Decide on bounds syntax for async closures (RFC 3668)

[traviscross]

Prior to stabilization, we need to decide which syntax to use for async closure bounds. This was left as an open question in RFC 3668.

The main known candidates are:

• F: async FnMut() -> T
• F: AsyncFnMut() -> T

Other candidates discussed have included:

• F: async mut fn() -> T
• F: async mut () -> T
• F: async mut || -> T
• F: async fn(&mut self) -> T
• F: AsyncFn(&mut self) -> T
• F: async |&mut self| -> T

Tracking:

• Tracking issue for #![feature(async_closure)] (RFC 2394 / RFC 3668) #62290

@rustbot labels +T-lang +I-lang-nominated

cc @compiler-errors @rust-lang/lang

[kennytm]

Why are the 3 alternatives even considered in the first place? It feels like the team is trying to setup strawmans to divert attention away from the pros and cons between async FnMut() -> T vs AsyncFnMut() -> T.

If F: async mut || -> T is valid I would have expected F: mut || -> T should also be introduced and this is so punctuation-heavy that hurts searchability and teachability. And F: async mut fn() -> T is doubly worse because fn() -> T is already a valid type not a trait. Even if these 3 are included as experiments I highly doubt if final consensus would land on them.

[compiler-errors]

I think at this point it's basically useless to consider anything other than async Fn or AsyncFn.

[nikomatsakis]

@kennytm

...It feels like the team is trying to setup strawmans to divert attention away...

I agree with you that we should just limit our focus to async Fn and AsyncFn.

However, this summary implies to me bad faith, and I don't think that's warranted. Those suggestions arose because there were some who felt that having some kind of syntax makes sense (for a variety of reasons, but mostly about forwards compatibility) but specifically didn't like async Fn. that said, one of the reasons I don't like those proposals is precisely that I'd expect them to also be used for synchronous functions, and I think that's a pretty invasive and unwarranted change.

[oriongonza]

As a user I expect that async prefacing a function desugars to the same thing but returning a future so the symmetry feels correct and more intuitive.

[casey]

I personally find AsyncFnMut() -> T to be a little more intuitive than async FnMut() -> T.

I think the reason is that the fn keyword isn't used in function types, it's Fn and friends. If a function type could be written as fn() -> T, then I think async fn() -> T would make a lot of sense, but since it's Fn() -> T, I think the non keyword version, AsyncFn() -> T is a bit more consistent.

[mehcode]

I think we need to take this and look in the future at other potential async traits.

If we go with F: async Fn() here I personally will want to see T: async Iterator<Item = i32>.

Likewise, if we go with F: AsyncFn(), I will want to see T: AsyncIterator<Item = i32>.

I'm not sure what I like more, to be honest. I can see T: async TRAIT to be a nice easy-to-understand syntax when it works, but from a beginner's point-of-view, I'd expect every trait to be able to be async by slapping async in front of the trait name.

[compiler-errors]

This thread is not the right place to have discussions. Please refer to zulip: https://rust-lang.zulipchat.com/#narrow/stream/213817-t-lang/topic/Async.20closures.20bounds.20syntax where a lot of this has been discussed at length.

This issue exists to track the requirement to decide this before stabilization, but like most tracking issues is not particularly a great place for discusions.

Please also refer to the RFC and the lang team notes where this was last discussed.