• Feature Name: async_closure[^rework]1
2

This RFC reworks the “async closures” section of RFC 2394

1

The original feature name was not pluralized, and though it’d be more correct in the plural, it’s probably too late to change at this point.

Summary

This RFC adds an async bound modifier to the Fn family of trait bounds. The combination currently desugars to a set of unstable AsyncFn{,Mut,Once} traits that parallel the current Fn{,Mut,Once} traits.

These traits give users the ability to express bounds for async callable types that are higher-ranked, and allow async closures to return futures which borrow from the closure’s captures.

This RFC also connects these traits to the async || {} closure syntax, as originally laid out in RFC 2394, and confirms the necessity of a first-class async closure syntax.

Motivation

Users hit two major pitfalls when writing async code that uses closures and Fn trait bounds:

  • The inability to express higher-ranked async function signatures.
  • That closures cannot return futures that borrow from the closure captures.

We’ll discuss each of these in the sections below.

Inability to express higher-ranked async function signatures

Users often employ Fn() trait bounds to write more functional code and reduce code duplication by pulling out specific logic into callbacks. When adapting these idioms into async Rust, users find that they need to split their Fn() trait bounds3 into two to account for the fact that async blocks and async functions return anonymous futures. E.g.:

3

Or return a concrete future type, like F: Fn() -> Pin<Box<dyn Future<Output = T>>>.

async fn for_each_city<F, Fut>(mut f: F)
where
    F: for<'c> FnMut(&'c str) -> Fut,
    Fut: Future<Output = ()>,
{
    for x in ["New York", "London", "Tokyo"] {
        f(x).await;
    }
}

However, when they try to call this code, users are often hit with mysterious higher-ranked lifetime errors, e.g.:

async fn do_something(city_name: &str) { todo!() }

async fn main() {
    for_each_city(do_something).await;
}
error: implementation of `FnMut` is not general enough
  --> src/main.rs
   |
   |     for_each_city(do_something);
   |     ^^^^^^^^^^^^^^^^^^^^^^^^^^^ implementation of `FnMut` is not general enough
   |
   = note: `for<'a> fn(&'a str) -> impl Future<Output = ()> {do_something}` must implement `FnMut<(&str,)>`
   = note: ...but it actually implements `FnMut<(&'0 str,)>`, for some specific lifetime `'0`

This happens because the type for the Fut generic parameter is chosen by the caller, in main, and it cannot reference the higher-ranked lifetime for<'c> in the FnMut trait bound, but the anonymous future produced by calling do_something does capture a generic lifetime parameter, and must capture it in order to use the &str argument.

Closures cannot return futures that borrow from their captures

When users wants to call a function that takes an async callback argument, they often reach for || async {} (a closure that returns an anonymous future) because closures can capture state from the local environment and are syntactically lightweight.

However, users are quickly met with the limitation that they cannot use any of the closure’s captures by reference in the async block. E.g.:

async fn parent_country(city_name: &str) -> String { todo!() }

async fn main() {
    // Collect the country names of each city in our list.
    let mut countries = vec![];
    for_each_city(|city_name| async {
        countries.push(parent_country(city_name).await);
    })
}
error: captured variable cannot escape `FnMut` closure body
  --> src/main.rs
   |
   |       let mut countries = vec![];
   |           ------------- variable defined here
   |       for_each_city(|city_name| async {
   |  _____________________________-_^
   | |                             |
   | |                             inferred to be a `FnMut` closure
   | |         countries.push(parent_country(city_name).await);
   | |         --------- variable captured here
   | |     }).await;
   | |_____^ returns an `async` block that contains a reference to a captured variable, which then escapes the closure body
   |
   = note: `FnMut` closures only have access to their captured variables while they are executing...
   = note: ...therefore, they cannot allow references to captured variables to escape

The future that is returned by the closure cannot reference any of the captures of the closure, which is a limitation that makes || async {} quite unusable today without needing to, for example, clone data and declare the async block move.

In order for this to work, the FnMut trait would need to be “lending”; however, there are complications with implementing general lending closures.

Guide Level Explanation

(note: See the naming blocking concern about async Fn* vs AsyncFn* syntax. This RFC uses the async Fn syntax for trait bounds to avoid duplicating explanations for two different proposed syntaxes, but the syntax remains an open question.)

