`#[diagnostic::on_unimplemented]` fails to trigger when certain impls are present

[joshlf]

Consider the following code:

#[diagnostic::on_unimplemented(note = "A custom Foo message!")]
trait Foo {}

#[diagnostic::on_unimplemented(note = "A custom Bar message!")]
trait Bar {}

impl<'a, T> Bar for &'a T {}

fn takes_foo<T: Foo>(_t: T) {}
fn takes_bar<T: Bar>(_t: T) {}

fn main() {
    takes_foo(());
    takes_bar(());
}

I would expect the #[diagnostic::on_unimplemented] note to be present in the error output for both calls in main. However, what I get instead is:

error[E0277]: the trait bound `(): Foo` is not satisfied
  --> src/main.rs:13:15
   |
13 |     takes_foo(());
   |     --------- ^^ the trait `Foo` is not implemented for `()`
   |     |
   |     required by a bound introduced by this call
   |
   = note: A custom Foo message!
help: this trait has no implementations, consider adding one
  --> src/main.rs:2:1
   |
2  | trait Foo {}
   | ^^^^^^^^^
note: required by a bound in `takes_foo`
  --> src/main.rs:9:17
   |
9  | fn takes_foo<T: Foo>(_t: T) {}
   |                 ^^^ required by this bound in `takes_foo`

error[E0277]: the trait bound `(): Bar` is not satisfied
  --> src/main.rs:14:15
   |
14 |     takes_bar(());
   |     --------- ^^ the trait `Bar` is not implemented for `()`
   |     |
   |     required by a bound introduced by this call
   |
note: required by a bound in `takes_bar`
  --> src/main.rs:10:17
   |
10 | fn takes_bar<T: Bar>(_t: T) {}
   |                 ^^^ required by this bound in `takes_bar`
help: consider borrowing here
   |
14 |     takes_bar(&());
   |               +

While the custom note is emitted for Foo, it's not for Bar. The culprit seems to be the impl of Bar for &T, which leads rustc down the wrong path and has it suggest only that borrowing is an option. However, that's not always the right suggestion.

cc google/zerocopy#1296, google/zerocopy#1682

[weiznich]

I think in the end this is quite similar to #125231, which means the compiler just sees the wild card impl and tries to prove that this one holds. In my opinion that's something that might be solveable by the crate author using the currently unstable #[diagonstic::do_not_recommend] on the wild card impl.

Also that sounds really similar to #130082

[joshlf]

#[diagnostic::do_not_recommend] sounds great! Do you know what the status of stabilization is? I just found your blog post about it, so I assume you're the expert 😛