Implementation of FnOnce is not general enough for `Flatten` iterator of futures outliving an .await point

[Veetaha]

I tried this code:

let _: &(dyn Future<Output = ()> + Send) = &async {
    let _drop_after_await = std::iter::empty()
        .map(|_: &()| [async {}])
        .flatten();

    async {}.await;
};

Playground link

I expected that the code should compile without errors.

However, it produces a compile error:

error: implementation of `FnOnce` is not general enough
  --> src/main.rs:4:48
   |
4  |       let _: &(dyn Future<Output = ()> + Send) = &async {
   |  ________________________________________________^
5  | |         let _drop_after_await = std::iter::empty()
6  | |             .map(|_: &()| [async {}])
7  | |             .flatten();
8  | |
9  | |         async {}.await;
10 | |     };
   | |_____^ implementation of `FnOnce` is not general enough
   |
   = note: closure with signature `fn(&'0 ()) -> [impl Future<Output = ()>; 1]` must implement `FnOnce<(&'1 (),)>`, for any two lifetimes `'0` and `'1`...
   = note: ...but it actually implements `FnOnce<(&(),)>`

The bug doesn't reproduce if we remove the Send requirement or .flatten() combinator 🤔.

Meta

rustc -V:

rustc 1.61.0 (fe5b13d68 2022-05-18)

[traviscross]

@rustbot labels +AsyncAwait-Triaged +WG-async

We reviewed this today in WG-async triage.

We believe this is a manifestation of what's being worked on and tracked in #110338.

[georglauterbach]

Just ran into too. Very unpleasant issue because it is not obvious that the .flatten() method is actually the culprit because with Rust 1.85, if you do not have await points after the flatten iterator, there is no issue raised by the compiler :)