Just as you can write functions which accept closures, you can write functions which accept async closures:

async fn takes_async_closure(f: impl async Fn(u64)) {
    f(0).await;
    f(1).await;
}

takes_async_closure(async |i| {
    core::future::ready(i).await;
    println!("done with {i}.");
});

We recommend for users to write async Fn()/async FnMut()/async FnOnce() and async || for async closures. This is more flexible than a closure returning a future for the reasons described elsewhere in this RFC.

Async closures act similarly to closures, and can have parts of their their signatures specified:

// They can have arguments annotated with types:
let arg = async |x: i32| { async_add(x, 1).await };

// They can have their return types annotated:
let ret = async || -> Vec<Id> { async_iterator.collect().await };

// They can be higher-ranked:
let hr = async |x: &str| { do_something(x).await };

// They can capture values by move:
let s = String::from("hello, world");
let c = async move || do_something(&s).await };

When called, they return an anonymous future type corresponding to the (not-yet-executed) body of the closure. These can be awaited like any other future.

The async Fn trait bound syntax can be used anywhere a trait bound is allowed, such as:

/// In return-position impl trait:
fn closure() -> impl async Fn() { async || {} }

/// In trait bounds:
trait Foo<F>: Sized
where
    F: async Fn()
{
    fn new(f: F) -> Self;
}

/// in GATs:
trait Gat {
    type AsyncHasher<T>: async Fn(T) -> i32;
}

Detailed Explanation

AsyncFn*

This RFC begins by introducing a family of AsyncFn traits for the purposes of demonstrating the lending behavior of async closures. These traits are intended to remain unstable to name or implement, just like the Fn traits. Nonetheless, we’ll describe the details of these traits so as to explain the user-facing features enabled by them.

The definition of the traits is (modulo rustc_ attributes, and the "rust-call" ABI):

[!NOTE] We omit some details about the "rust-call" calling convention and the fact that the Args parameter is enforced to be a tuple.

/// An async-aware version of the [`FnOnce`](crate::ops::FnOnce) trait.
///
/// All `async fn` and functions returning futures implement this trait.
pub trait AsyncFnOnce<Args> {
    /// Future returned by [`AsyncFnOnce::async_call_once`].
    type CallOnceFuture: Future<Output = Self::Output>;

    /// Output type of the called closure's future.
    type Output;

    /// Call the [`AsyncFnOnce`], returning a future which may move out of the called closure.
    fn async_call_once(self, args: Args) -> Self::CallOnceFuture;
}

/// An async-aware version of the [`FnMut`](crate::ops::FnMut) trait.
///
/// All `async fn` and functions returning futures implement this trait.
pub trait AsyncFnMut<Args>: AsyncFnOnce<Args> {
    /// Future returned by [`AsyncFnMut::async_call_mut`] and [`AsyncFn::async_call`].
    type CallRefFuture<'a>: Future<Output = Self::Output>
    where
        Self: 'a;

    /// Call the [`AsyncFnMut`], returning a future which may borrow from the called closure.
    fn async_call_mut(&mut self, args: Args) -> Self::CallRefFuture<'_>;
}

/// An async-aware version of the [`Fn`](crate::ops::Fn) trait.
///
/// All `async fn` and functions returning futures implement this trait.
pub trait AsyncFn<Args>: AsyncFnMut<Args> {
    /// Call the [`AsyncFn`], returning a future which may borrow from the called closure.
    fn async_call(&self, args: Args) -> Self::CallRefFuture<'_>;
}

Associated types of AsyncFn* traits are not nameable

Unlike what is true today with the current Fn* traits, this RFC reserves as an implementation detail the associated types of the AsyncFn* traits, and these will not be nameable as part of the stable interface specified by this RFC.

That is, using the existing FnOnce trait, we can write this today on stable Rust:

fn foo<F, T>()
where
    F: FnOnce() -> T,
    F::Output: Send, //~ OK
{
}

(We decided to allow this in #34365.)

