[Pre-RFC] Local Default Bounds

language design
1

Summary

This RFC proposes a mechanism for crates to define default bounds on generics, traits, and associated types. By specifying these defaults at the crate level we can reduce the need for verbose and repetitive ?Trait annotations while maintaining backward compatibility and enabling future language evolution.

Motivation

What are ?Trait bounds

Generic parameters on functions (fn foo<T>()), associated types in traits trait Foo { type Assoc; } and Self in traits trait Foo where Self ... can have where bounds.

This function expects any T that can be compared via == operator:

fn foo<T: PartialEq>(t: &T) {}

But Rust introduces some bounds by default. In the code above, T must be both PartialEq and Sized. To opt out of this, users need to write + ?Sized manually:

fn foo<T: PartialEq + ?Sized>(t: &T) {}

Use of ?Trait bounds for new features

A lot of new features (see #use-cases) require breaking old code by removing long-established assumptions like size = stride or the ability to skip the destructor of a type. To avoid breaking the code, they create a new trait representing an assumption and then define their feature as types that do not implement this trait. Here ?Trait bounds come in - old code has old assumptions, but new code can add ?Trait to opt out of them and support more types.

It is also important to note that in most cases those assumptions are not actually exercised by generic code, they are just already present in signatures - rarely code needs size = stride, or to skip the destructor (especially for a foreign type).

The problem

Quotes from "Size != Stride" Pre-RFC thread:

In order to be backwards compatible, this change requires a new implicit trait bound, applied everywhere. However, that makes this change substantially less useful. If that became the way things worked forever, then #[repr(compact)] types would be very difficult to use, as almost no generic functions would accept them. Very few functions actually need AlignSized, but every generic function would get it implicitly.

@scottmdcm

Note that every time this has come up -- ?Move, ?Pinned, etc -- the answer has been "we're not adding more of these".

What would an alternative look like that doesn't have the implicit trait bound?

In general, many abstractions can work with both Trait and !Trait types, and only a few actually require Trait. For example, Forget bound is necessary for only a few functions in std, such as forget and Box::leak, while Option can work with !Forget types too. However, if Rust were to introduce ?Forget, every generic parameter in std would need an explicit ?Forget bound. This would create excessive verbosity and does not scale well.

There is a more fundamental problem noted by @bjorn3: std would still need to have Forget bounds on all associated items of traits to maintain backward compatibility, as some code may depend on them. This makes !Forget types significantly harder to use and reduces their practicality. Fortunately, @Nadrieril proposed a solution to that problem, which resulted in that RFC.

See #guide-level-explanation for details.

Use cases

The expected outcome is an open road for new language features to enter the language in a backward-compatible way and allow users and libraries to adapt gradually.

Guide-level explanation

The syntax is to be bikeshedded, initially, it might be with a crate-level attributes.

#![default_generic_bounds(?Forget, PartialEq)]

The following example demonstrates how the compiler will understand the code. PartialEq is used just for illustration purposes. In reality, only a special set of traits would be allowed and would grow with new "breaking" traits, like Forget. PartialEq would not be one of them.

#![default_generic_bounds(?Forget, PartialEq)]

use std::ops::Deref;

trait Trait: Deref + ?PartialEq {
    type Assoc: Forget;
}

struct Qux;
struct Foo<T>(T);
struct Bar<T: ?PartialEq>(T);
struct Baz<T: Trait>(T, T::Target, T::Assoc);

impl Trait for &i32 {
    type Assoc = &'static str;
}

fn main() {
    let foo = Foo(Qux); //~ error[E0277]: the trait bound `Qux: PartialEq` is not satisfied
    let bar = Bar(Qux); // compiles as expected
    let baz = Baz(&3, 3, "assoc"); // compiles as expected
}

Code above will be observable as:

use std::ops::Deref;

trait Trait: Deref<Target: ?Forget + PartialEq>
{
    type Assoc: PartialEq;
}

struct Qux;
struct Foo<T: ?Forget + PartialEq>(T);
struct Bar<T: ?Forget + ?PartialEq>(T);
struct Baz<'a, T>(T, &'a T::Target, T::Assoc)
where
    T: ?Forget + PartialEq,
    T: Trait<Target: ?Forget + PartialEq, Assoc: Forget + PartialEq>
;

impl Trait for &i32 {
    type Assoc = &'static str;
}

fn main() {
    let foo = Foo(Qux);
    let bar = Bar(Qux);
    let baz = Baz(&3, 3, "assoc");
}

Introducing this feature is backward compatible and does not require an edition.

