[Clang] [PowerPC] Emit module flag for current float abi

[qiucf]

Summary

This is part of the efforts adding .gnu_attribute support for PowerPC. In Clang, an extra metadata field will be added as float-abi to show current long double format. So backend can emit .gnu_attribute section data from this metadata.

To avoid breaking existing behavior, the module metadata will only be emitted when this module makes use of long double.

[MaskRay]

This is part of the efforts adding .gnu_attribute support for PowerPC. In Clang, an extra metadata field will be added as float-abi to show current long double format. So backend can emit .gnu_attribute section data from this metadata.

How does .gnu_attribute work with the metadata?

clang/test/CodeGen/ppc64-float-abi-attr.c

10

![[#]]

[qiucf]

In D116016#3202148, @MaskRay wrote:

This is part of the efforts adding .gnu_attribute support for PowerPC. In Clang, an extra metadata field will be added as float-abi to show current long double format. So backend can emit .gnu_attribute section data from this metadata.

How does .gnu_attribute work with the metadata?

Planned way to do is check the module flag in ASM printer, like

if (flt == "doubledouble")
  OutStreamer->emitGNUAttribute(4, 5);
else if (flt == "ieeequad")
  OutStreamer->emitGNUAttribute(4, 13);
else if (flt == "ieeedouble")
  OutStreamer->emitGNUAttribute(4, 9);

Currently this doesn't look like the most elegant way, and doesn't fully match behavior of GCC. (In GCC, if no floating operations exist, the attributes won't be generated) But since the floating abi option applies to the whole module, adding them to module flags are more reasonable than function attributes.

[nemanjai]

We should not be emitting the attribute in modules that do not have any use of long double.

[MaskRay]

In D116016#3203005, @nemanjai wrote:

We should not be emitting the attribute in modules that do not have any use of long double.

Right, otherwise it would be a bit unfortunate that a library not using long double has to be compiled differently to use ThinLTO with other programs, since the module flags metadata's behavior is Error (1).

[qiucf]

Only emit the flags when any instance of long double is used.

[qiucf]

Gentle ping?

[nemanjai]

LGTM. Thanks for reviving this patch.

clang/lib/CodeGen/CodeGenTypes.h

class CodeGenTypes {
   llvm::FoldingSet<CGFunctionInfo> FunctionInfos{FunctionInfosLog2InitSize};
 
   llvm::SmallPtrSet<const CGFunctionInfo*, 4> FunctionsBeingProcessed;
 
   /// True if we didn't layout a function due to a being inside
   /// a recursive struct conversion, set this to true.
   bool SkippedLayout;
 
+  /// True if any instance of long double types are used.
+  bool LongDoubleReferenced;
+
   /// This map keeps cache of llvm::Types and maps clang::Type to
   /// corresponding llvm::Type.
   llvm::DenseMap<const Type *, llvm::Type *> TypeCache;
 
   llvm::DenseMap<const Type *, llvm::Type *> RecordsWithOpaqueMemberPointers;
 
   static constexpr unsigned FunctionInfosLog2InitSize = 9;
   /// Helper for ConvertType.
public:  // These are internal details of CGT that shouldn't be used externally.
   /// Check if the pointer type can be zero-initialized (in the C++ sense)
   /// with an LLVM zeroinitializer.
   bool isPointerZeroInitializable(QualType T);
 
   /// IsZeroInitializable - Return whether a record type can be
   /// zero-initialized (in the C++ sense) with an LLVM zeroinitializer.
   bool isZeroInitializable(const RecordDecl *RD);
 
+  bool isLongDoubleReferenced() const { return LongDoubleReferenced; }
   bool isRecordLayoutComplete(const Type *Ty) const;
   unsigned getTargetAddressSpace(QualType T) const;
 };
 
 }  // end namespace CodeGen
 }  // end namespace clang
 
 #endif

clang/lib/CodeGen/CodeGenTypes.cpp

 using namespace clang;
 using namespace CodeGen;
 
 CodeGenTypes::CodeGenTypes(CodeGenModule &cgm)
   : CGM(cgm), Context(cgm.getContext()), TheModule(cgm.getModule()),
     Target(cgm.getTarget()), TheCXXABI(cgm.getCXXABI()),
     TheABIInfo(cgm.getTargetCodeGenInfo().getABIInfo()) {
   SkippedLayout = false;
+  LongDoubleReferenced = false;
 }
 
