implementation of FnOnce is not general enough regressed into closure/generator type that references itself

[cynecx]

Code

I tried this code:

use core::future::Future;

trait AsyncFn<I, R>: FnMut(I) -> Self::Fut {
    type Fut: Future<Output = R>;
}

impl<F, I, R, Fut> AsyncFn<I, R> for F
where
    Fut: Future<Output = R>,
    F: FnMut(I) -> Fut,
{
    type Fut = Fut;
}

async fn call<C, R, Fn>(mut ctx: C, mut f: Fn) -> Result<R, ()>
where
    Fn: for<'a> AsyncFn<&'a mut C, Result<R, ()>>,
{
    loop {
        match f(&mut ctx).await {
            Ok(val) => return Ok(val),
            Err(_) => continue,
        }
    }
}

fn test(ctx: &mut usize) {
    call(ctx, |ctx| async move {
        let _ctx = ctx;
        Ok(1usize)
    });
}

I expected to see this happen: rustc produces implementation of `FnOnce` is not general enough

Instead, this happened: rustc produces closure/generator type that references itself

Version it worked on

1.65.0

Version with regression

1.67.0-nightly (2022-12-05 e1d8195)

@rustbot modify labels: +regression-from-stable-to-nightly -regression-untriaged

[jruderman]

Regression in nightly-2022-11-05 from 6718ea1 (#101834, @compiler-errors, @lcnr)

Bisect command used

cargo-bisect-rustc --preserve --script ./shout.sh --start 2022-08-01 --end 2022-12-05

with this shout.sh:

! rustc --edition=2021 src/main.rs 2>&1 | grep "that references itself"

[compiler-errors]

I'm not sure if I'd consider this a regression, since this neither is an error -> non-error, nor a non-error -> error... Certainly this is an interesting consequence of #101834, but we don't expect error messages to remain stable across versions.

[apiraino]

WG-prioritization assigning priority (Zulip discussion).

@rustbot label -I-prioritize +P-low

[apiraino]

@rustbot label -I-prioritize +P-low