Async closure argument inference is broken with functions taking `FnOnce() -> futures::TryFuture`

[peku33]

I'm writing this as a continuation of #127425

While the problem from first message in issue seems to be fixed, there are still cases when this fails.

One of examples is when using try_for_each instead of for_each, when closure is expected to return Result:

#![feature(async_closure)]

use anyhow::Error;
use futures::{stream::repeat_with, StreamExt, TryStreamExt};

fn accept_str(_: &str) {}
fn accept_string(_: &String) {}

#[tokio::main(flavor = "current_thread")]
async fn main() -> Result<(), Error> {

    repeat_with(|| "foo".to_owned())
        .take(1)
        .map(Result::<_, Error>::Ok)
        .try_for_each(async move |value| {
            // error on whole closure, rust thinks that type of value is `str`
            accept_str(&value);

            // type annotations needed, cannot infer type
            accept_str(value.as_str());

            // this works
            accept_string(&value); // ok
            
            Ok(())
        })
        .await?;

    Ok(())
}

I expected to see this happen: All options used to work in past versions of rust

Instead, this happened: Without explicit type hint it fails

Version it worked on

It most recently worked on: Not 100% sure, some explanations have been made in original issue

Version with regression

rustc 1.81.0-nightly (d9284afea 2024-07-14)
binary: rustc
commit-hash: d9284afea99e0969a0e692b9e9fd61ea4ba21366
commit-date: 2024-07-14
host: x86_64-pc-windows-msvc
release: 1.81.0-nightly
LLVM version: 18.1.7

[compiler-errors]

This is because the signature of try_for_each is:

fn try_for_each<Fut, F>(self, f: F) -> TryForEach<Self, Fut, F>
where
    F: FnMut(Self::Ok) -> Fut,
    Fut: TryFuture<Ok = (), Error = Self::Error>,
    Self: Sized,

We only extract an async closure bound from cases where there is Future<Output = ???> bound on the trait, which TryFuture is not. There isn't even a supertrait bound that constrains the output type.

I also do not expect this to ever be fixed, because that would require overly complicating the closure signature inference algorithm in ways that make me quite uncomfortable¹². The correct solution is to rework the futures crate to have first-class async closure support, so for now I'd recommend just ascribing your type names. Sorry!

Footnotes

1. Firstly, this would break if another blanket implementation for TryFuture were ever added. ↩

2. Secondly, the combinator shouldn't really be using FnMut bounds either -- it should be using async FnMut bounds. ↩

[compiler-errors]

Alternatively, rust-lang/futures-rs#2763 would fix this when paired with a slight adjustment to the closure signature inference that I'd be happy to make, but it seems like that PR is blocked.