Tracking Issue for complex generic constants: feature(generic_const_exprs)

[lcnr]

This is a tracking issue for complex generic expressions in constants which is still highly experimental.
The feature gate for the issue is #![feature(generic_const_exprs)].

Initial proposal: rust-lang/compiler-team#340
Design document: HackMD document

About tracking issues

Tracking issues are used to record the overall progress of implementation.
They are also used as hubs connecting to other relevant issues, e.g., bugs or open design questions.
A tracking issue is however not meant for large scale discussion, questions, or bug reports about a feature.
Instead, open a dedicated issue for the specific matter and add the relevant feature gate label.

Status

The design questions and implementation challenges are collected in the project-const-generics repository. The implementation is still far from ready but already available for experimentation.

[RalfJung]

The MCP mentions that

arbitrary function calls

are allowed and will be "unified". Isn't that incompatible with impure const fn such as what one will be able to write once const fn can perform heap allocations?

[RalfJung]

Also, in terms of arithmetic, it might be better to exclude floating-point arithmetic since at least some models propose to make that arithmetic non-deterministic (specifically, the NaN payload may be non-deterministic) -- also see rust-lang/unsafe-code-guidelines#237.

[lcnr]

The MCP mentions that

arbitrary function calls

are allowed and will be "unified". Isn't that incompatible with impure const fn such as what one will be able to write once const fn can perform heap allocations?

uh, now that's a challenge I didn't think of 😆 not yet sure

how do we expect to deal with [nondet_or_unpure_const_fn(); 4]?
will we rerun the the const fn 4 times?

[RalfJung]

[nondet_or_unpure_const_fn(); 4] only compiles if the return type is Copy; it will call the function once and copy the result 4 times. No CTFE is involved.

The interesting case is [const { nondet_or_unpure_const_fn() }; 4]. I expect it evaluates the function once in CTFE and use the result 4 times.

Note that the non-determinism is very restricted; it only arises in the AllocId of pointers. So maybe for the type system this simply does not matter (assuming "value trees" consist of integers only, no raw pointers). However, we have have some ptr comparison functions... they are rather weak but still, I wonder what one could do there.

[clarfonthey]

So I was looking through the current design for this feature and noticed that there isn't a clear definition of how const generics should interact with associated consts. I filed #80976 and #80977 for specific errors that show up when associated consts are used in const generics directly, but noticed that we don't actually have any defined syntax for how you would constrain associated consts in bounds, e.g. something like T: Trait<const CONST = X>. A more robust example might be a const-time matrix multiplication function:

fn mat_mul<L, R>(dest: &mut D, lhs: &L, rhs: &R)
where
    D: MatrixMut,
    L: Matrix<Scalar = D::Scalar, const ROWS = D::ROWS>,
    R: Matrix<Scalar = D::Scalar, const COLS = D::COLS, const ROWS = L::ROWS>
{
    // ...
}

I figure that this is probably something that needs to be figured out before this feature can be merged, even if the decision is that associated consts simply aren't allowed at all for MVP (which IMHO wouldn't be that great of an idea, but would still resolve the issues I mentioned).

[linclelinkpart5]

@clarfonthey This is a feature I'm in mighty need of in my personal project using const generics, I'd love to be able to use associated consts more heavily and do something like this (Cool Things I Want to Do(tm) are highlighted with !!!!!!!!):

trait Sample: Copy + Clone + PartialOrd + PartialEq {
    const EQUILIBRIUM: Self;
}

trait Frame: Copy + Clone + PartialEq {
    type Sample: Sample;

    const NUM_CHANNELS: usize;
    const EQUILIBRIUM: Self;

    // Returns a fixed array of of `Self::Sample`s of size `Self::NUM_CHANNELS`.
    fn to_array(self) -> [Self::Sample; { Self::NUM_CHANNELS }];                                    // !!!!!!!!
}

// All fixed-size arrays of samples are frames.
impl<S, const N: usize> Frame for [S; N]
where
    S: Sample,
{
    type Sample = S;

    const NUM_CHANNELS: usize = N;
    const EQUILIBRIUM: Self = [S::EQUILIBRIUM; Self::NUM_CHANNELS];                                 // !!!!!!!!

    // Already an array, just return `self`.
    // I would hope [S; N] and [S; Self::NUM_CHANNELS] are realized to be the same type!
    fn to_array(self) -> [Self::Sample; { Self::NUM_CHANNELS }] {                                   // !!!!!!!!
        self
    }
}

// Standalone function to show example of const generic bound restrictions.
fn pair_test<A, B, F>(frame_a: A, frame_b: B, mut test_func: F) -> bool
where
    A: Frame,
    // Same constant number of channels.
    B: Frame<const NUM_CHANNELS = A::NUM_CHANNELS>,                                                 // !!!!!!!!
    F: FnMut(&A::Sample, &B::Sample) -> bool,
{
    todo!("pretend this is a calculation that expects two arrays of the same size")
}

Right now however, using #[min_const_generics] on nightly, I have to resort to something like this:

trait Sample: Copy + Clone + PartialOrd + PartialEq {
    const EQUILIBRIUM: Self;
}

