Implement support for option 'fexcess-precision'.

Differential revision: https://reviews.llvm.org/D136176

clang/docs/UsersManual.rst

@@ -1746,6 +1746,47 @@ floating point semantic models: precise (the default), strict, and fast.
    has no effect because the optimizer is prohibited from making unsafe
    transformations.
 
+.. option:: -fexcess-precision:
+
+   The C and C++ standards allow floating-point expressions to be computed as if
+   intermediate results had more precision (and/or a wider range) than the type
+   of the expression strictly allows.  This is called excess precision
+   arithmetic.
+   Excess precision arithmetic can improve the accuracy of results (although not
+   always), and it can make computation significantly faster if the target lacks
+   direct hardware support for arithmetic in a particular type.  However, it can
+   also undermine strict floating-point reproducibility.
+
+   Under the standards, assignments and explicit casts force the operand to be
+   converted to its formal type, discarding any excess precision.  Because data
+   can only flow between statements via an assignment, this means that the use
+   of excess precision arithmetic is a reliable local property of a single
+   statement, and results do not change based on optimization.  However, when
+   excess precision arithmetic is in use, Clang does not guarantee strict
+   reproducibility, and future compiler releases may recognize more
+   opportunities to use excess precision arithmetic, e.g. with floating-point
+   builtins.
+
+   Clang does not use excess precision arithmetic for most types or on most
+   targets. For example, even on pre-SSE X86 targets where ``float`` and
+   ``double`` computations must be performed in the 80-bit X87 format, Clang
+   rounds all intermediate results correctly for their type.  Clang currently
+   uses excess precision arithmetic by default only for the following types and
+   targets:
+
+   * ``_Float16`` on X86 targets without ``AVX512-FP16``.
+
+   The ``-fexcess-precision=<value>`` option can be used to control the use of
+   excess precision arithmetic.  Valid values are:
+
+   * ``standard`` - The default.  Allow the use of excess precision arithmetic
+     under the constraints of the C and C++ standards. Has no effect except on
+     the types and targets listed above.
+   * ``fast`` - Accepted for GCC compatibility, but currently treated as an
+     alias for ``standard``.
+   * ``16`` - Forces ``_Float16`` operations to be emitted without using excess
+     precision arithmetic.
+
 .. _crtfastmath.o:
 
 A note about ``crtfastmath.o``

clang/include/clang/AST/Type.h

@@ -763,6 +763,8 @@ class QualType {
   unsigned getLocalFastQualifiers() const { return Value.getInt(); }
   void setLocalFastQualifiers(unsigned Quals) { Value.setInt(Quals); }
 
+  bool UseExcessPrecision(const ASTContext &Ctx);
+
   /// Retrieves a pointer to the underlying (unqualified) type.
   ///
   /// This function requires that the type not be NULL. If the type might be

clang/include/clang/Basic/FPOptions.def

@@ -25,4 +25,5 @@ OPTION(NoSignedZero, bool, 1, NoHonorInfs)
 OPTION(AllowReciprocal, bool, 1, NoSignedZero)
 OPTION(AllowApproxFunc, bool, 1, AllowReciprocal)
 OPTION(FPEvalMethod, LangOptions::FPEvalMethodKind, 2, AllowApproxFunc)
+OPTION(Float16ExcessPrecision, LangOptions::ExcessPrecisionKind, 2, FPEvalMethod)
 #undef OPTION

clang/include/clang/Basic/LangOptions.def

@@ -317,6 +317,7 @@ COMPATIBLE_LANGOPT(ExpStrictFP, 1, false, "Enable experimental strict floating p
 BENIGN_LANGOPT(RoundingMath, 1, false, "Do not assume default floating-point rounding behavior")
 BENIGN_ENUM_LANGOPT(FPExceptionMode, FPExceptionModeKind, 2, FPE_Default, "FP Exception Behavior Mode type")
 BENIGN_ENUM_LANGOPT(FPEvalMethod, FPEvalMethodKind, 2, FEM_UnsetOnCommandLine, "FP type used for floating point arithmetic")
+ENUM_LANGOPT(Float16ExcessPrecision, ExcessPrecisionKind, 2, FPP_Standard, "Intermediate truncation behavior for floating point arithmetic")
 LANGOPT(NoBitFieldTypeAlign , 1, 0, "bit-field type alignment")
 LANGOPT(HexagonQdsp6Compat , 1, 0, "hexagon-qdsp6 backward compatibility")
 LANGOPT(ObjCAutoRefCount , 1, 0, "Objective-C automated reference counting")

clang/include/clang/Basic/LangOptions.h

@@ -295,6 +295,8 @@ class LangOptions : public LangOptionsBase {
     FEM_UnsetOnCommandLine = 3
   };
 
