[clang] the "cleanup" attribute prevents type deduction

[shynur]

/* a.cpp */
int open() { return 0; }
void close(decltype(open()) *) {}
int main() {
  auto fd [[gnu::cleanup(close)]] = open();
}

$ clang-21 a.cpp
a.cpp:4:26: error: 'cleanup' function 'close' parameter has type 'decltype(open()) *' (aka 'int *') which is incompatible with type 'auto *'
    4 |   auto fd [[gnu::cleanup(close)]] = open();
      |                          ^
1 error generated.

---

I built Clang in the main branch when its HEAD is commit cb7030dbe7f3f1947c31b3059958ff3968cc22ff

[llvmbot]

@llvm/issue-subscribers-clang-frontend

Author: Shynur (shynur)

```cpp /* a.cpp */ int open() { return 0; } void close(decltype(open()) *) {} int main() { auto fd [[gnu::cleanup(close)]] = open(); } ```

$ clang-21 a.cpp
a.cpp:4:26: error: 'cleanup' function 'close' parameter has type 'decltype(open()) *' (aka 'int *') which is incompatible with type 'auto *'
    4 |   auto fd [[gnu::cleanup(close)]] = open();
      |                          ^
1 error generated.

[cor3ntin]

@AaronBallman @to268

[to268]

It's not a new bug, I can reproduce the bug as far back as Clang 14 (at least).
Godbolt link
BTW It's more a C++ bug rather than a C one, since the decltype keyword doesn't exist in C.

[AaronBallman]

You can hit the same issue in C: https://godbolt.org/z/sqxMGvbdW

[shynur]

The key point is not decltype or typeof. Obviously Clang knows decltype(open()) is int.

type 'decltype(open()) *' (aka 'int *') which is incompatible with type 'auto *'
                                ^^^^^                                    ^^^^^^

[to268]

I'll take a look.

[shynur]

I think the problem is here (cast<VarDecl>(D)->getType()):

llvm-project/clang/lib/Sema/SemaDeclAttr.cpp

Lines 3565 to 3574 in 69527b0

	// We're currently more strict than GCC about what function types we accept.
	// If this ever proves to be a problem it should be easy to fix.
	QualType Ty = S.Context.getPointerType(cast<VarDecl>(D)->getType());
	QualType ParamTy = FD->getParamDecl(0)->getType();
	if (S.CheckAssignmentConstraints(FD->getParamDecl(0)->getLocation(),
	ParamTy, Ty) != Sema::Compatible) {
	S.Diag(Loc, diag::err_attribute_cleanup_func_arg_incompatible_type)
	<< NI.getName() << ParamTy << Ty;
	return;
	}

We should first deduce the auto-type from the RHS.

[to268]

That's what I've seen so far.

[AaronBallman]

The issue is that we are processing this attribute too early. We've processed the variable declaration, but not its initializer. So when we get to calling handleCleanupAttr(), we don't actually know the type for the VarDecl yet because we've not processed the initializer.

I think we need to have some sort of "rechecking" that happens from Sema::AddInititalizerToDecl() for any attribute which does type-based checking of the declaration it appertains to. This may be something we want to mark via TableGen in Attr.td and handle more generically, because you can run into this same kind of failure with other attributes: https://godbolt.org/z/bE41j1xx6

[shynur]

Any progress?