// This trait now has a const generic parameter.
trait Frame<const N: usize>: Copy + Clone + PartialEq {
    type Sample: Sample;

    const NUM_CHANNELS: usize;
    const EQUILIBRIUM: Self;

    fn to_array(self) -> [Self::Sample; N];
}

impl<S, const N: usize> Frame<N> for [S; N]
where
    S: Sample,
{
    type Sample = S;

    const NUM_CHANNELS: usize = N;
    const EQUILIBRIUM: Self = [S::EQUILIBRIUM; N];

    fn to_array(self) -> [Self::Sample; N] {
        self
    }
}

// Now, I have to include a `const N: usize` bound, even though this new method
// doesn't directly use its value. The value of `N` is only used to ensure that
// the two frames have the same number of channels. This kind of const generic
// pollution is especially bad when returning values parameterized by types that
// have const generics.
fn pair_test<A, B, F, const N: usize>(frame_a: A, frame_b: B, mut test_func: F) -> bool
where
    A: Frame<N>,
    B: Frame<N>, // Same constant number of channels.
    F: FnMut(&A::Sample, &B::Sample) -> bool,
{
    todo!("pretend this is a calculation that expects two arrays of the same size")
}

[liarokapisv]

I would expect that the following deduction would work and so the extra bound on the final create function wouldn't be needed. Is this out of the scope of this RFC ?

[u8 ; size_assert::<HeapStorage, T>()] : Sized =>
[u8 ; HeapStorage::Asserter<T>::VALUE as usize - 1] : Sized =>
[u8 ; AlwaysTrueAsserter::VALUE as usize - 1] : Sized =>
[u8 ; true as usize - 1] : Sized =>
[u8 ; 0] : Sized

code:

#![no_std]

#![feature(generic_associated_types)]
#![feature(const_generics)]
#![feature(const_evaluatable_checked)]
#![feature(const_fn)]

#![allow(dead_code)]
#![allow(incomplete_features)]

use core::marker::PhantomData;
use core::mem::{size_of, align_of};

trait BoolConst {
    const VALUE : bool;
}

struct AlwaysTrueAsserter {}

impl BoolConst for AlwaysTrueAsserter {
    const VALUE : bool = true;
}

struct SizeAsserter<L,R>(PhantomData<(L,R)>);

impl<S,T> BoolConst for SizeAsserter<S, T> {
    const VALUE : bool = size_of::<S>() >= size_of::<T>() && 
                         align_of::<S>() % align_of::<T>() == 0;
}

trait Storage {
    type Asserter<T> : BoolConst;
}

struct HeapStorage {}

impl Storage for HeapStorage {
    type Asserter<T> = AlwaysTrueAsserter;
}

struct RawBox<T, S: Storage>(PhantomData<(T,S)>);

const fn size_assertion<S : Storage, T>() -> usize {
    S::Asserter::<T>::VALUE as usize - 1
}

impl<T, S : Storage> RawBox<T, S> {

    fn create(_val : T) -> Self where
        [u8 ; size_assertion::<S, T>()] : Sized
    {
        Self(PhantomData)
    }

}

fn create<T>(val : T) -> RawBox::<T, HeapStorage> 
    // this here is my issue
    where  [u8 ; size_assertion::<HeapStorage, T>()] : Sized
{
    RawBox::<T, HeapStorage>::create(val)
}

(btw replacing the final RawBox::<T, HeapStorage>::create(val) expression with RawBox::create(val) leads to an ICE)

[K4rakara]

I think I might have found a bug (playground link).

The following code:

#![feature(adt_const_params)]
#![feature(const_trait_impl)]
#![feature(generic_const_exprs)]

use std::marker::PhantomData;
use std::mem;

pub struct Assert<const T: bool> {
    __prevent_contstruction: PhantomData<()>,
}

impl const IsTrue for Assert<true> { }

#[const_trait]
pub trait IsTrue { }

pub struct Foo;

impl Foo {
    pub const fn bar(&self)
    where
        Assert<{ mem::size_of_val(self) == 0 }>: IsTrue,
    {
        todo!()
    }
}

Causes the compiler to emit the following error:

error[E0424]: expected value, found module `self`
  --> src/lib.rs:22:35
   |
20 | /     pub const fn bar(&self)
21 | |     where
22 | |         Assert<{ mem::size_of_val(self) == 0 }>: IsTrue,
   | |                                   ^^^^ `self` value is a keyword only available in methods with a `self` parameter
23 | |     {
24 | |         todo!()
25 | |     }
   | |_____- this function has a `self` parameter, but a macro invocation can only access identifiers it receives from parameters

Whether you think self parameters should be allowed in generic const exprs is probably a point of contention, but I definitely don't think this should be mentioning macros in its error message.

[oli-obk]

Tracking issues have a tendency to become unmanageable. Please open a dedicated new issue and label it with F-generic_const_exprs `#![feature(generic_const_exprs)]` for absolutely any topics you want to discuss or have questions about. See rust-lang/compiler-team#739 for details and discussions on this prospective policy.