Const Generic value declarations are not required to be used in struct definition

[jamesmunns]

I tried this code:

Playground link

// Doesn't compile: Unused parameter
// pub struct Nothing<ANY>
// {
//     nothing: ()
// }

// Use of N not enforced
pub struct JustNum<const N: usize> {
    nothing: ()
}

// Use of N not enforced
pub struct Generic<ANY, const N: usize>
{
    any: ANY
}

// Use of N not enforced
impl<const N: usize> JustNum<{ N }> {
    fn new() -> Self {
        Self { nothing: () } 
    }
}

// Does not compile: Extra "M" 
// impl<const N: usize, const M: usize> JustNum<{ N }, { M }> {
//     fn new() -> Self {
//         Self { nothing: () } 
//     }
// }

I expected to see this happen:

For the following definition:

pub struct JustNum<const N: usize> {
    nothing: ()
}

I expect a compiler error, due to the unused const-generic argument

Instead, this happened:

No error was emitted

Meta

This occurs in the current beta and nightly:

• 1.51.0-beta.8 (73f48e5 2021-03-18)
• 1.52.0-nightly (f5f33ec 2021-03-19)

I'm not sure if const generic arguments are required to be used, but I could see this as an "odd" way to tag extra parameters on without them being bound to the actual impls.

[jamesmunns]

@jonas-schievink mentions this is intended behavior due to consts not having variance (here on twitter), if this is correct, I'm happy to close the issue.

I can't see anything "wrong" this could allow (other than acting as an associated const implicitly, I guess?), it was just a surprise for me.

[JulianKnodt]

I think maybe that's from a very compiler-oriented standpoint, for consts at least having it not be required to be instantiated allows for compile-time configuration, which for me personally is convenient, thinking of mainly linear algebra use-cases or passing compile time flags with structs. If they did have to be used, they would always have to result in some sort of [T; N] somewhere.

[ehuss]

It was explicitly enabled in #60058 (see #60047). It would be nice to have an example of where that is actually useful. Otherwise, I don't see why it shouldn't at least trigger an unused lint just like an unused function parameter does.

[Lokathor]

Sure, here's an example of a type that uses a const but doesn't store the const in a field: https://github.com/rust-console/voladdress/blob/main/src/volblock.rs

[jamesmunns]

Awesome, thanks for the clarification! Closing as this is intended behavior.