However, this RFC reserves as an implementation detail the associated types of the traits specified above, so this does not work:

fn foo<F, T>()
where
    F: async FnOnce() -> T,
    F::Output: Send,
    //~^ ERROR use of unstable library feature
    F::CallOnceFuture: Send,
    //~^ ERROR use of unstable library feature
{
}

async bound modifier on Fn() trait bounds

(note: See the naming blocking concern, which reflects that this remains an open question. Repeating the blocking concern: within this RFC, we generally name the user-facing semantics of async trait bounds as async Fn*, and we use the name AsyncFn* for the internal details of the trait implementation for the purpose of demonstrating the lending behavior.)

The AsyncFn* traits specified above are nameable via a new async bound modifier that is allowed on Fn trait bounds. That is, async Fn*() -> T desugars to AsyncFn*() -> T in bounds, where Fn* is one of the three flavors of existing function traits: Fn/FnMut/FnOnce.

This RFC specifies the modification to the TraitBound nonterminal in the grammar:

Syntax TraitBound :     async? ?? ForLifetimes? TypePath
  | ( async? ?? ForLifetimes? TypePath )

note: The grammar specifies that any for<'a> higher-ranked lifetimes come after the ? trait polarity. This seems inconsistent, but should be changed independently from this RFC. There’s an open question about how to deal with the ordering problem of ?, for<'a>, and async, or if we want to separate async traits into their own production rule that enforces the right ordering of for<'a> async.

Since the grammar doesn’t distinguish parenthesized and angle-bracketed generics in _TypePath_, async as a trait bound modifier will be accepted in all trait bounds at parsing time, but it will be rejected by the compiler post-expansion if it’s not attached to a parenthesized Fn() trait bound. Similarly, the combination of async and ? is syntactically valid but semantically invalid, and will be rejected by the compiler post-expansion.

Users are able to write async Fn*() -> T trait bounds in all positions that trait bounds are allowed, for example:

fn test(f: F) where F: async Fn() -> i32 {}

fn apit(f: impl async Fn() -> i32) {}

trait Tr {
    type Callable: async Fn() -> i32;
}

// Allowed syntactically; not currently object-safe:
let _: Box<dyn async Fn()> = todo!();

When is async Fn*() implemented?

All currently-stable callable types (i.e., closures, function items, function pointers, and dyn Fn* trait objects) automatically implement async Fn*() -> T if they implement Fn*() -> Fut for some output type Fut, and Fut implements Future<Output = T>.

This is to make sure that async Fn*() trait bounds have maximum compatibility with existing callable types which return futures, such as async function items and closures which return boxed futures. Async closures also implement async Fn*(), but their relationship to this trait is detailed later in the RFC – specifically the relationship between the CallRefFuture and CallOnceFuture associated types.

These implementations are built-in, but can conceptually be understood as:

impl<F, Args, Fut, T> AsyncFnOnce<Args> for F
where
    F: FnOnce<A, Output = Fut>,
    Fut: Future<Output = T>,
{
    type Output = T;
    type CallOnceFuture = Fut;

    fn async_call_once(self, args: Args) -> Self::CallOnceFuture {
        FnOnce::call_once(self, args)
    }
}

And similarly for AsyncFnMut and AsyncFn, with the appropriate FnMut and Fn trait bounds, respectively.

NOTE: This only works currently for concrete callable types – for example, impl Fn() -> impl Future<Output = ()> does not implement impl async Fn(), due to the fact that these blanket impls do not exist in reality. This may be relaxed in the future. Users can work around this by wrapping their type in an async closure and calling it.

The reason that these implementations are built-in is because using blanket impls would cause overlap with the built-in implementation of AsyncFn* for async closures, which must have a distinct implementation to support self-borrowing futures.

Some stable types that implement async Fn() today include, e.g.:

// Async functions:
async fn foo() {}

// Functions that return a concrete future type:
fn foo() -> Pin<Box<dyn Future<Output = ()>>> { Box::pin(async {}) }

// Closures that return an async block:
let c = || async {};

Notably, we can now express higher-ranked async callback bounds:

// We could also use APIT: `mut f: impl async FnMut(&str)`.
async fn for_each_city<F>(mut f: F)
where
    F: async FnMut(&str),
//     ...which is sugar for:
//  F: for<'a> async FnMut(&'a str),
{
    for x in ["New York", "London", "Tokyo"] {
        f(x).await;
    }
}

async fn increment_city_population_db_query(city_name: &str) { todo!() }

async fn main() {
    // Works for `async fn` that is higher-ranked.
    for_each_city(increment_city_population_db_query).await;
}

Async closures

Async closures were first specified in RFC 2394. This RFC doesn’t affect them syntactically, but it does lay out new rules for how they interact with AsyncFn* traits.

