#[test] visibility changes can produce conflicts and type errors with glob use.

[djrenren]

Because #[test] marks tests as public so that they can be reexported (avoiding E0364), they can cause namespace pollutions that only occur in test builds.

Minimal repro:

mod A {
    #[test]
    fn foo() {}
}

mod B {
    pub fn foo() -> bool {
        true
    }
}

use B::foo;
use A::*;

fn conflict() {
    let x: bool = foo();
}

In a normal build, the only foo in scope is B::foo, but in a test build A::foo will shadow it. And produce the following error:

error[E0308]: mismatched types
  --> ./test.rs:20:19
   |
20 |     let x: bool = foo();
   |                   ^^^^^ expected bool, found ()
   |
   = note: expected type `bool`
              found type `()`

[djrenren]

I propose that instead of expanding:

#[test]
fn foo() {}

to:

pub fn foo(){}
pub mod __test_reexports {
   pub use super::foo;
}

we expand to:

pub mod __test_reexports {
    use super::*;
    pub fn foo() {...}
}

This is a breaking change but only for the semi-pathological case where you have test functions referencing each other. It also brings test into line with standard builds in that #[test] are not referenceable. (in debug/release due to removal, and in test due to a gensymed module).

[nrc]

the semi-pathological case where you have test functions referencing each other

I bet that's not uncommon enough to make a breaking change. However, I would expect test functions that reference each other to do so via relative paths, and this would only be breaking if absolute paths were used (iiuc). In any case we should probably just fix this and do a Crater run.

[nrc]

cc @petrochenkov and @rust-lang/compiler

We might be able to hack something in name resolution? Perhaps glob imports don't treat test functions as public? Or we give them some weird hygiene marker or something?

[djrenren]

Actually if we did:

use __test_reexports::*;

pub mod __test_reexports {
    use super::*;
    pub fn foo() {...}
}

I think we'd be compatible and safe

[djrenren]

Well, technically we should use or pub use each test in accordance with its visibility

[eddyb]

I much better prefer rust-lang/rfcs#2471 (comment) as a solution to all our testing woes. I'll maybe try to prototype it soon.

[djrenren]

Fixed by #52890