[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

9

![[#]]

[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/TargetInfo.cpp

 
   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 DefaultTargetCodeGenInfo {
 public:
   PPC64TargetCodeGenInfo(CodeGenTypes &CGT) : DefaultTargetCodeGenInfo(CGT) {}
 
   int getDwarfEHStackPointer(CodeGen::CodeGenModule &M) const override {
     // This is recovered from gcc output.

 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 {
+  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);
 }
 
 //===----------------------------------------------------------------------===//

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
+
+long double foo(long double a, long double b) {
+  return a + b;
+}
+
+// CHECK: !{{[0-9]+}} = !{i32 1, !"float-abi", !"doubledouble"}
+// IEEE: !{{[0-9]+}} = !{i32 1, !"float-abi", !"ieeequad"}
+// LDBL64: !{{[0-9]+}} = !{i32 1, !"float-abi", !"ieeedouble"}