Disallow implementation of UnsizedConstParamTy and ConstParamTy_ in user code

[HackingRepo]

Hello, when i try

#![feature(unsized_const_params)]

use std::marker::ConstParamTy_;
use std::marker::UnsizedConstParamTy;

#[derive(PartialEq, Eq)]
pub struct MyStruct;

impl UnsizedConstParamTy for MyStruct {}
impl ConstParamTy_ for MyStruct {}


// Allowed
pub struct Test<const PARAM: MyStruct> {}

It works, You should allow only allow ConstParamTy and UnsizedConstParamTy derive macro in nightly and not expose the internal trait to user code.

[SpriteOvO]

Note that trait UnsizedConstParamTy is removed in #145095. So in the latest nightly the example should be

#![feature(unsized_const_params)]
#![feature(adt_const_params)]

use std::marker::ConstParamTy_;

#[derive(PartialEq, Eq)]
pub struct MyStruct;

impl ConstParamTy_ for MyStruct {}

// Allowed
pub struct Test<const PARAM: MyStruct> {}

I didn't find any existing discussions on whether manual impl for ConstParamTy_ should be disallowed (sealed). (did I miss?)

[SpriteOvO]

I can't think of a convincing reason to seal it off.

In stdlib marker mod, besides this trait, we also have marker type Copy which can be used both in derived macro and implemented manually; Send Sync markers can only be implemented manually, but the compiler will auto impl them for each type if possible; and there is indeed a trait Variance which is sealed, but this is because it only allows impl for Phantom* types within stdlib.

CC @BoxyUwU

[BoxyUwU]

the ConstParamTy trait is intended to be eventually stabilized. it has use in combination with generic_const_parameter_types to allow defining const generics such as <T: ConstParamTy, const N: T>.

It's not part of the adt_const_params feature gate as stabilizing the trait would stabilize what supertraits it has. We may want to change this in the future, particularly around Sized and StructuralPartialEq (another unstable trait).

It being under the unsized_const_params feature gate is somewhat weird, it could probably do with its own feature gate, but it was simpler to put it in there than make a whole new feature 😅 how we support unsized types in const generics is also the main thing causing unknowns about the supertraits of ConstParamTy (i.e. whether it should be Sized or not)