Check use<..> in RPITIT for refinement by compiler-errors · Pull Request #132795 · rust-lang/rust (original) (raw)

…in-traits, r=oli-obk,traviscross

Stabilize #![feature(precise_capturing_in_traits)]

Precise capturing (+ use<> bounds) in traits - Stabilization Report

Fixes rust-lang#130044.

Stabilization summary

This report proposes the stabilization of use<> precise capturing bounds in return-position impl traits in traits (RPITITs). This completes a missing part of [RFC 3617 "Precise capturing"].

Precise capturing in traits was not ready for stabilization when the first subset was proposed for stabilization (namely, RPITs on free and inherent functions - rust-lang#127672) since this feature has a slightly different implementation, and it hadn't yet been implemented or tested at the time. It is now complete, and the type system implications of this stabilization are detailed below.

Motivation

Currently, RPITITs capture all in-scope lifetimes, according to the decision made in the "lifetime capture rules 2024" RFC. However, traits can be designed such that some lifetimes in arguments may not want to be captured. There is currently no way to express this.

Major design decisions since the RFC

No major decisions were made. This is simply an extension to the RFC that was understood as a follow-up from the original stabilization.

What is stabilized?

Users may write + use<'a, T> bounds on their RPITITs. This conceptually modifies the desugaring of the RPITIT to omit the lifetimes that we would copy over from the method. For example,

trait Foo {
    fn method<'a>(&'a self) -> impl Sized;

    // ... desugars to something like:
    type RPITIT_1<'a>: Sized;
    fn method_desugared<'a>(&'a self) -> Self::RPITIT_1<'a>;

    // ... whereas with precise capturing ...
    fn precise<'a>(&'a self) -> impl Sized + use<Self>;

    // ... desugars to something like:
    type RPITIT_2: Sized;
    fn precise_desugared<'a>(&'a self) -> Self::RPITIT_2;
}

And thus the GAT doesn't name 'a. In the compiler internals, it's not implemented exactly like this, but not in a way that users should expect to be able to observe.

Limitations on what generics must be captured

Currently, we require that all generics from the trait (including the Self) type are captured. This is because the generics from the trait are required to be invariant in order to do associated type normalization.

And like regular precise capturing bounds, all type and const generics in scope must be captured.

Thus, only the in-scope method lifetimes may be relaxed with this syntax today.

What isn't stabilized? (a.k.a. potential future work)

See section above. Relaxing the requirement to capture all type and const generics in scope may be relaxed when rust-lang#130043 is implemented, however it currently interacts with some underexplored corners of the type system (e.g. unconstrained type bivariance) so I don't expect it to come soon after.

Implementation summary

This functionality is implemented analogously to the way that opaque type precise capturing works.

Namely, we currently use variance to model the capturedness of lifetimes. However, since RPITITs are anonymous GATs instead of opaque types, we instead modify the type relation of GATs to consider variances for RPITITs (along with opaque types which it has done since rust-lang#103491).

https://github.com/rust-lang/rust/blob/30f168ef811aec63124eac677e14699baa9395bd/compiler/rustc_middle/src/ty/util.rs#L954-L976

https://github.com/rust-lang/rust/blob/30f168ef811aec63124eac677e14699baa9395bd/compiler/rustc_type_ir/src/relate.rs#L240-L244

Using variance to model capturedness is an implementation detail, and in the future it would be desirable if opaques and RPITITs simply did not include the uncaptured lifetimes in their generics. This can be changed in a forwards-compatible way, and almost certainly would not be observable by users (at least not negatively, since it may indeed fix some bugs along the way).

Tests

What other unstable features may be exposed by this feature?

I don't believe that this exposes any new unstable features indirectly.

Remaining bugs and open issues

Not aware of any open issues or bugs.

Tooling support

Rustfmt: ✅ Supports formatting + use<> everywhere.

Clippy: ✅ No support needed, unless specific clippy lints are impl'd to care for precise capturing itself.

Rustdoc: ✅ Rendering + use<> precise capturing bounds is supported.

Rust-analyzer: ✅ Parser support, and then lifetime support isn't needed rust-lang#138128 (comment) (previous: :question: There is parser support, but I am unsure of rust-analyzer's level of support for RPITITs in general.)

History

Tracking issue: rust-lang#130044