+  enum ExcessPrecisionKind { FPP_Standard, FPP_Fast, FPP_None };
+
   /// Possible exception handling behavior.
   enum class ExceptionHandlingKind { None, SjLj, WinEH, DwarfCFI, Wasm };
 

clang/include/clang/Basic/TargetInfo.h

@@ -933,8 +933,6 @@ class TargetInfo : public virtual TransferrableTargetInfo,
     return true;
   }
 
-  virtual bool shouldEmitFloat16WithExcessPrecision() const { return false; }
-
   /// Specify if mangling based on address space map should be used or
   /// not for language specific address spaces
   bool useAddressSpaceMapMangling() const {

clang/include/clang/Driver/Options.td

@@ -1576,8 +1576,22 @@ def exception_model_EQ : Joined<["-"], "exception-model=">,
 def fignore_exceptions : Flag<["-"], "fignore-exceptions">, Group<f_Group>, Flags<[CC1Option]>,
   HelpText<"Enable support for ignoring exception handling constructs">,
   MarshallingInfoFlag<LangOpts<"IgnoreExceptions">>;
-def fexcess_precision_EQ : Joined<["-"], "fexcess-precision=">,
-    Group<clang_ignored_gcc_optimization_f_Group>;
+def fexcess_precision_EQ : Joined<["-"], "fexcess-precision=">, Group<f_Group>,
+  HelpText<"Allows control over excess precision on targets where native "
+  "support for the precision types is not available. By default, excess "
+  "precision is used to calculate intermediate results following the "
+  "rules specified in ISO C99.">,
+  Values<"standard,fast,none">, NormalizedValuesScope<"LangOptions">,
+  NormalizedValues<["FPP_Standard", "FPP_Fast", "FPP_None"]>;
+def ffloat16_excess_precision_EQ : Joined<["-"], "ffloat16-excess-precision=">,
+  Group<f_Group>, Flags<[CC1Option, NoDriverOption]>,
+  HelpText<"Allows control over excess precision on targets where native "
+  "support for Float16 precision types is not available. By default, excess "
+  "precision is used to calculate intermediate results following the "
+  "rules specified in ISO C99.">,
+  Values<"standard,fast,none">, NormalizedValuesScope<"LangOptions">,
+  NormalizedValues<["FPP_Standard", "FPP_Fast", "FPP_None"]>,
+  MarshallingInfoEnum<LangOpts<"Float16ExcessPrecision">, "FPP_Standard">;
 def : Flag<["-"], "fexpensive-optimizations">, Group<clang_ignored_gcc_optimization_f_Group>;
 def : Flag<["-"], "fno-expensive-optimizations">, Group<clang_ignored_gcc_optimization_f_Group>;
 def fextdirs_EQ : Joined<["-"], "fextdirs=">, Group<f_Group>;

clang/lib/AST/Type.cpp

@@ -1483,6 +1483,25 @@ struct StripObjCKindOfTypeVisitor
 
 } // namespace
 
+bool QualType::UseExcessPrecision(const ASTContext &Ctx) {
+  const BuiltinType *BT = getTypePtr()->getAs<BuiltinType>();
+  if (BT) {
+    switch (BT->getKind()) {
+    case BuiltinType::Kind::Float16: {
+      const TargetInfo &TI = Ctx.getTargetInfo();
+      if (TI.hasFloat16Type() && !TI.hasLegalHalfType() &&
+          Ctx.getLangOpts().getFloat16ExcessPrecision() !=
+              Ctx.getLangOpts().ExcessPrecisionKind::FPP_None)
+        return true;
+      return false;
+    }
+    default:
+      return false;
+    }
+  }
+  return false;
+}
+
 /// Substitute the given type arguments for Objective-C type
 /// parameters within the given type, recursively.
 QualType QualType::substObjCTypeArgs(ASTContext &ctx,

clang/lib/Basic/Targets/X86.h

@@ -302,10 +302,6 @@ class LLVM_LIBRARY_VISIBILITY X86TargetInfo : public TargetInfo {
     return false;
   }
 
-  bool shouldEmitFloat16WithExcessPrecision() const override {
-    return HasFloat16 && !hasLegalHalfType();
-  }
-
   void getTargetDefines(const LangOptions &Opts,
                         MacroBuilder &Builder) const override;
 

clang/lib/CodeGen/CGExprComplex.cpp

@@ -275,18 +275,13 @@ class ComplexExprEmitter
                                         const BinOpInfo &Op);
 