Like async functions, async closures return futures which execute the code in the body of the closure. Like closures, they are allowed to capture variables from the surrounding environment if they are mentioned in the body of the closure. Each variable is captured in the most permissive capture mode allowed, and this capture analysis generally follows the same rules as closures, including allowing disjoint captures per RFC 2229.

Async closures allow self-borrows

However, since async closures return a future instead of executing their bodies directly, the future corresponding to the body must itself capture all of the closure’s captures. These are captured with the most permissive capture mode allowed, which (unless the captures are being consumed by-value) necessitates borrowing from the closure itself.

For example:

let vec: Vec<String> = vec![];

let closure = async || {
    vec.push(ready(String::from("")).await);
};

The closure captures vec with some &'closure mut Vec<String> which lives until the closure is dropped. Then every call to the closure reborrows that mutable reference &'call mut Vec<String> which lives until the future is dropped (e.g. awaited).

As another example:

let string: String = "Hello, world".into();

let closure = async move || {
    ready(&string).await;
};

The closure is marked with move, which means it takes ownership of the string by value. However, if the future also took ownership of the string from the closure, then the closure would only be callable once. This is not a problem, since according to the usage of string in the closure body, the future only needs take a reference to it to call ready. Therefore, the future captures &'call String for some lifetime which lives until the future is dropped.

Closure kind analysis

Similarly to regular closures, async closures always implement AsyncFnOnce. They additionally implement AsyncFnMut if they do not move any of their captured values, and AsyncFn if they additionally do not mutate their captured values.

Async closures unconditionally implement the (non-async) FnOnce trait. They implement FnMut and Fn if they do not move their captured values, mutate them, or borrow in the future any data from the closure. The future borrows data from the closure if the data being borrowed by the future is owned or if the borrow is mutable.

For example:

let s = String::from("hello, world");
// Implements `async Fn()` along with `FnMut` and `Fn`
// because it can copy the `&String` that it captures.
let _ = async || {
    println!("{s}");
};

let s = String::from("hello, world");
// Implements `async Fn()` but not `FnMut` or `Fn` because
// it moves and owns a value of type `String`, and therefore
// the future it returns needs to take a pointer to data
// owned by the closure.
let _ = async move || {
    println!("{s}");
};

let mut s = String::from("hello, world");
// Implements `async FnMut()` but not `FnMut` or `Fn`
// because it needs to reborrow a mutable pointer to `s`.
let _ = async move || {
    s.push('!');
};

Specifics about the AsyncFnOnce implementation, CallOnceFuture vs CallRefFuture

From a user’s perspective, it makes sense that if they have an async FnMut closure then they should be able to “call it only once” in a way that is uniform with an async FnOnce closure. This is because an FnOnce is seen as less restrictive to the callee than FnMut, and we preserve that distinction with the async Fn* trait bounds.

If the closure is inferred to be async Fn or async FnMut, then the compiler needs to synthesize an async FnOnce implementation for the closure which returns a future that doesn’t borrow any captured values from the closure, but instead moves those captured values into the future. Synthesizing a distinct future that is returned by async FnOnce is necessary because the trait consumes the closure when it is called (evident from the self receiver type in the method signature), meaning that a self-borrowing future would have references to dropped data. This is an interesting problem described in more detail in compiler-errors’ blog post written on async closures.

This is reflected in the unstable trait implementations by the fact that AsyncFnOnce::CallOnceFuture is a distinct type from AsyncFnMut::CallRefFuture. While the latter is a generic-associated-type (GAT) due to supporting self-borrows of the called async closure, the former is not, since it must own all of the captures mentioned in the async closures’ body.

For example:

let s = String::from("hello, world");

let closure = async move || {
    ready(&s);
};
// At this point, `s` is moved out of.  However, the
// allocation for `s` is still live.  It just lives as a
// captured field in `closure`.

