[clang] inherit GD to let the codegen add kcfi type for ifunc

[aokblast]

In FreeBSD, we use ifunc to select best performance function in load time. However, the resolver is also a function itself but not been tagged kcfi by clang. The problems is caused by

if (D)
CodeGenModule::SetFunctionAttribute()

called by CodeGenModule::GetOrCreateLLVMFunction where D is empty originally.

[llvmbot]

@llvm/pr-subscribers-clang-codegen

@llvm/pr-subscribers-clang

Author: None (aokblast)

Changes

In FreeBSD, we use ifunc to select best performance function in load time. However, the resolver is also a function itself but not been tagged kcfi by clang. The problems is caused by

if (D)
CodeGenModule::SetFunctionAttribute()

called by CodeGenModule::GetOrCreateLLVMFunction where D is empty originally.

---

Full diff: https://github.com/llvm/llvm-project/pull/96400.diff

1 Files Affected:

• (modified) clang/lib/CodeGen/CodeGenModule.cpp (+1-1)

diff --git a/clang/lib/CodeGen/CodeGenModule.cpp b/clang/lib/CodeGen/CodeGenModule.cpp
index c8898ce196c1e..3e1f650884a7a 100644
--- a/clang/lib/CodeGen/CodeGenModule.cpp
+++ b/clang/lib/CodeGen/CodeGenModule.cpp
@@ -5995,7 +5995,7 @@ void CodeGenModule::emitIFuncDefinition(GlobalDecl GD) {
   llvm::Type *DeclTy = getTypes().ConvertTypeForMem(D->getType());
   llvm::Type *ResolverTy = llvm::GlobalIFunc::getResolverFunctionType(DeclTy);
   llvm::Constant *Resolver =
-      GetOrCreateLLVMFunction(IFA->getResolver(), ResolverTy, {},
+      GetOrCreateLLVMFunction(IFA->getResolver(), ResolverTy, GD,
                               /*ForVTable=*/false);
   llvm::GlobalIFunc *GIF =
       llvm::GlobalIFunc::create(DeclTy, 0, llvm::Function::ExternalLinkage,

[efriedma-quic]

Please fix buildbot failure (it looks like clang/test/CodeGen/ifunc.c is failing).

[aokblast]

I think it is not the correct way to fix this issue. GD is type for ifunc not for resolver. So the CI brokes if we do things like this.

[efriedma-quic]

Probably you want a dedicated codepath for computing the right kcfi type for resolvers.

[aokblast]

I think the calculation has nothing wrong. The problem is !kcfi_type function attribute is not appear in IR.

The problem happens because ifunc can refer resolver that havn't been defined before ifunc attribute. But LLVM Function only add attribute when it constructed first time. So there are two possible path:

In the under case, no_sanitize cannot be add because it happens at ifunc("resolver") so the unit test failed

func_t resolver() {
}

int foo(int) __attribute__ ((ifunc("resolver")));

In the under case, kcfi attribute failed to add because it happens at function definition:

int foo(int) __attribute__ ((ifunc("resolver")));

func_t resolver() {
}

I try to fix it but other error happens, here is the patch file.

[efriedma-quic]

Oh, hmm, I see.

Maybe the right strategy here is to delay attaching the resolver to the ifunc until the end of the translation unit, when we know the definition is already emitted. That way, it should already have the right attributes. (We already do some delayed checking on aliases/ifuncs anyway, in checkAliasedGlobal().)

[MaskRay]

Oh, hmm, I see.

Maybe the right strategy here is to delay attaching the resolver to the ifunc until the end of the translation unit, when we know the definition is already emitted. That way, it should already have the right attributes. (We already do some delayed checking on aliases/ifuncs anyway, in checkAliasedGlobal().)

@efriedma-quic I've tried your suggestion: move setResolver to checkAliasedGlobal. However, I've run into some difficulty. When EmitGlobal(foo) is called, how to mark resolver as eagerly emitted? (Otherwise, resolver would not be emitted at all.)

int foo(int) __attribute__ ((ifunc("resolver")));
static void *resolver(void) { return 0; }   // MustBeEmitted(Global) return false

[efriedma-quic]

The usual mechanism for emitting deferred definitions involves CodeGenModule::EmitDeferred(): declarations get added to the list by addDeferredDeclToEmit(), then it goes through to emit all the declarations on the list. So it's a matter of making sure the resolver ends up on the list. You should be able to look up the GlobalDecl from the mangled name using CodeGenModule::DeferredDecls, I think?

[MaskRay]

The usual mechanism for emitting deferred definitions involves CodeGenModule::EmitDeferred(): declarations get added to the list by addDeferredDeclToEmit(), then it goes through to emit all the declarations on the list. So it's a matter of making sure the resolver ends up on the list. You should be able to look up the GlobalDecl from the mangled name using CodeGenModule::DeferredDecls, I think?

Created #98832 , which shall fix the problem.

This problem is less about deferred emission, but a normal GetOrCreateLLVMFunction ForDefinition call prematurely returns due to llvm::Type *ResolverTy = llvm::GlobalIFunc::getResolverFunctionType(DeclTy) and (Entry->getValueType() == Ty)

[aokblast]

Wow, it helps me alot. I make the thing too complicate. I think I can close this PR. Thanks you two very much!