next_larger_fp_type_same_size_than_fp); + return CGF.getContext().getComplexType(ElementType); + } + } + + QualType getPromotionType(QualType Ty, bool IsDivOpCode = false) { if (auto *CT = Ty->getAs()) { QualType ElementType = CT->getElementType(); + if (IsDivOpCode && ElementType->isFloatingType() && + CGF.getLangOpts().getComplexRange() == + LangOptions::ComplexRangeKind::CX_Promoted) + return HigherPrecisionTypeForComplexArithmetic(ElementType, + IsDivOpCode); if (ElementType.UseExcessPrecision(CGF.getContext())) return CGF.getContext().getComplexType(CGF.getContext().FloatTy); } @@ -300,11 +346,12 @@ public: #define HANDLEBINOP(OP) \ ComplexPairTy VisitBin##OP(const BinaryOperator *E) { \ - QualType promotionTy = getPromotionType(E->getType()); \ + QualType promotionTy = getPromotionType( \ + E->getType(), \ + (E->getOpcode() == BinaryOperatorKind::BO_Div) ? true : false); \ ComplexPairTy result = EmitBin##OP(EmitBinOps(E, promotionTy)); \ if (!promotionTy.isNull()) \ - result = \ - CGF.EmitUnPromotedValue(result, E->getType()); \ + result = CGF.EmitUnPromotedValue(result, E->getType()); \ return result; \ } @@ -794,8 +841,9 @@ ComplexPairTy ComplexExprEmitter::EmitBinMul(const BinOpInfo &Op) { ResR = Builder.CreateFSub(AC, BD, "mul_r"); ResI = Builder.CreateFAdd(AD, BC, "mul_i"); - if (Op.FPFeatures.getComplexRange() == LangOptions::CX_Limited || - Op.FPFeatures.getComplexRange() == LangOptions::CX_Fortran) + if (Op.FPFeatures.getComplexRange() == LangOptions::CX_Basic || + Op.FPFeatures.getComplexRange() == LangOptions::CX_Improved || + Op.FPFeatures.getComplexRange() == LangOptions::CX_Promoted) return ComplexPairTy(ResR, ResI); // Emit the test for the real part becoming NaN and create a branch to @@ -986,14 +1034,17 @@ ComplexPairTy ComplexExprEmitter::EmitBinDiv(const BinOpInfo &Op) { llvm::Value *OrigLHSi = LHSi; if (!LHSi) LHSi = llvm::Constant::getNullValue(RHSi->getType()); - if (Op.FPFeatures.getComplexRange() == LangOptions::CX_Fortran) + QualType ComplexElementTy = Op.Ty->castAs()->getElementType(); + if (Op.FPFeatures.getComplexRange() == LangOptions::CX_Improved || + (Op.FPFeatures.getComplexRange() == LangOptions::CX_Promoted && + FPHasBeenPromoted)) return EmitRangeReductionDiv(LHSr, LHSi, RHSr, RHSi); - else if (Op.FPFeatures.getComplexRange() == LangOptions::CX_Limited) + else if (Op.FPFeatures.getComplexRange() == LangOptions::CX_Basic || + Op.FPFeatures.getComplexRange() == LangOptions::CX_Promoted) return EmitAlgebraicDiv(LHSr, LHSi, RHSr, RHSi); - else if (!CGF.getLangOpts().FastMath || - // '-ffast-math' is used in the command line but followed by an - // '-fno-cx-limited-range'. - Op.FPFeatures.getComplexRange() == LangOptions::CX_Full) { + // '-ffast-math' is used in the command line but followed by an + // '-fno-cx-limited-range' or '-fcomplex-arithmetic=full'. + else if (Op.FPFeatures.getComplexRange() == LangOptions::CX_Full) { LHSi = OrigLHSi; // If we have a complex operand on the RHS and FastMath is not allowed, we // delegate to a libcall to handle all of the complexities and minimize -- cgit v1.2.3