Encode optimized MIR of generators when emitting metadata

[tmiasko]

No description provided.

[oli-obk]

@bors r+ p=1 (regression fix)

[bors]

📌 Commit ea4cbff has been approved by oli-obk

[hellow554]

hey @tmiasko

does your PR also fix this:

#![feature(type_alias_impl_trait)]

use std::future::Future;

pub struct Task<F: Future>(F);
impl<F: Future> Task<F> {
    fn new() -> Self {
        todo!()
    }
    fn spawn(&self, _: impl FnOnce() -> F) {
        todo!()
    }
}

fn main() {
    async fn cb() {
        let a = Foo;
    }

    type F = impl Future;
    static POOL: Task<F> = Task::new();
    Task::spawn(&POOL, || cb());
}

oh, BTW: the feature flag isn't needed to trigger the ICE itself, worth investigating?

[tmiasko]

Thanks, it looks like there are multiple issues there (and this PR fixes only one of them).

[hellow554]

Ok, I posted my code in #80998 and I see you altered your initial post. Thanks!

[bors]

⌛ Testing commit ea4cbff with merge 7bb1630...

[bors]

☀️ Test successful - checks-actions
Approved by: oli-obk
Pushing 7bb1630 to master...

[rylev]

@tmiasko @oli-obk this has some fairly significant performance regressions. Is there any way to mitigate that?

[oli-obk]

Not really, these performance regressions were "gained" in #80718 (comment) due to us forgetting to encode something that is definitely necessary... at least for libraries, not sure whether it makes sense to emit most of this stuff for binaries, and not sure whether we do that at all, but it could be an easy win if we aren't doing it already.