RFC tries to be consistent with already existing handling of Sized.

Example: Migrating to Forget

With this RFC, transitioning to Forget is straightforward for any #![forbid(unsafe)] crate:

  1. Set the appropriate bounds:
#![default_generic_bounds(?Forget)]

  1. Resolve any compilation errors by explicitly adding + Forget where needed.

  2. Optionally: Recurse into your dependencies, applying the same changes as needed.

Crates using unsafe code should beware of ptr::write and other unsafe ways of skipping destructors.

Implications on the libraries

Relax generic bound on public API

For migrated users it is equivalent to semver's minor change, while not migrated uses will observe it as patch change.

Weakening associated type bound and Self bound in traits

Bounds for associated types and Self in traits would be weakened in respect to the new traits from the start:

trait Foo: ?Trait {
    type Assoc: ?Trait;
}

This change would not be observable for not migrated crates, because default_generic_bounds would default to Trait. But if users start migrate before libraries, they will not lock them into old bounds.

#![default_generic_bounds(?Forget)]

async fn foo<T: other_crate::Trait>(bar: T) {
    let fut = bar.baz();
    // Compiler will emit an error, as `fut` maybe `!Forget`, because we set `default_generic_bounds`
    // to `?Forget`, and `default_assoc_bounds` in `other_crate` is already `?Forget`. Otherwise it
    // would have been a breaking change for `other_crate` to make future provided by `baz` `!Forget`,
    // as this code would've compiled now but not in the future.
    core::mem::forget(fut);
}

// Libary that has not migrated yet.
mod other_crate {
    trait Trait {
        async fn baz();
    }
}

Reference-level explanation

Introduce new trait level attibute: default_generic_bounds used to (non-exhaustively) enumerate overwrides of defaults for different types of bounds. Only a special set of traits would be allowed and would grow with new "breaking" traits, like Forget.

Every trait would initally have its unique default. In practice, bounds for all traits that are stable at the date of RFC except Sized would default to ?Trait. For new "breaking" traits, default would be Trait, except bounds for Self in traits and associated types in traits.

default_generic_bounds is applied for generic parameters. Effectively, it would be observable like that:

// crate `b` that has not migrated to `#![default_generic_bounds(?Forget)]`
mod b {
    fn foo<T>() {} // Observed as `T: Forget` by `b` and other crates that have not migrated.
    struct Bar<T>(T); // Observed as `T: Forget`
    // `Self` and `Qux` will be ovservable or other crates, that migrated, without `Forget` bounds
    trait Baz<T> { // Observed as `T: Forget`
        type Qux<U>; // `U` is observed as `U: Forget`
    }

    // Observed as `T: Forget`, `U: Forget`, `for<V: Forget> Baz<V>: Forget`.
    fn baz<T: Baz<U>, U>() {}

    trait Async {
        async fn method();
    }
    // Applies to RPITIT too where, so observed as `T::method(..): Forget`
    fn async_observer<T: Async>() {}

    trait DerefTrait: Deref { }

    // Associated types in generics are masked with `Forget` too.
    // So `<T as Deref<Target>>` observed as `Deref<Target: Forget>`
    fn deref_observer<T: DerefTrait>() {}

    trait RecursiveTrait {
        type Assoc: RecursiveTrait;
    }

    // All `<T as Trait>::Assoc`, `<<T as Trait>::Assoc as Trait>::Assoc`,
    // `<<<T as Trait>::Assoc as Trait>::Assoc as Trait>::Assoc` etc would be
    // observable as `: Forget`.
    // `T` is observed as `T: RecursiveTrait + Forget` too.
    fn recursive_observer<T: RecursiveTrait>() { }
}

Drawbacks

  • It may increase compilation time due to the increased complexity of trait solving.
  • It may make reading source files of crates harder, as the reader should first look at the top of the crate to see the defaults, and then remember them. It may increase cognitive load.
  • It may take some time for the ecosystem around the language to fully adapt !Trait, but it will not be a breaking change.

Rationale and alternatives

This design is simple yet powerful because it offers a backward-compatible way to evolve the language.