// Manually call `AsyncFnOnce` -- this isn't stable since
// `AsyncFnOnce` isn't stable, but it's useful for the demo.
let fut = AsyncFnOnce::call_once(closure, ());
// At this point, `closure` is dropped.  However, the
// allocation for `s` is still live.  It now lives as a
// captured field in `fut`.

fut.await;
// After the future is awaited, it's dropped.  At that
// point, the allocation for `s` is dropped.

For the purposes of the compiler implementation, although these are distinct futures, they still have the same Output type (in other words, their futures await to the same type), and for types that have async Fn* implementations, the two future types execute identically, since they execute the same future body. They only differ in their captures. Given that users usually do not care about the concrete future type itself, but only its final output type, and that both futures are fully anonymous, the fact that a different future is used when calling an async FnMut via async_call_mut vs async_call_once are not noticeable except for pathological examples.

Interaction with return-type notation, naming the future returned by calling

With async Fn() -> T trait bounds, we don’t know anything about the Future returned by calling the async closure other than that it’s a Future and awaiting that future returns T.

This is not always sufficient, for example, if you want to spawn a future onto another thread:

async fn foo(x: impl async Fn(&str)) -> Result<()> {
    tokio::spawn(x("hello, world")).await
}
error[E0277]: cannot be sent between threads safely
   --> src/lib.rs
    |
    |     tokio::spawn(x("hello, world")).await
    |     ------------ ^^^^^^^^^^^^^^^^^ cannot be sent between threads safely
    |     |
    |     required by a bound introduced by this call

With the acceptance of the RTN (return-type notation) RFC 3654, this RFC specifies that users will be allowed to add RTN-like bounds to type parameters that are also bounded by async Fn(). Concretely, this bound expands to bound both CallOnceFuture and CallRefFuture (if the latter exists):

async fn foo(x: F) -> Result<()>
where
    F: async Fn(&str) -> Result<()>,
    // The future from calling `F` is `Send` and `'static`.
    F(..): Send + 'static,
    // Which expands to two bounds:
    // `for<'a> <F as AsyncFnMut>::CallRefFuture<'a>: Send`
    // `<F as AsyncFnOnce>::CallOnceFuture: Send`
    // the latter is only if `F` is bounded with `async Fn` or `async FnMut`.
{
    tokio::spawn(x("hello, world")).await
}

This bound is only valid if there is a corresponding async Fn*() trait bound.

Rationale and alternatives

Why do we need a new set of AsyncFn* traits?

As demonstrated in the motivation section, we need a set of traits that are lending in order to represent futures which borrow from the closure’s captures. This is described in more detail in a blog post written on async closures.

We technically only need to add LendingFn and LendingFnMut to our lattice of Fn* traits to support the specifics about async closures’ self-borrowing pattern, leaving us with a hierarchy of traits like so:

flowchart LR

Fn
FnMut
FnOnce
LendingFn
LendingFnMut

Fn -- isa --> FnMut
FnMut -- isa --> FnOnce

LendingFn -- isa --> LendingFnMut

Fn -- isa --> LendingFn
FnMut -- isa --> LendingFnMut

In this case, async Fn() would desugar to a LendingFnMut trait bound and a FnOnce trait bound, like:

where F: async Fn() -> i32

// is

where F: for<'s> LendingFn<LendingOutput<'s>: Future<Output = i32>> + FnOnce<Output: Future<Output = i32>>

However, there are some concrete technical implementation details that limit our ability to use LendingFn ergonomically in the compiler today. These have to do with:

  • Closure signature inference.
  • Limitations around higher-ranked trait bounds.
  • Shortcomings with error messages.

These limitations, plus the fact that the underlying trait should have no effect on the user experience of async closures and async Fn trait bounds, leads us to AsyncFn* for now. To ensure we can eventually move to these more general traits, we reserved the precise AsyncFn* trait definitions (including the associated types) as an implementation detail.

Why can’t we just use || async {}?

async || is analogous with async fn, and has an intuitive, first-class way to declare the return type of the future:

let c = async || -> i32 { 0 };

There isn’t currently a way to annotate the future’s return type in a closure that returns a future:

let c = || -> /* ??? */ async { 0 };

We could reuse impl Future to give users the ability to annotate the type of the future returned by the closure in this position, but it would require giving yet another subtly different meaning to impl Trait, since async closures return a different type when being called by-ref or by-move.

