stabilize #![feature(min_const_generics)] in 1.51

[lcnr]

A new Kind
A Sort long Prophesized
Once Fragile, now Eternal

Stabilization report

This is the stabilization report for #![feature(min_const_generics)] (tracking issue #74878), a subset of #![feature(const_generics)] (tracking issue #44580), based on rust-lang/rfcs#2000.

The version target is 1.50 (2020-12-31 => beta, 2021-02-11 => stable) 1.51 (2021-02-11 => beta, 2021-03-25 => stable).

This report is a collaborative effort of @varkor, @shepmaster and @lcnr.

Summary

It is currently possible to parameterize functions, type aliases, types, traits and implementations by types and lifetimes.
With #![feature(min_const_generics)], it becomes possible, in addition, to parameterize these by constants.

This is done using the syntax const IDENT: Type in the parameter listing. Unlike full const generics, min_const_generics is limited to parameterization by integers, and constants of type char or bool.

We already use #![feature(min_const_generics)] on stable to implement many common traits for arrays. See the documentation for specific examples.

Generic const arguments, for now, are not permitted to involve computations depending on generic parameters. This means that const parameters may only be instantiated using either:

1. const expressions that do not depend on any generic parameters, e.g. { foo() + 1 }, where foo is a const fn
2. standalone const parameters, e.g. {N}

Example

#![feature(min_const_generics)]

trait Foo<const N: usize> {
    fn method<const M: usize>(&mut self, arr: [[u8; M]; N]);
}

struct Bar<T, const N: usize> {
    inner: [T; N],
}

impl<const N: usize> Foo<N> for Bar<u8, N> {
    fn method<const M: usize>(&mut self, arr: [[u8; M]; N]) {
        for (elem, s) in self.inner.iter_mut().zip(arr.iter()) {
            for &x in s {
                *elem &= x;
            }
        }
    }
}

fn function<const N: u16>() -> u16 {
    // Const parameters can be used freely inside of functions.
    (N + 1) / 2 * N
}

fn main() {
    let mut bar = Bar { inner: [0xff; 3] };
    // This infers the value of `M` from the type of the function argument.
    bar.method([[0b11_00, 0b01_00], [0b00_11, 0b00_01], [0b11_00, 0b00_11]]);
    assert_eq!(bar.inner, [0b01_00, 0b00_01, 0b00_00]);

    // You can also explicitly specify the value of `N`.
    assert_eq!(function::<17>(), 153);
}

Motivation

Rust has the built-in array type, which is parametric over a constant. Without const generics, this type can be quite cumbersome to use as it is not possible to generically implement a trait for arrays of different lengths. For example, this meant that, for a long time, the standard library only contained trait implementations for arrays up to a length of 32. This restriction has since been lifted through the use of const generics.

Const parameters allow users to naturally specify variants of a generic type which are more naturally parameterized by values, rather than by types. For example, using const generics, many of the uses of the crate typenum may now be replaced with const parameters, improving compilation time as well as code readability and diagnostics.

The subset described by min_const_generics is self-contained, but extensive enough to help with the most frequent issues: implementing traits for arrays and using arbitrarily-sized arrays inside of other types. Furthermore, it extends naturally to full const_generics once the remaining design and implementation questions have been resolved.

In-depth feature description

Declaring const parameters

Const parameters are allowed in all places where types and lifetimes are supported. They use the syntax const IDENT: Type. Currently, const parameters must be declared after lifetime and type parameters. Their scope is equal to the scope of other generic parameters. They live in the value namespace.

Type must be one of u8, u16, u32, u64, u128, usize, i8, i16, i32, i64, i128, isize, char and bool. This restriction is implemented in two places:

1. during name resolution, where we forbid generic parameters
2. during well-formedness checking, where we only allow the types listed above

The updated syntax of parameter listings is:

GenericParams:
    (OuterAttr* LifetimeParam),* (OuterAttr* TypeParam),* (OuterAttr* ConstParam),*

OuterAttr: '#[' ... ']'
LifetimeParam: ...
TypeParam: ...
ConstParam: 'const' IDENT ':' Type

Unlike type and lifetime parameters, const parameters of types can be used without being mentioned inside of a parameterized type because const parameters do not have issues concerning variance. This means that the following types are allowed:

struct Foo<const N: usize>;
enum Bar<const M: usize> { A, B }

Const arguments

Const parameters are instantiated using const arguments. Any concrete const expression or const parameter as a standalone argument can be used. When applying an expression as const parameter, most expressions must be contained within a block, with two exceptions:

1. literals and single-segment path expressions
2. array lengths

This syntactic restriction is necessary to avoid ambiguity, or requiring infinite lookahead when parsing an expression as a generic argument.

