ICE from mir validation on unsizing in opendal

[jdonszelmann]

crater report: https://crater-reports.s3.amazonaws.com/pr-133502-16/try%23149edce6c6ee840adb602803b96d083101ef3592/gh/macpie.glacon/log.txt

this one is super annoying, tried to minimize but accidentally removed the ICE after about an hour, might try again soon but would be cool if someone looked at this

just compiling https://github.com/apache/opendal/tree/main/core with default feature flags triggers it.

[jdonszelmann]

minimized it to:

trait Access {
    // has to have an associated type, but can be anything
    type Reader;

    fn read(&self) -> impl Future<Output = Self::Reader> + Send {
        async { loop {} }
    }
}

trait AccessDyn: Sync {}
impl Access for dyn AccessDyn {
    type Reader = ();
}

trait Stream {
    fn poll_next(s: &'static dyn AccessDyn);
}

// has to be a function in a trait impl, can't be a normal impl block or standalone fn
impl Stream for () {
    fn poll_next(s: &'static dyn AccessDyn) {
        // new async block is important
        is_dyn_send(&async {
            s.read().await;
        });
    }
}

fn is_dyn_send(_: &dyn Send) {}

[jdonszelmann]

so: I learned something. This exact piece of code already caused an ICE once having nothing to do with the next solver. See this comment: rust-lang/rust#135470 (comment)
Also imports opendal, and the minimization clearly came from there (and looks a lot like mine)

[jdonszelmann]

which then was mentioned just yesterday on this new solver issue, likely related: #245

[lcnr]

while #245 is related and likely what causes the failure here, this is still concerning as it's a HIR typeck vs MIR borrowck mismatch 🤔

This issue shoulüd never cause ICE, only incorrect errors. It might be that check_coroutine_obligations or check_potentially_region_dependent_goals would have errored here and the mir validator just doesn't check these before ICEing 🤔 this is a more general issue and somewhat annoying. However, I also don't get why MIR borrowck wouldn't detect this then