Async fn doubles argument size

[4teap]

Generator optimization in #60187 by @tmandry reused generator locals, but arguments are still duplicated whenever used across yield points.

For example (playground):

#![feature(async_await)]

async fn wait() {}

async fn test(arg: [u8; 8192]) {
    wait().await;
    drop(arg);
}

fn main() {
    println!("{}", std::mem::size_of_val(&test([0; 8192])));
}

Expected: 8200
Actual: 16392

When passing in futures, the future size can grow exponentially (playground):

#![feature(async_await)]

async fn test(_arg: [u8; 8192]) {}

async fn use_future(fut: impl std::future::Future<Output = ()>) {
    fut.await
}

fn main() {
    println!(
        "{}",
        std::mem::size_of_val(&use_future(use_future(use_future(use_future(use_future(
            use_future(use_future(use_future(use_future(use_future(test(
                [0; 8192]
            ))))))
        ))))))
    );
}

Expected: 8236
Actual: 8396796

I didn't find any note on this. But given how common arguments are used, I think it might be useful if they are included in the optimization.

[tmandry]

Implementation notes: In the compiler, these arguments get converted to generator upvars, which are then immediately moved into locals at the top of the generator MIR. The values are afterward used from these local vars.

We don't support overlapping upvar storage. @cramertj had a suggestion which makes sense to me, which is to make upvars part of the generator Unresumed variant. Then when we move them into locals in the first resume, we'll be able to overlap those locals with the upvar storage. As long as we hold invariant that every upvar is moved into a local at the top of the generator resume function, this should always work.

I know there's some code which makes assumptions about where upvars live in closures and generators. I'm not sure how much will have to change if we do this.

cc @eddyb @mw @Zoxc

[eddyb]

@tmandry I think we can just add them as regular locals, and whether they will only be in Unresumed would depend on whether they are moved or not.
Keep in mind you can do e.g. async fn async_drop<T>(_: T) {}.

[nikomatsakis]

Check-in from async-await wg meeting: marking as "deferred" as this does not block stabilization (though it'd still be great to fix!).

[steveklabnik]

Triage: first playpen now reports 16386, second reports 8390655

[tmandry]

In theory I think we can get away with not treating upvars specially, and remapping them to locals immediately, inside the generator transform.

We should remove prefix_tys and audit any uses of upvar_tys to make sure they don't make an assumption of where upvars are stored in the generator layout. In particular, the generator layout code will need to change so it doesn't handle upvars specially anymore (the upvar types should be included as field types in GeneratorLayout now).

There are probably some details I'm forgetting, but hopefully that's enough to get started.