In the cases where a generic argument could be resolved as either a type or const argument, we always interpret it as a type. This causes the following test to fail:

type N = u32;
struct Foo<const N: usize>;
fn foo<const N: usize>() -> Foo<N> { todo!() } // ERR

To circumvent this, the user may wrap the const parameter with braces, at which point it is unambiguously accepted.

type N = u32;
struct Foo<const N: usize>;
fn bar<const N: usize>() -> Foo<{ N }> { todo!() } // ok

Operations depending on generic parameters are not allowed, which is enforced during well-formedness checking. Allowing generic unevaluated constants would require a way to check if they would always evaluate successfully to prevent errors that are not caught at declaration time. This ability forms part of #![feature(const_evaluatable_checked)], which is not yet being stabilised.

Since we are not yet stabilizing #![feature(lazy_normalization_consts)], we must not supply the parent generics to anonymous constants except for repeat expressions. Doing so can cause cycle errors for arrays used in where-bounds. Not supplying the parent generics can however lead to ICEs occurring before well-formedness checking when trying to use a generic parameter. See #56445 for details.

Since we expect cases like this to occur more frequently once min_const_generics is stabilized, we have chosen to forbid generic parameters in anonymous constants during name resolution. While this changes the ICE in the situation above to an ordinary error, this is theoretically a breaking change, as early-bound lifetimes were previously permitted in repeat expressions but now are disallowed, causing the following snippet to break:

fn late_bound<'a>() {
    let _ = [0; {
        let _: &'a (); // ICE ==> ERR
        3
    }];
}

fn early_bound<'a>() where &'a (): Sized {
    let _ = [0; {
        let _: &'a (); // ok ==> ERR
        3
    }];
}

Using const parameters

Const parameters can be used almost everywhere ordinary constants are allowed, except that they may not be used in the construction of consts, statics, functions, or types inside a function body and are subject to the generic argument restrictions mentioned above.

Expressions containing const parameters are eligible for promotion:

fn test<const N: usize>() -> &'static usize {
    &(3 + N)
}

Symbol mangling

See the Rust symbol name mangling RFC for an overview. Generic const parameters take the form K[type][value] when the value is known, or Kp where the value is not known, where:

• [type] is any integral type, bool, or char.
• [value] is the unsigned hex value for integers, preceded by n when negative; is 0 or 1 for bool; is the hex value for char.

Exhaustiveness checking

We do not check the exhaustiveness of impls, meaning that the following example does not compile:

struct Foo<const B: bool>;
trait Bar {}
impl Bar for Foo<true> {}
impl Bar for Foo<false> {}

fn needs_bar(_: impl Bar) {}
fn generic<const B: bool>() {
    let v = Foo::<B>;
    needs_bar(v);
}

Type inference

The value of const parameters can be inferred during typeck. One interesting case is the length of generic arrays, which can also be inferred from patterns (implemented in #70562). Practical usage of this can be seen in #76825.

Equality of constants

#![feature(min_const_generics)] only permits generic parameters to be used as standalone generic arguments. We compare two parameters to be equal if they are literally the same generic parameter.

Associated constants

Associated constants can use const parameters without restriction, see #79135 (comment) for more details.

Future work

As this is a limited subset of rust-lang/rfcs#2000, there are quite a few extensions we will be looking into next.

Lazy normalization of constants

Stabilizing #![feature(lazy_normalization_consts)] (tracking issue #72219) will remove some special cases that are currently necessary for min_const_generics, and unblocks operations on const parameters.

Relaxing ordering requirements between const and type parameters

We currently restrict the order of generic parameters so that types must come before consts. We could relax this, as is currently done with const_generics. Without this it is not possible to use both type defaults and const parameters at the same time.

Unrestricting the order will require us to improve some diagnostics that expect there to be a strict order between type and const parameters.

Allowing more parameter types

We would like to support const parameters of more types, especially&str and user-defined types. Both are blocked on valtrees. There are also open questions regarding the design of structural_match concerning the latter. Supporting generic const parameter types such as struct Foo<T, const N: T> will be a lot harder and is unlikely to be implemented in the near future.

Default values of const parameters

We do not yet support default values for const parameters. There is work in progress to enable this on nightly (see #75384).

