Tracking issue for const generics (RFC 2000)

[withoutboats]

Tracking issue for rust-lang/rfcs#2000

cc @eddyb

Blocking stabilization:

• Design:
  • Resolving ordering of const and type parameters, with default parameters
  • Decide what the best UX / implementation cost balance is for unifying abstract const expressions.
  • How we determine well formedness of const expressions.
• Implementation
• Documentation
  • rustc guide

[eddyb]

#44275 added a ConstEvaluatable predicate, and WF([T; expr]) now requires ConstEvaluatable(expr), as expr is lazily evaluated (even if it's just an integer literal). Fulfilling this predicate requires the expression to evaluate successfully, while normalization ignores the error and simply leaves the Unevaluated expression it found, untouched, which is more or less what happens with associated type projections. I expect the same system to scale to const generics.

@EpicatSupercell has expressed interest in working on this, I'll mentor them through the initial implementation. However, we can't go too far because of the limitations described in #44275.

That is, we need @nikomatsakis' lazy normalization to allow constant expressions embedded in types to observe the bounds in scope (from the function / type definition / impl / etc. item they're in), without producing cyclic dependencies half the time.

[eddyb]

Implementation waypoints (for more direct mentoring, seek @eddyb on Gitter or eddyb on IRC):

• Declaration / Syntax:
  • Data structures: AST and HIR - add const parameters alongside type ones
  • Parsing: parse_generic_params - parse const IDENT: Type
  • Misc: visiting (AST, HIR), folding (AST), lowering (HIR), HIR map, DefId collector
• Declaration / Semantics
  • Data structures: ty::Generics - add const parameters alongside type ones
  • Conversion from HIR: generics_of - create ty::ConstParameterDef from hir::ConstParam
• Use / Name Resolution
  • Data structures: Def - add a variant for const parameters
  • Name resolution pass: with_type_parameter_rib - support both type and const generics
• Use / Semantics
  • ConstVal - add Param variant akin to ty::TyParam
  • Conversion from HIR: TyArray with a length that's an ExprPath that resolved to Def::ConstParam should use ConstVal::Param instead of ConstVal::Unevaluated - in a similar fashion to how Def::TyParam turns into ty::TyParam
  • subst::Kind - support &ty::Const and check as_const as well in where places where as_type and as_region are checked
• Inference
  • ConstVal - add InferVar variant akin to ty::ReVar / ty::TyInfer(TyVar(_))
  • InferCtxt - const counterparts to int_unification_table and its UnifyKey impl
  • ty::relate::{Relate,TypeRelation} - support relating ty::Const, and handle ConstVal::InferVar in a similar way to how super_combine_tys does it for types

Note that all of this should allow impl<T, const N: usize> Trait for [T; N] {...}, but not actually passing a constant expression to a type/function, e.g. ArrayVec<T, 3>.

[jplatte]

I'd like to take a stab at this :)

[samsartor]

@jplatte @eddyb Any news on this?

[jplatte]

@samsartor there was a refactoring that had to be done before the main implementation work. That is now almost done (I'm waiting for feedback currently). I don't actually know how much work there is after that. Parsing const params was what I started with initially, before the refactoring. It's not done but I should be able to get that done relatively soon.

[EdorianDark]

@jplatte It would be nice, if you could link the pull request. I was not able to find it.

[jplatte]

There is no PR yet. All my work can be found here.

[RobertWHurst]

Nice work so far @jplatte 🍻

[jplatte]

There is now a PR for the groundwork (Generics refactoring): #45930

[huhlig]

I think this is already included but it would be good to handle the C++ style folding for handling linear algebra style n dimensional arrays with varying lengths, I.E. a 4x4 [[f64;4];4]or a 4x3x5x6 dimensional array [[[[f64;6];5];3];4] and be able to properly wrap and generate specialized methods for both it AND proper trait implementations for scalars properly dimensioned vectors, etc. See https://gist.github.com/huhlig/8b21850b54a75254be4b093551f8c2cb for a rudamentary example.

[Ixrec]

I don't recall anyone proposing fold expressions for Rust before, much less as part of this RFC. But since this is Rust, is there any reason we couldn't implement fold expressions as an ordinary macro, rather than new dedicated syntax?

[mark-i-m]

#51880 is done 🎉 🎉

[eddyb]

@withoutboats @oli-obk What do you think about blocking proper unification of expressions like N + 1 (from e.g. [T; N + 1]) on getting some basic form of symbolic evaluation in miri?

I think we'd already have a mechanism to encode it, when @varkor adds ConstValue::{Infer,Param}, we can use that to track "(shallowly) valid but unknown" (symbolic) values, and then also include value (mainly integers) operations and calls on top of that.

Any non-assert control-flow decisions would still require known values, but assert itself could perhaps be reified as AssertThen(condition, assert message, success value).
(Both the condition and the success value could be symbolic!)

Being able to export the symbolic values into ty::Const means we can implement some of the unification outside miri, treating (most?) operations as opaque.

We have to be careful not to assume anything of inference variables, but for e.g. N + 1 I think we can support unifying two Add(Param(N), Bits(1)) together.

