"symbol is already defined" when combining `track_caller` with `no_mangle`

[Jules-Bertholet]

I tried this code:

#[unsafe(no_mangle)]
#[track_caller]
pub fn foo() {}

fn main() {
    let _a = foo as fn();
}

I expected to see this happen: I’m not sure, but not what actually happened

Instead, this happened:

error: symbol `foo` is already defined
 --> src/main.rs:3:1
  |
3 | pub fn foo() {}
  | ^^^^^^^^^^^^

Meta

rustc --version:

1.88.0

@rustbot label A-name-mangling A-linkage

[workingjubilee]

I believe this is because #[track_caller] secretly materializes secondary functions to call them with a modified ABI, and it is attempting to propagate the function attributes like no_mangle.

[Jules-Bertholet]

Yes. But it’s not clear what the correct behavior should be. Should the unmangled, exported symbol correspond to the caller-tracking function, or the non-tracking shim?

[workingjubilee]

I believe it has to be the "shim" version, because that is the one that actually maintains the ABI declared in the function signature, and the ABI modification is an implementation detail of track_caller.

[Jules-Bertholet]

OTOH, this compiles:

unsafe extern "Rust" {
    #[track_caller]
    fn foo();
}

So perhaps we could argue that track_caller is part of the ABI-declared function signature…

[workingjubilee]

I think we should take #[unsafe(no_mangle)] and #[track_caller] being allowed at the same time by HIR attribute checking, yet not working, as evidence that It's Not That Deep, people just didn't think about it.

But I invite you to actually separately compile a library that would use track_caller and report back what actually happens under different linkage situations if you wish to further advance that argument. If it actually works I will be amazed.