Generic const operations

With #![feature(min_const_generics)], only concrete const expressions and parameters as standalone arguments are allowed in types and repeat expressions. However, supporting generic const operations, such as N + 1 or std::mem::size_of::<T>() is highly desirable. This feature is in early development under #![feature(const_evaluatable_checked)].

Implementation history

Many people have contributed to the design and implementation of const generics over the last three years. See #44580 (comment) for a summary. Once again thank you to everybody who helped out here!

---

r? @varkor

[mark-i-m]

I don't even have enough words to say thank you to everyone that worked on this...

[lcnr]

@bors try

[bors]

⌛ Trying commit 504ac2eda6435080272e911f9822198c9b915229 with merge 90c27759f26e0dee460c146b45d69b36d2943c22...

[bors]

☀️ Try build successful - checks-actions
Build commit: 90c27759f26e0dee460c146b45d69b36d2943c22 (90c27759f26e0dee460c146b45d69b36d2943c22)

[lcnr]

Let's start with the crater run.

@craterbot check

[craterbot]

👌 Experiment pr-79135 created and queued.
🤖 Automatically detected try build 90c27759f26e0dee460c146b45d69b36d2943c22
🔍 You can check out the queue and this experiment's details.

ℹ️ Crater is a tool to run experiments across parts of the Rust ecosystem. Learn more

[varkor]

Implementation looks fine, so this can be merged after a FCP (which is blocked on a crater run).

[newpavlov]

It could be not a right place to ask this question, but I haven't seen it discussed in github issues.

As I understand from #68436, the team is currently leaning towards making well-formedness bounds mandatory, at least for initial versions of const generics. But the example provided in the OP contains the following code:

trait Foo<const N: usize> {
    fn method<const M: usize>(&mut self, arr: [[u8; M]; N]);
}

struct Bar<T, const N: usize> {
    inner: [T; N],
}

It is obvious that using N = M = usize::MAX will result in invalid types (in the sense as discussed in the linked issue). And the version of min_const_generics proposed for stabilization allows definition of such types. Trying to instantiate Bar<u8, {usize::MAX}> will result in a "the type is too big for the current architecture" compilation error. Is it not a clear violation of the mandatory bounds approach?

In the linked issue @oli-obk has wrote:

i do expect that you will have to write almost all the same bounds as those that you write with the typenum crate

But with generic-array + typenum we have to use bounds like this:

use generic_array::{GenericArray, ArrayLength};

pub struct Bar<N: ArrayLength<u8>> {
    inner: GenericArray<u8, N>,
}

Having exceptions for mandatory bounds looks like an incoherency to me. I think that either well-formedness bounds must be mandatory for all const expressions (including simple constants) used in type construction, or that they should not be mandatory in the first place as argued by me in the linked issue, since any CTFE panic (e.g. underflow and overflow errors) will result in a compilation error either way and the "too big" errors can appear in code which uses standalone const parameters.

[oli-obk]

You are right, that clashes with what is stated in #68436, but we have opted to make an exception for arrays that are too big for the platform long ago. As you can see in https://play.rust-lang.org/?version=stable&mode=debug&edition=2018&gist=3b64dfe7c917d8016e5ca089cb03ad37, which in some form compiles since 1.0 up to the point where it recognizes that the array size is too large and aborts compilation.

What we'll allow you to do at some point is to specify via a where bound (#76560), that you are expecting a specific array type to compile successfully. Once that is in place, we may decide to start emitting lints (or maybe just have clippy do it) to encourage you to add bounds for your arrays. But even with such analyses, constructs like https://play.rust-lang.org/?version=stable&mode=debug&edition=2018&gist=f6ab8b6261decf5f651b2869b431eff1 can still cause the same error and are essentially impossible to catch without a lot of work.

Also, note that [[u8; usize::MAX]; usize::MAX] is a well formed type as long as you don't try to use an instance of it.
See https://play.rust-lang.org/?version=nightly&mode=debug&edition=2018&gist=c353a806e9965615ff00b3c1cddb1504 for an example. We do not believe that checking the size of values of types is a responsibility of the type system.

So, my summary is basically that this is nothing new and this PR does not make the situation worse, even if easier to trigger. I reopened #25116 to track the bad diagnostics of this situation.

[newpavlov]

We do not believe that checking the size of values of types is a responsibility of the type system.

I agree with this position. I had a different impression because in the #68436 discussion I've assumed that mandatory bounds will catch the "too big" error as well and no one has corrected me.

To prevent adding noise to this PR, I will try to argue my position in #76560 a bit later.