The impact of not accepting this RFC is that language features requiring types like !Forget, MustMove, !AlignSized and many others will not be accepted.

Add fine-grained attributes

We may have four attributes: default_generic_bounds, default_foreign_assoc_bounds, default_trait_bounds and default_assoc_bounds for more fine-grained control over defaults. For example, Sized has following defaults:

#![default_generic_bounds(Sized)]
#![default_trait_bounds(?Sized)]
#![default_assoc_bounds(Sized)]
#![default_foreign_assoc_bounds(?Sized)]

Previous version of this RFC was exactly this, you can read it here.

Alternative syntax

We may have a single macro to declare all bounds:

declare_default_bounds! { Sized, ?Forget, PartialEq };

Use similar strategy of foreign associated types defaults, but over edition

It may be possible to use the same trick over an edition for traits that we want to remove from defaults. In the case of Forget, we may set default bound for crates of edition 2024 and earlier, and lift it for editions after 2024. In terms of this RFC, it would mean that editions would have different presets of default bounds, while users would not be able to manipulate them manually.

Prior art

Unresolved questions

  • How to handle GATs? Rustc currently does not support proving for<U> <T as Trait>::Assoc<U>: Forget.
  • How to solve recursive associated type bounds? trait Trait { type Assoc: Trait }
  • Syntax
  • How to display it in Rustdoc
  • Should we allow default ! bounds? What would it mean?
  • Maybe use the term "implicit" instead of "default".
  • Should we allow Sized.
  • Maybe have 4 different attributes for more fine-grained control?

Shiny future we are working towards

Less backward compatibility burden and more freedom to fix old mistakes, and propose new exciting features.

8 Likes
[Pre-RFC] `Forget` Marker Trait
2

I think this would be a bit easier to use if editions could set a default, which this attribute then overrides. This way, the attributes for doing this will eventually go away, and not be boilerplate that continually grows over time.

For example, if in 2026 we get the AlignSized trait, then in the 2024 edition and before, all generic bounds will default to requiring it, with your proposed #![default_generic_bounds(?AlignSized)]. But then, when the 2027 edition rolls around, not requiring AlignSized could become the default, and then code can opt back into requiring it by placing a #![default_generic_bounds(AlignSized)].

2 Likes
3

I feel like people would abuse this to solve their "infectious bounds" woes. And that's a problem that needs some sort of a solution, but adding a magic attribute does not seem to me like a good one.

4

only a special set of traits would be allowed and would grow with new "breaking" traits, like Forget. PartialEq would not be one of them.

Is this paragraph related to your concern?

5

This can't be expressed as purely a use-site desugaring. Counterexample:

pub mod a {
#![default_generic_bounds(?Sized)]

pub trait Trait {
    type Assoc: Trait;
}

}

pub mod b {

use super::a::Trait;
pub fn foo<T: Trait>() {
    let _ = size_of::<<<<T as Trait>::Assoc as Trait>::Assoc as Trait>::Assoc>();
}

}

Because Trait is recursively bound, fn foo would need to bound an infinite set of types as being Sized to bound every potentially namable type. It may be possible to instead have b refer instead to effectively an alias of

trait b::Trait = a::Trait where
    Self: Sized,
    <Self as a::Trait>::Assoc: Sized + b::Trait,
;

but this requires significant additional complexity not yet touched on by the pre-RFC.

3 Likes
6

Thanks for the great critique! I believe this is an implementation detail, and I shouldn't have even included the term “desugaring” in the RFC. Yeah, it cannot be as simple as desugaring... I added this point to the unresolved questions, I think it would boil down to the exact implementation in the new trait solver if the design itself is sound and beneficial enough.

7

I've wanted something like this before for default Send bounds on async fns in traits. I think we should consider this as a direction if we can have that too. Syntax wise this is also related to the idea of specifying a default error type for try fn.

Strawman syntax for all of the above...

default trait = ?Forget;
default async<T> = Future<Output = T> + Send;
default try<T> = Result<T, anyhow::Error>;

I consider this is more of a design exploration at this point, since I don't know if it's feasible to represent in our type system.

That said, I suspect solving the "default" problem is not sufficient. We also need a way to switch modes, which an effect system may be better suited for. This might imply something like a modifier syntax: P: ?Forget Deref, meaning "P implements the ?Forget version of Deref".