 CodeGenTypes::~CodeGenTypes() {
   for (llvm::FoldingSet<CGFunctionInfo>::iterator
        I = FunctionInfos.begin(), E = FunctionInfos.end(); I != E; )
     delete &*I++;
 }
 
case Type::Builtin: {
 
     case BuiltinType::Half:
       // Half FP can either be storage-only (lowered to i16) or native.
       ResultType = getTypeForFormat(
           getLLVMContext(), Context.getFloatTypeSemantics(T),
           Context.getLangOpts().NativeHalfType ||
               !Context.getTargetInfo().useFP16ConversionIntrinsics());
       break;
+    case BuiltinType::LongDouble:
+      LongDoubleReferenced = true;
+      LLVM_FALLTHROUGH;
     case BuiltinType::BFloat16:
     case BuiltinType::Float:
     case BuiltinType::Double:
-    case BuiltinType::LongDouble:
     case BuiltinType::Float128:
     case BuiltinType::Ibm128:
       ResultType = getTypeForFormat(getLLVMContext(),
                                     Context.getFloatTypeSemantics(T),
                                     /* UseNativeHalf = */ false);
       break;
 
     case BuiltinType::NullPtr:

clang/lib/CodeGen/Targets/PPC.cpp

public:
 
   int getDwarfEHStackPointer(CodeGen::CodeGenModule &M) const override {
     // This is recovered from gcc output.
     return 1; // r1 is the dedicated stack pointer
   }
 
   bool initDwarfEHRegSizeTable(CodeGen::CodeGenFunction &CGF,
                                llvm::Value *Address) const override;
+  void emitTargetMetadata(CodeGen::CodeGenModule &CGM,
+                          const llvm::MapVector<GlobalDecl, StringRef>
+                              &MangledDeclNames) const override;
 };
 
 class PPC64TargetCodeGenInfo : public TargetCodeGenInfo {
 public:
   PPC64TargetCodeGenInfo(CodeGenTypes &CGT)
       : TargetCodeGenInfo(std::make_unique<DefaultABIInfo>(CGT)) {}
 
   int getDwarfEHStackPointer(CodeGen::CodeGenModule &M) const override {

 bool
 PPC64_SVR4_TargetCodeGenInfo::initDwarfEHRegSizeTable(
   CodeGen::CodeGenFunction &CGF,
   llvm::Value *Address) const {
   return PPC_initDwarfEHRegSizeTable(CGF, Address, /*Is64Bit*/ true,
                                      /*IsAIX*/ false);
 }
 
+void PPC64_SVR4_TargetCodeGenInfo::emitTargetMetadata(
+    CodeGen::CodeGenModule &CGM,
+    const llvm::MapVector<GlobalDecl, StringRef> &MangledDeclNames) const {
+  if (CGM.getTypes().isLongDoubleReferenced()) {
+    llvm::LLVMContext &Ctx = CGM.getLLVMContext();
+    const auto *flt = &CGM.getTarget().getLongDoubleFormat();
+    if (flt == &llvm::APFloat::PPCDoubleDouble())
+      CGM.getModule().addModuleFlag(llvm::Module::Error, "float-abi",
+                                    llvm::MDString::get(Ctx, "doubledouble"));
+    else if (flt == &llvm::APFloat::IEEEquad())
+      CGM.getModule().addModuleFlag(llvm::Module::Error, "float-abi",
+                                    llvm::MDString::get(Ctx, "ieeequad"));
+    else if (flt == &llvm::APFloat::IEEEdouble())
+      CGM.getModule().addModuleFlag(llvm::Module::Error, "float-abi",
+                                    llvm::MDString::get(Ctx, "ieeedouble"));
+  }
+}
+
 bool
 PPC64TargetCodeGenInfo::initDwarfEHRegSizeTable(CodeGen::CodeGenFunction &CGF,
                                                 llvm::Value *Address) const {
   return PPC_initDwarfEHRegSizeTable(CGF, Address, /*Is64Bit*/ true,
                                      /*IsAIX*/ false);
 }
 
 std::unique_ptr<TargetCodeGenInfo>

clang/test/CodeGen/ppc64-float-abi-attr.c

+// RUN: %clang_cc1 -triple powerpc64le-unknown-linux-gnu %s -emit-llvm -o - | FileCheck %s
+// RUN: %clang_cc1 -triple powerpc64le-unknown-linux-gnu %s -emit-llvm -mabi=ieeelongdouble -o - | FileCheck %s --check-prefix=IEEE
+// RUN: %clang_cc1 -triple powerpc64le-unknown-linux-gnu %s -emit-llvm -mlong-double-64 -o - | FileCheck %s --check-prefix=LDBL64
+// RUN: %clang_cc1 -triple powerpc64le-unknown-linux-gnu %s -emit-llvm -DNOLDBL -o - | FileCheck %s --check-prefix=NOLDBL
+
+#ifndef NOLDBL
+long double foo(long double a, long double b) {
+  return a + b;
+}
+#endif
+
+int bar() { return 1; }
+
+// CHECK: ![[#]] = !{i32 1, !"float-abi", !"doubledouble"}
+// IEEE: ![[#]] = !{i32 1, !"float-abi", !"ieeequad"}
+// LDBL64: ![[#]] = !{i32 1, !"float-abi", !"ieeedouble"}
+// NOLDBL-NOT: ![[#]] = !{i32 1, !"float-abi"