This also would have subtle limitations, e.g.:

// Easy to reanalyze as an async closure.
let _ = || async { do_stuff().await };

// Not possible to reanalyze as an async closure without a lot more work.
let _ = || {
    let fut = async { do_stuff().await };
    fut
};

Why not F: AsyncFn() -> T, naming AsyncFn* directly?

(note: See the naming blocking concern, which reflects that this remains an open question.)

Reusing the async keyword allows users to understand what an async Fn() -> T trait bound does by analogy, since they already should know that adding async to some fn foo() -> T makes it return an impl Future<Output = T> instead of the type T.

Why do we even need AsyncFnOnce?

We could desugar async FnOnce() -> T directly to FnOnce<(), Output: Future<Output = T>>. It seems overly complicated for an implementation detail, since users should never care what’s behind the AsyncFnOnce trait bound.

Why do we recommend the AsyncFnOnce::Output type remains unstable, unlike FnOnce::Output?

As mentioned above, FnOnce::Output was stabilized in #34365 as an alternative to break ecosystem code when a bug was fixed to detect usage of unstable associated items in “type-dependent” associated type paths (e.g. T::Output that is not qualified with a trait). This allows the following code on stable:

fn foo<F, T>()
where
    F: FnOnce() -> T,
    F::Output: Send, //~ OK
{
}

However, the stabilization of the assoicated type did not actually enable new things to be expressed, and instead FnOnce::Output just serves as a type alias for an existing type that may already be named.

This is because uniquely to Fn* trait bounds (compared to the other std::ops::* traits that define Output associated types, like Add), the associated type for FnOnce::Output is always constrained by the parenthesized generic syntax. In other words, given F: Fn*() -> T, F::Output can always be replaced by some type T, since T is necessary to complete the parenthesized trait bound syntax4. In that way, naming a type via the Output associated type is not more general or flexible than just naming the type itself:

4

In fact, the ::Output syntax doesn’t even make it easier to name the return type of a higher-ranked Fn trait bound either: https://godbolt.org/z/1rTGhfr9x

fn foo<F, T>()
where
    F: FnOnce() -> T,
    F::Output: Send,
    // Should just be rewritten like:
    T: Send,
{
}

Exposing the Output type for AsyncFnOnce complicates eventually moving onto other desugarings for async Fn*. For example, if AsyncFnOnce is replaced by a trait alias for FnOnce, it may change the meaning of Output in a way that would require extending the language or adding a hack into the compiler to preserve its meaning.

Given that expressivity isn’t meaningfully impaired by keeping the Output associated type as unstable, we do not expect to stabilize this associated type at the same time as async closures, and a stabilization report for the associated type should mention how it affects future possibilities to change the desugaring of async Fn*.

Drawbacks

Users might confusedly write || async {} over async || {}

Users may be confused whether to write || async {} or async || {}. The fact that async || {} has extra “superpowers” with respect to lending may lead to users hitting unnecessary errors if they invert the ordering.

We should be able to detect when users write || async {} – and subsequently hit borrow checker issues – and give a useful error message to move the async keyword. We may also lint against || async {} in code that does pass, since it’s not as expressive.

Users might write F: Fn() -> Fut, Fut: Future<Output = T> over F: async Fn() -> T

A similar problem could occur if users try to write “old style” trait bounds with two generic parameters F: Fn() -> Fut and Fut: Future<Output = T>. For example:

async fn for_each_city<F, Fut>(cb: F)
where
    F: Fn(&str) -> Fut,
    Fut: Future<Output = ()>,
{
    for x in ["New York", "London", "Tokyo"] {
        cb(x).await;
    }
}

This is problematic for two reasons:

  1. Although the Fn trait bound may be higher-ranked, the future that is returned cannot be, since we need to infer a single type parameter substitution for the Future bound to hold.
  2. There’s no way for an async closure to be lending in this case, so the expressivity of the closure is limited.

We can similarly implement a lint to detect cases where users write these two-part bounds and suggest that they instead write a single async Fn() -> T bound. This comes with the normal caveats of removing a type parameter from the function signature, e.g. semver incompatibility (since the type parameter may be turbofished). However, when users are designing a new API, they should always reach for async Fn trait bounds when they want to be generic over a closure that returns a future.

