New `AsyncFn*` traits being `#[fundamental]` might come with significant forward-compatibility issues

[steffahn]

As far as I understand the current design of AsyncFn* traits allows them to become somewhat of a “sugar” for a more generalized approach in the future. Quoting from the RFC:

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.

In particular, the case for AsyncFnOnce seems pretty clear. It’s a future possibility (which is IMO very desirable) that AsyncFnOnce(Args…) -> R might be automatically implemented for any implementor of FnOnce(Args…) -> F that returns some F: Future<Output = R>; either by turning it into a sort of “alias” (in a way that avoids the shortcomings mentioned in the RFC) or by finding a way to re-structure the blanket impls.

In particular, IMO it’s a very relevant future possibility that Box<dyn FnOnce(Args…) -> Pin<Box<dyn Future<Output = R> [+ Send] [+Sync]>> could become an implementor of AsyncFnOnce(Args…) -> R in the future.

But #[fundamental] kills this possibility, as the following code compiles successfully, powered by the negative reasoning that #[fundamental] provides:

use std::pin::Pin;
use std::future::Future;

type BoxedFnOnceReturningBoxFuture = Box<dyn FnOnce() -> Pin<Box<dyn Future<Output = ()>>>>;

trait MyTraitNoOverlap {}
impl MyTraitNoOverlap for BoxedFnOnceReturningBoxFuture {}
impl<F: AsyncFnOnce()> MyTraitNoOverlap for F {}

(playground)

---

I was unable to find any prior discussion about the value of why these traits (AsyncFn, AsyncFnMut, AsyncFnOnce) are marked #[fundamental] in the first place. I can imagine it’s perhaps by analogy to Fn, FnMut, FnOnce being marked as such. But this decision is an important trade-off that should be considered.

As far as I can tell, there should be no issue in simply removing the #[fundamental] markers from AsyncFn, AsyncFnMut, AsyncFnOnce before they’re stabilized in 1.85. Testing this locally, std still compiles find, and UI tests pass.

It should always be backwards-compatible to add back the #[fundamental] marker later.

cc @compiler-errors

@rustbot label T-compiler, T-libs-api, F-async_closure, C-discussion

[compiler-errors]

Yeah there's no reason for them to need to be fundamental; it's not like we need it for coherence in core today. Probably just an oversight from when I initially mirrored the Fn* traits. I'll remove it.

[compiler-errors]

There's no reason to prioritize this issue tho.

@rustbot label -I-prioritize

[steffahn]

There's no reason to prioritize this issue tho.

That’s alright if there’s no need for it. I thought a future-compatibility issue on beta is a bit like a regression on beta 🤷🏻

[cuviper]

It should always be backwards-compatible to add back the #[fundamental] marker later.

Is it? At least for types, adding #[fundamental] can definitely break downstream code, like when we discussed tuples:
https://rust-lang.zulipchat.com/#narrow/channel/144729-t-types/topic/.5BCoherence.5D.20Fundamental.20tuples

... but I don't know if there's a similar risk for traits.

[compiler-errors]

For traits it's backwards-compatible to add fundamental, since it means something completely different. For types, it means that we ignore the outermost "type constructor" for the orphan check (e.g. we allow you to impl non-local traits for Box<MyLocalType>) -- adding that would open up the type for more overlap from downstream crates. For traits, it means that we assume the trait will have no new implementations added for existing types -- adding it is essentially "sealing" the trait at that point.

[cuviper]

Does that reasoning hold for the type dyn Trait as well?
(but I see AsyncFn* aren't dyn-compatible anyway)

[compiler-errors]

Only ADTs (struct/enum/union) can be fundamental (for types).