[eddyb]

@varkor A test-case I think should work with just lazy normalization, no unification:

/*const*/ fn append<const N: usize, T>(xs: [T; N - 1], x: T) -> [T; N] {
    let new = MaybeUninit::<[T; N]>::uninitialized();
    unsafe {
        let p = new.as_mut_ptr() as *mut T;
        (p as *mut _).write(xs);
        p.add(N - 1).write(x);
    }
    new.into_inner()
}

That would only work because:

• N - 1 shows up at the type level exactly once
  • everything can refer to that expression, opaquely
  • (potential workaround: type ArrayWithoutLast<T, const N: usize> = [T; N - 1];)
• it's in an argument position within the signature
  • (can't remember if return position would also work. @nikomatsakis?)
  • callers get to prove it's WF, by providing concrete values for N
  • "bubbling up" the WF requirement needs unification / ArrayWithoutLast
  • other uses would need "embedding in a where clause" e.g. [T; N - 1]: Sized
    • can even abuse where [T; N - 1]: (does that get ignored today? ouch!)
    • again bypass unification with where ArrayWithoutLast<T, N>: ...

So overall, we can probably rely on WF rules to force "validation" of const expressions onto the users, but we'd still want unification for other things (including not needing hacks like ArrayWithoutLast).

[mark-i-m]

#53645

[newpavlov]

Small question which originates from discussion of Units of Measure. What should we do if type generic over constant does not directly depend on it? For example:

struct Foo<T, const N: usize> {
    a: T,
    // Will such array be "the way" to handle this problem?
    // Or do we want some kind of `std:marker::PhantomConstData`? (which can be a simple
    // type alias to the same array)
    // Or maybe make `PhantomData` to accept constants?
    _phantom: [(), N],
}

[cuviper]

_phantom: [(), N],

I assume you meant [(); N], but this will only work for usize.

I think it would make sense to have a special marker::PhantomConst allowing any const type.

[Ekleog]

Hmm… Re-reading the RFC with the latest comments in mind, I'm wondering: would something like this be allowed?

struct Foo<T, const V: T> {
    _phantom: PhantomConst<T, V>,
}

I can't see it being banned anywhere, but would have thought it'd deserve at least an example.

[varkor]

@Ekleog: as far as I'm aware, const parameter types may not depend on type parameters initially (though it would definitely make sense for an extension). (The implementation is trickier if we allow this, so it makes sense to start with the simplest version.)

[Zauberklavier]

How is the progress on this? Do we know an approximate time when this should hit nightly?

[GabrielMajeri]

@Zauberklavier As soon as this pull request is done. To quote from it:

There's a long way to go

[eddyb]

@newpavlov I assumed constant parameters would be allowed without being used in fields.
The reason type parameters must be used is because of variance, but constants don't have this issue.

[withoutboats]

@eddyb that's what I recall from the RFC discussion also.

[Ekleog]

@varkor So this means PhantomConst proposed by @cuviper cannot exist in the current state of this RFC… right? though latest comments appear to point that there is no need for PhantomConst anyway.

[burdges]

Do const parameters impact the variance of type parameters they involve?

[eddyb]

Currently I believe const generics can't have type parameters in their type.

[rodrimati1992]

I am not clear on what the first working version of this will be able to do.

For example whether the code in this comment would compile:
rust-lang/rfcs#2581 (comment)

If that code compiled,it would be a way to enforce constraints for constants before getting a syntax for it.

[eddyb]

@rodrimati1992 You can probably just do (): IsTrue<{N < 128}> without a separate type.
I guess you want to reuse the constraint and have it actually work? (because copying the N < 128 expression won't make it work, initially)
You could use trait Lt128<const N: usize> = IsTrue<{N < 128}>;, I guess.
There's also the option of just writing where [(); 128 - N],, I think (but I'm not sure we guarantee that will panic).

[rodrimati1992]

@rodrimati1992 You can probably just do (): IsTrue<{N < 128}> without a separate type.
I guess you want to reuse the constraint and have it actually work? (because copying the N < 128 expression won't make it work, initially)
You could use trait Lt128<const N: usize> = IsTrue<{N < 128}>;, I guess.
There's also the option of just writing where [(); 128 - N],, I think (but I'm not sure we guarantee that will panic).

So,with trait aliases I could rewrite is to this?:

trait AssertLessThan128<const N:usize>=
    Assert<{N<=128}, (
        Str<"uint cannot be constructed with a size larger than 128,the passed size is:",
        Usize<N>>
    ) >;

The idea with the Assert<cond B:bool,Msg> trait is using a type error to print an error message embedded in a type,which in the case of AssertLessThan128 is:

(Str<"uint cannot be constructed with a size larger than 128,the passed size is:",Usize<N>>)

which I expect to be more helpful than where [(); 128 - N],,since it tells you in the error message why the compile-time error happened.

[eddyb]

You can also do if !(N < 128) { panic!("...") } in a constant, I think?

[rodrimati1992]

I thought the std::fmt architecture is not going to be there until const trait constraints (or something similar) arrive?

With this thing you can have error messages with multiple values in it (embedded in types).

Maybe it's better to wait for const generics to talk about this,since it'll be easier to show executable examples.

[Nemo157]

@rodrimati1992 see https://github.com/rust-lang/rfcs/blob/master/text/2345-const-panic.md, there's going to be a special case to get const panic!() before it would be possible via other features (at a guess it probably won't allow custom formatting of values at first).

[eddyb]

@rodrimati1992 Ahh, I see, that's indeed a clever trick!