Lack of a desugaring

It’s not possible to directly name the future returned by calling some generic T: async Fn(). This means that it’s not possible, for example, to convert futures-rs’s StreamExt::then combinator, since the output future is referenced in the definition of Then returned by the combinator.

For example, consider a Then combinator that allows mapping a stream under a future:

pub struct Then<St, F, Fut, U>
where
    St: Stream,
    F: async FnMut(St::Item) -> Fut::Output,
    Fut: Future,
{
    stream: St,
    fun: F,
    future: Option</* how do we name this future type? */>,

}

The first problem here is that the RTN RFC 3654 says that RTN is only allowed in trait bound positions, so we can’t use it to name the returned future in type position, like in this struct field, without further design work.

Secondly, even if we could name the CallRefFuture type directly, we still need a lifetime to plug into the GAT. Conceptually, the future lives for the transient period of processing a single element in the stream, which isn’t representable with a lifetime argument. We would need some sort of 'unsafe or unsafe binder type.

Fixing this is a follow-up goal that we’re interested in pursuing in the near future. Design work regarding naming the future types in struct position can be done additively on top of what is exposed in this RFC, and ties into the larger question of how to use RTN in struct fields and other non-inference type positions.

Prior art

RFC 2394 described async closures at a very high level, and expressed that users would very likely want this feature eventually. This RFC confirms that suspicion.

Unresolved questions

What do we call the trait?

There is some discussion about whether to call the bound T: AsyncFn() or T: async Fn(). As stated above, there is not full consensus about whether async Fn() is the syntax we want to commit to name these bounds, but for the purposes of decoupling the fact that async Fn is the user-observable trait family, and AsyncFn is the traits of the implementation detail, this RFC names them separately.

? for<'a> and its interaction with async

Currently on nightly, we parse the async trait bound modifier along with ? (called polarity) before the for<'a> lifetime binders. This probably should get fixed so that the binder occurs on the outside of the trait, like so:

where T: for<'a> async ?Trait

(Which is semantically invalid but syntactically valid.) This is currently proposed in rust-lang/rust#127054, which should be decided before stabilization, and the stabilization report can re-confirm the correct ordering of for<'a> and async.

Where exactly is async || {} not backwards with || async {}

The stabilization report for async closures should thoroughly note any cases where rewriting || async {} into async || {} causes errors, as they will be pitfalls for adoption of async closures.

One predicted shortcoming will likely be due to corner cases of closure signature inference and pre-async-closure trait bounds in a previous section. This is not necessarily a blocker, since as the ecosystem migrates to async Fn()-style trait bounds, closure signature inference will be restored.

Future possibilities

gen Fn(), async gen Fn()

The existence of other coroutine-like modifiers, e.g. gen (RFC 3513) and async gen, suggests that we should also think about supporting these in closures and Fn() trait bounds.

This shouldn’t be too difficult to support, and we can unify these further by moving on to a general LendingFn* trait. This has some implementation concerns, but should be doable in the long term.

async bound modifier on arbitrary traits

There has been previous discussion of allowing async trait bounds on arbitrary traits, possibly based off a ?async maybe-async genericity system.

This RFC neither requires this more general extension to the language to be implemented, nor does it necessarily preclude this being an eventual possibility, since AsyncFn* remains unstable to implement.

Making async Fn() object-safe

Future work should be done to make async Fn() object-safe, so it can be used in Box, etc. E.g.:

let handlers: HashMap<Id, Box<dyn async Fn()>> = todo!();

This work will likely take a similar approach to making async fn in traits object-safe, since the major problem is how to “erase” the future returned by the async closure or callable, which differs for each implementation of the trait.

Changing the underlying definition to use LendingFn*

As mentioned above, async Fn*() trait bounds can be adjusted to desugar to LendingFn* + FnOnce trait bounds, using associated-type-bounds like:

where F: async Fn() -> i32

// desugars to

where F: for<'s> LendingFn<LendingOutput<'s>: Future<Output = i32>> + FnOnce<Output: Future<Output = i32>>

This should be doable in a way that does not affect existing code, but remain blocked on improvements to higher-ranked trait bounds around GATs. Any changes along these lines remain implementation details unless we decide separately to stabilize more user-observable aspects of the AsyncFn* trait, which is not likely to happen soon.