   QualType getPromotionType(QualType Ty) {
-    if (CGF.getTarget().shouldEmitFloat16WithExcessPrecision()) {
-      if (Ty->isRealFloatingType()) {
-        if (Ty->isFloat16Type())
-          return CGF.getContext().FloatTy;
-      } else {
-        assert(Ty->isAnyComplexType() &&
-               "Expecting to promote a complex type!");
-        QualType ElementType = Ty->castAs<ComplexType>()->getElementType();
-        if (ElementType->isFloat16Type())
-          return CGF.getContext().getComplexType(CGF.getContext().FloatTy);
-      }
+    if (auto *CT = Ty->getAs<ComplexType>()) {
+      QualType ElementType = CT->getElementType();
+      if (ElementType.UseExcessPrecision(CGF.getContext()))
+        return CGF.getContext().getComplexType(CGF.getContext().FloatTy);
     }
+    if (Ty.UseExcessPrecision(CGF.getContext()))
+      return CGF.getContext().FloatTy;
     return QualType();
   }
 

clang/lib/CodeGen/CGExprScalar.cpp

@@ -814,15 +814,13 @@ class ScalarExprEmitter
                             Value *(ScalarExprEmitter::*F)(const BinOpInfo &));
 
   QualType getPromotionType(QualType Ty) {
-    if (CGF.getTarget().shouldEmitFloat16WithExcessPrecision()) {
-      if (Ty->isAnyComplexType()) {
-        QualType ElementType = Ty->castAs<ComplexType>()->getElementType();
-        if (ElementType->isFloat16Type())
-          return CGF.getContext().getComplexType(CGF.getContext().FloatTy);
-      }
-      if (Ty->isFloat16Type())
-        return CGF.getContext().FloatTy;
+    if (auto *CT = Ty->getAs<ComplexType>()) {
+      QualType ElementType = CT->getElementType();
+      if (ElementType.UseExcessPrecision(CGF.getContext()))
+        return CGF.getContext().getComplexType(CGF.getContext().FloatTy);
     }
+    if (Ty.UseExcessPrecision(CGF.getContext()))
+      return CGF.getContext().FloatTy;
     return QualType();
   }
 

clang/lib/Driver/ToolChains/Clang.cpp

@@ -2705,6 +2705,7 @@ static void RenderFloatingPointOptions(const ToolChain &TC, const Driver &D,
       !JA.isOffloading(Action::OFK_HIP))
     FPContract = "on";
   bool StrictFPModel = false;
+  StringRef Float16ExcessPrecision = "";
 
   if (const Arg *A = Args.getLastArg(options::OPT_flimited_precision_EQ)) {
     CmdArgs.push_back("-mlimit-float-precision");
@@ -2901,6 +2902,27 @@ static void RenderFloatingPointOptions(const ToolChain &TC, const Driver &D,
       break;
     }
 
+    case options::OPT_fexcess_precision_EQ: {
+      StringRef Val = A->getValue();
+      const llvm::Triple::ArchType Arch = TC.getArch();
+      if (Arch == llvm::Triple::x86 || Arch == llvm::Triple::x86_64) {
+        if (Val.equals("standard") || Val.equals("fast"))
+          Float16ExcessPrecision = Val;
+        // To make it GCC compatible, allow the value of "16" which
+        // means disable excess precision, the same meaning than clang's
+        // equivalent value "none".
+        else if (Val.equals("16"))
+          Float16ExcessPrecision = "none";
+        else
+          D.Diag(diag::err_drv_unsupported_option_argument)
+              << A->getSpelling() << Val;
+      } else {
+        if (!(Val.equals("standard") || Val.equals("fast")))
+          D.Diag(diag::err_drv_unsupported_option_argument)
+              << A->getSpelling() << Val;
+      }
+      break;
+    }
     case options::OPT_ffinite_math_only:
       HonorINFs = false;
       HonorNaNs = false;
@@ -3071,6 +3093,10 @@ static void RenderFloatingPointOptions(const ToolChain &TC, const Driver &D,
   if (!FPEvalMethod.empty())
     CmdArgs.push_back(Args.MakeArgString("-ffp-eval-method=" + FPEvalMethod));
 
+  if (!Float16ExcessPrecision.empty())
+    CmdArgs.push_back(Args.MakeArgString("-ffloat16-excess-precision=" +
+                                         Float16ExcessPrecision));
+
   ParseMRecip(D, Args, CmdArgs);
 
   // -ffast-math enables the __FAST_MATH__ preprocessor macro, but check for the